diff --git a/.cargo/config.toml b/.cargo/config.toml
new file mode 100644
index 000000000..8af59dd8c
--- /dev/null
+++ b/.cargo/config.toml
@@ -0,0 +1,2 @@
+[env]
+RUST_TEST_THREADS = "1"
diff --git a/.claude/rules/paddler-cache-dir.md b/.claude/rules/paddler-cache-dir.md
index 780ff411a..c6085e91e 100644
--- a/.claude/rules/paddler-cache-dir.md
+++ b/.claude/rules/paddler-cache-dir.md
@@ -5,7 +5,7 @@ paths:
 
 # Paddler Cache Directory Context
 
-- `paddler_cache_dir` is a root Paddler crate, it must not depend on any other Paddler crate
+- `paddler_cache_dir` is a root Paddler crate, it must not depend on any other Paddler crate, except dependency-free utility crates such as `paddler_sha256_hex`
 - `paddler_cache_dir` manages Paddler's global cache directory, and all its nuances
 - `paddler_cache_dir` resolves OS-related differences internally
 - `paddler_cache_dir` must use cache directory patterns idiomatic to the specific operating system
diff --git a/.claude/rules/paddler-client-python.md b/.claude/rules/paddler-client-python.md
deleted file mode 100644
index 076b143d9..000000000
--- a/.claude/rules/paddler-client-python.md
+++ /dev/null
@@ -1,12 +0,0 @@
----
-paths:
-  - "paddler_client_python/**"
----
-
-# Paddler Client Context
-
-- `paddler_client_python` provides JavaScript client that connects to `paddler_balancer`
-- It must provide a way to connect to only, specifically balancer's inference address (without the need to connect to management service at the same time)
-- It must provide a way to connect to only, specifically balancer's management address (without the need to connect to inference service at the same time)
-- Paddler client must support all Paddler's native endpoints
-- It must not implement OpenAI compatibility client
diff --git a/.claude/rules/paddler-frame-writer.md b/.claude/rules/paddler-frame-writer.md
new file mode 100644
index 000000000..d7d95b81e
--- /dev/null
+++ b/.claude/rules/paddler-frame-writer.md
@@ -0,0 +1,9 @@
+---
+paths:
+  - "paddler_frame_writer/**"
+---
+
+# Paddler Frame Writer Context
+
+- `paddler_frame_writer` is a root Paddler crate, it must not depend on any other Paddler crate
+- `paddler_frame_writer` writes queued frames to a connection's sink until the queue ends, a send fails, or the connection closes, so `paddler_agent` and `paddler_balancer` share one way to abandon the writes their peer stopped reading
diff --git a/.claude/rules/paddler-opencode-tests.md b/.claude/rules/paddler-opencode-tests.md
index ad119f43b..953fde22b 100644
--- a/.claude/rules/paddler-opencode-tests.md
+++ b/.claude/rules/paddler-opencode-tests.md
@@ -9,4 +9,4 @@ paths:
 - `paddler_opencode_tests` must reuse the cluster harness from `paddler_tests` / `paddler_test_cluster_harness` to start Paddler; it must not reimplement cluster startup.
 - The OpenCode binary path is always provided by the user via the `PADDLER_OPENCODE_BINARY` environment variable and is never guessed.
 - The end-to-end tests are gated behind both the `tests_that_use_opencode` and `tests_that_use_llms` features; when they run without `PADDLER_OPENCODE_BINARY`, they fail rather than skip.
-- The model is always `qwen3_5_0_8b` via the shared harness; only the OpenCode binary path is configurable.
+- The model is always `ministral_3_3b_reasoning` via the shared harness, since it is the smallest harness model that drives OpenCode 2's Code Mode `execute` tool reliably; only the OpenCode binary path is configurable.
diff --git a/.claude/rules/paddler-sha256-hex.md b/.claude/rules/paddler-sha256-hex.md
new file mode 100644
index 000000000..6c5fa295d
--- /dev/null
+++ b/.claude/rules/paddler-sha256-hex.md
@@ -0,0 +1,9 @@
+---
+paths:
+  - "paddler_sha256_hex/**"
+---
+
+# Paddler SHA-256 Hex Context
+
+- `paddler_sha256_hex` is a root Paddler crate, it must not depend on any other Paddler crate
+- `paddler_sha256_hex` encodes the SHA-256 digest of bytes as lowercase hexadecimal, so every crate that names something after a digest names it the same way
diff --git a/.claude/rules/paddler-spawn-resuming-panic.md b/.claude/rules/paddler-spawn-resuming-panic.md
new file mode 100644
index 000000000..7e3e8a0d7
--- /dev/null
+++ b/.claude/rules/paddler-spawn-resuming-panic.md
@@ -0,0 +1,9 @@
+---
+paths:
+  - "paddler_spawn_resuming_panic/**"
+---
+
+# Paddler Spawn Resuming Panic Context
+
+- `paddler_spawn_resuming_panic` is a root Paddler crate, it must not depend on any other Paddler crate
+- `paddler_spawn_resuming_panic` runs work on a spawned or blocking tokio task, and resumes the task's panic in the task that awaits it, so the work fails exactly as if it ran inline
diff --git a/.claude/rules/paddler-state-database.md b/.claude/rules/paddler-state-database.md
index f76f8892d..868b9d2e5 100644
--- a/.claude/rules/paddler-state-database.md
+++ b/.claude/rules/paddler-state-database.md
@@ -6,5 +6,5 @@ paths:
 # Paddler State Database Context
 
 - `paddler_state_database` persists the balancer desired state, in a file or in memory
-- `paddler_state_database` depends only on `paddler_messaging` among the Paddler crates
+- `paddler_state_database` depends only on `paddler_messaging` and the root `paddler_spawn_resuming_panic` among the Paddler crates
 - `paddler_state_database` requires absolute file paths and treats an unreadable or corrupt state file as an error instead of replacing it with defaults
diff --git a/.claude/rules/paddler-tcp-buffer-limits.md b/.claude/rules/paddler-tcp-buffer-limits.md
new file mode 100644
index 000000000..cba59be60
--- /dev/null
+++ b/.claude/rules/paddler-tcp-buffer-limits.md
@@ -0,0 +1,9 @@
+---
+paths:
+  - "paddler_tcp_buffer_limits/**"
+---
+
+# Paddler TCP Buffer Limits Context
+
+- `paddler_tcp_buffer_limits` is a test fixture crate, it must not depend on any other Paddler crate
+- `paddler_tcp_buffer_limits` reads the kernel TCP buffer limits, so tests can size the data that outlasts both socket buffers of a loopback connection instead of guessing
diff --git a/.claude/skills/running-all-tests/SKILL.md b/.claude/skills/running-all-tests/SKILL.md
index b541bb8fa..6a76f335c 100644
--- a/.claude/skills/running-all-tests/SKILL.md
+++ b/.claude/skills/running-all-tests/SKILL.md
@@ -32,27 +32,27 @@ Copy this checklist and tick each item as the suite completes:
 - [ ] JS client
 - [ ] JS client LLM
 - [ ] Python lint
-- [ ] Python client
-- [ ] Python client LLM
 - [ ] OpenAI Python client LLM
+- [ ] TypeSafe Python client LLM
+- [ ] kev converter Python LLM
 - [ ] Rust unit
 - [ ] Rust integration
+- [ ] OpenCode integration
 ```
 
-| # | Suite                    | Command (from the repo root)                          |
-|---|--------------------------|-------------------------------------------------------|
-| 1 | JS client                | `TEST_DEVICE=$DEVICE make test.client.js`             |
-| 2 | JS client LLM            | `TEST_DEVICE=$DEVICE make test.client.js.llm`         |
-| 3 | Python lint              | `make lint.client.python lint.openai.python`          |
-| 4 | Python client            | `TEST_DEVICE=$DEVICE make test.client.python`         |
-| 5 | Python client LLM        | `TEST_DEVICE=$DEVICE make test.client.python.llm`     |
-| 6 | OpenAI Python client LLM | `TEST_DEVICE=$DEVICE make test.openai.python.llm`     |
-| 7 | Rust unit                | `TEST_DEVICE=$DEVICE make test.unit`                  |
-| 8 | Rust integration         | `TEST_DEVICE=$DEVICE make test.integration`           |
-
-Run them in this order. Cheap suites (1, 3, 4, 7) surface bugs quickly; the GPU-bound suites (2, 5, 6, 8) load models.
-
-On NixOS the pinned `ruff` wheel is dynamically linked, so suite 3 needs `nix-ld`.
+| # | Suite                      | Command (from the repo root)                                                                     |
+|---|----------------------------|--------------------------------------------------------------------------------------------------|
+| 1 | JS client                  | `TEST_DEVICE=$DEVICE make test.client.js`                                                        |
+| 2 | JS client LLM              | `TEST_DEVICE=$DEVICE make test.client.js.llm`                                                    |
+| 3 | Python lint                | `make lint.openai.python lint.typesafe.python lint.kev_converter.python`                         |
+| 4 | OpenAI Python client LLM   | `TEST_DEVICE=$DEVICE make test.openai.python.llm`                                                |
+| 5 | TypeSafe Python client LLM | `TEST_DEVICE=$DEVICE make test.typesafe.python.llm`                                              |
+| 6 | kev converter Python LLM   | `make test.kev_converter.python.llm`                                                             |
+| 7 | Rust unit                  | `TEST_DEVICE=$DEVICE make test.unit`                                                             |
+| 8 | Rust integration           | `TEST_DEVICE=$DEVICE make test.integration`                                                      |
+| 9 | OpenCode integration       | `PADDLER_OPENCODE_BINARY=$(command -v opencode) TEST_DEVICE=$DEVICE make test.integration.opencode` |
+
+Run them in this order. Cheap suites (1, 3, 7) surface bugs quickly; the GPU-bound suites (2, 4, 5, 6, 8, 9) load models.
 
 ## Step 3: rules during the run
 
diff --git a/.claude/skills/running-coverage/SKILL.md b/.claude/skills/running-coverage/SKILL.md
index 607ee65f9..c74d1428c 100644
--- a/.claude/skills/running-coverage/SKILL.md
+++ b/.claude/skills/running-coverage/SKILL.md
@@ -30,15 +30,15 @@ Copy this checklist and tick each item as the suite completes:
 
 ```
 - [ ] Rust
-- [ ] Python client
+- [ ] kev converter Python
 - [ ] JS client
 ```
 
-| # | Suite         | Command (from the repo root)                           |
-|---|---------------|--------------------------------------------------------|
-| 1 | Rust          | `TEST_DEVICE=$DEVICE make test.coverage`               |
-| 2 | Python client | `TEST_DEVICE=$DEVICE make test.client.python.coverage` |
-| 3 | JS client     | `TEST_DEVICE=$DEVICE make test.client.js.coverage`     |
+| # | Suite                | Command (from the repo root)                                                            |
+|---|----------------------|-----------------------------------------------------------------------------------------|
+| 1 | Rust                 | `PADDLER_OPENCODE_BINARY=$(command -v opencode) TEST_DEVICE=$DEVICE make test.coverage` |
+| 2 | kev converter Python | `make test.kev_converter.python.llm`                                                    |
+| 3 | JS client            | `TEST_DEVICE=$DEVICE make test.client.js.coverage`                                      |
 
 ## Step 3: rules during the run
 
diff --git a/.config/nextest.toml b/.config/nextest.toml
index 731e90d13..f3b420207 100644
--- a/.config/nextest.toml
+++ b/.config/nextest.toml
@@ -3,6 +3,10 @@ serial = { max-threads = 1 }
 
 [profile.default]
 slow-timeout = { period = "10s", terminate-after = 3 }
+test-threads = 1
+
+[profile.cpu]
+test-threads = "num-cpus"
 
 [[profile.default.overrides]]
 filter = '(package(=paddler_tests) | package(=paddler_cli_tests)) and kind(test)'
diff --git a/.dockerignore b/.dockerignore
index 69782076f..a34664e22 100644
--- a/.dockerignore
+++ b/.dockerignore
@@ -1,7 +1,6 @@
 **/.coverage
 **/.mypy_cache
 **/.pytest_cache
-**/.ruff_cache
 **/.venv
 **/__pycache__
 **/dist
diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
index 40bac7256..0c268670e 100644
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -22,7 +22,7 @@ jobs:
     runs-on: ${{ matrix.os }}
     strategy:
       matrix:
-        os: [macos-latest, windows-latest]
+        os: [macos-latest]
 
     steps:
       - name: checkout code
@@ -32,10 +32,5 @@ jobs:
 
       - uses: Swatinem/rust-cache@v2
 
-      - name: Install GNU Make
-        if: runner.os == 'Windows'
-        run: choco install make --version=4.4.1 --no-progress
-
       - name: Build the cli binary
         run: make target/debug/paddler
-        shell: bash
diff --git a/.github/workflows/release.yml b/.github/workflows/release.yml
index de496e26c..f5e5b2e00 100644
--- a/.github/workflows/release.yml
+++ b/.github/workflows/release.yml
@@ -46,13 +46,11 @@ jobs:
 
       - name: build the cli binary
         run: make target/release/paddler
-        shell: bash
         env:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
 
       - name: rename the cli binary
         run: mv ./target/release/paddler ./${{ matrix.artifact_name }}
-        shell: bash
 
       - name: assert cli version matches tag
         uses: intentee/release-action/assert-version@8125f727e97f72bfef24299d7c6d3217af1453bc
@@ -66,61 +64,9 @@ jobs:
           name: ${{ matrix.artifact_name }}
           path: ./${{ matrix.artifact_name }}
 
-  build-gui:
-    name: build-gui
-    needs: resolve-version
-    runs-on: ${{ matrix.os }}
-    strategy:
-      matrix:
-        include:
-          - os: ubuntu-22.04
-            artifact_name: paddler-gui-bin-linux-libc2.35-amd64
-
-          - os: ubuntu-latest
-            artifact_name: paddler-gui-bin-linux-libc2.39-amd64
-
-          - os: macos-latest
-            artifact_name: paddler-gui-bin-macos-arm
-
-    steps:
-      - name: checkout code
-        uses: actions/checkout@v4
-
-      - name: install paddler_gui system dependencies (linux)
-        if: runner.os == 'Linux'
-        run: |
-          sudo apt-get update
-          sudo apt-get install -y \
-            libvulkan-dev libxkbcommon-dev libwayland-dev libgl-dev \
-            libx11-dev libxcursor-dev libxrandr-dev libxi-dev \
-            libxcb1-dev libfontconfig-dev pkg-config
-        shell: bash
-
-      - name: build the gui binary
-        run: make target/release/paddler_gui
-        shell: bash
-        env:
-          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-
-      - name: rename the gui binary
-        run: mv ./target/release/paddler_gui ./${{ matrix.artifact_name }}
-        shell: bash
-
-      - name: assert gui version matches tag
-        uses: intentee/release-action/assert-version@8125f727e97f72bfef24299d7c6d3217af1453bc
-        with:
-          filename: "./${{ matrix.artifact_name }}"
-          expected-version: "${{ needs.resolve-version.outputs.version }}"
-
-      - name: upload gui artifact
-        uses: actions/upload-artifact@v4
-        with:
-          name: ${{ matrix.artifact_name }}
-          path: ./${{ matrix.artifact_name }}
-
   release:
     name: release
-    needs: [resolve-version, build-cli, build-gui]
+    needs: [resolve-version, build-cli]
     runs-on: ubuntu-latest
     permissions:
       contents: write
diff --git a/.github/workflows/test.yml b/.github/workflows/test.yml
index 6497c8006..cf79b21f4 100644
--- a/.github/workflows/test.yml
+++ b/.github/workflows/test.yml
@@ -108,19 +108,15 @@ jobs:
       - name: Install Poetry
         run: pip install poetry==2.4.3
 
-      - name: Lint the Python client
-        run: make lint.client.python
-
       - name: Lint the OpenAI Python client suite
         run: make lint.openai.python
 
-  python:
-    name: python (${{ matrix.python-version }})
-    needs: paddler-binary
+      - name: Lint the TypeSafe Python client suite
+        run: make lint.typesafe.python
+
+  kev-converter-lint:
+    name: kev converter lint
     runs-on: ubuntu-latest
-    strategy:
-      matrix:
-        python-version: ["3.11", "3.13"]
 
     steps:
       - name: checkout code
@@ -129,21 +125,10 @@ jobs:
       - name: Set up Python
         uses: actions/setup-python@v5
         with:
-          python-version: ${{ matrix.python-version }}
+          python-version: "3.13"
 
       - name: Install Poetry
         run: pip install poetry==2.4.3
 
-      - name: Download the Paddler binary
-        uses: actions/download-artifact@v4
-        with:
-          name: paddler-binary
-          path: target/debug
-
-      - name: Mark the downloaded Paddler binary as built and executable
-        run: |
-          chmod +x target/debug/paddler
-          touch target/debug/paddler
-
-      - name: Run Python client tests
-        run: make test.client.python
+      - name: Lint the kev converter
+        run: make lint.kev_converter.python
diff --git a/Cargo.lock b/Cargo.lock
index 2bb5989e9..5566d095c 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -2,30 +2,14 @@
 # It is not intended for manual editing.
 version = 4
 
-[[package]]
-name = "ab_glyph"
-version = "0.2.32"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "01c0457472c38ea5bd1c3b5ada5e368271cb550be7a4ca4a0b4634e9913f6cc2"
-dependencies = [
- "ab_glyph_rasterizer",
- "owned_ttf_parser",
-]
-
-[[package]]
-name = "ab_glyph_rasterizer"
-version = "0.1.10"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "366ffbaa4442f4684d91e2cd7c5ea7c4ed8add41959a31447066e279e432b618"
-
 [[package]]
 name = "actix-codec"
-version = "0.5.2"
+version = "0.5.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5f7b0a21988c1bf877cf4759ef5ddaac04c1c9fe808c9142ecb78ba97d97a28a"
+checksum = "4c13df95297bcf9014dc89162b0cc69431e192e34e3b419612fc124cfcd45dbf"
 dependencies = [
- "bitflags 2.11.0",
- "bytes 1.11.1",
+ "bitflags 2.13.2",
+ "bytes",
  "futures-core",
  "futures-sink",
  "memchr",
@@ -37,9 +21,9 @@ dependencies = [
 
 [[package]]
 name = "actix-cors"
-version = "0.7.1"
+version = "0.7.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "daa239b93927be1ff123eebada5a3ff23e89f0124ccb8609234e5103d5a5ae6d"
+checksum = "2aff07ada3254fc02618cb7850da91dceb23b5dbda53c6676ccbb28ba504f150"
 dependencies = [
  "actix-utils",
  "actix-web",
@@ -52,25 +36,23 @@ dependencies = [
 
 [[package]]
 name = "actix-http"
-version = "3.12.1"
+version = "3.19.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "93acb4a42f64936f9b8cae4a433b237599dd6eb6ed06124eb67132ef8cc90662"
+checksum = "b796ba3c974e5133f6ca1a2c4e97c5941ea1a46d202df83344afaa18219f1a62"
 dependencies = [
  "actix-codec",
- "actix-rt",
  "actix-service",
  "actix-utils",
- "base64",
- "bitflags 2.11.0",
+ "base64 0.22.1",
+ "bitflags 2.13.2",
  "brotli",
- "bytes 1.11.1",
+ "bytes",
  "bytestring",
  "derive_more",
  "encoding_rs",
  "flate2",
- "foldhash 0.1.5",
+ "foldhash",
  "futures-core",
- "h2 0.3.27",
  "http 0.2.12",
  "httparse",
  "httpdate",
@@ -80,7 +62,7 @@ dependencies = [
  "mime",
  "percent-encoding",
  "pin-project-lite",
- "rand 0.10.1",
+ "rand 0.10.3",
  "sha1 0.11.0",
  "smallvec",
  "tokio",
@@ -91,12 +73,13 @@ dependencies = [
 
 [[package]]
 name = "actix-macros"
-version = "0.2.4"
+version = "0.2.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e01ed3140b2f8d422c68afa1ed2e85d996ea619c988ac834d255db32138655cb"
+checksum = "367f814ad4afbac74f07df5001214da65f65e185c90ef56c4dd8df23f8695b9b"
 dependencies = [
+ "proc-macro2",
  "quote",
- "syn",
+ "syn 3.0.9",
 ]
 
 [[package]]
@@ -120,16 +103,15 @@ version = "2.15.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "e5f794807f82bbd36430c12cd600c73bbab0f52fdde4f0ed49978df113f4807f"
 dependencies = [
- "actix-macros",
  "futures-core",
  "tokio",
 ]
 
 [[package]]
 name = "actix-server"
-version = "2.9.7"
+version = "2.10.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8db2cf6f034e227c0c79825fbf11f3cad48569109515f62542b84d8abedb9376"
+checksum = "5e296908c4d6f3a191e6b9873dbdfec8023350a89129d2b470f61d17233931ee"
 dependencies = [
  "actix-rt",
  "actix-service",
@@ -153,9 +135,9 @@ dependencies = [
 
 [[package]]
 name = "actix-utils"
-version = "3.0.1"
+version = "3.0.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "88a1dcdff1466e3c2488e1cb5c36a71822750ad43839937f85d2f4d9f8b705d8"
+checksum = "0128396dd7313f697ad05b21b1a7be7d4cbb81888704f55996e4a27db196bb4d"
 dependencies = [
  "local-waker",
  "pin-project-lite",
@@ -163,9 +145,9 @@ dependencies = [
 
 [[package]]
 name = "actix-web"
-version = "4.13.0"
+version = "4.18.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ff87453bc3b56e9b2b23c1cc0b1be8797184accf51d2abe0f8a33ec275d316bf"
+checksum = "81f8da0d2ae4200dde8773d72ba47a65c220b251ba7b9248db5bd263c247e959"
 dependencies = [
  "actix-codec",
  "actix-http",
@@ -176,16 +158,16 @@ dependencies = [
  "actix-service",
  "actix-utils",
  "actix-web-codegen",
- "bytes 1.11.1",
+ "bytes",
  "bytestring",
  "cfg-if",
  "cookie",
  "derive_more",
  "encoding_rs",
- "foldhash 0.1.5",
+ "foldhash",
  "futures-core",
  "futures-util",
- "impl-more 0.1.9",
+ "impl-more",
  "itoa",
  "language-tags",
  "log",
@@ -200,50 +182,50 @@ dependencies = [
  "smallvec",
  "socket2",
  "time",
+ "tokio",
  "tracing",
  "url",
 ]
 
 [[package]]
 name = "actix-web-codegen"
-version = "4.3.0"
+version = "4.5.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f591380e2e68490b5dfaf1dd1aa0ebe78d84ba7067078512b4ea6e4492d622b8"
+checksum = "f7c9f4c52b954a73c76757a3952d3437fca00a2fbcb49704f8a6ea9f767e73f0"
 dependencies = [
  "actix-router",
  "proc-macro2",
  "quote",
- "syn",
+ "syn 3.0.9",
 ]
 
 [[package]]
 name = "actix-web-lab"
-version = "0.26.0"
+version = "0.29.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1a89606b86fd2f8c0ddd72ef0be171e58c919ee0b6914aa7b52354063e5ee361"
+checksum = "1d634f96d3f94f293999b6acb4bd3c98b00e6f598be7a9b5649a35aecb90de1a"
 dependencies = [
  "actix-http",
  "actix-router",
  "actix-service",
- "actix-utils",
  "actix-web",
  "actix-web-lab-derive",
  "ahash",
  "arc-swap",
- "bytes 1.11.1",
+ "bytes",
  "bytestring",
  "csv",
  "derive_more",
  "form_urlencoded",
- "futures-core",
  "futures-util",
  "http 0.2.12",
- "impl-more 0.3.1",
- "itertools 0.14.0",
+ "impl-more",
+ "itertools 0.15.0",
  "local-channel",
  "mime",
  "pin-project-lite",
  "regex",
+ "russe",
  "serde",
  "serde_html_form",
  "serde_json",
@@ -256,12 +238,12 @@ dependencies = [
 
 [[package]]
 name = "actix-web-lab-derive"
-version = "0.26.0"
+version = "0.29.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2ac4d6e04e97fe707286509b4f338e99c5fb7249c770e1da074af5e27faa96b3"
+checksum = "f3e4b5439e27c9786f334a4ca6b9dc4d148fae4d6683e1e4f6d36575a61d616d"
 dependencies = [
  "quote",
- "syn",
+ "syn 3.0.9",
 ]
 
 [[package]]
@@ -302,9 +284,9 @@ dependencies = [
 
 [[package]]
 name = "aho-corasick"
-version = "1.1.4"
+version = "1.1.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ddd31a130427c27518df266943a5308ed92d4b226cc639f5a8f1002816174301"
+checksum = "c982642fa9e8606056828ee9a8505737230110bb1099153c79efe865c59d12ba"
 dependencies = [
  "memchr",
 ]
@@ -329,15 +311,15 @@ dependencies = [
 
 [[package]]
 name = "alloc-no-stdlib"
-version = "2.0.4"
+version = "3.0.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "cc7bb162ec39d46ab1ca8c77bf72e890535becd1751bb45f64c597edb4c8c6b3"
+checksum = "c2fb6cfd47bf496ff64095c20eaba0c201404ee38714d4142fcfa1dc334fcc7a"
 
 [[package]]
 name = "alloc-stdlib"
-version = "0.2.2"
+version = "0.3.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "94fb8275041c72129eb51b7d0322c29b8387a0386127718b096429201a5d6ece"
+checksum = "0b5c1865780388bfa186411ab5f247819487fc4864c6e9c3106611fa347586e1"
 dependencies = [
  "alloc-no-stdlib",
 ]
@@ -348,45 +330,11 @@ version = "0.2.21"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "683d7910e743518b0e34f1186f92494becacb047c7b6bf616c96772180fef923"
 
-[[package]]
-name = "android-activity"
-version = "0.6.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0f2a1bb052857d5dd49572219344a7332b31b76405648eabac5bc68978251bcd"
-dependencies = [
- "android-properties",
- "bitflags 2.11.0",
- "cc",
- "jni 0.22.4",
- "libc",
- "log",
- "ndk",
- "ndk-context",
- "ndk-sys",
- "num_enum",
- "thiserror 2.0.18",
-]
-
-[[package]]
-name = "android-build"
-version = "0.1.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8cac4c64175d504608cf239756339c07f6384a476f97f20a7043f92920b0b8fd"
-dependencies = [
- "windows-sys 0.52.0",
-]
-
-[[package]]
-name = "android-properties"
-version = "0.2.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "fc7eb209b1518d6bb87b283c20095f5228ecda460da70b44f0802523dea6da04"
-
 [[package]]
 name = "android_system_properties"
-version = "0.1.5"
+version = "0.1.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "819e7219dbd41043ac279b19830f2efc897156490d7fd6ea916720117ee66311"
+checksum = "ae221649c9976a6f6c56ae1facf410f3ddb33cc661c4b7b61020a912d4237fbc"
 dependencies = [
  "libc",
 ]
@@ -443,21 +391,30 @@ dependencies = [
 
 [[package]]
 name = "anyhow"
-version = "1.0.102"
+version = "1.0.104"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "330a5ed07fa54e4702c9d6c4174f74427fc0ef6e214bbd677ae50a5099946470"
+
+[[package]]
+name = "approx"
+version = "0.5.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7f202df86484c868dbad7eaa557ef785d5c66295e41b460ef922eca0723b842c"
+checksum = "cab112f0a86d568ea0e627cc1d6be74a1e9cd55214684db5561995f6dad897c6"
+dependencies = [
+ "num-traits",
+]
 
 [[package]]
 name = "arbitrary"
-version = "1.4.2"
+version = "1.5.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c3d036a3c4ab069c7b410a2ce876bd74808d2d0888a82667669f8e783a898bf1"
+checksum = "3bc62ac97cc33321f50863d514c3bc38a453947a8f9e781137e47c7401020aed"
 
 [[package]]
 name = "arc-swap"
-version = "1.9.0"
+version = "1.9.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a07d1f37ff60921c83bdfc7407723bdefe89b44b98a9b772f225c8f9d67141a6"
+checksum = "380962b3d93b99642e50b9febf72c071e97e8b32d86eae9cf4ef582bb960ae6e"
 dependencies = [
  "rustversion",
 ]
@@ -470,7 +427,7 @@ checksum = "0ae92a5119aa49cdbcf6b9f893fe4e1d98b04ccbf82ee0584ad948a44a734dea"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn",
+ "syn 2.0.119",
 ]
 
 [[package]]
@@ -487,15 +444,9 @@ checksum = "76a2e8124351fda1ef8aaaa3bbd7ebbcb486bbcd4225aca0aa0d84bb2db8fecb"
 
 [[package]]
 name = "arrayvec"
-version = "0.7.6"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7c02d123df017efcdfbd739ef81735b36c5ba83ec3c59c80a9d7ecc718f92e50"
-
-[[package]]
-name = "as-raw-xcb-connection"
-version = "1.0.1"
+version = "0.7.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "175571dd1d178ced59193a6fc02dde1b972eb0bc56c892cde9beeceac5bf0f6b"
+checksum = "d3fb67a6e08acf24fdeccbac2cb6ac4305825bd1f117462e0e6f2f193345ad56"
 
 [[package]]
 name = "as-slice"
@@ -506,22 +457,13 @@ dependencies = [
  "stable_deref_trait",
 ]
 
-[[package]]
-name = "ash"
-version = "0.38.0+1.3.281"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0bb44936d800fea8f016d7f2311c6a4f97aebd5dc86f09906139ec848cf3a46f"
-dependencies = [
- "libloading",
-]
-
 [[package]]
 name = "askama"
-version = "0.14.0"
+version = "0.16.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f75363874b771be265f4ffe307ca705ef6f3baa19011c149da8674a87f1b75c4"
+checksum = "6024d73179f43f15ccd2b881bfea6fee7f3a46ec53f33b52210dea749ebebaa4"
 dependencies = [
- "askama_derive",
+ "askama_macros",
  "itoa",
  "percent-encoding",
  "serde",
@@ -530,120 +472,64 @@ dependencies = [
 
 [[package]]
 name = "askama_derive"
-version = "0.14.0"
+version = "0.16.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "129397200fe83088e8a68407a8e2b1f826cf0086b21ccdb866a722c8bcd3a94f"
+checksum = "071ee5ebf2138e3ad180e0aacf6940c2cab5e6d8333741d9925c7bee2b153f39"
 dependencies = [
  "askama_parser",
  "basic-toml",
+ "glob",
  "memchr",
  "proc-macro2",
  "quote",
- "rustc-hash 2.1.2",
- "serde",
- "serde_derive",
- "syn",
-]
-
-[[package]]
-name = "askama_parser"
-version = "0.14.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d6ab5630b3d5eaf232620167977f95eb51f3432fc76852328774afbd242d4358"
-dependencies = [
- "memchr",
+ "rustc-hash",
  "serde",
  "serde_derive",
- "winnow 0.7.15",
-]
-
-[[package]]
-name = "async-broadcast"
-version = "0.7.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "435a87a52755b8f27fcf321ac4f04b2802e337c8c4872923137471ec39c37532"
-dependencies = [
- "event-listener",
- "event-listener-strategy",
- "futures-core",
- "pin-project-lite",
-]
-
-[[package]]
-name = "async-channel"
-version = "2.5.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "924ed96dd52d1b75e9c1a3e6275715fd320f5f9439fb5a4a11fa51f4221158d2"
-dependencies = [
- "concurrent-queue",
- "event-listener-strategy",
- "futures-core",
- "pin-project-lite",
-]
-
-[[package]]
-name = "async-executor"
-version = "1.14.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c96bf972d85afc50bf5ab8fe2d54d1586b4e0b46c97c50a0c9e71e2f7bcd812a"
-dependencies = [
- "async-task",
- "concurrent-queue",
- "fastrand",
- "futures-lite",
- "pin-project-lite",
- "slab",
+ "syn 3.0.9",
 ]
 
 [[package]]
-name = "async-io"
-version = "2.6.0"
+name = "askama_macros"
+version = "0.16.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "456b8a8feb6f42d237746d4b3e9a178494627745c3c56c6ea55d92ba50d026fc"
+checksum = "643e1c7cbb6aec1d920332fe51a7c0d8219e273dcb8602db03f5263e4d16487b"
 dependencies = [
- "autocfg",
- "cfg-if",
- "concurrent-queue",
- "futures-io",
- "futures-lite",
- "parking",
- "polling",
- "rustix 1.1.4",
- "slab",
- "windows-sys 0.61.2",
+ "askama_derive",
 ]
 
 [[package]]
-name = "async-lock"
-version = "3.4.2"
+name = "askama_parser"
+version = "0.16.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "290f7f2596bd5b78a9fec8088ccd89180d7f9f55b94b0576823bbbdc72ee8311"
+checksum = "2c5ae75772275d268b03ab8bdccdd12117b6169ee23256942b34e46c9f476583"
 dependencies = [
- "event-listener",
- "event-listener-strategy",
- "pin-project-lite",
+ "rustc-hash",
+ "serde",
+ "serde_derive",
+ "unicode-ident",
+ "winnow",
 ]
 
 [[package]]
 name = "async-openai"
-version = "0.40.3"
+version = "0.42.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "278af84d3d19995d440ea7b401a4b121c780c8d37f5eb6e883d987c17b523aa4"
+checksum = "174e979f1e2ffe94b88d584740c11fd071f6d829c29120b1ba88bcaeb3374063"
 dependencies = [
  "async-openai-macros",
- "base64",
- "bytes 1.11.1",
+ "base64 0.22.1",
+ "bytes",
  "derive_builder",
  "eventsource-stream",
- "futures 0.3.32",
+ "futures",
  "getrandom 0.3.4",
- "rand 0.9.4",
- "reqwest 0.13.4",
+ "rand 0.9.5",
+ "reqwest",
  "secrecy",
  "serde",
  "serde_json",
  "serde_urlencoded",
- "thiserror 2.0.18",
+ "thiserror",
  "tokio",
  "tokio-stream",
  "tokio-util",
@@ -660,93 +546,18 @@ checksum = "492a944774207eed3acf425214eadbd6ce84a2b89331164ff1c11bae92b26302"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn",
-]
-
-[[package]]
-name = "async-process"
-version = "2.5.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "fc50921ec0055cdd8a16de48773bfeec5c972598674347252c0399676be7da75"
-dependencies = [
- "async-channel",
- "async-io",
- "async-lock",
- "async-signal",
- "async-task",
- "blocking",
- "cfg-if",
- "event-listener",
- "futures-lite",
- "rustix 1.1.4",
-]
-
-[[package]]
-name = "async-recursion"
-version = "1.1.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3b43422f69d8ff38f95f1b2bb76517c91589a924d1559a0e935d7c8ce0274c11"
-dependencies = [
- "proc-macro2",
- "quote",
- "syn",
-]
-
-[[package]]
-name = "async-signal"
-version = "0.2.14"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "52b5aaafa020cf5053a01f2a60e8ff5dccf550f0f77ec54a4e47285ac2bab485"
-dependencies = [
- "async-io",
- "async-lock",
- "atomic-waker",
- "cfg-if",
- "futures-core",
- "futures-io",
- "rustix 1.1.4",
- "signal-hook-registry",
- "slab",
- "windows-sys 0.61.2",
-]
-
-[[package]]
-name = "async-stream"
-version = "0.3.6"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0b5a71a6f37880a80d1d7f19efd781e4b5de42c88f0722cc13bcb6cc2cfe8476"
-dependencies = [
- "async-stream-impl",
- "futures-core",
- "pin-project-lite",
-]
-
-[[package]]
-name = "async-stream-impl"
-version = "0.3.6"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c7c24de15d275a1ecfd47a380fb4d5ec9bfe0933f309ed5e705b775596a3574d"
-dependencies = [
- "proc-macro2",
- "quote",
- "syn",
+ "syn 2.0.119",
 ]
 
-[[package]]
-name = "async-task"
-version = "4.7.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8b75356056920673b02621b35afd0f7dda9306d03c79a30f5c56c44cf256e3de"
-
 [[package]]
 name = "async-trait"
-version = "0.1.89"
+version = "0.1.92"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9035ad2d096bed7955a320ee7e2230574d28fd3c3a0f186cbea1ff3c7eed5dbb"
+checksum = "82f6aeea286b8eb4dd3431a1be1b59d290ace00f5bfd8e2a159bc2a05e2c1667"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn",
+ "syn 3.0.9",
 ]
 
 [[package]]
@@ -757,9 +568,9 @@ checksum = "1505bd5d3d116872e7271a6d4e16d81d0c8570876c8de68093a09ac269d8aac0"
 
 [[package]]
 name = "autocfg"
-version = "1.5.0"
+version = "1.5.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c08606f8c3cbf4ce6ec8e28fb0014a2c086708fe954eaa885384a6165172e7e8"
+checksum = "f2032f911046de80f0a198e0901378627c33f59ea0ac00e363d481118bd70a53"
 
 [[package]]
 name = "av-scenechange"
@@ -776,7 +587,7 @@ dependencies = [
  "num-traits",
  "pastey",
  "rayon",
- "thiserror 2.0.18",
+ "thiserror",
  "v_frame",
  "y4m",
 ]
@@ -797,18 +608,18 @@ dependencies = [
 
 [[package]]
 name = "avif-serialize"
-version = "0.8.8"
+version = "0.8.9"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "375082f007bd67184fb9c0374614b29f9aaa604ec301635f72338bb65386a53d"
+checksum = "e7178fe5f7d460b13895ebb9dcb28a3a6216d2df2574a0806cb51b555d297f38"
 dependencies = [
  "arrayvec",
 ]
 
 [[package]]
 name = "aws-lc-rs"
-version = "1.18.1"
+version = "1.18.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b281d307588d634de920874890732659e2e7672f72b5e10e81badc1a8a83621e"
+checksum = "faac5829c2b74c28f830747e7818ccfb684261b5f48a1118b1e2a13d36dfab13"
 dependencies = [
  "aws-lc-sys",
  "zeroize",
@@ -816,9 +627,9 @@ dependencies = [
 
 [[package]]
 name = "aws-lc-sys"
-version = "0.45.0"
+version = "0.46.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9bff6c3b54fad79a2e60b8102caf565819711497c1f5f092f49508e2f5c31b27"
+checksum = "f1622d8446a2d4b2ce0c7eefc73dd43a99779028d5ee5c2dd8073a658ba8a2bc"
 dependencies = [
  "cc",
  "cmake",
@@ -833,6 +644,12 @@ version = "0.22.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "72b3254f16251a8381aa12e40e3c4d2f0199f8c6508fbecb9d91f575e0fbb8c6"
 
+[[package]]
+name = "base64"
+version = "0.23.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ac07cdecf99051d9a5238b80f35af32cdeba5b336e55d957b318b50137e18da5"
+
 [[package]]
 name = "basic-toml"
 version = "0.1.10"
@@ -848,18 +665,18 @@ version = "0.72.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "993776b509cfb49c750f11b8f07a46fa23e0a1386ffc01fb1e7d343efc387895"
 dependencies = [
- "bitflags 2.11.0",
+ "bitflags 2.13.2",
  "cexpr",
  "clang-sys",
  "itertools 0.13.0",
  "log",
- "prettyplease",
+ "prettyplease 0.2.37",
  "proc-macro2",
  "quote",
  "regex",
- "rustc-hash 2.1.2",
+ "rustc-hash",
  "shlex",
- "syn",
+ "syn 2.0.119",
 ]
 
 [[package]]
@@ -891,24 +708,31 @@ checksum = "bef38d45163c2f1dde094a7dfd33ccf595c92905c8f8f4fdc18d06fb1037718a"
 
 [[package]]
 name = "bitflags"
-version = "2.11.0"
+version = "2.13.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "843867be96c8daad0d758b57df9392b6d8d271134fce549de6ce169ff98a92af"
+checksum = "3ded4057c258ba199e2d26386d3af3780957ecaee6c4ef4041c6b4b8b97c0b06"
 
 [[package]]
 name = "bitstream-io"
-version = "4.9.0"
+version = "4.10.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "60d4bd9d1db2c6bdf285e223a7fa369d5ce98ec767dec949c6ca62863ce61757"
+checksum = "7eff00be299a18769011411c9def0d827e8f2d7bf0c3dbf53633147a8867fd1f"
 dependencies = [
- "core2",
+ "no_std_io2",
 ]
 
 [[package]]
-name = "block"
-version = "0.1.6"
+name = "blake3"
+version = "1.8.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0d8c1fef690941d3e7788d328517591fecc684c084084702d6ff1641e993699a"
+checksum = "6d9e454fc11f76977dc803893aff6304ed33d6a26efae8696573bea74baa27ae"
+dependencies = [
+ "arrayvec",
+ "cc",
+ "cfg-if",
+ "constant_time_eq",
+ "cpufeatures 0.3.1",
+]
 
 [[package]]
 name = "block-buffer"
@@ -921,42 +745,34 @@ dependencies = [
 
 [[package]]
 name = "block-buffer"
-version = "0.12.0"
+version = "0.12.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "cdd35008169921d80bc60d3d0ab416eecb028c4cd653352907921d95084790be"
+checksum = "d2f6c7dbe95a6ed67ad9f18e57daf93a2f034c524b99fd2b76d18fdfeb6660aa"
 dependencies = [
  "hybrid-array",
 ]
 
 [[package]]
-name = "block2"
-version = "0.5.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2c132eebf10f5cad5289222520a4a058514204aed6d791f1cf4fe8088b82d15f"
-dependencies = [
- "objc2 0.5.2",
-]
-
-[[package]]
-name = "block2"
-version = "0.6.2"
+name = "bon"
+version = "3.10.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "cdeb9d870516001442e364c5220d3574d2da8dc765554b4a617230d33fa58ef5"
+checksum = "a05442cd7005532d0cfedec676b5b9333d2b4c7c3afabe2b1a9ce2742fbe2ee6"
 dependencies = [
- "objc2 0.6.4",
+ "bon-macros",
 ]
 
 [[package]]
-name = "blocking"
-version = "1.6.2"
+name = "bon-macros"
+version = "3.10.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e83f8d02be6967315521be875afa792a316e28d57b5a2d401897e2a7921b7f21"
+checksum = "93833e97ac7c46ba97045afcf247238c27057538fcf2f06344662ddd5f567be6"
 dependencies = [
- "async-channel",
- "async-task",
- "futures-io",
- "futures-lite",
- "piper",
+ "darling 0.24.1",
+ "ident_case",
+ "prettyplease 0.3.0",
+ "proc-macro2",
+ "quote",
+ "syn 3.0.9",
 ]
 
 [[package]]
@@ -967,9 +783,9 @@ checksum = "dc0b364ead1874514c8c2855ab558056ebfeb775653e7ae45ff72f28f8f3166c"
 
 [[package]]
 name = "brotli"
-version = "8.0.2"
+version = "9.0.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "4bd8b9603c7aa97359dbd97ecf258968c95f3adddd6db2f7e7a5bef101c84560"
+checksum = "f8b851b75c23ca7873623d612fe49bd1989aeb03d08fb9432187eb253d3d4c6b"
 dependencies = [
  "alloc-no-stdlib",
  "alloc-stdlib",
@@ -978,25 +794,36 @@ dependencies = [
 
 [[package]]
 name = "brotli-decompressor"
-version = "5.0.0"
+version = "6.0.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "874bb8112abecc98cbd6d81ea4fa7e94fb9449648c93cc89aa40c81c24d7de03"
+checksum = "941cd9bd4ddab83cb46fa5a2d428f1c857b24ac78cb876cf7beb710840934bd7"
 dependencies = [
  "alloc-no-stdlib",
  "alloc-stdlib",
 ]
 
+[[package]]
+name = "bstr"
+version = "1.13.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6bb31b46c14244e20ee9984b11bf5c992b91fb6939fea616e3512c8baecdbe5f"
+dependencies = [
+ "memchr",
+ "regex-automata",
+ "serde_core",
+]
+
 [[package]]
 name = "built"
-version = "0.8.0"
+version = "0.8.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f4ad8f11f288f48ca24471bbd51ac257aaeaaa07adae295591266b792902ae64"
+checksum = "5c0e531d93d39c34eef561e929e8a7f86d77a5af08aac4f6d6e39976c51858e9"
 
 [[package]]
 name = "bumpalo"
-version = "3.20.2"
+version = "3.20.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5d20789868f4b01b2f2caec9f5c4e0213b41e3e5702a50157d699ae31ced2fcb"
+checksum = "72f5acc6cb2ba439de613abc23857ec3d78374d8ed5ac84e9d11336e87da8649"
 
 [[package]]
 name = "bytecount"
@@ -1006,30 +833,24 @@ checksum = "175812e0be2bccb6abe50bb8d566126198344f707e304f45c648fd8f2cc0365e"
 
 [[package]]
 name = "bytemuck"
-version = "1.25.0"
+version = "1.25.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c8efb64bd706a16a1bdde310ae86b351e4d21550d98d056f22f8a7f7a2183fec"
+checksum = "95832e849adfb21180ccb6826a99da14e5d266ae5c2e668e1602cf234f153797"
 dependencies = [
  "bytemuck_derive",
 ]
 
 [[package]]
 name = "bytemuck_derive"
-version = "1.10.2"
+version = "1.12.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f9abbd1bc6865053c427f7198e6af43bfdedc55ab791faed4fbd361d789575ff"
+checksum = "6a1f896587b6f2c069c73d2f0913e2d590c3990285cd2f0b6aa02b786b4c679c"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn",
+ "syn 3.0.9",
 ]
 
-[[package]]
-name = "byteorder"
-version = "1.5.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1fd0f2584146f6f2ef48085050886acf353beff7305ebd1ae69500e27c67f64b"
-
 [[package]]
 name = "byteorder-lite"
 version = "0.1.0"
@@ -1038,89 +859,28 @@ checksum = "8f1fe948ff07f4bd06c30984e69f5b4899c516a3ef74f34df92a2df2ab535495"
 
 [[package]]
 name = "bytes"
-version = "0.4.12"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "206fdffcfa2df7cbe15601ef46c813fce0965eb3286db6b56c583b814b51c81c"
-dependencies = [
- "byteorder",
- "iovec",
-]
-
-[[package]]
-name = "bytes"
-version = "1.11.1"
+version = "1.12.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1e748733b7cbc798e1434b6ac524f0c1ff2ab456fe201501e6497c8417a4fc33"
+checksum = "fc652a48c352aef3ea3aed32080501cf3ef6ed5da78602a020c991775b0aff04"
 
 [[package]]
 name = "bytestring"
-version = "1.5.0"
+version = "1.5.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "113b4343b5f6617e7ad401ced8de3cc8b012e73a594347c307b90db3e9271289"
+checksum = "86566c496f2f47d9b8147a4c8b02ffdb69c919fe0c2b2e7195d22cbba0e635c9"
 dependencies = [
- "bytes 1.11.1",
+ "bytes",
 ]
 
 [[package]]
 name = "cadence"
-version = "1.7.0"
+version = "1.8.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d7aff0c323415907f37007d645d7499c378df47efb3e33ffc1f397fa4e549b2e"
+checksum = "5ca08f6db9f0c963249cf23a27820f9d973d2acaad4d6bfaeb858b58ebd58a6a"
 dependencies = [
  "crossbeam-channel",
 ]
 
-[[package]]
-name = "calloop"
-version = "0.13.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b99da2f8558ca23c71f4fd15dc57c906239752dd27ff3c00a1d56b685b7cbfec"
-dependencies = [
- "bitflags 2.11.0",
- "log",
- "polling",
- "rustix 0.38.44",
- "slab",
- "thiserror 1.0.69",
-]
-
-[[package]]
-name = "calloop"
-version = "0.14.4"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "4dbf9978365bac10f54d1d4b04f7ce4427e51f71d61f2fe15e3fed5166474df7"
-dependencies = [
- "bitflags 2.11.0",
- "polling",
- "rustix 1.1.4",
- "slab",
- "tracing",
-]
-
-[[package]]
-name = "calloop-wayland-source"
-version = "0.3.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "95a66a987056935f7efce4ab5668920b5d0dac4a7c99991a67395f13702ddd20"
-dependencies = [
- "calloop 0.13.0",
- "rustix 0.38.44",
- "wayland-backend",
- "wayland-client",
-]
-
-[[package]]
-name = "calloop-wayland-source"
-version = "0.4.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "138efcf0940a02ebf0cc8d1eff41a1682a46b431630f4c52450d6265876021fa"
-dependencies = [
- "calloop 0.14.4",
- "rustix 1.1.4",
- "wayland-backend",
- "wayland-client",
-]
-
 [[package]]
 name = "cc"
 version = "1.2.58"
@@ -1133,12 +893,6 @@ dependencies = [
  "shlex",
 ]
 
-[[package]]
-name = "cesu8"
-version = "1.1.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6d43a04d8753f35258c91f8ec639f792891f748a1edbd759cf1dcea3382ad83c"
-
 [[package]]
 name = "cexpr"
 version = "0.6.0"
@@ -1150,43 +904,56 @@ dependencies = [
 
 [[package]]
 name = "cfg-if"
-version = "1.0.4"
+version = "1.0.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9330f8b2ff13f34540b44e946ef35111825727b38d33286ef986142615121801"
+checksum = "4e7648175b45a9a48536d676f68d918270699102aa8dab5496df06904c914600"
 
 [[package]]
 name = "cfg_aliases"
-version = "0.2.1"
+version = "0.2.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "613afe47fcd5fac7ccf1db93babcb082c5994d996f20b8b159f2ad1658eb5724"
+checksum = "f079e83a288787bcd14a6aea84cee5c87a67c5a3e660c30f557a3d24761b3527"
 
 [[package]]
 name = "chacha20"
-version = "0.10.0"
+version = "0.10.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6f8d983286843e49675a4b7a2d174efe136dc93a18d69130dd18198a6c167601"
+checksum = "65c35e4b699c7e15ccbe7ee35c005e4fc0a278d22238a2857e6ce2dadeda1b06"
 dependencies = [
  "cfg-if",
- "cpufeatures 0.3.0",
+ "cpufeatures 0.3.1",
  "rand_core 0.10.1",
 ]
 
 [[package]]
-name = "clang-sys"
-version = "1.8.1"
+name = "chrono"
+version = "0.4.45"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0b023947811758c97c59bf9d1c188fd619ad4718dcaa767947df1cadb14f39f4"
+checksum = "1aa79e62e7697b8e29b513a68abacf485adcd1fe8284a4316c5ae868e6633327"
 dependencies = [
- "glob",
- "libc",
+ "iana-time-zone",
+ "js-sys",
+ "num-traits",
+ "wasm-bindgen",
+ "windows-link",
+]
+
+[[package]]
+name = "clang-sys"
+version = "1.9.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "157a8ba7b480713b56f4c09fd13fc3e0a22a5dfab8097ba61cbc5feef950788a"
+dependencies = [
+ "glob",
+ "libc",
  "libloading",
 ]
 
 [[package]]
 name = "clap"
-version = "4.6.0"
+version = "4.6.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b193af5b67834b676abd72466a96c1024e6a6ad978a1f484bd90b85c94041351"
+checksum = "aa8876b300ab35ba921adea3dfd70157a46249b33f95c9084ae5709785478946"
 dependencies = [
  "clap_builder",
  "clap_derive",
@@ -1194,9 +961,9 @@ dependencies = [
 
 [[package]]
 name = "clap_builder"
-version = "4.6.0"
+version = "4.6.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "714a53001bf66416adb0e2ef5ac857140e7dc3a0c48fb28b2f10762fc4b5069f"
+checksum = "ec0797fb7aeb1406c84efac526901f7ec3ead2124f946b494e72879d4b54704d"
 dependencies = [
  "anstream",
  "anstyle",
@@ -1206,60 +973,21 @@ dependencies = [
 
 [[package]]
 name = "clap_derive"
-version = "4.6.0"
+version = "4.6.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1110bd8a634a1ab8cb04345d8d878267d57c3cf1b38d91b71af6686408bbca6a"
+checksum = "f9c751b79415d4e559e3d1fcf128e09e720eb673a06d26cf6f392d37d75b66e0"
 dependencies = [
  "heck",
  "proc-macro2",
  "quote",
- "syn",
+ "syn 3.0.9",
 ]
 
 [[package]]
 name = "clap_lex"
-version = "1.1.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c8d4a3bb8b1e0c1050499d1815f5ab16d04f0959b233085fb31653fbfc9d98f9"
-
-[[package]]
-name = "clipboard-win"
-version = "5.4.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "bde03770d3df201d4fb868f2c9c59e66a3e4e2bd06692a0fe701e7103c7e84d4"
-dependencies = [
- "error-code",
-]
-
-[[package]]
-name = "clipboard_macos"
-version = "0.1.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9b7f4aaa047ba3c3630b080bb9860894732ff23e2aee290a418909aa6d5df38f"
-dependencies = [
- "objc2 0.5.2",
- "objc2-app-kit 0.2.2",
- "objc2-foundation 0.2.2",
-]
-
-[[package]]
-name = "clipboard_wayland"
-version = "0.2.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "003f886bc4e2987729d10c1db3424e7f80809f3fc22dbc16c685738887cb37b8"
-dependencies = [
- "smithay-clipboard",
-]
-
-[[package]]
-name = "clipboard_x11"
-version = "0.4.3"
+version = "1.1.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "bd63e33452ffdafd39924c4f05a5dd1e94db646c779c6bd59148a3d95fff5ad4"
-dependencies = [
- "thiserror 2.0.18",
- "x11rb",
-]
+checksum = "1c133bc6a41be0d194c306b5506d15e6feeea7b1d6604bd3f8310dfb2ca96486"
 
 [[package]]
 name = "cmake"
@@ -1270,17 +998,6 @@ dependencies = [
  "cc",
 ]
 
-[[package]]
-name = "codespan-reporting"
-version = "0.12.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "fe6d2e5af09e8c8ad56c969f2157a3d4238cebc7c55f0a517728c38f7b200f81"
-dependencies = [
- "serde",
- "termcolor",
- "unicode-width",
-]
-
 [[package]]
 name = "color_quant"
 version = "1.1.0"
@@ -1293,21 +1010,30 @@ version = "1.0.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "1d07550c9036bf2ae0c684c4297d503f838287c83c53686d05370d0e139ae570"
 
+[[package]]
+name = "colored"
+version = "3.1.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "faf9468729b8cbcea668e36183cb69d317348c2e08e994829fb56ebfdfbaac34"
+dependencies = [
+ "windows-sys 0.61.2",
+]
+
 [[package]]
 name = "combine"
-version = "4.6.7"
+version = "4.6.9"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ba5a308b75df32fe02788e748662718f03fde005016435c444eea572398219fd"
+checksum = "4356eba6976435498d05a311ee9fc486da3bc81c200d3f0d7dd9703b41913f92"
 dependencies = [
- "bytes 1.11.1",
+ "bytes",
  "memchr",
 ]
 
 [[package]]
 name = "command_handler"
-version = "0.1.0"
+version = "0.2.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "49385f359e523e20d614c5f92195db2ba8e042e5ce8dadafb6174c2abd95adbb"
+checksum = "fc373209739ae365edb7c8c78f832f02919b32fb02cb603dd37cf2d4457d6ccc"
 dependencies = [
  "anyhow",
  "async-trait",
@@ -1317,32 +1043,31 @@ dependencies = [
 ]
 
 [[package]]
-name = "concurrent-queue"
-version = "2.5.0"
+name = "const-oid"
+version = "0.10.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "4ca0197aee26d1ae37445ee532fefce43251d24cc7c166799f4d46817f1d3973"
-dependencies = [
- "crossbeam-utils",
-]
+checksum = "a6ef517f0926dd24a1582492c791b6a4818a4d94e789a334894aa15b0d12f55c"
 
 [[package]]
-name = "console"
-version = "0.15.11"
+name = "const-str"
+version = "1.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "18f12cc9948ed9604230cdddc7c86e270f9401ccbe3c2e98a4378c5e7632212f"
+
+[[package]]
+name = "const_panic"
+version = "0.2.17"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "054ccb5b10f9f2cbf51eb355ca1d05c2d279ce1804688d0db74b4733a5aeafd8"
+checksum = "9603f79528ece8163c496f8932121cb36cfe46259e9c907bb3d8205139d7caa3"
 dependencies = [
- "encode_unicode",
- "libc",
- "once_cell",
- "unicode-width",
- "windows-sys 0.59.0",
+ "typewit",
 ]
 
 [[package]]
-name = "const-oid"
-version = "0.10.2"
+name = "constant_time_eq"
+version = "0.4.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a6ef517f0926dd24a1582492c791b6a4818a4d94e789a334894aa15b0d12f55c"
+checksum = "3d52eff69cd5e647efe296129160853a42795992097e8af39800e1060caeea9b"
 
 [[package]]
 name = "convert_case"
@@ -1391,80 +1116,12 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "773648b94d0e5d620f64f280777445740e61fe701025087ec8b57f45c791888b"
 
 [[package]]
-name = "core-graphics"
-version = "0.23.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c07782be35f9e1140080c6b96f0d44b739e2278479f64e02fdab4e32dfd8b081"
-dependencies = [
- "bitflags 1.3.2",
- "core-foundation 0.9.4",
- "core-graphics-types 0.1.3",
- "foreign-types 0.5.0",
- "libc",
-]
-
-[[package]]
-name = "core-graphics-types"
-version = "0.1.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "45390e6114f68f718cc7a830514a96f903cccd70d02a8f6d9f643ac4ba45afaf"
-dependencies = [
- "bitflags 1.3.2",
- "core-foundation 0.9.4",
- "libc",
-]
-
-[[package]]
-name = "core-graphics-types"
-version = "0.2.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3d44a101f213f6c4cdc1853d4b78aef6db6bdfa3468798cc1d9912f4735013eb"
-dependencies = [
- "bitflags 2.11.0",
- "core-foundation 0.10.1",
- "libc",
-]
-
-[[package]]
-name = "core2"
-version = "0.4.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b49ba7ef1ad6107f8824dbe97de947cbaac53c44e7f9756a1fba0d37c1eec505"
-dependencies = [
- "memchr",
-]
-
-[[package]]
-name = "core_maths"
-version = "0.1.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "77745e017f5edba1a9c1d854f6f3a52dac8a12dd5af5d2f54aecf61e43d80d30"
-dependencies = [
- "libm",
-]
-
-[[package]]
-name = "cosmic-text"
-version = "0.15.0"
+name = "countio"
+version = "0.3.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "173852283a9a57a3cbe365d86e74dc428a09c50421477d5ad6fe9d9509e37737"
+checksum = "b9702aee5d1d744c01d82f6915644f950f898e014903385464c773b96fefdecb"
 dependencies = [
- "bitflags 2.11.0",
- "fontdb",
- "harfrust",
- "linebender_resource_handle",
- "log",
- "rangemap",
- "rustc-hash 1.1.0",
- "self_cell",
- "skrifa 0.37.0",
- "smol_str",
- "swash",
- "sys-locale",
- "unicode-bidi",
- "unicode-linebreak",
- "unicode-script",
- "unicode-segmentation",
+ "futures-io",
 ]
 
 [[package]]
@@ -1478,36 +1135,36 @@ dependencies = [
 
 [[package]]
 name = "cpufeatures"
-version = "0.3.0"
+version = "0.3.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8b2a41393f66f16b0823bb79094d54ac5fbd34ab292ddafb9a0456ac9f87d201"
+checksum = "5ca28b0ae3115b884660db4118d803791fd6756b6e88f39c0f3f7859060d7566"
 dependencies = [
  "libc",
 ]
 
 [[package]]
 name = "crc32fast"
-version = "1.5.0"
+version = "1.5.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9481c1c90cbf2ac953f07c8d4a58aa3945c425b7185c9154d67a65e4230da511"
+checksum = "0ba106e844045ddc02c20768ac4b0b7cb45f2cc2ecbd833ac9f5dacb183c2055"
 dependencies = [
  "cfg-if",
 ]
 
 [[package]]
 name = "crossbeam-channel"
-version = "0.5.15"
+version = "0.5.17"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "82b8f8f868b36967f9606790d1903570de9ceaf870a7bf9fbbd3016d636a2cb2"
+checksum = "98b0cc327b5bc766e7fda9c9260cc0fa81b43a8e240440422dff70788e3f9ef1"
 dependencies = [
  "crossbeam-utils",
 ]
 
 [[package]]
 name = "crossbeam-deque"
-version = "0.8.6"
+version = "0.8.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9dd111b7b7f7d55b72c0a6ae361660ee5853c9af73f70c3c2ef6858b950e2e51"
+checksum = "622f3fc73690be383c7214310406f28a90e6edeadc3cea882f9d71e495b9711a"
 dependencies = [
  "crossbeam-epoch",
  "crossbeam-utils",
@@ -1524,9 +1181,9 @@ dependencies = [
 
 [[package]]
 name = "crossbeam-utils"
-version = "0.8.21"
+version = "0.8.23"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d0a5c400df2834b80a4c3327b3aad3a4c4cd4de0629063962b03235697506a28"
+checksum = "a31eee39dddec8330830986fcd7625edb5a24ec90ea038215273bbc3adb08ac6"
 
 [[package]]
 name = "crunchy"
@@ -1534,19 +1191,6 @@ version = "0.2.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "460fbee9c2c2f33933d720630a6a0bac33ba7053db5344fac858d4b8952d77d5"
 
-[[package]]
-name = "cryoglyph"
-version = "0.1.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "08bc795bdbccdbd461736fb163930a009da6597b226d6f6fce33e7a8eb6ec519"
-dependencies = [
- "cosmic-text",
- "etagere",
- "lru",
- "rustc-hash 2.1.2",
- "wgpu",
-]
-
 [[package]]
 name = "crypto-common"
 version = "0.1.7"
@@ -1559,9 +1203,9 @@ dependencies = [
 
 [[package]]
 name = "crypto-common"
-version = "0.2.1"
+version = "0.2.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "77727bb15fa921304124b128af125e7e3b968275d1b108b379190264f4423710"
+checksum = "ce6e4c961d6cd6c9a86db418387425e8bdeaf05b3c8bc1411e6dca4c252f1453"
 dependencies = [
  "hybrid-array",
 ]
@@ -1588,25 +1232,33 @@ dependencies = [
 ]
 
 [[package]]
-name = "ctor-lite"
-version = "0.1.2"
+name = "ctor"
+version = "1.0.13"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e162d0c2e2068eb736b71e5597eff0b9944e6b973cd9f37b6a288ab9bf20e300"
+checksum = "914a755b7c2d4af2bdcff7ce1739e2db9a1b81a9b07123d8015786ae03c0980d"
+dependencies = [
+ "link-section",
+ "linktime-proc-macro",
+]
 
 [[package]]
-name = "cursor-icon"
-version = "1.2.0"
+name = "darling"
+version = "0.20.11"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f27ae1dd37df86211c42e150270f82743308803d90a6f6e6651cd730d5e1732f"
+checksum = "fc7f46116c46ff9ab3eb1597a45688b6715c6e628b5c133e288e709a29bcb4ee"
+dependencies = [
+ "darling_core 0.20.11",
+ "darling_macro 0.20.11",
+]
 
 [[package]]
 name = "darling"
-version = "0.20.11"
+version = "0.24.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "fc7f46116c46ff9ab3eb1597a45688b6715c6e628b5c133e288e709a29bcb4ee"
+checksum = "ed17f5901b6630b993ca003def43f2f8ef4014fc13b047b57aad617ff32bc2ec"
 dependencies = [
- "darling_core",
- "darling_macro",
+ "darling_core 0.24.1",
+ "darling_macro 0.24.1",
 ]
 
 [[package]]
@@ -1620,7 +1272,20 @@ dependencies = [
  "proc-macro2",
  "quote",
  "strsim",
- "syn",
+ "syn 2.0.119",
+]
+
+[[package]]
+name = "darling_core"
+version = "0.24.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6837e2cf7485aaae18f86181d2f0e9a7ed297a025e220aeabf63fdebd3a2ddff"
+dependencies = [
+ "ident_case",
+ "proc-macro2",
+ "quote",
+ "strsim",
+ "syn 3.0.9",
 ]
 
 [[package]]
@@ -1629,30 +1294,41 @@ version = "0.20.11"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "fc34b93ccb385b40dc71c6fceac4b2ad23662c7eeb248cf10d529b7e055b6ead"
 dependencies = [
- "darling_core",
+ "darling_core 0.20.11",
+ "quote",
+ "syn 2.0.119",
+]
+
+[[package]]
+name = "darling_macro"
+version = "0.24.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2ac7135c3ef02b2f7833bbeb1be5ba7f966dcde8a87c6b87f65a778d71a02785"
+dependencies = [
+ "darling_core 0.24.1",
  "quote",
- "syn",
+ "syn 3.0.9",
 ]
 
 [[package]]
 name = "dashmap"
-version = "6.1.0"
+version = "6.2.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5041cc499144891f3790297212f32a74fb938e5136a14943f338ef9e0ae276cf"
+checksum = "e6361d5c062261c78a176addb82d4c821ae42bed6089de0e12603cd25de2059c"
 dependencies = [
  "cfg-if",
  "crossbeam-utils",
  "hashbrown 0.14.5",
- "lock_api",
+ "lock_api 0.4.14",
  "once_cell",
- "parking_lot_core",
+ "parking_lot_core 0.9.12",
 ]
 
 [[package]]
 name = "data-encoding"
-version = "2.10.0"
+version = "2.11.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d7a1e2f27636f116493b8b860f5546edb47c8d8f8ea73e1d2a20be88e28d1fea"
+checksum = "4583a4551df46e2792f82ceeac45e850d2e2d5debba0b91f102385cda5b11f06"
 
 [[package]]
 name = "data-url"
@@ -1661,15 +1337,42 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "be1e0bca6c3637f992fc1cc7cbc52a78c1ef6db076dbf1059c4323d6a2048376"
 
 [[package]]
-name = "deranged"
-version = "0.5.8"
+name = "defmt"
+version = "1.2.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7cd812cc2bc1d69d4764bd80df88b4317eaef9e773c75226407d9bc0876b211c"
+checksum = "e4a0795ac740cb54d7b3775c9c17345f6aaecf884c01733b39d6017738e0e3e1"
 dependencies = [
- "powerfmt",
- "serde_core",
+ "bitflags 1.3.2",
+ "defmt-macros",
+]
+
+[[package]]
+name = "defmt-macros"
+version = "1.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "840bb056bd99c495c9188198a258ab787c79f306d864db3e6e8451442d6d8366"
+dependencies = [
+ "defmt-parser",
+ "proc-macro2",
+ "quote",
+ "syn 3.0.9",
+]
+
+[[package]]
+name = "defmt-parser"
+version = "1.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "71958f69fb2de856c5dd51ac9377e986ca1aa5919f931bf5395a807277f889a0"
+dependencies = [
+ "thiserror",
 ]
 
+[[package]]
+name = "deranged"
+version = "0.5.9"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9e9de72ce2ad1f90dc62fa25f0f430ef85eb4b0d8fa0be4f30373bc40a21d28e"
+
 [[package]]
 name = "derive_builder"
 version = "0.20.2"
@@ -1685,10 +1388,10 @@ version = "0.20.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "2d5bcf7b024d6835cfb3d473887cd966994907effbe9227e8c8219824d06c4e8"
 dependencies = [
- "darling",
+ "darling 0.20.11",
  "proc-macro2",
  "quote",
- "syn",
+ "syn 2.0.119",
 ]
 
 [[package]]
@@ -1698,7 +1401,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "ab63b0e2bf4d5928aff72e83a7dace85d7bba5fe12dcc3c5a572d78caffd3f3c"
 dependencies = [
  "derive_builder_core",
- "syn",
+ "syn 2.0.119",
 ]
 
 [[package]]
@@ -1720,7 +1423,7 @@ dependencies = [
  "proc-macro2",
  "quote",
  "rustc_version",
- "syn",
+ "syn 2.0.119",
  "unicode-xid",
 ]
 
@@ -1736,15 +1439,9 @@ dependencies = [
  "bytemuck_derive",
  "hashbrown 0.15.5",
  "regex-syntax",
- "strum",
+ "strum 0.27.2",
 ]
 
-[[package]]
-name = "diff"
-version = "0.1.13"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "56254986775e3233ffa9c4d7d3faaf6d36a2c09d30b20687e9f88bc8bafc16c8"
-
 [[package]]
 name = "digest"
 version = "0.10.7"
@@ -1757,13 +1454,13 @@ dependencies = [
 
 [[package]]
 name = "digest"
-version = "0.11.2"
+version = "0.11.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "4850db49bf08e663084f7fb5c87d202ef91a3907271aff24a94eb97ff039153c"
+checksum = "f1dd6dbb5841937940781866fa1281a1ff7bd3bf827091440879f9994983d5c2"
 dependencies = [
- "block-buffer 0.12.0",
+ "block-buffer 0.12.1",
  "const-oid",
- "crypto-common 0.2.1",
+ "crypto-common 0.2.2",
 ]
 
 [[package]]
@@ -1787,63 +1484,27 @@ dependencies = [
  "windows-sys 0.61.2",
 ]
 
-[[package]]
-name = "dispatch"
-version = "0.2.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "bd0c93bb4b0c6d9b77f4435b0ae98c24d17f1c45b2ff844c6151a07256ca923b"
-
 [[package]]
 name = "dispatch2"
 version = "0.3.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "1e0e367e4e7da84520dedcac1901e4da967309406d1e51017ae1abfb97adbd38"
 dependencies = [
- "bitflags 2.11.0",
- "objc2 0.6.4",
+ "bitflags 2.13.2",
+ "objc2",
 ]
 
 [[package]]
 name = "displaydoc"
-version = "0.2.5"
+version = "0.2.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "97369cbbc041bc366949bc74d34658d6cda5621039731c6310521892a3a20ae0"
+checksum = "c6232dd377dcc64799954cbd3a9bb882e9cdc1308ccd87b1c098f1fb2eaf82a8"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn",
-]
-
-[[package]]
-name = "dlib"
-version = "0.5.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ab8ecd87370524b461f8557c119c405552c396ed91fc0a8eec68679eab26f94a"
-dependencies = [
- "libloading",
+ "syn 3.0.9",
 ]
 
-[[package]]
-name = "document-features"
-version = "0.2.12"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d4b8a88685455ed29a21542a33abd9cb6510b6b129abadabdcef0f4c55bc8f61"
-dependencies = [
- "litrs",
-]
-
-[[package]]
-name = "downcast-rs"
-version = "1.2.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "75b325c5dbd37f80359721ad39aca5a29fb04c89279657cffdda8736d0c0b9d2"
-
-[[package]]
-name = "dpi"
-version = "0.1.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d8b14ccef22fc6f5a8f4d7d768562a182c04ce9a3b3157b91390b52ddfdf1a76"
-
 [[package]]
 name = "dunce"
 version = "1.0.5"
@@ -1852,9 +1513,9 @@ checksum = "92773504d58c093f6de2459af4af33faa518c13451eb8f2b5698ed3d36e7c813"
 
 [[package]]
 name = "either"
-version = "1.15.0"
+version = "1.19.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "48c757948c5ede0e46177b7add2e67155f70e33c07fea8284df6576da70b3719"
+checksum = "0e9c71c2167ca323c882b99918929403426e2373ea17242ff5653e0d5e1058be"
 
 [[package]]
 name = "email_address"
@@ -1865,12 +1526,6 @@ dependencies = [
  "serde",
 ]
 
-[[package]]
-name = "encode_unicode"
-version = "1.0.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "34aa73646ffb006b8f5147f3dc182bd4bcb190227ce861fc4a4844bf8e3cb2c0"
-
 [[package]]
 name = "encoding_rs"
 version = "0.8.35"
@@ -1881,12 +1536,6 @@ dependencies = [
  "serde",
 ]
 
-[[package]]
-name = "endi"
-version = "1.1.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "66b7e2430c6dff6a955451e2cfc438f09cea1965a9d6f87f7e3b90decc014099"
-
 [[package]]
 name = "enumflags2"
 version = "0.7.12"
@@ -1894,7 +1543,6 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "1027f7680c853e056ebcec683615fb6fbbc07dbaa13b4d5d9442b146ded4ecef"
 dependencies = [
  "enumflags2_derive",
- "serde",
 ]
 
 [[package]]
@@ -1905,14 +1553,14 @@ checksum = "67c78a4d8fdf9953a5c9d458f9efe940fd97a0cab0941c075a813ac594733827"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn",
+ "syn 2.0.119",
 ]
 
 [[package]]
 name = "env_filter"
-version = "1.0.1"
+version = "2.0.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "32e90c2accc4b07a8456ea0debdc2e7587bdd890680d71173a15d4ae604f6eef"
+checksum = "900d271a03799a1ee8d1ca9b19893b48ca674a9284fefcfb85f05e74ed314217"
 dependencies = [
  "log",
  "regex",
@@ -1920,9 +1568,9 @@ dependencies = [
 
 [[package]]
 name = "env_logger"
-version = "0.11.10"
+version = "0.11.11"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0621c04f2196ac3f488dd583365b9c09be011a4ab8b9f37248ffcc8f6198b56a"
+checksum = "de671bd27a75a797dc9ae289ba1e77276e75e2026408aab65185384e2d5cd3f6"
 dependencies = [
  "anstream",
  "anstyle",
@@ -1948,14 +1596,14 @@ checksum = "44f23cf4b44bfce11a86ace86f8a73ffdec849c9fd00a386a53d278bd9e81fb3"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn",
+ "syn 2.0.119",
 ]
 
 [[package]]
 name = "equivalent"
-version = "1.0.2"
+version = "1.0.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "877a4ace8713b0bcf2a4e7eec82529c029f1d0619886d18145fea96c3ffe5c0f"
+checksum = "00d174d5400e5e8fd687ad1049e2f578285fa914201b1af7e8b112a4546bd826"
 
 [[package]]
 name = "errno"
@@ -1967,37 +1615,16 @@ dependencies = [
  "windows-sys 0.61.2",
 ]
 
-[[package]]
-name = "error-code"
-version = "3.3.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "dea2df4cf52843e0452895c455a1a2cfbb842a1e7329671acf418fdc53ed4c59"
-
 [[package]]
 name = "esbuild-metafile"
-version = "0.5.2"
+version = "0.11.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "309f09d28539a42ca467019db5a74893fc755a921558635964a4b729ea5f5f96"
+checksum = "ced883606dbccc023d30958eb4382e85fdea3bbd87462087ea8f566c01c9ea89"
 dependencies = [
- "actix-utils",
- "actix-web",
- "anyhow",
- "askama",
- "dashmap",
- "once_cell",
- "pretty_assertions",
+ "log",
  "serde",
  "serde_json",
-]
-
-[[package]]
-name = "etagere"
-version = "0.2.15"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "fc89bf99e5dc15954a60f707c1e09d7540e5cd9af85fa75caa0b510bc08c5342"
-dependencies = [
- "euclid",
- "svg_fmt",
+ "thiserror",
 ]
 
 [[package]]
@@ -2010,29 +1637,8 @@ dependencies = [
 ]
 
 [[package]]
-name = "event-listener"
-version = "5.4.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e13b66accf52311f30a0db42147dadea9850cb48cd070028831ae5f5d4b856ab"
-dependencies = [
- "concurrent-queue",
- "parking",
- "pin-project-lite",
-]
-
-[[package]]
-name = "event-listener-strategy"
-version = "0.5.4"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8be9f3dfaaffdae2972880079a491a1a8bb7cbed0b8dd7a347f668b4150a3b93"
-dependencies = [
- "event-listener",
- "pin-project-lite",
-]
-
-[[package]]
-name = "eventsource-stream"
-version = "0.2.3"
+name = "eventsource-stream"
+version = "0.2.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "74fef4569247a5f429d9156b9d0a2599914385dd189c539334c625d8099d90ab"
 dependencies = [
@@ -2043,14 +1649,16 @@ dependencies = [
 
 [[package]]
 name = "exr"
-version = "1.74.0"
+version = "1.74.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "4300e043a56aa2cb633c01af81ca8f699a321879a7854d3896a0ba89056363be"
+checksum = "711fe42c9964295e01ee3fba3f9fe0e1d24b98886950d68efe81b1c76e21adf3"
 dependencies = [
  "bit_field",
  "half",
  "lebe",
- "miniz_oxide",
+ "miniz_oxide 0.8.9",
+ "num-complex",
+ "pulp",
  "rayon-core",
  "smallvec",
  "zune-inflate",
@@ -2058,9 +1666,9 @@ dependencies = [
 
 [[package]]
 name = "fancy-regex"
-version = "0.16.2"
+version = "0.19.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "998b056554fbe42e03ae0e152895cd1a7e1002aec800fdc6635d20270260c46f"
+checksum = "d301f5bf187b3c295fce6468d3875037a0bccc5f6b151c63cac2f85babf21912"
 dependencies = [
  "bit-set",
  "regex-automata",
@@ -2069,29 +1677,15 @@ dependencies = [
 
 [[package]]
 name = "fastrand"
-version = "2.3.0"
+version = "2.5.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "37909eebbb50d72f9059c3b6d82c0463f2ff062c9e95845c43a6c9c0355411be"
+checksum = "da7c62ceae207dd37ea5b845da6a0696c799f85e97da1ab5b7910be3c1c80223"
 
 [[package]]
 name = "fax"
-version = "0.2.6"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f05de7d48f37cd6730705cbca900770cab77a89f413d23e100ad7fad7795a0ab"
-dependencies = [
- "fax_derive",
-]
-
-[[package]]
-name = "fax_derive"
-version = "0.2.0"
+version = "0.2.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a0aca10fb742cb43f9e7bb8467c91aa9bcb8e3ffbc6a6f7389bb93ffc920577d"
-dependencies = [
- "proc-macro2",
- "quote",
- "syn",
-]
+checksum = "caf1079563223d5d59d83c85886a56e586cfd5c1a26292e971a0fa266531ac5a"
 
 [[package]]
 name = "fdeflate"
@@ -2104,9 +1698,9 @@ dependencies = [
 
 [[package]]
 name = "find-msvc-tools"
-version = "0.1.9"
+version = "0.1.15"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5baebc0774151f905a1a2cc41989300b1e6fbb29aff0ceffa1064fdd3088d582"
+checksum = "1c66915cc5391097de7a2b0b972e8505aa33c0df401fd3a6811dbdb32881ba48"
 
 [[package]]
 name = "find_cuda_helper"
@@ -2119,12 +1713,13 @@ dependencies = [
 
 [[package]]
 name = "flate2"
-version = "1.1.9"
+version = "1.1.10"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "843fba2746e448b37e26a819579957415c8cef339bf08564fe8b7ddbd959573c"
+checksum = "6e634e2e0ebac1ee034020da1ca582e17ffe4e0f5e985823721e168928136dcb"
 dependencies = [
  "crc32fast",
- "miniz_oxide",
+ "miniz_oxide 0.9.1",
+ "zlib-rs",
 ]
 
 [[package]]
@@ -2161,12 +1756,6 @@ version = "1.0.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "3f9eec918d3f24069decb9af1554cad7c880e2da24a9afd88aca000531ab82c1"
 
-[[package]]
-name = "foldhash"
-version = "0.1.5"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d9c4f5dac5e15c24eb999c26181a6ca40b39fe946cbe4c263c7209467bc83af2"
-
 [[package]]
 name = "foldhash"
 version = "0.2.0"
@@ -2175,18 +1764,9 @@ checksum = "77ce24cb58228fbb8aa041425bb1050850ac19177686ea6e0f41a70416f56fdb"
 
 [[package]]
 name = "font-types"
-version = "0.10.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "39a654f404bbcbd48ea58c617c2993ee91d1cb63727a37bf2323a4edeed1b8c5"
-dependencies = [
- "bytemuck",
-]
-
-[[package]]
-name = "font-types"
-version = "0.11.3"
+version = "0.12.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5b38ad915f6dadd993ced50848a8291a543bd41ca62bc10740d5e64e2ab4cfd7"
+checksum = "835ae9d781464a79026cde387f9cbc650c9a4909d7cf7e658b127e3c35d2f35d"
 dependencies = [
  "bytemuck",
 ]
@@ -2202,60 +1782,17 @@ dependencies = [
 
 [[package]]
 name = "fontdb"
-version = "0.23.0"
+version = "0.24.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "457e789b3d1202543297a350643cf459f836cade38934e7a4cf6a39e7cde2905"
+checksum = "2660c5e9157bf76d2db1294e4a9feba604ef610819a3b591088d0d8392a3290f"
 dependencies = [
  "fontconfig-parser",
  "log",
  "memmap2",
  "slotmap",
  "tinyvec",
- "ttf-parser",
-]
-
-[[package]]
-name = "foreign-types"
-version = "0.3.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f6f339eb8adc052cd2ca78910fda869aefa38d22d5cb648e6485e4d3fc06f3b1"
-dependencies = [
- "foreign-types-shared 0.1.1",
-]
-
-[[package]]
-name = "foreign-types"
-version = "0.5.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d737d9aa519fb7b749cbc3b962edcf310a8dd1f4b67c91c4f83975dbdd17d965"
-dependencies = [
- "foreign-types-macros",
- "foreign-types-shared 0.3.1",
-]
-
-[[package]]
-name = "foreign-types-macros"
-version = "0.2.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1a5c6c585bc94aaf2c7b51dd4c2ba22680844aba4c687be581871a6f518c5742"
-dependencies = [
- "proc-macro2",
- "quote",
- "syn",
 ]
 
-[[package]]
-name = "foreign-types-shared"
-version = "0.1.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "00b0228411908ca8685dba7fc2cdd70ec9990a6e753e89b6ac91a84c40fbaf4b"
-
-[[package]]
-name = "foreign-types-shared"
-version = "0.3.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "aa9a19cbb55df58761df49b23516a86d432839add4af60fc256da840f66ed35b"
-
 [[package]]
 name = "form_urlencoded"
 version = "1.2.2"
@@ -2267,12 +1804,12 @@ dependencies = [
 
 [[package]]
 name = "fraction"
-version = "0.15.3"
+version = "0.17.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0f158e3ff0a1b334408dc9fb811cd99b446986f4d8b741bb08f9df1604085ae7"
+checksum = "e246562084dde8ebbcc943b261c406ce4f68e5032ec28029a251a47d6a295500"
 dependencies = [
- "lazy_static",
  "num",
+ "num-bigint",
 ]
 
 [[package]]
@@ -2293,15 +1830,9 @@ dependencies = [
 
 [[package]]
 name = "futures"
-version = "0.1.31"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3a471a38ef8ed83cd6e40aa59c1ffe17db6855c18e3604d9c4ed8c08ebc28678"
-
-[[package]]
-name = "futures"
-version = "0.3.32"
+version = "0.3.34"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8b147ee9d1f6d097cef9ce628cd2ee62288d963e16fb287bd9286455b241382d"
+checksum = "9a31d2a3fbaaeb2af2368bbdd904aa8e812d3c04a1ee10d3171f52d556e5d0a3"
 dependencies = [
  "futures-channel",
  "futures-core",
@@ -2314,9 +1845,9 @@ dependencies = [
 
 [[package]]
 name = "futures-channel"
-version = "0.3.32"
+version = "0.3.34"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "07bbe89c50d7a535e539b8c17bc0b49bdb77747034daa8087407d655f3f7cc1d"
+checksum = "b1f9e3d69d39e4862ffed03ed071a76f9a13ba1d9109d355b0f0aa6b15e393c4"
 dependencies = [
  "futures-core",
  "futures-sink",
@@ -2324,15 +1855,15 @@ dependencies = [
 
 [[package]]
 name = "futures-core"
-version = "0.3.32"
+version = "0.3.34"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7e3450815272ef58cec6d564423f6e755e25379b217b0bc688e295ba24df6b1d"
+checksum = "92d699e522242e69e3003b94ecc1f960f3a5e015aa7c5d7486e65ad01dd94f5e"
 
 [[package]]
 name = "futures-executor"
-version = "0.3.32"
+version = "0.3.34"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "baf29c38818342a3b26b5b923639e7b1f4a61fc5e76102d4b1981c6dc7a7579d"
+checksum = "031b47cf1a3c6cc8bc2fc76cd437f521619387907d469316e7c0bc278f1f5432"
 dependencies = [
  "futures-core",
  "futures-task",
@@ -2341,51 +1872,38 @@ dependencies = [
 
 [[package]]
 name = "futures-io"
-version = "0.3.32"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "cecba35d7ad927e23624b22ad55235f2239cfa44fd10428eecbeba6d6a717718"
-
-[[package]]
-name = "futures-lite"
-version = "2.6.1"
+version = "0.3.34"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f78e10609fe0e0b3f4157ffab1876319b5b0db102a2c60dc4626306dc46b44ad"
-dependencies = [
- "fastrand",
- "futures-core",
- "futures-io",
- "parking",
- "pin-project-lite",
-]
+checksum = "53c0fa8157de1303bfffdaa1cc2a673bfffb60102f76b0ef4441659124373fed"
 
 [[package]]
 name = "futures-macro"
-version = "0.3.32"
+version = "0.3.34"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e835b70203e41293343137df5c0664546da5745f82ec9b84d40be8336958447b"
+checksum = "9fb9654ba8355388abeb8dcb4fc62f511300867002afc858860463bdd9fe0c44"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn",
+ "syn 3.0.9",
 ]
 
 [[package]]
 name = "futures-sink"
-version = "0.3.32"
+version = "0.3.34"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c39754e157331b013978ec91992bde1ac089843443c49cbc7f46150b0fad0893"
+checksum = "1944426bf7d03f1d14f708785e4b33efd750b36d48a157b836b3efc15ede8e1d"
 
 [[package]]
 name = "futures-task"
-version = "0.3.32"
+version = "0.3.34"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "037711b3d59c33004d3856fbdc83b99d4ff37a24768fa1be9ce3538a1cde4393"
+checksum = "cd417de3d1d015fc3bfd2b1ea46dfc7bab72ef86f1cc7cc9c78e728b34a6d1fd"
 
 [[package]]
 name = "futures-util"
-version = "0.3.32"
+version = "0.3.34"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "389ca41296e6190b48053de0321d02a77f32f8a5d2461dd38762c0593805c6d6"
+checksum = "0d50a92467f8ba5dd6e3ee5d4bd04d73ab2e4e1c44474a0674821dfce14b79bc"
 dependencies = [
  "futures-channel",
  "futures-core",
@@ -2396,9 +1914,14 @@ dependencies = [
  "memchr",
  "pin-project-lite",
  "slab",
- "tokio-io",
 ]
 
+[[package]]
+name = "gearhash"
+version = "0.1.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "616e8f476a586f1b078d9eece21f7a071f27997e709e2b471f5697deea53bda9"
+
 [[package]]
 name = "generic-array"
 version = "0.14.7"
@@ -2409,16 +1932,6 @@ dependencies = [
  "version_check",
 ]
 
-[[package]]
-name = "gethostname"
-version = "1.1.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1bd49230192a3797a9a4d6abe9b3eed6f7fa4c8a8a4947977c6f80025f92cbd8"
-dependencies = [
- "rustix 1.1.4",
- "windows-link",
-]
-
 [[package]]
 name = "getrandom"
 version = "0.2.17"
@@ -2428,7 +1941,7 @@ dependencies = [
  "cfg-if",
  "js-sys",
  "libc",
- "wasi",
+ "wasi 0.11.1+wasi-snapshot-preview1",
  "wasm-bindgen",
 ]
 
@@ -2448,176 +1961,79 @@ dependencies = [
 
 [[package]]
 name = "getrandom"
-version = "0.4.2"
+version = "0.4.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0de51e6874e94e7bf76d726fc5d13ba782deca734ff60d5bb2fb2607c7406555"
+checksum = "300e883d756b2e4ec94e02791f39b04b522276138852cfc41d9fb7e904106099"
 dependencies = [
  "cfg-if",
  "js-sys",
  "libc",
  "r-efi 6.0.0",
  "rand_core 0.10.1",
- "wasip2",
- "wasip3",
  "wasm-bindgen",
 ]
 
 [[package]]
 name = "gif"
-version = "0.13.3"
+version = "0.14.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "4ae047235e33e2829703574b54fdec96bfbad892062d97fed2f76022287de61b"
+checksum = "ee8cfcc411d9adbbaba82fb72661cc1bcca13e8bba98b364e62b2dba8f960159"
 dependencies = [
  "color_quant",
  "weezl",
 ]
 
 [[package]]
-name = "gif"
-version = "0.14.1"
+name = "git-version"
+version = "0.3.9"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f5df2ba84018d80c213569363bdcd0c64e6933c67fe4c1d60ecf822971a3c35e"
+checksum = "1ad568aa3db0fcbc81f2f116137f263d7304f512a1209b35b85150d3ef88ad19"
 dependencies = [
- "color_quant",
- "weezl",
+ "git-version-macro",
 ]
 
 [[package]]
-name = "gl_generator"
-version = "0.14.0"
+name = "git-version-macro"
+version = "0.3.9"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1a95dfc23a2b4a9a2f5ab41d194f8bfda3cabec42af4e39f08c339eb2a0c124d"
+checksum = "53010ccb100b96a67bc32c0175f0ed1426b31b655d562898e57325f81c023ac0"
 dependencies = [
- "khronos_api",
- "log",
- "xml-rs",
+ "proc-macro2",
+ "quote",
+ "syn 2.0.119",
 ]
 
-[[package]]
-name = "glam"
-version = "0.25.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "151665d9be52f9bb40fc7966565d39666f2d1e69233571b71b87791c7e0528b3"
-
 [[package]]
 name = "glob"
-version = "0.3.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0cc23270f6e1808e30a928bdc84dea0b9b4136a8bc82338574f23baf47bbd280"
-
-[[package]]
-name = "glow"
-version = "0.16.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c5e5ea60d70410161c8bf5da3fdfeaa1c72ed2c15f8bbb9d19fe3a4fad085f08"
-dependencies = [
- "js-sys",
- "slotmap",
- "wasm-bindgen",
- "web-sys",
-]
-
-[[package]]
-name = "glutin_wgl_sys"
-version = "0.6.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2c4ee00b289aba7a9e5306d57c2d05499b2e5dc427f84ac708bd2c090212cf3e"
-dependencies = [
- "gl_generator",
-]
-
-[[package]]
-name = "gpu-alloc"
-version = "0.6.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "fbcd2dba93594b227a1f57ee09b8b9da8892c34d55aa332e034a228d0fe6a171"
-dependencies = [
- "bitflags 2.11.0",
- "gpu-alloc-types",
-]
-
-[[package]]
-name = "gpu-alloc-types"
-version = "0.3.0"
+version = "0.3.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "98ff03b468aa837d70984d55f5d3f846f6ec31fe34bbb97c4f85219caeee1ca4"
-dependencies = [
- "bitflags 2.11.0",
-]
+checksum = "e4eba85ea1d0a966a983acd07deee566e67395d2d96b6fb39e62b5a833f1eb0b"
 
 [[package]]
-name = "gpu-allocator"
-version = "0.27.0"
+name = "globset"
+version = "0.4.20"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c151a2a5ef800297b4e79efa4f4bec035c5f51d5ae587287c9b952bdf734cacd"
+checksum = "07c34a9410465b45bd9787443bc7370f37735bad04b0f0cd57ff1a3186c98988"
 dependencies = [
+ "aho-corasick",
+ "bstr",
  "log",
- "presser",
- "thiserror 1.0.69",
- "windows 0.58.0",
-]
-
-[[package]]
-name = "gpu-descriptor"
-version = "0.3.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b89c83349105e3732062a895becfc71a8f921bb71ecbbdd8ff99263e3b53a0ca"
-dependencies = [
- "bitflags 2.11.0",
- "gpu-descriptor-types",
- "hashbrown 0.15.5",
-]
-
-[[package]]
-name = "gpu-descriptor-types"
-version = "0.2.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "fdf242682df893b86f33a73828fb09ca4b2d3bb6cc95249707fc684d27484b91"
-dependencies = [
- "bitflags 2.11.0",
-]
-
-[[package]]
-name = "guillotiere"
-version = "0.6.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b62d5865c036cb1393e23c50693df631d3f5d7bcca4c04fe4cc0fd592e74a782"
-dependencies = [
- "euclid",
- "svg_fmt",
-]
-
-[[package]]
-name = "h2"
-version = "0.3.27"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0beca50380b1fc32983fc1cb4587bfa4bb9e78fc259aad4a0032d2080309222d"
-dependencies = [
- "bytes 1.11.1",
- "fnv",
- "futures-core",
- "futures-sink",
- "futures-util",
- "http 0.2.12",
- "indexmap",
- "slab",
- "tokio",
- "tokio-util",
- "tracing",
+ "regex-automata",
+ "regex-syntax",
 ]
 
 [[package]]
 name = "h2"
-version = "0.4.19"
+version = "0.4.20"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ef8e5e5a340588f4452631496976cf8636d4a7ecf600239fdc27615d2530bc16"
+checksum = "7d29020232d6aa3fb1daca64c1127cf662cf97f254ae16c18c05b8ab635fc118"
 dependencies = [
  "atomic-waker",
- "bytes 1.11.1",
+ "bytes",
  "fnv",
  "futures-core",
  "futures-sink",
- "http 1.4.0",
+ "http 1.5.0",
  "indexmap",
  "slab",
  "tokio",
@@ -2633,20 +2049,18 @@ checksum = "6ea2d84b969582b4b1864a92dc5d27cd2b77b622a8d79306834f1be5ba20d84b"
 dependencies = [
  "cfg-if",
  "crunchy",
- "num-traits",
  "zerocopy",
 ]
 
 [[package]]
 name = "harfrust"
-version = "0.3.2"
+version = "0.12.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "92c020db12c71d8a12a3fe7607873cade3a01a6287e29d540c8723276221b9d8"
+checksum = "c03d949a14aa089bbb282f7dd76a498a7f684428e4257202efc119ec010376f9"
 dependencies = [
- "bitflags 2.11.0",
+ "bitflags 2.13.2",
  "bytemuck",
- "core_maths",
- "read-fonts 0.35.0",
+ "read-fonts",
  "smallvec",
 ]
 
@@ -2661,43 +2075,40 @@ name = "hashbrown"
 version = "0.15.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "9229cfe53dfd69f0609a49f65461bd93001ea1ef889cd5529dd176593f5338a1"
-dependencies = [
- "foldhash 0.1.5",
-]
 
 [[package]]
 name = "hashbrown"
-version = "0.16.1"
+version = "0.17.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "841d1cc9bed7f9236f321df977030373f4a4163ae1a7dbfe1a51a2c1a51d9100"
+checksum = "ed5909b6e89a2db4456e54cd5f673791d7eca6732202bbf2a9cc504fe2f9b84a"
 dependencies = [
  "allocator-api2",
  "equivalent",
- "foldhash 0.2.0",
+ "foldhash",
 ]
 
 [[package]]
 name = "hashlink"
-version = "0.11.0"
+version = "0.12.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ea0b22561a9c04a7cb1a302c013e0259cd3b4bb619f145b32f72b8b4bcbed230"
+checksum = "a596f1b20ed2cc5ecac41a164aaebc7258057060f06c0cf7a2ba3991ee7990fb"
 dependencies = [
- "hashbrown 0.16.1",
+ "hashbrown 0.17.1",
 ]
 
 [[package]]
 name = "headers"
-version = "0.4.1"
+version = "0.4.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b3314d5adb5d94bcdf56771f2e50dbbc80bb4bdf88967526706205ac9eff24eb"
+checksum = "fc2700e3fe555c3310aa7286cac6167449f2c87e0eb58769c9208a1c58a1d106"
 dependencies = [
- "base64",
- "bytes 1.11.1",
+ "base64 0.22.1",
+ "bytes",
  "headers-core",
- "http 1.4.0",
+ "http 1.5.0",
  "httpdate",
  "mime",
- "sha1 0.10.6",
+ "sha1 0.10.7",
 ]
 
 [[package]]
@@ -2706,55 +2117,73 @@ version = "0.3.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "54b4a22553d4242c49fddb9ba998a99962b5cc6f22cb5a3482bec22522403ce4"
 dependencies = [
- "http 1.4.0",
+ "http 1.5.0",
 ]
 
 [[package]]
-name = "heck"
-version = "0.5.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2304e00983f87ffb38b55b444b5e3b60a884b5d30c0fca7d82fe33449bbe55ea"
-
-[[package]]
-name = "hermit-abi"
-version = "0.5.2"
+name = "heapify"
+version = "0.2.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "fc0fef456e4baa96da950455cd02c081ca953b141298e41db3fc7e36b1da849c"
+checksum = "0049b265b7f201ca9ab25475b22b47fe444060126a51abe00f77d986fc5cc52e"
 
 [[package]]
-name = "hex"
-version = "0.4.3"
+name = "heck"
+version = "0.5.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7f24254aa9a54b5c858eaee2f5bccdb46aaf0e486a595ed5fd8f86ba55232a70"
+checksum = "2304e00983f87ffb38b55b444b5e3b60a884b5d30c0fca7d82fe33449bbe55ea"
 
 [[package]]
-name = "hexf-parse"
-version = "0.2.1"
+name = "hf-hub"
+version = "1.0.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "dfa686283ad6dd069f105e5ab091b04c62850d3e4cf5d67debad1933f55023df"
+checksum = "e7ccb6bcc85dec15413ef5949879f9a5497ca4568ed702547eb83fac23376e5c"
+dependencies = [
+ "base64 0.22.1",
+ "bon",
+ "bytes",
+ "futures",
+ "getrandom 0.2.17",
+ "globset",
+ "hf-xet",
+ "hyper",
+ "pathdiff",
+ "percent-encoding",
+ "reqwest",
+ "serde",
+ "serde_json",
+ "sha2",
+ "thiserror",
+ "tokio",
+ "tokio-retry",
+ "tokio-util",
+ "tracing",
+ "url",
+ "wasm-bindgen-futures",
+]
 
 [[package]]
-name = "hf-hub"
-version = "0.4.3"
+name = "hf-xet"
+version = "1.7.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "629d8f3bbeda9d148036d6b0de0a3ab947abd08ce90626327fc3547a49d59d97"
+checksum = "f24754aef4f6c3d57bc4a0dc2648ca74487050eda01c361b3eefa11feeafd662"
 dependencies = [
- "dirs",
- "futures 0.3.32",
- "http 1.4.0",
- "indicatif",
- "libc",
- "log",
- "native-tls",
- "num_cpus",
- "rand 0.9.4",
- "reqwest 0.12.28",
+ "anyhow",
+ "async-trait",
+ "bytes",
+ "http 1.5.0",
+ "more-asserts",
  "serde",
  "serde_json",
- "thiserror 2.0.18",
+ "thiserror",
  "tokio",
- "ureq",
- "windows-sys 0.60.2",
+ "tokio-util",
+ "tokio_with_wasm",
+ "tracing",
+ "uuid",
+ "xet-client",
+ "xet-core-structures",
+ "xet-data",
+ "xet-runtime",
 ]
 
 [[package]]
@@ -2763,40 +2192,40 @@ version = "0.2.12"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "601cbb57e577e2f5ef5be8e7b83f0f63994f25aa94d673e54a92d5c516d101f1"
 dependencies = [
- "bytes 1.11.1",
+ "bytes",
  "fnv",
  "itoa",
 ]
 
 [[package]]
 name = "http"
-version = "1.4.0"
+version = "1.5.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e3ba2a386d7f85a81f119ad7498ebe444d2e22c2af0b86b069416ace48b3311a"
+checksum = "918d3568bebf352712bc2ef3d46a8bcf1a75b373be6539de198e9105cbbf9ce0"
 dependencies = [
- "bytes 1.11.1",
+ "bytes",
  "itoa",
 ]
 
 [[package]]
 name = "http-body"
-version = "1.0.1"
+version = "1.1.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1efedce1fb8e6913f23e0c92de8e62cd5b772a67e7b3946df930a62566c93184"
+checksum = "ca2a8f2913ee65f60facd6a5905613afaa448497a0230cc41ce022d93290bc2c"
 dependencies = [
- "bytes 1.11.1",
- "http 1.4.0",
+ "bytes",
+ "http 1.5.0",
 ]
 
 [[package]]
 name = "http-body-util"
-version = "0.1.3"
+version = "0.1.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b021d93e26becf5dc7e1b75b1bed1fd93124b374ceb73f43d4d4eafec896a64a"
+checksum = "23169fe34a5fbcdd3f3862e78fb9b6fccd5f02a6dc6f732547005d45631ce71c"
 dependencies = [
- "bytes 1.11.1",
+ "bytes",
  "futures-core",
- "http 1.4.0",
+ "http 1.5.0",
  "http-body",
  "pin-project-lite",
 ]
@@ -2813,27 +2242,33 @@ version = "1.0.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "df3b46402a9d5adb4c86a0cf463f42e19994e3ee891101b1841f30a545cb49a9"
 
+[[package]]
+name = "humantime"
+version = "2.4.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "15cdd26707701c53297e2fa6afb323d55fbc1d0810c3aec078ae3ef0424c3c15"
+
 [[package]]
 name = "hybrid-array"
-version = "0.4.10"
+version = "0.4.15"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3944cf8cf766b40e2a1a333ee5e9b563f854d5fa49d6a8ca2764e97c6eddb214"
+checksum = "27f864f10dfb56725ce5ce5472bc52252c8f93a4ab86327122cebf62c5f59a17"
 dependencies = [
  "typenum",
 ]
 
 [[package]]
 name = "hyper"
-version = "1.9.0"
+version = "1.12.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6299f016b246a94207e63da54dbe807655bf9e00044f73ded42c3ac5305fbcca"
+checksum = "2c3e324da4c95177d6291d4c8730197c0d1822f8a9766814a4a44fa5ab797c9c"
 dependencies = [
  "atomic-waker",
- "bytes 1.11.1",
+ "bytes",
  "futures-channel",
  "futures-core",
- "h2 0.4.19",
- "http 1.4.0",
+ "h2",
+ "http 1.5.0",
  "http-body",
  "httparse",
  "itoa",
@@ -2845,48 +2280,32 @@ dependencies = [
 
 [[package]]
 name = "hyper-rustls"
-version = "0.27.7"
+version = "0.27.10"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e3c93eb611681b207e1fe55d5a71ecf91572ec8a6705cdb6857f7d8d5242cf58"
+checksum = "dfa8e654703247911e29c23fbeaa261834bd9bb74efba2f9acddc37bfb127f53"
 dependencies = [
- "http 1.4.0",
+ "http 1.5.0",
  "hyper",
  "hyper-util",
  "rustls",
- "rustls-pki-types",
  "tokio",
  "tokio-rustls",
  "tower-service",
 ]
 
 [[package]]
-name = "hyper-tls"
-version = "0.6.0"
+name = "hyper-util"
+version = "0.1.21"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "70206fc6890eaca9fde8a0bf71caa2ddfc9fe045ac9e5c70df101a7dbde866e0"
+checksum = "ddc03d96684f9226b8a787cdb71488417b53ab5ea8fdb1dac946cb9431cc8bff"
 dependencies = [
- "bytes 1.11.1",
- "http-body-util",
- "hyper",
- "hyper-util",
- "native-tls",
- "tokio",
- "tokio-native-tls",
- "tower-service",
-]
-
-[[package]]
-name = "hyper-util"
-version = "0.1.20"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "96547c2556ec9d12fb1578c4eaf448b04993e7fb79cbaad930a656880a6bdfa0"
-dependencies = [
- "base64",
- "bytes 1.11.1",
- "futures-channel",
- "futures-util",
- "http 1.4.0",
- "http-body",
+ "base64 0.23.1",
+ "bytes",
+ "futures-channel",
+ "futures-util",
+ "http 1.5.0",
+ "http-body",
+ "httparse",
  "hyper",
  "ipnet",
  "libc",
@@ -2901,226 +2320,34 @@ dependencies = [
 ]
 
 [[package]]
-name = "iced"
-version = "0.14.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "000e01026c93ba643f8357a3db3ada0e6555265a377f6f9291c472f6dd701fb3"
-dependencies = [
- "iced_core",
- "iced_debug",
- "iced_futures",
- "iced_renderer",
- "iced_runtime",
- "iced_widget",
- "iced_winit",
- "image",
- "thiserror 2.0.18",
-]
-
-[[package]]
-name = "iced_core"
-version = "0.14.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "91ab1937d699403e7e69252ae743a902bcee9f4ab2052cc4c9a46fcf34729d85"
-dependencies = [
- "bitflags 2.11.0",
- "bytes 1.11.1",
- "glam",
- "lilt",
- "log",
- "num-traits",
- "rustc-hash 2.1.2",
- "smol_str",
- "thiserror 2.0.18",
- "web-time",
-]
-
-[[package]]
-name = "iced_debug"
-version = "0.14.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "25035ab0215a620e53f4103e36fc4e59a1fb2817e4bfc38a30ad27b4202ea0be"
-dependencies = [
- "iced_core",
- "iced_futures",
- "log",
-]
-
-[[package]]
-name = "iced_futures"
-version = "0.14.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8c0c85ccad42dfbec7293c36c018af0ea0dbcc52d137a4a9a0b0f6822a3fdf0a"
-dependencies = [
- "futures 0.3.32",
- "iced_core",
- "log",
- "rustc-hash 2.1.2",
- "tokio",
- "wasm-bindgen-futures",
- "wasmtimer",
-]
-
-[[package]]
-name = "iced_graphics"
-version = "0.14.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "234ca1c2cec4155055f68fa5fad1b5242c496ac8238d80a259bca382fb44a102"
-dependencies = [
- "bitflags 2.11.0",
- "bytemuck",
- "cosmic-text",
- "half",
- "iced_core",
- "iced_futures",
- "image",
- "kamadak-exif",
- "log",
- "raw-window-handle",
- "rustc-hash 2.1.2",
- "thiserror 2.0.18",
- "unicode-segmentation",
-]
-
-[[package]]
-name = "iced_program"
-version = "0.14.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6dfafec2947cda688d8eb00dac337ba11aa60f9ef6335aed343e189d26e4a673"
-dependencies = [
- "iced_graphics",
- "iced_runtime",
-]
-
-[[package]]
-name = "iced_renderer"
-version = "0.14.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "250cc0802408e8c077986ec56c7d07c65f423ee658a4b9fd795a1f2aae5dac05"
-dependencies = [
- "iced_graphics",
- "iced_tiny_skia",
- "iced_wgpu",
- "log",
- "thiserror 2.0.18",
-]
-
-[[package]]
-name = "iced_runtime"
-version = "0.14.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d1889b819ce4c06674183242e336c8d49465665441396914dc07cc86f44fa8d4"
-dependencies = [
- "bytes 1.11.1",
- "iced_core",
- "iced_futures",
- "raw-window-handle",
- "thiserror 2.0.18",
-]
-
-[[package]]
-name = "iced_selector"
-version = "0.14.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b1442ccc30959cc48b3a624233521ef7f3bc3e06528b6fe306c3a775fbddc4fe"
-dependencies = [
- "iced_core",
-]
-
-[[package]]
-name = "iced_test"
-version = "0.14.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b56c26cc6e63958ce6be06a68fa3136fd5f646da6302ee12abf1a0f311dd0166"
-dependencies = [
- "iced_futures",
- "iced_program",
- "iced_renderer",
- "iced_runtime",
- "iced_selector",
- "nom 8.0.0",
- "png 0.18.1",
- "sha2",
- "thiserror 2.0.18",
-]
-
-[[package]]
-name = "iced_tiny_skia"
-version = "0.14.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "fe0acf8b75a3bc914aff5f2329fdffc1b36eeaea29dda0e4bd232f1c62e9cc3d"
-dependencies = [
- "bytemuck",
- "cosmic-text",
- "iced_debug",
- "iced_graphics",
- "kurbo 0.10.4",
- "log",
- "resvg 0.45.1",
- "rustc-hash 2.1.2",
- "softbuffer",
- "tiny-skia",
-]
-
-[[package]]
-name = "iced_wgpu"
-version = "0.14.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ff144a999b0ca0f8a10257934500060240825c42e950ec0ebee9c8ae30561c13"
-dependencies = [
- "bitflags 2.11.0",
- "bytemuck",
- "cryoglyph",
- "futures 0.3.32",
- "glam",
- "guillotiere",
- "iced_debug",
- "iced_graphics",
- "log",
- "resvg 0.45.1",
- "rustc-hash 2.1.2",
- "thiserror 2.0.18",
- "wgpu",
-]
-
-[[package]]
-name = "iced_widget"
-version = "0.14.2"
+name = "iana-time-zone"
+version = "0.1.65"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b1596afa0d3109c2618e8bc12bae6c11d3064df8f95c42dfce570397dbe957ab"
+checksum = "e31bc9ad994ba00e440a8aa5c9ef0ec67d5cb5e5cb0cc7f8b744a35b389cc470"
 dependencies = [
- "iced_renderer",
+ "android_system_properties",
+ "core-foundation-sys",
+ "iana-time-zone-haiku",
+ "js-sys",
  "log",
- "num-traits",
- "rustc-hash 2.1.2",
- "thiserror 2.0.18",
- "unicode-segmentation",
+ "wasm-bindgen",
+ "windows-core",
 ]
 
 [[package]]
-name = "iced_winit"
-version = "0.14.0"
+name = "iana-time-zone-haiku"
+version = "0.1.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8b7dbedc47562d1de3b9707d939f678b88c382004b7ab5a18f7a7dd723162d75"
+checksum = "f31827a206f56af32e590ba56d5d2d085f558508192593743f16b2306495269f"
 dependencies = [
- "iced_debug",
- "iced_program",
- "log",
- "mundy",
- "rustc-hash 2.1.2",
- "thiserror 2.0.18",
- "tracing",
- "wasm-bindgen-futures",
- "web-sys",
- "window_clipboard",
- "winit",
+ "cc",
 ]
 
 [[package]]
 name = "icu_collections"
-version = "2.2.0"
+version = "2.4.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2984d1cd16c883d7935b9e07e44071dca8d917fd52ecc02c04d5fa0b5a3f191c"
+checksum = "fe090a42a0d3140fed8c63263ed1fecc02422cb2f4bfe55cd8de2de1b8e105dd"
 dependencies = [
  "displaydoc",
  "potential_utf",
@@ -3132,9 +2359,9 @@ dependencies = [
 
 [[package]]
 name = "icu_locale_core"
-version = "2.2.0"
+version = "2.4.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "92219b62b3e2b4d88ac5119f8904c10f8f61bf7e95b640d25ba3075e6cac2c29"
+checksum = "25d29b57c17c6528d0dc4c2141cfa2d08200465acb537a3909c8b9b8a4a984c0"
 dependencies = [
  "displaydoc",
  "litemap",
@@ -3145,30 +2372,34 @@ dependencies = [
 
 [[package]]
 name = "icu_normalizer"
-version = "2.2.0"
+version = "2.4.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c56e5ee99d6e3d33bd91c5d85458b6005a22140021cc324cea84dd0e72cff3b4"
+checksum = "404a3b8c486f205c87fe7f23ab6894d84a4ad76d6e4671ae1981ba7def87c599"
 dependencies = [
  "icu_collections",
  "icu_normalizer_data",
  "icu_properties",
  "icu_provider",
  "smallvec",
+ "utf16_iter",
+ "utf8_iter",
+ "write16",
  "zerovec",
 ]
 
 [[package]]
 name = "icu_normalizer_data"
-version = "2.2.0"
+version = "2.4.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "da3be0ae77ea334f4da67c12f149704f19f81d1adf7c51cf482943e84a2bad38"
+checksum = "20f0088bfb7441f7a37fa0c9e20cb1e9cba1cac6b9ce8cd2a6dc44294f8e8cbc"
 
 [[package]]
 name = "icu_properties"
-version = "2.2.0"
+version = "2.4.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "bee3b67d0ea5c2cca5003417989af8996f8604e34fb9ddf96208a033901e70de"
+checksum = "2ea83d991c56158f7cf898bf86436c1dfe7c07ee82fe88951f2695c9f37a379f"
 dependencies = [
+ "displaydoc",
  "icu_collections",
  "icu_locale_core",
  "icu_properties_data",
@@ -3179,15 +2410,15 @@ dependencies = [
 
 [[package]]
 name = "icu_properties_data"
-version = "2.2.0"
+version = "2.4.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8e2bbb201e0c04f7b4b3e14382af113e17ba4f63e2c9d2ee626b720cbce54a14"
+checksum = "b98f29af82d02b163fe50161388350e743bb3107cc8e8c2f7314cb505faddfcf"
 
 [[package]]
 name = "icu_provider"
-version = "2.2.0"
+version = "2.4.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "139c4cf31c8b5f33d7e199446eff9c1e02decfc2f0eec2c8d71f65befa45b421"
+checksum = "413099315636c588ee8f011190b329b971a09a8fae9d157e01dcae8379468514"
 dependencies = [
  "displaydoc",
  "icu_locale_core",
@@ -3198,12 +2429,6 @@ dependencies = [
  "zerovec",
 ]
 
-[[package]]
-name = "id-arena"
-version = "2.3.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3d3067d79b975e8844ca9eb072e16b31c3c1c36928edf9c6789548c524d0d954"
-
 [[package]]
 name = "ident_case"
 version = "1.0.1"
@@ -3223,24 +2448,14 @@ dependencies = [
 
 [[package]]
 name = "idna_adapter"
-version = "1.2.1"
+version = "1.2.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3acae9609540aa318d1bc588455225fb2085b9ed0c4f6bd0d9d5bcd86f1a0344"
+checksum = "cb68373c0d6620ef8105e855e7745e18b0d00d3bdb07fb532e434244cdb9a714"
 dependencies = [
  "icu_normalizer",
  "icu_properties",
 ]
 
-[[package]]
-name = "if-addrs"
-version = "0.13.4"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "69b2eeee38fef3aa9b4cc5f1beea8a2444fc00e7377cafae396de3f5c2065e24"
-dependencies = [
- "libc",
- "windows-sys 0.59.0",
-]
-
 [[package]]
 name = "image"
 version = "0.25.10"
@@ -3251,18 +2466,18 @@ dependencies = [
  "byteorder-lite",
  "color_quant",
  "exr",
- "gif 0.14.1",
+ "gif",
  "image-webp",
  "moxcms",
  "num-traits",
- "png 0.18.1",
+ "png",
  "qoi",
  "ravif",
  "rayon",
  "rgb",
  "tiff",
- "zune-core 0.5.1",
- "zune-jpeg 0.5.15",
+ "zune-core",
+ "zune-jpeg",
 ]
 
 [[package]]
@@ -3277,64 +2492,37 @@ dependencies = [
 
 [[package]]
 name = "imagesize"
-version = "0.13.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "edcd27d72f2f071c64249075f42e205ff93c9a4c5f6c6da53e79ed9f9832c285"
-
-[[package]]
-name = "imagesize"
-version = "0.14.0"
+version = "0.15.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "09e54e57b4c48b40f7aec75635392b12b3421fa26fe8b4332e63138ed278459c"
+checksum = "65b27460c2c92b037f3f94c538ed9a3342f3fdf923606781629ccb35f82d042a"
 
 [[package]]
 name = "imgref"
-version = "1.12.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e7c5cedc30da3a610cac6b4ba17597bdf7152cf974e8aab3afb3d54455e371c8"
-
-[[package]]
-name = "impl-more"
-version = "0.1.9"
+version = "1.12.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e8a5a9a0ff0086c7a148acb942baaabeadf9504d10400b5a05645853729b9cd2"
+checksum = "6e44b0a4eaa4c82f441d50a963f2d5f05a787240aeee097597033e72accfd22f"
 
 [[package]]
 name = "impl-more"
-version = "0.3.1"
+version = "0.3.10"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "35a84fd5aa25fae5c0f4a33d9cac2ca017fc622cbd089be2229993514990f870"
+checksum = "2d3a73c82a0b0747dba739b380c046a140b5ae747234bf701df3460282da7193"
 
 [[package]]
 name = "indexmap"
-version = "2.13.1"
+version = "2.14.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "45a8a2b9cb3e0b0c1803dbb0758ffac5de2f425b23c28f518faabd9d805342ff"
+checksum = "cc4e190f5d26ca7051642629da2c52fc03bde85a03197c99408dcd291734c855"
 dependencies = [
  "equivalent",
- "hashbrown 0.16.1",
- "serde",
- "serde_core",
-]
-
-[[package]]
-name = "indicatif"
-version = "0.17.11"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "183b3088984b400f4cfac3620d5e076c84da5364016b4f49473de574b2586235"
-dependencies = [
- "console",
- "number_prefix",
- "portable-atomic",
- "unicode-width",
- "web-time",
+ "hashbrown 0.17.1",
 ]
 
 [[package]]
 name = "indoc"
-version = "2.0.7"
+version = "2.0.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "79cf5c93f93228cf8efb3ba362535fb11199ac548a09ce117c9b1adc3030d706"
+checksum = "a37b2691796cffeb8a8cd305ac66e65841559f147f4e63231d0eafa4db5384d1"
 dependencies = [
  "rustversion",
 ]
@@ -3347,52 +2535,14 @@ checksum = "c34819042dc3d3971c46c2190835914dfbe0c3c13f61449b2997f4e9722dfa60"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn",
-]
-
-[[package]]
-name = "iovec"
-version = "0.1.4"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b2b3ea6ff95e175473f8ffe6a7eb7c00d054240321b84c57051175fe3c1e075e"
-dependencies = [
- "libc",
+ "syn 2.0.119",
 ]
 
 [[package]]
 name = "ipnet"
-version = "2.12.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d98f6fed1fde3f8c21bc40a1abb88dd75e67924f9cffc3ef95607bad8017f8e2"
-
-[[package]]
-name = "iri-string"
-version = "0.7.12"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "25e659a4bb38e810ebc252e53b5814ff908a8c58c2a9ce2fae1bbec24cbf4e20"
-dependencies = [
- "memchr",
- "serde",
-]
-
-[[package]]
-name = "is-docker"
-version = "0.2.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "928bae27f42bc99b60d9ac7334e3a21d10ad8f1835a4e12ec3ec0464765ed1b3"
-dependencies = [
- "once_cell",
-]
-
-[[package]]
-name = "is-wsl"
-version = "0.4.0"
+version = "2.12.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "173609498df190136aa7dea1a91db051746d339e18476eed5ca40521f02d7aa5"
-dependencies = [
- "is-docker",
- "once_cell",
-]
+checksum = "791930b43c0d5973160d90a8f3894509f2b273430f5c5c73b668636d0287c5c0"
 
 [[package]]
 name = "is_terminal_polyfill"
@@ -3418,6 +2568,15 @@ dependencies = [
  "either",
 ]
 
+[[package]]
+name = "itertools"
+version = "0.15.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8b4baf93f58d4425749ca49a51c50ebab072c5df6994d08fed93541c331481dc"
+dependencies = [
+ "either",
+]
+
 [[package]]
 name = "itoa"
 version = "1.0.18"
@@ -3426,10 +2585,12 @@ checksum = "8f42a60cbdf9a97f5d2305f08a87dc4e09308d1276d28c869c684d7777685682"
 
 [[package]]
 name = "jiff"
-version = "0.2.23"
+version = "0.2.38"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1a3546dc96b6d42c5f24902af9e2538e82e39ad350b0c766eb3fbf2d8f3d8359"
+checksum = "b2b005715dcbeb0089a3c0dab99f2ff1cc3b2525323552703d648585d342a383"
 dependencies = [
+ "defmt",
+ "jiff-core",
  "jiff-static",
  "log",
  "portable-atomic",
@@ -3438,30 +2599,25 @@ dependencies = [
 ]
 
 [[package]]
-name = "jiff-static"
-version = "0.2.23"
+name = "jiff-core"
+version = "0.1.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2a8c8b344124222efd714b73bb41f8b5120b27a7cc1c75593a6ff768d9d05aa4"
+checksum = "5e52fe76043ccecc9005d2305ebaadf7d7fc0cc89ca6baa10a94d6bc68c7128c"
 dependencies = [
- "proc-macro2",
- "quote",
- "syn",
+ "defmt",
+ "log",
 ]
 
 [[package]]
-name = "jni"
-version = "0.21.1"
+name = "jiff-static"
+version = "0.2.38"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1a87aa2bb7d2af34197c04845522473242e1aa17c12f4935d5856491a7fb8c97"
+checksum = "2cc9817253cf7c7ee4684451bd327e88d6f3658014e54a29198625590650695c"
 dependencies = [
- "cesu8",
- "cfg-if",
- "combine",
- "jni-sys 0.3.1",
- "log",
- "thiserror 1.0.69",
- "walkdir",
- "windows-sys 0.45.0",
+ "jiff-core",
+ "proc-macro2",
+ "quote",
+ "syn 2.0.119",
 ]
 
 [[package]]
@@ -3473,10 +2629,10 @@ dependencies = [
  "cfg-if",
  "combine",
  "jni-macros",
- "jni-sys 0.4.1",
+ "jni-sys",
  "log",
  "simd_cesu8",
- "thiserror 2.0.18",
+ "thiserror",
  "walkdir",
  "windows-link",
 ]
@@ -3491,16 +2647,7 @@ dependencies = [
  "quote",
  "rustc_version",
  "simd_cesu8",
- "syn",
-]
-
-[[package]]
-name = "jni-sys"
-version = "0.3.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "41a652e1f9b6e0275df1f15b32661cf0d4b78d4d87ddec5e0c3c20f097433258"
-dependencies = [
- "jni-sys 0.4.1",
+ "syn 2.0.119",
 ]
 
 [[package]]
@@ -3519,36 +2666,35 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "38c0b942f458fe50cdac086d2f946512305e5631e720728f2a61aabcd47a6264"
 dependencies = [
  "quote",
- "syn",
+ "syn 2.0.119",
 ]
 
 [[package]]
 name = "jobserver"
-version = "0.1.34"
+version = "0.1.35"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9afb3de4395d6b3e67a780b6de64b51c978ecf11cb9a462c66be7d4ca9039d33"
+checksum = "1c00acbd29eabad4a2392fa0e921c874934dbbf4194312ad20f04a0ed67a3cb3"
 dependencies = [
- "getrandom 0.3.4",
+ "getrandom 0.4.3",
  "libc",
 ]
 
 [[package]]
 name = "js-sys"
-version = "0.3.94"
+version = "0.3.106"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2e04e2ef80ce82e13552136fabeef8a5ed1f985a96805761cbb9a2c34e7664d9"
+checksum = "7883d941dae510fb2d978fc3fe018c71c9e2892fd38854de3e8b92c2e5ad9cc5"
 dependencies = [
  "cfg-if",
  "futures-util",
- "once_cell",
  "wasm-bindgen",
 ]
 
 [[package]]
 name = "jsonschema"
-version = "0.37.4"
+version = "0.60.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "73c9ffb2b5c56d58030e1b532d8e8389da94590515f118cf35b5cb68e4764a7e"
+checksum = "65de752665c4a06528af6759955eecf60ba73ae7442d35349963a09c8031cd71"
 dependencies = [
  "ahash",
  "bytecount",
@@ -3557,75 +2703,73 @@ dependencies = [
  "fancy-regex",
  "fraction",
  "getrandom 0.3.4",
- "idna",
  "itoa",
+ "jsonschema-regex",
+ "jsonschema-value",
  "num-cmp",
  "num-traits",
  "percent-encoding",
- "referencing 0.37.4",
+ "referencing 0.60.1",
  "regex",
- "regex-syntax",
  "serde",
  "serde_json",
+ "strum 0.28.0",
  "unicode-general-category",
  "uuid-simd",
 ]
 
 [[package]]
-name = "kamadak-exif"
-version = "0.6.1"
+name = "jsonschema-regex"
+version = "0.60.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1130d80c7374efad55a117d715a3af9368f0fa7a2c54573afc15a188cd984837"
+checksum = "6f8a848331614144c03e815fb931f238c9e342bd7ab501b29e35d647e7cd5688"
 dependencies = [
- "mutate_once",
+ "regex-syntax",
 ]
 
 [[package]]
-name = "khronos-egl"
-version = "6.0.0"
+name = "jsonschema-value"
+version = "0.60.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6aae1df220ece3c0ada96b8153459b67eebe9ae9212258bb0134ae60416fdf76"
+checksum = "5a5a947ef798eeacacb43edf29b1d0ab64bdf5e2c9fbcaad10ee8663d922b544"
 dependencies = [
- "libc",
- "libloading",
- "pkg-config",
+ "ahash",
+ "bytecount",
+ "fraction",
+ "getrandom 0.3.4",
+ "memchr",
+ "num-cmp",
+ "num-traits",
+ "serde_json",
+ "zmij",
 ]
 
 [[package]]
-name = "khronos_api"
-version = "3.1.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e2db585e1d738fc771bf08a151420d3ed193d9d895a36df7f6f8a9456b911ddc"
-
-[[package]]
-name = "kurbo"
-version = "0.10.4"
+name = "konst"
+version = "0.4.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1618d4ebd923e97d67e7cd363d80aef35fe961005cbbbb3d2dad8bdd1bc63440"
+checksum = "f660d5f887e3562f9ab6f4a14988795b694099d66b4f5dedc02d197ba9becb1d"
 dependencies = [
- "arrayvec",
- "smallvec",
+ "const_panic",
+ "konst_proc_macros",
+ "typewit",
 ]
 
 [[package]]
-name = "kurbo"
-version = "0.11.3"
+name = "konst_proc_macros"
+version = "0.4.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c62026ae44756f8a599ba21140f350303d4f08dcdcc71b5ad9c9bb8128c13c62"
-dependencies = [
- "arrayvec",
- "euclid",
- "smallvec",
-]
+checksum = "e037a2e1d8d5fdbd49b16a4ea09d5d6401c1f29eca5ff29d03d3824dba16256a"
 
 [[package]]
 name = "kurbo"
-version = "0.13.0"
+version = "0.13.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7564e90fe3c0d5771e1f0bc95322b21baaeaa0d9213fa6a0b61c99f8b17b3bfb"
+checksum = "4b60dfc32f652b926df6192e55525b16d186c69d47876c3ead4da5cc9f8450e2"
 dependencies = [
  "arrayvec",
  "euclid",
+ "polycool",
  "smallvec",
 ]
 
@@ -3637,15 +2781,9 @@ checksum = "d4345964bb142484797b161f473a503a434de77149dd8c7427788c6e13379388"
 
 [[package]]
 name = "lazy_static"
-version = "1.5.0"
+version = "1.5.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "bbd2bcb4c963f2ddae06a2efc7e9f3591312473c50c6685e1f298068316e66fe"
-
-[[package]]
-name = "leb128fmt"
-version = "0.1.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "09edd9e8b54e49e587e4f6295a7d29c3ea94d469cb40ab8ca70b288248a81db2"
+checksum = "20870f649af7073d53e38067b2a84312175d56ea15217e1b15bc83506ec50afb"
 
 [[package]]
 name = "lebe"
@@ -3655,15 +2793,15 @@ checksum = "7a79a3332a6609480d7d0c9eab957bca6b455b91bb84e66d19f5ff66294b85b8"
 
 [[package]]
 name = "libc"
-version = "0.2.184"
+version = "0.2.191"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "48f5d2a454e16a5ea0f4ced81bd44e4cfc7bd3a507b61887c99fd3538b28e4af"
+checksum = "65cd87c9a6d60536580b316ad985726104730a5151c99845f2490428ef91240a"
 
 [[package]]
 name = "libfuzzer-sys"
-version = "0.4.12"
+version = "0.4.13"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f12a681b7dd8ce12bff52488013ba614b869148d54dd79836ab85aafdd53f08d"
+checksum = "a9fd2f41a1cba099f79a0b6b6c35656cf7c03351a7bae8ff0f28f25270f929d2"
 dependencies = [
  "arbitrary",
  "cc",
@@ -3687,56 +2825,24 @@ checksum = "b6d2cec3eae94f9f509c767b45932f1ada8350c4bdb85af2fcab4a3c14807981"
 
 [[package]]
 name = "libredox"
-version = "0.1.15"
+version = "0.1.25"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7ddbf48fd451246b1f8c2610bd3b4ac0cc6e149d89832867093ab69a17194f08"
+checksum = "61ff90caf6077a803a240f62fdbe88645a890bbca49ef8174c3cb0404362171d"
 dependencies = [
- "bitflags 2.11.0",
  "libc",
- "plain",
- "redox_syscall 0.7.4",
-]
-
-[[package]]
-name = "lilt"
-version = "0.8.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f67562e5eff6b20553fa9be1c503356768420994e28f67e3eafe6f41910e57ad"
-dependencies = [
- "web-time",
-]
-
-[[package]]
-name = "linebender_resource_handle"
-version = "0.1.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d4a5ff6bcca6c4867b1c4fd4ef63e4db7436ef363e0ad7531d1558856bae64f4"
-
-[[package]]
-name = "linkme"
-version = "0.3.35"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5e3283ed2d0e50c06dd8602e0ab319bb048b6325d0bba739db64ed8205179898"
-dependencies = [
- "linkme-impl",
 ]
 
 [[package]]
-name = "linkme-impl"
-version = "0.3.35"
+name = "link-section"
+version = "0.19.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e5cec0ec4228b4853bb129c84dbf093a27e6c7a20526da046defc334a1b017f7"
-dependencies = [
- "proc-macro2",
- "quote",
- "syn",
-]
+checksum = "39c29a617ce3df32c08497bdc1ab6e2376e0b17948ac166a2fbe5977c5954cd9"
 
 [[package]]
-name = "linux-raw-sys"
-version = "0.4.15"
+name = "linktime-proc-macro"
+version = "0.2.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d26c52dbd32dccf2d10cac7725f8eae5296885fb5703b261f7d0a0739ec807ab"
+checksum = "7e57c38c1e860fd37c604281cdfb1dd2216977fd76a50f85ba2f388ef3219616"
 
 [[package]]
 name = "linux-raw-sys"
@@ -3746,21 +2852,15 @@ checksum = "32a66949e030da00e8c7d4434b251670a91556f4144941d37452769c25d58a53"
 
 [[package]]
 name = "litemap"
-version = "0.8.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "92daf443525c4cce67b150400bc2316076100ce0b3686209eb8cf3c31612e6f0"
-
-[[package]]
-name = "litrs"
-version = "1.0.0"
+version = "0.8.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "11d3d7f243d5c5a8b9bb5d6dd2b1602c0cb0b9db1621bafc7ed66e35ff9fe092"
+checksum = "47d9d19d1d6efa0109d2f65ff4c85cddd50bd572e5a00127ab10987290bcefae"
 
 [[package]]
 name = "llama-cpp-bindings"
-version = "0.16.0"
+version = "0.25.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b0cab532a93dac3f03c6e3923fd24d91434bdf4142dcec89af08842a5e791ad7"
+checksum = "8cba492d27e5561fe423205eb14ab51854151e48bf732bbc717f7986a5f64e56"
 dependencies = [
  "encoding_rs",
  "enumflags2",
@@ -3776,89 +2876,89 @@ dependencies = [
  "once_cell",
  "serde",
  "serde_json",
- "thiserror 2.0.18",
+ "thiserror",
  "toktrie",
 ]
 
 [[package]]
 name = "llama-cpp-bindings-build"
-version = "0.16.0"
+version = "0.25.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b281e0ae2b5ad140756ebe47c5e00e5e17f1a08e890cdb7601d856a2f2da59b4"
+checksum = "31cc4171a52ead66ab2480630dc2d589b6ca6aca1ed43949cda23abcd429943c"
 dependencies = [
  "bindgen",
  "cc",
  "cmake",
  "find_cuda_helper",
  "llama-cpp-wrapper-sources",
- "thiserror 2.0.18",
+ "thiserror",
 ]
 
 [[package]]
 name = "llama-cpp-bindings-sys"
-version = "0.16.0"
+version = "0.25.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7fe502d4a74ec5d93a2a78917c9c4223826e2b9466b2c5b869357e51fb7116a8"
+checksum = "7b45cb652a6c60f065b22e29be57a72c38b9d13247bb15c6edd150b18d4f9dc1"
 dependencies = [
  "llama-cpp-bindings-build",
 ]
 
 [[package]]
 name = "llama-cpp-bindings-types"
-version = "0.16.0"
+version = "0.25.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "002e2c1fc09748fb2b9914c660a8661368cfef18d1b205d2fd14f33a979e5b48"
+checksum = "f200360a50320173ea1eb8063cf51ef2ba5d0fee6168d528a6431ed5fc6e7bb2"
 dependencies = [
  "serde",
  "serde_json",
- "thiserror 2.0.18",
+ "thiserror",
 ]
 
 [[package]]
 name = "llama-cpp-error-recorder"
-version = "0.16.0"
+version = "0.25.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d181159c5b293f11fddb23d3fd012277110a8405e24a6c0a454329ce859627fe"
+checksum = "7de19c6c96c4d610b79fe1cfbbd40adff64e867cd8bf6bf6c69f75fcc4cc6d23"
 dependencies = [
  "log",
 ]
 
 [[package]]
 name = "llama-cpp-ffi-status"
-version = "0.16.0"
+version = "0.25.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7d1c9ab4956a5f38e11df94e3064a77af2dc3d66ec900720e9409d43435e315e"
+checksum = "942ab7989e058a65410f3cd332bed2a75c4c05f64b41392c1fb93d53f4c3e593"
 dependencies = [
  "llama-cpp-bindings-sys",
- "thiserror 2.0.18",
+ "thiserror",
 ]
 
 [[package]]
 name = "llama-cpp-gbnf"
-version = "0.16.0"
+version = "0.25.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "484337250ef83f370320f9a262cb98074fc7777545b13363dfa50925b518c785"
+checksum = "9456faa4fcb02bf221f176d088e2a0146cc9a56c35015cac89413a4f6056487b"
 dependencies = [
  "llama-cpp-bindings-sys",
  "llama-cpp-ffi-status",
- "thiserror 2.0.18",
+ "thiserror",
 ]
 
 [[package]]
 name = "llama-cpp-log-decoder"
-version = "0.16.0"
+version = "0.25.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "af82bb7e5f57250f5d4cf0b7ae93730f0f38676367360eb512994c1c823f8132"
+checksum = "a8129eec443ccdb124fc172cf33db7d0e87f975868493b8b0279b6d6407cc334"
 
 [[package]]
 name = "llama-cpp-wrapper-sources"
-version = "0.16.0"
+version = "0.25.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "4c490620f948a5f26cde3ce443c70fe9e91d04042bd5a195d0e6924340566130"
+checksum = "d7f93c691c97cef03d185f7f46815bbcb00f00c21bdcd5d16b62754f189e682c"
 dependencies = [
  "serde",
  "serde_json",
- "thiserror 2.0.18",
+ "thiserror",
 ]
 
 [[package]]
@@ -3904,6 +3004,15 @@ dependencies = [
  "scopeguard",
 ]
 
+[[package]]
+name = "lock_api"
+version = "0.5.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "65d91b22da179613dcf4d6ec8349563f52a363f6849e03b00dd0a185ec4a59af"
+dependencies = [
+ "scopeguard",
+]
+
 [[package]]
 name = "log"
 version = "0.4.29"
@@ -3919,35 +3028,28 @@ dependencies = [
  "imgref",
 ]
 
-[[package]]
-name = "lru"
-version = "0.16.4"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7f66e8d5d03f609abc3a39e6f08e4164ebf1447a732906d39eb9b99b7919ef39"
-
 [[package]]
 name = "lru-slab"
-version = "0.1.2"
+version = "0.1.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "112b39cec0b298b6c1999fee3e31427f74f676e4cb9879ed1a121b43661a4154"
+checksum = "4050469837a6ff301cd14c1f8f24f88549e6d548f24f64e2148eb0f72cebc51f"
 
 [[package]]
-name = "macro_registry"
-version = "0.6.9"
+name = "lz4_flex"
+version = "0.13.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "28c03fc749d06e1000766283015673e91aea121f30c60b7445681f2248e4994c"
+checksum = "7ef0d4ed8669f8f8826eb00dc878084aa8f253506c4fd5e8f58f5bce72ddb97e"
 dependencies = [
- "module_path_extractor",
- "syn",
+ "twox-hash",
 ]
 
 [[package]]
-name = "malloc_buf"
-version = "0.0.6"
+name = "matchers"
+version = "0.2.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "62bb907fe88d54d8d9ce32a3cceab4218ed2f6b7d35617cafe9adf84e43919cb"
+checksum = "d1525a2a28c7f4fa0fc98bb91ae755d1e2d1505079e05539e35bc876b5d65ae9"
 dependencies = [
- "libc",
+ "regex-automata",
 ]
 
 [[package]]
@@ -3962,25 +3064,19 @@ dependencies = [
 
 [[package]]
 name = "memchr"
-version = "2.8.0"
+version = "2.8.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f8ca58f447f06ed17d5fc4043ce1b10dd205e060fb3ce5b979b8ed8e59ff3f79"
+checksum = "cf8baf1c55e62ffcace7a9f06f4bd9cd3f0c4beb022d3b367256b91b87513d98"
 
 [[package]]
 name = "memmap2"
-version = "0.9.10"
+version = "0.9.12"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "714098028fe011992e1c3962653c96b2d578c4b4bce9036e15ff220319b1e0e3"
+checksum = "6b0ca1155a9ecc25aa12e8313a119f32edf0255fea2f44839c4ac0a12b2ac157"
 dependencies = [
  "libc",
 ]
 
-[[package]]
-name = "memo-map"
-version = "0.3.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "38d1115007560874e373613744c6fba374c17688327a71c1476d1a5954cc857b"
-
 [[package]]
 name = "memoffset"
 version = "0.9.1"
@@ -3991,19 +3087,10 @@ dependencies = [
 ]
 
 [[package]]
-name = "metal"
-version = "0.32.0"
+name = "micromap"
+version = "0.3.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "00c15a6f673ff72ddcc22394663290f870fb224c1bfce55734a75c414150e605"
-dependencies = [
- "bitflags 2.11.0",
- "block",
- "core-graphics-types 0.2.0",
- "foreign-types 0.5.0",
- "log",
- "objc",
- "paste",
-]
+checksum = "c2a86d3146ed3995b5913c414f6664344b9617457320782e64f0bb44afd49d74"
 
 [[package]]
 name = "mime"
@@ -4023,25 +3110,22 @@ dependencies = [
 
 [[package]]
 name = "minijinja"
-version = "2.19.0"
+version = "3.0.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "805bfd7352166bae857ee569628b52bcd85a1cecf7810861ebceb1686b72b75d"
+checksum = "90086695417a271bc96e6cb50618555614ca48fa3bb0d9aa96e3468b0f84431e"
 dependencies = [
- "memo-map",
  "serde",
- "serde_json",
 ]
 
 [[package]]
 name = "minijinja-contrib"
-version = "2.19.0"
+version = "3.0.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "45092d80391870622fcf3bd82f5d2af18f99533ea60debb4bc9db0c76f0e809a"
+checksum = "9c8e63b8c511fbb281cc19a1f731acd0903886ab6fbaf8399240ca68e5626568"
 dependencies = [
+ "jiff",
  "minijinja",
- "serde",
  "textwrap",
- "time",
  "unicode_categories",
 ]
 
@@ -4061,29 +3145,33 @@ dependencies = [
  "simd-adler32",
 ]
 
+[[package]]
+name = "miniz_oxide"
+version = "0.9.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b63fbc4a50860e98e7b2aa7804ded1db5cbc3aff9193adaff57a6931bf7c4b4c"
+dependencies = [
+ "adler2",
+ "simd-adler32",
+]
+
 [[package]]
 name = "mio"
-version = "1.2.0"
+version = "1.2.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "50b7e5b27aa02a74bac8c3f23f448f8d87ff11f92d3aac1a6ed369ee08cc56c1"
+checksum = "1788edb87fdc09c7e26304471e2f5be8cdefb1b6930d6e3985fc02ff53bf86ee"
 dependencies = [
  "libc",
  "log",
- "wasi",
+ "wasi 0.11.1+wasi-snapshot-preview1",
  "windows-sys 0.61.2",
 ]
 
 [[package]]
-name = "moddef"
-version = "0.2.6"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "4e519fd9c6131c1c9a4a67f8bdc4f32eb4105b16c1468adea1b8e68c98c85ec4"
-
-[[package]]
-name = "module_path_extractor"
-version = "0.6.9"
+name = "more-asserts"
+version = "0.3.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "41ec8ded8dfe4b676fb3dc4341830262395371d2113117bec13b40e830a1dea7"
+checksum = "1fafa6961cabd9c63bcd77a45d7e3b7f3b552b70417831fb0f56db717e72407e"
 
 [[package]]
 name = "moxcms"
@@ -4095,117 +3183,13 @@ dependencies = [
  "pxfm",
 ]
 
-[[package]]
-name = "mundy"
-version = "0.2.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "523813c9e194ec43693805214eb112551f99382115b67f38600d724a692e7e8b"
-dependencies = [
- "android-build",
- "async-io",
- "cfg-if",
- "dispatch",
- "futures-channel",
- "futures-lite",
- "jni 0.21.1",
- "ndk-context",
- "objc2 0.6.4",
- "objc2-app-kit 0.3.2",
- "objc2-foundation 0.3.2",
- "pin-project-lite",
- "wasm-bindgen",
- "wasm-bindgen-futures",
- "web-sys",
- "windows 0.62.2",
- "zbus",
-]
-
-[[package]]
-name = "mutate_once"
-version = "0.1.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "13d2233c9842d08cfe13f9eac96e207ca6a2ea10b80259ebe8ad0268be27d2af"
-
-[[package]]
-name = "naga"
-version = "27.0.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "066cf25f0e8b11ee0df221219010f213ad429855f57c494f995590c861a9a7d8"
-dependencies = [
- "arrayvec",
- "bit-set",
- "bitflags 2.11.0",
- "cfg-if",
- "cfg_aliases",
- "codespan-reporting",
- "half",
- "hashbrown 0.16.1",
- "hexf-parse",
- "indexmap",
- "libm",
- "log",
- "num-traits",
- "once_cell",
- "rustc-hash 1.1.0",
- "spirv",
- "thiserror 2.0.18",
- "unicode-ident",
-]
-
 [[package]]
 name = "nanoid"
-version = "0.4.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3ffa00dec017b5b1a8b7cf5e2c008bfda1aa7e0697ac1508b491fdf2622fb4d8"
-dependencies = [
- "rand 0.8.6",
-]
-
-[[package]]
-name = "native-tls"
-version = "0.2.18"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "465500e14ea162429d264d44189adc38b199b62b1c21eea9f69e4b73cb03bbf2"
-dependencies = [
- "libc",
- "log",
- "openssl",
- "openssl-probe",
- "openssl-sys",
- "schannel",
- "security-framework",
- "security-framework-sys",
- "tempfile",
-]
-
-[[package]]
-name = "ndk"
-version = "0.9.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c3f42e7bbe13d351b6bead8286a43aac9534b82bd3cc43e47037f012ebfd62d4"
-dependencies = [
- "bitflags 2.11.0",
- "jni-sys 0.3.1",
- "log",
- "ndk-sys",
- "num_enum",
- "raw-window-handle",
- "thiserror 1.0.69",
-]
-
-[[package]]
-name = "ndk-context"
-version = "0.1.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "27b02d87554356db9e9a873add8782d4ea6e3e58ea071a9adb9a2e8ddb884a8b"
-
-[[package]]
-name = "ndk-sys"
-version = "0.6.0+11769913"
+version = "0.5.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ee6cda3051665f1fb8d9e08fc35c96d5a244fb1be711a03b71118828afc9a873"
+checksum = "8628de41fe064cc3f0cf07f3d299ee3e73521adaff72278731d5c8cae3797873"
 dependencies = [
- "jni-sys 0.3.1",
+ "rand 0.9.5",
 ]
 
 [[package]]
@@ -4216,14 +3200,24 @@ checksum = "650eef8c711430f1a879fdd01d4745a7deea475becfb90269c06775983bbf086"
 
 [[package]]
 name = "nix"
-version = "0.30.1"
+version = "0.31.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "74523f3a35e05aba87a1d978330aef40f67b0304ac79c1c00b294c9830543db6"
+checksum = "cf20d2fde8ff38632c426f1165ed7436270b44f199fc55284c38276f9db47c3d"
 dependencies = [
- "bitflags 2.11.0",
+ "bitflags 2.13.2",
  "cfg-if",
  "cfg_aliases",
  "libc",
+ "memoffset",
+]
+
+[[package]]
+name = "no_std_io2"
+version = "0.9.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "418abd1b6d34fbf6cae440dc874771b0525a604428704c76e48b29a5e67b8003"
+dependencies = [
+ "memchr",
 ]
 
 [[package]]
@@ -4251,6 +3245,24 @@ version = "0.3.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "0676bb32a98c1a483ce53e500a81ad9c3d5b3f7c920c28c24e9cb0980d0b5bc8"
 
+[[package]]
+name = "ntapi"
+version = "0.4.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c3b335231dfd352ffb0f8017f3b6027a4917f7df785ea2143d8af2adc66980ae"
+dependencies = [
+ "winapi",
+]
+
+[[package]]
+name = "nu-ansi-term"
+version = "0.50.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7957b9740744892f114936ab4a57b3f487491bbeafaf8083688b16841a4240e5"
+dependencies = [
+ "windows-sys 0.61.2",
+]
+
 [[package]]
 name = "num"
 version = "0.4.3"
@@ -4267,9 +3279,9 @@ dependencies = [
 
 [[package]]
 name = "num-bigint"
-version = "0.4.6"
+version = "0.4.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a5e44f723f1133c9deac646763579fdb3ac745e418f2a7af9cd0c431da1f20b9"
+checksum = "c89e69e7e0f03bea5ef08013795c25018e101932225a656383bd384495ecc367"
 dependencies = [
  "num-integer",
  "num-traits",
@@ -4287,14 +3299,15 @@ version = "0.4.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "73f88a1307638156682bada9d7604135552957b7818057dcef22705b4d509495"
 dependencies = [
+ "bytemuck",
  "num-traits",
 ]
 
 [[package]]
 name = "num-conv"
-version = "0.2.1"
+version = "0.2.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c6673768db2d862beb9b39a78fdcb1a69439615d5794a1be50caa9bc92c81967"
+checksum = "521739c6d2bac4aa25192232afe6841231376b2b26d4d9fae5ecf8ca5772e441"
 
 [[package]]
 name = "num-derive"
@@ -4304,25 +3317,24 @@ checksum = "ed3955f1a9c7c0c15e092f9c887db08b1fc683305fdf6eb6684f22555355e202"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn",
+ "syn 2.0.119",
 ]
 
 [[package]]
 name = "num-integer"
-version = "0.1.46"
+version = "0.1.47"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7969661fd2958a5cb096e56c8e1ad0444ac2bbcd0061bd28660485a44879858f"
+checksum = "7ce2d95d4b3734dc35aa2f45e1aa22cd416814592a4f9d9205e11affd5b8e10b"
 dependencies = [
  "num-traits",
 ]
 
 [[package]]
 name = "num-iter"
-version = "0.1.45"
+version = "0.1.46"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1429034a0490724d0075ebb2bc9e875d6503c3cf69e235a8941aa757d83ef5bf"
+checksum = "c92800bd69a1eac91786bcfe9da64a897eb72911b8dc3095decbd07429e8048b"
 dependencies = [
- "autocfg",
  "num-integer",
  "num-traits",
 ]
@@ -4345,584 +3357,249 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "071dfc062690e90b734c0b2273ce72ad0ffa95f0c74596bc250dcfd960262841"
 dependencies = [
  "autocfg",
- "libm",
-]
-
-[[package]]
-name = "num_cpus"
-version = "1.17.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "91df4bbde75afed763b708b7eee1e8e7651e02d97f6d5dd763e89367e957b23b"
-dependencies = [
- "hermit-abi",
- "libc",
 ]
 
 [[package]]
-name = "num_enum"
-version = "0.7.6"
+name = "objc2"
+version = "0.6.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5d0bca838442ec211fa11de3a8b0e0e8f3a4522575b5c4c06ed722e005036f26"
+checksum = "08849bbd4767dfae9457696856ae1c84fe4e0281bbe4a7abff2d0e06fb7981f8"
 dependencies = [
- "num_enum_derive",
- "rustversion",
+ "objc2-encode",
 ]
 
 [[package]]
-name = "num_enum_derive"
-version = "0.7.6"
+name = "objc2-core-foundation"
+version = "0.3.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "680998035259dcfcafe653688bf2aa6d3e2dc05e98be6ab46afb089dc84f1df8"
+checksum = "2a180dd8642fa45cdb7dd721cd4c11b1cadd4929ce112ebd8b9f5803cc79d536"
 dependencies = [
- "proc-macro-crate",
- "proc-macro2",
- "quote",
- "syn",
+ "bitflags 2.13.2",
+ "dispatch2",
+ "objc2",
 ]
 
 [[package]]
-name = "number_prefix"
-version = "0.4.0"
+name = "objc2-encode"
+version = "4.1.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "830b246a0e5f20af87141b25c173cd1b609bd7779a4617d6ec582abaf90870f3"
+checksum = "ef25abbcd74fb2609453eb695bd2f860d389e457f67dc17cafc8b8cbc89d0c33"
 
 [[package]]
-name = "objc"
-version = "0.2.7"
+name = "objc2-foundation"
+version = "0.3.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "915b1b472bc21c53464d6c8461c9d3af805ba1ef837e1cac254428f4a77177b1"
+checksum = "e3e0adef53c21f888deb4fa59fc59f7eb17404926ee8a6f59f5df0fd7f9f3272"
 dependencies = [
- "malloc_buf",
+ "bitflags 2.13.2",
+ "objc2",
 ]
 
 [[package]]
-name = "objc-sys"
-version = "0.3.5"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "cdb91bdd390c7ce1a8607f35f3ca7151b65afc0ff5ff3b34fa350f7d7c7e4310"
-
-[[package]]
-name = "objc2"
-version = "0.5.2"
+name = "objc2-io-kit"
+version = "0.3.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "46a785d4eeff09c14c487497c162e92766fbb3e4059a71840cecc03d9a50b804"
+checksum = "33fafba39597d6dc1fb709123dfa8289d39406734be322956a69f0931c73bb15"
 dependencies = [
- "objc-sys",
- "objc2-encode",
+ "libc",
+ "objc2-core-foundation",
 ]
 
 [[package]]
-name = "objc2"
-version = "0.6.4"
+name = "objc2-open-directory"
+version = "0.3.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3a12a8ed07aefc768292f076dc3ac8c48f3781c8f2d5851dd3d98950e8c5a89f"
+checksum = "bb82bed227edf5201dfedf072bba4015a33d3d4a98519837295a90f0a23f676d"
 dependencies = [
- "objc2-encode",
+ "objc2",
+ "objc2-core-foundation",
+ "objc2-foundation",
 ]
 
 [[package]]
-name = "objc2-app-kit"
-version = "0.2.2"
+name = "objc2-system-configuration"
+version = "0.3.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e4e89ad9e3d7d297152b17d39ed92cd50ca8063a89a9fa569046d41568891eff"
+checksum = "7216bd11cbda54ccabcab84d523dc93b858ec75ecfb3a7d89513fa22464da396"
 dependencies = [
- "bitflags 2.11.0",
- "block2 0.5.1",
- "libc",
- "objc2 0.5.2",
- "objc2-core-data 0.2.2",
- "objc2-core-image 0.2.2",
- "objc2-foundation 0.2.2",
- "objc2-quartz-core 0.2.2",
+ "objc2-core-foundation",
 ]
 
 [[package]]
-name = "objc2-app-kit"
-version = "0.3.2"
+name = "once_cell"
+version = "1.21.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d49e936b501e5c5bf01fda3a9452ff86dc3ea98ad5f283e1455153142d97518c"
-dependencies = [
- "bitflags 2.11.0",
- "block2 0.6.2",
- "libc",
- "objc2 0.6.4",
- "objc2-cloud-kit 0.3.2",
- "objc2-core-data 0.3.2",
- "objc2-core-foundation",
- "objc2-core-graphics",
- "objc2-core-image 0.3.2",
- "objc2-core-text",
- "objc2-core-video",
- "objc2-foundation 0.3.2",
- "objc2-quartz-core 0.3.2",
-]
+checksum = "9f7c3e4beb33f85d45ae3e3a1792185706c8e16d043238c593331cc7cd313b50"
 
 [[package]]
-name = "objc2-cloud-kit"
-version = "0.2.2"
+name = "once_cell_polyfill"
+version = "1.70.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "74dd3b56391c7a0596a295029734d3c1c5e7e510a4cb30245f8221ccea96b009"
-dependencies = [
- "bitflags 2.11.0",
- "block2 0.5.1",
- "objc2 0.5.2",
- "objc2-core-location",
- "objc2-foundation 0.2.2",
-]
+checksum = "384b8ab6d37215f3c5301a95a4accb5d64aa607f1fcb26a11b5303878451b4fe"
 
 [[package]]
-name = "objc2-cloud-kit"
-version = "0.3.2"
+name = "oneshot"
+version = "0.1.13"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "73ad74d880bb43877038da939b7427bba67e9dd42004a18b809ba7d87cee241c"
-dependencies = [
- "bitflags 2.11.0",
- "objc2 0.6.4",
- "objc2-foundation 0.3.2",
-]
+checksum = "269bca4c2591a28585d6bf10d9ed0332b7d76900a1b02bec41bdc3a2cdcda107"
 
 [[package]]
-name = "objc2-contacts"
-version = "0.2.2"
+name = "openssl-probe"
+version = "0.2.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a5ff520e9c33812fd374d8deecef01d4a840e7b41862d849513de77e44aa4889"
-dependencies = [
- "block2 0.5.1",
- "objc2 0.5.2",
- "objc2-foundation 0.2.2",
-]
+checksum = "7c87def4c32ab89d880effc9e097653c8da5d6ef28e6b539d313baaacfbafcbe"
 
 [[package]]
-name = "objc2-core-data"
-version = "0.2.2"
+name = "option-ext"
+version = "0.2.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "617fbf49e071c178c0b24c080767db52958f716d9eabdf0890523aeae54773ef"
-dependencies = [
- "bitflags 2.11.0",
- "block2 0.5.1",
- "objc2 0.5.2",
- "objc2-foundation 0.2.2",
-]
+checksum = "04744f49eae99ab78e0d5c0b603ab218f515ea8cfe5a456d7629ad883a3b6e7d"
 
 [[package]]
-name = "objc2-core-data"
-version = "0.3.2"
+name = "os_str_bytes"
+version = "6.6.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0b402a653efbb5e82ce4df10683b6b28027616a2715e90009947d50b8dd298fa"
+checksum = "e2355d85b9a3786f481747ced0e0ff2ba35213a1f9bd406ed906554d7af805a1"
 dependencies = [
- "bitflags 2.11.0",
- "objc2 0.6.4",
- "objc2-foundation 0.3.2",
+ "memchr",
 ]
 
 [[package]]
-name = "objc2-core-foundation"
-version = "0.3.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2a180dd8642fa45cdb7dd721cd4c11b1cadd4929ce112ebd8b9f5803cc79d536"
-dependencies = [
- "bitflags 2.11.0",
- "dispatch2",
- "objc2 0.6.4",
-]
-
-[[package]]
-name = "objc2-core-graphics"
-version = "0.3.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e022c9d066895efa1345f8e33e584b9f958da2fd4cd116792e15e07e4720a807"
-dependencies = [
- "bitflags 2.11.0",
- "dispatch2",
- "objc2 0.6.4",
- "objc2-core-foundation",
- "objc2-io-surface",
-]
-
-[[package]]
-name = "objc2-core-image"
-version = "0.2.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "55260963a527c99f1819c4f8e3b47fe04f9650694ef348ffd2227e8196d34c80"
-dependencies = [
- "block2 0.5.1",
- "objc2 0.5.2",
- "objc2-foundation 0.2.2",
- "objc2-metal",
-]
-
-[[package]]
-name = "objc2-core-image"
-version = "0.3.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e5d563b38d2b97209f8e861173de434bd0214cf020e3423a52624cd1d989f006"
-dependencies = [
- "objc2 0.6.4",
- "objc2-foundation 0.3.2",
-]
-
-[[package]]
-name = "objc2-core-location"
-version = "0.2.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "000cfee34e683244f284252ee206a27953279d370e309649dc3ee317b37e5781"
-dependencies = [
- "block2 0.5.1",
- "objc2 0.5.2",
- "objc2-contacts",
- "objc2-foundation 0.2.2",
-]
-
-[[package]]
-name = "objc2-core-text"
-version = "0.3.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0cde0dfb48d25d2b4862161a4d5fcc0e3c24367869ad306b0c9ec0073bfed92d"
-dependencies = [
- "bitflags 2.11.0",
- "objc2 0.6.4",
- "objc2-core-foundation",
- "objc2-core-graphics",
-]
-
-[[package]]
-name = "objc2-core-video"
-version = "0.3.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d425caf1df73233f29fd8a5c3e5edbc30d2d4307870f802d18f00d83dc5141a6"
-dependencies = [
- "bitflags 2.11.0",
- "objc2 0.6.4",
- "objc2-core-foundation",
- "objc2-core-graphics",
- "objc2-io-surface",
-]
-
-[[package]]
-name = "objc2-encode"
-version = "4.1.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ef25abbcd74fb2609453eb695bd2f860d389e457f67dc17cafc8b8cbc89d0c33"
-
-[[package]]
-name = "objc2-foundation"
-version = "0.2.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0ee638a5da3799329310ad4cfa62fbf045d5f56e3ef5ba4149e7452dcf89d5a8"
-dependencies = [
- "bitflags 2.11.0",
- "block2 0.5.1",
- "dispatch",
- "libc",
- "objc2 0.5.2",
-]
-
-[[package]]
-name = "objc2-foundation"
-version = "0.3.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e3e0adef53c21f888deb4fa59fc59f7eb17404926ee8a6f59f5df0fd7f9f3272"
-dependencies = [
- "bitflags 2.11.0",
- "block2 0.6.2",
- "libc",
- "objc2 0.6.4",
- "objc2-core-foundation",
-]
-
-[[package]]
-name = "objc2-io-surface"
-version = "0.3.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "180788110936d59bab6bd83b6060ffdfffb3b922ba1396b312ae795e1de9d81d"
-dependencies = [
- "bitflags 2.11.0",
- "objc2 0.6.4",
- "objc2-core-foundation",
-]
-
-[[package]]
-name = "objc2-link-presentation"
-version = "0.2.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a1a1ae721c5e35be65f01a03b6d2ac13a54cb4fa70d8a5da293d7b0020261398"
-dependencies = [
- "block2 0.5.1",
- "objc2 0.5.2",
- "objc2-app-kit 0.2.2",
- "objc2-foundation 0.2.2",
-]
-
-[[package]]
-name = "objc2-metal"
-version = "0.2.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "dd0cba1276f6023976a406a14ffa85e1fdd19df6b0f737b063b95f6c8c7aadd6"
-dependencies = [
- "bitflags 2.11.0",
- "block2 0.5.1",
- "objc2 0.5.2",
- "objc2-foundation 0.2.2",
-]
-
-[[package]]
-name = "objc2-quartz-core"
-version = "0.2.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e42bee7bff906b14b167da2bac5efe6b6a07e6f7c0a21a7308d40c960242dc7a"
-dependencies = [
- "bitflags 2.11.0",
- "block2 0.5.1",
- "objc2 0.5.2",
- "objc2-foundation 0.2.2",
- "objc2-metal",
-]
-
-[[package]]
-name = "objc2-quartz-core"
-version = "0.3.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "96c1358452b371bf9f104e21ec536d37a650eb10f7ee379fff67d2e08d537f1f"
-dependencies = [
- "bitflags 2.11.0",
- "objc2 0.6.4",
- "objc2-core-foundation",
- "objc2-foundation 0.3.2",
-]
-
-[[package]]
-name = "objc2-symbols"
-version = "0.2.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0a684efe3dec1b305badae1a28f6555f6ddd3bb2c2267896782858d5a78404dc"
-dependencies = [
- "objc2 0.5.2",
- "objc2-foundation 0.2.2",
-]
-
-[[package]]
-name = "objc2-ui-kit"
-version = "0.2.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b8bb46798b20cd6b91cbd113524c490f1686f4c4e8f49502431415f3512e2b6f"
-dependencies = [
- "bitflags 2.11.0",
- "block2 0.5.1",
- "objc2 0.5.2",
- "objc2-cloud-kit 0.2.2",
- "objc2-core-data 0.2.2",
- "objc2-core-image 0.2.2",
- "objc2-core-location",
- "objc2-foundation 0.2.2",
- "objc2-link-presentation",
- "objc2-quartz-core 0.2.2",
- "objc2-symbols",
- "objc2-uniform-type-identifiers",
- "objc2-user-notifications",
-]
-
-[[package]]
-name = "objc2-uniform-type-identifiers"
-version = "0.2.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "44fa5f9748dbfe1ca6c0b79ad20725a11eca7c2218bceb4b005cb1be26273bfe"
-dependencies = [
- "block2 0.5.1",
- "objc2 0.5.2",
- "objc2-foundation 0.2.2",
-]
-
-[[package]]
-name = "objc2-user-notifications"
-version = "0.2.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "76cfcbf642358e8689af64cee815d139339f3ed8ad05103ed5eaf73db8d84cb3"
-dependencies = [
- "bitflags 2.11.0",
- "block2 0.5.1",
- "objc2 0.5.2",
- "objc2-core-location",
- "objc2-foundation 0.2.2",
-]
-
-[[package]]
-name = "once_cell"
-version = "1.21.4"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9f7c3e4beb33f85d45ae3e3a1792185706c8e16d043238c593331cc7cd313b50"
-
-[[package]]
-name = "once_cell_polyfill"
-version = "1.70.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "384b8ab6d37215f3c5301a95a4accb5d64aa607f1fcb26a11b5303878451b4fe"
-
-[[package]]
-name = "open"
-version = "5.3.4"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9f3bab717c29a857abf75fcef718d441ec7cb2725f937343c734740a985d37fd"
-dependencies = [
- "is-wsl",
- "libc",
- "pathdiff",
-]
-
-[[package]]
-name = "openssl"
-version = "0.10.81"
+name = "outref"
+version = "0.5.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "77823a27f0babb03091cb9ed9ef80af3b39dbc82f97e8fa530374b7dafd87a45"
-dependencies = [
- "bitflags 2.11.0",
- "cfg-if",
- "foreign-types 0.3.2",
- "libc",
- "openssl-macros",
- "openssl-sys",
-]
+checksum = "1a80800c0488c3a21695ea981a54918fbb37abf04f4d0720c453632255e2ff0e"
 
 [[package]]
-name = "openssl-macros"
-version = "0.1.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a948666b637a0f465e8564c73e89d4dde00d72d4d473cc972f390fc3dcee7d9c"
+name = "paddler_agent"
+version = "5.0.0"
 dependencies = [
- "proc-macro2",
- "quote",
- "syn",
+ "anyhow",
+ "async-trait",
+ "futures-util",
+ "http 1.5.0",
+ "log",
+ "nix",
+ "paddler_agent_decision",
+ "paddler_agent_embeddings",
+ "paddler_agent_runtime",
+ "paddler_agent_status",
+ "paddler_agent_text_generation",
+ "paddler_blocking_thread_fixture",
+ "paddler_cache_dir",
+ "paddler_download_manager",
+ "paddler_frame_writer",
+ "paddler_inference_parameters",
+ "paddler_local_http_fixture",
+ "paddler_messaging",
+ "paddler_model_source",
+ "paddler_outbox",
+ "paddler_request_registry",
+ "paddler_spawn_resuming_panic",
+ "serde",
+ "serde_json",
+ "tempfile",
+ "thiserror",
+ "tokio",
+ "tokio-tungstenite",
+ "tokio-util",
+ "trzcina",
 ]
 
 [[package]]
-name = "openssl-probe"
-version = "0.2.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7c87def4c32ab89d880effc9e097653c8da5d6ef28e6b539d313baaacfbafcbe"
-
-[[package]]
-name = "openssl-sys"
-version = "0.9.117"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b47e7e6bb2c38cd930d25a23b40fa52e068c10e85f3e03a7f5ba5aaca5713695"
+name = "paddler_agent_decision"
+version = "5.0.0"
 dependencies = [
- "cc",
- "libc",
- "pkg-config",
- "vcpkg",
+ "llama-cpp-bindings",
+ "log",
+ "paddler_agent_pointer_head",
+ "paddler_agent_runtime",
+ "paddler_agent_status",
+ "paddler_inference_parameters",
+ "paddler_messaging",
+ "thiserror",
+ "tokio",
+ "tokio-util",
 ]
 
 [[package]]
-name = "option-ext"
-version = "0.2.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "04744f49eae99ab78e0d5c0b603ab218f515ea8cfe5a456d7629ad883a3b6e7d"
-
-[[package]]
-name = "orbclient"
-version = "0.3.51"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "59aed3b33578edcfa1bc96a321d590d31832b6ad55a26f0313362ce687e9abd6"
+name = "paddler_agent_embeddings"
+version = "5.0.0"
 dependencies = [
- "libc",
- "libredox",
+ "llama-cpp-bindings",
+ "log",
+ "paddler_agent_runtime",
+ "paddler_agent_status",
+ "paddler_inference_parameters",
+ "paddler_messaging",
+ "thiserror",
+ "tokio",
+ "tokio-util",
 ]
 
 [[package]]
-name = "ordered-float"
-version = "5.3.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b7d950ca161dc355eaf28f82b11345ed76c6e1f6eb1f4f4479e0323b9e2fbd0e"
+name = "paddler_agent_pointer_head"
+version = "5.0.0"
 dependencies = [
- "num-traits",
+ "icu_normalizer",
+ "llama-cpp-bindings",
+ "thiserror",
 ]
 
 [[package]]
-name = "ordered-stream"
-version = "0.2.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9aa2b01e1d916879f73a53d01d1d6cee68adbb31d6d9177a8cfce093cced1d50"
+name = "paddler_agent_runtime"
+version = "5.0.0"
 dependencies = [
- "futures-core",
- "pin-project-lite",
+ "llama-cpp-bindings",
+ "llama-cpp-bindings-sys",
+ "log",
+ "paddler_agent_status",
+ "paddler_inference_parameters",
+ "paddler_messaging",
+ "paddler_spawn_resuming_panic",
+ "parking_lot 0.13.0",
+ "thiserror",
+ "tokio",
+ "tokio-util",
 ]
 
 [[package]]
-name = "outref"
-version = "0.5.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1a80800c0488c3a21695ea981a54918fbb37abf04f4d0720c453632255e2ff0e"
-
-[[package]]
-name = "owned_ttf_parser"
-version = "0.25.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "36820e9051aca1014ddc75770aab4d68bc1e9e632f0f5627c4086bc216fb583b"
+name = "paddler_agent_status"
+version = "5.0.0"
 dependencies = [
- "ttf-parser",
+ "dashmap",
+ "paddler_messaging",
+ "parking_lot 0.13.0",
+ "tokio",
+ "tokio-test",
 ]
 
 [[package]]
-name = "paddler_agent"
-version = "4.1.0"
+name = "paddler_agent_text_generation"
+version = "5.0.0"
 dependencies = [
- "anyhow",
- "async-stream",
- "async-trait",
- "bytes 1.11.1",
- "dashmap",
  "encoding_rs",
- "futures 0.3.32",
- "futures-util",
- "hf-hub",
- "indoc",
  "llama-cpp-bindings",
- "llama-cpp-bindings-sys",
- "llama-cpp-bindings-types",
  "log",
  "minijinja",
  "minijinja-contrib",
- "nanoid",
+ "paddler_agent_runtime",
  "paddler_agent_status",
- "paddler_cache_dir",
- "paddler_download_manager",
  "paddler_image_decoder",
  "paddler_inference_parameters",
- "paddler_local_http_fixture",
  "paddler_messaging",
- "paddler_model_source",
- "paddler_request_registry",
+ "paddler_sha256_hex",
  "paddler_tool_call_validator",
- "parking_lot",
- "rand 0.9.4",
- "reqwest 0.12.28",
- "serde",
+ "rand 0.10.3",
  "serde_json",
- "shellexpand",
- "tempfile",
- "thiserror 2.0.18",
+ "thiserror",
  "tokio",
- "tokio-stream",
- "tokio-test",
- "tokio-tungstenite",
  "tokio-util",
- "trzcina",
- "url",
-]
-
-[[package]]
-name = "paddler_agent_status"
-version = "4.1.0"
-dependencies = [
- "dashmap",
- "paddler_messaging",
- "parking_lot",
- "tokio",
- "tokio-test",
 ]
 
 [[package]]
 name = "paddler_balancer"
-version = "4.1.0"
+version = "5.0.0"
 dependencies = [
  "actix-cors",
  "actix-web",
@@ -4930,80 +3607,92 @@ dependencies = [
  "actix-ws",
  "anyhow",
  "askama",
- "async-stream",
  "async-trait",
- "bytes 1.11.1",
+ "bytes",
  "cadence",
  "dashmap",
- "esbuild-metafile",
- "futures 0.3.32",
+ "futures",
  "futures-util",
  "indoc",
  "llama-cpp-bindings-types",
  "log",
  "mime_guess",
  "nanoid",
+ "paddler_blocking_thread_fixture",
+ "paddler_dashboard_assets",
+ "paddler_frame_writer",
  "paddler_inference_parameters",
  "paddler_messaging",
  "paddler_openai_response_format_validator",
+ "paddler_openai_translation",
+ "paddler_outbox",
  "paddler_request_registry",
+ "paddler_spawn_resuming_panic",
  "paddler_state_database",
- "parking_lot",
+ "paddler_typesafe_translation",
+ "parking_lot 0.13.0",
+ "rand 0.10.3",
  "rust-embed",
  "serde",
  "serde_json",
  "socket2",
  "tempfile",
- "thiserror 2.0.18",
  "tokio",
  "tokio-stream",
- "tokio-test",
  "tokio-util",
  "trzcina",
 ]
 
+[[package]]
+name = "paddler_blocking_thread_fixture"
+version = "5.0.0"
+dependencies = [
+ "tokio",
+]
+
 [[package]]
 name = "paddler_bootstrap"
-version = "4.1.0"
+version = "5.0.0"
 dependencies = [
  "actix-web",
  "anyhow",
- "async-trait",
  "log",
  "nanoid",
  "paddler_agent",
- "paddler_agent_status",
+ "paddler_agent_runtime",
  "paddler_balancer",
- "paddler_inference_parameters",
+ "paddler_client",
  "paddler_messaging",
+ "paddler_model_source",
+ "paddler_outbox",
  "paddler_state_database",
  "paddler_test_cluster_harness",
- "reqwest 0.12.28",
+ "reqwest",
  "serde_json",
  "tempfile",
- "thiserror 2.0.18",
+ "thiserror",
  "tokio",
  "tokio-util",
  "trzcina",
+ "url",
 ]
 
 [[package]]
 name = "paddler_cache_dir"
-version = "4.1.0"
+version = "5.0.0"
 dependencies = [
  "fslock",
- "sha2",
+ "paddler_sha256_hex",
  "tempfile",
- "thiserror 2.0.18",
+ "thiserror",
  "tokio",
  "tokio-util",
 ]
 
 [[package]]
 name = "paddler_cli"
-version = "4.1.0"
+version = "5.0.0"
 dependencies = [
- "actix-rt",
  "actix-web",
  "anyhow",
  "async-trait",
@@ -5012,6 +3701,7 @@ dependencies = [
  "env_logger",
  "paddler_balancer",
  "paddler_bootstrap",
+ "paddler_model_source",
  "paddler_state_database",
  "tokio-util",
  "trzcina",
@@ -5019,43 +3709,52 @@ dependencies = [
 
 [[package]]
 name = "paddler_cli_tests"
-version = "4.1.0"
+version = "5.0.0"
 dependencies = [
+ "actix-web",
  "anyhow",
  "async-trait",
  "futures-util",
- "log",
+ "libc",
+ "llama-cpp-bindings",
  "nix",
+ "paddler_agent_text_generation",
  "paddler_balancer",
+ "paddler_bootstrap",
  "paddler_cli",
  "paddler_client",
  "paddler_inference_parameters",
  "paddler_messaging",
+ "paddler_openai_response_format_validator",
  "paddler_test_cluster_harness",
- "reqwest 0.12.28",
+ "paddler_tests",
+ "reqwest",
+ "seccompiler",
  "serde_json",
- "thiserror 2.0.18",
+ "tempfile",
+ "thiserror",
  "tokio",
+ "tokio-tungstenite",
  "tokio-util",
+ "trzcina",
  "url",
 ]
 
 [[package]]
 name = "paddler_client"
-version = "4.1.0"
+version = "5.0.0"
 dependencies = [
- "bytes 1.11.1",
+ "bytes",
  "futures-util",
- "http 1.4.0",
+ "http 1.5.0",
  "log",
  "nanoid",
- "paddler_local_http_fixture",
  "paddler_messaging",
- "parking_lot",
- "reqwest 0.12.28",
+ "parking_lot 0.13.0",
+ "reqwest",
  "serde",
  "serde_json",
- "thiserror 2.0.18",
+ "thiserror",
  "tokio",
  "tokio-stream",
  "tokio-tungstenite",
@@ -5063,118 +3762,92 @@ dependencies = [
  "url",
 ]
 
+[[package]]
+name = "paddler_dashboard_assets"
+version = "5.0.0"
+dependencies = [
+ "esbuild-metafile",
+ "tempfile",
+ "thiserror",
+]
+
 [[package]]
 name = "paddler_download_manager"
-version = "4.1.0"
+version = "5.0.0"
 dependencies = [
- "bytes 1.11.1",
+ "bytes",
  "futures-util",
  "headers",
- "http 1.4.0",
+ "http 1.5.0",
  "paddler_local_http_fixture",
- "reqwest 0.12.28",
+ "reqwest",
  "tempfile",
- "thiserror 2.0.18",
+ "thiserror",
  "tokio",
  "tokio-util",
  "url",
 ]
 
 [[package]]
-name = "paddler_gui"
-version = "4.1.0"
-dependencies = [
- "async-stream",
- "clap",
- "command_handler",
- "env_logger",
- "iced",
- "if-addrs",
- "log",
- "open",
- "paddler_balancer",
- "paddler_bootstrap",
- "paddler_inference_parameters",
- "paddler_messaging",
- "paddler_state_database",
- "paddler_test_cluster_harness",
- "statum",
- "thiserror 2.0.18",
- "tokio",
- "tokio-util",
- "trzcina",
-]
-
-[[package]]
-name = "paddler_gui_tests"
-version = "4.1.0"
+name = "paddler_frame_writer"
+version = "5.0.0"
 dependencies = [
- "futures 0.3.32",
- "iced_test",
- "paddler_balancer",
- "paddler_bootstrap",
- "paddler_gui",
- "paddler_messaging",
- "paddler_test_cluster_harness",
- "reqwest 0.12.28",
- "thiserror 2.0.18",
+ "futures",
+ "futures-util",
  "tokio",
  "tokio-util",
- "trzcina",
 ]
 
 [[package]]
 name = "paddler_image_decoder"
-version = "4.1.0"
+version = "5.0.0"
 dependencies = [
- "base64",
+ "base64 0.23.1",
  "image",
  "log",
  "paddler_messaging",
- "resvg 0.46.0",
- "thiserror 2.0.18",
+ "resvg",
+ "thiserror",
 ]
 
 [[package]]
 name = "paddler_inference_parameters"
-version = "4.1.0"
+version = "5.0.0"
 dependencies = [
+ "llama-cpp-bindings-types",
  "serde",
  "serde_json",
- "thiserror 2.0.18",
+ "thiserror",
 ]
 
 [[package]]
 name = "paddler_local_http_fixture"
-version = "4.1.0"
+version = "5.0.0"
 dependencies = [
- "http 1.4.0",
+ "http 1.5.0",
  "httparse",
+ "nix",
+ "paddler_tcp_buffer_limits",
  "tokio",
 ]
 
 [[package]]
 name = "paddler_messaging"
-version = "4.1.0"
+version = "5.0.0"
 dependencies = [
- "base64",
- "encoding_rs",
  "llama-cpp-bindings-types",
- "log",
- "nanoid",
  "paddler_inference_parameters",
  "serde",
  "serde_json",
- "thiserror 2.0.18",
+ "thiserror",
  "tokio",
  "url",
 ]
 
 [[package]]
 name = "paddler_model_source"
-version = "4.1.0"
+version = "5.0.0"
 dependencies = [
- "async-trait",
  "hf-hub",
  "log",
  "paddler_agent_status",
@@ -5182,9 +3855,10 @@ dependencies = [
  "paddler_download_manager",
  "paddler_local_http_fixture",
  "paddler_messaging",
- "reqwest 0.12.28",
+ "percent-encoding",
+ "reqwest",
  "tempfile",
- "thiserror 2.0.18",
+ "thiserror",
  "tokio",
  "tokio-util",
  "url",
@@ -5192,74 +3866,118 @@ dependencies = [
 
 [[package]]
 name = "paddler_openai_response_format_validator"
-version = "4.1.0"
+version = "5.0.0"
 dependencies = [
  "jsonschema",
  "serde_json",
- "thiserror 2.0.18",
+ "thiserror",
  "yaml-rust2",
 ]
 
+[[package]]
+name = "paddler_openai_translation"
+version = "5.0.0"
+dependencies = [
+ "llama-cpp-bindings-types",
+ "nanoid",
+ "paddler_inference_parameters",
+ "paddler_messaging",
+ "paddler_openai_response_format_validator",
+ "serde",
+ "serde_json",
+ "thiserror",
+]
+
 [[package]]
 name = "paddler_opencode_tests"
-version = "4.1.0"
+version = "5.0.0"
 dependencies = [
- "anyhow",
- "paddler_test_cluster_harness",
  "paddler_tests",
  "serde_json",
  "tempfile",
- "thiserror 2.0.18",
+ "thiserror",
  "tokio",
  "url",
 ]
 
+[[package]]
+name = "paddler_outbox"
+version = "5.0.0"
+dependencies = [
+ "parking_lot 0.13.0",
+ "tokio",
+]
+
 [[package]]
 name = "paddler_request_registry"
-version = "4.1.0"
+version = "5.0.0"
 dependencies = [
  "dashmap",
+ "nanoid",
+ "tokio",
+]
+
+[[package]]
+name = "paddler_sha256_hex"
+version = "5.0.0"
+dependencies = [
+ "sha2",
+]
+
+[[package]]
+name = "paddler_spawn_resuming_panic"
+version = "5.0.0"
+dependencies = [
  "tokio",
 ]
 
 [[package]]
 name = "paddler_state_database"
-version = "4.1.0"
+version = "5.0.0"
 dependencies = [
- "async-trait",
- "log",
- "paddler_inference_parameters",
  "paddler_messaging",
- "parking_lot",
+ "paddler_spawn_resuming_panic",
+ "parking_lot 0.13.0",
  "serde",
  "serde_json",
  "shellexpand",
  "tempfile",
- "thiserror 2.0.18",
+ "thiserror",
  "tokio",
  "url",
 ]
 
+[[package]]
+name = "paddler_tcp_buffer_limits"
+version = "5.0.0"
+dependencies = [
+ "tempfile",
+ "thiserror",
+]
+
 [[package]]
 name = "paddler_test_cluster_harness"
-version = "4.1.0"
+version = "5.0.0"
 dependencies = [
  "anyhow",
  "async-openai",
  "async-trait",
- "base64",
+ "base64 0.23.1",
  "data-url",
+ "futures",
  "futures-util",
  "hf-hub",
+ "paddler_agent_status",
  "paddler_balancer",
  "paddler_client",
  "paddler_inference_parameters",
  "paddler_messaging",
+ "paddler_model_source",
  "quick-xml",
- "serde",
+ "reqwest",
  "serde_json",
  "tempfile",
- "thiserror 2.0.18",
+ "thiserror",
  "tokio",
  "tokio-tungstenite",
  "tokio-util",
@@ -5269,49 +3987,72 @@ dependencies = [
 
 [[package]]
 name = "paddler_tests"
-version = "4.1.0"
+version = "5.0.0"
 dependencies = [
  "anyhow",
  "async-openai",
  "async-trait",
+ "clap",
  "futures-util",
- "http 1.4.0",
+ "http 1.5.0",
+ "llama-cpp-bindings",
+ "llama-cpp-bindings-sys",
  "llama-cpp-bindings-types",
  "log",
+ "paddler_agent",
+ "paddler_agent_decision",
+ "paddler_agent_embeddings",
+ "paddler_agent_pointer_head",
+ "paddler_agent_runtime",
+ "paddler_agent_status",
+ "paddler_agent_text_generation",
  "paddler_balancer",
+ "paddler_blocking_thread_fixture",
  "paddler_bootstrap",
+ "paddler_cache_dir",
  "paddler_client",
+ "paddler_download_manager",
  "paddler_inference_parameters",
  "paddler_local_http_fixture",
  "paddler_messaging",
+ "paddler_model_source",
  "paddler_openai_response_format_validator",
+ "paddler_openai_translation",
  "paddler_state_database",
+ "paddler_tcp_buffer_limits",
  "paddler_test_cluster_harness",
- "reqwest 0.12.28",
+ "paddler_tool_call_validator",
+ "paddler_typesafe_translation",
+ "reqwest",
+ "serde",
  "serde_json",
+ "syn 3.0.9",
  "tempfile",
  "tokio",
  "tokio-tungstenite",
  "tokio-util",
- "trzcina",
 ]
 
 [[package]]
 name = "paddler_tool_call_validator"
-version = "4.1.0"
+version = "5.0.0"
 dependencies = [
  "jsonschema",
  "llama-cpp-bindings-types",
  "paddler_messaging",
  "serde_json",
- "thiserror 2.0.18",
+ "thiserror",
 ]
 
 [[package]]
-name = "parking"
-version = "2.2.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f38d5652c16fde515bb1ecef450ab0f6a219d619a7274976324d5e377f7dceba"
+name = "paddler_typesafe_translation"
+version = "5.0.0"
+dependencies = [
+ "paddler_messaging",
+ "serde",
+ "serde_json",
+ "thiserror",
+]
 
 [[package]]
 name = "parking_lot"
@@ -5319,8 +4060,18 @@ version = "0.12.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "93857453250e3077bd71ff98b6a65ea6621a19bb0f559a85248955ac12c45a1a"
 dependencies = [
- "lock_api",
- "parking_lot_core",
+ "lock_api 0.4.14",
+ "parking_lot_core 0.9.12",
+]
+
+[[package]]
+name = "parking_lot"
+version = "0.13.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a9cf937ca4e1d0b7578784559063422fc39ae7c5a00f02471f2614d5486fd226"
+dependencies = [
+ "lock_api 0.5.0",
+ "parking_lot_core 0.10.0",
 ]
 
 [[package]]
@@ -5331,7 +4082,19 @@ checksum = "2621685985a2ebf1c516881c026032ac7deafcda1a2c9b7850dc81e3dfcb64c1"
 dependencies = [
  "cfg-if",
  "libc",
- "redox_syscall 0.5.18",
+ "redox_syscall 0.5.18",
+ "smallvec",
+ "windows-link",
+]
+
+[[package]]
+name = "parking_lot_core"
+version = "0.10.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9cfff060068cc18aa7815010fcaf80c68c7fb9e85fc7f5c0f9bf142c21d19826"
+dependencies = [
+ "libc",
+ "redox_syscall 0.9.5",
  "smallvec",
  "windows-link",
 ]
@@ -5368,22 +4131,22 @@ checksum = "5be167a7af36ee22fe3115051bc51f6e6c7054c9348e28deb4f49bd6f705a315"
 
 [[package]]
 name = "pin-project"
-version = "1.1.11"
+version = "1.1.13"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f1749c7ed4bcaf4c3d0a3efc28538844fb29bcdd7d2b67b2be7e20ba861ff517"
+checksum = "2466b2336ed02bcdca6b294417127b90ec92038d1d5c4fbeac971a922e0e0924"
 dependencies = [
  "pin-project-internal",
 ]
 
 [[package]]
 name = "pin-project-internal"
-version = "1.1.11"
+version = "1.1.13"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d9b20ed30f105399776b9c883e68e536ef602a16ae6f596d2c473591d6ad64c6"
+checksum = "c96395f0a926bc13b1c17622aaddda1ecb55d49c8f1bf9777e4d877800a43f8b"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn",
+ "syn 2.0.119",
 ]
 
 [[package]]
@@ -5392,47 +4155,11 @@ version = "0.2.17"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "a89322df9ebe1c1578d689c92318e070967d1042b512afbe49518723f4e6d5cd"
 
-[[package]]
-name = "pin-utils"
-version = "0.1.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8b870d8c151b6f2fb93e84a13146138f05d02ed11c7e7c54f8826aaaf7c9f184"
-
-[[package]]
-name = "piper"
-version = "0.2.5"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c835479a4443ded371d6c535cbfd8d31ad92c5d23ae9770a61bc155e4992a3c1"
-dependencies = [
- "atomic-waker",
- "fastrand",
- "futures-io",
-]
-
 [[package]]
 name = "pkg-config"
-version = "0.3.32"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7edddbd0b52d732b21ad9a5fab5c704c14cd949e5e9a1ec5929a24fded1b904c"
-
-[[package]]
-name = "plain"
-version = "0.2.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b4596b6d070b27117e987119b4dac604f3c58cfb0b191112e24771b2faeac1a6"
-
-[[package]]
-name = "png"
-version = "0.17.16"
+version = "0.3.34"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "82151a2fc869e011c153adc57cf2789ccb8d9906ce52c0b39a6b5697749d7526"
-dependencies = [
- "bitflags 1.3.2",
- "crc32fast",
- "fdeflate",
- "flate2",
- "miniz_oxide",
-]
+checksum = "f6b464fbc74e149a392436b17d523f769e057cb6877f6a5c4618bc6f11800548"
 
 [[package]]
 name = "png"
@@ -5440,56 +4167,51 @@ version = "0.18.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "60769b8b31b2a9f263dae2776c37b1b28ae246943cf719eb6946a1db05128a61"
 dependencies = [
- "bitflags 2.11.0",
+ "bitflags 2.13.2",
  "crc32fast",
  "fdeflate",
  "flate2",
- "miniz_oxide",
+ "miniz_oxide 0.8.9",
 ]
 
 [[package]]
-name = "polling"
-version = "3.11.0"
+name = "polycool"
+version = "0.4.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5d0e4f59085d47d8241c88ead0f274e8a0cb551f3625263c05eb8dd897c34218"
+checksum = "50596ddc09eb5ad5f75cacd40209568e66df71baf86e1499a0e99c4cff12a5a6"
 dependencies = [
- "cfg-if",
- "concurrent-queue",
- "hermit-abi",
- "pin-project-lite",
- "rustix 1.1.4",
- "windows-sys 0.61.2",
+ "arrayvec",
 ]
 
 [[package]]
 name = "portable-atomic"
-version = "1.13.1"
+version = "1.15.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c33a9471896f1c69cecef8d20cbe2f7accd12527ce60845ff44c153bb2a21b49"
+checksum = "05c8b63e8d9609db387f0324918f81d68fe27748f084ef092fb35954d0539a85"
 
 [[package]]
 name = "portable-atomic-util"
-version = "0.2.6"
+version = "0.2.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "091397be61a01d4be58e7841595bd4bfedb15f1cd54977d79b8271e94ed799a3"
+checksum = "10ab3eb7f3becc3a1cbc4f2c6f20267996cfc1a6467a873763411b136a122715"
 dependencies = [
  "portable-atomic",
 ]
 
 [[package]]
 name = "potential_utf"
-version = "0.1.5"
+version = "0.1.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0103b1cef7ec0cf76490e969665504990193874ea05c85ff9bab8b911d0a0564"
+checksum = "d83eb9bc6d8e5cf568e7a1101d60ee05e81ed50ea106026f3d18deeb046d7661"
 dependencies = [
  "zerovec",
 ]
 
 [[package]]
 name = "powerfmt"
-version = "0.2.0"
+version = "0.2.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "439ee305def115ba05938db6eb1644ff94165c5ab5e9420d1c1bcedbba909391"
+checksum = "4a6394b9e965e73d0a289ee54f589087e2c676aedf60885baf52c76b771e4958"
 
 [[package]]
 name = "ppv-lite86"
@@ -5500,22 +4222,6 @@ dependencies = [
  "zerocopy",
 ]
 
-[[package]]
-name = "presser"
-version = "0.3.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e8cf8e6a8aa66ce33f63993ffc4ea4271eb5b0530a9002db8455ea6050c77bfa"
-
-[[package]]
-name = "pretty_assertions"
-version = "1.4.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3ae130e2f271fbc2ac3a40fb1d07180839cdbbe443c7a27e1e3c13c5cac0116d"
-dependencies = [
- "diff",
- "yansi",
-]
-
 [[package]]
 name = "prettyplease"
 version = "0.2.37"
@@ -5523,51 +4229,75 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "479ca8adacdd7ce8f1fb39ce9ecccbfe93a3f1344b3d0d97f20bc0196208f62b"
 dependencies = [
  "proc-macro2",
- "syn",
+ "syn 2.0.119",
 ]
 
 [[package]]
-name = "proc-macro-crate"
-version = "3.5.0"
+name = "prettyplease"
+version = "0.3.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e67ba7e9b2b56446f1d419b1d807906278ffa1a658a8a5d8a39dcb1f5a78614f"
+checksum = "2bfe0f4c752e450fc2faf62654f1c134747922825d5b04ca717b8874f41a40c0"
 dependencies = [
- "toml_edit",
+ "proc-macro2",
+ "syn 3.0.9",
 ]
 
 [[package]]
 name = "proc-macro2"
-version = "1.0.106"
+version = "1.0.107"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8fd00f0bb2e90d81d1044c2b32617f68fcb9fa3bb7640c23e9c748e53fb30934"
+checksum = "985e7ec9bb745e6ce6535b544d84d6cd6f7ad8bd711c398938ae983b91a766d9"
 dependencies = [
  "unicode-ident",
 ]
 
 [[package]]
 name = "profiling"
-version = "1.0.17"
+version = "1.0.18"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3eb8486b569e12e2c32ad3e204dbaba5e4b5b216e9367044f25f1dba42341773"
+checksum = "3d595e54a326bc53c1c197b32d295e14b169e3cfeaa8dc82b529f947fba6bcf5"
 dependencies = [
  "profiling-procmacros",
 ]
 
 [[package]]
 name = "profiling-procmacros"
-version = "1.0.17"
+version = "1.0.18"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "52717f9a02b6965224f95ca2a81e2e0c5c43baacd28ca057577988930b6c3d5b"
+checksum = "4488a4a36b9a4ba6b9334a32a39971f77c1436ec82c38707bce707699cc3bbcb"
 dependencies = [
  "quote",
- "syn",
+ "syn 2.0.119",
+]
+
+[[package]]
+name = "pulp"
+version = "0.22.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "046aa45b989642ec2e4717c8e72d677b13edd831a4d3b6cf37d9a3e54912496a"
+dependencies = [
+ "bytemuck",
+ "cfg-if",
+ "libm",
+ "num-complex",
+ "paste",
+ "pulp-wasm-simd-flag",
+ "raw-cpuid",
+ "reborrow",
+ "version_check",
 ]
 
+[[package]]
+name = "pulp-wasm-simd-flag"
+version = "0.1.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1d8f70e07b9c3962945a74e59ca1c511bba65b6419468acc217c457d93f3c740"
+
 [[package]]
 name = "pxfm"
-version = "0.1.28"
+version = "0.1.30"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b5a041e753da8b807c9255f28de81879c78c876392ff2469cde94799b2896b9d"
+checksum = "d55d956fa96f5ec02be2e13af0e20391a5aa83d6a074e3ad368959d0fab299ea"
 
 [[package]]
 name = "qoi"
@@ -5586,28 +4316,28 @@ checksum = "a993555f31e5a609f617c12db6250dedcac1b0a85076912c436e6fc9b2c8e6a3"
 
 [[package]]
 name = "quick-xml"
-version = "0.39.2"
+version = "0.42.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "958f21e8e7ceb5a1aa7fa87fab28e7c75976e0bfe7e23ff069e0a260f894067d"
+checksum = "41b1177fdf999d2321d3fb46ff47159d9c1fb9ad66a4879f8c50a0b504615e9b"
 dependencies = [
  "memchr",
 ]
 
 [[package]]
 name = "quinn"
-version = "0.11.9"
+version = "0.11.12"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b9e20a958963c291dc322d98411f541009df2ced7b5a4f2bd52337638cfccf20"
+checksum = "4051e23e9185c255a7e33ef59cdbca87a22d359052eecd22fc6b901fb37d9d11"
 dependencies = [
- "bytes 1.11.1",
+ "bytes",
  "cfg_aliases",
  "pin-project-lite",
  "quinn-proto",
  "quinn-udp",
- "rustc-hash 2.1.2",
+ "rustc-hash",
  "rustls",
  "socket2",
- "thiserror 2.0.18",
+ "thiserror",
  "tokio",
  "tracing",
  "web-time",
@@ -5615,22 +4345,22 @@ dependencies = [
 
 [[package]]
 name = "quinn-proto"
-version = "0.11.18"
+version = "0.11.19"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a9746dbde176634f4f2f1faf2404e30a31b2bc1e9cafb5329c95d8177a18c9fc"
+checksum = "0e750cca55fe4f0439a15d0bb529da9651e79993e8e72c61a899a36d462befbe"
 dependencies = [
  "aws-lc-rs",
- "bytes 1.11.1",
- "getrandom 0.4.2",
+ "bytes",
+ "getrandom 0.4.3",
  "lru-slab",
- "rand 0.10.1",
+ "rand 0.10.3",
  "rand_pcg",
  "ring",
- "rustc-hash 2.1.2",
+ "rustc-hash",
  "rustls",
  "rustls-pki-types",
  "slab",
- "thiserror 2.0.18",
+ "thiserror",
  "tinyvec",
  "tracing",
  "web-time",
@@ -5638,23 +4368,23 @@ dependencies = [
 
 [[package]]
 name = "quinn-udp"
-version = "0.5.14"
+version = "0.5.16"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "addec6a0dcad8a8d96a771f815f0eaf55f9d1805756410b39f5fa81332574cbd"
+checksum = "af66907df18639dcf4db56ca65490cabc4b27a97dbadd96f2926cca73298f016"
 dependencies = [
  "cfg_aliases",
  "libc",
  "once_cell",
  "socket2",
  "tracing",
- "windows-sys 0.60.2",
+ "windows-sys 0.61.2",
 ]
 
 [[package]]
 name = "quote"
-version = "1.0.45"
+version = "1.0.47"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "41f2619966050689382d2b44f664f4bc593e129785a36d6ee376ddf37259b924"
+checksum = "1fbf4db142a473a8d80c26bbf18454ed458bf8d26c8219c331daecfdbd079001"
 dependencies = [
  "proc-macro2",
 ]
@@ -5673,46 +4403,25 @@ checksum = "f8dcc9c7d52a811697d2151c701e0d08956f92b0e24136cf4cf27b57a6a0d9bf"
 
 [[package]]
 name = "rand"
-version = "0.8.6"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5ca0ecfa931c29007047d1bc58e623ab12e5590e8c7cc53200d5202b69266d8a"
-dependencies = [
- "libc",
- "rand_chacha 0.3.1",
- "rand_core 0.6.4",
-]
-
-[[package]]
-name = "rand"
-version = "0.9.4"
+version = "0.9.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "44c5af06bb1b7d3216d91932aed5265164bf384dc89cd6ba05cf59a35f5f76ea"
+checksum = "b9ef1d0d795eb7d84685bca4f72f3649f064e6641543d3a8c415898726a57b41"
 dependencies = [
- "rand_chacha 0.9.0",
+ "rand_chacha",
  "rand_core 0.9.5",
 ]
 
 [[package]]
 name = "rand"
-version = "0.10.1"
+version = "0.10.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d2e8e8bcc7961af1fdac401278c6a831614941f6164ee3bf4ce61b7edb162207"
+checksum = "65c9fb96cbc91e3478eaae79a69fcd3f1ae4ad052e471fe6732fff548984b4af"
 dependencies = [
  "chacha20",
- "getrandom 0.4.2",
+ "getrandom 0.4.3",
  "rand_core 0.10.1",
 ]
 
-[[package]]
-name = "rand_chacha"
-version = "0.3.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e6c10a63a0fa32252be49d21e7709d4d4baf8d231c2dbce1eaa8141b9b127d88"
-dependencies = [
- "ppv-lite86",
- "rand_core 0.6.4",
-]
-
 [[package]]
 name = "rand_chacha"
 version = "0.9.0"
@@ -5723,15 +4432,6 @@ dependencies = [
  "rand_core 0.9.5",
 ]
 
-[[package]]
-name = "rand_core"
-version = "0.6.4"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ec0be4795e2f6a28069bec0b5ff3e2ac9bafc99e6a9a7dc3547996c5c816922c"
-dependencies = [
- "getrandom 0.2.17",
-]
-
 [[package]]
 name = "rand_core"
 version = "0.9.5"
@@ -5756,18 +4456,6 @@ dependencies = [
  "rand_core 0.10.1",
 ]
 
-[[package]]
-name = "range-alloc"
-version = "0.1.5"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ca45419789ae5a7899559e9512e58ca889e41f04f1f2445e9f4b290ceccd1d08"
-
-[[package]]
-name = "rangemap"
-version = "1.7.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "973443cf09a9c8656b574a866ab68dfa19f0867d0340648c7d2f6a71b8a8ea68"
-
 [[package]]
 name = "rav1e"
 version = "0.8.1"
@@ -5795,10 +4483,10 @@ dependencies = [
  "num-traits",
  "paste",
  "profiling",
- "rand 0.9.4",
- "rand_chacha 0.9.0",
+ "rand 0.9.5",
+ "rand_chacha",
  "simd_helpers",
- "thiserror 2.0.18",
+ "thiserror",
  "v_frame",
  "wasm-bindgen",
 ]
@@ -5819,16 +4507,19 @@ dependencies = [
 ]
 
 [[package]]
-name = "raw-window-handle"
-version = "0.6.2"
+name = "raw-cpuid"
+version = "11.6.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "20675572f6f24e9e76ef639bc5552774ed45f1c30e2951e1e99c59888861c539"
+checksum = "498cd0dc59d73224351ee52a95fee0f1a617a2eae0e7d9d720cc622c73a54186"
+dependencies = [
+ "bitflags 2.13.2",
+]
 
 [[package]]
 name = "rayon"
-version = "1.11.0"
+version = "1.12.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "368f01d005bf8fd9b1206fb6fa653e6c4a81ceb1466406b81792d87c5677a58f"
+checksum = "fb39b166781f92d482534ef4b4b1b2568f42613b53e5b6c160e24cfbfa30926d"
 dependencies = [
  "either",
  "rayon-core",
@@ -5846,32 +4537,28 @@ dependencies = [
 
 [[package]]
 name = "read-fonts"
-version = "0.35.0"
+version = "0.41.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6717cf23b488adf64b9d711329542ba34de147df262370221940dfabc2c91358"
+checksum = "046a7d674daf459825b32f5062056d6882db0d2f5a479fbd76ccfc870ac18709"
 dependencies = [
  "bytemuck",
- "core_maths",
- "font-types 0.10.1",
+ "font-types",
+ "once_cell",
 ]
 
 [[package]]
-name = "read-fonts"
-version = "0.37.0"
+name = "reborrow"
+version = "0.5.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7b634fabf032fab15307ffd272149b622260f55974d9fad689292a5d33df02e5"
-dependencies = [
- "bytemuck",
- "font-types 0.11.3",
-]
+checksum = "03251193000f4bd3b042892be858ee50e8b3719f2b08e5833ac4353724632430"
 
 [[package]]
-name = "redox_syscall"
-version = "0.4.1"
+name = "redb"
+version = "3.1.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "4722d768eff46b75989dd134e5c353f0d6296e5aaa3132e776cbdb56be7731aa"
+checksum = "4ba239c1c1693315d3cc0e601db3b3965543afbf48c41730fdca2f069f510f4a"
 dependencies = [
- "bitflags 1.3.2",
+ "libc",
 ]
 
 [[package]]
@@ -5880,47 +4567,46 @@ version = "0.5.18"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "ed2bf2547551a7053d6fdfafda3f938979645c44812fbfcda098faae3f1a362d"
 dependencies = [
- "bitflags 2.11.0",
+ "bitflags 2.13.2",
 ]
 
 [[package]]
 name = "redox_syscall"
-version = "0.7.4"
+version = "0.9.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f450ad9c3b1da563fb6948a8e0fb0fb9269711c9c73d9ea1de5058c79c8d643a"
+checksum = "580c875845b41ecb1564ad1ae4703044cef774145e3d7a94dc2b3fa7c97868fd"
 dependencies = [
- "bitflags 2.11.0",
+ "bitflags 2.13.2",
 ]
 
 [[package]]
 name = "redox_users"
-version = "0.5.2"
+version = "0.5.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a4e608c6638b9c18977b00b475ac1f28d14e84b27d8d42f70e0bf1e3dec127ac"
+checksum = "60dc65c0ff1a7ae1294b0c67b9f14baf70b644404010370171787bfac1038fc0"
 dependencies = [
- "getrandom 0.2.17",
  "libredox",
- "thiserror 2.0.18",
+ "thiserror",
 ]
 
 [[package]]
 name = "ref-cast"
-version = "1.0.25"
+version = "1.0.27"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f354300ae66f76f1c85c5f84693f0ce81d747e2c3f21a45fef496d89c960bf7d"
+checksum = "7e440fb4e4b4147295338efb76001ab9e4efc0e5839df2c47fc5ac2381d365c3"
 dependencies = [
  "ref-cast-impl",
 ]
 
 [[package]]
 name = "ref-cast-impl"
-version = "1.0.25"
+version = "1.0.27"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b7186006dcb21920990093f30e3dea63b7d6e977bf1256be20c3563a5db070da"
+checksum = "92ecd8964f8453721699a1ed72037b0db49ce2f5a5138486ee89bed6f67cdf3a"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn",
+ "syn 3.0.9",
 ]
 
 [[package]]
@@ -5932,31 +4618,33 @@ dependencies = [
  "ahash",
  "fluent-uri 0.3.2",
  "once_cell",
- "parking_lot",
+ "parking_lot 0.12.5",
  "percent-encoding",
  "serde_json",
 ]
 
 [[package]]
 name = "referencing"
-version = "0.37.4"
+version = "0.60.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "4283168a506f0dcbdce31c9f9cce3129c924da4c6bca46e46707fcb746d2d70c"
+checksum = "922268812b6f9e20b4ab0f1efe8e07036958eed7346ceddb31c777b5531fd11f"
 dependencies = [
  "ahash",
  "fluent-uri 0.4.1",
  "getrandom 0.3.4",
- "hashbrown 0.16.1",
- "parking_lot",
+ "hashbrown 0.17.1",
+ "itoa",
+ "micromap",
+ "parking_lot 0.12.5",
  "percent-encoding",
  "serde_json",
 ]
 
 [[package]]
 name = "regex"
-version = "1.12.3"
+version = "1.13.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e10754a14b9137dd7b1e3e5b0493cc9171fdd105e0ab477f51b72e7f3ac0e276"
+checksum = "f020237b6c8eed93db2e2cb53c00c60a8e1bc73da7d073199a1180401450218d"
 dependencies = [
  "aho-corasick",
  "memchr",
@@ -5966,9 +4654,9 @@ dependencies = [
 
 [[package]]
 name = "regex-automata"
-version = "0.4.14"
+version = "0.4.18"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6e1dd4122fc1595e8162618945476892eefca7b88c52820e74af6262213cae8f"
+checksum = "ad8553b9b26413251cbf30e620595c7a41b3887f03da04579c0e6b0d6a06b4b2"
 dependencies = [
  "aho-corasick",
  "memchr",
@@ -5983,77 +4671,31 @@ checksum = "cab834c73d247e67f4fae452806d17d3c7501756d98c8808d7c9c7aa7d18f973"
 
 [[package]]
 name = "regex-syntax"
-version = "0.8.10"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "dc897dd8d9e8bd1ed8cdad82b5966c3e0ecae09fb1907d58efaa013543185d0a"
-
-[[package]]
-name = "renderdoc-sys"
-version = "1.1.0"
+version = "0.8.11"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "19b30a45b0cd0bcca8037f3d0dc3421eaf95327a17cad11964fb8179b4fc4832"
+checksum = "d6f6ff9a378485b298a5286656da665ba74413d36db0979633275d2e708145d4"
 
 [[package]]
 name = "reqwest"
-version = "0.12.28"
+version = "0.13.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "eddd3ca559203180a307f12d114c268abf583f59b03cb906fd0b3ff8646c1147"
+checksum = "16a1cfa75cc186dd73d5818e510e042e40927bccc9c236b061cea97e1eb08029"
 dependencies = [
- "base64",
- "bytes 1.11.1",
+ "base64 0.23.1",
+ "bytes",
  "encoding_rs",
  "futures-core",
  "futures-util",
- "h2 0.4.19",
- "http 1.4.0",
+ "h2",
+ "http 1.5.0",
  "http-body",
  "http-body-util",
  "hyper",
  "hyper-rustls",
- "hyper-tls",
  "hyper-util",
  "js-sys",
  "log",
  "mime",
- "native-tls",
- "percent-encoding",
- "pin-project-lite",
- "rustls-pki-types",
- "serde",
- "serde_json",
- "serde_urlencoded",
- "sync_wrapper",
- "tokio",
- "tokio-native-tls",
- "tokio-util",
- "tower",
- "tower-http",
- "tower-service",
- "url",
- "wasm-bindgen",
- "wasm-bindgen-futures",
- "wasm-streams 0.4.2",
- "web-sys",
-]
-
-[[package]]
-name = "reqwest"
-version = "0.13.4"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "219c5811de6525e5416c7d5d53bb656d3afdbc6c5af816e0802bcfa42dbdc1c3"
-dependencies = [
- "base64",
- "bytes 1.11.1",
- "futures-core",
- "futures-util",
- "http 1.4.0",
- "http-body",
- "http-body-util",
- "hyper",
- "hyper-rustls",
- "hyper-util",
- "js-sys",
- "log",
  "mime_guess",
  "percent-encoding",
  "pin-project-lite",
@@ -6074,42 +4716,40 @@ dependencies = [
  "url",
  "wasm-bindgen",
  "wasm-bindgen-futures",
- "wasm-streams 0.5.0",
+ "wasm-streams",
  "web-sys",
 ]
 
 [[package]]
-name = "resvg"
-version = "0.45.1"
+name = "reqwest-middleware"
+version = "0.5.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a8928798c0a55e03c9ca6c4c6846f76377427d2c1e1f7e6de3c06ae57942df43"
+checksum = "07bc3f1384cffa4f274dad2d4ddd73aed32fed8f786d96c6be8aa4e5fd3c3b58"
 dependencies = [
- "gif 0.13.3",
- "image-webp",
- "log",
- "pico-args",
- "rgb",
- "svgtypes 0.15.3",
- "tiny-skia",
- "usvg 0.45.1",
- "zune-jpeg 0.4.21",
+ "anyhow",
+ "async-trait",
+ "http 1.5.0",
+ "reqwest",
+ "thiserror",
+ "tower-service",
 ]
 
 [[package]]
 name = "resvg"
-version = "0.46.0"
+version = "0.49.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b563218631706d614e23059436526d005b50ab5f2d506b55a17eb65c5eb83419"
+checksum = "1572844b6727a1bf842190e9947641a68ac44bed83f50afb4bd1407a2564bee8"
 dependencies = [
- "gif 0.14.1",
+ "bytemuck",
+ "gif",
  "image-webp",
  "log",
  "pico-args",
  "rgb",
- "svgtypes 0.16.1",
+ "svgtypes",
  "tiny-skia",
- "usvg 0.46.0",
- "zune-jpeg 0.5.15",
+ "usvg",
+ "zune-jpeg",
 ]
 
 [[package]]
@@ -6143,18 +4783,36 @@ checksum = "6c20b6793b5c2fa6553b250154b78d6d0db37e72700ae35fad9387a46f487c97"
 
 [[package]]
 name = "roxmltree"
-version = "0.21.1"
+version = "0.21.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9da7f59328b09405b181fee1018925fad50f157afb43c0c0f5c522c256dd5907"
+dependencies = [
+ "memchr",
+]
+
+[[package]]
+name = "russe"
+version = "0.3.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f1964b10c76125c36f8afe190065a4bf9a87bf324842c05701330bba9f1cacbb"
+checksum = "56b278184c4cf1f499ad5020876c9d31961b6166526fc7415ad11f113166f91b"
 dependencies = [
+ "aho-corasick",
+ "bytes",
+ "bytestring",
+ "futures-util",
+ "impl-more",
+ "itoa",
  "memchr",
+ "tokio",
+ "tokio-util",
+ "tracing",
 ]
 
 [[package]]
 name = "rust-embed"
-version = "8.11.0"
+version = "8.13.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "04113cb9355a377d83f06ef1f0a45b8ab8cd7d8b1288160717d66df5c7988d27"
+checksum = "19afa5b4b6a611de00bd1bdae6ae6f39084c9399f0679c3f52d8469cf335cc23"
 dependencies = [
  "rust-embed-impl",
  "rust-embed-utils",
@@ -6163,23 +4821,24 @@ dependencies = [
 
 [[package]]
 name = "rust-embed-impl"
-version = "8.11.0"
+version = "8.13.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "da0902e4c7c8e997159ab384e6d0fc91c221375f6894346ae107f47dd0f3ccaa"
+checksum = "e0d8afda6374eac59e066abee06d265247ebbaf3006cf878e2879e8356e34053"
 dependencies = [
+ "mime_guess",
  "proc-macro2",
  "quote",
  "rust-embed-utils",
  "shellexpand",
- "syn",
+ "syn 2.0.119",
  "walkdir",
 ]
 
 [[package]]
 name = "rust-embed-utils"
-version = "8.11.0"
+version = "8.13.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5bcdef0be6fe7f6fa333b1073c949729274b05f123a0ad7efcb8efd878e5c3b1"
+checksum = "0d84e8ba78bd384263e5922f084cbe1b081c3b7e69add59c8fb097b879ba968a"
 dependencies = [
  "sha2",
  "walkdir",
@@ -6187,15 +4846,9 @@ dependencies = [
 
 [[package]]
 name = "rustc-hash"
-version = "1.1.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "08d43f7aa6b08d49f382cde6a7982047c3426db949b1424bc4b7ec9ae12c6ce2"
-
-[[package]]
-name = "rustc-hash"
-version = "2.1.2"
+version = "2.1.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "94300abf3f1ae2e2b8ffb7b58043de3d399c73fa6f4b73826402a5c457614dbe"
+checksum = "6b1e7f9a428571be2dc5bc0505c13fb6bf936822b894ec87abf8a08a4e51742d"
 
 [[package]]
 name = "rustc_version"
@@ -6208,40 +4861,25 @@ dependencies = [
 
 [[package]]
 name = "rustix"
-version = "0.38.44"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "fdb5bc1ae2baa591800df16c9ca78619bf65c0488b41b96ccec5d11220d8c154"
-dependencies = [
- "bitflags 2.11.0",
- "errno",
- "libc",
- "linux-raw-sys 0.4.15",
- "windows-sys 0.59.0",
-]
-
-[[package]]
-name = "rustix"
-version = "1.1.4"
+version = "1.1.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b6fe4565b9518b83ef4f91bb47ce29620ca828bd32cb7e408f0062e9930ba190"
+checksum = "891efababe418670775f199f0d233d84843c227a0949a883ce15b37c78d6629d"
 dependencies = [
- "bitflags 2.11.0",
+ "bitflags 2.13.2",
  "errno",
  "libc",
- "linux-raw-sys 0.12.1",
+ "linux-raw-sys",
  "windows-sys 0.61.2",
 ]
 
 [[package]]
 name = "rustls"
-version = "0.23.45"
+version = "0.23.46"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0d41d731c7d2f962d1ccc364cec258de3c0e93b38c2fb3ba97ac74513048d634"
+checksum = "48e13bd8c0e9365c43cfa5c9e8f9ad49d3c8444926c9aac819e0e4dc503c8fdf"
 dependencies = [
  "aws-lc-rs",
- "log",
  "once_cell",
- "ring",
  "rustls-pki-types",
  "rustls-webpki",
  "subtle",
@@ -6250,9 +4888,9 @@ dependencies = [
 
 [[package]]
 name = "rustls-native-certs"
-version = "0.8.4"
+version = "0.8.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "dab5152771c58876a2146916e53e35057e1a4dfa2b9df0f0305b07f611fdea4d"
+checksum = "4a073f5dc7469f984c52ad2752b63b0807745133b6de880b7b64c1ac4c48aec4"
 dependencies = [
  "openssl-probe",
  "rustls-pki-types",
@@ -6262,9 +4900,9 @@ dependencies = [
 
 [[package]]
 name = "rustls-pki-types"
-version = "1.14.0"
+version = "1.15.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "be040f8b0a225e40375822a563fa9524378b9d63112f53e19ffff34df5d33fdd"
+checksum = "2f4925028c7eb5d1fcdaf196971378ed9d2c1c4efc7dc5d011256f76c99c0a96"
 dependencies = [
  "web-time",
  "zeroize",
@@ -6272,13 +4910,13 @@ dependencies = [
 
 [[package]]
 name = "rustls-platform-verifier"
-version = "0.7.0"
+version = "0.7.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "26d1e2536ce4f35f4846aa13bff16bd0ff40157cdb14cc056c7b14ba41233ba0"
+checksum = "1167586491e2b18b8bfbb293e8180ec17c201c4f076d7cb3070ca964e7598f98"
 dependencies = [
  "core-foundation 0.10.1",
  "core-foundation-sys",
- "jni 0.22.4",
+ "jni",
  "log",
  "once_cell",
  "rustls",
@@ -6293,15 +4931,15 @@ dependencies = [
 
 [[package]]
 name = "rustls-platform-verifier-android"
-version = "0.1.1"
+version = "0.2.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f87165f0995f63a9fbeea62b64d10b4d9d8e78ec6d7d51fb2125fda7bb36788f"
+checksum = "eec689c0bc40ff2458a5977b6619cb718087084a18e02a131c599b62d05e1a5f"
 
 [[package]]
 name = "rustls-webpki"
-version = "0.103.15"
+version = "0.103.16"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f3c3cf1d8b1e7d4927e2d154c3fcb02979afb9939629c62cd9048d4f07b60ac2"
+checksum = "4443b1b6d39d7c663139942cc4558262562d2c84123bb6282ef177009ca8b281"
 dependencies = [
  "aws-lc-rs",
  "ring",
@@ -6311,27 +4949,9 @@ dependencies = [
 
 [[package]]
 name = "rustversion"
-version = "1.0.22"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b39cdef0fa800fc44525c84ccb54a029961a8215f9619753635a9c0d2538d46d"
-
-[[package]]
-name = "rustybuzz"
-version = "0.20.1"
+version = "1.0.23"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "fd3c7c96f8a08ee34eff8857b11b49b07d71d1c3f4e88f8a88d4c9e9f90b1702"
-dependencies = [
- "bitflags 2.11.0",
- "bytemuck",
- "core_maths",
- "log",
- "smallvec",
- "ttf-parser",
- "unicode-bidi-mirroring",
- "unicode-ccc",
- "unicode-properties",
- "unicode-script",
-]
+checksum = "cf54715a573b99ac80df0bc206da022bcd442c974952c7b9720069370852e21f"
 
 [[package]]
 name = "ryu"
@@ -6339,6 +4959,12 @@ version = "1.0.23"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "9774ba4a74de5f7b1c1451ed6cd5285a32eddb5cccb8cc655a4e50009e06477f"
 
+[[package]]
+name = "safe-transmute"
+version = "0.11.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3944826ff8fa8093089aba3acb4ef44b9446a99a16f3bf4e74af3f77d340ab7d"
+
 [[package]]
 name = "same-file"
 version = "1.0.6"
@@ -6350,19 +4976,13 @@ dependencies = [
 
 [[package]]
 name = "schannel"
-version = "0.1.29"
+version = "0.1.30"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "91c1b7e4904c873ef0710c1f407dde2e6287de2bebc1bbbf7d430bb7cbffd939"
+checksum = "9cb23c424d8ee841bb4d813677ca237e2a610067d594fec130f87d6093495ab5"
 dependencies = [
  "windows-sys 0.61.2",
 ]
 
-[[package]]
-name = "scoped-tls"
-version = "1.0.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e1cf6437eb19a8f4a6cc0f7dca544973b0b78843adbfeb3683d1a94a0024a294"
-
 [[package]]
 name = "scopeguard"
 version = "1.2.0"
@@ -6370,16 +4990,12 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "94143f37725109f92c262ed2cf5e59bce7498c01bcc1502d7b9afe439a4e9f49"
 
 [[package]]
-name = "sctk-adwaita"
-version = "0.10.1"
+name = "seccompiler"
+version = "0.5.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b6277f0217056f77f1d8f49f2950ac6c278c0d607c45f5ee99328d792ede24ec"
+checksum = "a4ae55de56877481d112a559bbc12667635fdaf5e005712fd4e2b2fa50ffc884"
 dependencies = [
- "ab_glyph",
- "log",
- "memmap2",
- "smithay-client-toolkit 0.19.2",
- "tiny-skia",
+ "libc",
 ]
 
 [[package]]
@@ -6398,7 +5014,7 @@ version = "3.7.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "b7f4bc775c73d9a02cde8bf7b2ec4c9d12743edf609006c7facc23998404cd1d"
 dependencies = [
- "bitflags 2.11.0",
+ "bitflags 2.13.2",
  "core-foundation 0.10.1",
  "core-foundation-sys",
  "libc",
@@ -6415,17 +5031,11 @@ dependencies = [
  "libc",
 ]
 
-[[package]]
-name = "self_cell"
-version = "1.2.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b12e76d157a900eb52e81bc6e9f3069344290341720e9178cde2407113ac8d89"
-
 [[package]]
 name = "semver"
-version = "1.0.27"
+version = "1.0.28"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d767eb0aabc880b29956c35734170f26ed551a859dbd361d140cdbeca61ab1e2"
+checksum = "8a7852d02fc848982e0c167ef163aaff9cd91dc640ba85e263cb1ce46fae51cd"
 
 [[package]]
 name = "serde"
@@ -6454,20 +5064,20 @@ checksum = "d540f220d3187173da220f885ab66608367b6574e925011a9353e4badda91d79"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn",
+ "syn 2.0.119",
 ]
 
 [[package]]
 name = "serde_html_form"
-version = "0.4.0"
+version = "0.4.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0946d52b4b7e28823148aebbeceb901012c595ad737920d504fa8634bb099e6f"
+checksum = "8f0346d7a342ab90f405cfc08f25d15075f944f42fcabbc5eac923829fa6d228"
 dependencies = [
  "form_urlencoded",
  "indexmap",
  "itoa",
- "ryu",
  "serde_core",
+ "zmij",
 ]
 
 [[package]]
@@ -6497,13 +5107,13 @@ dependencies = [
 
 [[package]]
 name = "serde_repr"
-version = "0.1.20"
+version = "0.1.21"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "175ee3e80ae9982737ca543e96133087cbd9a485eecc3bc4de9c1a37b47ea59c"
+checksum = "8d3b1629de253c70a0508c3899572da79ca359fdab27c7920ff00406df418906"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn",
+ "syn 3.0.9",
 ]
 
 [[package]]
@@ -6520,9 +5130,9 @@ dependencies = [
 
 [[package]]
 name = "sha1"
-version = "0.10.6"
+version = "0.10.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e3bf829a2d51ab4a5ddf1352d8470c140cadc8301b2ae1789db023f01cedd6ba"
+checksum = "a978451301f4db1d02937a4ab3ccce137717b81826e79b7d49ffe3244a13c3b8"
 dependencies = [
  "cfg-if",
  "cpufeatures 0.2.17",
@@ -6536,19 +5146,28 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "aacc4cc499359472b4abe1bf11d0b12e688af9a805fa5e3016f9a386dc2d0214"
 dependencies = [
  "cfg-if",
- "cpufeatures 0.3.0",
- "digest 0.11.2",
+ "cpufeatures 0.3.1",
+ "digest 0.11.3",
 ]
 
 [[package]]
 name = "sha2"
-version = "0.10.9"
+version = "0.11.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a7507d819769d01a365ab707794a4084392c824f54a7a6a7862f8c3d0892b283"
+checksum = "47d7069beb7d6ac7b9acd1039986e73443f24234f41074da099d6f994ac9ad19"
 dependencies = [
  "cfg-if",
- "cpufeatures 0.2.17",
- "digest 0.10.7",
+ "cpufeatures 0.3.1",
+ "digest 0.11.3",
+]
+
+[[package]]
+name = "sharded-slab"
+version = "0.1.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f40ca3c46823713e0d4209592e8d6e826aa57e928f09752619fc696c499637f6"
+dependencies = [
+ "lazy_static",
 ]
 
 [[package]]
@@ -6557,7 +5176,9 @@ version = "3.1.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "32824fab5e16e6c4d86dc1ba84489390419a39f97699852b66480bb87d297ed8"
 dependencies = [
+ "bstr",
  "dirs",
+ "os_str_bytes",
 ]
 
 [[package]]
@@ -6578,15 +5199,15 @@ dependencies = [
 
 [[package]]
 name = "simd-adler32"
-version = "0.3.9"
+version = "0.3.10"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "703d5c7ef118737c72f1af64ad2f6f8c5e1921f818cdcb97b8fe6fc69bf66214"
+checksum = "3a219298ac11a56ea9a6d2120044824d6f01aeb034955e7af7bc16858527deea"
 
 [[package]]
 name = "simd_cesu8"
-version = "1.1.1"
+version = "1.2.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "94f90157bb87cddf702797c5dadfa0be7d266cdf49e22da2fcaa32eff75b2c33"
+checksum = "11031e251abf8611c80f460e19dbdeb54a66db918e49c65a7065b46ac7aec520"
 dependencies = [
  "rustc_version",
  "simdutf8",
@@ -6618,28 +5239,18 @@ dependencies = [
 
 [[package]]
 name = "siphasher"
-version = "1.0.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b2aa850e253778c88a04c3d7323b043aeda9d3e30d5971937c1855769763678e"
-
-[[package]]
-name = "skrifa"
-version = "0.37.0"
+version = "1.0.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8c31071dedf532758ecf3fed987cdb4bd9509f900e026ab684b4ecb81ea49841"
-dependencies = [
- "bytemuck",
- "read-fonts 0.35.0",
-]
+checksum = "33f4fe9184a62d842c9ef383018f3306d8ba224fd9d836f56d7288308847c256"
 
 [[package]]
 name = "skrifa"
-version = "0.40.0"
+version = "0.44.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7fbdfe3d2475fbd7ddd1f3e5cf8288a30eb3e5f95832829570cd88115a7434ac"
+checksum = "819ab7d62b1d3e72d9d9dea5650bac30424f9111364bb94928dbf5ecad1baa68"
 dependencies = [
  "bytemuck",
- "read-fonts 0.37.0",
+ "read-fonts",
 ]
 
 [[package]]
@@ -6659,147 +5270,24 @@ dependencies = [
 
 [[package]]
 name = "smallvec"
-version = "1.15.1"
+version = "1.16.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "67b1b7a3b5fe4f1376887184045fcf45c69e92af734b7aaddc05fb777b6fbd03"
+checksum = "d0d5d4b1981b8bbbb7aed204cae5726d683d5d6ac7e74ec54122ea28f315e7c8"
 
 [[package]]
 name = "smawk"
-version = "0.3.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b7c388c1b5e93756d0c740965c41e8822f866621d41acbdf6336a6a168f8840c"
-
-[[package]]
-name = "smithay-client-toolkit"
-version = "0.19.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3457dea1f0eb631b4034d61d4d8c32074caa6cd1ab2d59f2327bd8461e2c0016"
-dependencies = [
- "bitflags 2.11.0",
- "calloop 0.13.0",
- "calloop-wayland-source 0.3.0",
- "cursor-icon",
- "libc",
- "log",
- "memmap2",
- "rustix 0.38.44",
- "thiserror 1.0.69",
- "wayland-backend",
- "wayland-client",
- "wayland-csd-frame",
- "wayland-cursor",
- "wayland-protocols",
- "wayland-protocols-wlr",
- "wayland-scanner",
- "xkeysym",
-]
-
-[[package]]
-name = "smithay-client-toolkit"
-version = "0.20.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0512da38f5e2b31201a93524adb8d3136276fa4fe4aafab4e1f727a82b534cc0"
-dependencies = [
- "bitflags 2.11.0",
- "calloop 0.14.4",
- "calloop-wayland-source 0.4.1",
- "cursor-icon",
- "libc",
- "log",
- "memmap2",
- "rustix 1.1.4",
- "thiserror 2.0.18",
- "wayland-backend",
- "wayland-client",
- "wayland-csd-frame",
- "wayland-cursor",
- "wayland-protocols",
- "wayland-protocols-experimental",
- "wayland-protocols-misc",
- "wayland-protocols-wlr",
- "wayland-scanner",
- "xkeysym",
-]
-
-[[package]]
-name = "smithay-clipboard"
-version = "0.7.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "71704c03f739f7745053bde45fa203a46c58d25bc5c4efba1d9a60e9dba81226"
-dependencies = [
- "libc",
- "smithay-client-toolkit 0.20.0",
- "wayland-backend",
-]
-
-[[package]]
-name = "smol_str"
-version = "0.2.2"
+version = "0.3.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "dd538fb6910ac1099850255cf94a94df6551fbdd602454387d0adb2d1ca6dead"
-dependencies = [
- "serde",
-]
+checksum = "e8e2fb0f499abb4d162f2bedad68f5ef91a1682b5a03596ddb67efd37768d100"
 
 [[package]]
 name = "socket2"
-version = "0.6.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3a766e1110788c36f4fa1c2b71b387a7815aa65f88ce0229841826633d93723e"
-dependencies = [
- "libc",
- "windows-sys 0.61.2",
-]
-
-[[package]]
-name = "socks"
-version = "0.3.4"
+version = "0.6.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f0c3dbbd9ae980613c6dd8e28a9407b50509d3803b57624d5dfe8315218cd58b"
+checksum = "c3d1e2c7f27f8d4cb10542a02c49005dbd6e93095799d6f3be745fae9f8fedd4"
 dependencies = [
- "byteorder",
  "libc",
- "winapi",
-]
-
-[[package]]
-name = "softbuffer"
-version = "0.4.8"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "aac18da81ebbf05109ab275b157c22a653bb3c12cf884450179942f81bcbf6c3"
-dependencies = [
- "as-raw-xcb-connection",
- "bytemuck",
- "fastrand",
- "js-sys",
- "memmap2",
- "ndk",
- "objc2 0.6.4",
- "objc2-core-foundation",
- "objc2-core-graphics",
- "objc2-foundation 0.3.2",
- "objc2-quartz-core 0.3.2",
- "raw-window-handle",
- "redox_syscall 0.5.18",
- "rustix 1.1.4",
- "tiny-xlib",
- "tracing",
- "wasm-bindgen",
- "wayland-backend",
- "wayland-client",
- "wayland-sys",
- "web-sys",
  "windows-sys 0.61.2",
- "x11rb",
-]
-
-[[package]]
-name = "spirv"
-version = "0.3.0+sdk-1.3.268.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "eda41003dc44290527a59b13432d4a0379379fa074b70174882adfbdfd917844"
-dependencies = [
- "bitflags 2.11.0",
 ]
 
 [[package]]
@@ -6815,36 +5303,13 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "a2eb9349b6444b326872e140eb1cf5e7c522154d69e7a0ffb0fb81c06b37543f"
 
 [[package]]
-name = "statum"
-version = "0.6.9"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d7b355d3d23dd91652df5621cc0a41b3e831e40a8376a50b4e06020cf4922f01"
-dependencies = [
- "statum-core",
- "statum-macros",
-]
-
-[[package]]
-name = "statum-core"
-version = "0.6.9"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e80c81d094f726bee8a08b3acb08119c56ee9e03e9c96dc62bd0685757fc6b04"
-dependencies = [
- "futures 0.3.32",
- "linkme",
-]
-
-[[package]]
-name = "statum-macros"
-version = "0.6.9"
+name = "statrs"
+version = "0.18.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6f5843996c8d6017fb8046be3ccb3779c69ad41940b8d5c53421200688e18ca7"
+checksum = "2a3fe7c28c6512e766b0874335db33c94ad7b8f9054228ae1c2abd47ce7d335e"
 dependencies = [
- "macro_registry",
- "moddef",
- "proc-macro2",
- "quote",
- "syn",
+ "approx",
+ "num-traits",
 ]
 
 [[package]]
@@ -6868,7 +5333,16 @@ version = "0.27.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "af23d6f6c1a224baef9d3f61e287d2761385a5b88fdab4eb4c6f11aeb54c4bcf"
 dependencies = [
- "strum_macros",
+ "strum_macros 0.27.2",
+]
+
+[[package]]
+name = "strum"
+version = "0.28.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9628de9b8791db39ceda2b119bbe13134770b56c138ec1d3af810d045c04f9bd"
+dependencies = [
+ "strum_macros 0.28.0",
 ]
 
 [[package]]
@@ -6880,57 +5354,59 @@ dependencies = [
  "heck",
  "proc-macro2",
  "quote",
- "syn",
+ "syn 2.0.119",
 ]
 
 [[package]]
-name = "subtle"
-version = "2.6.1"
+name = "strum_macros"
+version = "0.28.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "13c2bddecc57b384dee18652358fb23172facb8a2c51ccc10d74c157bdea3292"
+checksum = "ab85eea0270ee17587ed4156089e10b9e6880ee688791d45a905f5b1ca36f664"
+dependencies = [
+ "heck",
+ "proc-macro2",
+ "quote",
+ "syn 2.0.119",
+]
 
 [[package]]
-name = "svg_fmt"
-version = "0.4.5"
+name = "subtle"
+version = "2.6.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0193cc4331cfd2f3d2011ef287590868599a2f33c3e69bc22c1a3d3acf9e02fb"
+checksum = "13c2bddecc57b384dee18652358fb23172facb8a2c51ccc10d74c157bdea3292"
 
 [[package]]
 name = "svgtypes"
-version = "0.15.3"
+version = "0.16.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "68c7541fff44b35860c1a7a47a7cadf3e4a304c457b58f9870d9706ece028afc"
+checksum = "695b5790b3131dafa99b3bbfd25a216edb3d216dad9ca208d4657bfb8f2abc3d"
 dependencies = [
- "kurbo 0.11.3",
+ "kurbo",
  "siphasher",
 ]
 
 [[package]]
-name = "svgtypes"
-version = "0.16.1"
+name = "symlink"
+version = "0.1.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "695b5790b3131dafa99b3bbfd25a216edb3d216dad9ca208d4657bfb8f2abc3d"
-dependencies = [
- "kurbo 0.13.0",
- "siphasher",
-]
+checksum = "a7973cce6668464ea31f176d85b13c7ab3bba2cb3b77a2ed26abd7801688010a"
 
 [[package]]
-name = "swash"
-version = "0.2.7"
+name = "syn"
+version = "2.0.119"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "842f3cd369c2ba38966204f983eaa5e54a8e84a7d7159ed36ade2b6c335aae64"
+checksum = "872831b642d1a07999a962a351ed35b955ea2cfc8f3862091e2a240a84f17297"
 dependencies = [
- "skrifa 0.40.0",
- "yazi",
- "zeno",
+ "proc-macro2",
+ "quote",
+ "unicode-ident",
 ]
 
 [[package]]
 name = "syn"
-version = "2.0.117"
+version = "3.0.9"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e665b8803e7b1d2a727f4023456bbbbe74da67099c585258af0ad9c5013b9b99"
+checksum = "d78c8dee4c7bf0e14673097256fed6142ce9d3b85a408189d07482442145823b"
 dependencies = [
  "proc-macro2",
  "quote",
@@ -6948,22 +5424,28 @@ dependencies = [
 
 [[package]]
 name = "synstructure"
-version = "0.13.2"
+version = "0.14.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "728a70f3dbaf5bab7f0c4b1ac8d7ae5ea60a4b5549c8a5914361c99147a709d2"
+checksum = "901704edd0dfe137f1987838ee4f259e4e063c31371bdb423f7ae38ec6f77f02"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn",
+ "syn 3.0.9",
 ]
 
 [[package]]
-name = "sys-locale"
-version = "0.3.2"
+name = "sysinfo"
+version = "0.39.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8eab9a99a024a169fe8a903cf9d4a3b3601109bcc13bd9e3c6fff259138626c4"
+checksum = "d2071df9448915b71c4fe6d25deaf1c22f12bd234f01540b77312bb8e41361e6"
 dependencies = [
  "libc",
+ "memchr",
+ "ntapi",
+ "objc2-core-foundation",
+ "objc2-io-kit",
+ "objc2-open-directory",
+ "windows",
 ]
 
 [[package]]
@@ -6972,7 +5454,7 @@ version = "0.7.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "a13f3d0daba03132c0aa9767f98351b3488edc2c100cda2d2ec2b04f3d8d3c8b"
 dependencies = [
- "bitflags 2.11.0",
+ "bitflags 2.13.2",
  "core-foundation 0.9.4",
  "system-configuration-sys",
 ]
@@ -6994,37 +5476,20 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "32497e9a4c7b38532efcdebeef879707aa9f794296a4f0244f6f69e9bc8574bd"
 dependencies = [
  "fastrand",
- "getrandom 0.4.2",
+ "getrandom 0.4.3",
  "once_cell",
- "rustix 1.1.4",
+ "rustix",
  "windows-sys 0.61.2",
 ]
 
-[[package]]
-name = "termcolor"
-version = "1.4.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "06794f8f6c5c898b3275aebefa6b8a1cb24cd2c6c79397ab15774837a0bc5755"
-dependencies = [
- "winapi-util",
-]
-
 [[package]]
 name = "textwrap"
-version = "0.16.2"
+version = "0.16.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c13547615a44dc9c452a8a534638acdf07120d4b6847c8178705da06306a3057"
+checksum = "6ecfad6c3abc80a577f2b91c1e412ee57e7a060d430b553c1b0c940974ebcd49"
 dependencies = [
  "smawk",
-]
-
-[[package]]
-name = "thiserror"
-version = "1.0.69"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b6aaf5339b578ea85b50e080feb250a3e8ae8cfcdff9a461c9ec2904bc923f52"
-dependencies = [
- "thiserror-impl 1.0.69",
+ "unicode-width",
 ]
 
 [[package]]
@@ -7033,29 +5498,27 @@ version = "2.0.18"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "4288b5bcbc7920c07a1149a35cf9590a2aa808e0bc1eafaade0b80947865fbc4"
 dependencies = [
- "thiserror-impl 2.0.18",
+ "thiserror-impl",
 ]
 
 [[package]]
 name = "thiserror-impl"
-version = "1.0.69"
+version = "2.0.18"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "4fee6c4efc90059e10f81e6d42c60a18f76588c3d74cb83a0b242a2b6c7504c1"
+checksum = "ebc4ee7f67670e9b64d05fa4253e753e016c6c95ff35b89b7941d6b856dec1d5"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn",
+ "syn 2.0.119",
 ]
 
 [[package]]
-name = "thiserror-impl"
-version = "2.0.18"
+name = "thread_local"
+version = "1.1.10"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ebc4ee7f67670e9b64d05fa4253e753e016c6c95ff35b89b7941d6b856dec1d5"
+checksum = "1ad99c4c6d32803332c548b1af0540b357b3f5fc0be8f6c6bfe8b2e6ae784070"
 dependencies = [
- "proc-macro2",
- "quote",
- "syn",
+ "cfg-if",
 ]
 
 [[package]]
@@ -7069,17 +5532,16 @@ dependencies = [
  "half",
  "quick-error",
  "weezl",
- "zune-jpeg 0.5.15",
+ "zune-jpeg",
 ]
 
 [[package]]
 name = "time"
-version = "0.3.47"
+version = "0.3.55"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "743bd48c283afc0388f9b8827b976905fb217ad9e647fae3a379a9283c4def2c"
+checksum = "cdb87b95ec50ddfa440816d227a17b2ccbdda963a316a727fda0fc4334f7d134"
 dependencies = [
  "deranged",
- "itoa",
  "num-conv",
  "powerfmt",
  "serde_core",
@@ -7089,15 +5551,15 @@ dependencies = [
 
 [[package]]
 name = "time-core"
-version = "0.1.8"
+version = "0.1.9"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7694e1cfe791f8d31026952abf09c69ca6f6fa4e1a1229e18988f06a04a12dca"
+checksum = "9e1c906769ad99c88eaa54e728060edef082f8e358ff32030cb7c7d315e81109"
 
 [[package]]
 name = "time-macros"
-version = "0.2.27"
+version = "0.2.32"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2e70e4c5a0e0a8a4823ad65dfe1a6930e4f4d756dcd9dd7939022b5e8c501215"
+checksum = "7e689342a48d2ea927c87ea50cabf8594854bf940e9310208848d680d668ed85"
 dependencies = [
  "num-conv",
  "time-core",
@@ -7105,48 +5567,35 @@ dependencies = [
 
 [[package]]
 name = "tiny-skia"
-version = "0.11.4"
+version = "0.12.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "83d13394d44dae3207b52a326c0c85a8bf87f1541f23b0d143811088497b09ab"
+checksum = "47ffee5eaaf5527f630fb0e356b90ebdec84d5d18d937c5e440350f88c5a91ea"
 dependencies = [
  "arrayref",
  "arrayvec",
  "bytemuck",
  "cfg-if",
  "log",
- "png 0.17.16",
+ "png",
  "tiny-skia-path",
 ]
 
 [[package]]
 name = "tiny-skia-path"
-version = "0.11.4"
+version = "0.12.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9c9e7fc0c2e86a30b117d0462aa261b72b7a99b7ebd7deb3a14ceda95c5bdc93"
+checksum = "edca365c3faccca67d06593c5980fa6c57687de727a03131735bb85f01fdeeb9"
 dependencies = [
  "arrayref",
  "bytemuck",
  "strict-num",
 ]
 
-[[package]]
-name = "tiny-xlib"
-version = "0.2.4"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0324504befd01cab6e0c994f34b2ffa257849ee019d3fb3b64fb2c858887d89e"
-dependencies = [
- "as-raw-xcb-connection",
- "ctor-lite",
- "libloading",
- "pkg-config",
- "tracing",
-]
-
 [[package]]
 name = "tinystr"
-version = "0.8.3"
+version = "0.8.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c8323304221c2a851516f22236c5722a72eaa19749016521d6dff0824447d96d"
+checksum = "abc840c238154d58f9e86b2792261edfde565a17696e1cda1a449de27979801a"
 dependencies = [
  "displaydoc",
  "zerovec",
@@ -7154,29 +5603,20 @@ dependencies = [
 
 [[package]]
 name = "tinyvec"
-version = "1.11.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3e61e67053d25a4e82c844e8424039d9745781b3fc4f32b8d55ed50f5f667ef3"
-dependencies = [
- "tinyvec_macros",
-]
-
-[[package]]
-name = "tinyvec_macros"
-version = "0.1.1"
+version = "1.13.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1f3ccbac311fea05f86f61904b462b55fb3df8837a366dfc601a0161d0532f20"
+checksum = "fd3ca314f692efd6c868f8408f53fe444634a845f96c028b97d35f6a1f79f0ee"
 
 [[package]]
 name = "tokio"
-version = "1.51.0"
+version = "1.54.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2bd1c4c0fc4a7ab90fc15ef6daaa3ec3b893f004f915f2392557ed23237820cd"
+checksum = "b52fd3255258d26fac7264ab5a76b53c13367136dc011d3d6ad4f8ed070396a3"
 dependencies = [
- "bytes 1.11.1",
+ "bytes",
  "libc",
  "mio",
- "parking_lot",
+ "parking_lot 0.12.5",
  "pin-project-lite",
  "signal-hook-registry",
  "socket2",
@@ -7184,43 +5624,33 @@ dependencies = [
  "windows-sys 0.61.2",
 ]
 
-[[package]]
-name = "tokio-io"
-version = "0.1.13"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "57fc868aae093479e3131e3d165c93b1c7474109d13c90ec0dda2a1bbfff0674"
-dependencies = [
- "bytes 0.4.12",
- "futures 0.1.31",
- "log",
-]
-
 [[package]]
 name = "tokio-macros"
-version = "2.7.0"
+version = "2.7.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "385a6cb71ab9ab790c5fe8d67f1645e6c450a7ce006a33de03daa956cf70a496"
+checksum = "78773a2a397f451582ce068015985c33193cf6dea8b74d2a639fe457b2f07b0e"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn",
+ "syn 3.0.9",
 ]
 
 [[package]]
-name = "tokio-native-tls"
-version = "0.3.1"
+name = "tokio-retry"
+version = "0.3.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "bbae76ab933c85776efabc971569dd6119c580d8f5d448769dec1764bf796ef2"
+checksum = "4a129d95275ebf4c493ec53bf0f8cd95f5ac161bc4f381700809a54f595d4470"
 dependencies = [
- "native-tls",
+ "pin-project-lite",
+ "rand 0.10.3",
  "tokio",
 ]
 
 [[package]]
 name = "tokio-rustls"
-version = "0.26.4"
+version = "0.26.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1729aa945f29d91ba541258c8df89027d5792d85a8841fb65e8bf0f4ede4ef61"
+checksum = "c9cc2678c2cdd569ef8215e2afd7954ada2ae20b4fdd2c5fe6139a3b02d105db"
 dependencies = [
  "rustls",
  "tokio",
@@ -7228,9 +5658,9 @@ dependencies = [
 
 [[package]]
 name = "tokio-stream"
-version = "0.1.18"
+version = "0.1.19"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "32da49809aab5c3bc678af03902d4ccddea2a87d028d86392a4b1560c6906c70"
+checksum = "a3d06f0b082ba57c26b79407372e57cf2a1e28124f78e9479fe80322cf53420b"
 dependencies = [
  "futures-core",
  "pin-project-lite",
@@ -7240,9 +5670,9 @@ dependencies = [
 
 [[package]]
 name = "tokio-test"
-version = "0.4.5"
+version = "0.4.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3f6d24790a10a7af737693a3e8f1d03faef7e6ca0cc99aae5066f533766de545"
+checksum = "12bc5db7778a8e08ee255090861e0e7aac55e784300077e7ed293c7be36168b2"
 dependencies = [
  "futures-core",
  "tokio",
@@ -7251,71 +5681,70 @@ dependencies = [
 
 [[package]]
 name = "tokio-tungstenite"
-version = "0.28.0"
+version = "0.30.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d25a406cddcc431a75d3d9afc6a7c0f7428d4891dd973e4d54c56b46127bf857"
+checksum = "17a073bfed563fa236697a068031408a93cd9522e08abf9933ead3e73411bd71"
 dependencies = [
  "futures-util",
  "log",
+ "rustls",
+ "rustls-native-certs",
+ "rustls-pki-types",
  "tokio",
+ "tokio-rustls",
  "tungstenite",
 ]
 
 [[package]]
 name = "tokio-util"
-version = "0.7.18"
+version = "0.7.20"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9ae9cec805b01e8fc3fd2fe289f89149a9b66dd16786abd8b19cfa7b48cb0098"
+checksum = "e464cf451ba96ebfc6f9b6542f17ee8b8956e33f1e40d9690624e59d7a7f8a4b"
 dependencies = [
- "bytes 1.11.1",
+ "bytes",
  "futures-core",
  "futures-sink",
  "futures-util",
+ "libc",
  "pin-project-lite",
  "tokio",
 ]
 
 [[package]]
-name = "toktrie"
-version = "1.7.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0310776df61c26b98e6ed7c3629de5131e0c8731439350b79152b7f52468229a"
-dependencies = [
- "anyhow",
- "bytemuck",
- "bytemuck_derive",
- "serde",
- "serde_json",
-]
-
-[[package]]
-name = "toml_datetime"
-version = "1.1.1+spec-1.1.0"
+name = "tokio_with_wasm"
+version = "0.8.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3165f65f62e28e0115a00b2ebdd37eb6f3b641855f9d636d3cd4103767159ad7"
+checksum = "34e40fbbbd95441133fe9483f522db15dbfd26dc636164ebd8f2dd28759a6aa6"
 dependencies = [
- "serde_core",
+ "js-sys",
+ "tokio",
+ "tokio_with_wasm_proc",
+ "wasm-bindgen",
+ "wasm-bindgen-futures",
+ "web-sys",
 ]
 
 [[package]]
-name = "toml_edit"
-version = "0.25.11+spec-1.1.0"
+name = "tokio_with_wasm_proc"
+version = "0.8.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0b59c4d22ed448339746c59b905d24568fcbb3ab65a500494f7b8c3e97739f2b"
+checksum = "d01145a2c788d6aae4cd653afec1e8332534d7d783d01897cefcafe4428de992"
 dependencies = [
- "indexmap",
- "toml_datetime",
- "toml_parser",
- "winnow 1.0.1",
+ "quote",
+ "syn 2.0.119",
 ]
 
 [[package]]
-name = "toml_parser"
-version = "1.1.2+spec-1.1.0"
+name = "toktrie"
+version = "1.7.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a2abe9b86193656635d2411dc43050282ca48aa31c2451210f4202550afb7526"
+checksum = "0310776df61c26b98e6ed7c3629de5131e0c8731439350b79152b7f52468229a"
 dependencies = [
- "winnow 1.0.1",
+ "anyhow",
+ "bytemuck",
+ "bytemuck_derive",
+ "serde",
+ "serde_json",
 ]
 
 [[package]]
@@ -7337,20 +5766,20 @@ dependencies = [
 
 [[package]]
 name = "tower-http"
-version = "0.6.8"
+version = "0.6.11"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d4e6559d53cc268e5031cd8429d05415bc4cb4aefc4aa5d6cc35fbf5b924a1f8"
+checksum = "4cfcf7e2740e6fc6d4d688b4ef00650406bb94adf4731e43c096c3a19fe40840"
 dependencies = [
- "bitflags 2.11.0",
- "bytes 1.11.1",
+ "bitflags 2.13.2",
+ "bytes",
  "futures-util",
- "http 1.4.0",
+ "http 1.5.0",
  "http-body",
- "iri-string",
  "pin-project-lite",
  "tower",
  "tower-layer",
  "tower-service",
+ "url",
 ]
 
 [[package]]
@@ -7377,6 +5806,19 @@ dependencies = [
  "tracing-core",
 ]
 
+[[package]]
+name = "tracing-appender"
+version = "0.2.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "050686193eb999b4bb3bc2acfa891a13da00f79734704c4b8b4ef1a10b368a3c"
+dependencies = [
+ "crossbeam-channel",
+ "symlink",
+ "thiserror",
+ "time",
+ "tracing-subscriber",
+]
+
 [[package]]
 name = "tracing-attributes"
 version = "0.1.31"
@@ -7385,16 +5827,59 @@ checksum = "7490cfa5ec963746568740651ac6781f701c9c5ea257c58e057f3ba8cf69e8da"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn",
+ "syn 2.0.119",
+]
+
+[[package]]
+name = "tracing-core"
+version = "0.1.36"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "db97caf9d906fbde555dd62fa95ddba9eecfd14cb388e4f491a66d74cd5fb79a"
+dependencies = [
+ "once_cell",
+ "valuable",
+]
+
+[[package]]
+name = "tracing-log"
+version = "0.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ee855f1f400bd0e5c02d150ae5de3840039a3f54b025156404e34c23c03f47c3"
+dependencies = [
+ "log",
+ "once_cell",
+ "tracing-core",
+]
+
+[[package]]
+name = "tracing-serde"
+version = "0.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "704b1aeb7be0d0a84fc9828cae51dab5970fee5088f83d1dd7ee6f6246fc6ff1"
+dependencies = [
+ "serde",
+ "tracing-core",
 ]
 
 [[package]]
-name = "tracing-core"
-version = "0.1.36"
+name = "tracing-subscriber"
+version = "0.3.23"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "db97caf9d906fbde555dd62fa95ddba9eecfd14cb388e4f491a66d74cd5fb79a"
+checksum = "cb7f578e5945fb242538965c2d0b04418d38ec25c79d160cd279bf0731c8d319"
 dependencies = [
+ "matchers",
+ "nu-ansi-term",
  "once_cell",
+ "regex-automata",
+ "serde",
+ "serde_json",
+ "sharded-slab",
+ "smallvec",
+ "thread_local",
+ "tracing",
+ "tracing-core",
+ "tracing-log",
+ "tracing-serde",
 ]
 
 [[package]]
@@ -7405,66 +5890,60 @@ checksum = "e421abadd41a4225275504ea4d6566923418b7f05506fbc9c0fe86ba7396114b"
 
 [[package]]
 name = "trzcina"
-version = "0.3.0"
+version = "0.4.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c58d685f2fd0c04ea9c0dcff193161de5aebf1eba5e1267807e5ae2a8562cd93"
+checksum = "0105cd7bb068e2c2befb3910c55ccf33b07b7ce0592161be5044cc6233366d4f"
 dependencies = [
  "anyhow",
  "async-trait",
  "futures-util",
  "log",
+ "thiserror",
  "tokio",
  "tokio-util",
 ]
 
-[[package]]
-name = "ttf-parser"
-version = "0.25.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d2df906b07856748fa3f6e0ad0cbaa047052d4a7dd609e231c4f72cee8c36f31"
-dependencies = [
- "core_maths",
-]
-
 [[package]]
 name = "tungstenite"
-version = "0.28.0"
+version = "0.30.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8628dcc84e5a09eb3d8423d6cb682965dea9133204e8fb3efee74c2a0c259442"
+checksum = "e48ac77174b19c110a50ab2128b24215ac9cb40e0e12e093fb602d175c569d22"
 dependencies = [
- "bytes 1.11.1",
+ "bytes",
  "data-encoding",
- "http 1.4.0",
+ "http 1.5.0",
  "httparse",
  "log",
- "rand 0.9.4",
- "sha1 0.10.6",
- "thiserror 2.0.18",
- "utf-8",
+ "rand 0.10.3",
+ "rustls",
+ "rustls-pki-types",
+ "sha1 0.11.0",
+ "thiserror",
 ]
 
+[[package]]
+name = "twox-hash"
+version = "2.1.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "86a801b3cea342a06d468c8710662aa29e5e05e4f5c0d62f00bbb7f2ad7941c2"
+
 [[package]]
 name = "typenum"
-version = "1.19.0"
+version = "1.20.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "562d481066bde0658276a35467c4af00bdc6ee726305698a55b86e61d7ad82bb"
+checksum = "b6f5e870be6c3b371b77fe0ee0bafb859fa4964b4404c27de1d380043c4dda20"
 
 [[package]]
-name = "uds_windows"
-version = "1.2.1"
+name = "typewit"
+version = "1.15.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f2f6fb2847f6742cd76af783a2a2c49e9375d0a111c7bef6f71cd9e738c72d6e"
-dependencies = [
- "memoffset",
- "tempfile",
- "windows-sys 0.61.2",
-]
+checksum = "214ca0b2191785cbc06209b9ca1861e048e39b5ba33574b3cedd58363d5bb5f6"
 
 [[package]]
 name = "unicase"
-version = "2.9.0"
+version = "2.10.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "dbc4bc3a9f746d862c45cb89d705aa10f187bb96c76001afab07a0d35ce60142"
+checksum = "357cc3acc6a036009fd6c973ed009037c732d60d0b4f6c673e9041497482a28f"
 
 [[package]]
 name = "unicode-bidi"
@@ -7472,18 +5951,6 @@ version = "0.3.18"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "5c1cb5db39152898a79168971543b1cb5020dff7fe43c8dc468b0885f5e29df5"
 
-[[package]]
-name = "unicode-bidi-mirroring"
-version = "0.4.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5dfa6e8c60bb66d49db113e0125ee8711b7647b5579dc7f5f19c42357ed039fe"
-
-[[package]]
-name = "unicode-ccc"
-version = "0.4.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ce61d488bcdc9bc8b5d1772c404828b17fc481c0a582b5581e95fb233aef503e"
-
 [[package]]
 name = "unicode-general-category"
 version = "1.1.0"
@@ -7492,21 +5959,9 @@ checksum = "0b993bddc193ae5bd0d623b49ec06ac3e9312875fdae725a975c51db1cc1677f"
 
 [[package]]
 name = "unicode-ident"
-version = "1.0.24"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e6e4313cd5fcd3dad5cafa179702e2b244f760991f45397d14d4ebf38247da75"
-
-[[package]]
-name = "unicode-linebreak"
-version = "0.1.5"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3b09c83c3c29d37506a3e260c08c03743a6bb66a9cd432c6934ab501a190571f"
-
-[[package]]
-name = "unicode-properties"
-version = "0.1.4"
+version = "1.0.27"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7df058c713841ad818f1dc5d3fd88063241cc61f49f5fbea4b951e8cf5a8d71d"
+checksum = "a2c754d6c33795a1c324727428e5a7dedb5b06195f9890bdbcba760d3e246563"
 
 [[package]]
 name = "unicode-script"
@@ -7516,9 +5971,9 @@ checksum = "383ad40bb927465ec0ce7720e033cb4ca06912855fc35db31b5755d0de75b1ee"
 
 [[package]]
 name = "unicode-segmentation"
-version = "1.13.2"
+version = "1.13.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9629274872b2bfaf8d66f5f15725007f635594914870f65218920345aa11aa8c"
+checksum = "c6f5d3c3b1bf09027a88a6bc961fc00497d651009560b5463668dc81b0fa87a8"
 
 [[package]]
 name = "unicode-vo"
@@ -7550,26 +6005,6 @@ version = "0.9.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "8ecb6da28b8a351d773b68d5825ac39017e680750f980f3a1a85cd8dd28a47c1"
 
-[[package]]
-name = "ureq"
-version = "2.12.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "02d1a66277ed75f640d608235660df48c8e3c19f3b4edb6a263315626cc3c01d"
-dependencies = [
- "base64",
- "flate2",
- "log",
- "native-tls",
- "once_cell",
- "rustls",
- "rustls-pki-types",
- "serde",
- "serde_json",
- "socks",
- "url",
- "webpki-roots 0.26.11",
-]
-
 [[package]]
 name = "url"
 version = "2.5.8"
@@ -7584,52 +6019,32 @@ dependencies = [
 ]
 
 [[package]]
-name = "usvg"
-version = "0.45.1"
+name = "urlencoding"
+version = "2.1.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "80be9b06fbae3b8b303400ab20778c80bbaf338f563afe567cf3c9eea17b47ef"
-dependencies = [
- "base64",
- "data-url",
- "flate2",
- "fontdb",
- "imagesize 0.13.0",
- "kurbo 0.11.3",
- "log",
- "pico-args",
- "roxmltree 0.20.0",
- "rustybuzz",
- "simplecss",
- "siphasher",
- "strict-num",
- "svgtypes 0.15.3",
- "tiny-skia-path",
- "unicode-bidi",
- "unicode-script",
- "unicode-vo",
- "xmlwriter",
-]
+checksum = "daf8dba3b7eb870caf1ddeed7bc9d2a049f3cfdfae7cb521b087cc33ae4c49da"
 
 [[package]]
 name = "usvg"
-version = "0.46.0"
+version = "0.49.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e419dff010bb12512b0ae9e3d2f318dfbdf0167fde7eb05465134d4e8756076f"
+checksum = "d09b73ee83641d390ad7ffd9000f347b16eb128e437675b6ff277ea5c7fcf22c"
 dependencies = [
- "base64",
+ "base64 0.23.1",
  "data-url",
  "flate2",
  "fontdb",
- "imagesize 0.14.0",
- "kurbo 0.13.0",
+ "harfrust",
+ "imagesize",
+ "kurbo",
  "log",
  "pico-args",
- "roxmltree 0.21.1",
- "rustybuzz",
+ "roxmltree 0.21.2",
  "simplecss",
  "siphasher",
+ "skrifa",
  "strict-num",
- "svgtypes 0.16.1",
+ "svgtypes",
  "tiny-skia-path",
  "unicode-bidi",
  "unicode-script",
@@ -7638,10 +6053,10 @@ dependencies = [
 ]
 
 [[package]]
-name = "utf-8"
-version = "0.7.6"
+name = "utf16_iter"
+version = "1.0.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "09cc8ee72d2a9becf2f2febe0205bbed8fc6615b7cb429ad062dc7b7ddd036a9"
+checksum = "c8232dd3cdaed5356e0f716d285e4b40b932ac434100fe9b7e0e8e935b9e6246"
 
 [[package]]
 name = "utf8_iter"
@@ -7657,12 +6072,12 @@ checksum = "06abde3611657adf66d383f00b093d7faecc7fa57071cce2578660c9f1010821"
 
 [[package]]
 name = "uuid"
-version = "1.23.1"
+version = "1.29.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ddd74a9687298c6858e9b88ec8935ec45d22e8fd5e6394fa1bd4e99a87789c76"
+checksum = "4a0e2d11620c50695fe12bbbfde96b3477fe606e7284c22d505c12ce51214514"
 dependencies = [
+ "getrandom 0.4.3",
  "js-sys",
- "serde_core",
  "wasm-bindgen",
 ]
 
@@ -7688,10 +6103,10 @@ dependencies = [
 ]
 
 [[package]]
-name = "vcpkg"
-version = "0.2.15"
+name = "valuable"
+version = "0.1.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "accd4ea62f7bb7a82fe23066fb0957d48ef677f6eeb8215f372f52e48bb32426"
+checksum = "ba73ea9cf16a25df0c8caa16c51acb937d5712a8429db78a3ee29d5dcacd3a65"
 
 [[package]]
 name = "version_check"
@@ -7717,9 +6132,9 @@ dependencies = [
 
 [[package]]
 name = "want"
-version = "0.3.1"
+version = "0.3.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "bfa7760aed19e106de2c7c0b581b509f2f25d3dacaf737cb82ac61bc6d760b0e"
+checksum = "ec4cdd0dd910afe868b7ef477227d8d538b46b3075031afee8a9f2acb0a2ed0b"
 dependencies = [
  "try-lock",
 ]
@@ -7730,29 +6145,38 @@ version = "0.11.1+wasi-snapshot-preview1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "ccf3ec651a847eb01de73ccad15eb7d99f80485de043efb2f370cd654f4ea44b"
 
+[[package]]
+name = "wasi"
+version = "0.14.7+wasi-0.2.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "883478de20367e224c0090af9cf5f9fa85bed63a95c1abf3afc5c083ebc06e8c"
+dependencies = [
+ "wasip2",
+]
+
 [[package]]
 name = "wasip2"
-version = "1.0.2+wasi-0.2.9"
+version = "1.0.4+wasi-0.2.12"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9517f9239f02c069db75e65f174b3da828fe5f5b945c4dd26bd25d89c03ebcf5"
+checksum = "b67efb37e106e55ce722a510d6b5f9c17f083e5fc79afc2badeb12cc313d9487"
 dependencies = [
  "wit-bindgen",
 ]
 
 [[package]]
-name = "wasip3"
-version = "0.4.0+wasi-0.3.0-rc-2026-01-06"
+name = "wasite"
+version = "1.0.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5428f8bf88ea5ddc08faddef2ac4a67e390b88186c703ce6dbd955e1c145aca5"
+checksum = "66fe902b4a6b8028a753d5424909b764ccf79b7a209eac9bf97e59cda9f71a42"
 dependencies = [
- "wit-bindgen",
+ "wasi 0.14.7+wasi-0.2.4",
 ]
 
 [[package]]
 name = "wasm-bindgen"
-version = "0.2.117"
+version = "0.2.129"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0551fc1bb415591e3372d0bc4780db7e587d84e2a7e79da121051c5c4b89d0b0"
+checksum = "9bb54f33acc68fd454578d9820b0bde1a1a3d17aa17bb7b6595806d02886d409"
 dependencies = [
  "cfg-if",
  "once_cell",
@@ -7762,457 +6186,106 @@ dependencies = [
 ]
 
 [[package]]
-name = "wasm-bindgen-futures"
-version = "0.4.67"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "03623de6905b7206edd0a75f69f747f134b7f0a2323392d664448bf2d3c5d87e"
-dependencies = [
- "js-sys",
- "wasm-bindgen",
-]
-
-[[package]]
-name = "wasm-bindgen-macro"
-version = "0.2.117"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7fbdf9a35adf44786aecd5ff89b4563a90325f9da0923236f6104e603c7e86be"
-dependencies = [
- "quote",
- "wasm-bindgen-macro-support",
-]
-
-[[package]]
-name = "wasm-bindgen-macro-support"
-version = "0.2.117"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "dca9693ef2bab6d4e6707234500350d8dad079eb508dca05530c85dc3a529ff2"
-dependencies = [
- "bumpalo",
- "proc-macro2",
- "quote",
- "syn",
- "wasm-bindgen-shared",
-]
-
-[[package]]
-name = "wasm-bindgen-shared"
-version = "0.2.117"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "39129a682a6d2d841b6c429d0c51e5cb0ed1a03829d8b3d1e69a011e62cb3d3b"
-dependencies = [
- "unicode-ident",
-]
-
-[[package]]
-name = "wasm-encoder"
-version = "0.244.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "990065f2fe63003fe337b932cfb5e3b80e0b4d0f5ff650e6985b1048f62c8319"
-dependencies = [
- "leb128fmt",
- "wasmparser",
-]
-
-[[package]]
-name = "wasm-metadata"
-version = "0.244.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "bb0e353e6a2fbdc176932bbaab493762eb1255a7900fe0fea1a2f96c296cc909"
-dependencies = [
- "anyhow",
- "indexmap",
- "wasm-encoder",
- "wasmparser",
-]
-
-[[package]]
-name = "wasm-streams"
-version = "0.4.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "15053d8d85c7eccdbefef60f06769760a563c7f0a9d6902a13d35c7800b0ad65"
-dependencies = [
- "futures-util",
- "js-sys",
- "wasm-bindgen",
- "wasm-bindgen-futures",
- "web-sys",
-]
-
-[[package]]
-name = "wasm-streams"
-version = "0.5.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9d1ec4f6517c9e11ae630e200b2b65d193279042e28edd4a2cda233e46670bbb"
-dependencies = [
- "futures-util",
- "js-sys",
- "wasm-bindgen",
- "wasm-bindgen-futures",
- "web-sys",
-]
-
-[[package]]
-name = "wasmparser"
-version = "0.244.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "47b807c72e1bac69382b3a6fb3dbe8ea4c0ed87ff5629b8685ae6b9a611028fe"
-dependencies = [
- "bitflags 2.11.0",
- "hashbrown 0.15.5",
- "indexmap",
- "semver",
-]
-
-[[package]]
-name = "wasmtimer"
-version = "0.4.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1c598d6b99ea013e35844697fc4670d08339d5cda15588f193c6beedd12f644b"
-dependencies = [
- "futures 0.3.32",
- "js-sys",
- "parking_lot",
- "pin-utils",
- "slab",
- "wasm-bindgen",
-]
-
-[[package]]
-name = "wayland-backend"
-version = "0.3.15"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2857dd20b54e916ec7253b3d6b4d5c4d7d4ca2c33c2e11c6c76a99bd8744755d"
-dependencies = [
- "cc",
- "downcast-rs",
- "rustix 1.1.4",
- "scoped-tls",
- "smallvec",
- "wayland-sys",
-]
-
-[[package]]
-name = "wayland-client"
-version = "0.31.14"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "645c7c96bb74690c3189b5c9cb4ca1627062bb23693a4fad9d8c3de958260144"
-dependencies = [
- "bitflags 2.11.0",
- "rustix 1.1.4",
- "wayland-backend",
- "wayland-scanner",
-]
-
-[[package]]
-name = "wayland-csd-frame"
-version = "0.3.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "625c5029dbd43d25e6aa9615e88b829a5cad13b2819c4ae129fdbb7c31ab4c7e"
-dependencies = [
- "bitflags 2.11.0",
- "cursor-icon",
- "wayland-backend",
-]
-
-[[package]]
-name = "wayland-cursor"
-version = "0.31.14"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "4a52d18780be9b1314328a3de5f930b73d2200112e3849ca6cb11822793fb34d"
-dependencies = [
- "rustix 1.1.4",
- "wayland-client",
- "xcursor",
-]
-
-[[package]]
-name = "wayland-protocols"
-version = "0.32.12"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "563a85523cade2429938e790815fd7319062103b9f4a2dc806e9b53b95982d8f"
-dependencies = [
- "bitflags 2.11.0",
- "wayland-backend",
- "wayland-client",
- "wayland-scanner",
-]
-
-[[package]]
-name = "wayland-protocols-experimental"
-version = "20250721.0.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "40a1f863128dcaaec790d7b4b396cc9b9a7a079e878e18c47e6c2d2c5a8dcbb1"
-dependencies = [
- "bitflags 2.11.0",
- "wayland-backend",
- "wayland-client",
- "wayland-protocols",
- "wayland-scanner",
-]
-
-[[package]]
-name = "wayland-protocols-misc"
-version = "0.3.12"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6e9567599ef23e09b8dad6e429e5738d4509dfc46b3b21f32841a304d16b29c8"
-dependencies = [
- "bitflags 2.11.0",
- "wayland-backend",
- "wayland-client",
- "wayland-protocols",
- "wayland-scanner",
-]
-
-[[package]]
-name = "wayland-protocols-plasma"
-version = "0.3.12"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2b6d8cf1eb2c1c31ed1f5643c88a6e53538129d4af80030c8cabd1f9fa884d91"
-dependencies = [
- "bitflags 2.11.0",
- "wayland-backend",
- "wayland-client",
- "wayland-protocols",
- "wayland-scanner",
-]
-
-[[package]]
-name = "wayland-protocols-wlr"
-version = "0.3.12"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "eb04e52f7836d7c7976c78ca0250d61e33873c34156a2a1fc9474828ec268234"
-dependencies = [
- "bitflags 2.11.0",
- "wayland-backend",
- "wayland-client",
- "wayland-protocols",
- "wayland-scanner",
-]
-
-[[package]]
-name = "wayland-scanner"
-version = "0.31.10"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9c324a910fd86ebdc364a3e61ec1f11737d3b1d6c273c0239ee8ff4bc0d24b4a"
-dependencies = [
- "proc-macro2",
- "quick-xml",
- "quote",
-]
-
-[[package]]
-name = "wayland-sys"
-version = "0.31.11"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d8eab23fefc9e41f8e841df4a9c707e8a8c4ed26e944ef69297184de2785e3be"
-dependencies = [
- "dlib",
- "log",
- "once_cell",
- "pkg-config",
-]
-
-[[package]]
-name = "web-sys"
-version = "0.3.94"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "cd70027e39b12f0849461e08ffc50b9cd7688d942c1c8e3c7b22273236b4dd0a"
-dependencies = [
- "js-sys",
- "wasm-bindgen",
-]
-
-[[package]]
-name = "web-time"
-version = "1.1.0"
+name = "wasm-bindgen-futures"
+version = "0.4.79"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5a6580f308b1fad9207618087a65c04e7a10bc77e02c8e84e9b00dd4b12fa0bb"
+checksum = "3cbab34de2d982e9b48e18d216d04c4a6f641066ff19ffb699980f591ee3610e"
 dependencies = [
  "js-sys",
+ "tokio",
  "wasm-bindgen",
 ]
 
 [[package]]
-name = "webpki-root-certs"
-version = "1.0.7"
+name = "wasm-bindgen-macro"
+version = "0.2.129"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f31141ce3fc3e300ae89b78c0dd67f9708061d1d2eda54b8209346fd6be9a92c"
+checksum = "2e29d0c35b16e224a7eeb5cd2d25e3e1968fbd65604117b44d3b789d00ee8535"
 dependencies = [
- "rustls-pki-types",
+ "quote",
+ "wasm-bindgen-macro-support",
 ]
 
 [[package]]
-name = "webpki-roots"
-version = "0.26.11"
+name = "wasm-bindgen-macro-support"
+version = "0.2.129"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "521bc38abb08001b01866da9f51eb7c5d647a19260e00054a8c7fd5f9e57f7a9"
+checksum = "6f501a8bc3719dba86ef8ae4728879c08001bea749eb1333ac5b91e040e2a6b7"
 dependencies = [
- "webpki-roots 1.0.6",
+ "bumpalo",
+ "proc-macro2",
+ "quote",
+ "syn 3.0.9",
+ "wasm-bindgen-shared",
 ]
 
 [[package]]
-name = "webpki-roots"
-version = "1.0.6"
+name = "wasm-bindgen-shared"
+version = "0.2.129"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "22cfaf3c063993ff62e73cb4311efde4db1efb31ab78a3e5c457939ad5cc0bed"
+checksum = "23f0c9c52aa7cd7d77769a4cfe2a9adb1b331f489a41d912ce14513d5ab995c6"
 dependencies = [
- "rustls-pki-types",
+ "unicode-ident",
 ]
 
 [[package]]
-name = "weezl"
-version = "0.1.12"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a28ac98ddc8b9274cb41bb4d9d4d5c425b6020c50c46f25559911905610b4a88"
-
-[[package]]
-name = "wgpu"
-version = "27.0.1"
+name = "wasm-streams"
+version = "0.5.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "bfe68bac7cde125de7a731c3400723cadaaf1703795ad3f4805f187459cd7a77"
+checksum = "9d1ec4f6517c9e11ae630e200b2b65d193279042e28edd4a2cda233e46670bbb"
 dependencies = [
- "arrayvec",
- "bitflags 2.11.0",
- "cfg-if",
- "cfg_aliases",
- "document-features",
- "hashbrown 0.16.1",
+ "futures-util",
  "js-sys",
- "log",
- "naga",
- "parking_lot",
- "portable-atomic",
- "profiling",
- "raw-window-handle",
- "smallvec",
- "static_assertions",
  "wasm-bindgen",
  "wasm-bindgen-futures",
  "web-sys",
- "wgpu-core",
- "wgpu-hal",
- "wgpu-types",
 ]
 
 [[package]]
-name = "wgpu-core"
-version = "27.0.3"
+name = "web-sys"
+version = "0.3.106"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "27a75de515543b1897b26119f93731b385a19aea165a1ec5f0e3acecc229cae7"
+checksum = "88261b9deccee56594c11a3460c462c41f58d148598fe70ad77070126a68aba4"
 dependencies = [
- "arrayvec",
- "bit-set",
- "bit-vec",
- "bitflags 2.11.0",
- "bytemuck",
- "cfg_aliases",
- "document-features",
- "hashbrown 0.16.1",
- "indexmap",
- "log",
- "naga",
- "once_cell",
- "parking_lot",
- "portable-atomic",
- "profiling",
- "raw-window-handle",
- "rustc-hash 1.1.0",
- "smallvec",
- "thiserror 2.0.18",
- "wgpu-core-deps-apple",
- "wgpu-core-deps-emscripten",
- "wgpu-core-deps-windows-linux-android",
- "wgpu-hal",
- "wgpu-types",
+ "js-sys",
+ "wasm-bindgen",
 ]
 
 [[package]]
-name = "wgpu-core-deps-apple"
-version = "27.0.0"
+name = "web-time"
+version = "1.1.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0772ae958e9be0c729561d5e3fd9a19679bcdfb945b8b1a1969d9bfe8056d233"
+checksum = "5a6580f308b1fad9207618087a65c04e7a10bc77e02c8e84e9b00dd4b12fa0bb"
 dependencies = [
- "wgpu-hal",
+ "js-sys",
+ "wasm-bindgen",
 ]
 
 [[package]]
-name = "wgpu-core-deps-emscripten"
-version = "27.0.0"
+name = "webpki-root-certs"
+version = "1.0.11"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b06ac3444a95b0813ecfd81ddb2774b66220b264b3e2031152a4a29fda4da6b5"
+checksum = "e49d0e65abc00c142e8f9428da991da9f1a250c39be97459901eb12b9b49f62a"
 dependencies = [
- "wgpu-hal",
+ "rustls-pki-types",
 ]
 
 [[package]]
-name = "wgpu-core-deps-windows-linux-android"
-version = "27.0.0"
+name = "weezl"
+version = "0.1.12"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "71197027d61a71748e4120f05a9242b2ad142e3c01f8c1b47707945a879a03c3"
-dependencies = [
- "wgpu-hal",
-]
+checksum = "a28ac98ddc8b9274cb41bb4d9d4d5c425b6020c50c46f25559911905610b4a88"
 
 [[package]]
-name = "wgpu-hal"
-version = "27.0.4"
+name = "whoami"
+version = "2.1.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5b21cb61c57ee198bc4aff71aeadff4cbb80b927beb912506af9c780d64313ce"
+checksum = "626c4bac6755d76ffc12cb01b2eac751db1996b9e0041de9aa02c8c211ddc82c"
 dependencies = [
- "android_system_properties",
- "arrayvec",
- "ash",
- "bit-set",
- "bitflags 2.11.0",
- "block",
- "bytemuck",
- "cfg-if",
- "cfg_aliases",
- "core-graphics-types 0.2.0",
- "glow",
- "glutin_wgl_sys",
- "gpu-alloc",
- "gpu-allocator",
- "gpu-descriptor",
- "hashbrown 0.16.1",
- "js-sys",
- "khronos-egl",
  "libc",
- "libloading",
- "log",
- "metal",
- "naga",
- "ndk-sys",
- "objc",
- "once_cell",
- "ordered-float",
- "parking_lot",
- "portable-atomic",
- "portable-atomic-util",
- "profiling",
- "range-alloc",
- "raw-window-handle",
- "renderdoc-sys",
- "smallvec",
- "thiserror 2.0.18",
- "wasm-bindgen",
- "web-sys",
- "wgpu-types",
- "windows 0.58.0",
- "windows-core 0.58.0",
-]
-
-[[package]]
-name = "wgpu-types"
-version = "27.0.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "afdcf84c395990db737f2dd91628706cb31e86d72e53482320d368e52b5da5eb"
-dependencies = [
- "bitflags 2.11.0",
- "bytemuck",
- "js-sys",
- "log",
- "thiserror 2.0.18",
+ "libredox",
+ "objc2-system-configuration",
+ "wasite",
  "web-sys",
 ]
 
@@ -8247,30 +6320,6 @@ version = "0.4.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "712e227841d057c1ee1cd2fb22fa7e5a5461ae8e48fa2ca79ec42cfc1931183f"
 
-[[package]]
-name = "window_clipboard"
-version = "0.5.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d5654226305eaf2dde8853fb482861d28e5dcecbbd40cb88e8393d94bb80d733"
-dependencies = [
- "clipboard-win",
- "clipboard_macos",
- "clipboard_wayland",
- "clipboard_x11",
- "raw-window-handle",
- "thiserror 2.0.18",
-]
-
-[[package]]
-name = "windows"
-version = "0.58.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "dd04d41d93c4992d421894c18c8b43496aa748dd4c081bac0dc93eb0489272b6"
-dependencies = [
- "windows-core 0.58.0",
- "windows-targets 0.52.6",
-]
-
 [[package]]
 name = "windows"
 version = "0.62.2"
@@ -8278,7 +6327,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "527fadee13e0c05939a6a05d5bd6eec6cd2e3dbd648b9f8e447c6518133d8580"
 dependencies = [
  "windows-collections",
- "windows-core 0.62.2",
+ "windows-core",
  "windows-future",
  "windows-numerics",
 ]
@@ -8289,20 +6338,7 @@ version = "0.3.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "23b2d95af1a8a14a3c7367e1ed4fc9c20e0a26e79551b1454d72583c97cc6610"
 dependencies = [
- "windows-core 0.62.2",
-]
-
-[[package]]
-name = "windows-core"
-version = "0.58.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6ba6d44ec8c2591c134257ce647b7ea6b20335bf6379a27dac5f1641fcf59f99"
-dependencies = [
- "windows-implement 0.58.0",
- "windows-interface 0.58.0",
- "windows-result 0.2.0",
- "windows-strings 0.1.0",
- "windows-targets 0.52.6",
+ "windows-core",
 ]
 
 [[package]]
@@ -8311,11 +6347,11 @@ version = "0.62.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "b8e83a14d34d0623b51dce9581199302a221863196a1dde71a7663a4c2be9deb"
 dependencies = [
- "windows-implement 0.60.2",
- "windows-interface 0.59.3",
+ "windows-implement",
+ "windows-interface",
  "windows-link",
- "windows-result 0.4.1",
- "windows-strings 0.5.1",
+ "windows-result",
+ "windows-strings",
 ]
 
 [[package]]
@@ -8324,22 +6360,11 @@ version = "0.3.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "e1d6f90251fe18a279739e78025bd6ddc52a7e22f921070ccdc67dde84c605cb"
 dependencies = [
- "windows-core 0.62.2",
+ "windows-core",
  "windows-link",
  "windows-threading",
 ]
 
-[[package]]
-name = "windows-implement"
-version = "0.58.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2bbd5b46c938e506ecbce286b6628a02171d56153ba733b6c741fc627ec9579b"
-dependencies = [
- "proc-macro2",
- "quote",
- "syn",
-]
-
 [[package]]
 name = "windows-implement"
 version = "0.60.2"
@@ -8348,18 +6373,7 @@ checksum = "053e2e040ab57b9dc951b72c264860db7eb3b0200ba345b4e4c3b14f67855ddf"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn",
-]
-
-[[package]]
-name = "windows-interface"
-version = "0.58.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "053c4c462dc91d3b1504c6fe5a726dd15e216ba718e84a0e46a88fbe5ded3515"
-dependencies = [
- "proc-macro2",
- "quote",
- "syn",
+ "syn 2.0.119",
 ]
 
 [[package]]
@@ -8370,7 +6384,7 @@ checksum = "3f316c4a2570ba26bbec722032c4099d8c8bc095efccdc15688708623367e358"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn",
+ "syn 2.0.119",
 ]
 
 [[package]]
@@ -8385,7 +6399,7 @@ version = "0.3.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "6e2e40844ac143cdb44aead537bbf727de9b044e107a0f1220392177d15b0f26"
 dependencies = [
- "windows-core 0.62.2",
+ "windows-core",
  "windows-link",
 ]
 
@@ -8396,17 +6410,8 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "02752bf7fbdcce7f2a27a742f798510f3e5ad88dbe84871e5168e2120c3d5720"
 dependencies = [
  "windows-link",
- "windows-result 0.4.1",
- "windows-strings 0.5.1",
-]
-
-[[package]]
-name = "windows-result"
-version = "0.2.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1d1043d8214f791817bab27572aaa8af63732e11bf84aa21a45a78d6c317ae0e"
-dependencies = [
- "windows-targets 0.52.6",
+ "windows-result",
+ "windows-strings",
 ]
 
 [[package]]
@@ -8418,16 +6423,6 @@ dependencies = [
  "windows-link",
 ]
 
-[[package]]
-name = "windows-strings"
-version = "0.1.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "4cd9b125c486025df0eabcb585e62173c6c9eddcec5d117d3b6e8c30e2ee4d10"
-dependencies = [
- "windows-result 0.2.0",
- "windows-targets 0.52.6",
-]
-
 [[package]]
 name = "windows-strings"
 version = "0.5.1"
@@ -8437,40 +6432,13 @@ dependencies = [
  "windows-link",
 ]
 
-[[package]]
-name = "windows-sys"
-version = "0.45.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "75283be5efb2831d37ea142365f009c02ec203cd29a3ebecbc093d52315b66d0"
-dependencies = [
- "windows-targets 0.42.2",
-]
-
 [[package]]
 name = "windows-sys"
 version = "0.52.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "282be5f36a8ce781fad8c8ae18fa3f9beff57ec1b52cb3de0789201425d9a33d"
 dependencies = [
- "windows-targets 0.52.6",
-]
-
-[[package]]
-name = "windows-sys"
-version = "0.59.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1e38bc4d79ed67fd075bcc251a1c39b32a1776bbe92e5bef1f0bf1f8c531853b"
-dependencies = [
- "windows-targets 0.52.6",
-]
-
-[[package]]
-name = "windows-sys"
-version = "0.60.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f2f500e4d28234f72040990ec9d39e3a6b950f9f22d3dba18416c35882612bcb"
-dependencies = [
- "windows-targets 0.53.5",
+ "windows-targets",
 ]
 
 [[package]]
@@ -8482,52 +6450,20 @@ dependencies = [
  "windows-link",
 ]
 
-[[package]]
-name = "windows-targets"
-version = "0.42.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8e5180c00cd44c9b1c88adb3693291f1cd93605ded80c250a75d472756b4d071"
-dependencies = [
- "windows_aarch64_gnullvm 0.42.2",
- "windows_aarch64_msvc 0.42.2",
- "windows_i686_gnu 0.42.2",
- "windows_i686_msvc 0.42.2",
- "windows_x86_64_gnu 0.42.2",
- "windows_x86_64_gnullvm 0.42.2",
- "windows_x86_64_msvc 0.42.2",
-]
-
 [[package]]
 name = "windows-targets"
 version = "0.52.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "9b724f72796e036ab90c1021d4780d4d3d648aca59e491e6b98e725b84e99973"
 dependencies = [
- "windows_aarch64_gnullvm 0.52.6",
- "windows_aarch64_msvc 0.52.6",
- "windows_i686_gnu 0.52.6",
- "windows_i686_gnullvm 0.52.6",
- "windows_i686_msvc 0.52.6",
- "windows_x86_64_gnu 0.52.6",
- "windows_x86_64_gnullvm 0.52.6",
- "windows_x86_64_msvc 0.52.6",
-]
-
-[[package]]
-name = "windows-targets"
-version = "0.53.5"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "4945f9f551b88e0d65f3db0bc25c33b8acea4d9e41163edf90dcd0b19f9069f3"
-dependencies = [
- "windows-link",
- "windows_aarch64_gnullvm 0.53.1",
- "windows_aarch64_msvc 0.53.1",
- "windows_i686_gnu 0.53.1",
- "windows_i686_gnullvm 0.53.1",
- "windows_i686_msvc 0.53.1",
- "windows_x86_64_gnu 0.53.1",
- "windows_x86_64_gnullvm 0.53.1",
- "windows_x86_64_msvc 0.53.1",
+ "windows_aarch64_gnullvm",
+ "windows_aarch64_msvc",
+ "windows_i686_gnu",
+ "windows_i686_gnullvm",
+ "windows_i686_msvc",
+ "windows_x86_64_gnu",
+ "windows_x86_64_gnullvm",
+ "windows_x86_64_msvc",
 ]
 
 [[package]]
@@ -8541,369 +6477,222 @@ dependencies = [
 
 [[package]]
 name = "windows_aarch64_gnullvm"
-version = "0.42.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "597a5118570b68bc08d8d59125332c54f1ba9d9adeedeef5b99b02ba2b0698f8"
-
-[[package]]
-name = "windows_aarch64_gnullvm"
-version = "0.52.6"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "32a4622180e7a0ec044bb555404c800bc9fd9ec262ec147edd5989ccd0c02cd3"
-
-[[package]]
-name = "windows_aarch64_gnullvm"
-version = "0.53.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a9d8416fa8b42f5c947f8482c43e7d89e73a173cead56d044f6a56104a6d1b53"
-
-[[package]]
-name = "windows_aarch64_msvc"
-version = "0.42.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e08e8864a60f06ef0d0ff4ba04124db8b0fb3be5776a5cd47641e942e58c4d43"
-
-[[package]]
-name = "windows_aarch64_msvc"
-version = "0.52.6"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "09ec2a7bb152e2252b53fa7803150007879548bc709c039df7627cabbd05d469"
-
-[[package]]
-name = "windows_aarch64_msvc"
-version = "0.53.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b9d782e804c2f632e395708e99a94275910eb9100b2114651e04744e9b125006"
-
-[[package]]
-name = "windows_i686_gnu"
-version = "0.42.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c61d927d8da41da96a81f029489353e68739737d3beca43145c8afec9a31a84f"
-
-[[package]]
-name = "windows_i686_gnu"
-version = "0.52.6"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8e9b5ad5ab802e97eb8e295ac6720e509ee4c243f69d781394014ebfe8bbfa0b"
-
-[[package]]
-name = "windows_i686_gnu"
-version = "0.53.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "960e6da069d81e09becb0ca57a65220ddff016ff2d6af6a223cf372a506593a3"
-
-[[package]]
-name = "windows_i686_gnullvm"
-version = "0.52.6"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0eee52d38c090b3caa76c563b86c3a4bd71ef1a819287c19d586d7334ae8ed66"
-
-[[package]]
-name = "windows_i686_gnullvm"
-version = "0.53.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "fa7359d10048f68ab8b09fa71c3daccfb0e9b559aed648a8f95469c27057180c"
-
-[[package]]
-name = "windows_i686_msvc"
-version = "0.42.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "44d840b6ec649f480a41c8d80f9c65108b92d89345dd94027bfe06ac444d1060"
-
-[[package]]
-name = "windows_i686_msvc"
-version = "0.52.6"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "240948bc05c5e7c6dabba28bf89d89ffce3e303022809e73deaefe4f6ec56c66"
-
-[[package]]
-name = "windows_i686_msvc"
-version = "0.53.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1e7ac75179f18232fe9c285163565a57ef8d3c89254a30685b57d83a38d326c2"
-
-[[package]]
-name = "windows_x86_64_gnu"
-version = "0.42.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8de912b8b8feb55c064867cf047dda097f92d51efad5b491dfb98f6bbb70cb36"
-
-[[package]]
-name = "windows_x86_64_gnu"
-version = "0.52.6"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "147a5c80aabfbf0c7d901cb5895d1de30ef2907eb21fbbab29ca94c5b08b1a78"
-
-[[package]]
-name = "windows_x86_64_gnu"
-version = "0.53.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9c3842cdd74a865a8066ab39c8a7a473c0778a3f29370b5fd6b4b9aa7df4a499"
-
-[[package]]
-name = "windows_x86_64_gnullvm"
-version = "0.42.2"
+version = "0.52.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "26d41b46a36d453748aedef1486d5c7a85db22e56aff34643984ea85514e94a3"
+checksum = "32a4622180e7a0ec044bb555404c800bc9fd9ec262ec147edd5989ccd0c02cd3"
 
 [[package]]
-name = "windows_x86_64_gnullvm"
+name = "windows_aarch64_msvc"
 version = "0.52.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "24d5b23dc417412679681396f2b49f3de8c1473deb516bd34410872eff51ed0d"
+checksum = "09ec2a7bb152e2252b53fa7803150007879548bc709c039df7627cabbd05d469"
 
 [[package]]
-name = "windows_x86_64_gnullvm"
-version = "0.53.1"
+name = "windows_i686_gnu"
+version = "0.52.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0ffa179e2d07eee8ad8f57493436566c7cc30ac536a3379fdf008f47f6bb7ae1"
+checksum = "8e9b5ad5ab802e97eb8e295ac6720e509ee4c243f69d781394014ebfe8bbfa0b"
 
 [[package]]
-name = "windows_x86_64_msvc"
-version = "0.42.2"
+name = "windows_i686_gnullvm"
+version = "0.52.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9aec5da331524158c6d1a4ac0ab1541149c0b9505fde06423b02f5ef0106b9f0"
+checksum = "0eee52d38c090b3caa76c563b86c3a4bd71ef1a819287c19d586d7334ae8ed66"
 
 [[package]]
-name = "windows_x86_64_msvc"
+name = "windows_i686_msvc"
 version = "0.52.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "589f6da84c646204747d1270a2a5661ea66ed1cced2631d546fdfb155959f9ec"
+checksum = "240948bc05c5e7c6dabba28bf89d89ffce3e303022809e73deaefe4f6ec56c66"
 
 [[package]]
-name = "windows_x86_64_msvc"
-version = "0.53.1"
+name = "windows_x86_64_gnu"
+version = "0.52.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d6bbff5f0aada427a1e5a6da5f1f98158182f26556f345ac9e04d36d0ebed650"
+checksum = "147a5c80aabfbf0c7d901cb5895d1de30ef2907eb21fbbab29ca94c5b08b1a78"
 
 [[package]]
-name = "winit"
-version = "0.30.13"
+name = "windows_x86_64_gnullvm"
+version = "0.52.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a6755fa58a9f8350bd1e472d4c3fcc25f824ec358933bba33306d0b63df5978d"
-dependencies = [
- "ahash",
- "android-activity",
- "atomic-waker",
- "bitflags 2.11.0",
- "block2 0.5.1",
- "bytemuck",
- "calloop 0.13.0",
- "cfg_aliases",
- "concurrent-queue",
- "core-foundation 0.9.4",
- "core-graphics",
- "cursor-icon",
- "dpi",
- "js-sys",
- "libc",
- "memmap2",
- "ndk",
- "objc2 0.5.2",
- "objc2-app-kit 0.2.2",
- "objc2-foundation 0.2.2",
- "objc2-ui-kit",
- "orbclient",
- "percent-encoding",
- "pin-project",
- "raw-window-handle",
- "redox_syscall 0.4.1",
- "rustix 0.38.44",
- "sctk-adwaita",
- "smithay-client-toolkit 0.19.2",
- "smol_str",
- "tracing",
- "unicode-segmentation",
- "wasm-bindgen",
- "wasm-bindgen-futures",
- "wayland-backend",
- "wayland-client",
- "wayland-protocols",
- "wayland-protocols-plasma",
- "web-sys",
- "web-time",
- "windows-sys 0.52.0",
- "x11-dl",
- "x11rb",
- "xkbcommon-dl",
-]
+checksum = "24d5b23dc417412679681396f2b49f3de8c1473deb516bd34410872eff51ed0d"
 
 [[package]]
-name = "winnow"
-version = "0.7.15"
+name = "windows_x86_64_msvc"
+version = "0.52.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "df79d97927682d2fd8adb29682d1140b343be4ac0f08fd68b7765d9c059d3945"
-dependencies = [
- "memchr",
-]
+checksum = "589f6da84c646204747d1270a2a5661ea66ed1cced2631d546fdfb155959f9ec"
 
 [[package]]
 name = "winnow"
-version = "1.0.1"
+version = "1.0.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "09dac053f1cd375980747450bfc7250c264eaae0583872e845c0c7cd578872b5"
+checksum = "23b97319f7b8343df12cc98938e5c3eb436064524c8d2b4e30a1d3a36eecdf81"
 dependencies = [
  "memchr",
 ]
 
 [[package]]
 name = "wit-bindgen"
-version = "0.51.0"
+version = "0.57.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d7249219f66ced02969388cf2bb044a09756a083d0fab1e566056b04d9fbcaa5"
-dependencies = [
- "wit-bindgen-rust-macro",
-]
+checksum = "1ebf944e87a7c253233ad6766e082e3cd714b5d03812acc24c318f549614536e"
 
 [[package]]
-name = "wit-bindgen-core"
-version = "0.51.0"
+name = "write16"
+version = "1.0.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ea61de684c3ea68cb082b7a88508a8b27fcc8b797d738bfc99a82facf1d752dc"
-dependencies = [
- "anyhow",
- "heck",
- "wit-parser",
-]
+checksum = "d1890f4022759daae28ed4fe62859b1236caebfc61ede2f63ed4e695f3f6d936"
 
 [[package]]
-name = "wit-bindgen-rust"
-version = "0.51.0"
+name = "writeable"
+version = "0.6.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b7c566e0f4b284dd6561c786d9cb0142da491f46a9fbed79ea69cdad5db17f21"
-dependencies = [
- "anyhow",
- "heck",
- "indexmap",
- "prettyplease",
- "syn",
- "wasm-metadata",
- "wit-bindgen-core",
- "wit-component",
-]
+checksum = "3ad82d2a33cdc9674dc7465672f271e096168fcdbe0f799d9e6db8c5892679dc"
 
 [[package]]
-name = "wit-bindgen-rust-macro"
-version = "0.51.0"
+name = "xet-client"
+version = "1.7.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0c0f9bfd77e6a48eccf51359e3ae77140a7f50b1e2ebfe62422d8afdaffab17a"
+checksum = "15500e5fc9e17be15224f9e34cb07d41fb2805d9dbbf7bd3507c982fd1eb3efd"
 dependencies = [
  "anyhow",
- "prettyplease",
- "proc-macro2",
- "quote",
- "syn",
- "wit-bindgen-core",
- "wit-bindgen-rust",
+ "async-trait",
+ "base64 0.22.1",
+ "bytes",
+ "chrono",
+ "crc32fast",
+ "futures",
+ "http 1.5.0",
+ "hyper",
+ "more-asserts",
+ "rand 0.10.3",
+ "redb",
+ "reqwest",
+ "reqwest-middleware",
+ "serde",
+ "serde_json",
+ "serde_repr",
+ "statrs",
+ "tempfile",
+ "thiserror",
+ "tokio",
+ "tokio-retry",
+ "tokio_with_wasm",
+ "tracing",
+ "url",
+ "urlencoding",
+ "uuid",
+ "web-time",
+ "xet-core-structures",
+ "xet-runtime",
 ]
 
 [[package]]
-name = "wit-component"
-version = "0.244.0"
+name = "xet-core-structures"
+version = "1.7.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9d66ea20e9553b30172b5e831994e35fbde2d165325bec84fc43dbf6f4eb9cb2"
+checksum = "5080d5803c8bf15fa56ef934050a46b225801bad3ea4cb8bb3c46c5c37976516"
 dependencies = [
- "anyhow",
- "bitflags 2.11.0",
- "indexmap",
- "log",
+ "async-trait",
+ "base64 0.22.1",
+ "blake3",
+ "bytemuck",
+ "bytes",
+ "countio",
+ "futures",
+ "futures-util",
+ "getrandom 0.4.3",
+ "heapify",
+ "itertools 0.14.0",
+ "lz4_flex",
+ "more-asserts",
+ "rand 0.10.3",
+ "regex",
+ "safe-transmute",
  "serde",
- "serde_derive",
- "serde_json",
- "wasm-encoder",
- "wasm-metadata",
- "wasmparser",
- "wit-parser",
+ "static_assertions",
+ "thiserror",
+ "tokio",
+ "tokio-util",
+ "tracing",
+ "uuid",
+ "web-time",
+ "xet-runtime",
 ]
 
 [[package]]
-name = "wit-parser"
-version = "0.244.0"
+name = "xet-data"
+version = "1.7.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ecc8ac4bc1dc3381b7f59c34f00b67e18f910c2c0f50015669dde7def656a736"
+checksum = "1287a1657fb67de846030f6605a9960dcdad401c467b490657e6e909d2e0d1c0"
 dependencies = [
  "anyhow",
- "id-arena",
- "indexmap",
- "log",
- "semver",
+ "async-trait",
+ "bytes",
+ "chrono",
+ "gearhash",
+ "http 1.5.0",
+ "itertools 0.14.0",
+ "more-asserts",
+ "rand 0.10.3",
  "serde",
- "serde_derive",
  "serde_json",
- "unicode-xid",
- "wasmparser",
-]
-
-[[package]]
-name = "writeable"
-version = "0.6.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1ffae5123b2d3fc086436f8834ae3ab053a283cfac8fe0a0b8eaae044768a4c4"
-
-[[package]]
-name = "x11-dl"
-version = "2.21.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "38735924fedd5314a6e548792904ed8c6de6636285cb9fec04d5b1db85c1516f"
-dependencies = [
- "libc",
- "once_cell",
- "pkg-config",
+ "sha2",
+ "tempfile",
+ "thiserror",
+ "tokio",
+ "tokio-util",
+ "tokio_with_wasm",
+ "tracing",
+ "url",
+ "uuid",
+ "web-time",
+ "xet-client",
+ "xet-core-structures",
+ "xet-runtime",
 ]
 
 [[package]]
-name = "x11rb"
-version = "0.13.2"
+name = "xet-runtime"
+version = "1.7.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9993aa5be5a26815fe2c3eacfc1fde061fc1a1f094bf1ad2a18bf9c495dd7414"
+checksum = "921953c885ce83e02b21e8d26444b75c78c9e83076699ced08d5c6fc1cdb48ac"
 dependencies = [
- "as-raw-xcb-connection",
- "gethostname",
+ "anyhow",
+ "async-trait",
+ "bytes",
+ "chrono",
+ "colored",
+ "const-str",
+ "ctor",
+ "dirs",
+ "futures",
+ "git-version",
+ "humantime",
+ "konst",
  "libc",
- "libloading",
- "once_cell",
- "rustix 1.1.4",
- "x11rb-protocol",
-]
-
-[[package]]
-name = "x11rb-protocol"
-version = "0.13.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ea6fc2961e4ef194dcbfe56bb845534d0dc8098940c7e5c012a258bfec6701bd"
-
-[[package]]
-name = "xcursor"
-version = "0.3.10"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "bec9e4a500ca8864c5b47b8b482a73d62e4237670e5b5f1d6b9e3cae50f28f2b"
-
-[[package]]
-name = "xkbcommon-dl"
-version = "0.4.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d039de8032a9a8856a6be89cea3e5d12fdd82306ab7c94d74e6deab2460651c5"
-dependencies = [
- "bitflags 2.11.0",
- "dlib",
- "log",
- "once_cell",
- "xkeysym",
+ "more-asserts",
+ "oneshot",
+ "pin-project",
+ "rand 0.10.3",
+ "reqwest",
+ "serde",
+ "serde_json",
+ "shellexpand",
+ "sysinfo",
+ "thiserror",
+ "tokio",
+ "tokio-util",
+ "tokio_with_wasm",
+ "tracing",
+ "tracing-appender",
+ "tracing-subscriber",
+ "web-time",
+ "whoami",
+ "winapi",
 ]
 
-[[package]]
-name = "xkeysym"
-version = "0.2.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b9cc00251562a284751c9973bace760d86c0276c471b4be569fe6b068ee97a56"
-
-[[package]]
-name = "xml-rs"
-version = "0.8.28"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3ae8337f8a065cfc972643663ea4279e04e7256de865aa66fe25cec5fb912d3f"
-
 [[package]]
 name = "xmlwriter"
 version = "0.1.0"
@@ -8918,32 +6707,20 @@ checksum = "7a5a4b21e1a62b67a2970e6831bc091d7b87e119e7f9791aef9702e3bef04448"
 
 [[package]]
 name = "yaml-rust2"
-version = "0.11.0"
+version = "0.13.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "631a50d867fafb7093e709d75aaee9e0e0d5deb934021fcea25ac2fe09edc51e"
+checksum = "57e5b818a27a4cd30884ea380857a5e56f7ec3ba24a3990a3cc0b95af3238e18"
 dependencies = [
  "arraydeque",
  "encoding_rs",
  "hashlink",
 ]
 
-[[package]]
-name = "yansi"
-version = "1.0.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "cfe53a6657fd280eaa890a3bc59152892ffa3e30101319d168b781ed6529b049"
-
-[[package]]
-name = "yazi"
-version = "0.2.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e01738255b5a16e78bbb83e7fbba0a1e7dd506905cfc53f4622d89015a03fbb5"
-
 [[package]]
 name = "yoke"
-version = "0.8.2"
+version = "0.8.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "abe8c5fda708d9ca3df187cae8bfb9ceda00dd96231bed36e445a1a48e66f9ca"
+checksum = "417ef3fcf6c300b7dbd25c85531a746931c02ad536f3cf4786ed1e88223340b3"
 dependencies = [
  "stable_deref_trait",
  "yoke-derive",
@@ -8952,135 +6729,68 @@ dependencies = [
 
 [[package]]
 name = "yoke-derive"
-version = "0.8.2"
+version = "0.8.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "de844c262c8848816172cef550288e7dc6c7b7814b4ee56b3e1553f275f1858e"
+checksum = "ec8ebde2db3681e8c9980cc27822030e68752690ddfa9473e739aeb4dbde6d71"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn",
+ "syn 3.0.9",
  "synstructure",
 ]
 
-[[package]]
-name = "zbus"
-version = "5.14.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ca82f95dbd3943a40a53cfded6c2d0a2ca26192011846a1810c4256ef92c60bc"
-dependencies = [
- "async-broadcast",
- "async-executor",
- "async-io",
- "async-lock",
- "async-process",
- "async-recursion",
- "async-task",
- "async-trait",
- "blocking",
- "enumflags2",
- "event-listener",
- "futures-core",
- "futures-lite",
- "hex",
- "libc",
- "ordered-stream",
- "rustix 1.1.4",
- "serde",
- "serde_repr",
- "tracing",
- "uds_windows",
- "uuid",
- "windows-sys 0.61.2",
- "winnow 0.7.15",
- "zbus_macros",
- "zbus_names",
- "zvariant",
-]
-
-[[package]]
-name = "zbus_macros"
-version = "5.14.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "897e79616e84aac4b2c46e9132a4f63b93105d54fe8c0e8f6bffc21fa8d49222"
-dependencies = [
- "proc-macro-crate",
- "proc-macro2",
- "quote",
- "syn",
- "zbus_names",
- "zvariant",
- "zvariant_utils",
-]
-
-[[package]]
-name = "zbus_names"
-version = "4.3.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ffd8af6d5b78619bab301ff3c560a5bd22426150253db278f164d6cf3b72c50f"
-dependencies = [
- "serde",
- "winnow 0.7.15",
- "zvariant",
-]
-
-[[package]]
-name = "zeno"
-version = "0.3.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6df3dc4292935e51816d896edcd52aa30bc297907c26167fec31e2b0c6a32524"
-
 [[package]]
 name = "zerocopy"
-version = "0.8.48"
+version = "0.8.63"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "eed437bf9d6692032087e337407a86f04cd8d6a16a37199ed57949d415bd68e9"
+checksum = "e5fe1f8f1b06191a00962174c61aa5005e0bb391a6d80d07e24d115c01a92ed8"
 dependencies = [
  "zerocopy-derive",
 ]
 
 [[package]]
 name = "zerocopy-derive"
-version = "0.8.48"
+version = "0.8.63"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "70e3cd084b1788766f53af483dd21f93881ff30d7320490ec3ef7526d203bad4"
+checksum = "863ad3ac83293fb4d740aedbfdc9240dd8d1a50c1099acd76ce80ce7c7230c7f"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn",
+ "syn 2.0.119",
 ]
 
 [[package]]
 name = "zerofrom"
-version = "0.1.7"
+version = "0.1.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "69faa1f2a1ea75661980b013019ed6687ed0e83d069bc1114e2cc74c6c04c4df"
+checksum = "0ec05a11813ea801ff6d75110ad09cd0824ddba17dfe17128ea0d5f68e6c5272"
 dependencies = [
  "zerofrom-derive",
 ]
 
 [[package]]
 name = "zerofrom-derive"
-version = "0.1.7"
+version = "0.1.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "11532158c46691caf0f2593ea8358fed6bbf68a0315e80aae9bd41fbade684a1"
+checksum = "f75b4683f6c7f45248d4d64056a24298c6281e0993356d7d1b4a1a962ef10d4a"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn",
+ "syn 3.0.9",
  "synstructure",
 ]
 
 [[package]]
 name = "zeroize"
-version = "1.8.2"
+version = "1.9.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b97154e67e32c85465826e8bcc1c59429aaaf107c1e4a9e53c8d8ccd5eff88d0"
+checksum = "e13084392c5e4bc371903e2935a5eaeed24905a7511356b883835e18a78f6879"
 
 [[package]]
 name = "zerotrie"
-version = "0.2.4"
+version = "0.2.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0f9152d31db0792fa83f70fb2f83148effb5c1f5b8c7686c3459e361d9bc20bf"
+checksum = "4ea269c3bd32f0a32c321907a2ae912ba6f4649bb0fc764a15627e99a7095a3f"
 dependencies = [
  "displaydoc",
  "yoke",
@@ -9089,9 +6799,9 @@ dependencies = [
 
 [[package]]
 name = "zerovec"
-version = "0.11.6"
+version = "0.11.9"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "90f911cbc359ab6af17377d242225f4d75119aec87ea711a880987b18cd7b239"
+checksum = "f58a2799cfe9a1670c7ea8edbd52ef8edb93cb37a7f83bf31ba9e35d9dae6a40"
 dependencies = [
  "yoke",
  "zerofrom",
@@ -9100,44 +6810,50 @@ dependencies = [
 
 [[package]]
 name = "zerovec-derive"
-version = "0.11.3"
+version = "0.11.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "625dc425cab0dca6dc3c3319506e6593dcb08a9f387ea3b284dbd52a92c40555"
+checksum = "c3743ddbf47738d929cd1ae2e199d40d0162979dfe31aef8ae646d2bf00bf97a"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn",
+ "syn 3.0.9",
 ]
 
+[[package]]
+name = "zlib-rs"
+version = "0.6.8"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b268e58e7c693d7c271f93ffc4ba3b380412554231c85bf61ca7af91042a4112"
+
 [[package]]
 name = "zmij"
-version = "1.0.21"
+version = "1.0.23"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b8848ee67ecc8aedbaf3e4122217aff892639231befc6a1b58d29fff4c2cabaa"
+checksum = "29666d0abbfad1e3dc4dcf6144730dd3a3ab225bbbdac83319345b1b44ccfc1b"
 
 [[package]]
 name = "zstd"
-version = "0.13.3"
+version = "0.14.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e91ee311a569c327171651566e07972200e76fcfe2242a4fa446149a3881c08a"
+checksum = "db3ae9bf90af6e40197bef767eeefc46ad6397420fe873830bd85e5fbbfb3c70"
 dependencies = [
  "zstd-safe",
 ]
 
 [[package]]
 name = "zstd-safe"
-version = "7.2.4"
+version = "8.2.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8f49c4d5f0abb602a93fb8736af2a4f4dd9512e36f7f570d66e65ff867ed3b9d"
+checksum = "ded520f16b29e1a9b771c722221b2b0a0398c2316d9bb3c42b458a5414d7434b"
 dependencies = [
  "zstd-sys",
 ]
 
 [[package]]
 name = "zstd-sys"
-version = "2.0.16+zstd.1.5.7"
+version = "2.1.2+zstd.1.5.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "91e19ebc2adc8f83e43039e79776e3fda8ca919132d68a1fed6a5faca2683748"
+checksum = "481ada9c543b2323cbb7bbdbb5249177fbf0188ed959ed2c15e4b22421c84ed3"
 dependencies = [
  "cc",
  "pkg-config",
@@ -9145,15 +6861,9 @@ dependencies = [
 
 [[package]]
 name = "zune-core"
-version = "0.4.12"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3f423a2c17029964870cfaabb1f13dfab7d092a62a29a89264f4d36990ca414a"
-
-[[package]]
-name = "zune-core"
-version = "0.5.1"
+version = "0.5.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "cb8a0807f7c01457d0379ba880ba6322660448ddebc890ce29bb64da71fb40f9"
+checksum = "d56377fd46368984a170bc5aac5567e52ca5da874caa60bea39fcbca78fb658b"
 
 [[package]]
 name = "zune-inflate"
@@ -9164,60 +6874,11 @@ dependencies = [
  "simd-adler32",
 ]
 
-[[package]]
-name = "zune-jpeg"
-version = "0.4.21"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "29ce2c8a9384ad323cf564b67da86e21d3cfdff87908bc1223ed5c99bc792713"
-dependencies = [
- "zune-core 0.4.12",
-]
-
 [[package]]
 name = "zune-jpeg"
 version = "0.5.15"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "27bc9d5b815bc103f142aa054f561d9187d191692ec7c2d1e2b4737f8dbd7296"
 dependencies = [
- "zune-core 0.5.1",
-]
-
-[[package]]
-name = "zvariant"
-version = "5.10.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5708299b21903bbe348e94729f22c49c55d04720a004aa350f1f9c122fd2540b"
-dependencies = [
- "endi",
- "enumflags2",
- "serde",
- "winnow 0.7.15",
- "zvariant_derive",
- "zvariant_utils",
-]
-
-[[package]]
-name = "zvariant_derive"
-version = "5.10.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5b59b012ebe9c46656f9cc08d8da8b4c726510aef12559da3e5f1bf72780752c"
-dependencies = [
- "proc-macro-crate",
- "proc-macro2",
- "quote",
- "syn",
- "zvariant_utils",
-]
-
-[[package]]
-name = "zvariant_utils"
-version = "3.3.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f75c23a64ef8f40f13a6989991e643554d9bef1d682a281160cf0c1bc389c5e9"
-dependencies = [
- "proc-macro2",
- "quote",
- "serde",
- "syn",
- "winnow 0.7.15",
+ "zune-core",
 ]
diff --git a/Cargo.toml b/Cargo.toml
index 7673c05c5..babd81b76 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -1,28 +1,40 @@
 [workspace]
 members = [
     "paddler_agent",
+    "paddler_agent_decision",
+    "paddler_agent_embeddings",
+    "paddler_agent_pointer_head",
+    "paddler_agent_runtime",
     "paddler_agent_status",
+    "paddler_agent_text_generation",
     "paddler_balancer",
+    "paddler_blocking_thread_fixture",
     "paddler_bootstrap",
     "paddler_cache_dir",
     "paddler_cli",
     "paddler_cli_tests",
     "paddler_client",
+    "paddler_dashboard_assets",
     "paddler_download_manager",
-    "paddler_gui",
-    "paddler_gui_tests",
+    "paddler_frame_writer",
     "paddler_image_decoder",
     "paddler_inference_parameters",
     "paddler_local_http_fixture",
     "paddler_messaging",
     "paddler_model_source",
     "paddler_openai_response_format_validator",
+    "paddler_openai_translation",
     "paddler_opencode_tests",
+    "paddler_outbox",
     "paddler_request_registry",
+    "paddler_sha256_hex",
+    "paddler_spawn_resuming_panic",
     "paddler_state_database",
+    "paddler_tcp_buffer_limits",
     "paddler_test_cluster_harness",
     "paddler_tests",
     "paddler_tool_call_validator",
+    "paddler_typesafe_translation",
 ]
 resolver = "2"
 
@@ -33,94 +45,105 @@ edition = "2024"
 homepage = "https://paddler.intentee.com/"
 license = "Apache-2.0"
 repository = "https://github.com/intentee/paddler"
-version = "4.1.0"
+version = "5.0.0"
 
 [workspace.dependencies]
 actix-cors = "0.7"
-actix-rt = "2.11"
-actix-web = "4.13"
-actix-web-lab = "0.26"
+actix-web = { version = "4.18", default-features = false, features = ["compat", "compress-brotli", "compress-gzip", "compress-zstd", "cookies", "macros", "unicode", "ws"] }
+actix-web-lab = "0.29"
 actix-ws = "0.4"
 anyhow = { version = "1.0", features = ["backtrace"] }
-askama = "0.14"
-async-openai = { version = "0.40", features = ["byot", "chat-completion", "responses"] }
-async-stream = "0.3"
+askama = "0.16"
+async-openai = { version = "0.42", features = ["byot", "chat-completion", "responses"] }
 async-trait = "0.1"
-bytes = "1.11"
-cadence = "1.6"
-clap = { version = "4.5", features = ["derive", "string"] }
-command_handler = "0.1"
-dashmap = "6.1"
+base64 = "0.23"
+bytes = "1.12"
+cadence = "1.8"
+clap = { version = "4.6", features = ["derive", "env", "string"] }
+command_handler = "0.2"
+dashmap = "6.2"
 data-url = "0.3"
 encoding_rs = { version = "0.8", features = ["serde"] }
 env_logger = "0.11"
-esbuild-metafile = "0.5"
+esbuild-metafile = "0.11"
 fslock = "0.2"
 futures = "0.3"
-futures-util = { version = "0.3", features = ["tokio-io"] }
+futures-util = "0.3"
 headers = "0.4"
-hf-hub = { version = "0.4", features = ["tokio"] }
-http = "1.4"
+hf-hub = "1.0"
+http = "1.5"
 httparse = "1.10"
+icu_normalizer = "2.4"
 image = "0.25"
 indoc = "2.0"
-jsonschema = { version = "0.37", default-features = false }
-llama-cpp-bindings = "0.16"
-llama-cpp-bindings-sys = "0.16"
-llama-cpp-bindings-types = "0.16"
-base64 = "0.22"
+jsonschema = { version = "0.60", default-features = false }
+libc = "0.2"
+llama-cpp-bindings = "0.25"
+llama-cpp-bindings-sys = "0.25"
+llama-cpp-bindings-types = "0.25"
 log = "0.4"
 mime_guess = "2.0"
-minijinja = { version = "2.12", features = ["builtins", "json", "loader"] }
-minijinja-contrib = { version = "2.12", features = ["datetime", "pycompat", "wordcount", "wordwrap"] }
-nanoid = "0.4"
-nix = { version = "0.30", features = ["resource", "signal"] }
-open = "5.3"
-parking_lot = "0.12"
-quick-xml = "0.39"
-rand = "0.9"
-reqwest = { version = "0.12", features = ["json", "stream"] }
-resvg = "0.46"
-rust-embed = { version = "8.9", features = ["interpolate-folder-path"] }
+minijinja = { version = "3.0", features = ["builtins", "json", "serde"] }
+minijinja-contrib = { version = "3.0", features = ["datetime", "pycompat", "wordcount", "wordwrap"] }
+nanoid = "0.5"
+nix = { version = "0.31", features = ["fs", "net", "process", "resource", "signal", "socket", "user"] }
+parking_lot = "0.13"
+percent-encoding = "2.3"
+quick-xml = "0.42"
+rand = "0.10"
+reqwest = { version = "0.13", features = ["json", "stream"] }
+resvg = "0.49"
+rust-embed = { version = "8.13", features = ["debug-embed", "interpolate-folder-path"] }
 serde = { version = "1.0", features = ["derive"] }
-serde_json = { version = "1.0", features = ["preserve_order"] }
-sha2 = "0.10"
+serde_json = { version = "1.0", features = ["preserve_order", "raw_value"] }
+seccompiler = "0.5"
+sha2 = "0.11"
 shellexpand = "3.1"
 socket2 = "0.6"
-iced = { version = "0.14", features = ["image", "svg", "tokio"] }
-iced_test = { version = "0.14" }
-if-addrs = "0.13"
-statum = "0.6"
-tempfile = "3.20"
-tokio = { version = "1.48", features = ["full"] }
+syn = { version = "3.0", features = ["full", "visit"] }
+tempfile = "3.27"
+tokio = { version = "1.54", features = ["full"] }
 tokio-stream = { version = "0.1", features = ["sync"] }
 tokio-test = "0.4"
-tokio-tungstenite = "0.28"
+tokio-tungstenite = { version = "0.30", features = ["rustls-tls-native-roots"] }
 tokio-util = { version = "0.7", features = ["rt"] }
 thiserror = "2.0"
-trzcina = "0.3"
+trzcina = "0.4"
 url = { version = "2.5", features = ["serde"] }
-yaml-rust2 = "0.11"
-paddler_agent = { version = "4.1", path = "paddler_agent" }
-paddler_agent_status = { version = "4.1", path = "paddler_agent_status" }
-paddler_balancer = { version = "4.1", path = "paddler_balancer" }
-paddler_bootstrap = { version = "4.1", path = "paddler_bootstrap" }
-paddler_cache_dir = { version = "4.1", path = "paddler_cache_dir" }
-paddler_cli = { version = "4.1", path = "paddler_cli" }
-paddler_client = { version = "4.1", path = "paddler_client" }
-paddler_download_manager = { version = "4.1", path = "paddler_download_manager" }
-paddler_gui = { version = "4.1", path = "paddler_gui" }
-paddler_image_decoder = { version = "4.1", path = "paddler_image_decoder" }
-paddler_inference_parameters = { version = "4.1", path = "paddler_inference_parameters" }
-paddler_local_http_fixture = { version = "4.1", path = "paddler_local_http_fixture" }
-paddler_messaging = { version = "4.1", path = "paddler_messaging" }
-paddler_model_source = { version = "4.1", path = "paddler_model_source" }
-paddler_openai_response_format_validator = { version = "4.1", path = "paddler_openai_response_format_validator" }
-paddler_request_registry = { version = "4.1", path = "paddler_request_registry" }
-paddler_state_database = { version = "4.1", path = "paddler_state_database" }
-paddler_test_cluster_harness = { version = "4.1", path = "paddler_test_cluster_harness" }
-paddler_tests = { version = "4.1", path = "paddler_tests" }
-paddler_tool_call_validator = { version = "4.1", path = "paddler_tool_call_validator" }
+yaml-rust2 = "0.13"
+paddler_agent = { version = "5.0", path = "paddler_agent" }
+paddler_agent_decision = { version = "5.0", path = "paddler_agent_decision" }
+paddler_agent_embeddings = { version = "5.0", path = "paddler_agent_embeddings" }
+paddler_agent_pointer_head = { version = "5.0", path = "paddler_agent_pointer_head" }
+paddler_agent_runtime = { version = "5.0", path = "paddler_agent_runtime" }
+paddler_agent_status = { version = "5.0", path = "paddler_agent_status" }
+paddler_agent_text_generation = { version = "5.0", path = "paddler_agent_text_generation" }
+paddler_balancer = { version = "5.0", path = "paddler_balancer" }
+paddler_blocking_thread_fixture = { version = "5.0", path = "paddler_blocking_thread_fixture" }
+paddler_bootstrap = { version = "5.0", path = "paddler_bootstrap" }
+paddler_cache_dir = { version = "5.0", path = "paddler_cache_dir" }
+paddler_cli = { version = "5.0", path = "paddler_cli" }
+paddler_client = { version = "5.0", path = "paddler_client" }
+paddler_dashboard_assets = { version = "5.0", path = "paddler_dashboard_assets" }
+paddler_download_manager = { version = "5.0", path = "paddler_download_manager" }
+paddler_frame_writer = { version = "5.0", path = "paddler_frame_writer" }
+paddler_image_decoder = { version = "5.0", path = "paddler_image_decoder" }
+paddler_inference_parameters = { version = "5.0", path = "paddler_inference_parameters" }
+paddler_local_http_fixture = { version = "5.0", path = "paddler_local_http_fixture" }
+paddler_messaging = { version = "5.0", path = "paddler_messaging" }
+paddler_model_source = { version = "5.0", path = "paddler_model_source" }
+paddler_openai_response_format_validator = { version = "5.0", path = "paddler_openai_response_format_validator" }
+paddler_openai_translation = { version = "5.0", path = "paddler_openai_translation" }
+paddler_outbox = { version = "5.0", path = "paddler_outbox" }
+paddler_request_registry = { version = "5.0", path = "paddler_request_registry" }
+paddler_sha256_hex = { version = "5.0", path = "paddler_sha256_hex" }
+paddler_spawn_resuming_panic = { version = "5.0", path = "paddler_spawn_resuming_panic" }
+paddler_state_database = { version = "5.0", path = "paddler_state_database" }
+paddler_tcp_buffer_limits = { version = "5.0", path = "paddler_tcp_buffer_limits" }
+paddler_test_cluster_harness = { version = "5.0", path = "paddler_test_cluster_harness" }
+paddler_tests = { version = "5.0", path = "paddler_tests" }
+paddler_tool_call_validator = { version = "5.0", path = "paddler_tool_call_validator" }
+paddler_typesafe_translation = { version = "5.0", path = "paddler_typesafe_translation" }
 
 [profile.release]
 lto = true
diff --git a/Dockerfile b/Dockerfile
index 7c2256c7a..9ebe461be 100644
--- a/Dockerfile
+++ b/Dockerfile
@@ -1,6 +1,6 @@
 FROM node:22-bookworm@sha256:363e1587494626837fa7f9a23bdb453d13b0ff3c67c705c2805cfc69c2d2fad7 AS node
 
-FROM rust:1.95.0-bookworm@sha256:6258907abe69656e41cd992e0b705cdcfabcbbe3db374f92ed2d47121282d4a1
+FROM rust:1.99.1-bookworm@sha256:e12566d5abc17a0da741628782cdb504b68b8961f5bfcf677936bf3e8b32de29
 
 COPY --from=node /usr/local/bin/node /usr/local/bin/node
 COPY --from=node /usr/local/lib/node_modules /usr/local/lib/node_modules
diff --git a/Makefile b/Makefile
index cd1a8c7e0..27662bf54 100644
--- a/Makefile
+++ b/Makefile
@@ -1,21 +1,40 @@
 .DEFAULT_GOAL := target/release/paddler
 
+.PRECIOUS: target/kev_%/backbone/config.json
+
+.NOTPARALLEL:
+
 RUST_LOG ?= debug
 
-PADDLER_CRATES := paddler_agent paddler_agent_status paddler_balancer paddler_bootstrap paddler_cache_dir paddler_cli paddler_client paddler_download_manager paddler_gui paddler_image_decoder paddler_inference_parameters paddler_messaging paddler_model_source paddler_request_registry paddler_state_database paddler_tool_call_validator
-PADDLER_SOURCES := $(shell find $(addsuffix /src,$(PADDLER_CRATES)) -name '*.rs') $(addsuffix /Cargo.toml,$(PADDLER_CRATES)) Cargo.toml Cargo.lock
+PADDLER_CRATES := paddler_agent paddler_agent_decision paddler_agent_embeddings paddler_agent_pointer_head paddler_agent_runtime paddler_agent_status paddler_agent_text_generation paddler_balancer paddler_bootstrap paddler_cache_dir paddler_cli paddler_client paddler_dashboard_assets paddler_download_manager paddler_frame_writer paddler_image_decoder paddler_inference_parameters paddler_messaging paddler_model_source paddler_openai_translation paddler_outbox paddler_request_registry paddler_sha256_hex paddler_spawn_resuming_panic paddler_state_database paddler_tool_call_validator paddler_typesafe_translation
+PADDLER_SOURCES := $(shell find $(addsuffix /src,$(PADDLER_CRATES)) -name '*.rs') $(wildcard $(addsuffix /build.rs,$(PADDLER_CRATES))) $(addsuffix /Cargo.toml,$(PADDLER_CRATES)) Cargo.toml Cargo.lock
 FRONTEND_SOURCES := $(shell find resources paddler_client_javascript/src -type f) paddler_client_javascript/package.json $(wildcard jarmuz/*.mjs)
 JS_CLIENT_SOURCES := $(shell find paddler_client_javascript/src -type f) paddler_client_javascript/package.json paddler_client_javascript/tsconfig.json
+KEV_CHECKPOINT_0_8b := jaredpalmer/kev-0.8b@788ddbdd65715bb03a56788c822f6c632c9a551d
+KEV_CHECKPOINT_27b := jaredpalmer/kev-27b@af0e6d551bdc2cc724f3e9d7a8bee1cd4fb8f7bf
+KEV_GGUF_OUTTYPE_0_8b := bf16
+KEV_GGUF_OUTTYPE_27b := q8_0
+KEV_CONVERTER_SOURCES := $(shell find paddler_kev_converter_python/paddler_kev_converter -name '*.py')
+QWEN3_5_BASE := Qwen/Qwen3.5-0.8B-Base
+QWEN3_5_BASE_REVISION := dc7cdfe2ee4154fa7e30f5b51ca41bfa40174e68
+QWEN3_8_BASE := Qwen/Qwen3.8-27B
+QWEN3_8_BASE_REVISION := 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0
+LLAMA_CPP_SOURCE_DIRECTORY = $(dir $(shell cargo metadata --format-version 1 | jq -r '.packages[] | select(.name == "llama-cpp-bindings-sys") | .manifest_path'))llama.cpp
+RUN_WITH_TEST_CLUSTER_SOURCES := $(shell find paddler_tests/src paddler_test_cluster_harness/src -name '*.rs') paddler_tests/Cargo.toml paddler_test_cluster_harness/Cargo.toml
 LLVM_COV_THIRD_PARTY_SOURCES := /\.cargo/(registry|git)/|/\.rustup/toolchains/|^/rustc/|^/nix/store/|^$(CURDIR)/target/
 
 TEST_DEVICE ?= cpu
 
 ifeq ($(TEST_DEVICE),cpu)
+NEXTEST_PROFILE := cpu
 PADDLER_TEST_BINARY := target/debug/paddler
+RUN_WITH_TEST_CLUSTER_BINARY := target/debug/run_with_test_cluster
 TEST_DEVICE_FEATURE_SUFFIX :=
 TEST_DEVICE_TARGET_DIR :=
 else
+NEXTEST_PROFILE := default
 PADDLER_TEST_BINARY := target/$(TEST_DEVICE)/debug/paddler
+RUN_WITH_TEST_CLUSTER_BINARY := target/$(TEST_DEVICE)/debug/run_with_test_cluster
 TEST_DEVICE_FEATURE_SUFFIX := ,$(TEST_DEVICE)
 TEST_DEVICE_TARGET_DIR := --target-dir target/$(TEST_DEVICE)
 endif
@@ -25,6 +44,48 @@ PADDLER_BINARY_ENVIRONMENT := PADDLER_BINARY=$(CURDIR)/$(PADDLER_TEST_BINARY)
 esbuild-meta.json: $(FRONTEND_SOURCES) jarmuz-static.mjs tsconfig.json package.json node_modules
 	./jarmuz-static.mjs
 
+fixtures/kev_%_parity_reference.json: $(KEV_CONVERTER_SOURCES) paddler_kev_converter_python/.venv
+	poetry -C paddler_kev_converter_python run paddler-kev-converter parity-reference --checkpoint $(KEV_CHECKPOINT_$*) --output $(CURDIR)/$@
+
+fixtures/pointer_head_with_f16_tensors.gguf: $(KEV_CONVERTER_SOURCES) paddler_kev_converter_python/.venv
+	poetry -C paddler_kev_converter_python run paddler-kev-converter malformed-pointer-head --input-normalization NFC --defect f16-tensors --hidden-size 64 --output $(CURDIR)/$@
+
+fixtures/pointer_head_with_mismatched_projections.gguf: $(KEV_CONVERTER_SOURCES) paddler_kev_converter_python/.venv
+	poetry -C paddler_kev_converter_python run paddler-kev-converter malformed-pointer-head --input-normalization NFC --defect mismatched-projections --hidden-size 64 --output $(CURDIR)/$@
+
+fixtures/pointer_head_with_zero_temperature.gguf: $(KEV_CONVERTER_SOURCES) paddler_kev_converter_python/.venv
+	poetry -C paddler_kev_converter_python run paddler-kev-converter malformed-pointer-head --input-normalization NFC --defect zero-temperature --hidden-size 64 --output $(CURDIR)/$@
+
+fixtures/pointer_head_without_delimiters.gguf: $(KEV_CONVERTER_SOURCES) paddler_kev_converter_python/.venv
+	poetry -C paddler_kev_converter_python run paddler-kev-converter malformed-pointer-head --input-normalization NFC --defect missing-delimiters --hidden-size 64 --output $(CURDIR)/$@
+
+fixtures/pointer_head_without_input_normalization.gguf: $(KEV_CONVERTER_SOURCES) paddler_kev_converter_python/.venv
+	poetry -C paddler_kev_converter_python run paddler-kev-converter malformed-pointer-head --input-normalization NFC --defect missing-input-normalization --hidden-size 64 --output $(CURDIR)/$@
+
+fixtures/pointer_head_without_temperature.gguf: $(KEV_CONVERTER_SOURCES) paddler_kev_converter_python/.venv
+	poetry -C paddler_kev_converter_python run paddler-kev-converter malformed-pointer-head --input-normalization NFC --defect missing-temperature --hidden-size 64 --output $(CURDIR)/$@
+
+fixtures/pointer_head_with_oversized_head_dim.gguf: $(KEV_CONVERTER_SOURCES) paddler_kev_converter_python/.venv
+	poetry -C paddler_kev_converter_python run paddler-kev-converter malformed-pointer-head --input-normalization NFC --defect oversized-head-dim --hidden-size 1 --output $(CURDIR)/$@
+
+fixtures/qwen3_5_0_8b_pointer_head_with_plain_text_delimiters.gguf: $(KEV_CONVERTER_SOURCES) paddler_kev_converter_python/.venv
+	poetry -C paddler_kev_converter_python run paddler-kev-converter malformed-pointer-head --input-normalization NFC --defect plain-text-delimiters --hidden-size 1024 --output $(CURDIR)/$@
+
+fixtures/qwen3_5_0_8b_synthetic_pointer_head.gguf: $(KEV_CONVERTER_SOURCES) paddler_kev_converter_python/.venv
+	poetry -C paddler_kev_converter_python run paddler-kev-converter synthetic-pointer-head --input-normalization NFC --hidden-size 1024 --output $(CURDIR)/$@
+
+fixtures/qwen3_5_0_8b_synthetic_pointer_head_with_wrong_hidden_size.gguf: $(KEV_CONVERTER_SOURCES) paddler_kev_converter_python/.venv
+	poetry -C paddler_kev_converter_python run paddler-kev-converter synthetic-pointer-head --input-normalization NFC --hidden-size 512 --output $(CURDIR)/$@
+
+fixtures/qwen3_8_27b_synthetic_pointer_head.gguf: $(KEV_CONVERTER_SOURCES) paddler_kev_converter_python/.venv
+	poetry -C paddler_kev_converter_python run paddler-kev-converter synthetic-pointer-head --input-normalization NFC --hidden-size 5120 --output $(CURDIR)/$@
+
+fixtures/qwen3_%_tokenizer_reference.json: $(KEV_CONVERTER_SOURCES) paddler_kev_converter_python/.venv
+	poetry -C paddler_kev_converter_python run paddler-kev-converter tokenizer-reference --base $(QWEN3_$*_BASE) --base-revision $(QWEN3_$*_BASE_REVISION) --output $(CURDIR)/$@
+
+fixtures/typesafe_reference.json: $(KEV_CONVERTER_SOURCES) paddler_kev_converter_python/.venv
+	poetry -C paddler_kev_converter_python run paddler-kev-converter typesafe-reference --output $(CURDIR)/$@
+
 node_modules: package-lock.json
 	npm ci
 	touch node_modules
@@ -36,45 +97,49 @@ paddler_client_javascript/dist: $(JS_CLIENT_SOURCES) node_modules
 	npm --workspace @intentee/paddler-client run build
 	touch $@
 
-paddler_client_python/.venv: paddler_client_python/poetry.lock
-	POETRY_VIRTUALENVS_IN_PROJECT=true poetry -C paddler_client_python install
+paddler_kev_converter_python/.venv: paddler_kev_converter_python/poetry.lock
+	POETRY_VIRTUALENVS_IN_PROJECT=true poetry -C paddler_kev_converter_python install
+	touch $@
+
+paddler_llama_cpp_converter_python/.venv: paddler_llama_cpp_converter_python/poetry.lock
+	POETRY_VIRTUALENVS_IN_PROJECT=true poetry -C paddler_llama_cpp_converter_python install
 	touch $@
 
 paddler_openai_client_python_test/.venv: paddler_openai_client_python_test/poetry.lock
 	POETRY_VIRTUALENVS_IN_PROJECT=true poetry -C paddler_openai_client_python_test install
 	touch $@
 
-paddler_test_cluster_python/.venv: paddler_test_cluster_python/poetry.lock
-	POETRY_VIRTUALENVS_IN_PROJECT=true poetry -C paddler_test_cluster_python install
+paddler_typesafe_client_python_test/.venv: paddler_typesafe_client_python_test/poetry.lock
+	POETRY_VIRTUALENVS_IN_PROJECT=true poetry -C paddler_typesafe_client_python_test install
 	touch $@
 
+$(RUN_WITH_TEST_CLUSTER_BINARY): $(PADDLER_SOURCES) $(RUN_WITH_TEST_CLUSTER_SOURCES)
+	cargo build -p paddler_tests --bin run_with_test_cluster --features tests_that_use_llms$(TEST_DEVICE_FEATURE_SUFFIX) $(TEST_DEVICE_TARGET_DIR)
+
 target/cuda/debug/paddler: $(PADDLER_SOURCES) esbuild-meta.json
 	cargo build -p paddler_cli --features cuda,web_admin_panel --target-dir target/cuda
 
 target/cuda/release/paddler: $(PADDLER_SOURCES) esbuild-meta.json
 	cargo build --release -p paddler_cli --features cuda,web_admin_panel --target-dir target/cuda
 
-target/cuda/release/paddler_gui: $(PADDLER_SOURCES) esbuild-meta.json
-	cargo build --release -p paddler_gui --features cuda,web_admin_panel --target-dir target/cuda
-
 target/debug/paddler: $(PADDLER_SOURCES)
 	cargo build -p paddler_cli
 
+target/kev_%/backbone/config.json target/kev_%/pointer_head.gguf: $(KEV_CONVERTER_SOURCES) paddler_kev_converter_python/.venv
+	poetry -C paddler_kev_converter_python run paddler-kev-converter convert --checkpoint $(KEV_CHECKPOINT_$*) --backbone-directory $(CURDIR)/target/kev_$*/backbone --pointer-head $(CURDIR)/target/kev_$*/pointer_head.gguf
+
+target/kev_%/model.gguf: target/kev_%/backbone/config.json paddler_llama_cpp_converter_python/.venv
+	poetry -C paddler_llama_cpp_converter_python run python $(LLAMA_CPP_SOURCE_DIRECTORY)/convert_hf_to_gguf.py $(CURDIR)/target/kev_$*/backbone --no-mtp --outtype $(KEV_GGUF_OUTTYPE_$*) --outfile $(CURDIR)/$@
+
 target/metal/debug/paddler: $(PADDLER_SOURCES) esbuild-meta.json
 	cargo build -p paddler_cli --features metal,web_admin_panel --target-dir target/metal
 
 target/metal/release/paddler: $(PADDLER_SOURCES) esbuild-meta.json
 	cargo build --release -p paddler_cli --features metal,web_admin_panel --target-dir target/metal
 
-target/metal/release/paddler_gui: $(PADDLER_SOURCES) esbuild-meta.json
-	cargo build --release -p paddler_gui --features metal,web_admin_panel --target-dir target/metal
-
 target/release/paddler: $(PADDLER_SOURCES) esbuild-meta.json
 	cargo build --release -p paddler_cli --features web_admin_panel
 
-target/release/paddler_gui: $(PADDLER_SOURCES) esbuild-meta.json
-	cargo build --release -p paddler_gui --features web_admin_panel
-
 target/test-model-cards.stamp: $(wildcard paddler_test_cluster_harness/src/model_card/*.rs) paddler_test_cluster_harness/tests/every_model_card_resolves_from_the_hugging_face_cache.rs
 	cargo test -p paddler_test_cluster_harness --features tests_that_use_llms --test every_model_card_resolves_from_the_hugging_face_cache
 	touch $@
@@ -100,23 +165,20 @@ clippy: esbuild-meta.json
 fmt: node_modules
 	./jarmuz-fmt.mjs
 
-.PHONY: lint.client.python
-lint.client.python: paddler_client_python/.venv paddler_test_cluster_python/.venv
-	poetry -C paddler_client_python run ruff check
-	poetry -C paddler_client_python run ruff format --check
-	poetry -C paddler_client_python run mypy .
-	poetry -C paddler_test_cluster_python run ruff check
-	poetry -C paddler_test_cluster_python run ruff format --check
-	poetry -C paddler_test_cluster_python run mypy .
+.PHONY: lint.kev_converter.python
+lint.kev_converter.python: paddler_kev_converter_python/.venv
+	poetry -C paddler_kev_converter_python run mypy .
 
 .PHONY: lint.openai.python
 lint.openai.python: paddler_openai_client_python_test/.venv
-	poetry -C paddler_openai_client_python_test run ruff check
-	poetry -C paddler_openai_client_python_test run ruff format --check
 	poetry -C paddler_openai_client_python_test run mypy .
 
+.PHONY: lint.typesafe.python
+lint.typesafe.python: paddler_typesafe_client_python_test/.venv
+	poetry -C paddler_typesafe_client_python_test run mypy .
+
 .PHONY: test
-test: test.client.js test.client.python test.unit test.integration
+test: test.client.js test.unit test.integration
 
 .PHONY: test.client.js
 test.client.js: $(PADDLER_TEST_BINARY) node_modules
@@ -130,53 +192,50 @@ test.client.js.coverage: $(PADDLER_TEST_BINARY) node_modules
 test.client.js.llm: $(PADDLER_TEST_BINARY) node_modules target/test-model-cards.stamp
 	$(PADDLER_BINARY_ENVIRONMENT) npm --workspace @intentee/paddler-client run test:llm
 
-.PHONY: test.client.python
-test.client.python: $(PADDLER_TEST_BINARY) paddler_client_python/.venv paddler_test_cluster_python/.venv
-	$(PADDLER_BINARY_ENVIRONMENT) poetry -C paddler_test_cluster_python run pytest tests/unit tests/cluster
-	$(PADDLER_BINARY_ENVIRONMENT) poetry -C paddler_client_python run pytest tests/unit tests/cluster
-
-.PHONY: test.client.python.coverage
-test.client.python.coverage: $(PADDLER_TEST_BINARY) paddler_client_python/.venv paddler_test_cluster_python/.venv
-	$(PADDLER_BINARY_ENVIRONMENT) poetry -C paddler_test_cluster_python run pytest --cov
-	$(PADDLER_BINARY_ENVIRONMENT) poetry -C paddler_client_python run pytest --cov
-
-.PHONY: test.client.python.llm
-test.client.python.llm: $(PADDLER_TEST_BINARY) paddler_client_python/.venv paddler_test_cluster_python/.venv target/test-model-cards.stamp
-	$(PADDLER_BINARY_ENVIRONMENT) poetry -C paddler_test_cluster_python run pytest tests/llm
-	$(PADDLER_BINARY_ENVIRONMENT) poetry -C paddler_client_python run pytest tests/llm
-
 .PHONY: test.coverage
-test.coverage: esbuild-meta.json node_modules target/test-model-cards.stamp
+test.coverage: esbuild-meta.json node_modules target/kev_0_8b/model.gguf target/kev_0_8b/pointer_head.gguf target/kev_27b/model.gguf target/kev_27b/pointer_head.gguf target/test-model-cards.stamp
 	cargo llvm-cov clean --workspace
-	cargo llvm-cov nextest --features tests_that_use_llms,web_admin_panel$(TEST_DEVICE_FEATURE_SUFFIX) --no-report --workspace
-	cargo llvm-cov report --no-default-ignore-filename-regex --ignore-filename-regex '$(LLVM_COV_THIRD_PARTY_SOURCES)' --json --output-path target/llvm-cov.json
-	cargo llvm-cov report --no-default-ignore-filename-regex --ignore-filename-regex '$(LLVM_COV_THIRD_PARTY_SOURCES)' --lcov --output-path target/lcov.info
-	cargo llvm-cov report --no-default-ignore-filename-regex --ignore-filename-regex '$(LLVM_COV_THIRD_PARTY_SOURCES)'
+	NEXTEST_PROFILE=$(NEXTEST_PROFILE) cargo llvm-cov nextest --features tests_that_use_llms,tests_that_use_opencode,web_admin_panel$(TEST_DEVICE_FEATURE_SUFFIX) --include-build-script --no-report --workspace
+	cargo llvm-cov report --include-build-script --no-default-ignore-filename-regex --ignore-filename-regex '$(LLVM_COV_THIRD_PARTY_SOURCES)' --json --output-path target/llvm-cov.json
+	cargo llvm-cov report --include-build-script --no-default-ignore-filename-regex --ignore-filename-regex '$(LLVM_COV_THIRD_PARTY_SOURCES)' --lcov --output-path target/lcov.info
+	cargo llvm-cov report --include-build-script --no-default-ignore-filename-regex --ignore-filename-regex '$(LLVM_COV_THIRD_PARTY_SOURCES)'
 	npx rust-coverage-check target/llvm-cov.json \
 		--workspace-root $(CURDIR) \
-		--gated paddler_agent=97 \
+		--gated paddler_agent=100 \
+		--gated paddler_agent_decision=100 \
+		--gated paddler_agent_embeddings=100 \
+		--gated paddler_agent_pointer_head=100 \
+		--gated paddler_agent_runtime=100 \
 		--gated paddler_agent_status=100 \
-		--gated paddler_balancer=99 \
+		--gated paddler_agent_text_generation=100 \
+		--gated paddler_balancer=100 \
+		--gated paddler_blocking_thread_fixture=100 \
 		--gated paddler_bootstrap=100 \
 		--gated paddler_cache_dir=100 \
-		--gated paddler_cli=99 \
-		--gated paddler_cli_tests=94 \
-		--gated paddler_client=99 \
-		--gated paddler_download_manager=99 \
-		--gated paddler_gui=91 \
-		--gated paddler_gui_tests=97 \
+		--gated paddler_cli=100 \
+		--gated paddler_cli_tests=100 \
+		--gated paddler_client=100 \
+		--gated paddler_dashboard_assets=100 \
+		--gated paddler_download_manager=100 \
+		--gated paddler_frame_writer=100 \
 		--gated paddler_image_decoder=100 \
 		--gated paddler_inference_parameters=100 \
-		--gated paddler_local_http_fixture=89 \
+		--gated paddler_local_http_fixture=100 \
 		--gated paddler_messaging=100 \
-		--gated paddler_model_source=99 \
+		--gated paddler_model_source=100 \
 		--gated paddler_openai_response_format_validator=100 \
-		--gated paddler_opencode_tests=79 \
+		--gated paddler_openai_translation=100 \
+		--gated paddler_opencode_tests=100 \
+		--gated paddler_outbox=100 \
 		--gated paddler_request_registry=100 \
-		--gated paddler_state_database=99 \
-		--gated paddler_test_cluster_harness=94 \
-		--gated paddler_tests=98 \
-		--gated paddler_tool_call_validator=100
+		--gated paddler_sha256_hex=100 \
+		--gated paddler_spawn_resuming_panic=100 \
+		--gated paddler_state_database=100 \
+		--gated paddler_tcp_buffer_limits=100 \
+		--gated paddler_test_cluster_harness=100 \
+		--gated paddler_tests=100 \
+		--gated paddler_tool_call_validator=100 \
+		--gated paddler_typesafe_translation=100
 
 .PHONY: test.coverage-clean
 test.coverage-clean:
@@ -185,20 +244,28 @@ test.coverage-clean:
 	rm -f target/llvm-cov.json target/lcov.info
 
 .PHONY: test.integration
-test.integration: target/test-model-cards.stamp
-	cargo nextest run -p paddler_tests -p paddler_cli_tests --features tests_that_use_llms$(TEST_DEVICE_FEATURE_SUFFIX) $(TEST_DEVICE_TARGET_DIR)
+test.integration: target/kev_0_8b/model.gguf target/kev_0_8b/pointer_head.gguf target/kev_27b/model.gguf target/kev_27b/pointer_head.gguf target/test-model-cards.stamp
+	NEXTEST_PROFILE=$(NEXTEST_PROFILE) cargo nextest run -p paddler_tests -p paddler_cli_tests --features tests_that_use_llms$(TEST_DEVICE_FEATURE_SUFFIX) $(TEST_DEVICE_TARGET_DIR)
 
 .PHONY: test.integration.opencode
 test.integration.opencode: target/test-model-cards.stamp
-	cargo nextest run -p paddler_opencode_tests --features tests_that_use_llms,tests_that_use_opencode$(TEST_DEVICE_FEATURE_SUFFIX) $(TEST_DEVICE_TARGET_DIR)
+	NEXTEST_PROFILE=$(NEXTEST_PROFILE) cargo nextest run -p paddler_opencode_tests --features tests_that_use_llms,tests_that_use_opencode$(TEST_DEVICE_FEATURE_SUFFIX) $(TEST_DEVICE_TARGET_DIR)
+
+.PHONY: test.kev_converter.python.llm
+test.kev_converter.python.llm: paddler_kev_converter_python/.venv
+	PADDLER_KEV_CHECKPOINT=$(KEV_CHECKPOINT_0_8b) poetry -C paddler_kev_converter_python run pytest --cov
 
 .PHONY: test.openai.python.llm
-test.openai.python.llm: $(PADDLER_TEST_BINARY) paddler_openai_client_python_test/.venv target/test-model-cards.stamp
-	$(PADDLER_BINARY_ENVIRONMENT) poetry -C paddler_openai_client_python_test run pytest
+test.openai.python.llm: $(RUN_WITH_TEST_CLUSTER_BINARY) paddler_openai_client_python_test/.venv target/test-model-cards.stamp
+	$(RUN_WITH_TEST_CLUSTER_BINARY) qwen3-0-6b -- poetry -C paddler_openai_client_python_test run pytest
+
+.PHONY: test.typesafe.python.llm
+test.typesafe.python.llm: $(RUN_WITH_TEST_CLUSTER_BINARY) paddler_typesafe_client_python_test/.venv target/kev_0_8b/model.gguf target/kev_0_8b/pointer_head.gguf
+	$(RUN_WITH_TEST_CLUSTER_BINARY) kev-0-8b -- poetry -C paddler_typesafe_client_python_test run pytest
 
 .PHONY: test.unit
 test.unit: esbuild-meta.json
-	cargo nextest run --features web_admin_panel$(TEST_DEVICE_FEATURE_SUFFIX) $(TEST_DEVICE_TARGET_DIR)
+	NEXTEST_PROFILE=$(NEXTEST_PROFILE) cargo nextest run --features web_admin_panel$(TEST_DEVICE_FEATURE_SUFFIX) $(TEST_DEVICE_TARGET_DIR)
 
 .PHONY: watch
 watch: node_modules
diff --git a/clippy.toml b/clippy.toml
index e2a2044f3..c392dd52c 100644
--- a/clippy.toml
+++ b/clippy.toml
@@ -2,3 +2,8 @@ absolute-paths-max-segments = 1
 allow-expect-in-tests = true
 allow-panic-in-tests = true
 allow-unwrap-in-tests = true
+disallowed-methods = [
+    { path = "llama_cpp_bindings::context::LlamaContext::from_model", reason = "create contexts through LoadedLlamaModel::create_llama_context, which claims the VRAM residency of their capability" },
+    { path = "llama_cpp_bindings::model::LlamaModel::load_from_file", reason = "load models through LoadedLlamaModel::load, which claims the VRAM residency of the model weights" },
+    { path = "llama_cpp_bindings::mtmd::mtmd_context::MtmdContext::init_from_file", reason = "load multimodal projections through the text generation pipeline, which claims the VRAM residency of the projection" },
+]
diff --git a/eslint.config.js b/eslint.config.js
index cdb124302..3877414cc 100644
--- a/eslint.config.js
+++ b/eslint.config.js
@@ -20,7 +20,7 @@ export default tseslint.config(
     extends: [
       eslintConfig,
       tseslint.configs.strictTypeChecked,
-      reactHooks.configs["recommended-latest"],
+      reactHooks.configs.flat.recommended,
     ],
     files: ["resources/ts/**/*.{ts,tsx}"],
     languageOptions: {
diff --git a/fixtures/kev_0_8b_parity_reference.json b/fixtures/kev_0_8b_parity_reference.json
new file mode 100644
index 000000000..7e4818283
--- /dev/null
+++ b/fixtures/kev_0_8b_parity_reference.json
@@ -0,0 +1,165 @@
+[
+  {
+    "state": "Shoes arrived two weeks late and in the wrong size. Also I see two charges on my card.",
+    "questions": [
+      {
+        "id": "department",
+        "instructions": "Which team should handle this?",
+        "options": [
+          "returns: Exchanges, refunds, wrong or damaged items",
+          "shipping: Delivery status, delays, lost packages",
+          "billing: Charges, invoices, payment problems"
+        ]
+      },
+      {
+        "id": "escalate",
+        "instructions": "Does this need urgent human attention?",
+        "options": [
+          "no",
+          "yes"
+        ]
+      },
+      {
+        "id": "frustration",
+        "instructions": "How frustrated is the customer?",
+        "options": [
+          "Calm",
+          "Frustrated",
+          "Very angry"
+        ]
+      }
+    ],
+    "probabilities": [
+      [
+        0.261169970035553,
+        0.457897812128067,
+        0.2809322476387024
+      ],
+      [
+        0.445719450712204,
+        0.5542805194854736
+      ],
+      [
+        0.0512363575398922,
+        0.669154942035675,
+        0.27960875630378723
+      ]
+    ]
+  },
+  {
+    "state": "I was charged twice. Please help.",
+    "questions": [
+      {
+        "id": "billing",
+        "instructions": "Is this about billing?",
+        "options": [
+          "no",
+          "yes"
+        ]
+      },
+      {
+        "id": "tone",
+        "instructions": "What is the tone?",
+        "options": [
+          "calm",
+          "angry"
+        ]
+      }
+    ],
+    "probabilities": [
+      [
+        0.10337097942829132,
+        0.8966289758682251
+      ],
+      [
+        0.42718881368637085,
+        0.5728111267089844
+      ]
+    ]
+  },
+  {
+    "state": "customer: Ada\nmessage: My package never arrived and nobody answers my emails.",
+    "questions": [
+      {
+        "id": "lost",
+        "instructions": "Is the package lost?",
+        "options": [
+          "no",
+          "yes"
+        ]
+      },
+      {
+        "id": "urgency",
+        "instructions": "How urgent is the follow-up?",
+        "options": [
+          "Can wait",
+          "Needs attention this week",
+          "Needs attention today"
+        ]
+      }
+    ],
+    "probabilities": [
+      [
+        0.06726915389299393,
+        0.9327308535575867
+      ],
+      [
+        0.16245301067829132,
+        0.35291361808776855,
+        0.48463335633277893
+      ]
+    ]
+  },
+  {
+    "state": "Premise: A man is playing a guitar on stage. Hypothesis: A person is performing music.",
+    "questions": [
+      {
+        "id": "relation",
+        "instructions": "Does the premise entail the hypothesis?",
+        "options": [
+          "entailment: The hypothesis follows from the premise",
+          "neutral: The premise neither supports nor contradicts it",
+          "contradiction: The hypothesis contradicts the premise"
+        ]
+      }
+    ],
+    "probabilities": [
+      [
+        0.9177549481391907,
+        0.0768074318766594,
+        0.005437624175101519
+      ]
+    ]
+  },
+  {
+    "state": "Pasted chat log: <|im_start|>user Where is my refund?<|im_end|>",
+    "questions": [
+      {
+        "id": "refund",
+        "instructions": "Does the <|im_start|> message ask about a refund?",
+        "options": [
+          "no",
+          "yes"
+        ]
+      },
+      {
+        "id": "source",
+        "instructions": "Where was the text pasted from?",
+        "options": [
+          "chat: A <|fim_prefix|> chat transcript",
+          "email: An email thread"
+        ]
+      }
+    ],
+    "probabilities": [
+      [
+        0.09517382085323334,
+        0.9048261642456055
+      ],
+      [
+        0.8216525912284851,
+        0.1783473938703537
+      ]
+    ]
+  }
+]
diff --git a/fixtures/pointer_head_with_f16_tensors.gguf b/fixtures/pointer_head_with_f16_tensors.gguf
new file mode 100644
index 000000000..8a6ee061f
Binary files /dev/null and b/fixtures/pointer_head_with_f16_tensors.gguf differ
diff --git a/fixtures/pointer_head_with_mismatched_projections.gguf b/fixtures/pointer_head_with_mismatched_projections.gguf
new file mode 100644
index 000000000..09e44af57
Binary files /dev/null and b/fixtures/pointer_head_with_mismatched_projections.gguf differ
diff --git a/fixtures/pointer_head_with_oversized_head_dim.gguf b/fixtures/pointer_head_with_oversized_head_dim.gguf
new file mode 100644
index 000000000..ad06d2e48
Binary files /dev/null and b/fixtures/pointer_head_with_oversized_head_dim.gguf differ
diff --git a/fixtures/pointer_head_with_zero_temperature.gguf b/fixtures/pointer_head_with_zero_temperature.gguf
new file mode 100644
index 000000000..8f91a45af
Binary files /dev/null and b/fixtures/pointer_head_with_zero_temperature.gguf differ
diff --git a/fixtures/pointer_head_without_delimiters.gguf b/fixtures/pointer_head_without_delimiters.gguf
new file mode 100644
index 000000000..1a0e3cc19
Binary files /dev/null and b/fixtures/pointer_head_without_delimiters.gguf differ
diff --git a/fixtures/pointer_head_without_input_normalization.gguf b/fixtures/pointer_head_without_input_normalization.gguf
new file mode 100644
index 000000000..1240063eb
Binary files /dev/null and b/fixtures/pointer_head_without_input_normalization.gguf differ
diff --git a/fixtures/pointer_head_without_temperature.gguf b/fixtures/pointer_head_without_temperature.gguf
new file mode 100644
index 000000000..e25adbf69
Binary files /dev/null and b/fixtures/pointer_head_without_temperature.gguf differ
diff --git a/fixtures/qwen3_5_0_8b_pointer_head_with_plain_text_delimiters.gguf b/fixtures/qwen3_5_0_8b_pointer_head_with_plain_text_delimiters.gguf
new file mode 100644
index 000000000..dc29a5412
Binary files /dev/null and b/fixtures/qwen3_5_0_8b_pointer_head_with_plain_text_delimiters.gguf differ
diff --git a/fixtures/qwen3_5_0_8b_synthetic_pointer_head.gguf b/fixtures/qwen3_5_0_8b_synthetic_pointer_head.gguf
new file mode 100644
index 000000000..114578d2a
Binary files /dev/null and b/fixtures/qwen3_5_0_8b_synthetic_pointer_head.gguf differ
diff --git a/fixtures/qwen3_5_0_8b_synthetic_pointer_head_with_wrong_hidden_size.gguf b/fixtures/qwen3_5_0_8b_synthetic_pointer_head_with_wrong_hidden_size.gguf
new file mode 100644
index 000000000..dc6b3f264
Binary files /dev/null and b/fixtures/qwen3_5_0_8b_synthetic_pointer_head_with_wrong_hidden_size.gguf differ
diff --git a/fixtures/qwen3_5_tokenizer_reference.json b/fixtures/qwen3_5_tokenizer_reference.json
new file mode 100644
index 000000000..3edb98b9e
--- /dev/null
+++ b/fixtures/qwen3_5_tokenizer_reference.json
@@ -0,0 +1,378 @@
+{
+  "entries": [
+    {
+      "text": "Hello world",
+      "token_ids": [
+        9419,
+        1814
+      ]
+    },
+    {
+      "text": "  leading and trailing spaces  ",
+      "token_ids": [
+        220,
+        6187,
+        321,
+        26849,
+        12258,
+        256
+      ]
+    },
+    {
+      "text": "several    spaces\tand\ttabs\nand\n\nnewlines",
+      "token_ids": [
+        323,
+        17018,
+        262,
+        12258,
+        50711,
+        3139,
+        3325,
+        198,
+        429,
+        271,
+        902,
+        7718
+      ]
+    },
+    {
+      "text": "1234567890 and 3.14159 and -42",
+      "token_ids": [
+        16,
+        17,
+        18,
+        19,
+        20,
+        21,
+        22,
+        23,
+        24,
+        15,
+        321,
+        220,
+        18,
+        13,
+        16,
+        19,
+        16,
+        20,
+        24,
+        321,
+        471,
+        19,
+        17
+      ]
+    },
+    {
+      "text": "Don't stop, I'm sure we'll go; they've said: it's done!",
+      "token_ids": [
+        7751,
+        914,
+        2842,
+        11,
+        353,
+        2688,
+        2617,
+        567,
+        3172,
+        708,
+        26,
+        781,
+        2908,
+        1018,
+        25,
+        424,
+        579,
+        2725,
+        0
+      ]
+    },
+    {
+      "text": "Zażółć gęślą jaźń",
+      "token_ids": [
+        57,
+        57897,
+        45813,
+        7485,
+        338,
+        4991,
+        156885,
+        5679,
+        11598,
+        38204,
+        18570
+      ]
+    },
+    {
+      "text": "Zażółć gęślą jaźń",
+      "token_ids": [
+        57,
+        57897,
+        45813,
+        7485,
+        338,
+        4991,
+        156885,
+        5679,
+        11598,
+        38204,
+        18570
+      ]
+    },
+    {
+      "text": "été and été",
+      "token_ids": [
+        37500,
+        321,
+        22899
+      ]
+    },
+    {
+      "text": "ＡＢＣ full width and ① circled",
+      "token_ids": [
+        169095,
+        169283,
+        169151,
+        2400,
+        2297,
+        321,
+        220,
+        46701,
+        254,
+        4086,
+        806
+      ]
+    },
+    {
+      "text": "日本語のテキストと中文文本",
+      "token_ids": [
+        247359,
+        15303,
+        210342,
+        18602,
+        99986,
+        109120
+      ]
+    },
+    {
+      "text": "مرحبا بالعالم",
+      "token_ids": [
+        148739,
+        28850,
+        150027,
+        182946,
+        149650
+      ]
+    },
+    {
+      "text": "family 👨‍👩‍👧‍👦",
+      "token_ids": [
+        18411,
+        59720,
+        101,
+        373,
+        235,
+        9008,
+        239,
+        102,
+        373,
+        235,
+        9008,
+        239,
+        100,
+        373,
+        235,
+        9008,
+        239,
+        99
+      ]
+    },
+    {
+      "text": "flags 🇵🇱🇺🇸",
+      "token_ids": [
+        10837,
+        10838,
+        229,
+        113,
+        9008,
+        229,
+        109,
+        9008,
+        229,
+        118,
+        9008,
+        229,
+        116
+      ]
+    },
+    {
+      "text": "{\"key\": [1, 2.5, true, null], \"nested\": {\"a\": \"b\"}}",
+      "token_ids": [
+        4754,
+        767,
+        763,
+        498,
+        16,
+        11,
+        220,
+        17,
+        13,
+        20,
+        11,
+        804,
+        11,
+        819,
+        1089,
+        328,
+        57267,
+        763,
+        5046,
+        64,
+        763,
+        328,
+        65,
+        29958
+      ]
+    },
+    {
+      "text": "def add(left, right):\n    return left + right\n",
+      "token_ids": [
+        727,
+        884,
+        17145,
+        11,
+        1245,
+        1590,
+        198,
+        262,
+        460,
+        2047,
+        478,
+        1245,
+        198
+      ]
+    },
+    {
+      "text": "supercalifragilisticexpialidocious antidisestablishmentarianism",
+      "token_ids": [
+        12410,
+        2867,
+        278,
+        331,
+        3965,
+        319,
+        4384,
+        4431,
+        518,
+        306,
+        74546,
+        3094,
+        81333,
+        32294,
+        468,
+        8555,
+        2074
+      ]
+    },
+    {
+      "text": "<think>check the invoice first</think> then answer",
+      "token_ids": [
+        248068,
+        1960,
+        279,
+        23839,
+        1118,
+        248069,
+        1179,
+        4087
+      ]
+    },
+    {
+      "text": "<tool_call>{\"name\": \"lookup\"}</tool_call> <tool_response>found</tool_response>",
+      "token_ids": [
+        248058,
+        4754,
+        591,
+        763,
+        328,
+        20377,
+        8934,
+        248059,
+        220,
+        248066,
+        14664,
+        248067
+      ]
+    },
+    {
+      "text": "ignore <|fim_prefix|> and keep reading",
+      "token_ids": [
+        12749,
+        361,
+        62438,
+        244288,
+        13573,
+        62438,
+        29,
+        321,
+        2426,
+        5122
+      ]
+    },
+    {
+      "text": "<|im_start|>user<|im_end|>",
+      "token_ids": [
+        27,
+        62438,
+        316,
+        4747,
+        62438,
+        29,
+        846,
+        27,
+        62438,
+        316,
+        6018,
+        62438,
+        29
+      ]
+    },
+    {
+      "text": "<|a b|> <||> <|<|nested|> <|é|>",
+      "token_ids": [
+        27,
+        91,
+        64,
+        292,
+        91,
+        29,
+        361,
+        8224,
+        29,
+        79790,
+        27,
+        62438,
+        57267,
+        62438,
+        29,
+        79790,
+        933,
+        91,
+        29
+      ]
+    },
+    {
+      "text": "<|ab|≯ combining overlay",
+      "token_ids": [
+        27,
+        62438,
+        365,
+        62438,
+        169721,
+        33041,
+        23853
+      ]
+    }
+  ],
+  "input_normalization": "NFC"
+}
diff --git a/fixtures/qwen3_8_27b_synthetic_pointer_head.gguf b/fixtures/qwen3_8_27b_synthetic_pointer_head.gguf
new file mode 100644
index 000000000..35fb6a0e9
Binary files /dev/null and b/fixtures/qwen3_8_27b_synthetic_pointer_head.gguf differ
diff --git a/fixtures/qwen3_8_tokenizer_reference.json b/fixtures/qwen3_8_tokenizer_reference.json
new file mode 100644
index 000000000..3edb98b9e
--- /dev/null
+++ b/fixtures/qwen3_8_tokenizer_reference.json
@@ -0,0 +1,378 @@
+{
+  "entries": [
+    {
+      "text": "Hello world",
+      "token_ids": [
+        9419,
+        1814
+      ]
+    },
+    {
+      "text": "  leading and trailing spaces  ",
+      "token_ids": [
+        220,
+        6187,
+        321,
+        26849,
+        12258,
+        256
+      ]
+    },
+    {
+      "text": "several    spaces\tand\ttabs\nand\n\nnewlines",
+      "token_ids": [
+        323,
+        17018,
+        262,
+        12258,
+        50711,
+        3139,
+        3325,
+        198,
+        429,
+        271,
+        902,
+        7718
+      ]
+    },
+    {
+      "text": "1234567890 and 3.14159 and -42",
+      "token_ids": [
+        16,
+        17,
+        18,
+        19,
+        20,
+        21,
+        22,
+        23,
+        24,
+        15,
+        321,
+        220,
+        18,
+        13,
+        16,
+        19,
+        16,
+        20,
+        24,
+        321,
+        471,
+        19,
+        17
+      ]
+    },
+    {
+      "text": "Don't stop, I'm sure we'll go; they've said: it's done!",
+      "token_ids": [
+        7751,
+        914,
+        2842,
+        11,
+        353,
+        2688,
+        2617,
+        567,
+        3172,
+        708,
+        26,
+        781,
+        2908,
+        1018,
+        25,
+        424,
+        579,
+        2725,
+        0
+      ]
+    },
+    {
+      "text": "Zażółć gęślą jaźń",
+      "token_ids": [
+        57,
+        57897,
+        45813,
+        7485,
+        338,
+        4991,
+        156885,
+        5679,
+        11598,
+        38204,
+        18570
+      ]
+    },
+    {
+      "text": "Zażółć gęślą jaźń",
+      "token_ids": [
+        57,
+        57897,
+        45813,
+        7485,
+        338,
+        4991,
+        156885,
+        5679,
+        11598,
+        38204,
+        18570
+      ]
+    },
+    {
+      "text": "été and été",
+      "token_ids": [
+        37500,
+        321,
+        22899
+      ]
+    },
+    {
+      "text": "ＡＢＣ full width and ① circled",
+      "token_ids": [
+        169095,
+        169283,
+        169151,
+        2400,
+        2297,
+        321,
+        220,
+        46701,
+        254,
+        4086,
+        806
+      ]
+    },
+    {
+      "text": "日本語のテキストと中文文本",
+      "token_ids": [
+        247359,
+        15303,
+        210342,
+        18602,
+        99986,
+        109120
+      ]
+    },
+    {
+      "text": "مرحبا بالعالم",
+      "token_ids": [
+        148739,
+        28850,
+        150027,
+        182946,
+        149650
+      ]
+    },
+    {
+      "text": "family 👨‍👩‍👧‍👦",
+      "token_ids": [
+        18411,
+        59720,
+        101,
+        373,
+        235,
+        9008,
+        239,
+        102,
+        373,
+        235,
+        9008,
+        239,
+        100,
+        373,
+        235,
+        9008,
+        239,
+        99
+      ]
+    },
+    {
+      "text": "flags 🇵🇱🇺🇸",
+      "token_ids": [
+        10837,
+        10838,
+        229,
+        113,
+        9008,
+        229,
+        109,
+        9008,
+        229,
+        118,
+        9008,
+        229,
+        116
+      ]
+    },
+    {
+      "text": "{\"key\": [1, 2.5, true, null], \"nested\": {\"a\": \"b\"}}",
+      "token_ids": [
+        4754,
+        767,
+        763,
+        498,
+        16,
+        11,
+        220,
+        17,
+        13,
+        20,
+        11,
+        804,
+        11,
+        819,
+        1089,
+        328,
+        57267,
+        763,
+        5046,
+        64,
+        763,
+        328,
+        65,
+        29958
+      ]
+    },
+    {
+      "text": "def add(left, right):\n    return left + right\n",
+      "token_ids": [
+        727,
+        884,
+        17145,
+        11,
+        1245,
+        1590,
+        198,
+        262,
+        460,
+        2047,
+        478,
+        1245,
+        198
+      ]
+    },
+    {
+      "text": "supercalifragilisticexpialidocious antidisestablishmentarianism",
+      "token_ids": [
+        12410,
+        2867,
+        278,
+        331,
+        3965,
+        319,
+        4384,
+        4431,
+        518,
+        306,
+        74546,
+        3094,
+        81333,
+        32294,
+        468,
+        8555,
+        2074
+      ]
+    },
+    {
+      "text": "<think>check the invoice first</think> then answer",
+      "token_ids": [
+        248068,
+        1960,
+        279,
+        23839,
+        1118,
+        248069,
+        1179,
+        4087
+      ]
+    },
+    {
+      "text": "<tool_call>{\"name\": \"lookup\"}</tool_call> <tool_response>found</tool_response>",
+      "token_ids": [
+        248058,
+        4754,
+        591,
+        763,
+        328,
+        20377,
+        8934,
+        248059,
+        220,
+        248066,
+        14664,
+        248067
+      ]
+    },
+    {
+      "text": "ignore <|fim_prefix|> and keep reading",
+      "token_ids": [
+        12749,
+        361,
+        62438,
+        244288,
+        13573,
+        62438,
+        29,
+        321,
+        2426,
+        5122
+      ]
+    },
+    {
+      "text": "<|im_start|>user<|im_end|>",
+      "token_ids": [
+        27,
+        62438,
+        316,
+        4747,
+        62438,
+        29,
+        846,
+        27,
+        62438,
+        316,
+        6018,
+        62438,
+        29
+      ]
+    },
+    {
+      "text": "<|a b|> <||> <|<|nested|> <|é|>",
+      "token_ids": [
+        27,
+        91,
+        64,
+        292,
+        91,
+        29,
+        361,
+        8224,
+        29,
+        79790,
+        27,
+        62438,
+        57267,
+        62438,
+        29,
+        79790,
+        933,
+        91,
+        29
+      ]
+    },
+    {
+      "text": "<|ab|≯ combining overlay",
+      "token_ids": [
+        27,
+        62438,
+        365,
+        62438,
+        169721,
+        33041,
+        23853
+      ]
+    }
+  ],
+  "input_normalization": "NFC"
+}
diff --git a/fixtures/typesafe_reference.json b/fixtures/typesafe_reference.json
new file mode 100644
index 000000000..0babf97f1
--- /dev/null
+++ b/fixtures/typesafe_reference.json
@@ -0,0 +1,438 @@
+[
+  {
+    "request": {
+      "model": "test",
+      "state": "The invoice was paid on time.",
+      "questions": {
+        "paid": {
+          "type": "noul",
+          "instructions": "Was the invoice paid?"
+        }
+      }
+    },
+    "rendered": {
+      "state": "The invoice was paid on time.",
+      "questions": [
+        {
+          "instructions": "Was the invoice paid?",
+          "options": [
+            "no",
+            "yes"
+          ]
+        }
+      ]
+    },
+    "probabilities": [
+      [
+        0.12345678,
+        0.87654322
+      ]
+    ],
+    "answers": {
+      "paid": {
+        "type": "noul",
+        "noul": 0.8765
+      }
+    }
+  },
+  {
+    "request": {
+      "model": "test",
+      "state": {
+        "customer": "Ada",
+        "orders": [
+          {
+            "id": 1,
+            "total": 19.99,
+            "shipped": true
+          },
+          {
+            "id": 2,
+            "total": 100.0,
+            "shipped": false,
+            "notes": null
+          }
+        ],
+        "tags": [],
+        "metadata": {},
+        "balance": -0.0,
+        "ratio": 1e-07,
+        "population": 1e+16,
+        "huge": 1.5e+300,
+        "count": 123456789012345678
+      },
+      "questions": {
+        "sentiment": {
+          "type": "choice",
+          "instructions": {
+            "task": "Classify",
+            "hints": [
+              "be brief",
+              2
+            ]
+          },
+          "criteria": {
+            "positive": "The customer is happy",
+            "neutral": null,
+            "negative": "",
+            "mixed": {
+              "when": "both",
+              "weight": 0.5
+            }
+          }
+        },
+        "risk": {
+          "type": "score",
+          "instructions": null,
+          "criteria": [
+            "no risk",
+            {
+              "level": "some",
+              "detail": [
+                "late",
+                "partial"
+              ]
+            },
+            "high risk"
+          ]
+        },
+        "vip": {
+          "type": "noul",
+          "instructions": 42,
+          "criteria": {
+            "true": "Spends a lot",
+            "false": {
+              "reason": "rare"
+            }
+          }
+        }
+      }
+    },
+    "rendered": {
+      "state": "customer: Ada\norders:\n  - id: 1\n    total: 19.99\n    shipped: True\n  - id: 2\n    total: 100.0\n    shipped: False\n    notes: \ntags:\n\nmetadata:\n\nbalance: -0.0\nratio: 1e-07\npopulation: 1e+16\nhuge: 1.5e+300\ncount: 123456789012345678",
+      "questions": [
+        {
+          "instructions": "task: Classify\nhints:\n  - be brief\n  - 2",
+          "options": [
+            "positive: The customer is happy",
+            "neutral",
+            "negative",
+            "mixed: when: both\nweight: 0.5"
+          ]
+        },
+        {
+          "instructions": "",
+          "options": [
+            "no risk",
+            "level: some\ndetail:\n  - late\n  - partial",
+            "high risk"
+          ]
+        },
+        {
+          "instructions": "42",
+          "options": [
+            "no: reason: rare",
+            "yes: Spends a lot"
+          ]
+        }
+      ]
+    },
+    "probabilities": [
+      [
+        0.25,
+        0.25,
+        0.4,
+        0.1
+      ],
+      [
+        0.2,
+        0.3,
+        0.5
+      ],
+      [
+        0.5,
+        0.5
+      ]
+    ],
+    "answers": {
+      "sentiment": {
+        "type": "choice",
+        "choice": "negative",
+        "confidence": 0.2,
+        "probabilities": {
+          "positive": 0.25,
+          "neutral": 0.25,
+          "negative": 0.4,
+          "mixed": 0.1
+        }
+      },
+      "risk": {
+        "type": "score",
+        "score": 1.3,
+        "legend": {
+          "0": "no risk",
+          "1": "level: some\ndetail:\n  - late\n  - partial",
+          "2": "high risk"
+        },
+        "probabilities": {
+          "0": 0.2,
+          "1": 0.3,
+          "2": 0.5
+        },
+        "confidence": 0.0
+      },
+      "vip": {
+        "type": "noul",
+        "noul": 0.5
+      }
+    }
+  },
+  {
+    "request": {
+      "model": "test",
+      "state": [
+        "first line\nsecond line",
+        3.5,
+        true,
+        null,
+        [
+          "nested",
+          1
+        ]
+      ],
+      "questions": {
+        "tie": {
+          "type": "choice",
+          "instructions": [
+            "pick",
+            "one"
+          ],
+          "criteria": {
+            "left": 1,
+            "right": false
+          }
+        },
+        "single": {
+          "type": "score",
+          "instructions": 2.5,
+          "criteria": [
+            "only level"
+          ]
+        }
+      }
+    },
+    "rendered": {
+      "state": "- first line\nsecond line\n- 3.5\n- True\n- \n- - nested\n  - 1",
+      "questions": [
+        {
+          "instructions": "- pick\n- one",
+          "options": [
+            "left: 1",
+            "right: False"
+          ]
+        },
+        {
+          "instructions": "2.5",
+          "options": [
+            "only level"
+          ]
+        }
+      ]
+    },
+    "probabilities": [
+      [
+        0.5,
+        0.5
+      ],
+      [
+        1.0
+      ]
+    ],
+    "answers": {
+      "tie": {
+        "type": "choice",
+        "choice": "left",
+        "confidence": 0.0,
+        "probabilities": {
+          "left": 0.5,
+          "right": 0.5
+        }
+      },
+      "single": {
+        "type": "score",
+        "score": 0.0,
+        "legend": {
+          "0": "only level"
+        },
+        "probabilities": {
+          "0": 1.0
+        },
+        "confidence": 1.0
+      }
+    }
+  },
+  {
+    "request": {
+      "model": "test",
+      "state": "Zażółć gęślą jaźń 👨‍👩",
+      "questions": {
+        "lonely": {
+          "type": "choice",
+          "instructions": true,
+          "criteria": {
+            "only": "the only option"
+          }
+        },
+        "uncertain": {
+          "type": "score",
+          "instructions": "How confident?",
+          "criteria": [
+            "low",
+            "medium",
+            "high",
+            "certain"
+          ]
+        }
+      }
+    },
+    "rendered": {
+      "state": "Zażółć gęślą jaźń 👨‍👩",
+      "questions": [
+        {
+          "instructions": "True",
+          "options": [
+            "only: the only option"
+          ]
+        },
+        {
+          "instructions": "How confident?",
+          "options": [
+            "low",
+            "medium",
+            "high",
+            "certain"
+          ]
+        }
+      ]
+    },
+    "probabilities": [
+      [
+        1.0
+      ],
+      [
+        0.25,
+        0.25,
+        0.25,
+        0.25
+      ]
+    ],
+    "answers": {
+      "lonely": {
+        "type": "choice",
+        "choice": "only",
+        "confidence": 1.0,
+        "probabilities": {
+          "only": 1.0
+        }
+      },
+      "uncertain": {
+        "type": "score",
+        "score": 1.5,
+        "legend": {
+          "0": "low",
+          "1": "medium",
+          "2": "high",
+          "3": "certain"
+        },
+        "probabilities": {
+          "0": 0.25,
+          "1": 0.25,
+          "2": 0.25,
+          "3": 0.25
+        },
+        "confidence": 0.0
+      }
+    }
+  },
+  {
+    "request": {
+      "model": "test",
+      "state": {
+        "precise": 123.45678901234567,
+        "halfway": 1000000000000.6562,
+        "power_of_two": 7.120236347223045e-307
+      },
+      "questions": {
+        "measure": {
+          "type": "score",
+          "instructions": 1000000000000.6562,
+          "criteria": [
+            123.45678901234567,
+            7.120236347223045e-307
+          ]
+        },
+        "pick": {
+          "type": "choice",
+          "instructions": "Which value is exact?",
+          "criteria": {
+            "halfway": 1000000000000.6562,
+            "precise": 123.45678901234567
+          }
+        }
+      }
+    },
+    "rendered": {
+      "state": "precise: 123.45678901234567\nhalfway: 1000000000000.6562\npower_of_two: 7.120236347223045e-307",
+      "questions": [
+        {
+          "instructions": "1000000000000.6562",
+          "options": [
+            "123.45678901234567",
+            "7.120236347223045e-307"
+          ]
+        },
+        {
+          "instructions": "Which value is exact?",
+          "options": [
+            "halfway: 1000000000000.6562",
+            "precise: 123.45678901234567"
+          ]
+        }
+      ]
+    },
+    "probabilities": [
+      [
+        0.5,
+        0.5
+      ],
+      [
+        0.5,
+        0.5
+      ]
+    ],
+    "answers": {
+      "measure": {
+        "type": "score",
+        "score": 0.5,
+        "legend": {
+          "0": "123.45678901234567",
+          "1": "7.120236347223045e-307"
+        },
+        "probabilities": {
+          "0": 0.5,
+          "1": 0.5
+        },
+        "confidence": 0.0
+      },
+      "pick": {
+        "type": "choice",
+        "choice": "halfway",
+        "confidence": 0.0,
+        "probabilities": {
+          "halfway": 0.5,
+          "precise": 0.5
+        }
+      }
+    }
+  }
+]
diff --git a/flake.lock b/flake.lock
index cc14f0c95..e49133b41 100644
--- a/flake.lock
+++ b/flake.lock
@@ -37,11 +37,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1784356753,
-        "narHash": "sha256-12KrbMiWLcf8m7pCvAtZh1ZrgF85ZXDXvfR/fWTKy84=",
+        "lastModified": 1791260994,
+        "narHash": "sha256-Miqqk/ammqnTUxaoCyvtwoPeLbI1ksFyLxi/CazZpWY=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "61b7c44c4073f0b827768aff0049561b5110ea5a",
+        "rev": "151fa4e8ddfdd8dd25d945ad94ed54a13de9f6e4",
         "type": "github"
       },
       "original": {
@@ -66,11 +66,11 @@
         ]
       },
       "locked": {
-        "lastModified": 1784438913,
-        "narHash": "sha256-NYF7ZM5ip0u+w1pBFDpIGEbrbgN/wpnLFAmBkWkYMXw=",
+        "lastModified": 1792395246,
+        "narHash": "sha256-SfVFE7TKuByeQRwhVbaVrvTV/lI41AYK9CYSfB+pIis=",
         "owner": "oxalica",
         "repo": "rust-overlay",
-        "rev": "afacd6819d3765a05814ee8e3de74c77d42ac799",
+        "rev": "550547cefc67e01a6fd11a09e4096a3c844ca840",
         "type": "github"
       },
       "original": {
diff --git a/flake.nix b/flake.nix
index 76cff4945..81030979a 100644
--- a/flake.nix
+++ b/flake.nix
@@ -23,7 +23,7 @@
       crane,
     }:
     let
-      version = "4.1.0";
+      version = "5.0.0";
 
       paddlerPkgs =
         {
@@ -41,7 +41,9 @@
           // (if cudaCapabilities == [ ] then { } else { inherit cudaCapabilities; });
         };
 
-      craneLibFor = pkgs: (crane.mkLib pkgs).overrideToolchain (pkgs.rust-bin.fromRustupToolchainFile ./rust-toolchain.toml);
+      craneLibFor =
+        pkgs:
+        (crane.mkLib pkgs).overrideToolchain (pkgs.rust-bin.fromRustupToolchainFile ./rust-toolchain.toml);
 
       defaultAccelerator =
         pkgs:
@@ -77,7 +79,7 @@
             pname = "paddler-web-admin-panel";
             inherit version;
             src = self;
-            npmDepsHash = "sha256-sBDdMf388qFQVIjQ3t/BL3KC/yAqF1qj47a/40axgF8=";
+            npmDepsHash = "sha256-QqCDIIulF4bZ4F+vP+Z2/Q2zHSjzGXXb6j+AcPjpvpQ=";
             dontNpmBuild = true;
             nativeBuildInputs = [ pkgs.nodejs ];
             buildPhase = ''
@@ -165,6 +167,7 @@
             doCheck = false;
             nativeBuildInputs = [
               pkgs.cmake
+              pkgs.jq
               pkgs.pkg-config
               pkgs.llvmPackages.clang
             ]
@@ -232,11 +235,21 @@
               pkgs.pkg-config
               pkgs.llvmPackages.clang
               pkgs.openssl
+              pkgs.poetry
+              pkgs.python313
             ];
+            shellHook = ''
+              export LD_LIBRARY_PATH="${
+                pkgs.lib.makeLibraryPath [
+                  pkgs.stdenv.cc.cc.lib
+                  pkgs.zlib
+                ]
+              }:${pkgs.addDriverRunpath.driverLink}/lib''${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}"
+            '';
             LIBCLANG_PATH = "${pkgs.llvmPackages.libclang.lib}/lib";
           };
 
-          formatter = pkgs.nixfmt-rfc-style;
+          formatter = pkgs.nixfmt;
         };
 
       flake =
@@ -300,6 +313,10 @@
                   "--compat-openai-addr"
                   balancer.openaiCompatAddr
                 ]
+                ++ lib.optionals (balancer.typesafeCompatAddr != null) [
+                  "--compat-typesafe-addr"
+                  balancer.typesafeCompatAddr
+                ]
                 ++ lib.concatMap (host: [
                   "--management-cors-allowed-host"
                   host
@@ -364,6 +381,12 @@
                     description = "Address of the OpenAI-compatible API server. When null it is disabled.";
                   };
 
+                  typesafeCompatAddr = lib.mkOption {
+                    type = lib.types.nullOr socketAddrType;
+                    default = null;
+                    description = "Address of the TypeSafe-compatible System One API server. When null it is disabled.";
+                  };
+
                   stateDatabase = lib.mkOption {
                     type = lib.types.str;
                     default = "file:///var/lib/paddler/state.db";
@@ -391,7 +414,7 @@
                   openFirewall = lib.mkOption {
                     type = lib.types.bool;
                     default = false;
-                    description = "Open the management, inference, web admin panel and OpenAI-compatible ports in the firewall.";
+                    description = "Open the management, inference, web admin panel, OpenAI-compatible and TypeSafe-compatible ports in the firewall.";
                   };
                 };
 
@@ -512,7 +535,8 @@
                         (portOf cfg.balancer.inferenceAddr)
                       ]
                       ++ lib.optional (cfg.balancer.webAdminPanelAddr != null) (portOf cfg.balancer.webAdminPanelAddr)
-                      ++ lib.optional (cfg.balancer.openaiCompatAddr != null) (portOf cfg.balancer.openaiCompatAddr);
+                      ++ lib.optional (cfg.balancer.openaiCompatAddr != null) (portOf cfg.balancer.openaiCompatAddr)
+                      ++ lib.optional (cfg.balancer.typesafeCompatAddr != null) (portOf cfg.balancer.typesafeCompatAddr);
                   };
                 })
 
diff --git a/jarmuz-fmt.mjs b/jarmuz-fmt.mjs
index 2bbd3ee81..4c13cec87 100755
--- a/jarmuz-fmt.mjs
+++ b/jarmuz-fmt.mjs
@@ -8,29 +8,41 @@ jarmuz({
   watch: [
     "jarmuz",
     "paddler_agent",
+    "paddler_agent_decision",
+    "paddler_agent_embeddings",
+    "paddler_agent_pointer_head",
+    "paddler_agent_runtime",
     "paddler_agent_status",
+    "paddler_agent_text_generation",
     "paddler_balancer",
+    "paddler_blocking_thread_fixture",
     "paddler_bootstrap",
     "paddler_cache_dir",
     "paddler_cli",
     "paddler_cli_tests",
     "paddler_client",
     "paddler_client_javascript",
+    "paddler_dashboard_assets",
     "paddler_download_manager",
-    "paddler_gui",
-    "paddler_gui_tests",
+    "paddler_frame_writer",
     "paddler_image_decoder",
     "paddler_inference_parameters",
     "paddler_local_http_fixture",
     "paddler_messaging",
     "paddler_model_source",
     "paddler_openai_response_format_validator",
+    "paddler_openai_translation",
     "paddler_opencode_tests",
+    "paddler_outbox",
     "paddler_request_registry",
+    "paddler_sha256_hex",
+    "paddler_spawn_resuming_panic",
     "paddler_state_database",
+    "paddler_tcp_buffer_limits",
     "paddler_test_cluster_harness",
     "paddler_tests",
     "paddler_tool_call_validator",
+    "paddler_typesafe_translation",
     "resources",
   ],
 }).decide(function ({ matches, schedule }) {
diff --git a/jarmuz/run-website.mjs b/jarmuz/run-website.mjs
index 43cb0bb81..ba14ddf99 100644
--- a/jarmuz/run-website.mjs
+++ b/jarmuz/run-website.mjs
@@ -8,7 +8,12 @@ export function run({ development, once = false, rustJobs }) {
     pipeline: ["stylelint", "tcm", "tsc", "eslint", esbuildJob, ...rustJobs],
     watch: [
       "paddler_agent",
+      "paddler_agent_decision",
+      "paddler_agent_embeddings",
+      "paddler_agent_pointer_head",
+      "paddler_agent_runtime",
       "paddler_agent_status",
+      "paddler_agent_text_generation",
       "paddler_balancer",
       "paddler_bootstrap",
       "paddler_cache_dir",
@@ -20,9 +25,11 @@ export function run({ development, once = false, rustJobs }) {
       "paddler_inference_parameters",
       "paddler_messaging",
       "paddler_model_source",
+      "paddler_openai_translation",
       "paddler_request_registry",
       "paddler_state_database",
       "paddler_tool_call_validator",
+      "paddler_typesafe_translation",
       "resources",
     ],
   }).decide(function ({ matches, schedule }) {
diff --git a/jarmuz/worker-paddler.mjs b/jarmuz/worker-paddler.mjs
index f7a9f3484..7069a6c27 100644
--- a/jarmuz/worker-paddler.mjs
+++ b/jarmuz/worker-paddler.mjs
@@ -19,6 +19,7 @@ spawner(async function ({ buildId, command }) {
     command(`
       target/debug/paddler balancer
         --compat-openai-addr 127.0.0.1:8063
+        --compat-typesafe-addr 127.0.0.1:8009
         --inference-addr 127.0.0.1:8061
         --inference-item-timeout 30000
         --management-addr 127.0.0.1:8060
diff --git a/package-lock.json b/package-lock.json
index 6cfd1c119..ac3eb41e3 100644
--- a/package-lock.json
+++ b/package-lock.json
@@ -10,50 +10,51 @@
         "paddler_client_javascript"
       ],
       "dependencies": {
-        "@codemirror/lang-jinja": "6.0.0",
-        "@uiw/react-codemirror": "4.24.2",
+        "@codemirror/lang-jinja": "6.0.1",
+        "@uiw/react-codemirror": "4.25.12",
         "clsx": "2.1.1",
-        "nanoid": "5.1.16",
-        "path-to-regexp": "8.4.2",
-        "react": "19.1.1",
-        "react-dom": "19.1.1",
+        "nanoid": "6.0.2",
+        "react": "19.3.1",
+        "react-dom": "19.3.1",
         "react-markdown": "10.1.0",
         "rxjs": "7.8.2",
-        "wouter": "3.7.1",
-        "zod": "4.0.17"
+        "wouter": "3.13.0",
+        "zod": "4.6.7"
       },
       "devDependencies": {
-        "@intentee/rust-coverage-check": "0.2.0",
-        "@types/hotwired__turbo": "8.0.4",
-        "@types/react": "19.1.10",
-        "@types/react-dom": "19.1.7",
-        "esbuild": "0.25.8",
-        "eslint": "9.33.0",
-        "eslint-plugin-react-hooks": "5.2.0",
-        "fs-extra": "11.3.1",
-        "glob": "11.1.0",
-        "jarmuz": "0.10.0",
+        "@eslint/js": "10.0.1",
+        "@intentee/rust-coverage-check": "0.3.1",
+        "@types/hotwired__turbo": "8.0.16",
+        "@types/react": "19.3.0",
+        "@types/react-dom": "19.3.0",
+        "esbuild": "0.28.2",
+        "eslint": "10.13.0",
+        "eslint-plugin-react-hooks": "7.1.1",
+        "fs-extra": "11.4.1",
+        "glob": "13.0.6",
+        "globals": "17.13.0",
+        "jarmuz": "0.11.0",
         "lodash": "4.18.1",
-        "node-notifier": "10.0.1",
-        "prettier": "3.6.2",
-        "prettier-plugin-organize-imports": "4.2.0",
-        "stylelint": "16.23.1",
-        "stylelint-config-recommended": "17.0.0",
-        "tempy": "3.1.0",
+        "node-notifier": "11.0.0",
+        "prettier": "3.9.10",
+        "prettier-plugin-organize-imports": "4.3.0",
+        "stylelint": "17.16.0",
+        "stylelint-config-recommended": "18.0.0",
+        "tempy": "3.2.0",
         "tslib": "2.8.1",
         "typed-css-modules": "0.9.1",
-        "typescript": "5.9.2",
-        "typescript-eslint": "8.39.1"
+        "typescript": "6.0.3",
+        "typescript-eslint": "8.72.0"
       }
     },
     "node_modules/@babel/code-frame": {
-      "version": "7.27.1",
-      "resolved": "https://registry.npmjs.org/@babel/code-frame/-/code-frame-7.27.1.tgz",
-      "integrity": "sha512-cjQ7ZlQ0Mv3b47hABuTevyTuYN4i+loJKGeV9flcCgIK37cCXRh+L1bd3iBHlynerhQ7BhCkn2BPbQUL+rGqFg==",
+      "version": "7.29.7",
+      "resolved": "https://registry.npmjs.org/@babel/code-frame/-/code-frame-7.29.7.tgz",
+      "integrity": "sha512-Aup7aUOfpbAUg2ROOJN6Iw5f9DMBlzu0mIkm/malLQFN/YQgO48wCj0Kxa3sEHJvPVFg7siR+qRInwXd2qhQKw==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
-        "@babel/helper-validator-identifier": "^7.27.1",
+        "@babel/helper-validator-identifier": "^7.29.7",
         "js-tokens": "^4.0.0",
         "picocolors": "^1.1.1"
       },
@@ -61,29 +62,268 @@
         "node": ">=6.9.0"
       }
     },
+    "node_modules/@babel/compat-data": {
+      "version": "7.29.7",
+      "resolved": "https://registry.npmjs.org/@babel/compat-data/-/compat-data-7.29.7.tgz",
+      "integrity": "sha512-locTkQyKvwIEgBzVrn8693ebc97F2U8ZHjbXwDXJ5Fn2TCpNwTlKcaKLkdHop5c/icOFE7qt7Q9JC5hnKNa6Gg==",
+      "dev": true,
+      "license": "MIT",
+      "engines": {
+        "node": ">=6.9.0"
+      }
+    },
+    "node_modules/@babel/core": {
+      "version": "7.29.7",
+      "resolved": "https://registry.npmjs.org/@babel/core/-/core-7.29.7.tgz",
+      "integrity": "sha512-RgHBCvtjbOK2gXSNBNIkNoEc9qoVEtau3hj8gEqKQuL3HZAibKarWFEI3Lfm6EYKkLalOh8eSrj9b+ch9H/VBA==",
+      "dev": true,
+      "license": "MIT",
+      "dependencies": {
+        "@babel/code-frame": "^7.29.7",
+        "@babel/generator": "^7.29.7",
+        "@babel/helper-compilation-targets": "^7.29.7",
+        "@babel/helper-module-transforms": "^7.29.7",
+        "@babel/helpers": "^7.29.7",
+        "@babel/parser": "^7.29.7",
+        "@babel/template": "^7.29.7",
+        "@babel/traverse": "^7.29.7",
+        "@babel/types": "^7.29.7",
+        "@jridgewell/remapping": "^2.3.5",
+        "convert-source-map": "^2.0.0",
+        "debug": "^4.1.0",
+        "gensync": "^1.0.0-beta.2",
+        "json5": "^2.2.3",
+        "semver": "^6.3.1"
+      },
+      "engines": {
+        "node": ">=6.9.0"
+      },
+      "funding": {
+        "type": "opencollective",
+        "url": "https://opencollective.com/babel"
+      }
+    },
+    "node_modules/@babel/generator": {
+      "version": "7.29.8",
+      "resolved": "https://registry.npmjs.org/@babel/generator/-/generator-7.29.8.tgz",
+      "integrity": "sha512-gZbepsdh3WDtgZKWL+vTPh71LSBrm/Y4/QDZBVCcYfmeTEEuoOYwlSy+G1StfJg+/Zy550u/3TATbm7qDbbMtg==",
+      "dev": true,
+      "license": "MIT",
+      "dependencies": {
+        "@babel/parser": "^7.29.8",
+        "@babel/types": "^7.29.8",
+        "@jridgewell/gen-mapping": "^0.3.12",
+        "@jridgewell/trace-mapping": "^0.3.28",
+        "jsesc": "^3.0.2"
+      },
+      "engines": {
+        "node": ">=6.9.0"
+      }
+    },
+    "node_modules/@babel/helper-compilation-targets": {
+      "version": "7.29.7",
+      "resolved": "https://registry.npmjs.org/@babel/helper-compilation-targets/-/helper-compilation-targets-7.29.7.tgz",
+      "integrity": "sha512-wem6WaBj4NaVYVdNhLPPVacES6ZJ+KBBfSkTMD3YZxbP3rm3Di85tJU5ljaUNhaOynt+Aj0xruhYuzQBt8n71g==",
+      "dev": true,
+      "license": "MIT",
+      "dependencies": {
+        "@babel/compat-data": "^7.29.7",
+        "@babel/helper-validator-option": "^7.29.7",
+        "browserslist": "^4.24.0",
+        "lru-cache": "^5.1.1",
+        "semver": "^6.3.1"
+      },
+      "engines": {
+        "node": ">=6.9.0"
+      }
+    },
+    "node_modules/@babel/helper-globals": {
+      "version": "7.29.7",
+      "resolved": "https://registry.npmjs.org/@babel/helper-globals/-/helper-globals-7.29.7.tgz",
+      "integrity": "sha512-3nQVUAtvkKH9zahfWgw96Jc/uFOmjACE1kQz82E2lqWmHBgjzbNlsC22nuQTfahmWeQtTq5nQ/4Nnd2A1wj4zA==",
+      "dev": true,
+      "license": "MIT",
+      "engines": {
+        "node": ">=6.9.0"
+      }
+    },
+    "node_modules/@babel/helper-module-imports": {
+      "version": "7.29.7",
+      "resolved": "https://registry.npmjs.org/@babel/helper-module-imports/-/helper-module-imports-7.29.7.tgz",
+      "integrity": "sha512-ejHwrQQYcm9xnTivShn2IDOlIzInN34AXskvq9QicvCtEzq1Vzclu/tKF8Jq1Cg8JG2GL6/EmjgsCT7lXepE3g==",
+      "dev": true,
+      "license": "MIT",
+      "dependencies": {
+        "@babel/traverse": "^7.29.7",
+        "@babel/types": "^7.29.7"
+      },
+      "engines": {
+        "node": ">=6.9.0"
+      }
+    },
+    "node_modules/@babel/helper-module-transforms": {
+      "version": "7.29.7",
+      "resolved": "https://registry.npmjs.org/@babel/helper-module-transforms/-/helper-module-transforms-7.29.7.tgz",
+      "integrity": "sha512-UPUVSyXbOh627KiCIGQSgwWzGeBKLkaJ9PJEdrngIwMSzxLR4jS4+f1f1jb7VzBbg8nFLaYotvVPFCTqdrmTAg==",
+      "dev": true,
+      "license": "MIT",
+      "dependencies": {
+        "@babel/helper-module-imports": "^7.29.7",
+        "@babel/helper-validator-identifier": "^7.29.7",
+        "@babel/traverse": "^7.29.7"
+      },
+      "engines": {
+        "node": ">=6.9.0"
+      },
+      "peerDependencies": {
+        "@babel/core": "^7.0.0"
+      }
+    },
+    "node_modules/@babel/helper-string-parser": {
+      "version": "7.29.11",
+      "resolved": "https://registry.npmjs.org/@babel/helper-string-parser/-/helper-string-parser-7.29.11.tgz",
+      "integrity": "sha512-snZbO7MEDOof+hvNa97EjEY7qD7h2kwvQr5AJC3ZhQudKW5dnej6k3kP9NQV88hJwaS8vDqRLFo9rivLEhle2g==",
+      "dev": true,
+      "license": "MIT",
+      "engines": {
+        "node": ">=6.9.0"
+      }
+    },
     "node_modules/@babel/helper-validator-identifier": {
-      "version": "7.27.1",
-      "resolved": "https://registry.npmjs.org/@babel/helper-validator-identifier/-/helper-validator-identifier-7.27.1.tgz",
-      "integrity": "sha512-D2hP9eA+Sqx1kBZgzxZh0y1trbuU+JoDkiEwqhQ36nodYqJwyEIhPSdMNd7lOm/4io72luTPWH20Yda0xOuUow==",
+      "version": "7.29.7",
+      "resolved": "https://registry.npmjs.org/@babel/helper-validator-identifier/-/helper-validator-identifier-7.29.7.tgz",
+      "integrity": "sha512-qehxGkRj55h/ff8EMaJ+cYhyaKlHIxqYDn682wQD7RNp9UujOQsHog2uS0r2vzr4pW+sXf90NeeayjcNaX3fFg==",
+      "dev": true,
+      "license": "MIT",
+      "engines": {
+        "node": ">=6.9.0"
+      }
+    },
+    "node_modules/@babel/helper-validator-option": {
+      "version": "7.29.7",
+      "resolved": "https://registry.npmjs.org/@babel/helper-validator-option/-/helper-validator-option-7.29.7.tgz",
+      "integrity": "sha512-N9ZErrD+yW5geCDtBqnOoxmR8+tNKiGuxKlDpuJxfsqpa2dFcexaziGAE/qoHLiDDreVNMupxGmSoNlyvsA3gw==",
+      "dev": true,
+      "license": "MIT",
+      "engines": {
+        "node": ">=6.9.0"
+      }
+    },
+    "node_modules/@babel/helpers": {
+      "version": "7.29.10",
+      "resolved": "https://registry.npmjs.org/@babel/helpers/-/helpers-7.29.10.tgz",
+      "integrity": "sha512-EMCx3AGZ3WfAl6ht7AyQoPm2Lh+BGYEMdc4Q6Hf7uSCDOZFuLq44qNwL+iaim8VlyYhRAo4X43+weadK8ZB4Yg==",
       "dev": true,
       "license": "MIT",
+      "dependencies": {
+        "@babel/template": "^7.29.7",
+        "@babel/types": "^7.29.8"
+      },
       "engines": {
         "node": ">=6.9.0"
       }
     },
+    "node_modules/@babel/parser": {
+      "version": "7.29.11",
+      "resolved": "https://registry.npmjs.org/@babel/parser/-/parser-7.29.11.tgz",
+      "integrity": "sha512-OZ6+oaK7GMyg3IfCpj8JOmn8n+w2rJbDT0LLfVDTLb1ZFeKDh3AiNELgCMUw175IJHSkLl74hG0M0VLsvx+gag==",
+      "dev": true,
+      "license": "MIT",
+      "dependencies": {
+        "@babel/types": "^7.29.11"
+      },
+      "bin": {
+        "parser": "bin/babel-parser.js"
+      },
+      "engines": {
+        "node": ">=6.0.0"
+      }
+    },
     "node_modules/@babel/runtime": {
-      "version": "7.28.2",
-      "resolved": "https://registry.npmjs.org/@babel/runtime/-/runtime-7.28.2.tgz",
-      "integrity": "sha512-KHp2IflsnGywDjBWDkR9iEqiWSpc8GIi0lgTT3mOElT0PP1tG26P4tmFI2YvAdzgq9RGyoHZQEIEdZy6Ec5xCA==",
+      "version": "7.29.10",
+      "resolved": "https://registry.npmjs.org/@babel/runtime/-/runtime-7.29.10.tgz",
+      "integrity": "sha512-NpugTzbKgGrNe45Ui32cN8vCFFMxVtP/9rtfwv0nP5FRDAjvYBxJcLVpYA9mr1jxP6cnIT7x9LUzuDTbWYYNRA==",
+      "license": "MIT",
+      "engines": {
+        "node": ">=6.9.0"
+      }
+    },
+    "node_modules/@babel/template": {
+      "version": "7.29.7",
+      "resolved": "https://registry.npmjs.org/@babel/template/-/template-7.29.7.tgz",
+      "integrity": "sha512-puq+Gf35oI24FeN11LkoUQFqv9uwNeWpxXZi/Ji3rRIoKAzKnxRaZ+Gkj0vKS9ZCiTESfng1N9LyOyXvo+m+Gg==",
+      "dev": true,
+      "license": "MIT",
+      "dependencies": {
+        "@babel/code-frame": "^7.29.7",
+        "@babel/parser": "^7.29.7",
+        "@babel/types": "^7.29.7"
+      },
+      "engines": {
+        "node": ">=6.9.0"
+      }
+    },
+    "node_modules/@babel/traverse": {
+      "version": "7.29.10",
+      "resolved": "https://registry.npmjs.org/@babel/traverse/-/traverse-7.29.10.tgz",
+      "integrity": "sha512-B3lYAs2nkKskbgQ1nfy8Ma9ao9Q1rntTAQHrzzyDuj+TSO+X3le6PfWwqDvodsb+OumE10eMFfuMCTQSQ5iuIQ==",
+      "dev": true,
+      "license": "MIT",
+      "dependencies": {
+        "@babel/code-frame": "^7.29.7",
+        "@babel/generator": "^7.29.8",
+        "@babel/helper-globals": "^7.29.7",
+        "@babel/parser": "^7.29.9",
+        "@babel/template": "^7.29.7",
+        "@babel/types": "^7.29.8",
+        "debug": "^4.3.1"
+      },
+      "engines": {
+        "node": ">=6.9.0"
+      }
+    },
+    "node_modules/@babel/types": {
+      "version": "7.29.11",
+      "resolved": "https://registry.npmjs.org/@babel/types/-/types-7.29.11.tgz",
+      "integrity": "sha512-WCsemW40Cpc6gouC1SrY6PBaZ+oLRlEVY9XBvZBI7il/HyUrb2nJ505wE1MHEFyDZNnrWDEuy4jL50WVHfTTSQ==",
+      "dev": true,
       "license": "MIT",
+      "dependencies": {
+        "@babel/helper-string-parser": "^7.29.11",
+        "@babel/helper-validator-identifier": "^7.29.7"
+      },
       "engines": {
         "node": ">=6.9.0"
       }
     },
+    "node_modules/@cacheable/memory": {
+      "version": "2.2.0",
+      "resolved": "https://registry.npmjs.org/@cacheable/memory/-/memory-2.2.0.tgz",
+      "integrity": "sha512-CTLKqLItRCEixEAewD3/j9DB3/o96gpTPD4eJ1v+DGOlxZRZncRQkGYqqnAGCscYd6RNeXfGeiuCphsPtqyIfQ==",
+      "dev": true,
+      "license": "MIT",
+      "dependencies": {
+        "@cacheable/utils": "^2.5.0",
+        "@keyv/bigmap": "^1.3.1",
+        "hookified": "^1.15.1",
+        "keyv": "^5.6.0"
+      }
+    },
+    "node_modules/@cacheable/utils": {
+      "version": "2.5.0",
+      "resolved": "https://registry.npmjs.org/@cacheable/utils/-/utils-2.5.0.tgz",
+      "integrity": "sha512-buipgOVDkkPXNR5+xBpDw7Zk2n1EvU7qBJCNUcL7rhQ//kfpOXPAvQ511Os0vpLYJ1pZnvudNytkQt2hst3wqA==",
+      "dev": true,
+      "license": "MIT",
+      "dependencies": {
+        "hashery": "^1.5.1",
+        "keyv": "^5.6.0"
+      }
+    },
     "node_modules/@codemirror/autocomplete": {
-      "version": "6.18.6",
-      "resolved": "https://registry.npmjs.org/@codemirror/autocomplete/-/autocomplete-6.18.6.tgz",
-      "integrity": "sha512-PHHBXFomUs5DF+9tCOM/UoW6XQ4R44lLNNhRaW9PKPTU0D7lIjRg3ElxaJnTwsl/oHiR93WSXDBrekhoUGCPtg==",
+      "version": "6.20.4",
+      "resolved": "https://registry.npmjs.org/@codemirror/autocomplete/-/autocomplete-6.20.4.tgz",
+      "integrity": "sha512-yF8RhOf2SXlg00+qtupzLhrglcas61boqaONiNcbPIebHCpJocpqTBA6dyrtUrQO47hqyTdeob/3XnYilT7TaQ==",
       "license": "MIT",
       "dependencies": {
         "@codemirror/language": "^6.0.0",
@@ -93,13 +333,13 @@
       }
     },
     "node_modules/@codemirror/commands": {
-      "version": "6.8.1",
-      "resolved": "https://registry.npmjs.org/@codemirror/commands/-/commands-6.8.1.tgz",
-      "integrity": "sha512-KlGVYufHMQzxbdQONiLyGQDUW0itrLZwq3CcY7xpv9ZLRHqzkBSoteocBHtMCoY7/Ci4xhzSrToIeLg7FxHuaw==",
+      "version": "6.11.1",
+      "resolved": "https://registry.npmjs.org/@codemirror/commands/-/commands-6.11.1.tgz",
+      "integrity": "sha512-O/4hG3SC1YwcmQ0d2UVNDs+AsaNWd1iHVxbTeEBuqH+6bExAiPK3iS/BvpY6rZGURALv4ZD3sIgcCmRvw3ehBg==",
       "license": "MIT",
       "dependencies": {
         "@codemirror/language": "^6.0.0",
-        "@codemirror/state": "^6.4.0",
+        "@codemirror/state": "^6.7.0",
         "@codemirror/view": "^6.27.0",
         "@lezer/common": "^1.1.0"
       }
@@ -118,9 +358,9 @@
       }
     },
     "node_modules/@codemirror/lang-html": {
-      "version": "6.4.9",
-      "resolved": "https://registry.npmjs.org/@codemirror/lang-html/-/lang-html-6.4.9.tgz",
-      "integrity": "sha512-aQv37pIMSlueybId/2PVSP6NPnmurFDVmZwzc7jszd2KAF8qd4VBbvNYPXWQq90WIARjsdVkPbw29pszmHws3Q==",
+      "version": "6.4.12",
+      "resolved": "https://registry.npmjs.org/@codemirror/lang-html/-/lang-html-6.4.12.tgz",
+      "integrity": "sha512-pw2ReWKUqSkbvh76RAT4NYxiogRu+PWkR2ukAwO9uOgrm8uipkzjtKKtNpyeAQwHOqxEeSvAXZ6vr3AfyB9y/w==",
       "license": "MIT",
       "dependencies": {
         "@codemirror/autocomplete": "^6.0.0",
@@ -131,13 +371,13 @@
         "@codemirror/view": "^6.17.0",
         "@lezer/common": "^1.0.0",
         "@lezer/css": "^1.1.0",
-        "@lezer/html": "^1.3.0"
+        "@lezer/html": "^1.3.12"
       }
     },
     "node_modules/@codemirror/lang-javascript": {
-      "version": "6.2.4",
-      "resolved": "https://registry.npmjs.org/@codemirror/lang-javascript/-/lang-javascript-6.2.4.tgz",
-      "integrity": "sha512-0WVmhp1QOqZ4Rt6GlVGwKJN3KW7Xh4H2q8ZZNGZaP6lRdxXJzmjm4FqvmOojVj6khWJHIb9sp7U/72W7xQgqAA==",
+      "version": "6.2.5",
+      "resolved": "https://registry.npmjs.org/@codemirror/lang-javascript/-/lang-javascript-6.2.5.tgz",
+      "integrity": "sha512-zD4e5mS+50htS7F+TYjBPsiIFGanfVqg4HyUz6WNFikgOPf2BgKlx+TQedI1w6n/IqRBVBbBWmGFdLB/7uxO4A==",
       "license": "MIT",
       "dependencies": {
         "@codemirror/autocomplete": "^6.0.0",
@@ -150,63 +390,73 @@
       }
     },
     "node_modules/@codemirror/lang-jinja": {
-      "version": "6.0.0",
-      "resolved": "https://registry.npmjs.org/@codemirror/lang-jinja/-/lang-jinja-6.0.0.tgz",
-      "integrity": "sha512-47MFmRcR8UAxd8DReVgj7WJN1WSAMT7OJnewwugZM4XiHWkOjgJQqvEM1NpMj9ALMPyxmlziEI1opH9IaEvmaw==",
+      "version": "6.0.1",
+      "resolved": "https://registry.npmjs.org/@codemirror/lang-jinja/-/lang-jinja-6.0.1.tgz",
+      "integrity": "sha512-P5kyHLObzjtbGj16h+hyvZTxJhSjBEeSx4wMjbnAf3b0uwTy2+F0zGjMZL4PQOm/mh2eGZ5xUDVZXgwP783Nsw==",
       "license": "MIT",
       "dependencies": {
+        "@codemirror/autocomplete": "^6.0.0",
         "@codemirror/lang-html": "^6.0.0",
         "@codemirror/language": "^6.0.0",
+        "@codemirror/state": "^6.0.0",
+        "@codemirror/view": "^6.0.0",
         "@lezer/common": "^1.2.0",
         "@lezer/highlight": "^1.2.0",
         "@lezer/lr": "^1.4.0"
       }
     },
     "node_modules/@codemirror/language": {
-      "version": "6.11.2",
-      "resolved": "https://registry.npmjs.org/@codemirror/language/-/language-6.11.2.tgz",
-      "integrity": "sha512-p44TsNArL4IVXDTbapUmEkAlvWs2CFQbcfc0ymDsis1kH2wh0gcY96AS29c/vp2d0y2Tquk1EDSaawpzilUiAw==",
+      "version": "6.13.1",
+      "resolved": "https://registry.npmjs.org/@codemirror/language/-/language-6.13.1.tgz",
+      "integrity": "sha512-zfvU2r3TyylQLMj/te+5CnQ0rSqZi8B1oTCBbQkagcLW4Xc6imab6qN+pEJhsuNOLaJqiRyBXp9UhWA1DhVvDw==",
       "license": "MIT",
       "dependencies": {
         "@codemirror/state": "^6.0.0",
+        "@codemirror/streamparser": "^6.0.0",
         "@codemirror/view": "^6.23.0",
-        "@lezer/common": "^1.1.0",
+        "@lezer/common": "^1.5.0",
         "@lezer/highlight": "^1.0.0",
         "@lezer/lr": "^1.0.0",
         "style-mod": "^4.0.0"
       }
     },
     "node_modules/@codemirror/lint": {
-      "version": "6.8.5",
-      "resolved": "https://registry.npmjs.org/@codemirror/lint/-/lint-6.8.5.tgz",
-      "integrity": "sha512-s3n3KisH7dx3vsoeGMxsbRAgKe4O1vbrnKBClm99PU0fWxmxsx5rR2PfqQgIt+2MMJBHbiJ5rfIdLYfB9NNvsA==",
+      "version": "6.9.7",
+      "resolved": "https://registry.npmjs.org/@codemirror/lint/-/lint-6.9.7.tgz",
+      "integrity": "sha512-28/+iWLYxKxsvGYhSYL7zaCZqLz5+FFFDq9tVsvGv9kv8RY4fFAchJ5WX9M3YrrRlTIsECjsXPqeNgnSmNP2dg==",
       "license": "MIT",
       "dependencies": {
         "@codemirror/state": "^6.0.0",
-        "@codemirror/view": "^6.35.0",
+        "@codemirror/view": "^6.42.0",
         "crelt": "^1.0.5"
       }
     },
     "node_modules/@codemirror/search": {
-      "version": "6.5.11",
-      "resolved": "https://registry.npmjs.org/@codemirror/search/-/search-6.5.11.tgz",
-      "integrity": "sha512-KmWepDE6jUdL6n8cAAqIpRmLPBZ5ZKnicE8oGU/s3QrAVID+0VhLFrzUucVKHG5035/BSykhExDL/Xm7dHthiA==",
+      "version": "6.7.2",
+      "resolved": "https://registry.npmjs.org/@codemirror/search/-/search-6.7.2.tgz",
+      "integrity": "sha512-gUYkYhT2+n/+VGZ+8EzE5WFkYZUZYm1VOKDudIsNqh42uRVQJ0a6Yss9sdKT3MeOYfuL1N6AZA57oza0Oyr0LA==",
       "license": "MIT",
       "dependencies": {
         "@codemirror/state": "^6.0.0",
-        "@codemirror/view": "^6.0.0",
+        "@codemirror/view": "^6.37.0",
         "crelt": "^1.0.5"
       }
     },
     "node_modules/@codemirror/state": {
-      "version": "6.5.2",
-      "resolved": "https://registry.npmjs.org/@codemirror/state/-/state-6.5.2.tgz",
-      "integrity": "sha512-FVqsPqtPWKVVL3dPSxy8wEF/ymIEuVzF1PK3VbUgrxXpJUSHQWWZz4JMToquRxnkw+36LTamCZG2iua2Ptq0fA==",
+      "version": "6.7.6",
+      "resolved": "https://registry.npmjs.org/@codemirror/state/-/state-6.7.6.tgz",
+      "integrity": "sha512-kAz+AncRtKuIknedxT1bq4XwXv4UowhbkHU1myPrtVb/jZtImWuV5BXzv5vK6i3kYACsdiZiQKFQQ5Mq7elW8w==",
       "license": "MIT",
       "dependencies": {
         "@marijn/find-cluster-break": "^1.0.0"
       }
     },
+    "node_modules/@codemirror/streamparser": {
+      "version": "6.0.0",
+      "resolved": "https://registry.npmjs.org/@codemirror/streamparser/-/streamparser-6.0.0.tgz",
+      "integrity": "sha512-JTFPF6xrfjt0mnfGPgzcLiMwYmvIyNaoizp8BYzv8FE8KrT6GBGDZ190rtt+O7Wke3gG/MFHyiY9YGNeimof4w==",
+      "license": "MIT"
+    },
     "node_modules/@codemirror/theme-one-dark": {
       "version": "6.1.3",
       "resolved": "https://registry.npmjs.org/@codemirror/theme-one-dark/-/theme-one-dark-6.1.3.tgz",
@@ -220,21 +470,45 @@
       }
     },
     "node_modules/@codemirror/view": {
-      "version": "6.38.1",
-      "resolved": "https://registry.npmjs.org/@codemirror/view/-/view-6.38.1.tgz",
-      "integrity": "sha512-RmTOkE7hRU3OVREqFVITWHz6ocgBjv08GoePscAakgVQfciA3SGCEk7mb9IzwW61cKKmlTpHXG6DUE5Ubx+MGQ==",
+      "version": "6.43.15",
+      "resolved": "https://registry.npmjs.org/@codemirror/view/-/view-6.43.15.tgz",
+      "integrity": "sha512-HHsfpqFeHyXLs983pSRe38Xajs0vFWZQ2BuX0m7Ki8w3jZ5iD8cuR3lBz2/5/X53Dhf32SBHKh6rtSTB6tKeGw==",
       "license": "MIT",
       "dependencies": {
-        "@codemirror/state": "^6.5.0",
+        "@codemirror/state": "^6.7.0",
         "crelt": "^1.0.6",
         "style-mod": "^4.1.0",
         "w3c-keyname": "^2.2.4"
       }
     },
+    "node_modules/@csstools/css-calc": {
+      "version": "3.4.3",
+      "resolved": "https://registry.npmjs.org/@csstools/css-calc/-/css-calc-3.4.3.tgz",
+      "integrity": "sha512-iex20d8CHVkyvg6B7UKV7uHnI2Bqo9g+EFfT9E0y+GvTvhZ/DwONJ+9aKb1dlqm0ZiGsL5RXjp0fCoJYnkeDjA==",
+      "dev": true,
+      "funding": [
+        {
+          "type": "github",
+          "url": "https://github.com/sponsors/csstools"
+        },
+        {
+          "type": "opencollective",
+          "url": "https://opencollective.com/csstools"
+        }
+      ],
+      "license": "MIT",
+      "engines": {
+        "node": ">=20.19.0"
+      },
+      "peerDependencies": {
+        "@csstools/css-parser-algorithms": "^4.0.2",
+        "@csstools/css-tokenizer": "^4.0.2"
+      }
+    },
     "node_modules/@csstools/css-parser-algorithms": {
-      "version": "3.0.5",
-      "resolved": "https://registry.npmjs.org/@csstools/css-parser-algorithms/-/css-parser-algorithms-3.0.5.tgz",
-      "integrity": "sha512-DaDeUkXZKjdGhgYaHNJTV9pV7Y9B3b644jCLs9Upc3VeNGg6LWARAT6O+Q+/COo+2gg/bM5rhpMAtf70WqfBdQ==",
+      "version": "4.0.2",
+      "resolved": "https://registry.npmjs.org/@csstools/css-parser-algorithms/-/css-parser-algorithms-4.0.2.tgz",
+      "integrity": "sha512-40cSKyMvK+tq4qz6Awrlye2WGuOKt3FwPgtGg6KTfbHOWNw+Rk1rzbAtZnZ6IBhsY491HLRnDXwoyBAijmmILA==",
       "dev": true,
       "funding": [
         {
@@ -248,16 +522,41 @@
       ],
       "license": "MIT",
       "engines": {
-        "node": ">=18"
+        "node": ">=20.19.0"
       },
       "peerDependencies": {
-        "@csstools/css-tokenizer": "^3.0.4"
+        "@csstools/css-tokenizer": "^4.0.2"
+      }
+    },
+    "node_modules/@csstools/css-syntax-patches-for-csstree": {
+      "version": "1.1.15",
+      "resolved": "https://registry.npmjs.org/@csstools/css-syntax-patches-for-csstree/-/css-syntax-patches-for-csstree-1.1.15.tgz",
+      "integrity": "sha512-J0u7HkVl2nzSlhsiTOp4AmwcUQ3D+mGEEKfBy/7To5/y7F2OHwyLrXfrhR0SMgr4p5Lo+eaMVSeai24zUcBIxA==",
+      "dev": true,
+      "funding": [
+        {
+          "type": "github",
+          "url": "https://github.com/sponsors/csstools"
+        },
+        {
+          "type": "opencollective",
+          "url": "https://opencollective.com/csstools"
+        }
+      ],
+      "license": "MIT-0",
+      "peerDependencies": {
+        "css-tree": "^3.2.1"
+      },
+      "peerDependenciesMeta": {
+        "css-tree": {
+          "optional": true
+        }
       }
     },
     "node_modules/@csstools/css-tokenizer": {
-      "version": "3.0.4",
-      "resolved": "https://registry.npmjs.org/@csstools/css-tokenizer/-/css-tokenizer-3.0.4.tgz",
-      "integrity": "sha512-Vd/9EVDiu6PPJt9yAh6roZP6El1xHrdvIVGjyBsHR0RYwNHgL7FJPyIIW4fANJNG6FtyZfvlRPpFI4ZM/lubvw==",
+      "version": "4.0.2",
+      "resolved": "https://registry.npmjs.org/@csstools/css-tokenizer/-/css-tokenizer-4.0.2.tgz",
+      "integrity": "sha512-OoKoR0f76dCY666JlcbhmVTs2drYj1GUXZTYTcbUgJjh9Nv41aFfZ21bPQTERm5+L5cBDo466NltB2lplS5GBw==",
       "dev": true,
       "funding": [
         {
@@ -271,13 +570,13 @@
       ],
       "license": "MIT",
       "engines": {
-        "node": ">=18"
+        "node": ">=20.19.0"
       }
     },
     "node_modules/@csstools/media-query-list-parser": {
-      "version": "4.0.3",
-      "resolved": "https://registry.npmjs.org/@csstools/media-query-list-parser/-/media-query-list-parser-4.0.3.tgz",
-      "integrity": "sha512-HAYH7d3TLRHDOUQK4mZKf9k9Ph/m8Akstg66ywKR4SFAigjs3yBiUeZtFxywiTm5moZMAp/5W/ZuFnNXXYLuuQ==",
+      "version": "5.0.2",
+      "resolved": "https://registry.npmjs.org/@csstools/media-query-list-parser/-/media-query-list-parser-5.0.2.tgz",
+      "integrity": "sha512-qx19O8AbbLLbDSAcI6YOby/Vq8rV51psgIvlF5OZT/L8fQJDCyQifL6k+kYU4a8ySpzJ4RBzZkIs530bAn8y4Q==",
       "dev": true,
       "funding": [
         {
@@ -291,17 +590,17 @@
       ],
       "license": "MIT",
       "engines": {
-        "node": ">=18"
+        "node": ">=20.19.0"
       },
       "peerDependencies": {
-        "@csstools/css-parser-algorithms": "^3.0.5",
-        "@csstools/css-tokenizer": "^3.0.4"
+        "@csstools/css-parser-algorithms": "^4.0.2",
+        "@csstools/css-tokenizer": "^4.0.2"
       }
     },
-    "node_modules/@csstools/selector-specificity": {
-      "version": "5.0.0",
-      "resolved": "https://registry.npmjs.org/@csstools/selector-specificity/-/selector-specificity-5.0.0.tgz",
-      "integrity": "sha512-PCqQV3c4CoVm3kdPhyeZ07VmBRdH2EpMFA/pd9OASpOEC3aXNGoqPDAZ80D0cLpMBxnmk0+yNhGsEx31hq7Gtw==",
+    "node_modules/@csstools/selector-resolve-nested": {
+      "version": "4.0.2",
+      "resolved": "https://registry.npmjs.org/@csstools/selector-resolve-nested/-/selector-resolve-nested-4.0.2.tgz",
+      "integrity": "sha512-JpjNBJFA+4OnAHKRK4AMp8xrYyLhI9XapOeBlP2PhRU/IylSw+1ep3BFNTksBJnrTS3KT6kqKNVfpzLLwZUW5Q==",
       "dev": true,
       "funding": [
         {
@@ -315,27 +614,39 @@
       ],
       "license": "MIT-0",
       "engines": {
-        "node": ">=18"
+        "node": ">=20.19.0"
       },
       "peerDependencies": {
-        "postcss-selector-parser": "^7.0.0"
+        "postcss-selector-parser": "^7.1.1"
       }
     },
-    "node_modules/@dual-bundle/import-meta-resolve": {
-      "version": "4.1.0",
-      "resolved": "https://registry.npmjs.org/@dual-bundle/import-meta-resolve/-/import-meta-resolve-4.1.0.tgz",
-      "integrity": "sha512-+nxncfwHM5SgAtrVzgpzJOI1ol0PkumhVo469KCf9lUi21IGcY90G98VuHm9VRrUypmAzawAHO9bs6hqeADaVg==",
+    "node_modules/@csstools/selector-specificity": {
+      "version": "6.0.1",
+      "resolved": "https://registry.npmjs.org/@csstools/selector-specificity/-/selector-specificity-6.0.1.tgz",
+      "integrity": "sha512-MBuqFIIvqk4S593MGAskTkPpEUUzZNvGzh8mMxEwl3oHwjQlXmqHEDxNbN3Y0Tsh+xjmem4GZjeg7I0qkwBivg==",
       "dev": true,
-      "license": "MIT",
-      "funding": {
-        "type": "github",
-        "url": "https://github.com/sponsors/wooorm"
+      "funding": [
+        {
+          "type": "github",
+          "url": "https://github.com/sponsors/csstools"
+        },
+        {
+          "type": "opencollective",
+          "url": "https://opencollective.com/csstools"
+        }
+      ],
+      "license": "MIT-0",
+      "engines": {
+        "node": ">=20.19.0"
+      },
+      "peerDependencies": {
+        "postcss-selector-parser": "^7.1.1"
       }
     },
     "node_modules/@esbuild/aix-ppc64": {
-      "version": "0.25.8",
-      "resolved": "https://registry.npmjs.org/@esbuild/aix-ppc64/-/aix-ppc64-0.25.8.tgz",
-      "integrity": "sha512-urAvrUedIqEiFR3FYSLTWQgLu5tb+m0qZw0NBEasUeo6wuqatkMDaRT+1uABiGXEu5vqgPd7FGE1BhsAIy9QVA==",
+      "version": "0.28.2",
+      "resolved": "https://registry.npmjs.org/@esbuild/aix-ppc64/-/aix-ppc64-0.28.2.tgz",
+      "integrity": "sha512-XExcO+dvLKvVtNTibSTBej1NCAbaGhWn9Ww1ZPx80qsahhPFe/8jgWP0IchNe0F3HwkU7n8ejhH8bjonqht8mQ==",
       "cpu": [
         "ppc64"
       ],
@@ -350,9 +661,9 @@
       }
     },
     "node_modules/@esbuild/android-arm": {
-      "version": "0.25.8",
-      "resolved": "https://registry.npmjs.org/@esbuild/android-arm/-/android-arm-0.25.8.tgz",
-      "integrity": "sha512-RONsAvGCz5oWyePVnLdZY/HHwA++nxYWIX1atInlaW6SEkwq6XkP3+cb825EUcRs5Vss/lGh/2YxAb5xqc07Uw==",
+      "version": "0.28.2",
+      "resolved": "https://registry.npmjs.org/@esbuild/android-arm/-/android-arm-0.28.2.tgz",
+      "integrity": "sha512-kXXoiPVVGQcnIYGOeaovwOURpniDBpSq4A03qkQ+BMQqtGG6HYap3xne9C1O1yo4TR3qxlCX5IqqmX6fFo2Lqg==",
       "cpu": [
         "arm"
       ],
@@ -367,9 +678,9 @@
       }
     },
     "node_modules/@esbuild/android-arm64": {
-      "version": "0.25.8",
-      "resolved": "https://registry.npmjs.org/@esbuild/android-arm64/-/android-arm64-0.25.8.tgz",
-      "integrity": "sha512-OD3p7LYzWpLhZEyATcTSJ67qB5D+20vbtr6vHlHWSQYhKtzUYrETuWThmzFpZtFsBIxRvhO07+UgVA9m0i/O1w==",
+      "version": "0.28.2",
+      "resolved": "https://registry.npmjs.org/@esbuild/android-arm64/-/android-arm64-0.28.2.tgz",
+      "integrity": "sha512-5YfKeeI8qWfBZIX+u2xZC3Zlb3Os/gLS2sbEKM+I4ZOcsWmHS2WLysCcQZDAFRslDUU5Oiq44gf6PYN1vGwG5A==",
       "cpu": [
         "arm64"
       ],
@@ -384,9 +695,9 @@
       }
     },
     "node_modules/@esbuild/android-x64": {
-      "version": "0.25.8",
-      "resolved": "https://registry.npmjs.org/@esbuild/android-x64/-/android-x64-0.25.8.tgz",
-      "integrity": "sha512-yJAVPklM5+4+9dTeKwHOaA+LQkmrKFX96BM0A/2zQrbS6ENCmxc4OVoBs5dPkCCak2roAD+jKCdnmOqKszPkjA==",
+      "version": "0.28.2",
+      "resolved": "https://registry.npmjs.org/@esbuild/android-x64/-/android-x64-0.28.2.tgz",
+      "integrity": "sha512-O387ite7SzUyCcy3JQX4P4bLtEA7bLLkx+esve5JHnyYfNTxcVpXZo9jhdB0lTKN44gztELTdU7nS8Nr16Fs1Q==",
       "cpu": [
         "x64"
       ],
@@ -401,9 +712,9 @@
       }
     },
     "node_modules/@esbuild/darwin-arm64": {
-      "version": "0.25.8",
-      "resolved": "https://registry.npmjs.org/@esbuild/darwin-arm64/-/darwin-arm64-0.25.8.tgz",
-      "integrity": "sha512-Jw0mxgIaYX6R8ODrdkLLPwBqHTtYHJSmzzd+QeytSugzQ0Vg4c5rDky5VgkoowbZQahCbsv1rT1KW72MPIkevw==",
+      "version": "0.28.2",
+      "resolved": "https://registry.npmjs.org/@esbuild/darwin-arm64/-/darwin-arm64-0.28.2.tgz",
+      "integrity": "sha512-n4KqkOQrraxHJcgjM1RvwbigfQKIKJVpM7xp+KsxiyUSrRdIXnt73VhrPAx0fV44hgfmIVKjxMN9J1t5jySVkw==",
       "cpu": [
         "arm64"
       ],
@@ -418,9 +729,9 @@
       }
     },
     "node_modules/@esbuild/darwin-x64": {
-      "version": "0.25.8",
-      "resolved": "https://registry.npmjs.org/@esbuild/darwin-x64/-/darwin-x64-0.25.8.tgz",
-      "integrity": "sha512-Vh2gLxxHnuoQ+GjPNvDSDRpoBCUzY4Pu0kBqMBDlK4fuWbKgGtmDIeEC081xi26PPjn+1tct+Bh8FjyLlw1Zlg==",
+      "version": "0.28.2",
+      "resolved": "https://registry.npmjs.org/@esbuild/darwin-x64/-/darwin-x64-0.28.2.tgz",
+      "integrity": "sha512-uq6suIWYP37qzGddBKPw5QEQPi6HiLGsO7UmkpfyaYNQ3D+rN6w6WfwH+nuqcGXWvawGwxOEroO4YGnFh95azw==",
       "cpu": [
         "x64"
       ],
@@ -435,9 +746,9 @@
       }
     },
     "node_modules/@esbuild/freebsd-arm64": {
-      "version": "0.25.8",
-      "resolved": "https://registry.npmjs.org/@esbuild/freebsd-arm64/-/freebsd-arm64-0.25.8.tgz",
-      "integrity": "sha512-YPJ7hDQ9DnNe5vxOm6jaie9QsTwcKedPvizTVlqWG9GBSq+BuyWEDazlGaDTC5NGU4QJd666V0yqCBL2oWKPfA==",
+      "version": "0.28.2",
+      "resolved": "https://registry.npmjs.org/@esbuild/freebsd-arm64/-/freebsd-arm64-0.28.2.tgz",
+      "integrity": "sha512-n+I0BTSRIoy+d6RPKnEVwql5UwBJolytvY4mAOIEJorKlqgPII8ix6slVVrfZ5Tnj7glIZvloylbB/EJPMWEXw==",
       "cpu": [
         "arm64"
       ],
@@ -452,9 +763,9 @@
       }
     },
     "node_modules/@esbuild/freebsd-x64": {
-      "version": "0.25.8",
-      "resolved": "https://registry.npmjs.org/@esbuild/freebsd-x64/-/freebsd-x64-0.25.8.tgz",
-      "integrity": "sha512-MmaEXxQRdXNFsRN/KcIimLnSJrk2r5H8v+WVafRWz5xdSVmWLoITZQXcgehI2ZE6gioE6HirAEToM/RvFBeuhw==",
+      "version": "0.28.2",
+      "resolved": "https://registry.npmjs.org/@esbuild/freebsd-x64/-/freebsd-x64-0.28.2.tgz",
+      "integrity": "sha512-78XJTJkvPs0kz2w61301PJjXl4g7q3JqiYMZ/M/yVI73EHBrCRTgkhu9oqG7vPqq+a/yadEW8aD+agKlk5xrmg==",
       "cpu": [
         "x64"
       ],
@@ -469,9 +780,9 @@
       }
     },
     "node_modules/@esbuild/linux-arm": {
-      "version": "0.25.8",
-      "resolved": "https://registry.npmjs.org/@esbuild/linux-arm/-/linux-arm-0.25.8.tgz",
-      "integrity": "sha512-FuzEP9BixzZohl1kLf76KEVOsxtIBFwCaLupVuk4eFVnOZfU+Wsn+x5Ryam7nILV2pkq2TqQM9EZPsOBuMC+kg==",
+      "version": "0.28.2",
+      "resolved": "https://registry.npmjs.org/@esbuild/linux-arm/-/linux-arm-0.28.2.tgz",
+      "integrity": "sha512-XlDnu2q5yoqems+xay6wSAcg9DDD7K9RLKZEBOMZm3ckNpJBvOX20tSfby8KfrrhINDyv9V2YVZKY/SpoGJI8w==",
       "cpu": [
         "arm"
       ],
@@ -486,9 +797,9 @@
       }
     },
     "node_modules/@esbuild/linux-arm64": {
-      "version": "0.25.8",
-      "resolved": "https://registry.npmjs.org/@esbuild/linux-arm64/-/linux-arm64-0.25.8.tgz",
-      "integrity": "sha512-WIgg00ARWv/uYLU7lsuDK00d/hHSfES5BzdWAdAig1ioV5kaFNrtK8EqGcUBJhYqotlUByUKz5Qo6u8tt7iD/w==",
+      "version": "0.28.2",
+      "resolved": "https://registry.npmjs.org/@esbuild/linux-arm64/-/linux-arm64-0.28.2.tgz",
+      "integrity": "sha512-pW4AC0P3it8c7do9MVM4p51FzHzdM/TZrerurgRcHJ2WTa1VQ1CIq18xncfpBJw4ojkiZZrKW2yIBWBP92j6Ug==",
       "cpu": [
         "arm64"
       ],
@@ -503,9 +814,9 @@
       }
     },
     "node_modules/@esbuild/linux-ia32": {
-      "version": "0.25.8",
-      "resolved": "https://registry.npmjs.org/@esbuild/linux-ia32/-/linux-ia32-0.25.8.tgz",
-      "integrity": "sha512-A1D9YzRX1i+1AJZuFFUMP1E9fMaYY+GnSQil9Tlw05utlE86EKTUA7RjwHDkEitmLYiFsRd9HwKBPEftNdBfjg==",
+      "version": "0.28.2",
+      "resolved": "https://registry.npmjs.org/@esbuild/linux-ia32/-/linux-ia32-0.28.2.tgz",
+      "integrity": "sha512-CYbnj78HsIeA+DhgUKgFCfvNsTHFhMMrinUrMZpDXJXKN8T3XViTZ/+wtHeVxEWY8ewSzTFN+nRmSwO2tZaLUQ==",
       "cpu": [
         "ia32"
       ],
@@ -520,9 +831,9 @@
       }
     },
     "node_modules/@esbuild/linux-loong64": {
-      "version": "0.25.8",
-      "resolved": "https://registry.npmjs.org/@esbuild/linux-loong64/-/linux-loong64-0.25.8.tgz",
-      "integrity": "sha512-O7k1J/dwHkY1RMVvglFHl1HzutGEFFZ3kNiDMSOyUrB7WcoHGf96Sh+64nTRT26l3GMbCW01Ekh/ThKM5iI7hQ==",
+      "version": "0.28.2",
+      "resolved": "https://registry.npmjs.org/@esbuild/linux-loong64/-/linux-loong64-0.28.2.tgz",
+      "integrity": "sha512-buwkd8nsph4R+ajRvw0qM5Hja/TXQow3ptzWO2EbG/cqcIkHloRrdlBtQlshyYGTNFvfkfJ5tpPLVkY4DtsPfQ==",
       "cpu": [
         "loong64"
       ],
@@ -537,9 +848,9 @@
       }
     },
     "node_modules/@esbuild/linux-mips64el": {
-      "version": "0.25.8",
-      "resolved": "https://registry.npmjs.org/@esbuild/linux-mips64el/-/linux-mips64el-0.25.8.tgz",
-      "integrity": "sha512-uv+dqfRazte3BzfMp8PAQXmdGHQt2oC/y2ovwpTteqrMx2lwaksiFZ/bdkXJC19ttTvNXBuWH53zy/aTj1FgGw==",
+      "version": "0.28.2",
+      "resolved": "https://registry.npmjs.org/@esbuild/linux-mips64el/-/linux-mips64el-0.28.2.tgz",
+      "integrity": "sha512-ZVykbDyk7519VwiNb9Lcj9m8XM6v5V9uKPvrEMkkEedVewf+0itkhahp4HDpgERXhwLRpWFypsGbG/J8s0QjJA==",
       "cpu": [
         "mips64el"
       ],
@@ -554,9 +865,9 @@
       }
     },
     "node_modules/@esbuild/linux-ppc64": {
-      "version": "0.25.8",
-      "resolved": "https://registry.npmjs.org/@esbuild/linux-ppc64/-/linux-ppc64-0.25.8.tgz",
-      "integrity": "sha512-GyG0KcMi1GBavP5JgAkkstMGyMholMDybAf8wF5A70CALlDM2p/f7YFE7H92eDeH/VBtFJA5MT4nRPDGg4JuzQ==",
+      "version": "0.28.2",
+      "resolved": "https://registry.npmjs.org/@esbuild/linux-ppc64/-/linux-ppc64-0.28.2.tgz",
+      "integrity": "sha512-CAXl+Dtd9UUuJd8pKKdwh6MLm3MUMiqMPmhZ3tTSXPqfyQ3vDl6R5hZdZ/kYojK4ofXtdfSv1tFq8XzWx3heNQ==",
       "cpu": [
         "ppc64"
       ],
@@ -571,9 +882,9 @@
       }
     },
     "node_modules/@esbuild/linux-riscv64": {
-      "version": "0.25.8",
-      "resolved": "https://registry.npmjs.org/@esbuild/linux-riscv64/-/linux-riscv64-0.25.8.tgz",
-      "integrity": "sha512-rAqDYFv3yzMrq7GIcen3XP7TUEG/4LK86LUPMIz6RT8A6pRIDn0sDcvjudVZBiiTcZCY9y2SgYX2lgK3AF+1eg==",
+      "version": "0.28.2",
+      "resolved": "https://registry.npmjs.org/@esbuild/linux-riscv64/-/linux-riscv64-0.28.2.tgz",
+      "integrity": "sha512-GeXCej4IQtU1B+QlDV8W/RRvbzI3O/Stss+/bCXv4lZls5WGRtu2a+3JkA3i4qIUlMXpcHebWpF8AkJhATowuA==",
       "cpu": [
         "riscv64"
       ],
@@ -588,9 +899,9 @@
       }
     },
     "node_modules/@esbuild/linux-s390x": {
-      "version": "0.25.8",
-      "resolved": "https://registry.npmjs.org/@esbuild/linux-s390x/-/linux-s390x-0.25.8.tgz",
-      "integrity": "sha512-Xutvh6VjlbcHpsIIbwY8GVRbwoviWT19tFhgdA7DlenLGC/mbc3lBoVb7jxj9Z+eyGqvcnSyIltYUrkKzWqSvg==",
+      "version": "0.28.2",
+      "resolved": "https://registry.npmjs.org/@esbuild/linux-s390x/-/linux-s390x-0.28.2.tgz",
+      "integrity": "sha512-3H1weTYZPxt/WOhByszQZybS9w5lKzUn1FDMsgEChbHWQwHYQQRfBxgCcZvPhjHfKyJjIievvMmEUawJrdY9Dg==",
       "cpu": [
         "s390x"
       ],
@@ -605,9 +916,9 @@
       }
     },
     "node_modules/@esbuild/linux-x64": {
-      "version": "0.25.8",
-      "resolved": "https://registry.npmjs.org/@esbuild/linux-x64/-/linux-x64-0.25.8.tgz",
-      "integrity": "sha512-ASFQhgY4ElXh3nDcOMTkQero4b1lgubskNlhIfJrsH5OKZXDpUAKBlNS0Kx81jwOBp+HCeZqmoJuihTv57/jvQ==",
+      "version": "0.28.2",
+      "resolved": "https://registry.npmjs.org/@esbuild/linux-x64/-/linux-x64-0.28.2.tgz",
+      "integrity": "sha512-4xTZr1FUmSoQW4XIWmit3tzQrUTZM+N3P0XV8xROKYF50XfI7xeO90+1bZvNwxIufQ9hDQVRJH5YhgPVF8A/HQ==",
       "cpu": [
         "x64"
       ],
@@ -622,9 +933,9 @@
       }
     },
     "node_modules/@esbuild/netbsd-arm64": {
-      "version": "0.25.8",
-      "resolved": "https://registry.npmjs.org/@esbuild/netbsd-arm64/-/netbsd-arm64-0.25.8.tgz",
-      "integrity": "sha512-d1KfruIeohqAi6SA+gENMuObDbEjn22olAR7egqnkCD9DGBG0wsEARotkLgXDu6c4ncgWTZJtN5vcgxzWRMzcw==",
+      "version": "0.28.2",
+      "resolved": "https://registry.npmjs.org/@esbuild/netbsd-arm64/-/netbsd-arm64-0.28.2.tgz",
+      "integrity": "sha512-sSATRjPeDBg3pdgHoQfoYBob11Kk1FGa9lui5RIHZCoCkJa9QKlvl3/vKz2usCmYYjs7ymJR/2Nnsqe+Hjt5nw==",
       "cpu": [
         "arm64"
       ],
@@ -639,11 +950,11 @@
       }
     },
     "node_modules/@esbuild/netbsd-x64": {
-      "version": "0.25.8",
-      "resolved": "https://registry.npmjs.org/@esbuild/netbsd-x64/-/netbsd-x64-0.25.8.tgz",
-      "integrity": "sha512-nVDCkrvx2ua+XQNyfrujIG38+YGyuy2Ru9kKVNyh5jAys6n+l44tTtToqHjino2My8VAY6Lw9H7RI73XFi66Cg==",
-      "cpu": [
-        "x64"
+      "version": "0.28.2",
+      "resolved": "https://registry.npmjs.org/@esbuild/netbsd-x64/-/netbsd-x64-0.28.2.tgz",
+      "integrity": "sha512-lqnzCV+mM0gIADaKihiCg6ifgfU2L3h5E33rNQBN1Y4MaVGnzryzmvvf7UHxprpQdE8hpqLolJ9Rl+SkIRDpyw==",
+      "cpu": [
+        "x64"
       ],
       "dev": true,
       "license": "MIT",
@@ -656,9 +967,9 @@
       }
     },
     "node_modules/@esbuild/openbsd-arm64": {
-      "version": "0.25.8",
-      "resolved": "https://registry.npmjs.org/@esbuild/openbsd-arm64/-/openbsd-arm64-0.25.8.tgz",
-      "integrity": "sha512-j8HgrDuSJFAujkivSMSfPQSAa5Fxbvk4rgNAS5i3K+r8s1X0p1uOO2Hl2xNsGFppOeHOLAVgYwDVlmxhq5h+SQ==",
+      "version": "0.28.2",
+      "resolved": "https://registry.npmjs.org/@esbuild/openbsd-arm64/-/openbsd-arm64-0.28.2.tgz",
+      "integrity": "sha512-AL2qJILH7lNjrDmCQDvdxMfAUIv8KMNZOvrwAQ8i8//ntL9FflhOyMJ8OZSMBb8/AWXe3/5v5S20y3zCoZWKoQ==",
       "cpu": [
         "arm64"
       ],
@@ -673,9 +984,9 @@
       }
     },
     "node_modules/@esbuild/openbsd-x64": {
-      "version": "0.25.8",
-      "resolved": "https://registry.npmjs.org/@esbuild/openbsd-x64/-/openbsd-x64-0.25.8.tgz",
-      "integrity": "sha512-1h8MUAwa0VhNCDp6Af0HToI2TJFAn1uqT9Al6DJVzdIBAd21m/G0Yfc77KDM3uF3T/YaOgQq3qTJHPbTOInaIQ==",
+      "version": "0.28.2",
+      "resolved": "https://registry.npmjs.org/@esbuild/openbsd-x64/-/openbsd-x64-0.28.2.tgz",
+      "integrity": "sha512-QtiuPytchRyC4rwUKhexJdQKvDuZ6hWloi3igqPQNUJCS1/v9EiO3UTOXR6A3FoMo4fnAKbWJdqaIwhOzh8qEw==",
       "cpu": [
         "x64"
       ],
@@ -690,9 +1001,9 @@
       }
     },
     "node_modules/@esbuild/openharmony-arm64": {
-      "version": "0.25.8",
-      "resolved": "https://registry.npmjs.org/@esbuild/openharmony-arm64/-/openharmony-arm64-0.25.8.tgz",
-      "integrity": "sha512-r2nVa5SIK9tSWd0kJd9HCffnDHKchTGikb//9c7HX+r+wHYCpQrSgxhlY6KWV1nFo1l4KFbsMlHk+L6fekLsUg==",
+      "version": "0.28.2",
+      "resolved": "https://registry.npmjs.org/@esbuild/openharmony-arm64/-/openharmony-arm64-0.28.2.tgz",
+      "integrity": "sha512-WkhYDmpTjLvGlScA1rwjRUmhl4k8oXR3cIbtqWmELgU/dFeHHlEllxDvdWcNJV9rbzCexB5vz8gtNewWLgCT7Q==",
       "cpu": [
         "arm64"
       ],
@@ -707,9 +1018,9 @@
       }
     },
     "node_modules/@esbuild/sunos-x64": {
-      "version": "0.25.8",
-      "resolved": "https://registry.npmjs.org/@esbuild/sunos-x64/-/sunos-x64-0.25.8.tgz",
-      "integrity": "sha512-zUlaP2S12YhQ2UzUfcCuMDHQFJyKABkAjvO5YSndMiIkMimPmxA+BYSBikWgsRpvyxuRnow4nS5NPnf9fpv41w==",
+      "version": "0.28.2",
+      "resolved": "https://registry.npmjs.org/@esbuild/sunos-x64/-/sunos-x64-0.28.2.tgz",
+      "integrity": "sha512-GPMSkTOtMnv2U2F8gxe4Io6qmVs+YKyp832Etqqxr0hFngmXQ3rzwytelm3GIn7T4VviRUlf3sOgBOiTdvaf7g==",
       "cpu": [
         "x64"
       ],
@@ -724,9 +1035,9 @@
       }
     },
     "node_modules/@esbuild/win32-arm64": {
-      "version": "0.25.8",
-      "resolved": "https://registry.npmjs.org/@esbuild/win32-arm64/-/win32-arm64-0.25.8.tgz",
-      "integrity": "sha512-YEGFFWESlPva8hGL+zvj2z/SaK+pH0SwOM0Nc/d+rVnW7GSTFlLBGzZkuSU9kFIGIo8q9X3ucpZhu8PDN5A2sQ==",
+      "version": "0.28.2",
+      "resolved": "https://registry.npmjs.org/@esbuild/win32-arm64/-/win32-arm64-0.28.2.tgz",
+      "integrity": "sha512-PIhhEkE9uPBleRBrQEJpUn7MBnibZzbGzYWPmY3x+YoVg/95zbjB4CxPPOQ8l5tYYM4mMaCthF8/1DIfBQQyWQ==",
       "cpu": [
         "arm64"
       ],
@@ -741,9 +1052,9 @@
       }
     },
     "node_modules/@esbuild/win32-ia32": {
-      "version": "0.25.8",
-      "resolved": "https://registry.npmjs.org/@esbuild/win32-ia32/-/win32-ia32-0.25.8.tgz",
-      "integrity": "sha512-hiGgGC6KZ5LZz58OL/+qVVoZiuZlUYlYHNAmczOm7bs2oE1XriPFi5ZHHrS8ACpV5EjySrnoCKmcbQMN+ojnHg==",
+      "version": "0.28.2",
+      "resolved": "https://registry.npmjs.org/@esbuild/win32-ia32/-/win32-ia32-0.28.2.tgz",
+      "integrity": "sha512-YmJbfTlvU7Sdn9BB+4PRES4oB6pxgS37MAONj+hBr/cpXS1aBPKXxNnDbu+QCWPj0o9dgyxeq79g6c5P8KeuYA==",
       "cpu": [
         "ia32"
       ],
@@ -758,9 +1069,9 @@
       }
     },
     "node_modules/@esbuild/win32-x64": {
-      "version": "0.25.8",
-      "resolved": "https://registry.npmjs.org/@esbuild/win32-x64/-/win32-x64-0.25.8.tgz",
-      "integrity": "sha512-cn3Yr7+OaaZq1c+2pe+8yxC8E144SReCQjN6/2ynubzYjvyqZjTXfQJpAcQpsdJq3My7XADANiYGHoFC69pLQw==",
+      "version": "0.28.2",
+      "resolved": "https://registry.npmjs.org/@esbuild/win32-x64/-/win32-x64-0.28.2.tgz",
+      "integrity": "sha512-5ebpxr3nWMzrL/rnUI755Jkuee0bHL/Gq0WTF9lvcpv73wAp5eu8MfBUgWK9bhWvZjj7yX8etf/8tI8Ney695g==",
       "cpu": [
         "x64"
       ],
@@ -775,9 +1086,9 @@
       }
     },
     "node_modules/@eslint-community/eslint-utils": {
-      "version": "4.7.0",
-      "resolved": "https://registry.npmjs.org/@eslint-community/eslint-utils/-/eslint-utils-4.7.0.tgz",
-      "integrity": "sha512-dyybb3AcajC7uha6CvhdVRJqaKyn7w2YKqKyAN37NKYgZT36w+iRb0Dymmc5qEJ549c/S31cMMSFd75bteCpCw==",
+      "version": "4.10.1",
+      "resolved": "https://registry.npmjs.org/@eslint-community/eslint-utils/-/eslint-utils-4.10.1.tgz",
+      "integrity": "sha512-cuadcxVFE8sDK6iWJbs8Sn0av2Nrh2QSGQhVlBW9AaAHqHwjWsZHT8LJ4hFGPh7ASBV2deFdM7H/DPjulmh8rg==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
@@ -807,9 +1118,9 @@
       }
     },
     "node_modules/@eslint-community/regexpp": {
-      "version": "4.12.1",
-      "resolved": "https://registry.npmjs.org/@eslint-community/regexpp/-/regexpp-4.12.1.tgz",
-      "integrity": "sha512-CCZCDJuduB9OUkFkY2IgppNZMi2lBQgD2qzwXkEia16cge2pijY/aXi96CJMquDMn3nJdlPV1A5KrJEXwfLNzQ==",
+      "version": "4.12.2",
+      "resolved": "https://registry.npmjs.org/@eslint-community/regexpp/-/regexpp-4.12.2.tgz",
+      "integrity": "sha512-EriSTlt5OC9/7SXkRSCAhfSxxoSUgBm33OH+IkwbdpgoqsSsUg7y3uh+IICI/Qg4BBWr3U2i39RpmycbxMq4ew==",
       "dev": true,
       "license": "MIT",
       "engines": {
@@ -817,102 +1128,89 @@
       }
     },
     "node_modules/@eslint/config-array": {
-      "version": "0.21.0",
-      "resolved": "https://registry.npmjs.org/@eslint/config-array/-/config-array-0.21.0.tgz",
-      "integrity": "sha512-ENIdc4iLu0d93HeYirvKmrzshzofPw6VkZRKQGe9Nv46ZnWUzcF1xV01dcvEg/1wXUR61OmmlSfyeyO7EvjLxQ==",
+      "version": "0.23.5",
+      "resolved": "https://registry.npmjs.org/@eslint/config-array/-/config-array-0.23.5.tgz",
+      "integrity": "sha512-Y3kKLvC1dvTOT+oGlqNQ1XLqK6D1HU2YXPc52NmAlJZbMMWDzGYXMiPRJ8TYD39muD/OTjlZmNJ4ib7dvSrMBA==",
       "dev": true,
       "license": "Apache-2.0",
       "dependencies": {
-        "@eslint/object-schema": "^2.1.6",
+        "@eslint/object-schema": "^3.0.5",
         "debug": "^4.3.1",
-        "minimatch": "^3.1.2"
+        "minimatch": "^10.2.4"
       },
       "engines": {
-        "node": "^18.18.0 || ^20.9.0 || >=21.1.0"
+        "node": "^20.19.0 || ^22.13.0 || >=24"
       }
     },
     "node_modules/@eslint/config-helpers": {
-      "version": "0.3.1",
-      "resolved": "https://registry.npmjs.org/@eslint/config-helpers/-/config-helpers-0.3.1.tgz",
-      "integrity": "sha512-xR93k9WhrDYpXHORXpxVL5oHj3Era7wo6k/Wd8/IsQNnZUTzkGS29lyn3nAT05v6ltUuTFVCCYDEGfy2Or/sPA==",
+      "version": "0.7.0",
+      "resolved": "https://registry.npmjs.org/@eslint/config-helpers/-/config-helpers-0.7.0.tgz",
+      "integrity": "sha512-DObd/KKUsU+FaFv4PLxSRenpXfQWmPXXP3pPZ6/K1PCrMu2vQpMDMuQe/BqYeoLcz8ro0bVDF1RxOJgfVEdhUw==",
       "dev": true,
       "license": "Apache-2.0",
+      "dependencies": {
+        "@eslint/core": "^1.2.1"
+      },
       "engines": {
-        "node": "^18.18.0 || ^20.9.0 || >=21.1.0"
+        "node": "^20.19.0 || ^22.13.0 || >=24"
       }
     },
     "node_modules/@eslint/core": {
-      "version": "0.15.2",
-      "resolved": "https://registry.npmjs.org/@eslint/core/-/core-0.15.2.tgz",
-      "integrity": "sha512-78Md3/Rrxh83gCxoUc0EiciuOHsIITzLy53m3d9UyiW8y9Dj2D29FeETqyKA+BRK76tnTp6RXWb3pCay8Oyomg==",
+      "version": "1.2.1",
+      "resolved": "https://registry.npmjs.org/@eslint/core/-/core-1.2.1.tgz",
+      "integrity": "sha512-MwcE1P+AZ4C6DWlpin/OmOA54mmIZ/+xZuJiQd4SyB29oAJjN30UW9wkKNptW2ctp4cEsvhlLY/CsQ1uoHDloQ==",
       "dev": true,
       "license": "Apache-2.0",
       "dependencies": {
         "@types/json-schema": "^7.0.15"
       },
       "engines": {
-        "node": "^18.18.0 || ^20.9.0 || >=21.1.0"
-      }
-    },
-    "node_modules/@eslint/eslintrc": {
-      "version": "3.3.1",
-      "resolved": "https://registry.npmjs.org/@eslint/eslintrc/-/eslintrc-3.3.1.tgz",
-      "integrity": "sha512-gtF186CXhIl1p4pJNGZw8Yc6RlshoePRvE0X91oPGb3vZ8pM3qOS9W9NGPat9LziaBV7XrJWGylNQXkGcnM3IQ==",
-      "dev": true,
-      "license": "MIT",
-      "dependencies": {
-        "ajv": "^6.12.4",
-        "debug": "^4.3.2",
-        "espree": "^10.0.1",
-        "globals": "^14.0.0",
-        "ignore": "^5.2.0",
-        "import-fresh": "^3.2.1",
-        "js-yaml": "^4.1.0",
-        "minimatch": "^3.1.2",
-        "strip-json-comments": "^3.1.1"
-      },
-      "engines": {
-        "node": "^18.18.0 || ^20.9.0 || >=21.1.0"
-      },
-      "funding": {
-        "url": "https://opencollective.com/eslint"
+        "node": "^20.19.0 || ^22.13.0 || >=24"
       }
     },
     "node_modules/@eslint/js": {
-      "version": "9.33.0",
-      "resolved": "https://registry.npmjs.org/@eslint/js/-/js-9.33.0.tgz",
-      "integrity": "sha512-5K1/mKhWaMfreBGJTwval43JJmkip0RmM+3+IuqupeSKNC/Th2Kc7ucaq5ovTSra/OOKB9c58CGSz3QMVbWt0A==",
+      "version": "10.0.1",
+      "resolved": "https://registry.npmjs.org/@eslint/js/-/js-10.0.1.tgz",
+      "integrity": "sha512-zeR9k5pd4gxjZ0abRoIaxdc7I3nDktoXZk2qOv9gCNWx3mVwEn32VRhyLaRsDiJjTs0xq/T8mfPtyuXu7GWBcA==",
       "dev": true,
       "license": "MIT",
       "engines": {
-        "node": "^18.18.0 || ^20.9.0 || >=21.1.0"
+        "node": "^20.19.0 || ^22.13.0 || >=24"
       },
       "funding": {
         "url": "https://eslint.org/donate"
+      },
+      "peerDependencies": {
+        "eslint": "^10.0.0"
+      },
+      "peerDependenciesMeta": {
+        "eslint": {
+          "optional": true
+        }
       }
     },
     "node_modules/@eslint/object-schema": {
-      "version": "2.1.6",
-      "resolved": "https://registry.npmjs.org/@eslint/object-schema/-/object-schema-2.1.6.tgz",
-      "integrity": "sha512-RBMg5FRL0I0gs51M/guSAj5/e14VQ4tpZnQNWwuDT66P14I43ItmPfIZRhO9fUVIPOAQXU47atlywZ/czoqFPA==",
+      "version": "3.0.5",
+      "resolved": "https://registry.npmjs.org/@eslint/object-schema/-/object-schema-3.0.5.tgz",
+      "integrity": "sha512-vqTaUEgxzm+YDSdElad6PiRoX4t8VGDjCtt05zn4nU810UIx/uNEV7/lZJ6KwFThKZOzOxzXy48da+No7HZaMw==",
       "dev": true,
       "license": "Apache-2.0",
       "engines": {
-        "node": "^18.18.0 || ^20.9.0 || >=21.1.0"
+        "node": "^20.19.0 || ^22.13.0 || >=24"
       }
     },
     "node_modules/@eslint/plugin-kit": {
-      "version": "0.3.5",
-      "resolved": "https://registry.npmjs.org/@eslint/plugin-kit/-/plugin-kit-0.3.5.tgz",
-      "integrity": "sha512-Z5kJ+wU3oA7MMIqVR9tyZRtjYPr4OC004Q4Rw7pgOKUOKkJfZ3O24nz3WYfGRpMDNmcOi3TwQOmgm7B7Tpii0w==",
+      "version": "0.7.3",
+      "resolved": "https://registry.npmjs.org/@eslint/plugin-kit/-/plugin-kit-0.7.3.tgz",
+      "integrity": "sha512-IkO+/KEUvwbVpiURZg+P7zF74z5Jxe0UgJxVni+RtoHQ6IZieXaO02kmadomap/q+l6bc/jdPGGqTjhuZnuz1Q==",
       "dev": true,
       "license": "Apache-2.0",
       "dependencies": {
-        "@eslint/core": "^0.15.2",
+        "@eslint/core": "^1.2.1",
         "levn": "^0.4.1"
       },
       "engines": {
-        "node": "^18.18.0 || ^20.9.0 || >=21.1.0"
+        "node": "^20.19.0 || ^22.13.0 || >=24"
       }
     },
     "node_modules/@humanfs/core": {
@@ -986,9 +1284,9 @@
       "link": true
     },
     "node_modules/@intentee/rust-coverage-check": {
-      "version": "0.2.0",
-      "resolved": "https://registry.npmjs.org/@intentee/rust-coverage-check/-/rust-coverage-check-0.2.0.tgz",
-      "integrity": "sha512-RRCHxYdYLk5SjmBw4YAOsaE3GwjJYuzhpY/Gaz8frW0dLqXZNBz5Xtd2k79dI8januHjLDjH12/vYNHOoHNsSw==",
+      "version": "0.3.1",
+      "resolved": "https://registry.npmjs.org/@intentee/rust-coverage-check/-/rust-coverage-check-0.3.1.tgz",
+      "integrity": "sha512-Ep4kv95XYjNESk1LYmhs+KPuzaQQ1fGq3ZSwB/Knwu5aZj23ICY7hLQDdU//BEQMZU0hwrhmv2gtQ8Z1bPcNrQ==",
       "dev": true,
       "license": "MIT",
       "bin": {
@@ -1016,6 +1314,19 @@
         "node": ">=12"
       }
     },
+    "node_modules/@isaacs/cliui/node_modules/ansi-styles": {
+      "version": "6.2.3",
+      "resolved": "https://registry.npmjs.org/ansi-styles/-/ansi-styles-6.2.3.tgz",
+      "integrity": "sha512-4Dj6M28JB+oAH8kFkTLUo+a2jwOFkuqb3yucU0CANcRRUbxS0cP0nZYCGjcc3BNXwRIsUVmDGgzawme7zvJHvg==",
+      "dev": true,
+      "license": "MIT",
+      "engines": {
+        "node": ">=12"
+      },
+      "funding": {
+        "url": "https://github.com/chalk/ansi-styles?sponsor=1"
+      }
+    },
     "node_modules/@isaacs/cliui/node_modules/emoji-regex": {
       "version": "9.2.2",
       "resolved": "https://registry.npmjs.org/emoji-regex/-/emoji-regex-9.2.2.tgz",
@@ -1041,23 +1352,108 @@
         "url": "https://github.com/sponsors/sindresorhus"
       }
     },
+    "node_modules/@isaacs/cliui/node_modules/wrap-ansi": {
+      "version": "8.1.0",
+      "resolved": "https://registry.npmjs.org/wrap-ansi/-/wrap-ansi-8.1.0.tgz",
+      "integrity": "sha512-si7QWI6zUMq56bESFvagtmzMdGOtoxfR+Sez11Mobfc7tm+VkUckk9bW2UeffTGVUbOksxmSw0AA2gs8g71NCQ==",
+      "dev": true,
+      "license": "MIT",
+      "dependencies": {
+        "ansi-styles": "^6.1.0",
+        "string-width": "^5.0.1",
+        "strip-ansi": "^7.0.1"
+      },
+      "engines": {
+        "node": ">=12"
+      },
+      "funding": {
+        "url": "https://github.com/chalk/wrap-ansi?sponsor=1"
+      }
+    },
+    "node_modules/@jridgewell/gen-mapping": {
+      "version": "0.3.13",
+      "resolved": "https://registry.npmjs.org/@jridgewell/gen-mapping/-/gen-mapping-0.3.13.tgz",
+      "integrity": "sha512-2kkt/7niJ6MgEPxF0bYdQ6etZaA+fQvDcLKckhy1yIQOzaoKjBBjSj63/aLVjYE3qhRt5dvM+uUyfCg6UKCBbA==",
+      "dev": true,
+      "license": "MIT",
+      "dependencies": {
+        "@jridgewell/sourcemap-codec": "^1.5.0",
+        "@jridgewell/trace-mapping": "^0.3.24"
+      }
+    },
+    "node_modules/@jridgewell/remapping": {
+      "version": "2.3.5",
+      "resolved": "https://registry.npmjs.org/@jridgewell/remapping/-/remapping-2.3.5.tgz",
+      "integrity": "sha512-LI9u/+laYG4Ds1TDKSJW2YPrIlcVYOwi2fUC6xB43lueCjgxV4lffOCZCtYFiH6TNOX+tQKXx97T4IKHbhyHEQ==",
+      "dev": true,
+      "license": "MIT",
+      "dependencies": {
+        "@jridgewell/gen-mapping": "^0.3.5",
+        "@jridgewell/trace-mapping": "^0.3.24"
+      }
+    },
+    "node_modules/@jridgewell/resolve-uri": {
+      "version": "3.1.2",
+      "resolved": "https://registry.npmjs.org/@jridgewell/resolve-uri/-/resolve-uri-3.1.2.tgz",
+      "integrity": "sha512-bRISgCIjP20/tbWSPWMEi54QVPRZExkuD9lJL+UIxUKtwVJA8wW1Trb1jMs1RFXo1CBTNZ/5hpC9QvmKWdopKw==",
+      "dev": true,
+      "license": "MIT",
+      "engines": {
+        "node": ">=6.0.0"
+      }
+    },
+    "node_modules/@jridgewell/sourcemap-codec": {
+      "version": "1.6.0",
+      "resolved": "https://registry.npmjs.org/@jridgewell/sourcemap-codec/-/sourcemap-codec-1.6.0.tgz",
+      "integrity": "sha512-T7jf+5zgsZHwNJ4lvQ7/aezbyk0nNX+zJVWpmHA7VYsEx7a7qr5Rg5IbtJFqkgze5Y2sruq1RUY8Q837Od7iFw==",
+      "dev": true,
+      "license": "MIT"
+    },
+    "node_modules/@jridgewell/trace-mapping": {
+      "version": "0.3.31",
+      "resolved": "https://registry.npmjs.org/@jridgewell/trace-mapping/-/trace-mapping-0.3.31.tgz",
+      "integrity": "sha512-zzNR+SdQSDJzc8joaeP8QQoCQr8NuYx2dIIytl1QeBEZHJ9uW6hebsrYgbz8hJwUQao3TWCMtmfV8Nu1twOLAw==",
+      "dev": true,
+      "license": "MIT",
+      "dependencies": {
+        "@jridgewell/resolve-uri": "^3.1.0",
+        "@jridgewell/sourcemap-codec": "^1.4.14"
+      }
+    },
+    "node_modules/@keyv/bigmap": {
+      "version": "1.3.1",
+      "resolved": "https://registry.npmjs.org/@keyv/bigmap/-/bigmap-1.3.1.tgz",
+      "integrity": "sha512-WbzE9sdmQtKy8vrNPa9BRnwZh5UF4s1KTmSK0KUVLo3eff5BlQNNWDnFOouNpKfPKDnms9xynJjsMYjMaT/aFQ==",
+      "dev": true,
+      "license": "MIT",
+      "dependencies": {
+        "hashery": "^1.4.0",
+        "hookified": "^1.15.0"
+      },
+      "engines": {
+        "node": ">= 18"
+      },
+      "peerDependencies": {
+        "keyv": "^5.6.0"
+      }
+    },
     "node_modules/@keyv/serialize": {
-      "version": "1.1.0",
-      "resolved": "https://registry.npmjs.org/@keyv/serialize/-/serialize-1.1.0.tgz",
-      "integrity": "sha512-RlDgexML7Z63Q8BSaqhXdCYNBy/JQnqYIwxofUrNLGCblOMHp+xux2Q8nLMLlPpgHQPoU0Do8Z6btCpRBEqZ8g==",
+      "version": "1.1.1",
+      "resolved": "https://registry.npmjs.org/@keyv/serialize/-/serialize-1.1.1.tgz",
+      "integrity": "sha512-dXn3FZhPv0US+7dtJsIi2R+c7qWYiReoEh5zUntWCf4oSpMNib8FDhSoed6m3QyZdx5hK7iLFkYk3rNxwt8vTA==",
       "dev": true,
       "license": "MIT"
     },
     "node_modules/@lezer/common": {
-      "version": "1.2.3",
-      "resolved": "https://registry.npmjs.org/@lezer/common/-/common-1.2.3.tgz",
-      "integrity": "sha512-w7ojc8ejBqr2REPsWxJjrMFsA/ysDCFICn8zEOR9mrqzOu2amhITYuLD8ag6XZf0CFXDrhKqw7+tW8cX66NaDA==",
+      "version": "1.5.4",
+      "resolved": "https://registry.npmjs.org/@lezer/common/-/common-1.5.4.tgz",
+      "integrity": "sha512-xt3Q4z64xM+xhg4M7NXqw1N//1O74bUFo5kz6fwMeC4qwxSuRdwkYSmjcqCEEKYkNw5e67f2hcNqBoNERAEU0w==",
       "license": "MIT"
     },
     "node_modules/@lezer/css": {
-      "version": "1.3.0",
-      "resolved": "https://registry.npmjs.org/@lezer/css/-/css-1.3.0.tgz",
-      "integrity": "sha512-pBL7hup88KbI7hXnZV3PQsn43DHy6TWyzuyk2AO9UyoXcDltvIdqWKE1dLL/45JVZ+YZkHe1WVHqO6wugZZWcw==",
+      "version": "1.3.9",
+      "resolved": "https://registry.npmjs.org/@lezer/css/-/css-1.3.9.tgz",
+      "integrity": "sha512-ZmzgvTOBdJWGIsY/ItAfN7UW8DV802oZG+jeG6E508SwEzz5YhuAajptpWPICvIyqYAD2+VjMdOS40GgGkyEOQ==",
       "license": "MIT",
       "dependencies": {
         "@lezer/common": "^1.2.0",
@@ -1066,18 +1462,18 @@
       }
     },
     "node_modules/@lezer/highlight": {
-      "version": "1.2.1",
-      "resolved": "https://registry.npmjs.org/@lezer/highlight/-/highlight-1.2.1.tgz",
-      "integrity": "sha512-Z5duk4RN/3zuVO7Jq0pGLJ3qynpxUVsh7IbUbGj88+uV2ApSAn6kWg2au3iJb+0Zi7kKtqffIESgNcRXWZWmSA==",
+      "version": "1.2.5",
+      "resolved": "https://registry.npmjs.org/@lezer/highlight/-/highlight-1.2.5.tgz",
+      "integrity": "sha512-O1GMVKgtf5YspFaRzpmqkVgMtIk2HG9uMAbH6wdAIdNh3TVzCsch/+cRi+6l2//UGDKe6NB9BW3cvuR5ezMQ0w==",
       "license": "MIT",
       "dependencies": {
-        "@lezer/common": "^1.0.0"
+        "@lezer/common": "^1.3.0"
       }
     },
     "node_modules/@lezer/html": {
-      "version": "1.3.10",
-      "resolved": "https://registry.npmjs.org/@lezer/html/-/html-1.3.10.tgz",
-      "integrity": "sha512-dqpT8nISx/p9Do3AchvYGV3qYc4/rKr3IBZxlHmpIKam56P47RSHkSF5f13Vu9hebS1jM0HmtJIwLbWz1VIY6w==",
+      "version": "1.3.13",
+      "resolved": "https://registry.npmjs.org/@lezer/html/-/html-1.3.13.tgz",
+      "integrity": "sha512-oI7n6NJml729m7pjm9lvLvmXbdoMoi2f+1pwSDJkl9d68zGr7a9Btz8NdHTGQZtW2DA25ybeuv/SyDb9D5tseg==",
       "license": "MIT",
       "dependencies": {
         "@lezer/common": "^1.2.0",
@@ -1086,9 +1482,9 @@
       }
     },
     "node_modules/@lezer/javascript": {
-      "version": "1.5.1",
-      "resolved": "https://registry.npmjs.org/@lezer/javascript/-/javascript-1.5.1.tgz",
-      "integrity": "sha512-ATOImjeVJuvgm3JQ/bpo2Tmv55HSScE2MTPnKRMRIPx2cLhHGyX2VnqpHhtIV1tVzIjZDbcWQm+NCTF40ggZVw==",
+      "version": "1.5.6",
+      "resolved": "https://registry.npmjs.org/@lezer/javascript/-/javascript-1.5.6.tgz",
+      "integrity": "sha512-sEvOqpl4vJQca0NugSQK02vQPbc5KrvoojnYGoqqFgFO0MtcYhwGiHKVEB+LuXYJnfyXo4QPA8ztbrEa4Fw5NA==",
       "license": "MIT",
       "dependencies": {
         "@lezer/common": "^1.2.0",
@@ -1097,18 +1493,18 @@
       }
     },
     "node_modules/@lezer/lr": {
-      "version": "1.4.2",
-      "resolved": "https://registry.npmjs.org/@lezer/lr/-/lr-1.4.2.tgz",
-      "integrity": "sha512-pu0K1jCIdnQ12aWNaAVU5bzi7Bd1w54J3ECgANPmYLtQKP0HBj2cE/5coBD66MT10xbtIuUr7tg0Shbsvk0mDA==",
+      "version": "1.4.11",
+      "resolved": "https://registry.npmjs.org/@lezer/lr/-/lr-1.4.11.tgz",
+      "integrity": "sha512-UwgZcl1R66H2hDzBrA7kWyYmX6s8p5TQ+e/t7CuvgCtBEPxbEPaHYHu+pJBvTvDotEJpuQkJjbXuHiw5Cixxgw==",
       "license": "MIT",
       "dependencies": {
         "@lezer/common": "^1.0.0"
       }
     },
     "node_modules/@marijn/find-cluster-break": {
-      "version": "1.0.2",
-      "resolved": "https://registry.npmjs.org/@marijn/find-cluster-break/-/find-cluster-break-1.0.2.tgz",
-      "integrity": "sha512-l0h88YhZFyKdXIFNfSWpyjStDjGHwZ/U7iobcK1cQQD8sejsONdQtTVU+1wVN1PBw40PiiHB1vA5S7VTfQiP9g==",
+      "version": "1.0.4",
+      "resolved": "https://registry.npmjs.org/@marijn/find-cluster-break/-/find-cluster-break-1.0.4.tgz",
+      "integrity": "sha512-Wy0V7+SGUjnF9/TkiM1hKVDPj7jKXduPNboMVtHTA8dySMURWqfg/JZ9E2Sq8JgSJmkl7k7Qe9FLeMSrSraWmQ==",
       "license": "MIT"
     },
     "node_modules/@nodelib/fs.scandir": {
@@ -1160,19 +1556,39 @@
         "node": ">=14"
       }
     },
+    "node_modules/@sindresorhus/merge-streams": {
+      "version": "4.0.0",
+      "resolved": "https://registry.npmjs.org/@sindresorhus/merge-streams/-/merge-streams-4.0.0.tgz",
+      "integrity": "sha512-tlqY9xq5ukxTUZBmoOp+m61cqwQD5pHJtFY3Mn8CA8ps6yghLH/Hw8UPdqg4OLmFW3IFlcXnQNmo/dh8HzXYIQ==",
+      "dev": true,
+      "license": "MIT",
+      "engines": {
+        "node": ">=18"
+      },
+      "funding": {
+        "url": "https://github.com/sponsors/sindresorhus"
+      }
+    },
     "node_modules/@types/debug": {
-      "version": "4.1.12",
-      "resolved": "https://registry.npmjs.org/@types/debug/-/debug-4.1.12.tgz",
-      "integrity": "sha512-vIChWdVG3LG1SMxEvI/AK+FWJthlrqlTu7fbrlywTkkaONwk/UAGaULXRlf8vkzFBLVm0zkMdCquhL5aOjhXPQ==",
+      "version": "4.1.13",
+      "resolved": "https://registry.npmjs.org/@types/debug/-/debug-4.1.13.tgz",
+      "integrity": "sha512-KSVgmQmzMwPlmtljOomayoR89W4FynCAi3E8PPs7vmDVPe84hT+vGPKkJfThkmXs0x0jAaa9U8uW8bbfyS2fWw==",
       "license": "MIT",
       "dependencies": {
         "@types/ms": "*"
       }
     },
+    "node_modules/@types/esrecurse": {
+      "version": "4.3.1",
+      "resolved": "https://registry.npmjs.org/@types/esrecurse/-/esrecurse-4.3.1.tgz",
+      "integrity": "sha512-xJBAbDifo5hpffDBuHl0Y8ywswbiAp/Wi7Y/GtAgSlZyIABppyurxVueOPE8LUQOxdlgi6Zqce7uoEpqNTeiUw==",
+      "dev": true,
+      "license": "MIT"
+    },
     "node_modules/@types/estree": {
-      "version": "1.0.8",
-      "resolved": "https://registry.npmjs.org/@types/estree/-/estree-1.0.8.tgz",
-      "integrity": "sha512-dWHzHa2WqEXI/O1E9OjrocMTKJl2mSrEolh1Iomrv6U+JuNwaHXsXx9bLu5gG7BUWFIN0skIQJQ/L1rIex4X6w==",
+      "version": "1.0.9",
+      "resolved": "https://registry.npmjs.org/@types/estree/-/estree-1.0.9.tgz",
+      "integrity": "sha512-GhdPgy1el4/ImP05X05Uw4cw2/M93BCUmnEvWZNStlCzEKME4Fkk+YpoA5OiHNQmoS7Cafb8Xa3Pya8m1Qrzeg==",
       "license": "MIT"
     },
     "node_modules/@types/estree-jsx": {
@@ -1185,18 +1601,18 @@
       }
     },
     "node_modules/@types/hast": {
-      "version": "3.0.4",
-      "resolved": "https://registry.npmjs.org/@types/hast/-/hast-3.0.4.tgz",
-      "integrity": "sha512-WPs+bbQw5aCj+x6laNGWLH3wviHtoCv/P3+otBhbOhJgG8qtpdAMlTCxLtsTWA7LH1Oh/bFCHsBn0TPS5m30EQ==",
+      "version": "3.0.5",
+      "resolved": "https://registry.npmjs.org/@types/hast/-/hast-3.0.5.tgz",
+      "integrity": "sha512-rp/ezSWaD1m44dPKICGhiskI13nVr7qTloFwDa/IYkhhf5nzwP+zIQcIJh3WIFSBOy/H1PzB40jPjMDksN4F+g==",
       "license": "MIT",
       "dependencies": {
         "@types/unist": "*"
       }
     },
     "node_modules/@types/hotwired__turbo": {
-      "version": "8.0.4",
-      "resolved": "https://registry.npmjs.org/@types/hotwired__turbo/-/hotwired__turbo-8.0.4.tgz",
-      "integrity": "sha512-kgEKOms0taywPxPQvXUh90+G7M+phBHTiUw95Dosw0FU/we86iFerMfgBGcf3eQcAGs5pTNYnLMhiIlfRw/AfA==",
+      "version": "8.0.16",
+      "resolved": "https://registry.npmjs.org/@types/hotwired__turbo/-/hotwired__turbo-8.0.16.tgz",
+      "integrity": "sha512-FOYSD+lq8reFMIiZAp8O5ooKf5Lv9hdrEi0OPKfAgBp4Y7KpNfCbPzyalWbnyR05FepztJatz4PHK7nbsGi+vQ==",
       "dev": true,
       "license": "MIT"
     },
@@ -1223,32 +1639,32 @@
       "license": "MIT"
     },
     "node_modules/@types/node": {
-      "version": "22.19.17",
-      "resolved": "https://registry.npmjs.org/@types/node/-/node-22.19.17.tgz",
-      "integrity": "sha512-wGdMcf+vPYM6jikpS/qhg6WiqSV/OhG+jeeHT/KlVqxYfD40iYJf9/AE1uQxVWFvU7MipKRkRv8NSHiCGgPr8Q==",
+      "version": "26.6.5",
+      "resolved": "https://registry.npmjs.org/@types/node/-/node-26.6.5.tgz",
+      "integrity": "sha512-uMzsv9stDdofIZBx4UGh4Z9ZgXnGC9seRDscilBcTHAeOFk67MZg/k9bWVfO+bk4qZmG+l52hRnLwUiH2LnJTg==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
-        "undici-types": "~6.21.0"
+        "undici-types": "~8.9.0"
       }
     },
     "node_modules/@types/react": {
-      "version": "19.1.10",
-      "resolved": "https://registry.npmjs.org/@types/react/-/react-19.1.10.tgz",
-      "integrity": "sha512-EhBeSYX0Y6ye8pNebpKrwFJq7BoQ8J5SO6NlvNwwHjSj6adXJViPQrKlsyPw7hLBLvckEMO1yxeGdR82YBBlDg==",
+      "version": "19.3.0",
+      "resolved": "https://registry.npmjs.org/@types/react/-/react-19.3.0.tgz",
+      "integrity": "sha512-N0rFCuH9YoxG9/m61l9MfpJKfmLOVU0em7ipIz6TRgSSkvReLB9vL85GB+yr8Bs5leqpvg96JSwF4ZS1s4viQg==",
       "license": "MIT",
       "dependencies": {
-        "csstype": "^3.0.2"
+        "csstype": "^3.2.2"
       }
     },
     "node_modules/@types/react-dom": {
-      "version": "19.1.7",
-      "resolved": "https://registry.npmjs.org/@types/react-dom/-/react-dom-19.1.7.tgz",
-      "integrity": "sha512-i5ZzwYpqjmrKenzkoLM2Ibzt6mAsM7pxB6BCIouEVVmgiqaMj1TjaK7hnA36hbW5aZv20kx7Lw6hWzPWg0Rurw==",
+      "version": "19.3.0",
+      "resolved": "https://registry.npmjs.org/@types/react-dom/-/react-dom-19.3.0.tgz",
+      "integrity": "sha512-ZI7bU42mZXXKHn/qNLEw2IrbiINU7X5+vfgdixBHkCNpYWXjKgfQ/P+uyGb5CjOLB9UcnTeg3rylQtV2hym44Q==",
       "dev": true,
       "license": "MIT",
       "peerDependencies": {
-        "@types/react": "^19.0.0"
+        "@types/react": "^19.3.0"
       }
     },
     "node_modules/@types/unist": {
@@ -1258,21 +1674,20 @@
       "license": "MIT"
     },
     "node_modules/@typescript-eslint/eslint-plugin": {
-      "version": "8.39.1",
-      "resolved": "https://registry.npmjs.org/@typescript-eslint/eslint-plugin/-/eslint-plugin-8.39.1.tgz",
-      "integrity": "sha512-yYegZ5n3Yr6eOcqgj2nJH8cH/ZZgF+l0YIdKILSDjYFRjgYQMgv/lRjV5Z7Up04b9VYUondt8EPMqg7kTWgJ2g==",
+      "version": "8.72.0",
+      "resolved": "https://registry.npmjs.org/@typescript-eslint/eslint-plugin/-/eslint-plugin-8.72.0.tgz",
+      "integrity": "sha512-aysKgysPX9XOv/dk95UkvUDe++Q3n3dJPG9CEWT3RUccxfkPJ+Qn6nhKPDhZXU6Ofo9kgwGWwKyQCnZIXTgngg==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
-        "@eslint-community/regexpp": "^4.10.0",
-        "@typescript-eslint/scope-manager": "8.39.1",
-        "@typescript-eslint/type-utils": "8.39.1",
-        "@typescript-eslint/utils": "8.39.1",
-        "@typescript-eslint/visitor-keys": "8.39.1",
-        "graphemer": "^1.4.0",
-        "ignore": "^7.0.0",
+        "@eslint-community/regexpp": "^4.12.2",
+        "@typescript-eslint/scope-manager": "8.72.0",
+        "@typescript-eslint/type-utils": "8.72.0",
+        "@typescript-eslint/utils": "8.72.0",
+        "@typescript-eslint/visitor-keys": "8.72.0",
+        "ignore": "^7.0.5",
         "natural-compare": "^1.4.0",
-        "ts-api-utils": "^2.1.0"
+        "ts-api-utils": "^2.5.0"
       },
       "engines": {
         "node": "^18.18.0 || ^20.9.0 || >=21.1.0"
@@ -1282,15 +1697,15 @@
         "url": "https://opencollective.com/typescript-eslint"
       },
       "peerDependencies": {
-        "@typescript-eslint/parser": "^8.39.1",
-        "eslint": "^8.57.0 || ^9.0.0",
-        "typescript": ">=4.8.4 <6.0.0"
+        "@typescript-eslint/parser": "^8.72.0",
+        "eslint": "^8.57.0 || ^9.0.0 || ^10.0.0",
+        "typescript": ">=4.8.4 <6.1.0"
       }
     },
     "node_modules/@typescript-eslint/eslint-plugin/node_modules/ignore": {
-      "version": "7.0.5",
-      "resolved": "https://registry.npmjs.org/ignore/-/ignore-7.0.5.tgz",
-      "integrity": "sha512-Hs59xBNfUIunMFgWAbGX5cq6893IbWg4KnrjbYwX3tx0ztorVgTDA6B2sxf8ejHJ4wz8BqGUMYlnzNBer5NvGg==",
+      "version": "7.0.14",
+      "resolved": "https://registry.npmjs.org/ignore/-/ignore-7.0.14.tgz",
+      "integrity": "sha512-wCBiZ77u1tuWde6G/nfOtLaCTYcFDU5yCzARzK10y+rLhzYoL7a7yEemcopcc2Mzow7gZOIjnV2ym1NYmdiJOQ==",
       "dev": true,
       "license": "MIT",
       "engines": {
@@ -1298,17 +1713,17 @@
       }
     },
     "node_modules/@typescript-eslint/parser": {
-      "version": "8.39.1",
-      "resolved": "https://registry.npmjs.org/@typescript-eslint/parser/-/parser-8.39.1.tgz",
-      "integrity": "sha512-pUXGCuHnnKw6PyYq93lLRiZm3vjuslIy7tus1lIQTYVK9bL8XBgJnCWm8a0KcTtHC84Yya1Q6rtll+duSMj0dg==",
+      "version": "8.72.0",
+      "resolved": "https://registry.npmjs.org/@typescript-eslint/parser/-/parser-8.72.0.tgz",
+      "integrity": "sha512-0O+x3ynciD6zuXb1CKSu165haQV6ey53DKhHHUnujP7U+XlltJG+/keT0Xhd1HXCze+IHf7Tvt39ovyepX/LNw==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
-        "@typescript-eslint/scope-manager": "8.39.1",
-        "@typescript-eslint/types": "8.39.1",
-        "@typescript-eslint/typescript-estree": "8.39.1",
-        "@typescript-eslint/visitor-keys": "8.39.1",
-        "debug": "^4.3.4"
+        "@typescript-eslint/scope-manager": "8.72.0",
+        "@typescript-eslint/types": "8.72.0",
+        "@typescript-eslint/typescript-estree": "8.72.0",
+        "@typescript-eslint/visitor-keys": "8.72.0",
+        "debug": "^4.4.3"
       },
       "engines": {
         "node": "^18.18.0 || ^20.9.0 || >=21.1.0"
@@ -1318,20 +1733,20 @@
         "url": "https://opencollective.com/typescript-eslint"
       },
       "peerDependencies": {
-        "eslint": "^8.57.0 || ^9.0.0",
-        "typescript": ">=4.8.4 <6.0.0"
+        "eslint": "^8.57.0 || ^9.0.0 || ^10.0.0",
+        "typescript": ">=4.8.4 <6.1.0"
       }
     },
     "node_modules/@typescript-eslint/project-service": {
-      "version": "8.39.1",
-      "resolved": "https://registry.npmjs.org/@typescript-eslint/project-service/-/project-service-8.39.1.tgz",
-      "integrity": "sha512-8fZxek3ONTwBu9ptw5nCKqZOSkXshZB7uAxuFF0J/wTMkKydjXCzqqga7MlFMpHi9DoG4BadhmTkITBcg8Aybw==",
+      "version": "8.72.0",
+      "resolved": "https://registry.npmjs.org/@typescript-eslint/project-service/-/project-service-8.72.0.tgz",
+      "integrity": "sha512-bzHngBbmIgzj6TsTjVUCntGx6FFLgE7mPjpO0PzHycOaZHXqS43XnrkFIHRS16LLr/xO6FHSXG3GJFo/F4bSBQ==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
-        "@typescript-eslint/tsconfig-utils": "^8.39.1",
-        "@typescript-eslint/types": "^8.39.1",
-        "debug": "^4.3.4"
+        "@typescript-eslint/tsconfig-utils": "^8.72.0",
+        "@typescript-eslint/types": "^8.72.0",
+        "debug": "^4.4.3"
       },
       "engines": {
         "node": "^18.18.0 || ^20.9.0 || >=21.1.0"
@@ -1341,18 +1756,18 @@
         "url": "https://opencollective.com/typescript-eslint"
       },
       "peerDependencies": {
-        "typescript": ">=4.8.4 <6.0.0"
+        "typescript": ">=4.8.4 <6.1.0"
       }
     },
     "node_modules/@typescript-eslint/scope-manager": {
-      "version": "8.39.1",
-      "resolved": "https://registry.npmjs.org/@typescript-eslint/scope-manager/-/scope-manager-8.39.1.tgz",
-      "integrity": "sha512-RkBKGBrjgskFGWuyUGz/EtD8AF/GW49S21J8dvMzpJitOF1slLEbbHnNEtAHtnDAnx8qDEdRrULRnWVx27wGBw==",
+      "version": "8.72.0",
+      "resolved": "https://registry.npmjs.org/@typescript-eslint/scope-manager/-/scope-manager-8.72.0.tgz",
+      "integrity": "sha512-JewPyJOYcXBSAS7PIgufscHOwZqhTGyBRiQJhhuDln66tMchxvvt7nJ90JKKIoOelmdHA3kAQbm9cEUHc4JSEA==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
-        "@typescript-eslint/types": "8.39.1",
-        "@typescript-eslint/visitor-keys": "8.39.1"
+        "@typescript-eslint/types": "8.72.0",
+        "@typescript-eslint/visitor-keys": "8.72.0"
       },
       "engines": {
         "node": "^18.18.0 || ^20.9.0 || >=21.1.0"
@@ -1363,9 +1778,9 @@
       }
     },
     "node_modules/@typescript-eslint/tsconfig-utils": {
-      "version": "8.39.1",
-      "resolved": "https://registry.npmjs.org/@typescript-eslint/tsconfig-utils/-/tsconfig-utils-8.39.1.tgz",
-      "integrity": "sha512-ePUPGVtTMR8XMU2Hee8kD0Pu4NDE1CN9Q1sxGSGd/mbOtGZDM7pnhXNJnzW63zk/q+Z54zVzj44HtwXln5CvHA==",
+      "version": "8.72.0",
+      "resolved": "https://registry.npmjs.org/@typescript-eslint/tsconfig-utils/-/tsconfig-utils-8.72.0.tgz",
+      "integrity": "sha512-0vbxCf7xLVPSjcOdm6eQbgViH8EDJ8xnl1laLT9yoqSz2YmcSfSg5Vdl0JUvZf6rcuJ0ZivJIuMD4edTg8obaA==",
       "dev": true,
       "license": "MIT",
       "engines": {
@@ -1376,21 +1791,21 @@
         "url": "https://opencollective.com/typescript-eslint"
       },
       "peerDependencies": {
-        "typescript": ">=4.8.4 <6.0.0"
+        "typescript": ">=4.8.4 <6.1.0"
       }
     },
     "node_modules/@typescript-eslint/type-utils": {
-      "version": "8.39.1",
-      "resolved": "https://registry.npmjs.org/@typescript-eslint/type-utils/-/type-utils-8.39.1.tgz",
-      "integrity": "sha512-gu9/ahyatyAdQbKeHnhT4R+y3YLtqqHyvkfDxaBYk97EcbfChSJXyaJnIL3ygUv7OuZatePHmQvuH5ru0lnVeA==",
+      "version": "8.72.0",
+      "resolved": "https://registry.npmjs.org/@typescript-eslint/type-utils/-/type-utils-8.72.0.tgz",
+      "integrity": "sha512-gEXxmmbSTSOVTC37IYYvremOM31v+UN2W772bJxmVYPhAOPSUxtacyb7wkm1EhkYKU6KzFJsdU44DCV4to0LWw==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
-        "@typescript-eslint/types": "8.39.1",
-        "@typescript-eslint/typescript-estree": "8.39.1",
-        "@typescript-eslint/utils": "8.39.1",
-        "debug": "^4.3.4",
-        "ts-api-utils": "^2.1.0"
+        "@typescript-eslint/types": "8.72.0",
+        "@typescript-eslint/typescript-estree": "8.72.0",
+        "@typescript-eslint/utils": "8.72.0",
+        "debug": "^4.4.3",
+        "ts-api-utils": "^2.5.0"
       },
       "engines": {
         "node": "^18.18.0 || ^20.9.0 || >=21.1.0"
@@ -1400,14 +1815,14 @@
         "url": "https://opencollective.com/typescript-eslint"
       },
       "peerDependencies": {
-        "eslint": "^8.57.0 || ^9.0.0",
-        "typescript": ">=4.8.4 <6.0.0"
+        "eslint": "^8.57.0 || ^9.0.0 || ^10.0.0",
+        "typescript": ">=4.8.4 <6.1.0"
       }
     },
     "node_modules/@typescript-eslint/types": {
-      "version": "8.39.1",
-      "resolved": "https://registry.npmjs.org/@typescript-eslint/types/-/types-8.39.1.tgz",
-      "integrity": "sha512-7sPDKQQp+S11laqTrhHqeAbsCfMkwJMrV7oTDvtDds4mEofJYir414bYKUEb8YPUm9QL3U+8f6L6YExSoAGdQw==",
+      "version": "8.72.0",
+      "resolved": "https://registry.npmjs.org/@typescript-eslint/types/-/types-8.72.0.tgz",
+      "integrity": "sha512-Adfoy59kxNEjT5E5fnyrp7GpHYuWs8ONbdohu4ez6IU46qDq8LltG90tLmkucZB2uhYL8mqBjtXqGd8yhfjXqQ==",
       "dev": true,
       "license": "MIT",
       "engines": {
@@ -1419,22 +1834,21 @@
       }
     },
     "node_modules/@typescript-eslint/typescript-estree": {
-      "version": "8.39.1",
-      "resolved": "https://registry.npmjs.org/@typescript-eslint/typescript-estree/-/typescript-estree-8.39.1.tgz",
-      "integrity": "sha512-EKkpcPuIux48dddVDXyQBlKdeTPMmALqBUbEk38McWv0qVEZwOpVJBi7ugK5qVNgeuYjGNQxrrnoM/5+TI/BPw==",
+      "version": "8.72.0",
+      "resolved": "https://registry.npmjs.org/@typescript-eslint/typescript-estree/-/typescript-estree-8.72.0.tgz",
+      "integrity": "sha512-gwLfFgWcw15tN2KXDCJvxEOQjPn+P3j/n3eSqkZ5sX4VpW9Cosodr+tWwBZRvJT3TcKx6HnPRGZMCNAiP1Xhug==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
-        "@typescript-eslint/project-service": "8.39.1",
-        "@typescript-eslint/tsconfig-utils": "8.39.1",
-        "@typescript-eslint/types": "8.39.1",
-        "@typescript-eslint/visitor-keys": "8.39.1",
-        "debug": "^4.3.4",
-        "fast-glob": "^3.3.2",
-        "is-glob": "^4.0.3",
-        "minimatch": "^9.0.4",
-        "semver": "^7.6.0",
-        "ts-api-utils": "^2.1.0"
+        "@typescript-eslint/project-service": "8.72.0",
+        "@typescript-eslint/tsconfig-utils": "8.72.0",
+        "@typescript-eslint/types": "8.72.0",
+        "@typescript-eslint/visitor-keys": "8.72.0",
+        "debug": "^4.4.3",
+        "minimatch": "^10.2.2",
+        "semver": "^7.7.3",
+        "tinyglobby": "^0.2.15",
+        "ts-api-utils": "^2.5.0"
       },
       "engines": {
         "node": "^18.18.0 || ^20.9.0 || >=21.1.0"
@@ -1444,46 +1858,33 @@
         "url": "https://opencollective.com/typescript-eslint"
       },
       "peerDependencies": {
-        "typescript": ">=4.8.4 <6.0.0"
-      }
-    },
-    "node_modules/@typescript-eslint/typescript-estree/node_modules/brace-expansion": {
-      "version": "2.1.7",
-      "resolved": "https://registry.npmjs.org/brace-expansion/-/brace-expansion-2.1.7.tgz",
-      "integrity": "sha512-uZbew1NqdmPDTMJ8ah1y+b+9QEJrfkXFk3RcTQw3X0jW/xRUvFKsg1CfQdSYGdTbXZWExtU3J3ccxtnfw1Fi0g==",
-      "dev": true,
-      "license": "MIT",
-      "dependencies": {
-        "balanced-match": "^1.0.0"
+        "typescript": ">=4.8.4 <6.1.0"
       }
     },
-    "node_modules/@typescript-eslint/typescript-estree/node_modules/minimatch": {
-      "version": "9.0.9",
-      "resolved": "https://registry.npmjs.org/minimatch/-/minimatch-9.0.9.tgz",
-      "integrity": "sha512-OBwBN9AL4dqmETlpS2zasx+vTeWclWzkblfZk7KTA5j3jeOONz/tRCnZomUyvNg83wL5Zv9Ss6HMJXAgL8R2Yg==",
+    "node_modules/@typescript-eslint/typescript-estree/node_modules/semver": {
+      "version": "7.8.5",
+      "resolved": "https://registry.npmjs.org/semver/-/semver-7.8.5.tgz",
+      "integrity": "sha512-Y7/KDsb8LjooZpwaqGyulO6DQlksgCncchHGk+sZIY4SBvUocMBEFH5Ur1fI4dV+Jvl0w6cjvucaIi40puRioA==",
       "dev": true,
       "license": "ISC",
-      "dependencies": {
-        "brace-expansion": "^2.0.2"
+      "bin": {
+        "semver": "bin/semver.js"
       },
       "engines": {
-        "node": ">=16 || 14 >=14.17"
-      },
-      "funding": {
-        "url": "https://github.com/sponsors/isaacs"
+        "node": ">=10"
       }
     },
     "node_modules/@typescript-eslint/utils": {
-      "version": "8.39.1",
-      "resolved": "https://registry.npmjs.org/@typescript-eslint/utils/-/utils-8.39.1.tgz",
-      "integrity": "sha512-VF5tZ2XnUSTuiqZFXCZfZs1cgkdd3O/sSYmdo2EpSyDlC86UM/8YytTmKnehOW3TGAlivqTDT6bS87B/GQ/jyg==",
+      "version": "8.72.0",
+      "resolved": "https://registry.npmjs.org/@typescript-eslint/utils/-/utils-8.72.0.tgz",
+      "integrity": "sha512-0up3oArJZBYnzOaEXKo8DmI43QvMZyPoFEt9LWzRJcyl/RcJjfKKPebknN1VbTgXBf3M/KggUh5uUMdURLnseg==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
-        "@eslint-community/eslint-utils": "^4.7.0",
-        "@typescript-eslint/scope-manager": "8.39.1",
-        "@typescript-eslint/types": "8.39.1",
-        "@typescript-eslint/typescript-estree": "8.39.1"
+        "@eslint-community/eslint-utils": "^4.9.1",
+        "@typescript-eslint/scope-manager": "8.72.0",
+        "@typescript-eslint/types": "8.72.0",
+        "@typescript-eslint/typescript-estree": "8.72.0"
       },
       "engines": {
         "node": "^18.18.0 || ^20.9.0 || >=21.1.0"
@@ -1493,19 +1894,19 @@
         "url": "https://opencollective.com/typescript-eslint"
       },
       "peerDependencies": {
-        "eslint": "^8.57.0 || ^9.0.0",
-        "typescript": ">=4.8.4 <6.0.0"
+        "eslint": "^8.57.0 || ^9.0.0 || ^10.0.0",
+        "typescript": ">=4.8.4 <6.1.0"
       }
     },
     "node_modules/@typescript-eslint/visitor-keys": {
-      "version": "8.39.1",
-      "resolved": "https://registry.npmjs.org/@typescript-eslint/visitor-keys/-/visitor-keys-8.39.1.tgz",
-      "integrity": "sha512-W8FQi6kEh2e8zVhQ0eeRnxdvIoOkAp/CPAahcNio6nO9dsIwb9b34z90KOlheoyuVf6LSOEdjlkxSkapNEc+4A==",
+      "version": "8.72.0",
+      "resolved": "https://registry.npmjs.org/@typescript-eslint/visitor-keys/-/visitor-keys-8.72.0.tgz",
+      "integrity": "sha512-E2mroviasaWecDzTGiGMtuEguuLVOeRggnrIn93CJuxYJzQBHbyymlT08aVQydbBal5538o+JkBKXh2sRuCyIA==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
-        "@typescript-eslint/types": "8.39.1",
-        "eslint-visitor-keys": "^4.2.1"
+        "@typescript-eslint/types": "8.72.0",
+        "eslint-visitor-keys": "^5.0.0"
       },
       "engines": {
         "node": "^18.18.0 || ^20.9.0 || >=21.1.0"
@@ -1516,9 +1917,9 @@
       }
     },
     "node_modules/@uiw/codemirror-extensions-basic-setup": {
-      "version": "4.24.2",
-      "resolved": "https://registry.npmjs.org/@uiw/codemirror-extensions-basic-setup/-/codemirror-extensions-basic-setup-4.24.2.tgz",
-      "integrity": "sha512-wW/gjLRvVUeYyhdh2TApn25cvdcR+Rhg6R/j3eTOvXQzU1HNzNYCVH4YKVIfgtfdM/Xs+N8fkk+rbr1YvBppCg==",
+      "version": "4.25.12",
+      "resolved": "https://registry.npmjs.org/@uiw/codemirror-extensions-basic-setup/-/codemirror-extensions-basic-setup-4.25.12.tgz",
+      "integrity": "sha512-wR5lz7jbVcBGyiiLEObAmDikcc39KNqrrd4bdUhav85WZu48ttgnSf1YV+grHuMsj5l/WCw6u+r9ZuSfURJgiA==",
       "license": "MIT",
       "dependencies": {
         "@codemirror/autocomplete": "^6.0.0",
@@ -1543,16 +1944,16 @@
       }
     },
     "node_modules/@uiw/react-codemirror": {
-      "version": "4.24.2",
-      "resolved": "https://registry.npmjs.org/@uiw/react-codemirror/-/react-codemirror-4.24.2.tgz",
-      "integrity": "sha512-kp7DhTq4RR+M2zJBQBrHn1dIkBrtOskcwJX4vVsKGByReOvfMrhqRkGTxYMRDTX6x75EG2mvBJPDKYcUQcHWBw==",
+      "version": "4.25.12",
+      "resolved": "https://registry.npmjs.org/@uiw/react-codemirror/-/react-codemirror-4.25.12.tgz",
+      "integrity": "sha512-RE0Gj/QesD7/JWPnJREG8Kq8brGv2NAH17iUECcBrLkqXzzEAgsj/4Qqvz/OlK8XWKmzlsenLRkObT9H+gYQ0w==",
       "license": "MIT",
       "dependencies": {
         "@babel/runtime": "^7.18.6",
         "@codemirror/commands": "^6.1.0",
         "@codemirror/state": "^6.1.1",
         "@codemirror/theme-one-dark": "^6.0.0",
-        "@uiw/codemirror-extensions-basic-setup": "4.24.2",
+        "@uiw/codemirror-extensions-basic-setup": "4.25.12",
         "codemirror": "^6.0.0"
       },
       "funding": {
@@ -1569,15 +1970,15 @@
       }
     },
     "node_modules/@ungap/structured-clone": {
-      "version": "1.3.0",
-      "resolved": "https://registry.npmjs.org/@ungap/structured-clone/-/structured-clone-1.3.0.tgz",
-      "integrity": "sha512-WmoN8qaIAo7WTYWbAZuG8PYEhn5fkz7dZrqTBZ7dtt//lL2Gwms1IcnQ5yHqjDfX8Ft5j4YzDM23f87zBfDe9g==",
+      "version": "1.4.0",
+      "resolved": "https://registry.npmjs.org/@ungap/structured-clone/-/structured-clone-1.4.0.tgz",
+      "integrity": "sha512-1mEZtMKPM09vDmQt5y7YvmN2+DFTP7Tg0EWXdic8/C6VRnpb33e4ghisCIE3WZjsE2N8mf+QV1Zqh7ZFYLWInQ==",
       "license": "ISC"
     },
     "node_modules/acorn": {
-      "version": "8.15.0",
-      "resolved": "https://registry.npmjs.org/acorn/-/acorn-8.15.0.tgz",
-      "integrity": "sha512-NZyJarBfL7nWwIq+FDL6Zp/yHEhePMNnnJ0y3qfieCrmNvYct8uvtiV41UvlSe6apAfk0fY1FbWx+NwfmpvtTg==",
+      "version": "8.19.0",
+      "resolved": "https://registry.npmjs.org/acorn/-/acorn-8.19.0.tgz",
+      "integrity": "sha512-oJlA3XiRm7Cyk6qFD2Jn8ak9B7jSy0qy00ADO3+8dpT0LSjFihQYv4C02LFCSYJV3Q37xYwwRy5m+IpIiUzqWw==",
       "dev": true,
       "license": "MIT",
       "bin": {
@@ -1598,9 +1999,9 @@
       }
     },
     "node_modules/ajv": {
-      "version": "6.14.0",
-      "resolved": "https://registry.npmjs.org/ajv/-/ajv-6.14.0.tgz",
-      "integrity": "sha512-IWrosm/yrn43eiKqkfkHis7QioDleaXQHdDVPKg0FSwwd/DuvyX79TZnFOnYpB7dcsFAMmtFztZuXPDvSePkFw==",
+      "version": "6.15.0",
+      "resolved": "https://registry.npmjs.org/ajv/-/ajv-6.15.0.tgz",
+      "integrity": "sha512-fgFx7Hfoq60ytK2c7DhnF8jIvzYgOMxfugjLOSMHjLIPgenqa7S7oaagATUq99mV6IYvN2tRmC0wnTYX6iPbMw==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
@@ -1615,9 +2016,9 @@
       }
     },
     "node_modules/ansi-regex": {
-      "version": "6.1.0",
-      "resolved": "https://registry.npmjs.org/ansi-regex/-/ansi-regex-6.1.0.tgz",
-      "integrity": "sha512-7HSX4QQb4CspciLpVFwyRe79O3xsIZDDLER21kERQ71oaPodF8jL725AgJMFAYbooIqolJoRLuM81SpeUkpkvA==",
+      "version": "6.4.0",
+      "resolved": "https://registry.npmjs.org/ansi-regex/-/ansi-regex-6.4.0.tgz",
+      "integrity": "sha512-KzTVk2tCWAHtYrvvvaP8bJKJq2pVinhLcGEQdtLIYPbmNGNyYe8QwNaTUYQp2J7/vIsUKt5QCqAfUkYyG9DkOw==",
       "dev": true,
       "license": "MIT",
       "engines": {
@@ -1628,13 +2029,16 @@
       }
     },
     "node_modules/ansi-styles": {
-      "version": "6.2.1",
-      "resolved": "https://registry.npmjs.org/ansi-styles/-/ansi-styles-6.2.1.tgz",
-      "integrity": "sha512-bN798gFfQX+viw3R7yrGWRqnrN2oRkEkUjjl4JNn4E8GxxbjtG3FbrEIIY3l8/hrwUwIeCZvi4QuOTP4MErVug==",
+      "version": "4.3.0",
+      "resolved": "https://registry.npmjs.org/ansi-styles/-/ansi-styles-4.3.0.tgz",
+      "integrity": "sha512-zbB9rCJAT1rbjiVDb2hqKFHNYLxgtk8NURxZ3IZwD3F6NtxbXZQCnnSi1Lkx+IDohdPlFp222wVALIheZJQSEg==",
       "dev": true,
       "license": "MIT",
+      "dependencies": {
+        "color-convert": "^2.0.1"
+      },
       "engines": {
-        "node": ">=12"
+        "node": ">=8"
       },
       "funding": {
         "url": "https://github.com/chalk/ansi-styles?sponsor=1"
@@ -1654,19 +2058,6 @@
         "node": ">= 8"
       }
     },
-    "node_modules/anymatch/node_modules/picomatch": {
-      "version": "2.3.2",
-      "resolved": "https://registry.npmjs.org/picomatch/-/picomatch-2.3.2.tgz",
-      "integrity": "sha512-V7+vQEJ06Z+c5tSye8S+nHUfI51xoXIXjHQ99cQtKUkQqqO1kO/KCJUfZXuB47h/YBlDhah2H3hdUGXn8ie0oA==",
-      "dev": true,
-      "license": "MIT",
-      "engines": {
-        "node": ">=8.6"
-      },
-      "funding": {
-        "url": "https://github.com/sponsors/jonschlinkert"
-      }
-    },
     "node_modules/argparse": {
       "version": "2.0.1",
       "resolved": "https://registry.npmjs.org/argparse/-/argparse-2.0.1.tgz",
@@ -1674,16 +2065,6 @@
       "dev": true,
       "license": "Python-2.0"
     },
-    "node_modules/array-union": {
-      "version": "2.1.0",
-      "resolved": "https://registry.npmjs.org/array-union/-/array-union-2.1.0.tgz",
-      "integrity": "sha512-HGyxoOTYUyCM6stUe6EJgnd4EoewAI7zMdfqO+kGjnlZmBDz/cR5pf8r/cR4Wq60sL/p0IkcjUEEPwS3GFrIyw==",
-      "dev": true,
-      "license": "MIT",
-      "engines": {
-        "node": ">=8"
-      }
-    },
     "node_modules/astral-regex": {
       "version": "2.0.0",
       "resolved": "https://registry.npmjs.org/astral-regex/-/astral-regex-2.0.0.tgz",
@@ -1705,11 +2086,27 @@
       }
     },
     "node_modules/balanced-match": {
-      "version": "1.0.2",
-      "resolved": "https://registry.npmjs.org/balanced-match/-/balanced-match-1.0.2.tgz",
-      "integrity": "sha512-3oSeUO0TMV67hN1AmbXsK4yaqU7tjiHlbxRDZOpH0KW9+CeX4bRAaX0Anxt0tx2MrpRpWwQaPwIlISEJhYU5Pw==",
+      "version": "4.0.4",
+      "resolved": "https://registry.npmjs.org/balanced-match/-/balanced-match-4.0.4.tgz",
+      "integrity": "sha512-BLrgEcRTwX2o6gGxGOCNyMvGSp35YofuYzw9h1IMTRmKqttAZZVU67bdb9Pr2vUHA8+j3i2tJfjO6C6+4myGTA==",
       "dev": true,
-      "license": "MIT"
+      "license": "MIT",
+      "engines": {
+        "node": "18 || 20 || >=22"
+      }
+    },
+    "node_modules/baseline-browser-mapping": {
+      "version": "2.11.32",
+      "resolved": "https://registry.npmjs.org/baseline-browser-mapping/-/baseline-browser-mapping-2.11.32.tgz",
+      "integrity": "sha512-UvuxFU9P3hM29LrQdwu0cD6w6/+SFr+UlB9s/knI7+O2rK9xFxKUwSDWxGdbbqQ9/V7BJ5s0NI5nQAE4uJp3gQ==",
+      "dev": true,
+      "license": "Apache-2.0",
+      "bin": {
+        "baseline-browser-mapping": "dist/cli.cjs"
+      },
+      "engines": {
+        "node": ">=6.0.0"
+      }
     },
     "node_modules/binary-extensions": {
       "version": "2.3.0",
@@ -1725,14 +2122,16 @@
       }
     },
     "node_modules/brace-expansion": {
-      "version": "1.1.21",
-      "resolved": "https://registry.npmjs.org/brace-expansion/-/brace-expansion-1.1.21.tgz",
-      "integrity": "sha512-9zeA+KLZNNzglF2TPKRQEDyx6Yby7daAkuy8MiPzpXPsYDWi/DRM8jmwUDxokQjYqBpv5DgPiwD4h4ZZSy1Ujw==",
+      "version": "5.0.12",
+      "resolved": "https://registry.npmjs.org/brace-expansion/-/brace-expansion-5.0.12.tgz",
+      "integrity": "sha512-YovQ3rzhaLMIrDjNDMkNS01tea93qhEhG5xy8f6+R0l+dw3Ki+5sCoIoI942iuLZTHWogWktgwVDhU09iNEimQ==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
-        "balanced-match": "^1.0.0",
-        "concat-map": "0.0.1"
+        "balanced-match": "^4.0.2"
+      },
+      "engines": {
+        "node": "20 || >=22"
       }
     },
     "node_modules/braces": {
@@ -1748,25 +2147,62 @@
         "node": ">=8"
       }
     },
-    "node_modules/cacheable": {
-      "version": "1.10.3",
-      "resolved": "https://registry.npmjs.org/cacheable/-/cacheable-1.10.3.tgz",
-      "integrity": "sha512-M6p10iJ/VT0wT7TLIGUnm958oVrU2cUK8pQAVU21Zu7h8rbk/PeRtRWrvHJBql97Bhzk3g1N6+2VKC+Rjxna9Q==",
+    "node_modules/browserslist": {
+      "version": "4.29.4",
+      "resolved": "https://registry.npmjs.org/browserslist/-/browserslist-4.29.4.tgz",
+      "integrity": "sha512-2TcbbiwXEq+N0kD36tykpCfyyQigUXmuwNQbnPp4Xk4PX9JOqVIKjJlIDjpH7uY6tmRGoAd+apeT1p60T8wsdA==",
       "dev": true,
+      "funding": [
+        {
+          "type": "opencollective",
+          "url": "https://opencollective.com/browserslist"
+        },
+        {
+          "type": "tidelift",
+          "url": "https://tidelift.com/funding/github/npm/browserslist"
+        },
+        {
+          "type": "github",
+          "url": "https://github.com/sponsors/ai"
+        }
+      ],
       "license": "MIT",
       "dependencies": {
-        "hookified": "^1.10.0",
-        "keyv": "^5.4.0"
+        "baseline-browser-mapping": "^2.11.29",
+        "caniuse-lite": "^1.0.30001816",
+        "electron-to-chromium": "^1.5.459",
+        "node-releases": "^2.0.59",
+        "update-browserslist-db": "^1.3.4"
+      },
+      "bin": {
+        "browserslist": "cli.js"
+      },
+      "engines": {
+        "node": "^6 || ^7 || ^8 || ^9 || ^10 || ^11 || ^12 || >=13.7"
       }
     },
-    "node_modules/cacheable/node_modules/keyv": {
-      "version": "5.4.0",
-      "resolved": "https://registry.npmjs.org/keyv/-/keyv-5.4.0.tgz",
-      "integrity": "sha512-TMckyVjEoacG5IteUpUrOBsFORtheqziVyyY2dLUwg1jwTb8u48LX4TgmtogkNl9Y9unaEJ1luj10fGyjMGFOQ==",
+    "node_modules/cacheable": {
+      "version": "2.5.0",
+      "resolved": "https://registry.npmjs.org/cacheable/-/cacheable-2.5.0.tgz",
+      "integrity": "sha512-60cyAOytib/OzBw1JNSoSV/boK1AtHryDIjvVBk7XbN4ugfkM3+Sry7fEjNgPMGgOjuaZPAp8ruZ0Cxafwyq9g==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
-        "@keyv/serialize": "^1.1.0"
+        "@cacheable/memory": "^2.2.0",
+        "@cacheable/utils": "^2.5.0",
+        "hookified": "^1.15.0",
+        "keyv": "^5.6.0",
+        "qified": "^0.10.1"
+      }
+    },
+    "node_modules/callsites": {
+      "version": "3.1.0",
+      "resolved": "https://registry.npmjs.org/callsites/-/callsites-3.1.0.tgz",
+      "integrity": "sha512-P8BjAsXvZS+VIDUI11hHCQEv74YT67YUi5JJFNWIqL235sBmjX4+qx9Muvls5ivyNENctx46xQLQ3aTuE7ssaQ==",
+      "dev": true,
+      "license": "MIT",
+      "engines": {
+        "node": ">=6"
       }
     },
     "node_modules/camelcase": {
@@ -1782,16 +2218,77 @@
         "url": "https://github.com/sponsors/sindresorhus"
       }
     },
-    "node_modules/ccount": {
-      "version": "2.0.1",
-      "resolved": "https://registry.npmjs.org/ccount/-/ccount-2.0.1.tgz",
-      "integrity": "sha512-eyrF0jiFpY+3drT6383f1qhkbGsLSifNAjA61IUjZjmLCWjItY6LB9ft9YhoDgwfmclB2zhu51Lc7+95b8NRAg==",
-      "license": "MIT",
+    "node_modules/caniuse-lite": {
+      "version": "1.0.30001819",
+      "resolved": "https://registry.npmjs.org/caniuse-lite/-/caniuse-lite-1.0.30001819.tgz",
+      "integrity": "sha512-LFUqIIAwpzsIuevnKnsR9qnStBcgzPShTX61lBwN9uFkkf8o8zLiebEjbW8AYtZiF+1YKDWfdt/pWDtMuG8Xwg==",
+      "dev": true,
+      "funding": [
+        {
+          "type": "opencollective",
+          "url": "https://opencollective.com/browserslist"
+        },
+        {
+          "type": "tidelift",
+          "url": "https://tidelift.com/funding/github/npm/caniuse-lite"
+        },
+        {
+          "type": "github",
+          "url": "https://github.com/sponsors/ai"
+        }
+      ],
+      "license": "CC-BY-4.0"
+    },
+    "node_modules/ccount": {
+      "version": "2.0.1",
+      "resolved": "https://registry.npmjs.org/ccount/-/ccount-2.0.1.tgz",
+      "integrity": "sha512-eyrF0jiFpY+3drT6383f1qhkbGsLSifNAjA61IUjZjmLCWjItY6LB9ft9YhoDgwfmclB2zhu51Lc7+95b8NRAg==",
+      "license": "MIT",
       "funding": {
         "type": "github",
         "url": "https://github.com/sponsors/wooorm"
       }
     },
+    "node_modules/chalk": {
+      "version": "4.1.2",
+      "resolved": "https://registry.npmjs.org/chalk/-/chalk-4.1.2.tgz",
+      "integrity": "sha512-oKnbhFyRIXpUuez8iBMmyEa4nbj4IOQyuhc/wy9kY7/WVPcwIO9VA668Pu8RkO7+0G76SLROeyw9CpQ061i4mA==",
+      "dev": true,
+      "license": "MIT",
+      "dependencies": {
+        "ansi-styles": "^4.1.0",
+        "supports-color": "^7.1.0"
+      },
+      "engines": {
+        "node": ">=10"
+      },
+      "funding": {
+        "url": "https://github.com/chalk/chalk?sponsor=1"
+      }
+    },
+    "node_modules/chalk/node_modules/has-flag": {
+      "version": "4.0.0",
+      "resolved": "https://registry.npmjs.org/has-flag/-/has-flag-4.0.0.tgz",
+      "integrity": "sha512-EykJT/Q1KjTWctppgIAgfSO0tKVuZUjhgMr17kqTumMl6Afv3EISleU7qZUzoXDFTAHTDC4NOoG/ZxU3EvlMPQ==",
+      "dev": true,
+      "license": "MIT",
+      "engines": {
+        "node": ">=8"
+      }
+    },
+    "node_modules/chalk/node_modules/supports-color": {
+      "version": "7.2.0",
+      "resolved": "https://registry.npmjs.org/supports-color/-/supports-color-7.2.0.tgz",
+      "integrity": "sha512-qpCAvRl9stuOHveKsn7HncJRvv501qIacKzQlO/+Lwxc9+0q2wLyv4Dfvt80/DPn2pqOBsJdDiogXGR9+OvwRw==",
+      "dev": true,
+      "license": "MIT",
+      "dependencies": {
+        "has-flag": "^4.0.0"
+      },
+      "engines": {
+        "node": ">=8"
+      }
+    },
     "node_modules/character-entities": {
       "version": "2.0.2",
       "resolved": "https://registry.npmjs.org/character-entities/-/character-entities-2.0.2.tgz",
@@ -1873,39 +2370,6 @@
         "node": ">=8"
       }
     },
-    "node_modules/cliui/node_modules/ansi-styles": {
-      "version": "4.3.0",
-      "resolved": "https://registry.npmjs.org/ansi-styles/-/ansi-styles-4.3.0.tgz",
-      "integrity": "sha512-zbB9rCJAT1rbjiVDb2hqKFHNYLxgtk8NURxZ3IZwD3F6NtxbXZQCnnSi1Lkx+IDohdPlFp222wVALIheZJQSEg==",
-      "dev": true,
-      "license": "MIT",
-      "dependencies": {
-        "color-convert": "^2.0.1"
-      },
-      "engines": {
-        "node": ">=8"
-      },
-      "funding": {
-        "url": "https://github.com/chalk/ansi-styles?sponsor=1"
-      }
-    },
-    "node_modules/cliui/node_modules/emoji-regex": {
-      "version": "8.0.0",
-      "resolved": "https://registry.npmjs.org/emoji-regex/-/emoji-regex-8.0.0.tgz",
-      "integrity": "sha512-MSjYzcWNOA0ewAHpz0MxpYFvwg6yjy1NG3xteoqz644VCo/RPgnr1/GGt+ic3iJTzQ8Eu3TdM14SawnVUmGE6A==",
-      "dev": true,
-      "license": "MIT"
-    },
-    "node_modules/cliui/node_modules/is-fullwidth-code-point": {
-      "version": "3.0.0",
-      "resolved": "https://registry.npmjs.org/is-fullwidth-code-point/-/is-fullwidth-code-point-3.0.0.tgz",
-      "integrity": "sha512-zymm5+u+sCsSWyD9qNaejV3DFvhCKclKdizYaJUuHA83RLjb7nSuGnddCHGv0hk+KY7BMAlsWeK4Ueg6EV6XQg==",
-      "dev": true,
-      "license": "MIT",
-      "engines": {
-        "node": ">=8"
-      }
-    },
     "node_modules/cliui/node_modules/string-width": {
       "version": "4.2.3",
       "resolved": "https://registry.npmjs.org/string-width/-/string-width-4.2.3.tgz",
@@ -1934,24 +2398,6 @@
         "node": ">=8"
       }
     },
-    "node_modules/cliui/node_modules/wrap-ansi": {
-      "version": "7.0.0",
-      "resolved": "https://registry.npmjs.org/wrap-ansi/-/wrap-ansi-7.0.0.tgz",
-      "integrity": "sha512-YVGIj2kamLSTxw6NsZjoBxfSwsn0ycdesmc4p+Q21c5zPuZ1pl+NfxVdxPtdHvmNVOQ6XSYG4AUtyt/Fi7D16Q==",
-      "dev": true,
-      "license": "MIT",
-      "dependencies": {
-        "ansi-styles": "^4.0.0",
-        "string-width": "^4.1.0",
-        "strip-ansi": "^6.0.0"
-      },
-      "engines": {
-        "node": ">=10"
-      },
-      "funding": {
-        "url": "https://github.com/chalk/wrap-ansi?sponsor=1"
-      }
-    },
     "node_modules/clsx": {
       "version": "2.1.1",
       "resolved": "https://registry.npmjs.org/clsx/-/clsx-2.1.1.tgz",
@@ -2013,17 +2459,17 @@
         "url": "https://github.com/sponsors/wooorm"
       }
     },
-    "node_modules/concat-map": {
-      "version": "0.0.1",
-      "resolved": "https://registry.npmjs.org/concat-map/-/concat-map-0.0.1.tgz",
-      "integrity": "sha512-/Srv4dswyQNBfohGpz9o6Yb3Gz3SrUDqBH5rTuhGR7ahtlbYKnVxw2bCFMRljaA7EXHaXZ8wsHdodFvbkhKmqg==",
+    "node_modules/convert-source-map": {
+      "version": "2.0.0",
+      "resolved": "https://registry.npmjs.org/convert-source-map/-/convert-source-map-2.0.0.tgz",
+      "integrity": "sha512-Kvp459HrV2FEJ1CAsi1Ku+MY3kasH19TFykTz2xWmMeq6bk2NU3XXvfJ+Q61m0xktWwt+1HSYf3JZsTms3aRJg==",
       "dev": true,
       "license": "MIT"
     },
     "node_modules/cosmiconfig": {
-      "version": "9.0.0",
-      "resolved": "https://registry.npmjs.org/cosmiconfig/-/cosmiconfig-9.0.0.tgz",
-      "integrity": "sha512-itvL5h8RETACmOTFc4UfIyB2RfEHi71Ax6E/PivVxq9NseKbOWpeyHEOIbmAw1rs8Ak0VursQNww7lf7YtUwzg==",
+      "version": "9.0.2",
+      "resolved": "https://registry.npmjs.org/cosmiconfig/-/cosmiconfig-9.0.2.tgz",
+      "integrity": "sha512-gtTZxTDau1wL7Y7zifc2dd8jHSK/k6BTx/2Xp/BpdlAdnlYWFVt7qhJqgwi7637yRwRQ3qL4ZidbB4I8tA5VOg==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
@@ -2048,9 +2494,9 @@
       }
     },
     "node_modules/crelt": {
-      "version": "1.0.6",
-      "resolved": "https://registry.npmjs.org/crelt/-/crelt-1.0.6.tgz",
-      "integrity": "sha512-VQ2MBenTq1fWZUH9DJNGti7kKv6EeAuYr3cLwxUWhIu1baTaXh4Ib5W2CqHVqib4/MqbYGJqiL3Zb8GJZr3l4g==",
+      "version": "1.0.7",
+      "resolved": "https://registry.npmjs.org/crelt/-/crelt-1.0.7.tgz",
+      "integrity": "sha512-aK6BbWfhf4U/wCcLHKPJl/xa6VkVstRaPywWtMKGwuOLc/wZTyQYuoxgvZnNsBvv7Kg3YTBQYYBCggcviQczuA==",
       "license": "MIT"
     },
     "node_modules/cross-spawn": {
@@ -2098,24 +2544,24 @@
       }
     },
     "node_modules/css-functions-list": {
-      "version": "3.2.3",
-      "resolved": "https://registry.npmjs.org/css-functions-list/-/css-functions-list-3.2.3.tgz",
-      "integrity": "sha512-IQOkD3hbR5KrN93MtcYuad6YPuTSUhntLHDuLEbFWE+ff2/XSZNdZG+LcbbIW5AXKg/WFIfYItIzVoHngHXZzA==",
+      "version": "3.3.3",
+      "resolved": "https://registry.npmjs.org/css-functions-list/-/css-functions-list-3.3.3.tgz",
+      "integrity": "sha512-8HFEBPKhOpJPEPu70wJJetjKta86Gw9+CCyCnB3sui2qQfOvRyqBy4IKLKKAwdMpWb2lHXWk9Wb4Z6AmaUT1Pg==",
       "dev": true,
       "license": "MIT",
       "engines": {
-        "node": ">=12 || >=16"
+        "node": ">=12"
       }
     },
     "node_modules/css-tree": {
-      "version": "3.1.0",
-      "resolved": "https://registry.npmjs.org/css-tree/-/css-tree-3.1.0.tgz",
-      "integrity": "sha512-0eW44TGN5SQXU1mWSkKwFstI/22X2bG1nYzZTYMAWjylYURhse752YgbE4Cx46AC+bAvI+/dYTPRk1LqSUnu6w==",
+      "version": "3.2.1",
+      "resolved": "https://registry.npmjs.org/css-tree/-/css-tree-3.2.1.tgz",
+      "integrity": "sha512-X7sjQzceUhu1u7Y/ylrRZFU2FS6LRiFVp6rKLPg23y3x3c3DOKAwuXGDp+PAGjh6CSnCjYeAul8pcT8bAl+lSA==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
-        "mdn-data": "2.12.2",
-        "source-map-js": "^1.0.1"
+        "mdn-data": "2.27.1",
+        "source-map-js": "^1.2.1"
       },
       "engines": {
         "node": "^10 || ^12.20.0 || ^14.13.0 || >=15.0.0"
@@ -2135,15 +2581,15 @@
       }
     },
     "node_modules/csstype": {
-      "version": "3.1.3",
-      "resolved": "https://registry.npmjs.org/csstype/-/csstype-3.1.3.tgz",
-      "integrity": "sha512-M1uQkMl8rQK/szD0LNhtqxIPLpimGm8sOBwU7lLnCpSbTyY3yeU1Vc7l4KT5zT4s/yOxHH5O7tIuuLOCnLADRw==",
+      "version": "3.2.3",
+      "resolved": "https://registry.npmjs.org/csstype/-/csstype-3.2.3.tgz",
+      "integrity": "sha512-z1HGKcYy2xA8AGQfwrn0PAy+PB7X/GSj3UVJW9qKyn43xWa+gl5nXmU4qqLMRzWVLFC8KusUX8T/0kCiOYpAIQ==",
       "license": "MIT"
     },
     "node_modules/debug": {
-      "version": "4.4.1",
-      "resolved": "https://registry.npmjs.org/debug/-/debug-4.4.1.tgz",
-      "integrity": "sha512-KcKCqiftBJcZr++7ykoDIEwSa3XWowTfNPo92BYxjXiyYEVrUQh2aLyhxBCwww+heortUFxEJYcRzosstTEBYQ==",
+      "version": "4.4.3",
+      "resolved": "https://registry.npmjs.org/debug/-/debug-4.4.3.tgz",
+      "integrity": "sha512-RGwwWnwQvkVfavKVt22FGLw+xYSdzARwm0ru6DhTVA3umU5hZc28V3kO4stgYryrTlLpuvgI9GiijltAjNbcqA==",
       "license": "MIT",
       "dependencies": {
         "ms": "^2.1.3"
@@ -2158,9 +2604,9 @@
       }
     },
     "node_modules/decode-named-character-reference": {
-      "version": "1.2.0",
-      "resolved": "https://registry.npmjs.org/decode-named-character-reference/-/decode-named-character-reference-1.2.0.tgz",
-      "integrity": "sha512-c6fcElNV6ShtZXmsgNgFFV5tVX2PaV4g+MOAkb8eXHvn6sryJBrZa9r0zV6+dtTyoCKxtDy5tyQ5ZwQuidtd+Q==",
+      "version": "1.3.0",
+      "resolved": "https://registry.npmjs.org/decode-named-character-reference/-/decode-named-character-reference-1.3.0.tgz",
+      "integrity": "sha512-GtpQYB283KrPp6nRw50q3U9/VfOutZOe103qlN7BPP6Ad27xYnOIWv4lPzo8HCAL+mMZofJ9KEy30fq6MfaK6Q==",
       "license": "MIT",
       "dependencies": {
         "character-entities": "^2.0.0"
@@ -2199,29 +2645,6 @@
         "url": "https://github.com/sponsors/wooorm"
       }
     },
-    "node_modules/dir-glob": {
-      "version": "3.0.1",
-      "resolved": "https://registry.npmjs.org/dir-glob/-/dir-glob-3.0.1.tgz",
-      "integrity": "sha512-WkrWp9GR4KXfKGYzOLmTuGVi1UWFfws377n9cc55/tb6DuqyF6pcQ5AbiHEshaDpY9v6oaSr2XCDidGmMwdzIA==",
-      "dev": true,
-      "license": "MIT",
-      "dependencies": {
-        "path-type": "^4.0.0"
-      },
-      "engines": {
-        "node": ">=8"
-      }
-    },
-    "node_modules/dir-glob/node_modules/path-type": {
-      "version": "4.0.0",
-      "resolved": "https://registry.npmjs.org/path-type/-/path-type-4.0.0.tgz",
-      "integrity": "sha512-gDKb8aZMDeD/tZWs9P6+q0J9Mwkdl6xMV8TjnGP3qJVJ06bdMgkbBlLU8IdfOsIsFz2BW1rNVT3XuNEl8zPAvw==",
-      "dev": true,
-      "license": "MIT",
-      "engines": {
-        "node": ">=8"
-      }
-    },
     "node_modules/eastasianwidth": {
       "version": "0.2.0",
       "resolved": "https://registry.npmjs.org/eastasianwidth/-/eastasianwidth-0.2.0.tgz",
@@ -2229,6 +2652,20 @@
       "dev": true,
       "license": "MIT"
     },
+    "node_modules/electron-to-chromium": {
+      "version": "1.5.464",
+      "resolved": "https://registry.npmjs.org/electron-to-chromium/-/electron-to-chromium-1.5.464.tgz",
+      "integrity": "sha512-nRa65aFge40viPAKxMKULIY8tcXKwKboJWNn4RoKwyWXJ6SP7BnrDGKUTzc64818jIciH0PK7LSF8QzZes19GA==",
+      "dev": true,
+      "license": "ISC"
+    },
+    "node_modules/emoji-regex": {
+      "version": "8.0.0",
+      "resolved": "https://registry.npmjs.org/emoji-regex/-/emoji-regex-8.0.0.tgz",
+      "integrity": "sha512-MSjYzcWNOA0ewAHpz0MxpYFvwg6yjy1NG3xteoqz644VCo/RPgnr1/GGt+ic3iJTzQ8Eu3TdM14SawnVUmGE6A==",
+      "dev": true,
+      "license": "MIT"
+    },
     "node_modules/env-paths": {
       "version": "2.2.1",
       "resolved": "https://registry.npmjs.org/env-paths/-/env-paths-2.2.1.tgz",
@@ -2240,9 +2677,9 @@
       }
     },
     "node_modules/error-ex": {
-      "version": "1.3.2",
-      "resolved": "https://registry.npmjs.org/error-ex/-/error-ex-1.3.2.tgz",
-      "integrity": "sha512-7dFHNmqeFSEt2ZBsCriorKnn3Z2pj+fd9kmI6QoWw4//DL+icEBfc0U7qJCisqrTsKTjw4fNFy2pW9OqStD84g==",
+      "version": "1.3.4",
+      "resolved": "https://registry.npmjs.org/error-ex/-/error-ex-1.3.4.tgz",
+      "integrity": "sha512-sqQamAnR14VgCr1A618A3sGrygcpK+HEbenA/HiEAkkUwcZIIB/tgWqHFxWgOyDh4nB4JCRimh79dR5Ywc9MDQ==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
@@ -2250,9 +2687,9 @@
       }
     },
     "node_modules/esbuild": {
-      "version": "0.25.8",
-      "resolved": "https://registry.npmjs.org/esbuild/-/esbuild-0.25.8.tgz",
-      "integrity": "sha512-vVC0USHGtMi8+R4Kz8rt6JhEWLxsv9Rnu/lGYbPR8u47B+DCBksq9JarW0zOO7bs37hyOK1l2/oqtbciutL5+Q==",
+      "version": "0.28.2",
+      "resolved": "https://registry.npmjs.org/esbuild/-/esbuild-0.28.2.tgz",
+      "integrity": "sha512-HKVLS8dvII+xoKW9kmqxbRKrnWEXfJJr/FZhhJmiqIB0e053QNYFqOBouTMO/k5sID4MvCiUCvv8b9M4h32wIA==",
       "dev": true,
       "hasInstallScript": true,
       "license": "MIT",
@@ -2263,32 +2700,32 @@
         "node": ">=18"
       },
       "optionalDependencies": {
-        "@esbuild/aix-ppc64": "0.25.8",
-        "@esbuild/android-arm": "0.25.8",
-        "@esbuild/android-arm64": "0.25.8",
-        "@esbuild/android-x64": "0.25.8",
-        "@esbuild/darwin-arm64": "0.25.8",
-        "@esbuild/darwin-x64": "0.25.8",
-        "@esbuild/freebsd-arm64": "0.25.8",
-        "@esbuild/freebsd-x64": "0.25.8",
-        "@esbuild/linux-arm": "0.25.8",
-        "@esbuild/linux-arm64": "0.25.8",
-        "@esbuild/linux-ia32": "0.25.8",
-        "@esbuild/linux-loong64": "0.25.8",
-        "@esbuild/linux-mips64el": "0.25.8",
-        "@esbuild/linux-ppc64": "0.25.8",
-        "@esbuild/linux-riscv64": "0.25.8",
-        "@esbuild/linux-s390x": "0.25.8",
-        "@esbuild/linux-x64": "0.25.8",
-        "@esbuild/netbsd-arm64": "0.25.8",
-        "@esbuild/netbsd-x64": "0.25.8",
-        "@esbuild/openbsd-arm64": "0.25.8",
-        "@esbuild/openbsd-x64": "0.25.8",
-        "@esbuild/openharmony-arm64": "0.25.8",
-        "@esbuild/sunos-x64": "0.25.8",
-        "@esbuild/win32-arm64": "0.25.8",
-        "@esbuild/win32-ia32": "0.25.8",
-        "@esbuild/win32-x64": "0.25.8"
+        "@esbuild/aix-ppc64": "0.28.2",
+        "@esbuild/android-arm": "0.28.2",
+        "@esbuild/android-arm64": "0.28.2",
+        "@esbuild/android-x64": "0.28.2",
+        "@esbuild/darwin-arm64": "0.28.2",
+        "@esbuild/darwin-x64": "0.28.2",
+        "@esbuild/freebsd-arm64": "0.28.2",
+        "@esbuild/freebsd-x64": "0.28.2",
+        "@esbuild/linux-arm": "0.28.2",
+        "@esbuild/linux-arm64": "0.28.2",
+        "@esbuild/linux-ia32": "0.28.2",
+        "@esbuild/linux-loong64": "0.28.2",
+        "@esbuild/linux-mips64el": "0.28.2",
+        "@esbuild/linux-ppc64": "0.28.2",
+        "@esbuild/linux-riscv64": "0.28.2",
+        "@esbuild/linux-s390x": "0.28.2",
+        "@esbuild/linux-x64": "0.28.2",
+        "@esbuild/netbsd-arm64": "0.28.2",
+        "@esbuild/netbsd-x64": "0.28.2",
+        "@esbuild/openbsd-arm64": "0.28.2",
+        "@esbuild/openbsd-x64": "0.28.2",
+        "@esbuild/openharmony-arm64": "0.28.2",
+        "@esbuild/sunos-x64": "0.28.2",
+        "@esbuild/win32-arm64": "0.28.2",
+        "@esbuild/win32-ia32": "0.28.2",
+        "@esbuild/win32-x64": "0.28.2"
       }
     },
     "node_modules/escalade": {
@@ -2315,45 +2752,43 @@
       }
     },
     "node_modules/eslint": {
-      "version": "9.33.0",
-      "resolved": "https://registry.npmjs.org/eslint/-/eslint-9.33.0.tgz",
-      "integrity": "sha512-TS9bTNIryDzStCpJN93aC5VRSW3uTx9sClUn4B87pwiCaJh220otoI0X8mJKr+VcPtniMdN8GKjlwgWGUv5ZKA==",
+      "version": "10.13.0",
+      "resolved": "https://registry.npmjs.org/eslint/-/eslint-10.13.0.tgz",
+      "integrity": "sha512-UohDjD4zSzEwhf58M3O7vc/ycc8O1I7dNlhtD0WxE8Px8EHNJKtUwzBqrzeGhisRuv2ag4sAYswszCJRgKXYUg==",
       "dev": true,
       "license": "MIT",
+      "workspaces": [
+        "packages/*"
+      ],
       "dependencies": {
-        "@eslint-community/eslint-utils": "^4.2.0",
-        "@eslint-community/regexpp": "^4.12.1",
-        "@eslint/config-array": "^0.21.0",
-        "@eslint/config-helpers": "^0.3.1",
-        "@eslint/core": "^0.15.2",
-        "@eslint/eslintrc": "^3.3.1",
-        "@eslint/js": "9.33.0",
-        "@eslint/plugin-kit": "^0.3.5",
+        "@eslint-community/eslint-utils": "^4.8.0",
+        "@eslint-community/regexpp": "^4.12.2",
+        "@eslint/config-array": "^0.23.5",
+        "@eslint/config-helpers": "^0.7.0",
+        "@eslint/core": "^1.2.1",
+        "@eslint/plugin-kit": "^0.7.3",
         "@humanfs/node": "^0.16.6",
         "@humanwhocodes/module-importer": "^1.0.1",
         "@humanwhocodes/retry": "^0.4.2",
         "@types/estree": "^1.0.6",
-        "@types/json-schema": "^7.0.15",
-        "ajv": "^6.12.4",
-        "chalk": "^4.0.0",
+        "ajv": "^6.14.0",
         "cross-spawn": "^7.0.6",
         "debug": "^4.3.2",
         "escape-string-regexp": "^4.0.0",
-        "eslint-scope": "^8.4.0",
-        "eslint-visitor-keys": "^4.2.1",
-        "espree": "^10.4.0",
-        "esquery": "^1.5.0",
+        "eslint-scope": "^9.1.2",
+        "eslint-visitor-keys": "^5.0.1",
+        "espree": "^11.2.0",
+        "esquery": "^1.7.0",
         "esutils": "^2.0.2",
         "fast-deep-equal": "^3.1.3",
-        "file-entry-cache": "^8.0.0",
+        "file-entry-cache": "11.1.5 || >11.1.6 <12",
         "find-up": "^5.0.0",
         "glob-parent": "^6.0.2",
         "ignore": "^5.2.0",
         "imurmurhash": "^0.1.4",
         "is-glob": "^4.0.0",
         "json-stable-stringify-without-jsonify": "^1.0.1",
-        "lodash.merge": "^4.6.2",
-        "minimatch": "^3.1.2",
+        "minimatch": "^10.2.5",
         "natural-compare": "^1.4.0",
         "optionator": "^0.9.3"
       },
@@ -2361,7 +2796,7 @@
         "eslint": "bin/eslint.js"
       },
       "engines": {
-        "node": "^18.18.0 || ^20.9.0 || >=21.1.0"
+        "node": "^20.19.0 || ^22.13.0 || >=24"
       },
       "funding": {
         "url": "https://eslint.org/donate"
@@ -2376,103 +2811,79 @@
       }
     },
     "node_modules/eslint-plugin-react-hooks": {
-      "version": "5.2.0",
-      "resolved": "https://registry.npmjs.org/eslint-plugin-react-hooks/-/eslint-plugin-react-hooks-5.2.0.tgz",
-      "integrity": "sha512-+f15FfK64YQwZdJNELETdn5ibXEUQmW1DZL6KXhNnc2heoy/sg9VJJeT7n8TlMWouzWqSWavFkIhHyIbIAEapg==",
+      "version": "7.1.1",
+      "resolved": "https://registry.npmjs.org/eslint-plugin-react-hooks/-/eslint-plugin-react-hooks-7.1.1.tgz",
+      "integrity": "sha512-f2I7Gw6JbvCexzIInuSbZpfdQ44D7iqdWX01FKLvrPgqxoE7oMj8clOfto8U6vYiz4yd5oKu39rRSVOe1zRu0g==",
       "dev": true,
       "license": "MIT",
+      "dependencies": {
+        "@babel/core": "^7.24.4",
+        "@babel/parser": "^7.24.4",
+        "hermes-parser": "^0.25.1",
+        "zod": "^3.25.0 || ^4.0.0",
+        "zod-validation-error": "^3.5.0 || ^4.0.0"
+      },
       "engines": {
-        "node": ">=10"
+        "node": ">=18"
       },
       "peerDependencies": {
-        "eslint": "^3.0.0 || ^4.0.0 || ^5.0.0 || ^6.0.0 || ^7.0.0 || ^8.0.0-0 || ^9.0.0"
+        "eslint": "^3.0.0 || ^4.0.0 || ^5.0.0 || ^6.0.0 || ^7.0.0 || ^8.0.0-0 || ^9.0.0 || ^10.0.0"
       }
     },
     "node_modules/eslint-scope": {
-      "version": "8.4.0",
-      "resolved": "https://registry.npmjs.org/eslint-scope/-/eslint-scope-8.4.0.tgz",
-      "integrity": "sha512-sNXOfKCn74rt8RICKMvJS7XKV/Xk9kA7DyJr8mJik3S7Cwgy3qlkkmyS2uQB3jiJg6VNdZd/pDBJu0nvG2NlTg==",
+      "version": "9.1.2",
+      "resolved": "https://registry.npmjs.org/eslint-scope/-/eslint-scope-9.1.2.tgz",
+      "integrity": "sha512-xS90H51cKw0jltxmvmHy2Iai1LIqrfbw57b79w/J7MfvDfkIkFZ+kj6zC3BjtUwh150HsSSdxXZcsuv72miDFQ==",
       "dev": true,
       "license": "BSD-2-Clause",
       "dependencies": {
+        "@types/esrecurse": "^4.3.1",
+        "@types/estree": "^1.0.8",
         "esrecurse": "^4.3.0",
         "estraverse": "^5.2.0"
       },
       "engines": {
-        "node": "^18.18.0 || ^20.9.0 || >=21.1.0"
+        "node": "^20.19.0 || ^22.13.0 || >=24"
       },
       "funding": {
         "url": "https://opencollective.com/eslint"
       }
     },
     "node_modules/eslint-visitor-keys": {
-      "version": "4.2.1",
-      "resolved": "https://registry.npmjs.org/eslint-visitor-keys/-/eslint-visitor-keys-4.2.1.tgz",
-      "integrity": "sha512-Uhdk5sfqcee/9H/rCOJikYz67o0a2Tw2hGRPOG2Y1R2dg7brRe1uG0yaNQDHu+TO/uQPF/5eCapvYSmHUjt7JQ==",
+      "version": "5.0.1",
+      "resolved": "https://registry.npmjs.org/eslint-visitor-keys/-/eslint-visitor-keys-5.0.1.tgz",
+      "integrity": "sha512-tD40eHxA35h0PEIZNeIjkHoDR4YjjJp34biM0mDvplBe//mB+IHCqHDGV7pxF+7MklTvighcCPPZC7ynWyjdTA==",
       "dev": true,
       "license": "Apache-2.0",
       "engines": {
-        "node": "^18.18.0 || ^20.9.0 || >=21.1.0"
+        "node": "^20.19.0 || ^22.13.0 || >=24"
       },
       "funding": {
         "url": "https://opencollective.com/eslint"
       }
     },
-    "node_modules/eslint/node_modules/ansi-styles": {
-      "version": "4.3.0",
-      "resolved": "https://registry.npmjs.org/ansi-styles/-/ansi-styles-4.3.0.tgz",
-      "integrity": "sha512-zbB9rCJAT1rbjiVDb2hqKFHNYLxgtk8NURxZ3IZwD3F6NtxbXZQCnnSi1Lkx+IDohdPlFp222wVALIheZJQSEg==",
-      "dev": true,
-      "license": "MIT",
-      "dependencies": {
-        "color-convert": "^2.0.1"
-      },
-      "engines": {
-        "node": ">=8"
-      },
-      "funding": {
-        "url": "https://github.com/chalk/ansi-styles?sponsor=1"
-      }
-    },
-    "node_modules/eslint/node_modules/chalk": {
-      "version": "4.1.2",
-      "resolved": "https://registry.npmjs.org/chalk/-/chalk-4.1.2.tgz",
-      "integrity": "sha512-oKnbhFyRIXpUuez8iBMmyEa4nbj4IOQyuhc/wy9kY7/WVPcwIO9VA668Pu8RkO7+0G76SLROeyw9CpQ061i4mA==",
-      "dev": true,
-      "license": "MIT",
-      "dependencies": {
-        "ansi-styles": "^4.1.0",
-        "supports-color": "^7.1.0"
-      },
-      "engines": {
-        "node": ">=10"
-      },
-      "funding": {
-        "url": "https://github.com/chalk/chalk?sponsor=1"
-      }
-    },
     "node_modules/espree": {
-      "version": "10.4.0",
-      "resolved": "https://registry.npmjs.org/espree/-/espree-10.4.0.tgz",
-      "integrity": "sha512-j6PAQ2uUr79PZhBjP5C5fhl8e39FmRnOjsD5lGnWrFU8i2G776tBK7+nP8KuQUTTyAZUwfQqXAgrVH5MbH9CYQ==",
+      "version": "11.2.0",
+      "resolved": "https://registry.npmjs.org/espree/-/espree-11.2.0.tgz",
+      "integrity": "sha512-7p3DrVEIopW1B1avAGLuCSh1jubc01H2JHc8B4qqGblmg5gI9yumBgACjWo4JlIc04ufug4xJ3SQI8HkS/Rgzw==",
       "dev": true,
       "license": "BSD-2-Clause",
       "dependencies": {
-        "acorn": "^8.15.0",
+        "acorn": "^8.16.0",
         "acorn-jsx": "^5.3.2",
-        "eslint-visitor-keys": "^4.2.1"
+        "eslint-visitor-keys": "^5.0.1"
       },
       "engines": {
-        "node": "^18.18.0 || ^20.9.0 || >=21.1.0"
+        "node": "^20.19.0 || ^22.13.0 || >=24"
       },
       "funding": {
         "url": "https://opencollective.com/eslint"
       }
     },
     "node_modules/esquery": {
-      "version": "1.6.0",
-      "resolved": "https://registry.npmjs.org/esquery/-/esquery-1.6.0.tgz",
-      "integrity": "sha512-ca9pw9fomFcKPvFLXhBKUK90ZvGibiGOvRJNbjljY7s7uq/5YO4BOzcYtJqExdx99rF6aAcnRxHmcUHcz6sQsg==",
+      "version": "1.7.0",
+      "resolved": "https://registry.npmjs.org/esquery/-/esquery-1.7.0.tgz",
+      "integrity": "sha512-Ap6G0WQwcU/LHsvLwON1fAQX9Zp0A2Y6Y/cJBl9r/JbW90Zyg4/zbG6zzKa2OTALELarYHmKu0GhpM5EO+7T0g==",
       "dev": true,
       "license": "BSD-3-Clause",
       "dependencies": {
@@ -2583,9 +2994,9 @@
       "license": "MIT"
     },
     "node_modules/fast-uri": {
-      "version": "3.1.8",
-      "resolved": "https://registry.npmjs.org/fast-uri/-/fast-uri-3.1.8.tgz",
-      "integrity": "sha512-GZMtZUTNRpOVIECoXwLNZS5xUGE+mVNbTB8h/7Rwh2TFWcBQiPzTgyZi05BF9UMZKkLJv8XBRJTlU7zg8+ZfMg==",
+      "version": "3.1.9",
+      "resolved": "https://registry.npmjs.org/fast-uri/-/fast-uri-3.1.9.tgz",
+      "integrity": "sha512-hS/0Wt3Ya043oKXs0sIxrMpKQITXhMLIrVAKKXi4gozGC9v7YRWVnXIZzjNPTO02M9BdxLCWqK1yRXgNziVyJQ==",
       "dev": true,
       "funding": [
         {
@@ -2610,9 +3021,9 @@
       }
     },
     "node_modules/fastq": {
-      "version": "1.19.1",
-      "resolved": "https://registry.npmjs.org/fastq/-/fastq-1.19.1.tgz",
-      "integrity": "sha512-GwLTyxkCXjXbxqIhTsMI2Nui8huMPtnxg7krajPJAjnEG/iiOS7i+zCtWGZR9G0NBKbXKh6X9m9UIsYX/N6vvQ==",
+      "version": "1.20.3",
+      "resolved": "https://registry.npmjs.org/fastq/-/fastq-1.20.3.tgz",
+      "integrity": "sha512-XKv5nnLs6nLF71NgiKJLIZFLkPyIEuOselLG7ujZnGrRfQK8HpvY+WqKhAJUAdLomwVHErVS4LfxFlPq0/FTAw==",
       "dev": true,
       "license": "ISC",
       "dependencies": {
@@ -2620,16 +3031,13 @@
       }
     },
     "node_modules/file-entry-cache": {
-      "version": "8.0.0",
-      "resolved": "https://registry.npmjs.org/file-entry-cache/-/file-entry-cache-8.0.0.tgz",
-      "integrity": "sha512-XXTUwCvisa5oacNGRP9SfNtYBNAMi+RPwBFmblZEF7N7swHYQS6/Zfk7SRwx4D5j3CH211YNRco1DEMNVfZCnQ==",
+      "version": "11.1.5",
+      "resolved": "https://registry.npmjs.org/file-entry-cache/-/file-entry-cache-11.1.5.tgz",
+      "integrity": "sha512-+PFTHITI08JIGhnNpGNI8T8inUpgZfk3GNEqfT9R2zZV2iFXg3CvqzSl/uEhs7TSGujYRELEANyDvS8Fj7+S7Q==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
-        "flat-cache": "^4.0.0"
-      },
-      "engines": {
-        "node": ">=16.0.0"
+        "flat-cache": "^6.1.23"
       }
     },
     "node_modules/fill-range": {
@@ -2663,23 +3071,21 @@
       }
     },
     "node_modules/flat-cache": {
-      "version": "4.0.1",
-      "resolved": "https://registry.npmjs.org/flat-cache/-/flat-cache-4.0.1.tgz",
-      "integrity": "sha512-f7ccFPK3SXFHpx15UIGyRJ/FJQctuKZ0zVuN3frBo4HnK3cay9VEW0R6yPYFHC0AgqhukPzKjq22t5DmAyqGyw==",
+      "version": "6.1.23",
+      "resolved": "https://registry.npmjs.org/flat-cache/-/flat-cache-6.1.23.tgz",
+      "integrity": "sha512-f++BY9pTk+983xK1FLzlLpmM0i0z+jHmx3QESGkURMXujQZz1k5wzwX6hjnQ8goaD0B+sYnDK1yZ6MTyZfUaqA==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
-        "flatted": "^3.2.9",
-        "keyv": "^4.5.4"
-      },
-      "engines": {
-        "node": ">=16"
+        "cacheable": "^2.5.0",
+        "flatted": "^3.4.2",
+        "hookified": "^1.15.0"
       }
     },
     "node_modules/flatted": {
-      "version": "3.4.2",
-      "resolved": "https://registry.npmjs.org/flatted/-/flatted-3.4.2.tgz",
-      "integrity": "sha512-PjDse7RzhcPkIJwy5t7KPWQSZ9cAbzQXcafsetQoD7sOJRQlGikNbx7yZp2OotDnJyrDcbyRq3Ttb18iYOqkxA==",
+      "version": "3.4.4",
+      "resolved": "https://registry.npmjs.org/flatted/-/flatted-3.4.4.tgz",
+      "integrity": "sha512-5+ybhBZANEJxaH3X5evAFatUxLfEHSr7n6kYJ+1Qd0mUqr4eu9gIf6GDbWHf8RJijHrjjO8G+la14SlL2SeS1Q==",
       "dev": true,
       "license": "ISC"
     },
@@ -2701,9 +3107,9 @@
       }
     },
     "node_modules/fs-extra": {
-      "version": "11.3.1",
-      "resolved": "https://registry.npmjs.org/fs-extra/-/fs-extra-11.3.1.tgz",
-      "integrity": "sha512-eXvGGwZ5CL17ZSwHWd3bbgk7UUpF6IFHtP57NYYakPvHOs8GDgDe5KJI36jIJzDkJ6eJjuzRA8eBQb6SkKue0g==",
+      "version": "11.4.1",
+      "resolved": "https://registry.npmjs.org/fs-extra/-/fs-extra-11.4.1.tgz",
+      "integrity": "sha512-KYAb4c9BJQI6QqGKthV68OHe0badztdXJWKo0WtBA9IuCFPTKvE5ZdUBglP833aMjhaSPNO4A5j/EkzZtGlKjA==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
@@ -2730,6 +3136,16 @@
         "node": "^8.16.0 || ^10.6.0 || >=11.0.0"
       }
     },
+    "node_modules/gensync": {
+      "version": "1.0.0-beta.2",
+      "resolved": "https://registry.npmjs.org/gensync/-/gensync-1.0.0-beta.2.tgz",
+      "integrity": "sha512-3hN7NaskYvMDLQY55gnW3NQ+mesEAepTqlg+VEbj7zzqEMBVNhzcGYYeqFo/TlYz6eQiFcp1HcsCZO+nGgS8zg==",
+      "dev": true,
+      "license": "MIT",
+      "engines": {
+        "node": ">=6.9.0"
+      }
+    },
     "node_modules/get-caller-file": {
       "version": "2.0.5",
       "resolved": "https://registry.npmjs.org/get-caller-file/-/get-caller-file-2.0.5.tgz",
@@ -2740,25 +3156,32 @@
         "node": "6.* || 8.* || >= 10.*"
       }
     },
+    "node_modules/get-east-asian-width": {
+      "version": "1.7.0",
+      "resolved": "https://registry.npmjs.org/get-east-asian-width/-/get-east-asian-width-1.7.0.tgz",
+      "integrity": "sha512-XjH1AECxf0giL2V1aU8vKyRR2ppRUb5c0EvT7zuJTokQ74bNo52zOtghqdWIqrhUD79fo3x0WfKZdOqxF6LG1Q==",
+      "dev": true,
+      "license": "MIT",
+      "engines": {
+        "node": ">=18"
+      },
+      "funding": {
+        "url": "https://github.com/sponsors/sindresorhus"
+      }
+    },
     "node_modules/glob": {
-      "version": "11.1.0",
-      "resolved": "https://registry.npmjs.org/glob/-/glob-11.1.0.tgz",
-      "integrity": "sha512-vuNwKSaKiqm7g0THUBu2x7ckSs3XJLXE+2ssL7/MfTGPLLcrJQ/4Uq1CjPTtO5cCIiRxqvN6Twy1qOwhL0Xjcw==",
+      "version": "13.0.6",
+      "resolved": "https://registry.npmjs.org/glob/-/glob-13.0.6.tgz",
+      "integrity": "sha512-Wjlyrolmm8uDpm/ogGyXZXb1Z+Ca2B8NbJwqBVg0axK9GbBeoS7yGV6vjXnYdGm6X53iehEuxxbyiKp8QmN4Vw==",
       "dev": true,
       "license": "BlueOak-1.0.0",
       "dependencies": {
-        "foreground-child": "^3.3.1",
-        "jackspeak": "^4.1.1",
-        "minimatch": "^10.1.1",
-        "minipass": "^7.1.2",
-        "package-json-from-dist": "^1.0.0",
-        "path-scurry": "^2.0.0"
-      },
-      "bin": {
-        "glob": "dist/esm/bin.mjs"
+        "minimatch": "^10.2.2",
+        "minipass": "^7.1.3",
+        "path-scurry": "^2.0.2"
       },
       "engines": {
-        "node": "20 || >=22"
+        "node": "18 || 20 || >=22"
       },
       "funding": {
         "url": "https://github.com/sponsors/isaacs"
@@ -2777,56 +3200,17 @@
         "node": ">=10.13.0"
       }
     },
-    "node_modules/glob/node_modules/balanced-match": {
-      "version": "4.0.4",
-      "resolved": "https://registry.npmjs.org/balanced-match/-/balanced-match-4.0.4.tgz",
-      "integrity": "sha512-BLrgEcRTwX2o6gGxGOCNyMvGSp35YofuYzw9h1IMTRmKqttAZZVU67bdb9Pr2vUHA8+j3i2tJfjO6C6+4myGTA==",
+    "node_modules/global-modules": {
+      "version": "2.0.0",
+      "resolved": "https://registry.npmjs.org/global-modules/-/global-modules-2.0.0.tgz",
+      "integrity": "sha512-NGbfmJBp9x8IxyJSd1P+otYK8vonoJactOogrVfFRIAEY1ukil8RSKDz2Yo7wh1oihl51l/r6W4epkeKJHqL8A==",
       "dev": true,
       "license": "MIT",
+      "dependencies": {
+        "global-prefix": "^3.0.0"
+      },
       "engines": {
-        "node": "18 || 20 || >=22"
-      }
-    },
-    "node_modules/glob/node_modules/brace-expansion": {
-      "version": "5.0.12",
-      "resolved": "https://registry.npmjs.org/brace-expansion/-/brace-expansion-5.0.12.tgz",
-      "integrity": "sha512-YovQ3rzhaLMIrDjNDMkNS01tea93qhEhG5xy8f6+R0l+dw3Ki+5sCoIoI942iuLZTHWogWktgwVDhU09iNEimQ==",
-      "dev": true,
-      "license": "MIT",
-      "dependencies": {
-        "balanced-match": "^4.0.2"
-      },
-      "engines": {
-        "node": "20 || >=22"
-      }
-    },
-    "node_modules/glob/node_modules/minimatch": {
-      "version": "10.2.4",
-      "resolved": "https://registry.npmjs.org/minimatch/-/minimatch-10.2.4.tgz",
-      "integrity": "sha512-oRjTw/97aTBN0RHbYCdtF1MQfvusSIBQM0IZEgzl6426+8jSC0nF1a/GmnVLpfB9yyr6g6FTqWqiZVbxrtaCIg==",
-      "dev": true,
-      "license": "BlueOak-1.0.0",
-      "dependencies": {
-        "brace-expansion": "^5.0.2"
-      },
-      "engines": {
-        "node": "18 || 20 || >=22"
-      },
-      "funding": {
-        "url": "https://github.com/sponsors/isaacs"
-      }
-    },
-    "node_modules/global-modules": {
-      "version": "2.0.0",
-      "resolved": "https://registry.npmjs.org/global-modules/-/global-modules-2.0.0.tgz",
-      "integrity": "sha512-NGbfmJBp9x8IxyJSd1P+otYK8vonoJactOogrVfFRIAEY1ukil8RSKDz2Yo7wh1oihl51l/r6W4epkeKJHqL8A==",
-      "dev": true,
-      "license": "MIT",
-      "dependencies": {
-        "global-prefix": "^3.0.0"
-      },
-      "engines": {
-        "node": ">=6"
+        "node": ">=6"
       }
     },
     "node_modules/global-prefix": {
@@ -2858,9 +3242,9 @@
       }
     },
     "node_modules/globals": {
-      "version": "14.0.0",
-      "resolved": "https://registry.npmjs.org/globals/-/globals-14.0.0.tgz",
-      "integrity": "sha512-oahGvuMGQlPw/ivIYBjVSrWAfWLBeku5tpPE2fOPLi+WHffIWbuh2tCjhyQhTBPMf5E9jDEH4FOmTYgYwbKwtQ==",
+      "version": "17.13.0",
+      "resolved": "https://registry.npmjs.org/globals/-/globals-17.13.0.tgz",
+      "integrity": "sha512-RwMTC61u7hrG3ZJMkZQOK4JLJrKS8QtoTii63EmWvFXHqycY9FObBJQCbsLxSO8kjKhWE5kV1GC+Vb1g3RI0Zg==",
       "dev": true,
       "license": "MIT",
       "engines": {
@@ -2870,6 +3254,38 @@
         "url": "https://github.com/sponsors/sindresorhus"
       }
     },
+    "node_modules/globby": {
+      "version": "16.2.4",
+      "resolved": "https://registry.npmjs.org/globby/-/globby-16.2.4.tgz",
+      "integrity": "sha512-c8B/VNLmxRcmqqenRA9t+9IyOjf9+V6lTxPaUJLqOCONdQkWZ0ETYgX0qbtJqPsgCNusT9MZ5Jeidw8Eb9tn2g==",
+      "dev": true,
+      "license": "MIT",
+      "dependencies": {
+        "@sindresorhus/merge-streams": "^4.0.0",
+        "fast-glob": "^3.3.3",
+        "ignore": "^7.0.5",
+        "is-path-inside": "^4.0.0",
+        "micromatch": "^4.0.8",
+        "slash": "^5.1.0",
+        "unicorn-magic": "^0.4.0"
+      },
+      "engines": {
+        "node": ">=20"
+      },
+      "funding": {
+        "url": "https://github.com/sponsors/sindresorhus"
+      }
+    },
+    "node_modules/globby/node_modules/ignore": {
+      "version": "7.0.14",
+      "resolved": "https://registry.npmjs.org/ignore/-/ignore-7.0.14.tgz",
+      "integrity": "sha512-wCBiZ77u1tuWde6G/nfOtLaCTYcFDU5yCzARzK10y+rLhzYoL7a7yEemcopcc2Mzow7gZOIjnV2ym1NYmdiJOQ==",
+      "dev": true,
+      "license": "MIT",
+      "engines": {
+        "node": ">= 4"
+      }
+    },
     "node_modules/globjoin": {
       "version": "0.1.4",
       "resolved": "https://registry.npmjs.org/globjoin/-/globjoin-0.1.4.tgz",
@@ -2884,13 +3300,6 @@
       "dev": true,
       "license": "ISC"
     },
-    "node_modules/graphemer": {
-      "version": "1.4.0",
-      "resolved": "https://registry.npmjs.org/graphemer/-/graphemer-1.4.0.tgz",
-      "integrity": "sha512-EtKwoO6kxCL9WO5xipiHTZlSzBm7WLT627TqC/uVRd0HKmq8NXyebnNYxDoBi7wt8eTWrUrKXCOVaFq9x1kgag==",
-      "dev": true,
-      "license": "MIT"
-    },
     "node_modules/growly": {
       "version": "1.3.0",
       "resolved": "https://registry.npmjs.org/growly/-/growly-1.3.0.tgz",
@@ -2899,13 +3308,29 @@
       "license": "MIT"
     },
     "node_modules/has-flag": {
-      "version": "4.0.0",
-      "resolved": "https://registry.npmjs.org/has-flag/-/has-flag-4.0.0.tgz",
-      "integrity": "sha512-EykJT/Q1KjTWctppgIAgfSO0tKVuZUjhgMr17kqTumMl6Afv3EISleU7qZUzoXDFTAHTDC4NOoG/ZxU3EvlMPQ==",
+      "version": "5.0.1",
+      "resolved": "https://registry.npmjs.org/has-flag/-/has-flag-5.0.1.tgz",
+      "integrity": "sha512-CsNUt5x9LUdx6hnk/E2SZLsDyvfqANZSUq4+D3D8RzDJ2M+HDTIkF60ibS1vHaK55vzgiZw1bEPFG9yH7l33wA==",
       "dev": true,
       "license": "MIT",
       "engines": {
-        "node": ">=8"
+        "node": ">=12"
+      },
+      "funding": {
+        "url": "https://github.com/sponsors/sindresorhus"
+      }
+    },
+    "node_modules/hashery": {
+      "version": "1.5.1",
+      "resolved": "https://registry.npmjs.org/hashery/-/hashery-1.5.1.tgz",
+      "integrity": "sha512-iZyKG96/JwPz1N55vj2Ie2vXbhu440zfUfJvSwEqEbeLluk7NnapfGqa7LH0mOsnDxTF85Mx8/dyR6HfqcbmbQ==",
+      "dev": true,
+      "license": "MIT",
+      "dependencies": {
+        "hookified": "^1.15.0"
+      },
+      "engines": {
+        "node": ">=20"
       }
     },
     "node_modules/hast-util-to-jsx-runtime": {
@@ -2948,21 +3373,38 @@
         "url": "https://opencollective.com/unified"
       }
     },
+    "node_modules/hermes-estree": {
+      "version": "0.25.1",
+      "resolved": "https://registry.npmjs.org/hermes-estree/-/hermes-estree-0.25.1.tgz",
+      "integrity": "sha512-0wUoCcLp+5Ev5pDW2OriHC2MJCbwLwuRx+gAqMTOkGKJJiBCLjtrvy4PWUGn6MIVefecRpzoOZ/UV6iGdOr+Cw==",
+      "dev": true,
+      "license": "MIT"
+    },
+    "node_modules/hermes-parser": {
+      "version": "0.25.1",
+      "resolved": "https://registry.npmjs.org/hermes-parser/-/hermes-parser-0.25.1.tgz",
+      "integrity": "sha512-6pEjquH3rqaI6cYAXYPcz9MS4rY6R4ngRgrgfDshRptUZIc3lw0MCIJIGDj9++mfySOuPTHB4nrSW99BCvOPIA==",
+      "dev": true,
+      "license": "MIT",
+      "dependencies": {
+        "hermes-estree": "0.25.1"
+      }
+    },
     "node_modules/hookified": {
-      "version": "1.11.0",
-      "resolved": "https://registry.npmjs.org/hookified/-/hookified-1.11.0.tgz",
-      "integrity": "sha512-aDdIN3GyU5I6wextPplYdfmWCo+aLmjjVbntmX6HLD5RCi/xKsivYEBhnRD+d9224zFf008ZpLMPlWF0ZodYZw==",
+      "version": "1.15.1",
+      "resolved": "https://registry.npmjs.org/hookified/-/hookified-1.15.1.tgz",
+      "integrity": "sha512-MvG/clsADq1GPM2KGo2nyfaWVyn9naPiXrqIe4jYjXNZQt238kWyOGrsyc/DmRAQ+Re6yeo6yX/yoNCG5KAEVg==",
       "dev": true,
       "license": "MIT"
     },
     "node_modules/html-tags": {
-      "version": "3.3.1",
-      "resolved": "https://registry.npmjs.org/html-tags/-/html-tags-3.3.1.tgz",
-      "integrity": "sha512-ztqyC3kLto0e9WbNp0aeP+M3kTt+nbaIveGmUxAtZa+8iFgKLUOD4YKM5j+f3QD89bra7UeumolZHKuOXnTmeQ==",
+      "version": "5.1.0",
+      "resolved": "https://registry.npmjs.org/html-tags/-/html-tags-5.1.0.tgz",
+      "integrity": "sha512-n6l5uca7/y5joxZ3LUePhzmBFUJ+U2YWzhMa8XUTecSeSlQiZdF5XAd/Q3/WUl0VsXgUwWi8I7CNIwdI5WN1SQ==",
       "dev": true,
       "license": "MIT",
       "engines": {
-        "node": ">=8"
+        "node": ">=20.10"
       },
       "funding": {
         "url": "https://github.com/sponsors/sindresorhus"
@@ -3025,14 +3467,15 @@
         "url": "https://github.com/sponsors/sindresorhus"
       }
     },
-    "node_modules/import-fresh/node_modules/resolve-from": {
-      "version": "4.0.0",
-      "resolved": "https://registry.npmjs.org/resolve-from/-/resolve-from-4.0.0.tgz",
-      "integrity": "sha512-pb/MYmXstAkysRFx8piNI1tGFNQIFA3vkE3Gq4EuA1dF6gHp/+vgZqsCGJapvy8N3Q+4o7FwvquPJcnZ7RYy4g==",
+    "node_modules/import-meta-resolve": {
+      "version": "4.2.0",
+      "resolved": "https://registry.npmjs.org/import-meta-resolve/-/import-meta-resolve-4.2.0.tgz",
+      "integrity": "sha512-Iqv2fzaTQN28s/FwZAoFq0ZSs/7hMAHJVX+w8PZl3cY19Pxk6jFFalxQoIfW2826i/fDLXv8IiEZRIT0lDuWcg==",
       "dev": true,
       "license": "MIT",
-      "engines": {
-        "node": ">=4"
+      "funding": {
+        "type": "github",
+        "url": "https://github.com/sponsors/wooorm"
       }
     },
     "node_modules/imurmurhash": {
@@ -3053,9 +3496,9 @@
       "license": "ISC"
     },
     "node_modules/inline-style-parser": {
-      "version": "0.2.4",
-      "resolved": "https://registry.npmjs.org/inline-style-parser/-/inline-style-parser-0.2.4.tgz",
-      "integrity": "sha512-0aO8FkhNZlj/ZIbNi7Lxxr12obT7cL1moPfE4tg1LkX7LlLfC6DeX4l2ZEud1ukP9jNQyNnfzQVqwbwmAATY4Q==",
+      "version": "0.2.7",
+      "resolved": "https://registry.npmjs.org/inline-style-parser/-/inline-style-parser-0.2.7.tgz",
+      "integrity": "sha512-Nb2ctOyNR8DqQoR0OwRG95uNWIC0C1lCgf5Naz5H6Ji72KZ8OcFZLz2P5sNgwlyoJ8Yif11oMuYs5pBQa86csA==",
       "license": "MIT"
     },
     "node_modules/is-alphabetical": {
@@ -3113,16 +3556,16 @@
       }
     },
     "node_modules/is-docker": {
-      "version": "2.2.1",
-      "resolved": "https://registry.npmjs.org/is-docker/-/is-docker-2.2.1.tgz",
-      "integrity": "sha512-F+i2BKsFrH66iaUFc0woD8sLy8getkwTwtOBjvs56Cx4CgJDeKQeqfz8wAYiSb8JOprWhHH5p77PbmYCvvUuXQ==",
+      "version": "3.0.0",
+      "resolved": "https://registry.npmjs.org/is-docker/-/is-docker-3.0.0.tgz",
+      "integrity": "sha512-eljcgEDlEns/7AXFosB5K/2nCM4P7FQPkGc/DWLy5rmFEWvZayGrik1d9/QIY5nJ4f9YsVvBkA6kJpHn9rISdQ==",
       "dev": true,
       "license": "MIT",
       "bin": {
         "is-docker": "cli.js"
       },
       "engines": {
-        "node": ">=8"
+        "node": "^12.20.0 || ^14.13.1 || >=16.0.0"
       },
       "funding": {
         "url": "https://github.com/sponsors/sindresorhus"
@@ -3138,6 +3581,16 @@
         "node": ">=0.10.0"
       }
     },
+    "node_modules/is-fullwidth-code-point": {
+      "version": "3.0.0",
+      "resolved": "https://registry.npmjs.org/is-fullwidth-code-point/-/is-fullwidth-code-point-3.0.0.tgz",
+      "integrity": "sha512-zymm5+u+sCsSWyD9qNaejV3DFvhCKclKdizYaJUuHA83RLjb7nSuGnddCHGv0hk+KY7BMAlsWeK4Ueg6EV6XQg==",
+      "dev": true,
+      "license": "MIT",
+      "engines": {
+        "node": ">=8"
+      }
+    },
     "node_modules/is-glob": {
       "version": "4.0.3",
       "resolved": "https://registry.npmjs.org/is-glob/-/is-glob-4.0.3.tgz",
@@ -3161,6 +3614,25 @@
         "url": "https://github.com/sponsors/wooorm"
       }
     },
+    "node_modules/is-inside-container": {
+      "version": "1.0.0",
+      "resolved": "https://registry.npmjs.org/is-inside-container/-/is-inside-container-1.0.0.tgz",
+      "integrity": "sha512-KIYLCCJghfHZxqjYBE7rEy0OBuTd5xCHS7tHVgvCLkx7StIoaxwNW3hCALgEUjFfeRk+MG/Qxmp/vtETEF3tRA==",
+      "dev": true,
+      "license": "MIT",
+      "dependencies": {
+        "is-docker": "^3.0.0"
+      },
+      "bin": {
+        "is-inside-container": "cli.js"
+      },
+      "engines": {
+        "node": ">=14.16"
+      },
+      "funding": {
+        "url": "https://github.com/sponsors/sindresorhus"
+      }
+    },
     "node_modules/is-number": {
       "version": "7.0.0",
       "resolved": "https://registry.npmjs.org/is-number/-/is-number-7.0.0.tgz",
@@ -3171,10 +3643,11 @@
         "node": ">=0.12.0"
       }
     },
-    "node_modules/is-plain-obj": {
-      "version": "4.1.0",
-      "resolved": "https://registry.npmjs.org/is-plain-obj/-/is-plain-obj-4.1.0.tgz",
-      "integrity": "sha512-+Pgi+vMuUNkJyExiMBt5IlFoMyKnr5zhJ4Uspz58WOhBF5QoIZkFyNHIbBAtHwzVAgk5RtndVNsDRN61/mmDqg==",
+    "node_modules/is-path-inside": {
+      "version": "4.0.0",
+      "resolved": "https://registry.npmjs.org/is-path-inside/-/is-path-inside-4.0.0.tgz",
+      "integrity": "sha512-lJJV/5dYS+RcL8uQdBDW9c9uWFLLBNRyFhnAKXw5tVqLlKZ4RMGZKv+YQ/IA3OhD+RpbJa1LLFM1FQPGyIXvOA==",
+      "dev": true,
       "license": "MIT",
       "engines": {
         "node": ">=12"
@@ -3183,14 +3656,16 @@
         "url": "https://github.com/sponsors/sindresorhus"
       }
     },
-    "node_modules/is-plain-object": {
-      "version": "5.0.0",
-      "resolved": "https://registry.npmjs.org/is-plain-object/-/is-plain-object-5.0.0.tgz",
-      "integrity": "sha512-VRSzKkbMm5jMDoKLbltAkFQ5Qr7VDiTFGXxYFXXowVj387GeGNOCsOH6Msy00SGZ3Fp84b1Naa1psqgcCIEP5Q==",
-      "dev": true,
+    "node_modules/is-plain-obj": {
+      "version": "4.1.0",
+      "resolved": "https://registry.npmjs.org/is-plain-obj/-/is-plain-obj-4.1.0.tgz",
+      "integrity": "sha512-+Pgi+vMuUNkJyExiMBt5IlFoMyKnr5zhJ4Uspz58WOhBF5QoIZkFyNHIbBAtHwzVAgk5RtndVNsDRN61/mmDqg==",
       "license": "MIT",
       "engines": {
-        "node": ">=0.10.0"
+        "node": ">=12"
+      },
+      "funding": {
+        "url": "https://github.com/sponsors/sindresorhus"
       }
     },
     "node_modules/is-stream": {
@@ -3214,16 +3689,19 @@
       "license": "MIT"
     },
     "node_modules/is-wsl": {
-      "version": "2.2.0",
-      "resolved": "https://registry.npmjs.org/is-wsl/-/is-wsl-2.2.0.tgz",
-      "integrity": "sha512-fKzAra0rGJUUBwGBgNkHZuToZcn+TtXHpeCgmkMJMMYx1sQDYaCSyjJBSCa2nH1DGm7s3n1oBnohoVTBaN7Lww==",
+      "version": "3.1.1",
+      "resolved": "https://registry.npmjs.org/is-wsl/-/is-wsl-3.1.1.tgz",
+      "integrity": "sha512-e6rvdUCiQCAuumZslxRJWR/Doq4VpPR82kqclvcS0efgt430SlGIk05vdCN58+VrzgtIcfNODjozVielycD4Sw==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
-        "is-docker": "^2.0.0"
+        "is-inside-container": "^1.0.0"
       },
       "engines": {
-        "node": ">=8"
+        "node": ">=16"
+      },
+      "funding": {
+        "url": "https://github.com/sponsors/sindresorhus"
       }
     },
     "node_modules/isexe": {
@@ -3234,66 +3712,46 @@
       "license": "ISC"
     },
     "node_modules/jackspeak": {
-      "version": "4.1.1",
-      "resolved": "https://registry.npmjs.org/jackspeak/-/jackspeak-4.1.1.tgz",
-      "integrity": "sha512-zptv57P3GpL+O0I7VdMJNBZCu+BPHVQUk55Ft8/QCJjTVxrnJHuVuX/0Bl2A6/+2oyR/ZMEuFKwmzqqZ/U5nPQ==",
+      "version": "3.4.3",
+      "resolved": "https://registry.npmjs.org/jackspeak/-/jackspeak-3.4.3.tgz",
+      "integrity": "sha512-OGlZQpz2yfahA/Rd1Y8Cd9SIEsqvXkLVoSw/cgwhnhFMDbsQFeZYoJJ7bIZBS9BcamUW96asq/npPWugM+RQBw==",
       "dev": true,
       "license": "BlueOak-1.0.0",
       "dependencies": {
         "@isaacs/cliui": "^8.0.2"
       },
-      "engines": {
-        "node": "20 || >=22"
-      },
       "funding": {
         "url": "https://github.com/sponsors/isaacs"
+      },
+      "optionalDependencies": {
+        "@pkgjs/parseargs": "^0.11.0"
       }
     },
     "node_modules/jarmuz": {
-      "version": "0.10.0",
-      "resolved": "https://registry.npmjs.org/jarmuz/-/jarmuz-0.10.0.tgz",
-      "integrity": "sha512-vWaE3NMjDTn5NF6guzaUlHvhwfMp1l81WaUvYiCN0ye0bzVjFrObJKGer5f6cYWWUq9bE/9AVM/M3+oI3d9VTw==",
-      "dev": true,
-      "license": "MIT",
-      "dependencies": {
-        "chokidar": "^4.0.3",
-        "cross-spawn": "^7.0.6",
-        "minimatch": "^10.0.3",
-        "nanoid": "^5.1.5",
-        "string-argv": "^0.3.2"
-      }
-    },
-    "node_modules/jarmuz/node_modules/balanced-match": {
-      "version": "4.0.4",
-      "resolved": "https://registry.npmjs.org/balanced-match/-/balanced-match-4.0.4.tgz",
-      "integrity": "sha512-BLrgEcRTwX2o6gGxGOCNyMvGSp35YofuYzw9h1IMTRmKqttAZZVU67bdb9Pr2vUHA8+j3i2tJfjO6C6+4myGTA==",
-      "dev": true,
-      "license": "MIT",
-      "engines": {
-        "node": "18 || 20 || >=22"
-      }
-    },
-    "node_modules/jarmuz/node_modules/brace-expansion": {
-      "version": "5.0.12",
-      "resolved": "https://registry.npmjs.org/brace-expansion/-/brace-expansion-5.0.12.tgz",
-      "integrity": "sha512-YovQ3rzhaLMIrDjNDMkNS01tea93qhEhG5xy8f6+R0l+dw3Ki+5sCoIoI942iuLZTHWogWktgwVDhU09iNEimQ==",
+      "version": "0.11.0",
+      "resolved": "https://registry.npmjs.org/jarmuz/-/jarmuz-0.11.0.tgz",
+      "integrity": "sha512-qBGM3pTj/gxyK8OsOx8gbSR3USOTqthg69t/qxJxCFYp8Bsa7y0KAuZy4WnagYBi4pcx+FT+APBm4oSpv29oPA==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
-        "balanced-match": "^4.0.2"
+        "chokidar": "4.0.3",
+        "cross-spawn": "7.0.6",
+        "minimatch": "10.2.5",
+        "nanoid": "5.1.5",
+        "string-argv": "0.3.2"
       },
       "engines": {
-        "node": "20 || >=22"
+        "node": ">=24.0.0"
       }
     },
     "node_modules/jarmuz/node_modules/minimatch": {
-      "version": "10.2.4",
-      "resolved": "https://registry.npmjs.org/minimatch/-/minimatch-10.2.4.tgz",
-      "integrity": "sha512-oRjTw/97aTBN0RHbYCdtF1MQfvusSIBQM0IZEgzl6426+8jSC0nF1a/GmnVLpfB9yyr6g6FTqWqiZVbxrtaCIg==",
+      "version": "10.2.5",
+      "resolved": "https://registry.npmjs.org/minimatch/-/minimatch-10.2.5.tgz",
+      "integrity": "sha512-MULkVLfKGYDFYejP07QOurDLLQpcjk7Fw+7jXS2R2czRQzR56yHRveU5NDJEOviH+hETZKSkIk5c+T23GjFUMg==",
       "dev": true,
       "license": "BlueOak-1.0.0",
       "dependencies": {
-        "brace-expansion": "^5.0.2"
+        "brace-expansion": "^5.0.5"
       },
       "engines": {
         "node": "18 || 20 || >=22"
@@ -3332,12 +3790,18 @@
         "js-yaml": "bin/js-yaml.js"
       }
     },
-    "node_modules/json-buffer": {
-      "version": "3.0.1",
-      "resolved": "https://registry.npmjs.org/json-buffer/-/json-buffer-3.0.1.tgz",
-      "integrity": "sha512-4bV5BfR2mqfQTJm+V5tPPdf+ZpuhiIvTuAB5g8kcrXOZpTT/QwwVRWBywX1ozr6lEuPdbHxwaJlm9G6mI2sfSQ==",
+    "node_modules/jsesc": {
+      "version": "3.1.0",
+      "resolved": "https://registry.npmjs.org/jsesc/-/jsesc-3.1.0.tgz",
+      "integrity": "sha512-/sM3dO2FOzXjKQhJuo0Q173wf2KOo8t4I8vHy6lF9poUp7bKT0/NHE8fPX23PwfhnykfqnC2xRxOnVw5XuGIaA==",
       "dev": true,
-      "license": "MIT"
+      "license": "MIT",
+      "bin": {
+        "jsesc": "bin/jsesc"
+      },
+      "engines": {
+        "node": ">=6"
+      }
     },
     "node_modules/json-parse-even-better-errors": {
       "version": "2.3.1",
@@ -3360,10 +3824,23 @@
       "dev": true,
       "license": "MIT"
     },
+    "node_modules/json5": {
+      "version": "2.2.3",
+      "resolved": "https://registry.npmjs.org/json5/-/json5-2.2.3.tgz",
+      "integrity": "sha512-XmOWe7eyHYH14cLdVPoyg+GOH3rYX++KpzrylJwSW98t3Nk+U8XOl8FWKOgwtzdb8lXGf6zYwDUzeHMWfxasyg==",
+      "dev": true,
+      "license": "MIT",
+      "bin": {
+        "json5": "lib/cli.js"
+      },
+      "engines": {
+        "node": ">=6"
+      }
+    },
     "node_modules/jsonfile": {
-      "version": "6.1.0",
-      "resolved": "https://registry.npmjs.org/jsonfile/-/jsonfile-6.1.0.tgz",
-      "integrity": "sha512-5dgndWOriYSm5cnYaJNhalLNDKOqFwyDB/rr1E9ZsGciGvKPs8R2xYGCacuf3z6K1YKDz182fd+fY3cn3pMqXQ==",
+      "version": "6.2.1",
+      "resolved": "https://registry.npmjs.org/jsonfile/-/jsonfile-6.2.1.tgz",
+      "integrity": "sha512-zwOTdL3rFQ/lRdBnntKVOX6k5cKJwEc1HdilT71BWEu7J41gXIB2MRp+vxduPSwZJPWBxEzv4yH1wYLJGUHX4Q==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
@@ -3374,13 +3851,13 @@
       }
     },
     "node_modules/keyv": {
-      "version": "4.5.4",
-      "resolved": "https://registry.npmjs.org/keyv/-/keyv-4.5.4.tgz",
-      "integrity": "sha512-oxVHkHR/EJf2CNXnWxRLW6mg7JyCCUcG0DtEGmL2ctUo1PNTin1PUil+r/+4r5MpVgC/fn1kjsx7mjSujKqIpw==",
+      "version": "5.6.0",
+      "resolved": "https://registry.npmjs.org/keyv/-/keyv-5.6.0.tgz",
+      "integrity": "sha512-CYDD3SOtsHtyXeEORYRx2qBtpDJFjRTGXUtmNEMGyzYOKj1TE3tycdlho7kA1Ufx9OYWZzg52QFBGALTirzDSw==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
-        "json-buffer": "3.0.1"
+        "@keyv/serialize": "^1.1.1"
       }
     },
     "node_modules/kind-of": {
@@ -3393,13 +3870,6 @@
         "node": ">=0.10.0"
       }
     },
-    "node_modules/known-css-properties": {
-      "version": "0.37.0",
-      "resolved": "https://registry.npmjs.org/known-css-properties/-/known-css-properties-0.37.0.tgz",
-      "integrity": "sha512-JCDrsP4Z1Sb9JwG0aJ8Eo2r7k4Ou5MwmThS/6lcIe1ICyb7UBJKGRIUUdqc2ASdE/42lgz6zFUnzAIhtXnBVrQ==",
-      "dev": true,
-      "license": "MIT"
-    },
     "node_modules/levn": {
       "version": "0.4.1",
       "resolved": "https://registry.npmjs.org/levn/-/levn-0.4.1.tgz",
@@ -3444,13 +3914,6 @@
       "dev": true,
       "license": "MIT"
     },
-    "node_modules/lodash.merge": {
-      "version": "4.6.2",
-      "resolved": "https://registry.npmjs.org/lodash.merge/-/lodash.merge-4.6.2.tgz",
-      "integrity": "sha512-0KpjqXRVvrYyCsX1swR/XTK0va6VQkQM6MNo7PqW77ByjAhoARA8EfrP1N4+KlKj8YS0ZUCtRT/YUuhyYDujIQ==",
-      "dev": true,
-      "license": "MIT"
-    },
     "node_modules/lodash.truncate": {
       "version": "4.4.2",
       "resolved": "https://registry.npmjs.org/lodash.truncate/-/lodash.truncate-4.4.2.tgz",
@@ -3469,19 +3932,19 @@
       }
     },
     "node_modules/lru-cache": {
-      "version": "11.1.0",
-      "resolved": "https://registry.npmjs.org/lru-cache/-/lru-cache-11.1.0.tgz",
-      "integrity": "sha512-QIXZUBJUx+2zHUdQujWejBkcD9+cs94tLn0+YL8UrCh+D5sCXZ4c7LaEH48pNwRY3MLDgqUFyhlCyjJPf1WP0A==",
+      "version": "5.1.1",
+      "resolved": "https://registry.npmjs.org/lru-cache/-/lru-cache-5.1.1.tgz",
+      "integrity": "sha512-KpNARQA3Iwv+jTA0utUVVbrh+Jlrr1Fv0e56GGzAFOXN7dk/FviaDW8LHmK52DlcH4WP2n6gI8vN1aesBFgo9w==",
       "dev": true,
       "license": "ISC",
-      "engines": {
-        "node": "20 || >=22"
+      "dependencies": {
+        "yallist": "^3.0.2"
       }
     },
     "node_modules/mathml-tag-names": {
-      "version": "2.1.3",
-      "resolved": "https://registry.npmjs.org/mathml-tag-names/-/mathml-tag-names-2.1.3.tgz",
-      "integrity": "sha512-APMBEanjybaPzUrfqU0IMU5I0AswKMH7k8OTLs0vvV4KZpExkTkY87nR/zpbuTPj+gARop7aGUbl11pnDfW6xg==",
+      "version": "4.0.0",
+      "resolved": "https://registry.npmjs.org/mathml-tag-names/-/mathml-tag-names-4.0.0.tgz",
+      "integrity": "sha512-aa6AU2Pcx0VP/XWnh8IGL0SYSgQHDT6Ucror2j2mXeFAlN3ahaNs8EZtG1YiticMkSLj3Gt6VPFfZogt7G5iFQ==",
       "dev": true,
       "license": "MIT",
       "funding": {
@@ -3490,9 +3953,9 @@
       }
     },
     "node_modules/mdast-util-from-markdown": {
-      "version": "2.0.2",
-      "resolved": "https://registry.npmjs.org/mdast-util-from-markdown/-/mdast-util-from-markdown-2.0.2.tgz",
-      "integrity": "sha512-uZhTV/8NBuw0WHkPTrCqDOl0zVe1BIng5ZtHoDk49ME1qqcjYmmLmOf0gELgcRMxN4w2iuIeVso5/6QymSrgmA==",
+      "version": "2.1.0",
+      "resolved": "https://registry.npmjs.org/mdast-util-from-markdown/-/mdast-util-from-markdown-2.1.0.tgz",
+      "integrity": "sha512-KRLgmPaP2H1ul2bpoJkOS+9pAcDAAj9Nu1mqLnh8NLeJG5tlvj42TH80PIGBpROWlRXLk0lDrWNuZCFy/tsB4A==",
       "license": "MIT",
       "dependencies": {
         "@types/mdast": "^4.0.0",
@@ -3609,9 +4072,9 @@
       }
     },
     "node_modules/mdast-util-to-markdown": {
-      "version": "2.1.2",
-      "resolved": "https://registry.npmjs.org/mdast-util-to-markdown/-/mdast-util-to-markdown-2.1.2.tgz",
-      "integrity": "sha512-xj68wMTvGXVOKonmog6LwyJKrYXZPvlwabaryTjLh9LuvovB/KAH+kvi8Gjj+7rJjsFi23nkUxRQv1KqSroMqA==",
+      "version": "2.2.0",
+      "resolved": "https://registry.npmjs.org/mdast-util-to-markdown/-/mdast-util-to-markdown-2.2.0.tgz",
+      "integrity": "sha512-2Jn/ADrHFFyyniBMi0bncTbJqU78K3/TEYvzvwrWWzD9MYPUa/Gm4r3nOeKFQg/PDM5BsbDYdc8ackW4NTY6NA==",
       "license": "MIT",
       "dependencies": {
         "@types/mdast": "^4.0.0",
@@ -3619,8 +4082,10 @@
         "longest-streak": "^3.0.0",
         "mdast-util-phrasing": "^4.0.0",
         "mdast-util-to-string": "^4.0.0",
+        "micromark-util-character": "^2.0.0",
         "micromark-util-classify-character": "^2.0.0",
         "micromark-util-decode-string": "^2.0.0",
+        "micromark-util-html-tag-name": "^2.0.0",
         "unist-util-visit": "^5.0.0",
         "zwitch": "^2.0.0"
       },
@@ -3643,20 +4108,20 @@
       }
     },
     "node_modules/mdn-data": {
-      "version": "2.12.2",
-      "resolved": "https://registry.npmjs.org/mdn-data/-/mdn-data-2.12.2.tgz",
-      "integrity": "sha512-IEn+pegP1aManZuckezWCO+XZQDplx1366JoVhTpMpBB1sPey/SbveZQUosKiKiGYjg1wH4pMlNgXbCiYgihQA==",
+      "version": "2.27.1",
+      "resolved": "https://registry.npmjs.org/mdn-data/-/mdn-data-2.27.1.tgz",
+      "integrity": "sha512-9Yubnt3e8A0OKwxYSXyhLymGW4sCufcLG6VdiDdUGVkPhpqLxlvP5vl1983gQjJl3tqbrM731mjaZaP68AgosQ==",
       "dev": true,
       "license": "CC0-1.0"
     },
     "node_modules/meow": {
-      "version": "13.2.0",
-      "resolved": "https://registry.npmjs.org/meow/-/meow-13.2.0.tgz",
-      "integrity": "sha512-pxQJQzB6djGPXh08dacEloMFopsOqGVRKFPYvPOt9XDZ1HasbgDZA74CJGreSU4G3Ak7EFJGoiH2auq+yXISgA==",
+      "version": "14.1.0",
+      "resolved": "https://registry.npmjs.org/meow/-/meow-14.1.0.tgz",
+      "integrity": "sha512-EDYo6VlmtnumlcBCbh1gLJ//9jvM/ndXHfVXIFrZVr6fGcwTUyCTFNTLCKuY3ffbK8L/+3Mzqnd58RojiZqHVw==",
       "dev": true,
       "license": "MIT",
       "engines": {
-        "node": ">=18"
+        "node": ">=20"
       },
       "funding": {
         "url": "https://github.com/sponsors/sindresorhus"
@@ -3673,9 +4138,9 @@
       }
     },
     "node_modules/micromark": {
-      "version": "4.0.2",
-      "resolved": "https://registry.npmjs.org/micromark/-/micromark-4.0.2.tgz",
-      "integrity": "sha512-zpe98Q6kvavpCr1NPVSCMebCKfD7CA2NqZ+rykeNhONIJBpc1tFKt9hucLGwha3jNTNI8lHpctWJWoimVF4PfA==",
+      "version": "4.0.3",
+      "resolved": "https://registry.npmjs.org/micromark/-/micromark-4.0.3.tgz",
+      "integrity": "sha512-oGYfQzHSG5dOMovQcJ3fyTmZlWAWpi0XA0sJwJs+i6OT88o1+Jtw/8z0CmdowSLxGhhFK89rQ/oXph/wN02PNw==",
       "funding": [
         {
           "type": "GitHub Sponsors",
@@ -3698,6 +4163,7 @@
         "micromark-util-chunked": "^2.0.0",
         "micromark-util-combine-extensions": "^2.0.0",
         "micromark-util-decode-numeric-character-reference": "^2.0.0",
+        "micromark-util-edit-map": "^1.0.0",
         "micromark-util-encode": "^2.0.0",
         "micromark-util-normalize-identifier": "^2.0.0",
         "micromark-util-resolve-all": "^2.0.0",
@@ -3708,9 +4174,9 @@
       }
     },
     "node_modules/micromark-core-commonmark": {
-      "version": "2.0.3",
-      "resolved": "https://registry.npmjs.org/micromark-core-commonmark/-/micromark-core-commonmark-2.0.3.tgz",
-      "integrity": "sha512-RDBrHEMSxVFLg6xvnXmb1Ayr2WzLAWjeSATAoxwKYJV94TeNavgoIdA0a9ytzDSVzBy2YKFK+emCPOEibLeCrg==",
+      "version": "2.0.4",
+      "resolved": "https://registry.npmjs.org/micromark-core-commonmark/-/micromark-core-commonmark-2.0.4.tgz",
+      "integrity": "sha512-wxEeE8v8XVvOrxn1TZj74qYhAtTQsSqufHVS3uNVyT1MupXxhyaHk8/9xDNPh9BmL77QNDgWTkZ3RsYHB2ry7w==",
       "funding": [
         {
           "type": "GitHub Sponsors",
@@ -3727,18 +4193,19 @@
         "devlop": "^1.0.0",
         "micromark-factory-destination": "^2.0.0",
         "micromark-factory-label": "^2.0.0",
-        "micromark-factory-space": "^2.0.0",
+        "micromark-factory-space": "^2.1.0",
         "micromark-factory-title": "^2.0.0",
         "micromark-factory-whitespace": "^2.0.0",
         "micromark-util-character": "^2.0.0",
         "micromark-util-chunked": "^2.0.0",
         "micromark-util-classify-character": "^2.0.0",
+        "micromark-util-edit-map": "^1.0.0",
         "micromark-util-html-tag-name": "^2.0.0",
         "micromark-util-normalize-identifier": "^2.0.0",
         "micromark-util-resolve-all": "^2.0.0",
         "micromark-util-subtokenize": "^2.0.0",
         "micromark-util-symbol": "^2.0.0",
-        "micromark-util-types": "^2.0.0"
+        "micromark-util-types": "^2.0.3"
       }
     },
     "node_modules/micromark-factory-destination": {
@@ -3785,9 +4252,9 @@
       }
     },
     "node_modules/micromark-factory-space": {
-      "version": "2.0.1",
-      "resolved": "https://registry.npmjs.org/micromark-factory-space/-/micromark-factory-space-2.0.1.tgz",
-      "integrity": "sha512-zRkxjtBxxLd2Sc0d+fbnEunsTj46SWXgXciZmHq0kDYGnck/ZSGj9/wULTV95uoeYiK5hRXP2mJ98Uo4cq/LQg==",
+      "version": "2.1.0",
+      "resolved": "https://registry.npmjs.org/micromark-factory-space/-/micromark-factory-space-2.1.0.tgz",
+      "integrity": "sha512-fS8hnLIjnjvdQIj39Geug8wWsR0HrYZ43KShKxNfwT7t2LOHo/LbWZEzEaSOjwtjdCsjoE6syHhonEzW0zv0+Q==",
       "funding": [
         {
           "type": "GitHub Sponsors",
@@ -3969,6 +4436,25 @@
         "micromark-util-symbol": "^2.0.0"
       }
     },
+    "node_modules/micromark-util-edit-map": {
+      "version": "1.0.0",
+      "resolved": "https://registry.npmjs.org/micromark-util-edit-map/-/micromark-util-edit-map-1.0.0.tgz",
+      "integrity": "sha512-Pa2ljlsEL6sVwFaYeyrOLSYbQt73JvGbPOYFq+9AElXiSnfI5Q4465RRZ1sHv7lDjDOnGRDDaqPXqClqWK/q1Q==",
+      "funding": [
+        {
+          "type": "GitHub Sponsors",
+          "url": "https://github.com/sponsors/unifiedjs"
+        },
+        {
+          "type": "OpenCollective",
+          "url": "https://opencollective.com/unified"
+        }
+      ],
+      "license": "MIT",
+      "dependencies": {
+        "micromark-util-types": "^2.0.0"
+      }
+    },
     "node_modules/micromark-util-encode": {
       "version": "2.0.1",
       "resolved": "https://registry.npmjs.org/micromark-util-encode/-/micromark-util-encode-2.0.1.tgz",
@@ -4099,9 +4585,9 @@
       "license": "MIT"
     },
     "node_modules/micromark-util-types": {
-      "version": "2.0.2",
-      "resolved": "https://registry.npmjs.org/micromark-util-types/-/micromark-util-types-2.0.2.tgz",
-      "integrity": "sha512-Yw0ECSpJoViF1qTU4DC6NwtC4aWGt1EkzaQB8KPPyCRR8z9TWeV0HbEFGTO+ZY1wB22zmxnJqhPyTpOVCpeHTA==",
+      "version": "2.0.3",
+      "resolved": "https://registry.npmjs.org/micromark-util-types/-/micromark-util-types-2.0.3.tgz",
+      "integrity": "sha512-oxB2Ik03hI0gv+VNn9tnh1t1YEe9MDPptViAEgfdf3YQHsn0pzGTgCdlSCJXcwhqm8phaEuM7zeEu3QQzVBrPg==",
       "funding": [
         {
           "type": "GitHub Sponsors",
@@ -4128,48 +4614,32 @@
         "node": ">=8.6"
       }
     },
-    "node_modules/micromatch/node_modules/picomatch": {
-      "version": "2.3.2",
-      "resolved": "https://registry.npmjs.org/picomatch/-/picomatch-2.3.2.tgz",
-      "integrity": "sha512-V7+vQEJ06Z+c5tSye8S+nHUfI51xoXIXjHQ99cQtKUkQqqO1kO/KCJUfZXuB47h/YBlDhah2H3hdUGXn8ie0oA==",
-      "dev": true,
-      "license": "MIT",
-      "engines": {
-        "node": ">=8.6"
-      },
-      "funding": {
-        "url": "https://github.com/sponsors/jonschlinkert"
-      }
-    },
     "node_modules/minimatch": {
-      "version": "3.1.5",
-      "resolved": "https://registry.npmjs.org/minimatch/-/minimatch-3.1.5.tgz",
-      "integrity": "sha512-VgjWUsnnT6n+NUk6eZq77zeFdpW2LWDzP6zFGrCbHXiYNul5Dzqk2HHQ5uFH2DNW5Xbp8+jVzaeNt94ssEEl4w==",
+      "version": "10.2.6",
+      "resolved": "https://registry.npmjs.org/minimatch/-/minimatch-10.2.6.tgz",
+      "integrity": "sha512-vpLQEs+VLCr1nU0BXS07maYoFwlDAH0gngQuuttxIwutDFEMHq2blX+8vpgxDdK3J1PwjCJiep77OitTZ4Ll1A==",
       "dev": true,
-      "license": "ISC",
+      "license": "BlueOak-1.0.0",
       "dependencies": {
-        "brace-expansion": "^1.1.7"
+        "brace-expansion": "^5.0.8"
       },
       "engines": {
-        "node": "*"
+        "node": "18 || 20 || >=22"
+      },
+      "funding": {
+        "url": "https://github.com/sponsors/isaacs"
       }
     },
     "node_modules/minipass": {
-      "version": "7.1.2",
-      "resolved": "https://registry.npmjs.org/minipass/-/minipass-7.1.2.tgz",
-      "integrity": "sha512-qOOzS1cBTWYF4BH8fVePDBOO9iptMnGUEZwNc/cMWnTV2nVLZ7VoNWEPHkYczZA0pdoA7dl6e7FL659nX9S2aw==",
+      "version": "7.1.3",
+      "resolved": "https://registry.npmjs.org/minipass/-/minipass-7.1.3.tgz",
+      "integrity": "sha512-tEBHqDnIoM/1rXME1zgka9g6Q2lcoCkxHLuc7ODJ5BxbP5d4c2Z5cGgtXAku59200Cx7diuHTOYfSBD8n6mm8A==",
       "dev": true,
-      "license": "ISC",
+      "license": "BlueOak-1.0.0",
       "engines": {
         "node": ">=16 || 14 >=14.17"
       }
     },
-    "node_modules/mitt": {
-      "version": "3.0.1",
-      "resolved": "https://registry.npmjs.org/mitt/-/mitt-3.0.1.tgz",
-      "integrity": "sha512-vKivATfr97l2/QBCYAkXYDbrIWPM2IIKEl7YPhjCvKlG3kE2gm+uBo6nEXK3M5/Ffh/FLpKExzOQ3JJoJGFKBw==",
-      "license": "MIT"
-    },
     "node_modules/mkdirp": {
       "version": "3.0.1",
       "resolved": "https://registry.npmjs.org/mkdirp/-/mkdirp-3.0.1.tgz",
@@ -4193,9 +4663,9 @@
       "license": "MIT"
     },
     "node_modules/nanoid": {
-      "version": "5.1.16",
-      "resolved": "https://registry.npmjs.org/nanoid/-/nanoid-5.1.16.tgz",
-      "integrity": "sha512-kVrnsrJqMR8+oLJnGEmSWw9BivK5mt7H3FZatVRjrc5wGqFYuBxX1yG7+A7Gi5AefkX6t/oCkizcQgpu0cY1dQ==",
+      "version": "6.0.2",
+      "resolved": "https://registry.npmjs.org/nanoid/-/nanoid-6.0.2.tgz",
+      "integrity": "sha512-hIvVOb9Qr1p/9CNNLOe0OlWvgx/ziTsVwKN/GDSsuArN2K0wDiafP3s0SWmsm7o2BdnyS8Z9Z64NFBUbn6ziBg==",
       "funding": [
         {
           "type": "github",
@@ -4207,7 +4677,7 @@
         "nanoid": "bin/nanoid.js"
       },
       "engines": {
-        "node": "^18 || >=20"
+        "node": "^22 || ^24 || >=26"
       }
     },
     "node_modules/natural-compare": {
@@ -4218,18 +4688,27 @@
       "license": "MIT"
     },
     "node_modules/node-notifier": {
-      "version": "10.0.1",
-      "resolved": "https://registry.npmjs.org/node-notifier/-/node-notifier-10.0.1.tgz",
-      "integrity": "sha512-YX7TSyDukOZ0g+gmzjB6abKu+hTGvO8+8+gIFDsRCU2t8fLV/P2unmt+LGFaIa4y64aX98Qksa97rgz4vMNeLQ==",
+      "version": "11.0.0",
+      "resolved": "https://registry.npmjs.org/node-notifier/-/node-notifier-11.0.0.tgz",
+      "integrity": "sha512-C4ePuKXvH+Xd4NpZ4iQyTCEAQDMdVrhH1ntJ4v3pLaIKNuODtM88XyvEcwVceqK54mBzMYlGjkiUHlvtqCdgYg==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
         "growly": "^1.3.0",
-        "is-wsl": "^2.2.0",
-        "semver": "^7.3.5",
-        "shellwords": "^0.1.1",
-        "uuid": "^8.3.2",
-        "which": "^2.0.2"
+        "is-wsl": "^3.1.1"
+      },
+      "engines": {
+        "node": "^22.22.2 || ^24.15.0 || >=26.0.0"
+      }
+    },
+    "node_modules/node-releases": {
+      "version": "2.0.60",
+      "resolved": "https://registry.npmjs.org/node-releases/-/node-releases-2.0.60.tgz",
+      "integrity": "sha512-OmO8+GM1JUZqXET720VrJCwkjUCXmCMQm0Z3UDDRbfNLHS4NEAvY60j5jiFv0tkKw2PNNlqOLhahVmwJIOLDIQ==",
+      "dev": true,
+      "license": "MIT",
+      "engines": {
+        "node": ">=18"
       }
     },
     "node_modules/normalize-path": {
@@ -4312,16 +4791,6 @@
         "node": ">=6"
       }
     },
-    "node_modules/parent-module/node_modules/callsites": {
-      "version": "3.1.0",
-      "resolved": "https://registry.npmjs.org/callsites/-/callsites-3.1.0.tgz",
-      "integrity": "sha512-P8BjAsXvZS+VIDUI11hHCQEv74YT67YUi5JJFNWIqL235sBmjX4+qx9Muvls5ivyNENctx46xQLQ3aTuE7ssaQ==",
-      "dev": true,
-      "license": "MIT",
-      "engines": {
-        "node": ">=6"
-      }
-    },
     "node_modules/parse-entities": {
       "version": "4.0.2",
       "resolved": "https://registry.npmjs.org/parse-entities/-/parse-entities-4.0.2.tgz",
@@ -4387,9 +4856,9 @@
       }
     },
     "node_modules/path-scurry": {
-      "version": "2.0.0",
-      "resolved": "https://registry.npmjs.org/path-scurry/-/path-scurry-2.0.0.tgz",
-      "integrity": "sha512-ypGJsmGtdXUOeM5u93TyeIEfEhM6s+ljAhrk5vAvSx8uyY/02OvrZnA0YNGUrPXfpJMgI1ODd3nwz8Npx4O4cg==",
+      "version": "2.0.2",
+      "resolved": "https://registry.npmjs.org/path-scurry/-/path-scurry-2.0.2.tgz",
+      "integrity": "sha512-3O/iVVsJAPsOnpwWIeD+d6z/7PmqApyQePUtCndjatj/9I5LylHvt5qluFaBT3I5h3r1ejfR056c+FCv+NnNXg==",
       "dev": true,
       "license": "BlueOak-1.0.0",
       "dependencies": {
@@ -4397,20 +4866,20 @@
         "minipass": "^7.1.2"
       },
       "engines": {
-        "node": "20 || >=22"
+        "node": "18 || 20 || >=22"
       },
       "funding": {
         "url": "https://github.com/sponsors/isaacs"
       }
     },
-    "node_modules/path-to-regexp": {
-      "version": "8.4.2",
-      "resolved": "https://registry.npmjs.org/path-to-regexp/-/path-to-regexp-8.4.2.tgz",
-      "integrity": "sha512-qRcuIdP69NPm4qbACK+aDogI5CBDMi1jKe0ry5rSQJz8JVLsC7jV8XpiJjGRLLol3N+R5ihGYcrPLTno6pAdBA==",
-      "license": "MIT",
-      "funding": {
-        "type": "opencollective",
-        "url": "https://opencollective.com/express"
+    "node_modules/path-scurry/node_modules/lru-cache": {
+      "version": "11.5.3",
+      "resolved": "https://registry.npmjs.org/lru-cache/-/lru-cache-11.5.3.tgz",
+      "integrity": "sha512-U4N8FgzmWxc8k1VH8Kr6lQg18U7Fjvby6wXHVRX/ZZ7IwWbRMgrRbP0Wrb5q5NVinryp4SQampHKdvtecItxUg==",
+      "dev": true,
+      "license": "BlueOak-1.0.0",
+      "engines": {
+        "node": "20 || >=22"
       }
     },
     "node_modules/picocolors": {
@@ -4420,10 +4889,23 @@
       "dev": true,
       "license": "ISC"
     },
+    "node_modules/picomatch": {
+      "version": "2.3.2",
+      "resolved": "https://registry.npmjs.org/picomatch/-/picomatch-2.3.2.tgz",
+      "integrity": "sha512-V7+vQEJ06Z+c5tSye8S+nHUfI51xoXIXjHQ99cQtKUkQqqO1kO/KCJUfZXuB47h/YBlDhah2H3hdUGXn8ie0oA==",
+      "dev": true,
+      "license": "MIT",
+      "engines": {
+        "node": ">=8.6"
+      },
+      "funding": {
+        "url": "https://github.com/sponsors/jonschlinkert"
+      }
+    },
     "node_modules/postcss": {
-      "version": "8.5.28",
-      "resolved": "https://registry.npmjs.org/postcss/-/postcss-8.5.28.tgz",
-      "integrity": "sha512-RRuzqDtt5Y9h3quz5hWhK+TPnsmVs6WwSU6LkJMeY4HstUEDuYTG8UJSdawMRzmzAtV+KEoG8N3Qg2qLy5vM/A==",
+      "version": "8.5.31",
+      "resolved": "https://registry.npmjs.org/postcss/-/postcss-8.5.31.tgz",
+      "integrity": "sha512-YUbWYCgErwnMAVkjUnYa+KlMBzGpHwiIt/DE7f0G33AuX2LH9UcTk6KgWxd3zRU+pN3u/C+Xl3EnMLM4Hea7Sg==",
       "dev": true,
       "funding": [
         {
@@ -4441,9 +4923,9 @@
       ],
       "license": "MIT",
       "dependencies": {
-        "nanoid": "^3.3.18",
+        "nanoid": "^3.3.20",
         "picocolors": "^1.1.1",
-        "source-map-js": "^1.2.1"
+        "source-map-js": "^1.2.2"
       },
       "engines": {
         "node": "^10 || ^12 || >=14"
@@ -4512,17 +4994,10 @@
         "postcss": "^8.1.0"
       }
     },
-    "node_modules/postcss-resolve-nested-selector": {
-      "version": "0.1.6",
-      "resolved": "https://registry.npmjs.org/postcss-resolve-nested-selector/-/postcss-resolve-nested-selector-0.1.6.tgz",
-      "integrity": "sha512-0sglIs9Wmkzbr8lQwEyIzlDOOC9bGmfVKcJTaxv3vMmd3uo4o4DerC3En0bnmgceeql9BfC8hRkp7cg0fjdVqw==",
-      "dev": true,
-      "license": "MIT"
-    },
     "node_modules/postcss-safe-parser": {
-      "version": "7.0.1",
-      "resolved": "https://registry.npmjs.org/postcss-safe-parser/-/postcss-safe-parser-7.0.1.tgz",
-      "integrity": "sha512-0AioNCJZ2DPYz5ABT6bddIqlhgwhpHZ/l65YAYo0BCIn0xiDpsnTHz0gnoTGk0OXZW0JRs+cDwL8u/teRdz+8A==",
+      "version": "7.1.0",
+      "resolved": "https://registry.npmjs.org/postcss-safe-parser/-/postcss-safe-parser-7.1.0.tgz",
+      "integrity": "sha512-1WzZxRLaAFwEh6Do+zyGpjWV3nGJNxxhuh7Ubu/q1ICImMgZJnLwhoaEpRbG4pJppp2Y1ncL9ffA2f+LhrefQg==",
       "dev": true,
       "funding": [
         {
@@ -4547,9 +5022,9 @@
       }
     },
     "node_modules/postcss-selector-parser": {
-      "version": "7.1.6",
-      "resolved": "https://registry.npmjs.org/postcss-selector-parser/-/postcss-selector-parser-7.1.6.tgz",
-      "integrity": "sha512-7qASPzhKF2l2KLboRZux8CCTRMdGiV08vWmyKzPz22qZ7ZjQBOeY7rNzNoCLSUiftJ7HUq0GERHmxw/t0dCdMw==",
+      "version": "7.1.7",
+      "resolved": "https://registry.npmjs.org/postcss-selector-parser/-/postcss-selector-parser-7.1.7.tgz",
+      "integrity": "sha512-wWBro/7hJ40ye6TPiXkUQoXKCw23xSsjriUvqd7Uls9Sso3oLiON1flc18cCdnjgCRXw4fBVqUJJeOdDTh3P3g==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
@@ -4568,9 +5043,9 @@
       "license": "MIT"
     },
     "node_modules/postcss/node_modules/nanoid": {
-      "version": "3.3.19",
-      "resolved": "https://registry.npmjs.org/nanoid/-/nanoid-3.3.19.tgz",
-      "integrity": "sha512-Y2tUNy4ouw6tq5oDSKeQYGOyhkUBhNOcGV/02KC+6kd9eDGqdZd++mjMiIDilrBYvjEnCYvVtsuHCuP+okSfug==",
+      "version": "3.3.20",
+      "resolved": "https://registry.npmjs.org/nanoid/-/nanoid-3.3.20.tgz",
+      "integrity": "sha512-uKdg2G3GNCKQn9byYOpxbGqrT2fGO5KRt5J/8b3pok8rT6qxGWF6hxMyJiEYtAf+FVyYuD9hRaDqX5uPFYJ4ZQ==",
       "dev": true,
       "funding": [
         {
@@ -4597,9 +5072,9 @@
       }
     },
     "node_modules/prettier": {
-      "version": "3.6.2",
-      "resolved": "https://registry.npmjs.org/prettier/-/prettier-3.6.2.tgz",
-      "integrity": "sha512-I7AIg5boAr5R0FFtJ6rCfD+LFsWHp81dolrFD8S79U9tb8Az2nGrJncnMSnys+bpQJfRUzqs9hnA81OAA3hCuQ==",
+      "version": "3.9.10",
+      "resolved": "https://registry.npmjs.org/prettier/-/prettier-3.9.10.tgz",
+      "integrity": "sha512-skGHTH1Yb9QvaTC4hWnRTzEYtCpApF0UMBPz+t/0ZPzP8ihefo/btKd+QA7pUE0fEwA1SbS/NmhyEcstdXOIdA==",
       "dev": true,
       "license": "MIT",
       "bin": {
@@ -4613,9 +5088,9 @@
       }
     },
     "node_modules/prettier-plugin-organize-imports": {
-      "version": "4.2.0",
-      "resolved": "https://registry.npmjs.org/prettier-plugin-organize-imports/-/prettier-plugin-organize-imports-4.2.0.tgz",
-      "integrity": "sha512-Zdy27UhlmyvATZi67BTnLcKTo8fm6Oik59Sz6H64PgZJVs6NJpPD1mT240mmJn62c98/QaL+r3kx9Q3gRpDajg==",
+      "version": "4.3.0",
+      "resolved": "https://registry.npmjs.org/prettier-plugin-organize-imports/-/prettier-plugin-organize-imports-4.3.0.tgz",
+      "integrity": "sha512-FxFz0qFhyBsGdIsb697f/EkvHzi5SZOhWAjxcx2dLt+Q532bAlhswcXGYB1yzjZ69kW8UoadFBw7TyNwlq96Iw==",
       "dev": true,
       "license": "MIT",
       "peerDependencies": {
@@ -4630,9 +5105,9 @@
       }
     },
     "node_modules/property-information": {
-      "version": "7.1.0",
-      "resolved": "https://registry.npmjs.org/property-information/-/property-information-7.1.0.tgz",
-      "integrity": "sha512-TwEZ+X+yCJmYfL7TPUOcvBZ4QfoT5YenQiJuX//0th53DE6w0xxLEtfK3iyryQFddXuvkIk51EEgrJQ0WJkOmQ==",
+      "version": "7.2.0",
+      "resolved": "https://registry.npmjs.org/property-information/-/property-information-7.2.0.tgz",
+      "integrity": "sha512-IAtzIB6sUiWaJYrX9smp3V46pBGbBeLFRGdh25kg1334VcBlD8HzhPeNIWQH9zhGmo2itIe25EHt9dQP7G5hmg==",
       "license": "MIT",
       "funding": {
         "type": "github",
@@ -4649,6 +5124,26 @@
         "node": ">=6"
       }
     },
+    "node_modules/qified": {
+      "version": "0.10.1",
+      "resolved": "https://registry.npmjs.org/qified/-/qified-0.10.1.tgz",
+      "integrity": "sha512-+Owyggi9IxT1ePKGafcI87ubSmxol6smwJ+RAHDQlx9+9cPwFWDiKFFCPuWhr9ignlGpZ9vDQLw67N4dcTVFEA==",
+      "dev": true,
+      "license": "MIT",
+      "dependencies": {
+        "hookified": "^2.1.1"
+      },
+      "engines": {
+        "node": ">=20"
+      }
+    },
+    "node_modules/qified/node_modules/hookified": {
+      "version": "2.2.0",
+      "resolved": "https://registry.npmjs.org/hookified/-/hookified-2.2.0.tgz",
+      "integrity": "sha512-p/LgFzRN5FeoD3DLS6bkUapeye6E4SI6yJs6KetENd18S+FBthqYq2amJUWpt5z0EQwwHemidjY5OqJGEKm5uA==",
+      "dev": true,
+      "license": "MIT"
+    },
     "node_modules/queue-microtask": {
       "version": "1.2.3",
       "resolved": "https://registry.npmjs.org/queue-microtask/-/queue-microtask-1.2.3.tgz",
@@ -4671,24 +5166,24 @@
       "license": "MIT"
     },
     "node_modules/react": {
-      "version": "19.1.1",
-      "resolved": "https://registry.npmjs.org/react/-/react-19.1.1.tgz",
-      "integrity": "sha512-w8nqGImo45dmMIfljjMwOGtbmC/mk4CMYhWIicdSflH91J9TyCyczcPFXJzrZ/ZXcgGRFeP6BU0BEJTw6tZdfQ==",
+      "version": "19.3.1",
+      "resolved": "https://registry.npmjs.org/react/-/react-19.3.1.tgz",
+      "integrity": "sha512-QNhWdt7nrgVWWQG58QuGsAR3jv3Hbe4mb4vz7w2csnHavciZJgda1CPNaldF6eVxv90ud3jUXpKlHL69gO38yw==",
       "license": "MIT",
       "engines": {
         "node": ">=0.10.0"
       }
     },
     "node_modules/react-dom": {
-      "version": "19.1.1",
-      "resolved": "https://registry.npmjs.org/react-dom/-/react-dom-19.1.1.tgz",
-      "integrity": "sha512-Dlq/5LAZgF0Gaz6yiqZCf6VCcZs1ghAJyrsu84Q/GT0gV+mCxbfmKNoGRKBYMJ8IEdGPqu49YWXD02GCknEDkw==",
+      "version": "19.3.1",
+      "resolved": "https://registry.npmjs.org/react-dom/-/react-dom-19.3.1.tgz",
+      "integrity": "sha512-Y0bARz53lhEC1NMCASs0/j+faB/zqt3R8hpvg1e8+61ZVJ4UL0UdC4cBbXfZMcIuhrfgeILTLe+ZeBCavQPtKQ==",
       "license": "MIT",
       "dependencies": {
-        "scheduler": "^0.26.0"
+        "scheduler": "^0.28.0"
       },
       "peerDependencies": {
-        "react": "^19.1.1"
+        "react": "^19.3.1"
       }
     },
     "node_modules/react-markdown": {
@@ -4795,13 +5290,13 @@
       }
     },
     "node_modules/resolve-from": {
-      "version": "5.0.0",
-      "resolved": "https://registry.npmjs.org/resolve-from/-/resolve-from-5.0.0.tgz",
-      "integrity": "sha512-qYg9KP24dD5qka9J47d0aVky0N+b4fTU89LN9iDnjB5waksiC49rvMB0PrUJQGoTmH50XPiqOvAjDfaijGxYZw==",
+      "version": "4.0.0",
+      "resolved": "https://registry.npmjs.org/resolve-from/-/resolve-from-4.0.0.tgz",
+      "integrity": "sha512-pb/MYmXstAkysRFx8piNI1tGFNQIFA3vkE3Gq4EuA1dF6gHp/+vgZqsCGJapvy8N3Q+4o7FwvquPJcnZ7RYy4g==",
       "dev": true,
       "license": "MIT",
       "engines": {
-        "node": ">=8"
+        "node": ">=4"
       }
     },
     "node_modules/reusify": {
@@ -4849,22 +5344,19 @@
       }
     },
     "node_modules/scheduler": {
-      "version": "0.26.0",
-      "resolved": "https://registry.npmjs.org/scheduler/-/scheduler-0.26.0.tgz",
-      "integrity": "sha512-NlHwttCI/l5gCPR3D1nNXtWABUmBwvZpEQiD4IXSbIDq8BzLIK/7Ir5gTFSGZDUu37K5cMNp0hFtzO38sC7gWA==",
+      "version": "0.28.0",
+      "resolved": "https://registry.npmjs.org/scheduler/-/scheduler-0.28.0.tgz",
+      "integrity": "sha512-juorfCmIkIw8tT+p5BXSm6PJjQF/ycEYmKyzURCIt/RaZIhL+PulbQ9Yu2z1HdOJDdqDTlxA1+xKBmHXJsczAw==",
       "license": "MIT"
     },
     "node_modules/semver": {
-      "version": "7.7.2",
-      "resolved": "https://registry.npmjs.org/semver/-/semver-7.7.2.tgz",
-      "integrity": "sha512-RF0Fw+rO5AMf9MAyaRXI4AV0Ulj5lMHqVxxdSgiVbixSCXoEmmX/jk0CuJw4+3SqroYO9VoUh+HcuJivvtJemA==",
+      "version": "6.3.1",
+      "resolved": "https://registry.npmjs.org/semver/-/semver-6.3.1.tgz",
+      "integrity": "sha512-BR7VvDCVHO+q2xBEWskxS6DJE1qRnb7DxzUrogb71CWoSficBxYsiAGd+Kl0mmq/MprG9yArRkyrQxTO6XjMzA==",
       "dev": true,
       "license": "ISC",
       "bin": {
         "semver": "bin/semver.js"
-      },
-      "engines": {
-        "node": ">=10"
       }
     },
     "node_modules/shebang-command": {
@@ -4890,13 +5382,6 @@
         "node": ">=8"
       }
     },
-    "node_modules/shellwords": {
-      "version": "0.1.1",
-      "resolved": "https://registry.npmjs.org/shellwords/-/shellwords-0.1.1.tgz",
-      "integrity": "sha512-vFwSUfQvqybiICwZY5+DAWIPLKsWO31Q91JSKl3UYv+K5c2QRPzn0qzec6QPu1Qc9eHYItiP3NdJqNVqetYAww==",
-      "dev": true,
-      "license": "MIT"
-    },
     "node_modules/signal-exit": {
       "version": "4.1.0",
       "resolved": "https://registry.npmjs.org/signal-exit/-/signal-exit-4.1.0.tgz",
@@ -4910,10 +5395,41 @@
         "url": "https://github.com/sponsors/isaacs"
       }
     },
+    "node_modules/slash": {
+      "version": "5.1.0",
+      "resolved": "https://registry.npmjs.org/slash/-/slash-5.1.0.tgz",
+      "integrity": "sha512-ZA6oR3T/pEyuqwMgAKT0/hAv8oAXckzbkmR0UkUosQ+Mc4RxGoJkRmwHgHufaenlyAgE1Mxgpdcrf75y6XcnDg==",
+      "dev": true,
+      "license": "MIT",
+      "engines": {
+        "node": ">=14.16"
+      },
+      "funding": {
+        "url": "https://github.com/sponsors/sindresorhus"
+      }
+    },
+    "node_modules/slice-ansi": {
+      "version": "4.0.0",
+      "resolved": "https://registry.npmjs.org/slice-ansi/-/slice-ansi-4.0.0.tgz",
+      "integrity": "sha512-qMCMfhY040cVHT43K9BFygqYbUPFZKHOg7K73mtTWJRb8pyP3fzf4Ixd5SzdEJQ6MRUg/WBnOLxghZtKKurENQ==",
+      "dev": true,
+      "license": "MIT",
+      "dependencies": {
+        "ansi-styles": "^4.0.0",
+        "astral-regex": "^2.0.0",
+        "is-fullwidth-code-point": "^3.0.0"
+      },
+      "engines": {
+        "node": ">=10"
+      },
+      "funding": {
+        "url": "https://github.com/chalk/slice-ansi?sponsor=1"
+      }
+    },
     "node_modules/source-map-js": {
-      "version": "1.2.1",
-      "resolved": "https://registry.npmjs.org/source-map-js/-/source-map-js-1.2.1.tgz",
-      "integrity": "sha512-UXWMKhLOwVKb728IUtQPXxfYU+usdybtUrK/8uGE8CQMvrhOpwvzDBwj0QhSL7MQc7vIsISBG8VQ8+IDQxpfQA==",
+      "version": "1.2.2",
+      "resolved": "https://registry.npmjs.org/source-map-js/-/source-map-js-1.2.2.tgz",
+      "integrity": "sha512-KGj/8Y43x35aZVDtt+J4mK1hoLGHULMYfSkODJNQjNDC3oW1PqPoxMwo0pLUsWM/UEGzON/NxeHywEfNXNP3Vw==",
       "dev": true,
       "license": "BSD-3-Clause",
       "engines": {
@@ -4940,6 +5456,23 @@
         "node": ">=0.6.19"
       }
     },
+    "node_modules/string-width": {
+      "version": "8.3.0",
+      "resolved": "https://registry.npmjs.org/string-width/-/string-width-8.3.0.tgz",
+      "integrity": "sha512-ZbmZM0JCihQN91dWnxoipT2KOEyHqEyfRXUyjuRhW8b/xnqPDoq4gWEVApTVa9db2wN8mmoikgFBbjh71+cGeQ==",
+      "dev": true,
+      "license": "MIT",
+      "dependencies": {
+        "get-east-asian-width": "^1.5.0",
+        "strip-ansi": "^7.1.2"
+      },
+      "engines": {
+        "node": ">=20"
+      },
+      "funding": {
+        "url": "https://github.com/sponsors/sindresorhus"
+      }
+    },
     "node_modules/string-width-cjs": {
       "name": "string-width",
       "version": "4.2.3",
@@ -4966,23 +5499,6 @@
         "node": ">=8"
       }
     },
-    "node_modules/string-width-cjs/node_modules/emoji-regex": {
-      "version": "8.0.0",
-      "resolved": "https://registry.npmjs.org/emoji-regex/-/emoji-regex-8.0.0.tgz",
-      "integrity": "sha512-MSjYzcWNOA0ewAHpz0MxpYFvwg6yjy1NG3xteoqz644VCo/RPgnr1/GGt+ic3iJTzQ8Eu3TdM14SawnVUmGE6A==",
-      "dev": true,
-      "license": "MIT"
-    },
-    "node_modules/string-width-cjs/node_modules/is-fullwidth-code-point": {
-      "version": "3.0.0",
-      "resolved": "https://registry.npmjs.org/is-fullwidth-code-point/-/is-fullwidth-code-point-3.0.0.tgz",
-      "integrity": "sha512-zymm5+u+sCsSWyD9qNaejV3DFvhCKclKdizYaJUuHA83RLjb7nSuGnddCHGv0hk+KY7BMAlsWeK4Ueg6EV6XQg==",
-      "dev": true,
-      "license": "MIT",
-      "engines": {
-        "node": ">=8"
-      }
-    },
     "node_modules/string-width-cjs/node_modules/strip-ansi": {
       "version": "6.0.1",
       "resolved": "https://registry.npmjs.org/strip-ansi/-/strip-ansi-6.0.1.tgz",
@@ -5011,13 +5527,13 @@
       }
     },
     "node_modules/strip-ansi": {
-      "version": "7.1.0",
-      "resolved": "https://registry.npmjs.org/strip-ansi/-/strip-ansi-7.1.0.tgz",
-      "integrity": "sha512-iq6eVVI64nQQTRYq2KtEg2d2uU7LElhTJwsH4YzIHZshxlgZms/wIc4VoDQTlG/IvVIrBKG06CrZnp0qv7hkcQ==",
+      "version": "7.2.0",
+      "resolved": "https://registry.npmjs.org/strip-ansi/-/strip-ansi-7.2.0.tgz",
+      "integrity": "sha512-yDPMNjp4WyfYBkHnjIRLfca1i6KMyGCtsVgoKe/z1+6vukgaENdgGBZt+ZmKPc4gavvEZ5OgHfHdrazhgNyG7w==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
-        "ansi-regex": "^6.0.1"
+        "ansi-regex": "^6.2.2"
       },
       "engines": {
         "node": ">=12"
@@ -5050,47 +5566,34 @@
         "node": ">=8"
       }
     },
-    "node_modules/strip-json-comments": {
-      "version": "3.1.1",
-      "resolved": "https://registry.npmjs.org/strip-json-comments/-/strip-json-comments-3.1.1.tgz",
-      "integrity": "sha512-6fPc+R4ihwqP6N/aIv2f1gMH8lOVtWQHoqC4yK6oSDVVocumAsfCqjkXnqiYMhmMwS/mEHLp7Vehlt3ql6lEig==",
-      "dev": true,
-      "license": "MIT",
-      "engines": {
-        "node": ">=8"
-      },
-      "funding": {
-        "url": "https://github.com/sponsors/sindresorhus"
-      }
-    },
     "node_modules/style-mod": {
-      "version": "4.1.2",
-      "resolved": "https://registry.npmjs.org/style-mod/-/style-mod-4.1.2.tgz",
-      "integrity": "sha512-wnD1HyVqpJUI2+eKZ+eo1UwghftP6yuFheBqqe+bWCotBjC2K1YnteJILRMs3SM4V/0dLEW1SC27MWP5y+mwmw==",
+      "version": "4.1.4",
+      "resolved": "https://registry.npmjs.org/style-mod/-/style-mod-4.1.4.tgz",
+      "integrity": "sha512-XXWIQt633/EpAFx8aZDOTjBzrCaGmhvEQlQo6MVPfa2OzO2cWo+4hV9h+6UkHYlXGfy+ODXKUdP7Pthmcu5ATw==",
       "license": "MIT"
     },
     "node_modules/style-to-js": {
-      "version": "1.1.17",
-      "resolved": "https://registry.npmjs.org/style-to-js/-/style-to-js-1.1.17.tgz",
-      "integrity": "sha512-xQcBGDxJb6jjFCTzvQtfiPn6YvvP2O8U1MDIPNfJQlWMYfktPy+iGsHE7cssjs7y84d9fQaK4UF3RIJaAHSoYA==",
+      "version": "1.1.21",
+      "resolved": "https://registry.npmjs.org/style-to-js/-/style-to-js-1.1.21.tgz",
+      "integrity": "sha512-RjQetxJrrUJLQPHbLku6U/ocGtzyjbJMP9lCNK7Ag0CNh690nSH8woqWH9u16nMjYBAok+i7JO1NP2pOy8IsPQ==",
       "license": "MIT",
       "dependencies": {
-        "style-to-object": "1.0.9"
+        "style-to-object": "1.0.14"
       }
     },
     "node_modules/style-to-object": {
-      "version": "1.0.9",
-      "resolved": "https://registry.npmjs.org/style-to-object/-/style-to-object-1.0.9.tgz",
-      "integrity": "sha512-G4qppLgKu/k6FwRpHiGiKPaPTFcG3g4wNVX/Qsfu+RqQM30E7Tyu/TEgxcL9PNLF5pdRLwQdE3YKKf+KF2Dzlw==",
+      "version": "1.0.14",
+      "resolved": "https://registry.npmjs.org/style-to-object/-/style-to-object-1.0.14.tgz",
+      "integrity": "sha512-LIN7rULI0jBscWQYaSswptyderlarFkjQ+t79nzty8tcIAceVomEVlLzH5VP4Cmsv6MtKhs7qaAiwlcp+Mgaxw==",
       "license": "MIT",
       "dependencies": {
-        "inline-style-parser": "0.2.4"
+        "inline-style-parser": "0.2.7"
       }
     },
     "node_modules/stylelint": {
-      "version": "16.23.1",
-      "resolved": "https://registry.npmjs.org/stylelint/-/stylelint-16.23.1.tgz",
-      "integrity": "sha512-dNvDTsKV1U2YtiUDfe9d2gp902veFeo3ecCWdGlmLm2WFrAV0+L5LoOj/qHSBABQwMsZPJwfC4bf39mQm1S5zw==",
+      "version": "17.16.0",
+      "resolved": "https://registry.npmjs.org/stylelint/-/stylelint-17.16.0.tgz",
+      "integrity": "sha512-ejlWnbJ13x1GU32+zZXa6sYQJsKKkwSdv2B4rAdO3OWdEmvhSJKcXi5pOphQYSzyaj1kXpuGJr7pcMBNfDIuPw==",
       "dev": true,
       "funding": [
         {
@@ -5104,56 +5607,53 @@
       ],
       "license": "MIT",
       "dependencies": {
-        "@csstools/css-parser-algorithms": "^3.0.5",
-        "@csstools/css-tokenizer": "^3.0.4",
-        "@csstools/media-query-list-parser": "^4.0.3",
-        "@csstools/selector-specificity": "^5.0.0",
-        "@dual-bundle/import-meta-resolve": "^4.1.0",
-        "balanced-match": "^2.0.0",
-        "colord": "^2.9.3",
-        "cosmiconfig": "^9.0.0",
-        "css-functions-list": "^3.2.3",
-        "css-tree": "^3.1.0",
-        "debug": "^4.4.1",
+        "@csstools/css-calc": "^3.3.0",
+        "@csstools/css-parser-algorithms": "^4.0.0",
+        "@csstools/css-syntax-patches-for-csstree": "^1.1.9",
+        "@csstools/css-tokenizer": "^4.0.0",
+        "@csstools/media-query-list-parser": "^5.0.0",
+        "@csstools/selector-resolve-nested": "^4.0.1",
+        "@csstools/selector-specificity": "^6.0.0",
+        "colord": "^2.10.0",
+        "cosmiconfig": "^9.0.2",
+        "css-functions-list": "^3.3.3",
+        "css-tree": "^3.2.1",
+        "debug": "^4.4.3",
         "fast-glob": "^3.3.3",
         "fastest-levenshtein": "^1.0.16",
-        "file-entry-cache": "^10.1.3",
+        "file-entry-cache": "^11.1.5",
         "global-modules": "^2.0.0",
-        "globby": "^11.1.0",
+        "globby": "^16.2.4",
         "globjoin": "^0.1.4",
-        "html-tags": "^3.3.1",
-        "ignore": "^7.0.5",
-        "imurmurhash": "^0.1.4",
-        "is-plain-object": "^5.0.0",
-        "known-css-properties": "^0.37.0",
-        "mathml-tag-names": "^2.1.3",
-        "meow": "^13.2.0",
+        "html-tags": "^5.1.0",
+        "ignore": "^7.0.6",
+        "import-meta-resolve": "^4.2.0",
+        "mathml-tag-names": "^4.0.0",
+        "meow": "^14.1.0",
         "micromatch": "^4.0.8",
         "normalize-path": "^3.0.0",
         "picocolors": "^1.1.1",
-        "postcss": "^8.5.6",
-        "postcss-resolve-nested-selector": "^0.1.6",
+        "postcss": "^8.5.26",
         "postcss-safe-parser": "^7.0.1",
-        "postcss-selector-parser": "^7.1.0",
+        "postcss-selector-parser": "^7.1.5",
         "postcss-value-parser": "^4.2.0",
-        "resolve-from": "^5.0.0",
-        "string-width": "^4.2.3",
-        "supports-hyperlinks": "^3.2.0",
+        "string-width": "^8.2.2",
+        "supports-hyperlinks": "^4.5.0",
         "svg-tags": "^1.0.0",
         "table": "^6.9.0",
-        "write-file-atomic": "^5.0.1"
+        "write-file-atomic": "^7.0.1"
       },
       "bin": {
         "stylelint": "bin/stylelint.mjs"
       },
       "engines": {
-        "node": ">=18.12.0"
+        "node": ">=20.19.0"
       }
     },
     "node_modules/stylelint-config-recommended": {
-      "version": "17.0.0",
-      "resolved": "https://registry.npmjs.org/stylelint-config-recommended/-/stylelint-config-recommended-17.0.0.tgz",
-      "integrity": "sha512-WaMSdEiPfZTSFVoYmJbxorJfA610O0tlYuU2aEwY33UQhSPgFbClrVJYWvy3jGJx+XW37O+LyNLiZOEXhKhJmA==",
+      "version": "18.0.0",
+      "resolved": "https://registry.npmjs.org/stylelint-config-recommended/-/stylelint-config-recommended-18.0.0.tgz",
+      "integrity": "sha512-mxgT2XY6YZ3HWWe3Di8umG6aBmWmHTblTgu/f10rqFXnyWxjKWwNdjSWkgkwCtxIKnqjSJzvFmPT5yabVIRxZg==",
       "dev": true,
       "funding": [
         {
@@ -5167,120 +5667,110 @@
       ],
       "license": "MIT",
       "engines": {
-        "node": ">=18.12.0"
+        "node": ">=20.19.0"
       },
       "peerDependencies": {
-        "stylelint": "^16.23.0"
+        "stylelint": "^17.0.0"
       }
     },
-    "node_modules/stylelint/node_modules/ansi-regex": {
-      "version": "5.0.1",
-      "resolved": "https://registry.npmjs.org/ansi-regex/-/ansi-regex-5.0.1.tgz",
-      "integrity": "sha512-quJQXlTSUGL2LH9SUXo8VwsY4soanhgo6LNSm84E1LBcE8s3O0wpdiRzyR9z/ZZJMlMWv37qOOb9pdJlMUEKFQ==",
+    "node_modules/stylelint/node_modules/ignore": {
+      "version": "7.0.14",
+      "resolved": "https://registry.npmjs.org/ignore/-/ignore-7.0.14.tgz",
+      "integrity": "sha512-wCBiZ77u1tuWde6G/nfOtLaCTYcFDU5yCzARzK10y+rLhzYoL7a7yEemcopcc2Mzow7gZOIjnV2ym1NYmdiJOQ==",
       "dev": true,
       "license": "MIT",
       "engines": {
-        "node": ">=8"
-      }
-    },
-    "node_modules/stylelint/node_modules/balanced-match": {
-      "version": "2.0.0",
-      "resolved": "https://registry.npmjs.org/balanced-match/-/balanced-match-2.0.0.tgz",
-      "integrity": "sha512-1ugUSr8BHXRnK23KfuYS+gVMC3LB8QGH9W1iGtDPsNWoQbgtXSExkBu2aDR4epiGWZOjZsj6lDl/N/AqqTC3UA==",
-      "dev": true,
-      "license": "MIT"
-    },
-    "node_modules/stylelint/node_modules/emoji-regex": {
-      "version": "8.0.0",
-      "resolved": "https://registry.npmjs.org/emoji-regex/-/emoji-regex-8.0.0.tgz",
-      "integrity": "sha512-MSjYzcWNOA0ewAHpz0MxpYFvwg6yjy1NG3xteoqz644VCo/RPgnr1/GGt+ic3iJTzQ8Eu3TdM14SawnVUmGE6A==",
-      "dev": true,
-      "license": "MIT"
-    },
-    "node_modules/stylelint/node_modules/file-entry-cache": {
-      "version": "10.1.3",
-      "resolved": "https://registry.npmjs.org/file-entry-cache/-/file-entry-cache-10.1.3.tgz",
-      "integrity": "sha512-D+w75Ub8T55yor7fPgN06rkCAUbAYw2vpxJmmjv/GDAcvCnv9g7IvHhIZoxzRZThrXPFI2maeY24pPbtyYU7Lg==",
-      "dev": true,
-      "license": "MIT",
-      "dependencies": {
-        "flat-cache": "^6.1.12"
+        "node": ">= 4"
       }
     },
-    "node_modules/stylelint/node_modules/flat-cache": {
-      "version": "6.1.12",
-      "resolved": "https://registry.npmjs.org/flat-cache/-/flat-cache-6.1.12.tgz",
-      "integrity": "sha512-U+HqqpZPPXP5d24bWuRzjGqVqUcw64k4nZAbruniDwdRg0H10tvN7H6ku1tjhA4rg5B9GS3siEvwO2qjJJ6f8Q==",
+    "node_modules/supports-color": {
+      "version": "10.2.2",
+      "resolved": "https://registry.npmjs.org/supports-color/-/supports-color-10.2.2.tgz",
+      "integrity": "sha512-SS+jx45GF1QjgEXQx4NJZV9ImqmO2NPz5FNsIHrsDjh2YsHnawpan7SNQ1o8NuhrbHZy9AZhIoCUiCeaW/C80g==",
       "dev": true,
       "license": "MIT",
-      "dependencies": {
-        "cacheable": "^1.10.3",
-        "flatted": "^3.3.3",
-        "hookified": "^1.10.0"
+      "engines": {
+        "node": ">=18"
+      },
+      "funding": {
+        "url": "https://github.com/chalk/supports-color?sponsor=1"
       }
     },
-    "node_modules/stylelint/node_modules/globby": {
-      "version": "11.1.0",
-      "resolved": "https://registry.npmjs.org/globby/-/globby-11.1.0.tgz",
-      "integrity": "sha512-jhIXaOzy1sb8IyocaruWSn1TjmnBVs8Ayhcy83rmxNJ8q2uWKCAj3CnJY+KpGSXCueAPc0i05kVvVKtP1t9S3g==",
+    "node_modules/supports-hyperlinks": {
+      "version": "4.6.0",
+      "resolved": "https://registry.npmjs.org/supports-hyperlinks/-/supports-hyperlinks-4.6.0.tgz",
+      "integrity": "sha512-FL/q/YePS4tFFTNiBOR+zPk2ZrKheqNtmTJ/9OUZ4e1wKHJofDbGQSmSsxS0XNnabSkqexryCaXGvEKz1vf9UQ==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
-        "array-union": "^2.1.0",
-        "dir-glob": "^3.0.1",
-        "fast-glob": "^3.2.9",
-        "ignore": "^5.2.0",
-        "merge2": "^1.4.1",
-        "slash": "^3.0.0"
+        "has-flag": "^5.0.1",
+        "supports-color": "^10.2.2"
       },
       "engines": {
-        "node": ">=10"
+        "node": ">=20"
       },
       "funding": {
-        "url": "https://github.com/sponsors/sindresorhus"
+        "url": "https://github.com/chalk/supports-hyperlinks?sponsor=1"
       }
     },
-    "node_modules/stylelint/node_modules/globby/node_modules/ignore": {
-      "version": "5.3.2",
-      "resolved": "https://registry.npmjs.org/ignore/-/ignore-5.3.2.tgz",
-      "integrity": "sha512-hsBTNUqQTDwkWtcdYI2i06Y/nUBEsNEDJKjWdigLvegy8kDuJAS8uRlpkkcQpyEXL0Z/pjDy5HBmMjRCJ2gq+g==",
+    "node_modules/svg-tags": {
+      "version": "1.0.0",
+      "resolved": "https://registry.npmjs.org/svg-tags/-/svg-tags-1.0.0.tgz",
+      "integrity": "sha512-ovssysQTa+luh7A5Weu3Rta6FJlFBBbInjOh722LIt6klpU2/HtdUbszju/G4devcvk8PGt7FCLv5wftu3THUA==",
+      "dev": true
+    },
+    "node_modules/table": {
+      "version": "6.9.0",
+      "resolved": "https://registry.npmjs.org/table/-/table-6.9.0.tgz",
+      "integrity": "sha512-9kY+CygyYM6j02t5YFHbNz2FN5QmYGv9zAjVp4lCDjlCw7amdckXlEt/bjMhUIfj4ThGRE4gCUH5+yGnNuPo5A==",
       "dev": true,
-      "license": "MIT",
+      "license": "BSD-3-Clause",
+      "dependencies": {
+        "ajv": "^8.0.1",
+        "lodash.truncate": "^4.4.2",
+        "slice-ansi": "^4.0.0",
+        "string-width": "^4.2.3",
+        "strip-ansi": "^6.0.1"
+      },
       "engines": {
-        "node": ">= 4"
+        "node": ">=10.0.0"
       }
     },
-    "node_modules/stylelint/node_modules/ignore": {
-      "version": "7.0.5",
-      "resolved": "https://registry.npmjs.org/ignore/-/ignore-7.0.5.tgz",
-      "integrity": "sha512-Hs59xBNfUIunMFgWAbGX5cq6893IbWg4KnrjbYwX3tx0ztorVgTDA6B2sxf8ejHJ4wz8BqGUMYlnzNBer5NvGg==",
+    "node_modules/table/node_modules/ajv": {
+      "version": "8.20.0",
+      "resolved": "https://registry.npmjs.org/ajv/-/ajv-8.20.0.tgz",
+      "integrity": "sha512-Thbli+OlOj+iMPYFBVBfJ3OmCAnaSyNn4M1vz9T6Gka5Jt9ba/HIR56joy65tY6kx/FCF5VXNB819Y7/GUrBGA==",
       "dev": true,
       "license": "MIT",
-      "engines": {
-        "node": ">= 4"
+      "dependencies": {
+        "fast-deep-equal": "^3.1.3",
+        "fast-uri": "^3.0.1",
+        "json-schema-traverse": "^1.0.0",
+        "require-from-string": "^2.0.2"
+      },
+      "funding": {
+        "type": "github",
+        "url": "https://github.com/sponsors/epoberezkin"
       }
     },
-    "node_modules/stylelint/node_modules/is-fullwidth-code-point": {
-      "version": "3.0.0",
-      "resolved": "https://registry.npmjs.org/is-fullwidth-code-point/-/is-fullwidth-code-point-3.0.0.tgz",
-      "integrity": "sha512-zymm5+u+sCsSWyD9qNaejV3DFvhCKclKdizYaJUuHA83RLjb7nSuGnddCHGv0hk+KY7BMAlsWeK4Ueg6EV6XQg==",
+    "node_modules/table/node_modules/ansi-regex": {
+      "version": "5.0.1",
+      "resolved": "https://registry.npmjs.org/ansi-regex/-/ansi-regex-5.0.1.tgz",
+      "integrity": "sha512-quJQXlTSUGL2LH9SUXo8VwsY4soanhgo6LNSm84E1LBcE8s3O0wpdiRzyR9z/ZZJMlMWv37qOOb9pdJlMUEKFQ==",
       "dev": true,
       "license": "MIT",
       "engines": {
         "node": ">=8"
       }
     },
-    "node_modules/stylelint/node_modules/slash": {
-      "version": "3.0.0",
-      "resolved": "https://registry.npmjs.org/slash/-/slash-3.0.0.tgz",
-      "integrity": "sha512-g9Q1haeby36OSStwb4ntCGGGaKsaVSjQ68fBxoQcutl5fS1vuY18H3wSt3jFyFtrkx+Kz0V1G85A4MyAdDMi2Q==",
+    "node_modules/table/node_modules/json-schema-traverse": {
+      "version": "1.0.0",
+      "resolved": "https://registry.npmjs.org/json-schema-traverse/-/json-schema-traverse-1.0.0.tgz",
+      "integrity": "sha512-NM8/P9n3XjXhIZn1lLhkFaACTOURQXjWhV4BA/RnOv8xvgqtqpAX9IO4mRQxSx1Rlo4tqzeqb0sOlruaOy3dug==",
       "dev": true,
-      "license": "MIT",
-      "engines": {
-        "node": ">=8"
-      }
+      "license": "MIT"
     },
-    "node_modules/stylelint/node_modules/string-width": {
+    "node_modules/table/node_modules/string-width": {
       "version": "4.2.3",
       "resolved": "https://registry.npmjs.org/string-width/-/string-width-4.2.3.tgz",
       "integrity": "sha512-wKyQRQpjJ0sIp62ErSZdGsjMJWsap5oRNihHhu6G7JVO/9jIB6UyevL+tXuOqrng8j/cxKTWyWUwvSTriiZz/g==",
@@ -5295,7 +5785,7 @@
         "node": ">=8"
       }
     },
-    "node_modules/stylelint/node_modules/strip-ansi": {
+    "node_modules/table/node_modules/strip-ansi": {
       "version": "6.0.1",
       "resolved": "https://registry.npmjs.org/strip-ansi/-/strip-ansi-6.0.1.tgz",
       "integrity": "sha512-Y38VPSHcqkFrCpFnQ9vuSXmquuv5oXOKpGeT6aGrr3o3Gc9AlVa6JBfUSOCnbxGGZF+/0ooI7KrPuUSztUdU5A==",
@@ -5308,226 +5798,81 @@
         "node": ">=8"
       }
     },
-    "node_modules/stylelint/node_modules/write-file-atomic": {
-      "version": "5.0.1",
-      "resolved": "https://registry.npmjs.org/write-file-atomic/-/write-file-atomic-5.0.1.tgz",
-      "integrity": "sha512-+QU2zd6OTD8XWIJCbffaiQeH9U73qIqafo1x6V1snCWYGJf6cVE0cDR4D8xRzcEnfI21IFrUPzPGtcPf8AC+Rw==",
+    "node_modules/temp-dir": {
+      "version": "3.0.0",
+      "resolved": "https://registry.npmjs.org/temp-dir/-/temp-dir-3.0.0.tgz",
+      "integrity": "sha512-nHc6S/bwIilKHNRgK/3jlhDoIHcp45YgyiwcAk46Tr0LfEqGBVpmiAyuiuxeVE44m3mXnEeVhaipLOEWmH+Njw==",
       "dev": true,
-      "license": "ISC",
-      "dependencies": {
-        "imurmurhash": "^0.1.4",
-        "signal-exit": "^4.0.1"
-      },
+      "license": "MIT",
       "engines": {
-        "node": "^14.17.0 || ^16.13.0 || >=18.0.0"
+        "node": ">=14.16"
       }
     },
-    "node_modules/supports-color": {
-      "version": "7.2.0",
-      "resolved": "https://registry.npmjs.org/supports-color/-/supports-color-7.2.0.tgz",
-      "integrity": "sha512-qpCAvRl9stuOHveKsn7HncJRvv501qIacKzQlO/+Lwxc9+0q2wLyv4Dfvt80/DPn2pqOBsJdDiogXGR9+OvwRw==",
-      "dev": true,
-      "license": "MIT",
-      "dependencies": {
-        "has-flag": "^4.0.0"
-      },
-      "engines": {
-        "node": ">=8"
-      }
-    },
-    "node_modules/supports-hyperlinks": {
+    "node_modules/tempy": {
       "version": "3.2.0",
-      "resolved": "https://registry.npmjs.org/supports-hyperlinks/-/supports-hyperlinks-3.2.0.tgz",
-      "integrity": "sha512-zFObLMyZeEwzAoKCyu1B91U79K2t7ApXuQfo8OuxwXLDgcKxuwM+YvcbIhm6QWqz7mHUH1TVytR1PwVVjEuMig==",
+      "resolved": "https://registry.npmjs.org/tempy/-/tempy-3.2.0.tgz",
+      "integrity": "sha512-d79HhZya5Djd7am0q+W4RTsSU+D/aJzM+4Y4AGJGuGlgM2L6sx5ZvOYTmZjqPhrDrV6xJTtRSm1JCLj6V6LHLQ==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
-        "has-flag": "^4.0.0",
-        "supports-color": "^7.0.0"
-      },
-      "engines": {
-        "node": ">=14.18"
-      },
-      "funding": {
-        "url": "https://github.com/chalk/supports-hyperlinks?sponsor=1"
-      }
-    },
-    "node_modules/svg-tags": {
-      "version": "1.0.0",
-      "resolved": "https://registry.npmjs.org/svg-tags/-/svg-tags-1.0.0.tgz",
-      "integrity": "sha512-ovssysQTa+luh7A5Weu3Rta6FJlFBBbInjOh722LIt6klpU2/HtdUbszju/G4devcvk8PGt7FCLv5wftu3THUA==",
-      "dev": true
-    },
-    "node_modules/table": {
-      "version": "6.9.0",
-      "resolved": "https://registry.npmjs.org/table/-/table-6.9.0.tgz",
-      "integrity": "sha512-9kY+CygyYM6j02t5YFHbNz2FN5QmYGv9zAjVp4lCDjlCw7amdckXlEt/bjMhUIfj4ThGRE4gCUH5+yGnNuPo5A==",
-      "dev": true,
-      "license": "BSD-3-Clause",
-      "dependencies": {
-        "ajv": "^8.0.1",
-        "lodash.truncate": "^4.4.2",
-        "slice-ansi": "^4.0.0",
-        "string-width": "^4.2.3",
-        "strip-ansi": "^6.0.1"
+        "is-stream": "^3.0.0",
+        "temp-dir": "^3.0.0",
+        "type-fest": "^2.12.2",
+        "unique-string": "^3.0.0"
       },
       "engines": {
-        "node": ">=10.0.0"
-      }
-    },
-    "node_modules/table/node_modules/ajv": {
-      "version": "8.18.0",
-      "resolved": "https://registry.npmjs.org/ajv/-/ajv-8.18.0.tgz",
-      "integrity": "sha512-PlXPeEWMXMZ7sPYOHqmDyCJzcfNrUr3fGNKtezX14ykXOEIvyK81d+qydx89KY5O71FKMPaQ2vBfBFI5NHR63A==",
-      "dev": true,
-      "license": "MIT",
-      "dependencies": {
-        "fast-deep-equal": "^3.1.3",
-        "fast-uri": "^3.0.1",
-        "json-schema-traverse": "^1.0.0",
-        "require-from-string": "^2.0.2"
+        "node": ">=14.16"
       },
       "funding": {
-        "type": "github",
-        "url": "https://github.com/sponsors/epoberezkin"
-      }
-    },
-    "node_modules/table/node_modules/ansi-regex": {
-      "version": "5.0.1",
-      "resolved": "https://registry.npmjs.org/ansi-regex/-/ansi-regex-5.0.1.tgz",
-      "integrity": "sha512-quJQXlTSUGL2LH9SUXo8VwsY4soanhgo6LNSm84E1LBcE8s3O0wpdiRzyR9z/ZZJMlMWv37qOOb9pdJlMUEKFQ==",
-      "dev": true,
-      "license": "MIT",
-      "engines": {
-        "node": ">=8"
+        "url": "https://github.com/sponsors/sindresorhus"
       }
     },
-    "node_modules/table/node_modules/ansi-styles": {
-      "version": "4.3.0",
-      "resolved": "https://registry.npmjs.org/ansi-styles/-/ansi-styles-4.3.0.tgz",
-      "integrity": "sha512-zbB9rCJAT1rbjiVDb2hqKFHNYLxgtk8NURxZ3IZwD3F6NtxbXZQCnnSi1Lkx+IDohdPlFp222wVALIheZJQSEg==",
+    "node_modules/tinyglobby": {
+      "version": "0.2.17",
+      "resolved": "https://registry.npmjs.org/tinyglobby/-/tinyglobby-0.2.17.tgz",
+      "integrity": "sha512-wXR/dYpcqKmfWpEdZjiKJOwCNFndD0DMnrW/cYjVGttEkBfVgcLFHoNrlj47mjOVic9yyNu65alsgF4NQyTa2g==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
-        "color-convert": "^2.0.1"
+        "fdir": "^6.5.0",
+        "picomatch": "^4.0.4"
       },
       "engines": {
-        "node": ">=8"
+        "node": ">=12.0.0"
       },
       "funding": {
-        "url": "https://github.com/chalk/ansi-styles?sponsor=1"
-      }
-    },
-    "node_modules/table/node_modules/emoji-regex": {
-      "version": "8.0.0",
-      "resolved": "https://registry.npmjs.org/emoji-regex/-/emoji-regex-8.0.0.tgz",
-      "integrity": "sha512-MSjYzcWNOA0ewAHpz0MxpYFvwg6yjy1NG3xteoqz644VCo/RPgnr1/GGt+ic3iJTzQ8Eu3TdM14SawnVUmGE6A==",
-      "dev": true,
-      "license": "MIT"
-    },
-    "node_modules/table/node_modules/is-fullwidth-code-point": {
-      "version": "3.0.0",
-      "resolved": "https://registry.npmjs.org/is-fullwidth-code-point/-/is-fullwidth-code-point-3.0.0.tgz",
-      "integrity": "sha512-zymm5+u+sCsSWyD9qNaejV3DFvhCKclKdizYaJUuHA83RLjb7nSuGnddCHGv0hk+KY7BMAlsWeK4Ueg6EV6XQg==",
-      "dev": true,
-      "license": "MIT",
-      "engines": {
-        "node": ">=8"
+        "url": "https://github.com/sponsors/SuperchupuDev"
       }
     },
-    "node_modules/table/node_modules/json-schema-traverse": {
-      "version": "1.0.0",
-      "resolved": "https://registry.npmjs.org/json-schema-traverse/-/json-schema-traverse-1.0.0.tgz",
-      "integrity": "sha512-NM8/P9n3XjXhIZn1lLhkFaACTOURQXjWhV4BA/RnOv8xvgqtqpAX9IO4mRQxSx1Rlo4tqzeqb0sOlruaOy3dug==",
-      "dev": true,
-      "license": "MIT"
-    },
-    "node_modules/table/node_modules/slice-ansi": {
-      "version": "4.0.0",
-      "resolved": "https://registry.npmjs.org/slice-ansi/-/slice-ansi-4.0.0.tgz",
-      "integrity": "sha512-qMCMfhY040cVHT43K9BFygqYbUPFZKHOg7K73mtTWJRb8pyP3fzf4Ixd5SzdEJQ6MRUg/WBnOLxghZtKKurENQ==",
+    "node_modules/tinyglobby/node_modules/fdir": {
+      "version": "6.5.0",
+      "resolved": "https://registry.npmjs.org/fdir/-/fdir-6.5.0.tgz",
+      "integrity": "sha512-tIbYtZbucOs0BRGqPJkshJUYdL+SDH7dVM8gjy+ERp3WAUjLEFJE+02kanyHtwjWOnwrKYBiwAmM0p4kLJAnXg==",
       "dev": true,
       "license": "MIT",
-      "dependencies": {
-        "ansi-styles": "^4.0.0",
-        "astral-regex": "^2.0.0",
-        "is-fullwidth-code-point": "^3.0.0"
-      },
       "engines": {
-        "node": ">=10"
-      },
-      "funding": {
-        "url": "https://github.com/chalk/slice-ansi?sponsor=1"
-      }
-    },
-    "node_modules/table/node_modules/string-width": {
-      "version": "4.2.3",
-      "resolved": "https://registry.npmjs.org/string-width/-/string-width-4.2.3.tgz",
-      "integrity": "sha512-wKyQRQpjJ0sIp62ErSZdGsjMJWsap5oRNihHhu6G7JVO/9jIB6UyevL+tXuOqrng8j/cxKTWyWUwvSTriiZz/g==",
-      "dev": true,
-      "license": "MIT",
-      "dependencies": {
-        "emoji-regex": "^8.0.0",
-        "is-fullwidth-code-point": "^3.0.0",
-        "strip-ansi": "^6.0.1"
+        "node": ">=12.0.0"
       },
-      "engines": {
-        "node": ">=8"
-      }
-    },
-    "node_modules/table/node_modules/strip-ansi": {
-      "version": "6.0.1",
-      "resolved": "https://registry.npmjs.org/strip-ansi/-/strip-ansi-6.0.1.tgz",
-      "integrity": "sha512-Y38VPSHcqkFrCpFnQ9vuSXmquuv5oXOKpGeT6aGrr3o3Gc9AlVa6JBfUSOCnbxGGZF+/0ooI7KrPuUSztUdU5A==",
-      "dev": true,
-      "license": "MIT",
-      "dependencies": {
-        "ansi-regex": "^5.0.1"
+      "peerDependencies": {
+        "picomatch": "^3 || ^4"
       },
-      "engines": {
-        "node": ">=8"
-      }
-    },
-    "node_modules/temp-dir": {
-      "version": "3.0.0",
-      "resolved": "https://registry.npmjs.org/temp-dir/-/temp-dir-3.0.0.tgz",
-      "integrity": "sha512-nHc6S/bwIilKHNRgK/3jlhDoIHcp45YgyiwcAk46Tr0LfEqGBVpmiAyuiuxeVE44m3mXnEeVhaipLOEWmH+Njw==",
-      "dev": true,
-      "license": "MIT",
-      "engines": {
-        "node": ">=14.16"
+      "peerDependenciesMeta": {
+        "picomatch": {
+          "optional": true
+        }
       }
     },
-    "node_modules/tempy": {
-      "version": "3.1.0",
-      "resolved": "https://registry.npmjs.org/tempy/-/tempy-3.1.0.tgz",
-      "integrity": "sha512-7jDLIdD2Zp0bDe5r3D2qtkd1QOCacylBuL7oa4udvN6v2pqr4+LcCr67C8DR1zkpaZ8XosF5m1yQSabKAW6f2g==",
+    "node_modules/tinyglobby/node_modules/picomatch": {
+      "version": "4.0.7",
+      "resolved": "https://registry.npmjs.org/picomatch/-/picomatch-4.0.7.tgz",
+      "integrity": "sha512-qcJu88Q2IWqJsDD529JKMdwGm/dvInW4HvQnRwiH9JtihJvzGOscDtHE3x1pBKeUOTysQ8kVmLnJ2kJu7yhcGA==",
       "dev": true,
       "license": "MIT",
-      "dependencies": {
-        "is-stream": "^3.0.0",
-        "temp-dir": "^3.0.0",
-        "type-fest": "^2.12.2",
-        "unique-string": "^3.0.0"
-      },
-      "engines": {
-        "node": ">=14.16"
-      },
-      "funding": {
-        "url": "https://github.com/sponsors/sindresorhus"
-      }
-    },
-    "node_modules/tempy/node_modules/type-fest": {
-      "version": "2.19.0",
-      "resolved": "https://registry.npmjs.org/type-fest/-/type-fest-2.19.0.tgz",
-      "integrity": "sha512-RAH822pAdBgcNMAfWnCBU3CFZcfZ/i1eZjwFU/dsLKumyuuP3niueg2UAukXYF0E2AAoc82ZSSf9J0WQBinzHA==",
-      "dev": true,
-      "license": "(MIT OR CC0-1.0)",
       "engines": {
-        "node": ">=12.20"
+        "node": ">=12"
       },
       "funding": {
-        "url": "https://github.com/sponsors/sindresorhus"
+        "url": "https://github.com/sponsors/jonschlinkert"
       }
     },
     "node_modules/to-regex-range": {
@@ -5564,9 +5909,9 @@
       }
     },
     "node_modules/ts-api-utils": {
-      "version": "2.1.0",
-      "resolved": "https://registry.npmjs.org/ts-api-utils/-/ts-api-utils-2.1.0.tgz",
-      "integrity": "sha512-CUgTZL1irw8u29bzrOD/nH85jqyc74D6SshFgujOIA7osm2Rz7dYH77agkx7H4FBNxDq7Cjf+IjaX/8zwFW+ZQ==",
+      "version": "2.5.0",
+      "resolved": "https://registry.npmjs.org/ts-api-utils/-/ts-api-utils-2.5.0.tgz",
+      "integrity": "sha512-OJ/ibxhPlqrMM0UiNHJ/0CKQkoKF243/AEmplt3qpRgkW8VG7IfOS41h7V8TjITqdByHzrjcS/2si+y4lIh8NA==",
       "dev": true,
       "license": "MIT",
       "engines": {
@@ -5584,505 +5929,21 @@
     },
     "node_modules/tsx": {
       "version": "4.23.15",
-      "resolved": "https://registry.npmjs.org/tsx/-/tsx-4.23.15.tgz",
-      "integrity": "sha512-Yiex1Ovn8z2xPpOWckIiysV1SSyRMY9BkLF++q0yKiDxCqRhosKfMg3janKkiLBwZ5c/YryloKwGZcrEmtwxKw==",
-      "dev": true,
-      "license": "MIT",
-      "dependencies": {
-        "esbuild": "~0.28.0"
-      },
-      "bin": {
-        "tsx": "dist/cli.mjs"
-      },
-      "engines": {
-        "node": ">=18.0.0"
-      },
-      "optionalDependencies": {
-        "fsevents": "~2.3.3"
-      }
-    },
-    "node_modules/tsx/node_modules/@esbuild/aix-ppc64": {
-      "version": "0.28.2",
-      "resolved": "https://registry.npmjs.org/@esbuild/aix-ppc64/-/aix-ppc64-0.28.2.tgz",
-      "integrity": "sha512-XExcO+dvLKvVtNTibSTBej1NCAbaGhWn9Ww1ZPx80qsahhPFe/8jgWP0IchNe0F3HwkU7n8ejhH8bjonqht8mQ==",
-      "cpu": [
-        "ppc64"
-      ],
-      "dev": true,
-      "license": "MIT",
-      "optional": true,
-      "os": [
-        "aix"
-      ],
-      "engines": {
-        "node": ">=18"
-      }
-    },
-    "node_modules/tsx/node_modules/@esbuild/android-arm": {
-      "version": "0.28.2",
-      "resolved": "https://registry.npmjs.org/@esbuild/android-arm/-/android-arm-0.28.2.tgz",
-      "integrity": "sha512-kXXoiPVVGQcnIYGOeaovwOURpniDBpSq4A03qkQ+BMQqtGG6HYap3xne9C1O1yo4TR3qxlCX5IqqmX6fFo2Lqg==",
-      "cpu": [
-        "arm"
-      ],
-      "dev": true,
-      "license": "MIT",
-      "optional": true,
-      "os": [
-        "android"
-      ],
-      "engines": {
-        "node": ">=18"
-      }
-    },
-    "node_modules/tsx/node_modules/@esbuild/android-arm64": {
-      "version": "0.28.2",
-      "resolved": "https://registry.npmjs.org/@esbuild/android-arm64/-/android-arm64-0.28.2.tgz",
-      "integrity": "sha512-5YfKeeI8qWfBZIX+u2xZC3Zlb3Os/gLS2sbEKM+I4ZOcsWmHS2WLysCcQZDAFRslDUU5Oiq44gf6PYN1vGwG5A==",
-      "cpu": [
-        "arm64"
-      ],
-      "dev": true,
-      "license": "MIT",
-      "optional": true,
-      "os": [
-        "android"
-      ],
-      "engines": {
-        "node": ">=18"
-      }
-    },
-    "node_modules/tsx/node_modules/@esbuild/android-x64": {
-      "version": "0.28.2",
-      "resolved": "https://registry.npmjs.org/@esbuild/android-x64/-/android-x64-0.28.2.tgz",
-      "integrity": "sha512-O387ite7SzUyCcy3JQX4P4bLtEA7bLLkx+esve5JHnyYfNTxcVpXZo9jhdB0lTKN44gztELTdU7nS8Nr16Fs1Q==",
-      "cpu": [
-        "x64"
-      ],
-      "dev": true,
-      "license": "MIT",
-      "optional": true,
-      "os": [
-        "android"
-      ],
-      "engines": {
-        "node": ">=18"
-      }
-    },
-    "node_modules/tsx/node_modules/@esbuild/darwin-arm64": {
-      "version": "0.28.2",
-      "resolved": "https://registry.npmjs.org/@esbuild/darwin-arm64/-/darwin-arm64-0.28.2.tgz",
-      "integrity": "sha512-n4KqkOQrraxHJcgjM1RvwbigfQKIKJVpM7xp+KsxiyUSrRdIXnt73VhrPAx0fV44hgfmIVKjxMN9J1t5jySVkw==",
-      "cpu": [
-        "arm64"
-      ],
-      "dev": true,
-      "license": "MIT",
-      "optional": true,
-      "os": [
-        "darwin"
-      ],
-      "engines": {
-        "node": ">=18"
-      }
-    },
-    "node_modules/tsx/node_modules/@esbuild/darwin-x64": {
-      "version": "0.28.2",
-      "resolved": "https://registry.npmjs.org/@esbuild/darwin-x64/-/darwin-x64-0.28.2.tgz",
-      "integrity": "sha512-uq6suIWYP37qzGddBKPw5QEQPi6HiLGsO7UmkpfyaYNQ3D+rN6w6WfwH+nuqcGXWvawGwxOEroO4YGnFh95azw==",
-      "cpu": [
-        "x64"
-      ],
-      "dev": true,
-      "license": "MIT",
-      "optional": true,
-      "os": [
-        "darwin"
-      ],
-      "engines": {
-        "node": ">=18"
-      }
-    },
-    "node_modules/tsx/node_modules/@esbuild/freebsd-arm64": {
-      "version": "0.28.2",
-      "resolved": "https://registry.npmjs.org/@esbuild/freebsd-arm64/-/freebsd-arm64-0.28.2.tgz",
-      "integrity": "sha512-n+I0BTSRIoy+d6RPKnEVwql5UwBJolytvY4mAOIEJorKlqgPII8ix6slVVrfZ5Tnj7glIZvloylbB/EJPMWEXw==",
-      "cpu": [
-        "arm64"
-      ],
-      "dev": true,
-      "license": "MIT",
-      "optional": true,
-      "os": [
-        "freebsd"
-      ],
-      "engines": {
-        "node": ">=18"
-      }
-    },
-    "node_modules/tsx/node_modules/@esbuild/freebsd-x64": {
-      "version": "0.28.2",
-      "resolved": "https://registry.npmjs.org/@esbuild/freebsd-x64/-/freebsd-x64-0.28.2.tgz",
-      "integrity": "sha512-78XJTJkvPs0kz2w61301PJjXl4g7q3JqiYMZ/M/yVI73EHBrCRTgkhu9oqG7vPqq+a/yadEW8aD+agKlk5xrmg==",
-      "cpu": [
-        "x64"
-      ],
-      "dev": true,
-      "license": "MIT",
-      "optional": true,
-      "os": [
-        "freebsd"
-      ],
-      "engines": {
-        "node": ">=18"
-      }
-    },
-    "node_modules/tsx/node_modules/@esbuild/linux-arm": {
-      "version": "0.28.2",
-      "resolved": "https://registry.npmjs.org/@esbuild/linux-arm/-/linux-arm-0.28.2.tgz",
-      "integrity": "sha512-XlDnu2q5yoqems+xay6wSAcg9DDD7K9RLKZEBOMZm3ckNpJBvOX20tSfby8KfrrhINDyv9V2YVZKY/SpoGJI8w==",
-      "cpu": [
-        "arm"
-      ],
-      "dev": true,
-      "license": "MIT",
-      "optional": true,
-      "os": [
-        "linux"
-      ],
-      "engines": {
-        "node": ">=18"
-      }
-    },
-    "node_modules/tsx/node_modules/@esbuild/linux-arm64": {
-      "version": "0.28.2",
-      "resolved": "https://registry.npmjs.org/@esbuild/linux-arm64/-/linux-arm64-0.28.2.tgz",
-      "integrity": "sha512-pW4AC0P3it8c7do9MVM4p51FzHzdM/TZrerurgRcHJ2WTa1VQ1CIq18xncfpBJw4ojkiZZrKW2yIBWBP92j6Ug==",
-      "cpu": [
-        "arm64"
-      ],
-      "dev": true,
-      "license": "MIT",
-      "optional": true,
-      "os": [
-        "linux"
-      ],
-      "engines": {
-        "node": ">=18"
-      }
-    },
-    "node_modules/tsx/node_modules/@esbuild/linux-ia32": {
-      "version": "0.28.2",
-      "resolved": "https://registry.npmjs.org/@esbuild/linux-ia32/-/linux-ia32-0.28.2.tgz",
-      "integrity": "sha512-CYbnj78HsIeA+DhgUKgFCfvNsTHFhMMrinUrMZpDXJXKN8T3XViTZ/+wtHeVxEWY8ewSzTFN+nRmSwO2tZaLUQ==",
-      "cpu": [
-        "ia32"
-      ],
-      "dev": true,
-      "license": "MIT",
-      "optional": true,
-      "os": [
-        "linux"
-      ],
-      "engines": {
-        "node": ">=18"
-      }
-    },
-    "node_modules/tsx/node_modules/@esbuild/linux-loong64": {
-      "version": "0.28.2",
-      "resolved": "https://registry.npmjs.org/@esbuild/linux-loong64/-/linux-loong64-0.28.2.tgz",
-      "integrity": "sha512-buwkd8nsph4R+ajRvw0qM5Hja/TXQow3ptzWO2EbG/cqcIkHloRrdlBtQlshyYGTNFvfkfJ5tpPLVkY4DtsPfQ==",
-      "cpu": [
-        "loong64"
-      ],
-      "dev": true,
-      "license": "MIT",
-      "optional": true,
-      "os": [
-        "linux"
-      ],
-      "engines": {
-        "node": ">=18"
-      }
-    },
-    "node_modules/tsx/node_modules/@esbuild/linux-mips64el": {
-      "version": "0.28.2",
-      "resolved": "https://registry.npmjs.org/@esbuild/linux-mips64el/-/linux-mips64el-0.28.2.tgz",
-      "integrity": "sha512-ZVykbDyk7519VwiNb9Lcj9m8XM6v5V9uKPvrEMkkEedVewf+0itkhahp4HDpgERXhwLRpWFypsGbG/J8s0QjJA==",
-      "cpu": [
-        "mips64el"
-      ],
-      "dev": true,
-      "license": "MIT",
-      "optional": true,
-      "os": [
-        "linux"
-      ],
-      "engines": {
-        "node": ">=18"
-      }
-    },
-    "node_modules/tsx/node_modules/@esbuild/linux-ppc64": {
-      "version": "0.28.2",
-      "resolved": "https://registry.npmjs.org/@esbuild/linux-ppc64/-/linux-ppc64-0.28.2.tgz",
-      "integrity": "sha512-CAXl+Dtd9UUuJd8pKKdwh6MLm3MUMiqMPmhZ3tTSXPqfyQ3vDl6R5hZdZ/kYojK4ofXtdfSv1tFq8XzWx3heNQ==",
-      "cpu": [
-        "ppc64"
-      ],
-      "dev": true,
-      "license": "MIT",
-      "optional": true,
-      "os": [
-        "linux"
-      ],
-      "engines": {
-        "node": ">=18"
-      }
-    },
-    "node_modules/tsx/node_modules/@esbuild/linux-riscv64": {
-      "version": "0.28.2",
-      "resolved": "https://registry.npmjs.org/@esbuild/linux-riscv64/-/linux-riscv64-0.28.2.tgz",
-      "integrity": "sha512-GeXCej4IQtU1B+QlDV8W/RRvbzI3O/Stss+/bCXv4lZls5WGRtu2a+3JkA3i4qIUlMXpcHebWpF8AkJhATowuA==",
-      "cpu": [
-        "riscv64"
-      ],
-      "dev": true,
-      "license": "MIT",
-      "optional": true,
-      "os": [
-        "linux"
-      ],
-      "engines": {
-        "node": ">=18"
-      }
-    },
-    "node_modules/tsx/node_modules/@esbuild/linux-s390x": {
-      "version": "0.28.2",
-      "resolved": "https://registry.npmjs.org/@esbuild/linux-s390x/-/linux-s390x-0.28.2.tgz",
-      "integrity": "sha512-3H1weTYZPxt/WOhByszQZybS9w5lKzUn1FDMsgEChbHWQwHYQQRfBxgCcZvPhjHfKyJjIievvMmEUawJrdY9Dg==",
-      "cpu": [
-        "s390x"
-      ],
-      "dev": true,
-      "license": "MIT",
-      "optional": true,
-      "os": [
-        "linux"
-      ],
-      "engines": {
-        "node": ">=18"
-      }
-    },
-    "node_modules/tsx/node_modules/@esbuild/linux-x64": {
-      "version": "0.28.2",
-      "resolved": "https://registry.npmjs.org/@esbuild/linux-x64/-/linux-x64-0.28.2.tgz",
-      "integrity": "sha512-4xTZr1FUmSoQW4XIWmit3tzQrUTZM+N3P0XV8xROKYF50XfI7xeO90+1bZvNwxIufQ9hDQVRJH5YhgPVF8A/HQ==",
-      "cpu": [
-        "x64"
-      ],
-      "dev": true,
-      "license": "MIT",
-      "optional": true,
-      "os": [
-        "linux"
-      ],
-      "engines": {
-        "node": ">=18"
-      }
-    },
-    "node_modules/tsx/node_modules/@esbuild/netbsd-arm64": {
-      "version": "0.28.2",
-      "resolved": "https://registry.npmjs.org/@esbuild/netbsd-arm64/-/netbsd-arm64-0.28.2.tgz",
-      "integrity": "sha512-sSATRjPeDBg3pdgHoQfoYBob11Kk1FGa9lui5RIHZCoCkJa9QKlvl3/vKz2usCmYYjs7ymJR/2Nnsqe+Hjt5nw==",
-      "cpu": [
-        "arm64"
-      ],
-      "dev": true,
-      "license": "MIT",
-      "optional": true,
-      "os": [
-        "netbsd"
-      ],
-      "engines": {
-        "node": ">=18"
-      }
-    },
-    "node_modules/tsx/node_modules/@esbuild/netbsd-x64": {
-      "version": "0.28.2",
-      "resolved": "https://registry.npmjs.org/@esbuild/netbsd-x64/-/netbsd-x64-0.28.2.tgz",
-      "integrity": "sha512-lqnzCV+mM0gIADaKihiCg6ifgfU2L3h5E33rNQBN1Y4MaVGnzryzmvvf7UHxprpQdE8hpqLolJ9Rl+SkIRDpyw==",
-      "cpu": [
-        "x64"
-      ],
-      "dev": true,
-      "license": "MIT",
-      "optional": true,
-      "os": [
-        "netbsd"
-      ],
-      "engines": {
-        "node": ">=18"
-      }
-    },
-    "node_modules/tsx/node_modules/@esbuild/openbsd-arm64": {
-      "version": "0.28.2",
-      "resolved": "https://registry.npmjs.org/@esbuild/openbsd-arm64/-/openbsd-arm64-0.28.2.tgz",
-      "integrity": "sha512-AL2qJILH7lNjrDmCQDvdxMfAUIv8KMNZOvrwAQ8i8//ntL9FflhOyMJ8OZSMBb8/AWXe3/5v5S20y3zCoZWKoQ==",
-      "cpu": [
-        "arm64"
-      ],
-      "dev": true,
-      "license": "MIT",
-      "optional": true,
-      "os": [
-        "openbsd"
-      ],
-      "engines": {
-        "node": ">=18"
-      }
-    },
-    "node_modules/tsx/node_modules/@esbuild/openbsd-x64": {
-      "version": "0.28.2",
-      "resolved": "https://registry.npmjs.org/@esbuild/openbsd-x64/-/openbsd-x64-0.28.2.tgz",
-      "integrity": "sha512-QtiuPytchRyC4rwUKhexJdQKvDuZ6hWloi3igqPQNUJCS1/v9EiO3UTOXR6A3FoMo4fnAKbWJdqaIwhOzh8qEw==",
-      "cpu": [
-        "x64"
-      ],
-      "dev": true,
-      "license": "MIT",
-      "optional": true,
-      "os": [
-        "openbsd"
-      ],
-      "engines": {
-        "node": ">=18"
-      }
-    },
-    "node_modules/tsx/node_modules/@esbuild/openharmony-arm64": {
-      "version": "0.28.2",
-      "resolved": "https://registry.npmjs.org/@esbuild/openharmony-arm64/-/openharmony-arm64-0.28.2.tgz",
-      "integrity": "sha512-WkhYDmpTjLvGlScA1rwjRUmhl4k8oXR3cIbtqWmELgU/dFeHHlEllxDvdWcNJV9rbzCexB5vz8gtNewWLgCT7Q==",
-      "cpu": [
-        "arm64"
-      ],
-      "dev": true,
-      "license": "MIT",
-      "optional": true,
-      "os": [
-        "openharmony"
-      ],
-      "engines": {
-        "node": ">=18"
-      }
-    },
-    "node_modules/tsx/node_modules/@esbuild/sunos-x64": {
-      "version": "0.28.2",
-      "resolved": "https://registry.npmjs.org/@esbuild/sunos-x64/-/sunos-x64-0.28.2.tgz",
-      "integrity": "sha512-GPMSkTOtMnv2U2F8gxe4Io6qmVs+YKyp832Etqqxr0hFngmXQ3rzwytelm3GIn7T4VviRUlf3sOgBOiTdvaf7g==",
-      "cpu": [
-        "x64"
-      ],
-      "dev": true,
-      "license": "MIT",
-      "optional": true,
-      "os": [
-        "sunos"
-      ],
-      "engines": {
-        "node": ">=18"
-      }
-    },
-    "node_modules/tsx/node_modules/@esbuild/win32-arm64": {
-      "version": "0.28.2",
-      "resolved": "https://registry.npmjs.org/@esbuild/win32-arm64/-/win32-arm64-0.28.2.tgz",
-      "integrity": "sha512-PIhhEkE9uPBleRBrQEJpUn7MBnibZzbGzYWPmY3x+YoVg/95zbjB4CxPPOQ8l5tYYM4mMaCthF8/1DIfBQQyWQ==",
-      "cpu": [
-        "arm64"
-      ],
-      "dev": true,
-      "license": "MIT",
-      "optional": true,
-      "os": [
-        "win32"
-      ],
-      "engines": {
-        "node": ">=18"
-      }
-    },
-    "node_modules/tsx/node_modules/@esbuild/win32-ia32": {
-      "version": "0.28.2",
-      "resolved": "https://registry.npmjs.org/@esbuild/win32-ia32/-/win32-ia32-0.28.2.tgz",
-      "integrity": "sha512-YmJbfTlvU7Sdn9BB+4PRES4oB6pxgS37MAONj+hBr/cpXS1aBPKXxNnDbu+QCWPj0o9dgyxeq79g6c5P8KeuYA==",
-      "cpu": [
-        "ia32"
-      ],
-      "dev": true,
-      "license": "MIT",
-      "optional": true,
-      "os": [
-        "win32"
-      ],
-      "engines": {
-        "node": ">=18"
-      }
-    },
-    "node_modules/tsx/node_modules/@esbuild/win32-x64": {
-      "version": "0.28.2",
-      "resolved": "https://registry.npmjs.org/@esbuild/win32-x64/-/win32-x64-0.28.2.tgz",
-      "integrity": "sha512-5ebpxr3nWMzrL/rnUI755Jkuee0bHL/Gq0WTF9lvcpv73wAp5eu8MfBUgWK9bhWvZjj7yX8etf/8tI8Ney695g==",
-      "cpu": [
-        "x64"
-      ],
-      "dev": true,
-      "license": "MIT",
-      "optional": true,
-      "os": [
-        "win32"
-      ],
-      "engines": {
-        "node": ">=18"
-      }
-    },
-    "node_modules/tsx/node_modules/esbuild": {
-      "version": "0.28.2",
-      "resolved": "https://registry.npmjs.org/esbuild/-/esbuild-0.28.2.tgz",
-      "integrity": "sha512-HKVLS8dvII+xoKW9kmqxbRKrnWEXfJJr/FZhhJmiqIB0e053QNYFqOBouTMO/k5sID4MvCiUCvv8b9M4h32wIA==",
+      "resolved": "https://registry.npmjs.org/tsx/-/tsx-4.23.15.tgz",
+      "integrity": "sha512-Yiex1Ovn8z2xPpOWckIiysV1SSyRMY9BkLF++q0yKiDxCqRhosKfMg3janKkiLBwZ5c/YryloKwGZcrEmtwxKw==",
       "dev": true,
-      "hasInstallScript": true,
       "license": "MIT",
+      "dependencies": {
+        "esbuild": "~0.28.0"
+      },
       "bin": {
-        "esbuild": "bin/esbuild"
+        "tsx": "dist/cli.mjs"
       },
       "engines": {
-        "node": ">=18"
+        "node": ">=18.0.0"
       },
       "optionalDependencies": {
-        "@esbuild/aix-ppc64": "0.28.2",
-        "@esbuild/android-arm": "0.28.2",
-        "@esbuild/android-arm64": "0.28.2",
-        "@esbuild/android-x64": "0.28.2",
-        "@esbuild/darwin-arm64": "0.28.2",
-        "@esbuild/darwin-x64": "0.28.2",
-        "@esbuild/freebsd-arm64": "0.28.2",
-        "@esbuild/freebsd-x64": "0.28.2",
-        "@esbuild/linux-arm": "0.28.2",
-        "@esbuild/linux-arm64": "0.28.2",
-        "@esbuild/linux-ia32": "0.28.2",
-        "@esbuild/linux-loong64": "0.28.2",
-        "@esbuild/linux-mips64el": "0.28.2",
-        "@esbuild/linux-ppc64": "0.28.2",
-        "@esbuild/linux-riscv64": "0.28.2",
-        "@esbuild/linux-s390x": "0.28.2",
-        "@esbuild/linux-x64": "0.28.2",
-        "@esbuild/netbsd-arm64": "0.28.2",
-        "@esbuild/netbsd-x64": "0.28.2",
-        "@esbuild/openbsd-arm64": "0.28.2",
-        "@esbuild/openbsd-x64": "0.28.2",
-        "@esbuild/openharmony-arm64": "0.28.2",
-        "@esbuild/sunos-x64": "0.28.2",
-        "@esbuild/win32-arm64": "0.28.2",
-        "@esbuild/win32-ia32": "0.28.2",
-        "@esbuild/win32-x64": "0.28.2"
+        "fsevents": "~2.3.3"
       }
     },
     "node_modules/type-check": {
@@ -6098,6 +5959,19 @@
         "node": ">= 0.8.0"
       }
     },
+    "node_modules/type-fest": {
+      "version": "2.19.0",
+      "resolved": "https://registry.npmjs.org/type-fest/-/type-fest-2.19.0.tgz",
+      "integrity": "sha512-RAH822pAdBgcNMAfWnCBU3CFZcfZ/i1eZjwFU/dsLKumyuuP3niueg2UAukXYF0E2AAoc82ZSSf9J0WQBinzHA==",
+      "dev": true,
+      "license": "(MIT OR CC0-1.0)",
+      "engines": {
+        "node": ">=12.20"
+      },
+      "funding": {
+        "url": "https://github.com/sponsors/sindresorhus"
+      }
+    },
     "node_modules/typed-css-modules": {
       "version": "0.9.1",
       "resolved": "https://registry.npmjs.org/typed-css-modules/-/typed-css-modules-0.9.1.tgz",
@@ -6126,21 +6000,12 @@
         "node": ">=18.0.0"
       }
     },
-    "node_modules/typed-css-modules/node_modules/ansi-styles": {
-      "version": "4.3.0",
-      "resolved": "https://registry.npmjs.org/ansi-styles/-/ansi-styles-4.3.0.tgz",
-      "integrity": "sha512-zbB9rCJAT1rbjiVDb2hqKFHNYLxgtk8NURxZ3IZwD3F6NtxbXZQCnnSi1Lkx+IDohdPlFp222wVALIheZJQSEg==",
+    "node_modules/typed-css-modules/node_modules/balanced-match": {
+      "version": "1.0.2",
+      "resolved": "https://registry.npmjs.org/balanced-match/-/balanced-match-1.0.2.tgz",
+      "integrity": "sha512-3oSeUO0TMV67hN1AmbXsK4yaqU7tjiHlbxRDZOpH0KW9+CeX4bRAaX0Anxt0tx2MrpRpWwQaPwIlISEJhYU5Pw==",
       "dev": true,
-      "license": "MIT",
-      "dependencies": {
-        "color-convert": "^2.0.1"
-      },
-      "engines": {
-        "node": ">=8"
-      },
-      "funding": {
-        "url": "https://github.com/chalk/ansi-styles?sponsor=1"
-      }
+      "license": "MIT"
     },
     "node_modules/typed-css-modules/node_modules/brace-expansion": {
       "version": "2.1.7",
@@ -6152,23 +6017,6 @@
         "balanced-match": "^1.0.0"
       }
     },
-    "node_modules/typed-css-modules/node_modules/chalk": {
-      "version": "4.1.2",
-      "resolved": "https://registry.npmjs.org/chalk/-/chalk-4.1.2.tgz",
-      "integrity": "sha512-oKnbhFyRIXpUuez8iBMmyEa4nbj4IOQyuhc/wy9kY7/WVPcwIO9VA668Pu8RkO7+0G76SLROeyw9CpQ061i4mA==",
-      "dev": true,
-      "license": "MIT",
-      "dependencies": {
-        "ansi-styles": "^4.1.0",
-        "supports-color": "^7.1.0"
-      },
-      "engines": {
-        "node": ">=10"
-      },
-      "funding": {
-        "url": "https://github.com/chalk/chalk?sponsor=1"
-      }
-    },
     "node_modules/typed-css-modules/node_modules/chokidar": {
       "version": "3.6.0",
       "resolved": "https://registry.npmjs.org/chokidar/-/chokidar-3.6.0.tgz",
@@ -6198,6 +6046,7 @@
       "version": "10.5.0",
       "resolved": "https://registry.npmjs.org/glob/-/glob-10.5.0.tgz",
       "integrity": "sha512-DfXN8DfhJ7NH3Oe7cFmu3NCu1wKbkReJ8TorzSAFbSKrlNaQSKfIzqYqVY8zlbs2NLBbWpRiU52GX2PbaBVNkg==",
+      "deprecated": "Old versions of glob are not supported, and contain widely publicized security vulnerabilities, which have been fixed in the current version. Please update. Support for old versions may be purchased (at exorbitant rates) by contacting i@izs.me",
       "dev": true,
       "license": "ISC",
       "dependencies": {
@@ -6228,22 +6077,6 @@
         "node": ">= 6"
       }
     },
-    "node_modules/typed-css-modules/node_modules/jackspeak": {
-      "version": "3.4.3",
-      "resolved": "https://registry.npmjs.org/jackspeak/-/jackspeak-3.4.3.tgz",
-      "integrity": "sha512-OGlZQpz2yfahA/Rd1Y8Cd9SIEsqvXkLVoSw/cgwhnhFMDbsQFeZYoJJ7bIZBS9BcamUW96asq/npPWugM+RQBw==",
-      "dev": true,
-      "license": "BlueOak-1.0.0",
-      "dependencies": {
-        "@isaacs/cliui": "^8.0.2"
-      },
-      "funding": {
-        "url": "https://github.com/sponsors/isaacs"
-      },
-      "optionalDependencies": {
-        "@pkgjs/parseargs": "^0.11.0"
-      }
-    },
     "node_modules/typed-css-modules/node_modules/lru-cache": {
       "version": "10.4.3",
       "resolved": "https://registry.npmjs.org/lru-cache/-/lru-cache-10.4.3.tgz",
@@ -6284,19 +6117,6 @@
         "url": "https://github.com/sponsors/isaacs"
       }
     },
-    "node_modules/typed-css-modules/node_modules/picomatch": {
-      "version": "2.3.2",
-      "resolved": "https://registry.npmjs.org/picomatch/-/picomatch-2.3.2.tgz",
-      "integrity": "sha512-V7+vQEJ06Z+c5tSye8S+nHUfI51xoXIXjHQ99cQtKUkQqqO1kO/KCJUfZXuB47h/YBlDhah2H3hdUGXn8ie0oA==",
-      "dev": true,
-      "license": "MIT",
-      "engines": {
-        "node": ">=8.6"
-      },
-      "funding": {
-        "url": "https://github.com/sponsors/jonschlinkert"
-      }
-    },
     "node_modules/typed-css-modules/node_modules/readdirp": {
       "version": "3.6.0",
       "resolved": "https://registry.npmjs.org/readdirp/-/readdirp-3.6.0.tgz",
@@ -6311,9 +6131,9 @@
       }
     },
     "node_modules/typescript": {
-      "version": "5.9.2",
-      "resolved": "https://registry.npmjs.org/typescript/-/typescript-5.9.2.tgz",
-      "integrity": "sha512-CWBzXQrc/qOkhidw1OzBTQuYRbfyxDXJMVJ1XNwUHGROVmuaeiEm3OslpZ1RV96d7SKKjZKrSJu3+t/xlw3R9A==",
+      "version": "6.0.3",
+      "resolved": "https://registry.npmjs.org/typescript/-/typescript-6.0.3.tgz",
+      "integrity": "sha512-y2TvuxSZPDyQakkFRPZHKFm+KKVqIisdg9/CZwm9ftvKXLP8NRWj38/ODjNbr43SsoXqNuAisEf1GdCxqWcdBw==",
       "dev": true,
       "license": "Apache-2.0",
       "bin": {
@@ -6325,16 +6145,16 @@
       }
     },
     "node_modules/typescript-eslint": {
-      "version": "8.39.1",
-      "resolved": "https://registry.npmjs.org/typescript-eslint/-/typescript-eslint-8.39.1.tgz",
-      "integrity": "sha512-GDUv6/NDYngUlNvwaHM1RamYftxf782IyEDbdj3SeaIHHv8fNQVRC++fITT7kUJV/5rIA/tkoRSSskt6osEfqg==",
+      "version": "8.72.0",
+      "resolved": "https://registry.npmjs.org/typescript-eslint/-/typescript-eslint-8.72.0.tgz",
+      "integrity": "sha512-r26Fkc8As+mZvqp7SLsMmNFyYUup/6CjK7/eHBHmhOw6GH6n6YFxHVFrG0jN1c5CMpDzelLPxvgeCOzx7BjN+A==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
-        "@typescript-eslint/eslint-plugin": "8.39.1",
-        "@typescript-eslint/parser": "8.39.1",
-        "@typescript-eslint/typescript-estree": "8.39.1",
-        "@typescript-eslint/utils": "8.39.1"
+        "@typescript-eslint/eslint-plugin": "8.72.0",
+        "@typescript-eslint/parser": "8.72.0",
+        "@typescript-eslint/typescript-estree": "8.72.0",
+        "@typescript-eslint/utils": "8.72.0"
       },
       "engines": {
         "node": "^18.18.0 || ^20.9.0 || >=21.1.0"
@@ -6344,17 +6164,30 @@
         "url": "https://opencollective.com/typescript-eslint"
       },
       "peerDependencies": {
-        "eslint": "^8.57.0 || ^9.0.0",
-        "typescript": ">=4.8.4 <6.0.0"
+        "eslint": "^8.57.0 || ^9.0.0 || ^10.0.0",
+        "typescript": ">=4.8.4 <6.1.0"
       }
     },
     "node_modules/undici-types": {
-      "version": "6.21.0",
-      "resolved": "https://registry.npmjs.org/undici-types/-/undici-types-6.21.0.tgz",
-      "integrity": "sha512-iwDZqg0QAGrg9Rav5H4n0M64c3mkR59cJ6wQp+7C4nI0gsmExaedaYLNO44eT4AtBBwjbTiGPMlt2Md0T9H9JQ==",
+      "version": "8.9.0",
+      "resolved": "https://registry.npmjs.org/undici-types/-/undici-types-8.9.0.tgz",
+      "integrity": "sha512-KTDyRTYX8sWmKXAikPHHSyc63CRPETMctyjKFupcC6OBLXT3xsN0e9aF7m+mIXutFWpUXuedtowG7iLOzp0kQg==",
       "dev": true,
       "license": "MIT"
     },
+    "node_modules/unicorn-magic": {
+      "version": "0.4.1",
+      "resolved": "https://registry.npmjs.org/unicorn-magic/-/unicorn-magic-0.4.1.tgz",
+      "integrity": "sha512-lzlXPoVB0Uy/FvTaUgX39RbixYiSKoc3JSoSf01+0c2B6FR3qdYIPN7Qjo/AV7n6sqLCk0509cT3LRj1oSZ1+A==",
+      "dev": true,
+      "license": "MIT",
+      "engines": {
+        "node": ">=20"
+      },
+      "funding": {
+        "url": "https://github.com/sponsors/sindresorhus"
+      }
+    },
     "node_modules/unified": {
       "version": "11.0.5",
       "resolved": "https://registry.npmjs.org/unified/-/unified-11.0.5.tgz",
@@ -6391,9 +6224,9 @@
       }
     },
     "node_modules/unist-util-is": {
-      "version": "6.0.0",
-      "resolved": "https://registry.npmjs.org/unist-util-is/-/unist-util-is-6.0.0.tgz",
-      "integrity": "sha512-2qCTHimwdxLfz+YzdGfkqNlH0tLi9xjTnHddPmJwtIG9MGsdbutfTc4P+haPD7l7Cjxf/WZj+we5qfVPvvxfYw==",
+      "version": "6.0.1",
+      "resolved": "https://registry.npmjs.org/unist-util-is/-/unist-util-is-6.0.1.tgz",
+      "integrity": "sha512-LsiILbtBETkDz8I9p1dQ0uyRUWuaQzd/cuEeS1hoRSyW5E5XGmTzlwY1OrNzzakGowI9Dr/I8HVaw4hTtnxy8g==",
       "license": "MIT",
       "dependencies": {
         "@types/unist": "^3.0.0"
@@ -6430,9 +6263,9 @@
       }
     },
     "node_modules/unist-util-visit": {
-      "version": "5.0.0",
-      "resolved": "https://registry.npmjs.org/unist-util-visit/-/unist-util-visit-5.0.0.tgz",
-      "integrity": "sha512-MR04uvD+07cwl/yhVuVWAtw+3GOR/knlL55Nd/wAdblk27GCVt3lqpTivy/tkJcZoNPzTwS1Y+KMojlLDhoTzg==",
+      "version": "5.1.0",
+      "resolved": "https://registry.npmjs.org/unist-util-visit/-/unist-util-visit-5.1.0.tgz",
+      "integrity": "sha512-m+vIdyeCOpdr/QeQCu2EzxX/ohgS8KbnPDgFni4dQsfSCtpz8UqDyY5GjRru8PDKuYn7Fq19j1CQ+nJSsGKOzg==",
       "license": "MIT",
       "dependencies": {
         "@types/unist": "^3.0.0",
@@ -6445,9 +6278,9 @@
       }
     },
     "node_modules/unist-util-visit-parents": {
-      "version": "6.0.1",
-      "resolved": "https://registry.npmjs.org/unist-util-visit-parents/-/unist-util-visit-parents-6.0.1.tgz",
-      "integrity": "sha512-L/PqWzfTP9lzzEa6CKs0k2nARxTdZduw3zyh8d2NVBnsyvHjSX4TWse388YrrQKbvI8w20fGjGlhgT96WwKykw==",
+      "version": "6.0.2",
+      "resolved": "https://registry.npmjs.org/unist-util-visit-parents/-/unist-util-visit-parents-6.0.2.tgz",
+      "integrity": "sha512-goh1s1TBrqSqukSc8wrjwWhL0hiJxgA8m4kFxGlQ+8FYQ3C/m11FcTs4YYem7V664AhHVvgoQLk890Ssdsr2IQ==",
       "license": "MIT",
       "dependencies": {
         "@types/unist": "^3.0.0",
@@ -6468,6 +6301,37 @@
         "node": ">= 10.0.0"
       }
     },
+    "node_modules/update-browserslist-db": {
+      "version": "1.3.4",
+      "resolved": "https://registry.npmjs.org/update-browserslist-db/-/update-browserslist-db-1.3.4.tgz",
+      "integrity": "sha512-t8wonIOdQyOjtPd2VktlBu6QjTVTks6v8bSi7kd0r2Vll8TK1JYDZtRRdbmCHOAMphkZgrzK+L+vXK4pyZ8t4A==",
+      "dev": true,
+      "funding": [
+        {
+          "type": "opencollective",
+          "url": "https://opencollective.com/browserslist"
+        },
+        {
+          "type": "tidelift",
+          "url": "https://tidelift.com/funding/github/npm/browserslist"
+        },
+        {
+          "type": "github",
+          "url": "https://github.com/sponsors/ai"
+        }
+      ],
+      "license": "MIT",
+      "dependencies": {
+        "escalade": "^3.2.0",
+        "picocolors": "^1.1.1"
+      },
+      "bin": {
+        "update-browserslist-db": "cli.js"
+      },
+      "peerDependencies": {
+        "browserslist": ">= 4.21.0"
+      }
+    },
     "node_modules/uri-js": {
       "version": "4.4.1",
       "resolved": "https://registry.npmjs.org/uri-js/-/uri-js-4.4.1.tgz",
@@ -6479,9 +6343,9 @@
       }
     },
     "node_modules/use-sync-external-store": {
-      "version": "1.5.0",
-      "resolved": "https://registry.npmjs.org/use-sync-external-store/-/use-sync-external-store-1.5.0.tgz",
-      "integrity": "sha512-Rb46I4cGGVBmjamjphe8L/UnvJD+uPPtTkNvX5mZgqdbavhI4EbgIWJiIHXJ8bc/i9EQGPRh4DwEURJ552Do0A==",
+      "version": "1.7.0",
+      "resolved": "https://registry.npmjs.org/use-sync-external-store/-/use-sync-external-store-1.7.0.tgz",
+      "integrity": "sha512-6L+EeigHMQhdaIPNIFUKwfWJSwWFQ8gJbJ2DLOs5sDIegTwR9fRxvnM3uciHKjIZhFz+KAv2emhWMRvDmMcY8A==",
       "license": "MIT",
       "peerDependencies": {
         "react": "^16.8.0 || ^17.0.0 || ^18.0.0 || ^19.0.0"
@@ -6494,20 +6358,6 @@
       "dev": true,
       "license": "MIT"
     },
-    "node_modules/uuid": {
-      "version": "14.0.0",
-      "resolved": "https://registry.npmjs.org/uuid/-/uuid-14.0.0.tgz",
-      "integrity": "sha512-Qo+uWgilfSmAhXCMav1uYFynlQO7fMFiMVZsQqZRMIXp0O7rR7qjkj+cPvBHLgBqi960QCoo/PH2/6ZtVqKvrg==",
-      "dev": true,
-      "funding": [
-        "https://github.com/sponsors/broofa",
-        "https://github.com/sponsors/ctavan"
-      ],
-      "license": "MIT",
-      "bin": {
-        "uuid": "dist-node/bin/uuid"
-      }
-    },
     "node_modules/vfile": {
       "version": "6.0.3",
       "resolved": "https://registry.npmjs.org/vfile/-/vfile-6.0.3.tgz",
@@ -6569,12 +6419,11 @@
       }
     },
     "node_modules/wouter": {
-      "version": "3.7.1",
-      "resolved": "https://registry.npmjs.org/wouter/-/wouter-3.7.1.tgz",
-      "integrity": "sha512-od5LGmndSUzntZkE2R5CHhoiJ7YMuTIbiXsa0Anytc2RATekgv4sfWRAxLEULBrp7ADzinWQw8g470lkT8+fOw==",
+      "version": "3.13.0",
+      "resolved": "https://registry.npmjs.org/wouter/-/wouter-3.13.0.tgz",
+      "integrity": "sha512-p1RoCuVyxwb3X5jLjDdkwDzlXj+/UGZqZb+806uaJOVEHvNfGkjET0KiQ/CGiDKC90rLsALMgMj4onNVKxgNdQ==",
       "license": "Unlicense",
       "dependencies": {
-        "mitt": "^3.0.1",
         "regexparam": "^3.0.0",
         "use-sync-external-store": "^1.0.0"
       },
@@ -6583,18 +6432,18 @@
       }
     },
     "node_modules/wrap-ansi": {
-      "version": "8.1.0",
-      "resolved": "https://registry.npmjs.org/wrap-ansi/-/wrap-ansi-8.1.0.tgz",
-      "integrity": "sha512-si7QWI6zUMq56bESFvagtmzMdGOtoxfR+Sez11Mobfc7tm+VkUckk9bW2UeffTGVUbOksxmSw0AA2gs8g71NCQ==",
+      "version": "7.0.0",
+      "resolved": "https://registry.npmjs.org/wrap-ansi/-/wrap-ansi-7.0.0.tgz",
+      "integrity": "sha512-YVGIj2kamLSTxw6NsZjoBxfSwsn0ycdesmc4p+Q21c5zPuZ1pl+NfxVdxPtdHvmNVOQ6XSYG4AUtyt/Fi7D16Q==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
-        "ansi-styles": "^6.1.0",
-        "string-width": "^5.0.1",
-        "strip-ansi": "^7.0.1"
+        "ansi-styles": "^4.0.0",
+        "string-width": "^4.1.0",
+        "strip-ansi": "^6.0.0"
       },
       "engines": {
-        "node": ">=12"
+        "node": ">=10"
       },
       "funding": {
         "url": "https://github.com/chalk/wrap-ansi?sponsor=1"
@@ -6629,40 +6478,45 @@
         "node": ">=8"
       }
     },
-    "node_modules/wrap-ansi-cjs/node_modules/ansi-styles": {
-      "version": "4.3.0",
-      "resolved": "https://registry.npmjs.org/ansi-styles/-/ansi-styles-4.3.0.tgz",
-      "integrity": "sha512-zbB9rCJAT1rbjiVDb2hqKFHNYLxgtk8NURxZ3IZwD3F6NtxbXZQCnnSi1Lkx+IDohdPlFp222wVALIheZJQSEg==",
+    "node_modules/wrap-ansi-cjs/node_modules/string-width": {
+      "version": "4.2.3",
+      "resolved": "https://registry.npmjs.org/string-width/-/string-width-4.2.3.tgz",
+      "integrity": "sha512-wKyQRQpjJ0sIp62ErSZdGsjMJWsap5oRNihHhu6G7JVO/9jIB6UyevL+tXuOqrng8j/cxKTWyWUwvSTriiZz/g==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
-        "color-convert": "^2.0.1"
+        "emoji-regex": "^8.0.0",
+        "is-fullwidth-code-point": "^3.0.0",
+        "strip-ansi": "^6.0.1"
       },
       "engines": {
         "node": ">=8"
-      },
-      "funding": {
-        "url": "https://github.com/chalk/ansi-styles?sponsor=1"
       }
     },
-    "node_modules/wrap-ansi-cjs/node_modules/emoji-regex": {
-      "version": "8.0.0",
-      "resolved": "https://registry.npmjs.org/emoji-regex/-/emoji-regex-8.0.0.tgz",
-      "integrity": "sha512-MSjYzcWNOA0ewAHpz0MxpYFvwg6yjy1NG3xteoqz644VCo/RPgnr1/GGt+ic3iJTzQ8Eu3TdM14SawnVUmGE6A==",
+    "node_modules/wrap-ansi-cjs/node_modules/strip-ansi": {
+      "version": "6.0.1",
+      "resolved": "https://registry.npmjs.org/strip-ansi/-/strip-ansi-6.0.1.tgz",
+      "integrity": "sha512-Y38VPSHcqkFrCpFnQ9vuSXmquuv5oXOKpGeT6aGrr3o3Gc9AlVa6JBfUSOCnbxGGZF+/0ooI7KrPuUSztUdU5A==",
       "dev": true,
-      "license": "MIT"
+      "license": "MIT",
+      "dependencies": {
+        "ansi-regex": "^5.0.1"
+      },
+      "engines": {
+        "node": ">=8"
+      }
     },
-    "node_modules/wrap-ansi-cjs/node_modules/is-fullwidth-code-point": {
-      "version": "3.0.0",
-      "resolved": "https://registry.npmjs.org/is-fullwidth-code-point/-/is-fullwidth-code-point-3.0.0.tgz",
-      "integrity": "sha512-zymm5+u+sCsSWyD9qNaejV3DFvhCKclKdizYaJUuHA83RLjb7nSuGnddCHGv0hk+KY7BMAlsWeK4Ueg6EV6XQg==",
+    "node_modules/wrap-ansi/node_modules/ansi-regex": {
+      "version": "5.0.1",
+      "resolved": "https://registry.npmjs.org/ansi-regex/-/ansi-regex-5.0.1.tgz",
+      "integrity": "sha512-quJQXlTSUGL2LH9SUXo8VwsY4soanhgo6LNSm84E1LBcE8s3O0wpdiRzyR9z/ZZJMlMWv37qOOb9pdJlMUEKFQ==",
       "dev": true,
       "license": "MIT",
       "engines": {
         "node": ">=8"
       }
     },
-    "node_modules/wrap-ansi-cjs/node_modules/string-width": {
+    "node_modules/wrap-ansi/node_modules/string-width": {
       "version": "4.2.3",
       "resolved": "https://registry.npmjs.org/string-width/-/string-width-4.2.3.tgz",
       "integrity": "sha512-wKyQRQpjJ0sIp62ErSZdGsjMJWsap5oRNihHhu6G7JVO/9jIB6UyevL+tXuOqrng8j/cxKTWyWUwvSTriiZz/g==",
@@ -6677,7 +6531,7 @@
         "node": ">=8"
       }
     },
-    "node_modules/wrap-ansi-cjs/node_modules/strip-ansi": {
+    "node_modules/wrap-ansi/node_modules/strip-ansi": {
       "version": "6.0.1",
       "resolved": "https://registry.npmjs.org/strip-ansi/-/strip-ansi-6.0.1.tgz",
       "integrity": "sha512-Y38VPSHcqkFrCpFnQ9vuSXmquuv5oXOKpGeT6aGrr3o3Gc9AlVa6JBfUSOCnbxGGZF+/0ooI7KrPuUSztUdU5A==",
@@ -6690,29 +6544,17 @@
         "node": ">=8"
       }
     },
-    "node_modules/wrap-ansi/node_modules/emoji-regex": {
-      "version": "9.2.2",
-      "resolved": "https://registry.npmjs.org/emoji-regex/-/emoji-regex-9.2.2.tgz",
-      "integrity": "sha512-L18DaJsXSUk2+42pv8mLs5jJT2hqFkFE4j21wOmgbUqsZ2hL72NsUU785g9RXgo3s0ZNgVl42TiHp3ZtOv/Vyg==",
-      "dev": true,
-      "license": "MIT"
-    },
-    "node_modules/wrap-ansi/node_modules/string-width": {
-      "version": "5.1.2",
-      "resolved": "https://registry.npmjs.org/string-width/-/string-width-5.1.2.tgz",
-      "integrity": "sha512-HnLOCR3vjcY8beoNLtcjZ5/nxn2afmME6lhrDrebokqMap+XbeW8n9TXpPDOqdGK5qcI3oT0GKTW6wC7EMiVqA==",
+    "node_modules/write-file-atomic": {
+      "version": "7.0.1",
+      "resolved": "https://registry.npmjs.org/write-file-atomic/-/write-file-atomic-7.0.1.tgz",
+      "integrity": "sha512-OTIk8iR8/aCRWBqvxrzxR0hgxWpnYBblY1S5hDWBQfk/VFmJwzmJgQFN3WsoUKHISv2eAwe+PpbUzyL1CKTLXg==",
       "dev": true,
-      "license": "MIT",
+      "license": "ISC",
       "dependencies": {
-        "eastasianwidth": "^0.2.0",
-        "emoji-regex": "^9.2.2",
-        "strip-ansi": "^7.0.1"
+        "signal-exit": "^4.0.1"
       },
       "engines": {
-        "node": ">=12"
-      },
-      "funding": {
-        "url": "https://github.com/sponsors/sindresorhus"
+        "node": "^20.17.0 || >=22.9.0"
       }
     },
     "node_modules/y18n": {
@@ -6725,10 +6567,17 @@
         "node": ">=10"
       }
     },
+    "node_modules/yallist": {
+      "version": "3.1.1",
+      "resolved": "https://registry.npmjs.org/yallist/-/yallist-3.1.1.tgz",
+      "integrity": "sha512-a4UGQaWPH59mOXUYnAG2ewncQS4i4F43Tv3JoAM+s2VDAmS9NsK8GpDMLrCHPksFT7h3K6TOoUNn2pb7RoXx4g==",
+      "dev": true,
+      "license": "ISC"
+    },
     "node_modules/yargs": {
-      "version": "17.7.2",
-      "resolved": "https://registry.npmjs.org/yargs/-/yargs-17.7.2.tgz",
-      "integrity": "sha512-7dSzzRQ++CKnNI/krKnYRV7JKKPUXMEh61soaHKg9mrWEhzFWhFnxPxGl+69cD1Ou63C13NUPCnmIcrvqCuM6w==",
+      "version": "17.7.3",
+      "resolved": "https://registry.npmjs.org/yargs/-/yargs-17.7.3.tgz",
+      "integrity": "sha512-GZtjxm/J/4TSxuL3FNYjCmLktBTnIw/rVmKSIyKeYAZpmJB2ig9VauCC5xsa82GNKVKDAqpOn3KVzNt0zmrU0g==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
@@ -6764,23 +6613,6 @@
         "node": ">=8"
       }
     },
-    "node_modules/yargs/node_modules/emoji-regex": {
-      "version": "8.0.0",
-      "resolved": "https://registry.npmjs.org/emoji-regex/-/emoji-regex-8.0.0.tgz",
-      "integrity": "sha512-MSjYzcWNOA0ewAHpz0MxpYFvwg6yjy1NG3xteoqz644VCo/RPgnr1/GGt+ic3iJTzQ8Eu3TdM14SawnVUmGE6A==",
-      "dev": true,
-      "license": "MIT"
-    },
-    "node_modules/yargs/node_modules/is-fullwidth-code-point": {
-      "version": "3.0.0",
-      "resolved": "https://registry.npmjs.org/is-fullwidth-code-point/-/is-fullwidth-code-point-3.0.0.tgz",
-      "integrity": "sha512-zymm5+u+sCsSWyD9qNaejV3DFvhCKclKdizYaJUuHA83RLjb7nSuGnddCHGv0hk+KY7BMAlsWeK4Ueg6EV6XQg==",
-      "dev": true,
-      "license": "MIT",
-      "engines": {
-        "node": ">=8"
-      }
-    },
     "node_modules/yargs/node_modules/string-width": {
       "version": "4.2.3",
       "resolved": "https://registry.npmjs.org/string-width/-/string-width-4.2.3.tgz",
@@ -6823,14 +6655,27 @@
       }
     },
     "node_modules/zod": {
-      "version": "4.0.17",
-      "resolved": "https://registry.npmjs.org/zod/-/zod-4.0.17.tgz",
-      "integrity": "sha512-1PHjlYRevNxxdy2JZ8JcNAw7rX8V9P1AKkP+x/xZfxB0K5FYfuV+Ug6P/6NVSR2jHQ+FzDDoDHS04nYUsOIyLQ==",
+      "version": "4.6.7",
+      "resolved": "https://registry.npmjs.org/zod/-/zod-4.6.7.tgz",
+      "integrity": "sha512-g8W2D2kU6a8untc053fX1uVBKzM7yqAxW/3zwk83+6RyF2+SKOuV+c0wBDhUwzV1PaGapmZH+rmcUzPf+3CGwg==",
       "license": "MIT",
       "funding": {
         "url": "https://github.com/sponsors/colinhacks"
       }
     },
+    "node_modules/zod-validation-error": {
+      "version": "4.0.2",
+      "resolved": "https://registry.npmjs.org/zod-validation-error/-/zod-validation-error-4.0.2.tgz",
+      "integrity": "sha512-Q6/nZLe6jxuU80qb/4uJ4t5v2VEZ44lzQjPDhYJNztRQ4wyWc6VF3D3Kb/fAuPetZQnhS3hnajCf9CsWesghLQ==",
+      "dev": true,
+      "license": "MIT",
+      "engines": {
+        "node": ">=18.0.0"
+      },
+      "peerDependencies": {
+        "zod": "^3.25.0 || ^4.0.0"
+      }
+    },
     "node_modules/zwitch": {
       "version": "2.0.4",
       "resolved": "https://registry.npmjs.org/zwitch/-/zwitch-2.0.4.tgz",
@@ -6843,15 +6688,15 @@
     },
     "paddler_client_javascript": {
       "name": "@intentee/paddler-client",
-      "version": "4.0.1",
+      "version": "5.0.0",
       "license": "Apache-2.0",
       "dependencies": {
-        "nanoid": "5.1.16"
+        "nanoid": "6.0.2"
       },
       "devDependencies": {
-        "@types/node": "22.19.17",
+        "@types/node": "26.6.5",
         "tsx": "4.23.15",
-        "typescript": "5.9.2"
+        "typescript": "6.0.3"
       },
       "engines": {
         "node": ">=22"
diff --git a/package.json b/package.json
index 69ebce29d..2688392ef 100644
--- a/package.json
+++ b/package.json
@@ -1,48 +1,52 @@
 {
   "devDependencies": {
-    "@intentee/rust-coverage-check": "0.2.0",
-    "@types/hotwired__turbo": "8.0.4",
-    "@types/react": "19.1.10",
-    "@types/react-dom": "19.1.7",
-    "esbuild": "0.25.8",
-    "eslint": "9.33.0",
-    "eslint-plugin-react-hooks": "5.2.0",
-    "fs-extra": "11.3.1",
-    "glob": "11.1.0",
-    "jarmuz": "0.10.0",
+    "@eslint/js": "10.0.1",
+    "@intentee/rust-coverage-check": "0.3.1",
+    "@types/hotwired__turbo": "8.0.16",
+    "@types/react": "19.3.0",
+    "@types/react-dom": "19.3.0",
+    "esbuild": "0.28.2",
+    "eslint": "10.13.0",
+    "eslint-plugin-react-hooks": "7.1.1",
+    "fs-extra": "11.4.1",
+    "glob": "13.0.6",
+    "globals": "17.13.0",
+    "jarmuz": "0.11.0",
     "lodash": "4.18.1",
-    "node-notifier": "10.0.1",
-    "prettier": "3.6.2",
-    "prettier-plugin-organize-imports": "4.2.0",
-    "stylelint": "16.23.1",
-    "stylelint-config-recommended": "17.0.0",
-    "tempy": "3.1.0",
+    "node-notifier": "11.0.0",
+    "prettier": "3.9.10",
+    "prettier-plugin-organize-imports": "4.3.0",
+    "stylelint": "17.16.0",
+    "stylelint-config-recommended": "18.0.0",
+    "tempy": "3.2.0",
     "tslib": "2.8.1",
     "typed-css-modules": "0.9.1",
-    "typescript": "5.9.2",
-    "typescript-eslint": "8.39.1"
+    "typescript": "6.0.3",
+    "typescript-eslint": "8.72.0"
   },
   "dependencies": {
-    "@codemirror/lang-jinja": "6.0.0",
-    "@uiw/react-codemirror": "4.24.2",
+    "@codemirror/lang-jinja": "6.0.1",
+    "@uiw/react-codemirror": "4.25.12",
     "clsx": "2.1.1",
-    "nanoid": "5.1.16",
-    "path-to-regexp": "8.4.2",
-    "react": "19.1.1",
-    "react-dom": "19.1.1",
+    "nanoid": "6.0.2",
+    "react": "19.3.1",
+    "react-dom": "19.3.1",
     "react-markdown": "10.1.0",
     "rxjs": "7.8.2",
-    "wouter": "3.7.1",
-    "zod": "4.0.17"
+    "wouter": "3.13.0",
+    "zod": "4.6.7"
   },
   "license": "Apache-2.0",
   "name": "paddler",
   "private": true,
   "overrides": {
-    "node-notifier": {
-      "uuid": "14.0.0"
+    "jarmuz": {
+      "nanoid": "6.0.2"
     }
   },
   "type": "module",
-  "workspaces": ["paddler_client_javascript"]
+  "workspaces": ["paddler_client_javascript"],
+  "allowScripts": {
+    "esbuild@0.28.2": true
+  }
 }
diff --git a/paddler_agent/Cargo.toml b/paddler_agent/Cargo.toml
index d9c84ab8f..2c686b75a 100644
--- a/paddler_agent/Cargo.toml
+++ b/paddler_agent/Cargo.toml
@@ -1,5 +1,6 @@
 [package]
 name = "paddler_agent"
+autotests = false
 authors.workspace = true
 description.workspace = true
 edition.workspace = true
@@ -8,57 +9,51 @@ license.workspace = true
 repository.workspace = true
 version.workspace = true
 
+[[test]]
+name = "integration"
+path = "tests/main.rs"
+
 [dependencies]
 anyhow = { workspace = true }
-async-stream = { workspace = true }
 async-trait = { workspace = true }
-bytes = { workspace = true }
-dashmap = { workspace = true }
-encoding_rs = { workspace = true }
-futures = { workspace = true }
 futures-util = { workspace = true }
-hf-hub = { workspace = true }
-llama-cpp-bindings = { workspace = true }
-llama-cpp-bindings-sys = { workspace = true }
-llama-cpp-bindings-types = { workspace = true }
 log = { workspace = true }
-minijinja = { workspace = true }
-minijinja-contrib = { workspace = true }
-nanoid = { workspace = true }
-paddler_cache_dir = { workspace = true }
-paddler_download_manager = { workspace = true }
-paddler_image_decoder = { workspace = true }
-paddler_inference_parameters = { workspace = true }
-paddler_agent_status = { workspace = true }
-paddler_messaging = { workspace = true }
-paddler_model_source = { workspace = true }
-paddler_request_registry = { workspace = true }
-paddler_tool_call_validator = { workspace = true }
-parking_lot = { workspace = true }
-rand = { workspace = true }
-reqwest = { workspace = true }
 serde = { workspace = true }
 serde_json = { workspace = true }
-shellexpand = { workspace = true }
 thiserror = { workspace = true }
 tokio = { workspace = true }
-tokio-stream = { workspace = true }
 tokio-tungstenite = { workspace = true }
 tokio-util = { workspace = true }
 trzcina = { workspace = true }
-url = { workspace = true }
+paddler_agent_decision = { workspace = true }
+paddler_agent_embeddings = { workspace = true }
+paddler_agent_runtime = { workspace = true }
+paddler_agent_status = { workspace = true }
+paddler_agent_text_generation = { workspace = true }
+paddler_frame_writer = { workspace = true }
+paddler_inference_parameters = { workspace = true }
+paddler_messaging = { workspace = true }
+paddler_model_source = { workspace = true }
+paddler_outbox = { workspace = true }
+paddler_request_registry = { workspace = true }
+paddler_spawn_resuming_panic = { workspace = true }
 
 [features]
 default = []
-cuda = ["llama-cpp-bindings/cuda"]
-metal = ["llama-cpp-bindings/metal"]
-vulkan = ["llama-cpp-bindings/vulkan"]
+cuda = ["paddler_agent_decision/cuda", "paddler_agent_embeddings/cuda", "paddler_agent_runtime/cuda", "paddler_agent_text_generation/cuda"]
+metal = ["paddler_agent_decision/metal", "paddler_agent_embeddings/metal", "paddler_agent_runtime/metal", "paddler_agent_text_generation/metal"]
+vulkan = ["paddler_agent_decision/vulkan", "paddler_agent_embeddings/vulkan", "paddler_agent_runtime/vulkan", "paddler_agent_text_generation/vulkan"]
 
 [dev-dependencies]
-indoc = { workspace = true }
-paddler_local_http_fixture = { workspace = true }
+http = { workspace = true }
+nix = { workspace = true }
 tempfile = { workspace = true }
-tokio-test = { workspace = true }
+tokio = { workspace = true, features = ["test-util"] }
+
+paddler_blocking_thread_fixture = { workspace = true }
+paddler_cache_dir = { workspace = true }
+paddler_download_manager = { workspace = true }
+paddler_local_http_fixture = { workspace = true }
 
 [lints]
 workspace = true
diff --git a/paddler_agent/src/agent_applicable_model.rs b/paddler_agent/src/agent_applicable_model.rs
deleted file mode 100644
index 80f0fbd5e..000000000
--- a/paddler_agent/src/agent_applicable_model.rs
+++ /dev/null
@@ -1,11 +0,0 @@
-use std::path::PathBuf;
-
-#[derive(Clone, Debug, Default, Eq, PartialEq)]
-pub enum AgentApplicableModel {
-    #[default]
-    NotConfigured,
-    Resolved {
-        model_path: PathBuf,
-        multimodal_projection_path: Option<PathBuf>,
-    },
-}
diff --git a/paddler_agent/src/agent_applicable_state.rs b/paddler_agent/src/agent_applicable_state.rs
index 43a41548d..7b5c36b82 100644
--- a/paddler_agent/src/agent_applicable_state.rs
+++ b/paddler_agent/src/agent_applicable_state.rs
@@ -1,11 +1,63 @@
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
+use paddler_agent_text_generation::text_generation_settings::TextGenerationSettings;
 use paddler_messaging::chat_template::ChatTemplate;
+use paddler_messaging::chat_template_source::ChatTemplateSource;
 
-use crate::agent_applicable_model::AgentApplicableModel;
+use crate::inference_capabilities::InferenceCapabilities;
+use crate::model_serving_settings::ModelServingSettings;
 
 #[derive(Clone, Debug, Default, PartialEq)]
-pub struct AgentApplicableState {
-    pub chat_template_override: Option<ChatTemplate>,
-    pub inference_parameters: InferenceParameters,
-    pub model: AgentApplicableModel,
+pub enum AgentApplicableState {
+    #[default]
+    NotConfigured,
+    WithModel(ModelServingSettings),
+    WithoutModel {
+        chat_template_source: ChatTemplateSource,
+    },
+}
+
+impl AgentApplicableState {
+    #[must_use]
+    pub fn chat_template_override(&self) -> Option<ChatTemplate> {
+        match self {
+            Self::WithModel(ModelServingSettings {
+                capabilities:
+                    InferenceCapabilities {
+                        text_generation:
+                            TextGenerationSettings {
+                                chat_template_source: ChatTemplateSource::Override(chat_template),
+                                ..
+                            },
+                        ..
+                    },
+                ..
+            })
+            | Self::WithoutModel {
+                chat_template_source: ChatTemplateSource::Override(chat_template),
+            } => Some(chat_template.clone()),
+            Self::NotConfigured | Self::WithModel(_) | Self::WithoutModel { .. } => None,
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use paddler_messaging::chat_template::ChatTemplate;
+    use paddler_messaging::chat_template_source::ChatTemplateSource;
+
+    use super::AgentApplicableState;
+
+    #[test]
+    fn a_state_without_a_model_overrides_its_chat_template() {
+        let chat_template = ChatTemplate {
+            content: "{{ messages }}".to_owned(),
+        };
+
+        assert_eq!(
+            AgentApplicableState::WithoutModel {
+                chat_template_source: ChatTemplateSource::Override(chat_template.clone()),
+            }
+            .chat_template_override(),
+            Some(chat_template)
+        );
+    }
 }
diff --git a/paddler_agent/src/agent_applicable_state_holder.rs b/paddler_agent/src/agent_applicable_state_holder.rs
index 8f8f96f6b..a8a192a5f 100644
--- a/paddler_agent/src/agent_applicable_state_holder.rs
+++ b/paddler_agent/src/agent_applicable_state_holder.rs
@@ -15,7 +15,15 @@ impl AgentApplicableStateHolder {
 
     pub fn set_agent_applicable_state(&self, agent_applicable_state: AgentApplicableState) {
         self.agent_applicable_state_tx
-            .send_replace(agent_applicable_state);
+            .send_if_modified(|current_agent_applicable_state| {
+                if *current_agent_applicable_state == agent_applicable_state {
+                    return false;
+                }
+
+                *current_agent_applicable_state = agent_applicable_state;
+
+                true
+            });
     }
 
     #[must_use]
@@ -27,35 +35,3 @@ impl AgentApplicableStateHolder {
         agent_applicable_state_rx
     }
 }
-
-#[cfg(test)]
-mod tests {
-    use std::path::PathBuf;
-
-    use paddler_inference_parameters::inference_parameters::InferenceParameters;
-
-    use super::AgentApplicableStateHolder;
-    use crate::agent_applicable_model::AgentApplicableModel;
-    use crate::agent_applicable_state::AgentApplicableState;
-
-    #[test]
-    fn a_new_subscriber_sees_the_state_applied_before_it_subscribed() {
-        let holder = AgentApplicableStateHolder::default();
-
-        let applied_state = AgentApplicableState {
-            chat_template_override: None,
-            inference_parameters: InferenceParameters::default(),
-            model: AgentApplicableModel::Resolved {
-                model_path: PathBuf::from("model.gguf"),
-                multimodal_projection_path: None,
-            },
-        };
-
-        holder.set_agent_applicable_state(applied_state.clone());
-
-        let agent_applicable_state_rx = holder.subscribe();
-
-        assert!(agent_applicable_state_rx.has_changed().unwrap());
-        assert_eq!(*agent_applicable_state_rx.borrow(), applied_state);
-    }
-}
diff --git a/paddler_agent/src/agent_desired_state_conversion.rs b/paddler_agent/src/agent_desired_state_conversion.rs
index 15e5fe8e4..6b97532af 100644
--- a/paddler_agent/src/agent_desired_state_conversion.rs
+++ b/paddler_agent/src/agent_desired_state_conversion.rs
@@ -1,5 +1,6 @@
 use crate::agent_applicable_state::AgentApplicableState;
 
+#[derive(Debug, PartialEq)]
 pub enum AgentDesiredStateConversion {
     Cancelled,
     Converted(AgentApplicableState),
diff --git a/paddler_agent/src/agent_desired_state_converter.rs b/paddler_agent/src/agent_desired_state_converter.rs
index 9908bb5a4..fa80c712d 100644
--- a/paddler_agent/src/agent_desired_state_converter.rs
+++ b/paddler_agent/src/agent_desired_state_converter.rs
@@ -5,19 +5,27 @@ use std::sync::Arc;
 use tokio_util::sync::CancellationToken;
 
 use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
+use paddler_agent_text_generation::multimodal_projection::MultimodalProjection;
+use paddler_agent_text_generation::text_generation_settings::TextGenerationSettings;
+use paddler_messaging::agent_decisions_settings::AgentDecisionsSettings;
 use paddler_messaging::agent_desired_model::AgentDesiredModel;
 use paddler_messaging::agent_desired_state::AgentDesiredState;
-use paddler_messaging::agent_issue::AgentIssue;
-use paddler_messaging::agent_issue_params::model_path::ModelPath;
+use paddler_messaging::balancer_text_generation_settings::BalancerTextGenerationSettings;
+use paddler_messaging::multimodal_settings::MultimodalSettings;
+use paddler_messaging::served_decision_settings::ServedDecisionSettings;
 use paddler_model_source::desired_model_resolution::DesiredModelResolution;
+use paddler_model_source::desired_model_resolver::DesiredModelResolver;
 use paddler_model_source::model_source_error::ModelSourceError;
-use paddler_model_source::resolve_desired_model::resolve_desired_model;
 
-use crate::agent_applicable_model::AgentApplicableModel;
 use crate::agent_applicable_state::AgentApplicableState;
 use crate::agent_desired_state_conversion::AgentDesiredStateConversion;
+use crate::applicable_decisions::ApplicableDecisions;
+use crate::desired_model_file::DesiredModelFile;
+use crate::inference_capabilities::InferenceCapabilities;
+use crate::model_serving_settings::ModelServingSettings;
 
 pub struct AgentDesiredStateConverter {
+    pub desired_model_resolver: DesiredModelResolver,
     pub cancellation_token: CancellationToken,
     pub slot_aggregated_status: Arc<SlotAggregatedStatus>,
 }
@@ -26,60 +34,127 @@ impl AgentDesiredStateConverter {
     pub async fn convert(
         &self,
         AgentDesiredState {
-            chat_template_override,
-            inference_parameters,
+            decisions,
             model,
-            multimodal_projection,
+            model_runtime_parameters,
+            pooling_type,
+            text_generation:
+                BalancerTextGenerationSettings {
+                    chat_template_source,
+                    multimodal:
+                        MultimodalSettings {
+                            image_resize_to_fit,
+                            projection,
+                        },
+                    sampling_parameters,
+                },
         }: &AgentDesiredState,
     ) -> Result<AgentDesiredStateConversion, ModelSourceError> {
         let ControlFlow::Continue(model_path) = self
-            .resolve_model_file(model, AgentIssue::ModelFileDoesNotExist)
+            .resolve_model_file(model, DesiredModelFile::Model)
             .await?
         else {
             return Ok(AgentDesiredStateConversion::Cancelled);
         };
+        let Some(model_path) = model_path else {
+            return Ok(AgentDesiredStateConversion::Converted(
+                AgentApplicableState::WithoutModel {
+                    chat_template_source: chat_template_source.clone(),
+                },
+            ));
+        };
         let ControlFlow::Continue(multimodal_projection_path) = self
-            .resolve_model_file(
-                multimodal_projection,
-                AgentIssue::MultimodalProjectionCannotBeLoaded,
-            )
+            .resolve_model_file(projection, DesiredModelFile::MultimodalProjection)
             .await?
         else {
             return Ok(AgentDesiredStateConversion::Cancelled);
         };
+        let ControlFlow::Continue(applicable_decisions) = self.resolve_decisions(decisions).await?
+        else {
+            return Ok(AgentDesiredStateConversion::Cancelled);
+        };
 
         Ok(AgentDesiredStateConversion::Converted(
-            AgentApplicableState {
-                chat_template_override: chat_template_override.clone(),
-                inference_parameters: inference_parameters.clone(),
-                model: model_path.map_or(AgentApplicableModel::NotConfigured, |model_path| {
-                    AgentApplicableModel::Resolved {
-                        model_path,
-                        multimodal_projection_path,
-                    }
-                }),
-            },
+            AgentApplicableState::WithModel(ModelServingSettings {
+                capabilities: InferenceCapabilities {
+                    decisions: applicable_decisions,
+                    embeddings: pooling_type.clone(),
+                    text_generation: TextGenerationSettings {
+                        chat_template_source: chat_template_source.clone(),
+                        image_resize_to_fit: *image_resize_to_fit,
+                        multimodal_projection: multimodal_projection_path.map_or(
+                            MultimodalProjection::NotConfigured,
+                            MultimodalProjection::File,
+                        ),
+                        sampling_parameters: sampling_parameters.clone(),
+                    },
+                },
+                model_path,
+                model_runtime_parameters: model_runtime_parameters.clone(),
+            }),
         ))
     }
 
+    async fn resolve_decisions(
+        &self,
+        decisions: &AgentDecisionsSettings,
+    ) -> Result<ControlFlow<(), ApplicableDecisions>, ModelSourceError> {
+        match decisions {
+            AgentDecisionsSettings::NotServed => {
+                Ok(ControlFlow::Continue(ApplicableDecisions::Disabled))
+            }
+            AgentDecisionsSettings::Served(ServedDecisionSettings { pointer_head_uri }) => Ok(
+                match self
+                    .resolve_model_uri(pointer_head_uri, DesiredModelFile::PointerHead)
+                    .await?
+                {
+                    ControlFlow::Break(()) => ControlFlow::Break(()),
+                    ControlFlow::Continue(pointer_head_path) => {
+                        ControlFlow::Continue(ApplicableDecisions::Enabled { pointer_head_path })
+                    }
+                },
+            ),
+        }
+    }
+
     async fn resolve_model_file(
         &self,
         desired_model: &AgentDesiredModel,
-        local_file_missing_issue: fn(ModelPath) -> AgentIssue,
+        desired_model_file: DesiredModelFile,
     ) -> Result<ControlFlow<(), Option<PathBuf>>, ModelSourceError> {
-        match resolve_desired_model(
-            &self.cancellation_token,
-            desired_model,
-            self.slot_aggregated_status.clone(),
-        )
-        .await?
+        match desired_model {
+            AgentDesiredModel::None => Ok(ControlFlow::Continue(None)),
+            AgentDesiredModel::Uri(uri) => Ok(
+                match self.resolve_model_uri(uri, desired_model_file).await? {
+                    ControlFlow::Break(()) => ControlFlow::Break(()),
+                    ControlFlow::Continue(path) => ControlFlow::Continue(Some(path)),
+                },
+            ),
+        }
+    }
+
+    async fn resolve_model_uri(
+        &self,
+        uri: &str,
+        desired_model_file: DesiredModelFile,
+    ) -> Result<ControlFlow<(), PathBuf>, ModelSourceError> {
+        match self
+            .desired_model_resolver
+            .resolve_uri(
+                &self.cancellation_token,
+                uri,
+                self.slot_aggregated_status.clone(),
+            )
+            .await?
         {
             DesiredModelResolution::Cancelled => Ok(ControlFlow::Break(())),
-            DesiredModelResolution::NotConfigured => Ok(ControlFlow::Continue(None)),
-            DesiredModelResolution::Resolved(path) => Ok(ControlFlow::Continue(Some(path))),
+            DesiredModelResolution::Resolved(path) => {
+                desired_model_file.report_found(&path, &self.slot_aggregated_status);
+
+                Ok(ControlFlow::Continue(path))
+            }
             DesiredModelResolution::LocalFileMissing(path) => {
-                self.slot_aggregated_status
-                    .register_issue(local_file_missing_issue(ModelPath::from(path.as_path())));
+                desired_model_file.report_missing(&path, &self.slot_aggregated_status);
 
                 Err(ModelSourceError::LocalFileMissing { path })
             }
@@ -89,31 +164,55 @@ impl AgentDesiredStateConverter {
 
 #[cfg(test)]
 mod tests {
-    use std::mem::discriminant;
+    use std::fs::write;
+    use std::path::Path;
     use std::path::PathBuf;
     use std::sync::Arc;
 
+    use tempfile::NamedTempFile;
     use tempfile::TempDir;
     use tempfile::tempdir;
     use tokio_util::sync::CancellationToken;
 
     use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
-    use paddler_inference_parameters::inference_parameters::InferenceParameters;
+    use paddler_agent_text_generation::multimodal_projection::MultimodalProjection;
+    use paddler_agent_text_generation::text_generation_settings::TextGenerationSettings;
+    use paddler_inference_parameters::model_runtime_parameters::ModelRuntimeParameters;
+    use paddler_inference_parameters::pooling_type::PoolingType;
+    use paddler_inference_parameters::sampling_parameters::SamplingParameters;
+    use paddler_messaging::agent_decisions_settings::AgentDecisionsSettings;
     use paddler_messaging::agent_desired_model::AgentDesiredModel;
     use paddler_messaging::agent_desired_state::AgentDesiredState;
     use paddler_messaging::agent_issue::AgentIssue;
     use paddler_messaging::agent_issue_params::model_path::ModelPath;
-    use paddler_messaging::huggingface_model_reference::HuggingFaceModelReference;
+    use paddler_messaging::balancer_text_generation_settings::BalancerTextGenerationSettings;
+    use paddler_messaging::chat_template::ChatTemplate;
+    use paddler_messaging::chat_template_source::ChatTemplateSource;
+    use paddler_messaging::multimodal_settings::MultimodalSettings;
+    use paddler_messaging::served_decision_settings::ServedDecisionSettings;
+    use paddler_model_source::agent_local_model_path::AgentLocalModelPath;
+    use paddler_model_source::desired_model_resolver::DesiredModelResolver;
+    use paddler_model_source::huggingface_default_endpoint::HUGGINGFACE_DEFAULT_ENDPOINT;
+    use paddler_model_source::huggingface_model_reference::HuggingFaceModelReference;
+    use paddler_model_source::model_source::ModelSource;
     use paddler_model_source::model_source_error::ModelSourceError;
 
+    use crate::agent_applicable_state::AgentApplicableState;
     use crate::agent_desired_state_conversion::AgentDesiredStateConversion;
     use crate::agent_desired_state_converter::AgentDesiredStateConverter;
+    use crate::applicable_decisions::ApplicableDecisions;
+    use crate::inference_capabilities::InferenceCapabilities;
+    use crate::model_serving_settings::ModelServingSettings;
 
     struct MissingLocalModel {
         _dir_guard: TempDir,
         path: PathBuf,
     }
 
+    fn process_env_desired_model_resolver() -> DesiredModelResolver {
+        DesiredModelResolver::from_process_env(HUGGINGFACE_DEFAULT_ENDPOINT.to_owned()).unwrap()
+    }
+
     fn fresh_status() -> Arc<SlotAggregatedStatus> {
         Arc::new(SlotAggregatedStatus::new(1))
     }
@@ -133,11 +232,98 @@ mod tests {
         multimodal_projection: AgentDesiredModel,
     ) -> AgentDesiredState {
         AgentDesiredState {
-            chat_template_override: None,
-            inference_parameters: InferenceParameters::default(),
+            decisions: AgentDecisionsSettings::NotServed,
             model,
-            multimodal_projection,
+            model_runtime_parameters: ModelRuntimeParameters::default(),
+            pooling_type: PoolingType::Last,
+            text_generation: BalancerTextGenerationSettings {
+                multimodal: MultimodalSettings {
+                    projection: multimodal_projection,
+                    ..MultimodalSettings::default()
+                },
+                ..BalancerTextGenerationSettings::default()
+            },
+        }
+    }
+
+    fn desired_state_serving_decisions(
+        model: AgentDesiredModel,
+        pointer_head: &ModelSource,
+    ) -> AgentDesiredState {
+        AgentDesiredState {
+            decisions: AgentDecisionsSettings::Served(ServedDecisionSettings {
+                pointer_head_uri: pointer_head.uri(),
+            }),
+            ..desired_state(model, AgentDesiredModel::None)
+        }
+    }
+
+    fn served_settings(
+        model_file: &NamedTempFile,
+        decisions: ApplicableDecisions,
+        multimodal_projection: MultimodalProjection,
+    ) -> AgentApplicableState {
+        AgentApplicableState::WithModel(ModelServingSettings {
+            capabilities: InferenceCapabilities {
+                decisions,
+                embeddings: PoolingType::Last,
+                text_generation: TextGenerationSettings {
+                    chat_template_source: ChatTemplateSource::default(),
+                    image_resize_to_fit: MultimodalSettings::default().image_resize_to_fit,
+                    multimodal_projection,
+                    sampling_parameters: SamplingParameters::default(),
+                },
+            },
+            model_path: model_file.path().to_path_buf(),
+            model_runtime_parameters: ModelRuntimeParameters::default(),
+        })
+    }
+
+    fn local_source_at(path: &Path) -> ModelSource {
+        ModelSource::LocalToAgent(AgentLocalModelPath::from_absolute_path(path).unwrap())
+    }
+
+    fn local_model_at(path: &Path) -> AgentDesiredModel {
+        local_source_at(path).into_agent_desired_model()
+    }
+
+    fn cancelled_token() -> CancellationToken {
+        let cancellation_token = CancellationToken::new();
+
+        cancellation_token.cancel();
+
+        cancellation_token
+    }
+
+    fn uncached_hugging_face_source() -> ModelSource {
+        ModelSource::HuggingFace(HuggingFaceModelReference {
+            filename: "uncached.gguf".to_owned(),
+            repository_name: "repo".to_owned(),
+            repository_owner: "owner".to_owned(),
+            revision: "main".to_owned(),
+        })
+    }
+
+    fn uncached_hugging_face_model() -> AgentDesiredModel {
+        uncached_hugging_face_source().into_agent_desired_model()
+    }
+
+    async fn conversion(
+        cancellation_token: CancellationToken,
+        desired: &AgentDesiredState,
+    ) -> AgentDesiredStateConversion {
+        AgentDesiredStateConverter {
+            desired_model_resolver: process_env_desired_model_resolver(),
+            cancellation_token,
+            slot_aggregated_status: fresh_status(),
         }
+        .convert(desired)
+        .await
+        .expect("the desired state must convert")
+    }
+
+    async fn converted(desired: &AgentDesiredState) -> AgentDesiredStateConversion {
+        conversion(CancellationToken::new(), desired).await
     }
 
     #[tokio::test]
@@ -147,11 +333,9 @@ mod tests {
             _dir_guard,
             path: missing_path,
         } = nonexistent_path_in_temp_dir("model");
-        let desired = desired_state(
-            AgentDesiredModel::LocalToAgent(missing_path.display().to_string()),
-            AgentDesiredModel::None,
-        );
+        let desired = desired_state(local_model_at(&missing_path), AgentDesiredModel::None);
         let converter = AgentDesiredStateConverter {
+            desired_model_resolver: process_env_desired_model_resolver(),
             cancellation_token: CancellationToken::new(),
             slot_aggregated_status: status.clone(),
         };
@@ -176,19 +360,77 @@ mod tests {
         );
     }
 
+    #[tokio::test]
+    async fn a_model_file_that_appears_on_a_retry_clears_its_missing_file_issue() {
+        let status = fresh_status();
+        let MissingLocalModel {
+            _dir_guard,
+            path: appearing_path,
+        } = nonexistent_path_in_temp_dir("model");
+        let desired = desired_state(local_model_at(&appearing_path), AgentDesiredModel::None);
+        let converter = AgentDesiredStateConverter {
+            desired_model_resolver: process_env_desired_model_resolver(),
+            cancellation_token: CancellationToken::new(),
+            slot_aggregated_status: status.clone(),
+        };
+
+        assert!(converter.convert(&desired).await.is_err());
+
+        write(&appearing_path, b"").unwrap();
+        converter.convert(&desired).await.unwrap();
+
+        assert!(
+            !status.has_issue(&AgentIssue::ModelFileDoesNotExist(ModelPath {
+                model_path: appearing_path.display().to_string(),
+            }))
+        );
+    }
+
+    #[tokio::test]
+    async fn a_pointer_head_that_appears_on_a_retry_clears_its_missing_file_issue() {
+        let status = fresh_status();
+        let model_file = NamedTempFile::new().unwrap();
+        let MissingLocalModel {
+            _dir_guard,
+            path: appearing_path,
+        } = nonexistent_path_in_temp_dir("pointer-head");
+        let desired = desired_state_serving_decisions(
+            local_model_at(model_file.path()),
+            &local_source_at(&appearing_path),
+        );
+        let converter = AgentDesiredStateConverter {
+            desired_model_resolver: process_env_desired_model_resolver(),
+            cancellation_token: CancellationToken::new(),
+            slot_aggregated_status: status.clone(),
+        };
+
+        assert!(converter.convert(&desired).await.is_err());
+
+        write(&appearing_path, b"").unwrap();
+        converter.convert(&desired).await.unwrap();
+
+        assert!(
+            !status.has_issue(&AgentIssue::PointerHeadCannotBeLoaded(ModelPath {
+                model_path: appearing_path.display().to_string(),
+            }))
+        );
+    }
+
     #[tokio::test]
     async fn local_missing_multimodal_projection_registers_multimodal_projection_cannot_be_loaded_and_errs()
      {
         let status = fresh_status();
+        let model_file = NamedTempFile::new().unwrap();
         let MissingLocalModel {
             _dir_guard,
             path: missing_path,
         } = nonexistent_path_in_temp_dir("projection");
         let desired = desired_state(
-            AgentDesiredModel::None,
-            AgentDesiredModel::LocalToAgent(missing_path.display().to_string()),
+            local_model_at(model_file.path()),
+            local_model_at(&missing_path),
         );
         let converter = AgentDesiredStateConverter {
+            desired_model_resolver: process_env_desired_model_resolver(),
             cancellation_token: CancellationToken::new(),
             slot_aggregated_status: status.clone(),
         };
@@ -213,29 +455,155 @@ mod tests {
         );
     }
 
+    #[tokio::test]
+    async fn local_missing_pointer_head_registers_pointer_head_cannot_be_loaded_and_errs() {
+        let status = fresh_status();
+        let model_file = NamedTempFile::new().unwrap();
+        let MissingLocalModel {
+            _dir_guard,
+            path: missing_path,
+        } = nonexistent_path_in_temp_dir("pointer-head");
+        let converter = AgentDesiredStateConverter {
+            desired_model_resolver: process_env_desired_model_resolver(),
+            cancellation_token: CancellationToken::new(),
+            slot_aggregated_status: status.clone(),
+        };
+
+        let outcome = converter
+            .convert(&desired_state_serving_decisions(
+                local_model_at(model_file.path()),
+                &local_source_at(&missing_path),
+            ))
+            .await;
+
+        assert_eq!(
+            outcome.err().as_ref().map(ToString::to_string),
+            Some(
+                ModelSourceError::LocalFileMissing {
+                    path: missing_path.clone(),
+                }
+                .to_string()
+            )
+        );
+        assert!(
+            status.has_issue(&AgentIssue::PointerHeadCannotBeLoaded(ModelPath {
+                model_path: missing_path.display().to_string(),
+            })),
+            "PointerHeadCannotBeLoaded must be registered for a missing local pointer head"
+        );
+    }
+
     #[tokio::test]
     async fn cancelling_while_the_multimodal_projection_resolves_cancels_the_conversion() {
-        let cancellation_token = CancellationToken::new();
-        let desired = desired_state(
-            AgentDesiredModel::None,
-            AgentDesiredModel::HuggingFace(HuggingFaceModelReference {
-                filename: "projection.gguf".to_owned(),
-                repo_id: "owner/repo".to_owned(),
-                revision: "main".to_owned(),
-            }),
+        let model_file = NamedTempFile::new().unwrap();
+
+        assert_eq!(
+            conversion(
+                cancelled_token(),
+                &desired_state(
+                    local_model_at(model_file.path()),
+                    uncached_hugging_face_model()
+                ),
+            )
+            .await,
+            AgentDesiredStateConversion::Cancelled
         );
+    }
 
-        cancellation_token.cancel();
+    #[tokio::test]
+    async fn cancelling_while_the_pointer_head_resolves_cancels_the_conversion() {
+        let model_file = NamedTempFile::new().unwrap();
 
-        let outcome = AgentDesiredStateConverter {
-            cancellation_token,
-            slot_aggregated_status: fresh_status(),
-        }
-        .convert(&desired)
-        .await;
+        assert_eq!(
+            conversion(
+                cancelled_token(),
+                &desired_state_serving_decisions(
+                    local_model_at(model_file.path()),
+                    &uncached_hugging_face_source(),
+                ),
+            )
+            .await,
+            AgentDesiredStateConversion::Cancelled
+        );
+    }
+
+    #[tokio::test]
+    async fn a_state_with_a_model_serves_text_generation_and_embeddings_with_its_resolved_files() {
+        let model_file = NamedTempFile::new().unwrap();
+        let projection_file = NamedTempFile::new().unwrap();
+
+        assert_eq!(
+            converted(&desired_state(
+                local_model_at(model_file.path()),
+                local_model_at(projection_file.path()),
+            ))
+            .await,
+            AgentDesiredStateConversion::Converted(served_settings(
+                &model_file,
+                ApplicableDecisions::Disabled,
+                MultimodalProjection::File(projection_file.path().to_path_buf()),
+            ))
+        );
+    }
 
-        assert!(outcome.is_ok_and(|conversion| {
-            discriminant(&conversion) == discriminant(&AgentDesiredStateConversion::Cancelled)
-        }));
+    #[tokio::test]
+    async fn a_state_without_a_model_keeps_its_chat_template() {
+        let chat_template = ChatTemplate {
+            content: "{{ messages }}".to_owned(),
+        };
+
+        assert_eq!(
+            converted(&AgentDesiredState {
+                text_generation: BalancerTextGenerationSettings {
+                    chat_template_source: ChatTemplateSource::Override(chat_template.clone()),
+                    ..BalancerTextGenerationSettings::default()
+                },
+                ..desired_state(AgentDesiredModel::None, AgentDesiredModel::None)
+            })
+            .await,
+            AgentDesiredStateConversion::Converted(AgentApplicableState::WithoutModel {
+                chat_template_source: ChatTemplateSource::Override(chat_template),
+            })
+        );
+    }
+
+    #[tokio::test]
+    async fn a_state_without_a_model_does_not_resolve_its_multimodal_projection() {
+        let MissingLocalModel {
+            _dir_guard,
+            path: missing_path,
+        } = nonexistent_path_in_temp_dir("projection");
+
+        assert_eq!(
+            converted(&desired_state(
+                AgentDesiredModel::None,
+                local_model_at(&missing_path)
+            ))
+            .await,
+            AgentDesiredStateConversion::Converted(AgentApplicableState::WithoutModel {
+                chat_template_source: ChatTemplateSource::EmbeddedInModel,
+            })
+        );
+    }
+
+    #[tokio::test]
+    async fn a_state_serving_decisions_carries_its_pointer_head() {
+        let model_file = NamedTempFile::new().unwrap();
+        let pointer_head_file = NamedTempFile::new().unwrap();
+
+        assert_eq!(
+            converted(&desired_state_serving_decisions(
+                local_model_at(model_file.path()),
+                &local_source_at(pointer_head_file.path()),
+            ))
+            .await,
+            AgentDesiredStateConversion::Converted(served_settings(
+                &model_file,
+                ApplicableDecisions::Enabled {
+                    pointer_head_path: pointer_head_file.path().to_path_buf(),
+                },
+                MultimodalProjection::NotConfigured,
+            ))
+        );
     }
 }
diff --git a/paddler_agent/src/agent_error.rs b/paddler_agent/src/agent_error.rs
new file mode 100644
index 000000000..bdc49f2db
--- /dev/null
+++ b/paddler_agent/src/agent_error.rs
@@ -0,0 +1,19 @@
+use paddler_agent_decision::decision_error::DecisionError;
+use paddler_agent_embeddings::embedding_error::EmbeddingError;
+use paddler_agent_runtime::agent_runtime_error::AgentRuntimeError;
+use paddler_agent_text_generation::text_generation_error::TextGenerationError;
+
+#[derive(Debug, thiserror::Error)]
+pub enum AgentError {
+    #[error("the decision pipeline failed: {0}")]
+    DecisionPipeline(#[from] DecisionError),
+
+    #[error("the embedding pipeline failed: {0}")]
+    EmbeddingPipeline(#[from] EmbeddingError),
+
+    #[error("the model failed to load: {0}")]
+    ModelLoad(#[from] AgentRuntimeError),
+
+    #[error("the text generation pipeline failed: {0}")]
+    TextGenerationPipeline(#[from] TextGenerationError),
+}
diff --git a/paddler_agent/src/agent_response_message.rs b/paddler_agent/src/agent_response_message.rs
new file mode 100644
index 000000000..2879632a4
--- /dev/null
+++ b/paddler_agent/src/agent_response_message.rs
@@ -0,0 +1,15 @@
+use paddler_messaging::jsonrpc::response_envelope::ResponseEnvelope;
+use paddler_messaging::management_socket::agent::response::Response as JsonRpcResponse;
+use paddler_messaging::management_socket::balancer::message::Message as ManagementJsonRpcMessage;
+
+#[must_use]
+pub const fn agent_response_message(
+    request_id: String,
+    response: JsonRpcResponse,
+) -> ManagementJsonRpcMessage {
+    ManagementJsonRpcMessage::Response(ResponseEnvelope {
+        generated_by: None,
+        request_id,
+        response,
+    })
+}
diff --git a/paddler_agent/src/applicable_decisions.rs b/paddler_agent/src/applicable_decisions.rs
new file mode 100644
index 000000000..8a4b73d0f
--- /dev/null
+++ b/paddler_agent/src/applicable_decisions.rs
@@ -0,0 +1,19 @@
+use std::path::PathBuf;
+
+use paddler_messaging::agent_decisions_serving::AgentDecisionsServing;
+
+#[derive(Clone, Debug, Eq, PartialEq)]
+pub enum ApplicableDecisions {
+    Disabled,
+    Enabled { pointer_head_path: PathBuf },
+}
+
+impl ApplicableDecisions {
+    #[must_use]
+    pub const fn serving(&self) -> AgentDecisionsServing {
+        match self {
+            Self::Disabled => AgentDecisionsServing::NotServed,
+            Self::Enabled { .. } => AgentDecisionsServing::Served,
+        }
+    }
+}
diff --git a/paddler_agent/src/balancer_message_context.rs b/paddler_agent/src/balancer_message_context.rs
index aa5aefa9b..09983ae46 100644
--- a/paddler_agent/src/balancer_message_context.rs
+++ b/paddler_agent/src/balancer_message_context.rs
@@ -1,22 +1,207 @@
 use std::sync::Arc;
 
+use futures_util::StreamExt as _;
+use log::debug;
+use log::error;
+use log::warn;
+use serde_json::from_str;
 use tokio::sync::mpsc;
+use tokio::sync::watch;
+use tokio_tungstenite::tungstenite::protocol::Message;
 
+use paddler_agent_runtime::agent_request::AgentRequest;
+use paddler_agent_runtime::model_metadata_holder::ModelMetadataHolder;
 use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
+use paddler_agent_status::slot_guard::SlotGuard;
 use paddler_messaging::agent_desired_state::AgentDesiredState;
+use paddler_messaging::decision_result::DecisionResult;
+use paddler_messaging::embedding_result::EmbeddingResult;
+use paddler_messaging::generated_token_result::GeneratedTokenResult;
+use paddler_messaging::jsonrpc::request_envelope::RequestEnvelope;
+use paddler_messaging::management_socket::agent::message::Message as JsonRpcMessage;
+use paddler_messaging::management_socket::agent::notification::Notification as JsonRpcNotification;
+use paddler_messaging::management_socket::agent::notification_params::version_params::VersionParams;
+use paddler_messaging::management_socket::agent::request::Request as JsonRpcRequest;
+use paddler_messaging::management_socket::agent::response::Response as JsonRpcResponse;
+use paddler_messaging::reserved_request::ReservedRequest;
+use paddler_outbox::outbox::Outbox;
+use paddler_request_registry::request_delivery::RequestDelivery;
 use paddler_request_registry::request_registry::RequestRegistry;
+use paddler_request_registry::request_registry_guard::RequestRegistryGuard;
 
 use crate::agent_applicable_state_holder::AgentApplicableStateHolder;
-use crate::continuous_batch_preparation_request::ContinuousBatchPreparationRequest;
-use crate::model_metadata_holder::ModelMetadataHolder;
+use crate::agent_response_message::agent_response_message;
+use crate::management_connection_end::ManagementConnectionEnd;
+use crate::management_connection_step::ManagementConnectionStep;
+use crate::pipeline_request::PipelineRequest;
+use crate::pipeline_response_stream::PipelineResponseStream;
 
-#[derive(Clone)]
 pub struct BalancerMessageContext {
-    pub agent_applicable_state_holder: Arc<AgentApplicableStateHolder>,
-    pub agent_desired_state_tx: mpsc::UnboundedSender<AgentDesiredState>,
-    pub continuous_batch_preparation_request_tx:
-        mpsc::UnboundedSender<ContinuousBatchPreparationRequest>,
+    pub agent_desired_state_tx: watch::Sender<AgentDesiredState>,
     pub model_metadata_holder: Arc<ModelMetadataHolder>,
+    pub pipeline_requests: Arc<Outbox<PipelineRequest>>,
     pub request_stoppers: Arc<RequestRegistry<mpsc::UnboundedSender<()>>>,
+    pub served_applicable_state_holder: Arc<AgentApplicableStateHolder>,
     pub slot_aggregated_status: Arc<SlotAggregatedStatus>,
 }
+
+impl BalancerMessageContext {
+    #[must_use]
+    pub fn handle_message(&self, message: Message) -> ManagementConnectionStep {
+        match message {
+            Message::Text(text) => match from_str::<JsonRpcMessage>(&text) {
+                Ok(deserialized_message) => self.handle_deserialized_message(deserialized_message),
+                Err(deserialization_error) => {
+                    error!(
+                        "Closing the management connection after an undeserializable message {text}: {deserialization_error}"
+                    );
+
+                    ManagementConnectionStep::End(
+                        ManagementConnectionEnd::BalancerMessageUndeserializable,
+                    )
+                }
+            },
+            Message::Binary(_) | Message::Frame(_) => {
+                error!("Closing the management connection after a non-text message");
+
+                ManagementConnectionStep::End(ManagementConnectionEnd::BalancerSentUnsupportedFrame)
+            }
+            Message::Close(_) | Message::Ping(_) | Message::Pong(_) => {
+                ManagementConnectionStep::Continue
+            }
+        }
+    }
+
+    fn handle_deserialized_message(
+        &self,
+        deserialized_message: JsonRpcMessage,
+    ) -> ManagementConnectionStep {
+        match deserialized_message {
+            JsonRpcMessage::Notification(JsonRpcNotification::SetState(set_state_params)) => {
+                let desired_state = set_state_params.desired_state;
+
+                self.agent_desired_state_tx
+                    .send_if_modified(|current_desired_state| {
+                        if *current_desired_state == desired_state {
+                            return false;
+                        }
+
+                        *current_desired_state = desired_state;
+
+                        true
+                    });
+
+                ManagementConnectionStep::Continue
+            }
+            JsonRpcMessage::Notification(JsonRpcNotification::StopRespondingTo(request_id)) => {
+                match self.request_stoppers.send_to(&request_id, ()) {
+                    RequestDelivery::Delivered => {}
+                    RequestDelivery::ReceiverDropped | RequestDelivery::RequestNotRegistered => {
+                        debug!("Request {request_id:?} finished before its stop arrived");
+                    }
+                }
+
+                ManagementConnectionStep::Continue
+            }
+            JsonRpcMessage::Notification(JsonRpcNotification::Version(VersionParams {
+                version,
+            })) => {
+                if version != env!("CARGO_PKG_VERSION") {
+                    warn!(
+                        "Version mismatch: server version is {version}, client version is {}",
+                        env!("CARGO_PKG_VERSION")
+                    );
+                }
+
+                ManagementConnectionStep::Continue
+            }
+            JsonRpcMessage::Request(RequestEnvelope {
+                id,
+                request:
+                    JsonRpcRequest::ContinueFromConversationHistory(
+                        continue_from_conversation_history_params,
+                    ),
+            }) => self.stream_responses::<_, GeneratedTokenResult>(
+                id,
+                continue_from_conversation_history_params,
+            ),
+            JsonRpcMessage::Request(RequestEnvelope {
+                id,
+                request: JsonRpcRequest::ContinueFromRawPrompt(continue_from_raw_prompt_params),
+            }) => self
+                .stream_responses::<_, GeneratedTokenResult>(id, continue_from_raw_prompt_params),
+            JsonRpcMessage::Request(RequestEnvelope {
+                id,
+                request: JsonRpcRequest::Decide(decide_params),
+            }) => self.stream_responses::<_, DecisionResult>(id, decide_params),
+            JsonRpcMessage::Request(RequestEnvelope {
+                id,
+                request: JsonRpcRequest::GenerateEmbeddingBatch(generate_embedding_batch_params),
+            }) => self.stream_responses::<_, EmbeddingResult>(id, generate_embedding_batch_params),
+            JsonRpcMessage::Request(RequestEnvelope {
+                id,
+                request: JsonRpcRequest::GetChatTemplateOverride,
+            }) => ManagementConnectionStep::Write(agent_response_message(
+                id,
+                JsonRpcResponse::ChatTemplateOverride(
+                    self.served_applicable_state_holder
+                        .get_agent_applicable_state()
+                        .chat_template_override(),
+                ),
+            )),
+            JsonRpcMessage::Request(RequestEnvelope {
+                id,
+                request: JsonRpcRequest::GetModelMetadata,
+            }) => ManagementConnectionStep::Write(agent_response_message(
+                id,
+                JsonRpcResponse::ModelMetadata(self.model_metadata_holder.get_model_metadata()),
+            )),
+        }
+    }
+
+    fn stream_responses<TParams, TResponse>(
+        &self,
+        request_id: String,
+        ReservedRequest {
+            params,
+            reserved_slots,
+        }: ReservedRequest<TParams>,
+    ) -> ManagementConnectionStep
+    where
+        AgentRequest<TParams, TResponse>: Into<PipelineRequest>,
+        TResponse: Into<JsonRpcResponse> + Send + 'static,
+    {
+        let (response_tx, response_rx) = mpsc::unbounded_channel::<TResponse>();
+        let (stop_tx, stop_rx) = mpsc::unbounded_channel::<()>();
+
+        let Some(stopper_guard) =
+            RequestRegistryGuard::register(&self.request_stoppers, request_id.clone(), stop_tx)
+        else {
+            error!(
+                "The balancer reused the id of request {request_id:?}, which is still in flight"
+            );
+
+            return ManagementConnectionStep::End(
+                ManagementConnectionEnd::BalancerReusedInFlightRequestId,
+            );
+        };
+
+        self.pipeline_requests.push(
+            AgentRequest {
+                params,
+                response_tx,
+                slot_guard: SlotGuard::new(self.slot_aggregated_status.clone(), reserved_slots),
+                stop_rx,
+            }
+            .into(),
+        );
+
+        ManagementConnectionStep::StreamResponses(
+            PipelineResponseStream {
+                response_rx,
+                stopper_guard,
+            }
+            .boxed(),
+        )
+    }
+}
diff --git a/paddler_agent/src/capability_plan.rs b/paddler_agent/src/capability_plan.rs
new file mode 100644
index 000000000..6534bd298
--- /dev/null
+++ b/paddler_agent/src/capability_plan.rs
@@ -0,0 +1,61 @@
+use paddler_agent_runtime::shuts_down::ShutsDown;
+
+pub enum CapabilityPlan<TServed> {
+    Keep(TServed),
+    Start,
+}
+
+impl<TServed> CapabilityPlan<TServed>
+where
+    TServed: ShutsDown + Send,
+{
+    pub async fn retaining<TSettings: PartialEq>(
+        served: TServed,
+        served_settings: &TSettings,
+        desired_settings: &TSettings,
+    ) -> Self {
+        if served_settings == desired_settings {
+            Self::Keep(served)
+        } else {
+            served.shut_down().await;
+
+            Self::Start
+        }
+    }
+}
+
+impl<TServed> ShutsDown for CapabilityPlan<TServed>
+where
+    TServed: ShutsDown + Send,
+{
+    async fn shut_down(self) {
+        match self {
+            Self::Keep(served) => served.shut_down().await,
+            Self::Start => {}
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::mem::discriminant;
+
+    use super::CapabilityPlan;
+    use crate::served_decisions::ServedDecisions;
+
+    #[tokio::test]
+    async fn keeps_a_capability_whose_settings_did_not_change() {
+        assert_eq!(
+            discriminant(&CapabilityPlan::retaining(ServedDecisions::Disabled, &1, &1).await),
+            discriminant(&CapabilityPlan::Keep(ServedDecisions::Disabled))
+        );
+    }
+
+    #[tokio::test]
+    async fn restarts_a_capability_whose_settings_changed() {
+        assert_eq!(
+            discriminant(&CapabilityPlan::retaining(ServedDecisions::Disabled, &1, &2).await),
+            discriminant(&CapabilityPlan::<ServedDecisions>::Start)
+        );
+    }
+}
diff --git a/paddler_agent/src/capability_starter.rs b/paddler_agent/src/capability_starter.rs
new file mode 100644
index 000000000..70e96d064
--- /dev/null
+++ b/paddler_agent/src/capability_starter.rs
@@ -0,0 +1,253 @@
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent_decision::decision_pipeline::DecisionPipeline;
+use paddler_agent_decision::decision_prerequisites::DecisionPrerequisites;
+use paddler_agent_embeddings::embedding_pipeline::EmbeddingPipeline;
+use paddler_agent_runtime::inference_runtime_context::InferenceRuntimeContext;
+use paddler_agent_runtime::llama_model_settings::LlamaModelSettings;
+use paddler_agent_runtime::loaded_llama_model::LoadedLlamaModel;
+use paddler_agent_runtime::shuts_down::ShutsDown as _;
+use paddler_agent_runtime::startup_outcome::StartupOutcome;
+use paddler_agent_text_generation::text_generation_pipeline::TextGenerationPipeline;
+use paddler_inference_parameters::model_runtime_parameters::ModelRuntimeParameters;
+use paddler_spawn_resuming_panic::spawn_blocking_resuming_panic::spawn_blocking_resuming_panic;
+
+use crate::agent_error::AgentError;
+use crate::applicable_decisions::ApplicableDecisions;
+use crate::capability_plan::CapabilityPlan;
+use crate::capability_startup_interruption::CapabilityStartupInterruption;
+use crate::inference_capabilities::InferenceCapabilities;
+use crate::loaded_model_serving_plan::LoadedModelServingPlan;
+use crate::model_serving_settings::ModelServingSettings;
+use crate::prepared_decisions::PreparedDecisions;
+use crate::served_capabilities::ServedCapabilities;
+use crate::served_decisions::ServedDecisions;
+use crate::served_embeddings::ServedEmbeddings;
+use crate::served_model::ServedModel;
+use crate::served_text_generation::ServedTextGeneration;
+use crate::serving_interruption::ServingInterruption;
+
+trait IntoEstablishedCapability<TStarted> {
+    fn into_established_capability(self) -> Result<TStarted, CapabilityStartupInterruption>;
+}
+
+impl<TStarted, TError> IntoEstablishedCapability<TStarted>
+    for Result<StartupOutcome<TStarted>, TError>
+where
+    AgentError: From<TError>,
+{
+    fn into_established_capability(self) -> Result<TStarted, CapabilityStartupInterruption> {
+        match self {
+            Ok(StartupOutcome::Ready(started)) => Ok(started),
+            Ok(StartupOutcome::Cancelled) => Err(CapabilityStartupInterruption::Cancelled),
+            Err(startup_error) => Err(CapabilityStartupInterruption::Failed(AgentError::from(
+                startup_error,
+            ))),
+        }
+    }
+}
+
+pub struct CapabilityStarter<'startup> {
+    pub cancellation_token: &'startup CancellationToken,
+    pub inference_runtime_context: &'startup InferenceRuntimeContext,
+}
+
+impl CapabilityStarter<'_> {
+    pub async fn prepare_decisions(
+        &self,
+        applicable_decisions: &ApplicableDecisions,
+    ) -> Result<PreparedDecisions, CapabilityStartupInterruption> {
+        match applicable_decisions {
+            ApplicableDecisions::Disabled => Ok(PreparedDecisions::Disabled),
+            ApplicableDecisions::Enabled { pointer_head_path } => {
+                let decision_prerequisites = DecisionPrerequisites {
+                    inference_runtime_context: self.inference_runtime_context.clone(),
+                    pointer_head_path: pointer_head_path.clone(),
+                };
+                let cancellation_token = self.cancellation_token.clone();
+
+                spawn_blocking_resuming_panic(move || {
+                    decision_prerequisites
+                        .load_pointer_head()
+                        .map(|loaded_pointer_head| {
+                            if cancellation_token.is_cancelled() {
+                                StartupOutcome::Cancelled
+                            } else {
+                                StartupOutcome::Ready(loaded_pointer_head)
+                            }
+                        })
+                })
+                .await
+                .into_established_capability()
+                .map(|loaded_pointer_head| {
+                    PreparedDecisions::Enabled(Box::new(loaded_pointer_head))
+                })
+            }
+        }
+    }
+
+    pub async fn load_model(
+        &self,
+        llama_model_settings: LlamaModelSettings,
+    ) -> Result<LoadedLlamaModel, CapabilityStartupInterruption> {
+        LoadedLlamaModel::load_unless_shutting_down(
+            self.cancellation_token,
+            self.inference_runtime_context.clone(),
+            llama_model_settings,
+        )
+        .await
+        .into_established_capability()
+    }
+
+    pub async fn serve(
+        &self,
+        LoadedModelServingPlan {
+            capability_plans:
+                InferenceCapabilities {
+                    decisions: decisions_plan,
+                    embeddings: embeddings_plan,
+                    text_generation: text_generation_plan,
+                },
+            loaded_llama_model,
+            model_serving_settings,
+        }: LoadedModelServingPlan,
+        prepared_decisions: PreparedDecisions,
+    ) -> Result<ServedModel, Box<ServingInterruption>> {
+        let decisions = match self
+            .establish_decisions(
+                &loaded_llama_model,
+                &model_serving_settings.model_runtime_parameters,
+                decisions_plan,
+                prepared_decisions,
+            )
+            .await
+        {
+            Ok(decisions) => decisions,
+            Err(interruption) => {
+                embeddings_plan.shut_down().await;
+                text_generation_plan.shut_down().await;
+
+                return Err(Box::new(ServingInterruption {
+                    interruption,
+                    loaded_llama_model,
+                }));
+            }
+        };
+        let embeddings = match self
+            .establish_embeddings(
+                &loaded_llama_model,
+                &model_serving_settings,
+                embeddings_plan,
+            )
+            .await
+        {
+            Ok(embeddings) => embeddings,
+            Err(interruption) => {
+                decisions.shut_down().await;
+                text_generation_plan.shut_down().await;
+
+                return Err(Box::new(ServingInterruption {
+                    interruption,
+                    loaded_llama_model,
+                }));
+            }
+        };
+        let text_generation = match self
+            .establish_text_generation(
+                &loaded_llama_model,
+                &model_serving_settings,
+                text_generation_plan,
+            )
+            .await
+        {
+            Ok(text_generation) => text_generation,
+            Err(interruption) => {
+                decisions.shut_down().await;
+                embeddings.shut_down().await;
+
+                return Err(Box::new(ServingInterruption {
+                    interruption,
+                    loaded_llama_model,
+                }));
+            }
+        };
+
+        Ok(ServedModel {
+            capabilities: ServedCapabilities {
+                decisions,
+                embeddings,
+                text_generation,
+            },
+            loaded_llama_model,
+            model_serving_settings,
+        })
+    }
+
+    async fn establish_decisions(
+        &self,
+        loaded_llama_model: &LoadedLlamaModel,
+        model_runtime_parameters: &ModelRuntimeParameters,
+        decisions_plan: CapabilityPlan<ServedDecisions>,
+        prepared_decisions: PreparedDecisions,
+    ) -> Result<ServedDecisions, CapabilityStartupInterruption> {
+        match (decisions_plan, prepared_decisions) {
+            (CapabilityPlan::Keep(served_decisions), _) => Ok(served_decisions),
+            (CapabilityPlan::Start, PreparedDecisions::Disabled) => Ok(ServedDecisions::Disabled),
+            (CapabilityPlan::Start, PreparedDecisions::Enabled(loaded_pointer_head)) => {
+                DecisionPipeline {
+                    inference_runtime_context: self.inference_runtime_context.clone(),
+                    loaded_llama_model: loaded_llama_model.clone(),
+                    loaded_pointer_head: *loaded_pointer_head,
+                    model_runtime_parameters: model_runtime_parameters.clone(),
+                }
+                .spawn(self.cancellation_token)
+                .await
+                .into_established_capability()
+                .map(ServedDecisions::Enabled)
+            }
+        }
+    }
+
+    async fn establish_embeddings(
+        &self,
+        loaded_llama_model: &LoadedLlamaModel,
+        model_serving_settings: &ModelServingSettings,
+        embeddings_plan: CapabilityPlan<ServedEmbeddings>,
+    ) -> Result<ServedEmbeddings, CapabilityStartupInterruption> {
+        match embeddings_plan {
+            CapabilityPlan::Keep(embeddings) => Ok(embeddings),
+            CapabilityPlan::Start => EmbeddingPipeline {
+                inference_runtime_context: self.inference_runtime_context.clone(),
+                loaded_llama_model: loaded_llama_model.clone(),
+                model_runtime_parameters: model_serving_settings.model_runtime_parameters.clone(),
+                pooling_type: model_serving_settings.capabilities.embeddings.clone(),
+            }
+            .spawn(self.cancellation_token)
+            .await
+            .into_established_capability(),
+        }
+    }
+
+    async fn establish_text_generation(
+        &self,
+        loaded_llama_model: &LoadedLlamaModel,
+        model_serving_settings: &ModelServingSettings,
+        text_generation_plan: CapabilityPlan<ServedTextGeneration>,
+    ) -> Result<ServedTextGeneration, CapabilityStartupInterruption> {
+        match text_generation_plan {
+            CapabilityPlan::Keep(text_generation) => Ok(text_generation),
+            CapabilityPlan::Start => TextGenerationPipeline {
+                inference_runtime_context: self.inference_runtime_context.clone(),
+                loaded_llama_model: loaded_llama_model.clone(),
+                model_runtime_parameters: model_serving_settings.model_runtime_parameters.clone(),
+                text_generation_settings: model_serving_settings
+                    .capabilities
+                    .text_generation
+                    .clone(),
+            }
+            .spawn(self.cancellation_token)
+            .await
+            .into_established_capability(),
+        }
+    }
+}
diff --git a/paddler_agent/src/capability_startup_interruption.rs b/paddler_agent/src/capability_startup_interruption.rs
new file mode 100644
index 000000000..086f7a9b6
--- /dev/null
+++ b/paddler_agent/src/capability_startup_interruption.rs
@@ -0,0 +1,7 @@
+use crate::agent_error::AgentError;
+
+#[derive(Debug)]
+pub enum CapabilityStartupInterruption {
+    Cancelled,
+    Failed(AgentError),
+}
diff --git a/paddler_agent/src/chat_prompt_render_request.rs b/paddler_agent/src/chat_prompt_render_request.rs
deleted file mode 100644
index df6981c4d..000000000
--- a/paddler_agent/src/chat_prompt_render_request.rs
+++ /dev/null
@@ -1,10 +0,0 @@
-use paddler_messaging::chat_template_message::ChatTemplateMessage;
-use paddler_messaging::request_params::continue_from_conversation_history_params::tool::Tool;
-use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters_schema::validated_parameters_schema::ValidatedParametersSchema;
-
-pub struct ChatPromptRenderRequest<'request> {
-    pub add_generation_prompt: bool,
-    pub enable_thinking: bool,
-    pub messages: &'request [ChatTemplateMessage],
-    pub tools: &'request [Tool<ValidatedParametersSchema>],
-}
diff --git a/paddler_agent/src/chat_template_load_status.rs b/paddler_agent/src/chat_template_load_status.rs
deleted file mode 100644
index 1852f5004..000000000
--- a/paddler_agent/src/chat_template_load_status.rs
+++ /dev/null
@@ -1,4 +0,0 @@
-pub enum ChatTemplateLoadStatus {
-    Loaded,
-    SkippedForEmbeddings,
-}
diff --git a/paddler_agent/src/chat_template_renderer/mod.rs b/paddler_agent/src/chat_template_renderer/mod.rs
deleted file mode 100644
index 4023ec51b..000000000
--- a/paddler_agent/src/chat_template_renderer/mod.rs
+++ /dev/null
@@ -1,176 +0,0 @@
-pub mod pyjinja_tojson;
-pub mod raise_exception;
-
-use minijinja::Environment;
-use minijinja::Error;
-use minijinja_contrib::add_to_environment;
-use minijinja_contrib::pycompat::unknown_method_callback;
-use serde::ser::Serialize;
-
-use paddler_messaging::chat_template::ChatTemplate;
-
-use self::pyjinja_tojson::pyjinja_tojson;
-use self::raise_exception::raise_exception;
-
-const CHAT_TEMPLATE_NAME: &str = "chat_template";
-
-pub struct ChatTemplateRenderer {
-    minijinja_env: Environment<'static>,
-}
-
-impl ChatTemplateRenderer {
-    pub fn new(ChatTemplate { content }: ChatTemplate) -> Result<Self, Error> {
-        let mut minijinja_env = Environment::new();
-
-        minijinja_env.add_function("raise_exception", raise_exception);
-        minijinja_env.add_template_owned(CHAT_TEMPLATE_NAME, content)?;
-        minijinja_env.set_unknown_method_callback(unknown_method_callback);
-
-        add_to_environment(&mut minijinja_env);
-        minijinja_env.add_filter("tojson", pyjinja_tojson);
-
-        Ok(Self { minijinja_env })
-    }
-
-    pub fn render<TContext: Serialize>(&self, context: TContext) -> Result<String, Error> {
-        self.minijinja_env
-            .get_template(CHAT_TEMPLATE_NAME)
-            .and_then(|template| template.render(context))
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use std::collections::HashMap;
-
-    use minijinja::ErrorKind;
-    use minijinja::context;
-
-    use paddler_messaging::chat_template::ChatTemplate;
-    use paddler_messaging::chat_template_message::ChatTemplateMessage;
-    use paddler_messaging::chat_template_message_content::ChatTemplateMessageContent;
-
-    use crate::chat_template_renderer::ChatTemplateRenderer;
-
-    #[test]
-    fn new_with_valid_template_succeeds() {
-        let template = ChatTemplate {
-            content: "Hello {{ name }}!".to_owned(),
-        };
-
-        assert!(ChatTemplateRenderer::new(template).is_ok());
-    }
-
-    #[test]
-    fn new_with_invalid_template_fails() {
-        let template = ChatTemplate {
-            content: "{% if unclosed %}".to_owned(),
-        };
-
-        assert!(
-            ChatTemplateRenderer::new(template)
-                .is_err_and(|compile_error| compile_error.kind() == ErrorKind::SyntaxError)
-        );
-    }
-
-    #[test]
-    fn renders_the_context_values_into_the_template() {
-        let template = ChatTemplate {
-            content: "Hello {{ name }}!".to_owned(),
-        };
-        let renderer = ChatTemplateRenderer::new(template).unwrap();
-        let mut context = HashMap::new();
-        context.insert("name", "world");
-
-        let result = renderer.render(context).unwrap();
-
-        assert_eq!(result, "Hello world!");
-    }
-
-    #[test]
-    fn renders_messages_loop_with_roles() {
-        let template = ChatTemplate {
-            content: "{% for message in messages %}{{ message.role }}:{{ message.content }}\n{% endfor %}".to_owned(),
-        };
-        let renderer = ChatTemplateRenderer::new(template).unwrap();
-        let messages = vec![
-            ChatTemplateMessage {
-                content: ChatTemplateMessageContent::Text("hi".to_owned()),
-                role: "user".to_owned(),
-            },
-            ChatTemplateMessage {
-                content: ChatTemplateMessageContent::Text("hello".to_owned()),
-                role: "assistant".to_owned(),
-            },
-        ];
-
-        let result = renderer.render(context! { messages => messages }).unwrap();
-
-        assert_eq!(result, "user:hi\nassistant:hello\n");
-    }
-
-    #[test]
-    fn add_generation_prompt_branch_changes_output() {
-        let template = ChatTemplate {
-            content: "A{% if add_generation_prompt %}B{% endif %}".to_owned(),
-        };
-        let renderer = ChatTemplateRenderer::new(template).unwrap();
-
-        let with_prompt = renderer
-            .render(context! { add_generation_prompt => true })
-            .unwrap();
-        let without_prompt = renderer
-            .render(context! { add_generation_prompt => false })
-            .unwrap();
-
-        assert_eq!(with_prompt, "AB");
-        assert_eq!(without_prompt, "A");
-    }
-
-    #[test]
-    fn registers_pyjinja_tojson_filter() {
-        let template = ChatTemplate {
-            content: "{{ value | tojson(ensure_ascii=False) }}".to_owned(),
-        };
-        let renderer = ChatTemplateRenderer::new(template).unwrap();
-
-        let result = renderer.render(context! { value => "café" }).unwrap();
-
-        assert_eq!(result, "\"café\"");
-    }
-
-    #[test]
-    fn registers_raise_exception_function() {
-        let template = ChatTemplate {
-            content: "{{ raise_exception('boom') }}".to_owned(),
-        };
-        let template_renderer = ChatTemplateRenderer::new(template).unwrap();
-
-        let render_error = template_renderer
-            .render(context! {})
-            .expect_err("raise_exception must turn rendering into an error");
-
-        assert_eq!(
-            render_error.detail(),
-            Some("Model's chat template raised an exception: 'boom'")
-        );
-    }
-
-    #[test]
-    fn render_fails_when_template_is_not_registered() {
-        let template = ChatTemplate {
-            content: "Hello {{ name }}!".to_owned(),
-        };
-        let mut renderer = ChatTemplateRenderer::new(template).unwrap();
-
-        renderer
-            .minijinja_env
-            .remove_template(super::CHAT_TEMPLATE_NAME);
-
-        let render_error = renderer
-            .render(context! {})
-            .expect_err("rendering must fail when the template is missing");
-
-        assert_eq!(render_error.kind(), ErrorKind::TemplateNotFound);
-    }
-}
diff --git a/paddler_agent/src/continuous_batch_arbiter.rs b/paddler_agent/src/continuous_batch_arbiter.rs
deleted file mode 100644
index 32b0656b4..000000000
--- a/paddler_agent/src/continuous_batch_arbiter.rs
+++ /dev/null
@@ -1,525 +0,0 @@
-use std::cmp::max;
-use std::path::PathBuf;
-use std::sync::Arc;
-use std::sync::mpsc::channel;
-use std::thread;
-use std::thread::available_parallelism;
-
-use anyhow::Context as _;
-use anyhow::Result;
-use encoding_rs::Decoder;
-use encoding_rs::UTF_8;
-use llama_cpp_bindings::SampledToken;
-use llama_cpp_bindings::context::LlamaContext;
-use llama_cpp_bindings::context::params::LlamaContextParams;
-use llama_cpp_bindings::error::TokenToStringError;
-use llama_cpp_bindings::llama_backend::LlamaBackend;
-use llama_cpp_bindings::llama_batch::LlamaBatch;
-use llama_cpp_bindings::model::LlamaModel;
-use llama_cpp_bindings::model::params::LlamaModelParams;
-use llama_cpp_bindings::mtmd::MtmdContext;
-use llama_cpp_bindings::mtmd::MtmdContextParams;
-use llama_cpp_bindings::mtmd::micro_batch_tokens;
-use llama_cpp_bindings::mtmd::mtmd_default_marker;
-use llama_cpp_bindings::token::LlamaToken;
-use llama_cpp_bindings_sys::LLAMA_FLASH_ATTN_TYPE_AUTO;
-use log::debug;
-use log::error;
-use log::info;
-use log::warn;
-use tokio::sync::oneshot;
-use tokio::task::spawn_blocking;
-use tokio_util::sync::CancellationToken;
-use tokio_util::task::TaskTracker;
-
-use paddler_agent_status::agent_issue_fix::AgentIssueFix;
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
-use paddler_messaging::agent_issue::AgentIssue;
-use paddler_messaging::agent_issue_params::chat_template_does_not_compile_params::ChatTemplateDoesNotCompileParams;
-use paddler_messaging::agent_issue_params::model_path::ModelPath;
-use paddler_messaging::agent_issue_params::slot_cannot_start_params::SlotCannotStartParams;
-use paddler_messaging::chat_template::ChatTemplate;
-use paddler_messaging::media_marker::MediaMarker;
-use paddler_messaging::model_metadata::ModelMetadata;
-
-use crate::agent_kv_cache_dtype::AgentKvCacheDtype;
-use crate::agent_pooling_type::AgentPoolingType;
-use crate::chat_prompt_renderer::ChatPromptRenderer;
-use crate::chat_template_load_status::ChatTemplateLoadStatus;
-use crate::chat_template_renderer::ChatTemplateRenderer;
-use crate::continuous_batch_arbiter_context::ContinuousBatchArbiterContext;
-use crate::continuous_batch_arbiter_handle::ContinuousBatchArbiterHandle;
-use crate::continuous_batch_arbiter_spawn_outcome::ContinuousBatchArbiterSpawnOutcome;
-use crate::continuous_batch_request_preparer::ContinuousBatchRequestPreparer;
-use crate::continuous_batch_scheduler::ContinuousBatchScheduler;
-use crate::continuous_batch_scheduler_context::ContinuousBatchSchedulerContext;
-use crate::continuous_batch_scheduler_params::ContinuousBatchSchedulerParams;
-use crate::converts_to_llama_kv_cache_dtype::ConvertsToLlamaKvCacheDtype;
-use crate::converts_to_llama_pooling_type::ConvertsToLlamaPoolingType;
-use crate::embedding_batch_preparer::EmbeddingBatchPreparer;
-use crate::generation_request_preparer::GenerationRequestPreparer;
-use crate::image_input::ImageInput;
-use crate::join_scheduler_thread::join_scheduler_thread;
-use crate::multimodal_prompt_support::MultimodalPromptSupport;
-use crate::prompt_tokenizer::PromptTokenizer;
-use crate::sampler_chain_factory::SamplerChainFactory;
-use crate::send_startup_signal::send_startup_signal;
-use crate::startup_signal_delivery::StartupSignalDelivery;
-use crate::token_generation::TokenGeneration;
-use crate::token_generation_support::TokenGenerationSupport;
-
-const LOGICAL_CORES_PER_PHYSICAL_CORE: i32 = 2;
-const MINIMUM_THREAD_COUNT: i32 = 2;
-
-fn special_token_text(
-    model: &LlamaModel,
-    token: LlamaToken,
-    decoder: &mut Decoder,
-) -> Result<String, TokenToStringError> {
-    model.token_to_piece(&SampledToken::Content(token), decoder, true, None)
-}
-
-pub struct ContinuousBatchArbiter {
-    pub chat_template_override: Option<ChatTemplate>,
-    pub context: ContinuousBatchArbiterContext,
-    pub inference_parameters: InferenceParameters,
-    pub model_path: PathBuf,
-    pub multimodal_projection_path: Option<PathBuf>,
-}
-
-impl ContinuousBatchArbiter {
-    pub async fn spawn(
-        &self,
-        cancellation_token: &CancellationToken,
-    ) -> Result<ContinuousBatchArbiterSpawnOutcome> {
-        if cancellation_token.is_cancelled() {
-            return Ok(ContinuousBatchArbiterSpawnOutcome::Cancelled);
-        }
-
-        let desired_slots_total = self.context.slot_aggregated_status.desired_slots_total;
-        let n_seq_max = u32::from(desired_slots_total);
-        let (chat_template_loaded_tx, chat_template_loaded_rx) =
-            oneshot::channel::<ChatTemplateLoadStatus>();
-        let (model_loaded_tx, model_loaded_rx) = oneshot::channel::<()>();
-        let (agent_warm_and_scheduler_running_tx, agent_warm_and_scheduler_running_rx) =
-            oneshot::channel::<Arc<ContinuousBatchRequestPreparer>>();
-
-        let available_parallelism_value: i32 = available_parallelism()?.get().try_into()?;
-        let thread_count = max(
-            MINIMUM_THREAD_COUNT,
-            available_parallelism_value / LOGICAL_CORES_PER_PHYSICAL_CORE,
-        );
-
-        info!("Using {thread_count} threads for generation and for batch processing");
-
-        let agent_name_clone = self.context.agent_name.clone();
-        let inference_parameters = self.inference_parameters.clone();
-        let model_metadata_holder = self.context.model_metadata_holder.clone();
-        let multimodal_projection_path = self.multimodal_projection_path.clone();
-        let model_path = self.model_path.clone();
-        let model_issue_path = ModelPath::from(self.model_path.as_path());
-        let chat_template_override = self.chat_template_override.clone();
-        let slot_aggregated_status = self.context.slot_aggregated_status.clone();
-        let agent_shutdown = cancellation_token.clone();
-
-        let scheduler_thread_handle = thread::spawn(move || -> Result<()> {
-            let llama_backend =
-                Arc::new(LlamaBackend::init().context("Unable to initialize llama.cpp backend")?);
-
-            let n_batch_tokens = inference_parameters.n_batch.tokens().get();
-
-            let default_context_params = LlamaContextParams::default();
-            let n_ubatch = if inference_parameters.enable_embeddings {
-                n_batch_tokens
-            } else {
-                default_context_params.n_ubatch()
-            };
-
-            let context_params = default_context_params
-                .with_embeddings(inference_parameters.enable_embeddings)
-                .with_n_ctx(Some(inference_parameters.context_size))
-                .with_n_batch(n_batch_tokens)
-                .with_n_ubatch(n_ubatch)
-                .with_flash_attention_policy(LLAMA_FLASH_ATTN_TYPE_AUTO)
-                .with_n_seq_max(n_seq_max)
-                .with_n_threads(thread_count)
-                .with_n_threads_batch(thread_count)
-                .with_pooling_type(
-                    AgentPoolingType(inference_parameters.pooling_type.clone())
-                        .to_llama_pooling_type(),
-                )
-                .with_type_k(
-                    AgentKvCacheDtype(inference_parameters.k_cache_dtype.clone())
-                        .to_llama_kv_cache_dtype(),
-                )
-                .with_type_v(
-                    AgentKvCacheDtype(inference_parameters.v_cache_dtype.clone())
-                        .to_llama_kv_cache_dtype(),
-                );
-
-            let model = Arc::new(
-                LlamaModel::load_from_file(
-                    &llama_backend,
-                    model_path.clone(),
-                    &LlamaModelParams::default()
-                        .with_n_gpu_layers(inference_parameters.n_gpu_layers),
-                )
-                .context("Unable to load model from file")?,
-            );
-
-            if matches!(
-                send_startup_signal(model_loaded_tx, ()),
-                StartupSignalDelivery::Abandoned
-            ) {
-                return Ok(());
-            }
-
-            let mut model_metadata = ModelMetadata::default();
-
-            for metadata_index in 0..model.meta_count() {
-                model_metadata.set_meta_field(
-                    model.meta_key_by_index(metadata_index)?,
-                    model.meta_val_str_by_index(metadata_index)?,
-                );
-            }
-
-            model_metadata_holder.set_model_metadata(model_metadata);
-
-            let token_generation = if inference_parameters.enable_embeddings
-                && chat_template_override.is_none()
-            {
-                if matches!(
-                    send_startup_signal(
-                        chat_template_loaded_tx,
-                        ChatTemplateLoadStatus::SkippedForEmbeddings,
-                    ),
-                    StartupSignalDelivery::Abandoned
-                ) {
-                    return Ok(());
-                }
-
-                TokenGeneration::DisabledForEmbeddings
-            } else {
-                let llama_chat_template_string = match chat_template_override {
-                    Some(chat_template) => chat_template.content,
-                    None => model
-                        .chat_template(None)
-                        .context(format!(
-                            "Failed to load chat template for model at path: {}",
-                            model_path.display()
-                        ))?
-                        .to_string()?,
-                };
-
-                if matches!(
-                    send_startup_signal(chat_template_loaded_tx, ChatTemplateLoadStatus::Loaded),
-                    StartupSignalDelivery::Abandoned
-                ) {
-                    return Ok(());
-                }
-
-                let chat_template_renderer = match ChatTemplateRenderer::new(ChatTemplate {
-                    content: llama_chat_template_string.clone(),
-                })
-                .context("Failed to create chat template renderer")
-                {
-                    Ok(renderer) => {
-                        slot_aggregated_status.register_fix(
-                            &AgentIssueFix::ChatTemplateIsCompiled(model_issue_path.clone()),
-                        );
-
-                        renderer
-                    }
-                    Err(err) => {
-                        slot_aggregated_status.register_issue(
-                            AgentIssue::ChatTemplateDoesNotCompile(
-                                ChatTemplateDoesNotCompileParams {
-                                    error: format!("{err}"),
-                                    model_path: model_issue_path.clone(),
-                                    template_content: llama_chat_template_string,
-                                },
-                            ),
-                        );
-
-                        return Err(err);
-                    }
-                };
-
-                let mut special_token_decoder = UTF_8.new_decoder();
-
-                TokenGeneration::Enabled(Box::new(TokenGenerationSupport {
-                    chat_prompt_renderer: ChatPromptRenderer {
-                        chat_template_renderer,
-                        media_marker: MediaMarker::new(mtmd_default_marker()?.to_owned()),
-                        token_bos_str: special_token_text(
-                            &model,
-                            model.token_bos(),
-                            &mut special_token_decoder,
-                        )?,
-                        token_eos_str: special_token_text(
-                            &model,
-                            model.token_eos(),
-                            &mut special_token_decoder,
-                        )?,
-                        token_nl_str: special_token_text(
-                            &model,
-                            model.token_nl(),
-                            &mut special_token_decoder,
-                        )?,
-                    },
-                    streaming_markers: model.streaming_markers()?,
-                }))
-            };
-
-            slot_aggregated_status.set_model_path(Some(model_issue_path.model_path.clone()));
-
-            let mut llama_context =
-                match LlamaContext::from_model(&model, &llama_backend, context_params)
-                    .context("Unable to create llama.cpp context")
-                {
-                    Ok(context) => context,
-                    Err(err) => {
-                        for slot_index in 0..n_seq_max {
-                            slot_aggregated_status.register_issue(AgentIssue::SlotCannotStart(
-                                SlotCannotStartParams {
-                                    error: format!("{err:#}"),
-                                    slot_index,
-                                },
-                            ));
-                        }
-
-                        return Err(err);
-                    }
-                };
-
-            let sequence_context_size = llama_context.n_ctx_seq();
-
-            let image_input = match multimodal_projection_path {
-                Some(multimodal_projection_path) => {
-                    let multimodal_projection_path_str =
-                        multimodal_projection_path.to_string_lossy();
-
-                    match MtmdContext::init_from_file(
-                        &multimodal_projection_path_str,
-                        &model,
-                        &MtmdContextParams::default(),
-                    ) {
-                        Ok(mtmd_context) => {
-                            slot_aggregated_status.register_fix(
-                                &AgentIssueFix::MultimodalProjectionIsLoaded(ModelPath::from(
-                                    multimodal_projection_path.as_path(),
-                                )),
-                            );
-
-                            info!(
-                                "Multimodal context initialized from: {}",
-                                multimodal_projection_path.display()
-                            );
-
-                            ImageInput::Supported(MultimodalPromptSupport {
-                                micro_batch_tokens: micro_batch_tokens(
-                                    &llama_context,
-                                    inference_parameters.n_batch.tokens(),
-                                ),
-                                multimodal_context: Arc::new(mtmd_context),
-                                n_batch: inference_parameters.n_batch.tokens_i32(),
-                                sequence_context_size,
-                            })
-                        }
-                        Err(err) => {
-                            slot_aggregated_status.register_issue(
-                                AgentIssue::MultimodalProjectionCannotBeLoaded(ModelPath::from(
-                                    multimodal_projection_path.as_path(),
-                                )),
-                            );
-
-                            return Err(err.into());
-                        }
-                    }
-                }
-                None => ImageInput::Unsupported,
-            };
-
-            let (scheduler_command_tx, scheduler_command_rx) = channel();
-
-            let request_preparer = Arc::new(ContinuousBatchRequestPreparer {
-                agent_name: agent_name_clone.clone(),
-                embedding_batch_preparer: EmbeddingBatchPreparer {
-                    enable_embeddings: inference_parameters.enable_embeddings,
-                    model: model.clone(),
-                    n_batch: inference_parameters.n_batch.tokens_usize(),
-                },
-                generation_request_preparer: GenerationRequestPreparer {
-                    image_input,
-                    image_resize_to_fit: inference_parameters.image_resize_to_fit,
-                    model: model.clone(),
-                    prompt_tokenizer: PromptTokenizer {
-                        model: model.clone(),
-                        sequence_context_size,
-                    },
-                    sampler_chain_factory: SamplerChainFactory {
-                        inference_parameters: inference_parameters.clone(),
-                        n_vocab: model.n_vocab(),
-                    },
-                    token_generation,
-                },
-                llama_backend: llama_backend.clone(),
-                preparation_tasks: TaskTracker::new(),
-                scheduler_command_tx,
-            });
-
-            let mut batch = LlamaBatch::new(
-                inference_parameters.n_batch.tokens_usize(),
-                i32::from(desired_slots_total),
-            )?;
-
-            Self::run_warmup_decode(&model, &mut llama_context, &mut batch, desired_slots_total);
-
-            let mut scheduler = ContinuousBatchScheduler::new(ContinuousBatchSchedulerParams {
-                agent_shutdown,
-                batch,
-                command_rx: scheduler_command_rx,
-                llama_context,
-                scheduler_context: ContinuousBatchSchedulerContext {
-                    agent_name: agent_name_clone,
-                    desired_slots_total,
-                    inference_parameters,
-                    model: model.clone(),
-                },
-            });
-
-            if matches!(
-                send_startup_signal(agent_warm_and_scheduler_running_tx, request_preparer),
-                StartupSignalDelivery::Abandoned
-            ) {
-                return Ok(());
-            }
-
-            scheduler.run();
-
-            Ok(())
-        });
-
-        let Some(startup_result) = cancellation_token
-            .run_until_cancelled(self.await_startup_signals(
-                model_loaded_rx,
-                chat_template_loaded_rx,
-                agent_warm_and_scheduler_running_rx,
-            ))
-            .await
-        else {
-            if let Err(err) = spawn_blocking(move || join_scheduler_thread(scheduler_thread_handle))
-                .await
-                .context("Failed to join the scheduler shutdown task")?
-            {
-                debug!("Scheduler thread ended while its spawn was being cancelled: {err}");
-            }
-
-            return Ok(ContinuousBatchArbiterSpawnOutcome::Cancelled);
-        };
-
-        let request_preparer = match startup_result {
-            Ok(request_preparer) => request_preparer,
-            Err(startup_error) => {
-                return spawn_blocking(move || join_scheduler_thread(scheduler_thread_handle))
-                    .await
-                    .context("Failed to join the scheduler thread after its startup failed")?
-                    .and(Err(startup_error));
-            }
-        };
-
-        for slot_index in 0..n_seq_max {
-            self.context.slot_aggregated_status.increment_total_slots();
-
-            self.context
-                .slot_aggregated_status
-                .register_fix(&AgentIssueFix::SlotStarted(slot_index));
-        }
-
-        Ok(ContinuousBatchArbiterSpawnOutcome::Ready(
-            ContinuousBatchArbiterHandle {
-                request_preparer,
-                scheduler_thread_handle,
-            },
-        ))
-    }
-
-    async fn await_startup_signals(
-        &self,
-        model_loaded_rx: oneshot::Receiver<()>,
-        chat_template_loaded_rx: oneshot::Receiver<ChatTemplateLoadStatus>,
-        agent_warm_and_scheduler_running_rx: oneshot::Receiver<Arc<ContinuousBatchRequestPreparer>>,
-    ) -> Result<Arc<ContinuousBatchRequestPreparer>> {
-        let model_issue_path = ModelPath::from(self.model_path.as_path());
-
-        match model_loaded_rx
-            .await
-            .context("Failed to receive model loaded signal")
-        {
-            Ok(()) => {
-                self.context
-                    .slot_aggregated_status
-                    .register_fix(&AgentIssueFix::ModelIsLoaded(model_issue_path.clone()));
-            }
-            Err(err) => {
-                error!("Failed to load model: {err}");
-
-                self.context
-                    .slot_aggregated_status
-                    .register_issue(AgentIssue::ModelCannotBeLoaded(model_issue_path.clone()));
-            }
-        }
-
-        match chat_template_loaded_rx
-            .await
-            .context("Failed to receive chat template loaded signal")
-        {
-            Ok(ChatTemplateLoadStatus::Loaded) => {
-                self.context.slot_aggregated_status.register_fix(
-                    &AgentIssueFix::ModelChatTemplateIsLoaded(model_issue_path.clone()),
-                );
-            }
-            Ok(ChatTemplateLoadStatus::SkippedForEmbeddings) => {}
-            Err(err) => {
-                error!("Failed to load chat template: {err}");
-
-                if !self
-                    .context
-                    .slot_aggregated_status
-                    .has_issue(&AgentIssue::ModelCannotBeLoaded(model_issue_path.clone()))
-                {
-                    self.context
-                        .slot_aggregated_status
-                        .register_issue(AgentIssue::UnableToFindChatTemplate(model_issue_path));
-                }
-            }
-        }
-
-        agent_warm_and_scheduler_running_rx.await.context(
-            "Scheduler thread did not signal agent-warm-and-scheduler-running before exiting",
-        )
-    }
-
-    fn run_warmup_decode(
-        model: &LlamaModel,
-        llama_context: &mut LlamaContext<'_>,
-        warmup_batch: &mut LlamaBatch<'static>,
-        desired_slots_total: u16,
-    ) {
-        let warmup_tokens = [model.token_bos(), model.token_eos()];
-
-        for sequence_index in 0..i32::from(desired_slots_total) {
-            if let Err(err) = warmup_batch.add_sequence(&warmup_tokens, sequence_index, true) {
-                warn!("Warmup batch add_sequence failed: {err:#}");
-                return;
-            }
-        }
-
-        llama_context.clear_kv_cache();
-        if let Err(err) = llama_context.decode(warmup_batch) {
-            warn!("Warmup decode failed: {err:#}");
-        }
-        llama_context.synchronize();
-        llama_context.clear_kv_cache();
-    }
-}
diff --git a/paddler_agent/src/continuous_batch_arbiter_context.rs b/paddler_agent/src/continuous_batch_arbiter_context.rs
deleted file mode 100644
index f84d239ef..000000000
--- a/paddler_agent/src/continuous_batch_arbiter_context.rs
+++ /dev/null
@@ -1,12 +0,0 @@
-use std::sync::Arc;
-
-use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
-
-use crate::model_metadata_holder::ModelMetadataHolder;
-
-#[derive(Clone)]
-pub struct ContinuousBatchArbiterContext {
-    pub agent_name: Option<String>,
-    pub model_metadata_holder: Arc<ModelMetadataHolder>,
-    pub slot_aggregated_status: Arc<SlotAggregatedStatus>,
-}
diff --git a/paddler_agent/src/continuous_batch_arbiter_handle.rs b/paddler_agent/src/continuous_batch_arbiter_handle.rs
deleted file mode 100644
index e4e60f4e7..000000000
--- a/paddler_agent/src/continuous_batch_arbiter_handle.rs
+++ /dev/null
@@ -1,35 +0,0 @@
-use std::sync::Arc;
-use std::thread;
-
-use anyhow::Context as _;
-use anyhow::Result;
-use tokio::task::spawn_blocking;
-
-use crate::continuous_batch_request_preparer::ContinuousBatchRequestPreparer;
-use crate::join_scheduler_thread::join_scheduler_thread;
-
-pub struct ContinuousBatchArbiterHandle {
-    pub request_preparer: Arc<ContinuousBatchRequestPreparer>,
-    pub scheduler_thread_handle: thread::JoinHandle<Result<()>>,
-}
-
-impl ContinuousBatchArbiterHandle {
-    pub async fn shutdown(self) -> Result<()> {
-        let Self {
-            request_preparer,
-            scheduler_thread_handle,
-        } = self;
-
-        request_preparer
-            .shut_down_scheduler_after_pending_preparations()
-            .await;
-
-        spawn_blocking(move || {
-            drop(request_preparer);
-
-            join_scheduler_thread(scheduler_thread_handle)
-        })
-        .await
-        .context("Failed to join the scheduler shutdown task")?
-    }
-}
diff --git a/paddler_agent/src/continuous_batch_arbiter_spawn_outcome.rs b/paddler_agent/src/continuous_batch_arbiter_spawn_outcome.rs
deleted file mode 100644
index 24de55e8f..000000000
--- a/paddler_agent/src/continuous_batch_arbiter_spawn_outcome.rs
+++ /dev/null
@@ -1,6 +0,0 @@
-use crate::continuous_batch_arbiter_handle::ContinuousBatchArbiterHandle;
-
-pub enum ContinuousBatchArbiterSpawnOutcome {
-    Cancelled,
-    Ready(ContinuousBatchArbiterHandle),
-}
diff --git a/paddler_agent/src/continuous_batch_arbiter_state.rs b/paddler_agent/src/continuous_batch_arbiter_state.rs
deleted file mode 100644
index f347f6c2d..000000000
--- a/paddler_agent/src/continuous_batch_arbiter_state.rs
+++ /dev/null
@@ -1,218 +0,0 @@
-use anyhow::Context as _;
-use anyhow::Result;
-use log::error;
-use log::info;
-use log::warn;
-use tokio_util::sync::CancellationToken;
-
-use paddler_messaging::agent_state_application_status::AgentStateApplicationStatus;
-
-use crate::agent_applicable_model::AgentApplicableModel;
-use crate::agent_applicable_state::AgentApplicableState;
-use crate::continuous_batch_arbiter::ContinuousBatchArbiter;
-use crate::continuous_batch_arbiter_context::ContinuousBatchArbiterContext;
-use crate::continuous_batch_arbiter_handle::ContinuousBatchArbiterHandle;
-use crate::continuous_batch_arbiter_spawn_outcome::ContinuousBatchArbiterSpawnOutcome;
-use crate::continuous_batch_preparation_request::ContinuousBatchPreparationRequest;
-
-pub enum ContinuousBatchArbiterState {
-    Idle,
-    Running(ContinuousBatchArbiterHandle),
-}
-
-impl ContinuousBatchArbiterState {
-    pub async fn apply(
-        self,
-        shutdown: &CancellationToken,
-        agent_applicable_state: AgentApplicableState,
-        arbiter_context: &ContinuousBatchArbiterContext,
-    ) -> Self {
-        match self
-            .try_to_apply(shutdown, agent_applicable_state, arbiter_context)
-            .await
-        {
-            Ok(next_state) => next_state,
-            Err(application_error) => {
-                error!("Failed to apply reconciled state change: {application_error}");
-
-                Self::Idle
-            }
-        }
-    }
-
-    pub fn forward(&self, agent_name: Option<&str>, request: ContinuousBatchPreparationRequest) {
-        match self {
-            Self::Idle => request.reject_because_no_model_is_loaded(agent_name),
-            Self::Running(arbiter_handle) => arbiter_handle.request_preparer.prepare(request),
-        }
-    }
-
-    pub async fn shut_down(self) -> Result<()> {
-        match self {
-            Self::Idle => Ok(()),
-            Self::Running(arbiter_handle) => arbiter_handle
-                .shutdown()
-                .await
-                .context("Arbiter shutdown returned an error"),
-        }
-    }
-
-    async fn try_to_apply(
-        self,
-        shutdown: &CancellationToken,
-        agent_applicable_state: AgentApplicableState,
-        arbiter_context: &ContinuousBatchArbiterContext,
-    ) -> Result<Self> {
-        arbiter_context.slot_aggregated_status.reset();
-        self.shut_down().await?;
-
-        let next_state = match agent_applicable_state {
-            AgentApplicableState {
-                chat_template_override,
-                inference_parameters,
-                model:
-                    AgentApplicableModel::Resolved {
-                        model_path,
-                        multimodal_projection_path,
-                    },
-            } => {
-                Self::spawn(
-                    shutdown,
-                    ContinuousBatchArbiter {
-                        chat_template_override,
-                        context: arbiter_context.clone(),
-                        inference_parameters,
-                        model_path,
-                        multimodal_projection_path,
-                    },
-                )
-                .await?
-            }
-            AgentApplicableState {
-                model: AgentApplicableModel::NotConfigured,
-                ..
-            } => {
-                warn!("No model configured in applicable state; skipping llama.cpp initialization");
-
-                Self::Idle
-            }
-        };
-
-        arbiter_context
-            .slot_aggregated_status
-            .set_state_application_status(AgentStateApplicationStatus::Applied);
-
-        Ok(next_state)
-    }
-
-    async fn spawn(shutdown: &CancellationToken, arbiter: ContinuousBatchArbiter) -> Result<Self> {
-        match arbiter.spawn(shutdown).await? {
-            ContinuousBatchArbiterSpawnOutcome::Ready(arbiter_handle) => {
-                info!("Reconciled state change applied successfully");
-
-                Ok(Self::Running(arbiter_handle))
-            }
-            ContinuousBatchArbiterSpawnOutcome::Cancelled => {
-                info!("Model load was cancelled by shutdown before it finished");
-
-                Ok(Self::Idle)
-            }
-        }
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use std::mem::discriminant;
-    use std::num::NonZeroU32;
-    use std::sync::Arc;
-
-    use tokio::sync::mpsc;
-    use tokio_util::sync::CancellationToken;
-
-    use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
-    use paddler_messaging::agent_state_application_status::AgentStateApplicationStatus;
-    use paddler_messaging::generated_token_result::GeneratedTokenResult;
-    use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
-
-    use super::ContinuousBatchArbiterState;
-    use crate::agent_applicable_state::AgentApplicableState;
-    use crate::agent_request::AgentRequest;
-    use crate::continuous_batch_arbiter_context::ContinuousBatchArbiterContext;
-    use crate::continuous_batch_preparation_request::ContinuousBatchPreparationRequest;
-    use crate::model_metadata_holder::ModelMetadataHolder;
-    use crate::slot_guard::SlotGuard;
-
-    fn arbiter_context() -> ContinuousBatchArbiterContext {
-        ContinuousBatchArbiterContext {
-            agent_name: Some("agent".to_owned()),
-            model_metadata_holder: Arc::new(ModelMetadataHolder::default()),
-            slot_aggregated_status: Arc::new(SlotAggregatedStatus::new(1)),
-        }
-    }
-
-    async fn applied(
-        agent_applicable_state: AgentApplicableState,
-        arbiter_context: &ContinuousBatchArbiterContext,
-    ) -> ContinuousBatchArbiterState {
-        ContinuousBatchArbiterState::Idle
-            .apply(
-                &CancellationToken::new(),
-                agent_applicable_state,
-                arbiter_context,
-            )
-            .await
-    }
-
-    #[tokio::test]
-    async fn an_idle_arbiter_rejects_a_request_because_no_model_is_loaded() {
-        let arbiter_context = arbiter_context();
-        let (generated_tokens_tx, mut generated_tokens_rx) = mpsc::unbounded_channel();
-        let (_generate_tokens_stop_tx, generate_tokens_stop_rx) = mpsc::unbounded_channel();
-
-        ContinuousBatchArbiterState::Idle.forward(
-            arbiter_context.agent_name.as_deref(),
-            ContinuousBatchPreparationRequest::ContinueFromRawPrompt(AgentRequest {
-                params: ContinueFromRawPromptParams {
-                    grammar: None,
-                    max_tokens: NonZeroU32::MIN,
-                    raw_prompt: "Hello".to_owned(),
-                },
-                response_tx: generated_tokens_tx,
-                slot_guard: SlotGuard::new(arbiter_context.slot_aggregated_status.clone()),
-                stop_rx: generate_tokens_stop_rx,
-            }),
-        );
-
-        assert_eq!(
-            generated_tokens_rx.recv().await,
-            Some(GeneratedTokenResult::ModelNotLoaded(
-                "agent: no model is loaded".to_owned()
-            ))
-        );
-        assert_eq!(
-            arbiter_context
-                .slot_aggregated_status
-                .slots_processing_count(),
-            0
-        );
-    }
-
-    #[tokio::test]
-    async fn applying_a_state_without_a_model_stays_idle_and_reports_applied() {
-        let arbiter_context = arbiter_context();
-
-        let arbiter_state = applied(AgentApplicableState::default(), &arbiter_context).await;
-
-        assert_eq!(
-            discriminant(&arbiter_state),
-            discriminant(&ContinuousBatchArbiterState::Idle)
-        );
-        assert_eq!(
-            arbiter_context
-                .slot_aggregated_status
-                .get_state_application_status(),
-            AgentStateApplicationStatus::Applied
-        );
-    }
-}
diff --git a/paddler_agent/src/continuous_batch_embedding_processor.rs b/paddler_agent/src/continuous_batch_embedding_processor.rs
deleted file mode 100644
index ff34ca78e..000000000
--- a/paddler_agent/src/continuous_batch_embedding_processor.rs
+++ /dev/null
@@ -1,155 +0,0 @@
-use llama_cpp_bindings::context::LlamaContext;
-use llama_cpp_bindings::llama_batch::LlamaBatch;
-use log::warn;
-use tokio::sync::mpsc;
-
-use paddler_messaging::embedding::Embedding;
-use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
-use paddler_messaging::embedding_result::EmbeddingResult;
-
-use crate::continuous_batch_scheduler_context::ContinuousBatchSchedulerContext;
-use crate::embedding_batch_rejection::EmbeddingBatchRejection;
-use crate::embedding_input_tokenized::EmbeddingInputTokenized;
-use crate::normalization::normalize_embedding::normalize_embedding;
-use crate::plan_embedding_batches::plan_embedding_batches;
-use crate::prepared_embedding_batch_request::PreparedEmbeddingBatchRequest;
-use crate::receives_stop_request::ReceivesStopRequest as _;
-
-pub struct ContinuousBatchEmbeddingProcessor<'context> {
-    batch: &'context mut LlamaBatch<'static>,
-    llama_context: &'context mut LlamaContext<'static>,
-    scheduler_context: &'context ContinuousBatchSchedulerContext,
-}
-
-impl<'context> ContinuousBatchEmbeddingProcessor<'context> {
-    pub const fn new(
-        batch: &'context mut LlamaBatch<'static>,
-        llama_context: &'context mut LlamaContext<'static>,
-        scheduler_context: &'context ContinuousBatchSchedulerContext,
-    ) -> Self {
-        Self {
-            batch,
-            llama_context,
-            scheduler_context,
-        }
-    }
-
-    pub fn process_embedding_batch(
-        &mut self,
-        PreparedEmbeddingBatchRequest {
-            mut generate_embedding_stop_rx,
-            generated_embedding_tx,
-            inputs,
-            normalization_method,
-            oversized_documents,
-            slot_guard,
-        }: PreparedEmbeddingBatchRequest,
-    ) -> Result<(), EmbeddingBatchRejection> {
-        let _slot_guard = slot_guard;
-
-        for oversized_document in oversized_documents {
-            warn!(
-                "{:?}: skipped embedding document {:?}: {} tokens exceeds n_batch {}",
-                self.scheduler_context.agent_name,
-                oversized_document.source_document_id,
-                oversized_document.document_tokens,
-                oversized_document.n_batch,
-            );
-
-            generated_embedding_tx
-                .send(EmbeddingResult::DocumentExceedsBatchSize(
-                    oversized_document,
-                ))
-                .map_err(EmbeddingBatchRejection::ClientDisconnected)?;
-        }
-
-        let n_batch = self
-            .scheduler_context
-            .inference_parameters
-            .n_batch
-            .tokens_usize();
-        let max_sequences_per_batch = self.scheduler_context.desired_slots_total;
-
-        let token_counts: Vec<usize> = inputs.iter().map(|input| input.tokens.len()).collect();
-        let planned_batches =
-            plan_embedding_batches(&token_counts, n_batch, max_sequences_per_batch);
-        self.batch.clear();
-
-        let mut embeddings_emitted: usize = 0;
-
-        for planned_batch in planned_batches {
-            if generate_embedding_stop_rx.is_stop_requested() {
-                break;
-            }
-
-            let batch_inputs: Vec<&EmbeddingInputTokenized> =
-                inputs[planned_batch].iter().collect();
-
-            for (sequence_index, input) in batch_inputs.iter().enumerate() {
-                self.batch
-                    .add_sequence(
-                        &input.tokens,
-                        i32::try_from(sequence_index)
-                            .map_err(EmbeddingBatchRejection::SequenceIndexOutOfRange)?,
-                        true,
-                    )
-                    .map_err(EmbeddingBatchRejection::BatchAssemblyFailed)?;
-            }
-
-            self.embedding_batch_decode(
-                &batch_inputs,
-                &generated_embedding_tx,
-                &normalization_method,
-            )?;
-
-            embeddings_emitted += batch_inputs.len();
-        }
-
-        generated_embedding_tx
-            .send(if embeddings_emitted == 0 {
-                EmbeddingResult::NoEmbeddingsProduced
-            } else {
-                EmbeddingResult::Done
-            })
-            .map_err(EmbeddingBatchRejection::ClientDisconnected)
-    }
-
-    fn embedding_batch_decode(
-        &mut self,
-        current_batch_embeddings: &[&EmbeddingInputTokenized],
-        generated_embedding_tx: &mpsc::UnboundedSender<EmbeddingResult>,
-        normalization_method: &EmbeddingNormalizationMethod,
-    ) -> Result<(), EmbeddingBatchRejection> {
-        self.llama_context.clear_kv_cache();
-        self.llama_context
-            .decode(self.batch)
-            .map_err(EmbeddingBatchRejection::DecodeFailed)?;
-
-        for (index, embedding_input_tokenized) in current_batch_embeddings.iter().enumerate() {
-            let embedding = self
-                .llama_context
-                .embeddings_seq_ith(
-                    i32::try_from(index)
-                        .map_err(EmbeddingBatchRejection::SequenceIndexOutOfRange)?,
-                )
-                .map_err(EmbeddingBatchRejection::EmbeddingsUnavailable)?;
-
-            generated_embedding_tx
-                .send(EmbeddingResult::Embedding(Embedding {
-                    embedding: normalize_embedding(embedding.to_vec(), normalization_method)?,
-                    normalization_method: normalization_method.clone(),
-                    pooling_type: self
-                        .scheduler_context
-                        .inference_parameters
-                        .pooling_type
-                        .clone(),
-                    source_document_id: embedding_input_tokenized.id.clone(),
-                }))
-                .map_err(EmbeddingBatchRejection::ClientDisconnected)?;
-        }
-
-        self.batch.clear();
-
-        Ok(())
-    }
-}
diff --git a/paddler_agent/src/continuous_batch_preparation_request.rs b/paddler_agent/src/continuous_batch_preparation_request.rs
deleted file mode 100644
index 7dcdb4853..000000000
--- a/paddler_agent/src/continuous_batch_preparation_request.rs
+++ /dev/null
@@ -1,142 +0,0 @@
-use paddler_messaging::embedding_result::EmbeddingResult;
-use paddler_messaging::generated_token_result::GeneratedTokenResult;
-use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
-use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters_schema::validated_parameters_schema::ValidatedParametersSchema;
-use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
-use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
-
-use crate::agent_request::AgentRequest;
-use crate::embedding_batch_rejection::EmbeddingBatchRejection;
-use crate::generation_request_rejection::GenerationRequestRejection;
-
-pub enum ContinuousBatchPreparationRequest {
-    ContinueFromConversationHistory(
-        AgentRequest<
-            ContinueFromConversationHistoryParams<ValidatedParametersSchema>,
-            GeneratedTokenResult,
-        >,
-    ),
-    ContinueFromRawPrompt(AgentRequest<ContinueFromRawPromptParams, GeneratedTokenResult>),
-    GenerateEmbeddingBatch(AgentRequest<GenerateEmbeddingBatchParams, EmbeddingResult>),
-}
-
-impl ContinuousBatchPreparationRequest {
-    pub fn reject_because_no_model_is_loaded(self, agent_name: Option<&str>) {
-        match self {
-            Self::ContinueFromConversationHistory(AgentRequest { response_tx, .. })
-            | Self::ContinueFromRawPrompt(AgentRequest { response_tx, .. }) => {
-                GenerationRequestRejection::ModelNotLoaded.report(agent_name, &response_tx);
-            }
-            Self::GenerateEmbeddingBatch(AgentRequest { response_tx, .. }) => {
-                EmbeddingBatchRejection::ModelNotLoaded.report(agent_name, &response_tx);
-            }
-        }
-    }
-}
-
-impl
-    From<
-        AgentRequest<
-            ContinueFromConversationHistoryParams<ValidatedParametersSchema>,
-            GeneratedTokenResult,
-        >,
-    > for ContinuousBatchPreparationRequest
-{
-    fn from(
-        request: AgentRequest<
-            ContinueFromConversationHistoryParams<ValidatedParametersSchema>,
-            GeneratedTokenResult,
-        >,
-    ) -> Self {
-        Self::ContinueFromConversationHistory(request)
-    }
-}
-
-impl From<AgentRequest<ContinueFromRawPromptParams, GeneratedTokenResult>>
-    for ContinuousBatchPreparationRequest
-{
-    fn from(request: AgentRequest<ContinueFromRawPromptParams, GeneratedTokenResult>) -> Self {
-        Self::ContinueFromRawPrompt(request)
-    }
-}
-
-impl From<AgentRequest<GenerateEmbeddingBatchParams, EmbeddingResult>>
-    for ContinuousBatchPreparationRequest
-{
-    fn from(request: AgentRequest<GenerateEmbeddingBatchParams, EmbeddingResult>) -> Self {
-        Self::GenerateEmbeddingBatch(request)
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use std::num::NonZeroU32;
-    use std::sync::Arc;
-
-    use tokio::sync::mpsc;
-
-    use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
-    use paddler_messaging::conversation_history::ConversationHistory;
-    use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
-    use paddler_messaging::embedding_result::EmbeddingResult;
-    use paddler_messaging::generated_token_result::GeneratedTokenResult;
-    use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
-    use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
-
-    use super::ContinuousBatchPreparationRequest;
-    use crate::agent_request::AgentRequest;
-    use crate::slot_guard::SlotGuard;
-
-    #[test]
-    fn rejects_a_conversation_because_no_model_is_loaded() {
-        let (generated_tokens_tx, mut generated_tokens_rx) = mpsc::unbounded_channel();
-        let (_generate_tokens_stop_tx, generate_tokens_stop_rx) = mpsc::unbounded_channel();
-
-        ContinuousBatchPreparationRequest::ContinueFromConversationHistory(AgentRequest {
-            params: ContinueFromConversationHistoryParams {
-                add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(Vec::new()),
-                enable_thinking: false,
-                grammar: None,
-                max_tokens: NonZeroU32::MIN,
-                parse_tool_calls: false,
-                tools: Vec::new(),
-            },
-            response_tx: generated_tokens_tx,
-            slot_guard: SlotGuard::new(Arc::new(SlotAggregatedStatus::new(1))),
-            stop_rx: generate_tokens_stop_rx,
-        })
-        .reject_because_no_model_is_loaded(Some("agent"));
-
-        assert_eq!(
-            generated_tokens_rx.try_recv(),
-            Ok(GeneratedTokenResult::ModelNotLoaded(
-                "agent: no model is loaded".to_owned()
-            ))
-        );
-    }
-
-    #[test]
-    fn rejects_an_embedding_batch_because_no_model_is_loaded() {
-        let (generated_embedding_tx, mut generated_embedding_rx) = mpsc::unbounded_channel();
-        let (_generate_embedding_stop_tx, generate_embedding_stop_rx) = mpsc::unbounded_channel();
-
-        ContinuousBatchPreparationRequest::GenerateEmbeddingBatch(AgentRequest {
-            params: GenerateEmbeddingBatchParams {
-                input_batch: Vec::new(),
-                normalization_method: EmbeddingNormalizationMethod::None,
-            },
-            response_tx: generated_embedding_tx,
-            slot_guard: SlotGuard::new(Arc::new(SlotAggregatedStatus::new(1))),
-            stop_rx: generate_embedding_stop_rx,
-        })
-        .reject_because_no_model_is_loaded(Some("agent"));
-
-        assert_eq!(
-            generated_embedding_rx.try_recv(),
-            Ok(EmbeddingResult::ModelNotLoaded(
-                "agent: no model is loaded".to_owned()
-            ))
-        );
-    }
-}
diff --git a/paddler_agent/src/continuous_batch_request_phase.rs b/paddler_agent/src/continuous_batch_request_phase.rs
deleted file mode 100644
index 5dbcf3766..000000000
--- a/paddler_agent/src/continuous_batch_request_phase.rs
+++ /dev/null
@@ -1,11 +0,0 @@
-use crate::continuous_batch_generation_step::ContinuousBatchGenerationStep;
-use crate::continuous_batch_terminal_outcome::ContinuousBatchTerminalOutcome;
-use crate::multimodal_prompt_ingestion::MultimodalPromptIngestion;
-
-#[derive(Debug)]
-pub enum ContinuousBatchRequestPhase {
-    IngestingText,
-    IngestingMultimodal(MultimodalPromptIngestion),
-    Generating(ContinuousBatchGenerationStep),
-    Completed(ContinuousBatchTerminalOutcome),
-}
diff --git a/paddler_agent/src/continuous_batch_request_preparer.rs b/paddler_agent/src/continuous_batch_request_preparer.rs
deleted file mode 100644
index 6d23c7adb..000000000
--- a/paddler_agent/src/continuous_batch_request_preparer.rs
+++ /dev/null
@@ -1,100 +0,0 @@
-use std::sync::Arc;
-use std::sync::mpsc::Sender;
-
-use llama_cpp_bindings::llama_backend::LlamaBackend;
-use tokio_util::task::TaskTracker;
-
-use paddler_messaging::embedding_result::EmbeddingResult;
-use paddler_messaging::generated_token_result::GeneratedTokenResult;
-use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
-
-use crate::agent_request::AgentRequest;
-use crate::continuous_batch_preparation_request::ContinuousBatchPreparationRequest;
-use crate::continuous_batch_scheduler_command::ContinuousBatchSchedulerCommand;
-use crate::embedding_batch_preparer::EmbeddingBatchPreparer;
-use crate::forward_scheduler_command::forward_scheduler_command;
-use crate::generation_request_preparer::GenerationRequestPreparer;
-use crate::generation_request_rejection::GenerationRequestRejection;
-use crate::prepared_generation_request::PreparedGenerationRequest;
-
-pub struct ContinuousBatchRequestPreparer {
-    pub agent_name: Option<String>,
-    pub embedding_batch_preparer: EmbeddingBatchPreparer,
-    pub generation_request_preparer: GenerationRequestPreparer,
-    pub llama_backend: Arc<LlamaBackend>,
-    pub preparation_tasks: TaskTracker,
-    pub scheduler_command_tx: Sender<ContinuousBatchSchedulerCommand>,
-}
-
-impl ContinuousBatchRequestPreparer {
-    pub fn prepare(self: &Arc<Self>, request: ContinuousBatchPreparationRequest) {
-        let preparer = self.clone();
-
-        self.preparation_tasks
-            .spawn_blocking(move || match request {
-                ContinuousBatchPreparationRequest::ContinueFromConversationHistory(request) => {
-                    preparer.accept_generation(
-                        request,
-                        GenerationRequestPreparer::prepare_conversation_history,
-                    );
-                }
-                ContinuousBatchPreparationRequest::ContinueFromRawPrompt(request) => {
-                    preparer
-                        .accept_generation(request, GenerationRequestPreparer::prepare_raw_prompt);
-                }
-                ContinuousBatchPreparationRequest::GenerateEmbeddingBatch(request) => {
-                    preparer.accept_embedding_batch(request);
-                }
-            });
-    }
-
-    pub async fn shut_down_scheduler_after_pending_preparations(&self) {
-        self.preparation_tasks.close();
-        self.preparation_tasks.wait().await;
-        self.forward(ContinuousBatchSchedulerCommand::Shutdown);
-    }
-
-    fn forward(&self, command: ContinuousBatchSchedulerCommand) {
-        forward_scheduler_command(
-            &self.scheduler_command_tx,
-            self.agent_name.as_deref(),
-            command,
-        );
-    }
-
-    fn accept_generation<TParams>(
-        &self,
-        request: AgentRequest<TParams, GeneratedTokenResult>,
-        prepare: fn(
-            &GenerationRequestPreparer,
-            AgentRequest<TParams, GeneratedTokenResult>,
-        ) -> Result<PreparedGenerationRequest, GenerationRequestRejection>,
-    ) {
-        let generated_tokens_tx = request.response_tx.clone();
-
-        match prepare(&self.generation_request_preparer, request) {
-            Ok(prepared) => self.forward(ContinuousBatchSchedulerCommand::Generate(Box::new(
-                prepared,
-            ))),
-            Err(rejection) => rejection.report(self.agent_name.as_deref(), &generated_tokens_tx),
-        }
-    }
-
-    fn accept_embedding_batch(
-        &self,
-        request: AgentRequest<GenerateEmbeddingBatchParams, EmbeddingResult>,
-    ) {
-        let generated_embedding_tx = request.response_tx.clone();
-
-        match self.embedding_batch_preparer.prepare(request) {
-            Ok(prepared) => {
-                self.forward(ContinuousBatchSchedulerCommand::GenerateEmbeddingBatch(
-                    prepared,
-                ));
-            }
-            Err(rejection) => {
-                rejection.report(self.agent_name.as_deref(), &generated_embedding_tx);
-            }
-        }
-    }
-}
diff --git a/paddler_agent/src/continuous_batch_scheduler/mod.rs b/paddler_agent/src/continuous_batch_scheduler/mod.rs
deleted file mode 100644
index c16e21fc6..000000000
--- a/paddler_agent/src/continuous_batch_scheduler/mod.rs
+++ /dev/null
@@ -1,529 +0,0 @@
-pub mod advance_generating_phase;
-pub mod advance_outcome;
-pub mod assemble_batch_phase;
-pub mod batch_pass;
-pub mod classified_token;
-pub mod commit_phase;
-pub mod completion_check_outcome;
-pub mod completion_check_phase;
-pub mod contributions;
-pub mod decode_failure_phase;
-pub mod emit_token_outcome;
-pub mod emit_token_phase;
-pub mod generating_contribution;
-pub mod ingesting_contribution;
-pub mod sequence_ordered_insertion_index;
-
-use std::collections::VecDeque;
-use std::mem::transmute;
-use std::slice;
-use std::sync::mpsc::Receiver;
-use std::sync::mpsc::TryRecvError;
-
-use llama_cpp_bindings::SampledTokenClassifier;
-use llama_cpp_bindings::batch_add_error::BatchAddError;
-use llama_cpp_bindings::context::LlamaContext;
-use llama_cpp_bindings::error::DecodeError;
-use llama_cpp_bindings::ingest_outcome::IngestOutcome;
-use llama_cpp_bindings::llama_batch::LlamaBatch;
-use llama_cpp_bindings::token::data_array::LlamaTokenDataArray;
-use log::debug;
-use log::error;
-use log::info;
-use tokio_util::sync::CancellationToken;
-
-use paddler_messaging::embedding_result::EmbeddingResult;
-use paddler_messaging::generated_token_result::GeneratedTokenResult;
-use paddler_messaging::generation_finish::GenerationFinish;
-use paddler_messaging::generation_summary::GenerationSummary;
-
-use self::advance_generating_phase::AdvanceGeneratingPhase;
-use self::assemble_batch_phase::AssembleBatchPhase;
-use self::batch_pass::BatchPass;
-use self::commit_phase::CommitPhase;
-use self::decode_failure_phase::DecodeFailurePhase;
-use self::sequence_ordered_insertion_index::sequence_ordered_insertion_index;
-use crate::continuous_batch_active_request::ContinuousBatchActiveRequest;
-use crate::continuous_batch_embedding_processor::ContinuousBatchEmbeddingProcessor;
-use crate::continuous_batch_request_phase::ContinuousBatchRequestPhase;
-use crate::continuous_batch_request_state::ContinuousBatchRequestState;
-use crate::continuous_batch_scheduler_command::ContinuousBatchSchedulerCommand;
-use crate::continuous_batch_scheduler_context::ContinuousBatchSchedulerContext;
-use crate::continuous_batch_scheduler_params::ContinuousBatchSchedulerParams;
-use crate::generation_request_rejection::GenerationRequestRejection;
-use crate::multimodal_ingestion_progress::MultimodalIngestionProgress;
-use crate::prepared_embedding_batch_request::PreparedEmbeddingBatchRequest;
-use crate::prepared_generation_request::PreparedGenerationRequest;
-use crate::prepared_prompt::PreparedPrompt;
-use crate::receives_stop_request::ReceivesStopRequest as _;
-use crate::send_result_or_warn::send_result_or_warn;
-use crate::sequence_id_guard::SequenceIdGuard;
-use crate::sequence_id_pool::SequenceIdPool;
-use crate::token_classification::TokenClassification;
-
-pub struct ContinuousBatchScheduler {
-    active_requests: Vec<ContinuousBatchActiveRequest>,
-    agent_shutdown: CancellationToken,
-    batch: LlamaBatch<'static>,
-    candidates: LlamaTokenDataArray,
-    command_rx: Receiver<ContinuousBatchSchedulerCommand>,
-    ingest_outcomes: Vec<IngestOutcome>,
-    llama_context: LlamaContext<'static>,
-    pending_embedding_requests: VecDeque<PreparedEmbeddingBatchRequest>,
-    scheduler_context: ContinuousBatchSchedulerContext,
-    sequence_id_pool: SequenceIdPool,
-    shutdown_requested: bool,
-}
-
-impl ContinuousBatchScheduler {
-    #[must_use]
-    #[expect(
-        unsafe_code,
-        reason = "required for FFI lifetime extension with llama.cpp"
-    )]
-    pub fn new(
-        ContinuousBatchSchedulerParams {
-            agent_shutdown,
-            batch,
-            command_rx,
-            llama_context,
-            scheduler_context,
-        }: ContinuousBatchSchedulerParams,
-    ) -> Self {
-        let llama_context =
-            unsafe { transmute::<LlamaContext<'_>, LlamaContext<'static>>(llama_context) };
-        let sequence_id_pool = SequenceIdPool::new(scheduler_context.desired_slots_total);
-
-        Self {
-            active_requests: Vec::new(),
-            agent_shutdown,
-            batch,
-            candidates: LlamaTokenDataArray::new(Vec::new(), false),
-            command_rx,
-            ingest_outcomes: Vec::new(),
-            llama_context,
-            pending_embedding_requests: VecDeque::new(),
-            scheduler_context,
-            sequence_id_pool,
-            shutdown_requested: false,
-        }
-    }
-
-    pub fn run(&mut self) {
-        info!(
-            "{:?}: continuous batch scheduler started",
-            self.scheduler_context.agent_name
-        );
-
-        while !self.agent_shutdown.is_cancelled() {
-            self.check_stop_signals();
-            self.remove_completed_requests();
-            self.accept_new_commands();
-
-            let has_active_requests = self.has_active_requests();
-
-            self.try_process_embedding_request(has_active_requests);
-
-            if has_active_requests {
-                if let Err(batch_add_error) = self.execute_one_iteration() {
-                    self.fail_unfinished_requests(GenerationRequestRejection::BatchAssemblyFailed(
-                        batch_add_error,
-                    ));
-                }
-            } else if self.pending_embedding_requests.is_empty() {
-                if self.shutdown_requested {
-                    break;
-                }
-
-                self.wait_for_next_command();
-            }
-        }
-
-        while !self.active_requests.is_empty() {
-            self.cleanup_completed_request(0);
-        }
-
-        self.llama_context.synchronize();
-        self.llama_context.detach_threadpool();
-
-        info!(
-            "{:?}: continuous batch scheduler stopped",
-            self.scheduler_context.agent_name
-        );
-    }
-
-    fn wait_for_next_command(&mut self) {
-        if let Ok(command) = self.command_rx.recv() {
-            self.process_command(command);
-        } else {
-            info!(
-                "{:?}: command channel closed, shutting down scheduler",
-                self.scheduler_context.agent_name
-            );
-            self.shutdown_requested = true;
-        }
-    }
-
-    fn accept_new_commands(&mut self) {
-        loop {
-            match self.command_rx.try_recv() {
-                Ok(command) => self.process_command(command),
-                Err(TryRecvError::Empty) => break,
-                Err(TryRecvError::Disconnected) => {
-                    self.shutdown_requested = true;
-
-                    break;
-                }
-            }
-        }
-    }
-
-    fn process_command(&mut self, command: ContinuousBatchSchedulerCommand) {
-        match command {
-            ContinuousBatchSchedulerCommand::Generate(request) => {
-                self.accept_generation_request(*request);
-            }
-            ContinuousBatchSchedulerCommand::GenerateEmbeddingBatch(request) => {
-                self.pending_embedding_requests.push_back(request);
-            }
-            ContinuousBatchSchedulerCommand::Shutdown => {
-                self.shutdown_requested = true;
-            }
-        }
-    }
-
-    #[expect(
-        unsafe_code,
-        reason = "the SchedulerContext owns the LlamaModel for the lifetime of the active_requests vec — same pattern as LlamaContext<'static> above"
-    )]
-    fn build_token_classifier_for_active_request(
-        &self,
-        TokenClassification {
-            bare_json_tool_calls,
-            streaming_markers,
-        }: TokenClassification,
-    ) -> SampledTokenClassifier<'static> {
-        let classifier = SampledTokenClassifier::new(
-            &self.scheduler_context.model,
-            streaming_markers,
-            bare_json_tool_calls,
-        );
-
-        unsafe {
-            transmute::<SampledTokenClassifier<'_>, SampledTokenClassifier<'static>>(classifier)
-        }
-    }
-
-    fn accept_generation_request(&mut self, request: PreparedGenerationRequest) {
-        let generated_tokens_tx = request.generated_tokens_tx.clone();
-
-        match self.admit_generation_request(request) {
-            Ok(active_request) => {
-                debug!(
-                    "{:?}: accepted generation request on sequence {}",
-                    self.scheduler_context.agent_name,
-                    active_request.sequence_id_guard.sequence_id()
-                );
-
-                let insertion_index = sequence_ordered_insertion_index(
-                    self.active_requests
-                        .iter()
-                        .map(|request| request.sequence_id_guard.sequence_id()),
-                    active_request.sequence_id_guard.sequence_id(),
-                );
-
-                self.active_requests.insert(insertion_index, active_request);
-            }
-            Err(rejection) => {
-                rejection.report(
-                    self.scheduler_context.agent_name.as_deref(),
-                    &generated_tokens_tx,
-                );
-            }
-        }
-    }
-
-    fn admit_generation_request(
-        &mut self,
-        PreparedGenerationRequest {
-            generate_tokens_stop_rx,
-            generated_tokens_tx,
-            max_tokens,
-            prompt,
-            slot_guard,
-            token_classification,
-            token_sampling,
-            tool_call_handling,
-        }: PreparedGenerationRequest,
-    ) -> Result<ContinuousBatchActiveRequest, GenerationRequestRejection> {
-        let sequence_id_guard = SequenceIdGuard::acquire(&self.sequence_id_pool)
-            .ok_or(GenerationRequestRejection::NoSequenceSlotAvailable)?;
-        let mut token_classifier =
-            self.build_token_classifier_for_active_request(token_classification);
-
-        self.clear_kv_cache_for_sequence(sequence_id_guard.sequence_id())?;
-
-        let mut state = ContinuousBatchRequestState {
-            current_token_position: 0,
-            last_outcome_section: token_classifier.current_section(),
-            max_tokens,
-            phase: ContinuousBatchRequestPhase::IngestingText,
-            prompt_tokens: Vec::new(),
-            prompt_tokens_ingested: 0,
-            sampled_tokens: 0,
-        };
-
-        match prompt {
-            PreparedPrompt::TextTokens(prompt_tokens) => {
-                token_classifier.record_prompt_tokens(prompt_tokens.len() as u64);
-                token_classifier.ingest_prompt_tokens(&prompt_tokens);
-                state.prompt_tokens = prompt_tokens;
-            }
-            PreparedPrompt::Multimodal(ingestion) => {
-                state.phase = ContinuousBatchRequestPhase::IngestingMultimodal(ingestion);
-            }
-        }
-
-        state.last_outcome_section = token_classifier.current_section();
-
-        Ok(ContinuousBatchActiveRequest {
-            state,
-            token_classifier,
-            token_sampling,
-            generated_tokens_tx,
-            generate_tokens_stop_rx,
-            sequence_id_guard,
-            slot_guard,
-            tool_call_handling,
-        })
-    }
-
-    fn ingest_next_multimodal_chunks(&mut self) {
-        for active_request in &mut self.active_requests {
-            let ContinuousBatchRequestPhase::IngestingMultimodal(ingestion) =
-                &mut active_request.state.phase
-            else {
-                continue;
-            };
-
-            match ingestion.ingest_next_chunk(
-                &mut active_request.token_classifier,
-                &self.llama_context,
-                active_request.sequence_id_guard.sequence_id(),
-                active_request.state.current_token_position,
-            ) {
-                Ok(MultimodalIngestionProgress::ChunksRemain { next_position }) => {
-                    active_request.state.current_token_position = next_position;
-                }
-                Ok(MultimodalIngestionProgress::PromptIngested { next_position }) => {
-                    active_request
-                        .state
-                        .begin_generating_after_multimodal_prompt(
-                            next_position,
-                            active_request.token_classifier.current_section(),
-                        );
-                    self.llama_context.mark_logits_initialized(-1);
-
-                    AdvanceGeneratingPhase {
-                        candidates: &mut self.candidates,
-                        ingest_outcomes: &mut self.ingest_outcomes,
-                        llama_context: &self.llama_context,
-                        scheduler_context: &self.scheduler_context,
-                    }
-                    .run(slice::from_mut(active_request));
-                }
-                Err(ingestion_error) => {
-                    active_request.complete_with_outcome(
-                        GenerationRequestRejection::MultimodalIngestionFailed(ingestion_error)
-                            .into_generated_token_result(
-                                self.scheduler_context.agent_name.as_deref(),
-                            ),
-                    );
-                }
-            }
-        }
-    }
-
-    fn check_stop_signals(&mut self) {
-        for active_request in &mut self.active_requests {
-            if matches!(
-                active_request.state.phase,
-                ContinuousBatchRequestPhase::Completed(_)
-            ) {
-                continue;
-            }
-
-            if active_request.generate_tokens_stop_rx.is_stop_requested() {
-                let summary = GenerationSummary {
-                    finish: GenerationFinish::StopRequested,
-                    usage: *active_request.token_classifier.usage(),
-                };
-
-                active_request.complete_with_outcome(GeneratedTokenResult::Done(summary));
-            }
-        }
-    }
-
-    fn try_process_embedding_request(&mut self, has_active_requests: bool) {
-        let Some(request) = self.pending_embedding_requests.pop_front() else {
-            return;
-        };
-
-        if has_active_requests {
-            send_result_or_warn(
-                self.scheduler_context.agent_name.as_deref(),
-                &request.generated_embedding_tx,
-                EmbeddingResult::EmbeddingRejectedDueToActiveTokenGeneration,
-            );
-
-            return;
-        }
-
-        let generated_embedding_tx = request.generated_embedding_tx.clone();
-        let mut processor = ContinuousBatchEmbeddingProcessor::new(
-            &mut self.batch,
-            &mut self.llama_context,
-            &self.scheduler_context,
-        );
-
-        if let Err(rejection) = processor.process_embedding_batch(request) {
-            rejection.report(
-                self.scheduler_context.agent_name.as_deref(),
-                &generated_embedding_tx,
-            );
-        }
-    }
-
-    fn has_active_requests(&self) -> bool {
-        self.active_requests.iter().any(|request| {
-            !matches!(
-                request.state.phase,
-                ContinuousBatchRequestPhase::Completed(_)
-            )
-        })
-    }
-
-    fn fail_unfinished_requests(&mut self, rejection: GenerationRequestRejection) {
-        let outcome =
-            rejection.into_generated_token_result(self.scheduler_context.agent_name.as_deref());
-
-        for active_request in &mut self.active_requests {
-            active_request.complete_with_outcome(outcome.clone());
-        }
-    }
-
-    fn execute_one_iteration(&mut self) -> Result<(), BatchAddError> {
-        self.advance_generating_requests();
-        self.ingest_next_multimodal_chunks();
-
-        let assemble_phase = AssembleBatchPhase {
-            n_batch: self
-                .scheduler_context
-                .inference_parameters
-                .n_batch
-                .tokens_usize(),
-        };
-
-        let mut pass = BatchPass::new(&mut self.batch);
-
-        assemble_phase.run(&mut pass, &mut self.active_requests)?;
-
-        if pass.is_empty() {
-            return Ok(());
-        }
-
-        debug!(
-            "{:?}: decoding batch with {} tokens for {} active requests",
-            self.scheduler_context.agent_name,
-            pass.batch.n_tokens(),
-            self.active_requests.len()
-        );
-
-        match self.llama_context.decode(pass.batch) {
-            Ok(()) => {
-                CommitPhase {
-                    requests: &mut self.active_requests,
-                }
-                .run(pass);
-            }
-            Err(DecodeError::Aborted) => {}
-            Err(decode_error) => {
-                let description = format!(
-                    "{:?}: llama.cpp rejected the batch: {decode_error}",
-                    self.scheduler_context.agent_name
-                );
-
-                error!("{description}");
-                DecodeFailurePhase {
-                    requests: &mut self.active_requests,
-                }
-                .run(pass, &description);
-            }
-        }
-
-        Ok(())
-    }
-
-    fn advance_generating_requests(&mut self) {
-        AdvanceGeneratingPhase {
-            candidates: &mut self.candidates,
-            ingest_outcomes: &mut self.ingest_outcomes,
-            llama_context: &self.llama_context,
-            scheduler_context: &self.scheduler_context,
-        }
-        .run(&mut self.active_requests);
-    }
-
-    fn remove_completed_requests(&mut self) {
-        let mut removal_index = 0;
-
-        while removal_index < self.active_requests.len() {
-            if matches!(
-                self.active_requests[removal_index].state.phase,
-                ContinuousBatchRequestPhase::Completed(_)
-            ) {
-                self.cleanup_completed_request(removal_index);
-            } else {
-                removal_index += 1;
-            }
-        }
-    }
-
-    fn clear_kv_cache_for_sequence(
-        &mut self,
-        sequence_id: i32,
-    ) -> Result<(), GenerationRequestRejection> {
-        let kv_cache_sequence_id = u32::try_from(sequence_id).map_err(|source| {
-            GenerationRequestRejection::SequenceIdOutOfRange {
-                sequence_id,
-                source,
-            }
-        })?;
-
-        self.llama_context
-            .clear_kv_cache_seq(Some(kv_cache_sequence_id), None, None)
-            .map_err(GenerationRequestRejection::KvCacheClearFailed)
-    }
-
-    fn cleanup_completed_request(&mut self, index: usize) {
-        let removed_request = self.active_requests.remove(index);
-        let sequence_id = removed_request.sequence_id_guard.sequence_id();
-        let usage = *removed_request.token_classifier.usage();
-        let terminal_delivery = removed_request.into_terminal_delivery();
-
-        if let Err(rejection) = self.clear_kv_cache_for_sequence(sequence_id) {
-            error!(
-                "{:?}: {rejection}; the next request admitted to sequence {sequence_id} clears it again",
-                self.scheduler_context.agent_name
-            );
-        }
-
-        debug!(
-            "{:?}: cleaned up sequence {sequence_id} ({} completion tokens generated)",
-            self.scheduler_context.agent_name,
-            usage.completion_tokens(),
-        );
-
-        terminal_delivery.deliver(self.scheduler_context.agent_name.as_deref());
-    }
-}
diff --git a/paddler_agent/src/continuous_batch_scheduler/sequence_ordered_insertion_index.rs b/paddler_agent/src/continuous_batch_scheduler/sequence_ordered_insertion_index.rs
deleted file mode 100644
index 3fb360d52..000000000
--- a/paddler_agent/src/continuous_batch_scheduler/sequence_ordered_insertion_index.rs
+++ /dev/null
@@ -1,29 +0,0 @@
-pub fn sequence_ordered_insertion_index(
-    active_sequence_ids: impl IntoIterator<Item = i32>,
-    sequence_id: i32,
-) -> usize {
-    active_sequence_ids
-        .into_iter()
-        .take_while(|active_sequence_id| *active_sequence_id < sequence_id)
-        .count()
-}
-
-#[cfg(test)]
-mod tests {
-    use super::sequence_ordered_insertion_index;
-
-    #[test]
-    fn places_a_sequence_before_every_higher_sequence_id() {
-        assert_eq!(sequence_ordered_insertion_index([1, 4, 6], 5), 2);
-    }
-
-    #[test]
-    fn places_the_lowest_sequence_id_first() {
-        assert_eq!(sequence_ordered_insertion_index([1, 4, 6], 0), 0);
-    }
-
-    #[test]
-    fn places_the_highest_sequence_id_last() {
-        assert_eq!(sequence_ordered_insertion_index([1, 4, 6], 7), 3);
-    }
-}
diff --git a/paddler_agent/src/continuous_batch_scheduler_command.rs b/paddler_agent/src/continuous_batch_scheduler_command.rs
deleted file mode 100644
index b806ffa0c..000000000
--- a/paddler_agent/src/continuous_batch_scheduler_command.rs
+++ /dev/null
@@ -1,8 +0,0 @@
-use crate::prepared_embedding_batch_request::PreparedEmbeddingBatchRequest;
-use crate::prepared_generation_request::PreparedGenerationRequest;
-
-pub enum ContinuousBatchSchedulerCommand {
-    Generate(Box<PreparedGenerationRequest>),
-    GenerateEmbeddingBatch(PreparedEmbeddingBatchRequest),
-    Shutdown,
-}
diff --git a/paddler_agent/src/continuous_batch_scheduler_context.rs b/paddler_agent/src/continuous_batch_scheduler_context.rs
deleted file mode 100644
index 008d9f55a..000000000
--- a/paddler_agent/src/continuous_batch_scheduler_context.rs
+++ /dev/null
@@ -1,12 +0,0 @@
-use std::sync::Arc;
-
-use llama_cpp_bindings::model::LlamaModel;
-
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
-
-pub struct ContinuousBatchSchedulerContext {
-    pub agent_name: Option<String>,
-    pub desired_slots_total: u16,
-    pub inference_parameters: InferenceParameters,
-    pub model: Arc<LlamaModel>,
-}
diff --git a/paddler_agent/src/continuous_batch_scheduler_params.rs b/paddler_agent/src/continuous_batch_scheduler_params.rs
deleted file mode 100644
index 1e5d296e1..000000000
--- a/paddler_agent/src/continuous_batch_scheduler_params.rs
+++ /dev/null
@@ -1,16 +0,0 @@
-use std::sync::mpsc::Receiver;
-
-use llama_cpp_bindings::context::LlamaContext;
-use llama_cpp_bindings::llama_batch::LlamaBatch;
-use tokio_util::sync::CancellationToken;
-
-use crate::continuous_batch_scheduler_command::ContinuousBatchSchedulerCommand;
-use crate::continuous_batch_scheduler_context::ContinuousBatchSchedulerContext;
-
-pub struct ContinuousBatchSchedulerParams<'model> {
-    pub agent_shutdown: CancellationToken,
-    pub batch: LlamaBatch<'static>,
-    pub command_rx: Receiver<ContinuousBatchSchedulerCommand>,
-    pub llama_context: LlamaContext<'model>,
-    pub scheduler_context: ContinuousBatchSchedulerContext,
-}
diff --git a/paddler_agent/src/desired_model_file.rs b/paddler_agent/src/desired_model_file.rs
new file mode 100644
index 000000000..70d23abd5
--- /dev/null
+++ b/paddler_agent/src/desired_model_file.rs
@@ -0,0 +1,41 @@
+use std::path::Path;
+
+use paddler_agent_status::agent_issue_fix::AgentIssueFix;
+use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
+use paddler_messaging::agent_issue::AgentIssue;
+use paddler_messaging::agent_issue_params::model_path::ModelPath;
+
+#[derive(Clone, Copy)]
+pub enum DesiredModelFile {
+    Model,
+    MultimodalProjection,
+    PointerHead,
+}
+
+impl DesiredModelFile {
+    pub fn report_found(self, path: &Path, slot_aggregated_status: &SlotAggregatedStatus) {
+        match self {
+            Self::Model => {
+                slot_aggregated_status
+                    .register_fix(&AgentIssueFix::ModelFileExists(ModelPath::from(path)));
+            }
+            Self::MultimodalProjection => {}
+            Self::PointerHead => {
+                slot_aggregated_status
+                    .register_fix(&AgentIssueFix::PointerHeadFileExists(ModelPath::from(path)));
+            }
+        }
+    }
+
+    pub fn report_missing(self, path: &Path, slot_aggregated_status: &SlotAggregatedStatus) {
+        let model_path = ModelPath::from(path);
+
+        slot_aggregated_status.register_desired_state_issue(match self {
+            Self::Model => AgentIssue::ModelFileDoesNotExist(model_path),
+            Self::MultimodalProjection => {
+                AgentIssue::MultimodalProjectionCannotBeLoaded(model_path)
+            }
+            Self::PointerHead => AgentIssue::PointerHeadCannotBeLoaded(model_path),
+        });
+    }
+}
diff --git a/paddler_agent/src/desired_state_reconciler.rs b/paddler_agent/src/desired_state_reconciler.rs
index 3952aa2ee..9b734b58e 100644
--- a/paddler_agent/src/desired_state_reconciler.rs
+++ b/paddler_agent/src/desired_state_reconciler.rs
@@ -3,17 +3,17 @@ use std::sync::Arc;
 use log::error;
 use tokio_util::sync::CancellationToken;
 
-use paddler_agent_status::agent_issue_fix::AgentIssueFix;
 use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
 use paddler_messaging::agent_desired_state::AgentDesiredState;
+use paddler_model_source::desired_model_resolver::DesiredModelResolver;
 
-use crate::agent_applicable_state_holder::AgentApplicableStateHolder;
 use crate::agent_desired_state_conversion::AgentDesiredStateConversion;
 use crate::agent_desired_state_converter::AgentDesiredStateConverter;
 use crate::desired_state_reconciliation::DesiredStateReconciliation;
 
+#[derive(Clone)]
 pub struct DesiredStateReconciler {
-    pub agent_applicable_state_holder: Arc<AgentApplicableStateHolder>,
+    pub desired_model_resolver: DesiredModelResolver,
     pub slot_aggregated_status: Arc<SlotAggregatedStatus>,
 }
 
@@ -24,6 +24,7 @@ impl DesiredStateReconciler {
         agent_desired_state: AgentDesiredState,
     ) -> DesiredStateReconciliation {
         let conversion = AgentDesiredStateConverter {
+            desired_model_resolver: self.desired_model_resolver.clone(),
             cancellation_token: cancellation_token.clone(),
             slot_aggregated_status: self.slot_aggregated_status.clone(),
         }
@@ -32,16 +33,7 @@ impl DesiredStateReconciler {
 
         match conversion {
             Ok(AgentDesiredStateConversion::Converted(applicable_state)) => {
-                let uses_chat_template_override = applicable_state.chat_template_override.is_some();
-
-                self.agent_applicable_state_holder
-                    .set_agent_applicable_state(applicable_state);
-                self.slot_aggregated_status
-                    .set_uses_chat_template_override(uses_chat_template_override);
-                self.slot_aggregated_status
-                    .register_fix(&AgentIssueFix::ModelStateIsReconciled);
-
-                DesiredStateReconciliation::Reconciled
+                DesiredStateReconciliation::Reconciled(Box::new(applicable_state))
             }
             Ok(AgentDesiredStateConversion::Cancelled) => {
                 DesiredStateReconciliation::Pending(Box::new(agent_desired_state))
@@ -54,3 +46,90 @@ impl DesiredStateReconciler {
         }
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use std::path::Path;
+    use std::sync::Arc;
+
+    use tokio_util::sync::CancellationToken;
+
+    use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
+    use paddler_messaging::agent_desired_model::AgentDesiredModel;
+    use paddler_messaging::agent_desired_state::AgentDesiredState;
+    use paddler_model_source::agent_local_model_path::AgentLocalModelPath;
+    use paddler_model_source::desired_model_resolver::DesiredModelResolver;
+    use paddler_model_source::huggingface_default_endpoint::HUGGINGFACE_DEFAULT_ENDPOINT;
+    use paddler_model_source::huggingface_model_reference::HuggingFaceModelReference;
+    use paddler_model_source::model_source::ModelSource;
+
+    use super::DesiredStateReconciler;
+    use crate::desired_state_reconciliation::DesiredStateReconciliation;
+
+    fn desired_state_with_model(model: AgentDesiredModel) -> AgentDesiredState {
+        AgentDesiredState {
+            model,
+            ..AgentDesiredState::default()
+        }
+    }
+
+    fn desired_state_reconciler() -> DesiredStateReconciler {
+        DesiredStateReconciler {
+            desired_model_resolver: DesiredModelResolver::from_process_env(
+                HUGGINGFACE_DEFAULT_ENDPOINT.to_owned(),
+            )
+            .unwrap(),
+            slot_aggregated_status: Arc::new(SlotAggregatedStatus::new(1)),
+        }
+    }
+
+    #[tokio::test]
+    async fn a_missing_local_model_leaves_the_state_unconverted() {
+        let reconciler = desired_state_reconciler();
+        let desired_state = desired_state_with_model(
+            ModelSource::LocalToAgent(
+                AgentLocalModelPath::from_absolute_path(Path::new(
+                    "/paddler-nonexistent-model-for-reconciliation.gguf",
+                ))
+                .unwrap(),
+            )
+            .into_agent_desired_model(),
+        );
+
+        let reconciliation = reconciler
+            .reconcile(&CancellationToken::new(), desired_state.clone())
+            .await;
+
+        assert_eq!(
+            reconciliation,
+            DesiredStateReconciliation::Pending(Box::new(desired_state))
+        );
+    }
+
+    #[tokio::test]
+    async fn a_model_download_cancelled_mid_conversion_leaves_the_state_unconverted() {
+        let cancellation_token = CancellationToken::new();
+
+        cancellation_token.cancel();
+
+        let reconciler = desired_state_reconciler();
+        let desired_state = desired_state_with_model(
+            ModelSource::HuggingFace(HuggingFaceModelReference {
+                filename: "model.gguf".to_owned(),
+                repository_name: "never-downloaded".to_owned(),
+                repository_owner: "paddler-tests".to_owned(),
+                revision: "main".to_owned(),
+            })
+            .into_agent_desired_model(),
+        );
+
+        let reconciliation = reconciler
+            .reconcile(&cancellation_token, desired_state.clone())
+            .await;
+
+        assert_eq!(
+            reconciliation,
+            DesiredStateReconciliation::Pending(Box::new(desired_state))
+        );
+    }
+}
diff --git a/paddler_agent/src/desired_state_reconciliation.rs b/paddler_agent/src/desired_state_reconciliation.rs
index ee4a9ba6e..7f9056048 100644
--- a/paddler_agent/src/desired_state_reconciliation.rs
+++ b/paddler_agent/src/desired_state_reconciliation.rs
@@ -1,7 +1,9 @@
 use paddler_messaging::agent_desired_state::AgentDesiredState;
 
+use crate::agent_applicable_state::AgentApplicableState;
+
 #[derive(Debug, PartialEq)]
 pub enum DesiredStateReconciliation {
     Pending(Box<AgentDesiredState>),
-    Reconciled,
+    Reconciled(Box<AgentApplicableState>),
 }
diff --git a/paddler_agent/src/desired_state_reconciliation_attempt.rs b/paddler_agent/src/desired_state_reconciliation_attempt.rs
new file mode 100644
index 000000000..30bf4d67b
--- /dev/null
+++ b/paddler_agent/src/desired_state_reconciliation_attempt.rs
@@ -0,0 +1,209 @@
+use std::future::Future;
+use std::future::pending;
+use std::pin::Pin;
+
+use log::error;
+use tokio::runtime::Builder;
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::agent_desired_state::AgentDesiredState;
+use paddler_spawn_resuming_panic::spawn_blocking_resuming_panic::spawn_blocking_resuming_panic;
+
+use crate::desired_state_reconciler::DesiredStateReconciler;
+use crate::desired_state_reconciliation::DesiredStateReconciliation;
+use crate::retry_delay::RetryDelay;
+use crate::retry_delay_outcome::RetryDelayOutcome;
+
+pub enum DesiredStateReconciliationAttempt {
+    Idle,
+    Reconciling {
+        cancellation_token: CancellationToken,
+        reconciliation: Pin<Box<dyn Future<Output = DesiredStateReconciliation> + Send>>,
+    },
+}
+
+impl DesiredStateReconciliationAttempt {
+    #[must_use]
+    pub fn start(
+        desired_state_reconciler: DesiredStateReconciler,
+        agent_desired_state: AgentDesiredState,
+        agent_shutdown: &CancellationToken,
+        retry_delay: RetryDelay,
+    ) -> Self {
+        let cancellation_token = agent_shutdown.child_token();
+        let reconciliation_cancellation_token = cancellation_token.clone();
+
+        Self::Reconciling {
+            cancellation_token,
+            reconciliation: Box::pin(async move {
+                if retry_delay.elapse(&reconciliation_cancellation_token).await
+                    == RetryDelayOutcome::Cancelled
+                {
+                    return DesiredStateReconciliation::Pending(Box::new(agent_desired_state));
+                }
+
+                match Builder::new_current_thread().enable_all().build() {
+                    Ok(reconciliation_runtime) => {
+                        spawn_blocking_resuming_panic(move || {
+                            reconciliation_runtime.block_on(
+                                desired_state_reconciler.reconcile(
+                                    &reconciliation_cancellation_token,
+                                    agent_desired_state,
+                                ),
+                            )
+                        })
+                        .await
+                    }
+                    Err(runtime_creation_error) => {
+                        error!(
+                            "Unable to create the runtime that reconciles the desired state: {runtime_creation_error}"
+                        );
+
+                        DesiredStateReconciliation::Pending(Box::new(agent_desired_state))
+                    }
+                }
+            }),
+        }
+    }
+
+    pub async fn abandon(&mut self) {
+        if let Self::Reconciling {
+            cancellation_token,
+            reconciliation,
+        } = self
+        {
+            cancellation_token.cancel();
+            reconciliation.as_mut().await;
+        }
+
+        *self = Self::Idle;
+    }
+
+    pub async fn finish(&mut self) -> DesiredStateReconciliation {
+        match self {
+            Self::Idle => pending().await,
+            Self::Reconciling { reconciliation, .. } => {
+                let desired_state_reconciliation = reconciliation.as_mut().await;
+
+                *self = Self::Idle;
+
+                desired_state_reconciliation
+            }
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::fs::File;
+    use std::fs::read_dir;
+    use std::os::fd::AsRawFd as _;
+    use std::sync::Arc;
+
+    use nix::sys::resource::Resource;
+    use nix::sys::resource::getrlimit;
+    use nix::sys::resource::setrlimit;
+    use tempfile::TempDir;
+    use tokio_util::sync::CancellationToken;
+
+    use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
+    use paddler_cache_dir::cache_dir::CacheDir;
+    use paddler_download_manager::download_manager::DownloadManager;
+    use paddler_local_http_fixture::fixture_response::FixtureResponse;
+    use paddler_local_http_fixture::local_http_fixture::LocalHttpFixture;
+    use paddler_messaging::agent_desired_model::AgentDesiredModel;
+    use paddler_messaging::agent_desired_state::AgentDesiredState;
+    use paddler_model_source::desired_model_resolver::DesiredModelResolver;
+    use paddler_model_source::huggingface_cache_location::HuggingFaceCacheLocation;
+    use paddler_model_source::huggingface_default_endpoint::HUGGINGFACE_DEFAULT_ENDPOINT;
+    use paddler_model_source::huggingface_hub::HuggingFaceHub;
+
+    use super::DesiredStateReconciliationAttempt;
+    use crate::desired_state_reconciler::DesiredStateReconciler;
+    use crate::desired_state_reconciliation::DesiredStateReconciliation;
+    use crate::retry_delay::RetryDelay;
+
+    #[tokio::test]
+    async fn a_reconciliation_that_cannot_create_its_runtime_stays_pending() {
+        let (soft_descriptor_limit, hard_descriptor_limit) =
+            getrlimit(Resource::RLIMIT_NOFILE).unwrap();
+        let lowest_free_descriptor = File::open("/dev/null").unwrap().as_raw_fd();
+        let mut reconciliation_attempt = DesiredStateReconciliationAttempt::start(
+            DesiredStateReconciler {
+                desired_model_resolver: DesiredModelResolver::from_process_env(
+                    HUGGINGFACE_DEFAULT_ENDPOINT.to_owned(),
+                )
+                .unwrap(),
+                slot_aggregated_status: Arc::new(SlotAggregatedStatus::new(1)),
+            },
+            AgentDesiredState::default(),
+            &CancellationToken::new(),
+            RetryDelay::Immediately,
+        );
+
+        setrlimit(
+            Resource::RLIMIT_NOFILE,
+            u64::try_from(lowest_free_descriptor).unwrap(),
+            hard_descriptor_limit,
+        )
+        .unwrap();
+
+        let desired_state_reconciliation = reconciliation_attempt.finish().await;
+
+        setrlimit(
+            Resource::RLIMIT_NOFILE,
+            soft_descriptor_limit,
+            hard_descriptor_limit,
+        )
+        .unwrap();
+
+        assert_eq!(
+            desired_state_reconciliation,
+            DesiredStateReconciliation::Pending(Box::default())
+        );
+    }
+
+    #[tokio::test(start_paused = true)]
+    async fn a_reconciliation_abandoned_during_its_retry_delay_never_resolves_its_model() {
+        let cache_dir = TempDir::new().unwrap();
+        let model_server = LocalHttpFixture::start(FixtureResponse::Ok(b"model".to_vec()))
+            .await
+            .unwrap();
+        let mut reconciliation_attempt = DesiredStateReconciliationAttempt::start(
+            DesiredStateReconciler {
+                desired_model_resolver: DesiredModelResolver {
+                    cache_dir: CacheDir {
+                        explicit: Some(cache_dir.path().to_path_buf()),
+                        home: None,
+                        xdg: None,
+                    },
+                    download_manager: DownloadManager::new().unwrap(),
+                    huggingface_client: HuggingFaceHub {
+                        cache_location: HuggingFaceCacheLocation::HubDefault,
+                        endpoint: HUGGINGFACE_DEFAULT_ENDPOINT.to_owned(),
+                    }
+                    .client()
+                    .unwrap(),
+                },
+                slot_aggregated_status: Arc::new(SlotAggregatedStatus::new(1)),
+            },
+            AgentDesiredState {
+                model: AgentDesiredModel::Uri(model_server.url("/model.gguf")),
+                ..AgentDesiredState::default()
+            },
+            &CancellationToken::new(),
+            RetryDelay::AfterRetryInterval,
+        );
+
+        reconciliation_attempt.abandon().await;
+
+        assert_eq!(read_dir(cache_dir.path()).unwrap().count(), 0);
+        assert_eq!(model_server.request_count(), 0);
+
+        model_server
+            .stop()
+            .await
+            .expect("the fixture task must not panic")
+            .expect("the fixture must serve every request");
+    }
+}
diff --git a/paddler_agent/src/embedding_batch_preparer.rs b/paddler_agent/src/embedding_batch_preparer.rs
deleted file mode 100644
index 4ef98112a..000000000
--- a/paddler_agent/src/embedding_batch_preparer.rs
+++ /dev/null
@@ -1,73 +0,0 @@
-use std::sync::Arc;
-
-use llama_cpp_bindings::model::AddBos;
-use llama_cpp_bindings::model::LlamaModel;
-
-use paddler_messaging::embedding_result::EmbeddingResult;
-use paddler_messaging::oversized_embedding_document_details::OversizedEmbeddingDocumentDetails;
-use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
-
-use crate::agent_request::AgentRequest;
-use crate::embedding_batch_rejection::EmbeddingBatchRejection;
-use crate::embedding_input_tokenized::EmbeddingInputTokenized;
-use crate::prepared_embedding_batch_request::PreparedEmbeddingBatchRequest;
-use crate::require_embeddings_enabled::require_embeddings_enabled;
-
-pub struct EmbeddingBatchPreparer {
-    pub enable_embeddings: bool,
-    pub model: Arc<LlamaModel>,
-    pub n_batch: usize,
-}
-
-impl EmbeddingBatchPreparer {
-    pub fn prepare(
-        &self,
-        AgentRequest {
-            params:
-                GenerateEmbeddingBatchParams {
-                    input_batch,
-                    normalization_method,
-                },
-            response_tx: generated_embedding_tx,
-            slot_guard,
-            stop_rx: generate_embedding_stop_rx,
-        }: AgentRequest<GenerateEmbeddingBatchParams, EmbeddingResult>,
-    ) -> Result<PreparedEmbeddingBatchRequest, EmbeddingBatchRejection> {
-        require_embeddings_enabled(self.enable_embeddings)?;
-
-        let mut inputs = Vec::with_capacity(input_batch.len());
-        let mut oversized_documents = Vec::new();
-
-        for input in input_batch {
-            let tokens = self
-                .model
-                .str_to_token(&input.content, AddBos::Always)
-                .map_err(|source| EmbeddingBatchRejection::InputTokenizationFailed {
-                    source_document_id: input.id.clone(),
-                    source,
-                })?;
-
-            if tokens.len() > self.n_batch {
-                oversized_documents.push(OversizedEmbeddingDocumentDetails {
-                    document_tokens: tokens.len(),
-                    n_batch: self.n_batch,
-                    source_document_id: input.id,
-                });
-            } else {
-                inputs.push(EmbeddingInputTokenized {
-                    id: input.id,
-                    tokens,
-                });
-            }
-        }
-
-        Ok(PreparedEmbeddingBatchRequest {
-            generate_embedding_stop_rx,
-            generated_embedding_tx,
-            inputs,
-            normalization_method,
-            oversized_documents,
-            slot_guard,
-        })
-    }
-}
diff --git a/paddler_agent/src/embedding_batch_rejection.rs b/paddler_agent/src/embedding_batch_rejection.rs
deleted file mode 100644
index 1e25cb031..000000000
--- a/paddler_agent/src/embedding_batch_rejection.rs
+++ /dev/null
@@ -1,129 +0,0 @@
-use std::num::TryFromIntError;
-
-use llama_cpp_bindings::batch_add_error::BatchAddError;
-use llama_cpp_bindings::error::DecodeError;
-use llama_cpp_bindings::error::EmbeddingsError;
-use llama_cpp_bindings::error::StringToTokenError;
-use log::error;
-use tokio::sync::mpsc;
-use tokio::sync::mpsc::error::SendError;
-
-use paddler_messaging::embedding_result::EmbeddingResult;
-
-use crate::rejection_description::rejection_description;
-use crate::send_result_or_warn::send_result_or_warn;
-
-#[derive(Debug, thiserror::Error)]
-pub enum EmbeddingBatchRejection {
-    #[error("embeddings are not enabled")]
-    EmbeddingsDisabled,
-
-    #[error("failed to tokenize embedding input {source_document_id:?}: {source}")]
-    InputTokenizationFailed {
-        source_document_id: String,
-        #[source]
-        source: StringToTokenError,
-    },
-
-    #[error("no model is loaded")]
-    ModelNotLoaded,
-
-    #[error("the scheduler is no longer accepting requests")]
-    SchedulerUnavailable,
-
-    #[error("failed to add an embedding input to the batch: {0}")]
-    BatchAssemblyFailed(#[source] BatchAddError),
-
-    #[error("the client stopped listening for embeddings: {0}")]
-    ClientDisconnected(#[source] SendError<EmbeddingResult>),
-
-    #[error("llama.cpp rejected the embedding batch: {0}")]
-    DecodeFailed(#[source] DecodeError),
-
-    #[error("{dimensions}-dimensional embedding is too long for RMS normalization: {source}")]
-    EmbeddingTooLongForRmsNormalization {
-        dimensions: usize,
-        #[source]
-        source: TryFromIntError,
-    },
-
-    #[error("failed to read the embedding of a batch sequence: {0}")]
-    EmbeddingsUnavailable(#[source] EmbeddingsError),
-
-    #[error("embedding batch sequence index does not fit in i32: {0}")]
-    SequenceIndexOutOfRange(#[source] TryFromIntError),
-}
-
-impl EmbeddingBatchRejection {
-    pub fn report(
-        self,
-        agent_name: Option<&str>,
-        generated_embedding_tx: &mpsc::UnboundedSender<EmbeddingResult>,
-    ) {
-        let message = rejection_description(agent_name, &self);
-
-        error!("{message}");
-
-        let result = match self {
-            Self::EmbeddingsDisabled => EmbeddingResult::EmbeddingsDisabled,
-            Self::ModelNotLoaded => EmbeddingResult::ModelNotLoaded(message),
-            Self::InputTokenizationFailed { .. }
-            | Self::SchedulerUnavailable
-            | Self::BatchAssemblyFailed(_)
-            | Self::ClientDisconnected(_)
-            | Self::DecodeFailed(_)
-            | Self::EmbeddingTooLongForRmsNormalization { .. }
-            | Self::EmbeddingsUnavailable(_)
-            | Self::SequenceIndexOutOfRange(_) => EmbeddingResult::Error(message),
-        };
-
-        send_result_or_warn(agent_name, generated_embedding_tx, result);
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use tokio::sync::mpsc;
-
-    use paddler_messaging::embedding_result::EmbeddingResult;
-
-    use super::EmbeddingBatchRejection;
-
-    #[test]
-    fn reports_disabled_embeddings() {
-        let (generated_embedding_tx, mut generated_embedding_rx) = mpsc::unbounded_channel();
-
-        EmbeddingBatchRejection::EmbeddingsDisabled.report(Some("agent"), &generated_embedding_tx);
-
-        assert_eq!(
-            generated_embedding_rx.try_recv(),
-            Ok(EmbeddingResult::EmbeddingsDisabled)
-        );
-    }
-
-    #[test]
-    fn reports_other_rejections_as_errors_naming_the_cause() {
-        let (generated_embedding_tx, mut generated_embedding_rx) = mpsc::unbounded_channel();
-
-        EmbeddingBatchRejection::SchedulerUnavailable
-            .report(Some("agent"), &generated_embedding_tx);
-
-        assert_eq!(
-            generated_embedding_rx.try_recv(),
-            Ok(EmbeddingResult::Error(
-                "agent: the scheduler is no longer accepting requests".to_owned()
-            ))
-        );
-    }
-
-    #[test]
-    fn tolerates_a_disconnected_client() {
-        let (generated_embedding_tx, generated_embedding_rx) = mpsc::unbounded_channel();
-
-        drop(generated_embedding_rx);
-
-        EmbeddingBatchRejection::EmbeddingsDisabled.report(None, &generated_embedding_tx);
-
-        assert!(generated_embedding_tx.is_closed());
-    }
-}
diff --git a/paddler_agent/src/enqueue_management_message.rs b/paddler_agent/src/enqueue_management_message.rs
new file mode 100644
index 000000000..a6a4f839e
--- /dev/null
+++ b/paddler_agent/src/enqueue_management_message.rs
@@ -0,0 +1,58 @@
+use log::error;
+use serde::Serialize;
+use serde_json::to_string;
+use tokio::sync::mpsc;
+use tokio_tungstenite::tungstenite::protocol::Message;
+
+use crate::enqueue_outbound_frame::enqueue_outbound_frame;
+use crate::management_connection_end::ManagementConnectionEnd;
+use crate::management_connection_step::ManagementConnectionStep;
+
+pub fn enqueue_management_message<TMessage>(
+    outbound_tx: &mpsc::UnboundedSender<Message>,
+    message: &TMessage,
+    step_after_enqueue: ManagementConnectionStep,
+) -> ManagementConnectionStep
+where
+    TMessage: Serialize,
+{
+    match to_string(message) {
+        Ok(serialized_message) => enqueue_outbound_frame(
+            outbound_tx,
+            Message::Text(serialized_message.into()),
+            step_after_enqueue,
+        ),
+        Err(serialization_error) => {
+            error!("Failed to serialize a management socket message: {serialization_error}");
+
+            ManagementConnectionStep::End(ManagementConnectionEnd::OutgoingMessageUnserializable)
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::collections::BTreeMap;
+
+    use tokio::sync::mpsc;
+
+    use super::enqueue_management_message;
+    use crate::management_connection_end::ManagementConnectionEnd;
+    use crate::management_connection_step::ManagementConnectionStep;
+
+    #[test]
+    fn ends_the_connection_when_the_message_cannot_be_serialized() {
+        let (outbound_tx, _outbound_rx) = mpsc::unbounded_channel();
+        let map_with_a_non_string_key = BTreeMap::from([(vec![1_u8], 1_u8)]);
+
+        assert!(matches!(
+            enqueue_management_message(
+                &outbound_tx,
+                &map_with_a_non_string_key,
+                ManagementConnectionStep::Continue,
+            ),
+            ManagementConnectionStep::End(management_connection_end)
+                if management_connection_end == ManagementConnectionEnd::OutgoingMessageUnserializable
+        ));
+    }
+}
diff --git a/paddler_agent/src/enqueue_outbound_frame.rs b/paddler_agent/src/enqueue_outbound_frame.rs
new file mode 100644
index 000000000..3c49bb464
--- /dev/null
+++ b/paddler_agent/src/enqueue_outbound_frame.rs
@@ -0,0 +1,50 @@
+use log::error;
+use tokio::sync::mpsc;
+use tokio::sync::mpsc::error::SendError;
+use tokio_tungstenite::tungstenite::protocol::Message;
+
+use crate::management_connection_end::ManagementConnectionEnd;
+use crate::management_connection_step::ManagementConnectionStep;
+
+pub fn enqueue_outbound_frame(
+    outbound_tx: &mpsc::UnboundedSender<Message>,
+    frame: Message,
+    step_after_enqueue: ManagementConnectionStep,
+) -> ManagementConnectionStep {
+    match outbound_tx.send(frame) {
+        Ok(()) => step_after_enqueue,
+        Err(SendError(undelivered_frame)) => {
+            error!("The management socket writer stopped before it could send {undelivered_frame}");
+
+            ManagementConnectionStep::End(ManagementConnectionEnd::OutgoingStreamFailed)
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use tokio::sync::mpsc;
+    use tokio_tungstenite::tungstenite::Bytes;
+    use tokio_tungstenite::tungstenite::protocol::Message;
+
+    use super::enqueue_outbound_frame;
+    use crate::management_connection_end::ManagementConnectionEnd;
+    use crate::management_connection_step::ManagementConnectionStep;
+
+    #[test]
+    fn ends_the_connection_when_the_writer_stopped() {
+        let (outbound_tx, outbound_rx) = mpsc::unbounded_channel();
+
+        drop(outbound_rx);
+
+        assert!(matches!(
+            enqueue_outbound_frame(
+                &outbound_tx,
+                Message::Ping(Bytes::new()),
+                ManagementConnectionStep::Continue,
+            ),
+            ManagementConnectionStep::End(management_connection_end)
+                if management_connection_end == ManagementConnectionEnd::OutgoingStreamFailed
+        ));
+    }
+}
diff --git a/paddler_agent/src/finished_serving_startup.rs b/paddler_agent/src/finished_serving_startup.rs
new file mode 100644
index 000000000..1df50fb7f
--- /dev/null
+++ b/paddler_agent/src/finished_serving_startup.rs
@@ -0,0 +1,7 @@
+use crate::agent_applicable_state::AgentApplicableState;
+use crate::replacement_outcome::ReplacementOutcome;
+
+pub struct FinishedServingStartup {
+    pub agent_applicable_state: AgentApplicableState,
+    pub outcome: ReplacementOutcome,
+}
diff --git a/paddler_agent/src/forward_management_socket_messages.rs b/paddler_agent/src/forward_management_socket_messages.rs
deleted file mode 100644
index 7cb64adaa..000000000
--- a/paddler_agent/src/forward_management_socket_messages.rs
+++ /dev/null
@@ -1,155 +0,0 @@
-use std::fmt::Display;
-
-use futures_util::Sink;
-use futures_util::SinkExt as _;
-use log::error;
-use serde_json::to_string;
-use tokio::select;
-use tokio::sync::mpsc;
-use tokio_tungstenite::tungstenite::protocol::Message;
-use tokio_util::sync::CancellationToken;
-
-use paddler_messaging::management_socket::balancer::message::Message as ManagementJsonRpcMessage;
-
-async fn write_message<TSink>(write: &mut TSink, message: ManagementJsonRpcMessage) -> bool
-where
-    TSink: Sink<Message> + Unpin,
-    TSink::Error: Display,
-{
-    match to_string(&message) {
-        Ok(serialized_message) => {
-            match write.send(Message::Text(serialized_message.into())).await {
-                Ok(()) => true,
-                Err(err) => {
-                    error!("Failed to send a management socket message: {err}");
-
-                    false
-                }
-            }
-        }
-        Err(err) => {
-            error!("Failed to serialize a management socket message: {err}");
-
-            false
-        }
-    }
-}
-
-pub async fn forward_management_socket_messages<TSink>(
-    connection_close: CancellationToken,
-    mut message_rx: mpsc::UnboundedReceiver<ManagementJsonRpcMessage>,
-    mut write: TSink,
-) -> TSink
-where
-    TSink: Sink<Message> + Unpin,
-    TSink::Error: Display,
-{
-    loop {
-        select! {
-            () = connection_close.cancelled() => {
-                while let Ok(message) = message_rx.try_recv()
-                    && write_message(&mut write, message).await
-                {}
-
-                break;
-            }
-            message = message_rx.recv() => {
-                let Some(message) = message else {
-                    break;
-                };
-
-                if !write_message(&mut write, message).await {
-                    break;
-                }
-            }
-        }
-    }
-
-    write
-}
-
-#[cfg(test)]
-mod tests {
-    use futures::channel::mpsc as sink_channel;
-    use futures_util::StreamExt as _;
-    use serde_json::to_string;
-    use tokio::sync::mpsc;
-    use tokio_tungstenite::tungstenite::protocol::Message;
-    use tokio_util::sync::CancellationToken;
-
-    use paddler_messaging::management_socket::balancer::message::Message as ManagementJsonRpcMessage;
-    use paddler_messaging::management_socket::balancer::notification::Notification;
-
-    use super::forward_management_socket_messages;
-
-    fn deregister_agent() -> ManagementJsonRpcMessage {
-        ManagementJsonRpcMessage::Notification(Notification::DeregisterAgent)
-    }
-
-    fn deregister_agent_text() -> Message {
-        Message::Text(
-            to_string(&deregister_agent())
-                .expect("a deregistration must serialize")
-                .into(),
-        )
-    }
-
-    #[tokio::test]
-    async fn forwards_queued_messages_until_every_sender_is_gone() {
-        let (sink_tx, sink_rx) = sink_channel::unbounded::<Message>();
-        let (message_tx, message_rx) = mpsc::unbounded_channel();
-
-        message_tx
-            .send(deregister_agent())
-            .expect("the forwarder must still receive messages");
-        drop(message_tx);
-
-        drop(
-            forward_management_socket_messages(CancellationToken::new(), message_rx, sink_tx).await,
-        );
-
-        assert_eq!(
-            sink_rx.collect::<Vec<Message>>().await,
-            vec![deregister_agent_text()]
-        );
-    }
-
-    #[tokio::test]
-    async fn drains_queued_messages_when_the_connection_closes() {
-        let (sink_tx, sink_rx) = sink_channel::unbounded::<Message>();
-        let (message_tx, message_rx) = mpsc::unbounded_channel();
-        let connection_close = CancellationToken::new();
-
-        for _ in 0..2 {
-            message_tx
-                .send(deregister_agent())
-                .expect("the forwarder must still receive messages");
-        }
-        connection_close.cancel();
-
-        drop(forward_management_socket_messages(connection_close, message_rx, sink_tx).await);
-
-        assert_eq!(
-            sink_rx.collect::<Vec<Message>>().await,
-            vec![deregister_agent_text(), deregister_agent_text()]
-        );
-        drop(message_tx);
-    }
-
-    #[tokio::test]
-    async fn stops_forwarding_when_the_socket_rejects_a_message() {
-        let (sink_tx, sink_rx) = sink_channel::unbounded::<Message>();
-        let (message_tx, message_rx) = mpsc::unbounded_channel();
-
-        drop(sink_rx);
-        message_tx
-            .send(deregister_agent())
-            .expect("the forwarder must still receive messages");
-
-        let returned_sink =
-            forward_management_socket_messages(CancellationToken::new(), message_rx, sink_tx).await;
-
-        assert!(returned_sink.is_closed());
-        drop(message_tx);
-    }
-}
diff --git a/paddler_agent/src/forward_scheduler_command.rs b/paddler_agent/src/forward_scheduler_command.rs
deleted file mode 100644
index 4dfaacac3..000000000
--- a/paddler_agent/src/forward_scheduler_command.rs
+++ /dev/null
@@ -1,174 +0,0 @@
-use std::sync::mpsc::SendError;
-use std::sync::mpsc::Sender;
-
-use log::info;
-
-use crate::continuous_batch_scheduler_command::ContinuousBatchSchedulerCommand;
-use crate::embedding_batch_rejection::EmbeddingBatchRejection;
-use crate::generation_request_rejection::GenerationRequestRejection;
-
-pub fn forward_scheduler_command(
-    scheduler_command_tx: &Sender<ContinuousBatchSchedulerCommand>,
-    agent_name: Option<&str>,
-    command: ContinuousBatchSchedulerCommand,
-) {
-    if let Err(SendError(unsent_command)) = scheduler_command_tx.send(command) {
-        match unsent_command {
-            ContinuousBatchSchedulerCommand::Generate(request) => {
-                GenerationRequestRejection::SchedulerUnavailable
-                    .report(agent_name, &request.generated_tokens_tx);
-            }
-            ContinuousBatchSchedulerCommand::GenerateEmbeddingBatch(request) => {
-                EmbeddingBatchRejection::SchedulerUnavailable
-                    .report(agent_name, &request.generated_embedding_tx);
-            }
-            ContinuousBatchSchedulerCommand::Shutdown => {
-                info!("{agent_name:?}: scheduler stopped before the shutdown command arrived");
-            }
-        }
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use std::mem::discriminant;
-    use std::num::NonZeroU32;
-    use std::sync::Arc;
-    use std::sync::mpsc::channel;
-
-    use llama_cpp_bindings::BareJsonToolCalls;
-    use llama_cpp_bindings::StreamingMarkers;
-    use tokio::sync::mpsc;
-
-    use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
-    use paddler_inference_parameters::inference_parameters::InferenceParameters;
-    use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
-    use paddler_messaging::embedding_result::EmbeddingResult;
-    use paddler_messaging::generated_token_result::GeneratedTokenResult;
-
-    use super::forward_scheduler_command;
-    use crate::continuous_batch_scheduler_command::ContinuousBatchSchedulerCommand;
-    use crate::grammar_sampling::GrammarSampling;
-    use crate::prepared_embedding_batch_request::PreparedEmbeddingBatchRequest;
-    use crate::prepared_generation_request::PreparedGenerationRequest;
-    use crate::prepared_prompt::PreparedPrompt;
-    use crate::sampler_chain_factory::SamplerChainFactory;
-    use crate::slot_guard::SlotGuard;
-    use crate::token_classification::TokenClassification;
-    use crate::token_sampling::TokenSampling;
-    use crate::tool_call_handling::ToolCallHandling;
-
-    fn slot_guard() -> SlotGuard {
-        SlotGuard::new(Arc::new(SlotAggregatedStatus::new(1)))
-    }
-
-    #[test]
-    fn delivers_command_to_a_running_scheduler() {
-        let (scheduler_command_tx, scheduler_command_rx) = channel();
-
-        forward_scheduler_command(
-            &scheduler_command_tx,
-            None,
-            ContinuousBatchSchedulerCommand::Shutdown,
-        );
-
-        assert_eq!(
-            scheduler_command_rx
-                .try_recv()
-                .map(|command| discriminant(&command)),
-            Ok(discriminant(&ContinuousBatchSchedulerCommand::Shutdown))
-        );
-    }
-
-    #[test]
-    fn rejects_generation_when_the_scheduler_stopped() {
-        let (scheduler_command_tx, scheduler_command_rx) = channel();
-        let (generated_tokens_tx, mut generated_tokens_rx) = mpsc::unbounded_channel();
-        let (_generate_tokens_stop_tx, generate_tokens_stop_rx) = mpsc::unbounded_channel();
-
-        drop(scheduler_command_rx);
-
-        forward_scheduler_command(
-            &scheduler_command_tx,
-            Some("agent"),
-            ContinuousBatchSchedulerCommand::Generate(Box::new(PreparedGenerationRequest {
-                generate_tokens_stop_rx,
-                generated_tokens_tx,
-                max_tokens: NonZeroU32::new(1).unwrap(),
-                prompt: PreparedPrompt::TextTokens(Vec::new()),
-                slot_guard: slot_guard(),
-                token_classification: TokenClassification {
-                    bare_json_tool_calls: BareJsonToolCalls::Ignore,
-                    streaming_markers: Arc::new(StreamingMarkers::default()),
-                },
-                token_sampling: TokenSampling {
-                    chain: SamplerChainFactory {
-                        inference_parameters: InferenceParameters::default(),
-                        n_vocab: 1,
-                    }
-                    .create(0)
-                    .unwrap(),
-                    grammar: GrammarSampling::Unconstrained,
-                },
-                tool_call_handling: ToolCallHandling::Streamed,
-            })),
-        );
-
-        assert_eq!(
-            generated_tokens_rx.try_recv(),
-            Ok(GeneratedTokenResult::SchedulerUnavailable(
-                "agent: the scheduler is no longer accepting requests".to_owned()
-            ))
-        );
-    }
-
-    #[test]
-    fn rejects_embedding_batch_when_the_scheduler_stopped() {
-        let (scheduler_command_tx, scheduler_command_rx) = channel();
-        let (generated_embedding_tx, mut generated_embedding_rx) = mpsc::unbounded_channel();
-        let (_generate_embedding_stop_tx, generate_embedding_stop_rx) = mpsc::unbounded_channel();
-
-        drop(scheduler_command_rx);
-
-        forward_scheduler_command(
-            &scheduler_command_tx,
-            Some("agent"),
-            ContinuousBatchSchedulerCommand::GenerateEmbeddingBatch(
-                PreparedEmbeddingBatchRequest {
-                    generate_embedding_stop_rx,
-                    generated_embedding_tx,
-                    inputs: Vec::new(),
-                    normalization_method: EmbeddingNormalizationMethod::None,
-                    oversized_documents: Vec::new(),
-                    slot_guard: slot_guard(),
-                },
-            ),
-        );
-
-        assert_eq!(
-            generated_embedding_rx.try_recv(),
-            Ok(EmbeddingResult::Error(
-                "agent: the scheduler is no longer accepting requests".to_owned()
-            ))
-        );
-    }
-
-    #[test]
-    fn accepts_shutdown_after_the_scheduler_stopped() {
-        let (scheduler_command_tx, scheduler_command_rx) = channel();
-
-        drop(scheduler_command_rx);
-
-        forward_scheduler_command(
-            &scheduler_command_tx,
-            None,
-            ContinuousBatchSchedulerCommand::Shutdown,
-        );
-
-        assert!(
-            scheduler_command_tx
-                .send(ContinuousBatchSchedulerCommand::Shutdown)
-                .is_err()
-        );
-    }
-}
diff --git a/paddler_agent/src/generation_request_preparer.rs b/paddler_agent/src/generation_request_preparer.rs
deleted file mode 100644
index a69b5b351..000000000
--- a/paddler_agent/src/generation_request_preparer.rs
+++ /dev/null
@@ -1,214 +0,0 @@
-use std::num::NonZeroU32;
-use std::sync::Arc;
-
-use llama_cpp_bindings::BareJsonToolCalls;
-use llama_cpp_bindings::ChatMessageParser;
-use llama_cpp_bindings::model::LlamaModel;
-use llama_cpp_bindings::mtmd::MtmdBitmap;
-use rand::Rng as _;
-use rand::rng;
-use serde_json::to_string;
-
-use paddler_image_decoder::decoded_image::DecodedImage;
-use paddler_messaging::chat_template_conversation::ChatTemplateConversation;
-use paddler_messaging::generated_token_result::GeneratedTokenResult;
-use paddler_messaging::grammar_constraint::GrammarConstraint;
-use paddler_messaging::image_url::ImageUrl;
-use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
-use paddler_messaging::request_params::continue_from_conversation_history_params::tool::Tool;
-use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters_schema::validated_parameters_schema::ValidatedParametersSchema;
-use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
-use paddler_tool_call_validator::tool_call_validator::ToolCallValidator;
-
-use crate::agent_request::AgentRequest;
-use crate::chat_prompt_render_request::ChatPromptRenderRequest;
-use crate::converts_to_mtmd_bitmap::ConvertsToMtmdBitmap;
-use crate::generation_request_rejection::GenerationRequestRejection;
-use crate::image_input::ImageInput;
-use crate::prepared_generation_request::PreparedGenerationRequest;
-use crate::prepared_prompt::PreparedPrompt;
-use crate::prompt_modality::PromptModality;
-use crate::prompt_tokenizer::PromptTokenizer;
-use crate::require_grammar_compatible_with_thinking::require_grammar_compatible_with_thinking;
-use crate::resolve_grammar::resolve_grammar;
-use crate::sampler_chain_factory::SamplerChainFactory;
-use crate::token_classification::TokenClassification;
-use crate::token_generation::TokenGeneration;
-use crate::token_generation_support::TokenGenerationSupport;
-use crate::token_sampling::TokenSampling;
-use crate::tool_call_handling::ToolCallHandling;
-use crate::tool_call_pipeline::ToolCallPipeline;
-
-pub struct GenerationRequestPreparer {
-    pub image_input: ImageInput,
-    pub image_resize_to_fit: NonZeroU32,
-    pub model: Arc<LlamaModel>,
-    pub prompt_tokenizer: PromptTokenizer,
-    pub sampler_chain_factory: SamplerChainFactory,
-    pub token_generation: TokenGeneration,
-}
-
-impl GenerationRequestPreparer {
-    pub fn prepare_raw_prompt(
-        &self,
-        AgentRequest {
-            params:
-                ContinueFromRawPromptParams {
-                    grammar,
-                    max_tokens,
-                    raw_prompt,
-                },
-            response_tx: generated_tokens_tx,
-            slot_guard,
-            stop_rx: generate_tokens_stop_rx,
-        }: AgentRequest<ContinueFromRawPromptParams, GeneratedTokenResult>,
-    ) -> Result<PreparedGenerationRequest, GenerationRequestRejection> {
-        let TokenGenerationSupport {
-            streaming_markers, ..
-        } = self.token_generation.require_enabled()?;
-        let prompt = PreparedPrompt::TextTokens(self.prompt_tokenizer.tokenize(&raw_prompt)?);
-        let token_sampling = self.build_token_sampling(grammar)?;
-
-        Ok(PreparedGenerationRequest {
-            generate_tokens_stop_rx,
-            generated_tokens_tx,
-            max_tokens,
-            prompt,
-            slot_guard,
-            token_classification: TokenClassification {
-                bare_json_tool_calls: BareJsonToolCalls::Ignore,
-                streaming_markers: streaming_markers.clone(),
-            },
-            token_sampling,
-            tool_call_handling: ToolCallHandling::Streamed,
-        })
-    }
-
-    pub fn prepare_conversation_history(
-        &self,
-        AgentRequest {
-            response_tx: generated_tokens_tx,
-            stop_rx: generate_tokens_stop_rx,
-            params:
-                ContinueFromConversationHistoryParams {
-                    add_generation_prompt,
-                    conversation_history,
-                    enable_thinking,
-                    grammar,
-                    max_tokens,
-                    parse_tool_calls,
-                    tools,
-                },
-            slot_guard,
-        }: AgentRequest<
-            ContinueFromConversationHistoryParams<ValidatedParametersSchema>,
-            GeneratedTokenResult,
-        >,
-    ) -> Result<PreparedGenerationRequest, GenerationRequestRejection> {
-        let TokenGenerationSupport {
-            chat_prompt_renderer,
-            streaming_markers,
-        } = self.token_generation.require_enabled()?;
-
-        require_grammar_compatible_with_thinking(grammar.as_ref(), enable_thinking)?;
-
-        let ChatTemplateConversation {
-            image_urls,
-            messages,
-        } = conversation_history
-            .into_chat_template_conversation(&chat_prompt_renderer.media_marker);
-        let prompt_modality = self.image_input.prompt_modality_for(&image_urls)?;
-
-        let raw_prompt = chat_prompt_renderer.render(ChatPromptRenderRequest {
-            add_generation_prompt,
-            enable_thinking,
-            messages: &messages,
-            tools: &tools,
-        })?;
-
-        let prompt = match prompt_modality {
-            PromptModality::Multimodal(multimodal_prompt_support) => {
-                let add_special_tokens = !self
-                    .prompt_tokenizer
-                    .renders_its_own_bos_token(&raw_prompt)?;
-
-                PreparedPrompt::Multimodal(
-                    multimodal_prompt_support.prepare_prompt(
-                        &image_urls
-                            .iter()
-                            .map(|image_url| self.decode_image_bitmap(image_url))
-                            .collect::<Result<Vec<MtmdBitmap>, GenerationRequestRejection>>()?,
-                        raw_prompt,
-                        add_special_tokens,
-                    )?,
-                )
-            }
-            PromptModality::TextOnly => {
-                PreparedPrompt::TextTokens(self.prompt_tokenizer.tokenize(&raw_prompt)?)
-            }
-        };
-
-        let token_sampling = self.build_token_sampling(grammar)?;
-        let tool_call_handling = if parse_tool_calls {
-            ToolCallHandling::Parsed(self.build_tool_call_pipeline(&tools)?)
-        } else {
-            ToolCallHandling::Streamed
-        };
-        let bare_json_tool_calls = if tools.is_empty() {
-            BareJsonToolCalls::Ignore
-        } else {
-            BareJsonToolCalls::Detect
-        };
-
-        Ok(PreparedGenerationRequest {
-            generate_tokens_stop_rx,
-            generated_tokens_tx,
-            max_tokens,
-            prompt,
-            slot_guard,
-            token_classification: TokenClassification {
-                bare_json_tool_calls,
-                streaming_markers: streaming_markers.clone(),
-            },
-            token_sampling,
-            tool_call_handling,
-        })
-    }
-
-    fn decode_image_bitmap(
-        &self,
-        image_url: &ImageUrl,
-    ) -> Result<MtmdBitmap, GenerationRequestRejection> {
-        DecodedImage::from_data_uri(image_url, self.image_resize_to_fit)
-            .map_err(GenerationRequestRejection::ImageDecodingFailed)?
-            .to_mtmd_bitmap()
-            .map_err(GenerationRequestRejection::ImageBitmapCreationFailed)
-    }
-
-    fn build_token_sampling(
-        &self,
-        grammar: Option<GrammarConstraint>,
-    ) -> Result<TokenSampling, GenerationRequestRejection> {
-        let grammar = resolve_grammar(grammar, &self.model)?;
-        let chain = self
-            .sampler_chain_factory
-            .create(rng().random::<u32>())
-            .map_err(GenerationRequestRejection::SamplerChainCreationFailed)?;
-
-        Ok(TokenSampling { chain, grammar })
-    }
-
-    fn build_tool_call_pipeline(
-        &self,
-        tools: &[Tool<ValidatedParametersSchema>],
-    ) -> Result<ToolCallPipeline, GenerationRequestRejection> {
-        let validator = ToolCallValidator::from_tools(tools)
-            .map_err(GenerationRequestRejection::ToolSchemaInvalid)?;
-        let tools_json =
-            to_string(tools).map_err(GenerationRequestRejection::ToolsSerializationFailed)?;
-        let chat_message_parser = ChatMessageParser::new(&self.model, &tools_json)
-            .map_err(GenerationRequestRejection::ToolCallParserCreationFailed)?;
-
-        Ok(ToolCallPipeline::new(chat_message_parser, validator))
-    }
-}
diff --git a/paddler_agent/src/generation_request_rejection.rs b/paddler_agent/src/generation_request_rejection.rs
deleted file mode 100644
index 03bcb002c..000000000
--- a/paddler_agent/src/generation_request_rejection.rs
+++ /dev/null
@@ -1,559 +0,0 @@
-use std::num::TryFromIntError;
-
-use llama_cpp_bindings::batch_add_error::BatchAddError;
-use llama_cpp_bindings::error::ClearKvCacheSeqError;
-use llama_cpp_bindings::error::EvalMultimodalChunksError;
-use llama_cpp_bindings::error::GrammarError;
-use llama_cpp_bindings::error::JsonSchemaToGrammarError;
-use llama_cpp_bindings::error::ParseChatMessageError;
-use llama_cpp_bindings::error::SamplingError;
-use llama_cpp_bindings::error::StringToTokenError;
-use llama_cpp_bindings::mtmd::MtmdBitmapError;
-use llama_cpp_bindings::mtmd::MtmdEvalError;
-use llama_cpp_bindings::mtmd::MtmdTokenizeError;
-use log::error;
-use minijinja::Error as MinijinjaError;
-use serde_json::Error;
-use tokio::sync::mpsc;
-
-use paddler_image_decoder::decoded_image_error::DecodedImageError;
-use paddler_messaging::generated_token_result::GeneratedTokenResult;
-use paddler_messaging::oversized_media_details::OversizedMediaDetails;
-use paddler_messaging::oversized_prompt_details::OversizedPromptDetails;
-use paddler_tool_call_validator::tool_call_validation_error::ToolCallValidationError;
-
-use crate::rejection_description::rejection_description;
-use crate::send_result_or_warn::send_result_or_warn;
-
-#[derive(Debug, thiserror::Error)]
-pub enum GenerationRequestRejection {
-    #[error("Grammar constraints are incompatible with thinking mode")]
-    GrammarIncompatibleWithThinking,
-
-    #[error("Failed to convert JSON schema to grammar: {0}")]
-    GrammarConversionFailed(#[source] JsonSchemaToGrammarError),
-
-    #[error("token generation is disabled because this agent is running in embeddings-only mode")]
-    TokenGenerationDisabled,
-
-    #[error("failed to decode images: {0}")]
-    ImageDecodingFailed(#[source] DecodedImageError),
-
-    #[error("failed to create image bitmap: {0}")]
-    ImageBitmapCreationFailed(#[source] MtmdBitmapError),
-
-    #[error("received images but model does not support multimodal input")]
-    MultimodalNotSupported,
-
-    #[error("no model is loaded")]
-    ModelNotLoaded,
-
-    #[error("failed to render chat template: {0}")]
-    ChatTemplateRenderingFailed(#[source] MinijinjaError),
-
-    #[error("{0}")]
-    ToolSchemaInvalid(#[source] ToolCallValidationError),
-
-    #[error("failed to serialize tools: {0}")]
-    ToolsSerializationFailed(#[source] Error),
-
-    #[error("failed to build this model's tool call parser: {0}")]
-    ToolCallParserCreationFailed(#[source] ParseChatMessageError),
-
-    #[error("failed to tokenize prompt: {0}")]
-    PromptTokenizationFailed(#[source] StringToTokenError),
-
-    #[error(
-        "prompt has {} tokens but each sequence holds {} tokens",
-        details.prompt_tokens,
-        details.sequence_context_size
-    )]
-    PromptExceedsContextSize { details: OversizedPromptDetails },
-
-    #[error("the scheduler is no longer accepting requests")]
-    SchedulerUnavailable,
-
-    #[error("no available sequence slots, all slots are busy")]
-    NoSequenceSlotAvailable,
-
-    #[error("invalid grammar: {0}")]
-    GrammarInvalid(#[source] GrammarError),
-    #[error("failed to initialize grammar sampler: {0}")]
-    GrammarSamplerInitializationFailed(#[source] GrammarError),
-
-    #[error("failed to create sampler chain: {0}")]
-    SamplerChainCreationFailed(#[source] SamplingError),
-
-    #[error("failed to tokenize multimodal input: {0}")]
-    MultimodalTokenizationFailed(#[source] MtmdTokenizeError),
-
-    #[error(
-        "media chunk has {} tokens but one micro batch holds {} tokens",
-        details.media_tokens,
-        details.micro_batch_tokens
-    )]
-    MediaExceedsMicroBatch { details: OversizedMediaDetails },
-    #[error("failed to check media chunks against the micro batch: {0}")]
-    MediaMicroBatchCheckFailed(#[source] MtmdEvalError),
-
-    #[error("failed to ingest multimodal prompt: {0}")]
-    MultimodalIngestionFailed(#[source] EvalMultimodalChunksError),
-
-    #[error("failed to add a token to the batch: {0}")]
-    BatchAssemblyFailed(#[source] BatchAddError),
-
-    #[error("failed to clear the KV cache of the sequence: {0}")]
-    KvCacheClearFailed(#[source] ClearKvCacheSeqError),
-
-    #[error("sequence id {sequence_id} cannot address a KV cache sequence: {source}")]
-    SequenceIdOutOfRange {
-        sequence_id: i32,
-        #[source]
-        source: TryFromIntError,
-    },
-}
-
-impl GenerationRequestRejection {
-    #[must_use]
-    pub const fn for_gbnf_grammar_error(grammar_error: GrammarError) -> Self {
-        match grammar_error {
-            GrammarError::GrammarContainsNul(_)
-            | GrammarError::GrammarRejected(_)
-            | GrammarError::RootContainsNul(_)
-            | GrammarError::RootNotFound => Self::GrammarInvalid(grammar_error),
-            GrammarError::FfiContract(_)
-            | GrammarError::FfiStatus(_)
-            | GrammarError::GrammarMalformed
-            | GrammarError::InvalidTriggerPattern { .. }
-            | GrammarError::LazyGrammarMalformed
-            | GrammarError::LlamaCppOutOfMemory
-            | GrammarError::LlguidanceFactoryUnavailable { .. }
-            | GrammarError::LlguidanceGrammarInvalid { .. }
-            | GrammarError::LlguidanceParserUnavailable { .. }
-            | GrammarError::NotEnoughMemory
-            | GrammarError::Reported { .. }
-            | GrammarError::SamplerInitialization(_)
-            | GrammarError::SequenceBreakerContainsNul(_)
-            | GrammarError::TokEnvUnavailable(_)
-            | GrammarError::TriggerPatternContainsNul(_) => {
-                Self::GrammarSamplerInitializationFailed(grammar_error)
-            }
-        }
-    }
-
-    #[must_use]
-    pub fn for_micro_batch_fit_error(fit_error: MtmdEvalError) -> Self {
-        match fit_error {
-            MtmdEvalError::NonCausalChunkExceedsMicroBatch(mismatch) => {
-                Self::MediaExceedsMicroBatch {
-                    details: OversizedMediaDetails {
-                        media_tokens: mismatch.chunk_tokens,
-                        micro_batch_tokens: mismatch.micro_batch_tokens,
-                    },
-                }
-            }
-            other_fit_error => Self::MediaMicroBatchCheckFailed(other_fit_error),
-        }
-    }
-
-    pub fn report(
-        self,
-        agent_name: Option<&str>,
-        generated_tokens_tx: &mpsc::UnboundedSender<GeneratedTokenResult>,
-    ) {
-        send_result_or_warn(
-            agent_name,
-            generated_tokens_tx,
-            self.into_generated_token_result(agent_name),
-        );
-    }
-
-    #[must_use]
-    pub fn into_generated_token_result(self, agent_name: Option<&str>) -> GeneratedTokenResult {
-        let message = rejection_description(agent_name, &self);
-
-        error!("{message}");
-
-        match self {
-            Self::GrammarIncompatibleWithThinking => {
-                GeneratedTokenResult::GrammarIncompatibleWithThinking(message)
-            }
-            Self::GrammarConversionFailed(_) | Self::GrammarInvalid(_) => {
-                GeneratedTokenResult::GrammarSyntaxError(message)
-            }
-            Self::TokenGenerationDisabled => GeneratedTokenResult::TokenGenerationDisabled(message),
-            Self::ImageDecodingFailed(_) | Self::ImageBitmapCreationFailed(_) => {
-                GeneratedTokenResult::ImageDecodingFailed(message)
-            }
-            Self::ModelNotLoaded => GeneratedTokenResult::ModelNotLoaded(message),
-            Self::MultimodalNotSupported => GeneratedTokenResult::MultimodalNotSupported(message),
-            Self::ChatTemplateRenderingFailed(_) | Self::ToolCallParserCreationFailed(_) => {
-                GeneratedTokenResult::ChatTemplateError(message)
-            }
-            Self::ToolSchemaInvalid(_) => GeneratedTokenResult::ToolSchemaInvalid(message),
-            Self::GrammarSamplerInitializationFailed(_) => {
-                GeneratedTokenResult::GrammarInitializationFailed(message)
-            }
-            Self::MediaExceedsMicroBatch { details } => {
-                GeneratedTokenResult::MediaExceedsMicroBatch(details)
-            }
-            Self::PromptExceedsContextSize { details } => {
-                GeneratedTokenResult::PromptExceedsContextSize(details)
-            }
-            Self::ToolsSerializationFailed(_) => {
-                GeneratedTokenResult::ToolsSerializationFailed(message)
-            }
-            Self::MediaMicroBatchCheckFailed(_) => {
-                GeneratedTokenResult::MediaMicroBatchCheckFailed(message)
-            }
-            Self::PromptTokenizationFailed(_) => {
-                GeneratedTokenResult::PromptTokenizationFailed(message)
-            }
-            Self::SchedulerUnavailable => GeneratedTokenResult::SchedulerUnavailable(message),
-            Self::NoSequenceSlotAvailable => GeneratedTokenResult::NoSequenceSlotAvailable(message),
-            Self::SamplerChainCreationFailed(_) => {
-                GeneratedTokenResult::SamplerChainCreationFailed(message)
-            }
-            Self::MultimodalTokenizationFailed(_) => {
-                GeneratedTokenResult::MultimodalTokenizationFailed(message)
-            }
-            Self::MultimodalIngestionFailed(_) => {
-                GeneratedTokenResult::MultimodalIngestionFailed(message)
-            }
-            Self::BatchAssemblyFailed(_) => GeneratedTokenResult::BatchAssemblyFailed(message),
-            Self::KvCacheClearFailed(_) => GeneratedTokenResult::KvCacheClearFailed(message),
-            Self::SequenceIdOutOfRange { .. } => {
-                GeneratedTokenResult::SequenceIdOutOfRange(message)
-            }
-        }
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use llama_cpp_bindings::batch_add_error::BatchAddError;
-    use llama_cpp_bindings::error::ClearKvCacheSeqError;
-    use llama_cpp_bindings::error::EvalMultimodalChunksError;
-    use llama_cpp_bindings::error::FfiContractError;
-    use llama_cpp_bindings::error::FfiStatusError;
-    use llama_cpp_bindings::error::GrammarError;
-    use llama_cpp_bindings::error::JsonSchemaToGrammarError;
-    use llama_cpp_bindings::error::ParseChatMessageError;
-    use llama_cpp_bindings::error::SamplingError;
-    use llama_cpp_bindings::mtmd::MtmdEvalError;
-    use llama_cpp_bindings::mtmd::NonCausalChunkMicroBatchMismatch;
-    use minijinja::Error as MinijinjaError;
-    use minijinja::ErrorKind as MinijinjaErrorKind;
-    use serde_json::Map;
-    use serde_json::Value;
-    use serde_json::from_str;
-    use serde_json::json;
-    use tokio::sync::mpsc;
-
-    use paddler_image_decoder::decoded_image_error::DecodedImageError;
-    use paddler_messaging::generated_token_result::GeneratedTokenResult;
-    use paddler_messaging::oversized_media_details::OversizedMediaDetails;
-    use paddler_messaging::oversized_prompt_details::OversizedPromptDetails;
-    use paddler_messaging::request_params::continue_from_conversation_history_params::tool::Tool;
-    use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::FunctionCall;
-    use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::function::Function;
-    use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters::Parameters;
-    use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters_schema::validated_parameters_schema::ValidatedParametersSchema;
-    use paddler_tool_call_validator::tool_call_validator::ToolCallValidator;
-
-    use super::GenerationRequestRejection;
-
-    struct ExpectedReport {
-        rejection: GenerationRequestRejection,
-        into_result: fn(String) -> GeneratedTokenResult,
-    }
-
-    fn reported(rejection: GenerationRequestRejection) -> GeneratedTokenResult {
-        let (generated_tokens_tx, mut generated_tokens_rx) = mpsc::unbounded_channel();
-
-        rejection.report(Some("agent"), &generated_tokens_tx);
-
-        generated_tokens_rx.try_recv().unwrap()
-    }
-
-    fn agent_message(description: &str) -> String {
-        format!("agent: {description}")
-    }
-
-    #[test]
-    fn reports_thinking_incompatibility_with_the_agent_name() {
-        assert_eq!(
-            reported(GenerationRequestRejection::GrammarIncompatibleWithThinking),
-            GeneratedTokenResult::GrammarIncompatibleWithThinking(agent_message(
-                "Grammar constraints are incompatible with thinking mode"
-            ))
-        );
-    }
-
-    #[test]
-    fn reports_grammar_conversion_failure_as_grammar_syntax_error() {
-        let conversion_error = JsonSchemaToGrammarError::NotEnoughMemory;
-        let expected_message = agent_message(&format!(
-            "Failed to convert JSON schema to grammar: {conversion_error}"
-        ));
-
-        assert_eq!(
-            reported(GenerationRequestRejection::GrammarConversionFailed(
-                conversion_error
-            )),
-            GeneratedTokenResult::GrammarSyntaxError(expected_message)
-        );
-    }
-
-    #[test]
-    fn reports_disabled_token_generation() {
-        assert_eq!(
-            reported(GenerationRequestRejection::TokenGenerationDisabled),
-            GeneratedTokenResult::TokenGenerationDisabled(agent_message(
-                "token generation is disabled because this agent is running in embeddings-only mode"
-            ))
-        );
-    }
-
-    #[test]
-    fn reports_image_failures_as_image_decoding_failure() {
-        assert_eq!(
-            reported(GenerationRequestRejection::ImageDecodingFailed(
-                DecodedImageError::MissingCommaSeparator,
-            )),
-            GeneratedTokenResult::ImageDecodingFailed(agent_message(
-                "failed to decode images: Invalid data URI: missing comma separator"
-            ))
-        );
-    }
-
-    #[test]
-    fn reports_missing_multimodal_support() {
-        assert_eq!(
-            reported(GenerationRequestRejection::MultimodalNotSupported),
-            GeneratedTokenResult::MultimodalNotSupported(agent_message(
-                "received images but model does not support multimodal input"
-            ))
-        );
-    }
-
-    #[test]
-    fn reports_prompt_rendering_failures_as_chat_template_error() {
-        let rendering_error =
-            MinijinjaError::new(MinijinjaErrorKind::UndefinedError, "missing variable");
-        let expected_message = agent_message(&format!(
-            "failed to render chat template: {rendering_error}"
-        ));
-
-        assert_eq!(
-            reported(GenerationRequestRejection::ChatTemplateRenderingFailed(
-                rendering_error
-            )),
-            GeneratedTokenResult::ChatTemplateError(expected_message)
-        );
-    }
-
-    #[test]
-    fn reports_invalid_tool_schema() {
-        let mut properties = Map::new();
-
-        properties.insert("location".to_owned(), json!({"type": 123}));
-
-        let schema_error = ToolCallValidator::from_tools(&[Tool::Function(FunctionCall {
-            function: Function {
-                name: "get_weather".to_owned(),
-                description: "fetch weather".to_owned(),
-                parameters: Parameters::Schema(ValidatedParametersSchema {
-                    schema_type: "object".to_owned(),
-                    properties: Some(properties),
-                    ..ValidatedParametersSchema::default()
-                }),
-            },
-        })])
-        .err()
-        .expect("a property typed as a number is not a valid JSON Schema");
-        let expected_message = agent_message(&schema_error.to_string());
-
-        assert_eq!(
-            reported(GenerationRequestRejection::ToolSchemaInvalid(schema_error)),
-            GeneratedTokenResult::ToolSchemaInvalid(expected_message)
-        );
-    }
-
-    #[test]
-    fn reports_grammar_sampler_initialization_failure() {
-        let grammar_error = GrammarError::FfiStatus(FfiStatusError {
-            operation: "llama_sampler_init_grammar",
-            code: 1,
-        });
-        let expected_message = agent_message(&format!(
-            "failed to initialize grammar sampler: {grammar_error}"
-        ));
-
-        assert_eq!(
-            reported(GenerationRequestRejection::GrammarSamplerInitializationFailed(grammar_error)),
-            GeneratedTokenResult::GrammarInitializationFailed(expected_message)
-        );
-    }
-
-    #[test]
-    fn reports_a_grammar_the_client_malformed_as_grammar_syntax_error() {
-        let expected_message =
-            agent_message(&format!("invalid grammar: {}", GrammarError::RootNotFound));
-
-        assert_eq!(
-            reported(GenerationRequestRejection::for_gbnf_grammar_error(
-                GrammarError::RootNotFound
-            )),
-            GeneratedTokenResult::GrammarSyntaxError(expected_message)
-        );
-    }
-
-    #[test]
-    fn reports_a_grammar_the_agent_failed_to_initialize_as_grammar_initialization_failure() {
-        let expected_message = agent_message(&format!(
-            "failed to initialize grammar sampler: {}",
-            GrammarError::NotEnoughMemory
-        ));
-
-        assert_eq!(
-            reported(GenerationRequestRejection::for_gbnf_grammar_error(
-                GrammarError::NotEnoughMemory
-            )),
-            GeneratedTokenResult::GrammarInitializationFailed(expected_message)
-        );
-    }
-
-    #[test]
-    fn reports_tool_call_parser_creation_failure_as_chat_template_error() {
-        let parser_error = ParseChatMessageError::ToolsNotAnArray;
-        let expected_message = agent_message(&format!(
-            "failed to build this model's tool call parser: {parser_error}"
-        ));
-
-        assert_eq!(
-            reported(GenerationRequestRejection::ToolCallParserCreationFailed(
-                parser_error
-            )),
-            GeneratedTokenResult::ChatTemplateError(expected_message)
-        );
-    }
-
-    #[test]
-    fn reports_media_exceeding_the_micro_batch_with_its_token_counts() {
-        let rejection = GenerationRequestRejection::for_micro_batch_fit_error(
-            MtmdEvalError::NonCausalChunkExceedsMicroBatch(NonCausalChunkMicroBatchMismatch {
-                chunk_tokens: 256,
-                micro_batch_tokens: 128,
-            }),
-        );
-
-        assert_eq!(
-            reported(rejection),
-            GeneratedTokenResult::MediaExceedsMicroBatch(OversizedMediaDetails {
-                media_tokens: 256,
-                micro_batch_tokens: 128,
-            })
-        );
-    }
-
-    #[test]
-    fn reports_other_micro_batch_check_failures_as_micro_batch_check_failures() {
-        let fit_error = MtmdEvalError::FfiContract(FfiContractError {
-            operation: "mtmd_input_chunks_get",
-            detail: "returned a null chunk within the chunk count",
-        });
-        let expected_message = agent_message(&format!(
-            "failed to check media chunks against the micro batch: {fit_error}"
-        ));
-
-        assert_eq!(
-            reported(GenerationRequestRejection::for_micro_batch_fit_error(
-                fit_error
-            )),
-            GeneratedTokenResult::MediaMicroBatchCheckFailed(expected_message)
-        );
-    }
-
-    #[test]
-    fn reports_oversized_prompt_with_its_token_counts() {
-        assert_eq!(
-            reported(GenerationRequestRejection::PromptExceedsContextSize {
-                details: OversizedPromptDetails {
-                    prompt_tokens: 9895,
-                    sequence_context_size: 8192,
-                },
-            }),
-            GeneratedTokenResult::PromptExceedsContextSize(OversizedPromptDetails {
-                prompt_tokens: 9895,
-                sequence_context_size: 8192,
-            })
-        );
-    }
-
-    #[test]
-    fn reports_multimodal_ingestion_failures_as_ingestion_failures() {
-        let rejection = GenerationRequestRejection::MultimodalIngestionFailed(
-            EvalMultimodalChunksError::ChunkOutOfBounds(3),
-        );
-
-        assert_eq!(
-            reported(rejection),
-            GeneratedTokenResult::MultimodalIngestionFailed(agent_message(
-                "failed to ingest multimodal prompt: chunk index 3 out of bounds during post-eval walk"
-            ))
-        );
-    }
-
-    #[test]
-    fn reports_internal_scheduler_failures_under_their_own_results() {
-        for ExpectedReport {
-            rejection,
-            into_result,
-        } in [
-            ExpectedReport {
-                rejection: GenerationRequestRejection::BatchAssemblyFailed(
-                    BatchAddError::EmptyBuffer,
-                ),
-                into_result: GeneratedTokenResult::BatchAssemblyFailed,
-            },
-            ExpectedReport {
-                rejection: GenerationRequestRejection::KvCacheClearFailed(
-                    ClearKvCacheSeqError::PartialSequenceNotRemoved {
-                        seq_id: 1,
-                        p0: 0,
-                        p1: 4,
-                    },
-                ),
-                into_result: GeneratedTokenResult::KvCacheClearFailed,
-            },
-            ExpectedReport {
-                rejection: GenerationRequestRejection::NoSequenceSlotAvailable,
-                into_result: GeneratedTokenResult::NoSequenceSlotAvailable,
-            },
-            ExpectedReport {
-                rejection: GenerationRequestRejection::SamplerChainCreationFailed(
-                    SamplingError::SamplerUnavailable { sampler: "grammar" },
-                ),
-                into_result: GeneratedTokenResult::SamplerChainCreationFailed,
-            },
-            ExpectedReport {
-                rejection: GenerationRequestRejection::SequenceIdOutOfRange {
-                    sequence_id: -1,
-                    source: u32::try_from(-1_i32).unwrap_err(),
-                },
-                into_result: GeneratedTokenResult::SequenceIdOutOfRange,
-            },
-            ExpectedReport {
-                rejection: GenerationRequestRejection::ToolsSerializationFailed(
-                    from_str::<Value>("{").unwrap_err(),
-                ),
-                into_result: GeneratedTokenResult::ToolsSerializationFailed,
-            },
-        ] {
-            let expected_message = agent_message(&rejection.to_string());
-
-            assert_eq!(reported(rejection), into_result(expected_message));
-        }
-    }
-}
diff --git a/paddler_agent/src/image_input.rs b/paddler_agent/src/image_input.rs
deleted file mode 100644
index 3cc28a6cd..000000000
--- a/paddler_agent/src/image_input.rs
+++ /dev/null
@@ -1,53 +0,0 @@
-use paddler_messaging::image_url::ImageUrl;
-
-use crate::generation_request_rejection::GenerationRequestRejection;
-use crate::multimodal_prompt_support::MultimodalPromptSupport;
-use crate::prompt_modality::PromptModality;
-
-pub enum ImageInput {
-    Supported(MultimodalPromptSupport),
-    Unsupported,
-}
-
-impl ImageInput {
-    pub const fn prompt_modality_for(
-        &self,
-        image_urls: &[ImageUrl],
-    ) -> Result<PromptModality<'_>, GenerationRequestRejection> {
-        if image_urls.is_empty() {
-            return Ok(PromptModality::TextOnly);
-        }
-
-        match self {
-            Self::Supported(multimodal_prompt_support) => {
-                Ok(PromptModality::Multimodal(multimodal_prompt_support))
-            }
-            Self::Unsupported => Err(GenerationRequestRejection::MultimodalNotSupported),
-        }
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use std::mem::discriminant;
-
-    use paddler_messaging::image_url::ImageUrl;
-
-    use super::ImageInput;
-    use crate::generation_request_rejection::GenerationRequestRejection;
-
-    #[test]
-    fn rejects_images_when_no_multimodal_projection_is_loaded() {
-        assert_eq!(
-            ImageInput::Unsupported
-                .prompt_modality_for(&[ImageUrl {
-                    url: "data:image/png;base64,AAAA".to_owned(),
-                }])
-                .err()
-                .map(|rejection| discriminant(&rejection)),
-            Some(discriminant(
-                &GenerationRequestRejection::MultimodalNotSupported
-            ))
-        );
-    }
-}
diff --git a/paddler_agent/src/inference_capabilities.rs b/paddler_agent/src/inference_capabilities.rs
new file mode 100644
index 000000000..90bcd5bd1
--- /dev/null
+++ b/paddler_agent/src/inference_capabilities.rs
@@ -0,0 +1,32 @@
+use tokio::join;
+
+use paddler_agent_runtime::shuts_down::ShutsDown;
+
+#[derive(Clone, Debug, Eq, PartialEq)]
+pub struct InferenceCapabilities<TDecisions, TEmbeddings, TTextGeneration> {
+    pub decisions: TDecisions,
+    pub embeddings: TEmbeddings,
+    pub text_generation: TTextGeneration,
+}
+
+impl<TDecisions, TEmbeddings, TTextGeneration> ShutsDown
+    for InferenceCapabilities<TDecisions, TEmbeddings, TTextGeneration>
+where
+    TDecisions: ShutsDown + Send,
+    TEmbeddings: ShutsDown + Send,
+    TTextGeneration: ShutsDown + Send,
+{
+    async fn shut_down(self) {
+        let Self {
+            decisions,
+            embeddings,
+            text_generation,
+        } = self;
+
+        join!(
+            decisions.shut_down(),
+            embeddings.shut_down(),
+            text_generation.shut_down()
+        );
+    }
+}
diff --git a/paddler_agent/src/last_announced_agent_status.rs b/paddler_agent/src/last_announced_agent_status.rs
new file mode 100644
index 000000000..b4e7876e6
--- /dev/null
+++ b/paddler_agent/src/last_announced_agent_status.rs
@@ -0,0 +1,75 @@
+use paddler_messaging::agent_status::AgentStatus;
+use paddler_messaging::management_socket::balancer::message::Message as ManagementJsonRpcMessage;
+use paddler_messaging::management_socket::balancer::notification::Notification as ManagementJsonRpcNotification;
+use paddler_messaging::management_socket::balancer::notification_params::update_agent_status_params::UpdateAgentStatusParams;
+use paddler_messaging::slot_aggregated_status_snapshot::SlotAggregatedStatusSnapshot;
+
+use crate::management_connection_step::ManagementConnectionStep;
+
+pub struct LastAnnouncedAgentStatus {
+    agent_status: AgentStatus,
+}
+
+impl LastAnnouncedAgentStatus {
+    #[must_use]
+    pub const fn new(agent_status: AgentStatus) -> Self {
+        Self { agent_status }
+    }
+
+    pub fn announce_if_changed(
+        &mut self,
+        slot_aggregated_status_snapshot: SlotAggregatedStatusSnapshot,
+    ) -> ManagementConnectionStep {
+        if self.agent_status == slot_aggregated_status_snapshot.status {
+            return ManagementConnectionStep::Continue;
+        }
+
+        self.agent_status = slot_aggregated_status_snapshot.status.clone();
+
+        ManagementConnectionStep::Write(ManagementJsonRpcMessage::Notification(
+            ManagementJsonRpcNotification::UpdateAgentStatus(UpdateAgentStatusParams {
+                slot_aggregated_status_snapshot,
+            }),
+        ))
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::mem::discriminant;
+
+    use paddler_messaging::agent_status::AgentStatus;
+    use paddler_messaging::management_socket::balancer::message::Message as ManagementJsonRpcMessage;
+    use paddler_messaging::management_socket::balancer::notification::Notification as ManagementJsonRpcNotification;
+    use paddler_messaging::management_socket::balancer::notification_params::update_agent_status_params::UpdateAgentStatusParams;
+    use paddler_messaging::slot_aggregated_status_snapshot::SlotAggregatedStatusSnapshot;
+
+    use super::LastAnnouncedAgentStatus;
+    use crate::management_connection_step::ManagementConnectionStep;
+
+    #[test]
+    fn a_changed_status_is_announced_once() {
+        let mut last_announced_agent_status = LastAnnouncedAgentStatus::new(AgentStatus::default());
+        let changed_snapshot = SlotAggregatedStatusSnapshot {
+            status: AgentStatus {
+                uses_chat_template_override: true,
+                ..AgentStatus::default()
+            },
+            version: 1,
+        };
+
+        assert!(matches!(
+            last_announced_agent_status.announce_if_changed(changed_snapshot.clone()),
+            ManagementConnectionStep::Write(announced_message)
+                if announced_message == ManagementJsonRpcMessage::Notification(
+                    ManagementJsonRpcNotification::UpdateAgentStatus(UpdateAgentStatusParams {
+                        slot_aggregated_status_snapshot: changed_snapshot.clone(),
+                    })
+                )
+        ));
+        assert_eq!(
+            discriminant(&last_announced_agent_status.announce_if_changed(changed_snapshot)),
+            discriminant(&ManagementConnectionStep::Continue)
+        );
+    }
+}
diff --git a/paddler_agent/src/lib.rs b/paddler_agent/src/lib.rs
index b447c0d35..8c42e31dc 100644
--- a/paddler_agent/src/lib.rs
+++ b/paddler_agent/src/lib.rs
@@ -1,83 +1,51 @@
-pub mod agent_applicable_model;
 pub mod agent_applicable_state;
 pub mod agent_applicable_state_holder;
 pub mod agent_desired_state_conversion;
 pub mod agent_desired_state_converter;
-pub mod agent_kv_cache_dtype;
-pub mod agent_pooling_type;
-pub mod agent_request;
+pub mod agent_error;
+pub mod agent_response_message;
+pub mod applicable_decisions;
 pub mod balancer_message_context;
-pub mod chat_prompt_render_request;
-pub mod chat_prompt_renderer;
-pub mod chat_template_load_status;
-pub mod chat_template_renderer;
-pub mod continuous_batch_active_request;
-pub mod continuous_batch_arbiter;
-pub mod continuous_batch_arbiter_context;
-pub mod continuous_batch_arbiter_handle;
-pub mod continuous_batch_arbiter_spawn_outcome;
-pub mod continuous_batch_arbiter_state;
-pub mod continuous_batch_embedding_processor;
-pub mod continuous_batch_generation_step;
-pub mod continuous_batch_preparation_request;
-pub mod continuous_batch_request_phase;
-pub mod continuous_batch_request_preparer;
-pub mod continuous_batch_request_state;
-pub mod continuous_batch_scheduler;
-pub mod continuous_batch_scheduler_command;
-pub mod continuous_batch_scheduler_context;
-pub mod continuous_batch_scheduler_params;
-pub mod continuous_batch_terminal_delivery;
-pub mod continuous_batch_terminal_outcome;
-pub mod converts_to_llama_kv_cache_dtype;
-pub mod converts_to_llama_pooling_type;
-pub mod converts_to_mtmd_bitmap;
+pub mod capability_plan;
+pub mod capability_starter;
+pub mod capability_startup_interruption;
+pub mod desired_model_file;
 pub mod desired_state_reconciler;
 pub mod desired_state_reconciliation;
-pub mod embedding_batch_preparer;
-pub mod embedding_batch_rejection;
-pub mod embedding_input_tokenized;
-pub mod forward_management_socket_messages;
-pub mod forward_scheduler_command;
-pub mod generation_request_preparer;
-pub mod generation_request_rejection;
-pub mod grammar_sampling;
-pub mod image_input;
-pub mod join_scheduler_thread;
+pub mod desired_state_reconciliation_attempt;
+pub mod enqueue_management_message;
+pub mod enqueue_outbound_frame;
+pub mod finished_serving_startup;
+pub mod inference_capabilities;
+pub mod last_announced_agent_status;
 pub mod llamacpp_arbiter_service;
+pub mod loaded_model_serving_plan;
+pub mod management_connection_end;
+pub mod management_connection_step;
 pub mod management_socket_client_service;
-pub mod model_bos_token;
-pub mod model_metadata_holder;
-pub mod multimodal_ingestion_progress;
-pub mod multimodal_prompt_ingestion;
-pub mod multimodal_prompt_support;
-pub mod normalization;
-pub mod plan_embedding_batches;
-pub mod prepared_embedding_batch_request;
-pub mod prepared_generation_request;
-pub mod prepared_prompt;
-pub mod prompt_modality;
-pub mod prompt_tokenizer;
-pub mod receives_stop_request;
+pub mod model_serving;
+pub mod model_serving_settings;
+pub mod pipeline_arbiter;
+pub mod pipeline_arbiter_params;
+pub mod pipeline_request;
+pub mod pipeline_response_stream;
+pub mod preparation_outcome;
+pub mod prepared_decisions;
+pub mod prepared_serving;
 pub mod reconciliation_service;
-pub mod rejection_description;
-pub mod require_embeddings_enabled;
-pub mod require_grammar_compatible_with_thinking;
-pub mod require_prompt_fits_sequence_context;
-pub mod resolve_grammar;
-pub mod sampler_chain_factory;
-pub mod sampling_outcome;
-pub mod send_result_or_warn;
-pub mod send_startup_signal;
-pub mod sequence_id_guard;
-pub mod sequence_id_pool;
-pub mod slot_guard;
-pub mod startup_signal_delivery;
-pub mod token_classification;
-pub mod token_generation;
-pub mod token_generation_support;
-pub mod token_sampling;
-pub mod tool_call_buffer;
-pub mod tool_call_event;
-pub mod tool_call_handling;
-pub mod tool_call_pipeline;
+pub mod replacement_outcome;
+pub mod retained_serving;
+pub mod retry_delay;
+pub mod retry_delay_outcome;
+pub mod served_capabilities;
+pub mod served_decisions;
+pub mod served_embeddings;
+pub mod served_model;
+pub mod served_text_generation;
+pub mod serving_interruption;
+pub mod serving_preparation;
+pub mod serving_replacement;
+pub mod serving_startup;
+pub mod serving_startup_progress;
+pub mod startup_phase;
+pub mod superseding_state;
diff --git a/paddler_agent/src/llamacpp_arbiter_service.rs b/paddler_agent/src/llamacpp_arbiter_service.rs
index a03c47843..13fffd018 100644
--- a/paddler_agent/src/llamacpp_arbiter_service.rs
+++ b/paddler_agent/src/llamacpp_arbiter_service.rs
@@ -2,101 +2,71 @@ use std::sync::Arc;
 
 use anyhow::Result;
 use async_trait::async_trait;
-use log::error;
 use tokio::select;
-use tokio::sync::mpsc;
-use tokio::time::Duration;
-use tokio::time::Instant;
-use tokio::time::MissedTickBehavior;
-use tokio::time::interval_at;
 use tokio_util::sync::CancellationToken;
 use trzcina::Service;
 
-use paddler_messaging::agent_state_application_status::AgentStateApplicationStatus;
+use paddler_agent_runtime::inference_runtime_context::InferenceRuntimeContext;
+use paddler_outbox::outbox::Outbox;
 
 use crate::agent_applicable_state_holder::AgentApplicableStateHolder;
-use crate::continuous_batch_arbiter_context::ContinuousBatchArbiterContext;
-use crate::continuous_batch_arbiter_state::ContinuousBatchArbiterState;
-use crate::continuous_batch_preparation_request::ContinuousBatchPreparationRequest;
-
-const STATE_APPLICATION_RETRY_INTERVAL: Duration = Duration::from_secs(1);
+use crate::pipeline_arbiter::PipelineArbiter;
+use crate::pipeline_arbiter_params::PipelineArbiterParams;
+use crate::pipeline_request::PipelineRequest;
 
 pub struct LlamaCppArbiterService {
     pub agent_applicable_state_holder: Arc<AgentApplicableStateHolder>,
-    pub arbiter_context: ContinuousBatchArbiterContext,
-    pub continuous_batch_preparation_request_rx:
-        mpsc::UnboundedReceiver<ContinuousBatchPreparationRequest>,
+    pub inference_runtime_context: InferenceRuntimeContext,
+    pub pipeline_requests: Arc<Outbox<PipelineRequest>>,
+    pub served_applicable_state_holder: Arc<AgentApplicableStateHolder>,
 }
 
-#[async_trait]
-impl Service for LlamaCppArbiterService {
-    fn name(&self) -> &'static str {
-        "agent::llamacpp_arbiter_service"
-    }
-
-    async fn run(self: Box<Self>, shutdown: CancellationToken) -> Result<()> {
+impl LlamaCppArbiterService {
+    async fn arbitrate_pipelines_until_shutdown(
+        self: Box<Self>,
+        shutdown: CancellationToken,
+    ) -> Result<()> {
         let Self {
             agent_applicable_state_holder,
-            arbiter_context,
-            mut continuous_batch_preparation_request_rx,
+            inference_runtime_context,
+            pipeline_requests,
+            served_applicable_state_holder,
         } = *self;
-
-        let mut arbiter_state = ContinuousBatchArbiterState::Idle;
-        let mut reconciled_state = agent_applicable_state_holder.subscribe();
-        let mut ticker = interval_at(
-            Instant::now() + STATE_APPLICATION_RETRY_INTERVAL,
-            STATE_APPLICATION_RETRY_INTERVAL,
-        );
-
-        ticker.set_missed_tick_behavior(MissedTickBehavior::Delay);
-
-        let shutdown_outcome = loop {
+        let mut agent_applicable_state_rx = agent_applicable_state_holder.subscribe();
+        let mut pipeline_arbiter = PipelineArbiter::new(PipelineArbiterParams {
+            agent_shutdown: shutdown.clone(),
+            inference_runtime_context,
+            served_applicable_state_holder,
+        });
+
+        let service_outcome = loop {
             select! {
                 biased;
                 () = shutdown.cancelled() => break Ok(()),
-                _ = ticker.tick() => {
-                    let slot_aggregated_status = &arbiter_context.slot_aggregated_status;
-                    let current_status = slot_aggregated_status.get_state_application_status();
-
-                    if current_status.should_try_to_apply() {
-                        slot_aggregated_status.set_state_application_status(
-                            if matches!(current_status, AgentStateApplicationStatus::AttemptedAndRetrying) {
-                                AgentStateApplicationStatus::Stuck
-                            } else {
-                                AgentStateApplicationStatus::AttemptedAndRetrying
-                            }
-                        );
-
-                        let agent_applicable_state = reconciled_state.borrow().clone();
-
-                        arbiter_state = arbiter_state
-                            .apply(&shutdown, agent_applicable_state, &arbiter_context)
-                            .await;
-                    }
+                Ok(()) = agent_applicable_state_rx.changed() => {
+                    pipeline_arbiter.apply(agent_applicable_state_rx.borrow_and_update().clone());
                 }
-                Ok(()) = reconciled_state.changed() => {
-                    let agent_applicable_state = reconciled_state.borrow_and_update().clone();
-
-                    arbiter_context
-                        .slot_aggregated_status
-                        .set_state_application_status(AgentStateApplicationStatus::Fresh);
-
-                    arbiter_state = arbiter_state
-                        .apply(&shutdown, agent_applicable_state, &arbiter_context)
-                        .await;
+                finished_pipeline_startup = pipeline_arbiter.next_finished_startup() => {
+                    pipeline_arbiter.conclude(finished_pipeline_startup);
                 }
-                request = continuous_batch_preparation_request_rx.recv() => match request {
-                    Some(request) => arbiter_state.forward(arbiter_context.agent_name.as_deref(), request),
-                    None => break Ok(()),
-                },
+                pipeline_request = pipeline_requests.next() => pipeline_arbiter.forward(pipeline_request),
             }
         };
 
-        if let Err(shutdown_error) = arbiter_state.shut_down().await {
-            error!("Failed to shut down arbiter cleanly: {shutdown_error:#}");
-        }
+        pipeline_arbiter.shut_down().await;
+
+        service_outcome
+    }
+}
+
+#[async_trait]
+impl Service for LlamaCppArbiterService {
+    fn name(&self) -> &'static str {
+        "agent::llamacpp_arbiter_service"
+    }
 
-        shutdown_outcome
+    async fn run(self: Box<Self>, shutdown: CancellationToken) -> Result<()> {
+        self.arbitrate_pipelines_until_shutdown(shutdown).await
     }
 }
 
@@ -104,36 +74,36 @@ impl Service for LlamaCppArbiterService {
 mod tests {
     use std::sync::Arc;
 
-    use tokio::sync::mpsc;
     use tokio_util::sync::CancellationToken;
     use trzcina::Service;
 
-    use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
+    use paddler_agent_runtime::inference_runtime_context::InferenceRuntimeContext;
+    use paddler_agent_runtime::inference_runtime_context_params::InferenceRuntimeContextParams;
+    use paddler_outbox::outbox::Outbox;
 
     use super::LlamaCppArbiterService;
     use crate::agent_applicable_state_holder::AgentApplicableStateHolder;
-    use crate::continuous_batch_arbiter_context::ContinuousBatchArbiterContext;
-    use crate::model_metadata_holder::ModelMetadataHolder;
 
     #[tokio::test(flavor = "multi_thread")]
-    async fn exits_when_its_request_channel_closes() {
-        let (continuous_batch_preparation_request_tx, continuous_batch_preparation_request_rx) =
-            mpsc::unbounded_channel();
+    async fn exits_when_the_agent_shuts_down() {
+        let agent_shutdown = CancellationToken::new();
         let service = LlamaCppArbiterService {
             agent_applicable_state_holder: Arc::new(AgentApplicableStateHolder::default()),
-            arbiter_context: ContinuousBatchArbiterContext {
-                agent_name: None,
-                model_metadata_holder: Arc::new(ModelMetadataHolder::default()),
-                slot_aggregated_status: Arc::new(SlotAggregatedStatus::new(1)),
-            },
-            continuous_batch_preparation_request_rx,
+            inference_runtime_context: InferenceRuntimeContext::new(
+                InferenceRuntimeContextParams {
+                    agent_name: None,
+                    slots: 1,
+                },
+            ),
+            pipeline_requests: Arc::new(Outbox::default()),
+            served_applicable_state_holder: Arc::new(AgentApplicableStateHolder::default()),
         };
 
-        drop(continuous_batch_preparation_request_tx);
+        agent_shutdown.cancel();
 
         Box::new(service)
-            .run(CancellationToken::new())
+            .run(agent_shutdown)
             .await
-            .expect("the service must finish cleanly once no request can arrive");
+            .expect("the service must finish cleanly once the agent shuts down");
     }
 }
diff --git a/paddler_agent/src/loaded_model_serving_plan.rs b/paddler_agent/src/loaded_model_serving_plan.rs
new file mode 100644
index 000000000..e05381336
--- /dev/null
+++ b/paddler_agent/src/loaded_model_serving_plan.rs
@@ -0,0 +1,36 @@
+use paddler_agent_runtime::loaded_llama_model::LoadedLlamaModel;
+
+use crate::capability_plan::CapabilityPlan;
+use crate::inference_capabilities::InferenceCapabilities;
+use crate::model_serving_settings::ModelServingSettings;
+use crate::served_decisions::ServedDecisions;
+use crate::served_embeddings::ServedEmbeddings;
+use crate::served_text_generation::ServedTextGeneration;
+
+pub struct LoadedModelServingPlan {
+    pub capability_plans: InferenceCapabilities<
+        CapabilityPlan<ServedDecisions>,
+        CapabilityPlan<ServedEmbeddings>,
+        CapabilityPlan<ServedTextGeneration>,
+    >,
+    pub loaded_llama_model: LoadedLlamaModel,
+    pub model_serving_settings: ModelServingSettings,
+}
+
+impl LoadedModelServingPlan {
+    #[must_use]
+    pub const fn starting_every_capability(
+        loaded_llama_model: LoadedLlamaModel,
+        model_serving_settings: ModelServingSettings,
+    ) -> Self {
+        Self {
+            capability_plans: InferenceCapabilities {
+                decisions: CapabilityPlan::Start,
+                embeddings: CapabilityPlan::Start,
+                text_generation: CapabilityPlan::Start,
+            },
+            loaded_llama_model,
+            model_serving_settings,
+        }
+    }
+}
diff --git a/paddler_agent/src/management_connection_end.rs b/paddler_agent/src/management_connection_end.rs
new file mode 100644
index 000000000..141439d19
--- /dev/null
+++ b/paddler_agent/src/management_connection_end.rs
@@ -0,0 +1,89 @@
+use tokio_tungstenite::tungstenite::Utf8Bytes;
+use tokio_tungstenite::tungstenite::protocol::CloseFrame;
+use tokio_tungstenite::tungstenite::protocol::frame::coding::CloseCode;
+
+const fn described_close_frame(code: CloseCode, reason: &'static str) -> CloseFrame {
+    CloseFrame {
+        code,
+        reason: Utf8Bytes::from_static(reason),
+    }
+}
+
+#[derive(Debug, Eq, PartialEq)]
+pub enum ManagementConnectionEnd {
+    AgentShuttingDown,
+    BalancerClosedConnection,
+    BalancerMessageUndeserializable,
+    BalancerReusedInFlightRequestId,
+    BalancerSentUnsupportedFrame,
+    BalancerStoppedAnsweringPings,
+    IncomingStreamFailed,
+    OutgoingMessageUnserializable,
+    OutgoingStreamFailed,
+}
+
+impl ManagementConnectionEnd {
+    #[must_use]
+    pub const fn close_frame(&self) -> Option<CloseFrame> {
+        match self {
+            Self::AgentShuttingDown => Some(described_close_frame(
+                CloseCode::Away,
+                "Agent shutting down",
+            )),
+            Self::BalancerClosedConnection
+            | Self::BalancerStoppedAnsweringPings
+            | Self::IncomingStreamFailed
+            | Self::OutgoingStreamFailed => None,
+            Self::BalancerMessageUndeserializable => Some(described_close_frame(
+                CloseCode::Invalid,
+                "Balancer message could not be deserialized",
+            )),
+            Self::BalancerReusedInFlightRequestId => Some(described_close_frame(
+                CloseCode::Policy,
+                "Balancer reused the id of a request in flight",
+            )),
+            Self::BalancerSentUnsupportedFrame => Some(described_close_frame(
+                CloseCode::Unsupported,
+                "Balancer sent a frame that is not text",
+            )),
+            Self::OutgoingMessageUnserializable => Some(described_close_frame(
+                CloseCode::Error,
+                "Agent message could not be serialized",
+            )),
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use tokio_tungstenite::tungstenite::protocol::frame::coding::CloseCode;
+
+    use super::ManagementConnectionEnd;
+
+    #[test]
+    fn an_abandoned_connection_tells_the_balancer_why_it_closes() {
+        let close_codes: Vec<Option<CloseCode>> = [
+            ManagementConnectionEnd::BalancerMessageUndeserializable,
+            ManagementConnectionEnd::BalancerReusedInFlightRequestId,
+            ManagementConnectionEnd::BalancerSentUnsupportedFrame,
+            ManagementConnectionEnd::OutgoingMessageUnserializable,
+        ]
+        .iter()
+        .map(|connection_end| {
+            connection_end
+                .close_frame()
+                .map(|close_frame| close_frame.code)
+        })
+        .collect();
+
+        assert_eq!(
+            close_codes,
+            vec![
+                Some(CloseCode::Invalid),
+                Some(CloseCode::Policy),
+                Some(CloseCode::Unsupported),
+                Some(CloseCode::Error),
+            ]
+        );
+    }
+}
diff --git a/paddler_agent/src/management_connection_step.rs b/paddler_agent/src/management_connection_step.rs
new file mode 100644
index 000000000..bd30b14b4
--- /dev/null
+++ b/paddler_agent/src/management_connection_step.rs
@@ -0,0 +1,13 @@
+use futures_util::stream::BoxStream;
+
+use paddler_messaging::management_socket::balancer::message::Message as ManagementJsonRpcMessage;
+
+use crate::management_connection_end::ManagementConnectionEnd;
+
+pub enum ManagementConnectionStep {
+    Continue,
+    Deregister,
+    End(ManagementConnectionEnd),
+    StreamResponses(BoxStream<'static, ManagementJsonRpcMessage>),
+    Write(ManagementJsonRpcMessage),
+}
diff --git a/paddler_agent/src/management_socket_client_service.rs b/paddler_agent/src/management_socket_client_service.rs
index f72775ec0..3996bdb28 100644
--- a/paddler_agent/src/management_socket_client_service.rs
+++ b/paddler_agent/src/management_socket_client_service.rs
@@ -1,415 +1,210 @@
-use anyhow::Context;
 use anyhow::Result;
 use async_trait::async_trait;
-use futures_util::SinkExt as _;
-use futures_util::StreamExt;
-use log::debug;
+use futures_util::StreamExt as _;
+use futures_util::stream::BoxStream;
+use futures_util::stream::SelectAll;
+use futures_util::stream::SplitStream;
 use log::error;
 use log::info;
-use log::warn;
-use serde_json::from_str;
+use tokio::join;
 use tokio::net::TcpStream;
 use tokio::select;
-use tokio::spawn;
 use tokio::sync::mpsc;
-use tokio::sync::mpsc::error::SendError;
-use tokio::sync::watch;
 use tokio::time::Duration;
+use tokio::time::Instant;
 use tokio::time::MissedTickBehavior;
 use tokio::time::interval;
+use tokio::time::interval_at;
 use tokio_tungstenite::MaybeTlsStream;
 use tokio_tungstenite::WebSocketStream;
 use tokio_tungstenite::connect_async_with_config;
+use tokio_tungstenite::tungstenite::Bytes;
 use tokio_tungstenite::tungstenite::protocol::Message;
 use tokio_util::sync::CancellationToken;
 use trzcina::Service;
 
-use paddler_messaging::balancer_connection::BalancerConnection;
-use paddler_messaging::embedding_result::EmbeddingResult;
-use paddler_messaging::generated_token_result::GeneratedTokenResult;
-use paddler_messaging::jsonrpc::request_envelope::RequestEnvelope;
-use paddler_messaging::jsonrpc::response_envelope::ResponseEnvelope;
-use paddler_messaging::management_socket::agent::message::Message as JsonRpcMessage;
-use paddler_messaging::management_socket::agent::notification::Notification as JsonRpcNotification;
-use paddler_messaging::management_socket::agent::notification_params::version_params::VersionParams;
-use paddler_messaging::management_socket::agent::request::Request as JsonRpcRequest;
-use paddler_messaging::management_socket::agent::response::Response as JsonRpcResponse;
+use paddler_frame_writer::frame_writing_end::FrameWritingEnd;
+use paddler_frame_writer::write_frames_until_connection_closes::write_frames_until_connection_closes;
 use paddler_messaging::management_socket::balancer::message::Message as ManagementJsonRpcMessage;
 use paddler_messaging::management_socket::balancer::notification::Notification as ManagementJsonRpcNotification;
 use paddler_messaging::management_socket::balancer::notification_params::register_agent_params::RegisterAgentParams;
-use paddler_messaging::management_socket::balancer::notification_params::update_agent_status_params::UpdateAgentStatusParams;
-use paddler_messaging::produces_snapshot::ProducesSnapshot;
+use paddler_messaging::ping_interval::PING_INTERVAL;
+use paddler_messaging::ping_liveness::PingLiveness;
+use paddler_messaging::ping_turn::PingTurn;
+use paddler_messaging::produces_snapshot::ProducesSnapshot as _;
 use paddler_messaging::subscribes_to_updates::SubscribesToUpdates as _;
-use paddler_request_registry::request_delivery::RequestDelivery;
-use paddler_request_registry::request_registry_guard::RequestRegistryGuard;
 
-use crate::agent_request::AgentRequest;
 use crate::balancer_message_context::BalancerMessageContext;
-use crate::continuous_batch_preparation_request::ContinuousBatchPreparationRequest;
-use crate::forward_management_socket_messages::forward_management_socket_messages;
-use crate::slot_guard::SlotGuard;
+use crate::enqueue_management_message::enqueue_management_message;
+use crate::enqueue_outbound_frame::enqueue_outbound_frame;
+use crate::last_announced_agent_status::LastAnnouncedAgentStatus;
+use crate::management_connection_end::ManagementConnectionEnd;
+use crate::management_connection_step::ManagementConnectionStep;
 
 const BALANCER_RECONNECT_INTERVAL: Duration = Duration::from_secs(1);
 
-struct IncomingMessageContext {
-    balancer_message_context: BalancerMessageContext,
-    connection_close: CancellationToken,
-    message_tx: mpsc::UnboundedSender<ManagementJsonRpcMessage>,
-}
-
-impl IncomingMessageContext {
-    fn generate_responses<TParams, TResponse>(self, id: String, params: TParams) -> Result<()>
-    where
-        AgentRequest<TParams, TResponse>: Into<ContinuousBatchPreparationRequest>,
-        TResponse: Into<JsonRpcResponse> + Send + 'static,
-    {
-        let Self {
-            balancer_message_context:
-                BalancerMessageContext {
-                    continuous_batch_preparation_request_tx,
-                    request_stoppers,
-                    slot_aggregated_status,
-                    ..
-                },
-            connection_close,
-            message_tx,
-        } = self;
-        let (response_tx, mut response_rx) = mpsc::unbounded_channel::<TResponse>();
-        let (stop_tx, stop_rx) = mpsc::unbounded_channel::<()>();
-
-        let stopper_guard = RequestRegistryGuard::register(&request_stoppers, id.clone(), stop_tx)
-            .context(format!("Failed to register stopper for request: {id}"))?;
-
-        continuous_batch_preparation_request_tx.send(
-            AgentRequest {
-                params,
-                response_tx,
-                slot_guard: SlotGuard::new(slot_aggregated_status),
-                stop_rx,
-            }
-            .into(),
-        )?;
-
-        spawn(async move {
-            let _stopper_guard = stopper_guard;
-
-            loop {
-                select! {
-                    () = connection_close.cancelled() => break,
-                    response = response_rx.recv() => {
-                        match response {
-                            Some(response) => {
-                                if let Err(err) = message_tx.send(
-                                    ManagementJsonRpcMessage::Response(
-                                        ResponseEnvelope {
-                                            generated_by: None,
-                                            request_id: id.clone(),
-                                            response: response.into(),
-                                        }
-                                    ),
-                                ) {
-                                    error!("Failed to forward response for request {id:?}: {err}");
-
-                                    break;
-                                }
-                            }
-                            None => break,
-                        }
-                    }
-                }
-            }
-        });
-
-        Ok(())
-    }
-
-    fn handle_deserialized_message(self, deserialized_message: JsonRpcMessage) -> Result<()> {
-        match deserialized_message {
-            JsonRpcMessage::Notification(JsonRpcNotification::SetState(set_state_params)) => {
-                self.balancer_message_context
-                    .agent_desired_state_tx
-                    .send(set_state_params.desired_state)?;
-
-                Ok(())
-            }
-            JsonRpcMessage::Notification(JsonRpcNotification::StopRespondingTo(request_id)) => {
-                debug!("Received StopGeneratingTokens notification for request ID: {request_id:?}");
-                match self
-                    .balancer_message_context
-                    .request_stoppers
-                    .send_to(&request_id, ())
-                {
-                    RequestDelivery::Delivered => Ok(()),
-                    RequestDelivery::ReceiverDropped | RequestDelivery::RequestNotRegistered => {
-                        debug!("Request {request_id:?} finished before its stop arrived");
-
-                        Ok(())
-                    }
-                }
-            }
-            JsonRpcMessage::Notification(JsonRpcNotification::Version(VersionParams {
-                version,
-            })) => {
-                if version != env!("CARGO_PKG_VERSION") {
-                    warn!(
-                        "Version mismatch: server version is {version}, client version is {}",
-                        env!("CARGO_PKG_VERSION")
-                    );
-                }
-
-                Ok(())
-            }
-            JsonRpcMessage::Request(RequestEnvelope {
-                id,
-                request:
-                    JsonRpcRequest::ContinueFromConversationHistory(
-                        continue_from_conversation_history_params,
-                    ),
-            }) => self.generate_responses::<_, GeneratedTokenResult>(
-                id,
-                continue_from_conversation_history_params,
-            ),
-            JsonRpcMessage::Request(RequestEnvelope {
-                id,
-                request: JsonRpcRequest::ContinueFromRawPrompt(generate_tokens_params),
-            }) => self.generate_responses::<_, GeneratedTokenResult>(id, generate_tokens_params),
-            JsonRpcMessage::Request(RequestEnvelope {
-                id,
-                request: JsonRpcRequest::GenerateEmbeddingBatch(generate_embedding_batch_params),
-            }) => {
-                self.generate_responses::<_, EmbeddingResult>(id, generate_embedding_batch_params)
-            }
-            JsonRpcMessage::Request(RequestEnvelope {
-                id,
-                request: JsonRpcRequest::GetChatTemplateOverride,
-            }) => {
-                Ok(self
-                    .message_tx
-                    .send(ManagementJsonRpcMessage::Response(ResponseEnvelope {
-                        generated_by: None,
-                        request_id: id,
-                        response: JsonRpcResponse::ChatTemplateOverride(
-                            self.balancer_message_context
-                                .agent_applicable_state_holder
-                                .get_agent_applicable_state()
-                                .chat_template_override,
-                        ),
-                    }))?)
-            }
-            JsonRpcMessage::Request(RequestEnvelope {
-                id,
-                request: JsonRpcRequest::GetModelMetadata,
-            }) => {
-                Ok(self
-                    .message_tx
-                    .send(ManagementJsonRpcMessage::Response(ResponseEnvelope {
-                        generated_by: None,
-                        request_id: id,
-                        response: JsonRpcResponse::ModelMetadata(
-                            self.balancer_message_context
-                                .model_metadata_holder
-                                .get_model_metadata(),
-                        ),
-                    }))?)
-            }
-        }
-    }
-
-    fn handle_incoming_message(self, msg: Message) {
-        match msg {
-            Message::Text(text) => match from_str::<JsonRpcMessage>(&text) {
-                Ok(deserialized_message) => {
-                    if let Err(err) = self.handle_deserialized_message(deserialized_message) {
-                        error!("Error handling incoming message: {err}");
-                    }
-                }
-                Err(deserialization_error) => {
-                    error!(
-                        "Closing the management connection after an undeserializable message {text}: {deserialization_error}"
-                    );
-
-                    self.connection_close.cancel();
-                }
-            },
-            Message::Binary(_) | Message::Frame(_) => {
-                error!("Closing the management connection after a non-text message");
-
-                self.connection_close.cancel();
-            }
-            Message::Close(_) => {
-                info!("Connection closed by server");
-            }
-            Message::Ping(_) | Message::Pong(_) => {}
-        }
-    }
-}
-
 pub struct ManagementSocketClientService {
-    pub balancer_connection_tx: watch::Sender<BalancerConnection>,
     pub balancer_message_context: BalancerMessageContext,
     pub name: Option<String>,
     pub socket_url: String,
 }
 
 impl ManagementSocketClientService {
-    async fn keep_connection_alive(&self, shutdown: CancellationToken) -> Result<()> {
-        match shutdown
-            .run_until_cancelled(
-                self.balancer_message_context
-                    .slot_aggregated_status
-                    .wait_until_no_slots_are_processing(),
-            )
-            .await
-        {
-            Some(idle_outcome) => idle_outcome
-                .context("Slot status stopped announcing updates while waiting for idle slots")?,
-            None => return Ok(()),
-        }
+    async fn connect_once_slots_are_idle(
+        &self,
+    ) -> Result<WebSocketStream<MaybeTlsStream<TcpStream>>> {
+        self.balancer_message_context
+            .slot_aggregated_status
+            .wait_until_no_slots_are_processing()
+            .await;
 
         info!("Connecting to management server at {}", self.socket_url);
 
-        let (ws_stream, _response) = match shutdown
-            .run_until_cancelled(connect_async_with_config(
-                self.socket_url.clone(),
-                None,
-                true,
-            ))
-            .await
-        {
-            Some(connect_outcome) => connect_outcome?,
-            None => return Ok(()),
-        };
+        let (ws_stream, _response) =
+            connect_async_with_config(self.socket_url.clone(), None, true).await?;
 
         info!("Connected to management server");
 
-        self.balancer_connection_tx
-            .send_replace(BalancerConnection::Connected);
-
-        let connection_outcome = self.serve_connection(ws_stream, shutdown).await;
+        Ok(ws_stream)
+    }
 
-        self.balancer_connection_tx
-            .send_replace(BalancerConnection::Connecting);
+    async fn keep_connection_alive(&self, shutdown: CancellationToken) -> Result<()> {
+        if let Some(connect_outcome) = shutdown
+            .run_until_cancelled(self.connect_once_slots_are_idle())
+            .await
+        {
+            self.serve_connection(connect_outcome?, shutdown).await;
+        }
 
-        connection_outcome
+        Ok(())
     }
 
     async fn serve_connection(
         &self,
         ws_stream: WebSocketStream<MaybeTlsStream<TcpStream>>,
         shutdown: CancellationToken,
-    ) -> Result<()> {
+    ) {
+        let (ws_sink, ws_source) = ws_stream.split();
+        let (outbound_tx, outbound_rx) = mpsc::unbounded_channel();
         let connection_close = CancellationToken::new();
-        let (message_tx, message_rx) = mpsc::unbounded_channel::<ManagementJsonRpcMessage>();
-        let (write, mut read) = ws_stream.split();
-
-        let message_forward_handle = spawn(forward_management_socket_messages(
-            connection_close.clone(),
-            message_rx,
-            write,
-        ));
 
-        let mut update_rx = self
-            .balancer_message_context
-            .slot_aggregated_status
-            .subscribe_to_updates();
-
-        message_tx
-            .send(ManagementJsonRpcMessage::Notification(
-                ManagementJsonRpcNotification::RegisterAgent(RegisterAgentParams {
-                    name: self.name.clone(),
-                    slot_aggregated_status_snapshot: self
-                        .balancer_message_context
-                        .slot_aggregated_status
-                        .make_snapshot(),
-                }),
-            ))
-            .context("The management socket writer stopped before the agent registered")?;
+        let ((), frame_writing_end) = join!(
+            self.read_until_the_connection_ends(
+                connection_close.clone(),
+                ws_source,
+                outbound_tx,
+                shutdown
+            ),
+            write_frames_until_connection_closes(connection_close, outbound_rx, ws_sink),
+        );
 
-        let agent_ends_connection = loop {
-            select! {
-                () = connection_close.cancelled() => {
-                    info!("Connection close signal received, shutting down");
+        match frame_writing_end {
+            FrameWritingEnd::ConnectionClosed => {
+                info!("Abandoned the management socket frames the balancer did not read");
+            }
+            FrameWritingEnd::EveryFrameSent => {}
+            FrameWritingEnd::SendFailed(send_error) => {
+                error!("Failed to send a management socket frame: {send_error}");
+            }
+        }
+    }
 
-                    break false;
-                }
-                () = shutdown.cancelled() => break true,
-                Ok(()) = update_rx.changed() => {
-                    let status_update = ManagementJsonRpcMessage::Notification(
-                        ManagementJsonRpcNotification::UpdateAgentStatus(UpdateAgentStatusParams {
-                            slot_aggregated_status_snapshot: self
-                                .balancer_message_context
-                                .slot_aggregated_status
-                                .make_snapshot(),
-                        }),
-                    );
+    async fn read_until_the_connection_ends(
+        &self,
+        connection_close: CancellationToken,
+        mut ws_source: SplitStream<WebSocketStream<MaybeTlsStream<TcpStream>>>,
+        outbound_tx: mpsc::UnboundedSender<Message>,
+        shutdown: CancellationToken,
+    ) {
+        let slot_aggregated_status = &self.balancer_message_context.slot_aggregated_status;
+        let mut update_rx = slot_aggregated_status.subscribe_to_updates();
+        let registration_snapshot = slot_aggregated_status.make_snapshot();
+        let mut last_announced_agent_status =
+            LastAnnouncedAgentStatus::new(registration_snapshot.status.clone());
+        let mut pipeline_responses =
+            SelectAll::<BoxStream<'static, ManagementJsonRpcMessage>>::new();
+        let mut ping_ticker = interval_at(Instant::now() + PING_INTERVAL, PING_INTERVAL);
+        let mut ping_liveness = PingLiveness::default();
+        let mut step = ManagementConnectionStep::Write(ManagementJsonRpcMessage::Notification(
+            ManagementJsonRpcNotification::RegisterAgent(RegisterAgentParams {
+                name: self.name.clone(),
+                slot_aggregated_status_snapshot: registration_snapshot,
+            }),
+        ));
 
-                    if let Err(SendError(_undelivered_status_update)) = message_tx.send(status_update) {
-                        info!("The management socket writer stopped, reconnecting");
+        ping_ticker.set_missed_tick_behavior(MissedTickBehavior::Delay);
 
-                        break false;
-                    }
-                }
-                msg = read.next() => {
-                    let should_close = match msg {
-                        Some(Ok(msg)) => {
-                            IncomingMessageContext {
-                                balancer_message_context: self.balancer_message_context.clone(),
-                                connection_close: connection_close.clone(),
-                                message_tx: message_tx.clone(),
-                            }
-                            .handle_incoming_message(msg);
+        let connection_end = loop {
+            step = match step {
+                ManagementConnectionStep::Continue => select! {
+                    biased;
+                    () = shutdown.cancelled() => ManagementConnectionStep::Deregister,
+                    incoming_message = ws_source.next() => match incoming_message {
+                        Some(Ok(Message::Pong(_))) => {
+                            ping_liveness.record_pong();
 
-                            false
+                            ManagementConnectionStep::Continue
                         }
-                        Some(Err(err)) => {
-                            error!("Error reading message: {err}");
+                        Some(Ok(message)) => self.balancer_message_context.handle_message(message),
+                        Some(Err(read_error)) => {
+                            error!("Failed to read from the management socket: {read_error}");
 
-                            true
+                            ManagementConnectionStep::End(ManagementConnectionEnd::IncomingStreamFailed)
                         }
-                        None => true,
-                    };
-
-                    if should_close {
-                        connection_close.cancel();
-
-                        break false;
+                        None => ManagementConnectionStep::End(ManagementConnectionEnd::BalancerClosedConnection),
+                    },
+                    _ = ping_ticker.tick() => match ping_liveness.next_ping() {
+                        PingTurn::PeerStoppedAnswering => {
+                            ManagementConnectionStep::End(ManagementConnectionEnd::BalancerStoppedAnsweringPings)
+                        }
+                        PingTurn::SendPing => enqueue_outbound_frame(
+                            &outbound_tx,
+                            Message::Ping(Bytes::new()),
+                            ManagementConnectionStep::Continue,
+                        ),
+                    },
+                    Ok(()) = update_rx.changed() => last_announced_agent_status
+                        .announce_if_changed(slot_aggregated_status.make_snapshot()),
+                    Some(pipeline_response) = pipeline_responses.next() => {
+                        ManagementConnectionStep::Write(pipeline_response)
                     }
+                },
+                ManagementConnectionStep::Deregister => enqueue_management_message(
+                    &outbound_tx,
+                    &ManagementJsonRpcMessage::Notification(
+                        ManagementJsonRpcNotification::DeregisterAgent,
+                    ),
+                    ManagementConnectionStep::End(ManagementConnectionEnd::AgentShuttingDown),
+                ),
+                ManagementConnectionStep::End(connection_end) => break connection_end,
+                ManagementConnectionStep::StreamResponses(pipeline_response_stream) => {
+                    pipeline_responses.push(pipeline_response_stream);
+
+                    ManagementConnectionStep::Continue
                 }
-            }
+                ManagementConnectionStep::Write(message) => enqueue_management_message(
+                    &outbound_tx,
+                    &message,
+                    ManagementConnectionStep::Continue,
+                ),
+            };
         };
 
-        if let Err(SendError(_undelivered_deregistration)) = message_tx.send(
-            ManagementJsonRpcMessage::Notification(ManagementJsonRpcNotification::DeregisterAgent),
-        ) {
-            info!(
-                "The management socket writer already stopped; the balancer deregisters the agent once the connection drops"
-            );
-        }
+        let close_frame_delivery = outbound_tx.send(Message::Close(connection_end.close_frame()));
 
         connection_close.cancel();
 
-        let mut write = message_forward_handle
-            .await
-            .context("Failed to join message forwarding task")?;
-
-        if agent_ends_connection {
-            write
-                .send(Message::Close(None))
-                .await
-                .context("Failed to close the management socket")?;
-        }
-
-        Ok(())
+        info!(
+            "Closing the management connection after {connection_end:?}: {close_frame_delivery:?}"
+        );
     }
 }
 
-#[async_trait]
-impl Service for ManagementSocketClientService {
-    fn name(&self) -> &'static str {
-        "agent::management_socket_client_service"
-    }
-
-    async fn run(self: Box<Self>, shutdown: CancellationToken) -> Result<()> {
+impl ManagementSocketClientService {
+    async fn reconnect_to_the_balancer_until_shutdown(
+        self: Box<Self>,
+        shutdown: CancellationToken,
+    ) -> Result<()> {
         let mut ticker = interval(BALANCER_RECONNECT_INTERVAL);
 
         ticker.set_missed_tick_behavior(MissedTickBehavior::Delay);
@@ -432,605 +227,14 @@ impl Service for ManagementSocketClientService {
     }
 }
 
-#[cfg(test)]
-mod tests {
-    use std::collections::BTreeMap;
-    use std::num::NonZeroU32;
-    use std::sync::Arc;
-
-    use futures_util::StreamExt as _;
-    use serde_json::from_str;
-    use serde_json::to_string;
-    use tokio::net::TcpListener;
-    use tokio::spawn;
-    use tokio::sync::mpsc;
-    use tokio::sync::oneshot;
-    use tokio::sync::watch;
-    use tokio::task::JoinHandle;
-    use tokio_tungstenite::accept_async;
-    use tokio_tungstenite::tungstenite::Bytes;
-    use tokio_tungstenite::tungstenite::protocol::Message;
-    use tokio_util::sync::CancellationToken;
-    use trzcina::Service;
-
-    use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
-    use paddler_messaging::agent_desired_state::AgentDesiredState;
-    use paddler_messaging::agent_status::AgentStatus;
-    use paddler_messaging::api_path::ApiPath;
-    use paddler_messaging::balancer_connection::BalancerConnection;
-    use paddler_messaging::generated_token_result::GeneratedTokenResult;
-    use paddler_messaging::jsonrpc::request_envelope::RequestEnvelope;
-    use paddler_messaging::jsonrpc::response_envelope::ResponseEnvelope;
-    use paddler_messaging::management_socket::agent::message::Message as JsonRpcMessage;
-    use paddler_messaging::management_socket::agent::notification::Notification as JsonRpcNotification;
-    use paddler_messaging::management_socket::agent::notification_params::set_state_params::SetStateParams;
-    use paddler_messaging::management_socket::agent::notification_params::version_params::VersionParams;
-    use paddler_messaging::management_socket::agent::request::Request as JsonRpcRequest;
-    use paddler_messaging::management_socket::agent::response::Response as JsonRpcResponse;
-    use paddler_messaging::management_socket::balancer::message::Message as ManagementJsonRpcMessage;
-    use paddler_messaging::management_socket::balancer::notification::Notification as ManagementJsonRpcNotification;
-    use paddler_messaging::management_socket::balancer::notification_params::update_agent_status_params::UpdateAgentStatusParams;
-    use paddler_messaging::model_metadata::ModelMetadata;
-    use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
-    use paddler_messaging::slot_aggregated_status_snapshot::SlotAggregatedStatusSnapshot;
-    use paddler_request_registry::request_delivery::RequestDelivery;
-    use paddler_request_registry::request_registry_guard::RequestRegistryGuard;
-
-    use super::IncomingMessageContext;
-    use super::ManagementSocketClientService;
-    use crate::agent_applicable_state_holder::AgentApplicableStateHolder;
-    use crate::balancer_message_context::BalancerMessageContext;
-    use crate::continuous_batch_preparation_request::ContinuousBatchPreparationRequest;
-    use crate::model_metadata_holder::ModelMetadataHolder;
-
-    struct IncomingMessageFixture {
-        agent_desired_state_rx: mpsc::UnboundedReceiver<AgentDesiredState>,
-        context: IncomingMessageContext,
-        continuous_batch_preparation_request_rx:
-            mpsc::UnboundedReceiver<ContinuousBatchPreparationRequest>,
-        message_rx: mpsc::UnboundedReceiver<ManagementJsonRpcMessage>,
-    }
-
-    fn incoming_message_fixture() -> IncomingMessageFixture {
-        let (agent_desired_state_tx, agent_desired_state_rx) = mpsc::unbounded_channel();
-        let (continuous_batch_preparation_request_tx, continuous_batch_preparation_request_rx) =
-            mpsc::unbounded_channel();
-        let (message_tx, message_rx) = mpsc::unbounded_channel();
-
-        IncomingMessageFixture {
-            agent_desired_state_rx,
-            context: IncomingMessageContext {
-                balancer_message_context: BalancerMessageContext {
-                    agent_applicable_state_holder: Arc::new(AgentApplicableStateHolder::default()),
-                    agent_desired_state_tx,
-                    continuous_batch_preparation_request_tx,
-                    model_metadata_holder: Arc::new(ModelMetadataHolder::new()),
-                    request_stoppers: Arc::default(),
-                    slot_aggregated_status: Arc::new(SlotAggregatedStatus::new(2)),
-                },
-                connection_close: CancellationToken::new(),
-                message_tx,
-            },
-            continuous_batch_preparation_request_rx,
-            message_rx,
-        }
-    }
-
-    fn service_with_socket_url(socket_url: String) -> ManagementSocketClientService {
-        let (balancer_connection_tx, _initial_balancer_connection_rx) =
-            watch::channel(BalancerConnection::Connecting);
-
-        ManagementSocketClientService {
-            balancer_connection_tx,
-            balancer_message_context: incoming_message_fixture().context.balancer_message_context,
-            name: None,
-            socket_url,
-        }
-    }
-
-    fn raw_prompt_params() -> ContinueFromRawPromptParams {
-        ContinueFromRawPromptParams {
-            grammar: None,
-            max_tokens: NonZeroU32::new(8).unwrap(),
-            raw_prompt: "hello".to_owned(),
-        }
-    }
-
-    fn set_state_message() -> JsonRpcMessage {
-        JsonRpcMessage::Notification(JsonRpcNotification::SetState(Box::new(SetStateParams {
-            desired_state: AgentDesiredState::default(),
-        })))
-    }
-
-    struct StalledHandshakeFixture {
-        accepted_rx: oneshot::Receiver<()>,
-        server: JoinHandle<()>,
-        shutdown: CancellationToken,
-    }
-
-    fn spawn_stalled_handshake_fixture(listener: TcpListener) -> StalledHandshakeFixture {
-        let (accepted_tx, accepted_rx) = oneshot::channel::<()>();
-        let shutdown = CancellationToken::new();
-        let server_shutdown = shutdown.clone();
-
-        let server = spawn(async move {
-            let (_stream, _peer_addr) = listener
-                .accept()
-                .await
-                .expect("the fixture balancer must accept the agent connection");
-
-            accepted_tx
-                .send(())
-                .expect("the test must still be waiting for the accept signal");
-
-            server_shutdown.cancelled().await;
-        });
-
-        StalledHandshakeFixture {
-            accepted_rx,
-            server,
-            shutdown,
-        }
-    }
-
-    #[tokio::test]
-    async fn keep_connection_alive_returns_when_shutdown_arrives_during_a_stalled_handshake() {
-        let listener = TcpListener::bind("127.0.0.1:0").await.unwrap();
-        let addr = listener.local_addr().unwrap();
-        let StalledHandshakeFixture {
-            accepted_rx,
-            server,
-            shutdown: fixture_shutdown,
-        } = spawn_stalled_handshake_fixture(listener);
-
-        let service = service_with_socket_url(format!(
-            "ws://{addr}{}",
-            ApiPath::agent_socket("test-agent")
-        ));
-        let shutdown = CancellationToken::new();
-        let keep_alive_shutdown = shutdown.clone();
-        let keep_alive_handle =
-            spawn(async move { service.keep_connection_alive(keep_alive_shutdown).await });
-
-        accepted_rx
-            .await
-            .expect("the agent's connect must reach the fixture balancer before shutdown");
-
-        shutdown.cancel();
-
-        let keep_alive_result = keep_alive_handle
-            .await
-            .expect("the keep_connection_alive task must not panic");
-
-        assert!(keep_alive_result.is_ok());
-
-        fixture_shutdown.cancel();
-        server
-            .await
-            .expect("the fixture balancer task must not panic");
-    }
-
-    #[tokio::test]
-    async fn keep_connection_alive_errors_when_the_balancer_refuses_the_connection() {
-        let service = service_with_socket_url(format!(
-            "ws://127.0.0.1:1{}",
-            ApiPath::agent_socket("test-agent")
-        ));
-        let shutdown = CancellationToken::new();
-
-        let keep_alive_result = service.keep_connection_alive(shutdown).await;
-
-        assert!(keep_alive_result.is_err());
-    }
-
-    #[tokio::test]
-    async fn run_returns_when_shutdown_arrives_during_a_stalled_handshake() {
-        let listener = TcpListener::bind("127.0.0.1:0").await.unwrap();
-        let addr = listener.local_addr().unwrap();
-        let StalledHandshakeFixture {
-            accepted_rx,
-            server,
-            shutdown: fixture_shutdown,
-        } = spawn_stalled_handshake_fixture(listener);
-
-        let service = service_with_socket_url(format!(
-            "ws://{addr}{}",
-            ApiPath::agent_socket("test-agent")
-        ));
-        let shutdown = CancellationToken::new();
-        let run_shutdown = shutdown.clone();
-        let run_handle = spawn(async move { Box::new(service).run(run_shutdown).await });
-
-        accepted_rx
-            .await
-            .expect("the agent's connect must reach the fixture balancer before shutdown");
-
-        shutdown.cancel();
-
-        let run_result = run_handle.await.expect("the run task must not panic");
-
-        assert!(run_result.is_ok());
-
-        fixture_shutdown.cancel();
-        server
-            .await
-            .expect("the fixture balancer task must not panic");
-    }
-
-    #[tokio::test]
-    async fn announces_every_status_change_to_the_balancer() {
-        let listener = TcpListener::bind("127.0.0.1:0").await.unwrap();
-        let addr = listener.local_addr().unwrap();
-        let service = service_with_socket_url(format!(
-            "ws://{addr}{}",
-            ApiPath::agent_socket("test-agent")
-        ));
-        let slot_aggregated_status = service
-            .balancer_message_context
-            .slot_aggregated_status
-            .clone();
-        let shutdown = CancellationToken::new();
-        let keep_alive_shutdown = shutdown.clone();
-        let keep_alive_handle =
-            spawn(async move { service.keep_connection_alive(keep_alive_shutdown).await });
-
-        let (stream, _peer_addr) = listener.accept().await.unwrap();
-        let mut balancer_socket = accept_async(stream).await.unwrap();
-        let registration = balancer_socket.next().await.unwrap().unwrap();
-
-        assert!(matches!(
-            from_str::<ManagementJsonRpcMessage>(registration.to_text().unwrap()).unwrap(),
-            ManagementJsonRpcMessage::Notification(ManagementJsonRpcNotification::RegisterAgent(
-                register_agent_params
-            )) if register_agent_params.name.is_none()
-        ));
-
-        slot_aggregated_status.set_uses_chat_template_override(true);
-
-        let status_update = balancer_socket.next().await.unwrap().unwrap();
-
-        assert!(matches!(
-            from_str::<ManagementJsonRpcMessage>(status_update.to_text().unwrap()).unwrap(),
-            ManagementJsonRpcMessage::Notification(
-                ManagementJsonRpcNotification::UpdateAgentStatus(UpdateAgentStatusParams {
-                    slot_aggregated_status_snapshot: SlotAggregatedStatusSnapshot {
-                        status: AgentStatus {
-                            uses_chat_template_override: true,
-                            ..
-                        },
-                        ..
-                    },
-                })
-            )
-        ));
-
-        shutdown.cancel();
-
-        assert!(keep_alive_handle.await.unwrap().is_ok());
+#[async_trait]
+impl Service for ManagementSocketClientService {
+    fn name(&self) -> &'static str {
+        "agent::management_socket_client_service"
     }
 
-    #[tokio::test]
-    async fn reports_the_balancer_connected_only_while_its_socket_is_open() {
-        let listener = TcpListener::bind("127.0.0.1:0").await.unwrap();
-        let addr = listener.local_addr().unwrap();
-        let service = service_with_socket_url(format!(
-            "ws://{addr}{}",
-            ApiPath::agent_socket("test-agent")
-        ));
-        let mut balancer_connection_rx = service.balancer_connection_tx.subscribe();
-        let keep_alive_handle = spawn(async move {
-            service
-                .keep_connection_alive(CancellationToken::new())
-                .await
-        });
-
-        let (stream, _peer_addr) = listener.accept().await.unwrap();
-        let balancer_socket = accept_async(stream).await.unwrap();
-
-        balancer_connection_rx
-            .wait_for(|balancer_connection| *balancer_connection == BalancerConnection::Connected)
-            .await
-            .unwrap();
-
-        drop(balancer_socket);
-
-        balancer_connection_rx
-            .wait_for(|balancer_connection| *balancer_connection == BalancerConnection::Connecting)
+    async fn run(self: Box<Self>, shutdown: CancellationToken) -> Result<()> {
+        self.reconnect_to_the_balancer_until_shutdown(shutdown)
             .await
-            .unwrap();
-
-        assert!(keep_alive_handle.await.unwrap().is_ok());
-    }
-
-    #[tokio::test]
-    async fn set_state_notification_forwards_desired_state() {
-        let mut fixture = incoming_message_fixture();
-
-        fixture
-            .context
-            .handle_deserialized_message(set_state_message())
-            .unwrap();
-
-        assert_eq!(
-            fixture.agent_desired_state_rx.try_recv().unwrap(),
-            AgentDesiredState::default()
-        );
-    }
-
-    #[tokio::test]
-    async fn set_state_notification_errors_when_receiver_dropped() {
-        let fixture = incoming_message_fixture();
-
-        drop(fixture.agent_desired_state_rx);
-
-        assert!(
-            fixture
-                .context
-                .handle_deserialized_message(set_state_message())
-                .is_err()
-        );
-    }
-
-    #[tokio::test]
-    async fn stop_responding_to_a_finished_request_is_not_an_error_and_retains_nothing() {
-        let fixture = incoming_message_fixture();
-        let request_stoppers = fixture
-            .context
-            .balancer_message_context
-            .request_stoppers
-            .clone();
-
-        fixture
-            .context
-            .handle_deserialized_message(JsonRpcMessage::Notification(
-                JsonRpcNotification::StopRespondingTo("already_finished".to_owned()),
-            ))
-            .expect("a stop for a request that already finished must not be an error");
-
-        assert!(
-            request_stoppers
-                .with_registered("already_finished", |_registered_stopper| ())
-                .is_none(),
-            "the stop must leave nothing behind; retaining it would leak one entry per cancelled \
-             request"
-        );
-    }
-
-    #[tokio::test]
-    async fn stop_responding_to_registered_request_signals_stopper() {
-        let fixture = incoming_message_fixture();
-        let (stop_tx, mut stop_rx) = mpsc::unbounded_channel::<()>();
-
-        let _stopper_guard = RequestRegistryGuard::register(
-            &fixture.context.balancer_message_context.request_stoppers,
-            "active_request".to_owned(),
-            stop_tx,
-        )
-        .expect("a fresh request id must register");
-
-        fixture
-            .context
-            .handle_deserialized_message(JsonRpcMessage::Notification(
-                JsonRpcNotification::StopRespondingTo("active_request".to_owned()),
-            ))
-            .unwrap();
-
-        assert_eq!(stop_rx.try_recv(), Ok(()));
-    }
-
-    #[tokio::test]
-    async fn mismatched_version_notification_is_acknowledged() {
-        let mut fixture = incoming_message_fixture();
-
-        fixture
-            .context
-            .handle_deserialized_message(JsonRpcMessage::Notification(
-                JsonRpcNotification::Version(VersionParams {
-                    version: "0.0.0-mismatch".to_owned(),
-                }),
-            ))
-            .unwrap();
-
-        assert!(fixture.message_rx.try_recv().is_err());
-    }
-
-    #[tokio::test]
-    async fn get_chat_template_override_without_applicable_state_responds_with_none() {
-        let mut fixture = incoming_message_fixture();
-
-        fixture
-            .context
-            .handle_deserialized_message(JsonRpcMessage::Request(RequestEnvelope {
-                id: "req_template".to_owned(),
-                request: JsonRpcRequest::GetChatTemplateOverride,
-            }))
-            .unwrap();
-
-        assert!(matches!(
-            fixture.message_rx.try_recv().unwrap(),
-            ManagementJsonRpcMessage::Response(ResponseEnvelope {
-                request_id,
-                response: JsonRpcResponse::ChatTemplateOverride(None),
-                ..
-            }) if request_id == "req_template"
-        ));
-    }
-
-    #[tokio::test]
-    async fn get_chat_template_override_errors_when_message_receiver_dropped() {
-        let fixture = incoming_message_fixture();
-
-        drop(fixture.message_rx);
-
-        assert!(
-            fixture
-                .context
-                .handle_deserialized_message(JsonRpcMessage::Request(RequestEnvelope {
-                    id: "req_template".to_owned(),
-                    request: JsonRpcRequest::GetChatTemplateOverride,
-                }))
-                .is_err()
-        );
-    }
-
-    #[tokio::test]
-    async fn get_model_metadata_responds_with_stored_metadata() {
-        let mut fixture = incoming_message_fixture();
-        let mut metadata = BTreeMap::new();
-
-        metadata.insert("architecture".to_owned(), "llama".to_owned());
-        fixture
-            .context
-            .balancer_message_context
-            .model_metadata_holder
-            .set_model_metadata(ModelMetadata {
-                metadata: metadata.clone(),
-            });
-
-        fixture
-            .context
-            .handle_deserialized_message(JsonRpcMessage::Request(RequestEnvelope {
-                id: "req_metadata".to_owned(),
-                request: JsonRpcRequest::GetModelMetadata,
-            }))
-            .unwrap();
-
-        assert!(matches!(
-            fixture.message_rx.try_recv().unwrap(),
-            ManagementJsonRpcMessage::Response(ResponseEnvelope {
-                response: JsonRpcResponse::ModelMetadata(Some(returned_metadata)),
-                ..
-            }) if returned_metadata.metadata == metadata
-        ));
-    }
-
-    #[tokio::test]
-    async fn get_model_metadata_errors_when_message_receiver_dropped() {
-        let fixture = incoming_message_fixture();
-
-        drop(fixture.message_rx);
-
-        assert!(
-            fixture
-                .context
-                .handle_deserialized_message(JsonRpcMessage::Request(RequestEnvelope {
-                    id: "req_metadata".to_owned(),
-                    request: JsonRpcRequest::GetModelMetadata,
-                }))
-                .is_err()
-        );
-    }
-
-    #[tokio::test]
-    async fn text_message_dispatches_deserialized_set_state() {
-        let mut fixture = incoming_message_fixture();
-
-        fixture.context.handle_incoming_message(Message::Text(
-            to_string(&set_state_message()).unwrap().into(),
-        ));
-
-        assert_eq!(
-            fixture.agent_desired_state_rx.recv().await.unwrap(),
-            AgentDesiredState::default()
-        );
-    }
-
-    #[tokio::test]
-    async fn undecodable_text_message_closes_the_connection() {
-        let fixture = incoming_message_fixture();
-        let connection_close = fixture.context.connection_close.clone();
-
-        fixture
-            .context
-            .handle_incoming_message(Message::Text("not json".into()));
-
-        assert!(connection_close.is_cancelled());
-    }
-
-    #[tokio::test]
-    async fn binary_message_closes_the_connection() {
-        let fixture = incoming_message_fixture();
-        let connection_close = fixture.context.connection_close.clone();
-
-        fixture
-            .context
-            .handle_incoming_message(Message::Binary(Bytes::from_static(b"\x00\x01")));
-
-        assert!(connection_close.is_cancelled());
-    }
-
-    #[tokio::test]
-    async fn generate_responses_registers_the_stopper_before_it_returns() {
-        let mut fixture = incoming_message_fixture();
-        let request_stoppers = fixture
-            .context
-            .balancer_message_context
-            .request_stoppers
-            .clone();
-
-        fixture
-            .context
-            .generate_responses::<_, GeneratedTokenResult>(
-                "req_generate".to_owned(),
-                raw_prompt_params(),
-            )
-            .expect("the request must be accepted");
-
-        let dispatched_request = fixture
-            .continuous_batch_preparation_request_rx
-            .try_recv()
-            .expect("the request must be dispatched to the arbiter");
-
-        assert!(matches!(
-            &dispatched_request,
-            ContinuousBatchPreparationRequest::ContinueFromRawPrompt(request)
-                if request.params.raw_prompt == "hello"
-        ));
-        assert_eq!(
-            request_stoppers.send_to("req_generate", ()),
-            RequestDelivery::Delivered,
-            "the stopper must be registered before generate_responses returns, so a stop arriving \
-             on the next frame cannot be lost"
-        );
-    }
-
-    #[tokio::test]
-    async fn generate_responses_errors_when_request_receiver_dropped() {
-        let fixture = incoming_message_fixture();
-
-        drop(fixture.continuous_batch_preparation_request_rx);
-
-        assert!(
-            fixture
-                .context
-                .generate_responses::<_, GeneratedTokenResult>(
-                    "req_generate".to_owned(),
-                    raw_prompt_params(),
-                )
-                .is_err()
-        );
-    }
-
-    #[tokio::test]
-    async fn generate_responses_errors_when_stopper_already_registered() {
-        let fixture = incoming_message_fixture();
-        let (existing_stop_tx, _existing_stop_rx) = mpsc::unbounded_channel::<()>();
-
-        let _existing_stopper_guard = RequestRegistryGuard::register(
-            &fixture.context.balancer_message_context.request_stoppers,
-            "req_generate".to_owned(),
-            existing_stop_tx,
-        )
-        .expect("a fresh request id must register");
-
-        assert!(
-            fixture
-                .context
-                .generate_responses::<_, GeneratedTokenResult>(
-                    "req_generate".to_owned(),
-                    raw_prompt_params(),
-                )
-                .is_err()
-        );
     }
 }
diff --git a/paddler_agent/src/model_bos_token.rs b/paddler_agent/src/model_bos_token.rs
deleted file mode 100644
index 10b86efa2..000000000
--- a/paddler_agent/src/model_bos_token.rs
+++ /dev/null
@@ -1,54 +0,0 @@
-use llama_cpp_bindings::model::LlamaModel;
-use llama_cpp_bindings::token::LlamaToken;
-
-pub struct ModelBosToken {
-    pub added_by_tokenizer: bool,
-    pub token: LlamaToken,
-}
-
-impl ModelBosToken {
-    #[must_use]
-    pub fn of(model: &LlamaModel) -> Self {
-        Self {
-            added_by_tokenizer: model.adds_bos_token(),
-            token: model.token_bos(),
-        }
-    }
-
-    #[must_use]
-    pub fn is_duplicated_at_start_of(&self, prompt_tokens: &[LlamaToken]) -> bool {
-        self.added_by_tokenizer && prompt_tokens.starts_with(&[self.token, self.token])
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use llama_cpp_bindings::token::LlamaToken;
-
-    use super::ModelBosToken;
-
-    const BOS: LlamaToken = LlamaToken(1);
-    const TEXT: LlamaToken = LlamaToken(42);
-
-    fn bos_added_by_tokenizer(added_by_tokenizer: bool) -> ModelBosToken {
-        ModelBosToken {
-            added_by_tokenizer,
-            token: BOS,
-        }
-    }
-
-    #[test]
-    fn detects_a_rendered_bos_after_the_one_the_tokenizer_added() {
-        assert!(bos_added_by_tokenizer(true).is_duplicated_at_start_of(&[BOS, BOS, TEXT]));
-    }
-
-    #[test]
-    fn accepts_a_single_leading_bos() {
-        assert!(!bos_added_by_tokenizer(true).is_duplicated_at_start_of(&[BOS, TEXT]));
-    }
-
-    #[test]
-    fn keeps_repeated_bos_tokens_the_tokenizer_did_not_add() {
-        assert!(!bos_added_by_tokenizer(false).is_duplicated_at_start_of(&[BOS, BOS, TEXT]));
-    }
-}
diff --git a/paddler_agent/src/model_metadata_holder.rs b/paddler_agent/src/model_metadata_holder.rs
deleted file mode 100644
index e00a7d597..000000000
--- a/paddler_agent/src/model_metadata_holder.rs
+++ /dev/null
@@ -1,63 +0,0 @@
-use parking_lot::RwLock;
-
-use paddler_messaging::model_metadata::ModelMetadata;
-
-pub struct ModelMetadataHolder {
-    model_metadata: RwLock<Option<ModelMetadata>>,
-}
-
-impl ModelMetadataHolder {
-    #[must_use]
-    pub fn new() -> Self {
-        Self::default()
-    }
-
-    pub fn set_model_metadata(&self, metadata: ModelMetadata) {
-        let mut lock = self.model_metadata.write();
-
-        *lock = Some(metadata);
-    }
-
-    pub fn get_model_metadata(&self) -> Option<ModelMetadata> {
-        let lock = self.model_metadata.read();
-
-        lock.clone()
-    }
-}
-
-impl Default for ModelMetadataHolder {
-    fn default() -> Self {
-        Self {
-            model_metadata: RwLock::new(None),
-        }
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use std::collections::BTreeMap;
-
-    use paddler_messaging::model_metadata::ModelMetadata;
-
-    use super::ModelMetadataHolder;
-
-    #[test]
-    fn new_holder_starts_empty() {
-        let holder = ModelMetadataHolder::new();
-
-        assert!(holder.get_model_metadata().is_none());
-    }
-
-    #[test]
-    fn stored_metadata_is_returned() {
-        let holder = ModelMetadataHolder::new();
-        let mut metadata = BTreeMap::new();
-        metadata.insert("architecture".to_owned(), "llama".to_owned());
-
-        holder.set_model_metadata(ModelMetadata { metadata });
-
-        let stored = holder.get_model_metadata().unwrap();
-
-        assert_eq!(stored.metadata.get("architecture").unwrap(), "llama");
-    }
-}
diff --git a/paddler_agent/src/model_serving.rs b/paddler_agent/src/model_serving.rs
new file mode 100644
index 000000000..9ab691742
--- /dev/null
+++ b/paddler_agent/src/model_serving.rs
@@ -0,0 +1,77 @@
+use paddler_agent_runtime::loaded_llama_model::LoadedLlamaModel;
+use paddler_agent_runtime::shuts_down::ShutsDown;
+
+use crate::agent_applicable_state::AgentApplicableState;
+use crate::applicable_decisions::ApplicableDecisions;
+use crate::loaded_model_serving_plan::LoadedModelServingPlan;
+use crate::model_serving_settings::ModelServingSettings;
+use crate::pipeline_request::PipelineRequest;
+use crate::retained_serving::RetainedServing;
+use crate::served_model::ServedModel;
+
+pub enum ModelServing {
+    Holding(LoadedLlamaModel),
+    Idle,
+    Serving(Box<ServedModel>),
+}
+
+impl ModelServing {
+    #[must_use]
+    pub fn decisions_to_prepare(
+        &self,
+        agent_applicable_state: &AgentApplicableState,
+    ) -> ApplicableDecisions {
+        match (self, agent_applicable_state) {
+            (Self::Serving(served_model), AgentApplicableState::WithModel(desired_settings))
+                if served_model.keeps_decisions_of(desired_settings) =>
+            {
+                ApplicableDecisions::Disabled
+            }
+            (_, AgentApplicableState::WithModel(desired_settings)) => {
+                desired_settings.capabilities.decisions.clone()
+            }
+            (
+                _,
+                AgentApplicableState::NotConfigured | AgentApplicableState::WithoutModel { .. },
+            ) => ApplicableDecisions::Disabled,
+        }
+    }
+
+    pub fn forward(&self, agent_name: Option<&str>, request: PipelineRequest) {
+        match self {
+            Self::Holding(_) | Self::Idle => request.reject_because_no_model_is_loaded(agent_name),
+            Self::Serving(served_model) => served_model.forward(agent_name, request),
+        }
+    }
+
+    pub async fn retain_for(self, desired_settings: ModelServingSettings) -> RetainedServing {
+        match self {
+            Self::Holding(loaded_llama_model)
+                if loaded_llama_model.llama_model_settings
+                    == desired_settings.llama_model_settings() =>
+            {
+                RetainedServing::LoadedModel(LoadedModelServingPlan::starting_every_capability(
+                    loaded_llama_model,
+                    desired_settings,
+                ))
+            }
+            Self::Holding(loaded_llama_model) => {
+                loaded_llama_model.release().await;
+
+                RetainedServing::ModelToLoad(desired_settings)
+            }
+            Self::Idle => RetainedServing::ModelToLoad(desired_settings),
+            Self::Serving(served_model) => served_model.retain_for(desired_settings).await,
+        }
+    }
+}
+
+impl ShutsDown for ModelServing {
+    async fn shut_down(self) {
+        match self {
+            Self::Holding(loaded_llama_model) => loaded_llama_model.release().await,
+            Self::Idle => {}
+            Self::Serving(served_model) => served_model.shut_down().await,
+        }
+    }
+}
diff --git a/paddler_agent/src/model_serving_settings.rs b/paddler_agent/src/model_serving_settings.rs
new file mode 100644
index 000000000..2776df1a1
--- /dev/null
+++ b/paddler_agent/src/model_serving_settings.rs
@@ -0,0 +1,27 @@
+use std::path::PathBuf;
+
+use paddler_agent_runtime::llama_model_settings::LlamaModelSettings;
+use paddler_agent_text_generation::text_generation_settings::TextGenerationSettings;
+use paddler_inference_parameters::model_runtime_parameters::ModelRuntimeParameters;
+use paddler_inference_parameters::pooling_type::PoolingType;
+
+use crate::applicable_decisions::ApplicableDecisions;
+use crate::inference_capabilities::InferenceCapabilities;
+
+#[derive(Clone, Debug, PartialEq)]
+pub struct ModelServingSettings {
+    pub capabilities:
+        InferenceCapabilities<ApplicableDecisions, PoolingType, TextGenerationSettings>,
+    pub model_path: PathBuf,
+    pub model_runtime_parameters: ModelRuntimeParameters,
+}
+
+impl ModelServingSettings {
+    #[must_use]
+    pub fn llama_model_settings(&self) -> LlamaModelSettings {
+        LlamaModelSettings {
+            model_path: self.model_path.clone(),
+            n_gpu_layers: self.model_runtime_parameters.n_gpu_layers,
+        }
+    }
+}
diff --git a/paddler_agent/src/multimodal_prompt_ingestion.rs b/paddler_agent/src/multimodal_prompt_ingestion.rs
deleted file mode 100644
index 95070a741..000000000
--- a/paddler_agent/src/multimodal_prompt_ingestion.rs
+++ /dev/null
@@ -1,53 +0,0 @@
-use std::sync::Arc;
-
-use llama_cpp_bindings::SampledTokenClassifier;
-use llama_cpp_bindings::context::LlamaContext;
-use llama_cpp_bindings::error::EvalMultimodalChunksError;
-use llama_cpp_bindings::ingest_prompt_chunk;
-use llama_cpp_bindings::mtmd::MtmdContext;
-use llama_cpp_bindings::mtmd::MtmdInputChunks;
-
-use crate::multimodal_ingestion_progress::MultimodalIngestionProgress;
-
-#[derive(Debug)]
-pub struct MultimodalPromptIngestion {
-    pub chunks: MtmdInputChunks,
-    pub multimodal_context: Arc<MtmdContext>,
-    pub n_batch: i32,
-    pub next_chunk_index: usize,
-}
-
-impl MultimodalPromptIngestion {
-    pub fn ingest_next_chunk(
-        &mut self,
-        token_classifier: &mut SampledTokenClassifier<'_>,
-        llama_context: &LlamaContext,
-        sequence_id: i32,
-        start_position: i32,
-    ) -> Result<MultimodalIngestionProgress, EvalMultimodalChunksError> {
-        let chunk_index = self.next_chunk_index;
-        let chunk = self
-            .chunks
-            .get(chunk_index)
-            .ok_or(EvalMultimodalChunksError::ChunkOutOfBounds(chunk_index))?;
-        let is_last_chunk = chunk_index + 1 == self.chunks.len();
-        let next_position = chunk.eval_single(
-            &self.multimodal_context,
-            llama_context,
-            start_position,
-            sequence_id,
-            self.n_batch,
-            is_last_chunk,
-        )?;
-
-        ingest_prompt_chunk(token_classifier, &chunk)?;
-
-        self.next_chunk_index += 1;
-
-        Ok(if is_last_chunk {
-            MultimodalIngestionProgress::PromptIngested { next_position }
-        } else {
-            MultimodalIngestionProgress::ChunksRemain { next_position }
-        })
-    }
-}
diff --git a/paddler_agent/src/multimodal_prompt_support.rs b/paddler_agent/src/multimodal_prompt_support.rs
deleted file mode 100644
index f39ce6caa..000000000
--- a/paddler_agent/src/multimodal_prompt_support.rs
+++ /dev/null
@@ -1,51 +0,0 @@
-use std::sync::Arc;
-
-use llama_cpp_bindings::mtmd::MtmdBitmap;
-use llama_cpp_bindings::mtmd::MtmdContext;
-use llama_cpp_bindings::mtmd::MtmdInputText;
-
-use crate::generation_request_rejection::GenerationRequestRejection;
-use crate::multimodal_prompt_ingestion::MultimodalPromptIngestion;
-use crate::require_prompt_fits_sequence_context::require_prompt_fits_sequence_context;
-
-pub struct MultimodalPromptSupport {
-    pub micro_batch_tokens: u32,
-    pub multimodal_context: Arc<MtmdContext>,
-    pub n_batch: i32,
-    pub sequence_context_size: u32,
-}
-
-impl MultimodalPromptSupport {
-    pub fn prepare_prompt(
-        &self,
-        bitmaps: &[MtmdBitmap],
-        text: String,
-        add_special_tokens: bool,
-    ) -> Result<MultimodalPromptIngestion, GenerationRequestRejection> {
-        let bitmap_refs: Vec<&MtmdBitmap> = bitmaps.iter().collect();
-        let chunks = self
-            .multimodal_context
-            .tokenize(
-                MtmdInputText {
-                    text,
-                    add_special: add_special_tokens,
-                    parse_special: true,
-                },
-                &bitmap_refs,
-            )
-            .map_err(GenerationRequestRejection::MultimodalTokenizationFailed)?;
-
-        require_prompt_fits_sequence_context(chunks.total_tokens(), self.sequence_context_size)?;
-
-        chunks
-            .fit_to_micro_batch(&self.multimodal_context, self.micro_batch_tokens)
-            .map_err(GenerationRequestRejection::for_micro_batch_fit_error)?;
-
-        Ok(MultimodalPromptIngestion {
-            chunks,
-            multimodal_context: self.multimodal_context.clone(),
-            n_batch: self.n_batch,
-            next_chunk_index: 0,
-        })
-    }
-}
diff --git a/paddler_agent/src/pipeline_arbiter.rs b/paddler_agent/src/pipeline_arbiter.rs
new file mode 100644
index 000000000..e91abbcf1
--- /dev/null
+++ b/paddler_agent/src/pipeline_arbiter.rs
@@ -0,0 +1,262 @@
+use std::mem;
+use std::sync::Arc;
+
+use log::error;
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent_runtime::inference_runtime_context::InferenceRuntimeContext;
+use paddler_agent_runtime::shuts_down::ShutsDown as _;
+
+use crate::agent_applicable_state::AgentApplicableState;
+use crate::agent_applicable_state_holder::AgentApplicableStateHolder;
+use crate::finished_serving_startup::FinishedServingStartup;
+use crate::model_serving::ModelServing;
+use crate::pipeline_arbiter_params::PipelineArbiterParams;
+use crate::pipeline_request::PipelineRequest;
+use crate::preparation_outcome::PreparationOutcome;
+use crate::prepared_serving::PreparedServing;
+use crate::replacement_outcome::ReplacementOutcome;
+use crate::retry_delay::RetryDelay;
+use crate::served_model::ServedModel;
+use crate::serving_preparation::ServingPreparation;
+use crate::serving_replacement::ServingReplacement;
+use crate::serving_startup::ServingStartup;
+use crate::serving_startup_progress::ServingStartupProgress;
+use crate::startup_phase::StartupPhase;
+use crate::superseding_state::SupersedingState;
+
+pub struct PipelineArbiter {
+    agent_shutdown: CancellationToken,
+    inference_runtime_context: InferenceRuntimeContext,
+    model_serving: ModelServing,
+    served_applicable_state_holder: Arc<AgentApplicableStateHolder>,
+    serving_startup: ServingStartup,
+    superseding_state: SupersedingState,
+}
+
+impl PipelineArbiter {
+    #[must_use]
+    pub fn new(
+        PipelineArbiterParams {
+            agent_shutdown,
+            inference_runtime_context,
+            served_applicable_state_holder,
+        }: PipelineArbiterParams,
+    ) -> Self {
+        Self {
+            agent_shutdown,
+            inference_runtime_context,
+            model_serving: ModelServing::Idle,
+            served_applicable_state_holder,
+            serving_startup: ServingStartup::Idle,
+            superseding_state: SupersedingState::NotSuperseded,
+        }
+    }
+
+    pub fn apply(&mut self, agent_applicable_state: AgentApplicableState) {
+        match self.serving_startup {
+            ServingStartup::Idle => self.start_pipeline(agent_applicable_state),
+            ServingStartup::Preparing(_) | ServingStartup::Replacing(_) => {
+                self.inference_runtime_context
+                    .slot_aggregated_status
+                    .begin_pipeline();
+                self.serving_startup.cancel();
+                self.superseding_state = SupersedingState::SupersededBy(agent_applicable_state);
+            }
+        }
+    }
+
+    pub async fn next_finished_startup(&mut self) -> ServingStartupProgress {
+        self.serving_startup.next_progress().await
+    }
+
+    pub fn conclude(&mut self, serving_startup_progress: ServingStartupProgress) {
+        match serving_startup_progress {
+            ServingStartupProgress::Prepared(preparation_outcome) => {
+                self.conclude_preparation(preparation_outcome);
+            }
+            ServingStartupProgress::Replaced(finished_serving_startup) => {
+                self.conclude_replacement(finished_serving_startup);
+            }
+        }
+    }
+
+    pub fn forward(&self, request: PipelineRequest) {
+        self.model_serving.forward(
+            self.inference_runtime_context.agent_name.as_deref(),
+            request,
+        );
+    }
+
+    pub async fn shut_down(self) {
+        let Self {
+            model_serving,
+            serving_startup,
+            ..
+        } = self;
+
+        serving_startup.shut_down().await;
+        model_serving.shut_down().await;
+    }
+
+    fn conclude_preparation(&mut self, preparation_outcome: PreparationOutcome) {
+        if let PreparationOutcome::Failed { error, .. } = &preparation_outcome {
+            error!("Failed to prepare the reconciled state: {error}");
+        }
+
+        if let SupersedingState::SupersededBy(superseding_applicable_state) =
+            mem::replace(&mut self.superseding_state, SupersedingState::NotSuperseded)
+        {
+            self.start_pipeline(superseding_applicable_state);
+
+            return;
+        }
+
+        match preparation_outcome {
+            PreparationOutcome::Cancelled => {}
+            PreparationOutcome::Failed {
+                agent_applicable_state,
+                ..
+            } => {
+                self.inference_runtime_context
+                    .slot_aggregated_status
+                    .retry_pipeline();
+                self.prepare(agent_applicable_state, RetryDelay::AfterRetryInterval);
+            }
+            PreparationOutcome::Prepared(prepared_serving) => self.replace(prepared_serving),
+        }
+    }
+
+    fn conclude_replacement(
+        &mut self,
+        FinishedServingStartup {
+            agent_applicable_state,
+            outcome,
+        }: FinishedServingStartup,
+    ) {
+        if let ReplacementOutcome::Failed { error, .. } = &outcome {
+            error!("Failed to serve the reconciled state: {error}");
+        }
+
+        if let SupersedingState::SupersededBy(superseding_applicable_state) =
+            mem::replace(&mut self.superseding_state, SupersedingState::NotSuperseded)
+        {
+            self.model_serving = outcome.into_model_serving();
+            self.start_pipeline(superseding_applicable_state);
+
+            return;
+        }
+
+        self.settle_replacement(agent_applicable_state, &outcome);
+        self.model_serving = outcome.into_model_serving();
+    }
+
+    fn prepare(&mut self, agent_applicable_state: AgentApplicableState, retry_delay: RetryDelay) {
+        let cancellation_token = self.agent_shutdown.child_token();
+        let decisions_to_prepare = self
+            .model_serving
+            .decisions_to_prepare(&agent_applicable_state);
+
+        self.serving_startup = ServingStartup::Preparing(StartupPhase {
+            cancellation_token: cancellation_token.clone(),
+            outcome: Box::pin(
+                ServingPreparation {
+                    agent_applicable_state,
+                    decisions_to_prepare,
+                    retry_delay,
+                }
+                .prepare(cancellation_token, self.inference_runtime_context.clone()),
+            ),
+        });
+    }
+
+    fn replace(&mut self, prepared_serving: PreparedServing) {
+        if let ModelServing::Serving(served_model) = &self.model_serving
+            && let AgentApplicableState::WithModel(desired_settings) =
+                &prepared_serving.agent_applicable_state
+            && served_model.model_serving_settings == *desired_settings
+        {
+            self.resume_serving(served_model, prepared_serving.agent_applicable_state);
+
+            return;
+        }
+
+        let held = self.withdraw_serving();
+        let cancellation_token = self.agent_shutdown.child_token();
+
+        self.serving_startup = ServingStartup::Replacing(StartupPhase {
+            cancellation_token: cancellation_token.clone(),
+            outcome: Box::pin(
+                ServingReplacement {
+                    held,
+                    prepared_serving,
+                }
+                .replace(cancellation_token, self.inference_runtime_context.clone()),
+            ),
+        });
+    }
+
+    fn resume_serving(
+        &self,
+        served_model: &ServedModel,
+        agent_applicable_state: AgentApplicableState,
+    ) {
+        self.inference_runtime_context
+            .slot_aggregated_status
+            .start_serving(
+                served_model
+                    .serving_report(agent_applicable_state.chat_template_override().is_some()),
+            );
+        self.inference_runtime_context
+            .model_metadata_holder
+            .set_model_metadata(served_model.loaded_llama_model.model_metadata.clone());
+        self.served_applicable_state_holder
+            .set_agent_applicable_state(agent_applicable_state);
+    }
+
+    fn settle_replacement(
+        &mut self,
+        agent_applicable_state: AgentApplicableState,
+        outcome: &ReplacementOutcome,
+    ) {
+        match outcome {
+            ReplacementOutcome::Cancelled { .. } => {}
+            ReplacementOutcome::Failed { .. } => {
+                self.inference_runtime_context
+                    .slot_aggregated_status
+                    .retry_pipeline();
+                self.prepare(agent_applicable_state, RetryDelay::AfterRetryInterval);
+            }
+            ReplacementOutcome::NothingToServe => {
+                self.inference_runtime_context
+                    .slot_aggregated_status
+                    .serve_nothing(agent_applicable_state.chat_template_override().is_some());
+                self.served_applicable_state_holder
+                    .set_agent_applicable_state(agent_applicable_state);
+            }
+            ReplacementOutcome::Serving(served_model) => {
+                self.resume_serving(served_model, agent_applicable_state);
+            }
+        }
+    }
+
+    fn start_pipeline(&mut self, agent_applicable_state: AgentApplicableState) {
+        self.inference_runtime_context
+            .slot_aggregated_status
+            .begin_pipeline();
+        self.prepare(agent_applicable_state, RetryDelay::Immediately);
+    }
+
+    fn withdraw_serving(&mut self) -> ModelServing {
+        self.inference_runtime_context
+            .slot_aggregated_status
+            .withdraw_serving();
+        self.inference_runtime_context
+            .model_metadata_holder
+            .forget_model_metadata();
+        self.served_applicable_state_holder
+            .set_agent_applicable_state(AgentApplicableState::NotConfigured);
+
+        mem::replace(&mut self.model_serving, ModelServing::Idle)
+    }
+}
diff --git a/paddler_agent/src/pipeline_arbiter_params.rs b/paddler_agent/src/pipeline_arbiter_params.rs
new file mode 100644
index 000000000..d72d45bb0
--- /dev/null
+++ b/paddler_agent/src/pipeline_arbiter_params.rs
@@ -0,0 +1,13 @@
+use std::sync::Arc;
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent_runtime::inference_runtime_context::InferenceRuntimeContext;
+
+use crate::agent_applicable_state_holder::AgentApplicableStateHolder;
+
+pub struct PipelineArbiterParams {
+    pub agent_shutdown: CancellationToken,
+    pub inference_runtime_context: InferenceRuntimeContext,
+    pub served_applicable_state_holder: Arc<AgentApplicableStateHolder>,
+}
diff --git a/paddler_agent/src/pipeline_request.rs b/paddler_agent/src/pipeline_request.rs
new file mode 100644
index 000000000..ff6e83f88
--- /dev/null
+++ b/paddler_agent/src/pipeline_request.rs
@@ -0,0 +1,78 @@
+use paddler_agent_decision::decision_error::DecisionError;
+use paddler_agent_decision::reports_decision_outcome::ReportsDecisionOutcome as _;
+use paddler_agent_embeddings::embedding_error::EmbeddingError;
+use paddler_agent_runtime::agent_request::AgentRequest;
+use paddler_agent_text_generation::text_generation_error::TextGenerationError;
+use paddler_messaging::decision_result::DecisionResult;
+use paddler_messaging::embedding_result::EmbeddingResult;
+use paddler_messaging::generated_token_result::GeneratedTokenResult;
+use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
+use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters_schema::validated_parameters_schema::ValidatedParametersSchema;
+use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
+use paddler_messaging::request_params::decide_params::DecideParams;
+use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
+
+pub enum PipelineRequest {
+    ContinueFromConversationHistory(
+        AgentRequest<
+            ContinueFromConversationHistoryParams<ValidatedParametersSchema>,
+            GeneratedTokenResult,
+        >,
+    ),
+    ContinueFromRawPrompt(AgentRequest<ContinueFromRawPromptParams, GeneratedTokenResult>),
+    Decide(AgentRequest<DecideParams, DecisionResult>),
+    GenerateEmbeddingBatch(AgentRequest<GenerateEmbeddingBatchParams, EmbeddingResult>),
+}
+
+impl PipelineRequest {
+    pub fn reject_because_no_model_is_loaded(self, agent_name: Option<&str>) {
+        match self {
+            Self::ContinueFromConversationHistory(AgentRequest { response_tx, .. })
+            | Self::ContinueFromRawPrompt(AgentRequest { response_tx, .. }) => {
+                TextGenerationError::ModelNotLoaded.report(agent_name, &response_tx);
+            }
+            Self::Decide(AgentRequest { response_tx, .. }) => {
+                DecisionError::ModelNotLoaded.report(agent_name, &response_tx);
+            }
+            Self::GenerateEmbeddingBatch(AgentRequest { response_tx, .. }) => {
+                EmbeddingError::ModelNotLoaded.report(agent_name, &response_tx);
+            }
+        }
+    }
+}
+
+impl
+    From<
+        AgentRequest<
+            ContinueFromConversationHistoryParams<ValidatedParametersSchema>,
+            GeneratedTokenResult,
+        >,
+    > for PipelineRequest
+{
+    fn from(
+        request: AgentRequest<
+            ContinueFromConversationHistoryParams<ValidatedParametersSchema>,
+            GeneratedTokenResult,
+        >,
+    ) -> Self {
+        Self::ContinueFromConversationHistory(request)
+    }
+}
+
+impl From<AgentRequest<ContinueFromRawPromptParams, GeneratedTokenResult>> for PipelineRequest {
+    fn from(request: AgentRequest<ContinueFromRawPromptParams, GeneratedTokenResult>) -> Self {
+        Self::ContinueFromRawPrompt(request)
+    }
+}
+
+impl From<AgentRequest<DecideParams, DecisionResult>> for PipelineRequest {
+    fn from(request: AgentRequest<DecideParams, DecisionResult>) -> Self {
+        Self::Decide(request)
+    }
+}
+
+impl From<AgentRequest<GenerateEmbeddingBatchParams, EmbeddingResult>> for PipelineRequest {
+    fn from(request: AgentRequest<GenerateEmbeddingBatchParams, EmbeddingResult>) -> Self {
+        Self::GenerateEmbeddingBatch(request)
+    }
+}
diff --git a/paddler_agent/src/pipeline_response_stream.rs b/paddler_agent/src/pipeline_response_stream.rs
new file mode 100644
index 000000000..d128fd8af
--- /dev/null
+++ b/paddler_agent/src/pipeline_response_stream.rs
@@ -0,0 +1,37 @@
+use std::pin::Pin;
+use std::task::Context;
+use std::task::Poll;
+
+use futures_util::Stream;
+use tokio::sync::mpsc;
+
+use paddler_messaging::management_socket::agent::response::Response as JsonRpcResponse;
+use paddler_messaging::management_socket::balancer::message::Message as ManagementJsonRpcMessage;
+use paddler_request_registry::request_registry_guard::RequestRegistryGuard;
+
+use crate::agent_response_message::agent_response_message;
+
+pub struct PipelineResponseStream<TResponse> {
+    pub response_rx: mpsc::UnboundedReceiver<TResponse>,
+    pub stopper_guard: RequestRegistryGuard<mpsc::UnboundedSender<()>>,
+}
+
+impl<TResponse> Stream for PipelineResponseStream<TResponse>
+where
+    TResponse: Into<JsonRpcResponse>,
+{
+    type Item = ManagementJsonRpcMessage;
+
+    fn poll_next(self: Pin<&mut Self>, context: &mut Context<'_>) -> Poll<Option<Self::Item>> {
+        let Self {
+            response_rx,
+            stopper_guard,
+        } = self.get_mut();
+
+        response_rx.poll_recv(context).map(|received_response| {
+            received_response.map(|response| {
+                agent_response_message(stopper_guard.request_id.clone(), response.into())
+            })
+        })
+    }
+}
diff --git a/paddler_agent/src/plan_embedding_batches.rs b/paddler_agent/src/plan_embedding_batches.rs
deleted file mode 100644
index a8f0b2dd9..000000000
--- a/paddler_agent/src/plan_embedding_batches.rs
+++ /dev/null
@@ -1,88 +0,0 @@
-use std::ops::Range;
-
-#[must_use]
-pub fn plan_embedding_batches(
-    token_counts: &[usize],
-    n_batch: usize,
-    max_sequences_per_batch: u16,
-) -> Vec<Range<usize>> {
-    let mut batches = Vec::new();
-    let mut start = 0usize;
-    let mut current_tokens: usize = 0;
-    let mut current_sequences: u16 = 0;
-
-    for (index, &token_count) in token_counts.iter().enumerate() {
-        let would_exceed_tokens = current_tokens + token_count > n_batch;
-        let would_exceed_sequences = current_sequences >= max_sequences_per_batch;
-
-        if (would_exceed_tokens || would_exceed_sequences) && current_sequences > 0 {
-            batches.push(start..index);
-            start = index;
-            current_tokens = 0;
-            current_sequences = 0;
-        }
-
-        current_tokens += token_count;
-        current_sequences += 1;
-    }
-
-    if current_sequences > 0 {
-        batches.push(start..token_counts.len());
-    }
-
-    batches
-}
-
-#[cfg(test)]
-mod tests {
-    use super::plan_embedding_batches;
-
-    #[test]
-    fn empty_input_yields_no_batches() {
-        let batches = plan_embedding_batches(&[], 512, 4);
-
-        assert!(batches.is_empty());
-    }
-
-    #[test]
-    fn inputs_under_both_caps_fit_in_single_batch() {
-        let batches = plan_embedding_batches(&[10, 20, 30], 512, 4);
-
-        assert_eq!(batches, vec![0..3]);
-    }
-
-    #[test]
-    fn batch_flushes_when_sequence_cap_reached() {
-        let batches = plan_embedding_batches(&[10, 10, 10, 10, 10], 512, 2);
-
-        assert_eq!(batches, vec![0..2, 2..4, 4..5]);
-    }
-
-    #[test]
-    fn batch_flushes_when_token_cap_would_be_exceeded() {
-        let batches = plan_embedding_batches(&[300, 300, 300], 512, 16);
-
-        assert_eq!(batches, vec![0..1, 1..2, 2..3]);
-    }
-
-    #[test]
-    fn oversized_single_input_gets_its_own_batch() {
-        let batches = plan_embedding_batches(&[50, 1000, 50], 512, 16);
-
-        assert_eq!(batches, vec![0..1, 1..2, 2..3]);
-    }
-
-    #[test]
-    fn batch_exactly_at_sequence_cap_starts_new_batch_on_next_input() {
-        let batches = plan_embedding_batches(&[10, 10, 10, 10], 512, 4);
-
-        assert_eq!(batches, vec![0..4]);
-    }
-
-    #[test]
-    fn input_count_three_times_slot_cap_splits_into_three_batches() {
-        let batches = plan_embedding_batches(&[10; 12], 512, 4);
-
-        assert_eq!(batches, vec![0..4, 4..8, 8..12]);
-    }
-}
diff --git a/paddler_agent/src/preparation_outcome.rs b/paddler_agent/src/preparation_outcome.rs
new file mode 100644
index 000000000..800d0f223
--- /dev/null
+++ b/paddler_agent/src/preparation_outcome.rs
@@ -0,0 +1,12 @@
+use crate::agent_applicable_state::AgentApplicableState;
+use crate::agent_error::AgentError;
+use crate::prepared_serving::PreparedServing;
+
+pub enum PreparationOutcome {
+    Cancelled,
+    Failed {
+        agent_applicable_state: AgentApplicableState,
+        error: AgentError,
+    },
+    Prepared(PreparedServing),
+}
diff --git a/paddler_agent/src/prepared_decisions.rs b/paddler_agent/src/prepared_decisions.rs
new file mode 100644
index 000000000..b6689bfba
--- /dev/null
+++ b/paddler_agent/src/prepared_decisions.rs
@@ -0,0 +1,6 @@
+use paddler_agent_decision::loaded_pointer_head::LoadedPointerHead;
+
+pub enum PreparedDecisions {
+    Disabled,
+    Enabled(Box<LoadedPointerHead>),
+}
diff --git a/paddler_agent/src/prepared_embedding_batch_request.rs b/paddler_agent/src/prepared_embedding_batch_request.rs
deleted file mode 100644
index eb3c42727..000000000
--- a/paddler_agent/src/prepared_embedding_batch_request.rs
+++ /dev/null
@@ -1,17 +0,0 @@
-use tokio::sync::mpsc;
-
-use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
-use paddler_messaging::embedding_result::EmbeddingResult;
-use paddler_messaging::oversized_embedding_document_details::OversizedEmbeddingDocumentDetails;
-
-use crate::embedding_input_tokenized::EmbeddingInputTokenized;
-use crate::slot_guard::SlotGuard;
-
-pub struct PreparedEmbeddingBatchRequest {
-    pub generate_embedding_stop_rx: mpsc::UnboundedReceiver<()>,
-    pub generated_embedding_tx: mpsc::UnboundedSender<EmbeddingResult>,
-    pub inputs: Vec<EmbeddingInputTokenized>,
-    pub normalization_method: EmbeddingNormalizationMethod,
-    pub oversized_documents: Vec<OversizedEmbeddingDocumentDetails>,
-    pub slot_guard: SlotGuard,
-}
diff --git a/paddler_agent/src/prepared_generation_request.rs b/paddler_agent/src/prepared_generation_request.rs
deleted file mode 100644
index bec20fedc..000000000
--- a/paddler_agent/src/prepared_generation_request.rs
+++ /dev/null
@@ -1,22 +0,0 @@
-use std::num::NonZeroU32;
-
-use tokio::sync::mpsc;
-
-use paddler_messaging::generated_token_result::GeneratedTokenResult;
-
-use crate::prepared_prompt::PreparedPrompt;
-use crate::slot_guard::SlotGuard;
-use crate::token_classification::TokenClassification;
-use crate::token_sampling::TokenSampling;
-use crate::tool_call_handling::ToolCallHandling;
-
-pub struct PreparedGenerationRequest {
-    pub generate_tokens_stop_rx: mpsc::UnboundedReceiver<()>,
-    pub generated_tokens_tx: mpsc::UnboundedSender<GeneratedTokenResult>,
-    pub max_tokens: NonZeroU32,
-    pub prompt: PreparedPrompt,
-    pub slot_guard: SlotGuard,
-    pub token_classification: TokenClassification,
-    pub token_sampling: TokenSampling,
-    pub tool_call_handling: ToolCallHandling,
-}
diff --git a/paddler_agent/src/prepared_serving.rs b/paddler_agent/src/prepared_serving.rs
new file mode 100644
index 000000000..ba8ca9655
--- /dev/null
+++ b/paddler_agent/src/prepared_serving.rs
@@ -0,0 +1,7 @@
+use crate::agent_applicable_state::AgentApplicableState;
+use crate::prepared_decisions::PreparedDecisions;
+
+pub struct PreparedServing {
+    pub agent_applicable_state: AgentApplicableState,
+    pub prepared_decisions: PreparedDecisions,
+}
diff --git a/paddler_agent/src/prompt_tokenizer.rs b/paddler_agent/src/prompt_tokenizer.rs
deleted file mode 100644
index 5dff1e64f..000000000
--- a/paddler_agent/src/prompt_tokenizer.rs
+++ /dev/null
@@ -1,47 +0,0 @@
-use std::sync::Arc;
-
-use llama_cpp_bindings::model::AddBos;
-use llama_cpp_bindings::model::LlamaModel;
-use llama_cpp_bindings::token::LlamaToken;
-
-use crate::generation_request_rejection::GenerationRequestRejection;
-use crate::model_bos_token::ModelBosToken;
-use crate::require_prompt_fits_sequence_context::require_prompt_fits_sequence_context;
-
-pub struct PromptTokenizer {
-    pub model: Arc<LlamaModel>,
-    pub sequence_context_size: u32,
-}
-
-impl PromptTokenizer {
-    pub fn renders_its_own_bos_token(
-        &self,
-        prompt: &str,
-    ) -> Result<bool, GenerationRequestRejection> {
-        self.tokens_with_special_tokens(prompt)
-            .map(|prompt_tokens| {
-                ModelBosToken::of(&self.model).is_duplicated_at_start_of(&prompt_tokens)
-            })
-    }
-
-    pub fn tokenize(&self, prompt: &str) -> Result<Vec<LlamaToken>, GenerationRequestRejection> {
-        let mut prompt_tokens = self.tokens_with_special_tokens(prompt)?;
-
-        if ModelBosToken::of(&self.model).is_duplicated_at_start_of(&prompt_tokens) {
-            prompt_tokens.remove(0);
-        }
-
-        require_prompt_fits_sequence_context(prompt_tokens.len(), self.sequence_context_size)?;
-
-        Ok(prompt_tokens)
-    }
-
-    fn tokens_with_special_tokens(
-        &self,
-        prompt: &str,
-    ) -> Result<Vec<LlamaToken>, GenerationRequestRejection> {
-        self.model
-            .str_to_token(prompt, AddBos::Always)
-            .map_err(GenerationRequestRejection::PromptTokenizationFailed)
-    }
-}
diff --git a/paddler_agent/src/reconciliation_service.rs b/paddler_agent/src/reconciliation_service.rs
index a37bc7b9e..5f5dc8ab2 100644
--- a/paddler_agent/src/reconciliation_service.rs
+++ b/paddler_agent/src/reconciliation_service.rs
@@ -1,67 +1,94 @@
+use std::sync::Arc;
+
 use anyhow::Result;
 use async_trait::async_trait;
 use log::error;
 use tokio::select;
-use tokio::sync::mpsc;
-use tokio::time::Duration;
-use tokio::time::MissedTickBehavior;
-use tokio::time::interval;
+use tokio::sync::watch;
 use tokio_util::sync::CancellationToken;
 use trzcina::Service;
 
 use paddler_messaging::agent_desired_state::AgentDesiredState;
 
+use crate::agent_applicable_state_holder::AgentApplicableStateHolder;
 use crate::desired_state_reconciler::DesiredStateReconciler;
 use crate::desired_state_reconciliation::DesiredStateReconciliation;
-
-const DESIRED_STATE_CONVERSION_RETRY_INTERVAL: Duration = Duration::from_secs(1);
+use crate::desired_state_reconciliation_attempt::DesiredStateReconciliationAttempt;
+use crate::retry_delay::RetryDelay;
 
 pub struct ReconciliationService {
-    pub agent_desired_state_rx: mpsc::UnboundedReceiver<AgentDesiredState>,
+    pub agent_applicable_state_holder: Arc<AgentApplicableStateHolder>,
+    pub agent_desired_state_rx: watch::Receiver<AgentDesiredState>,
     pub desired_state_reconciler: DesiredStateReconciler,
 }
 
-#[async_trait]
-impl Service for ReconciliationService {
-    fn name(&self) -> &'static str {
-        "agent::reconciliation_service"
-    }
-
-    async fn run(self: Box<Self>, shutdown: CancellationToken) -> Result<()> {
+impl ReconciliationService {
+    async fn reconcile_desired_states_until_shutdown(
+        self: Box<Self>,
+        shutdown: CancellationToken,
+    ) -> Result<()> {
         let Self {
+            agent_applicable_state_holder,
             mut agent_desired_state_rx,
             desired_state_reconciler,
         } = *self;
+        let mut reconciliation_attempt = DesiredStateReconciliationAttempt::Idle;
 
-        let mut reconciliation = DesiredStateReconciliation::Reconciled;
-        let mut ticker = interval(DESIRED_STATE_CONVERSION_RETRY_INTERVAL);
-
-        ticker.set_missed_tick_behavior(MissedTickBehavior::Delay);
-
-        loop {
+        let service_outcome = loop {
             select! {
                 biased;
                 () = shutdown.cancelled() => break Ok(()),
-                _ = ticker.tick() => {
-                    if let DesiredStateReconciliation::Pending(agent_desired_state) = reconciliation {
-                        reconciliation = desired_state_reconciler
-                            .reconcile(&shutdown, *agent_desired_state)
-                            .await;
-                    }
-                },
-                next_agent_desired_state = agent_desired_state_rx.recv() => {
-                    let Some(agent_desired_state) = next_agent_desired_state else {
+                desired_state_change = agent_desired_state_rx.changed() => {
+                    if desired_state_change.is_err() {
                         error!("Agent desired state channel closed, stopping reconciliation service.");
 
                         break Ok(());
-                    };
+                    }
 
-                    reconciliation = desired_state_reconciler
-                        .reconcile(&shutdown, agent_desired_state)
-                        .await;
+                    reconciliation_attempt.abandon().await;
+                    desired_state_reconciler
+                        .slot_aggregated_status
+                        .begin_desired_state();
+                    reconciliation_attempt = DesiredStateReconciliationAttempt::start(
+                        desired_state_reconciler.clone(),
+                        agent_desired_state_rx.borrow_and_update().clone(),
+                        &shutdown,
+                        RetryDelay::Immediately,
+                    );
+                }
+                desired_state_reconciliation = reconciliation_attempt.finish() => {
+                    match desired_state_reconciliation {
+                        DesiredStateReconciliation::Pending(agent_desired_state) => {
+                            reconciliation_attempt = DesiredStateReconciliationAttempt::start(
+                                desired_state_reconciler.clone(),
+                                *agent_desired_state,
+                                &shutdown,
+                                RetryDelay::AfterRetryInterval,
+                            );
+                        }
+                        DesiredStateReconciliation::Reconciled(agent_applicable_state) => {
+                            agent_applicable_state_holder
+                                .set_agent_applicable_state(*agent_applicable_state);
+                        }
+                    }
                 }
             }
-        }
+        };
+
+        reconciliation_attempt.abandon().await;
+
+        service_outcome
+    }
+}
+
+#[async_trait]
+impl Service for ReconciliationService {
+    fn name(&self) -> &'static str {
+        "agent::reconciliation_service"
+    }
+
+    async fn run(self: Box<Self>, shutdown: CancellationToken) -> Result<()> {
+        self.reconcile_desired_states_until_shutdown(shutdown).await
     }
 }
 
@@ -69,85 +96,93 @@ impl Service for ReconciliationService {
 mod tests {
     use std::sync::Arc;
 
+    use http::StatusCode;
+    use tempfile::TempDir;
+    use tokio::spawn;
+    use tokio::sync::watch;
+    use tokio::time::Instant;
     use tokio_util::sync::CancellationToken;
+    use trzcina::Service as _;
 
     use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
-    use paddler_inference_parameters::inference_parameters::InferenceParameters;
+    use paddler_cache_dir::cache_dir::CacheDir;
+    use paddler_download_manager::download_manager::DownloadManager;
+    use paddler_local_http_fixture::fixture_response::FixtureResponse;
+    use paddler_local_http_fixture::local_http_fixture::LocalHttpFixture;
     use paddler_messaging::agent_desired_model::AgentDesiredModel;
     use paddler_messaging::agent_desired_state::AgentDesiredState;
-    use paddler_messaging::huggingface_model_reference::HuggingFaceModelReference;
+    use paddler_model_source::desired_model_resolver::DesiredModelResolver;
+    use paddler_model_source::huggingface_cache_location::HuggingFaceCacheLocation;
+    use paddler_model_source::huggingface_default_endpoint::HUGGINGFACE_DEFAULT_ENDPOINT;
+    use paddler_model_source::huggingface_hub::HuggingFaceHub;
 
-    use crate::agent_applicable_state::AgentApplicableState;
+    use super::ReconciliationService;
     use crate::agent_applicable_state_holder::AgentApplicableStateHolder;
     use crate::desired_state_reconciler::DesiredStateReconciler;
-    use crate::desired_state_reconciliation::DesiredStateReconciliation;
 
     fn desired_state_with_model(model: AgentDesiredModel) -> AgentDesiredState {
         AgentDesiredState {
-            chat_template_override: None,
-            inference_parameters: InferenceParameters::default(),
             model,
-            multimodal_projection: AgentDesiredModel::None,
+            ..AgentDesiredState::default()
         }
     }
 
-    fn desired_state_reconciler() -> DesiredStateReconciler {
-        DesiredStateReconciler {
-            agent_applicable_state_holder: Arc::new(AgentApplicableStateHolder::default()),
-            slot_aggregated_status: Arc::new(SlotAggregatedStatus::new(1)),
-        }
-    }
-
-    #[tokio::test]
-    async fn a_missing_local_model_leaves_the_state_unconverted() {
-        let reconciler = desired_state_reconciler();
-        let desired_state = desired_state_with_model(AgentDesiredModel::LocalToAgent(
-            "/paddler-nonexistent-model-for-reconciliation.gguf".to_owned(),
-        ));
-
-        let reconciliation = reconciler
-            .reconcile(&CancellationToken::new(), desired_state.clone())
-            .await;
-
-        assert_eq!(
-            reconciliation,
-            DesiredStateReconciliation::Pending(Box::new(desired_state))
-        );
-        assert_eq!(
-            reconciler
-                .agent_applicable_state_holder
-                .get_agent_applicable_state(),
-            AgentApplicableState::default()
+    #[tokio::test(start_paused = true)]
+    async fn a_desired_state_arriving_during_retries_is_reconciled_without_waiting() {
+        let cache_dir = TempDir::new().unwrap();
+        let failing_model_server =
+            LocalHttpFixture::start(FixtureResponse::Status(StatusCode::INTERNAL_SERVER_ERROR))
+                .await
+                .unwrap();
+        let agent_applicable_state_holder = Arc::new(AgentApplicableStateHolder::default());
+        let mut agent_applicable_state_rx = agent_applicable_state_holder.subscribe();
+        let (agent_desired_state_tx, agent_desired_state_rx) =
+            watch::channel(AgentDesiredState::default());
+        let shutdown = CancellationToken::new();
+        let reconciliation = spawn(
+            Box::new(ReconciliationService {
+                agent_applicable_state_holder: agent_applicable_state_holder.clone(),
+                agent_desired_state_rx,
+                desired_state_reconciler: DesiredStateReconciler {
+                    desired_model_resolver: DesiredModelResolver {
+                        cache_dir: CacheDir {
+                            explicit: Some(cache_dir.path().to_path_buf()),
+                            home: None,
+                            xdg: None,
+                        },
+                        download_manager: DownloadManager::new().unwrap(),
+                        huggingface_client: HuggingFaceHub {
+                            cache_location: HuggingFaceCacheLocation::HubDefault,
+                            endpoint: HUGGINGFACE_DEFAULT_ENDPOINT.to_owned(),
+                        }
+                        .client()
+                        .unwrap(),
+                    },
+                    slot_aggregated_status: Arc::new(SlotAggregatedStatus::new(1)),
+                },
+            })
+            .run(shutdown.clone()),
         );
-    }
 
-    #[tokio::test]
-    async fn a_model_download_cancelled_mid_conversion_leaves_the_state_unconverted() {
-        let cancellation_token = CancellationToken::new();
+        agent_desired_state_tx.send_replace(desired_state_with_model(AgentDesiredModel::Uri(
+            failing_model_server.url("/model.gguf"),
+        )));
+        failing_model_server.wait_for_requests(2).await.unwrap();
 
-        cancellation_token.cancel();
+        let replaced_at = Instant::now();
 
-        let reconciler = desired_state_reconciler();
-        let desired_state =
-            desired_state_with_model(AgentDesiredModel::HuggingFace(HuggingFaceModelReference {
-                filename: "model.gguf".to_owned(),
-                repo_id: "paddler-tests/never-downloaded".to_owned(),
-                revision: "main".to_owned(),
-            }));
+        agent_desired_state_tx.send_replace(AgentDesiredState::default());
+        agent_applicable_state_rx.changed().await.unwrap();
 
-        let reconciliation = reconciler
-            .reconcile(&cancellation_token, desired_state.clone())
-            .await;
+        assert_eq!(Instant::now(), replaced_at);
+        assert_eq!(failing_model_server.request_count(), 2);
 
-        assert_eq!(
-            reconciliation,
-            DesiredStateReconciliation::Pending(Box::new(desired_state))
-        );
-        assert_eq!(
-            reconciler
-                .agent_applicable_state_holder
-                .get_agent_applicable_state(),
-            AgentApplicableState::default()
-        );
+        shutdown.cancel();
+        reconciliation.await.unwrap().unwrap();
+        failing_model_server
+            .stop()
+            .await
+            .expect("the fixture task must not panic")
+            .expect("the fixture must serve every request");
     }
 }
diff --git a/paddler_agent/src/replacement_outcome.rs b/paddler_agent/src/replacement_outcome.rs
new file mode 100644
index 000000000..85c581731
--- /dev/null
+++ b/paddler_agent/src/replacement_outcome.rs
@@ -0,0 +1,34 @@
+use paddler_agent_runtime::shuts_down::ShutsDown;
+
+use crate::agent_error::AgentError;
+use crate::model_serving::ModelServing;
+use crate::served_model::ServedModel;
+
+pub enum ReplacementOutcome {
+    Cancelled {
+        held: ModelServing,
+    },
+    Failed {
+        error: AgentError,
+        held: ModelServing,
+    },
+    NothingToServe,
+    Serving(Box<ServedModel>),
+}
+
+impl ReplacementOutcome {
+    #[must_use]
+    pub fn into_model_serving(self) -> ModelServing {
+        match self {
+            Self::Cancelled { held } | Self::Failed { held, .. } => held,
+            Self::NothingToServe => ModelServing::Idle,
+            Self::Serving(served_model) => ModelServing::Serving(served_model),
+        }
+    }
+}
+
+impl ShutsDown for ReplacementOutcome {
+    async fn shut_down(self) {
+        self.into_model_serving().shut_down().await;
+    }
+}
diff --git a/paddler_agent/src/require_embeddings_enabled.rs b/paddler_agent/src/require_embeddings_enabled.rs
deleted file mode 100644
index 2328b08f8..000000000
--- a/paddler_agent/src/require_embeddings_enabled.rs
+++ /dev/null
@@ -1,32 +0,0 @@
-use crate::embedding_batch_rejection::EmbeddingBatchRejection;
-
-pub const fn require_embeddings_enabled(
-    enable_embeddings: bool,
-) -> Result<(), EmbeddingBatchRejection> {
-    if enable_embeddings {
-        Ok(())
-    } else {
-        Err(EmbeddingBatchRejection::EmbeddingsDisabled)
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use std::mem::discriminant;
-
-    use super::require_embeddings_enabled;
-    use crate::embedding_batch_rejection::EmbeddingBatchRejection;
-
-    #[test]
-    fn accepts_batches_when_embeddings_are_enabled() {
-        assert!(require_embeddings_enabled(true).is_ok());
-    }
-
-    #[test]
-    fn rejects_batches_when_embeddings_are_disabled() {
-        assert!(
-            require_embeddings_enabled(false).is_err_and(|rejection| discriminant(&rejection)
-                == discriminant(&EmbeddingBatchRejection::EmbeddingsDisabled))
-        );
-    }
-}
diff --git a/paddler_agent/src/resolve_grammar.rs b/paddler_agent/src/resolve_grammar.rs
deleted file mode 100644
index a997a7c05..000000000
--- a/paddler_agent/src/resolve_grammar.rs
+++ /dev/null
@@ -1,32 +0,0 @@
-use llama_cpp_bindings::json_schema_to_grammar;
-use llama_cpp_bindings::model::LlamaModel;
-use llama_cpp_bindings::sampling::LlamaSampler;
-
-use paddler_messaging::grammar_constraint::GrammarConstraint;
-
-use crate::generation_request_rejection::GenerationRequestRejection;
-use crate::grammar_sampling::GrammarSampling;
-
-const JSON_SCHEMA_GRAMMAR_ROOT: &str = "root";
-
-pub fn resolve_grammar(
-    grammar: Option<GrammarConstraint>,
-    model: &LlamaModel,
-) -> Result<GrammarSampling, GenerationRequestRejection> {
-    match grammar {
-        None => Ok(GrammarSampling::Unconstrained),
-        Some(GrammarConstraint::Gbnf { grammar, root }) => {
-            LlamaSampler::grammar(model, &grammar, &root)
-                .map(GrammarSampling::Constrained)
-                .map_err(GenerationRequestRejection::for_gbnf_grammar_error)
-        }
-        Some(GrammarConstraint::JsonSchema { schema }) => {
-            let grammar = json_schema_to_grammar(&schema)
-                .map_err(GenerationRequestRejection::GrammarConversionFailed)?;
-
-            LlamaSampler::grammar(model, &grammar, JSON_SCHEMA_GRAMMAR_ROOT)
-                .map(GrammarSampling::Constrained)
-                .map_err(GenerationRequestRejection::GrammarSamplerInitializationFailed)
-        }
-    }
-}
diff --git a/paddler_agent/src/retained_serving.rs b/paddler_agent/src/retained_serving.rs
new file mode 100644
index 000000000..769f8b01e
--- /dev/null
+++ b/paddler_agent/src/retained_serving.rs
@@ -0,0 +1,7 @@
+use crate::loaded_model_serving_plan::LoadedModelServingPlan;
+use crate::model_serving_settings::ModelServingSettings;
+
+pub enum RetainedServing {
+    LoadedModel(LoadedModelServingPlan),
+    ModelToLoad(ModelServingSettings),
+}
diff --git a/paddler_agent/src/retry_delay.rs b/paddler_agent/src/retry_delay.rs
new file mode 100644
index 000000000..420e71bfe
--- /dev/null
+++ b/paddler_agent/src/retry_delay.rs
@@ -0,0 +1,32 @@
+use tokio::time::Duration;
+use tokio::time::sleep;
+use tokio_util::sync::CancellationToken;
+
+use crate::retry_delay_outcome::RetryDelayOutcome;
+
+const RETRY_INTERVAL: Duration = Duration::from_secs(1);
+
+#[derive(Clone, Copy)]
+pub enum RetryDelay {
+    AfterRetryInterval,
+    Immediately,
+}
+
+impl RetryDelay {
+    pub async fn elapse(self, cancellation_token: &CancellationToken) -> RetryDelayOutcome {
+        match self {
+            Self::AfterRetryInterval => {
+                cancellation_token
+                    .run_until_cancelled(sleep(RETRY_INTERVAL))
+                    .await;
+            }
+            Self::Immediately => {}
+        }
+
+        if cancellation_token.is_cancelled() {
+            RetryDelayOutcome::Cancelled
+        } else {
+            RetryDelayOutcome::Elapsed
+        }
+    }
+}
diff --git a/paddler_agent/src/retry_delay_outcome.rs b/paddler_agent/src/retry_delay_outcome.rs
new file mode 100644
index 000000000..04b6553a3
--- /dev/null
+++ b/paddler_agent/src/retry_delay_outcome.rs
@@ -0,0 +1,5 @@
+#[derive(Debug, Eq, PartialEq)]
+pub enum RetryDelayOutcome {
+    Cancelled,
+    Elapsed,
+}
diff --git a/paddler_agent/src/sampler_chain_factory.rs b/paddler_agent/src/sampler_chain_factory.rs
deleted file mode 100644
index f43c1d598..000000000
--- a/paddler_agent/src/sampler_chain_factory.rs
+++ /dev/null
@@ -1,68 +0,0 @@
-use llama_cpp_bindings::error::SamplingError;
-use llama_cpp_bindings::sampling::LlamaSampler;
-
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
-
-pub struct SamplerChainFactory {
-    pub inference_parameters: InferenceParameters,
-    pub n_vocab: i32,
-}
-
-impl SamplerChainFactory {
-    pub fn create(&self, seed: u32) -> Result<LlamaSampler, SamplingError> {
-        let inference_parameters = &self.inference_parameters;
-        let samplers = [
-            LlamaSampler::penalties(
-                self.n_vocab,
-                inference_parameters.penalty_last_n,
-                inference_parameters.penalty_repeat,
-                inference_parameters.penalty_frequency,
-                inference_parameters.penalty_presence,
-            ),
-            LlamaSampler::top_k(inference_parameters.top_k),
-            LlamaSampler::top_p(inference_parameters.top_p, 0),
-            LlamaSampler::min_p(inference_parameters.min_p, 0),
-            LlamaSampler::temp(inference_parameters.temperature),
-            LlamaSampler::dist(seed),
-        ];
-
-        LlamaSampler::chain_simple(samplers.into_iter().collect::<Result<Vec<_>, _>>()?)
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use llama_cpp_bindings::token::LlamaToken;
-    use llama_cpp_bindings::token::data::LlamaTokenData;
-    use llama_cpp_bindings::token::data_array::LlamaTokenDataArray;
-
-    use paddler_inference_parameters::inference_parameters::InferenceParameters;
-
-    use super::SamplerChainFactory;
-
-    #[test]
-    fn chain_honors_configured_top_k() {
-        let inference_parameters = InferenceParameters {
-            top_k: 1,
-            ..InferenceParameters::default()
-        };
-        let sampler_chain = SamplerChainFactory {
-            inference_parameters,
-            n_vocab: 3,
-        }
-        .create(7)
-        .unwrap();
-        let mut candidates = LlamaTokenDataArray::from_iter(
-            [
-                LlamaTokenData::new(LlamaToken(0), 0.5, 0.0),
-                LlamaTokenData::new(LlamaToken(1), 3.0, 0.0),
-                LlamaTokenData::new(LlamaToken(2), 1.0, 0.0),
-            ],
-            false,
-        );
-
-        candidates.apply_sampler(&sampler_chain).unwrap();
-
-        assert_eq!(candidates.selected_token(), Some(LlamaToken(1)));
-    }
-}
diff --git a/paddler_agent/src/send_startup_signal.rs b/paddler_agent/src/send_startup_signal.rs
deleted file mode 100644
index 83693b108..000000000
--- a/paddler_agent/src/send_startup_signal.rs
+++ /dev/null
@@ -1,51 +0,0 @@
-use log::debug;
-use tokio::sync::oneshot;
-
-use crate::startup_signal_delivery::StartupSignalDelivery;
-
-pub fn send_startup_signal<TSignal>(
-    startup_signal_tx: oneshot::Sender<TSignal>,
-    signal: TSignal,
-) -> StartupSignalDelivery {
-    match startup_signal_tx.send(signal) {
-        Ok(()) => StartupSignalDelivery::Delivered,
-        Err(_undelivered_signal) => {
-            debug!("The arbiter abandoned the scheduler startup; stopping the scheduler thread");
-
-            StartupSignalDelivery::Abandoned
-        }
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use std::mem::discriminant;
-
-    use tokio::sync::oneshot;
-
-    use super::send_startup_signal;
-    use crate::startup_signal_delivery::StartupSignalDelivery;
-
-    #[test]
-    fn delivers_the_signal_to_a_live_receiver() {
-        let (startup_signal_tx, startup_signal_rx) = oneshot::channel::<()>();
-
-        assert_eq!(
-            discriminant(&send_startup_signal(startup_signal_tx, ())),
-            discriminant(&StartupSignalDelivery::Delivered)
-        );
-        assert_eq!(startup_signal_rx.blocking_recv(), Ok(()));
-    }
-
-    #[test]
-    fn reports_an_abandoned_startup_when_the_receiver_was_dropped() {
-        let (startup_signal_tx, startup_signal_rx) = oneshot::channel::<()>();
-
-        drop(startup_signal_rx);
-
-        assert_eq!(
-            discriminant(&send_startup_signal(startup_signal_tx, ())),
-            discriminant(&StartupSignalDelivery::Abandoned)
-        );
-    }
-}
diff --git a/paddler_agent/src/served_capabilities.rs b/paddler_agent/src/served_capabilities.rs
new file mode 100644
index 000000000..ae57dc4cd
--- /dev/null
+++ b/paddler_agent/src/served_capabilities.rs
@@ -0,0 +1,7 @@
+use crate::inference_capabilities::InferenceCapabilities;
+use crate::served_decisions::ServedDecisions;
+use crate::served_embeddings::ServedEmbeddings;
+use crate::served_text_generation::ServedTextGeneration;
+
+pub type ServedCapabilities =
+    InferenceCapabilities<ServedDecisions, ServedEmbeddings, ServedTextGeneration>;
diff --git a/paddler_agent/src/served_decisions.rs b/paddler_agent/src/served_decisions.rs
new file mode 100644
index 000000000..e026d1199
--- /dev/null
+++ b/paddler_agent/src/served_decisions.rs
@@ -0,0 +1,37 @@
+use paddler_agent_decision::decision_error::DecisionError;
+use paddler_agent_decision::decision_request_preparer::DecisionRequestPreparer;
+use paddler_agent_decision::reports_decision_outcome::ReportsDecisionOutcome as _;
+use paddler_agent_runtime::agent_request::AgentRequest;
+use paddler_agent_runtime::scheduler_handle::SchedulerHandle;
+use paddler_agent_runtime::shuts_down::ShutsDown;
+use paddler_messaging::decision_result::DecisionResult;
+use paddler_messaging::request_params::decide_params::DecideParams;
+
+pub enum ServedDecisions {
+    Disabled,
+    Enabled(SchedulerHandle<DecisionRequestPreparer>),
+}
+
+impl ServedDecisions {
+    pub fn forward(
+        &self,
+        agent_name: Option<&str>,
+        request: AgentRequest<DecideParams, DecisionResult>,
+    ) {
+        match self {
+            Self::Disabled => {
+                DecisionError::DecisionsNotServed.report(agent_name, &request.response_tx);
+            }
+            Self::Enabled(scheduler_handle) => scheduler_handle.request_preparer.prepare(request),
+        }
+    }
+}
+
+impl ShutsDown for ServedDecisions {
+    async fn shut_down(self) {
+        match self {
+            Self::Disabled => {}
+            Self::Enabled(scheduler_handle) => scheduler_handle.shut_down().await,
+        }
+    }
+}
diff --git a/paddler_agent/src/served_embeddings.rs b/paddler_agent/src/served_embeddings.rs
new file mode 100644
index 000000000..58a400bdc
--- /dev/null
+++ b/paddler_agent/src/served_embeddings.rs
@@ -0,0 +1,4 @@
+use paddler_agent_embeddings::embedding_batch_preparer::EmbeddingBatchPreparer;
+use paddler_agent_runtime::scheduler_handle::SchedulerHandle;
+
+pub type ServedEmbeddings = SchedulerHandle<EmbeddingBatchPreparer>;
diff --git a/paddler_agent/src/served_model.rs b/paddler_agent/src/served_model.rs
new file mode 100644
index 000000000..f59a79ffb
--- /dev/null
+++ b/paddler_agent/src/served_model.rs
@@ -0,0 +1,160 @@
+use tokio::join;
+
+use paddler_agent_runtime::loaded_llama_model::LoadedLlamaModel;
+use paddler_agent_runtime::shuts_down::ShutsDown;
+use paddler_agent_status::serving_report::ServingReport;
+use paddler_agent_text_generation::text_generation_request::TextGenerationRequest;
+
+use crate::capability_plan::CapabilityPlan;
+use crate::inference_capabilities::InferenceCapabilities;
+use crate::loaded_model_serving_plan::LoadedModelServingPlan;
+use crate::model_serving_settings::ModelServingSettings;
+use crate::pipeline_request::PipelineRequest;
+use crate::retained_serving::RetainedServing;
+use crate::served_capabilities::ServedCapabilities;
+
+pub struct ServedModel {
+    pub capabilities: ServedCapabilities,
+    pub loaded_llama_model: LoadedLlamaModel,
+    pub model_serving_settings: ModelServingSettings,
+}
+
+impl ServedModel {
+    pub fn forward(&self, agent_name: Option<&str>, request: PipelineRequest) {
+        match request {
+            PipelineRequest::ContinueFromConversationHistory(request) => {
+                self.capabilities.text_generation.request_preparer.prepare(
+                    TextGenerationRequest::ContinueFromConversationHistory(request),
+                );
+            }
+            PipelineRequest::ContinueFromRawPrompt(request) => {
+                self.capabilities
+                    .text_generation
+                    .request_preparer
+                    .prepare(TextGenerationRequest::ContinueFromRawPrompt(request));
+            }
+            PipelineRequest::Decide(request) => {
+                self.capabilities.decisions.forward(agent_name, request);
+            }
+            PipelineRequest::GenerateEmbeddingBatch(request) => {
+                self.capabilities
+                    .embeddings
+                    .request_preparer
+                    .prepare(request);
+            }
+        }
+    }
+
+    #[must_use]
+    pub fn keeps_decisions_of(&self, desired_settings: &ModelServingSettings) -> bool {
+        self.serves_the_model_of(desired_settings)
+            && self.runs_the_runtime_parameters_of(desired_settings)
+            && self.model_serving_settings.capabilities.decisions
+                == desired_settings.capabilities.decisions
+    }
+
+    pub async fn retain_for(self, desired_settings: ModelServingSettings) -> RetainedServing {
+        let serves_the_desired_model = self.serves_the_model_of(&desired_settings);
+        let runs_the_desired_runtime_parameters =
+            self.runs_the_runtime_parameters_of(&desired_settings);
+        let Self {
+            capabilities,
+            loaded_llama_model,
+            model_serving_settings: served_settings,
+        } = self;
+
+        if !serves_the_desired_model {
+            capabilities.shut_down().await;
+            loaded_llama_model.release().await;
+
+            return RetainedServing::ModelToLoad(desired_settings);
+        }
+
+        if !runs_the_desired_runtime_parameters {
+            capabilities.shut_down().await;
+
+            return RetainedServing::LoadedModel(
+                LoadedModelServingPlan::starting_every_capability(
+                    loaded_llama_model,
+                    desired_settings,
+                ),
+            );
+        }
+
+        let InferenceCapabilities {
+            decisions,
+            embeddings,
+            text_generation,
+        } = capabilities;
+
+        let (decisions_plan, embeddings_plan, text_generation_plan) = join!(
+            CapabilityPlan::retaining(
+                decisions,
+                &served_settings.capabilities.decisions,
+                &desired_settings.capabilities.decisions,
+            ),
+            CapabilityPlan::retaining(
+                embeddings,
+                &served_settings.capabilities.embeddings,
+                &desired_settings.capabilities.embeddings,
+            ),
+            CapabilityPlan::retaining(
+                text_generation,
+                &served_settings.capabilities.text_generation,
+                &desired_settings.capabilities.text_generation,
+            ),
+        );
+
+        RetainedServing::LoadedModel(LoadedModelServingPlan {
+            capability_plans: InferenceCapabilities {
+                decisions: decisions_plan,
+                embeddings: embeddings_plan,
+                text_generation: text_generation_plan,
+            },
+            loaded_llama_model,
+            model_serving_settings: desired_settings,
+        })
+    }
+
+    fn serves_the_model_of(&self, desired_settings: &ModelServingSettings) -> bool {
+        self.loaded_llama_model.llama_model_settings == desired_settings.llama_model_settings()
+    }
+
+    fn runs_the_runtime_parameters_of(&self, desired_settings: &ModelServingSettings) -> bool {
+        self.model_serving_settings.model_runtime_parameters
+            == desired_settings.model_runtime_parameters
+    }
+
+    #[must_use]
+    pub fn serving_report(&self, uses_chat_template_override: bool) -> ServingReport {
+        ServingReport {
+            decisions: self.model_serving_settings.capabilities.decisions.serving(),
+            model_path: self
+                .loaded_llama_model
+                .llama_model_settings
+                .reported_model_path()
+                .model_path,
+            serving_issues: self
+                .capabilities
+                .text_generation
+                .request_preparer
+                .preparation
+                .conversation_support
+                .serving_issues(),
+            uses_chat_template_override,
+        }
+    }
+}
+
+impl ShutsDown for ServedModel {
+    async fn shut_down(self) {
+        let Self {
+            capabilities,
+            loaded_llama_model,
+            ..
+        } = self;
+
+        capabilities.shut_down().await;
+        loaded_llama_model.release().await;
+    }
+}
diff --git a/paddler_agent/src/served_text_generation.rs b/paddler_agent/src/served_text_generation.rs
new file mode 100644
index 000000000..a2d7c439d
--- /dev/null
+++ b/paddler_agent/src/served_text_generation.rs
@@ -0,0 +1,4 @@
+use paddler_agent_runtime::scheduler_handle::SchedulerHandle;
+use paddler_agent_text_generation::generation_request_preparer::GenerationRequestPreparer;
+
+pub type ServedTextGeneration = SchedulerHandle<GenerationRequestPreparer>;
diff --git a/paddler_agent/src/serving_interruption.rs b/paddler_agent/src/serving_interruption.rs
new file mode 100644
index 000000000..e2083cc66
--- /dev/null
+++ b/paddler_agent/src/serving_interruption.rs
@@ -0,0 +1,8 @@
+use paddler_agent_runtime::loaded_llama_model::LoadedLlamaModel;
+
+use crate::capability_startup_interruption::CapabilityStartupInterruption;
+
+pub struct ServingInterruption {
+    pub interruption: CapabilityStartupInterruption,
+    pub loaded_llama_model: LoadedLlamaModel,
+}
diff --git a/paddler_agent/src/serving_preparation.rs b/paddler_agent/src/serving_preparation.rs
new file mode 100644
index 000000000..fbee2cc43
--- /dev/null
+++ b/paddler_agent/src/serving_preparation.rs
@@ -0,0 +1,54 @@
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent_runtime::inference_runtime_context::InferenceRuntimeContext;
+
+use crate::agent_applicable_state::AgentApplicableState;
+use crate::applicable_decisions::ApplicableDecisions;
+use crate::capability_starter::CapabilityStarter;
+use crate::capability_startup_interruption::CapabilityStartupInterruption;
+use crate::preparation_outcome::PreparationOutcome;
+use crate::prepared_serving::PreparedServing;
+use crate::retry_delay::RetryDelay;
+use crate::retry_delay_outcome::RetryDelayOutcome;
+
+pub struct ServingPreparation {
+    pub agent_applicable_state: AgentApplicableState,
+    pub decisions_to_prepare: ApplicableDecisions,
+    pub retry_delay: RetryDelay,
+}
+
+impl ServingPreparation {
+    pub async fn prepare(
+        self,
+        cancellation_token: CancellationToken,
+        inference_runtime_context: InferenceRuntimeContext,
+    ) -> PreparationOutcome {
+        let Self {
+            agent_applicable_state,
+            decisions_to_prepare,
+            retry_delay,
+        } = self;
+
+        if retry_delay.elapse(&cancellation_token).await == RetryDelayOutcome::Cancelled {
+            return PreparationOutcome::Cancelled;
+        }
+
+        match (CapabilityStarter {
+            cancellation_token: &cancellation_token,
+            inference_runtime_context: &inference_runtime_context,
+        })
+        .prepare_decisions(&decisions_to_prepare)
+        .await
+        {
+            Ok(prepared_decisions) => PreparationOutcome::Prepared(PreparedServing {
+                agent_applicable_state,
+                prepared_decisions,
+            }),
+            Err(CapabilityStartupInterruption::Cancelled) => PreparationOutcome::Cancelled,
+            Err(CapabilityStartupInterruption::Failed(error)) => PreparationOutcome::Failed {
+                agent_applicable_state,
+                error,
+            },
+        }
+    }
+}
diff --git a/paddler_agent/src/serving_replacement.rs b/paddler_agent/src/serving_replacement.rs
new file mode 100644
index 000000000..19a9ed69f
--- /dev/null
+++ b/paddler_agent/src/serving_replacement.rs
@@ -0,0 +1,117 @@
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent_runtime::inference_runtime_context::InferenceRuntimeContext;
+use paddler_agent_runtime::shuts_down::ShutsDown as _;
+
+use crate::agent_applicable_state::AgentApplicableState;
+use crate::capability_starter::CapabilityStarter;
+use crate::capability_startup_interruption::CapabilityStartupInterruption;
+use crate::finished_serving_startup::FinishedServingStartup;
+use crate::loaded_model_serving_plan::LoadedModelServingPlan;
+use crate::model_serving::ModelServing;
+use crate::model_serving_settings::ModelServingSettings;
+use crate::prepared_decisions::PreparedDecisions;
+use crate::prepared_serving::PreparedServing;
+use crate::replacement_outcome::ReplacementOutcome;
+use crate::retained_serving::RetainedServing;
+use crate::serving_interruption::ServingInterruption;
+
+pub struct ServingReplacement {
+    pub held: ModelServing,
+    pub prepared_serving: PreparedServing,
+}
+
+impl ServingReplacement {
+    pub async fn replace(
+        self,
+        cancellation_token: CancellationToken,
+        inference_runtime_context: InferenceRuntimeContext,
+    ) -> FinishedServingStartup {
+        let Self {
+            held,
+            prepared_serving:
+                PreparedServing {
+                    agent_applicable_state,
+                    prepared_decisions,
+                },
+        } = self;
+        let capability_starter = CapabilityStarter {
+            cancellation_token: &cancellation_token,
+            inference_runtime_context: &inference_runtime_context,
+        };
+        let outcome = match &agent_applicable_state {
+            AgentApplicableState::NotConfigured | AgentApplicableState::WithoutModel { .. } => {
+                held.shut_down().await;
+
+                ReplacementOutcome::NothingToServe
+            }
+            AgentApplicableState::WithModel(desired_settings) => {
+                Self::serve(
+                    &capability_starter,
+                    held,
+                    desired_settings.clone(),
+                    prepared_decisions,
+                )
+                .await
+            }
+        };
+
+        FinishedServingStartup {
+            agent_applicable_state,
+            outcome,
+        }
+    }
+
+    async fn serve(
+        capability_starter: &CapabilityStarter<'_>,
+        held: ModelServing,
+        desired_settings: ModelServingSettings,
+        prepared_decisions: PreparedDecisions,
+    ) -> ReplacementOutcome {
+        let loaded_model_serving_plan = match held.retain_for(desired_settings).await {
+            RetainedServing::LoadedModel(loaded_model_serving_plan) => loaded_model_serving_plan,
+            RetainedServing::ModelToLoad(desired_settings) => match capability_starter
+                .load_model(desired_settings.llama_model_settings())
+                .await
+            {
+                Ok(loaded_llama_model) => LoadedModelServingPlan::starting_every_capability(
+                    loaded_llama_model,
+                    desired_settings,
+                ),
+                Err(CapabilityStartupInterruption::Cancelled) => {
+                    return ReplacementOutcome::Cancelled {
+                        held: ModelServing::Idle,
+                    };
+                }
+                Err(CapabilityStartupInterruption::Failed(error)) => {
+                    return ReplacementOutcome::Failed {
+                        error,
+                        held: ModelServing::Idle,
+                    };
+                }
+            },
+        };
+
+        match capability_starter
+            .serve(loaded_model_serving_plan, prepared_decisions)
+            .await
+        {
+            Ok(served_model) => ReplacementOutcome::Serving(Box::new(served_model)),
+            Err(serving_interruption) => match *serving_interruption {
+                ServingInterruption {
+                    interruption: CapabilityStartupInterruption::Cancelled,
+                    loaded_llama_model,
+                } => ReplacementOutcome::Cancelled {
+                    held: ModelServing::Holding(loaded_llama_model),
+                },
+                ServingInterruption {
+                    interruption: CapabilityStartupInterruption::Failed(error),
+                    loaded_llama_model,
+                } => ReplacementOutcome::Failed {
+                    error,
+                    held: ModelServing::Holding(loaded_llama_model),
+                },
+            },
+        }
+    }
+}
diff --git a/paddler_agent/src/serving_startup.rs b/paddler_agent/src/serving_startup.rs
new file mode 100644
index 000000000..a17eb7575
--- /dev/null
+++ b/paddler_agent/src/serving_startup.rs
@@ -0,0 +1,66 @@
+use std::future::pending;
+use std::mem;
+
+use log::error;
+
+use paddler_agent_runtime::shuts_down::ShutsDown;
+
+use crate::finished_serving_startup::FinishedServingStartup;
+use crate::preparation_outcome::PreparationOutcome;
+use crate::serving_startup_progress::ServingStartupProgress;
+use crate::startup_phase::StartupPhase;
+
+pub enum ServingStartup {
+    Idle,
+    Preparing(StartupPhase<PreparationOutcome>),
+    Replacing(StartupPhase<FinishedServingStartup>),
+}
+
+impl ServingStartup {
+    pub fn cancel(&self) {
+        match self {
+            Self::Idle => {}
+            Self::Preparing(StartupPhase {
+                cancellation_token, ..
+            })
+            | Self::Replacing(StartupPhase {
+                cancellation_token, ..
+            }) => cancellation_token.cancel(),
+        }
+    }
+
+    pub async fn next_progress(&mut self) -> ServingStartupProgress {
+        let serving_startup_progress = match self {
+            Self::Idle => pending().await,
+            Self::Preparing(StartupPhase { outcome, .. }) => {
+                ServingStartupProgress::Prepared(outcome.as_mut().await)
+            }
+            Self::Replacing(StartupPhase { outcome, .. }) => {
+                ServingStartupProgress::Replaced(outcome.as_mut().await)
+            }
+        };
+
+        drop(mem::replace(self, Self::Idle));
+
+        serving_startup_progress
+    }
+}
+
+impl ShutsDown for ServingStartup {
+    async fn shut_down(self) {
+        self.cancel();
+
+        match self {
+            Self::Idle => {}
+            Self::Preparing(StartupPhase { outcome, .. }) => match outcome.await {
+                PreparationOutcome::Cancelled | PreparationOutcome::Prepared(_) => {}
+                PreparationOutcome::Failed { error, .. } => {
+                    error!("Failed to prepare the serving that the shutdown stopped: {error}");
+                }
+            },
+            Self::Replacing(StartupPhase { outcome, .. }) => {
+                outcome.await.outcome.shut_down().await;
+            }
+        }
+    }
+}
diff --git a/paddler_agent/src/serving_startup_progress.rs b/paddler_agent/src/serving_startup_progress.rs
new file mode 100644
index 000000000..d1a3593f4
--- /dev/null
+++ b/paddler_agent/src/serving_startup_progress.rs
@@ -0,0 +1,7 @@
+use crate::finished_serving_startup::FinishedServingStartup;
+use crate::preparation_outcome::PreparationOutcome;
+
+pub enum ServingStartupProgress {
+    Prepared(PreparationOutcome),
+    Replaced(FinishedServingStartup),
+}
diff --git a/paddler_agent/src/slot_guard.rs b/paddler_agent/src/slot_guard.rs
deleted file mode 100644
index fb03ab3db..000000000
--- a/paddler_agent/src/slot_guard.rs
+++ /dev/null
@@ -1,49 +0,0 @@
-use std::sync::Arc;
-
-use paddler_agent_status::dispenses_slots::DispensesSlots as _;
-use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
-
-pub struct SlotGuard {
-    slot_aggregated_status: Arc<SlotAggregatedStatus>,
-}
-
-impl SlotGuard {
-    #[must_use]
-    pub fn new(slot_aggregated_status: Arc<SlotAggregatedStatus>) -> Self {
-        slot_aggregated_status.take_slot();
-
-        Self {
-            slot_aggregated_status,
-        }
-    }
-}
-
-impl Drop for SlotGuard {
-    fn drop(&mut self) {
-        self.slot_aggregated_status.release_slot();
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use std::sync::Arc;
-
-    use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
-
-    use crate::slot_guard::SlotGuard;
-
-    #[tokio::test]
-    async fn increments_slot_on_construct_and_releases_on_drop() {
-        let slot_aggregated_status = Arc::new(SlotAggregatedStatus::new(4));
-
-        assert_eq!(slot_aggregated_status.slots_processing_count(), 0);
-
-        {
-            let _guard = SlotGuard::new(slot_aggregated_status.clone());
-
-            assert_eq!(slot_aggregated_status.slots_processing_count(), 1);
-        }
-
-        assert_eq!(slot_aggregated_status.slots_processing_count(), 0);
-    }
-}
diff --git a/paddler_agent/src/startup_phase.rs b/paddler_agent/src/startup_phase.rs
new file mode 100644
index 000000000..81cf4f27b
--- /dev/null
+++ b/paddler_agent/src/startup_phase.rs
@@ -0,0 +1,9 @@
+use std::future::Future;
+use std::pin::Pin;
+
+use tokio_util::sync::CancellationToken;
+
+pub struct StartupPhase<TOutcome> {
+    pub cancellation_token: CancellationToken,
+    pub outcome: Pin<Box<dyn Future<Output = TOutcome> + Send>>,
+}
diff --git a/paddler_agent/src/superseding_state.rs b/paddler_agent/src/superseding_state.rs
new file mode 100644
index 000000000..876205142
--- /dev/null
+++ b/paddler_agent/src/superseding_state.rs
@@ -0,0 +1,6 @@
+use crate::agent_applicable_state::AgentApplicableState;
+
+pub enum SupersedingState {
+    NotSuperseded,
+    SupersededBy(AgentApplicableState),
+}
diff --git a/paddler_agent/src/token_generation.rs b/paddler_agent/src/token_generation.rs
deleted file mode 100644
index d196525f3..000000000
--- a/paddler_agent/src/token_generation.rs
+++ /dev/null
@@ -1,39 +0,0 @@
-use crate::generation_request_rejection::GenerationRequestRejection;
-use crate::token_generation_support::TokenGenerationSupport;
-
-pub enum TokenGeneration {
-    DisabledForEmbeddings,
-    Enabled(Box<TokenGenerationSupport>),
-}
-
-impl TokenGeneration {
-    pub const fn require_enabled(
-        &self,
-    ) -> Result<&TokenGenerationSupport, GenerationRequestRejection> {
-        match self {
-            Self::DisabledForEmbeddings => Err(GenerationRequestRejection::TokenGenerationDisabled),
-            Self::Enabled(token_generation_support) => Ok(token_generation_support),
-        }
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use std::mem::discriminant;
-
-    use super::TokenGeneration;
-    use crate::generation_request_rejection::GenerationRequestRejection;
-
-    #[test]
-    fn rejects_generation_in_embeddings_mode() {
-        assert_eq!(
-            TokenGeneration::DisabledForEmbeddings
-                .require_enabled()
-                .err()
-                .map(|rejection| discriminant(&rejection)),
-            Some(discriminant(
-                &GenerationRequestRejection::TokenGenerationDisabled
-            ))
-        );
-    }
-}
diff --git a/paddler_agent/src/token_generation_support.rs b/paddler_agent/src/token_generation_support.rs
deleted file mode 100644
index 75f691929..000000000
--- a/paddler_agent/src/token_generation_support.rs
+++ /dev/null
@@ -1,10 +0,0 @@
-use std::sync::Arc;
-
-use llama_cpp_bindings::StreamingMarkers;
-
-use crate::chat_prompt_renderer::ChatPromptRenderer;
-
-pub struct TokenGenerationSupport {
-    pub chat_prompt_renderer: ChatPromptRenderer,
-    pub streaming_markers: Arc<StreamingMarkers>,
-}
diff --git a/paddler_agent/src/token_sampling.rs b/paddler_agent/src/token_sampling.rs
deleted file mode 100644
index 90d87a0f8..000000000
--- a/paddler_agent/src/token_sampling.rs
+++ /dev/null
@@ -1,48 +0,0 @@
-use anyhow::Context as _;
-use anyhow::Result;
-use llama_cpp_bindings::context::LlamaContext;
-use llama_cpp_bindings::sampling::LlamaSampler;
-use llama_cpp_bindings::token::data_array::LlamaTokenDataArray;
-
-use crate::grammar_sampling::GrammarSampling;
-use crate::sampling_outcome::SamplingOutcome;
-
-pub struct TokenSampling {
-    pub chain: LlamaSampler,
-    pub grammar: GrammarSampling,
-}
-
-impl TokenSampling {
-    pub fn sample(
-        &mut self,
-        llama_context: &LlamaContext,
-        batch_index: i32,
-        candidates: &mut LlamaTokenDataArray,
-    ) -> Result<SamplingOutcome> {
-        llama_context
-            .fill_token_data_array_ith(batch_index, candidates)
-            .context("failed to read token data array for sampling")?;
-        self.grammar
-            .constrain(candidates)
-            .context("failed to apply grammar sampler to token data array")?;
-        candidates
-            .apply_sampler(&self.chain)
-            .context("failed to apply sampler chain to token data array")?;
-
-        let Some(token) = candidates.selected_token() else {
-            return Ok(SamplingOutcome::AllCandidatesEliminated);
-        };
-
-        self.chain
-            .accept(token)
-            .context("sampler chain failed to accept the selected token")?;
-
-        if let Err(grammar_rejection) = self.grammar.accept(token) {
-            return Ok(SamplingOutcome::GrammarRejectedModelOutput(
-                grammar_rejection.to_string(),
-            ));
-        }
-
-        Ok(SamplingOutcome::Token(token))
-    }
-}
diff --git a/paddler_agent/src/tool_call_buffer.rs b/paddler_agent/src/tool_call_buffer.rs
deleted file mode 100644
index e8ef7e13d..000000000
--- a/paddler_agent/src/tool_call_buffer.rs
+++ /dev/null
@@ -1,38 +0,0 @@
-use std::mem::take;
-
-#[derive(Debug, Default)]
-pub struct ToolCallBuffer {
-    accumulated: String,
-}
-
-impl ToolCallBuffer {
-    pub fn append(&mut self, fragment: &str) {
-        self.accumulated.push_str(fragment);
-    }
-
-    #[must_use]
-    pub fn take(&mut self) -> String {
-        take(&mut self.accumulated)
-    }
-
-    #[must_use]
-    pub const fn is_empty(&self) -> bool {
-        self.accumulated.is_empty()
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use super::ToolCallBuffer;
-
-    #[test]
-    fn take_returns_every_appended_fragment_and_empties_the_buffer() {
-        let mut buffer = ToolCallBuffer::default();
-
-        buffer.append("<tool_call>\n");
-        buffer.append("{\"name\":\"x\"}");
-
-        assert_eq!(buffer.take(), "<tool_call>\n{\"name\":\"x\"}");
-        assert!(buffer.is_empty());
-    }
-}
diff --git a/paddler_agent/src/tool_call_pipeline.rs b/paddler_agent/src/tool_call_pipeline.rs
deleted file mode 100644
index 3f9e72fcf..000000000
--- a/paddler_agent/src/tool_call_pipeline.rs
+++ /dev/null
@@ -1,72 +0,0 @@
-use llama_cpp_bindings::ChatMessageParseOutcome;
-use llama_cpp_bindings::ChatMessageParser;
-use llama_cpp_bindings::ParsedToolCall;
-use llama_cpp_bindings::RawChatMessage;
-
-use paddler_messaging::raw_tool_call_tokens::RawToolCallTokens;
-use paddler_tool_call_validator::tool_call_validator::ToolCallValidator;
-
-use crate::tool_call_buffer::ToolCallBuffer;
-use crate::tool_call_event::ToolCallEvent;
-
-pub struct ToolCallPipeline {
-    buffer: ToolCallBuffer,
-    chat_message_parser: ChatMessageParser,
-    validator: ToolCallValidator,
-}
-
-impl ToolCallPipeline {
-    #[must_use]
-    pub fn new(chat_message_parser: ChatMessageParser, validator: ToolCallValidator) -> Self {
-        Self {
-            buffer: ToolCallBuffer::default(),
-            chat_message_parser,
-            validator,
-        }
-    }
-
-    pub fn feed(&mut self, fragment: &str) {
-        self.buffer.append(fragment);
-    }
-
-    #[must_use]
-    pub const fn buffer_is_empty(&self) -> bool {
-        self.buffer.is_empty()
-    }
-
-    pub fn finalize(&mut self) -> ToolCallEvent {
-        let input = self.buffer.take();
-
-        if input.is_empty() {
-            return ToolCallEvent::Resolved(Vec::new());
-        }
-
-        match self.chat_message_parser.parse(&input, false) {
-            Ok(ChatMessageParseOutcome::Recognized(parsed)) => {
-                self.validate_resolved(parsed.tool_calls)
-            }
-            Ok(ChatMessageParseOutcome::Unrecognized(RawChatMessage {
-                text,
-                ffi_error_message,
-                ..
-            })) => ToolCallEvent::UnrecognizedFormat(RawToolCallTokens {
-                text,
-                ffi_error_message,
-            }),
-            Err(parse_error) => ToolCallEvent::ParseFailed(parse_error),
-        }
-    }
-
-    fn validate_resolved(&self, tool_calls: Vec<ParsedToolCall>) -> ToolCallEvent {
-        let validation_errors: Vec<_> = tool_calls
-            .iter()
-            .filter_map(|tool_call| self.validator.validate(tool_call).err())
-            .collect();
-
-        if validation_errors.is_empty() {
-            ToolCallEvent::Resolved(tool_calls)
-        } else {
-            ToolCallEvent::ValidationFailed(validation_errors)
-        }
-    }
-}
diff --git a/paddler_agent/tests/a_new_subscriber_sees_the_state_applied_before_it_subscribed.rs b/paddler_agent/tests/a_new_subscriber_sees_the_state_applied_before_it_subscribed.rs
new file mode 100644
index 000000000..2f6aa5c50
--- /dev/null
+++ b/paddler_agent/tests/a_new_subscriber_sees_the_state_applied_before_it_subscribed.rs
@@ -0,0 +1,16 @@
+use paddler_agent::agent_applicable_state_holder::AgentApplicableStateHolder;
+
+use crate::state_overriding_the_chat_template::state_overriding_the_chat_template;
+
+#[test]
+fn a_new_subscriber_sees_the_state_applied_before_it_subscribed() {
+    let holder = AgentApplicableStateHolder::default();
+    let applied_state = state_overriding_the_chat_template();
+
+    holder.set_agent_applicable_state(applied_state.clone());
+
+    let agent_applicable_state_rx = holder.subscribe();
+
+    assert!(agent_applicable_state_rx.has_changed().unwrap());
+    assert_eq!(*agent_applicable_state_rx.borrow(), applied_state);
+}
diff --git a/paddler_agent/tests/a_pipeline_that_fails_to_start_is_retried_until_it_is_stuck.rs b/paddler_agent/tests/a_pipeline_that_fails_to_start_is_retried_until_it_is_stuck.rs
new file mode 100644
index 000000000..5a3ca8a89
--- /dev/null
+++ b/paddler_agent/tests/a_pipeline_that_fails_to_start_is_retried_until_it_is_stuck.rs
@@ -0,0 +1,32 @@
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::agent_state_application_status::AgentStateApplicationStatus;
+
+use crate::arbiter_under_test::ArbiterUnderTest;
+use crate::state_whose_model_cannot_load::state_whose_model_cannot_load;
+
+#[tokio::test]
+async fn a_pipeline_that_fails_to_start_is_retried_until_it_is_stuck() {
+    let mut arbiter_under_test = ArbiterUnderTest::new(CancellationToken::new());
+
+    arbiter_under_test
+        .pipeline_arbiter
+        .apply(state_whose_model_cannot_load());
+
+    assert_eq!(
+        [
+            arbiter_under_test.conclude_next_progress().await,
+            arbiter_under_test.conclude_next_progress().await,
+            arbiter_under_test.conclude_next_progress().await,
+            arbiter_under_test.conclude_next_progress().await,
+        ],
+        [
+            AgentStateApplicationStatus::Fresh,
+            AgentStateApplicationStatus::AttemptedAndRetrying,
+            AgentStateApplicationStatus::AttemptedAndRetrying,
+            AgentStateApplicationStatus::Stuck,
+        ]
+    );
+
+    arbiter_under_test.pipeline_arbiter.shut_down().await;
+}
diff --git a/paddler_agent/tests/a_reconciliation_replaced_before_its_worker_runs_leaves_no_issue_behind.rs b/paddler_agent/tests/a_reconciliation_replaced_before_its_worker_runs_leaves_no_issue_behind.rs
new file mode 100644
index 000000000..cb669c628
--- /dev/null
+++ b/paddler_agent/tests/a_reconciliation_replaced_before_its_worker_runs_leaves_no_issue_behind.rs
@@ -0,0 +1,36 @@
+use std::path::Path;
+
+use paddler_messaging::agent_desired_state::AgentDesiredState;
+use paddler_messaging::agent_issue::AgentIssue;
+use paddler_messaging::agent_issue_params::model_path::ModelPath;
+use paddler_model_source::agent_local_model_path::AgentLocalModelPath;
+use paddler_model_source::model_source::ModelSource;
+
+use crate::reconciliation_after_replacement::ReconciliationAfterReplacement;
+use crate::reconciliation_replaced_before_its_worker_runs::ReconciliationReplacedBeforeItsWorkerRuns;
+
+const MISSING_MODEL_PATH: &str = "/paddler-missing-model-of-a-replaced-desired-state.gguf";
+
+#[test]
+fn a_reconciliation_replaced_before_its_worker_runs_leaves_no_issue_behind() {
+    let ReconciliationAfterReplacement {
+        slot_aggregated_status,
+        ..
+    } = ReconciliationReplacedBeforeItsWorkerRuns {
+        replaced_desired_state: AgentDesiredState {
+            model: ModelSource::LocalToAgent(
+                AgentLocalModelPath::from_absolute_path(Path::new(MISSING_MODEL_PATH))
+                    .expect("the missing model path must be absolute"),
+            )
+            .into_agent_desired_model(),
+            ..AgentDesiredState::default()
+        },
+    }
+    .run();
+
+    assert!(
+        !slot_aggregated_status.has_issue(&AgentIssue::ModelFileDoesNotExist(ModelPath {
+            model_path: MISSING_MODEL_PATH.to_owned(),
+        }))
+    );
+}
diff --git a/paddler_agent/tests/a_reconciliation_replaced_before_its_worker_runs_never_publishes_its_state.rs b/paddler_agent/tests/a_reconciliation_replaced_before_its_worker_runs_never_publishes_its_state.rs
new file mode 100644
index 000000000..42e61992e
--- /dev/null
+++ b/paddler_agent/tests/a_reconciliation_replaced_before_its_worker_runs_never_publishes_its_state.rs
@@ -0,0 +1,21 @@
+use paddler_agent::agent_applicable_state::AgentApplicableState;
+use paddler_messaging::agent_desired_state::AgentDesiredState;
+
+use crate::reconciliation_after_replacement::ReconciliationAfterReplacement;
+use crate::reconciliation_replaced_before_its_worker_runs::ReconciliationReplacedBeforeItsWorkerRuns;
+
+#[test]
+fn a_reconciliation_replaced_before_its_worker_runs_never_publishes_its_state() {
+    let ReconciliationAfterReplacement {
+        agent_applicable_state_holder,
+        ..
+    } = ReconciliationReplacedBeforeItsWorkerRuns {
+        replaced_desired_state: AgentDesiredState::default(),
+    }
+    .run();
+
+    assert_eq!(
+        agent_applicable_state_holder.get_agent_applicable_state(),
+        AgentApplicableState::NotConfigured
+    );
+}
diff --git a/paddler_agent/tests/a_replacement_cancelled_by_the_agent_shutting_down_changes_nothing.rs b/paddler_agent/tests/a_replacement_cancelled_by_the_agent_shutting_down_changes_nothing.rs
new file mode 100644
index 000000000..adc18cefa
--- /dev/null
+++ b/paddler_agent/tests/a_replacement_cancelled_by_the_agent_shutting_down_changes_nothing.rs
@@ -0,0 +1,35 @@
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::produces_snapshot::ProducesSnapshot as _;
+
+use crate::arbiter_under_test::ArbiterUnderTest;
+use crate::state_whose_model_cannot_load::state_whose_model_cannot_load;
+
+#[tokio::test]
+async fn a_replacement_cancelled_by_the_agent_shutting_down_changes_nothing() {
+    let agent_shutdown = CancellationToken::new();
+    let mut arbiter_under_test = ArbiterUnderTest::new(agent_shutdown.clone());
+
+    arbiter_under_test
+        .pipeline_arbiter
+        .apply(state_whose_model_cannot_load());
+    arbiter_under_test.conclude_next_progress().await;
+    agent_shutdown.cancel();
+
+    let status_before_the_replacement_concluded = arbiter_under_test
+        .inference_runtime_context
+        .slot_aggregated_status
+        .make_snapshot();
+
+    arbiter_under_test.conclude_next_progress().await;
+
+    assert_eq!(
+        arbiter_under_test
+            .inference_runtime_context
+            .slot_aggregated_status
+            .make_snapshot(),
+        status_before_the_replacement_concluded
+    );
+
+    arbiter_under_test.pipeline_arbiter.shut_down().await;
+}
diff --git a/paddler_agent/tests/a_replacement_whose_model_cannot_load_reports_nothing_as_served.rs b/paddler_agent/tests/a_replacement_whose_model_cannot_load_reports_nothing_as_served.rs
new file mode 100644
index 000000000..5bc894090
--- /dev/null
+++ b/paddler_agent/tests/a_replacement_whose_model_cannot_load_reports_nothing_as_served.rs
@@ -0,0 +1,41 @@
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent::agent_applicable_state::AgentApplicableState;
+use paddler_messaging::agent_state_application_status::AgentStateApplicationStatus;
+
+use crate::arbiter_under_test::ArbiterUnderTest;
+use crate::state_overriding_the_chat_template::state_overriding_the_chat_template;
+use crate::state_whose_model_cannot_load::state_whose_model_cannot_load;
+
+#[tokio::test]
+async fn a_replacement_whose_model_cannot_load_reports_nothing_as_served() {
+    let mut arbiter_under_test = ArbiterUnderTest::new(CancellationToken::new());
+
+    arbiter_under_test
+        .pipeline_arbiter
+        .apply(state_overriding_the_chat_template());
+    arbiter_under_test.conclude_next_progress().await;
+
+    assert_eq!(
+        arbiter_under_test.conclude_next_progress().await,
+        AgentStateApplicationStatus::Applied
+    );
+
+    arbiter_under_test
+        .pipeline_arbiter
+        .apply(state_whose_model_cannot_load());
+    arbiter_under_test.conclude_next_progress().await;
+
+    assert_eq!(
+        arbiter_under_test.conclude_next_progress().await,
+        AgentStateApplicationStatus::AttemptedAndRetrying
+    );
+    assert_eq!(
+        arbiter_under_test
+            .served_applicable_state_holder
+            .get_agent_applicable_state(),
+        AgentApplicableState::NotConfigured
+    );
+
+    arbiter_under_test.pipeline_arbiter.shut_down().await;
+}
diff --git a/paddler_agent/tests/a_startup_cancelled_by_the_agent_shutting_down_changes_nothing.rs b/paddler_agent/tests/a_startup_cancelled_by_the_agent_shutting_down_changes_nothing.rs
new file mode 100644
index 000000000..487cd6390
--- /dev/null
+++ b/paddler_agent/tests/a_startup_cancelled_by_the_agent_shutting_down_changes_nothing.rs
@@ -0,0 +1,34 @@
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::produces_snapshot::ProducesSnapshot as _;
+
+use crate::arbiter_under_test::ArbiterUnderTest;
+use crate::state_whose_model_cannot_load::state_whose_model_cannot_load;
+
+#[tokio::test]
+async fn a_startup_cancelled_by_the_agent_shutting_down_changes_nothing() {
+    let agent_shutdown = CancellationToken::new();
+    let mut arbiter_under_test = ArbiterUnderTest::new(agent_shutdown.clone());
+
+    agent_shutdown.cancel();
+    arbiter_under_test
+        .pipeline_arbiter
+        .apply(state_whose_model_cannot_load());
+
+    let status_before_the_startup_concluded = arbiter_under_test
+        .inference_runtime_context
+        .slot_aggregated_status
+        .make_snapshot();
+
+    arbiter_under_test.conclude_next_progress().await;
+
+    assert_eq!(
+        arbiter_under_test
+            .inference_runtime_context
+            .slot_aggregated_status
+            .make_snapshot(),
+        status_before_the_startup_concluded
+    );
+
+    arbiter_under_test.pipeline_arbiter.shut_down().await;
+}
diff --git a/paddler_agent/tests/a_state_superseded_while_it_is_prepared_yields_to_its_replacement.rs b/paddler_agent/tests/a_state_superseded_while_it_is_prepared_yields_to_its_replacement.rs
new file mode 100644
index 000000000..844961f75
--- /dev/null
+++ b/paddler_agent/tests/a_state_superseded_while_it_is_prepared_yields_to_its_replacement.rs
@@ -0,0 +1,30 @@
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent::agent_applicable_state::AgentApplicableState;
+
+use crate::arbiter_under_test::ArbiterUnderTest;
+use crate::state_overriding_the_chat_template::state_overriding_the_chat_template;
+
+#[tokio::test]
+async fn a_state_superseded_while_it_is_prepared_yields_to_its_replacement() {
+    let mut arbiter_under_test = ArbiterUnderTest::new(CancellationToken::new());
+
+    arbiter_under_test
+        .pipeline_arbiter
+        .apply(AgentApplicableState::NotConfigured);
+    arbiter_under_test
+        .pipeline_arbiter
+        .apply(state_overriding_the_chat_template());
+    arbiter_under_test.conclude_next_progress().await;
+    arbiter_under_test.conclude_next_progress().await;
+    arbiter_under_test.conclude_next_progress().await;
+
+    assert_eq!(
+        arbiter_under_test
+            .served_applicable_state_holder
+            .get_agent_applicable_state(),
+        state_overriding_the_chat_template()
+    );
+
+    arbiter_under_test.pipeline_arbiter.shut_down().await;
+}
diff --git a/paddler_agent/tests/a_state_superseded_while_its_model_loads_yields_to_its_replacement.rs b/paddler_agent/tests/a_state_superseded_while_its_model_loads_yields_to_its_replacement.rs
new file mode 100644
index 000000000..d952f17c6
--- /dev/null
+++ b/paddler_agent/tests/a_state_superseded_while_its_model_loads_yields_to_its_replacement.rs
@@ -0,0 +1,30 @@
+use tokio_util::sync::CancellationToken;
+
+use crate::arbiter_under_test::ArbiterUnderTest;
+use crate::state_overriding_the_chat_template::state_overriding_the_chat_template;
+use crate::state_whose_model_cannot_load::state_whose_model_cannot_load;
+
+#[tokio::test]
+async fn a_state_superseded_while_its_model_loads_yields_to_its_replacement() {
+    let mut arbiter_under_test = ArbiterUnderTest::new(CancellationToken::new());
+
+    arbiter_under_test
+        .pipeline_arbiter
+        .apply(state_whose_model_cannot_load());
+    arbiter_under_test.conclude_next_progress().await;
+    arbiter_under_test
+        .pipeline_arbiter
+        .apply(state_overriding_the_chat_template());
+    arbiter_under_test.conclude_next_progress().await;
+    arbiter_under_test.conclude_next_progress().await;
+    arbiter_under_test.conclude_next_progress().await;
+
+    assert_eq!(
+        arbiter_under_test
+            .served_applicable_state_holder
+            .get_agent_applicable_state(),
+        state_overriding_the_chat_template()
+    );
+
+    arbiter_under_test.pipeline_arbiter.shut_down().await;
+}
diff --git a/paddler_agent/tests/a_state_superseding_a_decisions_preparation_that_fails_leaves_no_issue_behind.rs b/paddler_agent/tests/a_state_superseding_a_decisions_preparation_that_fails_leaves_no_issue_behind.rs
new file mode 100644
index 000000000..379118f07
--- /dev/null
+++ b/paddler_agent/tests/a_state_superseding_a_decisions_preparation_that_fails_leaves_no_issue_behind.rs
@@ -0,0 +1,17 @@
+use crate::slots_insufficient_for_decisions::slots_insufficient_for_decisions;
+use crate::startup_superseded_while_its_blocking_work_waits::StartupSupersededWhileItsBlockingWorkWaits;
+use crate::state_enabling_decisions::state_enabling_decisions;
+
+#[test]
+fn a_state_superseding_a_decisions_preparation_that_fails_leaves_no_issue_behind() {
+    let slot_aggregated_status = StartupSupersededWhileItsBlockingWorkWaits {
+        superseded_state: state_enabling_decisions(),
+    }
+    .run();
+
+    assert!(
+        !slot_aggregated_status.has_issue(&slots_insufficient_for_decisions(
+            slot_aggregated_status.desired_slots_total
+        ))
+    );
+}
diff --git a/paddler_agent/tests/a_state_superseding_a_model_load_that_fails_leaves_no_issue_behind.rs b/paddler_agent/tests/a_state_superseding_a_model_load_that_fails_leaves_no_issue_behind.rs
new file mode 100644
index 000000000..98428d4ee
--- /dev/null
+++ b/paddler_agent/tests/a_state_superseding_a_model_load_that_fails_leaves_no_issue_behind.rs
@@ -0,0 +1,19 @@
+use paddler_messaging::agent_issue::AgentIssue;
+use paddler_messaging::agent_issue_params::model_path::ModelPath;
+
+use crate::settings_whose_model_cannot_load::settings_whose_model_cannot_load;
+use crate::startup_superseded_while_its_blocking_work_waits::StartupSupersededWhileItsBlockingWorkWaits;
+use crate::state_whose_model_cannot_load::state_whose_model_cannot_load;
+
+#[test]
+fn a_state_superseding_a_model_load_that_fails_leaves_no_issue_behind() {
+    assert!(
+        !StartupSupersededWhileItsBlockingWorkWaits {
+            superseded_state: state_whose_model_cannot_load(),
+        }
+        .run()
+        .has_issue(&AgentIssue::ModelCannotBeLoaded(ModelPath::from(
+            settings_whose_model_cannot_load().model_path.as_path()
+        )))
+    );
+}
diff --git a/paddler_agent/tests/a_state_without_a_model_serves_nothing_and_forgets_the_previous_model.rs b/paddler_agent/tests/a_state_without_a_model_serves_nothing_and_forgets_the_previous_model.rs
new file mode 100644
index 000000000..2bf90a434
--- /dev/null
+++ b/paddler_agent/tests/a_state_without_a_model_serves_nothing_and_forgets_the_previous_model.rs
@@ -0,0 +1,54 @@
+use std::collections::BTreeMap;
+use std::sync::Arc;
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent::agent_applicable_state::AgentApplicableState;
+use paddler_messaging::agent_runtime_status::AgentRuntimeStatus;
+use paddler_messaging::agent_state_application_status::AgentStateApplicationStatus;
+use paddler_messaging::model_metadata::ModelMetadata;
+use paddler_messaging::produces_snapshot::ProducesSnapshot as _;
+
+use crate::arbiter_under_test::ArbiterUnderTest;
+
+#[tokio::test]
+async fn a_state_without_a_model_serves_nothing_and_forgets_the_previous_model() {
+    let mut arbiter_under_test = ArbiterUnderTest::new(CancellationToken::new());
+
+    arbiter_under_test
+        .inference_runtime_context
+        .model_metadata_holder
+        .set_model_metadata(Arc::new(ModelMetadata {
+            metadata: BTreeMap::from([("general.name".to_owned(), "previous".to_owned())]),
+        }));
+    arbiter_under_test
+        .pipeline_arbiter
+        .apply(AgentApplicableState::NotConfigured);
+
+    assert_eq!(
+        arbiter_under_test.conclude_next_progress().await,
+        AgentStateApplicationStatus::Fresh
+    );
+    assert_eq!(
+        arbiter_under_test.conclude_next_progress().await,
+        AgentStateApplicationStatus::Applied
+    );
+    assert_eq!(
+        arbiter_under_test
+            .inference_runtime_context
+            .slot_aggregated_status
+            .make_snapshot()
+            .status
+            .runtime,
+        AgentRuntimeStatus::Idle
+    );
+    assert_eq!(
+        arbiter_under_test
+            .inference_runtime_context
+            .model_metadata_holder
+            .get_model_metadata(),
+        None
+    );
+
+    arbiter_under_test.pipeline_arbiter.shut_down().await;
+}
diff --git a/paddler_agent/tests/an_identical_state_is_not_applied_again.rs b/paddler_agent/tests/an_identical_state_is_not_applied_again.rs
new file mode 100644
index 000000000..759845abf
--- /dev/null
+++ b/paddler_agent/tests/an_identical_state_is_not_applied_again.rs
@@ -0,0 +1,18 @@
+use paddler_agent::agent_applicable_state_holder::AgentApplicableStateHolder;
+
+use crate::state_overriding_the_chat_template::state_overriding_the_chat_template;
+
+#[test]
+fn an_identical_state_is_not_applied_again() {
+    let holder = AgentApplicableStateHolder::default();
+    let applied_state = state_overriding_the_chat_template();
+
+    holder.set_agent_applicable_state(applied_state.clone());
+
+    let mut agent_applicable_state_rx = holder.subscribe();
+
+    agent_applicable_state_rx.borrow_and_update();
+    holder.set_agent_applicable_state(applied_state);
+
+    assert!(!agent_applicable_state_rx.has_changed().unwrap());
+}
diff --git a/paddler_agent/tests/arbiter_under_test.rs b/paddler_agent/tests/arbiter_under_test.rs
new file mode 100644
index 000000000..4b1229b38
--- /dev/null
+++ b/paddler_agent/tests/arbiter_under_test.rs
@@ -0,0 +1,47 @@
+use std::sync::Arc;
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent::agent_applicable_state_holder::AgentApplicableStateHolder;
+use paddler_agent::pipeline_arbiter::PipelineArbiter;
+use paddler_agent::pipeline_arbiter_params::PipelineArbiterParams;
+use paddler_agent_runtime::inference_runtime_context::InferenceRuntimeContext;
+use paddler_agent_runtime::inference_runtime_context_params::InferenceRuntimeContextParams;
+use paddler_messaging::agent_state_application_status::AgentStateApplicationStatus;
+
+pub struct ArbiterUnderTest {
+    pub inference_runtime_context: InferenceRuntimeContext,
+    pub pipeline_arbiter: PipelineArbiter,
+    pub served_applicable_state_holder: Arc<AgentApplicableStateHolder>,
+}
+
+impl ArbiterUnderTest {
+    pub fn new(agent_shutdown: CancellationToken) -> Self {
+        let inference_runtime_context =
+            InferenceRuntimeContext::new(InferenceRuntimeContextParams {
+                agent_name: Some("agent".to_owned()),
+                slots: 1,
+            });
+        let served_applicable_state_holder = Arc::new(AgentApplicableStateHolder::default());
+
+        Self {
+            pipeline_arbiter: PipelineArbiter::new(PipelineArbiterParams {
+                agent_shutdown,
+                inference_runtime_context: inference_runtime_context.clone(),
+                served_applicable_state_holder: served_applicable_state_holder.clone(),
+            }),
+            inference_runtime_context,
+            served_applicable_state_holder,
+        }
+    }
+
+    pub async fn conclude_next_progress(&mut self) -> AgentStateApplicationStatus {
+        let serving_startup_progress = self.pipeline_arbiter.next_finished_startup().await;
+
+        self.pipeline_arbiter.conclude(serving_startup_progress);
+
+        self.inference_runtime_context
+            .slot_aggregated_status
+            .get_state_application_status()
+    }
+}
diff --git a/paddler_agent/tests/main.rs b/paddler_agent/tests/main.rs
new file mode 100644
index 000000000..032ca201c
--- /dev/null
+++ b/paddler_agent/tests/main.rs
@@ -0,0 +1,33 @@
+mod a_new_subscriber_sees_the_state_applied_before_it_subscribed;
+mod a_pipeline_that_fails_to_start_is_retried_until_it_is_stuck;
+mod a_reconciliation_replaced_before_its_worker_runs_leaves_no_issue_behind;
+mod a_reconciliation_replaced_before_its_worker_runs_never_publishes_its_state;
+mod a_replacement_cancelled_by_the_agent_shutting_down_changes_nothing;
+mod a_replacement_whose_model_cannot_load_reports_nothing_as_served;
+mod a_startup_cancelled_by_the_agent_shutting_down_changes_nothing;
+mod a_state_superseded_while_it_is_prepared_yields_to_its_replacement;
+mod a_state_superseded_while_its_model_loads_yields_to_its_replacement;
+mod a_state_superseding_a_decisions_preparation_that_fails_leaves_no_issue_behind;
+mod a_state_superseding_a_model_load_that_fails_leaves_no_issue_behind;
+mod a_state_without_a_model_serves_nothing_and_forgets_the_previous_model;
+mod an_identical_state_is_not_applied_again;
+#[cfg(test)]
+mod arbiter_under_test;
+#[cfg(test)]
+mod reconciliation_after_replacement;
+#[cfg(test)]
+mod reconciliation_replaced_before_its_worker_runs;
+#[cfg(test)]
+mod settings_whose_model_cannot_load;
+mod shutting_down_waits_for_a_decisions_preparation_that_fails;
+#[cfg(test)]
+mod slots_insufficient_for_decisions;
+#[cfg(test)]
+mod startup_superseded_while_its_blocking_work_waits;
+#[cfg(test)]
+mod state_enabling_decisions;
+#[cfg(test)]
+mod state_overriding_the_chat_template;
+#[cfg(test)]
+mod state_whose_model_cannot_load;
+mod the_reconciliation_service_stops_once_its_desired_state_channel_closes;
diff --git a/paddler_agent/tests/reconciliation_after_replacement.rs b/paddler_agent/tests/reconciliation_after_replacement.rs
new file mode 100644
index 000000000..0e067713c
--- /dev/null
+++ b/paddler_agent/tests/reconciliation_after_replacement.rs
@@ -0,0 +1,9 @@
+use std::sync::Arc;
+
+use paddler_agent::agent_applicable_state_holder::AgentApplicableStateHolder;
+use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
+
+pub struct ReconciliationAfterReplacement {
+    pub agent_applicable_state_holder: Arc<AgentApplicableStateHolder>,
+    pub slot_aggregated_status: Arc<SlotAggregatedStatus>,
+}
diff --git a/paddler_agent/tests/reconciliation_replaced_before_its_worker_runs.rs b/paddler_agent/tests/reconciliation_replaced_before_its_worker_runs.rs
new file mode 100644
index 000000000..9998a535c
--- /dev/null
+++ b/paddler_agent/tests/reconciliation_replaced_before_its_worker_runs.rs
@@ -0,0 +1,112 @@
+use std::sync::Arc;
+
+use tempfile::TempDir;
+use tokio::spawn;
+use tokio::sync::watch;
+use tokio_util::sync::CancellationToken;
+use trzcina::Service as _;
+
+use paddler_agent::agent_applicable_state_holder::AgentApplicableStateHolder;
+use paddler_agent::desired_state_reconciler::DesiredStateReconciler;
+use paddler_agent::reconciliation_service::ReconciliationService;
+use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
+use paddler_blocking_thread_fixture::occupied_blocking_thread::OccupiedBlockingThread;
+use paddler_blocking_thread_fixture::single_blocking_thread_runtime::single_blocking_thread_runtime;
+use paddler_cache_dir::cache_dir::CacheDir;
+use paddler_download_manager::download_manager::DownloadManager;
+use paddler_local_http_fixture::fixture_response::FixtureResponse;
+use paddler_local_http_fixture::local_http_fixture::LocalHttpFixture;
+use paddler_messaging::agent_desired_model::AgentDesiredModel;
+use paddler_messaging::agent_desired_state::AgentDesiredState;
+use paddler_messaging::subscribes_to_updates::SubscribesToUpdates as _;
+use paddler_model_source::desired_model_resolver::DesiredModelResolver;
+use paddler_model_source::huggingface_cache_location::HuggingFaceCacheLocation;
+use paddler_model_source::huggingface_default_endpoint::HUGGINGFACE_DEFAULT_ENDPOINT;
+use paddler_model_source::huggingface_hub::HuggingFaceHub;
+
+use crate::reconciliation_after_replacement::ReconciliationAfterReplacement;
+
+pub struct ReconciliationReplacedBeforeItsWorkerRuns {
+    pub replaced_desired_state: AgentDesiredState,
+}
+
+impl ReconciliationReplacedBeforeItsWorkerRuns {
+    pub fn run(self) -> ReconciliationAfterReplacement {
+        single_blocking_thread_runtime()
+            .expect("the test runtime must build")
+            .block_on(self.run_on_one_blocking_thread())
+    }
+
+    async fn run_on_one_blocking_thread(self) -> ReconciliationAfterReplacement {
+        let cache_dir = TempDir::new().expect("the download cache must be creatable");
+        let stalled_model_server = LocalHttpFixture::start(FixtureResponse::StallBeforeHeaders)
+            .await
+            .expect("the stalled model server must start");
+        let occupied_blocking_thread = OccupiedBlockingThread::occupy();
+        let agent_applicable_state_holder = Arc::new(AgentApplicableStateHolder::default());
+        let slot_aggregated_status = Arc::new(SlotAggregatedStatus::new(1));
+        let mut status_updates = slot_aggregated_status.subscribe_to_updates();
+        let (agent_desired_state_tx, agent_desired_state_rx) =
+            watch::channel(AgentDesiredState::default());
+        let shutdown = CancellationToken::new();
+        let reconciliation = spawn(
+            Box::new(ReconciliationService {
+                agent_applicable_state_holder: agent_applicable_state_holder.clone(),
+                agent_desired_state_rx,
+                desired_state_reconciler: DesiredStateReconciler {
+                    desired_model_resolver: DesiredModelResolver {
+                        cache_dir: CacheDir {
+                            explicit: Some(cache_dir.path().to_path_buf()),
+                            home: None,
+                            xdg: None,
+                        },
+                        download_manager: DownloadManager::new()
+                            .expect("the model downloader must start"),
+                        huggingface_client: HuggingFaceHub {
+                            cache_location: HuggingFaceCacheLocation::HubDefault,
+                            endpoint: HUGGINGFACE_DEFAULT_ENDPOINT.to_owned(),
+                        }
+                        .client()
+                        .expect("the Hugging Face client must start"),
+                    },
+                    slot_aggregated_status: slot_aggregated_status.clone(),
+                },
+            })
+            .run(shutdown.clone()),
+        );
+
+        status_updates.borrow_and_update();
+        agent_desired_state_tx.send_replace(self.replaced_desired_state);
+        status_updates
+            .changed()
+            .await
+            .expect("the agent must begin the replaced desired state");
+        agent_desired_state_tx.send_replace(AgentDesiredState {
+            model: AgentDesiredModel::Uri(stalled_model_server.url("/model.gguf")),
+            ..AgentDesiredState::default()
+        });
+        occupied_blocking_thread
+            .release()
+            .await
+            .expect("the blocking thread occupant must be released");
+        stalled_model_server
+            .wait_for_requests(1)
+            .await
+            .expect("the replacement must start downloading after the replaced worker ran");
+        shutdown.cancel();
+        reconciliation
+            .await
+            .expect("the reconciliation service must not panic")
+            .expect("the reconciliation service must stop cleanly");
+        stalled_model_server
+            .stop()
+            .await
+            .expect("the stalled model server must not panic")
+            .expect("the stalled model server must serve every request");
+
+        ReconciliationAfterReplacement {
+            agent_applicable_state_holder,
+            slot_aggregated_status,
+        }
+    }
+}
diff --git a/paddler_agent/tests/settings_whose_model_cannot_load.rs b/paddler_agent/tests/settings_whose_model_cannot_load.rs
new file mode 100644
index 000000000..14961a359
--- /dev/null
+++ b/paddler_agent/tests/settings_whose_model_cannot_load.rs
@@ -0,0 +1,29 @@
+use std::num::NonZeroU32;
+use std::path::PathBuf;
+
+use paddler_agent::applicable_decisions::ApplicableDecisions;
+use paddler_agent::inference_capabilities::InferenceCapabilities;
+use paddler_agent::model_serving_settings::ModelServingSettings;
+use paddler_agent_text_generation::multimodal_projection::MultimodalProjection;
+use paddler_agent_text_generation::text_generation_settings::TextGenerationSettings;
+use paddler_inference_parameters::model_runtime_parameters::ModelRuntimeParameters;
+use paddler_inference_parameters::pooling_type::PoolingType;
+use paddler_inference_parameters::sampling_parameters::SamplingParameters;
+use paddler_messaging::chat_template_source::ChatTemplateSource;
+
+pub fn settings_whose_model_cannot_load() -> ModelServingSettings {
+    ModelServingSettings {
+        capabilities: InferenceCapabilities {
+            decisions: ApplicableDecisions::Disabled,
+            embeddings: PoolingType::Last,
+            text_generation: TextGenerationSettings {
+                chat_template_source: ChatTemplateSource::EmbeddedInModel,
+                image_resize_to_fit: NonZeroU32::MIN,
+                multimodal_projection: MultimodalProjection::NotConfigured,
+                sampling_parameters: SamplingParameters::default(),
+            },
+        },
+        model_path: PathBuf::from("/nonexistent/model.gguf"),
+        model_runtime_parameters: ModelRuntimeParameters::default(),
+    }
+}
diff --git a/paddler_agent/tests/shutting_down_waits_for_a_decisions_preparation_that_fails.rs b/paddler_agent/tests/shutting_down_waits_for_a_decisions_preparation_that_fails.rs
new file mode 100644
index 000000000..0f00d3f1a
--- /dev/null
+++ b/paddler_agent/tests/shutting_down_waits_for_a_decisions_preparation_that_fails.rs
@@ -0,0 +1,55 @@
+use futures_util::FutureExt as _;
+use tokio::join;
+use tokio_util::sync::CancellationToken;
+
+use paddler_blocking_thread_fixture::occupied_blocking_thread::OccupiedBlockingThread;
+use paddler_blocking_thread_fixture::single_blocking_thread_runtime::single_blocking_thread_runtime;
+
+use crate::arbiter_under_test::ArbiterUnderTest;
+use crate::slots_insufficient_for_decisions::slots_insufficient_for_decisions;
+use crate::state_enabling_decisions::state_enabling_decisions;
+
+#[test]
+fn shutting_down_waits_for_a_decisions_preparation_that_fails() {
+    single_blocking_thread_runtime()
+        .expect("the test runtime must build")
+        .block_on(async {
+            let occupied_blocking_thread = OccupiedBlockingThread::occupy();
+            let mut arbiter_under_test = ArbiterUnderTest::new(CancellationToken::new());
+
+            arbiter_under_test
+                .pipeline_arbiter
+                .apply(state_enabling_decisions());
+
+            assert!(
+                arbiter_under_test
+                    .pipeline_arbiter
+                    .next_finished_startup()
+                    .now_or_never()
+                    .is_none()
+            );
+
+            let ArbiterUnderTest {
+                inference_runtime_context,
+                pipeline_arbiter,
+                ..
+            } = arbiter_under_test;
+
+            join!(pipeline_arbiter.shut_down(), async {
+                occupied_blocking_thread
+                    .release()
+                    .await
+                    .expect("the blocking thread occupant must be released");
+            });
+
+            assert!(
+                inference_runtime_context.slot_aggregated_status.has_issue(
+                    &slots_insufficient_for_decisions(
+                        inference_runtime_context
+                            .slot_aggregated_status
+                            .desired_slots_total
+                    )
+                )
+            );
+        });
+}
diff --git a/paddler_agent/tests/slots_insufficient_for_decisions.rs b/paddler_agent/tests/slots_insufficient_for_decisions.rs
new file mode 100644
index 000000000..089a9688b
--- /dev/null
+++ b/paddler_agent/tests/slots_insufficient_for_decisions.rs
@@ -0,0 +1,10 @@
+use paddler_messaging::agent_issue::AgentIssue;
+use paddler_messaging::agent_issue_params::slots_insufficient_for_decisions_params::SlotsInsufficientForDecisionsParams;
+use paddler_messaging::decision_slots_minimum::DECISION_SLOTS_MINIMUM;
+
+pub fn slots_insufficient_for_decisions(desired_slots: u16) -> AgentIssue {
+    AgentIssue::SlotsInsufficientForDecisions(SlotsInsufficientForDecisionsParams {
+        desired_slots,
+        required_slots: DECISION_SLOTS_MINIMUM.get(),
+    })
+}
diff --git a/paddler_agent/tests/startup_superseded_while_its_blocking_work_waits.rs b/paddler_agent/tests/startup_superseded_while_its_blocking_work_waits.rs
new file mode 100644
index 000000000..9167a8d24
--- /dev/null
+++ b/paddler_agent/tests/startup_superseded_while_its_blocking_work_waits.rs
@@ -0,0 +1,62 @@
+use std::sync::Arc;
+
+use futures_util::FutureExt as _;
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent::agent_applicable_state::AgentApplicableState;
+use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
+use paddler_blocking_thread_fixture::occupied_blocking_thread::OccupiedBlockingThread;
+use paddler_blocking_thread_fixture::single_blocking_thread_runtime::single_blocking_thread_runtime;
+
+use crate::arbiter_under_test::ArbiterUnderTest;
+
+pub struct StartupSupersededWhileItsBlockingWorkWaits {
+    pub superseded_state: AgentApplicableState,
+}
+
+impl StartupSupersededWhileItsBlockingWorkWaits {
+    pub fn run(self) -> Arc<SlotAggregatedStatus> {
+        single_blocking_thread_runtime()
+            .expect("the test runtime must build")
+            .block_on(self.run_on_one_blocking_thread())
+    }
+
+    async fn run_on_one_blocking_thread(self) -> Arc<SlotAggregatedStatus> {
+        let occupied_blocking_thread = OccupiedBlockingThread::occupy();
+        let mut arbiter_under_test = ArbiterUnderTest::new(CancellationToken::new());
+
+        arbiter_under_test
+            .pipeline_arbiter
+            .apply(self.superseded_state);
+
+        while let Some(serving_startup_progress) = arbiter_under_test
+            .pipeline_arbiter
+            .next_finished_startup()
+            .now_or_never()
+        {
+            arbiter_under_test
+                .pipeline_arbiter
+                .conclude(serving_startup_progress);
+        }
+
+        arbiter_under_test
+            .pipeline_arbiter
+            .apply(AgentApplicableState::NotConfigured);
+        occupied_blocking_thread
+            .release()
+            .await
+            .expect("the blocking thread occupant must be released");
+        arbiter_under_test.conclude_next_progress().await;
+        arbiter_under_test.conclude_next_progress().await;
+        arbiter_under_test.conclude_next_progress().await;
+
+        let slot_aggregated_status = arbiter_under_test
+            .inference_runtime_context
+            .slot_aggregated_status
+            .clone();
+
+        arbiter_under_test.pipeline_arbiter.shut_down().await;
+
+        slot_aggregated_status
+    }
+}
diff --git a/paddler_agent/tests/state_enabling_decisions.rs b/paddler_agent/tests/state_enabling_decisions.rs
new file mode 100644
index 000000000..7294961de
--- /dev/null
+++ b/paddler_agent/tests/state_enabling_decisions.rs
@@ -0,0 +1,22 @@
+use std::path::PathBuf;
+
+use paddler_agent::agent_applicable_state::AgentApplicableState;
+use paddler_agent::applicable_decisions::ApplicableDecisions;
+use paddler_agent::inference_capabilities::InferenceCapabilities;
+use paddler_agent::model_serving_settings::ModelServingSettings;
+
+use crate::settings_whose_model_cannot_load::settings_whose_model_cannot_load;
+
+pub fn state_enabling_decisions() -> AgentApplicableState {
+    let settings_whose_model_cannot_load = settings_whose_model_cannot_load();
+
+    AgentApplicableState::WithModel(ModelServingSettings {
+        capabilities: InferenceCapabilities {
+            decisions: ApplicableDecisions::Enabled {
+                pointer_head_path: PathBuf::from("/nonexistent/pointer_head.gguf"),
+            },
+            ..settings_whose_model_cannot_load.capabilities
+        },
+        ..settings_whose_model_cannot_load
+    })
+}
diff --git a/paddler_agent/tests/state_overriding_the_chat_template.rs b/paddler_agent/tests/state_overriding_the_chat_template.rs
new file mode 100644
index 000000000..a7f7aae85
--- /dev/null
+++ b/paddler_agent/tests/state_overriding_the_chat_template.rs
@@ -0,0 +1,11 @@
+use paddler_agent::agent_applicable_state::AgentApplicableState;
+use paddler_messaging::chat_template::ChatTemplate;
+use paddler_messaging::chat_template_source::ChatTemplateSource;
+
+pub fn state_overriding_the_chat_template() -> AgentApplicableState {
+    AgentApplicableState::WithoutModel {
+        chat_template_source: ChatTemplateSource::Override(ChatTemplate {
+            content: "{{ messages }}".to_owned(),
+        }),
+    }
+}
diff --git a/paddler_agent/tests/state_whose_model_cannot_load.rs b/paddler_agent/tests/state_whose_model_cannot_load.rs
new file mode 100644
index 000000000..38cfc97ec
--- /dev/null
+++ b/paddler_agent/tests/state_whose_model_cannot_load.rs
@@ -0,0 +1,7 @@
+use paddler_agent::agent_applicable_state::AgentApplicableState;
+
+use crate::settings_whose_model_cannot_load::settings_whose_model_cannot_load;
+
+pub fn state_whose_model_cannot_load() -> AgentApplicableState {
+    AgentApplicableState::WithModel(settings_whose_model_cannot_load())
+}
diff --git a/paddler_agent/tests/the_reconciliation_service_stops_once_its_desired_state_channel_closes.rs b/paddler_agent/tests/the_reconciliation_service_stops_once_its_desired_state_channel_closes.rs
new file mode 100644
index 000000000..ee027532f
--- /dev/null
+++ b/paddler_agent/tests/the_reconciliation_service_stops_once_its_desired_state_channel_closes.rs
@@ -0,0 +1,50 @@
+use std::sync::Arc;
+
+use tokio::sync::watch;
+use tokio_util::sync::CancellationToken;
+use trzcina::Service as _;
+
+use paddler_agent::agent_applicable_state_holder::AgentApplicableStateHolder;
+use paddler_agent::desired_state_reconciler::DesiredStateReconciler;
+use paddler_agent::reconciliation_service::ReconciliationService;
+use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
+use paddler_cache_dir::cache_dir::CacheDir;
+use paddler_download_manager::download_manager::DownloadManager;
+use paddler_messaging::agent_desired_state::AgentDesiredState;
+use paddler_model_source::desired_model_resolver::DesiredModelResolver;
+use paddler_model_source::huggingface_cache_location::HuggingFaceCacheLocation;
+use paddler_model_source::huggingface_default_endpoint::HUGGINGFACE_DEFAULT_ENDPOINT;
+use paddler_model_source::huggingface_hub::HuggingFaceHub;
+
+#[tokio::test]
+async fn the_reconciliation_service_stops_once_its_desired_state_channel_closes() {
+    let (agent_desired_state_tx, agent_desired_state_rx) =
+        watch::channel(AgentDesiredState::default());
+
+    drop(agent_desired_state_tx);
+
+    Box::new(ReconciliationService {
+        agent_applicable_state_holder: Arc::new(AgentApplicableStateHolder::default()),
+        agent_desired_state_rx,
+        desired_state_reconciler: DesiredStateReconciler {
+            desired_model_resolver: DesiredModelResolver {
+                cache_dir: CacheDir {
+                    explicit: None,
+                    home: None,
+                    xdg: None,
+                },
+                download_manager: DownloadManager::new().expect("the model downloader must start"),
+                huggingface_client: HuggingFaceHub {
+                    cache_location: HuggingFaceCacheLocation::HubDefault,
+                    endpoint: HUGGINGFACE_DEFAULT_ENDPOINT.to_owned(),
+                }
+                .client()
+                .expect("the Hugging Face client must start"),
+            },
+            slot_aggregated_status: Arc::new(SlotAggregatedStatus::new(1)),
+        },
+    })
+    .run(CancellationToken::new())
+    .await
+    .expect("the reconciliation service must stop cleanly once nothing can send it desired states");
+}
diff --git a/paddler_agent_decision/Cargo.toml b/paddler_agent_decision/Cargo.toml
new file mode 100644
index 000000000..93e305b8c
--- /dev/null
+++ b/paddler_agent_decision/Cargo.toml
@@ -0,0 +1,30 @@
+[package]
+name = "paddler_agent_decision"
+authors.workspace = true
+description.workspace = true
+edition.workspace = true
+homepage.workspace = true
+license.workspace = true
+repository.workspace = true
+version.workspace = true
+
+[dependencies]
+llama-cpp-bindings = { workspace = true }
+log = { workspace = true }
+thiserror = { workspace = true }
+tokio = { workspace = true }
+tokio-util = { workspace = true }
+paddler_agent_pointer_head = { workspace = true }
+paddler_agent_runtime = { workspace = true }
+paddler_agent_status = { workspace = true }
+paddler_inference_parameters = { workspace = true }
+paddler_messaging = { workspace = true }
+
+[features]
+default = []
+cuda = ["llama-cpp-bindings/cuda", "paddler_agent_pointer_head/cuda", "paddler_agent_runtime/cuda"]
+metal = ["llama-cpp-bindings/metal", "paddler_agent_pointer_head/metal", "paddler_agent_runtime/metal"]
+vulkan = ["llama-cpp-bindings/vulkan", "paddler_agent_pointer_head/vulkan", "paddler_agent_runtime/vulkan"]
+
+[lints]
+workspace = true
diff --git a/paddler_agent_decision/src/active_decision_request.rs b/paddler_agent_decision/src/active_decision_request.rs
new file mode 100644
index 000000000..ace8525ec
--- /dev/null
+++ b/paddler_agent_decision/src/active_decision_request.rs
@@ -0,0 +1,249 @@
+use std::collections::VecDeque;
+use std::mem;
+use std::time::Instant;
+
+use llama_cpp_bindings::token::LlamaToken;
+use tokio::sync::mpsc;
+
+use paddler_agent_pointer_head::pointer_head::PointerHead;
+use paddler_agent_runtime::receives_stop_request::ReceivesStopRequest as _;
+use paddler_agent_runtime::sequence_id::SequenceId;
+use paddler_agent_runtime::sequence_id_guard::SequenceIdGuard;
+use paddler_agent_status::slot_guard::SlotGuard;
+use paddler_messaging::decision_answer::DecisionAnswer;
+use paddler_messaging::decision_result::DecisionResult;
+use paddler_messaging::decision_summary::DecisionSummary;
+
+use crate::decision_admission::DecisionAdmission;
+use crate::decision_advance::DecisionAdvance;
+use crate::decision_decoder::DecisionDecoder;
+use crate::decision_error::DecisionError;
+use crate::decision_question_row::DecisionQuestionRow;
+use crate::decision_request_phase::DecisionRequestPhase;
+use crate::decision_token_layout::DecisionTokenLayout;
+use crate::prepared_decision_request::PreparedDecisionRequest;
+use crate::question_lane::QuestionLane;
+use crate::reports_decision_outcome::ReportsDecisionOutcome;
+
+pub struct ActiveDecisionRequest {
+    pub current_question: DecisionQuestionRow,
+    pub decision_result_tx: mpsc::UnboundedSender<DecisionResult>,
+    pub decision_stop_rx: mpsc::UnboundedReceiver<()>,
+    pub hidden_states: Vec<Vec<f32>>,
+    pub input_tokens: usize,
+    pub phase: DecisionRequestPhase,
+    pub question_lane: QuestionLane,
+    pub remaining_questions: VecDeque<DecisionQuestionRow>,
+    pub slot_guard: SlotGuard,
+    pub started_at: Instant,
+    pub state: Vec<LlamaToken>,
+    pub state_sequence: SequenceIdGuard,
+}
+
+impl ActiveDecisionRequest {
+    #[must_use]
+    pub fn new(
+        PreparedDecisionRequest {
+            decision_result_tx,
+            decision_stop_rx,
+            layout,
+            slot_guard,
+            started_at,
+        }: PreparedDecisionRequest,
+        DecisionAdmission {
+            question_lane,
+            state_sequence,
+        }: DecisionAdmission,
+    ) -> Self {
+        let input_tokens = layout.total_tokens();
+        let DecisionTokenLayout {
+            first_question,
+            remaining_questions,
+            state,
+        } = layout;
+
+        Self {
+            current_question: first_question,
+            decision_result_tx,
+            decision_stop_rx,
+            hidden_states: Vec::new(),
+            input_tokens,
+            phase: DecisionRequestPhase::IngestingState { next_offset: 0 },
+            question_lane,
+            remaining_questions,
+            slot_guard,
+            started_at,
+            state,
+            state_sequence,
+        }
+    }
+
+    pub fn advance(
+        &mut self,
+        decoder: &mut DecisionDecoder,
+        pointer_head: &PointerHead,
+        agent_name: Option<&str>,
+    ) -> Result<DecisionAdvance, DecisionError> {
+        match self.phase {
+            DecisionRequestPhase::IngestingState { next_offset } => {
+                let next_offset = decoder.ingest_state_chunk(
+                    self.state_sequence.sequence_id(),
+                    &self.state,
+                    next_offset,
+                )?;
+
+                self.phase = if next_offset < self.state.len() {
+                    DecisionRequestPhase::IngestingState { next_offset }
+                } else {
+                    DecisionRequestPhase::AnsweringQuestion { next_offset: 0 }
+                };
+
+                Ok(DecisionAdvance::Continuing)
+            }
+            DecisionRequestPhase::AnsweringQuestion { next_offset } => match &self.question_lane {
+                QuestionLane::Reserved(lane) => {
+                    let lane_sequence_id = lane.sequence_id();
+
+                    self.answer_leading_question(
+                        decoder,
+                        pointer_head,
+                        agent_name,
+                        lane_sequence_id,
+                        next_offset,
+                    )
+                }
+                QuestionLane::Released => {
+                    self.answer_last_question(decoder, pointer_head, agent_name, next_offset)
+                }
+            },
+        }
+    }
+
+    pub fn finish_with<TOutcome: ReportsDecisionOutcome>(
+        mut self,
+        decoder: &mut DecisionDecoder,
+        agent_name: Option<&str>,
+        outcome: TOutcome,
+    ) {
+        self.state_sequence
+            .sequence_id()
+            .release_kv_cache_in(decoder.llama_context, agent_name);
+        self.release_question_lane(decoder, agent_name);
+
+        outcome.report(agent_name, &self.decision_result_tx);
+    }
+
+    pub fn is_stop_requested(&mut self) -> bool {
+        self.decision_stop_rx.is_stop_requested()
+    }
+
+    #[must_use]
+    pub fn summary(&self) -> DecisionResult {
+        DecisionResult::Done(DecisionSummary {
+            input_tokens: self.input_tokens,
+            processing_milliseconds: self.started_at.elapsed().as_millis() as u64,
+        })
+    }
+
+    fn answer_last_question(
+        &mut self,
+        decoder: &mut DecisionDecoder,
+        pointer_head: &PointerHead,
+        agent_name: Option<&str>,
+        next_offset: usize,
+    ) -> Result<DecisionAdvance, DecisionError> {
+        let next_offset =
+            self.decode_question_chunk(decoder, self.state_sequence.sequence_id(), next_offset)?;
+
+        if next_offset < self.current_question.tokens.len() {
+            self.phase = DecisionRequestPhase::AnsweringQuestion { next_offset };
+
+            return Ok(DecisionAdvance::Continuing);
+        }
+
+        self.send_answer(pointer_head)?;
+
+        Ok(self.finish_question(decoder, agent_name))
+    }
+
+    fn answer_leading_question(
+        &mut self,
+        decoder: &mut DecisionDecoder,
+        pointer_head: &PointerHead,
+        agent_name: Option<&str>,
+        lane_sequence_id: SequenceId,
+        next_offset: usize,
+    ) -> Result<DecisionAdvance, DecisionError> {
+        if next_offset == 0 {
+            decoder.copy_sequence(self.state_sequence.sequence_id(), lane_sequence_id)?;
+        }
+
+        let next_offset = self.decode_question_chunk(decoder, lane_sequence_id, next_offset)?;
+
+        if next_offset < self.current_question.tokens.len() {
+            self.phase = DecisionRequestPhase::AnsweringQuestion { next_offset };
+
+            return Ok(DecisionAdvance::Continuing);
+        }
+
+        self.send_answer(pointer_head)?;
+
+        Ok(self.finish_question(decoder, agent_name))
+    }
+
+    fn decode_question_chunk(
+        &mut self,
+        decoder: &mut DecisionDecoder,
+        sequence_id: SequenceId,
+        next_offset: usize,
+    ) -> Result<usize, DecisionError> {
+        decoder.decode_row_chunk(
+            sequence_id,
+            &self.current_question,
+            next_offset,
+            self.state.len(),
+            &mut self.hidden_states,
+        )
+    }
+
+    fn finish_question(
+        &mut self,
+        decoder: &mut DecisionDecoder,
+        agent_name: Option<&str>,
+    ) -> DecisionAdvance {
+        let Some(next_question) = self.remaining_questions.pop_front() else {
+            return DecisionAdvance::Completed;
+        };
+
+        self.current_question = next_question;
+
+        if self.remaining_questions.is_empty() {
+            self.release_question_lane(decoder, agent_name);
+        }
+
+        self.phase = DecisionRequestPhase::AnsweringQuestion { next_offset: 0 };
+
+        DecisionAdvance::Continuing
+    }
+
+    fn release_question_lane(&mut self, decoder: &mut DecisionDecoder, agent_name: Option<&str>) {
+        if let QuestionLane::Reserved(lane) =
+            mem::replace(&mut self.question_lane, QuestionLane::Released)
+        {
+            lane.sequence_id()
+                .release_kv_cache_in(decoder.llama_context, agent_name);
+        }
+    }
+
+    fn send_answer(&mut self, pointer_head: &PointerHead) -> Result<(), DecisionError> {
+        let probabilities = pointer_head.option_probabilities(&self.hidden_states);
+
+        self.hidden_states.clear();
+        self.decision_result_tx
+            .send(DecisionResult::QuestionAnswered(DecisionAnswer {
+                id: self.current_question.id.clone(),
+                probabilities,
+            }))
+            .map_err(DecisionError::ClientDisconnected)
+    }
+}
diff --git a/paddler_agent_decision/src/decision_admission.rs b/paddler_agent_decision/src/decision_admission.rs
new file mode 100644
index 000000000..7385e94ba
--- /dev/null
+++ b/paddler_agent_decision/src/decision_admission.rs
@@ -0,0 +1,8 @@
+use paddler_agent_runtime::sequence_id_guard::SequenceIdGuard;
+
+use crate::question_lane::QuestionLane;
+
+pub struct DecisionAdmission {
+    pub question_lane: QuestionLane,
+    pub state_sequence: SequenceIdGuard,
+}
diff --git a/paddler_agent_decision/src/decision_advance.rs b/paddler_agent_decision/src/decision_advance.rs
new file mode 100644
index 000000000..a64b25873
--- /dev/null
+++ b/paddler_agent_decision/src/decision_advance.rs
@@ -0,0 +1,4 @@
+pub enum DecisionAdvance {
+    Completed,
+    Continuing,
+}
diff --git a/paddler_agent_decision/src/decision_capacity_ledger.rs b/paddler_agent_decision/src/decision_capacity_ledger.rs
new file mode 100644
index 000000000..93ee0d5e8
--- /dev/null
+++ b/paddler_agent_decision/src/decision_capacity_ledger.rs
@@ -0,0 +1,70 @@
+use paddler_agent_runtime::sequence_id_guard::SequenceIdGuard;
+use paddler_agent_runtime::sequence_id_pool::SequenceIdPool;
+
+use crate::decision_admission::DecisionAdmission;
+use crate::question_lane::QuestionLane;
+use crate::question_lane_demand::QuestionLaneDemand;
+
+pub struct DecisionCapacityLedger {
+    sequence_id_pool: SequenceIdPool,
+}
+
+impl DecisionCapacityLedger {
+    #[must_use]
+    pub fn new(sequences: u16) -> Self {
+        Self {
+            sequence_id_pool: SequenceIdPool::new(sequences),
+        }
+    }
+
+    #[must_use]
+    pub fn admit(&self, question_lane_demand: QuestionLaneDemand) -> Option<DecisionAdmission> {
+        let state_sequence = SequenceIdGuard::acquire(&self.sequence_id_pool)?;
+        let question_lane = match question_lane_demand {
+            QuestionLaneDemand::Needed => {
+                QuestionLane::Reserved(SequenceIdGuard::acquire(&self.sequence_id_pool)?)
+            }
+            QuestionLaneDemand::NotNeeded => QuestionLane::Released,
+        };
+
+        Some(DecisionAdmission {
+            question_lane,
+            state_sequence,
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::DecisionCapacityLedger;
+    use crate::question_lane_demand::QuestionLaneDemand;
+
+    #[test]
+    fn a_single_question_decision_takes_only_its_state_sequence() {
+        let ledger = DecisionCapacityLedger::new(2);
+        let _first_admission = ledger
+            .admit(QuestionLaneDemand::NotNeeded)
+            .expect("one free sequence fits a single question");
+
+        assert!(ledger.admit(QuestionLaneDemand::NotNeeded).is_some());
+    }
+
+    #[test]
+    fn a_multi_question_decision_waits_for_a_free_question_lane() {
+        let ledger = DecisionCapacityLedger::new(1);
+
+        assert!(ledger.admit(QuestionLaneDemand::Needed).is_none());
+    }
+
+    #[test]
+    fn a_decision_runs_once_the_sequences_of_a_finished_one_are_returned() {
+        let ledger = DecisionCapacityLedger::new(2);
+        let first_admission = ledger.admit(QuestionLaneDemand::Needed);
+
+        assert!(ledger.admit(QuestionLaneDemand::NotNeeded).is_none());
+
+        drop(first_admission);
+
+        assert!(ledger.admit(QuestionLaneDemand::NotNeeded).is_some());
+    }
+}
diff --git a/paddler_agent_decision/src/decision_decoder.rs b/paddler_agent_decision/src/decision_decoder.rs
new file mode 100644
index 000000000..2fc80368e
--- /dev/null
+++ b/paddler_agent_decision/src/decision_decoder.rs
@@ -0,0 +1,134 @@
+use llama_cpp_bindings::SampledToken;
+use llama_cpp_bindings::context::LlamaContext;
+use llama_cpp_bindings::llama_batch::LlamaBatch;
+use llama_cpp_bindings::token::LlamaToken;
+
+use paddler_agent_runtime::sequence_id::SequenceId;
+
+use crate::decision_error::DecisionError;
+use crate::decision_question_row::DecisionQuestionRow;
+
+pub struct DecisionDecoder<'scheduler, 'model> {
+    pub batch: &'scheduler mut LlamaBatch<'static>,
+    pub llama_context: &'scheduler mut LlamaContext<'model>,
+    pub n_batch: usize,
+}
+
+impl DecisionDecoder<'_, '_> {
+    pub fn copy_sequence(
+        &mut self,
+        source_sequence_id: SequenceId,
+        destination_sequence_id: SequenceId,
+    ) -> Result<(), DecisionError> {
+        self.remove_sequence(destination_sequence_id)?;
+        self.llama_context
+            .copy_kv_cache_seq(
+                source_sequence_id.llama_sequence_id(),
+                destination_sequence_id.llama_sequence_id(),
+                None,
+                None,
+            )
+            .map_err(DecisionError::KvCacheCopyFailed)
+    }
+
+    pub fn ingest_state_chunk(
+        &mut self,
+        sequence_id: SequenceId,
+        state: &[LlamaToken],
+        next_offset: usize,
+    ) -> Result<usize, DecisionError> {
+        self.batch.clear();
+
+        for (position, token) in (0_i32..).zip(state).skip(next_offset).take(self.n_batch) {
+            self.batch
+                .add(
+                    &SampledToken::Content(*token),
+                    position,
+                    &[sequence_id.llama_sequence_id()],
+                    false,
+                )
+                .map_err(DecisionError::BatchAssemblyFailed)?;
+        }
+
+        self.llama_context
+            .decode(self.batch)
+            .map_err(DecisionError::DecodeFailed)?;
+
+        Ok(next_offset + self.batch.n_tokens() as usize)
+    }
+
+    pub fn decode_row_chunk(
+        &mut self,
+        sequence_id: SequenceId,
+        row: &DecisionQuestionRow,
+        next_offset: usize,
+        state_length: usize,
+        hidden_states: &mut Vec<Vec<f32>>,
+    ) -> Result<usize, DecisionError> {
+        self.batch.clear();
+
+        let mut output_batch_indices = Vec::new();
+
+        let row_chunk = row
+            .tokens
+            .iter()
+            .enumerate()
+            .skip(next_offset)
+            .take(self.n_batch);
+        let row_chunk_positions = (0_i32..).skip(state_length + next_offset);
+
+        for ((batch_index, position), (row_offset, token)) in
+            (0_i32..).zip(row_chunk_positions).zip(row_chunk)
+        {
+            let is_output = row.is_output(row_offset);
+
+            self.batch
+                .add(
+                    &SampledToken::Content(*token),
+                    position,
+                    &[sequence_id.llama_sequence_id()],
+                    is_output,
+                )
+                .map_err(DecisionError::BatchAssemblyFailed)?;
+
+            if is_output {
+                output_batch_indices.push(batch_index);
+            }
+        }
+
+        self.llama_context.set_embeddings(true);
+
+        let row_hidden_states = self.decode_row_hidden_states(&output_batch_indices);
+
+        self.llama_context.set_embeddings(false);
+
+        hidden_states.extend(row_hidden_states?);
+
+        Ok(next_offset + self.batch.n_tokens() as usize)
+    }
+
+    fn decode_row_hidden_states(
+        &mut self,
+        output_batch_indices: &[i32],
+    ) -> Result<Vec<Vec<f32>>, DecisionError> {
+        self.llama_context
+            .decode(self.batch)
+            .map_err(DecisionError::DecodeFailed)?;
+
+        output_batch_indices
+            .iter()
+            .map(|batch_index| {
+                self.llama_context
+                    .embeddings_ith(*batch_index)
+                    .map(<[f32]>::to_vec)
+                    .map_err(DecisionError::HiddenStateUnavailable)
+            })
+            .collect()
+    }
+
+    pub fn remove_sequence(&mut self, sequence_id: SequenceId) -> Result<(), DecisionError> {
+        sequence_id
+            .clear_kv_cache_in(self.llama_context)
+            .map_err(DecisionError::KvCacheRemovalFailed)
+    }
+}
diff --git a/paddler_agent_decision/src/decision_delimiter_tokens.rs b/paddler_agent_decision/src/decision_delimiter_tokens.rs
new file mode 100644
index 000000000..c623c0409
--- /dev/null
+++ b/paddler_agent_decision/src/decision_delimiter_tokens.rs
@@ -0,0 +1,142 @@
+use llama_cpp_bindings::llama_token_attr::LlamaTokenAttr;
+use llama_cpp_bindings::token::LlamaToken;
+
+use paddler_agent_pointer_head::pointer_head_delimiters::PointerHeadDelimiters;
+use paddler_messaging::agent_issue_params::pointer_head_incompatibility::PointerHeadIncompatibility;
+
+use crate::decision_error::DecisionError;
+use crate::delimiter_vocabulary::DelimiterVocabulary;
+
+fn resolve_control_token<TVocabulary: DelimiterVocabulary>(
+    vocabulary: &TVocabulary,
+    delimiter: &str,
+) -> Result<LlamaToken, DecisionError> {
+    let not_a_control_token = || DecisionError::PointerHeadIncompatibleWithModel {
+        incompatibility: PointerHeadIncompatibility::DelimiterIsNotAControlToken {
+            delimiter: delimiter.to_owned(),
+        },
+    };
+    let tokens = vocabulary.tokenize_delimiter(delimiter).map_err(|source| {
+        DecisionError::DelimiterTokenizationFailed {
+            delimiter: delimiter.to_owned(),
+            source,
+        }
+    })?;
+    let control_token = match tokens.as_slice() {
+        [token] => vocabulary
+            .token_attributes(*token)
+            .map_err(|source| DecisionError::DelimiterAttributesUnreadable {
+                delimiter: delimiter.to_owned(),
+                source,
+            })?
+            .contains(LlamaTokenAttr::Control)
+            .then_some(*token),
+        _ => None,
+    };
+
+    control_token.ok_or_else(not_a_control_token)
+}
+
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+pub struct DecisionDelimiterTokens {
+    pub decide: LlamaToken,
+    pub option_end: LlamaToken,
+    pub option_start: LlamaToken,
+    pub question: LlamaToken,
+    pub state: LlamaToken,
+}
+
+impl DecisionDelimiterTokens {
+    pub fn resolve<TVocabulary: DelimiterVocabulary>(
+        vocabulary: &TVocabulary,
+        PointerHeadDelimiters {
+            decide,
+            option_end,
+            option_start,
+            question,
+            state,
+        }: &PointerHeadDelimiters,
+    ) -> Result<Self, DecisionError> {
+        let mut control_tokens = [LlamaToken::new(0); 5];
+
+        for (control_token, delimiter) in
+            control_tokens
+                .iter_mut()
+                .zip([decide, option_end, option_start, question, state])
+        {
+            *control_token = resolve_control_token(vocabulary, delimiter)?;
+        }
+
+        let [decide, option_end, option_start, question, state] = control_tokens;
+
+        Ok(Self {
+            decide,
+            option_end,
+            option_start,
+            question,
+            state,
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use llama_cpp_bindings::error::StringToTokenError;
+    use llama_cpp_bindings::error::TokenAttrError;
+    use llama_cpp_bindings::llama_token_attrs::LlamaTokenAttrs;
+    use llama_cpp_bindings::llama_token_attrs_from_int_error::LlamaTokenAttrsFromIntError;
+    use llama_cpp_bindings::token::LlamaToken;
+
+    use paddler_agent_pointer_head::pointer_head_delimiters::PointerHeadDelimiters;
+
+    use super::DecisionDelimiterTokens;
+    use crate::decision_error::DecisionError;
+    use crate::delimiter_vocabulary::DelimiterVocabulary;
+
+    const UNKNOWN_ATTRIBUTE_BITS: u32 = 1 << 30;
+
+    struct VocabularyWithUnreadableAttributes;
+
+    impl DelimiterVocabulary for VocabularyWithUnreadableAttributes {
+        fn tokenize_delimiter(
+            &self,
+            _delimiter: &str,
+        ) -> Result<Vec<LlamaToken>, StringToTokenError> {
+            Ok(vec![LlamaToken::new(0)])
+        }
+
+        fn token_attributes(&self, _token: LlamaToken) -> Result<LlamaTokenAttrs, TokenAttrError> {
+            Err(TokenAttrError::UnknownAttributes(
+                LlamaTokenAttrsFromIntError::UnknownValue(UNKNOWN_ATTRIBUTE_BITS),
+            ))
+        }
+    }
+
+    #[test]
+    fn a_delimiter_whose_token_attributes_are_unreadable_is_refused() {
+        assert_eq!(
+            DecisionDelimiterTokens::resolve(
+                &VocabularyWithUnreadableAttributes,
+                &PointerHeadDelimiters {
+                    decide: "<decide>".to_owned(),
+                    option_end: "<option_end>".to_owned(),
+                    option_start: "<option_start>".to_owned(),
+                    question: "<question>".to_owned(),
+                    state: "<state>".to_owned(),
+                },
+            )
+            .err()
+            .as_ref()
+            .map(ToString::to_string),
+            Some(
+                DecisionError::DelimiterAttributesUnreadable {
+                    delimiter: "<decide>".to_owned(),
+                    source: TokenAttrError::UnknownAttributes(
+                        LlamaTokenAttrsFromIntError::UnknownValue(UNKNOWN_ATTRIBUTE_BITS),
+                    ),
+                }
+                .to_string()
+            )
+        );
+    }
+}
diff --git a/paddler_agent_decision/src/decision_error.rs b/paddler_agent_decision/src/decision_error.rs
new file mode 100644
index 000000000..0f3ed1a7e
--- /dev/null
+++ b/paddler_agent_decision/src/decision_error.rs
@@ -0,0 +1,234 @@
+use llama_cpp_bindings::batch_add_error::BatchAddError;
+use llama_cpp_bindings::error::ClearKvCacheSeqError;
+use llama_cpp_bindings::error::CopyKvCacheSeqError;
+use llama_cpp_bindings::error::DecodeError;
+use llama_cpp_bindings::error::EmbeddingsError;
+use llama_cpp_bindings::error::StringToTokenError;
+use llama_cpp_bindings::error::TokenAttrError;
+use log::error;
+use tokio::sync::mpsc;
+use tokio::sync::mpsc::error::SendError;
+
+use paddler_agent_pointer_head::pointer_head_error::PointerHeadError;
+use paddler_agent_runtime::agent_runtime_error::AgentRuntimeError;
+use paddler_agent_runtime::rejection_description::rejection_description;
+use paddler_agent_runtime::send_result_or_warn::send_result_or_warn;
+use paddler_messaging::agent_issue_params::pointer_head_incompatibility::PointerHeadIncompatibility;
+use paddler_messaging::decision_result::DecisionResult;
+use paddler_messaging::oversized_decision_details::OversizedDecisionDetails;
+
+use crate::reports_decision_outcome::ReportsDecisionOutcome;
+
+#[derive(Debug, thiserror::Error)]
+pub enum DecisionError {
+    #[error(transparent)]
+    AgentRuntime(#[from] AgentRuntimeError),
+
+    #[error("failed to add a decision token to the batch: {0}")]
+    BatchAssemblyFailed(#[source] BatchAddError),
+
+    #[error("the client stopped listening for decisions: {0}")]
+    ClientDisconnected(#[source] SendError<DecisionResult>),
+
+    #[error("no pointer head is configured, so the agent serves no decisions")]
+    DecisionsNotServed,
+
+    #[error("llama.cpp rejected the decision batch: {0}")]
+    DecodeFailed(#[source] DecodeError),
+
+    #[error("unable to read the attributes of the pointer head delimiter {delimiter}: {source}")]
+    DelimiterAttributesUnreadable {
+        delimiter: String,
+        #[source]
+        source: TokenAttrError,
+    },
+
+    #[error("unable to tokenize the pointer head delimiter {delimiter}: {source}")]
+    DelimiterTokenizationFailed {
+        delimiter: String,
+        #[source]
+        source: StringToTokenError,
+    },
+
+    #[error("unable to read the hidden state of a decision output: {0}")]
+    HiddenStateUnavailable(#[source] EmbeddingsError),
+
+    #[error("failed to tokenize the decision input: {0}")]
+    InputTokenizationFailed(#[source] StringToTokenError),
+
+    #[error("failed to copy the decision state into its question lane: {0}")]
+    KvCacheCopyFailed(#[source] CopyKvCacheSeqError),
+
+    #[error("failed to remove a decision sequence from the KV cache: {0}")]
+    KvCacheRemovalFailed(#[source] ClearKvCacheSeqError),
+
+    #[error("no model is loaded")]
+    ModelNotLoaded,
+
+    #[error("unable to load the pointer head: {0}")]
+    PointerHeadCannotBeLoaded(#[source] PointerHeadError),
+
+    #[error("the pointer head does not fit the model: {incompatibility:?}")]
+    PointerHeadIncompatibleWithModel {
+        incompatibility: PointerHeadIncompatibility,
+    },
+
+    #[error(
+        "the decision needs {} tokens in the cache, more than the {} the context holds",
+        details.required_tokens,
+        details.context_size
+    )]
+    RequestExceedsContext { details: OversizedDecisionDetails },
+
+    #[error("the scheduler is no longer accepting requests")]
+    SchedulerUnavailable,
+
+    #[error("decisions need {required_slots} slots, the agent runs {desired_slots}")]
+    SlotsInsufficient {
+        desired_slots: u16,
+        required_slots: u16,
+    },
+}
+
+impl ReportsDecisionOutcome for DecisionError {
+    fn report(
+        self,
+        agent_name: Option<&str>,
+        decision_result_tx: &mpsc::UnboundedSender<DecisionResult>,
+    ) {
+        let message = rejection_description(agent_name, &self);
+
+        error!("{message}");
+
+        let result = match self {
+            Self::AgentRuntime(_)
+            | Self::DelimiterAttributesUnreadable { .. }
+            | Self::DelimiterTokenizationFailed { .. }
+            | Self::ModelNotLoaded
+            | Self::PointerHeadCannotBeLoaded(_)
+            | Self::PointerHeadIncompatibleWithModel { .. }
+            | Self::SlotsInsufficient { .. } => DecisionResult::ModelNotLoaded(message),
+            Self::BatchAssemblyFailed(_) => DecisionResult::BatchAssemblyFailed(message),
+            Self::ClientDisconnected(_) => return,
+            Self::DecisionsNotServed => DecisionResult::DecisionsNotServed(message),
+            Self::DecodeFailed(_) => DecisionResult::DecodeFailed(message),
+            Self::HiddenStateUnavailable(_) => DecisionResult::HiddenStateUnavailable(message),
+            Self::InputTokenizationFailed(_) => DecisionResult::InputTokenizationFailed(message),
+            Self::KvCacheCopyFailed(_) => DecisionResult::KvCacheCopyFailed(message),
+            Self::KvCacheRemovalFailed(_) => DecisionResult::KvCacheRemovalFailed(message),
+            Self::RequestExceedsContext { details } => {
+                DecisionResult::RequestExceedsContext(details)
+            }
+            Self::SchedulerUnavailable => DecisionResult::SchedulerUnavailable(message),
+        };
+
+        send_result_or_warn(agent_name, decision_result_tx, result);
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::ffi::CString;
+    use std::mem::discriminant;
+
+    use llama_cpp_bindings::batch_add_error::BatchAddError;
+    use llama_cpp_bindings::error::ClearKvCacheSeqError;
+    use llama_cpp_bindings::error::CopyKvCacheSeqError;
+    use llama_cpp_bindings::error::DecodeError;
+    use llama_cpp_bindings::error::EmbeddingsError;
+    use llama_cpp_bindings::error::StringToTokenError;
+    use tokio::sync::mpsc;
+    use tokio::sync::mpsc::error::SendError;
+
+    use paddler_messaging::decision_result::DecisionResult;
+    use paddler_messaging::oversized_decision_details::OversizedDecisionDetails;
+
+    use super::DecisionError;
+    use crate::reports_decision_outcome::ReportsDecisionOutcome as _;
+
+    fn reported(decision_error: DecisionError) -> Option<DecisionResult> {
+        let (decision_result_tx, mut decision_result_rx) = mpsc::unbounded_channel();
+
+        decision_error.report(None, &decision_result_tx);
+
+        decision_result_rx.try_recv().ok()
+    }
+
+    #[test]
+    fn every_request_failure_reports_its_own_decision_result() {
+        let description = String::new;
+        let oversized = OversizedDecisionDetails {
+            context_size: 256,
+            required_tokens: 900,
+        };
+        let failures = [
+            (
+                DecisionError::BatchAssemblyFailed(BatchAddError::InsufficientSpace(1)),
+                DecisionResult::BatchAssemblyFailed(description()),
+            ),
+            (
+                DecisionError::DecisionsNotServed,
+                DecisionResult::DecisionsNotServed(description()),
+            ),
+            (
+                DecisionError::DecodeFailed(DecodeError::NoKvCacheSlot),
+                DecisionResult::DecodeFailed(description()),
+            ),
+            (
+                DecisionError::HiddenStateUnavailable(EmbeddingsError::LogitsNotEnabled),
+                DecisionResult::HiddenStateUnavailable(description()),
+            ),
+            (
+                DecisionError::InputTokenizationFailed(StringToTokenError::NulError(
+                    CString::new("a \0 byte").unwrap_err(),
+                )),
+                DecisionResult::InputTokenizationFailed(description()),
+            ),
+            (
+                DecisionError::KvCacheCopyFailed(CopyKvCacheSeqError::MemoryHandleUnavailable),
+                DecisionResult::KvCacheCopyFailed(description()),
+            ),
+            (
+                DecisionError::KvCacheRemovalFailed(
+                    ClearKvCacheSeqError::PartialSequenceNotRemoved {
+                        seq_id: 0,
+                        p0: -1,
+                        p1: -1,
+                    },
+                ),
+                DecisionResult::KvCacheRemovalFailed(description()),
+            ),
+            (
+                DecisionError::ModelNotLoaded,
+                DecisionResult::ModelNotLoaded(description()),
+            ),
+            (
+                DecisionError::RequestExceedsContext {
+                    details: oversized.clone(),
+                },
+                DecisionResult::RequestExceedsContext(oversized),
+            ),
+            (
+                DecisionError::SchedulerUnavailable,
+                DecisionResult::SchedulerUnavailable(description()),
+            ),
+        ];
+
+        for (decision_error, expected_result) in failures {
+            assert_eq!(
+                reported(decision_error).as_ref().map(discriminant),
+                Some(discriminant(&expected_result))
+            );
+        }
+    }
+
+    #[test]
+    fn a_disconnected_client_is_not_sent_a_result() {
+        assert_eq!(
+            reported(DecisionError::ClientDisconnected(SendError(
+                DecisionResult::SchedulerUnavailable(String::new())
+            ))),
+            None
+        );
+    }
+}
diff --git a/paddler_agent_decision/src/decision_llama_context_settings.rs b/paddler_agent_decision/src/decision_llama_context_settings.rs
new file mode 100644
index 000000000..4be105de7
--- /dev/null
+++ b/paddler_agent_decision/src/decision_llama_context_settings.rs
@@ -0,0 +1,30 @@
+use llama_cpp_bindings::context::params::LlamaPoolingType;
+
+use paddler_agent_runtime::inference_runtime_context::InferenceRuntimeContext;
+use paddler_agent_runtime::llama_context_settings::LlamaContextSettings;
+use paddler_agent_runtime::loaded_llama_model::LoadedLlamaModel;
+use paddler_agent_runtime::resident_llama_context::ResidentLlamaContext;
+use paddler_messaging::inference_capability::InferenceCapability;
+
+use crate::decision_error::DecisionError;
+
+pub struct DecisionLlamaContextSettings<'parameters> {
+    pub llama_context_settings: LlamaContextSettings<'parameters>,
+}
+
+impl DecisionLlamaContextSettings<'_> {
+    pub fn create_llama_context<'model>(
+        self,
+        loaded_llama_model: &'model LoadedLlamaModel,
+        inference_runtime_context: &InferenceRuntimeContext,
+    ) -> Result<ResidentLlamaContext<'model>, DecisionError> {
+        Ok(loaded_llama_model.create_llama_context(
+            inference_runtime_context,
+            InferenceCapability::Decisions,
+            self.llama_context_settings
+                .into_llama_context_params()
+                .with_kv_unified(true)
+                .with_pooling_type(LlamaPoolingType::None),
+        )?)
+    }
+}
diff --git a/paddler_agent_decision/src/decision_pipeline.rs b/paddler_agent_decision/src/decision_pipeline.rs
new file mode 100644
index 000000000..f644721c6
--- /dev/null
+++ b/paddler_agent_decision/src/decision_pipeline.rs
@@ -0,0 +1,191 @@
+use std::collections::VecDeque;
+use std::sync::Arc;
+
+use llama_cpp_bindings::llama_batch::LlamaBatch;
+use tokio::sync::mpsc::UnboundedSender;
+use tokio::sync::mpsc::unbounded_channel;
+use tokio_util::sync::CancellationToken;
+use tokio_util::task::TaskTracker;
+
+use paddler_agent_runtime::continue_startup_unless_shutting_down::continue_startup_unless_shutting_down;
+use paddler_agent_runtime::inference_runtime_context::InferenceRuntimeContext;
+use paddler_agent_runtime::loaded_llama_model::LoadedLlamaModel;
+use paddler_agent_runtime::run_scheduler_unless_startup_abandoned::run_scheduler_unless_startup_abandoned;
+use paddler_agent_runtime::scheduler_handle::SchedulerHandle;
+use paddler_agent_runtime::scheduler_inbox::SchedulerInbox;
+use paddler_agent_runtime::scheduler_ready::SchedulerReady;
+use paddler_agent_runtime::scheduler_request_preparer::SchedulerRequestPreparer;
+use paddler_agent_runtime::start_scheduler_thread::start_scheduler_thread;
+use paddler_agent_runtime::startup_outcome::StartupOutcome;
+use paddler_agent_status::agent_issue_fix::AgentIssueFix;
+use paddler_inference_parameters::model_runtime_parameters::ModelRuntimeParameters;
+use paddler_messaging::agent_issue::AgentIssue;
+use paddler_messaging::agent_issue_params::model_path::ModelPath;
+use paddler_messaging::agent_issue_params::pointer_head_incompatibility::PointerHeadIncompatibility;
+use paddler_messaging::agent_issue_params::pointer_head_incompatible_with_model_params::PointerHeadIncompatibleWithModelParams;
+
+use crate::decision_capacity_ledger::DecisionCapacityLedger;
+use crate::decision_delimiter_tokens::DecisionDelimiterTokens;
+use crate::decision_error::DecisionError;
+use crate::decision_llama_context_settings::DecisionLlamaContextSettings;
+use crate::decision_request_preparer::DecisionRequestPreparer;
+use crate::decision_scheduler::DecisionScheduler;
+use crate::decision_scheduler_context::DecisionSchedulerContext;
+use crate::decision_text_tokenizer::DecisionTextTokenizer;
+use crate::loaded_pointer_head::LoadedPointerHead;
+
+pub struct DecisionPipeline {
+    pub inference_runtime_context: InferenceRuntimeContext,
+    pub loaded_llama_model: LoadedLlamaModel,
+    pub loaded_pointer_head: LoadedPointerHead,
+    pub model_runtime_parameters: ModelRuntimeParameters,
+}
+
+impl DecisionPipeline {
+    pub async fn spawn(
+        self,
+        cancellation_token: &CancellationToken,
+    ) -> Result<StartupOutcome<SchedulerHandle<DecisionRequestPreparer>>, DecisionError> {
+        start_scheduler_thread(
+            cancellation_token,
+            move |agent_shutdown, scheduler_startup_tx| {
+                self.run_scheduler(agent_shutdown, scheduler_startup_tx)
+            },
+        )
+        .await
+    }
+
+    fn register_compatibility(&self) {
+        self.inference_runtime_context
+            .slot_aggregated_status
+            .register_fix(&AgentIssueFix::PointerHeadIsCompatibleWithModel(
+                ModelPath::from(self.loaded_pointer_head.pointer_head_path.as_path()),
+            ));
+    }
+
+    fn register_incompatibility(
+        &self,
+        incompatibility: PointerHeadIncompatibility,
+    ) -> DecisionError {
+        self.inference_runtime_context
+            .slot_aggregated_status
+            .register_pipeline_issue(AgentIssue::PointerHeadIncompatibleWithModel(
+                PointerHeadIncompatibleWithModelParams {
+                    incompatibility: incompatibility.clone(),
+                    pointer_head_path: ModelPath::from(
+                        self.loaded_pointer_head.pointer_head_path.as_path(),
+                    ),
+                },
+            ));
+
+        DecisionError::PointerHeadIncompatibleWithModel { incompatibility }
+    }
+
+    fn resolve_delimiter_tokens(&self) -> Result<DecisionDelimiterTokens, DecisionError> {
+        let model_hidden_size = self.loaded_llama_model.model.n_embd_out() as usize;
+        let pointer_head = &self.loaded_pointer_head.pointer_head;
+
+        if pointer_head.hidden_size() != model_hidden_size {
+            return Err(self.register_incompatibility(
+                PointerHeadIncompatibility::HiddenSizeMismatch {
+                    model_hidden_size,
+                    pointer_head_hidden_size: pointer_head.hidden_size(),
+                },
+            ));
+        }
+
+        match DecisionDelimiterTokens::resolve(
+            &self.loaded_llama_model.model.llama_model,
+            &pointer_head.delimiters,
+        ) {
+            Err(DecisionError::PointerHeadIncompatibleWithModel { incompatibility }) => {
+                Err(self.register_incompatibility(incompatibility))
+            }
+            resolution => resolution.inspect(|_delimiter_tokens| self.register_compatibility()),
+        }
+    }
+
+    fn run_scheduler(
+        self,
+        agent_shutdown: &CancellationToken,
+        scheduler_startup_tx: &UnboundedSender<
+            Result<SchedulerReady<DecisionRequestPreparer>, DecisionError>,
+        >,
+    ) -> Result<(), DecisionError> {
+        continue_startup_unless_shutting_down(agent_shutdown, || {
+            let delimiter_tokens = self.resolve_delimiter_tokens()?;
+            let desired_slots_total = self
+                .inference_runtime_context
+                .slot_aggregated_status
+                .desired_slots_total;
+            let n_batch = self.model_runtime_parameters.n_batch.tokens_usize();
+            let llama_context = DecisionLlamaContextSettings {
+                llama_context_settings: self.loaded_llama_model.llama_context_settings(
+                    &self.model_runtime_parameters,
+                    u32::from(desired_slots_total),
+                ),
+            }
+            .create_llama_context(&self.loaded_llama_model, &self.inference_runtime_context)?;
+            let batch = LlamaBatch::new(self.model_runtime_parameters.n_batch.capacity(), 1);
+
+            continue_startup_unless_shutting_down(agent_shutdown, || {
+                self.loaded_llama_model
+                    .warm_up_and_serve(
+                        llama_context,
+                        batch,
+                        desired_slots_total,
+                        |llama_context, batch| {
+                            let cells_per_slot =
+                                llama_context.n_ctx() / u32::from(desired_slots_total);
+                            let (scheduler_message_tx, scheduler_message_rx) = unbounded_channel();
+
+                            run_scheduler_unless_startup_abandoned(
+                                scheduler_startup_tx,
+                                Ok(SchedulerReady {
+                                    request_preparer: Arc::new(SchedulerRequestPreparer {
+                                        agent_name: self
+                                            .inference_runtime_context
+                                            .agent_name
+                                            .clone(),
+                                        preparation: DecisionRequestPreparer {
+                                            cells_per_slot,
+                                            delimiter_tokens,
+                                            text_tokenizer: DecisionTextTokenizer {
+                                                input_normalization: self
+                                                    .loaded_pointer_head
+                                                    .pointer_head
+                                                    .input_normalization,
+                                                model: self.loaded_llama_model.model.clone(),
+                                            },
+                                        },
+                                        preparation_tasks: TaskTracker::new(),
+                                        scheduler_message_tx,
+                                    }),
+                                }),
+                                || {
+                                    DecisionScheduler {
+                                        active_requests: VecDeque::new(),
+                                        agent_shutdown: agent_shutdown.clone(),
+                                        batch,
+                                        capacity_ledger: DecisionCapacityLedger::new(
+                                            desired_slots_total,
+                                        ),
+                                        llama_context,
+                                        pending_requests: VecDeque::new(),
+                                        scheduler_context: DecisionSchedulerContext {
+                                            agent_name: self.inference_runtime_context.agent_name,
+                                            n_batch,
+                                            pointer_head: self.loaded_pointer_head.pointer_head,
+                                        },
+                                        scheduler_inbox: SchedulerInbox::new(scheduler_message_rx),
+                                    }
+                                    .run();
+                                },
+                            );
+                        },
+                    )
+                    .map_err(DecisionError::from)
+            })
+        })
+    }
+}
diff --git a/paddler_agent_decision/src/decision_prerequisites.rs b/paddler_agent_decision/src/decision_prerequisites.rs
new file mode 100644
index 000000000..0f662f296
--- /dev/null
+++ b/paddler_agent_decision/src/decision_prerequisites.rs
@@ -0,0 +1,72 @@
+use std::path::PathBuf;
+
+use paddler_agent_pointer_head::pointer_head::PointerHead;
+use paddler_agent_runtime::inference_runtime_context::InferenceRuntimeContext;
+use paddler_agent_status::agent_issue_fix::AgentIssueFix;
+use paddler_messaging::agent_issue::AgentIssue;
+use paddler_messaging::agent_issue_params::model_path::ModelPath;
+use paddler_messaging::agent_issue_params::slots_insufficient_for_decisions_params::SlotsInsufficientForDecisionsParams;
+use paddler_messaging::decision_slots_minimum::DECISION_SLOTS_MINIMUM;
+
+use crate::decision_error::DecisionError;
+use crate::loaded_pointer_head::LoadedPointerHead;
+
+pub struct DecisionPrerequisites {
+    pub inference_runtime_context: InferenceRuntimeContext,
+    pub pointer_head_path: PathBuf,
+}
+
+impl DecisionPrerequisites {
+    pub fn load_pointer_head(self) -> Result<LoadedPointerHead, DecisionError> {
+        self.require_enough_slots()?;
+
+        match PointerHead::load(&self.pointer_head_path) {
+            Ok(pointer_head) => {
+                self.inference_runtime_context
+                    .slot_aggregated_status
+                    .register_fix(&AgentIssueFix::PointerHeadIsLoaded(ModelPath::from(
+                        self.pointer_head_path.as_path(),
+                    )));
+
+                Ok(LoadedPointerHead {
+                    pointer_head,
+                    pointer_head_path: self.pointer_head_path,
+                })
+            }
+            Err(source) => {
+                self.inference_runtime_context
+                    .slot_aggregated_status
+                    .register_pipeline_issue(AgentIssue::PointerHeadCannotBeLoaded(
+                        ModelPath::from(self.pointer_head_path.as_path()),
+                    ));
+
+                Err(DecisionError::PointerHeadCannotBeLoaded(source))
+            }
+        }
+    }
+
+    fn require_enough_slots(&self) -> Result<(), DecisionError> {
+        let desired_slots = self
+            .inference_runtime_context
+            .slot_aggregated_status
+            .desired_slots_total;
+
+        if desired_slots >= DECISION_SLOTS_MINIMUM.get() {
+            return Ok(());
+        }
+
+        self.inference_runtime_context
+            .slot_aggregated_status
+            .register_pipeline_issue(AgentIssue::SlotsInsufficientForDecisions(
+                SlotsInsufficientForDecisionsParams {
+                    desired_slots,
+                    required_slots: DECISION_SLOTS_MINIMUM.get(),
+                },
+            ));
+
+        Err(DecisionError::SlotsInsufficient {
+            desired_slots,
+            required_slots: DECISION_SLOTS_MINIMUM.get(),
+        })
+    }
+}
diff --git a/paddler_agent_decision/src/decision_question_row.rs b/paddler_agent_decision/src/decision_question_row.rs
new file mode 100644
index 000000000..764f753bc
--- /dev/null
+++ b/paddler_agent_decision/src/decision_question_row.rs
@@ -0,0 +1,38 @@
+use llama_cpp_bindings::token::LlamaToken;
+
+#[derive(Clone, Debug, Eq, PartialEq)]
+pub struct DecisionQuestionRow {
+    pub id: String,
+    pub option_end_offsets: Vec<usize>,
+    pub tokens: Vec<LlamaToken>,
+}
+
+impl DecisionQuestionRow {
+    #[must_use]
+    pub fn is_output(&self, row_offset: usize) -> bool {
+        row_offset + 1 == self.tokens.len() || self.option_end_offsets.contains(&row_offset)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use llama_cpp_bindings::token::LlamaToken;
+
+    use super::DecisionQuestionRow;
+
+    #[test]
+    fn option_ends_and_the_closing_decide_token_are_outputs() {
+        let row = DecisionQuestionRow {
+            id: "question".to_owned(),
+            option_end_offsets: vec![2],
+            tokens: (0..4).map(LlamaToken::new).collect(),
+        };
+
+        assert_eq!(
+            (0..4)
+                .map(|row_offset| row.is_output(row_offset))
+                .collect::<Vec<_>>(),
+            vec![false, false, true, true]
+        );
+    }
+}
diff --git a/paddler_agent_decision/src/decision_request_phase.rs b/paddler_agent_decision/src/decision_request_phase.rs
new file mode 100644
index 000000000..334a725c3
--- /dev/null
+++ b/paddler_agent_decision/src/decision_request_phase.rs
@@ -0,0 +1,5 @@
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+pub enum DecisionRequestPhase {
+    AnsweringQuestion { next_offset: usize },
+    IngestingState { next_offset: usize },
+}
diff --git a/paddler_agent_decision/src/decision_request_preparer.rs b/paddler_agent_decision/src/decision_request_preparer.rs
new file mode 100644
index 000000000..936db4b8d
--- /dev/null
+++ b/paddler_agent_decision/src/decision_request_preparer.rs
@@ -0,0 +1,130 @@
+use std::collections::HashMap;
+use std::collections::hash_map::Entry;
+use std::time::Instant;
+
+use llama_cpp_bindings::token::LlamaToken;
+
+use paddler_agent_runtime::agent_request::AgentRequest;
+use paddler_agent_runtime::prepares_scheduler_command::PreparesSchedulerCommand;
+use paddler_messaging::decision_question::DecisionQuestion;
+use paddler_messaging::decision_result::DecisionResult;
+use paddler_messaging::oversized_decision_details::OversizedDecisionDetails;
+use paddler_messaging::request_params::decide_params::DecideParams;
+
+use crate::decision_delimiter_tokens::DecisionDelimiterTokens;
+use crate::decision_error::DecisionError;
+use crate::decision_text_tokenizer::DecisionTextTokenizer;
+use crate::decision_token_layout::DecisionTokenLayout;
+use crate::prepared_decision_request::PreparedDecisionRequest;
+use crate::reports_decision_outcome::ReportsDecisionOutcome as _;
+use crate::tokenized_decision_question::TokenizedDecisionQuestion;
+
+pub struct DecisionRequestPreparer {
+    pub cells_per_slot: u32,
+    pub delimiter_tokens: DecisionDelimiterTokens,
+    pub text_tokenizer: DecisionTextTokenizer,
+}
+
+impl DecisionRequestPreparer {
+    fn tokenize_once(
+        &self,
+        tokens_by_text: &mut HashMap<String, Vec<LlamaToken>>,
+        text: String,
+    ) -> Result<Vec<LlamaToken>, DecisionError> {
+        match tokens_by_text.entry(text) {
+            Entry::Occupied(tokenized_text) => Ok(tokenized_text.get().clone()),
+            Entry::Vacant(untokenized_text) => {
+                let tokens = self.text_tokenizer.tokenize(untokenized_text.key())?;
+
+                Ok(untokenized_text.insert(tokens).clone())
+            }
+        }
+    }
+
+    fn tokenize_question(
+        &self,
+        tokens_by_text: &mut HashMap<String, Vec<LlamaToken>>,
+        DecisionQuestion {
+            id,
+            instructions,
+            options,
+        }: DecisionQuestion,
+    ) -> Result<TokenizedDecisionQuestion, DecisionError> {
+        Ok(TokenizedDecisionQuestion {
+            id,
+            instructions: self.tokenize_once(tokens_by_text, instructions)?,
+            options: options
+                .into_iter()
+                .map(|option| self.tokenize_once(tokens_by_text, option))
+                .collect::<Result<Vec<_>, _>>()?,
+        })
+    }
+
+    fn prepare(
+        &self,
+        AgentRequest {
+            params,
+            response_tx: decision_result_tx,
+            slot_guard,
+            stop_rx: decision_stop_rx,
+        }: AgentRequest<DecideParams, DecisionResult>,
+    ) -> Result<PreparedDecisionRequest, DecisionError> {
+        let started_at = Instant::now();
+        let context_size = u32::from(slot_guard.slots().get()) * self.cells_per_slot;
+        let DecideParams {
+            last_question,
+            leading_questions,
+            state,
+        } = params;
+        let mut tokens_by_text = HashMap::new();
+        let layout = DecisionTokenLayout::new(
+            &self.delimiter_tokens,
+            self.text_tokenizer.tokenize(&state)?,
+            leading_questions
+                .into_iter()
+                .map(|question| self.tokenize_question(&mut tokens_by_text, question))
+                .collect::<Result<Vec<_>, _>>()?,
+            self.tokenize_question(&mut tokens_by_text, last_question)?,
+        );
+        let required_tokens = layout.required_cells();
+
+        if required_tokens > context_size as usize {
+            return Err(DecisionError::RequestExceedsContext {
+                details: OversizedDecisionDetails {
+                    context_size,
+                    required_tokens,
+                },
+            });
+        }
+
+        Ok(PreparedDecisionRequest {
+            decision_result_tx,
+            decision_stop_rx,
+            layout,
+            slot_guard,
+            started_at,
+        })
+    }
+}
+
+impl PreparesSchedulerCommand for DecisionRequestPreparer {
+    type Command = PreparedDecisionRequest;
+    type Request = AgentRequest<DecideParams, DecisionResult>;
+
+    fn prepare_scheduler_command(
+        &self,
+        agent_name: Option<&str>,
+        request: Self::Request,
+    ) -> Option<Self::Command> {
+        let decision_result_tx = request.response_tx.clone();
+
+        match self.prepare(request) {
+            Ok(prepared) => Some(prepared),
+            Err(rejection) => {
+                rejection.report(agent_name, &decision_result_tx);
+
+                None
+            }
+        }
+    }
+}
diff --git a/paddler_agent_decision/src/decision_scheduler.rs b/paddler_agent_decision/src/decision_scheduler.rs
new file mode 100644
index 000000000..cd2d7be07
--- /dev/null
+++ b/paddler_agent_decision/src/decision_scheduler.rs
@@ -0,0 +1,158 @@
+use std::collections::VecDeque;
+
+use llama_cpp_bindings::llama_batch::LlamaBatch;
+use log::info;
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent_runtime::receives_stop_request::ReceivesStopRequest as _;
+use paddler_agent_runtime::resident_llama_context::ResidentLlamaContext;
+use paddler_agent_runtime::scheduler_inbox::SchedulerInbox;
+use paddler_messaging::decision_result::DecisionResult;
+
+use crate::active_decision_request::ActiveDecisionRequest;
+use crate::decision_advance::DecisionAdvance;
+use crate::decision_capacity_ledger::DecisionCapacityLedger;
+use crate::decision_decoder::DecisionDecoder;
+use crate::decision_error::DecisionError;
+use crate::decision_scheduler_context::DecisionSchedulerContext;
+use crate::prepared_decision_request::PreparedDecisionRequest;
+use crate::reports_decision_outcome::ReportsDecisionOutcome as _;
+
+pub struct DecisionScheduler<'model> {
+    pub active_requests: VecDeque<ActiveDecisionRequest>,
+    pub agent_shutdown: CancellationToken,
+    pub batch: LlamaBatch<'static>,
+    pub capacity_ledger: DecisionCapacityLedger,
+    pub llama_context: ResidentLlamaContext<'model>,
+    pub pending_requests: VecDeque<PreparedDecisionRequest>,
+    pub scheduler_context: DecisionSchedulerContext,
+    pub scheduler_inbox: SchedulerInbox<PreparedDecisionRequest>,
+}
+
+impl DecisionScheduler<'_> {
+    pub fn run(&mut self) {
+        info!(
+            "{:?}: decision scheduler started",
+            self.scheduler_context.agent_name
+        );
+
+        while !self.agent_shutdown.is_cancelled() {
+            if self.active_requests.is_empty() && self.pending_requests.is_empty() {
+                if self.scheduler_inbox.shutdown_requested {
+                    break;
+                }
+
+                self.wait_for_next_scheduler_message();
+            } else {
+                self.accept_new_scheduler_messages();
+            }
+
+            self.admit_pending_requests();
+            self.advance_next_active_request();
+        }
+
+        self.reject_unfinished_requests();
+        self.llama_context.synchronize();
+
+        info!(
+            "{:?}: decision scheduler stopped",
+            self.scheduler_context.agent_name
+        );
+    }
+
+    fn accept_new_scheduler_messages(&mut self) {
+        while let Some(request) = self.scheduler_inbox.take_next_waiting_command() {
+            self.pending_requests.push_back(request);
+        }
+    }
+
+    fn admit_pending_requests(&mut self) {
+        while let Some(mut request) = self.pending_requests.pop_front() {
+            if request.decision_stop_rx.is_stop_requested() {
+                DecisionResult::StopRequested.report(
+                    self.scheduler_context.agent_name.as_deref(),
+                    &request.decision_result_tx,
+                );
+
+                continue;
+            }
+
+            let Some(admission) = self
+                .capacity_ledger
+                .admit(request.layout.question_lane_demand())
+            else {
+                self.pending_requests.push_front(request);
+
+                break;
+            };
+
+            if let Err(clear_error) = admission
+                .state_sequence
+                .sequence_id()
+                .clear_kv_cache_in(&mut self.llama_context)
+            {
+                DecisionError::KvCacheRemovalFailed(clear_error).report(
+                    self.scheduler_context.agent_name.as_deref(),
+                    &request.decision_result_tx,
+                );
+
+                continue;
+            }
+
+            self.active_requests
+                .push_back(ActiveDecisionRequest::new(request, admission));
+        }
+    }
+
+    fn advance_next_active_request(&mut self) {
+        let Some(mut request) = self.active_requests.pop_front() else {
+            return;
+        };
+        let mut decoder = DecisionDecoder {
+            batch: &mut self.batch,
+            llama_context: &mut self.llama_context,
+            n_batch: self.scheduler_context.n_batch,
+        };
+        let agent_name = self.scheduler_context.agent_name.as_deref();
+
+        if request.is_stop_requested() {
+            request.finish_with(&mut decoder, agent_name, DecisionResult::StopRequested);
+
+            return;
+        }
+
+        match request.advance(
+            &mut decoder,
+            &self.scheduler_context.pointer_head,
+            agent_name,
+        ) {
+            Ok(DecisionAdvance::Continuing) => self.active_requests.push_back(request),
+            Ok(DecisionAdvance::Completed) => {
+                let summary = request.summary();
+
+                request.finish_with(&mut decoder, agent_name, summary);
+            }
+            Err(advance_error) => request.finish_with(&mut decoder, agent_name, advance_error),
+        }
+    }
+
+    fn reject_unfinished_requests(&mut self) {
+        let agent_name = self.scheduler_context.agent_name.as_deref();
+
+        for request in self.active_requests.drain(..) {
+            DecisionError::SchedulerUnavailable.report(agent_name, &request.decision_result_tx);
+        }
+
+        for request in self.pending_requests.drain(..) {
+            DecisionError::SchedulerUnavailable.report(agent_name, &request.decision_result_tx);
+        }
+
+        self.scheduler_inbox.reject_waiting_commands(agent_name);
+    }
+
+    fn wait_for_next_scheduler_message(&mut self) {
+        if let Some(request) = self.scheduler_inbox.wait_for_next_command() {
+            self.pending_requests.push_back(request);
+        }
+    }
+}
diff --git a/paddler_agent_decision/src/decision_scheduler_context.rs b/paddler_agent_decision/src/decision_scheduler_context.rs
new file mode 100644
index 000000000..d7bae17e1
--- /dev/null
+++ b/paddler_agent_decision/src/decision_scheduler_context.rs
@@ -0,0 +1,7 @@
+use paddler_agent_pointer_head::pointer_head::PointerHead;
+
+pub struct DecisionSchedulerContext {
+    pub agent_name: Option<String>,
+    pub n_batch: usize,
+    pub pointer_head: PointerHead,
+}
diff --git a/paddler_agent_decision/src/decision_text_tokenizer.rs b/paddler_agent_decision/src/decision_text_tokenizer.rs
new file mode 100644
index 000000000..3743f2077
--- /dev/null
+++ b/paddler_agent_decision/src/decision_text_tokenizer.rs
@@ -0,0 +1,31 @@
+use std::ops::Deref;
+use std::sync::Arc;
+
+use llama_cpp_bindings::model::AddBos;
+use llama_cpp_bindings::model::LlamaModel;
+use llama_cpp_bindings::model::ParseSpecialTokens;
+use llama_cpp_bindings::token::LlamaToken;
+
+use paddler_agent_pointer_head::pointer_head_input_normalization::PointerHeadInputNormalization;
+
+use crate::decision_error::DecisionError;
+use crate::escape_special_token_lookalikes::escape_special_token_lookalikes;
+
+pub struct DecisionTextTokenizer {
+    pub input_normalization: PointerHeadInputNormalization,
+    pub model: Arc<dyn Deref<Target = LlamaModel> + Send + Sync>,
+}
+
+impl DecisionTextTokenizer {
+    pub fn tokenize(&self, text: &str) -> Result<Vec<LlamaToken>, DecisionError> {
+        self.model
+            .str_to_token(
+                &self
+                    .input_normalization
+                    .normalize(&escape_special_token_lookalikes(text)),
+                AddBos::Never,
+                ParseSpecialTokens::Never,
+            )
+            .map_err(DecisionError::InputTokenizationFailed)
+    }
+}
diff --git a/paddler_agent_decision/src/decision_token_layout.rs b/paddler_agent_decision/src/decision_token_layout.rs
new file mode 100644
index 000000000..570a08a09
--- /dev/null
+++ b/paddler_agent_decision/src/decision_token_layout.rs
@@ -0,0 +1,205 @@
+use std::collections::VecDeque;
+
+use llama_cpp_bindings::token::LlamaToken;
+
+use crate::decision_delimiter_tokens::DecisionDelimiterTokens;
+use crate::decision_question_row::DecisionQuestionRow;
+use crate::question_lane_demand::QuestionLaneDemand;
+use crate::tokenized_decision_question::TokenizedDecisionQuestion;
+
+fn question_row(
+    delimiter_tokens: &DecisionDelimiterTokens,
+    TokenizedDecisionQuestion {
+        id,
+        instructions,
+        options,
+    }: TokenizedDecisionQuestion,
+) -> DecisionQuestionRow {
+    let mut tokens = Vec::with_capacity(
+        instructions.len() + options.iter().map(|option| option.len() + 2).sum::<usize>() + 2,
+    );
+    let mut option_end_offsets = Vec::with_capacity(options.len());
+
+    tokens.push(delimiter_tokens.question);
+    tokens.extend(instructions);
+
+    for option in options {
+        tokens.push(delimiter_tokens.option_start);
+        tokens.extend(option);
+        option_end_offsets.push(tokens.len());
+        tokens.push(delimiter_tokens.option_end);
+    }
+
+    tokens.push(delimiter_tokens.decide);
+
+    DecisionQuestionRow {
+        id,
+        option_end_offsets,
+        tokens,
+    }
+}
+
+#[derive(Clone, Debug, Eq, PartialEq)]
+pub struct DecisionTokenLayout {
+    pub first_question: DecisionQuestionRow,
+    pub remaining_questions: VecDeque<DecisionQuestionRow>,
+    pub state: Vec<LlamaToken>,
+}
+
+impl DecisionTokenLayout {
+    #[must_use]
+    pub fn new(
+        delimiter_tokens: &DecisionDelimiterTokens,
+        state: Vec<LlamaToken>,
+        leading_questions: Vec<TokenizedDecisionQuestion>,
+        last_question: TokenizedDecisionQuestion,
+    ) -> Self {
+        let mut state_tokens = Vec::with_capacity(state.len() + 1);
+
+        state_tokens.push(delimiter_tokens.state);
+        state_tokens.extend(state);
+
+        let mut remaining_questions: VecDeque<DecisionQuestionRow> = leading_questions
+            .into_iter()
+            .map(|question| question_row(delimiter_tokens, question))
+            .collect();
+        let last_question = question_row(delimiter_tokens, last_question);
+        let first_question = match remaining_questions.pop_front() {
+            Some(first_question) => {
+                remaining_questions.push_back(last_question);
+
+                first_question
+            }
+            None => last_question,
+        };
+
+        Self {
+            first_question,
+            remaining_questions,
+            state: state_tokens,
+        }
+    }
+
+    #[must_use]
+    pub fn question_lane_demand(&self) -> QuestionLaneDemand {
+        if self.remaining_questions.is_empty() {
+            QuestionLaneDemand::NotNeeded
+        } else {
+            QuestionLaneDemand::Needed
+        }
+    }
+
+    #[must_use]
+    pub fn required_cells(&self) -> usize {
+        self.state.len()
+            + self
+                .remaining_questions
+                .iter()
+                .map(|question| question.tokens.len())
+                .fold(self.first_question.tokens.len(), usize::max)
+    }
+
+    #[must_use]
+    pub fn total_tokens(&self) -> usize {
+        self.state.len()
+            + self.first_question.tokens.len()
+            + self
+                .remaining_questions
+                .iter()
+                .map(|question| question.tokens.len())
+                .sum::<usize>()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use llama_cpp_bindings::token::LlamaToken;
+
+    use super::DecisionTokenLayout;
+    use crate::decision_delimiter_tokens::DecisionDelimiterTokens;
+    use crate::decision_question_row::DecisionQuestionRow;
+    use crate::tokenized_decision_question::TokenizedDecisionQuestion;
+
+    fn tokens(ids: &[i32]) -> Vec<LlamaToken> {
+        ids.iter().copied().map(LlamaToken::new).collect()
+    }
+
+    fn delimiter_tokens() -> DecisionDelimiterTokens {
+        DecisionDelimiterTokens {
+            decide: LlamaToken::new(-5),
+            option_end: LlamaToken::new(-4),
+            option_start: LlamaToken::new(-3),
+            question: LlamaToken::new(-2),
+            state: LlamaToken::new(-1),
+        }
+    }
+
+    fn first_question() -> TokenizedDecisionQuestion {
+        TokenizedDecisionQuestion {
+            id: "first".to_owned(),
+            instructions: tokens(&[20]),
+            options: vec![tokens(&[30, 31]), tokens(&[32])],
+        }
+    }
+
+    fn last_question() -> TokenizedDecisionQuestion {
+        TokenizedDecisionQuestion {
+            id: "last".to_owned(),
+            instructions: Vec::new(),
+            options: vec![tokens(&[40])],
+        }
+    }
+
+    fn two_question_layout() -> DecisionTokenLayout {
+        DecisionTokenLayout::new(
+            &delimiter_tokens(),
+            tokens(&[10, 11]),
+            vec![first_question()],
+            last_question(),
+        )
+    }
+
+    #[test]
+    fn lays_out_the_state_and_one_row_per_question_with_kev_delimiters() {
+        assert_eq!(
+            two_question_layout(),
+            DecisionTokenLayout {
+                first_question: DecisionQuestionRow {
+                    id: "first".to_owned(),
+                    option_end_offsets: vec![5, 8],
+                    tokens: tokens(&[-2, 20, -3, 30, 31, -4, -3, 32, -4, -5]),
+                },
+                remaining_questions: [DecisionQuestionRow {
+                    id: "last".to_owned(),
+                    option_end_offsets: vec![3],
+                    tokens: tokens(&[-2, -3, 40, -4, -5]),
+                }]
+                .into(),
+                state: tokens(&[-1, 10, 11]),
+            }
+        );
+    }
+
+    #[test]
+    fn a_single_question_is_the_first_question() {
+        let layout = DecisionTokenLayout::new(
+            &delimiter_tokens(),
+            tokens(&[10]),
+            Vec::new(),
+            last_question(),
+        );
+
+        assert_eq!(layout.first_question.id, "last");
+        assert!(layout.remaining_questions.is_empty());
+    }
+
+    #[test]
+    fn a_decision_needs_its_state_and_its_longest_question_in_the_cache() {
+        assert_eq!(two_question_layout().required_cells(), 13);
+    }
+
+    #[test]
+    fn a_decision_decodes_its_state_once_and_every_question_row() {
+        assert_eq!(two_question_layout().total_tokens(), 18);
+    }
+}
diff --git a/paddler_agent_decision/src/delimiter_vocabulary.rs b/paddler_agent_decision/src/delimiter_vocabulary.rs
new file mode 100644
index 000000000..0faf1933c
--- /dev/null
+++ b/paddler_agent_decision/src/delimiter_vocabulary.rs
@@ -0,0 +1,23 @@
+use llama_cpp_bindings::error::StringToTokenError;
+use llama_cpp_bindings::error::TokenAttrError;
+use llama_cpp_bindings::llama_token_attrs::LlamaTokenAttrs;
+use llama_cpp_bindings::model::AddBos;
+use llama_cpp_bindings::model::LlamaModel;
+use llama_cpp_bindings::model::ParseSpecialTokens;
+use llama_cpp_bindings::token::LlamaToken;
+
+pub trait DelimiterVocabulary {
+    fn tokenize_delimiter(&self, delimiter: &str) -> Result<Vec<LlamaToken>, StringToTokenError>;
+
+    fn token_attributes(&self, token: LlamaToken) -> Result<LlamaTokenAttrs, TokenAttrError>;
+}
+
+impl DelimiterVocabulary for LlamaModel {
+    fn tokenize_delimiter(&self, delimiter: &str) -> Result<Vec<LlamaToken>, StringToTokenError> {
+        self.str_to_token(delimiter, AddBos::Never, ParseSpecialTokens::Always)
+    }
+
+    fn token_attributes(&self, token: LlamaToken) -> Result<LlamaTokenAttrs, TokenAttrError> {
+        self.token_attr(token)
+    }
+}
diff --git a/paddler_agent_decision/src/escape_special_token_lookalikes.rs b/paddler_agent_decision/src/escape_special_token_lookalikes.rs
new file mode 100644
index 000000000..8d776b1af
--- /dev/null
+++ b/paddler_agent_decision/src/escape_special_token_lookalikes.rs
@@ -0,0 +1,74 @@
+const ESCAPED_LOOKALIKE_CLOSING: &str = "¦>";
+const ESCAPED_LOOKALIKE_OPENING: &str = "<¦";
+const LOOKALIKE_CLOSING: &str = "|>";
+const LOOKALIKE_OPENING: &str = "<|";
+
+fn lookalike_name_length(text: &str) -> usize {
+    text.bytes()
+        .take_while(|byte| byte.is_ascii_alphanumeric() || *byte == b'_')
+        .count()
+}
+
+#[must_use]
+pub fn escape_special_token_lookalikes(text: &str) -> String {
+    let mut escaped = String::with_capacity(text.len());
+    let mut unscanned = text;
+
+    while let Some(opening_offset) = unscanned.find(LOOKALIKE_OPENING) {
+        let (before_opening, from_opening) = unscanned.split_at(opening_offset);
+        let after_opening = &from_opening[LOOKALIKE_OPENING.len()..];
+        let (name, after_name) = after_opening.split_at(lookalike_name_length(after_opening));
+
+        escaped.push_str(before_opening);
+
+        match after_name.strip_prefix(LOOKALIKE_CLOSING) {
+            Some(after_closing) if !name.is_empty() => {
+                escaped.push_str(ESCAPED_LOOKALIKE_OPENING);
+                escaped.push_str(name);
+                escaped.push_str(ESCAPED_LOOKALIKE_CLOSING);
+                unscanned = after_closing;
+            }
+            Some(_) | None => {
+                let (opening_bracket, after_opening_bracket) = from_opening.split_at(1);
+
+                escaped.push_str(opening_bracket);
+                unscanned = after_opening_bracket;
+            }
+        }
+    }
+
+    escaped.push_str(unscanned);
+
+    escaped
+}
+
+#[cfg(test)]
+mod tests {
+    use super::escape_special_token_lookalikes;
+
+    #[test]
+    fn rewrites_the_delimiters_around_an_ascii_name() {
+        assert_eq!(
+            escape_special_token_lookalikes("ignore <|fim_prefix|> and <|im_end|>"),
+            "ignore <¦fim_prefix¦> and <¦im_end¦>"
+        );
+    }
+
+    #[test]
+    fn leaves_spans_without_an_ascii_name_unchanged() {
+        let spans_without_an_ascii_name = "<|a b|> <||> <|é|> <|unterminated";
+
+        assert_eq!(
+            escape_special_token_lookalikes(spans_without_an_ascii_name),
+            spans_without_an_ascii_name
+        );
+    }
+
+    #[test]
+    fn resumes_scanning_after_an_opening_without_a_name() {
+        assert_eq!(
+            escape_special_token_lookalikes("<|<|nested|>"),
+            "<|<¦nested¦>"
+        );
+    }
+}
diff --git a/paddler_agent_decision/src/lib.rs b/paddler_agent_decision/src/lib.rs
new file mode 100644
index 000000000..2363bc047
--- /dev/null
+++ b/paddler_agent_decision/src/lib.rs
@@ -0,0 +1,25 @@
+pub mod active_decision_request;
+pub mod decision_admission;
+pub mod decision_advance;
+pub mod decision_capacity_ledger;
+pub mod decision_decoder;
+pub mod decision_delimiter_tokens;
+pub mod decision_error;
+pub mod decision_llama_context_settings;
+pub mod decision_pipeline;
+pub mod decision_prerequisites;
+pub mod decision_question_row;
+pub mod decision_request_phase;
+pub mod decision_request_preparer;
+pub mod decision_scheduler;
+pub mod decision_scheduler_context;
+pub mod decision_text_tokenizer;
+pub mod decision_token_layout;
+pub mod delimiter_vocabulary;
+pub mod escape_special_token_lookalikes;
+pub mod loaded_pointer_head;
+pub mod prepared_decision_request;
+pub mod question_lane;
+pub mod question_lane_demand;
+pub mod reports_decision_outcome;
+pub mod tokenized_decision_question;
diff --git a/paddler_agent_decision/src/loaded_pointer_head.rs b/paddler_agent_decision/src/loaded_pointer_head.rs
new file mode 100644
index 000000000..ee7e70a26
--- /dev/null
+++ b/paddler_agent_decision/src/loaded_pointer_head.rs
@@ -0,0 +1,8 @@
+use std::path::PathBuf;
+
+use paddler_agent_pointer_head::pointer_head::PointerHead;
+
+pub struct LoadedPointerHead {
+    pub pointer_head: PointerHead,
+    pub pointer_head_path: PathBuf,
+}
diff --git a/paddler_agent_decision/src/prepared_decision_request.rs b/paddler_agent_decision/src/prepared_decision_request.rs
new file mode 100644
index 000000000..63c0d448e
--- /dev/null
+++ b/paddler_agent_decision/src/prepared_decision_request.rs
@@ -0,0 +1,75 @@
+use std::time::Instant;
+
+use tokio::sync::mpsc;
+
+use paddler_agent_runtime::scheduler_command::SchedulerCommand;
+use paddler_agent_status::slot_guard::SlotGuard;
+use paddler_messaging::decision_result::DecisionResult;
+
+use crate::decision_error::DecisionError;
+use crate::decision_token_layout::DecisionTokenLayout;
+use crate::reports_decision_outcome::ReportsDecisionOutcome as _;
+
+pub struct PreparedDecisionRequest {
+    pub decision_result_tx: mpsc::UnboundedSender<DecisionResult>,
+    pub decision_stop_rx: mpsc::UnboundedReceiver<()>,
+    pub layout: DecisionTokenLayout,
+    pub slot_guard: SlotGuard,
+    pub started_at: Instant,
+}
+
+impl SchedulerCommand for PreparedDecisionRequest {
+    fn reject_because_the_scheduler_stopped(self, agent_name: Option<&str>) {
+        DecisionError::SchedulerUnavailable.report(agent_name, &self.decision_result_tx);
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::collections::VecDeque;
+    use std::num::NonZeroU16;
+    use std::sync::Arc;
+    use std::time::Instant;
+
+    use llama_cpp_bindings::token::LlamaToken;
+    use tokio::sync::mpsc;
+
+    use paddler_agent_runtime::scheduler_command::SchedulerCommand as _;
+    use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
+    use paddler_agent_status::slot_guard::SlotGuard;
+    use paddler_messaging::decision_result::DecisionResult;
+
+    use super::PreparedDecisionRequest;
+    use crate::decision_question_row::DecisionQuestionRow;
+    use crate::decision_token_layout::DecisionTokenLayout;
+
+    #[test]
+    fn rejects_a_decision_when_the_scheduler_stopped() {
+        let (decision_result_tx, mut decision_result_rx) = mpsc::unbounded_channel();
+        let (_decision_stop_tx, decision_stop_rx) = mpsc::unbounded_channel();
+
+        PreparedDecisionRequest {
+            decision_result_tx,
+            decision_stop_rx,
+            layout: DecisionTokenLayout {
+                first_question: DecisionQuestionRow {
+                    id: "paid".to_owned(),
+                    option_end_offsets: Vec::new(),
+                    tokens: vec![LlamaToken::new(0)],
+                },
+                remaining_questions: VecDeque::new(),
+                state: Vec::new(),
+            },
+            slot_guard: SlotGuard::new(Arc::new(SlotAggregatedStatus::new(1)), NonZeroU16::MIN),
+            started_at: Instant::now(),
+        }
+        .reject_because_the_scheduler_stopped(Some("agent"));
+
+        assert_eq!(
+            decision_result_rx.try_recv(),
+            Ok(DecisionResult::SchedulerUnavailable(
+                "agent: the scheduler is no longer accepting requests".to_owned()
+            ))
+        );
+    }
+}
diff --git a/paddler_agent_decision/src/question_lane.rs b/paddler_agent_decision/src/question_lane.rs
new file mode 100644
index 000000000..bfe995807
--- /dev/null
+++ b/paddler_agent_decision/src/question_lane.rs
@@ -0,0 +1,6 @@
+use paddler_agent_runtime::sequence_id_guard::SequenceIdGuard;
+
+pub enum QuestionLane {
+    Released,
+    Reserved(SequenceIdGuard),
+}
diff --git a/paddler_agent_decision/src/question_lane_demand.rs b/paddler_agent_decision/src/question_lane_demand.rs
new file mode 100644
index 000000000..b05a6eec9
--- /dev/null
+++ b/paddler_agent_decision/src/question_lane_demand.rs
@@ -0,0 +1,5 @@
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+pub enum QuestionLaneDemand {
+    NotNeeded,
+    Needed,
+}
diff --git a/paddler_agent_decision/src/reports_decision_outcome.rs b/paddler_agent_decision/src/reports_decision_outcome.rs
new file mode 100644
index 000000000..fba32e58f
--- /dev/null
+++ b/paddler_agent_decision/src/reports_decision_outcome.rs
@@ -0,0 +1,22 @@
+use tokio::sync::mpsc;
+
+use paddler_agent_runtime::send_result_or_warn::send_result_or_warn;
+use paddler_messaging::decision_result::DecisionResult;
+
+pub trait ReportsDecisionOutcome {
+    fn report(
+        self,
+        agent_name: Option<&str>,
+        decision_result_tx: &mpsc::UnboundedSender<DecisionResult>,
+    );
+}
+
+impl ReportsDecisionOutcome for DecisionResult {
+    fn report(
+        self,
+        agent_name: Option<&str>,
+        decision_result_tx: &mpsc::UnboundedSender<DecisionResult>,
+    ) {
+        send_result_or_warn(agent_name, decision_result_tx, self);
+    }
+}
diff --git a/paddler_agent_decision/src/tokenized_decision_question.rs b/paddler_agent_decision/src/tokenized_decision_question.rs
new file mode 100644
index 000000000..05fcbbe24
--- /dev/null
+++ b/paddler_agent_decision/src/tokenized_decision_question.rs
@@ -0,0 +1,7 @@
+use llama_cpp_bindings::token::LlamaToken;
+
+pub struct TokenizedDecisionQuestion {
+    pub id: String,
+    pub instructions: Vec<LlamaToken>,
+    pub options: Vec<Vec<LlamaToken>>,
+}
diff --git a/paddler_agent_embeddings/Cargo.toml b/paddler_agent_embeddings/Cargo.toml
new file mode 100644
index 000000000..3e6b1965f
--- /dev/null
+++ b/paddler_agent_embeddings/Cargo.toml
@@ -0,0 +1,29 @@
+[package]
+name = "paddler_agent_embeddings"
+authors.workspace = true
+description.workspace = true
+edition.workspace = true
+homepage.workspace = true
+license.workspace = true
+repository.workspace = true
+version.workspace = true
+
+[dependencies]
+llama-cpp-bindings = { workspace = true }
+log = { workspace = true }
+thiserror = { workspace = true }
+tokio = { workspace = true }
+tokio-util = { workspace = true }
+paddler_agent_runtime = { workspace = true }
+paddler_agent_status = { workspace = true }
+paddler_inference_parameters = { workspace = true }
+paddler_messaging = { workspace = true }
+
+[features]
+default = []
+cuda = ["llama-cpp-bindings/cuda", "paddler_agent_runtime/cuda"]
+metal = ["llama-cpp-bindings/metal", "paddler_agent_runtime/metal"]
+vulkan = ["llama-cpp-bindings/vulkan", "paddler_agent_runtime/vulkan"]
+
+[lints]
+workspace = true
diff --git a/paddler_agent_embeddings/src/active_embedding_request.rs b/paddler_agent_embeddings/src/active_embedding_request.rs
new file mode 100644
index 000000000..4a107dd9d
--- /dev/null
+++ b/paddler_agent_embeddings/src/active_embedding_request.rs
@@ -0,0 +1,68 @@
+use std::collections::VecDeque;
+
+use tokio::sync::mpsc;
+
+use paddler_agent_runtime::receives_stop_request::ReceivesStopRequest as _;
+use paddler_agent_runtime::send_result_or_warn::send_result_or_warn;
+use paddler_agent_status::slot_guard::SlotGuard;
+use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
+use paddler_messaging::embedding_result::EmbeddingResult;
+
+use crate::embedding_input_tokenized::EmbeddingInputTokenized;
+use crate::embedding_request_continuation::EmbeddingRequestContinuation;
+
+pub struct ActiveEmbeddingRequest {
+    pub embeddings_emitted: usize,
+    pub generate_embedding_stop_rx: mpsc::UnboundedReceiver<()>,
+    pub generated_embedding_tx: mpsc::UnboundedSender<EmbeddingResult>,
+    pub normalization_method: EmbeddingNormalizationMethod,
+    pub pending_inputs: VecDeque<EmbeddingInputTokenized>,
+    pub slot_guard: SlotGuard,
+}
+
+impl ActiveEmbeddingRequest {
+    #[must_use]
+    pub const fn sequences(&self) -> u16 {
+        self.slot_guard.slots().get()
+    }
+
+    #[must_use]
+    pub fn continue_unless_stop_requested(
+        mut self,
+        agent_name: Option<&str>,
+    ) -> EmbeddingRequestContinuation {
+        if self.generate_embedding_stop_rx.is_stop_requested() {
+            self.finish(agent_name);
+
+            return EmbeddingRequestContinuation::Concluded;
+        }
+
+        EmbeddingRequestContinuation::Continues(self)
+    }
+
+    #[must_use]
+    pub fn continue_unless_every_input_is_embedded(
+        self,
+        agent_name: Option<&str>,
+    ) -> EmbeddingRequestContinuation {
+        if self.pending_inputs.is_empty() {
+            self.finish(agent_name);
+
+            return EmbeddingRequestContinuation::Concluded;
+        }
+
+        EmbeddingRequestContinuation::Continues(self)
+    }
+
+    fn finish(self, agent_name: Option<&str>) {
+        send_result_or_warn(
+            agent_name,
+            &self.generated_embedding_tx,
+            if self.embeddings_emitted == 0 {
+                EmbeddingResult::NoEmbeddingsProduced
+            } else {
+                EmbeddingResult::Done
+            },
+        );
+    }
+}
diff --git a/paddler_agent/src/agent_pooling_type.rs b/paddler_agent_embeddings/src/agent_pooling_type.rs
similarity index 100%
rename from paddler_agent/src/agent_pooling_type.rs
rename to paddler_agent_embeddings/src/agent_pooling_type.rs
diff --git a/paddler_agent_embeddings/src/batched_embedding_input.rs b/paddler_agent_embeddings/src/batched_embedding_input.rs
new file mode 100644
index 000000000..0da661101
--- /dev/null
+++ b/paddler_agent_embeddings/src/batched_embedding_input.rs
@@ -0,0 +1,8 @@
+use crate::embedding_input_tokenized::EmbeddingInputTokenized;
+
+#[derive(Debug, Eq, PartialEq)]
+pub struct BatchedEmbeddingInput {
+    pub input: EmbeddingInputTokenized,
+    pub request_index: usize,
+    pub sequence_id: i32,
+}
diff --git a/paddler_agent/src/converts_to_llama_pooling_type.rs b/paddler_agent_embeddings/src/converts_to_llama_pooling_type.rs
similarity index 100%
rename from paddler_agent/src/converts_to_llama_pooling_type.rs
rename to paddler_agent_embeddings/src/converts_to_llama_pooling_type.rs
diff --git a/paddler_agent_embeddings/src/embedding_batch_preparer.rs b/paddler_agent_embeddings/src/embedding_batch_preparer.rs
new file mode 100644
index 000000000..3d234695c
--- /dev/null
+++ b/paddler_agent_embeddings/src/embedding_batch_preparer.rs
@@ -0,0 +1,118 @@
+use llama_cpp_bindings::model::AddBos;
+use llama_cpp_bindings::model::ParseSpecialTokens;
+
+use paddler_agent_runtime::agent_request::AgentRequest;
+use paddler_agent_runtime::loaded_llama_model::LoadedLlamaModel;
+use paddler_agent_runtime::prepares_scheduler_command::PreparesSchedulerCommand;
+use paddler_messaging::embedding_document_exceeding_sequence_context_details::EmbeddingDocumentExceedingSequenceContextDetails;
+use paddler_messaging::embedding_document_tokenized_to_nothing_details::EmbeddingDocumentTokenizedToNothingDetails;
+use paddler_messaging::embedding_result::EmbeddingResult;
+use paddler_messaging::oversized_embedding_document_details::OversizedEmbeddingDocumentDetails;
+use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
+
+use crate::embedding_error::EmbeddingError;
+use crate::embedding_input_tokenized::EmbeddingInputTokenized;
+use crate::prepared_embedding_batch_request::PreparedEmbeddingBatchRequest;
+use crate::sequence_context_limit::SequenceContextLimit;
+use crate::unembeddable_embedding_document::UnembeddableEmbeddingDocument;
+
+pub struct EmbeddingBatchPreparer {
+    pub loaded_llama_model: LoadedLlamaModel,
+    pub n_batch: usize,
+    pub sequence_context_limit: SequenceContextLimit,
+}
+
+impl EmbeddingBatchPreparer {
+    fn prepare(
+        &self,
+        AgentRequest {
+            params:
+                GenerateEmbeddingBatchParams {
+                    input_batch,
+                    normalization_method,
+                },
+            response_tx: generated_embedding_tx,
+            slot_guard,
+            stop_rx: generate_embedding_stop_rx,
+        }: AgentRequest<GenerateEmbeddingBatchParams, EmbeddingResult>,
+    ) -> Result<PreparedEmbeddingBatchRequest, EmbeddingError> {
+        let mut inputs = Vec::with_capacity(input_batch.len());
+        let mut unembeddable_documents = Vec::new();
+
+        for input in input_batch {
+            let tokens = self
+                .loaded_llama_model
+                .model
+                .str_to_token(&input.content, AddBos::Always, ParseSpecialTokens::Always)
+                .map_err(|source| EmbeddingError::InputTokenizationFailed {
+                    source_document_id: input.id.clone(),
+                    source,
+                })?;
+
+            let document_tokens = tokens.len();
+
+            if tokens.is_empty() {
+                unembeddable_documents.push(UnembeddableEmbeddingDocument::TokenizedToNothing(
+                    EmbeddingDocumentTokenizedToNothingDetails {
+                        source_document_id: input.id,
+                    },
+                ));
+            } else if document_tokens > self.n_batch {
+                unembeddable_documents.push(UnembeddableEmbeddingDocument::ExceedsBatchSize(
+                    OversizedEmbeddingDocumentDetails {
+                        document_tokens,
+                        n_batch: self.n_batch,
+                        source_document_id: input.id,
+                    },
+                ));
+            } else if let SequenceContextLimit::Tokens(sequence_context_size) =
+                self.sequence_context_limit
+                && document_tokens > sequence_context_size
+            {
+                unembeddable_documents.push(UnembeddableEmbeddingDocument::ExceedsSequenceContext(
+                    EmbeddingDocumentExceedingSequenceContextDetails {
+                        document_tokens,
+                        sequence_context_size,
+                        source_document_id: input.id,
+                    },
+                ));
+            } else {
+                inputs.push(EmbeddingInputTokenized {
+                    id: input.id,
+                    tokens,
+                });
+            }
+        }
+
+        Ok(PreparedEmbeddingBatchRequest {
+            generate_embedding_stop_rx,
+            generated_embedding_tx,
+            inputs,
+            normalization_method,
+            unembeddable_documents,
+            slot_guard,
+        })
+    }
+}
+
+impl PreparesSchedulerCommand for EmbeddingBatchPreparer {
+    type Command = PreparedEmbeddingBatchRequest;
+    type Request = AgentRequest<GenerateEmbeddingBatchParams, EmbeddingResult>;
+
+    fn prepare_scheduler_command(
+        &self,
+        agent_name: Option<&str>,
+        request: Self::Request,
+    ) -> Option<Self::Command> {
+        let generated_embedding_tx = request.response_tx.clone();
+
+        match self.prepare(request) {
+            Ok(prepared) => Some(prepared),
+            Err(rejection) => {
+                rejection.report(agent_name, &generated_embedding_tx);
+
+                None
+            }
+        }
+    }
+}
diff --git a/paddler_agent_embeddings/src/embedding_batch_processor.rs b/paddler_agent_embeddings/src/embedding_batch_processor.rs
new file mode 100644
index 000000000..25d05b61d
--- /dev/null
+++ b/paddler_agent_embeddings/src/embedding_batch_processor.rs
@@ -0,0 +1,167 @@
+use std::collections::BTreeSet;
+
+use llama_cpp_bindings::context::LlamaContext;
+use llama_cpp_bindings::llama_batch::LlamaBatch;
+
+use paddler_messaging::embedding::Embedding;
+use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
+use paddler_messaging::embedding_result::EmbeddingResult;
+use paddler_messaging::embedding_vector::EmbeddingVector;
+
+use crate::active_embedding_request::ActiveEmbeddingRequest;
+use crate::batched_embedding_input::BatchedEmbeddingInput;
+use crate::embedding_error::EmbeddingError;
+use crate::embedding_input_tokenized::EmbeddingInputTokenized;
+use crate::embedding_normalization::EmbeddingNormalization;
+use crate::embedding_request_step::EmbeddingRequestStep;
+use crate::embedding_scheduler_context::EmbeddingSchedulerContext;
+use crate::next_embedding_batch::next_embedding_batch;
+use crate::normalizes_embedding::NormalizesEmbedding;
+
+pub struct EmbeddingBatchProcessor<
+    'scheduler,
+    'model,
+    TNormalization: NormalizesEmbedding = EmbeddingNormalization,
+> {
+    pub batch: &'scheduler mut LlamaBatch<'static>,
+    pub llama_context: &'scheduler mut LlamaContext<'model>,
+    pub normalization: TNormalization,
+    pub scheduler_context: &'scheduler EmbeddingSchedulerContext,
+}
+
+impl<TNormalization: NormalizesEmbedding> EmbeddingBatchProcessor<'_, '_, TNormalization> {
+    #[must_use]
+    pub fn embed_next_batch(
+        &mut self,
+        active_requests: Vec<ActiveEmbeddingRequest>,
+    ) -> Vec<ActiveEmbeddingRequest> {
+        let agent_name = self.scheduler_context.agent_name.as_deref();
+        let mut active_requests: Vec<ActiveEmbeddingRequest> = active_requests
+            .into_iter()
+            .flat_map(|active_request| active_request.continue_unless_stop_requested(agent_name))
+            .collect();
+        let batched_inputs = next_embedding_batch(
+            &mut active_requests,
+            self.scheduler_context.n_batch,
+            self.scheduler_context.desired_slots_total,
+        );
+
+        if batched_inputs.is_empty() {
+            return active_requests;
+        }
+
+        let mut request_steps = vec![EmbeddingRequestStep::Continuing; active_requests.len()];
+
+        match self.decode(&batched_inputs) {
+            Ok(()) => {
+                self.deliver_embeddings(batched_inputs, &mut active_requests, &mut request_steps);
+            }
+            Err(batch_error) => {
+                let failed_request_indexes: BTreeSet<usize> = batched_inputs
+                    .iter()
+                    .map(|batched_input| batched_input.request_index)
+                    .collect();
+
+                batch_error.report_to_each(
+                    agent_name,
+                    failed_request_indexes.iter().map(|request_index| {
+                        &active_requests[*request_index].generated_embedding_tx
+                    }),
+                );
+
+                for request_index in failed_request_indexes {
+                    request_steps[request_index] = EmbeddingRequestStep::Ended;
+                }
+            }
+        }
+
+        active_requests
+            .into_iter()
+            .zip(request_steps)
+            .filter(|(_active_request, request_step)| {
+                *request_step == EmbeddingRequestStep::Continuing
+            })
+            .flat_map(|(active_request, _request_step)| {
+                active_request.continue_unless_every_input_is_embedded(agent_name)
+            })
+            .collect()
+    }
+
+    fn decode(&mut self, batched_inputs: &[BatchedEmbeddingInput]) -> Result<(), EmbeddingError> {
+        self.batch.clear();
+
+        for BatchedEmbeddingInput {
+            input, sequence_id, ..
+        } in batched_inputs
+        {
+            self.batch
+                .add_sequence(&input.tokens, *sequence_id, true)
+                .map_err(EmbeddingError::BatchAssemblyFailed)?;
+        }
+
+        self.llama_context.clear_kv_cache();
+        self.llama_context
+            .decode(self.batch)
+            .map_err(EmbeddingError::DecodeFailed)
+    }
+
+    fn deliver_embeddings(
+        &self,
+        batched_inputs: Vec<BatchedEmbeddingInput>,
+        active_requests: &mut [ActiveEmbeddingRequest],
+        request_steps: &mut [EmbeddingRequestStep],
+    ) {
+        let agent_name = self.scheduler_context.agent_name.as_deref();
+
+        for BatchedEmbeddingInput {
+            input,
+            request_index,
+            sequence_id,
+        } in batched_inputs
+        {
+            if request_steps[request_index] == EmbeddingRequestStep::Ended {
+                continue;
+            }
+
+            let active_request = &mut active_requests[request_index];
+            let delivery = self
+                .embedding_of(&input, sequence_id, &active_request.normalization_method)
+                .and_then(|embedding| {
+                    active_request
+                        .generated_embedding_tx
+                        .send(EmbeddingResult::Embedding(embedding))
+                        .map_err(EmbeddingError::ClientDisconnected)
+                });
+
+            match delivery {
+                Ok(()) => active_request.embeddings_emitted += 1,
+                Err(delivery_error) => {
+                    delivery_error.report(agent_name, &active_request.generated_embedding_tx);
+                    request_steps[request_index] = EmbeddingRequestStep::Ended;
+                }
+            }
+        }
+    }
+
+    fn embedding_of(
+        &self,
+        input: &EmbeddingInputTokenized,
+        sequence_id: i32,
+        normalization_method: &EmbeddingNormalizationMethod,
+    ) -> Result<Embedding, EmbeddingError> {
+        let embedding = self
+            .llama_context
+            .embeddings_seq_ith(sequence_id)
+            .map_err(EmbeddingError::EmbeddingsUnavailable)?;
+
+        Ok(Embedding {
+            embedding: EmbeddingVector::from(
+                self.normalization
+                    .normalize(embedding.to_vec(), normalization_method)?,
+            ),
+            normalization_method: normalization_method.clone(),
+            pooling_type: self.scheduler_context.pooling_type.clone(),
+            source_document_id: input.id.clone(),
+        })
+    }
+}
diff --git a/paddler_agent_embeddings/src/embedding_error.rs b/paddler_agent_embeddings/src/embedding_error.rs
new file mode 100644
index 000000000..39e456bf8
--- /dev/null
+++ b/paddler_agent_embeddings/src/embedding_error.rs
@@ -0,0 +1,209 @@
+use std::num::TryFromIntError;
+
+use llama_cpp_bindings::batch_add_error::BatchAddError;
+use llama_cpp_bindings::error::DecodeError;
+use llama_cpp_bindings::error::EmbeddingsError;
+use llama_cpp_bindings::error::StringToTokenError;
+use log::error;
+use tokio::sync::mpsc;
+use tokio::sync::mpsc::error::SendError;
+
+use paddler_agent_runtime::agent_runtime_error::AgentRuntimeError;
+use paddler_agent_runtime::rejection_description::rejection_description;
+use paddler_agent_runtime::send_result_or_warn::send_result_or_warn;
+use paddler_messaging::embedding_result::EmbeddingResult;
+
+#[derive(Debug, thiserror::Error)]
+pub enum EmbeddingError {
+    #[error(transparent)]
+    AgentRuntime(#[from] AgentRuntimeError),
+
+    #[error("failed to tokenize embedding input {source_document_id:?}: {source}")]
+    InputTokenizationFailed {
+        source_document_id: String,
+        #[source]
+        source: StringToTokenError,
+    },
+
+    #[error("no model is loaded")]
+    ModelNotLoaded,
+
+    #[error("the scheduler is no longer accepting requests")]
+    SchedulerUnavailable,
+
+    #[error("failed to add an embedding input to the batch: {0}")]
+    BatchAssemblyFailed(#[source] BatchAddError),
+
+    #[error("the client stopped listening for embeddings: {0}")]
+    ClientDisconnected(#[source] SendError<EmbeddingResult>),
+
+    #[error("llama.cpp rejected the embedding batch: {0}")]
+    DecodeFailed(#[source] DecodeError),
+
+    #[error("{dimensions}-dimensional embedding is too long for RMS normalization: {source}")]
+    EmbeddingTooLongForRmsNormalization {
+        dimensions: usize,
+        #[source]
+        source: TryFromIntError,
+    },
+
+    #[error("failed to read the embedding of a batch sequence: {0}")]
+    EmbeddingsUnavailable(#[source] EmbeddingsError),
+}
+
+impl EmbeddingError {
+    pub fn report(
+        self,
+        agent_name: Option<&str>,
+        generated_embedding_tx: &mpsc::UnboundedSender<EmbeddingResult>,
+    ) {
+        self.report_to_each(agent_name, [generated_embedding_tx]);
+    }
+
+    pub fn report_to_each<'sender>(
+        self,
+        agent_name: Option<&str>,
+        generated_embedding_txs: impl IntoIterator<
+            Item = &'sender mpsc::UnboundedSender<EmbeddingResult>,
+        >,
+    ) {
+        let message = rejection_description(agent_name, &self);
+
+        error!("{message}");
+
+        let result = match self {
+            Self::AgentRuntime(_) => EmbeddingResult::AgentRuntimeFailed(message),
+            Self::BatchAssemblyFailed(_) => EmbeddingResult::BatchAssemblyFailed(message),
+            Self::ClientDisconnected(_) => return,
+            Self::DecodeFailed(_) => EmbeddingResult::DecodeFailed(message),
+            Self::EmbeddingTooLongForRmsNormalization { .. } => {
+                EmbeddingResult::EmbeddingTooLongForRmsNormalization(message)
+            }
+            Self::EmbeddingsUnavailable(_) => EmbeddingResult::EmbeddingsUnavailable(message),
+            Self::InputTokenizationFailed { .. } => {
+                EmbeddingResult::InputTokenizationFailed(message)
+            }
+            Self::ModelNotLoaded => EmbeddingResult::ModelNotLoaded(message),
+            Self::SchedulerUnavailable => EmbeddingResult::SchedulerUnavailable(message),
+        };
+
+        for generated_embedding_tx in generated_embedding_txs {
+            send_result_or_warn(agent_name, generated_embedding_tx, result.clone());
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::ffi::CString;
+    use std::io;
+    use std::mem::discriminant;
+
+    use llama_cpp_bindings::batch_add_error::BatchAddError;
+    use llama_cpp_bindings::error::DecodeError;
+    use llama_cpp_bindings::error::EmbeddingsError;
+    use llama_cpp_bindings::error::StringToTokenError;
+    use tokio::sync::mpsc;
+    use tokio::sync::mpsc::error::SendError;
+    use tokio::sync::mpsc::error::TryRecvError;
+
+    use paddler_agent_runtime::agent_runtime_error::AgentRuntimeError;
+    use paddler_messaging::embedding_result::EmbeddingResult;
+
+    use super::EmbeddingError;
+
+    #[test]
+    fn every_embedding_failure_reports_its_own_embedding_result() {
+        let description = String::new;
+        let failures = [
+            (
+                EmbeddingError::AgentRuntime(AgentRuntimeError::AvailableParallelismUnknown(
+                    io::Error::other("parallelism unknown"),
+                )),
+                EmbeddingResult::AgentRuntimeFailed(description()),
+            ),
+            (
+                EmbeddingError::BatchAssemblyFailed(BatchAddError::EmptyBuffer),
+                EmbeddingResult::BatchAssemblyFailed(description()),
+            ),
+            (
+                EmbeddingError::DecodeFailed(DecodeError::NoKvCacheSlot),
+                EmbeddingResult::DecodeFailed(description()),
+            ),
+            (
+                EmbeddingError::EmbeddingTooLongForRmsNormalization {
+                    dimensions: usize::MAX,
+                    source: u16::try_from(usize::MAX).unwrap_err(),
+                },
+                EmbeddingResult::EmbeddingTooLongForRmsNormalization(description()),
+            ),
+            (
+                EmbeddingError::EmbeddingsUnavailable(EmbeddingsError::NextnEmbeddingsNotEnabled),
+                EmbeddingResult::EmbeddingsUnavailable(description()),
+            ),
+            (
+                EmbeddingError::InputTokenizationFailed {
+                    source_document_id: "document".to_owned(),
+                    source: StringToTokenError::NulError(CString::new("a \0 byte").unwrap_err()),
+                },
+                EmbeddingResult::InputTokenizationFailed(description()),
+            ),
+            (
+                EmbeddingError::ModelNotLoaded,
+                EmbeddingResult::ModelNotLoaded(description()),
+            ),
+            (
+                EmbeddingError::SchedulerUnavailable,
+                EmbeddingResult::SchedulerUnavailable(description()),
+            ),
+        ];
+
+        for (embedding_error, expected_result) in failures {
+            let (generated_embedding_tx, mut generated_embedding_rx) = mpsc::unbounded_channel();
+
+            embedding_error.report(None, &generated_embedding_tx);
+
+            assert_eq!(
+                generated_embedding_rx.try_recv().as_ref().map(discriminant),
+                Ok(discriminant(&expected_result))
+            );
+        }
+    }
+
+    #[test]
+    fn sends_nothing_to_a_client_that_already_stopped_listening() {
+        let (generated_embedding_tx, mut generated_embedding_rx) = mpsc::unbounded_channel();
+
+        EmbeddingError::ClientDisconnected(SendError(EmbeddingResult::Done))
+            .report(Some("agent"), &generated_embedding_tx);
+
+        assert_eq!(generated_embedding_rx.try_recv(), Err(TryRecvError::Empty));
+    }
+
+    #[test]
+    fn tolerates_a_disconnected_client() {
+        let (generated_embedding_tx, generated_embedding_rx) = mpsc::unbounded_channel();
+
+        drop(generated_embedding_rx);
+
+        EmbeddingError::SchedulerUnavailable.report(None, &generated_embedding_tx);
+
+        assert!(generated_embedding_tx.is_closed());
+    }
+
+    #[test]
+    fn reports_one_failure_to_every_request_it_affects() {
+        let (first_embedding_tx, mut first_embedding_rx) = mpsc::unbounded_channel();
+        let (second_embedding_tx, mut second_embedding_rx) = mpsc::unbounded_channel();
+
+        EmbeddingError::DecodeFailed(DecodeError::NoKvCacheSlot)
+            .report_to_each(None, [&first_embedding_tx, &second_embedding_tx]);
+
+        let expected_result = Ok(EmbeddingResult::DecodeFailed(
+            EmbeddingError::DecodeFailed(DecodeError::NoKvCacheSlot).to_string(),
+        ));
+
+        assert_eq!(first_embedding_rx.try_recv(), expected_result);
+        assert_eq!(second_embedding_rx.try_recv(), expected_result);
+    }
+}
diff --git a/paddler_agent/src/embedding_input_tokenized.rs b/paddler_agent_embeddings/src/embedding_input_tokenized.rs
similarity index 80%
rename from paddler_agent/src/embedding_input_tokenized.rs
rename to paddler_agent_embeddings/src/embedding_input_tokenized.rs
index 28653f8d6..daf6310e7 100644
--- a/paddler_agent/src/embedding_input_tokenized.rs
+++ b/paddler_agent_embeddings/src/embedding_input_tokenized.rs
@@ -1,5 +1,6 @@
 use llama_cpp_bindings::token::LlamaToken;
 
+#[derive(Debug, Eq, PartialEq)]
 pub struct EmbeddingInputTokenized {
     pub id: String,
     pub tokens: Vec<LlamaToken>,
diff --git a/paddler_agent_embeddings/src/embedding_llama_context_settings.rs b/paddler_agent_embeddings/src/embedding_llama_context_settings.rs
new file mode 100644
index 000000000..87e19fce2
--- /dev/null
+++ b/paddler_agent_embeddings/src/embedding_llama_context_settings.rs
@@ -0,0 +1,43 @@
+use paddler_agent_runtime::agent_runtime_error::AgentRuntimeError;
+use paddler_agent_runtime::inference_runtime_context::InferenceRuntimeContext;
+use paddler_agent_runtime::llama_context_settings::LlamaContextSettings;
+use paddler_agent_runtime::loaded_llama_model::LoadedLlamaModel;
+use paddler_agent_runtime::resident_llama_context::ResidentLlamaContext;
+use paddler_inference_parameters::pooling_type::PoolingType;
+use paddler_messaging::inference_capability::InferenceCapability;
+
+use crate::agent_pooling_type::AgentPoolingType;
+use crate::converts_to_llama_pooling_type::ConvertsToLlamaPoolingType as _;
+
+pub struct EmbeddingLlamaContextSettings<'parameters> {
+    pub llama_context_settings: LlamaContextSettings<'parameters>,
+    pub pooling_type: PoolingType,
+}
+
+impl EmbeddingLlamaContextSettings<'_> {
+    pub fn create_llama_context<'model>(
+        self,
+        loaded_llama_model: &'model LoadedLlamaModel,
+        inference_runtime_context: &InferenceRuntimeContext,
+    ) -> Result<ResidentLlamaContext<'model>, AgentRuntimeError> {
+        let Self {
+            llama_context_settings,
+            pooling_type,
+        } = self;
+        let n_ubatch = llama_context_settings
+            .model_runtime_parameters
+            .n_batch
+            .tokens()
+            .get();
+
+        loaded_llama_model.create_llama_context(
+            inference_runtime_context,
+            InferenceCapability::Embeddings,
+            llama_context_settings
+                .into_llama_context_params()
+                .with_embeddings(true)
+                .with_n_ubatch(n_ubatch)
+                .with_pooling_type(AgentPoolingType(pooling_type).to_llama_pooling_type()),
+        )
+    }
+}
diff --git a/paddler_agent_embeddings/src/embedding_normalization.rs b/paddler_agent_embeddings/src/embedding_normalization.rs
new file mode 100644
index 000000000..6b690b2ba
--- /dev/null
+++ b/paddler_agent_embeddings/src/embedding_normalization.rs
@@ -0,0 +1,17 @@
+use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
+
+use crate::embedding_error::EmbeddingError;
+use crate::normalization::normalize_embedding::normalize_embedding;
+use crate::normalizes_embedding::NormalizesEmbedding;
+
+pub struct EmbeddingNormalization;
+
+impl NormalizesEmbedding for EmbeddingNormalization {
+    fn normalize(
+        &self,
+        embedding: Vec<f32>,
+        normalization_method: &EmbeddingNormalizationMethod,
+    ) -> Result<Vec<f32>, EmbeddingError> {
+        normalize_embedding(embedding, normalization_method)
+    }
+}
diff --git a/paddler_agent_embeddings/src/embedding_pipeline.rs b/paddler_agent_embeddings/src/embedding_pipeline.rs
new file mode 100644
index 000000000..3e53c822a
--- /dev/null
+++ b/paddler_agent_embeddings/src/embedding_pipeline.rs
@@ -0,0 +1,130 @@
+use std::sync::Arc;
+
+use llama_cpp_bindings::llama_batch::LlamaBatch;
+use tokio::sync::mpsc::UnboundedSender;
+use tokio::sync::mpsc::unbounded_channel;
+use tokio_util::sync::CancellationToken;
+use tokio_util::task::TaskTracker;
+
+use paddler_agent_runtime::continue_startup_unless_shutting_down::continue_startup_unless_shutting_down;
+use paddler_agent_runtime::inference_runtime_context::InferenceRuntimeContext;
+use paddler_agent_runtime::loaded_llama_model::LoadedLlamaModel;
+use paddler_agent_runtime::run_scheduler_unless_startup_abandoned::run_scheduler_unless_startup_abandoned;
+use paddler_agent_runtime::scheduler_handle::SchedulerHandle;
+use paddler_agent_runtime::scheduler_inbox::SchedulerInbox;
+use paddler_agent_runtime::scheduler_ready::SchedulerReady;
+use paddler_agent_runtime::scheduler_request_preparer::SchedulerRequestPreparer;
+use paddler_agent_runtime::start_scheduler_thread::start_scheduler_thread;
+use paddler_agent_runtime::startup_outcome::StartupOutcome;
+use paddler_inference_parameters::model_runtime_parameters::ModelRuntimeParameters;
+use paddler_inference_parameters::pooling_type::PoolingType;
+
+use crate::embedding_batch_preparer::EmbeddingBatchPreparer;
+use crate::embedding_error::EmbeddingError;
+use crate::embedding_llama_context_settings::EmbeddingLlamaContextSettings;
+use crate::embedding_scheduler::EmbeddingScheduler;
+use crate::embedding_scheduler_context::EmbeddingSchedulerContext;
+use crate::sequence_context_limit::SequenceContextLimit;
+
+pub struct EmbeddingPipeline {
+    pub inference_runtime_context: InferenceRuntimeContext,
+    pub loaded_llama_model: LoadedLlamaModel,
+    pub model_runtime_parameters: ModelRuntimeParameters,
+    pub pooling_type: PoolingType,
+}
+
+impl EmbeddingPipeline {
+    pub async fn spawn(
+        self,
+        cancellation_token: &CancellationToken,
+    ) -> Result<StartupOutcome<SchedulerHandle<EmbeddingBatchPreparer>>, EmbeddingError> {
+        start_scheduler_thread(
+            cancellation_token,
+            move |agent_shutdown, scheduler_startup_tx| {
+                self.run_scheduler(agent_shutdown, scheduler_startup_tx)
+            },
+        )
+        .await
+    }
+
+    fn run_scheduler(
+        self,
+        agent_shutdown: &CancellationToken,
+        scheduler_startup_tx: &UnboundedSender<
+            Result<SchedulerReady<EmbeddingBatchPreparer>, EmbeddingError>,
+        >,
+    ) -> Result<(), EmbeddingError> {
+        let Self {
+            inference_runtime_context,
+            loaded_llama_model,
+            model_runtime_parameters,
+            pooling_type,
+        } = self;
+        let desired_slots_total = inference_runtime_context
+            .slot_aggregated_status
+            .desired_slots_total;
+        let n_batch = model_runtime_parameters.n_batch.tokens_usize();
+
+        continue_startup_unless_shutting_down(agent_shutdown, || {
+            let llama_context = EmbeddingLlamaContextSettings {
+                llama_context_settings: loaded_llama_model.llama_context_settings(
+                    &model_runtime_parameters,
+                    u32::from(desired_slots_total),
+                ),
+                pooling_type: pooling_type.clone(),
+            }
+            .create_llama_context(&loaded_llama_model, &inference_runtime_context)?;
+            let batch = LlamaBatch::new(
+                model_runtime_parameters.n_batch.capacity(),
+                i32::from(desired_slots_total),
+            );
+
+            continue_startup_unless_shutting_down(agent_shutdown, || {
+                loaded_llama_model
+                    .warm_up_and_serve(
+                        llama_context,
+                        batch,
+                        desired_slots_total,
+                        |llama_context, batch| {
+                            let (scheduler_message_tx, scheduler_message_rx) = unbounded_channel();
+
+                            run_scheduler_unless_startup_abandoned(
+                                scheduler_startup_tx,
+                                Ok(SchedulerReady {
+                                    request_preparer: Arc::new(SchedulerRequestPreparer {
+                                        agent_name: inference_runtime_context.agent_name.clone(),
+                                        preparation: EmbeddingBatchPreparer {
+                                            loaded_llama_model: loaded_llama_model.clone(),
+                                            n_batch,
+                                            sequence_context_limit: SequenceContextLimit::of(
+                                                &llama_context,
+                                            ),
+                                        },
+                                        preparation_tasks: TaskTracker::new(),
+                                        scheduler_message_tx,
+                                    }),
+                                }),
+                                || {
+                                    EmbeddingScheduler {
+                                        active_requests: Vec::new(),
+                                        agent_shutdown: agent_shutdown.clone(),
+                                        batch,
+                                        llama_context,
+                                        scheduler_context: EmbeddingSchedulerContext {
+                                            agent_name: inference_runtime_context.agent_name,
+                                            desired_slots_total,
+                                            n_batch,
+                                            pooling_type,
+                                        },
+                                        scheduler_inbox: SchedulerInbox::new(scheduler_message_rx),
+                                    }
+                                    .run();
+                                },
+                            );
+                        },
+                    )
+                    .map_err(EmbeddingError::from)
+            })
+        })
+    }
+}
diff --git a/paddler_agent_embeddings/src/embedding_request_continuation.rs b/paddler_agent_embeddings/src/embedding_request_continuation.rs
new file mode 100644
index 000000000..b73fd63c0
--- /dev/null
+++ b/paddler_agent_embeddings/src/embedding_request_continuation.rs
@@ -0,0 +1,20 @@
+use std::option::IntoIter;
+
+use crate::active_embedding_request::ActiveEmbeddingRequest;
+
+pub enum EmbeddingRequestContinuation {
+    Concluded,
+    Continues(ActiveEmbeddingRequest),
+}
+
+impl IntoIterator for EmbeddingRequestContinuation {
+    type IntoIter = IntoIter<ActiveEmbeddingRequest>;
+    type Item = ActiveEmbeddingRequest;
+
+    fn into_iter(self) -> Self::IntoIter {
+        match self {
+            Self::Concluded => None.into_iter(),
+            Self::Continues(active_request) => Some(active_request).into_iter(),
+        }
+    }
+}
diff --git a/paddler_agent_embeddings/src/embedding_request_step.rs b/paddler_agent_embeddings/src/embedding_request_step.rs
new file mode 100644
index 000000000..3e242db16
--- /dev/null
+++ b/paddler_agent_embeddings/src/embedding_request_step.rs
@@ -0,0 +1,5 @@
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+pub enum EmbeddingRequestStep {
+    Continuing,
+    Ended,
+}
diff --git a/paddler_agent_embeddings/src/embedding_scheduler.rs b/paddler_agent_embeddings/src/embedding_scheduler.rs
new file mode 100644
index 000000000..faa9a9c2f
--- /dev/null
+++ b/paddler_agent_embeddings/src/embedding_scheduler.rs
@@ -0,0 +1,92 @@
+use std::mem::take;
+
+use llama_cpp_bindings::llama_batch::LlamaBatch;
+use log::info;
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent_runtime::resident_llama_context::ResidentLlamaContext;
+use paddler_agent_runtime::scheduler_inbox::SchedulerInbox;
+
+use crate::active_embedding_request::ActiveEmbeddingRequest;
+use crate::embedding_batch_processor::EmbeddingBatchProcessor;
+use crate::embedding_error::EmbeddingError;
+use crate::embedding_normalization::EmbeddingNormalization;
+use crate::embedding_scheduler_context::EmbeddingSchedulerContext;
+use crate::prepared_embedding_batch_request::PreparedEmbeddingBatchRequest;
+
+pub struct EmbeddingScheduler<'model> {
+    pub active_requests: Vec<ActiveEmbeddingRequest>,
+    pub agent_shutdown: CancellationToken,
+    pub batch: LlamaBatch<'static>,
+    pub llama_context: ResidentLlamaContext<'model>,
+    pub scheduler_context: EmbeddingSchedulerContext,
+    pub scheduler_inbox: SchedulerInbox<PreparedEmbeddingBatchRequest>,
+}
+
+impl EmbeddingScheduler<'_> {
+    pub fn run(&mut self) {
+        info!(
+            "{:?}: embedding scheduler started",
+            self.scheduler_context.agent_name
+        );
+
+        while !self.agent_shutdown.is_cancelled() {
+            if self.active_requests.is_empty() {
+                if self.scheduler_inbox.shutdown_requested {
+                    break;
+                }
+
+                self.wait_for_next_scheduler_message();
+            } else {
+                self.accept_new_scheduler_messages();
+            }
+
+            let active_requests = take(&mut self.active_requests);
+
+            self.active_requests = EmbeddingBatchProcessor {
+                batch: &mut self.batch,
+                llama_context: &mut self.llama_context,
+                normalization: EmbeddingNormalization,
+                scheduler_context: &self.scheduler_context,
+            }
+            .embed_next_batch(active_requests);
+        }
+
+        self.reject_unfinished_requests();
+        self.llama_context.synchronize();
+        self.llama_context.detach_threadpool();
+
+        info!(
+            "{:?}: embedding scheduler stopped",
+            self.scheduler_context.agent_name
+        );
+    }
+
+    fn accept_new_scheduler_messages(&mut self) {
+        while let Some(request) = self.scheduler_inbox.take_next_waiting_command() {
+            self.admit(request);
+        }
+    }
+
+    fn admit(&mut self, request: PreparedEmbeddingBatchRequest) {
+        self.active_requests
+            .extend(request.admit(self.scheduler_context.agent_name.as_deref()));
+    }
+
+    fn reject_unfinished_requests(&mut self) {
+        let agent_name = self.scheduler_context.agent_name.as_deref();
+
+        for active_request in self.active_requests.drain(..) {
+            EmbeddingError::SchedulerUnavailable
+                .report(agent_name, &active_request.generated_embedding_tx);
+        }
+
+        self.scheduler_inbox.reject_waiting_commands(agent_name);
+    }
+
+    fn wait_for_next_scheduler_message(&mut self) {
+        if let Some(request) = self.scheduler_inbox.wait_for_next_command() {
+            self.admit(request);
+        }
+    }
+}
diff --git a/paddler_agent_embeddings/src/embedding_scheduler_context.rs b/paddler_agent_embeddings/src/embedding_scheduler_context.rs
new file mode 100644
index 000000000..f4dcf9069
--- /dev/null
+++ b/paddler_agent_embeddings/src/embedding_scheduler_context.rs
@@ -0,0 +1,8 @@
+use paddler_inference_parameters::pooling_type::PoolingType;
+
+pub struct EmbeddingSchedulerContext {
+    pub agent_name: Option<String>,
+    pub desired_slots_total: u16,
+    pub n_batch: usize,
+    pub pooling_type: PoolingType,
+}
diff --git a/paddler_agent_embeddings/src/lib.rs b/paddler_agent_embeddings/src/lib.rs
new file mode 100644
index 000000000..b424b9236
--- /dev/null
+++ b/paddler_agent_embeddings/src/lib.rs
@@ -0,0 +1,21 @@
+pub mod active_embedding_request;
+pub mod agent_pooling_type;
+pub mod batched_embedding_input;
+pub mod converts_to_llama_pooling_type;
+pub mod embedding_batch_preparer;
+pub mod embedding_batch_processor;
+pub mod embedding_error;
+pub mod embedding_input_tokenized;
+pub mod embedding_llama_context_settings;
+pub mod embedding_normalization;
+pub mod embedding_pipeline;
+pub mod embedding_request_continuation;
+pub mod embedding_request_step;
+pub mod embedding_scheduler;
+pub mod embedding_scheduler_context;
+pub mod next_embedding_batch;
+pub mod normalization;
+pub mod normalizes_embedding;
+pub mod prepared_embedding_batch_request;
+pub mod sequence_context_limit;
+pub mod unembeddable_embedding_document;
diff --git a/paddler_agent_embeddings/src/next_embedding_batch.rs b/paddler_agent_embeddings/src/next_embedding_batch.rs
new file mode 100644
index 000000000..b7f74254e
--- /dev/null
+++ b/paddler_agent_embeddings/src/next_embedding_batch.rs
@@ -0,0 +1,179 @@
+use crate::active_embedding_request::ActiveEmbeddingRequest;
+use crate::batched_embedding_input::BatchedEmbeddingInput;
+
+#[must_use]
+pub fn next_embedding_batch(
+    active_requests: &mut [ActiveEmbeddingRequest],
+    n_batch: usize,
+    sequences_in_batch: u16,
+) -> Vec<BatchedEmbeddingInput> {
+    let mut batched_inputs = Vec::new();
+    let mut batched_tokens = 0;
+    let mut next_sequence_id: u16 = 0;
+
+    for (request_index, active_request) in active_requests.iter_mut().enumerate() {
+        let mut request_sequences = 0;
+
+        while request_sequences < active_request.sequences()
+            && next_sequence_id < sequences_in_batch
+        {
+            let Some(input) = active_request.pending_inputs.pop_front() else {
+                break;
+            };
+
+            if batched_tokens + input.tokens.len() > n_batch {
+                active_request.pending_inputs.push_front(input);
+
+                break;
+            }
+
+            batched_tokens += input.tokens.len();
+            batched_inputs.push(BatchedEmbeddingInput {
+                input,
+                request_index,
+                sequence_id: i32::from(next_sequence_id),
+            });
+            request_sequences += 1;
+            next_sequence_id += 1;
+        }
+    }
+
+    batched_inputs
+}
+
+#[cfg(test)]
+mod tests {
+    use std::collections::VecDeque;
+    use std::num::NonZeroU16;
+    use std::sync::Arc;
+
+    use llama_cpp_bindings::token::LlamaToken;
+    use tokio::sync::mpsc;
+
+    use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
+    use paddler_agent_status::slot_guard::SlotGuard;
+    use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
+
+    use super::next_embedding_batch;
+    use crate::active_embedding_request::ActiveEmbeddingRequest;
+    use crate::batched_embedding_input::BatchedEmbeddingInput;
+    use crate::embedding_input_tokenized::EmbeddingInputTokenized;
+    use crate::prepared_embedding_batch_request::PreparedEmbeddingBatchRequest;
+
+    fn document(id: &str, tokens: usize) -> EmbeddingInputTokenized {
+        EmbeddingInputTokenized {
+            id: id.to_owned(),
+            tokens: vec![LlamaToken::new(0); tokens],
+        }
+    }
+
+    fn active_request(
+        sequences: u16,
+        inputs: Vec<EmbeddingInputTokenized>,
+    ) -> ActiveEmbeddingRequest {
+        let (generated_embedding_tx, _generated_embedding_rx) = mpsc::unbounded_channel();
+        let (_generate_embedding_stop_tx, generate_embedding_stop_rx) = mpsc::unbounded_channel();
+
+        PreparedEmbeddingBatchRequest {
+            generate_embedding_stop_rx,
+            generated_embedding_tx,
+            inputs,
+            normalization_method: EmbeddingNormalizationMethod::None,
+            unembeddable_documents: Vec::new(),
+            slot_guard: SlotGuard::new(
+                Arc::new(SlotAggregatedStatus::new(8)),
+                NonZeroU16::new(sequences).unwrap(),
+            ),
+        }
+        .admit(None)
+        .into_iter()
+        .next()
+        .expect("a request with inputs to embed must stay active")
+    }
+
+    fn batched(
+        request_index: usize,
+        input: EmbeddingInputTokenized,
+        sequence_id: i32,
+    ) -> BatchedEmbeddingInput {
+        BatchedEmbeddingInput {
+            input,
+            request_index,
+            sequence_id,
+        }
+    }
+
+    #[test]
+    fn every_request_contributes_up_to_its_reserved_sequences() {
+        let mut active_requests = [
+            active_request(1, vec![document("a1", 1), document("a2", 1)]),
+            active_request(
+                2,
+                vec![document("b1", 1), document("b2", 1), document("b3", 1)],
+            ),
+        ];
+
+        assert_eq!(
+            next_embedding_batch(&mut active_requests, 100, 8),
+            [
+                batched(0, document("a1", 1), 0),
+                batched(1, document("b1", 1), 1),
+                batched(1, document("b2", 1), 2),
+            ]
+        );
+        assert_eq!(
+            active_requests[1].pending_inputs,
+            VecDeque::from([document("b3", 1)])
+        );
+    }
+
+    #[test]
+    fn a_request_without_enough_documents_leaves_its_reserved_sequences_unused() {
+        let mut active_requests = [
+            active_request(3, vec![document("a1", 1)]),
+            active_request(1, vec![document("b1", 1)]),
+        ];
+
+        assert_eq!(
+            next_embedding_batch(&mut active_requests, 100, 8),
+            [
+                batched(0, document("a1", 1), 0),
+                batched(1, document("b1", 1), 1)
+            ]
+        );
+    }
+
+    #[test]
+    fn a_document_that_does_not_fit_the_remaining_tokens_waits_for_the_next_batch() {
+        let mut active_requests = [
+            active_request(1, vec![document("a1", 6)]),
+            active_request(2, vec![document("b1", 5), document("b2", 3)]),
+        ];
+
+        assert_eq!(
+            next_embedding_batch(&mut active_requests, 10, 8),
+            [batched(0, document("a1", 6), 0)]
+        );
+        assert_eq!(
+            active_requests[1].pending_inputs,
+            VecDeque::from([document("b1", 5), document("b2", 3)])
+        );
+    }
+
+    #[test]
+    fn a_batch_holds_no_more_sequences_than_the_context_serves() {
+        let mut active_requests = [
+            active_request(2, vec![document("a1", 1), document("a2", 1)]),
+            active_request(2, vec![document("b1", 1), document("b2", 1)]),
+        ];
+
+        assert_eq!(
+            next_embedding_batch(&mut active_requests, 100, 3),
+            [
+                batched(0, document("a1", 1), 0),
+                batched(0, document("a2", 1), 1),
+                batched(1, document("b1", 1), 2),
+            ]
+        );
+    }
+}
diff --git a/paddler_agent/src/normalization/l2.rs b/paddler_agent_embeddings/src/normalization/l2.rs
similarity index 100%
rename from paddler_agent/src/normalization/l2.rs
rename to paddler_agent_embeddings/src/normalization/l2.rs
diff --git a/paddler_agent/src/normalization/mod.rs b/paddler_agent_embeddings/src/normalization/mod.rs
similarity index 100%
rename from paddler_agent/src/normalization/mod.rs
rename to paddler_agent_embeddings/src/normalization/mod.rs
diff --git a/paddler_agent/src/normalization/normalize_embedding.rs b/paddler_agent_embeddings/src/normalization/normalize_embedding.rs
similarity index 86%
rename from paddler_agent/src/normalization/normalize_embedding.rs
rename to paddler_agent_embeddings/src/normalization/normalize_embedding.rs
index 67a8c774f..3d031ce7c 100644
--- a/paddler_agent/src/normalization/normalize_embedding.rs
+++ b/paddler_agent_embeddings/src/normalization/normalize_embedding.rs
@@ -1,19 +1,19 @@
 use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
 
-use crate::embedding_batch_rejection::EmbeddingBatchRejection;
+use crate::embedding_error::EmbeddingError;
 use crate::normalization::l2::l2;
 use crate::normalization::rms_norm::rms_norm;
 
 pub fn normalize_embedding(
     mut embedding: Vec<f32>,
     normalization_method: &EmbeddingNormalizationMethod,
-) -> Result<Vec<f32>, EmbeddingBatchRejection> {
+) -> Result<Vec<f32>, EmbeddingError> {
     match normalization_method {
         EmbeddingNormalizationMethod::None => {}
         EmbeddingNormalizationMethod::L2 => l2(&mut embedding),
         EmbeddingNormalizationMethod::RmsNorm { epsilon } => {
             rms_norm(&mut embedding, *epsilon).map_err(|source| {
-                EmbeddingBatchRejection::EmbeddingTooLongForRmsNormalization {
+                EmbeddingError::EmbeddingTooLongForRmsNormalization {
                     dimensions: embedding.len(),
                     source,
                 }
@@ -29,7 +29,7 @@ mod tests {
     use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
 
     use super::normalize_embedding;
-    use crate::embedding_batch_rejection::EmbeddingBatchRejection;
+    use crate::embedding_error::EmbeddingError;
 
     #[test]
     fn scales_the_embedding_to_unit_length_with_l2() {
@@ -68,7 +68,7 @@ mod tests {
 
         assert!(matches!(
             result,
-            Err(EmbeddingBatchRejection::EmbeddingTooLongForRmsNormalization {
+            Err(EmbeddingError::EmbeddingTooLongForRmsNormalization {
                 dimensions,
                 ..
             }) if dimensions == oversized_length
diff --git a/paddler_agent/src/normalization/rms_norm.rs b/paddler_agent_embeddings/src/normalization/rms_norm.rs
similarity index 98%
rename from paddler_agent/src/normalization/rms_norm.rs
rename to paddler_agent_embeddings/src/normalization/rms_norm.rs
index a742923e8..184432b7c 100644
--- a/paddler_agent/src/normalization/rms_norm.rs
+++ b/paddler_agent_embeddings/src/normalization/rms_norm.rs
@@ -89,7 +89,7 @@ mod tests {
         let mut embedding: Vec<f32> = Vec::new();
         rms_norm(&mut embedding, 0.0).unwrap();
 
-        assert!(embedding.is_empty());
+        assert_eq!(embedding, Vec::<f32>::new());
     }
 
     #[test]
diff --git a/paddler_agent_embeddings/src/normalizes_embedding.rs b/paddler_agent_embeddings/src/normalizes_embedding.rs
new file mode 100644
index 000000000..eb9296525
--- /dev/null
+++ b/paddler_agent_embeddings/src/normalizes_embedding.rs
@@ -0,0 +1,11 @@
+use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
+
+use crate::embedding_error::EmbeddingError;
+
+pub trait NormalizesEmbedding {
+    fn normalize(
+        &self,
+        embedding: Vec<f32>,
+        normalization_method: &EmbeddingNormalizationMethod,
+    ) -> Result<Vec<f32>, EmbeddingError>;
+}
diff --git a/paddler_agent_embeddings/src/prepared_embedding_batch_request.rs b/paddler_agent_embeddings/src/prepared_embedding_batch_request.rs
new file mode 100644
index 000000000..74dc90735
--- /dev/null
+++ b/paddler_agent_embeddings/src/prepared_embedding_batch_request.rs
@@ -0,0 +1,195 @@
+use log::warn;
+use tokio::sync::mpsc;
+
+use paddler_agent_runtime::scheduler_command::SchedulerCommand;
+use paddler_agent_status::slot_guard::SlotGuard;
+use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
+use paddler_messaging::embedding_result::EmbeddingResult;
+
+use crate::active_embedding_request::ActiveEmbeddingRequest;
+use crate::embedding_error::EmbeddingError;
+use crate::embedding_input_tokenized::EmbeddingInputTokenized;
+use crate::embedding_request_continuation::EmbeddingRequestContinuation;
+use crate::unembeddable_embedding_document::UnembeddableEmbeddingDocument;
+
+pub struct PreparedEmbeddingBatchRequest {
+    pub generate_embedding_stop_rx: mpsc::UnboundedReceiver<()>,
+    pub generated_embedding_tx: mpsc::UnboundedSender<EmbeddingResult>,
+    pub inputs: Vec<EmbeddingInputTokenized>,
+    pub normalization_method: EmbeddingNormalizationMethod,
+    pub unembeddable_documents: Vec<UnembeddableEmbeddingDocument>,
+    pub slot_guard: SlotGuard,
+}
+
+impl PreparedEmbeddingBatchRequest {
+    #[must_use]
+    pub fn admit(self, agent_name: Option<&str>) -> EmbeddingRequestContinuation {
+        let Self {
+            generate_embedding_stop_rx,
+            generated_embedding_tx,
+            inputs,
+            normalization_method,
+            unembeddable_documents,
+            slot_guard,
+        } = self;
+
+        for unembeddable_document in unembeddable_documents {
+            warn!("{agent_name:?}: skipped an unembeddable document: {unembeddable_document:?}");
+
+            if let Err(send_error) =
+                generated_embedding_tx.send(unembeddable_document.into_embedding_result())
+            {
+                EmbeddingError::ClientDisconnected(send_error)
+                    .report(agent_name, &generated_embedding_tx);
+
+                return EmbeddingRequestContinuation::Concluded;
+            }
+        }
+
+        ActiveEmbeddingRequest {
+            embeddings_emitted: 0,
+            generate_embedding_stop_rx,
+            generated_embedding_tx,
+            normalization_method,
+            pending_inputs: inputs.into(),
+            slot_guard,
+        }
+        .continue_unless_every_input_is_embedded(agent_name)
+    }
+}
+
+impl SchedulerCommand for PreparedEmbeddingBatchRequest {
+    fn reject_because_the_scheduler_stopped(self, agent_name: Option<&str>) {
+        EmbeddingError::SchedulerUnavailable.report(agent_name, &self.generated_embedding_tx);
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::mem::discriminant;
+    use std::num::NonZeroU16;
+    use std::sync::Arc;
+
+    use llama_cpp_bindings::token::LlamaToken;
+    use tokio::sync::mpsc;
+    use tokio::sync::mpsc::error::TryRecvError;
+
+    use paddler_agent_runtime::scheduler_command::SchedulerCommand as _;
+    use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
+    use paddler_agent_status::slot_guard::SlotGuard;
+    use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
+    use paddler_messaging::embedding_result::EmbeddingResult;
+    use paddler_messaging::oversized_embedding_document_details::OversizedEmbeddingDocumentDetails;
+
+    use super::PreparedEmbeddingBatchRequest;
+    use crate::embedding_input_tokenized::EmbeddingInputTokenized;
+    use crate::embedding_request_continuation::EmbeddingRequestContinuation;
+    use crate::unembeddable_embedding_document::UnembeddableEmbeddingDocument;
+
+    fn oversized_document_details() -> OversizedEmbeddingDocumentDetails {
+        OversizedEmbeddingDocumentDetails {
+            document_tokens: 4096,
+            n_batch: 2048,
+            source_document_id: "oversized".to_owned(),
+        }
+    }
+
+    fn prepared_request(
+        generated_embedding_tx: mpsc::UnboundedSender<EmbeddingResult>,
+        inputs: Vec<EmbeddingInputTokenized>,
+    ) -> PreparedEmbeddingBatchRequest {
+        let (_generate_embedding_stop_tx, generate_embedding_stop_rx) = mpsc::unbounded_channel();
+
+        PreparedEmbeddingBatchRequest {
+            generate_embedding_stop_rx,
+            generated_embedding_tx,
+            inputs,
+            normalization_method: EmbeddingNormalizationMethod::None,
+            unembeddable_documents: vec![UnembeddableEmbeddingDocument::ExceedsBatchSize(
+                oversized_document_details(),
+            )],
+            slot_guard: SlotGuard::new(Arc::new(SlotAggregatedStatus::new(1)), NonZeroU16::MIN),
+        }
+    }
+
+    #[test]
+    fn rejects_an_embedding_batch_when_the_scheduler_stopped() {
+        let (generated_embedding_tx, mut generated_embedding_rx) = mpsc::unbounded_channel();
+
+        prepared_request(generated_embedding_tx, Vec::new())
+            .reject_because_the_scheduler_stopped(Some("agent"));
+
+        assert_eq!(
+            generated_embedding_rx.try_recv(),
+            Ok(EmbeddingResult::SchedulerUnavailable(
+                "agent: the scheduler is no longer accepting requests".to_owned()
+            ))
+        );
+    }
+
+    #[test]
+    fn a_batch_of_only_oversized_documents_reports_them_and_produces_no_embeddings() {
+        let (generated_embedding_tx, mut generated_embedding_rx) = mpsc::unbounded_channel();
+
+        assert_eq!(
+            discriminant(&prepared_request(generated_embedding_tx, Vec::new()).admit(None)),
+            discriminant(&EmbeddingRequestContinuation::Concluded)
+        );
+        assert_eq!(
+            generated_embedding_rx.try_recv(),
+            Ok(EmbeddingResult::DocumentExceedsBatchSize(
+                oversized_document_details()
+            ))
+        );
+        assert_eq!(
+            generated_embedding_rx.try_recv(),
+            Ok(EmbeddingResult::NoEmbeddingsProduced)
+        );
+        assert_eq!(
+            generated_embedding_rx.try_recv(),
+            Err(TryRecvError::Disconnected)
+        );
+    }
+
+    #[test]
+    fn a_client_that_stopped_listening_concludes_before_its_documents_are_embedded() {
+        let (generated_embedding_tx, generated_embedding_rx) = mpsc::unbounded_channel();
+
+        drop(generated_embedding_rx);
+
+        assert_eq!(
+            discriminant(
+                &prepared_request(
+                    generated_embedding_tx,
+                    vec![EmbeddingInputTokenized {
+                        id: "document".to_owned(),
+                        tokens: vec![LlamaToken::new(0)],
+                    }],
+                )
+                .admit(None)
+            ),
+            discriminant(&EmbeddingRequestContinuation::Concluded)
+        );
+    }
+
+    #[test]
+    fn a_batch_with_documents_to_embed_stays_open_after_reporting_its_oversized_documents() {
+        let (generated_embedding_tx, mut generated_embedding_rx) = mpsc::unbounded_channel();
+        let _continuation = prepared_request(
+            generated_embedding_tx,
+            vec![EmbeddingInputTokenized {
+                id: "document".to_owned(),
+                tokens: vec![LlamaToken::new(0)],
+            }],
+        )
+        .admit(None);
+
+        assert_eq!(
+            generated_embedding_rx.try_recv(),
+            Ok(EmbeddingResult::DocumentExceedsBatchSize(
+                oversized_document_details()
+            ))
+        );
+        assert_eq!(generated_embedding_rx.try_recv(), Err(TryRecvError::Empty));
+    }
+}
diff --git a/paddler_agent_embeddings/src/sequence_context_limit.rs b/paddler_agent_embeddings/src/sequence_context_limit.rs
new file mode 100644
index 000000000..e1ae8076a
--- /dev/null
+++ b/paddler_agent_embeddings/src/sequence_context_limit.rs
@@ -0,0 +1,18 @@
+use llama_cpp_bindings::context::LlamaContext;
+
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+pub enum SequenceContextLimit {
+    Tokens(usize),
+    Unlimited,
+}
+
+impl SequenceContextLimit {
+    #[must_use]
+    pub fn of(llama_context: &LlamaContext<'_>) -> Self {
+        if llama_context.has_memory() {
+            Self::Tokens(llama_context.n_ctx_seq() as usize)
+        } else {
+            Self::Unlimited
+        }
+    }
+}
diff --git a/paddler_agent_embeddings/src/unembeddable_embedding_document.rs b/paddler_agent_embeddings/src/unembeddable_embedding_document.rs
new file mode 100644
index 000000000..1267b06fd
--- /dev/null
+++ b/paddler_agent_embeddings/src/unembeddable_embedding_document.rs
@@ -0,0 +1,26 @@
+use paddler_messaging::embedding_document_exceeding_sequence_context_details::EmbeddingDocumentExceedingSequenceContextDetails;
+use paddler_messaging::embedding_document_tokenized_to_nothing_details::EmbeddingDocumentTokenizedToNothingDetails;
+use paddler_messaging::embedding_result::EmbeddingResult;
+use paddler_messaging::oversized_embedding_document_details::OversizedEmbeddingDocumentDetails;
+
+#[derive(Debug)]
+pub enum UnembeddableEmbeddingDocument {
+    ExceedsBatchSize(OversizedEmbeddingDocumentDetails),
+    ExceedsSequenceContext(EmbeddingDocumentExceedingSequenceContextDetails),
+    TokenizedToNothing(EmbeddingDocumentTokenizedToNothingDetails),
+}
+
+impl UnembeddableEmbeddingDocument {
+    #[must_use]
+    pub fn into_embedding_result(self) -> EmbeddingResult {
+        match self {
+            Self::ExceedsBatchSize(details) => EmbeddingResult::DocumentExceedsBatchSize(details),
+            Self::ExceedsSequenceContext(details) => {
+                EmbeddingResult::DocumentExceedsSequenceContext(details)
+            }
+            Self::TokenizedToNothing(details) => {
+                EmbeddingResult::DocumentTokenizedToNothing(details)
+            }
+        }
+    }
+}
diff --git a/paddler_agent_pointer_head/Cargo.toml b/paddler_agent_pointer_head/Cargo.toml
new file mode 100644
index 000000000..f9f32f878
--- /dev/null
+++ b/paddler_agent_pointer_head/Cargo.toml
@@ -0,0 +1,28 @@
+[package]
+name = "paddler_agent_pointer_head"
+autotests = false
+authors.workspace = true
+description.workspace = true
+edition.workspace = true
+homepage.workspace = true
+license.workspace = true
+repository.workspace = true
+version.workspace = true
+
+[[test]]
+name = "integration"
+path = "tests/main.rs"
+
+[dependencies]
+icu_normalizer = { workspace = true }
+llama-cpp-bindings = { workspace = true }
+thiserror = { workspace = true }
+
+[features]
+default = []
+cuda = ["llama-cpp-bindings/cuda"]
+metal = ["llama-cpp-bindings/metal"]
+vulkan = ["llama-cpp-bindings/vulkan"]
+
+[lints]
+workspace = true
diff --git a/paddler_agent_pointer_head/src/lib.rs b/paddler_agent_pointer_head/src/lib.rs
new file mode 100644
index 000000000..c89686829
--- /dev/null
+++ b/paddler_agent_pointer_head/src/lib.rs
@@ -0,0 +1,5 @@
+pub mod pointer_head;
+pub mod pointer_head_delimiters;
+pub mod pointer_head_error;
+pub mod pointer_head_input_normalization;
+pub mod pointer_head_projection;
diff --git a/paddler_agent_pointer_head/src/pointer_head.rs b/paddler_agent_pointer_head/src/pointer_head.rs
new file mode 100644
index 000000000..602f20ffb
--- /dev/null
+++ b/paddler_agent_pointer_head/src/pointer_head.rs
@@ -0,0 +1,213 @@
+use std::path::Path;
+
+use llama_cpp_bindings::gguf_context::GgufContext;
+use llama_cpp_bindings::gguf_tensor_f32::GgufTensorF32;
+
+use crate::pointer_head_delimiters::PointerHeadDelimiters;
+use crate::pointer_head_error::PointerHeadError;
+use crate::pointer_head_input_normalization::PointerHeadInputNormalization;
+use crate::pointer_head_projection::PointerHeadProjection;
+
+const TEMPERATURE_KEY: &str = "pointer_head.temperature";
+const TENSORS: [&str; 4] = [
+    "pointer_head.key.bias",
+    "pointer_head.key.weight",
+    "pointer_head.query.bias",
+    "pointer_head.query.weight",
+];
+
+fn read_temperature(gguf_context: &GgufContext) -> Result<f32, PointerHeadError> {
+    let temperature = gguf_context
+        .find_key(TEMPERATURE_KEY)
+        .and_then(|key_id| gguf_context.val_f32(key_id))
+        .map_err(|source| PointerHeadError::KeyUnreadable {
+            key: TEMPERATURE_KEY.to_owned(),
+            source,
+        })?;
+
+    if temperature.is_finite() && temperature > 0.0 {
+        Ok(temperature)
+    } else {
+        Err(PointerHeadError::TemperatureOutOfRange { temperature })
+    }
+}
+
+fn softmax(logits: &[f32]) -> Vec<f32> {
+    let largest_logit = logits.iter().copied().fold(f32::NEG_INFINITY, f32::max);
+    let exponentials: Vec<f32> = logits
+        .iter()
+        .map(|logit| (logit - largest_logit).exp())
+        .collect();
+    let exponential_sum: f32 = exponentials.iter().sum();
+
+    exponentials
+        .iter()
+        .map(|exponential| exponential / exponential_sum)
+        .collect()
+}
+
+pub struct PointerHead {
+    pub delimiters: PointerHeadDelimiters,
+    pub input_normalization: PointerHeadInputNormalization,
+    pub key_projection: PointerHeadProjection,
+    pub logit_scale: f32,
+    pub query_projection: PointerHeadProjection,
+    pub temperature: f32,
+}
+
+impl PointerHead {
+    pub fn load(path: &Path) -> Result<Self, PointerHeadError> {
+        let gguf_context =
+            GgufContext::from_file(path).map_err(|source| PointerHeadError::FileUnreadable {
+                path: path.to_path_buf(),
+                source,
+            })?;
+        let temperature = read_temperature(&gguf_context)?;
+        let delimiters = PointerHeadDelimiters::load(&gguf_context)?;
+        let input_normalization = PointerHeadInputNormalization::load(&gguf_context)?;
+        let mut tensor_shapes = [[0; 4]; 4];
+        let mut tensor_values: [Vec<f32>; 4] = Default::default();
+
+        for ((tensor, tensor_shape), tensor_value) in TENSORS
+            .iter()
+            .zip(&mut tensor_shapes)
+            .zip(&mut tensor_values)
+        {
+            let GgufTensorF32 { shape, values } =
+                gguf_context.read_tensor_f32(tensor).map_err(|source| {
+                    PointerHeadError::TensorUnreadable {
+                        tensor: (*tensor).to_owned(),
+                        source,
+                    }
+                })?;
+
+            *tensor_shape = shape;
+            *tensor_value = values;
+        }
+
+        let [key_bias_shape, key_weight_shape, _, _] = tensor_shapes;
+        let head_dim = key_bias_shape[0];
+        let hidden_size = key_weight_shape[0];
+        let bias_shape = [head_dim, 1, 1, 1];
+        let weight_shape = [hidden_size, head_dim, 1, 1];
+
+        for ((tensor, actual), expected) in TENSORS.iter().zip(tensor_shapes).zip([
+            bias_shape,
+            weight_shape,
+            bias_shape,
+            weight_shape,
+        ]) {
+            if actual != expected {
+                return Err(PointerHeadError::TensorShapeMismatch {
+                    tensor: (*tensor).to_owned(),
+                    expected,
+                    actual,
+                });
+            }
+        }
+
+        let [key_bias, key_weight, query_bias, query_weight] = tensor_values;
+
+        let logit_scale =
+            f32::from(u16::try_from(head_dim).map_err(|source| {
+                PointerHeadError::HeadDimensionOutOfRange { head_dim, source }
+            })?)
+            .sqrt()
+            .recip();
+
+        Ok(Self {
+            delimiters,
+            input_normalization,
+            key_projection: PointerHeadProjection {
+                bias: key_bias,
+                hidden_size: hidden_size as usize,
+                weight: key_weight,
+            },
+            logit_scale,
+            query_projection: PointerHeadProjection {
+                bias: query_bias,
+                hidden_size: hidden_size as usize,
+                weight: query_weight,
+            },
+            temperature,
+        })
+    }
+
+    #[must_use]
+    pub const fn hidden_size(&self) -> usize {
+        self.key_projection.hidden_size
+    }
+
+    #[must_use]
+    pub fn option_probabilities(&self, output_hidden_states: &[Vec<f32>]) -> Vec<f32> {
+        let Some((decide_hidden_state, option_hidden_states)) = output_hidden_states.split_last()
+        else {
+            return Vec::new();
+        };
+        let query = self.query_projection.project(decide_hidden_state);
+
+        softmax(
+            &option_hidden_states
+                .iter()
+                .map(|option_hidden_state| {
+                    let key = self.key_projection.project(option_hidden_state);
+                    let alignment: f32 = key
+                        .iter()
+                        .zip(&query)
+                        .map(|(key_value, query_value)| key_value * query_value)
+                        .sum();
+
+                    alignment * self.logit_scale / self.temperature
+                })
+                .collect::<Vec<f32>>(),
+        )
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::PointerHead;
+    use crate::pointer_head_delimiters::PointerHeadDelimiters;
+    use crate::pointer_head_input_normalization::PointerHeadInputNormalization;
+    use crate::pointer_head_projection::PointerHeadProjection;
+
+    fn identity_projection() -> PointerHeadProjection {
+        PointerHeadProjection {
+            bias: vec![0.0, 0.0],
+            hidden_size: 2,
+            weight: vec![1.0, 0.0, 0.0, 1.0],
+        }
+    }
+
+    fn pointer_head() -> PointerHead {
+        PointerHead {
+            delimiters: PointerHeadDelimiters {
+                decide: "<decide>".to_owned(),
+                option_end: "</option>".to_owned(),
+                option_start: "<option>".to_owned(),
+                question: "<question>".to_owned(),
+                state: "<state>".to_owned(),
+            },
+            input_normalization: PointerHeadInputNormalization::Unnormalized,
+            key_projection: identity_projection(),
+            logit_scale: 0.5,
+            query_projection: identity_projection(),
+            temperature: 2.0,
+        }
+    }
+
+    #[test]
+    fn options_aligned_with_the_decision_get_the_most_probability() {
+        let probabilities =
+            pointer_head().option_probabilities(&[vec![1.0, 0.0], vec![0.0, 1.0], vec![4.0, 0.0]]);
+        let expected_aligned_probability = 1.0 / (1.0 + (-1.0_f32).exp());
+
+        assert!((probabilities[0] - expected_aligned_probability).abs() < f32::EPSILON);
+        assert!((probabilities[1] - (1.0 - expected_aligned_probability)).abs() < f32::EPSILON);
+    }
+
+    #[test]
+    fn a_question_without_outputs_has_no_probabilities() {
+        assert_eq!(pointer_head().option_probabilities(&[]), Vec::<f32>::new());
+    }
+}
diff --git a/paddler_agent_pointer_head/src/pointer_head_delimiters.rs b/paddler_agent_pointer_head/src/pointer_head_delimiters.rs
new file mode 100644
index 000000000..6eed94af1
--- /dev/null
+++ b/paddler_agent_pointer_head/src/pointer_head_delimiters.rs
@@ -0,0 +1,47 @@
+use llama_cpp_bindings::gguf_context::GgufContext;
+
+use crate::pointer_head_error::PointerHeadError;
+
+const DELIMITER_KEYS: [&str; 5] = [
+    "pointer_head.delimiter.decide",
+    "pointer_head.delimiter.option_end",
+    "pointer_head.delimiter.option_start",
+    "pointer_head.delimiter.question",
+    "pointer_head.delimiter.state",
+];
+
+#[derive(Clone, Debug, Eq, PartialEq)]
+pub struct PointerHeadDelimiters {
+    pub decide: String,
+    pub option_end: String,
+    pub option_start: String,
+    pub question: String,
+    pub state: String,
+}
+
+impl PointerHeadDelimiters {
+    pub fn load(gguf_context: &GgufContext) -> Result<Self, PointerHeadError> {
+        let mut delimiter_tokens: [String; 5] = Default::default();
+
+        for (delimiter_token, key) in delimiter_tokens.iter_mut().zip(DELIMITER_KEYS) {
+            gguf_context
+                .find_key(key)
+                .and_then(|key_id| gguf_context.val_str(key_id))
+                .map_err(|source| PointerHeadError::KeyUnreadable {
+                    key: key.to_owned(),
+                    source,
+                })?
+                .clone_into(delimiter_token);
+        }
+
+        let [decide, option_end, option_start, question, state] = delimiter_tokens;
+
+        Ok(Self {
+            decide,
+            option_end,
+            option_start,
+            question,
+            state,
+        })
+    }
+}
diff --git a/paddler_agent_pointer_head/src/pointer_head_error.rs b/paddler_agent_pointer_head/src/pointer_head_error.rs
new file mode 100644
index 000000000..cbd0a1161
--- /dev/null
+++ b/paddler_agent_pointer_head/src/pointer_head_error.rs
@@ -0,0 +1,50 @@
+use std::num::TryFromIntError;
+use std::path::PathBuf;
+
+use llama_cpp_bindings::gguf_context_error::GgufContextError;
+
+#[derive(Debug, thiserror::Error)]
+pub enum PointerHeadError {
+    #[error("unable to read the pointer head file {path}: {source}")]
+    FileUnreadable {
+        path: PathBuf,
+        #[source]
+        source: GgufContextError,
+    },
+
+    #[error("the pointer head dimension {head_dim} does not fit sixteen bits: {source}")]
+    HeadDimensionOutOfRange {
+        head_dim: i64,
+        #[source]
+        source: TryFromIntError,
+    },
+
+    #[error(
+        "the pointer head declares the input normalization {declared}, which is not NFC, NFD, NFKC, NFKD or None"
+    )]
+    InputNormalizationUnknown { declared: String },
+
+    #[error("unable to read the pointer head key {key}: {source}")]
+    KeyUnreadable {
+        key: String,
+        #[source]
+        source: GgufContextError,
+    },
+
+    #[error("the pointer head temperature {temperature} is not a finite positive number")]
+    TemperatureOutOfRange { temperature: f32 },
+
+    #[error("the pointer head tensor {tensor} has the shape {actual:?} instead of {expected:?}")]
+    TensorShapeMismatch {
+        tensor: String,
+        expected: [i64; 4],
+        actual: [i64; 4],
+    },
+
+    #[error("unable to read the pointer head tensor {tensor}: {source}")]
+    TensorUnreadable {
+        tensor: String,
+        #[source]
+        source: GgufContextError,
+    },
+}
diff --git a/paddler_agent_pointer_head/src/pointer_head_input_normalization.rs b/paddler_agent_pointer_head/src/pointer_head_input_normalization.rs
new file mode 100644
index 000000000..01f124381
--- /dev/null
+++ b/paddler_agent_pointer_head/src/pointer_head_input_normalization.rs
@@ -0,0 +1,116 @@
+use std::borrow::Cow;
+
+use icu_normalizer::ComposingNormalizerBorrowed;
+use icu_normalizer::DecomposingNormalizerBorrowed;
+use llama_cpp_bindings::gguf_context::GgufContext;
+
+use crate::pointer_head_error::PointerHeadError;
+
+const INPUT_NORMALIZATION_KEY: &str = "pointer_head.input_normalization";
+
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+pub enum PointerHeadInputNormalization {
+    Nfc,
+    Nfd,
+    Nfkc,
+    Nfkd,
+    Unnormalized,
+}
+
+impl PointerHeadInputNormalization {
+    pub fn load(gguf_context: &GgufContext) -> Result<Self, PointerHeadError> {
+        gguf_context
+            .find_key(INPUT_NORMALIZATION_KEY)
+            .and_then(|key_id| gguf_context.val_str(key_id))
+            .map_err(|source| PointerHeadError::KeyUnreadable {
+                key: INPUT_NORMALIZATION_KEY.to_owned(),
+                source,
+            })
+            .and_then(Self::try_from)
+    }
+
+    #[must_use]
+    pub fn normalize<'text>(&self, text: &'text str) -> Cow<'text, str> {
+        match self {
+            Self::Nfc => ComposingNormalizerBorrowed::new_nfc().normalize(text),
+            Self::Nfd => DecomposingNormalizerBorrowed::new_nfd().normalize(text),
+            Self::Nfkc => ComposingNormalizerBorrowed::new_nfkc().normalize(text),
+            Self::Nfkd => DecomposingNormalizerBorrowed::new_nfkd().normalize(text),
+            Self::Unnormalized => Cow::Borrowed(text),
+        }
+    }
+}
+
+impl TryFrom<&str> for PointerHeadInputNormalization {
+    type Error = PointerHeadError;
+
+    fn try_from(declared: &str) -> Result<Self, Self::Error> {
+        match declared {
+            "NFC" => Ok(Self::Nfc),
+            "NFD" => Ok(Self::Nfd),
+            "NFKC" => Ok(Self::Nfkc),
+            "NFKD" => Ok(Self::Nfkd),
+            "None" => Ok(Self::Unnormalized),
+            unknown => Err(PointerHeadError::InputNormalizationUnknown {
+                declared: unknown.to_owned(),
+            }),
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::PointerHeadInputNormalization;
+    use crate::pointer_head_error::PointerHeadError;
+
+    #[test]
+    fn every_declarable_form_is_read_by_its_name() {
+        assert_eq!(
+            ["NFC", "NFD", "NFKC", "NFKD", "None"]
+                .map(PointerHeadInputNormalization::try_from)
+                .map(Result::ok),
+            [
+                Some(PointerHeadInputNormalization::Nfc),
+                Some(PointerHeadInputNormalization::Nfd),
+                Some(PointerHeadInputNormalization::Nfkc),
+                Some(PointerHeadInputNormalization::Nfkd),
+                Some(PointerHeadInputNormalization::Unnormalized),
+            ]
+        );
+    }
+
+    #[test]
+    fn a_form_a_pointer_head_cannot_declare_is_refused() {
+        assert!(matches!(
+            PointerHeadInputNormalization::try_from("NFX"),
+            Err(PointerHeadError::InputNormalizationUnknown { declared }) if declared == "NFX"
+        ));
+    }
+
+    #[test]
+    fn each_declared_form_normalizes_text_the_way_it_names() {
+        let decomposed_e_acute_and_fi_ligature = "e\u{301}\u{fb01}";
+
+        assert_eq!(
+            [
+                PointerHeadInputNormalization::Nfc,
+                PointerHeadInputNormalization::Nfd,
+                PointerHeadInputNormalization::Nfkc,
+                PointerHeadInputNormalization::Nfkd,
+                PointerHeadInputNormalization::Unnormalized,
+            ]
+            .map(|input_normalization| {
+                input_normalization
+                    .normalize(decomposed_e_acute_and_fi_ligature)
+                    .into_owned()
+            }),
+            [
+                "\u{e9}\u{fb01}".to_owned(),
+                "e\u{301}\u{fb01}".to_owned(),
+                "\u{e9}fi".to_owned(),
+                "e\u{301}fi".to_owned(),
+                "e\u{301}\u{fb01}".to_owned(),
+            ]
+        );
+    }
+}
diff --git a/paddler_agent_pointer_head/src/pointer_head_projection.rs b/paddler_agent_pointer_head/src/pointer_head_projection.rs
new file mode 100644
index 000000000..0394d8797
--- /dev/null
+++ b/paddler_agent_pointer_head/src/pointer_head_projection.rs
@@ -0,0 +1,39 @@
+pub struct PointerHeadProjection {
+    pub bias: Vec<f32>,
+    pub hidden_size: usize,
+    pub weight: Vec<f32>,
+}
+
+impl PointerHeadProjection {
+    #[must_use]
+    pub fn project(&self, hidden_state: &[f32]) -> Vec<f32> {
+        self.weight
+            .chunks_exact(self.hidden_size)
+            .zip(&self.bias)
+            .map(|(weight_row, bias)| {
+                weight_row
+                    .iter()
+                    .zip(hidden_state)
+                    .map(|(weight, hidden_value)| weight * hidden_value)
+                    .sum::<f32>()
+                    + bias
+            })
+            .collect()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::PointerHeadProjection;
+
+    #[test]
+    fn projects_a_hidden_state_through_every_weight_row_and_adds_its_bias() {
+        let projection = PointerHeadProjection {
+            bias: vec![0.5, -1.0],
+            hidden_size: 3,
+            weight: vec![1.0, 2.0, 3.0, -1.0, 0.0, 1.0],
+        };
+
+        assert_eq!(projection.project(&[1.0, 1.0, 2.0]), vec![9.5, 0.0]);
+    }
+}
diff --git a/paddler_agent_pointer_head/tests/a_pointer_head_whose_head_dimension_exceeds_sixteen_bits_is_refused.rs b/paddler_agent_pointer_head/tests/a_pointer_head_whose_head_dimension_exceeds_sixteen_bits_is_refused.rs
new file mode 100644
index 000000000..3f2d3a964
--- /dev/null
+++ b/paddler_agent_pointer_head/tests/a_pointer_head_whose_head_dimension_exceeds_sixteen_bits_is_refused.rs
@@ -0,0 +1,24 @@
+use paddler_agent_pointer_head::pointer_head::PointerHead;
+use paddler_agent_pointer_head::pointer_head_error::PointerHeadError;
+
+use crate::fixture_path::fixture_path;
+
+const OVERSIZED_HEAD_DIM: i64 = 1 << 16;
+
+#[test]
+fn a_pointer_head_whose_head_dimension_exceeds_sixteen_bits_is_refused() {
+    assert_eq!(
+        PointerHead::load(&fixture_path("pointer_head_with_oversized_head_dim.gguf"))
+            .err()
+            .as_ref()
+            .map(ToString::to_string),
+        Some(
+            PointerHeadError::HeadDimensionOutOfRange {
+                head_dim: OVERSIZED_HEAD_DIM,
+                source: u16::try_from(OVERSIZED_HEAD_DIM)
+                    .expect_err("the head dimension must not fit sixteen bits"),
+            }
+            .to_string()
+        )
+    );
+}
diff --git a/paddler_agent_pointer_head/tests/a_pointer_head_with_f16_tensors_is_refused.rs b/paddler_agent_pointer_head/tests/a_pointer_head_with_f16_tensors_is_refused.rs
new file mode 100644
index 000000000..12b82074f
--- /dev/null
+++ b/paddler_agent_pointer_head/tests/a_pointer_head_with_f16_tensors_is_refused.rs
@@ -0,0 +1,40 @@
+use std::error::Error;
+
+use llama_cpp_bindings::context::kv_cache_type::KvCacheType;
+use llama_cpp_bindings::gguf_context_error::GgufContextError;
+
+use paddler_agent_pointer_head::pointer_head::PointerHead;
+use paddler_agent_pointer_head::pointer_head_error::PointerHeadError;
+
+use crate::fixture_path::fixture_path;
+
+fn key_bias_is_not_f32() -> GgufContextError {
+    GgufContextError::TensorIsNotF32 {
+        name: "pointer_head.key.bias".to_owned(),
+        ggml_type: KvCacheType::F16.into(),
+    }
+}
+
+#[test]
+fn a_pointer_head_with_f16_tensors_is_refused() {
+    let load_error = PointerHead::load(&fixture_path("pointer_head_with_f16_tensors.gguf")).err();
+
+    assert_eq!(
+        load_error.as_ref().map(ToString::to_string),
+        Some(
+            PointerHeadError::TensorUnreadable {
+                tensor: "pointer_head.key.bias".to_owned(),
+                source: key_bias_is_not_f32(),
+            }
+            .to_string()
+        )
+    );
+    assert_eq!(
+        load_error
+            .as_ref()
+            .and_then(Error::source)
+            .and_then(|source| source.downcast_ref::<GgufContextError>())
+            .map(ToString::to_string),
+        Some(key_bias_is_not_f32().to_string())
+    );
+}
diff --git a/paddler_agent_pointer_head/tests/a_pointer_head_with_mismatched_projections_is_refused.rs b/paddler_agent_pointer_head/tests/a_pointer_head_with_mismatched_projections_is_refused.rs
new file mode 100644
index 000000000..45a58404b
--- /dev/null
+++ b/paddler_agent_pointer_head/tests/a_pointer_head_with_mismatched_projections_is_refused.rs
@@ -0,0 +1,24 @@
+use paddler_agent_pointer_head::pointer_head::PointerHead;
+use paddler_agent_pointer_head::pointer_head_error::PointerHeadError;
+
+use crate::fixture_path::fixture_path;
+
+#[test]
+fn a_pointer_head_with_mismatched_projections_is_refused() {
+    assert_eq!(
+        PointerHead::load(&fixture_path(
+            "pointer_head_with_mismatched_projections.gguf"
+        ))
+        .err()
+        .as_ref()
+        .map(ToString::to_string),
+        Some(
+            PointerHeadError::TensorShapeMismatch {
+                tensor: "pointer_head.query.bias".to_owned(),
+                expected: [16, 1, 1, 1],
+                actual: [15, 1, 1, 1],
+            }
+            .to_string()
+        )
+    );
+}
diff --git a/paddler_agent_pointer_head/tests/a_pointer_head_with_zero_temperature_is_refused.rs b/paddler_agent_pointer_head/tests/a_pointer_head_with_zero_temperature_is_refused.rs
new file mode 100644
index 000000000..ebe218c08
--- /dev/null
+++ b/paddler_agent_pointer_head/tests/a_pointer_head_with_zero_temperature_is_refused.rs
@@ -0,0 +1,15 @@
+use paddler_agent_pointer_head::pointer_head::PointerHead;
+use paddler_agent_pointer_head::pointer_head_error::PointerHeadError;
+
+use crate::fixture_path::fixture_path;
+
+#[test]
+fn a_pointer_head_with_zero_temperature_is_refused() {
+    assert_eq!(
+        PointerHead::load(&fixture_path("pointer_head_with_zero_temperature.gguf"))
+            .err()
+            .as_ref()
+            .map(ToString::to_string),
+        Some(PointerHeadError::TemperatureOutOfRange { temperature: 0.0 }.to_string())
+    );
+}
diff --git a/paddler_agent_pointer_head/tests/a_pointer_head_without_its_delimiters_is_refused.rs b/paddler_agent_pointer_head/tests/a_pointer_head_without_its_delimiters_is_refused.rs
new file mode 100644
index 000000000..4a31f4257
--- /dev/null
+++ b/paddler_agent_pointer_head/tests/a_pointer_head_without_its_delimiters_is_refused.rs
@@ -0,0 +1,38 @@
+use std::error::Error;
+
+use llama_cpp_bindings::gguf_context_error::GgufContextError;
+
+use paddler_agent_pointer_head::pointer_head::PointerHead;
+use paddler_agent_pointer_head::pointer_head_error::PointerHeadError;
+
+use crate::fixture_path::fixture_path;
+
+fn decide_delimiter_not_found() -> GgufContextError {
+    GgufContextError::KeyNotFound {
+        key: "pointer_head.delimiter.decide".to_owned(),
+    }
+}
+
+#[test]
+fn a_pointer_head_without_its_delimiters_is_refused() {
+    let load_error = PointerHead::load(&fixture_path("pointer_head_without_delimiters.gguf")).err();
+
+    assert_eq!(
+        load_error.as_ref().map(ToString::to_string),
+        Some(
+            PointerHeadError::KeyUnreadable {
+                key: "pointer_head.delimiter.decide".to_owned(),
+                source: decide_delimiter_not_found(),
+            }
+            .to_string()
+        )
+    );
+    assert_eq!(
+        load_error
+            .as_ref()
+            .and_then(Error::source)
+            .and_then(|source| source.downcast_ref::<GgufContextError>())
+            .map(ToString::to_string),
+        Some(decide_delimiter_not_found().to_string())
+    );
+}
diff --git a/paddler_agent_pointer_head/tests/a_pointer_head_without_its_input_normalization_is_refused.rs b/paddler_agent_pointer_head/tests/a_pointer_head_without_its_input_normalization_is_refused.rs
new file mode 100644
index 000000000..41f440cc1
--- /dev/null
+++ b/paddler_agent_pointer_head/tests/a_pointer_head_without_its_input_normalization_is_refused.rs
@@ -0,0 +1,41 @@
+use std::error::Error;
+
+use llama_cpp_bindings::gguf_context_error::GgufContextError;
+
+use paddler_agent_pointer_head::pointer_head::PointerHead;
+use paddler_agent_pointer_head::pointer_head_error::PointerHeadError;
+
+use crate::fixture_path::fixture_path;
+
+fn input_normalization_not_found() -> GgufContextError {
+    GgufContextError::KeyNotFound {
+        key: "pointer_head.input_normalization".to_owned(),
+    }
+}
+
+#[test]
+fn a_pointer_head_without_its_input_normalization_is_refused() {
+    let load_error = PointerHead::load(&fixture_path(
+        "pointer_head_without_input_normalization.gguf",
+    ))
+    .err();
+
+    assert_eq!(
+        load_error.as_ref().map(ToString::to_string),
+        Some(
+            PointerHeadError::KeyUnreadable {
+                key: "pointer_head.input_normalization".to_owned(),
+                source: input_normalization_not_found(),
+            }
+            .to_string()
+        )
+    );
+    assert_eq!(
+        load_error
+            .as_ref()
+            .and_then(Error::source)
+            .and_then(|source| source.downcast_ref::<GgufContextError>())
+            .map(ToString::to_string),
+        Some(input_normalization_not_found().to_string())
+    );
+}
diff --git a/paddler_agent_pointer_head/tests/a_pointer_head_without_its_temperature_is_refused.rs b/paddler_agent_pointer_head/tests/a_pointer_head_without_its_temperature_is_refused.rs
new file mode 100644
index 000000000..775a6fa27
--- /dev/null
+++ b/paddler_agent_pointer_head/tests/a_pointer_head_without_its_temperature_is_refused.rs
@@ -0,0 +1,39 @@
+use std::error::Error;
+
+use llama_cpp_bindings::gguf_context_error::GgufContextError;
+
+use paddler_agent_pointer_head::pointer_head::PointerHead;
+use paddler_agent_pointer_head::pointer_head_error::PointerHeadError;
+
+use crate::fixture_path::fixture_path;
+
+fn temperature_not_found() -> GgufContextError {
+    GgufContextError::KeyNotFound {
+        key: "pointer_head.temperature".to_owned(),
+    }
+}
+
+#[test]
+fn a_pointer_head_without_its_temperature_is_refused() {
+    let load_error =
+        PointerHead::load(&fixture_path("pointer_head_without_temperature.gguf")).err();
+
+    assert_eq!(
+        load_error.as_ref().map(ToString::to_string),
+        Some(
+            PointerHeadError::KeyUnreadable {
+                key: "pointer_head.temperature".to_owned(),
+                source: temperature_not_found(),
+            }
+            .to_string()
+        )
+    );
+    assert_eq!(
+        load_error
+            .as_ref()
+            .and_then(Error::source)
+            .and_then(|source| source.downcast_ref::<GgufContextError>())
+            .map(ToString::to_string),
+        Some(temperature_not_found().to_string())
+    );
+}
diff --git a/paddler_agent_pointer_head/tests/an_unreadable_pointer_head_file_is_refused.rs b/paddler_agent_pointer_head/tests/an_unreadable_pointer_head_file_is_refused.rs
new file mode 100644
index 000000000..7d0574f0e
--- /dev/null
+++ b/paddler_agent_pointer_head/tests/an_unreadable_pointer_head_file_is_refused.rs
@@ -0,0 +1,34 @@
+use std::error::Error;
+
+use llama_cpp_bindings::gguf_context_error::GgufContextError;
+
+use paddler_agent_pointer_head::pointer_head::PointerHead;
+use paddler_agent_pointer_head::pointer_head_error::PointerHeadError;
+
+use crate::fixture_path::fixture_path;
+
+#[test]
+fn an_unreadable_pointer_head_file_is_refused() {
+    let invalid_file = fixture_path("invalid.gguf");
+
+    let load_error = PointerHead::load(&invalid_file).err();
+
+    assert_eq!(
+        load_error.as_ref().map(ToString::to_string),
+        Some(
+            PointerHeadError::FileUnreadable {
+                path: invalid_file.clone(),
+                source: GgufContextError::InitFailed(invalid_file.clone()),
+            }
+            .to_string()
+        )
+    );
+    assert_eq!(
+        load_error
+            .as_ref()
+            .and_then(Error::source)
+            .and_then(|source| source.downcast_ref::<GgufContextError>())
+            .map(ToString::to_string),
+        Some(GgufContextError::InitFailed(invalid_file).to_string())
+    );
+}
diff --git a/paddler_agent_pointer_head/tests/fixture_path.rs b/paddler_agent_pointer_head/tests/fixture_path.rs
new file mode 100644
index 000000000..1793cf55b
--- /dev/null
+++ b/paddler_agent_pointer_head/tests/fixture_path.rs
@@ -0,0 +1,8 @@
+use std::path::PathBuf;
+
+pub fn fixture_path(fixture_name: &str) -> PathBuf {
+    PathBuf::from(format!(
+        "{}/../fixtures/{fixture_name}",
+        env!("CARGO_MANIFEST_DIR")
+    ))
+}
diff --git a/paddler_agent_pointer_head/tests/loading_the_synthetic_pointer_head_reads_its_dimensions_scale_temperature_delimiters_and_input_normalization.rs b/paddler_agent_pointer_head/tests/loading_the_synthetic_pointer_head_reads_its_dimensions_scale_temperature_delimiters_and_input_normalization.rs
new file mode 100644
index 000000000..c024479a3
--- /dev/null
+++ b/paddler_agent_pointer_head/tests/loading_the_synthetic_pointer_head_reads_its_dimensions_scale_temperature_delimiters_and_input_normalization.rs
@@ -0,0 +1,32 @@
+use paddler_agent_pointer_head::pointer_head::PointerHead;
+use paddler_agent_pointer_head::pointer_head_delimiters::PointerHeadDelimiters;
+use paddler_agent_pointer_head::pointer_head_input_normalization::PointerHeadInputNormalization;
+
+use crate::fixture_path::fixture_path;
+
+#[test]
+fn loading_the_synthetic_pointer_head_reads_its_dimensions_scale_temperature_delimiters_and_input_normalization()
+ {
+    let pointer_head = PointerHead::load(&fixture_path("qwen3_5_0_8b_synthetic_pointer_head.gguf"))
+        .expect("the synthetic pointer head must load");
+
+    assert_eq!(pointer_head.hidden_size(), 1024);
+    assert_eq!(pointer_head.key_projection.bias.len(), 16);
+    assert_eq!(pointer_head.query_projection.weight.len(), 16 * 1024);
+    assert!((pointer_head.logit_scale - 0.25).abs() < f32::EPSILON);
+    assert!((pointer_head.temperature - 1.5).abs() < f32::EPSILON);
+    assert_eq!(
+        pointer_head.delimiters,
+        PointerHeadDelimiters {
+            decide: "<|fim_suffix|>".to_owned(),
+            option_end: "<|box_end|>".to_owned(),
+            option_start: "<|box_start|>".to_owned(),
+            question: "<|fim_middle|>".to_owned(),
+            state: "<|fim_prefix|>".to_owned(),
+        }
+    );
+    assert_eq!(
+        pointer_head.input_normalization,
+        PointerHeadInputNormalization::Nfc
+    );
+}
diff --git a/paddler_agent_pointer_head/tests/main.rs b/paddler_agent_pointer_head/tests/main.rs
new file mode 100644
index 000000000..1ade46fef
--- /dev/null
+++ b/paddler_agent_pointer_head/tests/main.rs
@@ -0,0 +1,10 @@
+mod a_pointer_head_whose_head_dimension_exceeds_sixteen_bits_is_refused;
+mod a_pointer_head_with_f16_tensors_is_refused;
+mod a_pointer_head_with_mismatched_projections_is_refused;
+mod a_pointer_head_with_zero_temperature_is_refused;
+mod a_pointer_head_without_its_delimiters_is_refused;
+mod a_pointer_head_without_its_input_normalization_is_refused;
+mod a_pointer_head_without_its_temperature_is_refused;
+mod an_unreadable_pointer_head_file_is_refused;
+mod fixture_path;
+mod loading_the_synthetic_pointer_head_reads_its_dimensions_scale_temperature_delimiters_and_input_normalization;
diff --git a/paddler_agent_runtime/Cargo.toml b/paddler_agent_runtime/Cargo.toml
new file mode 100644
index 000000000..18c386632
--- /dev/null
+++ b/paddler_agent_runtime/Cargo.toml
@@ -0,0 +1,36 @@
+[package]
+name = "paddler_agent_runtime"
+autotests = false
+authors.workspace = true
+description.workspace = true
+edition.workspace = true
+homepage.workspace = true
+license.workspace = true
+repository.workspace = true
+version.workspace = true
+
+[[test]]
+name = "integration"
+path = "tests/main.rs"
+
+[dependencies]
+llama-cpp-bindings = { workspace = true }
+llama-cpp-bindings-sys = { workspace = true }
+log = { workspace = true }
+parking_lot = { workspace = true }
+thiserror = { workspace = true }
+tokio = { workspace = true }
+tokio-util = { workspace = true }
+paddler_agent_status = { workspace = true }
+paddler_inference_parameters = { workspace = true }
+paddler_messaging = { workspace = true }
+paddler_spawn_resuming_panic = { workspace = true }
+
+[features]
+default = []
+cuda = ["llama-cpp-bindings/cuda"]
+metal = ["llama-cpp-bindings/metal"]
+vulkan = ["llama-cpp-bindings/vulkan"]
+
+[lints]
+workspace = true
diff --git a/paddler_agent/src/agent_kv_cache_dtype.rs b/paddler_agent_runtime/src/agent_kv_cache_dtype.rs
similarity index 100%
rename from paddler_agent/src/agent_kv_cache_dtype.rs
rename to paddler_agent_runtime/src/agent_kv_cache_dtype.rs
diff --git a/paddler_agent/src/agent_request.rs b/paddler_agent_runtime/src/agent_request.rs
similarity index 82%
rename from paddler_agent/src/agent_request.rs
rename to paddler_agent_runtime/src/agent_request.rs
index c1e807ff6..17cebb30d 100644
--- a/paddler_agent/src/agent_request.rs
+++ b/paddler_agent_runtime/src/agent_request.rs
@@ -1,6 +1,6 @@
 use tokio::sync::mpsc;
 
-use crate::slot_guard::SlotGuard;
+use paddler_agent_status::slot_guard::SlotGuard;
 
 pub struct AgentRequest<TParams, TResponse> {
     pub params: TParams,
diff --git a/paddler_agent_runtime/src/agent_runtime_error.rs b/paddler_agent_runtime/src/agent_runtime_error.rs
new file mode 100644
index 000000000..9ee7c7750
--- /dev/null
+++ b/paddler_agent_runtime/src/agent_runtime_error.rs
@@ -0,0 +1,54 @@
+use std::io;
+use std::num::TryFromIntError;
+use std::path::PathBuf;
+
+use llama_cpp_bindings::LlamaCppError;
+use llama_cpp_bindings::batch_add_error::BatchAddError;
+use llama_cpp_bindings::error::DecodeError;
+use llama_cpp_bindings::error::LlamaContextLoadError;
+use llama_cpp_bindings::error::LlamaModelLoadError;
+use llama_cpp_bindings::error::MetaValError;
+
+use paddler_messaging::inference_capability::InferenceCapability;
+
+#[derive(Debug, thiserror::Error)]
+pub enum AgentRuntimeError {
+    #[error("unable to determine how many threads this machine runs in parallel: {0}")]
+    AvailableParallelismUnknown(#[source] io::Error),
+
+    #[error("unable to initialize the llama.cpp backend: {0}")]
+    BackendInitializationFailed(#[source] LlamaCppError),
+
+    #[error("unable to create the llama.cpp context: {0}")]
+    ContextCreationFailed(#[source] LlamaContextLoadError),
+
+    #[error("a {inference_capability:?} context is already resident in VRAM")]
+    LlamaContextAlreadyResident {
+        inference_capability: InferenceCapability,
+    },
+
+    #[error("unable to load the model from {model_path}: {source}")]
+    ModelLoadFailed {
+        model_path: PathBuf,
+        #[source]
+        source: LlamaModelLoadError,
+    },
+
+    #[error("unable to read the metadata of the loaded model: {0}")]
+    ModelMetadataUnreadable(#[source] MetaValError),
+
+    #[error("model weights are already resident in VRAM")]
+    ModelWeightsAlreadyResident,
+
+    #[error("a multimodal projection is already resident in VRAM")]
+    MultimodalProjectionAlreadyResident,
+
+    #[error("the machine runs more threads in parallel than llama.cpp can address: {0}")]
+    ThreadCountOutOfRange(#[source] TryFromIntError),
+
+    #[error("unable to assemble the warm-up batch: {0}")]
+    WarmUpBatchAssemblyFailed(#[source] BatchAddError),
+
+    #[error("llama.cpp rejected the warm-up batch: {0}")]
+    WarmUpDecodeFailed(#[source] DecodeError),
+}
diff --git a/paddler_agent_runtime/src/agent_vram_residencies.rs b/paddler_agent_runtime/src/agent_vram_residencies.rs
new file mode 100644
index 000000000..4d4c19053
--- /dev/null
+++ b/paddler_agent_runtime/src/agent_vram_residencies.rs
@@ -0,0 +1,80 @@
+use paddler_messaging::inference_capability::InferenceCapability;
+
+use crate::vram_residency::VramResidency;
+use crate::vram_resident::VramResident;
+
+#[derive(Clone)]
+pub struct AgentVramResidencies {
+    pub decisions_context: VramResidency,
+    pub embeddings_context: VramResidency,
+    pub model_weights: VramResidency,
+    pub multimodal_projection: VramResidency,
+    pub text_generation_context: VramResidency,
+}
+
+impl AgentVramResidencies {
+    #[must_use]
+    pub const fn llama_context(&self, inference_capability: InferenceCapability) -> &VramResidency {
+        match inference_capability {
+            InferenceCapability::Decisions => &self.decisions_context,
+            InferenceCapability::Embeddings => &self.embeddings_context,
+            InferenceCapability::TextGeneration => &self.text_generation_context,
+        }
+    }
+}
+
+impl Default for AgentVramResidencies {
+    fn default() -> Self {
+        Self {
+            decisions_context: VramResidency::vacant(VramResident::LlamaContext(
+                InferenceCapability::Decisions,
+            )),
+            embeddings_context: VramResidency::vacant(VramResident::LlamaContext(
+                InferenceCapability::Embeddings,
+            )),
+            model_weights: VramResidency::vacant(VramResident::ModelWeights),
+            multimodal_projection: VramResidency::vacant(VramResident::MultimodalProjection),
+            text_generation_context: VramResidency::vacant(VramResident::LlamaContext(
+                InferenceCapability::TextGeneration,
+            )),
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use paddler_messaging::inference_capability::InferenceCapability;
+
+    use super::AgentVramResidencies;
+    use crate::agent_runtime_error::AgentRuntimeError;
+
+    #[test]
+    fn every_capability_occupies_a_context_residency_of_its_own() {
+        let inference_capabilities = [
+            InferenceCapability::Decisions,
+            InferenceCapability::Embeddings,
+            InferenceCapability::TextGeneration,
+        ];
+        let agent_vram_residencies = AgentVramResidencies::default();
+        let _context_occupancies = inference_capabilities.map(|inference_capability| {
+            agent_vram_residencies
+                .llama_context(inference_capability)
+                .try_occupy()
+                .unwrap()
+        });
+
+        assert_eq!(
+            inference_capabilities.map(|inference_capability| agent_vram_residencies
+                .llama_context(inference_capability)
+                .try_occupy()
+                .err()
+                .map(|error| error.to_string())),
+            inference_capabilities.map(|inference_capability| Some(
+                AgentRuntimeError::LlamaContextAlreadyResident {
+                    inference_capability
+                }
+                .to_string()
+            ))
+        );
+    }
+}
diff --git a/paddler_agent_runtime/src/await_scheduler_startup.rs b/paddler_agent_runtime/src/await_scheduler_startup.rs
new file mode 100644
index 000000000..f76a885b7
--- /dev/null
+++ b/paddler_agent_runtime/src/await_scheduler_startup.rs
@@ -0,0 +1,48 @@
+use std::thread::JoinHandle;
+
+use tokio::select;
+use tokio::sync::mpsc::UnboundedReceiver;
+use tokio_util::sync::CancellationToken;
+
+use crate::join_scheduler_thread_in_background::join_scheduler_thread_in_background;
+use crate::prepares_scheduler_command::PreparesSchedulerCommand;
+use crate::scheduler_handle::SchedulerHandle;
+use crate::scheduler_ready::SchedulerReady;
+use crate::startup_outcome::StartupOutcome;
+
+pub async fn await_scheduler_startup<TPreparation, TError>(
+    cancellation_token: &CancellationToken,
+    mut scheduler_startup_rx: UnboundedReceiver<Result<SchedulerReady<TPreparation>, TError>>,
+    scheduler_thread_handle: JoinHandle<()>,
+) -> Result<StartupOutcome<SchedulerHandle<TPreparation>>, TError>
+where
+    TPreparation: PreparesSchedulerCommand,
+    TError: Send + 'static,
+{
+    let scheduler_startup = select! {
+        biased;
+        () = cancellation_token.cancelled() => None,
+        scheduler_startup = scheduler_startup_rx.recv() => scheduler_startup,
+    };
+
+    match scheduler_startup {
+        Some(Ok(SchedulerReady { request_preparer })) => {
+            Ok(StartupOutcome::Ready(SchedulerHandle {
+                request_preparer,
+                scheduler_thread_handle,
+            }))
+        }
+        Some(Err(startup_error)) => {
+            join_scheduler_thread_in_background(scheduler_startup_rx, scheduler_thread_handle)
+                .await;
+
+            Err(startup_error)
+        }
+        None => {
+            join_scheduler_thread_in_background(scheduler_startup_rx, scheduler_thread_handle)
+                .await;
+
+            Ok(StartupOutcome::Cancelled)
+        }
+    }
+}
diff --git a/paddler_agent_runtime/src/continue_startup_unless_shutting_down.rs b/paddler_agent_runtime/src/continue_startup_unless_shutting_down.rs
new file mode 100644
index 000000000..d9e026900
--- /dev/null
+++ b/paddler_agent_runtime/src/continue_startup_unless_shutting_down.rs
@@ -0,0 +1,17 @@
+use log::debug;
+use tokio_util::sync::CancellationToken;
+
+pub fn continue_startup_unless_shutting_down<TError>(
+    agent_shutdown: &CancellationToken,
+    remaining_startup: impl FnOnce() -> Result<(), TError>,
+) -> Result<(), TError> {
+    if agent_shutdown.is_cancelled() {
+        debug!(
+            "Shutdown was requested during the scheduler startup; skipping its remaining stages"
+        );
+
+        return Ok(());
+    }
+
+    remaining_startup()
+}
diff --git a/paddler_agent/src/converts_to_llama_kv_cache_dtype.rs b/paddler_agent_runtime/src/converts_to_llama_kv_cache_dtype.rs
similarity index 100%
rename from paddler_agent/src/converts_to_llama_kv_cache_dtype.rs
rename to paddler_agent_runtime/src/converts_to_llama_kv_cache_dtype.rs
diff --git a/paddler_agent_runtime/src/inference_runtime_context.rs b/paddler_agent_runtime/src/inference_runtime_context.rs
new file mode 100644
index 000000000..8f63241f2
--- /dev/null
+++ b/paddler_agent_runtime/src/inference_runtime_context.rs
@@ -0,0 +1,36 @@
+use std::sync::Arc;
+
+use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
+
+use crate::agent_vram_residencies::AgentVramResidencies;
+use crate::inference_runtime_context_params::InferenceRuntimeContextParams;
+use crate::model_metadata_holder::ModelMetadataHolder;
+use crate::reports_available_parallelism::ReportsAvailableParallelism;
+use crate::shared_llama_backend::SharedLlamaBackend;
+use crate::system_parallelism::SystemParallelism;
+
+#[derive(Clone)]
+pub struct InferenceRuntimeContext {
+    pub agent_name: Option<String>,
+    pub available_parallelism: Arc<dyn ReportsAvailableParallelism>,
+    pub llama_backend: Arc<SharedLlamaBackend>,
+    pub model_metadata_holder: Arc<ModelMetadataHolder>,
+    pub slot_aggregated_status: Arc<SlotAggregatedStatus>,
+    pub vram_residencies: AgentVramResidencies,
+}
+
+impl InferenceRuntimeContext {
+    #[must_use]
+    pub fn new(
+        InferenceRuntimeContextParams { agent_name, slots }: InferenceRuntimeContextParams,
+    ) -> Self {
+        Self {
+            agent_name,
+            available_parallelism: Arc::new(SystemParallelism),
+            llama_backend: Arc::new(SharedLlamaBackend::default()),
+            model_metadata_holder: Arc::new(ModelMetadataHolder::default()),
+            slot_aggregated_status: Arc::new(SlotAggregatedStatus::new(slots)),
+            vram_residencies: AgentVramResidencies::default(),
+        }
+    }
+}
diff --git a/paddler_agent_runtime/src/inference_runtime_context_params.rs b/paddler_agent_runtime/src/inference_runtime_context_params.rs
new file mode 100644
index 000000000..cfcf24921
--- /dev/null
+++ b/paddler_agent_runtime/src/inference_runtime_context_params.rs
@@ -0,0 +1,4 @@
+pub struct InferenceRuntimeContextParams {
+    pub agent_name: Option<String>,
+    pub slots: u16,
+}
diff --git a/paddler_agent_runtime/src/inference_thread_count.rs b/paddler_agent_runtime/src/inference_thread_count.rs
new file mode 100644
index 000000000..aa86ace75
--- /dev/null
+++ b/paddler_agent_runtime/src/inference_thread_count.rs
@@ -0,0 +1,58 @@
+use std::cmp::max;
+use std::num::NonZeroUsize;
+
+use crate::agent_runtime_error::AgentRuntimeError;
+use crate::reports_available_parallelism::ReportsAvailableParallelism;
+
+const LOGICAL_CORES_PER_PHYSICAL_CORE: i32 = 2;
+const MINIMUM_THREAD_COUNT: i32 = 2;
+
+fn thread_count_for_logical_cores(logical_cores: NonZeroUsize) -> Result<i32, AgentRuntimeError> {
+    let logical_cores =
+        i32::try_from(logical_cores.get()).map_err(AgentRuntimeError::ThreadCountOutOfRange)?;
+
+    Ok(max(
+        MINIMUM_THREAD_COUNT,
+        logical_cores / LOGICAL_CORES_PER_PHYSICAL_CORE,
+    ))
+}
+
+pub fn inference_thread_count(
+    parallelism: &dyn ReportsAvailableParallelism,
+) -> Result<i32, AgentRuntimeError> {
+    thread_count_for_logical_cores(
+        parallelism
+            .available_parallelism()
+            .map_err(AgentRuntimeError::AvailableParallelismUnknown)?,
+    )
+}
+
+#[cfg(test)]
+mod tests {
+    use std::num::NonZeroUsize;
+
+    use super::MINIMUM_THREAD_COUNT;
+    use super::thread_count_for_logical_cores;
+
+    #[test]
+    fn never_runs_fewer_than_the_minimum_threads() {
+        assert_eq!(
+            thread_count_for_logical_cores(NonZeroUsize::MIN).unwrap(),
+            MINIMUM_THREAD_COUNT
+        );
+    }
+
+    #[test]
+    fn reports_a_parallelism_llama_cpp_cannot_address() {
+        assert_eq!(
+            thread_count_for_logical_cores(NonZeroUsize::MAX)
+                .err()
+                .as_ref()
+                .map(ToString::to_string),
+            Some(
+                "the machine runs more threads in parallel than llama.cpp can address: number too large to fit in target type"
+                    .to_owned()
+            )
+        );
+    }
+}
diff --git a/paddler_agent/src/join_scheduler_thread.rs b/paddler_agent_runtime/src/join_scheduler_thread.rs
similarity index 80%
rename from paddler_agent/src/join_scheduler_thread.rs
rename to paddler_agent_runtime/src/join_scheduler_thread.rs
index 96496ab41..c27a7b1c0 100644
--- a/paddler_agent/src/join_scheduler_thread.rs
+++ b/paddler_agent_runtime/src/join_scheduler_thread.rs
@@ -1,10 +1,8 @@
 use std::panic::resume_unwind;
 use std::thread::JoinHandle;
 
-use anyhow::Result;
-
-pub fn join_scheduler_thread(scheduler_thread_handle: JoinHandle<Result<()>>) -> Result<()> {
+pub fn join_scheduler_thread(scheduler_thread_handle: JoinHandle<()>) {
     scheduler_thread_handle
         .join()
-        .unwrap_or_else(|panic_payload| resume_unwind(panic_payload))
+        .unwrap_or_else(|panic_payload| resume_unwind(panic_payload));
 }
diff --git a/paddler_agent_runtime/src/join_scheduler_thread_in_background.rs b/paddler_agent_runtime/src/join_scheduler_thread_in_background.rs
new file mode 100644
index 000000000..bde422609
--- /dev/null
+++ b/paddler_agent_runtime/src/join_scheduler_thread_in_background.rs
@@ -0,0 +1,18 @@
+use std::thread::JoinHandle;
+
+use paddler_spawn_resuming_panic::spawn_blocking_resuming_panic::spawn_blocking_resuming_panic;
+
+use crate::join_scheduler_thread::join_scheduler_thread;
+
+pub async fn join_scheduler_thread_in_background<TReleasedFirst>(
+    released_first: TReleasedFirst,
+    scheduler_thread_handle: JoinHandle<()>,
+) where
+    TReleasedFirst: Send + 'static,
+{
+    spawn_blocking_resuming_panic(move || {
+        drop(released_first);
+        join_scheduler_thread(scheduler_thread_handle);
+    })
+    .await;
+}
diff --git a/paddler_agent_runtime/src/lib.rs b/paddler_agent_runtime/src/lib.rs
new file mode 100644
index 000000000..66198d8ad
--- /dev/null
+++ b/paddler_agent_runtime/src/lib.rs
@@ -0,0 +1,45 @@
+pub mod agent_kv_cache_dtype;
+pub mod agent_request;
+pub mod agent_runtime_error;
+pub mod agent_vram_residencies;
+pub mod await_scheduler_startup;
+pub mod continue_startup_unless_shutting_down;
+pub mod converts_to_llama_kv_cache_dtype;
+pub mod inference_runtime_context;
+pub mod inference_runtime_context_params;
+pub mod inference_thread_count;
+pub mod join_scheduler_thread;
+pub mod join_scheduler_thread_in_background;
+pub mod llama_backend_state;
+pub mod llama_context_settings;
+pub mod llama_model_settings;
+pub mod loaded_llama_model;
+pub mod loaded_model_metadata;
+pub mod model_metadata_holder;
+pub mod prepares_scheduler_command;
+pub mod receives_stop_request;
+pub mod rejection_description;
+pub mod reports_available_parallelism;
+pub mod resident_llama_context;
+pub mod resident_llama_model;
+pub mod run_scheduler_unless_startup_abandoned;
+pub mod scheduler_command;
+pub mod scheduler_handle;
+pub mod scheduler_inbox;
+pub mod scheduler_message;
+pub mod scheduler_ready;
+pub mod scheduler_request_preparer;
+pub mod send_result_or_warn;
+pub mod send_startup_signal;
+pub mod sequence_id;
+pub mod sequence_id_guard;
+pub mod sequence_id_pool;
+pub mod shared_llama_backend;
+pub mod shuts_down;
+pub mod start_scheduler_thread;
+pub mod startup_outcome;
+pub mod startup_signal_delivery;
+pub mod system_parallelism;
+pub mod vram_occupancy;
+pub mod vram_residency;
+pub mod vram_resident;
diff --git a/paddler_agent_runtime/src/llama_backend_state.rs b/paddler_agent_runtime/src/llama_backend_state.rs
new file mode 100644
index 000000000..f8359a1a8
--- /dev/null
+++ b/paddler_agent_runtime/src/llama_backend_state.rs
@@ -0,0 +1,8 @@
+use std::sync::Arc;
+
+use llama_cpp_bindings::llama_backend::LlamaBackend;
+
+pub enum LlamaBackendState {
+    Initialized(Arc<LlamaBackend>),
+    Uninitialized,
+}
diff --git a/paddler_agent_runtime/src/llama_context_settings.rs b/paddler_agent_runtime/src/llama_context_settings.rs
new file mode 100644
index 000000000..1376d4bc8
--- /dev/null
+++ b/paddler_agent_runtime/src/llama_context_settings.rs
@@ -0,0 +1,71 @@
+use llama_cpp_bindings::context::params::LlamaContextParams;
+use llama_cpp_bindings_sys::LLAMA_FLASH_ATTN_TYPE_AUTO;
+
+use paddler_inference_parameters::model_runtime_parameters::ModelRuntimeParameters;
+
+use crate::agent_kv_cache_dtype::AgentKvCacheDtype;
+use crate::converts_to_llama_kv_cache_dtype::ConvertsToLlamaKvCacheDtype as _;
+
+pub struct LlamaContextSettings<'parameters> {
+    pub model_runtime_parameters: &'parameters ModelRuntimeParameters,
+    pub n_seq_max: u32,
+    pub thread_count: i32,
+}
+
+impl LlamaContextSettings<'_> {
+    #[must_use]
+    pub fn into_llama_context_params(self) -> LlamaContextParams {
+        let Self {
+            model_runtime_parameters,
+            n_seq_max,
+            thread_count,
+        } = self;
+
+        LlamaContextParams::default()
+            .with_n_ctx(Some(model_runtime_parameters.context_size))
+            .with_n_batch(model_runtime_parameters.n_batch.tokens().get())
+            .with_flash_attention_policy(LLAMA_FLASH_ATTN_TYPE_AUTO)
+            .with_n_seq_max(n_seq_max)
+            .with_n_threads(thread_count)
+            .with_n_threads_batch(thread_count)
+            .with_type_k(
+                AgentKvCacheDtype(model_runtime_parameters.k_cache_dtype.clone())
+                    .to_llama_kv_cache_dtype(),
+            )
+            .with_type_v(
+                AgentKvCacheDtype(model_runtime_parameters.v_cache_dtype.clone())
+                    .to_llama_kv_cache_dtype(),
+            )
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use paddler_inference_parameters::model_runtime_parameters::ModelRuntimeParameters;
+
+    use super::LlamaContextSettings;
+
+    #[test]
+    fn carries_the_model_runtime_parameters_into_the_context() {
+        let model_runtime_parameters = ModelRuntimeParameters::default();
+
+        let llama_context_params = LlamaContextSettings {
+            model_runtime_parameters: &model_runtime_parameters,
+            n_seq_max: 4,
+            thread_count: 6,
+        }
+        .into_llama_context_params();
+
+        assert_eq!(
+            llama_context_params.n_ctx(),
+            Some(model_runtime_parameters.context_size)
+        );
+        assert_eq!(
+            llama_context_params.n_batch(),
+            model_runtime_parameters.n_batch.tokens().get()
+        );
+        assert_eq!(llama_context_params.n_seq_max(), 4);
+        assert_eq!(llama_context_params.n_threads(), 6);
+        assert_eq!(llama_context_params.n_threads_batch(), 6);
+    }
+}
diff --git a/paddler_agent_runtime/src/llama_model_settings.rs b/paddler_agent_runtime/src/llama_model_settings.rs
new file mode 100644
index 000000000..068735ef8
--- /dev/null
+++ b/paddler_agent_runtime/src/llama_model_settings.rs
@@ -0,0 +1,43 @@
+use std::path::PathBuf;
+
+use llama_cpp_bindings::model::params::LlamaModelParams;
+
+use paddler_messaging::agent_issue_params::model_path::ModelPath;
+
+#[derive(Clone, Debug, Eq, PartialEq)]
+pub struct LlamaModelSettings {
+    pub model_path: PathBuf,
+    pub n_gpu_layers: i32,
+}
+
+impl LlamaModelSettings {
+    #[must_use]
+    pub fn reported_model_path(&self) -> ModelPath {
+        ModelPath::from(self.model_path.as_path())
+    }
+
+    #[must_use]
+    pub fn to_llama_model_params(&self) -> LlamaModelParams {
+        LlamaModelParams::default().with_n_gpu_layers(self.n_gpu_layers)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::path::PathBuf;
+
+    use super::LlamaModelSettings;
+
+    #[test]
+    fn offloads_the_layers_it_was_given() {
+        assert_eq!(
+            LlamaModelSettings {
+                model_path: PathBuf::from("model.gguf"),
+                n_gpu_layers: 12,
+            }
+            .to_llama_model_params()
+            .n_gpu_layers(),
+            12
+        );
+    }
+}
diff --git a/paddler_agent_runtime/src/loaded_llama_model.rs b/paddler_agent_runtime/src/loaded_llama_model.rs
new file mode 100644
index 000000000..53ca6077e
--- /dev/null
+++ b/paddler_agent_runtime/src/loaded_llama_model.rs
@@ -0,0 +1,285 @@
+use std::sync::Arc;
+
+use llama_cpp_bindings::SampledToken;
+use llama_cpp_bindings::context::LlamaContext;
+use llama_cpp_bindings::context::params::LlamaContextParams;
+use llama_cpp_bindings::llama_batch::LlamaBatch;
+use llama_cpp_bindings::model::LlamaModel;
+use llama_cpp_bindings::token::LlamaToken;
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent_status::agent_issue_fix::AgentIssueFix;
+use paddler_inference_parameters::model_runtime_parameters::ModelRuntimeParameters;
+use paddler_messaging::agent_issue::AgentIssue;
+use paddler_messaging::agent_issue_params::slot_cannot_start_params::SlotCannotStartParams;
+use paddler_messaging::inference_capability::InferenceCapability;
+use paddler_messaging::model_metadata::ModelMetadata;
+use paddler_spawn_resuming_panic::spawn_blocking_resuming_panic::spawn_blocking_resuming_panic;
+
+use crate::agent_runtime_error::AgentRuntimeError;
+use crate::inference_runtime_context::InferenceRuntimeContext;
+use crate::inference_thread_count::inference_thread_count;
+use crate::llama_context_settings::LlamaContextSettings;
+use crate::llama_model_settings::LlamaModelSettings;
+use crate::resident_llama_context::ResidentLlamaContext;
+use crate::resident_llama_model::ResidentLlamaModel;
+use crate::startup_outcome::StartupOutcome;
+use crate::vram_occupancy::VramOccupancy;
+
+fn read_model_metadata(model: &LlamaModel) -> Result<ModelMetadata, AgentRuntimeError> {
+    let mut model_metadata = ModelMetadata::default();
+
+    for metadata_index in 0..model.meta_count() {
+        model_metadata.set_meta_field(
+            model
+                .meta_key_by_index(metadata_index)
+                .map_err(AgentRuntimeError::ModelMetadataUnreadable)?,
+            model
+                .meta_val_str_by_index(metadata_index)
+                .map_err(AgentRuntimeError::ModelMetadataUnreadable)?,
+        );
+    }
+
+    Ok(model_metadata)
+}
+
+#[derive(Clone)]
+pub struct LoadedLlamaModel {
+    pub inference_thread_count: i32,
+    pub llama_model_settings: LlamaModelSettings,
+    pub model: Arc<ResidentLlamaModel>,
+    pub model_metadata: Arc<ModelMetadata>,
+}
+
+impl LoadedLlamaModel {
+    pub fn load(
+        inference_runtime_context: &InferenceRuntimeContext,
+        llama_model_settings: LlamaModelSettings,
+    ) -> Result<Self, AgentRuntimeError> {
+        let inference_thread_count =
+            inference_thread_count(inference_runtime_context.available_parallelism.as_ref())?;
+        let vram_occupancy = inference_runtime_context
+            .vram_residencies
+            .model_weights
+            .try_occupy()?;
+        let model_issue_path = llama_model_settings.reported_model_path();
+        let loaded = Self::load_from_file(
+            inference_runtime_context,
+            inference_thread_count,
+            llama_model_settings,
+            vram_occupancy,
+        );
+
+        match &loaded {
+            Ok(_loaded_llama_model) => {
+                inference_runtime_context
+                    .slot_aggregated_status
+                    .register_fix(&AgentIssueFix::ModelIsLoaded(model_issue_path));
+            }
+            Err(_load_failure) => {
+                inference_runtime_context
+                    .slot_aggregated_status
+                    .register_pipeline_issue(AgentIssue::ModelCannotBeLoaded(model_issue_path));
+            }
+        }
+
+        loaded
+    }
+
+    pub async fn load_unless_shutting_down(
+        cancellation_token: &CancellationToken,
+        inference_runtime_context: InferenceRuntimeContext,
+        llama_model_settings: LlamaModelSettings,
+    ) -> Result<StartupOutcome<Self>, AgentRuntimeError> {
+        if cancellation_token.is_cancelled() {
+            return Ok(StartupOutcome::Cancelled);
+        }
+
+        let loaded_llama_model = spawn_blocking_resuming_panic(move || {
+            Self::load(&inference_runtime_context, llama_model_settings)
+        })
+        .await?;
+
+        if cancellation_token.is_cancelled() {
+            loaded_llama_model.release().await;
+
+            return Ok(StartupOutcome::Cancelled);
+        }
+
+        Ok(StartupOutcome::Ready(loaded_llama_model))
+    }
+
+    pub async fn release(self) {
+        spawn_blocking_resuming_panic(move || drop(self)).await;
+    }
+
+    #[must_use]
+    pub const fn llama_context_settings<'parameters>(
+        &self,
+        model_runtime_parameters: &'parameters ModelRuntimeParameters,
+        n_seq_max: u32,
+    ) -> LlamaContextSettings<'parameters> {
+        LlamaContextSettings {
+            model_runtime_parameters,
+            n_seq_max,
+            thread_count: self.inference_thread_count,
+        }
+    }
+
+    pub fn create_llama_context(
+        &self,
+        inference_runtime_context: &InferenceRuntimeContext,
+        inference_capability: InferenceCapability,
+        llama_context_params: LlamaContextParams,
+    ) -> Result<ResidentLlamaContext<'_>, AgentRuntimeError> {
+        let vram_occupancy = inference_runtime_context
+            .vram_residencies
+            .llama_context(inference_capability)
+            .try_occupy()?;
+        let n_seq_max = llama_context_params.n_seq_max();
+
+        #[expect(
+            clippy::disallowed_methods,
+            reason = "the context is created here, paired with the VRAM occupancy of its capability"
+        )]
+        let created_llama_context = LlamaContext::from_model(
+            &self.model.llama_model,
+            &self.model.llama_backend,
+            llama_context_params,
+        );
+
+        created_llama_context
+            .map(|llama_context| ResidentLlamaContext {
+                llama_context,
+                vram_occupancy,
+            })
+            .map_err(|source| {
+                let context_creation_failure = AgentRuntimeError::ContextCreationFailed(source);
+                let error = context_creation_failure.to_string();
+
+                for slot_index in 0..n_seq_max {
+                    inference_runtime_context
+                        .slot_aggregated_status
+                        .register_pipeline_issue(AgentIssue::SlotCannotStart(
+                            SlotCannotStartParams {
+                                error: error.clone(),
+                                inference_capability,
+                                slot_index,
+                            },
+                        ));
+                }
+
+                context_creation_failure
+            })
+    }
+
+    pub fn warm_up_llama_context(
+        &self,
+        llama_context: &mut LlamaContext<'_>,
+        warmup_batch: &mut LlamaBatch<'static>,
+        desired_slots_total: u16,
+    ) -> Result<(), AgentRuntimeError> {
+        let warmup_tokens: Vec<LlamaToken> = [self.model.token_bos(), self.model.token_eos()]
+            .into_iter()
+            .flatten()
+            .collect();
+
+        llama_context.clear_kv_cache();
+
+        let warm_up = Self::decode_every_warm_up_position(
+            llama_context,
+            warmup_batch,
+            &warmup_tokens,
+            desired_slots_total,
+        );
+
+        llama_context.synchronize();
+        llama_context.clear_kv_cache();
+
+        warm_up
+    }
+
+    pub fn warm_up_and_serve<'model, TServe>(
+        &self,
+        mut llama_context: ResidentLlamaContext<'model>,
+        mut warmup_batch: LlamaBatch<'static>,
+        desired_slots_total: u16,
+        serve: TServe,
+    ) -> Result<(), AgentRuntimeError>
+    where
+        TServe: FnOnce(ResidentLlamaContext<'model>, LlamaBatch<'static>),
+    {
+        self.warm_up_llama_context(&mut llama_context, &mut warmup_batch, desired_slots_total)?;
+
+        serve(llama_context, warmup_batch);
+
+        Ok(())
+    }
+
+    fn decode_every_warm_up_position(
+        llama_context: &mut LlamaContext<'_>,
+        warmup_batch: &mut LlamaBatch<'static>,
+        warmup_tokens: &[LlamaToken],
+        desired_slots_total: u16,
+    ) -> Result<(), AgentRuntimeError> {
+        for (position, warmup_token) in (0..).zip(warmup_tokens) {
+            warmup_batch.clear();
+
+            for sequence_index in 0..i32::from(desired_slots_total) {
+                warmup_batch
+                    .add(
+                        &SampledToken::Content(*warmup_token),
+                        position,
+                        &[sequence_index],
+                        true,
+                    )
+                    .map_err(AgentRuntimeError::WarmUpBatchAssemblyFailed)?;
+            }
+
+            llama_context
+                .decode(warmup_batch)
+                .map_err(AgentRuntimeError::WarmUpDecodeFailed)?;
+        }
+
+        Ok(())
+    }
+
+    fn load_from_file(
+        inference_runtime_context: &InferenceRuntimeContext,
+        inference_thread_count: i32,
+        llama_model_settings: LlamaModelSettings,
+        vram_occupancy: VramOccupancy,
+    ) -> Result<Self, AgentRuntimeError> {
+        let llama_backend = inference_runtime_context
+            .llama_backend
+            .get_or_init()
+            .map_err(AgentRuntimeError::BackendInitializationFailed)?;
+
+        #[expect(
+            clippy::disallowed_methods,
+            reason = "the model is loaded here, paired with the VRAM occupancy of the model weights"
+        )]
+        let loaded_model = LlamaModel::load_from_file(
+            &llama_backend,
+            &llama_model_settings.model_path,
+            &llama_model_settings.to_llama_model_params(),
+        );
+
+        let model = loaded_model.map_err(|source| AgentRuntimeError::ModelLoadFailed {
+            model_path: llama_model_settings.model_path.clone(),
+            source,
+        })?;
+        let model_metadata = read_model_metadata(&model)?;
+
+        Ok(Self {
+            inference_thread_count,
+            llama_model_settings,
+            model: Arc::new(ResidentLlamaModel {
+                llama_model: model,
+                llama_backend,
+                vram_occupancy,
+            }),
+            model_metadata: Arc::new(model_metadata),
+        })
+    }
+}
diff --git a/paddler_agent_runtime/src/loaded_model_metadata.rs b/paddler_agent_runtime/src/loaded_model_metadata.rs
new file mode 100644
index 000000000..d4779c4a1
--- /dev/null
+++ b/paddler_agent_runtime/src/loaded_model_metadata.rs
@@ -0,0 +1,9 @@
+use std::sync::Arc;
+
+use paddler_messaging::model_metadata::ModelMetadata;
+
+#[derive(Clone, Debug)]
+pub enum LoadedModelMetadata {
+    Loaded(Arc<ModelMetadata>),
+    NoModelLoaded,
+}
diff --git a/paddler_agent_runtime/src/model_metadata_holder.rs b/paddler_agent_runtime/src/model_metadata_holder.rs
new file mode 100644
index 000000000..c1059d8b0
--- /dev/null
+++ b/paddler_agent_runtime/src/model_metadata_holder.rs
@@ -0,0 +1,84 @@
+use std::sync::Arc;
+
+use parking_lot::RwLock;
+
+use paddler_messaging::model_metadata::ModelMetadata;
+
+use crate::loaded_model_metadata::LoadedModelMetadata;
+
+pub struct ModelMetadataHolder {
+    loaded_model_metadata: RwLock<LoadedModelMetadata>,
+}
+
+impl ModelMetadataHolder {
+    #[must_use]
+    pub fn new() -> Self {
+        Self::default()
+    }
+
+    pub fn set_model_metadata(&self, model_metadata: Arc<ModelMetadata>) {
+        *self.loaded_model_metadata.write() = LoadedModelMetadata::Loaded(model_metadata);
+    }
+
+    pub fn forget_model_metadata(&self) {
+        *self.loaded_model_metadata.write() = LoadedModelMetadata::NoModelLoaded;
+    }
+
+    #[must_use]
+    pub fn get_model_metadata(&self) -> Option<ModelMetadata> {
+        match &*self.loaded_model_metadata.read() {
+            LoadedModelMetadata::Loaded(model_metadata) => {
+                Some(ModelMetadata::clone(model_metadata))
+            }
+            LoadedModelMetadata::NoModelLoaded => None,
+        }
+    }
+}
+
+impl Default for ModelMetadataHolder {
+    fn default() -> Self {
+        Self {
+            loaded_model_metadata: RwLock::new(LoadedModelMetadata::NoModelLoaded),
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::collections::BTreeMap;
+    use std::sync::Arc;
+
+    use paddler_messaging::model_metadata::ModelMetadata;
+
+    use super::ModelMetadataHolder;
+
+    fn llama_metadata() -> ModelMetadata {
+        ModelMetadata {
+            metadata: BTreeMap::from([("architecture".to_owned(), "llama".to_owned())]),
+        }
+    }
+
+    #[test]
+    fn new_holder_starts_empty() {
+        assert_eq!(ModelMetadataHolder::new().get_model_metadata(), None);
+    }
+
+    #[test]
+    fn stored_metadata_is_returned() {
+        let holder = ModelMetadataHolder::new();
+
+        holder.set_model_metadata(Arc::new(llama_metadata()));
+
+        assert_eq!(holder.get_model_metadata(), Some(llama_metadata()));
+    }
+
+    #[test]
+    fn forgotten_metadata_is_no_longer_returned() {
+        let holder = ModelMetadataHolder::new();
+
+        holder.set_model_metadata(Arc::new(llama_metadata()));
+        holder.forget_model_metadata();
+
+        assert_eq!(holder.get_model_metadata(), None);
+    }
+}
diff --git a/paddler_agent_runtime/src/prepares_scheduler_command.rs b/paddler_agent_runtime/src/prepares_scheduler_command.rs
new file mode 100644
index 000000000..6162a843d
--- /dev/null
+++ b/paddler_agent_runtime/src/prepares_scheduler_command.rs
@@ -0,0 +1,12 @@
+use crate::scheduler_command::SchedulerCommand;
+
+pub trait PreparesSchedulerCommand: Send + Sync + 'static {
+    type Command: SchedulerCommand;
+    type Request: Send + 'static;
+
+    fn prepare_scheduler_command(
+        &self,
+        agent_name: Option<&str>,
+        request: Self::Request,
+    ) -> Option<Self::Command>;
+}
diff --git a/paddler_agent/src/receives_stop_request.rs b/paddler_agent_runtime/src/receives_stop_request.rs
similarity index 100%
rename from paddler_agent/src/receives_stop_request.rs
rename to paddler_agent_runtime/src/receives_stop_request.rs
diff --git a/paddler_agent/src/rejection_description.rs b/paddler_agent_runtime/src/rejection_description.rs
similarity index 100%
rename from paddler_agent/src/rejection_description.rs
rename to paddler_agent_runtime/src/rejection_description.rs
diff --git a/paddler_agent_runtime/src/reports_available_parallelism.rs b/paddler_agent_runtime/src/reports_available_parallelism.rs
new file mode 100644
index 000000000..de16f8516
--- /dev/null
+++ b/paddler_agent_runtime/src/reports_available_parallelism.rs
@@ -0,0 +1,6 @@
+use std::io;
+use std::num::NonZeroUsize;
+
+pub trait ReportsAvailableParallelism: Send + Sync {
+    fn available_parallelism(&self) -> io::Result<NonZeroUsize>;
+}
diff --git a/paddler_agent_runtime/src/resident_llama_context.rs b/paddler_agent_runtime/src/resident_llama_context.rs
new file mode 100644
index 000000000..33a429a58
--- /dev/null
+++ b/paddler_agent_runtime/src/resident_llama_context.rs
@@ -0,0 +1,25 @@
+use std::ops::Deref;
+use std::ops::DerefMut;
+
+use llama_cpp_bindings::context::LlamaContext;
+
+use crate::vram_occupancy::VramOccupancy;
+
+pub struct ResidentLlamaContext<'model> {
+    pub llama_context: LlamaContext<'model>,
+    pub vram_occupancy: VramOccupancy,
+}
+
+impl<'model> Deref for ResidentLlamaContext<'model> {
+    type Target = LlamaContext<'model>;
+
+    fn deref(&self) -> &LlamaContext<'model> {
+        &self.llama_context
+    }
+}
+
+impl<'model> DerefMut for ResidentLlamaContext<'model> {
+    fn deref_mut(&mut self) -> &mut LlamaContext<'model> {
+        &mut self.llama_context
+    }
+}
diff --git a/paddler_agent_runtime/src/resident_llama_model.rs b/paddler_agent_runtime/src/resident_llama_model.rs
new file mode 100644
index 000000000..4a203ca4c
--- /dev/null
+++ b/paddler_agent_runtime/src/resident_llama_model.rs
@@ -0,0 +1,21 @@
+use std::ops::Deref;
+use std::sync::Arc;
+
+use llama_cpp_bindings::llama_backend::LlamaBackend;
+use llama_cpp_bindings::model::LlamaModel;
+
+use crate::vram_occupancy::VramOccupancy;
+
+pub struct ResidentLlamaModel {
+    pub llama_model: LlamaModel,
+    pub llama_backend: Arc<LlamaBackend>,
+    pub vram_occupancy: VramOccupancy,
+}
+
+impl Deref for ResidentLlamaModel {
+    type Target = LlamaModel;
+
+    fn deref(&self) -> &LlamaModel {
+        &self.llama_model
+    }
+}
diff --git a/paddler_agent_runtime/src/run_scheduler_unless_startup_abandoned.rs b/paddler_agent_runtime/src/run_scheduler_unless_startup_abandoned.rs
new file mode 100644
index 000000000..2e480ec41
--- /dev/null
+++ b/paddler_agent_runtime/src/run_scheduler_unless_startup_abandoned.rs
@@ -0,0 +1,15 @@
+use tokio::sync::mpsc::UnboundedSender;
+
+use crate::send_startup_signal::send_startup_signal;
+use crate::startup_signal_delivery::StartupSignalDelivery;
+
+pub fn run_scheduler_unless_startup_abandoned<TSignal>(
+    scheduler_startup_tx: &UnboundedSender<TSignal>,
+    signal: TSignal,
+    run_scheduler: impl FnOnce(),
+) {
+    match send_startup_signal(scheduler_startup_tx, signal) {
+        StartupSignalDelivery::Abandoned => {}
+        StartupSignalDelivery::Delivered => run_scheduler(),
+    }
+}
diff --git a/paddler_agent_runtime/src/scheduler_command.rs b/paddler_agent_runtime/src/scheduler_command.rs
new file mode 100644
index 000000000..ea7974eb8
--- /dev/null
+++ b/paddler_agent_runtime/src/scheduler_command.rs
@@ -0,0 +1,3 @@
+pub trait SchedulerCommand: Send + 'static {
+    fn reject_because_the_scheduler_stopped(self, agent_name: Option<&str>);
+}
diff --git a/paddler_agent_runtime/src/scheduler_handle.rs b/paddler_agent_runtime/src/scheduler_handle.rs
new file mode 100644
index 000000000..edf539257
--- /dev/null
+++ b/paddler_agent_runtime/src/scheduler_handle.rs
@@ -0,0 +1,30 @@
+use std::sync::Arc;
+use std::thread::JoinHandle;
+
+use crate::join_scheduler_thread_in_background::join_scheduler_thread_in_background;
+use crate::prepares_scheduler_command::PreparesSchedulerCommand;
+use crate::scheduler_request_preparer::SchedulerRequestPreparer;
+use crate::shuts_down::ShutsDown;
+
+pub struct SchedulerHandle<TPreparation: PreparesSchedulerCommand> {
+    pub request_preparer: Arc<SchedulerRequestPreparer<TPreparation>>,
+    pub scheduler_thread_handle: JoinHandle<()>,
+}
+
+impl<TPreparation> ShutsDown for SchedulerHandle<TPreparation>
+where
+    TPreparation: PreparesSchedulerCommand,
+{
+    async fn shut_down(self) {
+        let Self {
+            request_preparer,
+            scheduler_thread_handle,
+        } = self;
+
+        request_preparer
+            .stop_scheduler_after_pending_preparations()
+            .await;
+
+        join_scheduler_thread_in_background(request_preparer, scheduler_thread_handle).await;
+    }
+}
diff --git a/paddler_agent_runtime/src/scheduler_inbox.rs b/paddler_agent_runtime/src/scheduler_inbox.rs
new file mode 100644
index 000000000..edbbdb9b5
--- /dev/null
+++ b/paddler_agent_runtime/src/scheduler_inbox.rs
@@ -0,0 +1,56 @@
+use tokio::sync::mpsc::UnboundedReceiver;
+use tokio::sync::mpsc::error::TryRecvError;
+
+use crate::scheduler_command::SchedulerCommand;
+use crate::scheduler_message::SchedulerMessage;
+
+pub struct SchedulerInbox<TCommand: SchedulerCommand> {
+    pub scheduler_message_rx: UnboundedReceiver<SchedulerMessage<TCommand>>,
+    pub shutdown_requested: bool,
+}
+
+impl<TCommand: SchedulerCommand> SchedulerInbox<TCommand> {
+    #[must_use]
+    pub const fn new(scheduler_message_rx: UnboundedReceiver<SchedulerMessage<TCommand>>) -> Self {
+        Self {
+            scheduler_message_rx,
+            shutdown_requested: false,
+        }
+    }
+
+    pub fn take_next_waiting_command(&mut self) -> Option<TCommand> {
+        match self.scheduler_message_rx.try_recv() {
+            Ok(SchedulerMessage::Command(command)) => Some(command),
+            Ok(SchedulerMessage::Shutdown) | Err(TryRecvError::Disconnected) => {
+                self.shutdown_requested = true;
+
+                None
+            }
+            Err(TryRecvError::Empty) => None,
+        }
+    }
+
+    pub fn wait_for_next_command(&mut self) -> Option<TCommand> {
+        match self.scheduler_message_rx.blocking_recv() {
+            Some(SchedulerMessage::Command(command)) => Some(command),
+            Some(SchedulerMessage::Shutdown) | None => {
+                self.shutdown_requested = true;
+
+                None
+            }
+        }
+    }
+
+    pub fn reject_waiting_commands(&mut self, agent_name: Option<&str>) {
+        self.scheduler_message_rx.close();
+
+        while let Ok(scheduler_message) = self.scheduler_message_rx.try_recv() {
+            match scheduler_message {
+                SchedulerMessage::Command(command) => {
+                    command.reject_because_the_scheduler_stopped(agent_name);
+                }
+                SchedulerMessage::Shutdown => {}
+            }
+        }
+    }
+}
diff --git a/paddler_agent_runtime/src/scheduler_message.rs b/paddler_agent_runtime/src/scheduler_message.rs
new file mode 100644
index 000000000..c2092f486
--- /dev/null
+++ b/paddler_agent_runtime/src/scheduler_message.rs
@@ -0,0 +1,4 @@
+pub enum SchedulerMessage<TCommand> {
+    Command(TCommand),
+    Shutdown,
+}
diff --git a/paddler_agent_runtime/src/scheduler_ready.rs b/paddler_agent_runtime/src/scheduler_ready.rs
new file mode 100644
index 000000000..f9278d813
--- /dev/null
+++ b/paddler_agent_runtime/src/scheduler_ready.rs
@@ -0,0 +1,8 @@
+use std::sync::Arc;
+
+use crate::prepares_scheduler_command::PreparesSchedulerCommand;
+use crate::scheduler_request_preparer::SchedulerRequestPreparer;
+
+pub struct SchedulerReady<TPreparation: PreparesSchedulerCommand> {
+    pub request_preparer: Arc<SchedulerRequestPreparer<TPreparation>>,
+}
diff --git a/paddler_agent_runtime/src/scheduler_request_preparer.rs b/paddler_agent_runtime/src/scheduler_request_preparer.rs
new file mode 100644
index 000000000..3cf24f361
--- /dev/null
+++ b/paddler_agent_runtime/src/scheduler_request_preparer.rs
@@ -0,0 +1,55 @@
+use log::info;
+use std::sync::Arc;
+use tokio::sync::mpsc::UnboundedSender;
+use tokio::sync::mpsc::error::SendError;
+use tokio_util::task::TaskTracker;
+
+use crate::prepares_scheduler_command::PreparesSchedulerCommand;
+use crate::scheduler_command::SchedulerCommand as _;
+use crate::scheduler_message::SchedulerMessage;
+
+pub struct SchedulerRequestPreparer<TPreparation: PreparesSchedulerCommand> {
+    pub agent_name: Option<String>,
+    pub preparation: TPreparation,
+    pub preparation_tasks: TaskTracker,
+    pub scheduler_message_tx: UnboundedSender<SchedulerMessage<TPreparation::Command>>,
+}
+
+impl<TPreparation: PreparesSchedulerCommand> SchedulerRequestPreparer<TPreparation> {
+    pub fn prepare(self: &Arc<Self>, request: TPreparation::Request) {
+        let request_preparer = self.clone();
+
+        self.preparation_tasks.spawn_blocking(move || {
+            if let Some(command) = request_preparer
+                .preparation
+                .prepare_scheduler_command(request_preparer.agent_name.as_deref(), request)
+            {
+                request_preparer.send(SchedulerMessage::Command(command));
+            }
+        });
+    }
+
+    pub async fn stop_scheduler_after_pending_preparations(&self) {
+        self.preparation_tasks.close();
+        self.preparation_tasks.wait().await;
+        self.send(SchedulerMessage::Shutdown);
+    }
+
+    fn send(&self, scheduler_message: SchedulerMessage<TPreparation::Command>) {
+        if let Err(SendError(undelivered_message)) =
+            self.scheduler_message_tx.send(scheduler_message)
+        {
+            match undelivered_message {
+                SchedulerMessage::Command(command) => {
+                    command.reject_because_the_scheduler_stopped(self.agent_name.as_deref());
+                }
+                SchedulerMessage::Shutdown => {
+                    info!(
+                        "{:?}: the scheduler stopped before its shutdown was requested",
+                        self.agent_name
+                    );
+                }
+            }
+        }
+    }
+}
diff --git a/paddler_agent/src/send_result_or_warn.rs b/paddler_agent_runtime/src/send_result_or_warn.rs
similarity index 100%
rename from paddler_agent/src/send_result_or_warn.rs
rename to paddler_agent_runtime/src/send_result_or_warn.rs
diff --git a/paddler_agent_runtime/src/send_startup_signal.rs b/paddler_agent_runtime/src/send_startup_signal.rs
new file mode 100644
index 000000000..a3baf1029
--- /dev/null
+++ b/paddler_agent_runtime/src/send_startup_signal.rs
@@ -0,0 +1,52 @@
+use log::debug;
+use tokio::sync::mpsc::UnboundedSender;
+use tokio::sync::mpsc::error::SendError;
+
+use crate::startup_signal_delivery::StartupSignalDelivery;
+
+pub fn send_startup_signal<TSignal>(
+    startup_signal_tx: &UnboundedSender<TSignal>,
+    signal: TSignal,
+) -> StartupSignalDelivery {
+    match startup_signal_tx.send(signal) {
+        Ok(()) => StartupSignalDelivery::Delivered,
+        Err(SendError(_undelivered_signal)) => {
+            debug!("The arbiter abandoned the scheduler startup; stopping the scheduler thread");
+
+            StartupSignalDelivery::Abandoned
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::mem::discriminant;
+
+    use tokio::sync::mpsc::unbounded_channel;
+
+    use super::send_startup_signal;
+    use crate::startup_signal_delivery::StartupSignalDelivery;
+
+    #[test]
+    fn delivers_the_signal_to_a_live_receiver() {
+        let (startup_signal_tx, mut startup_signal_rx) = unbounded_channel::<()>();
+
+        assert_eq!(
+            discriminant(&send_startup_signal(&startup_signal_tx, ())),
+            discriminant(&StartupSignalDelivery::Delivered)
+        );
+        assert_eq!(startup_signal_rx.try_recv(), Ok(()));
+    }
+
+    #[test]
+    fn reports_an_abandoned_startup_when_the_receiver_was_dropped() {
+        let (startup_signal_tx, startup_signal_rx) = unbounded_channel::<()>();
+
+        drop(startup_signal_rx);
+
+        assert_eq!(
+            discriminant(&send_startup_signal(&startup_signal_tx, ())),
+            discriminant(&StartupSignalDelivery::Abandoned)
+        );
+    }
+}
diff --git a/paddler_agent_runtime/src/sequence_id.rs b/paddler_agent_runtime/src/sequence_id.rs
new file mode 100644
index 000000000..e88de7d10
--- /dev/null
+++ b/paddler_agent_runtime/src/sequence_id.rs
@@ -0,0 +1,37 @@
+use llama_cpp_bindings::context::LlamaContext;
+use llama_cpp_bindings::error::ClearKvCacheSeqError;
+use log::error;
+
+#[derive(Clone, Copy, Debug, Eq, Ord, PartialEq, PartialOrd)]
+pub struct SequenceId(pub u16);
+
+impl SequenceId {
+    #[must_use]
+    pub fn llama_sequence_id(self) -> i32 {
+        i32::from(self.0)
+    }
+
+    #[must_use]
+    pub fn kv_cache_sequence_id(self) -> u32 {
+        u32::from(self.0)
+    }
+
+    pub fn clear_kv_cache_in(
+        self,
+        llama_context: &mut LlamaContext<'_>,
+    ) -> Result<(), ClearKvCacheSeqError> {
+        llama_context.clear_kv_cache_seq(Some(self.kv_cache_sequence_id()), None, None)
+    }
+
+    pub fn release_kv_cache_in(
+        self,
+        llama_context: &mut LlamaContext<'_>,
+        agent_name: Option<&str>,
+    ) {
+        if let Err(clear_error) = self.clear_kv_cache_in(llama_context) {
+            error!(
+                "{agent_name:?}: {clear_error}; the next request admitted to sequence {self:?} clears it again"
+            );
+        }
+    }
+}
diff --git a/paddler_agent/src/sequence_id_guard.rs b/paddler_agent_runtime/src/sequence_id_guard.rs
similarity index 86%
rename from paddler_agent/src/sequence_id_guard.rs
rename to paddler_agent_runtime/src/sequence_id_guard.rs
index 5c2b2400d..214e9f2a7 100644
--- a/paddler_agent/src/sequence_id_guard.rs
+++ b/paddler_agent_runtime/src/sequence_id_guard.rs
@@ -1,7 +1,8 @@
+use crate::sequence_id::SequenceId;
 use crate::sequence_id_pool::SequenceIdPool;
 
 pub struct SequenceIdGuard {
-    sequence_id: i32,
+    sequence_id: u16,
     sequence_id_pool: SequenceIdPool,
 }
 
@@ -15,8 +16,8 @@ impl SequenceIdGuard {
     }
 
     #[must_use]
-    pub const fn sequence_id(&self) -> i32 {
-        self.sequence_id
+    pub const fn sequence_id(&self) -> SequenceId {
+        SequenceId(self.sequence_id)
     }
 }
 
@@ -29,6 +30,7 @@ impl Drop for SequenceIdGuard {
 #[cfg(test)]
 mod tests {
     use super::SequenceIdGuard;
+    use crate::sequence_id::SequenceId;
     use crate::sequence_id_pool::SequenceIdPool;
 
     #[test]
@@ -37,7 +39,7 @@ mod tests {
 
         let guard = SequenceIdGuard::acquire(&sequence_id_pool).unwrap();
 
-        assert_eq!(guard.sequence_id(), 0);
+        assert_eq!(guard.sequence_id(), SequenceId(0));
         assert_eq!(sequence_id_pool.available_count(), 0);
     }
 
diff --git a/paddler_agent/src/sequence_id_pool.rs b/paddler_agent_runtime/src/sequence_id_pool.rs
similarity index 90%
rename from paddler_agent/src/sequence_id_pool.rs
rename to paddler_agent_runtime/src/sequence_id_pool.rs
index a2bce9303..8509fbae2 100644
--- a/paddler_agent/src/sequence_id_pool.rs
+++ b/paddler_agent_runtime/src/sequence_id_pool.rs
@@ -3,23 +3,23 @@ use std::rc::Rc;
 
 #[derive(Clone)]
 pub struct SequenceIdPool {
-    available_ids: Rc<RefCell<Vec<i32>>>,
+    available_ids: Rc<RefCell<Vec<u16>>>,
 }
 
 impl SequenceIdPool {
     #[must_use]
     pub fn new(max_sequences: u16) -> Self {
         Self {
-            available_ids: Rc::new(RefCell::new((0..i32::from(max_sequences)).rev().collect())),
+            available_ids: Rc::new(RefCell::new((0..max_sequences).rev().collect())),
         }
     }
 
     #[must_use]
-    pub fn acquire(&self) -> Option<i32> {
+    pub fn acquire(&self) -> Option<u16> {
         self.available_ids.borrow_mut().pop()
     }
 
-    pub fn release(&self, sequence_id: i32) {
+    pub fn release(&self, sequence_id: u16) {
         self.available_ids.borrow_mut().push(sequence_id);
     }
 
diff --git a/paddler_agent_runtime/src/shared_llama_backend.rs b/paddler_agent_runtime/src/shared_llama_backend.rs
new file mode 100644
index 000000000..5f3e713d0
--- /dev/null
+++ b/paddler_agent_runtime/src/shared_llama_backend.rs
@@ -0,0 +1,53 @@
+use std::sync::Arc;
+
+use llama_cpp_bindings::LlamaCppError;
+use llama_cpp_bindings::llama_backend::LlamaBackend;
+use parking_lot::Mutex;
+
+use crate::llama_backend_state::LlamaBackendState;
+
+pub struct SharedLlamaBackend {
+    llama_backend_state: Mutex<LlamaBackendState>,
+}
+
+impl SharedLlamaBackend {
+    pub fn get_or_init(&self) -> Result<Arc<LlamaBackend>, LlamaCppError> {
+        let mut llama_backend_state = self.llama_backend_state.lock();
+
+        match &*llama_backend_state {
+            LlamaBackendState::Initialized(llama_backend) => Ok(llama_backend.clone()),
+            LlamaBackendState::Uninitialized => {
+                let llama_backend = Arc::new(LlamaBackend::init()?);
+
+                *llama_backend_state = LlamaBackendState::Initialized(llama_backend.clone());
+
+                Ok(llama_backend)
+            }
+        }
+    }
+}
+
+impl Default for SharedLlamaBackend {
+    fn default() -> Self {
+        Self {
+            llama_backend_state: Mutex::new(LlamaBackendState::Uninitialized),
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::sync::Arc;
+
+    use super::SharedLlamaBackend;
+
+    #[test]
+    fn every_model_of_an_agent_shares_one_backend() {
+        let shared_llama_backend = SharedLlamaBackend::default();
+
+        assert!(Arc::ptr_eq(
+            &shared_llama_backend.get_or_init().unwrap(),
+            &shared_llama_backend.get_or_init().unwrap()
+        ));
+    }
+}
diff --git a/paddler_agent_runtime/src/shuts_down.rs b/paddler_agent_runtime/src/shuts_down.rs
new file mode 100644
index 000000000..5c4ea5461
--- /dev/null
+++ b/paddler_agent_runtime/src/shuts_down.rs
@@ -0,0 +1,5 @@
+use std::future::Future;
+
+pub trait ShutsDown {
+    fn shut_down(self) -> impl Future<Output = ()> + Send;
+}
diff --git a/paddler_agent_runtime/src/start_scheduler_thread.rs b/paddler_agent_runtime/src/start_scheduler_thread.rs
new file mode 100644
index 000000000..c212df78f
--- /dev/null
+++ b/paddler_agent_runtime/src/start_scheduler_thread.rs
@@ -0,0 +1,42 @@
+use std::thread::spawn;
+
+use tokio::sync::mpsc::UnboundedSender;
+use tokio::sync::mpsc::unbounded_channel;
+use tokio_util::sync::CancellationToken;
+
+use crate::await_scheduler_startup::await_scheduler_startup;
+use crate::prepares_scheduler_command::PreparesSchedulerCommand;
+use crate::scheduler_handle::SchedulerHandle;
+use crate::scheduler_ready::SchedulerReady;
+use crate::send_startup_signal::send_startup_signal;
+use crate::startup_outcome::StartupOutcome;
+
+pub async fn start_scheduler_thread<TPreparation, TError, TStartScheduler>(
+    cancellation_token: &CancellationToken,
+    start_scheduler: TStartScheduler,
+) -> Result<StartupOutcome<SchedulerHandle<TPreparation>>, TError>
+where
+    TPreparation: PreparesSchedulerCommand,
+    TError: Send + 'static,
+    TStartScheduler: FnOnce(
+            &CancellationToken,
+            &UnboundedSender<Result<SchedulerReady<TPreparation>, TError>>,
+        ) -> Result<(), TError>
+        + Send
+        + 'static,
+{
+    let (scheduler_startup_tx, scheduler_startup_rx) = unbounded_channel();
+    let agent_shutdown = cancellation_token.clone();
+    let scheduler_thread_handle = spawn(move || {
+        if let Err(startup_error) = start_scheduler(&agent_shutdown, &scheduler_startup_tx) {
+            send_startup_signal(&scheduler_startup_tx, Err(startup_error));
+        }
+    });
+
+    await_scheduler_startup(
+        cancellation_token,
+        scheduler_startup_rx,
+        scheduler_thread_handle,
+    )
+    .await
+}
diff --git a/paddler_agent_runtime/src/startup_outcome.rs b/paddler_agent_runtime/src/startup_outcome.rs
new file mode 100644
index 000000000..4a23b4bf5
--- /dev/null
+++ b/paddler_agent_runtime/src/startup_outcome.rs
@@ -0,0 +1,4 @@
+pub enum StartupOutcome<TStarted> {
+    Cancelled,
+    Ready(TStarted),
+}
diff --git a/paddler_agent/src/startup_signal_delivery.rs b/paddler_agent_runtime/src/startup_signal_delivery.rs
similarity index 100%
rename from paddler_agent/src/startup_signal_delivery.rs
rename to paddler_agent_runtime/src/startup_signal_delivery.rs
diff --git a/paddler_agent_runtime/src/system_parallelism.rs b/paddler_agent_runtime/src/system_parallelism.rs
new file mode 100644
index 000000000..0e0948c33
--- /dev/null
+++ b/paddler_agent_runtime/src/system_parallelism.rs
@@ -0,0 +1,13 @@
+use std::io;
+use std::num::NonZeroUsize;
+use std::thread::available_parallelism;
+
+use crate::reports_available_parallelism::ReportsAvailableParallelism;
+
+pub struct SystemParallelism;
+
+impl ReportsAvailableParallelism for SystemParallelism {
+    fn available_parallelism(&self) -> io::Result<NonZeroUsize> {
+        available_parallelism()
+    }
+}
diff --git a/paddler_agent_runtime/src/vram_occupancy.rs b/paddler_agent_runtime/src/vram_occupancy.rs
new file mode 100644
index 000000000..1be4f5a0a
--- /dev/null
+++ b/paddler_agent_runtime/src/vram_occupancy.rs
@@ -0,0 +1,31 @@
+use std::sync::Arc;
+use std::sync::atomic::AtomicBool;
+use std::sync::atomic::Ordering;
+
+use crate::agent_runtime_error::AgentRuntimeError;
+use crate::vram_resident::VramResident;
+
+#[derive(Debug)]
+pub struct VramOccupancy {
+    occupied: Arc<AtomicBool>,
+}
+
+impl VramOccupancy {
+    pub fn claim(
+        occupied: &Arc<AtomicBool>,
+        resident: VramResident,
+    ) -> Result<Self, AgentRuntimeError> {
+        match occupied.compare_exchange(false, true, Ordering::AcqRel, Ordering::Acquire) {
+            Ok(_was_vacant) => Ok(Self {
+                occupied: occupied.clone(),
+            }),
+            Err(_already_occupied) => Err(resident.already_resident()),
+        }
+    }
+}
+
+impl Drop for VramOccupancy {
+    fn drop(&mut self) {
+        self.occupied.store(false, Ordering::Release);
+    }
+}
diff --git a/paddler_agent_runtime/src/vram_residency.rs b/paddler_agent_runtime/src/vram_residency.rs
new file mode 100644
index 000000000..543b899ae
--- /dev/null
+++ b/paddler_agent_runtime/src/vram_residency.rs
@@ -0,0 +1,76 @@
+use std::sync::Arc;
+use std::sync::atomic::AtomicBool;
+
+use crate::agent_runtime_error::AgentRuntimeError;
+use crate::vram_occupancy::VramOccupancy;
+use crate::vram_resident::VramResident;
+
+#[derive(Clone)]
+pub struct VramResidency {
+    occupied: Arc<AtomicBool>,
+    resident: VramResident,
+}
+
+impl VramResidency {
+    #[must_use]
+    pub fn vacant(resident: VramResident) -> Self {
+        Self {
+            occupied: Arc::new(AtomicBool::new(false)),
+            resident,
+        }
+    }
+
+    pub fn try_occupy(&self) -> Result<VramOccupancy, AgentRuntimeError> {
+        VramOccupancy::claim(&self.occupied, self.resident)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use paddler_messaging::inference_capability::InferenceCapability;
+
+    use super::VramResidency;
+    use crate::agent_runtime_error::AgentRuntimeError;
+    use crate::vram_resident::VramResident;
+
+    #[test]
+    fn an_occupied_residency_refuses_a_second_occupant_naming_what_is_resident() {
+        for (resident, refusal) in [
+            (
+                VramResident::LlamaContext(InferenceCapability::Embeddings),
+                AgentRuntimeError::LlamaContextAlreadyResident {
+                    inference_capability: InferenceCapability::Embeddings,
+                },
+            ),
+            (
+                VramResident::ModelWeights,
+                AgentRuntimeError::ModelWeightsAlreadyResident,
+            ),
+            (
+                VramResident::MultimodalProjection,
+                AgentRuntimeError::MultimodalProjectionAlreadyResident,
+            ),
+        ] {
+            let vram_residency = VramResidency::vacant(resident);
+            let _first_occupancy = vram_residency.try_occupy().unwrap();
+
+            assert_eq!(
+                vram_residency
+                    .try_occupy()
+                    .err()
+                    .map(|error| error.to_string()),
+                Some(refusal.to_string())
+            );
+        }
+    }
+
+    #[test]
+    fn dropping_the_occupancy_vacates_the_residency_for_its_clones() {
+        let vram_residency = VramResidency::vacant(VramResident::ModelWeights);
+        let cloned_vram_residency = vram_residency.clone();
+
+        drop(vram_residency.try_occupy().unwrap());
+
+        assert!(cloned_vram_residency.try_occupy().is_ok());
+    }
+}
diff --git a/paddler_agent_runtime/src/vram_resident.rs b/paddler_agent_runtime/src/vram_resident.rs
new file mode 100644
index 000000000..6660d818e
--- /dev/null
+++ b/paddler_agent_runtime/src/vram_resident.rs
@@ -0,0 +1,25 @@
+use paddler_messaging::inference_capability::InferenceCapability;
+
+use crate::agent_runtime_error::AgentRuntimeError;
+
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+pub enum VramResident {
+    LlamaContext(InferenceCapability),
+    ModelWeights,
+    MultimodalProjection,
+}
+
+impl VramResident {
+    #[must_use]
+    pub const fn already_resident(self) -> AgentRuntimeError {
+        match self {
+            Self::LlamaContext(inference_capability) => {
+                AgentRuntimeError::LlamaContextAlreadyResident {
+                    inference_capability,
+                }
+            }
+            Self::ModelWeights => AgentRuntimeError::ModelWeightsAlreadyResident,
+            Self::MultimodalProjection => AgentRuntimeError::MultimodalProjectionAlreadyResident,
+        }
+    }
+}
diff --git a/paddler_agent_runtime/tests/a_model_load_requested_during_shutdown_loads_nothing.rs b/paddler_agent_runtime/tests/a_model_load_requested_during_shutdown_loads_nothing.rs
new file mode 100644
index 000000000..647307b94
--- /dev/null
+++ b/paddler_agent_runtime/tests/a_model_load_requested_during_shutdown_loads_nothing.rs
@@ -0,0 +1,44 @@
+use std::collections::BTreeSet;
+use std::mem::discriminant;
+use std::path::PathBuf;
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent_runtime::llama_model_settings::LlamaModelSettings;
+use paddler_agent_runtime::loaded_llama_model::LoadedLlamaModel;
+use paddler_agent_runtime::startup_outcome::StartupOutcome;
+use paddler_messaging::produces_snapshot::ProducesSnapshot as _;
+
+use crate::inference_runtime_context_fixture::inference_runtime_context_fixture;
+
+#[tokio::test]
+async fn a_model_load_requested_during_shutdown_loads_nothing() {
+    let inference_runtime_context = inference_runtime_context_fixture();
+    let agent_shutdown = CancellationToken::new();
+
+    agent_shutdown.cancel();
+
+    let load_outcome = LoadedLlamaModel::load_unless_shutting_down(
+        &agent_shutdown,
+        inference_runtime_context.clone(),
+        LlamaModelSettings {
+            model_path: PathBuf::from("model-that-is-never-read.gguf"),
+            n_gpu_layers: 0,
+        },
+    )
+    .await
+    .expect("a model load requested during shutdown must not attempt to load");
+
+    assert_eq!(
+        discriminant(&load_outcome),
+        discriminant(&StartupOutcome::Cancelled)
+    );
+    assert_eq!(
+        inference_runtime_context
+            .slot_aggregated_status
+            .make_snapshot()
+            .status
+            .issues,
+        BTreeSet::new()
+    );
+}
diff --git a/paddler_agent_runtime/tests/a_scheduler_that_fails_before_readiness_reports_its_error.rs b/paddler_agent_runtime/tests/a_scheduler_that_fails_before_readiness_reports_its_error.rs
new file mode 100644
index 000000000..58d99f5c0
--- /dev/null
+++ b/paddler_agent_runtime/tests/a_scheduler_that_fails_before_readiness_reports_its_error.rs
@@ -0,0 +1,19 @@
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent_runtime::start_scheduler_thread::start_scheduler_thread;
+
+use crate::even_number_preparation::EvenNumberPreparation;
+
+#[tokio::test]
+async fn a_scheduler_that_fails_before_readiness_reports_its_error() {
+    let startup_result = start_scheduler_thread::<EvenNumberPreparation, String, _>(
+        &CancellationToken::new(),
+        |_agent_shutdown, _scheduler_startup_tx| Err("the model could not be loaded".to_owned()),
+    )
+    .await;
+
+    assert_eq!(
+        startup_result.err(),
+        Some("the model could not be loaded".to_owned())
+    );
+}
diff --git a/paddler_agent_runtime/tests/a_scheduler_that_signals_readiness_starts.rs b/paddler_agent_runtime/tests/a_scheduler_that_signals_readiness_starts.rs
new file mode 100644
index 000000000..9df69119e
--- /dev/null
+++ b/paddler_agent_runtime/tests/a_scheduler_that_signals_readiness_starts.rs
@@ -0,0 +1,42 @@
+use std::sync::Arc;
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent_runtime::run_scheduler_unless_startup_abandoned::run_scheduler_unless_startup_abandoned;
+use paddler_agent_runtime::scheduler_ready::SchedulerReady;
+use paddler_agent_runtime::start_scheduler_thread::start_scheduler_thread;
+
+use crate::number_request_preparer::NumberRequestPreparer;
+use crate::reports_ready_request_preparer::ReportsReadyRequestPreparer as _;
+
+#[tokio::test]
+async fn a_scheduler_that_signals_readiness_starts() {
+    let NumberRequestPreparer {
+        request_preparer, ..
+    } = NumberRequestPreparer::new();
+    let signalled_request_preparer = request_preparer.clone();
+
+    let spawn_outcome = start_scheduler_thread::<_, String, _>(
+        &CancellationToken::new(),
+        move |_agent_shutdown, scheduler_startup_tx| {
+            run_scheduler_unless_startup_abandoned(
+                scheduler_startup_tx,
+                Ok(SchedulerReady { request_preparer }),
+                || {},
+            );
+
+            Ok(())
+        },
+    )
+    .await
+    .expect("a scheduler that signalled readiness must start");
+
+    assert!(
+        spawn_outcome
+            .ready_request_preparer()
+            .is_some_and(|ready_request_preparer| Arc::ptr_eq(
+                ready_request_preparer,
+                &signalled_request_preparer
+            ))
+    );
+}
diff --git a/paddler_agent_runtime/tests/a_scheduler_whose_startup_was_abandoned_never_runs.rs b/paddler_agent_runtime/tests/a_scheduler_whose_startup_was_abandoned_never_runs.rs
new file mode 100644
index 000000000..f92a32ced
--- /dev/null
+++ b/paddler_agent_runtime/tests/a_scheduler_whose_startup_was_abandoned_never_runs.rs
@@ -0,0 +1,12 @@
+use tokio::sync::mpsc::unbounded_channel;
+
+use crate::scheduler_ran_after_handover::scheduler_ran_after_handover;
+
+#[test]
+fn a_scheduler_whose_startup_was_abandoned_never_runs() {
+    let (scheduler_startup_tx, scheduler_startup_rx) = unbounded_channel();
+
+    drop(scheduler_startup_rx);
+
+    assert!(!scheduler_ran_after_handover(&scheduler_startup_tx));
+}
diff --git a/paddler_agent_runtime/tests/a_scheduler_whose_startup_was_delivered_runs.rs b/paddler_agent_runtime/tests/a_scheduler_whose_startup_was_delivered_runs.rs
new file mode 100644
index 000000000..2c688b821
--- /dev/null
+++ b/paddler_agent_runtime/tests/a_scheduler_whose_startup_was_delivered_runs.rs
@@ -0,0 +1,11 @@
+use tokio::sync::mpsc::unbounded_channel;
+
+use crate::scheduler_ran_after_handover::scheduler_ran_after_handover;
+
+#[test]
+fn a_scheduler_whose_startup_was_delivered_runs() {
+    let (scheduler_startup_tx, mut scheduler_startup_rx) = unbounded_channel();
+
+    assert!(scheduler_ran_after_handover(&scheduler_startup_tx));
+    assert_eq!(scheduler_startup_rx.try_recv(), Ok(()));
+}
diff --git a/paddler_agent_runtime/tests/a_startup_cancelled_while_its_scheduler_signals_readiness_is_cancelled.rs b/paddler_agent_runtime/tests/a_startup_cancelled_while_its_scheduler_signals_readiness_is_cancelled.rs
new file mode 100644
index 000000000..f402003d3
--- /dev/null
+++ b/paddler_agent_runtime/tests/a_startup_cancelled_while_its_scheduler_signals_readiness_is_cancelled.rs
@@ -0,0 +1,73 @@
+use std::future::Future as _;
+use std::future::poll_fn;
+use std::mem::discriminant;
+use std::pin::pin;
+use std::sync::mpsc::channel;
+use std::task::Poll;
+use std::thread::spawn;
+
+use tokio::sync::mpsc::unbounded_channel;
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent_runtime::await_scheduler_startup::await_scheduler_startup;
+use paddler_agent_runtime::scheduler_ready::SchedulerReady;
+use paddler_agent_runtime::send_startup_signal::send_startup_signal;
+use paddler_agent_runtime::startup_signal_delivery::StartupSignalDelivery;
+
+use crate::number_request_preparer::NumberRequestPreparer;
+use crate::reports_ready_request_preparer::ReportsReadyRequestPreparer as _;
+
+#[tokio::test]
+async fn a_startup_cancelled_while_its_scheduler_signals_readiness_is_cancelled() {
+    let NumberRequestPreparer {
+        request_preparer, ..
+    } = NumberRequestPreparer::new();
+    let (scheduler_startup_tx, scheduler_startup_rx) = unbounded_channel();
+    let (readiness_requested_tx, readiness_requested_rx) = channel();
+    let (readiness_signalled_tx, readiness_signalled_rx) = channel();
+    let scheduler_thread_handle = spawn(move || {
+        readiness_requested_rx
+            .recv()
+            .expect("the test must request the readiness signal");
+        readiness_signalled_tx
+            .send(send_startup_signal(
+                &scheduler_startup_tx,
+                Ok::<_, String>(SchedulerReady { request_preparer }),
+            ))
+            .expect("the test must wait for the readiness signal");
+    });
+    let cancellation_token = CancellationToken::new();
+    let mut startup = pin!(await_scheduler_startup(
+        &cancellation_token,
+        scheduler_startup_rx,
+        scheduler_thread_handle,
+    ));
+
+    assert!(
+        poll_fn(|context| Poll::Ready(startup.as_mut().poll(context).is_pending())).await,
+        "the startup must wait for its scheduler"
+    );
+
+    readiness_requested_tx
+        .send(())
+        .expect("the scheduler thread must wait for the readiness request");
+
+    assert_eq!(
+        discriminant(
+            &readiness_signalled_rx
+                .recv()
+                .expect("the scheduler thread must report its readiness signal")
+        ),
+        discriminant(&StartupSignalDelivery::Delivered)
+    );
+
+    cancellation_token.cancel();
+
+    assert!(
+        startup
+            .await
+            .expect("a cancelled startup must join its scheduler thread")
+            .ready_request_preparer()
+            .is_none()
+    );
+}
diff --git a/paddler_agent_runtime/tests/commands_for_a_stopped_scheduler_are_rejected.rs b/paddler_agent_runtime/tests/commands_for_a_stopped_scheduler_are_rejected.rs
new file mode 100644
index 000000000..d4e2b6b87
--- /dev/null
+++ b/paddler_agent_runtime/tests/commands_for_a_stopped_scheduler_are_rejected.rs
@@ -0,0 +1,22 @@
+use tokio::sync::mpsc::error::TryRecvError;
+
+use crate::number_request_preparer::NumberRequestPreparer;
+
+#[tokio::test]
+async fn commands_for_a_stopped_scheduler_are_rejected() {
+    let NumberRequestPreparer {
+        mut rejected_rx,
+        request_preparer,
+        scheduler_message_rx,
+    } = NumberRequestPreparer::new();
+
+    drop(scheduler_message_rx);
+
+    request_preparer.prepare(4);
+    request_preparer
+        .stop_scheduler_after_pending_preparations()
+        .await;
+
+    assert_eq!(rejected_rx.try_recv(), Ok(4));
+    assert_eq!(rejected_rx.try_recv(), Err(TryRecvError::Empty));
+}
diff --git a/paddler_agent_runtime/tests/drained_scheduler_messages.rs b/paddler_agent_runtime/tests/drained_scheduler_messages.rs
new file mode 100644
index 000000000..8e06bc1cb
--- /dev/null
+++ b/paddler_agent_runtime/tests/drained_scheduler_messages.rs
@@ -0,0 +1,17 @@
+use tokio::sync::mpsc::UnboundedReceiver;
+
+use paddler_agent_runtime::scheduler_message::SchedulerMessage;
+
+use crate::number_command::NumberCommand;
+
+pub fn drained_scheduler_messages(
+    mut scheduler_message_rx: UnboundedReceiver<SchedulerMessage<NumberCommand>>,
+) -> Vec<SchedulerMessage<NumberCommand>> {
+    let mut scheduler_messages = Vec::new();
+
+    while let Ok(scheduler_message) = scheduler_message_rx.try_recv() {
+        scheduler_messages.push(scheduler_message);
+    }
+
+    scheduler_messages
+}
diff --git a/paddler_agent_runtime/tests/even_number_preparation.rs b/paddler_agent_runtime/tests/even_number_preparation.rs
new file mode 100644
index 000000000..74265ee98
--- /dev/null
+++ b/paddler_agent_runtime/tests/even_number_preparation.rs
@@ -0,0 +1,25 @@
+use tokio::sync::mpsc::UnboundedSender;
+
+use paddler_agent_runtime::prepares_scheduler_command::PreparesSchedulerCommand;
+
+use crate::number_command::NumberCommand;
+
+pub struct EvenNumberPreparation {
+    pub rejected_tx: UnboundedSender<u32>,
+}
+
+impl PreparesSchedulerCommand for EvenNumberPreparation {
+    type Command = NumberCommand;
+    type Request = u32;
+
+    fn prepare_scheduler_command(
+        &self,
+        _agent_name: Option<&str>,
+        request: Self::Request,
+    ) -> Option<Self::Command> {
+        request.is_multiple_of(2).then(|| NumberCommand {
+            rejected_tx: self.rejected_tx.clone(),
+            value: request,
+        })
+    }
+}
diff --git a/paddler_agent_runtime/tests/inference_runtime_context_fixture.rs b/paddler_agent_runtime/tests/inference_runtime_context_fixture.rs
new file mode 100644
index 000000000..02a312c51
--- /dev/null
+++ b/paddler_agent_runtime/tests/inference_runtime_context_fixture.rs
@@ -0,0 +1,9 @@
+use paddler_agent_runtime::inference_runtime_context::InferenceRuntimeContext;
+use paddler_agent_runtime::inference_runtime_context_params::InferenceRuntimeContextParams;
+
+pub fn inference_runtime_context_fixture() -> InferenceRuntimeContext {
+    InferenceRuntimeContext::new(InferenceRuntimeContextParams {
+        agent_name: Some("agent".to_owned()),
+        slots: 1,
+    })
+}
diff --git a/paddler_agent_runtime/tests/joining_in_the_background_resumes_the_panic_of_the_scheduler_thread.rs b/paddler_agent_runtime/tests/joining_in_the_background_resumes_the_panic_of_the_scheduler_thread.rs
new file mode 100644
index 000000000..e3540342d
--- /dev/null
+++ b/paddler_agent_runtime/tests/joining_in_the_background_resumes_the_panic_of_the_scheduler_thread.rs
@@ -0,0 +1,13 @@
+use std::thread::spawn;
+
+use paddler_agent_runtime::join_scheduler_thread_in_background::join_scheduler_thread_in_background;
+
+#[tokio::test]
+#[should_panic(expected = "the scheduler crashed")]
+async fn joining_in_the_background_resumes_the_panic_of_the_scheduler_thread() {
+    let scheduler_thread_handle = spawn(|| {
+        panic!("the scheduler crashed");
+    });
+
+    join_scheduler_thread_in_background((), scheduler_thread_handle).await;
+}
diff --git a/paddler_agent_runtime/tests/joining_resumes_the_panic_of_the_scheduler_thread.rs b/paddler_agent_runtime/tests/joining_resumes_the_panic_of_the_scheduler_thread.rs
new file mode 100644
index 000000000..62f2e8109
--- /dev/null
+++ b/paddler_agent_runtime/tests/joining_resumes_the_panic_of_the_scheduler_thread.rs
@@ -0,0 +1,13 @@
+use std::thread::spawn;
+
+use paddler_agent_runtime::join_scheduler_thread::join_scheduler_thread;
+
+#[test]
+#[should_panic(expected = "the scheduler crashed")]
+fn joining_resumes_the_panic_of_the_scheduler_thread() {
+    let scheduler_thread_handle = spawn(|| {
+        panic!("the scheduler crashed");
+    });
+
+    join_scheduler_thread(scheduler_thread_handle);
+}
diff --git a/paddler_agent_runtime/tests/loading_a_model_while_another_backend_runs_reports_that_it_cannot_be_loaded.rs b/paddler_agent_runtime/tests/loading_a_model_while_another_backend_runs_reports_that_it_cannot_be_loaded.rs
new file mode 100644
index 000000000..032eab3f6
--- /dev/null
+++ b/paddler_agent_runtime/tests/loading_a_model_while_another_backend_runs_reports_that_it_cannot_be_loaded.rs
@@ -0,0 +1,40 @@
+use std::mem::discriminant;
+use std::path::PathBuf;
+
+use llama_cpp_bindings::LlamaCppError;
+use llama_cpp_bindings::llama_backend::LlamaBackend;
+
+use paddler_agent_runtime::agent_runtime_error::AgentRuntimeError;
+use paddler_agent_runtime::llama_model_settings::LlamaModelSettings;
+use paddler_agent_runtime::loaded_llama_model::LoadedLlamaModel;
+use paddler_messaging::agent_issue::AgentIssue;
+use paddler_messaging::agent_issue_params::model_path::ModelPath;
+
+use crate::inference_runtime_context_fixture::inference_runtime_context_fixture;
+
+#[test]
+fn loading_a_model_while_another_backend_runs_reports_that_it_cannot_be_loaded() {
+    let inference_runtime_context = inference_runtime_context_fixture();
+    let model_path = PathBuf::from("model-that-is-never-read.gguf");
+    let _running_backend = LlamaBackend::init().expect("the first backend must initialize");
+
+    let load_error = LoadedLlamaModel::load(
+        &inference_runtime_context,
+        LlamaModelSettings {
+            model_path: model_path.clone(),
+            n_gpu_layers: 0,
+        },
+    )
+    .err()
+    .expect("a second backend must not initialize");
+
+    assert_eq!(
+        discriminant(&load_error),
+        discriminant(&AgentRuntimeError::BackendInitializationFailed(
+            LlamaCppError::BackendAlreadyInitialized
+        ))
+    );
+    assert!(inference_runtime_context.slot_aggregated_status.has_issue(
+        &AgentIssue::ModelCannotBeLoaded(ModelPath::from(model_path.as_path()))
+    ));
+}
diff --git a/paddler_agent_runtime/tests/main.rs b/paddler_agent_runtime/tests/main.rs
new file mode 100644
index 000000000..41822721f
--- /dev/null
+++ b/paddler_agent_runtime/tests/main.rs
@@ -0,0 +1,20 @@
+mod a_model_load_requested_during_shutdown_loads_nothing;
+mod a_scheduler_that_fails_before_readiness_reports_its_error;
+mod a_scheduler_that_signals_readiness_starts;
+mod a_scheduler_whose_startup_was_abandoned_never_runs;
+mod a_scheduler_whose_startup_was_delivered_runs;
+mod a_startup_cancelled_while_its_scheduler_signals_readiness_is_cancelled;
+mod commands_for_a_stopped_scheduler_are_rejected;
+#[cfg(test)]
+mod drained_scheduler_messages;
+mod even_number_preparation;
+mod inference_runtime_context_fixture;
+mod joining_in_the_background_resumes_the_panic_of_the_scheduler_thread;
+mod joining_resumes_the_panic_of_the_scheduler_thread;
+mod loading_a_model_while_another_backend_runs_reports_that_it_cannot_be_loaded;
+mod number_command;
+mod number_request_preparer;
+mod prepared_requests_reach_the_scheduler_before_its_shutdown;
+mod reports_ready_request_preparer;
+mod requests_the_preparation_rejects_never_reach_the_scheduler;
+mod scheduler_ran_after_handover;
diff --git a/paddler_agent_runtime/tests/number_command.rs b/paddler_agent_runtime/tests/number_command.rs
new file mode 100644
index 000000000..0dfc0e083
--- /dev/null
+++ b/paddler_agent_runtime/tests/number_command.rs
@@ -0,0 +1,15 @@
+use tokio::sync::mpsc::UnboundedSender;
+
+use paddler_agent_runtime::scheduler_command::SchedulerCommand;
+use paddler_agent_runtime::send_result_or_warn::send_result_or_warn;
+
+pub struct NumberCommand {
+    pub rejected_tx: UnboundedSender<u32>,
+    pub value: u32,
+}
+
+impl SchedulerCommand for NumberCommand {
+    fn reject_because_the_scheduler_stopped(self, agent_name: Option<&str>) {
+        send_result_or_warn(agent_name, &self.rejected_tx, self.value);
+    }
+}
diff --git a/paddler_agent_runtime/tests/number_request_preparer.rs b/paddler_agent_runtime/tests/number_request_preparer.rs
new file mode 100644
index 000000000..bb786ab76
--- /dev/null
+++ b/paddler_agent_runtime/tests/number_request_preparer.rs
@@ -0,0 +1,35 @@
+use std::sync::Arc;
+
+use tokio::sync::mpsc::UnboundedReceiver;
+use tokio::sync::mpsc::unbounded_channel;
+use tokio_util::task::TaskTracker;
+
+use paddler_agent_runtime::scheduler_message::SchedulerMessage;
+use paddler_agent_runtime::scheduler_request_preparer::SchedulerRequestPreparer;
+
+use crate::even_number_preparation::EvenNumberPreparation;
+use crate::number_command::NumberCommand;
+
+pub struct NumberRequestPreparer {
+    pub rejected_rx: UnboundedReceiver<u32>,
+    pub request_preparer: Arc<SchedulerRequestPreparer<EvenNumberPreparation>>,
+    pub scheduler_message_rx: UnboundedReceiver<SchedulerMessage<NumberCommand>>,
+}
+
+impl NumberRequestPreparer {
+    pub fn new() -> Self {
+        let (rejected_tx, rejected_rx) = unbounded_channel();
+        let (scheduler_message_tx, scheduler_message_rx) = unbounded_channel();
+
+        Self {
+            rejected_rx,
+            request_preparer: Arc::new(SchedulerRequestPreparer {
+                agent_name: Some("agent".to_owned()),
+                preparation: EvenNumberPreparation { rejected_tx },
+                preparation_tasks: TaskTracker::new(),
+                scheduler_message_tx,
+            }),
+            scheduler_message_rx,
+        }
+    }
+}
diff --git a/paddler_agent_runtime/tests/prepared_requests_reach_the_scheduler_before_its_shutdown.rs b/paddler_agent_runtime/tests/prepared_requests_reach_the_scheduler_before_its_shutdown.rs
new file mode 100644
index 000000000..83607bfb4
--- /dev/null
+++ b/paddler_agent_runtime/tests/prepared_requests_reach_the_scheduler_before_its_shutdown.rs
@@ -0,0 +1,29 @@
+use paddler_agent_runtime::scheduler_message::SchedulerMessage;
+
+use crate::drained_scheduler_messages::drained_scheduler_messages;
+use crate::number_command::NumberCommand;
+use crate::number_request_preparer::NumberRequestPreparer;
+
+#[tokio::test]
+async fn prepared_requests_reach_the_scheduler_before_its_shutdown() {
+    let NumberRequestPreparer {
+        request_preparer,
+        scheduler_message_rx,
+        ..
+    } = NumberRequestPreparer::new();
+
+    request_preparer.prepare(2);
+    request_preparer
+        .stop_scheduler_after_pending_preparations()
+        .await;
+
+    let received: Vec<Option<u32>> = drained_scheduler_messages(scheduler_message_rx)
+        .into_iter()
+        .map(|scheduler_message| match scheduler_message {
+            SchedulerMessage::Command(NumberCommand { value, .. }) => Some(value),
+            SchedulerMessage::Shutdown => None,
+        })
+        .collect();
+
+    assert_eq!(received, vec![Some(2), None]);
+}
diff --git a/paddler_agent_runtime/tests/reports_ready_request_preparer.rs b/paddler_agent_runtime/tests/reports_ready_request_preparer.rs
new file mode 100644
index 000000000..a45ecab79
--- /dev/null
+++ b/paddler_agent_runtime/tests/reports_ready_request_preparer.rs
@@ -0,0 +1,23 @@
+use std::sync::Arc;
+
+use paddler_agent_runtime::prepares_scheduler_command::PreparesSchedulerCommand;
+use paddler_agent_runtime::scheduler_handle::SchedulerHandle;
+use paddler_agent_runtime::scheduler_request_preparer::SchedulerRequestPreparer;
+use paddler_agent_runtime::startup_outcome::StartupOutcome;
+
+pub trait ReportsReadyRequestPreparer<TPreparation: PreparesSchedulerCommand> {
+    fn ready_request_preparer(&self) -> Option<&Arc<SchedulerRequestPreparer<TPreparation>>>;
+}
+
+impl<TPreparation: PreparesSchedulerCommand> ReportsReadyRequestPreparer<TPreparation>
+    for StartupOutcome<SchedulerHandle<TPreparation>>
+{
+    fn ready_request_preparer(&self) -> Option<&Arc<SchedulerRequestPreparer<TPreparation>>> {
+        match self {
+            Self::Cancelled => None,
+            Self::Ready(SchedulerHandle {
+                request_preparer, ..
+            }) => Some(request_preparer),
+        }
+    }
+}
diff --git a/paddler_agent_runtime/tests/requests_the_preparation_rejects_never_reach_the_scheduler.rs b/paddler_agent_runtime/tests/requests_the_preparation_rejects_never_reach_the_scheduler.rs
new file mode 100644
index 000000000..f22fbfe62
--- /dev/null
+++ b/paddler_agent_runtime/tests/requests_the_preparation_rejects_never_reach_the_scheduler.rs
@@ -0,0 +1,31 @@
+use std::mem::discriminant;
+
+use paddler_agent_runtime::scheduler_message::SchedulerMessage;
+
+use crate::drained_scheduler_messages::drained_scheduler_messages;
+use crate::number_command::NumberCommand;
+use crate::number_request_preparer::NumberRequestPreparer;
+
+#[tokio::test]
+async fn requests_the_preparation_rejects_never_reach_the_scheduler() {
+    let NumberRequestPreparer {
+        request_preparer,
+        scheduler_message_rx,
+        ..
+    } = NumberRequestPreparer::new();
+
+    request_preparer.prepare(1);
+    request_preparer
+        .stop_scheduler_after_pending_preparations()
+        .await;
+
+    let received: Vec<_> = drained_scheduler_messages(scheduler_message_rx)
+        .into_iter()
+        .map(|scheduler_message| discriminant(&scheduler_message))
+        .collect();
+
+    assert_eq!(
+        received,
+        vec![discriminant(&SchedulerMessage::<NumberCommand>::Shutdown)]
+    );
+}
diff --git a/paddler_agent_runtime/tests/scheduler_ran_after_handover.rs b/paddler_agent_runtime/tests/scheduler_ran_after_handover.rs
new file mode 100644
index 000000000..b3d68c0d6
--- /dev/null
+++ b/paddler_agent_runtime/tests/scheduler_ran_after_handover.rs
@@ -0,0 +1,11 @@
+use tokio::sync::mpsc::UnboundedSender;
+
+use paddler_agent_runtime::run_scheduler_unless_startup_abandoned::run_scheduler_unless_startup_abandoned;
+
+pub fn scheduler_ran_after_handover(scheduler_startup_tx: &UnboundedSender<()>) -> bool {
+    let mut scheduler_ran = false;
+
+    run_scheduler_unless_startup_abandoned(scheduler_startup_tx, (), || scheduler_ran = true);
+
+    scheduler_ran
+}
diff --git a/paddler_agent_status/src/agent_issue_fix.rs b/paddler_agent_status/src/agent_issue_fix.rs
index 11753cc26..3fe39aeff 100644
--- a/paddler_agent_status/src/agent_issue_fix.rs
+++ b/paddler_agent_status/src/agent_issue_fix.rs
@@ -1,5 +1,7 @@
 use paddler_messaging::agent_issue::AgentIssue;
+use paddler_messaging::agent_issue_params::hugging_face_download_failed_params::HuggingFaceDownloadFailedParams;
 use paddler_messaging::agent_issue_params::model_path::ModelPath;
+use paddler_messaging::agent_issue_params::pointer_head_incompatible_with_model_params::PointerHeadIncompatibleWithModelParams;
 use paddler_messaging::agent_issue_params::slot_cannot_start_params::SlotCannotStartParams;
 
 #[derive(Debug)]
@@ -7,13 +9,14 @@ pub enum AgentIssueFix {
     ChatTemplateIsCompiled(ModelPath),
     HuggingFaceDownloadedModel(ModelPath),
     HuggingFaceStartedDownloading(ModelPath),
-    ModelChatTemplateIsLoaded(ModelPath),
     ModelFileExists(ModelPath),
     ModelIsLoaded(ModelPath),
     ModelDownloadCompleted(ModelPath),
     ModelDownloadStarted(ModelPath),
-    ModelStateIsReconciled,
     MultimodalProjectionIsLoaded(ModelPath),
+    PointerHeadFileExists(ModelPath),
+    PointerHeadIsCompatibleWithModel(ModelPath),
+    PointerHeadIsLoaded(ModelPath),
     SlotStarted(u32),
 }
 
@@ -25,7 +28,6 @@ impl AgentIssueFix {
                 Self::ChatTemplateIsCompiled(fix_model_path) => {
                     issue_params.model_path.eq(fix_model_path)
                 }
-                Self::ModelStateIsReconciled => true,
                 _ => false,
             },
             AgentIssue::HuggingFaceCannotAcquireLock(hugging_face_download_lock) => match self {
@@ -33,7 +35,16 @@ impl AgentIssueFix {
                 | Self::HuggingFaceStartedDownloading(fix_model_path) => {
                     hugging_face_download_lock.model_path.eq(fix_model_path)
                 }
-                Self::ModelStateIsReconciled => true,
+                _ => false,
+            },
+            AgentIssue::HuggingFaceDownloadFailed(HuggingFaceDownloadFailedParams {
+                failure: _,
+                model_path: issue_model_path,
+            }) => match self {
+                Self::HuggingFaceDownloadedModel(fix_model_path)
+                | Self::HuggingFaceStartedDownloading(fix_model_path) => {
+                    issue_model_path.eq(fix_model_path)
+                }
                 _ => false,
             },
             AgentIssue::HuggingFaceModelDoesNotExist(issue_model_path)
@@ -43,13 +54,20 @@ impl AgentIssueFix {
                 | Self::MultimodalProjectionIsLoaded(fix_model_path) => {
                     issue_model_path.eq(fix_model_path)
                 }
-                Self::ModelStateIsReconciled => true,
                 _ => false,
             },
             AgentIssue::ModelCannotBeLoaded(issue_model_path) => match self {
                 Self::ModelIsLoaded(fix_model_path) => issue_model_path.eq(fix_model_path),
                 _ => false,
             },
+            AgentIssue::LocalModelFileCannotBeChecked(issue_model_path) => match self {
+                Self::ModelFileExists(fix_model_path)
+                | Self::MultimodalProjectionIsLoaded(fix_model_path)
+                | Self::PointerHeadFileExists(fix_model_path) => {
+                    issue_model_path.eq(fix_model_path)
+                }
+                _ => false,
+            },
             AgentIssue::ModelFileDoesNotExist(issue_model_path) => match self {
                 Self::ModelFileExists(fix_model_path)
                 | Self::MultimodalProjectionIsLoaded(fix_model_path) => {
@@ -57,23 +75,41 @@ impl AgentIssueFix {
                 }
                 _ => false,
             },
+            AgentIssue::AgentModelUriIsNotAnAbsoluteLocalPath(_)
+            | AgentIssue::AgentModelUriIsNotCanonical(_)
+            | AgentIssue::HuggingFaceModelUriIsMalformed(_)
+            | AgentIssue::ModelUriIsUnparseable(_)
+            | AgentIssue::SlotsInsufficientForDecisions(_)
+            | AgentIssue::UnableToFindChatTemplate(_) => false,
             AgentIssue::MultimodalProjectionCannotBeLoaded(_) => {
                 matches!(self, Self::MultimodalProjectionIsLoaded(_))
             }
+            AgentIssue::PointerHeadCannotBeLoaded(issue_pointer_head_path) => match self {
+                Self::PointerHeadFileExists(fix_pointer_head_path)
+                | Self::PointerHeadIsLoaded(fix_pointer_head_path) => {
+                    issue_pointer_head_path.eq(fix_pointer_head_path)
+                }
+                _ => false,
+            },
+            AgentIssue::PointerHeadIncompatibleWithModel(
+                PointerHeadIncompatibleWithModelParams {
+                    incompatibility: _,
+                    pointer_head_path: issue_pointer_head_path,
+                },
+            ) => match self {
+                Self::PointerHeadIsCompatibleWithModel(fix_pointer_head_path) => {
+                    issue_pointer_head_path.eq(fix_pointer_head_path)
+                }
+                _ => false,
+            },
             AgentIssue::SlotCannotStart(SlotCannotStartParams {
                 error: _,
+                inference_capability: _,
                 slot_index,
             }) => match self {
                 Self::SlotStarted(started_slot_index) => slot_index == started_slot_index,
                 _ => false,
             },
-            AgentIssue::UnableToFindChatTemplate(issue_model_path) => match self {
-                Self::ModelChatTemplateIsLoaded(fix_model_path) => {
-                    issue_model_path.eq(fix_model_path)
-                }
-                Self::ModelStateIsReconciled => true,
-                _ => false,
-            },
             AgentIssue::CacheCannotAcquireLock(issue_model_path)
             | AgentIssue::CacheDirectoryIsNotWritable(issue_model_path)
             | AgentIssue::CacheStorageIsFull(issue_model_path)
@@ -87,7 +123,6 @@ impl AgentIssueFix {
             | AgentIssue::ModelDoesNotExistAtUrl(issue_model_path) => match self {
                 Self::ModelDownloadCompleted(fix_model_path)
                 | Self::ModelDownloadStarted(fix_model_path) => issue_model_path.eq(fix_model_path),
-                Self::ModelStateIsReconciled => true,
                 _ => false,
             },
         }
@@ -97,10 +132,17 @@ impl AgentIssueFix {
 #[cfg(test)]
 mod tests {
     use paddler_messaging::agent_issue::AgentIssue;
+    use paddler_messaging::agent_issue_params::agent_model_uri_not_canonical_params::AgentModelUriNotCanonicalParams;
     use paddler_messaging::agent_issue_params::chat_template_does_not_compile_params::ChatTemplateDoesNotCompileParams;
+    use paddler_messaging::agent_issue_params::hugging_face_download_failed_params::HuggingFaceDownloadFailedParams;
+    use paddler_messaging::agent_issue_params::hugging_face_download_failure::HuggingFaceDownloadFailure;
     use paddler_messaging::agent_issue_params::hugging_face_download_lock::HuggingFaceDownloadLock;
     use paddler_messaging::agent_issue_params::model_path::ModelPath;
+    use paddler_messaging::agent_issue_params::pointer_head_incompatibility::PointerHeadIncompatibility;
+    use paddler_messaging::agent_issue_params::pointer_head_incompatible_with_model_params::PointerHeadIncompatibleWithModelParams;
     use paddler_messaging::agent_issue_params::slot_cannot_start_params::SlotCannotStartParams;
+    use paddler_messaging::agent_issue_params::slots_insufficient_for_decisions_params::SlotsInsufficientForDecisionsParams;
+    use paddler_messaging::inference_capability::InferenceCapability;
 
     use super::AgentIssueFix;
 
@@ -110,6 +152,142 @@ mod tests {
         }
     }
 
+    #[test]
+    fn configuration_issues_are_cleared_only_with_their_scope() {
+        let issue_path = model_path("model_a");
+        let configuration_issues = [
+            AgentIssue::AgentModelUriIsNotAnAbsoluteLocalPath(issue_path.clone()),
+            AgentIssue::AgentModelUriIsNotCanonical(AgentModelUriNotCanonicalParams {
+                canonical_uri: "agent:///model_a".to_owned(),
+                model_path: issue_path.clone(),
+            }),
+            AgentIssue::HuggingFaceModelUriIsMalformed(issue_path.clone()),
+            AgentIssue::ModelUriIsUnparseable(issue_path.clone()),
+            AgentIssue::SlotsInsufficientForDecisions(SlotsInsufficientForDecisionsParams {
+                desired_slots: 1,
+                required_slots: 2,
+            }),
+            AgentIssue::UnableToFindChatTemplate(issue_path.clone()),
+        ];
+        let targeted_fixes = [
+            AgentIssueFix::ChatTemplateIsCompiled(issue_path.clone()),
+            AgentIssueFix::HuggingFaceDownloadedModel(issue_path.clone()),
+            AgentIssueFix::HuggingFaceStartedDownloading(issue_path.clone()),
+            AgentIssueFix::ModelDownloadCompleted(issue_path.clone()),
+            AgentIssueFix::ModelDownloadStarted(issue_path.clone()),
+            AgentIssueFix::ModelFileExists(issue_path.clone()),
+            AgentIssueFix::ModelIsLoaded(issue_path.clone()),
+            AgentIssueFix::MultimodalProjectionIsLoaded(issue_path.clone()),
+            AgentIssueFix::PointerHeadFileExists(issue_path.clone()),
+            AgentIssueFix::PointerHeadIsCompatibleWithModel(issue_path.clone()),
+            AgentIssueFix::PointerHeadIsLoaded(issue_path),
+            AgentIssueFix::SlotStarted(0),
+        ];
+
+        for issue in &configuration_issues {
+            for fix in &targeted_fixes {
+                assert!(!fix.can_fix(issue), "{fix:?} must not fix {issue:?}");
+            }
+        }
+    }
+
+    #[test]
+    fn pointer_head_file_exists_fixes_only_the_pointer_head_it_names() {
+        let fix = AgentIssueFix::PointerHeadFileExists(model_path("pointer_head_a"));
+
+        assert!(
+            fix.can_fix(&AgentIssue::PointerHeadCannotBeLoaded(model_path(
+                "pointer_head_a"
+            )))
+        );
+        assert!(
+            !fix.can_fix(&AgentIssue::PointerHeadCannotBeLoaded(model_path(
+                "pointer_head_b"
+            )))
+        );
+    }
+
+    #[test]
+    fn pointer_head_is_loaded_fixes_only_the_pointer_head_it_names() {
+        let fix = AgentIssueFix::PointerHeadIsLoaded(model_path("pointer_head_a"));
+
+        assert!(
+            fix.can_fix(&AgentIssue::PointerHeadCannotBeLoaded(model_path(
+                "pointer_head_a"
+            )))
+        );
+        assert!(
+            !fix.can_fix(&AgentIssue::PointerHeadCannotBeLoaded(model_path(
+                "pointer_head_b"
+            )))
+        );
+    }
+
+    #[test]
+    fn pointer_head_is_compatible_with_model_fixes_only_the_incompatibility_of_the_pointer_head_it_names()
+     {
+        let fix = AgentIssueFix::PointerHeadIsCompatibleWithModel(model_path("pointer_head_a"));
+        let incompatibility_of = |pointer_head_path: &str| {
+            AgentIssue::PointerHeadIncompatibleWithModel(PointerHeadIncompatibleWithModelParams {
+                incompatibility: PointerHeadIncompatibility::HiddenSizeMismatch {
+                    model_hidden_size: 1024,
+                    pointer_head_hidden_size: 512,
+                },
+                pointer_head_path: model_path(pointer_head_path),
+            })
+        };
+
+        assert!(fix.can_fix(&incompatibility_of("pointer_head_a")));
+        assert!(!fix.can_fix(&incompatibility_of("pointer_head_b")));
+        assert!(
+            !fix.can_fix(&AgentIssue::PointerHeadCannotBeLoaded(model_path(
+                "pointer_head_a"
+            )))
+        );
+    }
+
+    #[test]
+    fn a_local_model_file_that_cannot_be_checked_is_cleared_once_that_file_is_found() {
+        let found_file_fixes = [
+            AgentIssueFix::ModelFileExists(model_path("model_a")),
+            AgentIssueFix::MultimodalProjectionIsLoaded(model_path("model_a")),
+            AgentIssueFix::PointerHeadFileExists(model_path("model_a")),
+        ];
+
+        for fix in &found_file_fixes {
+            assert!(
+                fix.can_fix(&AgentIssue::LocalModelFileCannotBeChecked(model_path(
+                    "model_a"
+                )))
+            );
+            assert!(
+                !fix.can_fix(&AgentIssue::LocalModelFileCannotBeChecked(model_path(
+                    "model_b"
+                )))
+            );
+        }
+    }
+
+    #[test]
+    fn a_local_model_file_that_cannot_be_checked_is_not_cleared_by_fixes_that_do_not_find_it() {
+        let issue = AgentIssue::LocalModelFileCannotBeChecked(model_path("model_a"));
+        let unrelated_fixes = [
+            AgentIssueFix::ChatTemplateIsCompiled(model_path("model_a")),
+            AgentIssueFix::HuggingFaceDownloadedModel(model_path("model_a")),
+            AgentIssueFix::HuggingFaceStartedDownloading(model_path("model_a")),
+            AgentIssueFix::ModelDownloadCompleted(model_path("model_a")),
+            AgentIssueFix::ModelDownloadStarted(model_path("model_a")),
+            AgentIssueFix::ModelIsLoaded(model_path("model_a")),
+            AgentIssueFix::PointerHeadIsCompatibleWithModel(model_path("model_a")),
+            AgentIssueFix::PointerHeadIsLoaded(model_path("model_a")),
+            AgentIssueFix::SlotStarted(0),
+        ];
+
+        for fix in &unrelated_fixes {
+            assert!(!fix.can_fix(&issue), "{fix:?} must not fix {issue:?}");
+        }
+    }
+
     #[test]
     fn chat_template_is_compiled_fixes_matching_compile_issue() {
         let fix = AgentIssueFix::ChatTemplateIsCompiled(model_path("model_a"));
@@ -134,35 +312,17 @@ mod tests {
         assert!(!fix.can_fix(&issue));
     }
 
-    #[test]
-    fn model_state_is_reconciled_fixes_chat_template_issue() {
-        let fix = AgentIssueFix::ModelStateIsReconciled;
-        let issue = AgentIssue::ChatTemplateDoesNotCompile(ChatTemplateDoesNotCompileParams {
-            error: "error".to_owned(),
-            model_path: model_path("any_model"),
-            template_content: "template".to_owned(),
-        });
-
-        assert!(fix.can_fix(&issue));
-    }
-
-    #[test]
-    fn model_state_is_reconciled_fixes_unable_to_find_chat_template() {
-        let fix = AgentIssueFix::ModelStateIsReconciled;
-        let issue = AgentIssue::UnableToFindChatTemplate(model_path("any_model"));
-
-        assert!(fix.can_fix(&issue));
-    }
-
     #[test]
     fn slot_started_matches_by_slot_index() {
         let fix = AgentIssueFix::SlotStarted(3);
         let matching_issue = AgentIssue::SlotCannotStart(SlotCannotStartParams {
             error: "failed".to_owned(),
+            inference_capability: InferenceCapability::TextGeneration,
             slot_index: 3,
         });
         let non_matching_issue = AgentIssue::SlotCannotStart(SlotCannotStartParams {
             error: "failed".to_owned(),
+            inference_capability: InferenceCapability::Decisions,
             slot_index: 5,
         });
 
@@ -170,14 +330,6 @@ mod tests {
         assert!(!fix.can_fix(&non_matching_issue));
     }
 
-    #[test]
-    fn model_is_loaded_does_not_fix_unrelated_issue() {
-        let fix = AgentIssueFix::ModelIsLoaded(model_path("model_a"));
-        let issue = AgentIssue::UnableToFindChatTemplate(model_path("model_a"));
-
-        assert!(!fix.can_fix(&issue));
-    }
-
     #[test]
     fn model_is_loaded_fixes_model_cannot_be_loaded_with_same_path() {
         let fix = AgentIssueFix::ModelIsLoaded(model_path("model_a"));
@@ -277,14 +429,6 @@ mod tests {
         assert!(!fix.can_fix(&issue));
     }
 
-    #[test]
-    fn model_state_is_reconciled_fixes_model_cache_is_corrupted() {
-        let fix = AgentIssueFix::ModelStateIsReconciled;
-        let issue = AgentIssue::ModelCacheIsCorrupted(model_path("https://example.com/m.gguf"));
-
-        assert!(fix.can_fix(&issue));
-    }
-
     #[test]
     fn model_download_started_fixes_cache_cannot_acquire_lock() {
         let fix = AgentIssueFix::ModelDownloadStarted(model_path("https://example.com/m.gguf"));
@@ -333,20 +477,35 @@ mod tests {
         assert!(
             AgentIssueFix::HuggingFaceStartedDownloading(model_path("model_a")).can_fix(&issue)
         );
-        assert!(AgentIssueFix::ModelStateIsReconciled.can_fix(&issue));
         assert!(
             !AgentIssueFix::HuggingFaceStartedDownloading(model_path("model_b")).can_fix(&issue)
         );
         assert!(!AgentIssueFix::ModelIsLoaded(model_path("model_a")).can_fix(&issue));
     }
 
+    #[test]
+    fn a_failed_hugging_face_download_is_fixed_only_by_downloading_the_same_model() {
+        let issue = AgentIssue::HuggingFaceDownloadFailed(HuggingFaceDownloadFailedParams {
+            failure: HuggingFaceDownloadFailure::HubConnectionFailed,
+            model_path: model_path("model_a"),
+        });
+
+        assert!(
+            AgentIssueFix::HuggingFaceStartedDownloading(model_path("model_a")).can_fix(&issue)
+        );
+        assert!(AgentIssueFix::HuggingFaceDownloadedModel(model_path("model_a")).can_fix(&issue));
+        assert!(
+            !AgentIssueFix::HuggingFaceStartedDownloading(model_path("model_b")).can_fix(&issue)
+        );
+        assert!(!AgentIssueFix::ModelDownloadStarted(model_path("model_a")).can_fix(&issue));
+    }
+
     #[test]
     fn hugging_face_model_does_not_exist_fixes() {
         let issue = AgentIssue::HuggingFaceModelDoesNotExist(model_path("model_a"));
 
         assert!(AgentIssueFix::HuggingFaceDownloadedModel(model_path("model_a")).can_fix(&issue));
         assert!(AgentIssueFix::MultimodalProjectionIsLoaded(model_path("model_a")).can_fix(&issue));
-        assert!(AgentIssueFix::ModelStateIsReconciled.can_fix(&issue));
         assert!(!AgentIssueFix::ModelIsLoaded(model_path("model_a")).can_fix(&issue));
     }
 
@@ -387,20 +546,13 @@ mod tests {
         let fix = AgentIssueFix::ModelIsLoaded(model_path("model_a"));
         let issue = AgentIssue::SlotCannotStart(SlotCannotStartParams {
             error: "failed".to_owned(),
+            inference_capability: InferenceCapability::Embeddings,
             slot_index: 1,
         });
 
         assert!(!fix.can_fix(&issue));
     }
 
-    #[test]
-    fn unable_to_find_chat_template_fixed_by_model_chat_template_loaded() {
-        let issue = AgentIssue::UnableToFindChatTemplate(model_path("model_a"));
-
-        assert!(AgentIssueFix::ModelChatTemplateIsLoaded(model_path("model_a")).can_fix(&issue));
-        assert!(!AgentIssueFix::ModelChatTemplateIsLoaded(model_path("model_b")).can_fix(&issue));
-    }
-
     #[test]
     fn download_server_rejected_request_fixed_by_model_download_started() {
         let fix = AgentIssueFix::ModelDownloadStarted(model_path("https://example.com/m.gguf"));
diff --git a/paddler_agent_status/src/dispenses_slots.rs b/paddler_agent_status/src/dispenses_slots.rs
index 399e4bbf1..a0c5f8380 100644
--- a/paddler_agent_status/src/dispenses_slots.rs
+++ b/paddler_agent_status/src/dispenses_slots.rs
@@ -1,4 +1,6 @@
+use std::num::NonZeroU16;
+
 pub trait DispensesSlots {
-    fn release_slot(&self);
-    fn take_slot(&self);
+    fn release_slots(&self, slots: NonZeroU16);
+    fn take_slots(&self, slots: NonZeroU16);
 }
diff --git a/paddler_agent_status/src/lib.rs b/paddler_agent_status/src/lib.rs
index fa3f9075c..a335c090c 100644
--- a/paddler_agent_status/src/lib.rs
+++ b/paddler_agent_status/src/lib.rs
@@ -1,3 +1,5 @@
 pub mod agent_issue_fix;
 pub mod dispenses_slots;
+pub mod serving_report;
 pub mod slot_aggregated_status;
+pub mod slot_guard;
diff --git a/paddler_agent_status/src/serving_report.rs b/paddler_agent_status/src/serving_report.rs
new file mode 100644
index 000000000..f69a21458
--- /dev/null
+++ b/paddler_agent_status/src/serving_report.rs
@@ -0,0 +1,11 @@
+use std::collections::BTreeSet;
+
+use paddler_messaging::agent_decisions_serving::AgentDecisionsServing;
+use paddler_messaging::agent_issue::AgentIssue;
+
+pub struct ServingReport {
+    pub decisions: AgentDecisionsServing,
+    pub model_path: String,
+    pub serving_issues: BTreeSet<AgentIssue>,
+    pub uses_chat_template_override: bool,
+}
diff --git a/paddler_agent_status/src/slot_aggregated_status.rs b/paddler_agent_status/src/slot_aggregated_status.rs
index 45627cce0..c5d31848b 100644
--- a/paddler_agent_status/src/slot_aggregated_status.rs
+++ b/paddler_agent_status/src/slot_aggregated_status.rs
@@ -1,12 +1,15 @@
+use std::collections::BTreeSet;
+use std::num::NonZeroU16;
 use std::sync::atomic::AtomicBool;
 use std::sync::atomic::AtomicU64;
 
 use dashmap::DashSet;
 use parking_lot::RwLock;
+use tokio::sync::Notify;
 use tokio::sync::watch;
-use tokio::sync::watch::error::RecvError;
 
 use paddler_messaging::agent_issue::AgentIssue;
+use paddler_messaging::agent_runtime_status::AgentRuntimeStatus;
 use paddler_messaging::agent_state_application_status::AgentStateApplicationStatus;
 use paddler_messaging::agent_status::AgentStatus;
 use paddler_messaging::atomic_value::AtomicValue;
@@ -17,14 +20,18 @@ use paddler_messaging::subscribes_to_updates::SubscribesToUpdates;
 
 use crate::agent_issue_fix::AgentIssueFix;
 use crate::dispenses_slots::DispensesSlots;
+use crate::serving_report::ServingReport;
 
 pub struct SlotAggregatedStatus {
     pub desired_slots_total: u16,
+    desired_state_issues: DashSet<AgentIssue>,
     download_status: RwLock<ModelDownloadStatus>,
-    issues: DashSet<AgentIssue>,
     model_path: RwLock<Option<String>>,
-    slots_processing: AtomicValue<AtomicU64>,
-    slots_total: AtomicValue<AtomicU64>,
+    pipeline_issues: DashSet<AgentIssue>,
+    runtime: RwLock<AgentRuntimeStatus>,
+    serving_issues: RwLock<BTreeSet<AgentIssue>>,
+    slots_processing: RwLock<u64>,
+    slots_released: Notify,
     state_application_status: RwLock<AgentStateApplicationStatus>,
     update_tx: watch::Sender<()>,
     uses_chat_template_override: AtomicValue<AtomicBool>,
@@ -38,12 +45,15 @@ impl SlotAggregatedStatus {
 
         Self {
             desired_slots_total,
+            desired_state_issues: DashSet::new(),
             download_status: RwLock::new(ModelDownloadStatus::NotDownloading),
-            issues: DashSet::new(),
             model_path: RwLock::new(None),
+            pipeline_issues: DashSet::new(),
+            runtime: RwLock::new(AgentRuntimeStatus::Idle),
+            serving_issues: RwLock::new(BTreeSet::new()),
             state_application_status: RwLock::new(AgentStateApplicationStatus::Fresh),
-            slots_processing: AtomicValue::<AtomicU64>::new(0),
-            slots_total: AtomicValue::<AtomicU64>::new(0),
+            slots_processing: RwLock::new(0),
+            slots_released: Notify::new(),
             update_tx,
             uses_chat_template_override: AtomicValue::<AtomicBool>::new(false),
             version: AtomicValue::<AtomicU64>::new(0),
@@ -55,26 +65,25 @@ impl SlotAggregatedStatus {
         self.update_tx.send_replace(());
     }
 
-    pub fn decrement_total_slots(&self) {
-        self.slots_total.decrement();
-        self.announce_change();
-    }
-
     pub fn get_state_application_status(&self) -> AgentStateApplicationStatus {
         *self.state_application_status.read()
     }
 
     pub fn has_issue(&self, issue: &AgentIssue) -> bool {
-        self.issues.contains(issue)
+        self.desired_state_issues.contains(issue)
+            || self.pipeline_issues.contains(issue)
+            || self.serving_issues.read().contains(issue)
     }
 
     pub fn has_issue_like<TFunction>(&self, issue_like: TFunction) -> bool
     where
         TFunction: Fn(&AgentIssue) -> bool,
     {
-        self.issues
+        self.desired_state_issues
             .iter()
+            .chain(self.pipeline_issues.iter())
             .any(|ref_multi| issue_like(ref_multi.key()))
+            || self.serving_issues.read().iter().any(issue_like)
     }
 
     pub fn add_downloaded_bytes(&self, added_bytes: u64) {
@@ -84,37 +93,85 @@ impl SlotAggregatedStatus {
         self.announce_change();
     }
 
-    pub fn fail_download(&self, issue: AgentIssue) {
+    pub fn set_downloaded_bytes(&self, downloaded_bytes: u64) {
+        self.download_status
+            .write()
+            .set_downloaded_bytes(downloaded_bytes);
+        self.announce_change();
+    }
+
+    pub fn begin_desired_state(&self) {
+        self.desired_state_issues.clear();
         *self.download_status.write() = ModelDownloadStatus::NotDownloading;
-        self.issues.insert(issue);
         self.announce_change();
     }
 
-    pub fn increment_total_slots(&self) {
-        self.slots_total.increment();
+    pub fn begin_pipeline(&self) {
+        self.pipeline_issues.clear();
+        *self.state_application_status.write() = AgentStateApplicationStatus::Fresh;
+        self.announce_change();
+    }
+
+    pub fn retry_pipeline(&self) {
+        {
+            let mut state_application_status = self.state_application_status.write();
+
+            *state_application_status = match *state_application_status {
+                AgentStateApplicationStatus::AttemptedAndRetrying
+                | AgentStateApplicationStatus::Stuck => AgentStateApplicationStatus::Stuck,
+                AgentStateApplicationStatus::Applied | AgentStateApplicationStatus::Fresh => {
+                    AgentStateApplicationStatus::AttemptedAndRetrying
+                }
+            };
+        }
+
+        self.announce_change();
+    }
+
+    pub fn fail_download(&self, issue: AgentIssue) {
+        *self.download_status.write() = ModelDownloadStatus::NotDownloading;
+        self.desired_state_issues.insert(issue);
         self.announce_change();
     }
 
-    pub fn register_issue(&self, issue: AgentIssue) {
-        if self.issues.insert(issue) {
+    pub fn register_desired_state_issue(&self, issue: AgentIssue) {
+        if self.desired_state_issues.insert(issue) {
+            self.announce_change();
+        }
+    }
+
+    pub fn register_pipeline_issue(&self, issue: AgentIssue) {
+        if self.pipeline_issues.insert(issue) {
             self.announce_change();
         }
     }
 
     pub fn register_fix(&self, fix: &AgentIssueFix) {
-        let size_before = self.issues.len();
+        let issues_before = self.desired_state_issues.len() + self.pipeline_issues.len();
 
-        self.issues.retain(|issue| !fix.can_fix(issue));
+        self.desired_state_issues
+            .retain(|issue| !fix.can_fix(issue));
+        self.pipeline_issues.retain(|issue| !fix.can_fix(issue));
 
-        if self.issues.len() < size_before {
+        if self.desired_state_issues.len() + self.pipeline_issues.len() < issues_before {
             self.announce_change();
         }
     }
 
-    pub fn reset(&self) {
-        self.issues.clear();
-        self.set_model_path(None);
-        self.slots_total.reset();
+    pub fn serve_nothing(&self, uses_chat_template_override: bool) {
+        *self.model_path.write() = None;
+        *self.runtime.write() = AgentRuntimeStatus::Idle;
+        self.serving_issues.write().clear();
+        *self.state_application_status.write() = AgentStateApplicationStatus::Applied;
+        self.uses_chat_template_override
+            .set(uses_chat_template_override);
+        self.announce_change();
+    }
+
+    pub fn withdraw_serving(&self) {
+        *self.model_path.write() = None;
+        *self.runtime.write() = AgentRuntimeStatus::Idle;
+        self.serving_issues.write().clear();
         self.announce_change();
     }
 
@@ -123,48 +180,56 @@ impl SlotAggregatedStatus {
         self.announce_change();
     }
 
-    pub fn set_model_path(&self, model_path: Option<String>) {
-        {
-            let mut path_lock = self.model_path.write();
-
-            *path_lock = model_path;
+    pub fn start_serving(
+        &self,
+        ServingReport {
+            decisions,
+            model_path,
+            serving_issues,
+            uses_chat_template_override,
+        }: ServingReport,
+    ) {
+        for slot_index in 0..u32::from(self.desired_slots_total) {
+            self.register_fix(&AgentIssueFix::SlotStarted(slot_index));
         }
 
+        *self.model_path.write() = Some(model_path);
+        *self.runtime.write() = AgentRuntimeStatus::Serving {
+            decisions,
+            slots_total: u64::from(self.desired_slots_total),
+        };
+        *self.serving_issues.write() = serving_issues;
+        *self.state_application_status.write() = AgentStateApplicationStatus::Applied;
+        self.uses_chat_template_override
+            .set(uses_chat_template_override);
         self.announce_change();
     }
 
-    pub fn set_state_application_status(&self, status: AgentStateApplicationStatus) {
-        *self.state_application_status.write() = status;
-        self.announce_change();
+    pub fn slots_processing_count(&self) -> u64 {
+        *self.slots_processing.read()
     }
 
-    pub fn set_uses_chat_template_override(&self, uses: bool) {
-        self.uses_chat_template_override.set(uses);
-        self.announce_change();
-    }
+    pub async fn wait_until_no_slots_are_processing(&self) {
+        loop {
+            let slots_released = self.slots_released.notified();
 
-    pub fn slots_processing_count(&self) -> u64 {
-        self.slots_processing.get()
-    }
+            if self.slots_processing_count() == 0 {
+                return;
+            }
 
-    pub async fn wait_until_no_slots_are_processing(&self) -> Result<(), RecvError> {
-        self.update_tx
-            .subscribe()
-            .wait_for(|()| self.slots_processing.get() == 0)
-            .await
-            .map(drop)
+            slots_released.await;
+        }
     }
 }
 
 impl DispensesSlots for SlotAggregatedStatus {
-    fn release_slot(&self) {
-        self.slots_processing.decrement();
-        self.announce_change();
+    fn release_slots(&self, slots: NonZeroU16) {
+        *self.slots_processing.write() -= u64::from(slots.get());
+        self.slots_released.notify_waiters();
     }
 
-    fn take_slot(&self) {
-        self.slots_processing.increment();
-        self.announce_change();
+    fn take_slots(&self, slots: NonZeroU16) {
+        *self.slots_processing.write() += u64::from(slots.get());
     }
 }
 
@@ -182,9 +247,15 @@ impl ProducesSnapshot for SlotAggregatedStatus {
             status: AgentStatus {
                 desired_slots_total: self.desired_slots_total,
                 download_status: self.download_status.read().clone(),
-                issues: self.issues.iter().map(|item| item.clone()).collect(),
+                issues: self
+                    .desired_state_issues
+                    .iter()
+                    .chain(self.pipeline_issues.iter())
+                    .map(|issue| issue.clone())
+                    .chain(self.serving_issues.read().iter().cloned())
+                    .collect(),
                 model_path: self.model_path.read().clone(),
-                slots_total: self.slots_total.get(),
+                runtime: *self.runtime.read(),
                 state_application_status: self.get_state_application_status(),
                 uses_chat_template_override: self.uses_chat_template_override.get(),
             },
@@ -196,14 +267,18 @@ impl ProducesSnapshot for SlotAggregatedStatus {
 #[cfg(test)]
 mod tests {
     use std::collections::BTreeSet;
+    use std::num::NonZeroU16;
     use std::task::Poll;
 
     use tokio_test::task::spawn;
 
+    use paddler_messaging::agent_decisions_serving::AgentDecisionsServing;
     use paddler_messaging::agent_issue::AgentIssue;
     use paddler_messaging::agent_issue_params::model_path::ModelPath;
     use paddler_messaging::agent_issue_params::slot_cannot_start_params::SlotCannotStartParams;
+    use paddler_messaging::agent_runtime_status::AgentRuntimeStatus;
     use paddler_messaging::agent_state_application_status::AgentStateApplicationStatus;
+    use paddler_messaging::inference_capability::InferenceCapability;
     use paddler_messaging::model_download_status::ModelDownloadStatus;
     use paddler_messaging::produces_snapshot::ProducesSnapshot;
     use paddler_messaging::subscribes_to_updates::SubscribesToUpdates;
@@ -211,37 +286,60 @@ mod tests {
     use super::SlotAggregatedStatus;
     use crate::agent_issue_fix::AgentIssueFix;
     use crate::dispenses_slots::DispensesSlots;
+    use crate::serving_report::ServingReport;
+
+    fn serving_report(decisions: AgentDecisionsServing) -> ServingReport {
+        ServingReport {
+            decisions,
+            model_path: "model".to_owned(),
+            serving_issues: BTreeSet::new(),
+            uses_chat_template_override: false,
+        }
+    }
+
+    fn serving_without_a_chat_template() -> ServingReport {
+        ServingReport {
+            serving_issues: BTreeSet::from([AgentIssue::UnableToFindChatTemplate(model_path(
+                "model",
+            ))]),
+            ..serving_report(AgentDecisionsServing::NotServed)
+        }
+    }
+
+    fn model_path(path: &str) -> ModelPath {
+        ModelPath {
+            model_path: path.to_owned(),
+        }
+    }
+
+    fn is_slot_cannot_start(agent_issue: &AgentIssue) -> bool {
+        matches!(agent_issue, AgentIssue::SlotCannotStart(_))
+    }
 
     #[test]
     fn waiting_for_idle_slots_completes_once_the_last_slot_is_released() {
         let status = SlotAggregatedStatus::new(1);
 
-        status.take_slot();
+        status.take_slots(NonZeroU16::MIN);
 
         let mut idle_wait = spawn(status.wait_until_no_slots_are_processing());
 
         assert!(idle_wait.poll().is_pending());
 
-        status.release_slot();
+        status.release_slots(NonZeroU16::MIN);
 
         assert!(idle_wait.is_woken());
-        assert!(matches!(idle_wait.poll(), Poll::Ready(wait_result) if wait_result.is_ok()));
+        assert_eq!(idle_wait.poll(), Poll::Ready(()));
     }
 
     #[test]
-    fn take_slot_notifies_subscribers() {
+    fn taking_a_slot_does_not_announce_a_status_change() {
         let status = SlotAggregatedStatus::new(2);
         let update_rx = status.subscribe_to_updates();
 
-        status.take_slot();
+        status.take_slots(NonZeroU16::MIN);
 
-        assert!(update_rx.has_changed().unwrap());
-    }
-
-    fn model_path(path: &str) -> ModelPath {
-        ModelPath {
-            model_path: path.to_owned(),
-        }
+        assert!(!update_rx.has_changed().unwrap());
     }
 
     #[test]
@@ -251,7 +349,7 @@ mod tests {
 
         assert!(!status.has_issue(&issue));
 
-        status.register_issue(issue.clone());
+        status.register_desired_state_issue(issue.clone());
 
         assert!(status.has_issue(&issue));
     }
@@ -261,26 +359,25 @@ mod tests {
         let status = SlotAggregatedStatus::new(2);
         let issue = AgentIssue::ModelFileDoesNotExist(model_path("model_test"));
 
-        status.register_issue(issue.clone());
+        status.register_desired_state_issue(issue.clone());
         status.register_fix(&AgentIssueFix::ModelFileExists(model_path("model_test")));
 
         assert!(!status.has_issue(&issue));
     }
 
-    fn is_slot_cannot_start(agent_issue: &AgentIssue) -> bool {
-        matches!(agent_issue, AgentIssue::SlotCannotStart(_))
-    }
-
     #[test]
     fn has_issue_like_matches_with_predicate() {
         let status = SlotAggregatedStatus::new(2);
 
-        status.register_issue(AgentIssue::ModelFileDoesNotExist(model_path("model_test")));
+        status.register_desired_state_issue(AgentIssue::ModelFileDoesNotExist(model_path(
+            "model_test",
+        )));
 
         assert!(!status.has_issue_like(is_slot_cannot_start));
 
-        status.register_issue(AgentIssue::SlotCannotStart(SlotCannotStartParams {
+        status.register_pipeline_issue(AgentIssue::SlotCannotStart(SlotCannotStartParams {
             error: "failed".to_owned(),
+            inference_capability: InferenceCapability::Embeddings,
             slot_index: 3,
         }));
 
@@ -292,69 +389,81 @@ mod tests {
     }
 
     #[test]
-    fn snapshot_counts_started_minus_stopped_slots() {
+    fn serving_fixes_the_slot_start_issues_of_every_desired_slot() {
         let status = SlotAggregatedStatus::new(2);
 
-        status.increment_total_slots();
-        status.increment_total_slots();
-
-        assert_eq!(status.make_snapshot().status.slots_total, 2);
+        for slot_index in 0..2 {
+            status.register_pipeline_issue(AgentIssue::SlotCannotStart(SlotCannotStartParams {
+                error: "context creation failed".to_owned(),
+                inference_capability: InferenceCapability::TextGeneration,
+                slot_index,
+            }));
+        }
 
-        status.decrement_total_slots();
+        status.start_serving(serving_report(AgentDecisionsServing::NotServed));
 
-        assert_eq!(status.make_snapshot().status.slots_total, 1);
+        assert!(!status.has_issue_like(is_slot_cannot_start));
     }
 
     #[test]
-    fn version_increments_on_slot_changes() {
+    fn serving_nothing_makes_the_agent_idle() {
         let status = SlotAggregatedStatus::new(2);
 
-        let initial_version = status.make_snapshot().version;
+        status.start_serving(serving_report(AgentDecisionsServing::NotServed));
+        status.serve_nothing(false);
 
-        status.increment_total_slots();
-
-        let updated_version = status.make_snapshot().version;
-        assert!(updated_version > initial_version);
+        assert_eq!(
+            status.make_snapshot().status.runtime,
+            AgentRuntimeStatus::Idle
+        );
     }
 
     #[test]
-    fn snapshot_reports_the_desired_slots_and_the_model_path() {
+    fn starting_to_serve_reports_the_model_its_decisions_and_the_applied_state() {
         let status = SlotAggregatedStatus::new(4);
 
-        status.set_model_path(Some("test_model".to_owned()));
-        status.increment_total_slots();
-        status.increment_total_slots();
+        status.start_serving(serving_report(AgentDecisionsServing::Served));
 
         let snapshot = status.make_snapshot();
 
         assert_eq!(snapshot.status.desired_slots_total, 4);
-        assert_eq!(snapshot.status.model_path, Some("test_model".to_owned()));
-        assert_eq!(snapshot.status.slots_total, 2);
+        assert_eq!(snapshot.status.model_path, Some("model".to_owned()));
+        assert_eq!(
+            snapshot.status.runtime,
+            AgentRuntimeStatus::Serving {
+                decisions: AgentDecisionsServing::Served,
+                slots_total: 4,
+            }
+        );
         assert_eq!(
             snapshot.status.state_application_status,
-            AgentStateApplicationStatus::Fresh
+            AgentStateApplicationStatus::Applied
         );
     }
 
     #[test]
-    fn get_state_application_status_reflects_set_value() {
+    fn a_retried_pipeline_reports_retrying_then_stuck_until_the_next_pipeline() {
         let status = SlotAggregatedStatus::new(2);
+        let mut reported_statuses = vec![status.get_state_application_status()];
 
-        assert_eq!(
-            status.get_state_application_status(),
-            AgentStateApplicationStatus::Fresh
-        );
-
-        status.set_state_application_status(AgentStateApplicationStatus::Applied);
+        status.retry_pipeline();
+        reported_statuses.push(status.get_state_application_status());
+        status.retry_pipeline();
+        reported_statuses.push(status.get_state_application_status());
+        status.retry_pipeline();
+        reported_statuses.push(status.get_state_application_status());
+        status.begin_pipeline();
+        reported_statuses.push(status.make_snapshot().status.state_application_status);
 
         assert_eq!(
-            status.get_state_application_status(),
-            AgentStateApplicationStatus::Applied
-        );
-
-        assert_eq!(
-            status.make_snapshot().status.state_application_status,
-            AgentStateApplicationStatus::Applied
+            reported_statuses,
+            [
+                AgentStateApplicationStatus::Fresh,
+                AgentStateApplicationStatus::AttemptedAndRetrying,
+                AgentStateApplicationStatus::Stuck,
+                AgentStateApplicationStatus::Stuck,
+                AgentStateApplicationStatus::Fresh,
+            ]
         );
     }
 
@@ -363,8 +472,8 @@ mod tests {
         let status = SlotAggregatedStatus::new(2);
         let issue = AgentIssue::ModelFileDoesNotExist(model_path("model_test"));
 
-        status.register_issue(issue.clone());
-        status.register_issue(issue);
+        status.register_desired_state_issue(issue.clone());
+        status.register_desired_state_issue(issue);
 
         assert_eq!(status.make_snapshot().status.issues.len(), 1);
     }
@@ -374,7 +483,7 @@ mod tests {
         let status = SlotAggregatedStatus::new(2);
         let issue = AgentIssue::ModelFileDoesNotExist(model_path("model_test"));
 
-        status.register_issue(issue.clone());
+        status.register_desired_state_issue(issue.clone());
         status.register_fix(&AgentIssueFix::ModelFileExists(model_path("other_model")));
 
         assert!(status.has_issue(&issue));
@@ -386,32 +495,73 @@ mod tests {
 
         assert_eq!(status.slots_processing_count(), 0);
 
-        status.take_slot();
+        status.take_slots(NonZeroU16::MIN);
 
         assert_eq!(status.slots_processing_count(), 1);
 
-        status.release_slot();
+        status.release_slots(NonZeroU16::MIN);
 
         assert_eq!(status.slots_processing_count(), 0);
     }
 
     #[test]
-    fn reset_clears_state_but_keeps_requests_in_flight() {
+    fn serving_nothing_forgets_the_model_but_keeps_requests_in_flight() {
         let status = SlotAggregatedStatus::new(2);
 
-        status.set_model_path(Some("test_model".to_owned()));
-        status.increment_total_slots();
-        status.take_slot();
-        status.register_issue(AgentIssue::ModelFileDoesNotExist(model_path("model_test")));
+        status.start_serving(serving_report(AgentDecisionsServing::NotServed));
+        status.take_slots(NonZeroU16::MIN);
 
-        status.reset();
+        status.serve_nothing(false);
 
         let snapshot = status.make_snapshot();
 
-        assert_eq!(snapshot.status.slots_total, 0);
+        assert_eq!(snapshot.status.runtime, AgentRuntimeStatus::Idle);
         assert_eq!(status.slots_processing_count(), 1);
         assert_eq!(snapshot.status.model_path, None);
-        assert!(snapshot.status.issues.is_empty());
+    }
+
+    #[test]
+    fn beginning_a_desired_state_clears_only_the_desired_state_issues() {
+        let status = SlotAggregatedStatus::new(2);
+        let pipeline_issue = AgentIssue::ModelCannotBeLoaded(model_path("serving_model"));
+
+        status.set_download_status(ModelDownloadStatus::Downloading {
+            downloaded_bytes: 500,
+            model_path: "https://example.com/model.gguf".to_owned(),
+            total_bytes: 1000,
+        });
+        status.register_desired_state_issue(AgentIssue::ModelFileDoesNotExist(model_path(
+            "replaced_model",
+        )));
+        status.register_pipeline_issue(pipeline_issue.clone());
+
+        status.begin_desired_state();
+
+        let snapshot = status.make_snapshot();
+
+        assert_eq!(snapshot.status.issues, BTreeSet::from([pipeline_issue]));
+        assert_eq!(
+            snapshot.status.download_status,
+            ModelDownloadStatus::NotDownloading
+        );
+    }
+
+    #[test]
+    fn beginning_a_pipeline_clears_only_the_pipeline_issues() {
+        let status = SlotAggregatedStatus::new(2);
+        let desired_state_issue = AgentIssue::ModelFileDoesNotExist(model_path("desired_model"));
+
+        status.register_desired_state_issue(desired_state_issue.clone());
+        status.register_pipeline_issue(AgentIssue::ModelCannotBeLoaded(model_path(
+            "replaced_model",
+        )));
+
+        status.begin_pipeline();
+
+        assert_eq!(
+            status.make_snapshot().status.issues,
+            BTreeSet::from([desired_state_issue])
+        );
     }
 
     #[test]
@@ -436,17 +586,68 @@ mod tests {
     }
 
     #[test]
-    fn set_uses_chat_template_override() {
+    fn serving_reports_whether_the_served_state_overrides_the_chat_template() {
         let status = SlotAggregatedStatus::new(2);
 
         assert!(!status.make_snapshot().status.uses_chat_template_override);
 
-        status.set_uses_chat_template_override(true);
+        status.start_serving(ServingReport {
+            uses_chat_template_override: true,
+            ..serving_report(AgentDecisionsServing::NotServed)
+        });
 
         assert!(status.make_snapshot().status.uses_chat_template_override);
 
-        status.set_uses_chat_template_override(false);
+        status.serve_nothing(false);
 
         assert!(!status.make_snapshot().status.uses_chat_template_override);
     }
+
+    #[test]
+    fn serving_issues_stay_reported_while_a_new_pipeline_begins() {
+        let status = SlotAggregatedStatus::new(2);
+        let serving_issue = AgentIssue::UnableToFindChatTemplate(model_path("model"));
+
+        status.start_serving(serving_without_a_chat_template());
+        status.begin_pipeline();
+
+        assert!(status.has_issue(&serving_issue));
+        assert!(status.has_issue_like(|agent_issue| *agent_issue == serving_issue));
+        assert_eq!(
+            status.make_snapshot().status.issues,
+            BTreeSet::from([serving_issue])
+        );
+    }
+
+    #[test]
+    fn serving_again_replaces_the_previous_serving_issues() {
+        let status = SlotAggregatedStatus::new(2);
+
+        status.start_serving(serving_without_a_chat_template());
+        status.start_serving(serving_report(AgentDecisionsServing::NotServed));
+
+        assert_eq!(status.make_snapshot().status.issues, BTreeSet::new());
+    }
+
+    #[test]
+    fn serving_nothing_clears_the_serving_issues() {
+        let status = SlotAggregatedStatus::new(2);
+
+        status.start_serving(serving_without_a_chat_template());
+        status.serve_nothing(false);
+
+        assert_eq!(status.make_snapshot().status.issues, BTreeSet::new());
+    }
+
+    #[test]
+    fn released_slots_return_exactly_the_slots_that_were_taken() {
+        let status = SlotAggregatedStatus::new(4);
+        let two_slots = NonZeroU16::new(2).unwrap();
+
+        status.take_slots(two_slots);
+        status.take_slots(NonZeroU16::MIN);
+        status.release_slots(two_slots);
+
+        assert_eq!(status.slots_processing_count(), 1);
+    }
 }
diff --git a/paddler_agent_status/src/slot_guard.rs b/paddler_agent_status/src/slot_guard.rs
new file mode 100644
index 000000000..a3ab91694
--- /dev/null
+++ b/paddler_agent_status/src/slot_guard.rs
@@ -0,0 +1,56 @@
+use std::num::NonZeroU16;
+use std::sync::Arc;
+
+use crate::dispenses_slots::DispensesSlots as _;
+use crate::slot_aggregated_status::SlotAggregatedStatus;
+
+pub struct SlotGuard {
+    slot_aggregated_status: Arc<SlotAggregatedStatus>,
+    slots: NonZeroU16,
+}
+
+impl SlotGuard {
+    #[must_use]
+    pub fn new(slot_aggregated_status: Arc<SlotAggregatedStatus>, slots: NonZeroU16) -> Self {
+        slot_aggregated_status.take_slots(slots);
+
+        Self {
+            slot_aggregated_status,
+            slots,
+        }
+    }
+
+    #[must_use]
+    pub const fn slots(&self) -> NonZeroU16 {
+        self.slots
+    }
+}
+
+impl Drop for SlotGuard {
+    fn drop(&mut self) {
+        self.slot_aggregated_status.release_slots(self.slots);
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::num::NonZeroU16;
+    use std::sync::Arc;
+
+    use crate::slot_aggregated_status::SlotAggregatedStatus;
+    use crate::slot_guard::SlotGuard;
+
+    #[tokio::test]
+    async fn holds_its_reserved_slots_until_dropped() {
+        let slot_aggregated_status = Arc::new(SlotAggregatedStatus::new(4));
+
+        {
+            let guard = SlotGuard::new(slot_aggregated_status.clone(), NonZeroU16::new(3).unwrap());
+
+            assert_eq!(slot_aggregated_status.slots_processing_count(), 3);
+            assert_eq!(guard.slots(), NonZeroU16::new(3).unwrap());
+        }
+
+        assert_eq!(slot_aggregated_status.slots_processing_count(), 0);
+    }
+}
diff --git a/paddler_agent_text_generation/Cargo.toml b/paddler_agent_text_generation/Cargo.toml
new file mode 100644
index 000000000..7ae094031
--- /dev/null
+++ b/paddler_agent_text_generation/Cargo.toml
@@ -0,0 +1,37 @@
+[package]
+name = "paddler_agent_text_generation"
+authors.workspace = true
+description.workspace = true
+edition.workspace = true
+homepage.workspace = true
+license.workspace = true
+repository.workspace = true
+version.workspace = true
+
+[dependencies]
+encoding_rs = { workspace = true }
+llama-cpp-bindings = { workspace = true }
+log = { workspace = true }
+minijinja = { workspace = true }
+minijinja-contrib = { workspace = true }
+rand = { workspace = true }
+serde_json = { workspace = true }
+thiserror = { workspace = true }
+tokio = { workspace = true }
+tokio-util = { workspace = true }
+paddler_agent_runtime = { workspace = true }
+paddler_agent_status = { workspace = true }
+paddler_image_decoder = { workspace = true }
+paddler_inference_parameters = { workspace = true }
+paddler_messaging = { workspace = true }
+paddler_sha256_hex = { workspace = true }
+paddler_tool_call_validator = { workspace = true }
+
+[features]
+default = []
+cuda = ["llama-cpp-bindings/cuda", "paddler_agent_runtime/cuda"]
+metal = ["llama-cpp-bindings/metal", "paddler_agent_runtime/metal"]
+vulkan = ["llama-cpp-bindings/vulkan", "paddler_agent_runtime/vulkan"]
+
+[lints]
+workspace = true
diff --git a/paddler_agent_text_generation/src/chat_prompt_render_request.rs b/paddler_agent_text_generation/src/chat_prompt_render_request.rs
new file mode 100644
index 000000000..8419b1662
--- /dev/null
+++ b/paddler_agent_text_generation/src/chat_prompt_render_request.rs
@@ -0,0 +1,10 @@
+use serde_json::Value;
+
+use paddler_messaging::chat_template_message::ChatTemplateMessage;
+
+pub struct ChatPromptRenderRequest<'request> {
+    pub add_generation_prompt: bool,
+    pub enable_thinking: bool,
+    pub messages: &'request [ChatTemplateMessage],
+    pub tools: &'request Value,
+}
diff --git a/paddler_agent/src/chat_prompt_renderer.rs b/paddler_agent_text_generation/src/chat_prompt_renderer.rs
similarity index 75%
rename from paddler_agent/src/chat_prompt_renderer.rs
rename to paddler_agent_text_generation/src/chat_prompt_renderer.rs
index 0726825bd..5d3c570cb 100644
--- a/paddler_agent/src/chat_prompt_renderer.rs
+++ b/paddler_agent_text_generation/src/chat_prompt_renderer.rs
@@ -1,14 +1,12 @@
 use minijinja::context;
-
-use paddler_messaging::media_marker::MediaMarker;
+use minijinja::value::Serde;
 
 use crate::chat_prompt_render_request::ChatPromptRenderRequest;
 use crate::chat_template_renderer::ChatTemplateRenderer;
-use crate::generation_request_rejection::GenerationRequestRejection;
+use crate::text_generation_error::TextGenerationError;
 
 pub struct ChatPromptRenderer {
     pub chat_template_renderer: ChatTemplateRenderer,
-    pub media_marker: MediaMarker,
     pub token_bos_str: String,
     pub token_eos_str: String,
     pub token_nl_str: String,
@@ -23,29 +21,31 @@ impl ChatPromptRenderer {
             messages,
             tools,
         }: ChatPromptRenderRequest,
-    ) -> Result<String, GenerationRequestRejection> {
+    ) -> Result<String, TextGenerationError> {
         self.chat_template_renderer
             .render(context! {
                 add_generation_prompt,
-                bos_token => self.token_bos_str,
+                bos_token => self.token_bos_str.as_str(),
                 enable_thinking,
-                eos_token => self.token_eos_str,
-                messages,
-                nl_token => self.token_nl_str,
-                tools,
+                eos_token => self.token_eos_str.as_str(),
+                messages => Serde(messages),
+                nl_token => self.token_nl_str.as_str(),
+                tools => Serde(tools),
             })
-            .map_err(GenerationRequestRejection::ChatTemplateRenderingFailed)
+            .map_err(TextGenerationError::ChatTemplateRenderingFailed)
     }
 }
 
 #[cfg(test)]
 mod tests {
+    use serde_json::Value;
     use std::mem::discriminant;
 
     use minijinja::Error as MinijinjaError;
     use minijinja::ErrorKind as MinijinjaErrorKind;
 
     use paddler_messaging::chat_template::ChatTemplate;
+    use paddler_messaging::chat_template_content_form::ChatTemplateContentForm;
     use paddler_messaging::chat_template_conversation::ChatTemplateConversation;
     use paddler_messaging::conversation_history::ConversationHistory;
     use paddler_messaging::conversation_message::ConversationMessage;
@@ -57,7 +57,7 @@ mod tests {
     use super::ChatPromptRenderer;
     use crate::chat_prompt_render_request::ChatPromptRenderRequest;
     use crate::chat_template_renderer::ChatTemplateRenderer;
-    use crate::generation_request_rejection::GenerationRequestRejection;
+    use crate::text_generation_error::TextGenerationError;
 
     const SPECIAL_TOKENS_TEMPLATE: &str = "{{ bos_token }}{% for message in messages %}{% for part in message.content %}{{ part.text }}{{ nl_token }}{% endfor %}{% endfor %}{{ eos_token }}";
 
@@ -67,15 +67,14 @@ mod tests {
                 content: template_content.to_owned(),
             })
             .unwrap(),
-            media_marker: MediaMarker::new("<media>".to_owned()),
             token_bos_str: "<bos>".to_owned(),
             token_eos_str: "<eos>".to_owned(),
             token_nl_str: "<nl>".to_owned(),
         }
     }
 
-    fn image_then_text_conversation(renderer: &ChatPromptRenderer) -> ChatTemplateConversation {
-        ConversationHistory::new(vec![ConversationMessage {
+    fn image_then_text_conversation() -> ChatTemplateConversation {
+        ConversationHistory::new(vec![ConversationMessage::Message {
             content: ConversationMessageContent::Parts(vec![
                 ConversationMessageContentPart::ImageUrl {
                     image_url: ImageUrl {
@@ -88,17 +87,20 @@ mod tests {
             ]),
             role: "user".to_owned(),
         }])
-        .into_chat_template_conversation(&renderer.media_marker)
+        .into_chat_template_conversation(
+            ChatTemplateContentForm::TypedParts,
+            &MediaMarker::new("<media>".to_owned()),
+        )
     }
 
     fn render_image_then_text(
         renderer: &ChatPromptRenderer,
-    ) -> Result<String, GenerationRequestRejection> {
+    ) -> Result<String, TextGenerationError> {
         renderer.render(ChatPromptRenderRequest {
             add_generation_prompt: true,
             enable_thinking: false,
-            messages: &image_then_text_conversation(renderer).messages,
-            tools: &[],
+            messages: &image_then_text_conversation().messages,
+            tools: &Value::Array(Vec::new()),
         })
     }
 
@@ -121,7 +123,7 @@ mod tests {
                 .err()
                 .map(|rejection| discriminant(&rejection)),
             Some(discriminant(
-                &GenerationRequestRejection::ChatTemplateRenderingFailed(MinijinjaError::from(
+                &TextGenerationError::ChatTemplateRenderingFailed(MinijinjaError::from(
                     MinijinjaErrorKind::InvalidOperation
                 ))
             ))
diff --git a/paddler_agent_text_generation/src/chat_template_renderer/mod.rs b/paddler_agent_text_generation/src/chat_template_renderer/mod.rs
new file mode 100644
index 000000000..b1a59ec7f
--- /dev/null
+++ b/paddler_agent_text_generation/src/chat_template_renderer/mod.rs
@@ -0,0 +1,177 @@
+pub mod pyjinja_tojson;
+pub mod raise_exception;
+
+use minijinja::Environment;
+use minijinja::Error;
+use minijinja::Value;
+use minijinja_contrib::add_to_environment;
+use minijinja_contrib::pycompat::unknown_method_callback;
+
+use paddler_messaging::chat_template::ChatTemplate;
+
+use self::pyjinja_tojson::pyjinja_tojson;
+use self::raise_exception::raise_exception;
+
+const CHAT_TEMPLATE_NAME: &str = "chat_template";
+
+pub struct ChatTemplateRenderer {
+    minijinja_env: Environment<'static>,
+}
+
+impl ChatTemplateRenderer {
+    pub fn new(ChatTemplate { content }: ChatTemplate) -> Result<Self, Error> {
+        let mut minijinja_env = Environment::new();
+
+        minijinja_env.add_function("raise_exception", raise_exception);
+        minijinja_env.add_template_owned(CHAT_TEMPLATE_NAME, content)?;
+        minijinja_env.set_unknown_method_callback(unknown_method_callback);
+
+        add_to_environment(&mut minijinja_env);
+        minijinja_env.add_filter("tojson", pyjinja_tojson);
+
+        Ok(Self { minijinja_env })
+    }
+
+    pub fn render(&self, context: Value) -> Result<String, Error> {
+        self.minijinja_env
+            .get_template(CHAT_TEMPLATE_NAME)
+            .and_then(|template| template.render(context))
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use minijinja::ErrorKind;
+    use minijinja::context;
+    use minijinja::value::Serde;
+
+    use paddler_messaging::chat_template::ChatTemplate;
+    use paddler_messaging::chat_template_message::ChatTemplateMessage;
+    use paddler_messaging::chat_template_message_content::ChatTemplateMessageContent;
+
+    use crate::chat_template_renderer::ChatTemplateRenderer;
+
+    #[test]
+    fn new_with_valid_template_succeeds() {
+        let template = ChatTemplate {
+            content: "Hello {{ name }}!".to_owned(),
+        };
+
+        assert!(ChatTemplateRenderer::new(template).is_ok());
+    }
+
+    #[test]
+    fn new_with_invalid_template_fails() {
+        let template = ChatTemplate {
+            content: "{% if unclosed %}".to_owned(),
+        };
+
+        assert!(
+            ChatTemplateRenderer::new(template)
+                .is_err_and(|compile_error| compile_error.kind() == ErrorKind::SyntaxError)
+        );
+    }
+
+    #[test]
+    fn renders_the_context_values_into_the_template() {
+        let template = ChatTemplate {
+            content: "Hello {{ name }}!".to_owned(),
+        };
+        let renderer = ChatTemplateRenderer::new(template).unwrap();
+
+        let result = renderer.render(context! { name => "world" }).unwrap();
+
+        assert_eq!(result, "Hello world!");
+    }
+
+    #[test]
+    fn renders_messages_loop_with_roles() {
+        let template = ChatTemplate {
+            content: "{% for message in messages %}{{ message.role }}:{{ message.content }}\n{% endfor %}".to_owned(),
+        };
+        let renderer = ChatTemplateRenderer::new(template).unwrap();
+        let messages = vec![
+            ChatTemplateMessage::Message {
+                content: ChatTemplateMessageContent::Text("hi".to_owned()),
+                reasoning_content: None,
+                role: "user".to_owned(),
+            },
+            ChatTemplateMessage::Message {
+                content: ChatTemplateMessageContent::Text("hello".to_owned()),
+                reasoning_content: None,
+                role: "assistant".to_owned(),
+            },
+        ];
+
+        let result = renderer
+            .render(context! { messages => Serde(messages) })
+            .unwrap();
+
+        assert_eq!(result, "user:hi\nassistant:hello\n");
+    }
+
+    #[test]
+    fn add_generation_prompt_branch_changes_output() {
+        let template = ChatTemplate {
+            content: "A{% if add_generation_prompt %}B{% endif %}".to_owned(),
+        };
+        let renderer = ChatTemplateRenderer::new(template).unwrap();
+
+        let with_prompt = renderer
+            .render(context! { add_generation_prompt => true })
+            .unwrap();
+        let without_prompt = renderer
+            .render(context! { add_generation_prompt => false })
+            .unwrap();
+
+        assert_eq!(with_prompt, "AB");
+        assert_eq!(without_prompt, "A");
+    }
+
+    #[test]
+    fn registers_pyjinja_tojson_filter() {
+        let template = ChatTemplate {
+            content: "{{ value | tojson(ensure_ascii=False) }}".to_owned(),
+        };
+        let renderer = ChatTemplateRenderer::new(template).unwrap();
+
+        let result = renderer.render(context! { value => "café" }).unwrap();
+
+        assert_eq!(result, "\"café\"");
+    }
+
+    #[test]
+    fn registers_raise_exception_function() {
+        let template = ChatTemplate {
+            content: "{{ raise_exception('boom') }}".to_owned(),
+        };
+        let template_renderer = ChatTemplateRenderer::new(template).unwrap();
+
+        let render_error = template_renderer
+            .render(context! {})
+            .expect_err("raise_exception must turn rendering into an error");
+
+        assert_eq!(
+            render_error.detail(),
+            Some("Model's chat template raised an exception: 'boom'")
+        );
+    }
+
+    #[test]
+    fn render_fails_when_template_is_not_registered() {
+        let template = ChatTemplate {
+            content: "Hello {{ name }}!".to_owned(),
+        };
+        let mut renderer = ChatTemplateRenderer::new(template).unwrap();
+
+        renderer
+            .minijinja_env
+            .remove_template(super::CHAT_TEMPLATE_NAME);
+
+        let render_error = renderer
+            .render(context! {})
+            .expect_err("rendering must fail when the template is missing");
+
+        assert_eq!(render_error.kind(), ErrorKind::TemplateNotFound);
+    }
+}
diff --git a/paddler_agent/src/chat_template_renderer/pyjinja_tojson.rs b/paddler_agent_text_generation/src/chat_template_renderer/pyjinja_tojson.rs
similarity index 100%
rename from paddler_agent/src/chat_template_renderer/pyjinja_tojson.rs
rename to paddler_agent_text_generation/src/chat_template_renderer/pyjinja_tojson.rs
diff --git a/paddler_agent/src/chat_template_renderer/raise_exception.rs b/paddler_agent_text_generation/src/chat_template_renderer/raise_exception.rs
similarity index 100%
rename from paddler_agent/src/chat_template_renderer/raise_exception.rs
rename to paddler_agent_text_generation/src/chat_template_renderer/raise_exception.rs
diff --git a/paddler_agent_text_generation/src/chat_template_tools.rs b/paddler_agent_text_generation/src/chat_template_tools.rs
new file mode 100644
index 000000000..3c7dc8de2
--- /dev/null
+++ b/paddler_agent_text_generation/src/chat_template_tools.rs
@@ -0,0 +1,104 @@
+use serde_json::Map;
+use serde_json::Value;
+
+use paddler_messaging::request_params::continue_from_conversation_history_params::tool::Tool;
+use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::FunctionCall;
+use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::function::Function;
+use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters::Parameters;
+
+fn openai_function_tool(tool: &Tool<Value>) -> Value {
+    let Tool::Function(FunctionCall {
+        function:
+            Function {
+                name,
+                description,
+                parameters,
+            },
+    }) = tool;
+    let mut function = Map::new();
+
+    function.insert("name".to_owned(), Value::String(name.clone()));
+    function.insert("description".to_owned(), Value::String(description.clone()));
+
+    if let Parameters::Schema(schema) = parameters {
+        function.insert("parameters".to_owned(), schema.clone());
+    }
+
+    Value::Object(Map::from_iter([
+        ("type".to_owned(), Value::String("function".to_owned())),
+        ("function".to_owned(), Value::Object(function)),
+    ]))
+}
+
+pub struct ChatTemplateTools {
+    pub json: String,
+    pub tools: Value,
+}
+
+impl ChatTemplateTools {
+    #[must_use]
+    pub fn new(tools: &[Tool<Value>]) -> Self {
+        let tools = Value::Array(tools.iter().map(openai_function_tool).collect());
+
+        Self {
+            json: tools.to_string(),
+            tools,
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use serde_json::Value;
+    use serde_json::json;
+
+    use paddler_messaging::request_params::continue_from_conversation_history_params::tool::Tool;
+    use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::FunctionCall;
+    use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::function::Function;
+    use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters::Parameters;
+
+    use super::ChatTemplateTools;
+
+    fn tool(parameters: Parameters<Value>) -> Tool<Value> {
+        Tool::Function(FunctionCall {
+            function: Function {
+                name: "get_weather".to_owned(),
+                description: "Reports the weather".to_owned(),
+                parameters,
+            },
+        })
+    }
+
+    #[test]
+    fn a_tool_with_parameters_becomes_an_openai_function_with_its_schema() {
+        let schema = json!({"type": "object", "properties": {"location": {"type": "string"}}});
+
+        assert_eq!(
+            ChatTemplateTools::new(&[tool(Parameters::Schema(schema.clone()))]).json,
+            json!([{
+                "type": "function",
+                "function": {
+                    "name": "get_weather",
+                    "description": "Reports the weather",
+                    "parameters": schema,
+                },
+            }])
+            .to_string()
+        );
+    }
+
+    #[test]
+    fn a_tool_without_parameters_becomes_an_openai_function_without_them() {
+        assert_eq!(
+            ChatTemplateTools::new(&[tool(Parameters::Empty)]).json,
+            json!([{
+                "type": "function",
+                "function": {
+                    "name": "get_weather",
+                    "description": "Reports the weather",
+                },
+            }])
+            .to_string()
+        );
+    }
+}
diff --git a/paddler_agent/src/continuous_batch_active_request.rs b/paddler_agent_text_generation/src/continuous_batch_active_request.rs
similarity index 83%
rename from paddler_agent/src/continuous_batch_active_request.rs
rename to paddler_agent_text_generation/src/continuous_batch_active_request.rs
index 984ef678f..d2aa39173 100644
--- a/paddler_agent/src/continuous_batch_active_request.rs
+++ b/paddler_agent_text_generation/src/continuous_batch_active_request.rs
@@ -1,19 +1,19 @@
 use llama_cpp_bindings::SampledTokenClassifier;
 use tokio::sync::mpsc;
 
+use paddler_agent_runtime::sequence_id_guard::SequenceIdGuard;
+use paddler_agent_status::slot_guard::SlotGuard;
 use paddler_messaging::generated_token_result::GeneratedTokenResult;
 
 use crate::continuous_batch_request_state::ContinuousBatchRequestState;
 use crate::continuous_batch_terminal_delivery::ContinuousBatchTerminalDelivery;
 use crate::continuous_batch_terminal_outcome::ContinuousBatchTerminalOutcome;
-use crate::sequence_id_guard::SequenceIdGuard;
-use crate::slot_guard::SlotGuard;
 use crate::token_sampling::TokenSampling;
 use crate::tool_call_handling::ToolCallHandling;
 
-pub struct ContinuousBatchActiveRequest {
+pub struct ContinuousBatchActiveRequest<'model> {
     pub state: ContinuousBatchRequestState,
-    pub token_classifier: SampledTokenClassifier<'static>,
+    pub token_classifier: SampledTokenClassifier<'model>,
     pub token_sampling: TokenSampling,
     pub generated_tokens_tx: mpsc::UnboundedSender<GeneratedTokenResult>,
     pub generate_tokens_stop_rx: mpsc::UnboundedReceiver<()>,
@@ -22,7 +22,7 @@ pub struct ContinuousBatchActiveRequest {
     pub tool_call_handling: ToolCallHandling,
 }
 
-impl ContinuousBatchActiveRequest {
+impl ContinuousBatchActiveRequest<'_> {
     pub fn complete_with_outcome(&mut self, outcome: GeneratedTokenResult) {
         self.state
             .mark_completed(ContinuousBatchTerminalOutcome::EmitToClient(outcome));
diff --git a/paddler_agent/src/continuous_batch_generation_step.rs b/paddler_agent_text_generation/src/continuous_batch_generation_step.rs
similarity index 100%
rename from paddler_agent/src/continuous_batch_generation_step.rs
rename to paddler_agent_text_generation/src/continuous_batch_generation_step.rs
diff --git a/paddler_agent_text_generation/src/continuous_batch_request_phase.rs b/paddler_agent_text_generation/src/continuous_batch_request_phase.rs
new file mode 100644
index 000000000..f25b459e9
--- /dev/null
+++ b/paddler_agent_text_generation/src/continuous_batch_request_phase.rs
@@ -0,0 +1,34 @@
+use crate::continuous_batch_generation_step::ContinuousBatchGenerationStep;
+use crate::continuous_batch_terminal_outcome::ContinuousBatchTerminalOutcome;
+use crate::multimodal_prompt_ingestion::MultimodalPromptIngestion;
+
+#[derive(Debug)]
+pub enum ContinuousBatchRequestPhase {
+    IngestingText,
+    IngestingLeadingBos(MultimodalPromptIngestion),
+    IngestingMultimodal(MultimodalPromptIngestion),
+    Generating(ContinuousBatchGenerationStep),
+    Completed(ContinuousBatchTerminalOutcome),
+}
+
+impl ContinuousBatchRequestPhase {
+    #[must_use]
+    pub fn after_its_text_prompt(self, last_batch_position: i32) -> Self {
+        match self {
+            Self::IngestingLeadingBos(ingestion) => Self::IngestingMultimodal(ingestion),
+            Self::IngestingText
+            | Self::IngestingMultimodal(_)
+            | Self::Generating(_)
+            | Self::Completed(_) => {
+                Self::Generating(ContinuousBatchGenerationStep::ReadyToSample {
+                    batch_position: last_batch_position,
+                })
+            }
+        }
+    }
+
+    #[must_use]
+    pub const fn samples_after_its_text_prompt(&self) -> bool {
+        matches!(self, Self::IngestingText)
+    }
+}
diff --git a/paddler_agent/src/continuous_batch_request_state.rs b/paddler_agent_text_generation/src/continuous_batch_request_state.rs
similarity index 97%
rename from paddler_agent/src/continuous_batch_request_state.rs
rename to paddler_agent_text_generation/src/continuous_batch_request_state.rs
index b06aa9a89..62acf0bf6 100644
--- a/paddler_agent/src/continuous_batch_request_state.rs
+++ b/paddler_agent_text_generation/src/continuous_batch_request_state.rs
@@ -1,3 +1,4 @@
+use std::mem;
 use std::num::NonZeroU32;
 
 use llama_cpp_bindings::SampledToken;
@@ -49,11 +50,8 @@ impl ContinuousBatchRequestState {
         self.current_token_position = next_token_position;
 
         if is_last_chunk {
-            self.phase = ContinuousBatchRequestPhase::Generating(
-                ContinuousBatchGenerationStep::ReadyToSample {
-                    batch_position: last_batch_position,
-                },
-            );
+            self.phase = mem::replace(&mut self.phase, ContinuousBatchRequestPhase::IngestingText)
+                .after_its_text_prompt(last_batch_position);
         }
     }
 
@@ -94,6 +92,7 @@ impl ContinuousBatchRequestState {
             ContinuousBatchRequestPhase::Completed(terminal_outcome) => terminal_outcome,
             ContinuousBatchRequestPhase::Generating(_)
             | ContinuousBatchRequestPhase::IngestingText
+            | ContinuousBatchRequestPhase::IngestingLeadingBos(_)
             | ContinuousBatchRequestPhase::IngestingMultimodal(_) => {
                 ContinuousBatchTerminalOutcome::EmitNothing
             }
diff --git a/paddler_agent/src/continuous_batch_scheduler/advance_generating_phase.rs b/paddler_agent_text_generation/src/continuous_batch_scheduler/advance_generating_phase.rs
similarity index 86%
rename from paddler_agent/src/continuous_batch_scheduler/advance_generating_phase.rs
rename to paddler_agent_text_generation/src/continuous_batch_scheduler/advance_generating_phase.rs
index c00cfe9e4..1cf5db1f0 100644
--- a/paddler_agent/src/continuous_batch_scheduler/advance_generating_phase.rs
+++ b/paddler_agent_text_generation/src/continuous_batch_scheduler/advance_generating_phase.rs
@@ -26,11 +26,11 @@ pub struct AdvanceGeneratingPhase<'context> {
     pub candidates: &'context mut LlamaTokenDataArray,
     pub ingest_outcomes: &'context mut Vec<IngestOutcome>,
     pub llama_context: &'context LlamaContext<'context>,
-    pub scheduler_context: &'context ContinuousBatchSchedulerContext,
+    pub scheduler_context: &'context ContinuousBatchSchedulerContext<'context>,
 }
 
 impl AdvanceGeneratingPhase<'_> {
-    pub fn run(mut self, requests: &mut [ContinuousBatchActiveRequest]) {
+    pub fn run(mut self, requests: &mut [ContinuousBatchActiveRequest<'_>]) {
         for request in requests {
             let outcome = self.advance_one(request);
 
@@ -40,7 +40,7 @@ impl AdvanceGeneratingPhase<'_> {
 
     fn advance_one(
         &mut self,
-        request: &mut ContinuousBatchActiveRequest,
+        request: &mut ContinuousBatchActiveRequest<'_>,
     ) -> Option<AdvanceOutcome> {
         let ContinuousBatchRequestPhase::Generating(ContinuousBatchGenerationStep::ReadyToSample {
             batch_position,
@@ -57,7 +57,7 @@ impl AdvanceGeneratingPhase<'_> {
                 Ok(SamplingOutcome::Token(token)) => token,
                 Ok(SamplingOutcome::AllCandidatesEliminated) => {
                     error!(
-                        "{:?}: sequence {} sampling exhausted candidates",
+                        "{:?}: sequence {:?} sampling exhausted candidates",
                         self.scheduler_context.agent_name,
                         request.sequence_id_guard.sequence_id()
                     );
@@ -69,7 +69,7 @@ impl AdvanceGeneratingPhase<'_> {
                 }
                 Ok(SamplingOutcome::GrammarRejectedModelOutput(message)) => {
                     error!(
-                        "{:?}: sequence {} grammar rejected sampled token: {message}",
+                        "{:?}: sequence {:?} grammar rejected sampled token: {message}",
                         self.scheduler_context.agent_name,
                         request.sequence_id_guard.sequence_id()
                     );
@@ -78,15 +78,10 @@ impl AdvanceGeneratingPhase<'_> {
                     ));
                 }
                 Err(sampling_error) => {
-                    let message = format!("{sampling_error:#}");
-
-                    error!(
-                        "{:?}: sequence {} sampling error: {message}",
-                        self.scheduler_context.agent_name,
-                        request.sequence_id_guard.sequence_id()
-                    );
                     return Some(AdvanceOutcome::Completed(
-                        GeneratedTokenResult::SamplerError(message),
+                        sampling_error.into_generated_token_result(
+                            self.scheduler_context.agent_name.as_deref(),
+                        ),
                     ));
                 }
             };
@@ -101,7 +96,7 @@ impl AdvanceGeneratingPhase<'_> {
             Ok(progress) => progress,
             Err(detokenization_error) => {
                 error!(
-                    "{:?}: sequence {} token classification failed: {detokenization_error}",
+                    "{:?}: sequence {:?} token classification failed: {detokenization_error}",
                     self.scheduler_context.agent_name,
                     request.sequence_id_guard.sequence_id()
                 );
@@ -114,7 +109,7 @@ impl AdvanceGeneratingPhase<'_> {
 
         let completion = CompletionCheckPhase {
             request_state: &request.state,
-            sequence_context_size: self.llama_context.n_ctx_seq(),
+            sequence_context_size: self.scheduler_context.sequence_context_size,
             usage: request.token_classifier.usage(),
         }
         .run(progress);
@@ -151,7 +146,7 @@ impl AdvanceGeneratingPhase<'_> {
 
     fn client_disconnected_while_emitting(
         &mut self,
-        request: &mut ContinuousBatchActiveRequest,
+        request: &mut ContinuousBatchActiveRequest<'_>,
     ) -> bool {
         for ingest_outcome in self.ingest_outcomes.drain(..) {
             let classified =
@@ -173,7 +168,7 @@ impl AdvanceGeneratingPhase<'_> {
                 request.generated_tokens_tx.send(tool_call_result).is_err()
             }) {
                 warn!(
-                    "{:?}: sequence {} client disconnected (receiver dropped)",
+                    "{:?}: sequence {:?} client disconnected (receiver dropped)",
                     self.scheduler_context.agent_name,
                     request.sequence_id_guard.sequence_id()
                 );
@@ -187,14 +182,14 @@ impl AdvanceGeneratingPhase<'_> {
 
     fn complete_after_resolving_tool_calls(
         &self,
-        request: &mut ContinuousBatchActiveRequest,
+        request: &mut ContinuousBatchActiveRequest<'_>,
         finish: GenerationFinish,
     ) -> AdvanceOutcome {
         if let Some(tool_call_result) = request.tool_call_handling.resolve_at_completion()
             && request.generated_tokens_tx.send(tool_call_result).is_err()
         {
             warn!(
-                "{:?}: sequence {} client disconnected (receiver dropped) while resolving tool calls at completion",
+                "{:?}: sequence {:?} client disconnected (receiver dropped) while resolving tool calls at completion",
                 self.scheduler_context.agent_name,
                 request.sequence_id_guard.sequence_id()
             );
@@ -208,7 +203,10 @@ impl AdvanceGeneratingPhase<'_> {
         }))
     }
 
-    fn apply_outcome(request: &mut ContinuousBatchActiveRequest, outcome: Option<AdvanceOutcome>) {
+    fn apply_outcome(
+        request: &mut ContinuousBatchActiveRequest<'_>,
+        outcome: Option<AdvanceOutcome>,
+    ) {
         match outcome {
             None => {}
             Some(AdvanceOutcome::SampledAndStored(token)) => {
diff --git a/paddler_agent/src/continuous_batch_scheduler/advance_outcome.rs b/paddler_agent_text_generation/src/continuous_batch_scheduler/advance_outcome.rs
similarity index 100%
rename from paddler_agent/src/continuous_batch_scheduler/advance_outcome.rs
rename to paddler_agent_text_generation/src/continuous_batch_scheduler/advance_outcome.rs
diff --git a/paddler_agent/src/continuous_batch_scheduler/assemble_batch_phase.rs b/paddler_agent_text_generation/src/continuous_batch_scheduler/assemble_batch_phase.rs
similarity index 87%
rename from paddler_agent/src/continuous_batch_scheduler/assemble_batch_phase.rs
rename to paddler_agent_text_generation/src/continuous_batch_scheduler/assemble_batch_phase.rs
index ebd15008a..8b6ed83c3 100644
--- a/paddler_agent/src/continuous_batch_scheduler/assemble_batch_phase.rs
+++ b/paddler_agent_text_generation/src/continuous_batch_scheduler/assemble_batch_phase.rs
@@ -25,18 +25,17 @@ impl AssembleBatchPhase {
     pub fn run(
         &self,
         pass: &mut BatchPass,
-        requests: &mut [ContinuousBatchActiveRequest],
+        requests: &mut [ContinuousBatchActiveRequest<'_>],
     ) -> Result<(), BatchAddError> {
-        let added = self.fill_generating(pass, requests)?;
+        let added = Self::fill_generating(pass, requests)?;
         pass.contributions.current_batch_token_count += added;
         self.fill_ingesting(pass, requests)?;
         Ok(())
     }
 
     fn fill_generating(
-        &self,
         pass: &mut BatchPass,
-        requests: &[ContinuousBatchActiveRequest],
+        requests: &[ContinuousBatchActiveRequest<'_>],
     ) -> Result<usize, BatchAddError> {
         let mut tokens_added: usize = 0;
 
@@ -48,16 +47,12 @@ impl AssembleBatchPhase {
                 continue;
             };
 
-            if tokens_added >= self.n_batch {
-                break;
-            }
-
             let batch_position = pass.batch.n_tokens();
 
             pass.batch.add(
                 pending_token,
                 request.state.current_token_position,
-                &[request.sequence_id_guard.sequence_id()],
+                &[request.sequence_id_guard.sequence_id().llama_sequence_id()],
                 true,
             )?;
 
@@ -75,16 +70,19 @@ impl AssembleBatchPhase {
     fn fill_ingesting(
         &self,
         pass: &mut BatchPass,
-        requests: &[ContinuousBatchActiveRequest],
+        requests: &[ContinuousBatchActiveRequest<'_>],
     ) -> Result<(), BatchAddError> {
         for (request_index, request) in requests.iter().enumerate() {
             if !matches!(
                 request.state.phase,
                 ContinuousBatchRequestPhase::IngestingText
+                    | ContinuousBatchRequestPhase::IngestingLeadingBos(_)
             ) {
                 continue;
             }
 
+            let samples_after_its_text_prompt = request.state.phase.samples_after_its_text_prompt();
+
             let remaining = request.state.remaining_prompt_tokens();
             let chunk_size = compute_ingesting_chunk_size(
                 remaining.len(),
@@ -112,8 +110,8 @@ impl AssembleBatchPhase {
                 pass.batch.add(
                     &SampledToken::Content(*token),
                     position,
-                    &[request.sequence_id_guard.sequence_id()],
-                    is_last_token_of_prompt,
+                    &[request.sequence_id_guard.sequence_id().llama_sequence_id()],
+                    is_last_token_of_prompt && samples_after_its_text_prompt,
                 )?;
 
                 next_token_position = position + 1;
@@ -138,6 +136,8 @@ impl AssembleBatchPhase {
 mod tests {
     use llama_cpp_bindings::llama_batch::LlamaBatch;
 
+    use paddler_inference_parameters::batch_size::BatchSize;
+
     use super::AssembleBatchPhase;
     use super::compute_ingesting_chunk_size;
     use crate::continuous_batch_active_request::ContinuousBatchActiveRequest;
@@ -146,7 +146,7 @@ mod tests {
     #[test]
     fn run_over_empty_requests_leaves_batch_untouched() {
         let assemble_phase = AssembleBatchPhase { n_batch: 16 };
-        let mut batch = LlamaBatch::new(16, 1).unwrap();
+        let mut batch = LlamaBatch::new(BatchSize::try_from(16).unwrap().capacity(), 1);
         let mut pass = BatchPass::new(&mut batch);
         let mut requests: [ContinuousBatchActiveRequest; 0] = [];
 
diff --git a/paddler_agent/src/continuous_batch_scheduler/batch_pass.rs b/paddler_agent_text_generation/src/continuous_batch_scheduler/batch_pass.rs
similarity index 87%
rename from paddler_agent/src/continuous_batch_scheduler/batch_pass.rs
rename to paddler_agent_text_generation/src/continuous_batch_scheduler/batch_pass.rs
index 485c8dc33..b2fa89812 100644
--- a/paddler_agent/src/continuous_batch_scheduler/batch_pass.rs
+++ b/paddler_agent_text_generation/src/continuous_batch_scheduler/batch_pass.rs
@@ -29,11 +29,13 @@ mod tests {
     use llama_cpp_bindings::llama_batch::LlamaBatch;
     use llama_cpp_bindings::token::LlamaToken;
 
+    use paddler_inference_parameters::batch_size::BatchSize;
+
     use super::BatchPass;
 
     #[test]
     fn new_clears_tokens_left_by_the_previous_pass() {
-        let mut batch = LlamaBatch::new(16, 1).unwrap();
+        let mut batch = LlamaBatch::new(BatchSize::try_from(16).unwrap().capacity(), 1);
 
         batch
             .add(&SampledToken::Content(LlamaToken::new(1)), 0, &[0], true)
diff --git a/paddler_agent/src/continuous_batch_scheduler/classified_token.rs b/paddler_agent_text_generation/src/continuous_batch_scheduler/classified_token.rs
similarity index 100%
rename from paddler_agent/src/continuous_batch_scheduler/classified_token.rs
rename to paddler_agent_text_generation/src/continuous_batch_scheduler/classified_token.rs
diff --git a/paddler_agent/src/continuous_batch_scheduler/commit_phase.rs b/paddler_agent_text_generation/src/continuous_batch_scheduler/commit_phase.rs
similarity index 85%
rename from paddler_agent/src/continuous_batch_scheduler/commit_phase.rs
rename to paddler_agent_text_generation/src/continuous_batch_scheduler/commit_phase.rs
index 1571d4cbc..8c47ef11f 100644
--- a/paddler_agent/src/continuous_batch_scheduler/commit_phase.rs
+++ b/paddler_agent_text_generation/src/continuous_batch_scheduler/commit_phase.rs
@@ -2,11 +2,11 @@ use crate::continuous_batch_active_request::ContinuousBatchActiveRequest;
 use crate::continuous_batch_scheduler::batch_pass::BatchPass;
 use crate::continuous_batch_scheduler::contributions::Contributions;
 
-pub struct CommitPhase<'requests> {
-    pub requests: &'requests mut [ContinuousBatchActiveRequest],
+pub struct CommitPhase<'requests, 'model> {
+    pub requests: &'requests mut [ContinuousBatchActiveRequest<'model>],
 }
 
-impl CommitPhase<'_> {
+impl CommitPhase<'_, '_> {
     pub fn run(
         self,
         BatchPass {
diff --git a/paddler_agent/src/continuous_batch_scheduler/completion_check_outcome.rs b/paddler_agent_text_generation/src/continuous_batch_scheduler/completion_check_outcome.rs
similarity index 100%
rename from paddler_agent/src/continuous_batch_scheduler/completion_check_outcome.rs
rename to paddler_agent_text_generation/src/continuous_batch_scheduler/completion_check_outcome.rs
diff --git a/paddler_agent/src/continuous_batch_scheduler/completion_check_phase.rs b/paddler_agent_text_generation/src/continuous_batch_scheduler/completion_check_phase.rs
similarity index 100%
rename from paddler_agent/src/continuous_batch_scheduler/completion_check_phase.rs
rename to paddler_agent_text_generation/src/continuous_batch_scheduler/completion_check_phase.rs
diff --git a/paddler_agent/src/continuous_batch_scheduler/contributions.rs b/paddler_agent_text_generation/src/continuous_batch_scheduler/contributions.rs
similarity index 100%
rename from paddler_agent/src/continuous_batch_scheduler/contributions.rs
rename to paddler_agent_text_generation/src/continuous_batch_scheduler/contributions.rs
diff --git a/paddler_agent/src/continuous_batch_scheduler/decode_failure_phase.rs b/paddler_agent_text_generation/src/continuous_batch_scheduler/decode_failure_phase.rs
similarity index 87%
rename from paddler_agent/src/continuous_batch_scheduler/decode_failure_phase.rs
rename to paddler_agent_text_generation/src/continuous_batch_scheduler/decode_failure_phase.rs
index 4aeff8fad..54e1c20d6 100644
--- a/paddler_agent/src/continuous_batch_scheduler/decode_failure_phase.rs
+++ b/paddler_agent_text_generation/src/continuous_batch_scheduler/decode_failure_phase.rs
@@ -4,11 +4,11 @@ use crate::continuous_batch_active_request::ContinuousBatchActiveRequest;
 use crate::continuous_batch_scheduler::batch_pass::BatchPass;
 use crate::continuous_batch_scheduler::contributions::Contributions;
 
-pub struct DecodeFailurePhase<'requests> {
-    pub requests: &'requests mut [ContinuousBatchActiveRequest],
+pub struct DecodeFailurePhase<'requests, 'model> {
+    pub requests: &'requests mut [ContinuousBatchActiveRequest<'model>],
 }
 
-impl DecodeFailurePhase<'_> {
+impl DecodeFailurePhase<'_, '_> {
     pub fn run(
         self,
         BatchPass {
diff --git a/paddler_agent/src/continuous_batch_scheduler/emit_token_outcome.rs b/paddler_agent_text_generation/src/continuous_batch_scheduler/emit_token_outcome.rs
similarity index 100%
rename from paddler_agent/src/continuous_batch_scheduler/emit_token_outcome.rs
rename to paddler_agent_text_generation/src/continuous_batch_scheduler/emit_token_outcome.rs
diff --git a/paddler_agent/src/continuous_batch_scheduler/emit_token_phase.rs b/paddler_agent_text_generation/src/continuous_batch_scheduler/emit_token_phase.rs
similarity index 100%
rename from paddler_agent/src/continuous_batch_scheduler/emit_token_phase.rs
rename to paddler_agent_text_generation/src/continuous_batch_scheduler/emit_token_phase.rs
diff --git a/paddler_agent/src/continuous_batch_scheduler/generating_contribution.rs b/paddler_agent_text_generation/src/continuous_batch_scheduler/generating_contribution.rs
similarity index 100%
rename from paddler_agent/src/continuous_batch_scheduler/generating_contribution.rs
rename to paddler_agent_text_generation/src/continuous_batch_scheduler/generating_contribution.rs
diff --git a/paddler_agent/src/continuous_batch_scheduler/ingesting_contribution.rs b/paddler_agent_text_generation/src/continuous_batch_scheduler/ingesting_contribution.rs
similarity index 100%
rename from paddler_agent/src/continuous_batch_scheduler/ingesting_contribution.rs
rename to paddler_agent_text_generation/src/continuous_batch_scheduler/ingesting_contribution.rs
diff --git a/paddler_agent_text_generation/src/continuous_batch_scheduler/mod.rs b/paddler_agent_text_generation/src/continuous_batch_scheduler/mod.rs
new file mode 100644
index 000000000..4be18d76c
--- /dev/null
+++ b/paddler_agent_text_generation/src/continuous_batch_scheduler/mod.rs
@@ -0,0 +1,433 @@
+pub mod advance_generating_phase;
+pub mod advance_outcome;
+pub mod assemble_batch_phase;
+pub mod batch_pass;
+pub mod classified_token;
+pub mod commit_phase;
+pub mod completion_check_outcome;
+pub mod completion_check_phase;
+pub mod contributions;
+pub mod decode_failure_phase;
+pub mod emit_token_outcome;
+pub mod emit_token_phase;
+pub mod generating_contribution;
+pub mod ingesting_contribution;
+pub mod sequence_ordered_insertion_index;
+
+use std::mem;
+use std::slice;
+
+use llama_cpp_bindings::SampledTokenClassifier;
+use llama_cpp_bindings::batch_add_error::BatchAddError;
+use llama_cpp_bindings::ingest_outcome::IngestOutcome;
+use llama_cpp_bindings::llama_batch::LlamaBatch;
+use llama_cpp_bindings::token::data_array::LlamaTokenDataArray;
+use log::debug;
+use log::error;
+use log::info;
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent_runtime::receives_stop_request::ReceivesStopRequest as _;
+use paddler_agent_runtime::resident_llama_context::ResidentLlamaContext;
+use paddler_agent_runtime::scheduler_inbox::SchedulerInbox;
+use paddler_agent_runtime::sequence_id_guard::SequenceIdGuard;
+use paddler_agent_runtime::sequence_id_pool::SequenceIdPool;
+use paddler_messaging::generated_token_result::GeneratedTokenResult;
+use paddler_messaging::generation_finish::GenerationFinish;
+use paddler_messaging::generation_summary::GenerationSummary;
+
+use self::advance_generating_phase::AdvanceGeneratingPhase;
+use self::assemble_batch_phase::AssembleBatchPhase;
+use self::batch_pass::BatchPass;
+use self::commit_phase::CommitPhase;
+use self::decode_failure_phase::DecodeFailurePhase;
+use self::sequence_ordered_insertion_index::sequence_ordered_insertion_index;
+use crate::continuous_batch_active_request::ContinuousBatchActiveRequest;
+use crate::continuous_batch_request_phase::ContinuousBatchRequestPhase;
+use crate::continuous_batch_request_state::ContinuousBatchRequestState;
+use crate::continuous_batch_scheduler_context::ContinuousBatchSchedulerContext;
+use crate::continuous_batch_scheduler_params::ContinuousBatchSchedulerParams;
+use crate::leading_bos_ingestion::LeadingBosIngestion;
+use crate::multimodal_ingestion_progress::MultimodalIngestionProgress;
+use crate::prepared_generation_request::PreparedGenerationRequest;
+use crate::prepared_prompt::PreparedPrompt;
+use crate::text_generation_error::TextGenerationError;
+use crate::token_classification::TokenClassification;
+
+pub struct ContinuousBatchScheduler<'model> {
+    active_requests: Vec<ContinuousBatchActiveRequest<'model>>,
+    agent_shutdown: CancellationToken,
+    batch: LlamaBatch<'static>,
+    candidates: LlamaTokenDataArray,
+    ingest_outcomes: Vec<IngestOutcome>,
+    pub llama_context: ResidentLlamaContext<'model>,
+    scheduler_context: ContinuousBatchSchedulerContext<'model>,
+    scheduler_inbox: SchedulerInbox<PreparedGenerationRequest>,
+    sequence_id_pool: SequenceIdPool,
+}
+
+impl<'model> ContinuousBatchScheduler<'model> {
+    #[must_use]
+    pub fn new(
+        ContinuousBatchSchedulerParams {
+            agent_shutdown,
+            batch,
+            llama_context,
+            scheduler_context,
+            scheduler_message_rx,
+        }: ContinuousBatchSchedulerParams<'model>,
+    ) -> Self {
+        let sequence_id_pool = SequenceIdPool::new(scheduler_context.desired_slots_total);
+
+        Self {
+            active_requests: Vec::new(),
+            agent_shutdown,
+            batch,
+            candidates: LlamaTokenDataArray::new(Vec::new(), false),
+            ingest_outcomes: Vec::new(),
+            llama_context,
+            scheduler_context,
+            scheduler_inbox: SchedulerInbox::new(scheduler_message_rx),
+            sequence_id_pool,
+        }
+    }
+
+    pub fn run(&mut self) {
+        info!(
+            "{:?}: continuous batch scheduler started",
+            self.scheduler_context.agent_name
+        );
+
+        while !self.agent_shutdown.is_cancelled() {
+            self.check_stop_signals();
+            self.remove_completed_requests();
+            self.accept_new_scheduler_messages();
+
+            if self.has_active_requests() {
+                if let Err(batch_add_error) = self.execute_one_iteration() {
+                    self.fail_unfinished_requests(TextGenerationError::BatchAssemblyFailed(
+                        batch_add_error,
+                    ));
+                }
+            } else if self.scheduler_inbox.shutdown_requested {
+                break;
+            } else {
+                self.wait_for_next_scheduler_message();
+            }
+        }
+
+        while !self.active_requests.is_empty() {
+            self.cleanup_completed_request(0);
+        }
+
+        self.scheduler_inbox
+            .reject_waiting_commands(self.scheduler_context.agent_name.as_deref());
+
+        self.llama_context.synchronize();
+        self.llama_context.detach_threadpool();
+
+        info!(
+            "{:?}: continuous batch scheduler stopped",
+            self.scheduler_context.agent_name
+        );
+    }
+
+    fn wait_for_next_scheduler_message(&mut self) {
+        if let Some(request) = self.scheduler_inbox.wait_for_next_command() {
+            self.accept_generation_request(request);
+        }
+    }
+
+    fn accept_new_scheduler_messages(&mut self) {
+        while let Some(request) = self.scheduler_inbox.take_next_waiting_command() {
+            self.accept_generation_request(request);
+        }
+    }
+
+    fn build_token_classifier_for_active_request(
+        &self,
+        TokenClassification {
+            bare_json_tool_calls,
+            streaming_markers,
+        }: TokenClassification,
+    ) -> SampledTokenClassifier<'model> {
+        SampledTokenClassifier::new(
+            self.scheduler_context.model,
+            streaming_markers,
+            bare_json_tool_calls,
+        )
+    }
+
+    fn accept_generation_request(&mut self, request: PreparedGenerationRequest) {
+        let generated_tokens_tx = request.generated_tokens_tx.clone();
+
+        match self.admit_generation_request(request) {
+            Ok(active_request) => {
+                debug!(
+                    "{:?}: accepted generation request on sequence {:?}",
+                    self.scheduler_context.agent_name,
+                    active_request.sequence_id_guard.sequence_id()
+                );
+
+                let insertion_index = sequence_ordered_insertion_index(
+                    self.active_requests
+                        .iter()
+                        .map(|request| request.sequence_id_guard.sequence_id()),
+                    active_request.sequence_id_guard.sequence_id(),
+                );
+
+                self.active_requests.insert(insertion_index, active_request);
+            }
+            Err(rejection) => {
+                rejection.report(
+                    self.scheduler_context.agent_name.as_deref(),
+                    &generated_tokens_tx,
+                );
+            }
+        }
+    }
+
+    fn admit_generation_request(
+        &mut self,
+        PreparedGenerationRequest {
+            generate_tokens_stop_rx,
+            generated_tokens_tx,
+            max_tokens,
+            prompt,
+            slot_guard,
+            token_classification,
+            token_sampling,
+            tool_call_handling,
+        }: PreparedGenerationRequest,
+    ) -> Result<ContinuousBatchActiveRequest<'model>, TextGenerationError> {
+        let sequence_id_guard = SequenceIdGuard::acquire(&self.sequence_id_pool)
+            .ok_or(TextGenerationError::NoSequenceSlotAvailable)?;
+        let mut token_classifier =
+            self.build_token_classifier_for_active_request(token_classification);
+
+        sequence_id_guard
+            .sequence_id()
+            .clear_kv_cache_in(&mut self.llama_context)
+            .map_err(TextGenerationError::KvCacheClearFailed)?;
+
+        let mut state = ContinuousBatchRequestState {
+            current_token_position: 0,
+            last_outcome_section: token_classifier.current_section(),
+            max_tokens,
+            phase: ContinuousBatchRequestPhase::IngestingText,
+            prompt_tokens: Vec::new(),
+            prompt_tokens_ingested: 0,
+            sampled_tokens: 0,
+        };
+
+        state.phase = match prompt {
+            PreparedPrompt::TextTokens(prompt_tokens) => {
+                state.prompt_tokens = prompt_tokens;
+
+                ContinuousBatchRequestPhase::IngestingText
+            }
+            PreparedPrompt::Multimodal(mut ingestion) => match mem::replace(
+                &mut ingestion.leading_bos_ingestion,
+                LeadingBosIngestion::Skip,
+            ) {
+                LeadingBosIngestion::Ingest(bos_token) => {
+                    state.prompt_tokens = vec![bos_token];
+
+                    ContinuousBatchRequestPhase::IngestingLeadingBos(ingestion)
+                }
+                LeadingBosIngestion::Skip => {
+                    ContinuousBatchRequestPhase::IngestingMultimodal(ingestion)
+                }
+            },
+        };
+        token_classifier.record_prompt_tokens(state.prompt_tokens.len() as u64);
+        token_classifier.ingest_prompt_tokens(&state.prompt_tokens);
+        state.last_outcome_section = token_classifier.current_section();
+
+        Ok(ContinuousBatchActiveRequest {
+            state,
+            token_classifier,
+            token_sampling,
+            generated_tokens_tx,
+            generate_tokens_stop_rx,
+            sequence_id_guard,
+            slot_guard,
+            tool_call_handling,
+        })
+    }
+
+    fn ingest_next_multimodal_chunks(&mut self) {
+        for active_request in &mut self.active_requests {
+            let ContinuousBatchRequestPhase::IngestingMultimodal(ingestion) =
+                &mut active_request.state.phase
+            else {
+                continue;
+            };
+
+            match ingestion.ingest_next_chunk(
+                &mut active_request.token_classifier,
+                &self.llama_context,
+                active_request.sequence_id_guard.sequence_id(),
+                active_request.state.current_token_position,
+            ) {
+                Ok(MultimodalIngestionProgress::ChunksRemain { next_position }) => {
+                    active_request.state.current_token_position = next_position;
+                }
+                Ok(MultimodalIngestionProgress::PromptIngested { next_position }) => {
+                    active_request
+                        .state
+                        .begin_generating_after_multimodal_prompt(
+                            next_position,
+                            active_request.token_classifier.current_section(),
+                        );
+                    self.llama_context.mark_logits_initialized(-1);
+
+                    AdvanceGeneratingPhase {
+                        candidates: &mut self.candidates,
+                        ingest_outcomes: &mut self.ingest_outcomes,
+                        llama_context: &self.llama_context,
+                        scheduler_context: &self.scheduler_context,
+                    }
+                    .run(slice::from_mut(active_request));
+                }
+                Err(ingestion_error) => {
+                    active_request.complete_with_outcome(
+                        ingestion_error.into_generated_token_result(
+                            self.scheduler_context.agent_name.as_deref(),
+                        ),
+                    );
+                }
+            }
+        }
+    }
+
+    fn check_stop_signals(&mut self) {
+        for active_request in &mut self.active_requests {
+            if matches!(
+                active_request.state.phase,
+                ContinuousBatchRequestPhase::Completed(_)
+            ) {
+                continue;
+            }
+
+            if active_request.generate_tokens_stop_rx.is_stop_requested() {
+                let summary = GenerationSummary {
+                    finish: GenerationFinish::StopRequested,
+                    usage: *active_request.token_classifier.usage(),
+                };
+
+                active_request.complete_with_outcome(GeneratedTokenResult::Done(summary));
+            }
+        }
+    }
+
+    fn has_active_requests(&self) -> bool {
+        self.active_requests.iter().any(|request| {
+            !matches!(
+                request.state.phase,
+                ContinuousBatchRequestPhase::Completed(_)
+            )
+        })
+    }
+
+    fn fail_unfinished_requests(&mut self, rejection: TextGenerationError) {
+        let outcome =
+            rejection.into_generated_token_result(self.scheduler_context.agent_name.as_deref());
+
+        for active_request in &mut self.active_requests {
+            active_request.complete_with_outcome(outcome.clone());
+        }
+    }
+
+    fn execute_one_iteration(&mut self) -> Result<(), BatchAddError> {
+        self.advance_generating_requests();
+        self.ingest_next_multimodal_chunks();
+
+        let assemble_phase = AssembleBatchPhase {
+            n_batch: self.scheduler_context.n_batch.tokens_usize(),
+        };
+
+        let mut pass = BatchPass::new(&mut self.batch);
+
+        assemble_phase.run(&mut pass, &mut self.active_requests)?;
+
+        if pass.is_empty() {
+            return Ok(());
+        }
+
+        debug!(
+            "{:?}: decoding batch with {} tokens for {} active requests",
+            self.scheduler_context.agent_name,
+            pass.batch.n_tokens(),
+            self.active_requests.len()
+        );
+
+        match self.llama_context.decode(pass.batch) {
+            Ok(()) => {
+                CommitPhase {
+                    requests: &mut self.active_requests,
+                }
+                .run(pass);
+            }
+            Err(decode_error) => {
+                let description = format!(
+                    "{:?}: llama.cpp rejected the batch: {decode_error}",
+                    self.scheduler_context.agent_name
+                );
+
+                error!("{description}");
+                DecodeFailurePhase {
+                    requests: &mut self.active_requests,
+                }
+                .run(pass, &description);
+            }
+        }
+
+        Ok(())
+    }
+
+    fn advance_generating_requests(&mut self) {
+        AdvanceGeneratingPhase {
+            candidates: &mut self.candidates,
+            ingest_outcomes: &mut self.ingest_outcomes,
+            llama_context: &self.llama_context,
+            scheduler_context: &self.scheduler_context,
+        }
+        .run(&mut self.active_requests);
+    }
+
+    fn remove_completed_requests(&mut self) {
+        let mut removal_index = 0;
+
+        while removal_index < self.active_requests.len() {
+            if matches!(
+                self.active_requests[removal_index].state.phase,
+                ContinuousBatchRequestPhase::Completed(_)
+            ) {
+                self.cleanup_completed_request(removal_index);
+            } else {
+                removal_index += 1;
+            }
+        }
+    }
+
+    fn cleanup_completed_request(&mut self, index: usize) {
+        let removed_request = self.active_requests.remove(index);
+        let sequence_id = removed_request.sequence_id_guard.sequence_id();
+        let usage = *removed_request.token_classifier.usage();
+        let terminal_delivery = removed_request.into_terminal_delivery();
+
+        sequence_id.release_kv_cache_in(
+            &mut self.llama_context,
+            self.scheduler_context.agent_name.as_deref(),
+        );
+
+        debug!(
+            "{:?}: cleaned up sequence {sequence_id:?} ({} completion tokens generated)",
+            self.scheduler_context.agent_name,
+            usage.completion_tokens(),
+        );
+
+        terminal_delivery.deliver(self.scheduler_context.agent_name.as_deref());
+    }
+}
diff --git a/paddler_agent_text_generation/src/continuous_batch_scheduler/sequence_ordered_insertion_index.rs b/paddler_agent_text_generation/src/continuous_batch_scheduler/sequence_ordered_insertion_index.rs
new file mode 100644
index 000000000..1ac28eb84
--- /dev/null
+++ b/paddler_agent_text_generation/src/continuous_batch_scheduler/sequence_ordered_insertion_index.rs
@@ -0,0 +1,46 @@
+use paddler_agent_runtime::sequence_id::SequenceId;
+
+pub fn sequence_ordered_insertion_index(
+    active_sequence_ids: impl IntoIterator<Item = SequenceId>,
+    sequence_id: SequenceId,
+) -> usize {
+    active_sequence_ids
+        .into_iter()
+        .take_while(|active_sequence_id| *active_sequence_id < sequence_id)
+        .count()
+}
+
+#[cfg(test)]
+mod tests {
+    use paddler_agent_runtime::sequence_id::SequenceId;
+
+    use super::sequence_ordered_insertion_index;
+
+    fn active_sequence_ids() -> [SequenceId; 3] {
+        [SequenceId(1), SequenceId(4), SequenceId(6)]
+    }
+
+    #[test]
+    fn places_a_sequence_before_every_higher_sequence_id() {
+        assert_eq!(
+            sequence_ordered_insertion_index(active_sequence_ids(), SequenceId(5)),
+            2
+        );
+    }
+
+    #[test]
+    fn places_the_lowest_sequence_id_first() {
+        assert_eq!(
+            sequence_ordered_insertion_index(active_sequence_ids(), SequenceId(0)),
+            0
+        );
+    }
+
+    #[test]
+    fn places_the_highest_sequence_id_last() {
+        assert_eq!(
+            sequence_ordered_insertion_index(active_sequence_ids(), SequenceId(7)),
+            3
+        );
+    }
+}
diff --git a/paddler_agent_text_generation/src/continuous_batch_scheduler_context.rs b/paddler_agent_text_generation/src/continuous_batch_scheduler_context.rs
new file mode 100644
index 000000000..05ffb8345
--- /dev/null
+++ b/paddler_agent_text_generation/src/continuous_batch_scheduler_context.rs
@@ -0,0 +1,11 @@
+use llama_cpp_bindings::model::LlamaModel;
+
+use paddler_inference_parameters::batch_size::BatchSize;
+
+pub struct ContinuousBatchSchedulerContext<'model> {
+    pub agent_name: Option<String>,
+    pub desired_slots_total: u16,
+    pub model: &'model LlamaModel,
+    pub n_batch: BatchSize,
+    pub sequence_context_size: u32,
+}
diff --git a/paddler_agent_text_generation/src/continuous_batch_scheduler_params.rs b/paddler_agent_text_generation/src/continuous_batch_scheduler_params.rs
new file mode 100644
index 000000000..174589df9
--- /dev/null
+++ b/paddler_agent_text_generation/src/continuous_batch_scheduler_params.rs
@@ -0,0 +1,17 @@
+use llama_cpp_bindings::llama_batch::LlamaBatch;
+use tokio::sync::mpsc::UnboundedReceiver;
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent_runtime::resident_llama_context::ResidentLlamaContext;
+use paddler_agent_runtime::scheduler_message::SchedulerMessage;
+
+use crate::continuous_batch_scheduler_context::ContinuousBatchSchedulerContext;
+use crate::prepared_generation_request::PreparedGenerationRequest;
+
+pub struct ContinuousBatchSchedulerParams<'model> {
+    pub agent_shutdown: CancellationToken,
+    pub batch: LlamaBatch<'static>,
+    pub scheduler_message_rx: UnboundedReceiver<SchedulerMessage<PreparedGenerationRequest>>,
+    pub llama_context: ResidentLlamaContext<'model>,
+    pub scheduler_context: ContinuousBatchSchedulerContext<'model>,
+}
diff --git a/paddler_agent/src/continuous_batch_terminal_delivery.rs b/paddler_agent_text_generation/src/continuous_batch_terminal_delivery.rs
similarity index 90%
rename from paddler_agent/src/continuous_batch_terminal_delivery.rs
rename to paddler_agent_text_generation/src/continuous_batch_terminal_delivery.rs
index d2c7653ad..693e43ddd 100644
--- a/paddler_agent/src/continuous_batch_terminal_delivery.rs
+++ b/paddler_agent_text_generation/src/continuous_batch_terminal_delivery.rs
@@ -1,10 +1,10 @@
 use tokio::sync::mpsc;
 
+use paddler_agent_runtime::send_result_or_warn::send_result_or_warn;
+use paddler_agent_runtime::sequence_id_guard::SequenceIdGuard;
 use paddler_messaging::generated_token_result::GeneratedTokenResult;
 
 use crate::continuous_batch_terminal_outcome::ContinuousBatchTerminalOutcome;
-use crate::send_result_or_warn::send_result_or_warn;
-use crate::sequence_id_guard::SequenceIdGuard;
 
 pub struct ContinuousBatchTerminalDelivery {
     generated_tokens_tx: mpsc::UnboundedSender<GeneratedTokenResult>,
@@ -40,12 +40,12 @@ impl ContinuousBatchTerminalDelivery {
 mod tests {
     use tokio::sync::mpsc;
 
+    use paddler_agent_runtime::sequence_id_guard::SequenceIdGuard;
+    use paddler_agent_runtime::sequence_id_pool::SequenceIdPool;
     use paddler_messaging::generated_token_result::GeneratedTokenResult;
 
     use super::ContinuousBatchTerminalDelivery;
     use crate::continuous_batch_terminal_outcome::ContinuousBatchTerminalOutcome;
-    use crate::sequence_id_guard::SequenceIdGuard;
-    use crate::sequence_id_pool::SequenceIdPool;
 
     #[test]
     fn delivering_an_emit_to_client_outcome_sends_the_terminal_result() {
diff --git a/paddler_agent/src/continuous_batch_terminal_outcome.rs b/paddler_agent_text_generation/src/continuous_batch_terminal_outcome.rs
similarity index 100%
rename from paddler_agent/src/continuous_batch_terminal_outcome.rs
rename to paddler_agent_text_generation/src/continuous_batch_terminal_outcome.rs
diff --git a/paddler_agent_text_generation/src/conversation_prompt_support.rs b/paddler_agent_text_generation/src/conversation_prompt_support.rs
new file mode 100644
index 000000000..53f81e351
--- /dev/null
+++ b/paddler_agent_text_generation/src/conversation_prompt_support.rs
@@ -0,0 +1,7 @@
+use crate::chat_prompt_renderer::ChatPromptRenderer;
+use crate::image_input::ImageInput;
+
+pub struct ConversationPromptSupport {
+    pub chat_prompt_renderer: ChatPromptRenderer,
+    pub image_input: ImageInput,
+}
diff --git a/paddler_agent_text_generation/src/conversation_support.rs b/paddler_agent_text_generation/src/conversation_support.rs
new file mode 100644
index 000000000..81279776c
--- /dev/null
+++ b/paddler_agent_text_generation/src/conversation_support.rs
@@ -0,0 +1,74 @@
+use std::collections::BTreeSet;
+
+use paddler_messaging::agent_issue::AgentIssue;
+use paddler_messaging::agent_issue_params::model_path::ModelPath;
+
+use crate::conversation_prompt_support::ConversationPromptSupport;
+use crate::text_generation_error::TextGenerationError;
+
+pub enum ConversationSupport {
+    Available(Box<ConversationPromptSupport>),
+    ChatTemplateMissing(ModelPath),
+}
+
+impl ConversationSupport {
+    pub fn prompt_support(&self) -> Result<&ConversationPromptSupport, TextGenerationError> {
+        match self {
+            Self::Available(conversation_prompt_support) => Ok(conversation_prompt_support),
+            Self::ChatTemplateMissing(model_path) => {
+                Err(TextGenerationError::ConversationRequiresChatTemplate {
+                    model_path: model_path.clone(),
+                })
+            }
+        }
+    }
+
+    #[must_use]
+    pub fn serving_issues(&self) -> BTreeSet<AgentIssue> {
+        match self {
+            Self::Available(_) => BTreeSet::new(),
+            Self::ChatTemplateMissing(model_path) => {
+                BTreeSet::from([AgentIssue::UnableToFindChatTemplate(model_path.clone())])
+            }
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::collections::BTreeSet;
+
+    use paddler_messaging::agent_issue::AgentIssue;
+    use paddler_messaging::agent_issue_params::model_path::ModelPath;
+
+    use super::ConversationSupport;
+
+    fn model_path() -> ModelPath {
+        ModelPath {
+            model_path: "nomic.gguf".to_owned(),
+        }
+    }
+
+    #[test]
+    fn a_model_without_a_chat_template_refuses_conversation_prompts() {
+        assert_eq!(
+            ConversationSupport::ChatTemplateMissing(model_path())
+                .prompt_support()
+                .err()
+                .as_ref()
+                .map(ToString::to_string),
+            Some(
+                "the model nomic.gguf has no chat template, so it cannot continue a conversation"
+                    .to_owned()
+            )
+        );
+    }
+
+    #[test]
+    fn a_model_without_a_chat_template_reports_that_it_cannot_find_one() {
+        assert_eq!(
+            ConversationSupport::ChatTemplateMissing(model_path()).serving_issues(),
+            BTreeSet::from([AgentIssue::UnableToFindChatTemplate(model_path())])
+        );
+    }
+}
diff --git a/paddler_agent/src/converts_to_mtmd_bitmap.rs b/paddler_agent_text_generation/src/converts_to_mtmd_bitmap.rs
similarity index 100%
rename from paddler_agent/src/converts_to_mtmd_bitmap.rs
rename to paddler_agent_text_generation/src/converts_to_mtmd_bitmap.rs
diff --git a/paddler_agent_text_generation/src/generation_request_preparer.rs b/paddler_agent_text_generation/src/generation_request_preparer.rs
new file mode 100644
index 000000000..670d8e872
--- /dev/null
+++ b/paddler_agent_text_generation/src/generation_request_preparer.rs
@@ -0,0 +1,257 @@
+use std::num::NonZeroU32;
+use std::sync::Arc;
+
+use llama_cpp_bindings::BareJsonToolCalls;
+use llama_cpp_bindings::ChatMessageParser;
+use llama_cpp_bindings::StreamingMarkers;
+use llama_cpp_bindings::mtmd::MtmdBitmap;
+use rand::RngExt as _;
+use rand::rng;
+use serde_json::Value;
+
+use paddler_agent_runtime::agent_request::AgentRequest;
+use paddler_agent_runtime::loaded_llama_model::LoadedLlamaModel;
+use paddler_agent_runtime::prepares_scheduler_command::PreparesSchedulerCommand;
+use paddler_image_decoder::decoded_image::DecodedImage;
+use paddler_messaging::chat_template_conversation::ChatTemplateConversation;
+use paddler_messaging::generated_token_result::GeneratedTokenResult;
+use paddler_messaging::grammar_constraint::GrammarConstraint;
+use paddler_messaging::image_url::ImageUrl;
+use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
+use paddler_messaging::request_params::continue_from_conversation_history_params::tool::Tool;
+use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters_schema::validated_parameters_schema::ValidatedParametersSchema;
+use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
+use paddler_tool_call_validator::tool_call_validator::ToolCallValidator;
+
+use crate::chat_prompt_render_request::ChatPromptRenderRequest;
+use crate::chat_template_tools::ChatTemplateTools;
+use crate::conversation_prompt_support::ConversationPromptSupport;
+use crate::conversation_support::ConversationSupport;
+use crate::converts_to_mtmd_bitmap::ConvertsToMtmdBitmap;
+use crate::media_marked_prompt::MediaMarkedPrompt;
+use crate::prepared_generation_request::PreparedGenerationRequest;
+use crate::prepared_prompt::PreparedPrompt;
+use crate::prompt_modality::PromptModality;
+use crate::prompt_tokenizer::PromptTokenizer;
+use crate::request_media_marker::request_media_marker;
+use crate::require_grammar_compatible_with_thinking::require_grammar_compatible_with_thinking;
+use crate::resolve_grammar::resolve_grammar;
+use crate::sampler_chain_factory::SamplerChainFactory;
+use crate::text_generation_error::TextGenerationError;
+use crate::text_generation_request::TextGenerationRequest;
+use crate::token_classification::TokenClassification;
+use crate::token_sampling::TokenSampling;
+use crate::tool_call_handling::ToolCallHandling;
+use crate::tool_call_pipeline::ToolCallPipeline;
+
+pub struct GenerationRequestPreparer {
+    pub conversation_support: ConversationSupport,
+    pub image_resize_to_fit: NonZeroU32,
+    pub loaded_llama_model: LoadedLlamaModel,
+    pub prompt_tokenizer: PromptTokenizer,
+    pub sampler_chain_factory: SamplerChainFactory,
+    pub streaming_markers: Arc<StreamingMarkers>,
+}
+
+impl GenerationRequestPreparer {
+    pub fn prepare_raw_prompt(
+        &self,
+        AgentRequest {
+            params:
+                ContinueFromRawPromptParams {
+                    grammar,
+                    max_tokens,
+                    raw_prompt,
+                },
+            response_tx: generated_tokens_tx,
+            slot_guard,
+            stop_rx: generate_tokens_stop_rx,
+        }: AgentRequest<ContinueFromRawPromptParams, GeneratedTokenResult>,
+    ) -> Result<PreparedGenerationRequest, TextGenerationError> {
+        let prompt = PreparedPrompt::TextTokens(self.prompt_tokenizer.tokenize(&raw_prompt)?);
+        let token_sampling = self.build_token_sampling(grammar)?;
+
+        Ok(PreparedGenerationRequest {
+            generate_tokens_stop_rx,
+            generated_tokens_tx,
+            max_tokens,
+            prompt,
+            slot_guard,
+            token_classification: TokenClassification {
+                bare_json_tool_calls: BareJsonToolCalls::Ignore,
+                streaming_markers: self.streaming_markers.clone(),
+            },
+            token_sampling,
+            tool_call_handling: ToolCallHandling::Streamed,
+        })
+    }
+
+    pub fn prepare_conversation_history(
+        &self,
+        AgentRequest {
+            response_tx: generated_tokens_tx,
+            stop_rx: generate_tokens_stop_rx,
+            params:
+                ContinueFromConversationHistoryParams {
+                    add_generation_prompt,
+                    conversation_history,
+                    enable_thinking,
+                    grammar,
+                    max_tokens,
+                    parse_tool_calls,
+                    tools,
+                },
+            slot_guard,
+        }: AgentRequest<
+            ContinueFromConversationHistoryParams<ValidatedParametersSchema>,
+            GeneratedTokenResult,
+        >,
+    ) -> Result<PreparedGenerationRequest, TextGenerationError> {
+        let ConversationPromptSupport {
+            chat_prompt_renderer,
+            image_input,
+        } = self.conversation_support.prompt_support()?;
+
+        require_grammar_compatible_with_thinking(grammar.as_ref(), enable_thinking)?;
+
+        let tools_with_schema_documents: Vec<Tool<Value>> =
+            tools.into_iter().map(Tool::from).collect();
+        let chat_template_tools = ChatTemplateTools::new(&tools_with_schema_documents);
+        let media_marker = request_media_marker(&conversation_history, &chat_template_tools);
+        let ChatTemplateConversation {
+            image_urls,
+            messages,
+        } = conversation_history.into_chat_template_conversation(
+            image_input.chat_template_content_form(),
+            &media_marker,
+        );
+        let prompt_modality = image_input.prompt_modality_for(&image_urls)?;
+        let raw_prompt = chat_prompt_renderer.render(ChatPromptRenderRequest {
+            add_generation_prompt,
+            enable_thinking,
+            messages: &messages,
+            tools: &chat_template_tools.tools,
+        })?;
+
+        let prompt = match prompt_modality {
+            PromptModality::Multimodal(multimodal_prompt_support) => PreparedPrompt::Multimodal(
+                multimodal_prompt_support.prepare_prompt(
+                    &image_urls
+                        .iter()
+                        .map(|image_url| self.decode_image_bitmap(image_url))
+                        .collect::<Result<Vec<MtmdBitmap>, TextGenerationError>>()?,
+                    &MediaMarkedPrompt {
+                        media_marker: &media_marker,
+                        prompt: &raw_prompt,
+                    },
+                )?,
+            ),
+            PromptModality::TextOnly => {
+                PreparedPrompt::TextTokens(self.prompt_tokenizer.tokenize(&raw_prompt)?)
+            }
+        };
+
+        let token_sampling = self.build_token_sampling(grammar)?;
+        let tool_call_handling = if parse_tool_calls {
+            ToolCallHandling::Parsed(
+                self.build_tool_call_pipeline(&tools_with_schema_documents, &chat_template_tools)?,
+            )
+        } else {
+            ToolCallHandling::Streamed
+        };
+        let bare_json_tool_calls = if tools_with_schema_documents.is_empty() {
+            BareJsonToolCalls::Ignore
+        } else {
+            BareJsonToolCalls::Detect
+        };
+
+        Ok(PreparedGenerationRequest {
+            generate_tokens_stop_rx,
+            generated_tokens_tx,
+            max_tokens,
+            prompt,
+            slot_guard,
+            token_classification: TokenClassification {
+                bare_json_tool_calls,
+                streaming_markers: self.streaming_markers.clone(),
+            },
+            token_sampling,
+            tool_call_handling,
+        })
+    }
+
+    fn generation_command<TParams>(
+        &self,
+        agent_name: Option<&str>,
+        request: AgentRequest<TParams, GeneratedTokenResult>,
+        prepare: fn(
+            &Self,
+            AgentRequest<TParams, GeneratedTokenResult>,
+        ) -> Result<PreparedGenerationRequest, TextGenerationError>,
+    ) -> Option<PreparedGenerationRequest> {
+        let generated_tokens_tx = request.response_tx.clone();
+
+        match prepare(self, request) {
+            Ok(prepared) => Some(prepared),
+            Err(rejection) => {
+                rejection.report(agent_name, &generated_tokens_tx);
+
+                None
+            }
+        }
+    }
+
+    fn decode_image_bitmap(&self, image_url: &ImageUrl) -> Result<MtmdBitmap, TextGenerationError> {
+        DecodedImage::from_data_uri(image_url, self.image_resize_to_fit)
+            .map_err(TextGenerationError::ImageDecodingFailed)?
+            .to_mtmd_bitmap()
+            .map_err(TextGenerationError::ImageBitmapCreationFailed)
+    }
+
+    fn build_token_sampling(
+        &self,
+        grammar: Option<GrammarConstraint>,
+    ) -> Result<TokenSampling, TextGenerationError> {
+        let grammar = resolve_grammar(grammar, &self.loaded_llama_model.model)?;
+        let chain = self
+            .sampler_chain_factory
+            .create(rng().random::<u32>())
+            .map_err(TextGenerationError::SamplerChainCreationFailed)?;
+
+        Ok(TokenSampling { chain, grammar })
+    }
+
+    fn build_tool_call_pipeline(
+        &self,
+        tools: &[Tool<Value>],
+        chat_template_tools: &ChatTemplateTools,
+    ) -> Result<ToolCallPipeline, TextGenerationError> {
+        let validator =
+            ToolCallValidator::from_tools(tools).map_err(TextGenerationError::ToolSchemaInvalid)?;
+        let chat_message_parser =
+            ChatMessageParser::new(&self.loaded_llama_model.model, &chat_template_tools.json)
+                .map_err(TextGenerationError::ToolCallParserCreationFailed)?;
+
+        Ok(ToolCallPipeline::new(chat_message_parser, validator))
+    }
+}
+
+impl PreparesSchedulerCommand for GenerationRequestPreparer {
+    type Command = PreparedGenerationRequest;
+    type Request = TextGenerationRequest;
+
+    fn prepare_scheduler_command(
+        &self,
+        agent_name: Option<&str>,
+        request: Self::Request,
+    ) -> Option<Self::Command> {
+        match request {
+            TextGenerationRequest::ContinueFromConversationHistory(request) => {
+                self.generation_command(agent_name, request, Self::prepare_conversation_history)
+            }
+            TextGenerationRequest::ContinueFromRawPrompt(request) => {
+                self.generation_command(agent_name, request, Self::prepare_raw_prompt)
+            }
+        }
+    }
+}
diff --git a/paddler_agent/src/grammar_sampling.rs b/paddler_agent_text_generation/src/grammar_sampling.rs
similarity index 100%
rename from paddler_agent/src/grammar_sampling.rs
rename to paddler_agent_text_generation/src/grammar_sampling.rs
diff --git a/paddler_agent_text_generation/src/image_input.rs b/paddler_agent_text_generation/src/image_input.rs
new file mode 100644
index 000000000..0dd606c4e
--- /dev/null
+++ b/paddler_agent_text_generation/src/image_input.rs
@@ -0,0 +1,60 @@
+use paddler_messaging::chat_template_content_form::ChatTemplateContentForm;
+use paddler_messaging::image_url::ImageUrl;
+
+use crate::multimodal_prompt_support::MultimodalPromptSupport;
+use crate::prompt_modality::PromptModality;
+use crate::text_generation_error::TextGenerationError;
+
+pub enum ImageInput {
+    Supported(MultimodalPromptSupport),
+    Unsupported,
+}
+
+impl ImageInput {
+    #[must_use]
+    pub const fn chat_template_content_form(&self) -> ChatTemplateContentForm {
+        match self {
+            Self::Supported(_) => ChatTemplateContentForm::TypedParts,
+            Self::Unsupported => ChatTemplateContentForm::JoinedText,
+        }
+    }
+
+    pub const fn prompt_modality_for(
+        &self,
+        image_urls: &[ImageUrl],
+    ) -> Result<PromptModality<'_>, TextGenerationError> {
+        if image_urls.is_empty() {
+            return Ok(PromptModality::TextOnly);
+        }
+
+        match self {
+            Self::Supported(multimodal_prompt_support) => {
+                Ok(PromptModality::Multimodal(multimodal_prompt_support))
+            }
+            Self::Unsupported => Err(TextGenerationError::MultimodalNotSupported),
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::mem::discriminant;
+
+    use paddler_messaging::image_url::ImageUrl;
+
+    use super::ImageInput;
+    use crate::text_generation_error::TextGenerationError;
+
+    #[test]
+    fn rejects_images_when_no_multimodal_projection_is_loaded() {
+        assert_eq!(
+            ImageInput::Unsupported
+                .prompt_modality_for(&[ImageUrl {
+                    url: "data:image/png;base64,AAAA".to_owned(),
+                }])
+                .err()
+                .map(|rejection| discriminant(&rejection)),
+            Some(discriminant(&TextGenerationError::MultimodalNotSupported))
+        );
+    }
+}
diff --git a/paddler_agent_text_generation/src/leading_bos_ingestion.rs b/paddler_agent_text_generation/src/leading_bos_ingestion.rs
new file mode 100644
index 000000000..926c5271a
--- /dev/null
+++ b/paddler_agent_text_generation/src/leading_bos_ingestion.rs
@@ -0,0 +1,17 @@
+use llama_cpp_bindings::token::LlamaToken;
+
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+pub enum LeadingBosIngestion {
+    Ingest(LlamaToken),
+    Skip,
+}
+
+impl LeadingBosIngestion {
+    #[must_use]
+    pub const fn token_count(self) -> usize {
+        match self {
+            Self::Ingest(_) => 1,
+            Self::Skip => 0,
+        }
+    }
+}
diff --git a/paddler_agent_text_generation/src/lib.rs b/paddler_agent_text_generation/src/lib.rs
new file mode 100644
index 000000000..7f94dd735
--- /dev/null
+++ b/paddler_agent_text_generation/src/lib.rs
@@ -0,0 +1,47 @@
+pub mod chat_prompt_render_request;
+pub mod chat_prompt_renderer;
+pub mod chat_template_renderer;
+pub mod chat_template_tools;
+pub mod continuous_batch_active_request;
+pub mod continuous_batch_generation_step;
+pub mod continuous_batch_request_phase;
+pub mod continuous_batch_request_state;
+pub mod continuous_batch_scheduler;
+pub mod continuous_batch_scheduler_context;
+pub mod continuous_batch_scheduler_params;
+pub mod continuous_batch_terminal_delivery;
+pub mod continuous_batch_terminal_outcome;
+pub mod conversation_prompt_support;
+pub mod conversation_support;
+pub mod converts_to_mtmd_bitmap;
+pub mod generation_request_preparer;
+pub mod grammar_sampling;
+pub mod image_input;
+pub mod leading_bos_ingestion;
+pub mod media_marked_prompt;
+pub mod model_bos_token;
+pub mod multimodal_ingestion_progress;
+pub mod multimodal_projection;
+pub mod multimodal_prompt_ingestion;
+pub mod multimodal_prompt_support;
+pub mod prepared_generation_request;
+pub mod prepared_prompt;
+pub mod prompt_modality;
+pub mod prompt_tokenizer;
+pub mod request_media_marker;
+pub mod require_grammar_compatible_with_thinking;
+pub mod require_prompt_fits_sequence_context;
+pub mod resident_multimodal_context;
+pub mod resolve_grammar;
+pub mod sampler_chain_factory;
+pub mod sampling_outcome;
+pub mod text_generation_error;
+pub mod text_generation_pipeline;
+pub mod text_generation_request;
+pub mod text_generation_settings;
+pub mod token_classification;
+pub mod token_sampling;
+pub mod tool_call_buffer;
+pub mod tool_call_event;
+pub mod tool_call_handling;
+pub mod tool_call_pipeline;
diff --git a/paddler_agent_text_generation/src/media_marked_prompt.rs b/paddler_agent_text_generation/src/media_marked_prompt.rs
new file mode 100644
index 000000000..b9c12c768
--- /dev/null
+++ b/paddler_agent_text_generation/src/media_marked_prompt.rs
@@ -0,0 +1,115 @@
+use llama_cpp_bindings::mtmd::MtmdBitmap;
+use llama_cpp_bindings::mtmd::MtmdInputPart;
+use llama_cpp_bindings::mtmd::MtmdInputTextPart;
+
+use paddler_messaging::media_marker::MediaMarker;
+
+use crate::text_generation_error::TextGenerationError;
+
+fn text_part<'bitmap>(text_segment: &str) -> MtmdInputPart<'bitmap> {
+    MtmdInputPart::Text(MtmdInputTextPart {
+        parse_special: true,
+        text: text_segment.to_owned(),
+    })
+}
+
+pub struct MediaMarkedPrompt<'prompt> {
+    pub media_marker: &'prompt MediaMarker,
+    pub prompt: &'prompt str,
+}
+
+impl MediaMarkedPrompt<'_> {
+    pub fn interleave_with<'bitmap>(
+        &self,
+        bitmaps: &'bitmap [MtmdBitmap],
+    ) -> Result<Vec<MtmdInputPart<'bitmap>>, TextGenerationError> {
+        let text_segments: Vec<&str> = self
+            .prompt
+            .split(self.media_marker.marker.as_str())
+            .collect();
+
+        match text_segments.split_last() {
+            Some((trailing_text_segment, leading_text_segments))
+                if leading_text_segments.len() == bitmaps.len() =>
+            {
+                let mut parts = Vec::with_capacity(text_segments.len() + bitmaps.len());
+
+                for (text_segment, bitmap) in leading_text_segments.iter().zip(bitmaps) {
+                    parts.push(text_part(text_segment));
+                    parts.push(MtmdInputPart::Bitmap(bitmap));
+                }
+
+                parts.push(text_part(trailing_text_segment));
+
+                Ok(parts)
+            }
+            _ => Err(TextGenerationError::RenderedMediaMarkersMismatchImages {
+                images: bitmaps.len(),
+                rendered_media_markers: text_segments.len().saturating_sub(1),
+            }),
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use llama_cpp_bindings::mtmd::MtmdBitmap;
+    use llama_cpp_bindings::mtmd::MtmdInputPart;
+
+    use paddler_messaging::media_marker::MediaMarker;
+
+    use super::MediaMarkedPrompt;
+    use crate::text_generation_error::TextGenerationError;
+
+    const MEDIA_MARKER: &str = "<__media_test__>";
+
+    fn image() -> MtmdBitmap {
+        MtmdBitmap::from_image_data(2, 2, &[128; 2 * 2 * 3]).unwrap()
+    }
+
+    fn interleaving_failure(prompt: &str, images: &[MtmdBitmap]) -> Option<String> {
+        let media_marker = MediaMarker::new(MEDIA_MARKER.to_owned());
+
+        MediaMarkedPrompt {
+            media_marker: &media_marker,
+            prompt,
+        }
+        .interleave_with(images)
+        .err()
+        .as_ref()
+        .map(ToString::to_string)
+    }
+
+    #[test]
+    fn places_each_image_where_its_marker_was_rendered() {
+        let media_marker = MediaMarker::new(MEDIA_MARKER.to_owned());
+        let images = [image(), image()];
+
+        assert_eq!(
+            MediaMarkedPrompt {
+                media_marker: &media_marker,
+                prompt: &format!("Compare {MEDIA_MARKER} with {MEDIA_MARKER} please"),
+            }
+            .interleave_with(&images)
+            .unwrap()
+            .iter()
+            .map(|part| matches!(part, MtmdInputPart::Bitmap(_)))
+            .collect::<Vec<_>>(),
+            [false, true, false, true, false]
+        );
+    }
+
+    #[test]
+    fn a_rendering_that_lost_an_image_marker_is_rejected() {
+        assert_eq!(
+            interleaving_failure("Describe nothing", &[image()]),
+            Some(
+                TextGenerationError::RenderedMediaMarkersMismatchImages {
+                    images: 1,
+                    rendered_media_markers: 0,
+                }
+                .to_string()
+            )
+        );
+    }
+}
diff --git a/paddler_agent_text_generation/src/model_bos_token.rs b/paddler_agent_text_generation/src/model_bos_token.rs
new file mode 100644
index 000000000..e96aebe2c
--- /dev/null
+++ b/paddler_agent_text_generation/src/model_bos_token.rs
@@ -0,0 +1,88 @@
+use llama_cpp_bindings::model::LlamaModel;
+use llama_cpp_bindings::token::LlamaToken;
+
+use crate::leading_bos_ingestion::LeadingBosIngestion;
+
+#[derive(Clone, Copy)]
+pub enum ModelBosToken {
+    AddedByTokenizer(LlamaToken),
+    NotAddedByTokenizer,
+}
+
+impl ModelBosToken {
+    #[must_use]
+    pub fn of(model: &LlamaModel) -> Self {
+        match model.token_bos() {
+            Some(token) if model.adds_bos_token() => Self::AddedByTokenizer(token),
+            Some(_) | None => Self::NotAddedByTokenizer,
+        }
+    }
+
+    #[must_use]
+    pub fn leading_bos_ingestion(&self, prompt_tokens: &[LlamaToken]) -> LeadingBosIngestion {
+        match self {
+            Self::AddedByTokenizer(token) if prompt_tokens.first() != Some(token) => {
+                LeadingBosIngestion::Ingest(*token)
+            }
+            Self::AddedByTokenizer(_) | Self::NotAddedByTokenizer => LeadingBosIngestion::Skip,
+        }
+    }
+
+    #[must_use]
+    pub fn is_duplicated_at_start_of(&self, prompt_tokens: &[LlamaToken]) -> bool {
+        match self {
+            Self::AddedByTokenizer(token) => prompt_tokens.starts_with(&[*token, *token]),
+            Self::NotAddedByTokenizer => false,
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use llama_cpp_bindings::token::LlamaToken;
+
+    use super::ModelBosToken;
+    use crate::leading_bos_ingestion::LeadingBosIngestion;
+
+    const BOS: LlamaToken = LlamaToken(1);
+    const TEXT: LlamaToken = LlamaToken(42);
+
+    #[test]
+    fn detects_a_rendered_bos_after_the_one_the_tokenizer_added() {
+        assert!(ModelBosToken::AddedByTokenizer(BOS).is_duplicated_at_start_of(&[BOS, BOS, TEXT]));
+    }
+
+    #[test]
+    fn accepts_a_single_leading_bos() {
+        assert!(!ModelBosToken::AddedByTokenizer(BOS).is_duplicated_at_start_of(&[BOS, TEXT]));
+    }
+
+    #[test]
+    fn ingests_the_bos_a_prompt_does_not_start_with() {
+        assert_eq!(
+            ModelBosToken::AddedByTokenizer(BOS).leading_bos_ingestion(&[TEXT]),
+            LeadingBosIngestion::Ingest(BOS)
+        );
+    }
+
+    #[test]
+    fn skips_the_bos_a_prompt_already_starts_with() {
+        assert_eq!(
+            ModelBosToken::AddedByTokenizer(BOS).leading_bos_ingestion(&[BOS, TEXT]),
+            LeadingBosIngestion::Skip
+        );
+    }
+
+    #[test]
+    fn skips_the_bos_the_tokenizer_does_not_add() {
+        assert_eq!(
+            ModelBosToken::NotAddedByTokenizer.leading_bos_ingestion(&[TEXT]),
+            LeadingBosIngestion::Skip
+        );
+    }
+
+    #[test]
+    fn keeps_repeated_bos_tokens_the_tokenizer_did_not_add() {
+        assert!(!ModelBosToken::NotAddedByTokenizer.is_duplicated_at_start_of(&[BOS, BOS, TEXT]));
+    }
+}
diff --git a/paddler_agent/src/multimodal_ingestion_progress.rs b/paddler_agent_text_generation/src/multimodal_ingestion_progress.rs
similarity index 100%
rename from paddler_agent/src/multimodal_ingestion_progress.rs
rename to paddler_agent_text_generation/src/multimodal_ingestion_progress.rs
diff --git a/paddler_agent_text_generation/src/multimodal_projection.rs b/paddler_agent_text_generation/src/multimodal_projection.rs
new file mode 100644
index 000000000..90e470fbb
--- /dev/null
+++ b/paddler_agent_text_generation/src/multimodal_projection.rs
@@ -0,0 +1,7 @@
+use std::path::PathBuf;
+
+#[derive(Clone, Debug, Eq, PartialEq)]
+pub enum MultimodalProjection {
+    File(PathBuf),
+    NotConfigured,
+}
diff --git a/paddler_agent_text_generation/src/multimodal_prompt_ingestion.rs b/paddler_agent_text_generation/src/multimodal_prompt_ingestion.rs
new file mode 100644
index 000000000..9302de216
--- /dev/null
+++ b/paddler_agent_text_generation/src/multimodal_prompt_ingestion.rs
@@ -0,0 +1,68 @@
+use std::sync::Arc;
+
+use llama_cpp_bindings::SampledTokenClassifier;
+use llama_cpp_bindings::context::LlamaContext;
+use llama_cpp_bindings::error::EvalMultimodalChunksError;
+use llama_cpp_bindings::mtmd::FittedMtmdInputChunk;
+use llama_cpp_bindings::mtmd::FittedMtmdInputChunks;
+use llama_cpp_bindings::mtmd::MtmdInputChunks;
+
+use paddler_agent_runtime::sequence_id::SequenceId;
+
+use crate::leading_bos_ingestion::LeadingBosIngestion;
+use crate::multimodal_ingestion_progress::MultimodalIngestionProgress;
+use crate::resident_multimodal_context::ResidentMultimodalContext;
+use crate::text_generation_error::TextGenerationError;
+
+#[derive(Debug)]
+pub struct MultimodalPromptIngestion {
+    pub chunks: FittedMtmdInputChunks<MtmdInputChunks>,
+    pub leading_bos_ingestion: LeadingBosIngestion,
+    pub multimodal_context: Arc<ResidentMultimodalContext>,
+    pub n_batch: i32,
+    pub next_chunk_index: usize,
+}
+
+impl MultimodalPromptIngestion {
+    pub fn ingest_next_chunk(
+        &mut self,
+        token_classifier: &mut SampledTokenClassifier<'_>,
+        llama_context: &LlamaContext<'_>,
+        sequence_id: SequenceId,
+        start_position: i32,
+    ) -> Result<MultimodalIngestionProgress, TextGenerationError> {
+        self.ingest_next_fitted_chunk(token_classifier, llama_context, sequence_id, start_position)
+            .map_err(TextGenerationError::MultimodalIngestionFailed)
+    }
+
+    fn ingest_next_fitted_chunk(
+        &mut self,
+        token_classifier: &mut SampledTokenClassifier<'_>,
+        llama_context: &LlamaContext<'_>,
+        sequence_id: SequenceId,
+        start_position: i32,
+    ) -> Result<MultimodalIngestionProgress, EvalMultimodalChunksError> {
+        let chunk_index = self.next_chunk_index;
+        let chunk = FittedMtmdInputChunk::at(&self.chunks, chunk_index)
+            .ok_or(EvalMultimodalChunksError::ChunkOutOfBounds(chunk_index))?;
+        let is_last_chunk = chunk_index + 1 == self.chunks.len();
+        let next_position = chunk.eval_single(
+            &self.multimodal_context,
+            llama_context,
+            start_position,
+            sequence_id.llama_sequence_id(),
+            self.n_batch,
+            is_last_chunk,
+        )?;
+
+        token_classifier.ingest_fitted_chunk(&chunk);
+
+        self.next_chunk_index += 1;
+
+        Ok(if is_last_chunk {
+            MultimodalIngestionProgress::PromptIngested { next_position }
+        } else {
+            MultimodalIngestionProgress::ChunksRemain { next_position }
+        })
+    }
+}
diff --git a/paddler_agent_text_generation/src/multimodal_prompt_support.rs b/paddler_agent_text_generation/src/multimodal_prompt_support.rs
new file mode 100644
index 000000000..0b1a940f3
--- /dev/null
+++ b/paddler_agent_text_generation/src/multimodal_prompt_support.rs
@@ -0,0 +1,56 @@
+use std::sync::Arc;
+
+use llama_cpp_bindings::mtmd::FittedMtmdInputChunk;
+use llama_cpp_bindings::mtmd::FittedMtmdInputChunks;
+use llama_cpp_bindings::mtmd::MtmdBitmap;
+
+use crate::media_marked_prompt::MediaMarkedPrompt;
+use crate::model_bos_token::ModelBosToken;
+use crate::multimodal_prompt_ingestion::MultimodalPromptIngestion;
+use crate::require_prompt_fits_sequence_context::require_prompt_fits_sequence_context;
+use crate::resident_multimodal_context::ResidentMultimodalContext;
+use crate::text_generation_error::TextGenerationError;
+
+pub struct MultimodalPromptSupport {
+    pub micro_batch_tokens: u32,
+    pub model_bos_token: ModelBosToken,
+    pub multimodal_context: Arc<ResidentMultimodalContext>,
+    pub n_batch: i32,
+    pub sequence_context_size: u32,
+}
+
+impl MultimodalPromptSupport {
+    pub fn prepare_prompt(
+        &self,
+        bitmaps: &[MtmdBitmap],
+        media_marked_prompt: &MediaMarkedPrompt,
+    ) -> Result<MultimodalPromptIngestion, TextGenerationError> {
+        let chunks = self
+            .multimodal_context
+            .tokenize_from_parts(&media_marked_prompt.interleave_with(bitmaps)?, false)
+            .map_err(TextGenerationError::MultimodalTokenizationFailed)?;
+        let chunk_tokens = chunks.total_tokens();
+        let chunks =
+            FittedMtmdInputChunks::fit(chunks, &self.multimodal_context, self.micro_batch_tokens)
+                .map_err(TextGenerationError::for_micro_batch_fit_error)?;
+        let first_chunk = FittedMtmdInputChunk::at(&chunks, 0);
+        let leading_bos_ingestion = self.model_bos_token.leading_bos_ingestion(
+            first_chunk
+                .as_ref()
+                .map_or(&[], FittedMtmdInputChunk::text_tokens),
+        );
+
+        require_prompt_fits_sequence_context(
+            chunk_tokens + leading_bos_ingestion.token_count(),
+            self.sequence_context_size,
+        )?;
+
+        Ok(MultimodalPromptIngestion {
+            chunks,
+            leading_bos_ingestion,
+            multimodal_context: self.multimodal_context.clone(),
+            n_batch: self.n_batch,
+            next_chunk_index: 0,
+        })
+    }
+}
diff --git a/paddler_agent_text_generation/src/prepared_generation_request.rs b/paddler_agent_text_generation/src/prepared_generation_request.rs
new file mode 100644
index 000000000..328aee69f
--- /dev/null
+++ b/paddler_agent_text_generation/src/prepared_generation_request.rs
@@ -0,0 +1,91 @@
+use std::num::NonZeroU32;
+
+use tokio::sync::mpsc;
+
+use paddler_agent_runtime::scheduler_command::SchedulerCommand;
+use paddler_agent_status::slot_guard::SlotGuard;
+use paddler_messaging::generated_token_result::GeneratedTokenResult;
+
+use crate::prepared_prompt::PreparedPrompt;
+use crate::text_generation_error::TextGenerationError;
+use crate::token_classification::TokenClassification;
+use crate::token_sampling::TokenSampling;
+use crate::tool_call_handling::ToolCallHandling;
+
+pub struct PreparedGenerationRequest {
+    pub generate_tokens_stop_rx: mpsc::UnboundedReceiver<()>,
+    pub generated_tokens_tx: mpsc::UnboundedSender<GeneratedTokenResult>,
+    pub max_tokens: NonZeroU32,
+    pub prompt: PreparedPrompt,
+    pub slot_guard: SlotGuard,
+    pub token_classification: TokenClassification,
+    pub token_sampling: TokenSampling,
+    pub tool_call_handling: ToolCallHandling,
+}
+
+impl SchedulerCommand for PreparedGenerationRequest {
+    fn reject_because_the_scheduler_stopped(self, agent_name: Option<&str>) {
+        TextGenerationError::SchedulerUnavailable.report(agent_name, &self.generated_tokens_tx);
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::num::NonZeroU16;
+    use std::num::NonZeroU32;
+    use std::sync::Arc;
+
+    use llama_cpp_bindings::BareJsonToolCalls;
+    use llama_cpp_bindings::StreamingMarkers;
+    use tokio::sync::mpsc;
+
+    use paddler_agent_runtime::scheduler_command::SchedulerCommand as _;
+    use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
+    use paddler_agent_status::slot_guard::SlotGuard;
+    use paddler_inference_parameters::sampling_parameters::SamplingParameters;
+    use paddler_messaging::generated_token_result::GeneratedTokenResult;
+
+    use super::PreparedGenerationRequest;
+    use crate::grammar_sampling::GrammarSampling;
+    use crate::prepared_prompt::PreparedPrompt;
+    use crate::sampler_chain_factory::SamplerChainFactory;
+    use crate::token_classification::TokenClassification;
+    use crate::token_sampling::TokenSampling;
+    use crate::tool_call_handling::ToolCallHandling;
+
+    #[test]
+    fn rejects_generation_when_the_scheduler_stopped() {
+        let (generated_tokens_tx, mut generated_tokens_rx) = mpsc::unbounded_channel();
+        let (_generate_tokens_stop_tx, generate_tokens_stop_rx) = mpsc::unbounded_channel();
+
+        PreparedGenerationRequest {
+            generate_tokens_stop_rx,
+            generated_tokens_tx,
+            max_tokens: NonZeroU32::new(1).unwrap(),
+            prompt: PreparedPrompt::TextTokens(Vec::new()),
+            slot_guard: SlotGuard::new(Arc::new(SlotAggregatedStatus::new(1)), NonZeroU16::MIN),
+            token_classification: TokenClassification {
+                bare_json_tool_calls: BareJsonToolCalls::Ignore,
+                streaming_markers: Arc::new(StreamingMarkers::default()),
+            },
+            token_sampling: TokenSampling {
+                chain: SamplerChainFactory {
+                    sampling_parameters: SamplingParameters::default(),
+                    n_vocab: 1,
+                }
+                .create(0)
+                .unwrap(),
+                grammar: GrammarSampling::Unconstrained,
+            },
+            tool_call_handling: ToolCallHandling::Streamed,
+        }
+        .reject_because_the_scheduler_stopped(Some("agent"));
+
+        assert_eq!(
+            generated_tokens_rx.try_recv(),
+            Ok(GeneratedTokenResult::SchedulerUnavailable(
+                "agent: the scheduler is no longer accepting requests".to_owned()
+            ))
+        );
+    }
+}
diff --git a/paddler_agent/src/prepared_prompt.rs b/paddler_agent_text_generation/src/prepared_prompt.rs
similarity index 100%
rename from paddler_agent/src/prepared_prompt.rs
rename to paddler_agent_text_generation/src/prepared_prompt.rs
diff --git a/paddler_agent/src/prompt_modality.rs b/paddler_agent_text_generation/src/prompt_modality.rs
similarity index 100%
rename from paddler_agent/src/prompt_modality.rs
rename to paddler_agent_text_generation/src/prompt_modality.rs
diff --git a/paddler_agent_text_generation/src/prompt_tokenizer.rs b/paddler_agent_text_generation/src/prompt_tokenizer.rs
new file mode 100644
index 000000000..3a1b734b6
--- /dev/null
+++ b/paddler_agent_text_generation/src/prompt_tokenizer.rs
@@ -0,0 +1,47 @@
+use std::sync::Arc;
+
+use llama_cpp_bindings::model::AddBos;
+use llama_cpp_bindings::model::ParseSpecialTokens;
+use llama_cpp_bindings::token::LlamaToken;
+
+use paddler_agent_runtime::resident_llama_model::ResidentLlamaModel;
+
+use crate::model_bos_token::ModelBosToken;
+use crate::require_prompt_fits_sequence_context::require_prompt_fits_sequence_context;
+use crate::text_generation_error::TextGenerationError;
+
+pub struct PromptTokenizer {
+    pub model: Arc<ResidentLlamaModel>,
+    pub model_bos_token: ModelBosToken,
+    pub sequence_context_size: u32,
+}
+
+impl PromptTokenizer {
+    pub fn tokenize(&self, prompt: &str) -> Result<Vec<LlamaToken>, TextGenerationError> {
+        let mut prompt_tokens = self.tokens_with_special_tokens(prompt)?;
+
+        if self
+            .model_bos_token
+            .is_duplicated_at_start_of(&prompt_tokens)
+        {
+            prompt_tokens.remove(0);
+        }
+
+        if prompt_tokens.is_empty() {
+            return Err(TextGenerationError::PromptTokenizedToNothing);
+        }
+
+        require_prompt_fits_sequence_context(prompt_tokens.len(), self.sequence_context_size)?;
+
+        Ok(prompt_tokens)
+    }
+
+    fn tokens_with_special_tokens(
+        &self,
+        prompt: &str,
+    ) -> Result<Vec<LlamaToken>, TextGenerationError> {
+        self.model
+            .str_to_token(prompt, AddBos::Always, ParseSpecialTokens::Always)
+            .map_err(TextGenerationError::PromptTokenizationFailed)
+    }
+}
diff --git a/paddler_agent_text_generation/src/request_media_marker.rs b/paddler_agent_text_generation/src/request_media_marker.rs
new file mode 100644
index 000000000..fae66891a
--- /dev/null
+++ b/paddler_agent_text_generation/src/request_media_marker.rs
@@ -0,0 +1,72 @@
+use paddler_messaging::conversation_history::ConversationHistory;
+use paddler_messaging::media_marker::MediaMarker;
+use paddler_sha256_hex::sha256_lowercase_hex::sha256_lowercase_hex;
+
+use crate::chat_template_tools::ChatTemplateTools;
+
+#[must_use]
+pub fn request_media_marker(
+    conversation_history: &ConversationHistory,
+    chat_template_tools: &ChatTemplateTools,
+) -> MediaMarker {
+    MediaMarker::new(format!(
+        "<__media_{}__>",
+        sha256_lowercase_hex(
+            format!("{conversation_history:?}{}", chat_template_tools.json).as_bytes()
+        )
+    ))
+}
+
+#[cfg(test)]
+mod tests {
+    use serde_json::Value;
+    use serde_json::json;
+
+    use paddler_messaging::conversation_history::ConversationHistory;
+    use paddler_messaging::conversation_message::ConversationMessage;
+    use paddler_messaging::conversation_message_content::ConversationMessageContent;
+    use paddler_messaging::request_params::continue_from_conversation_history_params::tool::Tool;
+    use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::FunctionCall;
+    use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::function::Function;
+    use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters::Parameters;
+
+    use super::request_media_marker;
+    use crate::chat_template_tools::ChatTemplateTools;
+
+    fn conversation_asking(question: &str) -> ConversationHistory {
+        ConversationHistory::new(vec![ConversationMessage::Message {
+            content: ConversationMessageContent::Text(question.to_owned()),
+            role: "user".to_owned(),
+        }])
+    }
+
+    fn tools_described_as(description: &str) -> ChatTemplateTools {
+        ChatTemplateTools::new(&[Tool::<Value>::Function(FunctionCall {
+            function: Function {
+                name: "get_weather".to_owned(),
+                description: description.to_owned(),
+                parameters: Parameters::Schema(json!({"type": "object"})),
+            },
+        })])
+    }
+
+    #[test]
+    fn requests_asking_different_questions_get_different_markers() {
+        let tools = tools_described_as("Reports the weather");
+
+        assert_ne!(
+            request_media_marker(&conversation_asking("Describe the image"), &tools).marker,
+            request_media_marker(&conversation_asking("Count the llamas"), &tools).marker
+        );
+    }
+
+    #[test]
+    fn requests_describing_their_tools_differently_get_different_markers() {
+        let conversation = conversation_asking("Describe the image");
+
+        assert_ne!(
+            request_media_marker(&conversation, &tools_described_as("Reports the weather")).marker,
+            request_media_marker(&conversation, &tools_described_as("Reports the forecast")).marker
+        );
+    }
+}
diff --git a/paddler_agent/src/require_grammar_compatible_with_thinking.rs b/paddler_agent_text_generation/src/require_grammar_compatible_with_thinking.rs
similarity index 75%
rename from paddler_agent/src/require_grammar_compatible_with_thinking.rs
rename to paddler_agent_text_generation/src/require_grammar_compatible_with_thinking.rs
index b21219014..3dded5ff6 100644
--- a/paddler_agent/src/require_grammar_compatible_with_thinking.rs
+++ b/paddler_agent_text_generation/src/require_grammar_compatible_with_thinking.rs
@@ -1,13 +1,13 @@
 use paddler_messaging::grammar_constraint::GrammarConstraint;
 
-use crate::generation_request_rejection::GenerationRequestRejection;
+use crate::text_generation_error::TextGenerationError;
 
 pub const fn require_grammar_compatible_with_thinking(
     grammar: Option<&GrammarConstraint>,
     enable_thinking: bool,
-) -> Result<(), GenerationRequestRejection> {
+) -> Result<(), TextGenerationError> {
     if grammar.is_some() && enable_thinking {
-        Err(GenerationRequestRejection::GrammarIncompatibleWithThinking)
+        Err(TextGenerationError::GrammarIncompatibleWithThinking)
     } else {
         Ok(())
     }
@@ -20,7 +20,7 @@ mod tests {
     use paddler_messaging::grammar_constraint::GrammarConstraint;
 
     use super::require_grammar_compatible_with_thinking;
-    use crate::generation_request_rejection::GenerationRequestRejection;
+    use crate::text_generation_error::TextGenerationError;
 
     #[test]
     fn rejects_a_grammar_when_thinking_is_enabled() {
@@ -33,7 +33,7 @@ mod tests {
             require_grammar_compatible_with_thinking(Some(&grammar), true)
                 .map_err(|rejection| discriminant(&rejection)),
             Err(discriminant(
-                &GenerationRequestRejection::GrammarIncompatibleWithThinking
+                &TextGenerationError::GrammarIncompatibleWithThinking
             ))
         );
     }
diff --git a/paddler_agent/src/require_prompt_fits_sequence_context.rs b/paddler_agent_text_generation/src/require_prompt_fits_sequence_context.rs
similarity index 80%
rename from paddler_agent/src/require_prompt_fits_sequence_context.rs
rename to paddler_agent_text_generation/src/require_prompt_fits_sequence_context.rs
index 5884d70e7..57c725e0e 100644
--- a/paddler_agent/src/require_prompt_fits_sequence_context.rs
+++ b/paddler_agent_text_generation/src/require_prompt_fits_sequence_context.rs
@@ -1,15 +1,15 @@
 use paddler_messaging::oversized_prompt_details::OversizedPromptDetails;
 
-use crate::generation_request_rejection::GenerationRequestRejection;
+use crate::text_generation_error::TextGenerationError;
 
 pub const fn require_prompt_fits_sequence_context(
     prompt_tokens: usize,
     sequence_context_size: u32,
-) -> Result<(), GenerationRequestRejection> {
+) -> Result<(), TextGenerationError> {
     if prompt_tokens < sequence_context_size as usize {
         Ok(())
     } else {
-        Err(GenerationRequestRejection::PromptExceedsContextSize {
+        Err(TextGenerationError::PromptExceedsContextSize {
             details: OversizedPromptDetails {
                 prompt_tokens,
                 sequence_context_size,
@@ -23,12 +23,12 @@ mod tests {
     use paddler_messaging::oversized_prompt_details::OversizedPromptDetails;
 
     use super::require_prompt_fits_sequence_context;
-    use crate::generation_request_rejection::GenerationRequestRejection;
+    use crate::text_generation_error::TextGenerationError;
 
     fn rejects_as_oversized(prompt_tokens: usize, sequence_context_size: u32) -> bool {
         matches!(
             require_prompt_fits_sequence_context(prompt_tokens, sequence_context_size),
-            Err(GenerationRequestRejection::PromptExceedsContextSize { details })
+            Err(TextGenerationError::PromptExceedsContextSize { details })
                 if details == OversizedPromptDetails {
                     prompt_tokens,
                     sequence_context_size,
diff --git a/paddler_agent_text_generation/src/resident_multimodal_context.rs b/paddler_agent_text_generation/src/resident_multimodal_context.rs
new file mode 100644
index 000000000..da6fa093a
--- /dev/null
+++ b/paddler_agent_text_generation/src/resident_multimodal_context.rs
@@ -0,0 +1,46 @@
+use std::ops::Deref;
+use std::path::Path;
+
+use llama_cpp_bindings::model::LlamaModel;
+use llama_cpp_bindings::mtmd::MtmdContext;
+use llama_cpp_bindings::mtmd::MtmdContextParams;
+use llama_cpp_bindings::mtmd::MtmdInitError;
+
+use paddler_agent_runtime::vram_occupancy::VramOccupancy;
+
+#[derive(Debug)]
+pub struct ResidentMultimodalContext {
+    pub multimodal_context: MtmdContext,
+    pub vram_occupancy: VramOccupancy,
+}
+
+impl ResidentMultimodalContext {
+    pub fn load(
+        multimodal_projection_path: &Path,
+        llama_model: &LlamaModel,
+        vram_occupancy: VramOccupancy,
+    ) -> Result<Self, MtmdInitError> {
+        #[expect(
+            clippy::disallowed_methods,
+            reason = "the multimodal projection is loaded here, paired with its VRAM occupancy"
+        )]
+        let initialized_multimodal_context = MtmdContext::init_from_file(
+            &multimodal_projection_path.to_string_lossy(),
+            llama_model,
+            &MtmdContextParams::default(),
+        );
+
+        Ok(Self {
+            multimodal_context: initialized_multimodal_context?,
+            vram_occupancy,
+        })
+    }
+}
+
+impl Deref for ResidentMultimodalContext {
+    type Target = MtmdContext;
+
+    fn deref(&self) -> &MtmdContext {
+        &self.multimodal_context
+    }
+}
diff --git a/paddler_agent_text_generation/src/resolve_grammar.rs b/paddler_agent_text_generation/src/resolve_grammar.rs
new file mode 100644
index 000000000..17d0cd92d
--- /dev/null
+++ b/paddler_agent_text_generation/src/resolve_grammar.rs
@@ -0,0 +1,32 @@
+use llama_cpp_bindings::json_schema_to_grammar;
+use llama_cpp_bindings::model::LlamaModel;
+use llama_cpp_bindings::sampling::LlamaSampler;
+
+use paddler_messaging::grammar_constraint::GrammarConstraint;
+
+use crate::grammar_sampling::GrammarSampling;
+use crate::text_generation_error::TextGenerationError;
+
+const JSON_SCHEMA_GRAMMAR_ROOT: &str = "root";
+
+pub fn resolve_grammar(
+    grammar: Option<GrammarConstraint>,
+    model: &LlamaModel,
+) -> Result<GrammarSampling, TextGenerationError> {
+    match grammar {
+        None => Ok(GrammarSampling::Unconstrained),
+        Some(GrammarConstraint::Gbnf { grammar, root }) => {
+            LlamaSampler::grammar(model, &grammar, &root)
+                .map(GrammarSampling::Constrained)
+                .map_err(TextGenerationError::for_gbnf_grammar_error)
+        }
+        Some(GrammarConstraint::JsonSchema { schema }) => {
+            let grammar = json_schema_to_grammar(&schema)
+                .map_err(TextGenerationError::GrammarConversionFailed)?;
+
+            LlamaSampler::grammar(model, &grammar, JSON_SCHEMA_GRAMMAR_ROOT)
+                .map(GrammarSampling::Constrained)
+                .map_err(TextGenerationError::GrammarSamplerInitializationFailed)
+        }
+    }
+}
diff --git a/paddler_agent_text_generation/src/sampler_chain_factory.rs b/paddler_agent_text_generation/src/sampler_chain_factory.rs
new file mode 100644
index 000000000..4adec3d3a
--- /dev/null
+++ b/paddler_agent_text_generation/src/sampler_chain_factory.rs
@@ -0,0 +1,68 @@
+use llama_cpp_bindings::error::SamplingError;
+use llama_cpp_bindings::sampling::LlamaSampler;
+
+use paddler_inference_parameters::sampling_parameters::SamplingParameters;
+
+pub struct SamplerChainFactory {
+    pub sampling_parameters: SamplingParameters,
+    pub n_vocab: i32,
+}
+
+impl SamplerChainFactory {
+    pub fn create(&self, seed: u32) -> Result<LlamaSampler, SamplingError> {
+        let sampling_parameters = &self.sampling_parameters;
+        let samplers = [
+            LlamaSampler::penalties(
+                self.n_vocab,
+                sampling_parameters.penalty_last_n,
+                sampling_parameters.penalty_repeat,
+                sampling_parameters.penalty_frequency,
+                sampling_parameters.penalty_presence,
+            ),
+            LlamaSampler::top_k(sampling_parameters.top_k),
+            LlamaSampler::top_p(sampling_parameters.top_p, 0),
+            LlamaSampler::min_p(sampling_parameters.min_p, 0),
+            LlamaSampler::temp(sampling_parameters.temperature),
+            LlamaSampler::dist(seed),
+        ];
+
+        LlamaSampler::chain_simple(samplers.into_iter().collect::<Result<Vec<_>, _>>()?)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use llama_cpp_bindings::token::LlamaToken;
+    use llama_cpp_bindings::token::data::LlamaTokenData;
+    use llama_cpp_bindings::token::data_array::LlamaTokenDataArray;
+
+    use paddler_inference_parameters::sampling_parameters::SamplingParameters;
+
+    use super::SamplerChainFactory;
+
+    #[test]
+    fn chain_honors_configured_top_k() {
+        let sampling_parameters = SamplingParameters {
+            top_k: 1,
+            ..SamplingParameters::default()
+        };
+        let sampler_chain = SamplerChainFactory {
+            sampling_parameters,
+            n_vocab: 3,
+        }
+        .create(7)
+        .unwrap();
+        let mut candidates = LlamaTokenDataArray::from_iter(
+            [
+                LlamaTokenData::new(LlamaToken(0), 0.5, 0.0),
+                LlamaTokenData::new(LlamaToken(1), 3.0, 0.0),
+                LlamaTokenData::new(LlamaToken(2), 1.0, 0.0),
+            ],
+            false,
+        );
+
+        candidates.apply_sampler(&sampler_chain).unwrap();
+
+        assert_eq!(candidates.selected_token(), Some(LlamaToken(1)));
+    }
+}
diff --git a/paddler_agent/src/sampling_outcome.rs b/paddler_agent_text_generation/src/sampling_outcome.rs
similarity index 100%
rename from paddler_agent/src/sampling_outcome.rs
rename to paddler_agent_text_generation/src/sampling_outcome.rs
diff --git a/paddler_agent_text_generation/src/text_generation_error.rs b/paddler_agent_text_generation/src/text_generation_error.rs
new file mode 100644
index 000000000..f31f12c11
--- /dev/null
+++ b/paddler_agent_text_generation/src/text_generation_error.rs
@@ -0,0 +1,634 @@
+use std::path::PathBuf;
+
+use llama_cpp_bindings::batch_add_error::BatchAddError;
+use llama_cpp_bindings::error::ClearKvCacheSeqError;
+use llama_cpp_bindings::error::EvalMultimodalChunksError;
+use llama_cpp_bindings::error::GrammarError;
+use llama_cpp_bindings::error::JsonSchemaToGrammarError;
+use llama_cpp_bindings::error::LogitsError;
+use llama_cpp_bindings::error::MarkerDetectionError;
+use llama_cpp_bindings::error::ParseChatMessageError;
+use llama_cpp_bindings::error::SamplerAcceptError;
+use llama_cpp_bindings::error::SamplerApplyError;
+use llama_cpp_bindings::error::SamplingError;
+use llama_cpp_bindings::error::StringToTokenError;
+use llama_cpp_bindings::error::TokenToStringError;
+use llama_cpp_bindings::mtmd::MtmdBitmapError;
+use llama_cpp_bindings::mtmd::MtmdEvalError;
+use llama_cpp_bindings::mtmd::MtmdInitError;
+use llama_cpp_bindings::mtmd::MtmdTokenizeError;
+use log::error;
+use minijinja::Error as MinijinjaError;
+use tokio::sync::mpsc;
+
+use paddler_agent_runtime::agent_runtime_error::AgentRuntimeError;
+use paddler_agent_runtime::rejection_description::rejection_description;
+use paddler_agent_runtime::send_result_or_warn::send_result_or_warn;
+use paddler_image_decoder::decoded_image_error::DecodedImageError;
+use paddler_inference_parameters::invalid_inference_parameters::InvalidInferenceParameters;
+use paddler_messaging::agent_issue_params::model_path::ModelPath;
+use paddler_messaging::generated_token_result::GeneratedTokenResult;
+use paddler_messaging::oversized_media_details::OversizedMediaDetails;
+use paddler_messaging::oversized_prompt_details::OversizedPromptDetails;
+use paddler_tool_call_validator::tool_call_validation_error::ToolCallValidationError;
+
+#[derive(Debug, thiserror::Error)]
+pub enum TextGenerationError {
+    #[error(transparent)]
+    AgentRuntime(#[from] AgentRuntimeError),
+
+    #[error(
+        "the model {} has no chat template, so it cannot continue a conversation",
+        model_path.model_path
+    )]
+    ConversationRequiresChatTemplate { model_path: ModelPath },
+
+    #[error("the chat template does not compile: {0}")]
+    ChatTemplateDoesNotCompile(#[source] MinijinjaError),
+
+    #[error("unable to load the multimodal projection from {multimodal_projection_path}: {source}")]
+    MultimodalProjectionLoadFailed {
+        multimodal_projection_path: PathBuf,
+        #[source]
+        source: MtmdInitError,
+    },
+
+    #[error("failed to apply the grammar sampler to the token candidates: {0}")]
+    GrammarConstraintFailed(#[source] SamplerApplyError),
+
+    #[error("failed to apply the sampler chain to the token candidates: {0}")]
+    SamplerChainApplicationFailed(#[source] SamplerApplyError),
+
+    #[error("the sampler chain failed to accept the selected token: {0}")]
+    SamplerChainRejectedToken(#[source] SamplerAcceptError),
+
+    #[error("failed to read the token candidates for sampling: {0}")]
+    TokenDataUnreadable(#[source] LogitsError),
+
+    #[error("unable to render the model's special token: {0}")]
+    SpecialTokenUnrenderable(#[source] TokenToStringError),
+
+    #[error("unable to detect the model's streaming markers: {0}")]
+    StreamingMarkersUnavailable(#[source] MarkerDetectionError),
+
+    #[error("the sampling parameters do not fit the sequence context: {0}")]
+    SamplingParametersUnfitForTheSequenceContext(#[source] InvalidInferenceParameters),
+
+    #[error("Grammar constraints are incompatible with thinking mode")]
+    GrammarIncompatibleWithThinking,
+
+    #[error("Failed to convert JSON schema to grammar: {0}")]
+    GrammarConversionFailed(#[source] JsonSchemaToGrammarError),
+
+    #[error("failed to decode images: {0}")]
+    ImageDecodingFailed(#[source] DecodedImageError),
+
+    #[error("failed to create image bitmap: {0}")]
+    ImageBitmapCreationFailed(#[source] MtmdBitmapError),
+
+    #[error("received images but model does not support multimodal input")]
+    MultimodalNotSupported,
+
+    #[error("no model is loaded")]
+    ModelNotLoaded,
+
+    #[error("failed to render chat template: {0}")]
+    ChatTemplateRenderingFailed(#[source] MinijinjaError),
+
+    #[error("{0}")]
+    ToolSchemaInvalid(#[source] ToolCallValidationError),
+
+    #[error("failed to build this model's tool call parser: {0}")]
+    ToolCallParserCreationFailed(#[source] ParseChatMessageError),
+
+    #[error("failed to tokenize prompt: {0}")]
+    PromptTokenizationFailed(#[source] StringToTokenError),
+
+    #[error("the prompt tokenized to no tokens, so there is nothing to generate from")]
+    PromptTokenizedToNothing,
+
+    #[error(
+        "prompt has {} tokens but each sequence holds {} tokens",
+        details.prompt_tokens,
+        details.sequence_context_size
+    )]
+    PromptExceedsContextSize { details: OversizedPromptDetails },
+
+    #[error("the scheduler is no longer accepting requests")]
+    SchedulerUnavailable,
+
+    #[error("no available sequence slots, all slots are busy")]
+    NoSequenceSlotAvailable,
+
+    #[error("invalid grammar: {0}")]
+    GrammarInvalid(#[source] GrammarError),
+    #[error("failed to initialize grammar sampler: {0}")]
+    GrammarSamplerInitializationFailed(#[source] GrammarError),
+
+    #[error("failed to create sampler chain: {0}")]
+    SamplerChainCreationFailed(#[source] SamplingError),
+
+    #[error("failed to tokenize multimodal input: {0}")]
+    MultimodalTokenizationFailed(#[source] MtmdTokenizeError),
+
+    #[error(
+        "the chat template rendered {rendered_media_markers} media markers for {images} images"
+    )]
+    RenderedMediaMarkersMismatchImages {
+        images: usize,
+        rendered_media_markers: usize,
+    },
+
+    #[error(
+        "media chunk has {} tokens but one micro batch holds {} tokens",
+        details.media_tokens,
+        details.micro_batch_tokens
+    )]
+    MediaExceedsMicroBatch { details: OversizedMediaDetails },
+    #[error("failed to check media chunks against the micro batch: {0}")]
+    MediaMicroBatchCheckFailed(#[source] MtmdEvalError),
+
+    #[error("failed to ingest multimodal prompt: {0}")]
+    MultimodalIngestionFailed(#[source] EvalMultimodalChunksError),
+
+    #[error("failed to add a token to the batch: {0}")]
+    BatchAssemblyFailed(#[source] BatchAddError),
+
+    #[error("failed to clear the KV cache of the sequence: {0}")]
+    KvCacheClearFailed(#[source] ClearKvCacheSeqError),
+}
+
+impl TextGenerationError {
+    #[must_use]
+    pub const fn for_gbnf_grammar_error(grammar_error: GrammarError) -> Self {
+        match grammar_error {
+            GrammarError::GrammarContainsNul(_)
+            | GrammarError::GrammarRejected(_)
+            | GrammarError::RootContainsNul(_)
+            | GrammarError::RootNotFound => Self::GrammarInvalid(grammar_error),
+            GrammarError::FfiContract(_)
+            | GrammarError::FfiStatus(_)
+            | GrammarError::GrammarMalformed
+            | GrammarError::InvalidTriggerPattern { .. }
+            | GrammarError::LazyGrammarMalformed
+            | GrammarError::LlamaCppOutOfMemory
+            | GrammarError::LlguidanceFactoryUnavailable { .. }
+            | GrammarError::LlguidanceGrammarInvalid { .. }
+            | GrammarError::LlguidanceParserUnavailable { .. }
+            | GrammarError::NotEnoughMemory
+            | GrammarError::Reported { .. }
+            | GrammarError::SamplerInitialization(_)
+            | GrammarError::SequenceBreakerContainsNul(_)
+            | GrammarError::TokEnvUnavailable(_)
+            | GrammarError::TriggerPatternContainsNul(_) => {
+                Self::GrammarSamplerInitializationFailed(grammar_error)
+            }
+        }
+    }
+
+    #[must_use]
+    pub fn for_micro_batch_fit_error(fit_error: MtmdEvalError) -> Self {
+        match fit_error {
+            MtmdEvalError::NonCausalChunkExceedsMicroBatch(mismatch) => {
+                Self::MediaExceedsMicroBatch {
+                    details: OversizedMediaDetails {
+                        media_tokens: mismatch.chunk_tokens,
+                        micro_batch_tokens: mismatch.micro_batch_tokens,
+                    },
+                }
+            }
+            other_fit_error => Self::MediaMicroBatchCheckFailed(other_fit_error),
+        }
+    }
+
+    pub fn report(
+        self,
+        agent_name: Option<&str>,
+        generated_tokens_tx: &mpsc::UnboundedSender<GeneratedTokenResult>,
+    ) {
+        send_result_or_warn(
+            agent_name,
+            generated_tokens_tx,
+            self.into_generated_token_result(agent_name),
+        );
+    }
+
+    #[must_use]
+    pub fn into_generated_token_result(self, agent_name: Option<&str>) -> GeneratedTokenResult {
+        let message = rejection_description(agent_name, &self);
+
+        error!("{message}");
+
+        match self {
+            Self::GrammarIncompatibleWithThinking => {
+                GeneratedTokenResult::GrammarIncompatibleWithThinking(message)
+            }
+            Self::GrammarConversionFailed(_) | Self::GrammarInvalid(_) => {
+                GeneratedTokenResult::GrammarSyntaxError(message)
+            }
+            Self::ImageDecodingFailed(_) | Self::ImageBitmapCreationFailed(_) => {
+                GeneratedTokenResult::ImageDecodingFailed(message)
+            }
+            Self::AgentRuntime(_)
+            | Self::ModelNotLoaded
+            | Self::MultimodalProjectionLoadFailed { .. }
+            | Self::SamplingParametersUnfitForTheSequenceContext(_)
+            | Self::SpecialTokenUnrenderable(_)
+            | Self::StreamingMarkersUnavailable(_) => GeneratedTokenResult::ModelNotLoaded(message),
+            Self::MultimodalNotSupported => GeneratedTokenResult::MultimodalNotSupported(message),
+            Self::ChatTemplateDoesNotCompile(_)
+            | Self::ChatTemplateRenderingFailed(_)
+            | Self::ConversationRequiresChatTemplate { .. }
+            | Self::ToolCallParserCreationFailed(_) => {
+                GeneratedTokenResult::ChatTemplateError(message)
+            }
+            Self::ToolSchemaInvalid(_) => GeneratedTokenResult::ToolSchemaInvalid(message),
+            Self::GrammarSamplerInitializationFailed(_) => {
+                GeneratedTokenResult::GrammarInitializationFailed(message)
+            }
+            Self::MediaExceedsMicroBatch { details } => {
+                GeneratedTokenResult::MediaExceedsMicroBatch(details)
+            }
+            Self::PromptExceedsContextSize { details } => {
+                GeneratedTokenResult::PromptExceedsContextSize(details)
+            }
+            Self::MediaMicroBatchCheckFailed(_) => {
+                GeneratedTokenResult::MediaMicroBatchCheckFailed(message)
+            }
+            Self::PromptTokenizationFailed(_) => {
+                GeneratedTokenResult::PromptTokenizationFailed(message)
+            }
+            Self::PromptTokenizedToNothing => {
+                GeneratedTokenResult::PromptTokenizedToNothing(message)
+            }
+            Self::SchedulerUnavailable => GeneratedTokenResult::SchedulerUnavailable(message),
+            Self::NoSequenceSlotAvailable => GeneratedTokenResult::NoSequenceSlotAvailable(message),
+            Self::SamplerChainCreationFailed(_) => {
+                GeneratedTokenResult::SamplerChainCreationFailed(message)
+            }
+            Self::GrammarConstraintFailed(_)
+            | Self::SamplerChainApplicationFailed(_)
+            | Self::SamplerChainRejectedToken(_)
+            | Self::TokenDataUnreadable(_) => GeneratedTokenResult::SamplerError(message),
+            Self::MultimodalTokenizationFailed(_)
+            | Self::RenderedMediaMarkersMismatchImages { .. } => {
+                GeneratedTokenResult::MultimodalTokenizationFailed(message)
+            }
+            Self::MultimodalIngestionFailed(_) => {
+                GeneratedTokenResult::MultimodalIngestionFailed(message)
+            }
+            Self::BatchAssemblyFailed(_) => GeneratedTokenResult::BatchAssemblyFailed(message),
+            Self::KvCacheClearFailed(_) => GeneratedTokenResult::KvCacheClearFailed(message),
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::ffi::CString;
+
+    use llama_cpp_bindings::batch_add_error::BatchAddError;
+    use llama_cpp_bindings::error::ClearKvCacheSeqError;
+    use llama_cpp_bindings::error::EvalMultimodalChunksError;
+    use llama_cpp_bindings::error::FfiContractError;
+    use llama_cpp_bindings::error::FfiStatusError;
+    use llama_cpp_bindings::error::GrammarError;
+    use llama_cpp_bindings::error::JsonSchemaToGrammarError;
+    use llama_cpp_bindings::error::LogitsError;
+    use llama_cpp_bindings::error::ParseChatMessageError;
+    use llama_cpp_bindings::error::SamplingError;
+    use llama_cpp_bindings::error::StringToTokenError;
+    use llama_cpp_bindings::mtmd::MtmdEvalError;
+    use llama_cpp_bindings::mtmd::MtmdTokenizeError;
+    use llama_cpp_bindings::mtmd::NonCausalChunkMicroBatchMismatch;
+    use minijinja::Error as MinijinjaError;
+    use minijinja::ErrorKind as MinijinjaErrorKind;
+    use serde_json::Map;
+    use serde_json::json;
+    use tokio::sync::mpsc;
+
+    use paddler_image_decoder::decoded_image_error::DecodedImageError;
+    use paddler_messaging::agent_issue_params::model_path::ModelPath;
+    use paddler_messaging::generated_token_result::GeneratedTokenResult;
+    use paddler_messaging::oversized_media_details::OversizedMediaDetails;
+    use paddler_messaging::oversized_prompt_details::OversizedPromptDetails;
+    use paddler_messaging::request_params::continue_from_conversation_history_params::tool::Tool;
+    use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::FunctionCall;
+    use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::function::Function;
+    use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters::Parameters;
+    use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters_schema::validated_parameters_schema::ValidatedParametersSchema;
+    use paddler_tool_call_validator::tool_call_validator::ToolCallValidator;
+
+    use super::TextGenerationError;
+
+    struct ExpectedReport {
+        rejection: TextGenerationError,
+        into_result: fn(String) -> GeneratedTokenResult,
+    }
+
+    fn reported(rejection: TextGenerationError) -> GeneratedTokenResult {
+        let (generated_tokens_tx, mut generated_tokens_rx) = mpsc::unbounded_channel();
+
+        rejection.report(Some("agent"), &generated_tokens_tx);
+
+        generated_tokens_rx.try_recv().unwrap()
+    }
+
+    fn agent_message(description: &str) -> String {
+        format!("agent: {description}")
+    }
+
+    #[test]
+    fn reports_thinking_incompatibility_with_the_agent_name() {
+        assert_eq!(
+            reported(TextGenerationError::GrammarIncompatibleWithThinking),
+            GeneratedTokenResult::GrammarIncompatibleWithThinking(agent_message(
+                "Grammar constraints are incompatible with thinking mode"
+            ))
+        );
+    }
+
+    #[test]
+    fn reports_grammar_conversion_failure_as_grammar_syntax_error() {
+        let conversion_error = JsonSchemaToGrammarError::NotEnoughMemory;
+        let expected_message = agent_message(&format!(
+            "Failed to convert JSON schema to grammar: {conversion_error}"
+        ));
+
+        assert_eq!(
+            reported(TextGenerationError::GrammarConversionFailed(
+                conversion_error
+            )),
+            GeneratedTokenResult::GrammarSyntaxError(expected_message)
+        );
+    }
+
+    #[test]
+    fn reports_a_conversation_on_a_model_without_a_chat_template_as_a_chat_template_error() {
+        assert_eq!(
+            reported(TextGenerationError::ConversationRequiresChatTemplate {
+                model_path: ModelPath {
+                    model_path: "nomic.gguf".to_owned(),
+                },
+            }),
+            GeneratedTokenResult::ChatTemplateError(agent_message(
+                "the model nomic.gguf has no chat template, so it cannot continue a conversation"
+            ))
+        );
+    }
+
+    #[test]
+    fn reports_image_failures_as_image_decoding_failure() {
+        assert_eq!(
+            reported(TextGenerationError::ImageDecodingFailed(
+                DecodedImageError::MissingCommaSeparator,
+            )),
+            GeneratedTokenResult::ImageDecodingFailed(agent_message(
+                "failed to decode images: Invalid data URI: missing comma separator"
+            ))
+        );
+    }
+
+    #[test]
+    fn reports_missing_multimodal_support() {
+        assert_eq!(
+            reported(TextGenerationError::MultimodalNotSupported),
+            GeneratedTokenResult::MultimodalNotSupported(agent_message(
+                "received images but model does not support multimodal input"
+            ))
+        );
+    }
+
+    #[test]
+    fn reports_prompt_rendering_failures_as_chat_template_error() {
+        let rendering_error =
+            MinijinjaError::new(MinijinjaErrorKind::UndefinedError, "missing variable");
+        let expected_message = agent_message(&format!(
+            "failed to render chat template: {rendering_error}"
+        ));
+
+        assert_eq!(
+            reported(TextGenerationError::ChatTemplateRenderingFailed(
+                rendering_error
+            )),
+            GeneratedTokenResult::ChatTemplateError(expected_message)
+        );
+    }
+
+    #[test]
+    fn reports_invalid_tool_schema() {
+        let mut properties = Map::new();
+
+        properties.insert("location".to_owned(), json!({"type": 123}));
+
+        let schema_error =
+            ToolCallValidator::from_tools(&[Tool::from(Tool::Function(FunctionCall {
+                function: Function {
+                    name: "get_weather".to_owned(),
+                    description: "fetch weather".to_owned(),
+                    parameters: Parameters::Schema(ValidatedParametersSchema {
+                        schema_type: "object".to_owned(),
+                        properties: Some(properties),
+                        ..ValidatedParametersSchema::default()
+                    }),
+                },
+            }))])
+            .err()
+            .expect("a property typed as a number is not a valid JSON Schema");
+        let expected_message = agent_message(&schema_error.to_string());
+
+        assert_eq!(
+            reported(TextGenerationError::ToolSchemaInvalid(schema_error)),
+            GeneratedTokenResult::ToolSchemaInvalid(expected_message)
+        );
+    }
+
+    #[test]
+    fn reports_grammar_sampler_initialization_failure() {
+        let grammar_error = GrammarError::FfiStatus(FfiStatusError {
+            operation: "llama_sampler_init_grammar",
+            code: 1,
+        });
+        let expected_message = agent_message(&format!(
+            "failed to initialize grammar sampler: {grammar_error}"
+        ));
+
+        assert_eq!(
+            reported(TextGenerationError::GrammarSamplerInitializationFailed(
+                grammar_error
+            )),
+            GeneratedTokenResult::GrammarInitializationFailed(expected_message)
+        );
+    }
+
+    #[test]
+    fn reports_a_grammar_the_client_malformed_as_grammar_syntax_error() {
+        let expected_message =
+            agent_message(&format!("invalid grammar: {}", GrammarError::RootNotFound));
+
+        assert_eq!(
+            reported(TextGenerationError::for_gbnf_grammar_error(
+                GrammarError::RootNotFound
+            )),
+            GeneratedTokenResult::GrammarSyntaxError(expected_message)
+        );
+    }
+
+    #[test]
+    fn reports_a_grammar_the_agent_failed_to_initialize_as_grammar_initialization_failure() {
+        let expected_message = agent_message(&format!(
+            "failed to initialize grammar sampler: {}",
+            GrammarError::NotEnoughMemory
+        ));
+
+        assert_eq!(
+            reported(TextGenerationError::for_gbnf_grammar_error(
+                GrammarError::NotEnoughMemory
+            )),
+            GeneratedTokenResult::GrammarInitializationFailed(expected_message)
+        );
+    }
+
+    #[test]
+    fn reports_tool_call_parser_creation_failure_as_chat_template_error() {
+        let parser_error = ParseChatMessageError::ToolsNotAnArray;
+        let expected_message = agent_message(&format!(
+            "failed to build this model's tool call parser: {parser_error}"
+        ));
+
+        assert_eq!(
+            reported(TextGenerationError::ToolCallParserCreationFailed(
+                parser_error
+            )),
+            GeneratedTokenResult::ChatTemplateError(expected_message)
+        );
+    }
+
+    #[test]
+    fn reports_media_exceeding_the_micro_batch_with_its_token_counts() {
+        let rejection = TextGenerationError::for_micro_batch_fit_error(
+            MtmdEvalError::NonCausalChunkExceedsMicroBatch(NonCausalChunkMicroBatchMismatch {
+                chunk_tokens: 256,
+                micro_batch_tokens: 128,
+            }),
+        );
+
+        assert_eq!(
+            reported(rejection),
+            GeneratedTokenResult::MediaExceedsMicroBatch(OversizedMediaDetails {
+                media_tokens: 256,
+                micro_batch_tokens: 128,
+            })
+        );
+    }
+
+    #[test]
+    fn reports_other_micro_batch_check_failures_as_micro_batch_check_failures() {
+        let fit_error = MtmdEvalError::FfiContract(FfiContractError {
+            operation: "mtmd_input_chunks_get",
+            detail: "returned a null chunk within the chunk count",
+        });
+        let expected_message = agent_message(&format!(
+            "failed to check media chunks against the micro batch: {fit_error}"
+        ));
+
+        assert_eq!(
+            reported(TextGenerationError::for_micro_batch_fit_error(fit_error)),
+            GeneratedTokenResult::MediaMicroBatchCheckFailed(expected_message)
+        );
+    }
+
+    #[test]
+    fn reports_oversized_prompt_with_its_token_counts() {
+        assert_eq!(
+            reported(TextGenerationError::PromptExceedsContextSize {
+                details: OversizedPromptDetails {
+                    prompt_tokens: 9895,
+                    sequence_context_size: 8192,
+                },
+            }),
+            GeneratedTokenResult::PromptExceedsContextSize(OversizedPromptDetails {
+                prompt_tokens: 9895,
+                sequence_context_size: 8192,
+            })
+        );
+    }
+
+    #[test]
+    fn reports_multimodal_ingestion_failures_as_ingestion_failures() {
+        let rejection = TextGenerationError::MultimodalIngestionFailed(
+            EvalMultimodalChunksError::ChunkOutOfBounds(3),
+        );
+
+        assert_eq!(
+            reported(rejection),
+            GeneratedTokenResult::MultimodalIngestionFailed(agent_message(
+                "failed to ingest multimodal prompt: chunk index 3 out of bounds during post-eval walk"
+            ))
+        );
+    }
+
+    #[test]
+    fn reports_internal_scheduler_failures_under_their_own_results() {
+        for ExpectedReport {
+            rejection,
+            into_result,
+        } in [
+            ExpectedReport {
+                rejection: TextGenerationError::BatchAssemblyFailed(BatchAddError::EmptyBuffer),
+                into_result: GeneratedTokenResult::BatchAssemblyFailed,
+            },
+            ExpectedReport {
+                rejection: TextGenerationError::KvCacheClearFailed(
+                    ClearKvCacheSeqError::PartialSequenceNotRemoved {
+                        seq_id: 1,
+                        p0: 0,
+                        p1: 4,
+                    },
+                ),
+                into_result: GeneratedTokenResult::KvCacheClearFailed,
+            },
+            ExpectedReport {
+                rejection: TextGenerationError::ModelNotLoaded,
+                into_result: GeneratedTokenResult::ModelNotLoaded,
+            },
+            ExpectedReport {
+                rejection: TextGenerationError::MultimodalTokenizationFailed(
+                    MtmdTokenizeError::BitmapCountDoesNotMatchMarkerCount,
+                ),
+                into_result: GeneratedTokenResult::MultimodalTokenizationFailed,
+            },
+            ExpectedReport {
+                rejection: TextGenerationError::RenderedMediaMarkersMismatchImages {
+                    images: 2,
+                    rendered_media_markers: 1,
+                },
+                into_result: GeneratedTokenResult::MultimodalTokenizationFailed,
+            },
+            ExpectedReport {
+                rejection: TextGenerationError::NoSequenceSlotAvailable,
+                into_result: GeneratedTokenResult::NoSequenceSlotAvailable,
+            },
+            ExpectedReport {
+                rejection: TextGenerationError::PromptTokenizationFailed(
+                    StringToTokenError::NulError(CString::new("a \0 byte").unwrap_err()),
+                ),
+                into_result: GeneratedTokenResult::PromptTokenizationFailed,
+            },
+            ExpectedReport {
+                rejection: TextGenerationError::SamplerChainCreationFailed(
+                    SamplingError::SamplerUnavailable { sampler: "grammar" },
+                ),
+                into_result: GeneratedTokenResult::SamplerChainCreationFailed,
+            },
+            ExpectedReport {
+                rejection: TextGenerationError::TokenDataUnreadable(LogitsError::NullLogits),
+                into_result: GeneratedTokenResult::SamplerError,
+            },
+        ] {
+            let expected_message = agent_message(&rejection.to_string());
+
+            assert_eq!(reported(rejection), into_result(expected_message));
+        }
+    }
+}
diff --git a/paddler_agent_text_generation/src/text_generation_pipeline.rs b/paddler_agent_text_generation/src/text_generation_pipeline.rs
new file mode 100644
index 000000000..51a6daf35
--- /dev/null
+++ b/paddler_agent_text_generation/src/text_generation_pipeline.rs
@@ -0,0 +1,357 @@
+use std::sync::Arc;
+
+use encoding_rs::Decoder;
+use encoding_rs::UTF_8;
+use llama_cpp_bindings::SampledToken;
+use llama_cpp_bindings::context::LlamaContext;
+use llama_cpp_bindings::error::TokenToStringError;
+use llama_cpp_bindings::llama_batch::LlamaBatch;
+use llama_cpp_bindings::model::LlamaModel;
+use llama_cpp_bindings::mtmd::micro_batch_tokens;
+use llama_cpp_bindings::token::LlamaToken;
+use log::info;
+use log::warn;
+use tokio::sync::mpsc::UnboundedSender;
+use tokio::sync::mpsc::unbounded_channel;
+use tokio_util::sync::CancellationToken;
+use tokio_util::task::TaskTracker;
+
+use paddler_agent_runtime::continue_startup_unless_shutting_down::continue_startup_unless_shutting_down;
+use paddler_agent_runtime::inference_runtime_context::InferenceRuntimeContext;
+use paddler_agent_runtime::loaded_llama_model::LoadedLlamaModel;
+use paddler_agent_runtime::run_scheduler_unless_startup_abandoned::run_scheduler_unless_startup_abandoned;
+use paddler_agent_runtime::scheduler_handle::SchedulerHandle;
+use paddler_agent_runtime::scheduler_ready::SchedulerReady;
+use paddler_agent_runtime::scheduler_request_preparer::SchedulerRequestPreparer;
+use paddler_agent_runtime::start_scheduler_thread::start_scheduler_thread;
+use paddler_agent_runtime::startup_outcome::StartupOutcome;
+use paddler_agent_status::agent_issue_fix::AgentIssueFix;
+use paddler_inference_parameters::model_runtime_parameters::ModelRuntimeParameters;
+use paddler_messaging::agent_issue::AgentIssue;
+use paddler_messaging::agent_issue_params::chat_template_does_not_compile_params::ChatTemplateDoesNotCompileParams;
+use paddler_messaging::agent_issue_params::model_path::ModelPath;
+use paddler_messaging::chat_template::ChatTemplate;
+use paddler_messaging::chat_template_source::ChatTemplateSource;
+use paddler_messaging::inference_capability::InferenceCapability;
+
+use crate::chat_prompt_renderer::ChatPromptRenderer;
+use crate::chat_template_renderer::ChatTemplateRenderer;
+use crate::continuous_batch_scheduler::ContinuousBatchScheduler;
+use crate::continuous_batch_scheduler_context::ContinuousBatchSchedulerContext;
+use crate::continuous_batch_scheduler_params::ContinuousBatchSchedulerParams;
+use crate::conversation_prompt_support::ConversationPromptSupport;
+use crate::conversation_support::ConversationSupport;
+use crate::generation_request_preparer::GenerationRequestPreparer;
+use crate::image_input::ImageInput;
+use crate::model_bos_token::ModelBosToken;
+use crate::multimodal_projection::MultimodalProjection;
+use crate::multimodal_prompt_support::MultimodalPromptSupport;
+use crate::prompt_tokenizer::PromptTokenizer;
+use crate::resident_multimodal_context::ResidentMultimodalContext;
+use crate::sampler_chain_factory::SamplerChainFactory;
+use crate::text_generation_error::TextGenerationError;
+use crate::text_generation_settings::TextGenerationSettings;
+
+fn special_token_text(
+    model: &LlamaModel,
+    special_token: Option<LlamaToken>,
+    decoder: &mut Decoder,
+) -> Result<String, TokenToStringError> {
+    special_token.map_or_else(
+        || Ok(String::new()),
+        |token| model.token_to_piece(&SampledToken::Content(token), decoder, true, None),
+    )
+}
+
+pub struct TextGenerationPipeline {
+    pub inference_runtime_context: InferenceRuntimeContext,
+    pub loaded_llama_model: LoadedLlamaModel,
+    pub model_runtime_parameters: ModelRuntimeParameters,
+    pub text_generation_settings: TextGenerationSettings,
+}
+
+impl TextGenerationPipeline {
+    pub async fn spawn(
+        self,
+        cancellation_token: &CancellationToken,
+    ) -> Result<StartupOutcome<SchedulerHandle<GenerationRequestPreparer>>, TextGenerationError>
+    {
+        start_scheduler_thread(
+            cancellation_token,
+            move |agent_shutdown, scheduler_startup_tx| {
+                self.run_scheduler(agent_shutdown, scheduler_startup_tx)
+            },
+        )
+        .await
+    }
+
+    fn run_scheduler(
+        self,
+        agent_shutdown: &CancellationToken,
+        scheduler_startup_tx: &UnboundedSender<
+            Result<SchedulerReady<GenerationRequestPreparer>, TextGenerationError>,
+        >,
+    ) -> Result<(), TextGenerationError> {
+        let desired_slots_total = self
+            .inference_runtime_context
+            .slot_aggregated_status
+            .desired_slots_total;
+        let loaded_llama_model = self.loaded_llama_model.clone();
+
+        continue_startup_unless_shutting_down(agent_shutdown, || {
+            let streaming_markers = loaded_llama_model
+                .model
+                .streaming_markers()
+                .map_err(TextGenerationError::StreamingMarkersUnavailable)?;
+            let llama_context = loaded_llama_model.create_llama_context(
+                &self.inference_runtime_context,
+                InferenceCapability::TextGeneration,
+                loaded_llama_model
+                    .llama_context_settings(
+                        &self.model_runtime_parameters,
+                        u32::from(desired_slots_total),
+                    )
+                    .into_llama_context_params(),
+            )?;
+            let sequence_context_size = llama_context.n_ctx_seq();
+
+            self.text_generation_settings
+                .sampling_parameters
+                .ensure_penalty_window_fits(sequence_context_size)
+                .map_err(TextGenerationError::SamplingParametersUnfitForTheSequenceContext)?;
+
+            let prompt_tokenizer = PromptTokenizer {
+                model: loaded_llama_model.model.clone(),
+                model_bos_token: ModelBosToken::of(&loaded_llama_model.model),
+                sequence_context_size,
+            };
+
+            continue_startup_unless_shutting_down(agent_shutdown, || {
+                let conversation_support =
+                    self.conversation_support(&llama_context, &prompt_tokenizer)?;
+                let batch = LlamaBatch::new(
+                    self.model_runtime_parameters.n_batch.capacity(),
+                    i32::from(desired_slots_total),
+                );
+
+                continue_startup_unless_shutting_down(agent_shutdown, || {
+                    loaded_llama_model
+                        .warm_up_and_serve(
+                            llama_context,
+                            batch,
+                            desired_slots_total,
+                            |llama_context, batch| {
+                                let (scheduler_message_tx, scheduler_message_rx) =
+                                    unbounded_channel();
+
+                                run_scheduler_unless_startup_abandoned(
+                                    scheduler_startup_tx,
+                                    Ok(SchedulerReady {
+                                        request_preparer: Arc::new(SchedulerRequestPreparer {
+                                            agent_name: self
+                                                .inference_runtime_context
+                                                .agent_name
+                                                .clone(),
+                                            preparation: GenerationRequestPreparer {
+                                                conversation_support,
+                                                image_resize_to_fit: self
+                                                    .text_generation_settings
+                                                    .image_resize_to_fit,
+                                                loaded_llama_model: loaded_llama_model.clone(),
+                                                prompt_tokenizer,
+                                                sampler_chain_factory: SamplerChainFactory {
+                                                    sampling_parameters: self
+                                                        .text_generation_settings
+                                                        .sampling_parameters
+                                                        .clone(),
+                                                    n_vocab: loaded_llama_model.model.n_vocab(),
+                                                },
+                                                streaming_markers,
+                                            },
+                                            preparation_tasks: TaskTracker::new(),
+                                            scheduler_message_tx,
+                                        }),
+                                    }),
+                                    || {
+                                        ContinuousBatchScheduler::new(
+                                            ContinuousBatchSchedulerParams {
+                                                agent_shutdown: agent_shutdown.clone(),
+                                                batch,
+                                                scheduler_message_rx,
+                                                llama_context,
+                                                scheduler_context:
+                                                    ContinuousBatchSchedulerContext {
+                                                        agent_name: self
+                                                            .inference_runtime_context
+                                                            .agent_name,
+                                                        desired_slots_total,
+                                                        model: &loaded_llama_model.model,
+                                                        n_batch: self
+                                                            .model_runtime_parameters
+                                                            .n_batch,
+                                                        sequence_context_size,
+                                                    },
+                                            },
+                                        )
+                                        .run();
+                                    },
+                                );
+                            },
+                        )
+                        .map_err(TextGenerationError::from)
+                })
+            })
+        })
+    }
+
+    fn conversation_support(
+        &self,
+        llama_context: &LlamaContext<'_>,
+        prompt_tokenizer: &PromptTokenizer,
+    ) -> Result<ConversationSupport, TextGenerationError> {
+        let chat_template = match &self.text_generation_settings.chat_template_source {
+            ChatTemplateSource::Override(chat_template_override) => chat_template_override.clone(),
+            ChatTemplateSource::EmbeddedInModel => {
+                let Some(embedded_chat_template) = self
+                    .loaded_llama_model
+                    .model_metadata
+                    .embedded_chat_template()
+                else {
+                    warn!(
+                        "{:?}: the model embeds no chat template, so it serves no conversations",
+                        self.inference_runtime_context.agent_name
+                    );
+
+                    return Ok(ConversationSupport::ChatTemplateMissing(
+                        self.loaded_llama_model
+                            .llama_model_settings
+                            .reported_model_path(),
+                    ));
+                };
+
+                embedded_chat_template
+            }
+        };
+        let chat_prompt_renderer = self.chat_prompt_renderer(chat_template)?;
+
+        Ok(ConversationSupport::Available(Box::new(
+            ConversationPromptSupport {
+                image_input: self.image_input(llama_context, prompt_tokenizer)?,
+                chat_prompt_renderer,
+            },
+        )))
+    }
+
+    fn chat_prompt_renderer(
+        &self,
+        chat_template: ChatTemplate,
+    ) -> Result<ChatPromptRenderer, TextGenerationError> {
+        let model = &self.loaded_llama_model.model;
+        let chat_template_renderer = self.chat_template_renderer(chat_template)?;
+        let mut special_token_decoder = UTF_8.new_decoder();
+
+        Ok(ChatPromptRenderer {
+            chat_template_renderer,
+            token_bos_str: special_token_text(model, model.token_bos(), &mut special_token_decoder)
+                .map_err(TextGenerationError::SpecialTokenUnrenderable)?,
+            token_eos_str: special_token_text(model, model.token_eos(), &mut special_token_decoder)
+                .map_err(TextGenerationError::SpecialTokenUnrenderable)?,
+            token_nl_str: special_token_text(model, model.token_nl(), &mut special_token_decoder)
+                .map_err(TextGenerationError::SpecialTokenUnrenderable)?,
+        })
+    }
+
+    fn chat_template_renderer(
+        &self,
+        chat_template: ChatTemplate,
+    ) -> Result<ChatTemplateRenderer, TextGenerationError> {
+        let model_issue_path = self
+            .loaded_llama_model
+            .llama_model_settings
+            .reported_model_path();
+        let template_content = chat_template.content.clone();
+        let slot_aggregated_status = &self.inference_runtime_context.slot_aggregated_status;
+
+        match ChatTemplateRenderer::new(chat_template) {
+            Ok(chat_template_renderer) => {
+                slot_aggregated_status
+                    .register_fix(&AgentIssueFix::ChatTemplateIsCompiled(model_issue_path));
+
+                Ok(chat_template_renderer)
+            }
+            Err(compile_error) => {
+                slot_aggregated_status.register_pipeline_issue(
+                    AgentIssue::ChatTemplateDoesNotCompile(ChatTemplateDoesNotCompileParams {
+                        error: format!("{compile_error}"),
+                        model_path: model_issue_path,
+                        template_content,
+                    }),
+                );
+
+                Err(TextGenerationError::ChatTemplateDoesNotCompile(
+                    compile_error,
+                ))
+            }
+        }
+    }
+
+    fn image_input(
+        &self,
+        llama_context: &LlamaContext<'_>,
+        prompt_tokenizer: &PromptTokenizer,
+    ) -> Result<ImageInput, TextGenerationError> {
+        let MultimodalProjection::File(multimodal_projection_path) =
+            &self.text_generation_settings.multimodal_projection
+        else {
+            return Ok(ImageInput::Unsupported);
+        };
+        let multimodal_projection_issue_path =
+            ModelPath::from(multimodal_projection_path.as_path());
+        let slot_aggregated_status = &self.inference_runtime_context.slot_aggregated_status;
+        let vram_occupancy = self
+            .inference_runtime_context
+            .vram_residencies
+            .multimodal_projection
+            .try_occupy()?;
+
+        match ResidentMultimodalContext::load(
+            multimodal_projection_path,
+            &self.loaded_llama_model.model,
+            vram_occupancy,
+        ) {
+            Ok(resident_multimodal_context) => {
+                slot_aggregated_status.register_fix(&AgentIssueFix::MultimodalProjectionIsLoaded(
+                    multimodal_projection_issue_path,
+                ));
+
+                info!(
+                    "Multimodal context initialized from: {}",
+                    multimodal_projection_path.display()
+                );
+
+                Ok(ImageInput::Supported(MultimodalPromptSupport {
+                    model_bos_token: prompt_tokenizer.model_bos_token,
+                    micro_batch_tokens: micro_batch_tokens(
+                        llama_context,
+                        self.model_runtime_parameters.n_batch.tokens(),
+                    ),
+                    multimodal_context: Arc::new(resident_multimodal_context),
+                    n_batch: self.model_runtime_parameters.n_batch.tokens_i32(),
+                    sequence_context_size: prompt_tokenizer.sequence_context_size,
+                }))
+            }
+            Err(source) => {
+                slot_aggregated_status.register_pipeline_issue(
+                    AgentIssue::MultimodalProjectionCannotBeLoaded(
+                        multimodal_projection_issue_path,
+                    ),
+                );
+
+                Err(TextGenerationError::MultimodalProjectionLoadFailed {
+                    multimodal_projection_path: multimodal_projection_path.clone(),
+                    source,
+                })
+            }
+        }
+    }
+}
diff --git a/paddler_agent_text_generation/src/text_generation_request.rs b/paddler_agent_text_generation/src/text_generation_request.rs
new file mode 100644
index 000000000..baf0b0a70
--- /dev/null
+++ b/paddler_agent_text_generation/src/text_generation_request.rs
@@ -0,0 +1,15 @@
+use paddler_agent_runtime::agent_request::AgentRequest;
+use paddler_messaging::generated_token_result::GeneratedTokenResult;
+use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
+use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters_schema::validated_parameters_schema::ValidatedParametersSchema;
+use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
+
+pub enum TextGenerationRequest {
+    ContinueFromConversationHistory(
+        AgentRequest<
+            ContinueFromConversationHistoryParams<ValidatedParametersSchema>,
+            GeneratedTokenResult,
+        >,
+    ),
+    ContinueFromRawPrompt(AgentRequest<ContinueFromRawPromptParams, GeneratedTokenResult>),
+}
diff --git a/paddler_agent_text_generation/src/text_generation_settings.rs b/paddler_agent_text_generation/src/text_generation_settings.rs
new file mode 100644
index 000000000..feb2b622f
--- /dev/null
+++ b/paddler_agent_text_generation/src/text_generation_settings.rs
@@ -0,0 +1,14 @@
+use std::num::NonZeroU32;
+
+use paddler_inference_parameters::sampling_parameters::SamplingParameters;
+use paddler_messaging::chat_template_source::ChatTemplateSource;
+
+use crate::multimodal_projection::MultimodalProjection;
+
+#[derive(Clone, Debug, PartialEq)]
+pub struct TextGenerationSettings {
+    pub chat_template_source: ChatTemplateSource,
+    pub image_resize_to_fit: NonZeroU32,
+    pub multimodal_projection: MultimodalProjection,
+    pub sampling_parameters: SamplingParameters,
+}
diff --git a/paddler_agent/src/token_classification.rs b/paddler_agent_text_generation/src/token_classification.rs
similarity index 100%
rename from paddler_agent/src/token_classification.rs
rename to paddler_agent_text_generation/src/token_classification.rs
diff --git a/paddler_agent_text_generation/src/token_sampling.rs b/paddler_agent_text_generation/src/token_sampling.rs
new file mode 100644
index 000000000..13b803d33
--- /dev/null
+++ b/paddler_agent_text_generation/src/token_sampling.rs
@@ -0,0 +1,58 @@
+use llama_cpp_bindings::context::LlamaContext;
+use llama_cpp_bindings::error::SamplerApplyError;
+use llama_cpp_bindings::sampling::LlamaSampler;
+use llama_cpp_bindings::token::data_array::LlamaTokenDataArray;
+
+use crate::grammar_sampling::GrammarSampling;
+use crate::sampling_outcome::SamplingOutcome;
+use crate::text_generation_error::TextGenerationError;
+
+pub struct TokenSampling {
+    pub chain: LlamaSampler,
+    pub grammar: GrammarSampling,
+}
+
+impl TokenSampling {
+    pub fn sample(
+        &mut self,
+        llama_context: &LlamaContext,
+        batch_index: i32,
+        candidates: &mut LlamaTokenDataArray,
+    ) -> Result<SamplingOutcome, TextGenerationError> {
+        llama_context
+            .fill_token_data_array_ith(batch_index, candidates)
+            .map_err(TextGenerationError::TokenDataUnreadable)?;
+        match self.grammar.constrain(candidates) {
+            Ok(()) => {}
+            Err(SamplerApplyError::EveryCandidateExcluded) => {
+                return Ok(SamplingOutcome::GrammarRejectedModelOutput(
+                    SamplerApplyError::EveryCandidateExcluded.to_string(),
+                ));
+            }
+            Err(constraint_error) => {
+                return Err(TextGenerationError::GrammarConstraintFailed(
+                    constraint_error,
+                ));
+            }
+        }
+        candidates
+            .apply_sampler(&self.chain)
+            .map_err(TextGenerationError::SamplerChainApplicationFailed)?;
+
+        let Some(token) = candidates.selected_token() else {
+            return Ok(SamplingOutcome::AllCandidatesEliminated);
+        };
+
+        self.chain
+            .accept(token)
+            .map_err(TextGenerationError::SamplerChainRejectedToken)?;
+
+        if let Err(grammar_rejection) = self.grammar.accept(token) {
+            return Ok(SamplingOutcome::GrammarRejectedModelOutput(
+                grammar_rejection.to_string(),
+            ));
+        }
+
+        Ok(SamplingOutcome::Token(token))
+    }
+}
diff --git a/paddler_agent_text_generation/src/tool_call_buffer.rs b/paddler_agent_text_generation/src/tool_call_buffer.rs
new file mode 100644
index 000000000..101559da9
--- /dev/null
+++ b/paddler_agent_text_generation/src/tool_call_buffer.rs
@@ -0,0 +1,36 @@
+use llama_cpp_bindings::token_piece::TokenPiece;
+
+#[derive(Debug, Default)]
+pub struct ToolCallBuffer {
+    pub accumulated: String,
+    pub holds_visible_text: bool,
+}
+
+impl ToolCallBuffer {
+    pub fn append(&mut self, piece: &TokenPiece) {
+        self.accumulated.push_str(piece.raw());
+        self.holds_visible_text |= !piece.visible().is_empty();
+    }
+
+    #[must_use]
+    pub const fn is_empty(&self) -> bool {
+        self.accumulated.is_empty()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use llama_cpp_bindings::token_piece::TokenPiece;
+
+    use super::ToolCallBuffer;
+
+    #[test]
+    fn appending_keeps_the_raw_text_of_markers_and_of_visible_pieces() {
+        let mut buffer = ToolCallBuffer::default();
+
+        buffer.append(&TokenPiece::Marker("<tool_call>".to_owned()));
+        buffer.append(&TokenPiece::Visible("{\"name\":\"x\"}".to_owned()));
+
+        assert_eq!(buffer.accumulated, "<tool_call>{\"name\":\"x\"}");
+    }
+}
diff --git a/paddler_agent/src/tool_call_event.rs b/paddler_agent_text_generation/src/tool_call_event.rs
similarity index 85%
rename from paddler_agent/src/tool_call_event.rs
rename to paddler_agent_text_generation/src/tool_call_event.rs
index 31519d00d..03737f175 100644
--- a/paddler_agent/src/tool_call_event.rs
+++ b/paddler_agent_text_generation/src/tool_call_event.rs
@@ -1,13 +1,13 @@
-use llama_cpp_bindings::ParsedToolCall;
 use llama_cpp_bindings::error::ParseChatMessageError;
 
 use paddler_messaging::generated_token_result::GeneratedTokenResult;
+use paddler_messaging::generated_tool_call::GeneratedToolCall;
 use paddler_messaging::raw_tool_call_tokens::RawToolCallTokens;
 use paddler_tool_call_validator::tool_call_validation_error::ToolCallValidationError;
 
 #[derive(Debug)]
 pub enum ToolCallEvent {
-    Resolved(Vec<ParsedToolCall>),
+    Resolved(Vec<GeneratedToolCall>),
     ParseFailed(ParseChatMessageError),
     ValidationFailed(Vec<ToolCallValidationError>),
     UnrecognizedFormat(RawToolCallTokens),
@@ -36,12 +36,12 @@ impl ToolCallEvent {
 
 #[cfg(test)]
 mod tests {
-    use llama_cpp_bindings::ParsedToolCall;
     use llama_cpp_bindings::ToolCallArguments;
     use llama_cpp_bindings::error::ParseChatMessageError;
     use serde_json::json;
 
     use paddler_messaging::generated_token_result::GeneratedTokenResult;
+    use paddler_messaging::generated_tool_call::GeneratedToolCall;
     use paddler_messaging::raw_tool_call_tokens::RawToolCallTokens;
     use paddler_tool_call_validator::tool_call_validation_error::ToolCallValidationError;
 
@@ -49,15 +49,16 @@ mod tests {
 
     #[test]
     fn resolved_converts_to_tool_call_parsed() {
-        let parsed = ParsedToolCall::new(
-            "id".to_owned(),
-            "tool".to_owned(),
-            ToolCallArguments::ValidJson(json!({})),
-        );
+        let generated = GeneratedToolCall {
+            arguments: ToolCallArguments::ValidJson(json!({})),
+            id: "id".to_owned(),
+            index: 0,
+            name: "tool".to_owned(),
+        };
 
         assert_eq!(
-            ToolCallEvent::Resolved(vec![parsed.clone()]).into_generated_token_result(),
-            GeneratedTokenResult::ToolCallParsed(vec![parsed])
+            ToolCallEvent::Resolved(vec![generated.clone()]).into_generated_token_result(),
+            GeneratedTokenResult::ToolCallParsed(vec![generated])
         );
     }
 
diff --git a/paddler_agent/src/tool_call_handling.rs b/paddler_agent_text_generation/src/tool_call_handling.rs
similarity index 96%
rename from paddler_agent/src/tool_call_handling.rs
rename to paddler_agent_text_generation/src/tool_call_handling.rs
index 46fbf83f6..2607f370d 100644
--- a/paddler_agent/src/tool_call_handling.rs
+++ b/paddler_agent_text_generation/src/tool_call_handling.rs
@@ -15,7 +15,7 @@ impl ToolCallHandling {
         if let Self::Parsed(pipeline) = self
             && matches!(classified.sampled_token, SampledToken::ToolCall(_))
         {
-            pipeline.feed(classified.piece.raw());
+            pipeline.feed(&classified.piece);
         }
     }
 
diff --git a/paddler_agent_text_generation/src/tool_call_pipeline.rs b/paddler_agent_text_generation/src/tool_call_pipeline.rs
new file mode 100644
index 000000000..870708c45
--- /dev/null
+++ b/paddler_agent_text_generation/src/tool_call_pipeline.rs
@@ -0,0 +1,106 @@
+use std::mem::take;
+
+use llama_cpp_bindings::ChatMessageParseOutcome;
+use llama_cpp_bindings::ChatMessageParser;
+use llama_cpp_bindings::ParsedToolCall;
+use llama_cpp_bindings::RawChatMessage;
+use llama_cpp_bindings::token_piece::TokenPiece;
+
+use paddler_messaging::generated_tool_call::GeneratedToolCall;
+use paddler_messaging::raw_tool_call_tokens::RawToolCallTokens;
+use paddler_tool_call_validator::tool_call_validator::ToolCallValidator;
+
+use crate::tool_call_buffer::ToolCallBuffer;
+use crate::tool_call_event::ToolCallEvent;
+
+pub struct ToolCallPipeline {
+    buffer: ToolCallBuffer,
+    chat_message_parser: ChatMessageParser,
+    tool_calls_generated: usize,
+    validator: ToolCallValidator,
+}
+
+impl ToolCallPipeline {
+    #[must_use]
+    pub fn new(chat_message_parser: ChatMessageParser, validator: ToolCallValidator) -> Self {
+        Self {
+            buffer: ToolCallBuffer::default(),
+            chat_message_parser,
+            tool_calls_generated: 0,
+            validator,
+        }
+    }
+
+    pub fn feed(&mut self, piece: &TokenPiece) {
+        self.buffer.append(piece);
+    }
+
+    #[must_use]
+    pub const fn buffer_is_empty(&self) -> bool {
+        self.buffer.is_empty()
+    }
+
+    pub fn finalize(&mut self) -> ToolCallEvent {
+        let ToolCallBuffer {
+            accumulated: input,
+            holds_visible_text,
+        } = take(&mut self.buffer);
+
+        if input.is_empty() {
+            return ToolCallEvent::Resolved(Vec::new());
+        }
+
+        match self.chat_message_parser.parse(&input, false) {
+            Ok(ChatMessageParseOutcome::Recognized(parsed))
+                if parsed.tool_calls.is_empty() && holds_visible_text =>
+            {
+                ToolCallEvent::UnrecognizedFormat(RawToolCallTokens {
+                    text: input,
+                    ffi_error_message: String::new(),
+                })
+            }
+            Ok(ChatMessageParseOutcome::Recognized(parsed)) => {
+                self.validate_resolved(parsed.tool_calls)
+            }
+            Ok(ChatMessageParseOutcome::Unrecognized(RawChatMessage {
+                text,
+                ffi_error_message,
+                ..
+            })) => ToolCallEvent::UnrecognizedFormat(RawToolCallTokens {
+                text,
+                ffi_error_message,
+            }),
+            Err(parse_error) => ToolCallEvent::ParseFailed(parse_error),
+        }
+    }
+
+    fn validate_resolved(&mut self, tool_calls: Vec<ParsedToolCall>) -> ToolCallEvent {
+        let validation_errors: Vec<_> = tool_calls
+            .iter()
+            .filter_map(|tool_call| self.validator.validate(tool_call).err())
+            .collect();
+
+        if validation_errors.is_empty() {
+            ToolCallEvent::Resolved(self.number_across_the_generation(tool_calls))
+        } else {
+            ToolCallEvent::ValidationFailed(validation_errors)
+        }
+    }
+
+    fn number_across_the_generation(
+        &mut self,
+        tool_calls: Vec<ParsedToolCall>,
+    ) -> Vec<GeneratedToolCall> {
+        tool_calls
+            .into_iter()
+            .map(|tool_call| {
+                let generated_tool_call =
+                    GeneratedToolCall::at_index(tool_call, self.tool_calls_generated);
+
+                self.tool_calls_generated += 1;
+
+                generated_tool_call
+            })
+            .collect()
+    }
+}
diff --git a/paddler_balancer/Cargo.toml b/paddler_balancer/Cargo.toml
index 2028f9b34..6b4085a17 100644
--- a/paddler_balancer/Cargo.toml
+++ b/paddler_balancer/Cargo.toml
@@ -1,5 +1,6 @@
 [package]
 name = "paddler_balancer"
+autotests = false
 authors.workspace = true
 description.workspace = true
 edition.workspace = true
@@ -15,12 +16,10 @@ actix-web-lab = { workspace = true }
 actix-ws = { workspace = true }
 anyhow = { workspace = true }
 askama = { workspace = true, optional = true }
-async-stream = { workspace = true }
 async-trait = { workspace = true }
 bytes = { workspace = true }
 cadence = { workspace = true }
 dashmap = { workspace = true }
-esbuild-metafile = { workspace = true, optional = true }
 futures = { workspace = true }
 futures-util = { workspace = true }
 indoc = { workspace = true }
@@ -28,16 +27,21 @@ llama-cpp-bindings-types = { workspace = true }
 log = { workspace = true }
 mime_guess = { workspace = true, optional = true }
 nanoid = { workspace = true }
+paddler_frame_writer = { workspace = true }
 paddler_inference_parameters = { workspace = true }
 paddler_messaging = { workspace = true }
+paddler_openai_translation = { workspace = true }
+paddler_outbox = { workspace = true }
 paddler_request_registry = { workspace = true }
+paddler_spawn_resuming_panic = { workspace = true }
 paddler_state_database = { workspace = true }
+paddler_typesafe_translation = { workspace = true }
 parking_lot = { workspace = true }
+rand = { workspace = true }
 rust-embed = { workspace = true, optional = true }
 serde = { workspace = true }
 serde_json = { workspace = true }
 socket2 = { workspace = true }
-thiserror = { workspace = true }
 tokio = { workspace = true }
 tokio-stream = { workspace = true }
 tokio-util = { workspace = true }
@@ -47,15 +51,18 @@ trzcina = { workspace = true }
 default = []
 web_admin_panel = [
     "dep:askama",
-    "dep:esbuild-metafile",
     "dep:mime_guess",
+    "dep:paddler_dashboard_assets",
     "dep:rust-embed",
 ]
 
+[build-dependencies]
+paddler_dashboard_assets = { workspace = true, optional = true }
+
 [dev-dependencies]
+paddler_blocking_thread_fixture = { workspace = true }
 paddler_openai_response_format_validator = { workspace = true }
 tempfile = { workspace = true }
-tokio-test = { workspace = true }
 
 [lints]
 workspace = true
diff --git a/paddler_balancer/build.rs b/paddler_balancer/build.rs
new file mode 100644
index 000000000..c5be687c0
--- /dev/null
+++ b/paddler_balancer/build.rs
@@ -0,0 +1,23 @@
+#[cfg(feature = "web_admin_panel")]
+use std::env::var_os;
+#[cfg(feature = "web_admin_panel")]
+use std::io::stdout;
+#[cfg(feature = "web_admin_panel")]
+use std::path::Path;
+
+#[cfg(feature = "web_admin_panel")]
+use paddler_dashboard_assets::dashboard_assets_build::DashboardAssetsBuild;
+#[cfg(feature = "web_admin_panel")]
+use paddler_dashboard_assets::dashboard_assets_error::DashboardAssetsError;
+
+#[cfg(feature = "web_admin_panel")]
+fn main() -> Result<(), DashboardAssetsError> {
+    DashboardAssetsBuild {
+        metafile_path: Path::new(env!("CARGO_MANIFEST_DIR")).join("../esbuild-meta.json"),
+        out_dir: var_os("OUT_DIR"),
+    }
+    .run(stdout())
+}
+
+#[cfg(not(feature = "web_admin_panel"))]
+fn main() {}
diff --git a/paddler_balancer/src/accept_newer_status_report.rs b/paddler_balancer/src/accept_newer_status_report.rs
new file mode 100644
index 000000000..346c23647
--- /dev/null
+++ b/paddler_balancer/src/accept_newer_status_report.rs
@@ -0,0 +1,74 @@
+use paddler_messaging::slot_aggregated_status_snapshot::SlotAggregatedStatusSnapshot;
+
+use crate::agent_controller_update_result::AgentControllerUpdateResult;
+
+pub fn accept_newer_status_report(
+    current_snapshot: &mut SlotAggregatedStatusSnapshot,
+    reported_snapshot: SlotAggregatedStatusSnapshot,
+) -> AgentControllerUpdateResult {
+    if reported_snapshot.version < current_snapshot.version {
+        return AgentControllerUpdateResult::NoMeaningfulChanges;
+    }
+
+    let status_changed = reported_snapshot.status != current_snapshot.status;
+
+    *current_snapshot = reported_snapshot;
+
+    if status_changed {
+        AgentControllerUpdateResult::Updated
+    } else {
+        AgentControllerUpdateResult::NoMeaningfulChanges
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use paddler_messaging::agent_status::AgentStatus;
+    use paddler_messaging::slot_aggregated_status_snapshot::SlotAggregatedStatusSnapshot;
+
+    use super::accept_newer_status_report;
+    use crate::agent_controller_update_result::AgentControllerUpdateResult;
+
+    fn reported(version: u64, desired_slots_total: u16) -> SlotAggregatedStatusSnapshot {
+        SlotAggregatedStatusSnapshot {
+            status: AgentStatus {
+                desired_slots_total,
+                ..AgentStatus::default()
+            },
+            version,
+        }
+    }
+
+    #[test]
+    fn keeps_the_newer_status_when_an_older_report_arrives_late() {
+        let mut current_snapshot = reported(2, 4);
+
+        assert_eq!(
+            accept_newer_status_report(&mut current_snapshot, reported(1, 8)),
+            AgentControllerUpdateResult::NoMeaningfulChanges
+        );
+        assert_eq!(current_snapshot, reported(2, 4));
+    }
+
+    #[test]
+    fn reports_no_change_for_a_newer_report_with_the_same_status() {
+        let mut current_snapshot = reported(1, 4);
+
+        assert_eq!(
+            accept_newer_status_report(&mut current_snapshot, reported(2, 4)),
+            AgentControllerUpdateResult::NoMeaningfulChanges
+        );
+        assert_eq!(current_snapshot.version, 2);
+    }
+
+    #[test]
+    fn takes_a_newer_report_with_a_changed_status() {
+        let mut current_snapshot = reported(1, 4);
+
+        assert_eq!(
+            accept_newer_status_report(&mut current_snapshot, reported(2, 8)),
+            AgentControllerUpdateResult::Updated
+        );
+        assert_eq!(current_snapshot, reported(2, 8));
+    }
+}
diff --git a/paddler_balancer/src/agent_controller.rs b/paddler_balancer/src/agent_controller.rs
index 371674a07..b9f446e73 100644
--- a/paddler_balancer/src/agent_controller.rs
+++ b/paddler_balancer/src/agent_controller.rs
@@ -1,14 +1,12 @@
+use std::num::NonZeroU16;
 use std::sync::Arc;
 use std::sync::atomic::AtomicU64;
 
-use nanoid::nanoid;
+use parking_lot::Mutex;
 use parking_lot::RwLock;
-use tokio::sync::mpsc;
-use tokio::sync::mpsc::error::SendError;
 use tokio_util::sync::CancellationToken;
 
 use paddler_messaging::agent_controller_snapshot::AgentControllerSnapshot;
-use paddler_messaging::agent_desired_state::AgentDesiredState;
 use paddler_messaging::atomic_value::AtomicValue;
 use paddler_messaging::chat_template::ChatTemplate;
 use paddler_messaging::jsonrpc::request_envelope::RequestEnvelope;
@@ -18,71 +16,102 @@ use paddler_messaging::management_socket::agent::request::Request as AgentJsonRp
 use paddler_messaging::model_metadata::ModelMetadata;
 use paddler_messaging::produces_snapshot::ProducesSnapshot;
 use paddler_messaging::slot_aggregated_status_snapshot::SlotAggregatedStatusSnapshot;
+use paddler_outbox::outbox::Outbox;
 
+use crate::accept_newer_status_report::accept_newer_status_report;
+use crate::agent_controller_params::AgentControllerParams;
 use crate::agent_controller_update_result::AgentControllerUpdateResult;
 use crate::agent_desired_state_notification::agent_desired_state_notification;
-use crate::agent_relay_error::AgentRelayError;
 use crate::agent_response_receiver::AgentResponseReceiver;
 use crate::agent_response_senders::AgentResponseSenders;
 use crate::agent_streaming_request::AgentStreamingRequest;
-use crate::desired_state_delivery::DesiredStateDelivery;
+use crate::balancer_applicable_state::BalancerApplicableState;
 use crate::response_senders::ResponseSenders;
 
 pub struct AgentController {
-    pub agent_message_tx: mpsc::UnboundedSender<AgentJsonRpcMessage>,
+    pub agent_outbox: Outbox<AgentJsonRpcMessage>,
     pub agent_response_senders: AgentResponseSenders,
     pub connection_close: CancellationToken,
     pub id: String,
     pub name: Option<String>,
+    pub next_deliverable_desired_state_version: Mutex<u64>,
     pub reported_status: RwLock<SlotAggregatedStatusSnapshot>,
     pub slots_processing: AtomicValue<AtomicU64>,
 }
 
 impl AgentController {
-    pub fn get_chat_template_override(
-        &self,
-    ) -> Result<AgentResponseReceiver<Option<ChatTemplate>>, AgentRelayError> {
+    #[must_use]
+    pub fn new(
+        AgentControllerParams {
+            connection_close,
+            id,
+            name,
+            reported_status,
+        }: AgentControllerParams,
+    ) -> Self {
+        Self {
+            agent_outbox: Outbox::default(),
+            agent_response_senders: AgentResponseSenders::default(),
+            connection_close,
+            id,
+            name,
+            next_deliverable_desired_state_version: Mutex::new(0),
+            reported_status: RwLock::new(reported_status),
+            slots_processing: AtomicValue::<AtomicU64>::new(0),
+        }
+    }
+
+    #[must_use]
+    pub fn get_chat_template_override(&self) -> AgentResponseReceiver<Option<ChatTemplate>> {
         self.relay_request(
-            nanoid!(),
             AgentJsonRpcRequest::GetChatTemplateOverride,
             &self.agent_response_senders.chat_template_override,
         )
     }
 
-    pub fn get_model_metadata(
-        &self,
-    ) -> Result<AgentResponseReceiver<Option<ModelMetadata>>, AgentRelayError> {
+    #[must_use]
+    pub fn get_model_metadata(&self) -> AgentResponseReceiver<Option<ModelMetadata>> {
         self.relay_request(
-            nanoid!(),
             AgentJsonRpcRequest::GetModelMetadata,
             &self.agent_response_senders.model_metadata,
         )
     }
 
-    pub fn set_desired_state(&self, desired_state: AgentDesiredState) -> DesiredStateDelivery {
-        match self
-            .agent_message_tx
-            .send(agent_desired_state_notification(desired_state))
-        {
-            Ok(()) => DesiredStateDelivery::Delivered,
-            Err(SendError(_undelivered_message)) => DesiredStateDelivery::AgentDisconnected,
+    pub fn set_desired_state(
+        &self,
+        BalancerApplicableState {
+            agent_desired_state,
+            version,
+            ..
+        }: &BalancerApplicableState,
+    ) {
+        let mut next_deliverable_desired_state_version =
+            self.next_deliverable_desired_state_version.lock();
+
+        if *version < *next_deliverable_desired_state_version {
+            return;
         }
+
+        *next_deliverable_desired_state_version = version + 1;
+        self.agent_outbox.push(agent_desired_state_notification(
+            agent_desired_state.clone(),
+        ));
     }
 
-    pub fn stop_responding_to(&self, request_id: String) -> Result<(), AgentRelayError> {
-        self.send_to_agent(AgentJsonRpcMessage::Notification(
+    pub fn stop_responding_to(&self, request_id: String) {
+        self.agent_outbox.push(AgentJsonRpcMessage::Notification(
             AgentJsonRpcNotification::StopRespondingTo(request_id),
-        ))
+        ));
     }
 
+    #[must_use]
     pub fn stream_responses_to<TRequest: AgentStreamingRequest>(
         &self,
-        request_id: String,
         request: TRequest,
-    ) -> Result<AgentResponseReceiver<TRequest::Response>, AgentRelayError> {
+        reserved_slots: NonZeroU16,
+    ) -> AgentResponseReceiver<TRequest::Response> {
         self.relay_request(
-            request_id,
-            request.into(),
+            request.into_reserved_agent_request(reserved_slots),
             TRequest::response_senders(&self.agent_response_senders),
         )
     }
@@ -91,48 +120,26 @@ impl AgentController {
         &self,
         reported_snapshot: SlotAggregatedStatusSnapshot,
     ) -> AgentControllerUpdateResult {
-        let mut current_snapshot = self.reported_status.write();
-
-        if reported_snapshot.version < current_snapshot.version {
-            return AgentControllerUpdateResult::NoMeaningfulChanges;
-        }
-
-        let status_changed = reported_snapshot.status != current_snapshot.status;
-
-        *current_snapshot = reported_snapshot;
-
-        if status_changed {
-            AgentControllerUpdateResult::Updated
-        } else {
-            AgentControllerUpdateResult::NoMeaningfulChanges
-        }
+        accept_newer_status_report(&mut self.reported_status.write(), reported_snapshot)
     }
 
     fn relay_request<TResponse>(
         &self,
-        request_id: String,
         request: AgentJsonRpcRequest,
         response_senders: &Arc<ResponseSenders<TResponse>>,
-    ) -> Result<AgentResponseReceiver<TResponse>, AgentRelayError> {
-        AgentResponseReceiver::register(response_senders, request_id.clone()).and_then(
-            |agent_response_receiver| {
-                self.send_to_agent(AgentJsonRpcMessage::Request(RequestEnvelope {
-                    id: request_id,
-                    request,
-                }))
-                .map(|()| agent_response_receiver)
-            },
-        )
-    }
+    ) -> AgentResponseReceiver<TResponse> {
+        let agent_response_receiver = AgentResponseReceiver::register(response_senders);
+
+        self.agent_outbox
+            .push(AgentJsonRpcMessage::Request(RequestEnvelope {
+                id: agent_response_receiver
+                    .request_registration
+                    .request_id
+                    .clone(),
+                request,
+            }));
 
-    fn send_to_agent(&self, message: AgentJsonRpcMessage) -> Result<(), AgentRelayError> {
-        self.agent_message_tx
-            .send(message)
-            .map_err(
-                |SendError(_undelivered_message)| AgentRelayError::AgentSocketClosed {
-                    agent_id: self.id.clone(),
-                },
-            )
+        agent_response_receiver
     }
 }
 
@@ -151,132 +158,57 @@ impl ProducesSnapshot for AgentController {
 
 #[cfg(test)]
 mod tests {
-    use std::num::NonZeroU32;
-    use std::sync::atomic::AtomicU64;
-
-    use parking_lot::RwLock;
-    use tokio::sync::mpsc;
+    use futures::FutureExt as _;
+    use serde_json::Value;
+    use serde_json::to_value;
     use tokio_util::sync::CancellationToken;
 
-    use paddler_messaging::agent_status::AgentStatus;
-    use paddler_messaging::atomic_value::AtomicValue;
-    use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
+    use paddler_messaging::agent_desired_model::AgentDesiredModel;
+    use paddler_messaging::balancer_desired_state::BalancerDesiredState;
     use paddler_messaging::slot_aggregated_status_snapshot::SlotAggregatedStatusSnapshot;
 
     use super::AgentController;
-    use crate::agent_controller_update_result::AgentControllerUpdateResult;
-    use crate::agent_relay_error::AgentRelayError;
-    use crate::agent_response_senders::AgentResponseSenders;
-
-    fn reported(version: u64, slots_total: u64) -> SlotAggregatedStatusSnapshot {
-        SlotAggregatedStatusSnapshot {
-            status: AgentStatus {
-                slots_total,
-                ..AgentStatus::default()
-            },
-            version,
+    use crate::agent_controller_params::AgentControllerParams;
+    use crate::agent_desired_state_notification::agent_desired_state_notification;
+    use crate::balancer_applicable_state::BalancerApplicableState;
+
+    fn desired_state_with_model(model_uri: &str) -> BalancerDesiredState {
+        BalancerDesiredState {
+            model: AgentDesiredModel::Uri(model_uri.to_owned()),
+            ..BalancerDesiredState::default()
         }
     }
 
-    fn agent_controller_with(reported_status: SlotAggregatedStatusSnapshot) -> AgentController {
-        let (agent_message_tx, _agent_message_rx) = mpsc::unbounded_channel();
-
-        AgentController {
-            agent_message_tx,
-            agent_response_senders: AgentResponseSenders::default(),
+    #[test]
+    fn a_desired_state_arriving_after_a_newer_one_is_not_sent_to_the_agent() {
+        let replaced_state =
+            BalancerApplicableState::initial(desired_state_with_model("agent:///replaced.gguf"));
+        let newest_state =
+            replaced_state.superseded_by(desired_state_with_model("agent:///newest.gguf"));
+        let agent_controller = AgentController::new(AgentControllerParams {
             connection_close: CancellationToken::new(),
             id: "agent".to_owned(),
             name: None,
-            reported_status: RwLock::new(reported_status),
-            slots_processing: AtomicValue::<AtomicU64>::new(0),
-        }
-    }
+            reported_status: SlotAggregatedStatusSnapshot::default(),
+        });
 
-    #[test]
-    fn keeps_the_newer_status_when_an_older_report_arrives_late() {
-        let agent_controller = agent_controller_with(reported(2, 4));
-
-        assert_eq!(
-            agent_controller.update_from_slot_aggregated_status_snapshot(reported(1, 8)),
-            AgentControllerUpdateResult::NoMeaningfulChanges
-        );
-        assert_eq!(*agent_controller.reported_status.read(), reported(2, 4));
-    }
-
-    #[test]
-    fn reports_no_change_for_a_newer_report_with_the_same_status() {
-        let agent_controller = agent_controller_with(reported(1, 4));
-
-        assert_eq!(
-            agent_controller.update_from_slot_aggregated_status_snapshot(reported(2, 4)),
-            AgentControllerUpdateResult::NoMeaningfulChanges
-        );
-        assert_eq!(agent_controller.reported_status.read().version, 2);
-    }
-
-    #[test]
-    fn takes_a_newer_report_with_a_changed_status() {
-        let agent_controller = agent_controller_with(reported(1, 4));
+        agent_controller.set_desired_state(&newest_state);
+        agent_controller.set_desired_state(&replaced_state);
 
-        assert_eq!(
-            agent_controller.update_from_slot_aggregated_status_snapshot(reported(2, 8)),
-            AgentControllerUpdateResult::Updated
-        );
-        assert_eq!(*agent_controller.reported_status.read(), reported(2, 8));
-    }
+        let mut sent_messages = Vec::<Value>::new();
 
-    fn raw_prompt_params() -> ContinueFromRawPromptParams {
-        ContinueFromRawPromptParams {
-            grammar: None,
-            max_tokens: NonZeroU32::MIN,
-            raw_prompt: "hello".to_owned(),
+        while let Some(message) = agent_controller.agent_outbox.next().now_or_never() {
+            sent_messages.push(to_value(message).unwrap());
         }
-    }
-
-    #[test]
-    fn fails_to_ask_an_agent_whose_socket_is_gone_for_its_model_metadata() {
-        let agent_controller = agent_controller_with(reported(1, 4));
-
-        assert!(matches!(
-            agent_controller.get_model_metadata(),
-            Err(AgentRelayError::AgentSocketClosed { agent_id }) if agent_id == "agent"
-        ));
-    }
-
-    #[test]
-    fn releases_the_request_id_of_a_stream_to_an_agent_whose_socket_is_gone() {
-        let agent_controller = agent_controller_with(reported(1, 4));
 
-        let stream_result =
-            agent_controller.stream_responses_to("request".to_owned(), raw_prompt_params());
-
-        assert!(matches!(
-            stream_result,
-            Err(AgentRelayError::AgentSocketClosed { agent_id }) if agent_id == "agent"
-        ));
         assert_eq!(
-            agent_controller
-                .agent_response_senders
-                .generated_tokens
-                .with_registered("request", |_response_tx| ()),
-            None
+            sent_messages,
+            vec![
+                to_value(agent_desired_state_notification(
+                    newest_state.agent_desired_state
+                ))
+                .unwrap()
+            ]
         );
     }
-
-    #[test]
-    fn refuses_a_second_stream_under_a_request_id_in_flight() {
-        let (agent_message_tx, _agent_message_rx) = mpsc::unbounded_channel();
-        let agent_controller = AgentController {
-            agent_message_tx,
-            ..agent_controller_with(reported(1, 4))
-        };
-        let _in_flight_stream = agent_controller
-            .stream_responses_to("request".to_owned(), raw_prompt_params())
-            .expect("the first stream must reach the agent");
-
-        assert!(matches!(
-            agent_controller.stream_responses_to("request".to_owned(), raw_prompt_params()),
-            Err(AgentRelayError::RequestIdAlreadyRelayed { request_id }) if request_id == "request"
-        ));
-    }
 }
diff --git a/paddler_balancer/src/agent_controller_params.rs b/paddler_balancer/src/agent_controller_params.rs
new file mode 100644
index 000000000..b49d22580
--- /dev/null
+++ b/paddler_balancer/src/agent_controller_params.rs
@@ -0,0 +1,10 @@
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::slot_aggregated_status_snapshot::SlotAggregatedStatusSnapshot;
+
+pub struct AgentControllerParams {
+    pub connection_close: CancellationToken,
+    pub id: String,
+    pub name: Option<String>,
+    pub reported_status: SlotAggregatedStatusSnapshot,
+}
diff --git a/paddler_balancer/src/agent_controller_pool.rs b/paddler_balancer/src/agent_controller_pool.rs
index f5df7682d..32691a43d 100644
--- a/paddler_balancer/src/agent_controller_pool.rs
+++ b/paddler_balancer/src/agent_controller_pool.rs
@@ -1,47 +1,67 @@
+use std::collections::BTreeSet;
 use std::sync::Arc;
 
 use anyhow::Result;
 use dashmap::DashMap;
 use dashmap::mapref::entry::Entry;
-use log::debug;
+use tokio::sync::futures::Notified;
 use tokio::sync::watch;
 
 use paddler_messaging::agent_controller_pool_snapshot::AgentControllerPoolSnapshot;
-use paddler_messaging::agent_desired_state::AgentDesiredState;
+use paddler_messaging::inference_capability::InferenceCapability;
 use paddler_messaging::produces_snapshot::ProducesSnapshot;
 use paddler_messaging::subscribes_to_updates::SubscribesToUpdates;
 
 use super::agent_controller::AgentController;
 use super::agent_controller_pool_total_slots::AgentControllerPoolTotalSlots;
+use crate::agent_controller_pool_updates::AgentControllerPoolUpdates;
 use crate::agent_controller_registration::AgentControllerRegistration;
 use crate::agent_controller_slot_guard::AgentControllerSlotGuard;
-use crate::desired_state_delivery::DesiredStateDelivery;
+use crate::balancer_applicable_state::BalancerApplicableState;
+use crate::balancer_applicable_state_holder::BalancerApplicableStateHolder;
+use crate::buffered_request_counter::BufferedRequestCounter;
 use crate::dispatch_candidate::DispatchCandidate;
+use crate::dispatch_position::DispatchPosition;
+use crate::dispatch_requirement::DispatchRequirement;
+use crate::dispatch_waiting_line::DispatchWaitingLine;
 use crate::dispatched_agent::DispatchedAgent;
 use crate::registered_agent_controller_guard::RegisteredAgentControllerGuard;
+use crate::waiting_line_place::WaitingLinePlace;
 
 pub struct AgentControllerPool {
-    pub agents: DashMap<String, Arc<AgentController>>,
-    update_tx: watch::Sender<()>,
+    agents: DashMap<String, Arc<AgentController>>,
+    dispatch_waiting_line: DispatchWaitingLine,
+    updates: Arc<AgentControllerPoolUpdates>,
 }
 
 impl AgentControllerPool {
     #[must_use]
-    pub fn select_least_busy_with_capacity(&self) -> Option<DispatchCandidate> {
+    pub fn select_least_busy_with_capacity(
+        &self,
+        dispatch_requirement: DispatchRequirement,
+        dispatch_position: DispatchPosition,
+    ) -> Option<DispatchCandidate> {
+        let requirements_waiting_ahead = self
+            .dispatch_waiting_line
+            .requirements_ahead_of(dispatch_position);
         let mut best: Option<DispatchCandidate> = None;
 
         for entry in &self.agents {
             let agent_controller = entry.value().clone();
             let snapshot = agent_controller.slots_processing.get();
-
-            if snapshot >= agent_controller.reported_status.read().status.slots_total {
+            let agent_runtime_status = agent_controller.reported_status.read().status.runtime;
+            let Some(reserved_slots) = dispatch_requirement.reservable_slots(
+                &agent_runtime_status,
+                snapshot + requirements_waiting_ahead.slots_held_on(&agent_runtime_status),
+            ) else {
                 continue;
-            }
+            };
 
             best = Some(match best {
                 Some(current) if current.snapshot <= snapshot => current,
                 _ => DispatchCandidate {
                     agent_controller,
+                    reserved_slots,
                     snapshot,
                 },
             });
@@ -57,13 +77,17 @@ impl AgentControllerPool {
         if candidate
             .agent_controller
             .slots_processing
-            .compare_and_swap(candidate.snapshot, candidate.snapshot + 1)
+            .compare_and_swap(
+                candidate.snapshot,
+                candidate.snapshot + u64::from(candidate.reserved_slots.get()),
+            )
         {
-            self.update_tx.send_replace(());
+            self.updates.signal();
 
             let slot_guard = AgentControllerSlotGuard::new(
                 candidate.agent_controller.clone(),
-                self.update_tx.clone(),
+                self.updates.clone(),
+                candidate.reserved_slots,
             );
 
             Ok(DispatchedAgent::new(candidate.agent_controller, slot_guard))
@@ -72,10 +96,47 @@ impl AgentControllerPool {
         }
     }
 
+    pub fn agent_availability_changed(&self) -> Notified<'_> {
+        self.updates.agent_availability_changed()
+    }
+
+    #[must_use]
+    pub fn join_waiting_line_when_admitted(
+        &self,
+        dispatch_requirement: DispatchRequirement,
+        buffered_request_counter: &Arc<BufferedRequestCounter>,
+    ) -> Option<WaitingLinePlace<'_>> {
+        self.dispatch_waiting_line
+            .join_when_admitted(dispatch_requirement, buffered_request_counter)
+    }
+
+    pub async fn next_available_agent<'pool>(
+        &'pool self,
+        mut agent_availability_changed: Notified<'pool>,
+        dispatch_requirement: DispatchRequirement,
+        dispatch_position: DispatchPosition,
+    ) -> DispatchedAgent {
+        loop {
+            agent_availability_changed.await;
+            agent_availability_changed = self.agent_availability_changed();
+
+            if let Some(dispatched_agent) =
+                self.take_least_busy_agent_controller(dispatch_requirement, dispatch_position)
+            {
+                return dispatched_agent;
+            }
+        }
+    }
+
     #[must_use]
-    pub fn take_least_busy_agent_controller(&self) -> Option<DispatchedAgent> {
+    pub fn take_least_busy_agent_controller(
+        &self,
+        dispatch_requirement: DispatchRequirement,
+        dispatch_position: DispatchPosition,
+    ) -> Option<DispatchedAgent> {
         loop {
-            let candidate = self.select_least_busy_with_capacity()?;
+            let candidate =
+                self.select_least_busy_with_capacity(dispatch_requirement, dispatch_position)?;
 
             if let Ok(dispatched) = self.try_claim(candidate) {
                 return Some(dispatched);
@@ -91,12 +152,17 @@ impl AgentControllerPool {
     pub fn register_agent_controller(
         self: &Arc<Self>,
         agent_controller: Arc<AgentController>,
+        balancer_applicable_state_holder: &BalancerApplicableStateHolder,
     ) -> AgentControllerRegistration {
         match self.agents.entry(agent_controller.id.clone()) {
             Entry::Occupied(_registered_agent) => AgentControllerRegistration::DuplicateAgentId,
             Entry::Vacant(vacant_agent_slot) => {
-                vacant_agent_slot.insert(agent_controller.clone());
-                self.update_tx.send_replace(());
+                vacant_agent_slot
+                    .insert(agent_controller.clone())
+                    .set_desired_state(
+                        &balancer_applicable_state_holder.get_balancer_applicable_state(),
+                    );
+                self.updates.signal();
 
                 AgentControllerRegistration::Registered(RegisteredAgentControllerGuard {
                     agent_controller,
@@ -108,26 +174,42 @@ impl AgentControllerPool {
 
     pub fn remove_agent_controller(&self, agent_id: &str) {
         self.agents.remove(agent_id);
-        self.update_tx.send_replace(());
+        self.updates.signal();
     }
 
-    pub fn set_desired_state(&self, desired_state: &AgentDesiredState) {
+    pub fn set_desired_state(&self, balancer_applicable_state: &BalancerApplicableState) {
         for agent in &self.agents {
-            let agent_controller = agent.value();
+            agent.value().set_desired_state(balancer_applicable_state);
+        }
+    }
 
-            if matches!(
-                agent_controller.set_desired_state(desired_state.clone()),
-                DesiredStateDelivery::AgentDisconnected
-            ) {
-                let agent_id = &agent_controller.id;
+    #[must_use]
+    pub fn model_paths_serving(
+        &self,
+        inference_capability: InferenceCapability,
+    ) -> BTreeSet<String> {
+        self.agents
+            .iter()
+            .filter_map(|agent| {
+                let reported_status = agent.value().reported_status.read();
+
+                reported_status
+                    .status
+                    .runtime
+                    .serves(inference_capability)
+                    .then(|| reported_status.status.model_path.clone())
+                    .flatten()
+            })
+            .collect()
+    }
 
-                debug!("Skipping the desired state for disconnected agent {agent_id}");
-            }
-        }
+    #[must_use]
+    pub fn agent_count(&self) -> usize {
+        self.agents.len()
     }
 
     pub fn signal_update(&self) {
-        self.update_tx.send_replace(());
+        self.updates.signal();
     }
 
     #[must_use]
@@ -139,7 +221,7 @@ impl AgentControllerPool {
             let agent = entry.value();
 
             slots_processing += agent.slots_processing.get();
-            slots_total += agent.reported_status.read().status.slots_total;
+            slots_total += agent.reported_status.read().status.runtime.slots_total();
         }
 
         AgentControllerPoolTotalSlots {
@@ -151,18 +233,19 @@ impl AgentControllerPool {
 
 impl Default for AgentControllerPool {
     fn default() -> Self {
-        let (update_tx, _initial_rx) = watch::channel(());
+        let updates = Arc::new(AgentControllerPoolUpdates::default());
 
         Self {
             agents: DashMap::new(),
-            update_tx,
+            dispatch_waiting_line: DispatchWaitingLine::new(updates.clone()),
+            updates,
         }
     }
 }
 
 impl SubscribesToUpdates for AgentControllerPool {
     fn subscribe_to_updates(&self) -> watch::Receiver<()> {
-        self.update_tx.subscribe()
+        self.updates.subscribe()
     }
 }
 
@@ -183,70 +266,67 @@ impl ProducesSnapshot for AgentControllerPool {
 #[cfg(test)]
 mod tests {
     use std::sync::Arc;
-    use std::sync::atomic::AtomicU64;
 
-    use parking_lot::RwLock;
-    use tokio::sync::mpsc;
+    use futures::FutureExt as _;
+    use serde_json::Value;
+    use serde_json::to_value;
+    use tokio::sync::watch;
     use tokio_util::sync::CancellationToken;
 
+    use paddler_messaging::agent_desired_model::AgentDesiredModel;
     use paddler_messaging::agent_desired_state::AgentDesiredState;
-    use paddler_messaging::atomic_value::AtomicValue;
-    use paddler_messaging::management_socket::agent::message::Message as AgentJsonRpcMessage;
-    use paddler_messaging::management_socket::agent::notification::Notification as AgentJsonRpcNotification;
+    use paddler_messaging::balancer_desired_state::BalancerDesiredState;
     use paddler_messaging::slot_aggregated_status_snapshot::SlotAggregatedStatusSnapshot;
 
     use super::AgentControllerPool;
     use crate::agent_controller::AgentController;
-    use crate::agent_controller_registration::AgentControllerRegistration;
-    use crate::agent_response_senders::AgentResponseSenders;
-
-    struct AgentWithInbox {
-        controller: Arc<AgentController>,
-        inbox: mpsc::UnboundedReceiver<AgentJsonRpcMessage>,
-    }
-
-    fn agent_with_inbox(agent_id: &str) -> AgentWithInbox {
-        let (agent_message_tx, inbox) = mpsc::unbounded_channel();
-
-        AgentWithInbox {
-            controller: Arc::new(AgentController {
-                agent_message_tx,
-                agent_response_senders: AgentResponseSenders::default(),
-                connection_close: CancellationToken::new(),
-                id: agent_id.to_owned(),
-                name: None,
-                reported_status: RwLock::new(SlotAggregatedStatusSnapshot::default()),
-                slots_processing: AtomicValue::<AtomicU64>::new(0),
-            }),
-            inbox,
+    use crate::agent_controller_params::AgentControllerParams;
+    use crate::agent_desired_state_notification::agent_desired_state_notification;
+    use crate::balancer_applicable_state::BalancerApplicableState;
+    use crate::balancer_applicable_state_holder::BalancerApplicableStateHolder;
+
+    fn desired_state_with_model(model_uri: &str) -> BalancerDesiredState {
+        BalancerDesiredState {
+            model: AgentDesiredModel::Uri(model_uri.to_owned()),
+            ..BalancerDesiredState::default()
         }
     }
 
+    fn registering_agent_controller() -> Arc<AgentController> {
+        Arc::new(AgentController::new(AgentControllerParams {
+            connection_close: CancellationToken::new(),
+            id: "registering".to_owned(),
+            name: None,
+            reported_status: SlotAggregatedStatusSnapshot::default(),
+        }))
+    }
+
+    fn desired_state_notification_value(desired_state: BalancerDesiredState) -> Value {
+        to_value(agent_desired_state_notification(AgentDesiredState::from(
+            &desired_state,
+        )))
+        .unwrap()
+    }
+
     #[test]
-    fn delivers_the_desired_state_to_connected_agents_past_a_disconnected_one() {
-        let pool = Arc::new(AgentControllerPool::default());
-        let disconnected_agent = agent_with_inbox("disconnected");
-        let mut connected_agent = agent_with_inbox("connected");
-
-        drop(disconnected_agent.inbox);
-
-        let registrations = [
-            pool.register_agent_controller(disconnected_agent.controller),
-            pool.register_agent_controller(connected_agent.controller),
-        ];
-
-        assert!(registrations.iter().all(|registration| matches!(
-            registration,
-            AgentControllerRegistration::Registered(_)
-        )));
-
-        pool.set_desired_state(&AgentDesiredState::default());
-
-        assert!(matches!(
-            connected_agent.inbox.try_recv(),
-            Ok(AgentJsonRpcMessage::Notification(
-                AgentJsonRpcNotification::SetState(set_state_params)
-            )) if set_state_params.desired_state == AgentDesiredState::default()
+    fn registering_an_agent_sends_it_the_current_desired_state_before_anything_else() {
+        let current_desired_state = desired_state_with_model("agent:///current/model.gguf");
+        let (_applicable_state_tx, applicable_state_rx) = watch::channel(Arc::new(
+            BalancerApplicableState::initial(current_desired_state.clone()),
         ));
+        let agent_controller = registering_agent_controller();
+        let _registration = Arc::new(AgentControllerPool::default()).register_agent_controller(
+            agent_controller.clone(),
+            &BalancerApplicableStateHolder::new(applicable_state_rx),
+        );
+
+        assert_eq!(
+            agent_controller
+                .agent_outbox
+                .next()
+                .now_or_never()
+                .map(|first_message| to_value(first_message).unwrap()),
+            Some(desired_state_notification_value(current_desired_state))
+        );
     }
 }
diff --git a/paddler_balancer/src/agent_controller_pool_updates.rs b/paddler_balancer/src/agent_controller_pool_updates.rs
new file mode 100644
index 000000000..05470c732
--- /dev/null
+++ b/paddler_balancer/src/agent_controller_pool_updates.rs
@@ -0,0 +1,39 @@
+use tokio::sync::Notify;
+use tokio::sync::futures::Notified;
+use tokio::sync::watch;
+
+pub struct AgentControllerPoolUpdates {
+    agent_availability_change: Notify,
+    update_tx: watch::Sender<()>,
+}
+
+impl AgentControllerPoolUpdates {
+    pub fn agent_availability_changed(&self) -> Notified<'_> {
+        self.agent_availability_change.notified()
+    }
+
+    pub fn signal_waiting_line_change(&self) {
+        self.agent_availability_change.notify_waiters();
+    }
+
+    pub fn signal(&self) {
+        self.update_tx.send_replace(());
+        self.agent_availability_change.notify_waiters();
+    }
+
+    #[must_use]
+    pub fn subscribe(&self) -> watch::Receiver<()> {
+        self.update_tx.subscribe()
+    }
+}
+
+impl Default for AgentControllerPoolUpdates {
+    fn default() -> Self {
+        let (update_tx, _initial_rx) = watch::channel(());
+
+        Self {
+            agent_availability_change: Notify::new(),
+            update_tx,
+        }
+    }
+}
diff --git a/paddler_balancer/src/agent_controller_slot_guard.rs b/paddler_balancer/src/agent_controller_slot_guard.rs
index 4612db7ad..2943f1b8f 100644
--- a/paddler_balancer/src/agent_controller_slot_guard.rs
+++ b/paddler_balancer/src/agent_controller_slot_guard.rs
@@ -1,29 +1,39 @@
+use std::num::NonZeroU16;
 use std::sync::Arc;
 
-use tokio::sync::watch;
-
 use crate::agent_controller::AgentController;
+use crate::agent_controller_pool_updates::AgentControllerPoolUpdates;
 
 pub struct AgentControllerSlotGuard {
     agent_controller: Arc<AgentController>,
-    pool_update_tx: watch::Sender<()>,
+    pool_updates: Arc<AgentControllerPoolUpdates>,
+    reserved_slots: NonZeroU16,
 }
 
 impl AgentControllerSlotGuard {
     pub const fn new(
         agent_controller: Arc<AgentController>,
-        pool_update_tx: watch::Sender<()>,
+        pool_updates: Arc<AgentControllerPoolUpdates>,
+        reserved_slots: NonZeroU16,
     ) -> Self {
         Self {
             agent_controller,
-            pool_update_tx,
+            pool_updates,
+            reserved_slots,
         }
     }
+
+    #[must_use]
+    pub const fn reserved_slots(&self) -> NonZeroU16 {
+        self.reserved_slots
+    }
 }
 
 impl Drop for AgentControllerSlotGuard {
     fn drop(&mut self) {
-        self.agent_controller.slots_processing.decrement();
-        self.pool_update_tx.send_replace(());
+        self.agent_controller
+            .slots_processing
+            .subtract(u64::from(self.reserved_slots.get()));
+        self.pool_updates.signal();
     }
 }
diff --git a/paddler_balancer/src/agent_relay.rs b/paddler_balancer/src/agent_relay.rs
new file mode 100644
index 000000000..d5d2c0bfb
--- /dev/null
+++ b/paddler_balancer/src/agent_relay.rs
@@ -0,0 +1,100 @@
+use std::time::Duration;
+
+use tokio_util::sync::CancellationToken;
+
+use crate::agent_relay_conclusion::AgentRelayConclusion;
+use crate::agent_relay_event::AgentRelayEvent;
+use crate::agent_relay_phase::AgentRelayPhase;
+use crate::agent_relay_step::AgentRelayStep;
+use crate::agent_streaming_request::AgentStreamingRequest;
+use crate::awaiting_agent_relay::AwaitingAgentRelay;
+use crate::dispatched_agent::DispatchedAgent;
+use crate::dispatched_agent_relay::DispatchedAgentRelay;
+
+pub struct AgentRelay<TParams>
+where
+    TParams: AgentStreamingRequest,
+{
+    phase: AgentRelayPhase<TParams>,
+    settlement: CancellationToken,
+}
+
+impl<TParams> AgentRelay<TParams>
+where
+    TParams: AgentStreamingRequest,
+{
+    #[must_use]
+    pub fn awaiting(awaiting_agent_relay: AwaitingAgentRelay<TParams>) -> Self {
+        let settlement = CancellationToken::new();
+
+        Self {
+            phase: AgentRelayPhase::Awaiting {
+                awaiting_agent_relay,
+                settlement_guard: settlement.clone().drop_guard(),
+            },
+            settlement,
+        }
+    }
+
+    #[must_use]
+    pub fn dispatched(
+        dispatched_agent: DispatchedAgent,
+        params: TParams,
+        inference_item_timeout: Duration,
+        shutdown: CancellationToken,
+    ) -> Self {
+        let settlement = CancellationToken::new();
+
+        Self {
+            phase: AgentRelayPhase::Dispatched(DispatchedAgentRelay::start(
+                dispatched_agent,
+                params,
+                inference_item_timeout,
+                settlement.clone(),
+                shutdown,
+            )),
+            settlement,
+        }
+    }
+
+    #[must_use]
+    pub fn settlement(&self) -> CancellationToken {
+        self.settlement.clone()
+    }
+
+    pub async fn next(self) -> AgentRelayStep<TParams> {
+        let Self { phase, settlement } = self;
+        let dispatched_agent_relay = match phase {
+            AgentRelayPhase::Awaiting {
+                awaiting_agent_relay,
+                settlement_guard,
+            } => match awaiting_agent_relay.dispatch(settlement_guard).await {
+                Ok(dispatched_agent_relay) => dispatched_agent_relay,
+                Err(relay_failure) => {
+                    return AgentRelayStep::Concluded(AgentRelayConclusion::Failed(relay_failure));
+                }
+            },
+            AgentRelayPhase::Dispatched(dispatched_agent_relay) => dispatched_agent_relay,
+        };
+
+        match dispatched_agent_relay.agent_events.next().await {
+            AgentRelayEvent::Failed(relay_failure) => {
+                AgentRelayStep::Concluded(AgentRelayConclusion::Failed(relay_failure))
+            }
+            AgentRelayEvent::Final(final_result) => {
+                AgentRelayStep::Concluded(AgentRelayConclusion::AgentConcluded {
+                    final_result,
+                    generated_by: dispatched_agent_relay.generated_by,
+                })
+            }
+            AgentRelayEvent::Partial(partial_result) => AgentRelayStep::Partial {
+                generated_by: dispatched_agent_relay.generated_by.clone(),
+                partial_result,
+                relay: Self {
+                    phase: AgentRelayPhase::Dispatched(dispatched_agent_relay),
+                    settlement,
+                },
+            },
+        }
+    }
+}
diff --git a/paddler_balancer/src/agent_relay_conclusion.rs b/paddler_balancer/src/agent_relay_conclusion.rs
new file mode 100644
index 000000000..6c0b73870
--- /dev/null
+++ b/paddler_balancer/src/agent_relay_conclusion.rs
@@ -0,0 +1,9 @@
+use crate::relay_failure::RelayFailure;
+
+pub enum AgentRelayConclusion<TFinal> {
+    AgentConcluded {
+        final_result: TFinal,
+        generated_by: Option<String>,
+    },
+    Failed(RelayFailure),
+}
diff --git a/paddler_balancer/src/agent_relay_error.rs b/paddler_balancer/src/agent_relay_error.rs
deleted file mode 100644
index 58744bec0..000000000
--- a/paddler_balancer/src/agent_relay_error.rs
+++ /dev/null
@@ -1,16 +0,0 @@
-use serde_json::Error as SerdeJsonError;
-use thiserror::Error;
-
-use paddler_messaging::inference_client::message::Message as OutgoingMessage;
-
-#[derive(Debug, Error)]
-pub enum AgentRelayError {
-    #[error("The socket of agent {agent_id} is closed")]
-    AgentSocketClosed { agent_id: String },
-    #[error("Unable to serialize a message relayed to the client")]
-    MessageUnserializable(#[source] SerdeJsonError),
-    #[error("The client's protocol cannot carry this message: {message:?}")]
-    MessageNotRelayable { message: Box<OutgoingMessage> },
-    #[error("A response is already being relayed under request id {request_id}")]
-    RequestIdAlreadyRelayed { request_id: String },
-}
diff --git a/paddler_balancer/src/agent_relay_event.rs b/paddler_balancer/src/agent_relay_event.rs
new file mode 100644
index 000000000..2964b9178
--- /dev/null
+++ b/paddler_balancer/src/agent_relay_event.rs
@@ -0,0 +1,8 @@
+use crate::relay_failure::RelayFailure;
+
+#[derive(Debug, Eq, PartialEq)]
+pub enum AgentRelayEvent<TPartial, TFinal> {
+    Failed(RelayFailure),
+    Final(TFinal),
+    Partial(TPartial),
+}
diff --git a/paddler_balancer/src/agent_relay_phase.rs b/paddler_balancer/src/agent_relay_phase.rs
new file mode 100644
index 000000000..86979d7dd
--- /dev/null
+++ b/paddler_balancer/src/agent_relay_phase.rs
@@ -0,0 +1,16 @@
+use tokio_util::sync::DropGuard;
+
+use crate::agent_streaming_request::AgentStreamingRequest;
+use crate::awaiting_agent_relay::AwaitingAgentRelay;
+use crate::dispatched_agent_relay::DispatchedAgentRelay;
+
+pub enum AgentRelayPhase<TParams>
+where
+    TParams: AgentStreamingRequest,
+{
+    Awaiting {
+        awaiting_agent_relay: AwaitingAgentRelay<TParams>,
+        settlement_guard: DropGuard,
+    },
+    Dispatched(DispatchedAgentRelay<TParams::Response>),
+}
diff --git a/paddler_balancer/src/agent_relay_step.rs b/paddler_balancer/src/agent_relay_step.rs
new file mode 100644
index 000000000..3ab2f19da
--- /dev/null
+++ b/paddler_balancer/src/agent_relay_step.rs
@@ -0,0 +1,17 @@
+use paddler_messaging::streamable_result::StreamableResult;
+
+use crate::agent_relay::AgentRelay;
+use crate::agent_relay_conclusion::AgentRelayConclusion;
+use crate::agent_streaming_request::AgentStreamingRequest;
+
+pub enum AgentRelayStep<TParams>
+where
+    TParams: AgentStreamingRequest,
+{
+    Concluded(AgentRelayConclusion<<TParams::Response as StreamableResult>::Final>),
+    Partial {
+        generated_by: Option<String>,
+        partial_result: <TParams::Response as StreamableResult>::Partial,
+        relay: AgentRelay<TParams>,
+    },
+}
diff --git a/paddler_balancer/src/agent_response_forwarder.rs b/paddler_balancer/src/agent_response_forwarder.rs
new file mode 100644
index 000000000..87c09a2d9
--- /dev/null
+++ b/paddler_balancer/src/agent_response_forwarder.rs
@@ -0,0 +1,119 @@
+use std::pin::pin;
+use std::sync::Arc;
+use std::time::Duration;
+
+use tokio::select;
+use tokio::time::Instant;
+use tokio::time::sleep;
+use tokio_util::sync::CancellationToken;
+use tokio_util::sync::DropGuard;
+
+use paddler_messaging::streamable_result::StreamableResult;
+use paddler_outbox::outbox::Outbox;
+
+use crate::agent_relay_event::AgentRelayEvent;
+use crate::agent_response_forwarding_mode::AgentResponseForwardingMode;
+use crate::agent_response_receiver::AgentResponseReceiver;
+use crate::agent_stop_follow_up::AgentStopFollowUp;
+use crate::dispatched_agent::DispatchedAgent;
+use crate::forwarding_event::ForwardingEvent;
+use crate::forwarding_plan::ForwardingPlan;
+
+pub struct AgentResponseForwarder<TResponse>
+where
+    TResponse: StreamableResult,
+{
+    pub agent_events: Arc<Outbox<AgentRelayEvent<TResponse::Partial, TResponse::Final>>>,
+    pub dispatched_agent: DispatchedAgent,
+    pub inference_item_timeout: Duration,
+    pub relay_abandonment: CancellationToken,
+    pub response_receiver: AgentResponseReceiver<TResponse>,
+    pub settlement_guard: DropGuard,
+    pub shutdown: CancellationToken,
+}
+
+impl<TResponse> AgentResponseForwarder<TResponse>
+where
+    TResponse: StreamableResult,
+{
+    pub async fn forward(self) {
+        let Self {
+            agent_events,
+            dispatched_agent,
+            inference_item_timeout,
+            relay_abandonment,
+            response_receiver:
+                AgentResponseReceiver {
+                    request_registration,
+                    mut response_rx,
+                },
+            settlement_guard: _settlement_guard,
+            shutdown,
+        } = self;
+        let agent_controller = &dispatched_agent.agent_controller;
+        let agent_connection_close = agent_controller.connection_close.clone();
+        let mut forwarding_mode = AgentResponseForwardingMode::ForwardingToRelay;
+        let mut item_timeout = pin!(sleep(inference_item_timeout));
+
+        loop {
+            item_timeout
+                .as_mut()
+                .reset(Instant::now() + inference_item_timeout);
+
+            let is_forwarding_to_relay = matches!(
+                forwarding_mode,
+                AgentResponseForwardingMode::ForwardingToRelay
+            );
+
+            let forwarding_event = select! {
+                biased;
+
+                () = shutdown.cancelled() => ForwardingEvent::ShutdownRequested,
+                () = agent_connection_close.cancelled() => ForwardingEvent::AgentConnectionClosed,
+                () = relay_abandonment.cancelled(), if is_forwarding_to_relay => {
+                    ForwardingEvent::RelayAbandoned
+                }
+                () = &mut item_timeout => ForwardingEvent::ItemTimedOut,
+                Some(response) = response_rx.recv() => {
+                    ForwardingEvent::ResponseReceived(response.into_step())
+                }
+            };
+
+            let agent_stop_follow_up =
+                match forwarding_mode.plan_for(forwarding_event, inference_item_timeout) {
+                    ForwardingPlan::Finish => break,
+                    ForwardingPlan::IgnoreDrainedResponse => continue,
+                    ForwardingPlan::NotifyRelay(agent_relay_event) => {
+                        agent_events.push(agent_relay_event);
+
+                        continue;
+                    }
+                    ForwardingPlan::NotifyRelayThenFinish(agent_relay_event) => {
+                        agent_events.push(agent_relay_event);
+
+                        break;
+                    }
+                    ForwardingPlan::NotifyRelayThenStopAgent(agent_relay_event) => {
+                        agent_events.push(agent_relay_event);
+
+                        AgentStopFollowUp::DrainUntilAgentConfirms
+                    }
+                    ForwardingPlan::NotifyRelayThenStopAgentThenFinish(agent_relay_event) => {
+                        agent_events.push(agent_relay_event);
+
+                        AgentStopFollowUp::Finish
+                    }
+                    ForwardingPlan::StopAgent => AgentStopFollowUp::DrainUntilAgentConfirms,
+                };
+
+            agent_controller.stop_responding_to(request_registration.request_id.clone());
+
+            match agent_stop_follow_up {
+                AgentStopFollowUp::DrainUntilAgentConfirms => {
+                    forwarding_mode = AgentResponseForwardingMode::DrainingUntilAgentConfirms;
+                }
+                AgentStopFollowUp::Finish => break,
+            }
+        }
+    }
+}
diff --git a/paddler_balancer/src/agent_response_forwarding_mode.rs b/paddler_balancer/src/agent_response_forwarding_mode.rs
index c1f3b31d6..6e6ff9187 100644
--- a/paddler_balancer/src/agent_response_forwarding_mode.rs
+++ b/paddler_balancer/src/agent_response_forwarding_mode.rs
@@ -1,60 +1,55 @@
 use std::time::Duration;
 
-use paddler_messaging::jsonrpc::error::Error as JsonRpcError;
-use paddler_messaging::streamable_result::StreamableResult;
+use paddler_messaging::result_step::ResultStep;
 
-use crate::balancer_shutdown_error::balancer_shutdown_error;
+use crate::agent_relay_event::AgentRelayEvent;
 use crate::forwarding_event::ForwardingEvent;
 use crate::forwarding_plan::ForwardingPlan;
+use crate::relay_failure::RelayFailure;
 
 pub enum AgentResponseForwardingMode {
     DrainingUntilAgentConfirms,
-    ForwardingToClient,
+    ForwardingToRelay,
 }
 
 impl AgentResponseForwardingMode {
     #[must_use]
-    pub fn plan_for<TResponse: StreamableResult>(
+    pub fn plan_for<TPartial, TFinal>(
         &self,
-        event: ForwardingEvent<TResponse>,
+        event: ForwardingEvent<TPartial, TFinal>,
         inference_item_timeout: Duration,
-    ) -> ForwardingPlan<TResponse> {
+    ) -> ForwardingPlan<TPartial, TFinal> {
         match self {
-            Self::ForwardingToClient => match event {
-                ForwardingEvent::AgentConnectionClosed => {
-                    ForwardingPlan::ReplyWithErrorThenFinish(JsonRpcError {
-                        code: 502,
-                        description: "Agent controller connection closed".to_owned(),
-                    })
+            Self::ForwardingToRelay => match event {
+                ForwardingEvent::AgentConnectionClosed => ForwardingPlan::NotifyRelayThenFinish(
+                    AgentRelayEvent::Failed(RelayFailure::AgentDisconnected),
+                ),
+                ForwardingEvent::ItemTimedOut => ForwardingPlan::NotifyRelayThenStopAgent(
+                    AgentRelayEvent::Failed(RelayFailure::ItemTimedOut {
+                        inference_item_timeout,
+                    }),
+                ),
+                ForwardingEvent::RelayAbandoned => ForwardingPlan::StopAgent,
+                ForwardingEvent::ResponseReceived(ResultStep::Final(final_result)) => {
+                    ForwardingPlan::NotifyRelayThenFinish(AgentRelayEvent::Final(final_result))
                 }
-                ForwardingEvent::ClientConnectionClosed => ForwardingPlan::StopAgent,
-                ForwardingEvent::ItemTimedOut => {
-                    let timeout_ms = inference_item_timeout.as_millis();
-
-                    ForwardingPlan::ReplyWithErrorThenStopAgent(JsonRpcError {
-                        code: 504,
-                        description: format!(
-                            "Inference timed out after {timeout_ms}ms waiting for next token. \
-                            Increase --inference-item-timeout if the prompt requires longer processing."
-                        ),
-                    })
+                ForwardingEvent::ResponseReceived(ResultStep::Partial(partial_result)) => {
+                    ForwardingPlan::NotifyRelay(AgentRelayEvent::Partial(partial_result))
                 }
-                ForwardingEvent::ResponseReceived(response) => ForwardingPlan::ForwardResponse {
-                    is_done: response.is_done(),
-                    response,
-                },
                 ForwardingEvent::ShutdownRequested => {
-                    ForwardingPlan::ReplyWithErrorThenStopAgentThenFinish(balancer_shutdown_error())
+                    ForwardingPlan::NotifyRelayThenStopAgentThenFinish(AgentRelayEvent::Failed(
+                        RelayFailure::BalancerShuttingDown,
+                    ))
                 }
             },
             Self::DrainingUntilAgentConfirms => match event {
-                ForwardingEvent::ResponseReceived(response) if !response.is_done() => {
+                ForwardingEvent::ResponseReceived(ResultStep::Partial(_)) => {
                     ForwardingPlan::IgnoreDrainedResponse
                 }
                 ForwardingEvent::AgentConnectionClosed
-                | ForwardingEvent::ClientConnectionClosed
                 | ForwardingEvent::ItemTimedOut
-                | ForwardingEvent::ResponseReceived(_)
+                | ForwardingEvent::RelayAbandoned
+                | ForwardingEvent::ResponseReceived(ResultStep::Final(_))
                 | ForwardingEvent::ShutdownRequested => ForwardingPlan::Finish,
             },
         }
@@ -65,90 +60,114 @@ impl AgentResponseForwardingMode {
 mod tests {
     use std::time::Duration;
 
-    use paddler_messaging::generated_token_result::GeneratedTokenResult;
-    use paddler_messaging::jsonrpc::error::Error as JsonRpcError;
+    use llama_cpp_bindings_types::TokenUsage;
+
+    use paddler_messaging::generated_token_final_result::GeneratedTokenFinalResult;
+    use paddler_messaging::generated_token_partial_result::GeneratedTokenPartialResult;
+    use paddler_messaging::generation_finish::GenerationFinish;
+    use paddler_messaging::generation_summary::GenerationSummary;
+    use paddler_messaging::result_step::ResultStep;
 
     use super::AgentResponseForwardingMode;
+    use crate::agent_relay_event::AgentRelayEvent;
     use crate::forwarding_event::ForwardingEvent;
     use crate::forwarding_plan::ForwardingPlan;
+    use crate::relay_failure::RelayFailure;
 
     const INFERENCE_ITEM_TIMEOUT: Duration = Duration::from_millis(250);
 
-    fn forwarding(
-        event: ForwardingEvent<GeneratedTokenResult>,
-    ) -> ForwardingPlan<GeneratedTokenResult> {
-        AgentResponseForwardingMode::ForwardingToClient.plan_for(event, INFERENCE_ITEM_TIMEOUT)
+    type GeneratedTokenForwardingEvent =
+        ForwardingEvent<GeneratedTokenPartialResult, GeneratedTokenFinalResult>;
+
+    type GeneratedTokenForwardingPlan =
+        ForwardingPlan<GeneratedTokenPartialResult, GeneratedTokenFinalResult>;
+
+    fn forwarding(event: GeneratedTokenForwardingEvent) -> GeneratedTokenForwardingPlan {
+        AgentResponseForwardingMode::ForwardingToRelay.plan_for(event, INFERENCE_ITEM_TIMEOUT)
     }
 
-    fn draining(
-        event: ForwardingEvent<GeneratedTokenResult>,
-    ) -> ForwardingPlan<GeneratedTokenResult> {
+    fn draining(event: GeneratedTokenForwardingEvent) -> GeneratedTokenForwardingPlan {
         AgentResponseForwardingMode::DrainingUntilAgentConfirms
             .plan_for(event, INFERENCE_ITEM_TIMEOUT)
     }
 
+    fn content_token() -> GeneratedTokenPartialResult {
+        GeneratedTokenPartialResult::ContentToken("token".to_owned())
+    }
+
+    fn generation_done() -> GeneratedTokenFinalResult {
+        GeneratedTokenFinalResult::Done(GenerationSummary {
+            finish: GenerationFinish::EndOfGeneration,
+            usage: TokenUsage::new(),
+        })
+    }
+
     #[test]
-    fn tells_the_client_about_a_shutdown_and_stops_the_agent() {
+    fn tells_the_relay_about_a_shutdown_and_stops_the_agent() {
         assert_eq!(
             forwarding(ForwardingEvent::ShutdownRequested),
-            ForwardingPlan::ReplyWithErrorThenStopAgentThenFinish(JsonRpcError {
-                code: 503,
-                description: "balancer is shutting down".to_owned(),
-            })
+            ForwardingPlan::NotifyRelayThenStopAgentThenFinish(AgentRelayEvent::Failed(
+                RelayFailure::BalancerShuttingDown
+            ))
         );
     }
 
     #[test]
-    fn tells_the_client_when_the_agent_connection_closes() {
+    fn tells_the_relay_when_the_agent_connection_closes() {
         assert_eq!(
             forwarding(ForwardingEvent::AgentConnectionClosed),
-            ForwardingPlan::ReplyWithErrorThenFinish(JsonRpcError {
-                code: 502,
-                description: "Agent controller connection closed".to_owned(),
-            })
+            ForwardingPlan::NotifyRelayThenFinish(AgentRelayEvent::Failed(
+                RelayFailure::AgentDisconnected
+            ))
         );
     }
 
     #[test]
-    fn stops_the_agent_when_the_client_goes_away() {
+    fn stops_the_agent_when_the_relay_is_abandoned() {
         assert_eq!(
-            forwarding(ForwardingEvent::ClientConnectionClosed),
+            forwarding(ForwardingEvent::RelayAbandoned),
             ForwardingPlan::StopAgent
         );
     }
 
     #[test]
-    fn tells_the_client_about_an_item_timeout_and_stops_the_agent() {
+    fn tells_the_relay_about_an_item_timeout_and_stops_the_agent() {
         assert_eq!(
             forwarding(ForwardingEvent::ItemTimedOut),
-            ForwardingPlan::ReplyWithErrorThenStopAgent(JsonRpcError {
-                code: 504,
-                description: "Inference timed out after 250ms waiting for next token. \
-                    Increase --inference-item-timeout if the prompt requires longer processing."
-                    .to_owned(),
-            })
+            ForwardingPlan::NotifyRelayThenStopAgent(AgentRelayEvent::Failed(
+                RelayFailure::ItemTimedOut {
+                    inference_item_timeout: INFERENCE_ITEM_TIMEOUT,
+                }
+            ))
+        );
+    }
+
+    #[test]
+    fn forwards_partial_responses_to_the_relay() {
+        assert_eq!(
+            forwarding(ForwardingEvent::ResponseReceived(ResultStep::Partial(
+                content_token()
+            ))),
+            ForwardingPlan::NotifyRelay(AgentRelayEvent::Partial(content_token()))
         );
     }
 
     #[test]
-    fn forwards_responses_to_the_client() {
+    fn forwards_the_final_response_to_the_relay_and_finishes() {
         assert_eq!(
-            forwarding(ForwardingEvent::ResponseReceived(
-                GeneratedTokenResult::ContentToken("token".to_owned())
-            )),
-            ForwardingPlan::ForwardResponse {
-                is_done: false,
-                response: GeneratedTokenResult::ContentToken("token".to_owned()),
-            }
+            forwarding(ForwardingEvent::ResponseReceived(ResultStep::Final(
+                generation_done()
+            ))),
+            ForwardingPlan::NotifyRelayThenFinish(AgentRelayEvent::Final(generation_done()))
         );
     }
 
     #[test]
-    fn ignores_responses_while_the_agent_confirms_the_stop() {
+    fn ignores_partial_responses_while_the_agent_confirms_the_stop() {
         assert_eq!(
-            draining(ForwardingEvent::ResponseReceived(
-                GeneratedTokenResult::ContentToken("token".to_owned())
-            )),
+            draining(ForwardingEvent::ResponseReceived(ResultStep::Partial(
+                content_token()
+            ))),
             ForwardingPlan::IgnoreDrainedResponse
         );
     }
@@ -156,9 +175,9 @@ mod tests {
     #[test]
     fn finishes_draining_once_the_agent_confirms_the_stop() {
         assert_eq!(
-            draining(ForwardingEvent::ResponseReceived(
-                GeneratedTokenResult::ChatTemplateError("template failed".to_owned())
-            )),
+            draining(ForwardingEvent::ResponseReceived(ResultStep::Final(
+                generation_done()
+            ))),
             ForwardingPlan::Finish
         );
     }
diff --git a/paddler_balancer/src/agent_response_receiver.rs b/paddler_balancer/src/agent_response_receiver.rs
index 2c8304506..2c846de6d 100644
--- a/paddler_balancer/src/agent_response_receiver.rs
+++ b/paddler_balancer/src/agent_response_receiver.rs
@@ -4,7 +4,6 @@ use tokio::sync::mpsc;
 
 use paddler_request_registry::request_registry_guard::RequestRegistryGuard;
 
-use crate::agent_relay_error::AgentRelayError;
 use crate::response_senders::ResponseSenders;
 
 pub struct AgentResponseReceiver<TResponse> {
@@ -13,59 +12,16 @@ pub struct AgentResponseReceiver<TResponse> {
 }
 
 impl<TResponse> AgentResponseReceiver<TResponse> {
-    pub fn register(
-        response_senders: &Arc<ResponseSenders<TResponse>>,
-        request_id: String,
-    ) -> Result<Self, AgentRelayError> {
+    #[must_use]
+    pub fn register(response_senders: &Arc<ResponseSenders<TResponse>>) -> Self {
         let (response_tx, response_rx) = mpsc::unbounded_channel();
 
-        RequestRegistryGuard::register(response_senders, request_id.clone(), response_tx)
-            .map(|request_registration| Self {
-                request_registration,
-                response_rx,
-            })
-            .ok_or(AgentRelayError::RequestIdAlreadyRelayed { request_id })
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use std::sync::Arc;
-
-    use paddler_messaging::embedding_result::EmbeddingResult;
-    use paddler_request_registry::request_delivery::RequestDelivery;
-
-    use super::AgentResponseReceiver;
-    use crate::agent_relay_error::AgentRelayError;
-    use crate::response_senders::ResponseSenders;
-
-    #[test]
-    fn receives_the_responses_sent_to_its_request() {
-        let response_senders = Arc::new(ResponseSenders::default());
-        let mut agent_response_receiver =
-            AgentResponseReceiver::register(&response_senders, "request-1".to_owned())
-                .expect("a fresh request id must register");
-
-        assert_eq!(
-            response_senders.send_to("request-1", EmbeddingResult::Done),
-            RequestDelivery::Delivered
-        );
-        assert_eq!(
-            agent_response_receiver.response_rx.try_recv(),
-            Ok(EmbeddingResult::Done)
-        );
-    }
-
-    #[test]
-    fn refuses_a_second_receiver_for_the_same_request() {
-        let response_senders = Arc::new(ResponseSenders::<EmbeddingResult>::default());
-        let _first_receiver =
-            AgentResponseReceiver::register(&response_senders, "request-1".to_owned())
-                .expect("a fresh request id must register");
-
-        assert!(matches!(
-            AgentResponseReceiver::register(&response_senders, "request-1".to_owned()),
-            Err(AgentRelayError::RequestIdAlreadyRelayed { request_id }) if request_id == "request-1"
-        ));
+        Self {
+            request_registration: RequestRegistryGuard::register_fresh(
+                response_senders,
+                response_tx,
+            ),
+            response_rx,
+        }
     }
 }
diff --git a/paddler_balancer/src/agent_response_senders.rs b/paddler_balancer/src/agent_response_senders.rs
index 8e2cced33..0a2d82afe 100644
--- a/paddler_balancer/src/agent_response_senders.rs
+++ b/paddler_balancer/src/agent_response_senders.rs
@@ -1,16 +1,63 @@
 use std::sync::Arc;
 
+use log::warn;
+
 use paddler_messaging::chat_template::ChatTemplate;
+use paddler_messaging::decision_result::DecisionResult;
 use paddler_messaging::embedding_result::EmbeddingResult;
 use paddler_messaging::generated_token_result::GeneratedTokenResult;
+use paddler_messaging::management_socket::agent::response::Response as AgentJsonRpcResponse;
 use paddler_messaging::model_metadata::ModelMetadata;
+use paddler_request_registry::request_delivery::RequestDelivery;
 
 use crate::response_senders::ResponseSenders;
 
+fn forward_agent_response<TResponse>(
+    response_senders: &ResponseSenders<TResponse>,
+    request_id: &str,
+    response: TResponse,
+) {
+    match response_senders.send_to(request_id, response) {
+        RequestDelivery::Delivered => {}
+        RequestDelivery::ReceiverDropped | RequestDelivery::RequestNotRegistered => {
+            warn!(
+                "Dropped an agent response for request {request_id:?}, which is no longer relayed"
+            );
+        }
+    }
+}
+
 #[derive(Clone, Default)]
 pub struct AgentResponseSenders {
     pub chat_template_override: Arc<ResponseSenders<Option<ChatTemplate>>>,
+    pub decision: Arc<ResponseSenders<DecisionResult>>,
     pub embedding: Arc<ResponseSenders<EmbeddingResult>>,
     pub generated_tokens: Arc<ResponseSenders<GeneratedTokenResult>>,
     pub model_metadata: Arc<ResponseSenders<Option<ModelMetadata>>>,
 }
+
+impl AgentResponseSenders {
+    pub fn forward(&self, request_id: &str, response: AgentJsonRpcResponse) {
+        match response {
+            AgentJsonRpcResponse::ChatTemplateOverride(chat_template_override) => {
+                forward_agent_response(
+                    &self.chat_template_override,
+                    request_id,
+                    chat_template_override,
+                );
+            }
+            AgentJsonRpcResponse::Decision(decision_result) => {
+                forward_agent_response(&self.decision, request_id, decision_result);
+            }
+            AgentJsonRpcResponse::Embedding(embedding_result) => {
+                forward_agent_response(&self.embedding, request_id, embedding_result);
+            }
+            AgentJsonRpcResponse::GeneratedToken(generated_token_result) => {
+                forward_agent_response(&self.generated_tokens, request_id, generated_token_result);
+            }
+            AgentJsonRpcResponse::ModelMetadata(model_metadata) => {
+                forward_agent_response(&self.model_metadata, request_id, model_metadata);
+            }
+        }
+    }
+}
diff --git a/paddler_balancer/src/agent_streaming_request.rs b/paddler_balancer/src/agent_streaming_request.rs
index b77211d39..be8db02f8 100644
--- a/paddler_balancer/src/agent_streaming_request.rs
+++ b/paddler_balancer/src/agent_streaming_request.rs
@@ -1,25 +1,49 @@
+use std::num::NonZeroU16;
 use std::sync::Arc;
 
+use paddler_messaging::decision_result::DecisionResult;
 use paddler_messaging::embedding_result::EmbeddingResult;
 use paddler_messaging::generated_token_result::GeneratedTokenResult;
-use paddler_messaging::management_socket::agent::request::Request as AgentJsonRpcRequest;
+use paddler_messaging::inference_client::response::Response as OutgoingResponse;
+use paddler_messaging::inference_capability::InferenceCapability;
 use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
 use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters_schema::validated_parameters_schema::ValidatedParametersSchema;
 use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
+use paddler_messaging::request_params::decide_params::DecideParams;
 use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
+use paddler_messaging::streamable_result::StreamableResult;
 
 use crate::agent_response_senders::AgentResponseSenders;
+use crate::dispatch_requirement::DispatchRequirement;
+use crate::into_reserved_agent_request::IntoReservedAgentRequest;
 use crate::response_senders::ResponseSenders;
+use crate::slot_demand::SlotDemand;
 
-pub trait AgentStreamingRequest: Into<AgentJsonRpcRequest> {
-    type Response: Send + 'static;
+pub trait AgentStreamingRequest: IntoReservedAgentRequest {
+    const INFERENCE_CAPABILITY: InferenceCapability;
+
+    type Response: StreamableResult<Final: Send + 'static, Partial: Send + 'static>
+        + Into<OutgoingResponse>
+        + Send
+        + 'static;
 
     fn response_senders(
         agent_response_senders: &AgentResponseSenders,
     ) -> &Arc<ResponseSenders<Self::Response>>;
+
+    fn slot_demand(&self) -> SlotDemand;
+
+    fn dispatch_requirement(&self) -> DispatchRequirement {
+        DispatchRequirement {
+            inference_capability: Self::INFERENCE_CAPABILITY,
+            slot_demand: self.slot_demand(),
+        }
+    }
 }
 
 impl AgentStreamingRequest for ContinueFromConversationHistoryParams<ValidatedParametersSchema> {
+    const INFERENCE_CAPABILITY: InferenceCapability = InferenceCapability::TextGeneration;
+
     type Response = GeneratedTokenResult;
 
     fn response_senders(
@@ -27,9 +51,15 @@ impl AgentStreamingRequest for ContinueFromConversationHistoryParams<ValidatedPa
     ) -> &Arc<ResponseSenders<Self::Response>> {
         &agent_response_senders.generated_tokens
     }
+
+    fn slot_demand(&self) -> SlotDemand {
+        SlotDemand::Exactly(NonZeroU16::MIN)
+    }
 }
 
 impl AgentStreamingRequest for ContinueFromRawPromptParams {
+    const INFERENCE_CAPABILITY: InferenceCapability = InferenceCapability::TextGeneration;
+
     type Response = GeneratedTokenResult;
 
     fn response_senders(
@@ -37,9 +67,31 @@ impl AgentStreamingRequest for ContinueFromRawPromptParams {
     ) -> &Arc<ResponseSenders<Self::Response>> {
         &agent_response_senders.generated_tokens
     }
+
+    fn slot_demand(&self) -> SlotDemand {
+        SlotDemand::Exactly(NonZeroU16::MIN)
+    }
+}
+
+impl AgentStreamingRequest for DecideParams {
+    const INFERENCE_CAPABILITY: InferenceCapability = InferenceCapability::Decisions;
+
+    type Response = DecisionResult;
+
+    fn response_senders(
+        agent_response_senders: &AgentResponseSenders,
+    ) -> &Arc<ResponseSenders<Self::Response>> {
+        &agent_response_senders.decision
+    }
+
+    fn slot_demand(&self) -> SlotDemand {
+        SlotDemand::Exactly(self.required_slots())
+    }
 }
 
 impl AgentStreamingRequest for GenerateEmbeddingBatchParams {
+    const INFERENCE_CAPABILITY: InferenceCapability = InferenceCapability::Embeddings;
+
     type Response = EmbeddingResult;
 
     fn response_senders(
@@ -47,4 +99,117 @@ impl AgentStreamingRequest for GenerateEmbeddingBatchParams {
     ) -> &Arc<ResponseSenders<Self::Response>> {
         &agent_response_senders.embedding
     }
+
+    fn slot_demand(&self) -> SlotDemand {
+        SlotDemand::UpTo(u16::try_from(self.input_batch.len()).unwrap_or(u16::MAX))
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::num::NonZeroU16;
+    use std::num::NonZeroU32;
+
+    use paddler_messaging::conversation_history::ConversationHistory;
+    use paddler_messaging::decision_question::DecisionQuestion;
+    use paddler_messaging::decision_slots_minimum::DECISION_SLOTS_MINIMUM;
+    use paddler_messaging::embedding_input_document::EmbeddingInputDocument;
+    use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
+    use paddler_messaging::inference_capability::InferenceCapability;
+    use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
+    use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
+    use paddler_messaging::request_params::decide_params::DecideParams;
+    use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
+
+    use super::AgentStreamingRequest as _;
+    use crate::dispatch_requirement::DispatchRequirement;
+    use crate::slot_demand::SlotDemand;
+
+    fn decision_question(id: &str) -> DecisionQuestion {
+        DecisionQuestion {
+            id: id.to_owned(),
+            instructions: "Pick one".to_owned(),
+            options: vec!["yes".to_owned(), "no".to_owned()],
+        }
+    }
+
+    fn embedding_batch(documents: usize) -> GenerateEmbeddingBatchParams {
+        GenerateEmbeddingBatchParams {
+            input_batch: (0..documents)
+                .map(|document_index| EmbeddingInputDocument {
+                    content: "content".to_owned(),
+                    id: document_index.to_string(),
+                })
+                .collect(),
+            normalization_method: EmbeddingNormalizationMethod::None,
+        }
+    }
+
+    #[test]
+    fn a_text_generation_request_demands_exactly_one_text_generation_slot() {
+        let one_text_generation_slot = DispatchRequirement {
+            inference_capability: InferenceCapability::TextGeneration,
+            slot_demand: SlotDemand::Exactly(NonZeroU16::MIN),
+        };
+
+        assert_eq!(
+            [
+                ContinueFromRawPromptParams {
+                    grammar: None,
+                    max_tokens: NonZeroU32::MIN,
+                    raw_prompt: "prompt".to_owned(),
+                }
+                .dispatch_requirement(),
+                ContinueFromConversationHistoryParams {
+                    add_generation_prompt: true,
+                    conversation_history: ConversationHistory::new(Vec::new()),
+                    enable_thinking: false,
+                    grammar: None,
+                    max_tokens: NonZeroU32::MIN,
+                    parse_tool_calls: false,
+                    tools: Vec::new(),
+                }
+                .dispatch_requirement(),
+            ],
+            [one_text_generation_slot, one_text_generation_slot]
+        );
+    }
+
+    #[test]
+    fn a_decision_with_leading_questions_demands_its_state_and_question_lane_slots() {
+        assert_eq!(
+            DecideParams {
+                last_question: decision_question("last"),
+                leading_questions: vec![decision_question("leading")],
+                state: "state".to_owned(),
+            }
+            .dispatch_requirement(),
+            DispatchRequirement {
+                inference_capability: InferenceCapability::Decisions,
+                slot_demand: SlotDemand::Exactly(DECISION_SLOTS_MINIMUM),
+            }
+        );
+    }
+
+    #[test]
+    fn an_embedding_batch_demands_up_to_one_slot_per_document() {
+        assert_eq!(
+            embedding_batch(3).dispatch_requirement(),
+            DispatchRequirement {
+                inference_capability: InferenceCapability::Embeddings,
+                slot_demand: SlotDemand::UpTo(3),
+            }
+        );
+    }
+
+    #[test]
+    fn an_embedding_batch_longer_than_the_slot_counter_demands_every_slot_it_can_count() {
+        assert_eq!(
+            embedding_batch(usize::from(u16::MAX) + 1).dispatch_requirement(),
+            DispatchRequirement {
+                inference_capability: InferenceCapability::Embeddings,
+                slot_demand: SlotDemand::UpTo(u16::MAX),
+            }
+        );
+    }
 }
diff --git a/paddler_balancer/src/awaiting_agent_relay.rs b/paddler_balancer/src/awaiting_agent_relay.rs
new file mode 100644
index 000000000..ed9ad2033
--- /dev/null
+++ b/paddler_balancer/src/awaiting_agent_relay.rs
@@ -0,0 +1,73 @@
+use std::sync::Arc;
+use std::time::Duration;
+
+use log::warn;
+use tokio::select;
+use tokio_util::sync::CancellationToken;
+use tokio_util::sync::DropGuard;
+
+use crate::agent_streaming_request::AgentStreamingRequest;
+use crate::buffered_request_agent_wait_result::BufferedRequestAgentWaitResult;
+use crate::buffered_request_manager::BufferedRequestManager;
+use crate::dispatchable_request::DispatchableRequest;
+use crate::dispatched_agent_relay::DispatchedAgentRelay;
+use crate::relay_failure::RelayFailure;
+
+pub struct AwaitingAgentRelay<TParams> {
+    pub buffered_request_manager: Arc<BufferedRequestManager>,
+    pub dispatchable_request: DispatchableRequest<TParams>,
+    pub inference_item_timeout: Duration,
+    pub shutdown: CancellationToken,
+}
+
+impl<TParams> AwaitingAgentRelay<TParams>
+where
+    TParams: AgentStreamingRequest,
+{
+    pub async fn dispatch(
+        self,
+        settlement_guard: DropGuard,
+    ) -> Result<DispatchedAgentRelay<TParams::Response>, RelayFailure> {
+        let Self {
+            buffered_request_manager,
+            dispatchable_request:
+                DispatchableRequest {
+                    dispatch_requirement,
+                    params,
+                },
+            inference_item_timeout,
+            shutdown,
+        } = self;
+
+        let buffered_request_agent_wait_result = select! {
+            biased;
+
+            () = shutdown.cancelled() => return Err(RelayFailure::BalancerShuttingDown),
+            buffered_request_agent_wait_result = buffered_request_manager.wait_for_available_agent(dispatch_requirement) => {
+                buffered_request_agent_wait_result
+            }
+        };
+
+        match buffered_request_agent_wait_result {
+            BufferedRequestAgentWaitResult::Found(dispatched_agent) => {
+                Ok(DispatchedAgentRelay::start(
+                    dispatched_agent,
+                    params,
+                    inference_item_timeout,
+                    settlement_guard.disarm(),
+                    shutdown,
+                ))
+            }
+            BufferedRequestAgentWaitResult::BufferOverflow => {
+                warn!("Too many buffered requests, dropping a request");
+
+                Err(RelayFailure::BufferOverflow)
+            }
+            BufferedRequestAgentWaitResult::Timeout => {
+                warn!("A buffered request timed out waiting for an available slot");
+
+                Err(RelayFailure::BufferedRequestTimedOut)
+            }
+        }
+    }
+}
diff --git a/paddler_balancer/src/balancer_addresses.rs b/paddler_balancer/src/balancer_addresses.rs
index 0ded12423..762f843aa 100644
--- a/paddler_balancer/src/balancer_addresses.rs
+++ b/paddler_balancer/src/balancer_addresses.rs
@@ -10,6 +10,7 @@ use serde_json::to_writer;
 #[serde(deny_unknown_fields)]
 pub struct BalancerAddresses {
     pub compat_openai: Option<SocketAddr>,
+    pub compat_typesafe: Option<SocketAddr>,
     pub inference: SocketAddr,
     pub management: SocketAddr,
     pub web_admin_panel: Option<SocketAddr>,
@@ -34,6 +35,7 @@ mod tests {
     fn inference_and_management_only() -> BalancerAddresses {
         BalancerAddresses {
             compat_openai: None,
+            compat_typesafe: None,
             inference: SocketAddr::from((Ipv4Addr::LOCALHOST, 8061)),
             management: SocketAddr::from((Ipv4Addr::LOCALHOST, 8060)),
             web_admin_panel: None,
@@ -50,7 +52,7 @@ mod tests {
 
         assert_eq!(
             String::from_utf8(announcement).expect("the announcement must be UTF-8"),
-            "{\"compat_openai\":null,\"inference\":\"127.0.0.1:8061\",\"management\":\"127.0.0.1:8060\",\"web_admin_panel\":null}\n"
+            "{\"compat_openai\":null,\"compat_typesafe\":null,\"inference\":\"127.0.0.1:8061\",\"management\":\"127.0.0.1:8060\",\"web_admin_panel\":null}\n"
         );
     }
 
diff --git a/paddler_balancer/src/balancer_applicable_state.rs b/paddler_balancer/src/balancer_applicable_state.rs
index ae2bc531e..e0d22ec0e 100644
--- a/paddler_balancer/src/balancer_applicable_state.rs
+++ b/paddler_balancer/src/balancer_applicable_state.rs
@@ -4,69 +4,26 @@ use paddler_messaging::balancer_desired_state::BalancerDesiredState;
 #[derive(Clone, Debug)]
 pub struct BalancerApplicableState {
     pub agent_desired_state: AgentDesiredState,
+    pub balancer_desired_state: BalancerDesiredState,
+    pub version: u64,
 }
 
-impl From<BalancerDesiredState> for BalancerApplicableState {
-    fn from(
-        BalancerDesiredState {
-            chat_template_override,
-            inference_parameters,
-            model,
-            multimodal_projection,
-            use_chat_template_override,
-        }: BalancerDesiredState,
-    ) -> Self {
-        Self {
-            agent_desired_state: AgentDesiredState {
-                chat_template_override: if use_chat_template_override {
-                    chat_template_override
-                } else {
-                    None
-                },
-                inference_parameters,
-                model,
-                multimodal_projection,
-            },
-        }
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-    use paddler_messaging::chat_template::ChatTemplate;
-
-    use super::BalancerApplicableState;
-
-    fn desired_state_with_override(use_chat_template_override: bool) -> BalancerDesiredState {
-        BalancerDesiredState {
-            chat_template_override: Some(ChatTemplate {
-                content: "{{ messages }}".to_owned(),
-            }),
-            use_chat_template_override,
-            ..BalancerDesiredState::default()
-        }
+impl BalancerApplicableState {
+    #[must_use]
+    pub fn initial(balancer_desired_state: BalancerDesiredState) -> Self {
+        Self::at_version(balancer_desired_state, 0)
     }
 
-    #[test]
-    fn applies_the_chat_template_override_when_it_is_enabled() {
-        assert_eq!(
-            BalancerApplicableState::from(desired_state_with_override(true))
-                .agent_desired_state
-                .chat_template_override,
-            Some(ChatTemplate {
-                content: "{{ messages }}".to_owned(),
-            })
-        );
+    #[must_use]
+    pub fn superseded_by(&self, balancer_desired_state: BalancerDesiredState) -> Self {
+        Self::at_version(balancer_desired_state, self.version + 1)
     }
 
-    #[test]
-    fn ignores_the_chat_template_override_when_it_is_disabled() {
-        assert_eq!(
-            BalancerApplicableState::from(desired_state_with_override(false))
-                .agent_desired_state
-                .chat_template_override,
-            None
-        );
+    fn at_version(balancer_desired_state: BalancerDesiredState, version: u64) -> Self {
+        Self {
+            agent_desired_state: AgentDesiredState::from(&balancer_desired_state),
+            balancer_desired_state,
+            version,
+        }
     }
 }
diff --git a/paddler_balancer/src/balancer_applicable_state_holder.rs b/paddler_balancer/src/balancer_applicable_state_holder.rs
index a56c78ca2..bea4cf727 100644
--- a/paddler_balancer/src/balancer_applicable_state_holder.rs
+++ b/paddler_balancer/src/balancer_applicable_state_holder.rs
@@ -1,132 +1,105 @@
-use parking_lot::RwLock;
+use std::sync::Arc;
+
 use tokio::sync::watch;
 
+use paddler_messaging::agent_decisions_settings::AgentDecisionsSettings;
 use paddler_messaging::agent_desired_state::AgentDesiredState;
-use paddler_messaging::subscribes_to_updates::SubscribesToUpdates;
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
 
 use crate::balancer_applicable_state::BalancerApplicableState;
-use crate::cluster_token_generation_mode::ClusterTokenGenerationMode;
+use crate::cluster_serves_no_decisions::ClusterServesNoDecisions;
 
 pub struct BalancerApplicableStateHolder {
-    balancer_applicable_state: RwLock<BalancerApplicableState>,
-    update_tx: watch::Sender<()>,
+    balancer_applicable_state_rx: watch::Receiver<Arc<BalancerApplicableState>>,
 }
 
 impl BalancerApplicableStateHolder {
     #[must_use]
-    pub fn new(balancer_applicable_state: BalancerApplicableState) -> Self {
-        let (update_tx, _initial_rx) = watch::channel(());
-
+    pub const fn new(
+        balancer_applicable_state_rx: watch::Receiver<Arc<BalancerApplicableState>>,
+    ) -> Self {
         Self {
-            balancer_applicable_state: RwLock::new(balancer_applicable_state),
-            update_tx,
+            balancer_applicable_state_rx,
         }
     }
 
+    #[must_use]
     pub fn get_agent_desired_state(&self) -> AgentDesiredState {
-        self.balancer_applicable_state
-            .read()
+        self.balancer_applicable_state_rx
+            .borrow()
             .agent_desired_state
             .clone()
     }
 
-    pub fn get_balancer_applicable_state(&self) -> BalancerApplicableState {
-        self.balancer_applicable_state.read().clone()
+    #[must_use]
+    pub fn get_balancer_applicable_state(&self) -> Arc<BalancerApplicableState> {
+        self.balancer_applicable_state_rx.borrow().clone()
     }
 
     #[must_use]
-    pub fn token_generation_mode(&self) -> ClusterTokenGenerationMode {
-        if self
-            .balancer_applicable_state
-            .read()
+    pub fn get_balancer_desired_state(&self) -> BalancerDesiredState {
+        self.balancer_applicable_state_rx
+            .borrow()
+            .balancer_desired_state
+            .clone()
+    }
+
+    pub fn require_decisions(&self) -> Result<(), ClusterServesNoDecisions> {
+        match self
+            .balancer_applicable_state_rx
+            .borrow()
             .agent_desired_state
-            .inference_parameters
-            .enable_embeddings
+            .decisions
         {
-            ClusterTokenGenerationMode::DisabledForEmbeddings
-        } else {
-            ClusterTokenGenerationMode::Enabled
+            AgentDecisionsSettings::NotServed => Err(ClusterServesNoDecisions),
+            AgentDecisionsSettings::Served(_) => Ok(()),
         }
     }
 
-    pub fn set_balancer_applicable_state(
-        &self,
-        balancer_applicable_state: BalancerApplicableState,
-    ) {
-        *self.balancer_applicable_state.write() = balancer_applicable_state;
-
-        self.update_tx.send_replace(());
-    }
-}
-
-impl SubscribesToUpdates for BalancerApplicableStateHolder {
-    fn subscribe_to_updates(&self) -> watch::Receiver<()> {
-        self.update_tx.subscribe()
+    #[must_use]
+    pub fn subscribe(&self) -> watch::Receiver<Arc<BalancerApplicableState>> {
+        self.balancer_applicable_state_rx.clone()
     }
 }
 
 #[cfg(test)]
 mod tests {
-    use paddler_inference_parameters::inference_parameters::InferenceParameters;
+    use std::sync::Arc;
+
+    use tokio::sync::watch;
+
     use paddler_messaging::agent_desired_model::AgentDesiredModel;
-    use paddler_messaging::agent_desired_state::AgentDesiredState;
     use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-    use paddler_messaging::subscribes_to_updates::SubscribesToUpdates as _;
+    use paddler_messaging::decision_settings::DecisionSettings;
 
     use super::BalancerApplicableStateHolder;
     use crate::balancer_applicable_state::BalancerApplicableState;
-    use crate::cluster_token_generation_mode::ClusterTokenGenerationMode;
-
-    fn holder_with_embeddings(enable_embeddings: bool) -> BalancerApplicableStateHolder {
-        BalancerApplicableStateHolder::new(BalancerApplicableState {
-            agent_desired_state: AgentDesiredState {
-                chat_template_override: None,
-                inference_parameters: InferenceParameters {
-                    enable_embeddings,
-                    ..InferenceParameters::default()
-                },
-                model: AgentDesiredModel::LocalToAgent("model.gguf".to_owned()),
-                multimodal_projection: AgentDesiredModel::None,
-            },
-        })
-    }
-
-    #[test]
-    fn generates_tokens_when_embeddings_are_disabled() {
-        assert_eq!(
-            holder_with_embeddings(false).token_generation_mode(),
-            ClusterTokenGenerationMode::Enabled
+    use crate::cluster_serves_no_decisions::ClusterServesNoDecisions;
+
+    fn holder_of(pointer_head: AgentDesiredModel) -> BalancerApplicableStateHolder {
+        let (_balancer_applicable_state_tx, balancer_applicable_state_rx) = watch::channel(
+            Arc::new(BalancerApplicableState::initial(BalancerDesiredState {
+                decision: DecisionSettings { pointer_head },
+                ..BalancerDesiredState::default()
+            })),
         );
+
+        BalancerApplicableStateHolder::new(balancer_applicable_state_rx)
     }
 
     #[test]
-    fn disables_token_generation_when_embeddings_are_enabled() {
+    fn grants_decisions_while_a_pointer_head_is_configured() {
         assert_eq!(
-            holder_with_embeddings(true).token_generation_mode(),
-            ClusterTokenGenerationMode::DisabledForEmbeddings
+            holder_of(AgentDesiredModel::Uri("pointer_head.gguf".to_owned())).require_decisions(),
+            Ok(())
         );
     }
 
     #[test]
-    fn replacing_the_state_notifies_subscribers_and_is_read_back() {
-        let holder = BalancerApplicableStateHolder::new(BalancerApplicableState::from(
-            BalancerDesiredState::default(),
-        ));
-        let update_rx = holder.subscribe_to_updates();
-
-        holder.set_balancer_applicable_state(BalancerApplicableState::from(BalancerDesiredState {
-            model: AgentDesiredModel::LocalToAgent("model.gguf".to_owned()),
-            ..BalancerDesiredState::default()
-        }));
-
-        assert!(
-            update_rx
-                .has_changed()
-                .expect("the holder must keep its update sender")
-        );
+    fn refuses_decisions_while_no_pointer_head_is_configured() {
         assert_eq!(
-            holder.get_agent_desired_state().model,
-            AgentDesiredModel::LocalToAgent("model.gguf".to_owned())
+            holder_of(AgentDesiredModel::None).require_decisions(),
+            Err(ClusterServesNoDecisions)
         );
     }
 }
diff --git a/paddler_balancer/src/balancer_desired_state_store.rs b/paddler_balancer/src/balancer_desired_state_store.rs
new file mode 100644
index 000000000..5b16057d0
--- /dev/null
+++ b/paddler_balancer/src/balancer_desired_state_store.rs
@@ -0,0 +1,242 @@
+use std::sync::Arc;
+
+use tokio::sync::Mutex;
+use tokio::sync::watch;
+
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_spawn_resuming_panic::spawn_resuming_panic::spawn_resuming_panic;
+use paddler_state_database::state_database::StateDatabase;
+use paddler_state_database::state_database_error::StateDatabaseError;
+use paddler_state_database::stored_balancer_desired_state::StoredBalancerDesiredState;
+
+use crate::balancer_applicable_state::BalancerApplicableState;
+use crate::balancer_applicable_state_holder::BalancerApplicableStateHolder;
+
+pub struct BalancerDesiredStateStore {
+    balancer_applicable_state_tx: watch::Sender<Arc<BalancerApplicableState>>,
+    replacement_lock: Arc<Mutex<()>>,
+    state_database: Arc<StateDatabase>,
+}
+
+impl BalancerDesiredStateStore {
+    pub async fn load(state_database: Arc<StateDatabase>) -> Result<Self, StateDatabaseError> {
+        let balancer_desired_state = match state_database.read_balancer_desired_state().await? {
+            StoredBalancerDesiredState::NothingStored => BalancerDesiredState::default(),
+            StoredBalancerDesiredState::Stored(stored_desired_state) => *stored_desired_state,
+        };
+        let (balancer_applicable_state_tx, _initial_balancer_applicable_state_rx) = watch::channel(
+            Arc::new(BalancerApplicableState::initial(balancer_desired_state)),
+        );
+
+        Ok(Self {
+            balancer_applicable_state_tx,
+            replacement_lock: Arc::new(Mutex::new(())),
+            state_database,
+        })
+    }
+
+    #[must_use]
+    pub fn applicable_state_holder(&self) -> BalancerApplicableStateHolder {
+        BalancerApplicableStateHolder::new(self.balancer_applicable_state_tx.subscribe())
+    }
+
+    pub async fn replace(
+        &self,
+        balancer_desired_state: BalancerDesiredState,
+    ) -> Result<(), StateDatabaseError> {
+        let replacement_guard = self.replacement_lock.clone().lock_owned().await;
+        let balancer_applicable_state_tx = self.balancer_applicable_state_tx.clone();
+        let state_database = self.state_database.clone();
+
+        spawn_resuming_panic(async move {
+            if balancer_applicable_state_tx.borrow().balancer_desired_state
+                == balancer_desired_state
+            {
+                return Ok(());
+            }
+
+            state_database
+                .store_balancer_desired_state(&balancer_desired_state)
+                .await?;
+
+            let superseding_applicable_state = balancer_applicable_state_tx
+                .borrow()
+                .superseded_by(balancer_desired_state);
+
+            balancer_applicable_state_tx.send_replace(Arc::new(superseding_applicable_state));
+            drop(replacement_guard);
+
+            Ok(())
+        })
+        .await
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::sync::Arc;
+
+    use futures_util::FutureExt as _;
+    use tempfile::TempDir;
+
+    use paddler_blocking_thread_fixture::occupied_blocking_thread::OccupiedBlockingThread;
+    use paddler_blocking_thread_fixture::single_blocking_thread_runtime::single_blocking_thread_runtime;
+
+    use paddler_messaging::agent_desired_model::AgentDesiredModel;
+    use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+    use paddler_state_database::state_database::StateDatabase;
+    use paddler_state_database::state_database_file::StateDatabaseFile;
+    use paddler_state_database::stored_balancer_desired_state::StoredBalancerDesiredState;
+
+    use super::BalancerDesiredStateStore;
+
+    fn state_with_model(model_uri: &str) -> BalancerDesiredState {
+        BalancerDesiredState {
+            model: AgentDesiredModel::Uri(model_uri.to_owned()),
+            ..BalancerDesiredState::default()
+        }
+    }
+
+    #[tokio::test]
+    async fn starts_from_the_default_state_when_nothing_is_stored() {
+        let store = BalancerDesiredStateStore::load(Arc::new(StateDatabase::in_memory(
+            StoredBalancerDesiredState::NothingStored,
+        )))
+        .await
+        .unwrap();
+
+        assert_eq!(
+            store.applicable_state_holder().get_balancer_desired_state(),
+            BalancerDesiredState::default()
+        );
+    }
+
+    #[tokio::test]
+    async fn starts_from_the_stored_state() {
+        let store = BalancerDesiredStateStore::load(Arc::new(StateDatabase::in_memory(
+            StoredBalancerDesiredState::Stored(Box::new(state_with_model("stored_model_uri"))),
+        )))
+        .await
+        .unwrap();
+
+        assert_eq!(
+            store.applicable_state_holder().get_balancer_desired_state(),
+            state_with_model("stored_model_uri")
+        );
+    }
+
+    #[tokio::test]
+    async fn a_replaced_state_is_stored_and_published() {
+        let state_database = Arc::new(StateDatabase::in_memory(
+            StoredBalancerDesiredState::NothingStored,
+        ));
+        let store = BalancerDesiredStateStore::load(state_database.clone())
+            .await
+            .unwrap();
+        let mut applicable_state_rx = store.applicable_state_holder().subscribe();
+
+        store
+            .replace(state_with_model("replacing_model_uri"))
+            .await
+            .unwrap();
+
+        assert_eq!(
+            state_database.read_balancer_desired_state().await.unwrap(),
+            StoredBalancerDesiredState::Stored(Box::new(state_with_model("replacing_model_uri")))
+        );
+        assert_eq!(
+            applicable_state_rx
+                .borrow_and_update()
+                .balancer_desired_state,
+            state_with_model("replacing_model_uri")
+        );
+    }
+
+    #[tokio::test]
+    async fn an_identical_state_is_neither_stored_nor_published() {
+        let state_database = Arc::new(StateDatabase::in_memory(
+            StoredBalancerDesiredState::NothingStored,
+        ));
+        let store = BalancerDesiredStateStore::load(state_database.clone())
+            .await
+            .unwrap();
+        let applicable_state_rx = store.applicable_state_holder().subscribe();
+
+        store
+            .replace(BalancerDesiredState::default())
+            .await
+            .unwrap();
+
+        assert_eq!(
+            state_database.read_balancer_desired_state().await.unwrap(),
+            StoredBalancerDesiredState::NothingStored
+        );
+        assert!(!applicable_state_rx.has_changed().unwrap());
+    }
+
+    #[test]
+    fn a_replacement_its_caller_abandons_is_applied_before_the_next_one() {
+        single_blocking_thread_runtime().unwrap().block_on(async {
+            let temp_dir = TempDir::new().unwrap();
+            let state_path = temp_dir.path().join("state.json");
+            let store = BalancerDesiredStateStore::load(Arc::new(StateDatabase::File(
+                StateDatabaseFile::new(state_path.clone()),
+            )))
+            .await
+            .unwrap();
+            let occupied_blocking_thread = OccupiedBlockingThread::occupy();
+
+            assert!(
+                store
+                    .replace(state_with_model("abandoned_model_uri"))
+                    .now_or_never()
+                    .is_none()
+            );
+
+            occupied_blocking_thread.release().await.unwrap();
+            store
+                .replace(state_with_model("next_model_uri"))
+                .await
+                .unwrap();
+
+            assert_eq!(
+                store
+                    .applicable_state_holder()
+                    .get_balancer_applicable_state()
+                    .version,
+                2
+            );
+            assert_eq!(
+                StateDatabaseFile::new(state_path)
+                    .read_balancer_desired_state()
+                    .await
+                    .unwrap(),
+                StoredBalancerDesiredState::Stored(Box::new(state_with_model("next_model_uri")))
+            );
+        });
+    }
+
+    #[tokio::test]
+    async fn a_state_that_cannot_be_stored_is_not_published() {
+        let directory_in_place_of_the_state_file = TempDir::new().unwrap();
+        let store =
+            BalancerDesiredStateStore::load(Arc::new(StateDatabase::File(StateDatabaseFile::new(
+                directory_in_place_of_the_state_file
+                    .path()
+                    .join("state.json"),
+            ))))
+            .await
+            .unwrap();
+        let applicable_state_rx = store.applicable_state_holder().subscribe();
+
+        directory_in_place_of_the_state_file.close().unwrap();
+
+        assert!(
+            store
+                .replace(state_with_model("unstorable_model_uri"))
+                .await
+                .is_err()
+        );
+        assert!(!applicable_state_rx.has_changed().unwrap());
+    }
+}
diff --git a/paddler_balancer/src/balancer_shutdown_error.rs b/paddler_balancer/src/balancer_shutdown_error.rs
deleted file mode 100644
index 0dd5f3a5c..000000000
--- a/paddler_balancer/src/balancer_shutdown_error.rs
+++ /dev/null
@@ -1,9 +0,0 @@
-use paddler_messaging::jsonrpc::error::Error as JsonRpcError;
-
-#[must_use]
-pub fn balancer_shutdown_error() -> JsonRpcError {
-    JsonRpcError {
-        code: 503,
-        description: "balancer is shutting down".to_owned(),
-    }
-}
diff --git a/paddler_balancer/src/buffered_request_counter.rs b/paddler_balancer/src/buffered_request_counter.rs
index 441f05b86..95da444e4 100644
--- a/paddler_balancer/src/buffered_request_counter.rs
+++ b/paddler_balancer/src/buffered_request_counter.rs
@@ -23,7 +23,7 @@ impl BufferedRequestCounter {
     }
 
     pub fn decrement(&self) {
-        self.count.decrement();
+        self.count.subtract(1);
         self.update_tx.send_replace(());
     }
 
diff --git a/paddler_balancer/src/buffered_request_manager.rs b/paddler_balancer/src/buffered_request_manager.rs
index a4747c06b..d2c6d9759 100644
--- a/paddler_balancer/src/buffered_request_manager.rs
+++ b/paddler_balancer/src/buffered_request_manager.rs
@@ -1,8 +1,6 @@
 use std::sync::Arc;
 use std::time::Duration;
 
-use anyhow::Error;
-use anyhow::Result;
 use tokio::sync::watch;
 use tokio::time::error::Elapsed;
 use tokio::time::timeout;
@@ -14,6 +12,8 @@ use paddler_messaging::subscribes_to_updates::SubscribesToUpdates;
 use crate::agent_controller_pool::AgentControllerPool;
 use crate::buffered_request_agent_wait_result::BufferedRequestAgentWaitResult;
 use crate::buffered_request_counter::BufferedRequestCounter;
+use crate::dispatch_position::DispatchPosition;
+use crate::dispatch_requirement::DispatchRequirement;
 use crate::dispatched_agent::DispatchedAgent;
 
 pub struct BufferedRequestManager {
@@ -44,37 +44,43 @@ impl BufferedRequestManager {
     }
 
     #[must_use]
-    pub fn take_available_agent(&self) -> Option<DispatchedAgent> {
+    pub fn take_available_agent(
+        &self,
+        dispatch_requirement: DispatchRequirement,
+    ) -> Option<DispatchedAgent> {
         self.agent_controller_pool
-            .take_least_busy_agent_controller()
+            .take_least_busy_agent_controller(dispatch_requirement, DispatchPosition::Arriving)
     }
 
-    pub async fn wait_for_available_agent(&self) -> Result<BufferedRequestAgentWaitResult> {
-        if let Some(dispatched_agent) = self.take_available_agent() {
-            return Ok(BufferedRequestAgentWaitResult::Found(dispatched_agent));
+    pub async fn wait_for_available_agent(
+        &self,
+        dispatch_requirement: DispatchRequirement,
+    ) -> BufferedRequestAgentWaitResult {
+        let agent_availability_changed = self.agent_controller_pool.agent_availability_changed();
+
+        if let Some(dispatched_agent) = self.take_available_agent(dispatch_requirement) {
+            return BufferedRequestAgentWaitResult::Found(dispatched_agent);
         }
 
-        let Some(_buffered_request_count_guard) = self.buffered_request_counter.try_admit() else {
-            return Ok(BufferedRequestAgentWaitResult::BufferOverflow);
+        let Some(waiting_line_place) = self
+            .agent_controller_pool
+            .join_waiting_line_when_admitted(dispatch_requirement, &self.buffered_request_counter)
+        else {
+            return BufferedRequestAgentWaitResult::BufferOverflow;
         };
-        let agent_controller_pool = self.agent_controller_pool.clone();
-        let mut update_rx = agent_controller_pool.subscribe_to_updates();
-
-        match timeout(self.buffered_request_timeout, async {
-            loop {
-                if let Some(dispatched_agent) =
-                    agent_controller_pool.take_least_busy_agent_controller()
-                {
-                    return Ok::<_, Error>(BufferedRequestAgentWaitResult::Found(dispatched_agent));
-                }
 
-                update_rx.changed().await?;
-            }
-        })
+        match timeout(
+            self.buffered_request_timeout,
+            self.agent_controller_pool.next_available_agent(
+                agent_availability_changed,
+                dispatch_requirement,
+                waiting_line_place.dispatch_position(),
+            ),
+        )
         .await
         {
-            Ok(inner_result) => Ok(inner_result?),
-            Err(Elapsed { .. }) => Ok(BufferedRequestAgentWaitResult::Timeout),
+            Ok(dispatched_agent) => BufferedRequestAgentWaitResult::Found(dispatched_agent),
+            Err(Elapsed { .. }) => BufferedRequestAgentWaitResult::Timeout,
         }
     }
 }
@@ -99,23 +105,29 @@ impl SubscribesToUpdates for BufferedRequestManager {
 mod tests {
     use std::hint::spin_loop;
     use std::mem::discriminant;
+    use std::num::NonZeroU32;
     use std::sync::Arc;
     use std::sync::atomic::AtomicBool;
     use std::thread::available_parallelism;
     use std::thread::scope;
     use std::time::Duration;
 
+    use futures::FutureExt as _;
+    use tokio::pin;
     use tokio::runtime::Builder;
     use tokio::select;
     use tokio::sync::mpsc;
     use tokio_util::sync::CancellationToken;
 
     use paddler_messaging::atomic_value::AtomicValue;
+    use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
     use paddler_messaging::subscribes_to_updates::SubscribesToUpdates as _;
 
     use super::BufferedRequestManager;
     use crate::agent_controller_pool::AgentControllerPool;
+    use crate::agent_streaming_request::AgentStreamingRequest as _;
     use crate::buffered_request_agent_wait_result::BufferedRequestAgentWaitResult;
+    use crate::dispatch_requirement::DispatchRequirement;
 
     const BUFFER_CAPACITY: u64 = 1;
     const RACE_ROUNDS: usize = 64;
@@ -125,12 +137,22 @@ mod tests {
         overflowed_callers: usize,
     }
 
+    fn raw_prompt_requirement() -> DispatchRequirement {
+        ContinueFromRawPromptParams {
+            grammar: None,
+            max_tokens: NonZeroU32::MIN,
+            raw_prompt: "hello".to_owned(),
+        }
+        .dispatch_requirement()
+    }
+
     fn race_callers_for_one_buffered_place(racing_callers: usize) -> RaceOutcome {
         let buffered_request_manager = BufferedRequestManager::new(
             Arc::new(AgentControllerPool::default()),
             Duration::MAX,
             BUFFER_CAPACITY,
         );
+        let raw_prompt_requirement = raw_prompt_requirement();
         let start_line = AtomicValue::<AtomicBool>::new(false);
         let admitted_callers_release = CancellationToken::new();
         let (overflow_tx, mut overflow_rx) = mpsc::unbounded_channel();
@@ -156,14 +178,12 @@ mod tests {
                         caller_runtime.block_on(async {
                             if let Some(wait_result) = admitted_callers_release
                                 .run_until_cancelled(
-                                    buffered_request_manager.wait_for_available_agent(),
+                                    buffered_request_manager
+                                        .wait_for_available_agent(raw_prompt_requirement),
                                 )
                                 .await
-                                && discriminant(
-                                    &wait_result.expect("waiting for an agent must not fail"),
-                                ) == discriminant(
-                                    &BufferedRequestAgentWaitResult::BufferOverflow,
-                                )
+                                && discriminant(&wait_result)
+                                    == discriminant(&BufferedRequestAgentWaitResult::BufferOverflow)
                             {
                                 overflow_tx
                                     .send(())
@@ -233,4 +253,25 @@ mod tests {
             assert_eq!(overflowed_callers, racing_callers - 1);
         }
     }
+
+    #[tokio::test]
+    async fn a_request_the_full_buffer_refuses_wakes_no_waiting_request() {
+        let agent_controller_pool = Arc::new(AgentControllerPool::default());
+        let buffered_request_manager =
+            BufferedRequestManager::new(agent_controller_pool.clone(), Duration::MAX, 0);
+        let agent_availability_changed = agent_controller_pool.agent_availability_changed();
+
+        pin!(agent_availability_changed);
+        agent_availability_changed.as_mut().enable();
+
+        let wait_result = buffered_request_manager
+            .wait_for_available_agent(raw_prompt_requirement())
+            .await;
+
+        assert_eq!(
+            discriminant(&wait_result),
+            discriminant(&BufferedRequestAgentWaitResult::BufferOverflow)
+        );
+        assert_eq!(agent_availability_changed.now_or_never(), None);
+    }
 }
diff --git a/paddler_balancer/src/chunk_forwarding_session_controller/identity_transformer.rs b/paddler_balancer/src/chunk_forwarding_session_controller/identity_transformer.rs
index 74f92d6d0..087cfe7af 100644
--- a/paddler_balancer/src/chunk_forwarding_session_controller/identity_transformer.rs
+++ b/paddler_balancer/src/chunk_forwarding_session_controller/identity_transformer.rs
@@ -1,13 +1,9 @@
-use async_trait::async_trait;
-use serde_json::to_string;
-
 use paddler_messaging::inference_client::message::Message as OutgoingMessage;
 
 use super::transform_result::TransformResult;
 use super::transforms_outgoing_message::TransformsOutgoingMessage;
-use crate::agent_relay_error::AgentRelayError;
 
-#[derive(Clone, Default)]
+#[derive(Default)]
 pub struct IdentityTransformer;
 
 impl IdentityTransformer {
@@ -17,16 +13,10 @@ impl IdentityTransformer {
     }
 }
 
-#[async_trait]
 impl TransformsOutgoingMessage for IdentityTransformer {
     type Output = TransformResult;
 
-    async fn transform(
-        &self,
-        message: OutgoingMessage,
-    ) -> Result<Vec<TransformResult>, AgentRelayError> {
-        to_string(&message)
-            .map(|serialized| vec![TransformResult::Chunk(serialized)])
-            .map_err(AgentRelayError::MessageUnserializable)
+    fn transform(&self, message: OutgoingMessage) -> Vec<TransformResult> {
+        vec![TransformResult::serialized_from(&message)]
     }
 }
diff --git a/paddler_balancer/src/chunk_forwarding_session_controller/mod.rs b/paddler_balancer/src/chunk_forwarding_session_controller/mod.rs
index 49333d645..f4f59a522 100644
--- a/paddler_balancer/src/chunk_forwarding_session_controller/mod.rs
+++ b/paddler_balancer/src/chunk_forwarding_session_controller/mod.rs
@@ -2,6 +2,8 @@ pub mod identity_transformer;
 pub mod transform_result;
 pub mod transforms_outgoing_message;
 
+use std::future::ready;
+
 use anyhow::Result;
 use tokio::sync::mpsc;
 
@@ -10,10 +12,9 @@ use paddler_messaging::inference_client::message::Message as OutgoingMessage;
 use self::transforms_outgoing_message::TransformsOutgoingMessage;
 use crate::controls_session::ControlsSession;
 
-#[derive(Clone)]
 pub struct ChunkForwardingSessionController<TTransformsOutgoingMessage>
 where
-    TTransformsOutgoingMessage: Clone + TransformsOutgoingMessage + Send + Sync,
+    TTransformsOutgoingMessage: TransformsOutgoingMessage + Send + Sync,
 {
     chunk_tx: mpsc::UnboundedSender<TTransformsOutgoingMessage::Output>,
     transformer: TTransformsOutgoingMessage,
@@ -21,7 +22,7 @@ where
 
 impl<TTransformsOutgoingMessage> ChunkForwardingSessionController<TTransformsOutgoingMessage>
 where
-    TTransformsOutgoingMessage: Clone + TransformsOutgoingMessage + Send + Sync,
+    TTransformsOutgoingMessage: TransformsOutgoingMessage + Send + Sync,
 {
     pub const fn new(
         chunk_tx: mpsc::UnboundedSender<TTransformsOutgoingMessage::Output>,
@@ -32,18 +33,25 @@ where
             transformer,
         }
     }
+
+    fn forward_transformed(&self, message: OutgoingMessage) -> Result<()> {
+        for output in self.transformer.transform(message) {
+            self.chunk_tx.send(output)?;
+        }
+
+        Ok(())
+    }
 }
 
 impl<TTransformsOutgoingMessage> ControlsSession<OutgoingMessage>
     for ChunkForwardingSessionController<TTransformsOutgoingMessage>
 where
-    TTransformsOutgoingMessage: Clone + TransformsOutgoingMessage + Send + Sync,
+    TTransformsOutgoingMessage: TransformsOutgoingMessage + Send + Sync,
 {
-    async fn send_response(&mut self, message: OutgoingMessage) -> Result<()> {
-        for output in self.transformer.transform(message).await? {
-            self.chunk_tx.send(output)?;
-        }
-
-        Ok(())
+    fn send_response(
+        &mut self,
+        message: OutgoingMessage,
+    ) -> impl Future<Output = Result<()>> + Send {
+        ready(self.forward_transformed(message))
     }
 }
diff --git a/paddler_balancer/src/chunk_forwarding_session_controller/transform_result.rs b/paddler_balancer/src/chunk_forwarding_session_controller/transform_result.rs
index ad2ca455b..ac000a7e0 100644
--- a/paddler_balancer/src/chunk_forwarding_session_controller/transform_result.rs
+++ b/paddler_balancer/src/chunk_forwarding_session_controller/transform_result.rs
@@ -1,4 +1,7 @@
 use bytes::Bytes;
+use log::error;
+use serde::Serialize;
+use serde_json::to_string;
 
 #[derive(Debug, Eq, PartialEq)]
 pub enum TransformResult {
@@ -8,6 +11,20 @@ pub enum TransformResult {
 }
 
 impl TransformResult {
+    #[must_use]
+    pub fn serialized_from<TMessage: Serialize>(message: &TMessage) -> Self {
+        match to_string(message) {
+            Ok(serialized) => Self::Chunk(serialized),
+            Err(serialization_error) => {
+                error!(
+                    "Unable to serialize a message relayed to the client: {serialization_error}"
+                );
+
+                Self::Discard
+            }
+        }
+    }
+
     #[must_use]
     pub fn into_ndjson_line(self) -> Option<Bytes> {
         match self {
@@ -20,3 +37,22 @@ impl TransformResult {
         }
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use std::collections::BTreeMap;
+
+    use super::TransformResult;
+
+    #[test]
+    fn a_message_that_cannot_be_serialized_is_discarded() {
+        let mut message_with_a_non_string_key = BTreeMap::new();
+
+        message_with_a_non_string_key.insert(vec![0_u8], 0_u8);
+
+        assert_eq!(
+            TransformResult::serialized_from(&message_with_a_non_string_key),
+            TransformResult::Discard
+        );
+    }
+}
diff --git a/paddler_balancer/src/chunk_forwarding_session_controller/transforms_outgoing_message.rs b/paddler_balancer/src/chunk_forwarding_session_controller/transforms_outgoing_message.rs
index bc3a56502..f1156d34f 100644
--- a/paddler_balancer/src/chunk_forwarding_session_controller/transforms_outgoing_message.rs
+++ b/paddler_balancer/src/chunk_forwarding_session_controller/transforms_outgoing_message.rs
@@ -1,15 +1,7 @@
-use async_trait::async_trait;
-
 use paddler_messaging::inference_client::message::Message as OutgoingMessage;
 
-use crate::agent_relay_error::AgentRelayError;
-
-#[async_trait]
 pub trait TransformsOutgoingMessage {
     type Output: Send + Sync + 'static;
 
-    async fn transform(
-        &self,
-        message: OutgoingMessage,
-    ) -> Result<Vec<Self::Output>, AgentRelayError>;
+    fn transform(&self, message: OutgoingMessage) -> Vec<Self::Output>;
 }
diff --git a/paddler_balancer/src/cluster_serves_no_decisions.rs b/paddler_balancer/src/cluster_serves_no_decisions.rs
new file mode 100644
index 000000000..e4c580e7a
--- /dev/null
+++ b/paddler_balancer/src/cluster_serves_no_decisions.rs
@@ -0,0 +1,12 @@
+use std::fmt;
+use std::fmt::Display;
+use std::fmt::Formatter;
+
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+pub struct ClusterServesNoDecisions;
+
+impl Display for ClusterServesNoDecisions {
+    fn fmt(&self, formatter: &mut Formatter<'_>) -> fmt::Result {
+        formatter.write_str("No pointer head is configured, so the cluster serves no decisions")
+    }
+}
diff --git a/paddler_balancer/src/cluster_token_generation_mode.rs b/paddler_balancer/src/cluster_token_generation_mode.rs
deleted file mode 100644
index 49dfa5040..000000000
--- a/paddler_balancer/src/cluster_token_generation_mode.rs
+++ /dev/null
@@ -1,50 +0,0 @@
-use actix_web::Error;
-use actix_web::error::ErrorNotImplemented;
-
-pub const TOKEN_GENERATION_DISABLED_MESSAGE: &str =
-    "Token generation is disabled while the cluster is configured for embeddings";
-
-#[derive(Clone, Copy, Debug, Eq, PartialEq)]
-pub enum ClusterTokenGenerationMode {
-    Enabled,
-    DisabledForEmbeddings,
-}
-
-impl ClusterTokenGenerationMode {
-    pub fn require_enabled(self) -> Result<(), Error> {
-        match self {
-            Self::Enabled => Ok(()),
-            Self::DisabledForEmbeddings => {
-                Err(ErrorNotImplemented(TOKEN_GENERATION_DISABLED_MESSAGE))
-            }
-        }
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use actix_web::http::StatusCode;
-
-    use super::ClusterTokenGenerationMode;
-
-    #[test]
-    fn allows_requests_while_token_generation_is_enabled() {
-        assert!(
-            ClusterTokenGenerationMode::Enabled
-                .require_enabled()
-                .is_ok()
-        );
-    }
-
-    #[test]
-    fn rejects_requests_as_not_implemented_while_disabled_for_embeddings() {
-        let rejection = ClusterTokenGenerationMode::DisabledForEmbeddings
-            .require_enabled()
-            .expect_err("token generation must be rejected in embeddings mode");
-
-        assert_eq!(
-            rejection.as_response_error().status_code(),
-            StatusCode::NOT_IMPLEMENTED
-        );
-    }
-}
diff --git a/paddler_balancer/src/compatibility/attach_request_id.rs b/paddler_balancer/src/compatibility/attach_request_id.rs
new file mode 100644
index 000000000..095db279c
--- /dev/null
+++ b/paddler_balancer/src/compatibility/attach_request_id.rs
@@ -0,0 +1,30 @@
+use actix_web::Error;
+use actix_web::body::MessageBody;
+use actix_web::dev::ServiceRequest;
+use actix_web::dev::ServiceResponse;
+use actix_web::http::header::HeaderName;
+use actix_web::http::header::HeaderValue;
+use actix_web::middleware::Next;
+use rand::random;
+
+use crate::compatibility::serves_compatibility_layer::ServesCompatibilityLayer;
+
+pub async fn attach_request_id<TCompatibilityLayer: ServesCompatibilityLayer>(
+    request: ServiceRequest,
+    next: Next<impl MessageBody>,
+) -> Result<ServiceResponse<impl MessageBody>, Error> {
+    let request_id = request
+        .headers()
+        .get(TCompatibilityLayer::REQUEST_ID_HEADER)
+        .cloned()
+        .unwrap_or_else(|| HeaderValue::from(random::<u64>()));
+
+    next.call(request).await.map(|mut response| {
+        response.headers_mut().insert(
+            HeaderName::from_static(TCompatibilityLayer::REQUEST_ID_HEADER),
+            request_id,
+        );
+
+        response
+    })
+}
diff --git a/paddler_balancer/src/compatibility/compatibility_app_data.rs b/paddler_balancer/src/compatibility/compatibility_app_data.rs
new file mode 100644
index 000000000..3cc69314f
--- /dev/null
+++ b/paddler_balancer/src/compatibility/compatibility_app_data.rs
@@ -0,0 +1,34 @@
+use std::sync::Arc;
+
+use tokio_util::sync::CancellationToken;
+
+use crate::agent_controller_pool::AgentControllerPool;
+use crate::agent_relay::AgentRelay;
+use crate::agent_streaming_request::AgentStreamingRequest;
+use crate::awaiting_agent_relay::AwaitingAgentRelay;
+use crate::balancer_applicable_state_holder::BalancerApplicableStateHolder;
+use crate::buffered_request_manager::BufferedRequestManager;
+use crate::dispatchable_request::DispatchableRequest;
+use crate::inference_service::configuration::Configuration;
+
+pub struct CompatibilityAppData {
+    pub agent_controller_pool: Arc<AgentControllerPool>,
+    pub balancer_applicable_state_holder: Arc<BalancerApplicableStateHolder>,
+    pub buffered_request_manager: Arc<BufferedRequestManager>,
+    pub inference_service_configuration: Configuration,
+    pub shutdown: CancellationToken,
+}
+
+impl CompatibilityAppData {
+    pub fn agent_relay<TParams>(&self, request_params: TParams) -> AgentRelay<TParams>
+    where
+        TParams: AgentStreamingRequest,
+    {
+        AgentRelay::awaiting(AwaitingAgentRelay {
+            buffered_request_manager: self.buffered_request_manager.clone(),
+            dispatchable_request: DispatchableRequest::from(request_params),
+            inference_item_timeout: self.inference_service_configuration.inference_item_timeout,
+            shutdown: self.shutdown.clone(),
+        })
+    }
+}
diff --git a/paddler_balancer/src/compatibility/compatibility_service.rs b/paddler_balancer/src/compatibility/compatibility_service.rs
new file mode 100644
index 000000000..d02faab34
--- /dev/null
+++ b/paddler_balancer/src/compatibility/compatibility_service.rs
@@ -0,0 +1,82 @@
+use std::sync::Arc;
+
+use actix_web::App;
+use actix_web::middleware::from_fn;
+use actix_web::web::Data;
+use anyhow::Result;
+use async_trait::async_trait;
+use tokio_util::sync::CancellationToken;
+use trzcina::Service;
+
+use crate::agent_controller_pool::AgentControllerPool;
+use crate::balancer_applicable_state_holder::BalancerApplicableStateHolder;
+use crate::buffered_request_manager::BufferedRequestManager;
+use crate::compatibility::attach_request_id::attach_request_id;
+use crate::compatibility::compatibility_app_data::CompatibilityAppData;
+use crate::compatibility::create_compatibility_cors_middleware::create_compatibility_cors_middleware;
+use crate::compatibility::serves_compatibility_layer::ServesCompatibilityLayer;
+use crate::http_listener::HttpListener;
+use crate::http_route::get_health::get_health;
+use crate::inference_service::configuration::Configuration as InferenceServiceConfiguration;
+use crate::run_http_service::run_http_service;
+use crate::run_http_service_parameters::RunHttpServiceParameters;
+
+pub struct CompatibilityService<TCompatibilityLayer> {
+    pub agent_controller_pool: Arc<AgentControllerPool>,
+    pub balancer_applicable_state_holder: Arc<BalancerApplicableStateHolder>,
+    pub buffered_request_manager: Arc<BufferedRequestManager>,
+    pub compatibility_layer: TCompatibilityLayer,
+    pub http_listener: HttpListener,
+    pub inference_service_configuration: InferenceServiceConfiguration,
+}
+
+impl<TCompatibilityLayer: ServesCompatibilityLayer> CompatibilityService<TCompatibilityLayer> {
+    async fn serve_until_shutdown(self: Box<Self>, shutdown: CancellationToken) -> Result<()> {
+        let service_name = self.name();
+        let cors_allowed_hosts_arc = Arc::new(
+            self.inference_service_configuration
+                .cors_allowed_hosts
+                .clone(),
+        );
+        let app_data = Data::new(CompatibilityAppData {
+            agent_controller_pool: self.agent_controller_pool.clone(),
+            balancer_applicable_state_holder: self.balancer_applicable_state_holder.clone(),
+            buffered_request_manager: self.buffered_request_manager.clone(),
+            inference_service_configuration: self.inference_service_configuration.clone(),
+            shutdown: shutdown.clone(),
+        });
+
+        run_http_service(
+            shutdown,
+            RunHttpServiceParameters {
+                app_factory: move || {
+                    App::new()
+                        .wrap(from_fn(attach_request_id::<TCompatibilityLayer>))
+                        .wrap(create_compatibility_cors_middleware::<TCompatibilityLayer>(
+                            &cors_allowed_hosts_arc,
+                        ))
+                        .app_data(app_data.clone())
+                        .configure(get_health)
+                        .configure(TCompatibilityLayer::configure)
+                },
+                http_listener: self.http_listener,
+                service_name,
+                worker_count: 16,
+            },
+        )
+        .await
+    }
+}
+
+#[async_trait]
+impl<TCompatibilityLayer: ServesCompatibilityLayer> Service
+    for CompatibilityService<TCompatibilityLayer>
+{
+    fn name(&self) -> &'static str {
+        TCompatibilityLayer::SERVICE_NAME
+    }
+
+    async fn run(self: Box<Self>, shutdown: CancellationToken) -> Result<()> {
+        self.serve_until_shutdown(shutdown).await
+    }
+}
diff --git a/paddler_balancer/src/compatibility/compatibility_service_configuration.rs b/paddler_balancer/src/compatibility/compatibility_service_configuration.rs
new file mode 100644
index 000000000..7394bb2f4
--- /dev/null
+++ b/paddler_balancer/src/compatibility/compatibility_service_configuration.rs
@@ -0,0 +1,6 @@
+use crate::resolved_socket_addr::ResolvedSocketAddr;
+
+#[derive(Clone)]
+pub struct CompatibilityServiceConfiguration {
+    pub addr: ResolvedSocketAddr,
+}
diff --git a/paddler_balancer/src/compatibility/create_compatibility_cors_middleware.rs b/paddler_balancer/src/compatibility/create_compatibility_cors_middleware.rs
new file mode 100644
index 000000000..f29f04040
--- /dev/null
+++ b/paddler_balancer/src/compatibility/create_compatibility_cors_middleware.rs
@@ -0,0 +1,19 @@
+use std::sync::Arc;
+
+use actix_cors::Cors;
+
+use crate::compatibility::serves_compatibility_layer::ServesCompatibilityLayer;
+use crate::create_cors_middleware::create_cors_middleware;
+
+pub fn create_compatibility_cors_middleware<TCompatibilityLayer: ServesCompatibilityLayer>(
+    allowed_hosts: &Arc<Vec<String>>,
+) -> Cors {
+    create_cors_middleware(allowed_hosts)
+        .allowed_headers(
+            TCompatibilityLayer::EXTRA_ALLOWED_HEADERS
+                .iter()
+                .copied()
+                .chain([TCompatibilityLayer::REQUEST_ID_HEADER]),
+        )
+        .expose_headers([TCompatibilityLayer::REQUEST_ID_HEADER])
+}
diff --git a/paddler_balancer/src/compatibility/mod.rs b/paddler_balancer/src/compatibility/mod.rs
index 7cc07e9ba..b7fadde6d 100644
--- a/paddler_balancer/src/compatibility/mod.rs
+++ b/paddler_balancer/src/compatibility/mod.rs
@@ -1 +1,9 @@
+pub mod attach_request_id;
+pub mod compatibility_app_data;
+pub mod compatibility_service;
+pub mod compatibility_service_configuration;
+pub mod create_compatibility_cors_middleware;
 pub mod openai_service;
+pub mod serves_compatibility_layer;
+pub mod typesafe_service;
+pub mod upstream_failure;
diff --git a/paddler_balancer/src/compatibility/openai_service/app_data.rs b/paddler_balancer/src/compatibility/openai_service/app_data.rs
deleted file mode 100644
index f235efbd0..000000000
--- a/paddler_balancer/src/compatibility/openai_service/app_data.rs
+++ /dev/null
@@ -1,14 +0,0 @@
-use std::sync::Arc;
-
-use tokio_util::sync::CancellationToken;
-
-use crate::balancer_applicable_state_holder::BalancerApplicableStateHolder;
-use crate::buffered_request_manager::BufferedRequestManager;
-use crate::inference_service::configuration::Configuration;
-
-pub struct AppData {
-    pub balancer_applicable_state_holder: Arc<BalancerApplicableStateHolder>,
-    pub buffered_request_manager: Arc<BufferedRequestManager>,
-    pub inference_service_configuration: Configuration,
-    pub shutdown: CancellationToken,
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/chat_completion.rs b/paddler_balancer/src/compatibility/openai_service/chat_completion.rs
deleted file mode 100644
index e1c789646..000000000
--- a/paddler_balancer/src/compatibility/openai_service/chat_completion.rs
+++ /dev/null
@@ -1,111 +0,0 @@
-use llama_cpp_bindings_types::ParsedToolCall;
-use serde::Serialize;
-use serde::Serializer;
-
-use crate::compatibility::openai_service::assistant_role::ASSISTANT_ROLE;
-use crate::compatibility::openai_service::chat_completion_tool_call::ChatCompletionToolCall;
-use crate::compatibility::openai_service::openai_usage::OpenAIUsage;
-
-struct SerializedToolCalls<'calls>(&'calls [ParsedToolCall]);
-
-impl Serialize for SerializedToolCalls<'_> {
-    fn serialize<TSerializer>(
-        &self,
-        serializer: TSerializer,
-    ) -> Result<TSerializer::Ok, TSerializer::Error>
-    where
-        TSerializer: Serializer,
-    {
-        serializer.collect_seq(self.0.iter().map(ChatCompletionToolCall))
-    }
-}
-
-#[derive(Serialize)]
-#[serde(untagged)]
-enum SerializedMessage<'completion> {
-    Content {
-        role: &'static str,
-        content: &'completion str,
-        refusal: (),
-        annotations: [u8; 0],
-    },
-    ToolCalls {
-        role: &'static str,
-        content: Option<&'completion str>,
-        refusal: (),
-        annotations: [u8; 0],
-        tool_calls: SerializedToolCalls<'completion>,
-    },
-}
-
-#[derive(Serialize)]
-struct SerializedChoice<'completion> {
-    index: u8,
-    message: SerializedMessage<'completion>,
-    logprobs: (),
-    finish_reason: &'static str,
-}
-
-#[derive(Serialize)]
-struct SerializedCompletion<'completion> {
-    id: &'completion str,
-    object: &'static str,
-    created: u64,
-    model: &'completion str,
-    choices: [SerializedChoice<'completion>; 1],
-    usage: OpenAIUsage,
-    service_tier: &'static str,
-}
-
-pub struct ChatCompletion<'completion> {
-    pub content: &'completion str,
-    pub created: u64,
-    pub finish_reason: &'static str,
-    pub id: &'completion str,
-    pub model: &'completion str,
-    pub tool_calls: &'completion [ParsedToolCall],
-    pub usage: OpenAIUsage,
-}
-
-impl Serialize for ChatCompletion<'_> {
-    fn serialize<TSerializer>(
-        &self,
-        serializer: TSerializer,
-    ) -> Result<TSerializer::Ok, TSerializer::Error>
-    where
-        TSerializer: Serializer,
-    {
-        let message = if self.tool_calls.is_empty() {
-            SerializedMessage::Content {
-                role: ASSISTANT_ROLE,
-                content: self.content,
-                refusal: (),
-                annotations: [],
-            }
-        } else {
-            SerializedMessage::ToolCalls {
-                role: ASSISTANT_ROLE,
-                content: (!self.content.is_empty()).then_some(self.content),
-                refusal: (),
-                annotations: [],
-                tool_calls: SerializedToolCalls(self.tool_calls),
-            }
-        };
-
-        SerializedCompletion {
-            id: self.id,
-            object: "chat.completion",
-            created: self.created,
-            model: self.model,
-            choices: [SerializedChoice {
-                index: 0,
-                message,
-                logprobs: (),
-                finish_reason: self.finish_reason,
-            }],
-            usage: self.usage,
-            service_tier: "default",
-        }
-        .serialize(serializer)
-    }
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/chat_completion_chunk.rs b/paddler_balancer/src/compatibility/openai_service/chat_completion_chunk.rs
deleted file mode 100644
index 42b1eaabd..000000000
--- a/paddler_balancer/src/compatibility/openai_service/chat_completion_chunk.rs
+++ /dev/null
@@ -1,71 +0,0 @@
-use serde::Serialize;
-use serde::Serializer;
-
-use crate::compatibility::openai_service::chat_completion_chunk_choice::ChatCompletionChunkChoice;
-use crate::compatibility::openai_service::chat_completion_chunk_payload::ChatCompletionChunkPayload;
-use crate::compatibility::openai_service::openai_usage::OpenAIUsage;
-
-#[derive(Serialize)]
-struct ChunkEnvelope<'chunk, TBody> {
-    id: &'chunk str,
-    object: &'static str,
-    created: u64,
-    model: &'chunk str,
-    system_fingerprint: &'chunk str,
-    #[serde(flatten)]
-    body: TBody,
-}
-
-#[derive(Serialize)]
-struct ChoiceBody<'body, 'choice> {
-    choices: [&'body ChatCompletionChunkChoice<'choice>; 1],
-}
-
-#[derive(Serialize)]
-struct UsageBody {
-    choices: [u8; 0],
-    usage: OpenAIUsage,
-}
-
-pub struct ChatCompletionChunk<'chunk> {
-    pub created: u64,
-    pub id: &'chunk str,
-    pub model: &'chunk str,
-    pub payload: ChatCompletionChunkPayload<'chunk>,
-    pub system_fingerprint: &'chunk str,
-}
-
-impl<'chunk> ChatCompletionChunk<'chunk> {
-    const fn envelope<TBody>(&self, body: TBody) -> ChunkEnvelope<'chunk, TBody> {
-        ChunkEnvelope {
-            id: self.id,
-            object: "chat.completion.chunk",
-            created: self.created,
-            model: self.model,
-            system_fingerprint: self.system_fingerprint,
-            body,
-        }
-    }
-}
-
-impl Serialize for ChatCompletionChunk<'_> {
-    fn serialize<TSerializer>(
-        &self,
-        serializer: TSerializer,
-    ) -> Result<TSerializer::Ok, TSerializer::Error>
-    where
-        TSerializer: Serializer,
-    {
-        match &self.payload {
-            ChatCompletionChunkPayload::Choice(choice) => self
-                .envelope(ChoiceBody { choices: [choice] })
-                .serialize(serializer),
-            ChatCompletionChunkPayload::Usage(usage) => self
-                .envelope(UsageBody {
-                    choices: [],
-                    usage: OpenAIUsage(**usage),
-                })
-                .serialize(serializer),
-        }
-    }
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/chat_completion_chunk_choice.rs b/paddler_balancer/src/compatibility/openai_service/chat_completion_chunk_choice.rs
deleted file mode 100644
index 8f53c6879..000000000
--- a/paddler_balancer/src/compatibility/openai_service/chat_completion_chunk_choice.rs
+++ /dev/null
@@ -1,101 +0,0 @@
-use llama_cpp_bindings_types::ParsedToolCall;
-use serde::Serialize;
-use serde::Serializer;
-
-use crate::compatibility::openai_service::assistant_role::ASSISTANT_ROLE;
-use crate::compatibility::openai_service::chat_completion_tool_call::ChatCompletionToolCall;
-
-#[derive(Serialize)]
-struct ToolCallDelta<'call> {
-    index: usize,
-    #[serde(flatten)]
-    tool_call: ChatCompletionToolCall<'call>,
-}
-
-struct ToolCallsDelta<'calls>(&'calls [ParsedToolCall]);
-
-impl Serialize for ToolCallsDelta<'_> {
-    fn serialize<TSerializer>(
-        &self,
-        serializer: TSerializer,
-    ) -> Result<TSerializer::Ok, TSerializer::Error>
-    where
-        TSerializer: Serializer,
-    {
-        serializer.collect_seq(
-            self.0
-                .iter()
-                .enumerate()
-                .map(|(index, call)| ToolCallDelta {
-                    index,
-                    tool_call: ChatCompletionToolCall(call),
-                }),
-        )
-    }
-}
-
-#[derive(Serialize)]
-#[serde(untagged)]
-enum ChoiceDelta<'choice> {
-    Content {
-        role: &'static str,
-        content: &'choice str,
-    },
-    Finish {},
-    ToolCalls {
-        role: &'static str,
-        tool_calls: ToolCallsDelta<'choice>,
-    },
-}
-
-#[derive(Serialize)]
-struct SerializedChoice<'choice> {
-    index: u8,
-    delta: ChoiceDelta<'choice>,
-    logprobs: (),
-    finish_reason: Option<&'static str>,
-}
-
-pub enum ChatCompletionChunkChoice<'chunk> {
-    Content(&'chunk str),
-    Finish(&'static str),
-    ToolCalls(&'chunk [ParsedToolCall]),
-}
-
-impl Serialize for ChatCompletionChunkChoice<'_> {
-    fn serialize<TSerializer>(
-        &self,
-        serializer: TSerializer,
-    ) -> Result<TSerializer::Ok, TSerializer::Error>
-    where
-        TSerializer: Serializer,
-    {
-        match *self {
-            Self::Content(content) => SerializedChoice {
-                index: 0,
-                delta: ChoiceDelta::Content {
-                    role: ASSISTANT_ROLE,
-                    content,
-                },
-                logprobs: (),
-                finish_reason: None,
-            },
-            Self::Finish(finish_reason) => SerializedChoice {
-                index: 0,
-                delta: ChoiceDelta::Finish {},
-                logprobs: (),
-                finish_reason: Some(finish_reason),
-            },
-            Self::ToolCalls(parsed_calls) => SerializedChoice {
-                index: 0,
-                delta: ChoiceDelta::ToolCalls {
-                    role: ASSISTANT_ROLE,
-                    tool_calls: ToolCallsDelta(parsed_calls),
-                },
-                logprobs: (),
-                finish_reason: None,
-            },
-        }
-        .serialize(serializer)
-    }
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/chat_completion_chunk_payload.rs b/paddler_balancer/src/compatibility/openai_service/chat_completion_chunk_payload.rs
deleted file mode 100644
index ae238ac1f..000000000
--- a/paddler_balancer/src/compatibility/openai_service/chat_completion_chunk_payload.rs
+++ /dev/null
@@ -1,8 +0,0 @@
-use llama_cpp_bindings_types::TokenUsage;
-
-use crate::compatibility::openai_service::chat_completion_chunk_choice::ChatCompletionChunkChoice;
-
-pub enum ChatCompletionChunkPayload<'chunk> {
-    Choice(ChatCompletionChunkChoice<'chunk>),
-    Usage(&'chunk TokenUsage),
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/chat_completions_sse_response.rs b/paddler_balancer/src/compatibility/openai_service/chat_completions_sse_response.rs
index 6df2b8f99..6a6f2d694 100644
--- a/paddler_balancer/src/compatibility/openai_service/chat_completions_sse_response.rs
+++ b/paddler_balancer/src/compatibility/openai_service/chat_completions_sse_response.rs
@@ -1,65 +1,78 @@
-use std::convert::Infallible;
-
 use actix_web::HttpResponse;
-use actix_web::http::header;
 use actix_web_lab::sse;
+use futures::future::ready;
 use futures::stream::Stream;
 use futures::stream::StreamExt as _;
+use futures::stream::iter;
 use futures::stream::once;
+use nanoid::nanoid;
+use serde_json::Error as SerdeJsonError;
 
-use crate::chunk_forwarding_session_controller::transform_result::TransformResult;
+use paddler_messaging::generated_token_final_result::GeneratedTokenFinalResult;
+use paddler_openai_translation::chat_completion_stream::ChatCompletionStream;
+use paddler_openai_translation::generation_outcome::GenerationOutcome;
+use paddler_openai_translation::generation_progress::GenerationProgress;
 
-pub fn chat_completions_sse_response<TStream>(transform_results: TStream) -> HttpResponse
-where
-    TStream: Stream<Item = TransformResult> + 'static,
-{
-    let event_stream = transform_results
-        .filter_map(|transform_result| async move {
-            match transform_result {
-                TransformResult::Chunk(chunk) | TransformResult::Error(chunk) => {
-                    Some(Ok::<sse::Event, Infallible>(sse::Event::Data(
-                        sse::Data::new(chunk),
-                    )))
-                }
-                TransformResult::Discard => None,
-            }
-        })
-        .chain(once(async {
-            Ok::<sse::Event, Infallible>(sse::Event::Data(sse::Data::new("[DONE]")))
-        }));
+use crate::agent_relay_event::AgentRelayEvent;
+use crate::compatibility::openai_service::openai_failure::OpenAIFailure;
+use crate::compatibility::openai_service::openai_sse_http_response::openai_sse_http_response;
 
-    HttpResponse::Ok()
-        .content_type("text/event-stream")
-        .insert_header((header::CACHE_CONTROL, "no-cache"))
-        .body(sse::Sse::from_stream(event_stream))
+fn openai_failure_data(openai_failure: OpenAIFailure) -> Vec<Result<sse::Data, SerdeJsonError>> {
+    vec![sse::Data::new_json(openai_failure.into_error_body())]
 }
 
-#[cfg(test)]
-mod tests {
-    use actix_web::body;
-    use futures::stream::iter;
-
-    use super::chat_completions_sse_response;
-    use crate::chunk_forwarding_session_controller::transform_result::TransformResult;
-
-    #[actix_web::test]
-    async fn frames_each_chunk_as_an_sse_data_event_and_terminates_with_done() {
-        let response = chat_completions_sse_response(iter([
-            TransformResult::Chunk("{\"id\":1}".to_owned()),
-            TransformResult::Discard,
-            TransformResult::Error("{\"error\":{}}".to_owned()),
-        ]));
+fn chat_completion_sse_data(
+    chat_completion_stream: &mut ChatCompletionStream,
+    request_id: &str,
+    agent_relay_event: AgentRelayEvent<GenerationProgress, GeneratedTokenFinalResult>,
+) -> Vec<Result<sse::Data, SerdeJsonError>> {
+    let chat_completion_chunks = match agent_relay_event {
+        AgentRelayEvent::Failed(relay_failure) => {
+            return openai_failure_data(OpenAIFailure::Relay(relay_failure));
+        }
+        AgentRelayEvent::Final(final_result) => {
+            chat_completion_stream.conclude(request_id, GenerationOutcome::from(final_result))
+        }
+        AgentRelayEvent::Partial(generation_progress) => {
+            chat_completion_stream.advance(request_id, generation_progress)
+        }
+    };
 
-        assert_eq!(
-            response.headers().get("content-type").unwrap(),
-            "text/event-stream"
-        );
+    match chat_completion_chunks {
+        Ok(chat_completion_chunks) => chat_completion_chunks
+            .into_iter()
+            .map(sse::Data::new_json)
+            .collect(),
+        Err(generation_failure) => {
+            openai_failure_data(OpenAIFailure::Generation(generation_failure))
+        }
+    }
+}
 
-        let body_bytes = body::to_bytes(response.into_body()).await.unwrap();
+pub fn chat_completions_sse_response<TAgentRelayEvents>(
+    chat_completion_stream: ChatCompletionStream,
+    agent_relay_events: TAgentRelayEvents,
+) -> HttpResponse
+where
+    TAgentRelayEvents:
+        Stream<Item = AgentRelayEvent<GenerationProgress, GeneratedTokenFinalResult>> + 'static,
+{
+    let request_id: String = nanoid!();
 
-        assert_eq!(
-            body_bytes.as_ref(),
-            b"data: {\"id\":1}\n\ndata: {\"error\":{}}\n\ndata: [DONE]\n\n"
-        );
-    }
+    openai_sse_http_response(
+        agent_relay_events
+            .scan(
+                chat_completion_stream,
+                move |chat_completion_stream, agent_relay_event| {
+                    ready(Some(chat_completion_sse_data(
+                        chat_completion_stream,
+                        &request_id,
+                        agent_relay_event,
+                    )))
+                },
+            )
+            .flat_map(iter)
+            .chain(once(ready(Ok(sse::Data::new("[DONE]")))))
+            .map(|sse_data| sse_data.map(sse::Event::Data)),
+    )
 }
diff --git a/paddler_balancer/src/compatibility/openai_service/configuration.rs b/paddler_balancer/src/compatibility/openai_service/configuration.rs
deleted file mode 100644
index f92b8925c..000000000
--- a/paddler_balancer/src/compatibility/openai_service/configuration.rs
+++ /dev/null
@@ -1,6 +0,0 @@
-use crate::resolved_socket_addr::ResolvedSocketAddr;
-
-#[derive(Clone)]
-pub struct Configuration {
-    pub addr: ResolvedSocketAddr,
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/content_part_event.rs b/paddler_balancer/src/compatibility/openai_service/content_part_event.rs
deleted file mode 100644
index ceb71a6c9..000000000
--- a/paddler_balancer/src/compatibility/openai_service/content_part_event.rs
+++ /dev/null
@@ -1,12 +0,0 @@
-use serde::Serialize;
-
-use crate::compatibility::openai_service::responses_content_part::ResponsesContentPart;
-
-#[derive(Clone, Debug, Serialize)]
-pub struct ContentPartEvent {
-    pub sequence_number: u64,
-    pub item_id: String,
-    pub output_index: usize,
-    pub content_index: usize,
-    pub part: ResponsesContentPart,
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/generation_progress_events.rs b/paddler_balancer/src/compatibility/openai_service/generation_progress_events.rs
new file mode 100644
index 000000000..1df8ebde7
--- /dev/null
+++ b/paddler_balancer/src/compatibility/openai_service/generation_progress_events.rs
@@ -0,0 +1,106 @@
+use futures::stream::Stream;
+use futures::stream::unfold;
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::generated_token_final_result::GeneratedTokenFinalResult;
+use paddler_messaging::generated_token_result::GeneratedTokenResult;
+use paddler_openai_translation::generation_progress::GenerationProgress;
+
+use crate::agent_relay::AgentRelay;
+use crate::agent_relay_conclusion::AgentRelayConclusion;
+use crate::agent_relay_event::AgentRelayEvent;
+use crate::agent_relay_step::AgentRelayStep;
+use crate::agent_streaming_request::AgentStreamingRequest;
+
+enum GenerationProgressPhase<TParams>
+where
+    TParams: AgentStreamingRequest<Response = GeneratedTokenResult>,
+{
+    Relaying {
+        agent_relay: AgentRelay<TParams>,
+        settlement: CancellationToken,
+    },
+    ReleasingRelay {
+        relay: AgentRelay<TParams>,
+        settlement: CancellationToken,
+    },
+    Settling {
+        settlement: CancellationToken,
+    },
+}
+
+impl<TParams> GenerationProgressPhase<TParams>
+where
+    TParams: AgentStreamingRequest<Response = GeneratedTokenResult>,
+{
+    async fn advance(
+        self,
+    ) -> Option<(
+        AgentRelayEvent<GenerationProgress, GeneratedTokenFinalResult>,
+        Self,
+    )> {
+        match self {
+            Self::Relaying {
+                agent_relay,
+                settlement,
+            } => match agent_relay.next().await {
+                AgentRelayStep::Concluded(AgentRelayConclusion::AgentConcluded {
+                    final_result,
+                    ..
+                }) => Some((
+                    AgentRelayEvent::Final(final_result),
+                    Self::Settling { settlement },
+                )),
+                AgentRelayStep::Concluded(AgentRelayConclusion::Failed(relay_failure)) => Some((
+                    AgentRelayEvent::Failed(relay_failure),
+                    Self::Settling { settlement },
+                )),
+                AgentRelayStep::Partial {
+                    partial_result,
+                    relay,
+                    ..
+                } => match GenerationProgress::from(partial_result) {
+                    failed_generation @ GenerationProgress::Failed(_) => Some((
+                        AgentRelayEvent::Partial(failed_generation),
+                        Self::ReleasingRelay { relay, settlement },
+                    )),
+                    generation_progress => Some((
+                        AgentRelayEvent::Partial(generation_progress),
+                        Self::Relaying {
+                            agent_relay: relay,
+                            settlement,
+                        },
+                    )),
+                },
+            },
+            Self::ReleasingRelay { relay, settlement } => {
+                drop(relay);
+                settlement.cancelled().await;
+
+                None
+            }
+            Self::Settling { settlement } => {
+                settlement.cancelled().await;
+
+                None
+            }
+        }
+    }
+}
+
+pub fn generation_progress_events<TParams>(
+    agent_relay: AgentRelay<TParams>,
+) -> impl Stream<Item = AgentRelayEvent<GenerationProgress, GeneratedTokenFinalResult>>
+where
+    TParams: AgentStreamingRequest<Response = GeneratedTokenResult>,
+{
+    let settlement = agent_relay.settlement();
+
+    unfold(
+        GenerationProgressPhase::Relaying {
+            agent_relay,
+            settlement,
+        },
+        GenerationProgressPhase::advance,
+    )
+}
diff --git a/paddler_balancer/src/compatibility/openai_service/http_route/post_chat_completions.rs b/paddler_balancer/src/compatibility/openai_service/http_route/post_chat_completions.rs
index e8aa718ac..048ed5aec 100644
--- a/paddler_balancer/src/compatibility/openai_service/http_route/post_chat_completions.rs
+++ b/paddler_balancer/src/compatibility/openai_service/http_route/post_chat_completions.rs
@@ -1,135 +1,50 @@
-use std::sync::Arc;
 use std::time::SystemTime;
 
-use actix_web::Error;
 use actix_web::HttpResponse;
-use actix_web::error::ErrorInternalServerError;
-use actix_web::http::StatusCode;
 use actix_web::web;
 use actix_web::web::post;
-use nanoid::nanoid;
-use parking_lot::Mutex;
 
-use paddler_messaging::conversation_history::ConversationHistory;
-use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
-use paddler_messaging::validates::Validates;
+use paddler_openai_translation::chat_completion_delivery::ChatCompletionDelivery;
+use paddler_openai_translation::chat_completion_request::ChatCompletionRequest;
+use paddler_openai_translation::translated_chat_completion_request::TranslatedChatCompletionRequest;
 
-use crate::cluster_token_generation_mode::ClusterTokenGenerationMode;
-use crate::compatibility::openai_service::app_data::AppData;
+use crate::compatibility::compatibility_app_data::CompatibilityAppData;
 use crate::compatibility::openai_service::chat_completions_sse_response::chat_completions_sse_response;
+use crate::compatibility::openai_service::generation_progress_events::generation_progress_events;
 use crate::compatibility::openai_service::openai_api_path::OpenAIApiPath;
-use crate::compatibility::openai_service::openai_chat_completion_tool::OpenAIChatCompletionTool;
-use crate::compatibility::openai_service::openai_completion_request_params::OpenAICompletionRequestParams;
-use crate::compatibility::openai_service::openai_default_max_tokens::OPENAI_DEFAULT_MAX_TOKENS;
-use crate::compatibility::openai_service::openai_error::OpenAIError;
-use crate::compatibility::openai_service::openai_error_type::OpenAIErrorType;
+use crate::compatibility::openai_service::openai_failure::OpenAIFailure;
+use crate::compatibility::openai_service::openai_http_response::openai_http_response;
 use crate::compatibility::openai_service::openai_json_config::openai_json_config;
-use crate::compatibility::openai_service::openai_json_response::openai_json_response;
-use crate::compatibility::openai_service::openai_message::OpenAIMessage;
-use crate::compatibility::openai_service::openai_non_streaming_response_transformer::OpenAINonStreamingResponseTransformer;
-use crate::compatibility::openai_service::openai_non_streaming_state::OpenAINonStreamingState;
-use crate::compatibility::openai_service::openai_streaming_response_transformer::OpenAIStreamingResponseTransformer;
-use crate::compatibility::openai_service::openai_streaming_state::OpenAIStreamingState;
-use crate::compatibility::openai_service::timestamp_from::timestamp_from;
-use crate::unbounded_stream_from_agent::unbounded_stream_from_agent;
-use crate::unbounded_stream_from_agent_params::UnboundedStreamFromAgentParams;
 
 async fn respond(
-    app_data: web::Data<AppData>,
-    openai_params: web::Json<OpenAICompletionRequestParams>,
-) -> Result<HttpResponse, Error> {
-    if app_data
-        .balancer_applicable_state_holder
-        .token_generation_mode()
-        == ClusterTokenGenerationMode::DisabledForEmbeddings
+    app_data: web::Data<CompatibilityAppData>,
+    chat_completion_request: web::Json<ChatCompletionRequest>,
+) -> HttpResponse {
+    match chat_completion_request
+        .into_inner()
+        .translate(SystemTime::now())
     {
-        return Ok(OpenAIError {
-            error_type: OpenAIErrorType::ServerError,
-            message: "Chat completions are disabled while the cluster is configured for embeddings"
-                .to_owned(),
-        }
-        .to_http_response(StatusCode::NOT_IMPLEMENTED));
-    }
-
-    let openai_params = openai_params.into_inner();
-    let enable_thinking = openai_params.enables_thinking();
-    let max_tokens = openai_params
-        .requested_max_tokens()
-        .unwrap_or(OPENAI_DEFAULT_MAX_TOKENS);
-
-    let validated_tools = match openai_params
-        .tools
-        .into_iter()
-        .map(OpenAIChatCompletionTool::into_tool)
-        .map(Validates::validate)
-        .collect::<Result<Vec<_>, _>>()
-    {
-        Ok(tools) => tools,
-        Err(err) => {
-            return Ok(OpenAIError {
-                error_type: OpenAIErrorType::InvalidRequestError,
-                message: err.to_string(),
+        Ok(TranslatedChatCompletionRequest {
+            conversation_history_params,
+            delivery,
+        }) => {
+            let agent_relay = app_data.agent_relay(conversation_history_params);
+
+            match delivery {
+                ChatCompletionDelivery::Buffered(chat_completion_header) => {
+                    openai_http_response(&chat_completion_header, agent_relay).await
+                }
+                ChatCompletionDelivery::Streamed(chat_completion_stream) => {
+                    chat_completions_sse_response(
+                        chat_completion_stream,
+                        generation_progress_events(agent_relay),
+                    )
+                }
             }
-            .to_http_response(StatusCode::BAD_REQUEST));
         }
-    };
-
-    let parse_tool_calls = !validated_tools.is_empty();
-    let paddler_params = ContinueFromConversationHistoryParams {
-        add_generation_prompt: true,
-        conversation_history: ConversationHistory::new(
-            openai_params
-                .messages
-                .into_iter()
-                .map(OpenAIMessage::into_conversation_message)
-                .collect(),
-        ),
-        enable_thinking,
-        grammar: None,
-        max_tokens,
-        parse_tool_calls,
-        tools: validated_tools,
-    };
-
-    let created = timestamp_from(SystemTime::now()).map_err(ErrorInternalServerError)?;
-
-    if openai_params.stream.unwrap_or(false) {
-        let include_usage = openai_params
-            .stream_options
-            .as_ref()
-            .is_some_and(|options| options.include_usage);
-
-        Ok(chat_completions_sse_response(unbounded_stream_from_agent(
-            UnboundedStreamFromAgentParams {
-                buffered_request_manager: app_data.buffered_request_manager.clone(),
-                inference_service_configuration: app_data.inference_service_configuration.clone(),
-                request_params: paddler_params,
-                shutdown: app_data.shutdown.clone(),
-                transformer: OpenAIStreamingResponseTransformer {
-                    created,
-                    include_usage,
-                    model: openai_params.model.clone(),
-                    state: Arc::new(Mutex::new(OpenAIStreamingState::default())),
-                    system_fingerprint: nanoid!(),
-                },
-            },
-        )))
-    } else {
-        Ok(openai_json_response(
-            unbounded_stream_from_agent(UnboundedStreamFromAgentParams {
-                buffered_request_manager: app_data.buffered_request_manager.clone(),
-                inference_service_configuration: app_data.inference_service_configuration.clone(),
-                request_params: paddler_params,
-                shutdown: app_data.shutdown.clone(),
-                transformer: OpenAINonStreamingResponseTransformer {
-                    created,
-                    model: openai_params.model.clone(),
-                    state: Arc::new(Mutex::new(OpenAINonStreamingState::default())),
-                },
-            }),
-            "no completion produced",
-        )
-        .await)
+        Err(translation_error) => {
+            OpenAIFailure::RequestUntranslatable(translation_error).into_http_response()
+        }
     }
 }
 
@@ -156,15 +71,13 @@ mod tests {
     use serde_json::Value;
     use serde_json::from_slice;
     use serde_json::json;
+    use tokio::sync::watch;
     use tokio_util::sync::CancellationToken;
 
-    use paddler_inference_parameters::inference_parameters::InferenceParameters;
-    use paddler_messaging::agent_desired_model::AgentDesiredModel;
-    use paddler_messaging::agent_desired_state::AgentDesiredState;
     use paddler_messaging::balancer_desired_state::BalancerDesiredState;
     use paddler_openai_response_format_validator::openai_validator::OpenAIValidator;
 
-    use super::AppData;
+    use super::CompatibilityAppData;
     use super::post_chat_completions;
     use crate::agent_controller_pool::AgentControllerPool;
     use crate::balancer_applicable_state::BalancerApplicableState;
@@ -174,13 +87,19 @@ mod tests {
     use crate::inference_service::configuration::Configuration as InferenceServiceConfiguration;
     use crate::resolved_socket_addr::ResolvedSocketAddr;
 
-    fn app_data_without_agents(max_buffered_requests: u64) -> AppData {
-        AppData {
+    fn app_data_without_agents(max_buffered_requests: u64) -> CompatibilityAppData {
+        let agent_controller_pool = Arc::new(AgentControllerPool::default());
+
+        CompatibilityAppData {
+            agent_controller_pool: agent_controller_pool.clone(),
             balancer_applicable_state_holder: Arc::new(BalancerApplicableStateHolder::new(
-                BalancerApplicableState::from(BalancerDesiredState::default()),
+                watch::channel(Arc::new(BalancerApplicableState::initial(
+                    BalancerDesiredState::default(),
+                )))
+                .1,
             )),
             buffered_request_manager: Arc::new(BufferedRequestManager::new(
-                Arc::new(AgentControllerPool::default()),
+                agent_controller_pool,
                 Duration::ZERO,
                 max_buffered_requests,
             )),
@@ -193,68 +112,6 @@ mod tests {
         }
     }
 
-    fn app_data_with_embeddings_enabled() -> AppData {
-        let balancer_applicable_state_holder = Arc::new(BalancerApplicableStateHolder::new(
-            BalancerApplicableState::from(BalancerDesiredState::default()),
-        ));
-
-        balancer_applicable_state_holder.set_balancer_applicable_state(BalancerApplicableState {
-            agent_desired_state: AgentDesiredState {
-                chat_template_override: None,
-                inference_parameters: InferenceParameters {
-                    enable_embeddings: true,
-                    ..InferenceParameters::default()
-                },
-                model: AgentDesiredModel::LocalToAgent("model.gguf".to_owned()),
-                multimodal_projection: AgentDesiredModel::None,
-            },
-        });
-
-        AppData {
-            balancer_applicable_state_holder,
-            buffered_request_manager: Arc::new(BufferedRequestManager::new(
-                Arc::new(AgentControllerPool::default()),
-                Duration::ZERO,
-                0,
-            )),
-            inference_service_configuration: InferenceServiceConfiguration {
-                addr: ResolvedSocketAddr::from(SocketAddr::from((Ipv4Addr::LOCALHOST, 0))),
-                cors_allowed_hosts: Vec::new(),
-                inference_item_timeout: Duration::ZERO,
-            },
-            shutdown: CancellationToken::new(),
-        }
-    }
-
-    #[actix_web::test]
-    async fn rejects_chat_completion_when_embeddings_are_enabled() -> Result<()> {
-        let app = init_service(
-            App::new()
-                .app_data(Data::new(app_data_with_embeddings_enabled()))
-                .configure(post_chat_completions),
-        )
-        .await;
-
-        let request = TestRequest::post()
-            .uri(OpenAIApiPath::CHAT_COMPLETIONS)
-            .set_json(json!({
-                "model": "test-model",
-                "messages": [{"role": "user", "content": "hi"}]
-            }))
-            .to_request();
-
-        let response = call_service(&app, request).await;
-
-        assert_eq!(response.status(), StatusCode::NOT_IMPLEMENTED);
-
-        let body = read_body(response).await;
-        let envelope: Value = from_slice(&body)?;
-
-        OpenAIValidator::new()?.validate_error_response(&envelope)?;
-
-        Ok(())
-    }
-
     #[actix_web::test]
     async fn invalid_tool_schema_returns_bad_request() {
         let app = init_service(
@@ -338,52 +195,7 @@ mod tests {
     }
 
     #[actix_web::test]
-    async fn opencode_style_tools_are_accepted() {
-        let app = init_service(
-            App::new()
-                .app_data(Data::new(app_data_without_agents(0)))
-                .configure(post_chat_completions),
-        )
-        .await;
-
-        let request = TestRequest::post()
-            .uri(OpenAIApiPath::CHAT_COMPLETIONS)
-            .set_json(json!({
-                "model": "test-model",
-                "messages": [{"role": "user", "content": "hi"}],
-                "tools": [
-                    {
-                        "type": "function",
-                        "function": {
-                            "name": "glob",
-                            "description": "Fast file pattern matching tool",
-                            "parameters": {
-                                "$schema": "https://json-schema.org/draft/2020-12/schema",
-                                "type": "object",
-                                "properties": {
-                                    "pattern": {"type": "string", "description": "The glob pattern"},
-                                    "path": {"type": "string", "description": "The directory to search in"}
-                                },
-                                "required": ["pattern"]
-                            }
-                        }
-                    }
-                ]
-            }))
-            .to_request();
-
-        let response = call_service(&app, request).await;
-
-        assert_eq!(response.status(), StatusCode::INTERNAL_SERVER_ERROR);
-
-        let body = read_body(response).await;
-        let parsed: Value = from_slice(&body).unwrap();
-
-        assert_eq!(parsed["error"]["type"], "server_error");
-    }
-
-    #[actix_web::test]
-    async fn non_streaming_request_without_available_agent_returns_internal_server_error() {
+    async fn non_streaming_request_overflowing_the_buffer_returns_service_unavailable() {
         let app = init_service(
             App::new()
                 .app_data(Data::new(app_data_without_agents(0)))
@@ -401,7 +213,7 @@ mod tests {
 
         let response = call_service(&app, request).await;
 
-        assert_eq!(response.status(), StatusCode::INTERNAL_SERVER_ERROR);
+        assert_eq!(response.status(), StatusCode::SERVICE_UNAVAILABLE);
 
         let body = read_body(response).await;
         let parsed: Value = from_slice(&body).unwrap();
diff --git a/paddler_balancer/src/compatibility/openai_service/http_route/post_responses.rs b/paddler_balancer/src/compatibility/openai_service/http_route/post_responses.rs
index c940a0444..2577c0fda 100644
--- a/paddler_balancer/src/compatibility/openai_service/http_route/post_responses.rs
+++ b/paddler_balancer/src/compatibility/openai_service/http_route/post_responses.rs
@@ -1,108 +1,55 @@
-use std::sync::Arc;
 use std::time::SystemTime;
 
-use actix_web::Error;
 use actix_web::HttpResponse;
-use actix_web::error::ErrorInternalServerError;
-use actix_web::http::StatusCode;
 use actix_web::web;
 use actix_web::web::post;
-use nanoid::nanoid;
-use parking_lot::Mutex;
 
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
+use paddler_openai_translation::responses_delivery::ResponsesDelivery;
+use paddler_openai_translation::responses_request::ResponsesRequest;
+use paddler_openai_translation::translated_responses_request::TranslatedResponsesRequest;
 
-use crate::cluster_token_generation_mode::ClusterTokenGenerationMode;
-use crate::compatibility::openai_service::app_data::AppData;
+use crate::compatibility::compatibility_app_data::CompatibilityAppData;
+use crate::compatibility::openai_service::generation_progress_events::generation_progress_events;
 use crate::compatibility::openai_service::openai_api_path::OpenAIApiPath;
-use crate::compatibility::openai_service::openai_error::OpenAIError;
-use crate::compatibility::openai_service::openai_error_type::OpenAIErrorType;
+use crate::compatibility::openai_service::openai_failure::OpenAIFailure;
+use crate::compatibility::openai_service::openai_http_response::openai_http_response;
 use crate::compatibility::openai_service::openai_json_config::openai_json_config;
-use crate::compatibility::openai_service::openai_json_response::openai_json_response;
-use crate::compatibility::openai_service::openai_responses_request_params::OpenAIResponsesRequestParams;
-use crate::compatibility::openai_service::responses_non_streaming_response_transformer::ResponsesNonStreamingResponseTransformer;
-use crate::compatibility::openai_service::responses_non_streaming_state::ResponsesNonStreamingState;
-use crate::compatibility::openai_service::responses_response_header::ResponsesResponseHeader;
 use crate::compatibility::openai_service::responses_sse_response::responses_sse_response;
-use crate::compatibility::openai_service::responses_streaming_response_transformer::ResponsesStreamingResponseTransformer;
-use crate::compatibility::openai_service::responses_streaming_state::ResponsesStreamingState;
-use crate::compatibility::openai_service::timestamp_from::timestamp_from;
-use crate::unbounded_stream_from_agent::unbounded_stream_from_agent;
-use crate::unbounded_stream_from_agent_params::UnboundedStreamFromAgentParams;
 
 async fn respond(
-    app_data: web::Data<AppData>,
-    openai_params: web::Json<OpenAIResponsesRequestParams>,
-) -> Result<HttpResponse, Error> {
-    if app_data
+    app_data: web::Data<CompatibilityAppData>,
+    responses_request: web::Json<ResponsesRequest>,
+) -> HttpResponse {
+    let balancer_applicable_state = app_data
         .balancer_applicable_state_holder
-        .token_generation_mode()
-        == ClusterTokenGenerationMode::DisabledForEmbeddings
-    {
-        return Ok(OpenAIError {
-            error_type: OpenAIErrorType::ServerError,
-            message: "Responses are disabled while the cluster is configured for embeddings"
-                .to_owned(),
-        }
-        .to_http_response(StatusCode::NOT_IMPLEMENTED));
-    }
-
-    let prepared = match openai_params.into_inner().into_prepared() {
-        Ok(prepared) => prepared,
-        Err(err) => {
-            return Ok(OpenAIError {
-                error_type: OpenAIErrorType::InvalidRequestError,
-                message: err.to_string(),
+        .get_balancer_applicable_state();
+
+    match responses_request.into_inner().translate(
+        SystemTime::now(),
+        &balancer_applicable_state
+            .balancer_desired_state
+            .text_generation
+            .sampling_parameters,
+    ) {
+        Ok(TranslatedResponsesRequest {
+            conversation_history_params,
+            delivery,
+        }) => {
+            let agent_relay = app_data.agent_relay(conversation_history_params);
+
+            match delivery {
+                ResponsesDelivery::Buffered(responses_response_header) => {
+                    openai_http_response(&responses_response_header, agent_relay).await
+                }
+                ResponsesDelivery::Streamed(responses_stream) => responses_sse_response(
+                    responses_stream,
+                    generation_progress_events(agent_relay),
+                ),
             }
-            .to_http_response(StatusCode::BAD_REQUEST));
         }
-    };
-
-    let created_at = timestamp_from(SystemTime::now()).map_err(ErrorInternalServerError)?;
-
-    let InferenceParameters {
-        temperature, top_p, ..
-    } = app_data
-        .balancer_applicable_state_holder
-        .get_agent_desired_state()
-        .inference_parameters;
-    let header = ResponsesResponseHeader {
-        id: format!("resp_{}", nanoid!()),
-        created_at,
-        model: prepared.model,
-        instructions: prepared.instructions,
-        temperature,
-        top_p,
-    };
-
-    if prepared.stream {
-        Ok(responses_sse_response(unbounded_stream_from_agent(
-            UnboundedStreamFromAgentParams {
-                buffered_request_manager: app_data.buffered_request_manager.clone(),
-                inference_service_configuration: app_data.inference_service_configuration.clone(),
-                request_params: prepared.paddler_params,
-                shutdown: app_data.shutdown.clone(),
-                transformer: ResponsesStreamingResponseTransformer {
-                    header,
-                    state: Arc::new(Mutex::new(ResponsesStreamingState::default())),
-                },
-            },
-        )))
-    } else {
-        Ok(openai_json_response(
-            unbounded_stream_from_agent(UnboundedStreamFromAgentParams {
-                buffered_request_manager: app_data.buffered_request_manager.clone(),
-                inference_service_configuration: app_data.inference_service_configuration.clone(),
-                request_params: prepared.paddler_params,
-                shutdown: app_data.shutdown.clone(),
-                transformer: ResponsesNonStreamingResponseTransformer {
-                    header,
-                    state: Arc::new(Mutex::new(ResponsesNonStreamingState::default())),
-                },
-            }),
-            "no response produced",
-        )
-        .await)
+        Err(translation_error) => {
+            OpenAIFailure::RequestUntranslatable(translation_error).into_http_response()
+        }
     }
 }
 
diff --git a/paddler_balancer/src/compatibility/openai_service/mod.rs b/paddler_balancer/src/compatibility/openai_service/mod.rs
index 015ffd114..51822c69b 100644
--- a/paddler_balancer/src/compatibility/openai_service/mod.rs
+++ b/paddler_balancer/src/compatibility/openai_service/mod.rs
@@ -1,141 +1,13 @@
-pub mod app_data;
-pub mod arguments_to_tool_call_string;
-pub mod assistant_role;
-pub mod chat_completion;
-pub mod chat_completion_chunk;
-pub mod chat_completion_chunk_choice;
-pub mod chat_completion_chunk_payload;
-pub mod chat_completion_finish_reason;
-pub mod chat_completion_tool_call;
 pub mod chat_completions_sse_response;
-pub mod configuration;
-pub mod content_part_event;
-pub mod function_call_arguments_delta_event;
-pub mod function_call_arguments_done_event;
+pub mod generation_progress_events;
 pub mod http_route;
-pub mod open_item;
 pub mod openai_api_path;
-pub mod openai_chat_completion_tool;
-pub mod openai_completion_request_params;
-pub mod openai_default_max_tokens;
-pub mod openai_error;
+pub mod openai_compatibility_layer;
+pub mod openai_error_body;
 pub mod openai_error_type;
-pub mod openai_function_definition;
+pub mod openai_failure;
+pub mod openai_header;
+pub mod openai_http_response;
 pub mod openai_json_config;
-pub mod openai_json_response;
-pub mod openai_message;
-pub mod openai_non_streaming_response_transformer;
-pub mod openai_non_streaming_state;
-pub mod openai_reasoning_effort;
-pub mod openai_responses_function_call_item;
-pub mod openai_responses_function_call_output_item;
-pub mod openai_responses_function_output;
-pub mod openai_responses_function_output_part;
-pub mod openai_responses_input;
-pub mod openai_responses_input_content_part;
-pub mod openai_responses_input_item;
-pub mod openai_responses_message_content;
-pub mod openai_responses_message_item;
-pub mod openai_responses_reasoning;
-pub mod openai_responses_request_params;
-pub mod openai_responses_tagged_item;
-pub mod openai_responses_text_format;
-pub mod openai_responses_text_param;
-pub mod openai_responses_tool;
-pub mod openai_streaming_response_transformer;
-pub mod openai_streaming_state;
-pub mod openai_tool_parameters_schema;
-pub mod openai_usage;
-pub mod output_item_event;
-pub mod response_snapshot_event;
-pub mod responses_content_part;
-pub mod responses_item_status;
-pub mod responses_non_streaming_response_transformer;
-pub mod responses_non_streaming_state;
-pub mod responses_output_item;
-pub mod responses_output_item_kind;
-pub mod responses_prepared_request;
-pub mod responses_reasoning_part;
-pub mod responses_response;
-pub mod responses_response_header;
-pub mod responses_response_progress;
-pub mod responses_response_status;
+pub mod openai_sse_http_response;
 pub mod responses_sse_response;
-pub mod responses_stream_event;
-pub mod responses_streaming_response_transformer;
-pub mod responses_streaming_state;
-pub mod responses_usage;
-pub mod stream_options;
-pub mod text_delta_event;
-pub mod text_done_event;
-pub mod timestamp_from;
-pub mod try_universal_error_chunk;
-
-use std::sync::Arc;
-
-use actix_web::App;
-use actix_web::web::Data;
-use anyhow::Result;
-use async_trait::async_trait;
-use tokio_util::sync::CancellationToken;
-use trzcina::Service;
-
-use crate::balancer_applicable_state_holder::BalancerApplicableStateHolder;
-use crate::buffered_request_manager::BufferedRequestManager;
-use crate::compatibility::openai_service::app_data::AppData;
-use crate::compatibility::openai_service::http_route::post_chat_completions::post_chat_completions;
-use crate::compatibility::openai_service::http_route::post_responses::post_responses;
-use crate::create_cors_middleware::create_cors_middleware;
-use crate::http_listener::HttpListener;
-use crate::http_route::get_health::get_health;
-use crate::inference_service::configuration::Configuration as InferenceServiceConfiguration;
-use crate::run_http_service::run_http_service;
-use crate::run_http_service_parameters::RunHttpServiceParameters;
-
-pub struct OpenAIService {
-    pub balancer_applicable_state_holder: Arc<BalancerApplicableStateHolder>,
-    pub buffered_request_manager: Arc<BufferedRequestManager>,
-    pub http_listener: HttpListener,
-    pub inference_service_configuration: InferenceServiceConfiguration,
-}
-
-#[async_trait]
-impl Service for OpenAIService {
-    fn name(&self) -> &'static str {
-        "balancer::compatibility::openai_service"
-    }
-
-    async fn run(self: Box<Self>, shutdown: CancellationToken) -> Result<()> {
-        let service_name = self.name();
-        let cors_allowed_hosts_arc = Arc::new(
-            self.inference_service_configuration
-                .cors_allowed_hosts
-                .clone(),
-        );
-
-        let app_data = Data::new(AppData {
-            balancer_applicable_state_holder: self.balancer_applicable_state_holder.clone(),
-            buffered_request_manager: self.buffered_request_manager.clone(),
-            inference_service_configuration: self.inference_service_configuration.clone(),
-            shutdown: shutdown.clone(),
-        });
-
-        run_http_service(
-            shutdown,
-            RunHttpServiceParameters {
-                app_factory: move || {
-                    App::new()
-                        .wrap(create_cors_middleware(&cors_allowed_hosts_arc))
-                        .app_data(app_data.clone())
-                        .configure(get_health)
-                        .configure(post_chat_completions)
-                        .configure(post_responses)
-                },
-                http_listener: self.http_listener,
-                service_name,
-                worker_count: 16,
-            },
-        )
-        .await
-    }
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/openai_chat_completion_tool.rs b/paddler_balancer/src/compatibility/openai_service/openai_chat_completion_tool.rs
deleted file mode 100644
index 2562a6031..000000000
--- a/paddler_balancer/src/compatibility/openai_service/openai_chat_completion_tool.rs
+++ /dev/null
@@ -1,51 +0,0 @@
-use serde::Deserialize;
-
-use paddler_messaging::request_params::continue_from_conversation_history_params::tool::Tool;
-use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters_schema::raw_parameters_schema::RawParametersSchema;
-
-use crate::compatibility::openai_service::openai_function_definition::OpenAIFunctionDefinition;
-
-#[derive(Deserialize)]
-#[serde(tag = "type")]
-pub enum OpenAIChatCompletionTool {
-    #[serde(rename = "function")]
-    Function {
-        function: Box<OpenAIFunctionDefinition>,
-    },
-}
-
-impl OpenAIChatCompletionTool {
-    #[must_use]
-    pub fn into_tool(self) -> Tool<RawParametersSchema> {
-        let Self::Function { function } = self;
-
-        (*function).into_tool()
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use serde_json::from_value;
-    use serde_json::json;
-
-    use paddler_messaging::request_params::continue_from_conversation_history_params::tool::Tool;
-
-    use super::OpenAIChatCompletionTool;
-
-    #[test]
-    fn function_tool_converts_to_internal_tool() {
-        let tool: OpenAIChatCompletionTool = from_value(json!({
-            "type": "function",
-            "function": {
-                "name": "get_weather",
-                "description": "fetch weather",
-                "parameters": {"type": "object"}
-            }
-        }))
-        .unwrap();
-
-        let Tool::Function(function_call) = tool.into_tool();
-
-        assert_eq!(function_call.function.name, "get_weather");
-    }
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/openai_compatibility_layer.rs b/paddler_balancer/src/compatibility/openai_service/openai_compatibility_layer.rs
new file mode 100644
index 000000000..2bedb973e
--- /dev/null
+++ b/paddler_balancer/src/compatibility/openai_service/openai_compatibility_layer.rs
@@ -0,0 +1,20 @@
+use actix_web::web::ServiceConfig;
+
+use crate::compatibility::openai_service::http_route::post_chat_completions::post_chat_completions;
+use crate::compatibility::openai_service::http_route::post_responses::post_responses;
+use crate::compatibility::openai_service::openai_header::OpenAIHeader;
+use crate::compatibility::serves_compatibility_layer::ServesCompatibilityLayer;
+
+pub struct OpenAICompatibilityLayer;
+
+impl ServesCompatibilityLayer for OpenAICompatibilityLayer {
+    const EXTRA_ALLOWED_HEADERS: &'static [&'static str] = &[];
+    const REQUEST_ID_HEADER: &'static str = OpenAIHeader::REQUEST_ID;
+    const SERVICE_NAME: &'static str = "balancer::compatibility::openai_service";
+
+    fn configure(service_config: &mut ServiceConfig) {
+        service_config
+            .configure(post_chat_completions)
+            .configure(post_responses);
+    }
+}
diff --git a/paddler_balancer/src/compatibility/openai_service/openai_completion_request_params.rs b/paddler_balancer/src/compatibility/openai_service/openai_completion_request_params.rs
deleted file mode 100644
index 669d550e8..000000000
--- a/paddler_balancer/src/compatibility/openai_service/openai_completion_request_params.rs
+++ /dev/null
@@ -1,223 +0,0 @@
-use std::num::NonZeroU32;
-
-use serde::Deserialize;
-
-use crate::compatibility::openai_service::openai_chat_completion_tool::OpenAIChatCompletionTool;
-use crate::compatibility::openai_service::openai_message::OpenAIMessage;
-use crate::compatibility::openai_service::openai_reasoning_effort::OpenAIReasoningEffort;
-use crate::compatibility::openai_service::stream_options::StreamOptions;
-
-#[derive(Deserialize)]
-pub struct OpenAICompletionRequestParams {
-    pub max_completion_tokens: Option<NonZeroU32>,
-    pub max_tokens: Option<NonZeroU32>,
-    pub messages: Vec<OpenAIMessage>,
-    pub model: String,
-    pub reasoning_effort: Option<OpenAIReasoningEffort>,
-    pub stream: Option<bool>,
-    pub stream_options: Option<StreamOptions>,
-    #[serde(default)]
-    pub tools: Vec<OpenAIChatCompletionTool>,
-}
-
-impl OpenAICompletionRequestParams {
-    #[must_use]
-    pub fn enables_thinking(&self) -> bool {
-        self.reasoning_effort
-            .is_none_or(OpenAIReasoningEffort::enables_thinking)
-    }
-
-    #[must_use]
-    pub fn requested_max_tokens(&self) -> Option<NonZeroU32> {
-        self.max_completion_tokens.or(self.max_tokens)
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use std::num::NonZeroU32;
-
-    use serde_json::Value;
-    use serde_json::from_value;
-    use serde_json::json;
-
-    use super::OpenAICompletionRequestParams;
-
-    fn request_with_token_limits(token_limits: &Value) -> OpenAICompletionRequestParams {
-        let mut input = json!({
-            "model": "test-model",
-            "messages": [{"role": "user", "content": "hello"}]
-        });
-
-        input
-            .as_object_mut()
-            .unwrap()
-            .extend(token_limits.as_object().unwrap().clone());
-
-        from_value(input).unwrap()
-    }
-
-    #[test]
-    fn requested_max_tokens_honors_max_tokens() {
-        let params = request_with_token_limits(&json!({"max_tokens": 7}));
-
-        assert_eq!(params.requested_max_tokens(), NonZeroU32::new(7));
-    }
-
-    #[test]
-    fn requested_max_tokens_prefers_max_completion_tokens() {
-        let params =
-            request_with_token_limits(&json!({"max_tokens": 7, "max_completion_tokens": 3}));
-
-        assert_eq!(params.requested_max_tokens(), NonZeroU32::new(3));
-    }
-
-    #[test]
-    fn requested_max_tokens_is_absent_without_limits() {
-        let params = request_with_token_limits(&json!({}));
-
-        assert_eq!(params.requested_max_tokens(), None);
-    }
-
-    #[test]
-    fn deserialize_text_only_request() {
-        let input = json!({
-            "model": "test-model",
-            "messages": [
-                {"role": "user", "content": "hello"}
-            ]
-        });
-
-        let params: OpenAICompletionRequestParams = from_value(input).unwrap();
-
-        assert_eq!(params.model, "test-model");
-        assert_eq!(params.messages.len(), 1);
-        assert_eq!(params.messages[0].role, "user");
-        assert_eq!(params.messages[0].content.text_content(), "hello");
-    }
-
-    #[test]
-    fn deserialize_request_with_stream_options_include_usage_true() {
-        let input = json!({
-            "model": "test-model",
-            "messages": [{"role": "user", "content": "hi"}],
-            "stream": true,
-            "stream_options": {"include_usage": true}
-        });
-
-        let params: OpenAICompletionRequestParams = from_value(input).unwrap();
-
-        let stream_options = params.stream_options.unwrap();
-
-        assert!(stream_options.include_usage);
-    }
-
-    #[test]
-    fn deserialize_request_without_stream_options_defaults_to_none() {
-        let input = json!({
-            "model": "test-model",
-            "messages": [{"role": "user", "content": "hi"}],
-            "stream": true
-        });
-
-        let params: OpenAICompletionRequestParams = from_value(input).unwrap();
-
-        assert!(params.stream_options.is_none());
-    }
-
-    #[test]
-    fn deserialize_multimodal_request_with_image() {
-        let input = json!({
-            "model": "vision-model",
-            "messages": [
-                {
-                    "role": "user",
-                    "content": [
-                        {"type": "text", "text": "describe this image"},
-                        {"type": "image_url", "image_url": {"url": "data:image/jpeg;base64,/9j/4AAQ"}}
-                    ]
-                }
-            ]
-        });
-
-        let params: OpenAICompletionRequestParams = from_value(input).unwrap();
-
-        assert_eq!(params.messages.len(), 1);
-        assert_eq!(
-            params.messages[0].content.text_content(),
-            "describe this image"
-        );
-
-        let image_urls = params.messages[0].content.image_urls();
-
-        assert_eq!(image_urls.len(), 1);
-        assert_eq!(image_urls[0].url, "data:image/jpeg;base64,/9j/4AAQ");
-    }
-
-    #[test]
-    fn deserialize_multi_turn_conversation() {
-        let input = json!({
-            "model": "test-model",
-            "messages": [
-                {"role": "system", "content": "You are a helpful assistant."},
-                {"role": "user", "content": "What is 2+2?"},
-                {"role": "assistant", "content": "4"},
-                {"role": "user", "content": "And 3+3?"}
-            ]
-        });
-
-        let params: OpenAICompletionRequestParams = from_value(input).unwrap();
-
-        assert_eq!(params.messages.len(), 4);
-    }
-
-    #[test]
-    fn deserialize_request_with_opencode_style_tools() {
-        let input = json!({
-            "model": "test-model",
-            "messages": [{"role": "user", "content": "hi"}],
-            "tools": [
-                {
-                    "type": "function",
-                    "function": {
-                        "name": "glob",
-                        "description": "Fast file pattern matching tool",
-                        "parameters": {
-                            "$schema": "https://json-schema.org/draft/2020-12/schema",
-                            "type": "object",
-                            "properties": {
-                                "pattern": {"type": "string", "description": "The glob pattern"},
-                                "path": {"type": "string", "description": "The directory to search in"}
-                            },
-                            "required": ["pattern"]
-                        }
-                    }
-                },
-                {
-                    "type": "function",
-                    "function": {
-                        "name": "read",
-                        "description": "Read a file from the local filesystem",
-                        "parameters": {
-                            "$schema": "https://json-schema.org/draft/2020-12/schema",
-                            "type": "object",
-                            "properties": {
-                                "filePath": {"type": "string", "description": "The absolute path"},
-                                "offset": {
-                                    "minimum": 0,
-                                    "type": "integer",
-                                    "description": "The line number to start reading from"
-                                }
-                            },
-                            "required": ["filePath"]
-                        }
-                    }
-                }
-            ]
-        });
-
-        let params: OpenAICompletionRequestParams = from_value(input).unwrap();
-
-        assert_eq!(params.tools.len(), 2);
-    }
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/openai_default_max_tokens.rs b/paddler_balancer/src/compatibility/openai_service/openai_default_max_tokens.rs
deleted file mode 100644
index fb28f4dd3..000000000
--- a/paddler_balancer/src/compatibility/openai_service/openai_default_max_tokens.rs
+++ /dev/null
@@ -1,3 +0,0 @@
-use std::num::NonZeroU32;
-
-pub const OPENAI_DEFAULT_MAX_TOKENS: NonZeroU32 = NonZeroU32::MAX;
diff --git a/paddler_balancer/src/compatibility/openai_service/openai_error.rs b/paddler_balancer/src/compatibility/openai_service/openai_error.rs
deleted file mode 100644
index 967b7dc13..000000000
--- a/paddler_balancer/src/compatibility/openai_service/openai_error.rs
+++ /dev/null
@@ -1,371 +0,0 @@
-use actix_web::HttpResponse;
-use actix_web::http::StatusCode;
-use serde_json::Value;
-use serde_json::json;
-
-use paddler_messaging::generated_token_result::GeneratedTokenResult;
-use paddler_messaging::inference_client::message::Message as OutgoingMessage;
-use paddler_messaging::inference_client::response::Response as OutgoingResponse;
-use paddler_messaging::jsonrpc::error::Error as JsonRpcError;
-use paddler_messaging::jsonrpc::error_envelope::ErrorEnvelope;
-use paddler_messaging::jsonrpc::response_envelope::ResponseEnvelope;
-use paddler_messaging::oversized_media_details::OversizedMediaDetails;
-use paddler_messaging::oversized_prompt_details::OversizedPromptDetails;
-use paddler_messaging::raw_tool_call_tokens::RawToolCallTokens;
-
-use crate::compatibility::openai_service::openai_error_type::OpenAIErrorType;
-
-fn validation_failure_message(errors: &[String]) -> String {
-    errors.join("; ")
-}
-
-fn unrecognized_tool_call_format_message(raw: &RawToolCallTokens) -> String {
-    format!(
-        "model produced output the parser did not recognise as any registered tool-call format; \
-         FFI error: {}; raw text: {}",
-        raw.ffi_error_message, raw.text,
-    )
-}
-
-fn media_exceeds_micro_batch_message(details: &OversizedMediaDetails) -> String {
-    format!(
-        "media required {} tokens but one agent micro batch holds {} tokens",
-        details.media_tokens, details.micro_batch_tokens,
-    )
-}
-
-fn prompt_exceeds_context_size_message(details: &OversizedPromptDetails) -> String {
-    format!(
-        "prompt has {} tokens but each agent sequence holds {} tokens; shorten the prompt or raise context_size",
-        details.prompt_tokens, details.sequence_context_size,
-    )
-}
-
-const fn openai_error(error_type: OpenAIErrorType, message: String) -> OpenAIError {
-    OpenAIError {
-        error_type,
-        message,
-    }
-}
-
-fn server_error_from_token(token: &GeneratedTokenResult) -> Option<OpenAIError> {
-    match token {
-        GeneratedTokenResult::MediaExceedsMicroBatch(details) => Some(openai_error(
-            OpenAIErrorType::InvalidRequestError,
-            media_exceeds_micro_batch_message(details),
-        )),
-        GeneratedTokenResult::PromptExceedsContextSize(details) => Some(openai_error(
-            OpenAIErrorType::InvalidRequestError,
-            prompt_exceeds_context_size_message(details),
-        )),
-        GeneratedTokenResult::GrammarIncompatibleWithThinking(description)
-        | GeneratedTokenResult::GrammarSyntaxError(description)
-        | GeneratedTokenResult::ImageDecodingFailed(description)
-        | GeneratedTokenResult::MultimodalNotSupported(description)
-        | GeneratedTokenResult::MultimodalTokenizationFailed(description)
-        | GeneratedTokenResult::PromptTokenizationFailed(description)
-        | GeneratedTokenResult::ToolSchemaInvalid(description) => Some(openai_error(
-            OpenAIErrorType::InvalidRequestError,
-            description.clone(),
-        )),
-        GeneratedTokenResult::ToolCallValidationFailed(errors) => Some(openai_error(
-            OpenAIErrorType::ServerError,
-            validation_failure_message(errors),
-        )),
-        GeneratedTokenResult::UnrecognizedToolCallFormat(raw) => Some(openai_error(
-            OpenAIErrorType::ServerError,
-            unrecognized_tool_call_format_message(raw),
-        )),
-        GeneratedTokenResult::BatchAssemblyFailed(description)
-        | GeneratedTokenResult::ChatTemplateError(description)
-        | GeneratedTokenResult::DecodeFailed(description)
-        | GeneratedTokenResult::DetokenizationFailed(description)
-        | GeneratedTokenResult::GrammarInitializationFailed(description)
-        | GeneratedTokenResult::GrammarRejectedModelOutput(description)
-        | GeneratedTokenResult::KvCacheClearFailed(description)
-        | GeneratedTokenResult::MediaMicroBatchCheckFailed(description)
-        | GeneratedTokenResult::ModelNotLoaded(description)
-        | GeneratedTokenResult::MultimodalIngestionFailed(description)
-        | GeneratedTokenResult::NoSequenceSlotAvailable(description)
-        | GeneratedTokenResult::SamplerChainCreationFailed(description)
-        | GeneratedTokenResult::SamplerError(description)
-        | GeneratedTokenResult::SamplingCandidatesExhausted(description)
-        | GeneratedTokenResult::SchedulerUnavailable(description)
-        | GeneratedTokenResult::SequenceIdOutOfRange(description)
-        | GeneratedTokenResult::TokenGenerationDisabled(description)
-        | GeneratedTokenResult::ToolCallParseFailed(description)
-        | GeneratedTokenResult::ToolsSerializationFailed(description) => Some(openai_error(
-            OpenAIErrorType::ServerError,
-            description.clone(),
-        )),
-        GeneratedTokenResult::ContentToken(_)
-        | GeneratedTokenResult::Done(_)
-        | GeneratedTokenResult::ReasoningToken(_)
-        | GeneratedTokenResult::ToolCallParsed(_)
-        | GeneratedTokenResult::ToolCallToken(_)
-        | GeneratedTokenResult::UndeterminableToken(_) => None,
-    }
-}
-
-pub struct OpenAIError {
-    pub error_type: OpenAIErrorType,
-    pub message: String,
-}
-
-impl OpenAIError {
-    #[must_use]
-    pub fn classify(message: &OutgoingMessage) -> Option<Self> {
-        match message {
-            OutgoingMessage::Error(ErrorEnvelope {
-                error: JsonRpcError { description, .. },
-                ..
-            }) => Some(Self {
-                error_type: OpenAIErrorType::ServerError,
-                message: description.clone(),
-            }),
-            OutgoingMessage::Notification(_) => None,
-            OutgoingMessage::Response(ResponseEnvelope { response, .. }) => match response {
-                OutgoingResponse::GeneratedToken(token) => server_error_from_token(token),
-                OutgoingResponse::Embedding(_) => Some(Self {
-                    error_type: OpenAIErrorType::InvalidRequestError,
-                    message: "unexpected embedding response to a token generation request"
-                        .to_owned(),
-                }),
-            },
-        }
-    }
-
-    #[must_use]
-    pub fn to_envelope(&self) -> Value {
-        json!({
-            "error": {
-                "message": self.message,
-                "type": self.error_type,
-                "param": null,
-                "code": null
-            }
-        })
-    }
-
-    #[must_use]
-    pub fn to_http_response(&self, status_code: StatusCode) -> HttpResponse {
-        HttpResponse::build(status_code).json(self.to_envelope())
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use llama_cpp_bindings_types::ParsedToolCall;
-    use llama_cpp_bindings_types::TokenUsage;
-    use llama_cpp_bindings_types::ToolCallArguments;
-    use serde_json::json;
-
-    use paddler_messaging::embedding_result::EmbeddingResult;
-    use paddler_messaging::generated_token_result::GeneratedTokenResult;
-    use paddler_messaging::generation_finish::GenerationFinish;
-    use paddler_messaging::generation_summary::GenerationSummary;
-    use paddler_messaging::jsonrpc::error::Error as JsonRpcError;
-    use paddler_messaging::jsonrpc::error_envelope::ErrorEnvelope;
-    use paddler_messaging::oversized_media_details::OversizedMediaDetails;
-    use paddler_messaging::oversized_prompt_details::OversizedPromptDetails;
-
-    use super::OpenAIError;
-    use super::OutgoingMessage;
-    use super::OutgoingResponse;
-    use super::ResponseEnvelope;
-    use super::validation_failure_message;
-    use crate::compatibility::openai_service::openai_error_type::OpenAIErrorType;
-
-    fn token_message(token_result: GeneratedTokenResult) -> OutgoingMessage {
-        OutgoingMessage::Response(ResponseEnvelope {
-            generated_by: None,
-            request_id: "test-request".to_owned(),
-            response: OutgoingResponse::GeneratedToken(token_result),
-        })
-    }
-
-    #[test]
-    fn to_envelope_has_the_openai_error_shape() {
-        let envelope = OpenAIError {
-            error_type: OpenAIErrorType::ServerError,
-            message: "something went wrong".to_owned(),
-        }
-        .to_envelope();
-
-        assert_eq!(envelope["error"]["type"], "server_error");
-        assert_eq!(envelope["error"]["message"], "something went wrong");
-        assert!(envelope["error"]["param"].is_null());
-        assert!(envelope["error"]["code"].is_null());
-    }
-
-    #[test]
-    fn validation_failure_message_reports_every_error() {
-        let message =
-            validation_failure_message(&["first issue".to_owned(), "second issue".to_owned()]);
-
-        assert_eq!(message, "first issue; second issue");
-    }
-
-    #[test]
-    fn classifies_media_exceeding_the_micro_batch_as_invalid_request() {
-        let error = OpenAIError::classify(&token_message(
-            GeneratedTokenResult::MediaExceedsMicroBatch(OversizedMediaDetails {
-                media_tokens: 256,
-                micro_batch_tokens: 128,
-            }),
-        ))
-        .unwrap();
-
-        assert_eq!(error.error_type, OpenAIErrorType::InvalidRequestError);
-        assert_eq!(
-            error.message,
-            "media required 256 tokens but one agent micro batch holds 128 tokens"
-        );
-    }
-
-    #[test]
-    fn classifies_client_mistakes_as_invalid_requests() {
-        for client_mistake in [
-            GeneratedTokenResult::GrammarIncompatibleWithThinking("client mistake".to_owned()),
-            GeneratedTokenResult::GrammarSyntaxError("client mistake".to_owned()),
-            GeneratedTokenResult::ImageDecodingFailed("client mistake".to_owned()),
-            GeneratedTokenResult::MultimodalNotSupported("client mistake".to_owned()),
-            GeneratedTokenResult::MultimodalTokenizationFailed("client mistake".to_owned()),
-            GeneratedTokenResult::PromptTokenizationFailed("client mistake".to_owned()),
-            GeneratedTokenResult::ToolSchemaInvalid("client mistake".to_owned()),
-        ] {
-            let error = OpenAIError::classify(&token_message(client_mistake)).unwrap();
-
-            assert_eq!(error.error_type, OpenAIErrorType::InvalidRequestError);
-            assert_eq!(error.message, "client mistake");
-        }
-    }
-
-    #[test]
-    fn classifies_agent_failures_as_server_errors() {
-        for agent_failure in [
-            GeneratedTokenResult::BatchAssemblyFailed("agent failure".to_owned()),
-            GeneratedTokenResult::ChatTemplateError("agent failure".to_owned()),
-            GeneratedTokenResult::DecodeFailed("agent failure".to_owned()),
-            GeneratedTokenResult::DetokenizationFailed("agent failure".to_owned()),
-            GeneratedTokenResult::GrammarInitializationFailed("agent failure".to_owned()),
-            GeneratedTokenResult::GrammarRejectedModelOutput("agent failure".to_owned()),
-            GeneratedTokenResult::KvCacheClearFailed("agent failure".to_owned()),
-            GeneratedTokenResult::MediaMicroBatchCheckFailed("agent failure".to_owned()),
-            GeneratedTokenResult::ModelNotLoaded("agent failure".to_owned()),
-            GeneratedTokenResult::MultimodalIngestionFailed("agent failure".to_owned()),
-            GeneratedTokenResult::NoSequenceSlotAvailable("agent failure".to_owned()),
-            GeneratedTokenResult::SamplerChainCreationFailed("agent failure".to_owned()),
-            GeneratedTokenResult::SamplerError("agent failure".to_owned()),
-            GeneratedTokenResult::SamplingCandidatesExhausted("agent failure".to_owned()),
-            GeneratedTokenResult::SchedulerUnavailable("agent failure".to_owned()),
-            GeneratedTokenResult::SequenceIdOutOfRange("agent failure".to_owned()),
-            GeneratedTokenResult::TokenGenerationDisabled("agent failure".to_owned()),
-            GeneratedTokenResult::ToolCallParseFailed("agent failure".to_owned()),
-            GeneratedTokenResult::ToolsSerializationFailed("agent failure".to_owned()),
-        ] {
-            let error = OpenAIError::classify(&token_message(agent_failure)).unwrap();
-
-            assert_eq!(error.error_type, OpenAIErrorType::ServerError);
-            assert_eq!(error.message, "agent failure");
-        }
-    }
-
-    #[test]
-    fn classifies_jsonrpc_error_as_server_error() {
-        let message = OutgoingMessage::Error(ErrorEnvelope {
-            request_id: "test-request".to_owned(),
-            error: JsonRpcError {
-                code: 500,
-                description: "internal failure".to_owned(),
-            },
-        });
-
-        let classified = OpenAIError::classify(&message).unwrap();
-
-        assert_eq!(classified.error_type, OpenAIErrorType::ServerError);
-        assert_eq!(classified.message, "internal failure");
-    }
-
-    #[test]
-    fn classifies_oversized_prompt_as_invalid_request() {
-        let error = OpenAIError::classify(&token_message(
-            GeneratedTokenResult::PromptExceedsContextSize(OversizedPromptDetails {
-                prompt_tokens: 9895,
-                sequence_context_size: 8192,
-            }),
-        ))
-        .unwrap();
-
-        assert_eq!(error.error_type, OpenAIErrorType::InvalidRequestError);
-        assert_eq!(
-            error.message,
-            "prompt has 9895 tokens but each agent sequence holds 8192 tokens; shorten the prompt or raise context_size"
-        );
-    }
-
-    #[test]
-    fn classifies_tool_call_validation_failure_as_server_error() {
-        let classified = OpenAIError::classify(&token_message(
-            GeneratedTokenResult::ToolCallValidationFailed(vec!["missing field x".to_owned()]),
-        ))
-        .unwrap();
-
-        assert_eq!(classified.error_type, OpenAIErrorType::ServerError);
-        assert_eq!(classified.message, "missing field x");
-    }
-
-    #[test]
-    fn does_not_classify_a_content_token() {
-        assert!(
-            OpenAIError::classify(&token_message(GeneratedTokenResult::ContentToken(
-                "hello".to_owned()
-            )))
-            .is_none()
-        );
-    }
-
-    #[test]
-    fn does_not_classify_a_done_summary() {
-        assert!(
-            OpenAIError::classify(&token_message(GeneratedTokenResult::Done(
-                GenerationSummary {
-                    finish: GenerationFinish::EndOfGeneration,
-                    usage: TokenUsage::new(),
-                }
-            )))
-            .is_none()
-        );
-    }
-
-    #[test]
-    fn classifies_an_embedding_response_as_an_invalid_request() {
-        let message = OutgoingMessage::Response(ResponseEnvelope {
-            generated_by: None,
-            request_id: "test-request".to_owned(),
-            response: OutgoingResponse::Embedding(EmbeddingResult::Done),
-        });
-
-        let error = OpenAIError::classify(&message).unwrap();
-
-        assert_eq!(error.error_type, OpenAIErrorType::InvalidRequestError);
-        assert_eq!(
-            error.message,
-            "unexpected embedding response to a token generation request"
-        );
-    }
-
-    #[test]
-    fn does_not_classify_a_parsed_tool_call() {
-        let parsed = vec![ParsedToolCall::new(
-            "call_x".to_owned(),
-            "get_weather".to_owned(),
-            ToolCallArguments::ValidJson(json!({"location": "Paris"})),
-        )];
-
-        assert!(
-            OpenAIError::classify(&token_message(GeneratedTokenResult::ToolCallParsed(parsed)))
-                .is_none()
-        );
-    }
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/openai_error_body.rs b/paddler_balancer/src/compatibility/openai_service/openai_error_body.rs
new file mode 100644
index 000000000..1b7d41205
--- /dev/null
+++ b/paddler_balancer/src/compatibility/openai_service/openai_error_body.rs
@@ -0,0 +1,52 @@
+use actix_web::HttpResponse;
+use actix_web::http::StatusCode;
+use serde::Serialize;
+use serde::Serializer;
+
+use crate::compatibility::openai_service::openai_error_type::OpenAIErrorType;
+
+#[derive(Serialize)]
+struct SerializedError<'body> {
+    message: &'body str,
+    #[serde(rename = "type")]
+    error_type: OpenAIErrorType,
+    param: (),
+    code: (),
+}
+
+#[derive(Serialize)]
+struct SerializedErrorBody<'body> {
+    error: SerializedError<'body>,
+}
+
+pub struct OpenAIErrorBody {
+    pub error_type: OpenAIErrorType,
+    pub message: String,
+}
+
+impl OpenAIErrorBody {
+    #[must_use]
+    pub fn into_http_response(self, status_code: StatusCode) -> HttpResponse {
+        HttpResponse::build(status_code).json(self)
+    }
+}
+
+impl Serialize for OpenAIErrorBody {
+    fn serialize<TSerializer>(
+        &self,
+        serializer: TSerializer,
+    ) -> Result<TSerializer::Ok, TSerializer::Error>
+    where
+        TSerializer: Serializer,
+    {
+        SerializedErrorBody {
+            error: SerializedError {
+                message: &self.message,
+                error_type: self.error_type,
+                param: (),
+                code: (),
+            },
+        }
+        .serialize(serializer)
+    }
+}
diff --git a/paddler_balancer/src/compatibility/openai_service/openai_error_type.rs b/paddler_balancer/src/compatibility/openai_service/openai_error_type.rs
index 942753bfb..35456d7b9 100644
--- a/paddler_balancer/src/compatibility/openai_service/openai_error_type.rs
+++ b/paddler_balancer/src/compatibility/openai_service/openai_error_type.rs
@@ -1,3 +1,4 @@
+use actix_web::http::StatusCode;
 use serde::Serialize;
 
 #[derive(Clone, Copy, Debug, Eq, PartialEq, Serialize)]
@@ -6,3 +7,13 @@ pub enum OpenAIErrorType {
     InvalidRequestError,
     ServerError,
 }
+
+impl From<StatusCode> for OpenAIErrorType {
+    fn from(status_code: StatusCode) -> Self {
+        if status_code.is_client_error() {
+            Self::InvalidRequestError
+        } else {
+            Self::ServerError
+        }
+    }
+}
diff --git a/paddler_balancer/src/compatibility/openai_service/openai_failure.rs b/paddler_balancer/src/compatibility/openai_service/openai_failure.rs
new file mode 100644
index 000000000..96d442799
--- /dev/null
+++ b/paddler_balancer/src/compatibility/openai_service/openai_failure.rs
@@ -0,0 +1,249 @@
+use actix_web::HttpResponse;
+use actix_web::http::StatusCode;
+
+use paddler_openai_translation::generation_failure::GenerationFailure;
+use paddler_openai_translation::generation_failure_cause::GenerationFailureCause;
+use paddler_openai_translation::openai_translation_error::OpenAITranslationError;
+
+use crate::compatibility::openai_service::openai_error_body::OpenAIErrorBody;
+use crate::compatibility::openai_service::openai_error_type::OpenAIErrorType;
+use crate::compatibility::upstream_failure::UpstreamFailure;
+use crate::relay_failure::RelayFailure;
+
+pub enum OpenAIFailure {
+    Generation(GenerationFailure),
+    Relay(RelayFailure),
+    RequestUntranslatable(OpenAITranslationError),
+}
+
+impl OpenAIFailure {
+    #[must_use]
+    pub fn into_error_body(self) -> OpenAIErrorBody {
+        let status_code = self.status_code();
+
+        self.into_error_body_for(status_code)
+    }
+
+    #[must_use]
+    pub fn into_http_response(self) -> HttpResponse {
+        let status_code = self.status_code();
+
+        self.into_error_body_for(status_code)
+            .into_http_response(status_code)
+    }
+
+    fn into_error_body_for(self, status_code: StatusCode) -> OpenAIErrorBody {
+        OpenAIErrorBody {
+            error_type: OpenAIErrorType::from(status_code),
+            message: self.into_message(),
+        }
+    }
+
+    fn into_message(self) -> String {
+        match self {
+            Self::Generation(GenerationFailure { message, .. }) => message,
+            Self::Relay(relay_failure) => relay_failure.description(),
+            Self::RequestUntranslatable(translation_error) => translation_error.to_string(),
+        }
+    }
+
+    fn status_code(&self) -> StatusCode {
+        match self {
+            Self::Generation(GenerationFailure {
+                cause: GenerationFailureCause::InvalidRequest,
+                ..
+            })
+            | Self::RequestUntranslatable(
+                OpenAITranslationError::MessageWithoutContent { .. }
+                | OpenAITranslationError::ReasoningItemWithoutContent
+                | OpenAITranslationError::ReasoningOnNonAssistantMessage { .. }
+                | OpenAITranslationError::ToolCallArgumentsAreNotAJsonObject { .. }
+                | OpenAITranslationError::ToolCallsOnNonAssistantMessage { .. }
+                | OpenAITranslationError::ToolRejected(_)
+                | OpenAITranslationError::ToolResultWithoutToolCallId,
+            ) => StatusCode::BAD_REQUEST,
+            Self::Generation(GenerationFailure {
+                cause: GenerationFailureCause::Unavailable,
+                ..
+            }) => UpstreamFailure::Unavailable.status_code(),
+            Self::Generation(GenerationFailure {
+                cause: GenerationFailureCause::AgentFailed,
+                ..
+            }) => UpstreamFailure::AgentFailed.status_code(),
+            Self::RequestUntranslatable(OpenAITranslationError::ClockBeforeUnixEpoch(_)) => {
+                StatusCode::INTERNAL_SERVER_ERROR
+            }
+            Self::Relay(relay_failure) => UpstreamFailure::from(relay_failure).status_code(),
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::time::Duration;
+    use std::time::UNIX_EPOCH;
+
+    use actix_web::body::to_bytes;
+    use actix_web::http::StatusCode;
+    use serde_json::Map;
+    use serde_json::Value;
+    use serde_json::from_slice;
+    use serde_json::from_str;
+    use serde_json::json;
+
+    use paddler_messaging::request_params_validation_error::RequestParamsValidationError;
+    use paddler_openai_translation::chat_completion_message_role::ChatCompletionMessageRole;
+    use paddler_openai_translation::generation_failure::GenerationFailure;
+    use paddler_openai_translation::generation_failure_cause::GenerationFailureCause;
+    use paddler_openai_translation::openai_translation_error::OpenAITranslationError;
+
+    use super::OpenAIFailure;
+    use crate::relay_failure::RelayFailure;
+
+    fn error_body(message: &str, error_type: &str) -> Value {
+        json!({
+            "error": {
+                "message": message,
+                "type": error_type,
+                "param": null,
+                "code": null
+            }
+        })
+    }
+
+    fn generation_failure(cause: GenerationFailureCause, message: &str) -> OpenAIFailure {
+        OpenAIFailure::Generation(GenerationFailure {
+            cause,
+            message: message.to_owned(),
+        })
+    }
+
+    #[actix_web::test]
+    async fn answers_every_failure_with_its_status_and_openai_error_body() {
+        let clock_error = (UNIX_EPOCH - Duration::from_secs(1))
+            .duration_since(UNIX_EPOCH)
+            .unwrap_err();
+
+        for (openai_failure, expected_status, expected_body) in [
+            (
+                generation_failure(GenerationFailureCause::InvalidRequest, "client mistake"),
+                StatusCode::BAD_REQUEST,
+                error_body("client mistake", "invalid_request_error"),
+            ),
+            (
+                generation_failure(GenerationFailureCause::Unavailable, "no model is loaded"),
+                StatusCode::SERVICE_UNAVAILABLE,
+                error_body("no model is loaded", "server_error"),
+            ),
+            (
+                generation_failure(GenerationFailureCause::AgentFailed, "sampler blew up"),
+                StatusCode::INTERNAL_SERVER_ERROR,
+                error_body("sampler blew up", "server_error"),
+            ),
+            (
+                OpenAIFailure::RequestUntranslatable(OpenAITranslationError::ToolRejected(
+                    RequestParamsValidationError::RequiredFieldNotInProperties {
+                        field: "absent".to_owned(),
+                    },
+                )),
+                StatusCode::BAD_REQUEST,
+                error_body(
+                    "Required field 'absent' not found in properties",
+                    "invalid_request_error",
+                ),
+            ),
+            (
+                OpenAIFailure::RequestUntranslatable(
+                    OpenAITranslationError::MessageWithoutContent {
+                        role: ChatCompletionMessageRole::User,
+                    },
+                ),
+                StatusCode::BAD_REQUEST,
+                error_body(
+                    "a 'user' message must have content",
+                    "invalid_request_error",
+                ),
+            ),
+            (
+                OpenAIFailure::RequestUntranslatable(
+                    OpenAITranslationError::ReasoningOnNonAssistantMessage {
+                        role: ChatCompletionMessageRole::Developer,
+                    },
+                ),
+                StatusCode::BAD_REQUEST,
+                error_body(
+                    "only an assistant message can carry reasoning, not a 'developer' message",
+                    "invalid_request_error",
+                ),
+            ),
+            (
+                OpenAIFailure::RequestUntranslatable(
+                    OpenAITranslationError::ReasoningItemWithoutContent,
+                ),
+                StatusCode::BAD_REQUEST,
+                error_body(
+                    "a reasoning item must carry its reasoning text",
+                    "invalid_request_error",
+                ),
+            ),
+            (
+                OpenAIFailure::RequestUntranslatable(
+                    OpenAITranslationError::ToolCallArgumentsAreNotAJsonObject {
+                        tool_call_id: "call_0".to_owned(),
+                        source: from_str::<Map<String, Value>>("[]").unwrap_err(),
+                    },
+                ),
+                StatusCode::BAD_REQUEST,
+                error_body(
+                    "the arguments of the tool call 'call_0' are not a JSON object",
+                    "invalid_request_error",
+                ),
+            ),
+            (
+                OpenAIFailure::RequestUntranslatable(
+                    OpenAITranslationError::ToolCallsOnNonAssistantMessage {
+                        role: ChatCompletionMessageRole::User,
+                    },
+                ),
+                StatusCode::BAD_REQUEST,
+                error_body(
+                    "only an assistant message can request tool calls, not a 'user' message",
+                    "invalid_request_error",
+                ),
+            ),
+            (
+                OpenAIFailure::RequestUntranslatable(
+                    OpenAITranslationError::ToolResultWithoutToolCallId,
+                ),
+                StatusCode::BAD_REQUEST,
+                error_body(
+                    "a tool message must name the tool call it answers in 'tool_call_id'",
+                    "invalid_request_error",
+                ),
+            ),
+            (
+                OpenAIFailure::RequestUntranslatable(OpenAITranslationError::ClockBeforeUnixEpoch(
+                    clock_error,
+                )),
+                StatusCode::INTERNAL_SERVER_ERROR,
+                error_body(
+                    "the system clock reads a time before the Unix epoch",
+                    "server_error",
+                ),
+            ),
+            (
+                OpenAIFailure::Relay(RelayFailure::BufferedRequestTimedOut),
+                StatusCode::GATEWAY_TIMEOUT,
+                error_body("Waiting for available slot timed out", "server_error"),
+            ),
+        ] {
+            let response = openai_failure.into_http_response();
+
+            assert_eq!(response.status(), expected_status);
+            assert_eq!(
+                from_slice::<Value>(&to_bytes(response.into_body()).await.unwrap()).unwrap(),
+                expected_body
+            );
+        }
+    }
+}
diff --git a/paddler_balancer/src/compatibility/openai_service/openai_function_definition.rs b/paddler_balancer/src/compatibility/openai_service/openai_function_definition.rs
deleted file mode 100644
index 671b0eecb..000000000
--- a/paddler_balancer/src/compatibility/openai_service/openai_function_definition.rs
+++ /dev/null
@@ -1,79 +0,0 @@
-use serde::Deserialize;
-
-use paddler_messaging::request_params::continue_from_conversation_history_params::tool::Tool;
-use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::FunctionCall;
-use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::function::Function;
-use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters::Parameters;
-use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters_schema::raw_parameters_schema::RawParametersSchema;
-
-use crate::compatibility::openai_service::openai_tool_parameters_schema::OpenAIToolParametersSchema;
-
-#[derive(Deserialize)]
-pub struct OpenAIFunctionDefinition {
-    pub name: String,
-    #[serde(default)]
-    pub description: Option<String>,
-    #[serde(default)]
-    pub parameters: Option<OpenAIToolParametersSchema>,
-}
-
-impl OpenAIFunctionDefinition {
-    #[must_use]
-    pub fn into_tool(self) -> Tool<RawParametersSchema> {
-        let Self {
-            name,
-            description,
-            parameters,
-        } = self;
-
-        Tool::Function(FunctionCall {
-            function: Function {
-                name,
-                description: description.unwrap_or_default(),
-                parameters: parameters.map_or(Parameters::Empty, |parameters| {
-                    Parameters::Schema(parameters.into_raw_parameters_schema())
-                }),
-            },
-        })
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use serde_json::from_value;
-    use serde_json::json;
-
-    use paddler_messaging::request_params::continue_from_conversation_history_params::tool::Tool;
-
-    use super::OpenAIFunctionDefinition;
-
-    #[test]
-    fn conversion_ignores_unknown_fields_and_builds_function_with_schema() {
-        let function: OpenAIFunctionDefinition = from_value(json!({
-            "name": "get_weather",
-            "description": "fetch weather",
-            "parameters": {"type": "object", "properties": {"location": {"type": "string"}}},
-            "strict": true
-        }))
-        .unwrap();
-
-        let Tool::Function(function_call) = function.into_tool();
-
-        assert_eq!(function_call.function.name, "get_weather");
-        assert_eq!(function_call.function.description, "fetch weather");
-        assert!(!function_call.function.parameters.is_empty());
-    }
-
-    #[test]
-    fn conversion_defaults_missing_description_and_parameters() {
-        let function: OpenAIFunctionDefinition = from_value(json!({
-            "name": "noop"
-        }))
-        .unwrap();
-
-        let Tool::Function(function_call) = function.into_tool();
-
-        assert_eq!(function_call.function.description, "");
-        assert!(function_call.function.parameters.is_empty());
-    }
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/openai_header.rs b/paddler_balancer/src/compatibility/openai_service/openai_header.rs
new file mode 100644
index 000000000..7ffe1be19
--- /dev/null
+++ b/paddler_balancer/src/compatibility/openai_service/openai_header.rs
@@ -0,0 +1,5 @@
+pub struct OpenAIHeader;
+
+impl OpenAIHeader {
+    pub const REQUEST_ID: &str = "x-request-id";
+}
diff --git a/paddler_balancer/src/compatibility/openai_service/openai_http_response.rs b/paddler_balancer/src/compatibility/openai_service/openai_http_response.rs
new file mode 100644
index 000000000..af2c86386
--- /dev/null
+++ b/paddler_balancer/src/compatibility/openai_service/openai_http_response.rs
@@ -0,0 +1,69 @@
+use actix_web::HttpResponse;
+use nanoid::nanoid;
+
+use paddler_messaging::generated_token_result::GeneratedTokenResult;
+use paddler_openai_translation::completes_generation::CompletesGeneration;
+use paddler_openai_translation::generated_output::GeneratedOutput;
+use paddler_openai_translation::generation_outcome::GenerationOutcome;
+use paddler_openai_translation::generation_progress::GenerationProgress;
+
+use crate::agent_relay::AgentRelay;
+use crate::agent_relay_conclusion::AgentRelayConclusion;
+use crate::agent_relay_step::AgentRelayStep;
+use crate::agent_streaming_request::AgentStreamingRequest;
+use crate::compatibility::openai_service::openai_failure::OpenAIFailure;
+
+pub async fn openai_http_response<TCompletesGeneration, TParams>(
+    completes_generation: &TCompletesGeneration,
+    mut agent_relay: AgentRelay<TParams>,
+) -> HttpResponse
+where
+    TCompletesGeneration: CompletesGeneration,
+    TParams: AgentStreamingRequest<Response = GeneratedTokenResult>,
+{
+    let mut generated_output = GeneratedOutput::default();
+
+    loop {
+        match agent_relay.next().await {
+            AgentRelayStep::Concluded(AgentRelayConclusion::AgentConcluded {
+                final_result,
+                ..
+            }) => {
+                return match GenerationOutcome::from(final_result) {
+                    GenerationOutcome::Failed(generation_failure) => {
+                        OpenAIFailure::Generation(generation_failure).into_http_response()
+                    }
+                    GenerationOutcome::Finished(generation_summary) => {
+                        let request_id: String = nanoid!();
+
+                        HttpResponse::Ok().json(completes_generation.complete(
+                            &request_id,
+                            generated_output,
+                            &generation_summary,
+                        ))
+                    }
+                };
+            }
+            AgentRelayStep::Concluded(AgentRelayConclusion::Failed(relay_failure)) => {
+                return OpenAIFailure::Relay(relay_failure).into_http_response();
+            }
+            AgentRelayStep::Partial {
+                partial_result,
+                relay,
+                ..
+            } => {
+                match GenerationProgress::from(partial_result) {
+                    GenerationProgress::Failed(generation_failure) => {
+                        return OpenAIFailure::Generation(generation_failure).into_http_response();
+                    }
+                    GenerationProgress::Produced(generated_output_part) => {
+                        generated_output.append(generated_output_part);
+                    }
+                    GenerationProgress::ToolCallTokenProduced => {}
+                }
+
+                agent_relay = relay;
+            }
+        }
+    }
+}
diff --git a/paddler_balancer/src/compatibility/openai_service/openai_json_config.rs b/paddler_balancer/src/compatibility/openai_service/openai_json_config.rs
index 5e177d94e..8419b308c 100644
--- a/paddler_balancer/src/compatibility/openai_service/openai_json_config.rs
+++ b/paddler_balancer/src/compatibility/openai_service/openai_json_config.rs
@@ -2,17 +2,17 @@ use actix_web::error::InternalError;
 use actix_web::http::StatusCode;
 use actix_web::web::JsonConfig;
 
-use crate::compatibility::openai_service::openai_error::OpenAIError;
+use crate::compatibility::openai_service::openai_error_body::OpenAIErrorBody;
 use crate::compatibility::openai_service::openai_error_type::OpenAIErrorType;
 
 #[must_use]
 pub fn openai_json_config() -> JsonConfig {
     JsonConfig::default().error_handler(|json_payload_error, _request| {
-        let bad_request = OpenAIError {
+        let bad_request = OpenAIErrorBody {
             error_type: OpenAIErrorType::InvalidRequestError,
             message: json_payload_error.to_string(),
         }
-        .to_http_response(StatusCode::BAD_REQUEST);
+        .into_http_response(StatusCode::BAD_REQUEST);
 
         InternalError::from_response(json_payload_error, bad_request).into()
     })
diff --git a/paddler_balancer/src/compatibility/openai_service/openai_json_response.rs b/paddler_balancer/src/compatibility/openai_service/openai_json_response.rs
deleted file mode 100644
index 18535bd58..000000000
--- a/paddler_balancer/src/compatibility/openai_service/openai_json_response.rs
+++ /dev/null
@@ -1,113 +0,0 @@
-use actix_web::HttpResponse;
-use actix_web::http::StatusCode;
-use futures::stream::Stream;
-use futures::stream::StreamExt as _;
-
-use crate::chunk_forwarding_session_controller::transform_result::TransformResult;
-use crate::compatibility::openai_service::openai_error::OpenAIError;
-use crate::compatibility::openai_service::openai_error_type::OpenAIErrorType;
-
-pub async fn openai_json_response<TStream>(
-    transform_results: TStream,
-    missing_body_message: &str,
-) -> HttpResponse
-where
-    TStream: Stream<Item = TransformResult>,
-{
-    let mut json_bodies: Vec<String> = Vec::new();
-
-    for transform_result in transform_results.collect::<Vec<TransformResult>>().await {
-        match transform_result {
-            TransformResult::Chunk(json_body) => json_bodies.push(json_body),
-            TransformResult::Discard => {}
-            TransformResult::Error(error_json) => {
-                return HttpResponse::InternalServerError()
-                    .content_type("application/json")
-                    .body(error_json);
-            }
-        }
-    }
-
-    json_bodies.into_iter().next().map_or_else(
-        || {
-            OpenAIError {
-                error_type: OpenAIErrorType::ServerError,
-                message: missing_body_message.to_owned(),
-            }
-            .to_http_response(StatusCode::INTERNAL_SERVER_ERROR)
-        },
-        |json_body| {
-            HttpResponse::Ok()
-                .content_type("application/json")
-                .body(json_body)
-        },
-    )
-}
-
-#[cfg(test)]
-mod tests {
-    use actix_web::body::to_bytes;
-    use actix_web::http::StatusCode;
-    use futures::stream::iter;
-    use serde_json::Value;
-    use serde_json::from_slice;
-    use serde_json::json;
-
-    use super::openai_json_response;
-    use crate::chunk_forwarding_session_controller::transform_result::TransformResult;
-
-    #[actix_web::test]
-    async fn answers_with_the_completion_that_follows_discarded_results() {
-        let response = openai_json_response(
-            iter([
-                TransformResult::Discard,
-                TransformResult::Chunk(r#"{"id":"completion"}"#.to_owned()),
-            ]),
-            "no completion",
-        )
-        .await;
-
-        assert_eq!(response.status(), StatusCode::OK);
-        assert_eq!(
-            to_bytes(response.into_body()).await.unwrap(),
-            r#"{"id":"completion"}"#
-        );
-    }
-
-    #[actix_web::test]
-    async fn answers_with_the_error_even_after_a_completion() {
-        let response = openai_json_response(
-            iter([
-                TransformResult::Chunk(r#"{"id":"completion"}"#.to_owned()),
-                TransformResult::Error(r#"{"error":"agent failed"}"#.to_owned()),
-            ]),
-            "no completion",
-        )
-        .await;
-
-        assert_eq!(response.status(), StatusCode::INTERNAL_SERVER_ERROR);
-        assert_eq!(
-            to_bytes(response.into_body()).await.unwrap(),
-            r#"{"error":"agent failed"}"#
-        );
-    }
-
-    #[actix_web::test]
-    async fn answers_a_server_error_when_no_completion_was_produced() {
-        let response =
-            openai_json_response(iter([TransformResult::Discard]), "no completion").await;
-
-        assert_eq!(response.status(), StatusCode::INTERNAL_SERVER_ERROR);
-        assert_eq!(
-            from_slice::<Value>(&to_bytes(response.into_body()).await.unwrap()).unwrap(),
-            json!({
-                "error": {
-                    "message": "no completion",
-                    "type": "server_error",
-                    "param": null,
-                    "code": null
-                }
-            })
-        );
-    }
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/openai_message.rs b/paddler_balancer/src/compatibility/openai_service/openai_message.rs
deleted file mode 100644
index 85800f9d8..000000000
--- a/paddler_balancer/src/compatibility/openai_service/openai_message.rs
+++ /dev/null
@@ -1,46 +0,0 @@
-use serde::Deserialize;
-
-use paddler_messaging::conversation_message::ConversationMessage;
-use paddler_messaging::conversation_message_content::ConversationMessageContent;
-
-#[derive(Deserialize)]
-pub struct OpenAIMessage {
-    pub content: ConversationMessageContent,
-    pub role: String,
-}
-
-impl OpenAIMessage {
-    #[must_use]
-    pub fn into_conversation_message(self) -> ConversationMessage {
-        ConversationMessage {
-            content: self.content,
-            role: self.role,
-        }
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use serde_json::from_value;
-    use serde_json::json;
-
-    use super::OpenAIMessage;
-
-    #[test]
-    fn openai_message_converts_to_conversation_message() {
-        let input = json!({
-            "role": "user",
-            "content": [
-                {"type": "text", "text": "OCR this"},
-                {"type": "image_url", "image_url": {"url": "data:image/png;base64,abc"}}
-            ]
-        });
-
-        let openai_message: OpenAIMessage = from_value(input).unwrap();
-        let conversation_message = openai_message.into_conversation_message();
-
-        assert_eq!(conversation_message.role, "user");
-        assert_eq!(conversation_message.content.text_content(), "OCR this");
-        assert_eq!(conversation_message.content.image_urls().len(), 1);
-    }
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/openai_non_streaming_response_transformer.rs b/paddler_balancer/src/compatibility/openai_service/openai_non_streaming_response_transformer.rs
deleted file mode 100644
index b67d57018..000000000
--- a/paddler_balancer/src/compatibility/openai_service/openai_non_streaming_response_transformer.rs
+++ /dev/null
@@ -1,543 +0,0 @@
-use std::mem::take;
-use std::sync::Arc;
-
-use async_trait::async_trait;
-use llama_cpp_bindings_types::ParsedToolCall;
-use parking_lot::Mutex;
-use serde_json::to_string;
-
-use paddler_messaging::generated_token_result::GeneratedTokenResult;
-use paddler_messaging::generation_summary::GenerationSummary;
-use paddler_messaging::inference_client::message::Message as OutgoingMessage;
-use paddler_messaging::inference_client::response::Response as OutgoingResponse;
-use paddler_messaging::jsonrpc::response_envelope::ResponseEnvelope;
-
-use crate::agent_relay_error::AgentRelayError;
-use crate::chunk_forwarding_session_controller::transform_result::TransformResult;
-use crate::chunk_forwarding_session_controller::transforms_outgoing_message::TransformsOutgoingMessage;
-use crate::compatibility::openai_service::chat_completion::ChatCompletion;
-use crate::compatibility::openai_service::chat_completion_finish_reason::chat_completion_finish_reason;
-use crate::compatibility::openai_service::openai_non_streaming_state::OpenAINonStreamingState;
-use crate::compatibility::openai_service::openai_usage::OpenAIUsage;
-use crate::compatibility::openai_service::try_universal_error_chunk::try_universal_error_chunk;
-
-#[derive(Clone)]
-pub struct OpenAINonStreamingResponseTransformer {
-    pub created: u64,
-    pub model: String,
-    pub state: Arc<Mutex<OpenAINonStreamingState>>,
-}
-
-impl OpenAINonStreamingResponseTransformer {
-    fn append_content(&self, text: &str) {
-        self.state.lock().content.push_str(text);
-    }
-
-    fn append_tool_calls(&self, parsed_calls: Vec<ParsedToolCall>) {
-        self.state.lock().tool_calls.extend(parsed_calls);
-    }
-
-    fn build_done_chunk(
-        &self,
-        request_id: &str,
-        summary: &GenerationSummary,
-    ) -> Result<String, AgentRelayError> {
-        let snapshot = take(&mut *self.state.lock());
-
-        let has_tool_calls = !snapshot.tool_calls.is_empty();
-        let finish_reason = chat_completion_finish_reason(summary.finish, has_tool_calls);
-
-        to_string(&ChatCompletion {
-            content: &snapshot.content,
-            created: self.created,
-            finish_reason,
-            id: request_id,
-            model: &self.model,
-            tool_calls: &snapshot.tool_calls,
-            usage: OpenAIUsage(summary.usage),
-        })
-        .map_err(AgentRelayError::MessageUnserializable)
-    }
-}
-
-#[async_trait]
-impl TransformsOutgoingMessage for OpenAINonStreamingResponseTransformer {
-    type Output = TransformResult;
-
-    async fn transform(
-        &self,
-        message: OutgoingMessage,
-    ) -> Result<Vec<TransformResult>, AgentRelayError> {
-        match message {
-            OutgoingMessage::Response(ResponseEnvelope {
-                response:
-                    OutgoingResponse::GeneratedToken(
-                        GeneratedTokenResult::ContentToken(text)
-                        | GeneratedTokenResult::UndeterminableToken(text),
-                    ),
-                ..
-            }) => {
-                self.append_content(&text);
-                Ok(vec![])
-            }
-            OutgoingMessage::Response(ResponseEnvelope {
-                response:
-                    OutgoingResponse::GeneratedToken(
-                        GeneratedTokenResult::ReasoningToken(_)
-                        | GeneratedTokenResult::ToolCallToken(_),
-                    ),
-                ..
-            }) => Ok(vec![]),
-            OutgoingMessage::Response(ResponseEnvelope {
-                response:
-                    OutgoingResponse::GeneratedToken(GeneratedTokenResult::ToolCallParsed(parsed_calls)),
-                ..
-            }) => {
-                self.append_tool_calls(parsed_calls);
-                Ok(vec![])
-            }
-            OutgoingMessage::Response(ResponseEnvelope {
-                request_id,
-                response: OutgoingResponse::GeneratedToken(GeneratedTokenResult::Done(summary)),
-                ..
-            }) => Ok(vec![TransformResult::Chunk(
-                self.build_done_chunk(&request_id, &summary)?,
-            )]),
-            other => try_universal_error_chunk(&other)
-                .map(|error_chunk| vec![error_chunk])
-                .ok_or_else(|| AgentRelayError::MessageNotRelayable {
-                    message: Box::new(other),
-                }),
-        }
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use std::sync::Arc;
-
-    use llama_cpp_bindings_types::ParsedToolCall;
-    use llama_cpp_bindings_types::TokenUsage;
-    use llama_cpp_bindings_types::ToolCallArguments;
-    use parking_lot::Mutex;
-    use serde_json::json;
-
-    use paddler_messaging::embedding_result::EmbeddingResult;
-    use paddler_messaging::generated_token_result::GeneratedTokenResult;
-    use paddler_messaging::generation_finish::GenerationFinish;
-    use paddler_messaging::generation_summary::GenerationSummary;
-    use paddler_messaging::inference_client::message::Message as OutgoingMessage;
-    use paddler_messaging::inference_client::notification::Notification;
-    use paddler_messaging::inference_client::response::Response as OutgoingResponse;
-    use paddler_messaging::jsonrpc::error::Error as JsonRpcError;
-    use paddler_messaging::jsonrpc::error_envelope::ErrorEnvelope;
-    use paddler_messaging::jsonrpc::response_envelope::ResponseEnvelope;
-    use paddler_messaging::oversized_media_details::OversizedMediaDetails;
-    use paddler_messaging::raw_tool_call_tokens::RawToolCallTokens;
-
-    use super::OpenAINonStreamingResponseTransformer;
-    use super::OpenAINonStreamingState;
-    use crate::agent_relay_error::AgentRelayError;
-    use crate::chunk_forwarding_session_controller::transform_result::TransformResult;
-    use crate::chunk_forwarding_session_controller::transforms_outgoing_message::TransformsOutgoingMessage;
-
-    #[must_use]
-    pub fn token_message(token_result: GeneratedTokenResult) -> OutgoingMessage {
-        OutgoingMessage::Response(ResponseEnvelope {
-            generated_by: None,
-            request_id: "test-request".to_owned(),
-            response: OutgoingResponse::GeneratedToken(token_result),
-        })
-    }
-
-    #[must_use]
-    pub fn error_message(code: i32, description: &str) -> OutgoingMessage {
-        OutgoingMessage::Error(ErrorEnvelope {
-            request_id: "test-request".to_owned(),
-            error: JsonRpcError {
-                code,
-                description: description.to_owned(),
-            },
-        })
-    }
-
-    #[must_use]
-    pub fn response_message(response: OutgoingResponse) -> OutgoingMessage {
-        OutgoingMessage::Response(ResponseEnvelope {
-            generated_by: None,
-            request_id: "test-request".to_owned(),
-            response,
-        })
-    }
-
-    #[must_use]
-    pub fn summary_with_counts(
-        prompt_tokens: u64,
-        content_tokens: u64,
-        reasoning_tokens: u64,
-    ) -> GenerationSummary {
-        GenerationSummary {
-            finish: GenerationFinish::EndOfGeneration,
-            usage: TokenUsage {
-                prompt_tokens,
-                content_tokens,
-                reasoning_tokens,
-                ..TokenUsage::default()
-            },
-        }
-    }
-
-    #[must_use]
-    pub fn weather_call() -> ParsedToolCall {
-        ParsedToolCall::new(
-            "call_x".to_owned(),
-            "get_weather".to_owned(),
-            ToolCallArguments::ValidJson(json!({ "location": "Paris" })),
-        )
-    }
-
-    #[must_use]
-    pub fn invalid_json_call() -> ParsedToolCall {
-        ParsedToolCall::new(
-            "call_invalid".to_owned(),
-            "broken_tool".to_owned(),
-            ToolCallArguments::InvalidJson("{not valid json".to_owned()),
-        )
-    }
-
-    fn non_streaming_transformer() -> OpenAINonStreamingResponseTransformer {
-        OpenAINonStreamingResponseTransformer {
-            created: 0,
-            model: "test-model".to_owned(),
-            state: Arc::new(Mutex::new(OpenAINonStreamingState::default())),
-        }
-    }
-
-    #[tokio::test]
-    async fn non_streaming_aggregates_content_only_when_no_reasoning() {
-        let transformer = non_streaming_transformer();
-
-        transformer
-            .transform(token_message(GeneratedTokenResult::ContentToken(
-                "hel".to_owned(),
-            )))
-            .await
-            .expect("the transformer must accept the message");
-        transformer
-            .transform(token_message(GeneratedTokenResult::ContentToken(
-                "lo".to_owned(),
-            )))
-            .await
-            .expect("the transformer must accept the message");
-
-        let summary = summary_with_counts(4, 2, 0);
-        let final_chunks = transformer
-            .transform(token_message(GeneratedTokenResult::Done(summary)))
-            .await
-            .expect("the transformer must accept the message");
-
-        assert_eq!(
-            final_chunks,
-            vec![
-                TransformResult::Chunk(r#"{"id":"test-request","object":"chat.completion","created":0,"model":"test-model","choices":[{"index":0,"message":{"role":"assistant","content":"hello","refusal":null,"annotations":[]},"logprobs":null,"finish_reason":"stop"}],"usage":{"prompt_tokens":4,"completion_tokens":2,"total_tokens":6,"prompt_tokens_details":{"cached_tokens":0,"audio_tokens":0},"completion_tokens_details":{"reasoning_tokens":0}},"service_tier":"default"}"#.to_owned()),
-            ]
-        );
-    }
-
-    #[tokio::test]
-    async fn non_streaming_drops_reasoning_but_keeps_reasoning_token_count() {
-        let transformer = non_streaming_transformer();
-
-        transformer
-            .transform(token_message(GeneratedTokenResult::ReasoningToken(
-                "think".to_owned(),
-            )))
-            .await
-            .expect("the transformer must accept the message");
-        transformer
-            .transform(token_message(GeneratedTokenResult::ContentToken(
-                "answer".to_owned(),
-            )))
-            .await
-            .expect("the transformer must accept the message");
-
-        let summary = summary_with_counts(3, 1, 1);
-        let final_chunks = transformer
-            .transform(token_message(GeneratedTokenResult::Done(summary)))
-            .await
-            .expect("the transformer must accept the message");
-
-        assert_eq!(
-            final_chunks,
-            vec![
-                TransformResult::Chunk(r#"{"id":"test-request","object":"chat.completion","created":0,"model":"test-model","choices":[{"index":0,"message":{"role":"assistant","content":"answer","refusal":null,"annotations":[]},"logprobs":null,"finish_reason":"stop"}],"usage":{"prompt_tokens":3,"completion_tokens":2,"total_tokens":5,"prompt_tokens_details":{"cached_tokens":0,"audio_tokens":0},"completion_tokens_details":{"reasoning_tokens":1}},"service_tier":"default"}"#.to_owned()),
-            ]
-        );
-    }
-
-    #[tokio::test]
-    async fn non_streaming_undeterminable_routes_to_content() {
-        let transformer = non_streaming_transformer();
-
-        transformer
-            .transform(token_message(GeneratedTokenResult::UndeterminableToken(
-                "amb".to_owned(),
-            )))
-            .await
-            .expect("the transformer must accept the message");
-
-        let summary = summary_with_counts(2, 0, 0);
-        let final_chunks = transformer
-            .transform(token_message(GeneratedTokenResult::Done(summary)))
-            .await
-            .expect("the transformer must accept the message");
-
-        assert_eq!(
-            final_chunks,
-            vec![
-                TransformResult::Chunk(r#"{"id":"test-request","object":"chat.completion","created":0,"model":"test-model","choices":[{"index":0,"message":{"role":"assistant","content":"amb","refusal":null,"annotations":[]},"logprobs":null,"finish_reason":"stop"}],"usage":{"prompt_tokens":2,"completion_tokens":0,"total_tokens":2,"prompt_tokens_details":{"cached_tokens":0,"audio_tokens":0},"completion_tokens_details":{"reasoning_tokens":0}},"service_tier":"default"}"#.to_owned()),
-            ]
-        );
-    }
-
-    #[tokio::test]
-    async fn non_streaming_tool_call_parsed_populates_message_tool_calls() {
-        let transformer = non_streaming_transformer();
-
-        transformer
-            .transform(token_message(GeneratedTokenResult::ToolCallParsed(vec![
-                weather_call(),
-            ])))
-            .await
-            .expect("the transformer must accept the message");
-
-        let summary = summary_with_counts(4, 0, 0);
-        let final_chunks = transformer
-            .transform(token_message(GeneratedTokenResult::Done(summary)))
-            .await
-            .expect("the transformer must accept the message");
-
-        assert_eq!(
-            final_chunks,
-            vec![
-                TransformResult::Chunk(r#"{"id":"test-request","object":"chat.completion","created":0,"model":"test-model","choices":[{"index":0,"message":{"role":"assistant","content":null,"refusal":null,"annotations":[],"tool_calls":[{"id":"call_x","type":"function","function":{"name":"get_weather","arguments":"{\"location\":\"Paris\"}"}}]},"logprobs":null,"finish_reason":"tool_calls"}],"usage":{"prompt_tokens":4,"completion_tokens":0,"total_tokens":4,"prompt_tokens_details":{"cached_tokens":0,"audio_tokens":0},"completion_tokens_details":{"reasoning_tokens":0}},"service_tier":"default"}"#.to_owned()),
-            ]
-        );
-    }
-
-    #[tokio::test]
-    async fn non_streaming_tool_call_parse_failed_emits_server_error() {
-        let transformer = non_streaming_transformer();
-
-        let chunks = transformer
-            .transform(token_message(GeneratedTokenResult::ToolCallParseFailed(
-                "bad payload".to_owned(),
-            )))
-            .await
-            .expect("the transformer must accept the message");
-
-        assert_eq!(
-            chunks,
-            vec![
-                TransformResult::Error(r#"{"error":{"message":"bad payload","type":"server_error","param":null,"code":null}}"#.to_owned()),
-            ]
-        );
-    }
-
-    #[tokio::test]
-    async fn non_streaming_tool_call_validation_failed_emits_server_error() {
-        let transformer = non_streaming_transformer();
-
-        let chunks = transformer
-            .transform(token_message(
-                GeneratedTokenResult::ToolCallValidationFailed(vec!["bad shape".to_owned()]),
-            ))
-            .await
-            .expect("the transformer must accept the message");
-
-        assert_eq!(
-            chunks,
-            vec![
-                TransformResult::Error(r#"{"error":{"message":"bad shape","type":"server_error","param":null,"code":null}}"#.to_owned()),
-            ]
-        );
-    }
-
-    #[tokio::test]
-    async fn non_streaming_unrecognized_tool_call_format_emits_server_error() {
-        let transformer = non_streaming_transformer();
-
-        let chunks = transformer
-            .transform(token_message(
-                GeneratedTokenResult::UnrecognizedToolCallFormat(RawToolCallTokens {
-                    text: "<unknown_marker>blah</unknown_marker>".to_owned(),
-                    ffi_error_message: "common_chat_parse failed: no parser".to_owned(),
-                }),
-            ))
-            .await
-            .expect("the transformer must accept the message");
-
-        assert_eq!(
-            chunks,
-            vec![
-                TransformResult::Error(r#"{"error":{"message":"model produced output the parser did not recognise as any registered tool-call format; FFI error: common_chat_parse failed: no parser; raw text: <unknown_marker>blah</unknown_marker>","type":"server_error","param":null,"code":null}}"#.to_owned()),
-            ]
-        );
-    }
-
-    #[tokio::test]
-    async fn non_streaming_error_message_returns_server_error() {
-        let transformer = non_streaming_transformer();
-
-        let chunks = transformer
-            .transform(error_message(500, "internal server error"))
-            .await
-            .expect("the transformer must accept the message");
-
-        assert_eq!(
-            chunks,
-            vec![
-                TransformResult::Error(r#"{"error":{"message":"internal server error","type":"server_error","param":null,"code":null}}"#.to_owned()),
-            ]
-        );
-    }
-
-    #[tokio::test]
-    async fn non_streaming_chat_template_error_returns_server_error() {
-        let transformer = non_streaming_transformer();
-
-        let message = token_message(GeneratedTokenResult::ChatTemplateError(
-            "bad template".to_owned(),
-        ));
-        let chunks = transformer
-            .transform(message)
-            .await
-            .expect("the transformer must accept the message");
-
-        assert_eq!(
-            chunks,
-            vec![
-                TransformResult::Error(r#"{"error":{"message":"bad template","type":"server_error","param":null,"code":null}}"#.to_owned()),
-            ]
-        );
-    }
-
-    #[tokio::test]
-    async fn non_streaming_image_decoding_failed_returns_invalid_request_error() {
-        let transformer = non_streaming_transformer();
-
-        let message = token_message(GeneratedTokenResult::ImageDecodingFailed(
-            "unsupported format".to_owned(),
-        ));
-        let chunks = transformer
-            .transform(message)
-            .await
-            .expect("the transformer must accept the message");
-
-        assert_eq!(
-            chunks,
-            vec![
-                TransformResult::Error(r#"{"error":{"message":"unsupported format","type":"invalid_request_error","param":null,"code":null}}"#.to_owned()),
-            ]
-        );
-    }
-
-    #[tokio::test]
-    async fn non_streaming_multimodal_not_supported_returns_invalid_request_error() {
-        let transformer = non_streaming_transformer();
-
-        let message = token_message(GeneratedTokenResult::MultimodalNotSupported(
-            "model does not support images".to_owned(),
-        ));
-        let chunks = transformer
-            .transform(message)
-            .await
-            .expect("the transformer must accept the message");
-
-        assert_eq!(
-            chunks,
-            vec![
-                TransformResult::Error(r#"{"error":{"message":"model does not support images","type":"invalid_request_error","param":null,"code":null}}"#.to_owned()),
-            ]
-        );
-    }
-
-    #[tokio::test]
-    async fn non_streaming_media_exceeding_the_micro_batch_returns_invalid_request_error() {
-        let transformer = non_streaming_transformer();
-
-        let message = token_message(GeneratedTokenResult::MediaExceedsMicroBatch(
-            OversizedMediaDetails {
-                media_tokens: 256,
-                micro_batch_tokens: 128,
-            },
-        ));
-        let chunks = transformer
-            .transform(message)
-            .await
-            .expect("the transformer must accept the message");
-
-        assert_eq!(
-            chunks,
-            vec![
-                TransformResult::Error(r#"{"error":{"message":"media required 256 tokens but one agent micro batch holds 128 tokens","type":"invalid_request_error","param":null,"code":null}}"#.to_owned()),
-            ]
-        );
-    }
-
-    #[tokio::test]
-    async fn non_streaming_tool_call_with_invalid_json_arguments_passes_raw_string_through() {
-        let transformer = non_streaming_transformer();
-
-        transformer
-            .transform(token_message(GeneratedTokenResult::ToolCallParsed(vec![
-                invalid_json_call(),
-            ])))
-            .await
-            .unwrap();
-
-        let final_chunks = transformer
-            .transform(token_message(GeneratedTokenResult::Done(
-                summary_with_counts(3, 0, 0),
-            )))
-            .await
-            .unwrap();
-
-        assert_eq!(
-            final_chunks,
-            vec![
-                TransformResult::Chunk(r#"{"id":"test-request","object":"chat.completion","created":0,"model":"test-model","choices":[{"index":0,"message":{"role":"assistant","content":null,"refusal":null,"annotations":[],"tool_calls":[{"id":"call_invalid","type":"function","function":{"name":"broken_tool","arguments":"{not valid json"}}]},"logprobs":null,"finish_reason":"tool_calls"}],"usage":{"prompt_tokens":3,"completion_tokens":0,"total_tokens":3,"prompt_tokens_details":{"cached_tokens":0,"audio_tokens":0},"completion_tokens_details":{"reasoning_tokens":0}},"service_tier":"default"}"#.to_owned()),
-            ]
-        );
-    }
-
-    #[tokio::test]
-    async fn non_streaming_embedding_response_returns_invalid_request_error() {
-        let transformer = non_streaming_transformer();
-
-        let message = response_message(OutgoingResponse::Embedding(EmbeddingResult::Done));
-        let chunks = transformer.transform(message).await.unwrap();
-
-        assert_eq!(
-            chunks,
-            vec![
-                TransformResult::Error(r#"{"error":{"message":"unexpected embedding response to a token generation request","type":"invalid_request_error","param":null,"code":null}}"#.to_owned()),
-            ]
-        );
-    }
-
-    #[tokio::test]
-    async fn rejects_inference_socket_notifications() {
-        let transform_result = non_streaming_transformer()
-            .transform(OutgoingMessage::Notification(
-                Notification::TokenGenerationEnabled,
-            ))
-            .await;
-
-        assert!(matches!(
-            transform_result,
-            Err(AgentRelayError::MessageNotRelayable { message })
-                if matches!(*message, OutgoingMessage::Notification(Notification::TokenGenerationEnabled))
-        ));
-    }
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/openai_non_streaming_state.rs b/paddler_balancer/src/compatibility/openai_service/openai_non_streaming_state.rs
deleted file mode 100644
index 7ee5416a5..000000000
--- a/paddler_balancer/src/compatibility/openai_service/openai_non_streaming_state.rs
+++ /dev/null
@@ -1,7 +0,0 @@
-use llama_cpp_bindings_types::ParsedToolCall;
-
-#[derive(Clone, Default)]
-pub struct OpenAINonStreamingState {
-    pub content: String,
-    pub tool_calls: Vec<ParsedToolCall>,
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/openai_reasoning_effort.rs b/paddler_balancer/src/compatibility/openai_service/openai_reasoning_effort.rs
deleted file mode 100644
index 06f96bf60..000000000
--- a/paddler_balancer/src/compatibility/openai_service/openai_reasoning_effort.rs
+++ /dev/null
@@ -1,32 +0,0 @@
-use serde::Deserialize;
-
-#[derive(Clone, Copy, Debug, Deserialize, Eq, PartialEq)]
-#[serde(rename_all = "lowercase")]
-pub enum OpenAIReasoningEffort {
-    None,
-    Minimal,
-    Low,
-    Medium,
-    High,
-    Xhigh,
-}
-
-impl OpenAIReasoningEffort {
-    #[must_use]
-    pub const fn enables_thinking(self) -> bool {
-        match self {
-            Self::None => false,
-            Self::Minimal | Self::Low | Self::Medium | Self::High | Self::Xhigh => true,
-        }
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use super::OpenAIReasoningEffort;
-
-    #[test]
-    fn minimal_effort_still_enables_thinking() {
-        assert!(OpenAIReasoningEffort::Minimal.enables_thinking());
-    }
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/openai_responses_function_call_item.rs b/paddler_balancer/src/compatibility/openai_service/openai_responses_function_call_item.rs
deleted file mode 100644
index 69c8004ab..000000000
--- a/paddler_balancer/src/compatibility/openai_service/openai_responses_function_call_item.rs
+++ /dev/null
@@ -1,8 +0,0 @@
-use serde::Deserialize;
-
-#[derive(Deserialize)]
-pub struct OpenAIResponsesFunctionCallItem {
-    pub call_id: String,
-    pub name: String,
-    pub arguments: String,
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/openai_responses_function_call_output_item.rs b/paddler_balancer/src/compatibility/openai_service/openai_responses_function_call_output_item.rs
deleted file mode 100644
index 2b38ae883..000000000
--- a/paddler_balancer/src/compatibility/openai_service/openai_responses_function_call_output_item.rs
+++ /dev/null
@@ -1,8 +0,0 @@
-use serde::Deserialize;
-
-use crate::compatibility::openai_service::openai_responses_function_output::OpenAIResponsesFunctionOutput;
-
-#[derive(Deserialize)]
-pub struct OpenAIResponsesFunctionCallOutputItem {
-    pub output: OpenAIResponsesFunctionOutput,
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/openai_responses_function_output.rs b/paddler_balancer/src/compatibility/openai_service/openai_responses_function_output.rs
deleted file mode 100644
index af5c60b22..000000000
--- a/paddler_balancer/src/compatibility/openai_service/openai_responses_function_output.rs
+++ /dev/null
@@ -1,51 +0,0 @@
-use serde::Deserialize;
-
-use crate::compatibility::openai_service::openai_responses_function_output_part::OpenAIResponsesFunctionOutputPart;
-
-#[derive(Deserialize)]
-#[serde(untagged)]
-pub enum OpenAIResponsesFunctionOutput {
-    Text(String),
-    Parts(Vec<OpenAIResponsesFunctionOutputPart>),
-}
-
-impl OpenAIResponsesFunctionOutput {
-    #[must_use]
-    pub fn into_text(self) -> String {
-        match self {
-            Self::Text(text) => text,
-            Self::Parts(parts) => parts
-                .into_iter()
-                .map(|OpenAIResponsesFunctionOutputPart::InputText { text }| text)
-                .collect::<String>(),
-        }
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use super::OpenAIResponsesFunctionOutput;
-    use crate::compatibility::openai_service::openai_responses_function_output_part::OpenAIResponsesFunctionOutputPart;
-
-    #[test]
-    fn text_output_returns_its_text() {
-        assert_eq!(
-            OpenAIResponsesFunctionOutput::Text("done".to_owned()).into_text(),
-            "done"
-        );
-    }
-
-    #[test]
-    fn parts_output_concatenates_its_text() {
-        let output = OpenAIResponsesFunctionOutput::Parts(vec![
-            OpenAIResponsesFunctionOutputPart::InputText {
-                text: "foo".to_owned(),
-            },
-            OpenAIResponsesFunctionOutputPart::InputText {
-                text: "bar".to_owned(),
-            },
-        ]);
-
-        assert_eq!(output.into_text(), "foobar");
-    }
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/openai_responses_function_output_part.rs b/paddler_balancer/src/compatibility/openai_service/openai_responses_function_output_part.rs
deleted file mode 100644
index e7f25ce12..000000000
--- a/paddler_balancer/src/compatibility/openai_service/openai_responses_function_output_part.rs
+++ /dev/null
@@ -1,8 +0,0 @@
-use serde::Deserialize;
-
-#[derive(Deserialize)]
-#[serde(tag = "type")]
-pub enum OpenAIResponsesFunctionOutputPart {
-    #[serde(rename = "input_text")]
-    InputText { text: String },
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/openai_responses_input.rs b/paddler_balancer/src/compatibility/openai_service/openai_responses_input.rs
deleted file mode 100644
index a5e259a17..000000000
--- a/paddler_balancer/src/compatibility/openai_service/openai_responses_input.rs
+++ /dev/null
@@ -1,34 +0,0 @@
-use serde::Deserialize;
-use serde::Deserializer;
-
-use crate::compatibility::openai_service::openai_responses_input_item::OpenAIResponsesInputItem;
-
-pub enum OpenAIResponsesInput {
-    Text(String),
-    Items(Vec<OpenAIResponsesInputItem>),
-}
-
-impl Default for OpenAIResponsesInput {
-    fn default() -> Self {
-        Self::Items(Vec::new())
-    }
-}
-
-impl<'deserializer> Deserialize<'deserializer> for OpenAIResponsesInput {
-    fn deserialize<TDeserializer>(deserializer: TDeserializer) -> Result<Self, TDeserializer::Error>
-    where
-        TDeserializer: Deserializer<'deserializer>,
-    {
-        #[derive(Deserialize)]
-        #[serde(untagged)]
-        enum TextOrItems {
-            Text(String),
-            Items(Vec<OpenAIResponsesInputItem>),
-        }
-
-        Ok(match TextOrItems::deserialize(deserializer)? {
-            TextOrItems::Text(text) => Self::Text(text),
-            TextOrItems::Items(items) => Self::Items(items),
-        })
-    }
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/openai_responses_input_content_part.rs b/paddler_balancer/src/compatibility/openai_service/openai_responses_input_content_part.rs
deleted file mode 100644
index 8e246193b..000000000
--- a/paddler_balancer/src/compatibility/openai_service/openai_responses_input_content_part.rs
+++ /dev/null
@@ -1,55 +0,0 @@
-use serde::Deserialize;
-
-use paddler_messaging::conversation_message_content_part::ConversationMessageContentPart;
-use paddler_messaging::image_url::ImageUrl;
-
-#[derive(Deserialize)]
-#[serde(tag = "type")]
-pub enum OpenAIResponsesInputContentPart {
-    #[serde(rename = "input_text")]
-    InputText { text: String },
-    #[serde(rename = "input_image")]
-    InputImage { image_url: String },
-}
-
-impl OpenAIResponsesInputContentPart {
-    #[must_use]
-    pub fn into_conversation_part(self) -> ConversationMessageContentPart {
-        match self {
-            Self::InputText { text } => ConversationMessageContentPart::Text { text },
-            Self::InputImage { image_url } => ConversationMessageContentPart::ImageUrl {
-                image_url: ImageUrl { url: image_url },
-            },
-        }
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use paddler_messaging::conversation_message_content_part::ConversationMessageContentPart;
-
-    use super::OpenAIResponsesInputContentPart;
-
-    #[test]
-    fn input_text_becomes_text_part() {
-        assert!(matches!(
-            OpenAIResponsesInputContentPart::InputText {
-                text: "hello".to_owned(),
-            }
-            .into_conversation_part(),
-            ConversationMessageContentPart::Text { text } if text == "hello"
-        ));
-    }
-
-    #[test]
-    fn input_image_becomes_image_part() {
-        assert!(matches!(
-            OpenAIResponsesInputContentPart::InputImage {
-                image_url: "https://example.test/cat.png".to_owned(),
-            }
-            .into_conversation_part(),
-            ConversationMessageContentPart::ImageUrl { image_url }
-                if image_url.url == "https://example.test/cat.png"
-        ));
-    }
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/openai_responses_input_item.rs b/paddler_balancer/src/compatibility/openai_service/openai_responses_input_item.rs
deleted file mode 100644
index 66a55be7e..000000000
--- a/paddler_balancer/src/compatibility/openai_service/openai_responses_input_item.rs
+++ /dev/null
@@ -1,47 +0,0 @@
-use serde::Deserialize;
-use serde_json::json;
-
-use paddler_messaging::conversation_message::ConversationMessage;
-use paddler_messaging::conversation_message_content::ConversationMessageContent;
-
-use crate::compatibility::openai_service::assistant_role::ASSISTANT_ROLE;
-use crate::compatibility::openai_service::openai_responses_function_call_item::OpenAIResponsesFunctionCallItem;
-use crate::compatibility::openai_service::openai_responses_function_call_output_item::OpenAIResponsesFunctionCallOutputItem;
-use crate::compatibility::openai_service::openai_responses_message_item::OpenAIResponsesMessageItem;
-use crate::compatibility::openai_service::openai_responses_tagged_item::OpenAIResponsesTaggedItem;
-
-#[derive(Deserialize)]
-#[serde(untagged)]
-pub enum OpenAIResponsesInputItem {
-    Tagged(OpenAIResponsesTaggedItem),
-    Message(OpenAIResponsesMessageItem),
-}
-
-impl OpenAIResponsesInputItem {
-    #[must_use]
-    pub fn into_conversation_message(self) -> ConversationMessage {
-        match self {
-            Self::Message(message) | Self::Tagged(OpenAIResponsesTaggedItem::Message(message)) => {
-                message.into_conversation_message()
-            }
-            Self::Tagged(OpenAIResponsesTaggedItem::FunctionCall(
-                OpenAIResponsesFunctionCallItem {
-                    call_id,
-                    name,
-                    arguments,
-                },
-            )) => ConversationMessage {
-                content: ConversationMessageContent::Text(
-                    json!({ "call_id": call_id, "name": name, "arguments": arguments }).to_string(),
-                ),
-                role: ASSISTANT_ROLE.to_owned(),
-            },
-            Self::Tagged(OpenAIResponsesTaggedItem::FunctionCallOutput(
-                OpenAIResponsesFunctionCallOutputItem { output },
-            )) => ConversationMessage {
-                content: ConversationMessageContent::Text(output.into_text()),
-                role: "tool".to_owned(),
-            },
-        }
-    }
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/openai_responses_message_content.rs b/paddler_balancer/src/compatibility/openai_service/openai_responses_message_content.rs
deleted file mode 100644
index da3abdb7d..000000000
--- a/paddler_balancer/src/compatibility/openai_service/openai_responses_message_content.rs
+++ /dev/null
@@ -1,27 +0,0 @@
-use serde::Deserialize;
-
-use paddler_messaging::conversation_message_content::ConversationMessageContent;
-
-use crate::compatibility::openai_service::openai_responses_input_content_part::OpenAIResponsesInputContentPart;
-
-#[derive(Deserialize)]
-#[serde(untagged)]
-pub enum OpenAIResponsesMessageContent {
-    Text(String),
-    Parts(Vec<OpenAIResponsesInputContentPart>),
-}
-
-impl OpenAIResponsesMessageContent {
-    #[must_use]
-    pub fn into_conversation_content(self) -> ConversationMessageContent {
-        match self {
-            Self::Text(text) => ConversationMessageContent::Text(text),
-            Self::Parts(parts) => ConversationMessageContent::Parts(
-                parts
-                    .into_iter()
-                    .map(OpenAIResponsesInputContentPart::into_conversation_part)
-                    .collect(),
-            ),
-        }
-    }
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/openai_responses_message_item.rs b/paddler_balancer/src/compatibility/openai_service/openai_responses_message_item.rs
deleted file mode 100644
index 09ec37c03..000000000
--- a/paddler_balancer/src/compatibility/openai_service/openai_responses_message_item.rs
+++ /dev/null
@@ -1,29 +0,0 @@
-use serde::Deserialize;
-
-use paddler_messaging::conversation_message::ConversationMessage;
-
-use crate::compatibility::openai_service::openai_responses_message_content::OpenAIResponsesMessageContent;
-
-fn normalize_role(role: String) -> String {
-    if role == "developer" {
-        "system".to_owned()
-    } else {
-        role
-    }
-}
-
-#[derive(Deserialize)]
-pub struct OpenAIResponsesMessageItem {
-    pub role: String,
-    pub content: OpenAIResponsesMessageContent,
-}
-
-impl OpenAIResponsesMessageItem {
-    #[must_use]
-    pub fn into_conversation_message(self) -> ConversationMessage {
-        ConversationMessage {
-            content: self.content.into_conversation_content(),
-            role: normalize_role(self.role),
-        }
-    }
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/openai_responses_reasoning.rs b/paddler_balancer/src/compatibility/openai_service/openai_responses_reasoning.rs
deleted file mode 100644
index 5ac61a423..000000000
--- a/paddler_balancer/src/compatibility/openai_service/openai_responses_reasoning.rs
+++ /dev/null
@@ -1,16 +0,0 @@
-use serde::Deserialize;
-
-use crate::compatibility::openai_service::openai_reasoning_effort::OpenAIReasoningEffort;
-
-#[derive(Deserialize)]
-pub struct OpenAIResponsesReasoning {
-    pub effort: Option<OpenAIReasoningEffort>,
-}
-
-impl OpenAIResponsesReasoning {
-    #[must_use]
-    pub fn enables_thinking(&self) -> bool {
-        self.effort
-            .is_none_or(OpenAIReasoningEffort::enables_thinking)
-    }
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/openai_responses_request_params.rs b/paddler_balancer/src/compatibility/openai_service/openai_responses_request_params.rs
deleted file mode 100644
index 6f575cf79..000000000
--- a/paddler_balancer/src/compatibility/openai_service/openai_responses_request_params.rs
+++ /dev/null
@@ -1,390 +0,0 @@
-use std::num::NonZeroU32;
-
-use anyhow::Result;
-use serde::Deserialize;
-
-use paddler_messaging::conversation_history::ConversationHistory;
-use paddler_messaging::conversation_message::ConversationMessage;
-use paddler_messaging::conversation_message_content::ConversationMessageContent;
-use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
-use paddler_messaging::validates::Validates;
-
-use crate::compatibility::openai_service::openai_default_max_tokens::OPENAI_DEFAULT_MAX_TOKENS;
-use crate::compatibility::openai_service::openai_responses_input::OpenAIResponsesInput;
-use crate::compatibility::openai_service::openai_responses_input_item::OpenAIResponsesInputItem;
-use crate::compatibility::openai_service::openai_responses_reasoning::OpenAIResponsesReasoning;
-use crate::compatibility::openai_service::openai_responses_text_param::OpenAIResponsesTextParam;
-use crate::compatibility::openai_service::openai_responses_tool::OpenAIResponsesTool;
-use crate::compatibility::openai_service::responses_prepared_request::ResponsesPreparedRequest;
-
-#[derive(Deserialize)]
-pub struct OpenAIResponsesRequestParams {
-    pub model: String,
-    #[serde(default)]
-    pub input: OpenAIResponsesInput,
-    #[serde(default)]
-    pub instructions: Option<String>,
-    #[serde(default)]
-    pub stream: Option<bool>,
-    #[serde(default)]
-    pub max_output_tokens: Option<NonZeroU32>,
-    #[serde(default)]
-    pub tools: Vec<OpenAIResponsesTool>,
-    #[serde(default)]
-    pub text: Option<OpenAIResponsesTextParam>,
-    #[serde(default)]
-    pub reasoning: Option<OpenAIResponsesReasoning>,
-}
-
-impl OpenAIResponsesRequestParams {
-    pub fn into_prepared(self) -> Result<ResponsesPreparedRequest> {
-        let Self {
-            model,
-            input,
-            instructions,
-            stream,
-            max_output_tokens,
-            tools,
-            text,
-            reasoning,
-        } = self;
-
-        let mut messages: Vec<ConversationMessage> = Vec::new();
-
-        if let Some(instructions) = &instructions
-            && !instructions.is_empty()
-        {
-            messages.push(ConversationMessage {
-                content: ConversationMessageContent::Text(instructions.clone()),
-                role: "system".to_owned(),
-            });
-        }
-
-        match input {
-            OpenAIResponsesInput::Text(text) => messages.push(ConversationMessage {
-                content: ConversationMessageContent::Text(text),
-                role: "user".to_owned(),
-            }),
-            OpenAIResponsesInput::Items(items) => {
-                messages.extend(
-                    items
-                        .into_iter()
-                        .map(OpenAIResponsesInputItem::into_conversation_message),
-                );
-            }
-        }
-
-        let validated_tools = tools
-            .into_iter()
-            .map(|OpenAIResponsesTool::Function(function_definition)| {
-                function_definition.into_tool()
-            })
-            .map(Validates::validate)
-            .collect::<Result<Vec<_>, _>>()?;
-
-        let parse_tool_calls = !validated_tools.is_empty();
-
-        Ok(ResponsesPreparedRequest {
-            paddler_params: ContinueFromConversationHistoryParams {
-                add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(messages),
-                enable_thinking: reasoning
-                    .as_ref()
-                    .is_none_or(OpenAIResponsesReasoning::enables_thinking),
-                grammar: text.and_then(OpenAIResponsesTextParam::into_grammar_constraint),
-                max_tokens: max_output_tokens.unwrap_or(OPENAI_DEFAULT_MAX_TOKENS),
-                parse_tool_calls,
-                tools: validated_tools,
-            },
-            stream: stream.unwrap_or(false),
-            model,
-            instructions,
-        })
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use std::num::NonZeroU32;
-
-    use serde_json::Value;
-    use serde_json::error::Category;
-    use serde_json::from_str;
-    use serde_json::from_value;
-    use serde_json::json;
-
-    use paddler_messaging::conversation_message_content::ConversationMessageContent;
-    use paddler_messaging::conversation_message_content_part::ConversationMessageContentPart;
-    use paddler_messaging::grammar_constraint::GrammarConstraint;
-    use paddler_messaging::image_url::ImageUrl;
-    use paddler_messaging::request_params::continue_from_conversation_history_params::tool::Tool;
-    use paddler_messaging::request_params_validation_error::RequestParamsValidationError;
-
-    use super::OpenAIResponsesRequestParams;
-    use crate::compatibility::openai_service::responses_prepared_request::ResponsesPreparedRequest;
-
-    fn prepared_from(value: Value) -> ResponsesPreparedRequest {
-        let params: OpenAIResponsesRequestParams = from_value(value).unwrap();
-
-        params.into_prepared().unwrap()
-    }
-
-    #[test]
-    fn string_input_becomes_a_single_user_message() {
-        let prepared = prepared_from(json!({ "model": "test", "input": "Say hello" }));
-
-        let messages = &prepared.paddler_params.conversation_history.messages;
-
-        assert_eq!(messages.len(), 1);
-        assert_eq!(messages[0].role, "user");
-        assert_eq!(messages[0].content.text_content(), "Say hello");
-    }
-
-    #[test]
-    fn instructions_are_prepended_as_a_system_message() {
-        let prepared = prepared_from(json!({
-            "model": "test",
-            "instructions": "be terse",
-            "input": "hi"
-        }));
-
-        let messages = &prepared.paddler_params.conversation_history.messages;
-
-        assert_eq!(messages[0].role, "system");
-        assert_eq!(messages[0].content.text_content(), "be terse");
-        assert_eq!(messages[1].role, "user");
-    }
-
-    #[test]
-    fn function_call_output_item_becomes_a_tool_message() {
-        let prepared = prepared_from(json!({
-            "model": "test",
-            "input": [
-                { "type": "function_call_output", "call_id": "call_1", "output": "sunny" }
-            ]
-        }));
-
-        let messages = &prepared.paddler_params.conversation_history.messages;
-
-        assert_eq!(messages[0].role, "tool");
-        assert_eq!(messages[0].content.text_content(), "sunny");
-    }
-
-    #[test]
-    fn function_call_item_becomes_an_assistant_message_carrying_the_call() {
-        let prepared = prepared_from(json!({
-            "model": "test",
-            "input": [
-                { "type": "function_call", "call_id": "call_1", "name": "get_weather", "arguments": "{}" }
-            ]
-        }));
-
-        let messages = &prepared.paddler_params.conversation_history.messages;
-
-        assert_eq!(messages[0].role, "assistant");
-        assert_eq!(
-            from_str::<Value>(&messages[0].content.text_content()).unwrap(),
-            json!({ "call_id": "call_1", "name": "get_weather", "arguments": "{}" })
-        );
-    }
-
-    #[test]
-    fn message_content_parts_become_conversation_parts() {
-        let prepared = prepared_from(json!({
-            "model": "test",
-            "input": [
-                {
-                    "type": "message",
-                    "role": "user",
-                    "content": [
-                        { "type": "input_text", "text": "What is this?" },
-                        { "type": "input_image", "image_url": "https://example.test/cat.png" }
-                    ]
-                }
-            ]
-        }));
-
-        assert_eq!(
-            prepared.paddler_params.conversation_history.messages[0].content,
-            ConversationMessageContent::Parts(vec![
-                ConversationMessageContentPart::Text {
-                    text: "What is this?".to_owned(),
-                },
-                ConversationMessageContentPart::ImageUrl {
-                    image_url: ImageUrl {
-                        url: "https://example.test/cat.png".to_owned(),
-                    },
-                },
-            ])
-        );
-    }
-
-    #[test]
-    fn text_format_text_leaves_the_output_unconstrained() {
-        let prepared = prepared_from(json!({
-            "model": "test",
-            "input": "hi",
-            "text": { "format": { "type": "text" } }
-        }));
-
-        assert_eq!(prepared.paddler_params.grammar, None);
-    }
-
-    #[test]
-    fn developer_role_is_normalized_to_system() {
-        let prepared = prepared_from(json!({
-            "model": "test",
-            "input": [
-                { "type": "message", "role": "developer", "content": "rules" }
-            ]
-        }));
-
-        assert_eq!(
-            prepared.paddler_params.conversation_history.messages[0].role,
-            "system"
-        );
-    }
-
-    #[test]
-    fn flat_function_tool_maps_to_an_internal_tool_with_default_description() {
-        let prepared = prepared_from(json!({
-            "model": "test",
-            "input": "hi",
-            "tools": [
-                { "type": "function", "name": "get_weather", "parameters": { "type": "object" } }
-            ]
-        }));
-
-        assert!(prepared.paddler_params.parse_tool_calls);
-
-        let Tool::Function(function_call) = &prepared.paddler_params.tools[0];
-
-        assert_eq!(function_call.function.name, "get_weather");
-        assert_eq!(function_call.function.description, "");
-    }
-
-    #[test]
-    fn text_format_json_schema_becomes_a_grammar_constraint() {
-        let prepared = prepared_from(json!({
-            "model": "test",
-            "input": "hi",
-            "text": { "format": { "type": "json_schema", "name": "out", "schema": { "type": "object" } } }
-        }));
-
-        assert!(matches!(
-            &prepared.paddler_params.grammar,
-            Some(GrammarConstraint::JsonSchema { schema }) if schema == r#"{"type":"object"}"#
-        ));
-    }
-
-    #[test]
-    fn a_request_without_an_output_limit_generates_until_its_context_is_full() {
-        let prepared = prepared_from(json!({ "model": "test", "input": "hi" }));
-
-        assert_eq!(prepared.paddler_params.max_tokens, NonZeroU32::MAX);
-    }
-
-    #[test]
-    fn reasoning_effort_none_disables_thinking() {
-        let prepared = prepared_from(json!({
-            "model": "test",
-            "input": "hi",
-            "reasoning": { "effort": "none" }
-        }));
-
-        assert!(!prepared.paddler_params.enable_thinking);
-    }
-
-    #[test]
-    fn rejects_input_it_cannot_represent() {
-        let unrepresentable_requests = [
-            json!({ "model": "test", "input": "hi", "tools": [ { "type": "web_search" } ] }),
-            json!({ "model": "test", "input": [ { "type": "reasoning", "summary": [] } ] }),
-            json!({ "model": "test", "input": [ { "type": "message", "role": "user", "content": [ { "type": "input_file", "file_id": "file_1" } ] } ] }),
-            json!({ "model": "test", "input": [ { "type": "message", "role": "user", "content": [ { "type": "input_image", "file_id": "file_1" } ] } ] }),
-            json!({ "model": "test", "input": [ { "type": "function_call_output", "call_id": "call_1", "output": [ { "type": "input_image", "image_url": "https://example.test/cat.png" } ] } ] }),
-            json!({ "model": "test", "input": "hi", "text": { "format": { "type": "json_object" } } }),
-        ];
-
-        for unrepresentable_request in unrepresentable_requests {
-            let rejection = from_value::<OpenAIResponsesRequestParams>(unrepresentable_request)
-                .err()
-                .map(|deserialization_error| deserialization_error.classify());
-
-            assert_eq!(rejection, Some(Category::Data));
-        }
-    }
-
-    #[test]
-    fn unsupported_and_stateful_fields_are_ignored() {
-        let prepared = prepared_from(json!({
-            "model": "test",
-            "input": "hi",
-            "store": true,
-            "previous_response_id": "resp_prev",
-            "conversation": "conv_1",
-            "temperature": 0.5,
-            "tool_choice": "required"
-        }));
-
-        assert_eq!(
-            prepared.paddler_params.conversation_history.messages.len(),
-            1
-        );
-    }
-
-    #[test]
-    fn instructions_alone_become_the_whole_conversation() {
-        let prepared = prepared_from(json!({ "model": "test", "instructions": "be terse" }));
-
-        let messages = &prepared.paddler_params.conversation_history.messages;
-
-        assert_eq!(messages.len(), 1);
-        assert_eq!(messages[0].role, "system");
-        assert_eq!(messages[0].content.text_content(), "be terse");
-    }
-
-    #[test]
-    fn message_item_without_a_type_becomes_a_conversation_message() {
-        let prepared = prepared_from(json!({
-            "model": "test",
-            "input": [ { "role": "user", "content": "hi" } ]
-        }));
-
-        let messages = &prepared.paddler_params.conversation_history.messages;
-
-        assert_eq!(messages[0].role, "user");
-        assert_eq!(messages[0].content.text_content(), "hi");
-    }
-
-    #[test]
-    fn rejects_a_function_tool_that_requires_an_undeclared_property() {
-        let params: OpenAIResponsesRequestParams = from_value(json!({
-            "model": "test",
-            "input": "hi",
-            "tools": [
-                {
-                    "type": "function",
-                    "name": "broken",
-                    "parameters": {
-                        "type": "object",
-                        "properties": { "present": { "type": "string" } },
-                        "required": ["absent"]
-                    }
-                }
-            ]
-        }))
-        .unwrap();
-
-        let rejection = params.into_prepared().err().unwrap();
-
-        assert_eq!(
-            rejection.downcast_ref::<RequestParamsValidationError>(),
-            Some(
-                &RequestParamsValidationError::RequiredFieldNotInProperties {
-                    field: "absent".to_owned(),
-                }
-            )
-        );
-    }
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/openai_responses_tagged_item.rs b/paddler_balancer/src/compatibility/openai_service/openai_responses_tagged_item.rs
deleted file mode 100644
index e667b1d31..000000000
--- a/paddler_balancer/src/compatibility/openai_service/openai_responses_tagged_item.rs
+++ /dev/null
@@ -1,16 +0,0 @@
-use serde::Deserialize;
-
-use crate::compatibility::openai_service::openai_responses_function_call_item::OpenAIResponsesFunctionCallItem;
-use crate::compatibility::openai_service::openai_responses_function_call_output_item::OpenAIResponsesFunctionCallOutputItem;
-use crate::compatibility::openai_service::openai_responses_message_item::OpenAIResponsesMessageItem;
-
-#[derive(Deserialize)]
-#[serde(tag = "type")]
-pub enum OpenAIResponsesTaggedItem {
-    #[serde(rename = "message")]
-    Message(OpenAIResponsesMessageItem),
-    #[serde(rename = "function_call")]
-    FunctionCall(OpenAIResponsesFunctionCallItem),
-    #[serde(rename = "function_call_output")]
-    FunctionCallOutput(OpenAIResponsesFunctionCallOutputItem),
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/openai_responses_text_format.rs b/paddler_balancer/src/compatibility/openai_service/openai_responses_text_format.rs
deleted file mode 100644
index 10537f1a7..000000000
--- a/paddler_balancer/src/compatibility/openai_service/openai_responses_text_format.rs
+++ /dev/null
@@ -1,11 +0,0 @@
-use serde::Deserialize;
-use serde_json::Value;
-
-#[derive(Deserialize)]
-#[serde(tag = "type")]
-pub enum OpenAIResponsesTextFormat {
-    #[serde(rename = "text")]
-    Text,
-    #[serde(rename = "json_schema")]
-    JsonSchema { schema: Value },
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/openai_responses_text_param.rs b/paddler_balancer/src/compatibility/openai_service/openai_responses_text_param.rs
deleted file mode 100644
index 5c21f1fde..000000000
--- a/paddler_balancer/src/compatibility/openai_service/openai_responses_text_param.rs
+++ /dev/null
@@ -1,25 +0,0 @@
-use serde::Deserialize;
-
-use paddler_messaging::grammar_constraint::GrammarConstraint;
-
-use crate::compatibility::openai_service::openai_responses_text_format::OpenAIResponsesTextFormat;
-
-#[derive(Deserialize)]
-pub struct OpenAIResponsesTextParam {
-    #[serde(default)]
-    pub format: Option<OpenAIResponsesTextFormat>,
-}
-
-impl OpenAIResponsesTextParam {
-    #[must_use]
-    pub fn into_grammar_constraint(self) -> Option<GrammarConstraint> {
-        match self.format {
-            Some(OpenAIResponsesTextFormat::JsonSchema { schema }) => {
-                Some(GrammarConstraint::JsonSchema {
-                    schema: schema.to_string(),
-                })
-            }
-            Some(OpenAIResponsesTextFormat::Text) | None => None,
-        }
-    }
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/openai_responses_tool.rs b/paddler_balancer/src/compatibility/openai_service/openai_responses_tool.rs
deleted file mode 100644
index 02f00b0bc..000000000
--- a/paddler_balancer/src/compatibility/openai_service/openai_responses_tool.rs
+++ /dev/null
@@ -1,10 +0,0 @@
-use serde::Deserialize;
-
-use crate::compatibility::openai_service::openai_function_definition::OpenAIFunctionDefinition;
-
-#[derive(Deserialize)]
-#[serde(tag = "type")]
-pub enum OpenAIResponsesTool {
-    #[serde(rename = "function")]
-    Function(Box<OpenAIFunctionDefinition>),
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/openai_sse_http_response.rs b/paddler_balancer/src/compatibility/openai_service/openai_sse_http_response.rs
new file mode 100644
index 000000000..3b1c2b244
--- /dev/null
+++ b/paddler_balancer/src/compatibility/openai_service/openai_sse_http_response.rs
@@ -0,0 +1,15 @@
+use actix_web::HttpResponse;
+use actix_web::http::header;
+use actix_web_lab::sse;
+use futures::stream::Stream;
+use serde_json::Error as SerdeJsonError;
+
+pub fn openai_sse_http_response<TStream>(sse_events: TStream) -> HttpResponse
+where
+    TStream: Stream<Item = Result<sse::Event, SerdeJsonError>> + 'static,
+{
+    HttpResponse::Ok()
+        .content_type("text/event-stream")
+        .insert_header((header::CACHE_CONTROL, "no-cache"))
+        .body(sse::Sse::from_stream(sse_events))
+}
diff --git a/paddler_balancer/src/compatibility/openai_service/openai_streaming_response_transformer.rs b/paddler_balancer/src/compatibility/openai_service/openai_streaming_response_transformer.rs
deleted file mode 100644
index 671a21f5a..000000000
--- a/paddler_balancer/src/compatibility/openai_service/openai_streaming_response_transformer.rs
+++ /dev/null
@@ -1,770 +0,0 @@
-use std::sync::Arc;
-
-use async_trait::async_trait;
-use llama_cpp_bindings_types::ParsedToolCall;
-use parking_lot::Mutex;
-use serde_json::to_string;
-
-use paddler_messaging::generated_token_result::GeneratedTokenResult;
-use paddler_messaging::generation_summary::GenerationSummary;
-use paddler_messaging::inference_client::message::Message as OutgoingMessage;
-use paddler_messaging::inference_client::response::Response as OutgoingResponse;
-use paddler_messaging::jsonrpc::response_envelope::ResponseEnvelope;
-
-use crate::agent_relay_error::AgentRelayError;
-use crate::chunk_forwarding_session_controller::transform_result::TransformResult;
-use crate::chunk_forwarding_session_controller::transforms_outgoing_message::TransformsOutgoingMessage;
-use crate::compatibility::openai_service::chat_completion_chunk::ChatCompletionChunk;
-use crate::compatibility::openai_service::chat_completion_chunk_choice::ChatCompletionChunkChoice;
-use crate::compatibility::openai_service::chat_completion_chunk_payload::ChatCompletionChunkPayload;
-use crate::compatibility::openai_service::chat_completion_finish_reason::chat_completion_finish_reason;
-use crate::compatibility::openai_service::openai_streaming_state::OpenAIStreamingState;
-use crate::compatibility::openai_service::try_universal_error_chunk::try_universal_error_chunk;
-
-#[derive(Clone)]
-pub struct OpenAIStreamingResponseTransformer {
-    pub created: u64,
-    pub include_usage: bool,
-    pub model: String,
-    pub state: Arc<Mutex<OpenAIStreamingState>>,
-    pub system_fingerprint: String,
-}
-
-impl OpenAIStreamingResponseTransformer {
-    fn chunk(
-        &self,
-        request_id: &str,
-        payload: ChatCompletionChunkPayload<'_>,
-    ) -> Result<String, AgentRelayError> {
-        to_string(&ChatCompletionChunk {
-            created: self.created,
-            id: request_id,
-            model: &self.model,
-            payload,
-            system_fingerprint: &self.system_fingerprint,
-        })
-        .map_err(AgentRelayError::MessageUnserializable)
-    }
-
-    fn handle_content(
-        &self,
-        request_id: &str,
-        text: &str,
-    ) -> Result<Vec<TransformResult>, AgentRelayError> {
-        self.chunk(
-            request_id,
-            ChatCompletionChunkPayload::Choice(ChatCompletionChunkChoice::Content(text)),
-        )
-        .map(|chunk| vec![TransformResult::Chunk(chunk)])
-    }
-
-    fn handle_tool_call_parsed(
-        &self,
-        request_id: &str,
-        parsed_calls: &[ParsedToolCall],
-    ) -> Result<Vec<TransformResult>, AgentRelayError> {
-        if parsed_calls.is_empty() {
-            return Ok(vec![]);
-        }
-
-        self.state.lock().saw_tool_call = true;
-
-        self.chunk(
-            request_id,
-            ChatCompletionChunkPayload::Choice(ChatCompletionChunkChoice::ToolCalls(parsed_calls)),
-        )
-        .map(|chunk| vec![TransformResult::Chunk(chunk)])
-    }
-
-    fn handle_done(
-        &self,
-        request_id: &str,
-        summary: &GenerationSummary,
-    ) -> Result<Vec<TransformResult>, AgentRelayError> {
-        let saw_tool_call = self.state.lock().saw_tool_call;
-        let finish_reason = chat_completion_finish_reason(summary.finish, saw_tool_call);
-
-        self.chunk(
-            request_id,
-            ChatCompletionChunkPayload::Choice(ChatCompletionChunkChoice::Finish(finish_reason)),
-        )
-        .and_then(|finish_chunk| {
-            let finish = TransformResult::Chunk(finish_chunk);
-
-            if self.include_usage {
-                self.chunk(
-                    request_id,
-                    ChatCompletionChunkPayload::Usage(&summary.usage),
-                )
-                .map(|usage_chunk| vec![finish, TransformResult::Chunk(usage_chunk)])
-            } else {
-                Ok(vec![finish])
-            }
-        })
-    }
-}
-
-#[async_trait]
-impl TransformsOutgoingMessage for OpenAIStreamingResponseTransformer {
-    type Output = TransformResult;
-
-    async fn transform(
-        &self,
-        message: OutgoingMessage,
-    ) -> Result<Vec<TransformResult>, AgentRelayError> {
-        match message {
-            OutgoingMessage::Response(ResponseEnvelope {
-                request_id,
-                response:
-                    OutgoingResponse::GeneratedToken(
-                        GeneratedTokenResult::ContentToken(text)
-                        | GeneratedTokenResult::UndeterminableToken(text),
-                    ),
-                ..
-            }) => self.handle_content(&request_id, &text),
-            OutgoingMessage::Response(ResponseEnvelope {
-                response:
-                    OutgoingResponse::GeneratedToken(
-                        GeneratedTokenResult::ReasoningToken(_)
-                        | GeneratedTokenResult::ToolCallToken(_),
-                    ),
-                ..
-            }) => Ok(vec![]),
-            OutgoingMessage::Response(ResponseEnvelope {
-                request_id,
-                response:
-                    OutgoingResponse::GeneratedToken(GeneratedTokenResult::ToolCallParsed(parsed_calls)),
-                ..
-            }) => self.handle_tool_call_parsed(&request_id, &parsed_calls),
-            OutgoingMessage::Response(ResponseEnvelope {
-                request_id,
-                response: OutgoingResponse::GeneratedToken(GeneratedTokenResult::Done(summary)),
-                ..
-            }) => self.handle_done(&request_id, &summary),
-            other => try_universal_error_chunk(&other)
-                .map(|error_chunk| vec![error_chunk])
-                .ok_or_else(|| AgentRelayError::MessageNotRelayable {
-                    message: Box::new(other),
-                }),
-        }
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use std::sync::Arc;
-
-    use llama_cpp_bindings_types::ParsedToolCall;
-    use llama_cpp_bindings_types::TokenUsage;
-    use llama_cpp_bindings_types::ToolCallArguments;
-    use parking_lot::Mutex;
-    use serde_json::json;
-
-    use paddler_messaging::embedding_result::EmbeddingResult;
-    use paddler_messaging::generated_token_result::GeneratedTokenResult;
-    use paddler_messaging::generation_finish::GenerationFinish;
-    use paddler_messaging::generation_summary::GenerationSummary;
-    use paddler_messaging::inference_client::message::Message as OutgoingMessage;
-    use paddler_messaging::inference_client::notification::Notification;
-    use paddler_messaging::inference_client::response::Response as OutgoingResponse;
-    use paddler_messaging::jsonrpc::error::Error as JsonRpcError;
-    use paddler_messaging::jsonrpc::error_envelope::ErrorEnvelope;
-    use paddler_messaging::jsonrpc::response_envelope::ResponseEnvelope;
-    use paddler_messaging::oversized_media_details::OversizedMediaDetails;
-    use paddler_messaging::raw_tool_call_tokens::RawToolCallTokens;
-
-    use super::OpenAIStreamingResponseTransformer;
-    use super::OpenAIStreamingState;
-    use crate::agent_relay_error::AgentRelayError;
-    use crate::chunk_forwarding_session_controller::transform_result::TransformResult;
-    use crate::chunk_forwarding_session_controller::transforms_outgoing_message::TransformsOutgoingMessage;
-
-    #[must_use]
-    pub fn token_message(token_result: GeneratedTokenResult) -> OutgoingMessage {
-        OutgoingMessage::Response(ResponseEnvelope {
-            generated_by: None,
-            request_id: "test-request".to_owned(),
-            response: OutgoingResponse::GeneratedToken(token_result),
-        })
-    }
-
-    #[must_use]
-    pub fn error_message(code: i32, description: &str) -> OutgoingMessage {
-        OutgoingMessage::Error(ErrorEnvelope {
-            request_id: "test-request".to_owned(),
-            error: JsonRpcError {
-                code,
-                description: description.to_owned(),
-            },
-        })
-    }
-
-    #[must_use]
-    pub fn response_message(response: OutgoingResponse) -> OutgoingMessage {
-        OutgoingMessage::Response(ResponseEnvelope {
-            generated_by: None,
-            request_id: "test-request".to_owned(),
-            response,
-        })
-    }
-
-    #[must_use]
-    pub fn summary_with_counts(
-        prompt_tokens: u64,
-        content_tokens: u64,
-        reasoning_tokens: u64,
-    ) -> GenerationSummary {
-        GenerationSummary {
-            finish: GenerationFinish::EndOfGeneration,
-            usage: TokenUsage {
-                prompt_tokens,
-                content_tokens,
-                reasoning_tokens,
-                ..TokenUsage::default()
-            },
-        }
-    }
-
-    #[must_use]
-    pub fn weather_call() -> ParsedToolCall {
-        ParsedToolCall::new(
-            "call_x".to_owned(),
-            "get_weather".to_owned(),
-            ToolCallArguments::ValidJson(json!({ "location": "Paris" })),
-        )
-    }
-
-    #[must_use]
-    pub fn invalid_json_call() -> ParsedToolCall {
-        ParsedToolCall::new(
-            "call_invalid".to_owned(),
-            "broken_tool".to_owned(),
-            ToolCallArguments::InvalidJson("{not valid json".to_owned()),
-        )
-    }
-
-    fn streaming_transformer(include_usage: bool) -> OpenAIStreamingResponseTransformer {
-        OpenAIStreamingResponseTransformer {
-            created: 0,
-            include_usage,
-            model: "test-model".to_owned(),
-            state: Arc::new(Mutex::new(OpenAIStreamingState::default())),
-            system_fingerprint: "test-fingerprint".to_owned(),
-        }
-    }
-
-    #[tokio::test]
-    async fn streaming_content_token_emits_content_delta() {
-        let transformer = streaming_transformer(false);
-
-        let message = token_message(GeneratedTokenResult::ContentToken("hello".to_owned()));
-        let chunks = transformer
-            .transform(message)
-            .await
-            .expect("the transformer must accept the message");
-
-        assert_eq!(
-            chunks,
-            vec![
-                TransformResult::Chunk(r#"{"id":"test-request","object":"chat.completion.chunk","created":0,"model":"test-model","system_fingerprint":"test-fingerprint","choices":[{"index":0,"delta":{"role":"assistant","content":"hello"},"logprobs":null,"finish_reason":null}]}"#.to_owned()),
-            ]
-        );
-    }
-
-    #[tokio::test]
-    async fn streaming_reasoning_token_is_dropped() {
-        let transformer = streaming_transformer(false);
-
-        let message = token_message(GeneratedTokenResult::ReasoningToken("thought".to_owned()));
-        let chunks = transformer
-            .transform(message)
-            .await
-            .expect("the transformer must accept the message");
-
-        assert_eq!(chunks.len(), 0);
-    }
-
-    #[tokio::test]
-    async fn streaming_undeterminable_token_emits_content_delta() {
-        let transformer = streaming_transformer(false);
-
-        let message = token_message(GeneratedTokenResult::UndeterminableToken(
-            "ambig".to_owned(),
-        ));
-        let chunks = transformer
-            .transform(message)
-            .await
-            .expect("the transformer must accept the message");
-
-        assert_eq!(
-            chunks,
-            vec![
-                TransformResult::Chunk(r#"{"id":"test-request","object":"chat.completion.chunk","created":0,"model":"test-model","system_fingerprint":"test-fingerprint","choices":[{"index":0,"delta":{"role":"assistant","content":"ambig"},"logprobs":null,"finish_reason":null}]}"#.to_owned()),
-            ]
-        );
-    }
-
-    #[tokio::test]
-    async fn streaming_tool_call_token_is_silently_dropped() {
-        let transformer = streaming_transformer(false);
-
-        let chunks = transformer
-            .transform(token_message(GeneratedTokenResult::ToolCallToken(
-                "{".to_owned(),
-            )))
-            .await
-            .expect("the transformer must accept the message");
-
-        assert_eq!(chunks.len(), 0);
-    }
-
-    #[tokio::test]
-    async fn streaming_tool_call_parsed_emits_structured_tool_calls_chunk() {
-        let transformer = streaming_transformer(false);
-
-        let chunks = transformer
-            .transform(token_message(GeneratedTokenResult::ToolCallParsed(vec![
-                weather_call(),
-            ])))
-            .await
-            .expect("the transformer must accept the message");
-
-        assert_eq!(
-            chunks,
-            vec![
-                TransformResult::Chunk(r#"{"id":"test-request","object":"chat.completion.chunk","created":0,"model":"test-model","system_fingerprint":"test-fingerprint","choices":[{"index":0,"delta":{"role":"assistant","tool_calls":[{"index":0,"id":"call_x","type":"function","function":{"name":"get_weather","arguments":"{\"location\":\"Paris\"}"}}]},"logprobs":null,"finish_reason":null}]}"#.to_owned()),
-            ]
-        );
-    }
-
-    #[tokio::test]
-    async fn streaming_done_after_tool_call_uses_tool_calls_finish_reason() {
-        let transformer = streaming_transformer(false);
-
-        transformer
-            .transform(token_message(GeneratedTokenResult::ToolCallParsed(vec![
-                weather_call(),
-            ])))
-            .await
-            .expect("the transformer must accept the message");
-
-        let summary = summary_with_counts(2, 0, 0);
-        let chunks = transformer
-            .transform(token_message(GeneratedTokenResult::Done(summary)))
-            .await
-            .expect("the transformer must accept the message");
-
-        assert_eq!(
-            chunks,
-            vec![
-                TransformResult::Chunk(r#"{"id":"test-request","object":"chat.completion.chunk","created":0,"model":"test-model","system_fingerprint":"test-fingerprint","choices":[{"index":0,"delta":{},"logprobs":null,"finish_reason":"tool_calls"}]}"#.to_owned()),
-            ]
-        );
-    }
-
-    #[tokio::test]
-    async fn streaming_done_without_tool_call_uses_stop_finish_reason() {
-        let transformer = streaming_transformer(false);
-
-        transformer
-            .transform(token_message(GeneratedTokenResult::ContentToken(
-                "hi".to_owned(),
-            )))
-            .await
-            .expect("the transformer must accept the message");
-
-        let summary = summary_with_counts(2, 1, 0);
-        let chunks = transformer
-            .transform(token_message(GeneratedTokenResult::Done(summary)))
-            .await
-            .expect("the transformer must accept the message");
-
-        assert_eq!(
-            chunks,
-            vec![
-                TransformResult::Chunk(r#"{"id":"test-request","object":"chat.completion.chunk","created":0,"model":"test-model","system_fingerprint":"test-fingerprint","choices":[{"index":0,"delta":{},"logprobs":null,"finish_reason":"stop"}]}"#.to_owned()),
-            ]
-        );
-    }
-
-    #[tokio::test]
-    async fn streaming_done_with_include_usage_emits_finish_then_usage_chunk() {
-        let transformer = streaming_transformer(true);
-        let summary = summary_with_counts(7, 4, 1);
-
-        let chunks = transformer
-            .transform(token_message(GeneratedTokenResult::Done(summary)))
-            .await
-            .expect("the transformer must accept the message");
-
-        assert_eq!(
-            chunks,
-            vec![
-                TransformResult::Chunk(r#"{"id":"test-request","object":"chat.completion.chunk","created":0,"model":"test-model","system_fingerprint":"test-fingerprint","choices":[{"index":0,"delta":{},"logprobs":null,"finish_reason":"stop"}]}"#.to_owned()),
-                TransformResult::Chunk(r#"{"id":"test-request","object":"chat.completion.chunk","created":0,"model":"test-model","system_fingerprint":"test-fingerprint","choices":[],"usage":{"prompt_tokens":7,"completion_tokens":5,"total_tokens":12,"prompt_tokens_details":{"cached_tokens":0,"audio_tokens":0},"completion_tokens_details":{"reasoning_tokens":1}}}"#.to_owned()),
-            ]
-        );
-    }
-
-    #[tokio::test]
-    async fn streaming_done_without_include_usage_emits_only_finish_chunk() {
-        let transformer = streaming_transformer(false);
-        let summary = summary_with_counts(5, 3, 2);
-
-        let chunks = transformer
-            .transform(token_message(GeneratedTokenResult::Done(summary)))
-            .await
-            .expect("the transformer must accept the message");
-
-        assert_eq!(
-            chunks,
-            vec![
-                TransformResult::Chunk(r#"{"id":"test-request","object":"chat.completion.chunk","created":0,"model":"test-model","system_fingerprint":"test-fingerprint","choices":[{"index":0,"delta":{},"logprobs":null,"finish_reason":"stop"}]}"#.to_owned()),
-            ]
-        );
-    }
-
-    #[tokio::test]
-    async fn streaming_tool_call_parse_failed_emits_server_error() {
-        let transformer = streaming_transformer(false);
-
-        let chunks = transformer
-            .transform(token_message(GeneratedTokenResult::ToolCallParseFailed(
-                "bad payload".to_owned(),
-            )))
-            .await
-            .expect("the transformer must accept the message");
-
-        assert_eq!(
-            chunks,
-            vec![
-                TransformResult::Error(r#"{"error":{"message":"bad payload","type":"server_error","param":null,"code":null}}"#.to_owned()),
-            ]
-        );
-    }
-
-    #[tokio::test]
-    async fn streaming_tool_call_validation_failed_emits_server_error() {
-        let transformer = streaming_transformer(false);
-
-        let chunks = transformer
-            .transform(token_message(
-                GeneratedTokenResult::ToolCallValidationFailed(vec!["missing field x".to_owned()]),
-            ))
-            .await
-            .expect("the transformer must accept the message");
-
-        assert_eq!(
-            chunks,
-            vec![
-                TransformResult::Error(r#"{"error":{"message":"missing field x","type":"server_error","param":null,"code":null}}"#.to_owned()),
-            ]
-        );
-    }
-
-    #[tokio::test]
-    async fn streaming_unrecognized_tool_call_format_emits_server_error() {
-        let transformer = streaming_transformer(false);
-
-        let chunks = transformer
-            .transform(token_message(
-                GeneratedTokenResult::UnrecognizedToolCallFormat(RawToolCallTokens {
-                    text: "<unknown_marker>blah</unknown_marker>".to_owned(),
-                    ffi_error_message: "common_chat_parse failed: no parser".to_owned(),
-                }),
-            ))
-            .await
-            .expect("the transformer must accept the message");
-
-        assert_eq!(
-            chunks,
-            vec![
-                TransformResult::Error(r#"{"error":{"message":"model produced output the parser did not recognise as any registered tool-call format; FFI error: common_chat_parse failed: no parser; raw text: <unknown_marker>blah</unknown_marker>","type":"server_error","param":null,"code":null}}"#.to_owned()),
-            ]
-        );
-    }
-
-    #[tokio::test]
-    async fn streaming_error_message_returns_server_error() {
-        let transformer = streaming_transformer(false);
-
-        let message = error_message(500, "internal server error");
-        let chunks = transformer
-            .transform(message)
-            .await
-            .expect("the transformer must accept the message");
-
-        assert_eq!(
-            chunks,
-            vec![
-                TransformResult::Error(r#"{"error":{"message":"internal server error","type":"server_error","param":null,"code":null}}"#.to_owned()),
-            ]
-        );
-    }
-
-    #[tokio::test]
-    async fn streaming_chat_template_error_returns_server_error() {
-        let transformer = streaming_transformer(false);
-
-        let message = token_message(GeneratedTokenResult::ChatTemplateError(
-            "bad template".to_owned(),
-        ));
-        let chunks = transformer
-            .transform(message)
-            .await
-            .expect("the transformer must accept the message");
-
-        assert_eq!(
-            chunks,
-            vec![
-                TransformResult::Error(r#"{"error":{"message":"bad template","type":"server_error","param":null,"code":null}}"#.to_owned()),
-            ]
-        );
-    }
-
-    #[tokio::test]
-    async fn streaming_image_decoding_failed_returns_invalid_request_error() {
-        let transformer = streaming_transformer(false);
-
-        let message = token_message(GeneratedTokenResult::ImageDecodingFailed(
-            "unsupported format".to_owned(),
-        ));
-        let chunks = transformer
-            .transform(message)
-            .await
-            .expect("the transformer must accept the message");
-
-        assert_eq!(
-            chunks,
-            vec![
-                TransformResult::Error(r#"{"error":{"message":"unsupported format","type":"invalid_request_error","param":null,"code":null}}"#.to_owned()),
-            ]
-        );
-    }
-
-    #[tokio::test]
-    async fn streaming_multimodal_not_supported_returns_invalid_request_error() {
-        let transformer = streaming_transformer(false);
-
-        let message = token_message(GeneratedTokenResult::MultimodalNotSupported(
-            "model does not support images".to_owned(),
-        ));
-        let chunks = transformer
-            .transform(message)
-            .await
-            .expect("the transformer must accept the message");
-
-        assert_eq!(
-            chunks,
-            vec![
-                TransformResult::Error(r#"{"error":{"message":"model does not support images","type":"invalid_request_error","param":null,"code":null}}"#.to_owned()),
-            ]
-        );
-    }
-
-    #[tokio::test]
-    async fn streaming_media_exceeding_the_micro_batch_returns_invalid_request_error() {
-        let transformer = streaming_transformer(false);
-
-        let message = token_message(GeneratedTokenResult::MediaExceedsMicroBatch(
-            OversizedMediaDetails {
-                media_tokens: 256,
-                micro_batch_tokens: 128,
-            },
-        ));
-        let chunks = transformer
-            .transform(message)
-            .await
-            .expect("the transformer must accept the message");
-
-        assert_eq!(
-            chunks,
-            vec![
-                TransformResult::Error(r#"{"error":{"message":"media required 256 tokens but one agent micro batch holds 128 tokens","type":"invalid_request_error","param":null,"code":null}}"#.to_owned()),
-            ]
-        );
-    }
-
-    #[tokio::test]
-    async fn streaming_tool_call_with_invalid_json_arguments_passes_raw_string_through() {
-        let transformer = streaming_transformer(false);
-
-        let chunks = transformer
-            .transform(token_message(GeneratedTokenResult::ToolCallParsed(vec![
-                invalid_json_call(),
-            ])))
-            .await
-            .unwrap();
-
-        assert_eq!(
-            chunks,
-            vec![
-                TransformResult::Chunk(r#"{"id":"test-request","object":"chat.completion.chunk","created":0,"model":"test-model","system_fingerprint":"test-fingerprint","choices":[{"index":0,"delta":{"role":"assistant","tool_calls":[{"index":0,"id":"call_invalid","type":"function","function":{"name":"broken_tool","arguments":"{not valid json"}}]},"logprobs":null,"finish_reason":null}]}"#.to_owned()),
-            ]
-        );
-    }
-
-    #[tokio::test]
-    async fn streaming_empty_parsed_tool_calls_emit_no_chunks() {
-        let transformer = streaming_transformer(false);
-
-        let chunks = transformer
-            .transform(token_message(GeneratedTokenResult::ToolCallParsed(
-                Vec::new(),
-            )))
-            .await
-            .unwrap();
-
-        assert_eq!(chunks.len(), 0);
-    }
-
-    #[tokio::test]
-    async fn streaming_embedding_response_returns_invalid_request_error() {
-        let transformer = streaming_transformer(false);
-
-        let message = response_message(OutgoingResponse::Embedding(EmbeddingResult::Done));
-        let chunks = transformer.transform(message).await.unwrap();
-
-        assert_eq!(
-            chunks,
-            vec![
-                TransformResult::Error(r#"{"error":{"message":"unexpected embedding response to a token generation request","type":"invalid_request_error","param":null,"code":null}}"#.to_owned()),
-            ]
-        );
-    }
-
-    #[tokio::test]
-    async fn streaming_grammar_incompatible_with_thinking_returns_invalid_request_error() {
-        let transformer = streaming_transformer(false);
-
-        let chunks = transformer
-            .transform(token_message(
-                GeneratedTokenResult::GrammarIncompatibleWithThinking(
-                    "grammar conflicts with thinking".to_owned(),
-                ),
-            ))
-            .await
-            .unwrap();
-
-        assert_eq!(
-            chunks,
-            vec![
-                TransformResult::Error(r#"{"error":{"message":"grammar conflicts with thinking","type":"invalid_request_error","param":null,"code":null}}"#.to_owned()),
-            ]
-        );
-    }
-
-    #[tokio::test]
-    async fn streaming_grammar_rejected_model_output_returns_server_error() {
-        let transformer = streaming_transformer(false);
-
-        let chunks = transformer
-            .transform(token_message(
-                GeneratedTokenResult::GrammarRejectedModelOutput(
-                    "output rejected by grammar".to_owned(),
-                ),
-            ))
-            .await
-            .unwrap();
-
-        assert_eq!(
-            chunks,
-            vec![
-                TransformResult::Error(r#"{"error":{"message":"output rejected by grammar","type":"server_error","param":null,"code":null}}"#.to_owned()),
-            ]
-        );
-    }
-
-    #[tokio::test]
-    async fn streaming_grammar_initialization_failed_returns_server_error() {
-        let transformer = streaming_transformer(false);
-
-        let chunks = transformer
-            .transform(token_message(
-                GeneratedTokenResult::GrammarInitializationFailed(
-                    "could not build grammar".to_owned(),
-                ),
-            ))
-            .await
-            .unwrap();
-
-        assert_eq!(
-            chunks,
-            vec![
-                TransformResult::Error(r#"{"error":{"message":"could not build grammar","type":"server_error","param":null,"code":null}}"#.to_owned()),
-            ]
-        );
-    }
-
-    #[tokio::test]
-    async fn streaming_grammar_syntax_error_returns_invalid_request_error() {
-        let transformer = streaming_transformer(false);
-
-        let chunks = transformer
-            .transform(token_message(GeneratedTokenResult::GrammarSyntaxError(
-                "bad grammar syntax".to_owned(),
-            )))
-            .await
-            .unwrap();
-
-        assert_eq!(
-            chunks,
-            vec![
-                TransformResult::Error(r#"{"error":{"message":"bad grammar syntax","type":"invalid_request_error","param":null,"code":null}}"#.to_owned()),
-            ]
-        );
-    }
-
-    #[tokio::test]
-    async fn streaming_sampler_error_returns_server_error() {
-        let transformer = streaming_transformer(false);
-
-        let chunks = transformer
-            .transform(token_message(GeneratedTokenResult::SamplerError(
-                "sampler blew up".to_owned(),
-            )))
-            .await
-            .unwrap();
-
-        assert_eq!(
-            chunks,
-            vec![
-                TransformResult::Error(r#"{"error":{"message":"sampler blew up","type":"server_error","param":null,"code":null}}"#.to_owned()),
-            ]
-        );
-    }
-
-    #[tokio::test]
-    async fn streaming_tool_schema_invalid_returns_invalid_request_error() {
-        let transformer = streaming_transformer(false);
-
-        let chunks = transformer
-            .transform(token_message(GeneratedTokenResult::ToolSchemaInvalid(
-                "schema is not valid".to_owned(),
-            )))
-            .await
-            .unwrap();
-
-        assert_eq!(
-            chunks,
-            vec![
-                TransformResult::Error(r#"{"error":{"message":"schema is not valid","type":"invalid_request_error","param":null,"code":null}}"#.to_owned()),
-            ]
-        );
-    }
-
-    #[tokio::test]
-    async fn rejects_inference_socket_notifications() {
-        let transform_result = streaming_transformer(false)
-            .transform(OutgoingMessage::Notification(
-                Notification::TokenGenerationEnabled,
-            ))
-            .await;
-
-        assert!(matches!(
-            transform_result,
-            Err(AgentRelayError::MessageNotRelayable { message })
-                if matches!(*message, OutgoingMessage::Notification(Notification::TokenGenerationEnabled))
-        ));
-    }
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/openai_streaming_state.rs b/paddler_balancer/src/compatibility/openai_service/openai_streaming_state.rs
deleted file mode 100644
index b57f74f4d..000000000
--- a/paddler_balancer/src/compatibility/openai_service/openai_streaming_state.rs
+++ /dev/null
@@ -1,4 +0,0 @@
-#[derive(Default)]
-pub struct OpenAIStreamingState {
-    pub saw_tool_call: bool,
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/openai_tool_parameters_schema.rs b/paddler_balancer/src/compatibility/openai_service/openai_tool_parameters_schema.rs
deleted file mode 100644
index 91cda920e..000000000
--- a/paddler_balancer/src/compatibility/openai_service/openai_tool_parameters_schema.rs
+++ /dev/null
@@ -1,79 +0,0 @@
-use serde::Deserialize;
-use serde_json::Map;
-use serde_json::Value;
-
-use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters_schema::raw_parameters_schema::RawParametersSchema;
-
-#[derive(Deserialize)]
-pub struct OpenAIToolParametersSchema {
-    #[serde(default, rename = "type")]
-    pub schema_type: Option<String>,
-    #[serde(default)]
-    pub properties: Option<Map<String, Value>>,
-    #[serde(default)]
-    pub required: Option<Vec<String>>,
-    #[serde(default, rename = "additionalProperties")]
-    pub additional_properties: Option<Value>,
-}
-
-impl OpenAIToolParametersSchema {
-    #[must_use]
-    pub fn into_raw_parameters_schema(self) -> RawParametersSchema {
-        let Self {
-            schema_type,
-            properties,
-            required,
-            additional_properties,
-        } = self;
-
-        RawParametersSchema {
-            schema_type: schema_type.unwrap_or_else(|| "object".to_owned()),
-            properties,
-            required,
-            additional_properties,
-        }
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use serde_json::from_value;
-    use serde_json::json;
-
-    use super::OpenAIToolParametersSchema;
-
-    #[test]
-    fn conversion_ignores_unknown_keywords_and_keeps_recognized_fields() {
-        let schema: OpenAIToolParametersSchema = from_value(json!({
-            "type": "object",
-            "properties": {"location": {"type": "string"}},
-            "required": ["location"],
-            "additionalProperties": false,
-            "$schema": "http://json-schema.org/draft-07/schema#",
-            "title": "Weather",
-            "$defs": {"Unit": {"type": "string"}}
-        }))
-        .unwrap();
-
-        let raw = schema.into_raw_parameters_schema();
-
-        assert_eq!(raw.schema_type, "object");
-        assert_eq!(raw.properties.unwrap().len(), 1);
-        assert_eq!(raw.required, Some(vec!["location".to_owned()]));
-        assert_eq!(raw.additional_properties, Some(json!(false)));
-    }
-
-    #[test]
-    fn conversion_defaults_missing_type_to_object() {
-        let schema: OpenAIToolParametersSchema = from_value(json!({
-            "properties": {"location": {"type": "string"}}
-        }))
-        .unwrap();
-
-        let raw = schema.into_raw_parameters_schema();
-
-        assert_eq!(raw.schema_type, "object");
-        assert!(raw.required.is_none());
-        assert!(raw.additional_properties.is_none());
-    }
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/openai_usage.rs b/paddler_balancer/src/compatibility/openai_service/openai_usage.rs
deleted file mode 100644
index 29ef258e1..000000000
--- a/paddler_balancer/src/compatibility/openai_service/openai_usage.rs
+++ /dev/null
@@ -1,52 +0,0 @@
-use llama_cpp_bindings_types::TokenUsage;
-use serde::Serialize;
-use serde::Serializer;
-
-#[derive(Serialize)]
-struct PromptTokensDetails {
-    cached_tokens: u64,
-    audio_tokens: u64,
-}
-
-#[derive(Serialize)]
-struct CompletionTokensDetails {
-    reasoning_tokens: u64,
-}
-
-#[derive(Serialize)]
-struct SerializedUsage {
-    prompt_tokens: u64,
-    completion_tokens: u64,
-    total_tokens: u64,
-    prompt_tokens_details: PromptTokensDetails,
-    completion_tokens_details: CompletionTokensDetails,
-}
-
-#[derive(Clone, Copy, Debug)]
-pub struct OpenAIUsage(pub TokenUsage);
-
-impl Serialize for OpenAIUsage {
-    fn serialize<TSerializer>(
-        &self,
-        serializer: TSerializer,
-    ) -> Result<TSerializer::Ok, TSerializer::Error>
-    where
-        TSerializer: Serializer,
-    {
-        let Self(usage) = self;
-
-        SerializedUsage {
-            prompt_tokens: usage.prompt_tokens,
-            completion_tokens: usage.completion_tokens(),
-            total_tokens: usage.total_tokens(),
-            prompt_tokens_details: PromptTokensDetails {
-                cached_tokens: usage.cached_prompt_tokens,
-                audio_tokens: usage.input_audio_tokens,
-            },
-            completion_tokens_details: CompletionTokensDetails {
-                reasoning_tokens: usage.reasoning_tokens,
-            },
-        }
-        .serialize(serializer)
-    }
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/output_item_event.rs b/paddler_balancer/src/compatibility/openai_service/output_item_event.rs
deleted file mode 100644
index 61fdcf16c..000000000
--- a/paddler_balancer/src/compatibility/openai_service/output_item_event.rs
+++ /dev/null
@@ -1,10 +0,0 @@
-use serde::Serialize;
-
-use crate::compatibility::openai_service::responses_output_item::ResponsesOutputItem;
-
-#[derive(Clone, Debug, Serialize)]
-pub struct OutputItemEvent {
-    pub sequence_number: u64,
-    pub output_index: usize,
-    pub item: ResponsesOutputItem,
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/response_snapshot_event.rs b/paddler_balancer/src/compatibility/openai_service/response_snapshot_event.rs
deleted file mode 100644
index d31ee26c8..000000000
--- a/paddler_balancer/src/compatibility/openai_service/response_snapshot_event.rs
+++ /dev/null
@@ -1,9 +0,0 @@
-use serde::Serialize;
-
-use crate::compatibility::openai_service::responses_response::ResponsesResponse;
-
-#[derive(Clone, Debug, Serialize)]
-pub struct ResponseSnapshotEvent {
-    pub sequence_number: u64,
-    pub response: ResponsesResponse,
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/responses_non_streaming_response_transformer.rs b/paddler_balancer/src/compatibility/openai_service/responses_non_streaming_response_transformer.rs
deleted file mode 100644
index e08d979cc..000000000
--- a/paddler_balancer/src/compatibility/openai_service/responses_non_streaming_response_transformer.rs
+++ /dev/null
@@ -1,346 +0,0 @@
-use std::mem::take;
-use std::sync::Arc;
-
-use async_trait::async_trait;
-use parking_lot::Mutex;
-use serde_json::to_string;
-
-use paddler_messaging::generated_token_result::GeneratedTokenResult;
-use paddler_messaging::generation_summary::GenerationSummary;
-use paddler_messaging::inference_client::message::Message as OutgoingMessage;
-use paddler_messaging::inference_client::response::Response as OutgoingResponse;
-use paddler_messaging::jsonrpc::response_envelope::ResponseEnvelope;
-
-use crate::agent_relay_error::AgentRelayError;
-use crate::chunk_forwarding_session_controller::transform_result::TransformResult;
-use crate::chunk_forwarding_session_controller::transforms_outgoing_message::TransformsOutgoingMessage;
-use crate::compatibility::openai_service::responses_non_streaming_state::ResponsesNonStreamingState;
-use crate::compatibility::openai_service::responses_output_item::ResponsesOutputItem;
-use crate::compatibility::openai_service::responses_output_item_kind::ResponsesOutputItemKind;
-use crate::compatibility::openai_service::responses_response_header::ResponsesResponseHeader;
-use crate::compatibility::openai_service::try_universal_error_chunk::try_universal_error_chunk;
-
-#[derive(Clone)]
-pub struct ResponsesNonStreamingResponseTransformer {
-    pub header: ResponsesResponseHeader,
-    pub state: Arc<Mutex<ResponsesNonStreamingState>>,
-}
-
-impl ResponsesNonStreamingResponseTransformer {
-    fn build_completed(&self, summary: &GenerationSummary) -> Result<String, AgentRelayError> {
-        let snapshot = take(&mut *self.state.lock());
-
-        let mut output: Vec<ResponsesOutputItem> = Vec::new();
-
-        if !snapshot.reasoning.is_empty() {
-            output.push(ResponsesOutputItem::completed_reasoning(
-                ResponsesOutputItemKind::Reasoning.item_id(output.len()),
-                snapshot.reasoning,
-            ));
-        }
-
-        let has_tool_calls = !snapshot.tool_calls.is_empty();
-
-        if !snapshot.content.is_empty() || !has_tool_calls {
-            output.push(ResponsesOutputItem::completed_message(
-                ResponsesOutputItemKind::Message.item_id(output.len()),
-                snapshot.content,
-            ));
-        }
-
-        for call in &snapshot.tool_calls {
-            output.push(ResponsesOutputItem::completed_function_call(
-                ResponsesOutputItemKind::FunctionCall.item_id(output.len()),
-                call,
-            ));
-        }
-
-        to_string(&self.header.finished(output, &summary.usage, summary.finish))
-            .map_err(AgentRelayError::MessageUnserializable)
-    }
-}
-
-#[async_trait]
-impl TransformsOutgoingMessage for ResponsesNonStreamingResponseTransformer {
-    type Output = TransformResult;
-
-    async fn transform(
-        &self,
-        message: OutgoingMessage,
-    ) -> Result<Vec<TransformResult>, AgentRelayError> {
-        match message {
-            OutgoingMessage::Response(ResponseEnvelope {
-                response:
-                    OutgoingResponse::GeneratedToken(
-                        GeneratedTokenResult::ContentToken(text)
-                        | GeneratedTokenResult::UndeterminableToken(text),
-                    ),
-                ..
-            }) => {
-                self.state.lock().content.push_str(&text);
-                Ok(vec![])
-            }
-            OutgoingMessage::Response(ResponseEnvelope {
-                response:
-                    OutgoingResponse::GeneratedToken(GeneratedTokenResult::ReasoningToken(text)),
-                ..
-            }) => {
-                self.state.lock().reasoning.push_str(&text);
-                Ok(vec![])
-            }
-            OutgoingMessage::Response(ResponseEnvelope {
-                response: OutgoingResponse::GeneratedToken(GeneratedTokenResult::ToolCallToken(_)),
-                ..
-            }) => Ok(vec![]),
-            OutgoingMessage::Response(ResponseEnvelope {
-                response:
-                    OutgoingResponse::GeneratedToken(GeneratedTokenResult::ToolCallParsed(parsed_calls)),
-                ..
-            }) => {
-                self.state.lock().tool_calls.extend(parsed_calls);
-                Ok(vec![])
-            }
-            OutgoingMessage::Response(ResponseEnvelope {
-                response: OutgoingResponse::GeneratedToken(GeneratedTokenResult::Done(summary)),
-                ..
-            }) => Ok(vec![TransformResult::Chunk(
-                self.build_completed(&summary)?,
-            )]),
-            other => try_universal_error_chunk(&other)
-                .map(|error_chunk| vec![error_chunk])
-                .ok_or_else(|| AgentRelayError::MessageNotRelayable {
-                    message: Box::new(other),
-                }),
-        }
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use std::mem::discriminant;
-    use std::sync::Arc;
-
-    use llama_cpp_bindings_types::ParsedToolCall;
-    use llama_cpp_bindings_types::TokenUsage;
-    use llama_cpp_bindings_types::ToolCallArguments;
-    use parking_lot::Mutex;
-    use serde_json::Value;
-    use serde_json::from_slice;
-    use serde_json::json;
-
-    use paddler_messaging::generated_token_result::GeneratedTokenResult;
-    use paddler_messaging::generation_finish::GenerationFinish;
-    use paddler_messaging::generation_summary::GenerationSummary;
-    use paddler_messaging::inference_client::message::Message as OutgoingMessage;
-    use paddler_messaging::inference_client::notification::Notification;
-    use paddler_messaging::inference_client::response::Response as OutgoingResponse;
-    use paddler_messaging::jsonrpc::response_envelope::ResponseEnvelope;
-    use paddler_openai_response_format_validator::openai_validator::OpenAIValidator;
-
-    use super::ResponsesNonStreamingResponseTransformer;
-    use super::ResponsesNonStreamingState;
-    use crate::agent_relay_error::AgentRelayError;
-    use crate::chunk_forwarding_session_controller::transform_result::TransformResult;
-    use crate::chunk_forwarding_session_controller::transforms_outgoing_message::TransformsOutgoingMessage;
-    use crate::compatibility::openai_service::responses_response_header::ResponsesResponseHeader;
-
-    #[must_use]
-    pub fn token_message(token_result: GeneratedTokenResult) -> OutgoingMessage {
-        OutgoingMessage::Response(ResponseEnvelope {
-            generated_by: None,
-            request_id: "test-request".to_owned(),
-            response: OutgoingResponse::GeneratedToken(token_result),
-        })
-    }
-
-    #[must_use]
-    pub fn summary_with_counts(
-        prompt_tokens: u64,
-        content_tokens: u64,
-        reasoning_tokens: u64,
-    ) -> GenerationSummary {
-        GenerationSummary {
-            finish: GenerationFinish::EndOfGeneration,
-            usage: TokenUsage {
-                prompt_tokens,
-                content_tokens,
-                reasoning_tokens,
-                ..TokenUsage::default()
-            },
-        }
-    }
-
-    #[must_use]
-    pub fn weather_call() -> ParsedToolCall {
-        ParsedToolCall::new(
-            "call_x".to_owned(),
-            "get_weather".to_owned(),
-            ToolCallArguments::ValidJson(json!({ "location": "Paris" })),
-        )
-    }
-
-    #[must_use]
-    pub fn header() -> ResponsesResponseHeader {
-        ResponsesResponseHeader {
-            id: "resp_test".to_owned(),
-            created_at: 0,
-            model: "test-model".to_owned(),
-            instructions: None,
-            temperature: 0.25,
-            top_p: 0.5,
-        }
-    }
-
-    fn only_chunk(chunks: Vec<TransformResult>) -> Value {
-        let [chunk] = <[TransformResult; 1]>::try_from(chunks).unwrap();
-
-        assert_eq!(
-            discriminant(&chunk),
-            discriminant(&TransformResult::Chunk(String::new()))
-        );
-
-        from_slice(&chunk.into_ndjson_line().unwrap()).unwrap()
-    }
-
-    fn non_streaming_transformer() -> ResponsesNonStreamingResponseTransformer {
-        ResponsesNonStreamingResponseTransformer {
-            header: header(),
-            state: Arc::new(Mutex::new(ResponsesNonStreamingState::default())),
-        }
-    }
-
-    #[tokio::test]
-    async fn non_streaming_aggregates_content_into_a_message_item() {
-        let transformer = non_streaming_transformer();
-
-        transformer
-            .transform(token_message(GeneratedTokenResult::ContentToken(
-                "hel".to_owned(),
-            )))
-            .await
-            .unwrap();
-        transformer
-            .transform(token_message(GeneratedTokenResult::ContentToken(
-                "lo".to_owned(),
-            )))
-            .await
-            .unwrap();
-        let chunks = transformer
-            .transform(token_message(GeneratedTokenResult::Done(
-                summary_with_counts(3, 2, 0),
-            )))
-            .await
-            .unwrap();
-
-        let response = only_chunk(chunks);
-
-        assert_eq!(response["object"], "response");
-        assert_eq!(response["status"], "completed");
-        assert_eq!(response["output"][0]["type"], "message");
-        assert_eq!(response["output"][0]["content"][0]["text"], "hello");
-    }
-
-    #[tokio::test]
-    async fn non_streaming_surfaces_reasoning_and_tool_calls_in_output() {
-        let transformer = non_streaming_transformer();
-
-        transformer
-            .transform(token_message(GeneratedTokenResult::ReasoningToken(
-                "ponder".to_owned(),
-            )))
-            .await
-            .unwrap();
-        transformer
-            .transform(token_message(GeneratedTokenResult::ToolCallParsed(vec![
-                weather_call(),
-            ])))
-            .await
-            .unwrap();
-        let chunks = transformer
-            .transform(token_message(GeneratedTokenResult::Done(
-                summary_with_counts(3, 0, 1),
-            )))
-            .await
-            .unwrap();
-
-        let response = only_chunk(chunks);
-
-        assert_eq!(response["output"][0]["type"], "reasoning");
-        assert_eq!(response["output"][1]["type"], "function_call");
-        assert_eq!(response["output"][1]["name"], "get_weather");
-        assert_eq!(
-            response["usage"]["output_tokens_details"]["reasoning_tokens"],
-            1
-        );
-    }
-
-    #[tokio::test]
-    async fn non_streaming_error_returns_an_error_envelope() {
-        let transformer = non_streaming_transformer();
-
-        let chunks = transformer
-            .transform(token_message(GeneratedTokenResult::SamplerError(
-                "sampler blew up".to_owned(),
-            )))
-            .await
-            .unwrap();
-
-        assert_eq!(
-            chunks,
-            vec![TransformResult::Error(
-                r#"{"error":{"message":"sampler blew up","type":"server_error","param":null,"code":null}}"#
-                    .to_owned()
-            )]
-        );
-    }
-
-    #[tokio::test]
-    async fn the_non_streaming_response_conforms_to_the_schema() {
-        let validator = OpenAIValidator::new().unwrap();
-        let transformer = non_streaming_transformer();
-
-        transformer
-            .transform(token_message(GeneratedTokenResult::ReasoningToken(
-                "p".to_owned(),
-            )))
-            .await
-            .unwrap();
-        transformer
-            .transform(token_message(GeneratedTokenResult::ContentToken(
-                "hello".to_owned(),
-            )))
-            .await
-            .unwrap();
-        transformer
-            .transform(token_message(GeneratedTokenResult::ToolCallParsed(vec![
-                weather_call(),
-            ])))
-            .await
-            .unwrap();
-        let chunks = transformer
-            .transform(token_message(GeneratedTokenResult::Done(
-                summary_with_counts(5, 3, 2),
-            )))
-            .await
-            .unwrap();
-
-        let response = only_chunk(chunks);
-
-        validator.validate_responses_response(&response).unwrap();
-    }
-
-    #[tokio::test]
-    async fn rejects_inference_socket_notifications() {
-        let transform_result = non_streaming_transformer()
-            .transform(OutgoingMessage::Notification(
-                Notification::TokenGenerationEnabled,
-            ))
-            .await;
-
-        assert!(matches!(
-            transform_result,
-            Err(AgentRelayError::MessageNotRelayable { message })
-                if matches!(*message, OutgoingMessage::Notification(Notification::TokenGenerationEnabled))
-        ));
-    }
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/responses_non_streaming_state.rs b/paddler_balancer/src/compatibility/openai_service/responses_non_streaming_state.rs
deleted file mode 100644
index c7c8e4ac0..000000000
--- a/paddler_balancer/src/compatibility/openai_service/responses_non_streaming_state.rs
+++ /dev/null
@@ -1,8 +0,0 @@
-use llama_cpp_bindings_types::ParsedToolCall;
-
-#[derive(Clone, Default)]
-pub struct ResponsesNonStreamingState {
-    pub content: String,
-    pub reasoning: String,
-    pub tool_calls: Vec<ParsedToolCall>,
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/responses_output_item_kind.rs b/paddler_balancer/src/compatibility/openai_service/responses_output_item_kind.rs
deleted file mode 100644
index d9d0e5d08..000000000
--- a/paddler_balancer/src/compatibility/openai_service/responses_output_item_kind.rs
+++ /dev/null
@@ -1,37 +0,0 @@
-#[derive(Clone, Copy)]
-pub enum ResponsesOutputItemKind {
-    FunctionCall,
-    Message,
-    Reasoning,
-}
-
-impl ResponsesOutputItemKind {
-    #[must_use]
-    pub fn item_id(self, output_index: usize) -> String {
-        let prefix = match self {
-            Self::FunctionCall => "fc",
-            Self::Message => "msg",
-            Self::Reasoning => "rs",
-        };
-
-        format!("{prefix}_{output_index}")
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use super::ResponsesOutputItemKind;
-
-    #[test]
-    fn names_each_output_item_after_its_kind_and_position() {
-        assert_eq!(
-            [
-                ResponsesOutputItemKind::Reasoning,
-                ResponsesOutputItemKind::Message,
-                ResponsesOutputItemKind::FunctionCall,
-            ]
-            .map(|item_kind| item_kind.item_id(2)),
-            ["rs_2", "msg_2", "fc_2"]
-        );
-    }
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/responses_prepared_request.rs b/paddler_balancer/src/compatibility/openai_service/responses_prepared_request.rs
deleted file mode 100644
index 4175ef9c0..000000000
--- a/paddler_balancer/src/compatibility/openai_service/responses_prepared_request.rs
+++ /dev/null
@@ -1,9 +0,0 @@
-use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
-use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters_schema::validated_parameters_schema::ValidatedParametersSchema;
-
-pub struct ResponsesPreparedRequest {
-    pub paddler_params: ContinueFromConversationHistoryParams<ValidatedParametersSchema>,
-    pub stream: bool,
-    pub model: String,
-    pub instructions: Option<String>,
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/responses_response_header.rs b/paddler_balancer/src/compatibility/openai_service/responses_response_header.rs
deleted file mode 100644
index 6d46fd604..000000000
--- a/paddler_balancer/src/compatibility/openai_service/responses_response_header.rs
+++ /dev/null
@@ -1,138 +0,0 @@
-use llama_cpp_bindings_types::TokenUsage;
-
-use paddler_messaging::generation_finish::GenerationFinish;
-
-use crate::compatibility::openai_service::openai_error::OpenAIError;
-use crate::compatibility::openai_service::responses_output_item::ResponsesOutputItem;
-use crate::compatibility::openai_service::responses_response::ResponsesResponse;
-use crate::compatibility::openai_service::responses_response_progress::ResponsesResponseProgress;
-use crate::compatibility::openai_service::responses_usage::ResponsesUsage;
-
-#[derive(Clone, Debug)]
-pub struct ResponsesResponseHeader {
-    pub created_at: u64,
-    pub id: String,
-    pub instructions: Option<String>,
-    pub model: String,
-    pub temperature: f32,
-    pub top_p: f32,
-}
-
-impl ResponsesResponseHeader {
-    #[must_use]
-    pub fn in_progress(&self) -> ResponsesResponse {
-        self.with_progress(ResponsesResponseProgress::InProgress)
-    }
-
-    #[must_use]
-    pub fn finished(
-        &self,
-        output: Vec<ResponsesOutputItem>,
-        usage: &TokenUsage,
-        finish: GenerationFinish,
-    ) -> ResponsesResponse {
-        let usage = ResponsesUsage(*usage);
-
-        self.with_progress(if finish.reached_a_length_limit() {
-            ResponsesResponseProgress::Incomplete { output, usage }
-        } else {
-            ResponsesResponseProgress::Completed { output, usage }
-        })
-    }
-
-    #[must_use]
-    pub fn failed(&self, error: &OpenAIError) -> ResponsesResponse {
-        self.with_progress(ResponsesResponseProgress::Failed {
-            error_message: error.message.clone(),
-        })
-    }
-
-    fn with_progress(&self, progress: ResponsesResponseProgress) -> ResponsesResponse {
-        ResponsesResponse {
-            header: self.clone(),
-            progress,
-        }
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use llama_cpp_bindings_types::TokenUsage;
-    use serde_json::json;
-    use serde_json::to_value;
-
-    use paddler_messaging::generation_finish::GenerationFinish;
-
-    use super::ResponsesResponseHeader;
-    use crate::compatibility::openai_service::openai_error::OpenAIError;
-    use crate::compatibility::openai_service::openai_error_type::OpenAIErrorType;
-    use crate::compatibility::openai_service::responses_output_item::ResponsesOutputItem;
-
-    fn header() -> ResponsesResponseHeader {
-        ResponsesResponseHeader {
-            id: "resp_test".to_owned(),
-            created_at: 1_234,
-            model: "test-model".to_owned(),
-            instructions: None,
-            temperature: 0.25,
-            top_p: 0.5,
-        }
-    }
-
-    #[test]
-    fn in_progress_reports_in_progress_status_with_no_output() {
-        let response = to_value(header().in_progress()).unwrap();
-
-        assert_eq!(response["status"], "in_progress");
-        assert_eq!(response["object"], "response");
-        assert_eq!(response["output"], json!([]));
-        assert!(response.get("usage").is_none());
-    }
-
-    #[test]
-    fn reports_the_sampling_parameters_in_effect() {
-        let response = to_value(header().in_progress()).unwrap();
-
-        assert_eq!(response["temperature"], 0.25);
-        assert_eq!(response["top_p"], 0.5);
-    }
-
-    #[test]
-    fn failed_carries_the_error_message() {
-        let response = to_value(header().failed(&OpenAIError {
-            error_type: OpenAIErrorType::ServerError,
-            message: "boom".to_owned(),
-        }))
-        .unwrap();
-
-        assert_eq!(response["status"], "failed");
-        assert_eq!(response["error"]["message"], "boom");
-        assert!(response.get("usage").is_none());
-    }
-
-    #[test]
-    fn completed_includes_usage_and_output() {
-        let response = to_value(header().finished(
-            vec![ResponsesOutputItem::completed_message(
-                "msg_0".to_owned(),
-                "hello".to_owned(),
-            )],
-            &TokenUsage {
-                prompt_tokens: 3,
-                cached_prompt_tokens: 0,
-                input_image_tokens: 0,
-                input_audio_tokens: 0,
-                content_tokens: 5,
-                reasoning_tokens: 2,
-                tool_call_tokens: 0,
-                undeterminable_tokens: 0,
-            },
-            GenerationFinish::EndOfGeneration,
-        ))
-        .unwrap();
-
-        assert_eq!(response["status"], "completed");
-        assert_eq!(response["usage"]["input_tokens"], 3);
-        assert_eq!(response["output"][0]["type"], "message");
-    }
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/responses_response_progress.rs b/paddler_balancer/src/compatibility/openai_service/responses_response_progress.rs
deleted file mode 100644
index ba971848b..000000000
--- a/paddler_balancer/src/compatibility/openai_service/responses_response_progress.rs
+++ /dev/null
@@ -1,18 +0,0 @@
-use crate::compatibility::openai_service::responses_output_item::ResponsesOutputItem;
-use crate::compatibility::openai_service::responses_usage::ResponsesUsage;
-
-#[derive(Clone, Debug)]
-pub enum ResponsesResponseProgress {
-    Completed {
-        output: Vec<ResponsesOutputItem>,
-        usage: ResponsesUsage,
-    },
-    Failed {
-        error_message: String,
-    },
-    Incomplete {
-        output: Vec<ResponsesOutputItem>,
-        usage: ResponsesUsage,
-    },
-    InProgress,
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/responses_sse_response.rs b/paddler_balancer/src/compatibility/openai_service/responses_sse_response.rs
index 6e77c5f23..f0c94cd70 100644
--- a/paddler_balancer/src/compatibility/openai_service/responses_sse_response.rs
+++ b/paddler_balancer/src/compatibility/openai_service/responses_sse_response.rs
@@ -1,27 +1,60 @@
 use actix_web::HttpResponse;
-use actix_web::http::header;
 use actix_web_lab::sse;
+use futures::future::ready;
 use futures::stream::Stream;
 use futures::stream::StreamExt as _;
-use serde_json::Error;
+use futures::stream::iter;
+use serde_json::Error as SerdeJsonError;
 use serde_json::to_string;
 
-use crate::compatibility::openai_service::responses_stream_event::ResponsesStreamEvent;
+use paddler_messaging::generated_token_final_result::GeneratedTokenFinalResult;
+use paddler_openai_translation::generation_outcome::GenerationOutcome;
+use paddler_openai_translation::generation_progress::GenerationProgress;
+use paddler_openai_translation::responses_stream::ResponsesStream;
+use paddler_openai_translation::responses_stream_event::ResponsesStreamEvent;
 
-fn event_to_sse_data(event: &ResponsesStreamEvent) -> Result<sse::Data, Error> {
+use crate::agent_relay_event::AgentRelayEvent;
+use crate::compatibility::openai_service::openai_sse_http_response::openai_sse_http_response;
+
+fn responses_stream_events(
+    responses_stream: &mut ResponsesStream,
+    agent_relay_event: AgentRelayEvent<GenerationProgress, GeneratedTokenFinalResult>,
+) -> Vec<ResponsesStreamEvent> {
+    match agent_relay_event {
+        AgentRelayEvent::Failed(relay_failure) => {
+            responses_stream.fail(relay_failure.description())
+        }
+        AgentRelayEvent::Final(final_result) => {
+            responses_stream.conclude(GenerationOutcome::from(final_result))
+        }
+        AgentRelayEvent::Partial(generation_progress) => {
+            responses_stream.advance(generation_progress)
+        }
+    }
+}
+
+fn event_to_sse_data(event: &ResponsesStreamEvent) -> Result<sse::Data, SerdeJsonError> {
     to_string(event)
         .map(|serialized_event| sse::Data::new(serialized_event).event(event.event_name()))
 }
 
-pub fn responses_sse_response<TStream>(responses_stream_events: TStream) -> HttpResponse
+pub fn responses_sse_response<TAgentRelayEvents>(
+    responses_stream: ResponsesStream,
+    agent_relay_events: TAgentRelayEvents,
+) -> HttpResponse
 where
-    TStream: Stream<Item = ResponsesStreamEvent> + 'static,
+    TAgentRelayEvents:
+        Stream<Item = AgentRelayEvent<GenerationProgress, GeneratedTokenFinalResult>> + 'static,
 {
-    let event_stream =
-        responses_stream_events.map(|event| event_to_sse_data(&event).map(sse::Event::Data));
-
-    HttpResponse::Ok()
-        .content_type("text/event-stream")
-        .insert_header((header::CACHE_CONTROL, "no-cache"))
-        .body(sse::Sse::from_stream(event_stream))
+    openai_sse_http_response(
+        agent_relay_events
+            .scan(responses_stream, |responses_stream, agent_relay_event| {
+                ready(Some(responses_stream_events(
+                    responses_stream,
+                    agent_relay_event,
+                )))
+            })
+            .flat_map(iter)
+            .map(|event| event_to_sse_data(&event).map(sse::Event::Data)),
+    )
 }
diff --git a/paddler_balancer/src/compatibility/openai_service/responses_streaming_response_transformer.rs b/paddler_balancer/src/compatibility/openai_service/responses_streaming_response_transformer.rs
deleted file mode 100644
index f13a39a3d..000000000
--- a/paddler_balancer/src/compatibility/openai_service/responses_streaming_response_transformer.rs
+++ /dev/null
@@ -1,476 +0,0 @@
-use std::mem::take;
-use std::sync::Arc;
-
-use async_trait::async_trait;
-use parking_lot::Mutex;
-
-use paddler_messaging::generated_token_result::GeneratedTokenResult;
-use paddler_messaging::generation_summary::GenerationSummary;
-use paddler_messaging::inference_client::message::Message as OutgoingMessage;
-use paddler_messaging::inference_client::response::Response as OutgoingResponse;
-use paddler_messaging::jsonrpc::response_envelope::ResponseEnvelope;
-
-use crate::agent_relay_error::AgentRelayError;
-use crate::chunk_forwarding_session_controller::transforms_outgoing_message::TransformsOutgoingMessage;
-use crate::compatibility::openai_service::openai_error::OpenAIError;
-use crate::compatibility::openai_service::response_snapshot_event::ResponseSnapshotEvent;
-use crate::compatibility::openai_service::responses_response_header::ResponsesResponseHeader;
-use crate::compatibility::openai_service::responses_stream_event::ResponsesStreamEvent;
-use crate::compatibility::openai_service::responses_streaming_state::ResponsesStreamingState;
-
-#[derive(Clone)]
-pub struct ResponsesStreamingResponseTransformer {
-    pub header: ResponsesResponseHeader,
-    pub state: Arc<Mutex<ResponsesStreamingState>>,
-}
-
-impl ResponsesStreamingResponseTransformer {
-    fn ensure_preamble(
-        &self,
-        state: &mut ResponsesStreamingState,
-        events: &mut Vec<ResponsesStreamEvent>,
-    ) {
-        if state.started {
-            return;
-        }
-
-        state.started = true;
-
-        let created_sequence_number = state.next_sequence_number();
-        events.push(ResponsesStreamEvent::Created(ResponseSnapshotEvent {
-            sequence_number: created_sequence_number,
-            response: self.header.in_progress(),
-        }));
-
-        let in_progress_sequence_number = state.next_sequence_number();
-        events.push(ResponsesStreamEvent::InProgress(ResponseSnapshotEvent {
-            sequence_number: in_progress_sequence_number,
-            response: self.header.in_progress(),
-        }));
-    }
-
-    fn handle_done(
-        &self,
-        state: &mut ResponsesStreamingState,
-        events: &mut Vec<ResponsesStreamEvent>,
-        summary: &GenerationSummary,
-    ) {
-        state.close_open_item(events);
-
-        let output = take(&mut state.finalized_output);
-        let snapshot = ResponseSnapshotEvent {
-            sequence_number: state.next_sequence_number(),
-            response: self.header.finished(output, &summary.usage, summary.finish),
-        };
-
-        events.push(if summary.finish.reached_a_length_limit() {
-            ResponsesStreamEvent::Incomplete(snapshot)
-        } else {
-            ResponsesStreamEvent::Completed(snapshot)
-        });
-    }
-}
-
-#[async_trait]
-impl TransformsOutgoingMessage for ResponsesStreamingResponseTransformer {
-    type Output = ResponsesStreamEvent;
-
-    async fn transform(
-        &self,
-        message: OutgoingMessage,
-    ) -> Result<Vec<ResponsesStreamEvent>, AgentRelayError> {
-        let mut events: Vec<ResponsesStreamEvent> = Vec::new();
-        let mut state = self.state.lock();
-
-        match message {
-            OutgoingMessage::Response(ResponseEnvelope {
-                response:
-                    OutgoingResponse::GeneratedToken(
-                        GeneratedTokenResult::ContentToken(text)
-                        | GeneratedTokenResult::UndeterminableToken(text),
-                    ),
-                ..
-            }) => {
-                self.ensure_preamble(&mut state, &mut events);
-                state.handle_content(&mut events, &text);
-            }
-            OutgoingMessage::Response(ResponseEnvelope {
-                response:
-                    OutgoingResponse::GeneratedToken(GeneratedTokenResult::ReasoningToken(text)),
-                ..
-            }) => {
-                self.ensure_preamble(&mut state, &mut events);
-                state.handle_reasoning(&mut events, &text);
-            }
-            OutgoingMessage::Response(ResponseEnvelope {
-                response: OutgoingResponse::GeneratedToken(GeneratedTokenResult::ToolCallToken(_)),
-                ..
-            }) => {}
-            OutgoingMessage::Response(ResponseEnvelope {
-                response:
-                    OutgoingResponse::GeneratedToken(GeneratedTokenResult::ToolCallParsed(parsed_calls)),
-                ..
-            }) => {
-                self.ensure_preamble(&mut state, &mut events);
-                state.handle_tool_calls(&mut events, &parsed_calls);
-            }
-            OutgoingMessage::Response(ResponseEnvelope {
-                response: OutgoingResponse::GeneratedToken(GeneratedTokenResult::Done(summary)),
-                ..
-            }) => {
-                self.ensure_preamble(&mut state, &mut events);
-                self.handle_done(&mut state, &mut events, &summary);
-            }
-            other => {
-                let error = OpenAIError::classify(&other).ok_or_else(|| {
-                    AgentRelayError::MessageNotRelayable {
-                        message: Box::new(other),
-                    }
-                })?;
-
-                self.ensure_preamble(&mut state, &mut events);
-
-                let failed_sequence_number = state.next_sequence_number();
-                events.push(ResponsesStreamEvent::Failed(ResponseSnapshotEvent {
-                    sequence_number: failed_sequence_number,
-                    response: self.header.failed(&error),
-                }));
-            }
-        }
-
-        Ok(events)
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use std::sync::Arc;
-
-    use llama_cpp_bindings_types::ParsedToolCall;
-    use llama_cpp_bindings_types::TokenUsage;
-    use llama_cpp_bindings_types::ToolCallArguments;
-    use parking_lot::Mutex;
-    use serde_json::Value;
-    use serde_json::json;
-    use serde_json::to_value;
-
-    use paddler_messaging::generated_token_result::GeneratedTokenResult;
-    use paddler_messaging::generation_finish::GenerationFinish;
-    use paddler_messaging::generation_summary::GenerationSummary;
-    use paddler_messaging::inference_client::message::Message as OutgoingMessage;
-    use paddler_messaging::inference_client::notification::Notification;
-    use paddler_messaging::inference_client::response::Response as OutgoingResponse;
-    use paddler_messaging::jsonrpc::response_envelope::ResponseEnvelope;
-    use paddler_openai_response_format_validator::openai_validator::OpenAIValidator;
-
-    use super::ResponsesStreamingResponseTransformer;
-    use super::ResponsesStreamingState;
-    use crate::agent_relay_error::AgentRelayError;
-    use crate::chunk_forwarding_session_controller::transforms_outgoing_message::TransformsOutgoingMessage;
-    use crate::compatibility::openai_service::responses_response_header::ResponsesResponseHeader;
-    use crate::compatibility::openai_service::responses_stream_event::ResponsesStreamEvent;
-
-    fn serialized(event: &ResponsesStreamEvent) -> Value {
-        to_value(event).expect("a responses stream event must serialize")
-    }
-
-    #[must_use]
-    pub fn token_message(token_result: GeneratedTokenResult) -> OutgoingMessage {
-        OutgoingMessage::Response(ResponseEnvelope {
-            generated_by: None,
-            request_id: "test-request".to_owned(),
-            response: OutgoingResponse::GeneratedToken(token_result),
-        })
-    }
-
-    #[must_use]
-    pub fn summary_with_counts(
-        prompt_tokens: u64,
-        content_tokens: u64,
-        reasoning_tokens: u64,
-    ) -> GenerationSummary {
-        GenerationSummary {
-            finish: GenerationFinish::EndOfGeneration,
-            usage: TokenUsage {
-                prompt_tokens,
-                content_tokens,
-                reasoning_tokens,
-                ..TokenUsage::default()
-            },
-        }
-    }
-
-    #[must_use]
-    pub fn weather_call() -> ParsedToolCall {
-        ParsedToolCall::new(
-            "call_x".to_owned(),
-            "get_weather".to_owned(),
-            ToolCallArguments::ValidJson(json!({ "location": "Paris" })),
-        )
-    }
-
-    #[must_use]
-    pub fn header() -> ResponsesResponseHeader {
-        ResponsesResponseHeader {
-            id: "resp_test".to_owned(),
-            created_at: 0,
-            model: "test-model".to_owned(),
-            instructions: None,
-            temperature: 0.25,
-            top_p: 0.5,
-        }
-    }
-
-    fn streaming_transformer() -> ResponsesStreamingResponseTransformer {
-        ResponsesStreamingResponseTransformer {
-            header: header(),
-            state: Arc::new(Mutex::new(ResponsesStreamingState::default())),
-        }
-    }
-
-    fn names(events: &[ResponsesStreamEvent]) -> Vec<&'static str> {
-        events
-            .iter()
-            .map(ResponsesStreamEvent::event_name)
-            .collect()
-    }
-
-    #[tokio::test]
-    async fn streaming_first_content_token_emits_preamble_then_text_delta() {
-        let transformer = streaming_transformer();
-
-        let events = transformer
-            .transform(token_message(GeneratedTokenResult::ContentToken(
-                "hi".to_owned(),
-            )))
-            .await
-            .unwrap();
-
-        assert_eq!(
-            names(&events),
-            vec![
-                "response.created",
-                "response.in_progress",
-                "response.output_item.added",
-                "response.content_part.added",
-                "response.output_text.delta",
-            ]
-        );
-        assert_eq!(serialized(&events[0])["response"]["status"], "in_progress");
-        assert_eq!(serialized(&events[4])["delta"], "hi");
-    }
-
-    #[tokio::test]
-    async fn streaming_preamble_is_emitted_only_once() {
-        let transformer = streaming_transformer();
-
-        transformer
-            .transform(token_message(GeneratedTokenResult::ContentToken(
-                "a".to_owned(),
-            )))
-            .await
-            .unwrap();
-        let events = transformer
-            .transform(token_message(GeneratedTokenResult::ContentToken(
-                "b".to_owned(),
-            )))
-            .await
-            .unwrap();
-
-        assert_eq!(names(&events), vec!["response.output_text.delta"]);
-    }
-
-    #[tokio::test]
-    async fn streaming_done_finalizes_message_and_emits_completed_with_usage() {
-        let transformer = streaming_transformer();
-
-        transformer
-            .transform(token_message(GeneratedTokenResult::ContentToken(
-                "hello".to_owned(),
-            )))
-            .await
-            .unwrap();
-        let events = transformer
-            .transform(token_message(GeneratedTokenResult::Done(
-                summary_with_counts(7, 4, 1),
-            )))
-            .await
-            .unwrap();
-
-        assert_eq!(
-            names(&events),
-            vec![
-                "response.output_text.done",
-                "response.content_part.done",
-                "response.output_item.done",
-                "response.completed",
-            ]
-        );
-
-        let completed = serialized(&events[3]);
-
-        assert_eq!(completed["response"]["status"], "completed");
-        assert_eq!(completed["response"]["usage"]["input_tokens"], 7);
-        assert_eq!(completed["response"]["usage"]["total_tokens"], 12);
-        assert_eq!(
-            completed["response"]["output"][0]["content"][0]["text"],
-            "hello"
-        );
-        assert_eq!(
-            completed["response"]["output"][0]["content"][0]["logprobs"],
-            json!([])
-        );
-    }
-
-    #[tokio::test]
-    async fn streaming_reasoning_then_content_closes_the_reasoning_item_first() {
-        let transformer = streaming_transformer();
-
-        transformer
-            .transform(token_message(GeneratedTokenResult::ReasoningToken(
-                "think".to_owned(),
-            )))
-            .await
-            .unwrap();
-        let events = transformer
-            .transform(token_message(GeneratedTokenResult::ContentToken(
-                "answer".to_owned(),
-            )))
-            .await
-            .unwrap();
-
-        assert_eq!(
-            names(&events),
-            vec![
-                "response.reasoning_text.done",
-                "response.output_item.done",
-                "response.output_item.added",
-                "response.content_part.added",
-                "response.output_text.delta",
-            ]
-        );
-        assert_eq!(serialized(&events[1])["output_index"], 0);
-        assert_eq!(serialized(&events[2])["output_index"], 1);
-        assert_eq!(serialized(&events[1])["item"]["type"], "reasoning");
-    }
-
-    #[tokio::test]
-    async fn streaming_tool_call_emits_function_call_argument_events_without_content_index() {
-        let transformer = streaming_transformer();
-
-        let events = transformer
-            .transform(token_message(GeneratedTokenResult::ToolCallParsed(vec![
-                weather_call(),
-            ])))
-            .await
-            .unwrap();
-
-        assert_eq!(
-            names(&events),
-            vec![
-                "response.created",
-                "response.in_progress",
-                "response.output_item.added",
-                "response.function_call_arguments.delta",
-                "response.function_call_arguments.done",
-                "response.output_item.done",
-            ]
-        );
-
-        let delta_event = serialized(&events[3]);
-
-        assert_eq!(delta_event["delta"], "{\"location\":\"Paris\"}");
-        assert!(
-            delta_event.get("content_index").is_none(),
-            "function_call_arguments events must not carry a content_index"
-        );
-        assert_eq!(serialized(&events[4])["name"], "get_weather");
-        assert_eq!(serialized(&events[5])["item"]["call_id"], "call_x");
-    }
-
-    #[tokio::test]
-    async fn streaming_error_emits_preamble_then_failed() {
-        let transformer = streaming_transformer();
-
-        let events = transformer
-            .transform(token_message(GeneratedTokenResult::ChatTemplateError(
-                "boom".to_owned(),
-            )))
-            .await
-            .unwrap();
-
-        assert_eq!(
-            names(&events),
-            vec![
-                "response.created",
-                "response.in_progress",
-                "response.failed"
-            ]
-        );
-
-        let failed = serialized(&events[2]);
-
-        assert_eq!(failed["response"]["status"], "failed");
-        assert_eq!(failed["response"]["error"]["code"], "server_error");
-        assert_eq!(failed["response"]["error"]["message"], "boom");
-    }
-
-    #[tokio::test]
-    async fn every_emitted_streaming_event_conforms_to_the_schema() {
-        let validator = OpenAIValidator::new().unwrap();
-        let transformer = streaming_transformer();
-
-        let mut emitted: Vec<ResponsesStreamEvent> = Vec::new();
-
-        for token in [
-            GeneratedTokenResult::ReasoningToken("ponder".to_owned()),
-            GeneratedTokenResult::ContentToken("hello".to_owned()),
-            GeneratedTokenResult::ToolCallParsed(vec![weather_call()]),
-            GeneratedTokenResult::Done(summary_with_counts(5, 3, 2)),
-        ] {
-            emitted.extend(transformer.transform(token_message(token)).await.unwrap());
-        }
-
-        assert!(emitted.len() > 10);
-
-        for event in &emitted {
-            validator
-                .validate_responses_stream_event(&serialized(event))
-                .unwrap();
-        }
-    }
-
-    #[tokio::test]
-    async fn the_failed_streaming_event_conforms_to_the_schema() {
-        let validator = OpenAIValidator::new().unwrap();
-        let transformer = streaming_transformer();
-
-        let events = transformer
-            .transform(token_message(GeneratedTokenResult::SamplerError(
-                "boom".to_owned(),
-            )))
-            .await
-            .unwrap();
-
-        for event in &events {
-            validator
-                .validate_responses_stream_event(&serialized(event))
-                .unwrap();
-        }
-    }
-
-    #[tokio::test]
-    async fn rejects_inference_socket_notifications() {
-        let transform_result = streaming_transformer()
-            .transform(OutgoingMessage::Notification(
-                Notification::TokenGenerationEnabled,
-            ))
-            .await;
-
-        assert!(matches!(
-            transform_result,
-            Err(AgentRelayError::MessageNotRelayable { message })
-                if matches!(*message, OutgoingMessage::Notification(Notification::TokenGenerationEnabled))
-        ));
-    }
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/responses_streaming_state.rs b/paddler_balancer/src/compatibility/openai_service/responses_streaming_state.rs
deleted file mode 100644
index 34e01bf4f..000000000
--- a/paddler_balancer/src/compatibility/openai_service/responses_streaming_state.rs
+++ /dev/null
@@ -1,225 +0,0 @@
-use std::mem::take;
-
-use llama_cpp_bindings_types::ParsedToolCall;
-
-use crate::compatibility::openai_service::arguments_to_tool_call_string::arguments_to_tool_call_string;
-use crate::compatibility::openai_service::content_part_event::ContentPartEvent;
-use crate::compatibility::openai_service::function_call_arguments_delta_event::FunctionCallArgumentsDeltaEvent;
-use crate::compatibility::openai_service::function_call_arguments_done_event::FunctionCallArgumentsDoneEvent;
-use crate::compatibility::openai_service::open_item::OpenItem;
-use crate::compatibility::openai_service::output_item_event::OutputItemEvent;
-use crate::compatibility::openai_service::responses_content_part::ResponsesContentPart;
-use crate::compatibility::openai_service::responses_output_item::ResponsesOutputItem;
-use crate::compatibility::openai_service::responses_output_item_kind::ResponsesOutputItemKind;
-use crate::compatibility::openai_service::responses_stream_event::ResponsesStreamEvent;
-use crate::compatibility::openai_service::text_delta_event::TextDeltaEvent;
-use crate::compatibility::openai_service::text_done_event::TextDoneEvent;
-
-#[derive(Default)]
-pub struct ResponsesStreamingState {
-    pub finalized_output: Vec<ResponsesOutputItem>,
-    open: OpenItem,
-    output_index: usize,
-    sequence_number: u64,
-    pub started: bool,
-}
-
-impl ResponsesStreamingState {
-    pub const fn next_sequence_number(&mut self) -> u64 {
-        let sequence_number = self.sequence_number;
-        self.sequence_number += 1;
-
-        sequence_number
-    }
-
-    pub fn close_open_item(&mut self, events: &mut Vec<ResponsesStreamEvent>) {
-        let output_index = self.output_index;
-        let item = match take(&mut self.open) {
-            OpenItem::Nothing => return,
-            OpenItem::Reasoning { item_id, text } => {
-                let text_done_sequence_number = self.next_sequence_number();
-                events.push(ResponsesStreamEvent::ReasoningTextDone(TextDoneEvent {
-                    sequence_number: text_done_sequence_number,
-                    item_id: item_id.clone(),
-                    output_index,
-                    content_index: 0,
-                    text: text.clone(),
-                }));
-
-                ResponsesOutputItem::completed_reasoning(item_id, text)
-            }
-            OpenItem::Message { item_id, text } => {
-                let text_done_sequence_number = self.next_sequence_number();
-                events.push(ResponsesStreamEvent::OutputTextDone(TextDoneEvent {
-                    sequence_number: text_done_sequence_number,
-                    item_id: item_id.clone(),
-                    output_index,
-                    content_index: 0,
-                    text: text.clone(),
-                }));
-
-                let part_done_sequence_number = self.next_sequence_number();
-                events.push(ResponsesStreamEvent::ContentPartDone(ContentPartEvent {
-                    sequence_number: part_done_sequence_number,
-                    item_id: item_id.clone(),
-                    output_index,
-                    content_index: 0,
-                    part: ResponsesContentPart::output_text(text.clone()),
-                }));
-
-                ResponsesOutputItem::completed_message(item_id, text)
-            }
-        };
-
-        self.finish_item(events, item);
-    }
-
-    pub fn handle_reasoning(&mut self, events: &mut Vec<ResponsesStreamEvent>, text: &str) {
-        let item_id = if let OpenItem::Reasoning {
-            item_id,
-            text: reasoning_text,
-        } = &mut self.open
-        {
-            reasoning_text.push_str(text);
-
-            item_id.clone()
-        } else {
-            self.close_open_item(events);
-
-            let item_id = ResponsesOutputItemKind::Reasoning.item_id(self.output_index);
-
-            self.add_item(
-                events,
-                ResponsesOutputItem::reasoning_in_progress(item_id.clone()),
-            );
-            self.open = OpenItem::Reasoning {
-                item_id: item_id.clone(),
-                text: text.to_owned(),
-            };
-
-            item_id
-        };
-
-        let delta_sequence_number = self.next_sequence_number();
-        events.push(ResponsesStreamEvent::ReasoningTextDelta(TextDeltaEvent {
-            sequence_number: delta_sequence_number,
-            item_id,
-            output_index: self.output_index,
-            content_index: 0,
-            delta: text.to_owned(),
-        }));
-    }
-
-    pub fn handle_content(&mut self, events: &mut Vec<ResponsesStreamEvent>, text: &str) {
-        let item_id = if let OpenItem::Message {
-            item_id,
-            text: message_text,
-        } = &mut self.open
-        {
-            message_text.push_str(text);
-
-            item_id.clone()
-        } else {
-            self.close_open_item(events);
-
-            let item_id = ResponsesOutputItemKind::Message.item_id(self.output_index);
-
-            self.add_item(
-                events,
-                ResponsesOutputItem::message_in_progress(item_id.clone()),
-            );
-
-            let part_added_sequence_number = self.next_sequence_number();
-            events.push(ResponsesStreamEvent::ContentPartAdded(ContentPartEvent {
-                sequence_number: part_added_sequence_number,
-                item_id: item_id.clone(),
-                output_index: self.output_index,
-                content_index: 0,
-                part: ResponsesContentPart::output_text(String::new()),
-            }));
-            self.open = OpenItem::Message {
-                item_id: item_id.clone(),
-                text: text.to_owned(),
-            };
-
-            item_id
-        };
-
-        let delta_sequence_number = self.next_sequence_number();
-        events.push(ResponsesStreamEvent::OutputTextDelta(TextDeltaEvent {
-            sequence_number: delta_sequence_number,
-            item_id,
-            output_index: self.output_index,
-            content_index: 0,
-            delta: text.to_owned(),
-        }));
-    }
-
-    pub fn handle_tool_calls(
-        &mut self,
-        events: &mut Vec<ResponsesStreamEvent>,
-        parsed_calls: &[ParsedToolCall],
-    ) {
-        self.close_open_item(events);
-
-        for call in parsed_calls {
-            let output_index = self.output_index;
-            let item_id = ResponsesOutputItemKind::FunctionCall.item_id(output_index);
-            let arguments = arguments_to_tool_call_string(&call.arguments);
-
-            self.add_item(
-                events,
-                ResponsesOutputItem::function_call_in_progress(item_id.clone(), call),
-            );
-
-            let delta_sequence_number = self.next_sequence_number();
-            events.push(ResponsesStreamEvent::FunctionCallArgumentsDelta(
-                FunctionCallArgumentsDeltaEvent {
-                    sequence_number: delta_sequence_number,
-                    item_id: item_id.clone(),
-                    output_index,
-                    delta: arguments.clone(),
-                },
-            ));
-
-            let done_sequence_number = self.next_sequence_number();
-            events.push(ResponsesStreamEvent::FunctionCallArgumentsDone(
-                FunctionCallArgumentsDoneEvent {
-                    sequence_number: done_sequence_number,
-                    item_id: item_id.clone(),
-                    output_index,
-                    name: call.name.clone(),
-                    arguments,
-                },
-            ));
-
-            self.finish_item(
-                events,
-                ResponsesOutputItem::completed_function_call(item_id, call),
-            );
-        }
-    }
-
-    fn add_item(&mut self, events: &mut Vec<ResponsesStreamEvent>, item: ResponsesOutputItem) {
-        let added_sequence_number = self.next_sequence_number();
-
-        events.push(ResponsesStreamEvent::OutputItemAdded(OutputItemEvent {
-            sequence_number: added_sequence_number,
-            output_index: self.output_index,
-            item,
-        }));
-    }
-
-    fn finish_item(&mut self, events: &mut Vec<ResponsesStreamEvent>, item: ResponsesOutputItem) {
-        let item_done_sequence_number = self.next_sequence_number();
-
-        events.push(ResponsesStreamEvent::OutputItemDone(OutputItemEvent {
-            sequence_number: item_done_sequence_number,
-            output_index: self.output_index,
-            item: item.clone(),
-        }));
-
-        self.finalized_output.push(item);
-        self.output_index += 1;
-    }
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/stream_options.rs b/paddler_balancer/src/compatibility/openai_service/stream_options.rs
deleted file mode 100644
index 41de73fc4..000000000
--- a/paddler_balancer/src/compatibility/openai_service/stream_options.rs
+++ /dev/null
@@ -1,8 +0,0 @@
-use serde::Deserialize;
-
-#[derive(Default, Deserialize)]
-#[serde(deny_unknown_fields)]
-pub struct StreamOptions {
-    #[serde(default)]
-    pub include_usage: bool,
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/timestamp_from.rs b/paddler_balancer/src/compatibility/openai_service/timestamp_from.rs
deleted file mode 100644
index 1826d1ede..000000000
--- a/paddler_balancer/src/compatibility/openai_service/timestamp_from.rs
+++ /dev/null
@@ -1,35 +0,0 @@
-use std::time::SystemTime;
-use std::time::UNIX_EPOCH;
-
-use anyhow::Context as _;
-use anyhow::Result;
-
-pub fn timestamp_from(now: SystemTime) -> Result<u64> {
-    Ok(now
-        .duration_since(UNIX_EPOCH)
-        .context("system time is before the Unix epoch")?
-        .as_secs())
-}
-
-#[cfg(test)]
-mod tests {
-    use std::time::Duration;
-    use std::time::SystemTime;
-    use std::time::UNIX_EPOCH;
-
-    use super::timestamp_from;
-
-    #[test]
-    fn returns_seconds_since_epoch() {
-        let timestamp = timestamp_from(SystemTime::now()).unwrap();
-
-        assert!(timestamp > 0);
-    }
-
-    #[test]
-    fn errors_before_the_unix_epoch() {
-        let before_epoch = UNIX_EPOCH - Duration::from_secs(1);
-
-        assert!(timestamp_from(before_epoch).is_err());
-    }
-}
diff --git a/paddler_balancer/src/compatibility/openai_service/try_universal_error_chunk.rs b/paddler_balancer/src/compatibility/openai_service/try_universal_error_chunk.rs
deleted file mode 100644
index b00100956..000000000
--- a/paddler_balancer/src/compatibility/openai_service/try_universal_error_chunk.rs
+++ /dev/null
@@ -1,10 +0,0 @@
-use paddler_messaging::inference_client::message::Message as OutgoingMessage;
-
-use crate::chunk_forwarding_session_controller::transform_result::TransformResult;
-use crate::compatibility::openai_service::openai_error::OpenAIError;
-
-#[must_use]
-pub fn try_universal_error_chunk(message: &OutgoingMessage) -> Option<TransformResult> {
-    OpenAIError::classify(message)
-        .map(|error| TransformResult::Error(error.to_envelope().to_string()))
-}
diff --git a/paddler_balancer/src/compatibility/serves_compatibility_layer.rs b/paddler_balancer/src/compatibility/serves_compatibility_layer.rs
new file mode 100644
index 000000000..5b2a1fd2a
--- /dev/null
+++ b/paddler_balancer/src/compatibility/serves_compatibility_layer.rs
@@ -0,0 +1,9 @@
+use actix_web::web::ServiceConfig;
+
+pub trait ServesCompatibilityLayer: Send + Sync + 'static {
+    const EXTRA_ALLOWED_HEADERS: &'static [&'static str];
+    const REQUEST_ID_HEADER: &'static str;
+    const SERVICE_NAME: &'static str;
+
+    fn configure(service_config: &mut ServiceConfig);
+}
diff --git a/paddler_balancer/src/compatibility/typesafe_service/cluster_serves_no_decisions_http_response.rs b/paddler_balancer/src/compatibility/typesafe_service/cluster_serves_no_decisions_http_response.rs
new file mode 100644
index 000000000..5f2fe584b
--- /dev/null
+++ b/paddler_balancer/src/compatibility/typesafe_service/cluster_serves_no_decisions_http_response.rs
@@ -0,0 +1,15 @@
+use actix_web::HttpResponse;
+
+use crate::cluster_serves_no_decisions::ClusterServesNoDecisions;
+use crate::compatibility::typesafe_service::typesafe_error_body::TypeSafeErrorBody;
+use crate::compatibility::upstream_failure::UpstreamFailure;
+
+#[must_use]
+pub fn cluster_serves_no_decisions_http_response(
+    cluster_serves_no_decisions: ClusterServesNoDecisions,
+) -> HttpResponse {
+    TypeSafeErrorBody {
+        detail: cluster_serves_no_decisions.to_string(),
+    }
+    .into_http_response(UpstreamFailure::Unavailable.status_code())
+}
diff --git a/paddler_balancer/src/compatibility/typesafe_service/decision_outcome.rs b/paddler_balancer/src/compatibility/typesafe_service/decision_outcome.rs
new file mode 100644
index 000000000..12b3355e4
--- /dev/null
+++ b/paddler_balancer/src/compatibility/typesafe_service/decision_outcome.rs
@@ -0,0 +1,154 @@
+use actix_web::http::StatusCode;
+
+use paddler_messaging::decision_final_result::DecisionFinalResult;
+use paddler_messaging::decision_summary::DecisionSummary;
+use paddler_messaging::oversized_decision_details::OversizedDecisionDetails;
+
+use crate::compatibility::typesafe_service::typesafe_decision_failure::TypeSafeDecisionFailure;
+use crate::compatibility::upstream_failure::UpstreamFailure;
+
+const fn failure(status_code: StatusCode, detail: String) -> DecisionOutcome {
+    DecisionOutcome::Failed(TypeSafeDecisionFailure {
+        detail,
+        status_code,
+    })
+}
+
+#[derive(Debug, Eq, PartialEq)]
+pub enum DecisionOutcome {
+    Failed(TypeSafeDecisionFailure),
+    Finished(DecisionSummary),
+}
+
+impl From<DecisionFinalResult> for DecisionOutcome {
+    fn from(decision_final_result: DecisionFinalResult) -> Self {
+        match decision_final_result {
+            DecisionFinalResult::Done(summary) => Self::Finished(summary),
+            DecisionFinalResult::RequestExceedsContext(OversizedDecisionDetails {
+                context_size,
+                required_tokens,
+            }) => failure(
+                StatusCode::UNPROCESSABLE_ENTITY,
+                format!(
+                    "the request needs {required_tokens} tokens in the cache, more than the {context_size} the context holds"
+                ),
+            ),
+            DecisionFinalResult::InputTokenizationFailed(detail) => {
+                failure(StatusCode::UNPROCESSABLE_ENTITY, detail)
+            }
+            DecisionFinalResult::DecisionsNotServed(detail)
+            | DecisionFinalResult::ModelNotLoaded(detail)
+            | DecisionFinalResult::SchedulerUnavailable(detail) => {
+                failure(UpstreamFailure::Unavailable.status_code(), detail)
+            }
+            DecisionFinalResult::AgentRuntimeFailed(detail)
+            | DecisionFinalResult::BatchAssemblyFailed(detail)
+            | DecisionFinalResult::DecodeFailed(detail)
+            | DecisionFinalResult::HiddenStateUnavailable(detail)
+            | DecisionFinalResult::KvCacheCopyFailed(detail)
+            | DecisionFinalResult::KvCacheRemovalFailed(detail) => {
+                failure(UpstreamFailure::AgentFailed.status_code(), detail)
+            }
+            DecisionFinalResult::StopRequested => failure(
+                UpstreamFailure::RelayFailed.status_code(),
+                "the agent stopped before it finished the decision".to_owned(),
+            ),
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use actix_web::http::StatusCode;
+
+    use paddler_messaging::decision_final_result::DecisionFinalResult;
+    use paddler_messaging::decision_summary::DecisionSummary;
+    use paddler_messaging::oversized_decision_details::OversizedDecisionDetails;
+
+    use super::DecisionOutcome;
+    use crate::compatibility::typesafe_service::typesafe_decision_failure::TypeSafeDecisionFailure;
+
+    fn failed(status_code: StatusCode, detail: &str) -> DecisionOutcome {
+        DecisionOutcome::Failed(TypeSafeDecisionFailure {
+            detail: detail.to_owned(),
+            status_code,
+        })
+    }
+
+    #[test]
+    fn classifies_every_final_decision_result() {
+        let summary = DecisionSummary {
+            input_tokens: 1,
+            processing_milliseconds: 1,
+        };
+
+        for (decision_final_result, expected_outcome) in [
+            (
+                DecisionFinalResult::Done(summary.clone()),
+                DecisionOutcome::Finished(summary),
+            ),
+            (
+                DecisionFinalResult::RequestExceedsContext(OversizedDecisionDetails {
+                    context_size: 128,
+                    required_tokens: 410,
+                }),
+                failed(
+                    StatusCode::UNPROCESSABLE_ENTITY,
+                    "the request needs 410 tokens in the cache, more than the 128 the context holds",
+                ),
+            ),
+            (
+                DecisionFinalResult::InputTokenizationFailed("nul byte".to_owned()),
+                failed(StatusCode::UNPROCESSABLE_ENTITY, "nul byte"),
+            ),
+            (
+                DecisionFinalResult::DecisionsNotServed("no pointer head".to_owned()),
+                failed(StatusCode::SERVICE_UNAVAILABLE, "no pointer head"),
+            ),
+            (
+                DecisionFinalResult::ModelNotLoaded("no model".to_owned()),
+                failed(StatusCode::SERVICE_UNAVAILABLE, "no model"),
+            ),
+            (
+                DecisionFinalResult::SchedulerUnavailable("stopped".to_owned()),
+                failed(StatusCode::SERVICE_UNAVAILABLE, "stopped"),
+            ),
+            (
+                DecisionFinalResult::AgentRuntimeFailed("runtime".to_owned()),
+                failed(StatusCode::INTERNAL_SERVER_ERROR, "runtime"),
+            ),
+            (
+                DecisionFinalResult::BatchAssemblyFailed("batch".to_owned()),
+                failed(StatusCode::INTERNAL_SERVER_ERROR, "batch"),
+            ),
+            (
+                DecisionFinalResult::DecodeFailed("decode".to_owned()),
+                failed(StatusCode::INTERNAL_SERVER_ERROR, "decode"),
+            ),
+            (
+                DecisionFinalResult::HiddenStateUnavailable("hidden".to_owned()),
+                failed(StatusCode::INTERNAL_SERVER_ERROR, "hidden"),
+            ),
+            (
+                DecisionFinalResult::KvCacheCopyFailed("copy".to_owned()),
+                failed(StatusCode::INTERNAL_SERVER_ERROR, "copy"),
+            ),
+            (
+                DecisionFinalResult::KvCacheRemovalFailed("removal".to_owned()),
+                failed(StatusCode::INTERNAL_SERVER_ERROR, "removal"),
+            ),
+            (
+                DecisionFinalResult::StopRequested,
+                failed(
+                    StatusCode::BAD_GATEWAY,
+                    "the agent stopped before it finished the decision",
+                ),
+            ),
+        ] {
+            assert_eq!(
+                DecisionOutcome::from(decision_final_result),
+                expected_outcome
+            );
+        }
+    }
+}
diff --git a/paddler_balancer/src/compatibility/typesafe_service/http_route/get_models.rs b/paddler_balancer/src/compatibility/typesafe_service/http_route/get_models.rs
new file mode 100644
index 000000000..7d7aba126
--- /dev/null
+++ b/paddler_balancer/src/compatibility/typesafe_service/http_route/get_models.rs
@@ -0,0 +1,31 @@
+use actix_web::HttpResponse;
+use actix_web::web;
+use actix_web::web::get;
+
+use paddler_messaging::request_params::decide_params::DecideParams;
+
+use crate::agent_streaming_request::AgentStreamingRequest as _;
+use crate::compatibility::compatibility_app_data::CompatibilityAppData;
+use crate::compatibility::typesafe_service::cluster_serves_no_decisions_http_response::cluster_serves_no_decisions_http_response;
+use crate::compatibility::typesafe_service::typesafe_api_path::TypeSafeApiPath;
+use crate::compatibility::typesafe_service::typesafe_models::TypeSafeModels;
+
+async fn respond(app_data: web::Data<CompatibilityAppData>) -> HttpResponse {
+    match app_data
+        .balancer_applicable_state_holder
+        .require_decisions()
+    {
+        Ok(()) => HttpResponse::Ok().json(TypeSafeModels::from(
+            app_data
+                .agent_controller_pool
+                .model_paths_serving(DecideParams::INFERENCE_CAPABILITY),
+        )),
+        Err(cluster_serves_no_decisions) => {
+            cluster_serves_no_decisions_http_response(cluster_serves_no_decisions)
+        }
+    }
+}
+
+pub fn get_models(cfg: &mut web::ServiceConfig) {
+    cfg.route(TypeSafeApiPath::MODELS, get().to(respond));
+}
diff --git a/paddler_balancer/src/compatibility/typesafe_service/http_route/mod.rs b/paddler_balancer/src/compatibility/typesafe_service/http_route/mod.rs
new file mode 100644
index 000000000..48dde0af0
--- /dev/null
+++ b/paddler_balancer/src/compatibility/typesafe_service/http_route/mod.rs
@@ -0,0 +1,2 @@
+pub mod get_models;
+pub mod post_system_one;
diff --git a/paddler_balancer/src/compatibility/typesafe_service/http_route/post_system_one.rs b/paddler_balancer/src/compatibility/typesafe_service/http_route/post_system_one.rs
new file mode 100644
index 000000000..acad6c38d
--- /dev/null
+++ b/paddler_balancer/src/compatibility/typesafe_service/http_route/post_system_one.rs
@@ -0,0 +1,42 @@
+use actix_web::HttpResponse;
+use actix_web::http::StatusCode;
+use actix_web::web;
+use actix_web::web::post;
+
+use paddler_typesafe_translation::system_one_request::SystemOneRequest;
+use paddler_typesafe_translation::translated_system_one_request::TranslatedSystemOneRequest;
+
+use crate::compatibility::compatibility_app_data::CompatibilityAppData;
+use crate::compatibility::typesafe_service::cluster_serves_no_decisions_http_response::cluster_serves_no_decisions_http_response;
+use crate::compatibility::typesafe_service::system_one_http_response::system_one_http_response;
+use crate::compatibility::typesafe_service::typesafe_api_path::TypeSafeApiPath;
+use crate::compatibility::typesafe_service::typesafe_error_body::TypeSafeErrorBody;
+use crate::compatibility::typesafe_service::typesafe_json_config::typesafe_json_config;
+
+async fn respond(
+    app_data: web::Data<CompatibilityAppData>,
+    system_one_request: web::Json<SystemOneRequest>,
+) -> HttpResponse {
+    if let Err(cluster_serves_no_decisions) = app_data
+        .balancer_applicable_state_holder
+        .require_decisions()
+    {
+        return cluster_serves_no_decisions_http_response(cluster_serves_no_decisions);
+    }
+
+    match system_one_request.into_inner().translate() {
+        Ok(TranslatedSystemOneRequest {
+            decide_params,
+            response_layout,
+        }) => system_one_http_response(&response_layout, app_data.agent_relay(decide_params)).await,
+        Err(translation_error) => TypeSafeErrorBody {
+            detail: translation_error.to_string(),
+        }
+        .into_http_response(StatusCode::UNPROCESSABLE_ENTITY),
+    }
+}
+
+pub fn post_system_one(cfg: &mut web::ServiceConfig) {
+    cfg.app_data(typesafe_json_config())
+        .route(TypeSafeApiPath::SYSTEM_ONE, post().to(respond));
+}
diff --git a/paddler_balancer/src/compatibility/typesafe_service/mod.rs b/paddler_balancer/src/compatibility/typesafe_service/mod.rs
new file mode 100644
index 000000000..c5368ee6b
--- /dev/null
+++ b/paddler_balancer/src/compatibility/typesafe_service/mod.rs
@@ -0,0 +1,12 @@
+pub mod cluster_serves_no_decisions_http_response;
+pub mod decision_outcome;
+pub mod http_route;
+pub mod system_one_http_response;
+pub mod typesafe_api_path;
+pub mod typesafe_compatibility_layer;
+pub mod typesafe_decision_failure;
+pub mod typesafe_error_body;
+pub mod typesafe_header;
+pub mod typesafe_json_config;
+pub mod typesafe_model;
+pub mod typesafe_models;
diff --git a/paddler_balancer/src/compatibility/typesafe_service/system_one_http_response.rs b/paddler_balancer/src/compatibility/typesafe_service/system_one_http_response.rs
new file mode 100644
index 000000000..652b1ee0d
--- /dev/null
+++ b/paddler_balancer/src/compatibility/typesafe_service/system_one_http_response.rs
@@ -0,0 +1,56 @@
+use actix_web::HttpResponse;
+
+use paddler_messaging::decision_result::DecisionResult;
+use paddler_typesafe_translation::system_one_response_layout::SystemOneResponseLayout;
+
+use crate::agent_relay::AgentRelay;
+use crate::agent_relay_conclusion::AgentRelayConclusion;
+use crate::agent_relay_step::AgentRelayStep;
+use crate::agent_streaming_request::AgentStreamingRequest;
+use crate::compatibility::typesafe_service::decision_outcome::DecisionOutcome;
+use crate::compatibility::typesafe_service::typesafe_decision_failure::TypeSafeDecisionFailure;
+
+pub async fn system_one_http_response<TParams>(
+    response_layout: &SystemOneResponseLayout,
+    mut agent_relay: AgentRelay<TParams>,
+) -> HttpResponse
+where
+    TParams: AgentStreamingRequest<Response = DecisionResult>,
+{
+    let mut answers = Vec::new();
+
+    loop {
+        match agent_relay.next().await {
+            AgentRelayStep::Concluded(AgentRelayConclusion::AgentConcluded {
+                final_result,
+                ..
+            }) => {
+                return match DecisionOutcome::from(final_result) {
+                    DecisionOutcome::Failed(typesafe_decision_failure) => {
+                        typesafe_decision_failure.into_http_response()
+                    }
+                    DecisionOutcome::Finished(summary) => {
+                        match response_layout.respond(&answers, &summary) {
+                            Ok(system_one_response) => HttpResponse::Ok().json(system_one_response),
+                            Err(translation_error) => {
+                                TypeSafeDecisionFailure::from(translation_error)
+                                    .into_http_response()
+                            }
+                        }
+                    }
+                };
+            }
+            AgentRelayStep::Concluded(AgentRelayConclusion::Failed(relay_failure)) => {
+                return TypeSafeDecisionFailure::from(relay_failure).into_http_response();
+            }
+            AgentRelayStep::Partial {
+                partial_result: answer,
+                relay,
+                ..
+            } => {
+                answers.push(answer);
+                agent_relay = relay;
+            }
+        }
+    }
+}
diff --git a/paddler_balancer/src/compatibility/typesafe_service/typesafe_api_path.rs b/paddler_balancer/src/compatibility/typesafe_service/typesafe_api_path.rs
new file mode 100644
index 000000000..41f410978
--- /dev/null
+++ b/paddler_balancer/src/compatibility/typesafe_service/typesafe_api_path.rs
@@ -0,0 +1,6 @@
+pub struct TypeSafeApiPath;
+
+impl TypeSafeApiPath {
+    pub const MODELS: &str = "/v1/models";
+    pub const SYSTEM_ONE: &str = "/v1/systemone";
+}
diff --git a/paddler_balancer/src/compatibility/typesafe_service/typesafe_compatibility_layer.rs b/paddler_balancer/src/compatibility/typesafe_service/typesafe_compatibility_layer.rs
new file mode 100644
index 000000000..b70b463c7
--- /dev/null
+++ b/paddler_balancer/src/compatibility/typesafe_service/typesafe_compatibility_layer.rs
@@ -0,0 +1,24 @@
+use actix_web::web::ServiceConfig;
+
+use crate::compatibility::serves_compatibility_layer::ServesCompatibilityLayer;
+use crate::compatibility::typesafe_service::http_route::get_models::get_models;
+use crate::compatibility::typesafe_service::http_route::post_system_one::post_system_one;
+use crate::compatibility::typesafe_service::typesafe_header::TypeSafeHeader;
+
+pub struct TypeSafeCompatibilityLayer;
+
+impl ServesCompatibilityLayer for TypeSafeCompatibilityLayer {
+    const EXTRA_ALLOWED_HEADERS: &'static [&'static str] = &[
+        TypeSafeHeader::RETRY_COUNT,
+        TypeSafeHeader::RUNTIME,
+        TypeSafeHeader::SDK,
+    ];
+    const REQUEST_ID_HEADER: &'static str = TypeSafeHeader::REQUEST_ID;
+    const SERVICE_NAME: &'static str = "balancer::compatibility::typesafe_service";
+
+    fn configure(service_config: &mut ServiceConfig) {
+        service_config
+            .configure(get_models)
+            .configure(post_system_one);
+    }
+}
diff --git a/paddler_balancer/src/compatibility/typesafe_service/typesafe_decision_failure.rs b/paddler_balancer/src/compatibility/typesafe_service/typesafe_decision_failure.rs
new file mode 100644
index 000000000..966827fab
--- /dev/null
+++ b/paddler_balancer/src/compatibility/typesafe_service/typesafe_decision_failure.rs
@@ -0,0 +1,42 @@
+use actix_web::HttpResponse;
+use actix_web::http::StatusCode;
+
+use paddler_typesafe_translation::typesafe_translation_error::TypeSafeTranslationError;
+
+use crate::compatibility::typesafe_service::typesafe_error_body::TypeSafeErrorBody;
+use crate::compatibility::upstream_failure::UpstreamFailure;
+use crate::relay_failure::RelayFailure;
+
+#[derive(Debug, Eq, PartialEq)]
+pub struct TypeSafeDecisionFailure {
+    pub detail: String,
+    pub status_code: StatusCode,
+}
+
+impl TypeSafeDecisionFailure {
+    #[must_use]
+    pub fn into_http_response(self) -> HttpResponse {
+        TypeSafeErrorBody {
+            detail: self.detail,
+        }
+        .into_http_response(self.status_code)
+    }
+}
+
+impl From<RelayFailure> for TypeSafeDecisionFailure {
+    fn from(relay_failure: RelayFailure) -> Self {
+        Self {
+            detail: relay_failure.description(),
+            status_code: UpstreamFailure::from(&relay_failure).status_code(),
+        }
+    }
+}
+
+impl From<TypeSafeTranslationError> for TypeSafeDecisionFailure {
+    fn from(translation_error: TypeSafeTranslationError) -> Self {
+        Self {
+            detail: translation_error.to_string(),
+            status_code: UpstreamFailure::RelayFailed.status_code(),
+        }
+    }
+}
diff --git a/paddler_balancer/src/compatibility/typesafe_service/typesafe_error_body.rs b/paddler_balancer/src/compatibility/typesafe_service/typesafe_error_body.rs
new file mode 100644
index 000000000..babf4aaa1
--- /dev/null
+++ b/paddler_balancer/src/compatibility/typesafe_service/typesafe_error_body.rs
@@ -0,0 +1,15 @@
+use actix_web::HttpResponse;
+use actix_web::http::StatusCode;
+use serde::Serialize;
+
+#[derive(Serialize)]
+pub struct TypeSafeErrorBody {
+    pub detail: String,
+}
+
+impl TypeSafeErrorBody {
+    #[must_use]
+    pub fn into_http_response(self, status_code: StatusCode) -> HttpResponse {
+        HttpResponse::build(status_code).json(self)
+    }
+}
diff --git a/paddler_balancer/src/compatibility/typesafe_service/typesafe_header.rs b/paddler_balancer/src/compatibility/typesafe_service/typesafe_header.rs
new file mode 100644
index 000000000..83f72ed65
--- /dev/null
+++ b/paddler_balancer/src/compatibility/typesafe_service/typesafe_header.rs
@@ -0,0 +1,8 @@
+pub struct TypeSafeHeader;
+
+impl TypeSafeHeader {
+    pub const REQUEST_ID: &str = "x-typesafe-request-id";
+    pub const RETRY_COUNT: &str = "x-typesafe-retry-count";
+    pub const RUNTIME: &str = "x-typesafe-runtime";
+    pub const SDK: &str = "x-typesafe-sdk";
+}
diff --git a/paddler_balancer/src/compatibility/typesafe_service/typesafe_json_config.rs b/paddler_balancer/src/compatibility/typesafe_service/typesafe_json_config.rs
new file mode 100644
index 000000000..8f0cf1fd7
--- /dev/null
+++ b/paddler_balancer/src/compatibility/typesafe_service/typesafe_json_config.rs
@@ -0,0 +1,17 @@
+use actix_web::error::InternalError;
+use actix_web::http::StatusCode;
+use actix_web::web::JsonConfig;
+
+use crate::compatibility::typesafe_service::typesafe_error_body::TypeSafeErrorBody;
+
+#[must_use]
+pub fn typesafe_json_config() -> JsonConfig {
+    JsonConfig::default().error_handler(|json_payload_error, _request| {
+        let unprocessable_entity = TypeSafeErrorBody {
+            detail: json_payload_error.to_string(),
+        }
+        .into_http_response(StatusCode::UNPROCESSABLE_ENTITY);
+
+        InternalError::from_response(json_payload_error, unprocessable_entity).into()
+    })
+}
diff --git a/paddler_balancer/src/compatibility/typesafe_service/typesafe_model.rs b/paddler_balancer/src/compatibility/typesafe_service/typesafe_model.rs
new file mode 100644
index 000000000..fc3c63a3e
--- /dev/null
+++ b/paddler_balancer/src/compatibility/typesafe_service/typesafe_model.rs
@@ -0,0 +1,21 @@
+use serde::Serialize;
+
+const RELEASE_DATE_UNKNOWN: &str = "unknown";
+
+#[derive(Serialize)]
+pub struct TypeSafeModel {
+    pub name: String,
+    pub description: String,
+    pub release_date: String,
+}
+
+impl TypeSafeModel {
+    #[must_use]
+    pub fn named(name: String) -> Self {
+        Self {
+            description: format!("Decision model {name} served by Paddler"),
+            name,
+            release_date: RELEASE_DATE_UNKNOWN.to_owned(),
+        }
+    }
+}
diff --git a/paddler_balancer/src/compatibility/typesafe_service/typesafe_models.rs b/paddler_balancer/src/compatibility/typesafe_service/typesafe_models.rs
new file mode 100644
index 000000000..6af009084
--- /dev/null
+++ b/paddler_balancer/src/compatibility/typesafe_service/typesafe_models.rs
@@ -0,0 +1,21 @@
+use std::collections::BTreeSet;
+
+use serde::Serialize;
+
+use crate::compatibility::typesafe_service::typesafe_model::TypeSafeModel;
+
+#[derive(Serialize)]
+pub struct TypeSafeModels {
+    pub models: Vec<TypeSafeModel>,
+}
+
+impl From<BTreeSet<String>> for TypeSafeModels {
+    fn from(served_model_paths: BTreeSet<String>) -> Self {
+        Self {
+            models: served_model_paths
+                .into_iter()
+                .map(TypeSafeModel::named)
+                .collect(),
+        }
+    }
+}
diff --git a/paddler_balancer/src/compatibility/upstream_failure.rs b/paddler_balancer/src/compatibility/upstream_failure.rs
new file mode 100644
index 000000000..d6b2beadb
--- /dev/null
+++ b/paddler_balancer/src/compatibility/upstream_failure.rs
@@ -0,0 +1,90 @@
+use actix_web::http::StatusCode;
+
+use crate::relay_failure::RelayFailure;
+
+#[derive(Debug, Eq, PartialEq)]
+pub enum UpstreamFailure {
+    AgentFailed,
+    RelayFailed,
+    TimedOut,
+    Unavailable,
+}
+
+impl UpstreamFailure {
+    #[must_use]
+    pub const fn status_code(&self) -> StatusCode {
+        match self {
+            Self::AgentFailed => StatusCode::INTERNAL_SERVER_ERROR,
+            Self::RelayFailed => StatusCode::BAD_GATEWAY,
+            Self::TimedOut => StatusCode::GATEWAY_TIMEOUT,
+            Self::Unavailable => StatusCode::SERVICE_UNAVAILABLE,
+        }
+    }
+}
+
+impl From<&RelayFailure> for UpstreamFailure {
+    fn from(relay_failure: &RelayFailure) -> Self {
+        match relay_failure {
+            RelayFailure::AgentDisconnected => Self::RelayFailed,
+            RelayFailure::BalancerShuttingDown | RelayFailure::BufferOverflow => Self::Unavailable,
+            RelayFailure::BufferedRequestTimedOut | RelayFailure::ItemTimedOut { .. } => {
+                Self::TimedOut
+            }
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::time::Duration;
+
+    use actix_web::http::StatusCode;
+
+    use super::UpstreamFailure;
+    use crate::relay_failure::RelayFailure;
+
+    #[test]
+    fn maps_relay_failures_to_upstream_failures() {
+        for (relay_failure, expected_failure) in [
+            (
+                RelayFailure::AgentDisconnected,
+                UpstreamFailure::RelayFailed,
+            ),
+            (
+                RelayFailure::BalancerShuttingDown,
+                UpstreamFailure::Unavailable,
+            ),
+            (RelayFailure::BufferOverflow, UpstreamFailure::Unavailable),
+            (
+                RelayFailure::BufferedRequestTimedOut,
+                UpstreamFailure::TimedOut,
+            ),
+            (
+                RelayFailure::ItemTimedOut {
+                    inference_item_timeout: Duration::from_millis(250),
+                },
+                UpstreamFailure::TimedOut,
+            ),
+        ] {
+            assert_eq!(UpstreamFailure::from(&relay_failure), expected_failure);
+        }
+    }
+
+    #[test]
+    fn maps_upstream_failures_to_http_statuses() {
+        for (failure, expected_status) in [
+            (
+                UpstreamFailure::AgentFailed,
+                StatusCode::INTERNAL_SERVER_ERROR,
+            ),
+            (UpstreamFailure::RelayFailed, StatusCode::BAD_GATEWAY),
+            (UpstreamFailure::TimedOut, StatusCode::GATEWAY_TIMEOUT),
+            (
+                UpstreamFailure::Unavailable,
+                StatusCode::SERVICE_UNAVAILABLE,
+            ),
+        ] {
+            assert_eq!(failure.status_code(), expected_status);
+        }
+    }
+}
diff --git a/paddler_balancer/src/continuation_decision.rs b/paddler_balancer/src/continuation_decision.rs
index efe3fcadc..ae58dd2b7 100644
--- a/paddler_balancer/src/continuation_decision.rs
+++ b/paddler_balancer/src/continuation_decision.rs
@@ -1,5 +1,6 @@
 use crate::websocket_close_cause::WebSocketCloseCause;
 
+#[derive(Debug, Eq, PartialEq)]
 pub enum ContinuationDecision {
     Continue,
     Stop(WebSocketCloseCause),
diff --git a/paddler_balancer/src/controls_session.rs b/paddler_balancer/src/controls_session.rs
index 899c3f819..59c236032 100644
--- a/paddler_balancer/src/controls_session.rs
+++ b/paddler_balancer/src/controls_session.rs
@@ -14,7 +14,7 @@ where
     fn send_response_safe(&mut self, message: TResponse) -> impl Future<Output = ()> + Send {
         async move {
             if let Err(err) = self.send_response(message).await {
-                error!("Failed to send response: {err}");
+                error!("Failed to send response: {err:#}");
             }
         }
     }
diff --git a/paddler_balancer/src/controls_websocket_endpoint.rs b/paddler_balancer/src/controls_websocket_endpoint.rs
index 42a8b3cc8..decff16c8 100644
--- a/paddler_balancer/src/controls_websocket_endpoint.rs
+++ b/paddler_balancer/src/controls_websocket_endpoint.rs
@@ -8,7 +8,6 @@ use actix_web::rt;
 use actix_web::web::Payload;
 use actix_ws::AggregatedMessage;
 use actix_ws::ProtocolError;
-use actix_ws::Session;
 use actix_ws::handle;
 use anyhow::Result;
 use futures_util::StreamExt as _;
@@ -20,15 +19,23 @@ use serde_json::from_str;
 use tokio::select;
 use tokio::sync::mpsc;
 use tokio::sync::mpsc::UnboundedSender;
+use tokio::time::Instant;
+use tokio::time::MissedTickBehavior;
+use tokio::time::interval_at;
 use tokio_util::sync::CancellationToken;
 
+use paddler_messaging::ping_interval::PING_INTERVAL;
+use paddler_messaging::ping_liveness::PingLiveness;
+use paddler_messaging::ping_turn::PingTurn;
 use paddler_messaging::rpc_message::RpcMessage;
 
 use crate::continuation_decision::ContinuationDecision;
+use crate::enqueue_websocket_frame::enqueue_websocket_frame;
 use crate::max_websocket_message_size::MAX_WEBSOCKET_MESSAGE_SIZE;
-use crate::session_keep_alive::SessionKeepAlive;
 use crate::websocket_close_cause::WebSocketCloseCause;
+use crate::websocket_outbound_frame::WebSocketOutboundFrame;
 use crate::websocket_session_controller::WebSocketSessionController;
+use crate::write_websocket_session::write_websocket_session;
 
 pub trait ControlsWebSocketEndpoint: Send + Sync + 'static {
     type Context: Send + Sync + 'static;
@@ -54,7 +61,8 @@ pub trait ControlsWebSocketEndpoint: Send + Sync + 'static {
         connection_close: CancellationToken,
         context: Arc<Self::Context>,
         msg: Option<Result<AggregatedMessage, ProtocolError>>,
-        session: &mut Session,
+        outbound_tx: &UnboundedSender<WebSocketOutboundFrame>,
+        ping_liveness: &mut PingLiveness,
         continuation_stop_tx: UnboundedSender<WebSocketCloseCause>,
     ) -> impl Future<Output = ContinuationDecision> + Send {
         async move {
@@ -68,18 +76,21 @@ pub trait ControlsWebSocketEndpoint: Send + Sync + 'static {
                     ContinuationDecision::Stop(WebSocketCloseCause::PeerClosedConnection)
                 }
                 Some(Ok(AggregatedMessage::Ping(ping_payload))) => {
-                    session.pong(&ping_payload).await.map_or(
-                        ContinuationDecision::Stop(WebSocketCloseCause::SessionAlreadyClosed),
-                        |()| ContinuationDecision::Continue,
-                    )
+                    enqueue_websocket_frame(outbound_tx, WebSocketOutboundFrame::Pong(ping_payload))
+                }
+                Some(Ok(AggregatedMessage::Pong(_))) => {
+                    ping_liveness.record_pong();
+
+                    ContinuationDecision::Continue
                 }
-                Some(Ok(AggregatedMessage::Pong(_))) => ContinuationDecision::Continue,
                 Some(Ok(AggregatedMessage::Text(text))) => {
                     Self::handle_text_message(
                         connection_close,
                         context,
                         &text,
-                        WebSocketSessionController::<Self::OutgoingMessage>::new(session.clone()),
+                        WebSocketSessionController::<Self::OutgoingMessage>::new(
+                            outbound_tx.clone(),
+                        ),
                         continuation_stop_tx,
                     )
                     .await
@@ -139,7 +150,7 @@ pub trait ControlsWebSocketEndpoint: Send + Sync + 'static {
     fn on_connection_start(
         connection_close: CancellationToken,
         context: Arc<Self::Context>,
-        session: &mut Session,
+        outbound_tx: &UnboundedSender<WebSocketOutboundFrame>,
     ) -> impl Future<Output = ()> + Send;
 
     fn respond(
@@ -150,7 +161,7 @@ pub trait ControlsWebSocketEndpoint: Send + Sync + 'static {
     ) -> Result<HttpResponse, Error> {
         let connection_close = CancellationToken::new();
         let context = Arc::new(self.create_context());
-        let (res, mut session, msg_stream) = handle(&req, payload)?;
+        let (res, session, msg_stream) = handle(&req, payload)?;
 
         let mut aggregated_msg_stream = msg_stream
             .max_frame_size(MAX_WEBSOCKET_MESSAGE_SIZE)
@@ -158,19 +169,23 @@ pub trait ControlsWebSocketEndpoint: Send + Sync + 'static {
             .max_continuation_size(MAX_WEBSOCKET_MESSAGE_SIZE);
 
         rt::spawn(async move {
-            Self::on_connection_start(connection_close.clone(), context.clone(), &mut session)
-                .await;
+            let (outbound_tx, outbound_rx) = mpsc::unbounded_channel::<WebSocketOutboundFrame>();
 
-            rt::spawn(
-                SessionKeepAlive {
-                    connection_close: connection_close.clone(),
-                    session: session.clone(),
-                }
-                .run(),
-            );
+            rt::spawn(write_websocket_session(
+                connection_close.clone(),
+                outbound_rx,
+                session,
+            ));
+
+            Self::on_connection_start(connection_close.clone(), context.clone(), &outbound_tx)
+                .await;
 
             let (continuation_stop_tx, mut continuation_stop_rx) =
                 mpsc::unbounded_channel::<WebSocketCloseCause>();
+            let mut ping_liveness = PingLiveness::default();
+            let mut ping_ticker = interval_at(Instant::now() + PING_INTERVAL, PING_INTERVAL);
+
+            ping_ticker.set_missed_tick_behavior(MissedTickBehavior::Delay);
 
             let close_cause = loop {
                 select! {
@@ -179,12 +194,27 @@ pub trait ControlsWebSocketEndpoint: Send + Sync + 'static {
                             connection_close.clone(),
                             context.clone(),
                             msg,
-                            &mut session,
+                            &outbound_tx,
+                            &mut ping_liveness,
                             continuation_stop_tx.clone(),
                         ).await {
                             break close_cause;
                         }
                     }
+                    _ = ping_ticker.tick() => {
+                        let ping_decision = match ping_liveness.next_ping() {
+                            PingTurn::PeerStoppedAnswering => {
+                                ContinuationDecision::Stop(WebSocketCloseCause::PeerStoppedAnsweringPings)
+                            }
+                            PingTurn::SendPing => {
+                                enqueue_websocket_frame(&outbound_tx, WebSocketOutboundFrame::Ping)
+                            }
+                        };
+
+                        if let ContinuationDecision::Stop(close_cause) = ping_decision {
+                            break close_cause;
+                        }
+                    }
                     Some(close_cause) = continuation_stop_rx.recv() => {
                         break close_cause;
                     }
@@ -199,9 +229,10 @@ pub trait ControlsWebSocketEndpoint: Send + Sync + 'static {
 
             connection_close.cancel();
 
-            let close_outcome = session.close(close_cause.close_reason()).await;
+            let close_frame_delivery =
+                outbound_tx.send(WebSocketOutboundFrame::Close(close_cause.close_reason()));
 
-            debug!("Closed the WebSocket session after {close_cause:?}: {close_outcome:?}");
+            debug!("Closing the WebSocket session after {close_cause:?}: {close_frame_delivery:?}");
         });
 
         Ok(res)
diff --git a/paddler_balancer/src/desired_state_delivery.rs b/paddler_balancer/src/desired_state_delivery.rs
deleted file mode 100644
index 3396809f8..000000000
--- a/paddler_balancer/src/desired_state_delivery.rs
+++ /dev/null
@@ -1,4 +0,0 @@
-pub enum DesiredStateDelivery {
-    AgentDisconnected,
-    Delivered,
-}
diff --git a/paddler_balancer/src/dispatch_candidate.rs b/paddler_balancer/src/dispatch_candidate.rs
index 782a6dfe4..9d830bbbf 100644
--- a/paddler_balancer/src/dispatch_candidate.rs
+++ b/paddler_balancer/src/dispatch_candidate.rs
@@ -1,8 +1,10 @@
+use std::num::NonZeroU16;
 use std::sync::Arc;
 
 use crate::agent_controller::AgentController;
 
 pub struct DispatchCandidate {
     pub agent_controller: Arc<AgentController>,
+    pub reserved_slots: NonZeroU16,
     pub snapshot: u64,
 }
diff --git a/paddler_balancer/src/dispatch_position.rs b/paddler_balancer/src/dispatch_position.rs
new file mode 100644
index 000000000..2b6ecd5f9
--- /dev/null
+++ b/paddler_balancer/src/dispatch_position.rs
@@ -0,0 +1,35 @@
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+pub enum DispatchPosition {
+    Arriving,
+    Waiting { ticket: u64 },
+}
+
+impl DispatchPosition {
+    #[must_use]
+    pub const fn waits_behind(&self, other_ticket: u64) -> bool {
+        match self {
+            Self::Arriving => true,
+            Self::Waiting { ticket } => other_ticket < *ticket,
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::DispatchPosition;
+
+    #[test]
+    fn an_arriving_request_waits_behind_every_waiting_request() {
+        assert!(DispatchPosition::Arriving.waits_behind(u64::MAX));
+    }
+
+    #[test]
+    fn a_waiting_request_waits_behind_only_the_requests_that_joined_before_it() {
+        let dispatch_position = DispatchPosition::Waiting { ticket: 3 };
+
+        assert_eq!(
+            [2, 3, 4].map(|other_ticket| dispatch_position.waits_behind(other_ticket)),
+            [true, false, false]
+        );
+    }
+}
diff --git a/paddler_balancer/src/dispatch_requirement.rs b/paddler_balancer/src/dispatch_requirement.rs
new file mode 100644
index 000000000..d37057a83
--- /dev/null
+++ b/paddler_balancer/src/dispatch_requirement.rs
@@ -0,0 +1,127 @@
+use std::num::NonZeroU16;
+
+use paddler_messaging::agent_runtime_status::AgentRuntimeStatus;
+use paddler_messaging::inference_capability::InferenceCapability;
+
+use crate::slot_demand::SlotDemand;
+
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+pub struct DispatchRequirement {
+    pub inference_capability: InferenceCapability,
+    pub slot_demand: SlotDemand,
+}
+
+impl DispatchRequirement {
+    #[must_use]
+    pub fn reservable_slots(
+        &self,
+        agent_runtime_status: &AgentRuntimeStatus,
+        slots_unavailable: u64,
+    ) -> Option<NonZeroU16> {
+        self.slot_demand.reservable_from(
+            agent_runtime_status
+                .slots_serving(self.inference_capability)
+                .saturating_sub(slots_unavailable),
+        )
+    }
+
+    #[must_use]
+    pub fn slots_held_while_waiting(&self, agent_runtime_status: &AgentRuntimeStatus) -> u64 {
+        let minimum_slots = u64::from(self.slot_demand.minimum_slots());
+
+        if agent_runtime_status.slots_serving(self.inference_capability) >= minimum_slots {
+            minimum_slots
+        } else {
+            0
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::num::NonZeroU16;
+
+    use paddler_messaging::agent_decisions_serving::AgentDecisionsServing;
+    use paddler_messaging::agent_runtime_status::AgentRuntimeStatus;
+    use paddler_messaging::decision_question::DecisionQuestion;
+    use paddler_messaging::inference_capability::InferenceCapability;
+    use paddler_messaging::request_params::decide_params::DecideParams;
+
+    use super::DispatchRequirement;
+    use crate::agent_streaming_request::AgentStreamingRequest as _;
+    use crate::slot_demand::SlotDemand;
+
+    const SERVING_FOUR_SLOTS: AgentRuntimeStatus = AgentRuntimeStatus::Serving {
+        decisions: AgentDecisionsServing::NotServed,
+        slots_total: 4,
+    };
+
+    fn decision_question(id: &str) -> DecisionQuestion {
+        DecisionQuestion {
+            id: id.to_owned(),
+            instructions: "Pick one".to_owned(),
+            options: vec!["yes".to_owned(), "no".to_owned()],
+        }
+    }
+
+    fn decision_with_a_leading_question() -> DispatchRequirement {
+        DecideParams {
+            last_question: decision_question("last"),
+            leading_questions: vec![decision_question("leading")],
+            state: "state".to_owned(),
+        }
+        .dispatch_requirement()
+    }
+
+    fn requirement(inference_capability: InferenceCapability) -> DispatchRequirement {
+        DispatchRequirement {
+            inference_capability,
+            slot_demand: SlotDemand::UpTo(8),
+        }
+    }
+
+    #[test]
+    fn reserves_the_slots_left_after_the_ones_in_use() {
+        assert_eq!(
+            requirement(InferenceCapability::Embeddings).reservable_slots(&SERVING_FOUR_SLOTS, 1),
+            NonZeroU16::new(3)
+        );
+    }
+
+    #[test]
+    fn reserves_nothing_from_an_agent_using_more_slots_than_it_serves() {
+        assert_eq!(
+            requirement(InferenceCapability::Embeddings).reservable_slots(&SERVING_FOUR_SLOTS, 5),
+            None
+        );
+    }
+
+    #[test]
+    fn reserves_nothing_for_a_capability_the_agent_does_not_serve() {
+        assert_eq!(
+            requirement(InferenceCapability::Decisions).reservable_slots(&SERVING_FOUR_SLOTS, 0),
+            None
+        );
+    }
+
+    #[test]
+    fn a_waiting_requirement_holds_its_minimum_slots_on_an_agent_that_serves_it() {
+        assert_eq!(
+            decision_with_a_leading_question().slots_held_while_waiting(
+                &AgentRuntimeStatus::Serving {
+                    decisions: AgentDecisionsServing::Served,
+                    slots_total: 4,
+                }
+            ),
+            2
+        );
+    }
+
+    #[test]
+    fn a_waiting_requirement_holds_no_slots_on_an_agent_that_cannot_serve_it() {
+        assert_eq!(
+            decision_with_a_leading_question().slots_held_while_waiting(&SERVING_FOUR_SLOTS),
+            0
+        );
+    }
+}
diff --git a/paddler_balancer/src/dispatch_waiting_line.rs b/paddler_balancer/src/dispatch_waiting_line.rs
new file mode 100644
index 000000000..fe002833c
--- /dev/null
+++ b/paddler_balancer/src/dispatch_waiting_line.rs
@@ -0,0 +1,69 @@
+use std::collections::BTreeMap;
+use std::sync::Arc;
+use std::sync::atomic::AtomicU64;
+use std::sync::atomic::Ordering;
+
+use parking_lot::Mutex;
+
+use crate::agent_controller_pool_updates::AgentControllerPoolUpdates;
+use crate::buffered_request_counter::BufferedRequestCounter;
+use crate::dispatch_position::DispatchPosition;
+use crate::dispatch_requirement::DispatchRequirement;
+use crate::requirements_waiting_ahead::RequirementsWaitingAhead;
+use crate::waiting_line_place::WaitingLinePlace;
+
+pub struct DispatchWaitingLine {
+    next_ticket: AtomicU64,
+    updates: Arc<AgentControllerPoolUpdates>,
+    waiting_requirements: Mutex<BTreeMap<u64, DispatchRequirement>>,
+}
+
+impl DispatchWaitingLine {
+    #[must_use]
+    pub const fn new(updates: Arc<AgentControllerPoolUpdates>) -> Self {
+        Self {
+            next_ticket: AtomicU64::new(0),
+            updates,
+            waiting_requirements: Mutex::new(BTreeMap::new()),
+        }
+    }
+
+    pub fn join_when_admitted(
+        &self,
+        dispatch_requirement: DispatchRequirement,
+        buffered_request_counter: &Arc<BufferedRequestCounter>,
+    ) -> Option<WaitingLinePlace<'_>> {
+        let mut waiting_requirements = self.waiting_requirements.lock();
+        let buffered_request_count_guard = buffered_request_counter.try_admit()?;
+        let ticket = self.next_ticket.fetch_add(1, Ordering::Relaxed);
+
+        waiting_requirements.insert(ticket, dispatch_requirement);
+
+        Some(WaitingLinePlace {
+            buffered_request_count_guard,
+            dispatch_waiting_line: self,
+            ticket,
+        })
+    }
+
+    pub fn leave(&self, ticket: u64) {
+        self.waiting_requirements.lock().remove(&ticket);
+        self.updates.signal_waiting_line_change();
+    }
+
+    #[must_use]
+    pub fn requirements_ahead_of(
+        &self,
+        dispatch_position: DispatchPosition,
+    ) -> RequirementsWaitingAhead {
+        RequirementsWaitingAhead {
+            dispatch_requirements: self
+                .waiting_requirements
+                .lock()
+                .iter()
+                .filter(|(ticket, _dispatch_requirement)| dispatch_position.waits_behind(**ticket))
+                .map(|(_ticket, dispatch_requirement)| *dispatch_requirement)
+                .collect(),
+        }
+    }
+}
diff --git a/paddler_balancer/src/dispatchable_request.rs b/paddler_balancer/src/dispatchable_request.rs
new file mode 100644
index 000000000..beb3e770a
--- /dev/null
+++ b/paddler_balancer/src/dispatchable_request.rs
@@ -0,0 +1,16 @@
+use crate::agent_streaming_request::AgentStreamingRequest;
+use crate::dispatch_requirement::DispatchRequirement;
+
+pub struct DispatchableRequest<TParams> {
+    pub dispatch_requirement: DispatchRequirement,
+    pub params: TParams,
+}
+
+impl<TParams: AgentStreamingRequest> From<TParams> for DispatchableRequest<TParams> {
+    fn from(params: TParams) -> Self {
+        Self {
+            dispatch_requirement: params.dispatch_requirement(),
+            params,
+        }
+    }
+}
diff --git a/paddler_balancer/src/dispatched_agent.rs b/paddler_balancer/src/dispatched_agent.rs
index 42cdd632c..e3a2591ac 100644
--- a/paddler_balancer/src/dispatched_agent.rs
+++ b/paddler_balancer/src/dispatched_agent.rs
@@ -1,3 +1,4 @@
+use std::num::NonZeroU16;
 use std::sync::Arc;
 
 use crate::agent_controller::AgentController;
@@ -5,7 +6,7 @@ use crate::agent_controller_slot_guard::AgentControllerSlotGuard;
 
 pub struct DispatchedAgent {
     pub agent_controller: Arc<AgentController>,
-    _slot_guard: AgentControllerSlotGuard,
+    slot_guard: AgentControllerSlotGuard,
 }
 
 impl DispatchedAgent {
@@ -15,7 +16,12 @@ impl DispatchedAgent {
     ) -> Self {
         Self {
             agent_controller,
-            _slot_guard: slot_guard,
+            slot_guard,
         }
     }
+
+    #[must_use]
+    pub const fn reserved_slots(&self) -> NonZeroU16 {
+        self.slot_guard.reserved_slots()
+    }
 }
diff --git a/paddler_balancer/src/dispatched_agent_relay.rs b/paddler_balancer/src/dispatched_agent_relay.rs
new file mode 100644
index 000000000..976ec9179
--- /dev/null
+++ b/paddler_balancer/src/dispatched_agent_relay.rs
@@ -0,0 +1,65 @@
+use std::sync::Arc;
+use std::time::Duration;
+
+use tokio::spawn;
+use tokio_util::sync::CancellationToken;
+use tokio_util::sync::DropGuard;
+
+use paddler_messaging::streamable_result::StreamableResult;
+use paddler_outbox::outbox::Outbox;
+
+use crate::agent_relay_event::AgentRelayEvent;
+use crate::agent_response_forwarder::AgentResponseForwarder;
+use crate::agent_streaming_request::AgentStreamingRequest;
+use crate::dispatched_agent::DispatchedAgent;
+
+pub struct DispatchedAgentRelay<TResponse>
+where
+    TResponse: StreamableResult,
+{
+    pub agent_events: Arc<Outbox<AgentRelayEvent<TResponse::Partial, TResponse::Final>>>,
+    pub generated_by: Option<String>,
+    _relay_abandonment: DropGuard,
+}
+
+impl<TResponse> DispatchedAgentRelay<TResponse>
+where
+    TResponse: StreamableResult<Final: Send + 'static, Partial: Send + 'static> + Send + 'static,
+{
+    pub fn start<TParams>(
+        dispatched_agent: DispatchedAgent,
+        params: TParams,
+        inference_item_timeout: Duration,
+        settlement: CancellationToken,
+        shutdown: CancellationToken,
+    ) -> Self
+    where
+        TParams: AgentStreamingRequest<Response = TResponse>,
+    {
+        let response_receiver = dispatched_agent
+            .agent_controller
+            .stream_responses_to(params, dispatched_agent.reserved_slots());
+        let agent_events = Arc::new(Outbox::default());
+        let generated_by = dispatched_agent.agent_controller.name.clone();
+        let relay_abandonment = CancellationToken::new();
+
+        spawn(
+            AgentResponseForwarder {
+                agent_events: agent_events.clone(),
+                dispatched_agent,
+                inference_item_timeout,
+                relay_abandonment: relay_abandonment.clone(),
+                response_receiver,
+                settlement_guard: settlement.drop_guard(),
+                shutdown,
+            }
+            .forward(),
+        );
+
+        Self {
+            agent_events,
+            generated_by,
+            _relay_abandonment: relay_abandonment.drop_guard(),
+        }
+    }
+}
diff --git a/paddler_balancer/src/distinct_embedding_documents.rs b/paddler_balancer/src/distinct_embedding_documents.rs
new file mode 100644
index 000000000..a993b9a34
--- /dev/null
+++ b/paddler_balancer/src/distinct_embedding_documents.rs
@@ -0,0 +1,184 @@
+use std::collections::HashMap;
+use std::collections::HashSet;
+use std::collections::hash_map::Entry;
+
+use paddler_messaging::embedding_input_document::EmbeddingInputDocument;
+use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
+
+use crate::embedding_document_duplicates::EmbeddingDocumentDuplicates;
+
+struct DocumentsSharingContent {
+    duplicate_document_ids: Vec<String>,
+    first_document_id: String,
+    first_position: usize,
+}
+
+struct DistinctContent {
+    content: String,
+    documents_sharing_content: DocumentsSharingContent,
+}
+
+fn identifies_documents_unambiguously(input_batch: &[EmbeddingInputDocument]) -> bool {
+    let mut document_ids = HashSet::with_capacity(input_batch.len());
+
+    input_batch
+        .iter()
+        .all(|document| document_ids.insert(document.id.as_str()))
+}
+
+pub struct DistinctEmbeddingDocuments {
+    pub duplicates: EmbeddingDocumentDuplicates,
+    pub params: GenerateEmbeddingBatchParams,
+}
+
+impl From<GenerateEmbeddingBatchParams> for DistinctEmbeddingDocuments {
+    fn from(params: GenerateEmbeddingBatchParams) -> Self {
+        if !identifies_documents_unambiguously(&params.input_batch) {
+            return Self {
+                duplicates: EmbeddingDocumentDuplicates::default(),
+                params,
+            };
+        }
+
+        let GenerateEmbeddingBatchParams {
+            input_batch,
+            normalization_method,
+        } = params;
+        let mut documents_by_content: HashMap<String, DocumentsSharingContent> =
+            HashMap::with_capacity(input_batch.len());
+
+        for (position, EmbeddingInputDocument { content, id }) in
+            input_batch.into_iter().enumerate()
+        {
+            match documents_by_content.entry(content) {
+                Entry::Occupied(mut documents_sharing_content) => {
+                    documents_sharing_content
+                        .get_mut()
+                        .duplicate_document_ids
+                        .push(id);
+                }
+                Entry::Vacant(first_with_content) => {
+                    first_with_content.insert(DocumentsSharingContent {
+                        duplicate_document_ids: Vec::new(),
+                        first_document_id: id,
+                        first_position: position,
+                    });
+                }
+            }
+        }
+
+        let mut distinct_contents: Vec<DistinctContent> = documents_by_content
+            .into_iter()
+            .map(|(content, documents_sharing_content)| DistinctContent {
+                content,
+                documents_sharing_content,
+            })
+            .collect();
+
+        distinct_contents.sort_by_key(|distinct_content| {
+            distinct_content.documents_sharing_content.first_position
+        });
+
+        let mut duplicate_document_ids = HashMap::new();
+        let mut distinct_documents = Vec::with_capacity(distinct_contents.len());
+
+        for DistinctContent {
+            content,
+            documents_sharing_content:
+                DocumentsSharingContent {
+                    duplicate_document_ids: duplicates_of_first_document,
+                    first_document_id,
+                    ..
+                },
+        } in distinct_contents
+        {
+            if !duplicates_of_first_document.is_empty() {
+                duplicate_document_ids
+                    .insert(first_document_id.clone(), duplicates_of_first_document);
+            }
+
+            distinct_documents.push(EmbeddingInputDocument {
+                content,
+                id: first_document_id,
+            });
+        }
+
+        Self {
+            duplicates: EmbeddingDocumentDuplicates {
+                duplicate_document_ids,
+            },
+            params: GenerateEmbeddingBatchParams {
+                input_batch: distinct_documents,
+                normalization_method,
+            },
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::collections::HashMap;
+
+    use paddler_messaging::embedding_input_document::EmbeddingInputDocument;
+    use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
+    use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
+
+    use super::DistinctEmbeddingDocuments;
+    use crate::embedding_document_duplicates::EmbeddingDocumentDuplicates;
+
+    fn document(id: &str, content: &str) -> EmbeddingInputDocument {
+        EmbeddingInputDocument {
+            content: content.to_owned(),
+            id: id.to_owned(),
+        }
+    }
+
+    fn batch_of(input_batch: Vec<EmbeddingInputDocument>) -> GenerateEmbeddingBatchParams {
+        GenerateEmbeddingBatchParams {
+            input_batch,
+            normalization_method: EmbeddingNormalizationMethod::None,
+        }
+    }
+
+    #[test]
+    fn embeds_each_content_once_for_every_document_sharing_it() {
+        let DistinctEmbeddingDocuments { duplicates, params } =
+            DistinctEmbeddingDocuments::from(batch_of(vec![
+                document("first", "shared content"),
+                document("lone", "lone content"),
+                document("second", "shared content"),
+                document("third", "shared content"),
+            ]));
+
+        assert_eq!(
+            params.input_batch,
+            vec![
+                document("first", "shared content"),
+                document("lone", "lone content"),
+            ]
+        );
+        assert_eq!(
+            duplicates,
+            EmbeddingDocumentDuplicates {
+                duplicate_document_ids: HashMap::from([(
+                    "first".to_owned(),
+                    vec!["second".to_owned(), "third".to_owned()],
+                )]),
+            }
+        );
+    }
+
+    #[test]
+    fn embeds_every_document_of_a_batch_that_reuses_a_document_id() {
+        let ambiguous_batch = vec![
+            document("reused", "shared content"),
+            document("reused", "other content"),
+            document("distinct", "shared content"),
+        ];
+        let DistinctEmbeddingDocuments { duplicates, params } =
+            DistinctEmbeddingDocuments::from(batch_of(ambiguous_batch.clone()));
+
+        assert_eq!(params.input_batch, ambiguous_batch);
+        assert_eq!(duplicates, EmbeddingDocumentDuplicates::default());
+    }
+}
diff --git a/paddler_balancer/src/embedding_chunk_dispatch.rs b/paddler_balancer/src/embedding_chunk_dispatch.rs
index 3fb214bf3..7c136fd1a 100644
--- a/paddler_balancer/src/embedding_chunk_dispatch.rs
+++ b/paddler_balancer/src/embedding_chunk_dispatch.rs
@@ -1,11 +1,10 @@
 use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
 
+use crate::dispatchable_request::DispatchableRequest;
 use crate::dispatched_agent::DispatchedAgent;
 
 pub enum EmbeddingChunkDispatch {
-    Buffered {
-        batch: GenerateEmbeddingBatchParams,
-    },
+    Buffered(DispatchableRequest<GenerateEmbeddingBatchParams>),
     Claimed {
         batch: GenerateEmbeddingBatchParams,
         dispatched_agent: DispatchedAgent,
diff --git a/paddler_balancer/src/embedding_document_duplicates.rs b/paddler_balancer/src/embedding_document_duplicates.rs
new file mode 100644
index 000000000..4087b9646
--- /dev/null
+++ b/paddler_balancer/src/embedding_document_duplicates.rs
@@ -0,0 +1,277 @@
+use std::collections::HashMap;
+use std::iter::once;
+
+use paddler_messaging::embedding::Embedding;
+use paddler_messaging::embedding_document_exceeding_sequence_context_details::EmbeddingDocumentExceedingSequenceContextDetails;
+use paddler_messaging::embedding_document_tokenized_to_nothing_details::EmbeddingDocumentTokenizedToNothingDetails;
+use paddler_messaging::embedding_result::EmbeddingResult;
+use paddler_messaging::inference_client::message::Message as OutgoingMessage;
+use paddler_messaging::inference_client::response::Response as OutgoingResponse;
+use paddler_messaging::jsonrpc::response_envelope::ResponseEnvelope;
+use paddler_messaging::oversized_embedding_document_details::OversizedEmbeddingDocumentDetails;
+
+#[derive(Debug, Default, Eq, PartialEq)]
+pub struct EmbeddingDocumentDuplicates {
+    pub duplicate_document_ids: HashMap<String, Vec<String>>,
+}
+
+impl EmbeddingDocumentDuplicates {
+    #[must_use]
+    pub fn fan_out(&self, message: OutgoingMessage) -> Vec<OutgoingMessage> {
+        match message {
+            OutgoingMessage::Response(ResponseEnvelope {
+                generated_by,
+                request_id,
+                response: OutgoingResponse::Embedding(embedding_result),
+            }) => {
+                let duplicate_results = self.results_for_duplicates_of(&embedding_result);
+
+                once(embedding_result)
+                    .chain(duplicate_results)
+                    .map(|fanned_out_result| {
+                        OutgoingMessage::Response(ResponseEnvelope {
+                            generated_by: generated_by.clone(),
+                            request_id: request_id.clone(),
+                            response: OutgoingResponse::Embedding(fanned_out_result),
+                        })
+                    })
+                    .collect()
+            }
+            other_message => vec![other_message],
+        }
+    }
+
+    fn duplicates_of<'duplicates>(
+        &'duplicates self,
+        source_document_id: &str,
+    ) -> impl Iterator<Item = &'duplicates String> {
+        self.duplicate_document_ids
+            .get(source_document_id)
+            .into_iter()
+            .flatten()
+    }
+
+    fn results_for_duplicates_of(
+        &self,
+        embedding_result: &EmbeddingResult,
+    ) -> Vec<EmbeddingResult> {
+        match embedding_result {
+            EmbeddingResult::DocumentExceedsBatchSize(oversized_document) => self
+                .duplicates_of(&oversized_document.source_document_id)
+                .map(|duplicate_document_id| {
+                    EmbeddingResult::DocumentExceedsBatchSize(OversizedEmbeddingDocumentDetails {
+                        source_document_id: duplicate_document_id.clone(),
+                        ..oversized_document.clone()
+                    })
+                })
+                .collect(),
+            EmbeddingResult::DocumentExceedsSequenceContext(oversized_document) => self
+                .duplicates_of(&oversized_document.source_document_id)
+                .map(|duplicate_document_id| {
+                    EmbeddingResult::DocumentExceedsSequenceContext(
+                        EmbeddingDocumentExceedingSequenceContextDetails {
+                            source_document_id: duplicate_document_id.clone(),
+                            ..oversized_document.clone()
+                        },
+                    )
+                })
+                .collect(),
+            EmbeddingResult::DocumentTokenizedToNothing(empty_document) => self
+                .duplicates_of(&empty_document.source_document_id)
+                .map(|duplicate_document_id| {
+                    EmbeddingResult::DocumentTokenizedToNothing(
+                        EmbeddingDocumentTokenizedToNothingDetails {
+                            source_document_id: duplicate_document_id.clone(),
+                        },
+                    )
+                })
+                .collect(),
+            EmbeddingResult::Embedding(embedding) => self
+                .duplicates_of(&embedding.source_document_id)
+                .map(|duplicate_document_id| {
+                    EmbeddingResult::Embedding(Embedding {
+                        source_document_id: duplicate_document_id.clone(),
+                        ..embedding.clone()
+                    })
+                })
+                .collect(),
+            EmbeddingResult::AgentRuntimeFailed(_)
+            | EmbeddingResult::BatchAssemblyFailed(_)
+            | EmbeddingResult::DecodeFailed(_)
+            | EmbeddingResult::Done
+            | EmbeddingResult::EmbeddingTooLongForRmsNormalization(_)
+            | EmbeddingResult::EmbeddingsUnavailable(_)
+            | EmbeddingResult::InputTokenizationFailed(_)
+            | EmbeddingResult::ModelNotLoaded(_)
+            | EmbeddingResult::NoEmbeddingsProduced
+            | EmbeddingResult::SchedulerUnavailable(_) => Vec::new(),
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::collections::HashMap;
+
+    use serde_json::to_string;
+
+    use paddler_inference_parameters::pooling_type::PoolingType;
+    use paddler_messaging::embedding::Embedding;
+    use paddler_messaging::embedding_document_exceeding_sequence_context_details::EmbeddingDocumentExceedingSequenceContextDetails;
+    use paddler_messaging::embedding_document_tokenized_to_nothing_details::EmbeddingDocumentTokenizedToNothingDetails;
+    use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
+    use paddler_messaging::embedding_result::EmbeddingResult;
+    use paddler_messaging::embedding_vector::EmbeddingVector;
+    use paddler_messaging::inference_client::message::Message as OutgoingMessage;
+    use paddler_messaging::inference_client::response::Response as OutgoingResponse;
+    use paddler_messaging::jsonrpc::error::Error as JsonRpcError;
+    use paddler_messaging::jsonrpc::error_envelope::ErrorEnvelope;
+    use paddler_messaging::jsonrpc::response_envelope::ResponseEnvelope;
+    use paddler_messaging::oversized_embedding_document_details::OversizedEmbeddingDocumentDetails;
+
+    use super::EmbeddingDocumentDuplicates;
+
+    fn duplicates_of_first_document() -> EmbeddingDocumentDuplicates {
+        EmbeddingDocumentDuplicates {
+            duplicate_document_ids: HashMap::from([(
+                "first".to_owned(),
+                vec!["second".to_owned(), "third".to_owned()],
+            )]),
+        }
+    }
+
+    fn embedding_response(embedding_result: EmbeddingResult) -> OutgoingMessage {
+        OutgoingMessage::Response(ResponseEnvelope {
+            generated_by: Some("agent".to_owned()),
+            request_id: "request".to_owned(),
+            response: OutgoingResponse::Embedding(embedding_result),
+        })
+    }
+
+    fn embedding_of(source_document_id: &str) -> EmbeddingResult {
+        EmbeddingResult::Embedding(Embedding {
+            embedding: EmbeddingVector::from(vec![0.25, 0.75]),
+            normalization_method: EmbeddingNormalizationMethod::None,
+            pooling_type: PoolingType::Mean,
+            source_document_id: source_document_id.to_owned(),
+        })
+    }
+
+    fn oversized_document(source_document_id: &str) -> EmbeddingResult {
+        EmbeddingResult::DocumentExceedsBatchSize(OversizedEmbeddingDocumentDetails {
+            document_tokens: 900,
+            n_batch: 512,
+            source_document_id: source_document_id.to_owned(),
+        })
+    }
+
+    fn document_exceeding_sequence_context(source_document_id: &str) -> EmbeddingResult {
+        EmbeddingResult::DocumentExceedsSequenceContext(
+            EmbeddingDocumentExceedingSequenceContextDetails {
+                document_tokens: 300,
+                sequence_context_size: 256,
+                source_document_id: source_document_id.to_owned(),
+            },
+        )
+    }
+
+    fn document_tokenized_to_nothing(source_document_id: &str) -> EmbeddingResult {
+        EmbeddingResult::DocumentTokenizedToNothing(EmbeddingDocumentTokenizedToNothingDetails {
+            source_document_id: source_document_id.to_owned(),
+        })
+    }
+
+    fn serialized(messages: &[OutgoingMessage]) -> Vec<String> {
+        messages
+            .iter()
+            .map(|message| to_string(message).unwrap())
+            .collect()
+    }
+
+    #[test]
+    fn fans_an_embedding_out_to_every_document_sharing_its_content() {
+        assert_eq!(
+            serialized(
+                &duplicates_of_first_document().fan_out(embedding_response(embedding_of("first")))
+            ),
+            serialized(&[
+                embedding_response(embedding_of("first")),
+                embedding_response(embedding_of("second")),
+                embedding_response(embedding_of("third")),
+            ])
+        );
+    }
+
+    #[test]
+    fn fans_an_oversized_document_out_to_every_document_sharing_its_content() {
+        assert_eq!(
+            serialized(
+                &duplicates_of_first_document()
+                    .fan_out(embedding_response(oversized_document("first")))
+            ),
+            serialized(&[
+                embedding_response(oversized_document("first")),
+                embedding_response(oversized_document("second")),
+                embedding_response(oversized_document("third")),
+            ])
+        );
+    }
+
+    #[test]
+    fn fans_a_document_exceeding_its_sequence_context_out_to_every_document_sharing_its_content() {
+        assert_eq!(
+            serialized(&duplicates_of_first_document().fan_out(embedding_response(
+                document_exceeding_sequence_context("first")
+            ))),
+            serialized(&[
+                embedding_response(document_exceeding_sequence_context("first")),
+                embedding_response(document_exceeding_sequence_context("second")),
+                embedding_response(document_exceeding_sequence_context("third")),
+            ])
+        );
+    }
+
+    #[test]
+    fn fans_a_document_tokenized_to_nothing_out_to_every_document_sharing_its_content() {
+        assert_eq!(
+            serialized(
+                &duplicates_of_first_document()
+                    .fan_out(embedding_response(document_tokenized_to_nothing("first")))
+            ),
+            serialized(&[
+                embedding_response(document_tokenized_to_nothing("first")),
+                embedding_response(document_tokenized_to_nothing("second")),
+                embedding_response(document_tokenized_to_nothing("third")),
+            ])
+        );
+    }
+
+    #[test]
+    fn passes_a_result_of_a_whole_chunk_through_once() {
+        assert_eq!(
+            serialized(
+                &duplicates_of_first_document()
+                    .fan_out(embedding_response(EmbeddingResult::NoEmbeddingsProduced))
+            ),
+            serialized(&[embedding_response(EmbeddingResult::NoEmbeddingsProduced)])
+        );
+    }
+
+    #[test]
+    fn passes_an_error_through_once() {
+        let error = || {
+            OutgoingMessage::Error(ErrorEnvelope {
+                request_id: "request".to_owned(),
+                error: JsonRpcError {
+                    code: 504,
+                    description: "timed out".to_owned(),
+                },
+            })
+        };
+
+        assert_eq!(
+            serialized(&duplicates_of_first_document().fan_out(error())),
+            serialized(&[error()])
+        );
+    }
+}
diff --git a/paddler_balancer/src/enqueue_websocket_frame.rs b/paddler_balancer/src/enqueue_websocket_frame.rs
new file mode 100644
index 000000000..4036b1c82
--- /dev/null
+++ b/paddler_balancer/src/enqueue_websocket_frame.rs
@@ -0,0 +1,40 @@
+use tokio::sync::mpsc::UnboundedSender;
+use tokio::sync::mpsc::error::SendError;
+
+use crate::continuation_decision::ContinuationDecision;
+use crate::websocket_close_cause::WebSocketCloseCause;
+use crate::websocket_outbound_frame::WebSocketOutboundFrame;
+
+pub fn enqueue_websocket_frame(
+    outbound_tx: &UnboundedSender<WebSocketOutboundFrame>,
+    frame: WebSocketOutboundFrame,
+) -> ContinuationDecision {
+    match outbound_tx.send(frame) {
+        Ok(()) => ContinuationDecision::Continue,
+        Err(SendError(_unsent_frame)) => {
+            ContinuationDecision::Stop(WebSocketCloseCause::SessionAlreadyClosed)
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use tokio::sync::mpsc::unbounded_channel;
+
+    use super::enqueue_websocket_frame;
+    use crate::continuation_decision::ContinuationDecision;
+    use crate::websocket_close_cause::WebSocketCloseCause;
+    use crate::websocket_outbound_frame::WebSocketOutboundFrame;
+
+    #[test]
+    fn a_frame_for_a_writer_that_stopped_ends_the_connection_as_already_closed() {
+        let (outbound_tx, outbound_rx) = unbounded_channel();
+
+        drop(outbound_rx);
+
+        assert_eq!(
+            enqueue_websocket_frame(&outbound_tx, WebSocketOutboundFrame::Ping),
+            ContinuationDecision::Stop(WebSocketCloseCause::SessionAlreadyClosed)
+        );
+    }
+}
diff --git a/paddler_balancer/src/forward_agent_relay.rs b/paddler_balancer/src/forward_agent_relay.rs
new file mode 100644
index 000000000..0ceab2003
--- /dev/null
+++ b/paddler_balancer/src/forward_agent_relay.rs
@@ -0,0 +1,51 @@
+use tokio::select;
+use tokio_util::sync::CancellationToken;
+
+use crate::agent_relay::AgentRelay;
+use crate::agent_relay_step::AgentRelayStep;
+use crate::agent_streaming_request::AgentStreamingRequest;
+use crate::receives_agent_relay::ReceivesAgentRelay;
+
+async fn relay_to_receiver<TParams, TReceiver>(
+    mut agent_relay: AgentRelay<TParams>,
+    mut receiver: TReceiver,
+) where
+    TParams: AgentStreamingRequest,
+    TReceiver: ReceivesAgentRelay<TParams::Response>,
+{
+    loop {
+        match agent_relay.next().await {
+            AgentRelayStep::Concluded(conclusion) => {
+                return receiver.receive_conclusion(conclusion).await;
+            }
+            AgentRelayStep::Partial {
+                generated_by,
+                partial_result,
+                relay,
+            } => {
+                receiver.receive_partial(generated_by, partial_result).await;
+                agent_relay = relay;
+            }
+        }
+    }
+}
+
+pub async fn forward_agent_relay<TParams, TReceiver>(
+    agent_relay: AgentRelay<TParams>,
+    connection_close: CancellationToken,
+    receiver: TReceiver,
+) where
+    TParams: AgentStreamingRequest,
+    TReceiver: ReceivesAgentRelay<TParams::Response>,
+{
+    let settlement = agent_relay.settlement();
+
+    select! {
+        biased;
+
+        () = connection_close.cancelled() => {}
+        () = relay_to_receiver(agent_relay, receiver) => {}
+    }
+
+    settlement.cancelled().await;
+}
diff --git a/paddler_balancer/src/forward_responses_stream.rs b/paddler_balancer/src/forward_responses_stream.rs
deleted file mode 100644
index 63045845d..000000000
--- a/paddler_balancer/src/forward_responses_stream.rs
+++ /dev/null
@@ -1,154 +0,0 @@
-use std::fmt::Debug;
-use std::pin::pin;
-
-use log::debug;
-use log::error;
-use tokio::select;
-use tokio::time::Instant;
-use tokio::time::sleep;
-use tokio_util::sync::CancellationToken;
-
-use paddler_messaging::inference_client::message::Message as OutgoingMessage;
-use paddler_messaging::inference_client::response::Response as OutgoingResponse;
-use paddler_messaging::jsonrpc::response_envelope::ResponseEnvelope;
-use paddler_messaging::streamable_result::StreamableResult;
-
-use crate::agent_response_forwarding_mode::AgentResponseForwardingMode;
-use crate::agent_response_receiver::AgentResponseReceiver;
-use crate::agent_stop_follow_up::AgentStopFollowUp;
-use crate::controls_session::ControlsSession;
-use crate::dispatched_agent::DispatchedAgent;
-use crate::forwarding_event::ForwardingEvent;
-use crate::forwarding_plan::ForwardingPlan;
-use crate::inference_service::configuration::Configuration as InferenceServiceConfiguration;
-use crate::respond_with_error::respond_with_error;
-
-async fn send_response_to_client<TControlsSession, TResponse>(
-    generated_by: Option<String>,
-    response: TResponse,
-    request_id: String,
-    session_controller: &mut TControlsSession,
-) -> bool
-where
-    TControlsSession: ControlsSession<OutgoingMessage>,
-    TResponse: Into<OutgoingResponse> + Send,
-{
-    if let Err(err) = session_controller
-        .send_response(OutgoingMessage::Response(ResponseEnvelope {
-            generated_by,
-            request_id: request_id.clone(),
-            response: response.into(),
-        }))
-        .await
-    {
-        error!("Failed to send response for request {request_id:?}: {err}");
-
-        return false;
-    }
-
-    true
-}
-
-pub async fn forward_responses_stream<TControlsSession, TResponse>(
-    connection_close: CancellationToken,
-    dispatched_agent: DispatchedAgent,
-    inference_service_configuration: InferenceServiceConfiguration,
-    mut receive_response_controller: AgentResponseReceiver<TResponse>,
-    request_id: String,
-    mut session_controller: TControlsSession,
-    shutdown: CancellationToken,
-) where
-    TControlsSession: ControlsSession<OutgoingMessage>,
-    TResponse: Debug + Into<OutgoingResponse> + Send + StreamableResult,
-{
-    debug!("Found available agent controller for request: {request_id:?}");
-
-    let agent_controller = dispatched_agent.agent_controller.clone();
-    let agent_request_id = receive_response_controller
-        .request_registration
-        .request_id
-        .clone();
-    let agent_connection_close = agent_controller.connection_close.clone();
-    let inference_item_timeout = inference_service_configuration.inference_item_timeout;
-    let mut forwarding_mode = AgentResponseForwardingMode::ForwardingToClient;
-    let mut item_timeout = pin!(sleep(inference_item_timeout));
-
-    loop {
-        item_timeout
-            .as_mut()
-            .reset(Instant::now() + inference_item_timeout);
-
-        let is_forwarding_to_client = matches!(
-            forwarding_mode,
-            AgentResponseForwardingMode::ForwardingToClient
-        );
-
-        let forwarding_event = select! {
-            biased;
-
-            () = shutdown.cancelled() => ForwardingEvent::ShutdownRequested,
-            () = agent_connection_close.cancelled() => ForwardingEvent::AgentConnectionClosed,
-            () = connection_close.cancelled(), if is_forwarding_to_client => {
-                ForwardingEvent::ClientConnectionClosed
-            }
-            () = &mut item_timeout => ForwardingEvent::ItemTimedOut,
-            Some(response) = receive_response_controller.response_rx.recv() => {
-                ForwardingEvent::ResponseReceived(response)
-            }
-        };
-
-        let agent_stop_follow_up =
-            match forwarding_mode.plan_for(forwarding_event, inference_item_timeout) {
-                ForwardingPlan::Finish => break,
-                ForwardingPlan::IgnoreDrainedResponse => continue,
-                ForwardingPlan::ForwardResponse { is_done, response } => {
-                    let send_succeeded = send_response_to_client(
-                        agent_controller.name.clone(),
-                        response,
-                        request_id.clone(),
-                        &mut session_controller,
-                    )
-                    .await;
-
-                    if is_done {
-                        break;
-                    }
-
-                    if send_succeeded {
-                        continue;
-                    }
-
-                    AgentStopFollowUp::DrainUntilAgentConfirms
-                }
-                ForwardingPlan::ReplyWithErrorThenFinish(error) => {
-                    respond_with_error(error, request_id.clone(), &mut session_controller).await;
-
-                    break;
-                }
-                ForwardingPlan::ReplyWithErrorThenStopAgent(error) => {
-                    respond_with_error(error, request_id.clone(), &mut session_controller).await;
-
-                    AgentStopFollowUp::DrainUntilAgentConfirms
-                }
-                ForwardingPlan::ReplyWithErrorThenStopAgentThenFinish(error) => {
-                    respond_with_error(error, request_id.clone(), &mut session_controller).await;
-
-                    AgentStopFollowUp::Finish
-                }
-                ForwardingPlan::StopAgent => AgentStopFollowUp::DrainUntilAgentConfirms,
-            };
-
-        if let Err(stop_error) = agent_controller.stop_responding_to(agent_request_id.clone()) {
-            error!("Failed to stop responding to request {request_id:?}: {stop_error}");
-
-            break;
-        }
-
-        match agent_stop_follow_up {
-            AgentStopFollowUp::DrainUntilAgentConfirms => {
-                forwarding_mode = AgentResponseForwardingMode::DrainingUntilAgentConfirms;
-            }
-            AgentStopFollowUp::Finish => break,
-        }
-    }
-}
diff --git a/paddler_balancer/src/forwarding_event.rs b/paddler_balancer/src/forwarding_event.rs
index 2114fd762..9f294955f 100644
--- a/paddler_balancer/src/forwarding_event.rs
+++ b/paddler_balancer/src/forwarding_event.rs
@@ -1,7 +1,9 @@
-pub enum ForwardingEvent<TResponse> {
+use paddler_messaging::result_step::ResultStep;
+
+pub enum ForwardingEvent<TPartial, TFinal> {
     AgentConnectionClosed,
-    ClientConnectionClosed,
     ItemTimedOut,
-    ResponseReceived(TResponse),
+    RelayAbandoned,
+    ResponseReceived(ResultStep<TPartial, TFinal>),
     ShutdownRequested,
 }
diff --git a/paddler_balancer/src/forwarding_plan.rs b/paddler_balancer/src/forwarding_plan.rs
index 70c730778..444d00fe2 100644
--- a/paddler_balancer/src/forwarding_plan.rs
+++ b/paddler_balancer/src/forwarding_plan.rs
@@ -1,12 +1,12 @@
-use paddler_messaging::jsonrpc::error::Error as JsonRpcError;
+use crate::agent_relay_event::AgentRelayEvent;
 
 #[derive(Debug, Eq, PartialEq)]
-pub enum ForwardingPlan<TResponse> {
+pub enum ForwardingPlan<TPartial, TFinal> {
     Finish,
-    ForwardResponse { is_done: bool, response: TResponse },
     IgnoreDrainedResponse,
-    ReplyWithErrorThenFinish(JsonRpcError),
-    ReplyWithErrorThenStopAgent(JsonRpcError),
-    ReplyWithErrorThenStopAgentThenFinish(JsonRpcError),
+    NotifyRelay(AgentRelayEvent<TPartial, TFinal>),
+    NotifyRelayThenFinish(AgentRelayEvent<TPartial, TFinal>),
+    NotifyRelayThenStopAgent(AgentRelayEvent<TPartial, TFinal>),
+    NotifyRelayThenStopAgentThenFinish(AgentRelayEvent<TPartial, TFinal>),
     StopAgent,
 }
diff --git a/paddler_balancer/src/http_listener.rs b/paddler_balancer/src/http_listener.rs
index 76b113712..33779b483 100644
--- a/paddler_balancer/src/http_listener.rs
+++ b/paddler_balancer/src/http_listener.rs
@@ -18,7 +18,6 @@ impl HttpListener {
     pub fn bind(addr: SocketAddr) -> io::Result<Self> {
         let socket = Socket::new(Domain::for_address(addr), Type::STREAM, Some(Protocol::TCP))?;
 
-        #[cfg(not(windows))]
         socket.set_reuse_address(true)?;
 
         socket.bind(&addr.into())?;
diff --git a/paddler_balancer/src/inference_service/http_route/api/mod.rs b/paddler_balancer/src/inference_service/http_route/api/mod.rs
index eb41ede29..7dc97c201 100644
--- a/paddler_balancer/src/inference_service/http_route/api/mod.rs
+++ b/paddler_balancer/src/inference_service/http_route/api/mod.rs
@@ -1,4 +1,5 @@
 pub mod post_continue_from_conversation_history;
 pub mod post_continue_from_raw_prompt;
+pub mod post_decide;
 pub mod post_generate_embedding_batch;
 pub mod ws_inference_socket;
diff --git a/paddler_balancer/src/inference_service/http_route/api/post_continue_from_conversation_history.rs b/paddler_balancer/src/inference_service/http_route/api/post_continue_from_conversation_history.rs
index a69a7b74f..cd806dc04 100644
--- a/paddler_balancer/src/inference_service/http_route/api/post_continue_from_conversation_history.rs
+++ b/paddler_balancer/src/inference_service/http_route/api/post_continue_from_conversation_history.rs
@@ -20,11 +20,6 @@ async fn respond(
     app_data: web::Data<AppData>,
     params: web::Json<ContinueFromConversationHistoryParams<RawParametersSchema>>,
 ) -> Result<impl Responder, Error> {
-    app_data
-        .balancer_applicable_state_holder
-        .token_generation_mode()
-        .require_enabled()?;
-
     let validated_params = params.into_inner().validate().map_err(|validation_error| {
         ErrorBadRequest(invalid_request_parameters_description(&validation_error))
     })?;
diff --git a/paddler_balancer/src/inference_service/http_route/api/post_continue_from_raw_prompt.rs b/paddler_balancer/src/inference_service/http_route/api/post_continue_from_raw_prompt.rs
index 776a7207b..453d1d9ba 100644
--- a/paddler_balancer/src/inference_service/http_route/api/post_continue_from_raw_prompt.rs
+++ b/paddler_balancer/src/inference_service/http_route/api/post_continue_from_raw_prompt.rs
@@ -16,11 +16,6 @@ async fn respond(
     app_data: web::Data<AppData>,
     params: web::Json<ContinueFromRawPromptParams>,
 ) -> Result<impl Responder, Error> {
-    app_data
-        .balancer_applicable_state_holder
-        .token_generation_mode()
-        .require_enabled()?;
-
     Ok(ndjson_response(unbounded_stream_from_agent(
         UnboundedStreamFromAgentParams {
             buffered_request_manager: app_data.buffered_request_manager.clone(),
diff --git a/paddler_balancer/src/inference_service/http_route/api/post_decide.rs b/paddler_balancer/src/inference_service/http_route/api/post_decide.rs
new file mode 100644
index 000000000..bc8a2ecd4
--- /dev/null
+++ b/paddler_balancer/src/inference_service/http_route/api/post_decide.rs
@@ -0,0 +1,45 @@
+use actix_web::Error;
+use actix_web::Responder;
+use actix_web::error::ErrorBadRequest;
+use actix_web::error::ErrorServiceUnavailable;
+use actix_web::web;
+use actix_web::web::post;
+
+use paddler_messaging::api_path::ApiPath;
+use paddler_messaging::request_params::decide_params::raw_decide_params::RawDecideParams;
+use paddler_messaging::validates::Validates as _;
+
+use crate::chunk_forwarding_session_controller::identity_transformer::IdentityTransformer;
+use crate::inference_service::app_data::AppData;
+use crate::invalid_request_parameters_description::invalid_request_parameters_description;
+use crate::ndjson_response::ndjson_response;
+use crate::unbounded_stream_from_agent::unbounded_stream_from_agent;
+use crate::unbounded_stream_from_agent_params::UnboundedStreamFromAgentParams;
+
+async fn respond(
+    app_data: web::Data<AppData>,
+    params: web::Json<RawDecideParams>,
+) -> Result<impl Responder, Error> {
+    app_data
+        .balancer_applicable_state_holder
+        .require_decisions()
+        .map_err(ErrorServiceUnavailable)?;
+
+    let validated_params = params.into_inner().validate().map_err(|validation_error| {
+        ErrorBadRequest(invalid_request_parameters_description(&validation_error))
+    })?;
+
+    Ok(ndjson_response(unbounded_stream_from_agent(
+        UnboundedStreamFromAgentParams {
+            buffered_request_manager: app_data.buffered_request_manager.clone(),
+            inference_service_configuration: app_data.inference_service_configuration.clone(),
+            request_params: validated_params,
+            shutdown: app_data.shutdown.clone(),
+            transformer: IdentityTransformer::new(),
+        },
+    )))
+}
+
+pub fn post_decide(cfg: &mut web::ServiceConfig) {
+    cfg.route(ApiPath::DECIDE, post().to(respond));
+}
diff --git a/paddler_balancer/src/inference_service/http_route/api/post_generate_embedding_batch.rs b/paddler_balancer/src/inference_service/http_route/api/post_generate_embedding_batch.rs
index 264a128bc..688412b97 100644
--- a/paddler_balancer/src/inference_service/http_route/api/post_generate_embedding_batch.rs
+++ b/paddler_balancer/src/inference_service/http_route/api/post_generate_embedding_batch.rs
@@ -1,19 +1,15 @@
-use std::num::NonZeroUsize;
+use std::sync::Arc;
 
 use actix_web::Error;
 use actix_web::HttpResponse;
 use actix_web::Responder;
-use actix_web::error::ErrorNotImplemented;
-use actix_web::error::ErrorServiceUnavailable;
 use actix_web::http::header;
 use actix_web::rt;
 use actix_web::web;
 use actix_web::web::post;
 use anyhow::Result;
-use async_trait::async_trait;
 use futures::stream::StreamExt;
 use nanoid::nanoid;
-use serde_json::to_string;
 use tokio::sync::mpsc;
 use tokio::task::JoinSet;
 use tokio_stream::wrappers::UnboundedReceiverStream;
@@ -26,37 +22,31 @@ use paddler_messaging::inference_client::response::Response as OutgoingResponse;
 use paddler_messaging::jsonrpc::response_envelope::ResponseEnvelope;
 use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
 
-use crate::agent_relay_error::AgentRelayError;
+use crate::agent_relay::AgentRelay;
+use crate::awaiting_agent_relay::AwaitingAgentRelay;
 use crate::cancellation_token_stream_guard::CancellationTokenStreamGuard;
 use crate::chunk_forwarding_session_controller::ChunkForwardingSessionController;
 use crate::chunk_forwarding_session_controller::identity_transformer::IdentityTransformer;
 use crate::chunk_forwarding_session_controller::transform_result::TransformResult;
 use crate::chunk_forwarding_session_controller::transforms_outgoing_message::TransformsOutgoingMessage;
 use crate::controls_session::ControlsSession as _;
+use crate::dispatchable_request::DispatchableRequest;
+use crate::distinct_embedding_documents::DistinctEmbeddingDocuments;
 use crate::embedding_chunk_dispatch::EmbeddingChunkDispatch;
+use crate::embedding_document_duplicates::EmbeddingDocumentDuplicates;
+use crate::forward_agent_relay::forward_agent_relay;
 use crate::inference_service::app_data::AppData;
-use crate::request_from_agent::request_from_agent;
-use crate::request_from_dispatched_agent::request_from_dispatched_agent;
+use crate::outgoing_message_relay_receiver::OutgoingMessageRelayReceiver;
 
 async fn respond(
     app_data: web::Data<AppData>,
     params: web::Json<GenerateEmbeddingBatchParams>,
 ) -> Result<impl Responder, Error> {
-    let agent_desired_state = app_data
+    let embedding_batch_size = app_data
         .balancer_applicable_state_holder
-        .get_agent_desired_state();
-
-    if !agent_desired_state.inference_parameters.enable_embeddings {
-        return Err(ErrorNotImplemented(
-            "Embedding generation is not enabled in the inference parameters",
-        ));
-    }
-
-    let Some(agent_count) = NonZeroUsize::new(app_data.agent_controller_pool.agents.len()) else {
-        return Err(ErrorServiceUnavailable("No agents are currently connected"));
-    };
-    let embedding_batch_size = agent_desired_state
-        .inference_parameters
+        .get_balancer_applicable_state()
+        .balancer_desired_state
+        .embeddings
         .embedding_batch_size;
 
     let connection_close = CancellationToken::new();
@@ -65,50 +55,74 @@ async fn respond(
 
     let mut chunk_tasks: JoinSet<()> = JoinSet::new();
 
-    let batches = params
-        .into_inner()
-        .chunk_evenly_with_cap(agent_count, embedding_batch_size);
+    let DistinctEmbeddingDocuments {
+        duplicates,
+        params: distinct_params,
+    } = DistinctEmbeddingDocuments::from(params.into_inner());
+    let embedding_document_duplicates = Arc::new(duplicates);
+    let batches = distinct_params.chunk_evenly_with_cap(
+        app_data.agent_controller_pool.agent_count(),
+        embedding_batch_size,
+    );
 
     let chunk_dispatches: Vec<EmbeddingChunkDispatch> = batches
         .into_iter()
-        .map(
-            |batch| match app_data.buffered_request_manager.take_available_agent() {
+        .map(|batch| {
+            let dispatchable_batch = DispatchableRequest::from(batch);
+
+            match app_data
+                .buffered_request_manager
+                .take_available_agent(dispatchable_batch.dispatch_requirement)
+            {
                 Some(dispatched_agent) => EmbeddingChunkDispatch::Claimed {
-                    batch,
+                    batch: dispatchable_batch.params,
                     dispatched_agent,
                 },
-                None => EmbeddingChunkDispatch::Buffered { batch },
-            },
-        )
+                None => EmbeddingChunkDispatch::Buffered(dispatchable_batch),
+            }
+        })
         .collect();
 
+    let inference_item_timeout = app_data
+        .inference_service_configuration
+        .inference_item_timeout;
+
     for chunk_dispatch in chunk_dispatches {
-        let session_controller =
-            ChunkForwardingSessionController::new(chunk_tx.clone(), EmbeddingChunkBodyTransformer);
+        let session_controller = ChunkForwardingSessionController::new(
+            chunk_tx.clone(),
+            EmbeddingChunkBodyTransformer {
+                embedding_document_duplicates: embedding_document_duplicates.clone(),
+            },
+        );
 
-        match chunk_dispatch {
+        let agent_relay = match chunk_dispatch {
             EmbeddingChunkDispatch::Claimed {
                 batch,
                 dispatched_agent,
-            } => chunk_tasks.spawn(request_from_dispatched_agent(
+            } => AgentRelay::dispatched(
                 dispatched_agent,
-                connection_close.clone(),
-                app_data.inference_service_configuration.clone(),
                 batch,
-                request_id.clone(),
-                session_controller,
+                inference_item_timeout,
                 app_data.shutdown.clone(),
-            )),
-            EmbeddingChunkDispatch::Buffered { batch } => chunk_tasks.spawn(request_from_agent(
-                app_data.buffered_request_manager.clone(),
-                connection_close.clone(),
-                app_data.inference_service_configuration.clone(),
-                batch,
-                request_id.clone(),
-                session_controller,
-                app_data.shutdown.clone(),
-            )),
+            ),
+            EmbeddingChunkDispatch::Buffered(dispatchable_batch) => {
+                AgentRelay::awaiting(AwaitingAgentRelay {
+                    buffered_request_manager: app_data.buffered_request_manager.clone(),
+                    dispatchable_request: dispatchable_batch,
+                    inference_item_timeout,
+                    shutdown: app_data.shutdown.clone(),
+                })
+            }
         };
+
+        chunk_tasks.spawn(forward_agent_relay(
+            agent_relay,
+            connection_close.clone(),
+            OutgoingMessageRelayReceiver {
+                request_id: request_id.clone(),
+                session_controller,
+            },
+        ));
     }
 
     let final_done_chunk_tx = chunk_tx.clone();
@@ -142,28 +156,27 @@ async fn respond(
         .streaming(stream))
 }
 
-#[derive(Clone)]
-struct EmbeddingChunkBodyTransformer;
+struct EmbeddingChunkBodyTransformer {
+    embedding_document_duplicates: Arc<EmbeddingDocumentDuplicates>,
+}
 
-#[async_trait]
 impl TransformsOutgoingMessage for EmbeddingChunkBodyTransformer {
     type Output = TransformResult;
 
-    async fn transform(
-        &self,
-        message: OutgoingMessage,
-    ) -> Result<Vec<TransformResult>, AgentRelayError> {
+    fn transform(&self, message: OutgoingMessage) -> Vec<TransformResult> {
         if let OutgoingMessage::Response(ResponseEnvelope {
             response: OutgoingResponse::Embedding(EmbeddingResult::Done),
             ..
         }) = &message
         {
-            return Ok(vec![TransformResult::Discard]);
+            return vec![TransformResult::Discard];
         }
 
-        to_string(&message)
-            .map(|serialized| vec![TransformResult::Chunk(serialized)])
-            .map_err(AgentRelayError::MessageUnserializable)
+        self.embedding_document_duplicates
+            .fan_out(message)
+            .iter()
+            .map(TransformResult::serialized_from)
+            .collect()
     }
 }
 
diff --git a/paddler_balancer/src/inference_service/http_route/api/ws_inference_socket/inference_socket_controller_context.rs b/paddler_balancer/src/inference_service/http_route/api/ws_inference_socket/inference_socket_controller_context.rs
index c87d25681..35f76112d 100644
--- a/paddler_balancer/src/inference_service/http_route/api/ws_inference_socket/inference_socket_controller_context.rs
+++ b/paddler_balancer/src/inference_service/http_route/api/ws_inference_socket/inference_socket_controller_context.rs
@@ -4,12 +4,10 @@ use tokio_util::sync::CancellationToken;
 
 use paddler_request_registry::request_registry::RequestRegistry;
 
-use crate::balancer_applicable_state_holder::BalancerApplicableStateHolder;
 use crate::buffered_request_manager::BufferedRequestManager;
 use crate::inference_service::configuration::Configuration as InferenceServiceConfiguration;
 
 pub struct InferenceSocketControllerContext {
-    pub balancer_applicable_state_holder: Arc<BalancerApplicableStateHolder>,
     pub buffered_request_manager: Arc<BufferedRequestManager>,
     pub inference_service_configuration: InferenceServiceConfiguration,
     pub request_cancellation_tokens: Arc<RequestRegistry<CancellationToken>>,
diff --git a/paddler_balancer/src/inference_service/http_route/api/ws_inference_socket/mod.rs b/paddler_balancer/src/inference_service/http_route/api/ws_inference_socket/mod.rs
index dd56b76cd..df27645bf 100644
--- a/paddler_balancer/src/inference_service/http_route/api/ws_inference_socket/mod.rs
+++ b/paddler_balancer/src/inference_service/http_route/api/ws_inference_socket/mod.rs
@@ -1,7 +1,5 @@
 mod inference_socket_controller_context;
-mod spawn_token_generation_mode_watcher;
 
-use std::fmt::Debug;
 use std::sync::Arc;
 
 use actix_web::Error;
@@ -12,64 +10,43 @@ use actix_web::web::Data;
 use actix_web::web::Payload;
 use actix_web::web::ServiceConfig;
 use actix_web::web::get;
-use actix_ws::Session;
 use anyhow::Result;
 use log::debug;
 use log::error;
 use serde_json::Error as SerdeJsonError;
 use serde_json::from_str;
+use tokio::sync::mpsc::UnboundedSender;
 use tokio_util::sync::CancellationToken;
 
 use paddler_messaging::api_path::ApiPath;
-use paddler_messaging::generated_token_result::GeneratedTokenResult;
 use paddler_messaging::inference_client::message::Message as OutgoingMessage;
-use paddler_messaging::inference_client::response::Response as OutgoingResponse;
 use paddler_messaging::inference_server::identified_request::IdentifiedRequest;
 use paddler_messaging::inference_server::message::Message as InferenceServerMessage;
 use paddler_messaging::inference_server::notification::Notification as InferenceServerNotification;
 use paddler_messaging::inference_server::request::Request as InferenceServerRequest;
 use paddler_messaging::jsonrpc::error::Error as JsonRpcError;
 use paddler_messaging::jsonrpc::request_envelope::RequestEnvelope;
-use paddler_messaging::jsonrpc::response_envelope::ResponseEnvelope;
 use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters_schema::raw_parameters_schema::RawParametersSchema;
-use paddler_messaging::streamable_result::StreamableResult;
 use paddler_messaging::validates::Validates as _;
 use paddler_request_registry::request_registry_guard::RequestRegistryGuard;
 
 use self::inference_socket_controller_context::InferenceSocketControllerContext;
-use self::spawn_token_generation_mode_watcher::spawn_token_generation_mode_watcher;
+use crate::agent_relay::AgentRelay;
 use crate::agent_streaming_request::AgentStreamingRequest;
-use crate::balancer_applicable_state_holder::BalancerApplicableStateHolder;
+use crate::awaiting_agent_relay::AwaitingAgentRelay;
 use crate::buffered_request_manager::BufferedRequestManager;
-use crate::cluster_token_generation_mode::ClusterTokenGenerationMode;
-use crate::cluster_token_generation_mode::TOKEN_GENERATION_DISABLED_MESSAGE;
 use crate::continuation_decision::ContinuationDecision;
-use crate::controls_session::ControlsSession as _;
 use crate::controls_websocket_endpoint::ControlsWebSocketEndpoint;
+use crate::dispatchable_request::DispatchableRequest;
+use crate::forward_agent_relay::forward_agent_relay;
 use crate::inference_service::app_data::AppData;
 use crate::inference_service::configuration::Configuration as InferenceServiceConfiguration;
 use crate::invalid_request_parameters_description::invalid_request_parameters_description;
-use crate::request_from_agent::request_from_agent;
+use crate::outgoing_message_relay_receiver::OutgoingMessageRelayReceiver;
 use crate::respond_with_error::respond_with_error;
+use crate::websocket_outbound_frame::WebSocketOutboundFrame;
 use crate::websocket_session_controller::WebSocketSessionController;
 
-async fn send_token_generation_disabled(
-    request_id: String,
-    websocket_session_controller: &mut WebSocketSessionController<OutgoingMessage>,
-) {
-    websocket_session_controller
-        .send_response_safe(OutgoingMessage::Response(ResponseEnvelope {
-            generated_by: None,
-            request_id,
-            response: OutgoingResponse::GeneratedToken(
-                GeneratedTokenResult::TokenGenerationDisabled(
-                    TOKEN_GENERATION_DISABLED_MESSAGE.to_owned(),
-                ),
-            ),
-        }))
-        .await;
-}
-
 async fn handle_inference_request<TParams>(
     connection_close: &CancellationToken,
     context: Arc<InferenceSocketControllerContext>,
@@ -77,57 +54,52 @@ async fn handle_inference_request<TParams>(
     request_id: String,
     mut websocket_session_controller: WebSocketSessionController<OutgoingMessage>,
 ) where
-    TParams: AgentStreamingRequest + Debug + Send + 'static,
-    TParams::Response: Debug + Into<OutgoingResponse> + StreamableResult,
+    TParams: AgentStreamingRequest + Send + 'static,
 {
-    match context
-        .balancer_applicable_state_holder
-        .token_generation_mode()
-    {
-        ClusterTokenGenerationMode::DisabledForEmbeddings => {
-            send_token_generation_disabled(request_id, &mut websocket_session_controller).await;
-        }
-        ClusterTokenGenerationMode::Enabled => {
-            let request_close = connection_close.child_token();
-
-            let Some(request_registration) = RequestRegistryGuard::register(
-                &context.request_cancellation_tokens,
-                request_id.clone(),
-                request_close.clone(),
-            ) else {
-                error!("Rejecting duplicate inference request {request_id:?}");
-
-                respond_with_error(
-                    JsonRpcError {
-                        code: 400,
-                        description: format!(
-                            "Request id {request_id:?} is already in flight on this connection"
-                        ),
-                    },
-                    request_id,
-                    &mut websocket_session_controller,
-                )
-                .await;
-
-                return;
-            };
+    let request_close = connection_close.child_token();
+
+    let Some(request_registration) = RequestRegistryGuard::register(
+        &context.request_cancellation_tokens,
+        request_id.clone(),
+        request_close.clone(),
+    ) else {
+        error!("Rejecting duplicate inference request {request_id:?}");
+
+        respond_with_error(
+            JsonRpcError {
+                code: 400,
+                description: format!(
+                    "Request id {request_id:?} is already in flight on this connection"
+                ),
+            },
+            request_id,
+            &mut websocket_session_controller,
+        )
+        .await;
 
-            rt::spawn(async move {
-                let _request_registration = request_registration;
+        return;
+    };
 
-                request_from_agent(
-                    context.buffered_request_manager.clone(),
-                    request_close,
-                    context.inference_service_configuration.clone(),
-                    params,
-                    request_id,
-                    websocket_session_controller,
-                    context.shutdown.clone(),
-                )
-                .await;
-            });
-        }
-    }
+    rt::spawn(async move {
+        let _request_registration = request_registration;
+
+        forward_agent_relay(
+            AgentRelay::awaiting(AwaitingAgentRelay {
+                buffered_request_manager: context.buffered_request_manager.clone(),
+                dispatchable_request: DispatchableRequest::from(params),
+                inference_item_timeout: context
+                    .inference_service_configuration
+                    .inference_item_timeout,
+                shutdown: context.shutdown.clone(),
+            }),
+            request_close,
+            OutgoingMessageRelayReceiver {
+                request_id,
+                session_controller: websocket_session_controller,
+            },
+        )
+        .await;
+    });
 }
 
 async fn respond(
@@ -136,7 +108,6 @@ async fn respond(
     http_request: HttpRequest,
 ) -> Result<HttpResponse, Error> {
     let inference_socket_controller = InferenceSocketController {
-        balancer_applicable_state_holder: app_data.balancer_applicable_state_holder.clone(),
         buffered_request_manager: app_data.buffered_request_manager.clone(),
         inference_service_configuration: app_data.inference_service_configuration.clone(),
         shutdown: app_data.shutdown.clone(),
@@ -150,7 +121,6 @@ type InferenceJsonRpcMessage = InferenceServerMessage<RawParametersSchema>;
 type InferenceJsonRpcRequest = InferenceServerRequest<RawParametersSchema>;
 
 struct InferenceSocketController {
-    balancer_applicable_state_holder: Arc<BalancerApplicableStateHolder>,
     buffered_request_manager: Arc<BufferedRequestManager>,
     inference_service_configuration: InferenceServiceConfiguration,
     shutdown: CancellationToken,
@@ -163,7 +133,6 @@ impl ControlsWebSocketEndpoint for InferenceSocketController {
 
     fn create_context(&self) -> Self::Context {
         InferenceSocketControllerContext {
-            balancer_applicable_state_holder: self.balancer_applicable_state_holder.clone(),
             buffered_request_manager: self.buffered_request_manager.clone(),
             inference_service_configuration: self.inference_service_configuration.clone(),
             request_cancellation_tokens: Arc::default(),
@@ -270,15 +239,10 @@ impl ControlsWebSocketEndpoint for InferenceSocketController {
     }
 
     async fn on_connection_start(
-        connection_close: CancellationToken,
-        context: Arc<Self::Context>,
-        session: &mut Session,
+        _connection_close: CancellationToken,
+        _context: Arc<Self::Context>,
+        _outbound_tx: &UnboundedSender<WebSocketOutboundFrame>,
     ) {
-        spawn_token_generation_mode_watcher(
-            context.balancer_applicable_state_holder.clone(),
-            connection_close,
-            session.clone(),
-        );
     }
 }
 
diff --git a/paddler_balancer/src/inference_service/http_route/api/ws_inference_socket/spawn_token_generation_mode_watcher.rs b/paddler_balancer/src/inference_service/http_route/api/ws_inference_socket/spawn_token_generation_mode_watcher.rs
deleted file mode 100644
index a3a9112ea..000000000
--- a/paddler_balancer/src/inference_service/http_route/api/ws_inference_socket/spawn_token_generation_mode_watcher.rs
+++ /dev/null
@@ -1,81 +0,0 @@
-use std::sync::Arc;
-
-use actix_web::rt;
-use actix_ws::Session;
-use log::error;
-use tokio::select;
-use tokio_util::sync::CancellationToken;
-
-use paddler_messaging::inference_client::message::Message as OutgoingMessage;
-use paddler_messaging::inference_client::notification::Notification;
-use paddler_messaging::subscribes_to_updates::SubscribesToUpdates as _;
-
-use crate::balancer_applicable_state_holder::BalancerApplicableStateHolder;
-use crate::cluster_token_generation_mode::ClusterTokenGenerationMode;
-use crate::controls_session::ControlsSession as _;
-use crate::websocket_session_controller::WebSocketSessionController;
-
-async fn send_notification(
-    session_controller: &mut WebSocketSessionController<OutgoingMessage>,
-    notification: Notification,
-) -> bool {
-    match session_controller
-        .send_response(OutgoingMessage::Notification(notification))
-        .await
-    {
-        Ok(()) => true,
-        Err(err) => {
-            error!("Failed to push token generation mode notification: {err}");
-
-            false
-        }
-    }
-}
-
-pub fn spawn_token_generation_mode_watcher(
-    balancer_applicable_state_holder: Arc<BalancerApplicableStateHolder>,
-    connection_close: CancellationToken,
-    session: Session,
-) {
-    rt::spawn(async move {
-        let mut update_rx = balancer_applicable_state_holder.subscribe_to_updates();
-        let mut session_controller = WebSocketSessionController::<OutgoingMessage>::new(session);
-        let mut last_mode = balancer_applicable_state_holder.token_generation_mode();
-
-        if last_mode == ClusterTokenGenerationMode::DisabledForEmbeddings
-            && !send_notification(
-                &mut session_controller,
-                Notification::TokenGenerationDisabled,
-            )
-            .await
-        {
-            return;
-        }
-
-        loop {
-            select! {
-                () = connection_close.cancelled() => break,
-                Ok(()) = update_rx.changed() => {
-                    let current_mode = balancer_applicable_state_holder.token_generation_mode();
-
-                    if current_mode == last_mode {
-                        continue;
-                    }
-
-                    last_mode = current_mode;
-
-                    let notification = match current_mode {
-                        ClusterTokenGenerationMode::Enabled => Notification::TokenGenerationEnabled,
-                        ClusterTokenGenerationMode::DisabledForEmbeddings => {
-                            Notification::TokenGenerationDisabled
-                        }
-                    };
-
-                    if !send_notification(&mut session_controller, notification).await {
-                        break;
-                    }
-                }
-            }
-        }
-    });
-}
diff --git a/paddler_balancer/src/inference_service/mod.rs b/paddler_balancer/src/inference_service/mod.rs
index 293d4ace5..8572fe6e0 100644
--- a/paddler_balancer/src/inference_service/mod.rs
+++ b/paddler_balancer/src/inference_service/mod.rs
@@ -23,6 +23,7 @@ use crate::inference_service::app_data::AppData;
 use crate::inference_service::configuration::Configuration as InferenceServiceConfiguration;
 use crate::inference_service::http_route::api::post_continue_from_conversation_history::post_continue_from_conversation_history;
 use crate::inference_service::http_route::api::post_continue_from_raw_prompt::post_continue_from_raw_prompt;
+use crate::inference_service::http_route::api::post_decide::post_decide;
 use crate::inference_service::http_route::api::post_generate_embedding_batch::post_generate_embedding_batch;
 use crate::inference_service::http_route::api::ws_inference_socket::ws_inference_socket;
 use crate::run_http_service::run_http_service;
@@ -37,13 +38,8 @@ pub struct InferenceService {
     pub web_admin_panel_addr: Option<SocketAddr>,
 }
 
-#[async_trait]
-impl Service for InferenceService {
-    fn name(&self) -> &'static str {
-        "balancer::inference_service"
-    }
-
-    async fn run(self: Box<Self>, shutdown: CancellationToken) -> Result<()> {
+impl InferenceService {
+    async fn serve_until_shutdown(self: Box<Self>, shutdown: CancellationToken) -> Result<()> {
         let service_name = self.name();
         let cors_allowed_hosts_arc = cors_allowed_hosts_with_web_admin_panel(
             &self.configuration.cors_allowed_hosts,
@@ -68,6 +64,7 @@ impl Service for InferenceService {
                         .configure(get_health)
                         .configure(post_continue_from_conversation_history)
                         .configure(post_continue_from_raw_prompt)
+                        .configure(post_decide)
                         .configure(post_generate_embedding_batch)
                         .configure(ws_inference_socket)
                 },
@@ -79,3 +76,14 @@ impl Service for InferenceService {
         .await
     }
 }
+
+#[async_trait]
+impl Service for InferenceService {
+    fn name(&self) -> &'static str {
+        "balancer::inference_service"
+    }
+
+    async fn run(self: Box<Self>, shutdown: CancellationToken) -> Result<()> {
+        self.serve_until_shutdown(shutdown).await
+    }
+}
diff --git a/paddler_balancer/src/into_reserved_agent_request.rs b/paddler_balancer/src/into_reserved_agent_request.rs
new file mode 100644
index 000000000..40c589bc5
--- /dev/null
+++ b/paddler_balancer/src/into_reserved_agent_request.rs
@@ -0,0 +1,21 @@
+use std::num::NonZeroU16;
+
+use paddler_messaging::management_socket::agent::request::Request as AgentJsonRpcRequest;
+use paddler_messaging::reserved_request::ReservedRequest;
+
+pub trait IntoReservedAgentRequest: Sized {
+    fn into_reserved_agent_request(self, reserved_slots: NonZeroU16) -> AgentJsonRpcRequest;
+}
+
+impl<TParams> IntoReservedAgentRequest for TParams
+where
+    ReservedRequest<TParams>: Into<AgentJsonRpcRequest>,
+{
+    fn into_reserved_agent_request(self, reserved_slots: NonZeroU16) -> AgentJsonRpcRequest {
+        ReservedRequest {
+            params: self,
+            reserved_slots,
+        }
+        .into()
+    }
+}
diff --git a/paddler_balancer/src/lib.rs b/paddler_balancer/src/lib.rs
index c64a240ad..5b6d228dc 100644
--- a/paddler_balancer/src/lib.rs
+++ b/paddler_balancer/src/lib.rs
@@ -1,54 +1,73 @@
+pub mod accept_newer_status_report;
 pub mod agent_controller;
+pub mod agent_controller_params;
 pub mod agent_controller_pool;
 mod agent_controller_pool_total_slots;
+pub mod agent_controller_pool_updates;
 pub mod agent_controller_registration;
 pub mod agent_controller_slot_guard;
 pub mod agent_controller_update_result;
 pub mod agent_desired_state_notification;
 pub mod agent_id_path_params;
-pub mod agent_relay_error;
+pub mod agent_relay;
+pub mod agent_relay_conclusion;
+pub mod agent_relay_event;
+pub mod agent_relay_phase;
+pub mod agent_relay_step;
+pub mod agent_response_forwarder;
 mod agent_response_forwarding_mode;
 pub mod agent_response_receiver;
 pub mod agent_response_senders;
 mod agent_stop_follow_up;
 pub mod agent_streaming_request;
+pub mod awaiting_agent_relay;
 pub mod balancer_addresses;
 pub mod balancer_applicable_state;
 pub mod balancer_applicable_state_holder;
+pub mod balancer_desired_state_store;
 pub mod balancer_gauge;
 pub mod balancer_metrics;
-pub mod balancer_shutdown_error;
 mod buffered_request_agent_wait_result;
 mod buffered_request_count_guard;
 mod buffered_request_counter;
 pub mod buffered_request_manager;
 pub mod cancellation_token_stream_guard;
 pub mod chunk_forwarding_session_controller;
-pub mod cluster_token_generation_mode;
+pub mod cluster_serves_no_decisions;
 pub mod compatibility;
 pub mod continuation_decision;
 pub mod controls_session;
 pub mod controls_websocket_endpoint;
 pub mod cors_allowed_hosts_with_web_admin_panel;
 pub mod create_cors_middleware;
-pub mod desired_state_delivery;
 pub mod dispatch_candidate;
+pub mod dispatch_position;
+pub mod dispatch_requirement;
+pub mod dispatch_waiting_line;
+pub mod dispatchable_request;
 pub mod dispatched_agent;
+pub mod dispatched_agent_relay;
+pub mod distinct_embedding_documents;
 pub mod embedding_chunk_dispatch;
-pub mod forward_responses_stream;
+pub mod embedding_document_duplicates;
+pub mod enqueue_websocket_frame;
+pub mod forward_agent_relay;
 pub mod forwarding_event;
 pub mod forwarding_plan;
 pub mod http_listener;
 mod http_route;
 pub mod inference_service;
+pub mod into_reserved_agent_request;
 pub mod invalid_request_parameters_description;
 pub mod management_service;
 pub mod max_websocket_message_size;
 mod ndjson_response;
+pub mod outgoing_message_relay_receiver;
+pub mod receives_agent_relay;
 pub mod reconciliation_service;
 pub mod registered_agent_controller_guard;
-pub mod request_from_agent;
-pub mod request_from_dispatched_agent;
+pub mod relay_failure;
+pub mod requirements_waiting_ahead;
 pub mod resolved_socket_addr;
 pub mod respond_with_agent_response;
 pub mod respond_with_error;
@@ -58,7 +77,7 @@ pub mod response_senders;
 pub mod run_http_service;
 pub mod run_http_service_parameters;
 pub mod serve_http_until_shutdown;
-pub mod session_keep_alive;
+pub mod slot_demand;
 pub mod snapshots_stream;
 pub mod sse_response_from_snapshots;
 #[cfg(feature = "web_admin_panel")]
@@ -66,7 +85,10 @@ pub mod static_files;
 pub mod statsd_service;
 mod unbounded_stream_from_agent;
 mod unbounded_stream_from_agent_params;
+pub mod waiting_line_place;
 #[cfg(feature = "web_admin_panel")]
 pub mod web_admin_panel_service;
 pub mod websocket_close_cause;
+pub mod websocket_outbound_frame;
 pub mod websocket_session_controller;
+pub mod write_websocket_session;
diff --git a/paddler_balancer/src/management_service/app_data.rs b/paddler_balancer/src/management_service/app_data.rs
index 414973182..45b2ac3fc 100644
--- a/paddler_balancer/src/management_service/app_data.rs
+++ b/paddler_balancer/src/management_service/app_data.rs
@@ -2,19 +2,16 @@ use std::sync::Arc;
 
 use tokio_util::sync::CancellationToken;
 
-use paddler_state_database::state_database::StateDatabase;
-
 use crate::agent_controller_pool::AgentControllerPool;
-use crate::agent_response_senders::AgentResponseSenders;
 use crate::balancer_applicable_state_holder::BalancerApplicableStateHolder;
+use crate::balancer_desired_state_store::BalancerDesiredStateStore;
 use crate::buffered_request_manager::BufferedRequestManager;
 
 pub struct AppData {
     pub agent_controller_pool: Arc<AgentControllerPool>,
     pub balancer_applicable_state_holder: Arc<BalancerApplicableStateHolder>,
+    pub balancer_desired_state_store: Arc<BalancerDesiredStateStore>,
     pub buffered_request_manager: Arc<BufferedRequestManager>,
-    pub agent_response_senders: AgentResponseSenders,
     pub shutdown: CancellationToken,
-    pub state_database: Arc<dyn StateDatabase>,
     pub statsd_prefix: String,
 }
diff --git a/paddler_balancer/src/management_service/http_route/api/get_balancer_applicable_state.rs b/paddler_balancer/src/management_service/http_route/api/get_balancer_applicable_state.rs
index 65a505be2..ad227eed1 100644
--- a/paddler_balancer/src/management_service/http_route/api/get_balancer_applicable_state.rs
+++ b/paddler_balancer/src/management_service/http_route/api/get_balancer_applicable_state.rs
@@ -32,74 +32,58 @@ mod tests {
     use actix_web::test::init_service;
     use actix_web::test::read_body_json;
     use actix_web::web::Data;
-    use tokio::sync::watch;
     use tokio_util::sync::CancellationToken;
 
-    use paddler_inference_parameters::inference_parameters::InferenceParameters;
     use paddler_messaging::agent_desired_model::AgentDesiredModel;
     use paddler_messaging::agent_desired_state::AgentDesiredState;
     use paddler_messaging::api_path::ApiPath;
     use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-    use paddler_state_database::memory::Memory;
+    use paddler_state_database::state_database::StateDatabase;
+    use paddler_state_database::stored_balancer_desired_state::StoredBalancerDesiredState;
 
     use super::get_balancer_applicable_state;
     use crate::agent_controller_pool::AgentControllerPool;
-    use crate::agent_response_senders::AgentResponseSenders;
-    use crate::balancer_applicable_state::BalancerApplicableState;
-    use crate::balancer_applicable_state_holder::BalancerApplicableStateHolder;
+    use crate::balancer_desired_state_store::BalancerDesiredStateStore;
     use crate::buffered_request_manager::BufferedRequestManager;
     use crate::management_service::app_data::AppData;
 
-    fn build_app_data(
-        balancer_applicable_state_holder: Arc<BalancerApplicableStateHolder>,
-    ) -> Data<AppData> {
-        let (balancer_desired_state_notify_tx, _balancer_desired_state_notify_rx) =
-            watch::channel(BalancerDesiredState::default());
-
-        Data::new(AppData {
-            agent_controller_pool: Arc::new(AgentControllerPool::default()),
-            balancer_applicable_state_holder,
-            buffered_request_manager: Arc::new(BufferedRequestManager::new(
-                Arc::new(AgentControllerPool::default()),
-                Duration::from_secs(1),
-                10,
-            )),
-            agent_response_senders: AgentResponseSenders::default(),
-            shutdown: CancellationToken::new(),
-            state_database: Arc::new(Memory::new(
-                balancer_desired_state_notify_tx,
-                BalancerDesiredState::default(),
-            )),
-            statsd_prefix: "paddler".to_owned(),
-        })
-    }
-
     #[actix_web::test]
-    async fn responds_with_stored_agent_desired_state() {
-        let balancer_applicable_state_holder = Arc::new(BalancerApplicableStateHolder::new(
-            BalancerApplicableState::from(BalancerDesiredState::default()),
-        ));
-
-        balancer_applicable_state_holder.set_balancer_applicable_state(BalancerApplicableState {
-            agent_desired_state: AgentDesiredState {
-                chat_template_override: None,
-                inference_parameters: InferenceParameters::default(),
-                model: AgentDesiredModel::LocalToAgent("model.gguf".to_owned()),
-                multimodal_projection: AgentDesiredModel::None,
-            },
-        });
-
-        let app_data = build_app_data(balancer_applicable_state_holder);
+    async fn responds_with_the_agent_desired_state_derived_from_the_desired_state() {
+        let balancer_desired_state_store =
+            BalancerDesiredStateStore::load(Arc::new(StateDatabase::in_memory(
+                StoredBalancerDesiredState::Stored(Box::new(BalancerDesiredState {
+                    model: AgentDesiredModel::Uri("model.gguf".to_owned()),
+                    ..BalancerDesiredState::default()
+                })),
+            )))
+            .await
+            .expect("the desired state store must load");
         let app = init_service(
             App::new()
-                .app_data(app_data)
+                .app_data(Data::new(AppData {
+                    agent_controller_pool: Arc::new(AgentControllerPool::default()),
+                    balancer_applicable_state_holder: Arc::new(
+                        balancer_desired_state_store.applicable_state_holder(),
+                    ),
+                    balancer_desired_state_store: Arc::new(balancer_desired_state_store),
+                    buffered_request_manager: Arc::new(BufferedRequestManager::new(
+                        Arc::new(AgentControllerPool::default()),
+                        Duration::from_secs(1),
+                        10,
+                    )),
+                    shutdown: CancellationToken::new(),
+                    statsd_prefix: "paddler".to_owned(),
+                }))
                 .configure(get_balancer_applicable_state),
         )
         .await;
-        let request = TestRequest::get()
-            .uri(ApiPath::BALANCER_APPLICABLE_STATE)
-            .to_request();
-        let response = call_service(&app, request).await;
+        let response = call_service(
+            &app,
+            TestRequest::get()
+                .uri(ApiPath::BALANCER_APPLICABLE_STATE)
+                .to_request(),
+        )
+        .await;
 
         assert_eq!(response.status(), StatusCode::OK);
 
@@ -107,7 +91,7 @@ mod tests {
 
         assert_eq!(
             agent_desired_state.model,
-            AgentDesiredModel::LocalToAgent("model.gguf".to_owned())
+            AgentDesiredModel::Uri("model.gguf".to_owned())
         );
     }
 }
diff --git a/paddler_balancer/src/management_service/http_route/api/get_balancer_desired_state.rs b/paddler_balancer/src/management_service/http_route/api/get_balancer_desired_state.rs
index bc48887fa..aad3c014f 100644
--- a/paddler_balancer/src/management_service/http_route/api/get_balancer_desired_state.rs
+++ b/paddler_balancer/src/management_service/http_route/api/get_balancer_desired_state.rs
@@ -1,7 +1,5 @@
-use actix_web::Error;
 use actix_web::HttpResponse;
 use actix_web::Responder;
-use actix_web::error::ErrorInternalServerError;
 use actix_web::web;
 use actix_web::web::get;
 
@@ -9,14 +7,12 @@ use paddler_messaging::api_path::ApiPath;
 
 use crate::management_service::app_data::AppData;
 
-async fn respond(app_data: web::Data<AppData>) -> Result<impl Responder, Error> {
-    let desired_state = app_data
-        .state_database
-        .read_balancer_desired_state()
-        .await
-        .map_err(ErrorInternalServerError)?;
-
-    Ok(HttpResponse::Ok().json(desired_state))
+async fn respond(app_data: web::Data<AppData>) -> impl Responder {
+    HttpResponse::Ok().json(
+        app_data
+            .balancer_applicable_state_holder
+            .get_balancer_desired_state(),
+    )
 }
 
 pub fn get_balancer_desired_state(cfg: &mut web::ServiceConfig) {
@@ -35,70 +31,58 @@ mod tests {
     use actix_web::test::init_service;
     use actix_web::test::read_body_json;
     use actix_web::web::Data;
-    use tempfile::TempDir;
-    use tokio::sync::watch;
     use tokio_util::sync::CancellationToken;
 
-    use paddler_inference_parameters::inference_parameters::InferenceParameters;
     use paddler_messaging::agent_desired_model::AgentDesiredModel;
     use paddler_messaging::api_path::ApiPath;
     use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-    use paddler_state_database::file::File;
-    use paddler_state_database::memory::Memory;
     use paddler_state_database::state_database::StateDatabase;
+    use paddler_state_database::stored_balancer_desired_state::StoredBalancerDesiredState;
 
     use super::get_balancer_desired_state;
     use crate::agent_controller_pool::AgentControllerPool;
-    use crate::agent_response_senders::AgentResponseSenders;
-    use crate::balancer_applicable_state::BalancerApplicableState;
-    use crate::balancer_applicable_state_holder::BalancerApplicableStateHolder;
+    use crate::balancer_desired_state_store::BalancerDesiredStateStore;
     use crate::buffered_request_manager::BufferedRequestManager;
     use crate::management_service::app_data::AppData;
 
-    fn build_app_data(state_database: Arc<dyn StateDatabase>) -> Data<AppData> {
-        Data::new(AppData {
-            agent_controller_pool: Arc::new(AgentControllerPool::default()),
-            balancer_applicable_state_holder: Arc::new(BalancerApplicableStateHolder::new(
-                BalancerApplicableState::from(BalancerDesiredState::default()),
-            )),
-            buffered_request_manager: Arc::new(BufferedRequestManager::new(
-                Arc::new(AgentControllerPool::default()),
-                Duration::from_secs(1),
-                10,
-            )),
-            agent_response_senders: AgentResponseSenders::default(),
-            shutdown: CancellationToken::new(),
-            state_database,
-            statsd_prefix: "paddler".to_owned(),
-        })
-    }
-
     #[actix_web::test]
-    async fn responds_with_stored_desired_state() {
-        let (balancer_desired_state_notify_tx, _balancer_desired_state_notify_rx) =
-            watch::channel(BalancerDesiredState::default());
+    async fn responds_with_the_desired_state_the_balancer_holds() {
         let stored_state = BalancerDesiredState {
-            chat_template_override: None,
-            inference_parameters: InferenceParameters::default(),
-            model: AgentDesiredModel::LocalToAgent("model.gguf".to_owned()),
-            multimodal_projection: AgentDesiredModel::None,
-            use_chat_template_override: false,
+            model: AgentDesiredModel::Uri("model.gguf".to_owned()),
+            ..BalancerDesiredState::default()
         };
-        let state_database = Arc::new(Memory::new(
-            balancer_desired_state_notify_tx,
-            stored_state.clone(),
-        ));
-        let app_data = build_app_data(state_database);
+        let balancer_desired_state_store =
+            BalancerDesiredStateStore::load(Arc::new(StateDatabase::in_memory(
+                StoredBalancerDesiredState::Stored(Box::new(stored_state.clone())),
+            )))
+            .await
+            .expect("the desired state store must load");
         let app = init_service(
             App::new()
-                .app_data(app_data)
+                .app_data(Data::new(AppData {
+                    agent_controller_pool: Arc::new(AgentControllerPool::default()),
+                    balancer_applicable_state_holder: Arc::new(
+                        balancer_desired_state_store.applicable_state_holder(),
+                    ),
+                    balancer_desired_state_store: Arc::new(balancer_desired_state_store),
+                    buffered_request_manager: Arc::new(BufferedRequestManager::new(
+                        Arc::new(AgentControllerPool::default()),
+                        Duration::from_secs(1),
+                        10,
+                    )),
+                    shutdown: CancellationToken::new(),
+                    statsd_prefix: "paddler".to_owned(),
+                }))
                 .configure(get_balancer_desired_state),
         )
         .await;
-        let request = TestRequest::get()
-            .uri(ApiPath::BALANCER_DESIRED_STATE)
-            .to_request();
-        let response = call_service(&app, request).await;
+        let response = call_service(
+            &app,
+            TestRequest::get()
+                .uri(ApiPath::BALANCER_DESIRED_STATE)
+                .to_request(),
+        )
+        .await;
 
         assert_eq!(response.status(), StatusCode::OK);
 
@@ -106,28 +90,4 @@ mod tests {
 
         assert_eq!(desired_state, stored_state);
     }
-
-    #[actix_web::test]
-    async fn responds_with_internal_server_error_when_reading_state_fails() {
-        let (balancer_desired_state_notify_tx, _balancer_desired_state_notify_rx) =
-            watch::channel(BalancerDesiredState::default());
-        let temp_dir = TempDir::new().unwrap();
-        let state_database = Arc::new(File::new(
-            balancer_desired_state_notify_tx,
-            temp_dir.path().to_path_buf(),
-        ));
-        let app_data = build_app_data(state_database);
-        let app = init_service(
-            App::new()
-                .app_data(app_data)
-                .configure(get_balancer_desired_state),
-        )
-        .await;
-        let request = TestRequest::get()
-            .uri(ApiPath::BALANCER_DESIRED_STATE)
-            .to_request();
-        let response = call_service(&app, request).await;
-
-        assert_eq!(response.status(), StatusCode::INTERNAL_SERVER_ERROR);
-    }
 }
diff --git a/paddler_balancer/src/management_service/http_route/api/get_buffered_requests.rs b/paddler_balancer/src/management_service/http_route/api/get_buffered_requests.rs
index 2b0a5aa5f..60b866ff0 100644
--- a/paddler_balancer/src/management_service/http_route/api/get_buffered_requests.rs
+++ b/paddler_balancer/src/management_service/http_route/api/get_buffered_requests.rs
@@ -27,19 +27,16 @@ mod tests {
     use actix_web::test::init_service;
     use actix_web::test::read_body_json;
     use actix_web::web::Data;
-    use tokio::sync::watch;
     use tokio_util::sync::CancellationToken;
 
     use paddler_messaging::api_path::ApiPath;
-    use paddler_messaging::balancer_desired_state::BalancerDesiredState;
     use paddler_messaging::buffered_request_manager_snapshot::BufferedRequestManagerSnapshot;
-    use paddler_state_database::memory::Memory;
+    use paddler_state_database::state_database::StateDatabase;
+    use paddler_state_database::stored_balancer_desired_state::StoredBalancerDesiredState;
 
     use super::get_buffered_requests;
     use crate::agent_controller_pool::AgentControllerPool;
-    use crate::agent_response_senders::AgentResponseSenders;
-    use crate::balancer_applicable_state::BalancerApplicableState;
-    use crate::balancer_applicable_state_holder::BalancerApplicableStateHolder;
+    use crate::balancer_desired_state_store::BalancerDesiredStateStore;
     use crate::buffered_request_manager::BufferedRequestManager;
     use crate::management_service::app_data::AppData;
 
@@ -58,21 +55,19 @@ mod tests {
             .buffered_request_counter
             .try_admit();
 
-        let (balancer_desired_state_notify_tx, _balancer_desired_state_notify_rx) =
-            watch::channel(BalancerDesiredState::default());
-
+        let balancer_desired_state_store = BalancerDesiredStateStore::load(Arc::new(
+            StateDatabase::in_memory(StoredBalancerDesiredState::NothingStored),
+        ))
+        .await
+        .expect("the desired state store must load");
         let app_data = Data::new(AppData {
             agent_controller_pool: Arc::new(AgentControllerPool::default()),
-            balancer_applicable_state_holder: Arc::new(BalancerApplicableStateHolder::new(
-                BalancerApplicableState::from(BalancerDesiredState::default()),
-            )),
+            balancer_applicable_state_holder: Arc::new(
+                balancer_desired_state_store.applicable_state_holder(),
+            ),
+            balancer_desired_state_store: Arc::new(balancer_desired_state_store),
             buffered_request_manager,
-            agent_response_senders: AgentResponseSenders::default(),
             shutdown: CancellationToken::new(),
-            state_database: Arc::new(Memory::new(
-                balancer_desired_state_notify_tx,
-                BalancerDesiredState::default(),
-            )),
             statsd_prefix: "paddler".to_owned(),
         });
 
diff --git a/paddler_balancer/src/management_service/http_route/api/put_balancer_desired_state.rs b/paddler_balancer/src/management_service/http_route/api/put_balancer_desired_state.rs
index 444711666..5f5a81936 100644
--- a/paddler_balancer/src/management_service/http_route/api/put_balancer_desired_state.rs
+++ b/paddler_balancer/src/management_service/http_route/api/put_balancer_desired_state.rs
@@ -7,7 +7,6 @@ use actix_web::web::put;
 
 use paddler_messaging::api_path::ApiPath;
 use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-use paddler_messaging::subscribes_to_updates::SubscribesToUpdates as _;
 
 use crate::management_service::app_data::AppData;
 
@@ -15,18 +14,9 @@ async fn respond(
     app_data: web::Data<AppData>,
     balancer_desired_state: web::Json<BalancerDesiredState>,
 ) -> Result<impl Responder, Error> {
-    let mut applied_state_rx = app_data
-        .balancer_applicable_state_holder
-        .subscribe_to_updates();
-
     app_data
-        .state_database
-        .store_balancer_desired_state(&balancer_desired_state.into_inner())
-        .await
-        .map_err(ErrorInternalServerError)?;
-
-    applied_state_rx
-        .changed()
+        .balancer_desired_state_store
+        .replace(balancer_desired_state.into_inner())
         .await
         .map_err(ErrorInternalServerError)?;
 
@@ -44,7 +34,6 @@ mod tests {
 
     use actix_web::App;
     use actix_web::http::StatusCode;
-    use actix_web::rt::spawn;
     use actix_web::test::TestRequest;
     use actix_web::test::call_service;
     use actix_web::test::init_service;
@@ -52,68 +41,55 @@ mod tests {
     use serde_json::json;
     use serde_json::to_value;
     use tempfile::TempDir;
-    use tokio::sync::watch;
     use tokio_util::sync::CancellationToken;
-    use trzcina::Service as _;
 
     use paddler_messaging::agent_desired_model::AgentDesiredModel;
     use paddler_messaging::api_path::ApiPath;
     use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-    use paddler_state_database::file::File;
-    use paddler_state_database::memory::Memory;
     use paddler_state_database::state_database::StateDatabase;
+    use paddler_state_database::state_database_file::StateDatabaseFile;
+    use paddler_state_database::stored_balancer_desired_state::StoredBalancerDesiredState;
 
     use super::put_balancer_desired_state;
     use crate::agent_controller_pool::AgentControllerPool;
-    use crate::agent_response_senders::AgentResponseSenders;
-    use crate::balancer_applicable_state::BalancerApplicableState;
-    use crate::balancer_applicable_state_holder::BalancerApplicableStateHolder;
+    use crate::balancer_desired_state_store::BalancerDesiredStateStore;
     use crate::buffered_request_manager::BufferedRequestManager;
     use crate::management_service::app_data::AppData;
-    use crate::reconciliation_service::ReconciliationService;
 
-    fn build_app_data(state_database: Arc<dyn StateDatabase>) -> Data<AppData> {
+    async fn build_app_data(state_database: Arc<StateDatabase>) -> Data<AppData> {
+        let balancer_desired_state_store = BalancerDesiredStateStore::load(state_database)
+            .await
+            .expect("the desired state store must load");
+
         Data::new(AppData {
             agent_controller_pool: Arc::new(AgentControllerPool::default()),
-            balancer_applicable_state_holder: Arc::new(BalancerApplicableStateHolder::new(
-                BalancerApplicableState::from(BalancerDesiredState::default()),
-            )),
+            balancer_applicable_state_holder: Arc::new(
+                balancer_desired_state_store.applicable_state_holder(),
+            ),
+            balancer_desired_state_store: Arc::new(balancer_desired_state_store),
             buffered_request_manager: Arc::new(BufferedRequestManager::new(
                 Arc::new(AgentControllerPool::default()),
                 Duration::from_secs(1),
                 10,
             )),
-            agent_response_senders: AgentResponseSenders::default(),
             shutdown: CancellationToken::new(),
-            state_database,
             statsd_prefix: "paddler".to_owned(),
         })
     }
 
     #[actix_web::test]
-    async fn responds_once_the_stored_state_is_applicable() {
-        let (balancer_desired_state_notify_tx, balancer_desired_state_rx) =
-            watch::channel(BalancerDesiredState::default());
-        let app_data = build_app_data(Arc::new(Memory::new(
-            balancer_desired_state_notify_tx,
-            BalancerDesiredState::default(),
-        )));
-        let reconciliation_shutdown = CancellationToken::new();
-        let reconciliation = spawn(
-            Box::new(ReconciliationService {
-                agent_controller_pool: app_data.agent_controller_pool.clone(),
-                balancer_applicable_state_holder: app_data.balancer_applicable_state_holder.clone(),
-                balancer_desired_state_rx,
-            })
-            .run(reconciliation_shutdown.clone()),
-        );
+    async fn applies_the_desired_state_before_responding() {
+        let app_data = build_app_data(Arc::new(StateDatabase::in_memory(
+            StoredBalancerDesiredState::NothingStored,
+        )))
+        .await;
         let app = init_service(
             App::new()
                 .app_data(app_data.clone())
                 .configure(put_balancer_desired_state),
         )
         .await;
-        let applied_model = AgentDesiredModel::LocalToAgent("applied-model".to_owned());
+        let applied_model = AgentDesiredModel::Uri("applied-model".to_owned());
         let request = TestRequest::put()
             .uri(ApiPath::BALANCER_DESIRED_STATE)
             .set_json(BalancerDesiredState {
@@ -133,63 +109,68 @@ mod tests {
                 .model,
             applied_model
         );
-
-        reconciliation_shutdown.cancel();
-        reconciliation.await.unwrap().unwrap();
     }
 
     #[actix_web::test]
-    async fn responds_with_bad_request_when_inference_parameters_are_invalid() {
-        let (balancer_desired_state_notify_tx, balancer_desired_state_notify_rx) =
-            watch::channel(BalancerDesiredState::default());
-        let state_database = Arc::new(Memory::new(
-            balancer_desired_state_notify_tx,
-            BalancerDesiredState::default(),
-        ));
-        let app_data = build_app_data(state_database);
+    async fn responds_with_bad_request_when_model_runtime_parameters_are_invalid() {
         let app = init_service(
             App::new()
-                .app_data(app_data)
+                .app_data(
+                    build_app_data(Arc::new(StateDatabase::in_memory(
+                        StoredBalancerDesiredState::NothingStored,
+                    )))
+                    .await,
+                )
                 .configure(put_balancer_desired_state),
         )
         .await;
         let mut invalid_desired_state = to_value(BalancerDesiredState::default()).unwrap();
 
-        invalid_desired_state["inference_parameters"]["penalty_last_n"] = json!(-1);
+        invalid_desired_state["model_runtime_parameters"]["n_gpu_layers"] = json!(-2);
 
         let request = TestRequest::put()
             .uri(ApiPath::BALANCER_DESIRED_STATE)
             .set_json(invalid_desired_state)
             .to_request();
-        let response = call_service(&app, request).await;
-
-        assert_eq!(response.status(), StatusCode::BAD_REQUEST);
 
-        drop(balancer_desired_state_notify_rx);
+        assert_eq!(
+            call_service(&app, request).await.status(),
+            StatusCode::BAD_REQUEST
+        );
     }
 
     #[actix_web::test]
     async fn responds_with_internal_server_error_when_store_fails() {
-        let (balancer_desired_state_notify_tx, _balancer_desired_state_notify_rx) =
-            watch::channel(BalancerDesiredState::default());
         let directory_in_place_of_the_state_file =
             TempDir::new().expect("a temporary directory must be creatable");
-        let app_data = build_app_data(Arc::new(File::new(
-            balancer_desired_state_notify_tx,
-            directory_in_place_of_the_state_file.path().to_path_buf(),
-        )));
         let app = init_service(
             App::new()
-                .app_data(app_data)
+                .app_data(
+                    build_app_data(Arc::new(StateDatabase::File(StateDatabaseFile::new(
+                        directory_in_place_of_the_state_file
+                            .path()
+                            .join("state.json"),
+                    ))))
+                    .await,
+                )
                 .configure(put_balancer_desired_state),
         )
         .await;
         let request = TestRequest::put()
             .uri(ApiPath::BALANCER_DESIRED_STATE)
-            .set_json(BalancerDesiredState::default())
+            .set_json(BalancerDesiredState {
+                model: AgentDesiredModel::Uri("unstorable-model".to_owned()),
+                ..BalancerDesiredState::default()
+            })
             .to_request();
-        let response = call_service(&app, request).await;
 
-        assert_eq!(response.status(), StatusCode::INTERNAL_SERVER_ERROR);
+        directory_in_place_of_the_state_file
+            .close()
+            .expect("the temporary directory must be removable");
+
+        assert_eq!(
+            call_service(&app, request).await.status(),
+            StatusCode::INTERNAL_SERVER_ERROR
+        );
     }
 }
diff --git a/paddler_balancer/src/management_service/http_route/api/ws_agent_socket/agent_socket_controller_context.rs b/paddler_balancer/src/management_service/http_route/api/ws_agent_socket/agent_socket_controller_context.rs
index 2b34c4ae1..3436ced1c 100644
--- a/paddler_balancer/src/management_service/http_route/api/ws_agent_socket/agent_socket_controller_context.rs
+++ b/paddler_balancer/src/management_service/http_route/api/ws_agent_socket/agent_socket_controller_context.rs
@@ -1,16 +1,179 @@
 use std::sync::Arc;
 
+use actix_web::rt;
+use log::error;
+use log::info;
 use parking_lot::RwLock;
+use tokio::select;
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::jsonrpc::response_envelope::ResponseEnvelope;
+use paddler_messaging::management_socket::agent::message::Message as AgentJsonRpcMessage;
+use paddler_messaging::management_socket::agent::response::Response as AgentJsonRpcResponse;
+use paddler_messaging::management_socket::balancer::message::Message as ManagementJsonRpcMessage;
+use paddler_messaging::management_socket::balancer::notification::Notification as ManagementJsonRpcNotification;
+use paddler_messaging::management_socket::balancer::notification_params::register_agent_params::RegisterAgentParams;
+use paddler_messaging::management_socket::balancer::notification_params::update_agent_status_params::UpdateAgentStatusParams;
 
 use super::agent_socket_registration::AgentSocketRegistration;
+use crate::agent_controller::AgentController;
+use crate::agent_controller_params::AgentControllerParams;
 use crate::agent_controller_pool::AgentControllerPool;
-use crate::agent_response_senders::AgentResponseSenders;
+use crate::agent_controller_registration::AgentControllerRegistration;
+use crate::agent_controller_update_result::AgentControllerUpdateResult;
 use crate::balancer_applicable_state_holder::BalancerApplicableStateHolder;
+use crate::continuation_decision::ContinuationDecision;
+use crate::controls_session::ControlsSession as _;
+use crate::registered_agent_controller_guard::RegisteredAgentControllerGuard;
+use crate::websocket_close_cause::WebSocketCloseCause;
+use crate::websocket_session_controller::WebSocketSessionController;
 
 pub struct AgentSocketControllerContext {
     pub agent_controller_pool: Arc<AgentControllerPool>,
     pub agent_id: String,
-    pub agent_response_senders: AgentResponseSenders,
     pub balancer_applicable_state_holder: Arc<BalancerApplicableStateHolder>,
     pub registration: RwLock<AgentSocketRegistration>,
 }
+
+impl AgentSocketControllerContext {
+    pub fn handle_message(
+        &self,
+        connection_close: CancellationToken,
+        message: ManagementJsonRpcMessage,
+        websocket_session_controller: WebSocketSessionController<AgentJsonRpcMessage>,
+    ) -> ContinuationDecision {
+        match message {
+            ManagementJsonRpcMessage::Notification(
+                ManagementJsonRpcNotification::DeregisterAgent,
+            ) => {
+                connection_close.cancel();
+
+                ContinuationDecision::Stop(WebSocketCloseCause::AgentDeregistered)
+            }
+            ManagementJsonRpcMessage::Notification(
+                ManagementJsonRpcNotification::RegisterAgent(register_agent_params),
+            ) => self.register_agent(
+                connection_close,
+                register_agent_params,
+                websocket_session_controller,
+            ),
+            ManagementJsonRpcMessage::Notification(
+                ManagementJsonRpcNotification::UpdateAgentStatus(update_agent_status_params),
+            ) => self.update_agent_status(update_agent_status_params),
+            ManagementJsonRpcMessage::Response(ResponseEnvelope {
+                request_id,
+                response,
+                ..
+            }) => self.forward_response(&request_id, response),
+        }
+    }
+
+    fn forward_response(
+        &self,
+        request_id: &str,
+        response: AgentJsonRpcResponse,
+    ) -> ContinuationDecision {
+        match &*self.registration.read() {
+            AgentSocketRegistration::AwaitingRegistration => {
+                error!(
+                    "Rejecting a response from agent {} before its registration",
+                    self.agent_id
+                );
+
+                ContinuationDecision::Stop(WebSocketCloseCause::AgentResponseBeforeRegistration)
+            }
+            AgentSocketRegistration::Registered(agent_controller) => {
+                agent_controller
+                    .agent_response_senders
+                    .forward(request_id, response);
+
+                ContinuationDecision::Continue
+            }
+        }
+    }
+
+    fn register_agent(
+        &self,
+        connection_close: CancellationToken,
+        RegisterAgentParams {
+            name,
+            slot_aggregated_status_snapshot,
+        }: RegisterAgentParams,
+        mut websocket_session_controller: WebSocketSessionController<AgentJsonRpcMessage>,
+    ) -> ContinuationDecision {
+        let agent_controller = Arc::new(AgentController::new(AgentControllerParams {
+            connection_close: connection_close.clone(),
+            id: self.agent_id.clone(),
+            name,
+            reported_status: slot_aggregated_status_snapshot,
+        }));
+
+        let registered_agent_controller_guard =
+            match self.agent_controller_pool.register_agent_controller(
+                agent_controller.clone(),
+                &self.balancer_applicable_state_holder,
+            ) {
+                AgentControllerRegistration::DuplicateAgentId => {
+                    error!("Rejecting a second registration of agent {}", self.agent_id);
+
+                    return ContinuationDecision::Stop(WebSocketCloseCause::AgentAlreadyRegistered);
+                }
+                AgentControllerRegistration::Registered(registered_agent_controller_guard) => {
+                    registered_agent_controller_guard
+                }
+            };
+
+        *self.registration.write() = AgentSocketRegistration::Registered(agent_controller);
+
+        info!("Registered agent: {}", self.agent_id);
+
+        rt::spawn(async move {
+            let RegisteredAgentControllerGuard {
+                agent_controller, ..
+            } = &registered_agent_controller_guard;
+
+            loop {
+                select! {
+                    () = connection_close.cancelled() => {
+                        break;
+                    }
+                    message = agent_controller.agent_outbox.next() => {
+                        websocket_session_controller.send_response_safe(message).await;
+                    }
+                }
+            }
+        });
+
+        ContinuationDecision::Continue
+    }
+
+    fn update_agent_status(
+        &self,
+        UpdateAgentStatusParams {
+            slot_aggregated_status_snapshot,
+        }: UpdateAgentStatusParams,
+    ) -> ContinuationDecision {
+        match &*self.registration.read() {
+            AgentSocketRegistration::AwaitingRegistration => {
+                error!(
+                    "Rejecting a status update from agent {} before its registration",
+                    self.agent_id
+                );
+
+                ContinuationDecision::Stop(WebSocketCloseCause::AgentStatusBeforeRegistration)
+            }
+            AgentSocketRegistration::Registered(agent_controller) => {
+                match agent_controller
+                    .update_from_slot_aggregated_status_snapshot(slot_aggregated_status_snapshot)
+                {
+                    AgentControllerUpdateResult::NoMeaningfulChanges => {}
+                    AgentControllerUpdateResult::Updated => {
+                        self.agent_controller_pool.signal_update();
+                    }
+                }
+
+                ContinuationDecision::Continue
+            }
+        }
+    }
+}
diff --git a/paddler_balancer/src/management_service/http_route/api/ws_agent_socket/mod.rs b/paddler_balancer/src/management_service/http_route/api/ws_agent_socket/mod.rs
index 379bcdd0c..f9f99b398 100644
--- a/paddler_balancer/src/management_service/http_route/api/ws_agent_socket/mod.rs
+++ b/paddler_balancer/src/management_service/http_route/api/ws_agent_socket/mod.rs
@@ -1,76 +1,43 @@
 mod agent_socket_controller_context;
 mod agent_socket_registration;
 
+use std::future::Future;
+use std::future::ready;
 use std::sync::Arc;
-use std::sync::atomic::AtomicU64;
 
 use actix_web::Error;
 use actix_web::HttpRequest;
 use actix_web::HttpResponse;
-use actix_web::rt;
 use actix_web::web::Data;
 use actix_web::web::Path;
 use actix_web::web::Payload;
 use actix_web::web::ServiceConfig;
 use actix_web::web::get;
-use actix_ws::Session;
 use anyhow::Result;
-use log::error;
-use log::info;
-use log::warn;
 use parking_lot::RwLock;
 use serde_json::Error as SerdeJsonError;
-use tokio::select;
-use tokio::sync::mpsc;
+use tokio::sync::mpsc::UnboundedSender;
 use tokio_util::sync::CancellationToken;
 
 use paddler_messaging::api_path::ApiPath;
-use paddler_messaging::atomic_value::AtomicValue;
-use paddler_messaging::jsonrpc::response_envelope::ResponseEnvelope;
 use paddler_messaging::management_socket::agent::message::Message as AgentJsonRpcMessage;
 use paddler_messaging::management_socket::agent::notification::Notification as AgentJsonRpcNotification;
 use paddler_messaging::management_socket::agent::notification_params::version_params::VersionParams;
-use paddler_messaging::management_socket::agent::response::Response as AgentJsonRpcResponse;
 use paddler_messaging::management_socket::balancer::message::Message as ManagementJsonRpcMessage;
-use paddler_messaging::management_socket::balancer::notification::Notification as ManagementJsonRpcNotification;
-use paddler_messaging::management_socket::balancer::notification_params::register_agent_params::RegisterAgentParams;
-use paddler_messaging::management_socket::balancer::notification_params::update_agent_status_params::UpdateAgentStatusParams;
-use paddler_request_registry::request_delivery::RequestDelivery;
 
 use self::agent_socket_controller_context::AgentSocketControllerContext;
 use self::agent_socket_registration::AgentSocketRegistration;
-use crate::agent_controller::AgentController;
 use crate::agent_controller_pool::AgentControllerPool;
-use crate::agent_controller_registration::AgentControllerRegistration;
-use crate::agent_controller_update_result::AgentControllerUpdateResult;
-use crate::agent_desired_state_notification::agent_desired_state_notification;
 use crate::agent_id_path_params::AgentIdPathParams;
-use crate::agent_response_senders::AgentResponseSenders;
 use crate::balancer_applicable_state_holder::BalancerApplicableStateHolder;
 use crate::continuation_decision::ContinuationDecision;
 use crate::controls_session::ControlsSession as _;
 use crate::controls_websocket_endpoint::ControlsWebSocketEndpoint;
 use crate::management_service::app_data::AppData;
-use crate::response_senders::ResponseSenders;
 use crate::websocket_close_cause::WebSocketCloseCause;
+use crate::websocket_outbound_frame::WebSocketOutboundFrame;
 use crate::websocket_session_controller::WebSocketSessionController;
 
-fn forward_agent_response<TResponse>(
-    response_senders: &ResponseSenders<TResponse>,
-    request_id: &str,
-    response: TResponse,
-) {
-    match response_senders.send_to(request_id, response) {
-        RequestDelivery::Delivered => {}
-        RequestDelivery::ReceiverDropped => {
-            warn!("Dropped an agent response for request {request_id:?}: its receiver went away");
-        }
-        RequestDelivery::RequestNotRegistered => {
-            warn!("Dropped an agent response for unknown request {request_id:?}");
-        }
-    }
-}
-
 async fn respond(
     app_data: Data<AppData>,
     path_params: Path<AgentIdPathParams>,
@@ -81,7 +48,6 @@ async fn respond(
         agent_controller_pool: app_data.agent_controller_pool.clone(),
         agent_id: path_params.agent_id.clone(),
         balancer_applicable_state_holder: app_data.balancer_applicable_state_holder.clone(),
-        agent_response_senders: app_data.agent_response_senders.clone(),
     };
 
     agent_socket_controller.respond(payload, req, app_data.shutdown.clone())
@@ -91,7 +57,6 @@ struct AgentSocketController {
     agent_controller_pool: Arc<AgentControllerPool>,
     agent_id: String,
     balancer_applicable_state_holder: Arc<BalancerApplicableStateHolder>,
-    agent_response_senders: AgentResponseSenders,
 }
 
 impl ControlsWebSocketEndpoint for AgentSocketController {
@@ -103,191 +68,40 @@ impl ControlsWebSocketEndpoint for AgentSocketController {
         AgentSocketControllerContext {
             agent_controller_pool: self.agent_controller_pool.clone(),
             agent_id: self.agent_id.clone(),
-            agent_response_senders: self.agent_response_senders.clone(),
             balancer_applicable_state_holder: self.balancer_applicable_state_holder.clone(),
             registration: RwLock::new(AgentSocketRegistration::AwaitingRegistration),
         }
     }
 
-    async fn handle_deserialized_message(
+    fn handle_deserialized_message(
         connection_close: CancellationToken,
         context: Arc<Self::Context>,
         deserialized_message: Self::IncomingMessage,
-        mut websocket_session_controller: WebSocketSessionController<Self::OutgoingMessage>,
-    ) -> ContinuationDecision {
-        match deserialized_message {
-            ManagementJsonRpcMessage::Notification(
-                ManagementJsonRpcNotification::DeregisterAgent,
-            ) => {
-                connection_close.cancel();
-
-                ContinuationDecision::Stop(WebSocketCloseCause::AgentDeregistered)
-            }
-            ManagementJsonRpcMessage::Notification(
-                ManagementJsonRpcNotification::RegisterAgent(RegisterAgentParams {
-                    name,
-                    slot_aggregated_status_snapshot,
-                }),
-            ) => {
-                let (agent_message_tx, mut agent_message_rx) =
-                    mpsc::unbounded_channel::<AgentJsonRpcMessage>();
-                let agent_controller = Arc::new(AgentController {
-                    agent_message_tx,
-                    agent_response_senders: context.agent_response_senders.clone(),
-                    connection_close: connection_close.clone(),
-                    id: context.agent_id.clone(),
-                    name,
-                    reported_status: RwLock::new(slot_aggregated_status_snapshot),
-                    slots_processing: AtomicValue::<AtomicU64>::new(0),
-                });
-
-                let registered_agent_controller_guard = match context
-                    .agent_controller_pool
-                    .register_agent_controller(agent_controller.clone())
-                {
-                    AgentControllerRegistration::DuplicateAgentId => {
-                        error!(
-                            "Rejecting a second registration of agent {}",
-                            context.agent_id
-                        );
-
-                        return ContinuationDecision::Stop(
-                            WebSocketCloseCause::AgentAlreadyRegistered,
-                        );
-                    }
-                    AgentControllerRegistration::Registered(registered_agent_controller_guard) => {
-                        registered_agent_controller_guard
-                    }
-                };
-
-                *context.registration.write() =
-                    AgentSocketRegistration::Registered(agent_controller);
-
-                info!("Registered agent: {}", context.agent_id);
-
-                let agent_desired_state = agent_desired_state_notification(
-                    context
-                        .balancer_applicable_state_holder
-                        .get_agent_desired_state(),
-                );
-                let forwarder_close = connection_close;
-
-                rt::spawn(async move {
-                    let _registered_agent_controller_guard = registered_agent_controller_guard;
-
-                    websocket_session_controller
-                        .send_response_safe(agent_desired_state)
-                        .await;
-
-                    loop {
-                        select! {
-                            () = forwarder_close.cancelled() => {
-                                break;
-                            }
-                            Some(message) = agent_message_rx.recv() => {
-                                websocket_session_controller.send_response_safe(message).await;
-                            }
-                        }
-                    }
-                });
-
-                ContinuationDecision::Continue
-            }
-            ManagementJsonRpcMessage::Notification(
-                ManagementJsonRpcNotification::UpdateAgentStatus(UpdateAgentStatusParams {
-                    slot_aggregated_status_snapshot,
-                }),
-            ) => match &*context.registration.read() {
-                AgentSocketRegistration::AwaitingRegistration => {
-                    error!(
-                        "Rejecting a status update from agent {} before its registration",
-                        context.agent_id
-                    );
-
-                    ContinuationDecision::Stop(WebSocketCloseCause::AgentStatusBeforeRegistration)
-                }
-                AgentSocketRegistration::Registered(agent_controller) => {
-                    match agent_controller.update_from_slot_aggregated_status_snapshot(
-                        slot_aggregated_status_snapshot,
-                    ) {
-                        AgentControllerUpdateResult::NoMeaningfulChanges => {}
-                        AgentControllerUpdateResult::Updated => {
-                            context.agent_controller_pool.signal_update();
-                        }
-                    }
-
-                    ContinuationDecision::Continue
-                }
-            },
-            ManagementJsonRpcMessage::Response(ResponseEnvelope {
-                request_id,
-                response: AgentJsonRpcResponse::ChatTemplateOverride(chat_template_override),
-                ..
-            }) => {
-                forward_agent_response(
-                    &context.agent_response_senders.chat_template_override,
-                    &request_id,
-                    chat_template_override,
-                );
-
-                ContinuationDecision::Continue
-            }
-            ManagementJsonRpcMessage::Response(ResponseEnvelope {
-                request_id,
-                response: AgentJsonRpcResponse::Embedding(embedding_result),
-                ..
-            }) => {
-                forward_agent_response(
-                    &context.agent_response_senders.embedding,
-                    &request_id,
-                    embedding_result,
-                );
-
-                ContinuationDecision::Continue
-            }
-            ManagementJsonRpcMessage::Response(ResponseEnvelope {
-                request_id,
-                response: AgentJsonRpcResponse::GeneratedToken(generated_token_envelope),
-                ..
-            }) => {
-                forward_agent_response(
-                    &context.agent_response_senders.generated_tokens,
-                    &request_id,
-                    generated_token_envelope,
-                );
-
-                ContinuationDecision::Continue
-            }
-            ManagementJsonRpcMessage::Response(ResponseEnvelope {
-                request_id,
-                response: AgentJsonRpcResponse::ModelMetadata(model_metadata),
-                ..
-            }) => {
-                forward_agent_response(
-                    &context.agent_response_senders.model_metadata,
-                    &request_id,
-                    model_metadata,
-                );
-
-                ContinuationDecision::Continue
-            }
-        }
+        websocket_session_controller: WebSocketSessionController<Self::OutgoingMessage>,
+    ) -> impl Future<Output = ContinuationDecision> + Send {
+        ready(context.handle_message(
+            connection_close,
+            deserialized_message,
+            websocket_session_controller,
+        ))
     }
 
-    async fn handle_undeserializable_message(
+    fn handle_undeserializable_message(
         _text: &str,
         _deserialization_error: SerdeJsonError,
         _websocket_session_controller: WebSocketSessionController<Self::OutgoingMessage>,
-    ) -> ContinuationDecision {
-        ContinuationDecision::Stop(WebSocketCloseCause::AgentMessageUndeserializable)
+    ) -> impl Future<Output = ContinuationDecision> + Send {
+        ready(ContinuationDecision::Stop(
+            WebSocketCloseCause::AgentMessageUndeserializable,
+        ))
     }
 
     async fn on_connection_start(
         _connection_close: CancellationToken,
         _context: Arc<Self::Context>,
-        session: &mut Session,
+        outbound_tx: &UnboundedSender<WebSocketOutboundFrame>,
     ) {
-        WebSocketSessionController::new(session.clone())
+        WebSocketSessionController::new(outbound_tx.clone())
             .send_response_safe(AgentJsonRpcMessage::Notification(
                 AgentJsonRpcNotification::Version(VersionParams {
                     version: env!("CARGO_PKG_VERSION").to_owned(),
diff --git a/paddler_balancer/src/management_service/mod.rs b/paddler_balancer/src/management_service/mod.rs
index b2160dc56..b8c2a8c12 100644
--- a/paddler_balancer/src/management_service/mod.rs
+++ b/paddler_balancer/src/management_service/mod.rs
@@ -12,11 +12,9 @@ use async_trait::async_trait;
 use tokio_util::sync::CancellationToken;
 use trzcina::Service;
 
-use paddler_state_database::state_database::StateDatabase;
-
 use crate::agent_controller_pool::AgentControllerPool;
-use crate::agent_response_senders::AgentResponseSenders;
 use crate::balancer_applicable_state_holder::BalancerApplicableStateHolder;
+use crate::balancer_desired_state_store::BalancerDesiredStateStore;
 use crate::buffered_request_manager::BufferedRequestManager;
 use crate::cors_allowed_hosts_with_web_admin_panel::cors_allowed_hosts_with_web_admin_panel;
 use crate::create_cors_middleware::create_cors_middleware;
@@ -41,22 +39,16 @@ use crate::run_http_service_parameters::RunHttpServiceParameters;
 pub struct ManagementService {
     pub agent_controller_pool: Arc<AgentControllerPool>,
     pub balancer_applicable_state_holder: Arc<BalancerApplicableStateHolder>,
+    pub balancer_desired_state_store: Arc<BalancerDesiredStateStore>,
     pub buffered_request_manager: Arc<BufferedRequestManager>,
-    pub agent_response_senders: AgentResponseSenders,
     pub configuration: ManagementServiceConfiguration,
     pub http_listener: HttpListener,
-    pub state_database: Arc<dyn StateDatabase>,
     pub statsd_prefix: String,
     pub web_admin_panel_addr: Option<SocketAddr>,
 }
 
-#[async_trait]
-impl Service for ManagementService {
-    fn name(&self) -> &'static str {
-        "balancer::management_service"
-    }
-
-    async fn run(self: Box<Self>, shutdown: CancellationToken) -> Result<()> {
+impl ManagementService {
+    async fn serve_until_shutdown(self: Box<Self>, shutdown: CancellationToken) -> Result<()> {
         let service_name = self.name();
         let cors_allowed_hosts_arc = cors_allowed_hosts_with_web_admin_panel(
             &self.configuration.cors_allowed_hosts,
@@ -66,10 +58,9 @@ impl Service for ManagementService {
         let app_data = Data::new(AppData {
             agent_controller_pool: self.agent_controller_pool.clone(),
             balancer_applicable_state_holder: self.balancer_applicable_state_holder.clone(),
+            balancer_desired_state_store: self.balancer_desired_state_store.clone(),
             buffered_request_manager: self.buffered_request_manager.clone(),
-            agent_response_senders: self.agent_response_senders.clone(),
             shutdown: shutdown.clone(),
-            state_database: self.state_database.clone(),
             statsd_prefix: self.statsd_prefix.clone(),
         });
 
@@ -101,3 +92,14 @@ impl Service for ManagementService {
         .await
     }
 }
+
+#[async_trait]
+impl Service for ManagementService {
+    fn name(&self) -> &'static str {
+        "balancer::management_service"
+    }
+
+    async fn run(self: Box<Self>, shutdown: CancellationToken) -> Result<()> {
+        self.serve_until_shutdown(shutdown).await
+    }
+}
diff --git a/paddler_balancer/src/outgoing_message_relay_receiver.rs b/paddler_balancer/src/outgoing_message_relay_receiver.rs
new file mode 100644
index 000000000..ff7f07f27
--- /dev/null
+++ b/paddler_balancer/src/outgoing_message_relay_receiver.rs
@@ -0,0 +1,65 @@
+use paddler_messaging::inference_client::message::Message as OutgoingMessage;
+use paddler_messaging::inference_client::response::Response as OutgoingResponse;
+use paddler_messaging::jsonrpc::response_envelope::ResponseEnvelope;
+use paddler_messaging::streamable_result::StreamableResult;
+
+use crate::agent_relay_conclusion::AgentRelayConclusion;
+use crate::controls_session::ControlsSession;
+use crate::receives_agent_relay::ReceivesAgentRelay;
+use crate::respond_with_error::respond_with_error;
+
+pub struct OutgoingMessageRelayReceiver<TControlsSession> {
+    pub request_id: String,
+    pub session_controller: TControlsSession,
+}
+
+impl<TControlsSession> OutgoingMessageRelayReceiver<TControlsSession>
+where
+    TControlsSession: ControlsSession<OutgoingMessage>,
+{
+    async fn send_response(&mut self, generated_by: Option<String>, response: OutgoingResponse) {
+        self.session_controller
+            .send_response_safe(OutgoingMessage::Response(ResponseEnvelope {
+                generated_by,
+                request_id: self.request_id.clone(),
+                response,
+            }))
+            .await;
+    }
+}
+
+impl<TControlsSession, TResponse> ReceivesAgentRelay<TResponse>
+    for OutgoingMessageRelayReceiver<TControlsSession>
+where
+    TControlsSession: ControlsSession<OutgoingMessage>,
+    TResponse: StreamableResult<Final: Send, Partial: Send> + Into<OutgoingResponse>,
+{
+    async fn receive_conclusion(mut self, conclusion: AgentRelayConclusion<TResponse::Final>) {
+        match conclusion {
+            AgentRelayConclusion::AgentConcluded {
+                final_result,
+                generated_by,
+            } => {
+                self.send_response(generated_by, Into::<TResponse>::into(final_result).into())
+                    .await;
+            }
+            AgentRelayConclusion::Failed(relay_failure) => {
+                respond_with_error(
+                    relay_failure.into_jsonrpc_error(),
+                    self.request_id,
+                    &mut self.session_controller,
+                )
+                .await;
+            }
+        }
+    }
+
+    async fn receive_partial(
+        &mut self,
+        generated_by: Option<String>,
+        partial_result: TResponse::Partial,
+    ) {
+        self.send_response(generated_by, Into::<TResponse>::into(partial_result).into())
+            .await;
+    }
+}
diff --git a/paddler_balancer/src/receives_agent_relay.rs b/paddler_balancer/src/receives_agent_relay.rs
new file mode 100644
index 000000000..b68b27a2e
--- /dev/null
+++ b/paddler_balancer/src/receives_agent_relay.rs
@@ -0,0 +1,21 @@
+use std::future::Future;
+
+use paddler_messaging::streamable_result::StreamableResult;
+
+use crate::agent_relay_conclusion::AgentRelayConclusion;
+
+pub trait ReceivesAgentRelay<TResponse>
+where
+    TResponse: StreamableResult,
+{
+    fn receive_conclusion(
+        self,
+        conclusion: AgentRelayConclusion<TResponse::Final>,
+    ) -> impl Future<Output = ()> + Send;
+
+    fn receive_partial(
+        &mut self,
+        generated_by: Option<String>,
+        partial_result: TResponse::Partial,
+    ) -> impl Future<Output = ()> + Send;
+}
diff --git a/paddler_balancer/src/reconciliation_service.rs b/paddler_balancer/src/reconciliation_service.rs
index 048a4f2cf..03f5c4604 100644
--- a/paddler_balancer/src/reconciliation_service.rs
+++ b/paddler_balancer/src/reconciliation_service.rs
@@ -4,57 +4,56 @@ use anyhow::Context as _;
 use anyhow::Result;
 use async_trait::async_trait;
 use tokio::select;
-use tokio::sync::watch;
 use tokio_util::sync::CancellationToken;
 use trzcina::Service;
 
-use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-
 use crate::agent_controller_pool::AgentControllerPool;
-use crate::balancer_applicable_state::BalancerApplicableState;
 use crate::balancer_applicable_state_holder::BalancerApplicableStateHolder;
 
 pub struct ReconciliationService {
     pub agent_controller_pool: Arc<AgentControllerPool>,
     pub balancer_applicable_state_holder: Arc<BalancerApplicableStateHolder>,
-    pub balancer_desired_state_rx: watch::Receiver<BalancerDesiredState>,
 }
 
-#[async_trait]
-impl Service for ReconciliationService {
-    fn name(&self) -> &'static str {
-        "balancer::reconciliation_service"
-    }
-
-    async fn run(self: Box<Self>, shutdown: CancellationToken) -> Result<()> {
+impl ReconciliationService {
+    async fn apply_desired_states_until_shutdown(
+        self: Box<Self>,
+        shutdown: CancellationToken,
+    ) -> Result<()> {
         let Self {
             agent_controller_pool,
             balancer_applicable_state_holder,
-            mut balancer_desired_state_rx,
         } = *self;
+        let mut balancer_applicable_state_rx = balancer_applicable_state_holder.subscribe();
 
         loop {
             select! {
                 biased;
                 () = shutdown.cancelled() => break Ok(()),
-                changed = balancer_desired_state_rx.changed() => {
-                    changed.context("the state database stopped announcing desired states")?;
+                changed = balancer_applicable_state_rx.changed() => {
+                    changed.context("the desired state store stopped announcing applicable states")?;
 
-                    let balancer_applicable_state = BalancerApplicableState::from(
-                        balancer_desired_state_rx.borrow_and_update().clone(),
-                    );
+                    let balancer_applicable_state =
+                        balancer_applicable_state_rx.borrow_and_update().clone();
 
-                    let agent_desired_state = balancer_applicable_state.agent_desired_state.clone();
-
-                    balancer_applicable_state_holder
-                        .set_balancer_applicable_state(balancer_applicable_state);
-                    agent_controller_pool.set_desired_state(&agent_desired_state);
+                    agent_controller_pool.set_desired_state(&balancer_applicable_state);
                 }
             }
         }
     }
 }
 
+#[async_trait]
+impl Service for ReconciliationService {
+    fn name(&self) -> &'static str {
+        "balancer::reconciliation_service"
+    }
+
+    async fn run(self: Box<Self>, shutdown: CancellationToken) -> Result<()> {
+        self.apply_desired_states_until_shutdown(shutdown).await
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use std::sync::Arc;
@@ -70,28 +69,42 @@ mod tests {
     use crate::balancer_applicable_state::BalancerApplicableState;
     use crate::balancer_applicable_state_holder::BalancerApplicableStateHolder;
 
-    const RUNS_TO_EXERCISE_EVERY_BRANCH_ORDER: usize = 64;
+    fn reconciliation_service_of_a_closed_store() -> ReconciliationService {
+        let (balancer_applicable_state_tx, balancer_applicable_state_rx) =
+            watch::channel(Arc::new(BalancerApplicableState::initial(
+                BalancerDesiredState::default(),
+            )));
+
+        drop(balancer_applicable_state_tx);
+
+        ReconciliationService {
+            agent_controller_pool: Arc::new(AgentControllerPool::default()),
+            balancer_applicable_state_holder: Arc::new(BalancerApplicableStateHolder::new(
+                balancer_applicable_state_rx,
+            )),
+        }
+    }
 
     #[tokio::test]
-    async fn stops_cleanly_when_its_state_database_closes_during_shutdown() {
-        for _ in 0..RUNS_TO_EXERCISE_EVERY_BRANCH_ORDER {
-            let (balancer_desired_state_tx, balancer_desired_state_rx) =
-                watch::channel(BalancerDesiredState::default());
-            let shutdown = CancellationToken::new();
-
-            shutdown.cancel();
-            drop(balancer_desired_state_tx);
-
-            Box::new(ReconciliationService {
-                agent_controller_pool: Arc::new(AgentControllerPool::default()),
-                balancer_applicable_state_holder: Arc::new(BalancerApplicableStateHolder::new(
-                    BalancerApplicableState::from(BalancerDesiredState::default()),
-                )),
-                balancer_desired_state_rx,
-            })
+    async fn stops_cleanly_when_its_desired_state_store_closes_during_shutdown() {
+        let shutdown = CancellationToken::new();
+
+        shutdown.cancel();
+
+        Box::new(reconciliation_service_of_a_closed_store())
             .run(shutdown)
             .await
             .expect("a shutdown must stop the reconciliation service cleanly");
-        }
+    }
+
+    #[tokio::test]
+    async fn fails_once_its_desired_state_store_stops_announcing_applicable_states() {
+        assert_eq!(
+            Box::new(reconciliation_service_of_a_closed_store())
+                .run(CancellationToken::new())
+                .await
+                .map_err(|reconciliation_error| reconciliation_error.to_string()),
+            Err("the desired state store stopped announcing applicable states".to_owned())
+        );
     }
 }
diff --git a/paddler_balancer/src/relay_failure.rs b/paddler_balancer/src/relay_failure.rs
new file mode 100644
index 000000000..df4b41c99
--- /dev/null
+++ b/paddler_balancer/src/relay_failure.rs
@@ -0,0 +1,96 @@
+use std::time::Duration;
+
+use paddler_messaging::jsonrpc::error::Error as JsonRpcError;
+
+#[derive(Debug, Eq, PartialEq)]
+pub enum RelayFailure {
+    AgentDisconnected,
+    BalancerShuttingDown,
+    BufferOverflow,
+    BufferedRequestTimedOut,
+    ItemTimedOut { inference_item_timeout: Duration },
+}
+
+impl RelayFailure {
+    #[must_use]
+    pub fn description(&self) -> String {
+        match self {
+            Self::AgentDisconnected => "Agent controller connection closed".to_owned(),
+            Self::BalancerShuttingDown => "balancer is shutting down".to_owned(),
+            Self::BufferOverflow => "Buffered requests overflow".to_owned(),
+            Self::BufferedRequestTimedOut => "Waiting for available slot timed out".to_owned(),
+            Self::ItemTimedOut {
+                inference_item_timeout,
+            } => {
+                let timeout_ms = inference_item_timeout.as_millis();
+
+                format!(
+                    "Inference timed out after {timeout_ms}ms waiting for next token. \
+                    Increase --inference-item-timeout if the prompt requires longer processing."
+                )
+            }
+        }
+    }
+
+    #[must_use]
+    pub fn into_jsonrpc_error(self) -> JsonRpcError {
+        JsonRpcError {
+            code: match self {
+                Self::AgentDisconnected => 502,
+                Self::BalancerShuttingDown | Self::BufferOverflow => 503,
+                Self::BufferedRequestTimedOut | Self::ItemTimedOut { .. } => 504,
+            },
+            description: self.description(),
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::time::Duration;
+
+    use paddler_messaging::jsonrpc::error::Error as JsonRpcError;
+
+    use super::RelayFailure;
+
+    #[test]
+    fn every_relay_failure_reaches_the_client_as_its_jsonrpc_error() {
+        for (relay_failure, expected_code, expected_description) in [
+            (
+                RelayFailure::AgentDisconnected,
+                502,
+                "Agent controller connection closed",
+            ),
+            (
+                RelayFailure::BalancerShuttingDown,
+                503,
+                "balancer is shutting down",
+            ),
+            (
+                RelayFailure::BufferOverflow,
+                503,
+                "Buffered requests overflow",
+            ),
+            (
+                RelayFailure::BufferedRequestTimedOut,
+                504,
+                "Waiting for available slot timed out",
+            ),
+            (
+                RelayFailure::ItemTimedOut {
+                    inference_item_timeout: Duration::from_millis(250),
+                },
+                504,
+                "Inference timed out after 250ms waiting for next token. Increase --inference-item-timeout if the prompt requires longer processing.",
+            ),
+        ] {
+            assert_eq!(
+                relay_failure.into_jsonrpc_error(),
+                JsonRpcError {
+                    code: expected_code,
+                    description: expected_description.to_owned(),
+                }
+            );
+        }
+    }
+}
diff --git a/paddler_balancer/src/request_from_agent.rs b/paddler_balancer/src/request_from_agent.rs
deleted file mode 100644
index a0fd2013e..000000000
--- a/paddler_balancer/src/request_from_agent.rs
+++ /dev/null
@@ -1,139 +0,0 @@
-use std::fmt::Debug;
-use std::sync::Arc;
-
-use log::debug;
-use log::error;
-use log::warn;
-use tokio::select;
-use tokio_util::sync::CancellationToken;
-
-use paddler_messaging::inference_client::message::Message as OutgoingMessage;
-use paddler_messaging::inference_client::response::Response as OutgoingResponse;
-use paddler_messaging::jsonrpc::error::Error as JsonRpcError;
-use paddler_messaging::streamable_result::StreamableResult;
-
-use crate::agent_streaming_request::AgentStreamingRequest;
-use crate::balancer_shutdown_error::balancer_shutdown_error;
-use crate::buffered_request_agent_wait_result::BufferedRequestAgentWaitResult;
-use crate::buffered_request_manager::BufferedRequestManager;
-use crate::controls_session::ControlsSession;
-use crate::dispatched_agent::DispatchedAgent;
-use crate::inference_service::configuration::Configuration as InferenceServiceConfiguration;
-use crate::request_from_dispatched_agent::request_from_dispatched_agent;
-use crate::respond_with_error::respond_with_error;
-
-async fn wait_for_agent_controller<TControlsSession>(
-    buffered_request_manager: Arc<BufferedRequestManager>,
-    connection_close: CancellationToken,
-    request_id: String,
-    session_controller: &mut TControlsSession,
-    shutdown: CancellationToken,
-) -> Option<DispatchedAgent>
-where
-    TControlsSession: ControlsSession<OutgoingMessage>,
-{
-    let buffered_request_manager = buffered_request_manager.clone();
-
-    select! {
-        biased;
-
-        () = shutdown.cancelled() => {
-            respond_with_error(
-                balancer_shutdown_error(),
-                request_id.clone(),
-                session_controller,
-            ).await;
-
-            None
-        },
-        () = connection_close.cancelled() => {
-            debug!("Connection close signal received, stopping GenerateTokens loop.");
-
-            None
-        },
-        buffered_request_agent_wait_result = buffered_request_manager.wait_for_available_agent() => {
-            match buffered_request_agent_wait_result {
-                Ok(BufferedRequestAgentWaitResult::Found(dispatched_agent)) => Some(dispatched_agent),
-                Ok(BufferedRequestAgentWaitResult::BufferOverflow) => {
-                    warn!("Too many buffered requests, dropping request: {request_id:?}");
-
-                    respond_with_error(
-                        JsonRpcError {
-                            code: 503,
-                            description: "Buffered requests overflow".to_owned(),
-                        },
-                        request_id.clone(),
-                        session_controller,
-                    ).await;
-
-                    None
-                }
-                Ok(BufferedRequestAgentWaitResult::Timeout) => {
-                    warn!("Buffered request {request_id:?} timed out waiting for an available slot");
-
-                    respond_with_error(
-                        JsonRpcError {
-                            code: 504,
-                            description: "Waiting for available slot timed out".to_owned(),
-                        },
-                        request_id.clone(),
-                        session_controller,
-                    ).await;
-
-                    None
-                }
-                Err(err) => {
-                    error!("Error while waiting for available agent controller for GenerateTokens request: {err}");
-
-                    respond_with_error(
-                        JsonRpcError {
-                            code: 500,
-                            description: "Internal server error".to_owned(),
-                        },
-                        request_id.clone(),
-                        session_controller,
-                    ).await;
-
-                    None
-                }
-            }
-        }
-    }
-}
-
-pub async fn request_from_agent<TControlsSession, TParams>(
-    buffered_request_manager: Arc<BufferedRequestManager>,
-    connection_close: CancellationToken,
-    inference_service_configuration: InferenceServiceConfiguration,
-    params: TParams,
-    request_id: String,
-    mut session_controller: TControlsSession,
-    shutdown: CancellationToken,
-) where
-    TControlsSession: ControlsSession<OutgoingMessage>,
-    TParams: AgentStreamingRequest + Debug + Send,
-    TParams::Response: Debug + Into<OutgoingResponse> + StreamableResult,
-{
-    let Some(dispatched_agent) = wait_for_agent_controller(
-        buffered_request_manager.clone(),
-        connection_close.clone(),
-        request_id.clone(),
-        &mut session_controller,
-        shutdown.clone(),
-    )
-    .await
-    else {
-        return;
-    };
-
-    request_from_dispatched_agent(
-        dispatched_agent,
-        connection_close,
-        inference_service_configuration,
-        params,
-        request_id,
-        session_controller,
-        shutdown,
-    )
-    .await;
-}
diff --git a/paddler_balancer/src/request_from_dispatched_agent.rs b/paddler_balancer/src/request_from_dispatched_agent.rs
deleted file mode 100644
index 54ca540eb..000000000
--- a/paddler_balancer/src/request_from_dispatched_agent.rs
+++ /dev/null
@@ -1,64 +0,0 @@
-use std::fmt::Debug;
-
-use log::error;
-use nanoid::nanoid;
-use tokio_util::sync::CancellationToken;
-
-use paddler_messaging::inference_client::message::Message as OutgoingMessage;
-use paddler_messaging::inference_client::response::Response as OutgoingResponse;
-use paddler_messaging::jsonrpc::error::Error as JsonRpcError;
-use paddler_messaging::streamable_result::StreamableResult;
-
-use crate::agent_streaming_request::AgentStreamingRequest;
-use crate::controls_session::ControlsSession;
-use crate::dispatched_agent::DispatchedAgent;
-use crate::forward_responses_stream::forward_responses_stream;
-use crate::inference_service::configuration::Configuration as InferenceServiceConfiguration;
-use crate::respond_with_error::respond_with_error;
-
-pub async fn request_from_dispatched_agent<TControlsSession, TParams>(
-    dispatched_agent: DispatchedAgent,
-    connection_close: CancellationToken,
-    inference_service_configuration: InferenceServiceConfiguration,
-    params: TParams,
-    request_id: String,
-    mut session_controller: TControlsSession,
-    shutdown: CancellationToken,
-) where
-    TControlsSession: ControlsSession<OutgoingMessage>,
-    TParams: AgentStreamingRequest + Debug + Send,
-    TParams::Response: Debug + Into<OutgoingResponse> + StreamableResult,
-{
-    let receive_response_controller = match dispatched_agent
-        .agent_controller
-        .stream_responses_to(nanoid!(), params)
-    {
-        Ok(receive_response_controller) => receive_response_controller,
-        Err(err) => {
-            error!("Failed to handle request {request_id:?}: {err}");
-
-            respond_with_error(
-                JsonRpcError {
-                    code: 500,
-                    description: "Failed to generate response".to_owned(),
-                },
-                request_id.clone(),
-                &mut session_controller,
-            )
-            .await;
-
-            return;
-        }
-    };
-
-    forward_responses_stream(
-        connection_close,
-        dispatched_agent,
-        inference_service_configuration,
-        receive_response_controller,
-        request_id,
-        session_controller,
-        shutdown,
-    )
-    .await;
-}
diff --git a/paddler_balancer/src/requirements_waiting_ahead.rs b/paddler_balancer/src/requirements_waiting_ahead.rs
new file mode 100644
index 000000000..1121e9aaa
--- /dev/null
+++ b/paddler_balancer/src/requirements_waiting_ahead.rs
@@ -0,0 +1,19 @@
+use paddler_messaging::agent_runtime_status::AgentRuntimeStatus;
+
+use crate::dispatch_requirement::DispatchRequirement;
+
+pub struct RequirementsWaitingAhead {
+    pub dispatch_requirements: Vec<DispatchRequirement>,
+}
+
+impl RequirementsWaitingAhead {
+    #[must_use]
+    pub fn slots_held_on(&self, agent_runtime_status: &AgentRuntimeStatus) -> u64 {
+        self.dispatch_requirements
+            .iter()
+            .map(|dispatch_requirement| {
+                dispatch_requirement.slots_held_while_waiting(agent_runtime_status)
+            })
+            .sum()
+    }
+}
diff --git a/paddler_balancer/src/respond_with_agent_response.rs b/paddler_balancer/src/respond_with_agent_response.rs
index 452f9a2c9..6deb6c4f9 100644
--- a/paddler_balancer/src/respond_with_agent_response.rs
+++ b/paddler_balancer/src/respond_with_agent_response.rs
@@ -8,7 +8,6 @@ use tokio::time::sleep;
 
 use crate::agent_controller::AgentController;
 use crate::agent_controller_pool::AgentControllerPool;
-use crate::agent_relay_error::AgentRelayError;
 use crate::agent_response_receiver::AgentResponseReceiver;
 
 const AGENT_RESPONSE_TIMEOUT: Duration = Duration::from_secs(3);
@@ -20,8 +19,7 @@ pub async fn respond_with_agent_response<TResponse, TAskAgent>(
 ) -> HttpResponse
 where
     TResponse: Serialize,
-    TAskAgent:
-        FnOnce(Arc<AgentController>) -> Result<AgentResponseReceiver<TResponse>, AgentRelayError>,
+    TAskAgent: FnOnce(Arc<AgentController>) -> AgentResponseReceiver<TResponse>,
 {
     let Some(agent_controller) = agent_controller_pool.get_agent_controller(agent_id) else {
         return HttpResponse::NotFound().finish();
@@ -29,14 +27,13 @@ where
 
     let connection_close = agent_controller.connection_close.clone();
 
-    match ask_agent(agent_controller) {
-        Ok(mut agent_response_receiver) => select! {
-            () = connection_close.cancelled() => HttpResponse::BadGateway().finish(),
-            () = sleep(AGENT_RESPONSE_TIMEOUT) => HttpResponse::GatewayTimeout().finish(),
-            Some(agent_response) = agent_response_receiver.response_rx.recv() => {
-                HttpResponse::Ok().json(agent_response)
-            }
-        },
-        Err(err) => HttpResponse::InternalServerError().body(format!("{err}")),
+    let mut agent_response_receiver = ask_agent(agent_controller);
+
+    select! {
+        () = connection_close.cancelled() => HttpResponse::BadGateway().finish(),
+        () = sleep(AGENT_RESPONSE_TIMEOUT) => HttpResponse::GatewayTimeout().finish(),
+        Some(agent_response) = agent_response_receiver.response_rx.recv() => {
+            HttpResponse::Ok().json(agent_response)
+        }
     }
 }
diff --git a/paddler_balancer/src/respond_with_error.rs b/paddler_balancer/src/respond_with_error.rs
index 5a997cd74..f811db9b3 100644
--- a/paddler_balancer/src/respond_with_error.rs
+++ b/paddler_balancer/src/respond_with_error.rs
@@ -1,5 +1,3 @@
-use log::error;
-
 use paddler_messaging::inference_client::message::Message as OutgoingMessage;
 use paddler_messaging::jsonrpc::error::Error as JsonRpcError;
 use paddler_messaging::jsonrpc::error_envelope::ErrorEnvelope;
@@ -14,12 +12,6 @@ pub async fn respond_with_error<TControlsSession>(
     TControlsSession: ControlsSession<OutgoingMessage>,
 {
     session_controller
-        .send_response(OutgoingMessage::Error(ErrorEnvelope {
-            request_id: request_id.clone(),
-            error,
-        }))
-        .await
-        .unwrap_or_else(|err| {
-            error!("Failed to send response for request {request_id:?}: {err}");
-        });
+        .send_response_safe(OutgoingMessage::Error(ErrorEnvelope { request_id, error }))
+        .await;
 }
diff --git a/paddler_balancer/src/response/view_from_http_response_builder.rs b/paddler_balancer/src/response/view_from_http_response_builder.rs
index 743a8469a..5784d34a7 100644
--- a/paddler_balancer/src/response/view_from_http_response_builder.rs
+++ b/paddler_balancer/src/response/view_from_http_response_builder.rs
@@ -51,9 +51,9 @@ mod tests {
     }
 
     impl FastWritable for RenderingTemplate {
-        fn write_into<TWriter: fmt::Write + ?Sized>(
+        fn write_into(
             &self,
-            destination: &mut TWriter,
+            destination: &mut dyn fmt::Write,
             values: &dyn Values,
         ) -> AskamaResult<()> {
             self.render_into_with_values(destination, values)
@@ -63,9 +63,9 @@ mod tests {
     impl Template for RenderingTemplate {
         const SIZE_HINT: usize = 0;
 
-        fn render_into_with_values<TWriter: fmt::Write + ?Sized>(
+        fn render_into_with_values(
             &self,
-            writer: &mut TWriter,
+            writer: &mut dyn fmt::Write,
             _values: &dyn Values,
         ) -> AskamaResult<()> {
             writer.write_str("<p>rendered</p>")?;
@@ -84,9 +84,9 @@ mod tests {
     }
 
     impl FastWritable for FailingTemplate {
-        fn write_into<TWriter: fmt::Write + ?Sized>(
+        fn write_into(
             &self,
-            destination: &mut TWriter,
+            destination: &mut dyn fmt::Write,
             values: &dyn Values,
         ) -> AskamaResult<()> {
             self.render_into_with_values(destination, values)
@@ -96,9 +96,9 @@ mod tests {
     impl Template for FailingTemplate {
         const SIZE_HINT: usize = 0;
 
-        fn render_into_with_values<TWriter: fmt::Write + ?Sized>(
+        fn render_into_with_values(
             &self,
-            _writer: &mut TWriter,
+            _writer: &mut dyn fmt::Write,
             _values: &dyn Values,
         ) -> AskamaResult<()> {
             Err(AskamaError::ValueMissing)
diff --git a/paddler_balancer/src/session_keep_alive.rs b/paddler_balancer/src/session_keep_alive.rs
deleted file mode 100644
index fd28ee0b2..000000000
--- a/paddler_balancer/src/session_keep_alive.rs
+++ /dev/null
@@ -1,30 +0,0 @@
-use actix_ws::Session;
-use tokio::time::Duration;
-use tokio::time::MissedTickBehavior;
-use tokio::time::interval;
-use tokio_util::sync::CancellationToken;
-
-const PING_INTERVAL: Duration = Duration::from_secs(3);
-
-pub struct SessionKeepAlive {
-    pub connection_close: CancellationToken,
-    pub session: Session,
-}
-
-impl SessionKeepAlive {
-    pub async fn run(mut self) {
-        let mut ping_ticker = interval(PING_INTERVAL);
-
-        ping_ticker.set_missed_tick_behavior(MissedTickBehavior::Delay);
-
-        while self
-            .connection_close
-            .run_until_cancelled(ping_ticker.tick())
-            .await
-            .is_some()
-            && self.session.ping(b"").await.is_ok()
-        {}
-
-        self.connection_close.cancel();
-    }
-}
diff --git a/paddler_balancer/src/slot_demand.rs b/paddler_balancer/src/slot_demand.rs
new file mode 100644
index 000000000..7bd96b26a
--- /dev/null
+++ b/paddler_balancer/src/slot_demand.rs
@@ -0,0 +1,60 @@
+use std::num::NonZeroU16;
+
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+pub enum SlotDemand {
+    Exactly(NonZeroU16),
+    UpTo(u16),
+}
+
+impl SlotDemand {
+    #[must_use]
+    pub const fn minimum_slots(self) -> u16 {
+        match self {
+            Self::Exactly(slots) => slots.get(),
+            Self::UpTo(_slots) => 1,
+        }
+    }
+
+    #[must_use]
+    pub fn reservable_from(self, free_slots: u64) -> Option<NonZeroU16> {
+        match self {
+            Self::Exactly(slots) => (free_slots >= u64::from(slots.get())).then_some(slots),
+            Self::UpTo(slots) => NonZeroU16::new(
+                u16::try_from(free_slots).map_or(slots, |free_slots| free_slots.min(slots)),
+            ),
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::num::NonZeroU16;
+
+    use super::SlotDemand;
+
+    const TWO_SLOTS: NonZeroU16 = NonZeroU16::new(2).unwrap();
+
+    #[test]
+    fn an_exact_demand_reserves_its_slots_when_they_are_free() {
+        assert_eq!(
+            SlotDemand::Exactly(TWO_SLOTS).reservable_from(3),
+            Some(TWO_SLOTS)
+        );
+    }
+
+    #[test]
+    fn an_exact_demand_reserves_nothing_while_fewer_slots_are_free() {
+        assert_eq!(SlotDemand::Exactly(TWO_SLOTS).reservable_from(1), None);
+    }
+
+    #[test]
+    fn a_flexible_demand_reserves_every_free_slot_it_can_use() {
+        assert_eq!(SlotDemand::UpTo(5).reservable_from(2), Some(TWO_SLOTS));
+        assert_eq!(SlotDemand::UpTo(2).reservable_from(5), Some(TWO_SLOTS));
+    }
+
+    #[test]
+    fn a_flexible_demand_reserves_nothing_while_no_slot_is_free() {
+        assert_eq!(SlotDemand::UpTo(5).reservable_from(0), None);
+    }
+}
diff --git a/paddler_balancer/src/snapshots_stream.rs b/paddler_balancer/src/snapshots_stream.rs
index c56f4cb68..e72d026b0 100644
--- a/paddler_balancer/src/snapshots_stream.rs
+++ b/paddler_balancer/src/snapshots_stream.rs
@@ -1,13 +1,19 @@
 use std::sync::Arc;
 
-use async_stream::stream;
 use futures::Stream;
+use futures::stream::unfold;
 use tokio::select;
+use tokio::sync::watch;
 use tokio_util::sync::CancellationToken;
 
 use paddler_messaging::produces_snapshot::ProducesSnapshot;
 use paddler_messaging::subscribes_to_updates::SubscribesToUpdates;
 
+enum SnapshotsPhase {
+    Subscribing,
+    AwaitingUpdate(watch::Receiver<()>),
+}
+
 pub fn snapshots_stream<TProducer>(
     producer: Arc<TProducer>,
     shutdown: CancellationToken,
@@ -16,18 +22,27 @@ where
     TProducer: ProducesSnapshot + SubscribesToUpdates + Send + Sync + 'static,
     TProducer::Snapshot: Send + 'static,
 {
-    stream! {
-        let mut update_rx = producer.subscribe_to_updates();
-
-        loop {
-            yield producer.make_snapshot();
-
-            select! {
-                () = shutdown.cancelled() => break,
-                Ok(()) = update_rx.changed() => {}
-            }
+    unfold(SnapshotsPhase::Subscribing, move |phase| {
+        let producer = producer.clone();
+        let shutdown = shutdown.clone();
+
+        async move {
+            let update_rx = match phase {
+                SnapshotsPhase::Subscribing => producer.subscribe_to_updates(),
+                SnapshotsPhase::AwaitingUpdate(mut update_rx) => {
+                    select! {
+                        () = shutdown.cancelled() => return None,
+                        Ok(()) = update_rx.changed() => update_rx,
+                    }
+                }
+            };
+
+            Some((
+                producer.make_snapshot(),
+                SnapshotsPhase::AwaitingUpdate(update_rx),
+            ))
         }
-    }
+    })
 }
 
 #[cfg(test)]
diff --git a/paddler_balancer/src/statsd_service/mod.rs b/paddler_balancer/src/statsd_service/mod.rs
index d74b890fe..fdfd4a0bf 100644
--- a/paddler_balancer/src/statsd_service/mod.rs
+++ b/paddler_balancer/src/statsd_service/mod.rs
@@ -1,4 +1,5 @@
 pub mod configuration;
+pub mod statsd_udp_sink;
 
 use std::net::UdpSocket;
 use std::sync::Arc;
@@ -6,7 +7,6 @@ use std::sync::Arc;
 use anyhow::Result;
 use async_trait::async_trait;
 use cadence::StatsdClient;
-use cadence::UdpMetricSink;
 use log::error;
 use tokio::select;
 use tokio::time::MissedTickBehavior;
@@ -18,6 +18,7 @@ use crate::agent_controller_pool::AgentControllerPool;
 use crate::balancer_metrics::BalancerMetrics;
 use crate::buffered_request_manager::BufferedRequestManager;
 use crate::statsd_service::configuration::Configuration as StatsdServiceConfiguration;
+use crate::statsd_service::statsd_udp_sink::StatsdUdpSink;
 
 pub struct StatsdService {
     pub agent_controller_pool: Arc<AgentControllerPool>,
@@ -26,19 +27,15 @@ pub struct StatsdService {
     pub statsd_prefix: String,
 }
 
-#[async_trait]
-impl Service for StatsdService {
-    fn name(&self) -> &'static str {
-        "balancer::statsd_service"
-    }
-
-    async fn run(self: Box<Self>, shutdown: CancellationToken) -> Result<()> {
-        let statsd_sink_socket = UdpSocket::bind("0.0.0.0:0")?;
-        let statsd_sink = UdpMetricSink::from(
-            self.configuration.statsd_addr.socket_addr,
-            statsd_sink_socket,
-        )?;
-
+impl StatsdService {
+    async fn report_metrics_until_shutdown(
+        self: Box<Self>,
+        shutdown: CancellationToken,
+    ) -> Result<()> {
+        let statsd_sink = StatsdUdpSink {
+            socket: UdpSocket::bind("0.0.0.0:0")?,
+            statsd_addr: self.configuration.statsd_addr.socket_addr,
+        };
         let client = StatsdClient::builder(&self.statsd_prefix, statsd_sink).build();
 
         let mut ticker = interval(self.configuration.statsd_reporting_interval);
@@ -62,3 +59,14 @@ impl Service for StatsdService {
         }
     }
 }
+
+#[async_trait]
+impl Service for StatsdService {
+    fn name(&self) -> &'static str {
+        "balancer::statsd_service"
+    }
+
+    async fn run(self: Box<Self>, shutdown: CancellationToken) -> Result<()> {
+        self.report_metrics_until_shutdown(shutdown).await
+    }
+}
diff --git a/paddler_balancer/src/statsd_service/statsd_udp_sink.rs b/paddler_balancer/src/statsd_service/statsd_udp_sink.rs
new file mode 100644
index 000000000..62f624457
--- /dev/null
+++ b/paddler_balancer/src/statsd_service/statsd_udp_sink.rs
@@ -0,0 +1,16 @@
+use std::io;
+use std::net::SocketAddr;
+use std::net::UdpSocket;
+
+use cadence::MetricSink;
+
+pub struct StatsdUdpSink {
+    pub socket: UdpSocket,
+    pub statsd_addr: SocketAddr,
+}
+
+impl MetricSink for StatsdUdpSink {
+    fn emit(&self, metric: &str) -> io::Result<usize> {
+        self.socket.send_to(metric.as_bytes(), self.statsd_addr)
+    }
+}
diff --git a/paddler_balancer/src/unbounded_stream_from_agent.rs b/paddler_balancer/src/unbounded_stream_from_agent.rs
index 407ab38b3..c9419c053 100644
--- a/paddler_balancer/src/unbounded_stream_from_agent.rs
+++ b/paddler_balancer/src/unbounded_stream_from_agent.rs
@@ -1,5 +1,3 @@
-use std::fmt::Debug;
-
 use actix_web::rt;
 use futures_util::Stream;
 use nanoid::nanoid;
@@ -7,14 +5,15 @@ use tokio::sync::mpsc;
 use tokio_stream::wrappers::UnboundedReceiverStream;
 use tokio_util::sync::CancellationToken;
 
-use paddler_messaging::inference_client::response::Response as OutgoingResponse;
-use paddler_messaging::streamable_result::StreamableResult;
-
+use crate::agent_relay::AgentRelay;
 use crate::agent_streaming_request::AgentStreamingRequest;
+use crate::awaiting_agent_relay::AwaitingAgentRelay;
 use crate::cancellation_token_stream_guard::CancellationTokenStreamGuard;
 use crate::chunk_forwarding_session_controller::ChunkForwardingSessionController;
 use crate::chunk_forwarding_session_controller::transforms_outgoing_message::TransformsOutgoingMessage;
-use crate::request_from_agent::request_from_agent;
+use crate::dispatchable_request::DispatchableRequest;
+use crate::forward_agent_relay::forward_agent_relay;
+use crate::outgoing_message_relay_receiver::OutgoingMessageRelayReceiver;
 use crate::unbounded_stream_from_agent_params::UnboundedStreamFromAgentParams;
 
 pub fn unbounded_stream_from_agent<TParams, TTransformsOutgoingMessage>(
@@ -27,32 +26,25 @@ pub fn unbounded_stream_from_agent<TParams, TTransformsOutgoingMessage>(
     }: UnboundedStreamFromAgentParams<TParams, TTransformsOutgoingMessage>,
 ) -> impl Stream<Item = TTransformsOutgoingMessage::Output>
 where
-    TParams: AgentStreamingRequest + Debug + Send + 'static,
-    TParams::Response: Debug + Into<OutgoingResponse> + StreamableResult,
-    TTransformsOutgoingMessage: Clone + TransformsOutgoingMessage + Send + Sync + 'static,
+    TParams: AgentStreamingRequest + 'static,
+    TTransformsOutgoingMessage: TransformsOutgoingMessage + Send + Sync + 'static,
 {
-    let request_id: String = nanoid!();
     let connection_close = CancellationToken::new();
     let (chunk_tx, chunk_rx) = mpsc::unbounded_channel();
 
-    rt::spawn({
-        let connection_close = connection_close.clone();
-
-        async move {
-            let session_controller = ChunkForwardingSessionController::new(chunk_tx, transformer);
-
-            request_from_agent(
-                buffered_request_manager,
-                connection_close,
-                inference_service_configuration,
-                request_params,
-                request_id,
-                session_controller,
-                shutdown,
-            )
-            .await;
-        }
-    });
+    rt::spawn(forward_agent_relay(
+        AgentRelay::awaiting(AwaitingAgentRelay {
+            buffered_request_manager,
+            dispatchable_request: DispatchableRequest::from(request_params),
+            inference_item_timeout: inference_service_configuration.inference_item_timeout,
+            shutdown,
+        }),
+        connection_close.clone(),
+        OutgoingMessageRelayReceiver {
+            request_id: nanoid!(),
+            session_controller: ChunkForwardingSessionController::new(chunk_tx, transformer),
+        },
+    ));
 
     CancellationTokenStreamGuard::new(connection_close, UnboundedReceiverStream::new(chunk_rx))
 }
diff --git a/paddler_balancer/src/waiting_line_place.rs b/paddler_balancer/src/waiting_line_place.rs
new file mode 100644
index 000000000..51bc83b34
--- /dev/null
+++ b/paddler_balancer/src/waiting_line_place.rs
@@ -0,0 +1,24 @@
+use crate::buffered_request_count_guard::BufferedRequestCountGuard;
+use crate::dispatch_position::DispatchPosition;
+use crate::dispatch_waiting_line::DispatchWaitingLine;
+
+pub struct WaitingLinePlace<'line> {
+    pub buffered_request_count_guard: BufferedRequestCountGuard,
+    pub dispatch_waiting_line: &'line DispatchWaitingLine,
+    pub ticket: u64,
+}
+
+impl WaitingLinePlace<'_> {
+    #[must_use]
+    pub const fn dispatch_position(&self) -> DispatchPosition {
+        DispatchPosition::Waiting {
+            ticket: self.ticket,
+        }
+    }
+}
+
+impl Drop for WaitingLinePlace<'_> {
+    fn drop(&mut self) {
+        self.dispatch_waiting_line.leave(self.ticket);
+    }
+}
diff --git a/paddler_balancer/src/web_admin_panel_service/app_data.rs b/paddler_balancer/src/web_admin_panel_service/app_data.rs
index e87ebeb02..2b6f52f95 100644
--- a/paddler_balancer/src/web_admin_panel_service/app_data.rs
+++ b/paddler_balancer/src/web_admin_panel_service/app_data.rs
@@ -1,10 +1,5 @@
-use std::sync::Arc;
-
-use esbuild_metafile::EsbuildMetaFile;
-
 use crate::web_admin_panel_service::template_data::TemplateData;
 
 pub struct AppData {
-    pub esbuild_metafile: Arc<EsbuildMetaFile>,
     pub template_data: TemplateData,
 }
diff --git a/paddler_balancer/src/web_admin_panel_service/dashboard_assets_html.rs b/paddler_balancer/src/web_admin_panel_service/dashboard_assets_html.rs
new file mode 100644
index 000000000..bea5291bd
--- /dev/null
+++ b/paddler_balancer/src/web_admin_panel_service/dashboard_assets_html.rs
@@ -0,0 +1 @@
+pub const DASHBOARD_ASSETS_HTML: &str = include_str!(env!("PADDLER_DASHBOARD_ASSETS_HTML"));
diff --git a/paddler_balancer/src/web_admin_panel_service/esbuild_meta_contents.rs b/paddler_balancer/src/web_admin_panel_service/esbuild_meta_contents.rs
deleted file mode 100644
index 6ff94b662..000000000
--- a/paddler_balancer/src/web_admin_panel_service/esbuild_meta_contents.rs
+++ /dev/null
@@ -1 +0,0 @@
-pub const ESBUILD_META_CONTENTS: &str = include_str!("../../../esbuild-meta.json");
diff --git a/paddler_balancer/src/web_admin_panel_service/http_route/home.rs b/paddler_balancer/src/web_admin_panel_service/http_route/home.rs
index 2e139ca7a..10ef21547 100644
--- a/paddler_balancer/src/web_admin_panel_service/http_route/home.rs
+++ b/paddler_balancer/src/web_admin_panel_service/http_route/home.rs
@@ -2,25 +2,24 @@ use actix_web::Responder;
 use actix_web::web;
 use actix_web::web::get;
 use askama::Template;
-use esbuild_metafile::HttpPreloader;
-use esbuild_metafile::filters;
 
 use crate::response::view::view;
 use crate::web_admin_panel_service::app_data::AppData;
+use crate::web_admin_panel_service::dashboard_assets_html::DASHBOARD_ASSETS_HTML;
 use crate::web_admin_panel_service::template_data::TemplateData;
 
 async fn respond(app_data: web::Data<AppData>) -> impl Responder {
     view(WebAdminPanelTemplate {
-        preloads: HttpPreloader::new(app_data.esbuild_metafile.clone()),
+        dashboard_assets_html: DASHBOARD_ASSETS_HTML,
         template_data: &app_data.template_data,
     })
 }
 
 #[derive(Template)]
 #[template(path = "web_admin_panel.html")]
-struct WebAdminPanelTemplate<'template_data> {
-    preloads: HttpPreloader,
-    template_data: &'template_data TemplateData,
+struct WebAdminPanelTemplate<'app_data> {
+    dashboard_assets_html: &'static str,
+    template_data: &'app_data TemplateData,
 }
 
 pub fn home(cfg: &mut web::ServiceConfig) {
diff --git a/paddler_balancer/src/web_admin_panel_service/mod.rs b/paddler_balancer/src/web_admin_panel_service/mod.rs
index 5d55a376d..690869145 100644
--- a/paddler_balancer/src/web_admin_panel_service/mod.rs
+++ b/paddler_balancer/src/web_admin_panel_service/mod.rs
@@ -1,18 +1,13 @@
 pub mod app_data;
 pub mod configuration;
-pub mod esbuild_meta_contents;
+pub mod dashboard_assets_html;
 pub mod http_route;
 pub mod template_data;
 
-use std::str::FromStr as _;
-use std::sync::Arc;
-
 use actix_web::App;
 use actix_web::web::Data;
-use anyhow::Context as _;
 use anyhow::Result;
 use async_trait::async_trait;
-use esbuild_metafile::EsbuildMetaFile;
 use tokio_util::sync::CancellationToken;
 use trzcina::Service;
 
@@ -20,7 +15,6 @@ use crate::http_listener::HttpListener;
 use crate::run_http_service::run_http_service;
 use crate::run_http_service_parameters::RunHttpServiceParameters;
 use crate::web_admin_panel_service::app_data::AppData;
-use crate::web_admin_panel_service::esbuild_meta_contents::ESBUILD_META_CONTENTS;
 use crate::web_admin_panel_service::http_route::favicon::favicon;
 use crate::web_admin_panel_service::http_route::home::home;
 use crate::web_admin_panel_service::http_route::static_files::static_files;
@@ -31,18 +25,10 @@ pub struct WebAdminPanelService {
     pub template_data: TemplateData,
 }
 
-#[async_trait]
-impl Service for WebAdminPanelService {
-    fn name(&self) -> &'static str {
-        "balancer::web_admin_panel_service"
-    }
-
-    async fn run(self: Box<Self>, shutdown: CancellationToken) -> Result<()> {
+impl WebAdminPanelService {
+    async fn serve_until_shutdown(self: Box<Self>, shutdown: CancellationToken) -> Result<()> {
         let service_name = self.name();
-        let esbuild_metafile = EsbuildMetaFile::from_str(ESBUILD_META_CONTENTS)
-            .context("the embedded esbuild metafile does not describe the dashboard assets")?;
         let app_data: Data<AppData> = Data::new(AppData {
-            esbuild_metafile: Arc::new(esbuild_metafile),
             template_data: self.template_data,
         });
 
@@ -65,6 +51,17 @@ impl Service for WebAdminPanelService {
     }
 }
 
+#[async_trait]
+impl Service for WebAdminPanelService {
+    fn name(&self) -> &'static str {
+        "balancer::web_admin_panel_service"
+    }
+
+    async fn run(self: Box<Self>, shutdown: CancellationToken) -> Result<()> {
+        self.serve_until_shutdown(shutdown).await
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use std::net::SocketAddr;
@@ -77,6 +74,7 @@ mod tests {
 
     use super::WebAdminPanelService;
     use crate::http_listener::HttpListener;
+
     use crate::resolved_socket_addr::ResolvedSocketAddr;
     use crate::web_admin_panel_service::template_data::TemplateData;
 
@@ -89,6 +87,7 @@ mod tests {
             template_data: TemplateData {
                 buffered_request_timeout: Duration::from_secs(30),
                 compat_openai_addr: None,
+                compat_typesafe_addr: None,
                 inference_addr: ResolvedSocketAddr::from(ephemeral_loopback_addr),
                 management_addr: ResolvedSocketAddr::from(ephemeral_loopback_addr),
                 max_buffered_requests: 32,
diff --git a/paddler_balancer/src/web_admin_panel_service/template_data.rs b/paddler_balancer/src/web_admin_panel_service/template_data.rs
index 679a070da..4e392fd98 100644
--- a/paddler_balancer/src/web_admin_panel_service/template_data.rs
+++ b/paddler_balancer/src/web_admin_panel_service/template_data.rs
@@ -7,6 +7,7 @@ use crate::statsd_service::configuration::Configuration as StatsdServiceConfigur
 pub struct TemplateData {
     pub buffered_request_timeout: Duration,
     pub compat_openai_addr: Option<ResolvedSocketAddr>,
+    pub compat_typesafe_addr: Option<ResolvedSocketAddr>,
     pub inference_addr: ResolvedSocketAddr,
     pub management_addr: ResolvedSocketAddr,
     pub max_buffered_requests: u64,
diff --git a/paddler_balancer/src/websocket_close_cause.rs b/paddler_balancer/src/websocket_close_cause.rs
index fe22056d0..720f38747 100644
--- a/paddler_balancer/src/websocket_close_cause.rs
+++ b/paddler_balancer/src/websocket_close_cause.rs
@@ -11,16 +11,18 @@ const fn described_close_reason(code: CloseCode, description: String) -> CloseRe
     }
 }
 
-#[derive(Debug)]
+#[derive(Debug, Eq, PartialEq)]
 pub enum WebSocketCloseCause {
     AgentAlreadyRegistered,
     AgentDeregistered,
     AgentMessageUndeserializable,
+    AgentResponseBeforeRegistration,
     AgentStatusBeforeRegistration,
     ConnectionCloseRequested,
     IncomingMessageTooBig,
     IncomingStreamFailed,
     PeerClosedConnection,
+    PeerStoppedAnsweringPings,
     ProtocolViolated,
     ServerShuttingDown,
     SessionAlreadyClosed,
@@ -42,6 +44,10 @@ impl WebSocketCloseCause {
                 CloseCode::Invalid,
                 "Agent message could not be deserialized".to_owned(),
             )),
+            Self::AgentResponseBeforeRegistration => Some(described_close_reason(
+                CloseCode::Policy,
+                "Agent sent a response before registering".to_owned(),
+            )),
             Self::AgentStatusBeforeRegistration => Some(described_close_reason(
                 CloseCode::Policy,
                 "Agent sent its status before registering".to_owned(),
@@ -56,6 +62,10 @@ impl WebSocketCloseCause {
                 CloseCode::Error,
                 "Failed to read the incoming message stream".to_owned(),
             )),
+            Self::PeerStoppedAnsweringPings => Some(described_close_reason(
+                CloseCode::Policy,
+                "Peer stopped answering pings".to_owned(),
+            )),
             Self::ProtocolViolated => Some(described_close_reason(
                 CloseCode::Protocol,
                 "WebSocket protocol violated".to_owned(),
diff --git a/paddler_balancer/src/websocket_outbound_frame.rs b/paddler_balancer/src/websocket_outbound_frame.rs
new file mode 100644
index 000000000..9fccd487e
--- /dev/null
+++ b/paddler_balancer/src/websocket_outbound_frame.rs
@@ -0,0 +1,10 @@
+use actix_ws::CloseReason;
+use bytes::Bytes;
+
+#[derive(Debug, Eq, PartialEq)]
+pub enum WebSocketOutboundFrame {
+    Close(Option<CloseReason>),
+    Ping,
+    Pong(Bytes),
+    Text(String),
+}
diff --git a/paddler_balancer/src/websocket_session_controller.rs b/paddler_balancer/src/websocket_session_controller.rs
index 6bf012ece..8993d6d5b 100644
--- a/paddler_balancer/src/websocket_session_controller.rs
+++ b/paddler_balancer/src/websocket_session_controller.rs
@@ -1,18 +1,21 @@
+use std::future::Future;
+use std::future::ready;
 use std::marker::PhantomData;
 
-use actix_ws::Session;
 use anyhow::Result;
 use serde_json::to_string;
+use tokio::sync::mpsc::UnboundedSender;
 
 use paddler_messaging::rpc_message::RpcMessage;
 
 use crate::controls_session::ControlsSession;
+use crate::websocket_outbound_frame::WebSocketOutboundFrame;
 
 pub struct WebSocketSessionController<TResponse>
 where
     TResponse: RpcMessage + Sync,
 {
-    session: Session,
+    outbound_tx: UnboundedSender<WebSocketOutboundFrame>,
     _marker: PhantomData<TResponse>,
 }
 
@@ -21,23 +24,110 @@ where
     TResponse: RpcMessage + Sync,
 {
     #[must_use]
-    pub const fn new(session: Session) -> Self {
+    pub const fn new(outbound_tx: UnboundedSender<WebSocketOutboundFrame>) -> Self {
         Self {
-            session,
+            outbound_tx,
             _marker: PhantomData,
         }
     }
+
+    fn send_text_frame(&self, message: &TResponse) -> Result<()> {
+        self.outbound_tx
+            .send(WebSocketOutboundFrame::Text(to_string(message)?))?;
+
+        Ok(())
+    }
 }
 
 impl<TResponse> ControlsSession<TResponse> for WebSocketSessionController<TResponse>
 where
     TResponse: RpcMessage + Sync + 'static,
 {
-    async fn send_response(&mut self, message: TResponse) -> Result<()> {
-        let serialized_message = to_string(&message)?;
+    fn send_response(&mut self, message: TResponse) -> impl Future<Output = Result<()>> + Send {
+        ready(self.send_text_frame(&message))
+    }
+}
 
-        self.session.text(serialized_message).await?;
+#[cfg(test)]
+mod tests {
+    use serde::Serialize;
+    use serde::Serializer;
+    use serde::ser::Error as _;
+    use serde_json::to_string;
+    use tokio::sync::mpsc::error::SendError;
+    use tokio::sync::mpsc::error::TryRecvError;
+    use tokio::sync::mpsc::unbounded_channel;
 
-        Ok(())
+    use paddler_messaging::inference_client::message::Message as OutgoingMessage;
+    use paddler_messaging::jsonrpc::error::Error as JsonRpcError;
+    use paddler_messaging::jsonrpc::error_envelope::ErrorEnvelope;
+    use paddler_messaging::rpc_message::RpcMessage;
+
+    use super::WebSocketSessionController;
+    use crate::controls_session::ControlsSession as _;
+    use crate::websocket_outbound_frame::WebSocketOutboundFrame;
+
+    struct MessageThatCannotSerialize;
+
+    impl Serialize for MessageThatCannotSerialize {
+        fn serialize<TSerializer>(
+            &self,
+            _serializer: TSerializer,
+        ) -> Result<TSerializer::Ok, TSerializer::Error>
+        where
+            TSerializer: Serializer,
+        {
+            Err(TSerializer::Error::custom(
+                "the message cannot be serialized",
+            ))
+        }
+    }
+
+    impl RpcMessage for MessageThatCannotSerialize {}
+
+    fn request_rejection() -> OutgoingMessage {
+        OutgoingMessage::Error(ErrorEnvelope {
+            request_id: "request".to_owned(),
+            error: JsonRpcError {
+                code: 400,
+                description: "rejected".to_owned(),
+            },
+        })
+    }
+
+    #[tokio::test]
+    async fn a_response_for_a_writer_that_stopped_is_reported_as_undelivered() {
+        let (outbound_tx, outbound_rx) = unbounded_channel();
+
+        drop(outbound_rx);
+
+        let SendError(undelivered_frame) =
+            WebSocketSessionController::<OutgoingMessage>::new(outbound_tx)
+                .send_response(request_rejection())
+                .await
+                .unwrap_err()
+                .downcast::<SendError<WebSocketOutboundFrame>>()
+                .unwrap();
+
+        assert_eq!(
+            undelivered_frame,
+            WebSocketOutboundFrame::Text(to_string(&request_rejection()).unwrap())
+        );
+    }
+
+    #[tokio::test]
+    async fn a_response_that_cannot_be_serialized_never_reaches_the_writer() {
+        let (outbound_tx, mut outbound_rx) = unbounded_channel();
+
+        assert_eq!(
+            WebSocketSessionController::<MessageThatCannotSerialize>::new(outbound_tx)
+                .send_response(MessageThatCannotSerialize)
+                .await
+                .err()
+                .as_ref()
+                .map(ToString::to_string),
+            Some("the message cannot be serialized".to_owned())
+        );
+        assert_eq!(outbound_rx.try_recv(), Err(TryRecvError::Disconnected));
     }
 }
diff --git a/paddler_balancer/src/write_websocket_session.rs b/paddler_balancer/src/write_websocket_session.rs
new file mode 100644
index 000000000..9aca40f9e
--- /dev/null
+++ b/paddler_balancer/src/write_websocket_session.rs
@@ -0,0 +1,120 @@
+use std::pin::pin;
+
+use actix_ws::Closed;
+use actix_ws::Session;
+use futures_util::sink::unfold;
+use log::debug;
+use tokio::sync::mpsc::UnboundedReceiver;
+use tokio_util::sync::CancellationToken;
+
+use paddler_frame_writer::frame_writing_end::FrameWritingEnd;
+use paddler_frame_writer::write_frames_until_connection_closes::write_frames_until_connection_closes;
+
+use crate::websocket_outbound_frame::WebSocketOutboundFrame;
+
+async fn send_frame(session: &mut Session, frame: WebSocketOutboundFrame) -> Result<(), Closed> {
+    match frame {
+        WebSocketOutboundFrame::Close(close_reason) => session.clone().close(close_reason).await,
+        WebSocketOutboundFrame::Ping => session.ping(b"").await,
+        WebSocketOutboundFrame::Pong(payload) => session.pong(&payload).await,
+        WebSocketOutboundFrame::Text(text) => session.text(text).await,
+    }
+}
+
+pub async fn write_websocket_session(
+    connection_close: CancellationToken,
+    outbound_rx: UnboundedReceiver<WebSocketOutboundFrame>,
+    session: Session,
+) {
+    match write_frames_until_connection_closes(
+        connection_close,
+        outbound_rx,
+        pin!(unfold(session, |mut session, frame| async move {
+            send_frame(&mut session, frame).await.map(|()| session)
+        })),
+    )
+    .await
+    {
+        FrameWritingEnd::ConnectionClosed => {
+            debug!("Abandoning the WebSocket session writes because the peer is not reading them");
+        }
+        FrameWritingEnd::EveryFrameSent => {}
+        FrameWritingEnd::SendFailed(Closed) => {
+            debug!("The WebSocket session closed before all of its frames were sent");
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use actix_web::FromRequest as _;
+    use actix_web::HttpResponse;
+    use actix_web::http::header::CONNECTION;
+    use actix_web::http::header::SEC_WEBSOCKET_KEY;
+    use actix_web::http::header::SEC_WEBSOCKET_VERSION;
+    use actix_web::http::header::UPGRADE;
+    use actix_web::test::TestRequest;
+    use actix_web::web::Payload;
+    use actix_ws::Session;
+    use actix_ws::handle;
+    use tokio::spawn;
+    use tokio::sync::mpsc::unbounded_channel;
+    use tokio::task::yield_now;
+    use tokio_util::sync::CancellationToken;
+
+    use super::write_websocket_session;
+    use crate::websocket_outbound_frame::WebSocketOutboundFrame;
+
+    async fn open_websocket_session() -> (HttpResponse, Session) {
+        let (http_request, mut request_payload) = TestRequest::default()
+            .insert_header((UPGRADE, "websocket"))
+            .insert_header((CONNECTION, "upgrade"))
+            .insert_header((SEC_WEBSOCKET_VERSION, "13"))
+            .insert_header((SEC_WEBSOCKET_KEY, "dGhlIHNhbXBsZSBub25jZQ=="))
+            .to_http_parts();
+        let payload = Payload::from_request(&http_request, &mut request_payload)
+            .await
+            .unwrap();
+        let (handshake_response, session, _message_stream) =
+            handle(&http_request, payload).unwrap();
+
+        (handshake_response, session)
+    }
+
+    #[tokio::test]
+    async fn a_writer_abandons_the_writes_its_peer_does_not_read_once_the_connection_closes() {
+        let (_unread_handshake_response, session) = open_websocket_session().await;
+        let connection_close = CancellationToken::new();
+        let (outbound_tx, outbound_rx) = unbounded_channel();
+        let writer = spawn(write_websocket_session(
+            connection_close.clone(),
+            outbound_rx,
+            session,
+        ));
+
+        connection_close.cancel();
+
+        while !writer.is_finished() {
+            outbound_tx.send(WebSocketOutboundFrame::Ping).unwrap();
+            yield_now().await;
+        }
+
+        writer.await.unwrap();
+    }
+
+    #[tokio::test]
+    async fn a_writer_stops_once_its_session_closes() {
+        let (handshake_response, session) = open_websocket_session().await;
+        let (outbound_tx, outbound_rx) = unbounded_channel();
+        let writer = spawn(write_websocket_session(
+            CancellationToken::new(),
+            outbound_rx,
+            session,
+        ));
+
+        drop(handshake_response);
+        outbound_tx.send(WebSocketOutboundFrame::Ping).unwrap();
+
+        writer.await.unwrap();
+    }
+}
diff --git a/paddler_balancer/templates/web_admin_panel.html b/paddler_balancer/templates/web_admin_panel.html
index c33d2ddda..7bf1b3397 100644
--- a/paddler_balancer/templates/web_admin_panel.html
+++ b/paddler_balancer/templates/web_admin_panel.html
@@ -5,10 +5,7 @@
     <meta name="viewport" content="width=device-width, initial-scale=1">
     <title>Paddler Dashboard</title>
 
-    {{ "resources/css/page-admin-panel.css"|asset(preloads) }}
-    {{ "resources/ts/controller_dashboard.tsx"|asset(preloads) }}
-
-    {{ preloads|render_assets|safe }}
+    {{ dashboard_assets_html|safe }}
 </head>
 <body>
     <div
@@ -18,6 +15,9 @@
         {% if let Some(compat_openai_addr) = template_data.compat_openai_addr %}
         data-compat-openai-addr="{{ compat_openai_addr.input_addr }}"
         {% endif %}
+        {% if let Some(compat_typesafe_addr) = template_data.compat_typesafe_addr %}
+        data-compat-typesafe-addr="{{ compat_typesafe_addr.input_addr }}"
+        {% endif %}
         data-inference-addr="{{ template_data.inference_addr.input_addr }}"
         data-management-addr="{{ template_data.management_addr.input_addr }}"
         data-max-buffered-requests="{{ template_data.max_buffered_requests }}"
diff --git a/paddler_blocking_thread_fixture/Cargo.toml b/paddler_blocking_thread_fixture/Cargo.toml
new file mode 100644
index 000000000..49d5e3c14
--- /dev/null
+++ b/paddler_blocking_thread_fixture/Cargo.toml
@@ -0,0 +1,21 @@
+[package]
+name = "paddler_blocking_thread_fixture"
+autotests = false
+authors.workspace = true
+description = "Holds the only blocking thread of a test runtime so Paddler tests can keep blocking work queued"
+edition.workspace = true
+homepage.workspace = true
+license.workspace = true
+publish = false
+repository.workspace = true
+version.workspace = true
+
+[[test]]
+name = "integration"
+path = "tests/main.rs"
+
+[dependencies]
+tokio = { workspace = true }
+
+[lints]
+workspace = true
diff --git a/paddler_blocking_thread_fixture/src/lib.rs b/paddler_blocking_thread_fixture/src/lib.rs
new file mode 100644
index 000000000..3e4afafc9
--- /dev/null
+++ b/paddler_blocking_thread_fixture/src/lib.rs
@@ -0,0 +1,2 @@
+pub mod occupied_blocking_thread;
+pub mod single_blocking_thread_runtime;
diff --git a/paddler_blocking_thread_fixture/src/occupied_blocking_thread.rs b/paddler_blocking_thread_fixture/src/occupied_blocking_thread.rs
new file mode 100644
index 000000000..4fceea7b0
--- /dev/null
+++ b/paddler_blocking_thread_fixture/src/occupied_blocking_thread.rs
@@ -0,0 +1,38 @@
+use std::sync::Arc;
+
+use tokio::runtime::Handle;
+use tokio::sync::Notify;
+use tokio::task::JoinError;
+use tokio::task::JoinHandle;
+use tokio::task::spawn_blocking;
+
+pub struct OccupiedBlockingThread {
+    occupant: JoinHandle<()>,
+    release: Arc<Notify>,
+}
+
+impl OccupiedBlockingThread {
+    #[must_use]
+    pub fn occupy() -> Self {
+        let release = Arc::new(Notify::new());
+        let occupant_release = release.clone();
+        let runtime = Handle::current();
+
+        Self {
+            occupant: spawn_blocking(move || runtime.block_on(occupant_release.notified())),
+            release,
+        }
+    }
+
+    pub async fn release(mut self) -> Result<(), JoinError> {
+        self.release.notify_one();
+
+        (&mut self.occupant).await
+    }
+}
+
+impl Drop for OccupiedBlockingThread {
+    fn drop(&mut self) {
+        self.release.notify_one();
+    }
+}
diff --git a/paddler_blocking_thread_fixture/src/single_blocking_thread_runtime.rs b/paddler_blocking_thread_fixture/src/single_blocking_thread_runtime.rs
new file mode 100644
index 000000000..14a683f59
--- /dev/null
+++ b/paddler_blocking_thread_fixture/src/single_blocking_thread_runtime.rs
@@ -0,0 +1,13 @@
+use std::io::Result as IoResult;
+
+use tokio::runtime::Builder;
+use tokio::runtime::Runtime;
+
+const ONLY_BLOCKING_THREAD: usize = 1;
+
+pub fn single_blocking_thread_runtime() -> IoResult<Runtime> {
+    Builder::new_current_thread()
+        .max_blocking_threads(ONLY_BLOCKING_THREAD)
+        .enable_all()
+        .build()
+}
diff --git a/paddler_blocking_thread_fixture/tests/a_dropped_occupied_blocking_thread_lets_the_next_blocking_task_run.rs b/paddler_blocking_thread_fixture/tests/a_dropped_occupied_blocking_thread_lets_the_next_blocking_task_run.rs
new file mode 100644
index 000000000..1cb7fe619
--- /dev/null
+++ b/paddler_blocking_thread_fixture/tests/a_dropped_occupied_blocking_thread_lets_the_next_blocking_task_run.rs
@@ -0,0 +1,15 @@
+use tokio::task::spawn_blocking;
+
+use paddler_blocking_thread_fixture::occupied_blocking_thread::OccupiedBlockingThread;
+use paddler_blocking_thread_fixture::single_blocking_thread_runtime::single_blocking_thread_runtime;
+
+#[test]
+fn a_dropped_occupied_blocking_thread_lets_the_next_blocking_task_run() {
+    single_blocking_thread_runtime()
+        .expect("the test runtime must build")
+        .block_on(async {
+            drop(OccupiedBlockingThread::occupy());
+
+            assert_eq!(spawn_blocking(|| ()).await.ok(), Some(()));
+        });
+}
diff --git a/paddler_blocking_thread_fixture/tests/main.rs b/paddler_blocking_thread_fixture/tests/main.rs
new file mode 100644
index 000000000..dab8db074
--- /dev/null
+++ b/paddler_blocking_thread_fixture/tests/main.rs
@@ -0,0 +1 @@
+mod a_dropped_occupied_blocking_thread_lets_the_next_blocking_task_run;
diff --git a/paddler_bootstrap/Cargo.toml b/paddler_bootstrap/Cargo.toml
index e86ca7b67..9b5aac7e5 100644
--- a/paddler_bootstrap/Cargo.toml
+++ b/paddler_bootstrap/Cargo.toml
@@ -4,7 +4,7 @@ autotests = false
 version.workspace = true
 edition.workspace = true
 authors.workspace = true
-description = "Shared service wiring for Paddler CLI and desktop app"
+description = "Starts and runs the Paddler balancer and agent services"
 homepage.workspace = true
 license.workspace = true
 repository.workspace = true
@@ -16,14 +16,14 @@ path = "tests/main.rs"
 [dependencies]
 actix-web = { workspace = true }
 anyhow = { workspace = true }
-async-trait = { workspace = true }
 log = { workspace = true }
 nanoid = { workspace = true }
 paddler_agent = { workspace = true }
+paddler_agent_runtime = { workspace = true }
 paddler_balancer = { workspace = true }
-paddler_inference_parameters = { workspace = true }
-paddler_agent_status = { workspace = true }
 paddler_messaging = { workspace = true }
+paddler_model_source = { workspace = true }
+paddler_outbox = { workspace = true }
 paddler_state_database = { workspace = true }
 thiserror = { workspace = true }
 tokio = { workspace = true }
@@ -31,10 +31,12 @@ tokio-util = { workspace = true }
 trzcina = { workspace = true }
 
 [dev-dependencies]
+paddler_client = { workspace = true }
 paddler_test_cluster_harness = { workspace = true }
 reqwest = { workspace = true }
 serde_json = { workspace = true }
 tempfile = { workspace = true }
+url = { workspace = true }
 
 [lints]
 workspace = true
diff --git a/paddler_bootstrap/src/agent_bootstrap_config.rs b/paddler_bootstrap/src/agent_bootstrap_config.rs
index d61acf6b6..3f4c9983a 100644
--- a/paddler_bootstrap/src/agent_bootstrap_config.rs
+++ b/paddler_bootstrap/src/agent_bootstrap_config.rs
@@ -1,6 +1,8 @@
-#[derive(Debug, Eq, PartialEq)]
+use paddler_model_source::desired_model_resolver::DesiredModelResolver;
+
 pub struct AgentBootstrapConfig {
     pub agent_name: Option<String>,
+    pub desired_model_resolver: DesiredModelResolver,
     pub management_address: String,
     pub slots: u16,
 }
diff --git a/paddler_bootstrap/src/agent_runner.rs b/paddler_bootstrap/src/agent_runner.rs
index 488ad4f6a..be28a4215 100644
--- a/paddler_bootstrap/src/agent_runner.rs
+++ b/paddler_bootstrap/src/agent_runner.rs
@@ -1,11 +1,3 @@
-use std::sync::Arc;
-
-use anyhow::Result;
-use tokio::sync::watch;
-
-use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
-use paddler_messaging::balancer_connection::BalancerConnection;
-
 use crate::agent_runner_params::AgentRunnerParams;
 use crate::agent_service_bundle::AgentServiceBundle;
 use crate::bootstrap_error::BootstrapError;
@@ -13,8 +5,6 @@ use crate::run_service_manager::run_service_manager;
 use crate::service_thread::ServiceThread;
 
 pub struct AgentRunner {
-    pub balancer_connection_rx: watch::Receiver<BalancerConnection>,
-    pub slot_aggregated_status: Arc<SlotAggregatedStatus>,
     thread: ServiceThread,
 }
 
@@ -28,21 +18,15 @@ impl AgentRunner {
         }: AgentRunnerParams,
     ) -> Self {
         let bundle = AgentServiceBundle::new(bootstrap_config);
-        let balancer_connection_rx = bundle.balancer_connection_rx.clone();
-        let slot_aggregated_status = bundle.slot_aggregated_status.clone();
 
         let thread = ServiceThread::spawn(cancellation_token, move |task_shutdown| {
             run_service_manager(bundle, task_shutdown, shutdown_options)
         });
 
-        Self {
-            balancer_connection_rx,
-            slot_aggregated_status,
-            thread,
-        }
+        Self { thread }
     }
 
-    pub async fn wait_for_completion(self) -> Result<(), BootstrapError> {
+    pub async fn wait_for_completion(&mut self) -> Result<(), BootstrapError> {
         self.thread.wait_for_completion().await
     }
 
diff --git a/paddler_bootstrap/src/agent_service_bundle.rs b/paddler_bootstrap/src/agent_service_bundle.rs
index 09869c6ab..caa95798b 100644
--- a/paddler_bootstrap/src/agent_service_bundle.rs
+++ b/paddler_bootstrap/src/agent_service_bundle.rs
@@ -1,35 +1,28 @@
 use std::sync::Arc;
 
-use anyhow::Result;
-use async_trait::async_trait;
 use nanoid::nanoid;
-use tokio::sync::mpsc;
 use tokio::sync::watch;
-use trzcina::Service;
-use trzcina::ServiceBundle;
+use trzcina::ServiceManager;
 
 use paddler_agent::agent_applicable_state_holder::AgentApplicableStateHolder;
 use paddler_agent::balancer_message_context::BalancerMessageContext;
-use paddler_agent::continuous_batch_arbiter_context::ContinuousBatchArbiterContext;
-use paddler_agent::continuous_batch_preparation_request::ContinuousBatchPreparationRequest;
 use paddler_agent::desired_state_reconciler::DesiredStateReconciler;
 use paddler_agent::llamacpp_arbiter_service::LlamaCppArbiterService;
 use paddler_agent::management_socket_client_service::ManagementSocketClientService;
-use paddler_agent::model_metadata_holder::ModelMetadataHolder;
 use paddler_agent::reconciliation_service::ReconciliationService;
-use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
+use paddler_agent_runtime::inference_runtime_context::InferenceRuntimeContext;
+use paddler_agent_runtime::inference_runtime_context_params::InferenceRuntimeContextParams;
 use paddler_messaging::agent_desired_state::AgentDesiredState;
 use paddler_messaging::api_path::ApiPath;
-use paddler_messaging::balancer_connection::BalancerConnection;
+use paddler_outbox::outbox::Outbox;
 
 use crate::agent_bootstrap_config::AgentBootstrapConfig;
+use crate::registers_services::RegistersServices;
 
 pub struct AgentServiceBundle {
-    pub balancer_connection_rx: watch::Receiver<BalancerConnection>,
-    pub slot_aggregated_status: Arc<SlotAggregatedStatus>,
-    llamacpp_arbiter_service: LlamaCppArbiterService,
-    management_socket_client_service: ManagementSocketClientService,
-    reconciliation_service: ReconciliationService,
+    llamacpp_arbiter: LlamaCppArbiterService,
+    management_socket_client: ManagementSocketClientService,
+    reconciliation: ReconciliationService,
 }
 
 impl AgentServiceBundle {
@@ -37,39 +30,42 @@ impl AgentServiceBundle {
     pub fn new(
         AgentBootstrapConfig {
             agent_name,
+            desired_model_resolver,
             management_address,
             slots,
         }: AgentBootstrapConfig,
     ) -> Self {
-        let (balancer_connection_tx, balancer_connection_rx) =
-            watch::channel(BalancerConnection::Connecting);
-        let (agent_desired_state_tx, agent_desired_state_rx) =
-            mpsc::unbounded_channel::<AgentDesiredState>();
-        let (continuous_batch_preparation_request_tx, continuous_batch_preparation_request_rx) =
-            mpsc::unbounded_channel::<ContinuousBatchPreparationRequest>();
+        let (agent_desired_state_tx, mut agent_desired_state_rx) =
+            watch::channel(AgentDesiredState::default());
+
+        agent_desired_state_rx.mark_changed();
+
+        let pipeline_requests = Arc::new(Outbox::default());
 
         let agent_applicable_state_holder = Arc::new(AgentApplicableStateHolder::default());
-        let model_metadata_holder = Arc::new(ModelMetadataHolder::default());
-        let slot_aggregated_status = Arc::new(SlotAggregatedStatus::new(slots));
+        let served_applicable_state_holder = Arc::new(AgentApplicableStateHolder::default());
+        let inference_runtime_context =
+            InferenceRuntimeContext::new(InferenceRuntimeContextParams {
+                agent_name: agent_name.clone(),
+                slots,
+            });
+        let model_metadata_holder = inference_runtime_context.model_metadata_holder.clone();
+        let slot_aggregated_status = inference_runtime_context.slot_aggregated_status.clone();
 
-        let llamacpp_arbiter_service = LlamaCppArbiterService {
+        let llamacpp_arbiter = LlamaCppArbiterService {
             agent_applicable_state_holder: agent_applicable_state_holder.clone(),
-            arbiter_context: ContinuousBatchArbiterContext {
-                agent_name: agent_name.clone(),
-                model_metadata_holder: model_metadata_holder.clone(),
-                slot_aggregated_status: slot_aggregated_status.clone(),
-            },
-            continuous_batch_preparation_request_rx,
+            inference_runtime_context,
+            pipeline_requests: pipeline_requests.clone(),
+            served_applicable_state_holder: served_applicable_state_holder.clone(),
         };
 
-        let management_socket_client_service = ManagementSocketClientService {
-            balancer_connection_tx,
+        let management_socket_client = ManagementSocketClientService {
             balancer_message_context: BalancerMessageContext {
-                agent_applicable_state_holder: agent_applicable_state_holder.clone(),
                 agent_desired_state_tx,
-                continuous_batch_preparation_request_tx,
                 model_metadata_holder,
+                pipeline_requests,
                 request_stoppers: Arc::default(),
+                served_applicable_state_holder,
                 slot_aggregated_status: slot_aggregated_status.clone(),
             },
             name: agent_name,
@@ -79,31 +75,27 @@ impl AgentServiceBundle {
             ),
         };
 
-        let reconciliation_service = ReconciliationService {
+        let reconciliation = ReconciliationService {
+            agent_applicable_state_holder,
             agent_desired_state_rx,
             desired_state_reconciler: DesiredStateReconciler {
-                agent_applicable_state_holder,
-                slot_aggregated_status: slot_aggregated_status.clone(),
+                desired_model_resolver,
+                slot_aggregated_status,
             },
         };
 
         Self {
-            balancer_connection_rx,
-            slot_aggregated_status,
-            llamacpp_arbiter_service,
-            management_socket_client_service,
-            reconciliation_service,
+            llamacpp_arbiter,
+            management_socket_client,
+            reconciliation,
         }
     }
 }
 
-#[async_trait]
-impl ServiceBundle for AgentServiceBundle {
-    async fn services(self) -> Result<Vec<Box<dyn Service>>> {
-        Ok(vec![
-            Box::new(self.llamacpp_arbiter_service),
-            Box::new(self.management_socket_client_service),
-            Box::new(self.reconciliation_service),
-        ])
+impl RegistersServices for AgentServiceBundle {
+    fn register_services(self, service_manager: &mut ServiceManager) {
+        service_manager.register_service(self.llamacpp_arbiter);
+        service_manager.register_service(self.management_socket_client);
+        service_manager.register_service(self.reconciliation);
     }
 }
diff --git a/paddler_bootstrap/src/balancer_bootstrap_config.rs b/paddler_bootstrap/src/balancer_bootstrap_config.rs
index e1f6e5101..7917a75ae 100644
--- a/paddler_bootstrap/src/balancer_bootstrap_config.rs
+++ b/paddler_bootstrap/src/balancer_bootstrap_config.rs
@@ -1,6 +1,6 @@
 use std::time::Duration;
 
-use paddler_balancer::compatibility::openai_service::configuration::Configuration as OpenAIServiceConfiguration;
+use paddler_balancer::compatibility::compatibility_service_configuration::CompatibilityServiceConfiguration;
 use paddler_balancer::inference_service::configuration::Configuration as InferenceServiceConfiguration;
 use paddler_balancer::management_service::configuration::Configuration as ManagementServiceConfiguration;
 use paddler_balancer::statsd_service::configuration::Configuration as StatsdServiceConfiguration;
@@ -13,10 +13,11 @@ pub struct BalancerBootstrapConfig {
     pub inference_service_configuration: InferenceServiceConfiguration,
     pub management_service_configuration: ManagementServiceConfiguration,
     pub max_buffered_requests: u64,
-    pub openai_service_configuration: Option<OpenAIServiceConfiguration>,
+    pub openai_service_configuration: Option<CompatibilityServiceConfiguration>,
     pub state_database_type: StateDatabaseType,
     pub statsd_prefix: String,
     pub statsd_service_configuration: Option<StatsdServiceConfiguration>,
+    pub typesafe_service_configuration: Option<CompatibilityServiceConfiguration>,
     #[cfg(feature = "web_admin_panel")]
     pub web_admin_panel_service_configuration: Option<WebAdminPanelServiceConfiguration>,
 }
diff --git a/paddler_bootstrap/src/balancer_defaults.rs b/paddler_bootstrap/src/balancer_defaults.rs
index 213424105..ef3df3952 100644
--- a/paddler_bootstrap/src/balancer_defaults.rs
+++ b/paddler_bootstrap/src/balancer_defaults.rs
@@ -10,5 +10,4 @@ impl BalancerDefaults {
     pub const MAX_BUFFERED_REQUESTS: u64 = 30;
     pub const STATSD_PREFIX: &str = "paddler_";
     pub const STATSD_REPORTING_INTERVAL: Duration = Duration::from_secs(10);
-    pub const WEB_ADMIN_PANEL_PORT: u16 = 8062;
 }
diff --git a/paddler_bootstrap/src/balancer_runner.rs b/paddler_bootstrap/src/balancer_runner.rs
index 9e449054f..52c5e704c 100644
--- a/paddler_bootstrap/src/balancer_runner.rs
+++ b/paddler_bootstrap/src/balancer_runner.rs
@@ -1,12 +1,7 @@
 use std::sync::Arc;
 
-use anyhow::Result;
-use tokio::sync::watch;
-
-use paddler_balancer::agent_controller_pool::AgentControllerPool;
 use paddler_balancer::balancer_addresses::BalancerAddresses;
 use paddler_balancer::balancer_applicable_state_holder::BalancerApplicableStateHolder;
-use paddler_messaging::balancer_desired_state::BalancerDesiredState;
 
 use crate::balancer_runner_params::BalancerRunnerParams;
 use crate::balancer_service_bundle::BalancerServiceBundle;
@@ -16,9 +11,7 @@ use crate::service_thread::ServiceThread;
 
 pub struct BalancerRunner {
     pub addresses: BalancerAddresses,
-    pub agent_controller_pool: Arc<AgentControllerPool>,
     pub balancer_applicable_state_holder: Arc<BalancerApplicableStateHolder>,
-    pub balancer_desired_state_tx: watch::Sender<BalancerDesiredState>,
     thread: ServiceThread,
 }
 
@@ -33,9 +26,7 @@ impl BalancerRunner {
         let bundle = BalancerServiceBundle::new(bootstrap_config).await?;
 
         let addresses = bundle.addresses;
-        let agent_controller_pool = bundle.agent_controller_pool.clone();
         let balancer_applicable_state_holder = bundle.balancer_applicable_state_holder.clone();
-        let balancer_desired_state_tx = bundle.balancer_desired_state_tx.clone();
 
         let thread = ServiceThread::spawn(cancellation_token, move |task_shutdown| {
             run_service_manager(bundle, task_shutdown, shutdown_options)
@@ -43,14 +34,12 @@ impl BalancerRunner {
 
         Ok(Self {
             addresses,
-            agent_controller_pool,
             balancer_applicable_state_holder,
-            balancer_desired_state_tx,
             thread,
         })
     }
 
-    pub async fn wait_for_completion(self) -> Result<(), BootstrapError> {
+    pub async fn wait_for_completion(&mut self) -> Result<(), BootstrapError> {
         self.thread.wait_for_completion().await
     }
 
diff --git a/paddler_bootstrap/src/balancer_service_bundle.rs b/paddler_bootstrap/src/balancer_service_bundle.rs
index a9ce64a7c..bfac4bfb7 100644
--- a/paddler_bootstrap/src/balancer_service_bundle.rs
+++ b/paddler_bootstrap/src/balancer_service_bundle.rs
@@ -1,20 +1,17 @@
 use std::sync::Arc;
 
-use anyhow::Result;
-use async_trait::async_trait;
-use tokio::sync::watch;
-use trzcina::Service;
-use trzcina::ServiceBundle;
+use trzcina::ServiceManager;
 
 use paddler_balancer::agent_controller_pool::AgentControllerPool;
-use paddler_balancer::agent_response_senders::AgentResponseSenders;
 use paddler_balancer::balancer_addresses::BalancerAddresses;
-use paddler_balancer::balancer_applicable_state::BalancerApplicableState;
 use paddler_balancer::balancer_applicable_state_holder::BalancerApplicableStateHolder;
+use paddler_balancer::balancer_desired_state_store::BalancerDesiredStateStore;
 use paddler_balancer::buffered_request_manager::BufferedRequestManager;
-use paddler_balancer::compatibility::openai_service::OpenAIService;
+use paddler_balancer::compatibility::compatibility_service::CompatibilityService;
 #[cfg(feature = "web_admin_panel")]
-use paddler_balancer::compatibility::openai_service::configuration::Configuration as OpenAIServiceConfiguration;
+use paddler_balancer::compatibility::compatibility_service_configuration::CompatibilityServiceConfiguration;
+use paddler_balancer::compatibility::openai_service::openai_compatibility_layer::OpenAICompatibilityLayer;
+use paddler_balancer::compatibility::typesafe_service::typesafe_compatibility_layer::TypeSafeCompatibilityLayer;
 use paddler_balancer::http_listener::HttpListener;
 use paddler_balancer::inference_service::InferenceService;
 use paddler_balancer::management_service::ManagementService;
@@ -26,26 +23,21 @@ use paddler_balancer::web_admin_panel_service::WebAdminPanelService;
 use paddler_balancer::web_admin_panel_service::configuration::Configuration as WebAdminPanelServiceConfiguration;
 #[cfg(feature = "web_admin_panel")]
 use paddler_balancer::web_admin_panel_service::template_data::TemplateData;
-use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-use paddler_state_database::file::File;
-use paddler_state_database::memory::Memory;
 use paddler_state_database::state_database::StateDatabase;
-use paddler_state_database::state_database_type::StateDatabaseType;
 
 use crate::balancer_bootstrap_config::BalancerBootstrapConfig;
 use crate::bootstrap_error::BootstrapError;
+use crate::registers_services::RegistersServices;
 
 pub struct BalancerServiceBundle {
     pub addresses: BalancerAddresses,
-    pub agent_controller_pool: Arc<AgentControllerPool>,
     pub balancer_applicable_state_holder: Arc<BalancerApplicableStateHolder>,
-    pub balancer_desired_state_tx: watch::Sender<BalancerDesiredState>,
-    pub state_database: Arc<dyn StateDatabase>,
     inference_service: InferenceService,
     management_service: ManagementService,
     reconciliation_service: ReconciliationService,
-    openai_service: Option<OpenAIService>,
+    openai_service: Option<CompatibilityService<OpenAICompatibilityLayer>>,
     statsd_service: Option<StatsdService>,
+    typesafe_service: Option<CompatibilityService<TypeSafeCompatibilityLayer>>,
     #[cfg(feature = "web_admin_panel")]
     web_admin_panel_service: Option<WebAdminPanelService>,
 }
@@ -61,6 +53,7 @@ impl BalancerServiceBundle {
             state_database_type,
             statsd_prefix,
             statsd_service_configuration,
+            typesafe_service_configuration,
             #[cfg(feature = "web_admin_panel")]
             web_admin_panel_service_configuration,
         }: BalancerBootstrapConfig,
@@ -76,35 +69,20 @@ impl BalancerServiceBundle {
             return Err(BootstrapError::StatsdReportingIntervalIsZero);
         }
 
-        let (balancer_desired_state_tx, _initial_desired_state_rx) =
-            watch::channel(BalancerDesiredState::default());
-
         let agent_controller_pool = Arc::new(AgentControllerPool::default());
         let buffered_request_manager = Arc::new(BufferedRequestManager::new(
             agent_controller_pool.clone(),
             buffered_request_timeout,
             max_buffered_requests,
         ));
-        let state_database: Arc<dyn StateDatabase> = match state_database_type {
-            StateDatabaseType::File(path) => {
-                Arc::new(File::new(balancer_desired_state_tx.clone(), path))
-            }
-            StateDatabaseType::Memory(initial_desired_state) => Arc::new(Memory::new(
-                balancer_desired_state_tx.clone(),
-                *initial_desired_state,
-            )),
-        };
-
-        let initial_desired_state = state_database
-            .read_balancer_desired_state()
-            .await
-            .map_err(|source| BootstrapError::StateDatabaseReadFailed { source })?;
-
-        balancer_desired_state_tx.send_replace(initial_desired_state.clone());
-
-        let balancer_applicable_state_holder = Arc::new(BalancerApplicableStateHolder::new(
-            BalancerApplicableState::from(initial_desired_state),
-        ));
+        let state_database = Arc::new(StateDatabase::from(state_database_type));
+        let balancer_desired_state_store = Arc::new(
+            BalancerDesiredStateStore::load(state_database)
+                .await
+                .map_err(|source| BootstrapError::StateDatabaseReadFailed { source })?,
+        );
+        let balancer_applicable_state_holder =
+            Arc::new(balancer_desired_state_store.applicable_state_holder());
 
         let inference_addr = inference_service_configuration.addr.socket_addr;
         let inference_http_listener = HttpListener::bind(inference_addr).map_err(|source| {
@@ -133,6 +111,19 @@ impl BalancerServiceBundle {
                 })
             })
             .transpose()?;
+        let typesafe_http_listener = typesafe_service_configuration
+            .as_ref()
+            .map(|typesafe_service_configuration| {
+                let typesafe_addr = typesafe_service_configuration.addr.socket_addr;
+
+                HttpListener::bind(typesafe_addr).map_err(|source| {
+                    BootstrapError::CompatTypeSafeBindFailed {
+                        addr: typesafe_addr,
+                        source,
+                    }
+                })
+            })
+            .transpose()?;
         #[cfg(feature = "web_admin_panel")]
         let web_admin_panel_service = web_admin_panel_service_configuration
             .map(|WebAdminPanelServiceConfiguration { addr }| {
@@ -142,7 +133,9 @@ impl BalancerServiceBundle {
                         template_data: TemplateData {
                             buffered_request_timeout,
                             compat_openai_addr: openai_service_configuration
-                                .map(|OpenAIServiceConfiguration { addr }| addr),
+                                .map(|CompatibilityServiceConfiguration { addr }| addr),
+                            compat_typesafe_addr: typesafe_service_configuration
+                                .map(|CompatibilityServiceConfiguration { addr }| addr),
                             inference_addr: inference_service_configuration.addr.clone(),
                             management_addr: management_service_configuration.addr.clone(),
                             max_buffered_requests,
@@ -164,6 +157,9 @@ impl BalancerServiceBundle {
             compat_openai: openai_http_listener
                 .as_ref()
                 .map(|openai_http_listener| openai_http_listener.local_addr),
+            compat_typesafe: typesafe_http_listener
+                .as_ref()
+                .map(|typesafe_http_listener| typesafe_http_listener.local_addr),
             inference: inference_http_listener.local_addr,
             management: management_http_listener.local_addr,
             web_admin_panel: web_admin_panel_addr,
@@ -181,11 +177,10 @@ impl BalancerServiceBundle {
         let management_service = ManagementService {
             agent_controller_pool: agent_controller_pool.clone(),
             balancer_applicable_state_holder: balancer_applicable_state_holder.clone(),
-            agent_response_senders: AgentResponseSenders::default(),
             buffered_request_manager: buffered_request_manager.clone(),
             configuration: management_service_configuration,
+            balancer_desired_state_store,
             http_listener: management_http_listener,
-            state_database: state_database.clone(),
             statsd_prefix: statsd_prefix.clone(),
             web_admin_panel_addr,
         };
@@ -193,12 +188,22 @@ impl BalancerServiceBundle {
         let reconciliation_service = ReconciliationService {
             agent_controller_pool: agent_controller_pool.clone(),
             balancer_applicable_state_holder: balancer_applicable_state_holder.clone(),
-            balancer_desired_state_rx: balancer_desired_state_tx.subscribe(),
         };
 
-        let openai_service = openai_http_listener.map(|http_listener| OpenAIService {
+        let openai_service = openai_http_listener.map(|http_listener| CompatibilityService {
+            agent_controller_pool: agent_controller_pool.clone(),
             balancer_applicable_state_holder: balancer_applicable_state_holder.clone(),
             buffered_request_manager: buffered_request_manager.clone(),
+            compatibility_layer: OpenAICompatibilityLayer,
+            http_listener,
+            inference_service_configuration: inference_service_configuration.clone(),
+        });
+
+        let typesafe_service = typesafe_http_listener.map(|http_listener| CompatibilityService {
+            agent_controller_pool: agent_controller_pool.clone(),
+            balancer_applicable_state_holder: balancer_applicable_state_holder.clone(),
+            buffered_request_manager: buffered_request_manager.clone(),
+            compatibility_layer: TypeSafeCompatibilityLayer,
             http_listener,
             inference_service_configuration,
         });
@@ -212,44 +217,41 @@ impl BalancerServiceBundle {
 
         Ok(Self {
             addresses,
-            agent_controller_pool,
             balancer_applicable_state_holder,
-            balancer_desired_state_tx,
-            state_database,
             inference_service,
             management_service,
             reconciliation_service,
             openai_service,
             statsd_service,
+            typesafe_service,
             #[cfg(feature = "web_admin_panel")]
             web_admin_panel_service,
         })
     }
 }
 
-#[async_trait]
-impl ServiceBundle for BalancerServiceBundle {
-    async fn services(self) -> Result<Vec<Box<dyn Service>>> {
-        let mut services: Vec<Box<dyn Service>> = vec![
-            Box::new(self.inference_service),
-            Box::new(self.management_service),
-            Box::new(self.reconciliation_service),
-        ];
+impl RegistersServices for BalancerServiceBundle {
+    fn register_services(self, service_manager: &mut ServiceManager) {
+        service_manager.register_service(self.inference_service);
+        service_manager.register_service(self.management_service);
+        service_manager.register_service(self.reconciliation_service);
 
         if let Some(service) = self.openai_service {
-            services.push(Box::new(service));
+            service_manager.register_service(service);
         }
 
         if let Some(service) = self.statsd_service {
-            services.push(Box::new(service));
+            service_manager.register_service(service);
+        }
+
+        if let Some(service) = self.typesafe_service {
+            service_manager.register_service(service);
         }
 
         #[cfg(feature = "web_admin_panel")]
         if let Some(service) = self.web_admin_panel_service {
-            services.push(Box::new(service));
+            service_manager.register_service(service);
         }
-
-        Ok(services)
     }
 }
 
@@ -257,9 +259,7 @@ impl ServiceBundle for BalancerServiceBundle {
 mod tests {
     use std::time::Duration;
 
-    use trzcina::ServiceBundle as _;
-
-    use paddler_balancer::compatibility::openai_service::configuration::Configuration as OpenAIServiceConfiguration;
+    use paddler_balancer::compatibility::compatibility_service_configuration::CompatibilityServiceConfiguration;
     use paddler_balancer::inference_service::configuration::Configuration as InferenceServiceConfiguration;
     use paddler_balancer::management_service::configuration::Configuration as ManagementServiceConfiguration;
     use paddler_balancer::resolved_socket_addr::ResolvedSocketAddr;
@@ -267,17 +267,13 @@ mod tests {
     #[cfg(feature = "web_admin_panel")]
     use paddler_balancer::web_admin_panel_service::configuration::Configuration as WebAdminPanelServiceConfiguration;
     use paddler_state_database::state_database_type::StateDatabaseType;
+    use paddler_state_database::stored_balancer_desired_state::StoredBalancerDesiredState;
     use paddler_test_cluster_harness::ephemeral_loopback_addr::EPHEMERAL_LOOPBACK_ADDR;
 
     use super::BalancerServiceBundle;
     use crate::balancer_bootstrap_config::BalancerBootstrapConfig;
 
-    #[cfg(feature = "web_admin_panel")]
-    const EXPECTED_SERVICE_COUNT: usize = 6;
-    #[cfg(not(feature = "web_admin_panel"))]
-    const EXPECTED_SERVICE_COUNT: usize = 5;
-
-    fn fully_configured_bootstrap_config() -> BalancerBootstrapConfig {
+    fn fully_configured_runner_config() -> BalancerBootstrapConfig {
         BalancerBootstrapConfig {
             buffered_request_timeout: Duration::from_secs(10),
             inference_service_configuration: InferenceServiceConfiguration {
@@ -290,15 +286,20 @@ mod tests {
                 cors_allowed_hosts: vec![],
             },
             max_buffered_requests: 30,
-            openai_service_configuration: Some(OpenAIServiceConfiguration {
+            openai_service_configuration: Some(CompatibilityServiceConfiguration {
                 addr: ResolvedSocketAddr::from(EPHEMERAL_LOOPBACK_ADDR),
             }),
-            state_database_type: StateDatabaseType::Memory(Box::default()),
+            state_database_type: StateDatabaseType::Memory(
+                StoredBalancerDesiredState::NothingStored,
+            ),
             statsd_prefix: "paddler_bootstrap_test_".to_owned(),
             statsd_service_configuration: Some(StatsdServiceConfiguration {
                 statsd_addr: ResolvedSocketAddr::from(EPHEMERAL_LOOPBACK_ADDR),
                 statsd_reporting_interval: Duration::from_secs(10),
             }),
+            typesafe_service_configuration: Some(CompatibilityServiceConfiguration {
+                addr: ResolvedSocketAddr::from(EPHEMERAL_LOOPBACK_ADDR),
+            }),
             #[cfg(feature = "web_admin_panel")]
             web_admin_panel_service_configuration: Some(WebAdminPanelServiceConfiguration {
                 addr: EPHEMERAL_LOOPBACK_ADDR,
@@ -306,23 +307,9 @@ mod tests {
         }
     }
 
-    #[tokio::test]
-    async fn services_includes_every_optional_service_when_configured() {
-        let bundle = BalancerServiceBundle::new(fully_configured_bootstrap_config())
-            .await
-            .expect("a fully configured bundle must bind its ephemeral addresses");
-
-        let services = bundle
-            .services()
-            .await
-            .expect("a bound bundle must produce its services");
-
-        assert_eq!(services.len(), EXPECTED_SERVICE_COUNT);
-    }
-
     #[tokio::test]
     async fn reports_the_bound_address_of_every_configured_service() {
-        let bundle = BalancerServiceBundle::new(fully_configured_bootstrap_config())
+        let bundle = BalancerServiceBundle::new(fully_configured_runner_config())
             .await
             .expect("a fully configured bundle must bind its ephemeral addresses");
 
@@ -331,6 +318,10 @@ mod tests {
                 .addresses
                 .compat_openai
                 .map(|compat_openai| compat_openai.port()),
+            bundle
+                .addresses
+                .compat_typesafe
+                .map(|compat_typesafe| compat_typesafe.port()),
             Some(bundle.addresses.inference.port()),
             Some(bundle.addresses.management.port()),
             #[cfg(feature = "web_admin_panel")]
diff --git a/paddler_bootstrap/src/bootstrap_error.rs b/paddler_bootstrap/src/bootstrap_error.rs
index 23ac7e2b9..ead028c19 100644
--- a/paddler_bootstrap/src/bootstrap_error.rs
+++ b/paddler_bootstrap/src/bootstrap_error.rs
@@ -14,6 +14,12 @@ pub enum BootstrapError {
         #[source]
         source: io::Error,
     },
+    #[error("Unable to bind the TypeSafe compatibility service to {addr}")]
+    CompatTypeSafeBindFailed {
+        addr: SocketAddr,
+        #[source]
+        source: io::Error,
+    },
     #[error("Unable to bind the inference service to {addr}")]
     InferenceBindFailed {
         addr: SocketAddr,
diff --git a/paddler_bootstrap/src/lib.rs b/paddler_bootstrap/src/lib.rs
index fcba65f53..254e1f404 100644
--- a/paddler_bootstrap/src/lib.rs
+++ b/paddler_bootstrap/src/lib.rs
@@ -9,5 +9,6 @@ pub mod balancer_runner_params;
 pub mod balancer_service_bundle;
 pub mod bootstrap_error;
 pub mod llama_cpp_max_sequences;
+pub mod registers_services;
 pub mod run_service_manager;
 pub mod service_thread;
diff --git a/paddler_bootstrap/src/registers_services.rs b/paddler_bootstrap/src/registers_services.rs
new file mode 100644
index 000000000..34ea2e4c4
--- /dev/null
+++ b/paddler_bootstrap/src/registers_services.rs
@@ -0,0 +1,5 @@
+use trzcina::ServiceManager;
+
+pub trait RegistersServices {
+    fn register_services(self, service_manager: &mut ServiceManager);
+}
diff --git a/paddler_bootstrap/src/run_service_manager.rs b/paddler_bootstrap/src/run_service_manager.rs
index 9616ec329..6694f17d0 100644
--- a/paddler_bootstrap/src/run_service_manager.rs
+++ b/paddler_bootstrap/src/run_service_manager.rs
@@ -1,18 +1,19 @@
 use anyhow::Error;
 use anyhow::Result;
 use tokio_util::sync::CancellationToken;
-use trzcina::ServiceBundle;
 use trzcina::ServiceManager;
 use trzcina::ServiceShutdownOptions;
 
-pub async fn run_service_manager<TServiceBundle: ServiceBundle>(
-    bundle: TServiceBundle,
+use crate::registers_services::RegistersServices;
+
+pub async fn run_service_manager<TServices: RegistersServices>(
+    services: TServices,
     task_shutdown: CancellationToken,
     shutdown_options: ServiceShutdownOptions,
 ) -> Result<()> {
     let mut service_manager = ServiceManager::default();
 
-    service_manager.register_bundle(bundle).await?;
+    services.register_services(&mut service_manager);
     service_manager
         .start(task_shutdown)
         .run_to_completion(shutdown_options)
@@ -20,93 +21,3 @@ pub async fn run_service_manager<TServiceBundle: ServiceBundle>(
         .into_result()
         .map_err(Error::from)
 }
-
-#[cfg(test)]
-mod tests {
-    use anyhow::Result;
-    use async_trait::async_trait;
-    use thiserror::Error;
-    use tokio_util::sync::CancellationToken;
-    use trzcina::Service;
-    use trzcina::ServiceBundle;
-    use trzcina::ServiceShutdownError;
-    use trzcina::ServiceShutdownOptions;
-    use trzcina::ServiceShutdownOutcome;
-    use trzcina::ServiceShutdownOutcomeWithServiceName;
-
-    use super::run_service_manager;
-
-    #[derive(Debug, Error)]
-    #[error("the bundle could not produce its services")]
-    struct ServicesUnavailable;
-
-    #[derive(Debug, Error)]
-    #[error("the service could not run")]
-    struct ServiceRunFailed;
-
-    struct FailingServiceBundle;
-
-    #[async_trait]
-    impl ServiceBundle for FailingServiceBundle {
-        async fn services(self) -> Result<Vec<Box<dyn Service>>> {
-            Err(ServicesUnavailable.into())
-        }
-    }
-
-    struct FailingService;
-
-    #[async_trait]
-    impl Service for FailingService {
-        fn name(&self) -> &'static str {
-            "failing_service"
-        }
-
-        async fn run(self: Box<Self>, _shutdown: CancellationToken) -> Result<()> {
-            Err(ServiceRunFailed.into())
-        }
-    }
-
-    struct BundleWithFailingService;
-
-    #[async_trait]
-    impl ServiceBundle for BundleWithFailingService {
-        async fn services(self) -> Result<Vec<Box<dyn Service>>> {
-            Ok(vec![Box::new(FailingService)])
-        }
-    }
-
-    #[tokio::test]
-    async fn propagates_bundle_registration_error() {
-        let error = run_service_manager(
-            FailingServiceBundle,
-            CancellationToken::new(),
-            ServiceShutdownOptions::default(),
-        )
-        .await
-        .expect_err("a bundle without services must fail to start");
-
-        assert!(error.downcast_ref::<ServicesUnavailable>().is_some());
-    }
-
-    #[tokio::test]
-    async fn propagates_service_run_error() {
-        let error = run_service_manager(
-            BundleWithFailingService,
-            CancellationToken::new(),
-            ServiceShutdownOptions::default(),
-        )
-        .await
-        .expect_err("a failing service must fail the service manager");
-        let shutdown_error = error
-            .downcast_ref::<ServiceShutdownError>()
-            .expect("the service manager must report which services failed");
-
-        assert!(matches!(
-            shutdown_error.failed_outcomes(),
-            [ServiceShutdownOutcomeWithServiceName {
-                name: "failing_service",
-                outcome: ServiceShutdownOutcome::Errored(service_error),
-            }] if service_error.downcast_ref::<ServiceRunFailed>().is_some()
-        ));
-    }
-}
diff --git a/paddler_bootstrap/src/service_thread.rs b/paddler_bootstrap/src/service_thread.rs
index cdb77da19..8b1d1700e 100644
--- a/paddler_bootstrap/src/service_thread.rs
+++ b/paddler_bootstrap/src/service_thread.rs
@@ -39,7 +39,7 @@ impl ServiceThread {
         }
     }
 
-    pub async fn wait_for_completion(mut self) -> Result<(), BootstrapError> {
+    pub async fn wait_for_completion(&mut self) -> Result<(), BootstrapError> {
         let Err(_thread_finished) = (&mut self.finished_rx).await;
 
         self.join_thread()
@@ -78,12 +78,16 @@ mod tests {
     use std::sync::atomic::AtomicBool;
     use std::sync::atomic::Ordering;
 
-    use anyhow::Error;
+    use thiserror::Error;
     use tokio_util::sync::CancellationToken;
 
     use super::ServiceThread;
     use crate::bootstrap_error::BootstrapError;
 
+    #[derive(Debug, Error)]
+    #[error("the service could not run")]
+    struct ServiceRunFailed;
+
     #[tokio::test]
     async fn dropping_a_service_thread_stops_and_joins_it() {
         let run_finished = Arc::new(AtomicBool::new(false));
@@ -113,7 +117,7 @@ mod tests {
                 task_cancellation_token.cancelled().await;
                 run_finished_flag.store(true, Ordering::Release);
 
-                Err(Error::from(BootstrapError::StatsdReportingIntervalIsZero))
+                Err(ServiceRunFailed.into())
             },
         );
 
@@ -124,9 +128,9 @@ mod tests {
 
     #[tokio::test]
     async fn waiting_reports_the_error_of_a_failed_run() {
-        let service_thread =
+        let mut service_thread =
             ServiceThread::spawn(CancellationToken::new(), |_task_cancellation_token| async {
-                Err(Error::from(BootstrapError::StatsdReportingIntervalIsZero))
+                Err(ServiceRunFailed.into())
             });
 
         let completion_error = service_thread
@@ -137,14 +141,13 @@ mod tests {
         assert!(matches!(
             completion_error,
             BootstrapError::ServiceRunFailed { source }
-                if source.downcast_ref::<BootstrapError>().map(discriminant)
-                    == Some(discriminant(&BootstrapError::StatsdReportingIntervalIsZero))
+                if source.downcast_ref::<ServiceRunFailed>().is_some()
         ));
     }
 
     #[tokio::test]
     async fn wait_for_completion_errors_when_service_thread_panics() {
-        let service_thread =
+        let mut service_thread =
             ServiceThread::spawn(CancellationToken::new(), |_task_cancellation_token| async {
                 panic!("service thread crashed")
             });
@@ -159,7 +162,7 @@ mod tests {
 
     #[tokio::test]
     async fn cancel_stops_the_running_service_thread() {
-        let service_thread = ServiceThread::spawn(
+        let mut service_thread = ServiceThread::spawn(
             CancellationToken::new(),
             |task_cancellation_token| async move {
                 task_cancellation_token.cancelled().await;
diff --git a/paddler_bootstrap/tests/agent_runner_completes_after_cancel_when_management_server_is_unresponsive.rs b/paddler_bootstrap/tests/agent_runner_completes_after_cancel_when_management_server_is_unresponsive.rs
deleted file mode 100644
index f9dc0c639..000000000
--- a/paddler_bootstrap/tests/agent_runner_completes_after_cancel_when_management_server_is_unresponsive.rs
+++ /dev/null
@@ -1,39 +0,0 @@
-use tokio::net::TcpListener;
-use tokio_util::sync::CancellationToken;
-use trzcina::ServiceShutdownOptions;
-
-use paddler_bootstrap::agent_bootstrap_config::AgentBootstrapConfig;
-use paddler_bootstrap::agent_runner::AgentRunner;
-use paddler_bootstrap::agent_runner_params::AgentRunnerParams;
-use paddler_test_cluster_harness::ephemeral_loopback_addr::EPHEMERAL_LOOPBACK_ADDR;
-
-#[tokio::test]
-async fn agent_runner_completes_after_cancel_when_management_server_is_unresponsive() {
-    let unresponsive_management_server = TcpListener::bind(EPHEMERAL_LOOPBACK_ADDR)
-        .await
-        .expect("the unresponsive management server must bind an ephemeral port");
-    let management_addr = unresponsive_management_server
-        .local_addr()
-        .expect("a bound listener must report its address");
-
-    let runner = AgentRunner::start(AgentRunnerParams {
-        bootstrap_config: AgentBootstrapConfig {
-            agent_name: Some("test-agent".to_owned()),
-            management_address: management_addr.to_string(),
-            slots: 1,
-        },
-        cancellation_token: CancellationToken::new(),
-        shutdown_options: ServiceShutdownOptions::default(),
-    });
-
-    let (_held_connection, _peer_addr) = unresponsive_management_server
-        .accept()
-        .await
-        .expect("the agent must connect to the unresponsive management server");
-
-    runner.cancel();
-    runner
-        .wait_for_completion()
-        .await
-        .expect("a cancelled agent runner must complete cleanly");
-}
diff --git a/paddler_bootstrap/tests/agent_runner_completes_when_cancelled.rs b/paddler_bootstrap/tests/agent_runner_completes_when_cancelled.rs
index 221d5b4eb..a07c6e350 100644
--- a/paddler_bootstrap/tests/agent_runner_completes_when_cancelled.rs
+++ b/paddler_bootstrap/tests/agent_runner_completes_when_cancelled.rs
@@ -6,7 +6,7 @@ use crate::agent_runner_params_for_unreachable_balancer::agent_runner_params_for
 
 #[tokio::test]
 async fn agent_runner_completes_when_cancelled() {
-    let runner = AgentRunner::start(agent_runner_params_for_unreachable_balancer(
+    let mut runner = AgentRunner::start(agent_runner_params_for_unreachable_balancer(
         CancellationToken::new(),
     ));
 
diff --git a/paddler_bootstrap/tests/agent_runner_completes_when_its_parent_token_is_cancelled.rs b/paddler_bootstrap/tests/agent_runner_completes_when_its_parent_token_is_cancelled.rs
index 297d36dc1..b40521296 100644
--- a/paddler_bootstrap/tests/agent_runner_completes_when_its_parent_token_is_cancelled.rs
+++ b/paddler_bootstrap/tests/agent_runner_completes_when_its_parent_token_is_cancelled.rs
@@ -7,7 +7,7 @@ use crate::agent_runner_params_for_unreachable_balancer::agent_runner_params_for
 #[tokio::test]
 async fn agent_runner_completes_when_its_parent_token_is_cancelled() {
     let parent_token = CancellationToken::new();
-    let runner = AgentRunner::start(agent_runner_params_for_unreachable_balancer(
+    let mut runner = AgentRunner::start(agent_runner_params_for_unreachable_balancer(
         parent_token.clone(),
     ));
 
diff --git a/paddler_bootstrap/tests/agent_runner_leaves_its_balancer_when_dropped.rs b/paddler_bootstrap/tests/agent_runner_leaves_its_balancer_when_dropped.rs
new file mode 100644
index 000000000..ce9bb7cd3
--- /dev/null
+++ b/paddler_bootstrap/tests/agent_runner_leaves_its_balancer_when_dropped.rs
@@ -0,0 +1,45 @@
+use tokio_util::sync::CancellationToken;
+use url::Url;
+
+use paddler_bootstrap::agent_runner::AgentRunner;
+use paddler_bootstrap::balancer_runner::BalancerRunner;
+use paddler_client::client_management::ClientManagement;
+use paddler_test_cluster_harness::snapshots_watcher::SnapshotsWatcher;
+
+use crate::agent_runner_params_for_balancer_at::agent_runner_params_for_balancer_at;
+use crate::ephemeral_balancer_runner_params::ephemeral_balancer_runner_params;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agent_runner_leaves_its_balancer_when_dropped() {
+    let balancer_runner =
+        BalancerRunner::start(ephemeral_balancer_runner_params(CancellationToken::new()))
+            .await
+            .expect("a runner on ephemeral ports must start");
+    let mut agents_watcher = SnapshotsWatcher::of_agents(
+        ClientManagement::new(
+            Url::parse(&format!("http://{}", balancer_runner.addresses.management))
+                .expect("the management address must form a URL"),
+        )
+        .get_agents_stream(CancellationToken::new())
+        .await
+        .expect("the agents stream must open"),
+    )
+    .await
+    .expect("the agents stream must deliver its first snapshot");
+    let agent_runner = AgentRunner::start(agent_runner_params_for_balancer_at(
+        CancellationToken::new(),
+        balancer_runner.addresses.management.to_string(),
+    ));
+
+    agents_watcher
+        .until(|snapshot| snapshot.agents.len() == 1)
+        .await
+        .expect("the agent must register with its balancer");
+
+    drop(agent_runner);
+
+    agents_watcher
+        .until(|snapshot| snapshot.agents.is_empty())
+        .await
+        .expect("a dropped agent runner must leave its balancer");
+}
diff --git a/paddler_bootstrap/tests/agent_runner_params_for_balancer_at.rs b/paddler_bootstrap/tests/agent_runner_params_for_balancer_at.rs
new file mode 100644
index 000000000..04354ba93
--- /dev/null
+++ b/paddler_bootstrap/tests/agent_runner_params_for_balancer_at.rs
@@ -0,0 +1,26 @@
+use tokio_util::sync::CancellationToken;
+use trzcina::ServiceShutdownOptions;
+
+use paddler_bootstrap::agent_bootstrap_config::AgentBootstrapConfig;
+use paddler_bootstrap::agent_runner_params::AgentRunnerParams;
+use paddler_model_source::desired_model_resolver::DesiredModelResolver;
+use paddler_model_source::huggingface_default_endpoint::HUGGINGFACE_DEFAULT_ENDPOINT;
+
+pub fn agent_runner_params_for_balancer_at(
+    cancellation_token: CancellationToken,
+    management_address: String,
+) -> AgentRunnerParams {
+    AgentRunnerParams {
+        bootstrap_config: AgentBootstrapConfig {
+            agent_name: Some("test-agent".to_owned()),
+            desired_model_resolver: DesiredModelResolver::from_process_env(
+                HUGGINGFACE_DEFAULT_ENDPOINT.to_owned(),
+            )
+            .expect("the model downloader must start"),
+            management_address,
+            slots: 1,
+        },
+        cancellation_token,
+        shutdown_options: ServiceShutdownOptions::default(),
+    }
+}
diff --git a/paddler_bootstrap/tests/agent_runner_params_for_unreachable_balancer.rs b/paddler_bootstrap/tests/agent_runner_params_for_unreachable_balancer.rs
index 3ac5f465d..f09bc66eb 100644
--- a/paddler_bootstrap/tests/agent_runner_params_for_unreachable_balancer.rs
+++ b/paddler_bootstrap/tests/agent_runner_params_for_unreachable_balancer.rs
@@ -1,19 +1,11 @@
 use tokio_util::sync::CancellationToken;
-use trzcina::ServiceShutdownOptions;
 
-use paddler_bootstrap::agent_bootstrap_config::AgentBootstrapConfig;
 use paddler_bootstrap::agent_runner_params::AgentRunnerParams;
 
+use crate::agent_runner_params_for_balancer_at::agent_runner_params_for_balancer_at;
+
 pub fn agent_runner_params_for_unreachable_balancer(
     cancellation_token: CancellationToken,
 ) -> AgentRunnerParams {
-    AgentRunnerParams {
-        bootstrap_config: AgentBootstrapConfig {
-            agent_name: Some("test-agent".to_owned()),
-            management_address: "127.0.0.1:1".to_owned(),
-            slots: 1,
-        },
-        cancellation_token,
-        shutdown_options: ServiceShutdownOptions::default(),
-    }
+    agent_runner_params_for_balancer_at(cancellation_token, "127.0.0.1:1".to_owned())
 }
diff --git a/paddler_bootstrap/tests/agent_runner_reports_its_connection_to_the_balancer.rs b/paddler_bootstrap/tests/agent_runner_reports_its_connection_to_the_balancer.rs
deleted file mode 100644
index a5f89fc8f..000000000
--- a/paddler_bootstrap/tests/agent_runner_reports_its_connection_to_the_balancer.rs
+++ /dev/null
@@ -1,44 +0,0 @@
-use tokio_util::sync::CancellationToken;
-use trzcina::ServiceShutdownOptions;
-
-use paddler_bootstrap::agent_bootstrap_config::AgentBootstrapConfig;
-use paddler_bootstrap::agent_runner::AgentRunner;
-use paddler_bootstrap::agent_runner_params::AgentRunnerParams;
-use paddler_bootstrap::balancer_runner::BalancerRunner;
-use paddler_messaging::balancer_connection::BalancerConnection;
-
-use crate::ephemeral_balancer_runner_params::ephemeral_balancer_runner_params;
-
-#[tokio::test]
-async fn agent_runner_reports_its_connection_to_the_balancer() {
-    let balancer_runner =
-        BalancerRunner::start(ephemeral_balancer_runner_params(CancellationToken::new()))
-            .await
-            .expect("a runner on ephemeral ports must start");
-    let agent_runner = AgentRunner::start(AgentRunnerParams {
-        bootstrap_config: AgentBootstrapConfig {
-            agent_name: Some("test-agent".to_owned()),
-            management_address: balancer_runner.addresses.management.to_string(),
-            slots: 1,
-        },
-        cancellation_token: CancellationToken::new(),
-        shutdown_options: ServiceShutdownOptions::default(),
-    });
-    let mut balancer_connection_rx = agent_runner.balancer_connection_rx.clone();
-
-    balancer_connection_rx
-        .wait_for(|balancer_connection| *balancer_connection == BalancerConnection::Connected)
-        .await
-        .expect("the agent must report its connection to the balancer");
-
-    agent_runner.cancel();
-    agent_runner
-        .wait_for_completion()
-        .await
-        .expect("a cancelled agent runner must complete cleanly");
-    balancer_runner.cancel();
-    balancer_runner
-        .wait_for_completion()
-        .await
-        .expect("a cancelled balancer runner must complete cleanly");
-}
diff --git a/paddler_bootstrap/tests/agent_runner_stops_when_dropped.rs b/paddler_bootstrap/tests/agent_runner_stops_when_dropped.rs
deleted file mode 100644
index 38e617608..000000000
--- a/paddler_bootstrap/tests/agent_runner_stops_when_dropped.rs
+++ /dev/null
@@ -1,14 +0,0 @@
-use tokio_util::sync::CancellationToken;
-
-use paddler_bootstrap::agent_runner::AgentRunner;
-
-use crate::agent_runner_params_for_unreachable_balancer::agent_runner_params_for_unreachable_balancer;
-
-#[tokio::test]
-async fn agent_runner_stops_when_dropped() {
-    let runner = AgentRunner::start(agent_runner_params_for_unreachable_balancer(
-        CancellationToken::new(),
-    ));
-
-    drop(runner);
-}
diff --git a/paddler_bootstrap/tests/balancer_reports_its_gauges_to_statsd.rs b/paddler_bootstrap/tests/balancer_reports_its_gauges_to_statsd.rs
index ed6bad258..5ea5da33c 100644
--- a/paddler_bootstrap/tests/balancer_reports_its_gauges_to_statsd.rs
+++ b/paddler_bootstrap/tests/balancer_reports_its_gauges_to_statsd.rs
@@ -31,7 +31,7 @@ async fn balancer_reports_its_gauges_to_statsd() {
         statsd_reporting_interval: REPORTING_INTERVAL_LONGER_THAN_THE_TEST,
     });
 
-    let runner = BalancerRunner::start(params)
+    let mut runner = BalancerRunner::start(params)
         .await
         .expect("a runner reporting to statsd must start");
     let mut received_gauges = BTreeSet::new();
diff --git a/paddler_bootstrap/tests/balancer_runner_completes_when_cancelled.rs b/paddler_bootstrap/tests/balancer_runner_completes_when_cancelled.rs
index 44ca202fe..777a5bae1 100644
--- a/paddler_bootstrap/tests/balancer_runner_completes_when_cancelled.rs
+++ b/paddler_bootstrap/tests/balancer_runner_completes_when_cancelled.rs
@@ -6,9 +6,10 @@ use crate::ephemeral_balancer_runner_params::ephemeral_balancer_runner_params;
 
 #[tokio::test]
 async fn balancer_runner_completes_when_cancelled() {
-    let runner = BalancerRunner::start(ephemeral_balancer_runner_params(CancellationToken::new()))
-        .await
-        .expect("a runner on ephemeral ports must start");
+    let mut runner =
+        BalancerRunner::start(ephemeral_balancer_runner_params(CancellationToken::new()))
+            .await
+            .expect("a runner on ephemeral ports must start");
 
     runner.cancel();
 
diff --git a/paddler_bootstrap/tests/balancer_runner_completes_when_its_parent_token_is_cancelled.rs b/paddler_bootstrap/tests/balancer_runner_completes_when_its_parent_token_is_cancelled.rs
index 4f4f7ccca..2e713e61e 100644
--- a/paddler_bootstrap/tests/balancer_runner_completes_when_its_parent_token_is_cancelled.rs
+++ b/paddler_bootstrap/tests/balancer_runner_completes_when_its_parent_token_is_cancelled.rs
@@ -7,7 +7,7 @@ use crate::ephemeral_balancer_runner_params::ephemeral_balancer_runner_params;
 #[tokio::test]
 async fn balancer_runner_completes_when_its_parent_token_is_cancelled() {
     let parent_token = CancellationToken::new();
-    let runner = BalancerRunner::start(ephemeral_balancer_runner_params(parent_token.clone()))
+    let mut runner = BalancerRunner::start(ephemeral_balancer_runner_params(parent_token.clone()))
         .await
         .expect("a runner on ephemeral ports must start");
 
diff --git a/paddler_bootstrap/tests/balancer_runner_fails_to_start_when_an_address_is_taken.rs b/paddler_bootstrap/tests/balancer_runner_fails_to_start_when_an_address_is_taken.rs
deleted file mode 100644
index 047ea2d14..000000000
--- a/paddler_bootstrap/tests/balancer_runner_fails_to_start_when_an_address_is_taken.rs
+++ /dev/null
@@ -1,96 +0,0 @@
-use std::net::TcpListener;
-
-use tokio_util::sync::CancellationToken;
-
-use paddler_balancer::compatibility::openai_service::configuration::Configuration as OpenAIServiceConfiguration;
-use paddler_balancer::resolved_socket_addr::ResolvedSocketAddr;
-#[cfg(feature = "web_admin_panel")]
-use paddler_balancer::web_admin_panel_service::configuration::Configuration as WebAdminPanelServiceConfiguration;
-use paddler_bootstrap::balancer_runner::BalancerRunner;
-use paddler_bootstrap::balancer_runner_params::BalancerRunnerParams;
-use paddler_bootstrap::bootstrap_error::BootstrapError;
-use paddler_test_cluster_harness::ephemeral_loopback_addr::EPHEMERAL_LOOPBACK_ADDR;
-
-use crate::ephemeral_balancer_runner_params::ephemeral_balancer_runner_params;
-
-async fn start_error_of(params: BalancerRunnerParams) -> Option<BootstrapError> {
-    BalancerRunner::start(params).await.err()
-}
-
-#[tokio::test]
-async fn balancer_runner_fails_to_start_when_its_inference_address_is_taken() {
-    let occupying_listener =
-        TcpListener::bind(EPHEMERAL_LOOPBACK_ADDR).expect("an ephemeral loopback port must bind");
-    let taken_addr = occupying_listener
-        .local_addr()
-        .expect("a bound listener must report its address");
-    let mut params = ephemeral_balancer_runner_params(CancellationToken::new());
-
-    params.bootstrap_config.inference_service_configuration.addr =
-        ResolvedSocketAddr::from(taken_addr);
-
-    assert!(matches!(
-        start_error_of(params).await,
-        Some(BootstrapError::InferenceBindFailed { addr, .. }) if addr == taken_addr
-    ));
-}
-
-#[tokio::test]
-async fn balancer_runner_fails_to_start_when_its_management_address_is_taken() {
-    let occupying_listener =
-        TcpListener::bind(EPHEMERAL_LOOPBACK_ADDR).expect("an ephemeral loopback port must bind");
-    let taken_addr = occupying_listener
-        .local_addr()
-        .expect("a bound listener must report its address");
-    let mut params = ephemeral_balancer_runner_params(CancellationToken::new());
-
-    params
-        .bootstrap_config
-        .management_service_configuration
-        .addr = ResolvedSocketAddr::from(taken_addr);
-
-    assert!(matches!(
-        start_error_of(params).await,
-        Some(BootstrapError::ManagementBindFailed { addr, .. }) if addr == taken_addr
-    ));
-}
-
-#[tokio::test]
-async fn balancer_runner_fails_to_start_when_its_compat_openai_address_is_taken() {
-    let occupying_listener =
-        TcpListener::bind(EPHEMERAL_LOOPBACK_ADDR).expect("an ephemeral loopback port must bind");
-    let taken_addr = occupying_listener
-        .local_addr()
-        .expect("a bound listener must report its address");
-    let mut params = ephemeral_balancer_runner_params(CancellationToken::new());
-
-    params.bootstrap_config.openai_service_configuration = Some(OpenAIServiceConfiguration {
-        addr: ResolvedSocketAddr::from(taken_addr),
-    });
-
-    assert!(matches!(
-        start_error_of(params).await,
-        Some(BootstrapError::CompatOpenAIBindFailed { addr, .. }) if addr == taken_addr
-    ));
-}
-
-#[cfg(feature = "web_admin_panel")]
-#[tokio::test]
-async fn balancer_runner_fails_to_start_when_its_web_admin_panel_address_is_taken() {
-    let occupying_listener =
-        TcpListener::bind(EPHEMERAL_LOOPBACK_ADDR).expect("an ephemeral loopback port must bind");
-    let taken_addr = occupying_listener
-        .local_addr()
-        .expect("a bound listener must report its address");
-    let mut params = ephemeral_balancer_runner_params(CancellationToken::new());
-
-    params
-        .bootstrap_config
-        .web_admin_panel_service_configuration =
-        Some(WebAdminPanelServiceConfiguration { addr: taken_addr });
-
-    assert!(matches!(
-        start_error_of(params).await,
-        Some(BootstrapError::WebAdminPanelBindFailed { addr, .. }) if addr == taken_addr
-    ));
-}
diff --git a/paddler_bootstrap/tests/balancer_runner_fails_to_start_when_its_compat_openai_address_is_taken.rs b/paddler_bootstrap/tests/balancer_runner_fails_to_start_when_its_compat_openai_address_is_taken.rs
new file mode 100644
index 000000000..c718b7aef
--- /dev/null
+++ b/paddler_bootstrap/tests/balancer_runner_fails_to_start_when_its_compat_openai_address_is_taken.rs
@@ -0,0 +1,24 @@
+use std::io::ErrorKind;
+
+use paddler_balancer::compatibility::compatibility_service_configuration::CompatibilityServiceConfiguration;
+use paddler_balancer::resolved_socket_addr::ResolvedSocketAddr;
+use paddler_bootstrap::bootstrap_error::BootstrapError;
+
+use crate::balancer_runner_start_on_a_taken_address::BalancerRunnerStartOnATakenAddress;
+
+#[tokio::test]
+async fn balancer_runner_fails_to_start_when_its_compat_openai_address_is_taken() {
+    let start_attempt = BalancerRunnerStartOnATakenAddress::attempt(|params, taken_addr| {
+        params.bootstrap_config.openai_service_configuration =
+            Some(CompatibilityServiceConfiguration {
+                addr: ResolvedSocketAddr::from(taken_addr),
+            });
+    })
+    .await;
+
+    assert!(matches!(
+        start_attempt.start_error,
+        BootstrapError::CompatOpenAIBindFailed { addr, .. } if addr == start_attempt.taken_addr
+    ));
+    assert_eq!(start_attempt.bind_error_kind(), Some(ErrorKind::AddrInUse));
+}
diff --git a/paddler_bootstrap/tests/balancer_runner_fails_to_start_when_its_compat_typesafe_address_is_taken.rs b/paddler_bootstrap/tests/balancer_runner_fails_to_start_when_its_compat_typesafe_address_is_taken.rs
new file mode 100644
index 000000000..0186d7009
--- /dev/null
+++ b/paddler_bootstrap/tests/balancer_runner_fails_to_start_when_its_compat_typesafe_address_is_taken.rs
@@ -0,0 +1,29 @@
+use std::io::Error as IoError;
+use std::io::ErrorKind;
+
+use paddler_balancer::compatibility::compatibility_service_configuration::CompatibilityServiceConfiguration;
+use paddler_balancer::resolved_socket_addr::ResolvedSocketAddr;
+use paddler_bootstrap::bootstrap_error::BootstrapError;
+
+use crate::balancer_runner_start_on_a_taken_address::BalancerRunnerStartOnATakenAddress;
+
+#[tokio::test]
+async fn balancer_runner_fails_to_start_when_its_compat_typesafe_address_is_taken() {
+    let start_attempt = BalancerRunnerStartOnATakenAddress::attempt(|params, taken_addr| {
+        params.bootstrap_config.typesafe_service_configuration =
+            Some(CompatibilityServiceConfiguration {
+                addr: ResolvedSocketAddr::from(taken_addr),
+            });
+    })
+    .await;
+
+    assert_eq!(
+        start_attempt.start_error.to_string(),
+        BootstrapError::CompatTypeSafeBindFailed {
+            addr: start_attempt.taken_addr,
+            source: IoError::from(ErrorKind::AddrInUse),
+        }
+        .to_string()
+    );
+    assert_eq!(start_attempt.bind_error_kind(), Some(ErrorKind::AddrInUse));
+}
diff --git a/paddler_bootstrap/tests/balancer_runner_fails_to_start_when_its_inference_address_is_taken.rs b/paddler_bootstrap/tests/balancer_runner_fails_to_start_when_its_inference_address_is_taken.rs
new file mode 100644
index 000000000..6f3991af3
--- /dev/null
+++ b/paddler_bootstrap/tests/balancer_runner_fails_to_start_when_its_inference_address_is_taken.rs
@@ -0,0 +1,21 @@
+use std::io::ErrorKind;
+
+use paddler_balancer::resolved_socket_addr::ResolvedSocketAddr;
+use paddler_bootstrap::bootstrap_error::BootstrapError;
+
+use crate::balancer_runner_start_on_a_taken_address::BalancerRunnerStartOnATakenAddress;
+
+#[tokio::test]
+async fn balancer_runner_fails_to_start_when_its_inference_address_is_taken() {
+    let start_attempt = BalancerRunnerStartOnATakenAddress::attempt(|params, taken_addr| {
+        params.bootstrap_config.inference_service_configuration.addr =
+            ResolvedSocketAddr::from(taken_addr);
+    })
+    .await;
+
+    assert!(matches!(
+        start_attempt.start_error,
+        BootstrapError::InferenceBindFailed { addr, .. } if addr == start_attempt.taken_addr
+    ));
+    assert_eq!(start_attempt.bind_error_kind(), Some(ErrorKind::AddrInUse));
+}
diff --git a/paddler_bootstrap/tests/balancer_runner_fails_to_start_when_its_management_address_is_taken.rs b/paddler_bootstrap/tests/balancer_runner_fails_to_start_when_its_management_address_is_taken.rs
new file mode 100644
index 000000000..1a34cc52b
--- /dev/null
+++ b/paddler_bootstrap/tests/balancer_runner_fails_to_start_when_its_management_address_is_taken.rs
@@ -0,0 +1,23 @@
+use std::io::ErrorKind;
+
+use paddler_balancer::resolved_socket_addr::ResolvedSocketAddr;
+use paddler_bootstrap::bootstrap_error::BootstrapError;
+
+use crate::balancer_runner_start_on_a_taken_address::BalancerRunnerStartOnATakenAddress;
+
+#[tokio::test]
+async fn balancer_runner_fails_to_start_when_its_management_address_is_taken() {
+    let start_attempt = BalancerRunnerStartOnATakenAddress::attempt(|params, taken_addr| {
+        params
+            .bootstrap_config
+            .management_service_configuration
+            .addr = ResolvedSocketAddr::from(taken_addr);
+    })
+    .await;
+
+    assert!(matches!(
+        start_attempt.start_error,
+        BootstrapError::ManagementBindFailed { addr, .. } if addr == start_attempt.taken_addr
+    ));
+    assert_eq!(start_attempt.bind_error_kind(), Some(ErrorKind::AddrInUse));
+}
diff --git a/paddler_bootstrap/tests/balancer_runner_fails_to_start_when_its_web_admin_panel_address_is_taken.rs b/paddler_bootstrap/tests/balancer_runner_fails_to_start_when_its_web_admin_panel_address_is_taken.rs
new file mode 100644
index 000000000..ba65799a5
--- /dev/null
+++ b/paddler_bootstrap/tests/balancer_runner_fails_to_start_when_its_web_admin_panel_address_is_taken.rs
@@ -0,0 +1,23 @@
+use std::io::ErrorKind;
+
+use paddler_balancer::web_admin_panel_service::configuration::Configuration as WebAdminPanelServiceConfiguration;
+use paddler_bootstrap::bootstrap_error::BootstrapError;
+
+use crate::balancer_runner_start_on_a_taken_address::BalancerRunnerStartOnATakenAddress;
+
+#[tokio::test]
+async fn balancer_runner_fails_to_start_when_its_web_admin_panel_address_is_taken() {
+    let start_attempt = BalancerRunnerStartOnATakenAddress::attempt(|params, taken_addr| {
+        params
+            .bootstrap_config
+            .web_admin_panel_service_configuration =
+            Some(WebAdminPanelServiceConfiguration { addr: taken_addr });
+    })
+    .await;
+
+    assert!(matches!(
+        start_attempt.start_error,
+        BootstrapError::WebAdminPanelBindFailed { addr, .. } if addr == start_attempt.taken_addr
+    ));
+    assert_eq!(start_attempt.bind_error_kind(), Some(ErrorKind::AddrInUse));
+}
diff --git a/paddler_bootstrap/tests/balancer_runner_holds_the_desired_state_it_starts_with.rs b/paddler_bootstrap/tests/balancer_runner_holds_the_desired_state_it_starts_with.rs
new file mode 100644
index 000000000..437c15da5
--- /dev/null
+++ b/paddler_bootstrap/tests/balancer_runner_holds_the_desired_state_it_starts_with.rs
@@ -0,0 +1,43 @@
+use tokio_util::sync::CancellationToken;
+
+use paddler_bootstrap::balancer_runner::BalancerRunner;
+use paddler_messaging::agent_desired_model::AgentDesiredModel;
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_messaging::decision_settings::DecisionSettings;
+use paddler_state_database::state_database_type::StateDatabaseType;
+use paddler_state_database::stored_balancer_desired_state::StoredBalancerDesiredState;
+
+use crate::ephemeral_balancer_runner_params::ephemeral_balancer_runner_params;
+
+#[tokio::test]
+async fn balancer_runner_holds_the_desired_state_it_starts_with() {
+    let initial_desired_state = BalancerDesiredState {
+        decision: DecisionSettings {
+            pointer_head: AgentDesiredModel::Uri("initial-pointer-head".to_owned()),
+        },
+        model: AgentDesiredModel::Uri("initial-embedding-model".to_owned()),
+        ..BalancerDesiredState::default()
+    };
+    let mut params = ephemeral_balancer_runner_params(CancellationToken::new());
+
+    params.bootstrap_config.state_database_type = StateDatabaseType::Memory(
+        StoredBalancerDesiredState::Stored(Box::new(initial_desired_state.clone())),
+    );
+
+    let mut runner = BalancerRunner::start(params)
+        .await
+        .expect("a runner starting with a desired state must start");
+
+    assert_eq!(
+        runner
+            .balancer_applicable_state_holder
+            .get_balancer_desired_state(),
+        initial_desired_state
+    );
+
+    runner.cancel();
+    runner
+        .wait_for_completion()
+        .await
+        .expect("a cancelled runner must complete cleanly");
+}
diff --git a/paddler_bootstrap/tests/balancer_runner_preserves_the_persisted_desired_state.rs b/paddler_bootstrap/tests/balancer_runner_preserves_the_persisted_desired_state.rs
index 12291e17f..b9689b613 100644
--- a/paddler_bootstrap/tests/balancer_runner_preserves_the_persisted_desired_state.rs
+++ b/paddler_bootstrap/tests/balancer_runner_preserves_the_persisted_desired_state.rs
@@ -1,15 +1,15 @@
 use tempfile::TempDir;
-use tokio::sync::watch;
 use tokio_util::sync::CancellationToken;
 
 use paddler_bootstrap::balancer_runner::BalancerRunner;
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
 use paddler_messaging::agent_desired_model::AgentDesiredModel;
 use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_messaging::balancer_text_generation_settings::BalancerTextGenerationSettings;
 use paddler_messaging::chat_template::ChatTemplate;
-use paddler_state_database::file::File as StateDatabaseFile;
-use paddler_state_database::state_database::StateDatabase as _;
+use paddler_messaging::chat_template_source::ChatTemplateSource;
+use paddler_state_database::state_database_file::StateDatabaseFile;
 use paddler_state_database::state_database_type::StateDatabaseType;
+use paddler_state_database::stored_balancer_desired_state::StoredBalancerDesiredState;
 
 use crate::ephemeral_balancer_runner_params::ephemeral_balancer_runner_params;
 
@@ -19,18 +19,16 @@ async fn balancer_runner_preserves_the_persisted_desired_state() {
         TempDir::new().expect("a temporary state database directory must be creatable");
     let state_database_path = state_database_directory.path().join("state.json");
     let persisted_state = BalancerDesiredState {
-        chat_template_override: Some(ChatTemplate {
-            content: "persisted-chat-template".to_owned(),
-        }),
-        inference_parameters: InferenceParameters::default(),
-        model: AgentDesiredModel::LocalToAgent("persisted-model".to_owned()),
-        multimodal_projection: AgentDesiredModel::None,
-        use_chat_template_override: true,
+        model: AgentDesiredModel::Uri("persisted-model".to_owned()),
+        text_generation: BalancerTextGenerationSettings {
+            chat_template_source: ChatTemplateSource::Override(ChatTemplate {
+                content: "persisted-chat-template".to_owned(),
+            }),
+            ..BalancerTextGenerationSettings::default()
+        },
+        ..BalancerDesiredState::default()
     };
-    let (balancer_desired_state_tx, _balancer_desired_state_rx) =
-        watch::channel(BalancerDesiredState::default());
-    let state_database =
-        StateDatabaseFile::new(balancer_desired_state_tx, state_database_path.clone());
+    let state_database = StateDatabaseFile::new(state_database_path.clone());
 
     state_database
         .store_balancer_desired_state(&persisted_state)
@@ -41,11 +39,16 @@ async fn balancer_runner_preserves_the_persisted_desired_state() {
 
     params.bootstrap_config.state_database_type = StateDatabaseType::File(state_database_path);
 
-    let runner = BalancerRunner::start(params)
+    let mut runner = BalancerRunner::start(params)
         .await
         .expect("a runner with a valid state database must start");
 
-    assert_eq!(*runner.balancer_desired_state_tx.borrow(), persisted_state);
+    assert_eq!(
+        runner
+            .balancer_applicable_state_holder
+            .get_balancer_desired_state(),
+        persisted_state
+    );
 
     runner.cancel();
     runner
@@ -58,6 +61,6 @@ async fn balancer_runner_preserves_the_persisted_desired_state() {
             .read_balancer_desired_state()
             .await
             .expect("the state database must stay readable after the runner stops"),
-        persisted_state
+        StoredBalancerDesiredState::Stored(Box::new(persisted_state))
     );
 }
diff --git a/paddler_bootstrap/tests/balancer_runner_start_on_a_taken_address.rs b/paddler_bootstrap/tests/balancer_runner_start_on_a_taken_address.rs
new file mode 100644
index 000000000..66941d3e6
--- /dev/null
+++ b/paddler_bootstrap/tests/balancer_runner_start_on_a_taken_address.rs
@@ -0,0 +1,49 @@
+use std::error::Error;
+use std::io::Error as IoError;
+use std::io::ErrorKind;
+use std::net::SocketAddr;
+use std::net::TcpListener;
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_bootstrap::balancer_runner::BalancerRunner;
+use paddler_bootstrap::balancer_runner_params::BalancerRunnerParams;
+use paddler_bootstrap::bootstrap_error::BootstrapError;
+use paddler_test_cluster_harness::ephemeral_loopback_addr::EPHEMERAL_LOOPBACK_ADDR;
+
+use crate::ephemeral_balancer_runner_params::ephemeral_balancer_runner_params;
+
+pub struct BalancerRunnerStartOnATakenAddress {
+    pub start_error: BootstrapError,
+    pub taken_addr: SocketAddr,
+}
+
+impl BalancerRunnerStartOnATakenAddress {
+    pub async fn attempt(
+        serve_on_the_taken_addr: impl FnOnce(&mut BalancerRunnerParams, SocketAddr),
+    ) -> Self {
+        let occupying_listener = TcpListener::bind(EPHEMERAL_LOOPBACK_ADDR)
+            .expect("an ephemeral loopback port must bind");
+        let taken_addr = occupying_listener
+            .local_addr()
+            .expect("a bound listener must report its address");
+        let mut params = ephemeral_balancer_runner_params(CancellationToken::new());
+
+        serve_on_the_taken_addr(&mut params, taken_addr);
+
+        Self {
+            start_error: BalancerRunner::start(params)
+                .await
+                .err()
+                .expect("a runner must not start on a taken address"),
+            taken_addr,
+        }
+    }
+
+    pub fn bind_error_kind(&self) -> Option<ErrorKind> {
+        self.start_error
+            .source()
+            .and_then(|source| source.downcast_ref::<IoError>())
+            .map(IoError::kind)
+    }
+}
diff --git a/paddler_bootstrap/tests/balancer_shutdown_releases_buffered_request_with_shutdown_error.rs b/paddler_bootstrap/tests/balancer_shutdown_releases_buffered_request_with_shutdown_error.rs
index 1aa8951cc..a53247985 100644
--- a/paddler_bootstrap/tests/balancer_shutdown_releases_buffered_request_with_shutdown_error.rs
+++ b/paddler_bootstrap/tests/balancer_shutdown_releases_buffered_request_with_shutdown_error.rs
@@ -15,9 +15,10 @@ use crate::ephemeral_balancer_runner_params::ephemeral_balancer_runner_params;
 
 #[tokio::test]
 async fn balancer_shutdown_releases_buffered_request_with_shutdown_error() {
-    let runner = BalancerRunner::start(ephemeral_balancer_runner_params(CancellationToken::new()))
-        .await
-        .expect("a runner on ephemeral ports must start");
+    let mut runner =
+        BalancerRunner::start(ephemeral_balancer_runner_params(CancellationToken::new()))
+            .await
+            .expect("a runner on ephemeral ports must start");
     let held_response = Client::new()
         .post(format!(
             "http://{}{}",
diff --git a/paddler_bootstrap/tests/balancer_web_admin_panel_omits_the_services_it_does_not_run.rs b/paddler_bootstrap/tests/balancer_web_admin_panel_omits_the_services_it_does_not_run.rs
index 8ccf2b307..f4088d048 100644
--- a/paddler_bootstrap/tests/balancer_web_admin_panel_omits_the_services_it_does_not_run.rs
+++ b/paddler_bootstrap/tests/balancer_web_admin_panel_omits_the_services_it_does_not_run.rs
@@ -1,16 +1,13 @@
 use std::collections::BTreeMap;
 use std::time::Duration;
 
-use reqwest::get;
 use tokio_util::sync::CancellationToken;
 
 use paddler_balancer::resolved_socket_addr::ResolvedSocketAddr;
-use paddler_balancer::web_admin_panel_service::configuration::Configuration as WebAdminPanelServiceConfiguration;
-use paddler_bootstrap::balancer_runner::BalancerRunner;
-use paddler_test_cluster_harness::dashboard_attributes::dashboard_attributes;
 use paddler_test_cluster_harness::ephemeral_loopback_addr::EPHEMERAL_LOOPBACK_ADDR;
 
 use crate::ephemeral_balancer_runner_params::ephemeral_balancer_runner_params;
+use crate::rendered_dashboard_attributes::rendered_dashboard_attributes;
 
 #[tokio::test]
 async fn balancer_web_admin_panel_omits_the_services_it_does_not_run() {
@@ -29,36 +26,9 @@ async fn balancer_web_admin_panel_omits_the_services_it_does_not_run() {
         input_addr: "localhost:0".to_owned(),
         socket_addr: EPHEMERAL_LOOPBACK_ADDR,
     };
-    params
-        .bootstrap_config
-        .web_admin_panel_service_configuration = Some(WebAdminPanelServiceConfiguration {
-        addr: EPHEMERAL_LOOPBACK_ADDR,
-    });
-
-    let runner = BalancerRunner::start(params)
-        .await
-        .expect("a runner with the web admin panel must start");
-    let web_admin_panel_addr = runner
-        .addresses
-        .web_admin_panel
-        .expect("the runner must report the web admin panel address");
-    let dashboard = get(format!("http://{web_admin_panel_addr}/"))
-        .await
-        .expect("the web admin panel must accept connections")
-        .error_for_status()
-        .expect("the web admin panel must render the dashboard")
-        .text()
-        .await
-        .expect("the dashboard must have a body");
-
-    runner.cancel();
-    runner
-        .wait_for_completion()
-        .await
-        .expect("the runner must shut down cleanly");
 
     assert_eq!(
-        dashboard_attributes(&dashboard).expect("the dashboard must carry its configuration"),
+        rendered_dashboard_attributes(params).await,
         BTreeMap::from(
             [
                 ("data-buffered-request-timeout-millis", "2000"),
diff --git a/paddler_bootstrap/tests/balancer_web_admin_panel_shows_the_balancer_configuration.rs b/paddler_bootstrap/tests/balancer_web_admin_panel_shows_the_balancer_configuration.rs
index 72e2d9864..702ef8e61 100644
--- a/paddler_bootstrap/tests/balancer_web_admin_panel_shows_the_balancer_configuration.rs
+++ b/paddler_bootstrap/tests/balancer_web_admin_panel_shows_the_balancer_configuration.rs
@@ -1,18 +1,15 @@
 use std::collections::BTreeMap;
 use std::time::Duration;
 
-use reqwest::get;
 use tokio_util::sync::CancellationToken;
 
-use paddler_balancer::compatibility::openai_service::configuration::Configuration as OpenAIServiceConfiguration;
+use paddler_balancer::compatibility::compatibility_service_configuration::CompatibilityServiceConfiguration;
 use paddler_balancer::resolved_socket_addr::ResolvedSocketAddr;
 use paddler_balancer::statsd_service::configuration::Configuration as StatsdServiceConfiguration;
-use paddler_balancer::web_admin_panel_service::configuration::Configuration as WebAdminPanelServiceConfiguration;
-use paddler_bootstrap::balancer_runner::BalancerRunner;
-use paddler_test_cluster_harness::dashboard_attributes::dashboard_attributes;
 use paddler_test_cluster_harness::ephemeral_loopback_addr::EPHEMERAL_LOOPBACK_ADDR;
 
 use crate::ephemeral_balancer_runner_params::ephemeral_balancer_runner_params;
+use crate::rendered_dashboard_attributes::rendered_dashboard_attributes;
 
 fn entered_as(input_addr: &str) -> ResolvedSocketAddr {
     ResolvedSocketAddr {
@@ -32,48 +29,27 @@ async fn balancer_web_admin_panel_shows_the_balancer_configuration() {
         .management_service_configuration
         .addr = entered_as("[::ffff:127.0.0.1]:0");
     params.bootstrap_config.max_buffered_requests = 7;
-    params.bootstrap_config.openai_service_configuration = Some(OpenAIServiceConfiguration {
-        addr: entered_as("127.0.0.1:0"),
-    });
+    params.bootstrap_config.openai_service_configuration =
+        Some(CompatibilityServiceConfiguration {
+            addr: entered_as("127.0.0.1:0"),
+        });
     params.bootstrap_config.statsd_prefix = "dashboard_".to_owned();
     params.bootstrap_config.statsd_service_configuration = Some(StatsdServiceConfiguration {
         statsd_addr: entered_as("localhost:8125"),
         statsd_reporting_interval: Duration::from_hours(1),
     });
-    params
-        .bootstrap_config
-        .web_admin_panel_service_configuration = Some(WebAdminPanelServiceConfiguration {
-        addr: EPHEMERAL_LOOPBACK_ADDR,
-    });
-
-    let runner = BalancerRunner::start(params)
-        .await
-        .expect("a runner with the web admin panel must start");
-    let web_admin_panel_addr = runner
-        .addresses
-        .web_admin_panel
-        .expect("the runner must report the web admin panel address");
-    let dashboard = get(format!("http://{web_admin_panel_addr}/"))
-        .await
-        .expect("the web admin panel must accept connections")
-        .error_for_status()
-        .expect("the web admin panel must render the dashboard")
-        .text()
-        .await
-        .expect("the dashboard must have a body");
-
-    runner.cancel();
-    runner
-        .wait_for_completion()
-        .await
-        .expect("the runner must shut down cleanly");
+    params.bootstrap_config.typesafe_service_configuration =
+        Some(CompatibilityServiceConfiguration {
+            addr: entered_as("[::1]:0"),
+        });
 
     assert_eq!(
-        dashboard_attributes(&dashboard).expect("the dashboard must carry its configuration"),
+        rendered_dashboard_attributes(params).await,
         BTreeMap::from(
             [
                 ("data-buffered-request-timeout-millis", "1500"),
                 ("data-compat-openai-addr", "127.0.0.1:0"),
+                ("data-compat-typesafe-addr", "[::1]:0"),
                 ("data-inference-addr", "localhost:0"),
                 ("data-management-addr", "[::ffff:127.0.0.1]:0"),
                 ("data-max-buffered-requests", "7"),
diff --git a/paddler_bootstrap/tests/ephemeral_balancer_runner_params.rs b/paddler_bootstrap/tests/ephemeral_balancer_runner_params.rs
index c81fd4232..7e2c3ea08 100644
--- a/paddler_bootstrap/tests/ephemeral_balancer_runner_params.rs
+++ b/paddler_bootstrap/tests/ephemeral_balancer_runner_params.rs
@@ -6,6 +6,7 @@ use paddler_balancer::resolved_socket_addr::ResolvedSocketAddr;
 use paddler_bootstrap::balancer_bootstrap_config::BalancerBootstrapConfig;
 use paddler_bootstrap::balancer_runner_params::BalancerRunnerParams;
 use paddler_state_database::state_database_type::StateDatabaseType;
+use paddler_state_database::stored_balancer_desired_state::StoredBalancerDesiredState;
 use paddler_test_cluster_harness::ephemeral_loopback_addr::EPHEMERAL_LOOPBACK_ADDR;
 use paddler_test_cluster_harness::longer_than_any_test_run::LONGER_THAN_ANY_TEST_RUN;
 use paddler_test_cluster_harness::unexpiring_shutdown_options::unexpiring_shutdown_options;
@@ -27,9 +28,12 @@ pub fn ephemeral_balancer_runner_params(
             },
             max_buffered_requests: 30,
             openai_service_configuration: None,
-            state_database_type: StateDatabaseType::Memory(Box::default()),
+            state_database_type: StateDatabaseType::Memory(
+                StoredBalancerDesiredState::NothingStored,
+            ),
             statsd_prefix: "paddler_bootstrap_test_".to_owned(),
             statsd_service_configuration: None,
+            typesafe_service_configuration: None,
             #[cfg(feature = "web_admin_panel")]
             web_admin_panel_service_configuration: None,
         },
diff --git a/paddler_bootstrap/tests/main.rs b/paddler_bootstrap/tests/main.rs
index 7caa3bbeb..578855d2d 100644
--- a/paddler_bootstrap/tests/main.rs
+++ b/paddler_bootstrap/tests/main.rs
@@ -1,21 +1,34 @@
-mod agent_runner_completes_after_cancel_when_management_server_is_unresponsive;
 mod agent_runner_completes_when_cancelled;
 mod agent_runner_completes_when_its_parent_token_is_cancelled;
+mod agent_runner_leaves_its_balancer_when_dropped;
+#[cfg(test)]
+mod agent_runner_params_for_balancer_at;
+#[cfg(test)]
 mod agent_runner_params_for_unreachable_balancer;
-mod agent_runner_reports_its_connection_to_the_balancer;
-mod agent_runner_stops_when_dropped;
 mod balancer_reports_its_gauges_to_statsd;
 mod balancer_runner_completes_when_cancelled;
 mod balancer_runner_completes_when_its_parent_token_is_cancelled;
-mod balancer_runner_fails_to_start_when_an_address_is_taken;
+mod balancer_runner_fails_to_start_when_its_compat_openai_address_is_taken;
+mod balancer_runner_fails_to_start_when_its_compat_typesafe_address_is_taken;
+mod balancer_runner_fails_to_start_when_its_inference_address_is_taken;
+mod balancer_runner_fails_to_start_when_its_management_address_is_taken;
+#[cfg(feature = "web_admin_panel")]
+mod balancer_runner_fails_to_start_when_its_web_admin_panel_address_is_taken;
 mod balancer_runner_fails_to_start_when_state_database_file_is_corrupt;
 mod balancer_runner_fails_to_start_with_a_zero_statsd_reporting_interval;
+mod balancer_runner_holds_the_desired_state_it_starts_with;
 mod balancer_runner_preserves_the_persisted_desired_state;
 mod balancer_runner_releases_its_ports_when_dropped;
 mod balancer_runner_serves_on_the_addresses_it_reports;
+#[cfg(test)]
+mod balancer_runner_start_on_a_taken_address;
 mod balancer_shutdown_releases_buffered_request_with_shutdown_error;
 #[cfg(feature = "web_admin_panel")]
 mod balancer_web_admin_panel_omits_the_services_it_does_not_run;
 #[cfg(feature = "web_admin_panel")]
 mod balancer_web_admin_panel_shows_the_balancer_configuration;
+#[cfg(test)]
 mod ephemeral_balancer_runner_params;
+#[cfg(feature = "web_admin_panel")]
+#[cfg(test)]
+mod rendered_dashboard_attributes;
diff --git a/paddler_bootstrap/tests/rendered_dashboard_attributes.rs b/paddler_bootstrap/tests/rendered_dashboard_attributes.rs
new file mode 100644
index 000000000..e608a1cf3
--- /dev/null
+++ b/paddler_bootstrap/tests/rendered_dashboard_attributes.rs
@@ -0,0 +1,43 @@
+use std::collections::BTreeMap;
+
+use reqwest::get;
+
+use paddler_balancer::web_admin_panel_service::configuration::Configuration as WebAdminPanelServiceConfiguration;
+use paddler_bootstrap::balancer_runner::BalancerRunner;
+use paddler_bootstrap::balancer_runner_params::BalancerRunnerParams;
+use paddler_test_cluster_harness::dashboard_attributes::dashboard_attributes;
+use paddler_test_cluster_harness::ephemeral_loopback_addr::EPHEMERAL_LOOPBACK_ADDR;
+
+pub async fn rendered_dashboard_attributes(
+    mut params: BalancerRunnerParams,
+) -> BTreeMap<String, String> {
+    params
+        .bootstrap_config
+        .web_admin_panel_service_configuration = Some(WebAdminPanelServiceConfiguration {
+        addr: EPHEMERAL_LOOPBACK_ADDR,
+    });
+
+    let mut runner = BalancerRunner::start(params)
+        .await
+        .expect("a runner with the web admin panel must start");
+    let web_admin_panel_addr = runner
+        .addresses
+        .web_admin_panel
+        .expect("the runner must report the web admin panel address");
+    let dashboard = get(format!("http://{web_admin_panel_addr}/"))
+        .await
+        .expect("the web admin panel must accept connections")
+        .error_for_status()
+        .expect("the web admin panel must render the dashboard")
+        .text()
+        .await
+        .expect("the dashboard must have a body");
+
+    runner.cancel();
+    runner
+        .wait_for_completion()
+        .await
+        .expect("the runner must shut down cleanly");
+
+    dashboard_attributes(&dashboard).expect("the dashboard must carry its configuration")
+}
diff --git a/paddler_cache_dir/Cargo.toml b/paddler_cache_dir/Cargo.toml
index 541efa843..92dd943a8 100644
--- a/paddler_cache_dir/Cargo.toml
+++ b/paddler_cache_dir/Cargo.toml
@@ -10,11 +10,12 @@ version.workspace = true
 
 [dependencies]
 fslock = { workspace = true }
-sha2 = { workspace = true }
 thiserror = { workspace = true }
 tokio = { workspace = true }
 tokio-util = { workspace = true }
 
+paddler_sha256_hex = { workspace = true }
+
 [dev-dependencies]
 tempfile = { workspace = true }
 
diff --git a/paddler_cache_dir/src/cache_dir.rs b/paddler_cache_dir/src/cache_dir.rs
new file mode 100644
index 000000000..a0456c691
--- /dev/null
+++ b/paddler_cache_dir/src/cache_dir.rs
@@ -0,0 +1,96 @@
+use std::env::var_os;
+use std::path::PathBuf;
+
+use crate::cache_dir_error::CacheDirError;
+
+#[derive(Clone, Debug, Eq, PartialEq)]
+pub struct CacheDir {
+    pub explicit: Option<PathBuf>,
+    pub home: Option<PathBuf>,
+    pub xdg: Option<PathBuf>,
+}
+
+impl CacheDir {
+    #[must_use]
+    pub fn from_process_env() -> Self {
+        Self {
+            explicit: var_os("PADDLER_CACHE_DIR").map(PathBuf::from),
+            home: var_os("HOME").map(PathBuf::from),
+            xdg: var_os("XDG_CACHE_HOME").map(PathBuf::from),
+        }
+    }
+
+    pub fn resolve(&self) -> Result<PathBuf, CacheDirError> {
+        if let Some(explicit) = &self.explicit {
+            return Ok(explicit.clone());
+        }
+
+        if let Some(xdg) = &self.xdg {
+            return Ok(xdg.join("paddler"));
+        }
+
+        self.home
+            .as_ref()
+            .map(|home| home.join(".cache").join("paddler"))
+            .ok_or(CacheDirError::HomeVariableUnset)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::mem::discriminant;
+    use std::path::PathBuf;
+
+    use super::CacheDir;
+    use crate::cache_dir_error::CacheDirError;
+
+    #[test]
+    fn explicit_value_wins_over_xdg_and_home() {
+        let cache = CacheDir {
+            explicit: Some(PathBuf::from("/explicit/cache")),
+            home: Some(PathBuf::from("/home/user")),
+            xdg: Some(PathBuf::from("/xdg/cache")),
+        };
+        let path = cache.resolve().unwrap();
+
+        assert_eq!(path, PathBuf::from("/explicit/cache"));
+    }
+
+    #[test]
+    fn xdg_value_used_when_no_explicit() {
+        let cache = CacheDir {
+            explicit: None,
+            home: Some(PathBuf::from("/home/user")),
+            xdg: Some(PathBuf::from("/xdg/cache")),
+        };
+        let path = cache.resolve().unwrap();
+
+        assert_eq!(path, PathBuf::from("/xdg/cache/paddler"));
+    }
+
+    #[test]
+    fn falls_back_to_home_dot_cache_paddler() {
+        let cache = CacheDir {
+            explicit: None,
+            home: Some(PathBuf::from("/home/user")),
+            xdg: None,
+        };
+        let path = cache.resolve().unwrap();
+
+        assert_eq!(path, PathBuf::from("/home/user/.cache/paddler"));
+    }
+
+    #[test]
+    fn errors_when_no_env_set() {
+        let cache = CacheDir {
+            explicit: None,
+            home: None,
+            xdg: None,
+        };
+
+        assert_eq!(
+            cache.resolve().err().as_ref().map(discriminant),
+            Some(discriminant(&CacheDirError::HomeVariableUnset))
+        );
+    }
+}
diff --git a/paddler_cache_dir/src/cache_dir/unix.rs b/paddler_cache_dir/src/cache_dir/unix.rs
deleted file mode 100644
index 47960b0a1..000000000
--- a/paddler_cache_dir/src/cache_dir/unix.rs
+++ /dev/null
@@ -1,94 +0,0 @@
-use std::env::var_os;
-use std::path::PathBuf;
-
-use crate::cache_dir_error::CacheDirError;
-
-pub struct CacheDir {
-    pub explicit: Option<PathBuf>,
-    pub home: Option<PathBuf>,
-    pub xdg: Option<PathBuf>,
-}
-
-impl CacheDir {
-    #[must_use]
-    pub fn from_process_env() -> Self {
-        Self {
-            explicit: var_os("PADDLER_CACHE_DIR").map(PathBuf::from),
-            home: var_os("HOME").map(PathBuf::from),
-            xdg: var_os("XDG_CACHE_HOME").map(PathBuf::from),
-        }
-    }
-
-    pub fn resolve(&self) -> Result<PathBuf, CacheDirError> {
-        if let Some(explicit) = &self.explicit {
-            return Ok(explicit.clone());
-        }
-
-        if let Some(xdg) = &self.xdg {
-            return Ok(xdg.join("paddler"));
-        }
-
-        self.home
-            .as_ref()
-            .map(|home| home.join(".cache").join("paddler"))
-            .ok_or(CacheDirError::HomeVariableUnset { variable: "HOME" })
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use std::path::PathBuf;
-
-    use super::CacheDir;
-    use crate::cache_dir_error::CacheDirError;
-
-    #[test]
-    fn explicit_value_wins_over_xdg_and_home() {
-        let cache = CacheDir {
-            explicit: Some(PathBuf::from("/explicit/cache")),
-            home: Some(PathBuf::from("/home/user")),
-            xdg: Some(PathBuf::from("/xdg/cache")),
-        };
-        let path = cache.resolve().unwrap();
-
-        assert_eq!(path, PathBuf::from("/explicit/cache"));
-    }
-
-    #[test]
-    fn xdg_value_used_when_no_explicit() {
-        let cache = CacheDir {
-            explicit: None,
-            home: Some(PathBuf::from("/home/user")),
-            xdg: Some(PathBuf::from("/xdg/cache")),
-        };
-        let path = cache.resolve().unwrap();
-
-        assert_eq!(path, PathBuf::from("/xdg/cache/paddler"));
-    }
-
-    #[test]
-    fn falls_back_to_home_dot_cache_paddler() {
-        let cache = CacheDir {
-            explicit: None,
-            home: Some(PathBuf::from("/home/user")),
-            xdg: None,
-        };
-        let path = cache.resolve().unwrap();
-
-        assert_eq!(path, PathBuf::from("/home/user/.cache/paddler"));
-    }
-
-    #[test]
-    fn errors_when_no_env_set() {
-        let cache = CacheDir {
-            explicit: None,
-            home: None,
-            xdg: None,
-        };
-
-        assert!(matches!(
-            cache.resolve(),
-            Err(CacheDirError::HomeVariableUnset { variable }) if variable == "HOME"
-        ));
-    }
-}
diff --git a/paddler_cache_dir/src/cache_dir/windows.rs b/paddler_cache_dir/src/cache_dir/windows.rs
deleted file mode 100644
index 79ae50085..000000000
--- a/paddler_cache_dir/src/cache_dir/windows.rs
+++ /dev/null
@@ -1,96 +0,0 @@
-use std::env::var_os;
-use std::path::PathBuf;
-
-use crate::cache_dir_error::CacheDirError;
-
-pub struct CacheDir {
-    pub explicit: Option<PathBuf>,
-    pub localappdata: Option<PathBuf>,
-    pub userprofile: Option<PathBuf>,
-}
-
-impl CacheDir {
-    #[must_use]
-    pub fn from_process_env() -> Self {
-        Self {
-            explicit: var_os("PADDLER_CACHE_DIR").map(PathBuf::from),
-            localappdata: var_os("LOCALAPPDATA").map(PathBuf::from),
-            userprofile: var_os("USERPROFILE").map(PathBuf::from),
-        }
-    }
-
-    pub fn resolve(&self) -> Result<PathBuf, CacheDirError> {
-        if let Some(explicit) = &self.explicit {
-            return Ok(explicit.clone());
-        }
-
-        if let Some(localappdata) = &self.localappdata {
-            return Ok(localappdata.join("paddler"));
-        }
-
-        self.userprofile
-            .as_ref()
-            .map(|userprofile| userprofile.join("AppData").join("Local").join("paddler"))
-            .ok_or(CacheDirError::HomeVariableUnset {
-                variable: "USERPROFILE",
-            })
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use std::path::PathBuf;
-
-    use super::CacheDir;
-    use crate::cache_dir_error::CacheDirError;
-
-    #[test]
-    fn explicit_value_wins_over_localappdata_and_userprofile() {
-        let cache = CacheDir {
-            explicit: Some(PathBuf::from(r"D:\explicit\cache")),
-            localappdata: Some(PathBuf::from(r"C:\Users\user\AppData\Local")),
-            userprofile: Some(PathBuf::from(r"C:\Users\user")),
-        };
-        let path = cache.resolve().unwrap();
-
-        assert_eq!(path, PathBuf::from(r"D:\explicit\cache"));
-    }
-
-    #[test]
-    fn localappdata_used_when_no_explicit() {
-        let cache = CacheDir {
-            explicit: None,
-            localappdata: Some(PathBuf::from(r"C:\Users\user\AppData\Local")),
-            userprofile: Some(PathBuf::from(r"C:\Users\user")),
-        };
-        let path = cache.resolve().unwrap();
-
-        assert_eq!(path, PathBuf::from(r"C:\Users\user\AppData\Local\paddler"));
-    }
-
-    #[test]
-    fn falls_back_to_userprofile_appdata_local_paddler() {
-        let cache = CacheDir {
-            explicit: None,
-            localappdata: None,
-            userprofile: Some(PathBuf::from(r"C:\Users\user")),
-        };
-        let path = cache.resolve().unwrap();
-
-        assert_eq!(path, PathBuf::from(r"C:\Users\user\AppData\Local\paddler"));
-    }
-
-    #[test]
-    fn errors_when_no_env_set() {
-        let cache = CacheDir {
-            explicit: None,
-            localappdata: None,
-            userprofile: None,
-        };
-
-        assert!(matches!(
-            cache.resolve(),
-            Err(CacheDirError::HomeVariableUnset { variable }) if variable == "USERPROFILE"
-        ));
-    }
-}
diff --git a/paddler_cache_dir/src/cache_dir_error.rs b/paddler_cache_dir/src/cache_dir_error.rs
index 3aa447af7..46a38bc6d 100644
--- a/paddler_cache_dir/src/cache_dir_error.rs
+++ b/paddler_cache_dir/src/cache_dir_error.rs
@@ -2,6 +2,6 @@ use thiserror::Error;
 
 #[derive(Debug, Error)]
 pub enum CacheDirError {
-    #[error("{variable} is not set; cannot derive the paddler cache directory")]
-    HomeVariableUnset { variable: &'static str },
+    #[error("HOME is not set; cannot derive the paddler cache directory")]
+    HomeVariableUnset,
 }
diff --git a/paddler_cache_dir/src/cached_downloaded_model.rs b/paddler_cache_dir/src/cached_downloaded_model.rs
index 4ab420c29..4c5a5e20f 100644
--- a/paddler_cache_dir/src/cached_downloaded_model.rs
+++ b/paddler_cache_dir/src/cached_downloaded_model.rs
@@ -2,11 +2,11 @@ use std::io::Error;
 use std::path::PathBuf;
 
 use fslock::LockFile;
-use sha2::Digest;
-use sha2::Sha256;
 use tokio::fs::create_dir_all;
 use tokio::fs::try_exists;
 
+use paddler_sha256_hex::sha256_lowercase_hex::sha256_lowercase_hex;
+
 use crate::cache_dir::CacheDir;
 use crate::cache_dir_error::CacheDirError;
 use crate::cached_downloaded_model_lock::CachedDownloadedModelLock;
@@ -23,7 +23,7 @@ pub struct CachedDownloadedModel {
 impl CachedDownloadedModel {
     pub fn new(cache_dir: &CacheDir, url_string: &str) -> Result<Self, CacheDirError> {
         cache_dir.resolve().map(|cache_root| {
-            let basename = format!("{:x}", Sha256::digest(url_string.as_bytes()));
+            let basename = sha256_lowercase_hex(url_string.as_bytes());
             let cache_subdir = cache_root.join(DOWNLOADED_MODELS_SUBDIR);
 
             Self {
@@ -72,21 +72,10 @@ mod tests {
     use crate::download_lock_acquisition::DownloadLockAcquisition;
 
     fn cache_dir_at(path: &Path) -> CacheDir {
-        #[cfg(unix)]
-        {
-            CacheDir {
-                explicit: Some(path.to_path_buf()),
-                home: None,
-                xdg: None,
-            }
-        }
-        #[cfg(windows)]
-        {
-            CacheDir {
-                explicit: Some(path.to_path_buf()),
-                localappdata: None,
-                userprofile: None,
-            }
+        CacheDir {
+            explicit: Some(path.to_path_buf()),
+            home: None,
+            xdg: None,
         }
     }
 
@@ -222,31 +211,17 @@ mod tests {
 
     #[test]
     fn new_returns_error_when_cache_dir_cannot_resolve() {
-        let unresolvable;
-        #[cfg(unix)]
-        {
-            unresolvable = CacheDir {
-                explicit: None,
-                home: None,
-                xdg: None,
-            };
-        }
-        #[cfg(windows)]
-        {
-            unresolvable = CacheDir {
-                explicit: None,
-                localappdata: None,
-                userprofile: None,
-            };
-        }
+        let unresolvable = CacheDir {
+            explicit: None,
+            home: None,
+            xdg: None,
+        };
 
         let result = CachedDownloadedModel::new(&unresolvable, "https://host.example/m.gguf");
 
         assert_eq!(
             result.err().as_ref().map(discriminant),
-            Some(discriminant(&CacheDirError::HomeVariableUnset {
-                variable: "HOME"
-            }))
+            Some(discriminant(&CacheDirError::HomeVariableUnset))
         );
     }
 
diff --git a/paddler_cache_dir/src/lib.rs b/paddler_cache_dir/src/lib.rs
index 25b84c2f7..51cb1bc6c 100644
--- a/paddler_cache_dir/src/lib.rs
+++ b/paddler_cache_dir/src/lib.rs
@@ -1,8 +1,3 @@
-#[cfg(unix)]
-#[path = "cache_dir/unix.rs"]
-pub mod cache_dir;
-#[cfg(windows)]
-#[path = "cache_dir/windows.rs"]
 pub mod cache_dir;
 pub mod cache_dir_error;
 pub mod cached_downloaded_model;
diff --git a/paddler_cli/Cargo.toml b/paddler_cli/Cargo.toml
index 8dc01c0ac..acec01d4b 100644
--- a/paddler_cli/Cargo.toml
+++ b/paddler_cli/Cargo.toml
@@ -18,7 +18,6 @@ name = "integration"
 path = "tests/main.rs"
 
 [dependencies]
-actix-rt = { workspace = true }
 actix-web = { workspace = true }
 anyhow = { workspace = true }
 async-trait = { workspace = true }
@@ -27,6 +26,7 @@ command_handler = { workspace = true }
 env_logger = { workspace = true }
 paddler_balancer = { workspace = true }
 paddler_bootstrap = { workspace = true }
+paddler_model_source = { workspace = true }
 paddler_state_database = { workspace = true }
 tokio-util = { workspace = true }
 trzcina = { workspace = true }
diff --git a/paddler_cli/src/cmd/agent.rs b/paddler_cli/src/cmd/agent.rs
index 3d6dd5109..b19570443 100644
--- a/paddler_cli/src/cmd/agent.rs
+++ b/paddler_cli/src/cmd/agent.rs
@@ -1,3 +1,4 @@
+use anyhow::Context as _;
 use anyhow::Result;
 use async_trait::async_trait;
 use clap::Parser;
@@ -11,11 +12,17 @@ use paddler_bootstrap::agent_bootstrap_config::AgentBootstrapConfig;
 use paddler_bootstrap::agent_service_bundle::AgentServiceBundle;
 use paddler_bootstrap::llama_cpp_max_sequences::LLAMA_CPP_MAX_SEQUENCES;
 use paddler_bootstrap::run_service_manager::run_service_manager;
+use paddler_model_source::desired_model_resolver::DesiredModelResolver;
+use paddler_model_source::huggingface_default_endpoint::HUGGINGFACE_DEFAULT_ENDPOINT;
 
 use super::value_parser::parse_socket_addr::parse_socket_addr;
 
 #[derive(Parser)]
 pub struct Agent {
+    #[arg(long, env = "HF_ENDPOINT", default_value = HUGGINGFACE_DEFAULT_ENDPOINT)]
+    /// Hugging Face Hub endpoint that the agent downloads models from
+    huggingface_endpoint: String,
+
     #[arg(long, value_parser = parse_socket_addr)]
     /// Address of the management server that the agent will connect to
     management_addr: ResolvedSocketAddr,
@@ -29,10 +36,10 @@ pub struct Agent {
     slots: u16,
 }
 
-#[async_trait(?Send)]
-impl Handler for Agent {
-    async fn handle(self, shutdown: CancellationToken) -> Result<()> {
+impl Agent {
+    async fn run_agent(self, shutdown: CancellationToken) -> Result<()> {
         let Self {
+            huggingface_endpoint,
             management_addr,
             name,
             slots,
@@ -41,6 +48,10 @@ impl Handler for Agent {
         run_service_manager(
             AgentServiceBundle::new(AgentBootstrapConfig {
                 agent_name: name,
+                desired_model_resolver: DesiredModelResolver::from_process_env(
+                    huggingface_endpoint,
+                )
+                .context("Unable to set up the model downloader")?,
                 management_address: management_addr.socket_addr.to_string(),
                 slots,
             }),
@@ -51,6 +62,13 @@ impl Handler for Agent {
     }
 }
 
+#[async_trait(?Send)]
+impl Handler for Agent {
+    async fn handle(self, shutdown: CancellationToken) -> Result<()> {
+        self.run_agent(shutdown).await
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use clap::Parser as _;
diff --git a/paddler_cli/src/cmd/balancer.rs b/paddler_cli/src/cmd/balancer.rs
index e590675c9..e1cc9f828 100644
--- a/paddler_cli/src/cmd/balancer.rs
+++ b/paddler_cli/src/cmd/balancer.rs
@@ -10,7 +10,7 @@ use command_handler::handler::Handler;
 use tokio_util::sync::CancellationToken;
 use trzcina::ServiceShutdownOptions;
 
-use paddler_balancer::compatibility::openai_service::configuration::Configuration as OpenAIServiceConfiguration;
+use paddler_balancer::compatibility::compatibility_service_configuration::CompatibilityServiceConfiguration;
 use paddler_balancer::inference_service::configuration::Configuration as InferenceServiceConfiguration;
 use paddler_balancer::management_service::configuration::Configuration as ManagementServiceConfiguration;
 use paddler_balancer::resolved_socket_addr::ResolvedSocketAddr;
@@ -41,6 +41,10 @@ pub struct Balancer {
     /// Address of the OpenAI-compatible API server (enabled only if this address is specified)
     compat_openai_addr: Option<ResolvedSocketAddr>,
 
+    #[arg(long, value_parser = parse_socket_addr)]
+    /// Address of the TypeSafe-compatible API server (enabled only if this address is specified)
+    compat_typesafe_addr: Option<ResolvedSocketAddr>,
+
     #[arg(
         long,
         default_value = SocketAddr::from((Ipv4Addr::LOCALHOST, BalancerDefaults::INFERENCE_PORT)).to_string(),
@@ -85,7 +89,7 @@ pub struct Balancer {
     max_buffered_requests: u64,
 
     #[arg(long, default_value = "memory://")]
-    /// Balancer state database URL. Supported: memory, memory://, or <file:///path> (optional)
+    /// Balancer state database URL. Supported: memory:// or <file:///absolute/path> (optional)
     state_database: StateDatabaseType,
 
     #[arg(long, value_parser = parse_socket_addr)]
@@ -110,12 +114,12 @@ pub struct Balancer {
     web_admin_panel_addr: Option<ResolvedSocketAddr>,
 }
 
-#[async_trait(?Send)]
-impl Handler for Balancer {
-    async fn handle(self, shutdown: CancellationToken) -> Result<()> {
+impl Balancer {
+    async fn run_balancer(self, shutdown: CancellationToken) -> Result<()> {
         let Self {
             buffered_request_timeout,
             compat_openai_addr,
+            compat_typesafe_addr,
             inference_addr,
             inference_item_timeout,
             inference_cors_allowed_hosts,
@@ -143,7 +147,7 @@ impl Handler for Balancer {
             },
             max_buffered_requests,
             openai_service_configuration: compat_openai_addr
-                .map(|addr| OpenAIServiceConfiguration { addr }),
+                .map(|addr| CompatibilityServiceConfiguration { addr }),
             state_database_type: state_database,
             statsd_prefix,
             statsd_service_configuration: statsd_addr.map(|statsd_addr| {
@@ -152,6 +156,8 @@ impl Handler for Balancer {
                     statsd_reporting_interval,
                 }
             }),
+            typesafe_service_configuration: compat_typesafe_addr
+                .map(|addr| CompatibilityServiceConfiguration { addr }),
             #[cfg(feature = "web_admin_panel")]
             web_admin_panel_service_configuration: web_admin_panel_addr.map(
                 |web_admin_panel_addr| WebAdminPanelServiceConfiguration {
@@ -166,3 +172,10 @@ impl Handler for Balancer {
         run_service_manager(bundle, shutdown, ServiceShutdownOptions::default()).await
     }
 }
+
+#[async_trait(?Send)]
+impl Handler for Balancer {
+    async fn handle(self, shutdown: CancellationToken) -> Result<()> {
+        self.run_balancer(shutdown).await
+    }
+}
diff --git a/paddler_cli/src/lib.rs b/paddler_cli/src/lib.rs
index 99e3598bf..438cec752 100644
--- a/paddler_cli/src/lib.rs
+++ b/paddler_cli/src/lib.rs
@@ -28,7 +28,7 @@ struct Cli {
 
 #[derive(Subcommand)]
 enum Commands {
-    /// Generates tokens and embeddings; connects to the balancer
+    /// Runs inference for the balancer it connects to
     Agent(Agent),
     /// Distributes incoming requests among agents
     Balancer(Box<Balancer>),
diff --git a/paddler_cli/tests/main.rs b/paddler_cli/tests/main.rs
index 60190071f..13b2f1f75 100644
--- a/paddler_cli/tests/main.rs
+++ b/paddler_cli/tests/main.rs
@@ -1,3 +1 @@
-mod paddler_agent_rejects_more_slots_than_llama_cpp_can_serve;
 mod paddler_balancer_exits_with_a_failure_when_its_management_address_is_taken;
-mod paddler_without_a_subcommand_prints_help_and_exits_with_a_usage_error;
diff --git a/paddler_cli/tests/paddler_agent_rejects_more_slots_than_llama_cpp_can_serve.rs b/paddler_cli/tests/paddler_agent_rejects_more_slots_than_llama_cpp_can_serve.rs
deleted file mode 100644
index 0a8261d1d..000000000
--- a/paddler_cli/tests/paddler_agent_rejects_more_slots_than_llama_cpp_can_serve.rs
+++ /dev/null
@@ -1,18 +0,0 @@
-use std::process::Command;
-
-const CLAP_USAGE_ERROR_EXIT_CODE: i32 = 2;
-
-#[test]
-fn paddler_agent_rejects_more_slots_than_llama_cpp_can_serve() {
-    let agent_output = Command::new(env!("CARGO_BIN_EXE_paddler"))
-        .arg("agent")
-        .arg("--management-addr")
-        .arg("127.0.0.1:1")
-        .arg("--slots")
-        .arg("257")
-        .output()
-        .expect("paddler agent must run");
-
-    assert_eq!(agent_output.status.code(), Some(CLAP_USAGE_ERROR_EXIT_CODE));
-    assert!(agent_output.stdout.is_empty());
-}
diff --git a/paddler_cli/tests/paddler_balancer_exits_with_a_failure_when_its_management_address_is_taken.rs b/paddler_cli/tests/paddler_balancer_exits_with_a_failure_when_its_management_address_is_taken.rs
index e5110a7f3..86bab1933 100644
--- a/paddler_cli/tests/paddler_balancer_exits_with_a_failure_when_its_management_address_is_taken.rs
+++ b/paddler_cli/tests/paddler_balancer_exits_with_a_failure_when_its_management_address_is_taken.rs
@@ -21,5 +21,5 @@ fn paddler_balancer_exits_with_a_failure_when_its_management_address_is_taken()
         .expect("paddler balancer must run");
 
     assert_eq!(balancer_output.status.code(), Some(FAILED_MAIN_EXIT_CODE));
-    assert!(balancer_output.stdout.is_empty());
+    assert_eq!(balancer_output.stdout, Vec::<u8>::new());
 }
diff --git a/paddler_cli/tests/paddler_without_a_subcommand_prints_help_and_exits_with_a_usage_error.rs b/paddler_cli/tests/paddler_without_a_subcommand_prints_help_and_exits_with_a_usage_error.rs
deleted file mode 100644
index 5ccbda26b..000000000
--- a/paddler_cli/tests/paddler_without_a_subcommand_prints_help_and_exits_with_a_usage_error.rs
+++ /dev/null
@@ -1,18 +0,0 @@
-use std::process::Command;
-
-const CLAP_USAGE_ERROR_EXIT_CODE: i32 = 2;
-
-#[test]
-fn paddler_without_a_subcommand_prints_help_and_exits_with_a_usage_error() {
-    let help_output = Command::new(env!("CARGO_BIN_EXE_paddler"))
-        .arg("--help")
-        .output()
-        .expect("paddler --help must run");
-    let bare_output = Command::new(env!("CARGO_BIN_EXE_paddler"))
-        .output()
-        .expect("paddler without arguments must run");
-
-    assert_eq!(bare_output.status.code(), Some(CLAP_USAGE_ERROR_EXIT_CODE));
-    assert!(bare_output.stdout.is_empty());
-    assert_eq!(bare_output.stderr, help_output.stdout);
-}
diff --git a/paddler_cli_tests/Cargo.toml b/paddler_cli_tests/Cargo.toml
index e923a03d7..fbb7ff261 100644
--- a/paddler_cli_tests/Cargo.toml
+++ b/paddler_cli_tests/Cargo.toml
@@ -23,12 +23,13 @@ default = []
 cuda = ["paddler_cli/cuda"]
 metal = ["paddler_cli/metal"]
 tests_that_use_llms = []
-web_admin_panel = ["paddler_cli/web_admin_panel"]
+web_admin_panel = ["paddler_cli/web_admin_panel", "paddler_tests/web_admin_panel"]
 
 [dependencies]
 anyhow = { workspace = true }
 async-trait = { workspace = true }
-log = { workspace = true }
+futures-util = { workspace = true }
+nix = { workspace = true }
 paddler_balancer = { workspace = true }
 paddler_cli = { workspace = true }
 paddler_inference_parameters = { workspace = true }
@@ -39,14 +40,23 @@ thiserror = { workspace = true }
 tokio = { workspace = true }
 tokio-util = { workspace = true }
 
-[target.'cfg(unix)'.dependencies]
-nix = { workspace = true }
-
 [dev-dependencies]
-futures-util = { workspace = true }
+actix-web = { workspace = true }
+llama-cpp-bindings = { workspace = true }
+paddler_agent_text_generation = { workspace = true }
+paddler_bootstrap = { workspace = true }
 paddler_client = { workspace = true }
+paddler_openai_response_format_validator = { workspace = true }
+paddler_tests = { workspace = true }
 reqwest = { workspace = true }
+tempfile = { workspace = true }
+tokio-tungstenite = { workspace = true }
+trzcina = { workspace = true }
 url = { workspace = true }
 
+[target.'cfg(target_os = "linux")'.dev-dependencies]
+libc = { workspace = true }
+seccompiler = { workspace = true }
+
 [lints]
 workspace = true
diff --git a/paddler_cli_tests/src/agent_exited_before_ready_with_code.rs b/paddler_cli_tests/src/agent_exited_before_ready_with_code.rs
new file mode 100644
index 000000000..481be85bc
--- /dev/null
+++ b/paddler_cli_tests/src/agent_exited_before_ready_with_code.rs
@@ -0,0 +1,21 @@
+use anyhow::Error;
+
+use paddler_test_cluster_harness::cluster_harness_error::ClusterHarnessError;
+use paddler_test_cluster_harness::process_end::ProcessEnd;
+
+use crate::subprocess_cluster_error::SubprocessClusterError;
+
+#[must_use]
+pub fn agent_exited_before_ready_with_code(error: &Error, expected_exit_code: i32) -> bool {
+    matches!(
+        error.downcast_ref::<ClusterHarnessError>(),
+        Some(ClusterHarnessError::AgentExitedBeforeReady {
+            process_end: ProcessEnd::Failed(exit_failure),
+            ..
+        }) if matches!(
+            exit_failure.downcast_ref::<SubprocessClusterError>(),
+            Some(SubprocessClusterError::ProcessExitedWithFailure { exit_status })
+                if exit_status.code() == Some(expected_exit_code)
+        )
+    )
+}
diff --git a/paddler_cli_tests/src/clap_usage_error_exit_code.rs b/paddler_cli_tests/src/clap_usage_error_exit_code.rs
new file mode 100644
index 000000000..6517220e3
--- /dev/null
+++ b/paddler_cli_tests/src/clap_usage_error_exit_code.rs
@@ -0,0 +1 @@
+pub const CLAP_USAGE_ERROR_EXIT_CODE: i32 = 2;
diff --git a/paddler_cli_tests/src/connect_subprocess_balancer.rs b/paddler_cli_tests/src/connect_subprocess_balancer.rs
new file mode 100644
index 000000000..3257ef959
--- /dev/null
+++ b/paddler_cli_tests/src/connect_subprocess_balancer.rs
@@ -0,0 +1,41 @@
+use std::ffi::OsStr;
+
+use anyhow::Result;
+use tokio_util::sync::CancellationToken;
+
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
+
+use crate::signalled_subprocess::SignalledSubprocess;
+use crate::spawn_balancer_subprocess::spawn_balancer_subprocess;
+use crate::spawned_balancer_subprocess::SpawnedBalancerSubprocess;
+use crate::subprocess_agents_cluster_connection::SubprocessAgentsClusterConnection;
+use crate::subprocess_cluster::SubprocessCluster;
+
+pub async fn connect_subprocess_balancer<TArguments, TArgument>(
+    binary_path: &str,
+    balancer_arguments: TArguments,
+    desired_state: &ClusterDesiredState,
+) -> Result<SubprocessCluster>
+where
+    TArguments: IntoIterator<Item = TArgument>,
+    TArgument: AsRef<OsStr>,
+{
+    let SpawnedBalancerSubprocess {
+        running_balancer,
+        signals: balancer_signals,
+    } = spawn_balancer_subprocess(binary_path, balancer_arguments, SignalledSubprocess::new)
+        .await?;
+    let cluster = SubprocessAgentsClusterConnection {
+        agent_binary_path: binary_path,
+        cancellation_token: CancellationToken::new(),
+        desired_state,
+        running_balancer,
+    }
+    .connect()
+    .await?;
+
+    Ok(SubprocessCluster {
+        balancer_signals,
+        cluster,
+    })
+}
diff --git a/paddler_cli_tests/src/lib.rs b/paddler_cli_tests/src/lib.rs
index 4769d9969..2a849665c 100644
--- a/paddler_cli_tests/src/lib.rs
+++ b/paddler_cli_tests/src/lib.rs
@@ -1,12 +1,31 @@
+pub mod agent_exited_before_ready_with_code;
+pub mod clap_usage_error_exit_code;
+pub mod connect_subprocess_balancer;
+pub mod missing_binary_path;
 pub mod paddler_command;
+pub mod pausable_agent;
+pub mod pausable_agent_cluster;
+pub mod pausable_agent_cluster_params;
+pub mod pausable_agent_params;
 pub mod read_balancer_addresses;
+pub mod restart_balancer;
+pub mod signal_as_already_reaped;
+pub mod signalled_subprocess;
 pub mod spawn_agent_subprocess;
 pub mod spawn_agent_subprocess_params;
 pub mod spawn_balancer_subprocess;
+pub mod spawn_signalled_agent_subprocess;
+pub mod spawned_balancer_subprocess;
+pub mod stallable_tcp_relay;
+pub mod start_pausable_agent_cluster;
+pub mod start_signalled_subprocess_cluster;
 pub mod start_subprocess_cluster;
 pub mod start_subprocess_cluster_with_qwen3;
 pub mod start_subprocess_embedding_cluster;
 pub mod subprocess_agent_spawner;
+pub mod subprocess_agents_cluster_connection;
+pub mod subprocess_cluster;
 pub mod subprocess_cluster_error;
 pub mod subprocess_process;
-pub mod terminate_child;
+pub mod subprocess_signals;
+pub mod unreachable_management_addr;
diff --git a/paddler_cli_tests/src/missing_binary_path.rs b/paddler_cli_tests/src/missing_binary_path.rs
new file mode 100644
index 000000000..ef495fe85
--- /dev/null
+++ b/paddler_cli_tests/src/missing_binary_path.rs
@@ -0,0 +1 @@
+pub const MISSING_BINARY_PATH: &str = "/paddler/definitely/missing/paddler";
diff --git a/paddler_cli_tests/src/pausable_agent.rs b/paddler_cli_tests/src/pausable_agent.rs
new file mode 100644
index 000000000..3bca2bc33
--- /dev/null
+++ b/paddler_cli_tests/src/pausable_agent.rs
@@ -0,0 +1,53 @@
+use anyhow::Result;
+
+use paddler_test_cluster_harness::agent_readiness::AgentReadiness;
+use paddler_test_cluster_harness::cluster::Cluster;
+use paddler_test_cluster_harness::running_agent::RunningAgent;
+
+use crate::pausable_agent_params::PausableAgentParams;
+use crate::signalled_subprocess::SignalledSubprocess;
+use crate::spawn_agent_subprocess_params::SpawnAgentSubprocessParams;
+use crate::spawn_signalled_agent_subprocess::spawn_signalled_agent_subprocess;
+use crate::subprocess_signals::SubprocessSignals;
+
+pub struct PausableAgent {
+    pub id: String,
+    pub signals: SubprocessSignals,
+}
+
+impl PausableAgent {
+    pub async fn join(
+        cluster: &mut Cluster,
+        PausableAgentParams {
+            binary_path,
+            config,
+            expected_slots_total,
+            management_addr,
+        }: PausableAgentParams,
+    ) -> Result<Self> {
+        let SignalledSubprocess {
+            process: agent_process,
+            signals,
+        } = spawn_signalled_agent_subprocess(SpawnAgentSubprocessParams {
+            binary_path,
+            management_addr,
+            name: config.name.clone(),
+            slots: config.slot_count,
+        })?;
+        let mut running_agent = RunningAgent::new(config, Box::new(agent_process));
+        let ready_agent = running_agent
+            .wait_until_ready(
+                &mut cluster.agents_watcher,
+                AgentReadiness::SlotsReady(expected_slots_total),
+            )
+            .await?;
+
+        cluster.agent_ids.push(ready_agent.id.clone());
+        cluster.agents.push(running_agent);
+
+        Ok(Self {
+            id: ready_agent.id,
+            signals,
+        })
+    }
+}
diff --git a/paddler_cli_tests/src/pausable_agent_cluster.rs b/paddler_cli_tests/src/pausable_agent_cluster.rs
new file mode 100644
index 000000000..07a5f3c62
--- /dev/null
+++ b/paddler_cli_tests/src/pausable_agent_cluster.rs
@@ -0,0 +1,8 @@
+use paddler_test_cluster_harness::cluster::Cluster;
+
+use crate::pausable_agent::PausableAgent;
+
+pub struct PausableAgentCluster {
+    pub cluster: Cluster,
+    pub pausable_agent: PausableAgent,
+}
diff --git a/paddler_cli_tests/src/pausable_agent_cluster_params.rs b/paddler_cli_tests/src/pausable_agent_cluster_params.rs
new file mode 100644
index 000000000..a03becc70
--- /dev/null
+++ b/paddler_cli_tests/src/pausable_agent_cluster_params.rs
@@ -0,0 +1,8 @@
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
+
+pub struct PausableAgentClusterParams {
+    pub binary_path: String,
+    pub desired_state: ClusterDesiredState,
+    pub expected_slots_total: u16,
+    pub slot_count: u16,
+}
diff --git a/paddler_cli_tests/src/pausable_agent_params.rs b/paddler_cli_tests/src/pausable_agent_params.rs
new file mode 100644
index 000000000..a0a20cbd8
--- /dev/null
+++ b/paddler_cli_tests/src/pausable_agent_params.rs
@@ -0,0 +1,10 @@
+use std::net::SocketAddr;
+
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+
+pub struct PausableAgentParams {
+    pub binary_path: String,
+    pub config: AgentConfig,
+    pub expected_slots_total: u16,
+    pub management_addr: SocketAddr,
+}
diff --git a/paddler_cli_tests/src/read_balancer_addresses.rs b/paddler_cli_tests/src/read_balancer_addresses.rs
index dab942a5c..5438de7cb 100644
--- a/paddler_cli_tests/src/read_balancer_addresses.rs
+++ b/paddler_cli_tests/src/read_balancer_addresses.rs
@@ -9,38 +9,39 @@ use crate::subprocess_cluster_error::SubprocessClusterError;
 
 pub async fn read_balancer_addresses<TReader>(
     reader: TReader,
-) -> Result<BalancerAddresses, SubprocessClusterError>
+) -> Result<Option<BalancerAddresses>, SubprocessClusterError>
 where
     TReader: AsyncRead + Unpin,
 {
-    let announcement = BufReader::new(reader)
+    let Some(announcement) = BufReader::new(reader)
         .lines()
         .next_line()
         .await
         .map_err(|source| SubprocessClusterError::AnnouncementUnreadable { source })?
-        .ok_or(SubprocessClusterError::StdoutClosedBeforeAnnouncement)?;
+    else {
+        return Ok(None);
+    };
 
-    from_str(&announcement).map_err(|source| SubprocessClusterError::AnnouncementInvalid {
-        announcement,
-        source,
+    from_str(&announcement).map(Some).map_err(|source| {
+        SubprocessClusterError::AnnouncementInvalid {
+            announcement,
+            source,
+        }
     })
 }
 
 #[cfg(test)]
 mod tests {
+    use std::error::Error as _;
+    use std::io;
+    use std::mem::discriminant;
+
     use super::read_balancer_addresses;
     use crate::subprocess_cluster_error::SubprocessClusterError;
 
     #[tokio::test]
-    async fn reports_stdout_closed_before_the_announcement() {
-        let read_error = read_balancer_addresses(&b""[..])
-            .await
-            .expect_err("an empty stdout must not yield balancer addresses");
-
-        assert!(matches!(
-            read_error,
-            SubprocessClusterError::StdoutClosedBeforeAnnouncement
-        ));
+    async fn yields_no_addresses_when_stdout_closes_before_the_announcement() {
+        assert_eq!(read_balancer_addresses(&b""[..]).await.ok(), Some(None));
     }
 
     #[tokio::test]
@@ -61,9 +62,18 @@ mod tests {
             .await
             .expect_err("a line that is not UTF-8 must not be readable");
 
-        assert!(matches!(
-            read_error,
-            SubprocessClusterError::AnnouncementUnreadable { .. }
-        ));
+        assert_eq!(
+            discriminant(&read_error),
+            discriminant(&SubprocessClusterError::AnnouncementUnreadable {
+                source: io::Error::from(io::ErrorKind::InvalidData)
+            })
+        );
+        assert_eq!(
+            read_error
+                .source()
+                .and_then(|source| source.downcast_ref::<io::Error>())
+                .map(io::Error::kind),
+            Some(io::ErrorKind::InvalidData)
+        );
     }
 }
diff --git a/paddler_cli_tests/src/restart_balancer.rs b/paddler_cli_tests/src/restart_balancer.rs
new file mode 100644
index 000000000..d0694419d
--- /dev/null
+++ b/paddler_cli_tests/src/restart_balancer.rs
@@ -0,0 +1,26 @@
+use std::net::SocketAddr;
+
+use anyhow::Result;
+
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
+use paddler_test_cluster_harness::ephemeral_loopback_addr::EPHEMERAL_LOOPBACK_ADDR;
+
+use crate::connect_subprocess_balancer::connect_subprocess_balancer;
+use crate::subprocess_cluster::SubprocessCluster;
+
+pub async fn restart_balancer(
+    binary_path: &str,
+    management_addr: SocketAddr,
+) -> Result<SubprocessCluster> {
+    connect_subprocess_balancer(
+        binary_path,
+        [
+            "--inference-addr".to_owned(),
+            EPHEMERAL_LOOPBACK_ADDR.to_string(),
+            "--management-addr".to_owned(),
+            management_addr.to_string(),
+        ],
+        &ClusterDesiredState::KeepStored,
+    )
+    .await
+}
diff --git a/paddler_cli_tests/src/signal_as_already_reaped.rs b/paddler_cli_tests/src/signal_as_already_reaped.rs
new file mode 100644
index 000000000..e108b5a16
--- /dev/null
+++ b/paddler_cli_tests/src/signal_as_already_reaped.rs
@@ -0,0 +1,10 @@
+use tokio::process::Child;
+
+use crate::signalled_subprocess::SignalledSubprocess;
+use crate::subprocess_cluster_error::SubprocessClusterError;
+
+pub fn signal_as_already_reaped(
+    _subprocess: Child,
+) -> Result<SignalledSubprocess, SubprocessClusterError> {
+    Err(SubprocessClusterError::ProcessAlreadyReaped)
+}
diff --git a/paddler_cli_tests/src/signalled_subprocess.rs b/paddler_cli_tests/src/signalled_subprocess.rs
new file mode 100644
index 000000000..4a4cfbe12
--- /dev/null
+++ b/paddler_cli_tests/src/signalled_subprocess.rs
@@ -0,0 +1,21 @@
+use tokio::process::Child;
+
+use crate::subprocess_cluster_error::SubprocessClusterError;
+use crate::subprocess_process::SubprocessProcess;
+use crate::subprocess_signals::SubprocessSignals;
+
+pub struct SignalledSubprocess {
+    pub process: SubprocessProcess,
+    pub signals: SubprocessSignals,
+}
+
+impl SignalledSubprocess {
+    pub fn new(child: Child) -> Result<Self, SubprocessClusterError> {
+        let process = SubprocessProcess::new(child);
+        let signals = process
+            .signals()
+            .ok_or(SubprocessClusterError::ProcessAlreadyReaped)?;
+
+        Ok(Self { process, signals })
+    }
+}
diff --git a/paddler_cli_tests/src/spawn_agent_subprocess.rs b/paddler_cli_tests/src/spawn_agent_subprocess.rs
index daeeac0d1..8e58b582b 100644
--- a/paddler_cli_tests/src/spawn_agent_subprocess.rs
+++ b/paddler_cli_tests/src/spawn_agent_subprocess.rs
@@ -1,11 +1,10 @@
 use std::process::Stdio;
 
-use anyhow::Context as _;
-use anyhow::Result;
 use tokio::process::Child;
 
 use crate::paddler_command::paddler_command;
 use crate::spawn_agent_subprocess_params::SpawnAgentSubprocessParams;
+use crate::subprocess_cluster_error::SubprocessClusterError;
 
 pub fn spawn_agent_subprocess(
     SpawnAgentSubprocessParams {
@@ -14,7 +13,7 @@ pub fn spawn_agent_subprocess(
         name,
         slots,
     }: SpawnAgentSubprocessParams,
-) -> Result<Child> {
+) -> Result<Child, SubprocessClusterError> {
     paddler_command(&binary_path)
         .arg("agent")
         .arg("--management-addr")
@@ -26,5 +25,5 @@ pub fn spawn_agent_subprocess(
         .stdout(Stdio::null())
         .stderr(Stdio::null())
         .spawn()
-        .context("failed to spawn paddler agent subprocess")
+        .map_err(SubprocessClusterError::AgentSpawnFailed)
 }
diff --git a/paddler_cli_tests/src/spawn_balancer_subprocess.rs b/paddler_cli_tests/src/spawn_balancer_subprocess.rs
index c18704777..55876796a 100644
--- a/paddler_cli_tests/src/spawn_balancer_subprocess.rs
+++ b/paddler_cli_tests/src/spawn_balancer_subprocess.rs
@@ -3,37 +3,60 @@ use std::process::Stdio;
 
 use anyhow::Context as _;
 use anyhow::Result;
+use futures_util::TryFutureExt as _;
+use futures_util::future::ready;
+use tokio::process::Child;
 
 use paddler_test_cluster_harness::running_balancer::RunningBalancer;
 
 use crate::paddler_command::paddler_command;
 use crate::read_balancer_addresses::read_balancer_addresses;
+use crate::signalled_subprocess::SignalledSubprocess;
+use crate::spawned_balancer_subprocess::SpawnedBalancerSubprocess;
 use crate::subprocess_cluster_error::SubprocessClusterError;
-use crate::subprocess_process::SubprocessProcess;
 
-pub async fn spawn_balancer_subprocess<TArguments, TArgument>(
+pub async fn spawn_balancer_subprocess<TArguments, TArgument, TSignalSubprocess>(
     binary_path: &str,
     arguments: TArguments,
-) -> Result<RunningBalancer>
+    signal_subprocess: TSignalSubprocess,
+) -> Result<SpawnedBalancerSubprocess>
 where
     TArguments: IntoIterator<Item = TArgument>,
     TArgument: AsRef<OsStr>,
+    TSignalSubprocess: FnOnce(Child) -> Result<SignalledSubprocess, SubprocessClusterError>,
 {
     let mut balancer_subprocess = paddler_command(binary_path)
         .arg("balancer")
         .args(arguments)
         .stdout(Stdio::piped())
-        .stderr(Stdio::null())
+        .stderr(Stdio::inherit())
         .spawn()
         .context("failed to spawn paddler balancer subprocess")?;
-    let balancer_stdout = balancer_subprocess
-        .stdout
-        .take()
-        .ok_or(SubprocessClusterError::StdoutNotPiped)?;
-    let addresses = read_balancer_addresses(balancer_stdout).await?;
+    let Some(addresses) = ready(
+        balancer_subprocess
+            .stdout
+            .take()
+            .ok_or(SubprocessClusterError::StdoutNotPiped),
+    )
+    .and_then(read_balancer_addresses)
+    .await?
+    else {
+        let exit_status = balancer_subprocess
+            .wait()
+            .await
+            .map_err(SubprocessClusterError::ProcessExitUnobservable)?;
 
-    Ok(RunningBalancer::new(
-        addresses,
-        Box::new(SubprocessProcess::new(balancer_subprocess)),
-    ))
+        return Err(SubprocessClusterError::BalancerExitedBeforeAnnouncing { exit_status }.into());
+    };
+    let SignalledSubprocess {
+        process: balancer_process,
+        signals,
+    } = signal_subprocess(balancer_subprocess)?;
+
+    RunningBalancer::new(addresses, Box::new(balancer_process))
+        .map(|running_balancer| SpawnedBalancerSubprocess {
+            running_balancer,
+            signals,
+        })
+        .context("failed to address the balancer subprocess")
 }
diff --git a/paddler_cli_tests/src/spawn_signalled_agent_subprocess.rs b/paddler_cli_tests/src/spawn_signalled_agent_subprocess.rs
new file mode 100644
index 000000000..67afc8a5b
--- /dev/null
+++ b/paddler_cli_tests/src/spawn_signalled_agent_subprocess.rs
@@ -0,0 +1,10 @@
+use crate::signalled_subprocess::SignalledSubprocess;
+use crate::spawn_agent_subprocess::spawn_agent_subprocess;
+use crate::spawn_agent_subprocess_params::SpawnAgentSubprocessParams;
+use crate::subprocess_cluster_error::SubprocessClusterError;
+
+pub fn spawn_signalled_agent_subprocess(
+    spawn_agent_subprocess_params: SpawnAgentSubprocessParams,
+) -> Result<SignalledSubprocess, SubprocessClusterError> {
+    SignalledSubprocess::new(spawn_agent_subprocess(spawn_agent_subprocess_params)?)
+}
diff --git a/paddler_cli_tests/src/spawned_balancer_subprocess.rs b/paddler_cli_tests/src/spawned_balancer_subprocess.rs
new file mode 100644
index 000000000..493db5f3d
--- /dev/null
+++ b/paddler_cli_tests/src/spawned_balancer_subprocess.rs
@@ -0,0 +1,8 @@
+use paddler_test_cluster_harness::running_balancer::RunningBalancer;
+
+use crate::subprocess_signals::SubprocessSignals;
+
+pub struct SpawnedBalancerSubprocess {
+    pub running_balancer: RunningBalancer,
+    pub signals: SubprocessSignals,
+}
diff --git a/paddler_cli_tests/src/stallable_tcp_relay.rs b/paddler_cli_tests/src/stallable_tcp_relay.rs
new file mode 100644
index 000000000..7365dadb3
--- /dev/null
+++ b/paddler_cli_tests/src/stallable_tcp_relay.rs
@@ -0,0 +1,116 @@
+use std::future::pending;
+use std::io;
+use std::net::SocketAddr;
+
+use futures_util::StreamExt as _;
+use futures_util::stream::FuturesUnordered;
+use tokio::io::copy_bidirectional;
+use tokio::net::TcpListener;
+use tokio::net::TcpStream;
+use tokio::select;
+use tokio::spawn;
+use tokio::sync::watch;
+use tokio::sync::watch::error::RecvError;
+use tokio::task::JoinError;
+use tokio::task::JoinHandle;
+use tokio_util::sync::CancellationToken;
+
+async fn relay_until_stalled(
+    mut downstream: TcpStream,
+    upstream_addr: SocketAddr,
+    stall: CancellationToken,
+) -> io::Result<()> {
+    let mut upstream = TcpStream::connect(upstream_addr).await?;
+
+    select! {
+        relay_outcome = copy_bidirectional(&mut downstream, &mut upstream) => relay_outcome.map(drop),
+        () = stall.cancelled() => pending().await,
+    }
+}
+
+async fn relay_accepted_connections(
+    listener: TcpListener,
+    upstream_addr: SocketAddr,
+    accepted_connection_count_tx: watch::Sender<u32>,
+    stall_rx: watch::Receiver<CancellationToken>,
+    stop: CancellationToken,
+) -> io::Result<()> {
+    let mut relayed_connections = FuturesUnordered::new();
+
+    loop {
+        select! {
+            () = stop.cancelled() => break Ok(()),
+            accepted_connection = listener.accept() => {
+                let (downstream, _downstream_addr) = accepted_connection?;
+
+                accepted_connection_count_tx.send_modify(|accepted_connection_count| {
+                    *accepted_connection_count += 1;
+                });
+                relayed_connections.push(relay_until_stalled(
+                    downstream,
+                    upstream_addr,
+                    stall_rx.borrow().clone(),
+                ));
+            }
+            Some(relayed_connection) = relayed_connections.next() => relayed_connection?,
+        }
+    }
+}
+
+pub struct StallableTcpRelay {
+    accepted_connection_count_rx: watch::Receiver<u32>,
+    relaying_task: JoinHandle<io::Result<()>>,
+    stall_tx: watch::Sender<CancellationToken>,
+    stop: CancellationToken,
+}
+
+impl StallableTcpRelay {
+    #[must_use]
+    pub fn start(listener: TcpListener, upstream_addr: SocketAddr) -> Self {
+        let (accepted_connection_count_tx, accepted_connection_count_rx) = watch::channel(0);
+        let (stall_tx, stall_rx) = watch::channel(CancellationToken::new());
+        let stop = CancellationToken::new();
+        let relaying_task = spawn(relay_accepted_connections(
+            listener,
+            upstream_addr,
+            accepted_connection_count_tx,
+            stall_rx,
+            stop.clone(),
+        ));
+
+        Self {
+            accepted_connection_count_rx,
+            relaying_task,
+            stall_tx,
+            stop,
+        }
+    }
+
+    pub fn stall_established_connections(&self) {
+        self.stall_tx
+            .send_replace(CancellationToken::new())
+            .cancel();
+    }
+
+    pub async fn wait_for_accepted_connections(
+        &self,
+        expected_accepted_connection_count: u32,
+    ) -> Result<(), RecvError> {
+        self.accepted_connection_count_rx
+            .clone()
+            .wait_for(|accepted_connection_count| {
+                *accepted_connection_count >= expected_accepted_connection_count
+            })
+            .await
+            .map(drop)
+    }
+
+    pub async fn wait_until_relaying_ends(self) -> Result<io::Result<()>, JoinError> {
+        self.relaying_task.await
+    }
+
+    pub async fn stop(self) -> Result<io::Result<()>, JoinError> {
+        self.stop.cancel();
+        self.relaying_task.await
+    }
+}
diff --git a/paddler_cli_tests/src/start_pausable_agent_cluster.rs b/paddler_cli_tests/src/start_pausable_agent_cluster.rs
new file mode 100644
index 000000000..096874a49
--- /dev/null
+++ b/paddler_cli_tests/src/start_pausable_agent_cluster.rs
@@ -0,0 +1,44 @@
+use anyhow::Result;
+
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+
+use crate::pausable_agent::PausableAgent;
+use crate::pausable_agent_cluster::PausableAgentCluster;
+use crate::pausable_agent_cluster_params::PausableAgentClusterParams;
+use crate::pausable_agent_params::PausableAgentParams;
+use crate::start_subprocess_cluster::start_subprocess_cluster;
+
+pub async fn start_pausable_agent_cluster(
+    PausableAgentClusterParams {
+        binary_path,
+        desired_state,
+        expected_slots_total,
+        slot_count,
+    }: PausableAgentClusterParams,
+) -> Result<PausableAgentCluster> {
+    let mut cluster = start_subprocess_cluster(
+        &binary_path,
+        ClusterParams {
+            desired_state,
+            ..ClusterParams::without_agents()
+        },
+    )
+    .await?;
+    let management_addr = cluster.balancer.addresses.management;
+    let pausable_agent = PausableAgent::join(
+        &mut cluster,
+        PausableAgentParams {
+            binary_path,
+            config: AgentConfig::single(slot_count),
+            expected_slots_total,
+            management_addr,
+        },
+    )
+    .await?;
+
+    Ok(PausableAgentCluster {
+        cluster,
+        pausable_agent,
+    })
+}
diff --git a/paddler_cli_tests/src/start_signalled_subprocess_cluster.rs b/paddler_cli_tests/src/start_signalled_subprocess_cluster.rs
new file mode 100644
index 000000000..7a1c4815c
--- /dev/null
+++ b/paddler_cli_tests/src/start_signalled_subprocess_cluster.rs
@@ -0,0 +1,64 @@
+use anyhow::Result;
+
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_test_cluster_harness::ephemeral_loopback_addr::EPHEMERAL_LOOPBACK_ADDR;
+
+use crate::connect_subprocess_balancer::connect_subprocess_balancer;
+use crate::subprocess_cluster::SubprocessCluster;
+
+pub async fn start_signalled_subprocess_cluster(
+    binary_path: &str,
+    ClusterParams {
+        agents,
+        buffered_request_timeout,
+        desired_state,
+        inference_cors_allowed_hosts,
+        inference_item_timeout,
+        management_cors_allowed_hosts,
+        max_buffered_requests,
+        state_database_url,
+        wait_for_slots_ready,
+    }: ClusterParams,
+) -> Result<SubprocessCluster> {
+    let ephemeral_loopback_addr = EPHEMERAL_LOOPBACK_ADDR.to_string();
+    let mut balancer_arguments = vec![
+        "--compat-openai-addr".to_owned(),
+        ephemeral_loopback_addr.clone(),
+        "--compat-typesafe-addr".to_owned(),
+        ephemeral_loopback_addr.clone(),
+        "--inference-addr".to_owned(),
+        ephemeral_loopback_addr.clone(),
+        "--management-addr".to_owned(),
+        ephemeral_loopback_addr,
+        "--state-database".to_owned(),
+        state_database_url,
+        "--max-buffered-requests".to_owned(),
+        max_buffered_requests.to_string(),
+        "--buffered-request-timeout".to_owned(),
+        buffered_request_timeout.as_millis().to_string(),
+        "--inference-item-timeout".to_owned(),
+        inference_item_timeout.as_millis().to_string(),
+    ];
+
+    for allowed_host in inference_cors_allowed_hosts {
+        balancer_arguments.extend(["--inference-cors-allowed-host".to_owned(), allowed_host]);
+    }
+
+    for allowed_host in management_cors_allowed_hosts {
+        balancer_arguments.extend(["--management-cors-allowed-host".to_owned(), allowed_host]);
+    }
+
+    let SubprocessCluster {
+        balancer_signals,
+        mut cluster,
+    } = connect_subprocess_balancer(binary_path, balancer_arguments, &desired_state).await?;
+
+    cluster
+        .register_agents(&agents, wait_for_slots_ready)
+        .await?;
+
+    Ok(SubprocessCluster {
+        balancer_signals,
+        cluster,
+    })
+}
diff --git a/paddler_cli_tests/src/start_subprocess_cluster.rs b/paddler_cli_tests/src/start_subprocess_cluster.rs
index 56d614e6a..e0c97ae2a 100644
--- a/paddler_cli_tests/src/start_subprocess_cluster.rs
+++ b/paddler_cli_tests/src/start_subprocess_cluster.rs
@@ -1,70 +1,17 @@
 use anyhow::Result;
-use tokio_util::sync::CancellationToken;
 
 use paddler_test_cluster_harness::cluster::Cluster;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
-use paddler_test_cluster_harness::ephemeral_loopback_addr::EPHEMERAL_LOOPBACK_ADDR;
 
-use crate::spawn_balancer_subprocess::spawn_balancer_subprocess;
-use crate::subprocess_agent_spawner::SubprocessAgentSpawner;
+use crate::start_signalled_subprocess_cluster::start_signalled_subprocess_cluster;
 
 pub async fn start_subprocess_cluster(
     binary_path: &str,
-    ClusterParams {
-        agents,
-        buffered_request_timeout,
-        desired_state,
-        inference_cors_allowed_hosts,
-        inference_item_timeout,
-        management_cors_allowed_hosts,
-        max_buffered_requests,
-        state_database_url,
-        wait_for_slots_ready,
-    }: ClusterParams,
+    cluster_params: ClusterParams,
 ) -> Result<Cluster> {
-    let ephemeral_loopback_addr = EPHEMERAL_LOOPBACK_ADDR.to_string();
-    let mut balancer_arguments = vec![
-        "--inference-addr".to_owned(),
-        ephemeral_loopback_addr.clone(),
-        "--management-addr".to_owned(),
-        ephemeral_loopback_addr.clone(),
-        "--compat-openai-addr".to_owned(),
-        ephemeral_loopback_addr,
-        "--state-database".to_owned(),
-        state_database_url,
-        "--max-buffered-requests".to_owned(),
-        max_buffered_requests.to_string(),
-        "--buffered-request-timeout".to_owned(),
-        buffered_request_timeout.as_millis().to_string(),
-        "--inference-item-timeout".to_owned(),
-        inference_item_timeout.as_millis().to_string(),
-    ];
-
-    for allowed_host in inference_cors_allowed_hosts {
-        balancer_arguments.extend(["--inference-cors-allowed-host".to_owned(), allowed_host]);
-    }
-
-    for allowed_host in management_cors_allowed_hosts {
-        balancer_arguments.extend(["--management-cors-allowed-host".to_owned(), allowed_host]);
-    }
-
-    let running_balancer = spawn_balancer_subprocess(binary_path, balancer_arguments).await?;
-    let management_addr = running_balancer.addresses.management;
-
-    let mut cluster = Cluster::connect(
-        CancellationToken::new(),
-        running_balancer,
-        Box::new(SubprocessAgentSpawner::new(
-            binary_path.to_owned(),
-            management_addr,
-        )),
-        desired_state.as_ref(),
+    Ok(
+        start_signalled_subprocess_cluster(binary_path, cluster_params)
+            .await?
+            .cluster,
     )
-    .await?;
-
-    cluster
-        .register_agents(&agents, wait_for_slots_ready)
-        .await?;
-
-    Ok(cluster)
 }
diff --git a/paddler_cli_tests/src/start_subprocess_cluster_with_qwen3.rs b/paddler_cli_tests/src/start_subprocess_cluster_with_qwen3.rs
index 9f45f22c9..832bfe64e 100644
--- a/paddler_cli_tests/src/start_subprocess_cluster_with_qwen3.rs
+++ b/paddler_cli_tests/src/start_subprocess_cluster_with_qwen3.rs
@@ -2,6 +2,7 @@ use anyhow::Result;
 
 use paddler_test_cluster_harness::agent_config::AgentConfig;
 use paddler_test_cluster_harness::cluster::Cluster;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
 use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
 
@@ -15,7 +16,7 @@ pub async fn start_subprocess_cluster_with_qwen3(
         binary_path,
         ClusterParams {
             agents,
-            desired_state: Some(qwen3_0_6b().into_desired_state()),
+            desired_state: ClusterDesiredState::Apply(Box::new(qwen3_0_6b().into_desired_state())),
             wait_for_slots_ready: true,
             ..ClusterParams::default()
         },
diff --git a/paddler_cli_tests/src/subprocess_agents_cluster_connection.rs b/paddler_cli_tests/src/subprocess_agents_cluster_connection.rs
new file mode 100644
index 000000000..c3193162c
--- /dev/null
+++ b/paddler_cli_tests/src/subprocess_agents_cluster_connection.rs
@@ -0,0 +1,32 @@
+use anyhow::Result;
+use tokio_util::sync::CancellationToken;
+
+use paddler_test_cluster_harness::cluster::Cluster;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
+use paddler_test_cluster_harness::running_balancer::RunningBalancer;
+
+use crate::subprocess_agent_spawner::SubprocessAgentSpawner;
+
+pub struct SubprocessAgentsClusterConnection<'connection> {
+    pub agent_binary_path: &'connection str,
+    pub cancellation_token: CancellationToken,
+    pub desired_state: &'connection ClusterDesiredState,
+    pub running_balancer: RunningBalancer,
+}
+
+impl SubprocessAgentsClusterConnection<'_> {
+    pub async fn connect(self) -> Result<Cluster> {
+        let management_addr = self.running_balancer.addresses.management;
+
+        Cluster::connect(
+            self.cancellation_token,
+            self.running_balancer,
+            Box::new(SubprocessAgentSpawner::new(
+                self.agent_binary_path.to_owned(),
+                management_addr,
+            )),
+            self.desired_state,
+        )
+        .await
+    }
+}
diff --git a/paddler_cli_tests/src/subprocess_cluster.rs b/paddler_cli_tests/src/subprocess_cluster.rs
new file mode 100644
index 000000000..eb12d1fea
--- /dev/null
+++ b/paddler_cli_tests/src/subprocess_cluster.rs
@@ -0,0 +1,8 @@
+use paddler_test_cluster_harness::cluster::Cluster;
+
+use crate::subprocess_signals::SubprocessSignals;
+
+pub struct SubprocessCluster {
+    pub balancer_signals: SubprocessSignals,
+    pub cluster: Cluster,
+}
diff --git a/paddler_cli_tests/src/subprocess_cluster_error.rs b/paddler_cli_tests/src/subprocess_cluster_error.rs
index 83d9d7703..9f0d9ace7 100644
--- a/paddler_cli_tests/src/subprocess_cluster_error.rs
+++ b/paddler_cli_tests/src/subprocess_cluster_error.rs
@@ -1,17 +1,40 @@
 use std::io;
-use std::num::TryFromIntError;
+use std::process::ExitStatus;
 
+use nix::errno::Errno;
+use nix::sys::signal::Signal;
+use nix::sys::wait::WaitStatus;
+use nix::unistd::Pid;
 use serde_json::Error as SerdeJsonError;
 use thiserror::Error;
 
 #[derive(Debug, Error)]
 pub enum SubprocessClusterError {
-    #[error("The subprocess PID {raw_pid} does not fit into a signal target")]
-    ProcessIdOutOfRange {
-        raw_pid: u32,
+    #[error("Unable to spawn the paddler agent subprocess")]
+    AgentSpawnFailed(#[source] io::Error),
+    #[error("The balancer subprocess exited with {exit_status} before announcing its addresses")]
+    BalancerExitedBeforeAnnouncing { exit_status: ExitStatus },
+    #[error("The subprocess {pid} reported {status:?} instead of stopping")]
+    ProcessDidNotStop { pid: Pid, status: WaitStatus },
+    #[error("The subprocess was already reaped, so it can no longer be signalled")]
+    ProcessAlreadyReaped,
+    #[error("Unable to observe the state of subprocess {pid}")]
+    ProcessUnobservable {
+        pid: Pid,
         #[source]
-        source: TryFromIntError,
+        source: Errno,
     },
+    #[error("Unable to send {signal} to subprocess {pid}")]
+    SignalUndeliverable {
+        pid: Pid,
+        signal: Signal,
+        #[source]
+        source: Errno,
+    },
+    #[error("The subprocess exited with {exit_status}")]
+    ProcessExitedWithFailure { exit_status: ExitStatus },
+    #[error("Unable to observe how the subprocess exited")]
+    ProcessExitUnobservable(#[source] io::Error),
     #[error(
         "The balancer subprocess announcement is not valid balancer addresses: {announcement:?}"
     )]
@@ -25,8 +48,6 @@ pub enum SubprocessClusterError {
         #[source]
         source: io::Error,
     },
-    #[error("The balancer subprocess closed its stdout before announcing its addresses")]
-    StdoutClosedBeforeAnnouncement,
     #[error("The balancer subprocess stdout is not piped")]
     StdoutNotPiped,
 }
diff --git a/paddler_cli_tests/src/subprocess_process.rs b/paddler_cli_tests/src/subprocess_process.rs
index 02464eb27..1e487cad5 100644
--- a/paddler_cli_tests/src/subprocess_process.rs
+++ b/paddler_cli_tests/src/subprocess_process.rs
@@ -1,37 +1,70 @@
+use std::sync::Arc;
+use std::sync::atomic::AtomicBool;
+use std::sync::atomic::Ordering;
+
 use anyhow::Result;
 use async_trait::async_trait;
-use log::warn;
+use nix::unistd::Pid;
 use tokio::process::Child;
 
 use paddler_test_cluster_harness::managed_process::ManagedProcess;
+use paddler_test_cluster_harness::process_end::ProcessEnd;
 
-use crate::terminate_child::terminate_child;
+use crate::subprocess_cluster_error::SubprocessClusterError;
+use crate::subprocess_signals::SubprocessSignals;
 
 pub struct SubprocessProcess {
     child: Child,
+    kill_requested: Arc<AtomicBool>,
 }
 
 impl SubprocessProcess {
     #[must_use]
-    pub const fn new(child: Child) -> Self {
-        Self { child }
+    pub fn new(child: Child) -> Self {
+        Self {
+            child,
+            kill_requested: Arc::default(),
+        }
+    }
+
+    #[must_use]
+    pub fn signals(&self) -> Option<SubprocessSignals> {
+        self.child.id().map(|raw_pid| SubprocessSignals {
+            kill_requested: self.kill_requested.clone(),
+            pid: Pid::from_raw(raw_pid.cast_signed()),
+        })
+    }
+}
+
+impl SubprocessProcess {
+    async fn wait_for_exit(&mut self) -> ProcessEnd {
+        match self.child.wait().await {
+            Ok(exit_status)
+                if exit_status.success() || self.kill_requested.load(Ordering::Acquire) =>
+            {
+                ProcessEnd::Clean
+            }
+            Ok(exit_status) => {
+                ProcessEnd::failed(SubprocessClusterError::ProcessExitedWithFailure { exit_status })
+            }
+            Err(source) => {
+                ProcessEnd::failed(SubprocessClusterError::ProcessExitUnobservable(source))
+            }
+        }
     }
 }
 
 #[async_trait]
 impl ManagedProcess for SubprocessProcess {
-    async fn shutdown(mut self: Box<Self>) -> Result<()> {
-        terminate_child(&mut self.child)?;
-        self.child.wait().await?;
+    fn terminate(&mut self) -> Result<()> {
+        let Some(signals) = self.signals() else {
+            return Ok(());
+        };
 
-        Ok(())
+        signals.terminate().map_err(Into::into)
     }
-}
 
-impl Drop for SubprocessProcess {
-    fn drop(&mut self) {
-        if let Err(error) = terminate_child(&mut self.child) {
-            warn!("SubprocessProcess drop: failed to terminate subprocess: {error:#}");
-        }
+    async fn exited(&mut self) -> ProcessEnd {
+        self.wait_for_exit().await
     }
 }
diff --git a/paddler_cli_tests/src/subprocess_signals.rs b/paddler_cli_tests/src/subprocess_signals.rs
new file mode 100644
index 000000000..b5d1a2562
--- /dev/null
+++ b/paddler_cli_tests/src/subprocess_signals.rs
@@ -0,0 +1,158 @@
+use std::sync::Arc;
+use std::sync::atomic::AtomicBool;
+use std::sync::atomic::Ordering;
+
+use nix::errno::Errno;
+use nix::sys::signal::Signal;
+use nix::sys::signal::kill;
+use nix::sys::wait::WaitPidFlag;
+use nix::sys::wait::WaitStatus;
+use nix::sys::wait::waitpid;
+use nix::unistd::Pid;
+
+use crate::subprocess_cluster_error::SubprocessClusterError;
+
+#[derive(Clone, Debug)]
+pub struct SubprocessSignals {
+    pub kill_requested: Arc<AtomicBool>,
+    pub pid: Pid,
+}
+
+impl SubprocessSignals {
+    pub fn pause(&self) -> Result<(), SubprocessClusterError> {
+        self.send(Signal::SIGSTOP)?;
+
+        match waitpid(self.pid, Some(WaitPidFlag::WUNTRACED)) {
+            Ok(WaitStatus::Stopped(..)) => Ok(()),
+            Ok(status) => Err(SubprocessClusterError::ProcessDidNotStop {
+                pid: self.pid,
+                status,
+            }),
+            Err(source) => Err(SubprocessClusterError::ProcessUnobservable {
+                pid: self.pid,
+                source,
+            }),
+        }
+    }
+
+    pub fn resume(&self) -> Result<(), SubprocessClusterError> {
+        self.send(Signal::SIGCONT)
+    }
+
+    pub fn kill(&self) -> Result<(), SubprocessClusterError> {
+        self.kill_requested.store(true, Ordering::Release);
+
+        self.send(Signal::SIGKILL)
+    }
+
+    pub fn terminate(&self) -> Result<(), SubprocessClusterError> {
+        match self.send(Signal::SIGTERM) {
+            Ok(())
+            | Err(SubprocessClusterError::SignalUndeliverable {
+                source: Errno::ESRCH,
+                ..
+            }) => Ok(()),
+            Err(signal_error) => Err(signal_error),
+        }
+    }
+
+    fn send(&self, signal: Signal) -> Result<(), SubprocessClusterError> {
+        kill(self.pid, signal).map_err(|source| SubprocessClusterError::SignalUndeliverable {
+            pid: self.pid,
+            signal,
+            source,
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::error::Error as _;
+    use std::io::BufRead as _;
+    use std::io::BufReader;
+    use std::process::Command;
+    use std::process::Stdio;
+    use std::sync::Arc;
+
+    use nix::errno::Errno;
+    use nix::sys::signal::Signal;
+    use nix::sys::signal::kill;
+    use nix::unistd::Pid;
+    use nix::unistd::Uid;
+
+    use super::SubprocessSignals;
+    use crate::subprocess_cluster_error::SubprocessClusterError;
+
+    fn signals_for(pid: Pid) -> SubprocessSignals {
+        SubprocessSignals {
+            kill_requested: Arc::default(),
+            pid,
+        }
+    }
+
+    #[test]
+    fn terminating_a_process_this_user_may_not_signal_is_refused() {
+        assert!(
+            !Uid::effective().is_root(),
+            "only an unprivileged user is refused when signalling the init process"
+        );
+
+        let terminate_error = signals_for(Pid::from_raw(1))
+            .terminate()
+            .expect_err("signalling the init process must be refused");
+
+        assert_eq!(
+            terminate_error.to_string(),
+            SubprocessClusterError::SignalUndeliverable {
+                pid: Pid::from_raw(1),
+                signal: Signal::SIGTERM,
+                source: Errno::EPERM,
+            }
+            .to_string()
+        );
+        assert_eq!(
+            terminate_error
+                .source()
+                .and_then(|source| source.downcast_ref::<Errno>()),
+            Some(&Errno::EPERM)
+        );
+    }
+
+    #[test]
+    fn pausing_a_process_this_harness_did_not_spawn_is_unobservable() {
+        let mut shell = Command::new("sh")
+            .args(["-c", "sleep 60 & echo $!; wait"])
+            .stdout(Stdio::piped())
+            .spawn()
+            .unwrap();
+        let mut grandchild_pid = String::new();
+
+        BufReader::new(shell.stdout.take().unwrap())
+            .read_line(&mut grandchild_pid)
+            .unwrap();
+
+        let grandchild = Pid::from_raw(grandchild_pid.trim_end().parse().unwrap());
+        let pause_result = signals_for(grandchild).pause();
+
+        kill(grandchild, Signal::SIGKILL).unwrap();
+        shell.wait().unwrap();
+
+        let pause_error =
+            pause_result.expect_err("pausing a process this harness did not spawn must fail");
+
+        assert_eq!(
+            pause_error.to_string(),
+            SubprocessClusterError::ProcessUnobservable {
+                pid: grandchild,
+                source: Errno::ECHILD,
+            }
+            .to_string()
+        );
+        assert_eq!(
+            pause_error
+                .source()
+                .and_then(|source| source.downcast_ref::<Errno>()),
+            Some(&Errno::ECHILD)
+        );
+    }
+}
diff --git a/paddler_cli_tests/src/terminate_child.rs b/paddler_cli_tests/src/terminate_child.rs
deleted file mode 100644
index 2bfa3120c..000000000
--- a/paddler_cli_tests/src/terminate_child.rs
+++ /dev/null
@@ -1,44 +0,0 @@
-use anyhow::Context as _;
-use anyhow::Error;
-use anyhow::Result;
-#[cfg(unix)]
-use nix::errno::Errno;
-#[cfg(unix)]
-use nix::sys::signal::Signal;
-#[cfg(unix)]
-use nix::sys::signal::kill;
-#[cfg(unix)]
-use nix::unistd::Pid;
-use tokio::process::Child;
-
-#[cfg(unix)]
-use crate::subprocess_cluster_error::SubprocessClusterError;
-
-#[cfg(unix)]
-pub fn terminate_child(child: &mut Child) -> Result<()> {
-    let Some(raw_pid) = child.id() else {
-        return Ok(());
-    };
-
-    let pid = Pid::from_raw(
-        i32::try_from(raw_pid)
-            .map_err(|source| SubprocessClusterError::ProcessIdOutOfRange { raw_pid, source })?,
-    );
-
-    match kill(pid, Signal::SIGTERM) {
-        Ok(()) | Err(Errno::ESRCH) => Ok(()),
-        Err(errno) => Err(Error::new(errno))
-            .with_context(|| format!("failed to send SIGTERM to process {raw_pid}")),
-    }
-}
-
-#[cfg(windows)]
-pub fn terminate_child(child: &mut Child) -> Result<()> {
-    if child.id().is_none() {
-        return Ok(());
-    }
-
-    child
-        .start_kill()
-        .context("failed to terminate child process")
-}
diff --git a/paddler_cli_tests/src/unreachable_management_addr.rs b/paddler_cli_tests/src/unreachable_management_addr.rs
new file mode 100644
index 000000000..8f3371d28
--- /dev/null
+++ b/paddler_cli_tests/src/unreachable_management_addr.rs
@@ -0,0 +1,6 @@
+use std::net::Ipv4Addr;
+use std::net::SocketAddr;
+use std::net::SocketAddrV4;
+
+pub const UNREACHABLE_MANAGEMENT_ADDR: SocketAddr =
+    SocketAddr::V4(SocketAddrV4::new(Ipv4Addr::LOCALHOST, 1));
diff --git a/paddler_cli_tests/tests/a_balancer_drops_an_agent_that_stops_answering_pings.rs b/paddler_cli_tests/tests/a_balancer_drops_an_agent_that_stops_answering_pings.rs
new file mode 100644
index 000000000..43fa7d84e
--- /dev/null
+++ b/paddler_cli_tests/tests/a_balancer_drops_an_agent_that_stops_answering_pings.rs
@@ -0,0 +1,37 @@
+use paddler_cli_tests::pausable_agent_cluster::PausableAgentCluster;
+use paddler_cli_tests::pausable_agent_cluster_params::PausableAgentClusterParams;
+use paddler_cli_tests::start_pausable_agent_cluster::start_pausable_agent_cluster;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn a_balancer_drops_an_agent_that_stops_answering_pings() {
+    let PausableAgentCluster {
+        mut cluster,
+        pausable_agent,
+    } = start_pausable_agent_cluster(PausableAgentClusterParams {
+        binary_path: env!("CARGO_BIN_EXE_paddler_cluster_node").to_owned(),
+        desired_state: ClusterDesiredState::KeepStored,
+        expected_slots_total: 0,
+        slot_count: 1,
+    })
+    .await
+    .expect("a cluster with a model-less agent must start");
+
+    pausable_agent
+        .signals
+        .pause()
+        .expect("the agent must stop answering pings");
+    cluster
+        .wait_for_agent_count(0)
+        .await
+        .expect("the balancer must drop the agent that stopped answering pings");
+
+    pausable_agent
+        .signals
+        .resume()
+        .expect("the agent must continue so that it can shut down");
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_cli_tests/tests/a_balancer_stops_with_its_statsd_service_failed_when_it_cannot_open_a_udp_socket.rs b/paddler_cli_tests/tests/a_balancer_stops_with_its_statsd_service_failed_when_it_cannot_open_a_udp_socket.rs
new file mode 100644
index 000000000..2bff30a61
--- /dev/null
+++ b/paddler_cli_tests/tests/a_balancer_stops_with_its_statsd_service_failed_when_it_cannot_open_a_udp_socket.rs
@@ -0,0 +1,87 @@
+use std::error::Error as _;
+use std::io;
+
+use libc::EPERM;
+use libc::SOCK_CLOEXEC;
+use libc::SOCK_DGRAM;
+use libc::SOCK_NONBLOCK;
+use libc::SYS_socket;
+use seccompiler::SeccompCmpArgLen;
+use seccompiler::SeccompCmpOp;
+use seccompiler::SeccompCondition;
+use seccompiler::SeccompRule;
+use trzcina::ServiceShutdownError;
+use trzcina::ServiceShutdownOutcome;
+use trzcina::ServiceShutdownOutcomeWithServiceName;
+
+use paddler_balancer::resolved_socket_addr::ResolvedSocketAddr;
+use paddler_balancer::statsd_service::configuration::Configuration as StatsdServiceConfiguration;
+use paddler_bootstrap::balancer_bootstrap_config::BalancerBootstrapConfig;
+use paddler_bootstrap::balancer_runner::BalancerRunner;
+use paddler_bootstrap::balancer_runner_params::BalancerRunnerParams;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_test_cluster_harness::ephemeral_loopback_addr::EPHEMERAL_LOOPBACK_ADDR;
+use paddler_test_cluster_harness::longer_than_any_test_run::LONGER_THAN_ANY_TEST_RUN;
+use paddler_tests::in_process_balancer_runner_params::in_process_balancer_runner_params;
+
+use crate::refused_syscall::RefusedSyscall;
+use crate::refused_syscalls::RefusedSyscalls;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn a_balancer_stops_with_its_statsd_service_failed_when_it_cannot_open_a_udp_socket() {
+    RefusedSyscalls {
+        errno: EPERM,
+        refused_syscalls: vec![RefusedSyscall {
+            syscall: SYS_socket,
+            rules: vec![
+                SeccompRule::new(vec![
+                    SeccompCondition::new(
+                        1,
+                        SeccompCmpArgLen::Dword,
+                        SeccompCmpOp::MaskedEq(u64::from(
+                            (!(SOCK_CLOEXEC | SOCK_NONBLOCK)).cast_unsigned(),
+                        )),
+                        u64::from(SOCK_DGRAM.cast_unsigned()),
+                    )
+                    .expect("the datagram socket condition must be valid"),
+                ])
+                .expect("the datagram socket rule must be valid"),
+            ],
+        }],
+    }
+    .apply_to_every_thread();
+
+    let balancer_runner_params =
+        in_process_balancer_runner_params(&ClusterParams::without_agents())
+            .expect("the balancer parameters must be valid");
+    let mut balancer_runner = BalancerRunner::start(BalancerRunnerParams {
+        bootstrap_config: BalancerBootstrapConfig {
+            statsd_service_configuration: Some(StatsdServiceConfiguration {
+                statsd_addr: ResolvedSocketAddr::from(EPHEMERAL_LOOPBACK_ADDR),
+                statsd_reporting_interval: LONGER_THAN_ANY_TEST_RUN,
+            }),
+            ..balancer_runner_params.bootstrap_config
+        },
+        ..balancer_runner_params
+    })
+    .await
+    .expect("the balancer must bind its listeners");
+
+    assert_eq!(
+        balancer_runner
+            .wait_for_completion()
+            .await
+            .expect_err("a balancer whose statsd service fails must stop with an error")
+            .source()
+            .map(ToString::to_string),
+        Some(
+            ServiceShutdownError::new(vec![ServiceShutdownOutcomeWithServiceName {
+                name: "balancer::statsd_service",
+                outcome: ServiceShutdownOutcome::Errored(
+                    io::Error::from_raw_os_error(EPERM).into()
+                ),
+            }])
+            .to_string()
+        )
+    );
+}
diff --git a/paddler_cli_tests/tests/a_sampler_chain_that_cannot_allocate_its_penalty_window_fails_the_generation_preparation.rs b/paddler_cli_tests/tests/a_sampler_chain_that_cannot_allocate_its_penalty_window_fails_the_generation_preparation.rs
new file mode 100644
index 000000000..c0ef13618
--- /dev/null
+++ b/paddler_cli_tests/tests/a_sampler_chain_that_cannot_allocate_its_penalty_window_fails_the_generation_preparation.rs
@@ -0,0 +1,116 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use libc::ENOMEM;
+use libc::SYS_brk;
+use libc::SYS_mmap;
+use llama_cpp_bindings::error::SamplingError;
+use seccompiler::SeccompCmpArgLen;
+use seccompiler::SeccompCmpOp;
+use seccompiler::SeccompCondition;
+use seccompiler::SeccompRule;
+
+use paddler_agent_text_generation::conversation_support::ConversationSupport;
+use paddler_agent_text_generation::generation_request_preparer::GenerationRequestPreparer;
+use paddler_agent_text_generation::model_bos_token::ModelBosToken;
+use paddler_agent_text_generation::prompt_tokenizer::PromptTokenizer;
+use paddler_agent_text_generation::sampler_chain_factory::SamplerChainFactory;
+use paddler_agent_text_generation::text_generation_error::TextGenerationError;
+use paddler_inference_parameters::sampling_parameters::SamplingParameters;
+use paddler_messaging::agent_issue_params::model_path::ModelPath;
+use paddler_messaging::multimodal_settings::MultimodalSettings;
+use paddler_tests::agent_inference_runtime_context::agent_inference_runtime_context;
+use paddler_tests::kev_checkpoint::KevCheckpoint;
+use paddler_tests::llama_runtime::LlamaRuntime;
+use paddler_tests::raw_prompt_generation::RawPromptGeneration;
+
+use crate::refused_syscall::RefusedSyscall;
+use crate::refused_syscalls::RefusedSyscalls;
+
+const PENALTY_WINDOW_OF_FOUR_GIBIBYTES_OF_TOKENS: i32 = 1 << 30;
+const SMALLEST_REFUSED_MAPPING_LENGTH: u64 = 1 << 32;
+const ERRNO_THAT_MAKES_BRK_REPORT_A_NULL_BREAK: i32 = 0;
+
+#[test]
+fn a_sampler_chain_that_cannot_allocate_its_penalty_window_fails_the_generation_preparation() {
+    let model_path = KevCheckpoint::Kev0_8b.target_path("model.gguf");
+    let llama_runtime = LlamaRuntime::load(&model_path, agent_inference_runtime_context(1))
+        .expect("the kev model must load");
+    let sequence_context_size = llama_runtime
+        .create_text_generation_llama_context(1)
+        .expect("the context must be created")
+        .n_ctx_seq();
+    let model = &llama_runtime.loaded_llama_model.model;
+    let request_preparer = GenerationRequestPreparer {
+        conversation_support: ConversationSupport::ChatTemplateMissing(ModelPath {
+            model_path: model_path.display().to_string(),
+        }),
+        image_resize_to_fit: MultimodalSettings::default().image_resize_to_fit,
+        loaded_llama_model: llama_runtime.loaded_llama_model.clone(),
+        prompt_tokenizer: PromptTokenizer {
+            model: model.clone(),
+            model_bos_token: ModelBosToken::of(model),
+            sequence_context_size,
+        },
+        sampler_chain_factory: SamplerChainFactory {
+            sampling_parameters: SamplingParameters {
+                penalty_last_n: PENALTY_WINDOW_OF_FOUR_GIBIBYTES_OF_TOKENS,
+                penalty_repeat: 1.1,
+                ..SamplingParameters::default()
+            },
+            n_vocab: model.n_vocab(),
+        },
+        streaming_markers: model
+            .streaming_markers()
+            .expect("the model must report its streaming markers"),
+    };
+    let RawPromptGeneration { agent_request, .. } = RawPromptGeneration::new(
+        "Paddler",
+        llama_runtime
+            .inference_runtime_context
+            .slot_aggregated_status
+            .clone(),
+    );
+
+    RefusedSyscalls {
+        errno: ENOMEM,
+        refused_syscalls: vec![RefusedSyscall {
+            syscall: SYS_mmap,
+            rules: vec![
+                SeccompRule::new(vec![
+                    SeccompCondition::new(
+                        1,
+                        SeccompCmpArgLen::Qword,
+                        SeccompCmpOp::Ge,
+                        SMALLEST_REFUSED_MAPPING_LENGTH,
+                    )
+                    .expect("the mapping length condition must be valid"),
+                ])
+                .expect("the mapping length rule must be valid"),
+            ],
+        }],
+    }
+    .apply_to_every_thread();
+
+    RefusedSyscalls {
+        errno: ERRNO_THAT_MAKES_BRK_REPORT_A_NULL_BREAK,
+        refused_syscalls: vec![RefusedSyscall {
+            syscall: SYS_brk,
+            rules: vec![],
+        }],
+    }
+    .apply_to_every_thread();
+
+    assert_eq!(
+        request_preparer
+            .prepare_raw_prompt(agent_request)
+            .err()
+            .as_ref()
+            .map(ToString::to_string),
+        Some(
+            TextGenerationError::SamplerChainCreationFailed(SamplingError::LlamaCppOutOfMemory {
+                sampler: "penalties"
+            })
+            .to_string()
+        )
+    );
+}
diff --git a/paddler_cli_tests/tests/agent_exits_with_failure_when_it_cannot_set_up_its_model_downloader.rs b/paddler_cli_tests/tests/agent_exits_with_failure_when_it_cannot_set_up_its_model_downloader.rs
new file mode 100644
index 000000000..07d49423c
--- /dev/null
+++ b/paddler_cli_tests/tests/agent_exits_with_failure_when_it_cannot_set_up_its_model_downloader.rs
@@ -0,0 +1,26 @@
+use std::process::Stdio;
+
+use paddler_cli_tests::paddler_command::paddler_command;
+use paddler_cli_tests::unreachable_management_addr::UNREACHABLE_MANAGEMENT_ADDR;
+use paddler_test_cluster_harness::ca_bundle_without_certificates::CA_BUNDLE_WITHOUT_CERTIFICATES;
+use paddler_test_cluster_harness::failed_run_exit_code::FAILED_RUN_EXIT_CODE;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agent_exits_with_failure_when_it_cannot_set_up_its_model_downloader() {
+    assert_eq!(
+        paddler_command(env!("CARGO_BIN_EXE_paddler_cluster_node"))
+            .arg("agent")
+            .arg("--management-addr")
+            .arg(UNREACHABLE_MANAGEMENT_ADDR.to_string())
+            .arg("--slots")
+            .arg("1")
+            .env("SSL_CERT_FILE", CA_BUNDLE_WITHOUT_CERTIFICATES)
+            .env_remove("SSL_CERT_DIR")
+            .stderr(Stdio::null())
+            .status()
+            .await
+            .expect("the agent must run")
+            .code(),
+        Some(FAILED_RUN_EXIT_CODE)
+    );
+}
diff --git a/paddler_cli_tests/tests/agent_reconnects_to_its_balancer_after_the_balancer_is_killed.rs b/paddler_cli_tests/tests/agent_reconnects_to_its_balancer_after_the_balancer_is_killed.rs
new file mode 100644
index 000000000..e1794e8bb
--- /dev/null
+++ b/paddler_cli_tests/tests/agent_reconnects_to_its_balancer_after_the_balancer_is_killed.rs
@@ -0,0 +1,56 @@
+use paddler_cli_tests::restart_balancer::restart_balancer;
+use paddler_cli_tests::start_signalled_subprocess_cluster::start_signalled_subprocess_cluster;
+use paddler_cli_tests::subprocess_cluster::SubprocessCluster;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster::Cluster;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agent_reconnects_to_its_balancer_after_the_balancer_is_killed() {
+    let SubprocessCluster {
+        balancer_signals,
+        cluster,
+    } = start_signalled_subprocess_cluster(
+        env!("CARGO_BIN_EXE_paddler_cluster_node"),
+        ClusterParams {
+            agents: vec![AgentConfig::single(1)],
+            wait_for_slots_ready: false,
+            ..ClusterParams::default()
+        },
+    )
+    .await
+    .expect("a cluster with a model-less agent must start");
+
+    balancer_signals
+        .kill()
+        .expect("the balancer must be killed without closing its connections");
+
+    let Cluster {
+        agents, balancer, ..
+    } = cluster;
+    let management_addr = balancer.addresses.management;
+
+    balancer
+        .shutdown()
+        .await
+        .expect("the balancer must shut down before it restarts");
+
+    let SubprocessCluster {
+        cluster: mut restarted_cluster,
+        ..
+    } = restart_balancer(env!("CARGO_BIN_EXE_paddler_cluster_node"), management_addr)
+        .await
+        .expect("the balancer must restart on its management address");
+
+    restarted_cluster.agents = agents;
+
+    restarted_cluster
+        .wait_for_agent_count(1)
+        .await
+        .expect("the agent must reconnect to the restarted balancer");
+
+    restarted_cluster
+        .shutdown()
+        .await
+        .expect("the restarted cluster must shut down cleanly");
+}
diff --git a/paddler_cli_tests/tests/agent_reconnects_to_its_balancer_after_the_balancer_restarts.rs b/paddler_cli_tests/tests/agent_reconnects_to_its_balancer_after_the_balancer_restarts.rs
new file mode 100644
index 000000000..d1c957f44
--- /dev/null
+++ b/paddler_cli_tests/tests/agent_reconnects_to_its_balancer_after_the_balancer_restarts.rs
@@ -0,0 +1,49 @@
+use paddler_cli_tests::restart_balancer::restart_balancer;
+use paddler_cli_tests::start_subprocess_cluster::start_subprocess_cluster;
+use paddler_cli_tests::subprocess_cluster::SubprocessCluster;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster::Cluster;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agent_reconnects_to_its_balancer_after_the_balancer_restarts() {
+    let cluster = start_subprocess_cluster(
+        env!("CARGO_BIN_EXE_paddler_cluster_node"),
+        ClusterParams {
+            agents: vec![AgentConfig::single(1)],
+            wait_for_slots_ready: false,
+            ..ClusterParams::default()
+        },
+    )
+    .await
+    .expect("a cluster with a model-less agent must start");
+
+    let Cluster {
+        agents, balancer, ..
+    } = cluster;
+    let management_addr = balancer.addresses.management;
+
+    balancer
+        .shutdown()
+        .await
+        .expect("the balancer must shut down before it restarts");
+
+    let SubprocessCluster {
+        cluster: mut restarted_cluster,
+        ..
+    } = restart_balancer(env!("CARGO_BIN_EXE_paddler_cluster_node"), management_addr)
+        .await
+        .expect("the balancer must restart on its management address");
+
+    restarted_cluster.agents = agents;
+
+    restarted_cluster
+        .wait_for_agent_count(1)
+        .await
+        .expect("the agent must reconnect to the restarted balancer");
+
+    restarted_cluster
+        .shutdown()
+        .await
+        .expect("the restarted cluster must shut down cleanly");
+}
diff --git a/paddler_cli_tests/tests/agent_shuts_down_cleanly_while_its_balancer_is_unresponsive.rs b/paddler_cli_tests/tests/agent_shuts_down_cleanly_while_its_balancer_is_unresponsive.rs
new file mode 100644
index 000000000..013ce2604
--- /dev/null
+++ b/paddler_cli_tests/tests/agent_shuts_down_cleanly_while_its_balancer_is_unresponsive.rs
@@ -0,0 +1,41 @@
+use paddler_cli_tests::start_signalled_subprocess_cluster::start_signalled_subprocess_cluster;
+use paddler_cli_tests::subprocess_cluster::SubprocessCluster;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agent_shuts_down_cleanly_while_its_balancer_is_unresponsive() {
+    let SubprocessCluster {
+        balancer_signals,
+        mut cluster,
+    } = start_signalled_subprocess_cluster(
+        env!("CARGO_BIN_EXE_paddler_cluster_node"),
+        ClusterParams {
+            agents: AgentConfig::uniform(1, 1),
+            wait_for_slots_ready: false,
+            ..ClusterParams::default()
+        },
+    )
+    .await
+    .expect("a cluster with a model-less agent must start");
+
+    balancer_signals
+        .pause()
+        .expect("the balancer must stop responding");
+
+    cluster
+        .agents
+        .pop()
+        .expect("the cluster must have an agent")
+        .shutdown()
+        .await
+        .expect("the agent must shut down cleanly while its balancer does not respond");
+
+    balancer_signals
+        .resume()
+        .expect("the balancer must continue so that it can shut down");
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_cli_tests/tests/agent_socket_of_a_vanished_balancer.rs b/paddler_cli_tests/tests/agent_socket_of_a_vanished_balancer.rs
new file mode 100644
index 000000000..937eabeab
--- /dev/null
+++ b/paddler_cli_tests/tests/agent_socket_of_a_vanished_balancer.rs
@@ -0,0 +1,52 @@
+use paddler_cli_tests::start_signalled_subprocess_cluster::start_signalled_subprocess_cluster;
+use paddler_cli_tests::subprocess_cluster::SubprocessCluster;
+use paddler_test_cluster_harness::cluster::Cluster;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_test_cluster_harness::untrusted_agent_socket_client::UntrustedAgentSocketClient;
+
+const AGENT_ID_OF_A_VANISHED_BALANCER: &str = "agent-of-a-vanished-balancer";
+
+pub struct AgentSocketOfAVanishedBalancer {
+    pub agent_socket: UntrustedAgentSocketClient,
+    pub cluster: Cluster,
+}
+
+impl AgentSocketOfAVanishedBalancer {
+    pub async fn connect() -> Self {
+        let SubprocessCluster {
+            balancer_signals,
+            cluster,
+        } = start_signalled_subprocess_cluster(
+            env!("CARGO_BIN_EXE_paddler_cluster_node"),
+            ClusterParams::without_agents(),
+        )
+        .await
+        .expect("a cluster without agents must start");
+        let mut agent_socket = UntrustedAgentSocketClient::connect(
+            cluster.balancer.addresses.management,
+            AGENT_ID_OF_A_VANISHED_BALANCER,
+        )
+        .await
+        .expect("the agent socket must reach the balancer");
+
+        agent_socket
+            .next_text()
+            .await
+            .expect("the balancer must announce its version before it is killed");
+        balancer_signals
+            .kill()
+            .expect("the balancer must be killed");
+
+        Self {
+            agent_socket,
+            cluster,
+        }
+    }
+
+    pub async fn shut_down(self) {
+        self.cluster
+            .shutdown()
+            .await
+            .expect("the cluster must shut down after its balancer was killed");
+    }
+}
diff --git a/paddler_cli_tests/tests/an_agent_reconnects_when_the_balancer_stops_answering_pings.rs b/paddler_cli_tests/tests/an_agent_reconnects_when_the_balancer_stops_answering_pings.rs
new file mode 100644
index 000000000..f6fb221f1
--- /dev/null
+++ b/paddler_cli_tests/tests/an_agent_reconnects_when_the_balancer_stops_answering_pings.rs
@@ -0,0 +1,56 @@
+use std::net::Ipv4Addr;
+use std::net::SocketAddr;
+
+use tokio::net::TcpListener;
+
+use paddler_cli_tests::pausable_agent::PausableAgent;
+use paddler_cli_tests::pausable_agent_params::PausableAgentParams;
+use paddler_cli_tests::stallable_tcp_relay::StallableTcpRelay;
+use paddler_cli_tests::start_subprocess_cluster::start_subprocess_cluster;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn an_agent_reconnects_when_the_balancer_stops_answering_pings() {
+    let mut cluster = start_subprocess_cluster(
+        env!("CARGO_BIN_EXE_paddler_cluster_node"),
+        ClusterParams::without_agents(),
+    )
+    .await
+    .expect("a cluster without agents must start");
+    let relay_listener = TcpListener::bind(SocketAddr::from((Ipv4Addr::LOCALHOST, 0)))
+        .await
+        .expect("the relay must bind a loopback port");
+    let relay_addr = relay_listener
+        .local_addr()
+        .expect("the relay listener must know its address");
+    let relay = StallableTcpRelay::start(relay_listener, cluster.balancer.addresses.management);
+
+    PausableAgent::join(
+        &mut cluster,
+        PausableAgentParams {
+            binary_path: env!("CARGO_BIN_EXE_paddler_cluster_node").to_owned(),
+            config: AgentConfig::single(1),
+            expected_slots_total: 0,
+            management_addr: relay_addr,
+        },
+    )
+    .await
+    .expect("the agent must register through the relay");
+
+    relay.stall_established_connections();
+    relay
+        .wait_for_accepted_connections(2)
+        .await
+        .expect("the agent must reconnect once the balancer stops answering its pings");
+
+    relay
+        .stop()
+        .await
+        .expect("the relay task must not panic")
+        .expect("the relay must stop without a relaying failure");
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_cli_tests/tests/an_http_service_reports_a_listener_it_cannot_make_nonblocking.rs b/paddler_cli_tests/tests/an_http_service_reports_a_listener_it_cannot_make_nonblocking.rs
new file mode 100644
index 000000000..ba0f21110
--- /dev/null
+++ b/paddler_cli_tests/tests/an_http_service_reports_a_listener_it_cannot_make_nonblocking.rs
@@ -0,0 +1,68 @@
+use std::io;
+use std::net::Ipv4Addr;
+use std::net::SocketAddr;
+
+use actix_web::App;
+use libc::EPERM;
+use libc::FIONBIO;
+use libc::SYS_ioctl;
+use seccompiler::SeccompCmpArgLen;
+use seccompiler::SeccompCmpOp;
+use seccompiler::SeccompCondition;
+use seccompiler::SeccompRule;
+use tokio_util::sync::CancellationToken;
+
+use paddler_balancer::http_listener::HttpListener;
+use paddler_balancer::run_http_service::run_http_service;
+use paddler_balancer::run_http_service_parameters::RunHttpServiceParameters;
+
+use crate::refused_syscall::RefusedSyscall;
+use crate::refused_syscalls::RefusedSyscalls;
+
+const SERVICE_NAME: &str = "balancer::service_that_cannot_listen";
+
+#[actix_web::test]
+async fn an_http_service_reports_a_listener_it_cannot_make_nonblocking() {
+    let http_listener = HttpListener::bind(SocketAddr::from((Ipv4Addr::LOCALHOST, 0)))
+        .expect("the listener must bind before nonblocking mode is refused");
+    let local_addr = http_listener.local_addr;
+
+    RefusedSyscalls {
+        errno: EPERM,
+        refused_syscalls: vec![RefusedSyscall {
+            syscall: SYS_ioctl,
+            rules: vec![
+                SeccompRule::new(vec![
+                    SeccompCondition::new(1, SeccompCmpArgLen::Dword, SeccompCmpOp::Eq, FIONBIO)
+                        .expect("the nonblocking request condition must be valid"),
+                ])
+                .expect("the nonblocking request rule must be valid"),
+            ],
+        }],
+    }
+    .apply_to_every_thread();
+
+    let service_error = run_http_service(
+        CancellationToken::new(),
+        RunHttpServiceParameters {
+            app_factory: App::new,
+            http_listener,
+            service_name: SERVICE_NAME,
+            worker_count: 1,
+        },
+    )
+    .await
+    .expect_err("a listener that cannot become nonblocking must not be served");
+
+    assert_eq!(
+        service_error.to_string(),
+        format!("Unable to listen for {SERVICE_NAME} on {local_addr}")
+    );
+    assert_eq!(
+        service_error
+            .root_cause()
+            .downcast_ref::<io::Error>()
+            .and_then(io::Error::raw_os_error),
+        Some(EPERM)
+    );
+}
diff --git a/paddler_cli_tests/tests/balancer_answers_gateway_timeout_when_a_paused_agent_never_answers.rs b/paddler_cli_tests/tests/balancer_answers_gateway_timeout_when_a_paused_agent_never_answers.rs
new file mode 100644
index 000000000..c0f3ba786
--- /dev/null
+++ b/paddler_cli_tests/tests/balancer_answers_gateway_timeout_when_a_paused_agent_never_answers.rs
@@ -0,0 +1,60 @@
+use reqwest::StatusCode;
+use tokio_util::sync::CancellationToken;
+
+use paddler_cli_tests::pausable_agent_cluster::PausableAgentCluster;
+use paddler_cli_tests::pausable_agent_cluster_params::PausableAgentClusterParams;
+use paddler_cli_tests::start_pausable_agent_cluster::start_pausable_agent_cluster;
+use paddler_client::error::Error as ClientError;
+use paddler_messaging::api_path::ApiPath;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn balancer_answers_gateway_timeout_when_a_paused_agent_never_answers() {
+    let PausableAgentCluster {
+        cluster,
+        pausable_agent,
+    } = start_pausable_agent_cluster(PausableAgentClusterParams {
+        binary_path: env!("CARGO_BIN_EXE_paddler_cluster_node").to_owned(),
+        desired_state: ClusterDesiredState::KeepStored,
+        expected_slots_total: 0,
+        slot_count: 1,
+    })
+    .await
+    .expect("a cluster with a model-less agent must start");
+
+    pausable_agent
+        .signals
+        .pause()
+        .expect("the agent must stop before it is asked anything");
+
+    let model_metadata_url = cluster
+        .balancer
+        .management_base_url
+        .join(&ApiPath::agent_model_metadata(&pausable_agent.id))
+        .expect("the model metadata path must extend the management address");
+    let metadata_result = cluster
+        .client_management
+        .get_model_metadata(CancellationToken::new(), &pausable_agent.id)
+        .await;
+
+    assert_eq!(
+        metadata_result.err().as_ref().map(ToString::to_string),
+        Some(
+            ClientError::UnexpectedResponseStatus {
+                message: String::new(),
+                status: StatusCode::GATEWAY_TIMEOUT,
+                url: model_metadata_url.to_string(),
+            }
+            .to_string()
+        )
+    );
+
+    pausable_agent
+        .signals
+        .resume()
+        .expect("the agent must continue so that it can shut down");
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_cli_tests/tests/balancer_dispatches_requests_only_to_agents_that_serve.rs b/paddler_cli_tests/tests/balancer_dispatches_requests_only_to_agents_that_serve.rs
new file mode 100644
index 000000000..850de0d54
--- /dev/null
+++ b/paddler_cli_tests/tests/balancer_dispatches_requests_only_to_agents_that_serve.rs
@@ -0,0 +1,43 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::num::NonZeroU32;
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
+
+use crate::subprocess_cluster_with_a_serving_and_a_stuck_agent::SERVING_AGENT_NAME;
+use crate::subprocess_cluster_with_a_serving_and_a_stuck_agent::subprocess_cluster_with_a_serving_and_a_stuck_agent;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn balancer_dispatches_requests_only_to_agents_that_serve() {
+    let cluster = subprocess_cluster_with_a_serving_and_a_stuck_agent().await;
+
+    let collected = cluster
+        .continue_from_raw_prompt(
+            CancellationToken::new(),
+            &ContinueFromRawPromptParams {
+                grammar: None,
+                max_tokens: NonZeroU32::MIN,
+                raw_prompt: "The capital of France is".to_owned(),
+            },
+        )
+        .await
+        .expect("the generation must be answered");
+
+    assert!(
+        collected
+            .token_results
+            .iter()
+            .all(|token_result| token_result.generated_by.as_deref() == Some(SERVING_AGENT_NAME))
+    );
+
+    collected
+        .summary()
+        .expect("the serving agent must finish the generation");
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_cli_tests/tests/balancer_distributes_buffered_requests_across_two_agents.rs b/paddler_cli_tests/tests/balancer_distributes_buffered_requests_across_two_agents.rs
deleted file mode 100644
index 5c42d368f..000000000
--- a/paddler_cli_tests/tests/balancer_distributes_buffered_requests_across_two_agents.rs
+++ /dev/null
@@ -1,82 +0,0 @@
-#![cfg(feature = "tests_that_use_llms")]
-
-use std::num::NonZeroU32;
-
-use futures_util::StreamExt as _;
-use tokio_util::sync::CancellationToken;
-
-use paddler_cli_tests::start_subprocess_cluster::start_subprocess_cluster;
-use paddler_messaging::inference_client::message::Message;
-use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
-use paddler_test_cluster_harness::agent_config::AgentConfig;
-use paddler_test_cluster_harness::cluster_params::ClusterParams;
-use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn balancer_distributes_buffered_requests_across_two_agents() {
-    let cluster = start_subprocess_cluster(
-        env!("CARGO_BIN_EXE_paddler_cluster_node"),
-        ClusterParams {
-            agents: vec![
-                AgentConfig {
-                    name: "distributed-agent-0".to_owned(),
-                    slot_count: 2,
-                },
-                AgentConfig {
-                    name: "distributed-agent-1".to_owned(),
-                    slot_count: 2,
-                },
-            ],
-            wait_for_slots_ready: true,
-            max_buffered_requests: 10,
-            desired_state: Some(qwen3_0_6b().into_desired_state()),
-            ..ClusterParams::default()
-        },
-    )
-    .await
-    .expect("the cluster must start");
-
-    let mut streams = Vec::new();
-
-    for _ in 0..5 {
-        let stream = cluster
-            .continue_from_raw_prompt_stream(
-                CancellationToken::new(),
-                &ContinueFromRawPromptParams {
-                    grammar: None,
-                    max_tokens: NonZeroU32::new(10).unwrap(),
-                    raw_prompt: "Hello".to_owned(),
-                },
-            )
-            .await
-            .expect("the inference request must be accepted");
-
-        streams.push(stream);
-    }
-
-    let mut successful_responses = 0;
-
-    for mut stream in streams {
-        if let Some(item) = stream.next().await {
-            match item.expect("the message must be readable") {
-                Message::Response(_) => successful_responses += 1,
-                Message::Error(envelope) => {
-                    panic!(
-                        "expected success, got error {}: {}",
-                        envelope.error.code, envelope.error.description
-                    );
-                }
-                Message::Notification(_) => {
-                    panic!("unexpected token-generation-mode notification");
-                }
-            }
-        }
-    }
-
-    assert_eq!(successful_responses, 5);
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_cli_tests/tests/balancer_distributes_embedding_batch_across_agents.rs b/paddler_cli_tests/tests/balancer_distributes_embedding_batch_across_agents.rs
index f7ad3fff2..432891f5b 100644
--- a/paddler_cli_tests/tests/balancer_distributes_embedding_batch_across_agents.rs
+++ b/paddler_cli_tests/tests/balancer_distributes_embedding_batch_across_agents.rs
@@ -5,9 +5,9 @@ use std::collections::BTreeSet;
 use tokio_util::sync::CancellationToken;
 
 use paddler_cli_tests::start_subprocess_embedding_cluster::start_subprocess_embedding_cluster;
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
 use paddler_messaging::embedding_input_document::EmbeddingInputDocument;
 use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
+use paddler_messaging::embedding_result::EmbeddingResult;
 use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
 use paddler_test_cluster_harness::embedding_cluster_params::EmbeddingClusterParams;
@@ -18,10 +18,7 @@ async fn balancer_distributes_embedding_batch_across_agents() {
         env!("CARGO_BIN_EXE_paddler_cluster_node"),
         EmbeddingClusterParams {
             agents: AgentConfig::uniform(2, 4),
-            inference_parameters: InferenceParameters {
-                enable_embeddings: true,
-                ..InferenceParameters::deterministic()
-            },
+
             ..EmbeddingClusterParams::default()
         },
     )
@@ -47,7 +44,7 @@ async fn balancer_distributes_embedding_batch_across_agents() {
 
     assert_eq!(collected.embeddings.len(), 12);
     assert!(collected.saw_done);
-    assert!(collected.errors.is_empty());
+    assert_eq!(collected.failures, Vec::<EmbeddingResult>::new());
 
     let producers: BTreeSet<&str> = collected
         .embeddings
diff --git a/paddler_cli_tests/tests/balancer_distributes_embedding_batch_across_agents_with_uneven_slots.rs b/paddler_cli_tests/tests/balancer_distributes_embedding_batch_across_agents_with_uneven_slots.rs
index b2abb0bda..e745f86be 100644
--- a/paddler_cli_tests/tests/balancer_distributes_embedding_batch_across_agents_with_uneven_slots.rs
+++ b/paddler_cli_tests/tests/balancer_distributes_embedding_batch_across_agents_with_uneven_slots.rs
@@ -5,9 +5,9 @@ use std::collections::BTreeSet;
 use tokio_util::sync::CancellationToken;
 
 use paddler_cli_tests::start_subprocess_embedding_cluster::start_subprocess_embedding_cluster;
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
 use paddler_messaging::embedding_input_document::EmbeddingInputDocument;
 use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
+use paddler_messaging::embedding_result::EmbeddingResult;
 use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
 use paddler_test_cluster_harness::embedding_cluster_params::EmbeddingClusterParams;
@@ -35,10 +35,7 @@ async fn balancer_distributes_embedding_batch_across_agents_with_uneven_slots()
                     slot_count: 1,
                 },
             ],
-            inference_parameters: InferenceParameters {
-                enable_embeddings: true,
-                ..InferenceParameters::deterministic()
-            },
+
             ..EmbeddingClusterParams::default()
         },
     )
@@ -65,7 +62,7 @@ async fn balancer_distributes_embedding_batch_across_agents_with_uneven_slots()
 
     assert_eq!(collected.embeddings.len(), 8);
     assert!(collected.saw_done);
-    assert!(collected.errors.is_empty());
+    assert_eq!(collected.failures, Vec::<EmbeddingResult>::new());
 
     let returned_document_ids: BTreeSet<String> = collected
         .embeddings
diff --git a/paddler_cli_tests/tests/balancer_distributes_embedding_burst_evenly_across_agents.rs b/paddler_cli_tests/tests/balancer_distributes_embedding_burst_evenly_across_agents.rs
deleted file mode 100644
index 2b7e29189..000000000
--- a/paddler_cli_tests/tests/balancer_distributes_embedding_burst_evenly_across_agents.rs
+++ /dev/null
@@ -1,83 +0,0 @@
-#![cfg(feature = "tests_that_use_llms")]
-
-use std::collections::BTreeSet;
-
-use futures_util::future;
-use tokio_util::sync::CancellationToken;
-
-use paddler_cli_tests::start_subprocess_embedding_cluster::start_subprocess_embedding_cluster;
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
-use paddler_messaging::embedding_input_document::EmbeddingInputDocument;
-use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
-use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
-use paddler_test_cluster_harness::agent_config::AgentConfig;
-use paddler_test_cluster_harness::embedding_cluster_params::EmbeddingClusterParams;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn balancer_distributes_embedding_burst_evenly_across_agents() {
-    const AGENT_COUNT: usize = 4;
-    const SLOTS_PER_AGENT: u16 = 2;
-    const CONCURRENT_REQUESTS: usize = 8;
-
-    let cluster = start_subprocess_embedding_cluster(
-        env!("CARGO_BIN_EXE_paddler_cluster_node"),
-        EmbeddingClusterParams {
-            agents: AgentConfig::uniform(AGENT_COUNT, SLOTS_PER_AGENT),
-            inference_parameters: InferenceParameters {
-                enable_embeddings: true,
-                ..InferenceParameters::deterministic()
-            },
-            max_buffered_requests: 32,
-            ..EmbeddingClusterParams::default()
-        },
-    )
-    .await
-    .expect("the cluster must start");
-
-    let collection_futures = (0..CONCURRENT_REQUESTS).map(|request_index| {
-        let input_batch: Vec<EmbeddingInputDocument> = (0..4)
-            .map(|document_index| EmbeddingInputDocument {
-                content: format!(
-                    "Burst request {request_index}, document {document_index}: \
-                     provide an embedding for evaluation."
-                ),
-                id: format!("req-{request_index}-doc-{document_index}"),
-            })
-            .collect();
-
-        cluster.generate_embedding_batch(
-            CancellationToken::new(),
-            &GenerateEmbeddingBatchParams {
-                input_batch,
-                normalization_method: EmbeddingNormalizationMethod::None,
-            },
-        )
-    });
-
-    let collected_streams = future::try_join_all(collection_futures)
-        .await
-        .expect("every concurrent request must succeed");
-
-    let producers_across_streams: BTreeSet<&str> = collected_streams
-        .iter()
-        .flat_map(|collected| collected.embeddings.iter())
-        .filter_map(|produced| produced.generated_by.as_deref())
-        .collect();
-
-    assert_eq!(
-        producers_across_streams.len(),
-        AGENT_COUNT,
-        "burst of {CONCURRENT_REQUESTS} embedding batches across {AGENT_COUNT} agents must reach every agent, but saw producers: {producers_across_streams:?}",
-    );
-
-    for collected in &collected_streams {
-        assert!(collected.saw_done);
-        assert!(collected.errors.is_empty());
-        assert_eq!(collected.embeddings.len(), 4);
-    }
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_cli_tests/tests/balancer_distributes_token_burst_evenly_across_agents.rs b/paddler_cli_tests/tests/balancer_distributes_token_burst_evenly_across_agents.rs
index 155fe4afc..c63e83f53 100644
--- a/paddler_cli_tests/tests/balancer_distributes_token_burst_evenly_across_agents.rs
+++ b/paddler_cli_tests/tests/balancer_distributes_token_burst_evenly_across_agents.rs
@@ -4,12 +4,12 @@ use std::collections::BTreeSet;
 
 use futures_util::StreamExt as _;
 use futures_util::future;
+use futures_util::stream::iter;
 use tokio_util::sync::CancellationToken;
 
 use paddler_cli_tests::start_subprocess_cluster_with_qwen3::start_subprocess_cluster_with_qwen3;
-use paddler_messaging::inference_client::message::Message as InferenceMessage;
-use paddler_messaging::jsonrpc::response_envelope::ResponseEnvelope;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::collect_generated_tokens::collect_generated_tokens;
 use paddler_test_cluster_harness::unending_generation::unending_generation;
 
 const AGENT_COUNT: usize = 4;
@@ -17,9 +17,10 @@ const SLOTS_PER_AGENT: u16 = 1;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn balancer_distributes_token_burst_evenly_across_agents() {
+    let agent_configs = AgentConfig::uniform(AGENT_COUNT, SLOTS_PER_AGENT);
     let cluster = start_subprocess_cluster_with_qwen3(
         env!("CARGO_BIN_EXE_paddler_cluster_node"),
-        AgentConfig::uniform(AGENT_COUNT, SLOTS_PER_AGENT),
+        agent_configs.clone(),
     )
     .await
     .expect("the cluster must start");
@@ -32,31 +33,38 @@ async fn balancer_distributes_token_burst_evenly_across_agents() {
     }))
     .await
     .expect("every concurrent request must succeed");
-    let mut producers = BTreeSet::new();
+    let mut first_streamed_items = Vec::new();
 
     for stream in &mut streams {
-        match stream
-            .next()
-            .await
-            .expect("every held generation must stream a first message")
-            .expect("the message must be readable")
-        {
-            InferenceMessage::Response(ResponseEnvelope {
-                generated_by: Some(producer),
-                ..
-            }) => {
-                producers.insert(producer);
-            }
-            unexpected_message => {
-                panic!("a held generation began with {unexpected_message:?}");
-            }
-        }
+        first_streamed_items.push(
+            stream
+                .next()
+                .await
+                .expect("every held generation must stream a first message"),
+        );
     }
 
     generation_cancellation.cancel();
     drop(streams);
 
-    assert_eq!(producers.len(), AGENT_COUNT);
+    let producers =
+        future::try_join_all(first_streamed_items.into_iter().map(|first_streamed_item| {
+            collect_generated_tokens(Box::pin(iter([first_streamed_item])))
+        }))
+        .await
+        .expect("every held generation must begin with a generated token")
+        .into_iter()
+        .flat_map(|collected| collected.token_results)
+        .map(|token_result_with_producer| token_result_with_producer.generated_by)
+        .collect::<BTreeSet<_>>();
+
+    assert_eq!(
+        producers,
+        agent_configs
+            .into_iter()
+            .map(|agent_config| Some(agent_config.name))
+            .collect::<BTreeSet<_>>()
+    );
 
     cluster
         .shutdown()
diff --git a/paddler_cli_tests/tests/balancer_emits_overflow_errors_when_embedding_burst_exceeds_max_buffered_requests.rs b/paddler_cli_tests/tests/balancer_emits_overflow_errors_when_embedding_burst_exceeds_max_buffered_requests.rs
index 32a2cf7ac..c75929d59 100644
--- a/paddler_cli_tests/tests/balancer_emits_overflow_errors_when_embedding_burst_exceeds_max_buffered_requests.rs
+++ b/paddler_cli_tests/tests/balancer_emits_overflow_errors_when_embedding_burst_exceeds_max_buffered_requests.rs
@@ -5,7 +5,7 @@ use std::num::NonZeroUsize;
 use tokio_util::sync::CancellationToken;
 
 use paddler_cli_tests::start_subprocess_embedding_cluster::start_subprocess_embedding_cluster;
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
+use paddler_inference_parameters::embedding_parameters::EmbeddingParameters;
 use paddler_messaging::embedding_input_document::EmbeddingInputDocument;
 use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
 use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
@@ -20,10 +20,9 @@ async fn balancer_emits_overflow_errors_when_embedding_burst_exceeds_max_buffere
         env!("CARGO_BIN_EXE_paddler_cluster_node"),
         EmbeddingClusterParams {
             agents: AgentConfig::uniform(4, 1),
-            inference_parameters: InferenceParameters {
+            embedding_parameters: EmbeddingParameters {
                 embedding_batch_size: NonZeroUsize::MIN,
-                enable_embeddings: true,
-                ..InferenceParameters::deterministic()
+                ..EmbeddingParameters::default()
             },
             max_buffered_requests: 4,
             ..EmbeddingClusterParams::default()
@@ -71,13 +70,14 @@ async fn balancer_emits_overflow_errors_when_embedding_burst_exceeds_max_buffere
         "stream must terminate cleanly with Done even when some sub-batches overflow",
     );
 
+    let embedding_count = collected.embeddings.len();
+    let wire_error_count = collected.wire_errors.len();
+    let overflow_error_count = overflow_errors.len();
+
     assert_eq!(
-        collected.embeddings.len() + collected.wire_errors.len(),
+        embedding_count + wire_error_count,
         TOTAL_DOCUMENTS,
-        "every sub-batch must be accounted for as either a successful embedding or a wire error (503 overflow or 504 timeout): {} embeddings + {} wire errors ({} of which are 503 overflow) ≠ {TOTAL_DOCUMENTS}",
-        collected.embeddings.len(),
-        collected.wire_errors.len(),
-        overflow_errors.len(),
+        "every sub-batch must be accounted for as either a successful embedding or a wire error (503 overflow or 504 timeout): {embedding_count} embeddings + {wire_error_count} wire errors ({overflow_error_count} of which are 503 overflow) ≠ {TOTAL_DOCUMENTS}",
     );
 
     cluster
diff --git a/paddler_cli_tests/tests/balancer_exits_with_failure_when_it_cannot_announce_its_addresses.rs b/paddler_cli_tests/tests/balancer_exits_with_failure_when_it_cannot_announce_its_addresses.rs
new file mode 100644
index 000000000..c1144914b
--- /dev/null
+++ b/paddler_cli_tests/tests/balancer_exits_with_failure_when_it_cannot_announce_its_addresses.rs
@@ -0,0 +1,29 @@
+use std::fs::OpenOptions;
+use std::os::unix::process::ExitStatusExt as _;
+use std::process::ExitStatus;
+use std::process::Stdio;
+
+use paddler_cli_tests::paddler_command::paddler_command;
+use paddler_test_cluster_harness::failed_run_exit_code::FAILED_RUN_EXIT_CODE;
+
+use crate::ephemeral_balancer_arguments::ephemeral_balancer_arguments;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn balancer_exits_with_failure_when_it_cannot_announce_its_addresses() {
+    assert_eq!(
+        paddler_command(env!("CARGO_BIN_EXE_paddler_cluster_node"))
+            .arg("balancer")
+            .args(ephemeral_balancer_arguments())
+            .stdout(
+                OpenOptions::new()
+                    .write(true)
+                    .open("/dev/full")
+                    .expect("the full device must open for writing"),
+            )
+            .stderr(Stdio::null())
+            .status()
+            .await
+            .expect("the balancer must run"),
+        ExitStatus::from_raw(FAILED_RUN_EXIT_CODE << 8)
+    );
+}
diff --git a/paddler_cli_tests/tests/balancer_fans_out_embedding_batch_to_all_agents.rs b/paddler_cli_tests/tests/balancer_fans_out_embedding_batch_to_all_agents.rs
index 9addc9ba6..6d8a1dcea 100644
--- a/paddler_cli_tests/tests/balancer_fans_out_embedding_batch_to_all_agents.rs
+++ b/paddler_cli_tests/tests/balancer_fans_out_embedding_batch_to_all_agents.rs
@@ -5,9 +5,9 @@ use std::collections::BTreeSet;
 use tokio_util::sync::CancellationToken;
 
 use paddler_cli_tests::start_subprocess_embedding_cluster::start_subprocess_embedding_cluster;
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
 use paddler_messaging::embedding_input_document::EmbeddingInputDocument;
 use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
+use paddler_messaging::embedding_result::EmbeddingResult;
 use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
 use paddler_test_cluster_harness::embedding_cluster_params::EmbeddingClusterParams;
@@ -20,10 +20,7 @@ async fn balancer_fans_out_embedding_batch_to_all_agents() {
         env!("CARGO_BIN_EXE_paddler_cluster_node"),
         EmbeddingClusterParams {
             agents: AgentConfig::uniform(agent_count, 2),
-            inference_parameters: InferenceParameters {
-                enable_embeddings: true,
-                ..InferenceParameters::deterministic()
-            },
+
             ..EmbeddingClusterParams::default()
         },
     )
@@ -49,7 +46,7 @@ async fn balancer_fans_out_embedding_batch_to_all_agents() {
 
     assert_eq!(collected.embeddings.len(), 16);
     assert!(collected.saw_done);
-    assert!(collected.errors.is_empty());
+    assert_eq!(collected.failures, Vec::<EmbeddingResult>::new());
 
     let producers: BTreeSet<&str> = collected
         .embeddings
diff --git a/paddler_cli_tests/tests/balancer_inference_cors_flag_allows_its_origin.rs b/paddler_cli_tests/tests/balancer_inference_cors_flag_allows_its_origin.rs
new file mode 100644
index 000000000..152ea788b
--- /dev/null
+++ b/paddler_cli_tests/tests/balancer_inference_cors_flag_allows_its_origin.rs
@@ -0,0 +1,43 @@
+use paddler_cli_tests::start_subprocess_cluster::start_subprocess_cluster;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_tests::send_cors_preflight::send_cors_preflight;
+
+const ALLOWED_ORIGIN: &str = "http://example.com";
+
+#[tokio::test(flavor = "multi_thread")]
+async fn balancer_inference_cors_flag_allows_its_origin() {
+    let cluster = start_subprocess_cluster(
+        env!("CARGO_BIN_EXE_paddler_cluster_node"),
+        ClusterParams {
+            inference_cors_allowed_hosts: vec![ALLOWED_ORIGIN.to_owned()],
+            ..ClusterParams::without_agents()
+        },
+    )
+    .await
+    .expect("a balancer subprocess without agents must start");
+    let health_url = cluster
+        .balancer
+        .inference_base_url
+        .join("health")
+        .expect("the health path must join onto the base URL");
+
+    let response = send_cors_preflight(health_url, ALLOWED_ORIGIN)
+        .await
+        .expect("the preflight request must succeed");
+
+    assert_eq!(response.status(), 200);
+    assert_eq!(
+        response
+            .headers()
+            .get("access-control-allow-origin")
+            .expect("the preflight response must allow an origin")
+            .to_str()
+            .expect("the allowed origin must be ASCII"),
+        ALLOWED_ORIGIN
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_cli_tests/tests/balancer_keeps_an_agent_connected_after_it_answers_a_request_that_timed_out.rs b/paddler_cli_tests/tests/balancer_keeps_an_agent_connected_after_it_answers_a_request_that_timed_out.rs
new file mode 100644
index 000000000..8d2852ecc
--- /dev/null
+++ b/paddler_cli_tests/tests/balancer_keeps_an_agent_connected_after_it_answers_a_request_that_timed_out.rs
@@ -0,0 +1,72 @@
+use reqwest::StatusCode;
+use tokio_util::sync::CancellationToken;
+
+use paddler_cli_tests::pausable_agent_cluster::PausableAgentCluster;
+use paddler_cli_tests::pausable_agent_cluster_params::PausableAgentClusterParams;
+use paddler_cli_tests::start_pausable_agent_cluster::start_pausable_agent_cluster;
+use paddler_client::error::Error as ClientError;
+use paddler_messaging::api_path::ApiPath;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn balancer_keeps_an_agent_connected_after_it_answers_a_request_that_timed_out() {
+    let PausableAgentCluster {
+        cluster,
+        pausable_agent,
+    } = start_pausable_agent_cluster(PausableAgentClusterParams {
+        binary_path: env!("CARGO_BIN_EXE_paddler_cluster_node").to_owned(),
+        desired_state: ClusterDesiredState::KeepStored,
+        expected_slots_total: 0,
+        slot_count: 1,
+    })
+    .await
+    .expect("a cluster with a model-less agent must start");
+
+    pausable_agent
+        .signals
+        .pause()
+        .expect("the agent must stop before it is asked anything");
+
+    let model_metadata_url = cluster
+        .balancer
+        .management_base_url
+        .join(&ApiPath::agent_model_metadata(&pausable_agent.id))
+        .expect("the model metadata path must extend the management address");
+
+    assert_eq!(
+        cluster
+            .client_management
+            .get_model_metadata(CancellationToken::new(), &pausable_agent.id)
+            .await
+            .err()
+            .as_ref()
+            .map(ToString::to_string),
+        Some(
+            ClientError::UnexpectedResponseStatus {
+                message: String::new(),
+                status: StatusCode::GATEWAY_TIMEOUT,
+                url: model_metadata_url.to_string(),
+            }
+            .to_string()
+        )
+    );
+
+    pausable_agent
+        .signals
+        .resume()
+        .expect("the agent must continue and answer the request that already timed out");
+
+    assert_eq!(
+        cluster
+            .client_management
+            .get_model_metadata(CancellationToken::new(), &pausable_agent.id)
+            .await
+            .expect("the balancer must relay the answer of the agent that stayed connected"),
+        None
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_cli_tests/tests/balancer_management_cors_flag_allows_its_origin.rs b/paddler_cli_tests/tests/balancer_management_cors_flag_allows_its_origin.rs
new file mode 100644
index 000000000..90a1b7a51
--- /dev/null
+++ b/paddler_cli_tests/tests/balancer_management_cors_flag_allows_its_origin.rs
@@ -0,0 +1,43 @@
+use paddler_cli_tests::start_subprocess_cluster::start_subprocess_cluster;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_tests::send_cors_preflight::send_cors_preflight;
+
+const ALLOWED_ORIGIN: &str = "http://example.com";
+
+#[tokio::test(flavor = "multi_thread")]
+async fn balancer_management_cors_flag_allows_its_origin() {
+    let cluster = start_subprocess_cluster(
+        env!("CARGO_BIN_EXE_paddler_cluster_node"),
+        ClusterParams {
+            management_cors_allowed_hosts: vec![ALLOWED_ORIGIN.to_owned()],
+            ..ClusterParams::without_agents()
+        },
+    )
+    .await
+    .expect("a balancer subprocess without agents must start");
+    let health_url = cluster
+        .balancer
+        .management_base_url
+        .join("health")
+        .expect("the health path must join onto the base URL");
+
+    let response = send_cors_preflight(health_url, ALLOWED_ORIGIN)
+        .await
+        .expect("the preflight request must succeed");
+
+    assert_eq!(response.status(), 200);
+    assert_eq!(
+        response
+            .headers()
+            .get("access-control-allow-origin")
+            .expect("the preflight response must allow an origin")
+            .to_str()
+            .expect("the allowed origin must be ASCII"),
+        ALLOWED_ORIGIN
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_cli_tests/tests/balancer_registers_multiple_agents_over_time.rs b/paddler_cli_tests/tests/balancer_registers_multiple_agents_over_time.rs
index 4cd0e5ce3..05440dc51 100644
--- a/paddler_cli_tests/tests/balancer_registers_multiple_agents_over_time.rs
+++ b/paddler_cli_tests/tests/balancer_registers_multiple_agents_over_time.rs
@@ -6,11 +6,7 @@ use paddler_test_cluster_harness::cluster_params::ClusterParams;
 async fn balancer_registers_multiple_agents_over_time() {
     let mut cluster = start_subprocess_cluster(
         env!("CARGO_BIN_EXE_paddler_cluster_node"),
-        ClusterParams {
-            agents: Vec::new(),
-            wait_for_slots_ready: false,
-            ..ClusterParams::default()
-        },
+        ClusterParams::without_agents(),
     )
     .await
     .expect("the cluster must start");
diff --git a/paddler_cli_tests/tests/balancer_serves_every_buffered_request_with_two_agents.rs b/paddler_cli_tests/tests/balancer_serves_every_buffered_request_with_two_agents.rs
new file mode 100644
index 000000000..edf8c2849
--- /dev/null
+++ b/paddler_cli_tests/tests/balancer_serves_every_buffered_request_with_two_agents.rs
@@ -0,0 +1,69 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::num::NonZeroU32;
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_cli_tests::start_subprocess_cluster::start_subprocess_cluster;
+use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_test_cluster_harness::collect_generated_tokens::collect_generated_tokens;
+use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn balancer_serves_every_buffered_request_with_two_agents() {
+    let cluster = start_subprocess_cluster(
+        env!("CARGO_BIN_EXE_paddler_cluster_node"),
+        ClusterParams {
+            agents: vec![
+                AgentConfig {
+                    name: "distributed-agent-0".to_owned(),
+                    slot_count: 2,
+                },
+                AgentConfig {
+                    name: "distributed-agent-1".to_owned(),
+                    slot_count: 2,
+                },
+            ],
+            wait_for_slots_ready: true,
+            max_buffered_requests: 10,
+            desired_state: ClusterDesiredState::Apply(Box::new(qwen3_0_6b().into_desired_state())),
+            ..ClusterParams::default()
+        },
+    )
+    .await
+    .expect("the cluster must start");
+
+    let mut streams = Vec::new();
+
+    for _ in 0..5 {
+        let stream = cluster
+            .continue_from_raw_prompt_stream(
+                CancellationToken::new(),
+                &ContinueFromRawPromptParams {
+                    grammar: None,
+                    max_tokens: NonZeroU32::new(10).unwrap(),
+                    raw_prompt: "Hello".to_owned(),
+                },
+            )
+            .await
+            .expect("the inference request must be accepted");
+
+        streams.push(stream);
+    }
+
+    for stream in streams {
+        collect_generated_tokens(stream)
+            .await
+            .expect("every buffered request must be served")
+            .summary()
+            .expect("every served request must finish with a summary");
+    }
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_cli_tests/tests/balancer_serves_every_service_within_the_default_macos_open_file_limit.rs b/paddler_cli_tests/tests/balancer_serves_every_service_within_the_default_macos_open_file_limit.rs
index 7ef415cb3..bc9d1fd10 100644
--- a/paddler_cli_tests/tests/balancer_serves_every_service_within_the_default_macos_open_file_limit.rs
+++ b/paddler_cli_tests/tests/balancer_serves_every_service_within_the_default_macos_open_file_limit.rs
@@ -6,6 +6,7 @@ use reqwest::StatusCode;
 use reqwest::get;
 use tokio_util::sync::CancellationToken;
 
+use paddler_cli_tests::signalled_subprocess::SignalledSubprocess;
 use paddler_cli_tests::spawn_balancer_subprocess::spawn_balancer_subprocess;
 use paddler_client::client_health::ClientHealth;
 use paddler_client::reports_health::ReportsHealth as _;
@@ -31,6 +32,8 @@ async fn balancer_serves_every_service_within_the_default_macos_open_file_limit(
         [
             "--compat-openai-addr",
             &ephemeral_loopback_addr,
+            "--compat-typesafe-addr",
+            &ephemeral_loopback_addr,
             "--inference-addr",
             &ephemeral_loopback_addr,
             "--management-addr",
@@ -38,24 +41,25 @@ async fn balancer_serves_every_service_within_the_default_macos_open_file_limit(
             "--web-admin-panel-addr",
             &ephemeral_loopback_addr,
         ],
+        SignalledSubprocess::new,
     )
     .await
-    .expect("the balancer must start and announce its addresses");
+    .expect("the balancer must start and announce its addresses")
+    .running_balancer;
     let web_admin_panel_addr = running_balancer
         .addresses
         .web_admin_panel
         .expect("the balancer must serve the web admin panel");
 
     let health_base_urls = [
-        running_balancer
-            .inference_base_url()
-            .expect("the inference address must form a base URL"),
-        running_balancer
-            .management_base_url()
-            .expect("the management address must form a base URL"),
+        running_balancer.inference_base_url.clone(),
+        running_balancer.management_base_url.clone(),
         running_balancer
             .compat_openai_base_url()
             .expect("the OpenAI compatibility address must form a base URL"),
+        running_balancer
+            .compat_typesafe_base_url()
+            .expect("the TypeSafe compatibility address must form a base URL"),
     ];
     let mut health_checks = Vec::new();
 
@@ -78,6 +82,6 @@ async fn balancer_serves_every_service_within_the_default_macos_open_file_limit(
         .await
         .expect("the balancer subprocess must shut down");
 
-    assert_eq!(health_checks, ["OK", "OK", "OK"]);
+    assert_eq!(health_checks, ["OK", "OK", "OK", "OK"]);
     assert_eq!(dashboard_status, StatusCode::OK);
 }
diff --git a/paddler_cli_tests/tests/balancer_serves_on_the_addresses_it_announces.rs b/paddler_cli_tests/tests/balancer_serves_on_the_addresses_it_announces.rs
index e8680c875..08713e9e3 100644
--- a/paddler_cli_tests/tests/balancer_serves_on_the_addresses_it_announces.rs
+++ b/paddler_cli_tests/tests/balancer_serves_on_the_addresses_it_announces.rs
@@ -8,11 +8,7 @@ use paddler_test_cluster_harness::cluster_params::ClusterParams;
 async fn balancer_serves_on_the_addresses_it_announces() {
     let cluster = start_subprocess_cluster(
         env!("CARGO_BIN_EXE_paddler_cluster_node"),
-        ClusterParams {
-            agents: Vec::new(),
-            wait_for_slots_ready: false,
-            ..ClusterParams::default()
-        },
+        ClusterParams::without_agents(),
     )
     .await
     .expect("the balancer subprocess must start and announce its addresses");
@@ -23,13 +19,21 @@ async fn balancer_serves_on_the_addresses_it_announces() {
         .await
         .expect("the announced inference address must serve health checks");
     let compat_openai_health = cluster
-        .client_compat_openai_health
+        .compat_openai_health_client()
+        .expect("the balancer must serve OpenAI compatibility")
         .get_health(CancellationToken::new())
         .await
         .expect("the announced OpenAI compatibility address must serve health checks");
+    let compat_typesafe_health = cluster
+        .compat_typesafe_health_client()
+        .expect("the balancer must serve TypeSafe compatibility")
+        .get_health(CancellationToken::new())
+        .await
+        .expect("the announced TypeSafe compatibility address must serve health checks");
 
     assert_eq!(inference_health, "OK");
     assert_eq!(compat_openai_health, "OK");
+    assert_eq!(compat_typesafe_health, "OK");
 
     cluster
         .shutdown()
diff --git a/paddler_cli_tests/tests/balancer_spreads_an_embedding_burst_across_every_agent.rs b/paddler_cli_tests/tests/balancer_spreads_an_embedding_burst_across_every_agent.rs
new file mode 100644
index 000000000..97c12694d
--- /dev/null
+++ b/paddler_cli_tests/tests/balancer_spreads_an_embedding_burst_across_every_agent.rs
@@ -0,0 +1,80 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::collections::BTreeSet;
+
+use futures_util::future;
+use tokio_util::sync::CancellationToken;
+
+use paddler_cli_tests::start_subprocess_embedding_cluster::start_subprocess_embedding_cluster;
+use paddler_messaging::embedding_input_document::EmbeddingInputDocument;
+use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
+use paddler_messaging::embedding_result::EmbeddingResult;
+use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::embedding_cluster_params::EmbeddingClusterParams;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn balancer_spreads_an_embedding_burst_across_every_agent() {
+    const AGENT_COUNT: usize = 4;
+    const SLOTS_PER_AGENT: u16 = 2;
+    const CONCURRENT_REQUESTS: usize = 8;
+
+    let cluster = start_subprocess_embedding_cluster(
+        env!("CARGO_BIN_EXE_paddler_cluster_node"),
+        EmbeddingClusterParams {
+            agents: AgentConfig::uniform(AGENT_COUNT, SLOTS_PER_AGENT),
+
+            max_buffered_requests: 32,
+            ..EmbeddingClusterParams::default()
+        },
+    )
+    .await
+    .expect("the cluster must start");
+
+    let collection_futures = (0..CONCURRENT_REQUESTS).map(|request_index| {
+        let input_batch: Vec<EmbeddingInputDocument> = (0..4)
+            .map(|document_index| EmbeddingInputDocument {
+                content: format!(
+                    "Burst request {request_index}, document {document_index}: \
+                     provide an embedding for evaluation."
+                ),
+                id: format!("req-{request_index}-doc-{document_index}"),
+            })
+            .collect();
+
+        cluster.generate_embedding_batch(
+            CancellationToken::new(),
+            &GenerateEmbeddingBatchParams {
+                input_batch,
+                normalization_method: EmbeddingNormalizationMethod::None,
+            },
+        )
+    });
+
+    let collected_streams = future::try_join_all(collection_futures)
+        .await
+        .expect("every concurrent request must succeed");
+
+    let producers_across_streams: BTreeSet<&str> = collected_streams
+        .iter()
+        .flat_map(|collected| collected.embeddings.iter())
+        .filter_map(|produced| produced.generated_by.as_deref())
+        .collect();
+
+    assert_eq!(
+        producers_across_streams.len(),
+        AGENT_COUNT,
+        "burst of {CONCURRENT_REQUESTS} embedding batches across {AGENT_COUNT} agents must reach every agent, but saw producers: {producers_across_streams:?}",
+    );
+
+    for collected in &collected_streams {
+        assert!(collected.saw_done);
+        assert_eq!(collected.failures, Vec::<EmbeddingResult>::new());
+        assert_eq!(collected.embeddings.len(), 4);
+    }
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_cli_tests/tests/balancer_web_admin_panel_shows_its_command_line_configuration.rs b/paddler_cli_tests/tests/balancer_web_admin_panel_shows_its_command_line_configuration.rs
index 712681037..fbf9b5b87 100644
--- a/paddler_cli_tests/tests/balancer_web_admin_panel_shows_its_command_line_configuration.rs
+++ b/paddler_cli_tests/tests/balancer_web_admin_panel_shows_its_command_line_configuration.rs
@@ -3,6 +3,7 @@ use std::collections::BTreeMap;
 use reqwest::get;
 use tokio::net::UdpSocket;
 
+use paddler_cli_tests::signalled_subprocess::SignalledSubprocess;
 use paddler_cli_tests::spawn_balancer_subprocess::spawn_balancer_subprocess;
 use paddler_test_cluster_harness::dashboard_attributes::dashboard_attributes;
 use paddler_test_cluster_harness::ephemeral_loopback_addr::EPHEMERAL_LOOPBACK_ADDR;
@@ -24,6 +25,8 @@ async fn balancer_web_admin_panel_shows_its_command_line_configuration() {
             "1500",
             "--compat-openai-addr",
             &ephemeral_loopback_addr,
+            "--compat-typesafe-addr",
+            &ephemeral_loopback_addr,
             "--inference-addr",
             &ephemeral_loopback_addr,
             "--management-addr",
@@ -39,9 +42,11 @@ async fn balancer_web_admin_panel_shows_its_command_line_configuration() {
             "--web-admin-panel-addr",
             &ephemeral_loopback_addr,
         ],
+        SignalledSubprocess::new,
     )
     .await
-    .expect("the balancer must start and announce its addresses");
+    .expect("the balancer must start and announce its addresses")
+    .running_balancer;
     let web_admin_panel_addr = running_balancer
         .addresses
         .web_admin_panel
@@ -66,6 +71,10 @@ async fn balancer_web_admin_panel_shows_its_command_line_configuration() {
             [
                 ("data-buffered-request-timeout-millis", "1500"),
                 ("data-compat-openai-addr", ephemeral_loopback_addr.as_str()),
+                (
+                    "data-compat-typesafe-addr",
+                    ephemeral_loopback_addr.as_str()
+                ),
                 ("data-inference-addr", ephemeral_loopback_addr.as_str()),
                 ("data-management-addr", ephemeral_loopback_addr.as_str()),
                 ("data-max-buffered-requests", "7"),
diff --git a/paddler_cli_tests/tests/balancer_without_compatibility_flags_announces_no_compatibility_addresses.rs b/paddler_cli_tests/tests/balancer_without_compatibility_flags_announces_no_compatibility_addresses.rs
new file mode 100644
index 000000000..41bf2d586
--- /dev/null
+++ b/paddler_cli_tests/tests/balancer_without_compatibility_flags_announces_no_compatibility_addresses.rs
@@ -0,0 +1,33 @@
+use std::mem::discriminant;
+
+use paddler_balancer::compatibility::openai_service::openai_api_path::OpenAIApiPath;
+use paddler_balancer::compatibility::typesafe_service::typesafe_api_path::TypeSafeApiPath;
+use paddler_test_cluster_harness::cluster_harness_error::ClusterHarnessError;
+
+use crate::running_balancer_subprocess::running_balancer_subprocess;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn balancer_without_compatibility_flags_announces_no_compatibility_addresses() {
+    let running_balancer = running_balancer_subprocess().await;
+
+    let compat_openai_url = running_balancer.compat_openai_url(OpenAIApiPath::CHAT_COMPLETIONS);
+    let compat_typesafe_url = running_balancer.compat_typesafe_url(TypeSafeApiPath::SYSTEM_ONE);
+
+    running_balancer
+        .shutdown()
+        .await
+        .expect("the balancer subprocess must shut down");
+
+    assert_eq!(
+        compat_openai_url.err().as_ref().map(discriminant),
+        Some(discriminant(
+            &ClusterHarnessError::CompatOpenAIServiceNotServed
+        ))
+    );
+    assert_eq!(
+        compat_typesafe_url.err().as_ref().map(discriminant),
+        Some(discriminant(
+            &ClusterHarnessError::CompatTypeSafeServiceNotServed
+        ))
+    );
+}
diff --git a/paddler_cli_tests/tests/cluster_with_a_missing_agent_binary.rs b/paddler_cli_tests/tests/cluster_with_a_missing_agent_binary.rs
new file mode 100644
index 000000000..25adc7ec3
--- /dev/null
+++ b/paddler_cli_tests/tests/cluster_with_a_missing_agent_binary.rs
@@ -0,0 +1,20 @@
+use tokio_util::sync::CancellationToken;
+
+use paddler_cli_tests::missing_binary_path::MISSING_BINARY_PATH;
+use paddler_cli_tests::subprocess_agents_cluster_connection::SubprocessAgentsClusterConnection;
+use paddler_test_cluster_harness::cluster::Cluster;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
+
+use crate::running_balancer_subprocess::running_balancer_subprocess;
+
+pub async fn cluster_with_a_missing_agent_binary() -> Cluster {
+    SubprocessAgentsClusterConnection {
+        agent_binary_path: MISSING_BINARY_PATH,
+        cancellation_token: CancellationToken::new(),
+        desired_state: &ClusterDesiredState::Apply(Box::default()),
+        running_balancer: running_balancer_subprocess().await,
+    }
+    .connect()
+    .await
+    .expect("the cluster must connect to the balancer")
+}
diff --git a/paddler_cli_tests/tests/ephemeral_balancer_arguments.rs b/paddler_cli_tests/tests/ephemeral_balancer_arguments.rs
new file mode 100644
index 000000000..5c83f69f6
--- /dev/null
+++ b/paddler_cli_tests/tests/ephemeral_balancer_arguments.rs
@@ -0,0 +1,12 @@
+use paddler_test_cluster_harness::ephemeral_loopback_addr::EPHEMERAL_LOOPBACK_ADDR;
+
+pub fn ephemeral_balancer_arguments() -> [String; 4] {
+    let ephemeral_loopback_addr = EPHEMERAL_LOOPBACK_ADDR.to_string();
+
+    [
+        "--inference-addr".to_owned(),
+        ephemeral_loopback_addr.clone(),
+        "--management-addr".to_owned(),
+        ephemeral_loopback_addr,
+    ]
+}
diff --git a/paddler_cli_tests/tests/http_client_reports_a_balancer_killed_before_responding_as_a_failed_request.rs b/paddler_cli_tests/tests/http_client_reports_a_balancer_killed_before_responding_as_a_failed_request.rs
new file mode 100644
index 000000000..3afabd92c
--- /dev/null
+++ b/paddler_cli_tests/tests/http_client_reports_a_balancer_killed_before_responding_as_a_failed_request.rs
@@ -0,0 +1,70 @@
+use serde_json::json;
+use tokio::spawn;
+use tokio_util::sync::CancellationToken;
+
+use paddler_balancer::compatibility::openai_service::openai_api_path::OpenAIApiPath;
+use paddler_cli_tests::start_signalled_subprocess_cluster::start_signalled_subprocess_cluster;
+use paddler_cli_tests::subprocess_cluster::SubprocessCluster;
+use paddler_client::http_client::HttpClient;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn http_client_reports_a_balancer_killed_before_responding_as_a_failed_request() {
+    let SubprocessCluster {
+        balancer_signals,
+        mut cluster,
+    } = start_signalled_subprocess_cluster(
+        env!("CARGO_BIN_EXE_paddler_cluster_node"),
+        ClusterParams::without_agents(),
+    )
+    .await
+    .expect("a balancer without agents must start");
+    let http_client = HttpClient::new(
+        cluster
+            .balancer
+            .compat_openai_base_url()
+            .expect("the balancer must serve the OpenAI compatibility service"),
+    );
+    let chat_completion = spawn(async move {
+        http_client
+            .post_json(
+                CancellationToken::new(),
+                OpenAIApiPath::CHAT_COMPLETIONS,
+                &json!({
+                    "model": "test-model",
+                    "messages": [{"role": "user", "content": "hi"}]
+                }),
+            )
+            .await
+    });
+
+    cluster
+        .wait_for_buffered_request_count(1)
+        .await
+        .expect("the chat completion must wait in the buffer while no agent is available");
+    balancer_signals
+        .kill()
+        .expect("the balancer must be killed before it responds");
+
+    let chat_completions_url = cluster
+        .balancer
+        .compat_openai_url(OpenAIApiPath::CHAT_COMPLETIONS)
+        .expect("the balancer must serve the OpenAI compatibility service");
+
+    assert_eq!(
+        chat_completion
+            .await
+            .expect("the chat completion task must not panic")
+            .err()
+            .as_ref()
+            .map(ToString::to_string),
+        Some(format!(
+            "HTTP request failed: error sending request for url ({chat_completions_url})"
+        ))
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_cli_tests/tests/http_listener_reports_a_bound_address_it_cannot_read.rs b/paddler_cli_tests/tests/http_listener_reports_a_bound_address_it_cannot_read.rs
new file mode 100644
index 000000000..4c78ef250
--- /dev/null
+++ b/paddler_cli_tests/tests/http_listener_reports_a_bound_address_it_cannot_read.rs
@@ -0,0 +1,31 @@
+use std::io;
+use std::net::Ipv4Addr;
+use std::net::SocketAddr;
+
+use libc::EPERM;
+use libc::SYS_getsockname;
+
+use paddler_balancer::http_listener::HttpListener;
+
+use crate::refused_syscall::RefusedSyscall;
+use crate::refused_syscalls::RefusedSyscalls;
+
+#[test]
+fn http_listener_reports_a_bound_address_it_cannot_read() {
+    RefusedSyscalls {
+        errno: EPERM,
+        refused_syscalls: vec![RefusedSyscall {
+            syscall: SYS_getsockname,
+            rules: Vec::new(),
+        }],
+    }
+    .apply_to_every_thread();
+
+    assert_eq!(
+        HttpListener::bind(SocketAddr::from((Ipv4Addr::LOCALHOST, 0)))
+            .err()
+            .as_ref()
+            .and_then(io::Error::raw_os_error),
+        Some(EPERM)
+    );
+}
diff --git a/paddler_cli_tests/tests/http_listener_reports_a_reuse_address_option_it_cannot_set.rs b/paddler_cli_tests/tests/http_listener_reports_a_reuse_address_option_it_cannot_set.rs
new file mode 100644
index 000000000..f024e814a
--- /dev/null
+++ b/paddler_cli_tests/tests/http_listener_reports_a_reuse_address_option_it_cannot_set.rs
@@ -0,0 +1,31 @@
+use std::io;
+use std::net::Ipv4Addr;
+use std::net::SocketAddr;
+
+use libc::EPERM;
+use libc::SYS_setsockopt;
+
+use paddler_balancer::http_listener::HttpListener;
+
+use crate::refused_syscall::RefusedSyscall;
+use crate::refused_syscalls::RefusedSyscalls;
+
+#[test]
+fn http_listener_reports_a_reuse_address_option_it_cannot_set() {
+    RefusedSyscalls {
+        errno: EPERM,
+        refused_syscalls: vec![RefusedSyscall {
+            syscall: SYS_setsockopt,
+            rules: Vec::new(),
+        }],
+    }
+    .apply_to_every_thread();
+
+    assert_eq!(
+        HttpListener::bind(SocketAddr::from((Ipv4Addr::LOCALHOST, 0)))
+            .err()
+            .as_ref()
+            .and_then(io::Error::raw_os_error),
+        Some(EPERM)
+    );
+}
diff --git a/paddler_cli_tests/tests/http_listener_reports_a_socket_it_cannot_listen_on.rs b/paddler_cli_tests/tests/http_listener_reports_a_socket_it_cannot_listen_on.rs
new file mode 100644
index 000000000..94e2f1e76
--- /dev/null
+++ b/paddler_cli_tests/tests/http_listener_reports_a_socket_it_cannot_listen_on.rs
@@ -0,0 +1,31 @@
+use std::io;
+use std::net::Ipv4Addr;
+use std::net::SocketAddr;
+
+use libc::EPERM;
+use libc::SYS_listen;
+
+use paddler_balancer::http_listener::HttpListener;
+
+use crate::refused_syscall::RefusedSyscall;
+use crate::refused_syscalls::RefusedSyscalls;
+
+#[test]
+fn http_listener_reports_a_socket_it_cannot_listen_on() {
+    RefusedSyscalls {
+        errno: EPERM,
+        refused_syscalls: vec![RefusedSyscall {
+            syscall: SYS_listen,
+            rules: Vec::new(),
+        }],
+    }
+    .apply_to_every_thread();
+
+    assert_eq!(
+        HttpListener::bind(SocketAddr::from((Ipv4Addr::LOCALHOST, 0)))
+            .err()
+            .as_ref()
+            .and_then(io::Error::raw_os_error),
+        Some(EPERM)
+    );
+}
diff --git a/paddler_cli_tests/tests/http_listener_reports_a_socket_it_cannot_open.rs b/paddler_cli_tests/tests/http_listener_reports_a_socket_it_cannot_open.rs
new file mode 100644
index 000000000..8c4bd9274
--- /dev/null
+++ b/paddler_cli_tests/tests/http_listener_reports_a_socket_it_cannot_open.rs
@@ -0,0 +1,43 @@
+use std::fs::File;
+use std::io;
+use std::net::Ipv4Addr;
+use std::net::SocketAddr;
+use std::os::fd::AsRawFd as _;
+
+use nix::errno::Errno;
+use nix::sys::resource::Resource;
+use nix::sys::resource::getrlimit;
+use nix::sys::resource::setrlimit;
+
+use paddler_balancer::http_listener::HttpListener;
+
+#[test]
+fn http_listener_reports_a_socket_it_cannot_open() {
+    let lowest_free_descriptor = File::open("/dev/null")
+        .expect("a probe descriptor must open")
+        .as_raw_fd();
+    let (soft_descriptor_limit, hard_descriptor_limit) =
+        getrlimit(Resource::RLIMIT_NOFILE).expect("the descriptor limit must be readable");
+
+    setrlimit(
+        Resource::RLIMIT_NOFILE,
+        u64::try_from(lowest_free_descriptor).expect("descriptors are never negative"),
+        hard_descriptor_limit,
+    )
+    .expect("the descriptor limit must be lowered");
+
+    let bind_errno = HttpListener::bind(SocketAddr::from((Ipv4Addr::LOCALHOST, 0)))
+        .err()
+        .as_ref()
+        .and_then(io::Error::raw_os_error)
+        .map(Errno::from_raw);
+
+    setrlimit(
+        Resource::RLIMIT_NOFILE,
+        soft_descriptor_limit,
+        hard_descriptor_limit,
+    )
+    .expect("the descriptor limit must be restored");
+
+    assert_eq!(bind_errno, Some(Errno::EMFILE));
+}
diff --git a/paddler_cli_tests/tests/inference_socket_connection_drops_pending_requests_when_the_balancer_is_killed.rs b/paddler_cli_tests/tests/inference_socket_connection_drops_pending_requests_when_the_balancer_is_killed.rs
new file mode 100644
index 000000000..d9b2eaf26
--- /dev/null
+++ b/paddler_cli_tests/tests/inference_socket_connection_drops_pending_requests_when_the_balancer_is_killed.rs
@@ -0,0 +1,70 @@
+use std::num::NonZeroU32;
+
+use paddler_cli_tests::start_signalled_subprocess_cluster::start_signalled_subprocess_cluster;
+use paddler_cli_tests::subprocess_cluster::SubprocessCluster;
+use paddler_client::error::Error;
+use paddler_client::inference_socket::connection::Connection;
+use paddler_messaging::inference_server::message::Message as InferenceServerMessage;
+use paddler_messaging::inference_server::request::Request as InferenceServerRequest;
+use paddler_messaging::jsonrpc::request_envelope::RequestEnvelope;
+use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters_schema::validated_parameters_schema::ValidatedParametersSchema;
+use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+
+const PENDING_REQUEST_ID: &str = "pending-request";
+
+#[tokio::test(flavor = "multi_thread")]
+async fn inference_socket_connection_drops_pending_requests_when_the_balancer_is_killed() {
+    let SubprocessCluster {
+        balancer_signals,
+        mut cluster,
+    } = start_signalled_subprocess_cluster(
+        env!("CARGO_BIN_EXE_paddler_cluster_node"),
+        ClusterParams::without_agents(),
+    )
+    .await
+    .expect("a balancer without agents must start");
+    let connection = Connection::connect(cluster.balancer.inference_base_url.clone())
+        .await
+        .expect("the inference socket must connect");
+    let request: InferenceServerMessage<ValidatedParametersSchema> =
+        InferenceServerMessage::Request(RequestEnvelope {
+            id: PENDING_REQUEST_ID.to_owned(),
+            request: InferenceServerRequest::ContinueFromRawPrompt(ContinueFromRawPromptParams {
+                grammar: None,
+                max_tokens: NonZeroU32::MIN,
+                raw_prompt: "wait for an agent".to_owned(),
+            }),
+        });
+    let mut response_rx = connection
+        .send(PENDING_REQUEST_ID.to_owned(), &request)
+        .expect("an open connection must accept the request");
+
+    cluster
+        .wait_for_buffered_request_count(1)
+        .await
+        .expect("the request must wait in the buffer while no agent is available");
+    balancer_signals
+        .kill()
+        .expect("the balancer must be killed while the request is pending");
+
+    assert_eq!(
+        response_rx
+            .recv()
+            .await
+            .and_then(Result::err)
+            .as_ref()
+            .map(ToString::to_string),
+        Some(
+            Error::ConnectionDropped {
+                request_id: PENDING_REQUEST_ID.to_owned(),
+            }
+            .to_string()
+        )
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_cli_tests/tests/main.rs b/paddler_cli_tests/tests/main.rs
index 45d06d68f..a1b77c168 100644
--- a/paddler_cli_tests/tests/main.rs
+++ b/paddler_cli_tests/tests/main.rs
@@ -1,14 +1,93 @@
-mod balancer_distributes_buffered_requests_across_two_agents;
+mod a_balancer_drops_an_agent_that_stops_answering_pings;
+#[cfg(target_os = "linux")]
+mod a_balancer_stops_with_its_statsd_service_failed_when_it_cannot_open_a_udp_socket;
+#[cfg(target_os = "linux")]
+mod a_sampler_chain_that_cannot_allocate_its_penalty_window_fails_the_generation_preparation;
+#[cfg(target_os = "linux")]
+mod agent_exits_with_failure_when_it_cannot_set_up_its_model_downloader;
+mod agent_reconnects_to_its_balancer_after_the_balancer_is_killed;
+mod agent_reconnects_to_its_balancer_after_the_balancer_restarts;
+mod agent_shuts_down_cleanly_while_its_balancer_is_unresponsive;
+#[cfg(test)]
+mod agent_socket_of_a_vanished_balancer;
+mod an_agent_reconnects_when_the_balancer_stops_answering_pings;
+#[cfg(target_os = "linux")]
+mod an_http_service_reports_a_listener_it_cannot_make_nonblocking;
+mod balancer_answers_gateway_timeout_when_a_paused_agent_never_answers;
+mod balancer_dispatches_requests_only_to_agents_that_serve;
 mod balancer_distributes_embedding_batch_across_agents;
 mod balancer_distributes_embedding_batch_across_agents_with_uneven_slots;
-mod balancer_distributes_embedding_burst_evenly_across_agents;
 mod balancer_distributes_token_burst_evenly_across_agents;
 mod balancer_emits_overflow_errors_when_embedding_burst_exceeds_max_buffered_requests;
+#[cfg(target_os = "linux")]
+mod balancer_exits_with_failure_when_it_cannot_announce_its_addresses;
 mod balancer_fans_out_embedding_batch_to_all_agents;
+mod balancer_inference_cors_flag_allows_its_origin;
+mod balancer_keeps_an_agent_connected_after_it_answers_a_request_that_timed_out;
+mod balancer_management_cors_flag_allows_its_origin;
 mod balancer_registers_multiple_agents_over_time;
-#[cfg(all(unix, feature = "web_admin_panel"))]
+mod balancer_serves_every_buffered_request_with_two_agents;
+#[cfg(feature = "web_admin_panel")]
 mod balancer_serves_every_service_within_the_default_macos_open_file_limit;
 mod balancer_serves_on_the_addresses_it_announces;
+mod balancer_spreads_an_embedding_burst_across_every_agent;
 #[cfg(feature = "web_admin_panel")]
 mod balancer_web_admin_panel_shows_its_command_line_configuration;
+mod balancer_without_compatibility_flags_announces_no_compatibility_addresses;
+#[cfg(test)]
+mod cluster_with_a_missing_agent_binary;
+#[cfg(test)]
+mod ephemeral_balancer_arguments;
+mod http_client_reports_a_balancer_killed_before_responding_as_a_failed_request;
+#[cfg(target_os = "linux")]
+mod http_listener_reports_a_bound_address_it_cannot_read;
+#[cfg(target_os = "linux")]
+mod http_listener_reports_a_reuse_address_option_it_cannot_set;
+#[cfg(target_os = "linux")]
+mod http_listener_reports_a_socket_it_cannot_listen_on;
+mod http_listener_reports_a_socket_it_cannot_open;
+mod inference_socket_connection_drops_pending_requests_when_the_balancer_is_killed;
+mod management_metrics_count_only_the_slots_of_serving_agents;
 mod management_metrics_endpoint_sums_the_slots_of_registered_agents;
+mod openai_chat_completion_reports_an_agent_killed_mid_generation_as_a_bad_gateway;
+mod paddler_agent_rejects_more_slots_than_llama_cpp_can_serve;
+mod paddler_without_a_subcommand_prints_help_and_exits_with_a_usage_error;
+mod pausable_agent_cluster_reports_a_binary_that_cannot_be_spawned;
+mod pausable_agent_cluster_reports_an_agent_that_exits_before_it_is_ready;
+mod pausable_agent_reports_a_binary_that_cannot_be_spawned;
+mod raw_inference_socket_waiting_for_an_answer_reports_a_balancer_that_vanished;
+#[cfg(test)]
+mod refused_syscall;
+#[cfg(test)]
+mod refused_syscalls;
+#[cfg(test)]
+mod running_balancer_subprocess;
+mod signalling_a_subprocess_refuses_one_already_reaped;
+#[cfg(test)]
+mod slots_beyond_what_llama_cpp_can_serve;
+mod spawned_balancer_reports_a_subprocess_it_cannot_signal;
+mod spawned_balancer_reports_an_announcement_that_is_not_its_addresses;
+mod spawned_balancer_reports_an_exit_it_cannot_observe;
+mod stallable_tcp_relay_relays_a_connection_until_both_sides_close_it;
+mod stallable_tcp_relay_reports_a_connection_it_cannot_accept;
+mod stallable_tcp_relay_reports_an_upstream_it_cannot_reach;
+mod subprocess_cluster_reports_a_balancer_that_exits_before_announcing;
+mod subprocess_cluster_reports_a_connection_cancelled_before_the_balancer_is_healthy;
+mod subprocess_cluster_reports_a_desired_state_the_balancer_cannot_store;
+mod subprocess_cluster_reports_an_additional_agent_binary_that_cannot_be_spawned;
+mod subprocess_cluster_reports_an_agent_binary_that_cannot_be_spawned;
+mod subprocess_cluster_reports_an_agent_that_exits_before_it_is_ready;
+mod subprocess_cluster_shutdown_reports_an_agent_that_exited_with_a_failure;
+#[cfg(test)]
+mod subprocess_cluster_with_a_serving_and_a_stuck_agent;
+mod subprocess_exit_reaped_elsewhere_is_reported_as_unobservable;
+mod subprocess_shutdown_reports_a_failed_exit;
+#[cfg(target_os = "linux")]
+mod subprocess_shutdown_reports_a_termination_it_cannot_signal;
+mod subprocess_signals_report_a_killed_process_that_cannot_be_paused;
+mod subprocess_signals_report_a_reaped_process_that_cannot_be_signalled;
+#[cfg(target_os = "linux")]
+mod the_cli_reports_a_shutdown_signal_it_cannot_listen_for;
+mod typesafe_system_one_reports_an_agent_killed_mid_decision_as_a_bad_gateway;
+mod untrusted_agent_socket_waiting_for_a_close_frame_reports_a_balancer_that_vanished;
+mod untrusted_agent_socket_waiting_for_text_reports_a_balancer_that_vanished;
diff --git a/paddler_cli_tests/tests/management_metrics_count_only_the_slots_of_serving_agents.rs b/paddler_cli_tests/tests/management_metrics_count_only_the_slots_of_serving_agents.rs
new file mode 100644
index 000000000..b6ab42723
--- /dev/null
+++ b/paddler_cli_tests/tests/management_metrics_count_only_the_slots_of_serving_agents.rs
@@ -0,0 +1,28 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use tokio_util::sync::CancellationToken;
+
+use crate::subprocess_cluster_with_a_serving_and_a_stuck_agent::SERVING_AGENT_SLOTS;
+use crate::subprocess_cluster_with_a_serving_and_a_stuck_agent::subprocess_cluster_with_a_serving_and_a_stuck_agent;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn management_metrics_count_only_the_slots_of_serving_agents() {
+    let cluster = subprocess_cluster_with_a_serving_and_a_stuck_agent().await;
+
+    let metrics = cluster
+        .client_management
+        .get_metrics(CancellationToken::new())
+        .await
+        .expect("the metrics endpoint must answer");
+
+    assert!(
+        metrics
+            .lines()
+            .any(|line| line == format!("paddler_slots_total {SERVING_AGENT_SLOTS}"))
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_cli_tests/tests/openai_chat_completion_reports_an_agent_killed_mid_generation_as_a_bad_gateway.rs b/paddler_cli_tests/tests/openai_chat_completion_reports_an_agent_killed_mid_generation_as_a_bad_gateway.rs
new file mode 100644
index 000000000..26c9c5c0d
--- /dev/null
+++ b/paddler_cli_tests/tests/openai_chat_completion_reports_an_agent_killed_mid_generation_as_a_bad_gateway.rs
@@ -0,0 +1,74 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use reqwest::StatusCode;
+use serde_json::Value;
+use tokio::spawn;
+
+use paddler_balancer::compatibility::openai_service::openai_api_path::OpenAIApiPath;
+use paddler_cli_tests::pausable_agent_cluster::PausableAgentCluster;
+use paddler_cli_tests::pausable_agent_cluster_params::PausableAgentClusterParams;
+use paddler_cli_tests::start_pausable_agent_cluster::start_pausable_agent_cluster;
+use paddler_openai_response_format_validator::openai_validator::OpenAIValidator;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
+use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
+use paddler_tests::minimal_openai_chat_completion::minimal_openai_chat_completion;
+
+const TEXT_GENERATION_SLOTS: u16 = 1;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn openai_chat_completion_reports_an_agent_killed_mid_generation_as_a_bad_gateway() {
+    let PausableAgentCluster {
+        mut cluster,
+        pausable_agent,
+    } = start_pausable_agent_cluster(PausableAgentClusterParams {
+        binary_path: env!("CARGO_BIN_EXE_paddler_cluster_node").to_owned(),
+        desired_state: ClusterDesiredState::Apply(Box::new(qwen3_0_6b().into_desired_state())),
+        expected_slots_total: TEXT_GENERATION_SLOTS,
+        slot_count: TEXT_GENERATION_SLOTS,
+    })
+    .await
+    .expect("a text generation cluster must start");
+
+    pausable_agent
+        .signals
+        .pause()
+        .expect("the agent must stop before the generation reaches it");
+
+    let chat_completion = spawn(minimal_openai_chat_completion(
+        cluster
+            .balancer
+            .compat_openai_url(OpenAIApiPath::CHAT_COMPLETIONS)
+            .expect("the balancer must serve the OpenAI compatibility service"),
+    ));
+
+    cluster
+        .wait_for_slots_processing(&pausable_agent.id, 1)
+        .await
+        .expect("the balancer must dispatch the generation to the paused agent");
+    pausable_agent
+        .signals
+        .kill()
+        .expect("the agent must be killed while it holds the generation");
+
+    let response = chat_completion
+        .await
+        .expect("the chat completion task must not panic")
+        .expect("the balancer must answer the chat completion");
+
+    assert_eq!(response.status(), StatusCode::BAD_GATEWAY);
+
+    OpenAIValidator::new()
+        .expect("the OpenAI specification must load")
+        .validate_error_response(
+            &response
+                .json::<Value>()
+                .await
+                .expect("the error body must be JSON"),
+        )
+        .expect("the error body must be an OpenAI error");
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_cli_tests/tests/paddler_agent_rejects_more_slots_than_llama_cpp_can_serve.rs b/paddler_cli_tests/tests/paddler_agent_rejects_more_slots_than_llama_cpp_can_serve.rs
new file mode 100644
index 000000000..188a3994a
--- /dev/null
+++ b/paddler_cli_tests/tests/paddler_agent_rejects_more_slots_than_llama_cpp_can_serve.rs
@@ -0,0 +1,20 @@
+use std::process::Command;
+
+use paddler_cli_tests::clap_usage_error_exit_code::CLAP_USAGE_ERROR_EXIT_CODE;
+
+use crate::slots_beyond_what_llama_cpp_can_serve::SLOTS_BEYOND_WHAT_LLAMA_CPP_CAN_SERVE;
+
+#[test]
+fn paddler_agent_rejects_more_slots_than_llama_cpp_can_serve() {
+    let agent_output = Command::new(env!("CARGO_BIN_EXE_paddler_cluster_node"))
+        .arg("agent")
+        .arg("--management-addr")
+        .arg("127.0.0.1:1")
+        .arg("--slots")
+        .arg(SLOTS_BEYOND_WHAT_LLAMA_CPP_CAN_SERVE.to_string())
+        .output()
+        .expect("paddler agent must run");
+
+    assert_eq!(agent_output.status.code(), Some(CLAP_USAGE_ERROR_EXIT_CODE));
+    assert_eq!(agent_output.stdout, Vec::<u8>::new());
+}
diff --git a/paddler_cli_tests/tests/paddler_without_a_subcommand_prints_help_and_exits_with_a_usage_error.rs b/paddler_cli_tests/tests/paddler_without_a_subcommand_prints_help_and_exits_with_a_usage_error.rs
new file mode 100644
index 000000000..7bb34d81d
--- /dev/null
+++ b/paddler_cli_tests/tests/paddler_without_a_subcommand_prints_help_and_exits_with_a_usage_error.rs
@@ -0,0 +1,18 @@
+use std::process::Command;
+
+use paddler_cli_tests::clap_usage_error_exit_code::CLAP_USAGE_ERROR_EXIT_CODE;
+
+#[test]
+fn paddler_without_a_subcommand_prints_help_and_exits_with_a_usage_error() {
+    let help_output = Command::new(env!("CARGO_BIN_EXE_paddler_cluster_node"))
+        .arg("--help")
+        .output()
+        .expect("paddler --help must run");
+    let bare_output = Command::new(env!("CARGO_BIN_EXE_paddler_cluster_node"))
+        .output()
+        .expect("paddler without arguments must run");
+
+    assert_eq!(bare_output.status.code(), Some(CLAP_USAGE_ERROR_EXIT_CODE));
+    assert_eq!(bare_output.stdout, Vec::<u8>::new());
+    assert_eq!(bare_output.stderr, help_output.stdout);
+}
diff --git a/paddler_cli_tests/tests/pausable_agent_cluster_reports_a_binary_that_cannot_be_spawned.rs b/paddler_cli_tests/tests/pausable_agent_cluster_reports_a_binary_that_cannot_be_spawned.rs
new file mode 100644
index 000000000..12202efe2
--- /dev/null
+++ b/paddler_cli_tests/tests/pausable_agent_cluster_reports_a_binary_that_cannot_be_spawned.rs
@@ -0,0 +1,28 @@
+use std::io;
+use std::io::ErrorKind;
+
+use paddler_cli_tests::missing_binary_path::MISSING_BINARY_PATH;
+use paddler_cli_tests::pausable_agent_cluster_params::PausableAgentClusterParams;
+use paddler_cli_tests::start_pausable_agent_cluster::start_pausable_agent_cluster;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn pausable_agent_cluster_reports_a_binary_that_cannot_be_spawned() {
+    let start_error = start_pausable_agent_cluster(PausableAgentClusterParams {
+        binary_path: MISSING_BINARY_PATH.to_owned(),
+        desired_state: ClusterDesiredState::KeepStored,
+        expected_slots_total: 1,
+        slot_count: 1,
+    })
+    .await
+    .err()
+    .expect("a missing binary must not start a cluster");
+
+    assert_eq!(
+        start_error
+            .root_cause()
+            .downcast_ref::<io::Error>()
+            .map(io::Error::kind),
+        Some(ErrorKind::NotFound)
+    );
+}
diff --git a/paddler_cli_tests/tests/pausable_agent_cluster_reports_an_agent_that_exits_before_it_is_ready.rs b/paddler_cli_tests/tests/pausable_agent_cluster_reports_an_agent_that_exits_before_it_is_ready.rs
new file mode 100644
index 000000000..673355e1c
--- /dev/null
+++ b/paddler_cli_tests/tests/pausable_agent_cluster_reports_an_agent_that_exits_before_it_is_ready.rs
@@ -0,0 +1,25 @@
+use paddler_cli_tests::agent_exited_before_ready_with_code::agent_exited_before_ready_with_code;
+use paddler_cli_tests::clap_usage_error_exit_code::CLAP_USAGE_ERROR_EXIT_CODE;
+use paddler_cli_tests::pausable_agent_cluster_params::PausableAgentClusterParams;
+use paddler_cli_tests::start_pausable_agent_cluster::start_pausable_agent_cluster;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
+
+const SLOTS_THE_AGENT_REJECTS: u16 = 0;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn pausable_agent_cluster_reports_an_agent_that_exits_before_it_is_ready() {
+    let start_error = start_pausable_agent_cluster(PausableAgentClusterParams {
+        binary_path: env!("CARGO_BIN_EXE_paddler_cluster_node").to_owned(),
+        desired_state: ClusterDesiredState::KeepStored,
+        expected_slots_total: SLOTS_THE_AGENT_REJECTS,
+        slot_count: SLOTS_THE_AGENT_REJECTS,
+    })
+    .await
+    .err()
+    .expect("an agent that exits must fail the cluster start");
+
+    assert!(agent_exited_before_ready_with_code(
+        &start_error,
+        CLAP_USAGE_ERROR_EXIT_CODE
+    ));
+}
diff --git a/paddler_cli_tests/tests/pausable_agent_reports_a_binary_that_cannot_be_spawned.rs b/paddler_cli_tests/tests/pausable_agent_reports_a_binary_that_cannot_be_spawned.rs
new file mode 100644
index 000000000..e9bf6ad5c
--- /dev/null
+++ b/paddler_cli_tests/tests/pausable_agent_reports_a_binary_that_cannot_be_spawned.rs
@@ -0,0 +1,46 @@
+use std::io;
+use std::io::ErrorKind;
+
+use paddler_cli_tests::missing_binary_path::MISSING_BINARY_PATH;
+use paddler_cli_tests::pausable_agent::PausableAgent;
+use paddler_cli_tests::pausable_agent_params::PausableAgentParams;
+use paddler_cli_tests::start_subprocess_cluster::start_subprocess_cluster;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn pausable_agent_reports_a_binary_that_cannot_be_spawned() {
+    let mut cluster = start_subprocess_cluster(
+        env!("CARGO_BIN_EXE_paddler_cluster_node"),
+        ClusterParams::without_agents(),
+    )
+    .await
+    .expect("a cluster without agents must start");
+
+    let management_addr = cluster.balancer.addresses.management;
+    let join_error = PausableAgent::join(
+        &mut cluster,
+        PausableAgentParams {
+            binary_path: MISSING_BINARY_PATH.to_owned(),
+            config: AgentConfig::single(1),
+            expected_slots_total: 1,
+            management_addr,
+        },
+    )
+    .await
+    .err()
+    .expect("a missing binary must not join the cluster");
+
+    assert_eq!(
+        join_error
+            .root_cause()
+            .downcast_ref::<io::Error>()
+            .map(io::Error::kind),
+        Some(ErrorKind::NotFound)
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_cli_tests/tests/raw_inference_socket_waiting_for_an_answer_reports_a_balancer_that_vanished.rs b/paddler_cli_tests/tests/raw_inference_socket_waiting_for_an_answer_reports_a_balancer_that_vanished.rs
new file mode 100644
index 000000000..566d65d8d
--- /dev/null
+++ b/paddler_cli_tests/tests/raw_inference_socket_waiting_for_an_answer_reports_a_balancer_that_vanished.rs
@@ -0,0 +1,56 @@
+use std::error::Error as _;
+use std::mem::discriminant;
+
+use tokio_tungstenite::tungstenite::Error as WebSocketError;
+use tokio_tungstenite::tungstenite::error::ProtocolError;
+
+use paddler_cli_tests::start_signalled_subprocess_cluster::start_signalled_subprocess_cluster;
+use paddler_cli_tests::subprocess_cluster::SubprocessCluster;
+use paddler_test_cluster_harness::cluster_harness_error::ClusterHarnessError;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_test_cluster_harness::raw_inference_socket::RawInferenceSocket;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn raw_inference_socket_waiting_for_an_answer_reports_a_balancer_that_vanished() {
+    let SubprocessCluster {
+        balancer_signals,
+        cluster,
+    } = start_signalled_subprocess_cluster(
+        env!("CARGO_BIN_EXE_paddler_cluster_node"),
+        ClusterParams::without_agents(),
+    )
+    .await
+    .expect("a cluster without agents must start");
+    let mut raw_inference_socket =
+        RawInferenceSocket::connect(cluster.balancer.addresses.inference)
+            .await
+            .expect("the raw socket must reach the inference socket");
+
+    balancer_signals
+        .kill()
+        .expect("the balancer must be killed");
+
+    let receive_error = raw_inference_socket
+        .next_answer()
+        .await
+        .expect_err("a vanished balancer must not answer");
+
+    assert_eq!(
+        discriminant(&receive_error),
+        discriminant(&ClusterHarnessError::InferenceSocketReceiveFailed(
+            WebSocketError::ConnectionClosed
+        ))
+    );
+    assert_eq!(
+        receive_error
+            .source()
+            .and_then(|websocket_error| websocket_error.source())
+            .and_then(|protocol_error| protocol_error.downcast_ref::<ProtocolError>()),
+        Some(&ProtocolError::ResetWithoutClosingHandshake)
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down after its balancer was killed");
+}
diff --git a/paddler_cli_tests/tests/refused_syscall.rs b/paddler_cli_tests/tests/refused_syscall.rs
new file mode 100644
index 000000000..a5a64ae60
--- /dev/null
+++ b/paddler_cli_tests/tests/refused_syscall.rs
@@ -0,0 +1,8 @@
+#![cfg(target_os = "linux")]
+
+use seccompiler::SeccompRule;
+
+pub struct RefusedSyscall {
+    pub syscall: i64,
+    pub rules: Vec<SeccompRule>,
+}
diff --git a/paddler_cli_tests/tests/refused_syscalls.rs b/paddler_cli_tests/tests/refused_syscalls.rs
new file mode 100644
index 000000000..c116a1111
--- /dev/null
+++ b/paddler_cli_tests/tests/refused_syscalls.rs
@@ -0,0 +1,40 @@
+#![cfg(target_os = "linux")]
+
+use std::env::consts::ARCH;
+
+use seccompiler::BpfProgram;
+use seccompiler::SeccompAction;
+use seccompiler::SeccompFilter;
+use seccompiler::apply_filter_all_threads;
+
+use crate::refused_syscall::RefusedSyscall;
+
+pub struct RefusedSyscalls {
+    pub errno: i32,
+    pub refused_syscalls: Vec<RefusedSyscall>,
+}
+
+impl RefusedSyscalls {
+    pub fn apply_to_every_thread(self) {
+        let Self {
+            errno,
+            refused_syscalls,
+        } = self;
+        let refusing_filter: BpfProgram = SeccompFilter::new(
+            refused_syscalls
+                .into_iter()
+                .map(|RefusedSyscall { syscall, rules }| (syscall, rules))
+                .collect(),
+            SeccompAction::Allow,
+            SeccompAction::Errno(errno.cast_unsigned()),
+            ARCH.try_into()
+                .expect("seccompiler must support the architecture the tests run on"),
+        )
+        .expect("the refusing filter must be valid")
+        .try_into()
+        .expect("the refusing filter must compile into BPF");
+
+        apply_filter_all_threads(&refusing_filter)
+            .expect("the refusing filter must apply to every thread of the test process");
+    }
+}
diff --git a/paddler_cli_tests/tests/running_balancer_subprocess.rs b/paddler_cli_tests/tests/running_balancer_subprocess.rs
new file mode 100644
index 000000000..7eec77249
--- /dev/null
+++ b/paddler_cli_tests/tests/running_balancer_subprocess.rs
@@ -0,0 +1,16 @@
+use paddler_cli_tests::signalled_subprocess::SignalledSubprocess;
+use paddler_cli_tests::spawn_balancer_subprocess::spawn_balancer_subprocess;
+use paddler_test_cluster_harness::running_balancer::RunningBalancer;
+
+use crate::ephemeral_balancer_arguments::ephemeral_balancer_arguments;
+
+pub async fn running_balancer_subprocess() -> RunningBalancer {
+    spawn_balancer_subprocess(
+        env!("CARGO_BIN_EXE_paddler_cluster_node"),
+        ephemeral_balancer_arguments(),
+        SignalledSubprocess::new,
+    )
+    .await
+    .expect("the balancer must start and announce its addresses")
+    .running_balancer
+}
diff --git a/paddler_cli_tests/tests/signalling_a_subprocess_refuses_one_already_reaped.rs b/paddler_cli_tests/tests/signalling_a_subprocess_refuses_one_already_reaped.rs
new file mode 100644
index 000000000..8f39260f7
--- /dev/null
+++ b/paddler_cli_tests/tests/signalling_a_subprocess_refuses_one_already_reaped.rs
@@ -0,0 +1,29 @@
+use paddler_cli_tests::signalled_subprocess::SignalledSubprocess;
+use paddler_cli_tests::spawn_agent_subprocess::spawn_agent_subprocess;
+use paddler_cli_tests::spawn_agent_subprocess_params::SpawnAgentSubprocessParams;
+use paddler_cli_tests::subprocess_cluster_error::SubprocessClusterError;
+use paddler_cli_tests::unreachable_management_addr::UNREACHABLE_MANAGEMENT_ADDR;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn signalling_a_subprocess_refuses_one_already_reaped() {
+    let mut agent = spawn_agent_subprocess(SpawnAgentSubprocessParams {
+        binary_path: env!("CARGO_BIN_EXE_paddler_cluster_node").to_owned(),
+        management_addr: UNREACHABLE_MANAGEMENT_ADDR,
+        name: "agent-that-is-reaped-first".to_owned(),
+        slots: 1,
+    })
+    .expect("the agent subprocess must start");
+
+    agent
+        .kill()
+        .await
+        .expect("the agent must be killed and reaped");
+
+    assert_eq!(
+        SignalledSubprocess::new(agent)
+            .err()
+            .as_ref()
+            .map(ToString::to_string),
+        Some(SubprocessClusterError::ProcessAlreadyReaped.to_string())
+    );
+}
diff --git a/paddler_cli_tests/tests/slots_beyond_what_llama_cpp_can_serve.rs b/paddler_cli_tests/tests/slots_beyond_what_llama_cpp_can_serve.rs
new file mode 100644
index 000000000..455e23d3f
--- /dev/null
+++ b/paddler_cli_tests/tests/slots_beyond_what_llama_cpp_can_serve.rs
@@ -0,0 +1,3 @@
+use paddler_bootstrap::llama_cpp_max_sequences::LLAMA_CPP_MAX_SEQUENCES;
+
+pub const SLOTS_BEYOND_WHAT_LLAMA_CPP_CAN_SERVE: u16 = LLAMA_CPP_MAX_SEQUENCES + 1;
diff --git a/paddler_cli_tests/tests/spawned_balancer_reports_a_subprocess_it_cannot_signal.rs b/paddler_cli_tests/tests/spawned_balancer_reports_a_subprocess_it_cannot_signal.rs
new file mode 100644
index 000000000..0aa33f28c
--- /dev/null
+++ b/paddler_cli_tests/tests/spawned_balancer_reports_a_subprocess_it_cannot_signal.rs
@@ -0,0 +1,22 @@
+use paddler_cli_tests::signal_as_already_reaped::signal_as_already_reaped;
+use paddler_cli_tests::spawn_balancer_subprocess::spawn_balancer_subprocess;
+use paddler_cli_tests::subprocess_cluster_error::SubprocessClusterError;
+
+use crate::ephemeral_balancer_arguments::ephemeral_balancer_arguments;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn spawned_balancer_reports_a_subprocess_it_cannot_signal() {
+    assert_eq!(
+        spawn_balancer_subprocess(
+            env!("CARGO_BIN_EXE_paddler_cluster_node"),
+            ephemeral_balancer_arguments(),
+            signal_as_already_reaped,
+        )
+        .await
+        .err()
+        .as_ref()
+        .and_then(|spawn_error| spawn_error.downcast_ref::<SubprocessClusterError>())
+        .map(ToString::to_string),
+        Some(SubprocessClusterError::ProcessAlreadyReaped.to_string())
+    );
+}
diff --git a/paddler_cli_tests/tests/spawned_balancer_reports_an_announcement_that_is_not_its_addresses.rs b/paddler_cli_tests/tests/spawned_balancer_reports_an_announcement_that_is_not_its_addresses.rs
new file mode 100644
index 000000000..0e8d8648f
--- /dev/null
+++ b/paddler_cli_tests/tests/spawned_balancer_reports_an_announcement_that_is_not_its_addresses.rs
@@ -0,0 +1,33 @@
+use std::mem::discriminant;
+
+use serde_json::Value;
+use serde_json::from_str;
+
+use paddler_cli_tests::signalled_subprocess::SignalledSubprocess;
+use paddler_cli_tests::spawn_balancer_subprocess::spawn_balancer_subprocess;
+use paddler_cli_tests::subprocess_cluster_error::SubprocessClusterError;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn spawned_balancer_reports_an_announcement_that_is_not_its_addresses() {
+    let spawn_error = spawn_balancer_subprocess(
+        env!("CARGO_BIN_EXE_paddler_cluster_node"),
+        ["--help"],
+        SignalledSubprocess::new,
+    )
+    .await
+    .err()
+    .expect("a balancer that prints its help must not be taken for an announcement");
+
+    let help_text_error =
+        from_str::<Value>("help text").expect_err("help text must not be valid JSON");
+
+    assert_eq!(
+        spawn_error
+            .downcast_ref::<SubprocessClusterError>()
+            .map(discriminant),
+        Some(discriminant(&SubprocessClusterError::AnnouncementInvalid {
+            announcement: String::new(),
+            source: help_text_error,
+        }))
+    );
+}
diff --git a/paddler_cli_tests/tests/spawned_balancer_reports_an_exit_it_cannot_observe.rs b/paddler_cli_tests/tests/spawned_balancer_reports_an_exit_it_cannot_observe.rs
new file mode 100644
index 000000000..267bc2c2c
--- /dev/null
+++ b/paddler_cli_tests/tests/spawned_balancer_reports_an_exit_it_cannot_observe.rs
@@ -0,0 +1,64 @@
+use std::error::Error as _;
+use std::fs::OpenOptions;
+use std::future::poll_fn;
+use std::io;
+use std::pin::pin;
+use std::task::Poll;
+
+use nix::errno::Errno;
+use nix::sys::stat::Mode;
+use nix::sys::wait::waitpid;
+use nix::unistd::mkfifo;
+use tempfile::TempDir;
+
+use paddler_cli_tests::signalled_subprocess::SignalledSubprocess;
+use paddler_cli_tests::spawn_balancer_subprocess::spawn_balancer_subprocess;
+use paddler_cli_tests::subprocess_cluster_error::SubprocessClusterError;
+
+#[tokio::test]
+async fn spawned_balancer_reports_an_exit_it_cannot_observe() {
+    let state_directory = TempDir::new().expect("the state directory must be created");
+    let state_fifo_path = state_directory.path().join("state.fifo");
+
+    mkfifo(&state_fifo_path, Mode::S_IRWXU).expect("the state fifo must be created");
+
+    let mut balancer_spawn = pin!(spawn_balancer_subprocess(
+        env!("CARGO_BIN_EXE_paddler_cluster_node"),
+        [format!(
+            "--state-database=file://{}",
+            state_fifo_path.display()
+        )],
+        SignalledSubprocess::new,
+    ));
+
+    assert!(
+        poll_fn(|context| Poll::Ready(balancer_spawn.as_mut().poll(context).is_pending())).await,
+        "the balancer must wait for its state before it can announce"
+    );
+
+    drop(
+        OpenOptions::new()
+            .write(true)
+            .open(&state_fifo_path)
+            .expect("the balancer must open its state fifo"),
+    );
+    waitpid(None, None).expect("the balancer must exit and be reaped before the harness waits");
+
+    let spawn_error = balancer_spawn
+        .await
+        .err()
+        .expect("a balancer reaped elsewhere must not be observed exiting");
+
+    assert_eq!(
+        spawn_error.to_string(),
+        SubprocessClusterError::ProcessExitUnobservable(io::Error::from(Errno::ECHILD)).to_string()
+    );
+    assert_eq!(
+        spawn_error
+            .downcast_ref::<SubprocessClusterError>()
+            .and_then(|subprocess_cluster_error| subprocess_cluster_error.source())
+            .and_then(|source| source.downcast_ref::<io::Error>())
+            .and_then(io::Error::raw_os_error),
+        Some(Errno::ECHILD as i32)
+    );
+}
diff --git a/paddler_cli_tests/tests/stallable_tcp_relay_relays_a_connection_until_both_sides_close_it.rs b/paddler_cli_tests/tests/stallable_tcp_relay_relays_a_connection_until_both_sides_close_it.rs
new file mode 100644
index 000000000..0acdc4fed
--- /dev/null
+++ b/paddler_cli_tests/tests/stallable_tcp_relay_relays_a_connection_until_both_sides_close_it.rs
@@ -0,0 +1,75 @@
+use std::io;
+use std::net::Ipv4Addr;
+use std::net::SocketAddr;
+
+use tokio::io::AsyncReadExt as _;
+use tokio::io::AsyncWriteExt as _;
+use tokio::net::TcpListener;
+use tokio::net::TcpStream;
+use tokio::spawn;
+
+use paddler_cli_tests::stallable_tcp_relay::StallableTcpRelay;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn stallable_tcp_relay_relays_a_connection_until_both_sides_close_it() {
+    let upstream_listener = TcpListener::bind(SocketAddr::from((Ipv4Addr::LOCALHOST, 0)))
+        .await
+        .expect("the upstream must bind a loopback port");
+    let upstream_addr = upstream_listener
+        .local_addr()
+        .expect("the upstream listener must know its address");
+    let upstream = spawn(async move {
+        let (mut relayed_connection, _relay_addr) = upstream_listener
+            .accept()
+            .await
+            .expect("the upstream must accept the relayed connection");
+        let mut request = Vec::new();
+
+        relayed_connection
+            .read_to_end(&mut request)
+            .await
+            .expect("the upstream must read the relayed request");
+        relayed_connection
+            .write_all(&request)
+            .await
+            .expect("the upstream must echo the request");
+        relayed_connection
+            .shutdown()
+            .await
+            .expect("the upstream must end its echo");
+    });
+    let relay_listener = TcpListener::bind(SocketAddr::from((Ipv4Addr::LOCALHOST, 0)))
+        .await
+        .expect("the relay must bind a loopback port");
+    let relay_addr = relay_listener
+        .local_addr()
+        .expect("the relay listener must know its address");
+    let relay = StallableTcpRelay::start(relay_listener, upstream_addr);
+    let mut downstream = TcpStream::connect(relay_addr)
+        .await
+        .expect("the relay must accept a connection");
+    let mut bytes_relayed_back = Vec::new();
+
+    downstream
+        .write_all(b"ping")
+        .await
+        .expect("the request must reach the relay");
+    downstream.shutdown().await.expect("the request must end");
+    downstream
+        .read_to_end(&mut bytes_relayed_back)
+        .await
+        .expect("the relay must relay the echo back");
+
+    assert_eq!(bytes_relayed_back, b"ping");
+    upstream.await.expect("the upstream task must not panic");
+    assert_eq!(
+        relay
+            .stop()
+            .await
+            .expect("the relay task must not panic")
+            .err()
+            .as_ref()
+            .map(io::Error::kind),
+        None
+    );
+}
diff --git a/paddler_cli_tests/tests/stallable_tcp_relay_reports_a_connection_it_cannot_accept.rs b/paddler_cli_tests/tests/stallable_tcp_relay_reports_a_connection_it_cannot_accept.rs
new file mode 100644
index 000000000..5c7337f11
--- /dev/null
+++ b/paddler_cli_tests/tests/stallable_tcp_relay_reports_a_connection_it_cannot_accept.rs
@@ -0,0 +1,73 @@
+use std::fs::File;
+use std::io;
+use std::net::Ipv4Addr;
+use std::net::SocketAddr;
+use std::net::SocketAddrV4;
+use std::os::fd::AsRawFd as _;
+
+use nix::errno::Errno;
+use nix::sys::resource::Resource;
+use nix::sys::resource::getrlimit;
+use nix::sys::resource::setrlimit;
+use nix::sys::socket::AddressFamily;
+use nix::sys::socket::SockFlag;
+use nix::sys::socket::SockType;
+use nix::sys::socket::SockaddrIn;
+use nix::sys::socket::connect;
+use nix::sys::socket::socket;
+use tokio::net::TcpListener;
+
+use paddler_cli_tests::stallable_tcp_relay::StallableTcpRelay;
+use paddler_cli_tests::unreachable_management_addr::UNREACHABLE_MANAGEMENT_ADDR;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn stallable_tcp_relay_reports_a_connection_it_cannot_accept() {
+    let relay_listener = TcpListener::bind(SocketAddr::from((Ipv4Addr::LOCALHOST, 0)))
+        .await
+        .expect("the relay must bind a loopback port");
+    let relay_addr_v4 = SocketAddrV4::new(
+        Ipv4Addr::LOCALHOST,
+        relay_listener
+            .local_addr()
+            .expect("the relay listener must know its address")
+            .port(),
+    );
+    let relay = StallableTcpRelay::start(relay_listener, UNREACHABLE_MANAGEMENT_ADDR);
+    let downstream_socket = socket(
+        AddressFamily::Inet,
+        SockType::Stream,
+        SockFlag::empty(),
+        None,
+    )
+    .expect("the downstream socket must open");
+    let lowest_free_descriptor = File::open("/dev/null")
+        .expect("a probe descriptor must open")
+        .as_raw_fd();
+    let (_soft_descriptor_limit, hard_descriptor_limit) =
+        getrlimit(Resource::RLIMIT_NOFILE).expect("the descriptor limit must be readable");
+
+    setrlimit(
+        Resource::RLIMIT_NOFILE,
+        u64::try_from(lowest_free_descriptor).expect("descriptors are never negative"),
+        hard_descriptor_limit,
+    )
+    .expect("the descriptor limit must be lowered");
+
+    connect(
+        downstream_socket.as_raw_fd(),
+        &SockaddrIn::from(relay_addr_v4),
+    )
+    .expect("the kernel must complete the handshake without a new descriptor");
+
+    assert_eq!(
+        relay
+            .wait_until_relaying_ends()
+            .await
+            .expect("the relay task must not panic")
+            .err()
+            .as_ref()
+            .and_then(io::Error::raw_os_error)
+            .map(Errno::from_raw),
+        Some(Errno::EMFILE)
+    );
+}
diff --git a/paddler_cli_tests/tests/stallable_tcp_relay_reports_an_upstream_it_cannot_reach.rs b/paddler_cli_tests/tests/stallable_tcp_relay_reports_an_upstream_it_cannot_reach.rs
new file mode 100644
index 000000000..49072a923
--- /dev/null
+++ b/paddler_cli_tests/tests/stallable_tcp_relay_reports_an_upstream_it_cannot_reach.rs
@@ -0,0 +1,42 @@
+use std::io;
+use std::io::ErrorKind;
+use std::net::Ipv4Addr;
+use std::net::SocketAddr;
+
+use tokio::io::AsyncReadExt as _;
+use tokio::net::TcpListener;
+use tokio::net::TcpStream;
+
+use paddler_cli_tests::stallable_tcp_relay::StallableTcpRelay;
+use paddler_cli_tests::unreachable_management_addr::UNREACHABLE_MANAGEMENT_ADDR;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn stallable_tcp_relay_reports_an_upstream_it_cannot_reach() {
+    let relay_listener = TcpListener::bind(SocketAddr::from((Ipv4Addr::LOCALHOST, 0)))
+        .await
+        .expect("the relay must bind a loopback port");
+    let relay_addr = relay_listener
+        .local_addr()
+        .expect("the relay listener must know its address");
+    let relay = StallableTcpRelay::start(relay_listener, UNREACHABLE_MANAGEMENT_ADDR);
+    let mut downstream = TcpStream::connect(relay_addr)
+        .await
+        .expect("the relay must accept a connection");
+    let mut bytes_relayed_back = Vec::new();
+
+    downstream
+        .read_to_end(&mut bytes_relayed_back)
+        .await
+        .expect("the relay must close a connection it cannot relay");
+
+    assert_eq!(
+        relay
+            .stop()
+            .await
+            .expect("the relay task must not panic")
+            .err()
+            .as_ref()
+            .map(io::Error::kind),
+        Some(ErrorKind::ConnectionRefused)
+    );
+}
diff --git a/paddler_cli_tests/tests/subprocess_cluster_reports_a_balancer_that_exits_before_announcing.rs b/paddler_cli_tests/tests/subprocess_cluster_reports_a_balancer_that_exits_before_announcing.rs
new file mode 100644
index 000000000..76a02a490
--- /dev/null
+++ b/paddler_cli_tests/tests/subprocess_cluster_reports_a_balancer_that_exits_before_announcing.rs
@@ -0,0 +1,33 @@
+use std::os::unix::process::ExitStatusExt as _;
+use std::process::ExitStatus;
+
+use paddler_cli_tests::clap_usage_error_exit_code::CLAP_USAGE_ERROR_EXIT_CODE;
+use paddler_cli_tests::start_subprocess_cluster::start_subprocess_cluster;
+use paddler_cli_tests::subprocess_cluster_error::SubprocessClusterError;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn subprocess_cluster_reports_a_balancer_that_exits_before_announcing() {
+    let start_error = start_subprocess_cluster(
+        env!("CARGO_BIN_EXE_paddler_cluster_node"),
+        ClusterParams {
+            state_database_url: "unsupported-scheme://state".to_owned(),
+            ..ClusterParams::without_agents()
+        },
+    )
+    .await
+    .err()
+    .expect("a balancer that rejects its arguments must fail the cluster start");
+
+    assert_eq!(
+        start_error
+            .downcast_ref::<SubprocessClusterError>()
+            .map(ToString::to_string),
+        Some(
+            SubprocessClusterError::BalancerExitedBeforeAnnouncing {
+                exit_status: ExitStatus::from_raw(CLAP_USAGE_ERROR_EXIT_CODE << 8),
+            }
+            .to_string()
+        )
+    );
+}
diff --git a/paddler_cli_tests/tests/subprocess_cluster_reports_a_connection_cancelled_before_the_balancer_is_healthy.rs b/paddler_cli_tests/tests/subprocess_cluster_reports_a_connection_cancelled_before_the_balancer_is_healthy.rs
new file mode 100644
index 000000000..a2d88818a
--- /dev/null
+++ b/paddler_cli_tests/tests/subprocess_cluster_reports_a_connection_cancelled_before_the_balancer_is_healthy.rs
@@ -0,0 +1,34 @@
+use std::mem::discriminant;
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_cli_tests::subprocess_agents_cluster_connection::SubprocessAgentsClusterConnection;
+use paddler_client::error::Error as ClientError;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
+
+use crate::running_balancer_subprocess::running_balancer_subprocess;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn subprocess_cluster_reports_a_connection_cancelled_before_the_balancer_is_healthy() {
+    let cancelled_connection = CancellationToken::new();
+
+    cancelled_connection.cancel();
+
+    assert_eq!(
+        SubprocessAgentsClusterConnection {
+            agent_binary_path: env!("CARGO_BIN_EXE_paddler_cluster_node"),
+            cancellation_token: cancelled_connection,
+            desired_state: &ClusterDesiredState::Apply(Box::default()),
+            running_balancer: running_balancer_subprocess().await,
+        }
+        .connect()
+        .await
+        .err()
+        .as_ref()
+        .and_then(|connection_error| connection_error.root_cause().downcast_ref::<ClientError>())
+        .map(discriminant),
+        Some(discriminant(&ClientError::RequestCancelled {
+            url: String::new()
+        }))
+    );
+}
diff --git a/paddler_cli_tests/tests/subprocess_cluster_reports_a_desired_state_the_balancer_cannot_store.rs b/paddler_cli_tests/tests/subprocess_cluster_reports_a_desired_state_the_balancer_cannot_store.rs
new file mode 100644
index 000000000..0c97cd74f
--- /dev/null
+++ b/paddler_cli_tests/tests/subprocess_cluster_reports_a_desired_state_the_balancer_cannot_store.rs
@@ -0,0 +1,60 @@
+use std::fs::Permissions;
+use std::fs::set_permissions;
+use std::os::unix::fs::PermissionsExt as _;
+
+use reqwest::StatusCode;
+use tempfile::TempDir;
+use url::Url;
+
+use paddler_cli_tests::start_subprocess_cluster::start_subprocess_cluster;
+use paddler_client::error::Error as ClientError;
+use paddler_messaging::agent_desired_model::AgentDesiredModel;
+use paddler_messaging::api_path::ApiPath;
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+
+const READ_ONLY_DIRECTORY_MODE: u32 = 0o555;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn subprocess_cluster_reports_a_desired_state_the_balancer_cannot_store() {
+    let state_directory = TempDir::new().expect("a temporary directory must be created");
+
+    set_permissions(
+        state_directory.path(),
+        Permissions::from_mode(READ_ONLY_DIRECTORY_MODE),
+    )
+    .expect("the state directory must become read-only");
+
+    let start_error = start_subprocess_cluster(
+        env!("CARGO_BIN_EXE_paddler_cluster_node"),
+        ClusterParams {
+            desired_state: ClusterDesiredState::Apply(Box::new(BalancerDesiredState {
+                model: AgentDesiredModel::Uri("agent:///unstorable-model.gguf".to_owned()),
+                ..BalancerDesiredState::default()
+            })),
+            state_database_url: format!(
+                "file://{}",
+                state_directory.path().join("state.json").display()
+            ),
+            ..ClusterParams::without_agents()
+        },
+    )
+    .await
+    .err()
+    .expect("a desired state the balancer cannot store must fail the cluster start");
+
+    assert!(matches!(
+        start_error.downcast_ref::<ClientError>(),
+        Some(ClientError::UnexpectedResponseStatus { message, status, url })
+            if *status == StatusCode::INTERNAL_SERVER_ERROR
+                && *message
+                    == format!(
+                        "Unable to create a replacement of the state database file in '{}'",
+                        state_directory.path().display()
+                    )
+                && Url::parse(url).is_ok_and(|desired_state_url| {
+                    desired_state_url.path() == ApiPath::BALANCER_DESIRED_STATE
+                })
+    ));
+}
diff --git a/paddler_cli_tests/tests/subprocess_cluster_reports_an_additional_agent_binary_that_cannot_be_spawned.rs b/paddler_cli_tests/tests/subprocess_cluster_reports_an_additional_agent_binary_that_cannot_be_spawned.rs
new file mode 100644
index 000000000..d87e4f527
--- /dev/null
+++ b/paddler_cli_tests/tests/subprocess_cluster_reports_an_additional_agent_binary_that_cannot_be_spawned.rs
@@ -0,0 +1,26 @@
+use std::io;
+use std::io::ErrorKind;
+
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+
+use crate::cluster_with_a_missing_agent_binary::cluster_with_a_missing_agent_binary;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn subprocess_cluster_reports_an_additional_agent_binary_that_cannot_be_spawned() {
+    let mut cluster = cluster_with_a_missing_agent_binary().await;
+
+    assert_eq!(
+        cluster
+            .spawn_additional_agent(&AgentConfig::single(1))
+            .err()
+            .as_ref()
+            .and_then(|spawn_error| spawn_error.root_cause().downcast_ref::<io::Error>())
+            .map(io::Error::kind),
+        Some(ErrorKind::NotFound)
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_cli_tests/tests/subprocess_cluster_reports_an_agent_binary_that_cannot_be_spawned.rs b/paddler_cli_tests/tests/subprocess_cluster_reports_an_agent_binary_that_cannot_be_spawned.rs
new file mode 100644
index 000000000..0406eef25
--- /dev/null
+++ b/paddler_cli_tests/tests/subprocess_cluster_reports_an_agent_binary_that_cannot_be_spawned.rs
@@ -0,0 +1,29 @@
+use std::io;
+use std::io::ErrorKind;
+
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+
+use crate::cluster_with_a_missing_agent_binary::cluster_with_a_missing_agent_binary;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn subprocess_cluster_reports_an_agent_binary_that_cannot_be_spawned() {
+    let mut cluster = cluster_with_a_missing_agent_binary().await;
+
+    let registration_error = cluster
+        .register_agents(&[AgentConfig::single(1)], false)
+        .await
+        .expect_err("an agent binary that is missing must not register");
+
+    assert_eq!(
+        registration_error
+            .root_cause()
+            .downcast_ref::<io::Error>()
+            .map(io::Error::kind),
+        Some(ErrorKind::NotFound)
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_cli_tests/tests/subprocess_cluster_reports_an_agent_that_exits_before_it_is_ready.rs b/paddler_cli_tests/tests/subprocess_cluster_reports_an_agent_that_exits_before_it_is_ready.rs
new file mode 100644
index 000000000..f177028ad
--- /dev/null
+++ b/paddler_cli_tests/tests/subprocess_cluster_reports_an_agent_that_exits_before_it_is_ready.rs
@@ -0,0 +1,27 @@
+use paddler_cli_tests::agent_exited_before_ready_with_code::agent_exited_before_ready_with_code;
+use paddler_cli_tests::clap_usage_error_exit_code::CLAP_USAGE_ERROR_EXIT_CODE;
+use paddler_cli_tests::start_subprocess_cluster::start_subprocess_cluster;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+
+const SLOTS_THE_AGENT_REJECTS: u16 = 0;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn subprocess_cluster_reports_an_agent_that_exits_before_it_is_ready() {
+    let start_error = start_subprocess_cluster(
+        env!("CARGO_BIN_EXE_paddler_cluster_node"),
+        ClusterParams {
+            agents: vec![AgentConfig::single(SLOTS_THE_AGENT_REJECTS)],
+            wait_for_slots_ready: false,
+            ..ClusterParams::default()
+        },
+    )
+    .await
+    .err()
+    .expect("an agent that exits must fail the cluster start");
+
+    assert!(agent_exited_before_ready_with_code(
+        &start_error,
+        CLAP_USAGE_ERROR_EXIT_CODE
+    ));
+}
diff --git a/paddler_cli_tests/tests/subprocess_cluster_shutdown_reports_an_agent_that_exited_with_a_failure.rs b/paddler_cli_tests/tests/subprocess_cluster_shutdown_reports_an_agent_that_exited_with_a_failure.rs
new file mode 100644
index 000000000..586762f85
--- /dev/null
+++ b/paddler_cli_tests/tests/subprocess_cluster_shutdown_reports_an_agent_that_exited_with_a_failure.rs
@@ -0,0 +1,39 @@
+use std::mem::discriminant;
+use std::os::unix::process::ExitStatusExt as _;
+use std::process::ExitStatus;
+
+use paddler_cli_tests::start_subprocess_cluster::start_subprocess_cluster;
+use paddler_cli_tests::subprocess_cluster_error::SubprocessClusterError;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+
+use crate::slots_beyond_what_llama_cpp_can_serve::SLOTS_BEYOND_WHAT_LLAMA_CPP_CAN_SERVE;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn subprocess_cluster_shutdown_reports_an_agent_that_exited_with_a_failure() {
+    let mut cluster = start_subprocess_cluster(
+        env!("CARGO_BIN_EXE_paddler_cluster_node"),
+        ClusterParams::without_agents(),
+    )
+    .await
+    .expect("a cluster without agents must start");
+
+    cluster
+        .spawn_additional_agent(&AgentConfig::single(SLOTS_BEYOND_WHAT_LLAMA_CPP_CAN_SERVE))
+        .expect("the agent process must spawn");
+
+    assert_eq!(
+        cluster
+            .shutdown()
+            .await
+            .err()
+            .as_ref()
+            .and_then(|shutdown_error| shutdown_error.downcast_ref::<SubprocessClusterError>())
+            .map(discriminant),
+        Some(discriminant(
+            &SubprocessClusterError::ProcessExitedWithFailure {
+                exit_status: ExitStatus::from_raw(0),
+            }
+        ))
+    );
+}
diff --git a/paddler_cli_tests/tests/subprocess_cluster_with_a_serving_and_a_stuck_agent.rs b/paddler_cli_tests/tests/subprocess_cluster_with_a_serving_and_a_stuck_agent.rs
new file mode 100644
index 000000000..c15b7cb37
--- /dev/null
+++ b/paddler_cli_tests/tests/subprocess_cluster_with_a_serving_and_a_stuck_agent.rs
@@ -0,0 +1,62 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use paddler_cli_tests::start_subprocess_cluster::start_subprocess_cluster;
+use paddler_inference_parameters::batch_size::BatchSize;
+use paddler_inference_parameters::model_runtime_parameters::ModelRuntimeParameters;
+use paddler_messaging::agent_issue::AgentIssue;
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster::Cluster;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
+
+pub const SERVING_AGENT_NAME: &str = "serving-agent";
+pub const SERVING_AGENT_SLOTS: u16 = 1;
+
+pub async fn subprocess_cluster_with_a_serving_and_a_stuck_agent() -> Cluster {
+    let qwen3_desired_state = qwen3_0_6b().into_desired_state();
+    let mut cluster = start_subprocess_cluster(
+        env!("CARGO_BIN_EXE_paddler_cluster_node"),
+        ClusterParams {
+            desired_state: ClusterDesiredState::Apply(Box::new(BalancerDesiredState {
+                model_runtime_parameters: ModelRuntimeParameters {
+                    n_batch: BatchSize::try_from(u32::from(SERVING_AGENT_SLOTS))
+                        .expect("a batch as large as the serving agent's slots is valid"),
+                    ..qwen3_desired_state.model_runtime_parameters
+                },
+                ..qwen3_desired_state
+            })),
+            ..ClusterParams::without_agents()
+        },
+    )
+    .await
+    .expect("the balancer must start");
+
+    cluster
+        .register_agents(
+            &[AgentConfig {
+                name: "stuck-agent".to_owned(),
+                slot_count: SERVING_AGENT_SLOTS + 1,
+            }],
+            false,
+        )
+        .await
+        .expect("the stuck agent must register");
+    cluster
+        .register_agents(
+            &[AgentConfig {
+                name: SERVING_AGENT_NAME.to_owned(),
+                slot_count: SERVING_AGENT_SLOTS,
+            }],
+            true,
+        )
+        .await
+        .expect("the serving agent must register and serve");
+    cluster
+        .wait_for_first_agent_issue(|issue| matches!(issue, AgentIssue::SlotCannotStart(_)))
+        .await
+        .expect("the stuck agent must report that its slots cannot start in a smaller batch");
+
+    cluster
+}
diff --git a/paddler_cli_tests/tests/subprocess_exit_reaped_elsewhere_is_reported_as_unobservable.rs b/paddler_cli_tests/tests/subprocess_exit_reaped_elsewhere_is_reported_as_unobservable.rs
new file mode 100644
index 000000000..46b34e623
--- /dev/null
+++ b/paddler_cli_tests/tests/subprocess_exit_reaped_elsewhere_is_reported_as_unobservable.rs
@@ -0,0 +1,44 @@
+use std::io;
+use std::process::Stdio;
+
+use nix::errno::Errno;
+use nix::sys::wait::waitpid;
+use nix::unistd::Pid;
+
+use paddler_cli_tests::paddler_command::paddler_command;
+use paddler_cli_tests::subprocess_cluster_error::SubprocessClusterError;
+use paddler_cli_tests::subprocess_process::SubprocessProcess;
+use paddler_test_cluster_harness::managed_process::ManagedProcess as _;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn subprocess_exit_reaped_elsewhere_is_reported_as_unobservable() {
+    let rejected_invocation = paddler_command(env!("CARGO_BIN_EXE_paddler_cluster_node"))
+        .arg("--flag-paddler-does-not-have")
+        .stdout(Stdio::null())
+        .stderr(Stdio::null())
+        .spawn()
+        .expect("paddler must start");
+    let rejected_invocation_pid = Pid::from_raw(
+        rejected_invocation
+            .id()
+            .expect("a running invocation must have a pid")
+            .cast_signed(),
+    );
+
+    waitpid(rejected_invocation_pid, None)
+        .expect("the rejected invocation must exit and be reaped before the harness waits");
+
+    assert_eq!(
+        SubprocessProcess::new(rejected_invocation)
+            .exited()
+            .await
+            .into_result()
+            .err()
+            .as_ref()
+            .map(ToString::to_string),
+        Some(
+            SubprocessClusterError::ProcessExitUnobservable(io::Error::from(Errno::ECHILD))
+                .to_string()
+        )
+    );
+}
diff --git a/paddler_cli_tests/tests/subprocess_shutdown_reports_a_failed_exit.rs b/paddler_cli_tests/tests/subprocess_shutdown_reports_a_failed_exit.rs
new file mode 100644
index 000000000..829f13e91
--- /dev/null
+++ b/paddler_cli_tests/tests/subprocess_shutdown_reports_a_failed_exit.rs
@@ -0,0 +1,40 @@
+use std::os::unix::process::ExitStatusExt as _;
+use std::process::ExitStatus;
+use std::process::Stdio;
+
+use paddler_cli_tests::clap_usage_error_exit_code::CLAP_USAGE_ERROR_EXIT_CODE;
+use paddler_cli_tests::paddler_command::paddler_command;
+use paddler_cli_tests::subprocess_cluster_error::SubprocessClusterError;
+use paddler_cli_tests::subprocess_process::SubprocessProcess;
+use paddler_test_cluster_harness::managed_process::ManagedProcess;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn subprocess_shutdown_reports_a_failed_exit() {
+    let mut rejected_invocation = paddler_command(env!("CARGO_BIN_EXE_paddler_cluster_node"))
+        .arg("--flag-paddler-does-not-have")
+        .stdout(Stdio::null())
+        .stderr(Stdio::null())
+        .spawn()
+        .expect("paddler must start");
+    let _rejected_invocation_status = rejected_invocation
+        .wait()
+        .await
+        .expect("the rejected invocation must exit");
+
+    let shutdown_error =
+        <dyn ManagedProcess>::shutdown(Box::new(SubprocessProcess::new(rejected_invocation)))
+            .await
+            .expect_err("a failed exit must be reported at shutdown");
+
+    assert_eq!(
+        shutdown_error
+            .downcast_ref::<SubprocessClusterError>()
+            .map(ToString::to_string),
+        Some(
+            SubprocessClusterError::ProcessExitedWithFailure {
+                exit_status: ExitStatus::from_raw(CLAP_USAGE_ERROR_EXIT_CODE << 8),
+            }
+            .to_string()
+        )
+    );
+}
diff --git a/paddler_cli_tests/tests/subprocess_shutdown_reports_a_termination_it_cannot_signal.rs b/paddler_cli_tests/tests/subprocess_shutdown_reports_a_termination_it_cannot_signal.rs
new file mode 100644
index 000000000..62bc4874b
--- /dev/null
+++ b/paddler_cli_tests/tests/subprocess_shutdown_reports_a_termination_it_cannot_signal.rs
@@ -0,0 +1,62 @@
+use libc::EPERM;
+use libc::SYS_kill;
+use nix::errno::Errno;
+use nix::sys::signal::Signal;
+
+use paddler_cli_tests::spawn_agent_subprocess::spawn_agent_subprocess;
+use paddler_cli_tests::spawn_agent_subprocess_params::SpawnAgentSubprocessParams;
+use paddler_cli_tests::subprocess_cluster_error::SubprocessClusterError;
+use paddler_cli_tests::subprocess_process::SubprocessProcess;
+use paddler_cli_tests::unreachable_management_addr::UNREACHABLE_MANAGEMENT_ADDR;
+use paddler_test_cluster_harness::managed_process::ManagedProcess;
+
+use crate::refused_syscall::RefusedSyscall;
+use crate::refused_syscalls::RefusedSyscalls;
+
+const SLOTS_THE_AGENT_REJECTS_ON_ITS_OWN: u16 = 0;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn subprocess_shutdown_reports_a_termination_it_cannot_signal() {
+    let agent_process = SubprocessProcess::new(
+        spawn_agent_subprocess(SpawnAgentSubprocessParams {
+            binary_path: env!("CARGO_BIN_EXE_paddler_cluster_node").to_owned(),
+            management_addr: UNREACHABLE_MANAGEMENT_ADDR,
+            name: "agent-that-cannot-be-terminated".to_owned(),
+            slots: SLOTS_THE_AGENT_REJECTS_ON_ITS_OWN,
+        })
+        .expect("the agent subprocess must start"),
+    );
+    let agent_pid = agent_process
+        .signals()
+        .expect("a running agent must be signallable")
+        .pid;
+
+    RefusedSyscalls {
+        errno: EPERM,
+        refused_syscalls: vec![RefusedSyscall {
+            syscall: SYS_kill,
+            rules: vec![],
+        }],
+    }
+    .apply_to_every_thread();
+
+    let shutdown_error = <dyn ManagedProcess>::shutdown(Box::new(agent_process))
+        .await
+        .expect_err("an agent that cannot be signalled must not shut down");
+
+    assert_eq!(
+        shutdown_error.to_string(),
+        SubprocessClusterError::SignalUndeliverable {
+            pid: agent_pid,
+            signal: Signal::SIGTERM,
+            source: Errno::EPERM,
+        }
+        .to_string()
+    );
+    assert_eq!(
+        shutdown_error
+            .source()
+            .and_then(|source| source.downcast_ref::<Errno>()),
+        Some(&Errno::EPERM)
+    );
+}
diff --git a/paddler_cli_tests/tests/subprocess_signals_report_a_killed_process_that_cannot_be_paused.rs b/paddler_cli_tests/tests/subprocess_signals_report_a_killed_process_that_cannot_be_paused.rs
new file mode 100644
index 000000000..fb9275289
--- /dev/null
+++ b/paddler_cli_tests/tests/subprocess_signals_report_a_killed_process_that_cannot_be_paused.rs
@@ -0,0 +1,37 @@
+use nix::sys::signal::Signal;
+use nix::sys::wait::WaitStatus;
+
+use paddler_cli_tests::spawn_agent_subprocess::spawn_agent_subprocess;
+use paddler_cli_tests::spawn_agent_subprocess_params::SpawnAgentSubprocessParams;
+use paddler_cli_tests::subprocess_cluster_error::SubprocessClusterError;
+use paddler_cli_tests::subprocess_process::SubprocessProcess;
+use paddler_cli_tests::unreachable_management_addr::UNREACHABLE_MANAGEMENT_ADDR;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn subprocess_signals_report_a_killed_process_that_cannot_be_paused() {
+    let agent_process = SubprocessProcess::new(
+        spawn_agent_subprocess(SpawnAgentSubprocessParams {
+            binary_path: env!("CARGO_BIN_EXE_paddler_cluster_node").to_owned(),
+            management_addr: UNREACHABLE_MANAGEMENT_ADDR,
+            name: "agent-that-is-killed".to_owned(),
+            slots: 1,
+        })
+        .expect("the agent subprocess must start"),
+    );
+    let signals = agent_process
+        .signals()
+        .expect("a running agent must be signallable");
+
+    signals.kill().expect("the agent must be killed");
+
+    assert_eq!(
+        signals.pause().err().as_ref().map(ToString::to_string),
+        Some(
+            SubprocessClusterError::ProcessDidNotStop {
+                pid: signals.pid,
+                status: WaitStatus::Signaled(signals.pid, Signal::SIGKILL, false),
+            }
+            .to_string()
+        )
+    );
+}
diff --git a/paddler_cli_tests/tests/subprocess_signals_report_a_reaped_process_that_cannot_be_signalled.rs b/paddler_cli_tests/tests/subprocess_signals_report_a_reaped_process_that_cannot_be_signalled.rs
new file mode 100644
index 000000000..493058d74
--- /dev/null
+++ b/paddler_cli_tests/tests/subprocess_signals_report_a_reaped_process_that_cannot_be_signalled.rs
@@ -0,0 +1,53 @@
+use std::error::Error as _;
+
+use nix::errno::Errno;
+use nix::sys::signal::Signal;
+
+use paddler_cli_tests::spawn_agent_subprocess::spawn_agent_subprocess;
+use paddler_cli_tests::spawn_agent_subprocess_params::SpawnAgentSubprocessParams;
+use paddler_cli_tests::subprocess_cluster_error::SubprocessClusterError;
+use paddler_cli_tests::subprocess_process::SubprocessProcess;
+use paddler_cli_tests::unreachable_management_addr::UNREACHABLE_MANAGEMENT_ADDR;
+use paddler_test_cluster_harness::managed_process::ManagedProcess;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn subprocess_signals_report_a_reaped_process_that_cannot_be_signalled() {
+    let agent_process = SubprocessProcess::new(
+        spawn_agent_subprocess(SpawnAgentSubprocessParams {
+            binary_path: env!("CARGO_BIN_EXE_paddler_cluster_node").to_owned(),
+            management_addr: UNREACHABLE_MANAGEMENT_ADDR,
+            name: "agent-that-is-reaped".to_owned(),
+            slots: 1,
+        })
+        .expect("the agent subprocess must start"),
+    );
+    let signals = agent_process
+        .signals()
+        .expect("a running agent must be signallable");
+
+    signals.kill().expect("the agent must be killed");
+
+    <dyn ManagedProcess>::shutdown(Box::new(agent_process))
+        .await
+        .expect("the killed agent must be reaped");
+
+    let pause_error = signals
+        .pause()
+        .expect_err("a reaped process must not be signallable");
+
+    assert_eq!(
+        pause_error.to_string(),
+        SubprocessClusterError::SignalUndeliverable {
+            pid: signals.pid,
+            signal: Signal::SIGSTOP,
+            source: Errno::ESRCH,
+        }
+        .to_string()
+    );
+    assert_eq!(
+        pause_error
+            .source()
+            .and_then(|source| source.downcast_ref::<Errno>()),
+        Some(&Errno::ESRCH)
+    );
+}
diff --git a/paddler_cli_tests/tests/the_cli_reports_a_shutdown_signal_it_cannot_listen_for.rs b/paddler_cli_tests/tests/the_cli_reports_a_shutdown_signal_it_cannot_listen_for.rs
new file mode 100644
index 000000000..63f8d2bd4
--- /dev/null
+++ b/paddler_cli_tests/tests/the_cli_reports_a_shutdown_signal_it_cannot_listen_for.rs
@@ -0,0 +1,48 @@
+use std::io;
+
+use libc::EPERM;
+use libc::SIGTERM;
+use libc::SYS_rt_sigaction;
+use seccompiler::SeccompCmpArgLen;
+use seccompiler::SeccompCmpOp;
+use seccompiler::SeccompCondition;
+use seccompiler::SeccompRule;
+
+use paddler_cli::run;
+
+use crate::refused_syscall::RefusedSyscall;
+use crate::refused_syscalls::RefusedSyscalls;
+
+#[test]
+fn the_cli_reports_a_shutdown_signal_it_cannot_listen_for() {
+    RefusedSyscalls {
+        errno: EPERM,
+        refused_syscalls: vec![RefusedSyscall {
+            syscall: SYS_rt_sigaction,
+            rules: vec![
+                SeccompRule::new(vec![
+                    SeccompCondition::new(
+                        0,
+                        SeccompCmpArgLen::Dword,
+                        SeccompCmpOp::Eq,
+                        u64::from(SIGTERM.cast_unsigned()),
+                    )
+                    .expect("the termination signal condition must be valid"),
+                ])
+                .expect("the termination signal rule must be valid"),
+            ],
+        }],
+    }
+    .apply_to_every_thread();
+
+    let cli_error = run().expect_err("a CLI that cannot listen for SIGTERM must not start");
+
+    assert_eq!(cli_error.to_string(), "failed to listen for SIGTERM");
+    assert_eq!(
+        cli_error
+            .root_cause()
+            .downcast_ref::<io::Error>()
+            .and_then(io::Error::raw_os_error),
+        Some(EPERM)
+    );
+}
diff --git a/paddler_cli_tests/tests/typesafe_system_one_reports_an_agent_killed_mid_decision_as_a_bad_gateway.rs b/paddler_cli_tests/tests/typesafe_system_one_reports_an_agent_killed_mid_decision_as_a_bad_gateway.rs
new file mode 100644
index 000000000..c4ab6c231
--- /dev/null
+++ b/paddler_cli_tests/tests/typesafe_system_one_reports_an_agent_killed_mid_decision_as_a_bad_gateway.rs
@@ -0,0 +1,64 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use reqwest::StatusCode;
+use tokio::spawn;
+
+use paddler_cli_tests::pausable_agent_cluster::PausableAgentCluster;
+use paddler_cli_tests::pausable_agent_cluster_params::PausableAgentClusterParams;
+use paddler_cli_tests::start_pausable_agent_cluster::start_pausable_agent_cluster;
+use paddler_messaging::decision_slots_minimum::DECISION_SLOTS_MINIMUM;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
+use paddler_test_cluster_harness::model_card::qwen3_5_0_8b::qwen3_5_0_8b;
+use paddler_test_cluster_harness::noul_system_one_request::noul_system_one_request;
+use paddler_test_cluster_harness::pointer_head_fixture::pointer_head_fixture;
+use paddler_test_cluster_harness::synthetic_pointer_head_fixture::QWEN3_5_0_8B_SYNTHETIC_POINTER_HEAD_FIXTURE;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn typesafe_system_one_reports_an_agent_killed_mid_decision_as_a_bad_gateway() {
+    let PausableAgentCluster {
+        mut cluster,
+        pausable_agent,
+    } = start_pausable_agent_cluster(PausableAgentClusterParams {
+        binary_path: env!("CARGO_BIN_EXE_paddler_cluster_node").to_owned(),
+        desired_state: ClusterDesiredState::Apply(Box::new(
+            qwen3_5_0_8b().into_decision_desired_state(
+                pointer_head_fixture(QWEN3_5_0_8B_SYNTHETIC_POINTER_HEAD_FIXTURE)
+                    .expect("the pointer head fixture path must be absolute"),
+            ),
+        )),
+        expected_slots_total: DECISION_SLOTS_MINIMUM.get(),
+        slot_count: DECISION_SLOTS_MINIMUM.get(),
+    })
+    .await
+    .expect("a decision cluster must start");
+
+    pausable_agent
+        .signals
+        .pause()
+        .expect("the agent must stop before the decision reaches it");
+
+    let decision = spawn(cluster.typesafe_system_one("killed-request", &noul_system_one_request()));
+
+    cluster
+        .wait_for_slots_processing(&pausable_agent.id, 1)
+        .await
+        .expect("the balancer must dispatch the decision to the paused agent");
+    pausable_agent
+        .signals
+        .kill()
+        .expect("the agent must be killed while it holds the decision");
+
+    assert_eq!(
+        decision
+            .await
+            .expect("the decision task must not panic")
+            .expect("the decision must be answered")
+            .status,
+        StatusCode::BAD_GATEWAY
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_cli_tests/tests/untrusted_agent_socket_waiting_for_a_close_frame_reports_a_balancer_that_vanished.rs b/paddler_cli_tests/tests/untrusted_agent_socket_waiting_for_a_close_frame_reports_a_balancer_that_vanished.rs
new file mode 100644
index 000000000..629c2e0a7
--- /dev/null
+++ b/paddler_cli_tests/tests/untrusted_agent_socket_waiting_for_a_close_frame_reports_a_balancer_that_vanished.rs
@@ -0,0 +1,48 @@
+use std::error::Error as _;
+use std::mem::discriminant;
+
+use tokio_tungstenite::tungstenite::Error as WebSocketError;
+use tokio_tungstenite::tungstenite::error::ProtocolError;
+
+use paddler_test_cluster_harness::cluster_harness_error::ClusterHarnessError;
+
+use crate::agent_socket_of_a_vanished_balancer::AgentSocketOfAVanishedBalancer;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn untrusted_agent_socket_waiting_for_a_close_frame_reports_a_balancer_that_vanished() {
+    let mut agent_socket_of_a_vanished_balancer = AgentSocketOfAVanishedBalancer::connect().await;
+
+    let receive_error = agent_socket_of_a_vanished_balancer
+        .agent_socket
+        .next_close_frame()
+        .await
+        .expect_err("a vanished balancer must not send a close frame");
+
+    assert_eq!(
+        discriminant(&receive_error),
+        discriminant(&ClusterHarnessError::AgentSocketReceiveFailed(
+            WebSocketError::ConnectionClosed
+        ))
+    );
+    assert_eq!(
+        receive_error
+            .source()
+            .and_then(|websocket_error| websocket_error.source())
+            .and_then(|protocol_error| protocol_error.downcast_ref::<ProtocolError>()),
+        Some(&ProtocolError::ResetWithoutClosingHandshake)
+    );
+    assert_eq!(
+        agent_socket_of_a_vanished_balancer
+            .agent_socket
+            .next_close_frame()
+            .await
+            .err()
+            .as_ref()
+            .map(discriminant),
+        Some(discriminant(
+            &ClusterHarnessError::AgentSocketEndedWithoutClosing
+        ))
+    );
+
+    agent_socket_of_a_vanished_balancer.shut_down().await;
+}
diff --git a/paddler_cli_tests/tests/untrusted_agent_socket_waiting_for_text_reports_a_balancer_that_vanished.rs b/paddler_cli_tests/tests/untrusted_agent_socket_waiting_for_text_reports_a_balancer_that_vanished.rs
new file mode 100644
index 000000000..95ea512b8
--- /dev/null
+++ b/paddler_cli_tests/tests/untrusted_agent_socket_waiting_for_text_reports_a_balancer_that_vanished.rs
@@ -0,0 +1,48 @@
+use std::error::Error as _;
+use std::mem::discriminant;
+
+use tokio_tungstenite::tungstenite::Error as WebSocketError;
+use tokio_tungstenite::tungstenite::error::ProtocolError;
+
+use paddler_test_cluster_harness::cluster_harness_error::ClusterHarnessError;
+
+use crate::agent_socket_of_a_vanished_balancer::AgentSocketOfAVanishedBalancer;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn untrusted_agent_socket_waiting_for_text_reports_a_balancer_that_vanished() {
+    let mut agent_socket_of_a_vanished_balancer = AgentSocketOfAVanishedBalancer::connect().await;
+
+    let receive_error = agent_socket_of_a_vanished_balancer
+        .agent_socket
+        .next_text()
+        .await
+        .expect_err("a vanished balancer must not send text");
+
+    assert_eq!(
+        discriminant(&receive_error),
+        discriminant(&ClusterHarnessError::AgentSocketReceiveFailed(
+            WebSocketError::ConnectionClosed
+        ))
+    );
+    assert_eq!(
+        receive_error
+            .source()
+            .and_then(|websocket_error| websocket_error.source())
+            .and_then(|protocol_error| protocol_error.downcast_ref::<ProtocolError>()),
+        Some(&ProtocolError::ResetWithoutClosingHandshake)
+    );
+    assert_eq!(
+        agent_socket_of_a_vanished_balancer
+            .agent_socket
+            .next_text()
+            .await
+            .err()
+            .as_ref()
+            .map(discriminant),
+        Some(discriminant(
+            &ClusterHarnessError::AgentSocketEndedWithoutText
+        ))
+    );
+
+    agent_socket_of_a_vanished_balancer.shut_down().await;
+}
diff --git a/paddler_client/Cargo.toml b/paddler_client/Cargo.toml
index 7497135aa..50536c12f 100644
--- a/paddler_client/Cargo.toml
+++ b/paddler_client/Cargo.toml
@@ -27,7 +27,6 @@ url = { workspace = true }
 
 [dev-dependencies]
 http = { workspace = true }
-paddler_local_http_fixture = { workspace = true }
 tokio = { workspace = true, features = ["test-util"] }
 
 [lints]
diff --git a/paddler_client/src/client_inference.rs b/paddler_client/src/client_inference.rs
index bc7053c5a..a13ed79bc 100644
--- a/paddler_client/src/client_inference.rs
+++ b/paddler_client/src/client_inference.rs
@@ -2,18 +2,17 @@ use std::sync::Arc;
 
 use nanoid::nanoid;
 use serde::Serialize;
-use tokio::sync::broadcast;
 use tokio_util::sync::CancellationToken;
 
 use paddler_messaging::api_path::ApiPath;
 use paddler_messaging::inference_client::message::Message as InferenceMessage;
-use paddler_messaging::inference_client::notification::Notification;
 use paddler_messaging::inference_server::message::Message as InferenceServerMessage;
 use paddler_messaging::inference_server::request::Request as InferenceServerRequest;
 use paddler_messaging::jsonrpc::request_envelope::RequestEnvelope;
 use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
 use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters_schema::validated_parameters_schema::ValidatedParametersSchema;
 use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
+use paddler_messaging::request_params::decide_params::raw_decide_params::RawDecideParams;
 use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
 
 use crate::client_inference_params::ClientInferenceParams;
@@ -80,11 +79,6 @@ impl ClientInference {
             .await
     }
 
-    #[must_use]
-    pub fn subscribe_to_token_generation_mode(&self) -> broadcast::Receiver<Notification> {
-        self.inference_socket_pool.subscribe_to_notifications()
-    }
-
     pub async fn continue_from_conversation_history(
         &self,
         cancellation_token: CancellationToken,
@@ -135,6 +129,15 @@ impl ClientInference {
         .await
     }
 
+    pub async fn post_decide(
+        &self,
+        cancellation_token: CancellationToken,
+        params: &RawDecideParams,
+    ) -> Result<InferenceMessageStream> {
+        self.post_streaming(cancellation_token, ApiPath::DECIDE, params)
+            .await
+    }
+
     pub async fn post_generate_embedding_batch(
         &self,
         cancellation_token: CancellationToken,
@@ -157,6 +160,7 @@ impl ReportsHealth for ClientInference {
 
 #[cfg(test)]
 mod tests {
+    use std::mem::discriminant;
     use std::num::NonZeroU32;
     use std::num::NonZeroUsize;
 
@@ -244,12 +248,21 @@ use paddler_messaging::request_params::generate_embedding_batch_params::Generate
 
         let rejection = unreachable_client()
             .continue_from_raw_prompt(cancellation_token, raw_prompt_params())
-            .await;
-
-        assert!(matches!(
-            rejection,
-            Err(Error::InferenceRequestCancelled { request_id }) if !request_id.is_empty()
-        ));
+            .await
+            .err()
+            .expect("a cancelled request must be rejected");
+        let rejection_without_a_request_id = Error::InferenceRequestCancelled {
+            request_id: String::new(),
+        };
+
+        assert_eq!(
+            discriminant(&rejection),
+            discriminant(&rejection_without_a_request_id)
+        );
+        assert_ne!(
+            rejection.to_string(),
+            rejection_without_a_request_id.to_string()
+        );
     }
 
     #[tokio::test]
diff --git a/paddler_client/src/format_api_url.rs b/paddler_client/src/format_api_url.rs
index 3adf21591..0471bb9f0 100644
--- a/paddler_client/src/format_api_url.rs
+++ b/paddler_client/src/format_api_url.rs
@@ -1,5 +1,6 @@
 use url::Url;
 
+#[must_use]
 pub fn format_api_url(base_url: &Url, path: &str) -> String {
     format!("{}{}", base_url.as_str().trim_end_matches('/'), path)
 }
diff --git a/paddler_client/src/http_client.rs b/paddler_client/src/http_client.rs
index 769431159..6bdf8fd9f 100644
--- a/paddler_client/src/http_client.rs
+++ b/paddler_client/src/http_client.rs
@@ -15,6 +15,10 @@ use crate::format_api_url::format_api_url;
 use crate::send_checked_request::send_checked_request;
 use crate::stream::sse::Sse;
 
+async fn response_text(response: Response) -> Result<String> {
+    Ok(response.text().await?)
+}
+
 #[derive(Clone)]
 pub struct HttpClient {
     reqwest_client: Client,
@@ -62,7 +66,7 @@ impl HttpClient {
         cancellation_token: CancellationToken,
         path: &str,
     ) -> Result<String> {
-        Ok(self.get(cancellation_token, path).await?.text().await?)
+        response_text(self.get(cancellation_token, path).await?).await
     }
 
     pub async fn post_json<TBody: Serialize + Sync + ?Sized>(
@@ -92,16 +96,21 @@ impl HttpClient {
 
 #[cfg(test)]
 mod tests {
-    use futures_util::StreamExt as _;
-    use http::StatusCode;
+    use std::io::Error as IoError;
+    use std::io::ErrorKind;
+    use std::io::Result as IoResult;
+
+    use futures_util::stream::iter;
+    use http::Response as HttpResponse;
+    use reqwest::Body;
+    use reqwest::Response;
     use tokio_util::sync::CancellationToken;
     use url::Url;
 
-    use paddler_local_http_fixture::fixture_response::FixtureResponse;
-    use paddler_local_http_fixture::local_http_fixture::LocalHttpFixture;
     use paddler_messaging::api_path::ApiPath;
 
     use super::HttpClient;
+    use super::response_text;
     use crate::error::Error;
 
     const UNREACHABLE_BASE_URL: &str = "http://127.0.0.1:1";
@@ -110,16 +119,6 @@ mod tests {
         HttpClient::new(Url::parse(UNREACHABLE_BASE_URL).expect("the test URL must be valid"))
     }
 
-    fn client_of(fixture: &LocalHttpFixture) -> HttpClient {
-        HttpClient::new(Url::parse(&fixture.url("/")).expect("the fixture URL must be valid"))
-    }
-
-    async fn fixture_serving(response: FixtureResponse) -> LocalHttpFixture {
-        LocalHttpFixture::start(response)
-            .await
-            .expect("the fixture server must start")
-    }
-
     fn cancelled_token() -> CancellationToken {
         let cancellation_token = CancellationToken::new();
 
@@ -193,44 +192,14 @@ mod tests {
     }
 
     #[tokio::test]
-    async fn a_body_that_is_not_json_is_reported_as_undecodable() {
-        let fixture = fixture_serving(FixtureResponse::Ok(b"not json".to_vec())).await;
+    async fn a_text_body_that_breaks_off_is_reported_as_unreadable() {
+        let chunks: [IoResult<Vec<u8>>; 2] = [
+            Ok(b"paddler_".to_vec()),
+            Err(IoError::from(ErrorKind::UnexpectedEof)),
+        ];
 
         assert!(matches!(
-            client_of(&fixture)
-                .get_json::<String>(CancellationToken::new(), ApiPath::AGENTS)
-                .await,
-            Err(Error::Http(source)) if source.is_decode()
-        ));
-    }
-
-    #[tokio::test]
-    async fn a_server_sent_event_that_is_not_json_is_reported_as_undecodable() {
-        let fixture = fixture_serving(FixtureResponse::Ok(b"data: not json\n\n".to_vec())).await;
-        let mut events = client_of(&fixture)
-            .get_sse_json::<String>(CancellationToken::new(), ApiPath::AGENTS_STREAM)
-            .await
-            .expect("the event stream must open");
-
-        assert!(matches!(
-            events.next().await,
-            Some(Err(Error::Json(source))) if source.is_syntax()
-        ));
-    }
-
-    #[tokio::test]
-    async fn a_truncated_text_body_is_reported_as_unreadable() {
-        let fixture = fixture_serving(FixtureResponse::TruncatedBody {
-            sent_body: b"paddler_".to_vec(),
-            status: StatusCode::OK,
-            withheld_byte_count: 8,
-        })
-        .await;
-
-        assert!(matches!(
-            client_of(&fixture)
-                .get_text(CancellationToken::new(), ApiPath::METRICS)
-                .await,
+            response_text(Response::from(HttpResponse::new(Body::wrap_stream(iter(chunks))))).await,
             Err(Error::Http(source)) if source.is_decode()
         ));
     }
diff --git a/paddler_client/src/inference_socket/awaited_response.rs b/paddler_client/src/inference_socket/awaited_response.rs
new file mode 100644
index 000000000..f0186685d
--- /dev/null
+++ b/paddler_client/src/inference_socket/awaited_response.rs
@@ -0,0 +1,10 @@
+use tokio::sync::mpsc::UnboundedSender;
+
+use paddler_messaging::inference_client::message::Message as InferenceMessage;
+
+use crate::error::Result;
+
+pub struct AwaitedResponse {
+    pub response_tx: UnboundedSender<Result<InferenceMessage>>,
+    pub stop_frame: String,
+}
diff --git a/paddler_client/src/inference_socket/connection.rs b/paddler_client/src/inference_socket/connection.rs
index c408b9bd4..21d6e10cb 100644
--- a/paddler_client/src/inference_socket/connection.rs
+++ b/paddler_client/src/inference_socket/connection.rs
@@ -1,87 +1,67 @@
 use std::sync::Arc;
 
 use futures_util::StreamExt;
-use serde_json::to_string;
-use tokio::sync::broadcast;
-use tokio::sync::mpsc;
+use serde::Serialize;
 use tokio::sync::mpsc::UnboundedReceiver;
-use tokio::sync::mpsc::UnboundedSender;
 use tokio::task::JoinHandle;
 use tokio_tungstenite::connect_async;
 use url::Url;
 
 use paddler_messaging::inference_client::message::Message as InferenceMessage;
-use paddler_messaging::inference_client::notification::Notification;
-use paddler_messaging::inference_server::message::Message as InferenceServerMessage;
-use paddler_messaging::inference_server::notification::Notification as InferenceServerNotification;
-use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters_schema::validated_parameters_schema::ValidatedParametersSchema;
 
-use crate::error::Error;
 use crate::error::Result;
-use crate::inference_socket::pending_requests::PendingRequests;
+use crate::inference_socket::inbound_message_router::InboundMessageRouter;
+use crate::inference_socket::inference_socket_exchange::InferenceSocketExchange;
+use crate::inference_socket::outbound_request::OutboundRequest;
 use crate::inference_socket::spawn_read_task::spawn_read_task;
 use crate::inference_socket::spawn_write_task::spawn_write_task;
 use crate::inference_socket::url::url;
 
 pub struct Connection {
-    pending: Arc<PendingRequests>,
+    exchange: Arc<InferenceSocketExchange>,
     read_task: JoinHandle<()>,
     write_task: JoinHandle<()>,
-    write_tx: UnboundedSender<String>,
 }
 
 impl Connection {
-    pub async fn connect(
-        connection_url: Url,
-        notification_tx: broadcast::Sender<Notification>,
-    ) -> Result<Self> {
+    pub async fn connect(connection_url: Url) -> Result<Self> {
         let ws_url = url(connection_url)?;
         let (ws_stream, _) = connect_async(ws_url.as_str()).await?;
         let (ws_write, ws_read) = ws_stream.split();
-        let pending = Arc::new(PendingRequests::default());
-        let (write_tx, write_rx) = mpsc::unbounded_channel::<String>();
+        let exchange = Arc::new(InferenceSocketExchange::default());
 
         Ok(Self {
-            read_task: spawn_read_task(ws_read, pending.clone(), notification_tx),
-            write_task: spawn_write_task(ws_write, write_rx, pending.clone()),
-            pending,
-            write_tx,
+            read_task: spawn_read_task(
+                ws_read,
+                InboundMessageRouter {
+                    exchange: exchange.clone(),
+                },
+            ),
+            write_task: spawn_write_task(ws_write, exchange.clone()),
+            exchange,
         })
     }
 
+    pub async fn disconnected(&self) {
+        self.exchange.closed().await;
+    }
+
     #[must_use]
     pub fn is_disconnected(&self) -> bool {
-        self.pending.is_closed()
+        self.exchange.is_closed()
     }
 
-    pub fn send(
+    pub fn send<TMessage: Serialize>(
         &self,
         request_id: String,
-        json: String,
+        message: &TMessage,
     ) -> Result<UnboundedReceiver<Result<InferenceMessage>>> {
-        let response_rx = self.pending.register(request_id.clone())?;
-
-        if self.write_tx.send(json).is_err() {
-            self.pending.remove(&request_id);
-
-            return Err(Error::ConnectionDropped { request_id });
-        }
-
-        Ok(response_rx)
+        OutboundRequest::serialize(request_id, message)
+            .and_then(|outbound_request| self.exchange.register(outbound_request))
     }
 
-    pub fn stop_responding_to(&self, request_id: String) -> Result<()> {
-        self.pending.remove(&request_id);
-
-        let stop_responding_to: InferenceServerMessage<ValidatedParametersSchema> =
-            InferenceServerMessage::Notification(InferenceServerNotification::StopRespondingTo(
-                request_id.clone(),
-            ));
-        let json = to_string(&stop_responding_to)?;
-
-        self.write_tx
-            .send(json)
-            .map_err(|_closed_channel| Error::ConnectionDropped { request_id })
+    pub fn stop_responding_to(&self, request_id: &str) {
+        self.exchange.stop_responding_to(request_id);
     }
 }
 
@@ -96,84 +76,35 @@ impl Drop for Connection {
 mod tests {
     use std::io::ErrorKind;
 
-    use futures_util::SinkExt as _;
-    use futures_util::StreamExt as _;
+    use tokio::io::AsyncReadExt as _;
+    use tokio::io::AsyncWriteExt as _;
+    use tokio::join;
     use tokio::net::TcpListener;
-    use tokio::net::TcpStream;
-    use tokio::spawn;
-    use tokio::sync::broadcast;
-    use tokio::task::JoinHandle;
-    use tokio_tungstenite::WebSocketStream;
-    use tokio_tungstenite::accept_async;
     use tokio_tungstenite::tungstenite::Error as WebSocketError;
-    use tokio_tungstenite::tungstenite::Message as WsMessage;
     use url::Url;
 
     use super::Connection;
     use crate::error::Error;
 
-    struct WebSocketFixture {
-        server: JoinHandle<WebSocketStream<TcpStream>>,
-        url: Url,
-    }
+    const PLAINTEXT_HTTP_REPLY_START: &[u8] = b"HTTP/";
+    const TLS_HANDSHAKE_RECORD_CONTENT_TYPE: u8 = 22;
 
-    async fn websocket_fixture_answering_the_first_request_with(
-        answer: Option<WsMessage>,
-    ) -> WebSocketFixture {
-        let listener = TcpListener::bind("127.0.0.1:0")
-            .await
-            .expect("the fixture socket must bind");
-        let url = Url::parse(&format!(
-            "http://{}",
-            listener
-                .local_addr()
-                .expect("the fixture socket must report its address")
-        ))
-        .expect("the fixture URL must be valid");
-        let server = spawn(async move {
-            let (stream, _peer_addr) = listener
-                .accept()
-                .await
-                .expect("the fixture must accept the client");
-            let mut websocket = accept_async(stream)
-                .await
-                .expect("the fixture must complete the handshake");
-
-            websocket
-                .next()
-                .await
-                .expect("the client must send its request")
-                .expect("the client request must be readable");
-
-            if let Some(answer) = answer {
-                websocket
-                    .send(answer)
-                    .await
-                    .expect("the fixture must send its answer");
-            }
-
-            websocket
-        });
-
-        WebSocketFixture { server, url }
-    }
-
-    async fn connection_to(url: Url) -> Connection {
-        let (notification_tx, _notification_rx) = broadcast::channel(1);
-
-        Connection::connect(url, notification_tx)
-            .await
-            .expect("the client must connect to the fixture")
+    #[tokio::test]
+    async fn connect_rejects_a_url_that_cannot_reach_an_inference_socket() {
+        assert!(matches!(
+            Connection::connect(
+                Url::parse("ftp://127.0.0.1/models").expect("the test URL must be valid"),
+            )
+            .await,
+            Err(Error::InferenceSocketUrlSchemeRejected { scheme, .. }) if scheme == "ftp"
+        ));
     }
 
     #[tokio::test]
     async fn connect_fails_for_an_unreachable_server() {
-        let (notification_tx, _notification_rx) = broadcast::channel(1);
-
         assert!(matches!(
             Connection::connect(
                 Url::parse("http://127.0.0.1:1").expect("the test URL must be valid"),
-                notification_tx,
             )
             .await,
             Err(Error::WebSocket(WebSocketError::Io(io_error)))
@@ -182,38 +113,45 @@ mod tests {
     }
 
     #[tokio::test]
-    async fn a_binary_frame_from_the_server_drops_every_pending_request() {
-        let WebSocketFixture { server, url } = websocket_fixture_answering_the_first_request_with(
-            Some(WsMessage::Binary(vec![0].into())),
-        )
-        .await;
-        let connection = connection_to(url).await;
-        let mut response_rx = connection
-            .send("request-1".to_owned(), "{}".to_owned())
-            .expect("an open connection must accept a request");
-
-        assert!(matches!(
-            response_rx.recv().await,
-            Some(Err(Error::ConnectionDropped { request_id })) if request_id == "request-1"
-        ));
-
-        drop(server.await.expect("the fixture must not panic"));
-    }
+    async fn connect_opens_a_tls_handshake_with_an_https_balancer() {
+        let listener = TcpListener::bind("127.0.0.1:0")
+            .await
+            .expect("the test listener must bind");
+        let listener_address = listener
+            .local_addr()
+            .expect("the test listener must have an address");
+        let plaintext_server = async move {
+            let (mut accepted_stream, _peer_address) =
+                listener.accept().await.expect("the client must connect");
+
+            accepted_stream
+                .write_all(PLAINTEXT_HTTP_REPLY_START)
+                .await
+                .expect("the plaintext reply must reach the client");
 
-    #[tokio::test]
-    async fn a_connection_reset_without_a_closing_handshake_drops_every_pending_request() {
-        let WebSocketFixture { server, url } =
-            websocket_fixture_answering_the_first_request_with(None).await;
-        let connection = connection_to(url).await;
-        let mut response_rx = connection
-            .send("request-1".to_owned(), "{}".to_owned())
-            .expect("an open connection must accept a request");
+            accepted_stream
+        };
+        let connection = Connection::connect(
+            Url::parse(&format!("https://{listener_address}")).expect("the test URL must be valid"),
+        );
 
-        drop(server.await.expect("the fixture must not panic"));
+        let (mut accepted_stream, connection_outcome) = join!(plaintext_server, connection);
 
-        assert!(matches!(
-            response_rx.recv().await,
-            Some(Err(Error::ConnectionDropped { request_id })) if request_id == "request-1"
-        ));
+        assert_eq!(
+            accepted_stream
+                .read_u8()
+                .await
+                .expect("the client must have sent a record"),
+            TLS_HANDSHAKE_RECORD_CONTENT_TYPE
+        );
+        assert_eq!(
+            connection_outcome
+                .err()
+                .map(|connect_error| connect_error.to_string()),
+            Some(
+                "WebSocket error: IO error: received corrupt message of type InvalidContentType"
+                    .to_owned()
+            )
+        );
     }
 }
diff --git a/paddler_client/src/inference_socket/inbound_message_router.rs b/paddler_client/src/inference_socket/inbound_message_router.rs
index 999a5f4a0..2a48e4f99 100644
--- a/paddler_client/src/inference_socket/inbound_message_router.rs
+++ b/paddler_client/src/inference_socket/inbound_message_router.rs
@@ -1,35 +1,36 @@
 use std::sync::Arc;
 
-use log::debug;
 use log::error;
 use log::warn;
 use serde_json::from_str;
-use tokio::sync::broadcast;
 
 use paddler_messaging::inference_client::identified_message::IdentifiedMessage;
 use paddler_messaging::inference_client::message::Message as InferenceMessage;
-use paddler_messaging::inference_client::notification::Notification;
 use paddler_messaging::inference_client::response::Response;
-use paddler_messaging::streamable_result::StreamableResult;
+use paddler_messaging::jsonrpc::response_envelope::ResponseEnvelope;
+use paddler_messaging::result_step::ResultStep;
 
 use crate::error::Error;
-use crate::inference_socket::pending_requests::PendingRequests;
-
-fn response_is_terminal(response: &Response) -> bool {
-    match response {
-        Response::Embedding(result) => result.is_done(),
-        Response::GeneratedToken(result) => result.is_done(),
-    }
-}
+use crate::inference_socket::inference_socket_exchange::InferenceSocketExchange;
 
 struct RequestScopedMessage {
     is_done: bool,
+    message: InferenceMessage,
     request_id: String,
 }
 
+impl RequestScopedMessage {
+    fn of_response(is_done: bool, envelope: ResponseEnvelope<Response>) -> Self {
+        Self {
+            is_done,
+            request_id: envelope.request_id.clone(),
+            message: InferenceMessage::Response(envelope),
+        }
+    }
+}
+
 pub struct InboundMessageRouter {
-    pub notification_tx: broadcast::Sender<Notification>,
-    pub pending: Arc<PendingRequests>,
+    pub exchange: Arc<InferenceSocketExchange>,
 }
 
 impl InboundMessageRouter {
@@ -42,7 +43,7 @@ impl InboundMessageRouter {
 
                     error!("Failed to decode a message for request {request_id}: {decoding_error}");
 
-                    if !self.pending.fail(
+                    if !self.exchange.fail(
                         &request_id,
                         Error::UndecodableInferenceMessage {
                             request_id: request_id.clone(),
@@ -64,27 +65,39 @@ impl InboundMessageRouter {
     }
 
     fn route(&self, message: InferenceMessage) {
-        let request_scoped_message = match &message {
+        let request_scoped_message = match message {
             InferenceMessage::Error(envelope) => RequestScopedMessage {
                 is_done: true,
                 request_id: envelope.request_id.clone(),
+                message: InferenceMessage::Error(envelope),
             },
-            InferenceMessage::Notification(notification) => {
-                if self.notification_tx.send(notification.clone()).is_err() {
-                    debug!("Dropped inference notification: no active subscribers");
-                }
-
-                return;
-            }
-            InferenceMessage::Response(envelope) => RequestScopedMessage {
-                is_done: response_is_terminal(&envelope.response),
-                request_id: envelope.request_id.clone(),
+            InferenceMessage::Response(ResponseEnvelope {
+                generated_by,
+                request_id,
+                response,
+            }) => match response.into_step() {
+                ResultStep::Final(final_response) => RequestScopedMessage::of_response(
+                    true,
+                    ResponseEnvelope {
+                        generated_by,
+                        request_id,
+                        response: final_response,
+                    },
+                ),
+                ResultStep::Partial(partial_response) => RequestScopedMessage::of_response(
+                    false,
+                    ResponseEnvelope {
+                        generated_by,
+                        request_id,
+                        response: partial_response,
+                    },
+                ),
             },
         };
 
-        if !self.pending.deliver(
+        if !self.exchange.deliver(
             &request_scoped_message.request_id,
-            message,
+            request_scoped_message.message,
             request_scoped_message.is_done,
         ) {
             warn!(
@@ -100,9 +113,9 @@ mod tests {
     use std::sync::Arc;
 
     use serde_json::json;
-    use tokio::sync::broadcast;
     use tokio::sync::mpsc::error::TryRecvError;
 
+    use paddler_messaging::decision_result::DecisionResult;
     use paddler_messaging::embedding_result::EmbeddingResult;
     use paddler_messaging::inference_client::message::Message as InferenceMessage;
     use paddler_messaging::inference_client::response::Response;
@@ -110,21 +123,26 @@ mod tests {
 
     use super::InboundMessageRouter;
     use crate::error::Error;
-    use crate::inference_socket::pending_requests::PendingRequests;
-
-    #[test]
-    fn fails_the_request_whose_message_cannot_be_decoded() {
-        let pending = Arc::new(PendingRequests::default());
-        let mut response_rx = pending
-            .register("request-1".to_owned())
-            .expect("an open registry must accept a request");
-        let (notification_tx, _notification_rx) = broadcast::channel(1);
+    use crate::inference_socket::inference_socket_exchange::InferenceSocketExchange;
+    use crate::inference_socket::outbound_request::OutboundRequest;
 
+    fn router_for(exchange: &Arc<InferenceSocketExchange>) -> InboundMessageRouter {
         InboundMessageRouter {
-            notification_tx,
-            pending: Arc::clone(&pending),
+            exchange: Arc::clone(exchange),
         }
-        .route_text(
+    }
+
+    #[test]
+    fn fails_the_request_whose_message_cannot_be_decoded() {
+        let exchange = Arc::new(InferenceSocketExchange::default());
+        let mut response_rx = exchange
+            .register(
+                OutboundRequest::serialize("request-1".to_owned(), &json!({}))
+                    .expect("a JSON request must serialize"),
+            )
+            .expect("an open exchange must accept a request");
+
+        router_for(&exchange).route_text(
             &json!({
                 "Response": {
                     "generated_by": null,
@@ -143,17 +161,15 @@ mod tests {
 
     #[test]
     fn leaves_other_requests_pending_when_an_undecodable_message_names_an_unknown_request() {
-        let pending = Arc::new(PendingRequests::default());
-        let mut response_rx = pending
-            .register("request-1".to_owned())
-            .expect("an open registry must accept a request");
-        let (notification_tx, _notification_rx) = broadcast::channel(1);
-
-        InboundMessageRouter {
-            notification_tx,
-            pending: Arc::clone(&pending),
-        }
-        .route_text(
+        let exchange = Arc::new(InferenceSocketExchange::default());
+        let mut response_rx = exchange
+            .register(
+                OutboundRequest::serialize("request-1".to_owned(), &json!({}))
+                    .expect("a JSON request must serialize"),
+            )
+            .expect("an open exchange must accept a request");
+
+        router_for(&exchange).route_text(
             &json!({ "Error": { "request_id": "request-2", "error": { "NoSuchError": {} } } })
                 .to_string(),
         );
@@ -167,17 +183,15 @@ mod tests {
 
     #[test]
     fn leaves_requests_pending_when_an_undecodable_message_names_no_request() {
-        let pending = Arc::new(PendingRequests::default());
-        let mut response_rx = pending
-            .register("request-1".to_owned())
-            .expect("an open registry must accept a request");
-        let (notification_tx, _notification_rx) = broadcast::channel(1);
+        let exchange = Arc::new(InferenceSocketExchange::default());
+        let mut response_rx = exchange
+            .register(
+                OutboundRequest::serialize("request-1".to_owned(), &json!({}))
+                    .expect("a JSON request must serialize"),
+            )
+            .expect("an open exchange must accept a request");
 
-        InboundMessageRouter {
-            notification_tx,
-            pending: Arc::clone(&pending),
-        }
-        .route_text("not json");
+        router_for(&exchange).route_text("not json");
 
         assert!(
             response_rx
@@ -187,62 +201,49 @@ mod tests {
     }
 
     #[test]
-    fn a_finished_embedding_batch_releases_its_request() {
-        let pending = Arc::new(PendingRequests::default());
-        let mut response_rx = pending
-            .register("request-1".to_owned())
-            .expect("an open registry must accept a request");
-        let (notification_tx, _notification_rx) = broadcast::channel(1);
-
-        InboundMessageRouter {
-            notification_tx,
-            pending: Arc::clone(&pending),
-        }
-        .route_text(
-            &json!({
-                "Response": {
-                    "generated_by": null,
-                    "request_id": "request-1",
-                    "response": { "Embedding": "Done" }
-                }
-            })
-            .to_string(),
-        );
-
-        assert!(matches!(
-            response_rx.try_recv(),
-            Ok(Ok(InferenceMessage::Response(ResponseEnvelope {
-                response: Response::Embedding(embedding_result),
-                ..
-            }))) if embedding_result == EmbeddingResult::Done
-        ));
-        assert!(
-            response_rx
-                .try_recv()
-                .is_err_and(|receive_error| receive_error == TryRecvError::Disconnected)
-        );
-    }
-
-    #[test]
-    fn a_notification_nobody_listens_to_leaves_requests_pending() {
-        let pending = Arc::new(PendingRequests::default());
-        let mut response_rx = pending
-            .register("request-1".to_owned())
-            .expect("an open registry must accept a request");
-        let (notification_tx, notification_rx) = broadcast::channel(1);
-
-        drop(notification_rx);
+    fn a_terminal_response_releases_its_request() {
+        for (wire_response, terminal_response) in [
+            (
+                json!({ "Embedding": "Done" }),
+                Response::Embedding(EmbeddingResult::Done),
+            ),
+            (
+                json!({ "Decision": "StopRequested" }),
+                Response::Decision(DecisionResult::StopRequested),
+            ),
+        ] {
+            let exchange = Arc::new(InferenceSocketExchange::default());
+            let mut response_rx = exchange
+                .register(
+                    OutboundRequest::serialize("request-1".to_owned(), &json!({}))
+                        .expect("a JSON request must serialize"),
+                )
+                .expect("an open exchange must accept a request");
+
+            router_for(&exchange).route_text(
+                &json!({
+                    "Response": {
+                        "generated_by": null,
+                        "request_id": "request-1",
+                        "response": wire_response
+                    }
+                })
+                .to_string(),
+            );
 
-        InboundMessageRouter {
-            notification_tx,
-            pending: Arc::clone(&pending),
+            assert_eq!(
+                response_rx.try_recv().ok().and_then(Result::ok),
+                Some(InferenceMessage::Response(ResponseEnvelope {
+                    generated_by: None,
+                    request_id: "request-1".to_owned(),
+                    response: terminal_response,
+                }))
+            );
+            assert!(
+                response_rx
+                    .try_recv()
+                    .is_err_and(|receive_error| receive_error == TryRecvError::Disconnected)
+            );
         }
-        .route_text(&json!({ "Notification": "TokenGenerationDisabled" }).to_string());
-
-        assert!(
-            response_rx
-                .try_recv()
-                .is_err_and(|receive_error| receive_error == TryRecvError::Empty)
-        );
     }
 }
diff --git a/paddler_client/src/inference_socket/inference_socket_exchange.rs b/paddler_client/src/inference_socket/inference_socket_exchange.rs
new file mode 100644
index 000000000..3cbf97605
--- /dev/null
+++ b/paddler_client/src/inference_socket/inference_socket_exchange.rs
@@ -0,0 +1,357 @@
+use std::collections::hash_map::Entry;
+use std::mem;
+use std::pin::pin;
+
+use log::debug;
+use parking_lot::Mutex;
+use tokio::sync::Notify;
+use tokio::sync::mpsc;
+use tokio::sync::mpsc::UnboundedReceiver;
+
+use paddler_messaging::inference_client::message::Message as InferenceMessage;
+
+use crate::error::Error;
+use crate::error::Result;
+use crate::inference_socket::awaited_response::AwaitedResponse;
+use crate::inference_socket::inference_socket_exchange_state::InferenceSocketExchangeState;
+use crate::inference_socket::open_inference_socket_exchange::OpenInferenceSocketExchange;
+use crate::inference_socket::outbound_frames::OutboundFrames;
+use crate::inference_socket::outbound_request::OutboundRequest;
+
+pub struct InferenceSocketExchange {
+    closing: Notify,
+    frames_queued: Notify,
+    state: Mutex<InferenceSocketExchangeState>,
+}
+
+impl InferenceSocketExchange {
+    pub fn close(&self) {
+        let previous_state = mem::replace(
+            &mut *self.state.lock(),
+            InferenceSocketExchangeState::Closed,
+        );
+
+        self.closing.notify_waiters();
+        self.frames_queued.notify_waiters();
+
+        if let InferenceSocketExchangeState::Open(OpenInferenceSocketExchange {
+            awaiting_responses,
+            ..
+        }) = previous_state
+        {
+            for (request_id, AwaitedResponse { response_tx, .. }) in awaiting_responses {
+                if response_tx
+                    .send(Err(Error::ConnectionDropped {
+                        request_id: request_id.clone(),
+                    }))
+                    .is_err()
+                {
+                    debug!("Receiver already dropped for request: {request_id}");
+                }
+            }
+        }
+    }
+
+    pub async fn closed(&self) {
+        let mut closing = pin!(self.closing.notified());
+
+        closing.as_mut().enable();
+
+        if !self.is_closed() {
+            closing.await;
+        }
+    }
+
+    pub fn deliver(&self, request_id: &str, message: InferenceMessage, is_done: bool) -> bool {
+        if let InferenceSocketExchangeState::Open(open_exchange) = &mut *self.state.lock()
+            && let Some(AwaitedResponse { response_tx, .. }) =
+                open_exchange.awaiting_responses.get(request_id)
+        {
+            let send_failed = response_tx.send(Ok(message)).is_err();
+
+            if is_done || send_failed {
+                open_exchange.awaiting_responses.remove(request_id);
+            }
+
+            return true;
+        }
+
+        false
+    }
+
+    pub fn fail(&self, request_id: &str, error: Error) -> bool {
+        if let InferenceSocketExchangeState::Open(open_exchange) = &mut *self.state.lock()
+            && let Some(AwaitedResponse {
+                response_tx,
+                stop_frame,
+            }) = open_exchange.awaiting_responses.remove(request_id)
+        {
+            if response_tx.send(Err(error)).is_err() {
+                debug!("Receiver already dropped for request: {request_id}");
+            }
+
+            open_exchange.queued_frames.push(stop_frame);
+            self.frames_queued.notify_waiters();
+
+            return true;
+        }
+
+        false
+    }
+
+    #[must_use]
+    pub fn is_closed(&self) -> bool {
+        matches!(*self.state.lock(), InferenceSocketExchangeState::Closed)
+    }
+
+    pub async fn next_outbound_frames(&self) -> OutboundFrames {
+        loop {
+            let mut frames_queued = pin!(self.frames_queued.notified());
+
+            frames_queued.as_mut().enable();
+
+            match &mut *self.state.lock() {
+                InferenceSocketExchangeState::Closed => return OutboundFrames::ExchangeClosed,
+                InferenceSocketExchangeState::Open(open_exchange)
+                    if !open_exchange.queued_frames.is_empty() =>
+                {
+                    return OutboundFrames::Queued(mem::take(&mut open_exchange.queued_frames));
+                }
+                InferenceSocketExchangeState::Open(_) => {}
+            }
+
+            frames_queued.await;
+        }
+    }
+
+    pub fn register(
+        &self,
+        OutboundRequest {
+            request_frame,
+            request_id,
+            stop_frame,
+        }: OutboundRequest,
+    ) -> Result<UnboundedReceiver<Result<InferenceMessage>>> {
+        match &mut *self.state.lock() {
+            InferenceSocketExchangeState::Closed => Err(Error::ConnectionDropped { request_id }),
+            InferenceSocketExchangeState::Open(open_exchange) => {
+                match open_exchange.awaiting_responses.entry(request_id) {
+                    Entry::Occupied(in_flight_request) => Err(Error::InferenceRequestIdInFlight {
+                        request_id: in_flight_request.key().clone(),
+                    }),
+                    Entry::Vacant(vacant_request) => {
+                        let (response_tx, response_rx) = mpsc::unbounded_channel();
+
+                        vacant_request.insert(AwaitedResponse {
+                            response_tx,
+                            stop_frame,
+                        });
+                        open_exchange.queued_frames.push(request_frame);
+                        self.frames_queued.notify_waiters();
+
+                        Ok(response_rx)
+                    }
+                }
+            }
+        }
+    }
+
+    pub fn stop_responding_to(&self, request_id: &str) {
+        if let InferenceSocketExchangeState::Open(open_exchange) = &mut *self.state.lock()
+            && let Some(AwaitedResponse { stop_frame, .. }) =
+                open_exchange.awaiting_responses.remove(request_id)
+        {
+            open_exchange.queued_frames.push(stop_frame);
+            self.frames_queued.notify_waiters();
+        }
+    }
+}
+
+impl Default for InferenceSocketExchange {
+    fn default() -> Self {
+        Self {
+            closing: Notify::new(),
+            frames_queued: Notify::new(),
+            state: Mutex::new(InferenceSocketExchangeState::Open(
+                OpenInferenceSocketExchange::default(),
+            )),
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use futures_util::FutureExt as _;
+    use serde_json::json;
+    use tokio::sync::mpsc::error::TryRecvError;
+
+    use super::InferenceSocketExchange;
+    use crate::error::Error;
+    use crate::inference_socket::outbound_frames::OutboundFrames;
+    use crate::inference_socket::outbound_request::OutboundRequest;
+
+    fn outbound_request(request_id: &str) -> OutboundRequest {
+        OutboundRequest::serialize(request_id.to_owned(), &json!({ "request": request_id }))
+            .expect("a JSON request must serialize")
+    }
+
+    #[test]
+    fn closing_notifies_open_requests_after_one_receiver_was_dropped() {
+        let exchange = InferenceSocketExchange::default();
+        let abandoned_request = exchange
+            .register(outbound_request("abandoned"))
+            .expect("an open exchange must accept a request");
+        let mut awaited_request = exchange
+            .register(outbound_request("awaited"))
+            .expect("an open exchange must accept a request");
+
+        drop(abandoned_request);
+        exchange.close();
+
+        assert!(matches!(
+            awaited_request.try_recv(),
+            Ok(Err(Error::ConnectionDropped { request_id })) if request_id == "awaited"
+        ));
+    }
+
+    #[test]
+    fn closing_an_already_closed_exchange_tells_each_request_only_once() {
+        let exchange = InferenceSocketExchange::default();
+        let mut awaited_request = exchange
+            .register(outbound_request("awaited"))
+            .expect("an open exchange must accept a request");
+
+        exchange.close();
+        exchange.close();
+
+        assert!(matches!(
+            awaited_request.try_recv(),
+            Ok(Err(Error::ConnectionDropped { request_id })) if request_id == "awaited"
+        ));
+        assert_eq!(
+            awaited_request.try_recv().err(),
+            Some(TryRecvError::Disconnected)
+        );
+    }
+
+    #[test]
+    fn waiting_for_an_already_closed_exchange_resolves_at_once() {
+        let exchange = InferenceSocketExchange::default();
+
+        exchange.close();
+
+        assert_eq!(exchange.closed().now_or_never(), Some(()));
+    }
+
+    #[test]
+    fn a_closed_exchange_tells_its_writer_to_stop() {
+        let exchange = InferenceSocketExchange::default();
+
+        exchange.close();
+
+        assert_eq!(
+            exchange.next_outbound_frames().now_or_never(),
+            Some(OutboundFrames::ExchangeClosed)
+        );
+    }
+
+    #[test]
+    fn stopping_an_awaited_request_queues_its_stop_frame_after_its_request_frame() {
+        let exchange = InferenceSocketExchange::default();
+        let stopped_request = outbound_request("stopped");
+        let expected_frames = vec![
+            stopped_request.request_frame.clone(),
+            stopped_request.stop_frame.clone(),
+        ];
+        let stopped_response_rx = exchange
+            .register(stopped_request)
+            .expect("an open exchange must accept a request");
+
+        exchange.stop_responding_to("stopped");
+        drop(stopped_response_rx);
+
+        assert_eq!(
+            exchange.next_outbound_frames().now_or_never(),
+            Some(OutboundFrames::Queued(expected_frames))
+        );
+    }
+
+    #[test]
+    fn failing_a_request_queues_its_stop_frame_after_its_request_frame() {
+        let exchange = InferenceSocketExchange::default();
+        let failed_request = outbound_request("failed");
+        let expected_frames = vec![
+            failed_request.request_frame.clone(),
+            failed_request.stop_frame.clone(),
+        ];
+        let _failed_response_rx = exchange
+            .register(failed_request)
+            .expect("an open exchange must accept a request");
+
+        exchange.fail(
+            "failed",
+            Error::InferenceRequestCancelled {
+                request_id: "failed".to_owned(),
+            },
+        );
+
+        assert_eq!(
+            exchange.next_outbound_frames().now_or_never(),
+            Some(OutboundFrames::Queued(expected_frames))
+        );
+    }
+
+    #[test]
+    fn failing_a_request_that_is_not_pending_fails_nothing() {
+        let exchange = InferenceSocketExchange::default();
+
+        assert!(!exchange.fail(
+            "unknown",
+            Error::InferenceRequestCancelled {
+                request_id: "unknown".to_owned(),
+            },
+        ));
+    }
+
+    #[test]
+    fn failing_an_abandoned_request_forgets_it() {
+        let exchange = InferenceSocketExchange::default();
+        let abandoned_request = exchange
+            .register(outbound_request("abandoned"))
+            .expect("an open exchange must accept a request");
+
+        drop(abandoned_request);
+
+        assert!(exchange.fail(
+            "abandoned",
+            Error::InferenceRequestCancelled {
+                request_id: "abandoned".to_owned(),
+            },
+        ));
+        assert!(!exchange.fail(
+            "abandoned",
+            Error::InferenceRequestCancelled {
+                request_id: "abandoned".to_owned(),
+            },
+        ));
+    }
+
+    #[test]
+    fn a_second_registration_of_an_in_flight_request_keeps_the_first_one() {
+        let exchange = InferenceSocketExchange::default();
+        let mut first_registration = exchange
+            .register(outbound_request("request-1"))
+            .expect("an open exchange must accept a request");
+        assert!(matches!(
+            exchange.register(outbound_request("request-1")),
+            Err(Error::InferenceRequestIdInFlight { request_id }) if request_id == "request-1"
+        ));
+
+        exchange.close();
+
+        assert!(matches!(
+            first_registration.try_recv(),
+            Ok(Err(Error::ConnectionDropped { request_id })) if request_id == "request-1"
+        ));
+    }
+}
diff --git a/paddler_client/src/inference_socket/inference_socket_exchange_state.rs b/paddler_client/src/inference_socket/inference_socket_exchange_state.rs
new file mode 100644
index 000000000..a910e70ef
--- /dev/null
+++ b/paddler_client/src/inference_socket/inference_socket_exchange_state.rs
@@ -0,0 +1,6 @@
+use crate::inference_socket::open_inference_socket_exchange::OpenInferenceSocketExchange;
+
+pub enum InferenceSocketExchangeState {
+    Closed,
+    Open(OpenInferenceSocketExchange),
+}
diff --git a/paddler_client/src/inference_socket/mod.rs b/paddler_client/src/inference_socket/mod.rs
index 5b0ab2ff0..39fe050da 100644
--- a/paddler_client/src/inference_socket/mod.rs
+++ b/paddler_client/src/inference_socket/mod.rs
@@ -1,8 +1,12 @@
+pub mod awaited_response;
 pub mod connection;
 pub mod connection_slot;
 pub mod inbound_message_router;
-pub mod pending_requests;
-pub mod pending_requests_state;
+pub mod inference_socket_exchange;
+pub mod inference_socket_exchange_state;
+pub mod open_inference_socket_exchange;
+pub mod outbound_frames;
+pub mod outbound_request;
 pub mod pool;
 pub mod response_stream;
 pub mod spawn_read_task;
diff --git a/paddler_client/src/inference_socket/open_inference_socket_exchange.rs b/paddler_client/src/inference_socket/open_inference_socket_exchange.rs
new file mode 100644
index 000000000..3b5d68b25
--- /dev/null
+++ b/paddler_client/src/inference_socket/open_inference_socket_exchange.rs
@@ -0,0 +1,9 @@
+use std::collections::HashMap;
+
+use crate::inference_socket::awaited_response::AwaitedResponse;
+
+#[derive(Default)]
+pub struct OpenInferenceSocketExchange {
+    pub awaiting_responses: HashMap<String, AwaitedResponse>,
+    pub queued_frames: Vec<String>,
+}
diff --git a/paddler_client/src/inference_socket/outbound_frames.rs b/paddler_client/src/inference_socket/outbound_frames.rs
new file mode 100644
index 000000000..f23e394b9
--- /dev/null
+++ b/paddler_client/src/inference_socket/outbound_frames.rs
@@ -0,0 +1,5 @@
+#[derive(Debug, Eq, PartialEq)]
+pub enum OutboundFrames {
+    ExchangeClosed,
+    Queued(Vec<String>),
+}
diff --git a/paddler_client/src/inference_socket/outbound_request.rs b/paddler_client/src/inference_socket/outbound_request.rs
new file mode 100644
index 000000000..31ee3f6cf
--- /dev/null
+++ b/paddler_client/src/inference_socket/outbound_request.rs
@@ -0,0 +1,34 @@
+use serde::Serialize;
+use serde_json::to_string;
+
+use paddler_messaging::inference_server::message::Message as InferenceServerMessage;
+use paddler_messaging::inference_server::notification::Notification as InferenceServerNotification;
+use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters_schema::validated_parameters_schema::ValidatedParametersSchema;
+
+use crate::error::Error;
+use crate::error::Result;
+
+pub struct OutboundRequest {
+    pub request_frame: String,
+    pub request_id: String,
+    pub stop_frame: String,
+}
+
+impl OutboundRequest {
+    pub fn serialize<TMessage: Serialize>(request_id: String, message: &TMessage) -> Result<Self> {
+        to_string(message)
+            .and_then(|request_frame| {
+                to_string(
+                    &InferenceServerMessage::<ValidatedParametersSchema>::Notification(
+                        InferenceServerNotification::StopRespondingTo(request_id.clone()),
+                    ),
+                )
+                .map(|stop_frame| Self {
+                    request_frame,
+                    request_id,
+                    stop_frame,
+                })
+            })
+            .map_err(Error::Json)
+    }
+}
diff --git a/paddler_client/src/inference_socket/pending_requests.rs b/paddler_client/src/inference_socket/pending_requests.rs
deleted file mode 100644
index 90592c98f..000000000
--- a/paddler_client/src/inference_socket/pending_requests.rs
+++ /dev/null
@@ -1,186 +0,0 @@
-use std::collections::HashMap;
-use std::collections::hash_map::Entry;
-use std::mem;
-
-use log::debug;
-use parking_lot::Mutex;
-use tokio::sync::mpsc;
-use tokio::sync::mpsc::UnboundedReceiver;
-
-use paddler_messaging::inference_client::message::Message as InferenceMessage;
-
-use crate::error::Error;
-use crate::error::Result;
-use crate::inference_socket::pending_requests_state::PendingRequestsState;
-
-pub struct PendingRequests {
-    state: Mutex<PendingRequestsState>,
-}
-
-impl PendingRequests {
-    pub fn close(&self) {
-        let previous_state = mem::replace(&mut *self.state.lock(), PendingRequestsState::Closed);
-
-        if let PendingRequestsState::Open(requests) = previous_state {
-            for (request_id, response_tx) in requests {
-                if response_tx
-                    .send(Err(Error::ConnectionDropped {
-                        request_id: request_id.clone(),
-                    }))
-                    .is_err()
-                {
-                    debug!("Receiver already dropped for request: {request_id}");
-                }
-            }
-        }
-    }
-
-    pub fn deliver(&self, request_id: &str, message: InferenceMessage, is_done: bool) -> bool {
-        if let PendingRequestsState::Open(requests) = &mut *self.state.lock()
-            && let Some(response_tx) = requests.get(request_id)
-        {
-            let send_failed = response_tx.send(Ok(message)).is_err();
-
-            if is_done || send_failed {
-                requests.remove(request_id);
-            }
-
-            return true;
-        }
-
-        false
-    }
-
-    pub fn fail(&self, request_id: &str, error: Error) -> bool {
-        if let PendingRequestsState::Open(requests) = &mut *self.state.lock()
-            && let Some(response_tx) = requests.remove(request_id)
-        {
-            if response_tx.send(Err(error)).is_err() {
-                debug!("Receiver already dropped for request: {request_id}");
-            }
-
-            return true;
-        }
-
-        false
-    }
-
-    #[must_use]
-    pub fn is_closed(&self) -> bool {
-        matches!(*self.state.lock(), PendingRequestsState::Closed)
-    }
-
-    pub fn register(
-        &self,
-        request_id: String,
-    ) -> Result<UnboundedReceiver<Result<InferenceMessage>>> {
-        match &mut *self.state.lock() {
-            PendingRequestsState::Closed => Err(Error::ConnectionDropped { request_id }),
-            PendingRequestsState::Open(requests) => match requests.entry(request_id) {
-                Entry::Occupied(in_flight_request) => Err(Error::InferenceRequestIdInFlight {
-                    request_id: in_flight_request.key().clone(),
-                }),
-                Entry::Vacant(vacant_request) => {
-                    let (response_tx, response_rx) = mpsc::unbounded_channel();
-
-                    vacant_request.insert(response_tx);
-
-                    Ok(response_rx)
-                }
-            },
-        }
-    }
-
-    pub fn remove(&self, request_id: &str) {
-        if let PendingRequestsState::Open(requests) = &mut *self.state.lock() {
-            requests.remove(request_id);
-        }
-    }
-}
-
-impl Default for PendingRequests {
-    fn default() -> Self {
-        Self {
-            state: Mutex::new(PendingRequestsState::Open(HashMap::new())),
-        }
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use super::PendingRequests;
-    use crate::error::Error;
-
-    #[test]
-    fn closing_notifies_open_requests_after_one_receiver_was_dropped() {
-        let pending_requests = PendingRequests::default();
-        let abandoned_request = pending_requests
-            .register("abandoned".to_owned())
-            .expect("an open registry must accept a request");
-        let mut awaited_request = pending_requests
-            .register("awaited".to_owned())
-            .expect("an open registry must accept a request");
-
-        drop(abandoned_request);
-        pending_requests.close();
-
-        assert!(matches!(
-            awaited_request.try_recv(),
-            Ok(Err(Error::ConnectionDropped { request_id })) if request_id == "awaited"
-        ));
-    }
-
-    #[test]
-    fn failing_a_request_that_is_not_pending_fails_nothing() {
-        let pending_requests = PendingRequests::default();
-
-        assert!(!pending_requests.fail(
-            "unknown",
-            Error::InferenceRequestCancelled {
-                request_id: "unknown".to_owned(),
-            },
-        ));
-    }
-
-    #[test]
-    fn failing_an_abandoned_request_forgets_it() {
-        let pending_requests = PendingRequests::default();
-        let abandoned_request = pending_requests
-            .register("abandoned".to_owned())
-            .expect("an open registry must accept a request");
-
-        drop(abandoned_request);
-
-        assert!(pending_requests.fail(
-            "abandoned",
-            Error::InferenceRequestCancelled {
-                request_id: "abandoned".to_owned(),
-            },
-        ));
-        assert!(!pending_requests.fail(
-            "abandoned",
-            Error::InferenceRequestCancelled {
-                request_id: "abandoned".to_owned(),
-            },
-        ));
-    }
-
-    #[test]
-    fn a_second_registration_of_an_in_flight_request_keeps_the_first_one() {
-        let pending_requests = PendingRequests::default();
-        let mut first_registration = pending_requests
-            .register("request-1".to_owned())
-            .expect("an open registry must accept a request");
-        assert!(matches!(
-            pending_requests.register("request-1".to_owned()),
-            Err(Error::InferenceRequestIdInFlight { request_id }) if request_id == "request-1"
-        ));
-
-        pending_requests.close();
-
-        assert!(matches!(
-            first_registration.try_recv(),
-            Ok(Err(Error::ConnectionDropped { request_id })) if request_id == "request-1"
-        ));
-    }
-}
diff --git a/paddler_client/src/inference_socket/pending_requests_state.rs b/paddler_client/src/inference_socket/pending_requests_state.rs
deleted file mode 100644
index 865775403..000000000
--- a/paddler_client/src/inference_socket/pending_requests_state.rs
+++ /dev/null
@@ -1,12 +0,0 @@
-use std::collections::HashMap;
-
-use tokio::sync::mpsc::UnboundedSender;
-
-use paddler_messaging::inference_client::message::Message as InferenceMessage;
-
-use crate::error::Result;
-
-pub enum PendingRequestsState {
-    Closed,
-    Open(HashMap<String, UnboundedSender<Result<InferenceMessage>>>),
-}
diff --git a/paddler_client/src/inference_socket/pool.rs b/paddler_client/src/inference_socket/pool.rs
index 2138d13df..2b442061d 100644
--- a/paddler_client/src/inference_socket/pool.rs
+++ b/paddler_client/src/inference_socket/pool.rs
@@ -4,15 +4,12 @@ use std::sync::atomic::AtomicUsize;
 use std::sync::atomic::Ordering;
 
 use serde::Serialize;
-use serde_json::to_string;
 use tokio::sync::Mutex;
-use tokio::sync::broadcast;
 use tokio::sync::mpsc::UnboundedReceiver;
 use tokio_util::sync::CancellationToken;
 use url::Url;
 
 use paddler_messaging::inference_client::message::Message as InferenceMessage;
-use paddler_messaging::inference_client::notification::Notification;
 
 use crate::error::Error;
 use crate::error::Result;
@@ -29,39 +26,29 @@ struct EstablishedRequest {
 pub struct Pool {
     connection_slots: Vec<Mutex<ConnectionSlot>>,
     next_slot_index: AtomicUsize,
-    notification_tx: broadcast::Sender<Notification>,
     url: Url,
 }
 
 impl Pool {
     #[must_use]
     pub fn new(url: Url, capacity: NonZeroUsize) -> Self {
-        let (notification_tx, _initial_notification_rx) = broadcast::channel(capacity.get());
-
         Self {
             connection_slots: (0..capacity.get())
                 .map(|_slot_index| Mutex::new(ConnectionSlot::Empty))
                 .collect(),
             next_slot_index: AtomicUsize::new(0),
-            notification_tx,
             url,
         }
     }
 
-    pub fn subscribe_to_notifications(&self) -> broadcast::Receiver<Notification> {
-        self.notification_tx.subscribe()
-    }
-
-    pub async fn send_request<TMessage: Serialize>(
+    pub async fn send_request<TMessage: Serialize + Sync>(
         &self,
         cancellation_token: CancellationToken,
         request_id: String,
         message: TMessage,
     ) -> Result<InferenceMessageStream> {
-        let json = to_string(&message)?;
-
         let Some(established_request_result) = cancellation_token
-            .run_until_cancelled(self.establish_request(json, request_id.clone()))
+            .run_until_cancelled(self.establish_request(&message, request_id.clone()))
             .await
         else {
             return Err(Error::InferenceRequestCancelled { request_id });
@@ -89,21 +76,20 @@ impl Pool {
             return Ok(connection.clone());
         }
 
-        let connection =
-            Arc::new(Connection::connect(self.url.clone(), self.notification_tx.clone()).await?);
+        let connection = Arc::new(Connection::connect(self.url.clone()).await?);
 
         *connection_slot = ConnectionSlot::Connected(connection.clone());
 
         Ok(connection)
     }
 
-    async fn establish_request(
+    async fn establish_request<TMessage: Serialize + Sync>(
         &self,
-        json: String,
+        message: &TMessage,
         request_id: String,
     ) -> Result<EstablishedRequest> {
         let connection = self.connected(self.next_connection_slot()).await?;
-        let response_rx = connection.send(request_id, json)?;
+        let response_rx = connection.send(request_id, message)?;
 
         Ok(EstablishedRequest {
             connection,
diff --git a/paddler_client/src/inference_socket/response_stream.rs b/paddler_client/src/inference_socket/response_stream.rs
index 398fa3e5f..6d2c682ff 100644
--- a/paddler_client/src/inference_socket/response_stream.rs
+++ b/paddler_client/src/inference_socket/response_stream.rs
@@ -2,7 +2,6 @@ use std::sync::Arc;
 
 use futures_util::Stream;
 use futures_util::stream::unfold;
-use log::debug;
 use tokio::sync::mpsc::UnboundedReceiver;
 use tokio_util::sync::CancellationToken;
 
@@ -27,15 +26,8 @@ impl StreamState {
 
 impl Drop for StreamState {
     fn drop(&mut self) {
-        if self.is_complete {
-            return;
-        }
-
-        if let Err(stop_error) = self.connection.stop_responding_to(self.request_id.clone()) {
-            debug!(
-                "Could not ask the balancer to stop responding to request {}: {stop_error}",
-                self.request_id
-            );
+        if !self.is_complete {
+            self.connection.stop_responding_to(&self.request_id);
         }
     }
 }
diff --git a/paddler_client/src/inference_socket/spawn_read_task.rs b/paddler_client/src/inference_socket/spawn_read_task.rs
index 87a983ef2..4989842e0 100644
--- a/paddler_client/src/inference_socket/spawn_read_task.rs
+++ b/paddler_client/src/inference_socket/spawn_read_task.rs
@@ -1,35 +1,23 @@
-use std::sync::Arc;
-
+use futures_util::Stream;
 use futures_util::StreamExt;
-use futures_util::stream::SplitStream;
 use log::error;
-use tokio::net::TcpStream;
 use tokio::spawn;
-use tokio::sync::broadcast;
 use tokio::task::JoinHandle;
-use tokio_tungstenite::MaybeTlsStream;
-use tokio_tungstenite::WebSocketStream;
+use tokio_tungstenite::tungstenite::Error as WebSocketError;
 use tokio_tungstenite::tungstenite::Message as WsMessage;
 
-use paddler_messaging::inference_client::notification::Notification;
-
 use crate::inference_socket::inbound_message_router::InboundMessageRouter;
-use crate::inference_socket::pending_requests::PendingRequests;
-
-type WebSocketReadStream = SplitStream<WebSocketStream<MaybeTlsStream<TcpStream>>>;
 
 #[must_use]
-pub fn spawn_read_task(
-    ws_read: WebSocketReadStream,
-    pending: Arc<PendingRequests>,
-    notification_tx: broadcast::Sender<Notification>,
-) -> JoinHandle<()> {
+pub fn spawn_read_task<TWebSocketReadStream>(
+    ws_read: TWebSocketReadStream,
+    router: InboundMessageRouter,
+) -> JoinHandle<()>
+where
+    TWebSocketReadStream: Stream<Item = Result<WsMessage, WebSocketError>> + Send + Unpin + 'static,
+{
     spawn(async move {
         let mut ws_read = ws_read;
-        let router = InboundMessageRouter {
-            notification_tx,
-            pending,
-        };
 
         while let Some(msg_result) = ws_read.next().await {
             match msg_result {
@@ -49,6 +37,102 @@ pub fn spawn_read_task(
             }
         }
 
-        router.pending.close();
+        router.exchange.close();
     })
 }
+
+#[cfg(test)]
+mod tests {
+    use std::sync::Arc;
+
+    use futures_util::StreamExt as _;
+    use futures_util::stream::iter;
+    use futures_util::stream::pending;
+    use serde_json::json;
+    use tokio_tungstenite::tungstenite::Bytes;
+    use tokio_tungstenite::tungstenite::Message as WsMessage;
+
+    use paddler_messaging::embedding_result::EmbeddingResult;
+    use paddler_messaging::inference_client::message::Message as InferenceMessage;
+    use paddler_messaging::inference_client::response::Response;
+    use paddler_messaging::jsonrpc::response_envelope::ResponseEnvelope;
+
+    use super::spawn_read_task;
+    use crate::error::Error;
+    use crate::inference_socket::inbound_message_router::InboundMessageRouter;
+    use crate::inference_socket::inference_socket_exchange::InferenceSocketExchange;
+    use crate::inference_socket::outbound_request::OutboundRequest;
+
+    #[tokio::test]
+    async fn a_binary_frame_drops_every_pending_request_without_waiting_for_more_frames() {
+        let exchange = Arc::new(InferenceSocketExchange::default());
+        let mut response_rx = exchange
+            .register(
+                OutboundRequest::serialize("request-1".to_owned(), &json!({}))
+                    .expect("a JSON request must serialize"),
+            )
+            .expect("an open connection must accept a request");
+
+        spawn_read_task(
+            iter([Ok(WsMessage::Binary(vec![0].into()))]).chain(pending()),
+            InboundMessageRouter { exchange },
+        )
+        .await
+        .expect("the read task must not panic");
+
+        assert_eq!(
+            response_rx
+                .recv()
+                .await
+                .and_then(Result::err)
+                .as_ref()
+                .map(ToString::to_string),
+            Some(
+                Error::ConnectionDropped {
+                    request_id: "request-1".to_owned(),
+                }
+                .to_string()
+            )
+        );
+    }
+
+    #[tokio::test]
+    async fn a_ping_between_frames_leaves_the_following_response_for_its_request() {
+        let exchange = Arc::new(InferenceSocketExchange::default());
+        let mut response_rx = exchange
+            .register(
+                OutboundRequest::serialize("request-1".to_owned(), &json!({}))
+                    .expect("a JSON request must serialize"),
+            )
+            .expect("an open connection must accept a request");
+
+        spawn_read_task(
+            iter([
+                Ok(WsMessage::Ping(Bytes::new())),
+                Ok(WsMessage::Text(
+                    json!({
+                        "Response": {
+                            "generated_by": null,
+                            "request_id": "request-1",
+                            "response": { "Embedding": "Done" }
+                        }
+                    })
+                    .to_string()
+                    .into(),
+                )),
+            ]),
+            InboundMessageRouter { exchange },
+        )
+        .await
+        .expect("the read task must not panic");
+
+        assert_eq!(
+            response_rx.recv().await.and_then(Result::ok),
+            Some(InferenceMessage::Response(ResponseEnvelope {
+                generated_by: None,
+                request_id: "request-1".to_owned(),
+                response: Response::Embedding(EmbeddingResult::Done),
+            }))
+        );
+    }
+}
diff --git a/paddler_client/src/inference_socket/spawn_write_task.rs b/paddler_client/src/inference_socket/spawn_write_task.rs
index e99c6c32b..024a3b836 100644
--- a/paddler_client/src/inference_socket/spawn_write_task.rs
+++ b/paddler_client/src/inference_socket/spawn_write_task.rs
@@ -1,35 +1,74 @@
+use std::fmt::Display;
 use std::sync::Arc;
 
+use futures_util::Sink;
 use futures_util::SinkExt;
-use futures_util::stream::SplitSink;
 use log::error;
-use tokio::net::TcpStream;
 use tokio::spawn;
-use tokio::sync::mpsc::UnboundedReceiver;
 use tokio::task::JoinHandle;
-use tokio_tungstenite::MaybeTlsStream;
-use tokio_tungstenite::WebSocketStream;
 use tokio_tungstenite::tungstenite::Message as WsMessage;
 
-use crate::inference_socket::pending_requests::PendingRequests;
+use crate::inference_socket::inference_socket_exchange::InferenceSocketExchange;
+use crate::inference_socket::outbound_frames::OutboundFrames;
 
-type WebSocketWriteSink = SplitSink<WebSocketStream<MaybeTlsStream<TcpStream>>, WsMessage>;
-
-pub fn spawn_write_task(
-    ws_write: WebSocketWriteSink,
-    write_rx: UnboundedReceiver<String>,
-    pending: Arc<PendingRequests>,
-) -> JoinHandle<()> {
+pub fn spawn_write_task<TWebSocketWriteSink>(
+    ws_write: TWebSocketWriteSink,
+    exchange: Arc<InferenceSocketExchange>,
+) -> JoinHandle<()>
+where
+    TWebSocketWriteSink: Sink<WsMessage> + Send + Unpin + 'static,
+    TWebSocketWriteSink::Error: Display,
+{
     spawn(async move {
         let mut ws_write = ws_write;
-        let mut write_rx = write_rx;
 
-        while let Some(message) = write_rx.recv().await {
-            if let Err(err) = ws_write.send(WsMessage::Text(message.into())).await {
-                error!("WebSocket write error: {err}");
-                pending.close();
-                break;
+        while let OutboundFrames::Queued(frames) = exchange.next_outbound_frames().await {
+            for frame in frames {
+                if let Err(write_error) = ws_write.send(WsMessage::Text(frame.into())).await {
+                    error!("WebSocket write error: {write_error}");
+                    exchange.close();
+
+                    return;
+                }
             }
         }
     })
 }
+
+#[cfg(test)]
+mod tests {
+    use std::sync::Arc;
+
+    use serde_json::json;
+    use tokio::sync::mpsc::channel;
+    use tokio_tungstenite::tungstenite::Message as WsMessage;
+    use tokio_util::sync::PollSender;
+
+    use super::spawn_write_task;
+    use crate::error::Error;
+    use crate::inference_socket::inference_socket_exchange::InferenceSocketExchange;
+    use crate::inference_socket::outbound_request::OutboundRequest;
+
+    #[tokio::test]
+    async fn a_failed_write_drops_every_awaiting_request() {
+        let exchange = Arc::new(InferenceSocketExchange::default());
+        let (frames_tx, frames_rx) = channel::<WsMessage>(1);
+        let mut response_rx = exchange
+            .register(
+                OutboundRequest::serialize("request-1".to_owned(), &json!({}))
+                    .expect("a JSON request must serialize"),
+            )
+            .expect("an open exchange must accept a request");
+
+        drop(frames_rx);
+
+        spawn_write_task(PollSender::new(frames_tx), exchange)
+            .await
+            .expect("the write task must not panic");
+
+        assert!(matches!(
+            response_rx.recv().await,
+            Some(Err(Error::ConnectionDropped { request_id })) if request_id == "request-1"
+        ));
+    }
+}
diff --git a/paddler_client/src/inference_socket/url.rs b/paddler_client/src/inference_socket/url.rs
index 767e5d5c2..3cfeb97a8 100644
--- a/paddler_client/src/inference_socket/url.rs
+++ b/paddler_client/src/inference_socket/url.rs
@@ -5,27 +5,31 @@ use paddler_messaging::api_path::ApiPath;
 use crate::error::Error;
 use crate::error::Result;
 
+fn websocket_scheme(scheme: &str) -> Option<&'static str> {
+    match scheme {
+        "http" | "ws" => Some("ws"),
+        "https" | "wss" => Some("wss"),
+        _ => None,
+    }
+}
+
 pub fn url(input: Url) -> Result<Url> {
-    let websocket_scheme = match input.scheme() {
-        "http" | "ws" => "ws",
-        "https" | "wss" => "wss",
-        rejected_scheme => {
-            return Err(Error::InferenceSocketUrlSchemeRejected {
-                scheme: rejected_scheme.to_owned(),
-                url: input.to_string(),
-            });
-        }
-    };
     let mut url = input;
 
-    url.set_scheme(websocket_scheme)
-        .map_err(|()| Error::InferenceSocketUrlSchemeRejected {
-            scheme: websocket_scheme.to_owned(),
-            url: url.to_string(),
-        })?;
-    url.set_path(ApiPath::INFERENCE_SOCKET);
+    match websocket_scheme(url.scheme())
+        .ok_or(())
+        .and_then(|scheme| url.set_scheme(scheme))
+    {
+        Ok(()) => {
+            url.set_path(ApiPath::INFERENCE_SOCKET);
 
-    Ok(url)
+            Ok(url)
+        }
+        Err(()) => Err(Error::InferenceSocketUrlSchemeRejected {
+            scheme: url.scheme().to_owned(),
+            url: url.to_string(),
+        }),
+    }
 }
 
 #[cfg(test)]
diff --git a/paddler_client/src/lib.rs b/paddler_client/src/lib.rs
index 71f225355..6e58a163e 100644
--- a/paddler_client/src/lib.rs
+++ b/paddler_client/src/lib.rs
@@ -5,7 +5,7 @@ pub mod client_inference;
 pub mod client_inference_params;
 pub mod client_management;
 pub mod error;
-mod format_api_url;
+pub mod format_api_url;
 pub mod http_client;
 pub mod inference_message_stream;
 pub mod inference_socket;
diff --git a/paddler_client/src/send_checked_request.rs b/paddler_client/src/send_checked_request.rs
index a6b565e66..2156541d8 100644
--- a/paddler_client/src/send_checked_request.rs
+++ b/paddler_client/src/send_checked_request.rs
@@ -21,6 +21,14 @@ async fn rejection(response: Response, url: String) -> Result<Error> {
     }
 }
 
+async fn checked_response(response: Response, url: String) -> Result<Response> {
+    if response.status().is_success() {
+        return Ok(response);
+    }
+
+    Err(rejection(response, url).await?)
+}
+
 pub async fn send_checked_request(
     cancellation_token: CancellationToken,
     url: String,
@@ -34,8 +42,7 @@ pub async fn send_checked_request(
     };
 
     match send_result {
-        Ok(response) if response.status().is_success() => Ok(response),
-        Ok(response) => Err(rejection(response, url).await?),
+        Ok(response) => checked_response(response, url).await,
         Err(source) if source.is_connect() => Err(Error::Connect { url, source }),
         Err(source) => Err(Error::Http(source)),
     }
@@ -43,31 +50,23 @@ pub async fn send_checked_request(
 
 #[cfg(test)]
 mod tests {
+    use std::io::Error as IoError;
+    use std::io::ErrorKind;
+    use std::io::Result as IoResult;
+
+    use futures_util::stream::iter;
+    use http::Response as HttpResponse;
     use http::StatusCode;
+    use reqwest::Body;
     use reqwest::Client;
-    use reqwest::Response;
     use tokio_util::sync::CancellationToken;
 
-    use paddler_local_http_fixture::fixture_response::FixtureResponse;
-    use paddler_local_http_fixture::local_http_fixture::LocalHttpFixture;
-    use paddler_messaging::api_path::ApiPath;
-
+    use super::checked_response;
     use super::send_checked_request;
     use crate::error::Error;
-    use crate::error::Result;
 
     const UNREACHABLE_HEALTH_URL: &str = "http://127.0.0.1:1/health";
 
-    async fn response_from(fixture_response: FixtureResponse) -> Result<Response> {
-        let fixture = LocalHttpFixture::start(fixture_response)
-            .await
-            .expect("the fixture server must start");
-        let url = fixture.url(ApiPath::HEALTH);
-        let request_builder = Client::new().get(&url);
-
-        send_checked_request(CancellationToken::new(), url, request_builder).await
-    }
-
     #[tokio::test]
     async fn a_refused_connection_maps_to_the_connect_variant() {
         let request_builder = Client::new().get(UNREACHABLE_HEALTH_URL);
@@ -103,22 +102,18 @@ mod tests {
 
     #[tokio::test]
     async fn a_rejection_whose_body_cannot_be_read_is_reported_as_undecodable() {
-        assert!(matches!(
-            response_from(FixtureResponse::TruncatedBody {
-                sent_body: b"inter".to_vec(),
-                status: StatusCode::INTERNAL_SERVER_ERROR,
-                withheld_byte_count: 5,
-            })
-            .await,
-            Err(Error::Http(source)) if source.is_decode()
-        ));
-    }
+        let chunks: [IoResult<Vec<u8>>; 2] = [
+            Ok(b"inter".to_vec()),
+            Err(IoError::from(ErrorKind::UnexpectedEof)),
+        ];
+        let rejection = HttpResponse::builder()
+            .status(StatusCode::INTERNAL_SERVER_ERROR)
+            .body(Body::wrap_stream(iter(chunks)))
+            .expect("the rejection must be buildable");
 
-    #[tokio::test]
-    async fn a_connection_closed_before_the_response_is_reported_as_a_failed_request() {
         assert!(matches!(
-            response_from(FixtureResponse::CloseBeforeHeaders).await,
-            Err(Error::Http(source)) if source.is_request()
+            checked_response(rejection.into(), UNREACHABLE_HEALTH_URL.to_owned()).await,
+            Err(Error::Http(source)) if source.is_decode()
         ));
     }
 }
diff --git a/paddler_client_javascript/README.md b/paddler_client_javascript/README.md
index b1b01c9b3..f600c58c7 100644
--- a/paddler_client_javascript/README.md
+++ b/paddler_client_javascript/README.md
@@ -24,8 +24,8 @@ const { continueConversation } = inferenceSocketClient({ webSocket });
 continueConversation({
   enableThinking: true,
   messages: [
-    { role: "system", content: "You are a helpful assistant." },
-    { role: "user", content: "Hello" },
+    { Message: { role: "system", content: "You are a helpful assistant." } },
+    { Message: { role: "user", content: "Hello" } },
   ],
 }).subscribe((chunk) => {
   if (chunk.error) {
diff --git a/paddler_client_javascript/package.json b/paddler_client_javascript/package.json
index 8e90efead..d1d85b909 100644
--- a/paddler_client_javascript/package.json
+++ b/paddler_client_javascript/package.json
@@ -1,6 +1,6 @@
 {
   "name": "@intentee/paddler-client",
-  "version": "4.0.1",
+  "version": "5.0.0",
   "description": "JavaScript/TypeScript client for the Paddler LLM load balancer",
   "license": "Apache-2.0",
   "homepage": "https://github.com/intentee/paddler",
@@ -20,7 +20,7 @@
   "files": ["dist", "README.md", "LICENSE"],
   "scripts": {
     "build": "tsc -p .",
-    "test": "tsc -p tsconfig.test.json && node --experimental-eventsource --enable-source-maps --import tsx --test 'tests/unit/**/*.test.ts' 'tests/cluster/**/*.test.ts'",
+    "test": "tsc -p tsconfig.test.json && node --experimental-eventsource --enable-source-maps --import tsx --test --test-concurrency=1 'tests/unit/**/*.test.ts' 'tests/cluster/**/*.test.ts'",
     "test:coverage": "tsc -p tsconfig.test.json && node --experimental-eventsource --enable-source-maps --import tsx --test --test-concurrency=1 --experimental-test-coverage --test-coverage-exclude='**/node_modules/**' --test-coverage-branches=100 --test-coverage-functions=100 --test-coverage-lines=100 'tests/**/*.test.ts'",
     "test:llm": "tsc -p tsconfig.test.json && node --experimental-eventsource --enable-source-maps --import tsx --test --test-concurrency=1 'tests/llm/**/*.test.ts'",
     "prepack": "tsc -p ."
@@ -30,12 +30,12 @@
     "zod": "^4"
   },
   "dependencies": {
-    "nanoid": "5.1.16"
+    "nanoid": "6.0.2"
   },
   "devDependencies": {
-    "@types/node": "22.19.17",
+    "@types/node": "26.6.5",
     "tsx": "4.23.15",
-    "typescript": "5.9.2"
+    "typescript": "6.0.3"
   },
   "engines": {
     "node": ">=22"
diff --git a/paddler_client_javascript/src/InvalidHuggingFaceUrlError.ts b/paddler_client_javascript/src/InvalidHuggingFaceUrlError.ts
deleted file mode 100644
index 6ea5e669e..000000000
--- a/paddler_client_javascript/src/InvalidHuggingFaceUrlError.ts
+++ /dev/null
@@ -1,9 +0,0 @@
-import { PaddlerError } from "./PaddlerError";
-
-export class InvalidHuggingFaceUrlError extends PaddlerError {
-  override name = "InvalidHuggingFaceUrlError";
-
-  constructor(public readonly pathname: string) {
-    super(`Invalid Hugging Face URL format: ${pathname}`);
-  }
-}
diff --git a/paddler_client_javascript/src/UnsupportedModelUrlError.ts b/paddler_client_javascript/src/UnsupportedModelUrlError.ts
deleted file mode 100644
index 4bbd4a726..000000000
--- a/paddler_client_javascript/src/UnsupportedModelUrlError.ts
+++ /dev/null
@@ -1,9 +0,0 @@
-import { PaddlerError } from "./PaddlerError";
-
-export class UnsupportedModelUrlError extends PaddlerError {
-  override name = "UnsupportedModelUrlError";
-
-  constructor(public readonly url: string) {
-    super("Unsupported URL format");
-  }
-}
diff --git a/paddler_client_javascript/src/eventSourceMessageState.ts b/paddler_client_javascript/src/eventSourceMessageState.ts
index 8beec143d..7b84403ca 100644
--- a/paddler_client_javascript/src/eventSourceMessageState.ts
+++ b/paddler_client_javascript/src/eventSourceMessageState.ts
@@ -13,8 +13,7 @@ export function eventSourceMessageState<TSchema extends z.ZodType>({
   data: string;
   schema: TSchema;
 }):
-  | EventSourceDataSnapshotState<TSchema>
-  | EventSourceDeserializationErrorState {
+  EventSourceDataSnapshotState<TSchema> | EventSourceDeserializationErrorState {
   let parsedJson: unknown;
 
   try {
diff --git a/paddler_client_javascript/src/extractHuggingFaceUrlParts.ts b/paddler_client_javascript/src/extractHuggingFaceUrlParts.ts
deleted file mode 100644
index f53958b59..000000000
--- a/paddler_client_javascript/src/extractHuggingFaceUrlParts.ts
+++ /dev/null
@@ -1,28 +0,0 @@
-import { InvalidHuggingFaceUrlError } from "./InvalidHuggingFaceUrlError";
-import type { HuggingFaceModelReference } from "./schemas/HuggingFaceModelReference";
-
-export function extractHuggingFaceUrlParts({
-  pathname,
-}: URL): HuggingFaceModelReference {
-  const [owner, repo, resourceKind, revision, ...filenameSegments] = pathname
-    .split("/")
-    .filter(function (segment) {
-      return segment.length > 0;
-    });
-
-  if (
-    owner === undefined ||
-    repo === undefined ||
-    revision === undefined ||
-    (resourceKind !== "blob" && resourceKind !== "resolve") ||
-    filenameSegments.length === 0
-  ) {
-    throw new InvalidHuggingFaceUrlError(pathname);
-  }
-
-  return {
-    filename: filenameSegments.join("/"),
-    repo_id: `${owner}/${repo}`,
-    revision,
-  };
-}
diff --git a/paddler_client_javascript/src/inferenceSocketClient.ts b/paddler_client_javascript/src/inferenceSocketClient.ts
index cbf24e4e6..a9a547c8f 100644
--- a/paddler_client_javascript/src/inferenceSocketClient.ts
+++ b/paddler_client_javascript/src/inferenceSocketClient.ts
@@ -2,31 +2,18 @@ import { nanoid } from "nanoid";
 import { filter, fromEvent, map, takeWhile, type Observable } from "rxjs";
 
 import type { ConversationMessage } from "./schemas/ConversationMessage";
-import {
-  InferenceNotificationSchema,
-  type InferenceNotification,
-} from "./schemas/InferenceNotification";
 import {
   InferenceServiceGenerateTokensResponseSchema,
   type InferenceServiceGenerateTokensResponse,
 } from "./schemas/InferenceServiceGenerateTokensResponse";
 
 export interface InferenceSocketClient {
-  clusterTokenGenerationMode$: Observable<InferenceNotification>;
   continueConversation(params: {
     enableThinking: boolean;
     messages: ConversationMessage[];
   }): Observable<InferenceServiceGenerateTokensResponse>;
 }
 
-function isNotificationFrame(parsedFrame: unknown): boolean {
-  return (
-    "object" === typeof parsedFrame &&
-    null !== parsedFrame &&
-    "Notification" in parsedFrame
-  );
-}
-
 export function inferenceSocketClient({
   webSocket,
 }: {
@@ -47,14 +34,6 @@ export function inferenceSocketClient({
     }),
   );
 
-  const clusterTokenGenerationMode$: Observable<InferenceNotification> =
-    parsedFrames$.pipe(
-      filter(isNotificationFrame),
-      map(function (parsedFrame: unknown): InferenceNotification {
-        return InferenceNotificationSchema.parse(parsedFrame).Notification;
-      }),
-    );
-
   function continueConversation({
     enableThinking,
     messages,
@@ -64,9 +43,6 @@ export function inferenceSocketClient({
   }): Observable<InferenceServiceGenerateTokensResponse> {
     const requestId = nanoid();
     const tokenStream = parsedFrames$.pipe(
-      filter(function (parsedFrame) {
-        return !isNotificationFrame(parsedFrame);
-      }),
       map(function (parsedFrame: unknown) {
         return InferenceServiceGenerateTokensResponseSchema.parse(parsedFrame);
       }),
@@ -98,7 +74,6 @@ export function inferenceSocketClient({
   }
 
   return Object.freeze({
-    clusterTokenGenerationMode$,
     continueConversation,
   });
 }
diff --git a/paddler_client_javascript/src/schemas/AgentDecisionsServing.ts b/paddler_client_javascript/src/schemas/AgentDecisionsServing.ts
new file mode 100644
index 000000000..43dbb8ae5
--- /dev/null
+++ b/paddler_client_javascript/src/schemas/AgentDecisionsServing.ts
@@ -0,0 +1,5 @@
+import { z } from "zod";
+
+export const AgentDecisionsServingSchema = z.enum(["NotServed", "Served"]);
+
+export type AgentDecisionsServing = z.infer<typeof AgentDecisionsServingSchema>;
diff --git a/paddler_client_javascript/src/schemas/AgentDesiredModel.ts b/paddler_client_javascript/src/schemas/AgentDesiredModel.ts
index b330b4cf6..7c1bdc779 100644
--- a/paddler_client_javascript/src/schemas/AgentDesiredModel.ts
+++ b/paddler_client_javascript/src/schemas/AgentDesiredModel.ts
@@ -1,17 +1,8 @@
 import { z } from "zod";
 
-import { HuggingFaceModelReferenceSchema } from "./HuggingFaceModelReference";
-import { UrlModelReferenceSchema } from "./UrlModelReference";
-
 export const AgentDesiredModelSchema = z.union([
   z.object({
-    HuggingFace: HuggingFaceModelReferenceSchema,
-  }),
-  z.object({
-    LocalToAgent: z.string(),
-  }),
-  z.object({
-    Url: UrlModelReferenceSchema,
+    Uri: z.string(),
   }),
   z.literal("None"),
 ]);
diff --git a/paddler_client_javascript/src/schemas/AgentIssue.ts b/paddler_client_javascript/src/schemas/AgentIssue.ts
index b12472730..4ec23800c 100644
--- a/paddler_client_javascript/src/schemas/AgentIssue.ts
+++ b/paddler_client_javascript/src/schemas/AgentIssue.ts
@@ -2,8 +2,18 @@ import { z } from "zod";
 
 import { AgentIssueModelPathSchema } from "./AgentIssueModelPath";
 import { HuggingFaceDownloadLockSchema } from "./HuggingFaceDownloadLock";
+import { InferenceCapabilitySchema } from "./InferenceCapability";
 
 export const AgentIssueSchema = z.union([
+  z.object({
+    AgentModelUriIsNotAnAbsoluteLocalPath: AgentIssueModelPathSchema,
+  }),
+  z.object({
+    AgentModelUriIsNotCanonical: z.object({
+      canonical_uri: z.string(),
+      model_path: AgentIssueModelPathSchema,
+    }),
+  }),
   z.object({
     CacheCannotAcquireLock: AgentIssueModelPathSchema,
   }),
@@ -41,12 +51,33 @@ export const AgentIssueSchema = z.union([
   z.object({
     HuggingFaceCannotAcquireLock: HuggingFaceDownloadLockSchema,
   }),
+  z.object({
+    HuggingFaceDownloadFailed: z.object({
+      failure: z.union([
+        z.literal("CacheIoFailed"),
+        z.literal("HubConnectionFailed"),
+        z.object({
+          HubRespondedWithStatus: z.object({ status: z.number() }),
+        }),
+        z.literal("HubResponseMalformed"),
+        z.literal("UnexpectedHubClientFailure"),
+        z.literal("XetTransferFailed"),
+      ]),
+      model_path: AgentIssueModelPathSchema,
+    }),
+  }),
   z.object({
     HuggingFaceModelDoesNotExist: AgentIssueModelPathSchema,
   }),
+  z.object({
+    HuggingFaceModelUriIsMalformed: AgentIssueModelPathSchema,
+  }),
   z.object({
     HuggingFacePermissions: AgentIssueModelPathSchema,
   }),
+  z.object({
+    LocalModelFileCannotBeChecked: AgentIssueModelPathSchema,
+  }),
   z.object({
     ModelCacheIsCorrupted: AgentIssueModelPathSchema,
   }),
@@ -59,15 +90,44 @@ export const AgentIssueSchema = z.union([
   z.object({
     ModelFileDoesNotExist: AgentIssueModelPathSchema,
   }),
+  z.object({
+    ModelUriIsUnparseable: AgentIssueModelPathSchema,
+  }),
   z.object({
     MultimodalProjectionCannotBeLoaded: AgentIssueModelPathSchema,
   }),
+  z.object({
+    PointerHeadCannotBeLoaded: AgentIssueModelPathSchema,
+  }),
+  z.object({
+    PointerHeadIncompatibleWithModel: z.object({
+      incompatibility: z.union([
+        z.object({
+          DelimiterIsNotAControlToken: z.object({ delimiter: z.string() }),
+        }),
+        z.object({
+          HiddenSizeMismatch: z.object({
+            model_hidden_size: z.number(),
+            pointer_head_hidden_size: z.number(),
+          }),
+        }),
+      ]),
+      pointer_head_path: AgentIssueModelPathSchema,
+    }),
+  }),
   z.object({
     SlotCannotStart: z.object({
       error: z.string(),
+      inference_capability: InferenceCapabilitySchema,
       slot_index: z.number(),
     }),
   }),
+  z.object({
+    SlotsInsufficientForDecisions: z.object({
+      desired_slots: z.number(),
+      required_slots: z.number(),
+    }),
+  }),
   z.object({
     UnableToFindChatTemplate: AgentIssueModelPathSchema,
   }),
diff --git a/paddler_client_javascript/src/schemas/AgentRuntimeStatus.ts b/paddler_client_javascript/src/schemas/AgentRuntimeStatus.ts
new file mode 100644
index 000000000..50534c184
--- /dev/null
+++ b/paddler_client_javascript/src/schemas/AgentRuntimeStatus.ts
@@ -0,0 +1,19 @@
+import { z } from "zod";
+
+import { AgentDecisionsServingSchema } from "./AgentDecisionsServing";
+
+export const AgentRuntimeStatusSchema = z.union([
+  z.literal("Idle"),
+  z
+    .object({
+      Serving: z
+        .object({
+          decisions: AgentDecisionsServingSchema,
+          slots_total: z.number(),
+        })
+        .strict(),
+    })
+    .strict(),
+]);
+
+export type AgentRuntimeStatus = z.infer<typeof AgentRuntimeStatusSchema>;
diff --git a/paddler_client_javascript/src/schemas/AgentStatus.ts b/paddler_client_javascript/src/schemas/AgentStatus.ts
index 3325a2a31..7df84a53c 100644
--- a/paddler_client_javascript/src/schemas/AgentStatus.ts
+++ b/paddler_client_javascript/src/schemas/AgentStatus.ts
@@ -1,6 +1,7 @@
 import { z } from "zod";
 
 import { AgentIssueSchema } from "./AgentIssue";
+import { AgentRuntimeStatusSchema } from "./AgentRuntimeStatus";
 import { ModelDownloadStatusSchema } from "./ModelDownloadStatus";
 
 export const AgentStatusSchema = z
@@ -9,10 +10,9 @@ export const AgentStatusSchema = z
     download_status: ModelDownloadStatusSchema,
     issues: z.array(AgentIssueSchema),
     model_path: z.string().nullable(),
-    slots_total: z.number(),
+    runtime: AgentRuntimeStatusSchema,
     state_application_status: z.enum([
       "Applied",
-      "AttemptedAndNotAppliable",
       "AttemptedAndRetrying",
       "Fresh",
       "Stuck",
diff --git a/paddler_client_javascript/src/schemas/BalancerDesiredState.ts b/paddler_client_javascript/src/schemas/BalancerDesiredState.ts
index 7134d08f9..5c36f7386 100644
--- a/paddler_client_javascript/src/schemas/BalancerDesiredState.ts
+++ b/paddler_client_javascript/src/schemas/BalancerDesiredState.ts
@@ -1,16 +1,18 @@
 import { z } from "zod";
 
 import { AgentDesiredModelSchema } from "./AgentDesiredModel";
-import { ChatTemplateSchema } from "./ChatTemplate";
-import { InferenceParametersSchema } from "./InferenceParameters";
+import { BalancerTextGenerationSettingsSchema } from "./BalancerTextGenerationSettings";
+import { DecisionSettingsSchema } from "./DecisionSettings";
+import { EmbeddingParametersSchema } from "./EmbeddingParameters";
+import { ModelRuntimeParametersSchema } from "./ModelRuntimeParameters";
 
 export const BalancerDesiredStateSchema = z
   .object({
-    chat_template_override: ChatTemplateSchema.nullable(),
-    inference_parameters: InferenceParametersSchema,
+    decision: DecisionSettingsSchema,
+    embeddings: EmbeddingParametersSchema,
     model: AgentDesiredModelSchema,
-    multimodal_projection: AgentDesiredModelSchema,
-    use_chat_template_override: z.boolean(),
+    model_runtime_parameters: ModelRuntimeParametersSchema,
+    text_generation: BalancerTextGenerationSettingsSchema,
   })
   .strict();
 
diff --git a/paddler_client_javascript/src/schemas/BalancerTextGenerationSettings.ts b/paddler_client_javascript/src/schemas/BalancerTextGenerationSettings.ts
new file mode 100644
index 000000000..d101fb283
--- /dev/null
+++ b/paddler_client_javascript/src/schemas/BalancerTextGenerationSettings.ts
@@ -0,0 +1,17 @@
+import { z } from "zod";
+
+import { ChatTemplateSourceSchema } from "./ChatTemplateSource";
+import { MultimodalSettingsSchema } from "./MultimodalSettings";
+import { SamplingParametersSchema } from "./SamplingParameters";
+
+export const BalancerTextGenerationSettingsSchema = z
+  .object({
+    chat_template_source: ChatTemplateSourceSchema,
+    multimodal: MultimodalSettingsSchema,
+    sampling_parameters: SamplingParametersSchema,
+  })
+  .strict();
+
+export type BalancerTextGenerationSettings = z.infer<
+  typeof BalancerTextGenerationSettingsSchema
+>;
diff --git a/paddler_client_javascript/src/schemas/ChatTemplateSource.ts b/paddler_client_javascript/src/schemas/ChatTemplateSource.ts
new file mode 100644
index 000000000..ee18383ed
--- /dev/null
+++ b/paddler_client_javascript/src/schemas/ChatTemplateSource.ts
@@ -0,0 +1,14 @@
+import { z } from "zod";
+
+import { ChatTemplateSchema } from "./ChatTemplate";
+
+export const ChatTemplateSourceSchema = z.union([
+  z
+    .object({
+      Override: ChatTemplateSchema,
+    })
+    .strict(),
+  z.literal("EmbeddedInModel"),
+]);
+
+export type ChatTemplateSource = z.infer<typeof ChatTemplateSourceSchema>;
diff --git a/paddler_client_javascript/src/schemas/ConversationMessage.ts b/paddler_client_javascript/src/schemas/ConversationMessage.ts
index 15948f7d3..30e52906b 100644
--- a/paddler_client_javascript/src/schemas/ConversationMessage.ts
+++ b/paddler_client_javascript/src/schemas/ConversationMessage.ts
@@ -1,10 +1,37 @@
 import { z } from "zod";
 
 import { ConversationMessageContentPartSchema } from "./ConversationMessageContentPart";
+import { ConversationToolCallSchema } from "./ConversationToolCall";
 
-export const ConversationMessageSchema = z.object({
-  role: z.string(),
-  content: z.union([z.string(), z.array(ConversationMessageContentPartSchema)]),
-});
+const ConversationMessageContentSchema = z.union([
+  z.string(),
+  z.array(ConversationMessageContentPartSchema),
+]);
+
+export const ConversationMessageSchema = z.union([
+  z.strictObject({
+    AssistantReasoning: z.strictObject({
+      reasoning_content: z.string(),
+    }),
+  }),
+  z.strictObject({
+    AssistantToolCalls: z.strictObject({
+      content: ConversationMessageContentSchema,
+      tool_calls: z.array(ConversationToolCallSchema),
+    }),
+  }),
+  z.strictObject({
+    Message: z.strictObject({
+      content: ConversationMessageContentSchema,
+      role: z.string(),
+    }),
+  }),
+  z.strictObject({
+    ToolResult: z.strictObject({
+      content: ConversationMessageContentSchema,
+      tool_call_id: z.string(),
+    }),
+  }),
+]);
 
 export type ConversationMessage = z.infer<typeof ConversationMessageSchema>;
diff --git a/paddler_client_javascript/src/schemas/ConversationToolCall.ts b/paddler_client_javascript/src/schemas/ConversationToolCall.ts
new file mode 100644
index 000000000..5f916249c
--- /dev/null
+++ b/paddler_client_javascript/src/schemas/ConversationToolCall.ts
@@ -0,0 +1,9 @@
+import { z } from "zod";
+
+export const ConversationToolCallSchema = z.strictObject({
+  arguments: z.record(z.string(), z.unknown()),
+  id: z.string(),
+  name: z.string(),
+});
+
+export type ConversationToolCall = z.infer<typeof ConversationToolCallSchema>;
diff --git a/paddler_client_javascript/src/schemas/DecideParams.ts b/paddler_client_javascript/src/schemas/DecideParams.ts
new file mode 100644
index 000000000..bda6e111e
--- /dev/null
+++ b/paddler_client_javascript/src/schemas/DecideParams.ts
@@ -0,0 +1,12 @@
+import { z } from "zod";
+
+import { DecisionQuestionSchema } from "./DecisionQuestion";
+
+export const DecideParamsSchema = z
+  .object({
+    questions: z.array(DecisionQuestionSchema).min(1),
+    state: z.string(),
+  })
+  .strict();
+
+export type DecideParams = z.infer<typeof DecideParamsSchema>;
diff --git a/paddler_client_javascript/src/schemas/DecisionQuestion.ts b/paddler_client_javascript/src/schemas/DecisionQuestion.ts
new file mode 100644
index 000000000..34e3e553e
--- /dev/null
+++ b/paddler_client_javascript/src/schemas/DecisionQuestion.ts
@@ -0,0 +1,11 @@
+import { z } from "zod";
+
+export const DecisionQuestionSchema = z
+  .object({
+    id: z.string(),
+    instructions: z.string(),
+    options: z.array(z.string()).min(1),
+  })
+  .strict();
+
+export type DecisionQuestion = z.infer<typeof DecisionQuestionSchema>;
diff --git a/paddler_client_javascript/src/schemas/DecisionSettings.ts b/paddler_client_javascript/src/schemas/DecisionSettings.ts
new file mode 100644
index 000000000..4fa0676f8
--- /dev/null
+++ b/paddler_client_javascript/src/schemas/DecisionSettings.ts
@@ -0,0 +1,11 @@
+import { z } from "zod";
+
+import { AgentDesiredModelSchema } from "./AgentDesiredModel";
+
+export const DecisionSettingsSchema = z
+  .object({
+    pointer_head: AgentDesiredModelSchema,
+  })
+  .strict();
+
+export type DecisionSettings = z.infer<typeof DecisionSettingsSchema>;
diff --git a/paddler_client_javascript/src/schemas/EmbeddingParameters.ts b/paddler_client_javascript/src/schemas/EmbeddingParameters.ts
new file mode 100644
index 000000000..e34b04821
--- /dev/null
+++ b/paddler_client_javascript/src/schemas/EmbeddingParameters.ts
@@ -0,0 +1,18 @@
+import { z } from "zod";
+
+export const poolingTypes = [
+  "Cls",
+  "Last",
+  "Mean",
+  "Rank",
+  "Unspecified",
+] as const;
+
+export const EmbeddingParametersSchema = z
+  .object({
+    embedding_batch_size: z.number().int(),
+    pooling_type: z.enum(poolingTypes),
+  })
+  .strict();
+
+export type EmbeddingParameters = z.infer<typeof EmbeddingParametersSchema>;
diff --git a/paddler_client_javascript/src/schemas/GenerateEmbeddingBatchParams.ts b/paddler_client_javascript/src/schemas/GenerateEmbeddingBatchParams.ts
index 0d3194d4e..5c60afd52 100644
--- a/paddler_client_javascript/src/schemas/GenerateEmbeddingBatchParams.ts
+++ b/paddler_client_javascript/src/schemas/GenerateEmbeddingBatchParams.ts
@@ -4,7 +4,7 @@ import { EmbeddingInputDocumentSchema } from "./EmbeddingInputDocument";
 import { EmbeddingNormalizationMethodSchema } from "./EmbeddingNormalizationMethod";
 
 export const GenerateEmbeddingBatchParamsSchema = z.object({
-  input_documents: z.array(EmbeddingInputDocumentSchema),
+  input_batch: z.array(EmbeddingInputDocumentSchema),
   normalization_method: EmbeddingNormalizationMethodSchema,
 });
 
diff --git a/paddler_client_javascript/src/schemas/GeneratedToolCall.ts b/paddler_client_javascript/src/schemas/GeneratedToolCall.ts
new file mode 100644
index 000000000..42090ddd0
--- /dev/null
+++ b/paddler_client_javascript/src/schemas/GeneratedToolCall.ts
@@ -0,0 +1,16 @@
+import { z } from "zod";
+
+export const ToolCallArgumentsSchema = z.union([
+  z.strictObject({ InvalidJson: z.string() }),
+  z.strictObject({ ValidJson: z.unknown() }),
+]);
+
+export const GeneratedToolCallSchema = z.object({
+  arguments: ToolCallArgumentsSchema,
+  id: z.string(),
+  index: z.number(),
+  name: z.string(),
+});
+
+export type GeneratedToolCall = z.infer<typeof GeneratedToolCallSchema>;
+export type ToolCallArguments = z.infer<typeof ToolCallArgumentsSchema>;
diff --git a/paddler_client_javascript/src/schemas/HuggingFaceModelReference.ts b/paddler_client_javascript/src/schemas/HuggingFaceModelReference.ts
deleted file mode 100644
index cf60ec752..000000000
--- a/paddler_client_javascript/src/schemas/HuggingFaceModelReference.ts
+++ /dev/null
@@ -1,11 +0,0 @@
-import { z } from "zod";
-
-export const HuggingFaceModelReferenceSchema = z.object({
-  filename: z.string(),
-  repo_id: z.string(),
-  revision: z.string(),
-});
-
-export type HuggingFaceModelReference = z.infer<
-  typeof HuggingFaceModelReferenceSchema
->;
diff --git a/paddler_client_javascript/src/schemas/InferenceCapability.ts b/paddler_client_javascript/src/schemas/InferenceCapability.ts
new file mode 100644
index 000000000..b2898ff63
--- /dev/null
+++ b/paddler_client_javascript/src/schemas/InferenceCapability.ts
@@ -0,0 +1,9 @@
+import { z } from "zod";
+
+export const InferenceCapabilitySchema = z.enum([
+  "Decisions",
+  "Embeddings",
+  "TextGeneration",
+]);
+
+export type InferenceCapability = z.infer<typeof InferenceCapabilitySchema>;
diff --git a/paddler_client_javascript/src/schemas/InferenceNotification.ts b/paddler_client_javascript/src/schemas/InferenceNotification.ts
deleted file mode 100644
index ea7cb6e06..000000000
--- a/paddler_client_javascript/src/schemas/InferenceNotification.ts
+++ /dev/null
@@ -1,9 +0,0 @@
-import { z } from "zod";
-
-export const InferenceNotificationSchema = z.object({
-  Notification: z.enum(["TokenGenerationDisabled", "TokenGenerationEnabled"]),
-});
-
-export type InferenceNotification = z.infer<
-  typeof InferenceNotificationSchema
->["Notification"];
diff --git a/paddler_client_javascript/src/schemas/InferenceParameters.ts b/paddler_client_javascript/src/schemas/InferenceParameters.ts
deleted file mode 100644
index 0ac3e61ee..000000000
--- a/paddler_client_javascript/src/schemas/InferenceParameters.ts
+++ /dev/null
@@ -1,81 +0,0 @@
-import { z } from "zod";
-
-export const cacheDtypes = [
-  "F32",
-  "F16",
-  "BF16",
-  "Q8_0",
-  "Q4_0",
-  "Q4_1",
-  "IQ4_NL",
-  "Q5_0",
-  "Q5_1",
-] as const;
-
-export const poolingTypes = [
-  "Cls",
-  "Last",
-  "Mean",
-  "None",
-  "Rank",
-  "Unspecified",
-] as const;
-
-const I32_MAX = 2_147_483_647;
-const NEUTRAL_PENALTY_REPEAT = 1;
-
-const ProbabilitySchema = z.number().min(0).max(1);
-
-export const InferenceParametersSchema = z
-  .object({
-    n_batch: z.number().int().min(1).max(I32_MAX),
-    context_size: z.number().int().min(1),
-    embedding_batch_size: z.number().int().min(1),
-    enable_embeddings: z.boolean(),
-    image_resize_to_fit: z.number().int().min(1),
-    k_cache_dtype: z.enum(cacheDtypes),
-    v_cache_dtype: z.enum(cacheDtypes),
-    min_p: ProbabilitySchema,
-    n_gpu_layers: z.number().int().min(-1),
-    penalty_frequency: z.number(),
-    penalty_last_n: z.number().int().min(0),
-    penalty_presence: z.number(),
-    penalty_repeat: z.number().gt(0),
-    pooling_type: z.enum(poolingTypes),
-    temperature: z.number().min(0),
-    top_k: z.number().int().min(0),
-    top_p: ProbabilitySchema,
-  })
-  .strict()
-  .superRefine(function (parameters, context) {
-    const penaltiesAreNeutral =
-      parameters.penalty_repeat === NEUTRAL_PENALTY_REPEAT &&
-      parameters.penalty_frequency === 0 &&
-      parameters.penalty_presence === 0;
-
-    if (parameters.n_batch > parameters.context_size) {
-      context.addIssue({
-        code: "custom",
-        message: "n_batch must not exceed context_size",
-        path: ["n_batch"],
-      });
-    }
-
-    if (parameters.penalty_last_n === 0 && !penaltiesAreNeutral) {
-      context.addIssue({
-        code: "custom",
-        message: "penalty strengths require a penalty_last_n window",
-        path: ["penalty_last_n"],
-      });
-    }
-
-    if (parameters.penalty_last_n > 0 && penaltiesAreNeutral) {
-      context.addIssue({
-        code: "custom",
-        message: "penalty_last_n requires at least one penalty strength",
-        path: ["penalty_last_n"],
-      });
-    }
-  });
-
-export type InferenceParameters = z.infer<typeof InferenceParametersSchema>;
diff --git a/paddler_client_javascript/src/schemas/InferenceServiceDecideResponse.ts b/paddler_client_javascript/src/schemas/InferenceServiceDecideResponse.ts
new file mode 100644
index 000000000..16693010f
--- /dev/null
+++ b/paddler_client_javascript/src/schemas/InferenceServiceDecideResponse.ts
@@ -0,0 +1,177 @@
+import { z } from "zod";
+
+const DecisionAnswerSchema = z.object({
+  id: z.string(),
+  probabilities: z.array(z.number()),
+});
+
+const DecisionSummarySchema = z.object({
+  input_tokens: z.number(),
+  processing_milliseconds: z.number(),
+});
+
+const OversizedDecisionDetailsSchema = z.object({
+  context_size: z.number(),
+  required_tokens: z.number(),
+});
+
+const decisionFailureCodes = {
+  AgentRuntimeFailed: 500,
+  BatchAssemblyFailed: 500,
+  DecisionsNotServed: 503,
+  DecodeFailed: 500,
+  HiddenStateUnavailable: 500,
+  InputTokenizationFailed: 400,
+  KvCacheCopyFailed: 500,
+  KvCacheRemovalFailed: 500,
+  ModelNotLoaded: 503,
+  SchedulerUnavailable: 503,
+} as const;
+
+type DecisionFailure = keyof typeof decisionFailureCodes;
+
+const DecisionFailureSchema = z.union([
+  z.object({ AgentRuntimeFailed: z.string() }).strict(),
+  z.object({ BatchAssemblyFailed: z.string() }).strict(),
+  z.object({ DecisionsNotServed: z.string() }).strict(),
+  z.object({ DecodeFailed: z.string() }).strict(),
+  z.object({ HiddenStateUnavailable: z.string() }).strict(),
+  z.object({ InputTokenizationFailed: z.string() }).strict(),
+  z.object({ KvCacheCopyFailed: z.string() }).strict(),
+  z.object({ KvCacheRemovalFailed: z.string() }).strict(),
+  z.object({ ModelNotLoaded: z.string() }).strict(),
+  z.object({ SchedulerUnavailable: z.string() }).strict(),
+]);
+
+const DecisionResultSchema = z.union([
+  z.object({ Done: DecisionSummarySchema }).strict(),
+  z.object({ QuestionAnswered: DecisionAnswerSchema }).strict(),
+  z.object({ RequestExceedsContext: OversizedDecisionDetailsSchema }).strict(),
+  DecisionFailureSchema,
+]);
+
+type DecisionMessage =
+  | {
+      answer: z.infer<typeof DecisionAnswerSchema>;
+      done: false;
+      error: null;
+      generated_by: string | null;
+      request_id: string;
+      summary: null;
+    }
+  | {
+      answer: null;
+      done: true;
+      error: null;
+      generated_by: string | null;
+      request_id: string;
+      summary: z.infer<typeof DecisionSummarySchema>;
+    }
+  | {
+      answer: null;
+      done: true;
+      error: { code: number; description: string };
+      generated_by: string | null;
+      request_id: string;
+      summary: null;
+    };
+
+function failedDecision(
+  request_id: string,
+  generated_by: string | null,
+  code: number,
+  description: string,
+): DecisionMessage {
+  return Object.freeze({
+    answer: null,
+    done: true,
+    error: Object.freeze({ code, description }),
+    generated_by,
+    request_id,
+    summary: null,
+  });
+}
+
+const DecisionStreamMessageSchema = z.union([
+  z.object({
+    Error: z.object({
+      error: z.object({ code: z.number(), description: z.string() }),
+      request_id: z.string(),
+    }),
+  }),
+  z.object({
+    Response: z.object({
+      generated_by: z.string().nullable(),
+      request_id: z.string(),
+      response: z.object({ Decision: DecisionResultSchema }),
+    }),
+  }),
+]);
+
+function decisionMessage(
+  message: z.infer<typeof DecisionStreamMessageSchema>,
+): DecisionMessage {
+  if ("Error" in message) {
+    return failedDecision(
+      message.Error.request_id,
+      null,
+      message.Error.error.code,
+      message.Error.error.description,
+    );
+  }
+
+  const { generated_by, request_id, response } = message.Response;
+  const decisionResult = response.Decision;
+
+  if ("Done" in decisionResult) {
+    return Object.freeze({
+      answer: null,
+      done: true,
+      error: null,
+      generated_by,
+      request_id,
+      summary: decisionResult.Done,
+    });
+  }
+
+  if ("QuestionAnswered" in decisionResult) {
+    return Object.freeze({
+      answer: decisionResult.QuestionAnswered,
+      done: false,
+      error: null,
+      generated_by,
+      request_id,
+      summary: null,
+    });
+  }
+
+  if ("RequestExceedsContext" in decisionResult) {
+    const { context_size, required_tokens } =
+      decisionResult.RequestExceedsContext;
+
+    return failedDecision(
+      request_id,
+      generated_by,
+      400,
+      `the decision needs ${required_tokens} tokens in the cache, more than the ${context_size} the context holds`,
+    );
+  }
+
+  const [[failure, description]] = Object.entries(decisionResult) as [
+    [DecisionFailure, string],
+  ];
+
+  return failedDecision(
+    request_id,
+    generated_by,
+    decisionFailureCodes[failure],
+    description,
+  );
+}
+
+export const InferenceServiceDecideResponseSchema =
+  DecisionStreamMessageSchema.transform(decisionMessage);
+
+export type InferenceServiceDecideResponse = z.infer<
+  typeof InferenceServiceDecideResponseSchema
+>;
diff --git a/paddler_client_javascript/src/schemas/InferenceServiceGenerateTokensResponse.ts b/paddler_client_javascript/src/schemas/InferenceServiceGenerateTokensResponse.ts
index edf6fd3e2..8169c4828 100644
--- a/paddler_client_javascript/src/schemas/InferenceServiceGenerateTokensResponse.ts
+++ b/paddler_client_javascript/src/schemas/InferenceServiceGenerateTokensResponse.ts
@@ -1,12 +1,9 @@
 import { z } from "zod";
 
-import { ParsedToolCallSchema } from "./ParsedToolCall";
+import { GeneratedToolCallSchema } from "./GeneratedToolCall";
 
 export type GeneratedTokenKind =
-  | "content"
-  | "reasoning"
-  | "tool_call"
-  | "undeterminable";
+  "content" | "reasoning" | "tool_call" | "undeterminable";
 
 const TokenUsageSchema = z.object({
   prompt_tokens: z.number(),
@@ -69,14 +66,12 @@ const GeneratedTokenResultSchema = z.union([
   z.object({ MultimodalTokenizationFailed: z.string() }),
   z.object({ NoSequenceSlotAvailable: z.string() }),
   z.object({ PromptTokenizationFailed: z.string() }),
+  z.object({ PromptTokenizedToNothing: z.string() }),
   z.object({ SamplerChainCreationFailed: z.string() }),
   z.object({ SamplingCandidatesExhausted: z.string() }),
   z.object({ SchedulerUnavailable: z.string() }),
-  z.object({ SequenceIdOutOfRange: z.string() }),
-  z.object({ ToolsSerializationFailed: z.string() }),
   z.object({ SamplerError: z.string() }),
-  z.object({ TokenGenerationDisabled: z.string() }),
-  z.object({ ToolCallParsed: z.array(ParsedToolCallSchema) }),
+  z.object({ ToolCallParsed: z.array(GeneratedToolCallSchema) }),
   z.object({ ToolCallParseFailed: z.string() }),
   z.object({ ToolCallValidationFailed: z.array(z.string()) }),
   z.object({ ToolSchemaInvalid: z.string() }),
@@ -118,7 +113,7 @@ type Normalised =
       summary: null;
       token: null;
       tokenKind: null;
-      toolCalls: ReadonlyArray<z.infer<typeof ParsedToolCallSchema>>;
+      toolCalls: ReadonlyArray<z.infer<typeof GeneratedToolCallSchema>>;
     }
   | {
       done: false;
@@ -435,6 +430,15 @@ export const InferenceServiceGenerateTokensResponseSchema = z
       );
     }
 
+    if ("PromptTokenizedToNothing" in variant) {
+      return terminalError(
+        request_id,
+        generated_by,
+        400,
+        variant.PromptTokenizedToNothing,
+      );
+    }
+
     if ("SamplerChainCreationFailed" in variant) {
       return terminalError(
         request_id,
@@ -462,24 +466,6 @@ export const InferenceServiceGenerateTokensResponseSchema = z
       );
     }
 
-    if ("SequenceIdOutOfRange" in variant) {
-      return terminalError(
-        request_id,
-        generated_by,
-        500,
-        variant.SequenceIdOutOfRange,
-      );
-    }
-
-    if ("ToolsSerializationFailed" in variant) {
-      return terminalError(
-        request_id,
-        generated_by,
-        500,
-        variant.ToolsSerializationFailed,
-      );
-    }
-
     if ("DecodeFailed" in variant) {
       return terminalError(request_id, generated_by, 500, variant.DecodeFailed);
     }
@@ -585,15 +571,6 @@ export const InferenceServiceGenerateTokensResponseSchema = z
       );
     }
 
-    if ("TokenGenerationDisabled" in variant) {
-      return terminalError(
-        request_id,
-        generated_by,
-        501,
-        variant.TokenGenerationDisabled,
-      );
-    }
-
     return terminalError(request_id, generated_by, 500, variant.SamplerError);
   });
 
diff --git a/paddler_client_javascript/src/schemas/ModelRuntimeParameters.ts b/paddler_client_javascript/src/schemas/ModelRuntimeParameters.ts
new file mode 100644
index 000000000..fd58143f4
--- /dev/null
+++ b/paddler_client_javascript/src/schemas/ModelRuntimeParameters.ts
@@ -0,0 +1,27 @@
+import { z } from "zod";
+
+export const cacheDtypes = [
+  "F32",
+  "F16",
+  "BF16",
+  "Q8_0",
+  "Q4_0",
+  "Q4_1",
+  "IQ4_NL",
+  "Q5_0",
+  "Q5_1",
+] as const;
+
+export const ModelRuntimeParametersSchema = z
+  .object({
+    context_size: z.number().int(),
+    k_cache_dtype: z.enum(cacheDtypes),
+    n_batch: z.number().int(),
+    n_gpu_layers: z.number().int(),
+    v_cache_dtype: z.enum(cacheDtypes),
+  })
+  .strict();
+
+export type ModelRuntimeParameters = z.infer<
+  typeof ModelRuntimeParametersSchema
+>;
diff --git a/paddler_client_javascript/src/schemas/MultimodalSettings.ts b/paddler_client_javascript/src/schemas/MultimodalSettings.ts
new file mode 100644
index 000000000..a7736ab63
--- /dev/null
+++ b/paddler_client_javascript/src/schemas/MultimodalSettings.ts
@@ -0,0 +1,12 @@
+import { z } from "zod";
+
+import { AgentDesiredModelSchema } from "./AgentDesiredModel";
+
+export const MultimodalSettingsSchema = z
+  .object({
+    image_resize_to_fit: z.number().int(),
+    projection: AgentDesiredModelSchema,
+  })
+  .strict();
+
+export type MultimodalSettings = z.infer<typeof MultimodalSettingsSchema>;
diff --git a/paddler_client_javascript/src/schemas/ParsedToolCall.ts b/paddler_client_javascript/src/schemas/ParsedToolCall.ts
deleted file mode 100644
index 073ece927..000000000
--- a/paddler_client_javascript/src/schemas/ParsedToolCall.ts
+++ /dev/null
@@ -1,15 +0,0 @@
-import { z } from "zod";
-
-export const ToolCallArgumentsSchema = z.union([
-  z.strictObject({ InvalidJson: z.string() }),
-  z.strictObject({ ValidJson: z.unknown() }),
-]);
-
-export const ParsedToolCallSchema = z.object({
-  id: z.string(),
-  name: z.string(),
-  arguments: ToolCallArgumentsSchema,
-});
-
-export type ParsedToolCall = z.infer<typeof ParsedToolCallSchema>;
-export type ToolCallArguments = z.infer<typeof ToolCallArgumentsSchema>;
diff --git a/paddler_client_javascript/src/schemas/SamplingParameters.ts b/paddler_client_javascript/src/schemas/SamplingParameters.ts
new file mode 100644
index 000000000..4dd09450a
--- /dev/null
+++ b/paddler_client_javascript/src/schemas/SamplingParameters.ts
@@ -0,0 +1,16 @@
+import { z } from "zod";
+
+export const SamplingParametersSchema = z
+  .object({
+    min_p: z.number(),
+    penalty_frequency: z.number(),
+    penalty_last_n: z.number().int(),
+    penalty_presence: z.number(),
+    penalty_repeat: z.number(),
+    temperature: z.number(),
+    top_k: z.number().int(),
+    top_p: z.number(),
+  })
+  .strict();
+
+export type SamplingParameters = z.infer<typeof SamplingParametersSchema>;
diff --git a/paddler_client_javascript/src/schemas/UrlModelReference.ts b/paddler_client_javascript/src/schemas/UrlModelReference.ts
deleted file mode 100644
index dce78a9a5..000000000
--- a/paddler_client_javascript/src/schemas/UrlModelReference.ts
+++ /dev/null
@@ -1,7 +0,0 @@
-import { z } from "zod";
-
-export const UrlModelReferenceSchema = z.object({
-  url: z.string(),
-});
-
-export type UrlModelReference = z.infer<typeof UrlModelReferenceSchema>;
diff --git a/paddler_client_javascript/src/urlToAgentDesiredModel.ts b/paddler_client_javascript/src/urlToAgentDesiredModel.ts
deleted file mode 100644
index a41b6cead..000000000
--- a/paddler_client_javascript/src/urlToAgentDesiredModel.ts
+++ /dev/null
@@ -1,25 +0,0 @@
-import { extractHuggingFaceUrlParts } from "./extractHuggingFaceUrlParts";
-import type { AgentDesiredModel } from "./schemas/AgentDesiredModel";
-import { UnsupportedModelUrlError } from "./UnsupportedModelUrlError";
-
-export function urlToAgentDesiredModel(url: URL): AgentDesiredModel {
-  if (url.hostname === "huggingface.co") {
-    return {
-      HuggingFace: extractHuggingFaceUrlParts(url),
-    };
-  }
-
-  if (url.protocol === "agent:") {
-    return {
-      LocalToAgent: url.pathname,
-    };
-  }
-
-  if (url.protocol === "http:" || url.protocol === "https:") {
-    return {
-      Url: { url: url.toString() },
-    };
-  }
-
-  throw new UnsupportedModelUrlError(url.toString());
-}
diff --git a/paddler_client_javascript/tests/AgentExitedBeforeReadyError.ts b/paddler_client_javascript/tests/AgentExitedBeforeReadyError.ts
new file mode 100644
index 000000000..bf57540fe
--- /dev/null
+++ b/paddler_client_javascript/tests/AgentExitedBeforeReadyError.ts
@@ -0,0 +1,12 @@
+export class AgentExitedBeforeReadyError extends Error {
+  override name = "AgentExitedBeforeReadyError";
+
+  constructor(
+    public readonly agentName: string,
+    public readonly exitCode: number | null,
+  ) {
+    super(
+      `Agent ${agentName} exited with code ${String(exitCode)} before it was ready`,
+    );
+  }
+}
diff --git a/paddler_client_javascript/tests/BalancerAddresses.ts b/paddler_client_javascript/tests/BalancerAddresses.ts
index 3d366393d..11a557a28 100644
--- a/paddler_client_javascript/tests/BalancerAddresses.ts
+++ b/paddler_client_javascript/tests/BalancerAddresses.ts
@@ -3,6 +3,7 @@ import { z } from "zod";
 export const BalancerAddressesSchema = z
   .object({
     compat_openai: z.string().nullable(),
+    compat_typesafe: z.string().nullable(),
     inference: z.string(),
     management: z.string(),
     web_admin_panel: z.string().nullable(),
diff --git a/paddler_client_javascript/tests/SpawnedBalancer.ts b/paddler_client_javascript/tests/SpawnedBalancer.ts
index 99110eb2a..1044952cd 100644
--- a/paddler_client_javascript/tests/SpawnedBalancer.ts
+++ b/paddler_client_javascript/tests/SpawnedBalancer.ts
@@ -28,12 +28,14 @@ export class SpawnedBalancer {
       paddlerBinaryPath(),
       [
         "balancer",
+        "--compat-openai-addr",
+        EPHEMERAL_LOOPBACK_ADDRESS,
+        "--compat-typesafe-addr",
+        EPHEMERAL_LOOPBACK_ADDRESS,
         "--inference-addr",
         EPHEMERAL_LOOPBACK_ADDRESS,
         "--management-addr",
         EPHEMERAL_LOOPBACK_ADDRESS,
-        "--compat-openai-addr",
-        EPHEMERAL_LOOPBACK_ADDRESS,
         "--buffered-request-timeout",
         String(bufferedRequestTimeoutMilliseconds),
       ],
diff --git a/paddler_client_javascript/tests/agentsStreamUrl.ts b/paddler_client_javascript/tests/agentsStreamUrl.ts
new file mode 100644
index 000000000..79d1276c3
--- /dev/null
+++ b/paddler_client_javascript/tests/agentsStreamUrl.ts
@@ -0,0 +1,3 @@
+export function agentsStreamUrl(managementAddress: string): string {
+  return `http://${managementAddress}/api/v1/agents/stream`;
+}
diff --git a/paddler_client_javascript/tests/assertFetchJsonReadsStoredDesiredState.ts b/paddler_client_javascript/tests/assertFetchJsonReadsStoredDesiredState.ts
new file mode 100644
index 000000000..470066059
--- /dev/null
+++ b/paddler_client_javascript/tests/assertFetchJsonReadsStoredDesiredState.ts
@@ -0,0 +1,36 @@
+import { deepStrictEqual } from "node:assert/strict";
+
+import { fetchJson } from "../src/fetchJson";
+import {
+  BalancerDesiredStateSchema,
+  type BalancerDesiredState,
+} from "../src/schemas/BalancerDesiredState";
+import { putBalancerDesiredState } from "./putBalancerDesiredState";
+
+export async function assertFetchJsonReadsStoredDesiredState({
+  managementAddress,
+  updateDesiredState,
+}: {
+  managementAddress: string;
+  updateDesiredState(
+    this: void,
+    storedDesiredState: BalancerDesiredState,
+  ): BalancerDesiredState;
+}): Promise<void> {
+  function fetchDesiredState(): Promise<BalancerDesiredState> {
+    return fetchJson({
+      schema: BalancerDesiredStateSchema,
+      signal: new AbortController().signal,
+      url: `http://${managementAddress}/api/v1/balancer_desired_state`,
+    });
+  }
+
+  const updatedDesiredState = updateDesiredState(await fetchDesiredState());
+
+  await putBalancerDesiredState({
+    desiredState: updatedDesiredState,
+    managementAddress,
+  });
+
+  deepStrictEqual(await fetchDesiredState(), updatedDesiredState);
+}
diff --git a/paddler_client_javascript/tests/assertFetchJsonReadsStoredInferenceParameters.ts b/paddler_client_javascript/tests/assertFetchJsonReadsStoredInferenceParameters.ts
deleted file mode 100644
index c77ea5232..000000000
--- a/paddler_client_javascript/tests/assertFetchJsonReadsStoredInferenceParameters.ts
+++ /dev/null
@@ -1,41 +0,0 @@
-import { deepStrictEqual } from "node:assert/strict";
-
-import { fetchJson } from "../src/fetchJson";
-import {
-  BalancerDesiredStateSchema,
-  type BalancerDesiredState,
-} from "../src/schemas/BalancerDesiredState";
-import type { InferenceParameters } from "../src/schemas/InferenceParameters";
-import { putBalancerDesiredState } from "./putBalancerDesiredState";
-
-export async function assertFetchJsonReadsStoredInferenceParameters({
-  inferenceParameters,
-  managementAddress,
-}: {
-  inferenceParameters: Partial<InferenceParameters>;
-  managementAddress: string;
-}): Promise<void> {
-  function fetchDesiredState(): Promise<BalancerDesiredState> {
-    return fetchJson({
-      schema: BalancerDesiredStateSchema,
-      signal: new AbortController().signal,
-      url: `http://${managementAddress}/api/v1/balancer_desired_state`,
-    });
-  }
-
-  const storedDesiredState = await fetchDesiredState();
-  const updatedDesiredState = {
-    ...storedDesiredState,
-    inference_parameters: {
-      ...storedDesiredState.inference_parameters,
-      ...inferenceParameters,
-    },
-  };
-
-  await putBalancerDesiredState({
-    desiredState: updatedDesiredState,
-    managementAddress,
-  });
-
-  deepStrictEqual(await fetchDesiredState(), updatedDesiredState);
-}
diff --git a/paddler_client_javascript/tests/clapUsageErrorExitCode.ts b/paddler_client_javascript/tests/clapUsageErrorExitCode.ts
new file mode 100644
index 000000000..877d0e089
--- /dev/null
+++ b/paddler_client_javascript/tests/clapUsageErrorExitCode.ts
@@ -0,0 +1 @@
+export const CLAP_USAGE_ERROR_EXIT_CODE = 2;
diff --git a/paddler_client_javascript/tests/cluster/fetchJsonReadsADesiredStateThatOffloadsAllLayers.test.ts b/paddler_client_javascript/tests/cluster/fetchJsonReadsADesiredStateThatOffloadsAllLayers.test.ts
index 66f64102d..ecbba5fe1 100644
--- a/paddler_client_javascript/tests/cluster/fetchJsonReadsADesiredStateThatOffloadsAllLayers.test.ts
+++ b/paddler_client_javascript/tests/cluster/fetchJsonReadsADesiredStateThatOffloadsAllLayers.test.ts
@@ -1,7 +1,7 @@
 import { test } from "node:test";
 
 import { ALL_GPU_LAYERS } from "../allGpuLayers";
-import { assertFetchJsonReadsStoredInferenceParameters } from "../assertFetchJsonReadsStoredInferenceParameters";
+import { assertFetchJsonReadsStoredDesiredState } from "../assertFetchJsonReadsStoredDesiredState";
 import { PADDLER_DEFAULT_BUFFERED_REQUEST_TIMEOUT_MILLISECONDS } from "../SpawnedBalancer";
 import { withSpawnedBalancer } from "../withSpawnedBalancer";
 
@@ -12,9 +12,17 @@ test("fetchJson reads a desired state that offloads all layers", async function
         PADDLER_DEFAULT_BUFFERED_REQUEST_TIMEOUT_MILLISECONDS,
     },
     async function ({ management }) {
-      await assertFetchJsonReadsStoredInferenceParameters({
-        inferenceParameters: { n_gpu_layers: ALL_GPU_LAYERS },
+      await assertFetchJsonReadsStoredDesiredState({
         managementAddress: management,
+        updateDesiredState(storedDesiredState) {
+          return {
+            ...storedDesiredState,
+            model_runtime_parameters: {
+              ...storedDesiredState.model_runtime_parameters,
+              n_gpu_layers: ALL_GPU_LAYERS,
+            },
+          };
+        },
       });
     },
   );
diff --git a/paddler_client_javascript/tests/cluster/fetchJsonReadsTheDesiredStateTheBalancerStores.test.ts b/paddler_client_javascript/tests/cluster/fetchJsonReadsTheDesiredStateTheBalancerStores.test.ts
index 73395b03c..0ed84da24 100644
--- a/paddler_client_javascript/tests/cluster/fetchJsonReadsTheDesiredStateTheBalancerStores.test.ts
+++ b/paddler_client_javascript/tests/cluster/fetchJsonReadsTheDesiredStateTheBalancerStores.test.ts
@@ -1,6 +1,6 @@
 import { test } from "node:test";
 
-import { assertFetchJsonReadsStoredInferenceParameters } from "../assertFetchJsonReadsStoredInferenceParameters";
+import { assertFetchJsonReadsStoredDesiredState } from "../assertFetchJsonReadsStoredDesiredState";
 import { PADDLER_DEFAULT_BUFFERED_REQUEST_TIMEOUT_MILLISECONDS } from "../SpawnedBalancer";
 import { withSpawnedBalancer } from "../withSpawnedBalancer";
 
@@ -11,9 +11,26 @@ test("fetchJson reads the desired state the balancer stores", async function ()
         PADDLER_DEFAULT_BUFFERED_REQUEST_TIMEOUT_MILLISECONDS,
     },
     async function ({ management }) {
-      await assertFetchJsonReadsStoredInferenceParameters({
-        inferenceParameters: { enable_embeddings: true },
+      await assertFetchJsonReadsStoredDesiredState({
         managementAddress: management,
+        updateDesiredState(storedDesiredState) {
+          return {
+            ...storedDesiredState,
+            decision: {
+              pointer_head: {
+                Uri: "agent:///pointer-head-no-agent-loads.gguf",
+              },
+            },
+            text_generation: {
+              ...storedDesiredState.text_generation,
+              chat_template_source: {
+                Override: {
+                  content: "{{ messages }}",
+                },
+              },
+            },
+          };
+        },
       });
     },
   );
diff --git a/paddler_client_javascript/tests/cluster/inferenceSocketClientDeliversATimeoutWithoutAgents.test.ts b/paddler_client_javascript/tests/cluster/inferenceSocketClientDeliversATimeoutWithoutAgents.test.ts
index af3a69e91..3c4b54e1e 100644
--- a/paddler_client_javascript/tests/cluster/inferenceSocketClientDeliversATimeoutWithoutAgents.test.ts
+++ b/paddler_client_javascript/tests/cluster/inferenceSocketClientDeliversATimeoutWithoutAgents.test.ts
@@ -26,7 +26,7 @@ test("inferenceSocketClient delivers a timeout without agents", async function (
           inferenceSocketClient({ webSocket })
             .continueConversation({
               enableThinking: false,
-              messages: [{ content: "Hello", role: "user" }],
+              messages: [{ Message: { content: "Hello", role: "user" } }],
             })
             .pipe(toArray()),
         );
diff --git a/paddler_client_javascript/tests/cluster/inferenceSocketClientReportsDisabledTokenGeneration.test.ts b/paddler_client_javascript/tests/cluster/inferenceSocketClientReportsDisabledTokenGeneration.test.ts
deleted file mode 100644
index c77c79dc1..000000000
--- a/paddler_client_javascript/tests/cluster/inferenceSocketClientReportsDisabledTokenGeneration.test.ts
+++ /dev/null
@@ -1,71 +0,0 @@
-import { deepStrictEqual, strictEqual } from "node:assert/strict";
-import { once } from "node:events";
-import { test } from "node:test";
-import { firstValueFrom, lastValueFrom, toArray } from "rxjs";
-
-import { fetchJson } from "../../src/fetchJson";
-import { inferenceSocketClient } from "../../src/inferenceSocketClient";
-import { BalancerDesiredStateSchema } from "../../src/schemas/BalancerDesiredState";
-import { inferenceSocketUrl } from "../inferenceSocketUrl";
-import { putBalancerDesiredState } from "../putBalancerDesiredState";
-import { PADDLER_DEFAULT_BUFFERED_REQUEST_TIMEOUT_MILLISECONDS } from "../SpawnedBalancer";
-import { withSpawnedBalancer } from "../withSpawnedBalancer";
-
-const NOT_IMPLEMENTED = 501;
-
-test("inferenceSocketClient reports disabled token generation", async function () {
-  await withSpawnedBalancer(
-    {
-      bufferedRequestTimeoutMilliseconds:
-        PADDLER_DEFAULT_BUFFERED_REQUEST_TIMEOUT_MILLISECONDS,
-    },
-    async function (addresses) {
-      const storedDesiredState = await fetchJson({
-        schema: BalancerDesiredStateSchema,
-        signal: new AbortController().signal,
-        url: `http://${addresses.management}/api/v1/balancer_desired_state`,
-      });
-
-      await putBalancerDesiredState({
-        desiredState: {
-          ...storedDesiredState,
-          inference_parameters: {
-            ...storedDesiredState.inference_parameters,
-            enable_embeddings: true,
-          },
-        },
-        managementAddress: addresses.management,
-      });
-
-      const webSocket = new WebSocket(inferenceSocketUrl(addresses));
-
-      try {
-        const { clusterTokenGenerationMode$, continueConversation } =
-          inferenceSocketClient({ webSocket });
-        const firstTokenGenerationMode = firstValueFrom(
-          clusterTokenGenerationMode$,
-        );
-
-        await once(webSocket, "open");
-
-        strictEqual(await firstTokenGenerationMode, "TokenGenerationDisabled");
-
-        const responses = await lastValueFrom(
-          continueConversation({
-            enableThinking: false,
-            messages: [{ content: "Hello", role: "user" }],
-          }).pipe(toArray()),
-        );
-
-        deepStrictEqual(
-          responses.map(function ({ done, error }) {
-            return { done, errorCode: error?.code };
-          }),
-          [{ done: true, errorCode: NOT_IMPLEMENTED }],
-        );
-      } finally {
-        webSocket.close();
-      }
-    },
-  );
-});
diff --git a/paddler_client_javascript/tests/cluster/putBalancerDesiredStateRejectsAnInvalidState.test.ts b/paddler_client_javascript/tests/cluster/putBalancerDesiredStateRejectsAnInvalidState.test.ts
index 9ea60a5d4..9d750bebc 100644
--- a/paddler_client_javascript/tests/cluster/putBalancerDesiredStateRejectsAnInvalidState.test.ts
+++ b/paddler_client_javascript/tests/cluster/putBalancerDesiredStateRejectsAnInvalidState.test.ts
@@ -27,9 +27,9 @@ test("putBalancerDesiredState rejects an invalid state", async function () {
         putBalancerDesiredState({
           desiredState: {
             ...storedDesiredState,
-            inference_parameters: {
-              ...storedDesiredState.inference_parameters,
-              embedding_batch_size: 0,
+            model_runtime_parameters: {
+              ...storedDesiredState.model_runtime_parameters,
+              n_gpu_layers: -2,
             },
           },
           managementAddress: management,
diff --git a/paddler_client_javascript/tests/cluster/spawnedBalancerReportsAnExitBeforeAnnouncing.test.ts b/paddler_client_javascript/tests/cluster/spawnedBalancerReportsAnExitBeforeAnnouncing.test.ts
index 4dbb99d95..c2a90eeec 100644
--- a/paddler_client_javascript/tests/cluster/spawnedBalancerReportsAnExitBeforeAnnouncing.test.ts
+++ b/paddler_client_javascript/tests/cluster/spawnedBalancerReportsAnExitBeforeAnnouncing.test.ts
@@ -2,13 +2,14 @@ import { rejects } from "node:assert/strict";
 import { test } from "node:test";
 
 import { BalancerExitedBeforeAnnouncingError } from "../BalancerExitedBeforeAnnouncingError";
+import { CLAP_USAGE_ERROR_EXIT_CODE } from "../clapUsageErrorExitCode";
 import { SpawnedBalancer } from "../SpawnedBalancer";
 
-const CLAP_USAGE_ERROR_EXIT_CODE = 2;
-
 test("spawned balancer reports an exit before announcing", async function () {
   await rejects(
-    SpawnedBalancer.spawn({ bufferedRequestTimeoutMilliseconds: -1 }),
+    SpawnedBalancer.spawn({
+      bufferedRequestTimeoutMilliseconds: -1,
+    }),
     function (error: unknown) {
       return (
         error instanceof BalancerExitedBeforeAnnouncingError &&
diff --git a/paddler_client_javascript/tests/cluster/spawnedProcessReportsAFailedExit.test.ts b/paddler_client_javascript/tests/cluster/spawnedProcessReportsAFailedExit.test.ts
index cfb1b4765..5e8d3190f 100644
--- a/paddler_client_javascript/tests/cluster/spawnedProcessReportsAFailedExit.test.ts
+++ b/paddler_client_javascript/tests/cluster/spawnedProcessReportsAFailedExit.test.ts
@@ -2,12 +2,11 @@ import { rejects } from "node:assert/strict";
 import { spawn } from "node:child_process";
 import { test } from "node:test";
 
+import { CLAP_USAGE_ERROR_EXIT_CODE } from "../clapUsageErrorExitCode";
 import { paddlerBinaryPath } from "../paddlerBinaryPath";
 import { ProcessExitedWithFailureError } from "../ProcessExitedWithFailureError";
 import { SpawnedProcess } from "../SpawnedProcess";
 
-const CLAP_USAGE_ERROR_EXIT_CODE = 2;
-
 test("spawned process reports a failed exit", async function () {
   const spawnedProcess = new SpawnedProcess(
     spawn(paddlerBinaryPath(), ["--flag-paddler-does-not-have"], {
diff --git a/paddler_client_javascript/tests/cluster/streamEventSourceConnectsAndDeliversSnapshots.test.ts b/paddler_client_javascript/tests/cluster/streamEventSourceConnectsAndDeliversSnapshots.test.ts
index a6ccfa754..f796eb80c 100644
--- a/paddler_client_javascript/tests/cluster/streamEventSourceConnectsAndDeliversSnapshots.test.ts
+++ b/paddler_client_javascript/tests/cluster/streamEventSourceConnectsAndDeliversSnapshots.test.ts
@@ -6,6 +6,7 @@ import { eventSourceConnectedState } from "../../src/EventSourceConnectedState";
 import { eventSourceInitialState } from "../../src/EventSourceInitialState";
 import { AgentsResponseSchema } from "../../src/schemas/AgentsResponse";
 import { streamEventSource } from "../../src/streamEventSource";
+import { agentsStreamUrl } from "../agentsStreamUrl";
 import { PADDLER_DEFAULT_BUFFERED_REQUEST_TIMEOUT_MILLISECONDS } from "../SpawnedBalancer";
 import { withSpawnedBalancer } from "../withSpawnedBalancer";
 
@@ -20,7 +21,7 @@ test("streamEventSource connects and delivers snapshots", async function () {
         await firstValueFrom(
           streamEventSource({
             schema: AgentsResponseSchema,
-            url: `http://${management}/api/v1/agents/stream`,
+            url: agentsStreamUrl(management),
           }).pipe(take(3), toArray()),
         ),
         [
diff --git a/paddler_client_javascript/tests/cluster/streamEventSourceReportsAConnectionError.test.ts b/paddler_client_javascript/tests/cluster/streamEventSourceReportsAConnectionError.test.ts
index d4b179cd7..bbfbb873b 100644
--- a/paddler_client_javascript/tests/cluster/streamEventSourceReportsAConnectionError.test.ts
+++ b/paddler_client_javascript/tests/cluster/streamEventSourceReportsAConnectionError.test.ts
@@ -6,18 +6,18 @@ import { eventSourceConnectionErrorState } from "../../src/EventSourceConnection
 import { eventSourceInitialState } from "../../src/EventSourceInitialState";
 import { AgentsResponseSchema } from "../../src/schemas/AgentsResponse";
 import { streamEventSource } from "../../src/streamEventSource";
-import { withConnectionClosingListener } from "../withConnectionClosingListener";
+import { agentsStreamUrl } from "../agentsStreamUrl";
+
+const UNREACHABLE_ADDRESS = "127.0.0.1:1";
 
 test("streamEventSource reports a connection error", async function () {
-  await withConnectionClosingListener(async function (address) {
-    deepStrictEqual(
-      await firstValueFrom(
-        streamEventSource({
-          schema: AgentsResponseSchema,
-          url: `http://${address}/api/v1/agents/stream`,
-        }).pipe(take(2), toArray()),
-      ),
-      [eventSourceInitialState, eventSourceConnectionErrorState],
-    );
-  });
+  deepStrictEqual(
+    await firstValueFrom(
+      streamEventSource({
+        schema: AgentsResponseSchema,
+        url: agentsStreamUrl(UNREACHABLE_ADDRESS),
+      }).pipe(take(2), toArray()),
+    ),
+    [eventSourceInitialState, eventSourceConnectionErrorState],
+  );
 });
diff --git a/paddler_client_javascript/tests/cluster/streamHttpNdjsonCompletesWhenItsRequestIsAborted.test.ts b/paddler_client_javascript/tests/cluster/streamHttpNdjsonCompletesWhenItsRequestIsAborted.test.ts
index ffde504d8..1eabfae9c 100644
--- a/paddler_client_javascript/tests/cluster/streamHttpNdjsonCompletesWhenItsRequestIsAborted.test.ts
+++ b/paddler_client_javascript/tests/cluster/streamHttpNdjsonCompletesWhenItsRequestIsAborted.test.ts
@@ -21,7 +21,9 @@ test("streamHttpNdjson completes when its request is aborted", async function ()
         streamHttpNdjson({
           body: {
             add_generation_prompt: true,
-            conversation_history: [{ content: "Hello", role: "user" }],
+            conversation_history: [
+              { Message: { content: "Hello", role: "user" } },
+            ],
             enable_thinking: false,
             max_tokens: 1,
           },
diff --git a/paddler_client_javascript/tests/cluster/streamHttpNdjsonDeliversADecisionTimeoutWithoutAgents.test.ts b/paddler_client_javascript/tests/cluster/streamHttpNdjsonDeliversADecisionTimeoutWithoutAgents.test.ts
new file mode 100644
index 000000000..a4b1c6a24
--- /dev/null
+++ b/paddler_client_javascript/tests/cluster/streamHttpNdjsonDeliversADecisionTimeoutWithoutAgents.test.ts
@@ -0,0 +1,62 @@
+import { deepStrictEqual } from "node:assert/strict";
+import { test } from "node:test";
+import { lastValueFrom, toArray } from "rxjs";
+
+import { fetchJson } from "../../src/fetchJson";
+import { BalancerDesiredStateSchema } from "../../src/schemas/BalancerDesiredState";
+import type { DecideParams } from "../../src/schemas/DecideParams";
+import { InferenceServiceDecideResponseSchema } from "../../src/schemas/InferenceServiceDecideResponse";
+import { streamHttpNdjson } from "../../src/streamHttpNdjson";
+import { putBalancerDesiredState } from "../putBalancerDesiredState";
+import { SHORT_BUFFERED_REQUEST_TIMEOUT_MILLISECONDS } from "../shortBufferedRequestTimeout";
+import { withSpawnedBalancer } from "../withSpawnedBalancer";
+
+const GATEWAY_TIMEOUT = 504;
+
+test("streamHttpNdjson delivers a decision timeout without agents", async function () {
+  await withSpawnedBalancer(
+    {
+      bufferedRequestTimeoutMilliseconds:
+        SHORT_BUFFERED_REQUEST_TIMEOUT_MILLISECONDS,
+    },
+    async function ({ inference, management }) {
+      const storedDesiredState = await fetchJson({
+        schema: BalancerDesiredStateSchema,
+        signal: new AbortController().signal,
+        url: `http://${management}/api/v1/balancer_desired_state`,
+      });
+
+      await putBalancerDesiredState({
+        desiredState: {
+          ...storedDesiredState,
+          decision: {
+            pointer_head: { Uri: "agent:///pointer-head-no-agent-loads.gguf" },
+          },
+        },
+        managementAddress: management,
+      });
+
+      const decideParams: DecideParams = {
+        questions: [
+          { id: "paid", instructions: "Was it paid?", options: ["no", "yes"] },
+        ],
+        state: "The invoice was paid on time.",
+      };
+      const responses = await lastValueFrom(
+        streamHttpNdjson({
+          body: decideParams,
+          schema: InferenceServiceDecideResponseSchema,
+          signal: new AbortController().signal,
+          url: `http://${inference}/api/v1/decide`,
+        }).pipe(toArray()),
+      );
+
+      deepStrictEqual(
+        responses.map(function ({ done, error }) {
+          return { done, errorCode: error?.code };
+        }),
+        [{ done: true, errorCode: GATEWAY_TIMEOUT }],
+      );
+    },
+  );
+});
diff --git a/paddler_client_javascript/tests/cluster/streamHttpNdjsonDeliversATimeoutWithoutAgents.test.ts b/paddler_client_javascript/tests/cluster/streamHttpNdjsonDeliversATimeoutWithoutAgents.test.ts
index 27b26f30a..fd2dd6b6a 100644
--- a/paddler_client_javascript/tests/cluster/streamHttpNdjsonDeliversATimeoutWithoutAgents.test.ts
+++ b/paddler_client_javascript/tests/cluster/streamHttpNdjsonDeliversATimeoutWithoutAgents.test.ts
@@ -20,7 +20,9 @@ test("streamHttpNdjson delivers a timeout without agents", async function () {
         streamHttpNdjson({
           body: {
             add_generation_prompt: true,
-            conversation_history: [{ content: "Hello", role: "user" }],
+            conversation_history: [
+              { Message: { content: "Hello", role: "user" } },
+            ],
             enable_thinking: false,
             max_tokens: 1,
           },
diff --git a/paddler_client_javascript/tests/cluster/waitForAgentReadyReportsAClosedStream.test.ts b/paddler_client_javascript/tests/cluster/waitForAgentReadyReportsAClosedStream.test.ts
index 7ce694607..4876fb10f 100644
--- a/paddler_client_javascript/tests/cluster/waitForAgentReadyReportsAClosedStream.test.ts
+++ b/paddler_client_javascript/tests/cluster/waitForAgentReadyReportsAClosedStream.test.ts
@@ -4,26 +4,45 @@ import { test } from "node:test";
 import { AgentsResponseSchema } from "../../src/schemas/AgentsResponse";
 import { streamEventSource } from "../../src/streamEventSource";
 import { AgentsStreamClosedError } from "../AgentsStreamClosedError";
+import { agentsStreamUrl } from "../agentsStreamUrl";
+import {
+  PADDLER_DEFAULT_BUFFERED_REQUEST_TIMEOUT_MILLISECONDS,
+  SpawnedBalancer,
+} from "../SpawnedBalancer";
 import { waitForAgentReady } from "../waitForAgentReady";
-import { withConnectionClosingListener } from "../withConnectionClosingListener";
+import { withReadyModelLessAgent } from "../withReadyModelLessAgent";
+
+const AGENT = { name: "agent-that-never-gets-its-slots", slots: 1 };
 
 test("waitForAgentReady reports a closed stream", async function () {
-  await withConnectionClosingListener(async function (address) {
-    await rejects(
-      waitForAgentReady({
-        agentName: "agent-that-never-joined",
-        agentsStates: streamEventSource({
-          schema: AgentsResponseSchema,
-          url: `http://${address}/api/v1/agents/stream`,
-        }),
-        expectedSlotsTotal: 1,
-      }),
-      function (error: unknown) {
-        return (
-          error instanceof AgentsStreamClosedError &&
-          error.agentName === "agent-that-never-joined"
-        );
-      },
-    );
+  const balancer = await SpawnedBalancer.spawn({
+    bufferedRequestTimeoutMilliseconds:
+      PADDLER_DEFAULT_BUFFERED_REQUEST_TIMEOUT_MILLISECONDS,
   });
+
+  await withReadyModelLessAgent(
+    { agent: AGENT, managementAddress: balancer.addresses.management },
+    async function (agentProcess) {
+      const closedStreamRejection = rejects(
+        waitForAgentReady({
+          agentName: AGENT.name,
+          agentProcess,
+          agentsStates: streamEventSource({
+            schema: AgentsResponseSchema,
+            url: agentsStreamUrl(balancer.addresses.management),
+          }),
+          expectedSlotsTotal: AGENT.slots,
+        }),
+        function (error: unknown) {
+          return (
+            error instanceof AgentsStreamClosedError &&
+            error.agentName === AGENT.name
+          );
+        },
+      );
+
+      await balancer.process.terminate();
+      await closedStreamRejection;
+    },
+  );
 });
diff --git a/paddler_client_javascript/tests/cluster/waitForAgentReadyReportsAgentIssues.test.ts b/paddler_client_javascript/tests/cluster/waitForAgentReadyReportsAgentIssues.test.ts
index 755f6bdc1..746749232 100644
--- a/paddler_client_javascript/tests/cluster/waitForAgentReadyReportsAgentIssues.test.ts
+++ b/paddler_client_javascript/tests/cluster/waitForAgentReadyReportsAgentIssues.test.ts
@@ -7,6 +7,7 @@ import { AgentsResponseSchema } from "../../src/schemas/AgentsResponse";
 import { BalancerDesiredStateSchema } from "../../src/schemas/BalancerDesiredState";
 import { streamEventSource } from "../../src/streamEventSource";
 import { AgentReportedIssuesError } from "../AgentReportedIssuesError";
+import { agentsStreamUrl } from "../agentsStreamUrl";
 import { putBalancerDesiredState } from "../putBalancerDesiredState";
 import { spawnAgent } from "../spawnAgent";
 import { PADDLER_DEFAULT_BUFFERED_REQUEST_TIMEOUT_MILLISECONDS } from "../SpawnedBalancer";
@@ -31,7 +32,7 @@ test("waitForAgentReady reports agent issues", async function () {
       await putBalancerDesiredState({
         desiredState: {
           ...storedDesiredState,
-          model: { LocalToAgent: "/nonexistent/model.gguf" },
+          model: { Uri: "agent:///nonexistent/model.gguf" },
         },
         managementAddress: management,
       });
@@ -45,9 +46,10 @@ test("waitForAgentReady reports agent issues", async function () {
         await rejects(
           waitForAgentReady({
             agentName: AGENT.name,
+            agentProcess,
             agentsStates: streamEventSource({
               schema: AgentsResponseSchema,
-              url: `http://${management}/api/v1/agents/stream`,
+              url: agentsStreamUrl(management),
             }),
             expectedSlotsTotal: AGENT.slots,
           }),
diff --git a/paddler_client_javascript/tests/cluster/waitForAgentReadyReportsAnAgentThatExits.test.ts b/paddler_client_javascript/tests/cluster/waitForAgentReadyReportsAnAgentThatExits.test.ts
new file mode 100644
index 000000000..1c43c2c91
--- /dev/null
+++ b/paddler_client_javascript/tests/cluster/waitForAgentReadyReportsAnAgentThatExits.test.ts
@@ -0,0 +1,46 @@
+import { rejects } from "node:assert/strict";
+import { test } from "node:test";
+
+import { AgentsResponseSchema } from "../../src/schemas/AgentsResponse";
+import { streamEventSource } from "../../src/streamEventSource";
+import { AgentExitedBeforeReadyError } from "../AgentExitedBeforeReadyError";
+import { agentsStreamUrl } from "../agentsStreamUrl";
+import { CLAP_USAGE_ERROR_EXIT_CODE } from "../clapUsageErrorExitCode";
+import { spawnAgent } from "../spawnAgent";
+import { PADDLER_DEFAULT_BUFFERED_REQUEST_TIMEOUT_MILLISECONDS } from "../SpawnedBalancer";
+import { waitForAgentReady } from "../waitForAgentReady";
+import { withSpawnedBalancer } from "../withSpawnedBalancer";
+
+const AGENT = { name: "agent-with-slots-it-rejects", slots: 0 };
+
+test("waitForAgentReady reports an agent that exits", async function () {
+  await withSpawnedBalancer(
+    {
+      bufferedRequestTimeoutMilliseconds:
+        PADDLER_DEFAULT_BUFFERED_REQUEST_TIMEOUT_MILLISECONDS,
+    },
+    async function ({ management }) {
+      await rejects(
+        waitForAgentReady({
+          agentName: AGENT.name,
+          agentProcess: spawnAgent({
+            agent: AGENT,
+            managementAddress: management,
+          }),
+          agentsStates: streamEventSource({
+            schema: AgentsResponseSchema,
+            url: agentsStreamUrl(management),
+          }),
+          expectedSlotsTotal: AGENT.slots,
+        }),
+        function (error: unknown) {
+          return (
+            error instanceof AgentExitedBeforeReadyError &&
+            error.agentName === AGENT.name &&
+            error.exitCode === CLAP_USAGE_ERROR_EXIT_CODE
+          );
+        },
+      );
+    },
+  );
+});
diff --git a/paddler_client_javascript/tests/cluster/waitForAgentReadyReportsAnUndeserializableSnapshot.test.ts b/paddler_client_javascript/tests/cluster/waitForAgentReadyReportsAnUndeserializableSnapshot.test.ts
index 326d0409d..d692cb0b3 100644
--- a/paddler_client_javascript/tests/cluster/waitForAgentReadyReportsAnUndeserializableSnapshot.test.ts
+++ b/paddler_client_javascript/tests/cluster/waitForAgentReadyReportsAnUndeserializableSnapshot.test.ts
@@ -6,8 +6,11 @@ import { streamEventSource } from "../../src/streamEventSource";
 import { AgentsSnapshotUndeserializableError } from "../AgentsSnapshotUndeserializableError";
 import { PADDLER_DEFAULT_BUFFERED_REQUEST_TIMEOUT_MILLISECONDS } from "../SpawnedBalancer";
 import { waitForAgentReady } from "../waitForAgentReady";
+import { withReadyModelLessAgent } from "../withReadyModelLessAgent";
 import { withSpawnedBalancer } from "../withSpawnedBalancer";
 
+const AGENT = { name: "agent-awaited-on-the-wrong-stream", slots: 1 };
+
 test("waitForAgentReady reports an undeserializable snapshot", async function () {
   await withSpawnedBalancer(
     {
@@ -15,16 +18,22 @@ test("waitForAgentReady reports an undeserializable snapshot", async function ()
         PADDLER_DEFAULT_BUFFERED_REQUEST_TIMEOUT_MILLISECONDS,
     },
     async function ({ management }) {
-      await rejects(
-        waitForAgentReady({
-          agentName: "agent-that-never-joined",
-          agentsStates: streamEventSource({
-            schema: AgentsResponseSchema,
-            url: `http://${management}/api/v1/buffered_requests/stream`,
-          }),
-          expectedSlotsTotal: 1,
-        }),
-        AgentsSnapshotUndeserializableError,
+      await withReadyModelLessAgent(
+        { agent: AGENT, managementAddress: management },
+        async function (agentProcess) {
+          await rejects(
+            waitForAgentReady({
+              agentName: AGENT.name,
+              agentProcess,
+              agentsStates: streamEventSource({
+                schema: AgentsResponseSchema,
+                url: `http://${management}/api/v1/buffered_requests/stream`,
+              }),
+              expectedSlotsTotal: AGENT.slots,
+            }),
+            AgentsSnapshotUndeserializableError,
+          );
+        },
       );
     },
   );
diff --git a/paddler_client_javascript/tests/llm/inferenceSocketClientStreamsTokensUntilDone.test.ts b/paddler_client_javascript/tests/llm/inferenceSocketClientStreamsTokensUntilDone.test.ts
index cca6029d8..d1d4b2428 100644
--- a/paddler_client_javascript/tests/llm/inferenceSocketClientStreamsTokensUntilDone.test.ts
+++ b/paddler_client_javascript/tests/llm/inferenceSocketClientStreamsTokensUntilDone.test.ts
@@ -24,7 +24,7 @@ test("inferenceSocketClient streams tokens until done", async function () {
           inferenceSocketClient({ webSocket })
             .continueConversation({
               enableThinking: false,
-              messages: [{ content: "Say hello.", role: "user" }],
+              messages: [{ Message: { content: "Say hello.", role: "user" } }],
             })
             .pipe(toArray()),
         );
diff --git a/paddler_client_javascript/tests/llm/streamHttpNdjsonDeliversAnInvalidToolSchemaRejection.test.ts b/paddler_client_javascript/tests/llm/streamHttpNdjsonDeliversAnInvalidToolSchemaRejection.test.ts
index e516d21b6..cd32594b0 100644
--- a/paddler_client_javascript/tests/llm/streamHttpNdjsonDeliversAnInvalidToolSchemaRejection.test.ts
+++ b/paddler_client_javascript/tests/llm/streamHttpNdjsonDeliversAnInvalidToolSchemaRejection.test.ts
@@ -21,7 +21,12 @@ test("streamHttpNdjson delivers an invalid tool schema rejection", async functio
           body: {
             add_generation_prompt: true,
             conversation_history: [
-              { content: "What is the weather in Paris?", role: "user" },
+              {
+                Message: {
+                  content: "What is the weather in Paris?",
+                  role: "user",
+                },
+              },
             ],
             enable_thinking: false,
             max_tokens: 64,
diff --git a/paddler_client_javascript/tests/qwen3DesiredState.ts b/paddler_client_javascript/tests/qwen3DesiredState.ts
index 69736d5f4..04c21bea7 100644
--- a/paddler_client_javascript/tests/qwen3DesiredState.ts
+++ b/paddler_client_javascript/tests/qwen3DesiredState.ts
@@ -6,22 +6,25 @@ export function qwen3DesiredState(
 ): BalancerDesiredState {
   return {
     ...storedDesiredState,
-    inference_parameters: {
-      ...storedDesiredState.inference_parameters,
-      min_p: 0,
+    decision: { pointer_head: "None" },
+    model: {
+      Uri: "https://huggingface.co/Qwen/Qwen3-0.6B-GGUF/blob/main/Qwen3-0.6B-Q8_0.gguf",
+    },
+    model_runtime_parameters: {
+      ...storedDesiredState.model_runtime_parameters,
       n_gpu_layers: ALL_GPU_LAYERS,
-      penalty_frequency: 0,
-      penalty_presence: 0,
-      penalty_repeat: 1,
-      temperature: 0,
-      top_k: 1,
-      top_p: 1,
     },
-    model: {
-      HuggingFace: {
-        filename: "Qwen3-0.6B-Q8_0.gguf",
-        repo_id: "Qwen/Qwen3-0.6B-GGUF",
-        revision: "main",
+    text_generation: {
+      ...storedDesiredState.text_generation,
+      sampling_parameters: {
+        min_p: 0,
+        penalty_frequency: 0,
+        penalty_last_n: 0,
+        penalty_presence: 0,
+        penalty_repeat: 1,
+        temperature: 0,
+        top_k: 1,
+        top_p: 1,
       },
     },
   };
diff --git a/paddler_client_javascript/tests/unit/CountSchema.ts b/paddler_client_javascript/tests/unit/CountSchema.ts
new file mode 100644
index 000000000..0b0985a98
--- /dev/null
+++ b/paddler_client_javascript/tests/unit/CountSchema.ts
@@ -0,0 +1,3 @@
+import { z } from "zod";
+
+export const CountSchema = z.object({ count: z.number() });
diff --git a/paddler_client_javascript/tests/unit/NdjsonDecoder.test.ts b/paddler_client_javascript/tests/unit/NdjsonDecoder.test.ts
deleted file mode 100644
index c512c5cf1..000000000
--- a/paddler_client_javascript/tests/unit/NdjsonDecoder.test.ts
+++ /dev/null
@@ -1,41 +0,0 @@
-import { deepStrictEqual, throws } from "node:assert/strict";
-import { test } from "node:test";
-
-import { JsonError } from "../../src/JsonError";
-import { NdjsonDecoder } from "../../src/NdjsonDecoder";
-
-test("joins a value split across pushed chunks", function () {
-  const ndjsonDecoder = new NdjsonDecoder();
-
-  deepStrictEqual(ndjsonDecoder.push('{"index":'), []);
-  deepStrictEqual(ndjsonDecoder.push('0}\n{"index":1}\n'), [
-    { index: 0 },
-    { index: 1 },
-  ]);
-});
-
-test("skips blank lines", function () {
-  const ndjsonDecoder = new NdjsonDecoder();
-
-  deepStrictEqual(ndjsonDecoder.push('\n  \r\n{"index":0}\n'), [{ index: 0 }]);
-});
-
-test("decodes a trailing value without a newline when finished", function () {
-  const ndjsonDecoder = new NdjsonDecoder();
-
-  deepStrictEqual(ndjsonDecoder.push('{"index":0}'), []);
-  deepStrictEqual(ndjsonDecoder.finish(), [{ index: 0 }]);
-});
-
-test("rejects a line that is not JSON with the raw line attached", function () {
-  const ndjsonDecoder = new NdjsonDecoder();
-
-  throws(
-    function () {
-      ndjsonDecoder.push("not json\n");
-    },
-    function (error: unknown) {
-      return error instanceof JsonError && error.raw === "not json";
-    },
-  );
-});
diff --git a/paddler_client_javascript/tests/unit/eventSourceMessageState.test.ts b/paddler_client_javascript/tests/unit/eventSourceMessageState.test.ts
deleted file mode 100644
index 96420795a..000000000
--- a/paddler_client_javascript/tests/unit/eventSourceMessageState.test.ts
+++ /dev/null
@@ -1,36 +0,0 @@
-import { deepStrictEqual, strictEqual } from "node:assert/strict";
-import { test } from "node:test";
-import { z } from "zod";
-
-import { eventSourceDeserializationErrorState } from "../../src/EventSourceDeserializationErrorState";
-import { eventSourceMessageState } from "../../src/eventSourceMessageState";
-
-const Schema = z.object({ count: z.number() });
-
-test("a message matching the schema becomes a data snapshot", function () {
-  deepStrictEqual(
-    eventSourceMessageState({ data: '{"count":7}', schema: Schema }),
-    {
-      data: { count: 7 },
-      isConnected: true,
-      isConnectionError: false,
-      isDeserializationError: false,
-      isInitial: false,
-      isOk: true,
-    },
-  );
-});
-
-test("a message that is not JSON is a deserialization error", function () {
-  strictEqual(
-    eventSourceMessageState({ data: "not json", schema: Schema }),
-    eventSourceDeserializationErrorState,
-  );
-});
-
-test("a message that does not match the schema is a deserialization error", function () {
-  strictEqual(
-    eventSourceMessageState({ data: '{"count":"seven"}', schema: Schema }),
-    eventSourceDeserializationErrorState,
-  );
-});
diff --git a/paddler_client_javascript/tests/unit/eventSourceMessageStateReportsAMessageNotMatchingTheSchemaAsADeserializationError.test.ts b/paddler_client_javascript/tests/unit/eventSourceMessageStateReportsAMessageNotMatchingTheSchemaAsADeserializationError.test.ts
new file mode 100644
index 000000000..d3ab0b864
--- /dev/null
+++ b/paddler_client_javascript/tests/unit/eventSourceMessageStateReportsAMessageNotMatchingTheSchemaAsADeserializationError.test.ts
@@ -0,0 +1,13 @@
+import { strictEqual } from "node:assert/strict";
+import { test } from "node:test";
+
+import { eventSourceDeserializationErrorState } from "../../src/EventSourceDeserializationErrorState";
+import { eventSourceMessageState } from "../../src/eventSourceMessageState";
+import { CountSchema } from "./CountSchema";
+
+test("eventSourceMessageState reports a message not matching the schema as a deserialization error", function () {
+  strictEqual(
+    eventSourceMessageState({ data: '{"count":"seven"}', schema: CountSchema }),
+    eventSourceDeserializationErrorState,
+  );
+});
diff --git a/paddler_client_javascript/tests/unit/eventSourceMessageStateReportsAMessageThatIsNotJsonAsADeserializationError.test.ts b/paddler_client_javascript/tests/unit/eventSourceMessageStateReportsAMessageThatIsNotJsonAsADeserializationError.test.ts
new file mode 100644
index 000000000..d0e9cd6ce
--- /dev/null
+++ b/paddler_client_javascript/tests/unit/eventSourceMessageStateReportsAMessageThatIsNotJsonAsADeserializationError.test.ts
@@ -0,0 +1,13 @@
+import { strictEqual } from "node:assert/strict";
+import { test } from "node:test";
+
+import { eventSourceDeserializationErrorState } from "../../src/EventSourceDeserializationErrorState";
+import { eventSourceMessageState } from "../../src/eventSourceMessageState";
+import { CountSchema } from "./CountSchema";
+
+test("eventSourceMessageState reports a message that is not JSON as a deserialization error", function () {
+  strictEqual(
+    eventSourceMessageState({ data: "not json", schema: CountSchema }),
+    eventSourceDeserializationErrorState,
+  );
+});
diff --git a/paddler_client_javascript/tests/unit/eventSourceMessageStateTurnsAMessageMatchingTheSchemaIntoADataSnapshot.test.ts b/paddler_client_javascript/tests/unit/eventSourceMessageStateTurnsAMessageMatchingTheSchemaIntoADataSnapshot.test.ts
new file mode 100644
index 000000000..eb452f47d
--- /dev/null
+++ b/paddler_client_javascript/tests/unit/eventSourceMessageStateTurnsAMessageMatchingTheSchemaIntoADataSnapshot.test.ts
@@ -0,0 +1,19 @@
+import { deepStrictEqual } from "node:assert/strict";
+import { test } from "node:test";
+
+import { eventSourceMessageState } from "../../src/eventSourceMessageState";
+import { CountSchema } from "./CountSchema";
+
+test("eventSourceMessageState turns a message matching the schema into a data snapshot", function () {
+  deepStrictEqual(
+    eventSourceMessageState({ data: '{"count":7}', schema: CountSchema }),
+    {
+      data: { count: 7 },
+      isConnected: true,
+      isConnectionError: false,
+      isDeserializationError: false,
+      isInitial: false,
+      isOk: true,
+    },
+  );
+});
diff --git a/paddler_client_javascript/tests/unit/extractHuggingFaceUrlParts.test.ts b/paddler_client_javascript/tests/unit/extractHuggingFaceUrlParts.test.ts
deleted file mode 100644
index df0d99b4e..000000000
--- a/paddler_client_javascript/tests/unit/extractHuggingFaceUrlParts.test.ts
+++ /dev/null
@@ -1,57 +0,0 @@
-import { deepStrictEqual, throws } from "node:assert/strict";
-import { test } from "node:test";
-
-import { extractHuggingFaceUrlParts } from "../../src/extractHuggingFaceUrlParts";
-import { InvalidHuggingFaceUrlError } from "../../src/InvalidHuggingFaceUrlError";
-
-test("blob URL extracts owner, repo, revision and filename", function () {
-  const url = new URL(
-    "https://huggingface.co/Qwen/Qwen3-0.6B-GGUF/blob/main/Qwen3-0.6B-Q8_0.gguf",
-  );
-
-  deepStrictEqual(extractHuggingFaceUrlParts(url), {
-    filename: "Qwen3-0.6B-Q8_0.gguf",
-    repo_id: "Qwen/Qwen3-0.6B-GGUF",
-    revision: "main",
-  });
-});
-
-test("resolve URL extracts the same fields", function () {
-  const url = new URL(
-    "https://huggingface.co/Qwen/Qwen3-0.6B-GGUF/resolve/main/Qwen3-0.6B-Q8_0.gguf",
-  );
-
-  deepStrictEqual(extractHuggingFaceUrlParts(url), {
-    filename: "Qwen3-0.6B-Q8_0.gguf",
-    repo_id: "Qwen/Qwen3-0.6B-GGUF",
-    revision: "main",
-  });
-});
-
-test("nested filename paths preserve every segment", function () {
-  const url = new URL(
-    "https://huggingface.co/owner/repo/blob/main/dir/sub/file.gguf",
-  );
-
-  deepStrictEqual(extractHuggingFaceUrlParts(url), {
-    filename: "dir/sub/file.gguf",
-    repo_id: "owner/repo",
-    revision: "main",
-  });
-});
-
-test("malformed URLs throw", function () {
-  const url = new URL("https://huggingface.co/owner/repo");
-
-  throws(function () {
-    extractHuggingFaceUrlParts(url);
-  }, InvalidHuggingFaceUrlError);
-});
-
-test("URLs that do not point at a file blob throw", function () {
-  const url = new URL("https://huggingface.co/owner/repo/tree/main/file.gguf");
-
-  throws(function () {
-    extractHuggingFaceUrlParts(url);
-  }, InvalidHuggingFaceUrlError);
-});
diff --git a/paddler_client_javascript/tests/unit/ndjsonDecoderDecodesATrailingValueWithoutANewlineWhenFinished.test.ts b/paddler_client_javascript/tests/unit/ndjsonDecoderDecodesATrailingValueWithoutANewlineWhenFinished.test.ts
new file mode 100644
index 000000000..9162b5f49
--- /dev/null
+++ b/paddler_client_javascript/tests/unit/ndjsonDecoderDecodesATrailingValueWithoutANewlineWhenFinished.test.ts
@@ -0,0 +1,11 @@
+import { deepStrictEqual } from "node:assert/strict";
+import { test } from "node:test";
+
+import { NdjsonDecoder } from "../../src/NdjsonDecoder";
+
+test("NdjsonDecoder decodes a trailing value without a newline when finished", function () {
+  const ndjsonDecoder = new NdjsonDecoder();
+
+  deepStrictEqual(ndjsonDecoder.push('{"index":0}'), []);
+  deepStrictEqual(ndjsonDecoder.finish(), [{ index: 0 }]);
+});
diff --git a/paddler_client_javascript/tests/unit/ndjsonDecoderJoinsAValueSplitAcrossPushedChunks.test.ts b/paddler_client_javascript/tests/unit/ndjsonDecoderJoinsAValueSplitAcrossPushedChunks.test.ts
new file mode 100644
index 000000000..2cf19fa3d
--- /dev/null
+++ b/paddler_client_javascript/tests/unit/ndjsonDecoderJoinsAValueSplitAcrossPushedChunks.test.ts
@@ -0,0 +1,14 @@
+import { deepStrictEqual } from "node:assert/strict";
+import { test } from "node:test";
+
+import { NdjsonDecoder } from "../../src/NdjsonDecoder";
+
+test("NdjsonDecoder joins a value split across pushed chunks", function () {
+  const ndjsonDecoder = new NdjsonDecoder();
+
+  deepStrictEqual(ndjsonDecoder.push('{"index":'), []);
+  deepStrictEqual(ndjsonDecoder.push('0}\n{"index":1}\n'), [
+    { index: 0 },
+    { index: 1 },
+  ]);
+});
diff --git a/paddler_client_javascript/tests/unit/ndjsonDecoderRejectsALineThatIsNotJsonWithTheRawLineAttached.test.ts b/paddler_client_javascript/tests/unit/ndjsonDecoderRejectsALineThatIsNotJsonWithTheRawLineAttached.test.ts
new file mode 100644
index 000000000..212aa8e3a
--- /dev/null
+++ b/paddler_client_javascript/tests/unit/ndjsonDecoderRejectsALineThatIsNotJsonWithTheRawLineAttached.test.ts
@@ -0,0 +1,18 @@
+import { throws } from "node:assert/strict";
+import { test } from "node:test";
+
+import { JsonError } from "../../src/JsonError";
+import { NdjsonDecoder } from "../../src/NdjsonDecoder";
+
+test("NdjsonDecoder rejects a line that is not JSON with the raw line attached", function () {
+  const ndjsonDecoder = new NdjsonDecoder();
+
+  throws(
+    function () {
+      ndjsonDecoder.push("not json\n");
+    },
+    function (error: unknown) {
+      return error instanceof JsonError && error.raw === "not json";
+    },
+  );
+});
diff --git a/paddler_client_javascript/tests/unit/ndjsonDecoderSkipsBlankLines.test.ts b/paddler_client_javascript/tests/unit/ndjsonDecoderSkipsBlankLines.test.ts
new file mode 100644
index 000000000..f39d07359
--- /dev/null
+++ b/paddler_client_javascript/tests/unit/ndjsonDecoderSkipsBlankLines.test.ts
@@ -0,0 +1,10 @@
+import { deepStrictEqual } from "node:assert/strict";
+import { test } from "node:test";
+
+import { NdjsonDecoder } from "../../src/NdjsonDecoder";
+
+test("NdjsonDecoder skips blank lines", function () {
+  const ndjsonDecoder = new NdjsonDecoder();
+
+  deepStrictEqual(ndjsonDecoder.push('\n  \r\n{"index":0}\n'), [{ index: 0 }]);
+});
diff --git a/paddler_client_javascript/tests/unit/schemas/Agent.test.ts b/paddler_client_javascript/tests/unit/schemas/Agent.test.ts
deleted file mode 100644
index 1e9f5173e..000000000
--- a/paddler_client_javascript/tests/unit/schemas/Agent.test.ts
+++ /dev/null
@@ -1,54 +0,0 @@
-import { deepStrictEqual, throws } from "node:assert/strict";
-import { test } from "node:test";
-
-import { AgentSchema } from "../../../src/schemas/Agent";
-
-test("parses an agent that is downloading its model", function () {
-  const parsed = AgentSchema.parse({
-    id: "agent-0",
-    name: "agent-0",
-    slots_processing: 1,
-    status: {
-      desired_slots_total: 4,
-      download_status: {
-        Downloading: {
-          downloaded_bytes: 100,
-          model_path: "https://example.com/qwen.gguf",
-          total_bytes: 400,
-        },
-      },
-      issues: [],
-      model_path: "/models/qwen.gguf",
-      slots_total: 4,
-      state_application_status: "Applied",
-      uses_chat_template_override: false,
-    },
-  });
-
-  deepStrictEqual(parsed.status.download_status, {
-    Downloading: {
-      downloaded_bytes: 100,
-      model_path: "https://example.com/qwen.gguf",
-      total_bytes: 400,
-    },
-  });
-});
-
-test("rejects an unknown state_application_status", function () {
-  throws(function () {
-    AgentSchema.parse({
-      id: "agent-x",
-      name: null,
-      slots_processing: 0,
-      status: {
-        desired_slots_total: 1,
-        download_status: "NotDownloading",
-        issues: [],
-        model_path: null,
-        slots_total: 1,
-        state_application_status: "Unknown",
-        uses_chat_template_override: false,
-      },
-    });
-  });
-});
diff --git a/paddler_client_javascript/tests/unit/schemas/AgentsResponse.test.ts b/paddler_client_javascript/tests/unit/schemas/AgentsResponse.test.ts
deleted file mode 100644
index d23fbec84..000000000
--- a/paddler_client_javascript/tests/unit/schemas/AgentsResponse.test.ts
+++ /dev/null
@@ -1,54 +0,0 @@
-import { deepStrictEqual } from "node:assert/strict";
-import { test } from "node:test";
-
-import { AgentsResponseSchema } from "../../../src/schemas/AgentsResponse";
-
-function agent({ id, name }: { id: string; name: string | null }): unknown {
-  return {
-    id,
-    name,
-    slots_processing: 0,
-    status: {
-      desired_slots_total: 1,
-      download_status: "NotDownloading",
-      issues: [],
-      model_path: null,
-      slots_total: 0,
-      state_application_status: "Applied",
-      uses_chat_template_override: false,
-    },
-  };
-}
-
-test("orders agents by name", function () {
-  const { agents } = AgentsResponseSchema.parse({
-    agents: [
-      agent({ id: "second", name: "second" }),
-      agent({ id: "first", name: "first" }),
-    ],
-  });
-
-  deepStrictEqual(
-    agents.map(function ({ name }) {
-      return name;
-    }),
-    ["first", "second"],
-  );
-});
-
-test("orders an unnamed agent by its id among the named ones", function () {
-  const { agents } = AgentsResponseSchema.parse({
-    agents: [
-      agent({ id: "bravo-id", name: "bravo" }),
-      agent({ id: "alpha-id", name: null }),
-      agent({ id: "charlie-id", name: "charlie" }),
-    ],
-  });
-
-  deepStrictEqual(
-    agents.map(function ({ id }) {
-      return id;
-    }),
-    ["alpha-id", "bravo-id", "charlie-id"],
-  );
-});
diff --git a/paddler_client_javascript/tests/unit/schemas/ContinueFromConversationHistoryParams.test.ts b/paddler_client_javascript/tests/unit/schemas/ContinueFromConversationHistoryParams.test.ts
deleted file mode 100644
index 7ffa13e7a..000000000
--- a/paddler_client_javascript/tests/unit/schemas/ContinueFromConversationHistoryParams.test.ts
+++ /dev/null
@@ -1,42 +0,0 @@
-import { strictEqual } from "node:assert/strict";
-import { test } from "node:test";
-
-import {
-  ContinueFromConversationHistoryParamsSchema,
-  type ContinueFromConversationHistoryParams,
-} from "../../../src/schemas/ContinueFromConversationHistoryParams";
-import { assertSchemaRejectsPaths } from "../../assertSchemaRejectsPaths";
-
-const paramsWithoutTools: ContinueFromConversationHistoryParams = {
-  add_generation_prompt: true,
-  conversation_history: [{ content: "Hello!", role: "user" }],
-  enable_thinking: false,
-  max_tokens: 100,
-  parse_tool_calls: true,
-};
-
-test("accepts tool call parsing with a declared tool", function () {
-  strictEqual(
-    ContinueFromConversationHistoryParamsSchema.safeParse({
-      ...paramsWithoutTools,
-      tools: [
-        {
-          function: {
-            description: "Get the current weather for a location",
-            name: "get_weather",
-          },
-          type: "function",
-        },
-      ],
-    }).success,
-    true,
-  );
-});
-
-test("rejects tool call parsing without tools", function () {
-  assertSchemaRejectsPaths({
-    expectedPaths: [["parse_tool_calls"]],
-    input: paramsWithoutTools,
-    schema: ContinueFromConversationHistoryParamsSchema,
-  });
-});
diff --git a/paddler_client_javascript/tests/unit/schemas/InferenceNotification.test.ts b/paddler_client_javascript/tests/unit/schemas/InferenceNotification.test.ts
deleted file mode 100644
index a65bcdf10..000000000
--- a/paddler_client_javascript/tests/unit/schemas/InferenceNotification.test.ts
+++ /dev/null
@@ -1,18 +0,0 @@
-import { strictEqual, throws } from "node:assert/strict";
-import { test } from "node:test";
-
-import { InferenceNotificationSchema } from "../../../src/schemas/InferenceNotification";
-
-test("parses a cluster token-generation-mode notification frame", function () {
-  const parsed = InferenceNotificationSchema.parse({
-    Notification: "TokenGenerationEnabled",
-  });
-
-  strictEqual(parsed.Notification, "TokenGenerationEnabled");
-});
-
-test("rejects an unknown notification value", function () {
-  throws(function () {
-    InferenceNotificationSchema.parse({ Notification: "SomethingElse" });
-  });
-});
diff --git a/paddler_client_javascript/tests/unit/schemas/InferenceParameters.test.ts b/paddler_client_javascript/tests/unit/schemas/InferenceParameters.test.ts
deleted file mode 100644
index 71a041bfe..000000000
--- a/paddler_client_javascript/tests/unit/schemas/InferenceParameters.test.ts
+++ /dev/null
@@ -1,63 +0,0 @@
-import { strictEqual } from "node:assert/strict";
-import { test } from "node:test";
-
-import {
-  InferenceParametersSchema,
-  type InferenceParameters,
-} from "../../../src/schemas/InferenceParameters";
-import { assertSchemaRejectsPaths } from "../../assertSchemaRejectsPaths";
-
-const validParameters: InferenceParameters = {
-  n_batch: 512,
-  context_size: 4096,
-  embedding_batch_size: 64,
-  enable_embeddings: false,
-  image_resize_to_fit: 512,
-  k_cache_dtype: "F16",
-  v_cache_dtype: "F16",
-  min_p: 0.05,
-  n_gpu_layers: -1,
-  penalty_frequency: 0.5,
-  penalty_last_n: 64,
-  penalty_presence: 0,
-  penalty_repeat: 1.1,
-  pooling_type: "Mean",
-  temperature: 0.7,
-  top_k: 40,
-  top_p: 0.9,
-};
-
-test("accepts parameters that satisfy every rule", function () {
-  strictEqual(
-    InferenceParametersSchema.safeParse(validParameters).success,
-    true,
-  );
-});
-
-test("rejects a batch larger than the context", function () {
-  assertSchemaRejectsPaths({
-    expectedPaths: [["n_batch"]],
-    input: { ...validParameters, n_batch: 8192 },
-    schema: InferenceParametersSchema,
-  });
-});
-
-test("rejects penalty strengths without a penalty window", function () {
-  assertSchemaRejectsPaths({
-    expectedPaths: [["penalty_last_n"]],
-    input: { ...validParameters, penalty_last_n: 0 },
-    schema: InferenceParametersSchema,
-  });
-});
-
-test("rejects a penalty window without penalty strengths", function () {
-  assertSchemaRejectsPaths({
-    expectedPaths: [["penalty_last_n"]],
-    input: {
-      ...validParameters,
-      penalty_frequency: 0,
-      penalty_repeat: 1,
-    },
-    schema: InferenceParametersSchema,
-  });
-});
diff --git a/paddler_client_javascript/tests/unit/schemas/InferenceServiceGenerateTokensResponse.test.ts b/paddler_client_javascript/tests/unit/schemas/InferenceServiceGenerateTokensResponse.test.ts
deleted file mode 100644
index ad1308216..000000000
--- a/paddler_client_javascript/tests/unit/schemas/InferenceServiceGenerateTokensResponse.test.ts
+++ /dev/null
@@ -1,290 +0,0 @@
-import { deepStrictEqual, strictEqual } from "node:assert/strict";
-import { test } from "node:test";
-
-import { InferenceServiceGenerateTokensResponseSchema } from "../../../src/schemas/InferenceServiceGenerateTokensResponse";
-
-function generatedToken(generatedTokenResult: unknown): unknown {
-  return {
-    Response: {
-      generated_by: null,
-      request_id: "req-1",
-      response: { GeneratedToken: generatedTokenResult },
-    },
-  };
-}
-
-const terminalErrorCodes = [
-  { code: 500, variant: "BatchAssemblyFailed" },
-  { code: 500, variant: "ChatTemplateError" },
-  { code: 500, variant: "DecodeFailed" },
-  { code: 500, variant: "DetokenizationFailed" },
-  { code: 400, variant: "GrammarIncompatibleWithThinking" },
-  { code: 500, variant: "GrammarInitializationFailed" },
-  { code: 500, variant: "GrammarRejectedModelOutput" },
-  { code: 400, variant: "GrammarSyntaxError" },
-  { code: 400, variant: "ImageDecodingFailed" },
-  { code: 500, variant: "KvCacheClearFailed" },
-  { code: 500, variant: "MediaMicroBatchCheckFailed" },
-  { code: 503, variant: "ModelNotLoaded" },
-  { code: 500, variant: "MultimodalIngestionFailed" },
-  { code: 400, variant: "MultimodalNotSupported" },
-  { code: 400, variant: "MultimodalTokenizationFailed" },
-  { code: 503, variant: "NoSequenceSlotAvailable" },
-  { code: 400, variant: "PromptTokenizationFailed" },
-  { code: 500, variant: "SamplerChainCreationFailed" },
-  { code: 500, variant: "SamplerError" },
-  { code: 500, variant: "SamplingCandidatesExhausted" },
-  { code: 503, variant: "SchedulerUnavailable" },
-  { code: 500, variant: "SequenceIdOutOfRange" },
-  { code: 501, variant: "TokenGenerationDisabled" },
-  { code: 400, variant: "ToolSchemaInvalid" },
-  { code: 500, variant: "ToolsSerializationFailed" },
-] as const;
-
-for (const { code, variant } of terminalErrorCodes) {
-  test(`${variant} normalises to a terminal error with code ${code}`, function () {
-    const parsed = InferenceServiceGenerateTokensResponseSchema.parse(
-      generatedToken({ [variant]: "described failure" }),
-    );
-
-    strictEqual(parsed.done, true);
-    deepStrictEqual(parsed.error, { code, description: "described failure" });
-  });
-}
-
-const streamingTokenKinds = [
-  { tokenKind: "tool_call", variant: "ToolCallToken" },
-  { tokenKind: "undeterminable", variant: "UndeterminableToken" },
-] as const;
-
-for (const { tokenKind, variant } of streamingTokenKinds) {
-  test(`${variant} normalises into a streaming token with ${tokenKind} kind`, function () {
-    const parsed = InferenceServiceGenerateTokensResponseSchema.parse(
-      generatedToken({ [variant]: "piece" }),
-    );
-
-    strictEqual(parsed.done, false);
-    strictEqual(parsed.token, "piece");
-    strictEqual(parsed.tokenKind, tokenKind);
-  });
-}
-
-test("ToolCallParsed carries the parsed calls without ending the stream", function () {
-  const parsed = InferenceServiceGenerateTokensResponseSchema.parse(
-    generatedToken({
-      ToolCallParsed: [
-        {
-          arguments: { ValidJson: { location: "Paris" } },
-          id: "call_42",
-          name: "get_weather",
-        },
-      ],
-    }),
-  );
-
-  strictEqual(parsed.done, false);
-  deepStrictEqual(parsed.toolCalls, [
-    {
-      arguments: { ValidJson: { location: "Paris" } },
-      id: "call_42",
-      name: "get_weather",
-    },
-  ]);
-});
-
-test("ToolCallParseFailed is a non-terminal unprocessable error", function () {
-  const parsed = InferenceServiceGenerateTokensResponseSchema.parse(
-    generatedToken({ ToolCallParseFailed: "syntax error at 12" }),
-  );
-
-  strictEqual(parsed.done, false);
-  deepStrictEqual(parsed.error, {
-    code: 422,
-    description: "syntax error at 12",
-  });
-});
-
-test("ToolCallValidationFailed joins its errors into a non-terminal unprocessable error", function () {
-  const parsed = InferenceServiceGenerateTokensResponseSchema.parse(
-    generatedToken({
-      ToolCallValidationFailed: [
-        "missing field 'location'",
-        "extra field 'foo'",
-      ],
-    }),
-  );
-
-  strictEqual(parsed.done, false);
-  deepStrictEqual(parsed.error, {
-    code: 422,
-    description: "missing field 'location'; extra field 'foo'",
-  });
-});
-
-test("ContentToken normalises into a streaming token with content kind", function () {
-  const parsed = InferenceServiceGenerateTokensResponseSchema.parse({
-    Response: {
-      generated_by: null,
-      request_id: "req-1",
-      response: { GeneratedToken: { ContentToken: "Hello" } },
-    },
-  });
-
-  strictEqual(parsed.done, false);
-  strictEqual(parsed.error, null);
-  strictEqual(parsed.token, "Hello");
-  strictEqual(parsed.tokenKind, "content");
-  strictEqual(parsed.toolCalls, null);
-});
-
-test("ReasoningToken maps to reasoning kind", function () {
-  const parsed = InferenceServiceGenerateTokensResponseSchema.parse({
-    Response: {
-      generated_by: null,
-      request_id: "req-2",
-      response: { GeneratedToken: { ReasoningToken: "thinking..." } },
-    },
-  });
-
-  strictEqual(parsed.token, "thinking...");
-  strictEqual(parsed.tokenKind, "reasoning");
-});
-
-test("Done normalises with the end reason and the full usage summary", function () {
-  const parsed = InferenceServiceGenerateTokensResponseSchema.parse({
-    Response: {
-      generated_by: null,
-      request_id: "req-3",
-      response: {
-        GeneratedToken: {
-          Done: {
-            finish: "MaxTokens",
-            usage: {
-              prompt_tokens: 10,
-              cached_prompt_tokens: 0,
-              input_image_tokens: 0,
-              input_audio_tokens: 0,
-              content_tokens: 5,
-              reasoning_tokens: 0,
-              tool_call_tokens: 0,
-              undeterminable_tokens: 0,
-            },
-          },
-        },
-      },
-    },
-  });
-
-  strictEqual(parsed.done, true);
-  strictEqual(parsed.error, null);
-  strictEqual(parsed.summary?.finish, "MaxTokens");
-  deepStrictEqual(parsed.summary?.usage.prompt_tokens, 10);
-});
-
-test("Top-level Error envelope normalises to terminal error", function () {
-  const parsed = InferenceServiceGenerateTokensResponseSchema.parse({
-    Error: {
-      request_id: "req-5",
-      error: { code: 500, description: "boom" },
-    },
-  });
-
-  strictEqual(parsed.done, true);
-  deepStrictEqual(parsed.error, { code: 500, description: "boom" });
-});
-
-test("UnrecognizedToolCallFormat preserves text and FFI error message", function () {
-  const parsed = InferenceServiceGenerateTokensResponseSchema.parse({
-    Response: {
-      generated_by: null,
-      request_id: "req-6",
-      response: {
-        GeneratedToken: {
-          UnrecognizedToolCallFormat: {
-            text: "<unknown>raw</unknown>",
-            ffi_error_message: "common_chat_parse failed: no parser",
-          },
-        },
-      },
-    },
-  });
-
-  strictEqual(parsed.done, false);
-  strictEqual(parsed.error, null);
-  strictEqual(parsed.ok, true);
-  strictEqual(parsed.token, null);
-  strictEqual(parsed.tokenKind, null);
-  strictEqual(parsed.toolCalls, null);
-  deepStrictEqual(parsed.rawToolCallTokens, {
-    text: "<unknown>raw</unknown>",
-    ffi_error_message: "common_chat_parse failed: no parser",
-  });
-});
-
-test("PromptExceedsContextSize is terminal and describes token counts", function () {
-  const parsed = InferenceServiceGenerateTokensResponseSchema.parse({
-    Response: {
-      generated_by: null,
-      request_id: "req-8",
-      response: {
-        GeneratedToken: {
-          PromptExceedsContextSize: {
-            prompt_tokens: 9895,
-            sequence_context_size: 8192,
-          },
-        },
-      },
-    },
-  });
-
-  deepStrictEqual(parsed, {
-    done: true,
-    error: {
-      code: 400,
-      description:
-        "prompt has 9895 tokens but each agent sequence holds 8192 tokens",
-    },
-    generated_by: null,
-    ok: false,
-    rawToolCallTokens: null,
-    request_id: "req-8",
-    summary: null,
-    token: null,
-    tokenKind: null,
-    toolCalls: null,
-  });
-});
-
-test("MediaExceedsMicroBatch is terminal and describes token counts", function () {
-  const parsed = InferenceServiceGenerateTokensResponseSchema.parse({
-    Response: {
-      generated_by: null,
-      request_id: "req-7",
-      response: {
-        GeneratedToken: {
-          MediaExceedsMicroBatch: {
-            media_tokens: 368,
-            micro_batch_tokens: 100,
-          },
-        },
-      },
-    },
-  });
-
-  deepStrictEqual(parsed, {
-    done: true,
-    error: {
-      code: 400,
-      description:
-        "media required 368 tokens but one agent micro batch holds 100 tokens",
-    },
-    generated_by: null,
-    ok: false,
-    rawToolCallTokens: null,
-    request_id: "req-7",
-    summary: null,
-    token: null,
-    tokenKind: null,
-    toolCalls: null,
-  });
-});
diff --git a/paddler_client_javascript/tests/unit/schemas/ParsedToolCall.test.ts b/paddler_client_javascript/tests/unit/schemas/ParsedToolCall.test.ts
deleted file mode 100644
index fae6d2338..000000000
--- a/paddler_client_javascript/tests/unit/schemas/ParsedToolCall.test.ts
+++ /dev/null
@@ -1,36 +0,0 @@
-import { deepStrictEqual, strictEqual, throws } from "node:assert/strict";
-import { test } from "node:test";
-
-import { ParsedToolCallSchema } from "../../../src/schemas/ParsedToolCall";
-
-test("ValidJson arguments parse with the inner JSON kept intact", function () {
-  const parsed = ParsedToolCallSchema.parse({
-    id: "call_0",
-    name: "get_weather",
-    arguments: { ValidJson: { location: "Paris" } },
-  });
-
-  strictEqual(parsed.id, "call_0");
-  strictEqual(parsed.name, "get_weather");
-  deepStrictEqual(parsed.arguments, { ValidJson: { location: "Paris" } });
-});
-
-test("InvalidJson arguments preserve the raw string", function () {
-  const parsed = ParsedToolCallSchema.parse({
-    id: "call_1",
-    name: "get_weather",
-    arguments: { InvalidJson: "not json" },
-  });
-
-  deepStrictEqual(parsed.arguments, { InvalidJson: "not json" });
-});
-
-test("rejects payloads missing the discriminated arguments wrapper", function () {
-  throws(function () {
-    ParsedToolCallSchema.parse({
-      id: "call_2",
-      name: "get_weather",
-      arguments: { location: "Paris" },
-    });
-  });
-});
diff --git a/paddler_client_javascript/tests/unit/schemas/agentSchemaParsesAnAgentThatIsDownloadingItsModel.test.ts b/paddler_client_javascript/tests/unit/schemas/agentSchemaParsesAnAgentThatIsDownloadingItsModel.test.ts
new file mode 100644
index 000000000..e75a38fcc
--- /dev/null
+++ b/paddler_client_javascript/tests/unit/schemas/agentSchemaParsesAnAgentThatIsDownloadingItsModel.test.ts
@@ -0,0 +1,40 @@
+import { deepStrictEqual } from "node:assert/strict";
+import { test } from "node:test";
+
+import { AgentSchema } from "../../../src/schemas/Agent";
+
+test("AgentSchema parses an agent that is downloading its model", function () {
+  const parsed = AgentSchema.parse({
+    id: "agent-0",
+    name: "agent-0",
+    slots_processing: 1,
+    status: {
+      desired_slots_total: 4,
+      download_status: {
+        Downloading: {
+          downloaded_bytes: 100,
+          model_path: "https://example.com/qwen.gguf",
+          total_bytes: 400,
+        },
+      },
+      issues: [],
+      model_path: "/models/qwen.gguf",
+      runtime: {
+        Serving: {
+          decisions: "NotServed",
+          slots_total: 4,
+        },
+      },
+      state_application_status: "Applied",
+      uses_chat_template_override: false,
+    },
+  });
+
+  deepStrictEqual(parsed.status.download_status, {
+    Downloading: {
+      downloaded_bytes: 100,
+      model_path: "https://example.com/qwen.gguf",
+      total_bytes: 400,
+    },
+  });
+});
diff --git a/paddler_client_javascript/tests/unit/schemas/agentSchemaRejectsAnUnknownStateApplicationStatus.test.ts b/paddler_client_javascript/tests/unit/schemas/agentSchemaRejectsAnUnknownStateApplicationStatus.test.ts
new file mode 100644
index 000000000..5ce8811f6
--- /dev/null
+++ b/paddler_client_javascript/tests/unit/schemas/agentSchemaRejectsAnUnknownStateApplicationStatus.test.ts
@@ -0,0 +1,16 @@
+import { throws } from "node:assert/strict";
+import { test } from "node:test";
+
+import { AgentSchema } from "../../../src/schemas/Agent";
+import { idleAgent } from "./idleAgent";
+
+test("AgentSchema rejects an unknown state_application_status", function () {
+  const agent = idleAgent({ id: "agent-x", name: null });
+
+  throws(function () {
+    AgentSchema.parse({
+      ...agent,
+      status: { ...agent.status, state_application_status: "Unknown" },
+    });
+  });
+});
diff --git a/paddler_client_javascript/tests/unit/schemas/agentsResponseSchemaOrdersAgentsByName.test.ts b/paddler_client_javascript/tests/unit/schemas/agentsResponseSchemaOrdersAgentsByName.test.ts
new file mode 100644
index 000000000..2f01e85d4
--- /dev/null
+++ b/paddler_client_javascript/tests/unit/schemas/agentsResponseSchemaOrdersAgentsByName.test.ts
@@ -0,0 +1,21 @@
+import { deepStrictEqual } from "node:assert/strict";
+import { test } from "node:test";
+
+import { AgentsResponseSchema } from "../../../src/schemas/AgentsResponse";
+import { idleAgent } from "./idleAgent";
+
+test("AgentsResponseSchema orders agents by name", function () {
+  const { agents } = AgentsResponseSchema.parse({
+    agents: [
+      idleAgent({ id: "second", name: "second" }),
+      idleAgent({ id: "first", name: "first" }),
+    ],
+  });
+
+  deepStrictEqual(
+    agents.map(function ({ name }) {
+      return name;
+    }),
+    ["first", "second"],
+  );
+});
diff --git a/paddler_client_javascript/tests/unit/schemas/agentsResponseSchemaOrdersAnUnnamedAgentByItsIdAmongTheNamedOnes.test.ts b/paddler_client_javascript/tests/unit/schemas/agentsResponseSchemaOrdersAnUnnamedAgentByItsIdAmongTheNamedOnes.test.ts
new file mode 100644
index 000000000..fb7fb2cae
--- /dev/null
+++ b/paddler_client_javascript/tests/unit/schemas/agentsResponseSchemaOrdersAnUnnamedAgentByItsIdAmongTheNamedOnes.test.ts
@@ -0,0 +1,22 @@
+import { deepStrictEqual } from "node:assert/strict";
+import { test } from "node:test";
+
+import { AgentsResponseSchema } from "../../../src/schemas/AgentsResponse";
+import { idleAgent } from "./idleAgent";
+
+test("AgentsResponseSchema orders an unnamed agent by its id among the named ones", function () {
+  const { agents } = AgentsResponseSchema.parse({
+    agents: [
+      idleAgent({ id: "bravo-id", name: "bravo" }),
+      idleAgent({ id: "alpha-id", name: null }),
+      idleAgent({ id: "charlie-id", name: "charlie" }),
+    ],
+  });
+
+  deepStrictEqual(
+    agents.map(function ({ id }) {
+      return id;
+    }),
+    ["alpha-id", "bravo-id", "charlie-id"],
+  );
+});
diff --git a/paddler_client_javascript/tests/unit/schemas/continueFromConversationHistoryParamsSchemaAcceptsToolCallParsingWithADeclaredTool.test.ts b/paddler_client_javascript/tests/unit/schemas/continueFromConversationHistoryParamsSchemaAcceptsToolCallParsingWithADeclaredTool.test.ts
new file mode 100644
index 000000000..e1117b9c8
--- /dev/null
+++ b/paddler_client_javascript/tests/unit/schemas/continueFromConversationHistoryParamsSchemaAcceptsToolCallParsingWithADeclaredTool.test.ts
@@ -0,0 +1,23 @@
+import { strictEqual } from "node:assert/strict";
+import { test } from "node:test";
+
+import { ContinueFromConversationHistoryParamsSchema } from "../../../src/schemas/ContinueFromConversationHistoryParams";
+import { continueFromConversationHistoryParamsWithoutTools } from "./continueFromConversationHistoryParamsWithoutTools";
+
+test("ContinueFromConversationHistoryParamsSchema accepts tool call parsing with a declared tool", function () {
+  strictEqual(
+    ContinueFromConversationHistoryParamsSchema.safeParse({
+      ...continueFromConversationHistoryParamsWithoutTools,
+      tools: [
+        {
+          function: {
+            description: "Get the current weather for a location",
+            name: "get_weather",
+          },
+          type: "function",
+        },
+      ],
+    }).success,
+    true,
+  );
+});
diff --git a/paddler_client_javascript/tests/unit/schemas/continueFromConversationHistoryParamsSchemaRejectsToolCallParsingWithoutTools.test.ts b/paddler_client_javascript/tests/unit/schemas/continueFromConversationHistoryParamsSchemaRejectsToolCallParsingWithoutTools.test.ts
new file mode 100644
index 000000000..a2df0d862
--- /dev/null
+++ b/paddler_client_javascript/tests/unit/schemas/continueFromConversationHistoryParamsSchemaRejectsToolCallParsingWithoutTools.test.ts
@@ -0,0 +1,13 @@
+import { test } from "node:test";
+
+import { ContinueFromConversationHistoryParamsSchema } from "../../../src/schemas/ContinueFromConversationHistoryParams";
+import { assertSchemaRejectsPaths } from "../../assertSchemaRejectsPaths";
+import { continueFromConversationHistoryParamsWithoutTools } from "./continueFromConversationHistoryParamsWithoutTools";
+
+test("ContinueFromConversationHistoryParamsSchema rejects tool call parsing without tools", function () {
+  assertSchemaRejectsPaths({
+    expectedPaths: [["parse_tool_calls"]],
+    input: continueFromConversationHistoryParamsWithoutTools,
+    schema: ContinueFromConversationHistoryParamsSchema,
+  });
+});
diff --git a/paddler_client_javascript/tests/unit/schemas/continueFromConversationHistoryParamsWithoutTools.ts b/paddler_client_javascript/tests/unit/schemas/continueFromConversationHistoryParamsWithoutTools.ts
new file mode 100644
index 000000000..b8c2cfeb9
--- /dev/null
+++ b/paddler_client_javascript/tests/unit/schemas/continueFromConversationHistoryParamsWithoutTools.ts
@@ -0,0 +1,10 @@
+import type { ContinueFromConversationHistoryParams } from "../../../src/schemas/ContinueFromConversationHistoryParams";
+
+export const continueFromConversationHistoryParamsWithoutTools: ContinueFromConversationHistoryParams =
+  {
+    add_generation_prompt: true,
+    conversation_history: [{ Message: { content: "Hello!", role: "user" } }],
+    enable_thinking: false,
+    max_tokens: 100,
+    parse_tool_calls: true,
+  };
diff --git a/paddler_client_javascript/tests/unit/schemas/conversationMessageSchemaAcceptsAToolCallTurnAndItsToolResult.test.ts b/paddler_client_javascript/tests/unit/schemas/conversationMessageSchemaAcceptsAToolCallTurnAndItsToolResult.test.ts
new file mode 100644
index 000000000..94983c741
--- /dev/null
+++ b/paddler_client_javascript/tests/unit/schemas/conversationMessageSchemaAcceptsAToolCallTurnAndItsToolResult.test.ts
@@ -0,0 +1,29 @@
+import { deepEqual } from "node:assert/strict";
+import { test } from "node:test";
+
+import { ConversationMessageSchema } from "../../../src/schemas/ConversationMessage";
+
+test("ConversationMessageSchema accepts a tool call turn and its tool result", function () {
+  const toolCallTurn = {
+    AssistantToolCalls: {
+      content: "",
+      tool_calls: [
+        {
+          arguments: { location: "Paris" },
+          id: "call_weather",
+          name: "get_weather",
+        },
+      ],
+    },
+  };
+  const toolResult = {
+    ToolResult: { content: "sunny", tool_call_id: "call_weather" },
+  };
+
+  deepEqual(
+    [toolCallTurn, toolResult].map(function (message) {
+      return ConversationMessageSchema.parse(message);
+    }),
+    [toolCallTurn, toolResult],
+  );
+});
diff --git a/paddler_client_javascript/tests/unit/schemas/conversationMessageSchemaRejectsAMessageWithoutItsKind.test.ts b/paddler_client_javascript/tests/unit/schemas/conversationMessageSchemaRejectsAMessageWithoutItsKind.test.ts
new file mode 100644
index 000000000..2bb2c38e2
--- /dev/null
+++ b/paddler_client_javascript/tests/unit/schemas/conversationMessageSchemaRejectsAMessageWithoutItsKind.test.ts
@@ -0,0 +1,10 @@
+import { throws } from "node:assert/strict";
+import { test } from "node:test";
+
+import { ConversationMessageSchema } from "../../../src/schemas/ConversationMessage";
+
+test("ConversationMessageSchema rejects a message without its kind", function () {
+  throws(function () {
+    ConversationMessageSchema.parse({ content: "Hello", role: "user" });
+  });
+});
diff --git a/paddler_client_javascript/tests/unit/schemas/decisionResponse.ts b/paddler_client_javascript/tests/unit/schemas/decisionResponse.ts
new file mode 100644
index 000000000..0c9b940cc
--- /dev/null
+++ b/paddler_client_javascript/tests/unit/schemas/decisionResponse.ts
@@ -0,0 +1,9 @@
+export function decisionResponse(decisionResult: unknown): unknown {
+  return {
+    Response: {
+      generated_by: "agent-1",
+      request_id: "req-decision",
+      response: { Decision: decisionResult },
+    },
+  };
+}
diff --git a/paddler_client_javascript/tests/unit/schemas/generatedToken.ts b/paddler_client_javascript/tests/unit/schemas/generatedToken.ts
new file mode 100644
index 000000000..af59f08df
--- /dev/null
+++ b/paddler_client_javascript/tests/unit/schemas/generatedToken.ts
@@ -0,0 +1,9 @@
+export function generatedToken(generatedTokenResult: unknown): unknown {
+  return {
+    Response: {
+      generated_by: null,
+      request_id: "req-1",
+      response: { GeneratedToken: generatedTokenResult },
+    },
+  };
+}
diff --git a/paddler_client_javascript/tests/unit/schemas/generatedToolCallSchemaKeepsTheInnerJsonOfValidJsonArgumentsIntact.test.ts b/paddler_client_javascript/tests/unit/schemas/generatedToolCallSchemaKeepsTheInnerJsonOfValidJsonArgumentsIntact.test.ts
new file mode 100644
index 000000000..9c8110cb9
--- /dev/null
+++ b/paddler_client_javascript/tests/unit/schemas/generatedToolCallSchemaKeepsTheInnerJsonOfValidJsonArgumentsIntact.test.ts
@@ -0,0 +1,17 @@
+import { deepStrictEqual, strictEqual } from "node:assert/strict";
+import { test } from "node:test";
+
+import { GeneratedToolCallSchema } from "../../../src/schemas/GeneratedToolCall";
+
+test("GeneratedToolCallSchema keeps the inner JSON of ValidJson arguments intact", function () {
+  const parsed = GeneratedToolCallSchema.parse({
+    id: "call_0",
+    index: 0,
+    name: "get_weather",
+    arguments: { ValidJson: { location: "Paris" } },
+  });
+
+  strictEqual(parsed.id, "call_0");
+  strictEqual(parsed.name, "get_weather");
+  deepStrictEqual(parsed.arguments, { ValidJson: { location: "Paris" } });
+});
diff --git a/paddler_client_javascript/tests/unit/schemas/generatedToolCallSchemaPreservesTheRawStringOfInvalidJsonArguments.test.ts b/paddler_client_javascript/tests/unit/schemas/generatedToolCallSchemaPreservesTheRawStringOfInvalidJsonArguments.test.ts
new file mode 100644
index 000000000..9d6fef199
--- /dev/null
+++ b/paddler_client_javascript/tests/unit/schemas/generatedToolCallSchemaPreservesTheRawStringOfInvalidJsonArguments.test.ts
@@ -0,0 +1,15 @@
+import { deepStrictEqual } from "node:assert/strict";
+import { test } from "node:test";
+
+import { GeneratedToolCallSchema } from "../../../src/schemas/GeneratedToolCall";
+
+test("GeneratedToolCallSchema preserves the raw string of InvalidJson arguments", function () {
+  const parsed = GeneratedToolCallSchema.parse({
+    id: "call_1",
+    index: 1,
+    name: "get_weather",
+    arguments: { InvalidJson: "not json" },
+  });
+
+  deepStrictEqual(parsed.arguments, { InvalidJson: "not json" });
+});
diff --git a/paddler_client_javascript/tests/unit/schemas/generatedToolCallSchemaRejectsArgumentsMissingTheirDiscriminatedWrapper.test.ts b/paddler_client_javascript/tests/unit/schemas/generatedToolCallSchemaRejectsArgumentsMissingTheirDiscriminatedWrapper.test.ts
new file mode 100644
index 000000000..7fd659074
--- /dev/null
+++ b/paddler_client_javascript/tests/unit/schemas/generatedToolCallSchemaRejectsArgumentsMissingTheirDiscriminatedWrapper.test.ts
@@ -0,0 +1,15 @@
+import { throws } from "node:assert/strict";
+import { test } from "node:test";
+
+import { GeneratedToolCallSchema } from "../../../src/schemas/GeneratedToolCall";
+
+test("GeneratedToolCallSchema rejects arguments missing their discriminated wrapper", function () {
+  throws(function () {
+    GeneratedToolCallSchema.parse({
+      id: "call_2",
+      index: 2,
+      name: "get_weather",
+      arguments: { location: "Paris" },
+    });
+  });
+});
diff --git a/paddler_client_javascript/tests/unit/schemas/idleAgent.ts b/paddler_client_javascript/tests/unit/schemas/idleAgent.ts
new file mode 100644
index 000000000..694d87b60
--- /dev/null
+++ b/paddler_client_javascript/tests/unit/schemas/idleAgent.ts
@@ -0,0 +1,24 @@
+import type { Agent } from "../../../src/schemas/Agent";
+
+export function idleAgent({
+  id,
+  name,
+}: {
+  id: string;
+  name: string | null;
+}): Agent {
+  return {
+    id,
+    name,
+    slots_processing: 0,
+    status: {
+      desired_slots_total: 1,
+      download_status: "NotDownloading",
+      issues: [],
+      model_path: null,
+      runtime: "Idle",
+      state_application_status: "Applied",
+      uses_chat_template_override: false,
+    },
+  };
+}
diff --git a/paddler_client_javascript/tests/unit/schemas/inferenceServiceDecideResponseSchemaCarriesTheSummaryOfAFinishedDecision.test.ts b/paddler_client_javascript/tests/unit/schemas/inferenceServiceDecideResponseSchemaCarriesTheSummaryOfAFinishedDecision.test.ts
new file mode 100644
index 000000000..9fd4d9b58
--- /dev/null
+++ b/paddler_client_javascript/tests/unit/schemas/inferenceServiceDecideResponseSchemaCarriesTheSummaryOfAFinishedDecision.test.ts
@@ -0,0 +1,16 @@
+import { deepStrictEqual } from "node:assert/strict";
+import { test } from "node:test";
+
+import { InferenceServiceDecideResponseSchema } from "../../../src/schemas/InferenceServiceDecideResponse";
+import { decisionResponse } from "./decisionResponse";
+
+test("InferenceServiceDecideResponseSchema carries the summary of a finished decision", function () {
+  deepStrictEqual(
+    InferenceServiceDecideResponseSchema.parse(
+      decisionResponse({
+        Done: { input_tokens: 85, processing_milliseconds: 12 },
+      }),
+    ).summary,
+    { input_tokens: 85, processing_milliseconds: 12 },
+  );
+});
diff --git a/paddler_client_javascript/tests/unit/schemas/inferenceServiceDecideResponseSchemaEndsTheStreamOnAWireError.test.ts b/paddler_client_javascript/tests/unit/schemas/inferenceServiceDecideResponseSchemaEndsTheStreamOnAWireError.test.ts
new file mode 100644
index 000000000..2d2dcf16b
--- /dev/null
+++ b/paddler_client_javascript/tests/unit/schemas/inferenceServiceDecideResponseSchemaEndsTheStreamOnAWireError.test.ts
@@ -0,0 +1,16 @@
+import { deepStrictEqual } from "node:assert/strict";
+import { test } from "node:test";
+
+import { InferenceServiceDecideResponseSchema } from "../../../src/schemas/InferenceServiceDecideResponse";
+
+test("InferenceServiceDecideResponseSchema ends the stream on a wire error", function () {
+  deepStrictEqual(
+    InferenceServiceDecideResponseSchema.parse({
+      Error: {
+        error: { code: 504, description: "timed out" },
+        request_id: "req-decision",
+      },
+    }).error,
+    { code: 504, description: "timed out" },
+  );
+});
diff --git a/paddler_client_javascript/tests/unit/schemas/inferenceServiceDecideResponseSchemaEndsTheStreamWithTheStatusCodeOfEveryDecisionFailure.test.ts b/paddler_client_javascript/tests/unit/schemas/inferenceServiceDecideResponseSchemaEndsTheStreamWithTheStatusCodeOfEveryDecisionFailure.test.ts
new file mode 100644
index 000000000..e5bc24a88
--- /dev/null
+++ b/paddler_client_javascript/tests/unit/schemas/inferenceServiceDecideResponseSchemaEndsTheStreamWithTheStatusCodeOfEveryDecisionFailure.test.ts
@@ -0,0 +1,33 @@
+import { deepStrictEqual, strictEqual } from "node:assert/strict";
+import { test } from "node:test";
+
+import { InferenceServiceDecideResponseSchema } from "../../../src/schemas/InferenceServiceDecideResponse";
+import { decisionResponse } from "./decisionResponse";
+
+const decisionFailureCodes = [
+  { code: 500, variant: "AgentRuntimeFailed" },
+  { code: 500, variant: "BatchAssemblyFailed" },
+  { code: 503, variant: "DecisionsNotServed" },
+  { code: 500, variant: "DecodeFailed" },
+  { code: 500, variant: "HiddenStateUnavailable" },
+  { code: 400, variant: "InputTokenizationFailed" },
+  { code: 500, variant: "KvCacheCopyFailed" },
+  { code: 500, variant: "KvCacheRemovalFailed" },
+  { code: 503, variant: "ModelNotLoaded" },
+  { code: 503, variant: "SchedulerUnavailable" },
+];
+
+test("InferenceServiceDecideResponseSchema ends the stream with the status code of every decision failure", function () {
+  for (const { code, variant } of decisionFailureCodes) {
+    const message = InferenceServiceDecideResponseSchema.parse(
+      decisionResponse({ [variant]: `agent-1: ${variant}` }),
+    );
+
+    strictEqual(message.done, true, variant);
+    deepStrictEqual(
+      message.error,
+      { code, description: `agent-1: ${variant}` },
+      variant,
+    );
+  }
+});
diff --git a/paddler_client_javascript/tests/unit/schemas/inferenceServiceDecideResponseSchemaFailsADecisionExceedingTheContextAsABadRequest.test.ts b/paddler_client_javascript/tests/unit/schemas/inferenceServiceDecideResponseSchemaFailsADecisionExceedingTheContextAsABadRequest.test.ts
new file mode 100644
index 000000000..ff8b4dea9
--- /dev/null
+++ b/paddler_client_javascript/tests/unit/schemas/inferenceServiceDecideResponseSchemaFailsADecisionExceedingTheContextAsABadRequest.test.ts
@@ -0,0 +1,20 @@
+import { deepStrictEqual } from "node:assert/strict";
+import { test } from "node:test";
+
+import { InferenceServiceDecideResponseSchema } from "../../../src/schemas/InferenceServiceDecideResponse";
+import { decisionResponse } from "./decisionResponse";
+
+test("InferenceServiceDecideResponseSchema fails a decision exceeding the context as a bad request", function () {
+  deepStrictEqual(
+    InferenceServiceDecideResponseSchema.parse(
+      decisionResponse({
+        RequestExceedsContext: { context_size: 256, required_tokens: 900 },
+      }),
+    ).error,
+    {
+      code: 400,
+      description:
+        "the decision needs 900 tokens in the cache, more than the 256 the context holds",
+    },
+  );
+});
diff --git a/paddler_client_javascript/tests/unit/schemas/inferenceServiceDecideResponseSchemaKeepsTheStreamOpenAfterAnAnsweredQuestion.test.ts b/paddler_client_javascript/tests/unit/schemas/inferenceServiceDecideResponseSchemaKeepsTheStreamOpenAfterAnAnsweredQuestion.test.ts
new file mode 100644
index 000000000..b492e250a
--- /dev/null
+++ b/paddler_client_javascript/tests/unit/schemas/inferenceServiceDecideResponseSchemaKeepsTheStreamOpenAfterAnAnsweredQuestion.test.ts
@@ -0,0 +1,23 @@
+import { deepStrictEqual } from "node:assert/strict";
+import { test } from "node:test";
+
+import { InferenceServiceDecideResponseSchema } from "../../../src/schemas/InferenceServiceDecideResponse";
+import { decisionResponse } from "./decisionResponse";
+
+test("InferenceServiceDecideResponseSchema keeps the stream open after an answered question", function () {
+  deepStrictEqual(
+    InferenceServiceDecideResponseSchema.parse(
+      decisionResponse({
+        QuestionAnswered: { id: "paid", probabilities: [0.25, 0.75] },
+      }),
+    ),
+    {
+      answer: { id: "paid", probabilities: [0.25, 0.75] },
+      done: false,
+      error: null,
+      generated_by: "agent-1",
+      request_id: "req-decision",
+      summary: null,
+    },
+  );
+});
diff --git a/paddler_client_javascript/tests/unit/schemas/inferenceServiceDecideResponseSchemaRefusesADecisionResultWithMoreThanOneFailure.test.ts b/paddler_client_javascript/tests/unit/schemas/inferenceServiceDecideResponseSchemaRefusesADecisionResultWithMoreThanOneFailure.test.ts
new file mode 100644
index 000000000..d73c5c761
--- /dev/null
+++ b/paddler_client_javascript/tests/unit/schemas/inferenceServiceDecideResponseSchemaRefusesADecisionResultWithMoreThanOneFailure.test.ts
@@ -0,0 +1,14 @@
+import { strictEqual } from "node:assert/strict";
+import { test } from "node:test";
+
+import { InferenceServiceDecideResponseSchema } from "../../../src/schemas/InferenceServiceDecideResponse";
+import { decisionResponse } from "./decisionResponse";
+
+test("InferenceServiceDecideResponseSchema refuses a decision result with more than one failure", function () {
+  strictEqual(
+    InferenceServiceDecideResponseSchema.safeParse(
+      decisionResponse({ DecodeFailed: "first", ModelNotLoaded: "second" }),
+    ).success,
+    false,
+  );
+});
diff --git a/paddler_client_javascript/tests/unit/schemas/inferenceServiceGenerateTokensResponseSchemaCarriesParsedToolCallsWithoutEndingTheStream.test.ts b/paddler_client_javascript/tests/unit/schemas/inferenceServiceGenerateTokensResponseSchemaCarriesParsedToolCallsWithoutEndingTheStream.test.ts
new file mode 100644
index 000000000..58c1de06b
--- /dev/null
+++ b/paddler_client_javascript/tests/unit/schemas/inferenceServiceGenerateTokensResponseSchemaCarriesParsedToolCallsWithoutEndingTheStream.test.ts
@@ -0,0 +1,30 @@
+import { deepStrictEqual, strictEqual } from "node:assert/strict";
+import { test } from "node:test";
+
+import { InferenceServiceGenerateTokensResponseSchema } from "../../../src/schemas/InferenceServiceGenerateTokensResponse";
+import { generatedToken } from "./generatedToken";
+
+test("InferenceServiceGenerateTokensResponseSchema carries parsed tool calls without ending the stream", function () {
+  const parsed = InferenceServiceGenerateTokensResponseSchema.parse(
+    generatedToken({
+      ToolCallParsed: [
+        {
+          arguments: { ValidJson: { location: "Paris" } },
+          id: "call_42",
+          index: 42,
+          name: "get_weather",
+        },
+      ],
+    }),
+  );
+
+  strictEqual(parsed.done, false);
+  deepStrictEqual(parsed.toolCalls, [
+    {
+      arguments: { ValidJson: { location: "Paris" } },
+      id: "call_42",
+      index: 42,
+      name: "get_weather",
+    },
+  ]);
+});
diff --git a/paddler_client_javascript/tests/unit/schemas/inferenceServiceGenerateTokensResponseSchemaEndsTheStreamOnAWireError.test.ts b/paddler_client_javascript/tests/unit/schemas/inferenceServiceGenerateTokensResponseSchemaEndsTheStreamOnAWireError.test.ts
new file mode 100644
index 000000000..bf81f43fb
--- /dev/null
+++ b/paddler_client_javascript/tests/unit/schemas/inferenceServiceGenerateTokensResponseSchemaEndsTheStreamOnAWireError.test.ts
@@ -0,0 +1,16 @@
+import { deepStrictEqual, strictEqual } from "node:assert/strict";
+import { test } from "node:test";
+
+import { InferenceServiceGenerateTokensResponseSchema } from "../../../src/schemas/InferenceServiceGenerateTokensResponse";
+
+test("InferenceServiceGenerateTokensResponseSchema ends the stream on a wire error", function () {
+  const parsed = InferenceServiceGenerateTokensResponseSchema.parse({
+    Error: {
+      request_id: "req-5",
+      error: { code: 500, description: "boom" },
+    },
+  });
+
+  strictEqual(parsed.done, true);
+  deepStrictEqual(parsed.error, { code: 500, description: "boom" });
+});
diff --git a/paddler_client_javascript/tests/unit/schemas/inferenceServiceGenerateTokensResponseSchemaEndsTheStreamWithTheFinishReasonAndTheUsageSummary.test.ts b/paddler_client_javascript/tests/unit/schemas/inferenceServiceGenerateTokensResponseSchemaEndsTheStreamWithTheFinishReasonAndTheUsageSummary.test.ts
new file mode 100644
index 000000000..e0f374c32
--- /dev/null
+++ b/paddler_client_javascript/tests/unit/schemas/inferenceServiceGenerateTokensResponseSchemaEndsTheStreamWithTheFinishReasonAndTheUsageSummary.test.ts
@@ -0,0 +1,30 @@
+import { deepStrictEqual, strictEqual } from "node:assert/strict";
+import { test } from "node:test";
+
+import { InferenceServiceGenerateTokensResponseSchema } from "../../../src/schemas/InferenceServiceGenerateTokensResponse";
+import { generatedToken } from "./generatedToken";
+
+test("InferenceServiceGenerateTokensResponseSchema ends the stream with the finish reason and the usage summary", function () {
+  const parsed = InferenceServiceGenerateTokensResponseSchema.parse(
+    generatedToken({
+      Done: {
+        finish: "MaxTokens",
+        usage: {
+          prompt_tokens: 10,
+          cached_prompt_tokens: 0,
+          input_image_tokens: 0,
+          input_audio_tokens: 0,
+          content_tokens: 5,
+          reasoning_tokens: 0,
+          tool_call_tokens: 0,
+          undeterminable_tokens: 0,
+        },
+      },
+    }),
+  );
+
+  strictEqual(parsed.done, true);
+  strictEqual(parsed.error, null);
+  strictEqual(parsed.summary?.finish, "MaxTokens");
+  deepStrictEqual(parsed.summary?.usage.prompt_tokens, 10);
+});
diff --git a/paddler_client_javascript/tests/unit/schemas/inferenceServiceGenerateTokensResponseSchemaEndsTheStreamWithTheStatusCodeOfEveryTerminalFailure.test.ts b/paddler_client_javascript/tests/unit/schemas/inferenceServiceGenerateTokensResponseSchemaEndsTheStreamWithTheStatusCodeOfEveryTerminalFailure.test.ts
new file mode 100644
index 000000000..3b7ec80dc
--- /dev/null
+++ b/paddler_client_javascript/tests/unit/schemas/inferenceServiceGenerateTokensResponseSchemaEndsTheStreamWithTheStatusCodeOfEveryTerminalFailure.test.ts
@@ -0,0 +1,46 @@
+import { deepStrictEqual, strictEqual } from "node:assert/strict";
+import { test } from "node:test";
+
+import { InferenceServiceGenerateTokensResponseSchema } from "../../../src/schemas/InferenceServiceGenerateTokensResponse";
+import { generatedToken } from "./generatedToken";
+
+const terminalErrorCodes = [
+  { code: 500, variant: "BatchAssemblyFailed" },
+  { code: 500, variant: "ChatTemplateError" },
+  { code: 500, variant: "DecodeFailed" },
+  { code: 500, variant: "DetokenizationFailed" },
+  { code: 400, variant: "GrammarIncompatibleWithThinking" },
+  { code: 500, variant: "GrammarInitializationFailed" },
+  { code: 500, variant: "GrammarRejectedModelOutput" },
+  { code: 400, variant: "GrammarSyntaxError" },
+  { code: 400, variant: "ImageDecodingFailed" },
+  { code: 500, variant: "KvCacheClearFailed" },
+  { code: 500, variant: "MediaMicroBatchCheckFailed" },
+  { code: 503, variant: "ModelNotLoaded" },
+  { code: 500, variant: "MultimodalIngestionFailed" },
+  { code: 400, variant: "MultimodalNotSupported" },
+  { code: 400, variant: "MultimodalTokenizationFailed" },
+  { code: 503, variant: "NoSequenceSlotAvailable" },
+  { code: 400, variant: "PromptTokenizationFailed" },
+  { code: 400, variant: "PromptTokenizedToNothing" },
+  { code: 500, variant: "SamplerChainCreationFailed" },
+  { code: 500, variant: "SamplerError" },
+  { code: 500, variant: "SamplingCandidatesExhausted" },
+  { code: 503, variant: "SchedulerUnavailable" },
+  { code: 400, variant: "ToolSchemaInvalid" },
+] as const;
+
+test("InferenceServiceGenerateTokensResponseSchema ends the stream with the status code of every terminal failure", function () {
+  for (const { code, variant } of terminalErrorCodes) {
+    const parsed = InferenceServiceGenerateTokensResponseSchema.parse(
+      generatedToken({ [variant]: "described failure" }),
+    );
+
+    strictEqual(parsed.done, true, variant);
+    deepStrictEqual(
+      parsed.error,
+      { code, description: "described failure" },
+      variant,
+    );
+  }
+});
diff --git a/paddler_client_javascript/tests/unit/schemas/inferenceServiceGenerateTokensResponseSchemaJoinsToolCallValidationErrorsIntoANonTerminalUnprocessableError.test.ts b/paddler_client_javascript/tests/unit/schemas/inferenceServiceGenerateTokensResponseSchemaJoinsToolCallValidationErrorsIntoANonTerminalUnprocessableError.test.ts
new file mode 100644
index 000000000..ab4666365
--- /dev/null
+++ b/paddler_client_javascript/tests/unit/schemas/inferenceServiceGenerateTokensResponseSchemaJoinsToolCallValidationErrorsIntoANonTerminalUnprocessableError.test.ts
@@ -0,0 +1,22 @@
+import { deepStrictEqual, strictEqual } from "node:assert/strict";
+import { test } from "node:test";
+
+import { InferenceServiceGenerateTokensResponseSchema } from "../../../src/schemas/InferenceServiceGenerateTokensResponse";
+import { generatedToken } from "./generatedToken";
+
+test("InferenceServiceGenerateTokensResponseSchema joins tool call validation errors into a non-terminal unprocessable error", function () {
+  const parsed = InferenceServiceGenerateTokensResponseSchema.parse(
+    generatedToken({
+      ToolCallValidationFailed: [
+        "missing field 'location'",
+        "extra field 'foo'",
+      ],
+    }),
+  );
+
+  strictEqual(parsed.done, false);
+  deepStrictEqual(parsed.error, {
+    code: 422,
+    description: "missing field 'location'; extra field 'foo'",
+  });
+});
diff --git a/paddler_client_javascript/tests/unit/schemas/inferenceServiceGenerateTokensResponseSchemaPreservesTheTextAndFfiErrorOfAnUnrecognizedToolCallFormat.test.ts b/paddler_client_javascript/tests/unit/schemas/inferenceServiceGenerateTokensResponseSchemaPreservesTheTextAndFfiErrorOfAnUnrecognizedToolCallFormat.test.ts
new file mode 100644
index 000000000..7df0e0a95
--- /dev/null
+++ b/paddler_client_javascript/tests/unit/schemas/inferenceServiceGenerateTokensResponseSchemaPreservesTheTextAndFfiErrorOfAnUnrecognizedToolCallFormat.test.ts
@@ -0,0 +1,27 @@
+import { deepStrictEqual, strictEqual } from "node:assert/strict";
+import { test } from "node:test";
+
+import { InferenceServiceGenerateTokensResponseSchema } from "../../../src/schemas/InferenceServiceGenerateTokensResponse";
+import { generatedToken } from "./generatedToken";
+
+test("InferenceServiceGenerateTokensResponseSchema preserves the text and FFI error of an unrecognized tool call format", function () {
+  const parsed = InferenceServiceGenerateTokensResponseSchema.parse(
+    generatedToken({
+      UnrecognizedToolCallFormat: {
+        text: "<unknown>raw</unknown>",
+        ffi_error_message: "common_chat_parse failed: no parser",
+      },
+    }),
+  );
+
+  strictEqual(parsed.done, false);
+  strictEqual(parsed.error, null);
+  strictEqual(parsed.ok, true);
+  strictEqual(parsed.token, null);
+  strictEqual(parsed.tokenKind, null);
+  strictEqual(parsed.toolCalls, null);
+  deepStrictEqual(parsed.rawToolCallTokens, {
+    text: "<unknown>raw</unknown>",
+    ffi_error_message: "common_chat_parse failed: no parser",
+  });
+});
diff --git a/paddler_client_javascript/tests/unit/schemas/inferenceServiceGenerateTokensResponseSchemaReportsAToolCallParseFailureAsANonTerminalUnprocessableError.test.ts b/paddler_client_javascript/tests/unit/schemas/inferenceServiceGenerateTokensResponseSchemaReportsAToolCallParseFailureAsANonTerminalUnprocessableError.test.ts
new file mode 100644
index 000000000..3dd6d2dd6
--- /dev/null
+++ b/paddler_client_javascript/tests/unit/schemas/inferenceServiceGenerateTokensResponseSchemaReportsAToolCallParseFailureAsANonTerminalUnprocessableError.test.ts
@@ -0,0 +1,17 @@
+import { deepStrictEqual, strictEqual } from "node:assert/strict";
+import { test } from "node:test";
+
+import { InferenceServiceGenerateTokensResponseSchema } from "../../../src/schemas/InferenceServiceGenerateTokensResponse";
+import { generatedToken } from "./generatedToken";
+
+test("InferenceServiceGenerateTokensResponseSchema reports a tool call parse failure as a non-terminal unprocessable error", function () {
+  const parsed = InferenceServiceGenerateTokensResponseSchema.parse(
+    generatedToken({ ToolCallParseFailed: "syntax error at 12" }),
+  );
+
+  strictEqual(parsed.done, false);
+  deepStrictEqual(parsed.error, {
+    code: 422,
+    description: "syntax error at 12",
+  });
+});
diff --git a/paddler_client_javascript/tests/unit/schemas/inferenceServiceGenerateTokensResponseSchemaStreamsEveryTokenVariantWithItsKind.test.ts b/paddler_client_javascript/tests/unit/schemas/inferenceServiceGenerateTokensResponseSchemaStreamsEveryTokenVariantWithItsKind.test.ts
new file mode 100644
index 000000000..42e5819eb
--- /dev/null
+++ b/paddler_client_javascript/tests/unit/schemas/inferenceServiceGenerateTokensResponseSchemaStreamsEveryTokenVariantWithItsKind.test.ts
@@ -0,0 +1,30 @@
+import { strictEqual } from "node:assert/strict";
+import { test } from "node:test";
+
+import { InferenceServiceGenerateTokensResponseSchema } from "../../../src/schemas/InferenceServiceGenerateTokensResponse";
+import { generatedToken } from "./generatedToken";
+
+const streamingTokenKinds = [
+  { token: "Hello", tokenKind: "content", variant: "ContentToken" },
+  { token: "thinking...", tokenKind: "reasoning", variant: "ReasoningToken" },
+  { token: "piece", tokenKind: "tool_call", variant: "ToolCallToken" },
+  {
+    token: "piece",
+    tokenKind: "undeterminable",
+    variant: "UndeterminableToken",
+  },
+] as const;
+
+test("InferenceServiceGenerateTokensResponseSchema streams every token variant with its kind", function () {
+  for (const { token, tokenKind, variant } of streamingTokenKinds) {
+    const parsed = InferenceServiceGenerateTokensResponseSchema.parse(
+      generatedToken({ [variant]: token }),
+    );
+
+    strictEqual(parsed.done, false, variant);
+    strictEqual(parsed.error, null, variant);
+    strictEqual(parsed.token, token, variant);
+    strictEqual(parsed.tokenKind, tokenKind, variant);
+    strictEqual(parsed.toolCalls, null, variant);
+  }
+});
diff --git a/paddler_client_javascript/tests/unit/schemas/inferenceServiceGenerateTokensResponseSchemaTerminatesAPromptExceedingTheContextWithItsTokenCounts.test.ts b/paddler_client_javascript/tests/unit/schemas/inferenceServiceGenerateTokensResponseSchemaTerminatesAPromptExceedingTheContextWithItsTokenCounts.test.ts
new file mode 100644
index 000000000..853b12fe9
--- /dev/null
+++ b/paddler_client_javascript/tests/unit/schemas/inferenceServiceGenerateTokensResponseSchemaTerminatesAPromptExceedingTheContextWithItsTokenCounts.test.ts
@@ -0,0 +1,33 @@
+import { deepStrictEqual } from "node:assert/strict";
+import { test } from "node:test";
+
+import { InferenceServiceGenerateTokensResponseSchema } from "../../../src/schemas/InferenceServiceGenerateTokensResponse";
+import { generatedToken } from "./generatedToken";
+
+test("InferenceServiceGenerateTokensResponseSchema terminates a prompt exceeding the context with its token counts", function () {
+  const parsed = InferenceServiceGenerateTokensResponseSchema.parse(
+    generatedToken({
+      PromptExceedsContextSize: {
+        prompt_tokens: 9895,
+        sequence_context_size: 8192,
+      },
+    }),
+  );
+
+  deepStrictEqual(parsed, {
+    done: true,
+    error: {
+      code: 400,
+      description:
+        "prompt has 9895 tokens but each agent sequence holds 8192 tokens",
+    },
+    generated_by: null,
+    ok: false,
+    rawToolCallTokens: null,
+    request_id: "req-1",
+    summary: null,
+    token: null,
+    tokenKind: null,
+    toolCalls: null,
+  });
+});
diff --git a/paddler_client_javascript/tests/unit/schemas/inferenceServiceGenerateTokensResponseSchemaTerminatesMediaExceedingTheMicroBatchWithItsTokenCounts.test.ts b/paddler_client_javascript/tests/unit/schemas/inferenceServiceGenerateTokensResponseSchemaTerminatesMediaExceedingTheMicroBatchWithItsTokenCounts.test.ts
new file mode 100644
index 000000000..54ff1b897
--- /dev/null
+++ b/paddler_client_javascript/tests/unit/schemas/inferenceServiceGenerateTokensResponseSchemaTerminatesMediaExceedingTheMicroBatchWithItsTokenCounts.test.ts
@@ -0,0 +1,33 @@
+import { deepStrictEqual } from "node:assert/strict";
+import { test } from "node:test";
+
+import { InferenceServiceGenerateTokensResponseSchema } from "../../../src/schemas/InferenceServiceGenerateTokensResponse";
+import { generatedToken } from "./generatedToken";
+
+test("InferenceServiceGenerateTokensResponseSchema terminates media exceeding the micro batch with its token counts", function () {
+  const parsed = InferenceServiceGenerateTokensResponseSchema.parse(
+    generatedToken({
+      MediaExceedsMicroBatch: {
+        media_tokens: 368,
+        micro_batch_tokens: 100,
+      },
+    }),
+  );
+
+  deepStrictEqual(parsed, {
+    done: true,
+    error: {
+      code: 400,
+      description:
+        "media required 368 tokens but one agent micro batch holds 100 tokens",
+    },
+    generated_by: null,
+    ok: false,
+    rawToolCallTokens: null,
+    request_id: "req-1",
+    summary: null,
+    token: null,
+    tokenKind: null,
+    toolCalls: null,
+  });
+});
diff --git a/paddler_client_javascript/tests/unit/urlToAgentDesiredModel.test.ts b/paddler_client_javascript/tests/unit/urlToAgentDesiredModel.test.ts
deleted file mode 100644
index b323b761d..000000000
--- a/paddler_client_javascript/tests/unit/urlToAgentDesiredModel.test.ts
+++ /dev/null
@@ -1,65 +0,0 @@
-import { deepStrictEqual, throws } from "node:assert/strict";
-import { test } from "node:test";
-
-import { UnsupportedModelUrlError } from "../../src/UnsupportedModelUrlError";
-import { urlToAgentDesiredModel } from "../../src/urlToAgentDesiredModel";
-
-test("recognizes Hugging Face URLs as HuggingFace variant", function () {
-  const url = new URL(
-    "https://huggingface.co/Qwen/Qwen3-0.6B-GGUF/blob/main/Qwen3-0.6B-Q8_0.gguf",
-  );
-
-  deepStrictEqual(urlToAgentDesiredModel(url), {
-    HuggingFace: {
-      filename: "Qwen3-0.6B-Q8_0.gguf",
-      repo_id: "Qwen/Qwen3-0.6B-GGUF",
-      revision: "main",
-    },
-  });
-});
-
-test("agent: URLs become LocalToAgent variant", function () {
-  const url = new URL("agent:///home/user/models/Qwen3-0.6B-Q8_0.gguf");
-
-  deepStrictEqual(urlToAgentDesiredModel(url), {
-    LocalToAgent: "/home/user/models/Qwen3-0.6B-Q8_0.gguf",
-  });
-});
-
-test("non-http(s), non-agent URLs throw", function () {
-  const url = new URL("ftp://example.com/file.gguf");
-
-  throws(function () {
-    urlToAgentDesiredModel(url);
-  }, UnsupportedModelUrlError);
-});
-
-test("blob URLs whose repository and file names contain dots route to HuggingFace", function () {
-  const url = new URL(
-    "https://huggingface.co/unsloth/Qwen3.6-35B-A3B-GGUF/blob/main/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf",
-  );
-
-  deepStrictEqual(urlToAgentDesiredModel(url), {
-    HuggingFace: {
-      filename: "Qwen3.6-35B-A3B-UD-Q4_K_M.gguf",
-      repo_id: "unsloth/Qwen3.6-35B-A3B-GGUF",
-      revision: "main",
-    },
-  });
-});
-
-test("https URLs off huggingface.co route to the Url variant", function () {
-  const url = new URL("https://example.com/path/to/model.gguf");
-
-  deepStrictEqual(urlToAgentDesiredModel(url), {
-    Url: { url: "https://example.com/path/to/model.gguf" },
-  });
-});
-
-test("plain http URLs route to the Url variant", function () {
-  const url = new URL("http://mirror.example.org/Qwen3-0.6B.gguf");
-
-  deepStrictEqual(urlToAgentDesiredModel(url), {
-    Url: { url: "http://mirror.example.org/Qwen3-0.6B.gguf" },
-  });
-});
diff --git a/paddler_client_javascript/tests/unit/webSocketProtocol.test.ts b/paddler_client_javascript/tests/unit/webSocketProtocol.test.ts
deleted file mode 100644
index fdcbb8a28..000000000
--- a/paddler_client_javascript/tests/unit/webSocketProtocol.test.ts
+++ /dev/null
@@ -1,16 +0,0 @@
-import { strictEqual } from "node:assert/strict";
-import { test } from "node:test";
-
-import { webSocketProtocol } from "../../src/webSocketProtocol";
-
-test("https: maps to wss:", function () {
-  strictEqual(webSocketProtocol("https:"), "wss:");
-});
-
-test("http: maps to ws:", function () {
-  strictEqual(webSocketProtocol("http:"), "ws:");
-});
-
-test("anything other than https: maps to ws:", function () {
-  strictEqual(webSocketProtocol("file:"), "ws:");
-});
diff --git a/paddler_client_javascript/tests/unit/webSocketProtocolMapsHttpsToWssAndEveryOtherProtocolToWs.test.ts b/paddler_client_javascript/tests/unit/webSocketProtocolMapsHttpsToWssAndEveryOtherProtocolToWs.test.ts
new file mode 100644
index 000000000..05e1c4e00
--- /dev/null
+++ b/paddler_client_javascript/tests/unit/webSocketProtocolMapsHttpsToWssAndEveryOtherProtocolToWs.test.ts
@@ -0,0 +1,20 @@
+import { strictEqual } from "node:assert/strict";
+import { test } from "node:test";
+
+import { webSocketProtocol } from "../../src/webSocketProtocol";
+
+const protocolMappings = [
+  { expectedWebSocketProtocol: "wss:", httpProtocol: "https:" },
+  { expectedWebSocketProtocol: "ws:", httpProtocol: "http:" },
+  { expectedWebSocketProtocol: "ws:", httpProtocol: "file:" },
+] as const;
+
+test("webSocketProtocol maps https: to wss: and every other protocol to ws:", function () {
+  for (const { expectedWebSocketProtocol, httpProtocol } of protocolMappings) {
+    strictEqual(
+      webSocketProtocol(httpProtocol),
+      expectedWebSocketProtocol,
+      httpProtocol,
+    );
+  }
+});
diff --git a/paddler_client_javascript/tests/waitForAgentReady.ts b/paddler_client_javascript/tests/waitForAgentReady.ts
index 9525291ca..f60604dff 100644
--- a/paddler_client_javascript/tests/waitForAgentReady.ts
+++ b/paddler_client_javascript/tests/waitForAgentReady.ts
@@ -2,6 +2,9 @@ import {
   filter,
   first,
   firstValueFrom,
+  from,
+  map,
+  merge,
   mergeMap,
   tap,
   type Observable,
@@ -9,33 +12,50 @@ import {
 
 import type { EventSourceState } from "../src/EventSourceState";
 import type { Agent } from "../src/schemas/Agent";
+import type { AgentRuntimeStatus } from "../src/schemas/AgentRuntimeStatus";
 import type { AgentsResponseSchema } from "../src/schemas/AgentsResponse";
+import { AgentExitedBeforeReadyError } from "./AgentExitedBeforeReadyError";
 import { AgentReportedIssuesError } from "./AgentReportedIssuesError";
 import { AgentsSnapshotUndeserializableError } from "./AgentsSnapshotUndeserializableError";
 import { AgentsStreamClosedError } from "./AgentsStreamClosedError";
+import type { SpawnedProcess } from "./SpawnedProcess";
+
+function slotsTotalOf(runtime: AgentRuntimeStatus): number {
+  return runtime === "Idle" ? 0 : runtime.Serving.slots_total;
+}
 
 export function waitForAgentReady({
   agentName,
+  agentProcess,
   agentsStates,
   expectedSlotsTotal,
 }: {
   agentName: string;
+  agentProcess: SpawnedProcess;
   agentsStates: Observable<EventSourceState<typeof AgentsResponseSchema>>;
   expectedSlotsTotal: number;
 }): Promise<Agent> {
   return firstValueFrom(
-    agentsStates.pipe(
-      mergeMap(function (agentsState) {
-        if (agentsState.isConnectionError) {
-          throw new AgentsStreamClosedError(agentName);
-        }
+    merge(
+      agentsStates.pipe(
+        mergeMap(function (agentsState) {
+          if (agentsState.isConnectionError) {
+            throw new AgentsStreamClosedError(agentName);
+          }
 
-        if (agentsState.isDeserializationError) {
-          throw new AgentsSnapshotUndeserializableError(agentName);
-        }
+          if (agentsState.isDeserializationError) {
+            throw new AgentsSnapshotUndeserializableError(agentName);
+          }
 
-        return agentsState.isOk ? agentsState.data.agents : [];
-      }),
+          return agentsState.isOk ? agentsState.data.agents : [];
+        }),
+      ),
+      from(agentProcess.exitCode()).pipe(
+        map(function (exitCode): never {
+          throw new AgentExitedBeforeReadyError(agentName, exitCode);
+        }),
+      ),
+    ).pipe(
       filter(function (agent) {
         return agent.name === agentName;
       }),
@@ -47,7 +67,7 @@ export function waitForAgentReady({
       first(function (agent) {
         return (
           agent.status.state_application_status === "Applied" &&
-          agent.status.slots_total === expectedSlotsTotal
+          slotsTotalOf(agent.status.runtime) === expectedSlotsTotal
         );
       }),
     ),
diff --git a/paddler_client_javascript/tests/withConnectionClosingListener.ts b/paddler_client_javascript/tests/withConnectionClosingListener.ts
deleted file mode 100644
index de8ee5936..000000000
--- a/paddler_client_javascript/tests/withConnectionClosingListener.ts
+++ /dev/null
@@ -1,25 +0,0 @@
-import { ok } from "node:assert/strict";
-import { once } from "node:events";
-import { createServer } from "node:net";
-
-export async function withConnectionClosingListener<TResult>(
-  body: (address: string) => Promise<TResult>,
-): Promise<TResult> {
-  const server = createServer(function (socket) {
-    socket.destroy();
-  });
-
-  server.listen(0, "127.0.0.1");
-
-  await once(server, "listening");
-
-  const listenedAddress = server.address();
-
-  try {
-    ok(typeof listenedAddress === "object" && listenedAddress !== null);
-
-    return await body(`${listenedAddress.address}:${listenedAddress.port}`);
-  } finally {
-    server.close();
-  }
-}
diff --git a/paddler_client_javascript/tests/withPaddlerCluster.ts b/paddler_client_javascript/tests/withPaddlerCluster.ts
index 80a809d36..000b15b0f 100644
--- a/paddler_client_javascript/tests/withPaddlerCluster.ts
+++ b/paddler_client_javascript/tests/withPaddlerCluster.ts
@@ -6,6 +6,7 @@ import {
 } from "../src/schemas/BalancerDesiredState";
 import { streamEventSource } from "../src/streamEventSource";
 import type { AgentSpec } from "./AgentSpec";
+import { agentsStreamUrl } from "./agentsStreamUrl";
 import type { BalancerAddresses } from "./BalancerAddresses";
 import { putBalancerDesiredState } from "./putBalancerDesiredState";
 import { spawnAgent } from "./spawnAgent";
@@ -47,13 +48,16 @@ export function withPaddlerCluster<TResult>(
         await putBalancerDesiredState({ desiredState, managementAddress });
 
         for (const agent of agents) {
-          agentProcesses.push(spawnAgent({ agent, managementAddress }));
+          const agentProcess = spawnAgent({ agent, managementAddress });
+
+          agentProcesses.push(agentProcess);
 
           await waitForAgentReady({
             agentName: agent.name,
+            agentProcess,
             agentsStates: streamEventSource({
               schema: AgentsResponseSchema,
-              url: `http://${managementAddress}/api/v1/agents/stream`,
+              url: agentsStreamUrl(managementAddress),
             }),
             expectedSlotsTotal: desiredState.model === "None" ? 0 : agent.slots,
           });
diff --git a/paddler_client_javascript/tests/withReadyModelLessAgent.ts b/paddler_client_javascript/tests/withReadyModelLessAgent.ts
new file mode 100644
index 000000000..4b8c2275b
--- /dev/null
+++ b/paddler_client_javascript/tests/withReadyModelLessAgent.ts
@@ -0,0 +1,38 @@
+import { AgentsResponseSchema } from "../src/schemas/AgentsResponse";
+import { streamEventSource } from "../src/streamEventSource";
+import type { AgentSpec } from "./AgentSpec";
+import { agentsStreamUrl } from "./agentsStreamUrl";
+import { spawnAgent } from "./spawnAgent";
+import type { SpawnedProcess } from "./SpawnedProcess";
+import { waitForAgentReady } from "./waitForAgentReady";
+
+const MODEL_LESS_SLOTS_TOTAL = 0;
+
+export async function withReadyModelLessAgent<TResult>(
+  {
+    agent,
+    managementAddress,
+  }: {
+    agent: AgentSpec;
+    managementAddress: string;
+  },
+  body: (agentProcess: SpawnedProcess) => Promise<TResult>,
+): Promise<TResult> {
+  const agentProcess = spawnAgent({ agent, managementAddress });
+
+  try {
+    await waitForAgentReady({
+      agentName: agent.name,
+      agentProcess,
+      agentsStates: streamEventSource({
+        schema: AgentsResponseSchema,
+        url: agentsStreamUrl(managementAddress),
+      }),
+      expectedSlotsTotal: MODEL_LESS_SLOTS_TOTAL,
+    });
+
+    return await body(agentProcess);
+  } finally {
+    await agentProcess.terminate();
+  }
+}
diff --git a/paddler_client_javascript/tsconfig.test.json b/paddler_client_javascript/tsconfig.test.json
index 908772ecf..66d2f91bb 100644
--- a/paddler_client_javascript/tsconfig.test.json
+++ b/paddler_client_javascript/tsconfig.test.json
@@ -2,7 +2,8 @@
   "extends": "./tsconfig.json",
   "compilerOptions": {
     "noEmit": true,
-    "rootDir": "."
+    "rootDir": ".",
+    "types": ["node"]
   },
   "include": ["src/**/*", "tests/**/*"]
 }
diff --git a/paddler_client_python/.gitignore b/paddler_client_python/.gitignore
deleted file mode 100644
index 4cf161822..000000000
--- a/paddler_client_python/.gitignore
+++ /dev/null
@@ -1,21 +0,0 @@
-__pycache__/
-*.py[cod]
-*$py.class
-*.so
-dist/
-build/
-*.egg-info/
-*.egg
-.eggs/
-*.whl
-.venv
-.venv/
-venv/
-.mypy_cache/
-.pytest_cache/
-.ruff_cache/
-*.cover
-*.log
-.coverage
-htmlcov/
-target/
diff --git a/paddler_client_python/LICENSE b/paddler_client_python/LICENSE
deleted file mode 100644
index 261eeb9e9..000000000
--- a/paddler_client_python/LICENSE
+++ /dev/null
@@ -1,201 +0,0 @@
-                                 Apache License
-                           Version 2.0, January 2004
-                        http://www.apache.org/licenses/
-
-   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
-
-   1. Definitions.
-
-      "License" shall mean the terms and conditions for use, reproduction,
-      and distribution as defined by Sections 1 through 9 of this document.
-
-      "Licensor" shall mean the copyright owner or entity authorized by
-      the copyright owner that is granting the License.
-
-      "Legal Entity" shall mean the union of the acting entity and all
-      other entities that control, are controlled by, or are under common
-      control with that entity. For the purposes of this definition,
-      "control" means (i) the power, direct or indirect, to cause the
-      direction or management of such entity, whether by contract or
-      otherwise, or (ii) ownership of fifty percent (50%) or more of the
-      outstanding shares, or (iii) beneficial ownership of such entity.
-
-      "You" (or "Your") shall mean an individual or Legal Entity
-      exercising permissions granted by this License.
-
-      "Source" form shall mean the preferred form for making modifications,
-      including but not limited to software source code, documentation
-      source, and configuration files.
-
-      "Object" form shall mean any form resulting from mechanical
-      transformation or translation of a Source form, including but
-      not limited to compiled object code, generated documentation,
-      and conversions to other media types.
-
-      "Work" shall mean the work of authorship, whether in Source or
-      Object form, made available under the License, as indicated by a
-      copyright notice that is included in or attached to the work
-      (an example is provided in the Appendix below).
-
-      "Derivative Works" shall mean any work, whether in Source or Object
-      form, that is based on (or derived from) the Work and for which the
-      editorial revisions, annotations, elaborations, or other modifications
-      represent, as a whole, an original work of authorship. For the purposes
-      of this License, Derivative Works shall not include works that remain
-      separable from, or merely link (or bind by name) to the interfaces of,
-      the Work and Derivative Works thereof.
-
-      "Contribution" shall mean any work of authorship, including
-      the original version of the Work and any modifications or additions
-      to that Work or Derivative Works thereof, that is intentionally
-      submitted to Licensor for inclusion in the Work by the copyright owner
-      or by an individual or Legal Entity authorized to submit on behalf of
-      the copyright owner. For the purposes of this definition, "submitted"
-      means any form of electronic, verbal, or written communication sent
-      to the Licensor or its representatives, including but not limited to
-      communication on electronic mailing lists, source code control systems,
-      and issue tracking systems that are managed by, or on behalf of, the
-      Licensor for the purpose of discussing and improving the Work, but
-      excluding communication that is conspicuously marked or otherwise
-      designated in writing by the copyright owner as "Not a Contribution."
-
-      "Contributor" shall mean Licensor and any individual or Legal Entity
-      on behalf of whom a Contribution has been received by Licensor and
-      subsequently incorporated within the Work.
-
-   2. Grant of Copyright License. Subject to the terms and conditions of
-      this License, each Contributor hereby grants to You a perpetual,
-      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
-      copyright license to reproduce, prepare Derivative Works of,
-      publicly display, publicly perform, sublicense, and distribute the
-      Work and such Derivative Works in Source or Object form.
-
-   3. Grant of Patent License. Subject to the terms and conditions of
-      this License, each Contributor hereby grants to You a perpetual,
-      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
-      (except as stated in this section) patent license to make, have made,
-      use, offer to sell, sell, import, and otherwise transfer the Work,
-      where such license applies only to those patent claims licensable
-      by such Contributor that are necessarily infringed by their
-      Contribution(s) alone or by combination of their Contribution(s)
-      with the Work to which such Contribution(s) was submitted. If You
-      institute patent litigation against any entity (including a
-      cross-claim or counterclaim in a lawsuit) alleging that the Work
-      or a Contribution incorporated within the Work constitutes direct
-      or contributory patent infringement, then any patent licenses
-      granted to You under this License for that Work shall terminate
-      as of the date such litigation is filed.
-
-   4. Redistribution. You may reproduce and distribute copies of the
-      Work or Derivative Works thereof in any medium, with or without
-      modifications, and in Source or Object form, provided that You
-      meet the following conditions:
-
-      (a) You must give any other recipients of the Work or
-          Derivative Works a copy of this License; and
-
-      (b) You must cause any modified files to carry prominent notices
-          stating that You changed the files; and
-
-      (c) You must retain, in the Source form of any Derivative Works
-          that You distribute, all copyright, patent, trademark, and
-          attribution notices from the Source form of the Work,
-          excluding those notices that do not pertain to any part of
-          the Derivative Works; and
-
-      (d) If the Work includes a "NOTICE" text file as part of its
-          distribution, then any Derivative Works that You distribute must
-          include a readable copy of the attribution notices contained
-          within such NOTICE file, excluding those notices that do not
-          pertain to any part of the Derivative Works, in at least one
-          of the following places: within a NOTICE text file distributed
-          as part of the Derivative Works; within the Source form or
-          documentation, if provided along with the Derivative Works; or,
-          within a display generated by the Derivative Works, if and
-          wherever such third-party notices normally appear. The contents
-          of the NOTICE file are for informational purposes only and
-          do not modify the License. You may add Your own attribution
-          notices within Derivative Works that You distribute, alongside
-          or as an addendum to the NOTICE text from the Work, provided
-          that such additional attribution notices cannot be construed
-          as modifying the License.
-
-      You may add Your own copyright statement to Your modifications and
-      may provide additional or different license terms and conditions
-      for use, reproduction, or distribution of Your modifications, or
-      for any such Derivative Works as a whole, provided Your use,
-      reproduction, and distribution of the Work otherwise complies with
-      the conditions stated in this License.
-
-   5. Submission of Contributions. Unless You explicitly state otherwise,
-      any Contribution intentionally submitted for inclusion in the Work
-      by You to the Licensor shall be under the terms and conditions of
-      this License, without any additional terms or conditions.
-      Notwithstanding the above, nothing herein shall supersede or modify
-      the terms of any separate license agreement you may have executed
-      with Licensor regarding such Contributions.
-
-   6. Trademarks. This License does not grant permission to use the trade
-      names, trademarks, service marks, or product names of the Licensor,
-      except as required for reasonable and customary use in describing the
-      origin of the Work and reproducing the content of the NOTICE file.
-
-   7. Disclaimer of Warranty. Unless required by applicable law or
-      agreed to in writing, Licensor provides the Work (and each
-      Contributor provides its Contributions) on an "AS IS" BASIS,
-      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
-      implied, including, without limitation, any warranties or conditions
-      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
-      PARTICULAR PURPOSE. You are solely responsible for determining the
-      appropriateness of using or redistributing the Work and assume any
-      risks associated with Your exercise of permissions under this License.
-
-   8. Limitation of Liability. In no event and under no legal theory,
-      whether in tort (including negligence), contract, or otherwise,
-      unless required by applicable law (such as deliberate and grossly
-      negligent acts) or agreed to in writing, shall any Contributor be
-      liable to You for damages, including any direct, indirect, special,
-      incidental, or consequential damages of any character arising as a
-      result of this License or out of the use or inability to use the
-      Work (including but not limited to damages for loss of goodwill,
-      work stoppage, computer failure or malfunction, or any and all
-      other commercial damages or losses), even if such Contributor
-      has been advised of the possibility of such damages.
-
-   9. Accepting Warranty or Additional Liability. While redistributing
-      the Work or Derivative Works thereof, You may choose to offer,
-      and charge a fee for, acceptance of support, warranty, indemnity,
-      or other liability obligations and/or rights consistent with this
-      License. However, in accepting such obligations, You may act only
-      on Your own behalf and on Your sole responsibility, not on behalf
-      of any other Contributor, and only if You agree to indemnify,
-      defend, and hold each Contributor harmless for any liability
-      incurred by, or claims asserted against, such Contributor by reason
-      of your accepting any such warranty or additional liability.
-
-   END OF TERMS AND CONDITIONS
-
-   APPENDIX: How to apply the Apache License to your work.
-
-      To apply the Apache License to your work, attach the following
-      boilerplate notice, with the fields enclosed by brackets "[]"
-      replaced with your own identifying information. (Don't include
-      the brackets!)  The text should be enclosed in the appropriate
-      comment syntax for the file format. We also recommend that a
-      file or class name and description of purpose be included on the
-      same "printed page" as the copyright notice for easier
-      identification within third-party archives.
-
-   Copyright [yyyy] [name of copyright owner]
-
-   Licensed under the Apache License, Version 2.0 (the "License");
-   you may not use this file except in compliance with the License.
-   You may obtain a copy of the License at
-
-       http://www.apache.org/licenses/LICENSE-2.0
-
-   Unless required by applicable law or agreed to in writing, software
-   distributed under the License is distributed on an "AS IS" BASIS,
-   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-   See the License for the specific language governing permissions and
-   limitations under the License.
diff --git a/paddler_client_python/README.md b/paddler_client_python/README.md
deleted file mode 100644
index 1aa6845f8..000000000
--- a/paddler_client_python/README.md
+++ /dev/null
@@ -1,95 +0,0 @@
-# paddler-client
-
-Python client for the [Paddler](https://github.com/intentee/paddler) LLM load balancer.
-
-## Install
-
-```sh
-pip install paddler-client
-```
-
-Requires Python 3.11+. Built on `httpx`, `websockets`, and `pydantic` v2.
-
-## Quick start
-
-### WebSocket inference (multiplexed, request-id-keyed)
-
-```python
-import asyncio
-
-from paddler_client import ClientInference
-from paddler_client.continue_from_conversation_history_params import (
-    ContinueFromConversationHistoryParams,
-)
-from paddler_client.conversation_message import ConversationMessage
-
-
-async def main() -> None:
-    params = ContinueFromConversationHistoryParams(
-        add_generation_prompt=True,
-        conversation_history=[
-            ConversationMessage(content="Say hello.", role="user"),
-        ],
-        enable_thinking=False,
-        max_tokens=128,
-    )
-
-    async with ClientInference(url="http://localhost:8061") as client:
-        stream = await client.continue_from_conversation_history(params)
-
-        async for message in stream:
-            if message.is_token:
-                print(message.token, end="", flush=True)
-            elif message.is_terminal:
-                break
-
-
-asyncio.run(main())
-```
-
-### HTTP NDJSON streaming
-
-```python
-async with ClientInference(url="http://localhost:8061") as client:
-    async for message in client.post_continue_from_conversation_history(params):
-        if message.is_token:
-            print(message.token, end="", flush=True)
-        elif message.is_terminal:
-            break
-```
-
-`ClientInference` also provides `continue_from_raw_prompt` and
-`generate_embedding_batch`.
-
-### Management
-
-```python
-from paddler_client import ClientManagement
-
-
-async with ClientManagement(url="http://localhost:8060") as client:
-    snapshot = await client.get_agents()
-
-    for agent in snapshot.agents:
-        print(agent.id, agent.slots_total)
-
-    state = await client.get_balancer_desired_state()
-    await client.put_balancer_desired_state(state)
-
-    metrics = await client.get_metrics()
-```
-
-`ClientManagement` also provides `agents_stream`, `get_buffered_requests`,
-`buffered_requests_stream`, `get_chat_template_override`, and
-`get_model_metadata`.
-
-## Coverage
-
-- Transport: WebSocket (multiplexed), HTTP NDJSON, HTTP JSON, Server-Sent Events
-- Models: every Paddler wire-format type (validated via pydantic)
-- Fully typed (`py.typed`), async-first API
-- Specialized error types per failure mode
-
-## License
-
-Apache-2.0
diff --git a/paddler_client_python/paddler_client/__init__.py b/paddler_client_python/paddler_client/__init__.py
deleted file mode 100644
index f94da0a3e..000000000
--- a/paddler_client_python/paddler_client/__init__.py
+++ /dev/null
@@ -1,4 +0,0 @@
-from paddler_client.client_inference import ClientInference
-from paddler_client.client_management import ClientManagement
-
-__all__ = ["ClientInference", "ClientManagement"]
diff --git a/paddler_client_python/paddler_client/agent_controller_pool_snapshot.py b/paddler_client_python/paddler_client/agent_controller_pool_snapshot.py
deleted file mode 100644
index d8cc796cc..000000000
--- a/paddler_client_python/paddler_client/agent_controller_pool_snapshot.py
+++ /dev/null
@@ -1,7 +0,0 @@
-from pydantic import BaseModel
-
-from paddler_client.agent_controller_snapshot import AgentControllerSnapshot
-
-
-class AgentControllerPoolSnapshot(BaseModel):
-    agents: list[AgentControllerSnapshot]
diff --git a/paddler_client_python/paddler_client/agent_controller_snapshot.py b/paddler_client_python/paddler_client/agent_controller_snapshot.py
deleted file mode 100644
index c0009314a..000000000
--- a/paddler_client_python/paddler_client/agent_controller_snapshot.py
+++ /dev/null
@@ -1,10 +0,0 @@
-from pydantic import BaseModel
-
-from paddler_client.agent_status import AgentStatus
-
-
-class AgentControllerSnapshot(BaseModel):
-    id: str
-    name: str | None = None
-    slots_processing: int
-    status: AgentStatus
diff --git a/paddler_client_python/paddler_client/agent_desired_model.py b/paddler_client_python/paddler_client/agent_desired_model.py
deleted file mode 100644
index 34921f3e3..000000000
--- a/paddler_client_python/paddler_client/agent_desired_model.py
+++ /dev/null
@@ -1,95 +0,0 @@
-from typing import Any, cast
-
-from pydantic import BaseModel, ConfigDict, model_serializer, model_validator
-
-from paddler_client.error import (
-    AgentDesiredModelLocalPathMissingError,
-    AgentDesiredModelUrlMissingError,
-    InvalidAgentDesiredModelError,
-    UnknownAgentDesiredModelVariantError,
-)
-from paddler_client.huggingface_model_reference import (
-    HuggingFaceModelReference,
-)
-from paddler_client.url_model_reference import UrlModelReference
-
-
-class AgentDesiredModel(BaseModel):
-    model_config = ConfigDict(frozen=True)
-
-    variant: str
-    huggingface: HuggingFaceModelReference | None = None
-    local_path: str | None = None
-    url: UrlModelReference | None = None
-
-    @model_validator(mode="before")
-    @classmethod
-    def from_serde(cls, data: Any) -> dict[str, Any]:
-        if isinstance(data, str) and data == "None":
-            return {"variant": "None"}
-
-        if isinstance(data, dict):
-            typed_data = cast("dict[str, Any]", data)
-
-            if "HuggingFace" in typed_data:
-                return {
-                    "variant": "HuggingFace",
-                    "huggingface": typed_data["HuggingFace"],
-                }
-
-            if "LocalToAgent" in typed_data:
-                return {
-                    "variant": "LocalToAgent",
-                    "local_path": typed_data["LocalToAgent"],
-                }
-
-            if "Url" in typed_data:
-                return {
-                    "variant": "Url",
-                    "url": typed_data["Url"],
-                }
-
-            if "variant" in typed_data:
-                return typed_data
-
-        raise InvalidAgentDesiredModelError(data)
-
-    @model_serializer
-    def to_serde(self) -> str | dict[str, Any]:
-        if self.variant == "None":
-            return "None"
-
-        if self.variant == "HuggingFace" and self.huggingface is not None:
-            return {"HuggingFace": self.huggingface.model_dump()}
-
-        if self.variant == "LocalToAgent":
-            if self.local_path is None:
-                raise AgentDesiredModelLocalPathMissingError
-
-            return {"LocalToAgent": self.local_path}
-
-        if self.variant == "Url":
-            if self.url is None:
-                raise AgentDesiredModelUrlMissingError
-
-            return {"Url": self.url.model_dump()}
-
-        raise UnknownAgentDesiredModelVariantError(self.variant)
-
-    @classmethod
-    def none(cls) -> "AgentDesiredModel":
-        return cls(variant="None")
-
-    @classmethod
-    def from_huggingface(
-        cls, reference: HuggingFaceModelReference
-    ) -> "AgentDesiredModel":
-        return cls(variant="HuggingFace", huggingface=reference)
-
-    @classmethod
-    def local_to_agent(cls, path: str) -> "AgentDesiredModel":
-        return cls(variant="LocalToAgent", local_path=path)
-
-    @classmethod
-    def from_url(cls, reference: UrlModelReference) -> "AgentDesiredModel":
-        return cls(variant="Url", url=reference)
diff --git a/paddler_client_python/paddler_client/agent_issue.py b/paddler_client_python/paddler_client/agent_issue.py
deleted file mode 100644
index 07040959d..000000000
--- a/paddler_client_python/paddler_client/agent_issue.py
+++ /dev/null
@@ -1,39 +0,0 @@
-from typing import Any, cast
-
-from pydantic import BaseModel, Field, model_serializer, model_validator
-
-from paddler_client.error import InvalidAgentIssueError
-
-
-class AgentIssue(BaseModel):
-    variant: str
-    params: dict[str, Any] = Field(default_factory=dict)
-
-    @model_validator(mode="before")
-    @classmethod
-    def from_serde(cls, data: Any) -> dict[str, Any]:
-        if isinstance(data, str):
-            return {"variant": data, "params": {}}
-
-        if isinstance(data, dict):
-            typed_data = cast("dict[str, Any]", data)
-
-            if "variant" in typed_data:
-                return typed_data
-
-            if len(typed_data) == 1:
-                variant, params = next(iter(typed_data.items()))
-
-                if isinstance(params, dict):
-                    return {"variant": variant, "params": params}
-
-                return {"variant": variant, "params": {"value": params}}
-
-        raise InvalidAgentIssueError(data)
-
-    @model_serializer
-    def to_serde(self) -> str | dict[str, Any]:
-        if not self.params:
-            return self.variant
-
-        return {self.variant: self.params}
diff --git a/paddler_client_python/paddler_client/agent_state_application_status.py b/paddler_client_python/paddler_client/agent_state_application_status.py
deleted file mode 100644
index dd4a52dd0..000000000
--- a/paddler_client_python/paddler_client/agent_state_application_status.py
+++ /dev/null
@@ -1,9 +0,0 @@
-from enum import StrEnum
-
-
-class AgentStateApplicationStatus(StrEnum):
-    APPLIED = "Applied"
-    ATTEMPTED_AND_NOT_APPLIABLE = "AttemptedAndNotAppliable"
-    ATTEMPTED_AND_RETRYING = "AttemptedAndRetrying"
-    FRESH = "Fresh"
-    STUCK = "Stuck"
diff --git a/paddler_client_python/paddler_client/agent_status.py b/paddler_client_python/paddler_client/agent_status.py
deleted file mode 100644
index e04bc5f59..000000000
--- a/paddler_client_python/paddler_client/agent_status.py
+++ /dev/null
@@ -1,17 +0,0 @@
-from pydantic import BaseModel
-
-from paddler_client.agent_issue import AgentIssue
-from paddler_client.agent_state_application_status import (
-    AgentStateApplicationStatus,
-)
-from paddler_client.model_download_status import ModelDownloadStatus
-
-
-class AgentStatus(BaseModel):
-    desired_slots_total: int
-    download_status: ModelDownloadStatus
-    issues: list[AgentIssue] = []
-    model_path: str | None = None
-    slots_total: int
-    state_application_status: AgentStateApplicationStatus
-    uses_chat_template_override: bool
diff --git a/paddler_client_python/paddler_client/balancer_desired_state.py b/paddler_client_python/paddler_client/balancer_desired_state.py
deleted file mode 100644
index f9a602ee7..000000000
--- a/paddler_client_python/paddler_client/balancer_desired_state.py
+++ /dev/null
@@ -1,17 +0,0 @@
-from pydantic import BaseModel, Field
-
-from paddler_client.agent_desired_model import AgentDesiredModel
-from paddler_client.chat_template import ChatTemplate
-from paddler_client.inference_parameters import InferenceParameters
-
-
-class BalancerDesiredState(BaseModel):
-    chat_template_override: ChatTemplate | None = None
-    inference_parameters: InferenceParameters = Field(
-        default_factory=InferenceParameters,
-    )
-    model: AgentDesiredModel = Field(default_factory=AgentDesiredModel.none)
-    multimodal_projection: AgentDesiredModel = Field(
-        default_factory=AgentDesiredModel.none,
-    )
-    use_chat_template_override: bool = False
diff --git a/paddler_client_python/paddler_client/buffered_request_manager_snapshot.py b/paddler_client_python/paddler_client/buffered_request_manager_snapshot.py
deleted file mode 100644
index dc7a2b96e..000000000
--- a/paddler_client_python/paddler_client/buffered_request_manager_snapshot.py
+++ /dev/null
@@ -1,5 +0,0 @@
-from pydantic import BaseModel
-
-
-class BufferedRequestManagerSnapshot(BaseModel):
-    buffered_requests_current: int
diff --git a/paddler_client_python/paddler_client/chat_template.py b/paddler_client_python/paddler_client/chat_template.py
deleted file mode 100644
index bc3a4d35e..000000000
--- a/paddler_client_python/paddler_client/chat_template.py
+++ /dev/null
@@ -1,5 +0,0 @@
-from pydantic import BaseModel
-
-
-class ChatTemplate(BaseModel):
-    content: str
diff --git a/paddler_client_python/paddler_client/client_inference.py b/paddler_client_python/paddler_client/client_inference.py
deleted file mode 100644
index 3de8f67bc..000000000
--- a/paddler_client_python/paddler_client/client_inference.py
+++ /dev/null
@@ -1,162 +0,0 @@
-from __future__ import annotations
-
-import uuid
-from typing import TYPE_CHECKING, Literal, Self
-
-import httpx
-
-from paddler_client.error import HttpError
-from paddler_client.inference_socket_pool import InferenceSocketPool
-from paddler_client.inference_socket_url import inference_socket_url
-from paddler_client.stream_ndjson import stream_ndjson_inference_messages
-
-if TYPE_CHECKING:
-    from collections.abc import AsyncGenerator
-
-    from pydantic import BaseModel
-
-    from paddler_client.continue_from_conversation_history_params import (
-        ContinueFromConversationHistoryParams,
-    )
-    from paddler_client.continue_from_raw_prompt_params import (
-        ContinueFromRawPromptParams,
-    )
-    from paddler_client.generate_embedding_batch_params import (
-        GenerateEmbeddingBatchParams,
-    )
-    from paddler_client.inference_message import InferenceMessage
-    from paddler_client.response_stream import ResponseStream
-
-InferenceRequestVariant = Literal[
-    "ContinueFromConversationHistory",
-    "ContinueFromRawPrompt",
-]
-
-
-class ClientInference:
-    def __init__(
-        self,
-        url: str,
-        socket_pool_size: int = 4,
-        http_client: httpx.AsyncClient | None = None,
-    ) -> None:
-        self._url = url.rstrip("/")
-        self._socket_pool_size = socket_pool_size
-        self._http_client = (
-            http_client
-            if http_client is not None
-            else httpx.AsyncClient(timeout=httpx.Timeout(timeout=300.0))
-        )
-        self._socket_pool: InferenceSocketPool | None = None
-
-    def _get_socket_pool(self) -> InferenceSocketPool:
-        if self._socket_pool is None:
-            ws_url = inference_socket_url(self._url)
-            self._socket_pool = InferenceSocketPool(
-                url=ws_url,
-                pool_size=self._socket_pool_size,
-            )
-
-        return self._socket_pool
-
-    async def get_health(self) -> str:
-        url = f"{self._url}/health"
-        response = await self._http_client.get(url)
-
-        if not response.is_success:
-            raise HttpError(response.status_code, response.text)
-
-        return response.text
-
-    async def continue_from_conversation_history(
-        self,
-        params: ContinueFromConversationHistoryParams,
-    ) -> ResponseStream:
-        return await self._send_ws_request(
-            "ContinueFromConversationHistory",
-            params,
-        )
-
-    async def continue_from_raw_prompt(
-        self,
-        params: ContinueFromRawPromptParams,
-    ) -> ResponseStream:
-        return await self._send_ws_request(
-            "ContinueFromRawPrompt",
-            params,
-        )
-
-    async def post_continue_from_conversation_history(
-        self,
-        params: ContinueFromConversationHistoryParams,
-    ) -> AsyncGenerator[InferenceMessage, None]:
-        async for message in self._stream_ndjson_post(
-            "/api/v1/continue_from_conversation_history",
-            params,
-        ):
-            yield message
-
-    async def generate_embedding_batch(
-        self,
-        params: GenerateEmbeddingBatchParams,
-    ) -> AsyncGenerator[InferenceMessage, None]:
-        async for message in self._stream_ndjson_post(
-            "/api/v1/generate_embedding_batch",
-            params,
-        ):
-            yield message
-
-    async def close(self) -> None:
-        if self._socket_pool is not None:
-            await self._socket_pool.close()
-
-        await self._http_client.aclose()
-
-    async def __aenter__(self) -> Self:
-        return self
-
-    async def __aexit__(self, *args: object) -> None:
-        await self.close()
-
-    async def _send_ws_request(
-        self,
-        variant: InferenceRequestVariant,
-        params: BaseModel,
-    ) -> ResponseStream:
-        request_id = str(uuid.uuid4())
-        message: dict[str, object] = {
-            "Request": {
-                "id": request_id,
-                "request": {
-                    variant: params.model_dump(
-                        mode="json",
-                        exclude_none=True,
-                        by_alias=True,
-                    ),
-                },
-            },
-        }
-        pool = self._get_socket_pool()
-
-        return await pool.send_request(request_id, message)
-
-    async def _stream_ndjson_post(
-        self,
-        path: str,
-        params: BaseModel,
-    ) -> AsyncGenerator[InferenceMessage, None]:
-        url = f"{self._url}{path}"
-        request_body = params.model_dump(
-            mode="json",
-            exclude_none=True,
-            by_alias=True,
-        )
-
-        async with self._http_client.stream(
-            "POST",
-            url,
-            json=request_body,
-            headers={"Content-Type": "application/json"},
-        ) as response:
-            async for message in stream_ndjson_inference_messages(response):
-                yield message
diff --git a/paddler_client_python/paddler_client/client_management.py b/paddler_client_python/paddler_client/client_management.py
deleted file mode 100644
index 40873a04c..000000000
--- a/paddler_client_python/paddler_client/client_management.py
+++ /dev/null
@@ -1,161 +0,0 @@
-from __future__ import annotations
-
-from typing import TYPE_CHECKING, Self
-
-import httpx
-
-from paddler_client.agent_controller_pool_snapshot import (
-    AgentControllerPoolSnapshot,
-)
-from paddler_client.balancer_desired_state import BalancerDesiredState
-from paddler_client.buffered_request_manager_snapshot import (
-    BufferedRequestManagerSnapshot,
-)
-from paddler_client.chat_template import ChatTemplate
-from paddler_client.error import HttpError
-from paddler_client.model_metadata import ModelMetadata
-from paddler_client.stream_sse import stream_sse
-
-if TYPE_CHECKING:
-    from collections.abc import AsyncGenerator
-
-
-class ClientManagement:
-    def __init__(
-        self,
-        url: str,
-        http_client: httpx.AsyncClient | None = None,
-    ) -> None:
-        self._url = url.rstrip("/")
-        self._http_client = (
-            http_client if http_client is not None else httpx.AsyncClient()
-        )
-
-    async def get_health(self) -> str:
-        response = await self._http_client.get(f"{self._url}/health")
-
-        if not response.is_success:
-            raise HttpError(response.status_code, response.text)
-
-        return response.text
-
-    async def get_agents(self) -> AgentControllerPoolSnapshot:
-        response = await self._http_client.get(
-            f"{self._url}/api/v1/agents",
-        )
-
-        if not response.is_success:
-            raise HttpError(response.status_code, response.text)
-
-        return AgentControllerPoolSnapshot.model_validate_json(
-            response.content,
-        )
-
-    async def agents_stream(
-        self,
-    ) -> AsyncGenerator[AgentControllerPoolSnapshot, None]:
-        url = f"{self._url}/api/v1/agents/stream"
-
-        async with self._http_client.stream("GET", url) as response:
-            async for data in stream_sse(response):
-                yield AgentControllerPoolSnapshot.model_validate_json(data)
-
-    async def get_balancer_desired_state(self) -> BalancerDesiredState:
-        response = await self._http_client.get(
-            f"{self._url}/api/v1/balancer_desired_state",
-        )
-
-        if not response.is_success:
-            raise HttpError(response.status_code, response.text)
-
-        return BalancerDesiredState.model_validate_json(response.content)
-
-    async def put_balancer_desired_state(
-        self,
-        state: BalancerDesiredState,
-    ) -> None:
-        response = await self._http_client.put(
-            f"{self._url}/api/v1/balancer_desired_state",
-            content=state.model_dump_json(
-                exclude_none=True,
-                by_alias=True,
-            ),
-            headers={"Content-Type": "application/json"},
-        )
-
-        if not response.is_success:
-            raise HttpError(response.status_code, response.text)
-
-    async def get_buffered_requests(
-        self,
-    ) -> BufferedRequestManagerSnapshot:
-        response = await self._http_client.get(
-            f"{self._url}/api/v1/buffered_requests",
-        )
-
-        if not response.is_success:
-            raise HttpError(response.status_code, response.text)
-
-        return BufferedRequestManagerSnapshot.model_validate_json(
-            response.content,
-        )
-
-    async def buffered_requests_stream(
-        self,
-    ) -> AsyncGenerator[BufferedRequestManagerSnapshot, None]:
-        url = f"{self._url}/api/v1/buffered_requests/stream"
-
-        async with self._http_client.stream("GET", url) as response:
-            async for data in stream_sse(response):
-                yield BufferedRequestManagerSnapshot.model_validate_json(
-                    data,
-                )
-
-    async def get_chat_template_override(
-        self,
-        agent_id: str,
-    ) -> ChatTemplate | None:
-        response = await self._http_client.get(
-            f"{self._url}/api/v1/agent/{agent_id}/chat_template_override",
-        )
-
-        if not response.is_success:
-            raise HttpError(response.status_code, response.text)
-
-        if not response.content or response.text == "null":
-            return None
-
-        return ChatTemplate.model_validate_json(response.content)
-
-    async def get_model_metadata(
-        self,
-        agent_id: str,
-    ) -> ModelMetadata | None:
-        response = await self._http_client.get(
-            f"{self._url}/api/v1/agent/{agent_id}/model_metadata",
-        )
-
-        if not response.is_success:
-            raise HttpError(response.status_code, response.text)
-
-        if not response.content or response.text == "null":
-            return None
-
-        return ModelMetadata.model_validate_json(response.content)
-
-    async def get_metrics(self) -> str:
-        response = await self._http_client.get(f"{self._url}/metrics")
-
-        if not response.is_success:
-            raise HttpError(response.status_code, response.text)
-
-        return response.text
-
-    async def close(self) -> None:
-        await self._http_client.aclose()
-
-    async def __aenter__(self) -> Self:
-        return self
-
-    async def __aexit__(self, *args: object) -> None:
-        await self.close()
diff --git a/paddler_client_python/paddler_client/continue_from_conversation_history_params.py b/paddler_client_python/paddler_client/continue_from_conversation_history_params.py
deleted file mode 100644
index 77a2992b5..000000000
--- a/paddler_client_python/paddler_client/continue_from_conversation_history_params.py
+++ /dev/null
@@ -1,25 +0,0 @@
-from pydantic import BaseModel, PositiveInt, model_validator
-
-from paddler_client.conversation_message import ConversationMessage
-from paddler_client.error import ToolCallParsingWithoutToolsError
-from paddler_client.grammar_constraint import GrammarConstraint
-from paddler_client.tool import Tool
-
-
-class ContinueFromConversationHistoryParams(BaseModel):
-    add_generation_prompt: bool
-    conversation_history: list[ConversationMessage]
-    enable_thinking: bool
-    grammar: GrammarConstraint | None = None
-    max_tokens: PositiveInt
-    parse_tool_calls: bool = False
-    tools: list[Tool] = []
-
-    @model_validator(mode="after")
-    def reject_tool_call_parsing_without_tools(
-        self,
-    ) -> "ContinueFromConversationHistoryParams":
-        if self.parse_tool_calls and not self.tools:
-            raise ToolCallParsingWithoutToolsError
-
-        return self
diff --git a/paddler_client_python/paddler_client/continue_from_raw_prompt_params.py b/paddler_client_python/paddler_client/continue_from_raw_prompt_params.py
deleted file mode 100644
index 972dee08e..000000000
--- a/paddler_client_python/paddler_client/continue_from_raw_prompt_params.py
+++ /dev/null
@@ -1,9 +0,0 @@
-from pydantic import BaseModel, PositiveInt
-
-from paddler_client.grammar_constraint import GrammarConstraint
-
-
-class ContinueFromRawPromptParams(BaseModel):
-    grammar: GrammarConstraint | None = None
-    max_tokens: PositiveInt
-    raw_prompt: str
diff --git a/paddler_client_python/paddler_client/conversation_message.py b/paddler_client_python/paddler_client/conversation_message.py
deleted file mode 100644
index aee5a4ef0..000000000
--- a/paddler_client_python/paddler_client/conversation_message.py
+++ /dev/null
@@ -1,12 +0,0 @@
-from pydantic import BaseModel
-
-from paddler_client.conversation_message_content_part import (
-    ConversationMessageContentPart,
-)
-
-ConversationMessageContent = str | list[ConversationMessageContentPart]
-
-
-class ConversationMessage(BaseModel):
-    content: ConversationMessageContent
-    role: str
diff --git a/paddler_client_python/paddler_client/conversation_message_content_part.py b/paddler_client_python/paddler_client/conversation_message_content_part.py
deleted file mode 100644
index c9229f01a..000000000
--- a/paddler_client_python/paddler_client/conversation_message_content_part.py
+++ /dev/null
@@ -1,22 +0,0 @@
-from typing import Annotated, Literal
-
-from pydantic import BaseModel, Discriminator, Tag
-
-from paddler_client.image_url import ImageUrl
-
-
-class TextContentPart(BaseModel):
-    type: Literal["text"] = "text"
-    text: str
-
-
-class ImageUrlContentPart(BaseModel):
-    type: Literal["image_url"] = "image_url"
-    image_url: ImageUrl
-
-
-ConversationMessageContentPart = Annotated[
-    Annotated[TextContentPart, Tag("text")]
-    | Annotated[ImageUrlContentPart, Tag("image_url")],
-    Discriminator("type"),
-]
diff --git a/paddler_client_python/paddler_client/embedding.py b/paddler_client_python/paddler_client/embedding.py
deleted file mode 100644
index 54282ffd0..000000000
--- a/paddler_client_python/paddler_client/embedding.py
+++ /dev/null
@@ -1,15 +0,0 @@
-from pydantic import BaseModel, ConfigDict
-
-from paddler_client.embedding_normalization_method import (
-    EmbeddingNormalizationMethod,
-)
-from paddler_client.pooling_type import PoolingType
-
-
-class Embedding(BaseModel):
-    model_config = ConfigDict(frozen=True)
-
-    embedding: list[float]
-    normalization_method: EmbeddingNormalizationMethod
-    pooling_type: PoolingType
-    source_document_id: str
diff --git a/paddler_client_python/paddler_client/embedding_input_document.py b/paddler_client_python/paddler_client/embedding_input_document.py
deleted file mode 100644
index af67c8e95..000000000
--- a/paddler_client_python/paddler_client/embedding_input_document.py
+++ /dev/null
@@ -1,6 +0,0 @@
-from pydantic import BaseModel
-
-
-class EmbeddingInputDocument(BaseModel):
-    content: str
-    id: str
diff --git a/paddler_client_python/paddler_client/embedding_normalization_method.py b/paddler_client_python/paddler_client/embedding_normalization_method.py
deleted file mode 100644
index da0f6fd67..000000000
--- a/paddler_client_python/paddler_client/embedding_normalization_method.py
+++ /dev/null
@@ -1,63 +0,0 @@
-from typing import Any, Self, cast
-
-from pydantic import BaseModel, ConfigDict, model_serializer, model_validator
-
-from paddler_client.error import (
-    InvalidEmbeddingNormalizationMethodError,
-    InvalidRmsNormPayloadError,
-    RmsNormEpsilonMissingError,
-)
-
-
-class EmbeddingNormalizationMethod(BaseModel):
-    model_config = ConfigDict(frozen=True)
-
-    variant: str
-    epsilon: float | None = None
-
-    @model_validator(mode="before")
-    @classmethod
-    def from_serde(cls, data: Any) -> dict[str, Any]:
-        if isinstance(data, str):
-            return {"variant": data}
-
-        if isinstance(data, dict):
-            typed_data = cast("dict[str, Any]", data)
-
-            if "RmsNorm" in typed_data:
-                inner: object = typed_data["RmsNorm"]
-
-                if not isinstance(inner, dict) or "epsilon" not in inner:
-                    raise InvalidRmsNormPayloadError(data)
-
-                return {
-                    "variant": "RmsNorm",
-                    "epsilon": inner["epsilon"],
-                }
-
-            if "variant" in typed_data:
-                return typed_data
-
-        raise InvalidEmbeddingNormalizationMethodError(data)
-
-    @model_serializer
-    def to_serde(self) -> str | dict[str, Any]:
-        if self.variant == "RmsNorm":
-            if self.epsilon is None:
-                raise RmsNormEpsilonMissingError
-
-            return {"RmsNorm": {"epsilon": self.epsilon}}
-
-        return self.variant
-
-    @classmethod
-    def l2(cls) -> Self:
-        return cls(variant="L2")
-
-    @classmethod
-    def none(cls) -> Self:
-        return cls(variant="None")
-
-    @classmethod
-    def rms_norm(cls, epsilon: float) -> Self:
-        return cls(variant="RmsNorm", epsilon=epsilon)
diff --git a/paddler_client_python/paddler_client/error.py b/paddler_client_python/paddler_client/error.py
deleted file mode 100644
index 7161995a0..000000000
--- a/paddler_client_python/paddler_client/error.py
+++ /dev/null
@@ -1,205 +0,0 @@
-class PaddlerError(Exception):
-    pass
-
-
-class HttpError(PaddlerError):
-    def __init__(self, status_code: int, message: str) -> None:
-        self.status_code = status_code
-        self.message = message
-        super().__init__(f"HTTP {status_code}: {message}")
-
-
-class JsonError(PaddlerError):
-    def __init__(self, message: str, raw_data: str) -> None:
-        self.raw_data = raw_data
-        super().__init__(message)
-
-
-class InvalidSocketPoolSizeError(PaddlerError, ValueError):
-    def __init__(self, pool_size: int) -> None:
-        self.pool_size = pool_size
-        super().__init__(f"pool_size must be >= 1, got {pool_size}")
-
-
-class ConnectionDroppedError(PaddlerError):
-    def __init__(self, request_id: str) -> None:
-        self.request_id = request_id
-        super().__init__(f"WebSocket connection dropped for request {request_id}")
-
-
-class RequestIdInFlightError(PaddlerError, ValueError):
-    def __init__(self, request_id: str) -> None:
-        self.request_id = request_id
-        super().__init__(
-            f"Request {request_id} is already in flight on this connection"
-        )
-
-
-class InvalidAgentDesiredModelError(PaddlerError, ValueError):
-    def __init__(self, data: object) -> None:
-        self.data = data
-        super().__init__(f"Invalid AgentDesiredModel: {data}")
-
-
-class AgentDesiredModelLocalPathMissingError(PaddlerError, ValueError):
-    def __init__(self) -> None:
-        super().__init__("local_path is required for LocalToAgent")
-
-
-class AgentDesiredModelUrlMissingError(PaddlerError, ValueError):
-    def __init__(self) -> None:
-        super().__init__("url is required for Url")
-
-
-class UnknownAgentDesiredModelVariantError(PaddlerError, ValueError):
-    def __init__(self, variant: str) -> None:
-        self.variant = variant
-        super().__init__(f"Unknown AgentDesiredModel variant: {variant}")
-
-
-class InvalidAgentIssueError(PaddlerError, ValueError):
-    def __init__(self, data: object) -> None:
-        self.data = data
-        super().__init__(f"Invalid AgentIssue: {data}")
-
-
-class InvalidModelDownloadStatusError(PaddlerError, ValueError):
-    def __init__(self, data: object) -> None:
-        self.data = data
-        super().__init__(f"Invalid ModelDownloadStatus: {data}")
-
-
-class InvalidRmsNormPayloadError(PaddlerError, ValueError):
-    def __init__(self, data: object) -> None:
-        self.data = data
-        super().__init__(f"Invalid RmsNorm payload: {data}")
-
-
-class InvalidEmbeddingNormalizationMethodError(PaddlerError, ValueError):
-    def __init__(self, data: object) -> None:
-        self.data = data
-        super().__init__(f"Invalid EmbeddingNormalizationMethod: {data}")
-
-
-class RmsNormEpsilonMissingError(PaddlerError, ValueError):
-    def __init__(self) -> None:
-        super().__init__("epsilon is required for RmsNorm")
-
-
-class ApiPathWithoutLeadingSlashError(PaddlerError, ValueError):
-    def __init__(self, path: str) -> None:
-        self.path = path
-        super().__init__(f"Path must start with '/': {path}")
-
-
-class UnsupportedUrlSchemeError(PaddlerError, ValueError):
-    def __init__(self, scheme: str) -> None:
-        self.scheme = scheme
-        super().__init__(f"Unsupported URL scheme: {scheme}")
-
-
-class UnknownToolCallArgumentsError(PaddlerError, ValueError):
-    def __init__(self, payload: object) -> None:
-        self.payload = payload
-        super().__init__(f"Unknown ToolCallArguments shape: {payload}")
-
-
-class ToolCallArgumentsNotAnObjectError(PaddlerError, TypeError):
-    def __init__(self, arguments: object) -> None:
-        self.arguments = arguments
-        super().__init__(
-            f"arguments field must be a dict (tagged enum), got: {arguments!r}"
-        )
-
-
-class InferenceClientMessageNotAnObjectError(PaddlerError, TypeError):
-    def __init__(self, data: object) -> None:
-        self.data = data
-        super().__init__(f"Unknown inference client message format: {data}")
-
-
-class UnknownInferenceClientMessageError(PaddlerError, ValueError):
-    def __init__(self, data: object) -> None:
-        self.data = data
-        super().__init__(f"Unknown inference client message format: {data}")
-
-
-class UnknownResponseVariantError(PaddlerError, ValueError):
-    def __init__(self, response: object) -> None:
-        self.response = response
-        super().__init__(f"Unknown response variant: {response}")
-
-
-class GeneratedTokenResultNotAnObjectError(PaddlerError, TypeError):
-    def __init__(self, data: object) -> None:
-        self.data = data
-        super().__init__(f"Unknown GeneratedTokenResult: {data}")
-
-
-class UnknownGeneratedTokenResultError(PaddlerError, ValueError):
-    def __init__(self, data: object) -> None:
-        self.data = data
-        super().__init__(f"Unknown GeneratedTokenResult: {data}")
-
-
-class UnknownEmbeddingResultError(PaddlerError, ValueError):
-    def __init__(self, data: object) -> None:
-        self.data = data
-        super().__init__(f"Unknown EmbeddingResult: {data}")
-
-
-class ToolCallParsedPayloadNotAListError(PaddlerError, TypeError):
-    def __init__(self, payload: object) -> None:
-        self.payload = payload
-        super().__init__(f"ToolCallParsed payload is not a list: {payload}")
-
-
-class ToolCallValidationFailedPayloadNotAListError(PaddlerError, TypeError):
-    def __init__(self, payload: object) -> None:
-        self.payload = payload
-        super().__init__(f"ToolCallValidationFailed payload is not a list: {payload}")
-
-
-class UnrecognizedToolCallFormatPayloadNotAnObjectError(PaddlerError, TypeError):
-    def __init__(self, payload: object) -> None:
-        self.payload = payload
-        super().__init__(
-            f"UnrecognizedToolCallFormat payload is not a dict: {payload!r}"
-        )
-
-
-class MediaExceedsMicroBatchPayloadNotAnObjectError(PaddlerError, TypeError):
-    def __init__(self, payload: object) -> None:
-        self.payload = payload
-        super().__init__(f"MediaExceedsMicroBatch payload is not a dict: {payload!r}")
-
-
-class PromptExceedsContextSizePayloadNotAnObjectError(PaddlerError, TypeError):
-    def __init__(self, payload: object) -> None:
-        self.payload = payload
-        super().__init__(f"PromptExceedsContextSize payload is not a dict: {payload!r}")
-
-
-class BatchSizeExceedsContextSizeError(PaddlerError, ValueError):
-    def __init__(self, n_batch: int, context_size: int) -> None:
-        self.n_batch = n_batch
-        self.context_size = context_size
-        super().__init__(f"n_batch {n_batch} exceeds context_size {context_size}")
-
-
-class PenaltiesWithoutWindowError(PaddlerError, ValueError):
-    def __init__(self) -> None:
-        super().__init__("penalty strengths require a penalty_last_n window")
-
-
-class PenaltyWindowWithoutPenaltiesError(PaddlerError, ValueError):
-    def __init__(self, penalty_last_n: int) -> None:
-        self.penalty_last_n = penalty_last_n
-        super().__init__(
-            f"penalty_last_n {penalty_last_n} requires at least one penalty strength"
-        )
-
-
-class ToolCallParsingWithoutToolsError(PaddlerError, ValueError):
-    def __init__(self) -> None:
-        super().__init__("parse_tool_calls requires at least one tool")
diff --git a/paddler_client_python/paddler_client/format_api_url.py b/paddler_client_python/paddler_client/format_api_url.py
deleted file mode 100644
index 252211f76..000000000
--- a/paddler_client_python/paddler_client/format_api_url.py
+++ /dev/null
@@ -1,8 +0,0 @@
-from paddler_client.error import ApiPathWithoutLeadingSlashError
-
-
-def format_api_url(base_url: str, path: str) -> str:
-    if not path.startswith("/"):
-        raise ApiPathWithoutLeadingSlashError(path)
-
-    return base_url.rstrip("/") + path
diff --git a/paddler_client_python/paddler_client/generate_embedding_batch_params.py b/paddler_client_python/paddler_client/generate_embedding_batch_params.py
deleted file mode 100644
index a51119f8b..000000000
--- a/paddler_client_python/paddler_client/generate_embedding_batch_params.py
+++ /dev/null
@@ -1,11 +0,0 @@
-from pydantic import BaseModel
-
-from paddler_client.embedding_input_document import EmbeddingInputDocument
-from paddler_client.embedding_normalization_method import (
-    EmbeddingNormalizationMethod,
-)
-
-
-class GenerateEmbeddingBatchParams(BaseModel):
-    input_batch: list[EmbeddingInputDocument]
-    normalization_method: EmbeddingNormalizationMethod
diff --git a/paddler_client_python/paddler_client/generation_finish.py b/paddler_client_python/paddler_client/generation_finish.py
deleted file mode 100644
index 1cc2f8414..000000000
--- a/paddler_client_python/paddler_client/generation_finish.py
+++ /dev/null
@@ -1,8 +0,0 @@
-from enum import StrEnum
-
-
-class GenerationFinish(StrEnum):
-    CONTEXT_FULL = "ContextFull"
-    END_OF_GENERATION = "EndOfGeneration"
-    MAX_TOKENS = "MaxTokens"
-    STOP_REQUESTED = "StopRequested"
diff --git a/paddler_client_python/paddler_client/grammar_constraint.py b/paddler_client_python/paddler_client/grammar_constraint.py
deleted file mode 100644
index 7dce46d86..000000000
--- a/paddler_client_python/paddler_client/grammar_constraint.py
+++ /dev/null
@@ -1,19 +0,0 @@
-from typing import Annotated, Literal
-
-from pydantic import BaseModel, ConfigDict, Field
-
-
-class GbnfGrammarConstraint(BaseModel):
-    type: Literal["gbnf"] = "gbnf"
-    grammar: str
-    root: str
-
-
-class JsonSchemaGrammarConstraint(BaseModel):
-    model_config = ConfigDict(populate_by_name=True)
-
-    type: Literal["json_schema"] = "json_schema"
-    schema_value: Annotated[str, Field(alias="schema")]
-
-
-GrammarConstraint = GbnfGrammarConstraint | JsonSchemaGrammarConstraint
diff --git a/paddler_client_python/paddler_client/huggingface_model_reference.py b/paddler_client_python/paddler_client/huggingface_model_reference.py
deleted file mode 100644
index 323a32841..000000000
--- a/paddler_client_python/paddler_client/huggingface_model_reference.py
+++ /dev/null
@@ -1,7 +0,0 @@
-from pydantic import BaseModel
-
-
-class HuggingFaceModelReference(BaseModel):
-    filename: str
-    repo_id: str
-    revision: str
diff --git a/paddler_client_python/paddler_client/image_url.py b/paddler_client_python/paddler_client/image_url.py
deleted file mode 100644
index 53738dd02..000000000
--- a/paddler_client_python/paddler_client/image_url.py
+++ /dev/null
@@ -1,5 +0,0 @@
-from pydantic import BaseModel
-
-
-class ImageUrl(BaseModel):
-    url: str
diff --git a/paddler_client_python/paddler_client/inference_message.py b/paddler_client_python/paddler_client/inference_message.py
deleted file mode 100644
index 9400da74d..000000000
--- a/paddler_client_python/paddler_client/inference_message.py
+++ /dev/null
@@ -1,539 +0,0 @@
-from __future__ import annotations
-
-import json
-from collections.abc import Callable
-from dataclasses import dataclass
-from enum import StrEnum
-from typing import Any, cast
-
-from paddler_client.embedding import Embedding
-from paddler_client.error import (
-    GeneratedTokenResultNotAnObjectError,
-    InferenceClientMessageNotAnObjectError,
-    MediaExceedsMicroBatchPayloadNotAnObjectError,
-    PromptExceedsContextSizePayloadNotAnObjectError,
-    ToolCallParsedPayloadNotAListError,
-    ToolCallValidationFailedPayloadNotAListError,
-    UnknownEmbeddingResultError,
-    UnknownGeneratedTokenResultError,
-    UnknownInferenceClientMessageError,
-    UnknownResponseVariantError,
-    UnrecognizedToolCallFormatPayloadNotAnObjectError,
-)
-from paddler_client.generation_finish import GenerationFinish
-from paddler_client.oversized_embedding_document_details import (
-    OversizedEmbeddingDocumentDetails,
-)
-from paddler_client.oversized_media_details import OversizedMediaDetails
-from paddler_client.oversized_prompt_details import OversizedPromptDetails
-from paddler_client.parsed_tool_call import ParsedToolCall
-from paddler_client.raw_tool_call_tokens import RawToolCallTokens
-
-
-class InferenceMessageKind(StrEnum):
-    BATCH_ASSEMBLY_FAILED = "batch_assembly_failed"
-    CHAT_TEMPLATE_ERROR = "chat_template_error"
-    CONTENT_TOKEN = "content_token"
-    DECODE_FAILED = "decode_failed"
-    DETOKENIZATION_FAILED = "detokenization_failed"
-    DONE = "done"
-    EMBEDDING = "embedding"
-    EMBEDDINGS_DISABLED = "embeddings_disabled"
-    EMBEDDING_DOCUMENT_EXCEEDS_BATCH_SIZE = "embedding_document_exceeds_batch_size"
-    EMBEDDING_DONE = "embedding_done"
-    EMBEDDING_ERROR = "embedding_error"
-    EMBEDDING_NO_EMBEDDINGS_PRODUCED = "embedding_no_embeddings_produced"
-    EMBEDDING_REJECTED_DUE_TO_ACTIVE_TOKEN_GENERATION = (
-        "embedding_rejected_due_to_active_token_generation"
-    )
-    GRAMMAR_INCOMPATIBLE_WITH_THINKING = "grammar_incompatible_with_thinking"
-    GRAMMAR_INITIALIZATION_FAILED = "grammar_initialization_failed"
-    GRAMMAR_REJECTED_MODEL_OUTPUT = "grammar_rejected_model_output"
-    GRAMMAR_SYNTAX_ERROR = "grammar_syntax_error"
-    IMAGE_DECODING_FAILED = "image_decoding_failed"
-    KV_CACHE_CLEAR_FAILED = "kv_cache_clear_failed"
-    MEDIA_EXCEEDS_MICRO_BATCH = "media_exceeds_micro_batch"
-    MEDIA_MICRO_BATCH_CHECK_FAILED = "media_micro_batch_check_failed"
-    MODEL_NOT_LOADED = "model_not_loaded"
-    MULTIMODAL_INGESTION_FAILED = "multimodal_ingestion_failed"
-    MULTIMODAL_NOT_SUPPORTED = "multimodal_not_supported"
-    MULTIMODAL_TOKENIZATION_FAILED = "multimodal_tokenization_failed"
-    NO_SEQUENCE_SLOT_AVAILABLE = "no_sequence_slot_available"
-    PROMPT_EXCEEDS_CONTEXT_SIZE = "prompt_exceeds_context_size"
-    PROMPT_TOKENIZATION_FAILED = "prompt_tokenization_failed"
-    REASONING_TOKEN = "reasoning_token"
-    SAMPLER_CHAIN_CREATION_FAILED = "sampler_chain_creation_failed"
-    SAMPLER_ERROR = "sampler_error"
-    SAMPLING_CANDIDATES_EXHAUSTED = "sampling_candidates_exhausted"
-    SCHEDULER_UNAVAILABLE = "scheduler_unavailable"
-    SEQUENCE_ID_OUT_OF_RANGE = "sequence_id_out_of_range"
-    SERVER_ERROR = "server_error"
-    TOKEN_GENERATION_DISABLED = "token_generation_disabled"
-    TOOLS_SERIALIZATION_FAILED = "tools_serialization_failed"
-    TOOL_CALL_PARSED = "tool_call_parsed"
-    TOOL_CALL_PARSE_FAILED = "tool_call_parse_failed"
-    TOOL_CALL_TOKEN = "tool_call_token"
-    TOOL_CALL_VALIDATION_FAILED = "tool_call_validation_failed"
-    TOOL_SCHEMA_INVALID = "tool_schema_invalid"
-    UNDETERMINABLE_TOKEN = "undeterminable_token"
-    UNRECOGNIZED_TOOL_CALL_FORMAT = "unrecognized_tool_call_format"
-
-
-_TOKEN_KINDS: frozenset[InferenceMessageKind] = frozenset(
-    {
-        InferenceMessageKind.CONTENT_TOKEN,
-        InferenceMessageKind.REASONING_TOKEN,
-        InferenceMessageKind.TOOL_CALL_TOKEN,
-        InferenceMessageKind.UNDETERMINABLE_TOKEN,
-    },
-)
-
-_NON_TERMINAL_KINDS: frozenset[InferenceMessageKind] = _TOKEN_KINDS | frozenset(
-    {
-        InferenceMessageKind.EMBEDDING,
-        InferenceMessageKind.EMBEDDING_DOCUMENT_EXCEEDS_BATCH_SIZE,
-        InferenceMessageKind.TOOL_CALL_PARSED,
-        InferenceMessageKind.TOOL_CALL_PARSE_FAILED,
-        InferenceMessageKind.TOOL_CALL_VALIDATION_FAILED,
-        InferenceMessageKind.UNRECOGNIZED_TOOL_CALL_FORMAT,
-    },
-)
-
-
-@dataclass(frozen=True)
-class TokenUsage:
-    prompt_tokens: int = 0
-    cached_prompt_tokens: int = 0
-    input_image_tokens: int = 0
-    input_audio_tokens: int = 0
-    content_tokens: int = 0
-    reasoning_tokens: int = 0
-    tool_call_tokens: int = 0
-    undeterminable_tokens: int = 0
-
-    @property
-    def completion_tokens(self) -> int:
-        return (
-            self.content_tokens
-            + self.reasoning_tokens
-            + self.tool_call_tokens
-            + self.undeterminable_tokens
-        )
-
-    @property
-    def total_tokens(self) -> int:
-        return self.prompt_tokens + self.completion_tokens
-
-    @classmethod
-    def from_dict(cls, data: dict[str, Any]) -> TokenUsage:
-        return cls(
-            prompt_tokens=int(data["prompt_tokens"]),
-            cached_prompt_tokens=int(data["cached_prompt_tokens"]),
-            input_image_tokens=int(data["input_image_tokens"]),
-            input_audio_tokens=int(data["input_audio_tokens"]),
-            content_tokens=int(data["content_tokens"]),
-            reasoning_tokens=int(data["reasoning_tokens"]),
-            tool_call_tokens=int(data["tool_call_tokens"]),
-            undeterminable_tokens=int(data["undeterminable_tokens"]),
-        )
-
-
-@dataclass(frozen=True)
-class GenerationSummary:
-    finish: GenerationFinish
-    usage: TokenUsage
-
-    @classmethod
-    def from_dict(cls, data: dict[str, Any]) -> GenerationSummary:
-        return cls(
-            finish=GenerationFinish(data["finish"]),
-            usage=TokenUsage.from_dict(data["usage"]),
-        )
-
-
-@dataclass(frozen=True)
-class InferenceMessage:
-    request_id: str
-    kind: InferenceMessageKind
-    token: str | None = None
-    embedding_data: Embedding | None = None
-    error_message: str | None = None
-    error_code: int | None = None
-    summary: GenerationSummary | None = None
-    parsed_tool_calls: list[ParsedToolCall] | None = None
-    raw_tool_call_tokens: RawToolCallTokens | None = None
-    oversized_media_details: OversizedMediaDetails | None = None
-    oversized_prompt_details: OversizedPromptDetails | None = None
-    oversized_embedding_document_details: OversizedEmbeddingDocumentDetails | None = (
-        None
-    )
-    generated_by: str | None = None
-
-    @property
-    def is_token(self) -> bool:
-        return self.kind in _TOKEN_KINDS
-
-    @property
-    def is_done(self) -> bool:
-        return self.kind == InferenceMessageKind.DONE
-
-    @property
-    def is_terminal(self) -> bool:
-        return self.kind not in _NON_TERMINAL_KINDS
-
-
-def parse_inference_client_message(
-    data: str | dict[str, Any],
-) -> InferenceMessage:
-    if isinstance(data, str):
-        data = json.loads(data)
-
-    if not isinstance(data, dict):
-        raise InferenceClientMessageNotAnObjectError(data)
-
-    if "Error" in data:
-        return _parse_error_envelope(data["Error"])
-
-    if "Response" in data:
-        response_envelope = data["Response"]
-
-        return _parse_response(
-            response_envelope["request_id"],
-            response_envelope["response"],
-            response_envelope["generated_by"],
-        )
-
-    raise UnknownInferenceClientMessageError(data)
-
-
-def _parse_error_envelope(
-    error_envelope: dict[str, Any],
-) -> InferenceMessage:
-    error = error_envelope["error"]
-
-    return InferenceMessage(
-        request_id=error_envelope["request_id"],
-        kind=InferenceMessageKind.SERVER_ERROR,
-        error_code=error["code"],
-        error_message=error["description"],
-    )
-
-
-def _parse_response(
-    request_id: str,
-    response: str | dict[str, Any],
-    generated_by: str | None,
-) -> InferenceMessage:
-    if isinstance(response, dict):
-        if "GeneratedToken" in response:
-            return _parse_generated_token_result(
-                request_id,
-                response["GeneratedToken"],
-                generated_by,
-            )
-
-        if "Embedding" in response:
-            return _parse_embedding_result(
-                request_id,
-                response["Embedding"],
-                generated_by,
-            )
-
-    raise UnknownResponseVariantError(response)
-
-
-_GENERATED_TOKEN_ERROR_KINDS: dict[str, InferenceMessageKind] = {
-    "ChatTemplateError": InferenceMessageKind.CHAT_TEMPLATE_ERROR,
-    "DecodeFailed": InferenceMessageKind.DECODE_FAILED,
-    "DetokenizationFailed": InferenceMessageKind.DETOKENIZATION_FAILED,
-    "GrammarIncompatibleWithThinking": (
-        InferenceMessageKind.GRAMMAR_INCOMPATIBLE_WITH_THINKING
-    ),
-    "GrammarInitializationFailed": InferenceMessageKind.GRAMMAR_INITIALIZATION_FAILED,
-    "GrammarRejectedModelOutput": InferenceMessageKind.GRAMMAR_REJECTED_MODEL_OUTPUT,
-    "GrammarSyntaxError": InferenceMessageKind.GRAMMAR_SYNTAX_ERROR,
-    "ImageDecodingFailed": InferenceMessageKind.IMAGE_DECODING_FAILED,
-    "ModelNotLoaded": InferenceMessageKind.MODEL_NOT_LOADED,
-    "MultimodalNotSupported": InferenceMessageKind.MULTIMODAL_NOT_SUPPORTED,
-    "SamplerError": InferenceMessageKind.SAMPLER_ERROR,
-    "TokenGenerationDisabled": InferenceMessageKind.TOKEN_GENERATION_DISABLED,
-    "ToolSchemaInvalid": InferenceMessageKind.TOOL_SCHEMA_INVALID,
-    "BatchAssemblyFailed": InferenceMessageKind.BATCH_ASSEMBLY_FAILED,
-    "KvCacheClearFailed": InferenceMessageKind.KV_CACHE_CLEAR_FAILED,
-    "MediaMicroBatchCheckFailed": InferenceMessageKind.MEDIA_MICRO_BATCH_CHECK_FAILED,
-    "MultimodalIngestionFailed": InferenceMessageKind.MULTIMODAL_INGESTION_FAILED,
-    "MultimodalTokenizationFailed": InferenceMessageKind.MULTIMODAL_TOKENIZATION_FAILED,
-    "NoSequenceSlotAvailable": InferenceMessageKind.NO_SEQUENCE_SLOT_AVAILABLE,
-    "PromptTokenizationFailed": InferenceMessageKind.PROMPT_TOKENIZATION_FAILED,
-    "SamplerChainCreationFailed": InferenceMessageKind.SAMPLER_CHAIN_CREATION_FAILED,
-    "SamplingCandidatesExhausted": InferenceMessageKind.SAMPLING_CANDIDATES_EXHAUSTED,
-    "SchedulerUnavailable": InferenceMessageKind.SCHEDULER_UNAVAILABLE,
-    "SequenceIdOutOfRange": InferenceMessageKind.SEQUENCE_ID_OUT_OF_RANGE,
-    "ToolsSerializationFailed": InferenceMessageKind.TOOLS_SERIALIZATION_FAILED,
-}
-
-
-_GENERATED_TOKEN_KINDS: dict[str, InferenceMessageKind] = {
-    "ContentToken": InferenceMessageKind.CONTENT_TOKEN,
-    "ReasoningToken": InferenceMessageKind.REASONING_TOKEN,
-    "ToolCallToken": InferenceMessageKind.TOOL_CALL_TOKEN,
-    "UndeterminableToken": InferenceMessageKind.UNDETERMINABLE_TOKEN,
-}
-
-
-def _build_done_message(
-    request_id: str,
-    payload: Any,
-    generated_by: str | None,
-) -> InferenceMessage:
-    return InferenceMessage(
-        request_id=request_id,
-        kind=InferenceMessageKind.DONE,
-        summary=GenerationSummary.from_dict(payload),
-        generated_by=generated_by,
-    )
-
-
-def _build_tool_call_parsed_message(
-    request_id: str,
-    payload: Any,
-    generated_by: str | None,
-) -> InferenceMessage:
-    if not isinstance(payload, list):
-        raise ToolCallParsedPayloadNotAListError(payload)
-    typed_calls = cast("list[dict[str, Any]]", payload)
-    parsed_calls: list[ParsedToolCall] = [
-        ParsedToolCall.from_dict(call) for call in typed_calls
-    ]
-    return InferenceMessage(
-        request_id=request_id,
-        kind=InferenceMessageKind.TOOL_CALL_PARSED,
-        parsed_tool_calls=parsed_calls,
-        generated_by=generated_by,
-    )
-
-
-def _build_tool_call_parse_failed_message(
-    request_id: str,
-    payload: Any,
-    generated_by: str | None,
-) -> InferenceMessage:
-    return InferenceMessage(
-        request_id=request_id,
-        kind=InferenceMessageKind.TOOL_CALL_PARSE_FAILED,
-        error_message=str(payload),
-        generated_by=generated_by,
-    )
-
-
-def _build_tool_call_validation_failed_message(
-    request_id: str,
-    payload: Any,
-    generated_by: str | None,
-) -> InferenceMessage:
-    if not isinstance(payload, list):
-        raise ToolCallValidationFailedPayloadNotAListError(payload)
-    typed_errors = cast("list[object]", payload)
-    joined_errors: str = "; ".join(str(error) for error in typed_errors)
-    return InferenceMessage(
-        request_id=request_id,
-        kind=InferenceMessageKind.TOOL_CALL_VALIDATION_FAILED,
-        error_message=joined_errors,
-        generated_by=generated_by,
-    )
-
-
-def _build_unrecognized_tool_call_format_message(
-    request_id: str,
-    payload: Any,
-    generated_by: str | None,
-) -> InferenceMessage:
-    if not isinstance(payload, dict):
-        raise UnrecognizedToolCallFormatPayloadNotAnObjectError(payload)
-    typed_raw = cast("dict[str, Any]", payload)
-    return InferenceMessage(
-        request_id=request_id,
-        kind=InferenceMessageKind.UNRECOGNIZED_TOOL_CALL_FORMAT,
-        raw_tool_call_tokens=RawToolCallTokens.from_dict(typed_raw),
-        generated_by=generated_by,
-    )
-
-
-def _build_media_exceeds_micro_batch_message(
-    request_id: str,
-    payload: Any,
-    generated_by: str | None,
-) -> InferenceMessage:
-    if not isinstance(payload, dict):
-        raise MediaExceedsMicroBatchPayloadNotAnObjectError(payload)
-    typed_details = cast("dict[str, Any]", payload)
-    return InferenceMessage(
-        request_id=request_id,
-        kind=InferenceMessageKind.MEDIA_EXCEEDS_MICRO_BATCH,
-        oversized_media_details=OversizedMediaDetails.from_dict(typed_details),
-        generated_by=generated_by,
-    )
-
-
-def _build_prompt_exceeds_context_size_message(
-    request_id: str,
-    payload: Any,
-    generated_by: str | None,
-) -> InferenceMessage:
-    if not isinstance(payload, dict):
-        raise PromptExceedsContextSizePayloadNotAnObjectError(payload)
-    typed_details = cast("dict[str, Any]", payload)
-    return InferenceMessage(
-        request_id=request_id,
-        kind=InferenceMessageKind.PROMPT_EXCEEDS_CONTEXT_SIZE,
-        oversized_prompt_details=OversizedPromptDetails.from_dict(typed_details),
-        generated_by=generated_by,
-    )
-
-
-def _build_token_kind_message(
-    request_id: str,
-    kind: InferenceMessageKind,
-    payload: Any,
-    generated_by: str | None,
-) -> InferenceMessage:
-    return InferenceMessage(
-        request_id=request_id,
-        kind=kind,
-        token=payload,
-        generated_by=generated_by,
-    )
-
-
-def _build_error_kind_message(
-    request_id: str,
-    kind: InferenceMessageKind,
-    payload: Any,
-    generated_by: str | None,
-) -> InferenceMessage:
-    return InferenceMessage(
-        request_id=request_id,
-        kind=kind,
-        error_message=payload,
-        generated_by=generated_by,
-    )
-
-
-_StructuredHandler = Callable[[str, Any, str | None], InferenceMessage]
-
-_STRUCTURED_HANDLERS: dict[str, _StructuredHandler] = {
-    "Done": _build_done_message,
-    "ToolCallParsed": _build_tool_call_parsed_message,
-    "ToolCallParseFailed": _build_tool_call_parse_failed_message,
-    "ToolCallValidationFailed": _build_tool_call_validation_failed_message,
-    "UnrecognizedToolCallFormat": _build_unrecognized_tool_call_format_message,
-    "MediaExceedsMicroBatch": _build_media_exceeds_micro_batch_message,
-    "PromptExceedsContextSize": _build_prompt_exceeds_context_size_message,
-}
-
-
-def _parse_generated_token_result(
-    request_id: str,
-    data: str | dict[str, Any],
-    generated_by: str | None,
-) -> InferenceMessage:
-    if not isinstance(data, dict):
-        raise GeneratedTokenResultNotAnObjectError(data)
-    for structured_key, handler in _STRUCTURED_HANDLERS.items():
-        if structured_key in data:
-            return handler(request_id, data[structured_key], generated_by)
-    for token_key, token_kind in _GENERATED_TOKEN_KINDS.items():
-        if token_key in data:
-            return _build_token_kind_message(
-                request_id,
-                token_kind,
-                data[token_key],
-                generated_by,
-            )
-    for error_key, error_kind in _GENERATED_TOKEN_ERROR_KINDS.items():
-        if error_key in data:
-            return _build_error_kind_message(
-                request_id,
-                error_kind,
-                data[error_key],
-                generated_by,
-            )
-    raise UnknownGeneratedTokenResultError(data)
-
-
-_EMBEDDING_UNIT_KINDS: dict[str, InferenceMessageKind] = {
-    "Done": InferenceMessageKind.EMBEDDING_DONE,
-    "EmbeddingRejectedDueToActiveTokenGeneration": (
-        InferenceMessageKind.EMBEDDING_REJECTED_DUE_TO_ACTIVE_TOKEN_GENERATION
-    ),
-    "EmbeddingsDisabled": InferenceMessageKind.EMBEDDINGS_DISABLED,
-    "NoEmbeddingsProduced": InferenceMessageKind.EMBEDDING_NO_EMBEDDINGS_PRODUCED,
-}
-
-
-def _build_embedding_message(
-    request_id: str,
-    payload: Any,
-    generated_by: str | None,
-) -> InferenceMessage:
-    return InferenceMessage(
-        request_id=request_id,
-        kind=InferenceMessageKind.EMBEDDING,
-        embedding_data=Embedding.model_validate(payload),
-        generated_by=generated_by,
-    )
-
-
-def _build_embedding_document_exceeds_batch_size_message(
-    request_id: str,
-    payload: Any,
-    generated_by: str | None,
-) -> InferenceMessage:
-    return InferenceMessage(
-        request_id=request_id,
-        kind=InferenceMessageKind.EMBEDDING_DOCUMENT_EXCEEDS_BATCH_SIZE,
-        oversized_embedding_document_details=(
-            OversizedEmbeddingDocumentDetails.from_dict(payload)
-        ),
-        generated_by=generated_by,
-    )
-
-
-_EMBEDDING_STRUCTURED_HANDLERS: dict[str, _StructuredHandler] = {
-    "DocumentExceedsBatchSize": _build_embedding_document_exceeds_batch_size_message,
-    "Embedding": _build_embedding_message,
-}
-
-
-_EMBEDDING_ERROR_KINDS: dict[str, InferenceMessageKind] = {
-    "Error": InferenceMessageKind.EMBEDDING_ERROR,
-    "ModelNotLoaded": InferenceMessageKind.MODEL_NOT_LOADED,
-}
-
-
-def _parse_embedding_result(
-    request_id: str,
-    data: str | dict[str, Any],
-    generated_by: str | None,
-) -> InferenceMessage:
-    if isinstance(data, str) and data in _EMBEDDING_UNIT_KINDS:
-        return InferenceMessage(
-            request_id=request_id,
-            kind=_EMBEDDING_UNIT_KINDS[data],
-            generated_by=generated_by,
-        )
-
-    if isinstance(data, dict):
-        for structured_key, handler in _EMBEDDING_STRUCTURED_HANDLERS.items():
-            if structured_key in data:
-                return handler(request_id, data[structured_key], generated_by)
-        for error_key, error_kind in _EMBEDDING_ERROR_KINDS.items():
-            if error_key in data:
-                return _build_error_kind_message(
-                    request_id,
-                    error_kind,
-                    data[error_key],
-                    generated_by,
-                )
-
-    raise UnknownEmbeddingResultError(data)
diff --git a/paddler_client_python/paddler_client/inference_parameters.py b/paddler_client_python/paddler_client/inference_parameters.py
deleted file mode 100644
index 28bb8b5e9..000000000
--- a/paddler_client_python/paddler_client/inference_parameters.py
+++ /dev/null
@@ -1,53 +0,0 @@
-from pydantic import BaseModel, ConfigDict, Field, model_validator
-
-from paddler_client.error import (
-    BatchSizeExceedsContextSizeError,
-    PenaltiesWithoutWindowError,
-    PenaltyWindowWithoutPenaltiesError,
-)
-from paddler_client.kv_cache_dtype import KvCacheDtype
-from paddler_client.pooling_type import PoolingType
-
-I32_MAX = 2_147_483_647
-NEUTRAL_PENALTY_REPEAT = 1.0
-
-
-class InferenceParameters(BaseModel):
-    model_config = ConfigDict(allow_inf_nan=False)
-
-    n_batch: int = Field(default=2048, ge=1, le=I32_MAX)
-    context_size: int = Field(default=8192, ge=1)
-    embedding_batch_size: int = Field(default=256, ge=1)
-    enable_embeddings: bool = False
-    image_resize_to_fit: int = Field(default=1024, ge=1)
-    k_cache_dtype: KvCacheDtype = KvCacheDtype.Q8_0
-    v_cache_dtype: KvCacheDtype = KvCacheDtype.Q8_0
-    min_p: float = Field(default=0.05, ge=0.0, le=1.0)
-    n_gpu_layers: int = Field(default=0, ge=-1)
-    penalty_frequency: float = 0.0
-    penalty_last_n: int = Field(default=0, ge=0)
-    penalty_presence: float = 0.0
-    penalty_repeat: float = Field(default=NEUTRAL_PENALTY_REPEAT, gt=0.0)
-    pooling_type: PoolingType = PoolingType.LAST
-    temperature: float = Field(default=0.8, ge=0.0)
-    top_k: int = Field(default=80, ge=0)
-    top_p: float = Field(default=0.8, ge=0.0, le=1.0)
-
-    @model_validator(mode="after")
-    def reject_ambiguous_combinations(self) -> "InferenceParameters":
-        penalties_are_neutral = (
-            self.penalty_repeat == NEUTRAL_PENALTY_REPEAT
-            and self.penalty_frequency == 0.0
-            and self.penalty_presence == 0.0
-        )
-
-        if self.n_batch > self.context_size:
-            raise BatchSizeExceedsContextSizeError(self.n_batch, self.context_size)
-
-        if self.penalty_last_n == 0 and not penalties_are_neutral:
-            raise PenaltiesWithoutWindowError
-
-        if self.penalty_last_n > 0 and penalties_are_neutral:
-            raise PenaltyWindowWithoutPenaltiesError(self.penalty_last_n)
-
-        return self
diff --git a/paddler_client_python/paddler_client/inference_socket_connection.py b/paddler_client_python/paddler_client/inference_socket_connection.py
deleted file mode 100644
index 72a41583d..000000000
--- a/paddler_client_python/paddler_client/inference_socket_connection.py
+++ /dev/null
@@ -1,83 +0,0 @@
-from __future__ import annotations
-
-import asyncio
-import contextlib
-import logging
-from typing import TYPE_CHECKING, Self
-
-import websockets
-from websockets.asyncio.client import ClientConnection, connect
-
-from paddler_client.error import ConnectionDroppedError
-from paddler_client.pending_requests import PendingRequests
-
-if TYPE_CHECKING:
-    from paddler_client.response_stream import ResponseStream
-
-logger = logging.getLogger(__name__)
-
-
-class InferenceSocketConnection:
-    def __init__(self, websocket: ClientConnection) -> None:
-        self._websocket = websocket
-        self._pending = PendingRequests()
-        self._write_queue: asyncio.Queue[str] = asyncio.Queue()
-        self._connected = True
-        self._read_task = asyncio.create_task(self._read_loop())
-        self._write_task = asyncio.create_task(self._write_loop())
-
-    @classmethod
-    async def connect(cls, url: str) -> Self:
-        return cls(await connect(url))
-
-    @property
-    def is_connected(self) -> bool:
-        return self._connected
-
-    async def send(
-        self,
-        request_id: str,
-        json_message: str,
-    ) -> ResponseStream:
-        if not self._connected:
-            raise ConnectionDroppedError(request_id)
-
-        response_stream = self._pending.register(request_id)
-        await self._write_queue.put(json_message)
-
-        return response_stream
-
-    async def close(self) -> None:
-        self._connected = False
-
-        await self._websocket.close()
-
-        for task in (self._write_task, self._read_task):
-            task.cancel()
-
-            with contextlib.suppress(
-                asyncio.CancelledError, websockets.ConnectionClosed
-            ):
-                await task
-
-    async def _read_loop(self) -> None:
-        try:
-            while True:
-                self._pending.dispatch(await self._websocket.recv(decode=True))
-        except websockets.ConnectionClosed:
-            logger.debug("WebSocket connection closed")
-        finally:
-            self._drop_connection()
-            self._write_task.cancel()
-
-    async def _write_loop(self) -> None:
-        try:
-            while True:
-                await self._websocket.send(await self._write_queue.get())
-        finally:
-            self._drop_connection()
-            self._read_task.cancel()
-
-    def _drop_connection(self) -> None:
-        self._connected = False
-        self._pending.drain()
diff --git a/paddler_client_python/paddler_client/inference_socket_pool.py b/paddler_client_python/paddler_client/inference_socket_pool.py
deleted file mode 100644
index 5ed14ca51..000000000
--- a/paddler_client_python/paddler_client/inference_socket_pool.py
+++ /dev/null
@@ -1,62 +0,0 @@
-from __future__ import annotations
-
-import asyncio
-import json
-from typing import TYPE_CHECKING
-
-from paddler_client.error import InvalidSocketPoolSizeError
-from paddler_client.inference_socket_connection import InferenceSocketConnection
-
-if TYPE_CHECKING:
-    from paddler_client.response_stream import ResponseStream
-
-
-class InferenceSocketPool:
-    def __init__(self, url: str, pool_size: int) -> None:
-        if pool_size < 1:
-            raise InvalidSocketPoolSizeError(pool_size)
-
-        self._url = url
-        self._pool_size = pool_size
-        self._connections: list[InferenceSocketConnection | None] = [None] * pool_size
-        self._next_connection_index = 0
-        self._lock = asyncio.Lock()
-
-    async def send_request(
-        self,
-        request_id: str,
-        message: dict[str, object],
-    ) -> ResponseStream:
-        serialized_message = json.dumps(message)
-
-        async with self._lock:
-            connection_index = self._next_connection_index
-            self._next_connection_index = (
-                self._next_connection_index + 1
-            ) % self._pool_size
-            connection = await self._ensure_connected(connection_index)
-
-        return await connection.send(request_id, serialized_message)
-
-    async def close(self) -> None:
-        for connection in self._connections:
-            if connection is not None:
-                await connection.close()
-
-        self._connections = [None] * self._pool_size
-
-    async def _ensure_connected(
-        self, connection_index: int
-    ) -> InferenceSocketConnection:
-        connection = self._connections[connection_index]
-
-        if connection is not None and connection.is_connected:
-            return connection
-
-        if connection is not None:
-            await connection.close()
-
-        new_connection = await InferenceSocketConnection.connect(self._url)
-        self._connections[connection_index] = new_connection
-
-        return new_connection
diff --git a/paddler_client_python/paddler_client/inference_socket_url.py b/paddler_client_python/paddler_client/inference_socket_url.py
deleted file mode 100644
index 185f7765f..000000000
--- a/paddler_client_python/paddler_client/inference_socket_url.py
+++ /dev/null
@@ -1,29 +0,0 @@
-from urllib.parse import urlparse, urlunparse
-
-from paddler_client.error import UnsupportedUrlSchemeError
-
-_SCHEME_MAP: dict[str, str] = {
-    "http": "ws",
-    "https": "wss",
-    "ws": "ws",
-    "wss": "wss",
-}
-
-
-def inference_socket_url(url: str) -> str:
-    parsed = urlparse(url)
-    new_scheme = _SCHEME_MAP.get(parsed.scheme)
-
-    if new_scheme is None:
-        raise UnsupportedUrlSchemeError(parsed.scheme)
-
-    return urlunparse(
-        (
-            new_scheme,
-            parsed.netloc,
-            "/api/v1/inference_socket",
-            "",
-            parsed.query,
-            "",
-        )
-    )
diff --git a/paddler_client_python/paddler_client/kv_cache_dtype.py b/paddler_client_python/paddler_client/kv_cache_dtype.py
deleted file mode 100644
index 44650eda3..000000000
--- a/paddler_client_python/paddler_client/kv_cache_dtype.py
+++ /dev/null
@@ -1,13 +0,0 @@
-from enum import StrEnum
-
-
-class KvCacheDtype(StrEnum):
-    F32 = "F32"
-    F16 = "F16"
-    BF16 = "BF16"
-    Q8_0 = "Q8_0"
-    Q4_0 = "Q4_0"
-    Q4_1 = "Q4_1"
-    IQ4_NL = "IQ4_NL"
-    Q5_0 = "Q5_0"
-    Q5_1 = "Q5_1"
diff --git a/paddler_client_python/paddler_client/model_download_status.py b/paddler_client_python/paddler_client/model_download_status.py
deleted file mode 100644
index 64ffcd612..000000000
--- a/paddler_client_python/paddler_client/model_download_status.py
+++ /dev/null
@@ -1,36 +0,0 @@
-from typing import Any, cast
-
-from pydantic import BaseModel, ConfigDict, model_validator
-
-from paddler_client.error import InvalidModelDownloadStatusError
-
-DOWNLOADING_VARIANTS = ("Downloading", "DownloadingWithUnknownSize")
-
-
-class ModelDownloadStatus(BaseModel):
-    model_config = ConfigDict(frozen=True)
-
-    variant: str
-    downloaded_bytes: int | None = None
-    model_path: str | None = None
-    total_bytes: int | None = None
-
-    @model_validator(mode="before")
-    @classmethod
-    def from_serde(cls, data: Any) -> dict[str, Any]:
-        if data == "NotDownloading":
-            return {"variant": "NotDownloading"}
-
-        if isinstance(data, dict):
-            typed_data = cast("dict[str, Any]", data)
-
-            if "variant" in typed_data:
-                return typed_data
-
-            if len(typed_data) == 1:
-                variant, progress = next(iter(typed_data.items()))
-
-                if variant in DOWNLOADING_VARIANTS and isinstance(progress, dict):
-                    return {"variant": variant, **cast("dict[str, Any]", progress)}
-
-        raise InvalidModelDownloadStatusError(data)
diff --git a/paddler_client_python/paddler_client/model_metadata.py b/paddler_client_python/paddler_client/model_metadata.py
deleted file mode 100644
index dae606715..000000000
--- a/paddler_client_python/paddler_client/model_metadata.py
+++ /dev/null
@@ -1,5 +0,0 @@
-from pydantic import BaseModel, Field
-
-
-class ModelMetadata(BaseModel):
-    metadata: dict[str, str] = Field(default_factory=dict)
diff --git a/paddler_client_python/paddler_client/ndjson_inference_messages.py b/paddler_client_python/paddler_client/ndjson_inference_messages.py
deleted file mode 100644
index ac1bb6429..000000000
--- a/paddler_client_python/paddler_client/ndjson_inference_messages.py
+++ /dev/null
@@ -1,22 +0,0 @@
-from __future__ import annotations
-
-import json
-from typing import TYPE_CHECKING
-
-from paddler_client.inference_message import (
-    InferenceMessage,
-    parse_inference_client_message,
-)
-
-if TYPE_CHECKING:
-    from collections.abc import AsyncGenerator, AsyncIterable
-
-
-async def ndjson_inference_messages(
-    lines: AsyncIterable[str],
-) -> AsyncGenerator[InferenceMessage, None]:
-    async for line in lines:
-        stripped_line = line.strip()
-
-        if stripped_line:
-            yield parse_inference_client_message(json.loads(stripped_line))
diff --git a/paddler_client_python/paddler_client/oversized_embedding_document_details.py b/paddler_client_python/paddler_client/oversized_embedding_document_details.py
deleted file mode 100644
index f8bb78b5c..000000000
--- a/paddler_client_python/paddler_client/oversized_embedding_document_details.py
+++ /dev/null
@@ -1,19 +0,0 @@
-from __future__ import annotations
-
-from dataclasses import dataclass
-from typing import Any
-
-
-@dataclass(frozen=True)
-class OversizedEmbeddingDocumentDetails:
-    document_tokens: int
-    n_batch: int
-    source_document_id: str
-
-    @classmethod
-    def from_dict(cls, data: dict[str, Any]) -> OversizedEmbeddingDocumentDetails:
-        return cls(
-            document_tokens=int(data["document_tokens"]),
-            n_batch=int(data["n_batch"]),
-            source_document_id=str(data["source_document_id"]),
-        )
diff --git a/paddler_client_python/paddler_client/oversized_media_details.py b/paddler_client_python/paddler_client/oversized_media_details.py
deleted file mode 100644
index c88bb1d9b..000000000
--- a/paddler_client_python/paddler_client/oversized_media_details.py
+++ /dev/null
@@ -1,17 +0,0 @@
-from __future__ import annotations
-
-from dataclasses import dataclass
-from typing import Any
-
-
-@dataclass(frozen=True)
-class OversizedMediaDetails:
-    media_tokens: int
-    micro_batch_tokens: int
-
-    @classmethod
-    def from_dict(cls, data: dict[str, Any]) -> OversizedMediaDetails:
-        return cls(
-            media_tokens=int(data["media_tokens"]),
-            micro_batch_tokens=int(data["micro_batch_tokens"]),
-        )
diff --git a/paddler_client_python/paddler_client/oversized_prompt_details.py b/paddler_client_python/paddler_client/oversized_prompt_details.py
deleted file mode 100644
index 7f14d54fa..000000000
--- a/paddler_client_python/paddler_client/oversized_prompt_details.py
+++ /dev/null
@@ -1,17 +0,0 @@
-from __future__ import annotations
-
-from dataclasses import dataclass
-from typing import Any
-
-
-@dataclass(frozen=True)
-class OversizedPromptDetails:
-    prompt_tokens: int
-    sequence_context_size: int
-
-    @classmethod
-    def from_dict(cls, data: dict[str, Any]) -> OversizedPromptDetails:
-        return cls(
-            prompt_tokens=int(data["prompt_tokens"]),
-            sequence_context_size=int(data["sequence_context_size"]),
-        )
diff --git a/paddler_client_python/paddler_client/parsed_tool_call.py b/paddler_client_python/paddler_client/parsed_tool_call.py
deleted file mode 100644
index c05600820..000000000
--- a/paddler_client_python/paddler_client/parsed_tool_call.py
+++ /dev/null
@@ -1,29 +0,0 @@
-from __future__ import annotations
-
-from dataclasses import dataclass
-from typing import Any, cast
-
-from paddler_client.error import ToolCallArgumentsNotAnObjectError
-from paddler_client.tool_call_arguments import (
-    ToolCallArguments,
-    parse_tool_call_arguments,
-)
-
-
-@dataclass(frozen=True)
-class ParsedToolCall:
-    id: str
-    name: str
-    arguments: ToolCallArguments
-
-    @classmethod
-    def from_dict(cls, data: dict[str, Any]) -> ParsedToolCall:
-        arguments_payload = data["arguments"]
-        if not isinstance(arguments_payload, dict):
-            raise ToolCallArgumentsNotAnObjectError(arguments_payload)
-        typed_payload = cast("dict[str, Any]", arguments_payload)
-        return cls(
-            id=str(data["id"]),
-            name=str(data["name"]),
-            arguments=parse_tool_call_arguments(typed_payload),
-        )
diff --git a/paddler_client_python/paddler_client/pending_requests.py b/paddler_client_python/paddler_client/pending_requests.py
deleted file mode 100644
index 797fda2b3..000000000
--- a/paddler_client_python/paddler_client/pending_requests.py
+++ /dev/null
@@ -1,83 +0,0 @@
-from __future__ import annotations
-
-import asyncio
-import json
-import logging
-
-from paddler_client.error import (
-    ConnectionDroppedError,
-    JsonError,
-    RequestIdInFlightError,
-)
-from paddler_client.inference_message import (
-    InferenceMessage,
-    parse_inference_client_message,
-)
-from paddler_client.response_stream import ResponseStream
-from paddler_client.unparsed_message_request_id import unparsed_message_request_id
-
-logger = logging.getLogger(__name__)
-
-
-class PendingRequests:
-    def __init__(self) -> None:
-        self._queues: dict[str, asyncio.Queue[InferenceMessage | Exception]] = {}
-
-    def register(self, request_id: str) -> ResponseStream:
-        if request_id in self._queues:
-            raise RequestIdInFlightError(request_id)
-
-        queue: asyncio.Queue[InferenceMessage | Exception] = asyncio.Queue()
-        self._queues[request_id] = queue
-
-        return ResponseStream(queue)
-
-    def dispatch(self, raw_message: str) -> None:
-        try:
-            data = json.loads(raw_message)
-        except json.JSONDecodeError:
-            logger.exception("Received a WebSocket message that is not JSON")
-
-            return
-
-        if isinstance(data, dict) and "Notification" in data:
-            return
-
-        try:
-            message = parse_inference_client_message(data)
-        except (KeyError, TypeError, ValueError):
-            logger.exception("Failed to parse WebSocket message")
-            self._fail(
-                unparsed_message_request_id(data),
-                JsonError("Failed to parse message", raw_data=raw_message),
-            )
-
-            return
-
-        queue = self._queues.get(message.request_id)
-
-        if queue is None:
-            logger.warning(
-                "Received message for unknown request: %s",
-                message.request_id,
-            )
-
-            return
-
-        queue.put_nowait(message)
-
-        if message.is_terminal:
-            del self._queues[message.request_id]
-
-    def drain(self) -> None:
-        queues = self._queues
-        self._queues = {}
-
-        for request_id, queue in queues.items():
-            queue.put_nowait(ConnectionDroppedError(request_id))
-
-    def _fail(self, request_id: str | None, error: Exception) -> None:
-        queue = self._queues.pop(request_id, None) if request_id else None
-
-        if queue is not None:
-            queue.put_nowait(error)
diff --git a/paddler_client_python/paddler_client/pooling_type.py b/paddler_client_python/paddler_client/pooling_type.py
deleted file mode 100644
index 8e8047c38..000000000
--- a/paddler_client_python/paddler_client/pooling_type.py
+++ /dev/null
@@ -1,10 +0,0 @@
-from enum import StrEnum
-
-
-class PoolingType(StrEnum):
-    UNSPECIFIED = "Unspecified"
-    NONE = "None"
-    MEAN = "Mean"
-    CLS = "Cls"
-    LAST = "Last"
-    RANK = "Rank"
diff --git a/paddler_client_python/paddler_client/raise_for_streaming_error.py b/paddler_client_python/paddler_client/raise_for_streaming_error.py
deleted file mode 100644
index f15b4018b..000000000
--- a/paddler_client_python/paddler_client/raise_for_streaming_error.py
+++ /dev/null
@@ -1,18 +0,0 @@
-from __future__ import annotations
-
-from typing import TYPE_CHECKING
-
-from paddler_client.error import HttpError
-
-if TYPE_CHECKING:
-    import httpx
-
-
-async def raise_for_streaming_error(response: httpx.Response) -> None:
-    if not response.is_success:
-        body = await response.aread()
-
-        raise HttpError(
-            response.status_code,
-            body.decode("utf-8", errors="replace"),
-        )
diff --git a/paddler_client_python/paddler_client/raw_tool_call_tokens.py b/paddler_client_python/paddler_client/raw_tool_call_tokens.py
deleted file mode 100644
index 37f4404cf..000000000
--- a/paddler_client_python/paddler_client/raw_tool_call_tokens.py
+++ /dev/null
@@ -1,17 +0,0 @@
-from __future__ import annotations
-
-from dataclasses import dataclass
-from typing import Any
-
-
-@dataclass(frozen=True)
-class RawToolCallTokens:
-    text: str
-    ffi_error_message: str
-
-    @classmethod
-    def from_dict(cls, data: dict[str, Any]) -> RawToolCallTokens:
-        return cls(
-            text=str(data["text"]),
-            ffi_error_message=str(data["ffi_error_message"]),
-        )
diff --git a/paddler_client_python/paddler_client/response_stream.py b/paddler_client_python/paddler_client/response_stream.py
deleted file mode 100644
index 0443a0a1f..000000000
--- a/paddler_client_python/paddler_client/response_stream.py
+++ /dev/null
@@ -1,35 +0,0 @@
-from __future__ import annotations
-
-from typing import TYPE_CHECKING
-
-if TYPE_CHECKING:
-    import asyncio
-
-    from paddler_client.inference_message import InferenceMessage
-
-
-class ResponseStream:
-    def __init__(
-        self,
-        queue: asyncio.Queue[InferenceMessage | Exception],
-    ) -> None:
-        self._queue = queue
-        self._done = False
-
-    def __aiter__(self) -> ResponseStream:
-        return self
-
-    async def __anext__(self) -> InferenceMessage:
-        if self._done:
-            raise StopAsyncIteration
-
-        item = await self._queue.get()
-
-        if isinstance(item, Exception):
-            self._done = True
-            raise item
-
-        if item.is_terminal:
-            self._done = True
-
-        return item
diff --git a/paddler_client_python/paddler_client/stream_ndjson.py b/paddler_client_python/paddler_client/stream_ndjson.py
deleted file mode 100644
index f3c498b58..000000000
--- a/paddler_client_python/paddler_client/stream_ndjson.py
+++ /dev/null
@@ -1,22 +0,0 @@
-from __future__ import annotations
-
-from typing import TYPE_CHECKING
-
-from paddler_client.ndjson_inference_messages import ndjson_inference_messages
-from paddler_client.raise_for_streaming_error import raise_for_streaming_error
-
-if TYPE_CHECKING:
-    from collections.abc import AsyncGenerator
-
-    import httpx
-
-    from paddler_client.inference_message import InferenceMessage
-
-
-async def stream_ndjson_inference_messages(
-    response: httpx.Response,
-) -> AsyncGenerator[InferenceMessage, None]:
-    await raise_for_streaming_error(response)
-
-    async for message in ndjson_inference_messages(response.aiter_lines()):
-        yield message
diff --git a/paddler_client_python/paddler_client/stream_sse.py b/paddler_client_python/paddler_client/stream_sse.py
deleted file mode 100644
index 7c203104c..000000000
--- a/paddler_client_python/paddler_client/stream_sse.py
+++ /dev/null
@@ -1,22 +0,0 @@
-from __future__ import annotations
-
-from typing import TYPE_CHECKING
-
-from paddler_client.raise_for_streaming_error import raise_for_streaming_error
-
-if TYPE_CHECKING:
-    from collections.abc import AsyncGenerator
-
-    import httpx
-
-
-async def stream_sse(
-    response: httpx.Response,
-) -> AsyncGenerator[str, None]:
-    await raise_for_streaming_error(response)
-
-    async for raw_line in response.aiter_lines():
-        stripped_line = raw_line.rstrip("\r")
-
-        if stripped_line.startswith("data: "):
-            yield stripped_line[6:]
diff --git a/paddler_client_python/paddler_client/tool.py b/paddler_client_python/paddler_client/tool.py
deleted file mode 100644
index 9c4e07e9b..000000000
--- a/paddler_client_python/paddler_client/tool.py
+++ /dev/null
@@ -1,18 +0,0 @@
-from typing import Literal
-
-from pydantic import BaseModel
-
-from paddler_client.validated_parameters_schema import (
-    ValidatedParametersSchema,
-)
-
-
-class Function(BaseModel):
-    name: str
-    description: str
-    parameters: ValidatedParametersSchema | None = None
-
-
-class Tool(BaseModel):
-    type: Literal["function"] = "function"
-    function: Function
diff --git a/paddler_client_python/paddler_client/tool_call_arguments.py b/paddler_client_python/paddler_client/tool_call_arguments.py
deleted file mode 100644
index 2e90aae3a..000000000
--- a/paddler_client_python/paddler_client/tool_call_arguments.py
+++ /dev/null
@@ -1,27 +0,0 @@
-from __future__ import annotations
-
-from dataclasses import dataclass
-from typing import Any
-
-from paddler_client.error import UnknownToolCallArgumentsError
-
-
-@dataclass(frozen=True)
-class ValidJson:
-    value: Any
-
-
-@dataclass(frozen=True)
-class InvalidJson:
-    raw: str
-
-
-ToolCallArguments = ValidJson | InvalidJson
-
-
-def parse_tool_call_arguments(payload: dict[str, Any]) -> ToolCallArguments:
-    if "ValidJson" in payload:
-        return ValidJson(payload["ValidJson"])
-    if "InvalidJson" in payload:
-        return InvalidJson(str(payload["InvalidJson"]))
-    raise UnknownToolCallArgumentsError(payload)
diff --git a/paddler_client_python/paddler_client/unparsed_message_request_id.py b/paddler_client_python/paddler_client/unparsed_message_request_id.py
deleted file mode 100644
index 776ca0ef8..000000000
--- a/paddler_client_python/paddler_client/unparsed_message_request_id.py
+++ /dev/null
@@ -1,16 +0,0 @@
-from __future__ import annotations
-
-from typing import Any
-
-
-def unparsed_message_request_id(data: Any) -> str | None:
-    if not isinstance(data, dict):
-        return None
-
-    for envelope_name in ("Response", "Error"):
-        envelope: Any = data.get(envelope_name)
-
-        if isinstance(envelope, dict) and "request_id" in envelope:
-            return str(envelope["request_id"])
-
-    return None
diff --git a/paddler_client_python/paddler_client/url_model_reference.py b/paddler_client_python/paddler_client/url_model_reference.py
deleted file mode 100644
index a760f80c9..000000000
--- a/paddler_client_python/paddler_client/url_model_reference.py
+++ /dev/null
@@ -1,5 +0,0 @@
-from pydantic import BaseModel
-
-
-class UrlModelReference(BaseModel):
-    url: str
diff --git a/paddler_client_python/paddler_client/validated_parameters_schema.py b/paddler_client_python/paddler_client/validated_parameters_schema.py
deleted file mode 100644
index 8ddf006eb..000000000
--- a/paddler_client_python/paddler_client/validated_parameters_schema.py
+++ /dev/null
@@ -1,15 +0,0 @@
-from typing import Any
-
-from pydantic import BaseModel, ConfigDict, Field
-
-
-class ValidatedParametersSchema(BaseModel):
-    model_config = ConfigDict(populate_by_name=True)
-
-    schema_type: str = Field(alias="type")
-    properties: dict[str, Any] | None = None
-    required: list[str] | None = None
-    additional_properties: Any | None = Field(
-        default=None,
-        alias="additionalProperties",
-    )
diff --git a/paddler_client_python/poetry.lock b/paddler_client_python/poetry.lock
deleted file mode 100644
index da2abf71b..000000000
--- a/paddler_client_python/poetry.lock
+++ /dev/null
@@ -1,904 +0,0 @@
-# This file is automatically @generated by Poetry 2.4.3 and should not be changed by hand.
-
-[[package]]
-name = "annotated-types"
-version = "0.7.0"
-description = "Reusable constraint types to use with typing.Annotated"
-optional = false
-python-versions = ">=3.8"
-groups = ["main"]
-files = [
-    {file = "annotated_types-0.7.0-py3-none-any.whl", hash = "sha256:1f02e8b43a8fbbc3f3e0d4f0f4bfc8131bcb4eebe8849b8e5c773f3a1c582a53"},
-    {file = "annotated_types-0.7.0.tar.gz", hash = "sha256:aff07c09a53a08bc8cfccb9c85b05f1aa9a2a6f23728d790723543408344ce89"},
-]
-
-[[package]]
-name = "anyio"
-version = "4.15.1"
-description = "High-level concurrency and networking framework on top of asyncio or Trio"
-optional = false
-python-versions = ">=3.10"
-groups = ["main"]
-files = [
-    {file = "anyio-4.15.1-py3-none-any.whl", hash = "sha256:6152fdbbf9a77fdec97731721bebf7c4c44f7c29b424b0065826173efc7ed101"},
-    {file = "anyio-4.15.1.tar.gz", hash = "sha256:9f28306018cbd6d329e64a36d58256edff76dd996fe423bc957326e578b82a94"},
-]
-
-[package.dependencies]
-idna = ">=2.8"
-typing_extensions = {version = ">=4.16.0", markers = "python_version < \"3.15\""}
-
-[package.extras]
-trio = ["trio (>=0.32.0)"]
-
-[[package]]
-name = "certifi"
-version = "2026.2.25"
-description = "Python package for providing Mozilla's CA Bundle."
-optional = false
-python-versions = ">=3.7"
-groups = ["main"]
-files = [
-    {file = "certifi-2026.2.25-py3-none-any.whl", hash = "sha256:027692e4402ad994f1c42e52a4997a9763c646b73e4096e4d5d6db8af1d6f0fa"},
-    {file = "certifi-2026.2.25.tar.gz", hash = "sha256:e887ab5cee78ea814d3472169153c2d12cd43b14bd03329a39a9c6e2e80bfba7"},
-]
-
-[[package]]
-name = "colorama"
-version = "0.4.6"
-description = "Cross-platform colored terminal text."
-optional = false
-python-versions = "!=3.0.*,!=3.1.*,!=3.2.*,!=3.3.*,!=3.4.*,!=3.5.*,!=3.6.*,>=2.7"
-groups = ["dev"]
-markers = "sys_platform == \"win32\""
-files = [
-    {file = "colorama-0.4.6-py2.py3-none-any.whl", hash = "sha256:4f1d9991f5acc0ca119f9d443620b77f9d6b33703e51011c16baf57afb285fc6"},
-    {file = "colorama-0.4.6.tar.gz", hash = "sha256:08695f5cb7ed6e0531a20572697297273c47b8cae5a63ffc6d6ed5c201be6e44"},
-]
-
-[[package]]
-name = "coverage"
-version = "7.13.5"
-description = "Code coverage measurement for Python"
-optional = false
-python-versions = ">=3.10"
-groups = ["dev"]
-files = [
-    {file = "coverage-7.13.5-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:e0723d2c96324561b9aa76fb982406e11d93cdb388a7a7da2b16e04719cf7ca5"},
-    {file = "coverage-7.13.5-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:52f444e86475992506b32d4e5ca55c24fc88d73bcbda0e9745095b28ef4dc0cf"},
-    {file = "coverage-7.13.5-cp310-cp310-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:704de6328e3d612a8f6c07000a878ff38181ec3263d5a11da1db294fa6a9bdf8"},
-    {file = "coverage-7.13.5-cp310-cp310-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:a1a6d79a14e1ec1832cabc833898636ad5f3754a678ef8bb4908515208bf84f4"},
-    {file = "coverage-7.13.5-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:79060214983769c7ba3f0cee10b54c97609dca4d478fa1aa32b914480fd5738d"},
-    {file = "coverage-7.13.5-cp310-cp310-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:356e76b46783a98c2a2fe81ec79df4883a1e62895ea952968fb253c114e7f930"},
-    {file = "coverage-7.13.5-cp310-cp310-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:0cef0cdec915d11254a7f549c1170afecce708d30610c6abdded1f74e581666d"},
-    {file = "coverage-7.13.5-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:dc022073d063b25a402454e5712ef9e007113e3a676b96c5f29b2bda29352f40"},
-    {file = "coverage-7.13.5-cp310-cp310-musllinux_1_2_i686.whl", hash = "sha256:9b74db26dfea4f4e50d48a4602207cd1e78be33182bc9cbf22da94f332f99878"},
-    {file = "coverage-7.13.5-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:ad146744ca4fd09b50c482650e3c1b1f4dfa1d4792e0a04a369c7f23336f0400"},
-    {file = "coverage-7.13.5-cp310-cp310-musllinux_1_2_riscv64.whl", hash = "sha256:c555b48be1853fe3997c11c4bd521cdd9a9612352de01fa4508f16ec341e6fe0"},
-    {file = "coverage-7.13.5-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:7034b5c56a58ae5e85f23949d52c14aca2cfc6848a31764995b7de88f13a1ea0"},
-    {file = "coverage-7.13.5-cp310-cp310-win32.whl", hash = "sha256:eb7fdf1ef130660e7415e0253a01a7d5a88c9c4d158bcf75cbbd922fd65a5b58"},
-    {file = "coverage-7.13.5-cp310-cp310-win_amd64.whl", hash = "sha256:3e1bb5f6c78feeb1be3475789b14a0f0a5b47d505bfc7267126ccbd50289999e"},
-    {file = "coverage-7.13.5-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:66a80c616f80181f4d643b0f9e709d97bcea413ecd9631e1dedc7401c8e6695d"},
-    {file = "coverage-7.13.5-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:145ede53ccbafb297c1c9287f788d1bc3efd6c900da23bf6931b09eafc931587"},
-    {file = "coverage-7.13.5-cp311-cp311-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:0672854dc733c342fa3e957e0605256d2bf5934feeac328da9e0b5449634a642"},
-    {file = "coverage-7.13.5-cp311-cp311-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:ec10e2a42b41c923c2209b846126c6582db5e43a33157e9870ba9fb70dc7854b"},
-    {file = "coverage-7.13.5-cp311-cp311-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:be3d4bbad9d4b037791794ddeedd7d64a56f5933a2c1373e18e9e568b9141686"},
-    {file = "coverage-7.13.5-cp311-cp311-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:4d2afbc5cc54d286bfb54541aa50b64cdb07a718227168c87b9e2fb8f25e1743"},
-    {file = "coverage-7.13.5-cp311-cp311-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:3ad050321264c49c2fa67bb599100456fc51d004b82534f379d16445da40fb75"},
-    {file = "coverage-7.13.5-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:7300c8a6d13335b29bb76d7651c66af6bd8658517c43499f110ddc6717bfc209"},
-    {file = "coverage-7.13.5-cp311-cp311-musllinux_1_2_i686.whl", hash = "sha256:eb07647a5738b89baab047f14edd18ded523de60f3b30e75c2acc826f79c839a"},
-    {file = "coverage-7.13.5-cp311-cp311-musllinux_1_2_ppc64le.whl", hash = "sha256:9adb6688e3b53adffefd4a52d72cbd8b02602bfb8f74dcd862337182fd4d1a4e"},
-    {file = "coverage-7.13.5-cp311-cp311-musllinux_1_2_riscv64.whl", hash = "sha256:7c8d4bc913dd70b93488d6c496c77f3aff5ea99a07e36a18f865bca55adef8bd"},
-    {file = "coverage-7.13.5-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:0e3c426ffc4cd952f54ee9ffbdd10345709ecc78a3ecfd796a57236bfad0b9b8"},
-    {file = "coverage-7.13.5-cp311-cp311-win32.whl", hash = "sha256:259b69bb83ad9894c4b25be2528139eecba9a82646ebdda2d9db1ba28424a6bf"},
-    {file = "coverage-7.13.5-cp311-cp311-win_amd64.whl", hash = "sha256:258354455f4e86e3e9d0d17571d522e13b4e1e19bf0f8596bcf9476d61e7d8a9"},
-    {file = "coverage-7.13.5-cp311-cp311-win_arm64.whl", hash = "sha256:bff95879c33ec8da99fc9b6fe345ddb5be6414b41d6d1ad1c8f188d26f36e028"},
-    {file = "coverage-7.13.5-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:460cf0114c5016fa841214ff5564aa4864f11948da9440bc97e21ad1f4ba1e01"},
-    {file = "coverage-7.13.5-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:0e223ce4b4ed47f065bfb123687686512e37629be25cc63728557ae7db261422"},
-    {file = "coverage-7.13.5-cp312-cp312-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:6e3370441f4513c6252bf042b9c36d22491142385049243253c7e48398a15a9f"},
-    {file = "coverage-7.13.5-cp312-cp312-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:03ccc709a17a1de074fb1d11f217342fb0d2b1582ed544f554fc9fc3f07e95f5"},
-    {file = "coverage-7.13.5-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:3f4818d065964db3c1c66dc0fbdac5ac692ecbc875555e13374fdbe7eedb4376"},
-    {file = "coverage-7.13.5-cp312-cp312-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:012d5319e66e9d5a218834642d6c35d265515a62f01157a45bcc036ecf947256"},
-    {file = "coverage-7.13.5-cp312-cp312-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:8dd02af98971bdb956363e4827d34425cb3df19ee550ef92855b0acb9c7ce51c"},
-    {file = "coverage-7.13.5-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:f08fd75c50a760c7eb068ae823777268daaf16a80b918fa58eea888f8e3919f5"},
-    {file = "coverage-7.13.5-cp312-cp312-musllinux_1_2_i686.whl", hash = "sha256:843ea8643cf967d1ac7e8ecd4bb00c99135adf4816c0c0593fdcc47b597fcf09"},
-    {file = "coverage-7.13.5-cp312-cp312-musllinux_1_2_ppc64le.whl", hash = "sha256:9d44d7aa963820b1b971dbecd90bfe5fe8f81cff79787eb6cca15750bd2f79b9"},
-    {file = "coverage-7.13.5-cp312-cp312-musllinux_1_2_riscv64.whl", hash = "sha256:7132bed4bd7b836200c591410ae7d97bf7ae8be6fc87d160b2bd881df929e7bf"},
-    {file = "coverage-7.13.5-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:a698e363641b98843c517817db75373c83254781426e94ada3197cabbc2c919c"},
-    {file = "coverage-7.13.5-cp312-cp312-win32.whl", hash = "sha256:bdba0a6b8812e8c7df002d908a9a2ea3c36e92611b5708633c50869e6d922fdf"},
-    {file = "coverage-7.13.5-cp312-cp312-win_amd64.whl", hash = "sha256:d2c87e0c473a10bffe991502eac389220533024c8082ec1ce849f4218dded810"},
-    {file = "coverage-7.13.5-cp312-cp312-win_arm64.whl", hash = "sha256:bf69236a9a81bdca3bff53796237aab096cdbf8d78a66ad61e992d9dac7eb2de"},
-    {file = "coverage-7.13.5-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:5ec4af212df513e399cf11610cc27063f1586419e814755ab362e50a85ea69c1"},
-    {file = "coverage-7.13.5-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:941617e518602e2d64942c88ec8499f7fbd49d3f6c4327d3a71d43a1973032f3"},
-    {file = "coverage-7.13.5-cp313-cp313-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:da305e9937617ee95c2e39d8ff9f040e0487cbf1ac174f777ed5eddd7a7c1f26"},
-    {file = "coverage-7.13.5-cp313-cp313-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:78e696e1cc714e57e8b25760b33a8b1026b7048d270140d25dafe1b0a1ee05a3"},
-    {file = "coverage-7.13.5-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:02ca0eed225b2ff301c474aeeeae27d26e2537942aa0f87491d3e147e784a82b"},
-    {file = "coverage-7.13.5-cp313-cp313-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:04690832cbea4e4663d9149e05dba142546ca05cb1848816760e7f58285c970a"},
-    {file = "coverage-7.13.5-cp313-cp313-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:0590e44dd2745c696a778f7bab6aa95256de2cbc8b8cff4f7db8ff09813d6969"},
-    {file = "coverage-7.13.5-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:d7cfad2d6d81dd298ab6b89fe72c3b7b05ec7544bdda3b707ddaecff8d25c161"},
-    {file = "coverage-7.13.5-cp313-cp313-musllinux_1_2_i686.whl", hash = "sha256:e092b9499de38ae0fbfbc603a74660eb6ff3e869e507b50d85a13b6db9863e15"},
-    {file = "coverage-7.13.5-cp313-cp313-musllinux_1_2_ppc64le.whl", hash = "sha256:48c39bc4a04d983a54a705a6389512883d4a3b9862991b3617d547940e9f52b1"},
-    {file = "coverage-7.13.5-cp313-cp313-musllinux_1_2_riscv64.whl", hash = "sha256:2d3807015f138ffea1ed9afeeb8624fd781703f2858b62a8dd8da5a0994c57b6"},
-    {file = "coverage-7.13.5-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:ee2aa19e03161671ec964004fb74b2257805d9710bf14a5c704558b9d8dbaf17"},
-    {file = "coverage-7.13.5-cp313-cp313-win32.whl", hash = "sha256:ce1998c0483007608c8382f4ff50164bfc5bd07a2246dd272aa4043b75e61e85"},
-    {file = "coverage-7.13.5-cp313-cp313-win_amd64.whl", hash = "sha256:631efb83f01569670a5e866ceb80fe483e7c159fac6f167e6571522636104a0b"},
-    {file = "coverage-7.13.5-cp313-cp313-win_arm64.whl", hash = "sha256:f4cd16206ad171cbc2470dbea9103cf9a7607d5fe8c242fdf1edf36174020664"},
-    {file = "coverage-7.13.5-cp313-cp313t-macosx_10_13_x86_64.whl", hash = "sha256:0428cbef5783ad91fe240f673cc1f76b25e74bbfe1a13115e4aa30d3f538162d"},
-    {file = "coverage-7.13.5-cp313-cp313t-macosx_11_0_arm64.whl", hash = "sha256:e0b216a19534b2427cc201a26c25da4a48633f29a487c61258643e89d28200c0"},
-    {file = "coverage-7.13.5-cp313-cp313t-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:972a9cd27894afe4bc2b1480107054e062df08e671df7c2f18c205e805ccd806"},
-    {file = "coverage-7.13.5-cp313-cp313t-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:4b59148601efcd2bac8c4dbf1f0ad6391693ccf7a74b8205781751637076aee3"},
-    {file = "coverage-7.13.5-cp313-cp313t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:505d7083c8b0c87a8fa8c07370c285847c1f77739b22e299ad75a6af6c32c5c9"},
-    {file = "coverage-7.13.5-cp313-cp313t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:60365289c3741e4db327e7baff2a4aaacf22f788e80fa4683393891b70a89fbd"},
-    {file = "coverage-7.13.5-cp313-cp313t-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:1b88c69c8ef5d4b6fe7dea66d6636056a0f6a7527c440e890cf9259011f5e606"},
-    {file = "coverage-7.13.5-cp313-cp313t-musllinux_1_2_aarch64.whl", hash = "sha256:5b13955d31d1633cf9376908089b7cebe7d15ddad7aeaabcbe969a595a97e95e"},
-    {file = "coverage-7.13.5-cp313-cp313t-musllinux_1_2_i686.whl", hash = "sha256:f70c9ab2595c56f81a89620e22899eea8b212a4041bd728ac6f4a28bf5d3ddd0"},
-    {file = "coverage-7.13.5-cp313-cp313t-musllinux_1_2_ppc64le.whl", hash = "sha256:084b84a8c63e8d6fc7e3931b316a9bcafca1458d753c539db82d31ed20091a87"},
-    {file = "coverage-7.13.5-cp313-cp313t-musllinux_1_2_riscv64.whl", hash = "sha256:ad14385487393e386e2ea988b09d62dd42c397662ac2dabc3832d71253eee479"},
-    {file = "coverage-7.13.5-cp313-cp313t-musllinux_1_2_x86_64.whl", hash = "sha256:7f2c47b36fe7709a6e83bfadf4eefb90bd25fbe4014d715224c4316f808e59a2"},
-    {file = "coverage-7.13.5-cp313-cp313t-win32.whl", hash = "sha256:67e9bc5449801fad0e5dff329499fb090ba4c5800b86805c80617b4e29809b2a"},
-    {file = "coverage-7.13.5-cp313-cp313t-win_amd64.whl", hash = "sha256:da86cdcf10d2519e10cabb8ac2de03da1bcb6e4853790b7fbd48523332e3a819"},
-    {file = "coverage-7.13.5-cp313-cp313t-win_arm64.whl", hash = "sha256:0ecf12ecb326fe2c339d93fc131816f3a7367d223db37817208905c89bded911"},
-    {file = "coverage-7.13.5-cp314-cp314-macosx_10_15_x86_64.whl", hash = "sha256:fbabfaceaeb587e16f7008f7795cd80d20ec548dc7f94fbb0d4ec2e038ce563f"},
-    {file = "coverage-7.13.5-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:9bb2a28101a443669a423b665939381084412b81c3f8c0fcfbac57f4e30b5b8e"},
-    {file = "coverage-7.13.5-cp314-cp314-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:bd3a2fbc1c6cccb3c5106140d87cc6a8715110373ef42b63cf5aea29df8c217a"},
-    {file = "coverage-7.13.5-cp314-cp314-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:6c36ddb64ed9d7e496028d1d00dfec3e428e0aabf4006583bb1839958d280510"},
-    {file = "coverage-7.13.5-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:380e8e9084d8eb38db3a9176a1a4f3c0082c3806fa0dc882d1d87abc3c789247"},
-    {file = "coverage-7.13.5-cp314-cp314-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:e808af52a0513762df4d945ea164a24b37f2f518cbe97e03deaa0ee66139b4d6"},
-    {file = "coverage-7.13.5-cp314-cp314-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:e301d30dd7e95ae068671d746ba8c34e945a82682e62918e41b2679acd2051a0"},
-    {file = "coverage-7.13.5-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:800bc829053c80d240a687ceeb927a94fd108bbdc68dfbe505d0d75ab578a882"},
-    {file = "coverage-7.13.5-cp314-cp314-musllinux_1_2_i686.whl", hash = "sha256:0b67af5492adb31940ee418a5a655c28e48165da5afab8c7fa6fd72a142f8740"},
-    {file = "coverage-7.13.5-cp314-cp314-musllinux_1_2_ppc64le.whl", hash = "sha256:c9136ff29c3a91e25b1d1552b5308e53a1e0653a23e53b6366d7c2dcbbaf8a16"},
-    {file = "coverage-7.13.5-cp314-cp314-musllinux_1_2_riscv64.whl", hash = "sha256:cff784eef7f0b8f6cb28804fbddcfa99f89efe4cc35fb5627e3ac58f91ed3ac0"},
-    {file = "coverage-7.13.5-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:68a4953be99b17ac3c23b6efbc8a38330d99680c9458927491d18700ef23ded0"},
-    {file = "coverage-7.13.5-cp314-cp314-win32.whl", hash = "sha256:35a31f2b1578185fbe6aa2e74cea1b1d0bbf4c552774247d9160d29b80ed56cc"},
-    {file = "coverage-7.13.5-cp314-cp314-win_amd64.whl", hash = "sha256:2aa055ae1857258f9e0045be26a6d62bdb47a72448b62d7b55f4820f361a2633"},
-    {file = "coverage-7.13.5-cp314-cp314-win_arm64.whl", hash = "sha256:1b11eef33edeae9d142f9b4358edb76273b3bfd30bc3df9a4f95d0e49caf94e8"},
-    {file = "coverage-7.13.5-cp314-cp314t-macosx_10_15_x86_64.whl", hash = "sha256:10a0c37f0b646eaff7cce1874c31d1f1ccb297688d4c747291f4f4c70741cc8b"},
-    {file = "coverage-7.13.5-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:b5db73ba3c41c7008037fa731ad5459fc3944cb7452fc0aa9f822ad3533c583c"},
-    {file = "coverage-7.13.5-cp314-cp314t-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:750db93a81e3e5a9831b534be7b1229df848b2e125a604fe6651e48aa070e5f9"},
-    {file = "coverage-7.13.5-cp314-cp314t-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:9ddb4f4a5479f2539644be484da179b653273bca1a323947d48ab107b3ed1f29"},
-    {file = "coverage-7.13.5-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:d8a7a2049c14f413163e2bdabd37e41179b1d1ccb10ffc6ccc4b7a718429c607"},
-    {file = "coverage-7.13.5-cp314-cp314t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:e1c85e0b6c05c592ea6d8768a66a254bfb3874b53774b12d4c89c481eb78cb90"},
-    {file = "coverage-7.13.5-cp314-cp314t-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:777c4d1eff1b67876139d24288aaf1817f6c03d6bae9c5cc8d27b83bcfe38fe3"},
-    {file = "coverage-7.13.5-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:6697e29b93707167687543480a40f0db8f356e86d9f67ddf2e37e2dfd91a9dab"},
-    {file = "coverage-7.13.5-cp314-cp314t-musllinux_1_2_i686.whl", hash = "sha256:8fdf453a942c3e4d99bd80088141c4c6960bb232c409d9c3558e2dbaa3998562"},
-    {file = "coverage-7.13.5-cp314-cp314t-musllinux_1_2_ppc64le.whl", hash = "sha256:32ca0c0114c9834a43f045a87dcebd69d108d8ffb666957ea65aa132f50332e2"},
-    {file = "coverage-7.13.5-cp314-cp314t-musllinux_1_2_riscv64.whl", hash = "sha256:8769751c10f339021e2638cd354e13adeac54004d1941119b2c96fe5276d45ea"},
-    {file = "coverage-7.13.5-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:cec2d83125531bd153175354055cdb7a09987af08a9430bd173c937c6d0fba2a"},
-    {file = "coverage-7.13.5-cp314-cp314t-win32.whl", hash = "sha256:0cd9ed7a8b181775459296e402ca4fb27db1279740a24e93b3b41942ebe4b215"},
-    {file = "coverage-7.13.5-cp314-cp314t-win_amd64.whl", hash = "sha256:301e3b7dfefecaca37c9f1aa6f0049b7d4ab8dd933742b607765d757aca77d43"},
-    {file = "coverage-7.13.5-cp314-cp314t-win_arm64.whl", hash = "sha256:9dacc2ad679b292709e0f5fc1ac74a6d4d5562e424058962c7bb0c658ad25e45"},
-    {file = "coverage-7.13.5-py3-none-any.whl", hash = "sha256:34b02417cf070e173989b3db962f7ed56d2f644307b2cf9d5a0f258e13084a61"},
-    {file = "coverage-7.13.5.tar.gz", hash = "sha256:c81f6515c4c40141f83f502b07bbfa5c240ba25bbe73da7b33f1e5b6120ff179"},
-]
-
-[package.extras]
-toml = ["tomli ; python_full_version <= \"3.11.0a6\""]
-
-[[package]]
-name = "h11"
-version = "0.16.0"
-description = "A pure-Python, bring-your-own-I/O implementation of HTTP/1.1"
-optional = false
-python-versions = ">=3.8"
-groups = ["main"]
-files = [
-    {file = "h11-0.16.0-py3-none-any.whl", hash = "sha256:63cf8bbe7522de3bf65932fda1d9c2772064ffb3dae62d55932da54b31cb6c86"},
-    {file = "h11-0.16.0.tar.gz", hash = "sha256:4e35b956cf45792e4caa5885e69fba00bdbc6ffafbfa020300e549b208ee5ff1"},
-]
-
-[[package]]
-name = "httpcore"
-version = "1.0.9"
-description = "A minimal low-level HTTP client."
-optional = false
-python-versions = ">=3.8"
-groups = ["main"]
-files = [
-    {file = "httpcore-1.0.9-py3-none-any.whl", hash = "sha256:2d400746a40668fc9dec9810239072b40b4484b640a8c38fd654a024c7a1bf55"},
-    {file = "httpcore-1.0.9.tar.gz", hash = "sha256:6e34463af53fd2ab5d807f399a9b45ea31c3dfa2276f15a2c3f00afff6e176e8"},
-]
-
-[package.dependencies]
-certifi = "*"
-h11 = ">=0.16"
-
-[package.extras]
-asyncio = ["anyio (>=4.0,<5.0)"]
-http2 = ["h2 (>=3,<5)"]
-socks = ["socksio (==1.*)"]
-trio = ["trio (>=0.22.0,<1.0)"]
-
-[[package]]
-name = "httpx"
-version = "0.28.1"
-description = "The next generation HTTP client."
-optional = false
-python-versions = ">=3.8"
-groups = ["main"]
-files = [
-    {file = "httpx-0.28.1-py3-none-any.whl", hash = "sha256:d909fcccc110f8c7faf814ca82a9a4d816bc5a6dbfea25d6591d6985b8ba59ad"},
-    {file = "httpx-0.28.1.tar.gz", hash = "sha256:75e98c5f16b0f35b567856f597f06ff2270a374470a5c2392242528e3e3e42fc"},
-]
-
-[package.dependencies]
-anyio = "*"
-certifi = "*"
-httpcore = "==1.*"
-idna = "*"
-
-[package.extras]
-brotli = ["brotli ; platform_python_implementation == \"CPython\"", "brotlicffi ; platform_python_implementation != \"CPython\""]
-cli = ["click (==8.*)", "pygments (==2.*)", "rich (>=10,<14)"]
-http2 = ["h2 (>=3,<5)"]
-socks = ["socksio (==1.*)"]
-zstd = ["zstandard (>=0.18.0)"]
-
-[[package]]
-name = "idna"
-version = "3.20"
-description = "Internationalized Domain Names in Applications (IDNA)"
-optional = false
-python-versions = ">=3.9"
-groups = ["main"]
-files = [
-    {file = "idna-3.20-py3-none-any.whl", hash = "sha256:ab7ae7122974553370f0bdb919e1a960b2cd1bc1ef0276416d896db81c14582c"},
-    {file = "idna-3.20.tar.gz", hash = "sha256:a7db850025b95ded1eae8a46181a1a6c56c92c96f0e2b005d9ff8dc0210cab44"},
-]
-
-[package.extras]
-all = ["coverage (>=7.10.0)", "hypothesis (>=6.141.1)", "mypy (>=1.11.2)", "pytest (>=8.3.2)", "ruff (>=0.16.0)", "ty (>=0.0.37)"]
-
-[[package]]
-name = "iniconfig"
-version = "2.3.0"
-description = "brain-dead simple config-ini parsing"
-optional = false
-python-versions = ">=3.10"
-groups = ["dev"]
-files = [
-    {file = "iniconfig-2.3.0-py3-none-any.whl", hash = "sha256:f631c04d2c48c52b84d0d0549c99ff3859c98df65b3101406327ecc7d53fbf12"},
-    {file = "iniconfig-2.3.0.tar.gz", hash = "sha256:c76315c77db068650d49c5b56314774a7804df16fee4402c1f19d6d15d8c4730"},
-]
-
-[[package]]
-name = "librt"
-version = "0.8.1"
-description = "Mypyc runtime library"
-optional = false
-python-versions = ">=3.9"
-groups = ["dev"]
-markers = "platform_python_implementation != \"PyPy\""
-files = [
-    {file = "librt-0.8.1-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:81fd938344fecb9373ba1b155968c8a329491d2ce38e7ddb76f30ffb938f12dc"},
-    {file = "librt-0.8.1-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:5db05697c82b3a2ec53f6e72b2ed373132b0c2e05135f0696784e97d7f5d48e7"},
-    {file = "librt-0.8.1-cp310-cp310-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:d56bc4011975f7460bea7b33e1ff425d2f1adf419935ff6707273c77f8a4ada6"},
-    {file = "librt-0.8.1-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:5cdc0f588ff4b663ea96c26d2a230c525c6fc62b28314edaaaca8ed5af931ad0"},
-    {file = "librt-0.8.1-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:97c2b54ff6717a7a563b72627990bec60d8029df17df423f0ed37d56a17a176b"},
-    {file = "librt-0.8.1-cp310-cp310-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:8f1125e6bbf2f1657d9a2f3ccc4a2c9b0c8b176965bb565dd4d86be67eddb4b6"},
-    {file = "librt-0.8.1-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:8f4bb453f408137d7581be309b2fbc6868a80e7ef60c88e689078ee3a296ae71"},
-    {file = "librt-0.8.1-cp310-cp310-musllinux_1_2_i686.whl", hash = "sha256:c336d61d2fe74a3195edc1646d53ff1cddd3a9600b09fa6ab75e5514ba4862a7"},
-    {file = "librt-0.8.1-cp310-cp310-musllinux_1_2_riscv64.whl", hash = "sha256:eb5656019db7c4deacf0c1a55a898c5bb8f989be904597fcb5232a2f4828fa05"},
-    {file = "librt-0.8.1-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:c25d9e338d5bed46c1632f851babf3d13c78f49a225462017cf5e11e845c5891"},
-    {file = "librt-0.8.1-cp310-cp310-win32.whl", hash = "sha256:aaab0e307e344cb28d800957ef3ec16605146ef0e59e059a60a176d19543d1b7"},
-    {file = "librt-0.8.1-cp310-cp310-win_amd64.whl", hash = "sha256:56e04c14b696300d47b3bc5f1d10a00e86ae978886d0cee14e5714fafb5df5d2"},
-    {file = "librt-0.8.1-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:681dc2451d6d846794a828c16c22dc452d924e9f700a485b7ecb887a30aad1fd"},
-    {file = "librt-0.8.1-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:a3b4350b13cc0e6f5bec8fa7caf29a8fb8cdc051a3bae45cfbfd7ce64f009965"},
-    {file = "librt-0.8.1-cp311-cp311-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:ac1e7817fd0ed3d14fd7c5df91daed84c48e4c2a11ee99c0547f9f62fdae13da"},
-    {file = "librt-0.8.1-cp311-cp311-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:747328be0c5b7075cde86a0e09d7a9196029800ba75a1689332348e998fb85c0"},
-    {file = "librt-0.8.1-cp311-cp311-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:f0af2bd2bc204fa27f3d6711d0f360e6b8c684a035206257a81673ab924aa11e"},
-    {file = "librt-0.8.1-cp311-cp311-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:d480de377f5b687b6b1bc0c0407426da556e2a757633cc7e4d2e1a057aa688f3"},
-    {file = "librt-0.8.1-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:d0ee06b5b5291f609ddb37b9750985b27bc567791bc87c76a569b3feed8481ac"},
-    {file = "librt-0.8.1-cp311-cp311-musllinux_1_2_i686.whl", hash = "sha256:9e2c6f77b9ad48ce5603b83b7da9ee3e36b3ab425353f695cba13200c5d96596"},
-    {file = "librt-0.8.1-cp311-cp311-musllinux_1_2_riscv64.whl", hash = "sha256:439352ba9373f11cb8e1933da194dcc6206daf779ff8df0ed69c5e39113e6a99"},
-    {file = "librt-0.8.1-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:82210adabbc331dbb65d7868b105185464ef13f56f7f76688565ad79f648b0fe"},
-    {file = "librt-0.8.1-cp311-cp311-win32.whl", hash = "sha256:52c224e14614b750c0a6d97368e16804a98c684657c7518752c356834fff83bb"},
-    {file = "librt-0.8.1-cp311-cp311-win_amd64.whl", hash = "sha256:c00e5c884f528c9932d278d5c9cbbea38a6b81eb62c02e06ae53751a83a4d52b"},
-    {file = "librt-0.8.1-cp311-cp311-win_arm64.whl", hash = "sha256:f7cdf7f26c2286ffb02e46d7bac56c94655540b26347673bea15fa52a6af17e9"},
-    {file = "librt-0.8.1-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:a28f2612ab566b17f3698b0da021ff9960610301607c9a5e8eaca62f5e1c350a"},
-    {file = "librt-0.8.1-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:60a78b694c9aee2a0f1aaeaa7d101cf713e92e8423a941d2897f4fa37908dab9"},
-    {file = "librt-0.8.1-cp312-cp312-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:758509ea3f1eba2a57558e7e98f4659d0ea7670bff49673b0dde18a3c7e6c0eb"},
-    {file = "librt-0.8.1-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:039b9f2c506bd0ab0f8725aa5ba339c6f0cd19d3b514b50d134789809c24285d"},
-    {file = "librt-0.8.1-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:5bb54f1205a3a6ab41a6fd71dfcdcbd278670d3a90ca502a30d9da583105b6f7"},
-    {file = "librt-0.8.1-cp312-cp312-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:05bd41cdee35b0c59c259f870f6da532a2c5ca57db95b5f23689fcb5c9e42440"},
-    {file = "librt-0.8.1-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:adfab487facf03f0d0857b8710cf82d0704a309d8ffc33b03d9302b4c64e91a9"},
-    {file = "librt-0.8.1-cp312-cp312-musllinux_1_2_i686.whl", hash = "sha256:153188fe98a72f206042be10a2c6026139852805215ed9539186312d50a8e972"},
-    {file = "librt-0.8.1-cp312-cp312-musllinux_1_2_riscv64.whl", hash = "sha256:dd3c41254ee98604b08bd5b3af5bf0a89740d4ee0711de95b65166bf44091921"},
-    {file = "librt-0.8.1-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:e0d138c7ae532908cbb342162b2611dbd4d90c941cd25ab82084aaf71d2c0bd0"},
-    {file = "librt-0.8.1-cp312-cp312-win32.whl", hash = "sha256:43353b943613c5d9c49a25aaffdba46f888ec354e71e3529a00cca3f04d66a7a"},
-    {file = "librt-0.8.1-cp312-cp312-win_amd64.whl", hash = "sha256:ff8baf1f8d3f4b6b7257fcb75a501f2a5499d0dda57645baa09d4d0d34b19444"},
-    {file = "librt-0.8.1-cp312-cp312-win_arm64.whl", hash = "sha256:0f2ae3725904f7377e11cc37722d5d401e8b3d5851fb9273d7f4fe04f6b3d37d"},
-    {file = "librt-0.8.1-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:7e6bad1cd94f6764e1e21950542f818a09316645337fd5ab9a7acc45d99a8f35"},
-    {file = "librt-0.8.1-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:cf450f498c30af55551ba4f66b9123b7185362ec8b625a773b3d39aa1a717583"},
-    {file = "librt-0.8.1-cp313-cp313-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:eca45e982fa074090057132e30585a7e8674e9e885d402eae85633e9f449ce6c"},
-    {file = "librt-0.8.1-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:0c3811485fccfda840861905b8c70bba5ec094e02825598bb9d4ca3936857a04"},
-    {file = "librt-0.8.1-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:5e4af413908f77294605e28cfd98063f54b2c790561383971d2f52d113d9c363"},
-    {file = "librt-0.8.1-cp313-cp313-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:5212a5bd7fae98dae95710032902edcd2ec4dc994e883294f75c857b83f9aba0"},
-    {file = "librt-0.8.1-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:e692aa2d1d604e6ca12d35e51fdc36f4cda6345e28e36374579f7ef3611b3012"},
-    {file = "librt-0.8.1-cp313-cp313-musllinux_1_2_i686.whl", hash = "sha256:4be2a5c926b9770c9e08e717f05737a269b9d0ebc5d2f0060f0fe3fe9ce47acb"},
-    {file = "librt-0.8.1-cp313-cp313-musllinux_1_2_riscv64.whl", hash = "sha256:fd1a720332ea335ceb544cf0a03f81df92abd4bb887679fd1e460976b0e6214b"},
-    {file = "librt-0.8.1-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:93c2af9e01e0ef80d95ae3c720be101227edae5f2fe7e3dc63d8857fadfc5a1d"},
-    {file = "librt-0.8.1-cp313-cp313-win32.whl", hash = "sha256:086a32dbb71336627e78cc1d6ee305a68d038ef7d4c39aaff41ae8c9aa46e91a"},
-    {file = "librt-0.8.1-cp313-cp313-win_amd64.whl", hash = "sha256:e11769a1dbda4da7b00a76cfffa67aa47cfa66921d2724539eee4b9ede780b79"},
-    {file = "librt-0.8.1-cp313-cp313-win_arm64.whl", hash = "sha256:924817ab3141aca17893386ee13261f1d100d1ef410d70afe4389f2359fea4f0"},
-    {file = "librt-0.8.1-cp314-cp314-macosx_10_13_x86_64.whl", hash = "sha256:6cfa7fe54fd4d1f47130017351a959fe5804bda7a0bc7e07a2cdbc3fdd28d34f"},
-    {file = "librt-0.8.1-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:228c2409c079f8c11fb2e5d7b277077f694cb93443eb760e00b3b83cb8b3176c"},
-    {file = "librt-0.8.1-cp314-cp314-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:7aae78ab5e3206181780e56912d1b9bb9f90a7249ce12f0e8bf531d0462dd0fc"},
-    {file = "librt-0.8.1-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:172d57ec04346b047ca6af181e1ea4858086c80bdf455f61994c4aa6fc3f866c"},
-    {file = "librt-0.8.1-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:6b1977c4ea97ce5eb7755a78fae68d87e4102e4aaf54985e8b56806849cc06a3"},
-    {file = "librt-0.8.1-cp314-cp314-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:10c42e1f6fd06733ef65ae7bebce2872bcafd8d6e6b0a08fe0a05a23b044fb14"},
-    {file = "librt-0.8.1-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:4c8dfa264b9193c4ee19113c985c95f876fae5e51f731494fc4e0cf594990ba7"},
-    {file = "librt-0.8.1-cp314-cp314-musllinux_1_2_i686.whl", hash = "sha256:01170b6729a438f0dedc4a26ed342e3dc4f02d1000b4b19f980e1877f0c297e6"},
-    {file = "librt-0.8.1-cp314-cp314-musllinux_1_2_riscv64.whl", hash = "sha256:7b02679a0d783bdae30d443025b94465d8c3dc512f32f5b5031f93f57ac32071"},
-    {file = "librt-0.8.1-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:190b109bb69592a3401fe1ffdea41a2e73370ace2ffdc4a0e8e2b39cdea81b78"},
-    {file = "librt-0.8.1-cp314-cp314-win32.whl", hash = "sha256:e70a57ecf89a0f64c24e37f38d3fe217a58169d2fe6ed6d70554964042474023"},
-    {file = "librt-0.8.1-cp314-cp314-win_amd64.whl", hash = "sha256:7e2f3edca35664499fbb36e4770650c4bd4a08abc1f4458eab9df4ec56389730"},
-    {file = "librt-0.8.1-cp314-cp314-win_arm64.whl", hash = "sha256:0d2f82168e55ddefd27c01c654ce52379c0750ddc31ee86b4b266bcf4d65f2a3"},
-    {file = "librt-0.8.1-cp314-cp314t-macosx_10_13_x86_64.whl", hash = "sha256:2c74a2da57a094bd48d03fa5d196da83d2815678385d2978657499063709abe1"},
-    {file = "librt-0.8.1-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:a355d99c4c0d8e5b770313b8b247411ed40949ca44e33e46a4789b9293a907ee"},
-    {file = "librt-0.8.1-cp314-cp314t-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:2eb345e8b33fb748227409c9f1233d4df354d6e54091f0e8fc53acdb2ffedeb7"},
-    {file = "librt-0.8.1-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:9be2f15e53ce4e83cc08adc29b26fb5978db62ef2a366fbdf716c8a6c8901040"},
-    {file = "librt-0.8.1-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:785ae29c1f5c6e7c2cde2c7c0e148147f4503da3abc5d44d482068da5322fd9e"},
-    {file = "librt-0.8.1-cp314-cp314t-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:1d3a7da44baf692f0c6aeb5b2a09c5e6fc7a703bca9ffa337ddd2e2da53f7732"},
-    {file = "librt-0.8.1-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:5fc48998000cbc39ec0d5311312dda93ecf92b39aaf184c5e817d5d440b29624"},
-    {file = "librt-0.8.1-cp314-cp314t-musllinux_1_2_i686.whl", hash = "sha256:e96baa6820280077a78244b2e06e416480ed859bbd8e5d641cf5742919d8beb4"},
-    {file = "librt-0.8.1-cp314-cp314t-musllinux_1_2_riscv64.whl", hash = "sha256:31362dbfe297b23590530007062c32c6f6176f6099646bb2c95ab1b00a57c382"},
-    {file = "librt-0.8.1-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:cc3656283d11540ab0ea01978378e73e10002145117055e03722417aeab30994"},
-    {file = "librt-0.8.1-cp314-cp314t-win32.whl", hash = "sha256:738f08021b3142c2918c03692608baed43bc51144c29e35807682f8070ee2a3a"},
-    {file = "librt-0.8.1-cp314-cp314t-win_amd64.whl", hash = "sha256:89815a22daf9c51884fb5dbe4f1ef65ee6a146e0b6a8df05f753e2e4a9359bf4"},
-    {file = "librt-0.8.1-cp314-cp314t-win_arm64.whl", hash = "sha256:bf512a71a23504ed08103a13c941f763db13fb11177beb3d9244c98c29fb4a61"},
-    {file = "librt-0.8.1-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:3dff3d3ca8db20e783b1bc7de49c0a2ab0b8387f31236d6a026597d07fcd68ac"},
-    {file = "librt-0.8.1-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:08eec3a1fc435f0d09c87b6bf1ec798986a3544f446b864e4099633a56fcd9ed"},
-    {file = "librt-0.8.1-cp39-cp39-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:e3f0a41487fd5fad7e760b9e8a90e251e27c2816fbc2cff36a22a0e6bcbbd9dd"},
-    {file = "librt-0.8.1-cp39-cp39-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:bacdb58d9939d95cc557b4dbaa86527c9db2ac1ed76a18bc8d26f6dc8647d851"},
-    {file = "librt-0.8.1-cp39-cp39-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:b6d7ab1f01aa753188605b09a51faa44a3327400b00b8cce424c71910fc0a128"},
-    {file = "librt-0.8.1-cp39-cp39-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:4998009e7cb9e896569f4be7004f09d0ed70d386fa99d42b6d363f6d200501ac"},
-    {file = "librt-0.8.1-cp39-cp39-musllinux_1_2_aarch64.whl", hash = "sha256:2cc68eeeef5e906839c7bb0815748b5b0a974ec27125beefc0f942715785b551"},
-    {file = "librt-0.8.1-cp39-cp39-musllinux_1_2_i686.whl", hash = "sha256:0bf69d79a23f4f40b8673a947a234baeeb133b5078b483b7297c5916539cf5d5"},
-    {file = "librt-0.8.1-cp39-cp39-musllinux_1_2_riscv64.whl", hash = "sha256:22b46eabd76c1986ee7d231b0765ad387d7673bbd996aa0d0d054b38ac65d8f6"},
-    {file = "librt-0.8.1-cp39-cp39-musllinux_1_2_x86_64.whl", hash = "sha256:237796479f4d0637d6b9cbcb926ff424a97735e68ade6facf402df4ec93375ed"},
-    {file = "librt-0.8.1-cp39-cp39-win32.whl", hash = "sha256:4beb04b8c66c6ae62f8c1e0b2f097c1ebad9295c929a8d5286c05eae7c2fc7dc"},
-    {file = "librt-0.8.1-cp39-cp39-win_amd64.whl", hash = "sha256:64548cde61b692dc0dc379f4b5f59a2f582c2ebe7890d09c1ae3b9e66fa015b7"},
-    {file = "librt-0.8.1.tar.gz", hash = "sha256:be46a14693955b3bd96014ccbdb8339ee8c9346fbe11c1b78901b55125f14c73"},
-]
-
-[[package]]
-name = "mypy"
-version = "1.19.1"
-description = "Optional static typing for Python"
-optional = false
-python-versions = ">=3.9"
-groups = ["dev"]
-files = [
-    {file = "mypy-1.19.1-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:5f05aa3d375b385734388e844bc01733bd33c644ab48e9684faa54e5389775ec"},
-    {file = "mypy-1.19.1-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:022ea7279374af1a5d78dfcab853fe6a536eebfda4b59deab53cd21f6cd9f00b"},
-    {file = "mypy-1.19.1-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:ee4c11e460685c3e0c64a4c5de82ae143622410950d6be863303a1c4ba0e36d6"},
-    {file = "mypy-1.19.1-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:de759aafbae8763283b2ee5869c7255391fbc4de3ff171f8f030b5ec48381b74"},
-    {file = "mypy-1.19.1-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:ab43590f9cd5108f41aacf9fca31841142c786827a74ab7cc8a2eacb634e09a1"},
-    {file = "mypy-1.19.1-cp310-cp310-win_amd64.whl", hash = "sha256:2899753e2f61e571b3971747e302d5f420c3fd09650e1951e99f823bc3089dac"},
-    {file = "mypy-1.19.1-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:d8dfc6ab58ca7dda47d9237349157500468e404b17213d44fc1cb77bce532288"},
-    {file = "mypy-1.19.1-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:e3f276d8493c3c97930e354b2595a44a21348b320d859fb4a2b9f66da9ed27ab"},
-    {file = "mypy-1.19.1-cp311-cp311-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:2abb24cf3f17864770d18d673c85235ba52456b36a06b6afc1e07c1fdcd3d0e6"},
-    {file = "mypy-1.19.1-cp311-cp311-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:a009ffa5a621762d0c926a078c2d639104becab69e79538a494bcccb62cc0331"},
-    {file = "mypy-1.19.1-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:f7cee03c9a2e2ee26ec07479f38ea9c884e301d42c6d43a19d20fb014e3ba925"},
-    {file = "mypy-1.19.1-cp311-cp311-win_amd64.whl", hash = "sha256:4b84a7a18f41e167f7995200a1d07a4a6810e89d29859df936f1c3923d263042"},
-    {file = "mypy-1.19.1-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:a8174a03289288c1f6c46d55cef02379b478bfbc8e358e02047487cad44c6ca1"},
-    {file = "mypy-1.19.1-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:ffcebe56eb09ff0c0885e750036a095e23793ba6c2e894e7e63f6d89ad51f22e"},
-    {file = "mypy-1.19.1-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:b64d987153888790bcdb03a6473d321820597ab8dd9243b27a92153c4fa50fd2"},
-    {file = "mypy-1.19.1-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:c35d298c2c4bba75feb2195655dfea8124d855dfd7343bf8b8c055421eaf0cf8"},
-    {file = "mypy-1.19.1-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:34c81968774648ab5ac09c29a375fdede03ba253f8f8287847bd480782f73a6a"},
-    {file = "mypy-1.19.1-cp312-cp312-win_amd64.whl", hash = "sha256:b10e7c2cd7870ba4ad9b2d8a6102eb5ffc1f16ca35e3de6bfa390c1113029d13"},
-    {file = "mypy-1.19.1-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:e3157c7594ff2ef1634ee058aafc56a82db665c9438fd41b390f3bde1ab12250"},
-    {file = "mypy-1.19.1-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:bdb12f69bcc02700c2b47e070238f42cb87f18c0bc1fc4cdb4fb2bc5fd7a3b8b"},
-    {file = "mypy-1.19.1-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:f859fb09d9583a985be9a493d5cfc5515b56b08f7447759a0c5deaf68d80506e"},
-    {file = "mypy-1.19.1-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:c9a6538e0415310aad77cb94004ca6482330fece18036b5f360b62c45814c4ef"},
-    {file = "mypy-1.19.1-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:da4869fc5e7f62a88f3fe0b5c919d1d9f7ea3cef92d3689de2823fd27e40aa75"},
-    {file = "mypy-1.19.1-cp313-cp313-win_amd64.whl", hash = "sha256:016f2246209095e8eda7538944daa1d60e1e8134d98983b9fc1e92c1fc0cb8dd"},
-    {file = "mypy-1.19.1-cp314-cp314-macosx_10_15_x86_64.whl", hash = "sha256:06e6170bd5836770e8104c8fdd58e5e725cfeb309f0a6c681a811f557e97eac1"},
-    {file = "mypy-1.19.1-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:804bd67b8054a85447c8954215a906d6eff9cabeabe493fb6334b24f4bfff718"},
-    {file = "mypy-1.19.1-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:21761006a7f497cb0d4de3d8ef4ca70532256688b0523eee02baf9eec895e27b"},
-    {file = "mypy-1.19.1-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:28902ee51f12e0f19e1e16fbe2f8f06b6637f482c459dd393efddd0ec7f82045"},
-    {file = "mypy-1.19.1-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:481daf36a4c443332e2ae9c137dfee878fcea781a2e3f895d54bd3002a900957"},
-    {file = "mypy-1.19.1-cp314-cp314-win_amd64.whl", hash = "sha256:8bb5c6f6d043655e055be9b542aa5f3bdd30e4f3589163e85f93f3640060509f"},
-    {file = "mypy-1.19.1-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:7bcfc336a03a1aaa26dfce9fff3e287a3ba99872a157561cbfcebe67c13308e3"},
-    {file = "mypy-1.19.1-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:b7951a701c07ea584c4fe327834b92a30825514c868b1f69c30445093fdd9d5a"},
-    {file = "mypy-1.19.1-cp39-cp39-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:b13cfdd6c87fc3efb69ea4ec18ef79c74c3f98b4e5498ca9b85ab3b2c2329a67"},
-    {file = "mypy-1.19.1-cp39-cp39-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:4f28f99c824ecebcdaa2e55d82953e38ff60ee5ec938476796636b86afa3956e"},
-    {file = "mypy-1.19.1-cp39-cp39-musllinux_1_2_x86_64.whl", hash = "sha256:c608937067d2fc5a4dd1a5ce92fd9e1398691b8c5d012d66e1ddd430e9244376"},
-    {file = "mypy-1.19.1-cp39-cp39-win_amd64.whl", hash = "sha256:409088884802d511ee52ca067707b90c883426bd95514e8cfda8281dc2effe24"},
-    {file = "mypy-1.19.1-py3-none-any.whl", hash = "sha256:f1235f5ea01b7db5468d53ece6aaddf1ad0b88d9e7462b86ef96fe04995d7247"},
-    {file = "mypy-1.19.1.tar.gz", hash = "sha256:19d88bb05303fe63f71dd2c6270daca27cb9401c4ca8255fe50d1d920e0eb9ba"},
-]
-
-[package.dependencies]
-librt = {version = ">=0.6.2", markers = "platform_python_implementation != \"PyPy\""}
-mypy_extensions = ">=1.0.0"
-pathspec = ">=0.9.0"
-typing_extensions = ">=4.6.0"
-
-[package.extras]
-dmypy = ["psutil (>=4.0)"]
-faster-cache = ["orjson"]
-install-types = ["pip"]
-mypyc = ["setuptools (>=50)"]
-reports = ["lxml"]
-
-[[package]]
-name = "mypy-extensions"
-version = "1.1.0"
-description = "Type system extensions for programs checked with the mypy type checker."
-optional = false
-python-versions = ">=3.8"
-groups = ["dev"]
-files = [
-    {file = "mypy_extensions-1.1.0-py3-none-any.whl", hash = "sha256:1be4cccdb0f2482337c4743e60421de3a356cd97508abadd57d47403e94f5505"},
-    {file = "mypy_extensions-1.1.0.tar.gz", hash = "sha256:52e68efc3284861e772bbcd66823fde5ae21fd2fdb51c62a211403730b916558"},
-]
-
-[[package]]
-name = "nodeenv"
-version = "1.10.0"
-description = "Node.js virtual environment builder"
-optional = false
-python-versions = "!=3.0.*,!=3.1.*,!=3.2.*,!=3.3.*,!=3.4.*,!=3.5.*,!=3.6.*,>=2.7"
-groups = ["dev"]
-files = [
-    {file = "nodeenv-1.10.0-py2.py3-none-any.whl", hash = "sha256:5bb13e3eed2923615535339b3c620e76779af4cb4c6a90deccc9e36b274d3827"},
-    {file = "nodeenv-1.10.0.tar.gz", hash = "sha256:996c191ad80897d076bdfba80a41994c2b47c68e224c542b48feba42ba00f8bb"},
-]
-
-[[package]]
-name = "packaging"
-version = "26.0"
-description = "Core utilities for Python packages"
-optional = false
-python-versions = ">=3.8"
-groups = ["dev"]
-files = [
-    {file = "packaging-26.0-py3-none-any.whl", hash = "sha256:b36f1fef9334a5588b4166f8bcd26a14e521f2b55e6b9de3aaa80d3ff7a37529"},
-    {file = "packaging-26.0.tar.gz", hash = "sha256:00243ae351a257117b6a241061796684b084ed1c516a08c48a3f7e147a9d80b4"},
-]
-
-[[package]]
-name = "paddler-test-cluster"
-version = "4.0.1"
-description = "Spawns real Paddler balancers and agents for the Python test suites"
-optional = false
-python-versions = "^3.11"
-groups = ["dev"]
-files = []
-develop = true
-
-[package.dependencies]
-paddler-client = {path = "../paddler_client_python", develop = true}
-
-[package.source]
-type = "directory"
-url = "../paddler_test_cluster_python"
-
-[[package]]
-name = "pathspec"
-version = "1.0.4"
-description = "Utility library for gitignore style pattern matching of file paths."
-optional = false
-python-versions = ">=3.9"
-groups = ["dev"]
-files = [
-    {file = "pathspec-1.0.4-py3-none-any.whl", hash = "sha256:fb6ae2fd4e7c921a165808a552060e722767cfa526f99ca5156ed2ce45a5c723"},
-    {file = "pathspec-1.0.4.tar.gz", hash = "sha256:0210e2ae8a21a9137c0d470578cb0e595af87edaa6ebf12ff176f14a02e0e645"},
-]
-
-[package.extras]
-hyperscan = ["hyperscan (>=0.7)"]
-optional = ["typing-extensions (>=4)"]
-re2 = ["google-re2 (>=1.1)"]
-tests = ["pytest (>=9)", "typing-extensions (>=4.15)"]
-
-[[package]]
-name = "pluggy"
-version = "1.6.0"
-description = "plugin and hook calling mechanisms for python"
-optional = false
-python-versions = ">=3.9"
-groups = ["dev"]
-files = [
-    {file = "pluggy-1.6.0-py3-none-any.whl", hash = "sha256:e920276dd6813095e9377c0bc5566d94c932c33b27a3e3945d8389c374dd4746"},
-    {file = "pluggy-1.6.0.tar.gz", hash = "sha256:7dcc130b76258d33b90f61b658791dede3486c3e6bfb003ee5c9bfb396dd22f3"},
-]
-
-[package.extras]
-dev = ["pre-commit", "tox"]
-testing = ["coverage", "pytest", "pytest-benchmark"]
-
-[[package]]
-name = "pydantic"
-version = "2.12.5"
-description = "Data validation using Python type hints"
-optional = false
-python-versions = ">=3.9"
-groups = ["main"]
-files = [
-    {file = "pydantic-2.12.5-py3-none-any.whl", hash = "sha256:e561593fccf61e8a20fc46dfc2dfe075b8be7d0188df33f221ad1f0139180f9d"},
-    {file = "pydantic-2.12.5.tar.gz", hash = "sha256:4d351024c75c0f085a9febbb665ce8c0c6ec5d30e903bdb6394b7ede26aebb49"},
-]
-
-[package.dependencies]
-annotated-types = ">=0.6.0"
-pydantic-core = "2.41.5"
-typing-extensions = ">=4.14.1"
-typing-inspection = ">=0.4.2"
-
-[package.extras]
-email = ["email-validator (>=2.0.0)"]
-timezone = ["tzdata ; python_version >= \"3.9\" and platform_system == \"Windows\""]
-
-[[package]]
-name = "pydantic-core"
-version = "2.41.5"
-description = "Core functionality for Pydantic validation and serialization"
-optional = false
-python-versions = ">=3.9"
-groups = ["main"]
-files = [
-    {file = "pydantic_core-2.41.5-cp310-cp310-macosx_10_12_x86_64.whl", hash = "sha256:77b63866ca88d804225eaa4af3e664c5faf3568cea95360d21f4725ab6e07146"},
-    {file = "pydantic_core-2.41.5-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:dfa8a0c812ac681395907e71e1274819dec685fec28273a28905df579ef137e2"},
-    {file = "pydantic_core-2.41.5-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:5921a4d3ca3aee735d9fd163808f5e8dd6c6972101e4adbda9a4667908849b97"},
-    {file = "pydantic_core-2.41.5-cp310-cp310-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:e25c479382d26a2a41b7ebea1043564a937db462816ea07afa8a44c0866d52f9"},
-    {file = "pydantic_core-2.41.5-cp310-cp310-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:f547144f2966e1e16ae626d8ce72b4cfa0caedc7fa28052001c94fb2fcaa1c52"},
-    {file = "pydantic_core-2.41.5-cp310-cp310-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:6f52298fbd394f9ed112d56f3d11aabd0d5bd27beb3084cc3d8ad069483b8941"},
-    {file = "pydantic_core-2.41.5-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:100baa204bb412b74fe285fb0f3a385256dad1d1879f0a5cb1499ed2e83d132a"},
-    {file = "pydantic_core-2.41.5-cp310-cp310-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:05a2c8852530ad2812cb7914dc61a1125dc4e06252ee98e5638a12da6cc6fb6c"},
-    {file = "pydantic_core-2.41.5-cp310-cp310-musllinux_1_1_aarch64.whl", hash = "sha256:29452c56df2ed968d18d7e21f4ab0ac55e71dc59524872f6fc57dcf4a3249ed2"},
-    {file = "pydantic_core-2.41.5-cp310-cp310-musllinux_1_1_armv7l.whl", hash = "sha256:d5160812ea7a8a2ffbe233d8da666880cad0cbaf5d4de74ae15c313213d62556"},
-    {file = "pydantic_core-2.41.5-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:df3959765b553b9440adfd3c795617c352154e497a4eaf3752555cfb5da8fc49"},
-    {file = "pydantic_core-2.41.5-cp310-cp310-win32.whl", hash = "sha256:1f8d33a7f4d5a7889e60dc39856d76d09333d8a6ed0f5f1190635cbec70ec4ba"},
-    {file = "pydantic_core-2.41.5-cp310-cp310-win_amd64.whl", hash = "sha256:62de39db01b8d593e45871af2af9e497295db8d73b085f6bfd0b18c83c70a8f9"},
-    {file = "pydantic_core-2.41.5-cp311-cp311-macosx_10_12_x86_64.whl", hash = "sha256:a3a52f6156e73e7ccb0f8cced536adccb7042be67cb45f9562e12b319c119da6"},
-    {file = "pydantic_core-2.41.5-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:7f3bf998340c6d4b0c9a2f02d6a400e51f123b59565d74dc60d252ce888c260b"},
-    {file = "pydantic_core-2.41.5-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:378bec5c66998815d224c9ca994f1e14c0c21cb95d2f52b6021cc0b2a58f2a5a"},
-    {file = "pydantic_core-2.41.5-cp311-cp311-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:e7b576130c69225432866fe2f4a469a85a54ade141d96fd396dffcf607b558f8"},
-    {file = "pydantic_core-2.41.5-cp311-cp311-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:6cb58b9c66f7e4179a2d5e0f849c48eff5c1fca560994d6eb6543abf955a149e"},
-    {file = "pydantic_core-2.41.5-cp311-cp311-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:88942d3a3dff3afc8288c21e565e476fc278902ae4d6d134f1eeda118cc830b1"},
-    {file = "pydantic_core-2.41.5-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:f31d95a179f8d64d90f6831d71fa93290893a33148d890ba15de25642c5d075b"},
-    {file = "pydantic_core-2.41.5-cp311-cp311-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:c1df3d34aced70add6f867a8cf413e299177e0c22660cc767218373d0779487b"},
-    {file = "pydantic_core-2.41.5-cp311-cp311-musllinux_1_1_aarch64.whl", hash = "sha256:4009935984bd36bd2c774e13f9a09563ce8de4abaa7226f5108262fa3e637284"},
-    {file = "pydantic_core-2.41.5-cp311-cp311-musllinux_1_1_armv7l.whl", hash = "sha256:34a64bc3441dc1213096a20fe27e8e128bd3ff89921706e83c0b1ac971276594"},
-    {file = "pydantic_core-2.41.5-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:c9e19dd6e28fdcaa5a1de679aec4141f691023916427ef9bae8584f9c2fb3b0e"},
-    {file = "pydantic_core-2.41.5-cp311-cp311-win32.whl", hash = "sha256:2c010c6ded393148374c0f6f0bf89d206bf3217f201faa0635dcd56bd1520f6b"},
-    {file = "pydantic_core-2.41.5-cp311-cp311-win_amd64.whl", hash = "sha256:76ee27c6e9c7f16f47db7a94157112a2f3a00e958bc626e2f4ee8bec5c328fbe"},
-    {file = "pydantic_core-2.41.5-cp311-cp311-win_arm64.whl", hash = "sha256:4bc36bbc0b7584de96561184ad7f012478987882ebf9f9c389b23f432ea3d90f"},
-    {file = "pydantic_core-2.41.5-cp312-cp312-macosx_10_12_x86_64.whl", hash = "sha256:f41a7489d32336dbf2199c8c0a215390a751c5b014c2c1c5366e817202e9cdf7"},
-    {file = "pydantic_core-2.41.5-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:070259a8818988b9a84a449a2a7337c7f430a22acc0859c6b110aa7212a6d9c0"},
-    {file = "pydantic_core-2.41.5-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:e96cea19e34778f8d59fe40775a7a574d95816eb150850a85a7a4c8f4b94ac69"},
-    {file = "pydantic_core-2.41.5-cp312-cp312-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:ed2e99c456e3fadd05c991f8f437ef902e00eedf34320ba2b0842bd1c3ca3a75"},
-    {file = "pydantic_core-2.41.5-cp312-cp312-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:65840751b72fbfd82c3c640cff9284545342a4f1eb1586ad0636955b261b0b05"},
-    {file = "pydantic_core-2.41.5-cp312-cp312-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:e536c98a7626a98feb2d3eaf75944ef6f3dbee447e1f841eae16f2f0a72d8ddc"},
-    {file = "pydantic_core-2.41.5-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:eceb81a8d74f9267ef4081e246ffd6d129da5d87e37a77c9bde550cb04870c1c"},
-    {file = "pydantic_core-2.41.5-cp312-cp312-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:d38548150c39b74aeeb0ce8ee1d8e82696f4a4e16ddc6de7b1d8823f7de4b9b5"},
-    {file = "pydantic_core-2.41.5-cp312-cp312-musllinux_1_1_aarch64.whl", hash = "sha256:c23e27686783f60290e36827f9c626e63154b82b116d7fe9adba1fda36da706c"},
-    {file = "pydantic_core-2.41.5-cp312-cp312-musllinux_1_1_armv7l.whl", hash = "sha256:482c982f814460eabe1d3bb0adfdc583387bd4691ef00b90575ca0d2b6fe2294"},
-    {file = "pydantic_core-2.41.5-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:bfea2a5f0b4d8d43adf9d7b8bf019fb46fdd10a2e5cde477fbcb9d1fa08c68e1"},
-    {file = "pydantic_core-2.41.5-cp312-cp312-win32.whl", hash = "sha256:b74557b16e390ec12dca509bce9264c3bbd128f8a2c376eaa68003d7f327276d"},
-    {file = "pydantic_core-2.41.5-cp312-cp312-win_amd64.whl", hash = "sha256:1962293292865bca8e54702b08a4f26da73adc83dd1fcf26fbc875b35d81c815"},
-    {file = "pydantic_core-2.41.5-cp312-cp312-win_arm64.whl", hash = "sha256:1746d4a3d9a794cacae06a5eaaccb4b8643a131d45fbc9af23e353dc0a5ba5c3"},
-    {file = "pydantic_core-2.41.5-cp313-cp313-macosx_10_12_x86_64.whl", hash = "sha256:941103c9be18ac8daf7b7adca8228f8ed6bb7a1849020f643b3a14d15b1924d9"},
-    {file = "pydantic_core-2.41.5-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:112e305c3314f40c93998e567879e887a3160bb8689ef3d2c04b6cc62c33ac34"},
-    {file = "pydantic_core-2.41.5-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:0cbaad15cb0c90aa221d43c00e77bb33c93e8d36e0bf74760cd00e732d10a6a0"},
-    {file = "pydantic_core-2.41.5-cp313-cp313-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:03ca43e12fab6023fc79d28ca6b39b05f794ad08ec2feccc59a339b02f2b3d33"},
-    {file = "pydantic_core-2.41.5-cp313-cp313-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:dc799088c08fa04e43144b164feb0c13f9a0bc40503f8df3e9fde58a3c0c101e"},
-    {file = "pydantic_core-2.41.5-cp313-cp313-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:97aeba56665b4c3235a0e52b2c2f5ae9cd071b8a8310ad27bddb3f7fb30e9aa2"},
-    {file = "pydantic_core-2.41.5-cp313-cp313-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:406bf18d345822d6c21366031003612b9c77b3e29ffdb0f612367352aab7d586"},
-    {file = "pydantic_core-2.41.5-cp313-cp313-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:b93590ae81f7010dbe380cdeab6f515902ebcbefe0b9327cc4804d74e93ae69d"},
-    {file = "pydantic_core-2.41.5-cp313-cp313-musllinux_1_1_aarch64.whl", hash = "sha256:01a3d0ab748ee531f4ea6c3e48ad9dac84ddba4b0d82291f87248f2f9de8d740"},
-    {file = "pydantic_core-2.41.5-cp313-cp313-musllinux_1_1_armv7l.whl", hash = "sha256:6561e94ba9dacc9c61bce40e2d6bdc3bfaa0259d3ff36ace3b1e6901936d2e3e"},
-    {file = "pydantic_core-2.41.5-cp313-cp313-musllinux_1_1_x86_64.whl", hash = "sha256:915c3d10f81bec3a74fbd4faebe8391013ba61e5a1a8d48c4455b923bdda7858"},
-    {file = "pydantic_core-2.41.5-cp313-cp313-win32.whl", hash = "sha256:650ae77860b45cfa6e2cdafc42618ceafab3a2d9a3811fcfbd3bbf8ac3c40d36"},
-    {file = "pydantic_core-2.41.5-cp313-cp313-win_amd64.whl", hash = "sha256:79ec52ec461e99e13791ec6508c722742ad745571f234ea6255bed38c6480f11"},
-    {file = "pydantic_core-2.41.5-cp313-cp313-win_arm64.whl", hash = "sha256:3f84d5c1b4ab906093bdc1ff10484838aca54ef08de4afa9de0f5f14d69639cd"},
-    {file = "pydantic_core-2.41.5-cp314-cp314-macosx_10_12_x86_64.whl", hash = "sha256:3f37a19d7ebcdd20b96485056ba9e8b304e27d9904d233d7b1015db320e51f0a"},
-    {file = "pydantic_core-2.41.5-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:1d1d9764366c73f996edd17abb6d9d7649a7eb690006ab6adbda117717099b14"},
-    {file = "pydantic_core-2.41.5-cp314-cp314-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:25e1c2af0fce638d5f1988b686f3b3ea8cd7de5f244ca147c777769e798a9cd1"},
-    {file = "pydantic_core-2.41.5-cp314-cp314-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:506d766a8727beef16b7adaeb8ee6217c64fc813646b424d0804d67c16eddb66"},
-    {file = "pydantic_core-2.41.5-cp314-cp314-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:4819fa52133c9aa3c387b3328f25c1facc356491e6135b459f1de698ff64d869"},
-    {file = "pydantic_core-2.41.5-cp314-cp314-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:2b761d210c9ea91feda40d25b4efe82a1707da2ef62901466a42492c028553a2"},
-    {file = "pydantic_core-2.41.5-cp314-cp314-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:22f0fb8c1c583a3b6f24df2470833b40207e907b90c928cc8d3594b76f874375"},
-    {file = "pydantic_core-2.41.5-cp314-cp314-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:2782c870e99878c634505236d81e5443092fba820f0373997ff75f90f68cd553"},
-    {file = "pydantic_core-2.41.5-cp314-cp314-musllinux_1_1_aarch64.whl", hash = "sha256:0177272f88ab8312479336e1d777f6b124537d47f2123f89cb37e0accea97f90"},
-    {file = "pydantic_core-2.41.5-cp314-cp314-musllinux_1_1_armv7l.whl", hash = "sha256:63510af5e38f8955b8ee5687740d6ebf7c2a0886d15a6d65c32814613681bc07"},
-    {file = "pydantic_core-2.41.5-cp314-cp314-musllinux_1_1_x86_64.whl", hash = "sha256:e56ba91f47764cc14f1daacd723e3e82d1a89d783f0f5afe9c364b8bb491ccdb"},
-    {file = "pydantic_core-2.41.5-cp314-cp314-win32.whl", hash = "sha256:aec5cf2fd867b4ff45b9959f8b20ea3993fc93e63c7363fe6851424c8a7e7c23"},
-    {file = "pydantic_core-2.41.5-cp314-cp314-win_amd64.whl", hash = "sha256:8e7c86f27c585ef37c35e56a96363ab8de4e549a95512445b85c96d3e2f7c1bf"},
-    {file = "pydantic_core-2.41.5-cp314-cp314-win_arm64.whl", hash = "sha256:e672ba74fbc2dc8eea59fb6d4aed6845e6905fc2a8afe93175d94a83ba2a01a0"},
-    {file = "pydantic_core-2.41.5-cp314-cp314t-macosx_10_12_x86_64.whl", hash = "sha256:8566def80554c3faa0e65ac30ab0932b9e3a5cd7f8323764303d468e5c37595a"},
-    {file = "pydantic_core-2.41.5-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:b80aa5095cd3109962a298ce14110ae16b8c1aece8b72f9dafe81cf597ad80b3"},
-    {file = "pydantic_core-2.41.5-cp314-cp314t-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:3006c3dd9ba34b0c094c544c6006cc79e87d8612999f1a5d43b769b89181f23c"},
-    {file = "pydantic_core-2.41.5-cp314-cp314t-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:72f6c8b11857a856bcfa48c86f5368439f74453563f951e473514579d44aa612"},
-    {file = "pydantic_core-2.41.5-cp314-cp314t-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:5cb1b2f9742240e4bb26b652a5aeb840aa4b417c7748b6f8387927bc6e45e40d"},
-    {file = "pydantic_core-2.41.5-cp314-cp314t-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:bd3d54f38609ff308209bd43acea66061494157703364ae40c951f83ba99a1a9"},
-    {file = "pydantic_core-2.41.5-cp314-cp314t-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:2ff4321e56e879ee8d2a879501c8e469414d948f4aba74a2d4593184eb326660"},
-    {file = "pydantic_core-2.41.5-cp314-cp314t-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:d0d2568a8c11bf8225044aa94409e21da0cb09dcdafe9ecd10250b2baad531a9"},
-    {file = "pydantic_core-2.41.5-cp314-cp314t-musllinux_1_1_aarch64.whl", hash = "sha256:a39455728aabd58ceabb03c90e12f71fd30fa69615760a075b9fec596456ccc3"},
-    {file = "pydantic_core-2.41.5-cp314-cp314t-musllinux_1_1_armv7l.whl", hash = "sha256:239edca560d05757817c13dc17c50766136d21f7cd0fac50295499ae24f90fdf"},
-    {file = "pydantic_core-2.41.5-cp314-cp314t-musllinux_1_1_x86_64.whl", hash = "sha256:2a5e06546e19f24c6a96a129142a75cee553cc018ffee48a460059b1185f4470"},
-    {file = "pydantic_core-2.41.5-cp314-cp314t-win32.whl", hash = "sha256:b4ececa40ac28afa90871c2cc2b9ffd2ff0bf749380fbdf57d165fd23da353aa"},
-    {file = "pydantic_core-2.41.5-cp314-cp314t-win_amd64.whl", hash = "sha256:80aa89cad80b32a912a65332f64a4450ed00966111b6615ca6816153d3585a8c"},
-    {file = "pydantic_core-2.41.5-cp314-cp314t-win_arm64.whl", hash = "sha256:35b44f37a3199f771c3eaa53051bc8a70cd7b54f333531c59e29fd4db5d15008"},
-    {file = "pydantic_core-2.41.5-cp39-cp39-macosx_10_12_x86_64.whl", hash = "sha256:8bfeaf8735be79f225f3fefab7f941c712aaca36f1128c9d7e2352ee1aa87bdf"},
-    {file = "pydantic_core-2.41.5-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:346285d28e4c8017da95144c7f3acd42740d637ff41946af5ce6e5e420502dd5"},
-    {file = "pydantic_core-2.41.5-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:a75dafbf87d6276ddc5b2bf6fae5254e3d0876b626eb24969a574fff9149ee5d"},
-    {file = "pydantic_core-2.41.5-cp39-cp39-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:7b93a4d08587e2b7e7882de461e82b6ed76d9026ce91ca7915e740ecc7855f60"},
-    {file = "pydantic_core-2.41.5-cp39-cp39-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:e8465ab91a4bd96d36dde3263f06caa6a8a6019e4113f24dc753d79a8b3a3f82"},
-    {file = "pydantic_core-2.41.5-cp39-cp39-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:299e0a22e7ae2b85c1a57f104538b2656e8ab1873511fd718a1c1c6f149b77b5"},
-    {file = "pydantic_core-2.41.5-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:707625ef0983fcfb461acfaf14de2067c5942c6bb0f3b4c99158bed6fedd3cf3"},
-    {file = "pydantic_core-2.41.5-cp39-cp39-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:f41eb9797986d6ebac5e8edff36d5cef9de40def462311b3eb3eeded1431e425"},
-    {file = "pydantic_core-2.41.5-cp39-cp39-musllinux_1_1_aarch64.whl", hash = "sha256:0384e2e1021894b1ff5a786dbf94771e2986ebe2869533874d7e43bc79c6f504"},
-    {file = "pydantic_core-2.41.5-cp39-cp39-musllinux_1_1_armv7l.whl", hash = "sha256:f0cd744688278965817fd0839c4a4116add48d23890d468bc436f78beb28abf5"},
-    {file = "pydantic_core-2.41.5-cp39-cp39-musllinux_1_1_x86_64.whl", hash = "sha256:753e230374206729bf0a807954bcc6c150d3743928a73faffee51ac6557a03c3"},
-    {file = "pydantic_core-2.41.5-cp39-cp39-win32.whl", hash = "sha256:873e0d5b4fb9b89ef7c2d2a963ea7d02879d9da0da8d9d4933dee8ee86a8b460"},
-    {file = "pydantic_core-2.41.5-cp39-cp39-win_amd64.whl", hash = "sha256:e4f4a984405e91527a0d62649ee21138f8e3d0ef103be488c1dc11a80d7f184b"},
-    {file = "pydantic_core-2.41.5-graalpy311-graalpy242_311_native-macosx_10_12_x86_64.whl", hash = "sha256:b96d5f26b05d03cc60f11a7761a5ded1741da411e7fe0909e27a5e6a0cb7b034"},
-    {file = "pydantic_core-2.41.5-graalpy311-graalpy242_311_native-macosx_11_0_arm64.whl", hash = "sha256:634e8609e89ceecea15e2d61bc9ac3718caaaa71963717bf3c8f38bfde64242c"},
-    {file = "pydantic_core-2.41.5-graalpy311-graalpy242_311_native-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:93e8740d7503eb008aa2df04d3b9735f845d43ae845e6dcd2be0b55a2da43cd2"},
-    {file = "pydantic_core-2.41.5-graalpy311-graalpy242_311_native-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:f15489ba13d61f670dcc96772e733aad1a6f9c429cc27574c6cdaed82d0146ad"},
-    {file = "pydantic_core-2.41.5-graalpy312-graalpy250_312_native-macosx_10_12_x86_64.whl", hash = "sha256:7da7087d756b19037bc2c06edc6c170eeef3c3bafcb8f532ff17d64dc427adfd"},
-    {file = "pydantic_core-2.41.5-graalpy312-graalpy250_312_native-macosx_11_0_arm64.whl", hash = "sha256:aabf5777b5c8ca26f7824cb4a120a740c9588ed58df9b2d196ce92fba42ff8dc"},
-    {file = "pydantic_core-2.41.5-graalpy312-graalpy250_312_native-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:c007fe8a43d43b3969e8469004e9845944f1a80e6acd47c150856bb87f230c56"},
-    {file = "pydantic_core-2.41.5-graalpy312-graalpy250_312_native-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:76d0819de158cd855d1cbb8fcafdf6f5cf1eb8e470abe056d5d161106e38062b"},
-    {file = "pydantic_core-2.41.5-pp310-pypy310_pp73-macosx_10_12_x86_64.whl", hash = "sha256:b5819cd790dbf0c5eb9f82c73c16b39a65dd6dd4d1439dcdea7816ec9adddab8"},
-    {file = "pydantic_core-2.41.5-pp310-pypy310_pp73-macosx_11_0_arm64.whl", hash = "sha256:5a4e67afbc95fa5c34cf27d9089bca7fcab4e51e57278d710320a70b956d1b9a"},
-    {file = "pydantic_core-2.41.5-pp310-pypy310_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ece5c59f0ce7d001e017643d8d24da587ea1f74f6993467d85ae8a5ef9d4f42b"},
-    {file = "pydantic_core-2.41.5-pp310-pypy310_pp73-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:16f80f7abe3351f8ea6858914ddc8c77e02578544a0ebc15b4c2e1a0e813b0b2"},
-    {file = "pydantic_core-2.41.5-pp310-pypy310_pp73-musllinux_1_1_aarch64.whl", hash = "sha256:33cb885e759a705b426baada1fe68cbb0a2e68e34c5d0d0289a364cf01709093"},
-    {file = "pydantic_core-2.41.5-pp310-pypy310_pp73-musllinux_1_1_armv7l.whl", hash = "sha256:c8d8b4eb992936023be7dee581270af5c6e0697a8559895f527f5b7105ecd36a"},
-    {file = "pydantic_core-2.41.5-pp310-pypy310_pp73-musllinux_1_1_x86_64.whl", hash = "sha256:242a206cd0318f95cd21bdacff3fcc3aab23e79bba5cac3db5a841c9ef9c6963"},
-    {file = "pydantic_core-2.41.5-pp310-pypy310_pp73-win_amd64.whl", hash = "sha256:d3a978c4f57a597908b7e697229d996d77a6d3c94901e9edee593adada95ce1a"},
-    {file = "pydantic_core-2.41.5-pp311-pypy311_pp73-macosx_10_12_x86_64.whl", hash = "sha256:b2379fa7ed44ddecb5bfe4e48577d752db9fc10be00a6b7446e9663ba143de26"},
-    {file = "pydantic_core-2.41.5-pp311-pypy311_pp73-macosx_11_0_arm64.whl", hash = "sha256:266fb4cbf5e3cbd0b53669a6d1b039c45e3ce651fd5442eff4d07c2cc8d66808"},
-    {file = "pydantic_core-2.41.5-pp311-pypy311_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:58133647260ea01e4d0500089a8c4f07bd7aa6ce109682b1426394988d8aaacc"},
-    {file = "pydantic_core-2.41.5-pp311-pypy311_pp73-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:287dad91cfb551c363dc62899a80e9e14da1f0e2b6ebde82c806612ca2a13ef1"},
-    {file = "pydantic_core-2.41.5-pp311-pypy311_pp73-musllinux_1_1_aarch64.whl", hash = "sha256:03b77d184b9eb40240ae9fd676ca364ce1085f203e1b1256f8ab9984dca80a84"},
-    {file = "pydantic_core-2.41.5-pp311-pypy311_pp73-musllinux_1_1_armv7l.whl", hash = "sha256:a668ce24de96165bb239160b3d854943128f4334822900534f2fe947930e5770"},
-    {file = "pydantic_core-2.41.5-pp311-pypy311_pp73-musllinux_1_1_x86_64.whl", hash = "sha256:f14f8f046c14563f8eb3f45f499cc658ab8d10072961e07225e507adb700e93f"},
-    {file = "pydantic_core-2.41.5-pp311-pypy311_pp73-win_amd64.whl", hash = "sha256:56121965f7a4dc965bff783d70b907ddf3d57f6eba29b6d2e5dabfaf07799c51"},
-    {file = "pydantic_core-2.41.5.tar.gz", hash = "sha256:08daa51ea16ad373ffd5e7606252cc32f07bc72b28284b6bc9c6df804816476e"},
-]
-
-[package.dependencies]
-typing-extensions = ">=4.14.1"
-
-[[package]]
-name = "pygments"
-version = "2.20.0"
-description = "Pygments is a syntax highlighting package written in Python."
-optional = false
-python-versions = ">=3.9"
-groups = ["dev"]
-files = [
-    {file = "pygments-2.20.0-py3-none-any.whl", hash = "sha256:81a9e26dd42fd28a23a2d169d86d7ac03b46e2f8b59ed4698fb4785f946d0176"},
-    {file = "pygments-2.20.0.tar.gz", hash = "sha256:6757cd03768053ff99f3039c1a36d6c0aa0b263438fcab17520b30a303a82b5f"},
-]
-
-[package.extras]
-windows-terminal = ["colorama (>=0.4.6)"]
-
-[[package]]
-name = "pyright"
-version = "1.1.408"
-description = "Command line wrapper for pyright"
-optional = false
-python-versions = ">=3.7"
-groups = ["dev"]
-files = [
-    {file = "pyright-1.1.408-py3-none-any.whl", hash = "sha256:090b32865f4fdb1e0e6cd82bf5618480d48eecd2eb2e70f960982a3d9a4c17c1"},
-    {file = "pyright-1.1.408.tar.gz", hash = "sha256:f28f2321f96852fa50b5829ea492f6adb0e6954568d1caa3f3af3a5f555eb684"},
-]
-
-[package.dependencies]
-nodeenv = ">=1.6.0"
-typing-extensions = ">=4.1"
-
-[package.extras]
-all = ["nodejs-wheel-binaries", "twine (>=3.4.1)"]
-dev = ["twine (>=3.4.1)"]
-nodejs = ["nodejs-wheel-binaries"]
-
-[[package]]
-name = "pytest"
-version = "9.0.3"
-description = "pytest: simple powerful testing with Python"
-optional = false
-python-versions = ">=3.10"
-groups = ["dev"]
-files = [
-    {file = "pytest-9.0.3-py3-none-any.whl", hash = "sha256:2c5efc453d45394fdd706ade797c0a81091eccd1d6e4bccfcd476e2b8e0ab5d9"},
-    {file = "pytest-9.0.3.tar.gz", hash = "sha256:b86ada508af81d19edeb213c681b1d48246c1a91d304c6c81a427674c17eb91c"},
-]
-
-[package.dependencies]
-colorama = {version = ">=0.4", markers = "sys_platform == \"win32\""}
-iniconfig = ">=1.0.1"
-packaging = ">=22"
-pluggy = ">=1.5,<2"
-pygments = ">=2.7.2"
-
-[package.extras]
-dev = ["argcomplete", "attrs (>=19.2)", "hypothesis (>=3.56)", "mock", "requests", "setuptools", "xmlschema"]
-
-[[package]]
-name = "pytest-asyncio"
-version = "1.3.0"
-description = "Pytest support for asyncio"
-optional = false
-python-versions = ">=3.10"
-groups = ["dev"]
-files = [
-    {file = "pytest_asyncio-1.3.0-py3-none-any.whl", hash = "sha256:611e26147c7f77640e6d0a92a38ed17c3e9848063698d5c93d5aa7aa11cebff5"},
-    {file = "pytest_asyncio-1.3.0.tar.gz", hash = "sha256:d7f52f36d231b80ee124cd216ffb19369aa168fc10095013c6b014a34d3ee9e5"},
-]
-
-[package.dependencies]
-pytest = ">=8.2,<10"
-typing-extensions = {version = ">=4.12", markers = "python_version < \"3.13\""}
-
-[package.extras]
-docs = ["sphinx (>=5.3)", "sphinx-rtd-theme (>=1)"]
-testing = ["coverage (>=6.2)", "hypothesis (>=5.7.1)"]
-
-[[package]]
-name = "pytest-cov"
-version = "7.1.0"
-description = "Pytest plugin for measuring coverage."
-optional = false
-python-versions = ">=3.9"
-groups = ["dev"]
-files = [
-    {file = "pytest_cov-7.1.0-py3-none-any.whl", hash = "sha256:a0461110b7865f9a271aa1b51e516c9a95de9d696734a2f71e3e78f46e1d4678"},
-    {file = "pytest_cov-7.1.0.tar.gz", hash = "sha256:30674f2b5f6351aa09702a9c8c364f6a01c27aae0c1366ae8016160d1efc56b2"},
-]
-
-[package.dependencies]
-coverage = {version = ">=7.10.6", extras = ["toml"]}
-pluggy = ">=1.2"
-pytest = ">=7"
-
-[package.extras]
-testing = ["process-tests", "pytest-xdist", "virtualenv"]
-
-[[package]]
-name = "ruff"
-version = "0.15.12"
-description = "An extremely fast Python linter and code formatter, written in Rust."
-optional = false
-python-versions = ">=3.7"
-groups = ["dev"]
-files = [
-    {file = "ruff-0.15.12-py3-none-linux_armv6l.whl", hash = "sha256:f86f176e188e94d6bdbc09f09bfd9dc729059ad93d0e7390b5a73efe19f8861c"},
-    {file = "ruff-0.15.12-py3-none-macosx_10_12_x86_64.whl", hash = "sha256:e3bcd123364c3770b8e1b7baaf343cc99a35f197c5c6e8af79015c666c423a6c"},
-    {file = "ruff-0.15.12-py3-none-macosx_11_0_arm64.whl", hash = "sha256:fe87510d000220aa1ed530d4448a7c696a0cae1213e5ec30e5874287b66557b5"},
-    {file = "ruff-0.15.12-py3-none-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:84a1630093121375a3e2a95b4a6dc7b59e2b4ee76216e32d81aae550a832d002"},
-    {file = "ruff-0.15.12-py3-none-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:fb129f40f114f089ebe0ca56c0d251cf2061b17651d464bb6478dc01e69f11f5"},
-    {file = "ruff-0.15.12-py3-none-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:b0c862b172d695db7598426b8af465e7e9ac00a3ea2a3630ee67eb82e366aaa6"},
-    {file = "ruff-0.15.12-py3-none-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:2849ea9f3484c3aca43a82f484210370319e7170df4dfe4843395ddf6c57bc33"},
-    {file = "ruff-0.15.12-py3-none-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:9e77c7e51c07fe396826d5969a5b846d9cd4c402535835fb6e21ce8b28fef847"},
-    {file = "ruff-0.15.12-py3-none-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:83b2f4f2f3b1026b5fb449b467d9264bf22067b600f7b6f41fc5958909f449d0"},
-    {file = "ruff-0.15.12-py3-none-manylinux_2_31_riscv64.whl", hash = "sha256:9ba3b8f1afd7e2e43d8943e55f249e13f9682fde09711644a6e7290eb4f3e339"},
-    {file = "ruff-0.15.12-py3-none-musllinux_1_2_aarch64.whl", hash = "sha256:e852ba9fdc890655e1d78f2df1499efbe0e54126bd405362154a75e2bde159c5"},
-    {file = "ruff-0.15.12-py3-none-musllinux_1_2_armv7l.whl", hash = "sha256:dd8aed930da53780d22fc70bdf84452c843cf64f8cb4eb38984319c24c5cd5fd"},
-    {file = "ruff-0.15.12-py3-none-musllinux_1_2_i686.whl", hash = "sha256:01da3988d225628b709493d7dc67c3b9b12c0210016b08690ef9bd27970b262b"},
-    {file = "ruff-0.15.12-py3-none-musllinux_1_2_x86_64.whl", hash = "sha256:9cae0f92bd5700d1213188b31cd3bdd2b315361296d10b96b8e2337d3d11f53e"},
-    {file = "ruff-0.15.12-py3-none-win32.whl", hash = "sha256:d0185894e038d7043ba8fd6aee7499ece6462dc0ea9f1e260c7451807c714c20"},
-    {file = "ruff-0.15.12-py3-none-win_amd64.whl", hash = "sha256:c87a162d61ab3adca47c03f7f717c68672edec7d1b5499e652331780fe74950d"},
-    {file = "ruff-0.15.12-py3-none-win_arm64.whl", hash = "sha256:a538f7a82d061cee7be55542aca1d86d1393d55d81d4fcc314370f4340930d4f"},
-    {file = "ruff-0.15.12.tar.gz", hash = "sha256:ecea26adb26b4232c0c2ca19ccbc0083a68344180bba2a600605538ce51a40a6"},
-]
-
-[[package]]
-name = "typing-extensions"
-version = "4.16.0"
-description = "Backported and Experimental Type Hints for Python 3.9+"
-optional = false
-python-versions = ">=3.9"
-groups = ["main", "dev"]
-files = [
-    {file = "typing_extensions-4.16.0-py3-none-any.whl", hash = "sha256:481caa481374e813c1b176ada14e97f1f67a4539ce9cfeb3f350d78d6370c2e8"},
-    {file = "typing_extensions-4.16.0.tar.gz", hash = "sha256:dc983d19a509c94dba722ee6abd33940f7c05a89e243c47e907eb4db6f1a43e5"},
-]
-
-[[package]]
-name = "typing-inspection"
-version = "0.4.2"
-description = "Runtime typing introspection tools"
-optional = false
-python-versions = ">=3.9"
-groups = ["main"]
-files = [
-    {file = "typing_inspection-0.4.2-py3-none-any.whl", hash = "sha256:4ed1cacbdc298c220f1bd249ed5287caa16f34d44ef4e9c3d0cbad5b521545e7"},
-    {file = "typing_inspection-0.4.2.tar.gz", hash = "sha256:ba561c48a67c5958007083d386c3295464928b01faa735ab8547c5692e87f464"},
-]
-
-[package.dependencies]
-typing-extensions = ">=4.12.0"
-
-[[package]]
-name = "websockets"
-version = "15.0.1"
-description = "An implementation of the WebSocket Protocol (RFC 6455 & 7692)"
-optional = false
-python-versions = ">=3.9"
-groups = ["main"]
-files = [
-    {file = "websockets-15.0.1-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:d63efaa0cd96cf0c5fe4d581521d9fa87744540d4bc999ae6e08595a1014b45b"},
-    {file = "websockets-15.0.1-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:ac60e3b188ec7574cb761b08d50fcedf9d77f1530352db4eef1707fe9dee7205"},
-    {file = "websockets-15.0.1-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:5756779642579d902eed757b21b0164cd6fe338506a8083eb58af5c372e39d9a"},
-    {file = "websockets-15.0.1-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:0fdfe3e2a29e4db3659dbd5bbf04560cea53dd9610273917799f1cde46aa725e"},
-    {file = "websockets-15.0.1-cp310-cp310-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:4c2529b320eb9e35af0fa3016c187dffb84a3ecc572bcee7c3ce302bfeba52bf"},
-    {file = "websockets-15.0.1-cp310-cp310-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ac1e5c9054fe23226fb11e05a6e630837f074174c4c2f0fe442996112a6de4fb"},
-    {file = "websockets-15.0.1-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:5df592cd503496351d6dc14f7cdad49f268d8e618f80dce0cd5a36b93c3fc08d"},
-    {file = "websockets-15.0.1-cp310-cp310-musllinux_1_2_i686.whl", hash = "sha256:0a34631031a8f05657e8e90903e656959234f3a04552259458aac0b0f9ae6fd9"},
-    {file = "websockets-15.0.1-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:3d00075aa65772e7ce9e990cab3ff1de702aa09be3940d1dc88d5abf1ab8a09c"},
-    {file = "websockets-15.0.1-cp310-cp310-win32.whl", hash = "sha256:1234d4ef35db82f5446dca8e35a7da7964d02c127b095e172e54397fb6a6c256"},
-    {file = "websockets-15.0.1-cp310-cp310-win_amd64.whl", hash = "sha256:39c1fec2c11dc8d89bba6b2bf1556af381611a173ac2b511cf7231622058af41"},
-    {file = "websockets-15.0.1-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:823c248b690b2fd9303ba00c4f66cd5e2d8c3ba4aa968b2779be9532a4dad431"},
-    {file = "websockets-15.0.1-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:678999709e68425ae2593acf2e3ebcbcf2e69885a5ee78f9eb80e6e371f1bf57"},
-    {file = "websockets-15.0.1-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:d50fd1ee42388dcfb2b3676132c78116490976f1300da28eb629272d5d93e905"},
-    {file = "websockets-15.0.1-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:d99e5546bf73dbad5bf3547174cd6cb8ba7273062a23808ffea025ecb1cf8562"},
-    {file = "websockets-15.0.1-cp311-cp311-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:66dd88c918e3287efc22409d426c8f729688d89a0c587c88971a0faa2c2f3792"},
-    {file = "websockets-15.0.1-cp311-cp311-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:8dd8327c795b3e3f219760fa603dcae1dcc148172290a8ab15158cf85a953413"},
-    {file = "websockets-15.0.1-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:8fdc51055e6ff4adeb88d58a11042ec9a5eae317a0a53d12c062c8a8865909e8"},
-    {file = "websockets-15.0.1-cp311-cp311-musllinux_1_2_i686.whl", hash = "sha256:693f0192126df6c2327cce3baa7c06f2a117575e32ab2308f7f8216c29d9e2e3"},
-    {file = "websockets-15.0.1-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:54479983bd5fb469c38f2f5c7e3a24f9a4e70594cd68cd1fa6b9340dadaff7cf"},
-    {file = "websockets-15.0.1-cp311-cp311-win32.whl", hash = "sha256:16b6c1b3e57799b9d38427dda63edcbe4926352c47cf88588c0be4ace18dac85"},
-    {file = "websockets-15.0.1-cp311-cp311-win_amd64.whl", hash = "sha256:27ccee0071a0e75d22cb35849b1db43f2ecd3e161041ac1ee9d2352ddf72f065"},
-    {file = "websockets-15.0.1-cp312-cp312-macosx_10_13_universal2.whl", hash = "sha256:3e90baa811a5d73f3ca0bcbf32064d663ed81318ab225ee4f427ad4e26e5aff3"},
-    {file = "websockets-15.0.1-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:592f1a9fe869c778694f0aa806ba0374e97648ab57936f092fd9d87f8bc03665"},
-    {file = "websockets-15.0.1-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:0701bc3cfcb9164d04a14b149fd74be7347a530ad3bbf15ab2c678a2cd3dd9a2"},
-    {file = "websockets-15.0.1-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:e8b56bdcdb4505c8078cb6c7157d9811a85790f2f2b3632c7d1462ab5783d215"},
-    {file = "websockets-15.0.1-cp312-cp312-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:0af68c55afbd5f07986df82831c7bff04846928ea8d1fd7f30052638788bc9b5"},
-    {file = "websockets-15.0.1-cp312-cp312-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:64dee438fed052b52e4f98f76c5790513235efaa1ef7f3f2192c392cd7c91b65"},
-    {file = "websockets-15.0.1-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:d5f6b181bb38171a8ad1d6aa58a67a6aa9d4b38d0f8c5f496b9e42561dfc62fe"},
-    {file = "websockets-15.0.1-cp312-cp312-musllinux_1_2_i686.whl", hash = "sha256:5d54b09eba2bada6011aea5375542a157637b91029687eb4fdb2dab11059c1b4"},
-    {file = "websockets-15.0.1-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:3be571a8b5afed347da347bfcf27ba12b069d9d7f42cb8c7028b5e98bbb12597"},
-    {file = "websockets-15.0.1-cp312-cp312-win32.whl", hash = "sha256:c338ffa0520bdb12fbc527265235639fb76e7bc7faafbb93f6ba80d9c06578a9"},
-    {file = "websockets-15.0.1-cp312-cp312-win_amd64.whl", hash = "sha256:fcd5cf9e305d7b8338754470cf69cf81f420459dbae8a3b40cee57417f4614a7"},
-    {file = "websockets-15.0.1-cp313-cp313-macosx_10_13_universal2.whl", hash = "sha256:ee443ef070bb3b6ed74514f5efaa37a252af57c90eb33b956d35c8e9c10a1931"},
-    {file = "websockets-15.0.1-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:5a939de6b7b4e18ca683218320fc67ea886038265fd1ed30173f5ce3f8e85675"},
-    {file = "websockets-15.0.1-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:746ee8dba912cd6fc889a8147168991d50ed70447bf18bcda7039f7d2e3d9151"},
-    {file = "websockets-15.0.1-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:595b6c3969023ecf9041b2936ac3827e4623bfa3ccf007575f04c5a6aa318c22"},
-    {file = "websockets-15.0.1-cp313-cp313-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:3c714d2fc58b5ca3e285461a4cc0c9a66bd0e24c5da9911e30158286c9b5be7f"},
-    {file = "websockets-15.0.1-cp313-cp313-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:0f3c1e2ab208db911594ae5b4f79addeb3501604a165019dd221c0bdcabe4db8"},
-    {file = "websockets-15.0.1-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:229cf1d3ca6c1804400b0a9790dc66528e08a6a1feec0d5040e8b9eb14422375"},
-    {file = "websockets-15.0.1-cp313-cp313-musllinux_1_2_i686.whl", hash = "sha256:756c56e867a90fb00177d530dca4b097dd753cde348448a1012ed6c5131f8b7d"},
-    {file = "websockets-15.0.1-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:558d023b3df0bffe50a04e710bc87742de35060580a293c2a984299ed83bc4e4"},
-    {file = "websockets-15.0.1-cp313-cp313-win32.whl", hash = "sha256:ba9e56e8ceeeedb2e080147ba85ffcd5cd0711b89576b83784d8605a7df455fa"},
-    {file = "websockets-15.0.1-cp313-cp313-win_amd64.whl", hash = "sha256:e09473f095a819042ecb2ab9465aee615bd9c2028e4ef7d933600a8401c79561"},
-    {file = "websockets-15.0.1-cp39-cp39-macosx_10_9_universal2.whl", hash = "sha256:5f4c04ead5aed67c8a1a20491d54cdfba5884507a48dd798ecaf13c74c4489f5"},
-    {file = "websockets-15.0.1-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:abdc0c6c8c648b4805c5eacd131910d2a7f6455dfd3becab248ef108e89ab16a"},
-    {file = "websockets-15.0.1-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:a625e06551975f4b7ea7102bc43895b90742746797e2e14b70ed61c43a90f09b"},
-    {file = "websockets-15.0.1-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:d591f8de75824cbb7acad4e05d2d710484f15f29d4a915092675ad3456f11770"},
-    {file = "websockets-15.0.1-cp39-cp39-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:47819cea040f31d670cc8d324bb6435c6f133b8c7a19ec3d61634e62f8d8f9eb"},
-    {file = "websockets-15.0.1-cp39-cp39-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ac017dd64572e5c3bd01939121e4d16cf30e5d7e110a119399cf3133b63ad054"},
-    {file = "websockets-15.0.1-cp39-cp39-musllinux_1_2_aarch64.whl", hash = "sha256:4a9fac8e469d04ce6c25bb2610dc535235bd4aa14996b4e6dbebf5e007eba5ee"},
-    {file = "websockets-15.0.1-cp39-cp39-musllinux_1_2_i686.whl", hash = "sha256:363c6f671b761efcb30608d24925a382497c12c506b51661883c3e22337265ed"},
-    {file = "websockets-15.0.1-cp39-cp39-musllinux_1_2_x86_64.whl", hash = "sha256:2034693ad3097d5355bfdacfffcbd3ef5694f9718ab7f29c29689a9eae841880"},
-    {file = "websockets-15.0.1-cp39-cp39-win32.whl", hash = "sha256:3b1ac0d3e594bf121308112697cf4b32be538fb1444468fb0a6ae4feebc83411"},
-    {file = "websockets-15.0.1-cp39-cp39-win_amd64.whl", hash = "sha256:b7643a03db5c95c799b89b31c036d5f27eeb4d259c798e878d6937d71832b1e4"},
-    {file = "websockets-15.0.1-pp310-pypy310_pp73-macosx_10_15_x86_64.whl", hash = "sha256:0c9e74d766f2818bb95f84c25be4dea09841ac0f734d1966f415e4edfc4ef1c3"},
-    {file = "websockets-15.0.1-pp310-pypy310_pp73-macosx_11_0_arm64.whl", hash = "sha256:1009ee0c7739c08a0cd59de430d6de452a55e42d6b522de7aa15e6f67db0b8e1"},
-    {file = "websockets-15.0.1-pp310-pypy310_pp73-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:76d1f20b1c7a2fa82367e04982e708723ba0e7b8d43aa643d3dcd404d74f1475"},
-    {file = "websockets-15.0.1-pp310-pypy310_pp73-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:f29d80eb9a9263b8d109135351caf568cc3f80b9928bccde535c235de55c22d9"},
-    {file = "websockets-15.0.1-pp310-pypy310_pp73-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:b359ed09954d7c18bbc1680f380c7301f92c60bf924171629c5db97febb12f04"},
-    {file = "websockets-15.0.1-pp310-pypy310_pp73-win_amd64.whl", hash = "sha256:cad21560da69f4ce7658ca2cb83138fb4cf695a2ba3e475e0559e05991aa8122"},
-    {file = "websockets-15.0.1-pp39-pypy39_pp73-macosx_10_15_x86_64.whl", hash = "sha256:7f493881579c90fc262d9cdbaa05a6b54b3811c2f300766748db79f098db9940"},
-    {file = "websockets-15.0.1-pp39-pypy39_pp73-macosx_11_0_arm64.whl", hash = "sha256:47b099e1f4fbc95b701b6e85768e1fcdaf1630f3cbe4765fa216596f12310e2e"},
-    {file = "websockets-15.0.1-pp39-pypy39_pp73-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:67f2b6de947f8c757db2db9c71527933ad0019737ec374a8a6be9a956786aaf9"},
-    {file = "websockets-15.0.1-pp39-pypy39_pp73-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:d08eb4c2b7d6c41da6ca0600c077e93f5adcfd979cd777d747e9ee624556da4b"},
-    {file = "websockets-15.0.1-pp39-pypy39_pp73-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:4b826973a4a2ae47ba357e4e82fa44a463b8f168e1ca775ac64521442b19e87f"},
-    {file = "websockets-15.0.1-pp39-pypy39_pp73-win_amd64.whl", hash = "sha256:21c1fa28a6a7e3cbdc171c694398b6df4744613ce9b36b1a498e816787e28123"},
-    {file = "websockets-15.0.1-py3-none-any.whl", hash = "sha256:f7a866fbc1e97b5c617ee4116daaa09b722101d4a3c170c787450ba409f9736f"},
-    {file = "websockets-15.0.1.tar.gz", hash = "sha256:82544de02076bafba038ce055ee6412d68da13ab47f0c60cab827346de828dee"},
-]
-
-[metadata]
-lock-version = "2.1"
-python-versions = "^3.11"
-content-hash = "d36bec21c1c3072e8cc9fa7852316bf6046a8ea7c61f73315dc16e6f1e3f281c"
diff --git a/paddler_client_python/pyproject.toml b/paddler_client_python/pyproject.toml
deleted file mode 100644
index 24a6efac4..000000000
--- a/paddler_client_python/pyproject.toml
+++ /dev/null
@@ -1,89 +0,0 @@
-[tool.poetry]
-name = "paddler-client"
-version = "4.0.1"
-description = "Python client for Paddler LLM load balancer"
-authors = ["Intentee"]
-license = "Apache-2.0"
-readme = "README.md"
-homepage = "https://github.com/intentee/paddler"
-repository = "https://github.com/intentee/paddler"
-keywords = ["llm", "load-balancer", "inference", "paddler"]
-classifiers = [
-    "Typing :: Typed",
-    "Framework :: Pydantic :: 2",
-]
-packages = [{include = "paddler_client"}]
-
-[tool.poetry.dependencies]
-python = "^3.11"
-httpx = "==0.28.1"
-websockets = "==15.0.1"
-pydantic = "==2.12.5"
-
-[tool.poetry.group.dev.dependencies]
-paddler-test-cluster = { path = "../paddler_test_cluster_python", develop = true }
-pytest = "9.0.3"
-pytest-asyncio = "1.3.0"
-pytest-cov = "7.1.0"
-pygments = "2.20.0"
-pyright = "1.1.408"
-ruff = "0.15.12"
-mypy = "1.19.1"
-
-[build-system]
-requires = ["poetry-core"]
-build-backend = "poetry.core.masonry.api"
-
-[tool.mypy]
-python_version = "3.11"
-strict = true
-extra_checks = true
-plugins = ["pydantic.mypy"]
-
-[tool.pydantic-mypy]
-init_forbid_extra = true
-init_typed = true
-warn_required_dynamic_aliases = true
-warn_untyped_fields = true
-
-[tool.pyright]
-pythonVersion = "3.11"
-pythonPlatform = "All"
-typeCheckingMode = "strict"
-
-[tool.ruff]
-line-length = 88
-target-version = "py311"
-
-[tool.ruff.lint]
-select = ["ALL"]
-ignore = [
-    "ANN401",  # Any type allowed for JSON data
-    "COM812",  # trailing comma - conflicts with formatter
-    "D",       # docstrings - project prefers self-documenting code
-    "FBT",     # boolean arguments - standard for pydantic models and config
-    "ISC001",  # implicit string concatenation - conflicts with formatter
-    "S105",    # false positive on enum member names like TOKEN
-    "TRY003",  # long exception messages - fine for ValueError with context
-]
-
-[tool.ruff.lint.per-file-ignores]
-"tests/**/*.py" = [
-    "PERF401", # async for with append - clearer than comprehension in tests
-    "PLR2004", # magic values in assertions
-    "S101",    # assert statements
-    "S106",    # false positive on token keyword argument
-    "SIM117",  # nested with statements - async with can't combine easily
-]
-
-[tool.coverage.report]
-fail_under = 100
-show_missing = true
-
-[tool.coverage.run]
-branch = true
-source = ["paddler_client", "tests"]
-
-[tool.pytest.ini_options]
-asyncio_mode = "auto"
-addopts = "--cov-report=term-missing"
diff --git a/paddler_client_python/tests/cluster/conftest.py b/paddler_client_python/tests/cluster/conftest.py
deleted file mode 100644
index d8e666a70..000000000
--- a/paddler_client_python/tests/cluster/conftest.py
+++ /dev/null
@@ -1,28 +0,0 @@
-from collections.abc import AsyncIterator
-from datetime import timedelta
-
-import pytest
-from paddler_test_cluster.agent_spec import AgentSpec
-from paddler_test_cluster.balancer_addresses import BalancerAddresses
-from paddler_test_cluster.paddler_cluster import paddler_cluster
-
-from paddler_client.balancer_desired_state import BalancerDesiredState
-
-MODEL_LESS_AGENT = AgentSpec(name="model-less-agent", slots=1)
-SHORT_BUFFERED_REQUEST_TIMEOUT = timedelta(milliseconds=50)
-
-
-@pytest.fixture
-async def cluster_with_a_model_less_agent() -> AsyncIterator[BalancerAddresses]:
-    async with paddler_cluster(BalancerDesiredState(), [MODEL_LESS_AGENT]) as addresses:
-        yield addresses
-
-
-@pytest.fixture
-async def cluster_without_agents() -> AsyncIterator[BalancerAddresses]:
-    async with paddler_cluster(
-        BalancerDesiredState(),
-        [],
-        buffered_request_timeout=SHORT_BUFFERED_REQUEST_TIMEOUT,
-    ) as addresses:
-        yield addresses
diff --git a/paddler_client_python/tests/cluster/test_inference_over_http_times_out_without_agents.py b/paddler_client_python/tests/cluster/test_inference_over_http_times_out_without_agents.py
deleted file mode 100644
index f70b40799..000000000
--- a/paddler_client_python/tests/cluster/test_inference_over_http_times_out_without_agents.py
+++ /dev/null
@@ -1,34 +0,0 @@
-from paddler_test_cluster.balancer_addresses import BalancerAddresses
-
-from paddler_client.client_inference import ClientInference
-from paddler_client.continue_from_conversation_history_params import (
-    ContinueFromConversationHistoryParams,
-)
-from paddler_client.conversation_message import ConversationMessage
-from paddler_client.inference_message import InferenceMessageKind
-
-GATEWAY_TIMEOUT = 504
-
-
-async def test_inference_over_http_times_out_without_agents(
-    cluster_without_agents: BalancerAddresses,
-) -> None:
-    async with ClientInference(url=cluster_without_agents.inference_url) as client:
-        messages = [
-            message
-            async for message in client.post_continue_from_conversation_history(
-                ContinueFromConversationHistoryParams(
-                    add_generation_prompt=True,
-                    conversation_history=[
-                        ConversationMessage(content="Hello", role="user")
-                    ],
-                    enable_thinking=False,
-                    max_tokens=1,
-                )
-            )
-        ]
-
-    [message] = messages
-
-    assert message.kind == InferenceMessageKind.SERVER_ERROR
-    assert message.error_code == GATEWAY_TIMEOUT
diff --git a/paddler_client_python/tests/cluster/test_inference_reports_its_health.py b/paddler_client_python/tests/cluster/test_inference_reports_its_health.py
deleted file mode 100644
index 0261bdc73..000000000
--- a/paddler_client_python/tests/cluster/test_inference_reports_its_health.py
+++ /dev/null
@@ -1,10 +0,0 @@
-from paddler_test_cluster.balancer_addresses import BalancerAddresses
-
-from paddler_client.client_inference import ClientInference
-
-
-async def test_inference_reports_its_health(
-    cluster_without_agents: BalancerAddresses,
-) -> None:
-    async with ClientInference(url=cluster_without_agents.inference_url) as client:
-        assert await client.get_health() == "OK"
diff --git a/paddler_client_python/tests/cluster/test_inference_reports_unsuccessful_responses.py b/paddler_client_python/tests/cluster/test_inference_reports_unsuccessful_responses.py
deleted file mode 100644
index e8662bba8..000000000
--- a/paddler_client_python/tests/cluster/test_inference_reports_unsuccessful_responses.py
+++ /dev/null
@@ -1,69 +0,0 @@
-from typing import TYPE_CHECKING
-
-import pytest
-from paddler_test_cluster.balancer_addresses import BalancerAddresses
-
-from paddler_client.client_inference import ClientInference
-from paddler_client.continue_from_conversation_history_params import (
-    ContinueFromConversationHistoryParams,
-)
-from paddler_client.conversation_message import ConversationMessage
-from paddler_client.embedding_input_document import EmbeddingInputDocument
-from paddler_client.embedding_normalization_method import (
-    EmbeddingNormalizationMethod,
-)
-from paddler_client.error import HttpError
-from paddler_client.generate_embedding_batch_params import (
-    GenerateEmbeddingBatchParams,
-)
-
-if TYPE_CHECKING:
-    from collections.abc import Awaitable, Callable
-
-NOT_FOUND = 404
-
-
-async def test_inference_reports_unsuccessful_responses(
-    cluster_without_agents: BalancerAddresses,
-) -> None:
-    async with ClientInference(
-        url=f"{cluster_without_agents.inference_url}/not-a-paddler-route"
-    ) as client:
-
-        async def first_generated_message() -> object:
-            return await anext(
-                client.post_continue_from_conversation_history(
-                    ContinueFromConversationHistoryParams(
-                        add_generation_prompt=True,
-                        conversation_history=[
-                            ConversationMessage(content="Hello", role="user")
-                        ],
-                        enable_thinking=False,
-                        max_tokens=1,
-                    )
-                )
-            )
-
-        async def first_embedding_message() -> object:
-            return await anext(
-                client.generate_embedding_batch(
-                    GenerateEmbeddingBatchParams(
-                        input_batch=[
-                            EmbeddingInputDocument(content="Hello", id="document")
-                        ],
-                        normalization_method=EmbeddingNormalizationMethod.none(),
-                    )
-                )
-            )
-
-        requests: list[Callable[[], Awaitable[object]]] = [
-            client.get_health,
-            first_generated_message,
-            first_embedding_message,
-        ]
-
-        for request in requests:
-            with pytest.raises(HttpError) as unsuccessful:
-                await request()
-
-            assert unsuccessful.value.status_code == NOT_FOUND
diff --git a/paddler_client_python/tests/cluster/test_inference_socket_answers_consecutive_requests_without_agents.py b/paddler_client_python/tests/cluster/test_inference_socket_answers_consecutive_requests_without_agents.py
deleted file mode 100644
index d5f990802..000000000
--- a/paddler_client_python/tests/cluster/test_inference_socket_answers_consecutive_requests_without_agents.py
+++ /dev/null
@@ -1,28 +0,0 @@
-from paddler_test_cluster.balancer_addresses import BalancerAddresses
-
-from paddler_client.client_inference import ClientInference
-from paddler_client.continue_from_raw_prompt_params import (
-    ContinueFromRawPromptParams,
-)
-
-GATEWAY_TIMEOUT = 504
-
-
-async def test_inference_socket_answers_consecutive_requests_without_agents(
-    cluster_without_agents: BalancerAddresses,
-) -> None:
-    async with ClientInference(
-        url=cluster_without_agents.inference_url,
-        socket_pool_size=1,
-    ) as client:
-        answers = [
-            [
-                message.error_code
-                async for message in await client.continue_from_raw_prompt(
-                    ContinueFromRawPromptParams(max_tokens=1, raw_prompt=raw_prompt)
-                )
-            ]
-            for raw_prompt in ("first", "second")
-        ]
-
-    assert answers == [[GATEWAY_TIMEOUT], [GATEWAY_TIMEOUT]]
diff --git a/paddler_client_python/tests/cluster/test_inference_socket_connection_rejects_requests_after_it_closes.py b/paddler_client_python/tests/cluster/test_inference_socket_connection_rejects_requests_after_it_closes.py
deleted file mode 100644
index 107780421..000000000
--- a/paddler_client_python/tests/cluster/test_inference_socket_connection_rejects_requests_after_it_closes.py
+++ /dev/null
@@ -1,21 +0,0 @@
-import pytest
-from paddler_test_cluster.balancer_addresses import BalancerAddresses
-
-from paddler_client.error import ConnectionDroppedError
-from paddler_client.inference_socket_connection import InferenceSocketConnection
-from paddler_client.inference_socket_url import inference_socket_url
-
-
-async def test_inference_socket_connection_rejects_requests_after_it_closes(
-    cluster_without_agents: BalancerAddresses,
-) -> None:
-    connection = await InferenceSocketConnection.connect(
-        inference_socket_url(cluster_without_agents.inference_url)
-    )
-
-    await connection.close()
-
-    with pytest.raises(ConnectionDroppedError) as connection_dropped:
-        await connection.send("request-after-close", "{}")
-
-    assert connection_dropped.value.request_id == "request-after-close"
diff --git a/paddler_client_python/tests/cluster/test_inference_socket_reconnects_after_the_balancer_closes_it.py b/paddler_client_python/tests/cluster/test_inference_socket_reconnects_after_the_balancer_closes_it.py
deleted file mode 100644
index 864dc2911..000000000
--- a/paddler_client_python/tests/cluster/test_inference_socket_reconnects_after_the_balancer_closes_it.py
+++ /dev/null
@@ -1,38 +0,0 @@
-import pytest
-from paddler_test_cluster.balancer_addresses import BalancerAddresses
-
-from paddler_client.client_inference import ClientInference
-from paddler_client.continue_from_raw_prompt_params import (
-    ContinueFromRawPromptParams,
-)
-from paddler_client.error import ConnectionDroppedError
-
-BALANCER_MAX_WEBSOCKET_FRAME_BYTES = 50 * 1024 * 1024
-GATEWAY_TIMEOUT = 504
-
-
-async def test_inference_socket_reconnects_after_the_balancer_closes_it(
-    cluster_without_agents: BalancerAddresses,
-) -> None:
-    async with ClientInference(
-        url=cluster_without_agents.inference_url,
-        socket_pool_size=1,
-    ) as client:
-        oversized_request = await client.continue_from_raw_prompt(
-            ContinueFromRawPromptParams(
-                max_tokens=1,
-                raw_prompt="x" * BALANCER_MAX_WEBSOCKET_FRAME_BYTES,
-            )
-        )
-
-        with pytest.raises(ConnectionDroppedError):
-            await anext(oversized_request)
-
-        follow_up_answer = [
-            message.error_code
-            async for message in await client.continue_from_raw_prompt(
-                ContinueFromRawPromptParams(max_tokens=1, raw_prompt="Hello")
-            )
-        ]
-
-    assert follow_up_answer == [GATEWAY_TIMEOUT]
diff --git a/paddler_client_python/tests/cluster/test_inference_socket_reports_disabled_token_generation.py b/paddler_client_python/tests/cluster/test_inference_socket_reports_disabled_token_generation.py
deleted file mode 100644
index 0d785a37a..000000000
--- a/paddler_client_python/tests/cluster/test_inference_socket_reports_disabled_token_generation.py
+++ /dev/null
@@ -1,42 +0,0 @@
-import json
-
-from paddler_test_cluster.paddler_cluster import paddler_cluster
-from websockets.asyncio.client import connect
-
-from paddler_client.balancer_desired_state import BalancerDesiredState
-from paddler_client.client_inference import ClientInference
-from paddler_client.continue_from_raw_prompt_params import (
-    ContinueFromRawPromptParams,
-)
-from paddler_client.inference_message import InferenceMessageKind
-from paddler_client.inference_parameters import InferenceParameters
-from paddler_client.inference_socket_url import inference_socket_url
-
-
-async def wait_until_token_generation_is_disabled(inference_url: str) -> None:
-    async with connect(inference_socket_url(inference_url)) as websocket:
-        assert json.loads(await websocket.recv()) == {
-            "Notification": "TokenGenerationDisabled"
-        }
-
-
-async def test_inference_socket_reports_disabled_token_generation() -> None:
-    async with (
-        paddler_cluster(
-            BalancerDesiredState(
-                inference_parameters=InferenceParameters(enable_embeddings=True)
-            ),
-            [],
-        ) as addresses,
-        ClientInference(url=addresses.inference_url, socket_pool_size=1) as client,
-    ):
-        await wait_until_token_generation_is_disabled(addresses.inference_url)
-
-        kinds = [
-            message.kind
-            async for message in await client.continue_from_raw_prompt(
-                ContinueFromRawPromptParams(max_tokens=1, raw_prompt="Hello")
-            )
-        ]
-
-    assert kinds == [InferenceMessageKind.TOKEN_GENERATION_DISABLED]
diff --git a/paddler_client_python/tests/cluster/test_management_agents_stream_ends_when_the_balancer_shuts_down.py b/paddler_client_python/tests/cluster/test_management_agents_stream_ends_when_the_balancer_shuts_down.py
deleted file mode 100644
index 8e77949d3..000000000
--- a/paddler_client_python/tests/cluster/test_management_agents_stream_ends_when_the_balancer_shuts_down.py
+++ /dev/null
@@ -1,23 +0,0 @@
-from contextlib import aclosing
-
-from paddler_test_cluster.spawned_balancer import SpawnedBalancer
-
-from paddler_client.client_management import ClientManagement
-
-
-async def test_management_agents_stream_ends_when_the_balancer_shuts_down() -> None:
-    balancer = await SpawnedBalancer.spawn()
-
-    try:
-        async with (
-            ClientManagement(url=balancer.addresses.management_url) as client,
-            aclosing(client.agents_stream()) as snapshots,
-        ):
-            await anext(snapshots)
-            await balancer.process.terminate()
-
-            remaining_snapshots = [snapshot async for snapshot in snapshots]
-    finally:
-        await balancer.process.terminate()
-
-    assert remaining_snapshots == []
diff --git a/paddler_client_python/tests/cluster/test_management_exposes_prometheus_gauges.py b/paddler_client_python/tests/cluster/test_management_exposes_prometheus_gauges.py
deleted file mode 100644
index 286003eff..000000000
--- a/paddler_client_python/tests/cluster/test_management_exposes_prometheus_gauges.py
+++ /dev/null
@@ -1,24 +0,0 @@
-from paddler_test_cluster.balancer_addresses import BalancerAddresses
-
-from paddler_client.client_management import ClientManagement
-
-IDLE_CLUSTER_METRICS = """\
-# HELP paddler_slots_processing Number of processing slots
-# TYPE paddler_slots_processing gauge
-paddler_slots_processing 0
-
-# HELP paddler_slots_total Number of total slots
-# TYPE paddler_slots_total gauge
-paddler_slots_total 0
-
-# HELP paddler_requests_buffered Number of buffered requests
-# TYPE paddler_requests_buffered gauge
-paddler_requests_buffered 0
-"""
-
-
-async def test_management_exposes_prometheus_gauges(
-    cluster_without_agents: BalancerAddresses,
-) -> None:
-    async with ClientManagement(url=cluster_without_agents.management_url) as client:
-        assert await client.get_metrics() == IDLE_CLUSTER_METRICS
diff --git a/paddler_client_python/tests/cluster/test_management_lists_registered_agents.py b/paddler_client_python/tests/cluster/test_management_lists_registered_agents.py
deleted file mode 100644
index eb0a5b229..000000000
--- a/paddler_client_python/tests/cluster/test_management_lists_registered_agents.py
+++ /dev/null
@@ -1,31 +0,0 @@
-from paddler_test_cluster.balancer_addresses import BalancerAddresses
-
-from paddler_client.agent_state_application_status import (
-    AgentStateApplicationStatus,
-)
-from paddler_client.agent_status import AgentStatus
-from paddler_client.client_management import ClientManagement
-from paddler_client.model_download_status import ModelDownloadStatus
-from tests.cluster.conftest import MODEL_LESS_AGENT
-
-
-async def test_management_lists_registered_agents(
-    cluster_with_a_model_less_agent: BalancerAddresses,
-) -> None:
-    async with ClientManagement(
-        url=cluster_with_a_model_less_agent.management_url
-    ) as client:
-        snapshot = await client.get_agents()
-
-    assert [agent.name for agent in snapshot.agents] == [MODEL_LESS_AGENT.name]
-    assert [agent.status for agent in snapshot.agents] == [
-        AgentStatus(
-            desired_slots_total=MODEL_LESS_AGENT.slots,
-            download_status=ModelDownloadStatus(variant="NotDownloading"),
-            issues=[],
-            model_path=None,
-            slots_total=0,
-            state_application_status=AgentStateApplicationStatus.APPLIED,
-            uses_chat_template_override=False,
-        )
-    ]
diff --git a/paddler_client_python/tests/cluster/test_management_reports_its_health.py b/paddler_client_python/tests/cluster/test_management_reports_its_health.py
deleted file mode 100644
index ac4e9fc2f..000000000
--- a/paddler_client_python/tests/cluster/test_management_reports_its_health.py
+++ /dev/null
@@ -1,10 +0,0 @@
-from paddler_test_cluster.balancer_addresses import BalancerAddresses
-
-from paddler_client.client_management import ClientManagement
-
-
-async def test_management_reports_its_health(
-    cluster_without_agents: BalancerAddresses,
-) -> None:
-    async with ClientManagement(url=cluster_without_agents.management_url) as client:
-        assert await client.get_health() == "OK"
diff --git a/paddler_client_python/tests/cluster/test_management_reports_no_buffered_requests_while_idle.py b/paddler_client_python/tests/cluster/test_management_reports_no_buffered_requests_while_idle.py
deleted file mode 100644
index e07d40a1a..000000000
--- a/paddler_client_python/tests/cluster/test_management_reports_no_buffered_requests_while_idle.py
+++ /dev/null
@@ -1,12 +0,0 @@
-from paddler_test_cluster.balancer_addresses import BalancerAddresses
-
-from paddler_client.client_management import ClientManagement
-
-
-async def test_management_reports_no_buffered_requests_while_idle(
-    cluster_without_agents: BalancerAddresses,
-) -> None:
-    async with ClientManagement(url=cluster_without_agents.management_url) as client:
-        snapshot = await client.get_buffered_requests()
-
-    assert snapshot.buffered_requests_current == 0
diff --git a/paddler_client_python/tests/cluster/test_management_reports_no_chat_template_override_for_an_agent_without_one.py b/paddler_client_python/tests/cluster/test_management_reports_no_chat_template_override_for_an_agent_without_one.py
deleted file mode 100644
index 94239aa7d..000000000
--- a/paddler_client_python/tests/cluster/test_management_reports_no_chat_template_override_for_an_agent_without_one.py
+++ /dev/null
@@ -1,14 +0,0 @@
-from paddler_test_cluster.balancer_addresses import BalancerAddresses
-
-from paddler_client.client_management import ClientManagement
-
-
-async def test_management_reports_no_chat_template_override_for_an_agent_without_one(
-    cluster_with_a_model_less_agent: BalancerAddresses,
-) -> None:
-    async with ClientManagement(
-        url=cluster_with_a_model_less_agent.management_url
-    ) as client:
-        [agent] = (await client.get_agents()).agents
-
-        assert await client.get_chat_template_override(agent.id) is None
diff --git a/paddler_client_python/tests/cluster/test_management_reports_no_model_metadata_for_an_agent_without_a_model.py b/paddler_client_python/tests/cluster/test_management_reports_no_model_metadata_for_an_agent_without_a_model.py
deleted file mode 100644
index ead0fab65..000000000
--- a/paddler_client_python/tests/cluster/test_management_reports_no_model_metadata_for_an_agent_without_a_model.py
+++ /dev/null
@@ -1,14 +0,0 @@
-from paddler_test_cluster.balancer_addresses import BalancerAddresses
-
-from paddler_client.client_management import ClientManagement
-
-
-async def test_management_reports_no_model_metadata_for_an_agent_without_a_model(
-    cluster_with_a_model_less_agent: BalancerAddresses,
-) -> None:
-    async with ClientManagement(
-        url=cluster_with_a_model_less_agent.management_url
-    ) as client:
-        [agent] = (await client.get_agents()).agents
-
-        assert await client.get_model_metadata(agent.id) is None
diff --git a/paddler_client_python/tests/cluster/test_management_reports_the_chat_template_override_an_agent_applies.py b/paddler_client_python/tests/cluster/test_management_reports_the_chat_template_override_an_agent_applies.py
deleted file mode 100644
index 5371305eb..000000000
--- a/paddler_client_python/tests/cluster/test_management_reports_the_chat_template_override_an_agent_applies.py
+++ /dev/null
@@ -1,24 +0,0 @@
-from paddler_test_cluster.paddler_cluster import paddler_cluster
-
-from paddler_client.balancer_desired_state import BalancerDesiredState
-from paddler_client.chat_template import ChatTemplate
-from paddler_client.client_management import ClientManagement
-from tests.cluster.conftest import MODEL_LESS_AGENT
-
-CHAT_TEMPLATE = ChatTemplate(content="{{ messages[0].content }}")
-
-
-async def test_management_reports_the_chat_template_override_an_agent_applies() -> None:
-    async with (
-        paddler_cluster(
-            BalancerDesiredState(
-                chat_template_override=CHAT_TEMPLATE,
-                use_chat_template_override=True,
-            ),
-            [MODEL_LESS_AGENT],
-        ) as addresses,
-        ClientManagement(url=addresses.management_url) as client,
-    ):
-        [agent] = (await client.get_agents()).agents
-
-        assert await client.get_chat_template_override(agent.id) == CHAT_TEMPLATE
diff --git a/paddler_client_python/tests/cluster/test_management_reports_unsuccessful_responses.py b/paddler_client_python/tests/cluster/test_management_reports_unsuccessful_responses.py
deleted file mode 100644
index 65c0152e1..000000000
--- a/paddler_client_python/tests/cluster/test_management_reports_unsuccessful_responses.py
+++ /dev/null
@@ -1,49 +0,0 @@
-from contextlib import aclosing
-from typing import TYPE_CHECKING
-
-import pytest
-from paddler_test_cluster.balancer_addresses import BalancerAddresses
-
-from paddler_client.balancer_desired_state import BalancerDesiredState
-from paddler_client.client_management import ClientManagement
-from paddler_client.error import HttpError
-
-if TYPE_CHECKING:
-    from collections.abc import Awaitable, Callable
-
-NOT_FOUND = 404
-
-
-async def test_management_reports_unsuccessful_responses(
-    cluster_without_agents: BalancerAddresses,
-) -> None:
-    async with ClientManagement(
-        url=f"{cluster_without_agents.management_url}/not-a-paddler-route"
-    ) as client:
-
-        async def first_agents_snapshot() -> object:
-            async with aclosing(client.agents_stream()) as snapshots:
-                return await anext(snapshots)
-
-        async def first_buffered_requests_snapshot() -> object:
-            async with aclosing(client.buffered_requests_stream()) as snapshots:
-                return await anext(snapshots)
-
-        requests: list[Callable[[], Awaitable[object]]] = [
-            client.get_health,
-            client.get_agents,
-            first_agents_snapshot,
-            client.get_balancer_desired_state,
-            lambda: client.put_balancer_desired_state(BalancerDesiredState()),
-            client.get_buffered_requests,
-            first_buffered_requests_snapshot,
-            lambda: client.get_chat_template_override("agent-id"),
-            lambda: client.get_model_metadata("agent-id"),
-            client.get_metrics,
-        ]
-
-        for request in requests:
-            with pytest.raises(HttpError) as unsuccessful:
-                await request()
-
-            assert unsuccessful.value.status_code == NOT_FOUND
diff --git a/paddler_client_python/tests/cluster/test_management_serves_the_desired_state_it_stores.py b/paddler_client_python/tests/cluster/test_management_serves_the_desired_state_it_stores.py
deleted file mode 100644
index 218664e31..000000000
--- a/paddler_client_python/tests/cluster/test_management_serves_the_desired_state_it_stores.py
+++ /dev/null
@@ -1,18 +0,0 @@
-from paddler_test_cluster.balancer_addresses import BalancerAddresses
-
-from paddler_client.agent_desired_model import AgentDesiredModel
-from paddler_client.balancer_desired_state import BalancerDesiredState
-from paddler_client.client_management import ClientManagement
-
-
-async def test_management_serves_the_desired_state_it_stores(
-    cluster_without_agents: BalancerAddresses,
-) -> None:
-    desired_state = BalancerDesiredState(
-        model=AgentDesiredModel.local_to_agent("/models/stored.gguf"),
-    )
-
-    async with ClientManagement(url=cluster_without_agents.management_url) as client:
-        await client.put_balancer_desired_state(desired_state)
-
-        assert await client.get_balancer_desired_state() == desired_state
diff --git a/paddler_client_python/tests/cluster/test_management_streams_agent_snapshots.py b/paddler_client_python/tests/cluster/test_management_streams_agent_snapshots.py
deleted file mode 100644
index d18a787df..000000000
--- a/paddler_client_python/tests/cluster/test_management_streams_agent_snapshots.py
+++ /dev/null
@@ -1,18 +0,0 @@
-from contextlib import aclosing
-
-from paddler_test_cluster.balancer_addresses import BalancerAddresses
-
-from paddler_client.client_management import ClientManagement
-from tests.cluster.conftest import MODEL_LESS_AGENT
-
-
-async def test_management_streams_agent_snapshots(
-    cluster_with_a_model_less_agent: BalancerAddresses,
-) -> None:
-    async with (
-        ClientManagement(url=cluster_with_a_model_less_agent.management_url) as client,
-        aclosing(client.agents_stream()) as snapshots,
-    ):
-        snapshot = await anext(snapshots)
-
-    assert [agent.name for agent in snapshot.agents] == [MODEL_LESS_AGENT.name]
diff --git a/paddler_client_python/tests/cluster/test_management_streams_buffered_request_snapshots.py b/paddler_client_python/tests/cluster/test_management_streams_buffered_request_snapshots.py
deleted file mode 100644
index c4fbdd31f..000000000
--- a/paddler_client_python/tests/cluster/test_management_streams_buffered_request_snapshots.py
+++ /dev/null
@@ -1,17 +0,0 @@
-from contextlib import aclosing
-
-from paddler_test_cluster.balancer_addresses import BalancerAddresses
-
-from paddler_client.client_management import ClientManagement
-
-
-async def test_management_streams_buffered_request_snapshots(
-    cluster_without_agents: BalancerAddresses,
-) -> None:
-    async with (
-        ClientManagement(url=cluster_without_agents.management_url) as client,
-        aclosing(client.buffered_requests_stream()) as snapshots,
-    ):
-        snapshot = await anext(snapshots)
-
-    assert snapshot.buffered_requests_current == 0
diff --git a/paddler_client_python/tests/llm/conftest.py b/paddler_client_python/tests/llm/conftest.py
deleted file mode 100644
index 896d4b557..000000000
--- a/paddler_client_python/tests/llm/conftest.py
+++ /dev/null
@@ -1,27 +0,0 @@
-from collections.abc import AsyncIterator
-
-import pytest
-from paddler_test_cluster.agent_spec import AgentSpec
-from paddler_test_cluster.balancer_addresses import BalancerAddresses
-from paddler_test_cluster.nomic_embed_text_v1_5_desired_state import (
-    NOMIC_EMBED_TEXT_V1_5_DESIRED_STATE,
-)
-from paddler_test_cluster.paddler_cluster import paddler_cluster
-from paddler_test_cluster.qwen3_0_6b_desired_state import QWEN3_0_6B_DESIRED_STATE
-
-QWEN3_AGENT = AgentSpec(name="qwen3-agent", slots=2)
-NOMIC_EMBED_AGENT = AgentSpec(name="nomic-embed-agent", slots=1)
-
-
-@pytest.fixture
-async def qwen3_cluster() -> AsyncIterator[BalancerAddresses]:
-    async with paddler_cluster(QWEN3_0_6B_DESIRED_STATE, [QWEN3_AGENT]) as addresses:
-        yield addresses
-
-
-@pytest.fixture
-async def nomic_embed_cluster() -> AsyncIterator[BalancerAddresses]:
-    async with paddler_cluster(
-        NOMIC_EMBED_TEXT_V1_5_DESIRED_STATE, [NOMIC_EMBED_AGENT]
-    ) as addresses:
-        yield addresses
diff --git a/paddler_client_python/tests/llm/test_conversation_history_streams_tokens_over_http.py b/paddler_client_python/tests/llm/test_conversation_history_streams_tokens_over_http.py
deleted file mode 100644
index 64cf541c8..000000000
--- a/paddler_client_python/tests/llm/test_conversation_history_streams_tokens_over_http.py
+++ /dev/null
@@ -1,35 +0,0 @@
-from paddler_test_cluster.balancer_addresses import BalancerAddresses
-
-from paddler_client.client_inference import ClientInference
-from paddler_client.continue_from_conversation_history_params import (
-    ContinueFromConversationHistoryParams,
-)
-from paddler_client.conversation_message import ConversationMessage
-from paddler_client.inference_message import InferenceMessageKind
-
-MAX_TOKENS = 8
-
-
-async def test_conversation_history_streams_tokens_over_http(
-    qwen3_cluster: BalancerAddresses,
-) -> None:
-    async with ClientInference(url=qwen3_cluster.inference_url) as client:
-        messages = [
-            message
-            async for message in client.post_continue_from_conversation_history(
-                ContinueFromConversationHistoryParams(
-                    add_generation_prompt=True,
-                    conversation_history=[
-                        ConversationMessage(content="Say hello.", role="user")
-                    ],
-                    enable_thinking=False,
-                    max_tokens=MAX_TOKENS,
-                )
-            )
-        ]
-
-    *tokens, done = messages
-
-    assert {token.kind for token in tokens} == {InferenceMessageKind.CONTENT_TOKEN}
-    assert done.kind == InferenceMessageKind.DONE
-    assert len(tokens) == MAX_TOKENS
diff --git a/paddler_client_python/tests/llm/test_conversation_history_streams_tokens_over_the_inference_socket.py b/paddler_client_python/tests/llm/test_conversation_history_streams_tokens_over_the_inference_socket.py
deleted file mode 100644
index 23787c896..000000000
--- a/paddler_client_python/tests/llm/test_conversation_history_streams_tokens_over_the_inference_socket.py
+++ /dev/null
@@ -1,35 +0,0 @@
-from paddler_test_cluster.balancer_addresses import BalancerAddresses
-
-from paddler_client.client_inference import ClientInference
-from paddler_client.continue_from_conversation_history_params import (
-    ContinueFromConversationHistoryParams,
-)
-from paddler_client.conversation_message import ConversationMessage
-from paddler_client.inference_message import InferenceMessageKind
-
-MAX_TOKENS = 8
-
-
-async def test_conversation_history_streams_tokens_over_the_inference_socket(
-    qwen3_cluster: BalancerAddresses,
-) -> None:
-    async with ClientInference(url=qwen3_cluster.inference_url) as client:
-        messages = [
-            message
-            async for message in await client.continue_from_conversation_history(
-                ContinueFromConversationHistoryParams(
-                    add_generation_prompt=True,
-                    conversation_history=[
-                        ConversationMessage(content="Say hello.", role="user")
-                    ],
-                    enable_thinking=False,
-                    max_tokens=MAX_TOKENS,
-                )
-            )
-        ]
-
-    *tokens, done = messages
-
-    assert {token.kind for token in tokens} == {InferenceMessageKind.CONTENT_TOKEN}
-    assert done.kind == InferenceMessageKind.DONE
-    assert len(tokens) == MAX_TOKENS
diff --git a/paddler_client_python/tests/llm/test_conversation_with_an_invalid_tool_schema_is_rejected.py b/paddler_client_python/tests/llm/test_conversation_with_an_invalid_tool_schema_is_rejected.py
deleted file mode 100644
index 395f757cb..000000000
--- a/paddler_client_python/tests/llm/test_conversation_with_an_invalid_tool_schema_is_rejected.py
+++ /dev/null
@@ -1,46 +0,0 @@
-from paddler_test_cluster.balancer_addresses import BalancerAddresses
-
-from paddler_client.client_inference import ClientInference
-from paddler_client.continue_from_conversation_history_params import (
-    ContinueFromConversationHistoryParams,
-)
-from paddler_client.conversation_message import ConversationMessage
-from paddler_client.inference_message import InferenceMessageKind
-from paddler_client.tool import Function, Tool
-from paddler_client.validated_parameters_schema import ValidatedParametersSchema
-
-
-async def test_conversation_with_an_invalid_tool_schema_is_rejected(
-    qwen3_cluster: BalancerAddresses,
-) -> None:
-    async with ClientInference(url=qwen3_cluster.inference_url) as client:
-        kinds = [
-            message.kind
-            async for message in client.post_continue_from_conversation_history(
-                ContinueFromConversationHistoryParams(
-                    add_generation_prompt=True,
-                    conversation_history=[
-                        ConversationMessage(
-                            content="What is the weather in Paris?", role="user"
-                        )
-                    ],
-                    enable_thinking=False,
-                    max_tokens=64,
-                    parse_tool_calls=True,
-                    tools=[
-                        Tool(
-                            function=Function(
-                                name="get_weather",
-                                description="Get the current weather for a location",
-                                parameters=ValidatedParametersSchema(
-                                    schema_type="object",
-                                    properties={"location": {"type": 123}},
-                                ),
-                            )
-                        )
-                    ],
-                )
-            )
-        ]
-
-    assert kinds == [InferenceMessageKind.TOOL_SCHEMA_INVALID]
diff --git a/paddler_client_python/tests/llm/test_embedding_batch_reports_documents_exceeding_the_batch_size.py b/paddler_client_python/tests/llm/test_embedding_batch_reports_documents_exceeding_the_batch_size.py
deleted file mode 100644
index 8bd5431df..000000000
--- a/paddler_client_python/tests/llm/test_embedding_batch_reports_documents_exceeding_the_batch_size.py
+++ /dev/null
@@ -1,65 +0,0 @@
-from paddler_test_cluster.nomic_embed_text_v1_5_desired_state import (
-    NOMIC_EMBED_TEXT_V1_5_DESIRED_STATE,
-)
-from paddler_test_cluster.paddler_cluster import paddler_cluster
-
-from paddler_client.client_inference import ClientInference
-from paddler_client.embedding_input_document import EmbeddingInputDocument
-from paddler_client.embedding_normalization_method import (
-    EmbeddingNormalizationMethod,
-)
-from paddler_client.generate_embedding_batch_params import (
-    GenerateEmbeddingBatchParams,
-)
-from paddler_client.inference_message import InferenceMessageKind
-from tests.llm.conftest import NOMIC_EMBED_AGENT
-
-N_BATCH = 64
-
-
-async def test_embedding_batch_reports_documents_exceeding_the_batch_size() -> None:
-    desired_state = NOMIC_EMBED_TEXT_V1_5_DESIRED_STATE.model_copy(
-        update={
-            "inference_parameters": (
-                NOMIC_EMBED_TEXT_V1_5_DESIRED_STATE.inference_parameters.model_copy(
-                    update={"n_batch": N_BATCH}
-                )
-            )
-        }
-    )
-
-    async with (
-        paddler_cluster(desired_state, [NOMIC_EMBED_AGENT]) as addresses,
-        ClientInference(url=addresses.inference_url) as client,
-    ):
-        messages = [
-            message
-            async for message in client.generate_embedding_batch(
-                GenerateEmbeddingBatchParams(
-                    input_batch=[
-                        EmbeddingInputDocument(
-                            content="The quick brown fox jumps over the lazy dog. "
-                            * 40,
-                            id="oversized",
-                        )
-                    ],
-                    normalization_method=EmbeddingNormalizationMethod.none(),
-                )
-            )
-        ]
-
-    oversized_documents = [
-        message.oversized_embedding_document_details
-        for message in messages
-        if message.kind == InferenceMessageKind.EMBEDDING_DOCUMENT_EXCEEDS_BATCH_SIZE
-    ]
-
-    assert [
-        details.source_document_id
-        for details in oversized_documents
-        if details is not None
-    ] == ["oversized"]
-    assert [
-        details.n_batch for details in oversized_documents if details is not None
-    ] == [N_BATCH]
-    assert messages[-1].kind == InferenceMessageKind.EMBEDDING_DONE
diff --git a/paddler_client_python/tests/llm/test_embedding_batch_returns_one_embedding_per_document.py b/paddler_client_python/tests/llm/test_embedding_batch_returns_one_embedding_per_document.py
deleted file mode 100644
index 08ee5b9d9..000000000
--- a/paddler_client_python/tests/llm/test_embedding_batch_returns_one_embedding_per_document.py
+++ /dev/null
@@ -1,42 +0,0 @@
-from paddler_test_cluster.balancer_addresses import BalancerAddresses
-
-from paddler_client.client_inference import ClientInference
-from paddler_client.embedding_input_document import EmbeddingInputDocument
-from paddler_client.embedding_normalization_method import (
-    EmbeddingNormalizationMethod,
-)
-from paddler_client.generate_embedding_batch_params import (
-    GenerateEmbeddingBatchParams,
-)
-from paddler_client.inference_message import InferenceMessageKind
-
-DOCUMENT_IDS = ["first", "second", "third"]
-
-
-async def test_embedding_batch_returns_one_embedding_per_document(
-    nomic_embed_cluster: BalancerAddresses,
-) -> None:
-    async with ClientInference(url=nomic_embed_cluster.inference_url) as client:
-        messages = [
-            message
-            async for message in client.generate_embedding_batch(
-                GenerateEmbeddingBatchParams(
-                    input_batch=[
-                        EmbeddingInputDocument(
-                            content=f"Document {document_id}", id=document_id
-                        )
-                        for document_id in DOCUMENT_IDS
-                    ],
-                    normalization_method=EmbeddingNormalizationMethod.l2(),
-                )
-            )
-        ]
-
-    embedded_document_ids = sorted(
-        message.embedding_data.source_document_id
-        for message in messages
-        if message.kind == InferenceMessageKind.EMBEDDING and message.embedding_data
-    )
-
-    assert embedded_document_ids == sorted(DOCUMENT_IDS)
-    assert messages[-1].kind == InferenceMessageKind.EMBEDDING_DONE
diff --git a/paddler_client_python/tests/llm/test_management_reports_the_model_metadata_of_a_loaded_model.py b/paddler_client_python/tests/llm/test_management_reports_the_model_metadata_of_a_loaded_model.py
deleted file mode 100644
index da087872a..000000000
--- a/paddler_client_python/tests/llm/test_management_reports_the_model_metadata_of_a_loaded_model.py
+++ /dev/null
@@ -1,14 +0,0 @@
-from paddler_test_cluster.balancer_addresses import BalancerAddresses
-
-from paddler_client.client_management import ClientManagement
-
-
-async def test_management_reports_the_model_metadata_of_a_loaded_model(
-    qwen3_cluster: BalancerAddresses,
-) -> None:
-    async with ClientManagement(url=qwen3_cluster.management_url) as client:
-        [agent] = (await client.get_agents()).agents
-        model_metadata = await client.get_model_metadata(agent.id)
-
-    assert model_metadata is not None
-    assert model_metadata.metadata["general.architecture"] == "qwen3"
diff --git a/paddler_client_python/tests/llm/test_raw_prompt_streams_tokens_over_the_inference_socket.py b/paddler_client_python/tests/llm/test_raw_prompt_streams_tokens_over_the_inference_socket.py
deleted file mode 100644
index 6f23b7bc3..000000000
--- a/paddler_client_python/tests/llm/test_raw_prompt_streams_tokens_over_the_inference_socket.py
+++ /dev/null
@@ -1,30 +0,0 @@
-from paddler_test_cluster.balancer_addresses import BalancerAddresses
-
-from paddler_client.client_inference import ClientInference
-from paddler_client.continue_from_raw_prompt_params import (
-    ContinueFromRawPromptParams,
-)
-from paddler_client.inference_message import InferenceMessageKind
-
-MAX_TOKENS = 8
-
-
-async def test_raw_prompt_streams_tokens_over_the_inference_socket(
-    qwen3_cluster: BalancerAddresses,
-) -> None:
-    async with ClientInference(url=qwen3_cluster.inference_url) as client:
-        messages = [
-            message
-            async for message in await client.continue_from_raw_prompt(
-                ContinueFromRawPromptParams(
-                    max_tokens=MAX_TOKENS,
-                    raw_prompt="The capital of France is",
-                )
-            )
-        ]
-
-    *tokens, done = messages
-
-    assert all(token.is_token for token in tokens)
-    assert done.kind == InferenceMessageKind.DONE
-    assert len(tokens) == MAX_TOKENS
diff --git a/paddler_client_python/tests/llm/test_tool_call_over_the_inference_socket_streams_until_done.py b/paddler_client_python/tests/llm/test_tool_call_over_the_inference_socket_streams_until_done.py
deleted file mode 100644
index f1f7c615a..000000000
--- a/paddler_client_python/tests/llm/test_tool_call_over_the_inference_socket_streams_until_done.py
+++ /dev/null
@@ -1,65 +0,0 @@
-from paddler_test_cluster.balancer_addresses import BalancerAddresses
-
-from paddler_client.client_inference import ClientInference
-from paddler_client.continue_from_conversation_history_params import (
-    ContinueFromConversationHistoryParams,
-)
-from paddler_client.conversation_message import ConversationMessage
-from paddler_client.inference_message import InferenceMessageKind
-from paddler_client.tool import Function, Tool
-from paddler_client.validated_parameters_schema import ValidatedParametersSchema
-
-
-async def test_tool_call_over_the_inference_socket_streams_until_done(
-    qwen3_cluster: BalancerAddresses,
-) -> None:
-    async with ClientInference(url=qwen3_cluster.inference_url) as client:
-        messages = [
-            message
-            async for message in await client.continue_from_conversation_history(
-                ContinueFromConversationHistoryParams(
-                    add_generation_prompt=True,
-                    conversation_history=[
-                        ConversationMessage(
-                            content=(
-                                "What is the weather in Paris? "
-                                "Use the get_weather tool to find out."
-                            ),
-                            role="user",
-                        )
-                    ],
-                    enable_thinking=False,
-                    max_tokens=400,
-                    parse_tool_calls=True,
-                    tools=[
-                        Tool(
-                            function=Function(
-                                name="get_weather",
-                                description="Get the current weather for a location",
-                                parameters=ValidatedParametersSchema(
-                                    schema_type="object",
-                                    properties={
-                                        "location": {
-                                            "type": "string",
-                                            "description": "The city name",
-                                        }
-                                    },
-                                    required=["location"],
-                                    additional_properties=False,
-                                ),
-                            )
-                        )
-                    ],
-                )
-            )
-        ]
-
-    parsed_tool_call_names = [
-        tool_call.name
-        for message in messages
-        if message.parsed_tool_calls is not None
-        for tool_call in message.parsed_tool_calls
-    ]
-
-    assert parsed_tool_call_names == ["get_weather"]
-    assert messages[-1].kind == InferenceMessageKind.DONE
diff --git a/paddler_client_python/tests/unit/__init__.py b/paddler_client_python/tests/unit/__init__.py
deleted file mode 100644
index e69de29bb..000000000
diff --git a/paddler_client_python/tests/unit/test_agent_controller_pool_snapshot.py b/paddler_client_python/tests/unit/test_agent_controller_pool_snapshot.py
deleted file mode 100644
index 3ef80d099..000000000
--- a/paddler_client_python/tests/unit/test_agent_controller_pool_snapshot.py
+++ /dev/null
@@ -1,26 +0,0 @@
-from paddler_client.agent_controller_pool_snapshot import (
-    AgentControllerPoolSnapshot,
-)
-
-
-def test_agent_controller_pool_snapshot_reads_every_agent() -> None:
-    pool = AgentControllerPoolSnapshot.model_validate(
-        {
-            "agents": [
-                {
-                    "id": "a1",
-                    "slots_processing": 0,
-                    "status": {
-                        "desired_slots_total": 2,
-                        "download_status": "NotDownloading",
-                        "slots_total": 2,
-                        "state_application_status": "Applied",
-                        "uses_chat_template_override": False,
-                    },
-                }
-            ]
-        }
-    )
-
-    assert len(pool.agents) == 1
-    assert pool.agents[0].id == "a1"
diff --git a/paddler_client_python/tests/unit/test_agent_controller_snapshot.py b/paddler_client_python/tests/unit/test_agent_controller_snapshot.py
deleted file mode 100644
index 61182cb85..000000000
--- a/paddler_client_python/tests/unit/test_agent_controller_snapshot.py
+++ /dev/null
@@ -1,27 +0,0 @@
-from paddler_client.agent_controller_snapshot import AgentControllerSnapshot
-from paddler_client.agent_state_application_status import (
-    AgentStateApplicationStatus,
-)
-
-
-def test_agent_controller_snapshot_deserialization() -> None:
-    snapshot = AgentControllerSnapshot.model_validate(
-        {
-            "id": "agent-1",
-            "name": "my-agent",
-            "slots_processing": 1,
-            "status": {
-                "desired_slots_total": 4,
-                "download_status": "NotDownloading",
-                "issues": [{"SlotCannotStart": {"error": "OOM", "slot_index": 0}}],
-                "model_path": "/models/test.gguf",
-                "slots_total": 4,
-                "state_application_status": "Fresh",
-                "uses_chat_template_override": True,
-            },
-        }
-    )
-
-    assert snapshot.id == "agent-1"
-    assert [issue.variant for issue in snapshot.status.issues] == ["SlotCannotStart"]
-    assert snapshot.status.state_application_status == AgentStateApplicationStatus.FRESH
diff --git a/paddler_client_python/tests/unit/test_agent_desired_model.py b/paddler_client_python/tests/unit/test_agent_desired_model.py
deleted file mode 100644
index 3dd67700d..000000000
--- a/paddler_client_python/tests/unit/test_agent_desired_model.py
+++ /dev/null
@@ -1,143 +0,0 @@
-import pytest
-from pydantic import ValidationError
-from pydantic_core import PydanticSerializationError
-
-from paddler_client.agent_desired_model import AgentDesiredModel
-from paddler_client.error import (
-    AgentDesiredModelLocalPathMissingError,
-    AgentDesiredModelUrlMissingError,
-    InvalidAgentDesiredModelError,
-    UnknownAgentDesiredModelVariantError,
-)
-from paddler_client.huggingface_model_reference import (
-    HuggingFaceModelReference,
-)
-from paddler_client.url_model_reference import UrlModelReference
-from tests.unit.validation_error_cause import validation_error_cause
-
-
-def test_agent_desired_model_none_serialization() -> None:
-    model = AgentDesiredModel.none()
-    serialized = model.model_dump(mode="json")
-
-    assert isinstance(serialized, str)
-    assert serialized == "None"
-
-
-def test_agent_desired_model_none_deserialization() -> None:
-    model = AgentDesiredModel.model_validate("None")
-
-    assert model.variant == "None"
-
-
-def test_agent_desired_model_huggingface_serialization() -> None:
-    reference = HuggingFaceModelReference(
-        filename="model.gguf",
-        repo_id="org/model",
-        revision="main",
-    )
-    model = AgentDesiredModel.from_huggingface(reference)
-    dumped = model.model_dump(mode="json")
-
-    assert "HuggingFace" in dumped
-    assert dumped["HuggingFace"]["repo_id"] == "org/model"
-
-
-def test_agent_desired_model_huggingface_deserialization() -> None:
-    model = AgentDesiredModel.model_validate(
-        {
-            "HuggingFace": {
-                "filename": "model.gguf",
-                "repo_id": "org/model",
-                "revision": "main",
-            }
-        }
-    )
-
-    assert model.variant == "HuggingFace"
-    assert model.huggingface is not None
-    assert model.huggingface.repo_id == "org/model"
-
-
-def test_agent_desired_model_local_to_agent_serialization() -> None:
-    model = AgentDesiredModel.local_to_agent("/path/to/model.gguf")
-    dumped = model.model_dump(mode="json")
-
-    assert dumped == {"LocalToAgent": "/path/to/model.gguf"}
-
-
-def test_agent_desired_model_local_to_agent_deserialization() -> None:
-    model = AgentDesiredModel.model_validate({"LocalToAgent": "/path/to/model.gguf"})
-
-    assert model.variant == "LocalToAgent"
-    assert model.local_path == "/path/to/model.gguf"
-
-
-def test_agent_desired_model_invalid_data_raises() -> None:
-    with pytest.raises(ValidationError) as rejection:
-        AgentDesiredModel.model_validate(42)
-
-    assert isinstance(
-        validation_error_cause(rejection.value), InvalidAgentDesiredModelError
-    )
-
-
-def test_agent_desired_model_unknown_variant_raises() -> None:
-    with pytest.raises(ValidationError) as rejection:
-        AgentDesiredModel.model_validate({"Unknown": "value"})
-
-    assert isinstance(
-        validation_error_cause(rejection.value), InvalidAgentDesiredModelError
-    )
-
-
-def test_agent_desired_model_serializes_a_local_path_to_the_wire_format() -> None:
-    model = AgentDesiredModel.local_to_agent("/path/to/model")
-    dumped = model.model_dump(mode="json")
-
-    assert dumped == {"LocalToAgent": "/path/to/model"}
-
-
-def test_agent_desired_model_unknown_variant_serialization_raises() -> None:
-    model = AgentDesiredModel(variant="Unknown")
-
-    with pytest.raises(PydanticSerializationError) as rejection:
-        model.model_dump(mode="json")
-
-    assert isinstance(rejection.value.__cause__, UnknownAgentDesiredModelVariantError)
-
-
-def test_agent_desired_model_local_to_agent_missing_path_raises() -> None:
-    model = AgentDesiredModel(variant="LocalToAgent", local_path=None)
-
-    with pytest.raises(PydanticSerializationError) as rejection:
-        model.model_dump(mode="json")
-
-    assert isinstance(rejection.value.__cause__, AgentDesiredModelLocalPathMissingError)
-
-
-def test_agent_desired_model_url_serialization() -> None:
-    reference = UrlModelReference(url="https://example.com/model.gguf")
-    model = AgentDesiredModel.from_url(reference)
-    dumped = model.model_dump(mode="json")
-
-    assert dumped == {"Url": {"url": "https://example.com/model.gguf"}}
-
-
-def test_agent_desired_model_url_deserialization() -> None:
-    model = AgentDesiredModel.model_validate(
-        {"Url": {"url": "https://example.com/model.gguf"}}
-    )
-
-    assert model.variant == "Url"
-    assert model.url is not None
-    assert model.url.url == "https://example.com/model.gguf"
-
-
-def test_agent_desired_model_url_missing_reference_raises() -> None:
-    model = AgentDesiredModel(variant="Url", url=None)
-
-    with pytest.raises(PydanticSerializationError) as rejection:
-        model.model_dump(mode="json")
-
-    assert isinstance(rejection.value.__cause__, AgentDesiredModelUrlMissingError)
diff --git a/paddler_client_python/tests/unit/test_agent_issue.py b/paddler_client_python/tests/unit/test_agent_issue.py
deleted file mode 100644
index b37bd160a..000000000
--- a/paddler_client_python/tests/unit/test_agent_issue.py
+++ /dev/null
@@ -1,60 +0,0 @@
-import pytest
-from pydantic import ValidationError
-
-from paddler_client.agent_issue import AgentIssue
-from paddler_client.error import InvalidAgentIssueError
-from tests.unit.validation_error_cause import validation_error_cause
-
-
-def test_agent_issue_deserialization() -> None:
-    issue = AgentIssue.model_validate(
-        {"ModelFileDoesNotExist": {"model_path": "/path/to/model"}}
-    )
-
-    assert issue.variant == "ModelFileDoesNotExist"
-    assert issue.params["model_path"] == "/path/to/model"
-
-
-def test_agent_issue_serialization() -> None:
-    issue = AgentIssue(
-        variant="ModelFileDoesNotExist",
-        params={"model_path": "/path/to/model"},
-    )
-    dumped = issue.model_dump(mode="json")
-
-    assert dumped == {"ModelFileDoesNotExist": {"model_path": "/path/to/model"}}
-
-
-def test_agent_issue_string_variant_deserialization() -> None:
-    issue = AgentIssue.model_validate("SimpleIssue")
-
-    assert issue.variant == "SimpleIssue"
-    assert issue.params == {}
-
-
-def test_agent_issue_non_dict_params_deserialization() -> None:
-    issue = AgentIssue.model_validate({"SomeIssue": "a string value"})
-
-    assert issue.variant == "SomeIssue"
-    assert issue.params == {"value": "a string value"}
-
-
-def test_agent_issue_invalid_data_raises() -> None:
-    with pytest.raises(ValidationError) as rejection:
-        AgentIssue.model_validate(42)
-
-    assert isinstance(validation_error_cause(rejection.value), InvalidAgentIssueError)
-
-
-def test_agent_issue_with_more_than_one_variant_raises() -> None:
-    with pytest.raises(ValidationError) as rejection:
-        AgentIssue.model_validate({"First": "value", "Second": "value"})
-
-    assert isinstance(validation_error_cause(rejection.value), InvalidAgentIssueError)
-
-
-def test_agent_issue_empty_params_serializes_as_string() -> None:
-    issue = AgentIssue(variant="SimpleIssue")
-    serialized = issue.model_dump_json()
-
-    assert serialized == '"SimpleIssue"'
diff --git a/paddler_client_python/tests/unit/test_agent_state_application_status.py b/paddler_client_python/tests/unit/test_agent_state_application_status.py
deleted file mode 100644
index 2821603a9..000000000
--- a/paddler_client_python/tests/unit/test_agent_state_application_status.py
+++ /dev/null
@@ -1,13 +0,0 @@
-from paddler_client.agent_state_application_status import (
-    AgentStateApplicationStatus,
-)
-
-
-def test_agent_state_application_status_uses_the_wire_names() -> None:
-    assert AgentStateApplicationStatus.APPLIED.value == "Applied"
-    assert (
-        AgentStateApplicationStatus.ATTEMPTED_AND_NOT_APPLIABLE.value
-        == "AttemptedAndNotAppliable"
-    )
-    assert AgentStateApplicationStatus.FRESH.value == "Fresh"
-    assert AgentStateApplicationStatus.STUCK.value == "Stuck"
diff --git a/paddler_client_python/tests/unit/test_balancer_desired_state.py b/paddler_client_python/tests/unit/test_balancer_desired_state.py
deleted file mode 100644
index db31f0cb5..000000000
--- a/paddler_client_python/tests/unit/test_balancer_desired_state.py
+++ /dev/null
@@ -1,23 +0,0 @@
-from paddler_client.balancer_desired_state import BalancerDesiredState
-from paddler_client.chat_template import ChatTemplate
-
-
-def test_balancer_desired_state_defaults_to_no_model_and_no_override() -> None:
-    state = BalancerDesiredState()
-    dumped = state.model_dump(mode="json")
-
-    assert dumped["model"] == "None"
-    assert dumped["multimodal_projection"] == "None"
-    assert dumped["use_chat_template_override"] is False
-    assert dumped["inference_parameters"]["temperature"] == 0.8
-
-
-def test_balancer_desired_state_with_chat_template() -> None:
-    state = BalancerDesiredState(
-        chat_template_override=ChatTemplate(content="{{ messages }}"),
-        use_chat_template_override=True,
-    )
-    dumped = state.model_dump(mode="json")
-
-    assert dumped["chat_template_override"] == {"content": "{{ messages }}"}
-    assert dumped["use_chat_template_override"] is True
diff --git a/paddler_client_python/tests/unit/test_buffered_request_manager_snapshot.py b/paddler_client_python/tests/unit/test_buffered_request_manager_snapshot.py
deleted file mode 100644
index eb3e5b814..000000000
--- a/paddler_client_python/tests/unit/test_buffered_request_manager_snapshot.py
+++ /dev/null
@@ -1,11 +0,0 @@
-from paddler_client.buffered_request_manager_snapshot import (
-    BufferedRequestManagerSnapshot,
-)
-
-
-def test_buffered_request_manager_snapshot_deserialization() -> None:
-    snapshot = BufferedRequestManagerSnapshot.model_validate(
-        {"buffered_requests_current": 12}
-    )
-
-    assert snapshot.buffered_requests_current == 12
diff --git a/paddler_client_python/tests/unit/test_chat_template.py b/paddler_client_python/tests/unit/test_chat_template.py
deleted file mode 100644
index 8223abf8e..000000000
--- a/paddler_client_python/tests/unit/test_chat_template.py
+++ /dev/null
@@ -1,8 +0,0 @@
-from paddler_client.chat_template import ChatTemplate
-
-
-def test_chat_template_serialization() -> None:
-    template = ChatTemplate(content="{% for msg in messages %}...")
-    dumped = template.model_dump(mode="json")
-
-    assert dumped == {"content": "{% for msg in messages %}..."}
diff --git a/paddler_client_python/tests/unit/test_continue_from_conversation_history_params.py b/paddler_client_python/tests/unit/test_continue_from_conversation_history_params.py
deleted file mode 100644
index 9f82b45d2..000000000
--- a/paddler_client_python/tests/unit/test_continue_from_conversation_history_params.py
+++ /dev/null
@@ -1,42 +0,0 @@
-from paddler_client.continue_from_conversation_history_params import (
-    ContinueFromConversationHistoryParams,
-)
-from paddler_client.conversation_message import ConversationMessage
-from paddler_client.grammar_constraint import JsonSchemaGrammarConstraint
-
-
-def test_serializes_the_conversation_history_request() -> None:
-    params = ContinueFromConversationHistoryParams(
-        add_generation_prompt=True,
-        conversation_history=[
-            ConversationMessage(content="Hello!", role="user"),
-        ],
-        enable_thinking=False,
-        max_tokens=100,
-    )
-    dumped = params.model_dump(mode="json")
-
-    assert dumped["add_generation_prompt"] is True
-    assert dumped["conversation_history"] == [{"content": "Hello!", "role": "user"}]
-    assert dumped["enable_thinking"] is False
-    assert dumped["grammar"] is None
-    assert dumped["max_tokens"] == 100
-    assert dumped["tools"] == []
-
-
-def test_serializes_a_json_schema_grammar_under_the_schema_key() -> None:
-    params = ContinueFromConversationHistoryParams(
-        add_generation_prompt=True,
-        conversation_history=[
-            ConversationMessage(content="hi", role="user"),
-        ],
-        enable_thinking=False,
-        grammar=JsonSchemaGrammarConstraint(
-            schema_value='{"type": "object"}',
-        ),
-        max_tokens=50,
-    )
-    dumped = params.model_dump(mode="json", by_alias=True)
-
-    assert dumped["grammar"]["type"] == "json_schema"
-    assert dumped["grammar"]["schema"] == '{"type": "object"}'
diff --git a/paddler_client_python/tests/unit/test_continue_from_conversation_history_params_rejects_tool_call_parsing_without_tools.py b/paddler_client_python/tests/unit/test_continue_from_conversation_history_params_rejects_tool_call_parsing_without_tools.py
deleted file mode 100644
index 2cc554001..000000000
--- a/paddler_client_python/tests/unit/test_continue_from_conversation_history_params_rejects_tool_call_parsing_without_tools.py
+++ /dev/null
@@ -1,27 +0,0 @@
-import pytest
-from pydantic import ValidationError
-
-from paddler_client.continue_from_conversation_history_params import (
-    ContinueFromConversationHistoryParams,
-)
-from paddler_client.conversation_message import ConversationMessage
-from paddler_client.error import ToolCallParsingWithoutToolsError
-from tests.unit.validation_error_cause import validation_error_cause
-
-
-def test_rejects_tool_call_parsing_without_tools() -> None:
-    with pytest.raises(ValidationError) as validation_error:
-        ContinueFromConversationHistoryParams(
-            add_generation_prompt=True,
-            conversation_history=[
-                ConversationMessage(content="Hello!", role="user"),
-            ],
-            enable_thinking=False,
-            max_tokens=100,
-            parse_tool_calls=True,
-        )
-
-    assert isinstance(
-        validation_error_cause(validation_error.value),
-        ToolCallParsingWithoutToolsError,
-    )
diff --git a/paddler_client_python/tests/unit/test_continue_from_raw_prompt_params.py b/paddler_client_python/tests/unit/test_continue_from_raw_prompt_params.py
deleted file mode 100644
index 63e12c6ee..000000000
--- a/paddler_client_python/tests/unit/test_continue_from_raw_prompt_params.py
+++ /dev/null
@@ -1,33 +0,0 @@
-from paddler_client.continue_from_raw_prompt_params import (
-    ContinueFromRawPromptParams,
-)
-from paddler_client.grammar_constraint import GbnfGrammarConstraint
-
-
-def test_serializes_the_raw_prompt_request() -> None:
-    params = ContinueFromRawPromptParams(
-        max_tokens=50,
-        raw_prompt="Once upon a time",
-    )
-    dumped = params.model_dump(mode="json")
-
-    assert dumped == {
-        "grammar": None,
-        "max_tokens": 50,
-        "raw_prompt": "Once upon a time",
-    }
-
-
-def test_serializes_a_gbnf_grammar_with_its_type_tag() -> None:
-    params = ContinueFromRawPromptParams(
-        grammar=GbnfGrammarConstraint(
-            grammar='root ::= "yes" | "no"',
-            root="root",
-        ),
-        max_tokens=10,
-        raw_prompt="Answer yes or no",
-    )
-    dumped = params.model_dump(mode="json")
-
-    assert dumped["grammar"]["type"] == "gbnf"
-    assert dumped["grammar"]["grammar"] == 'root ::= "yes" | "no"'
diff --git a/paddler_client_python/tests/unit/test_conversation_message.py b/paddler_client_python/tests/unit/test_conversation_message.py
deleted file mode 100644
index 55e2e96a9..000000000
--- a/paddler_client_python/tests/unit/test_conversation_message.py
+++ /dev/null
@@ -1,39 +0,0 @@
-from paddler_client.conversation_message import ConversationMessage
-from paddler_client.conversation_message_content_part import (
-    ImageUrlContentPart,
-    TextContentPart,
-)
-from paddler_client.image_url import ImageUrl
-
-
-def test_conversation_message_text_content() -> None:
-    message = ConversationMessage(content="hello", role="user")
-    dumped = message.model_dump(mode="json")
-
-    assert dumped == {"content": "hello", "role": "user"}
-
-
-def test_conversation_message_parts_content() -> None:
-    message = ConversationMessage(
-        content=[
-            TextContentPart(text="hello"),
-            ImageUrlContentPart(
-                image_url=ImageUrl(url="http://example.com/img.png"),
-            ),
-        ],
-        role="user",
-    )
-    dumped = message.model_dump(mode="json")
-
-    assert dumped["content"][0] == {"type": "text", "text": "hello"}
-    assert dumped["content"][1] == {
-        "type": "image_url",
-        "image_url": {"url": "http://example.com/img.png"},
-    }
-
-
-def test_conversation_message_text_deserialization() -> None:
-    message = ConversationMessage.model_validate({"content": "hi", "role": "assistant"})
-
-    assert message.content == "hi"
-    assert message.role == "assistant"
diff --git a/paddler_client_python/tests/unit/test_embedding.py b/paddler_client_python/tests/unit/test_embedding.py
deleted file mode 100644
index 70210d9ef..000000000
--- a/paddler_client_python/tests/unit/test_embedding.py
+++ /dev/null
@@ -1,32 +0,0 @@
-from paddler_client.embedding import Embedding
-from paddler_client.pooling_type import PoolingType
-
-
-def test_embedding_deserialization() -> None:
-    embedding = Embedding.model_validate(
-        {
-            "embedding": [0.1, 0.2, 0.3],
-            "normalization_method": "L2",
-            "pooling_type": "Mean",
-            "source_document_id": "doc-1",
-        }
-    )
-
-    assert embedding.embedding == [0.1, 0.2, 0.3]
-    assert embedding.normalization_method.variant == "L2"
-    assert embedding.pooling_type == PoolingType.MEAN
-    assert embedding.source_document_id == "doc-1"
-
-
-def test_embedding_with_rms_norm_deserialization() -> None:
-    embedding = Embedding.model_validate(
-        {
-            "embedding": [1.0],
-            "normalization_method": {"RmsNorm": {"epsilon": 1e-6}},
-            "pooling_type": "Cls",
-            "source_document_id": "doc-2",
-        }
-    )
-
-    assert embedding.normalization_method.variant == "RmsNorm"
-    assert embedding.normalization_method.epsilon == 1e-6
diff --git a/paddler_client_python/tests/unit/test_embedding_input_document.py b/paddler_client_python/tests/unit/test_embedding_input_document.py
deleted file mode 100644
index be7e7ce7d..000000000
--- a/paddler_client_python/tests/unit/test_embedding_input_document.py
+++ /dev/null
@@ -1,8 +0,0 @@
-from paddler_client.embedding_input_document import EmbeddingInputDocument
-
-
-def test_embedding_input_document_serialization() -> None:
-    doc = EmbeddingInputDocument(content="hello world", id="d1")
-    dumped = doc.model_dump(mode="json")
-
-    assert dumped == {"content": "hello world", "id": "d1"}
diff --git a/paddler_client_python/tests/unit/test_embedding_normalization_method.py b/paddler_client_python/tests/unit/test_embedding_normalization_method.py
deleted file mode 100644
index 503611621..000000000
--- a/paddler_client_python/tests/unit/test_embedding_normalization_method.py
+++ /dev/null
@@ -1,95 +0,0 @@
-import pytest
-from pydantic import ValidationError
-from pydantic_core import PydanticSerializationError
-
-from paddler_client.embedding_normalization_method import (
-    EmbeddingNormalizationMethod,
-)
-from paddler_client.error import (
-    InvalidEmbeddingNormalizationMethodError,
-    InvalidRmsNormPayloadError,
-    RmsNormEpsilonMissingError,
-)
-from tests.unit.validation_error_cause import validation_error_cause
-
-
-def test_embedding_normalization_method_l2_serialization() -> None:
-    method = EmbeddingNormalizationMethod.l2()
-    serialized = method.model_dump(mode="json")
-
-    assert isinstance(serialized, str)
-    assert serialized == "L2"
-
-
-def test_embedding_normalization_method_none_serialization() -> None:
-    method = EmbeddingNormalizationMethod.none()
-    serialized = method.model_dump(mode="json")
-
-    assert isinstance(serialized, str)
-    assert serialized == "None"
-
-
-def test_embedding_normalization_method_rms_norm_serialization() -> None:
-    method = EmbeddingNormalizationMethod.rms_norm(epsilon=0.001)
-
-    assert method.model_dump(mode="json") == {"RmsNorm": {"epsilon": 0.001}}
-
-
-def test_embedding_normalization_method_l2_deserialization() -> None:
-    method = EmbeddingNormalizationMethod.model_validate("L2")
-
-    assert method.variant == "L2"
-
-
-def test_embedding_normalization_method_rms_norm_deserialization() -> None:
-    method = EmbeddingNormalizationMethod.model_validate(
-        {"RmsNorm": {"epsilon": 0.001}}
-    )
-
-    assert method.variant == "RmsNorm"
-    assert method.epsilon == 0.001
-
-
-def test_embedding_normalization_method_rms_norm_missing_epsilon_raises() -> None:
-    method = EmbeddingNormalizationMethod(variant="RmsNorm", epsilon=None)
-
-    with pytest.raises(PydanticSerializationError) as rejection:
-        method.model_dump(mode="json")
-
-    assert isinstance(rejection.value.__cause__, RmsNormEpsilonMissingError)
-
-
-def test_embedding_normalization_method_invalid_rms_norm_raises() -> None:
-    with pytest.raises(ValidationError) as rejection:
-        EmbeddingNormalizationMethod.model_validate({"RmsNorm": "not a dict"})
-
-    assert isinstance(
-        validation_error_cause(rejection.value), InvalidRmsNormPayloadError
-    )
-
-
-def test_embedding_normalization_method_invalid_data_raises() -> None:
-    with pytest.raises(ValidationError) as rejection:
-        EmbeddingNormalizationMethod.model_validate(42)
-
-    assert isinstance(
-        validation_error_cause(rejection.value),
-        InvalidEmbeddingNormalizationMethodError,
-    )
-
-
-def test_embedding_normalization_method_unknown_variant_raises() -> None:
-    with pytest.raises(ValidationError) as rejection:
-        EmbeddingNormalizationMethod.model_validate({"Unknown": {}})
-
-    assert isinstance(
-        validation_error_cause(rejection.value),
-        InvalidEmbeddingNormalizationMethodError,
-    )
-
-
-def test_rms_norm_serializes_to_the_wire_format() -> None:
-    method = EmbeddingNormalizationMethod.rms_norm(epsilon=1e-6)
-    dumped = method.model_dump(mode="json")
-
-    assert dumped == {"RmsNorm": {"epsilon": 1e-6}}
diff --git a/paddler_client_python/tests/unit/test_format_api_url.py b/paddler_client_python/tests/unit/test_format_api_url.py
deleted file mode 100644
index a25435601..000000000
--- a/paddler_client_python/tests/unit/test_format_api_url.py
+++ /dev/null
@@ -1,27 +0,0 @@
-import pytest
-
-from paddler_client.error import ApiPathWithoutLeadingSlashError
-from paddler_client.format_api_url import format_api_url
-
-
-def test_joins_the_base_url_and_the_path() -> None:
-    result = format_api_url("http://localhost:8080", "/health")
-
-    assert result == "http://localhost:8080/health"
-
-
-def test_strips_trailing_slash() -> None:
-    result = format_api_url("http://localhost:8080/", "/health")
-
-    assert result == "http://localhost:8080/health"
-
-
-def test_path_must_start_with_slash() -> None:
-    with pytest.raises(ApiPathWithoutLeadingSlashError):
-        format_api_url("http://localhost:8080", "health")
-
-
-def test_multiple_trailing_slashes() -> None:
-    result = format_api_url("http://localhost:8080///", "/health")
-
-    assert result == "http://localhost:8080/health"
diff --git a/paddler_client_python/tests/unit/test_generate_embedding_batch_params.py b/paddler_client_python/tests/unit/test_generate_embedding_batch_params.py
deleted file mode 100644
index 050eed82c..000000000
--- a/paddler_client_python/tests/unit/test_generate_embedding_batch_params.py
+++ /dev/null
@@ -1,21 +0,0 @@
-from paddler_client.embedding_input_document import EmbeddingInputDocument
-from paddler_client.embedding_normalization_method import (
-    EmbeddingNormalizationMethod,
-)
-from paddler_client.generate_embedding_batch_params import (
-    GenerateEmbeddingBatchParams,
-)
-
-
-def test_serializes_the_embedding_batch_request() -> None:
-    params = GenerateEmbeddingBatchParams(
-        input_batch=[
-            EmbeddingInputDocument(content="hello world", id="doc-1"),
-        ],
-        normalization_method=EmbeddingNormalizationMethod.l2(),
-    )
-    dumped = params.model_dump(mode="json")
-
-    assert len(dumped["input_batch"]) == 1
-    assert dumped["input_batch"][0]["id"] == "doc-1"
-    assert dumped["normalization_method"] == "L2"
diff --git a/paddler_client_python/tests/unit/test_grammar_constraint.py b/paddler_client_python/tests/unit/test_grammar_constraint.py
deleted file mode 100644
index 736bec94a..000000000
--- a/paddler_client_python/tests/unit/test_grammar_constraint.py
+++ /dev/null
@@ -1,30 +0,0 @@
-from paddler_client.grammar_constraint import (
-    GbnfGrammarConstraint,
-    JsonSchemaGrammarConstraint,
-)
-
-
-def test_gbnf_grammar_constraint_serialization() -> None:
-    constraint = GbnfGrammarConstraint(
-        grammar='root ::= "yes" | "no"',
-        root="root",
-    )
-    dumped = constraint.model_dump(mode="json")
-
-    assert dumped == {
-        "type": "gbnf",
-        "grammar": 'root ::= "yes" | "no"',
-        "root": "root",
-    }
-
-
-def test_json_schema_grammar_constraint_serialization() -> None:
-    constraint = JsonSchemaGrammarConstraint(
-        schema_value='{"type": "object"}',
-    )
-    dumped = constraint.model_dump(mode="json", by_alias=True)
-
-    assert dumped == {
-        "type": "json_schema",
-        "schema": '{"type": "object"}',
-    }
diff --git a/paddler_client_python/tests/unit/test_inference_message.py b/paddler_client_python/tests/unit/test_inference_message.py
deleted file mode 100644
index 609f1685a..000000000
--- a/paddler_client_python/tests/unit/test_inference_message.py
+++ /dev/null
@@ -1,750 +0,0 @@
-import pytest
-
-from paddler_client.error import (
-    GeneratedTokenResultNotAnObjectError,
-    InferenceClientMessageNotAnObjectError,
-    MediaExceedsMicroBatchPayloadNotAnObjectError,
-    PromptExceedsContextSizePayloadNotAnObjectError,
-    ToolCallParsedPayloadNotAListError,
-    ToolCallValidationFailedPayloadNotAListError,
-    UnknownEmbeddingResultError,
-    UnknownGeneratedTokenResultError,
-    UnknownInferenceClientMessageError,
-    UnknownResponseVariantError,
-    UnrecognizedToolCallFormatPayloadNotAnObjectError,
-)
-from paddler_client.generation_finish import GenerationFinish
-from paddler_client.inference_message import (
-    InferenceMessageKind,
-    parse_inference_client_message,
-)
-
-
-def test_parse_content_token_response() -> None:
-    data = {
-        "Response": {
-            "generated_by": None,
-            "request_id": "req-1",
-            "response": {"GeneratedToken": {"ContentToken": "hello"}},
-        }
-    }
-    message = parse_inference_client_message(data)
-
-    assert message.request_id == "req-1"
-    assert message.kind == InferenceMessageKind.CONTENT_TOKEN
-    assert message.token == "hello"
-    assert message.is_token
-    assert not message.is_terminal
-
-
-def test_parse_reasoning_token_response() -> None:
-    data = {
-        "Response": {
-            "generated_by": None,
-            "request_id": "req-1",
-            "response": {"GeneratedToken": {"ReasoningToken": "thinking"}},
-        }
-    }
-    message = parse_inference_client_message(data)
-
-    assert message.kind == InferenceMessageKind.REASONING_TOKEN
-    assert message.token == "thinking"
-    assert message.is_token
-    assert not message.is_terminal
-
-
-def test_parse_tool_call_token_response() -> None:
-    data = {
-        "Response": {
-            "generated_by": None,
-            "request_id": "req-1",
-            "response": {"GeneratedToken": {"ToolCallToken": '{"name":'}},
-        }
-    }
-    message = parse_inference_client_message(data)
-
-    assert message.kind == InferenceMessageKind.TOOL_CALL_TOKEN
-    assert message.token == '{"name":'
-    assert message.is_token
-    assert not message.is_terminal
-
-
-def test_parse_tool_call_parsed_response_carries_structured_calls() -> None:
-    data = {
-        "Response": {
-            "generated_by": None,
-            "request_id": "req-1",
-            "response": {
-                "GeneratedToken": {
-                    "ToolCallParsed": [
-                        {
-                            "id": "call_42",
-                            "name": "get_weather",
-                            "arguments": {"ValidJson": {"location": "Paris"}},
-                        },
-                    ],
-                },
-            },
-        },
-    }
-    message = parse_inference_client_message(data)
-
-    assert message.kind == InferenceMessageKind.TOOL_CALL_PARSED
-    assert message.parsed_tool_calls is not None
-    assert len(message.parsed_tool_calls) == 1
-    assert message.parsed_tool_calls[0].id == "call_42"
-    assert message.parsed_tool_calls[0].name == "get_weather"
-    assert not message.is_token
-    assert not message.is_terminal
-
-
-def test_parse_tool_call_parse_failed_response_carries_error() -> None:
-    data = {
-        "Response": {
-            "generated_by": None,
-            "request_id": "req-1",
-            "response": {
-                "GeneratedToken": {"ToolCallParseFailed": "syntax error at 12"},
-            },
-        },
-    }
-    message = parse_inference_client_message(data)
-
-    assert message.kind == InferenceMessageKind.TOOL_CALL_PARSE_FAILED
-    assert message.error_message == "syntax error at 12"
-    assert not message.is_terminal
-
-
-def test_parse_tool_call_validation_failed_response_joins_errors() -> None:
-    data = {
-        "Response": {
-            "generated_by": None,
-            "request_id": "req-1",
-            "response": {
-                "GeneratedToken": {
-                    "ToolCallValidationFailed": [
-                        "missing field 'location'",
-                        "extra field 'foo'",
-                    ],
-                },
-            },
-        },
-    }
-    message = parse_inference_client_message(data)
-
-    assert message.kind == InferenceMessageKind.TOOL_CALL_VALIDATION_FAILED
-    assert message.error_message == "missing field 'location'; extra field 'foo'"
-    assert not message.is_terminal
-
-
-def test_parse_tool_call_parsed_with_non_list_payload_raises() -> None:
-    data = {
-        "Response": {
-            "generated_by": None,
-            "request_id": "req-1",
-            "response": {"GeneratedToken": {"ToolCallParsed": "not a list"}},
-        },
-    }
-
-    with pytest.raises(ToolCallParsedPayloadNotAListError):
-        parse_inference_client_message(data)
-
-
-def test_parse_tool_call_validation_failed_with_non_list_payload_raises() -> None:
-    data = {
-        "Response": {
-            "generated_by": None,
-            "request_id": "req-1",
-            "response": {"GeneratedToken": {"ToolCallValidationFailed": "oops"}},
-        },
-    }
-
-    with pytest.raises(ToolCallValidationFailedPayloadNotAListError):
-        parse_inference_client_message(data)
-
-
-def test_parse_unrecognized_tool_call_format_response_carries_text_and_ffi_error() -> (
-    None
-):
-    data = {
-        "Response": {
-            "generated_by": None,
-            "request_id": "req-1",
-            "response": {
-                "GeneratedToken": {
-                    "UnrecognizedToolCallFormat": {
-                        "text": "<unknown_marker>blah</unknown_marker>",
-                        "ffi_error_message": "common_chat_parse failed: no parser",
-                    },
-                },
-            },
-        },
-    }
-    message = parse_inference_client_message(data)
-
-    assert message.kind == InferenceMessageKind.UNRECOGNIZED_TOOL_CALL_FORMAT
-    assert message.raw_tool_call_tokens is not None
-    assert message.raw_tool_call_tokens.text == "<unknown_marker>blah</unknown_marker>"
-    assert (
-        message.raw_tool_call_tokens.ffi_error_message
-        == "common_chat_parse failed: no parser"
-    )
-    assert not message.is_token
-    assert not message.is_terminal
-
-
-def test_parse_unrecognized_tool_call_format_with_non_dict_payload_raises() -> None:
-    data = {
-        "Response": {
-            "generated_by": None,
-            "request_id": "req-1",
-            "response": {
-                "GeneratedToken": {"UnrecognizedToolCallFormat": "raw text only"},
-            },
-        },
-    }
-
-    with pytest.raises(UnrecognizedToolCallFormatPayloadNotAnObjectError):
-        parse_inference_client_message(data)
-
-
-def test_parse_media_exceeds_micro_batch_response_carries_token_counts() -> None:
-    data = {
-        "Response": {
-            "generated_by": None,
-            "request_id": "req-1",
-            "response": {
-                "GeneratedToken": {
-                    "MediaExceedsMicroBatch": {
-                        "media_tokens": 368,
-                        "micro_batch_tokens": 100,
-                    },
-                },
-            },
-        },
-    }
-    message = parse_inference_client_message(data)
-
-    assert message.kind == InferenceMessageKind.MEDIA_EXCEEDS_MICRO_BATCH
-    assert message.oversized_media_details is not None
-    assert message.oversized_media_details.media_tokens == 368
-    assert message.oversized_media_details.micro_batch_tokens == 100
-    assert not message.is_token
-
-
-def test_parse_media_exceeds_micro_batch_with_non_dict_payload_raises() -> None:
-    data = {
-        "Response": {
-            "generated_by": None,
-            "request_id": "req-1",
-            "response": {
-                "GeneratedToken": {"MediaExceedsMicroBatch": "scalar payload"},
-            },
-        },
-    }
-
-    with pytest.raises(MediaExceedsMicroBatchPayloadNotAnObjectError):
-        parse_inference_client_message(data)
-
-
-def test_parse_prompt_exceeds_context_size_response_carries_token_counts() -> None:
-    data = {
-        "Response": {
-            "generated_by": None,
-            "request_id": "req-1",
-            "response": {
-                "GeneratedToken": {
-                    "PromptExceedsContextSize": {
-                        "prompt_tokens": 9895,
-                        "sequence_context_size": 8192,
-                    },
-                },
-            },
-        },
-    }
-    message = parse_inference_client_message(data)
-
-    assert message.kind == InferenceMessageKind.PROMPT_EXCEEDS_CONTEXT_SIZE
-    assert message.oversized_prompt_details is not None
-    assert message.oversized_prompt_details.prompt_tokens == 9895
-    assert message.oversized_prompt_details.sequence_context_size == 8192
-    assert not message.is_token
-
-
-def test_parse_prompt_exceeds_context_size_with_non_dict_payload_raises() -> None:
-    data = {
-        "Response": {
-            "generated_by": None,
-            "request_id": "req-1",
-            "response": {
-                "GeneratedToken": {"PromptExceedsContextSize": "scalar payload"},
-            },
-        },
-    }
-
-    with pytest.raises(PromptExceedsContextSizePayloadNotAnObjectError):
-        parse_inference_client_message(data)
-
-
-def test_parse_undeterminable_token_response() -> None:
-    data = {
-        "Response": {
-            "generated_by": None,
-            "request_id": "req-1",
-            "response": {"GeneratedToken": {"UndeterminableToken": "raw"}},
-        }
-    }
-    message = parse_inference_client_message(data)
-
-    assert message.kind == InferenceMessageKind.UNDETERMINABLE_TOKEN
-    assert message.token == "raw"
-    assert message.is_token
-
-
-def test_parse_done_response_carries_summary() -> None:
-    data = {
-        "Response": {
-            "generated_by": None,
-            "request_id": "req-1",
-            "response": {
-                "GeneratedToken": {
-                    "Done": {
-                        "finish": "MaxTokens",
-                        "usage": {
-                            "prompt_tokens": 4,
-                            "cached_prompt_tokens": 0,
-                            "input_image_tokens": 0,
-                            "input_audio_tokens": 0,
-                            "content_tokens": 6,
-                            "reasoning_tokens": 1,
-                            "tool_call_tokens": 0,
-                            "undeterminable_tokens": 0,
-                        },
-                    }
-                }
-            },
-        }
-    }
-    message = parse_inference_client_message(data)
-
-    assert message.kind == InferenceMessageKind.DONE
-    assert message.is_done
-    assert message.is_terminal
-    assert message.summary is not None
-    assert message.summary.finish == GenerationFinish.MAX_TOKENS
-    assert message.summary.usage.prompt_tokens == 4
-    assert message.summary.usage.content_tokens == 6
-    assert message.summary.usage.reasoning_tokens == 1
-    assert message.summary.usage.completion_tokens == 7
-    assert message.summary.usage.total_tokens == 11
-
-
-def test_parse_server_error() -> None:
-    data = {
-        "Error": {
-            "request_id": "req-1",
-            "error": {"code": 500, "description": "Internal error"},
-        }
-    }
-    message = parse_inference_client_message(data)
-
-    assert message.kind == InferenceMessageKind.SERVER_ERROR
-    assert message.error_code == 500
-    assert message.error_message == "Internal error"
-    assert message.is_terminal
-
-
-def test_parse_chat_template_error() -> None:
-    data = {
-        "Response": {
-            "generated_by": None,
-            "request_id": "req-1",
-            "response": {"GeneratedToken": {"ChatTemplateError": "bad template"}},
-        }
-    }
-    message = parse_inference_client_message(data)
-
-    assert message.kind == InferenceMessageKind.CHAT_TEMPLATE_ERROR
-    assert message.error_message == "bad template"
-    assert message.is_terminal
-
-
-def test_parse_grammar_incompatible_with_thinking() -> None:
-    data = {
-        "Response": {
-            "generated_by": None,
-            "request_id": "req-1",
-            "response": {"GeneratedToken": {"GrammarIncompatibleWithThinking": "err"}},
-        }
-    }
-    message = parse_inference_client_message(data)
-
-    assert message.kind == InferenceMessageKind.GRAMMAR_INCOMPATIBLE_WITH_THINKING
-    assert message.error_message == "err"
-    assert message.is_terminal
-
-
-def test_parse_grammar_initialization_failed() -> None:
-    data = {
-        "Response": {
-            "generated_by": None,
-            "request_id": "req-1",
-            "response": {
-                "GeneratedToken": {"GrammarInitializationFailed": "null grammar"}
-            },
-        }
-    }
-    message = parse_inference_client_message(data)
-
-    assert message.kind == InferenceMessageKind.GRAMMAR_INITIALIZATION_FAILED
-    assert message.error_message == "null grammar"
-    assert message.is_terminal
-
-
-def test_parse_grammar_rejected_model_output() -> None:
-    data = {
-        "Response": {
-            "generated_by": None,
-            "request_id": "req-1",
-            "response": {
-                "GeneratedToken": {"GrammarRejectedModelOutput": "token rejected"}
-            },
-        }
-    }
-    message = parse_inference_client_message(data)
-
-    assert message.kind == InferenceMessageKind.GRAMMAR_REJECTED_MODEL_OUTPUT
-    assert message.error_message == "token rejected"
-    assert message.is_terminal
-
-
-def test_parse_grammar_syntax_error() -> None:
-    data = {
-        "Response": {
-            "generated_by": None,
-            "request_id": "req-1",
-            "response": {"GeneratedToken": {"GrammarSyntaxError": "invalid schema"}},
-        }
-    }
-    message = parse_inference_client_message(data)
-
-    assert message.kind == InferenceMessageKind.GRAMMAR_SYNTAX_ERROR
-    assert message.error_message == "invalid schema"
-    assert message.is_terminal
-
-
-def test_parse_tool_schema_invalid_response_carries_error() -> None:
-    data = {
-        "Response": {
-            "generated_by": None,
-            "request_id": "req-1",
-            "response": {
-                "GeneratedToken": {
-                    "ToolSchemaInvalid": (
-                        'tool "get_weather" parameters are not a valid JSON Schema'
-                    ),
-                },
-            },
-        },
-    }
-    message = parse_inference_client_message(data)
-
-    assert message.kind == InferenceMessageKind.TOOL_SCHEMA_INVALID
-    assert message.error_message == (
-        'tool "get_weather" parameters are not a valid JSON Schema'
-    )
-    assert message.is_terminal
-
-
-def test_parse_image_decoding_failed() -> None:
-    data = {
-        "Response": {
-            "generated_by": None,
-            "request_id": "req-1",
-            "response": {"GeneratedToken": {"ImageDecodingFailed": "corrupt image"}},
-        }
-    }
-    message = parse_inference_client_message(data)
-
-    assert message.kind == InferenceMessageKind.IMAGE_DECODING_FAILED
-    assert message.error_message == "corrupt image"
-    assert message.is_terminal
-
-
-def test_parse_multimodal_not_supported() -> None:
-    data = {
-        "Response": {
-            "generated_by": None,
-            "request_id": "req-1",
-            "response": {"GeneratedToken": {"MultimodalNotSupported": "no multimodal"}},
-        }
-    }
-    message = parse_inference_client_message(data)
-
-    assert message.kind == InferenceMessageKind.MULTIMODAL_NOT_SUPPORTED
-    assert message.error_message == "no multimodal"
-    assert message.is_terminal
-
-
-def test_parse_model_not_loaded() -> None:
-    data = {
-        "Response": {
-            "generated_by": None,
-            "request_id": "req-1",
-            "response": {"GeneratedToken": {"ModelNotLoaded": "no model is loaded"}},
-        }
-    }
-    message = parse_inference_client_message(data)
-
-    assert message.kind == InferenceMessageKind.MODEL_NOT_LOADED
-    assert message.error_message == "no model is loaded"
-    assert message.is_terminal
-
-
-def test_parse_sampler_error() -> None:
-    data = {
-        "Response": {
-            "generated_by": None,
-            "request_id": "req-1",
-            "response": {"GeneratedToken": {"SamplerError": "no candidates"}},
-        }
-    }
-    message = parse_inference_client_message(data)
-
-    assert message.kind == InferenceMessageKind.SAMPLER_ERROR
-    assert message.error_message == "no candidates"
-    assert message.is_terminal
-
-
-def test_parse_embedding_response() -> None:
-    data = {
-        "Response": {
-            "generated_by": None,
-            "request_id": "req-1",
-            "response": {
-                "Embedding": {
-                    "Embedding": {
-                        "embedding": [1.0, 2.0, 3.0],
-                        "normalization_method": "None",
-                        "pooling_type": "Mean",
-                        "source_document_id": "doc-1",
-                    }
-                }
-            },
-        }
-    }
-    message = parse_inference_client_message(data)
-
-    assert message.kind == InferenceMessageKind.EMBEDDING
-    assert message.embedding_data is not None
-    assert message.embedding_data.embedding == [1.0, 2.0, 3.0]
-    assert message.embedding_data.source_document_id == "doc-1"
-    assert not message.is_terminal
-
-
-def test_parse_embedding_done() -> None:
-    data = {
-        "Response": {
-            "generated_by": None,
-            "request_id": "req-1",
-            "response": {"Embedding": "Done"},
-        }
-    }
-    message = parse_inference_client_message(data)
-
-    assert message.kind == InferenceMessageKind.EMBEDDING_DONE
-    assert message.is_terminal
-
-
-def test_parse_embedding_error() -> None:
-    data = {
-        "Response": {
-            "generated_by": None,
-            "request_id": "req-1",
-            "response": {"Embedding": {"Error": "embedding failed"}},
-        }
-    }
-    message = parse_inference_client_message(data)
-
-    assert message.kind == InferenceMessageKind.EMBEDDING_ERROR
-    assert message.error_message == "embedding failed"
-    assert message.is_terminal
-
-
-def test_parse_embedding_model_not_loaded() -> None:
-    data = {
-        "Response": {
-            "generated_by": None,
-            "request_id": "req-1",
-            "response": {"Embedding": {"ModelNotLoaded": "no model is loaded"}},
-        }
-    }
-    message = parse_inference_client_message(data)
-
-    assert message.kind == InferenceMessageKind.MODEL_NOT_LOADED
-    assert message.error_message == "no model is loaded"
-    assert message.is_terminal
-
-
-def test_parse_detokenization_failed_response_carries_error() -> None:
-    data = {
-        "Response": {
-            "generated_by": None,
-            "request_id": "req-1",
-            "response": {
-                "GeneratedToken": {"DetokenizationFailed": "invalid UTF-8 piece"},
-            },
-        }
-    }
-    message = parse_inference_client_message(data)
-
-    assert message.kind == InferenceMessageKind.DETOKENIZATION_FAILED
-    assert message.error_message == "invalid UTF-8 piece"
-    assert message.is_terminal
-
-
-def test_parse_embeddings_disabled() -> None:
-    data = {
-        "Response": {
-            "generated_by": None,
-            "request_id": "req-1",
-            "response": {"Embedding": "EmbeddingsDisabled"},
-        }
-    }
-    message = parse_inference_client_message(data)
-
-    assert message.kind == InferenceMessageKind.EMBEDDINGS_DISABLED
-    assert message.is_terminal
-
-
-def test_parse_embedding_no_embeddings_produced() -> None:
-    data = {
-        "Response": {
-            "generated_by": None,
-            "request_id": "req-1",
-            "response": {"Embedding": "NoEmbeddingsProduced"},
-        }
-    }
-    message = parse_inference_client_message(data)
-
-    assert message.kind == InferenceMessageKind.EMBEDDING_NO_EMBEDDINGS_PRODUCED
-    assert message.is_terminal
-
-
-def test_parse_embedding_rejected_due_to_active_token_generation() -> None:
-    data = {
-        "Response": {
-            "generated_by": None,
-            "request_id": "req-1",
-            "response": {"Embedding": "EmbeddingRejectedDueToActiveTokenGeneration"},
-        }
-    }
-    message = parse_inference_client_message(data)
-
-    assert (
-        message.kind
-        == InferenceMessageKind.EMBEDDING_REJECTED_DUE_TO_ACTIVE_TOKEN_GENERATION
-    )
-    assert message.is_terminal
-
-
-def test_parses_a_message_given_as_a_json_string() -> None:
-    serialized_message = (
-        '{"Response": {"generated_by": null, "request_id": "req-1", '
-        '"response": {"GeneratedToken": '
-        '{"Done": {"finish": "EndOfGeneration", "usage": {"prompt_tokens": 0, '
-        '"cached_prompt_tokens": 0, '
-        '"input_image_tokens": 0, "input_audio_tokens": 0, "content_tokens": 0, '
-        '"reasoning_tokens": 0, "tool_call_tokens": 0, "undeterminable_tokens": 0}}}}}}'
-    )
-    message = parse_inference_client_message(serialized_message)
-
-    assert message.kind == InferenceMessageKind.DONE
-
-
-def test_parse_unknown_format_raises() -> None:
-    with pytest.raises(UnknownInferenceClientMessageError):
-        parse_inference_client_message({"Unknown": {}})
-
-
-def test_parse_json_that_is_not_an_object_raises_type_error() -> None:
-    with pytest.raises(InferenceClientMessageNotAnObjectError):
-        parse_inference_client_message("42")
-
-
-def test_parse_unknown_response_variant_raises() -> None:
-    data = {
-        "Response": {
-            "generated_by": None,
-            "request_id": "req-1",
-            "response": "SomethingUnexpected",
-        }
-    }
-
-    with pytest.raises(UnknownResponseVariantError):
-        parse_inference_client_message(data)
-
-
-def test_parse_unknown_response_dict_raises() -> None:
-    data = {
-        "Response": {
-            "generated_by": None,
-            "request_id": "req-1",
-            "response": {"UnknownKey": "value"},
-        }
-    }
-
-    with pytest.raises(UnknownResponseVariantError):
-        parse_inference_client_message(data)
-
-
-def test_parse_unknown_generated_token_result_raises() -> None:
-    data = {
-        "Response": {
-            "generated_by": None,
-            "request_id": "req-1",
-            "response": {"GeneratedToken": {"UnknownVariant": "data"}},
-        }
-    }
-
-    with pytest.raises(UnknownGeneratedTokenResultError):
-        parse_inference_client_message(data)
-
-
-def test_parse_string_generated_token_result_raises() -> None:
-    data = {
-        "Response": {
-            "generated_by": None,
-            "request_id": "req-1",
-            "response": {"GeneratedToken": "Done"},
-        }
-    }
-
-    with pytest.raises(GeneratedTokenResultNotAnObjectError):
-        parse_inference_client_message(data)
-
-
-def test_parse_unknown_embedding_result_raises() -> None:
-    data = {
-        "Response": {
-            "generated_by": None,
-            "request_id": "req-1",
-            "response": {"Embedding": {"UnknownVariant": "data"}},
-        }
-    }
-
-    with pytest.raises(UnknownEmbeddingResultError):
-        parse_inference_client_message(data)
-
-
-def test_parse_unknown_unit_embedding_result_raises() -> None:
-    data = {
-        "Response": {
-            "generated_by": None,
-            "request_id": "req-1",
-            "response": {"Embedding": "UnknownVariant"},
-        }
-    }
-
-    with pytest.raises(UnknownEmbeddingResultError):
-        parse_inference_client_message(data)
diff --git a/paddler_client_python/tests/unit/test_inference_parameters.py b/paddler_client_python/tests/unit/test_inference_parameters.py
deleted file mode 100644
index f07291398..000000000
--- a/paddler_client_python/tests/unit/test_inference_parameters.py
+++ /dev/null
@@ -1,58 +0,0 @@
-import pytest
-from pydantic import ValidationError
-
-from paddler_client.error import (
-    BatchSizeExceedsContextSizeError,
-    PenaltiesWithoutWindowError,
-    PenaltyWindowWithoutPenaltiesError,
-)
-from paddler_client.inference_parameters import InferenceParameters
-from paddler_client.pooling_type import PoolingType
-from tests.unit.validation_error_cause import validation_error_cause
-
-
-def test_inference_parameters_defaults() -> None:
-    params = InferenceParameters()
-
-    assert params.temperature == 0.8
-    assert params.context_size == 8192
-    assert params.pooling_type == PoolingType.LAST
-    assert params.enable_embeddings is False
-
-
-def test_inference_parameters_serialization() -> None:
-    params = InferenceParameters(temperature=0.5, top_k=40)
-    dumped = params.model_dump(mode="json")
-
-    assert dumped["temperature"] == 0.5
-    assert dumped["top_k"] == 40
-
-
-def test_inference_parameters_reject_a_batch_larger_than_the_context() -> None:
-    with pytest.raises(ValidationError) as validation_error:
-        InferenceParameters(n_batch=8192, context_size=4096)
-
-    assert isinstance(
-        validation_error_cause(validation_error.value),
-        BatchSizeExceedsContextSizeError,
-    )
-
-
-def test_inference_parameters_reject_penalty_strengths_without_a_window() -> None:
-    with pytest.raises(ValidationError) as validation_error:
-        InferenceParameters(penalty_presence=0.8)
-
-    assert isinstance(
-        validation_error_cause(validation_error.value),
-        PenaltiesWithoutWindowError,
-    )
-
-
-def test_inference_parameters_reject_a_window_without_penalty_strengths() -> None:
-    with pytest.raises(ValidationError) as validation_error:
-        InferenceParameters(penalty_last_n=64)
-
-    assert isinstance(
-        validation_error_cause(validation_error.value),
-        PenaltyWindowWithoutPenaltiesError,
-    )
diff --git a/paddler_client_python/tests/unit/test_inference_socket_pool_rejects_an_empty_pool.py b/paddler_client_python/tests/unit/test_inference_socket_pool_rejects_an_empty_pool.py
deleted file mode 100644
index 3a03acc2e..000000000
--- a/paddler_client_python/tests/unit/test_inference_socket_pool_rejects_an_empty_pool.py
+++ /dev/null
@@ -1,11 +0,0 @@
-import pytest
-
-from paddler_client.error import InvalidSocketPoolSizeError
-from paddler_client.inference_socket_pool import InferenceSocketPool
-
-
-def test_inference_socket_pool_rejects_an_empty_pool() -> None:
-    with pytest.raises(InvalidSocketPoolSizeError) as rejected:
-        InferenceSocketPool(url="ws://127.0.0.1:8061", pool_size=0)
-
-    assert rejected.value.pool_size == 0
diff --git a/paddler_client_python/tests/unit/test_inference_socket_url.py b/paddler_client_python/tests/unit/test_inference_socket_url.py
deleted file mode 100644
index 0d6053e2e..000000000
--- a/paddler_client_python/tests/unit/test_inference_socket_url.py
+++ /dev/null
@@ -1,45 +0,0 @@
-import pytest
-
-from paddler_client.error import UnsupportedUrlSchemeError
-from paddler_client.inference_socket_url import inference_socket_url
-
-
-def test_http_to_ws() -> None:
-    result = inference_socket_url("http://localhost:8080")
-
-    assert result == "ws://localhost:8080/api/v1/inference_socket"
-
-
-def test_https_to_wss() -> None:
-    result = inference_socket_url("https://localhost:8080")
-
-    assert result == "wss://localhost:8080/api/v1/inference_socket"
-
-
-def test_ws_stays_ws() -> None:
-    result = inference_socket_url("ws://localhost:8080")
-
-    assert result == "ws://localhost:8080/api/v1/inference_socket"
-
-
-def test_wss_stays_wss() -> None:
-    result = inference_socket_url("wss://localhost:8080")
-
-    assert result == "wss://localhost:8080/api/v1/inference_socket"
-
-
-def test_replaces_existing_path() -> None:
-    result = inference_socket_url("http://example.com:9090/ignored/path")
-
-    assert result == "ws://example.com:9090/api/v1/inference_socket"
-
-
-def test_preserves_query_parameters() -> None:
-    result = inference_socket_url("http://localhost:8080?token=abc")
-
-    assert result == "ws://localhost:8080/api/v1/inference_socket?token=abc"
-
-
-def test_rejects_a_scheme_that_is_neither_http_nor_websocket() -> None:
-    with pytest.raises(UnsupportedUrlSchemeError):
-        inference_socket_url("ftp://localhost:8080")
diff --git a/paddler_client_python/tests/unit/test_kv_cache_dtype.py b/paddler_client_python/tests/unit/test_kv_cache_dtype.py
deleted file mode 100644
index da3d26af6..000000000
--- a/paddler_client_python/tests/unit/test_kv_cache_dtype.py
+++ /dev/null
@@ -1,13 +0,0 @@
-from paddler_client.kv_cache_dtype import KvCacheDtype
-
-
-def test_kv_cache_dtype_uses_the_wire_names() -> None:
-    assert KvCacheDtype.F32.value == "F32"
-    assert KvCacheDtype.F16.value == "F16"
-    assert KvCacheDtype.BF16.value == "BF16"
-    assert KvCacheDtype.Q8_0.value == "Q8_0"
-    assert KvCacheDtype.Q4_0.value == "Q4_0"
-    assert KvCacheDtype.Q4_1.value == "Q4_1"
-    assert KvCacheDtype.IQ4_NL.value == "IQ4_NL"
-    assert KvCacheDtype.Q5_0.value == "Q5_0"
-    assert KvCacheDtype.Q5_1.value == "Q5_1"
diff --git a/paddler_client_python/tests/unit/test_model_download_status.py b/paddler_client_python/tests/unit/test_model_download_status.py
deleted file mode 100644
index 8aa61ca60..000000000
--- a/paddler_client_python/tests/unit/test_model_download_status.py
+++ /dev/null
@@ -1,67 +0,0 @@
-import pytest
-from pydantic import ValidationError
-
-from paddler_client.error import InvalidModelDownloadStatusError
-from paddler_client.model_download_status import ModelDownloadStatus
-from tests.unit.validation_error_cause import validation_error_cause
-
-
-def test_model_download_status_reads_a_download_of_known_size() -> None:
-    download_status = ModelDownloadStatus.model_validate(
-        {
-            "Downloading": {
-                "downloaded_bytes": 100,
-                "model_path": "https://example.com/model.gguf",
-                "total_bytes": 1000,
-            }
-        }
-    )
-
-    assert download_status == ModelDownloadStatus(
-        variant="Downloading",
-        downloaded_bytes=100,
-        model_path="https://example.com/model.gguf",
-        total_bytes=1000,
-    )
-
-
-def test_model_download_status_reads_no_download() -> None:
-    download_status = ModelDownloadStatus.model_validate("NotDownloading")
-
-    assert download_status == ModelDownloadStatus(variant="NotDownloading")
-
-
-def test_model_download_status_rejects_an_unknown_variant() -> None:
-    with pytest.raises(ValidationError) as rejection:
-        ModelDownloadStatus.model_validate({"Paused": {"downloaded_bytes": 100}})
-
-    assert isinstance(
-        validation_error_cause(rejection.value), InvalidModelDownloadStatusError
-    )
-
-
-def test_model_download_status_rejects_a_download_without_its_progress() -> None:
-    with pytest.raises(ValidationError) as rejection:
-        ModelDownloadStatus.model_validate("Downloading")
-
-    assert isinstance(
-        validation_error_cause(rejection.value), InvalidModelDownloadStatusError
-    )
-
-
-def test_model_download_status_rejects_a_status_naming_two_variants() -> None:
-    with pytest.raises(ValidationError) as rejection:
-        ModelDownloadStatus.model_validate(
-            {
-                "Downloading": {
-                    "downloaded_bytes": 1,
-                    "model_path": "a",
-                    "total_bytes": 2,
-                },
-                "NotDownloading": {},
-            }
-        )
-
-    assert isinstance(
-        validation_error_cause(rejection.value), InvalidModelDownloadStatusError
-    )
diff --git a/paddler_client_python/tests/unit/test_model_metadata.py b/paddler_client_python/tests/unit/test_model_metadata.py
deleted file mode 100644
index af4599355..000000000
--- a/paddler_client_python/tests/unit/test_model_metadata.py
+++ /dev/null
@@ -1,16 +0,0 @@
-from paddler_client.model_metadata import ModelMetadata
-
-
-def test_model_metadata_deserialization() -> None:
-    metadata = ModelMetadata.model_validate(
-        {"metadata": {"architecture": "llama", "params": "7B"}}
-    )
-
-    assert metadata.metadata["architecture"] == "llama"
-    assert metadata.metadata["params"] == "7B"
-
-
-def test_model_metadata_defaults_to_no_entries() -> None:
-    metadata = ModelMetadata()
-
-    assert metadata.metadata == {}
diff --git a/paddler_client_python/tests/unit/test_ndjson_inference_messages_skip_blank_lines.py b/paddler_client_python/tests/unit/test_ndjson_inference_messages_skip_blank_lines.py
deleted file mode 100644
index b8fa07d4f..000000000
--- a/paddler_client_python/tests/unit/test_ndjson_inference_messages_skip_blank_lines.py
+++ /dev/null
@@ -1,26 +0,0 @@
-import json
-from collections.abc import AsyncGenerator
-
-from paddler_client.ndjson_inference_messages import ndjson_inference_messages
-
-
-async def blank_line_then_message_line() -> AsyncGenerator[str, None]:
-    yield "  \r"
-    yield json.dumps(
-        {
-            "Response": {
-                "generated_by": None,
-                "request_id": "req-1",
-                "response": {"Embedding": "NoEmbeddingsProduced"},
-            }
-        }
-    )
-
-
-async def test_ndjson_inference_messages_skip_blank_lines() -> None:
-    messages = [
-        message
-        async for message in ndjson_inference_messages(blank_line_then_message_line())
-    ]
-
-    assert [message.request_id for message in messages] == ["req-1"]
diff --git a/paddler_client_python/tests/unit/test_parsed_tool_call.py b/paddler_client_python/tests/unit/test_parsed_tool_call.py
deleted file mode 100644
index 245efbdbb..000000000
--- a/paddler_client_python/tests/unit/test_parsed_tool_call.py
+++ /dev/null
@@ -1,42 +0,0 @@
-import pytest
-
-from paddler_client.error import ToolCallArgumentsNotAnObjectError
-from paddler_client.parsed_tool_call import ParsedToolCall
-from paddler_client.tool_call_arguments import InvalidJson, ValidJson
-
-
-def test_from_dict_with_valid_json_arguments() -> None:
-    parsed = ParsedToolCall.from_dict(
-        {
-            "id": "call_42",
-            "name": "get_weather",
-            "arguments": {"ValidJson": {"location": "Paris"}},
-        },
-    )
-
-    assert parsed.id == "call_42"
-    assert parsed.name == "get_weather"
-    assert parsed.arguments == ValidJson({"location": "Paris"})
-
-
-def test_from_dict_with_invalid_json_arguments() -> None:
-    parsed = ParsedToolCall.from_dict(
-        {
-            "id": "call_99",
-            "name": "freeform",
-            "arguments": {"InvalidJson": "{half a json"},
-        },
-    )
-
-    assert parsed.arguments == InvalidJson("{half a json")
-
-
-def test_from_dict_with_non_dict_arguments_raises() -> None:
-    with pytest.raises(ToolCallArgumentsNotAnObjectError):
-        ParsedToolCall.from_dict(
-            {
-                "id": "x",
-                "name": "y",
-                "arguments": "not a dict",
-            },
-        )
diff --git a/paddler_client_python/tests/unit/test_pending_requests_delivers_a_message_to_its_request.py b/paddler_client_python/tests/unit/test_pending_requests_delivers_a_message_to_its_request.py
deleted file mode 100644
index 23e21e771..000000000
--- a/paddler_client_python/tests/unit/test_pending_requests_delivers_a_message_to_its_request.py
+++ /dev/null
@@ -1,25 +0,0 @@
-import json
-
-from paddler_client.inference_message import InferenceMessageKind
-from paddler_client.pending_requests import PendingRequests
-
-
-async def test_pending_requests_delivers_a_message_to_its_request() -> None:
-    pending_requests = PendingRequests()
-    response_stream = pending_requests.register("request-1")
-
-    pending_requests.dispatch(
-        json.dumps(
-            {
-                "Error": {
-                    "request_id": "request-1",
-                    "error": {"code": 504, "description": "timed out"},
-                }
-            }
-        )
-    )
-
-    [message] = [message async for message in response_stream]
-
-    assert message.kind == InferenceMessageKind.SERVER_ERROR
-    assert message.error_code == 504
diff --git a/paddler_client_python/tests/unit/test_pending_requests_drops_every_request_when_drained.py b/paddler_client_python/tests/unit/test_pending_requests_drops_every_request_when_drained.py
deleted file mode 100644
index 272ae3c01..000000000
--- a/paddler_client_python/tests/unit/test_pending_requests_drops_every_request_when_drained.py
+++ /dev/null
@@ -1,22 +0,0 @@
-import pytest
-
-from paddler_client.error import ConnectionDroppedError
-from paddler_client.pending_requests import PendingRequests
-
-
-async def test_pending_requests_drops_every_request_when_drained() -> None:
-    pending_requests = PendingRequests()
-    response_streams = [
-        pending_requests.register(request_id)
-        for request_id in ("request-1", "request-2")
-    ]
-
-    pending_requests.drain()
-
-    for request_id, response_stream in zip(
-        ("request-1", "request-2"), response_streams, strict=True
-    ):
-        with pytest.raises(ConnectionDroppedError) as connection_dropped:
-            await anext(response_stream)
-
-        assert connection_dropped.value.request_id == request_id
diff --git a/paddler_client_python/tests/unit/test_pending_requests_fails_a_request_whose_error_cannot_be_parsed.py b/paddler_client_python/tests/unit/test_pending_requests_fails_a_request_whose_error_cannot_be_parsed.py
deleted file mode 100644
index 453226ec6..000000000
--- a/paddler_client_python/tests/unit/test_pending_requests_fails_a_request_whose_error_cannot_be_parsed.py
+++ /dev/null
@@ -1,19 +0,0 @@
-import json
-
-import pytest
-
-from paddler_client.error import JsonError
-from paddler_client.pending_requests import PendingRequests
-
-
-async def test_pending_requests_fails_a_request_whose_error_cannot_be_parsed() -> None:
-    pending_requests = PendingRequests()
-    response_stream = pending_requests.register("request-1")
-    raw_message = json.dumps({"Error": {"request_id": "request-1"}})
-
-    pending_requests.dispatch(raw_message)
-
-    with pytest.raises(JsonError) as json_error:
-        await anext(response_stream)
-
-    assert json_error.value.raw_data == raw_message
diff --git a/paddler_client_python/tests/unit/test_pending_requests_fails_a_request_whose_response_cannot_be_parsed.py b/paddler_client_python/tests/unit/test_pending_requests_fails_a_request_whose_response_cannot_be_parsed.py
deleted file mode 100644
index 0381b12c9..000000000
--- a/paddler_client_python/tests/unit/test_pending_requests_fails_a_request_whose_response_cannot_be_parsed.py
+++ /dev/null
@@ -1,29 +0,0 @@
-import json
-
-import pytest
-
-from paddler_client.error import JsonError
-from paddler_client.pending_requests import PendingRequests
-
-
-async def test_pending_requests_fails_a_request_whose_response_cannot_be_parsed() -> (
-    None
-):
-    pending_requests = PendingRequests()
-    response_stream = pending_requests.register("request-1")
-    raw_message = json.dumps(
-        {
-            "Response": {
-                "generated_by": None,
-                "request_id": "request-1",
-                "response": "NotAVariant",
-            }
-        }
-    )
-
-    pending_requests.dispatch(raw_message)
-
-    with pytest.raises(JsonError) as json_error:
-        await anext(response_stream)
-
-    assert json_error.value.raw_data == raw_message
diff --git a/paddler_client_python/tests/unit/test_pending_requests_ignores_a_message_for_an_unknown_request.py b/paddler_client_python/tests/unit/test_pending_requests_ignores_a_message_for_an_unknown_request.py
deleted file mode 100644
index 485ac09be..000000000
--- a/paddler_client_python/tests/unit/test_pending_requests_ignores_a_message_for_an_unknown_request.py
+++ /dev/null
@@ -1,24 +0,0 @@
-import json
-
-from paddler_client.pending_requests import PendingRequests
-
-
-def timeout_error_for(request_id: str) -> str:
-    return json.dumps(
-        {
-            "Error": {
-                "request_id": request_id,
-                "error": {"code": 504, "description": "timed out"},
-            }
-        }
-    )
-
-
-async def test_pending_requests_ignores_a_message_for_an_unknown_request() -> None:
-    pending_requests = PendingRequests()
-    response_stream = pending_requests.register("request-1")
-
-    pending_requests.dispatch(timeout_error_for("request-that-was-never-sent"))
-    pending_requests.dispatch(timeout_error_for("request-1"))
-
-    assert [message.request_id async for message in response_stream] == ["request-1"]
diff --git a/paddler_client_python/tests/unit/test_pending_requests_ignores_malformed_messages_it_cannot_attribute.py b/paddler_client_python/tests/unit/test_pending_requests_ignores_malformed_messages_it_cannot_attribute.py
deleted file mode 100644
index 22f533d9e..000000000
--- a/paddler_client_python/tests/unit/test_pending_requests_ignores_malformed_messages_it_cannot_attribute.py
+++ /dev/null
@@ -1,26 +0,0 @@
-import json
-
-from paddler_client.pending_requests import PendingRequests
-
-
-async def test_pending_requests_ignores_malformed_messages_it_cannot_attribute() -> (
-    None
-):
-    pending_requests = PendingRequests()
-    response_stream = pending_requests.register("request-1")
-
-    pending_requests.dispatch("not json")
-    pending_requests.dispatch("[1, 2, 3]")
-    pending_requests.dispatch('{"Unknown": {}}')
-    pending_requests.dispatch(
-        json.dumps(
-            {
-                "Error": {
-                    "request_id": "request-1",
-                    "error": {"code": 504, "description": "timed out"},
-                }
-            }
-        )
-    )
-
-    assert [message.error_code async for message in response_stream] == [504]
diff --git a/paddler_client_python/tests/unit/test_pending_requests_keeps_the_first_of_two_registrations_of_a_request.py b/paddler_client_python/tests/unit/test_pending_requests_keeps_the_first_of_two_registrations_of_a_request.py
deleted file mode 100644
index 8a29aa77d..000000000
--- a/paddler_client_python/tests/unit/test_pending_requests_keeps_the_first_of_two_registrations_of_a_request.py
+++ /dev/null
@@ -1,22 +0,0 @@
-import pytest
-
-from paddler_client.error import ConnectionDroppedError, RequestIdInFlightError
-from paddler_client.pending_requests import PendingRequests
-
-
-async def test_pending_requests_keeps_the_first_of_two_registrations_of_a_request() -> (
-    None
-):
-    pending_requests = PendingRequests()
-    first_response_stream = pending_requests.register("request-1")
-
-    with pytest.raises(RequestIdInFlightError) as request_id_in_flight:
-        pending_requests.register("request-1")
-
-    pending_requests.drain()
-
-    with pytest.raises(ConnectionDroppedError) as connection_dropped:
-        await anext(first_response_stream)
-
-    assert request_id_in_flight.value.request_id == "request-1"
-    assert connection_dropped.value.request_id == "request-1"
diff --git a/paddler_client_python/tests/unit/test_pooling_type.py b/paddler_client_python/tests/unit/test_pooling_type.py
deleted file mode 100644
index 37b270d88..000000000
--- a/paddler_client_python/tests/unit/test_pooling_type.py
+++ /dev/null
@@ -1,10 +0,0 @@
-from paddler_client.pooling_type import PoolingType
-
-
-def test_pooling_type_uses_the_wire_names() -> None:
-    assert PoolingType.MEAN.value == "Mean"
-    assert PoolingType.CLS.value == "Cls"
-    assert PoolingType.LAST.value == "Last"
-    assert PoolingType.NONE.value == "None"
-    assert PoolingType.RANK.value == "Rank"
-    assert PoolingType.UNSPECIFIED.value == "Unspecified"
diff --git a/paddler_client_python/tests/unit/test_tool.py b/paddler_client_python/tests/unit/test_tool.py
deleted file mode 100644
index bf96cfc26..000000000
--- a/paddler_client_python/tests/unit/test_tool.py
+++ /dev/null
@@ -1,58 +0,0 @@
-from paddler_client.tool import Function, Tool
-from paddler_client.validated_parameters_schema import (
-    ValidatedParametersSchema,
-)
-
-
-def test_tool_serialization() -> None:
-    tool = Tool(
-        function=Function(
-            name="get_weather",
-            description="Get weather",
-        )
-    )
-    dumped = tool.model_dump(mode="json", exclude_none=True)
-
-    assert dumped == {
-        "type": "function",
-        "function": {
-            "name": "get_weather",
-            "description": "Get weather",
-        },
-    }
-
-
-def test_tool_with_parameters_serialization() -> None:
-    tool = Tool(
-        function=Function(
-            name="get_weather",
-            description="Get weather",
-            parameters=ValidatedParametersSchema.model_validate(
-                {
-                    "type": "object",
-                    "properties": {"location": {"type": "string"}},
-                    "required": ["location"],
-                }
-            ),
-        )
-    )
-    dumped = tool.model_dump(
-        mode="json",
-        exclude_none=True,
-        by_alias=True,
-    )
-
-    assert dumped["function"]["parameters"]["type"] == "object"
-    assert "location" in dumped["function"]["parameters"]["properties"]
-    assert dumped["function"]["parameters"]["required"] == ["location"]
-
-
-def test_tool_deserialization() -> None:
-    tool = Tool.model_validate(
-        {
-            "type": "function",
-            "function": {"name": "test", "description": "desc"},
-        }
-    )
-
-    assert tool.function.name == "test"
diff --git a/paddler_client_python/tests/unit/test_tool_call_arguments.py b/paddler_client_python/tests/unit/test_tool_call_arguments.py
deleted file mode 100644
index cb4d1167f..000000000
--- a/paddler_client_python/tests/unit/test_tool_call_arguments.py
+++ /dev/null
@@ -1,31 +0,0 @@
-import pytest
-
-from paddler_client.error import UnknownToolCallArgumentsError
-from paddler_client.tool_call_arguments import (
-    InvalidJson,
-    ValidJson,
-    parse_tool_call_arguments,
-)
-
-
-def test_parse_valid_json_with_object() -> None:
-    result = parse_tool_call_arguments({"ValidJson": {"location": "Paris"}})
-
-    assert result == ValidJson({"location": "Paris"})
-
-
-def test_parse_valid_json_with_array() -> None:
-    result = parse_tool_call_arguments({"ValidJson": [1, 2, 3]})
-
-    assert result == ValidJson([1, 2, 3])
-
-
-def test_parse_invalid_json_carries_raw_text() -> None:
-    result = parse_tool_call_arguments({"InvalidJson": "{half a json"})
-
-    assert result == InvalidJson("{half a json")
-
-
-def test_parse_unknown_shape_raises() -> None:
-    with pytest.raises(UnknownToolCallArgumentsError):
-        parse_tool_call_arguments({"SomethingElse": "x"})
diff --git a/paddler_client_python/tests/unit/validation_error_cause.py b/paddler_client_python/tests/unit/validation_error_cause.py
deleted file mode 100644
index 89d7aa097..000000000
--- a/paddler_client_python/tests/unit/validation_error_cause.py
+++ /dev/null
@@ -1,7 +0,0 @@
-from pydantic import ValidationError
-
-
-def validation_error_cause(validation_error: ValidationError) -> object:
-    [error_details] = validation_error.errors()
-
-    return error_details["ctx"]["error"]
diff --git a/paddler_dashboard_assets/Cargo.toml b/paddler_dashboard_assets/Cargo.toml
new file mode 100644
index 000000000..39c37d6e0
--- /dev/null
+++ b/paddler_dashboard_assets/Cargo.toml
@@ -0,0 +1,19 @@
+[package]
+name = "paddler_dashboard_assets"
+authors.workspace = true
+description = "Renders the asset tags of the Paddler web admin panel dashboard from its esbuild metafile"
+edition.workspace = true
+homepage.workspace = true
+license.workspace = true
+repository.workspace = true
+version.workspace = true
+
+[dependencies]
+esbuild-metafile = { workspace = true }
+thiserror = { workspace = true }
+
+[dev-dependencies]
+tempfile = { workspace = true }
+
+[lints]
+workspace = true
diff --git a/paddler_dashboard_assets/fixtures/dashboard_metafile.json b/paddler_dashboard_assets/fixtures/dashboard_metafile.json
new file mode 100644
index 000000000..8d0242a41
--- /dev/null
+++ b/paddler_dashboard_assets/fixtures/dashboard_metafile.json
@@ -0,0 +1,31 @@
+{
+    "outputs": {
+        "static/controller_dashboard.js": {
+            "imports": [{ "path": "static/icon.svg" }],
+            "cssBundle": "static/controller_dashboard.css",
+            "entryPoint": "resources/ts/controller_dashboard.tsx"
+        },
+        "static/controller_dashboard.css": {
+            "imports": [{ "path": "static/icon.svg" }],
+            "cssBundle": null,
+            "entryPoint": null
+        },
+        "static/page-admin-panel.css": {
+            "imports": [{ "path": "static/font.ttf" }],
+            "cssBundle": null,
+            "entryPoint": "resources/css/page-admin-panel.css"
+        },
+        "static/icon.svg": {
+            "imports": [],
+            "cssBundle": null,
+            "entryPoint": null,
+            "inputs": { "resources/icons/icon.svg": {} }
+        },
+        "static/font.ttf": {
+            "imports": [],
+            "cssBundle": null,
+            "entryPoint": null,
+            "inputs": { "resources/fonts/font.ttf": {} }
+        }
+    }
+}
diff --git a/paddler_dashboard_assets/src/dashboard_assets.rs b/paddler_dashboard_assets/src/dashboard_assets.rs
new file mode 100644
index 000000000..266d7bb62
--- /dev/null
+++ b/paddler_dashboard_assets/src/dashboard_assets.rs
@@ -0,0 +1,85 @@
+use std::collections::BTreeSet;
+
+use esbuild_metafile::asset::Asset;
+use esbuild_metafile::esbuild_metafile::EsbuildMetafile;
+use esbuild_metafile::input_lookup::InputLookup;
+use esbuild_metafile::input_properties::InputProperties;
+use esbuild_metafile::path_renderer::PathRenderer;
+use esbuild_metafile::preloadable_asset::PreloadableAsset;
+
+use crate::dashboard_assets_error::DashboardAssetsError;
+
+const DASHBOARD_INPUT_PATHS: [&str; 2] = [
+    "resources/css/page-admin-panel.css",
+    "resources/ts/controller_dashboard.tsx",
+];
+
+pub struct DashboardAssets {
+    pub html: String,
+}
+
+impl DashboardAssets {
+    pub fn of(esbuild_metafile: &EsbuildMetafile) -> Result<Self, DashboardAssetsError> {
+        let path_renderer = PathRenderer {};
+        let mut rendered_includes = BTreeSet::new();
+        let mut rendered_preloads = BTreeSet::new();
+
+        for input_path in DASHBOARD_INPUT_PATHS {
+            let InputLookup::Found(InputProperties {
+                outputs, preloads, ..
+            }) = esbuild_metafile.input(input_path)
+            else {
+                return Err(DashboardAssetsError::DashboardInputMissing { input_path });
+            };
+
+            rendered_preloads.extend(preloads.into_iter().map(|preload_path| {
+                PreloadableAsset::from_path(preload_path).render(&path_renderer)
+            }));
+
+            for output_path in outputs {
+                match Asset::from_path(output_path) {
+                    Asset::Unknown(_) => {}
+                    includable_asset => {
+                        rendered_includes.insert(includable_asset.render(&path_renderer));
+                    }
+                }
+            }
+        }
+
+        let mut html = String::new();
+
+        for rendered_tag in rendered_preloads.into_iter().chain(rendered_includes) {
+            html.push_str(&rendered_tag);
+            html.push('\n');
+        }
+
+        Ok(Self { html })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::str::FromStr as _;
+
+    use esbuild_metafile::esbuild_metafile::EsbuildMetafile;
+
+    use super::DashboardAssets;
+
+    const DASHBOARD_METAFILE: &str = include_str!("../fixtures/dashboard_metafile.json");
+
+    #[test]
+    fn renders_every_preload_before_the_includes_of_both_dashboard_inputs() {
+        let esbuild_metafile = EsbuildMetafile::from_str(DASHBOARD_METAFILE).unwrap();
+
+        assert_eq!(
+            DashboardAssets::of(&esbuild_metafile).unwrap().html,
+            concat!(
+                "<link rel=\"preload\" href=\"/static/font.ttf\" as=\"font\" crossorigin>\n",
+                "<link rel=\"preload\" href=\"/static/icon.svg\" as=\"image\">\n",
+                "<link rel=\"stylesheet\" href=\"/static/controller_dashboard.css\">\n",
+                "<link rel=\"stylesheet\" href=\"/static/page-admin-panel.css\">\n",
+                "<script async src=\"/static/controller_dashboard.js\" type=\"module\"></script>\n",
+            )
+        );
+    }
+}
diff --git a/paddler_dashboard_assets/src/dashboard_assets_build.rs b/paddler_dashboard_assets/src/dashboard_assets_build.rs
new file mode 100644
index 000000000..eeec64ee7
--- /dev/null
+++ b/paddler_dashboard_assets/src/dashboard_assets_build.rs
@@ -0,0 +1,201 @@
+use std::ffi::OsString;
+use std::fs::read_to_string;
+use std::fs::write;
+use std::io::Write;
+use std::path::Path;
+use std::path::PathBuf;
+use std::str::FromStr as _;
+
+use esbuild_metafile::esbuild_metafile::EsbuildMetafile;
+
+use crate::dashboard_assets::DashboardAssets;
+use crate::dashboard_assets_error::DashboardAssetsError;
+
+const DASHBOARD_ASSETS_HTML_FILE_NAME: &str = "dashboard_assets.html";
+
+pub struct DashboardAssetsBuild {
+    pub metafile_path: PathBuf,
+    pub out_dir: Option<OsString>,
+}
+
+impl DashboardAssetsBuild {
+    pub fn run(&self, mut cargo_instructions: impl Write) -> Result<(), DashboardAssetsError> {
+        let metafile_contents = read_to_string(&self.metafile_path).map_err(|source| {
+            DashboardAssetsError::MetafileUnreadable {
+                path: self.metafile_path.clone(),
+                source,
+            }
+        })?;
+        let esbuild_metafile = EsbuildMetafile::from_str(&metafile_contents).map_err(|source| {
+            DashboardAssetsError::MetafileMalformed {
+                path: self.metafile_path.clone(),
+                source,
+            }
+        })?;
+        let DashboardAssets { html } = DashboardAssets::of(&esbuild_metafile)?;
+        let html_path = Path::new(
+            self.out_dir
+                .as_ref()
+                .ok_or(DashboardAssetsError::OutDirMissing)?,
+        )
+        .join(DASHBOARD_ASSETS_HTML_FILE_NAME);
+
+        write(&html_path, html).map_err(|source| DashboardAssetsError::HtmlUnwritable {
+            path: html_path.clone(),
+            source,
+        })?;
+
+        write!(
+            cargo_instructions,
+            "cargo::rerun-if-changed={}\ncargo::rustc-env=PADDLER_DASHBOARD_ASSETS_HTML={}\n",
+            self.metafile_path.display(),
+            html_path.display()
+        )
+        .map_err(DashboardAssetsError::CargoInstructionsUnwritable)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::ffi::OsString;
+    #[cfg(target_os = "linux")]
+    use std::fs::OpenOptions;
+    use std::fs::read_to_string;
+    use std::fs::write;
+    use std::io::ErrorKind;
+    use std::mem::discriminant;
+    use std::path::Path;
+    use std::path::PathBuf;
+    use std::str::FromStr as _;
+
+    use esbuild_metafile::esbuild_metafile::EsbuildMetafile;
+    use tempfile::TempDir;
+
+    use super::DashboardAssetsBuild;
+    use crate::dashboard_assets::DashboardAssets;
+    use crate::dashboard_assets_error::DashboardAssetsError;
+
+    const DASHBOARD_METAFILE: &str = include_str!("../fixtures/dashboard_metafile.json");
+
+    fn dashboard_metafile_path() -> PathBuf {
+        Path::new(env!("CARGO_MANIFEST_DIR")).join("fixtures/dashboard_metafile.json")
+    }
+
+    fn metafile_containing(directory: &TempDir, contents: &str) -> PathBuf {
+        let metafile_path = directory.path().join("esbuild-meta.json");
+
+        write(&metafile_path, contents).unwrap();
+
+        metafile_path
+    }
+
+    fn build_into(out_dir: &Path, metafile_path: PathBuf) -> DashboardAssetsBuild {
+        DashboardAssetsBuild {
+            metafile_path,
+            out_dir: Some(OsString::from(out_dir)),
+        }
+    }
+
+    #[test]
+    fn writes_the_dashboard_assets_and_points_cargo_at_them() {
+        let out_dir = TempDir::new().unwrap();
+        let mut cargo_instructions = Vec::new();
+        let html_path = out_dir.path().join("dashboard_assets.html");
+
+        build_into(out_dir.path(), dashboard_metafile_path())
+            .run(&mut cargo_instructions)
+            .unwrap();
+
+        assert_eq!(
+            read_to_string(&html_path).unwrap(),
+            DashboardAssets::of(&EsbuildMetafile::from_str(DASHBOARD_METAFILE).unwrap())
+                .unwrap()
+                .html
+        );
+        assert_eq!(
+            String::from_utf8(cargo_instructions).unwrap(),
+            format!(
+                "cargo::rerun-if-changed={}\ncargo::rustc-env=PADDLER_DASHBOARD_ASSETS_HTML={}\n",
+                dashboard_metafile_path().display(),
+                html_path.display()
+            )
+        );
+    }
+
+    #[test]
+    fn a_missing_metafile_cannot_be_read() {
+        let directory = TempDir::new().unwrap();
+        let missing_metafile_path = directory.path().join("esbuild-meta.json");
+
+        assert!(matches!(
+            build_into(directory.path(), missing_metafile_path.clone()).run(Vec::new()),
+            Err(DashboardAssetsError::MetafileUnreadable { path, source })
+                if path == missing_metafile_path && source.kind() == ErrorKind::NotFound
+        ));
+    }
+
+    #[test]
+    fn a_metafile_that_is_not_json_is_malformed() {
+        let directory = TempDir::new().unwrap();
+        let metafile_path = metafile_containing(&directory, "not an esbuild metafile");
+
+        assert!(matches!(
+            build_into(directory.path(), metafile_path.clone()).run(Vec::new()),
+            Err(DashboardAssetsError::MetafileMalformed { path, .. }) if path == metafile_path
+        ));
+    }
+
+    #[test]
+    fn a_metafile_without_the_dashboard_inputs_describes_no_dashboard() {
+        let directory = TempDir::new().unwrap();
+        let metafile_path = metafile_containing(&directory, r#"{ "outputs": {} }"#);
+
+        assert!(matches!(
+            build_into(directory.path(), metafile_path).run(Vec::new()),
+            Err(DashboardAssetsError::DashboardInputMissing {
+                input_path: "resources/css/page-admin-panel.css"
+            })
+        ));
+    }
+
+    #[test]
+    fn a_build_without_an_out_dir_has_nowhere_to_write_the_dashboard_assets() {
+        assert_eq!(
+            DashboardAssetsBuild {
+                metafile_path: dashboard_metafile_path(),
+                out_dir: None,
+            }
+            .run(Vec::new())
+            .err()
+            .as_ref()
+            .map(discriminant),
+            Some(discriminant(&DashboardAssetsError::OutDirMissing))
+        );
+    }
+
+    #[test]
+    fn an_out_dir_that_does_not_exist_cannot_hold_the_dashboard_assets() {
+        let directory = TempDir::new().unwrap();
+        let missing_out_dir = directory.path().join("missing");
+
+        assert!(matches!(
+            build_into(&missing_out_dir, dashboard_metafile_path()).run(Vec::new()),
+            Err(DashboardAssetsError::HtmlUnwritable { path, source })
+                if path == missing_out_dir.join("dashboard_assets.html")
+                    && source.kind() == ErrorKind::NotFound
+        ));
+    }
+
+    #[cfg(target_os = "linux")]
+    #[test]
+    fn cargo_instructions_sent_to_a_full_device_cannot_be_written() {
+        let out_dir = TempDir::new().unwrap();
+
+        assert!(matches!(
+            build_into(out_dir.path(), dashboard_metafile_path())
+                .run(OpenOptions::new().write(true).open("/dev/full").unwrap()),
+            Err(DashboardAssetsError::CargoInstructionsUnwritable(source))
+                if source.kind() == ErrorKind::StorageFull
+        ));
+    }
+}
diff --git a/paddler_dashboard_assets/src/dashboard_assets_error.rs b/paddler_dashboard_assets/src/dashboard_assets_error.rs
new file mode 100644
index 000000000..7d6e17b55
--- /dev/null
+++ b/paddler_dashboard_assets/src/dashboard_assets_error.rs
@@ -0,0 +1,33 @@
+use std::io;
+use std::path::PathBuf;
+
+use esbuild_metafile::error::Error as EsbuildMetafileError;
+use thiserror::Error;
+
+#[derive(Debug, Error)]
+pub enum DashboardAssetsError {
+    #[error("Unable to write the Cargo instructions of the dashboard assets build")]
+    CargoInstructionsUnwritable(#[source] io::Error),
+    #[error("The esbuild metafile does not describe the dashboard input '{input_path}'")]
+    DashboardInputMissing { input_path: &'static str },
+    #[error("Unable to write the dashboard assets to '{path_display}'", path_display = path.display())]
+    HtmlUnwritable {
+        path: PathBuf,
+        #[source]
+        source: io::Error,
+    },
+    #[error("The esbuild metafile '{path_display}' is not an esbuild metafile", path_display = path.display())]
+    MetafileMalformed {
+        path: PathBuf,
+        #[source]
+        source: EsbuildMetafileError,
+    },
+    #[error("Unable to read the esbuild metafile '{path_display}'", path_display = path.display())]
+    MetafileUnreadable {
+        path: PathBuf,
+        #[source]
+        source: io::Error,
+    },
+    #[error("Cargo did not provide the OUT_DIR to write the dashboard assets to")]
+    OutDirMissing,
+}
diff --git a/paddler_dashboard_assets/src/lib.rs b/paddler_dashboard_assets/src/lib.rs
new file mode 100644
index 000000000..828db7fb8
--- /dev/null
+++ b/paddler_dashboard_assets/src/lib.rs
@@ -0,0 +1,3 @@
+pub mod dashboard_assets;
+pub mod dashboard_assets_build;
+pub mod dashboard_assets_error;
diff --git a/paddler_download_manager/src/download_manager.rs b/paddler_download_manager/src/download_manager.rs
index 7305aad4e..a97ace7d2 100644
--- a/paddler_download_manager/src/download_manager.rs
+++ b/paddler_download_manager/src/download_manager.rs
@@ -20,6 +20,7 @@ use crate::download_outcome::DownloadOutcome;
 use crate::download_progress::DownloadProgress;
 use crate::download_url::DownloadUrl;
 use crate::partial_file::PartialFile;
+use crate::read_timeout::READ_TIMEOUT;
 use crate::response_classification::ResponseClassification;
 
 struct PartialFileDownload<'download, TProgressHandler> {
@@ -197,6 +198,7 @@ where
     }
 }
 
+#[derive(Clone)]
 pub struct DownloadManager {
     client: Client,
 }
@@ -205,7 +207,7 @@ impl DownloadManager {
     pub fn new() -> Result<Self, ReqwestError> {
         let client = Client::builder()
             .connect_timeout(Duration::from_secs(10))
-            .read_timeout(Duration::from_secs(10))
+            .read_timeout(READ_TIMEOUT)
             .build()?;
 
         Ok(Self { client })
diff --git a/paddler_download_manager/src/download_url.rs b/paddler_download_manager/src/download_url.rs
index 277c52ab9..251c897b3 100644
--- a/paddler_download_manager/src/download_url.rs
+++ b/paddler_download_manager/src/download_url.rs
@@ -2,24 +2,17 @@ use url::Url;
 
 use crate::download_error::DownloadError;
 
+#[derive(Debug, Eq, PartialEq)]
 pub struct DownloadUrl {
     url: Url,
 }
 
 impl DownloadUrl {
     pub fn parse(url: &str) -> Result<Self, DownloadError> {
-        let parsed_url = Url::parse(url).map_err(|source| DownloadError::InvalidUrl {
+        Self::try_from(Url::parse(url).map_err(|source| DownloadError::InvalidUrl {
             url: url.to_owned(),
             source,
-        })?;
-
-        match parsed_url.scheme() {
-            "http" | "https" => Ok(Self { url: parsed_url }),
-            unsupported_scheme => Err(DownloadError::UnsupportedUrlScheme {
-                url: url.to_owned(),
-                scheme: unsupported_scheme.to_owned(),
-            }),
-        }
+        })?)
     }
 
     #[must_use]
@@ -27,3 +20,17 @@ impl DownloadUrl {
         self.url.as_str()
     }
 }
+
+impl TryFrom<Url> for DownloadUrl {
+    type Error = DownloadError;
+
+    fn try_from(url: Url) -> Result<Self, DownloadError> {
+        match url.scheme() {
+            "http" | "https" => Ok(Self { url }),
+            unsupported_scheme => Err(DownloadError::UnsupportedUrlScheme {
+                scheme: unsupported_scheme.to_owned(),
+                url: url.to_string(),
+            }),
+        }
+    }
+}
diff --git a/paddler_download_manager/src/lib.rs b/paddler_download_manager/src/lib.rs
index d4aaa5ca7..5ae4f25d5 100644
--- a/paddler_download_manager/src/lib.rs
+++ b/paddler_download_manager/src/lib.rs
@@ -4,4 +4,5 @@ pub mod download_outcome;
 pub mod download_progress;
 pub mod download_url;
 pub mod partial_file;
+pub mod read_timeout;
 pub mod response_classification;
diff --git a/paddler_download_manager/src/partial_file.rs b/paddler_download_manager/src/partial_file.rs
index 5d4b0f193..c1a192ea8 100644
--- a/paddler_download_manager/src/partial_file.rs
+++ b/paddler_download_manager/src/partial_file.rs
@@ -101,28 +101,20 @@ impl PartialFile {
 
 #[cfg(test)]
 mod tests {
-    #[cfg(unix)]
     use std::io::ErrorKind;
-    #[cfg(unix)]
     use std::os::unix::fs::PermissionsExt;
 
     use tempfile::TempDir;
-    #[cfg(unix)]
     use tokio::fs::create_dir;
-    #[cfg(unix)]
     use tokio::fs::create_dir_all;
-    #[cfg(unix)]
     use tokio::fs::metadata;
     use tokio::fs::read;
-    #[cfg(unix)]
     use tokio::fs::remove_dir_all;
-    #[cfg(unix)]
     use tokio::fs::set_permissions;
     use tokio::fs::try_exists;
     use tokio::fs::write;
     use tokio::io::AsyncWriteExt;
 
-    #[cfg(unix)]
     use crate::download_error::DownloadError;
     use crate::partial_file::PartialFile;
 
@@ -222,7 +214,6 @@ mod tests {
         partial.remove().await.unwrap();
     }
 
-    #[cfg(unix)]
     #[tokio::test]
     async fn current_size_propagates_non_notfound_error() {
         let directory = TempDir::new().unwrap();
@@ -239,7 +230,6 @@ mod tests {
         ));
     }
 
-    #[cfg(unix)]
     #[tokio::test]
     async fn truncate_returns_io_error_when_partial_is_a_directory() {
         let directory = TempDir::new().unwrap();
@@ -251,7 +241,6 @@ mod tests {
         assert_eq!(result.unwrap_err().kind(), ErrorKind::IsADirectory);
     }
 
-    #[cfg(unix)]
     #[tokio::test]
     async fn open_for_append_returns_io_error_when_partial_is_a_directory() {
         let directory = TempDir::new().unwrap();
@@ -263,7 +252,6 @@ mod tests {
         assert_eq!(result.unwrap_err().kind(), ErrorKind::IsADirectory);
     }
 
-    #[cfg(unix)]
     #[tokio::test]
     async fn finalize_returns_io_error_when_final_is_a_non_empty_directory() {
         let directory = TempDir::new().unwrap();
@@ -279,7 +267,6 @@ mod tests {
         assert_eq!(result.unwrap_err().kind(), ErrorKind::IsADirectory);
     }
 
-    #[cfg(unix)]
     #[tokio::test]
     async fn remove_propagates_non_notfound_error() {
         let directory = TempDir::new().unwrap();
@@ -300,7 +287,6 @@ mod tests {
         assert_eq!(result.unwrap_err().kind(), ErrorKind::PermissionDenied);
     }
 
-    #[cfg(unix)]
     #[tokio::test]
     async fn open_for_append_fails_when_parent_blocked_by_file() {
         let directory = TempDir::new().unwrap();
@@ -313,7 +299,6 @@ mod tests {
         assert_eq!(result.unwrap_err().kind(), ErrorKind::NotADirectory);
     }
 
-    #[cfg(unix)]
     #[tokio::test]
     async fn truncate_fails_when_parent_blocked_by_file() {
         let directory = TempDir::new().unwrap();
@@ -326,7 +311,6 @@ mod tests {
         assert_eq!(result.unwrap_err().kind(), ErrorKind::NotADirectory);
     }
 
-    #[cfg(unix)]
     #[tokio::test]
     async fn finalize_returns_io_error_when_parent_was_deleted_mid_download() {
         let directory = TempDir::new().unwrap();
diff --git a/paddler_download_manager/src/read_timeout.rs b/paddler_download_manager/src/read_timeout.rs
new file mode 100644
index 000000000..53286f5f5
--- /dev/null
+++ b/paddler_download_manager/src/read_timeout.rs
@@ -0,0 +1,3 @@
+use std::time::Duration;
+
+pub const READ_TIMEOUT: Duration = Duration::from_secs(10);
diff --git a/paddler_download_manager/tests/cancelling_mid_stream_keeps_the_partial_file.rs b/paddler_download_manager/tests/cancelling_mid_stream_keeps_the_partial_file.rs
new file mode 100644
index 000000000..5ae43101d
--- /dev/null
+++ b/paddler_download_manager/tests/cancelling_mid_stream_keeps_the_partial_file.rs
@@ -0,0 +1,80 @@
+use tempfile::TempDir;
+use tokio::fs::try_exists;
+use tokio::spawn;
+use tokio::task::yield_now;
+use tokio_util::sync::CancellationToken;
+
+use paddler_download_manager::download_manager::DownloadManager;
+use paddler_download_manager::download_outcome::DownloadOutcome;
+use paddler_download_manager::download_progress::DownloadProgress;
+use paddler_local_http_fixture::fixture_response::FixtureResponse;
+use paddler_local_http_fixture::local_http_fixture::LocalHttpFixture;
+
+use crate::recorded_download::RecordedDownload;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn cancelling_mid_stream_keeps_the_partial_file() {
+    let download_manager =
+        DownloadManager::new().expect("the download manager must build its HTTP client");
+    let directory = TempDir::new().expect("a temporary directory must be creatable");
+    let destination = directory.path().join("model.gguf");
+    let partial_path = destination.with_extension("partial");
+    let fixture = LocalHttpFixture::start(FixtureResponse::StalledBody {
+        sent_body: vec![7_u8; 16],
+        withheld_byte_count: 8176,
+    })
+    .await
+    .expect("the local HTTP fixture must start");
+    let url = fixture.url("/model.gguf");
+    let cancellation_token = CancellationToken::new();
+    let download_cancellation_token = cancellation_token.clone();
+    let download_destination = destination.clone();
+    let download = spawn(async move {
+        RecordedDownload::download(
+            &download_manager,
+            &download_cancellation_token,
+            &url,
+            &download_destination,
+        )
+        .await
+    });
+
+    while !try_exists(&partial_path)
+        .await
+        .expect("the partial path must be inspectable")
+    {
+        yield_now().await;
+    }
+
+    cancellation_token.cancel();
+
+    let recorded_download = download.await.expect("the download task must not panic");
+
+    assert_eq!(
+        recorded_download
+            .result
+            .as_ref()
+            .expect("a cancelled download must not fail"),
+        &DownloadOutcome::Cancelled
+    );
+    assert!(
+        try_exists(&partial_path)
+            .await
+            .expect("the partial path must be inspectable")
+    );
+    assert!(
+        !try_exists(&destination)
+            .await
+            .expect("the destination must be inspectable")
+    );
+    assert_ne!(
+        recorded_download.progress.last(),
+        Some(&DownloadProgress::Finished)
+    );
+
+    fixture
+        .stop()
+        .await
+        .expect("the fixture task must not panic")
+        .expect("the fixture must serve every request");
+}
diff --git a/paddler_download_manager/tests/cancelling_mid_stream_leaves_a_resumable_partial_file.rs b/paddler_download_manager/tests/cancelling_mid_stream_leaves_a_resumable_partial_file.rs
deleted file mode 100644
index 84ce8ec09..000000000
--- a/paddler_download_manager/tests/cancelling_mid_stream_leaves_a_resumable_partial_file.rs
+++ /dev/null
@@ -1,74 +0,0 @@
-use tempfile::TempDir;
-use tokio::fs::try_exists;
-use tokio::spawn;
-use tokio::task::yield_now;
-use tokio_util::sync::CancellationToken;
-
-use paddler_download_manager::download_manager::DownloadManager;
-use paddler_download_manager::download_outcome::DownloadOutcome;
-use paddler_download_manager::download_progress::DownloadProgress;
-use paddler_local_http_fixture::fixture_response::FixtureResponse;
-use paddler_local_http_fixture::local_http_fixture::LocalHttpFixture;
-
-use crate::recorded_download::RecordedDownload;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn cancelling_mid_stream_leaves_a_resumable_partial_file() {
-    let download_manager =
-        DownloadManager::new().expect("the download manager must build its HTTP client");
-    let directory = TempDir::new().expect("a temporary directory must be creatable");
-    let destination = directory.path().join("model.gguf");
-    let partial_path = destination.with_extension("partial");
-    let fixture = LocalHttpFixture::start(FixtureResponse::StalledBody {
-        sent_body: vec![7_u8; 16],
-        withheld_byte_count: 8176,
-    })
-    .await
-    .expect("the local HTTP fixture must start");
-    let url = fixture.url("/model.gguf");
-    let cancellation_token = CancellationToken::new();
-    let download_cancellation_token = cancellation_token.clone();
-    let download_destination = destination.clone();
-    let download = spawn(async move {
-        RecordedDownload::download(
-            &download_manager,
-            &download_cancellation_token,
-            &url,
-            &download_destination,
-        )
-        .await
-    });
-
-    while !try_exists(&partial_path)
-        .await
-        .expect("the partial path must be inspectable")
-    {
-        yield_now().await;
-    }
-
-    cancellation_token.cancel();
-
-    let recorded_download = download.await.expect("the download task must not panic");
-
-    assert_eq!(
-        recorded_download
-            .result
-            .as_ref()
-            .expect("a cancelled download must not fail"),
-        &DownloadOutcome::Cancelled
-    );
-    assert!(
-        try_exists(&partial_path)
-            .await
-            .expect("the partial path must be inspectable")
-    );
-    assert!(
-        !try_exists(&destination)
-            .await
-            .expect("the destination must be inspectable")
-    );
-    assert_ne!(
-        recorded_download.progress.last(),
-        Some(&DownloadProgress::Finished)
-    );
-}
diff --git a/paddler_download_manager/tests/downloads_again_after_the_cache_directory_was_deleted.rs b/paddler_download_manager/tests/downloads_again_after_the_cache_directory_was_deleted.rs
index fa9026008..ccd695a93 100644
--- a/paddler_download_manager/tests/downloads_again_after_the_cache_directory_was_deleted.rs
+++ b/paddler_download_manager/tests/downloads_again_after_the_cache_directory_was_deleted.rs
@@ -51,4 +51,10 @@ async fn downloads_again_after_the_cache_directory_was_deleted() {
         body
     );
     assert_eq!(fixture.request_count(), 2);
+
+    fixture
+        .stop()
+        .await
+        .expect("the fixture task must not panic")
+        .expect("the fixture must serve every request");
 }
diff --git a/paddler_download_manager/tests/downloads_the_body_and_reports_started_chunks_and_finished.rs b/paddler_download_manager/tests/downloads_the_body_and_reports_started_chunks_and_finished.rs
index d9e7fa360..960a159e0 100644
--- a/paddler_download_manager/tests/downloads_the_body_and_reports_started_chunks_and_finished.rs
+++ b/paddler_download_manager/tests/downloads_the_body_and_reports_started_chunks_and_finished.rs
@@ -54,4 +54,10 @@ async fn downloads_the_body_and_reports_started_chunks_and_finished() {
         recorded_download.progress.last(),
         Some(&DownloadProgress::Finished)
     );
+
+    fixture
+        .stop()
+        .await
+        .expect("the fixture task must not panic")
+        .expect("the fixture must serve every request");
 }
diff --git a/paddler_download_manager/tests/main.rs b/paddler_download_manager/tests/main.rs
index 8309e9415..6226780f5 100644
--- a/paddler_download_manager/tests/main.rs
+++ b/paddler_download_manager/tests/main.rs
@@ -1,13 +1,13 @@
-mod cancelling_mid_stream_leaves_a_resumable_partial_file;
+mod cancelling_mid_stream_keeps_the_partial_file;
 mod downloads_again_after_the_cache_directory_was_deleted;
 mod downloads_the_body_and_reports_started_chunks_and_finished;
 mod recorded_download;
 mod rejects_a_directory_at_the_partial_path;
 mod rejects_a_final_path_without_a_file_name_without_a_request;
-mod rejects_an_invalid_url_without_a_request;
-mod rejects_an_unsupported_url_scheme_without_a_request;
+mod rejects_an_invalid_url;
+mod rejects_an_unsupported_url_scheme;
 mod reports_cache_disk_full_when_a_body_chunk_cannot_be_written;
-mod reports_cache_disk_full_when_the_partial_file_is_on_a_full_device;
+mod reports_cache_disk_full_when_the_written_body_cannot_be_flushed;
 mod reports_cache_permission_denied_when_a_stale_partial_cannot_be_removed;
 mod reports_cache_permission_denied_when_the_cache_directory_is_read_only;
 mod reports_cache_permission_denied_when_the_partial_file_cannot_be_truncated;
diff --git a/paddler_download_manager/tests/rejects_a_directory_at_the_partial_path.rs b/paddler_download_manager/tests/rejects_a_directory_at_the_partial_path.rs
index 6cadfd8b9..3991fc754 100644
--- a/paddler_download_manager/tests/rejects_a_directory_at_the_partial_path.rs
+++ b/paddler_download_manager/tests/rejects_a_directory_at_the_partial_path.rs
@@ -1,5 +1,3 @@
-#![cfg(unix)]
-
 use tempfile::TempDir;
 use tokio::fs::create_dir;
 use tokio_util::sync::CancellationToken;
@@ -41,4 +39,10 @@ async fn rejects_a_directory_at_the_partial_path() {
         Err(DownloadError::PartialPathIsADirectory { partial_path: rejected_path })
             if rejected_path == partial_path
     ));
+
+    fixture
+        .stop()
+        .await
+        .expect("the fixture task must not panic")
+        .expect("the fixture must serve every request");
 }
diff --git a/paddler_download_manager/tests/rejects_a_final_path_without_a_file_name_without_a_request.rs b/paddler_download_manager/tests/rejects_a_final_path_without_a_file_name_without_a_request.rs
index e3dade816..4edf2c5ac 100644
--- a/paddler_download_manager/tests/rejects_a_final_path_without_a_file_name_without_a_request.rs
+++ b/paddler_download_manager/tests/rejects_a_final_path_without_a_file_name_without_a_request.rs
@@ -31,4 +31,10 @@ async fn rejects_a_final_path_without_a_file_name_without_a_request() {
         recorded_download.result,
         Err(DownloadError::FinalPathHasNoFileName { final_path }) if final_path == destination
     ));
+
+    fixture
+        .stop()
+        .await
+        .expect("the fixture task must not panic")
+        .expect("the fixture must serve every request");
 }
diff --git a/paddler_download_manager/tests/rejects_an_invalid_url.rs b/paddler_download_manager/tests/rejects_an_invalid_url.rs
new file mode 100644
index 000000000..61339d68d
--- /dev/null
+++ b/paddler_download_manager/tests/rejects_an_invalid_url.rs
@@ -0,0 +1,27 @@
+use tempfile::TempDir;
+use tokio_util::sync::CancellationToken;
+
+use paddler_download_manager::download_error::DownloadError;
+use paddler_download_manager::download_manager::DownloadManager;
+
+use crate::recorded_download::RecordedDownload;
+
+#[tokio::test]
+async fn rejects_an_invalid_url() {
+    let download_manager =
+        DownloadManager::new().expect("the download manager must build its HTTP client");
+    let directory = TempDir::new().expect("a temporary directory must be creatable");
+
+    let recorded_download = RecordedDownload::download(
+        &download_manager,
+        &CancellationToken::new(),
+        "not a valid url",
+        &directory.path().join("model.gguf"),
+    )
+    .await;
+
+    assert!(matches!(
+        recorded_download.result,
+        Err(DownloadError::InvalidUrl { url, .. }) if url == "not a valid url"
+    ));
+}
diff --git a/paddler_download_manager/tests/rejects_an_invalid_url_without_a_request.rs b/paddler_download_manager/tests/rejects_an_invalid_url_without_a_request.rs
deleted file mode 100644
index 740bfa9b4..000000000
--- a/paddler_download_manager/tests/rejects_an_invalid_url_without_a_request.rs
+++ /dev/null
@@ -1,27 +0,0 @@
-use tempfile::TempDir;
-use tokio_util::sync::CancellationToken;
-
-use paddler_download_manager::download_error::DownloadError;
-use paddler_download_manager::download_manager::DownloadManager;
-
-use crate::recorded_download::RecordedDownload;
-
-#[tokio::test]
-async fn rejects_an_invalid_url_without_a_request() {
-    let download_manager =
-        DownloadManager::new().expect("the download manager must build its HTTP client");
-    let directory = TempDir::new().expect("a temporary directory must be creatable");
-
-    let recorded_download = RecordedDownload::download(
-        &download_manager,
-        &CancellationToken::new(),
-        "not a valid url",
-        &directory.path().join("model.gguf"),
-    )
-    .await;
-
-    assert!(matches!(
-        recorded_download.result,
-        Err(DownloadError::InvalidUrl { url, .. }) if url == "not a valid url"
-    ));
-}
diff --git a/paddler_download_manager/tests/rejects_an_unsupported_url_scheme.rs b/paddler_download_manager/tests/rejects_an_unsupported_url_scheme.rs
new file mode 100644
index 000000000..55d9997f6
--- /dev/null
+++ b/paddler_download_manager/tests/rejects_an_unsupported_url_scheme.rs
@@ -0,0 +1,27 @@
+use tempfile::TempDir;
+use tokio_util::sync::CancellationToken;
+
+use paddler_download_manager::download_error::DownloadError;
+use paddler_download_manager::download_manager::DownloadManager;
+
+use crate::recorded_download::RecordedDownload;
+
+#[tokio::test]
+async fn rejects_an_unsupported_url_scheme() {
+    let download_manager =
+        DownloadManager::new().expect("the download manager must build its HTTP client");
+    let directory = TempDir::new().expect("a temporary directory must be creatable");
+
+    let recorded_download = RecordedDownload::download(
+        &download_manager,
+        &CancellationToken::new(),
+        "ftp://example.invalid/model.gguf",
+        &directory.path().join("model.gguf"),
+    )
+    .await;
+
+    assert!(matches!(
+        recorded_download.result,
+        Err(DownloadError::UnsupportedUrlScheme { scheme, .. }) if scheme == "ftp"
+    ));
+}
diff --git a/paddler_download_manager/tests/rejects_an_unsupported_url_scheme_without_a_request.rs b/paddler_download_manager/tests/rejects_an_unsupported_url_scheme_without_a_request.rs
deleted file mode 100644
index fc03d75fa..000000000
--- a/paddler_download_manager/tests/rejects_an_unsupported_url_scheme_without_a_request.rs
+++ /dev/null
@@ -1,27 +0,0 @@
-use tempfile::TempDir;
-use tokio_util::sync::CancellationToken;
-
-use paddler_download_manager::download_error::DownloadError;
-use paddler_download_manager::download_manager::DownloadManager;
-
-use crate::recorded_download::RecordedDownload;
-
-#[tokio::test]
-async fn rejects_an_unsupported_url_scheme_without_a_request() {
-    let download_manager =
-        DownloadManager::new().expect("the download manager must build its HTTP client");
-    let directory = TempDir::new().expect("a temporary directory must be creatable");
-
-    let recorded_download = RecordedDownload::download(
-        &download_manager,
-        &CancellationToken::new(),
-        "ftp://example.invalid/model.gguf",
-        &directory.path().join("model.gguf"),
-    )
-    .await;
-
-    assert!(matches!(
-        recorded_download.result,
-        Err(DownloadError::UnsupportedUrlScheme { scheme, .. }) if scheme == "ftp"
-    ));
-}
diff --git a/paddler_download_manager/tests/reports_cache_disk_full_when_a_body_chunk_cannot_be_written.rs b/paddler_download_manager/tests/reports_cache_disk_full_when_a_body_chunk_cannot_be_written.rs
index 4a486f64d..f608b6113 100644
--- a/paddler_download_manager/tests/reports_cache_disk_full_when_a_body_chunk_cannot_be_written.rs
+++ b/paddler_download_manager/tests/reports_cache_disk_full_when_a_body_chunk_cannot_be_written.rs
@@ -14,6 +14,7 @@ use paddler_local_http_fixture::local_http_fixture::LocalHttpFixture;
 use crate::recorded_download::RecordedDownload;
 
 const HYPER_DEFAULT_MAX_READ_BUFFER_BYTES: usize = 8192 + 4096 * 100;
+const BODY_LENGTH: usize = 2 * HYPER_DEFAULT_MAX_READ_BUFFER_BYTES + 1;
 
 #[tokio::test]
 async fn reports_cache_disk_full_when_a_body_chunk_cannot_be_written() {
@@ -25,13 +26,9 @@ async fn reports_cache_disk_full_when_a_body_chunk_cannot_be_written() {
     symlink("/dev/full", destination.with_extension("partial"))
         .expect("the partial path must link to the full device");
 
-    let fixture = LocalHttpFixture::start(FixtureResponse::Ok(vec![
-        0_u8;
-        2 * HYPER_DEFAULT_MAX_READ_BUFFER_BYTES
-            + 1
-    ]))
-    .await
-    .expect("the local HTTP fixture must start");
+    let fixture = LocalHttpFixture::start(FixtureResponse::Ok(vec![0_u8; BODY_LENGTH]))
+        .await
+        .expect("the local HTTP fixture must start");
 
     let recorded_download = RecordedDownload::download(
         &download_manager,
@@ -42,7 +39,14 @@ async fn reports_cache_disk_full_when_a_body_chunk_cannot_be_written() {
     .await;
 
     assert!(matches!(
-        recorded_download.result,
+        &recorded_download.result,
         Err(DownloadError::CacheDiskFull { source, .. }) if source.kind() == ErrorKind::StorageFull
     ));
+    assert!(recorded_download.written_byte_count() < BODY_LENGTH as u64);
+
+    fixture
+        .stop()
+        .await
+        .expect("the fixture task must not panic")
+        .expect("the fixture must serve every request");
 }
diff --git a/paddler_download_manager/tests/reports_cache_disk_full_when_the_partial_file_is_on_a_full_device.rs b/paddler_download_manager/tests/reports_cache_disk_full_when_the_partial_file_is_on_a_full_device.rs
deleted file mode 100644
index 959b54461..000000000
--- a/paddler_download_manager/tests/reports_cache_disk_full_when_the_partial_file_is_on_a_full_device.rs
+++ /dev/null
@@ -1,44 +0,0 @@
-#![cfg(target_os = "linux")]
-
-use std::io::ErrorKind;
-use std::os::unix::fs::symlink;
-
-use tempfile::TempDir;
-use tokio_util::sync::CancellationToken;
-
-use paddler_download_manager::download_error::DownloadError;
-use paddler_download_manager::download_manager::DownloadManager;
-use paddler_local_http_fixture::fixture_response::FixtureResponse;
-use paddler_local_http_fixture::local_http_fixture::LocalHttpFixture;
-
-use crate::recorded_download::RecordedDownload;
-
-#[tokio::test]
-async fn reports_cache_disk_full_when_the_partial_file_is_on_a_full_device() {
-    let download_manager =
-        DownloadManager::new().expect("the download manager must build its HTTP client");
-    let directory = TempDir::new().expect("a temporary directory must be creatable");
-    let destination = directory.path().join("model.gguf");
-
-    symlink("/dev/full", destination.with_extension("partial"))
-        .expect("the partial path must link to the full device");
-
-    let fixture = LocalHttpFixture::start(FixtureResponse::Ok(
-        b"this body will fail to write because /dev/full".to_vec(),
-    ))
-    .await
-    .expect("the local HTTP fixture must start");
-
-    let recorded_download = RecordedDownload::download(
-        &download_manager,
-        &CancellationToken::new(),
-        &fixture.url("/model.gguf"),
-        &destination,
-    )
-    .await;
-
-    assert!(matches!(
-        recorded_download.result,
-        Err(DownloadError::CacheDiskFull { source, .. }) if source.kind() == ErrorKind::StorageFull
-    ));
-}
diff --git a/paddler_download_manager/tests/reports_cache_disk_full_when_the_written_body_cannot_be_flushed.rs b/paddler_download_manager/tests/reports_cache_disk_full_when_the_written_body_cannot_be_flushed.rs
new file mode 100644
index 000000000..77c9d9d0b
--- /dev/null
+++ b/paddler_download_manager/tests/reports_cache_disk_full_when_the_written_body_cannot_be_flushed.rs
@@ -0,0 +1,51 @@
+#![cfg(target_os = "linux")]
+
+use std::io::ErrorKind;
+use std::os::unix::fs::symlink;
+
+use tempfile::TempDir;
+use tokio_util::sync::CancellationToken;
+
+use paddler_download_manager::download_error::DownloadError;
+use paddler_download_manager::download_manager::DownloadManager;
+use paddler_local_http_fixture::fixture_response::FixtureResponse;
+use paddler_local_http_fixture::local_http_fixture::LocalHttpFixture;
+
+use crate::recorded_download::RecordedDownload;
+
+const BODY: &[u8] = b"this body will fail to write because /dev/full";
+
+#[tokio::test]
+async fn reports_cache_disk_full_when_the_written_body_cannot_be_flushed() {
+    let download_manager =
+        DownloadManager::new().expect("the download manager must build its HTTP client");
+    let directory = TempDir::new().expect("a temporary directory must be creatable");
+    let destination = directory.path().join("model.gguf");
+
+    symlink("/dev/full", destination.with_extension("partial"))
+        .expect("the partial path must link to the full device");
+
+    let fixture = LocalHttpFixture::start(FixtureResponse::Ok(BODY.to_vec()))
+        .await
+        .expect("the local HTTP fixture must start");
+
+    let recorded_download = RecordedDownload::download(
+        &download_manager,
+        &CancellationToken::new(),
+        &fixture.url("/model.gguf"),
+        &destination,
+    )
+    .await;
+
+    assert!(matches!(
+        &recorded_download.result,
+        Err(DownloadError::CacheDiskFull { source, .. }) if source.kind() == ErrorKind::StorageFull
+    ));
+    assert_eq!(recorded_download.written_byte_count(), BODY.len() as u64);
+
+    fixture
+        .stop()
+        .await
+        .expect("the fixture task must not panic")
+        .expect("the fixture must serve every request");
+}
diff --git a/paddler_download_manager/tests/reports_cache_permission_denied_when_a_stale_partial_cannot_be_removed.rs b/paddler_download_manager/tests/reports_cache_permission_denied_when_a_stale_partial_cannot_be_removed.rs
index 970b938ea..140e7280d 100644
--- a/paddler_download_manager/tests/reports_cache_permission_denied_when_a_stale_partial_cannot_be_removed.rs
+++ b/paddler_download_manager/tests/reports_cache_permission_denied_when_a_stale_partial_cannot_be_removed.rs
@@ -1,5 +1,3 @@
-#![cfg(unix)]
-
 use std::fs::Permissions;
 use std::os::unix::fs::PermissionsExt as _;
 
@@ -57,4 +55,10 @@ async fn reports_cache_permission_denied_when_a_stale_partial_cannot_be_removed(
         Err(DownloadError::CachePermissionDenied { path, .. })
             if path == destination.with_extension("partial")
     ));
+
+    fixture
+        .stop()
+        .await
+        .expect("the fixture task must not panic")
+        .expect("the fixture must serve every request");
 }
diff --git a/paddler_download_manager/tests/reports_cache_permission_denied_when_the_cache_directory_is_read_only.rs b/paddler_download_manager/tests/reports_cache_permission_denied_when_the_cache_directory_is_read_only.rs
index 342847cea..f95bbf4cc 100644
--- a/paddler_download_manager/tests/reports_cache_permission_denied_when_the_cache_directory_is_read_only.rs
+++ b/paddler_download_manager/tests/reports_cache_permission_denied_when_the_cache_directory_is_read_only.rs
@@ -1,5 +1,3 @@
-#![cfg(unix)]
-
 use std::fs::Permissions;
 use std::io::ErrorKind;
 use std::os::unix::fs::PermissionsExt as _;
@@ -50,4 +48,10 @@ async fn reports_cache_permission_denied_when_the_cache_directory_is_read_only()
         recorded_download.result,
         Err(DownloadError::CachePermissionDenied { source, .. }) if source.kind() == ErrorKind::PermissionDenied
     ));
+
+    fixture
+        .stop()
+        .await
+        .expect("the fixture task must not panic")
+        .expect("the fixture must serve every request");
 }
diff --git a/paddler_download_manager/tests/reports_cache_permission_denied_when_the_partial_file_cannot_be_truncated.rs b/paddler_download_manager/tests/reports_cache_permission_denied_when_the_partial_file_cannot_be_truncated.rs
index 9f70aef3c..4b003e6f1 100644
--- a/paddler_download_manager/tests/reports_cache_permission_denied_when_the_partial_file_cannot_be_truncated.rs
+++ b/paddler_download_manager/tests/reports_cache_permission_denied_when_the_partial_file_cannot_be_truncated.rs
@@ -1,5 +1,3 @@
-#![cfg(unix)]
-
 use std::fs::Permissions;
 use std::os::unix::fs::PermissionsExt as _;
 
@@ -47,4 +45,10 @@ async fn reports_cache_permission_denied_when_the_partial_file_cannot_be_truncat
         recorded_download.result,
         Err(DownloadError::CachePermissionDenied { path, .. }) if path == partial_path
     ));
+
+    fixture
+        .stop()
+        .await
+        .expect("the fixture task must not panic")
+        .expect("the fixture must serve every request");
 }
diff --git a/paddler_download_manager/tests/reports_interrupted_when_the_body_stream_drops.rs b/paddler_download_manager/tests/reports_interrupted_when_the_body_stream_drops.rs
index 12b27c5a1..5b51b4085 100644
--- a/paddler_download_manager/tests/reports_interrupted_when_the_body_stream_drops.rs
+++ b/paddler_download_manager/tests/reports_interrupted_when_the_body_stream_drops.rs
@@ -44,4 +44,10 @@ async fn reports_interrupted_when_the_body_stream_drops() {
         recorded_download.progress.last(),
         Some(&DownloadProgress::Finished)
     );
+
+    fixture
+        .stop()
+        .await
+        .expect("the fixture task must not panic")
+        .expect("the fixture must serve every request");
 }
diff --git a/paddler_download_manager/tests/reports_io_error_when_the_cache_directory_cannot_be_created.rs b/paddler_download_manager/tests/reports_io_error_when_the_cache_directory_cannot_be_created.rs
index 56c283a8b..b82ab1feb 100644
--- a/paddler_download_manager/tests/reports_io_error_when_the_cache_directory_cannot_be_created.rs
+++ b/paddler_download_manager/tests/reports_io_error_when_the_cache_directory_cannot_be_created.rs
@@ -42,4 +42,10 @@ async fn reports_io_error_when_the_cache_directory_cannot_be_created() {
             if path == destination.with_extension("partial")
                 && source.kind() == ErrorKind::NotADirectory
     ));
+
+    fixture
+        .stop()
+        .await
+        .expect("the fixture task must not panic")
+        .expect("the fixture must serve every request");
 }
diff --git a/paddler_download_manager/tests/reports_io_error_when_the_destination_is_a_non_empty_directory.rs b/paddler_download_manager/tests/reports_io_error_when_the_destination_is_a_non_empty_directory.rs
index fd7ef20dc..07c145e7d 100644
--- a/paddler_download_manager/tests/reports_io_error_when_the_destination_is_a_non_empty_directory.rs
+++ b/paddler_download_manager/tests/reports_io_error_when_the_destination_is_a_non_empty_directory.rs
@@ -1,5 +1,3 @@
-#![cfg(unix)]
-
 use std::io::ErrorKind;
 
 use tempfile::TempDir;
@@ -46,4 +44,10 @@ async fn reports_io_error_when_the_destination_is_a_non_empty_directory() {
             if path == destination.with_extension("partial")
                 && source.kind() == ErrorKind::IsADirectory
     ));
+
+    fixture
+        .stop()
+        .await
+        .expect("the fixture task must not panic")
+        .expect("the fixture must serve every request");
 }
diff --git a/paddler_download_manager/tests/reports_not_found_for_404_without_retrying.rs b/paddler_download_manager/tests/reports_not_found_for_404_without_retrying.rs
index 435a6ade7..ea2b51659 100644
--- a/paddler_download_manager/tests/reports_not_found_for_404_without_retrying.rs
+++ b/paddler_download_manager/tests/reports_not_found_for_404_without_retrying.rs
@@ -4,6 +4,7 @@ use tokio_util::sync::CancellationToken;
 
 use paddler_download_manager::download_error::DownloadError;
 use paddler_download_manager::download_manager::DownloadManager;
+use paddler_download_manager::download_progress::DownloadProgress;
 use paddler_local_http_fixture::fixture_response::FixtureResponse;
 use paddler_local_http_fixture::local_http_fixture::LocalHttpFixture;
 
@@ -31,5 +32,11 @@ async fn reports_not_found_for_404_without_retrying() {
         Err(DownloadError::NotFound { url }) if url == fixture.url("/model.gguf")
     ));
     assert_eq!(fixture.request_count(), 1);
-    assert!(recorded_download.progress.is_empty());
+    assert_eq!(recorded_download.progress, Vec::<DownloadProgress>::new());
+
+    fixture
+        .stop()
+        .await
+        .expect("the fixture task must not panic")
+        .expect("the fixture must serve every request");
 }
diff --git a/paddler_download_manager/tests/reports_permission_denied_without_retrying.rs b/paddler_download_manager/tests/reports_permission_denied_without_retrying.rs
index 5bd0b003b..cf014670c 100644
--- a/paddler_download_manager/tests/reports_permission_denied_without_retrying.rs
+++ b/paddler_download_manager/tests/reports_permission_denied_without_retrying.rs
@@ -4,6 +4,7 @@ use tokio_util::sync::CancellationToken;
 
 use paddler_download_manager::download_error::DownloadError;
 use paddler_download_manager::download_manager::DownloadManager;
+use paddler_download_manager::download_progress::DownloadProgress;
 use paddler_local_http_fixture::fixture_response::FixtureResponse;
 use paddler_local_http_fixture::local_http_fixture::LocalHttpFixture;
 
@@ -31,5 +32,11 @@ async fn reports_permission_denied_without_retrying() {
         Err(DownloadError::PermissionDenied { status, .. }) if status == StatusCode::FORBIDDEN
     ));
     assert_eq!(fixture.request_count(), 1);
-    assert!(recorded_download.progress.is_empty());
+    assert_eq!(recorded_download.progress, Vec::<DownloadProgress>::new());
+
+    fixture
+        .stop()
+        .await
+        .expect("the fixture task must not panic")
+        .expect("the fixture must serve every request");
 }
diff --git a/paddler_download_manager/tests/reports_rejected_request_for_other_client_errors_without_retrying.rs b/paddler_download_manager/tests/reports_rejected_request_for_other_client_errors_without_retrying.rs
index 02b00cf37..70cadbc5b 100644
--- a/paddler_download_manager/tests/reports_rejected_request_for_other_client_errors_without_retrying.rs
+++ b/paddler_download_manager/tests/reports_rejected_request_for_other_client_errors_without_retrying.rs
@@ -4,6 +4,7 @@ use tokio_util::sync::CancellationToken;
 
 use paddler_download_manager::download_error::DownloadError;
 use paddler_download_manager::download_manager::DownloadManager;
+use paddler_download_manager::download_progress::DownloadProgress;
 use paddler_local_http_fixture::fixture_response::FixtureResponse;
 use paddler_local_http_fixture::local_http_fixture::LocalHttpFixture;
 
@@ -31,5 +32,11 @@ async fn reports_rejected_request_for_other_client_errors_without_retrying() {
         Err(DownloadError::DownloadServerRejectedRequest { status, .. }) if status == StatusCode::BAD_REQUEST
     ));
     assert_eq!(fixture.request_count(), 1);
-    assert!(recorded_download.progress.is_empty());
+    assert_eq!(recorded_download.progress, Vec::<DownloadProgress>::new());
+
+    fixture
+        .stop()
+        .await
+        .expect("the fixture task must not panic")
+        .expect("the fixture must serve every request");
 }
diff --git a/paddler_download_manager/tests/reports_server_errored_for_server_errors_without_retrying.rs b/paddler_download_manager/tests/reports_server_errored_for_server_errors_without_retrying.rs
index 4f8554a83..008429ad1 100644
--- a/paddler_download_manager/tests/reports_server_errored_for_server_errors_without_retrying.rs
+++ b/paddler_download_manager/tests/reports_server_errored_for_server_errors_without_retrying.rs
@@ -4,6 +4,7 @@ use tokio_util::sync::CancellationToken;
 
 use paddler_download_manager::download_error::DownloadError;
 use paddler_download_manager::download_manager::DownloadManager;
+use paddler_download_manager::download_progress::DownloadProgress;
 use paddler_local_http_fixture::fixture_response::FixtureResponse;
 use paddler_local_http_fixture::local_http_fixture::LocalHttpFixture;
 
@@ -32,5 +33,11 @@ async fn reports_server_errored_for_server_errors_without_retrying() {
         Err(DownloadError::DownloadServerErrored { status, .. }) if status == StatusCode::INTERNAL_SERVER_ERROR
     ));
     assert_eq!(fixture.request_count(), 1);
-    assert!(recorded_download.progress.is_empty());
+    assert_eq!(recorded_download.progress, Vec::<DownloadProgress>::new());
+
+    fixture
+        .stop()
+        .await
+        .expect("the fixture task must not panic")
+        .expect("the fixture must serve every request");
 }
diff --git a/paddler_download_manager/tests/reports_stale_partial_for_416_and_removes_the_partial_file.rs b/paddler_download_manager/tests/reports_stale_partial_for_416_and_removes_the_partial_file.rs
index a4e606c68..a31f2083e 100644
--- a/paddler_download_manager/tests/reports_stale_partial_for_416_and_removes_the_partial_file.rs
+++ b/paddler_download_manager/tests/reports_stale_partial_for_416_and_removes_the_partial_file.rs
@@ -45,4 +45,10 @@ async fn reports_stale_partial_for_416_and_removes_the_partial_file() {
             .await
             .expect("the partial path must be inspectable")
     );
+
+    fixture
+        .stop()
+        .await
+        .expect("the fixture task must not panic")
+        .expect("the fixture must serve every request");
 }
diff --git a/paddler_download_manager/tests/reports_stale_partial_for_416_without_a_partial_file.rs b/paddler_download_manager/tests/reports_stale_partial_for_416_without_a_partial_file.rs
index 79c604a9b..c6f77e70e 100644
--- a/paddler_download_manager/tests/reports_stale_partial_for_416_without_a_partial_file.rs
+++ b/paddler_download_manager/tests/reports_stale_partial_for_416_without_a_partial_file.rs
@@ -40,4 +40,10 @@ async fn reports_stale_partial_for_416_without_a_partial_file() {
             .await
             .expect("the partial path must be inspectable")
     );
+
+    fixture
+        .stop()
+        .await
+        .expect("the fixture task must not panic")
+        .expect("the fixture must serve every request");
 }
diff --git a/paddler_download_manager/tests/reports_unreachable_when_the_server_stalls_before_headers.rs b/paddler_download_manager/tests/reports_unreachable_when_the_server_stalls_before_headers.rs
index 396a974a4..bf0a470a6 100644
--- a/paddler_download_manager/tests/reports_unreachable_when_the_server_stalls_before_headers.rs
+++ b/paddler_download_manager/tests/reports_unreachable_when_the_server_stalls_before_headers.rs
@@ -3,6 +3,7 @@ use tokio_util::sync::CancellationToken;
 
 use paddler_download_manager::download_error::DownloadError;
 use paddler_download_manager::download_manager::DownloadManager;
+use paddler_download_manager::download_progress::DownloadProgress;
 use paddler_local_http_fixture::fixture_response::FixtureResponse;
 use paddler_local_http_fixture::local_http_fixture::LocalHttpFixture;
 
@@ -32,5 +33,11 @@ async fn reports_unreachable_when_the_server_stalls_before_headers() {
         Err(DownloadError::DownloadServerIsUnreachable { url: reported_url, .. })
             if reported_url == url
     ));
-    assert!(recorded_download.progress.is_empty());
+    assert_eq!(recorded_download.progress, Vec::<DownloadProgress>::new());
+
+    fixture
+        .stop()
+        .await
+        .expect("the fixture task must not panic")
+        .expect("the fixture must serve every request");
 }
diff --git a/paddler_download_manager/tests/resumes_a_partial_file_with_a_range_request.rs b/paddler_download_manager/tests/resumes_a_partial_file_with_a_range_request.rs
index c4ea7857c..7b9753998 100644
--- a/paddler_download_manager/tests/resumes_a_partial_file_with_a_range_request.rs
+++ b/paddler_download_manager/tests/resumes_a_partial_file_with_a_range_request.rs
@@ -55,4 +55,10 @@ async fn resumes_a_partial_file_with_a_range_request() {
             total_bytes: Some(22),
         })
     );
+
+    fixture
+        .stop()
+        .await
+        .expect("the fixture task must not panic")
+        .expect("the fixture must serve every request");
 }
diff --git a/paddler_download_manager/tests/returns_cancelled_without_a_request_for_a_cancelled_token.rs b/paddler_download_manager/tests/returns_cancelled_without_a_request_for_a_cancelled_token.rs
index 2ed5a5a35..fe6fa5b53 100644
--- a/paddler_download_manager/tests/returns_cancelled_without_a_request_for_a_cancelled_token.rs
+++ b/paddler_download_manager/tests/returns_cancelled_without_a_request_for_a_cancelled_token.rs
@@ -42,4 +42,10 @@ async fn returns_cancelled_without_a_request_for_a_cancelled_token() {
             .expect("the destination must be inspectable")
     );
     assert_eq!(fixture.request_count(), 0);
+
+    fixture
+        .stop()
+        .await
+        .expect("the fixture task must not panic")
+        .expect("the fixture must serve every request");
 }
diff --git a/paddler_download_manager/tests/sends_no_range_header_for_a_fresh_download.rs b/paddler_download_manager/tests/sends_no_range_header_for_a_fresh_download.rs
index f8f30f980..de07a2585 100644
--- a/paddler_download_manager/tests/sends_no_range_header_for_a_fresh_download.rs
+++ b/paddler_download_manager/tests/sends_no_range_header_for_a_fresh_download.rs
@@ -28,4 +28,10 @@ async fn sends_no_range_header_for_a_fresh_download() {
 
     assert_eq!(fixture.request_count(), 1);
     assert_eq!(fixture.last_range_header(), None);
+
+    fixture
+        .stop()
+        .await
+        .expect("the fixture task must not panic")
+        .expect("the fixture must serve every request");
 }
diff --git a/paddler_download_manager/tests/starts_over_when_the_server_ignores_the_range_request.rs b/paddler_download_manager/tests/starts_over_when_the_server_ignores_the_range_request.rs
index a4f4b3b23..a51a36521 100644
--- a/paddler_download_manager/tests/starts_over_when_the_server_ignores_the_range_request.rs
+++ b/paddler_download_manager/tests/starts_over_when_the_server_ignores_the_range_request.rs
@@ -55,4 +55,10 @@ async fn starts_over_when_the_server_ignores_the_range_request() {
             total_bytes: Some(body.len() as u64),
         })
     );
+
+    fixture
+        .stop()
+        .await
+        .expect("the fixture task must not panic")
+        .expect("the fixture must serve every request");
 }
diff --git a/paddler_download_manager/tests/treats_a_mismatched_content_range_as_a_stale_partial_file.rs b/paddler_download_manager/tests/treats_a_mismatched_content_range_as_a_stale_partial_file.rs
index d0c6f920a..570bc1a49 100644
--- a/paddler_download_manager/tests/treats_a_mismatched_content_range_as_a_stale_partial_file.rs
+++ b/paddler_download_manager/tests/treats_a_mismatched_content_range_as_a_stale_partial_file.rs
@@ -5,6 +5,7 @@ use tokio_util::sync::CancellationToken;
 
 use paddler_download_manager::download_error::DownloadError;
 use paddler_download_manager::download_manager::DownloadManager;
+use paddler_download_manager::download_progress::DownloadProgress;
 use paddler_local_http_fixture::fixture_response::FixtureResponse;
 use paddler_local_http_fixture::local_http_fixture::LocalHttpFixture;
 
@@ -46,5 +47,11 @@ async fn treats_a_mismatched_content_range_as_a_stale_partial_file() {
             .await
             .expect("the partial path must be inspectable")
     );
-    assert!(recorded_download.progress.is_empty());
+    assert_eq!(recorded_download.progress, Vec::<DownloadProgress>::new());
+
+    fixture
+        .stop()
+        .await
+        .expect("the fixture task must not panic")
+        .expect("the fixture must serve every request");
 }
diff --git a/paddler_frame_writer/Cargo.toml b/paddler_frame_writer/Cargo.toml
new file mode 100644
index 000000000..f221218e2
--- /dev/null
+++ b/paddler_frame_writer/Cargo.toml
@@ -0,0 +1,20 @@
+[package]
+name = "paddler_frame_writer"
+authors.workspace = true
+description.workspace = true
+edition.workspace = true
+homepage.workspace = true
+license.workspace = true
+repository.workspace = true
+version.workspace = true
+
+[dependencies]
+futures-util = { workspace = true }
+tokio = { workspace = true }
+tokio-util = { workspace = true }
+
+[dev-dependencies]
+futures = { workspace = true }
+
+[lints]
+workspace = true
diff --git a/paddler_frame_writer/src/frame_writing_end.rs b/paddler_frame_writer/src/frame_writing_end.rs
new file mode 100644
index 000000000..0bdf3ff7d
--- /dev/null
+++ b/paddler_frame_writer/src/frame_writing_end.rs
@@ -0,0 +1,6 @@
+#[derive(Debug, Eq, PartialEq)]
+pub enum FrameWritingEnd<TSendError> {
+    ConnectionClosed,
+    EveryFrameSent,
+    SendFailed(TSendError),
+}
diff --git a/paddler_frame_writer/src/lib.rs b/paddler_frame_writer/src/lib.rs
new file mode 100644
index 000000000..bec780b40
--- /dev/null
+++ b/paddler_frame_writer/src/lib.rs
@@ -0,0 +1,2 @@
+pub mod frame_writing_end;
+pub mod write_frames_until_connection_closes;
diff --git a/paddler_frame_writer/src/write_frames_until_connection_closes.rs b/paddler_frame_writer/src/write_frames_until_connection_closes.rs
new file mode 100644
index 000000000..2f8b2dc27
--- /dev/null
+++ b/paddler_frame_writer/src/write_frames_until_connection_closes.rs
@@ -0,0 +1,90 @@
+use futures_util::Sink;
+use futures_util::SinkExt as _;
+use tokio::select;
+use tokio::sync::mpsc::UnboundedReceiver;
+use tokio_util::sync::CancellationToken;
+
+use crate::frame_writing_end::FrameWritingEnd;
+
+pub async fn write_frames_until_connection_closes<TFrame, TSink>(
+    connection_close: CancellationToken,
+    mut outbound_rx: UnboundedReceiver<TFrame>,
+    mut sink: TSink,
+) -> FrameWritingEnd<TSink::Error>
+where
+    TSink: Sink<TFrame> + Unpin,
+{
+    while let Some(frame) = outbound_rx.recv().await {
+        select! {
+            biased;
+            send_outcome = sink.send(frame) => {
+                if let Err(send_error) = send_outcome {
+                    return FrameWritingEnd::SendFailed(send_error);
+                }
+            }
+            () = connection_close.cancelled() => return FrameWritingEnd::ConnectionClosed,
+        }
+    }
+
+    FrameWritingEnd::EveryFrameSent
+}
+
+#[cfg(test)]
+mod tests {
+    use futures::channel::mpsc as sink_channel;
+    use tokio::spawn;
+    use tokio::sync::mpsc;
+    use tokio_util::sync::CancellationToken;
+
+    use super::write_frames_until_connection_closes;
+    use crate::frame_writing_end::FrameWritingEnd;
+
+    #[tokio::test]
+    async fn a_send_its_peer_never_completes_is_abandoned_once_the_connection_closes() {
+        let (sink, _undrained_sink_receiver) = sink_channel::channel::<u32>(0);
+        let (outbound_tx, outbound_rx) = mpsc::unbounded_channel();
+        let connection_close = CancellationToken::new();
+        let writer = spawn(write_frames_until_connection_closes(
+            connection_close.clone(),
+            outbound_rx,
+            sink,
+        ));
+
+        outbound_tx.send(1).unwrap();
+        outbound_tx.send(2).unwrap();
+        connection_close.cancel();
+
+        assert_eq!(writer.await.unwrap(), FrameWritingEnd::ConnectionClosed);
+    }
+
+    #[tokio::test]
+    async fn a_frame_the_sink_rejects_stops_the_writer() {
+        let (sink, sink_receiver) = sink_channel::unbounded::<u32>();
+        let (outbound_tx, outbound_rx) = mpsc::unbounded_channel();
+
+        drop(sink_receiver);
+        outbound_tx.send(1).unwrap();
+
+        assert!(matches!(
+            write_frames_until_connection_closes(CancellationToken::new(), outbound_rx, sink).await,
+            FrameWritingEnd::SendFailed(send_error) if send_error.is_disconnected()
+        ));
+    }
+
+    #[tokio::test]
+    async fn a_writer_sends_every_queued_frame_until_its_queue_closes() {
+        let (sink, mut sink_receiver) = sink_channel::unbounded::<u32>();
+        let (outbound_tx, outbound_rx) = mpsc::unbounded_channel();
+
+        outbound_tx.send(1).unwrap();
+        outbound_tx.send(2).unwrap();
+        drop(outbound_tx);
+
+        assert_eq!(
+            write_frames_until_connection_closes(CancellationToken::new(), outbound_rx, sink).await,
+            FrameWritingEnd::EveryFrameSent
+        );
+        assert_eq!(sink_receiver.try_recv().unwrap(), 1);
+        assert_eq!(sink_receiver.try_recv().unwrap(), 2);
+    }
+}
diff --git a/paddler_gui/Cargo.toml b/paddler_gui/Cargo.toml
deleted file mode 100644
index cec4ae721..000000000
--- a/paddler_gui/Cargo.toml
+++ /dev/null
@@ -1,53 +0,0 @@
-[package]
-name = "paddler_gui"
-version.workspace = true
-edition.workspace = true
-authors.workspace = true
-description = "Paddler desktop application"
-homepage.workspace = true
-license.workspace = true
-publish = false
-repository.workspace = true
-
-[lib]
-name = "paddler_gui"
-path = "src/lib.rs"
-
-[[bin]]
-name = "paddler_gui"
-path = "src/main.rs"
-
-[dependencies]
-async-stream = { workspace = true }
-clap = { workspace = true }
-command_handler = { workspace = true }
-env_logger = { workspace = true }
-iced = { workspace = true }
-if-addrs = { workspace = true }
-log = { workspace = true }
-open = { workspace = true }
-paddler_balancer = { workspace = true }
-paddler_bootstrap = { workspace = true }
-paddler_inference_parameters = { workspace = true }
-paddler_messaging = { workspace = true }
-paddler_state_database = { workspace = true }
-statum = { workspace = true }
-thiserror = { workspace = true }
-tokio = { workspace = true }
-tokio-util = { workspace = true }
-trzcina = { workspace = true }
-
-[dev-dependencies]
-paddler_test_cluster_harness = { workspace = true }
-
-[lints]
-workspace = true
-
-[features]
-default = []
-cuda = ["paddler_bootstrap/cuda"]
-metal = ["paddler_bootstrap/metal"]
-web_admin_panel = [
-    "paddler_balancer/web_admin_panel",
-    "paddler_bootstrap/web_admin_panel",
-]
diff --git a/paddler_gui/shell.nix b/paddler_gui/shell.nix
deleted file mode 100644
index 965a896f0..000000000
--- a/paddler_gui/shell.nix
+++ /dev/null
@@ -1,27 +0,0 @@
-{ pkgs ? import <nixpkgs> {} }:
-
-let
-  icedRuntimeLibs = with pkgs; [
-    vulkan-loader
-    libxkbcommon
-    wayland
-    libGL
-    libx11
-    libxcursor
-    libxrandr
-    libxi
-    libxcb
-  ];
-in
-pkgs.mkShell {
-  nativeBuildInputs = with pkgs; [
-    pkg-config
-  ];
-
-  buildInputs = icedRuntimeLibs ++ (with pkgs; [
-    fontconfig
-    xvfb
-  ]);
-
-  LD_LIBRARY_PATH = pkgs.lib.makeLibraryPath icedRuntimeLibs;
-}
diff --git a/paddler_gui/src/address_placeholder.rs b/paddler_gui/src/address_placeholder.rs
deleted file mode 100644
index 2afd5b1b6..000000000
--- a/paddler_gui/src/address_placeholder.rs
+++ /dev/null
@@ -1 +0,0 @@
-pub const ADDRESS_PLACEHOLDER: &str = "IP:port";
diff --git a/paddler_gui/src/agent_runner_messages.rs b/paddler_gui/src/agent_runner_messages.rs
deleted file mode 100644
index df62e19c7..000000000
--- a/paddler_gui/src/agent_runner_messages.rs
+++ /dev/null
@@ -1,61 +0,0 @@
-use std::sync::Arc;
-
-use async_stream::stream;
-use iced::futures::Stream;
-use tokio::pin;
-use tokio::select;
-
-use paddler_bootstrap::agent_runner::AgentRunner;
-use paddler_bootstrap::agent_runner_params::AgentRunnerParams;
-use paddler_messaging::produces_snapshot::ProducesSnapshot as _;
-use paddler_messaging::subscribes_to_updates::SubscribesToUpdates as _;
-
-use crate::agent_running_message::AgentRunningMessage;
-use crate::message::Message;
-
-pub fn agent_runner_messages(
-    agent_runner_params: AgentRunnerParams,
-) -> impl Stream<Item = Message> {
-    stream! {
-        let runner = AgentRunner::start(agent_runner_params);
-        let slot_aggregated_status = runner.slot_aggregated_status.clone();
-        let mut balancer_connection_rx = runner.balancer_connection_rx.clone();
-        let mut status_update_rx = slot_aggregated_status.subscribe_to_updates();
-        let completion = runner.wait_for_completion();
-
-        pin!(completion);
-
-        let agent_status_updated = || {
-            Message::AgentRunning(AgentRunningMessage::AgentStatusUpdated {
-                slots_processing: slot_aggregated_status.slots_processing_count(),
-                status: slot_aggregated_status.make_snapshot().status,
-            })
-        };
-
-        let initial_balancer_connection = *balancer_connection_rx.borrow_and_update();
-
-        yield agent_status_updated();
-        yield Message::AgentRunning(AgentRunningMessage::BalancerConnectionChanged(
-            initial_balancer_connection,
-        ));
-
-        let completion_result = loop {
-            let message = select! {
-                Ok(()) = status_update_rx.changed() => agent_status_updated(),
-                Ok(()) = balancer_connection_rx.changed() => Message::AgentRunning(
-                    AgentRunningMessage::BalancerConnectionChanged(
-                        *balancer_connection_rx.borrow_and_update(),
-                    ),
-                ),
-                completion_result = &mut completion => break completion_result,
-            };
-
-            yield message;
-        };
-
-        yield match completion_result {
-            Ok(()) => Message::AgentStopped,
-            Err(error) => Message::AgentFailed(Arc::new(error)),
-        };
-    }
-}
diff --git a/paddler_gui/src/agent_running_action.rs b/paddler_gui/src/agent_running_action.rs
deleted file mode 100644
index eeda04d1b..000000000
--- a/paddler_gui/src/agent_running_action.rs
+++ /dev/null
@@ -1,5 +0,0 @@
-#[derive(Debug, Eq, PartialEq)]
-pub enum AgentRunningAction {
-    None,
-    Disconnect,
-}
diff --git a/paddler_gui/src/agent_running_data.rs b/paddler_gui/src/agent_running_data.rs
deleted file mode 100644
index e582c8b3b..000000000
--- a/paddler_gui/src/agent_running_data.rs
+++ /dev/null
@@ -1,123 +0,0 @@
-use tokio_util::sync::CancellationToken;
-
-use paddler_messaging::agent_status::AgentStatus;
-use paddler_messaging::balancer_connection::BalancerConnection;
-
-use crate::agent_running_action::AgentRunningAction;
-use crate::agent_running_message::AgentRunningMessage;
-
-pub struct AgentRunningData {
-    pub balancer_address: String,
-    pub balancer_connection: BalancerConnection,
-    pub cancellation_token: CancellationToken,
-    pub name: Option<String>,
-    pub slots_processing: u64,
-    pub status: AgentStatus,
-}
-
-impl AgentRunningData {
-    pub fn update(&mut self, message: AgentRunningMessage) -> AgentRunningAction {
-        match message {
-            AgentRunningMessage::AgentStatusUpdated {
-                slots_processing,
-                status,
-            } => {
-                self.slots_processing = slots_processing;
-                self.status = status;
-
-                AgentRunningAction::None
-            }
-            AgentRunningMessage::BalancerConnectionChanged(balancer_connection) => {
-                self.balancer_connection = balancer_connection;
-
-                AgentRunningAction::None
-            }
-            AgentRunningMessage::Disconnect => {
-                self.cancellation_token.cancel();
-
-                AgentRunningAction::Disconnect
-            }
-        }
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use tokio_util::sync::CancellationToken;
-
-    use paddler_messaging::agent_state_application_status::AgentStateApplicationStatus;
-    use paddler_messaging::agent_status::AgentStatus;
-    use paddler_messaging::balancer_connection::BalancerConnection;
-
-    use super::AgentRunningData;
-    use crate::agent_running_action::AgentRunningAction;
-    use crate::agent_running_message::AgentRunningMessage;
-
-    const AGENT_NAME: &str = "gpu-box";
-
-    fn agent_waiting_for_its_first_status() -> AgentRunningData {
-        AgentRunningData {
-            balancer_address: "127.0.0.1:8060".to_owned(),
-            balancer_connection: BalancerConnection::Connecting,
-            cancellation_token: CancellationToken::new(),
-            name: Some(AGENT_NAME.to_owned()),
-            slots_processing: 0,
-            status: AgentStatus::default(),
-        }
-    }
-
-    #[test]
-    fn a_reported_status_updates_the_agent_under_its_own_name() {
-        let mut agent = agent_waiting_for_its_first_status();
-        let reported_status = AgentStatus {
-            desired_slots_total: 4,
-            model_path: Some("/models/qwen3.gguf".to_owned()),
-            slots_total: 4,
-            state_application_status: AgentStateApplicationStatus::Applied,
-            ..AgentStatus::default()
-        };
-
-        let action = agent.update(AgentRunningMessage::AgentStatusUpdated {
-            slots_processing: 1,
-            status: reported_status.clone(),
-        });
-
-        assert_eq!(action, AgentRunningAction::None);
-        assert_eq!(agent.name.as_deref(), Some(AGENT_NAME));
-        assert_eq!(agent.slots_processing, 1);
-        assert_eq!(agent.status, reported_status);
-    }
-
-    #[test]
-    fn a_reported_status_alone_does_not_mark_the_agent_connected() {
-        let mut agent = agent_waiting_for_its_first_status();
-
-        agent.update(AgentRunningMessage::AgentStatusUpdated {
-            slots_processing: 0,
-            status: AgentStatus::default(),
-        });
-
-        assert_eq!(agent.balancer_connection, BalancerConnection::Connecting);
-    }
-
-    #[test]
-    fn an_opened_balancer_connection_marks_the_agent_connected() {
-        let mut agent = agent_waiting_for_its_first_status();
-
-        agent.update(AgentRunningMessage::BalancerConnectionChanged(
-            BalancerConnection::Connected,
-        ));
-
-        assert_eq!(agent.balancer_connection, BalancerConnection::Connected);
-    }
-
-    #[test]
-    fn disconnecting_cancels_the_running_agent() {
-        let mut agent = agent_waiting_for_its_first_status();
-
-        let action = agent.update(AgentRunningMessage::Disconnect);
-
-        assert_eq!(action, AgentRunningAction::Disconnect);
-        assert!(agent.cancellation_token.is_cancelled());
-    }
-}
diff --git a/paddler_gui/src/agent_running_message.rs b/paddler_gui/src/agent_running_message.rs
deleted file mode 100644
index 0484793e5..000000000
--- a/paddler_gui/src/agent_running_message.rs
+++ /dev/null
@@ -1,12 +0,0 @@
-use paddler_messaging::agent_status::AgentStatus;
-use paddler_messaging::balancer_connection::BalancerConnection;
-
-#[derive(Debug, Clone)]
-pub enum AgentRunningMessage {
-    AgentStatusUpdated {
-        slots_processing: u64,
-        status: AgentStatus,
-    },
-    BalancerConnectionChanged(BalancerConnection),
-    Disconnect,
-}
diff --git a/paddler_gui/src/app.rs b/paddler_gui/src/app.rs
deleted file mode 100644
index 7190545c5..000000000
--- a/paddler_gui/src/app.rs
+++ /dev/null
@@ -1,340 +0,0 @@
-use std::mem;
-use std::sync::LazyLock;
-
-use async_stream::stream;
-use command_handler::shutdown_signal::register_shutdown_signals;
-use iced::Bottom;
-use iced::Center;
-use iced::Element;
-use iced::Fill;
-use iced::Right;
-use iced::Subscription;
-use iced::Task;
-use iced::clipboard::write;
-use iced::exit;
-use iced::futures::Stream;
-use iced::keyboard::Event as KeyboardEvent;
-use iced::keyboard::Key;
-use iced::keyboard::key::Named;
-use iced::keyboard::listen;
-use iced::widget::column;
-use iced::widget::container;
-use iced::widget::image;
-use iced::widget::image::Handle as ImageHandle;
-use iced::widget::operation;
-use iced::widget::stack;
-use iced::window;
-use log::error;
-use log::info;
-use log::warn;
-use open::that;
-use tokio_util::sync::CancellationToken;
-use trzcina::ServiceShutdownOptions;
-
-use paddler_bootstrap::agent_bootstrap_config::AgentBootstrapConfig;
-use paddler_bootstrap::agent_runner_params::AgentRunnerParams;
-use paddler_bootstrap::balancer_runner_params::BalancerRunnerParams;
-
-use crate::agent_runner_messages::agent_runner_messages;
-use crate::agent_running_action::AgentRunningAction;
-use crate::balancer_launch::BalancerLaunch;
-use crate::balancer_runner_messages::balancer_runner_messages;
-use crate::current_screen::CurrentScreen;
-use crate::home_message::HomeMessage;
-use crate::join_balancer_form_action::JoinBalancerFormAction;
-use crate::message::Message;
-use crate::running_balancer_action::RunningBalancerAction;
-use crate::screen::AgentRunning;
-use crate::screen::Screen;
-use crate::start_balancer_form_action::StartBalancerFormAction;
-use crate::ui::variables::SPACING_2X;
-use crate::ui::variables::SPACING_BASE;
-
-static BETA_IMAGE: LazyLock<ImageHandle> = LazyLock::new(|| {
-    ImageHandle::from_bytes(include_bytes!("../../resources/images/beta.png").as_slice())
-});
-
-fn shutdown_signal_stream() -> impl Stream<Item = Message> {
-    stream! {
-        match register_shutdown_signals() {
-            Ok(shutdown_signals) => match shutdown_signals.wait().await {
-                Ok(()) => yield Message::Quit,
-                Err(error) => error!("shutdown signal listener failed: {error}"),
-            },
-            Err(error) => error!("failed to register shutdown signal handlers: {error}"),
-        }
-    }
-}
-
-pub struct App {
-    screen: CurrentScreen,
-    shutdown: CancellationToken,
-}
-
-impl App {
-    pub fn new() -> (Self, Task<Message>) {
-        let app = Self {
-            screen: CurrentScreen::default(),
-            shutdown: CancellationToken::new(),
-        };
-
-        (app, Task::done(Message::IcedEventLoopReady))
-    }
-
-    pub fn update(&mut self, message: Message) -> Task<Message> {
-        let screen = mem::take(&mut self.screen);
-
-        match (screen, message) {
-            (screen, Message::IcedEventLoopReady) => {
-                info!("paddler_gui: iced event loop ready");
-                self.screen = screen;
-
-                Task::none()
-            }
-            (_, Message::Quit) => {
-                self.shutdown.cancel();
-
-                exit()
-            }
-            (CurrentScreen::Home(home), Message::Home(HomeMessage::StartBalancer)) => {
-                self.screen = CurrentScreen::StartBalancerForm(home.start_balancer());
-
-                Task::none()
-            }
-            (CurrentScreen::Home(home), Message::Home(HomeMessage::JoinBalancer)) => {
-                self.screen = CurrentScreen::JoinBalancerForm(home.join_balancer());
-
-                Task::none()
-            }
-            (CurrentScreen::JoinBalancerForm(mut form), Message::JoinBalancerForm(msg)) => {
-                let action = form.state_data.update(msg);
-
-                match action {
-                    JoinBalancerFormAction::None => {
-                        self.screen = CurrentScreen::JoinBalancerForm(form);
-
-                        Task::none()
-                    }
-                    JoinBalancerFormAction::Cancel => {
-                        self.screen = CurrentScreen::Home(form.cancel());
-
-                        Task::none()
-                    }
-                    JoinBalancerFormAction::ConnectAgent(agent_bootstrap_config) => self
-                        .spawn_agent(
-                            form.connect(self.shutdown.child_token()),
-                            agent_bootstrap_config,
-                        ),
-                }
-            }
-            (CurrentScreen::StartBalancerForm(mut form), Message::StartBalancerForm(msg)) => {
-                let action = form.state_data.update(msg);
-
-                match action {
-                    StartBalancerFormAction::None => {
-                        self.screen = CurrentScreen::StartBalancerForm(form);
-
-                        Task::none()
-                    }
-                    StartBalancerFormAction::Cancel => {
-                        if let BalancerLaunch::Starting(cancellation_token) =
-                            &form.state_data.launch
-                        {
-                            cancellation_token.cancel();
-                        }
-                        self.screen = CurrentScreen::Home(form.cancel());
-
-                        Task::none()
-                    }
-                    StartBalancerFormAction::StartBalancer(bootstrap_config) => {
-                        let cancellation_token = self.shutdown.child_token();
-
-                        form.state_data.launch =
-                            BalancerLaunch::Starting(cancellation_token.clone());
-                        self.screen = CurrentScreen::StartBalancerForm(form);
-
-                        Task::stream(balancer_runner_messages(BalancerRunnerParams {
-                            bootstrap_config: *bootstrap_config,
-                            cancellation_token,
-                            shutdown_options: ServiceShutdownOptions::default(),
-                        }))
-                    }
-                }
-            }
-            (
-                CurrentScreen::StartBalancerForm(form),
-                Message::BalancerStarted {
-                    addresses,
-                    cancellation_token,
-                    snapshot,
-                },
-            ) => {
-                self.screen = CurrentScreen::RunningBalancer(form.balancer_started(
-                    addresses,
-                    cancellation_token,
-                    snapshot,
-                ));
-
-                Task::none()
-            }
-            (CurrentScreen::StartBalancerForm(form), Message::BalancerFailed(error)) => {
-                error!("Balancer failed to start: {error}");
-                self.screen = CurrentScreen::Home(form.balancer_failed(error));
-
-                Task::none()
-            }
-            (CurrentScreen::RunningBalancer(mut running), Message::RunningBalancer(msg)) => {
-                let action = running.state_data.update(msg);
-
-                match action {
-                    RunningBalancerAction::None => {
-                        self.screen = CurrentScreen::RunningBalancer(running);
-
-                        Task::none()
-                    }
-                    RunningBalancerAction::CopyToClipboard(content) => {
-                        self.screen = CurrentScreen::RunningBalancer(running);
-
-                        write::<Message>(content).discard()
-                    }
-                    RunningBalancerAction::OpenUrl(url) => {
-                        self.screen = CurrentScreen::RunningBalancer(running);
-
-                        if let Err(error) = that(&url) {
-                            error!("Failed to open URL {url}: {error}");
-                        }
-
-                        Task::none()
-                    }
-                }
-            }
-            (CurrentScreen::RunningBalancer(running), Message::BalancerStopped) => {
-                self.screen = CurrentScreen::Home(running.balancer_stopped());
-
-                Task::none()
-            }
-            (CurrentScreen::RunningBalancer(running), Message::BalancerFailed(error)) => {
-                error!("Balancer failed unexpectedly: {error}");
-                self.screen = CurrentScreen::Home(running.balancer_failed(error));
-
-                Task::none()
-            }
-            (CurrentScreen::AgentRunning(mut running), Message::AgentRunning(msg)) => {
-                let action = running.state_data.update(msg);
-
-                match action {
-                    AgentRunningAction::None => {
-                        self.screen = CurrentScreen::AgentRunning(running);
-
-                        Task::none()
-                    }
-                    AgentRunningAction::Disconnect => {
-                        self.screen = CurrentScreen::Home(running.disconnect());
-
-                        Task::none()
-                    }
-                }
-            }
-            (CurrentScreen::AgentRunning(running), Message::AgentStopped) => {
-                info!("Agent stopped");
-                self.screen = CurrentScreen::Home(running.disconnect());
-
-                Task::none()
-            }
-            (CurrentScreen::AgentRunning(running), Message::AgentFailed(error)) => {
-                error!("Agent failed: {error}");
-                self.screen = CurrentScreen::Home(running.agent_failed(error));
-
-                Task::none()
-            }
-            (screen, Message::TabPressed { shift }) => {
-                self.screen = screen;
-
-                if shift {
-                    operation::focus_previous()
-                } else {
-                    operation::focus_next()
-                }
-            }
-            (screen, message) => {
-                warn!("Unhandled message {message:?} for current screen");
-                self.screen = screen;
-
-                Task::none()
-            }
-        }
-    }
-
-    pub fn subscription(&self) -> Subscription<Message> {
-        Subscription::batch([
-            listen().filter_map(|event| match event {
-                KeyboardEvent::KeyPressed {
-                    key: Key::Named(Named::Tab),
-                    modifiers,
-                    ..
-                } => Some(Message::TabPressed {
-                    shift: modifiers.shift(),
-                }),
-                _ => None,
-            }),
-            window::close_requests().map(|_closed_window_id| Message::Quit),
-            Subscription::run(shutdown_signal_stream),
-        ])
-    }
-
-    pub fn view(&self) -> Element<'_, Message> {
-        let screen_content = match &self.screen {
-            CurrentScreen::AgentRunning(screen) => {
-                screen.state_data.view().map(Message::AgentRunning)
-            }
-            CurrentScreen::Home(screen) => screen.state_data.view().map(Message::Home),
-            CurrentScreen::JoinBalancerForm(screen) => {
-                screen.state_data.view().map(Message::JoinBalancerForm)
-            }
-            CurrentScreen::StartBalancerForm(screen) => {
-                screen.state_data.view().map(Message::StartBalancerForm)
-            }
-            CurrentScreen::RunningBalancer(screen) => {
-                screen.state_data.view().map(Message::RunningBalancer)
-            }
-        };
-
-        let content_column = column![screen_content]
-            .max_width(700)
-            .padding([SPACING_2X * 2.0, SPACING_BASE])
-            .spacing(SPACING_BASE)
-            .align_x(Center);
-
-        let base_view = container(content_column).center_x(Fill).height(Fill);
-
-        if matches!(self.screen, CurrentScreen::Home(_)) {
-            let beta_image = image(BETA_IMAGE.clone()).width(100).height(100);
-
-            let beta_overlay = container(beta_image)
-                .width(Fill)
-                .height(Fill)
-                .align_x(Right)
-                .align_y(Bottom);
-
-            stack![base_view, beta_overlay].into()
-        } else {
-            base_view.into()
-        }
-    }
-
-    fn spawn_agent(
-        &mut self,
-        screen: Screen<AgentRunning>,
-        bootstrap_config: AgentBootstrapConfig,
-    ) -> Task<Message> {
-        let cancellation_token = screen.state_data.cancellation_token.clone();
-
-        self.screen = CurrentScreen::AgentRunning(screen);
-
-        Task::stream(agent_runner_messages(AgentRunnerParams {
-            bootstrap_config,
-            cancellation_token,
-            shutdown_options: ServiceShutdownOptions::default(),
-        }))
-    }
-}
diff --git a/paddler_gui/src/balancer_launch.rs b/paddler_gui/src/balancer_launch.rs
deleted file mode 100644
index 02330756d..000000000
--- a/paddler_gui/src/balancer_launch.rs
+++ /dev/null
@@ -1,6 +0,0 @@
-use tokio_util::sync::CancellationToken;
-
-pub enum BalancerLaunch {
-    NotRequested,
-    Starting(CancellationToken),
-}
diff --git a/paddler_gui/src/balancer_runner_messages.rs b/paddler_gui/src/balancer_runner_messages.rs
deleted file mode 100644
index 6af5a86d9..000000000
--- a/paddler_gui/src/balancer_runner_messages.rs
+++ /dev/null
@@ -1,73 +0,0 @@
-use std::sync::Arc;
-
-use async_stream::stream;
-use iced::futures::Stream;
-use tokio::pin;
-use tokio::select;
-use tokio::sync::watch;
-
-use paddler_bootstrap::balancer_runner::BalancerRunner;
-use paddler_bootstrap::balancer_runner_params::BalancerRunnerParams;
-use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-use paddler_messaging::subscribes_to_updates::SubscribesToUpdates as _;
-
-use crate::message::Message;
-use crate::running_balancer_message::RunningBalancerMessage;
-use crate::running_balancer_snapshot::RunningBalancerSnapshot;
-
-pub fn balancer_runner_messages(
-    balancer_runner_params: BalancerRunnerParams,
-) -> impl Stream<Item = Message> {
-    stream! {
-        let cancellation_token = balancer_runner_params.cancellation_token.clone();
-        let runner = match BalancerRunner::start(balancer_runner_params).await {
-            Ok(runner) => runner,
-            Err(error) => {
-                yield Message::BalancerFailed(Arc::new(error));
-
-                return;
-            }
-        };
-        let addresses = runner.addresses;
-        let agent_controller_pool = runner.agent_controller_pool.clone();
-        let balancer_applicable_state_holder = runner.balancer_applicable_state_holder.clone();
-        let mut desired_state_rx = runner.balancer_desired_state_tx.subscribe();
-        let mut pool_update_rx = agent_controller_pool.subscribe_to_updates();
-        let mut applicable_state_update_rx = balancer_applicable_state_holder.subscribe_to_updates();
-        let completion = runner.wait_for_completion();
-
-        pin!(completion);
-
-        let current_snapshot = |desired_state_rx: &mut watch::Receiver<BalancerDesiredState>| {
-            Box::new(RunningBalancerSnapshot::build(
-                &agent_controller_pool,
-                &balancer_applicable_state_holder,
-                desired_state_rx.borrow_and_update().clone(),
-            ))
-        };
-
-        yield Message::BalancerStarted {
-            addresses,
-            cancellation_token,
-            snapshot: current_snapshot(&mut desired_state_rx),
-        };
-
-        let completion_result = loop {
-            select! {
-                Ok(()) = pool_update_rx.changed() => {}
-                Ok(()) = applicable_state_update_rx.changed() => {}
-                Ok(()) = desired_state_rx.changed() => {}
-                completion_result = &mut completion => break completion_result,
-            }
-
-            yield Message::RunningBalancer(RunningBalancerMessage::SnapshotUpdated(
-                current_snapshot(&mut desired_state_rx),
-            ));
-        };
-
-        yield match completion_result {
-            Ok(()) => Message::BalancerStopped,
-            Err(error) => Message::BalancerFailed(Arc::new(error)),
-        };
-    }
-}
diff --git a/paddler_gui/src/current_screen.rs b/paddler_gui/src/current_screen.rs
deleted file mode 100644
index 58c6d5fb6..000000000
--- a/paddler_gui/src/current_screen.rs
+++ /dev/null
@@ -1,25 +0,0 @@
-use crate::home_data::HomeData;
-use crate::screen::AgentRunning;
-use crate::screen::Home;
-use crate::screen::JoinBalancerForm;
-use crate::screen::RunningBalancer;
-use crate::screen::Screen;
-use crate::screen::StartBalancerForm;
-
-pub enum CurrentScreen {
-    AgentRunning(Screen<AgentRunning>),
-    Home(Screen<Home>),
-    JoinBalancerForm(Screen<JoinBalancerForm>),
-    StartBalancerForm(Screen<StartBalancerForm>),
-    RunningBalancer(Screen<RunningBalancer>),
-}
-
-impl Default for CurrentScreen {
-    fn default() -> Self {
-        Self::Home(
-            Screen::<Home>::builder()
-                .state_data(HomeData { error: None })
-                .build(),
-        )
-    }
-}
diff --git a/paddler_gui/src/detect_network_interfaces.rs b/paddler_gui/src/detect_network_interfaces.rs
deleted file mode 100644
index 6360515f6..000000000
--- a/paddler_gui/src/detect_network_interfaces.rs
+++ /dev/null
@@ -1,30 +0,0 @@
-use if_addrs::get_if_addrs;
-use log::error;
-
-use crate::network_interface_address::NetworkInterfaceAddress;
-
-#[must_use]
-pub fn detect_network_interfaces() -> Vec<NetworkInterfaceAddress> {
-    let interfaces = match get_if_addrs() {
-        Ok(interfaces) => interfaces,
-        Err(error) => {
-            error!("Failed to detect network interfaces: {error}");
-
-            return Vec::new();
-        }
-    };
-
-    interfaces
-        .into_iter()
-        .filter(|interface| !interface.is_loopback())
-        .filter(|interface| interface.ip().is_ipv4())
-        .map(|interface| {
-            let ip_address = interface.ip();
-
-            NetworkInterfaceAddress {
-                interface_name: interface.name,
-                ip_address,
-            }
-        })
-        .collect()
-}
diff --git a/paddler_gui/src/form_field_error.rs b/paddler_gui/src/form_field_error.rs
deleted file mode 100644
index 97434eca1..000000000
--- a/paddler_gui/src/form_field_error.rs
+++ /dev/null
@@ -1,40 +0,0 @@
-use std::io;
-use std::net::AddrParseError;
-use std::net::SocketAddr;
-use std::num::ParseIntError;
-
-use thiserror::Error;
-
-#[derive(Debug, Error)]
-pub enum FormFieldError {
-    #[error("Cannot bind to {addr}: {source}")]
-    AddressCannotBeBound {
-        addr: SocketAddr,
-        #[source]
-        source: io::Error,
-    },
-    #[error("Port {port} is already in use")]
-    AddressPortInUse { port: u16 },
-    #[error("Address is required.")]
-    AddressRequired,
-    #[error("Invalid address ({source}), expected format: IP:port")]
-    AddressUnparsable {
-        #[source]
-        source: AddrParseError,
-    },
-    #[error("Cluster address is required.")]
-    ClusterAddressRequired,
-    #[error("Please select a model.")]
-    ModelNotSelected,
-    #[error("Invalid number of slots.")]
-    SlotsInvalid {
-        #[source]
-        source: ParseIntError,
-    },
-    #[error("Invalid number of slots (the number should be greater than zero).")]
-    SlotsNotPositive,
-    #[error("Number of slots is required.")]
-    SlotsRequired,
-    #[error("Number of slots cannot exceed {limit}.")]
-    SlotsAboveLimit { limit: u16 },
-}
diff --git a/paddler_gui/src/home_data.rs b/paddler_gui/src/home_data.rs
deleted file mode 100644
index 1b7aa69d9..000000000
--- a/paddler_gui/src/home_data.rs
+++ /dev/null
@@ -1,7 +0,0 @@
-use std::sync::Arc;
-
-use paddler_bootstrap::bootstrap_error::BootstrapError;
-
-pub struct HomeData {
-    pub error: Option<Arc<BootstrapError>>,
-}
diff --git a/paddler_gui/src/home_message.rs b/paddler_gui/src/home_message.rs
deleted file mode 100644
index db652521f..000000000
--- a/paddler_gui/src/home_message.rs
+++ /dev/null
@@ -1,5 +0,0 @@
-#[derive(Debug, Clone, Copy)]
-pub enum HomeMessage {
-    JoinBalancer,
-    StartBalancer,
-}
diff --git a/paddler_gui/src/join_balancer_form_action.rs b/paddler_gui/src/join_balancer_form_action.rs
deleted file mode 100644
index c649201e2..000000000
--- a/paddler_gui/src/join_balancer_form_action.rs
+++ /dev/null
@@ -1,8 +0,0 @@
-use paddler_bootstrap::agent_bootstrap_config::AgentBootstrapConfig;
-
-#[derive(Debug, Eq, PartialEq)]
-pub enum JoinBalancerFormAction {
-    None,
-    Cancel,
-    ConnectAgent(AgentBootstrapConfig),
-}
diff --git a/paddler_gui/src/join_balancer_form_data.rs b/paddler_gui/src/join_balancer_form_data.rs
deleted file mode 100644
index da3538250..000000000
--- a/paddler_gui/src/join_balancer_form_data.rs
+++ /dev/null
@@ -1,260 +0,0 @@
-use std::net::SocketAddr;
-use std::num::IntErrorKind;
-
-use paddler_bootstrap::agent_bootstrap_config::AgentBootstrapConfig;
-use paddler_bootstrap::llama_cpp_max_sequences::LLAMA_CPP_MAX_SEQUENCES;
-
-use crate::form_field_error::FormFieldError;
-use crate::join_balancer_form_action::JoinBalancerFormAction;
-use crate::join_balancer_form_message::JoinBalancerFormMessage;
-
-#[derive(Default)]
-pub struct JoinBalancerFormData {
-    pub agent_name: String,
-    pub balancer_address: String,
-    pub balancer_address_error: Option<FormFieldError>,
-    pub slots_count: String,
-    pub slots_error: Option<FormFieldError>,
-}
-
-impl JoinBalancerFormData {
-    #[must_use]
-    pub fn entered_agent_name(&self) -> Option<String> {
-        (!self.agent_name.is_empty()).then(|| self.agent_name.clone())
-    }
-
-    pub fn update(&mut self, message: JoinBalancerFormMessage) -> JoinBalancerFormAction {
-        match message {
-            JoinBalancerFormMessage::SetAgentName(name) => {
-                self.agent_name = name;
-
-                JoinBalancerFormAction::None
-            }
-            JoinBalancerFormMessage::SetBalancerAddress(address) => {
-                self.balancer_address = address;
-                self.balancer_address_error = None;
-
-                JoinBalancerFormAction::None
-            }
-            JoinBalancerFormMessage::SetSlotsCount(slots) => {
-                if slots.is_empty() || slots.chars().all(|character| character.is_ascii_digit()) {
-                    self.slots_count = slots;
-                    self.slots_error = None;
-                }
-
-                JoinBalancerFormAction::None
-            }
-            JoinBalancerFormMessage::Connect => self.validate_and_connect(),
-            JoinBalancerFormMessage::Cancel => JoinBalancerFormAction::Cancel,
-        }
-    }
-
-    fn validated_balancer_address(&self) -> Result<String, FormFieldError> {
-        if self.balancer_address.is_empty() {
-            return Err(FormFieldError::ClusterAddressRequired);
-        }
-
-        self.balancer_address
-            .parse::<SocketAddr>()
-            .map(|_balancer_address| self.balancer_address.clone())
-            .map_err(|source| FormFieldError::AddressUnparsable { source })
-    }
-
-    fn validated_slots(&self) -> Result<u16, FormFieldError> {
-        match self.slots_count.parse::<u16>() {
-            Ok(0) => Err(FormFieldError::SlotsNotPositive),
-            Ok(slots) if slots <= LLAMA_CPP_MAX_SEQUENCES => Ok(slots),
-            Ok(_slots_above_limit) => Err(FormFieldError::SlotsAboveLimit {
-                limit: LLAMA_CPP_MAX_SEQUENCES,
-            }),
-            Err(source) => Err(match source.kind() {
-                IntErrorKind::Empty => FormFieldError::SlotsRequired,
-                IntErrorKind::PosOverflow => FormFieldError::SlotsAboveLimit {
-                    limit: LLAMA_CPP_MAX_SEQUENCES,
-                },
-                _other_kind => FormFieldError::SlotsInvalid { source },
-            }),
-        }
-    }
-
-    fn validate_and_connect(&mut self) -> JoinBalancerFormAction {
-        match (self.validated_balancer_address(), self.validated_slots()) {
-            (Ok(management_address), Ok(slots)) => {
-                JoinBalancerFormAction::ConnectAgent(AgentBootstrapConfig {
-                    agent_name: self.entered_agent_name(),
-                    management_address,
-                    slots,
-                })
-            }
-            (management_address, slots) => {
-                self.balancer_address_error = management_address.err();
-                self.slots_error = slots.err();
-
-                JoinBalancerFormAction::None
-            }
-        }
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use std::mem::discriminant;
-    use std::net::SocketAddr;
-
-    use paddler_bootstrap::agent_bootstrap_config::AgentBootstrapConfig;
-    use paddler_bootstrap::llama_cpp_max_sequences::LLAMA_CPP_MAX_SEQUENCES;
-
-    use super::JoinBalancerFormData;
-    use crate::form_field_error::FormFieldError;
-    use crate::join_balancer_form_action::JoinBalancerFormAction;
-    use crate::join_balancer_form_message::JoinBalancerFormMessage;
-
-    fn form_with(balancer_address: &str, slots_count: &str) -> JoinBalancerFormData {
-        JoinBalancerFormData {
-            balancer_address: balancer_address.to_owned(),
-            slots_count: slots_count.to_owned(),
-            ..JoinBalancerFormData::default()
-        }
-    }
-
-    #[test]
-    fn ignores_slot_input_that_is_not_a_number() {
-        let mut form = form_with("", "4");
-
-        form.update(JoinBalancerFormMessage::SetSlotsCount("4a".to_owned()));
-
-        assert_eq!(form.slots_count, "4");
-    }
-
-    #[test]
-    fn connects_with_the_entered_address_slots_and_agent_name() {
-        let mut form = form_with("127.0.0.1:8060", "4");
-
-        form.update(JoinBalancerFormMessage::SetAgentName("gpu-box".to_owned()));
-
-        assert_eq!(
-            form.update(JoinBalancerFormMessage::Connect),
-            JoinBalancerFormAction::ConnectAgent(AgentBootstrapConfig {
-                agent_name: Some("gpu-box".to_owned()),
-                management_address: "127.0.0.1:8060".to_owned(),
-                slots: 4,
-            })
-        );
-    }
-
-    #[test]
-    fn connects_without_an_agent_name_when_none_was_entered() {
-        let mut form = form_with("127.0.0.1:8060", "1");
-
-        assert_eq!(
-            form.update(JoinBalancerFormMessage::Connect),
-            JoinBalancerFormAction::ConnectAgent(AgentBootstrapConfig {
-                agent_name: None,
-                management_address: "127.0.0.1:8060".to_owned(),
-                slots: 1,
-            })
-        );
-    }
-
-    #[test]
-    fn requires_both_the_address_and_the_slot_count() {
-        let mut form = form_with("", "");
-
-        assert_eq!(
-            form.update(JoinBalancerFormMessage::Connect),
-            JoinBalancerFormAction::None
-        );
-        assert_eq!(
-            form.balancer_address_error.as_ref().map(discriminant),
-            Some(discriminant(&FormFieldError::ClusterAddressRequired))
-        );
-        assert_eq!(
-            form.slots_error.as_ref().map(discriminant),
-            Some(discriminant(&FormFieldError::SlotsRequired))
-        );
-    }
-
-    #[test]
-    fn rejects_an_address_without_a_port() {
-        let mut form = form_with("127.0.0.1", "1");
-
-        assert_eq!(
-            form.update(JoinBalancerFormMessage::Connect),
-            JoinBalancerFormAction::None
-        );
-        assert!(matches!(
-            &form.balancer_address_error,
-            Some(FormFieldError::AddressUnparsable { source })
-                if "127.0.0.1".parse::<SocketAddr>().err().as_ref() == Some(source)
-        ));
-    }
-
-    #[test]
-    fn rejects_zero_slots() {
-        let mut form = form_with("127.0.0.1:8060", "0");
-
-        assert_eq!(
-            form.update(JoinBalancerFormMessage::Connect),
-            JoinBalancerFormAction::None
-        );
-        assert_eq!(
-            form.slots_error.as_ref().map(discriminant),
-            Some(discriminant(&FormFieldError::SlotsNotPositive))
-        );
-    }
-
-    #[test]
-    fn rejects_a_slot_count_that_does_not_fit() {
-        let mut form = form_with("127.0.0.1:8060", "99999999999");
-
-        assert_eq!(
-            form.update(JoinBalancerFormMessage::Connect),
-            JoinBalancerFormAction::None
-        );
-        assert!(matches!(
-            form.slots_error,
-            Some(FormFieldError::SlotsAboveLimit { limit }) if limit == LLAMA_CPP_MAX_SEQUENCES
-        ));
-    }
-
-    #[test]
-    fn rejects_a_slot_count_above_what_llama_cpp_can_serve() {
-        let mut form = form_with("127.0.0.1:8060", "257");
-
-        assert_eq!(
-            form.update(JoinBalancerFormMessage::Connect),
-            JoinBalancerFormAction::None
-        );
-        assert!(matches!(
-            form.slots_error,
-            Some(FormFieldError::SlotsAboveLimit { limit }) if limit == LLAMA_CPP_MAX_SEQUENCES
-        ));
-    }
-
-    #[test]
-    fn rejects_a_slot_count_that_is_not_a_number() {
-        let mut form = form_with("127.0.0.1:8060", "+");
-
-        assert_eq!(
-            form.update(JoinBalancerFormMessage::Connect),
-            JoinBalancerFormAction::None
-        );
-        assert!(matches!(
-            &form.slots_error,
-            Some(FormFieldError::SlotsInvalid { source })
-                if "+".parse::<u16>().err().as_ref() == Some(source)
-        ));
-    }
-
-    #[test]
-    fn clears_the_address_error_when_the_address_changes() {
-        let mut form = form_with("", "1");
-
-        form.update(JoinBalancerFormMessage::Connect);
-        form.update(JoinBalancerFormMessage::SetBalancerAddress(
-            "127.0.0.1:8060".to_owned(),
-        ));
-
-        assert!(form.balancer_address_error.is_none());
-    }
-}
diff --git a/paddler_gui/src/join_balancer_form_message.rs b/paddler_gui/src/join_balancer_form_message.rs
deleted file mode 100644
index 36815e36d..000000000
--- a/paddler_gui/src/join_balancer_form_message.rs
+++ /dev/null
@@ -1,8 +0,0 @@
-#[derive(Debug, Clone)]
-pub enum JoinBalancerFormMessage {
-    SetAgentName(String),
-    SetBalancerAddress(String),
-    SetSlotsCount(String),
-    Connect,
-    Cancel,
-}
diff --git a/paddler_gui/src/lib.rs b/paddler_gui/src/lib.rs
deleted file mode 100644
index 09e3827a8..000000000
--- a/paddler_gui/src/lib.rs
+++ /dev/null
@@ -1,32 +0,0 @@
-pub mod address_placeholder;
-pub mod agent_runner_messages;
-pub mod agent_running_action;
-pub mod agent_running_data;
-pub mod agent_running_message;
-pub mod app;
-pub mod balancer_launch;
-pub mod balancer_runner_messages;
-pub mod current_screen;
-pub mod detect_network_interfaces;
-pub mod form_field_error;
-pub mod home_data;
-pub mod home_message;
-pub mod join_balancer_form_action;
-pub mod join_balancer_form_data;
-pub mod join_balancer_form_message;
-pub mod message;
-pub mod model_preset;
-pub mod network_interface_address;
-pub mod paddler_application;
-pub mod port_check;
-pub mod running_balancer_action;
-pub mod running_balancer_data;
-pub mod running_balancer_message;
-pub mod running_balancer_snapshot;
-#[expect(unsafe_code, reason = "statum macros generate link_section statics")]
-pub mod screen;
-pub mod sort_agent_snapshots_by_label;
-pub mod start_balancer_form_action;
-pub mod start_balancer_form_data;
-pub mod start_balancer_form_message;
-pub mod ui;
diff --git a/paddler_gui/src/main.rs b/paddler_gui/src/main.rs
deleted file mode 100644
index 403a08b2f..000000000
--- a/paddler_gui/src/main.rs
+++ /dev/null
@@ -1,35 +0,0 @@
-use clap::Parser;
-use clap::Subcommand;
-use env_logger::Builder;
-use env_logger::Env;
-use iced::Result as IcedResult;
-use log::info;
-
-use paddler_gui::paddler_application::paddler_application;
-
-fn launch_gui() -> IcedResult {
-    Builder::from_env(Env::default().default_filter_or("info")).init();
-
-    info!("paddler_gui: ready");
-
-    paddler_application().run()
-}
-
-#[derive(Parser)]
-#[command(version, about, long_about = None)]
-struct Cli {
-    #[command(subcommand)]
-    command: Option<Commands>,
-}
-
-#[derive(Subcommand)]
-enum Commands {
-    /// Launch the desktop GUI application (default if no subcommand is given)
-    Launch,
-}
-
-fn main() -> IcedResult {
-    match Cli::parse().command {
-        Some(Commands::Launch) | None => launch_gui(),
-    }
-}
diff --git a/paddler_gui/src/message.rs b/paddler_gui/src/message.rs
deleted file mode 100644
index b31cc391e..000000000
--- a/paddler_gui/src/message.rs
+++ /dev/null
@@ -1,36 +0,0 @@
-use std::sync::Arc;
-
-use tokio_util::sync::CancellationToken;
-
-use paddler_balancer::balancer_addresses::BalancerAddresses;
-use paddler_bootstrap::bootstrap_error::BootstrapError;
-
-use crate::agent_running_message::AgentRunningMessage;
-use crate::home_message::HomeMessage;
-use crate::join_balancer_form_message::JoinBalancerFormMessage;
-use crate::running_balancer_message::RunningBalancerMessage;
-use crate::running_balancer_snapshot::RunningBalancerSnapshot;
-use crate::start_balancer_form_message::StartBalancerFormMessage;
-
-#[derive(Debug, Clone)]
-pub enum Message {
-    Home(HomeMessage),
-    StartBalancerForm(StartBalancerFormMessage),
-    JoinBalancerForm(JoinBalancerFormMessage),
-    RunningBalancer(RunningBalancerMessage),
-    AgentRunning(AgentRunningMessage),
-    BalancerStarted {
-        addresses: BalancerAddresses,
-        cancellation_token: CancellationToken,
-        snapshot: Box<RunningBalancerSnapshot>,
-    },
-    BalancerStopped,
-    BalancerFailed(Arc<BootstrapError>),
-    AgentStopped,
-    AgentFailed(Arc<BootstrapError>),
-    IcedEventLoopReady,
-    Quit,
-    TabPressed {
-        shift: bool,
-    },
-}
diff --git a/paddler_gui/src/model_preset.rs b/paddler_gui/src/model_preset.rs
deleted file mode 100644
index e619544b9..000000000
--- a/paddler_gui/src/model_preset.rs
+++ /dev/null
@@ -1,92 +0,0 @@
-use std::fmt;
-
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
-use paddler_messaging::agent_desired_model::AgentDesiredModel;
-use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-use paddler_messaging::huggingface_model_reference::HuggingFaceModelReference;
-
-fn huggingface_model(repo_id: &str, filename: &str) -> HuggingFaceModelReference {
-    HuggingFaceModelReference {
-        repo_id: repo_id.to_owned(),
-        filename: filename.to_owned(),
-        revision: "main".to_owned(),
-    }
-}
-
-#[derive(Clone, Copy, Debug, Eq, PartialEq)]
-pub enum ModelPreset {
-    Qwen3_0_6B,
-    Qwen3_5_0_8B,
-}
-
-impl ModelPreset {
-    pub const ALL: [Self; 2] = [Self::Qwen3_0_6B, Self::Qwen3_5_0_8B];
-
-    #[must_use]
-    pub fn model(self) -> HuggingFaceModelReference {
-        match self {
-            Self::Qwen3_0_6B => {
-                huggingface_model("unsloth/Qwen3-0.6B-GGUF", "Qwen3-0.6B-Q8_0.gguf")
-            }
-            Self::Qwen3_5_0_8B => {
-                huggingface_model("unsloth/Qwen3.5-0.8B-GGUF", "Qwen3.5-0.8B-Q4_K_M.gguf")
-            }
-        }
-    }
-
-    #[must_use]
-    pub fn multimodal_projection(self) -> Option<HuggingFaceModelReference> {
-        match self {
-            Self::Qwen3_0_6B => None,
-            Self::Qwen3_5_0_8B => Some(huggingface_model(
-                "unsloth/Qwen3.5-0.8B-GGUF",
-                "mmproj-F16.gguf",
-            )),
-        }
-    }
-
-    #[must_use]
-    pub fn to_balancer_desired_state(self) -> BalancerDesiredState {
-        BalancerDesiredState {
-            chat_template_override: None,
-            inference_parameters: InferenceParameters::default(),
-            model: AgentDesiredModel::HuggingFace(self.model()),
-            multimodal_projection: self
-                .multimodal_projection()
-                .map_or(AgentDesiredModel::None, AgentDesiredModel::HuggingFace),
-            use_chat_template_override: false,
-        }
-    }
-}
-
-impl fmt::Display for ModelPreset {
-    fn fmt(&self, formatter: &mut fmt::Formatter<'_>) -> fmt::Result {
-        formatter.write_str(match self {
-            Self::Qwen3_0_6B => "Qwen 3 0.6B",
-            Self::Qwen3_5_0_8B => "Qwen 3.5 0.8B",
-        })
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use paddler_messaging::agent_desired_model::AgentDesiredModel;
-    use paddler_messaging::huggingface_model_reference::HuggingFaceModelReference;
-
-    use super::ModelPreset;
-
-    #[test]
-    fn a_text_only_preset_runs_without_a_multimodal_projection() {
-        let desired_state = ModelPreset::Qwen3_0_6B.to_balancer_desired_state();
-
-        assert_eq!(
-            desired_state.model,
-            AgentDesiredModel::HuggingFace(HuggingFaceModelReference {
-                repo_id: "unsloth/Qwen3-0.6B-GGUF".to_owned(),
-                filename: "Qwen3-0.6B-Q8_0.gguf".to_owned(),
-                revision: "main".to_owned(),
-            })
-        );
-        assert_eq!(desired_state.multimodal_projection, AgentDesiredModel::None);
-    }
-}
diff --git a/paddler_gui/src/network_interface_address.rs b/paddler_gui/src/network_interface_address.rs
deleted file mode 100644
index f8eeae47b..000000000
--- a/paddler_gui/src/network_interface_address.rs
+++ /dev/null
@@ -1,7 +0,0 @@
-use std::net::IpAddr;
-
-#[derive(Clone, Debug, PartialEq, Eq)]
-pub struct NetworkInterfaceAddress {
-    pub interface_name: String,
-    pub ip_address: IpAddr,
-}
diff --git a/paddler_gui/src/paddler_application.rs b/paddler_gui/src/paddler_application.rs
deleted file mode 100644
index 34c3c7ee4..000000000
--- a/paddler_gui/src/paddler_application.rs
+++ /dev/null
@@ -1,23 +0,0 @@
-use iced::Program;
-use iced::Size;
-use iced::Theme;
-use iced::application;
-use iced::application::Application;
-
-use crate::app::App;
-use crate::message::Message;
-
-#[must_use]
-pub fn paddler_application()
--> Application<impl Program<State = App, Message = Message, Theme = Theme>> {
-    application(App::new, App::update, App::view)
-        .font(include_bytes!(
-            "../../resources/fonts/JetBrainsMono-Regular.ttf"
-        ))
-        .font(include_bytes!(
-            "../../resources/fonts/JetBrainsMono-Bold.ttf"
-        ))
-        .theme(Theme::Light)
-        .window_size(Size::new(800.0, 800.0))
-        .subscription(App::subscription)
-}
diff --git a/paddler_gui/src/port_check.rs b/paddler_gui/src/port_check.rs
deleted file mode 100644
index 4e2f895cd..000000000
--- a/paddler_gui/src/port_check.rs
+++ /dev/null
@@ -1,20 +0,0 @@
-use std::io;
-use std::net::SocketAddr;
-use std::net::TcpListener;
-
-pub enum PortCheck {
-    Available,
-    BindFailed(io::Error),
-    InUse,
-}
-
-impl PortCheck {
-    #[must_use]
-    pub fn of(address: &SocketAddr) -> Self {
-        match TcpListener::bind(address) {
-            Ok(_listener) => Self::Available,
-            Err(error) if error.kind() == io::ErrorKind::AddrInUse => Self::InUse,
-            Err(error) => Self::BindFailed(error),
-        }
-    }
-}
diff --git a/paddler_gui/src/running_balancer_action.rs b/paddler_gui/src/running_balancer_action.rs
deleted file mode 100644
index c7b21b64f..000000000
--- a/paddler_gui/src/running_balancer_action.rs
+++ /dev/null
@@ -1,6 +0,0 @@
-#[derive(Debug, Eq, PartialEq)]
-pub enum RunningBalancerAction {
-    None,
-    CopyToClipboard(String),
-    OpenUrl(String),
-}
diff --git a/paddler_gui/src/running_balancer_data.rs b/paddler_gui/src/running_balancer_data.rs
deleted file mode 100644
index 0f38a7eae..000000000
--- a/paddler_gui/src/running_balancer_data.rs
+++ /dev/null
@@ -1,96 +0,0 @@
-use tokio_util::sync::CancellationToken;
-
-use paddler_balancer::balancer_addresses::BalancerAddresses;
-
-use crate::running_balancer_action::RunningBalancerAction;
-use crate::running_balancer_message::RunningBalancerMessage;
-use crate::running_balancer_snapshot::RunningBalancerSnapshot;
-
-pub struct RunningBalancerData {
-    pub addresses: BalancerAddresses,
-    pub cancellation_token: CancellationToken,
-    pub snapshot: Box<RunningBalancerSnapshot>,
-    pub stopping: bool,
-}
-
-impl RunningBalancerData {
-    pub fn update(&mut self, message: RunningBalancerMessage) -> RunningBalancerAction {
-        match message {
-            RunningBalancerMessage::SnapshotUpdated(snapshot) => {
-                self.snapshot = snapshot;
-
-                RunningBalancerAction::None
-            }
-            RunningBalancerMessage::Stop => {
-                self.cancellation_token.cancel();
-                self.stopping = true;
-
-                RunningBalancerAction::None
-            }
-            RunningBalancerMessage::CopyToClipboard(content) => {
-                RunningBalancerAction::CopyToClipboard(content)
-            }
-            RunningBalancerMessage::OpenUrl(url) => RunningBalancerAction::OpenUrl(url),
-        }
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use std::net::Ipv4Addr;
-    use std::net::SocketAddr;
-
-    use tokio_util::sync::CancellationToken;
-
-    use paddler_balancer::balancer_addresses::BalancerAddresses;
-    use paddler_balancer::balancer_applicable_state::BalancerApplicableState;
-    use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-
-    use super::RunningBalancerData;
-    use crate::running_balancer_action::RunningBalancerAction;
-    use crate::running_balancer_message::RunningBalancerMessage;
-    use crate::running_balancer_snapshot::RunningBalancerSnapshot;
-
-    fn running_balancer() -> RunningBalancerData {
-        RunningBalancerData {
-            addresses: BalancerAddresses {
-                compat_openai: None,
-                inference: SocketAddr::from((Ipv4Addr::LOCALHOST, 8061)),
-                management: SocketAddr::from((Ipv4Addr::LOCALHOST, 8060)),
-                web_admin_panel: None,
-            },
-            cancellation_token: CancellationToken::new(),
-            snapshot: Box::new(RunningBalancerSnapshot {
-                agent_snapshots: Vec::new(),
-                balancer_applicable_state: BalancerApplicableState::from(
-                    BalancerDesiredState::default(),
-                ),
-                balancer_desired_state: BalancerDesiredState::default(),
-            }),
-            stopping: false,
-        }
-    }
-
-    #[test]
-    fn stopping_cancels_the_running_balancer() {
-        let mut balancer = running_balancer();
-
-        let action = balancer.update(RunningBalancerMessage::Stop);
-
-        assert_eq!(action, RunningBalancerAction::None);
-        assert!(balancer.cancellation_token.is_cancelled());
-        assert!(balancer.stopping);
-    }
-
-    #[test]
-    fn following_a_link_asks_to_open_it() {
-        let mut balancer = running_balancer();
-
-        assert_eq!(
-            balancer.update(RunningBalancerMessage::OpenUrl(
-                "http://127.0.0.1:8062".to_owned()
-            )),
-            RunningBalancerAction::OpenUrl("http://127.0.0.1:8062".to_owned())
-        );
-    }
-}
diff --git a/paddler_gui/src/running_balancer_message.rs b/paddler_gui/src/running_balancer_message.rs
deleted file mode 100644
index 5d634947d..000000000
--- a/paddler_gui/src/running_balancer_message.rs
+++ /dev/null
@@ -1,9 +0,0 @@
-use crate::running_balancer_snapshot::RunningBalancerSnapshot;
-
-#[derive(Debug, Clone)]
-pub enum RunningBalancerMessage {
-    SnapshotUpdated(Box<RunningBalancerSnapshot>),
-    Stop,
-    CopyToClipboard(String),
-    OpenUrl(String),
-}
diff --git a/paddler_gui/src/running_balancer_snapshot.rs b/paddler_gui/src/running_balancer_snapshot.rs
deleted file mode 100644
index d61ebed40..000000000
--- a/paddler_gui/src/running_balancer_snapshot.rs
+++ /dev/null
@@ -1,73 +0,0 @@
-use paddler_balancer::agent_controller_pool::AgentControllerPool;
-use paddler_balancer::balancer_applicable_state::BalancerApplicableState;
-use paddler_balancer::balancer_applicable_state_holder::BalancerApplicableStateHolder;
-use paddler_messaging::agent_controller_snapshot::AgentControllerSnapshot;
-use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-use paddler_messaging::produces_snapshot::ProducesSnapshot as _;
-
-use crate::sort_agent_snapshots_by_label::sort_agent_snapshots_by_label;
-
-#[derive(Clone, Debug)]
-pub struct RunningBalancerSnapshot {
-    pub agent_snapshots: Vec<AgentControllerSnapshot>,
-    pub balancer_applicable_state: BalancerApplicableState,
-    pub balancer_desired_state: BalancerDesiredState,
-}
-
-impl RunningBalancerSnapshot {
-    pub fn build(
-        agent_controller_pool: &AgentControllerPool,
-        balancer_applicable_state_holder: &BalancerApplicableStateHolder,
-        balancer_desired_state: BalancerDesiredState,
-    ) -> Self {
-        let mut agent_snapshots = agent_controller_pool.make_snapshot().agents;
-
-        sort_agent_snapshots_by_label(&mut agent_snapshots);
-
-        let balancer_applicable_state =
-            balancer_applicable_state_holder.get_balancer_applicable_state();
-
-        Self {
-            agent_snapshots,
-            balancer_applicable_state,
-            balancer_desired_state,
-        }
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use paddler_balancer::agent_controller_pool::AgentControllerPool;
-    use paddler_balancer::balancer_applicable_state::BalancerApplicableState;
-    use paddler_balancer::balancer_applicable_state_holder::BalancerApplicableStateHolder;
-    use paddler_messaging::agent_desired_model::AgentDesiredModel;
-    use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-
-    use super::RunningBalancerSnapshot;
-
-    #[test]
-    fn carries_the_applicable_state_apart_from_the_desired_state() {
-        let applied_model = AgentDesiredModel::LocalToAgent("configured_model".to_owned());
-        let requested_model = AgentDesiredModel::LocalToAgent("requested_model".to_owned());
-
-        let snapshot = RunningBalancerSnapshot::build(
-            &AgentControllerPool::default(),
-            &BalancerApplicableStateHolder::new(BalancerApplicableState::from(
-                BalancerDesiredState {
-                    model: applied_model.clone(),
-                    ..BalancerDesiredState::default()
-                },
-            )),
-            BalancerDesiredState {
-                model: requested_model.clone(),
-                ..BalancerDesiredState::default()
-            },
-        );
-
-        assert_eq!(
-            snapshot.balancer_applicable_state.agent_desired_state.model,
-            applied_model
-        );
-        assert_eq!(snapshot.balancer_desired_state.model, requested_model);
-    }
-}
diff --git a/paddler_gui/src/screen.rs b/paddler_gui/src/screen.rs
deleted file mode 100644
index 4a3116a01..000000000
--- a/paddler_gui/src/screen.rs
+++ /dev/null
@@ -1,126 +0,0 @@
-use std::sync::Arc;
-
-use statum::machine;
-use statum::state;
-use statum::transition;
-use tokio_util::sync::CancellationToken;
-
-use paddler_balancer::balancer_addresses::BalancerAddresses;
-use paddler_bootstrap::bootstrap_error::BootstrapError;
-use paddler_messaging::agent_status::AgentStatus;
-use paddler_messaging::balancer_connection::BalancerConnection;
-
-use crate::agent_running_data::AgentRunningData;
-use crate::detect_network_interfaces::detect_network_interfaces;
-use crate::home_data::HomeData;
-use crate::join_balancer_form_data::JoinBalancerFormData;
-use crate::running_balancer_data::RunningBalancerData;
-use crate::running_balancer_snapshot::RunningBalancerSnapshot;
-use crate::start_balancer_form_data::StartBalancerFormData;
-
-#[state]
-pub enum ScreenState {
-    AgentRunning(AgentRunningData),
-    Home(HomeData),
-    JoinBalancerForm(JoinBalancerFormData),
-    StartBalancerForm(StartBalancerFormData),
-    RunningBalancer(RunningBalancerData),
-}
-
-#[machine]
-pub struct Screen<ScreenState> {}
-
-#[transition]
-impl Screen<Home> {
-    #[must_use]
-    pub fn join_balancer(self) -> Screen<JoinBalancerForm> {
-        self.transition_with(JoinBalancerFormData::default())
-    }
-
-    #[must_use]
-    pub fn start_balancer(self) -> Screen<StartBalancerForm> {
-        self.transition_with(StartBalancerFormData::suggesting_addresses_on(
-            detect_network_interfaces()
-                .first()
-                .map(|interface| interface.ip_address),
-        ))
-    }
-}
-
-#[transition]
-impl Screen<JoinBalancerForm> {
-    #[must_use]
-    pub fn cancel(self) -> Screen<Home> {
-        self.transition_with(HomeData { error: None })
-    }
-
-    #[must_use]
-    pub fn connect(self, cancellation_token: CancellationToken) -> Screen<AgentRunning> {
-        self.transition_map(|form_data: JoinBalancerFormData| {
-            let name = form_data.entered_agent_name();
-
-            AgentRunningData {
-                balancer_address: form_data.balancer_address,
-                balancer_connection: BalancerConnection::Connecting,
-                cancellation_token,
-                name,
-                slots_processing: 0,
-                status: AgentStatus::default(),
-            }
-        })
-    }
-}
-
-#[transition]
-impl Screen<AgentRunning> {
-    #[must_use]
-    pub fn disconnect(self) -> Screen<Home> {
-        self.transition_with(HomeData { error: None })
-    }
-
-    #[must_use]
-    pub fn agent_failed(self, error: Arc<BootstrapError>) -> Screen<Home> {
-        self.transition_with(HomeData { error: Some(error) })
-    }
-}
-
-#[transition]
-impl Screen<StartBalancerForm> {
-    #[must_use]
-    pub fn cancel(self) -> Screen<Home> {
-        self.transition_with(HomeData { error: None })
-    }
-
-    #[must_use]
-    pub fn balancer_started(
-        self,
-        addresses: BalancerAddresses,
-        cancellation_token: CancellationToken,
-        snapshot: Box<RunningBalancerSnapshot>,
-    ) -> Screen<RunningBalancer> {
-        self.transition_with(RunningBalancerData {
-            addresses,
-            cancellation_token,
-            snapshot,
-            stopping: false,
-        })
-    }
-
-    #[must_use]
-    pub fn balancer_failed(self, error: Arc<BootstrapError>) -> Screen<Home> {
-        self.transition_with(HomeData { error: Some(error) })
-    }
-}
-
-#[transition]
-impl Screen<RunningBalancer> {
-    #[must_use]
-    pub fn balancer_stopped(self) -> Screen<Home> {
-        self.transition_with(HomeData { error: None })
-    }
-
-    #[must_use]
-    pub fn balancer_failed(self, error: Arc<BootstrapError>) -> Screen<Home> {
-        self.transition_with(HomeData { error: Some(error) })
-    }
-}
diff --git a/paddler_gui/src/sort_agent_snapshots_by_label.rs b/paddler_gui/src/sort_agent_snapshots_by_label.rs
deleted file mode 100644
index 19766225e..000000000
--- a/paddler_gui/src/sort_agent_snapshots_by_label.rs
+++ /dev/null
@@ -1,46 +0,0 @@
-use paddler_messaging::agent_controller_snapshot::AgentControllerSnapshot;
-
-pub fn sort_agent_snapshots_by_label(agent_snapshots: &mut [AgentControllerSnapshot]) {
-    agent_snapshots.sort_by(|left_agent, right_agent| {
-        let left_label = left_agent.name.as_deref().unwrap_or(&left_agent.id);
-        let right_label = right_agent.name.as_deref().unwrap_or(&right_agent.id);
-
-        left_label.cmp(right_label)
-    });
-}
-
-#[cfg(test)]
-mod tests {
-    use paddler_messaging::agent_controller_snapshot::AgentControllerSnapshot;
-    use paddler_messaging::agent_status::AgentStatus;
-
-    use super::sort_agent_snapshots_by_label;
-
-    fn agent_snapshot(id: &str, name: Option<&str>) -> AgentControllerSnapshot {
-        AgentControllerSnapshot {
-            id: id.to_owned(),
-            name: name.map(str::to_owned),
-            slots_processing: 0,
-            status: AgentStatus::default(),
-        }
-    }
-
-    #[test]
-    fn orders_agents_by_name_falling_back_to_their_id() {
-        let mut agent_snapshots = vec![
-            agent_snapshot("id_z", Some("alpha")),
-            agent_snapshot("id_m", None),
-            agent_snapshot("id_a", Some("bravo")),
-        ];
-
-        sort_agent_snapshots_by_label(&mut agent_snapshots);
-
-        assert_eq!(
-            agent_snapshots
-                .iter()
-                .map(|agent_snapshot| agent_snapshot.id.as_str())
-                .collect::<Vec<_>>(),
-            vec!["id_z", "id_a", "id_m"]
-        );
-    }
-}
diff --git a/paddler_gui/src/start_balancer_form_action.rs b/paddler_gui/src/start_balancer_form_action.rs
deleted file mode 100644
index dd11f8d93..000000000
--- a/paddler_gui/src/start_balancer_form_action.rs
+++ /dev/null
@@ -1,7 +0,0 @@
-use paddler_bootstrap::balancer_bootstrap_config::BalancerBootstrapConfig;
-
-pub enum StartBalancerFormAction {
-    None,
-    Cancel,
-    StartBalancer(Box<BalancerBootstrapConfig>),
-}
diff --git a/paddler_gui/src/start_balancer_form_data.rs b/paddler_gui/src/start_balancer_form_data.rs
deleted file mode 100644
index 4bf203927..000000000
--- a/paddler_gui/src/start_balancer_form_data.rs
+++ /dev/null
@@ -1,428 +0,0 @@
-use std::net::IpAddr;
-use std::net::SocketAddr;
-
-use paddler_balancer::inference_service::configuration::Configuration as InferenceServiceConfiguration;
-use paddler_balancer::management_service::configuration::Configuration as ManagementServiceConfiguration;
-use paddler_balancer::resolved_socket_addr::ResolvedSocketAddr;
-#[cfg(feature = "web_admin_panel")]
-use paddler_balancer::web_admin_panel_service::configuration::Configuration as WebAdminPanelServiceConfiguration;
-use paddler_bootstrap::balancer_bootstrap_config::BalancerBootstrapConfig;
-use paddler_bootstrap::balancer_defaults::BalancerDefaults;
-use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-use paddler_state_database::state_database_type::StateDatabaseType;
-
-use crate::address_placeholder::ADDRESS_PLACEHOLDER;
-use crate::balancer_launch::BalancerLaunch;
-use crate::form_field_error::FormFieldError;
-use crate::model_preset::ModelPreset;
-use crate::port_check::PortCheck;
-use crate::start_balancer_form_action::StartBalancerFormAction;
-use crate::start_balancer_form_message::StartBalancerFormMessage;
-
-fn balancer_bootstrap_config(
-    desired_state: BalancerDesiredState,
-    management_addr: SocketAddr,
-    inference_addr: SocketAddr,
-    #[cfg_attr(
-        not(feature = "web_admin_panel"),
-        expect(
-            unused_variables,
-            reason = "the web admin panel is only served when the feature is enabled"
-        )
-    )]
-    web_admin_panel_addr: Option<SocketAddr>,
-) -> BalancerBootstrapConfig {
-    BalancerBootstrapConfig {
-        buffered_request_timeout: BalancerDefaults::BUFFERED_REQUEST_TIMEOUT,
-        inference_service_configuration: InferenceServiceConfiguration {
-            addr: ResolvedSocketAddr::from(inference_addr),
-            cors_allowed_hosts: vec![],
-            inference_item_timeout: BalancerDefaults::INFERENCE_ITEM_TIMEOUT,
-        },
-        management_service_configuration: ManagementServiceConfiguration {
-            addr: ResolvedSocketAddr::from(management_addr),
-            cors_allowed_hosts: vec![],
-        },
-        max_buffered_requests: BalancerDefaults::MAX_BUFFERED_REQUESTS,
-        openai_service_configuration: None,
-        state_database_type: StateDatabaseType::Memory(Box::new(desired_state)),
-        statsd_prefix: BalancerDefaults::STATSD_PREFIX.to_owned(),
-        statsd_service_configuration: None,
-        #[cfg(feature = "web_admin_panel")]
-        web_admin_panel_service_configuration: web_admin_panel_addr
-            .map(|addr| WebAdminPanelServiceConfiguration { addr }),
-    }
-}
-
-fn validate_optional_address(raw: &str) -> Result<Option<SocketAddr>, FormFieldError> {
-    if raw.is_empty() {
-        return Ok(None);
-    }
-
-    let addr = raw
-        .parse::<SocketAddr>()
-        .map_err(|source| FormFieldError::AddressUnparsable { source })?;
-
-    match PortCheck::of(&addr) {
-        PortCheck::Available => Ok(Some(addr)),
-        PortCheck::InUse => Err(FormFieldError::AddressPortInUse { port: addr.port() }),
-        PortCheck::BindFailed(source) => Err(FormFieldError::AddressCannotBeBound { addr, source }),
-    }
-}
-
-fn validate_required_address(raw: &str) -> Result<SocketAddr, FormFieldError> {
-    validate_optional_address(raw)?.ok_or(FormFieldError::AddressRequired)
-}
-
-pub struct StartBalancerFormData {
-    pub add_model_later: bool,
-    pub balancer_address: String,
-    pub balancer_address_error: Option<FormFieldError>,
-    pub inference_address: String,
-    pub inference_address_error: Option<FormFieldError>,
-    pub launch: BalancerLaunch,
-    pub model_error: Option<FormFieldError>,
-    pub selected_model: Option<ModelPreset>,
-    pub web_admin_panel_address: String,
-    pub web_admin_panel_address_error: Option<FormFieldError>,
-    pub web_admin_panel_address_placeholder: String,
-}
-
-impl StartBalancerFormData {
-    #[must_use]
-    pub fn suggesting_addresses_on(interface_address: Option<IpAddr>) -> Self {
-        let suggested_address_on = |port| {
-            interface_address.map_or_else(String::new, |ip_address| {
-                SocketAddr::new(ip_address, port).to_string()
-            })
-        };
-
-        Self {
-            add_model_later: false,
-            balancer_address: suggested_address_on(BalancerDefaults::MANAGEMENT_PORT),
-            balancer_address_error: None,
-            inference_address: suggested_address_on(BalancerDefaults::INFERENCE_PORT),
-            inference_address_error: None,
-            launch: BalancerLaunch::NotRequested,
-            model_error: None,
-            selected_model: None,
-            web_admin_panel_address: String::new(),
-            web_admin_panel_address_error: None,
-            web_admin_panel_address_placeholder: interface_address.map_or_else(
-                || ADDRESS_PLACEHOLDER.to_owned(),
-                |ip_address| {
-                    SocketAddr::new(ip_address, BalancerDefaults::WEB_ADMIN_PANEL_PORT).to_string()
-                },
-            ),
-        }
-    }
-
-    pub fn update(&mut self, message: StartBalancerFormMessage) -> StartBalancerFormAction {
-        match message {
-            StartBalancerFormMessage::SelectModel(preset) => {
-                self.selected_model = Some(preset);
-                self.model_error = None;
-
-                StartBalancerFormAction::None
-            }
-            StartBalancerFormMessage::SetBalancerAddress(address) => {
-                self.balancer_address = address;
-                self.balancer_address_error = None;
-
-                StartBalancerFormAction::None
-            }
-            StartBalancerFormMessage::SetInferenceAddress(address) => {
-                self.inference_address = address;
-                self.inference_address_error = None;
-
-                StartBalancerFormAction::None
-            }
-            StartBalancerFormMessage::SetWebAdminPanelAddress(address) => {
-                self.web_admin_panel_address = address;
-                self.web_admin_panel_address_error = None;
-
-                StartBalancerFormAction::None
-            }
-            StartBalancerFormMessage::ToggleAddModelLater(add_later) => {
-                self.add_model_later = add_later;
-
-                if add_later {
-                    self.model_error = None;
-                }
-
-                StartBalancerFormAction::None
-            }
-            StartBalancerFormMessage::Confirm => self.validate_and_confirm(),
-            StartBalancerFormMessage::Cancel => StartBalancerFormAction::Cancel,
-        }
-    }
-
-    fn desired_state(&self) -> Result<BalancerDesiredState, FormFieldError> {
-        if self.add_model_later {
-            return Ok(BalancerDesiredState::default());
-        }
-
-        self.selected_model
-            .map(ModelPreset::to_balancer_desired_state)
-            .ok_or(FormFieldError::ModelNotSelected)
-    }
-
-    fn validate_and_confirm(&mut self) -> StartBalancerFormAction {
-        match (
-            self.desired_state(),
-            validate_required_address(&self.balancer_address),
-            validate_required_address(&self.inference_address),
-            validate_optional_address(&self.web_admin_panel_address),
-        ) {
-            (
-                Ok(desired_state),
-                Ok(management_addr),
-                Ok(inference_addr),
-                Ok(web_admin_panel_addr),
-            ) => StartBalancerFormAction::StartBalancer(Box::new(balancer_bootstrap_config(
-                desired_state,
-                management_addr,
-                inference_addr,
-                web_admin_panel_addr,
-            ))),
-            (desired_state, management_addr, inference_addr, web_admin_panel_addr) => {
-                self.model_error = desired_state.err();
-                self.balancer_address_error = management_addr.err();
-                self.inference_address_error = inference_addr.err();
-                self.web_admin_panel_address_error = web_admin_panel_addr.err();
-
-                StartBalancerFormAction::None
-            }
-        }
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use std::mem::discriminant;
-    use std::net::IpAddr;
-    use std::net::Ipv4Addr;
-    use std::net::SocketAddr;
-    use std::net::TcpListener;
-
-    use paddler_messaging::agent_desired_model::AgentDesiredModel;
-    use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-    use paddler_state_database::state_database_type::StateDatabaseType;
-    use paddler_test_cluster_harness::ephemeral_loopback_addr::EPHEMERAL_LOOPBACK_ADDR;
-
-    use super::StartBalancerFormData;
-    use crate::balancer_launch::BalancerLaunch;
-    use crate::form_field_error::FormFieldError;
-    use crate::model_preset::ModelPreset;
-    use crate::start_balancer_form_action::StartBalancerFormAction;
-    use crate::start_balancer_form_message::StartBalancerFormMessage;
-
-    const UNASSIGNED_TEST_NET_ADDRESS: &str = "192.0.2.1:0";
-
-    fn form_on_free_ports() -> StartBalancerFormData {
-        StartBalancerFormData {
-            add_model_later: false,
-            balancer_address: EPHEMERAL_LOOPBACK_ADDR.to_string(),
-            balancer_address_error: None,
-            inference_address: EPHEMERAL_LOOPBACK_ADDR.to_string(),
-            inference_address_error: None,
-            launch: BalancerLaunch::NotRequested,
-            model_error: None,
-            selected_model: None,
-            web_admin_panel_address: String::new(),
-            web_admin_panel_address_error: None,
-            web_admin_panel_address_placeholder: String::new(),
-        }
-    }
-
-    fn multimodal_preset() -> ModelPreset {
-        ModelPreset::ALL
-            .into_iter()
-            .find(|preset| preset.multimodal_projection().is_some())
-            .expect("a multimodal preset must be available")
-    }
-
-    #[test]
-    fn requires_a_model_unless_it_is_added_later() {
-        let mut form = form_on_free_ports();
-
-        assert_eq!(
-            discriminant(&form.update(StartBalancerFormMessage::Confirm)),
-            discriminant(&StartBalancerFormAction::None)
-        );
-        assert_eq!(
-            form.model_error.as_ref().map(discriminant),
-            Some(discriminant(&FormFieldError::ModelNotSelected))
-        );
-    }
-
-    #[test]
-    fn starts_without_a_model_when_it_is_added_later() {
-        let mut form = form_on_free_ports();
-
-        form.update(StartBalancerFormMessage::ToggleAddModelLater(true));
-
-        assert!(matches!(
-            form.update(StartBalancerFormMessage::Confirm),
-            StartBalancerFormAction::StartBalancer(bootstrap_config)
-                if matches!(
-                    &bootstrap_config.state_database_type,
-                    StateDatabaseType::Memory(desired_state)
-                        if **desired_state == BalancerDesiredState::default()
-                )
-        ));
-    }
-
-    #[test]
-    fn requires_a_model_again_once_it_is_no_longer_added_later() {
-        let mut form = form_on_free_ports();
-
-        form.update(StartBalancerFormMessage::ToggleAddModelLater(true));
-        form.update(StartBalancerFormMessage::ToggleAddModelLater(false));
-        form.update(StartBalancerFormMessage::Confirm);
-
-        assert_eq!(
-            form.model_error.as_ref().map(discriminant),
-            Some(discriminant(&FormFieldError::ModelNotSelected))
-        );
-    }
-
-    #[cfg(feature = "web_admin_panel")]
-    #[test]
-    fn serves_no_web_admin_panel_without_its_address() {
-        let mut form = form_on_free_ports();
-
-        form.update(StartBalancerFormMessage::ToggleAddModelLater(true));
-
-        assert!(matches!(
-            form.update(StartBalancerFormMessage::Confirm),
-            StartBalancerFormAction::StartBalancer(bootstrap_config)
-                if bootstrap_config.web_admin_panel_service_configuration.is_none()
-        ));
-    }
-
-    #[test]
-    fn starts_with_the_selected_model_and_its_multimodal_projection() {
-        let mut form = form_on_free_ports();
-        let preset = multimodal_preset();
-        let expected_model = AgentDesiredModel::HuggingFace(preset.model());
-        let expected_multimodal_projection = AgentDesiredModel::HuggingFace(
-            preset
-                .multimodal_projection()
-                .expect("the multimodal preset must carry a projection"),
-        );
-
-        form.update(StartBalancerFormMessage::SelectModel(preset));
-
-        assert!(matches!(
-            form.update(StartBalancerFormMessage::Confirm),
-            StartBalancerFormAction::StartBalancer(bootstrap_config)
-                if matches!(
-                    &bootstrap_config.state_database_type,
-                    StateDatabaseType::Memory(desired_state)
-                        if desired_state.model == expected_model
-                            && desired_state.multimodal_projection == expected_multimodal_projection
-                )
-        ));
-    }
-
-    #[test]
-    fn reports_every_invalid_address_at_once() {
-        let mut form = form_on_free_ports();
-
-        form.update(StartBalancerFormMessage::ToggleAddModelLater(true));
-        form.update(StartBalancerFormMessage::SetBalancerAddress(String::new()));
-        form.update(StartBalancerFormMessage::SetInferenceAddress(
-            "not an address".to_owned(),
-        ));
-        form.update(StartBalancerFormMessage::SetWebAdminPanelAddress(
-            "127.0.0.1".to_owned(),
-        ));
-
-        assert_eq!(
-            discriminant(&form.update(StartBalancerFormMessage::Confirm)),
-            discriminant(&StartBalancerFormAction::None)
-        );
-        assert_eq!(
-            form.balancer_address_error.as_ref().map(discriminant),
-            Some(discriminant(&FormFieldError::AddressRequired))
-        );
-        assert!(matches!(
-            &form.inference_address_error,
-            Some(FormFieldError::AddressUnparsable { source })
-                if "not an address".parse::<SocketAddr>().err().as_ref() == Some(source)
-        ));
-        assert!(matches!(
-            &form.web_admin_panel_address_error,
-            Some(FormFieldError::AddressUnparsable { source })
-                if "127.0.0.1".parse::<SocketAddr>().err().as_ref() == Some(source)
-        ));
-    }
-
-    #[test]
-    fn reports_an_address_another_listener_holds_as_in_use() {
-        let competing_listener =
-            TcpListener::bind(EPHEMERAL_LOOPBACK_ADDR).expect("the competing listener must bind");
-        let taken_address = competing_listener
-            .local_addr()
-            .expect("the competing listener must report its address");
-        let mut form = form_on_free_ports();
-
-        form.update(StartBalancerFormMessage::ToggleAddModelLater(true));
-        form.update(StartBalancerFormMessage::SetInferenceAddress(
-            taken_address.to_string(),
-        ));
-
-        assert_eq!(
-            discriminant(&form.update(StartBalancerFormMessage::Confirm)),
-            discriminant(&StartBalancerFormAction::None)
-        );
-        assert!(matches!(
-            form.inference_address_error,
-            Some(FormFieldError::AddressPortInUse { port }) if port == taken_address.port()
-        ));
-    }
-
-    #[test]
-    fn reports_why_an_address_cannot_be_bound() {
-        let bind_error = TcpListener::bind(UNASSIGNED_TEST_NET_ADDRESS)
-            .expect_err("an address assigned to no interface must not bind");
-        let mut form = form_on_free_ports();
-
-        form.update(StartBalancerFormMessage::ToggleAddModelLater(true));
-        form.update(StartBalancerFormMessage::SetWebAdminPanelAddress(
-            UNASSIGNED_TEST_NET_ADDRESS.to_owned(),
-        ));
-
-        assert_eq!(
-            discriminant(&form.update(StartBalancerFormMessage::Confirm)),
-            discriminant(&StartBalancerFormAction::None)
-        );
-        assert!(matches!(
-            form.web_admin_panel_address_error,
-            Some(FormFieldError::AddressCannotBeBound { addr, source })
-                if addr.to_string() == UNASSIGNED_TEST_NET_ADDRESS
-                    && source.kind() == bind_error.kind()
-        ));
-    }
-
-    #[test]
-    fn suggests_the_default_ports_on_the_detected_interface() {
-        let form = StartBalancerFormData::suggesting_addresses_on(Some(IpAddr::V4(Ipv4Addr::new(
-            192, 168, 1, 7,
-        ))));
-
-        assert_eq!(form.balancer_address, "192.168.1.7:8060");
-        assert_eq!(form.inference_address, "192.168.1.7:8061");
-        assert_eq!(form.web_admin_panel_address_placeholder, "192.168.1.7:8062");
-    }
-
-    #[test]
-    fn suggests_no_address_without_a_detected_interface() {
-        let form = StartBalancerFormData::suggesting_addresses_on(None);
-
-        assert_eq!(form.balancer_address, "");
-        assert_eq!(form.inference_address, "");
-        assert_eq!(form.web_admin_panel_address_placeholder, "IP:port");
-    }
-}
diff --git a/paddler_gui/src/start_balancer_form_message.rs b/paddler_gui/src/start_balancer_form_message.rs
deleted file mode 100644
index d306824cb..000000000
--- a/paddler_gui/src/start_balancer_form_message.rs
+++ /dev/null
@@ -1,12 +0,0 @@
-use crate::model_preset::ModelPreset;
-
-#[derive(Debug, Clone)]
-pub enum StartBalancerFormMessage {
-    SetBalancerAddress(String),
-    SetInferenceAddress(String),
-    SetWebAdminPanelAddress(String),
-    SelectModel(ModelPreset),
-    ToggleAddModelLater(bool),
-    Confirm,
-    Cancel,
-}
diff --git a/paddler_gui/src/ui/font.rs b/paddler_gui/src/ui/font.rs
deleted file mode 100644
index fcd5316ec..000000000
--- a/paddler_gui/src/ui/font.rs
+++ /dev/null
@@ -1,19 +0,0 @@
-use iced::Font;
-use iced::font::Family;
-use iced::font::Stretch;
-use iced::font::Style;
-use iced::font::Weight;
-
-pub const REGULAR: Font = Font {
-    family: Family::Name("JetBrains Mono"),
-    weight: Weight::Normal,
-    stretch: Stretch::Normal,
-    style: Style::Normal,
-};
-
-pub const BOLD: Font = Font {
-    family: Family::Name("JetBrains Mono"),
-    weight: Weight::Bold,
-    stretch: Stretch::Normal,
-    style: Style::Normal,
-};
diff --git a/paddler_gui/src/ui/mod.rs b/paddler_gui/src/ui/mod.rs
deleted file mode 100644
index 0333df897..000000000
--- a/paddler_gui/src/ui/mod.rs
+++ /dev/null
@@ -1,22 +0,0 @@
-mod font;
-mod style_agent_container;
-mod style_button_disconnect;
-mod style_button_primary;
-mod style_card_container;
-mod style_download_progress_bar;
-mod style_field_checkbox;
-mod style_field_container;
-mod style_field_pick_list;
-mod style_field_pick_list_menu;
-mod style_field_text_input;
-mod style_status_indicator;
-pub mod variables;
-mod view_agent_card;
-mod view_agent_running;
-mod view_error_message;
-mod view_form_field;
-mod view_home;
-mod view_join_balancer_form;
-mod view_running_balancer;
-mod view_start_balancer_form;
-mod view_stop_button;
diff --git a/paddler_gui/src/ui/style_agent_container.rs b/paddler_gui/src/ui/style_agent_container.rs
deleted file mode 100644
index bfc389862..000000000
--- a/paddler_gui/src/ui/style_agent_container.rs
+++ /dev/null
@@ -1,16 +0,0 @@
-use iced::Background;
-use iced::Theme;
-use iced::widget::container;
-
-use super::variables::BORDER_OUTLINE;
-use super::variables::COLOR_AGENT_BACKGROUND;
-
-pub fn style_agent_container(theme: &Theme) -> container::Style {
-    let base = container::transparent(theme);
-
-    container::Style {
-        background: Some(Background::Color(COLOR_AGENT_BACKGROUND)),
-        border: BORDER_OUTLINE,
-        ..base
-    }
-}
diff --git a/paddler_gui/src/ui/style_button_disconnect.rs b/paddler_gui/src/ui/style_button_disconnect.rs
deleted file mode 100644
index 7973c5ca1..000000000
--- a/paddler_gui/src/ui/style_button_disconnect.rs
+++ /dev/null
@@ -1,22 +0,0 @@
-use iced::Background;
-use iced::Border;
-use iced::Color;
-use iced::Theme;
-use iced::widget::button;
-
-use super::variables::COLOR_ERROR;
-
-pub fn style_button_disconnect(theme: &Theme, status: button::Status) -> button::Style {
-    let base = button::primary(theme, status);
-
-    button::Style {
-        background: Some(Background::Color(COLOR_ERROR)),
-        text_color: Color::WHITE,
-        border: Border {
-            color: COLOR_ERROR,
-            width: 0.0,
-            radius: 0.into(),
-        },
-        ..base
-    }
-}
diff --git a/paddler_gui/src/ui/style_button_primary.rs b/paddler_gui/src/ui/style_button_primary.rs
deleted file mode 100644
index 9ed1bcadd..000000000
--- a/paddler_gui/src/ui/style_button_primary.rs
+++ /dev/null
@@ -1,22 +0,0 @@
-use iced::Background;
-use iced::Border;
-use iced::Theme;
-use iced::widget::button;
-
-use super::variables::COLOR_BODY_BACKGROUND;
-use super::variables::COLOR_BORDER;
-
-pub fn style_button_primary(theme: &Theme, status: button::Status) -> button::Style {
-    let base = button::primary(theme, status);
-
-    button::Style {
-        background: Some(Background::Color(COLOR_BORDER)),
-        text_color: COLOR_BODY_BACKGROUND,
-        border: Border {
-            color: COLOR_BORDER,
-            width: 0.0,
-            radius: 0.into(),
-        },
-        ..base
-    }
-}
diff --git a/paddler_gui/src/ui/style_card_container.rs b/paddler_gui/src/ui/style_card_container.rs
deleted file mode 100644
index 5183d3047..000000000
--- a/paddler_gui/src/ui/style_card_container.rs
+++ /dev/null
@@ -1,16 +0,0 @@
-use iced::Background;
-use iced::Theme;
-use iced::widget::container;
-
-use super::variables::BORDER_OUTLINE;
-use super::variables::COLOR_BODY_BACKGROUND;
-
-pub fn style_card_container(theme: &Theme) -> container::Style {
-    let base = container::transparent(theme);
-
-    container::Style {
-        background: Some(Background::Color(COLOR_BODY_BACKGROUND)),
-        border: BORDER_OUTLINE,
-        ..base
-    }
-}
diff --git a/paddler_gui/src/ui/style_download_progress_bar.rs b/paddler_gui/src/ui/style_download_progress_bar.rs
deleted file mode 100644
index 975a0b4c0..000000000
--- a/paddler_gui/src/ui/style_download_progress_bar.rs
+++ /dev/null
@@ -1,15 +0,0 @@
-use iced::Background;
-use iced::Theme;
-use iced::widget::progress_bar;
-
-use super::variables::BORDER_OUTLINE;
-use super::variables::COLOR_BODY_BACKGROUND;
-use super::variables::COLOR_BORDER;
-
-pub const fn style_download_progress_bar(_theme: &Theme) -> progress_bar::Style {
-    progress_bar::Style {
-        background: Background::Color(COLOR_BODY_BACKGROUND),
-        bar: Background::Color(COLOR_BORDER),
-        border: BORDER_OUTLINE,
-    }
-}
diff --git a/paddler_gui/src/ui/style_field_checkbox.rs b/paddler_gui/src/ui/style_field_checkbox.rs
deleted file mode 100644
index c4393ceab..000000000
--- a/paddler_gui/src/ui/style_field_checkbox.rs
+++ /dev/null
@@ -1,68 +0,0 @@
-use iced::Background;
-use iced::Theme;
-use iced::widget::checkbox;
-
-use super::variables::BORDER_OUTLINE;
-use super::variables::COLOR_BODY_BACKGROUND;
-use super::variables::COLOR_BODY_FONT;
-use super::variables::COLOR_BORDER;
-
-pub const fn style_field_checkbox(_theme: &Theme, status: checkbox::Status) -> checkbox::Style {
-    let is_checked = matches!(
-        status,
-        checkbox::Status::Active { is_checked: true }
-            | checkbox::Status::Hovered { is_checked: true }
-            | checkbox::Status::Disabled { is_checked: true }
-    );
-
-    let background = if is_checked {
-        COLOR_BORDER
-    } else {
-        COLOR_BODY_BACKGROUND
-    };
-    let icon_color = if is_checked {
-        COLOR_BODY_BACKGROUND
-    } else {
-        COLOR_BODY_FONT
-    };
-
-    checkbox::Style {
-        background: Background::Color(background),
-        border: BORDER_OUTLINE,
-        icon_color,
-        text_color: Some(COLOR_BODY_FONT),
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use iced::Background;
-    use iced::Theme;
-    use iced::widget::checkbox::Status;
-
-    use super::style_field_checkbox;
-    use crate::ui::variables::COLOR_BODY_BACKGROUND;
-    use crate::ui::variables::COLOR_BODY_FONT;
-    use crate::ui::variables::COLOR_BORDER;
-
-    #[test]
-    fn fills_a_checked_box_and_inverts_its_mark() {
-        let checked_style =
-            style_field_checkbox(&Theme::Light, Status::Hovered { is_checked: true });
-
-        assert_eq!(checked_style.background, Background::Color(COLOR_BORDER));
-        assert_eq!(checked_style.icon_color, COLOR_BODY_BACKGROUND);
-    }
-
-    #[test]
-    fn leaves_an_unchecked_box_blank() {
-        let unchecked_style =
-            style_field_checkbox(&Theme::Light, Status::Active { is_checked: false });
-
-        assert_eq!(
-            unchecked_style.background,
-            Background::Color(COLOR_BODY_BACKGROUND)
-        );
-        assert_eq!(unchecked_style.icon_color, COLOR_BODY_FONT);
-    }
-}
diff --git a/paddler_gui/src/ui/style_field_container.rs b/paddler_gui/src/ui/style_field_container.rs
deleted file mode 100644
index df18be3d5..000000000
--- a/paddler_gui/src/ui/style_field_container.rs
+++ /dev/null
@@ -1,19 +0,0 @@
-use iced::Shadow;
-use iced::Theme;
-use iced::Vector;
-use iced::widget::container;
-
-use super::variables::COLOR_BORDER;
-
-pub fn style_field_container(theme: &Theme) -> container::Style {
-    let base = container::transparent(theme);
-
-    container::Style {
-        shadow: Shadow {
-            color: COLOR_BORDER,
-            offset: Vector::new(4.0, 4.0),
-            blur_radius: 0.0,
-        },
-        ..base
-    }
-}
diff --git a/paddler_gui/src/ui/style_field_pick_list.rs b/paddler_gui/src/ui/style_field_pick_list.rs
deleted file mode 100644
index d043dd04a..000000000
--- a/paddler_gui/src/ui/style_field_pick_list.rs
+++ /dev/null
@@ -1,19 +0,0 @@
-use iced::Background;
-use iced::Theme;
-use iced::widget::pick_list;
-
-use super::variables::BORDER_OUTLINE;
-use super::variables::COLOR_BODY_BACKGROUND;
-use super::variables::COLOR_BODY_FONT;
-
-pub fn style_field_pick_list(theme: &Theme, status: pick_list::Status) -> pick_list::Style {
-    let base = pick_list::default(theme, status);
-
-    pick_list::Style {
-        text_color: COLOR_BODY_FONT,
-        placeholder_color: base.placeholder_color,
-        handle_color: COLOR_BODY_FONT,
-        background: Background::Color(COLOR_BODY_BACKGROUND),
-        border: BORDER_OUTLINE,
-    }
-}
diff --git a/paddler_gui/src/ui/style_field_pick_list_menu.rs b/paddler_gui/src/ui/style_field_pick_list_menu.rs
deleted file mode 100644
index bce4cb4a5..000000000
--- a/paddler_gui/src/ui/style_field_pick_list_menu.rs
+++ /dev/null
@@ -1,26 +0,0 @@
-use iced::Background;
-use iced::Border;
-use iced::Color;
-use iced::Theme;
-use iced::overlay::menu;
-
-use super::variables::COLOR_BODY_BACKGROUND;
-use super::variables::COLOR_BODY_FONT;
-use super::variables::COLOR_BORDER;
-
-pub fn style_field_pick_list_menu(theme: &Theme) -> menu::Style {
-    let base = menu::default(theme);
-
-    menu::Style {
-        background: Background::Color(COLOR_BODY_BACKGROUND),
-        border: Border {
-            color: COLOR_BORDER,
-            width: 1.0,
-            radius: 0.into(),
-        },
-        text_color: COLOR_BODY_FONT,
-        selected_text_color: COLOR_BODY_FONT,
-        selected_background: Background::Color(Color::from_rgb(0.9, 0.9, 0.9)),
-        ..base
-    }
-}
diff --git a/paddler_gui/src/ui/style_field_text_input.rs b/paddler_gui/src/ui/style_field_text_input.rs
deleted file mode 100644
index 7552324bc..000000000
--- a/paddler_gui/src/ui/style_field_text_input.rs
+++ /dev/null
@@ -1,20 +0,0 @@
-use iced::Background;
-use iced::Theme;
-use iced::widget::text_input;
-
-use super::variables::BORDER_OUTLINE;
-use super::variables::COLOR_BODY_BACKGROUND;
-use super::variables::COLOR_BODY_FONT;
-
-pub fn style_field_text_input(theme: &Theme, status: text_input::Status) -> text_input::Style {
-    let base = text_input::default(theme, status);
-
-    text_input::Style {
-        background: Background::Color(COLOR_BODY_BACKGROUND),
-        border: BORDER_OUTLINE,
-        icon: COLOR_BODY_FONT,
-        placeholder: base.placeholder,
-        value: COLOR_BODY_FONT,
-        ..base
-    }
-}
diff --git a/paddler_gui/src/ui/style_status_indicator.rs b/paddler_gui/src/ui/style_status_indicator.rs
deleted file mode 100644
index b80b49c75..000000000
--- a/paddler_gui/src/ui/style_status_indicator.rs
+++ /dev/null
@@ -1,19 +0,0 @@
-use iced::Background;
-use iced::Border;
-use iced::Color;
-use iced::Theme;
-use iced::widget::container;
-
-pub fn style_status_indicator(theme: &Theme) -> container::Style {
-    let base = container::transparent(theme);
-
-    container::Style {
-        background: Some(Background::Color(Color::from_rgb8(0xEE, 0xFF, 0xEE))),
-        border: Border {
-            color: Color::from_rgb8(0xCC, 0xDD, 0xCC),
-            width: 2.0,
-            radius: 8.into(),
-        },
-        ..base
-    }
-}
diff --git a/paddler_gui/src/ui/variables.rs b/paddler_gui/src/ui/variables.rs
deleted file mode 100644
index 5ede1e084..000000000
--- a/paddler_gui/src/ui/variables.rs
+++ /dev/null
@@ -1,53 +0,0 @@
-use iced::Border;
-use iced::Color;
-use iced::border::Radius;
-
-const FONT_SIZE_L1: f32 = 1.5 * FONT_SIZE_BASE;
-
-pub const FONT_SIZE_BASE: f32 = 14.0;
-
-pub const FONT_SIZE_L2: f32 = 1.5 * FONT_SIZE_L1;
-
-pub const FORM_WIDTH: f32 = 400.0;
-
-pub const SPACING_BASE: f32 = 16.0;
-
-pub const SPACING_2X: f32 = 2.0 * SPACING_BASE;
-
-pub const SPACING_HALF: f32 = 0.5 * SPACING_BASE;
-
-pub const COLOR_BODY_BACKGROUND: Color = Color::WHITE;
-
-pub const COLOR_BODY_FONT: Color = Color {
-    r: 0.067,
-    g: 0.067,
-    b: 0.067,
-    a: 1.0,
-};
-
-pub const COLOR_AGENT_BACKGROUND: Color = Color {
-    r: 250.0 / 255.0,
-    g: 240.0 / 255.0,
-    b: 230.0 / 255.0,
-    a: 1.0,
-};
-
-pub const COLOR_BORDER: Color = Color::BLACK;
-
-pub const COLOR_ERROR: Color = Color {
-    r: 204.0 / 255.0,
-    g: 51.0 / 255.0,
-    b: 51.0 / 255.0,
-    a: 1.0,
-};
-
-pub const BORDER_OUTLINE: Border = Border {
-    color: COLOR_BORDER,
-    width: 2.0,
-    radius: Radius {
-        top_left: 0.0,
-        top_right: 0.0,
-        bottom_right: 0.0,
-        bottom_left: 0.0,
-    },
-};
diff --git a/paddler_gui/src/ui/view_agent_card.rs b/paddler_gui/src/ui/view_agent_card.rs
deleted file mode 100644
index 5fca39495..000000000
--- a/paddler_gui/src/ui/view_agent_card.rs
+++ /dev/null
@@ -1,217 +0,0 @@
-use std::path::Path;
-
-use iced::Element;
-use iced::Fill;
-use iced::widget::column;
-use iced::widget::container;
-use iced::widget::progress_bar;
-use iced::widget::row;
-use iced::widget::text;
-
-use paddler_messaging::agent_state_application_status::AgentStateApplicationStatus;
-use paddler_messaging::agent_status::AgentStatus;
-use paddler_messaging::model_download_status::ModelDownloadStatus;
-
-use super::font::BOLD;
-use super::font::REGULAR;
-use super::style_agent_container::style_agent_container;
-use super::style_download_progress_bar::style_download_progress_bar;
-use super::variables::SPACING_BASE;
-use super::variables::SPACING_HALF;
-
-fn display_last_path_part(path: &str) -> String {
-    Path::new(path)
-        .file_name()
-        .and_then(|name| name.to_str())
-        .unwrap_or(path)
-        .to_owned()
-}
-
-fn view_download_progress_bar<'card, TMessage: 'static>(
-    downloaded_bytes: u64,
-    total_bytes: u64,
-) -> Element<'card, TMessage> {
-    #[expect(
-        clippy::cast_precision_loss,
-        reason = "download sizes fit in f32 mantissa"
-    )]
-    progress_bar(0.0..=total_bytes as f32, downloaded_bytes as f32)
-        .girth(12)
-        .style(style_download_progress_bar)
-        .into()
-}
-
-fn agent_status_label(status: &AgentStatus) -> String {
-    match &status.download_status {
-        ModelDownloadStatus::Downloading {
-            downloaded_bytes,
-            total_bytes,
-            ..
-        } => {
-            #[expect(
-                clippy::cast_precision_loss,
-                reason = "download sizes fit in f32 mantissa"
-            )]
-            let percentage = (*downloaded_bytes as f32 / *total_bytes as f32) * 100.0;
-
-            format!("Downloading ({percentage:.0}%)")
-        }
-        ModelDownloadStatus::DownloadingWithUnknownSize { .. } => "Downloading...".to_owned(),
-        ModelDownloadStatus::NotDownloading if status.model_path.is_none() => {
-            "Waiting for model...".to_owned()
-        }
-        ModelDownloadStatus::NotDownloading => match &status.state_application_status {
-            AgentStateApplicationStatus::Applied => "OK".to_owned(),
-            AgentStateApplicationStatus::Fresh => "Pending".to_owned(),
-            AgentStateApplicationStatus::AttemptedAndRetrying => "Retrying".to_owned(),
-            AgentStateApplicationStatus::Stuck => "Retrying, but seems stuck?".to_owned(),
-            AgentStateApplicationStatus::AttemptedAndNotAppliable => "Needs your help".to_owned(),
-        },
-    }
-}
-
-fn model_label(status: &AgentStatus) -> String {
-    status.model_path.as_ref().map_or_else(
-        || "No model loaded".to_owned(),
-        |path| display_last_path_part(path),
-    )
-}
-
-pub fn view_agent_card<'card, TMessage: 'static>(
-    name: Option<&'card str>,
-    slots_processing: u64,
-    status: &'card AgentStatus,
-) -> Element<'card, TMessage> {
-    let name_cell =
-        container(name.map_or_else(|| text(""), |agent_name| text(agent_name).font(BOLD)))
-            .width(Fill);
-
-    let download_cell = match &status.download_status {
-        ModelDownloadStatus::Downloading {
-            downloaded_bytes,
-            total_bytes,
-            ..
-        } => view_download_progress_bar(*downloaded_bytes, *total_bytes),
-        ModelDownloadStatus::DownloadingWithUnknownSize { .. }
-        | ModelDownloadStatus::NotDownloading => text(model_label(status)).font(REGULAR).into(),
-    };
-    let status_label = agent_status_label(status);
-
-    let name_row = row![name_cell, download_cell];
-
-    let mut status_row_left = column![].spacing(SPACING_HALF);
-
-    status_row_left = status_row_left.push(text(format!("Status: {status_label}")).font(REGULAR));
-
-    if !status.issues.is_empty() {
-        status_row_left =
-            status_row_left.push(text(format!("{} issues", status.issues.len())).font(REGULAR));
-    }
-
-    let slots_label = format!(
-        "{slots_processing}/{}/{}",
-        status.slots_total, status.desired_slots_total,
-    );
-
-    let status_row_content = row![
-        container(status_row_left).width(Fill),
-        text(format!("Slots: {slots_label}")).font(REGULAR),
-    ];
-
-    let card_content = column![name_row, status_row_content,].spacing(SPACING_BASE);
-
-    container(card_content)
-        .width(Fill)
-        .padding(SPACING_BASE)
-        .style(style_agent_container)
-        .into()
-}
-
-#[cfg(test)]
-mod tests {
-    use paddler_messaging::agent_state_application_status::AgentStateApplicationStatus;
-    use paddler_messaging::agent_status::AgentStatus;
-    use paddler_messaging::model_download_status::ModelDownloadStatus;
-
-    use super::agent_status_label;
-    use super::model_label;
-
-    fn status_with_model(state_application_status: AgentStateApplicationStatus) -> AgentStatus {
-        AgentStatus {
-            model_path: Some("/models/qwen3.gguf".to_owned()),
-            state_application_status,
-            ..AgentStatus::default()
-        }
-    }
-
-    #[test]
-    fn labels_a_download_of_known_size_with_its_progress() {
-        let status = AgentStatus {
-            download_status: ModelDownloadStatus::Downloading {
-                downloaded_bytes: 250,
-                model_path: "owner/repo/main/model.gguf".to_owned(),
-                total_bytes: 1000,
-            },
-            ..AgentStatus::default()
-        };
-
-        assert_eq!(agent_status_label(&status), "Downloading (25%)");
-    }
-
-    #[test]
-    fn labels_a_download_of_unknown_size_without_progress() {
-        let status = AgentStatus {
-            download_status: ModelDownloadStatus::DownloadingWithUnknownSize {
-                downloaded_bytes: 250,
-                model_path: "https://example.com/model.gguf".to_owned(),
-            },
-            ..AgentStatus::default()
-        };
-
-        assert_eq!(agent_status_label(&status), "Downloading...");
-    }
-
-    #[test]
-    fn labels_an_agent_without_a_model_as_waiting() {
-        assert_eq!(
-            agent_status_label(&AgentStatus::default()),
-            "Waiting for model..."
-        );
-    }
-
-    #[test]
-    fn labels_each_state_application_status_of_a_loaded_agent() {
-        assert_eq!(
-            [
-                AgentStateApplicationStatus::Applied,
-                AgentStateApplicationStatus::Fresh,
-                AgentStateApplicationStatus::AttemptedAndRetrying,
-                AgentStateApplicationStatus::Stuck,
-                AgentStateApplicationStatus::AttemptedAndNotAppliable,
-            ]
-            .map(|state_application_status| {
-                agent_status_label(&status_with_model(state_application_status))
-            }),
-            [
-                "OK",
-                "Pending",
-                "Retrying",
-                "Retrying, but seems stuck?",
-                "Needs your help"
-            ]
-        );
-    }
-
-    #[test]
-    fn names_the_loaded_model_by_its_file_name() {
-        assert_eq!(
-            model_label(&status_with_model(AgentStateApplicationStatus::Applied)),
-            "qwen3.gguf"
-        );
-    }
-
-    #[test]
-    fn names_a_missing_model() {
-        assert_eq!(model_label(&AgentStatus::default()), "No model loaded");
-    }
-}
diff --git a/paddler_gui/src/ui/view_agent_running.rs b/paddler_gui/src/ui/view_agent_running.rs
deleted file mode 100644
index 2c129f410..000000000
--- a/paddler_gui/src/ui/view_agent_running.rs
+++ /dev/null
@@ -1,51 +0,0 @@
-use iced::Center;
-use iced::Element;
-use iced::Fill;
-use iced::widget::column;
-use iced::widget::container;
-use iced::widget::row;
-use iced::widget::text;
-
-use paddler_messaging::balancer_connection::BalancerConnection;
-
-use super::font::BOLD;
-use super::font::REGULAR;
-use super::variables::FONT_SIZE_L2;
-use super::variables::SPACING_2X;
-use super::variables::SPACING_BASE;
-use super::view_agent_card::view_agent_card;
-use super::view_stop_button::view_stop_button;
-use crate::agent_running_data::AgentRunningData;
-use crate::agent_running_message::AgentRunningMessage;
-
-impl AgentRunningData {
-    #[must_use]
-    pub fn view(&self) -> Element<'_, AgentRunningMessage> {
-        let disconnect_button =
-            view_stop_button("Disconnect", Some(AgentRunningMessage::Disconnect));
-
-        let connection_status = match self.balancer_connection {
-            BalancerConnection::Connected => text(format!(
-                "Connected to the cluster at {}",
-                self.balancer_address
-            ))
-            .font(REGULAR),
-            BalancerConnection::Connecting => text("Connecting to the cluster...").font(REGULAR),
-        };
-
-        let status_row = container(
-            row![container(connection_status).width(Fill), disconnect_button,].align_y(Center),
-        )
-        .padding([0.0, SPACING_BASE]);
-
-        column![
-            container(text("Your agent").size(FONT_SIZE_L2).font(BOLD))
-                .padding([0.0, SPACING_BASE]),
-            container(text("Agent details").font(BOLD)).padding([0.0, SPACING_BASE]),
-            view_agent_card(self.name.as_deref(), self.slots_processing, &self.status),
-            status_row,
-        ]
-        .spacing(SPACING_2X)
-        .into()
-    }
-}
diff --git a/paddler_gui/src/ui/view_error_message.rs b/paddler_gui/src/ui/view_error_message.rs
deleted file mode 100644
index 9c52eeda2..000000000
--- a/paddler_gui/src/ui/view_error_message.rs
+++ /dev/null
@@ -1,13 +0,0 @@
-use iced::Element;
-use iced::widget::container;
-use iced::widget::text;
-
-use super::font::REGULAR;
-use super::variables::COLOR_ERROR;
-use super::variables::SPACING_BASE;
-
-pub fn view_error_message<TMessage: 'static>(error: &str) -> Element<'static, TMessage> {
-    container(text(error.to_owned()).font(REGULAR).color(COLOR_ERROR))
-        .padding([0.0, SPACING_BASE])
-        .into()
-}
diff --git a/paddler_gui/src/ui/view_form_field.rs b/paddler_gui/src/ui/view_form_field.rs
deleted file mode 100644
index b093ecde3..000000000
--- a/paddler_gui/src/ui/view_form_field.rs
+++ /dev/null
@@ -1,29 +0,0 @@
-use iced::Element;
-use iced::widget::column;
-use iced::widget::container;
-use iced::widget::text;
-
-use super::font::BOLD;
-use super::style_field_container::style_field_container;
-use super::variables::SPACING_BASE;
-use super::variables::SPACING_HALF;
-use super::view_error_message::view_error_message;
-use crate::form_field_error::FormFieldError;
-
-pub fn view_form_field<'element, TMessage: 'static>(
-    label: &str,
-    input: Element<'element, TMessage>,
-    error: Option<&FormFieldError>,
-) -> Element<'element, TMessage> {
-    let mut field = column![
-        container(text(label.to_owned()).font(BOLD)).padding([0.0, SPACING_BASE]),
-        container(input).style(style_field_container),
-    ]
-    .spacing(SPACING_HALF);
-
-    if let Some(error) = error {
-        field = field.push(view_error_message(&error.to_string()));
-    }
-
-    field.into()
-}
diff --git a/paddler_gui/src/ui/view_home.rs b/paddler_gui/src/ui/view_home.rs
deleted file mode 100644
index b57ea6d73..000000000
--- a/paddler_gui/src/ui/view_home.rs
+++ /dev/null
@@ -1,74 +0,0 @@
-use std::sync::LazyLock;
-
-use iced::Center;
-use iced::Element;
-use iced::widget::button;
-use iced::widget::column;
-use iced::widget::container;
-use iced::widget::image;
-use iced::widget::image::Handle as ImageHandle;
-use iced::widget::row;
-use iced::widget::text;
-
-use super::font::BOLD;
-use super::style_button_primary::style_button_primary;
-use super::variables::FONT_SIZE_L2;
-use super::variables::SPACING_2X;
-use super::variables::SPACING_BASE;
-use super::variables::SPACING_HALF;
-use super::view_error_message::view_error_message;
-use crate::home_data::HomeData;
-use crate::home_message::HomeMessage;
-
-static CREATE_CLUSTER_IMAGE: LazyLock<ImageHandle> = LazyLock::new(|| {
-    ImageHandle::from_bytes(
-        include_bytes!("../../../resources/images/create_a_cluster.png").as_slice(),
-    )
-});
-
-static JOIN_CLUSTER_IMAGE: LazyLock<ImageHandle> = LazyLock::new(|| {
-    ImageHandle::from_bytes(
-        include_bytes!("../../../resources/images/join_a_cluster.png").as_slice(),
-    )
-});
-
-impl HomeData {
-    pub fn view(&self) -> Element<'_, HomeMessage> {
-        let create_image = image(CREATE_CLUSTER_IMAGE.clone()).width(200).height(200);
-
-        let join_image = image(JOIN_CLUSTER_IMAGE.clone()).width(200).height(200);
-
-        let start_button = button(text("Start a cluster").font(BOLD))
-            .padding([SPACING_HALF, SPACING_BASE])
-            .style(style_button_primary)
-            .on_press(HomeMessage::StartBalancer);
-
-        let join_button = button(text("Join a cluster").font(BOLD))
-            .padding([SPACING_HALF, SPACING_BASE])
-            .style(style_button_primary)
-            .on_press(HomeMessage::JoinBalancer);
-
-        let start_column = column![create_image, start_button]
-            .spacing(SPACING_BASE)
-            .align_x(Center);
-
-        let join_column = column![join_image, join_button]
-            .spacing(SPACING_BASE)
-            .align_x(Center);
-
-        let options_row = row![start_column, join_column].spacing(SPACING_2X);
-
-        let mut content = column![
-            container(text("Paddler App").size(FONT_SIZE_L2).font(BOLD))
-                .padding([0.0, SPACING_BASE]),
-            container(options_row).align_x(Center),
-        ]
-        .spacing(SPACING_2X);
-
-        if let Some(error) = &self.error {
-            content = content.push(view_error_message(&error.to_string()));
-        }
-
-        content.into()
-    }
-}
diff --git a/paddler_gui/src/ui/view_join_balancer_form.rs b/paddler_gui/src/ui/view_join_balancer_form.rs
deleted file mode 100644
index 9d181dbdf..000000000
--- a/paddler_gui/src/ui/view_join_balancer_form.rs
+++ /dev/null
@@ -1,79 +0,0 @@
-use iced::Center;
-use iced::Element;
-use iced::alignment::Horizontal;
-use iced::widget::button;
-use iced::widget::column;
-use iced::widget::container;
-use iced::widget::row;
-use iced::widget::text;
-use iced::widget::text_input;
-
-use super::font::BOLD;
-use super::style_button_primary::style_button_primary;
-use super::style_field_text_input::style_field_text_input;
-use super::variables::FONT_SIZE_L2;
-use super::variables::FORM_WIDTH;
-use super::variables::SPACING_2X;
-use super::variables::SPACING_BASE;
-use super::variables::SPACING_HALF;
-use super::view_form_field::view_form_field;
-use crate::address_placeholder::ADDRESS_PLACEHOLDER;
-use crate::join_balancer_form_data::JoinBalancerFormData;
-use crate::join_balancer_form_message::JoinBalancerFormMessage;
-
-impl JoinBalancerFormData {
-    pub fn view(&self) -> Element<'_, JoinBalancerFormMessage> {
-        let confirm_button = button(text("Connect").font(BOLD))
-            .padding([SPACING_HALF, SPACING_BASE])
-            .style(style_button_primary)
-            .on_press(JoinBalancerFormMessage::Connect);
-
-        let cancel_button = button(text("Cancel").font(BOLD))
-            .style(button::text)
-            .on_press(JoinBalancerFormMessage::Cancel);
-
-        let balancer_address_input = text_input(ADDRESS_PLACEHOLDER, &self.balancer_address)
-            .on_input(JoinBalancerFormMessage::SetBalancerAddress)
-            .padding(SPACING_BASE)
-            .style(style_field_text_input)
-            .into();
-
-        let agent_name_input = text_input("my-agent", &self.agent_name)
-            .on_input(JoinBalancerFormMessage::SetAgentName)
-            .padding(SPACING_BASE)
-            .style(style_field_text_input)
-            .into();
-
-        let slots_input = text_input("e.g. 1", &self.slots_count)
-            .on_input(JoinBalancerFormMessage::SetSlotsCount)
-            .padding(SPACING_BASE)
-            .style(style_field_text_input)
-            .into();
-
-        column![
-            container(text("Join a cluster").size(FONT_SIZE_L2).font(BOLD))
-                .padding([0.0, SPACING_BASE]),
-            container(
-                column![
-                    view_form_field(
-                        "Cluster address",
-                        balancer_address_input,
-                        self.balancer_address_error.as_ref()
-                    ),
-                    view_form_field("Agent name (optional)", agent_name_input, None),
-                    view_form_field("Slots", slots_input, self.slots_error.as_ref()),
-                    container(
-                        row![cancel_button, confirm_button]
-                            .align_y(Center)
-                            .spacing(SPACING_BASE),
-                    )
-                    .align_x(Horizontal::Right),
-                ]
-                .spacing(SPACING_2X),
-            )
-            .width(FORM_WIDTH),
-        ]
-        .spacing(SPACING_2X)
-        .into()
-    }
-}
diff --git a/paddler_gui/src/ui/view_running_balancer.rs b/paddler_gui/src/ui/view_running_balancer.rs
deleted file mode 100644
index e20b20b0a..000000000
--- a/paddler_gui/src/ui/view_running_balancer.rs
+++ /dev/null
@@ -1,199 +0,0 @@
-use iced::Center;
-use iced::Element;
-use iced::Fill;
-use iced::widget::button;
-use iced::widget::column;
-use iced::widget::container;
-use iced::widget::row;
-use iced::widget::svg;
-use iced::widget::svg::Handle as SvgHandle;
-use iced::widget::text;
-
-use paddler_messaging::agent_desired_model::AgentDesiredModel;
-
-use super::font::BOLD;
-use super::font::REGULAR;
-use super::style_card_container::style_card_container;
-use super::style_status_indicator::style_status_indicator;
-use super::variables::FONT_SIZE_L2;
-use super::variables::SPACING_2X;
-use super::variables::SPACING_BASE;
-use super::variables::SPACING_HALF;
-use super::view_agent_card::view_agent_card;
-use super::view_stop_button::view_stop_button;
-use crate::running_balancer_data::RunningBalancerData;
-use crate::running_balancer_message::RunningBalancerMessage;
-
-fn format_desired_model(desired_model: &AgentDesiredModel) -> String {
-    match desired_model {
-        AgentDesiredModel::HuggingFace(reference) => {
-            format!(
-                "HuggingFace {}/{} ({})",
-                reference.repo_id, reference.filename, reference.revision,
-            )
-        }
-        AgentDesiredModel::LocalToAgent(path) => format!("Local: {path}"),
-        AgentDesiredModel::Url(reference) => format!("URL: {}", reference.url),
-        AgentDesiredModel::None => "(not set)".to_owned(),
-    }
-}
-
-impl RunningBalancerData {
-    pub fn view(&self) -> Element<'_, RunningBalancerMessage> {
-        let copy_icon = svg(SvgHandle::from_memory(
-            include_bytes!("../../../resources/icons/copy.svg").as_slice(),
-        ))
-        .width(16)
-        .height(16);
-
-        let desired_model_label = format_desired_model(&self.snapshot.balancer_desired_state.model);
-        let applied_model_label = format_desired_model(
-            &self
-                .snapshot
-                .balancer_applicable_state
-                .agent_desired_state
-                .model,
-        );
-
-        let address_row = container(
-            column![
-                row![
-                    container(
-                        text(format!("Cluster address: {}", self.addresses.management))
-                            .font(REGULAR)
-                    )
-                    .width(Fill),
-                    button(
-                        row![copy_icon, text("Copy address").font(BOLD)]
-                            .spacing(SPACING_HALF)
-                            .align_y(Center),
-                    )
-                    .style(button::text)
-                    .on_press(RunningBalancerMessage::CopyToClipboard(
-                        self.addresses.management.to_string()
-                    )),
-                ]
-                .align_y(Center),
-                text(format!("Configured model: {desired_model_label}")).font(REGULAR),
-                text(format!("Applied model: {applied_model_label}")).font(REGULAR),
-            ]
-            .spacing(SPACING_HALF)
-            .padding(SPACING_BASE),
-        )
-        .style(style_card_container);
-
-        let status_indicator = container("")
-            .width(16)
-            .height(16)
-            .style(style_status_indicator);
-
-        let stop_button = if self.stopping {
-            view_stop_button("Stopping...", None)
-        } else {
-            view_stop_button("Stop cluster", Some(RunningBalancerMessage::Stop))
-        };
-
-        let status_row = container(
-            row![
-                container(
-                    row![text("Cluster is running").font(REGULAR), status_indicator,]
-                        .spacing(SPACING_HALF)
-                        .align_y(Center),
-                )
-                .width(Fill),
-                stop_button,
-            ]
-            .align_y(Center),
-        )
-        .padding([SPACING_HALF, SPACING_BASE]);
-
-        let mut content = column![
-            container(text("Your cluster").size(FONT_SIZE_L2).font(BOLD))
-                .padding([0.0, SPACING_BASE]),
-            address_row,
-        ]
-        .spacing(SPACING_2X);
-
-        if let Some(address) = self.addresses.web_admin_panel {
-            let open_in_new_icon = svg(SvgHandle::from_memory(
-                include_bytes!("../../../resources/icons/open_in_new.svg").as_slice(),
-            ))
-            .width(16)
-            .height(16);
-
-            content = content.push(
-                container(
-                    row![
-                        container(text(format!("Web admin panel: {address}")).font(REGULAR))
-                            .width(Fill),
-                        button(
-                            row![open_in_new_icon, text("Open in browser").font(BOLD)]
-                                .spacing(SPACING_HALF)
-                                .align_y(Center),
-                        )
-                        .style(button::text)
-                        .on_press(RunningBalancerMessage::OpenUrl(format!("http://{address}"))),
-                    ]
-                    .align_y(Center)
-                    .padding(SPACING_BASE),
-                )
-                .style(style_card_container),
-            );
-        }
-
-        content = content.push(status_row);
-        content = content
-            .push(container(text("Connected agents").font(BOLD)).padding([0.0, SPACING_BASE]));
-
-        if self.snapshot.agent_snapshots.is_empty() {
-            content = content.push(
-                container(text("Waiting for agents to connect...").font(REGULAR))
-                    .padding([0.0, SPACING_BASE]),
-            );
-        } else {
-            for agent_snapshot in &self.snapshot.agent_snapshots {
-                content = content.push(view_agent_card(
-                    agent_snapshot.name.as_deref(),
-                    agent_snapshot.slots_processing,
-                    &agent_snapshot.status,
-                ));
-            }
-        }
-
-        content.into()
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use paddler_messaging::agent_desired_model::AgentDesiredModel;
-    use paddler_messaging::huggingface_model_reference::HuggingFaceModelReference;
-    use paddler_messaging::url_model_reference::UrlModelReference;
-
-    use super::format_desired_model;
-
-    #[test]
-    fn describes_each_model_source() {
-        assert_eq!(
-            [
-                AgentDesiredModel::HuggingFace(HuggingFaceModelReference {
-                    repo_id: "unsloth/Qwen3-0.6B-GGUF".to_owned(),
-                    filename: "Qwen3-0.6B-Q8_0.gguf".to_owned(),
-                    revision: "main".to_owned(),
-                }),
-                AgentDesiredModel::LocalToAgent("/models/qwen3.gguf".to_owned()),
-                AgentDesiredModel::Url(UrlModelReference {
-                    url: "https://example.com/qwen3.gguf".to_owned(),
-                }),
-                AgentDesiredModel::None,
-            ]
-            .map(|desired_model| format_desired_model(&desired_model)),
-            [
-                "HuggingFace unsloth/Qwen3-0.6B-GGUF/Qwen3-0.6B-Q8_0.gguf (main)",
-                "Local: /models/qwen3.gguf",
-                "URL: https://example.com/qwen3.gguf",
-                "(not set)",
-            ]
-        );
-    }
-}
diff --git a/paddler_gui/src/ui/view_start_balancer_form.rs b/paddler_gui/src/ui/view_start_balancer_form.rs
deleted file mode 100644
index b03d75731..000000000
--- a/paddler_gui/src/ui/view_start_balancer_form.rs
+++ /dev/null
@@ -1,146 +0,0 @@
-use iced::Center;
-use iced::Element;
-use iced::Fill;
-use iced::Padding;
-use iced::widget::button;
-use iced::widget::checkbox;
-use iced::widget::column;
-use iced::widget::container;
-use iced::widget::pick_list;
-use iced::widget::row;
-use iced::widget::text;
-use iced::widget::text_input;
-
-use super::font::BOLD;
-use super::font::REGULAR;
-use super::style_button_primary::style_button_primary;
-use super::style_field_checkbox::style_field_checkbox;
-use super::style_field_pick_list::style_field_pick_list;
-use super::style_field_pick_list_menu::style_field_pick_list_menu;
-use super::style_field_text_input::style_field_text_input;
-use super::variables::FONT_SIZE_BASE;
-use super::variables::FONT_SIZE_L2;
-use super::variables::FORM_WIDTH;
-use super::variables::SPACING_2X;
-use super::variables::SPACING_BASE;
-use super::variables::SPACING_HALF;
-use super::view_form_field::view_form_field;
-use crate::address_placeholder::ADDRESS_PLACEHOLDER;
-use crate::balancer_launch::BalancerLaunch;
-use crate::model_preset::ModelPreset;
-use crate::start_balancer_form_data::StartBalancerFormData;
-use crate::start_balancer_form_message::StartBalancerFormMessage;
-
-impl StartBalancerFormData {
-    pub fn view(&self) -> Element<'_, StartBalancerFormMessage> {
-        let confirm_button = match self.launch {
-            BalancerLaunch::Starting(_) => button(text("Starting...").font(BOLD))
-                .padding([SPACING_HALF, SPACING_BASE])
-                .style(style_button_primary),
-            BalancerLaunch::NotRequested => button(text("Start a cluster").font(BOLD))
-                .padding([SPACING_HALF, SPACING_BASE])
-                .style(style_button_primary)
-                .on_press(StartBalancerFormMessage::Confirm),
-        };
-
-        let cancel_button = button(text("Cancel").font(BOLD))
-            .padding([SPACING_HALF, SPACING_BASE])
-            .style(button::text)
-            .on_press(StartBalancerFormMessage::Cancel);
-
-        let balancer_address_input = text_input(ADDRESS_PLACEHOLDER, &self.balancer_address)
-            .on_input(StartBalancerFormMessage::SetBalancerAddress)
-            .padding(SPACING_BASE)
-            .style(style_field_text_input)
-            .into();
-
-        let inference_address_input = text_input(ADDRESS_PLACEHOLDER, &self.inference_address)
-            .on_input(StartBalancerFormMessage::SetInferenceAddress)
-            .padding(SPACING_BASE)
-            .style(style_field_text_input)
-            .into();
-
-        let web_admin_panel_address_input = text_input(
-            &self.web_admin_panel_address_placeholder,
-            &self.web_admin_panel_address,
-        )
-        .on_input(StartBalancerFormMessage::SetWebAdminPanelAddress)
-        .padding(SPACING_BASE)
-        .style(style_field_text_input)
-        .into();
-
-        let model_input: Element<'_, StartBalancerFormMessage> = if self.add_model_later {
-            text_input("Model will be added later", "")
-                .padding(SPACING_BASE)
-                .style(style_field_text_input)
-                .into()
-        } else {
-            pick_list(
-                ModelPreset::ALL,
-                self.selected_model,
-                StartBalancerFormMessage::SelectModel,
-            )
-            .placeholder("Choose a model")
-            .width(Fill)
-            .padding(SPACING_BASE)
-            .style(style_field_pick_list)
-            .menu_style(style_field_pick_list_menu)
-            .into()
-        };
-
-        let add_model_later_checkbox: Element<'_, StartBalancerFormMessage> = container(
-            checkbox(self.add_model_later)
-                .label("Add a model later")
-                .font(REGULAR)
-                .size(FONT_SIZE_BASE)
-                .text_size(FONT_SIZE_BASE)
-                .on_toggle(StartBalancerFormMessage::ToggleAddModelLater)
-                .style(style_field_checkbox),
-        )
-        .padding(Padding {
-            top: 4.0,
-            right: SPACING_BASE,
-            bottom: 0.0,
-            left: SPACING_BASE,
-        })
-        .into();
-
-        let model_field = column![
-            view_form_field("Model", model_input, self.model_error.as_ref()),
-            add_model_later_checkbox,
-        ]
-        .spacing(SPACING_HALF);
-
-        column![
-            container(text("Start a cluster").size(FONT_SIZE_L2).font(BOLD))
-                .padding([0.0, SPACING_BASE]),
-            container(
-                column![
-                    view_form_field(
-                        "Cluster address",
-                        balancer_address_input,
-                        self.balancer_address_error.as_ref()
-                    ),
-                    view_form_field(
-                        "Inference address",
-                        inference_address_input,
-                        self.inference_address_error.as_ref()
-                    ),
-                    view_form_field(
-                        "Web admin panel (optional)",
-                        web_admin_panel_address_input,
-                        self.web_admin_panel_address_error.as_ref()
-                    ),
-                    model_field,
-                    row![cancel_button, confirm_button]
-                        .align_y(Center)
-                        .spacing(SPACING_BASE),
-                ]
-                .spacing(SPACING_2X),
-            )
-            .width(FORM_WIDTH),
-        ]
-        .spacing(SPACING_2X)
-        .into()
-    }
-}
diff --git a/paddler_gui/src/ui/view_stop_button.rs b/paddler_gui/src/ui/view_stop_button.rs
deleted file mode 100644
index b684087e6..000000000
--- a/paddler_gui/src/ui/view_stop_button.rs
+++ /dev/null
@@ -1,33 +0,0 @@
-use iced::Center;
-use iced::Element;
-use iced::widget::button;
-use iced::widget::row;
-use iced::widget::svg;
-use iced::widget::svg::Handle as SvgHandle;
-use iced::widget::text;
-
-use super::font::BOLD;
-use super::style_button_disconnect::style_button_disconnect;
-use super::variables::SPACING_BASE;
-use super::variables::SPACING_HALF;
-
-pub fn view_stop_button<TMessage: Clone + 'static>(
-    label: &str,
-    on_press: Option<TMessage>,
-) -> Element<'static, TMessage> {
-    let stop_icon = svg(SvgHandle::from_memory(
-        include_bytes!("../../../resources/icons/stop.svg").as_slice(),
-    ))
-    .width(16)
-    .height(16);
-
-    button(
-        row![stop_icon, text(label.to_owned()).font(BOLD)]
-            .spacing(SPACING_HALF)
-            .align_y(Center),
-    )
-    .padding([SPACING_HALF, SPACING_BASE])
-    .style(style_button_disconnect)
-    .on_press_maybe(on_press)
-    .into()
-}
diff --git a/paddler_gui_tests/Cargo.toml b/paddler_gui_tests/Cargo.toml
deleted file mode 100644
index 820989b16..000000000
--- a/paddler_gui_tests/Cargo.toml
+++ /dev/null
@@ -1,38 +0,0 @@
-[package]
-name = "paddler_gui_tests"
-autotests = false
-authors.workspace = true
-description = "Headless interface tests for the Paddler desktop application"
-edition.workspace = true
-homepage.workspace = true
-license.workspace = true
-publish = false
-repository.workspace = true
-version.workspace = true
-
-[[test]]
-name = "integration"
-path = "tests/main.rs"
-
-[features]
-default = []
-web_admin_panel = ["paddler_gui/web_admin_panel"]
-
-[dependencies]
-futures = { workspace = true }
-iced_test = { workspace = true }
-paddler_balancer = { workspace = true }
-paddler_gui = { workspace = true }
-paddler_test_cluster_harness = { workspace = true }
-thiserror = { workspace = true }
-
-[dev-dependencies]
-paddler_bootstrap = { workspace = true }
-paddler_messaging = { workspace = true }
-reqwest = { workspace = true }
-tokio = { workspace = true }
-tokio-util = { workspace = true }
-trzcina = { workspace = true }
-
-[lints]
-workspace = true
diff --git a/paddler_gui_tests/src/app_driver.rs b/paddler_gui_tests/src/app_driver.rs
deleted file mode 100644
index 0a766f49e..000000000
--- a/paddler_gui_tests/src/app_driver.rs
+++ /dev/null
@@ -1,136 +0,0 @@
-use iced_test::runtime::Task;
-use iced_test::simulator;
-
-use paddler_gui::app::App;
-use paddler_gui::home_message::HomeMessage;
-use paddler_gui::join_balancer_form_message::JoinBalancerFormMessage;
-use paddler_gui::message::Message;
-use paddler_gui::start_balancer_form_message::StartBalancerFormMessage;
-use paddler_test_cluster_harness::ephemeral_loopback_addr::EPHEMERAL_LOOPBACK_ADDR;
-
-use crate::gui_test_error::GuiTestError;
-use crate::started_cluster::StartedCluster;
-use crate::task_actions::TaskActions;
-
-pub struct AppDriver {
-    app: App,
-}
-
-impl AppDriver {
-    pub async fn open() -> Result<Self, GuiTestError> {
-        let (mut app, boot_task) = App::new();
-        let boot_message = TaskActions::of(boot_task)?.next_message().await?;
-        let _event_loop_ready_task = app.update(boot_message);
-
-        Ok(Self { app })
-    }
-
-    pub fn click(&mut self, label: &str) -> Result<Task<Message>, GuiTestError> {
-        self.messages_from_clicking(label).map(|messages| {
-            Task::batch(messages.into_iter().map(|message| self.app.update(message)))
-        })
-    }
-
-    pub async fn deliver_until<TPredicate>(
-        &mut self,
-        task_actions: &mut TaskActions,
-        is_awaited_message: TPredicate,
-    ) -> Result<(), GuiTestError>
-    where
-        TPredicate: Fn(&Message) -> bool,
-    {
-        loop {
-            let message = task_actions.next_message().await?;
-            let is_awaited = is_awaited_message(&message);
-            let _follow_up_task = self.app.update(message);
-
-            if is_awaited {
-                return Ok(());
-            }
-        }
-    }
-
-    pub fn find(&self, text: &str) -> Result<(), GuiTestError> {
-        simulator(self.app.view())
-            .find(text)
-            .map(|_found_target| ())
-            .map_err(GuiTestError::from)
-    }
-
-    pub fn join_cluster(
-        &mut self,
-        balancer_address: String,
-    ) -> Result<Task<Message>, GuiTestError> {
-        for message in [
-            Message::Home(HomeMessage::JoinBalancer),
-            Message::JoinBalancerForm(JoinBalancerFormMessage::SetBalancerAddress(
-                balancer_address,
-            )),
-            Message::JoinBalancerForm(JoinBalancerFormMessage::SetSlotsCount("1".to_owned())),
-        ] {
-            let _form_task = self.app.update(message);
-        }
-
-        self.click("Connect")
-    }
-
-    pub fn messages_from_clicking(&self, label: &str) -> Result<Vec<Message>, GuiTestError> {
-        let mut interface = simulator(self.app.view());
-        let click_result = interface.click(label);
-        let messages = interface.into_messages().collect();
-
-        click_result
-            .map(|_clicked_target| messages)
-            .map_err(GuiTestError::from)
-    }
-
-    pub fn confirm_cluster_start(
-        &mut self,
-        web_admin_panel_address: String,
-    ) -> Result<TaskActions, GuiTestError> {
-        for message in [
-            Message::Home(HomeMessage::StartBalancer),
-            Message::StartBalancerForm(StartBalancerFormMessage::SetBalancerAddress(
-                EPHEMERAL_LOOPBACK_ADDR.to_string(),
-            )),
-            Message::StartBalancerForm(StartBalancerFormMessage::SetInferenceAddress(
-                EPHEMERAL_LOOPBACK_ADDR.to_string(),
-            )),
-            Message::StartBalancerForm(StartBalancerFormMessage::SetWebAdminPanelAddress(
-                web_admin_panel_address,
-            )),
-            Message::StartBalancerForm(StartBalancerFormMessage::ToggleAddModelLater(true)),
-        ] {
-            let _form_task = self.app.update(message);
-        }
-
-        TaskActions::of(self.app.update(Message::StartBalancerForm(
-            StartBalancerFormMessage::Confirm,
-        )))
-    }
-
-    pub async fn start_cluster(
-        &mut self,
-        web_admin_panel_address: String,
-    ) -> Result<StartedCluster, GuiTestError> {
-        let mut messages = self.confirm_cluster_start(web_admin_panel_address)?;
-
-        match messages.next_message().await? {
-            balancer_started @ Message::BalancerStarted { addresses, .. } => {
-                let _running_task = self.app.update(balancer_started);
-
-                Ok(StartedCluster {
-                    addresses,
-                    messages,
-                })
-            }
-            message => Err(GuiTestError::ClusterDidNotStart {
-                message: Box::new(message),
-            }),
-        }
-    }
-
-    pub fn update(&mut self, message: Message) -> Task<Message> {
-        self.app.update(message)
-    }
-}
diff --git a/paddler_gui_tests/src/app_subscriptions.rs b/paddler_gui_tests/src/app_subscriptions.rs
deleted file mode 100644
index 827f40904..000000000
--- a/paddler_gui_tests/src/app_subscriptions.rs
+++ /dev/null
@@ -1,53 +0,0 @@
-use futures::StreamExt as _;
-use futures::channel::mpsc;
-use iced_test::core::Event;
-use iced_test::core::event::Status;
-use iced_test::core::window::Id as WindowId;
-use iced_test::futures::Runtime;
-use iced_test::futures::backend::default::Executor;
-use iced_test::futures::subscription::Event as SubscriptionEvent;
-use iced_test::futures::subscription::into_recipes;
-
-use paddler_gui::app::App;
-use paddler_gui::message::Message;
-
-use crate::gui_test_error::GuiTestError;
-
-pub struct AppSubscriptions {
-    messages: mpsc::Receiver<Message>,
-    runtime: Runtime<Executor, mpsc::Sender<Message>, Message>,
-    window: WindowId,
-}
-
-impl AppSubscriptions {
-    pub fn of(app: &App) -> Result<Self, GuiTestError> {
-        let executor = Executor::new().map_err(GuiTestError::ExecutorUnavailable)?;
-        let (message_tx, messages) = mpsc::channel(0);
-        let mut runtime = Runtime::new(executor, message_tx);
-        let recipes = into_recipes(runtime.enter(|| app.subscription()));
-
-        runtime.track(recipes);
-
-        Ok(Self {
-            messages,
-            runtime,
-            window: WindowId::unique(),
-        })
-    }
-
-    pub fn broadcast_unhandled(&mut self, event: Event) {
-        self.runtime.broadcast(SubscriptionEvent::Interaction {
-            window: self.window,
-            event,
-            status: Status::Ignored,
-        });
-    }
-
-    pub fn next_message(&mut self) -> Result<Message, GuiTestError> {
-        let messages = &mut self.messages;
-
-        self.runtime
-            .block_on(messages.next())
-            .ok_or(GuiTestError::SubscriptionsEnded)
-    }
-}
diff --git a/paddler_gui_tests/src/emulated_app.rs b/paddler_gui_tests/src/emulated_app.rs
deleted file mode 100644
index c56f82f24..000000000
--- a/paddler_gui_tests/src/emulated_app.rs
+++ /dev/null
@@ -1,94 +0,0 @@
-use futures::StreamExt as _;
-use futures::channel::mpsc;
-use futures::executor::block_on;
-use iced_test::Emulator;
-use iced_test::Instruction;
-use iced_test::core::Size;
-use iced_test::core::mouse::Button;
-use iced_test::emulator::Event as EmulatorEvent;
-use iced_test::emulator::Mode;
-use iced_test::instruction::Expectation;
-use iced_test::instruction::Interaction;
-use iced_test::instruction::Keyboard;
-use iced_test::instruction::Mouse;
-use iced_test::instruction::Target;
-use iced_test::program::Program;
-
-use crate::gui_test_error::GuiTestError;
-
-pub struct EmulatedApp<TProgram: Program + 'static> {
-    emulator: Emulator<TProgram>,
-    events: mpsc::Receiver<EmulatorEvent<TProgram>>,
-    program: TProgram,
-}
-
-impl<TProgram: Program + 'static> EmulatedApp<TProgram> {
-    pub fn boot(program: TProgram, window_size: Size) -> Result<Self, GuiTestError> {
-        let (event_tx, events) = mpsc::channel(0);
-        let emulator = Emulator::new(event_tx, &program, Mode::Immediate, window_size);
-        let mut emulated_app = Self {
-            emulator,
-            events,
-            program,
-        };
-
-        emulated_app.wait_until_ready().map(|()| emulated_app)
-    }
-
-    pub fn click(&mut self, label: &str) -> Result<(), GuiTestError> {
-        self.run(Instruction::Interact(Interaction::Mouse(Mouse::Click {
-            button: Button::Left,
-            target: Some(Target::Text(label.to_owned())),
-        })))
-    }
-
-    pub fn expect_text(&mut self, text: &str) -> Result<(), GuiTestError> {
-        self.run(Instruction::Expect(Expectation::Text(text.to_owned())))
-    }
-
-    pub fn perform_next_action(&mut self) -> Result<(), GuiTestError> {
-        match block_on(self.events.next()) {
-            Some(EmulatorEvent::Action(action)) => {
-                self.emulator.perform(&self.program, action);
-
-                Ok(())
-            }
-            Some(EmulatorEvent::Failed(instruction)) => {
-                Err(GuiTestError::InstructionFailed { instruction })
-            }
-            Some(EmulatorEvent::Ready) => Err(GuiTestError::EmulatorReadyWithoutAction),
-            None => Err(GuiTestError::EmulatorStopped),
-        }
-    }
-
-    pub fn typewrite(&mut self, text: &str) -> Result<(), GuiTestError> {
-        self.run(Instruction::Interact(Interaction::Keyboard(
-            Keyboard::Typewrite(text.to_owned()),
-        )))
-    }
-
-    pub fn update(&mut self, message: TProgram::Message) {
-        self.emulator.update(&self.program, message);
-    }
-
-    fn run(&mut self, instruction: Instruction) -> Result<(), GuiTestError> {
-        self.emulator.run(&self.program, instruction);
-
-        self.wait_until_ready()
-    }
-
-    fn wait_until_ready(&mut self) -> Result<(), GuiTestError> {
-        loop {
-            match block_on(self.events.next()) {
-                Some(EmulatorEvent::Action(action)) => {
-                    self.emulator.perform(&self.program, action);
-                }
-                Some(EmulatorEvent::Failed(instruction)) => {
-                    return Err(GuiTestError::InstructionFailed { instruction });
-                }
-                Some(EmulatorEvent::Ready) => return Ok(()),
-                None => return Err(GuiTestError::EmulatorStopped),
-            }
-        }
-    }
-}
diff --git a/paddler_gui_tests/src/gui_test_error.rs b/paddler_gui_tests/src/gui_test_error.rs
deleted file mode 100644
index 0eabe219b..000000000
--- a/paddler_gui_tests/src/gui_test_error.rs
+++ /dev/null
@@ -1,34 +0,0 @@
-use std::io;
-
-use iced_test::Error as InterfaceError;
-use iced_test::Instruction;
-use iced_test::runtime::Action;
-use thiserror::Error;
-
-use paddler_gui::message::Message;
-
-#[derive(Debug, Error)]
-pub enum GuiTestError {
-    #[error("the task performed {action:?} instead of writing to the clipboard")]
-    ActionIsNotAClipboardWrite { action: Box<Action<Message>> },
-    #[error("the task performed {action:?} instead of producing a message")]
-    ActionIsNotAMessage { action: Box<Action<Message>> },
-    #[error("the cluster reported {message:?} instead of starting")]
-    ClusterDidNotStart { message: Box<Message> },
-    #[error("the emulator became ready without performing the expected action")]
-    EmulatorReadyWithoutAction,
-    #[error("the emulator stopped before becoming ready")]
-    EmulatorStopped,
-    #[error("the subscription executor could not start")]
-    ExecutorUnavailable(#[source] io::Error),
-    #[error("the emulated instruction failed: {instruction}")]
-    InstructionFailed { instruction: Instruction },
-    #[error("the interface did not respond as expected")]
-    Interface(#[from] InterfaceError),
-    #[error("the subscriptions ended before producing the expected message")]
-    SubscriptionsEnded,
-    #[error("the task ended before performing the expected action")]
-    TaskEnded,
-    #[error("the task performs no actions")]
-    TaskPerformsNothing,
-}
diff --git a/paddler_gui_tests/src/lib.rs b/paddler_gui_tests/src/lib.rs
deleted file mode 100644
index 7e12833c1..000000000
--- a/paddler_gui_tests/src/lib.rs
+++ /dev/null
@@ -1,8 +0,0 @@
-pub mod app_driver;
-pub mod app_subscriptions;
-pub mod emulated_app;
-pub mod gui_test_error;
-pub mod loopback_balancer_addresses;
-pub mod named_key_press;
-pub mod started_cluster;
-pub mod task_actions;
diff --git a/paddler_gui_tests/src/loopback_balancer_addresses.rs b/paddler_gui_tests/src/loopback_balancer_addresses.rs
deleted file mode 100644
index dad2f670a..000000000
--- a/paddler_gui_tests/src/loopback_balancer_addresses.rs
+++ /dev/null
@@ -1,12 +0,0 @@
-use std::net::IpAddr;
-use std::net::Ipv4Addr;
-use std::net::SocketAddr;
-
-use paddler_balancer::balancer_addresses::BalancerAddresses;
-
-pub const LOOPBACK_BALANCER_ADDRESSES: BalancerAddresses = BalancerAddresses {
-    compat_openai: None,
-    inference: SocketAddr::new(IpAddr::V4(Ipv4Addr::LOCALHOST), 8061),
-    management: SocketAddr::new(IpAddr::V4(Ipv4Addr::LOCALHOST), 8060),
-    web_admin_panel: None,
-};
diff --git a/paddler_gui_tests/src/named_key_press.rs b/paddler_gui_tests/src/named_key_press.rs
deleted file mode 100644
index c675be5f0..000000000
--- a/paddler_gui_tests/src/named_key_press.rs
+++ /dev/null
@@ -1,21 +0,0 @@
-use iced_test::core::Event;
-use iced_test::core::keyboard::Event as KeyboardEvent;
-use iced_test::core::keyboard::Key;
-use iced_test::core::keyboard::Location;
-use iced_test::core::keyboard::Modifiers;
-use iced_test::core::keyboard::key::Code;
-use iced_test::core::keyboard::key::Named;
-use iced_test::core::keyboard::key::Physical;
-
-#[must_use]
-pub const fn named_key_press(named_key: Named, key_code: Code, modifiers: Modifiers) -> Event {
-    Event::Keyboard(KeyboardEvent::KeyPressed {
-        key: Key::Named(named_key),
-        modified_key: Key::Named(named_key),
-        physical_key: Physical::Code(key_code),
-        location: Location::Standard,
-        modifiers,
-        text: None,
-        repeat: false,
-    })
-}
diff --git a/paddler_gui_tests/src/started_cluster.rs b/paddler_gui_tests/src/started_cluster.rs
deleted file mode 100644
index d57470c4f..000000000
--- a/paddler_gui_tests/src/started_cluster.rs
+++ /dev/null
@@ -1,8 +0,0 @@
-use paddler_balancer::balancer_addresses::BalancerAddresses;
-
-use crate::task_actions::TaskActions;
-
-pub struct StartedCluster {
-    pub addresses: BalancerAddresses,
-    pub messages: TaskActions,
-}
diff --git a/paddler_gui_tests/src/task_actions.rs b/paddler_gui_tests/src/task_actions.rs
deleted file mode 100644
index 7ffdc2842..000000000
--- a/paddler_gui_tests/src/task_actions.rs
+++ /dev/null
@@ -1,42 +0,0 @@
-use futures::StreamExt as _;
-use futures::stream::BoxStream;
-use iced_test::runtime::Action;
-use iced_test::runtime::Task;
-use iced_test::runtime::clipboard::Action as ClipboardAction;
-use iced_test::runtime::task::into_stream;
-
-use paddler_gui::message::Message;
-
-use crate::gui_test_error::GuiTestError;
-
-pub struct TaskActions {
-    actions: BoxStream<'static, Action<Message>>,
-}
-
-impl TaskActions {
-    pub fn of(task: Task<Message>) -> Result<Self, GuiTestError> {
-        into_stream(task)
-            .map(|actions| Self { actions })
-            .ok_or(GuiTestError::TaskPerformsNothing)
-    }
-
-    pub async fn next_clipboard_write(&mut self) -> Result<String, GuiTestError> {
-        match self.actions.next().await {
-            Some(Action::Clipboard(ClipboardAction::Write { contents, .. })) => Ok(contents),
-            Some(action) => Err(GuiTestError::ActionIsNotAClipboardWrite {
-                action: Box::new(action),
-            }),
-            None => Err(GuiTestError::TaskEnded),
-        }
-    }
-
-    pub async fn next_message(&mut self) -> Result<Message, GuiTestError> {
-        match self.actions.next().await {
-            Some(Action::Output(message)) => Ok(message),
-            Some(action) => Err(GuiTestError::ActionIsNotAMessage {
-                action: Box::new(action),
-            }),
-            None => Err(GuiTestError::TaskEnded),
-        }
-    }
-}
diff --git a/paddler_gui_tests/tests/a_cluster_that_cannot_bind_its_address_shows_why_on_the_home_screen.rs b/paddler_gui_tests/tests/a_cluster_that_cannot_bind_its_address_shows_why_on_the_home_screen.rs
deleted file mode 100644
index a3fb84aa7..000000000
--- a/paddler_gui_tests/tests/a_cluster_that_cannot_bind_its_address_shows_why_on_the_home_screen.rs
+++ /dev/null
@@ -1,62 +0,0 @@
-use std::net::TcpListener;
-
-use futures::StreamExt as _;
-use tokio_util::sync::CancellationToken;
-use trzcina::ServiceShutdownOptions;
-
-use paddler_balancer::resolved_socket_addr::ResolvedSocketAddr;
-use paddler_bootstrap::balancer_runner_params::BalancerRunnerParams;
-use paddler_gui::balancer_runner_messages::balancer_runner_messages;
-use paddler_gui::home_message::HomeMessage;
-use paddler_gui::message::Message;
-use paddler_gui::start_balancer_form_action::StartBalancerFormAction;
-use paddler_gui::start_balancer_form_data::StartBalancerFormData;
-use paddler_gui::start_balancer_form_message::StartBalancerFormMessage;
-use paddler_gui_tests::app_driver::AppDriver;
-use paddler_test_cluster_harness::ephemeral_loopback_addr::EPHEMERAL_LOOPBACK_ADDR;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn a_cluster_that_cannot_bind_its_address_shows_why_on_the_home_screen() {
-    let occupying_listener =
-        TcpListener::bind(EPHEMERAL_LOOPBACK_ADDR).expect("the occupying listener must bind");
-    let occupied_addr = occupying_listener
-        .local_addr()
-        .expect("the occupying listener must report its address");
-    let mut start_form = StartBalancerFormData::suggesting_addresses_on(None);
-
-    for message in [
-        StartBalancerFormMessage::SetBalancerAddress(EPHEMERAL_LOOPBACK_ADDR.to_string()),
-        StartBalancerFormMessage::SetInferenceAddress(EPHEMERAL_LOOPBACK_ADDR.to_string()),
-        StartBalancerFormMessage::ToggleAddModelLater(true),
-    ] {
-        let _form_action = start_form.update(message);
-    }
-
-    let StartBalancerFormAction::StartBalancer(mut bootstrap_config) =
-        start_form.update(StartBalancerFormMessage::Confirm)
-    else {
-        panic!("a valid start form must start the cluster");
-    };
-
-    bootstrap_config.management_service_configuration.addr =
-        ResolvedSocketAddr::from(occupied_addr);
-
-    let cluster_messages = balancer_runner_messages(BalancerRunnerParams {
-        bootstrap_config: *bootstrap_config,
-        cancellation_token: CancellationToken::new(),
-        shutdown_options: ServiceShutdownOptions::default(),
-    })
-    .collect::<Vec<Message>>()
-    .await;
-    let mut app = AppDriver::open().await.expect("the app must open");
-    let _form_task = app.update(Message::Home(HomeMessage::StartBalancer));
-
-    for message in cluster_messages {
-        let _failure_task = app.update(message);
-    }
-
-    app.find(&format!(
-        "Unable to bind the management service to {occupied_addr}"
-    ))
-    .expect("a cluster that cannot bind its address must show why on the home screen");
-}
diff --git a/paddler_gui_tests/tests/a_cluster_that_fails_while_running_shows_why_on_the_home_screen.rs b/paddler_gui_tests/tests/a_cluster_that_fails_while_running_shows_why_on_the_home_screen.rs
deleted file mode 100644
index 2cfd073ec..000000000
--- a/paddler_gui_tests/tests/a_cluster_that_fails_while_running_shows_why_on_the_home_screen.rs
+++ /dev/null
@@ -1,44 +0,0 @@
-use std::sync::Arc;
-
-use tokio_util::sync::CancellationToken;
-
-use paddler_balancer::balancer_applicable_state::BalancerApplicableState;
-use paddler_bootstrap::bootstrap_error::BootstrapError;
-use paddler_gui::home_message::HomeMessage;
-use paddler_gui::message::Message;
-use paddler_gui::running_balancer_snapshot::RunningBalancerSnapshot;
-use paddler_gui_tests::app_driver::AppDriver;
-use paddler_gui_tests::loopback_balancer_addresses::LOOPBACK_BALANCER_ADDRESSES;
-use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn a_cluster_that_fails_while_running_shows_why_on_the_home_screen() {
-    let mut app = AppDriver::open().await.expect("the app must open");
-
-    for message in [
-        Message::Home(HomeMessage::StartBalancer),
-        Message::BalancerStarted {
-            addresses: LOOPBACK_BALANCER_ADDRESSES,
-            cancellation_token: CancellationToken::new(),
-            snapshot: Box::new(RunningBalancerSnapshot {
-                agent_snapshots: Vec::new(),
-                balancer_applicable_state: BalancerApplicableState::from(
-                    BalancerDesiredState::default(),
-                ),
-                balancer_desired_state: BalancerDesiredState::default(),
-            }),
-        },
-    ] {
-        let _transition_task = app.update(message);
-    }
-
-    app.find("Cluster is running")
-        .expect("a started cluster must show that it is running");
-
-    let _failure_task = app.update(Message::BalancerFailed(Arc::new(
-        BootstrapError::ServiceThreadPanicked,
-    )));
-
-    app.find("A Paddler service thread panicked")
-        .expect("a cluster that fails while running must show why on the home screen");
-}
diff --git a/paddler_gui_tests/tests/a_failed_agent_shows_why_on_the_home_screen.rs b/paddler_gui_tests/tests/a_failed_agent_shows_why_on_the_home_screen.rs
deleted file mode 100644
index 257f2f76d..000000000
--- a/paddler_gui_tests/tests/a_failed_agent_shows_why_on_the_home_screen.rs
+++ /dev/null
@@ -1,20 +0,0 @@
-use std::sync::Arc;
-
-use paddler_bootstrap::bootstrap_error::BootstrapError;
-use paddler_gui::message::Message;
-use paddler_gui_tests::app_driver::AppDriver;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn a_failed_agent_shows_why_on_the_home_screen() {
-    let mut app = AppDriver::open().await.expect("the app must open");
-    let _agent_task = app
-        .join_cluster("127.0.0.1:8060".to_owned())
-        .expect("the join form must offer to connect");
-
-    let _failed_task = app.update(Message::AgentFailed(Arc::new(
-        BootstrapError::ServiceThreadPanicked,
-    )));
-
-    app.find("A Paddler service thread panicked")
-        .expect("a failed agent must show why on the home screen");
-}
diff --git a/paddler_gui_tests/tests/a_joining_agent_says_it_is_connecting.rs b/paddler_gui_tests/tests/a_joining_agent_says_it_is_connecting.rs
deleted file mode 100644
index c848a5d06..000000000
--- a/paddler_gui_tests/tests/a_joining_agent_says_it_is_connecting.rs
+++ /dev/null
@@ -1,23 +0,0 @@
-use iced_test::simulator;
-use tokio_util::sync::CancellationToken;
-
-use paddler_gui::agent_running_data::AgentRunningData;
-use paddler_messaging::agent_status::AgentStatus;
-use paddler_messaging::balancer_connection::BalancerConnection;
-
-#[test]
-fn a_joining_agent_says_it_is_connecting() {
-    let agent_running_data = AgentRunningData {
-        balancer_address: "127.0.0.1:8060".to_owned(),
-        balancer_connection: BalancerConnection::Connecting,
-        cancellation_token: CancellationToken::new(),
-        name: Some("gpu-box".to_owned()),
-        slots_processing: 0,
-        status: AgentStatus::default(),
-    };
-    let mut running_agent = simulator(agent_running_data.view());
-
-    running_agent
-        .find("Connecting to the cluster...")
-        .expect("a joining agent must say that it is connecting");
-}
diff --git a/paddler_gui_tests/tests/a_message_for_another_screen_leaves_the_screen_unchanged.rs b/paddler_gui_tests/tests/a_message_for_another_screen_leaves_the_screen_unchanged.rs
deleted file mode 100644
index 22723eb79..000000000
--- a/paddler_gui_tests/tests/a_message_for_another_screen_leaves_the_screen_unchanged.rs
+++ /dev/null
@@ -1,12 +0,0 @@
-use paddler_gui::message::Message;
-use paddler_gui_tests::app_driver::AppDriver;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn a_message_for_another_screen_leaves_the_screen_unchanged() {
-    let mut app = AppDriver::open().await.expect("the app must open");
-
-    let _ignored_task = app.update(Message::BalancerStopped);
-
-    app.find("Join a cluster")
-        .expect("a message for another screen must leave the home screen in place");
-}
diff --git a/paddler_gui_tests/tests/a_running_cluster_lists_its_agents.rs b/paddler_gui_tests/tests/a_running_cluster_lists_its_agents.rs
deleted file mode 100644
index f786d0925..000000000
--- a/paddler_gui_tests/tests/a_running_cluster_lists_its_agents.rs
+++ /dev/null
@@ -1,52 +0,0 @@
-use iced_test::simulator;
-use tokio_util::sync::CancellationToken;
-
-use paddler_balancer::balancer_applicable_state::BalancerApplicableState;
-use paddler_gui::running_balancer_data::RunningBalancerData;
-use paddler_gui::running_balancer_snapshot::RunningBalancerSnapshot;
-use paddler_gui_tests::loopback_balancer_addresses::LOOPBACK_BALANCER_ADDRESSES;
-use paddler_messaging::agent_controller_snapshot::AgentControllerSnapshot;
-use paddler_messaging::agent_state_application_status::AgentStateApplicationStatus;
-use paddler_messaging::agent_status::AgentStatus;
-use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-
-#[test]
-fn a_running_cluster_lists_its_agents() {
-    let running_balancer_data = RunningBalancerData {
-        addresses: LOOPBACK_BALANCER_ADDRESSES,
-        cancellation_token: CancellationToken::new(),
-        snapshot: Box::new(RunningBalancerSnapshot {
-            agent_snapshots: vec![AgentControllerSnapshot {
-                id: "agent_id".to_owned(),
-                name: Some("gpu-box".to_owned()),
-                slots_processing: 1,
-                status: AgentStatus {
-                    desired_slots_total: 4,
-                    model_path: Some("/models/qwen3.gguf".to_owned()),
-                    slots_total: 4,
-                    state_application_status: AgentStateApplicationStatus::Applied,
-                    ..AgentStatus::default()
-                },
-            }],
-            balancer_applicable_state: BalancerApplicableState::from(
-                BalancerDesiredState::default(),
-            ),
-            balancer_desired_state: BalancerDesiredState::default(),
-        }),
-        stopping: false,
-    };
-    let mut running_cluster = simulator(running_balancer_data.view());
-
-    running_cluster
-        .find("gpu-box")
-        .expect("the running cluster must name its agent");
-    running_cluster
-        .find("qwen3.gguf")
-        .expect("the running cluster must show the model its agent loaded");
-    running_cluster
-        .find("Status: OK")
-        .expect("the running cluster must show the status of its agent");
-    running_cluster
-        .find("Slots: 1/4/4")
-        .expect("the running cluster must show the slots of its agent");
-}
diff --git a/paddler_gui_tests/tests/a_running_cluster_shows_its_configured_and_applied_models.rs b/paddler_gui_tests/tests/a_running_cluster_shows_its_configured_and_applied_models.rs
deleted file mode 100644
index 7fe96dd6a..000000000
--- a/paddler_gui_tests/tests/a_running_cluster_shows_its_configured_and_applied_models.rs
+++ /dev/null
@@ -1,36 +0,0 @@
-use iced_test::simulator;
-use tokio_util::sync::CancellationToken;
-
-use paddler_balancer::balancer_applicable_state::BalancerApplicableState;
-use paddler_gui::running_balancer_data::RunningBalancerData;
-use paddler_gui::running_balancer_snapshot::RunningBalancerSnapshot;
-use paddler_gui_tests::loopback_balancer_addresses::LOOPBACK_BALANCER_ADDRESSES;
-use paddler_messaging::agent_desired_model::AgentDesiredModel;
-use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-
-#[test]
-fn a_running_cluster_shows_its_configured_and_applied_models() {
-    let running_balancer_data = RunningBalancerData {
-        addresses: LOOPBACK_BALANCER_ADDRESSES,
-        cancellation_token: CancellationToken::new(),
-        snapshot: Box::new(RunningBalancerSnapshot {
-            agent_snapshots: Vec::new(),
-            balancer_applicable_state: BalancerApplicableState::from(
-                BalancerDesiredState::default(),
-            ),
-            balancer_desired_state: BalancerDesiredState {
-                model: AgentDesiredModel::LocalToAgent("/models/qwen3.gguf".to_owned()),
-                ..BalancerDesiredState::default()
-            },
-        }),
-        stopping: false,
-    };
-    let mut running_cluster = simulator(running_balancer_data.view());
-
-    running_cluster
-        .find("Configured model: Local: /models/qwen3.gguf")
-        .expect("the running cluster must show the model it was configured with");
-    running_cluster
-        .find("Applied model: (not set)")
-        .expect("the running cluster must show the model its agents apply");
-}
diff --git a/paddler_gui_tests/tests/a_running_cluster_without_agents_waits_for_them.rs b/paddler_gui_tests/tests/a_running_cluster_without_agents_waits_for_them.rs
deleted file mode 100644
index f6c56a293..000000000
--- a/paddler_gui_tests/tests/a_running_cluster_without_agents_waits_for_them.rs
+++ /dev/null
@@ -1,29 +0,0 @@
-use iced_test::simulator;
-use tokio_util::sync::CancellationToken;
-
-use paddler_balancer::balancer_applicable_state::BalancerApplicableState;
-use paddler_gui::running_balancer_data::RunningBalancerData;
-use paddler_gui::running_balancer_snapshot::RunningBalancerSnapshot;
-use paddler_gui_tests::loopback_balancer_addresses::LOOPBACK_BALANCER_ADDRESSES;
-use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-
-#[test]
-fn a_running_cluster_without_agents_waits_for_them() {
-    let running_balancer_data = RunningBalancerData {
-        addresses: LOOPBACK_BALANCER_ADDRESSES,
-        cancellation_token: CancellationToken::new(),
-        snapshot: Box::new(RunningBalancerSnapshot {
-            agent_snapshots: Vec::new(),
-            balancer_applicable_state: BalancerApplicableState::from(
-                BalancerDesiredState::default(),
-            ),
-            balancer_desired_state: BalancerDesiredState::default(),
-        }),
-        stopping: false,
-    };
-    let mut running_cluster = simulator(running_balancer_data.view());
-
-    running_cluster
-        .find("Waiting for agents to connect...")
-        .expect("a running cluster without agents must wait for them");
-}
diff --git a/paddler_gui_tests/tests/a_stopped_agent_returns_to_the_home_screen.rs b/paddler_gui_tests/tests/a_stopped_agent_returns_to_the_home_screen.rs
deleted file mode 100644
index b8f68509b..000000000
--- a/paddler_gui_tests/tests/a_stopped_agent_returns_to_the_home_screen.rs
+++ /dev/null
@@ -1,15 +0,0 @@
-use paddler_gui::message::Message;
-use paddler_gui_tests::app_driver::AppDriver;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn a_stopped_agent_returns_to_the_home_screen() {
-    let mut app = AppDriver::open().await.expect("the app must open");
-    let _agent_task = app
-        .join_cluster("127.0.0.1:8060".to_owned())
-        .expect("the join form must offer to connect");
-
-    let _stopped_task = app.update(Message::AgentStopped);
-
-    app.find("Join a cluster")
-        .expect("a stopped agent must return to the home screen");
-}
diff --git a/paddler_gui_tests/tests/an_agent_counts_its_issues.rs b/paddler_gui_tests/tests/an_agent_counts_its_issues.rs
deleted file mode 100644
index 8d33a503c..000000000
--- a/paddler_gui_tests/tests/an_agent_counts_its_issues.rs
+++ /dev/null
@@ -1,36 +0,0 @@
-use std::collections::BTreeSet;
-
-use iced_test::simulator;
-use tokio_util::sync::CancellationToken;
-
-use paddler_gui::agent_running_data::AgentRunningData;
-use paddler_messaging::agent_issue::AgentIssue;
-use paddler_messaging::agent_issue_params::model_path::ModelPath;
-use paddler_messaging::agent_status::AgentStatus;
-use paddler_messaging::balancer_connection::BalancerConnection;
-
-#[test]
-fn an_agent_counts_its_issues() {
-    let missing_model = ModelPath {
-        model_path: "/models/qwen3.gguf".to_owned(),
-    };
-    let agent_running_data = AgentRunningData {
-        balancer_address: "127.0.0.1:8060".to_owned(),
-        balancer_connection: BalancerConnection::Connected,
-        cancellation_token: CancellationToken::new(),
-        name: Some("gpu-box".to_owned()),
-        slots_processing: 0,
-        status: AgentStatus {
-            issues: BTreeSet::from([
-                AgentIssue::ModelFileDoesNotExist(missing_model.clone()),
-                AgentIssue::UnableToFindChatTemplate(missing_model),
-            ]),
-            ..AgentStatus::default()
-        },
-    };
-    let mut running_agent = simulator(agent_running_data.view());
-
-    running_agent
-        .find("2 issues")
-        .expect("an agent with issues must count them");
-}
diff --git a/paddler_gui_tests/tests/an_agent_joins_the_cluster_at_the_copied_address.rs b/paddler_gui_tests/tests/an_agent_joins_the_cluster_at_the_copied_address.rs
deleted file mode 100644
index ca6a913dc..000000000
--- a/paddler_gui_tests/tests/an_agent_joins_the_cluster_at_the_copied_address.rs
+++ /dev/null
@@ -1,66 +0,0 @@
-use paddler_gui::agent_running_message::AgentRunningMessage;
-use paddler_gui::message::Message;
-use paddler_gui_tests::app_driver::AppDriver;
-use paddler_gui_tests::task_actions::TaskActions;
-use paddler_messaging::balancer_connection::BalancerConnection;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn an_agent_joins_the_cluster_at_the_copied_address() {
-    let mut cluster_app = AppDriver::open().await.expect("the cluster app must open");
-    let mut cluster = cluster_app
-        .start_cluster(String::new())
-        .await
-        .expect("the cluster must start");
-    let copied_address = TaskActions::of(
-        cluster_app
-            .click("Copy address")
-            .expect("the running cluster must offer to copy its address"),
-    )
-    .expect("copying the address must write to the clipboard")
-    .next_clipboard_write()
-    .await
-    .expect("copying the address must write it to the clipboard");
-
-    let mut agent_app = AppDriver::open().await.expect("the agent app must open");
-
-    let mut agent = TaskActions::of(
-        agent_app
-            .join_cluster(copied_address.clone())
-            .expect("the join form must offer to connect"),
-    )
-    .expect("connecting must run the agent");
-
-    agent_app
-        .deliver_until(&mut agent, |message| {
-            matches!(
-                message,
-                Message::AgentRunning(AgentRunningMessage::BalancerConnectionChanged(
-                    BalancerConnection::Connected
-                ))
-            )
-        })
-        .await
-        .expect("the agent must connect to the copied address");
-    agent_app
-        .find(&format!("Connected to the cluster at {copied_address}"))
-        .expect("the agent must show the cluster it joined");
-
-    let _disconnect_task = agent_app
-        .click("Disconnect")
-        .expect("the running agent must offer to disconnect");
-    agent_app
-        .deliver_until(&mut agent, |message| {
-            matches!(message, Message::AgentStopped)
-        })
-        .await
-        .expect("the disconnected agent must stop");
-    let _stop_task = cluster_app
-        .click("Stop cluster")
-        .expect("the running cluster must offer to stop");
-    cluster_app
-        .deliver_until(&mut cluster.messages, |message| {
-            matches!(message, Message::BalancerStopped)
-        })
-        .await
-        .expect("the stopped cluster must report its stop");
-}
diff --git a/paddler_gui_tests/tests/an_agent_shows_its_download_progress.rs b/paddler_gui_tests/tests/an_agent_shows_its_download_progress.rs
deleted file mode 100644
index 4b78ab296..000000000
--- a/paddler_gui_tests/tests/an_agent_shows_its_download_progress.rs
+++ /dev/null
@@ -1,31 +0,0 @@
-use iced_test::simulator;
-use tokio_util::sync::CancellationToken;
-
-use paddler_gui::agent_running_data::AgentRunningData;
-use paddler_messaging::agent_status::AgentStatus;
-use paddler_messaging::balancer_connection::BalancerConnection;
-use paddler_messaging::model_download_status::ModelDownloadStatus;
-
-#[test]
-fn an_agent_shows_its_download_progress() {
-    let agent_running_data = AgentRunningData {
-        balancer_address: "127.0.0.1:8060".to_owned(),
-        balancer_connection: BalancerConnection::Connected,
-        cancellation_token: CancellationToken::new(),
-        name: None,
-        slots_processing: 0,
-        status: AgentStatus {
-            download_status: ModelDownloadStatus::Downloading {
-                downloaded_bytes: 250,
-                model_path: "owner/repo/main/model.gguf".to_owned(),
-                total_bytes: 1000,
-            },
-            ..AgentStatus::default()
-        },
-    };
-    let mut running_agent = simulator(agent_running_data.view());
-
-    running_agent
-        .find("Status: Downloading (25%)")
-        .expect("a downloading agent must show its download progress");
-}
diff --git a/paddler_gui_tests/tests/cancelling_a_starting_cluster_stops_it.rs b/paddler_gui_tests/tests/cancelling_a_starting_cluster_stops_it.rs
deleted file mode 100644
index 95f43257a..000000000
--- a/paddler_gui_tests/tests/cancelling_a_starting_cluster_stops_it.rs
+++ /dev/null
@@ -1,25 +0,0 @@
-use paddler_gui::message::Message;
-use paddler_gui_tests::app_driver::AppDriver;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn cancelling_a_starting_cluster_stops_it() {
-    let mut app = AppDriver::open().await.expect("the app must open");
-    let mut cluster = app
-        .confirm_cluster_start(String::new())
-        .expect("confirming the form must start the cluster");
-
-    app.find("Starting...")
-        .expect("the start form must show that the cluster is starting");
-
-    let _cancel_task = app
-        .click("Cancel")
-        .expect("the starting cluster must offer to cancel");
-
-    app.deliver_until(&mut cluster, |message| {
-        matches!(message, Message::BalancerStopped)
-    })
-    .await
-    .expect("cancelling a starting cluster must stop it");
-    app.find("Join a cluster")
-        .expect("a cancelled cluster must leave the app on the home screen");
-}
diff --git a/paddler_gui_tests/tests/cancelling_an_agent_reports_that_it_stopped.rs b/paddler_gui_tests/tests/cancelling_an_agent_reports_that_it_stopped.rs
deleted file mode 100644
index c4e77de8f..000000000
--- a/paddler_gui_tests/tests/cancelling_an_agent_reports_that_it_stopped.rs
+++ /dev/null
@@ -1,37 +0,0 @@
-use std::net::TcpListener;
-
-use futures::StreamExt as _;
-use tokio_util::sync::CancellationToken;
-use trzcina::ServiceShutdownOptions;
-
-use paddler_bootstrap::agent_bootstrap_config::AgentBootstrapConfig;
-use paddler_bootstrap::agent_runner_params::AgentRunnerParams;
-use paddler_gui::agent_runner_messages::agent_runner_messages;
-use paddler_gui::message::Message;
-use paddler_test_cluster_harness::ephemeral_loopback_addr::EPHEMERAL_LOOPBACK_ADDR;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn cancelling_an_agent_reports_that_it_stopped() {
-    let unresponsive_balancer =
-        TcpListener::bind(EPHEMERAL_LOOPBACK_ADDR).expect("the unresponsive balancer must bind");
-    let cancellation_token = CancellationToken::new();
-    let agent_messages = agent_runner_messages(AgentRunnerParams {
-        bootstrap_config: AgentBootstrapConfig {
-            agent_name: None,
-            management_address: unresponsive_balancer
-                .local_addr()
-                .expect("the unresponsive balancer must report its address")
-                .to_string(),
-            slots: 1,
-        },
-        cancellation_token: cancellation_token.clone(),
-        shutdown_options: ServiceShutdownOptions::default(),
-    });
-
-    cancellation_token.cancel();
-
-    assert!(matches!(
-        agent_messages.collect::<Vec<_>>().await.pop(),
-        Some(Message::AgentStopped)
-    ));
-}
diff --git a/paddler_gui_tests/tests/cancelling_the_join_form_returns_to_the_home_screen.rs b/paddler_gui_tests/tests/cancelling_the_join_form_returns_to_the_home_screen.rs
deleted file mode 100644
index 16d356388..000000000
--- a/paddler_gui_tests/tests/cancelling_the_join_form_returns_to_the_home_screen.rs
+++ /dev/null
@@ -1,16 +0,0 @@
-use paddler_gui::home_message::HomeMessage;
-use paddler_gui::message::Message;
-use paddler_gui_tests::app_driver::AppDriver;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn cancelling_the_join_form_returns_to_the_home_screen() {
-    let mut app = AppDriver::open().await.expect("the app must open");
-    let _form_task = app.update(Message::Home(HomeMessage::JoinBalancer));
-
-    let _cancel_task = app
-        .click("Cancel")
-        .expect("the join form must offer to cancel");
-
-    app.find("Start a cluster")
-        .expect("cancelling the join form must return to the home screen");
-}
diff --git a/paddler_gui_tests/tests/cancelling_the_start_form_returns_to_the_home_screen.rs b/paddler_gui_tests/tests/cancelling_the_start_form_returns_to_the_home_screen.rs
deleted file mode 100644
index 08113bd8b..000000000
--- a/paddler_gui_tests/tests/cancelling_the_start_form_returns_to_the_home_screen.rs
+++ /dev/null
@@ -1,16 +0,0 @@
-use paddler_gui::home_message::HomeMessage;
-use paddler_gui::message::Message;
-use paddler_gui_tests::app_driver::AppDriver;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn cancelling_the_start_form_returns_to_the_home_screen() {
-    let mut app = AppDriver::open().await.expect("the app must open");
-    let _form_task = app.update(Message::Home(HomeMessage::StartBalancer));
-
-    let _cancel_task = app
-        .click("Cancel")
-        .expect("the start form must offer to cancel");
-
-    app.find("Join a cluster")
-        .expect("cancelling the start form must return to the home screen");
-}
diff --git a/paddler_gui_tests/tests/closing_the_window_quits_the_app.rs b/paddler_gui_tests/tests/closing_the_window_quits_the_app.rs
deleted file mode 100644
index 22452b7b2..000000000
--- a/paddler_gui_tests/tests/closing_the_window_quits_the_app.rs
+++ /dev/null
@@ -1,22 +0,0 @@
-use iced_test::core::Event;
-use iced_test::core::window::Event as WindowEvent;
-
-use paddler_gui::app::App;
-use paddler_gui::message::Message;
-use paddler_gui_tests::app_subscriptions::AppSubscriptions;
-
-#[test]
-fn closing_the_window_quits_the_app() {
-    let (app, _boot_task) = App::new();
-    let mut subscriptions =
-        AppSubscriptions::of(&app).expect("the app must subscribe to its events");
-
-    subscriptions.broadcast_unhandled(Event::Window(WindowEvent::CloseRequested));
-
-    assert!(matches!(
-        subscriptions
-            .next_message()
-            .expect("closing the window must produce a message"),
-        Message::Quit
-    ));
-}
diff --git a/paddler_gui_tests/tests/disconnecting_an_agent_returns_to_the_home_screen.rs b/paddler_gui_tests/tests/disconnecting_an_agent_returns_to_the_home_screen.rs
deleted file mode 100644
index e0dae4ac0..000000000
--- a/paddler_gui_tests/tests/disconnecting_an_agent_returns_to_the_home_screen.rs
+++ /dev/null
@@ -1,16 +0,0 @@
-use paddler_gui_tests::app_driver::AppDriver;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn disconnecting_an_agent_returns_to_the_home_screen() {
-    let mut app = AppDriver::open().await.expect("the app must open");
-    let _agent_task = app
-        .join_cluster("127.0.0.1:8060".to_owned())
-        .expect("the join form must offer to connect");
-
-    let _disconnect_task = app
-        .click("Disconnect")
-        .expect("the running agent must offer to disconnect");
-
-    app.find("Join a cluster")
-        .expect("disconnecting must return to the home screen");
-}
diff --git a/paddler_gui_tests/tests/home_screen_offers_to_join_a_cluster.rs b/paddler_gui_tests/tests/home_screen_offers_to_join_a_cluster.rs
deleted file mode 100644
index f21940cde..000000000
--- a/paddler_gui_tests/tests/home_screen_offers_to_join_a_cluster.rs
+++ /dev/null
@@ -1,19 +0,0 @@
-use iced_test::simulator;
-
-use paddler_gui::home_data::HomeData;
-use paddler_gui::home_message::HomeMessage;
-
-#[test]
-fn home_screen_offers_to_join_a_cluster() {
-    let home_data = HomeData { error: None };
-    let mut home_screen = simulator(home_data.view());
-
-    home_screen
-        .click("Join a cluster")
-        .expect("the home screen must offer to join a cluster");
-
-    assert!(matches!(
-        home_screen.into_messages().collect::<Vec<_>>().as_slice(),
-        [HomeMessage::JoinBalancer]
-    ));
-}
diff --git a/paddler_gui_tests/tests/home_screen_offers_to_start_a_cluster.rs b/paddler_gui_tests/tests/home_screen_offers_to_start_a_cluster.rs
deleted file mode 100644
index 13d48b3a4..000000000
--- a/paddler_gui_tests/tests/home_screen_offers_to_start_a_cluster.rs
+++ /dev/null
@@ -1,19 +0,0 @@
-use iced_test::simulator;
-
-use paddler_gui::home_data::HomeData;
-use paddler_gui::home_message::HomeMessage;
-
-#[test]
-fn home_screen_offers_to_start_a_cluster() {
-    let home_data = HomeData { error: None };
-    let mut home_screen = simulator(home_data.view());
-
-    home_screen
-        .click("Start a cluster")
-        .expect("the home screen must offer to start a cluster");
-
-    assert!(matches!(
-        home_screen.into_messages().collect::<Vec<_>>().as_slice(),
-        [HomeMessage::StartBalancer]
-    ));
-}
diff --git a/paddler_gui_tests/tests/home_screen_shows_the_error_that_brought_the_user_back.rs b/paddler_gui_tests/tests/home_screen_shows_the_error_that_brought_the_user_back.rs
deleted file mode 100644
index b7f507d6b..000000000
--- a/paddler_gui_tests/tests/home_screen_shows_the_error_that_brought_the_user_back.rs
+++ /dev/null
@@ -1,18 +0,0 @@
-use std::sync::Arc;
-
-use iced_test::simulator;
-
-use paddler_bootstrap::bootstrap_error::BootstrapError;
-use paddler_gui::home_data::HomeData;
-
-#[test]
-fn home_screen_shows_the_error_that_brought_the_user_back() {
-    let home_data = HomeData {
-        error: Some(Arc::new(BootstrapError::ServiceThreadPanicked)),
-    };
-    let mut home_screen = simulator(home_data.view());
-
-    home_screen
-        .find("A Paddler service thread panicked")
-        .expect("the home screen must show the error that brought the user back");
-}
diff --git a/paddler_gui_tests/tests/join_form_edits_the_agent_name.rs b/paddler_gui_tests/tests/join_form_edits_the_agent_name.rs
deleted file mode 100644
index e0176ec3c..000000000
--- a/paddler_gui_tests/tests/join_form_edits_the_agent_name.rs
+++ /dev/null
@@ -1,23 +0,0 @@
-use iced_test::simulator;
-
-use paddler_gui::join_balancer_form_data::JoinBalancerFormData;
-use paddler_gui::join_balancer_form_message::JoinBalancerFormMessage;
-
-#[test]
-fn join_form_edits_the_agent_name() {
-    let join_form_data = JoinBalancerFormData {
-        agent_name: "gpu-box".to_owned(),
-        ..JoinBalancerFormData::default()
-    };
-    let mut join_form = simulator(join_form_data.view());
-
-    join_form
-        .click("gpu-box")
-        .expect("the join form must show the entered agent name");
-    join_form.typewrite("2");
-
-    assert!(matches!(
-        join_form.into_messages().last(),
-        Some(JoinBalancerFormMessage::SetAgentName(agent_name)) if agent_name == "gpu-box2"
-    ));
-}
diff --git a/paddler_gui_tests/tests/join_form_shows_why_it_refused_to_connect.rs b/paddler_gui_tests/tests/join_form_shows_why_it_refused_to_connect.rs
deleted file mode 100644
index 0a260a9fa..000000000
--- a/paddler_gui_tests/tests/join_form_shows_why_it_refused_to_connect.rs
+++ /dev/null
@@ -1,18 +0,0 @@
-use iced_test::simulator;
-
-use paddler_gui::join_balancer_form_data::JoinBalancerFormData;
-use paddler_gui::join_balancer_form_message::JoinBalancerFormMessage;
-
-#[test]
-fn join_form_shows_why_it_refused_to_connect() {
-    let mut join_form_data = JoinBalancerFormData::default();
-    let _refused_connection = join_form_data.update(JoinBalancerFormMessage::Connect);
-    let mut join_form = simulator(join_form_data.view());
-
-    join_form
-        .find("Cluster address is required.")
-        .expect("the join form must explain the missing cluster address");
-    join_form
-        .find("Number of slots is required.")
-        .expect("the join form must explain the missing slot count");
-}
diff --git a/paddler_gui_tests/tests/main.rs b/paddler_gui_tests/tests/main.rs
deleted file mode 100644
index 54a880d3c..000000000
--- a/paddler_gui_tests/tests/main.rs
+++ /dev/null
@@ -1,34 +0,0 @@
-mod a_cluster_that_cannot_bind_its_address_shows_why_on_the_home_screen;
-mod a_cluster_that_fails_while_running_shows_why_on_the_home_screen;
-mod a_failed_agent_shows_why_on_the_home_screen;
-mod a_joining_agent_says_it_is_connecting;
-mod a_message_for_another_screen_leaves_the_screen_unchanged;
-mod a_running_cluster_lists_its_agents;
-mod a_running_cluster_shows_its_configured_and_applied_models;
-mod a_running_cluster_without_agents_waits_for_them;
-mod a_stopped_agent_returns_to_the_home_screen;
-mod an_agent_counts_its_issues;
-mod an_agent_joins_the_cluster_at_the_copied_address;
-mod an_agent_shows_its_download_progress;
-mod cancelling_a_starting_cluster_stops_it;
-mod cancelling_an_agent_reports_that_it_stopped;
-mod cancelling_the_join_form_returns_to_the_home_screen;
-mod cancelling_the_start_form_returns_to_the_home_screen;
-mod closing_the_window_quits_the_app;
-mod disconnecting_an_agent_returns_to_the_home_screen;
-mod home_screen_offers_to_join_a_cluster;
-mod home_screen_offers_to_start_a_cluster;
-mod home_screen_shows_the_error_that_brought_the_user_back;
-mod join_form_edits_the_agent_name;
-mod join_form_shows_why_it_refused_to_connect;
-mod other_key_presses_leave_the_focus_alone;
-mod pressing_shift_tab_asks_to_focus_the_previous_field;
-mod pressing_tab_asks_to_focus_the_next_field;
-mod quitting_stops_the_running_agent;
-mod quitting_stops_the_running_cluster;
-mod start_form_cannot_be_submitted_while_the_cluster_starts;
-mod start_form_lets_the_model_be_chosen_after_all;
-mod start_form_offers_to_add_the_model_later;
-mod stopping_a_running_cluster_returns_to_the_home_screen;
-mod the_join_form_can_be_filled_with_the_keyboard;
-mod the_running_cluster_links_to_its_web_admin_panel;
diff --git a/paddler_gui_tests/tests/other_key_presses_leave_the_focus_alone.rs b/paddler_gui_tests/tests/other_key_presses_leave_the_focus_alone.rs
deleted file mode 100644
index 1d58620d2..000000000
--- a/paddler_gui_tests/tests/other_key_presses_leave_the_focus_alone.rs
+++ /dev/null
@@ -1,29 +0,0 @@
-use iced_test::core::keyboard::Modifiers;
-use iced_test::core::keyboard::key::Code;
-use iced_test::core::keyboard::key::Named;
-
-use paddler_gui::app::App;
-use paddler_gui::message::Message;
-use paddler_gui_tests::app_subscriptions::AppSubscriptions;
-use paddler_gui_tests::named_key_press::named_key_press;
-
-#[test]
-fn other_key_presses_leave_the_focus_alone() {
-    let (app, _boot_task) = App::new();
-    let mut subscriptions =
-        AppSubscriptions::of(&app).expect("the app must subscribe to its events");
-
-    subscriptions.broadcast_unhandled(named_key_press(
-        Named::Escape,
-        Code::Escape,
-        Modifiers::empty(),
-    ));
-    subscriptions.broadcast_unhandled(named_key_press(Named::Tab, Code::Tab, Modifiers::SHIFT));
-
-    assert!(matches!(
-        subscriptions
-            .next_message()
-            .expect("the tab press after the other key must produce a message"),
-        Message::TabPressed { shift: true }
-    ));
-}
diff --git a/paddler_gui_tests/tests/pressing_shift_tab_asks_to_focus_the_previous_field.rs b/paddler_gui_tests/tests/pressing_shift_tab_asks_to_focus_the_previous_field.rs
deleted file mode 100644
index 6658d3ce3..000000000
--- a/paddler_gui_tests/tests/pressing_shift_tab_asks_to_focus_the_previous_field.rs
+++ /dev/null
@@ -1,24 +0,0 @@
-use iced_test::core::keyboard::Modifiers;
-use iced_test::core::keyboard::key::Code;
-use iced_test::core::keyboard::key::Named;
-
-use paddler_gui::app::App;
-use paddler_gui::message::Message;
-use paddler_gui_tests::app_subscriptions::AppSubscriptions;
-use paddler_gui_tests::named_key_press::named_key_press;
-
-#[test]
-fn pressing_shift_tab_asks_to_focus_the_previous_field() {
-    let (app, _boot_task) = App::new();
-    let mut subscriptions =
-        AppSubscriptions::of(&app).expect("the app must subscribe to its events");
-
-    subscriptions.broadcast_unhandled(named_key_press(Named::Tab, Code::Tab, Modifiers::SHIFT));
-
-    assert!(matches!(
-        subscriptions
-            .next_message()
-            .expect("pressing shift tab must produce a message"),
-        Message::TabPressed { shift: true }
-    ));
-}
diff --git a/paddler_gui_tests/tests/pressing_tab_asks_to_focus_the_next_field.rs b/paddler_gui_tests/tests/pressing_tab_asks_to_focus_the_next_field.rs
deleted file mode 100644
index 5fc4fc86e..000000000
--- a/paddler_gui_tests/tests/pressing_tab_asks_to_focus_the_next_field.rs
+++ /dev/null
@@ -1,24 +0,0 @@
-use iced_test::core::keyboard::Modifiers;
-use iced_test::core::keyboard::key::Code;
-use iced_test::core::keyboard::key::Named;
-
-use paddler_gui::app::App;
-use paddler_gui::message::Message;
-use paddler_gui_tests::app_subscriptions::AppSubscriptions;
-use paddler_gui_tests::named_key_press::named_key_press;
-
-#[test]
-fn pressing_tab_asks_to_focus_the_next_field() {
-    let (app, _boot_task) = App::new();
-    let mut subscriptions =
-        AppSubscriptions::of(&app).expect("the app must subscribe to its events");
-
-    subscriptions.broadcast_unhandled(named_key_press(Named::Tab, Code::Tab, Modifiers::empty()));
-
-    assert!(matches!(
-        subscriptions
-            .next_message()
-            .expect("pressing tab must produce a message"),
-        Message::TabPressed { shift: false }
-    ));
-}
diff --git a/paddler_gui_tests/tests/quitting_stops_the_running_agent.rs b/paddler_gui_tests/tests/quitting_stops_the_running_agent.rs
deleted file mode 100644
index faa320dc8..000000000
--- a/paddler_gui_tests/tests/quitting_stops_the_running_agent.rs
+++ /dev/null
@@ -1,44 +0,0 @@
-use std::net::TcpListener;
-
-use paddler_gui::agent_running_message::AgentRunningMessage;
-use paddler_gui::message::Message;
-use paddler_gui_tests::app_driver::AppDriver;
-use paddler_gui_tests::task_actions::TaskActions;
-use paddler_messaging::balancer_connection::BalancerConnection;
-use paddler_test_cluster_harness::ephemeral_loopback_addr::EPHEMERAL_LOOPBACK_ADDR;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn quitting_stops_the_running_agent() {
-    let unresponsive_balancer =
-        TcpListener::bind(EPHEMERAL_LOOPBACK_ADDR).expect("the unresponsive balancer must bind");
-    let mut app = AppDriver::open().await.expect("the app must open");
-    let mut agent = TaskActions::of(
-        app.join_cluster(
-            unresponsive_balancer
-                .local_addr()
-                .expect("the unresponsive balancer must report its address")
-                .to_string(),
-        )
-        .expect("the join form must offer to connect"),
-    )
-    .expect("connecting must run the agent");
-
-    app.deliver_until(&mut agent, |message| {
-        matches!(
-            message,
-            Message::AgentRunning(AgentRunningMessage::BalancerConnectionChanged(
-                BalancerConnection::Connecting
-            ))
-        )
-    })
-    .await
-    .expect("the agent must start connecting");
-
-    let _exit_task = app.update(Message::Quit);
-
-    app.deliver_until(&mut agent, |message| {
-        matches!(message, Message::AgentStopped)
-    })
-    .await
-    .expect("quitting must stop the running agent");
-}
diff --git a/paddler_gui_tests/tests/quitting_stops_the_running_cluster.rs b/paddler_gui_tests/tests/quitting_stops_the_running_cluster.rs
deleted file mode 100644
index 51773fa61..000000000
--- a/paddler_gui_tests/tests/quitting_stops_the_running_cluster.rs
+++ /dev/null
@@ -1,19 +0,0 @@
-use paddler_gui::message::Message;
-use paddler_gui_tests::app_driver::AppDriver;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn quitting_stops_the_running_cluster() {
-    let mut app = AppDriver::open().await.expect("the app must open");
-    let mut cluster = app
-        .start_cluster(String::new())
-        .await
-        .expect("the cluster must start");
-
-    let _exit_task = app.update(Message::Quit);
-
-    app.deliver_until(&mut cluster.messages, |message| {
-        matches!(message, Message::BalancerStopped)
-    })
-    .await
-    .expect("quitting must stop the running cluster");
-}
diff --git a/paddler_gui_tests/tests/start_form_cannot_be_submitted_while_the_cluster_starts.rs b/paddler_gui_tests/tests/start_form_cannot_be_submitted_while_the_cluster_starts.rs
deleted file mode 100644
index 62af94a28..000000000
--- a/paddler_gui_tests/tests/start_form_cannot_be_submitted_while_the_cluster_starts.rs
+++ /dev/null
@@ -1,20 +0,0 @@
-use iced_test::simulator;
-use tokio_util::sync::CancellationToken;
-
-use paddler_gui::balancer_launch::BalancerLaunch;
-use paddler_gui::start_balancer_form_data::StartBalancerFormData;
-
-#[test]
-fn start_form_cannot_be_submitted_while_the_cluster_starts() {
-    let start_form_data = StartBalancerFormData {
-        launch: BalancerLaunch::Starting(CancellationToken::new()),
-        ..StartBalancerFormData::suggesting_addresses_on(None)
-    };
-    let mut start_form = simulator(start_form_data.view());
-
-    start_form
-        .click("Starting...")
-        .expect("the start form must show that the cluster is starting");
-
-    assert!(start_form.into_messages().next().is_none());
-}
diff --git a/paddler_gui_tests/tests/start_form_lets_the_model_be_chosen_after_all.rs b/paddler_gui_tests/tests/start_form_lets_the_model_be_chosen_after_all.rs
deleted file mode 100644
index 538b95554..000000000
--- a/paddler_gui_tests/tests/start_form_lets_the_model_be_chosen_after_all.rs
+++ /dev/null
@@ -1,22 +0,0 @@
-use iced_test::simulator;
-
-use paddler_gui::start_balancer_form_data::StartBalancerFormData;
-use paddler_gui::start_balancer_form_message::StartBalancerFormMessage;
-
-#[test]
-fn start_form_lets_the_model_be_chosen_after_all() {
-    let start_form_data = StartBalancerFormData {
-        add_model_later: true,
-        ..StartBalancerFormData::suggesting_addresses_on(None)
-    };
-    let mut start_form = simulator(start_form_data.view());
-
-    start_form
-        .click("Add a model later")
-        .expect("the start form must let the model be chosen after all");
-
-    assert!(matches!(
-        start_form.into_messages().collect::<Vec<_>>().as_slice(),
-        [StartBalancerFormMessage::ToggleAddModelLater(false)]
-    ));
-}
diff --git a/paddler_gui_tests/tests/start_form_offers_to_add_the_model_later.rs b/paddler_gui_tests/tests/start_form_offers_to_add_the_model_later.rs
deleted file mode 100644
index 91648afc5..000000000
--- a/paddler_gui_tests/tests/start_form_offers_to_add_the_model_later.rs
+++ /dev/null
@@ -1,19 +0,0 @@
-use iced_test::simulator;
-
-use paddler_gui::start_balancer_form_data::StartBalancerFormData;
-use paddler_gui::start_balancer_form_message::StartBalancerFormMessage;
-
-#[test]
-fn start_form_offers_to_add_the_model_later() {
-    let start_form_data = StartBalancerFormData::suggesting_addresses_on(None);
-    let mut start_form = simulator(start_form_data.view());
-
-    start_form
-        .click("Add a model later")
-        .expect("the start form must offer to add the model later");
-
-    assert!(matches!(
-        start_form.into_messages().collect::<Vec<_>>().as_slice(),
-        [StartBalancerFormMessage::ToggleAddModelLater(true)]
-    ));
-}
diff --git a/paddler_gui_tests/tests/stopping_a_running_cluster_returns_to_the_home_screen.rs b/paddler_gui_tests/tests/stopping_a_running_cluster_returns_to_the_home_screen.rs
deleted file mode 100644
index 1aad1f36c..000000000
--- a/paddler_gui_tests/tests/stopping_a_running_cluster_returns_to_the_home_screen.rs
+++ /dev/null
@@ -1,24 +0,0 @@
-use paddler_gui::message::Message;
-use paddler_gui_tests::app_driver::AppDriver;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn stopping_a_running_cluster_returns_to_the_home_screen() {
-    let mut app = AppDriver::open().await.expect("the app must open");
-    let mut cluster = app
-        .start_cluster(String::new())
-        .await
-        .expect("the cluster must start");
-
-    let _stop_task = app
-        .click("Stop cluster")
-        .expect("the running cluster must offer to stop");
-    app.find("Stopping...")
-        .expect("a stopping cluster must say that it is stopping");
-    app.deliver_until(&mut cluster.messages, |message| {
-        matches!(message, Message::BalancerStopped)
-    })
-    .await
-    .expect("the stopped cluster must report its stop");
-    app.find("Join a cluster")
-        .expect("a stopped cluster must return to the home screen");
-}
diff --git a/paddler_gui_tests/tests/the_join_form_can_be_filled_with_the_keyboard.rs b/paddler_gui_tests/tests/the_join_form_can_be_filled_with_the_keyboard.rs
deleted file mode 100644
index 795300bb3..000000000
--- a/paddler_gui_tests/tests/the_join_form_can_be_filled_with_the_keyboard.rs
+++ /dev/null
@@ -1,49 +0,0 @@
-use std::net::TcpListener;
-
-use iced_test::core::Size;
-use iced_test::program::Program;
-
-use paddler_gui::message::Message;
-use paddler_gui::paddler_application::paddler_application;
-use paddler_gui_tests::emulated_app::EmulatedApp;
-use paddler_gui_tests::gui_test_error::GuiTestError;
-use paddler_test_cluster_harness::ephemeral_loopback_addr::EPHEMERAL_LOOPBACK_ADDR;
-
-fn press_tab<TProgram>(app: &mut EmulatedApp<TProgram>, shift: bool) -> Result<(), GuiTestError>
-where
-    TProgram: Program<Message = Message> + 'static,
-{
-    app.update(Message::TabPressed { shift });
-    app.perform_next_action()
-}
-
-#[test]
-fn the_join_form_can_be_filled_with_the_keyboard() {
-    let unresponsive_balancer =
-        TcpListener::bind(EPHEMERAL_LOOPBACK_ADDR).expect("the unresponsive balancer must bind");
-    let balancer_address = unresponsive_balancer
-        .local_addr()
-        .expect("the unresponsive balancer must report its address")
-        .to_string();
-    let mut app = EmulatedApp::boot(paddler_application(), Size::new(800.0, 800.0))
-        .expect("the app must boot");
-
-    app.click("Join a cluster")
-        .expect("the home screen must offer to join a cluster");
-    press_tab(&mut app, false).expect("tab must focus the cluster address");
-    app.typewrite(&balancer_address)
-        .expect("the cluster address must accept typing");
-    press_tab(&mut app, false).expect("tab must focus the agent name");
-    press_tab(&mut app, false).expect("tab must focus the slots");
-    press_tab(&mut app, true).expect("shift tab must focus the agent name again");
-    app.typewrite("gpu-box")
-        .expect("the agent name must accept typing");
-    press_tab(&mut app, false).expect("tab must focus the slots again");
-    app.typewrite("2").expect("the slots must accept typing");
-    app.click("Connect")
-        .expect("the join form must offer to connect");
-    app.expect_text("Connecting to the cluster...")
-        .expect("the filled form must connect the agent");
-    app.expect_text("gpu-box")
-        .expect("the connecting agent must carry the typed name");
-}
diff --git a/paddler_gui_tests/tests/the_running_cluster_links_to_its_web_admin_panel.rs b/paddler_gui_tests/tests/the_running_cluster_links_to_its_web_admin_panel.rs
deleted file mode 100644
index a9fe5bef4..000000000
--- a/paddler_gui_tests/tests/the_running_cluster_links_to_its_web_admin_panel.rs
+++ /dev/null
@@ -1,40 +0,0 @@
-#![cfg(feature = "web_admin_panel")]
-
-use reqwest::get;
-
-use paddler_gui::message::Message;
-use paddler_gui::running_balancer_message::RunningBalancerMessage;
-use paddler_gui_tests::app_driver::AppDriver;
-use paddler_test_cluster_harness::ephemeral_loopback_addr::EPHEMERAL_LOOPBACK_ADDR;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn the_running_cluster_links_to_its_web_admin_panel() {
-    let mut app = AppDriver::open().await.expect("the app must open");
-    let mut cluster = app
-        .start_cluster(EPHEMERAL_LOOPBACK_ADDR.to_string())
-        .await
-        .expect("the cluster must start");
-    let link_messages = app
-        .messages_from_clicking("Open in browser")
-        .expect("the running cluster must link to its web admin panel");
-    let [Message::RunningBalancer(RunningBalancerMessage::OpenUrl(web_admin_panel_url))] =
-        link_messages.as_slice()
-    else {
-        panic!("following the link must only open the web admin panel: {link_messages:?}");
-    };
-
-    get(web_admin_panel_url)
-        .await
-        .expect("the linked web admin panel must accept connections")
-        .error_for_status()
-        .expect("the linked web admin panel must render");
-
-    let _stop_task = app
-        .click("Stop cluster")
-        .expect("the running cluster must offer to stop");
-    app.deliver_until(&mut cluster.messages, |message| {
-        matches!(message, Message::BalancerStopped)
-    })
-    .await
-    .expect("the stopped cluster must report its stop");
-}
diff --git a/paddler_image_decoder/src/decoded_image.rs b/paddler_image_decoder/src/decoded_image.rs
index 4b6296342..2fea27d40 100644
--- a/paddler_image_decoder/src/decoded_image.rs
+++ b/paddler_image_decoder/src/decoded_image.rs
@@ -63,7 +63,9 @@ fn rasterize_svg(
         height: target_height,
         rgb_pixels: pixmap
             .data()
-            .chunks_exact(RGBA_BYTES_PER_PIXEL)
+            .as_chunks::<RGBA_BYTES_PER_PIXEL>()
+            .0
+            .iter()
             .flat_map(|rgba_pixel| &rgba_pixel[..RGB_BYTES_PER_PIXEL])
             .copied()
             .collect(),
diff --git a/paddler_inference_parameters/Cargo.toml b/paddler_inference_parameters/Cargo.toml
index de07fb3df..6b5e54e05 100644
--- a/paddler_inference_parameters/Cargo.toml
+++ b/paddler_inference_parameters/Cargo.toml
@@ -9,6 +9,7 @@ repository.workspace = true
 version.workspace = true
 
 [dependencies]
+llama-cpp-bindings-types = { workspace = true }
 serde = { workspace = true }
 thiserror = { workspace = true }
 
diff --git a/paddler_inference_parameters/src/batch_size.rs b/paddler_inference_parameters/src/batch_size.rs
index d9658d05a..04b129029 100644
--- a/paddler_inference_parameters/src/batch_size.rs
+++ b/paddler_inference_parameters/src/batch_size.rs
@@ -1,32 +1,49 @@
+use std::num::NonZeroU16;
 use std::num::NonZeroU32;
 
+use llama_cpp_bindings_types::BatchCapacity;
+use llama_cpp_bindings_types::BatchCapacityError;
 use serde::Deserialize;
 use serde::Serialize;
 
 use crate::invalid_inference_parameters::InvalidInferenceParameters;
 
-const MAX_BATCH_SIZE: u32 = i32::MAX.unsigned_abs();
+const DEFAULT_TOKENS: NonZeroU16 = NonZeroU16::new(2048).unwrap();
 
 #[derive(Clone, Copy, Debug, Deserialize, Eq, PartialEq, Serialize)]
 #[serde(try_from = "u32", into = "u32")]
-pub struct BatchSize(NonZeroU32);
+pub struct BatchSize(BatchCapacity);
 
 impl BatchSize {
-    pub const DEFAULT: Self = Self(NonZeroU32::new(2048).unwrap());
+    #[must_use]
+    pub fn at_most(self, tokens: NonZeroU32) -> Self {
+        Self(self.0.at_most(tokens))
+    }
 
     #[must_use]
-    pub const fn tokens(self) -> NonZeroU32 {
+    pub const fn capacity(self) -> BatchCapacity {
         self.0
     }
 
+    #[must_use]
+    pub const fn tokens(self) -> NonZeroU32 {
+        self.0.tokens()
+    }
+
     #[must_use]
     pub const fn tokens_i32(self) -> i32 {
-        self.0.get().cast_signed()
+        self.0.tokens_i32()
     }
 
     #[must_use]
     pub const fn tokens_usize(self) -> usize {
-        self.0.get() as usize
+        self.0.tokens_usize()
+    }
+}
+
+impl Default for BatchSize {
+    fn default() -> Self {
+        Self(BatchCapacity::from(DEFAULT_TOKENS))
     }
 }
 
@@ -34,30 +51,43 @@ impl TryFrom<u32> for BatchSize {
     type Error = InvalidInferenceParameters;
 
     fn try_from(n_batch: u32) -> Result<Self, Self::Error> {
-        if n_batch > MAX_BATCH_SIZE {
-            return Err(InvalidInferenceParameters::BatchSizeTooLarge {
-                n_batch,
-                max_batch_size: MAX_BATCH_SIZE,
-            });
+        match BatchCapacity::try_from(n_batch) {
+            Ok(capacity) => Ok(Self(capacity)),
+            Err(BatchCapacityError::Zero) => Err(InvalidInferenceParameters::BatchSizeZero),
+            Err(BatchCapacityError::TooLarge { tokens, max_tokens }) => {
+                Err(InvalidInferenceParameters::BatchSizeTooLarge {
+                    n_batch: tokens,
+                    max_batch_size: max_tokens,
+                })
+            }
         }
-
-        NonZeroU32::new(n_batch)
-            .map(Self)
-            .ok_or(InvalidInferenceParameters::BatchSizeZero)
     }
 }
 
 impl From<BatchSize> for u32 {
     fn from(batch_size: BatchSize) -> Self {
-        batch_size.0.get()
+        batch_size.0.tokens().get()
     }
 }
 
 #[cfg(test)]
 mod tests {
+    use std::num::NonZeroU32;
+
     use super::BatchSize;
     use crate::invalid_inference_parameters::InvalidInferenceParameters;
 
+    #[test]
+    fn a_batch_never_exceeds_the_token_bound_it_is_capped_at() {
+        assert_eq!(
+            [
+                BatchSize::default().at_most(NonZeroU32::new(512).unwrap()),
+                BatchSize::default().at_most(NonZeroU32::new(4096).unwrap()),
+            ],
+            [BatchSize::try_from(512).unwrap(), BatchSize::default()]
+        );
+    }
+
     #[test]
     fn rejects_a_zero_batch() {
         assert_eq!(
diff --git a/paddler_inference_parameters/src/embedding_parameters.rs b/paddler_inference_parameters/src/embedding_parameters.rs
new file mode 100644
index 000000000..51181f686
--- /dev/null
+++ b/paddler_inference_parameters/src/embedding_parameters.rs
@@ -0,0 +1,69 @@
+use std::num::NonZeroUsize;
+
+use serde::Deserialize;
+use serde::Serialize;
+
+use crate::invalid_inference_parameters::InvalidInferenceParameters;
+use crate::pooling_type::PoolingType;
+use crate::raw_embedding_parameters::RawEmbeddingParameters;
+
+const DEFAULT_EMBEDDING_BATCH_SIZE: NonZeroUsize = NonZeroUsize::new(256).unwrap();
+
+#[derive(Clone, Debug, Deserialize, Eq, PartialEq, Serialize)]
+#[serde(try_from = "RawEmbeddingParameters")]
+pub struct EmbeddingParameters {
+    pub embedding_batch_size: NonZeroUsize,
+    pub pooling_type: PoolingType,
+}
+
+impl Default for EmbeddingParameters {
+    fn default() -> Self {
+        Self {
+            embedding_batch_size: DEFAULT_EMBEDDING_BATCH_SIZE,
+            pooling_type: PoolingType::Unspecified,
+        }
+    }
+}
+
+impl TryFrom<RawEmbeddingParameters> for EmbeddingParameters {
+    type Error = InvalidInferenceParameters;
+
+    fn try_from(
+        RawEmbeddingParameters {
+            embedding_batch_size,
+            pooling_type,
+        }: RawEmbeddingParameters,
+    ) -> Result<Self, Self::Error> {
+        if pooling_type == PoolingType::None {
+            return Err(InvalidInferenceParameters::PoolingProducesNoEmbeddings);
+        }
+
+        Ok(Self {
+            embedding_batch_size,
+            pooling_type,
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use serde_json::from_value;
+    use serde_json::json;
+
+    use super::EmbeddingParameters;
+    use crate::invalid_inference_parameters::InvalidInferenceParameters;
+
+    #[test]
+    fn deserialization_rejects_pooling_that_produces_no_embeddings() {
+        assert_eq!(
+            from_value::<EmbeddingParameters>(json!({
+                "embedding_batch_size": 256,
+                "pooling_type": "None",
+            }))
+            .err()
+            .as_ref()
+            .map(ToString::to_string),
+            Some(InvalidInferenceParameters::PoolingProducesNoEmbeddings.to_string())
+        );
+    }
+}
diff --git a/paddler_inference_parameters/src/inference_parameters.rs b/paddler_inference_parameters/src/inference_parameters.rs
deleted file mode 100644
index 2f80b0252..000000000
--- a/paddler_inference_parameters/src/inference_parameters.rs
+++ /dev/null
@@ -1,420 +0,0 @@
-use std::cmp::Ordering;
-use std::num::NonZeroU32;
-use std::num::NonZeroUsize;
-
-use serde::Deserialize;
-use serde::Serialize;
-
-use crate::all_gpu_layers::ALL_GPU_LAYERS;
-use crate::batch_size::BatchSize;
-use crate::invalid_inference_parameters::InvalidInferenceParameters;
-use crate::kv_cache_dtype::KvCacheDtype;
-use crate::pooling_type::PoolingType;
-use crate::raw_inference_parameters::RawInferenceParameters;
-
-const DEFAULT_BATCH_SIZE: BatchSize = BatchSize::DEFAULT;
-const DEFAULT_CONTEXT_SIZE: NonZeroU32 = NonZeroU32::new(8192).unwrap();
-const DEFAULT_EMBEDDING_BATCH_SIZE: NonZeroUsize = NonZeroUsize::new(256).unwrap();
-const DEFAULT_IMAGE_RESIZE_TO_FIT: NonZeroU32 = NonZeroU32::new(1024).unwrap();
-const NEUTRAL_PENALTY_REPEAT: f32 = 1.0;
-
-fn is_probability(value: f32) -> bool {
-    (0.0..=1.0).contains(&value)
-}
-
-fn penalties_are_neutral(
-    penalty_repeat: f32,
-    penalty_frequency: f32,
-    penalty_presence: f32,
-) -> bool {
-    penalty_repeat.total_cmp(&NEUTRAL_PENALTY_REPEAT) == Ordering::Equal
-        && penalty_frequency == 0.0
-        && penalty_presence == 0.0
-}
-
-#[derive(Clone, Debug, Deserialize, PartialEq, Serialize)]
-#[serde(try_from = "RawInferenceParameters")]
-pub struct InferenceParameters {
-    pub n_batch: BatchSize,
-    pub context_size: NonZeroU32,
-    pub embedding_batch_size: NonZeroUsize,
-    pub enable_embeddings: bool,
-    pub image_resize_to_fit: NonZeroU32,
-    pub k_cache_dtype: KvCacheDtype,
-    pub v_cache_dtype: KvCacheDtype,
-    /// The minimum probability for a token to be considered, relative to the probability of the most likely token
-    pub min_p: f32,
-    /// Number of model layers to offload to GPU. 0 = CPU-only, -1 = offload all layers.
-    /// Set to a value >= the model's transformer block count for full GPU offload.
-    pub n_gpu_layers: i32,
-    pub penalty_frequency: f32,
-    /// How many tokens to scan for repetitions (0 = disabled)
-    pub penalty_last_n: i32,
-    pub penalty_presence: f32,
-    /// Penalty for repeating tokens (1.0 = disabled)
-    pub penalty_repeat: f32,
-    pub pooling_type: PoolingType,
-    /// Adjust the randomness of the generated text (0.0 = greedy/deterministic)
-    pub temperature: f32,
-    /// Limit the next token selection to the K most probable tokens
-    pub top_k: i32,
-    /// Limit the next token selection to a subset of tokens with a cumulative probability above a threshold P
-    pub top_p: f32,
-}
-
-impl InferenceParameters {
-    #[must_use]
-    pub fn deterministic() -> Self {
-        Self {
-            min_p: 0.0,
-            penalty_frequency: 0.0,
-            penalty_last_n: 0,
-            penalty_presence: 0.0,
-            penalty_repeat: NEUTRAL_PENALTY_REPEAT,
-            temperature: 0.0,
-            top_k: 1,
-            top_p: 1.0,
-            ..Self::default()
-        }
-    }
-}
-
-impl Default for InferenceParameters {
-    fn default() -> Self {
-        Self {
-            n_batch: DEFAULT_BATCH_SIZE,
-            context_size: DEFAULT_CONTEXT_SIZE,
-            embedding_batch_size: DEFAULT_EMBEDDING_BATCH_SIZE,
-            enable_embeddings: false,
-            image_resize_to_fit: DEFAULT_IMAGE_RESIZE_TO_FIT,
-            k_cache_dtype: KvCacheDtype::Q80,
-            v_cache_dtype: KvCacheDtype::Q80,
-            min_p: 0.05,
-            n_gpu_layers: 0,
-            penalty_frequency: 0.0,
-            penalty_last_n: 0,
-            penalty_presence: 0.0,
-            penalty_repeat: NEUTRAL_PENALTY_REPEAT,
-            pooling_type: PoolingType::Last,
-            temperature: 0.8,
-            top_k: 80,
-            top_p: 0.8,
-        }
-    }
-}
-
-impl TryFrom<RawInferenceParameters> for InferenceParameters {
-    type Error = InvalidInferenceParameters;
-
-    fn try_from(
-        RawInferenceParameters {
-            n_batch,
-            context_size,
-            embedding_batch_size,
-            enable_embeddings,
-            image_resize_to_fit,
-            k_cache_dtype,
-            v_cache_dtype,
-            min_p,
-            n_gpu_layers,
-            penalty_frequency,
-            penalty_last_n,
-            penalty_presence,
-            penalty_repeat,
-            pooling_type,
-            temperature,
-            top_k,
-            top_p,
-        }: RawInferenceParameters,
-    ) -> Result<Self, Self::Error> {
-        if n_batch.tokens() > context_size {
-            return Err(InvalidInferenceParameters::BatchSizeExceedsContextSize {
-                n_batch: n_batch.tokens().get(),
-                context_size: context_size.get(),
-            });
-        }
-
-        if n_gpu_layers < ALL_GPU_LAYERS {
-            return Err(InvalidInferenceParameters::GpuLayersBelowAllLayers { n_gpu_layers });
-        }
-
-        if !is_probability(min_p) {
-            return Err(InvalidInferenceParameters::MinPOutOfRange { min_p });
-        }
-
-        if !is_probability(top_p) {
-            return Err(InvalidInferenceParameters::TopPOutOfRange { top_p });
-        }
-
-        if !(temperature.is_finite() && temperature >= 0.0) {
-            return Err(InvalidInferenceParameters::TemperatureNegative { temperature });
-        }
-
-        if top_k < 0 {
-            return Err(InvalidInferenceParameters::TopKNegative { top_k });
-        }
-
-        if penalty_last_n < 0 {
-            return Err(InvalidInferenceParameters::PenaltyLastNNegative { penalty_last_n });
-        }
-
-        if !(penalty_repeat.is_finite() && penalty_repeat > 0.0) {
-            return Err(InvalidInferenceParameters::PenaltyRepeatNotPositive { penalty_repeat });
-        }
-
-        if !penalty_frequency.is_finite() {
-            return Err(InvalidInferenceParameters::PenaltyFrequencyNotFinite {
-                penalty_frequency,
-            });
-        }
-
-        if !penalty_presence.is_finite() {
-            return Err(InvalidInferenceParameters::PenaltyPresenceNotFinite { penalty_presence });
-        }
-
-        let penalties_are_neutral =
-            penalties_are_neutral(penalty_repeat, penalty_frequency, penalty_presence);
-
-        if penalty_last_n == 0 && !penalties_are_neutral {
-            return Err(InvalidInferenceParameters::PenaltiesWithoutWindow {
-                penalty_frequency,
-                penalty_presence,
-                penalty_repeat,
-            });
-        }
-
-        if penalty_last_n > 0 && penalties_are_neutral {
-            return Err(InvalidInferenceParameters::PenaltyWindowWithoutPenalties {
-                penalty_last_n,
-            });
-        }
-
-        Ok(Self {
-            n_batch,
-            context_size,
-            embedding_batch_size,
-            enable_embeddings,
-            image_resize_to_fit,
-            k_cache_dtype,
-            v_cache_dtype,
-            min_p,
-            n_gpu_layers,
-            penalty_frequency,
-            penalty_last_n,
-            penalty_presence,
-            penalty_repeat,
-            pooling_type,
-            temperature,
-            top_k,
-            top_p,
-        })
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use std::num::NonZeroU32;
-    use std::num::NonZeroUsize;
-
-    use serde_json::from_value;
-    use serde_json::json;
-    use serde_json::to_value;
-
-    use super::InferenceParameters;
-    use crate::batch_size::BatchSize;
-    use crate::invalid_inference_parameters::InvalidInferenceParameters;
-    use crate::kv_cache_dtype::KvCacheDtype;
-    use crate::pooling_type::PoolingType;
-    use crate::raw_inference_parameters::RawInferenceParameters;
-
-    fn valid_raw() -> RawInferenceParameters {
-        RawInferenceParameters {
-            n_batch: BatchSize::try_from(512).unwrap(),
-            context_size: NonZeroU32::new(4096).unwrap(),
-            embedding_batch_size: NonZeroUsize::new(64).unwrap(),
-            enable_embeddings: false,
-            image_resize_to_fit: NonZeroU32::new(512).unwrap(),
-            k_cache_dtype: KvCacheDtype::F16,
-            v_cache_dtype: KvCacheDtype::F16,
-            min_p: 0.05,
-            n_gpu_layers: -1,
-            penalty_frequency: 0.5,
-            penalty_last_n: 64,
-            penalty_presence: 0.0,
-            penalty_repeat: 1.1,
-            pooling_type: PoolingType::Mean,
-            temperature: 0.7,
-            top_k: 40,
-            top_p: 0.9,
-        }
-    }
-
-    fn rejection(raw: RawInferenceParameters) -> Option<InvalidInferenceParameters> {
-        InferenceParameters::try_from(raw).err()
-    }
-
-    #[test]
-    fn accepts_parameters_that_satisfy_every_rule() {
-        assert!(InferenceParameters::try_from(valid_raw()).is_ok());
-    }
-
-    #[test]
-    fn deserialization_rejects_parameters_that_fail_validation() {
-        let mut parameters = to_value(InferenceParameters::default()).unwrap();
-
-        parameters["top_k"] = json!(-1);
-
-        assert_eq!(
-            from_value::<InferenceParameters>(parameters)
-                .unwrap_err()
-                .to_string(),
-            InvalidInferenceParameters::TopKNegative { top_k: -1 }.to_string()
-        );
-    }
-
-    #[test]
-    fn rejects_a_batch_larger_than_the_context() {
-        assert_eq!(
-            rejection(RawInferenceParameters {
-                n_batch: BatchSize::try_from(8192).unwrap(),
-                ..valid_raw()
-            }),
-            Some(InvalidInferenceParameters::BatchSizeExceedsContextSize {
-                n_batch: 8192,
-                context_size: 4096,
-            })
-        );
-    }
-
-    #[test]
-    fn rejects_gpu_layers_below_the_all_layers_value() {
-        assert_eq!(
-            rejection(RawInferenceParameters {
-                n_gpu_layers: -2,
-                ..valid_raw()
-            }),
-            Some(InvalidInferenceParameters::GpuLayersBelowAllLayers { n_gpu_layers: -2 })
-        );
-    }
-
-    #[test]
-    fn rejects_min_p_outside_zero_to_one() {
-        assert_eq!(
-            rejection(RawInferenceParameters {
-                min_p: 1.5,
-                ..valid_raw()
-            }),
-            Some(InvalidInferenceParameters::MinPOutOfRange { min_p: 1.5 })
-        );
-    }
-
-    #[test]
-    fn rejects_top_p_outside_zero_to_one() {
-        assert_eq!(
-            rejection(RawInferenceParameters {
-                top_p: -0.1,
-                ..valid_raw()
-            }),
-            Some(InvalidInferenceParameters::TopPOutOfRange { top_p: -0.1 })
-        );
-    }
-
-    #[test]
-    fn rejects_a_negative_temperature() {
-        assert_eq!(
-            rejection(RawInferenceParameters {
-                temperature: -0.5,
-                ..valid_raw()
-            }),
-            Some(InvalidInferenceParameters::TemperatureNegative { temperature: -0.5 })
-        );
-    }
-
-    #[test]
-    fn rejects_a_negative_top_k() {
-        assert_eq!(
-            rejection(RawInferenceParameters {
-                top_k: -1,
-                ..valid_raw()
-            }),
-            Some(InvalidInferenceParameters::TopKNegative { top_k: -1 })
-        );
-    }
-
-    #[test]
-    fn rejects_a_negative_penalty_window() {
-        assert_eq!(
-            rejection(RawInferenceParameters {
-                penalty_last_n: -1,
-                ..valid_raw()
-            }),
-            Some(InvalidInferenceParameters::PenaltyLastNNegative { penalty_last_n: -1 })
-        );
-    }
-
-    #[test]
-    fn rejects_a_repeat_penalty_that_is_not_positive() {
-        assert_eq!(
-            rejection(RawInferenceParameters {
-                penalty_repeat: 0.0,
-                ..valid_raw()
-            }),
-            Some(InvalidInferenceParameters::PenaltyRepeatNotPositive {
-                penalty_repeat: 0.0
-            })
-        );
-    }
-
-    #[test]
-    fn rejects_an_infinite_frequency_penalty() {
-        assert_eq!(
-            rejection(RawInferenceParameters {
-                penalty_frequency: f32::INFINITY,
-                ..valid_raw()
-            }),
-            Some(InvalidInferenceParameters::PenaltyFrequencyNotFinite {
-                penalty_frequency: f32::INFINITY,
-            })
-        );
-    }
-
-    #[test]
-    fn rejects_an_infinite_presence_penalty() {
-        assert_eq!(
-            rejection(RawInferenceParameters {
-                penalty_presence: f32::NEG_INFINITY,
-                ..valid_raw()
-            }),
-            Some(InvalidInferenceParameters::PenaltyPresenceNotFinite {
-                penalty_presence: f32::NEG_INFINITY,
-            })
-        );
-    }
-
-    #[test]
-    fn rejects_penalty_strengths_without_a_penalty_window() {
-        assert_eq!(
-            rejection(RawInferenceParameters {
-                penalty_last_n: 0,
-                ..valid_raw()
-            }),
-            Some(InvalidInferenceParameters::PenaltiesWithoutWindow {
-                penalty_frequency: 0.5,
-                penalty_presence: 0.0,
-                penalty_repeat: 1.1,
-            })
-        );
-    }
-
-    #[test]
-    fn rejects_a_penalty_window_without_penalty_strengths() {
-        assert_eq!(
-            rejection(RawInferenceParameters {
-                penalty_frequency: 0.0,
-                penalty_repeat: 1.0,
-                ..valid_raw()
-            }),
-            Some(InvalidInferenceParameters::PenaltyWindowWithoutPenalties { penalty_last_n: 64 })
-        );
-    }
-}
diff --git a/paddler_inference_parameters/src/invalid_inference_parameters.rs b/paddler_inference_parameters/src/invalid_inference_parameters.rs
index ccc0e2018..95d89f78e 100644
--- a/paddler_inference_parameters/src/invalid_inference_parameters.rs
+++ b/paddler_inference_parameters/src/invalid_inference_parameters.rs
@@ -30,6 +30,17 @@ pub enum InvalidInferenceParameters {
         "penalty_last_n {penalty_last_n} is set, but every penalty strength is neutral, so no penalty applies"
     )]
     PenaltyWindowWithoutPenalties { penalty_last_n: i32 },
+    #[error(
+        "penalty_last_n {penalty_last_n} exceeds the sequence context of {sequence_context_size} tokens"
+    )]
+    PenaltyWindowExceedsSequenceContext {
+        penalty_last_n: i32,
+        sequence_context_size: u32,
+    },
+    #[error(
+        "pooling_type None produces no embeddings; choose Mean, Cls, Last, Rank, or Unspecified"
+    )]
+    PoolingProducesNoEmbeddings,
     #[error("temperature {temperature} must be a finite number of at least 0")]
     TemperatureNegative { temperature: f32 },
     #[error("top_k {top_k} is negative; use 0 to disable top-k sampling")]
diff --git a/paddler_inference_parameters/src/lib.rs b/paddler_inference_parameters/src/lib.rs
index 99718ea39..ed81372db 100644
--- a/paddler_inference_parameters/src/lib.rs
+++ b/paddler_inference_parameters/src/lib.rs
@@ -1,7 +1,11 @@
 pub mod all_gpu_layers;
 pub mod batch_size;
-pub mod inference_parameters;
+pub mod embedding_parameters;
 pub mod invalid_inference_parameters;
 pub mod kv_cache_dtype;
+pub mod model_runtime_parameters;
 pub mod pooling_type;
-pub mod raw_inference_parameters;
+pub mod raw_embedding_parameters;
+pub mod raw_model_runtime_parameters;
+pub mod raw_sampling_parameters;
+pub mod sampling_parameters;
diff --git a/paddler_inference_parameters/src/model_runtime_parameters.rs b/paddler_inference_parameters/src/model_runtime_parameters.rs
new file mode 100644
index 000000000..9594f5679
--- /dev/null
+++ b/paddler_inference_parameters/src/model_runtime_parameters.rs
@@ -0,0 +1,140 @@
+use std::num::NonZeroU32;
+
+use serde::Deserialize;
+use serde::Serialize;
+
+use crate::all_gpu_layers::ALL_GPU_LAYERS;
+use crate::batch_size::BatchSize;
+use crate::invalid_inference_parameters::InvalidInferenceParameters;
+use crate::kv_cache_dtype::KvCacheDtype;
+use crate::raw_model_runtime_parameters::RawModelRuntimeParameters;
+
+const DEFAULT_CONTEXT_SIZE: NonZeroU32 = NonZeroU32::new(8192).unwrap();
+
+#[derive(Clone, Debug, Deserialize, Eq, PartialEq, Serialize)]
+#[serde(try_from = "RawModelRuntimeParameters")]
+pub struct ModelRuntimeParameters {
+    pub context_size: NonZeroU32,
+    pub k_cache_dtype: KvCacheDtype,
+    pub n_batch: BatchSize,
+    /// Number of model layers to offload to GPU. 0 = CPU-only, -1 = offload all layers.
+    /// Set to a value >= the model's transformer block count for full GPU offload.
+    pub n_gpu_layers: i32,
+    pub v_cache_dtype: KvCacheDtype,
+}
+
+impl Default for ModelRuntimeParameters {
+    fn default() -> Self {
+        Self {
+            context_size: DEFAULT_CONTEXT_SIZE,
+            k_cache_dtype: KvCacheDtype::Q80,
+            n_batch: BatchSize::default(),
+            n_gpu_layers: 0,
+            v_cache_dtype: KvCacheDtype::Q80,
+        }
+    }
+}
+
+impl TryFrom<RawModelRuntimeParameters> for ModelRuntimeParameters {
+    type Error = InvalidInferenceParameters;
+
+    fn try_from(
+        RawModelRuntimeParameters {
+            context_size,
+            k_cache_dtype,
+            n_batch,
+            n_gpu_layers,
+            v_cache_dtype,
+        }: RawModelRuntimeParameters,
+    ) -> Result<Self, Self::Error> {
+        if n_batch.tokens() > context_size {
+            return Err(InvalidInferenceParameters::BatchSizeExceedsContextSize {
+                n_batch: n_batch.tokens().get(),
+                context_size: context_size.get(),
+            });
+        }
+
+        if n_gpu_layers < ALL_GPU_LAYERS {
+            return Err(InvalidInferenceParameters::GpuLayersBelowAllLayers { n_gpu_layers });
+        }
+
+        Ok(Self {
+            context_size,
+            k_cache_dtype,
+            n_batch,
+            n_gpu_layers,
+            v_cache_dtype,
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::num::NonZeroU32;
+
+    use serde_json::from_value;
+    use serde_json::json;
+    use serde_json::to_value;
+
+    use super::ModelRuntimeParameters;
+    use crate::batch_size::BatchSize;
+    use crate::invalid_inference_parameters::InvalidInferenceParameters;
+    use crate::kv_cache_dtype::KvCacheDtype;
+    use crate::raw_model_runtime_parameters::RawModelRuntimeParameters;
+
+    fn valid_raw() -> RawModelRuntimeParameters {
+        RawModelRuntimeParameters {
+            context_size: NonZeroU32::new(4096).unwrap(),
+            k_cache_dtype: KvCacheDtype::F16,
+            n_batch: BatchSize::try_from(512).unwrap(),
+            n_gpu_layers: -1,
+            v_cache_dtype: KvCacheDtype::F16,
+        }
+    }
+
+    #[test]
+    fn accepts_parameters_that_satisfy_every_rule() {
+        assert!(ModelRuntimeParameters::try_from(valid_raw()).is_ok());
+    }
+
+    #[test]
+    fn deserialization_rejects_parameters_that_fail_validation() {
+        let mut parameters = to_value(ModelRuntimeParameters::default()).unwrap();
+
+        parameters["n_gpu_layers"] = json!(-2);
+
+        assert_eq!(
+            from_value::<ModelRuntimeParameters>(parameters)
+                .unwrap_err()
+                .to_string(),
+            InvalidInferenceParameters::GpuLayersBelowAllLayers { n_gpu_layers: -2 }.to_string()
+        );
+    }
+
+    #[test]
+    fn rejects_a_batch_larger_than_the_context() {
+        assert_eq!(
+            ModelRuntimeParameters::try_from(RawModelRuntimeParameters {
+                n_batch: BatchSize::try_from(8192).unwrap(),
+                ..valid_raw()
+            })
+            .err(),
+            Some(InvalidInferenceParameters::BatchSizeExceedsContextSize {
+                n_batch: 8192,
+                context_size: 4096,
+            })
+        );
+    }
+
+    #[test]
+    fn rejects_gpu_layers_below_the_all_layers_value() {
+        assert_eq!(
+            ModelRuntimeParameters::try_from(RawModelRuntimeParameters {
+                n_gpu_layers: -2,
+                ..valid_raw()
+            })
+            .err(),
+            Some(InvalidInferenceParameters::GpuLayersBelowAllLayers { n_gpu_layers: -2 })
+        );
+    }
+}
diff --git a/paddler_inference_parameters/src/raw_embedding_parameters.rs b/paddler_inference_parameters/src/raw_embedding_parameters.rs
new file mode 100644
index 000000000..efb7c84d1
--- /dev/null
+++ b/paddler_inference_parameters/src/raw_embedding_parameters.rs
@@ -0,0 +1,12 @@
+use std::num::NonZeroUsize;
+
+use serde::Deserialize;
+
+use crate::pooling_type::PoolingType;
+
+#[derive(Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct RawEmbeddingParameters {
+    pub embedding_batch_size: NonZeroUsize,
+    pub pooling_type: PoolingType,
+}
diff --git a/paddler_inference_parameters/src/raw_inference_parameters.rs b/paddler_inference_parameters/src/raw_inference_parameters.rs
deleted file mode 100644
index 8398f1a50..000000000
--- a/paddler_inference_parameters/src/raw_inference_parameters.rs
+++ /dev/null
@@ -1,30 +0,0 @@
-use std::num::NonZeroU32;
-use std::num::NonZeroUsize;
-
-use serde::Deserialize;
-
-use crate::batch_size::BatchSize;
-use crate::kv_cache_dtype::KvCacheDtype;
-use crate::pooling_type::PoolingType;
-
-#[derive(Deserialize)]
-#[serde(deny_unknown_fields)]
-pub struct RawInferenceParameters {
-    pub n_batch: BatchSize,
-    pub context_size: NonZeroU32,
-    pub embedding_batch_size: NonZeroUsize,
-    pub enable_embeddings: bool,
-    pub image_resize_to_fit: NonZeroU32,
-    pub k_cache_dtype: KvCacheDtype,
-    pub v_cache_dtype: KvCacheDtype,
-    pub min_p: f32,
-    pub n_gpu_layers: i32,
-    pub penalty_frequency: f32,
-    pub penalty_last_n: i32,
-    pub penalty_presence: f32,
-    pub penalty_repeat: f32,
-    pub pooling_type: PoolingType,
-    pub temperature: f32,
-    pub top_k: i32,
-    pub top_p: f32,
-}
diff --git a/paddler_inference_parameters/src/raw_model_runtime_parameters.rs b/paddler_inference_parameters/src/raw_model_runtime_parameters.rs
new file mode 100644
index 000000000..155d828bc
--- /dev/null
+++ b/paddler_inference_parameters/src/raw_model_runtime_parameters.rs
@@ -0,0 +1,16 @@
+use std::num::NonZeroU32;
+
+use serde::Deserialize;
+
+use crate::batch_size::BatchSize;
+use crate::kv_cache_dtype::KvCacheDtype;
+
+#[derive(Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct RawModelRuntimeParameters {
+    pub context_size: NonZeroU32,
+    pub k_cache_dtype: KvCacheDtype,
+    pub n_batch: BatchSize,
+    pub n_gpu_layers: i32,
+    pub v_cache_dtype: KvCacheDtype,
+}
diff --git a/paddler_inference_parameters/src/raw_sampling_parameters.rs b/paddler_inference_parameters/src/raw_sampling_parameters.rs
new file mode 100644
index 000000000..7affec0b1
--- /dev/null
+++ b/paddler_inference_parameters/src/raw_sampling_parameters.rs
@@ -0,0 +1,14 @@
+use serde::Deserialize;
+
+#[derive(Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct RawSamplingParameters {
+    pub min_p: f32,
+    pub penalty_frequency: f32,
+    pub penalty_last_n: i32,
+    pub penalty_presence: f32,
+    pub penalty_repeat: f32,
+    pub temperature: f32,
+    pub top_k: i32,
+    pub top_p: f32,
+}
diff --git a/paddler_inference_parameters/src/sampling_parameters.rs b/paddler_inference_parameters/src/sampling_parameters.rs
new file mode 100644
index 000000000..a3a76d712
--- /dev/null
+++ b/paddler_inference_parameters/src/sampling_parameters.rs
@@ -0,0 +1,361 @@
+use std::cmp::Ordering;
+
+use serde::Deserialize;
+use serde::Serialize;
+
+use crate::invalid_inference_parameters::InvalidInferenceParameters;
+use crate::raw_sampling_parameters::RawSamplingParameters;
+
+const NEUTRAL_PENALTY_REPEAT: f32 = 1.0;
+
+fn is_probability(value: f32) -> bool {
+    (0.0..=1.0).contains(&value)
+}
+
+fn penalties_are_neutral(
+    penalty_repeat: f32,
+    penalty_frequency: f32,
+    penalty_presence: f32,
+) -> bool {
+    penalty_repeat.total_cmp(&NEUTRAL_PENALTY_REPEAT) == Ordering::Equal
+        && penalty_frequency == 0.0
+        && penalty_presence == 0.0
+}
+
+#[derive(Clone, Debug, Deserialize, PartialEq, Serialize)]
+#[serde(try_from = "RawSamplingParameters")]
+pub struct SamplingParameters {
+    /// The minimum probability for a token to be considered, relative to the probability of the most likely token
+    pub min_p: f32,
+    pub penalty_frequency: f32,
+    /// How many tokens to scan for repetitions (0 = disabled)
+    pub penalty_last_n: i32,
+    pub penalty_presence: f32,
+    /// Penalty for repeating tokens (1.0 = disabled)
+    pub penalty_repeat: f32,
+    /// Adjust the randomness of the generated text (0.0 = greedy/deterministic)
+    pub temperature: f32,
+    /// Limit the next token selection to the K most probable tokens
+    pub top_k: i32,
+    /// Limit the next token selection to a subset of tokens with a cumulative probability above a threshold P
+    pub top_p: f32,
+}
+
+impl SamplingParameters {
+    #[must_use]
+    pub const fn deterministic() -> Self {
+        Self {
+            min_p: 0.0,
+            penalty_frequency: 0.0,
+            penalty_last_n: 0,
+            penalty_presence: 0.0,
+            penalty_repeat: NEUTRAL_PENALTY_REPEAT,
+            temperature: 0.0,
+            top_k: 1,
+            top_p: 1.0,
+        }
+    }
+
+    pub const fn ensure_penalty_window_fits(
+        &self,
+        sequence_context_size: u32,
+    ) -> Result<(), InvalidInferenceParameters> {
+        if self.penalty_last_n as i64 > sequence_context_size as i64 {
+            return Err(
+                InvalidInferenceParameters::PenaltyWindowExceedsSequenceContext {
+                    penalty_last_n: self.penalty_last_n,
+                    sequence_context_size,
+                },
+            );
+        }
+
+        Ok(())
+    }
+}
+
+impl Default for SamplingParameters {
+    fn default() -> Self {
+        Self {
+            min_p: 0.05,
+            penalty_frequency: 0.0,
+            penalty_last_n: 0,
+            penalty_presence: 0.0,
+            penalty_repeat: NEUTRAL_PENALTY_REPEAT,
+            temperature: 0.8,
+            top_k: 80,
+            top_p: 0.8,
+        }
+    }
+}
+
+impl TryFrom<RawSamplingParameters> for SamplingParameters {
+    type Error = InvalidInferenceParameters;
+
+    fn try_from(
+        RawSamplingParameters {
+            min_p,
+            penalty_frequency,
+            penalty_last_n,
+            penalty_presence,
+            penalty_repeat,
+            temperature,
+            top_k,
+            top_p,
+        }: RawSamplingParameters,
+    ) -> Result<Self, Self::Error> {
+        if !is_probability(min_p) {
+            return Err(InvalidInferenceParameters::MinPOutOfRange { min_p });
+        }
+
+        if !is_probability(top_p) {
+            return Err(InvalidInferenceParameters::TopPOutOfRange { top_p });
+        }
+
+        if !(temperature.is_finite() && temperature >= 0.0) {
+            return Err(InvalidInferenceParameters::TemperatureNegative { temperature });
+        }
+
+        if top_k < 0 {
+            return Err(InvalidInferenceParameters::TopKNegative { top_k });
+        }
+
+        if penalty_last_n < 0 {
+            return Err(InvalidInferenceParameters::PenaltyLastNNegative { penalty_last_n });
+        }
+
+        if !(penalty_repeat.is_finite() && penalty_repeat > 0.0) {
+            return Err(InvalidInferenceParameters::PenaltyRepeatNotPositive { penalty_repeat });
+        }
+
+        if !penalty_frequency.is_finite() {
+            return Err(InvalidInferenceParameters::PenaltyFrequencyNotFinite {
+                penalty_frequency,
+            });
+        }
+
+        if !penalty_presence.is_finite() {
+            return Err(InvalidInferenceParameters::PenaltyPresenceNotFinite { penalty_presence });
+        }
+
+        let penalties_are_neutral =
+            penalties_are_neutral(penalty_repeat, penalty_frequency, penalty_presence);
+
+        if penalty_last_n == 0 && !penalties_are_neutral {
+            return Err(InvalidInferenceParameters::PenaltiesWithoutWindow {
+                penalty_frequency,
+                penalty_presence,
+                penalty_repeat,
+            });
+        }
+
+        if penalty_last_n > 0 && penalties_are_neutral {
+            return Err(InvalidInferenceParameters::PenaltyWindowWithoutPenalties {
+                penalty_last_n,
+            });
+        }
+
+        Ok(Self {
+            min_p,
+            penalty_frequency,
+            penalty_last_n,
+            penalty_presence,
+            penalty_repeat,
+            temperature,
+            top_k,
+            top_p,
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use serde_json::from_value;
+    use serde_json::json;
+    use serde_json::to_value;
+
+    use super::SamplingParameters;
+    use crate::invalid_inference_parameters::InvalidInferenceParameters;
+    use crate::raw_sampling_parameters::RawSamplingParameters;
+
+    fn valid_raw() -> RawSamplingParameters {
+        RawSamplingParameters {
+            min_p: 0.05,
+            penalty_frequency: 0.5,
+            penalty_last_n: 64,
+            penalty_presence: 0.0,
+            penalty_repeat: 1.1,
+            temperature: 0.7,
+            top_k: 40,
+            top_p: 0.9,
+        }
+    }
+
+    fn rejection(raw: RawSamplingParameters) -> Option<InvalidInferenceParameters> {
+        SamplingParameters::try_from(raw).err()
+    }
+
+    #[test]
+    fn accepts_a_penalty_window_as_long_as_the_sequence_context() {
+        assert_eq!(
+            SamplingParameters::try_from(valid_raw())
+                .unwrap()
+                .ensure_penalty_window_fits(64),
+            Ok(())
+        );
+    }
+
+    #[test]
+    fn rejects_a_penalty_window_longer_than_the_sequence_context() {
+        assert_eq!(
+            SamplingParameters::try_from(valid_raw())
+                .unwrap()
+                .ensure_penalty_window_fits(63),
+            Err(
+                InvalidInferenceParameters::PenaltyWindowExceedsSequenceContext {
+                    penalty_last_n: 64,
+                    sequence_context_size: 63,
+                }
+            )
+        );
+    }
+
+    #[test]
+    fn accepts_parameters_that_satisfy_every_rule() {
+        assert!(SamplingParameters::try_from(valid_raw()).is_ok());
+    }
+
+    #[test]
+    fn deserialization_rejects_parameters_that_fail_validation() {
+        let mut parameters = to_value(SamplingParameters::default()).unwrap();
+
+        parameters["top_k"] = json!(-1);
+
+        assert_eq!(
+            from_value::<SamplingParameters>(parameters)
+                .unwrap_err()
+                .to_string(),
+            InvalidInferenceParameters::TopKNegative { top_k: -1 }.to_string()
+        );
+    }
+
+    #[test]
+    fn rejects_min_p_outside_zero_to_one() {
+        assert_eq!(
+            rejection(RawSamplingParameters {
+                min_p: 1.5,
+                ..valid_raw()
+            }),
+            Some(InvalidInferenceParameters::MinPOutOfRange { min_p: 1.5 })
+        );
+    }
+
+    #[test]
+    fn rejects_top_p_outside_zero_to_one() {
+        assert_eq!(
+            rejection(RawSamplingParameters {
+                top_p: -0.1,
+                ..valid_raw()
+            }),
+            Some(InvalidInferenceParameters::TopPOutOfRange { top_p: -0.1 })
+        );
+    }
+
+    #[test]
+    fn rejects_a_negative_temperature() {
+        assert_eq!(
+            rejection(RawSamplingParameters {
+                temperature: -0.5,
+                ..valid_raw()
+            }),
+            Some(InvalidInferenceParameters::TemperatureNegative { temperature: -0.5 })
+        );
+    }
+
+    #[test]
+    fn rejects_a_negative_top_k() {
+        assert_eq!(
+            rejection(RawSamplingParameters {
+                top_k: -1,
+                ..valid_raw()
+            }),
+            Some(InvalidInferenceParameters::TopKNegative { top_k: -1 })
+        );
+    }
+
+    #[test]
+    fn rejects_a_negative_penalty_window() {
+        assert_eq!(
+            rejection(RawSamplingParameters {
+                penalty_last_n: -1,
+                ..valid_raw()
+            }),
+            Some(InvalidInferenceParameters::PenaltyLastNNegative { penalty_last_n: -1 })
+        );
+    }
+
+    #[test]
+    fn rejects_a_repeat_penalty_that_is_not_positive() {
+        assert_eq!(
+            rejection(RawSamplingParameters {
+                penalty_repeat: 0.0,
+                ..valid_raw()
+            }),
+            Some(InvalidInferenceParameters::PenaltyRepeatNotPositive {
+                penalty_repeat: 0.0
+            })
+        );
+    }
+
+    #[test]
+    fn rejects_an_infinite_frequency_penalty() {
+        assert_eq!(
+            rejection(RawSamplingParameters {
+                penalty_frequency: f32::INFINITY,
+                ..valid_raw()
+            }),
+            Some(InvalidInferenceParameters::PenaltyFrequencyNotFinite {
+                penalty_frequency: f32::INFINITY,
+            })
+        );
+    }
+
+    #[test]
+    fn rejects_an_infinite_presence_penalty() {
+        assert_eq!(
+            rejection(RawSamplingParameters {
+                penalty_presence: f32::NEG_INFINITY,
+                ..valid_raw()
+            }),
+            Some(InvalidInferenceParameters::PenaltyPresenceNotFinite {
+                penalty_presence: f32::NEG_INFINITY,
+            })
+        );
+    }
+
+    #[test]
+    fn rejects_penalty_strengths_without_a_penalty_window() {
+        assert_eq!(
+            rejection(RawSamplingParameters {
+                penalty_last_n: 0,
+                ..valid_raw()
+            }),
+            Some(InvalidInferenceParameters::PenaltiesWithoutWindow {
+                penalty_frequency: 0.5,
+                penalty_presence: 0.0,
+                penalty_repeat: 1.1,
+            })
+        );
+    }
+
+    #[test]
+    fn rejects_a_penalty_window_without_penalty_strengths() {
+        assert_eq!(
+            rejection(RawSamplingParameters {
+                penalty_frequency: 0.0,
+                penalty_repeat: 1.0,
+                ..valid_raw()
+            }),
+            Some(InvalidInferenceParameters::PenaltyWindowWithoutPenalties { penalty_last_n: 64 })
+        );
+    }
+}
diff --git a/paddler_kev_converter_python/.gitignore b/paddler_kev_converter_python/.gitignore
new file mode 100644
index 000000000..d1e99ba9e
--- /dev/null
+++ b/paddler_kev_converter_python/.gitignore
@@ -0,0 +1,20 @@
+__pycache__/
+*.py[cod]
+*$py.class
+*.so
+dist/
+build/
+*.egg-info/
+*.egg
+.eggs/
+*.whl
+.venv
+.venv/
+venv/
+.mypy_cache/
+.pytest_cache/
+*.cover
+*.log
+.coverage
+htmlcov/
+target/
diff --git a/paddler_client_python/tests/__init__.py b/paddler_kev_converter_python/paddler_kev_converter/__init__.py
similarity index 100%
rename from paddler_client_python/tests/__init__.py
rename to paddler_kev_converter_python/paddler_kev_converter/__init__.py
diff --git a/paddler_kev_converter_python/paddler_kev_converter/command_line.py b/paddler_kev_converter_python/paddler_kev_converter/command_line.py
new file mode 100644
index 000000000..e40d4907a
--- /dev/null
+++ b/paddler_kev_converter_python/paddler_kev_converter/command_line.py
@@ -0,0 +1,94 @@
+from argparse import ArgumentParser, Namespace
+from pathlib import Path
+
+from paddler_kev_converter.convertible_kev_checkpoint import ConvertibleKevCheckpoint
+from paddler_kev_converter.input_normalization import InputNormalization
+from paddler_kev_converter.malformed_pointer_head import (
+    PointerHeadDefect,
+    write_malformed_pointer_head,
+)
+from paddler_kev_converter.parity_reference import write_parity_reference
+from paddler_kev_converter.synthetic_pointer_head import synthetic_pointer_head
+from paddler_kev_converter.tokenizer_reference import write_tokenizer_reference
+from paddler_kev_converter.typesafe_reference import write_typesafe_reference
+
+
+def _convert(parsed_arguments: Namespace) -> None:
+    checkpoint = ConvertibleKevCheckpoint.open(parsed_arguments.checkpoint)
+    checkpoint.write_backbone(parsed_arguments.backbone_directory)
+    checkpoint.pointer_head().write_gguf(parsed_arguments.pointer_head)
+
+
+def _write_malformed_pointer_head(parsed_arguments: Namespace) -> None:
+    write_malformed_pointer_head(
+        parsed_arguments.defect,
+        parsed_arguments.hidden_size,
+        parsed_arguments.input_normalization,
+        parsed_arguments.output,
+    )
+
+
+def _write_parity_reference(parsed_arguments: Namespace) -> None:
+    write_parity_reference(parsed_arguments.checkpoint, parsed_arguments.output)
+
+
+def _write_synthetic_pointer_head(parsed_arguments: Namespace) -> None:
+    synthetic_pointer_head(
+        parsed_arguments.hidden_size, parsed_arguments.input_normalization
+    ).write_gguf(parsed_arguments.output)
+
+
+def _write_tokenizer_reference(parsed_arguments: Namespace) -> None:
+    write_tokenizer_reference(
+        parsed_arguments.base, parsed_arguments.base_revision, parsed_arguments.output
+    )
+
+
+def _write_typesafe_reference(parsed_arguments: Namespace) -> None:
+    write_typesafe_reference(parsed_arguments.output)
+
+
+def main(arguments: list[str] | None = None) -> None:
+    parser = ArgumentParser(prog="paddler-kev-converter")
+    commands = parser.add_subparsers(required=True)
+
+    convert = commands.add_parser("convert")
+    convert.add_argument("--checkpoint", required=True)
+    convert.add_argument("--backbone-directory", required=True, type=Path)
+    convert.add_argument("--pointer-head", required=True, type=Path)
+    convert.set_defaults(handle=_convert)
+
+    malformed = commands.add_parser("malformed-pointer-head")
+    malformed.add_argument("--defect", required=True, type=PointerHeadDefect)
+    malformed.add_argument("--hidden-size", required=True, type=int)
+    malformed.add_argument(
+        "--input-normalization", required=True, type=InputNormalization
+    )
+    malformed.add_argument("--output", required=True, type=Path)
+    malformed.set_defaults(handle=_write_malformed_pointer_head)
+
+    parity = commands.add_parser("parity-reference")
+    parity.add_argument("--checkpoint", required=True)
+    parity.add_argument("--output", required=True, type=Path)
+    parity.set_defaults(handle=_write_parity_reference)
+
+    synthetic = commands.add_parser("synthetic-pointer-head")
+    synthetic.add_argument("--hidden-size", required=True, type=int)
+    synthetic.add_argument(
+        "--input-normalization", required=True, type=InputNormalization
+    )
+    synthetic.add_argument("--output", required=True, type=Path)
+    synthetic.set_defaults(handle=_write_synthetic_pointer_head)
+
+    tokenizer = commands.add_parser("tokenizer-reference")
+    tokenizer.add_argument("--base", required=True)
+    tokenizer.add_argument("--base-revision", required=True)
+    tokenizer.add_argument("--output", required=True, type=Path)
+    tokenizer.set_defaults(handle=_write_tokenizer_reference)
+
+    typesafe = commands.add_parser("typesafe-reference")
+    typesafe.add_argument("--output", required=True, type=Path)
+    typesafe.set_defaults(handle=_write_typesafe_reference)
+
+    parsed_arguments = parser.parse_args(arguments)
+    parsed_arguments.handle(parsed_arguments)
diff --git a/paddler_kev_converter_python/paddler_kev_converter/convertible_kev_checkpoint.py b/paddler_kev_converter_python/paddler_kev_converter/convertible_kev_checkpoint.py
new file mode 100644
index 000000000..037a18f84
--- /dev/null
+++ b/paddler_kev_converter_python/paddler_kev_converter/convertible_kev_checkpoint.py
@@ -0,0 +1,89 @@
+from dataclasses import dataclass
+from pathlib import Path
+
+import torch
+import transformers
+from kev.checkpoint import Checkpoint, LoadOptions
+from kev.model import SPECIAL
+from peft import LoraConfig
+from transformers import PreTrainedModel, PreTrainedTokenizerBase
+from transformers.models.auto.modeling_auto import MODEL_FOR_CAUSAL_LM_MAPPING_NAMES
+
+from paddler_kev_converter.error import (
+    BackboneDtypeUnsupportedError,
+    OptionIsolationUnsupportedError,
+    TrainedTokenEmbeddingsUnsupportedError,
+)
+from paddler_kev_converter.pointer_head import PointerHead
+from paddler_kev_converter.tokenizer_input_normalization import (
+    tokenizer_input_normalization,
+)
+
+EXACTLY_MERGING_BACKBONE_DTYPE = "fp32"
+
+
+def _require_exactly_mergeable_adapter(checkpoint: Checkpoint) -> None:
+    if checkpoint.meta.weights_dtype != EXACTLY_MERGING_BACKBONE_DTYPE:
+        raise BackboneDtypeUnsupportedError(
+            checkpoint.path, checkpoint.meta.weights_dtype
+        )
+
+    if LoraConfig.from_pretrained(checkpoint.path).trainable_token_indices:
+        raise TrainedTokenEmbeddingsUnsupportedError(checkpoint.path)
+
+
+@dataclass(frozen=True)
+class ConvertibleKevCheckpoint:
+    base_tokenizer: PreTrainedTokenizerBase
+    checkpoint: Checkpoint
+    decision_backbone: PreTrainedModel
+
+    @classmethod
+    def open(cls, run: str) -> "ConvertibleKevCheckpoint":
+        checkpoint = Checkpoint(run)
+
+        if checkpoint.meta.option_isolation:
+            raise OptionIsolationUnsupportedError(checkpoint.path)
+
+        if not checkpoint.full:
+            _require_exactly_mergeable_adapter(checkpoint)
+
+        base_tokenizer, decision_model = checkpoint.load("cpu", LoadOptions())
+
+        return cls(
+            base_tokenizer=base_tokenizer,
+            checkpoint=checkpoint,
+            decision_backbone=decision_model.lm,
+        )
+
+    def pointer_head(self) -> PointerHead:
+        meta = self.checkpoint.meta
+
+        return PointerHead(
+            delimiters=SPECIAL,
+            input_normalization=tokenizer_input_normalization(self.base_tokenizer),
+            key_bias=meta.head["k.bias"].numpy(),
+            key_weight=meta.head["k.weight"].numpy(),
+            query_bias=meta.head["q.bias"].numpy(),
+            query_weight=meta.head["q.weight"].numpy(),
+            temperature=meta.temperature,
+        )
+
+    def write_backbone(self, directory: Path) -> None:
+        backbone_config = self.decision_backbone.config
+        causal_language_model_class: type[PreTrainedModel] = getattr(
+            transformers, MODEL_FOR_CAUSAL_LM_MAPPING_NAMES[backbone_config.model_type]
+        )
+        backbone_config.tie_word_embeddings = True
+
+        with torch.device("meta"):
+            causal_language_model = causal_language_model_class(backbone_config)
+
+        setattr(
+            causal_language_model,
+            causal_language_model.base_model_prefix,
+            self.decision_backbone,
+        )
+        causal_language_model.tie_weights()
+        causal_language_model.save_pretrained(directory)
+        self.base_tokenizer.save_pretrained(directory)
diff --git a/paddler_kev_converter_python/paddler_kev_converter/error.py b/paddler_kev_converter_python/paddler_kev_converter/error.py
new file mode 100644
index 000000000..d8181b9dd
--- /dev/null
+++ b/paddler_kev_converter_python/paddler_kev_converter/error.py
@@ -0,0 +1,50 @@
+class KevConversionError(Exception):
+    pass
+
+
+class BackboneDtypeUnsupportedError(KevConversionError):
+    def __init__(self, checkpoint_path: str, weights_dtype: str) -> None:
+        self.checkpoint_path = checkpoint_path
+        self.weights_dtype = weights_dtype
+        super().__init__(
+            f"{checkpoint_path} was trained on a {weights_dtype} backbone; "
+            "only fp32 backbones merge exactly"
+        )
+
+
+class OptionIsolationUnsupportedError(KevConversionError):
+    def __init__(self, checkpoint_path: str) -> None:
+        self.checkpoint_path = checkpoint_path
+        super().__init__(
+            f"{checkpoint_path} isolates its options, "
+            "which needs a block-causal mask Paddler does not build"
+        )
+
+
+class TrainedTokenEmbeddingsUnsupportedError(KevConversionError):
+    def __init__(self, checkpoint_path: str) -> None:
+        self.checkpoint_path = checkpoint_path
+        super().__init__(
+            f"{checkpoint_path} trains token embeddings, "
+            "which cannot be merged into the backbone"
+        )
+
+
+class TokenizerNormalizerUnsupportedError(KevConversionError):
+    def __init__(self, tokenizer: str, normalizer: str) -> None:
+        self.tokenizer = tokenizer
+        self.normalizer = normalizer
+        super().__init__(
+            f"{tokenizer} normalizes its input with {normalizer}, "
+            "which a pointer head cannot declare"
+        )
+
+
+class TokenizerNormalizerUninspectableError(KevConversionError):
+    def __init__(self, tokenizer: str, tokenizer_class: str) -> None:
+        self.tokenizer = tokenizer
+        self.tokenizer_class = tokenizer_class
+        super().__init__(
+            f"{tokenizer} is a {tokenizer_class}, "
+            "which does not expose the normalizer it applies"
+        )
diff --git a/paddler_kev_converter_python/paddler_kev_converter/input_normalization.py b/paddler_kev_converter_python/paddler_kev_converter/input_normalization.py
new file mode 100644
index 000000000..f57f75cc2
--- /dev/null
+++ b/paddler_kev_converter_python/paddler_kev_converter/input_normalization.py
@@ -0,0 +1,9 @@
+from enum import StrEnum
+
+
+class InputNormalization(StrEnum):
+    NFC = "NFC"
+    NFD = "NFD"
+    NFKC = "NFKC"
+    NFKD = "NFKD"
+    NONE = "None"
diff --git a/paddler_kev_converter_python/paddler_kev_converter/malformed_pointer_head.py b/paddler_kev_converter_python/paddler_kev_converter/malformed_pointer_head.py
new file mode 100644
index 000000000..f6b31bd29
--- /dev/null
+++ b/paddler_kev_converter_python/paddler_kev_converter/malformed_pointer_head.py
@@ -0,0 +1,122 @@
+from collections.abc import Callable
+from dataclasses import replace
+from enum import StrEnum
+from pathlib import Path
+
+import numpy as np
+
+from paddler_kev_converter.input_normalization import InputNormalization
+from paddler_kev_converter.pointer_head import PointerHead
+from paddler_kev_converter.synthetic_pointer_head import synthetic_pointer_head
+
+OVERSIZED_HEAD_DIM = 2**16
+PLAIN_TEXT_DELIMITERS = ["state", "question", "option", "end", "decide"]
+
+
+class PointerHeadDefect(StrEnum):
+    F16_TENSORS = "f16-tensors"
+    MISMATCHED_PROJECTIONS = "mismatched-projections"
+    MISSING_DELIMITERS = "missing-delimiters"
+    MISSING_INPUT_NORMALIZATION = "missing-input-normalization"
+    MISSING_TEMPERATURE = "missing-temperature"
+    OVERSIZED_HEAD_DIM = "oversized-head-dim"
+    PLAIN_TEXT_DELIMITERS = "plain-text-delimiters"
+    ZERO_TEMPERATURE = "zero-temperature"
+
+
+def _write_with_f16_tensors(well_formed_head: PointerHead, path: Path) -> None:
+    replace(
+        well_formed_head,
+        key_bias=well_formed_head.key_bias.astype(np.float16),
+        key_weight=well_formed_head.key_weight.astype(np.float16),
+        query_bias=well_formed_head.query_bias.astype(np.float16),
+        query_weight=well_formed_head.query_weight.astype(np.float16),
+    ).write_gguf(path)
+
+
+def _write_with_mismatched_projections(
+    well_formed_head: PointerHead, path: Path
+) -> None:
+    replace(
+        well_formed_head,
+        query_bias=well_formed_head.query_bias[1:],
+        query_weight=well_formed_head.query_weight[1:],
+    ).write_gguf(path)
+
+
+def _write_without_delimiters(well_formed_head: PointerHead, path: Path) -> None:
+    well_formed_head.write_gguf_sections(
+        path,
+        [
+            well_formed_head.add_temperature,
+            well_formed_head.add_input_normalization,
+            well_formed_head.add_tensors,
+        ],
+    )
+
+
+def _write_without_input_normalization(
+    well_formed_head: PointerHead, path: Path
+) -> None:
+    well_formed_head.write_gguf_sections(
+        path,
+        [
+            well_formed_head.add_temperature,
+            well_formed_head.add_delimiters,
+            well_formed_head.add_tensors,
+        ],
+    )
+
+
+def _write_without_temperature(well_formed_head: PointerHead, path: Path) -> None:
+    well_formed_head.write_gguf_sections(
+        path,
+        [
+            well_formed_head.add_delimiters,
+            well_formed_head.add_input_normalization,
+            well_formed_head.add_tensors,
+        ],
+    )
+
+
+def _write_with_oversized_head_dim(well_formed_head: PointerHead, path: Path) -> None:
+    hidden_size = well_formed_head.key_weight.shape[1]
+
+    replace(
+        well_formed_head,
+        key_bias=np.zeros(OVERSIZED_HEAD_DIM, dtype=np.float32),
+        key_weight=np.zeros((OVERSIZED_HEAD_DIM, hidden_size), dtype=np.float32),
+        query_bias=np.zeros(OVERSIZED_HEAD_DIM, dtype=np.float32),
+        query_weight=np.zeros((OVERSIZED_HEAD_DIM, hidden_size), dtype=np.float32),
+    ).write_gguf(path)
+
+
+def _write_with_plain_text_delimiters(
+    well_formed_head: PointerHead, path: Path
+) -> None:
+    replace(well_formed_head, delimiters=PLAIN_TEXT_DELIMITERS).write_gguf(path)
+
+
+def _write_with_zero_temperature(well_formed_head: PointerHead, path: Path) -> None:
+    replace(well_formed_head, temperature=0.0).write_gguf(path)
+
+
+DEFECT_WRITERS: dict[PointerHeadDefect, Callable[[PointerHead, Path], None]] = {
+    PointerHeadDefect.F16_TENSORS: _write_with_f16_tensors,
+    PointerHeadDefect.MISMATCHED_PROJECTIONS: _write_with_mismatched_projections,
+    PointerHeadDefect.MISSING_DELIMITERS: _write_without_delimiters,
+    PointerHeadDefect.MISSING_INPUT_NORMALIZATION: _write_without_input_normalization,
+    PointerHeadDefect.MISSING_TEMPERATURE: _write_without_temperature,
+    PointerHeadDefect.OVERSIZED_HEAD_DIM: _write_with_oversized_head_dim,
+    PointerHeadDefect.PLAIN_TEXT_DELIMITERS: _write_with_plain_text_delimiters,
+    PointerHeadDefect.ZERO_TEMPERATURE: _write_with_zero_temperature,
+}
+
+
+def write_malformed_pointer_head(
+    defect: PointerHeadDefect,
+    hidden_size: int,
+    input_normalization: InputNormalization,
+    path: Path,
+) -> None:
+    DEFECT_WRITERS[defect](synthetic_pointer_head(hidden_size, input_normalization), path)
diff --git a/paddler_kev_converter_python/paddler_kev_converter/parity_reference.py b/paddler_kev_converter_python/paddler_kev_converter/parity_reference.py
new file mode 100644
index 000000000..39e3c577b
--- /dev/null
+++ b/paddler_kev_converter_python/paddler_kev_converter/parity_reference.py
@@ -0,0 +1,134 @@
+import json
+from pathlib import Path
+from typing import Any
+
+import torch
+from kev.api import SystemOneRequest, to_record
+from kev.checkpoint import Checkpoint, LoadOptions
+
+PARITY_REQUESTS: list[dict[str, Any]] = [
+    {
+        "state": (
+            "Shoes arrived two weeks late and in the wrong size. "
+            "Also I see two charges on my card."
+        ),
+        "questions": {
+            "department": {
+                "type": "choice",
+                "instructions": "Which team should handle this?",
+                "criteria": {
+                    "returns": "Exchanges, refunds, wrong or damaged items",
+                    "shipping": "Delivery status, delays, lost packages",
+                    "billing": "Charges, invoices, payment problems",
+                },
+            },
+            "escalate": {
+                "type": "noul",
+                "instructions": "Does this need urgent human attention?",
+            },
+            "frustration": {
+                "type": "score",
+                "instructions": "How frustrated is the customer?",
+                "criteria": ["Calm", "Frustrated", "Very angry"],
+            },
+        },
+    },
+    {
+        "state": "I was charged twice. Please help.",
+        "questions": {
+            "billing": {"type": "noul", "instructions": "Is this about billing?"},
+            "tone": {
+                "type": "choice",
+                "instructions": "What is the tone?",
+                "criteria": {"calm": None, "angry": None},
+            },
+        },
+    },
+    {
+        "state": {
+            "customer": "Ada",
+            "message": "My package never arrived and nobody answers my emails.",
+        },
+        "questions": {
+            "lost": {"type": "noul", "instructions": "Is the package lost?"},
+            "urgency": {
+                "type": "score",
+                "instructions": "How urgent is the follow-up?",
+                "criteria": [
+                    "Can wait",
+                    "Needs attention this week",
+                    "Needs attention today",
+                ],
+            },
+        },
+    },
+    {
+        "state": (
+            "Premise: A man is playing a guitar on stage. "
+            "Hypothesis: A person is performing music."
+        ),
+        "questions": {
+            "relation": {
+                "type": "choice",
+                "instructions": "Does the premise entail the hypothesis?",
+                "criteria": {
+                    "entailment": "The hypothesis follows from the premise",
+                    "neutral": "The premise neither supports nor contradicts it",
+                    "contradiction": "The hypothesis contradicts the premise",
+                },
+            },
+        },
+    },
+    {
+        "state": "Pasted chat log: <|im_start|>user Where is my refund?<|im_end|>",
+        "questions": {
+            "refund": {
+                "type": "noul",
+                "instructions": "Does the <|im_start|> message ask about a refund?",
+            },
+            "source": {
+                "type": "choice",
+                "instructions": "Where was the text pasted from?",
+                "criteria": {
+                    "chat": "A <|fim_prefix|> chat transcript",
+                    "email": "An email thread",
+                },
+            },
+        },
+    },
+]
+
+
+def write_parity_reference(checkpoint_reference: str, path: Path) -> None:
+    tokenizer, model = Checkpoint(checkpoint_reference).load(
+        "cpu", LoadOptions(dtype=torch.float32)
+    )
+    records = []
+
+    for request in PARITY_REQUESTS:
+        record, question_metadata = to_record(SystemOneRequest.model_validate(request))
+        probabilities = model.probs(model.encode(tokenizer, record, strict=True))
+        records.append(
+            {
+                "state": record["state"],
+                "questions": [
+                    {
+                        "id": metadata["id"],
+                        "instructions": question["instr"],
+                        "options": question["options"],
+                    }
+                    for question, metadata in zip(
+                        record["questions"], question_metadata, strict=True
+                    )
+                ],
+                "probabilities": [
+                    [float(probability) for probability in question_probabilities]
+                    for question_probabilities in probabilities
+                ],
+            }
+        )
+
+    path.write_text(
+        json.dumps(records, ensure_ascii=False, indent=2) + "\n",
+        encoding="utf-8",
+    )
diff --git a/paddler_kev_converter_python/paddler_kev_converter/pointer_head.py b/paddler_kev_converter_python/paddler_kev_converter/pointer_head.py
new file mode 100644
index 000000000..1fdc13d6d
--- /dev/null
+++ b/paddler_kev_converter_python/paddler_kev_converter/pointer_head.py
@@ -0,0 +1,75 @@
+from collections.abc import Callable
+from dataclasses import dataclass
+from pathlib import Path
+
+import numpy as np
+from gguf import GGUFWriter
+from numpy.typing import NDArray
+
+from paddler_kev_converter.input_normalization import InputNormalization
+
+type ProjectionArray = NDArray[np.float16] | NDArray[np.float32]
+
+POINTER_HEAD_ARCHITECTURE = "pointer_head"
+DELIMITER_NAMES = ["state", "question", "option_start", "option_end", "decide"]
+
+
+@dataclass(frozen=True)
+class PointerHead:
+    delimiters: list[str]
+    input_normalization: InputNormalization
+    key_bias: ProjectionArray
+    key_weight: ProjectionArray
+    query_bias: ProjectionArray
+    query_weight: ProjectionArray
+    temperature: float
+
+    def add_delimiters(self, writer: GGUFWriter) -> None:
+        for delimiter_name, delimiter_token in zip(
+            DELIMITER_NAMES, self.delimiters, strict=True
+        ):
+            writer.add_string(
+                f"{POINTER_HEAD_ARCHITECTURE}.delimiter.{delimiter_name}",
+                delimiter_token,
+            )
+
+    def add_input_normalization(self, writer: GGUFWriter) -> None:
+        writer.add_string(
+            f"{POINTER_HEAD_ARCHITECTURE}.input_normalization",
+            self.input_normalization.value,
+        )
+
+    def add_temperature(self, writer: GGUFWriter) -> None:
+        writer.add_float32(f"{POINTER_HEAD_ARCHITECTURE}.temperature", self.temperature)
+
+    def add_tensors(self, writer: GGUFWriter) -> None:
+        writer.add_tensor(f"{POINTER_HEAD_ARCHITECTURE}.key.bias", self.key_bias)
+        writer.add_tensor(f"{POINTER_HEAD_ARCHITECTURE}.key.weight", self.key_weight)
+        writer.add_tensor(f"{POINTER_HEAD_ARCHITECTURE}.query.bias", self.query_bias)
+        writer.add_tensor(
+            f"{POINTER_HEAD_ARCHITECTURE}.query.weight", self.query_weight
+        )
+
+    def write_gguf(self, path: Path) -> None:
+        self.write_gguf_sections(
+            path,
+            [
+                self.add_temperature,
+                self.add_delimiters,
+                self.add_input_normalization,
+                self.add_tensors,
+            ],
+        )
+
+    def write_gguf_sections(
+        self, path: Path, sections: list[Callable[[GGUFWriter], None]]
+    ) -> None:
+        writer = GGUFWriter(path, POINTER_HEAD_ARCHITECTURE)
+
+        for add_section in sections:
+            add_section(writer)
+
+        writer.write_header_to_file()
+        writer.write_kv_data_to_file()
+        writer.write_tensors_to_file()
+        writer.close()
diff --git a/paddler_kev_converter_python/paddler_kev_converter/synthetic_pointer_head.py b/paddler_kev_converter_python/paddler_kev_converter/synthetic_pointer_head.py
new file mode 100644
index 000000000..2205615aa
--- /dev/null
+++ b/paddler_kev_converter_python/paddler_kev_converter/synthetic_pointer_head.py
@@ -0,0 +1,37 @@
+from math import sqrt
+
+import numpy as np
+from kev.model import SPECIAL
+from numpy.typing import NDArray
+
+from paddler_kev_converter.input_normalization import InputNormalization
+from paddler_kev_converter.pointer_head import PointerHead
+
+SYNTHETIC_HEAD_DIM = 16
+SYNTHETIC_SEED = 20261006
+SYNTHETIC_TEMPERATURE = 1.5
+
+
+def synthetic_pointer_head(
+    hidden_size: int, input_normalization: InputNormalization
+) -> PointerHead:
+    generator = np.random.default_rng(SYNTHETIC_SEED)
+    weight_scale = 1 / sqrt(hidden_size)
+
+    def projection_weight() -> NDArray[np.float32]:
+        return (
+            generator.standard_normal((SYNTHETIC_HEAD_DIM, hidden_size)) * weight_scale
+        ).astype(np.float32)
+
+    def projection_bias() -> NDArray[np.float32]:
+        return generator.standard_normal(SYNTHETIC_HEAD_DIM).astype(np.float32)
+
+    return PointerHead(
+        delimiters=SPECIAL,
+        input_normalization=input_normalization,
+        key_bias=projection_bias(),
+        key_weight=projection_weight(),
+        query_bias=projection_bias(),
+        query_weight=projection_weight(),
+        temperature=SYNTHETIC_TEMPERATURE,
+    )
diff --git a/paddler_kev_converter_python/paddler_kev_converter/tokenizer_input_normalization.py b/paddler_kev_converter_python/paddler_kev_converter/tokenizer_input_normalization.py
new file mode 100644
index 000000000..d8768b9a8
--- /dev/null
+++ b/paddler_kev_converter_python/paddler_kev_converter/tokenizer_input_normalization.py
@@ -0,0 +1,33 @@
+from tokenizers.normalizers import NFC, NFD, NFKC, NFKD
+from transformers import PreTrainedTokenizerBase, PreTrainedTokenizerFast
+
+from paddler_kev_converter.error import (
+    TokenizerNormalizerUninspectableError,
+    TokenizerNormalizerUnsupportedError,
+)
+from paddler_kev_converter.input_normalization import InputNormalization
+
+
+def tokenizer_input_normalization(
+    tokenizer: PreTrainedTokenizerBase,
+) -> InputNormalization:
+    if not isinstance(tokenizer, PreTrainedTokenizerFast):
+        raise TokenizerNormalizerUninspectableError(
+            tokenizer.name_or_path, type(tokenizer).__name__
+        )
+
+    match tokenizer.backend_tokenizer.normalizer:
+        case None:
+            return InputNormalization.NONE
+        case NFC():
+            return InputNormalization.NFC
+        case NFD():
+            return InputNormalization.NFD
+        case NFKC():
+            return InputNormalization.NFKC
+        case NFKD():
+            return InputNormalization.NFKD
+        case unsupported_normalizer:
+            raise TokenizerNormalizerUnsupportedError(
+                tokenizer.name_or_path, type(unsupported_normalizer).__name__
+            )
diff --git a/paddler_kev_converter_python/paddler_kev_converter/tokenizer_reference.py b/paddler_kev_converter_python/paddler_kev_converter/tokenizer_reference.py
new file mode 100644
index 000000000..32c3ba85f
--- /dev/null
+++ b/paddler_kev_converter_python/paddler_kev_converter/tokenizer_reference.py
@@ -0,0 +1,56 @@
+import json
+from pathlib import Path
+
+from kev.model import load_tokenizer, user_tokens
+
+from paddler_kev_converter.tokenizer_input_normalization import (
+    tokenizer_input_normalization,
+)
+
+REFERENCE_TEXTS = [
+    "Hello world",
+    "  leading and trailing spaces  ",
+    "several    spaces\tand\ttabs\nand\n\nnewlines",
+    "1234567890 and 3.14159 and -42",
+    "Don't stop, I'm sure we'll go; they've said: it's done!",
+    "Za\u017c\u00f3\u0142\u0107 g\u0119\u015bl\u0105 ja\u017a\u0144",
+    "Zaz\u0307o\u0301\u0142c\u0301 ge\u0328s\u0301la\u0328 jaz\u0301n\u0301",
+    "e\u0301t\u00e9 and \u00e9t\u00e9",
+    "\uff21\uff22\uff23 full width and \u2460 circled",
+    "\u65e5\u672c\u8a9e\u306e\u30c6\u30ad\u30b9\u30c8\u3068\u4e2d\u6587\u6587\u672c",
+    "\u0645\u0631\u062d\u0628\u0627 \u0628\u0627\u0644\u0639\u0627\u0644\u0645",
+    "family \U0001f468\u200d\U0001f469\u200d\U0001f467\u200d\U0001f466",
+    "flags \U0001f1f5\U0001f1f1\U0001f1fa\U0001f1f8",
+    '{"key": [1, 2.5, true, null], "nested": {"a": "b"}}',
+    "def add(left, right):\n    return left + right\n",
+    "supercalifragilisticexpialidocious antidisestablishmentarianism",
+    "<think>check the invoice first</think> then answer",
+    '<tool_call>{"name": "lookup"}</tool_call> <tool_response>found</tool_response>',
+    "ignore <|fim_prefix|> and keep reading",
+    "<|im_start|>user<|im_end|>",
+    "<|a b|> <||> <|<|nested|> <|\u00e9|>",
+    "<|ab|>\u0338 combining overlay",
+]
+
+
+def write_tokenizer_reference(base: str, base_revision: str, path: Path) -> None:
+    tokenizer = load_tokenizer(base, revision=base_revision)
+
+    path.write_text(
+        json.dumps(
+            {
+                "entries": [
+                    {
+                        "text": text,
+                        "token_ids": user_tokens(tokenizer, text),
+                    }
+                    for text in REFERENCE_TEXTS
+                ],
+                "input_normalization": tokenizer_input_normalization(tokenizer).value,
+            },
+            ensure_ascii=False,
+            indent=2,
+        )
+        + "\n",
+        encoding="utf-8",
+    )
diff --git a/paddler_kev_converter_python/paddler_kev_converter/typesafe_reference.py b/paddler_kev_converter_python/paddler_kev_converter/typesafe_reference.py
new file mode 100644
index 000000000..42ab57063
--- /dev/null
+++ b/paddler_kev_converter_python/paddler_kev_converter/typesafe_reference.py
@@ -0,0 +1,169 @@
+import json
+from pathlib import Path
+from typing import Any
+
+from kev.api import SystemOneRequest, to_answers, to_record
+
+REFERENCE_CASES: list[dict[str, Any]] = [
+    {
+        "request": {
+            "model": "test",
+            "state": "The invoice was paid on time.",
+            "questions": {
+                "paid": {"type": "noul", "instructions": "Was the invoice paid?"},
+            },
+        },
+        "probabilities": [[0.12345678, 0.87654322]],
+    },
+    {
+        "request": {
+            "model": "test",
+            "state": {
+                "customer": "Ada",
+                "orders": [
+                    {"id": 1, "total": 19.99, "shipped": True},
+                    {"id": 2, "total": 100.0, "shipped": False, "notes": None},
+                ],
+                "tags": [],
+                "metadata": {},
+                "balance": -0.0,
+                "ratio": 1e-07,
+                "population": 1e16,
+                "huge": 1.5e300,
+                "count": 123456789012345678,
+            },
+            "questions": {
+                "sentiment": {
+                    "type": "choice",
+                    "instructions": {"task": "Classify", "hints": ["be brief", 2]},
+                    "criteria": {
+                        "positive": "The customer is happy",
+                        "neutral": None,
+                        "negative": "",
+                        "mixed": {"when": "both", "weight": 0.5},
+                    },
+                },
+                "risk": {
+                    "type": "score",
+                    "instructions": None,
+                    "criteria": [
+                        "no risk",
+                        {"level": "some", "detail": ["late", "partial"]},
+                        "high risk",
+                    ],
+                },
+                "vip": {
+                    "type": "noul",
+                    "instructions": 42,
+                    "criteria": {"true": "Spends a lot", "false": {"reason": "rare"}},
+                },
+            },
+        },
+        "probabilities": [
+            [0.25, 0.25, 0.4, 0.1],
+            [0.2, 0.3, 0.5],
+            [0.5, 0.5],
+        ],
+    },
+    {
+        "request": {
+            "model": "test",
+            "state": ["first line\nsecond line", 3.5, True, None, ["nested", 1]],
+            "questions": {
+                "tie": {
+                    "type": "choice",
+                    "instructions": ["pick", "one"],
+                    "criteria": {"left": 1, "right": False},
+                },
+                "single": {
+                    "type": "score",
+                    "instructions": 2.5,
+                    "criteria": ["only level"],
+                },
+            },
+        },
+        "probabilities": [[0.5, 0.5], [1.0]],
+    },
+    {
+        "request": {
+            "model": "test",
+            "state": (
+                "Za\u017c\u00f3\u0142\u0107 g\u0119\u015bl\u0105 ja\u017a\u0144 "
+                "\U0001f468\u200d\U0001f469"
+            ),
+            "questions": {
+                "lonely": {
+                    "type": "choice",
+                    "instructions": True,
+                    "criteria": {"only": "the only option"},
+                },
+                "uncertain": {
+                    "type": "score",
+                    "instructions": "How confident?",
+                    "criteria": ["low", "medium", "high", "certain"],
+                },
+            },
+        },
+        "probabilities": [[1.0], [0.25, 0.25, 0.25, 0.25]],
+    },
+    {
+        "request": {
+            "model": "test",
+            "state": {
+                "precise": 123.45678901234567,
+                "halfway": 1000000000000.6562,
+                "power_of_two": 7.120236347223045e-307,
+            },
+            "questions": {
+                "measure": {
+                    "type": "score",
+                    "instructions": 1000000000000.6562,
+                    "criteria": [123.45678901234567, 7.120236347223045e-307],
+                },
+                "pick": {
+                    "type": "choice",
+                    "instructions": "Which value is exact?",
+                    "criteria": {
+                        "halfway": 1000000000000.6562,
+                        "precise": 123.45678901234567,
+                    },
+                },
+            },
+        },
+        "probabilities": [[0.5, 0.5], [0.5, 0.5]],
+    },
+]
+
+
+def _rendered(record: dict[str, Any]) -> dict[str, Any]:
+    return {
+        "state": record["state"],
+        "questions": [
+            {"instructions": question["instr"], "options": question["options"]}
+            for question in record["questions"]
+        ],
+    }
+
+
+def write_typesafe_reference(path: Path) -> None:
+    references = []
+
+    for reference_case in REFERENCE_CASES:
+        record, question_metadata = to_record(
+            SystemOneRequest.model_validate(reference_case["request"])
+        )
+        references.append(
+            {
+                "request": reference_case["request"],
+                "rendered": _rendered(record),
+                "probabilities": reference_case["probabilities"],
+                "answers": to_answers(
+                    reference_case["probabilities"], question_metadata
+                ),
+            }
+        )
+
+    path.write_text(
+        json.dumps(references, ensure_ascii=False, indent=2) + "\n",
+        encoding="utf-8",
+    )
diff --git a/paddler_kev_converter_python/poetry.lock b/paddler_kev_converter_python/poetry.lock
new file mode 100644
index 000000000..5fd775bcd
--- /dev/null
+++ b/paddler_kev_converter_python/poetry.lock
@@ -0,0 +1,4138 @@
+# This file is automatically @generated by Poetry 2.4.3 and should not be changed by hand.
+
+[[package]]
+name = "accelerate"
+version = "1.15.0"
+description = "Accelerate"
+optional = false
+python-versions = ">=3.10.0"
+groups = ["main"]
+files = [
+    {file = "accelerate-1.15.0-py3-none-any.whl", hash = "sha256:97eacca0b73e45cb867dbf8c5d5d4dc32219544300e0c8992c7334dc2ef33cec"},
+    {file = "accelerate-1.15.0.tar.gz", hash = "sha256:5654f8c5eaa0d4fa68b33e287a97765da6849bf6d51dcac874e73fbbddfb6134"},
+]
+
+[package.dependencies]
+huggingface_hub = ">=0.21.0"
+numpy = ">=1.17"
+packaging = ">=20.0"
+psutil = "*"
+pyyaml = "*"
+safetensors = ">=0.4.3"
+torch = ">=2.0.0"
+
+[package.extras]
+deepspeed = ["deepspeed"]
+dev = ["bitsandbytes", "datasets", "diffusers", "evaluate", "parameterized", "peft", "pytest (>=7.2.0)", "pytest-order", "pytest-subtests", "pytest-xdist", "rich", "ruff (==0.13.1)", "scikit-learn", "scipy", "timm", "torchdata (>=0.8.0)", "torchpippy (>=0.2.0)", "tqdm", "transformers"]
+quality = ["ruff (==0.13.1)"]
+rich = ["rich"]
+sagemaker = ["sagemaker"]
+test-dev = ["bitsandbytes", "datasets", "diffusers", "evaluate", "peft", "scikit-learn", "scipy", "timm", "torchdata (>=0.8.0)", "torchpippy (>=0.2.0)", "tqdm", "transformers"]
+test-fp8 = ["torchao"]
+test-prod = ["parameterized", "pytest (>=7.2.0)", "pytest-order", "pytest-subtests", "pytest-xdist"]
+test-trackers = ["dvclive", "matplotlib", "swanlab[dashboard]", "tensorboard", "trackio", "wandb"]
+testing = ["bitsandbytes", "datasets", "diffusers", "evaluate", "parameterized", "peft", "pytest (>=7.2.0)", "pytest-order", "pytest-subtests", "pytest-xdist", "scikit-learn", "scipy", "timm", "torchdata (>=0.8.0)", "torchpippy (>=0.2.0)", "tqdm", "transformers"]
+
+[[package]]
+name = "aiohappyeyeballs"
+version = "2.7.1"
+description = "Happy Eyeballs for asyncio"
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "aiohappyeyeballs-2.7.1-py3-none-any.whl", hash = "sha256:9243213661e29250eb41368e5daa826fc017156c3b8a11440826b2e3ed376472"},
+    {file = "aiohappyeyeballs-2.7.1.tar.gz", hash = "sha256:065665c041c42a5938ed220bdcd7230f22527fbec085e1853d2402c8a3615d9d"},
+]
+
+[[package]]
+name = "aiohttp"
+version = "3.14.4"
+description = "Async http client/server framework (asyncio)"
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "aiohttp-3.14.4-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:dea3b3653c1d0d6babd8ce09cb9e5a7776864315d9cfb1c569f926389874a261"},
+    {file = "aiohttp-3.14.4-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:3099a389e5b5900fc79c6f613015ff3f8fa06a519744e6001cc68a31ba418cac"},
+    {file = "aiohttp-3.14.4-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:70247156fc76e92919f9aacb820556cb30289a54784299a31714d030bb604d84"},
+    {file = "aiohttp-3.14.4-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:fdf2d2a6ad402797422949f4a200a60d43e3a021695c8a7804aa1d5bbef22652"},
+    {file = "aiohttp-3.14.4-cp310-cp310-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:ba835fad8d4b5eaea0d34f9461cf499737a720a5f8bc6768d2f6d98a1d3ba8b5"},
+    {file = "aiohttp-3.14.4-cp310-cp310-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:0d4415ca5c9c8b24bc2f5733f870d4eab4929e680203559ba60131dd1a27852a"},
+    {file = "aiohttp-3.14.4-cp310-cp310-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:f6757e02c2c2b0d7985f57be87cf4a7c5b35708f849e1feaea3f77656ddb9dff"},
+    {file = "aiohttp-3.14.4-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:2a950bd8418ef54eccf5f9f517f08e37c1adab4bc50a470a0422ee1cc1032e29"},
+    {file = "aiohttp-3.14.4-cp310-cp310-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:34cf6f2fe4fc1d7183e8cc0d8c48bbdda828a740bc372d8278126149714d343c"},
+    {file = "aiohttp-3.14.4-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:bc0fbe4f79c3b156d754dd062718ffe404c3b31ca878d9d2b88ecdbc72ef8369"},
+    {file = "aiohttp-3.14.4-cp310-cp310-musllinux_1_2_armv7l.whl", hash = "sha256:d3d50a5a07a2b31a4369b86dd4a3acf3c2577c45091dbd5e884b198dec56480b"},
+    {file = "aiohttp-3.14.4-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:de15c46a63f46da881d219e44a04778c40114c756e5575b5ea38143b20eb3713"},
+    {file = "aiohttp-3.14.4-cp310-cp310-musllinux_1_2_riscv64.whl", hash = "sha256:caa4c0fc9b25e8c679baac68337e9c00deb6baf6601fe6f07e8511469cc1ceba"},
+    {file = "aiohttp-3.14.4-cp310-cp310-musllinux_1_2_s390x.whl", hash = "sha256:3647f8c090632c70abb1e80826e4ee24faea3f895f0283b60e91a0b17aea5e34"},
+    {file = "aiohttp-3.14.4-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:3f3e8030169843fb652a60a3d32c206e03f260198aa398985a48b41cd7e42f0a"},
+    {file = "aiohttp-3.14.4-cp310-cp310-win32.whl", hash = "sha256:a1205cbdddfca1428129bb6d23e0b292b09e681df951fa828145d83a63ffc0ca"},
+    {file = "aiohttp-3.14.4-cp310-cp310-win_amd64.whl", hash = "sha256:e90980e0454a043f81b435309a089ac5deab883779557eb58d004f33990f61e4"},
+    {file = "aiohttp-3.14.4-cp310-cp310-win_arm64.whl", hash = "sha256:e63eecf9a4b670a24055d5c0486128af16f83b02d9817abd685d3238940f8752"},
+    {file = "aiohttp-3.14.4-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:eb0093ca0817019d539f0d27bdf98e823944e05bb7efd9d2337a45b92e4d1b06"},
+    {file = "aiohttp-3.14.4-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:f482dc9c8309801b4259dd218609d3dae0a7bcdedeac56d90500eabaa0c23465"},
+    {file = "aiohttp-3.14.4-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:aa800bb7c1d00e166493931f2ddacb982db913292503dc2f4884db1f2bc5f105"},
+    {file = "aiohttp-3.14.4-cp311-cp311-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:d90e4dae71b26597f6d1fd12af7d092ae67551edf0e524b73e0687214b79ee33"},
+    {file = "aiohttp-3.14.4-cp311-cp311-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:5b681054b8c3b86aa6a0384b6ad7f5cb78dccbabab1f4a3a6793f98ad1378959"},
+    {file = "aiohttp-3.14.4-cp311-cp311-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:6dfe80b41bcf8d80d2125fb99a171adeb86db2fcb35c1a68e0226acba0ba7f03"},
+    {file = "aiohttp-3.14.4-cp311-cp311-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:71e4a59c6a8c5a6ae8b0b696c32635913e8f526cc01448762eaf5a3c59df8a4f"},
+    {file = "aiohttp-3.14.4-cp311-cp311-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:f887c3b7fb3058fc1ccd917f52078540c18d938d12018cb9edd64b6811d20c57"},
+    {file = "aiohttp-3.14.4-cp311-cp311-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:0da4525e9ba145a11617d2cd7e44fe1acaf471f9c72a0cad28e72879c1020fbb"},
+    {file = "aiohttp-3.14.4-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:f7e8d3c5caa44fdbee3eba1aa1417aa58c74c95d4089fe3c451911117fcbb2f3"},
+    {file = "aiohttp-3.14.4-cp311-cp311-musllinux_1_2_armv7l.whl", hash = "sha256:09092cf906c18c824b6b16881da5a89cc9cfa9b15344e496bcf880e9045bee33"},
+    {file = "aiohttp-3.14.4-cp311-cp311-musllinux_1_2_ppc64le.whl", hash = "sha256:dfd144837e92264878bba6d3a7ffa8206668b21e87748740afd45d193e48c87e"},
+    {file = "aiohttp-3.14.4-cp311-cp311-musllinux_1_2_riscv64.whl", hash = "sha256:9413cffe4e0d654b9b524f9c99dde8cbdb37e0ed9a2c4edd2e63adbac6b374e8"},
+    {file = "aiohttp-3.14.4-cp311-cp311-musllinux_1_2_s390x.whl", hash = "sha256:f50d719f97ba488dfb8e306cf8c8cc172ff0683ec0af53917b288229aa555ba5"},
+    {file = "aiohttp-3.14.4-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:68ed4824d96b7afec1f3a8d7fba190fe282718041c5772e20298003e1e3e5d66"},
+    {file = "aiohttp-3.14.4-cp311-cp311-win32.whl", hash = "sha256:d7a41d1427136828e6b11d3f3fd194d4357d5205a5b610be7ae07b1ba52ab974"},
+    {file = "aiohttp-3.14.4-cp311-cp311-win_amd64.whl", hash = "sha256:2efbdb87e79d596325c4eefaef0f4495d701145e882351310ebc538e854dc7a3"},
+    {file = "aiohttp-3.14.4-cp311-cp311-win_arm64.whl", hash = "sha256:ac6e6f90d9360e460c873f945f11dc8698db0a71fec6612823f1aaaab2c11faa"},
+    {file = "aiohttp-3.14.4-cp312-cp312-macosx_10_13_universal2.whl", hash = "sha256:55b0b335982b0db7117dff7f6a8e3aad43c4105f9e392ca04f9fc8d958a95cce"},
+    {file = "aiohttp-3.14.4-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:2d0d8e435582b0d53009dac99e26a3d2be932ed7c9dbfcfc70f3fce1cf943bd5"},
+    {file = "aiohttp-3.14.4-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:bf734016932d68e1324cfb638cac7a0866a83933a33badf15e836a7e24efb8c0"},
+    {file = "aiohttp-3.14.4-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:322e3d741b3133bcbb981a41d4f9b8ac3545c18de49c662c2b1f4e1c770b98ad"},
+    {file = "aiohttp-3.14.4-cp312-cp312-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:15a3bb5f90a4e515071f3617a45dfe26560beb834a6b6e108cdf4bc9e9719ec3"},
+    {file = "aiohttp-3.14.4-cp312-cp312-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:6c044e52df1a466af82cad3518c92ea80cc4ddcfea86966ff083bb34bab29bf4"},
+    {file = "aiohttp-3.14.4-cp312-cp312-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:d613d7d51bf06fe9a5e3aab86ce06d835bb0ceb1a31ea9e8766df4a403ae62a6"},
+    {file = "aiohttp-3.14.4-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:c5e93ab6e6330dce40f5b43996f1fc5f2543296d263a5e4b59ff09e73870b7d8"},
+    {file = "aiohttp-3.14.4-cp312-cp312-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:20134dc8e68c67e7478124d516449428684c0545b2d3faa575691c8bcbd56678"},
+    {file = "aiohttp-3.14.4-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:f0daef5ef012369dddc391c61f604111424aa9a96313a69fb899018b09d1b01e"},
+    {file = "aiohttp-3.14.4-cp312-cp312-musllinux_1_2_armv7l.whl", hash = "sha256:19290a108de0718b73bc69b86e788975c6a59855c671ebe2a98a07a6f968328f"},
+    {file = "aiohttp-3.14.4-cp312-cp312-musllinux_1_2_ppc64le.whl", hash = "sha256:1494877625baeffab669f4a25a09ec4eea721b49e35363081a2a8e4231fe141f"},
+    {file = "aiohttp-3.14.4-cp312-cp312-musllinux_1_2_riscv64.whl", hash = "sha256:a779ce4bf099ce4ad2cdfed62f1ca536d435f53237c54ff40172fddada24d1b5"},
+    {file = "aiohttp-3.14.4-cp312-cp312-musllinux_1_2_s390x.whl", hash = "sha256:0b4cec9fca876e2d4f4c6cb32b177390132d04618867689f43fefabf3f5c3980"},
+    {file = "aiohttp-3.14.4-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:217707313e23aedde9bea67cf8365c7b138b549ae179121ef3f476cf67e3173a"},
+    {file = "aiohttp-3.14.4-cp312-cp312-win32.whl", hash = "sha256:18a9fb9a3f6e6e63f4d27f9adeef313a2d3c69a84d31170a844623824e722534"},
+    {file = "aiohttp-3.14.4-cp312-cp312-win_amd64.whl", hash = "sha256:69dcb02d33dfe415d5ee342cc63b7d320efcaf4b761701c7bd7bfb91c715481a"},
+    {file = "aiohttp-3.14.4-cp312-cp312-win_arm64.whl", hash = "sha256:d36b0263e7c2fbf1750b9f35d4e7e48b0442d8c45e4a89e3e74168821fb014bb"},
+    {file = "aiohttp-3.14.4-cp313-cp313-android_24_arm64_v8a.whl", hash = "sha256:264b5a7568306590c44ae4a0237ad1715e35d447ea44ebf324e4c8fd9f78f67a"},
+    {file = "aiohttp-3.14.4-cp313-cp313-android_24_x86_64.whl", hash = "sha256:17e5d7c7775d8dd8e894c7ef6ebfa26509de36868a9a6c7c4fcfaf6a1bf43d60"},
+    {file = "aiohttp-3.14.4-cp313-cp313-ios_13_0_arm64_iphoneos.whl", hash = "sha256:e91dc8fe9cbd052d16f7f1267e285cf8bbac4d2194bd5741f05ef201c77ed09e"},
+    {file = "aiohttp-3.14.4-cp313-cp313-ios_13_0_arm64_iphonesimulator.whl", hash = "sha256:8df6157d31703972aec9e3b93963f039ff06318301e30b007c3e941ff8a7ac42"},
+    {file = "aiohttp-3.14.4-cp313-cp313-ios_13_0_x86_64_iphonesimulator.whl", hash = "sha256:220912b351549dc736c59d104dc443fa615753102bb778a4e98f86e4d94bff6a"},
+    {file = "aiohttp-3.14.4-cp313-cp313-macosx_10_13_universal2.whl", hash = "sha256:2efa42b39bb3f524d5eca3760639b584afdd1143a6b047fda29910ffb75e4993"},
+    {file = "aiohttp-3.14.4-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:7e7356059c8262d2c1fd95242e2b7b70fbc6eee3135bf470049cbec3e968b5b0"},
+    {file = "aiohttp-3.14.4-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:fac247d0cc956d732df1211ac2d8a9999ec8dc1a92a90401de9d52e4ead546f7"},
+    {file = "aiohttp-3.14.4-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:6e2eb0f8b03b4f2bd154ed8cf7c1a119a1ad04808c258bf58f90138c83b8c6d5"},
+    {file = "aiohttp-3.14.4-cp313-cp313-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:7fbe827d27e0e369fd7bb88ceeaa5b46ea006e73e015b9dd77e3723a53162293"},
+    {file = "aiohttp-3.14.4-cp313-cp313-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:a70387146d03af9f047c629b71aa27d9594b3b31ec4c9c91e1a73d0d8bf38e0e"},
+    {file = "aiohttp-3.14.4-cp313-cp313-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:0d0934926fc65744e2fdd44ce68b2d79d5ee608a1e23f0596b35dd696519bb7c"},
+    {file = "aiohttp-3.14.4-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:ff4b366510c733974adb7a4102c094e4a6308e6f369315a4cf14d2d0a619cf82"},
+    {file = "aiohttp-3.14.4-cp313-cp313-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:85fac7c99e0ac3dcbd6ae1c33774e0603bfc732ddb967a6fce4a731103c1888e"},
+    {file = "aiohttp-3.14.4-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:ef1e9c3b4a2300023a5dc865889ce532cfaf6798ba4ef78fbbefe738a9d23efa"},
+    {file = "aiohttp-3.14.4-cp313-cp313-musllinux_1_2_armv7l.whl", hash = "sha256:97e65916588be0f952b3307aaa8460c0771fa98dbea2dd27dabc9f5d84d53c3b"},
+    {file = "aiohttp-3.14.4-cp313-cp313-musllinux_1_2_ppc64le.whl", hash = "sha256:840775cb39a9424f9edf142133a1192263ac5bd79117ad2a16602cef6a0c1e9d"},
+    {file = "aiohttp-3.14.4-cp313-cp313-musllinux_1_2_riscv64.whl", hash = "sha256:54f9256430ca040d58affbd551513de13a5cff7d065204fc52b9a1b0af09a049"},
+    {file = "aiohttp-3.14.4-cp313-cp313-musllinux_1_2_s390x.whl", hash = "sha256:6112d5931bf12c776188cab81b82bbb17d86e2e1381f2ade6edc43e6b240f0cb"},
+    {file = "aiohttp-3.14.4-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:facc6df39047c4732cbd885bb62e40df20885fcd709fc399e11f81b45ce0fff9"},
+    {file = "aiohttp-3.14.4-cp313-cp313-win32.whl", hash = "sha256:7968634fa3a967a2bc0b9aeba1d5a1955501dcd7d0de44b96f7ebfb662797735"},
+    {file = "aiohttp-3.14.4-cp313-cp313-win_amd64.whl", hash = "sha256:c0a894b0265d139cd4c2c8bd4643822cc289a4a5331e21e810881bd000ffb2dc"},
+    {file = "aiohttp-3.14.4-cp313-cp313-win_arm64.whl", hash = "sha256:da16c3037178e72589d91de59536d271dffc92c2be2b47c38a45538cba54fd29"},
+    {file = "aiohttp-3.14.4-cp314-cp314-android_24_arm64_v8a.whl", hash = "sha256:749bfc43bb1df71bbe6ed215044d5412d786698fc5771d2037436ca49622f3c9"},
+    {file = "aiohttp-3.14.4-cp314-cp314-android_24_x86_64.whl", hash = "sha256:c113e4b489d711f606ca278cd8dc89179746244ea841aa0638b1b134600afad0"},
+    {file = "aiohttp-3.14.4-cp314-cp314-ios_13_0_arm64_iphoneos.whl", hash = "sha256:50969522bc7022026a2c7968e67da5c789f6b716b988edac55dc630390efd5df"},
+    {file = "aiohttp-3.14.4-cp314-cp314-ios_13_0_arm64_iphonesimulator.whl", hash = "sha256:cbf92cd7a1c800cee244558dd5630abb45c1c1093ef7b0d546a99f526ff7a9eb"},
+    {file = "aiohttp-3.14.4-cp314-cp314-ios_13_0_x86_64_iphonesimulator.whl", hash = "sha256:fc76be08ff407fba717a769aff04ca1b4bf9ecfdcb5caadef84a563fdbbd146d"},
+    {file = "aiohttp-3.14.4-cp314-cp314-macosx_10_15_universal2.whl", hash = "sha256:9b416f995b68a921069e6ec61fa07f2ac3ff033d92e97a69f673af89f5bb954d"},
+    {file = "aiohttp-3.14.4-cp314-cp314-macosx_10_15_x86_64.whl", hash = "sha256:2c9585101cfca10e74a07cc0ef2bdc6d90616e6750926b20456f429d0608f740"},
+    {file = "aiohttp-3.14.4-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:3c4dcf611b095e7983421f49d01ffbed7ed574891533d606ea115fc23e7c4e90"},
+    {file = "aiohttp-3.14.4-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:f8f61a9b5fa56c58e3441f9dc64137f786232d40fb89127190d0e0f84e1e5865"},
+    {file = "aiohttp-3.14.4-cp314-cp314-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:d43e538dfc85b705fd832620d050970ef5613b3869a7169b44a554b9dab9eee9"},
+    {file = "aiohttp-3.14.4-cp314-cp314-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:4135c4bb4b7682a7633fadc7170ea969d843f83aa7224e70111a20675105c985"},
+    {file = "aiohttp-3.14.4-cp314-cp314-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:9055324f156a94fdf6dc9a825ae60f9a2947b88f591ea81ee80826633de62ae7"},
+    {file = "aiohttp-3.14.4-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:fb22c8e4b60385a9b414c70c6e8b2d0f4b23ff7fc4bb572ceb58297b228169f1"},
+    {file = "aiohttp-3.14.4-cp314-cp314-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:3b6945559a260112742ed58cde2165488d2fc51aeea34afbaa99d9371b43e1b5"},
+    {file = "aiohttp-3.14.4-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:8f4078a90b174f15efb2c8725e8720c40dbd3b2cf6350dbc085c964a938d9de7"},
+    {file = "aiohttp-3.14.4-cp314-cp314-musllinux_1_2_armv7l.whl", hash = "sha256:f433c5654e32f72de3ef074d715eb5c7604eb6d28a5c2f3c9e5056f20fe6b306"},
+    {file = "aiohttp-3.14.4-cp314-cp314-musllinux_1_2_ppc64le.whl", hash = "sha256:871035fe79c636b6bc126ed78df37d72b830dfc751612cb4b1040360edf2719f"},
+    {file = "aiohttp-3.14.4-cp314-cp314-musllinux_1_2_riscv64.whl", hash = "sha256:2617132524f66338f6c1d218cd16081576aefc40df4f2bcba338a99fbf67fec1"},
+    {file = "aiohttp-3.14.4-cp314-cp314-musllinux_1_2_s390x.whl", hash = "sha256:d8bc326133272deee8ef337581550450e4bb735c0520be893b950a1f5baaf41a"},
+    {file = "aiohttp-3.14.4-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:3ec57223e261b7fa9979c12dbce009ca53439be76cac3af27f860d3e4b1d6f7a"},
+    {file = "aiohttp-3.14.4-cp314-cp314-win32.whl", hash = "sha256:7242043e71fc449a19a47d92f4e9fd1e0ae8d3f88488982381f422bf6d651218"},
+    {file = "aiohttp-3.14.4-cp314-cp314-win_amd64.whl", hash = "sha256:019587c7b3a44917dd61bcb312fd5c034f341dc9851b83e7b41b65e08b68ca0f"},
+    {file = "aiohttp-3.14.4-cp314-cp314-win_arm64.whl", hash = "sha256:60dc71d38db0988f37e78f6ad4a010a2f6355cfac3fff5f130cf5ef2231f3f0d"},
+    {file = "aiohttp-3.14.4-cp314-cp314t-macosx_10_15_universal2.whl", hash = "sha256:674f53ab9d2794e2dd767619a3384bf1006f8d075e43c19d449a645f8ff641c2"},
+    {file = "aiohttp-3.14.4-cp314-cp314t-macosx_10_15_x86_64.whl", hash = "sha256:6d243221f73e5ac546c2ed9255100423e4bc092b08b944f23498e8cf3096c26c"},
+    {file = "aiohttp-3.14.4-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:0daa016eb74e7888ed5f9ceec6eea3158999c1c4a115ef5cbe88baa8d924f94c"},
+    {file = "aiohttp-3.14.4-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:70c031a622aa20add05ebd234b90adaf6bf51c1c88cc305ec760721af83308c3"},
+    {file = "aiohttp-3.14.4-cp314-cp314t-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:7df363fa7c38945a896e12c69af4a1a56cbd4152791d9e41fd0bac396b72de38"},
+    {file = "aiohttp-3.14.4-cp314-cp314t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:02a508ba5ad91c4d730523584c0adc65882274bba5e37c33bc2f475e37577256"},
+    {file = "aiohttp-3.14.4-cp314-cp314t-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:0cb3edece6fe6a944eaf97d803b88f606da6bc1600b1894c2969e78ffdc1fc29"},
+    {file = "aiohttp-3.14.4-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:21ce1223b461bcc5fb389a7204001578f1c0b85c2da60a84d915ebb021294245"},
+    {file = "aiohttp-3.14.4-cp314-cp314t-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:35d5432f0dc1c8b71307e9112461b52f336da225e344283eec4779222c0a1dff"},
+    {file = "aiohttp-3.14.4-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:d6ec364493057c118d08292f7cd60865965c209167d8657b45643a26aef77495"},
+    {file = "aiohttp-3.14.4-cp314-cp314t-musllinux_1_2_armv7l.whl", hash = "sha256:70492983b7aed9e61bda82fe1ce80bcba9199610886443133783a288638923c0"},
+    {file = "aiohttp-3.14.4-cp314-cp314t-musllinux_1_2_ppc64le.whl", hash = "sha256:a77776673331577147e9d25cd7833a72e8891b47556336cff06ba1569e488b99"},
+    {file = "aiohttp-3.14.4-cp314-cp314t-musllinux_1_2_riscv64.whl", hash = "sha256:73799000dd6fd5247094fd8e6d3e1eb871e41aab5cc35dff9d9f654cd230bc24"},
+    {file = "aiohttp-3.14.4-cp314-cp314t-musllinux_1_2_s390x.whl", hash = "sha256:27b370fb532e707e8c96d04fa0c96301fb7dba2d50aaca57f6c456cdfffec5bd"},
+    {file = "aiohttp-3.14.4-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:884e0d1c707b6217fee3080b9b4029eee7551af788ee761010df19be158414b7"},
+    {file = "aiohttp-3.14.4-cp314-cp314t-win32.whl", hash = "sha256:f2e3d34edb151d3e27f93a8338e2da33e2c7261cd4c879ce5f5fe280cc190ea5"},
+    {file = "aiohttp-3.14.4-cp314-cp314t-win_amd64.whl", hash = "sha256:88cf889e51092537fe44adfbff96c2a22d6479d6f92b97b3d413fbfd0a4decf4"},
+    {file = "aiohttp-3.14.4-cp314-cp314t-win_arm64.whl", hash = "sha256:ca450c6d42fda7c0bfaf0e200f32f77c5f2c960a33be1e14d197c84588b7f5f6"},
+    {file = "aiohttp-3.14.4-cp315-cp315-android_24_arm64_v8a.whl", hash = "sha256:3b10799df72f66c30935bc661cb89e5ac2918debfc575e350e226fbaa846cbb7"},
+    {file = "aiohttp-3.14.4-cp315-cp315-android_24_x86_64.whl", hash = "sha256:7ffd7f9dd39caccc6b2f68ec0339435ecad2b20f6ea2e2b3de91bbe464ac7137"},
+    {file = "aiohttp-3.14.4-cp315-cp315-ios_13_0_arm64_iphoneos.whl", hash = "sha256:a0a7a227353f7693f73878ad14cd2623cafadd656ababcd20e2adfb7225172ea"},
+    {file = "aiohttp-3.14.4-cp315-cp315-ios_13_0_arm64_iphonesimulator.whl", hash = "sha256:1c43b6af8e4708e8b3812a9ee4f79895207e12d555b7f6d70deb838e0011e12d"},
+    {file = "aiohttp-3.14.4-cp315-cp315-ios_13_0_x86_64_iphonesimulator.whl", hash = "sha256:97416b2b997e5e9b817e322b2f0a89b76c54cde22471f51ac17ac15e10901ba0"},
+    {file = "aiohttp-3.14.4-cp315-cp315-macosx_10_15_universal2.whl", hash = "sha256:ee2a5aa31e507c17841d592364f2148393b17f83bc2766e9211e1eb31ea72a70"},
+    {file = "aiohttp-3.14.4-cp315-cp315-macosx_10_15_x86_64.whl", hash = "sha256:5d5d9843c34866e986dd24eef713452609e304550c471320ce7b54b3cfb1391e"},
+    {file = "aiohttp-3.14.4-cp315-cp315-macosx_11_0_arm64.whl", hash = "sha256:efcbf93e31b52c21665b6ebc489eb6fdb0dfd86b592edc677b5cffdfa7b11677"},
+    {file = "aiohttp-3.14.4-cp315-cp315-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:2f1a79977794592dcf6485ed964ec2448db527a1bc0c72ec89663658dad3a739"},
+    {file = "aiohttp-3.14.4-cp315-cp315-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:515e7ac6d27509dc3ba2c35fb88970b00b9e7b07ee7cc947c8733dcd54ef5def"},
+    {file = "aiohttp-3.14.4-cp315-cp315-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:185154ffa3b54765b5de96f464f41cfa476f8815a03854ae5e0fe4d48fb06e9c"},
+    {file = "aiohttp-3.14.4-cp315-cp315-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:6aece59f77959628a22cd2ffe48b102cda9611d680e090f204385b3998a1190f"},
+    {file = "aiohttp-3.14.4-cp315-cp315-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:d7205aa6589f1ad1306c6994ff2e1ab5efabacf938f9ffd7682fb225c197b06a"},
+    {file = "aiohttp-3.14.4-cp315-cp315-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:f38c7b3ce60d55ac0f2a9fd2edc194789d043fefea15cfb7dcb40b7d5f561146"},
+    {file = "aiohttp-3.14.4-cp315-cp315-musllinux_1_2_aarch64.whl", hash = "sha256:b0a1274ee4ee6203c15a341cce87f9f3df8151d5496047b3608bb62bd969cffc"},
+    {file = "aiohttp-3.14.4-cp315-cp315-musllinux_1_2_armv7l.whl", hash = "sha256:69ce2407951516f413adf3132b88451eababd710ce631f83c81a9603a3640253"},
+    {file = "aiohttp-3.14.4-cp315-cp315-musllinux_1_2_ppc64le.whl", hash = "sha256:b84d3d057c60eb8e0c2b457a1203e868d82709d3f0c272a4975dfd88c2414939"},
+    {file = "aiohttp-3.14.4-cp315-cp315-musllinux_1_2_riscv64.whl", hash = "sha256:4ecc84f69c9e217984f6877e832e4786a74567432c0bf2b7512919fbdfe18ade"},
+    {file = "aiohttp-3.14.4-cp315-cp315-musllinux_1_2_s390x.whl", hash = "sha256:878f2d7950238b2b0a8cbb88054444a125f0cfbc77f54c811aca0483132a9be3"},
+    {file = "aiohttp-3.14.4-cp315-cp315-musllinux_1_2_x86_64.whl", hash = "sha256:6ff9d3a0e6935f1cb0a6685f9f4b3f0d3e406456d7a434a03cf309b8cd5d7a2e"},
+    {file = "aiohttp-3.14.4-cp315-cp315-win32.whl", hash = "sha256:e07dfd7cd360f20be26dc2487ad4c7f21b0b39fdd593e9ebf37d2424ad52defb"},
+    {file = "aiohttp-3.14.4-cp315-cp315-win_amd64.whl", hash = "sha256:921fc4f1ad549091bfc39cb5e93925b04f307f83517756a3111843851130efce"},
+    {file = "aiohttp-3.14.4-cp315-cp315-win_arm64.whl", hash = "sha256:5be8af106f96fd625f6c264fa9faaaa9e0104089684d9015d2fa5b6eba6407c7"},
+    {file = "aiohttp-3.14.4-cp315-cp315t-macosx_10_15_universal2.whl", hash = "sha256:e69727ec04a4608200ff32ec29478299df7657e2eb78c416ae338d0ca6662596"},
+    {file = "aiohttp-3.14.4-cp315-cp315t-macosx_10_15_x86_64.whl", hash = "sha256:f8b658dd3ef2ebd7708b311ed96ff7cb5edaca001a30018a091504ef492bcb1c"},
+    {file = "aiohttp-3.14.4-cp315-cp315t-macosx_11_0_arm64.whl", hash = "sha256:3749e018123df1205161a73ba96d336e17e9cf6b947295e53a43571de420e7b1"},
+    {file = "aiohttp-3.14.4-cp315-cp315t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:8f28a3c4fca436b2e594090e7cd22558f5e93a30ca18516c436de982da6fc25a"},
+    {file = "aiohttp-3.14.4-cp315-cp315t-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:e099aae21246992a1d84303ff262262d054c474e46cace33b0d5a69810c71877"},
+    {file = "aiohttp-3.14.4-cp315-cp315t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:fafb521891e646d9cf5da89f01c75e7159d08be7254a024149630a698be28251"},
+    {file = "aiohttp-3.14.4-cp315-cp315t-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:1d385db2cd154bce0b3dd451cf2a52bec91ebb7be2c3a5d06b14b4391631d807"},
+    {file = "aiohttp-3.14.4-cp315-cp315t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:ae8a244889d8a53549851bf595b9f382f64acca57e8762da1ecb26ba786828f7"},
+    {file = "aiohttp-3.14.4-cp315-cp315t-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:681dde68ff8d8d5e7d5abcad4feb45bfb457a218310c919825e21fc36cac3d0b"},
+    {file = "aiohttp-3.14.4-cp315-cp315t-musllinux_1_2_aarch64.whl", hash = "sha256:b91a6441fcaed88ca7bca67726b189332df3ceb6ac0dfc98dcb9654bb723288c"},
+    {file = "aiohttp-3.14.4-cp315-cp315t-musllinux_1_2_armv7l.whl", hash = "sha256:f15dbbac49bc15eaff95c0b8166525c4050996af578c7be141e145f40ee2cf24"},
+    {file = "aiohttp-3.14.4-cp315-cp315t-musllinux_1_2_ppc64le.whl", hash = "sha256:7abae043e87692d1c2adda1fab501bb0612d28c4a23eb082005c1e684f792dac"},
+    {file = "aiohttp-3.14.4-cp315-cp315t-musllinux_1_2_riscv64.whl", hash = "sha256:21f4624fb1051fc1aa10d57a636a80e829f7d0417462d6f1ce380b08fe1a904c"},
+    {file = "aiohttp-3.14.4-cp315-cp315t-musllinux_1_2_s390x.whl", hash = "sha256:2e8dc3a0fdaf9d12b954d3d06cc4c355d91b17fb6c671715b593624ca4c379f8"},
+    {file = "aiohttp-3.14.4-cp315-cp315t-musllinux_1_2_x86_64.whl", hash = "sha256:1410b6f6e3d52ea7c02c950496522ce6e85d5543bf154bc80a6141e505251cfa"},
+    {file = "aiohttp-3.14.4-cp315-cp315t-win32.whl", hash = "sha256:a65189a89f3e621f7c27a968b5131bbca29126e2b8517dc72b205747a39d4c96"},
+    {file = "aiohttp-3.14.4-cp315-cp315t-win_amd64.whl", hash = "sha256:2b2c95f5f769eec92b552969db1f79ea0282058ce5131ed0f2fa540b4af3f14e"},
+    {file = "aiohttp-3.14.4-cp315-cp315t-win_arm64.whl", hash = "sha256:54209fff79346ee1ef0d5cbf92fa80a5bb37396406bb4d932e5107f9c29b2d3f"},
+    {file = "aiohttp-3.14.4-py3-none-any.whl", hash = "sha256:5c6758ba62aea282c537179cfc8474a90f2b5f7b089cc5ff66d8920d86a9bfdd"},
+    {file = "aiohttp-3.14.4.tar.gz", hash = "sha256:831fc5bd39ec2517851e348f613ddb5447a47cf4b71cb09845af7ad7ed45d8f9"},
+]
+
+[package.dependencies]
+aiohappyeyeballs = ">=2.5.0"
+aiosignal = ">=1.4.0"
+attrs = ">=17.3.0"
+frozenlist = ">=1.1.1"
+multidict = ">=4.5,<7.0"
+propcache = ">=0.2.0"
+typing_extensions = {version = ">=4.4", markers = "python_version < \"3.13\""}
+yarl = ">=1.25.1,<2.0"
+
+[package.extras]
+speedups = ["Brotli (>=1.2) ; platform_python_implementation == \"CPython\" and sys_platform != \"android\" and sys_platform != \"ios\"", "aiodns (>=3.3.0) ; sys_platform != \"android\" and sys_platform != \"ios\"", "backports.zstd ; platform_python_implementation == \"CPython\" and python_version < \"3.14\" and sys_platform != \"android\" and sys_platform != \"ios\"", "brotlicffi (>=1.2) ; platform_python_implementation != \"CPython\""]
+
+[[package]]
+name = "aiosignal"
+version = "1.4.0"
+description = "aiosignal: a list of registered asynchronous callbacks"
+optional = false
+python-versions = ">=3.9"
+groups = ["main"]
+files = [
+    {file = "aiosignal-1.4.0-py3-none-any.whl", hash = "sha256:053243f8b92b990551949e63930a839ff0cf0b0ebbe0597b0f3fb19e1a0fe82e"},
+    {file = "aiosignal-1.4.0.tar.gz", hash = "sha256:f47eecd9468083c2029cc99945502cb7708b082c232f9aca65da147157b251c7"},
+]
+
+[package.dependencies]
+frozenlist = ">=1.1.0"
+typing-extensions = {version = ">=4.2", markers = "python_version < \"3.13\""}
+
+[[package]]
+name = "annotated-doc"
+version = "0.0.5"
+description = "Document parameters, class attributes, return types, and variables inline, with Annotated."
+optional = false
+python-versions = ">=3.9"
+groups = ["main"]
+files = [
+    {file = "annotated_doc-0.0.5-py3-none-any.whl", hash = "sha256:117bac03a25ede5df5440e855b32d556049ca169ead221505badf432fed4b101"},
+    {file = "annotated_doc-0.0.5.tar.gz", hash = "sha256:c7e58ce09192557605d8bbd92836d7e1d520ac9580096042c0bfd197efacf1bb"},
+]
+
+[[package]]
+name = "annotated-types"
+version = "0.8.0"
+description = "Reusable constraint types to use with typing.Annotated"
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "annotated_types-0.8.0-py3-none-any.whl", hash = "sha256:f072f4d804ea359e4eaf198b1af7a8b0943881a87f31bb764f8bf219bb9419e0"},
+    {file = "annotated_types-0.8.0.tar.gz", hash = "sha256:13b2beaad985e05e2d6407ee4c4f35590b11f8d693a258a561055cac8f64cab7"},
+]
+
+[[package]]
+name = "anyio"
+version = "4.15.1"
+description = "High-level concurrency and networking framework on top of asyncio or Trio"
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+markers = "sys_platform != \"emscripten\""
+files = [
+    {file = "anyio-4.15.1-py3-none-any.whl", hash = "sha256:6152fdbbf9a77fdec97731721bebf7c4c44f7c29b424b0065826173efc7ed101"},
+    {file = "anyio-4.15.1.tar.gz", hash = "sha256:9f28306018cbd6d329e64a36d58256edff76dd996fe423bc957326e578b82a94"},
+]
+
+[package.dependencies]
+idna = ">=2.8"
+typing_extensions = {version = ">=4.16.0", markers = "python_version < \"3.15\""}
+
+[package.extras]
+trio = ["trio (>=0.32.0)"]
+
+[[package]]
+name = "ast-serialize"
+version = "0.13.1"
+description = "Python bindings for mypy AST serialization"
+optional = false
+python-versions = ">=3.7"
+groups = ["dev"]
+files = [
+    {file = "ast_serialize-0.13.1-cp314-cp314-pyemscripten_2026_0_wasm32.whl", hash = "sha256:3e129d44b0cffbea24a7f40117ce49e30993c3d119e584d4817ebd6f87246456"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-macosx_10_12_x86_64.whl", hash = "sha256:523d45bab16ddaf1399c04c4fe81bbf180f81e33c2f7bcbe76be56b0f7311ab4"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:638333fb8ae72c62753f9e794fda20290883ddc65b0d85e26e08f8bb227567ad"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:cc34c670747d85d4232834665e247009413eaebe47c69eaf60a5c439aaecaed9"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:8c8ebd419811c435812d242fb2a173e57dfede198c9ac9a0a584666ebc9b9331"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:8899c038b1c15ae8b98e379057fb3ef398ceb0b21d1737a8156dd5a0923b7e51"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:e14c4de98f965647a74e6eaf37e20207124e45451ccd36e84a6f6ebab45e0372"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:7f3e8a3eff69af640db985360fef35073fb46956dfc3e5c614cf00920d2746dd"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-manylinux_2_31_riscv64.whl", hash = "sha256:75d7750fe98f6f35a96c43aba8d98a2e623e22c778a0c17ab9ede92eb10f5d82"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:43e8ddd63e80fc280c2f790157386f27f9b9f6ccb9c4f8e803b863cbfb085769"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:9221e09496fb942901993c1812ee87950cc0ecbb514c7c0f1f5fa6d78f701ed8"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-musllinux_1_2_armv7l.whl", hash = "sha256:7a5c38bdb93669c596f94fb7204dd6ebe4fadb19058ac9a2faf2948c2a175f96"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-musllinux_1_2_i686.whl", hash = "sha256:acf4e9f055d75058ab01536da330648746f2f82b7697a85230547f2d5ebe895c"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-musllinux_1_2_ppc64le.whl", hash = "sha256:87427913c325adc811e81e7c9a757b061f47ef1e29114ac479aa0953a2687a98"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-musllinux_1_2_riscv64.whl", hash = "sha256:ad88921f95b92bc144d9ad059a0e16f8c1ae69f5239f2d3c6536fd2774f73dfe"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:f8b17e2921e56621bcb46792bd3e0b82ca3b85de78d5936e7b8174ba1758b0aa"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-win32.whl", hash = "sha256:2f9a9ce8b0994f757b941f7f7fe384a6666b4162b941ebc5c248081baea18e47"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-win_amd64.whl", hash = "sha256:5d99910e6b9a9b9360bcca2ae99aee36cc912ce058df2a57d1c33099b6cb2593"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-win_arm64.whl", hash = "sha256:a789cbe14d0a0e9ec9b32a456e5abc81ad8799479c3cfd88a29e3442262aebe6"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-macosx_10_12_x86_64.whl", hash = "sha256:23d768369b89d551077dc173da6892fad00443338417523b6732f4e8ba4a7ed5"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-macosx_11_0_arm64.whl", hash = "sha256:68f8829d114dab6c9db4a8bbb8d0e342958faee1b5e15a29022678df639960f8"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:4e385b17d5bde892a6184003a9aff3518387908b73f8f5ea539313c07c8c0ede"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:02686b281ec9b8e9614f6f0db27e12e4cac515f336c5d05b678bc9eb73ff2af5"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:a8ab2c2585b2e22e48cd4aff2136fce4eab4ced569b85f150ef9a5a5d6f34d49"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:c652d9f5d52d8b81eff3ff2c95c9dd0f281de45f6ad5d3fa95527550c396454a"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:121e1d08d58b75c15491cc3003609cbad71ac2b0016043c36d30a6b2032b6605"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-manylinux_2_31_riscv64.whl", hash = "sha256:30cf89b6a6844542d116c758725d13de5b97b9410a51ee577715f761270d0c1c"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:732aaae9e498c973fc63fd7f308009b57c4afa6ff59d0b674d894e57d0bca01a"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-musllinux_1_2_aarch64.whl", hash = "sha256:d9c1a920b2aa0c1e016bed7721689717034ca0dba64ac3a3ff243b003d4e8dd8"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-musllinux_1_2_armv7l.whl", hash = "sha256:847cad4c22953823f5e76d0b126c6f9da670cfa45b1df41b9324e22189d77611"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-musllinux_1_2_i686.whl", hash = "sha256:aca07de3ca8e3d9f2b716c3c31f4f384fe4bd2d03e09912417cff7a3e08036f7"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-musllinux_1_2_ppc64le.whl", hash = "sha256:5db12d843d931b9d1f1519ce66b9cbfe72d8563b75c1204d255322254d1964dd"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-musllinux_1_2_riscv64.whl", hash = "sha256:fb8021f4dc3296de18737e3b8b34994411da725374d74a0b9eb32a738f2a23d0"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-musllinux_1_2_x86_64.whl", hash = "sha256:0d764913374ba21df8cd618045b371ed104ce55c6276de7c59e554a4f5ba26fd"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-win32.whl", hash = "sha256:2bad302ae2427454ef9dab71bd5c3ad67188860c39d36daee38f72031edb39c4"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-win_amd64.whl", hash = "sha256:032b09d86c785c3984e975378b7861f2a68ac44e22ba7b07b93a536b2524eeb8"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-win_arm64.whl", hash = "sha256:929eda0c419e622ab49df3a4c45f81c57dcc10874bbad55d0470a4a03324dac3"},
+    {file = "ast_serialize-0.13.1-cp315-cp315-pyemscripten_2026_5_wasm32.whl", hash = "sha256:bf53a3ba1cbe45399fc1e4ccd684b8f2a11cca450f89706b79122f478534c6f5"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-macosx_10_12_x86_64.whl", hash = "sha256:4bcf1d5f4eefecaa325a34ecf079b2741189e7d0821523f92cb343a389a67489"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-macosx_11_0_arm64.whl", hash = "sha256:1b158f1c9e3a0d386913924c57995d6df78224868a4d0f215bdd41587c97b34a"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:77a6201b69ce42e919f2fd9a559138d5410b5d8afff7c3a6b1f500b0481ff7d8"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:6eb0be64c9b8f1ed26637eef345abe75ea90351c92c85a8aec7838b3af971011"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:7f2427bddfca208e84de700044fadf9aef2769a70bae740a1bdabb89ee9e1117"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:88df960e67184a71874e5b2ee014f87d1254a0dbc1f903c7bd7b1ab7733ecaef"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:9f99786e0d62082a5b121b3437f3167562133711df22128de1e9886f57f0b226"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-manylinux_2_31_riscv64.whl", hash = "sha256:a01b515b25a03184eff4acef25b981ae007916379682e908fc8785b342a0bf3e"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:29422d7bcd9e7f39411571c36cec45ce8a7a9b3ad4b2e7aa64497c15a9e397f6"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:6ffb0f548a7dc126067fdfb7b2325ab5dfcdeecc29faf5e0f0461c96d6ebfa3d"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-musllinux_1_2_armv7l.whl", hash = "sha256:2876444a5f87a7c76f371af3c07c2fda563a08bc8af52568df812e8aba4bf568"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-musllinux_1_2_i686.whl", hash = "sha256:885a62ec6481ce58ab24c6357360d7a82023b62dd7e5e7ceccf59e9662a3272b"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-musllinux_1_2_ppc64le.whl", hash = "sha256:2015fd0165e463aac3b1aadc6a1fa1e5820c14ca950821f4fe02f8a3f6a388d6"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-musllinux_1_2_riscv64.whl", hash = "sha256:eb41e1906502ac80d24ff343f45008552a3f5b5158607959fe6a07dbfc6eb037"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:ca9acc67f796aa2554fa3037dccc022a4b08b38b2ad153edfac9c9bb7cb1f3c6"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-win32.whl", hash = "sha256:4f98da562e5901d69c7dd30885594c5438f859edf354e9ac7517d188fb026c3a"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-win_amd64.whl", hash = "sha256:ddf7cbc9c8abf70102c4a088efe154bb86bfb825b0a10e556d2b15972cf99bd7"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-win_arm64.whl", hash = "sha256:48f11d96c9ec71119612f8e79a8c55c12f2b136716d68347433c561d27a25f43"},
+    {file = "ast_serialize-0.13.1.tar.gz", hash = "sha256:401cc84e8888d4d76f2c8d296fa006ed1a5ff993b6aff90dd777c03973efd025"},
+]
+
+[[package]]
+name = "attrs"
+version = "26.1.0"
+description = "Classes Without Boilerplate"
+optional = false
+python-versions = ">=3.9"
+groups = ["main"]
+files = [
+    {file = "attrs-26.1.0-py3-none-any.whl", hash = "sha256:c647aa4a12dfbad9333ca4e71fe62ddc36f4e63b2d260a37a8b83d2f043ac309"},
+    {file = "attrs-26.1.0.tar.gz", hash = "sha256:d03ceb89cb322a8fd706d4fb91940737b6642aa36998fe130a9bc96c985eff32"},
+]
+
+[[package]]
+name = "certifi"
+version = "2026.7.22"
+description = "Python package for providing Mozilla's CA Bundle."
+optional = false
+python-versions = ">=3.7"
+groups = ["main"]
+files = [
+    {file = "certifi-2026.7.22-py3-none-any.whl", hash = "sha256:62f22742b58a1a33014a2b6b706588a8d7e2a88ae7bd1a6ebe8c992928483775"},
+    {file = "certifi-2026.7.22.tar.gz", hash = "sha256:741e2c3b351ddf169a738da9f2c048608ff7f2c5cc02f1ebc6b118bb090d5d55"},
+]
+
+[[package]]
+name = "charset-normalizer"
+version = "3.5.2"
+description = "The Real First Universal Charset Detector. Open, modern and actively maintained alternative to Chardet."
+optional = false
+python-versions = ">=3.7"
+groups = ["main"]
+files = [
+    {file = "charset_normalizer-3.5.2-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:195c26fb65950f8fce54e26349852b7bdd7c5f120aeefbcc440b8a20faaed4a3"},
+    {file = "charset_normalizer-3.5.2-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:9373ad13ef0d2c0fb761e04e55bfdee5a08b52cef2c882c8fbe9935b1517152e"},
+    {file = "charset_normalizer-3.5.2-cp310-cp310-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:ddf19c062bea7a0cc80f519243d2c01dd091be0cf952a0750d4ad576709559f5"},
+    {file = "charset_normalizer-3.5.2-cp310-cp310-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:3d14b50de6bf4d0edf857a9386836846f982b8f524e188e2e68b96d702bcf4aa"},
+    {file = "charset_normalizer-3.5.2-cp310-cp310-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:28a15fdad492a99b6eccfaaed66ef3f74050680545ea61ec8b2f4c538f1f1320"},
+    {file = "charset_normalizer-3.5.2-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:8a893cc101149f80a653f82062ebc95b34525a2614382e1da5458fe7c6997249"},
+    {file = "charset_normalizer-3.5.2-cp310-cp310-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:619799369eeef6366ed3e8755a5670f4f2f0fb6b30a0fd7264dc0fdc2357058e"},
+    {file = "charset_normalizer-3.5.2-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:447441e76ec720b15e64418d32e092297340387053047c7c694f579efb0ee1d9"},
+    {file = "charset_normalizer-3.5.2-cp310-cp310-musllinux_1_2_armv7l.whl", hash = "sha256:62588a277bfb59def052abd940703fa35107152bf479781a878617d60faf8fb5"},
+    {file = "charset_normalizer-3.5.2-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:44bd4fbb29dfbeba60e7d2bd000c59e4b21ddb3cc53912b14048d37092706d7c"},
+    {file = "charset_normalizer-3.5.2-cp310-cp310-musllinux_1_2_riscv64.whl", hash = "sha256:30fcd120b732aa79317f08dee04d7de0847822e4cf7ee0e9f445bb958832252c"},
+    {file = "charset_normalizer-3.5.2-cp310-cp310-musllinux_1_2_s390x.whl", hash = "sha256:50e3adfb96fc189eb27b1cf62d3b598b89b4bb0420d93a3d3e42e137409011be"},
+    {file = "charset_normalizer-3.5.2-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:b736353c0a625bbd5fcec108576e2385db3496f4f771f785ff32e108d3c3bc45"},
+    {file = "charset_normalizer-3.5.2-cp310-cp310-win32.whl", hash = "sha256:f5833ad231be5eb6553de524a70f48d71b2c8563101750531e0b80184e175cd4"},
+    {file = "charset_normalizer-3.5.2-cp310-cp310-win_amd64.whl", hash = "sha256:1461ac396c4fdb983a675f20aa555624f0ee18ac83d832b9244ffff3d8055275"},
+    {file = "charset_normalizer-3.5.2-cp310-cp310-win_arm64.whl", hash = "sha256:c6708715abcf3c73b99508253e961a9967f02fe536532834149574eda6de0d1c"},
+    {file = "charset_normalizer-3.5.2-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:3d21b8b13c7592db2ac5e544a6d83187b995257472b0c9e8351b6d507ae37ed6"},
+    {file = "charset_normalizer-3.5.2-cp311-cp311-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:d760fe2a4d7c3b226cb9026d6a842868d52a7901bd98420e1baf14e80da85cf5"},
+    {file = "charset_normalizer-3.5.2-cp311-cp311-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:c9790464842f85f437dbbb54417eda1e0e6bfc52dd8d22d6fd1c994b73b2dc74"},
+    {file = "charset_normalizer-3.5.2-cp311-cp311-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:4685902cf26edf013ed7a3da0f426ebba7a00ebb9541386d835afbf002c11cab"},
+    {file = "charset_normalizer-3.5.2-cp311-cp311-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:4495c5002a7b28557e7e222e77e0b661183e432b7d6d2e788101e3f240e05b8c"},
+    {file = "charset_normalizer-3.5.2-cp311-cp311-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:211d5a3eb6af8f513b8d4ca19a8c1b7accab1b5f0d3175f9826b03c1a920dc1f"},
+    {file = "charset_normalizer-3.5.2-cp311-cp311-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:ef4fcbf3327382cd4c9f540babd61248208af7b93eec4de397b4d5f58a09e288"},
+    {file = "charset_normalizer-3.5.2-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:bd16aabe4a02a297c23417aa17ac6299dbd8c49f673bcd645b4929b11f5a4400"},
+    {file = "charset_normalizer-3.5.2-cp311-cp311-musllinux_1_2_armv7l.whl", hash = "sha256:fb9e68df06293761f9fe66ade60a9bc6d0f5e42b8acf2939a9158af86ab0e5bd"},
+    {file = "charset_normalizer-3.5.2-cp311-cp311-musllinux_1_2_ppc64le.whl", hash = "sha256:59f63901b0031c3136cf64704dcb21de0bbae62ce2c9529bc39d27665463de37"},
+    {file = "charset_normalizer-3.5.2-cp311-cp311-musllinux_1_2_riscv64.whl", hash = "sha256:304d5463e65a35d7bb0850550e0780395395f6fcf452f04db7d5ca7cecc425ac"},
+    {file = "charset_normalizer-3.5.2-cp311-cp311-musllinux_1_2_s390x.whl", hash = "sha256:9cf9b1a857e25c4baceeb3624e92a56df3668f398c4acba74e174d81fb4d1d3a"},
+    {file = "charset_normalizer-3.5.2-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:114e4d0c92d618409ed82a99e22b5c5e768fe995f2973f78265f4524f49d4640"},
+    {file = "charset_normalizer-3.5.2-cp311-cp311-win32.whl", hash = "sha256:2625388c6c754520c37abaf3b41eb34d1cc4a373f457898f08606c8e362b891d"},
+    {file = "charset_normalizer-3.5.2-cp311-cp311-win_amd64.whl", hash = "sha256:87e50a3e7cb90af586b6c5faf23e302a970415ac73bd7bd90a515a04b427ef96"},
+    {file = "charset_normalizer-3.5.2-cp311-cp311-win_arm64.whl", hash = "sha256:254eb48b9fa5ee9898a3c445825a1f340fe53712a098904b39b0bddba8ea3cb1"},
+    {file = "charset_normalizer-3.5.2-cp312-cp312-macosx_10_13_universal2.whl", hash = "sha256:ed2a239c0ea213acc1908150a3037257083c7c083128f1a4cec2ec4b97dca491"},
+    {file = "charset_normalizer-3.5.2-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:b91363207bd9dc966a691e959bb47f64b30f7ac4b072be9968b366982f7db77c"},
+    {file = "charset_normalizer-3.5.2-cp312-cp312-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:38a873987f3be698494da8b2e3085e29da02da7b633dce73e79c699a113d7bf0"},
+    {file = "charset_normalizer-3.5.2-cp312-cp312-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:355ad8011081dec5412240c087a9a0c9d4d5039f3ed11a3f13e18c2b29b56c51"},
+    {file = "charset_normalizer-3.5.2-cp312-cp312-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:ee21e28f0430bd6dc9086c6e525d5e818a44a5ad19720c8a0ef766792f3eb5e5"},
+    {file = "charset_normalizer-3.5.2-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:3d31298449090ab8d47b7b1b2a555ff73cac7ed438a08b7ac160980c7ebed649"},
+    {file = "charset_normalizer-3.5.2-cp312-cp312-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:5cde776b7cc66e4f6c99612cea4aa7269aa65863f7a15841b2c264f103822f4e"},
+    {file = "charset_normalizer-3.5.2-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:ae4f5fea5b8b8ccff88238cc8569303e5ee95efae67fa62922a311397a71f346"},
+    {file = "charset_normalizer-3.5.2-cp312-cp312-musllinux_1_2_armv7l.whl", hash = "sha256:f7d486c83842422badd511868fd8a9a20e9407ace71564b6af47ce7e60a336c1"},
+    {file = "charset_normalizer-3.5.2-cp312-cp312-musllinux_1_2_ppc64le.whl", hash = "sha256:11a4d68a6ecda3292cb1e50239e111543ba5d709bb62a6b4ea1afcfa729d8875"},
+    {file = "charset_normalizer-3.5.2-cp312-cp312-musllinux_1_2_riscv64.whl", hash = "sha256:d6734d2ef8a50fbf8445c139477da401f50d62a0606bf00e20ec6d87773fefb1"},
+    {file = "charset_normalizer-3.5.2-cp312-cp312-musllinux_1_2_s390x.whl", hash = "sha256:a815775b6c38d4e0ff7bcffbeba67feded90202bb6a226b8dd35f1c855217413"},
+    {file = "charset_normalizer-3.5.2-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:23851fb4e1b85ed3f6c2a27b777cdfe2e19fb5b38429a8faf38c7542b7665869"},
+    {file = "charset_normalizer-3.5.2-cp312-cp312-win32.whl", hash = "sha256:db19d07e2e0129e974a0e65d0064fc222a446cd5122c2fd4184d2af9fc734a9e"},
+    {file = "charset_normalizer-3.5.2-cp312-cp312-win_amd64.whl", hash = "sha256:780fbe7cab297b81dad9fb8dc5eb003c0468ffb0d9e5f65068c53a34661a96bc"},
+    {file = "charset_normalizer-3.5.2-cp312-cp312-win_arm64.whl", hash = "sha256:e2af3aad578aa6bd1384bcf4750fc285e5a9de53f40b7d41e5a0bf748edeb2b3"},
+    {file = "charset_normalizer-3.5.2-cp313-cp313-android_24_arm64_v8a.whl", hash = "sha256:ed905975ab14056a2e5eb1c376cb2e1ebc5396baf84163939c518556fccde9f5"},
+    {file = "charset_normalizer-3.5.2-cp313-cp313-android_24_x86_64.whl", hash = "sha256:a66c3bc5ab1f0ff2164fc9965ddd611ff0802173f4b9d24554c563f6ab7e1d6e"},
+    {file = "charset_normalizer-3.5.2-cp313-cp313-ios_13_0_arm64_iphoneos.whl", hash = "sha256:d2374b62878abb00cd8309b32af6c0b715cd02dec0ca74ef12e5069bdc64144a"},
+    {file = "charset_normalizer-3.5.2-cp313-cp313-ios_13_0_arm64_iphonesimulator.whl", hash = "sha256:d376bbd28b3a8999db1a103b3b388aee6f1ddeb3e51bc2172993efdcd86e064d"},
+    {file = "charset_normalizer-3.5.2-cp313-cp313-macosx_10_13_universal2.whl", hash = "sha256:6045373d5a89a5ec71afde535db987ca28e76dfa276c2d4c818265b375d4b055"},
+    {file = "charset_normalizer-3.5.2-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:849df64e889b2e17230d58410a03dba311a65b163508fd33679b2b737d4b7858"},
+    {file = "charset_normalizer-3.5.2-cp313-cp313-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:15c44f7edfd477b06f517a5cc317fc1707edb9de2c865f43d4b6513907473234"},
+    {file = "charset_normalizer-3.5.2-cp313-cp313-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:a89012d6d5476ee112d20d998570ed58df2260a852afb1758809cd6900411d21"},
+    {file = "charset_normalizer-3.5.2-cp313-cp313-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:0c951d5e6dd9c2ff60609476752bee49da4206adde960ebc247766937f72e718"},
+    {file = "charset_normalizer-3.5.2-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:7218e8f32b0956cfcd048fd42d9d5779809745ca1d86113ca56f66e7ae1549c4"},
+    {file = "charset_normalizer-3.5.2-cp313-cp313-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:a19a731138fc27d5682277d3b9df22855cea1239bce7fcec5f78f42ef2d1f3c3"},
+    {file = "charset_normalizer-3.5.2-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:62603db9a7caa0802eaa28c1c46fecd7b3a263a774069c24c3c28c302448721c"},
+    {file = "charset_normalizer-3.5.2-cp313-cp313-musllinux_1_2_armv7l.whl", hash = "sha256:b6856554c4f44d79fc2307d5768854310a8f0096e501c75637542c82292b0429"},
+    {file = "charset_normalizer-3.5.2-cp313-cp313-musllinux_1_2_ppc64le.whl", hash = "sha256:1bc0baf5ef96b6ede57d47f4b8fe4d9d84019c3bfcbeb20a41edc6a6ee341f1f"},
+    {file = "charset_normalizer-3.5.2-cp313-cp313-musllinux_1_2_riscv64.whl", hash = "sha256:56bc200a365efb37383b7852e4cc5898d3b2da5987289b543956cf8cad71018a"},
+    {file = "charset_normalizer-3.5.2-cp313-cp313-musllinux_1_2_s390x.whl", hash = "sha256:2c9ad19a6cfcd5ea5c0d41161d22f9df1dcc277e9bef2751391334546a314c00"},
+    {file = "charset_normalizer-3.5.2-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:e243bd13217235fc7290c621941c3f5cc8b66e4872495be821d7436ba2fb838d"},
+    {file = "charset_normalizer-3.5.2-cp313-cp313-pyemscripten_2025_0_wasm32.whl", hash = "sha256:a090bb2c68df85450502e3e20d665e3a5af9c65a84d6508ed477badd49166fd3"},
+    {file = "charset_normalizer-3.5.2-cp313-cp313-win32.whl", hash = "sha256:2b7b3bbfb4fe8ef40600792d762fbaa9057559f9d3fad209525b7a22b99e91fd"},
+    {file = "charset_normalizer-3.5.2-cp313-cp313-win_amd64.whl", hash = "sha256:78456a747de8dc58360ffa581f30a002baf5aa28cb262536545e91f113ed7639"},
+    {file = "charset_normalizer-3.5.2-cp313-cp313-win_arm64.whl", hash = "sha256:11912e4bb14baae7c5d8791aa55ba0a3a03ec6729073307b0f57270abaa713d3"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314-android_24_arm64_v8a.whl", hash = "sha256:1afb975bd5d68d5ce9f6b6d44fdf2f7e34b895a35e95708a7a91b20a3b51d187"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314-android_24_x86_64.whl", hash = "sha256:bbbfc8e28816f19d7c0f1816664980c0a9875d01b27cdf8eedddb639d9e108ad"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314-ios_13_0_arm64_iphoneos.whl", hash = "sha256:7967d08cf06dee78443b874f98c98036f624f3a4e73e11f9f64f5be4d25393cf"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314-ios_13_0_arm64_iphonesimulator.whl", hash = "sha256:4c2b5031f63e331e3839b40aed2dd6f191e9c07edbde303e7876846ea1946995"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314-macosx_10_15_universal2.whl", hash = "sha256:fcff63213e8e6e47770541a4607175404f47cbb3ebea7b6058cc82d524a0e424"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:8d86d6fc60743dc916eb79e2eb1ec4818e21e427731543af40a3021851174a13"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:7a881931aa470808df94a8c380eed2bbbc76cd9dc622310f99665658c821eb6d"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:8024d00c3faf3fc0c16e07a69f4405e8eac7cc0ab15f65fe6cf43827c4cf72b4"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:4d48f2d08b9de5864e2c8744d4461b862fb149a18274abc8b698c45975573438"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:34276fd796040bf0993ab33a369aa572e6979c7aab225a88893667ad8eac8f7a"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:0521c5665880b33d603717defa76c094048900010897909952397feb3039da56"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:eff0ac9dbe711a4aee69bf04a83896aa9b85f19641264053a9f6d48573abb7dd"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314-musllinux_1_2_armv7l.whl", hash = "sha256:1503bccbeb36d5527790c3930327704c39af22de3112f1b1666a9f3ce15ee204"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314-musllinux_1_2_ppc64le.whl", hash = "sha256:52aa6992700996af31f375de0c6bacd402b0097fe40b53c426b9f51a90ebabc7"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314-musllinux_1_2_riscv64.whl", hash = "sha256:e09a3942ecbdee5cce73ea9d42da82b81b72ac1bf031ce069b93b5adf4eac8cd"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314-musllinux_1_2_s390x.whl", hash = "sha256:c7c9ab723cde841fefb34efbad91e87f00a674b1fe1cd0784fde742bf2c154dc"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:ddc7dacc8ece3a182e7f15cb862d1fd616b46d076cb1ae9dd232b2c38b655874"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314-pyemscripten_2026_0_wasm32.whl", hash = "sha256:ee43c17b173d46a3212baa6ead3ae258eeabdae48c263a01ccf0218c366dd655"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314-win32.whl", hash = "sha256:4f87960d57feabfb618e4e0af6e7371645fa26a277860739d6e5d6e0012c92f0"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314-win_amd64.whl", hash = "sha256:e4e81e09c1578b8df602e3db08b0b3ea0a6947ad612f52bf8dc5ea8d47691f0c"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314-win_arm64.whl", hash = "sha256:80d02b6f04e92601a081dd97b23d3128033098bff5d35d392ddcc0476ea11253"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314t-macosx_10_15_universal2.whl", hash = "sha256:dca9ab98072a5a54ebacebdc45f53e645336b320c667410b061be1ca588ae709"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:f0aa869112ef88429ae17820d99c3dd9504c9e9c671d3c246f3d7442cb051084"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314t-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:c0afc6800ba57ccc350374c5bd6150419915d95ce93cdbab2d783d75eaf30ecb"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:7dcd882da75ef9adf94903b1e3b9419e8aa8fb4c7396822b834b9ef7fb96954f"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314t-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:2e06a3a98f916dd41d27f3105e02e7a40181c98c94b9158733d03a6f80506c09"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:6bd128f206a7752ae1f2ab6c61bf8a24ba28913a10df8b14c2637b973ff97a80"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314t-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:c8f3d67aeaf55f017982b73683f0e7342ba2f6635a78f69ce89ebb26aa411e5c"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:fe9753dfee015c570d73df76f899f18444d41388bffcde097deba51c4fadbb9f"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314t-musllinux_1_2_armv7l.whl", hash = "sha256:92888bb3187c5ba50500b00b3b310c9f2c651709d28036077680cb5255450a03"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314t-musllinux_1_2_ppc64le.whl", hash = "sha256:d008d90a7f2471519aef0c90dfbe73b3e6e4d5e66ac48e19154c17e89e98b604"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314t-musllinux_1_2_riscv64.whl", hash = "sha256:31f3930700408d211f13378ccbe1c40845d8da54bd0681fac3a9b5aae81c7aa8"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314t-musllinux_1_2_s390x.whl", hash = "sha256:2a925889534b3748302dae5dead07cc13480de1dac3aea80a941b729b471ef93"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:f5ec61164adcec446f8969a3358ec3f9b26bbda3b9213e5586d219afa8df2915"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314t-win32.whl", hash = "sha256:598a11a2c7ebaa5334bf698bf29568c9c390abac6a154d8170fedecd1cea38c5"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314t-win_amd64.whl", hash = "sha256:7fdde2c9fd9e3eca40631e024664cf2584272cc8f96308cbe5fdfc930f51d8bc"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314t-win_arm64.whl", hash = "sha256:d1befeed746d247c81127bb14de9dc3d30edb6e5976d34f83f86ed262b1d9105"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315-macosx_10_15_universal2.whl", hash = "sha256:87475fabc8d9996fd9c27debb395e642e8c838d78a00b6e932227a0e06b81e26"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:9409a8bf35cf78353942504b24a57de3d75b708997a1e4bd8db71ac8633ce364"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:498dc3188ca05a68231ac3fdbfc7f57eb67e1343c30e0fea17f8218c1599b253"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:e242bb1c5e76e97dfa9e7f209a71e93a01d7f19ffdd5cfbb2e2d55b4f08f8ab0"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:def79fa35ef0cef8d2accec024f4fdc7ead3012ff02f5215c783f39f03ef8cfc"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:3df041de8887954562c9b261cba85ca0e9ded74048daf125f45edcfaa4832229"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:04851f73ae72b8413dddadb16a49dfee95263553741fd42d546f7d66907e6be5"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315-musllinux_1_2_aarch64.whl", hash = "sha256:183b88127acdb4fabe59d951ab424faf1af7b63cdbb5f776186c1ea2ffcaed98"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315-musllinux_1_2_armv7l.whl", hash = "sha256:16fa0eccf81304b79c5cd87f9271c3b85dd9dd99245e4422ae9c0dd45e0f99d3"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315-musllinux_1_2_ppc64le.whl", hash = "sha256:7441d755b7ab94f8d4eb3e43ec05482d760842fd263d003a99102d742cd835e2"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315-musllinux_1_2_riscv64.whl", hash = "sha256:ca403d7e4798f525fdfc78e258820419cbbd0f0ecbab9de7840e3c017cf6b8cf"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315-musllinux_1_2_s390x.whl", hash = "sha256:df29a0a7107f7011e77f4eebdddec4c7331e24d787a0b21a46d63bdf7445da95"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315-musllinux_1_2_x86_64.whl", hash = "sha256:f3c96f633825733f735c5a9cf21d21a257d8e1edf0b1cee0a064b9c424ca0f7d"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315-win32.whl", hash = "sha256:281cb91036248400f4cc957495cccd44c275c2e0c5854f7e45ac5cf7dc193847"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315-win_amd64.whl", hash = "sha256:89b53f3cda69831909888e0494f4fa0bcd3537e3e138dabeb620bd6ad946bae8"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315-win_arm64.whl", hash = "sha256:6be488a102b8cf28d0391d8c4ba7748938ae28b78ad901f8585520fca33ead1a"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315t-macosx_10_15_universal2.whl", hash = "sha256:915563965d418f986e7e145accc592eae9e1a1be3566ff98a05d7a9ec42a76e1"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:65cd72beeeca9d3aaea1201e5923859f308f952f9c71de93f06063c79f0f7a3b"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315t-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:b7fd005a73d9e657273b7a10dc71a9e03c8fb9ee6999798d6918ce095b81ac7f"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:e54da4baf05720032d527874d40b65fa4d7e5c6c6a43d0c3adbeffcaf275a2b3"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315t-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:124fbf1a8ff966d87ae05bb8bd45a71f966055ed8bba320d0c7cf450bc5f4d0e"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:28b4f0d66fb834ff90f28209ac7bce77868c45d8c93e26f906709d9b7c2e1af9"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315t-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:58ca3755ee7ff7f59b57789ec9833c9de9ea275405cdd240eda1f193112e398a"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315t-musllinux_1_2_aarch64.whl", hash = "sha256:443eae2bf318abeaf6f15d785138f71fd6de770e99a92158b8b814265e079115"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315t-musllinux_1_2_armv7l.whl", hash = "sha256:58f361dcbab699cf8f42db3f47c8e7fd1036f138c23a5d08de9fde5f425a730c"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315t-musllinux_1_2_ppc64le.whl", hash = "sha256:1b4cbc7c3491ccb4aa17fcd8165649d01cf39f76de1696da8631b5f71b85401d"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315t-musllinux_1_2_riscv64.whl", hash = "sha256:ba0b1d2620edf869789c3879223f52bf2afc5d31b3cb47cc57b3a12c05e2aa9d"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315t-musllinux_1_2_s390x.whl", hash = "sha256:5e2b6b57e9733d39f0c9fd3185efa6b8e29652c4cd8fe94180272cf6ed9a78c4"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315t-musllinux_1_2_x86_64.whl", hash = "sha256:51cf45226a9b588d0d2b4880c62d686934b63ab0bd79ca23ab0e9762eb27441b"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315t-win32.whl", hash = "sha256:5fb29fb8cd1a46c27a1bf9613ad5ec2599310d46b4025d9556404a6b6a292800"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315t-win_amd64.whl", hash = "sha256:a192e2c40070d92c3ccf777e3a5c4ff515573cd2bb7ed0c537fdadbbec5bbf21"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315t-win_arm64.whl", hash = "sha256:749e97e1b32313717a565abbe321bc2190bc8b35f1a67e4cdbc7c56c8d8ffe58"},
+    {file = "charset_normalizer-3.5.2-cp37-abi3-macosx_10_9_universal2.whl", hash = "sha256:4275811936e2f06feff5e598fb42a1b7ae852da8e39605211892b56b81a34efd"},
+    {file = "charset_normalizer-3.5.2-cp37-abi3-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:1c50fe28bbc2ced33386f298650d91218076c05420e6cbd790b913adc41659e7"},
+    {file = "charset_normalizer-3.5.2-cp37-abi3-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:d19fbd981a488e22cd04883659ca6b08f50b5974f9fd7c95655ef6a043e5893f"},
+    {file = "charset_normalizer-3.5.2-cp37-abi3-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:0fed1d06615f022ee3b13caf5e8b180cfea32bb2c5aded8a9d44277afc040f93"},
+    {file = "charset_normalizer-3.5.2-cp37-abi3-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:838dcc90063569a0448120554591a1d6c4a4ffe11babf048908793154ab86ade"},
+    {file = "charset_normalizer-3.5.2-cp37-abi3-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:2ce45c6627b22c47e390bc91a41c3d13032192e699fa0bea96e9671b373d69b0"},
+    {file = "charset_normalizer-3.5.2-cp37-abi3-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:0774bf9bf620249fee3e0b8b9fd3065de213be30f3aa94ce2494b3b638949e26"},
+    {file = "charset_normalizer-3.5.2-cp37-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:1db38f4c5496827c1a501846d64d14c3b80c7e6714e406cd7dc36a9899fa1011"},
+    {file = "charset_normalizer-3.5.2-cp37-abi3-musllinux_1_2_armv7l.whl", hash = "sha256:304d8e4d493af723536393eee0c689eb7813f4a474c8b479dee63f1fdd98f621"},
+    {file = "charset_normalizer-3.5.2-cp37-abi3-musllinux_1_2_ppc64le.whl", hash = "sha256:9b7f416ff0978e2f2249330527f0ad6fa02f4932e6199692d3b52da2048c19e4"},
+    {file = "charset_normalizer-3.5.2-cp37-abi3-musllinux_1_2_riscv64.whl", hash = "sha256:01077390b03f7988f11d700a2194e69b119741a86b1a638b1db88891e3eced8e"},
+    {file = "charset_normalizer-3.5.2-cp37-abi3-musllinux_1_2_s390x.whl", hash = "sha256:7e841fb9010836c992c9f12fcbd43a831de93a5f726fc1ccd8ca1d0268c5014c"},
+    {file = "charset_normalizer-3.5.2-cp37-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:9cae88599c7219005d879f98e5ed53341e9a122af585e1091200358a3003d2a0"},
+    {file = "charset_normalizer-3.5.2-cp37-abi3-win32.whl", hash = "sha256:01b0c0d2262a9e28e8484a278c7e1b5d650e3ac8cf2683d2967e25899f208bdf"},
+    {file = "charset_normalizer-3.5.2-cp37-abi3-win_amd64.whl", hash = "sha256:9f56f72050826f63dcee7a7f55b0a77168cb3bfc553fd405e7f8f9ece75a4036"},
+    {file = "charset_normalizer-3.5.2-cp37-abi3-win_arm64.whl", hash = "sha256:40ab6bffa02ae10a0581e6c198be7d2d8ca5c2a0c64e4ed3465d766df457573e"},
+    {file = "charset_normalizer-3.5.2-cp39-cp39-macosx_10_9_universal2.whl", hash = "sha256:75a3ceed0724d625d64b86ca20aba182e4df462e04c2414fc941c0f523f06aac"},
+    {file = "charset_normalizer-3.5.2-cp39-cp39-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:0891b9d3903c5571c03771ca669a4b0ec5618ca722a5c957d3d29cd4e5062848"},
+    {file = "charset_normalizer-3.5.2-cp39-cp39-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:fc14a032f813bf5fe624d991960ea83e9715adc27e4c1830a2361eb1d02ac341"},
+    {file = "charset_normalizer-3.5.2-cp39-cp39-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:8b2bfab86aa71ae13aa41a6a26aab338e0db2b8bc75434b05aea89e011ff35a4"},
+    {file = "charset_normalizer-3.5.2-cp39-cp39-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:9bde855991b7e362c146535e3136a50bfaffc0487d38b33ca7e5edefc6e23849"},
+    {file = "charset_normalizer-3.5.2-cp39-cp39-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:55ea99acb17b9325618de155a0cd6a2e8f5d10be008113e1d433bbb58db543b2"},
+    {file = "charset_normalizer-3.5.2-cp39-cp39-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:68eb192d85ab8e5f6ec69c2bc6ac0179fbf04a5ac1569d12fbef74883fe102d0"},
+    {file = "charset_normalizer-3.5.2-cp39-cp39-musllinux_1_2_aarch64.whl", hash = "sha256:d913de495d90407cd859d263bee2e5d1a4ed3eb6573c04e70d9ec619a7cbed7f"},
+    {file = "charset_normalizer-3.5.2-cp39-cp39-musllinux_1_2_armv7l.whl", hash = "sha256:3ddacd27458c45bdacd6bd6db644bfb730efbf9e830310186e3045c9c5be8fb2"},
+    {file = "charset_normalizer-3.5.2-cp39-cp39-musllinux_1_2_ppc64le.whl", hash = "sha256:588461c2e8384d309bd63e5826019b6977bc66d629b99ac8737bb795d7b2cb5a"},
+    {file = "charset_normalizer-3.5.2-cp39-cp39-musllinux_1_2_riscv64.whl", hash = "sha256:e80e6c2f55656b4824d72065abb4ddd6a525c74bd78a0aab5d9fc2cf4fb5af50"},
+    {file = "charset_normalizer-3.5.2-cp39-cp39-musllinux_1_2_s390x.whl", hash = "sha256:d4a7319f304a774bed22115bc891618e45f85065ab44ea6acd07d274e750519a"},
+    {file = "charset_normalizer-3.5.2-cp39-cp39-musllinux_1_2_x86_64.whl", hash = "sha256:fd1fbe0f116b6e55da77aca2c6ddcddcfac2186cbf78bdebf40fc156efca389d"},
+    {file = "charset_normalizer-3.5.2-cp39-cp39-win32.whl", hash = "sha256:93223adc95033dd47133a46ccfc316a0139176fd79085762e27202ec56018f03"},
+    {file = "charset_normalizer-3.5.2-cp39-cp39-win_amd64.whl", hash = "sha256:15bb4005af6320d259dc7593ca84a38d7fe06a421dbcf7b910ae23979101e787"},
+    {file = "charset_normalizer-3.5.2-cp39-cp39-win_arm64.whl", hash = "sha256:2cc961b171b3f3440f410489ab3573e86aea8736134ebbb40ea1338b7f0831bc"},
+    {file = "charset_normalizer-3.5.2-py3-none-any.whl", hash = "sha256:b6b751274acb69d77b3323d6b7dbaa3c7fdfc1eb829b7eb61d262f32e1af9685"},
+    {file = "charset_normalizer-3.5.2.tar.gz", hash = "sha256:39de2a259fc954455c57274dc94c79d5842774e1247a016aff30bc0efed0f4ef"},
+]
+
+[[package]]
+name = "click"
+version = "8.5.0"
+description = "Composable command line interface toolkit"
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "click-8.5.0-py3-none-any.whl", hash = "sha256:255bc9599cf7748b4b1a446ccc735421bd08a2ae529a8b88597d3de5664ee360"},
+    {file = "click-8.5.0.tar.gz", hash = "sha256:ba0d2089de75ea0310e2dde03160e6ca10009947fb95a182f9b54021bb272e34"},
+]
+
+[[package]]
+name = "cloudpickle"
+version = "3.1.2"
+description = "Pickler class to extend the standard pickle.Pickler functionality"
+optional = false
+python-versions = ">=3.8"
+groups = ["main"]
+files = [
+    {file = "cloudpickle-3.1.2-py3-none-any.whl", hash = "sha256:9acb47f6afd73f60dc1df93bb801b472f05ff42fa6c84167d25cb206be1fbf4a"},
+    {file = "cloudpickle-3.1.2.tar.gz", hash = "sha256:7fda9eb655c9c230dab534f1983763de5835249750e85fbcef43aaa30a9a2414"},
+]
+
+[[package]]
+name = "colorama"
+version = "0.4.6"
+description = "Cross-platform colored terminal text."
+optional = false
+python-versions = "!=3.0.*,!=3.1.*,!=3.2.*,!=3.3.*,!=3.4.*,!=3.5.*,!=3.6.*,>=2.7"
+groups = ["main", "dev"]
+files = [
+    {file = "colorama-0.4.6-py2.py3-none-any.whl", hash = "sha256:4f1d9991f5acc0ca119f9d443620b77f9d6b33703e51011c16baf57afb285fc6"},
+    {file = "colorama-0.4.6.tar.gz", hash = "sha256:08695f5cb7ed6e0531a20572697297273c47b8cae5a63ffc6d6ed5c201be6e44"},
+]
+markers = {main = "platform_system == \"Windows\"", dev = "sys_platform == \"win32\""}
+
+[[package]]
+name = "coverage"
+version = "7.16.2"
+description = "Code coverage measurement for Python"
+optional = false
+python-versions = ">=3.10"
+groups = ["dev"]
+files = [
+    {file = "coverage-7.16.2-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:23219888477edd736b6fcaec1272d47d93b926e999641ffea7e53a1738e70b2b"},
+    {file = "coverage-7.16.2-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:40c0f00899fe6181ae7f434ceb200e51f5ee4b8ed10e3b5f0b605f0cae15da87"},
+    {file = "coverage-7.16.2-cp310-cp310-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:a4624f80732f6b427ac58f1f59c577a0994a12e8174b5af6a027b4b58795d4c3"},
+    {file = "coverage-7.16.2-cp310-cp310-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:191803c4996b499fcd78c2ad5e5f767dcc53cb4dc6de6d6a741b443a1821ef02"},
+    {file = "coverage-7.16.2-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:9fd670ac43b709c575aefc25bf52d8a598a3bc5017bddfd0a179152ab06a2deb"},
+    {file = "coverage-7.16.2-cp310-cp310-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:705e5af11d34647efdc170c7840b6857c81cf74be96419a553f237e68e62cb72"},
+    {file = "coverage-7.16.2-cp310-cp310-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:8afd9bf35cc6a1f22eb3634808fa8e0b91902459c5721ef2e4461dfe771d7f08"},
+    {file = "coverage-7.16.2-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:3f43bac1856ba269b905302778d4df433d6006489a192174ad77ac528e395032"},
+    {file = "coverage-7.16.2-cp310-cp310-musllinux_1_2_i686.whl", hash = "sha256:f8475460aa33ee28ac896ab1156d0bb3b6c639f7f8383c2677d3359eb35f8205"},
+    {file = "coverage-7.16.2-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:d6276d78f6fca7d0ac066d5da4165c5acd07829e8305c2cb900b738fb3a75a72"},
+    {file = "coverage-7.16.2-cp310-cp310-musllinux_1_2_riscv64.whl", hash = "sha256:736fde09ea39646d11f8e3b76bd3425c075aa4dd45f24891970bb77c14ff20f5"},
+    {file = "coverage-7.16.2-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:c85d54e7e8a2ca932fe8399301af9b8d5907ea2a455ffaff6e7d1208db83b943"},
+    {file = "coverage-7.16.2-cp310-cp310-win32.whl", hash = "sha256:5139009b5efd2194fc168ee9362f0e191ba612ef5d29242f9269c22f9b8f80c7"},
+    {file = "coverage-7.16.2-cp310-cp310-win_amd64.whl", hash = "sha256:c3305c38a2fa21a4254f2ace7dd9ef5fc569c9a558b66e7017650b3d637fb95e"},
+    {file = "coverage-7.16.2-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:732d950e51f3ba4fb6209c73250f3e8924fefca42953ee04a9e65d8c02414d7d"},
+    {file = "coverage-7.16.2-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:5dca0bb66b4c3d624ba047887bf70270030c150692d543cb501293dc38a9f4b5"},
+    {file = "coverage-7.16.2-cp311-cp311-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:af2a2a8c7c74de0559e0c368d94c8def9e16c58faaee33a0bf081057c4227e3b"},
+    {file = "coverage-7.16.2-cp311-cp311-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:db5f8394e17f877a625b257f2ba0ce8e728a499c2c1579ad66220272cd3df510"},
+    {file = "coverage-7.16.2-cp311-cp311-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:5b3146d2317c75f70df2509066d979dadd941f7021cdf9b5db4bcd8568258e25"},
+    {file = "coverage-7.16.2-cp311-cp311-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:9e1d0ced76318bab499693ff25f64faa343415187cb2e4d7befdfdd391a1cf6a"},
+    {file = "coverage-7.16.2-cp311-cp311-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:af98ad5ed9d6daaca956201e00bb429a7eb2b080426686f70a20353e0f9839f5"},
+    {file = "coverage-7.16.2-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:1d56e4d21c56d2046447733f8b118409597db48c01efe898ee9ac24e858ec2d6"},
+    {file = "coverage-7.16.2-cp311-cp311-musllinux_1_2_i686.whl", hash = "sha256:1d5d0e3b660506fb84f995814e3118a21efdc0c8eb80127da1be627d90093c17"},
+    {file = "coverage-7.16.2-cp311-cp311-musllinux_1_2_ppc64le.whl", hash = "sha256:17228fbca0f22976f797be94e975dcd237799c657d49551c7de1e0654d1202e9"},
+    {file = "coverage-7.16.2-cp311-cp311-musllinux_1_2_riscv64.whl", hash = "sha256:bc0b0ac781d489304b741269857f1f8338b7a26b1b89c06c0344658001ec0035"},
+    {file = "coverage-7.16.2-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:bf1bd822ec4e387ed245bed0d71151582cf7be9e5309bc4145eefe36083d5878"},
+    {file = "coverage-7.16.2-cp311-cp311-win32.whl", hash = "sha256:7ed238d227e23cc300c3d464babdaf9f6ddc740aa1b15a77ae96136e6a7c4516"},
+    {file = "coverage-7.16.2-cp311-cp311-win_amd64.whl", hash = "sha256:a90700f743e29aa3d75a6ff5f01953176a889c00e526194bc4d281731b88d99d"},
+    {file = "coverage-7.16.2-cp311-cp311-win_arm64.whl", hash = "sha256:a336eec40e3520d369b8a6cdabb4f596e69a8b42927ca074aa1452fed943238a"},
+    {file = "coverage-7.16.2-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:218d742afca2b5ad5ca759e93eddedfbcc6eadf8322f080dcefc40b7bd4e2d48"},
+    {file = "coverage-7.16.2-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:a9a638be322a8d76a41cdb17781c7f82aaee6a66493d8ffb7e2c09ee22423d99"},
+    {file = "coverage-7.16.2-cp312-cp312-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:724bd0f1e81856b35e59fc98cf7b4e544a3cb662e4e0864dca73d4326ee9d808"},
+    {file = "coverage-7.16.2-cp312-cp312-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:5375ebd99038021b35e99dc88255022912c06565d316212f4a576e4b08d30f5d"},
+    {file = "coverage-7.16.2-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:7a076277ca9f5750cc230f0f578ebd2620cec60255b25707361699fef6fb465c"},
+    {file = "coverage-7.16.2-cp312-cp312-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:58d4a54c6ea672afef66d49be922a2c69826c5ae1a42a9cd94f0c9c2bacdf800"},
+    {file = "coverage-7.16.2-cp312-cp312-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:0dcbcfcc059117284c603ff8cb61a65872512882f84a8cf0339241f7f7c2f148"},
+    {file = "coverage-7.16.2-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:afdf43b72ef3876c1fe66423b91466e37877c9e81e8cec70542b7e8525b9d1b7"},
+    {file = "coverage-7.16.2-cp312-cp312-musllinux_1_2_i686.whl", hash = "sha256:9acc7f7ec4a1b5f89bd929fde5b8a714f6fafdc6cc18725413d510aa082b47ad"},
+    {file = "coverage-7.16.2-cp312-cp312-musllinux_1_2_ppc64le.whl", hash = "sha256:80d3f7b48d43ee8fc5e8707a8adb43d743a5a1a85256c25a24f9d6d0e2238fa6"},
+    {file = "coverage-7.16.2-cp312-cp312-musllinux_1_2_riscv64.whl", hash = "sha256:126d1af8804d7224421fe991ff65d3ce649081560df7a98b1a5ffff07f9923bd"},
+    {file = "coverage-7.16.2-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:c19cd6d025c1673f22afcd22c7df8a662d779e05d8e3fa6820c22afb895b0206"},
+    {file = "coverage-7.16.2-cp312-cp312-win32.whl", hash = "sha256:152877cdc8a07264882cfcd503ba56a3ef6cba56a70e8c70f6eb8ffd7384789a"},
+    {file = "coverage-7.16.2-cp312-cp312-win_amd64.whl", hash = "sha256:e6c52d3307824ff93b39efd99e4185d557db40bd841452abfb32e5d9151ca162"},
+    {file = "coverage-7.16.2-cp312-cp312-win_arm64.whl", hash = "sha256:a678c0b6b22086ec2427359d22e37445d4a792f5fdbbc744112c7dade65cad02"},
+    {file = "coverage-7.16.2-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:1a37c6e478cf687e1aa30a593d19c92c02fad9d122b51ab73f51b8dc7a0c0fc9"},
+    {file = "coverage-7.16.2-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:0993d0e90858c03943d3cb152e068a20dd4707924deec84dd2230261baae3b1b"},
+    {file = "coverage-7.16.2-cp313-cp313-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:bb2fc905bbf4e6b7f40806ea79e31515abf6349594cdf0adf27c4215f0463204"},
+    {file = "coverage-7.16.2-cp313-cp313-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:4358b9c8c0125b460407f3017c6cce8156e904b32772c5630d27112f52bdbfe5"},
+    {file = "coverage-7.16.2-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:1f15254427c9b33eedac4f198eaf9e356eb4f6214551afb43da6194a2c088ad7"},
+    {file = "coverage-7.16.2-cp313-cp313-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:9a75a4704ff640e46170042eec1f984385a121227c505d5a16ad8e495f452541"},
+    {file = "coverage-7.16.2-cp313-cp313-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:14253fc7bb15749b849795a06f5d3b6d8bc3fb8a4b5ddc341faf7a89dce205fc"},
+    {file = "coverage-7.16.2-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:921415102a90637fcc2e3f169f61dad7699ecf690e8639fc21b813acbedc0967"},
+    {file = "coverage-7.16.2-cp313-cp313-musllinux_1_2_i686.whl", hash = "sha256:cce2bc991293f15cc4084ca116827b5900c5f34e1a54dfe83f10ab5c43162eb7"},
+    {file = "coverage-7.16.2-cp313-cp313-musllinux_1_2_ppc64le.whl", hash = "sha256:e1fa594c887365b69745f25a416806e61085dd07b94c9eae68a6e20730629b23"},
+    {file = "coverage-7.16.2-cp313-cp313-musllinux_1_2_riscv64.whl", hash = "sha256:11e597173af1dc33d5f8a7332ada544199269a223af1ee1770ddd5e245ad0fe8"},
+    {file = "coverage-7.16.2-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:3e7f99698ba3a7d13988bdd984b7ebf13af4dbe2166dc8502eef90d77603b0a4"},
+    {file = "coverage-7.16.2-cp313-cp313-win32.whl", hash = "sha256:f80bd9f9633eafc73d0a913ba2645c96ba58bba1befc30590f7c0fbfde59d865"},
+    {file = "coverage-7.16.2-cp313-cp313-win_amd64.whl", hash = "sha256:8be099e979fc42559328a21828281b4578304191ae46ed4e80a407048a82eee6"},
+    {file = "coverage-7.16.2-cp313-cp313-win_arm64.whl", hash = "sha256:28ff850182a67d117990fa2ce5ea1032836d8c9630dae867e8bdd3bff4533b79"},
+    {file = "coverage-7.16.2-cp314-cp314-macosx_10_15_x86_64.whl", hash = "sha256:4ee546b9e4872ffa194bf07ac87bfa1202ebb824d0795dc1ef22f175545ca90a"},
+    {file = "coverage-7.16.2-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:a2fac6895eb299a2e52d7bbb8fb3903502b9da8d3f5309ceb16ec40c646b58ee"},
+    {file = "coverage-7.16.2-cp314-cp314-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:57ff3783f99d75a1e81dd56a9737eb5665e6736a5d93258ba596b6dcad8fd05b"},
+    {file = "coverage-7.16.2-cp314-cp314-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:35f37886699cb9abd29958247d718628d5bc6f39e623dff66a09e546c42a7e03"},
+    {file = "coverage-7.16.2-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:0fd7a86fdda7cb6d616d178654bd0ad6bc0f3f33c2e478aa598500a1a9e34eda"},
+    {file = "coverage-7.16.2-cp314-cp314-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:ac0f3b379c94acc2f7dce5f5f0b24d44fa1cc6a509717ef83dfee07450c2117c"},
+    {file = "coverage-7.16.2-cp314-cp314-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:7d0732c83746bc24123c581a85d9dd96b70ddb538c9076020aa1a041790361e9"},
+    {file = "coverage-7.16.2-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:7b451c68218c150f616bc9649783ec8de76a59792c759b43aa0c9c0466a465e4"},
+    {file = "coverage-7.16.2-cp314-cp314-musllinux_1_2_i686.whl", hash = "sha256:a56ac4fa5a75c7e182e8f62600cfb4aff43c5ed7356a034f3557659c3bec1d90"},
+    {file = "coverage-7.16.2-cp314-cp314-musllinux_1_2_ppc64le.whl", hash = "sha256:4cc4f73aa3fabc36e32046d6cd2971405948d8a903636508a3d3b2f9128b3a95"},
+    {file = "coverage-7.16.2-cp314-cp314-musllinux_1_2_riscv64.whl", hash = "sha256:723dcdab91357159b722935b500ee8abc0a66c8c432e1e9fabf4cc7598952de8"},
+    {file = "coverage-7.16.2-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:5397e21a90dde0e9c6896b77ded8f0be26b66f8b22b33aed41f6043ed95d55e6"},
+    {file = "coverage-7.16.2-cp314-cp314-win32.whl", hash = "sha256:848893e1d361448c113dc2f0913503522a6f7be231d0e38333d2a22d9698a011"},
+    {file = "coverage-7.16.2-cp314-cp314-win_amd64.whl", hash = "sha256:5a27b731c171e43dc8b5f32b76a5051dde2ec9b9366c87028f08a7088ebc2c7b"},
+    {file = "coverage-7.16.2-cp314-cp314-win_arm64.whl", hash = "sha256:1c569a9fd25505f1cd6bea90588818f90373ce90e2632e2cacf19ddbd6e14fdb"},
+    {file = "coverage-7.16.2-cp314-cp314t-macosx_10_15_x86_64.whl", hash = "sha256:d93db87adb6b1c1b408dce4763314b55d76a9f589e96783a84ac9e7689e48bdf"},
+    {file = "coverage-7.16.2-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:aa62c85046473959c13ba9edca9dc90a77d5c1095b1ba313556314d77fe5b036"},
+    {file = "coverage-7.16.2-cp314-cp314t-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:db76506aa5416081f3e8974ae0f7965c58ada0bb0ef7339ac86099588dbb20d3"},
+    {file = "coverage-7.16.2-cp314-cp314t-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:a0f2285329dac10ab08f79cb11f5692c497018e6c7c511f95e6fd63a70b8f831"},
+    {file = "coverage-7.16.2-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:382d3346d56b0eec1b793d53a4c88799c8053f516aa3a8d7c44315696954bacf"},
+    {file = "coverage-7.16.2-cp314-cp314t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:648352b94507179d82637292e7ae8802508d95f78e2f00a705a50b6c48011681"},
+    {file = "coverage-7.16.2-cp314-cp314t-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:fb2bde05838fffae1a1bf75e5d411a6cac3e4e9bb97e6640fed8cd47888b33f0"},
+    {file = "coverage-7.16.2-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:6a75180829efb8ae62b4aded25be6ddca1c888d138d2d82e21d93bfbd88f41cb"},
+    {file = "coverage-7.16.2-cp314-cp314t-musllinux_1_2_i686.whl", hash = "sha256:99704f73721e23859112072d522076e11c31744fc96b5652e5dd2018aa4359f7"},
+    {file = "coverage-7.16.2-cp314-cp314t-musllinux_1_2_ppc64le.whl", hash = "sha256:29309ccc86b7f33df7db12813c299f215bbbc470ed6292d0bedd63ffae1ebf64"},
+    {file = "coverage-7.16.2-cp314-cp314t-musllinux_1_2_riscv64.whl", hash = "sha256:30c1b65d529e46569899fadca59e4a87c1faf2886923f1307ba61e654d4f3c20"},
+    {file = "coverage-7.16.2-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:dcf4bc2aab4e16b1c4c0c2005918f23a7dd5d7821ddae82caed9e3342dc2fcce"},
+    {file = "coverage-7.16.2-cp314-cp314t-win32.whl", hash = "sha256:a9cd3de0a5bfe7b0e21ee10e1a14e3d61bf52efc88217ab1d95d6ace6970bd46"},
+    {file = "coverage-7.16.2-cp314-cp314t-win_amd64.whl", hash = "sha256:611a44e5229a59d7483ce830160e1a0e85f700562c7a5651c7c63fb8f4eb528c"},
+    {file = "coverage-7.16.2-cp314-cp314t-win_arm64.whl", hash = "sha256:22957cef43ce038641de78ba995de7568d2d6a37c6ddbf7fa0fd7d1ae2344d91"},
+    {file = "coverage-7.16.2-cp315-cp315-macosx_10_15_x86_64.whl", hash = "sha256:414c26dfdb96aac2d570a54e03008f001e32eb2d413705365503648c6bd361d8"},
+    {file = "coverage-7.16.2-cp315-cp315-macosx_11_0_arm64.whl", hash = "sha256:00d3eb96e9988c45f50cccd1f1496571ac5c1f91386ac02c4d55516eeda19a24"},
+    {file = "coverage-7.16.2-cp315-cp315-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:4dbbd1155ca46e6e0b6b89d204428c56ef6a459af21333f365d135a2820e5a09"},
+    {file = "coverage-7.16.2-cp315-cp315-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:8fc15cc8d0d06e873c00ef18e1372d605f9aaf3de27d8c24e50782e75bc8b843"},
+    {file = "coverage-7.16.2-cp315-cp315-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:9c6afdd69218202bc1758c9a14b86b8cf1084f37ed2ca143e567a103772b16d1"},
+    {file = "coverage-7.16.2-cp315-cp315-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:aba5c63b7afdc749cc9eae943d5b868cba2b261a176378fa1c5a30bc8bc89982"},
+    {file = "coverage-7.16.2-cp315-cp315-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:9174f0af24e5eff248b9dbfe76ec5275a3d19d37edbc2810543f12cf97347a34"},
+    {file = "coverage-7.16.2-cp315-cp315-musllinux_1_2_aarch64.whl", hash = "sha256:80e9fdb4c3d926b6ba721d4bf7435bdb869c3527ae7803290361d0ab73db13b6"},
+    {file = "coverage-7.16.2-cp315-cp315-musllinux_1_2_i686.whl", hash = "sha256:7b3bce4a0d05401d70b7d0d5ca783e686bc9d30e81dbd7d980d532609bf809e4"},
+    {file = "coverage-7.16.2-cp315-cp315-musllinux_1_2_ppc64le.whl", hash = "sha256:44f21e407b278efdfc1ee5e481e00518bd1d500310a30a5fbf2bcbedfef4aaf0"},
+    {file = "coverage-7.16.2-cp315-cp315-musllinux_1_2_riscv64.whl", hash = "sha256:59c3926585e1cd1f2190f4b2ac9014de1bbeaf0d5d0587b0dc6b0aa90d17896a"},
+    {file = "coverage-7.16.2-cp315-cp315-musllinux_1_2_x86_64.whl", hash = "sha256:066429634299e14dd2d511e1e85f8f9cecc500781f6b41907c0dd6f1baea7e63"},
+    {file = "coverage-7.16.2-cp315-cp315-win32.whl", hash = "sha256:893ea9cf86cb8d2546812ac93d973aaf2ee1fb45110a873b014214fd23e3725e"},
+    {file = "coverage-7.16.2-cp315-cp315-win_amd64.whl", hash = "sha256:01c6908bc613b420c26c818fe948e1b97dfd041a53c98b01c63bd8321f5c9aae"},
+    {file = "coverage-7.16.2-cp315-cp315-win_arm64.whl", hash = "sha256:967d72c835d7a8cf0af99ec813a2d06e3db6df706402f1fe85b31b437645f495"},
+    {file = "coverage-7.16.2-cp315-cp315t-macosx_10_15_x86_64.whl", hash = "sha256:98d9c97f51b334b0adce7b964442a9af33c1a00c6ac856984cc5dc8d18f81c75"},
+    {file = "coverage-7.16.2-cp315-cp315t-macosx_11_0_arm64.whl", hash = "sha256:3e861f1071dcc2fec1e88bef0920f6b1eaa66a143555b4f8ab79ba2b0f30ef55"},
+    {file = "coverage-7.16.2-cp315-cp315t-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:fb9d92ecfe2d5b494367c67f7446f8b75b68d8d0c8cf3bc3e6997478be25d9e2"},
+    {file = "coverage-7.16.2-cp315-cp315t-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:eb57acff4a74246ae513c142d4b36e18c389c3aed8661914a53f7cd0071031b2"},
+    {file = "coverage-7.16.2-cp315-cp315t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:444889f7f66b74e4455c0a97e0e166dd41177f1dca8c0239a47cff25e05ba7e1"},
+    {file = "coverage-7.16.2-cp315-cp315t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:a740ea6f083c6db7b926534d159508f80ba275ab35e722522de0d18d0f56e55f"},
+    {file = "coverage-7.16.2-cp315-cp315t-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:8e209591f7c41ae4a9171335cf6156afda0b21de73b02f73f5aa95b2d5fbb08d"},
+    {file = "coverage-7.16.2-cp315-cp315t-musllinux_1_2_aarch64.whl", hash = "sha256:396bb16e04ce04efbb3df91456ae4e3da918e69ecdf67fb711b0a0fdf35ccce0"},
+    {file = "coverage-7.16.2-cp315-cp315t-musllinux_1_2_i686.whl", hash = "sha256:9cdf19874e0d247f32f03609200370343c3c7aa260b191d8c2bb251d36198283"},
+    {file = "coverage-7.16.2-cp315-cp315t-musllinux_1_2_ppc64le.whl", hash = "sha256:fd3d72233eb8b48acc94fa57d44e2d32ce8e7abed02882ccb6d855ccc4ed33ec"},
+    {file = "coverage-7.16.2-cp315-cp315t-musllinux_1_2_riscv64.whl", hash = "sha256:bb4ffe96aa663cee727659db5a2afeb38c95f8677b747d447b90d6d4874ea2c5"},
+    {file = "coverage-7.16.2-cp315-cp315t-musllinux_1_2_x86_64.whl", hash = "sha256:dba2edfb054f6d4a08df9d1637c39a5aa3865bca6617c13c86be21e45658a59c"},
+    {file = "coverage-7.16.2-cp315-cp315t-win32.whl", hash = "sha256:251aed777c47c77aba047096d4542889db089227655711dfc2b9c54ef0e15e35"},
+    {file = "coverage-7.16.2-cp315-cp315t-win_amd64.whl", hash = "sha256:2aca0bdfa9e91621d5b09d815357bf63def4fc0e9cb66da67bf2cf93f3b1a6f5"},
+    {file = "coverage-7.16.2-cp315-cp315t-win_arm64.whl", hash = "sha256:b88841e654f09732804809e435b3e005a929ffd9998b872b7b213957b8759cb8"},
+    {file = "coverage-7.16.2-py3-none-any.whl", hash = "sha256:11d28e9123a9156cb405d8d27b44256c9a58fb5decc2073a8f17862057e3aa0f"},
+    {file = "coverage-7.16.2.tar.gz", hash = "sha256:ca64d9f1f384f151b9511bec01126072acd2f313439f8ed015a22d8790aab6fa"},
+]
+
+[package.extras]
+toml = ["tomli ; python_full_version <= \"3.11.0a6\""]
+
+[[package]]
+name = "datasets"
+version = "5.1.0"
+description = "HuggingFace community-driven open-source library of datasets"
+optional = false
+python-versions = ">=3.10.0"
+groups = ["main"]
+files = [
+    {file = "datasets-5.1.0-py3-none-any.whl", hash = "sha256:86fd08b280485aab82894d2c81842c4151d93b575efea0054315104cb7bc3cec"},
+    {file = "datasets-5.1.0.tar.gz", hash = "sha256:1561347daa8841b51351ef2ad1647b4ee4af406852c0f7bd1df587ac1ca7888e"},
+]
+
+[package.dependencies]
+dill = ">=0.3.0,<0.4.2"
+filelock = "*"
+fsspec = {version = ">=2023.1.0,<=2026.7.0", extras = ["http"]}
+huggingface-hub = ">=1.31.0,<3.0"
+multiprocess = "<0.70.20"
+numpy = ">=1.17"
+packaging = "*"
+pandas = "*"
+pyarrow = ">=24.0.0"
+pyyaml = ">=5.1"
+tqdm = ">=4.66.3"
+xxhash = "*"
+
+[package.extras]
+audio = ["torch (>=2.8.0)", "torchcodec (>=0.6.0)"]
+benchmarks = ["tensorflow (==2.12.0)", "torch (==2.0.1)", "transformers (==4.30.1)"]
+bio = ["biopython (>=1.80)"]
+dev = ["Pillow (>=9.4.0)", "absl-py", "aiohttp", "biopython (>=1.80)", "decorator", "elasticsearch (>=7.17.12,<8.0.0)", "faiss-cpu (>=1.8.0.post1)", "h5py", "jax (>=0.3.14) ; sys_platform != \"win32\"", "jaxlib (>=0.3.14) ; sys_platform != \"win32\"", "joblib (<1.3.0)", "joblibspark ; python_version < \"3.14\"", "lz4 ; python_version < \"3.14\"", "moto[server]", "nibabel (>=5.3.1)", "numba (>=0.56.4) ; python_version < \"3.14\"", "polars[timezone] (>=0.20.0)", "protobuf (<4.0.0) ; python_version < \"3.14\"", "py7zr", "pyiceberg[pyarrow,sql-sqlite]", "pylance", "pyspark (>=3.4)", "pytest", "pytest-datadir", "pytest-xdist", "rarfile (>=4.0)", "ruff (>=0.3.0)", "sqlalchemy", "teich (==0.1.5)", "tensorflow (>=2.16.0) ; python_version >= \"3.10\" and sys_platform != \"win32\" and python_version < \"3.14\"", "tensorflow (>=2.6.0)", "tensorflow (>=2.6.0) ; python_version < \"3.10\" and sys_platform != \"win32\"", "tiktoken", "tomli (>=1.1.0) ; python_version < \"3.11\"", "torch", "torch (>=2.8.0)", "torchcodec (>=0.7.0) ; python_version < \"3.14\"", "torchdata", "transformers", "transformers (>=4.42.0)", "trimesh (>=4.10.0)", "vortex-data ; python_version >= \"3.11\" and sys_platform != \"win32\"", "zstandard"]
+docs = ["tensorflow (>=2.6.0)", "torch", "transformers"]
+iceberg = ["pyiceberg (>=0.7.0)"]
+jax = ["jax (>=0.3.14)", "jaxlib (>=0.3.14)"]
+mesh = ["trimesh (>=4.10.0)"]
+nibabel = ["ipyniivue (==2.4.2)", "nibabel (>=5.3.2)"]
+pdfs = ["pdfplumber (>=0.11.4)"]
+quality = ["ruff (>=0.3.0)"]
+tensorflow = ["tensorflow (>=2.6.0)"]
+tensorflow-gpu = ["tensorflow (>=2.6.0)"]
+tests = ["Pillow (>=9.4.0)", "absl-py", "aiohttp", "biopython (>=1.80)", "decorator", "elasticsearch (>=7.17.12,<8.0.0)", "faiss-cpu (>=1.8.0.post1)", "h5py", "jax (>=0.3.14) ; sys_platform != \"win32\"", "jaxlib (>=0.3.14) ; sys_platform != \"win32\"", "joblib (<1.3.0)", "joblibspark ; python_version < \"3.14\"", "lz4 ; python_version < \"3.14\"", "moto[server]", "nibabel (>=5.3.1)", "numba (>=0.56.4) ; python_version < \"3.14\"", "polars[timezone] (>=0.20.0)", "protobuf (<4.0.0) ; python_version < \"3.14\"", "py7zr", "pyiceberg[pyarrow,sql-sqlite]", "pylance", "pyspark (>=3.4)", "pytest", "pytest-datadir", "pytest-xdist", "rarfile (>=4.0)", "sqlalchemy", "teich (==0.1.5)", "tensorflow (>=2.16.0) ; python_version >= \"3.10\" and sys_platform != \"win32\" and python_version < \"3.14\"", "tensorflow (>=2.6.0) ; python_version < \"3.10\" and sys_platform != \"win32\"", "tiktoken", "tomli (>=1.1.0) ; python_version < \"3.11\"", "torch (>=2.8.0)", "torchcodec (>=0.7.0) ; python_version < \"3.14\"", "torchdata", "transformers (>=4.42.0)", "trimesh (>=4.10.0)", "vortex-data ; python_version >= \"3.11\" and sys_platform != \"win32\"", "zstandard"]
+tests-numpy2 = ["Pillow (>=9.4.0)", "absl-py", "aiohttp", "biopython (>=1.80)", "decorator", "elasticsearch (>=7.17.12,<8.0.0)", "h5py", "jax (>=0.3.14) ; sys_platform != \"win32\"", "jaxlib (>=0.3.14) ; sys_platform != \"win32\"", "joblib (<1.3.0)", "joblibspark ; python_version < \"3.14\"", "lz4 ; python_version < \"3.14\"", "moto[server]", "nibabel (>=5.3.1)", "numba (>=0.56.4) ; python_version < \"3.14\"", "polars[timezone] (>=0.20.0)", "protobuf (<4.0.0) ; python_version < \"3.14\"", "py7zr", "pyiceberg[pyarrow,sql-sqlite]", "pylance", "pyspark (>=3.4)", "pytest", "pytest-datadir", "pytest-xdist", "rarfile (>=4.0)", "sqlalchemy", "teich (==0.1.5)", "tiktoken", "tomli (>=1.1.0) ; python_version < \"3.11\"", "torch (>=2.8.0)", "torchcodec (>=0.7.0) ; python_version < \"3.14\"", "torchdata", "transformers (>=4.42.0)", "trimesh (>=4.10.0)", "vortex-data ; python_version >= \"3.11\" and sys_platform != \"win32\"", "zstandard"]
+torch = ["torch"]
+vision = ["Pillow (>=9.4.0)"]
+
+[[package]]
+name = "dill"
+version = "0.4.1"
+description = "serialize all of Python"
+optional = false
+python-versions = ">=3.9"
+groups = ["main"]
+files = [
+    {file = "dill-0.4.1-py3-none-any.whl", hash = "sha256:1e1ce33e978ae97fcfcff5638477032b801c46c7c65cf717f95fbc2248f79a9d"},
+    {file = "dill-0.4.1.tar.gz", hash = "sha256:423092df4182177d4d8ba8290c8a5b640c66ab35ec7da59ccfa00f6fa3eea5fa"},
+]
+
+[package.extras]
+graph = ["objgraph (>=1.7.2)"]
+profile = ["gprof2dot (>=2022.7.29)"]
+
+[[package]]
+name = "filelock"
+version = "4.0.12"
+description = "A platform independent file lock."
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "filelock-4.0.12-py3-none-any.whl", hash = "sha256:5f17ee83ecee8a6f3e389c75822fb70a1c2f0506b99438a6dffa1d56793588c8"},
+    {file = "filelock-4.0.12.tar.gz", hash = "sha256:cf42711a7ac791818b299fab0332a088c65aeeefa36290de98db92c434303b0c"},
+]
+
+[[package]]
+name = "frozenlist"
+version = "1.8.0"
+description = "A list-like structure which implements collections.abc.MutableSequence"
+optional = false
+python-versions = ">=3.9"
+groups = ["main"]
+files = [
+    {file = "frozenlist-1.8.0-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:b37f6d31b3dcea7deb5e9696e529a6aa4a898adc33db82da12e4c60a7c4d2011"},
+    {file = "frozenlist-1.8.0-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:ef2b7b394f208233e471abc541cc6991f907ffd47dc72584acee3147899d6565"},
+    {file = "frozenlist-1.8.0-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:a88f062f072d1589b7b46e951698950e7da00442fc1cacbe17e19e025dc327ad"},
+    {file = "frozenlist-1.8.0-cp310-cp310-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:f57fb59d9f385710aa7060e89410aeb5058b99e62f4d16b08b91986b9a2140c2"},
+    {file = "frozenlist-1.8.0-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:799345ab092bee59f01a915620b5d014698547afd011e691a208637312db9186"},
+    {file = "frozenlist-1.8.0-cp310-cp310-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:c23c3ff005322a6e16f71bf8692fcf4d5a304aaafe1e262c98c6d4adc7be863e"},
+    {file = "frozenlist-1.8.0-cp310-cp310-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:8a76ea0f0b9dfa06f254ee06053d93a600865b3274358ca48a352ce4f0798450"},
+    {file = "frozenlist-1.8.0-cp310-cp310-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:c7366fe1418a6133d5aa824ee53d406550110984de7637d65a178010f759c6ef"},
+    {file = "frozenlist-1.8.0-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:13d23a45c4cebade99340c4165bd90eeb4a56c6d8a9d8aa49568cac19a6d0dc4"},
+    {file = "frozenlist-1.8.0-cp310-cp310-musllinux_1_2_armv7l.whl", hash = "sha256:e4a3408834f65da56c83528fb52ce7911484f0d1eaf7b761fc66001db1646eff"},
+    {file = "frozenlist-1.8.0-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:42145cd2748ca39f32801dad54aeea10039da6f86e303659db90db1c4b614c8c"},
+    {file = "frozenlist-1.8.0-cp310-cp310-musllinux_1_2_s390x.whl", hash = "sha256:e2de870d16a7a53901e41b64ffdf26f2fbb8917b3e6ebf398098d72c5b20bd7f"},
+    {file = "frozenlist-1.8.0-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:20e63c9493d33ee48536600d1a5c95eefc870cd71e7ab037763d1fbb89cc51e7"},
+    {file = "frozenlist-1.8.0-cp310-cp310-win32.whl", hash = "sha256:adbeebaebae3526afc3c96fad434367cafbfd1b25d72369a9e5858453b1bb71a"},
+    {file = "frozenlist-1.8.0-cp310-cp310-win_amd64.whl", hash = "sha256:667c3777ca571e5dbeb76f331562ff98b957431df140b54c85fd4d52eea8d8f6"},
+    {file = "frozenlist-1.8.0-cp310-cp310-win_arm64.whl", hash = "sha256:80f85f0a7cc86e7a54c46d99c9e1318ff01f4687c172ede30fd52d19d1da1c8e"},
+    {file = "frozenlist-1.8.0-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:09474e9831bc2b2199fad6da3c14c7b0fbdd377cce9d3d77131be28906cb7d84"},
+    {file = "frozenlist-1.8.0-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:17c883ab0ab67200b5f964d2b9ed6b00971917d5d8a92df149dc2c9779208ee9"},
+    {file = "frozenlist-1.8.0-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:fa47e444b8ba08fffd1c18e8cdb9a75db1b6a27f17507522834ad13ed5922b93"},
+    {file = "frozenlist-1.8.0-cp311-cp311-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:2552f44204b744fba866e573be4c1f9048d6a324dfe14475103fd51613eb1d1f"},
+    {file = "frozenlist-1.8.0-cp311-cp311-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:957e7c38f250991e48a9a73e6423db1bb9dd14e722a10f6b8bb8e16a0f55f695"},
+    {file = "frozenlist-1.8.0-cp311-cp311-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:8585e3bb2cdea02fc88ffa245069c36555557ad3609e83be0ec71f54fd4abb52"},
+    {file = "frozenlist-1.8.0-cp311-cp311-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:edee74874ce20a373d62dc28b0b18b93f645633c2943fd90ee9d898550770581"},
+    {file = "frozenlist-1.8.0-cp311-cp311-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:c9a63152fe95756b85f31186bddf42e4c02c6321207fd6601a1c89ebac4fe567"},
+    {file = "frozenlist-1.8.0-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:b6db2185db9be0a04fecf2f241c70b63b1a242e2805be291855078f2b404dd6b"},
+    {file = "frozenlist-1.8.0-cp311-cp311-musllinux_1_2_armv7l.whl", hash = "sha256:f4be2e3d8bc8aabd566f8d5b8ba7ecc09249d74ba3c9ed52e54dc23a293f0b92"},
+    {file = "frozenlist-1.8.0-cp311-cp311-musllinux_1_2_ppc64le.whl", hash = "sha256:c8d1634419f39ea6f5c427ea2f90ca85126b54b50837f31497f3bf38266e853d"},
+    {file = "frozenlist-1.8.0-cp311-cp311-musllinux_1_2_s390x.whl", hash = "sha256:1a7fa382a4a223773ed64242dbe1c9c326ec09457e6b8428efb4118c685c3dfd"},
+    {file = "frozenlist-1.8.0-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:11847b53d722050808926e785df837353bd4d75f1d494377e59b23594d834967"},
+    {file = "frozenlist-1.8.0-cp311-cp311-win32.whl", hash = "sha256:27c6e8077956cf73eadd514be8fb04d77fc946a7fe9f7fe167648b0b9085cc25"},
+    {file = "frozenlist-1.8.0-cp311-cp311-win_amd64.whl", hash = "sha256:ac913f8403b36a2c8610bbfd25b8013488533e71e62b4b4adce9c86c8cea905b"},
+    {file = "frozenlist-1.8.0-cp311-cp311-win_arm64.whl", hash = "sha256:d4d3214a0f8394edfa3e303136d0575eece0745ff2b47bd2cb2e66dd92d4351a"},
+    {file = "frozenlist-1.8.0-cp312-cp312-macosx_10_13_universal2.whl", hash = "sha256:78f7b9e5d6f2fdb88cdde9440dc147259b62b9d3b019924def9f6478be254ac1"},
+    {file = "frozenlist-1.8.0-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:229bf37d2e4acdaf808fd3f06e854a4a7a3661e871b10dc1f8f1896a3b05f18b"},
+    {file = "frozenlist-1.8.0-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:f833670942247a14eafbb675458b4e61c82e002a148f49e68257b79296e865c4"},
+    {file = "frozenlist-1.8.0-cp312-cp312-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:494a5952b1c597ba44e0e78113a7266e656b9794eec897b19ead706bd7074383"},
+    {file = "frozenlist-1.8.0-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:96f423a119f4777a4a056b66ce11527366a8bb92f54e541ade21f2374433f6d4"},
+    {file = "frozenlist-1.8.0-cp312-cp312-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:3462dd9475af2025c31cc61be6652dfa25cbfb56cbbf52f4ccfe029f38decaf8"},
+    {file = "frozenlist-1.8.0-cp312-cp312-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:c4c800524c9cd9bac5166cd6f55285957fcfc907db323e193f2afcd4d9abd69b"},
+    {file = "frozenlist-1.8.0-cp312-cp312-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:d6a5df73acd3399d893dafc71663ad22534b5aa4f94e8a2fabfe856c3c1b6a52"},
+    {file = "frozenlist-1.8.0-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:405e8fe955c2280ce66428b3ca55e12b3c4e9c336fb2103a4937e891c69a4a29"},
+    {file = "frozenlist-1.8.0-cp312-cp312-musllinux_1_2_armv7l.whl", hash = "sha256:908bd3f6439f2fef9e85031b59fd4f1297af54415fb60e4254a95f75b3cab3f3"},
+    {file = "frozenlist-1.8.0-cp312-cp312-musllinux_1_2_ppc64le.whl", hash = "sha256:294e487f9ec720bd8ffcebc99d575f7eff3568a08a253d1ee1a0378754b74143"},
+    {file = "frozenlist-1.8.0-cp312-cp312-musllinux_1_2_s390x.whl", hash = "sha256:74c51543498289c0c43656701be6b077f4b265868fa7f8a8859c197006efb608"},
+    {file = "frozenlist-1.8.0-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:776f352e8329135506a1d6bf16ac3f87bc25b28e765949282dcc627af36123aa"},
+    {file = "frozenlist-1.8.0-cp312-cp312-win32.whl", hash = "sha256:433403ae80709741ce34038da08511d4a77062aa924baf411ef73d1146e74faf"},
+    {file = "frozenlist-1.8.0-cp312-cp312-win_amd64.whl", hash = "sha256:34187385b08f866104f0c0617404c8eb08165ab1272e884abc89c112e9c00746"},
+    {file = "frozenlist-1.8.0-cp312-cp312-win_arm64.whl", hash = "sha256:fe3c58d2f5db5fbd18c2987cba06d51b0529f52bc3a6cdc33d3f4eab725104bd"},
+    {file = "frozenlist-1.8.0-cp313-cp313-macosx_10_13_universal2.whl", hash = "sha256:8d92f1a84bb12d9e56f818b3a746f3efba93c1b63c8387a73dde655e1e42282a"},
+    {file = "frozenlist-1.8.0-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:96153e77a591c8adc2ee805756c61f59fef4cf4073a9275ee86fe8cba41241f7"},
+    {file = "frozenlist-1.8.0-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:f21f00a91358803399890ab167098c131ec2ddd5f8f5fd5fe9c9f2c6fcd91e40"},
+    {file = "frozenlist-1.8.0-cp313-cp313-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:fb30f9626572a76dfe4293c7194a09fb1fe93ba94c7d4f720dfae3b646b45027"},
+    {file = "frozenlist-1.8.0-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:eaa352d7047a31d87dafcacbabe89df0aa506abb5b1b85a2fb91bc3faa02d822"},
+    {file = "frozenlist-1.8.0-cp313-cp313-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:03ae967b4e297f58f8c774c7eabcce57fe3c2434817d4385c50661845a058121"},
+    {file = "frozenlist-1.8.0-cp313-cp313-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:f6292f1de555ffcc675941d65fffffb0a5bcd992905015f85d0592201793e0e5"},
+    {file = "frozenlist-1.8.0-cp313-cp313-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:29548f9b5b5e3460ce7378144c3010363d8035cea44bc0bf02d57f5a685e084e"},
+    {file = "frozenlist-1.8.0-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:ec3cc8c5d4084591b4237c0a272cc4f50a5b03396a47d9caaf76f5d7b38a4f11"},
+    {file = "frozenlist-1.8.0-cp313-cp313-musllinux_1_2_armv7l.whl", hash = "sha256:517279f58009d0b1f2e7c1b130b377a349405da3f7621ed6bfae50b10adf20c1"},
+    {file = "frozenlist-1.8.0-cp313-cp313-musllinux_1_2_ppc64le.whl", hash = "sha256:db1e72ede2d0d7ccb213f218df6a078a9c09a7de257c2fe8fcef16d5925230b1"},
+    {file = "frozenlist-1.8.0-cp313-cp313-musllinux_1_2_s390x.whl", hash = "sha256:b4dec9482a65c54a5044486847b8a66bf10c9cb4926d42927ec4e8fd5db7fed8"},
+    {file = "frozenlist-1.8.0-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:21900c48ae04d13d416f0e1e0c4d81f7931f73a9dfa0b7a8746fb2fe7dd970ed"},
+    {file = "frozenlist-1.8.0-cp313-cp313-win32.whl", hash = "sha256:8b7b94a067d1c504ee0b16def57ad5738701e4ba10cec90529f13fa03c833496"},
+    {file = "frozenlist-1.8.0-cp313-cp313-win_amd64.whl", hash = "sha256:878be833caa6a3821caf85eb39c5ba92d28e85df26d57afb06b35b2efd937231"},
+    {file = "frozenlist-1.8.0-cp313-cp313-win_arm64.whl", hash = "sha256:44389d135b3ff43ba8cc89ff7f51f5a0bb6b63d829c8300f79a2fe4fe61bcc62"},
+    {file = "frozenlist-1.8.0-cp313-cp313t-macosx_10_13_universal2.whl", hash = "sha256:e25ac20a2ef37e91c1b39938b591457666a0fa835c7783c3a8f33ea42870db94"},
+    {file = "frozenlist-1.8.0-cp313-cp313t-macosx_10_13_x86_64.whl", hash = "sha256:07cdca25a91a4386d2e76ad992916a85038a9b97561bf7a3fd12d5d9ce31870c"},
+    {file = "frozenlist-1.8.0-cp313-cp313t-macosx_11_0_arm64.whl", hash = "sha256:4e0c11f2cc6717e0a741f84a527c52616140741cd812a50422f83dc31749fb52"},
+    {file = "frozenlist-1.8.0-cp313-cp313t-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:b3210649ee28062ea6099cfda39e147fa1bc039583c8ee4481cb7811e2448c51"},
+    {file = "frozenlist-1.8.0-cp313-cp313t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:581ef5194c48035a7de2aefc72ac6539823bb71508189e5de01d60c9dcd5fa65"},
+    {file = "frozenlist-1.8.0-cp313-cp313t-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:3ef2d026f16a2b1866e1d86fc4e1291e1ed8a387b2c333809419a2f8b3a77b82"},
+    {file = "frozenlist-1.8.0-cp313-cp313t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:5500ef82073f599ac84d888e3a8c1f77ac831183244bfd7f11eaa0289fb30714"},
+    {file = "frozenlist-1.8.0-cp313-cp313t-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:50066c3997d0091c411a66e710f4e11752251e6d2d73d70d8d5d4c76442a199d"},
+    {file = "frozenlist-1.8.0-cp313-cp313t-musllinux_1_2_aarch64.whl", hash = "sha256:5c1c8e78426e59b3f8005e9b19f6ff46e5845895adbde20ece9218319eca6506"},
+    {file = "frozenlist-1.8.0-cp313-cp313t-musllinux_1_2_armv7l.whl", hash = "sha256:eefdba20de0d938cec6a89bd4d70f346a03108a19b9df4248d3cf0d88f1b0f51"},
+    {file = "frozenlist-1.8.0-cp313-cp313t-musllinux_1_2_ppc64le.whl", hash = "sha256:cf253e0e1c3ceb4aaff6df637ce033ff6535fb8c70a764a8f46aafd3d6ab798e"},
+    {file = "frozenlist-1.8.0-cp313-cp313t-musllinux_1_2_s390x.whl", hash = "sha256:032efa2674356903cd0261c4317a561a6850f3ac864a63fc1583147fb05a79b0"},
+    {file = "frozenlist-1.8.0-cp313-cp313t-musllinux_1_2_x86_64.whl", hash = "sha256:6da155091429aeba16851ecb10a9104a108bcd32f6c1642867eadaee401c1c41"},
+    {file = "frozenlist-1.8.0-cp313-cp313t-win32.whl", hash = "sha256:0f96534f8bfebc1a394209427d0f8a63d343c9779cda6fc25e8e121b5fd8555b"},
+    {file = "frozenlist-1.8.0-cp313-cp313t-win_amd64.whl", hash = "sha256:5d63a068f978fc69421fb0e6eb91a9603187527c86b7cd3f534a5b77a592b888"},
+    {file = "frozenlist-1.8.0-cp313-cp313t-win_arm64.whl", hash = "sha256:bf0a7e10b077bf5fb9380ad3ae8ce20ef919a6ad93b4552896419ac7e1d8e042"},
+    {file = "frozenlist-1.8.0-cp314-cp314-macosx_10_13_universal2.whl", hash = "sha256:cee686f1f4cadeb2136007ddedd0aaf928ab95216e7691c63e50a8ec066336d0"},
+    {file = "frozenlist-1.8.0-cp314-cp314-macosx_10_13_x86_64.whl", hash = "sha256:119fb2a1bd47307e899c2fac7f28e85b9a543864df47aa7ec9d3c1b4545f096f"},
+    {file = "frozenlist-1.8.0-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:4970ece02dbc8c3a92fcc5228e36a3e933a01a999f7094ff7c23fbd2beeaa67c"},
+    {file = "frozenlist-1.8.0-cp314-cp314-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:cba69cb73723c3f329622e34bdbf5ce1f80c21c290ff04256cff1cd3c2036ed2"},
+    {file = "frozenlist-1.8.0-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:778a11b15673f6f1df23d9586f83c4846c471a8af693a22e066508b77d201ec8"},
+    {file = "frozenlist-1.8.0-cp314-cp314-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:0325024fe97f94c41c08872db482cf8ac4800d80e79222c6b0b7b162d5b13686"},
+    {file = "frozenlist-1.8.0-cp314-cp314-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:97260ff46b207a82a7567b581ab4190bd4dfa09f4db8a8b49d1a958f6aa4940e"},
+    {file = "frozenlist-1.8.0-cp314-cp314-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:54b2077180eb7f83dd52c40b2750d0a9f175e06a42e3213ce047219de902717a"},
+    {file = "frozenlist-1.8.0-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:2f05983daecab868a31e1da44462873306d3cbfd76d1f0b5b69c473d21dbb128"},
+    {file = "frozenlist-1.8.0-cp314-cp314-musllinux_1_2_armv7l.whl", hash = "sha256:33f48f51a446114bc5d251fb2954ab0164d5be02ad3382abcbfe07e2531d650f"},
+    {file = "frozenlist-1.8.0-cp314-cp314-musllinux_1_2_ppc64le.whl", hash = "sha256:154e55ec0655291b5dd1b8731c637ecdb50975a2ae70c606d100750a540082f7"},
+    {file = "frozenlist-1.8.0-cp314-cp314-musllinux_1_2_s390x.whl", hash = "sha256:4314debad13beb564b708b4a496020e5306c7333fa9a3ab90374169a20ffab30"},
+    {file = "frozenlist-1.8.0-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:073f8bf8becba60aa931eb3bc420b217bb7d5b8f4750e6f8b3be7f3da85d38b7"},
+    {file = "frozenlist-1.8.0-cp314-cp314-win32.whl", hash = "sha256:bac9c42ba2ac65ddc115d930c78d24ab8d4f465fd3fc473cdedfccadb9429806"},
+    {file = "frozenlist-1.8.0-cp314-cp314-win_amd64.whl", hash = "sha256:3e0761f4d1a44f1d1a47996511752cf3dcec5bbdd9cc2b4fe595caf97754b7a0"},
+    {file = "frozenlist-1.8.0-cp314-cp314-win_arm64.whl", hash = "sha256:d1eaff1d00c7751b7c6662e9c5ba6eb2c17a2306ba5e2a37f24ddf3cc953402b"},
+    {file = "frozenlist-1.8.0-cp314-cp314t-macosx_10_13_universal2.whl", hash = "sha256:d3bb933317c52d7ea5004a1c442eef86f426886fba134ef8cf4226ea6ee1821d"},
+    {file = "frozenlist-1.8.0-cp314-cp314t-macosx_10_13_x86_64.whl", hash = "sha256:8009897cdef112072f93a0efdce29cd819e717fd2f649ee3016efd3cd885a7ed"},
+    {file = "frozenlist-1.8.0-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:2c5dcbbc55383e5883246d11fd179782a9d07a986c40f49abe89ddf865913930"},
+    {file = "frozenlist-1.8.0-cp314-cp314t-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:39ecbc32f1390387d2aa4f5a995e465e9e2f79ba3adcac92d68e3e0afae6657c"},
+    {file = "frozenlist-1.8.0-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:92db2bf818d5cc8d9c1f1fc56b897662e24ea5adb36ad1f1d82875bd64e03c24"},
+    {file = "frozenlist-1.8.0-cp314-cp314t-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:2dc43a022e555de94c3b68a4ef0b11c4f747d12c024a520c7101709a2144fb37"},
+    {file = "frozenlist-1.8.0-cp314-cp314t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:cb89a7f2de3602cfed448095bab3f178399646ab7c61454315089787df07733a"},
+    {file = "frozenlist-1.8.0-cp314-cp314t-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:33139dc858c580ea50e7e60a1b0ea003efa1fd42e6ec7fdbad78fff65fad2fd2"},
+    {file = "frozenlist-1.8.0-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:168c0969a329b416119507ba30b9ea13688fafffac1b7822802537569a1cb0ef"},
+    {file = "frozenlist-1.8.0-cp314-cp314t-musllinux_1_2_armv7l.whl", hash = "sha256:28bd570e8e189d7f7b001966435f9dac6718324b5be2990ac496cf1ea9ddb7fe"},
+    {file = "frozenlist-1.8.0-cp314-cp314t-musllinux_1_2_ppc64le.whl", hash = "sha256:b2a095d45c5d46e5e79ba1e5b9cb787f541a8dee0433836cea4b96a2c439dcd8"},
+    {file = "frozenlist-1.8.0-cp314-cp314t-musllinux_1_2_s390x.whl", hash = "sha256:eab8145831a0d56ec9c4139b6c3e594c7a83c2c8be25d5bcf2d86136a532287a"},
+    {file = "frozenlist-1.8.0-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:974b28cf63cc99dfb2188d8d222bc6843656188164848c4f679e63dae4b0708e"},
+    {file = "frozenlist-1.8.0-cp314-cp314t-win32.whl", hash = "sha256:342c97bf697ac5480c0a7ec73cd700ecfa5a8a40ac923bd035484616efecc2df"},
+    {file = "frozenlist-1.8.0-cp314-cp314t-win_amd64.whl", hash = "sha256:06be8f67f39c8b1dc671f5d83aaefd3358ae5cdcf8314552c57e7ed3e6475bdd"},
+    {file = "frozenlist-1.8.0-cp314-cp314t-win_arm64.whl", hash = "sha256:102e6314ca4da683dca92e3b1355490fed5f313b768500084fbe6371fddfdb79"},
+    {file = "frozenlist-1.8.0-cp39-cp39-macosx_10_9_universal2.whl", hash = "sha256:d8b7138e5cd0647e4523d6685b0eac5d4be9a184ae9634492f25c6eb38c12a47"},
+    {file = "frozenlist-1.8.0-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:a6483e309ca809f1efd154b4d37dc6d9f61037d6c6a81c2dc7a15cb22c8c5dca"},
+    {file = "frozenlist-1.8.0-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:1b9290cf81e95e93fdf90548ce9d3c1211cf574b8e3f4b3b7cb0537cf2227068"},
+    {file = "frozenlist-1.8.0-cp39-cp39-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:59a6a5876ca59d1b63af8cd5e7ffffb024c3dc1e9cf9301b21a2e76286505c95"},
+    {file = "frozenlist-1.8.0-cp39-cp39-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:6dc4126390929823e2d2d9dc79ab4046ed74680360fc5f38b585c12c66cdf459"},
+    {file = "frozenlist-1.8.0-cp39-cp39-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:332db6b2563333c5671fecacd085141b5800cb866be16d5e3eb15a2086476675"},
+    {file = "frozenlist-1.8.0-cp39-cp39-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:9ff15928d62a0b80bb875655c39bf517938c7d589554cbd2669be42d97c2cb61"},
+    {file = "frozenlist-1.8.0-cp39-cp39-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:7bf6cdf8e07c8151fba6fe85735441240ec7f619f935a5205953d58009aef8c6"},
+    {file = "frozenlist-1.8.0-cp39-cp39-musllinux_1_2_aarch64.whl", hash = "sha256:48e6d3f4ec5c7273dfe83ff27c91083c6c9065af655dc2684d2c200c94308bb5"},
+    {file = "frozenlist-1.8.0-cp39-cp39-musllinux_1_2_armv7l.whl", hash = "sha256:1a7607e17ad33361677adcd1443edf6f5da0ce5e5377b798fba20fae194825f3"},
+    {file = "frozenlist-1.8.0-cp39-cp39-musllinux_1_2_ppc64le.whl", hash = "sha256:5a3a935c3a4e89c733303a2d5a7c257ea44af3a56c8202df486b7f5de40f37e1"},
+    {file = "frozenlist-1.8.0-cp39-cp39-musllinux_1_2_s390x.whl", hash = "sha256:940d4a017dbfed9daf46a3b086e1d2167e7012ee297fef9e1c545c4d022f5178"},
+    {file = "frozenlist-1.8.0-cp39-cp39-musllinux_1_2_x86_64.whl", hash = "sha256:b9be22a69a014bc47e78072d0ecae716f5eb56c15238acca0f43d6eb8e4a5bda"},
+    {file = "frozenlist-1.8.0-cp39-cp39-win32.whl", hash = "sha256:1aa77cb5697069af47472e39612976ed05343ff2e84a3dcf15437b232cbfd087"},
+    {file = "frozenlist-1.8.0-cp39-cp39-win_amd64.whl", hash = "sha256:7398c222d1d405e796970320036b1b563892b65809d9e5261487bb2c7f7b5c6a"},
+    {file = "frozenlist-1.8.0-cp39-cp39-win_arm64.whl", hash = "sha256:b4f3b365f31c6cd4af24545ca0a244a53688cad8834e32f56831c4923b50a103"},
+    {file = "frozenlist-1.8.0-py3-none-any.whl", hash = "sha256:0c18a16eab41e82c295618a77502e17b195883241c563b00f0aa5106fc4eaa0d"},
+    {file = "frozenlist-1.8.0.tar.gz", hash = "sha256:3ede829ed8d842f6cd48fc7081d7a41001a56f1f38603f9d49bf3020d59a31ad"},
+]
+
+[[package]]
+name = "fsspec"
+version = "2026.7.0"
+description = "File-system specification"
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "fsspec-2026.7.0-py3-none-any.whl", hash = "sha256:b57ddbafedfaef7018c1ecab32aa200a9d7ca26b77965f64e48b70061249d279"},
+    {file = "fsspec-2026.7.0.tar.gz", hash = "sha256:c803c40f4cf860b49dea58ee3e1c33cb9c790520e233537e1340049f89b82a88"},
+]
+
+[package.dependencies]
+aiohttp = {version = "<4.0.0a0 || >4.0.0a0,<4.0.0a1 || >4.0.0a1", optional = true, markers = "extra == \"http\""}
+
+[package.extras]
+abfs = ["adlfs"]
+adl = ["adlfs"]
+arrow = ["pyarrow (>=1)"]
+dask = ["dask", "distributed"]
+dev = ["pre-commit", "ruff (>=0.5)"]
+doc = ["numpydoc", "sphinx", "sphinx-design", "sphinx-rtd-theme", "yarl"]
+dropbox = ["dropbox", "dropboxdrivefs", "requests"]
+full = ["adlfs", "aiohttp (!=4.0.0a0,!=4.0.0a1)", "dask", "distributed", "dropbox", "dropboxdrivefs", "fusepy", "gcsfs (>=2026.4.0)", "libarchive-c", "ocifs", "panel", "paramiko", "pyarrow (>=1)", "pygit2", "requests", "s3fs (>=2026.6.0)", "smbprotocol", "tqdm"]
+fuse = ["fusepy"]
+gcs = ["gcsfs (>=2026.4.0)"]
+git = ["pygit2"]
+github = ["requests"]
+gs = ["gcsfs (>=2026.4.0)"]
+gui = ["panel"]
+hdfs = ["pyarrow (>=1)"]
+http = ["aiohttp (!=4.0.0a0,!=4.0.0a1)"]
+libarchive = ["libarchive-c"]
+oci = ["ocifs"]
+s3 = ["s3fs (>=2026.6.0)"]
+sftp = ["paramiko"]
+smb = ["smbprotocol"]
+ssh = ["paramiko"]
+test = ["aiohttp (!=4.0.0a0,!=4.0.0a1)", "numpy", "pytest", "pytest-asyncio (!=0.22.0)", "pytest-benchmark", "pytest-cov", "pytest-mock", "pytest-recording", "pytest-rerunfailures", "requests"]
+test-downstream = ["aiobotocore (>=2.5.4,<3.0.0)", "dask[dataframe,test]", "moto[server] (>4,<5)", "pytest-timeout", "xarray"]
+test-full = ["adlfs", "aiohttp (!=4.0.0a0,!=4.0.0a1)", "backports-zstd ; python_version < \"3.14\"", "cloudpickle", "dask", "distributed", "dropbox", "dropboxdrivefs", "fastparquet", "fusepy", "gcsfs (>=2026.4.0)", "jinja2", "kerchunk", "libarchive-c", "lz4", "notebook", "numpy", "ocifs", "pandas (<3.0.0)", "panel", "paramiko", "pyarrow (>=1)", "pyftpdlib", "pygit2", "pytest", "pytest-asyncio (!=0.22.0)", "pytest-benchmark", "pytest-cov", "pytest-mock", "pytest-recording", "pytest-rerunfailures", "python-snappy", "requests", "s3fs (>=2026.6.0)", "smbprotocol", "tqdm", "urllib3", "zarr (<3.2.0)", "zstandard ; python_version < \"3.14\""]
+tqdm = ["tqdm"]
+
+[[package]]
+name = "gguf"
+version = "0.19.0"
+description = "Read and write ML models in GGUF for GGML"
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = []
+develop = false
+
+[package.dependencies]
+numpy = ">=2.2.6"
+pyyaml = ">=5.1"
+requests = ">=2.25"
+tqdm = ">=4.27"
+
+[package.extras]
+gui = ["PySide6 (>=6.9,<7.0) ; python_version >= \"3.9\" and python_version < \"3.14\""]
+
+[package.source]
+type = "git"
+url = "https://github.com/ggml-org/llama.cpp.git"
+reference = "3f03b17724ad8b69040825f45c43c59b092a0244"
+resolved_reference = "3f03b17724ad8b69040825f45c43c59b092a0244"
+subdirectory = "gguf-py"
+
+[[package]]
+name = "h11"
+version = "0.16.0"
+description = "A pure-Python, bring-your-own-I/O implementation of HTTP/1.1"
+optional = false
+python-versions = ">=3.8"
+groups = ["main"]
+markers = "sys_platform != \"emscripten\""
+files = [
+    {file = "h11-0.16.0-py3-none-any.whl", hash = "sha256:63cf8bbe7522de3bf65932fda1d9c2772064ffb3dae62d55932da54b31cb6c86"},
+    {file = "h11-0.16.0.tar.gz", hash = "sha256:4e35b956cf45792e4caa5885e69fba00bdbc6ffafbfa020300e549b208ee5ff1"},
+]
+
+[[package]]
+name = "hf-xet"
+version = "1.6.0"
+description = "Fast transfer of large files with the Hugging Face Hub."
+optional = false
+python-versions = ">=3.8"
+groups = ["main"]
+markers = "platform_machine == \"x86_64\" or platform_machine == \"amd64\" or platform_machine == \"AMD64\" or platform_machine == \"arm64\" or platform_machine == \"ARM64\" or platform_machine == \"aarch64\""
+files = [
+    {file = "hf_xet-1.6.0-cp314-cp314t-macosx_10_12_x86_64.whl", hash = "sha256:70cbb9c896901600128cb9b6f06e132954fbede1db30f31f7c6c63f84cb7c31d"},
+    {file = "hf_xet-1.6.0-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:23379c2f9ec8696d952b16414a2bae72cad86a52df869b050698ba60f538c675"},
+    {file = "hf_xet-1.6.0-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:f2f7278c05c22fd60cb436cda1269649b3e81db65ecdc8496e5e164aa4143e7b"},
+    {file = "hf_xet-1.6.0-cp314-cp314t-manylinux_2_28_aarch64.whl", hash = "sha256:948f15d3a9545cfe5932f6bd8b440f6ae630aee108f14b7bd6c561f7c2dcc522"},
+    {file = "hf_xet-1.6.0-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:5153e6bb103ad49d6ea9f1b2e230db5a2ea32551ad09a706d2f61d7c7c80d80e"},
+    {file = "hf_xet-1.6.0-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:35cec30d75c6f9eb9c16a77cef68e85a103b72e24d4b473714ec9ff06428bab9"},
+    {file = "hf_xet-1.6.0-cp314-cp314t-win_amd64.whl", hash = "sha256:5789835d7c6bc9436962853192082374297fb72d7eff7e7762ec25ceb7e25338"},
+    {file = "hf_xet-1.6.0-cp314-cp314t-win_arm64.whl", hash = "sha256:75765820ce4700db3750c94acc8fe27c5fae4c9ec000a0dbac3ca082acf97765"},
+    {file = "hf_xet-1.6.0-cp38-abi3-macosx_10_12_x86_64.whl", hash = "sha256:633dc0cd71d32da58ab8c03ad38e2fac452c15c2b0a2866ebf6ededfe0a5061d"},
+    {file = "hf_xet-1.6.0-cp38-abi3-macosx_11_0_arm64.whl", hash = "sha256:f0906082d9932ae0c0057fa194041c22b4e2cdb46b2592ef3b91f020d62a081a"},
+    {file = "hf_xet-1.6.0-cp38-abi3-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:d62671bb130879cef0ee4c9ebe47a14af6c66ec53e6d84dc15936e5ffdfac82f"},
+    {file = "hf_xet-1.6.0-cp38-abi3-manylinux_2_28_aarch64.whl", hash = "sha256:0e6e21fa3cdfcdcd76748564bf593870a5e013f47d97cf10aed63aa222cff5b7"},
+    {file = "hf_xet-1.6.0-cp38-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:4fc74352a17015bd0ee90038bc9efe38db894cde45f268b6712b04fce8cd0acb"},
+    {file = "hf_xet-1.6.0-cp38-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:8fb4f71cba6129110c3374a33f919001ff130488fc23553698e34cc1c2a1198c"},
+    {file = "hf_xet-1.6.0-cp38-abi3-win_amd64.whl", hash = "sha256:fb4fadde1b2b70bf4c0c14a6dccbe7194b1c28947fefd5bbe3fed9d940676c3b"},
+    {file = "hf_xet-1.6.0-cp38-abi3-win_arm64.whl", hash = "sha256:3dc3e35441ba395006af5aaacc40ef2e603c51ef46c3530b9156185f00935ea3"},
+    {file = "hf_xet-1.6.0.tar.gz", hash = "sha256:2e58454a340b3556dfa4972d5451aff4fba8dd42a236600ba1a1d2b1514f0fef"},
+]
+
+[package.extras]
+tests = ["pytest"]
+
+[[package]]
+name = "httpcore2"
+version = "2.16.0"
+description = "A minimal low-level HTTP client."
+optional = false
+python-versions = ">=3.11"
+groups = ["main"]
+markers = "sys_platform != \"emscripten\""
+files = [
+    {file = "httpcore2-2.16.0-py3-none-any.whl", hash = "sha256:5c4358319f2602d984ad93182c9fee5160f22e9da7babdda870a5e7c4e2dc34f"},
+    {file = "httpcore2-2.16.0.tar.gz", hash = "sha256:c72cd0413563b17e1cd8354677dff887466f8ec2551679759e93126534428643"},
+]
+
+[package.dependencies]
+h11 = ">=0.16"
+truststore = ">=0.10"
+
+[package.extras]
+asyncio = ["anyio (>=4.5.0,<5.0)"]
+http2 = ["h2 (>=3,<5)"]
+socks = ["socksio (==1.*)"]
+trio = ["trio (>=0.34.0,<1.0)"]
+
+[[package]]
+name = "httpx2"
+version = "2.16.0"
+description = "The next generation HTTP client."
+optional = false
+python-versions = ">=3.11"
+groups = ["main"]
+files = [
+    {file = "httpx2-2.16.0-py3-none-any.whl", hash = "sha256:fd2c5d7909e8d86fe92d53020c310ca0e4e431b838c207c5754b9bbf0308d49a"},
+    {file = "httpx2-2.16.0.tar.gz", hash = "sha256:0d54c98128e8f2ff7fbbde0fca8c359e5f032edeaaaac443cdb7f31f685bcd7f"},
+]
+
+[package.dependencies]
+anyio = {version = ">=4.10", markers = "sys_platform != \"emscripten\""}
+httpcore2 = {version = "2.16.0", markers = "sys_platform != \"emscripten\""}
+httpx2-jsfetch = {version = "*", markers = "sys_platform == \"emscripten\" and python_version >= \"3.12\""}
+idna = ">=3.18"
+truststore = {version = ">=0.10", markers = "sys_platform != \"emscripten\""}
+typing-extensions = {version = ">=4.5.0", markers = "python_version < \"3.13\""}
+
+[package.extras]
+brotli = ["brotli (>=1.2.0) ; platform_python_implementation == \"CPython\"", "brotlicffi (>=1.2.0.2) ; platform_python_implementation != \"CPython\""]
+cli = ["click (>=8.4.2)", "pygments (==2.*)", "rich (>=10,<16)"]
+http2 = ["h2 (>=3,<5)"]
+socks = ["socksio (==1.*)"]
+ws = ["wsproto (>=1.2)"]
+zstd = ["backports-zstd (>=1.0.0) ; python_version <= \"3.13\""]
+
+[[package]]
+name = "httpx2-jsfetch"
+version = "1.0"
+description = "httpx2 transports for Emscripten/Pyodide, backed by the JavaScript fetch API."
+optional = false
+python-versions = ">=3.12"
+groups = ["main"]
+markers = "sys_platform == \"emscripten\""
+files = [
+    {file = "httpx2_jsfetch-1.0-py3-none-any.whl", hash = "sha256:cb916b707601e69a07721aabc8f3f6659be3a6893bc1ff5c6f9e02241df2da32"},
+    {file = "httpx2_jsfetch-1.0.tar.gz", hash = "sha256:70a0e3eabfef7cce5ad9c629f7d01ca05e418f586646f4ddf14782e4c1454c60"},
+]
+
+[[package]]
+name = "huggingface-hub"
+version = "2.3.0"
+description = "Client library to download and publish models, datasets and other repos on the huggingface.co hub"
+optional = false
+python-versions = ">=3.10.0"
+groups = ["main"]
+files = [
+    {file = "huggingface_hub-2.3.0-py3-none-any.whl", hash = "sha256:5da369600233568a126eb5578754935b2add9446b42d9e2fc0927d56840ac1a5"},
+    {file = "huggingface_hub-2.3.0.tar.gz", hash = "sha256:0b783b37f200f1d25bef2344a997c7d410f4437fbe9d8fd965f958e75cec7aef"},
+]
+
+[package.dependencies]
+click = ">=8.4.2,<9.0.0"
+filelock = ">=3.10.0"
+fsspec = ">=2023.5.0"
+hf-xet = {version = ">=1.6.0,<2.0.0", markers = "platform_machine == \"x86_64\" or platform_machine == \"amd64\" or platform_machine == \"AMD64\" or platform_machine == \"arm64\" or platform_machine == \"ARM64\" or platform_machine == \"aarch64\""}
+httpx2 = ">=2.0.0,<3"
+packaging = ">=20.9"
+pyyaml = ">=5.1"
+tqdm = ">=4.42.1"
+typing-extensions = ">=4.1.0"
+
+[package.extras]
+all = ["Jinja2", "Pillow", "authlib (>=1.8.0)", "duckdb ; python_version < \"3.15\"", "fastapi", "fastapi", "itsdangerous", "jedi", "libcst (>=1.4.0)", "mypy (==1.15.0)", "numpy", "pytest (>=8.4.2)", "pytest-asyncio", "pytest-cov", "pytest-env", "pytest-mock", "pytest-rerunfailures (>=16.2)", "pytest-vcr", "pytest-xdist", "ruff (>=0.9.0)", "soundfile", "ty", "types-PyYAML", "types-simplejson", "types-toml", "types-tqdm", "types-urllib3", "typing-extensions (>=4.8.0)", "urllib3 (<2.0)"]
+dev = ["Jinja2", "Pillow", "authlib (>=1.8.0)", "duckdb ; python_version < \"3.15\"", "fastapi", "fastapi", "itsdangerous", "jedi", "libcst (>=1.4.0)", "mypy (==1.15.0)", "numpy", "pytest (>=8.4.2)", "pytest-asyncio", "pytest-cov", "pytest-env", "pytest-mock", "pytest-rerunfailures (>=16.2)", "pytest-vcr", "pytest-xdist", "ruff (>=0.9.0)", "soundfile", "ty", "types-PyYAML", "types-simplejson", "types-toml", "types-tqdm", "types-urllib3", "typing-extensions (>=4.8.0)", "urllib3 (<2.0)"]
+fastai = ["fastai (>=2.4)", "fastcore (>=1.3.27)", "toml"]
+gradio = ["gradio (>=5.0.0)", "requests"]
+hf-xet = ["hf-xet (>=1.6.0,<2.0.0)"]
+mcp = ["mcp (>=1.9.4,<2.0.0)"]
+oauth = ["authlib (>=1.8.0)", "fastapi", "itsdangerous"]
+quality = ["libcst (>=1.4.0)", "mypy (==1.15.0)", "ruff (>=0.9.0)", "ty"]
+testing = ["Jinja2", "Pillow", "authlib (>=1.8.0)", "duckdb ; python_version < \"3.15\"", "fastapi", "fastapi", "itsdangerous", "jedi", "numpy", "pytest (>=8.4.2)", "pytest-asyncio", "pytest-cov", "pytest-env", "pytest-mock", "pytest-rerunfailures (>=16.2)", "pytest-vcr", "pytest-xdist", "soundfile", "urllib3 (<2.0)"]
+torch = ["safetensors[torch]", "torch"]
+typing = ["types-PyYAML", "types-simplejson", "types-toml", "types-tqdm", "types-urllib3", "typing-extensions (>=4.8.0)"]
+
+[[package]]
+name = "idna"
+version = "3.20"
+description = "Internationalized Domain Names in Applications (IDNA)"
+optional = false
+python-versions = ">=3.9"
+groups = ["main"]
+files = [
+    {file = "idna-3.20-py3-none-any.whl", hash = "sha256:ab7ae7122974553370f0bdb919e1a960b2cd1bc1ef0276416d896db81c14582c"},
+    {file = "idna-3.20.tar.gz", hash = "sha256:a7db850025b95ded1eae8a46181a1a6c56c92c96f0e2b005d9ff8dc0210cab44"},
+]
+
+[package.extras]
+all = ["coverage (>=7.10.0)", "hypothesis (>=6.141.1)", "mypy (>=1.11.2)", "pytest (>=8.3.2)", "ruff (>=0.16.0)", "ty (>=0.0.37)"]
+
+[[package]]
+name = "iniconfig"
+version = "2.3.0"
+description = "brain-dead simple config-ini parsing"
+optional = false
+python-versions = ">=3.10"
+groups = ["dev"]
+files = [
+    {file = "iniconfig-2.3.0-py3-none-any.whl", hash = "sha256:f631c04d2c48c52b84d0d0549c99ff3859c98df65b3101406327ecc7d53fbf12"},
+    {file = "iniconfig-2.3.0.tar.gz", hash = "sha256:c76315c77db068650d49c5b56314774a7804df16fee4402c1f19d6d15d8c4730"},
+]
+
+[[package]]
+name = "jinja2"
+version = "3.1.6"
+description = "A very fast and expressive template engine."
+optional = false
+python-versions = ">=3.7"
+groups = ["main"]
+files = [
+    {file = "jinja2-3.1.6-py3-none-any.whl", hash = "sha256:85ece4451f492d0c13c5dd7c13a64681a86afae63a5f347908daf103ce6d2f67"},
+    {file = "jinja2-3.1.6.tar.gz", hash = "sha256:0137fb05990d35f1275a587e9aee6d56da821fc83491a0fb838183be43f66d6d"},
+]
+
+[package.dependencies]
+MarkupSafe = ">=2.0"
+
+[package.extras]
+i18n = ["Babel (>=2.7)"]
+
+[[package]]
+name = "joblib"
+version = "1.6.0"
+description = "Lightweight pipelining with Python functions"
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "joblib-1.6.0-py3-none-any.whl", hash = "sha256:3dbbf9f6e4b592a2357b854608e980fe6390d131d7a82f011a377ef2ebef7aba"},
+    {file = "joblib-1.6.0.tar.gz", hash = "sha256:2ccc96785b12046c08fd6d55839c12857831b54a3c1673ffadd2f04bfc4eda03"},
+]
+
+[package.dependencies]
+cloudpickle = ">=3.0"
+
+[package.extras]
+docs = ["distributed", "lz4", "matplotlib", "numpy", "numpydoc", "pandas", "psutil", "pydata-sphinx-theme", "sphinx", "sphinx-copybutton", "sphinx-design", "sphinx-gallery", "tqdm"]
+test = ["distributed", "lz4", "memory_profiler", "numpy", "pytest", "pytest-asyncio", "pytest-cov", "pytest-run-parallel", "pytest-timeout", "threadpoolctl"]
+
+[[package]]
+name = "kev"
+version = "0.1.0"
+description = "Laptop-scale reconstruction of a Jev-style decision model with a TypeSafe-compatible /v1/systemone API"
+optional = false
+python-versions = ">=3.12,<3.14"
+groups = ["main"]
+files = []
+develop = false
+
+[package.dependencies]
+accelerate = ">=1.15.0"
+datasets = ">=3.0"
+numpy = ">=2.5.3"
+peft = ">=0.21"
+pydantic = ">=2.9"
+scikit-learn = ">=1.9.1"
+torch = ">=2.6,<2.9"
+transformers = ">=5.17,<6"
+
+[package.extras]
+mlx = ["mlx-lm (>=0.31.3,<0.32) ; sys_platform == \"darwin\" and platform_machine == \"arm64\""]
+serve = ["fastapi (>=0.115)", "mlx-lm (>=0.31.3,<0.32) ; sys_platform == \"darwin\" and platform_machine == \"arm64\"", "typesafe-sdk (>=0.6.0)", "uvicorn (>=0.30)"]
+
+[package.source]
+type = "git"
+url = "https://github.com/jaredpalmer/kev.git"
+reference = "2074cc44256200bb37540103fa4572901faa10f3"
+resolved_reference = "2074cc44256200bb37540103fa4572901faa10f3"
+
+[[package]]
+name = "librt"
+version = "0.16.0"
+description = "Mypyc runtime library"
+optional = false
+python-versions = ">=3.9"
+groups = ["dev"]
+markers = "platform_python_implementation != \"PyPy\""
+files = [
+    {file = "librt-0.16.0-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:96f576f2711f8519152ec76d0e599243555c1f07679fa73606ca8c8c868c0be6"},
+    {file = "librt-0.16.0-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:2bec3818c7da7c96ceae0ef5915a3d16c52dd08f3ea913bf1fe8568c447c7978"},
+    {file = "librt-0.16.0-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:64c79520414a3fdfc6aabd7593e6169afa14d5f8d9908d4b498db068868b08dd"},
+    {file = "librt-0.16.0-cp310-cp310-manylinux2014_i686.manylinux_2_17_i686.manylinux_2_28_i686.whl", hash = "sha256:f81b5b19ce748ef68d4746656b7929762eb2fe99269b266e4be07e2ee4de7144"},
+    {file = "librt-0.16.0-cp310-cp310-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:c71d1b76210a36729fedfc5115069b50a3d8619054745f8758fe5d6f19e86671"},
+    {file = "librt-0.16.0-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:473eebc7866bb0a0c8849a292b5e7157c1aba5d14d0f0f610c52158d6d964262"},
+    {file = "librt-0.16.0-cp310-cp310-manylinux_2_34_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:4eb1313a19847089ee81e88742abedf285c60538816640b99742d8534b81d26a"},
+    {file = "librt-0.16.0-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:ebefd60b42e2a82b32d136bb5f7c94eadfcd29f772b547df6f3291d1ed855a1c"},
+    {file = "librt-0.16.0-cp310-cp310-musllinux_1_2_i686.whl", hash = "sha256:6c5da27e8056439f927ea896735da60e616c477a8293feaa3233d4e7781a6726"},
+    {file = "librt-0.16.0-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:67e718c7a43f8db325abbbf1404e2d535f12f8f7a1a82259568385cc5274b82a"},
+    {file = "librt-0.16.0-cp310-cp310-musllinux_1_2_riscv64.whl", hash = "sha256:314e703f0c19320dc8094e7a784b9cf29e1b67402515abf580069a6363c0b4f1"},
+    {file = "librt-0.16.0-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:943c6bbecbdf7fa575a4f2952fcfd848c88ef95507c3fca411e89d4ac3ff8143"},
+    {file = "librt-0.16.0-cp310-cp310-win32.whl", hash = "sha256:7cc365f006891afb006b52d5ee5ee74c09306ffa20e2f8705a32a4450af2f3ba"},
+    {file = "librt-0.16.0-cp310-cp310-win_amd64.whl", hash = "sha256:0314058469f4d2fd279ce7c62ac274ac82c3918ef7db62ef0697c4c359370155"},
+    {file = "librt-0.16.0-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:fe4372c52d4849096c6cc1cda2817d293ec51440c890474ed59ef38d46556f18"},
+    {file = "librt-0.16.0-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:c72c5295a84bd249526da9bdca38f2e176d15c31c13bb0063c5053f4ca023421"},
+    {file = "librt-0.16.0-cp311-cp311-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:be56ba9c884143495b517f23fe794ae367d58cd89ea0fdd6d437e3c024a87f9f"},
+    {file = "librt-0.16.0-cp311-cp311-manylinux2014_i686.manylinux_2_17_i686.manylinux_2_28_i686.whl", hash = "sha256:ef46c1a29ffb8c72e882e22618ec618778eacd0578fb22c6e7cf9c11d15f357b"},
+    {file = "librt-0.16.0-cp311-cp311-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:d3c94211ee0c4f8d649ec06b7c115c0ec4eadb873a0e3154ca15cef3f814b071"},
+    {file = "librt-0.16.0-cp311-cp311-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:94aed6a8308818b91677957d1bd03188869cd7aeb23c5dba7912a6c0402f7602"},
+    {file = "librt-0.16.0-cp311-cp311-manylinux_2_34_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:349c0bcb87ebd07481b6ff781e25cdc699723dbe2212e57dabb27f7a13b7b87d"},
+    {file = "librt-0.16.0-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:001bfd59a7d45b17e3e75f2a8c6405280b35e7b84471792778e718c4f368950e"},
+    {file = "librt-0.16.0-cp311-cp311-musllinux_1_2_i686.whl", hash = "sha256:28e038895b998d7a0c7798922ce8a1dc157675df5cf1c9ef0aca809ed804b7a1"},
+    {file = "librt-0.16.0-cp311-cp311-musllinux_1_2_ppc64le.whl", hash = "sha256:0dbe4096a7ecc00fa835d24510ad8545a4efef738dac96e0e63516783ccde905"},
+    {file = "librt-0.16.0-cp311-cp311-musllinux_1_2_riscv64.whl", hash = "sha256:5cd5b092441053364af968ea12084692cb9d4a22f3ce9524e377880bf028761e"},
+    {file = "librt-0.16.0-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:3ddeb3c9dedb461bb457c6c7d9aa7fbf35329da313d1a7543d00c8d0f3473c96"},
+    {file = "librt-0.16.0-cp311-cp311-win32.whl", hash = "sha256:e05108e0849966f53a8d2d3112a7af881d0efaa479bc735bba91108f9f2350a7"},
+    {file = "librt-0.16.0-cp311-cp311-win_amd64.whl", hash = "sha256:5f49cff01bd608ef7d97104cb035c75455e79c2d70bf4a506cf773338ac1860d"},
+    {file = "librt-0.16.0-cp311-cp311-win_arm64.whl", hash = "sha256:1d28ae980ae2218f9c5b95d191e947296f918c9bf0b400d467a9430275bbe678"},
+    {file = "librt-0.16.0-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:fe52bf4641069e7978a14253b036cb9002def1926317e710f2e249f8a8c47742"},
+    {file = "librt-0.16.0-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:5bcc2c4726ced915b00de0c9856a4eeabfb3fddb93e10e0b8f735b7709358b6d"},
+    {file = "librt-0.16.0-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:ff7baa55f8e7c69851419e50a666015d02a74198716fd45c0125a2112e0a389f"},
+    {file = "librt-0.16.0-cp312-cp312-manylinux2014_i686.manylinux_2_17_i686.manylinux_2_28_i686.whl", hash = "sha256:b95d5d92ab83d39e760a52091bb1baba664f3a2351e39b1e16801e5747c2f0e9"},
+    {file = "librt-0.16.0-cp312-cp312-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:b6d085d70bce51d43c5c7c36d63490770180d8779e71c49305c87b4213918de7"},
+    {file = "librt-0.16.0-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:36e53948e99bbe3ffea257124cfcae1cfb01831555c9a9c903c9f9a72db7fd07"},
+    {file = "librt-0.16.0-cp312-cp312-manylinux_2_34_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:54d11f726aae9df5a6ffbbf0a03a52449bbac84a53ef03669cb41cdfd4ae41bf"},
+    {file = "librt-0.16.0-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:4323193ac0cd025f85af531df8ba91bf24d1973b401697347a6282e8fd3fcf5e"},
+    {file = "librt-0.16.0-cp312-cp312-musllinux_1_2_i686.whl", hash = "sha256:e42f8e098b9c5396fefa05fb1cc7e33b0e08fc51da106b5de4a45fd22aac6743"},
+    {file = "librt-0.16.0-cp312-cp312-musllinux_1_2_ppc64le.whl", hash = "sha256:39ec1d5a14e37baf1450a6cabf03fe552340808bf1ad9d71824ab90117716459"},
+    {file = "librt-0.16.0-cp312-cp312-musllinux_1_2_riscv64.whl", hash = "sha256:d1aabe3925cbb4a08d15b7b20ba4011b53019da0c4173a25155139b7b1baed65"},
+    {file = "librt-0.16.0-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:300c3ffdc459f4a779a8411ecb188e3ac0b1ff3a3a7b099642555dedae06c69b"},
+    {file = "librt-0.16.0-cp312-cp312-win32.whl", hash = "sha256:c17194318e4c0c0348b36f36c2ec7534436fe0a4c15582403162a4f08c80797a"},
+    {file = "librt-0.16.0-cp312-cp312-win_amd64.whl", hash = "sha256:25a58a19ea8d83b68209f04912df765e9260635ef77646542ed4b4abe6bc7940"},
+    {file = "librt-0.16.0-cp312-cp312-win_arm64.whl", hash = "sha256:f7be7cf555bc30ec12622e9447299cc4a9b8ff307548b634794353db0c2065dc"},
+    {file = "librt-0.16.0-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:c5e6144e68b577f157519f2ba88ca20e3ed61c29b00e5cdfa76cd2d45acf059a"},
+    {file = "librt-0.16.0-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:33f41443a1f4e1f099331b3d8120e409fbff84b9760bc1cc9ea496f37ddaa5cc"},
+    {file = "librt-0.16.0-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:7e510b7770bee609617a3374a96548eb114cae048023e3f049ee449e7ff2db32"},
+    {file = "librt-0.16.0-cp313-cp313-manylinux2014_i686.manylinux_2_17_i686.manylinux_2_28_i686.whl", hash = "sha256:efc49c462d4516b8a58b00b490078fa64689fd1fe66970cc190131d7afb8027e"},
+    {file = "librt-0.16.0-cp313-cp313-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:92caf82ebef5e12d21c72242b70d1e92536f1711cf2a727a4c276de4b4469087"},
+    {file = "librt-0.16.0-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:17bac7f7a16b328fff77e440287693eb017abde913595b5827ebccbc21ecd8a6"},
+    {file = "librt-0.16.0-cp313-cp313-manylinux_2_34_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:5b976054553670829985ed767feb78fb6bcede0175327c4844dd5c281c1be659"},
+    {file = "librt-0.16.0-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:0058f9d68721094105917254c72ac0569117bb7b13b9769cf45d26d89f9d21cd"},
+    {file = "librt-0.16.0-cp313-cp313-musllinux_1_2_i686.whl", hash = "sha256:30b7beaf3f4487b7d8adef1f158b49067cb4d5a19fa7a3bf31a4e7a820e435c5"},
+    {file = "librt-0.16.0-cp313-cp313-musllinux_1_2_ppc64le.whl", hash = "sha256:468df902df016a06eb0e40b0747dc8d14e47d7a38b18b63b1fb167d85cb94d63"},
+    {file = "librt-0.16.0-cp313-cp313-musllinux_1_2_riscv64.whl", hash = "sha256:aea7b1f2b125dad5de85f049136651bff256c883c65e6b9209b2da0a1ac3cdef"},
+    {file = "librt-0.16.0-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:a8afb6557920860b7a3a596eb804cf37e09e7cf8a803db2478c202acc72d8c2e"},
+    {file = "librt-0.16.0-cp313-cp313-pyemscripten_2025_0_wasm32.whl", hash = "sha256:77c7a2b4fe2c1369e0d5aa1cade26740a7b14be32fbc9a5535d617d20065c39d"},
+    {file = "librt-0.16.0-cp313-cp313-win32.whl", hash = "sha256:02d89c813d5ff74b17df72d3a34819d132cd168e56b81bf755b809bd9e46b8c4"},
+    {file = "librt-0.16.0-cp313-cp313-win_amd64.whl", hash = "sha256:14ed6ebe3e4f85f326d7920011ad30ff49ed9334e62cf88caef9ba973d9e3a92"},
+    {file = "librt-0.16.0-cp313-cp313-win_arm64.whl", hash = "sha256:83d4041a3d9b2fd053a8a4e1f22878b3e5833e2712956382d5c048d791454e91"},
+    {file = "librt-0.16.0-cp314-cp314-macosx_10_15_x86_64.whl", hash = "sha256:931a0bb0fcac88f263e269e46eb30ba8e21402cd3c62ca40cb97034c0693fab1"},
+    {file = "librt-0.16.0-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:1bc17e54e5305f8d40b7ca203671ff5a9e59c1d0f8ea0f625dcca53a3984de11"},
+    {file = "librt-0.16.0-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:877698bf6bca5721d8be345f2fe09778e40ecadea8b58c73075f2b1a53666bf2"},
+    {file = "librt-0.16.0-cp314-cp314-manylinux2014_i686.manylinux_2_17_i686.manylinux_2_28_i686.whl", hash = "sha256:5981c011b306781ce561e18e14230a14524a3d8109b97553666c942c18f31a96"},
+    {file = "librt-0.16.0-cp314-cp314-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:afced3dfc17cd805ecf7a3d77996a71cf5f2c75aa66eb0c21a9930f4fc992f86"},
+    {file = "librt-0.16.0-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:ca8052401c55d7511dda6760719fda7618067e83535d7d0010096d216c34b667"},
+    {file = "librt-0.16.0-cp314-cp314-manylinux_2_34_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:1e511762a074005bb0aa569166779834e75e438370226930d0ce1866d4b6a33b"},
+    {file = "librt-0.16.0-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:f1e8591bd8a5a628cd7f07954c6a1592359a878bf032957a8e9057a41d644311"},
+    {file = "librt-0.16.0-cp314-cp314-musllinux_1_2_i686.whl", hash = "sha256:a4aaefb4ba6c07e1aeebb2795c8958148f1d6f9af3b555b53d23d766edb6d67a"},
+    {file = "librt-0.16.0-cp314-cp314-musllinux_1_2_ppc64le.whl", hash = "sha256:d92db7a0f6aee44f1baee94750457e8d2d1c6ccea41842de6268d34e8dc7eddd"},
+    {file = "librt-0.16.0-cp314-cp314-musllinux_1_2_riscv64.whl", hash = "sha256:378dfaffb38e59c24a87cde5713cd865d51ff7383fa12947f3907f306ea1ca55"},
+    {file = "librt-0.16.0-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:3e0c39bdc85370422e8b637be76eb1fd07d30967551b03e62267dd156f553152"},
+    {file = "librt-0.16.0-cp314-cp314-pyemscripten_2026_0_wasm32.whl", hash = "sha256:1b384b90ab79a7bc30b566895809a636e0666f21f3cf12b54823d025b7e83839"},
+    {file = "librt-0.16.0-cp314-cp314-win32.whl", hash = "sha256:52327da75a94012e7f932f913d20d3876bed3c102be00e6c3e8600ff7bdd58a7"},
+    {file = "librt-0.16.0-cp314-cp314-win_amd64.whl", hash = "sha256:3f0b8114c44b2ac06ff5dacd08e07e8e807ff4f46083f2a1602685122559be41"},
+    {file = "librt-0.16.0-cp314-cp314-win_arm64.whl", hash = "sha256:8caf96a4ef8fb27d0ac0d1ad8337d26a240acd4a02fe4345d0a8f264753e8f99"},
+    {file = "librt-0.16.0-cp314-cp314t-macosx_10_15_x86_64.whl", hash = "sha256:953107e2f68d0f3512c48f898b0dbf0ce5cc52bba0f318d847c985dc555ee4cc"},
+    {file = "librt-0.16.0-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:ad37d5b9abd49c9a655dcda7ea52a8a752884062ef1ee71ae17c2f2a0f81fe6a"},
+    {file = "librt-0.16.0-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:2c4aa329c17bd1aaea4f6e89335d8ccd494b3a5830b6654462273e50e11023f0"},
+    {file = "librt-0.16.0-cp314-cp314t-manylinux2014_i686.manylinux_2_17_i686.manylinux_2_28_i686.whl", hash = "sha256:0ead24d2562a49473dddd9efef8581f020007eb0054389c3ee3ffad38b1ca4c9"},
+    {file = "librt-0.16.0-cp314-cp314t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:02118f56a9c36ddd07dfd9b919d9ecc117ba20a90987d56aa4c429fa34509188"},
+    {file = "librt-0.16.0-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:4e29522c62e28595ff7e324c6834ade51127707f0e255b18d1c1cf03d39c1048"},
+    {file = "librt-0.16.0-cp314-cp314t-manylinux_2_34_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:3ff4b2367926b69c6215635902cccb04048e73094e9862900d27cb2c6bbff143"},
+    {file = "librt-0.16.0-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:c6f1b27bf1632a7e016af9f145f82be95e1edd7721a646505c21059257cb5a04"},
+    {file = "librt-0.16.0-cp314-cp314t-musllinux_1_2_i686.whl", hash = "sha256:5696d7f52e7b37217cb3a8f92c744fe835942602fdd4c1a8bc4741d3bfdce15e"},
+    {file = "librt-0.16.0-cp314-cp314t-musllinux_1_2_ppc64le.whl", hash = "sha256:6072e92dd876ff6ceeb6cf371e35e51f479349837391341f479b08df4564242b"},
+    {file = "librt-0.16.0-cp314-cp314t-musllinux_1_2_riscv64.whl", hash = "sha256:39ca4f2f2fe05de8e63493da592d84311adabe5bef52b193851981da9816b302"},
+    {file = "librt-0.16.0-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:f9807485a908f00355820f18e91e045ffdcdc5adb68aaec40a1e2b88c5f7bba1"},
+    {file = "librt-0.16.0-cp314-cp314t-win32.whl", hash = "sha256:94be5cb7bca4df6201f4183e9e4fa2086c655283d20b38cd84500a69057575a7"},
+    {file = "librt-0.16.0-cp314-cp314t-win_amd64.whl", hash = "sha256:d46ca272b251d033dd4527b0dec5f261a28a52bd5fa0f99c117b0a1f8588cc2d"},
+    {file = "librt-0.16.0-cp314-cp314t-win_arm64.whl", hash = "sha256:b9d6d4b14e92d876f8026b54c20c445f36425214c1081dc76f74e40db386b82b"},
+    {file = "librt-0.16.0-cp315-cp315-macosx_10_15_x86_64.whl", hash = "sha256:6fe436af2eaf630474f491af5d032cbe45f93fcff5c3b9fe4ab194a7255b20ff"},
+    {file = "librt-0.16.0-cp315-cp315-macosx_11_0_arm64.whl", hash = "sha256:8ff5d26c529336be9bd7ae04483235d77778ee7d6444a95353102b542601ce81"},
+    {file = "librt-0.16.0-cp315-cp315-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:909d8e3c1faee44cb762b1c519ff8613dcc5ceae5c99987a00917b5a31fd1d6a"},
+    {file = "librt-0.16.0-cp315-cp315-manylinux2014_i686.manylinux_2_17_i686.manylinux_2_28_i686.whl", hash = "sha256:6d4a64283ee61824b5790de882bc68e2d9d7a5143537cb7a966f7354f71646d4"},
+    {file = "librt-0.16.0-cp315-cp315-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:5810ba811297fdf37a1531a57667cb8ace0842013ca8606bf9eb7c24cf4be154"},
+    {file = "librt-0.16.0-cp315-cp315-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:e56aaf8c167548dc8e5d6f3bd0f48dcdd299a23c73be3f744aab79d99e9c7f5d"},
+    {file = "librt-0.16.0-cp315-cp315-manylinux_2_34_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:8f36c58e33b304b525c6c9c5076399c6ebf1109e17b9051a05a407b091b9215b"},
+    {file = "librt-0.16.0-cp315-cp315-musllinux_1_2_aarch64.whl", hash = "sha256:242e00b3d4fa37c3d3c1ca5f5c9adb7d909ddb1eac9c41f2787320d00caa0af2"},
+    {file = "librt-0.16.0-cp315-cp315-musllinux_1_2_i686.whl", hash = "sha256:0253721561787b8df8443eb347b7a6461015354e5bdd37ee38a41fef220d2bb0"},
+    {file = "librt-0.16.0-cp315-cp315-musllinux_1_2_ppc64le.whl", hash = "sha256:3e483a8d69ede8067db70c0e83007423b6925de6fd53afed01d66160f2e9398c"},
+    {file = "librt-0.16.0-cp315-cp315-musllinux_1_2_riscv64.whl", hash = "sha256:69ba927445cfaaffb4081003ef5224c55a5c2ab67ef956f416ef744916e44121"},
+    {file = "librt-0.16.0-cp315-cp315-musllinux_1_2_x86_64.whl", hash = "sha256:d6a365f2ab45a984d0e00eee0dd17f599ceab8cadab6ea07b6111c8132fc0e42"},
+    {file = "librt-0.16.0-cp315-cp315-pyemscripten_2026_5_wasm32.whl", hash = "sha256:f01f3805f2dae4781c0c34b440e31740d082950bdaf89a6f601ad589a28af57a"},
+    {file = "librt-0.16.0-cp315-cp315-win32.whl", hash = "sha256:b0e3e721c75d2e79a76d4422c79d7ba705fe1bbafec907037fe7a657a480a0e3"},
+    {file = "librt-0.16.0-cp315-cp315-win_amd64.whl", hash = "sha256:bc02954b1295de798bbdb0b4e2d8a28c2117de8b5c73dcbeb27dc32572dfb971"},
+    {file = "librt-0.16.0-cp315-cp315-win_arm64.whl", hash = "sha256:c5db585d43449a5f54303d4b2774e45e1babd975cfe1630a3d708c0b80c3e560"},
+    {file = "librt-0.16.0-cp315-cp315t-macosx_10_15_x86_64.whl", hash = "sha256:f06c689cb14afd9b612727553a5ec5a40febf113ca41c4413a2b0b334285884b"},
+    {file = "librt-0.16.0-cp315-cp315t-macosx_11_0_arm64.whl", hash = "sha256:13b4e8aba90b0b1c82474e9844aa9ffe7ad3faa484350e1da64cb8188d903134"},
+    {file = "librt-0.16.0-cp315-cp315t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:5a269c46ae327d8e6f8c1f85f7516cb52c0fa48127565a1105a4f4a05ff2a0b4"},
+    {file = "librt-0.16.0-cp315-cp315t-manylinux2014_i686.manylinux_2_17_i686.manylinux_2_28_i686.whl", hash = "sha256:a33e0dae1f8592146a4764d54ce842b278732d21a84e17c3bbe6b1bc158a2248"},
+    {file = "librt-0.16.0-cp315-cp315t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:47ada6ea32636492c61aa8ad27ae3b9404bfe7a97e3ba946d1984236cc741da0"},
+    {file = "librt-0.16.0-cp315-cp315t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:c43bd6e642d8a248c114327f98dd25ac5a7cb5aa168ef02f0559b91874df16b8"},
+    {file = "librt-0.16.0-cp315-cp315t-manylinux_2_34_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:c3d1bb7841a816ace6449bb26d3f9560dbfa20e71c568d23f0f62bf1e68f50b1"},
+    {file = "librt-0.16.0-cp315-cp315t-musllinux_1_2_aarch64.whl", hash = "sha256:3931f7a3db322e7f44e02a280e3949326ce9579ad388ee8d691dc7c76da9fb70"},
+    {file = "librt-0.16.0-cp315-cp315t-musllinux_1_2_i686.whl", hash = "sha256:f4462528b6000afe8f16907b5c7c2553abf1df005ba5140e6eb394541c3624c3"},
+    {file = "librt-0.16.0-cp315-cp315t-musllinux_1_2_ppc64le.whl", hash = "sha256:80039ba9b6a7d5f1a0175a4cca6bbefead87bd854c80abad1cb30afe47a830db"},
+    {file = "librt-0.16.0-cp315-cp315t-musllinux_1_2_riscv64.whl", hash = "sha256:7a1d272724b581bb6bc769dfdafed6da2ecc9886ba2450311de55a4ac2e1e9cd"},
+    {file = "librt-0.16.0-cp315-cp315t-musllinux_1_2_x86_64.whl", hash = "sha256:fbe4fb8c5445f7496d7f7f6bb0807875d09d47e6771ffa175fb2df2895fb86ba"},
+    {file = "librt-0.16.0-cp315-cp315t-win32.whl", hash = "sha256:375bfe6b572a8f6cfc398709356046173bf27e64c4c5edaf5f7062f051fb4bf9"},
+    {file = "librt-0.16.0-cp315-cp315t-win_amd64.whl", hash = "sha256:bd3150023d3dc2bc70f3784e59ffa1140d56ddba3d8125b3d6f9f85221279bfc"},
+    {file = "librt-0.16.0-cp315-cp315t-win_arm64.whl", hash = "sha256:8ceafb70f2a4f0826f11031942e59c0728fd98da112dc346d4352bde1e486866"},
+    {file = "librt-0.16.0-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:e1967e36ac4cae0c7e9615ad32e1a513cdacff79f9e8afb28bedc91caf48b4f3"},
+    {file = "librt-0.16.0-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:20fe0bf9053885e21c62b3e091fb5e73e1c54d1daf2e70eb388d70763bcd4220"},
+    {file = "librt-0.16.0-cp39-cp39-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:6a63610fa76524edfa605b5b259a603915c7a6e10e54f003e5503030506e81de"},
+    {file = "librt-0.16.0-cp39-cp39-manylinux2014_i686.manylinux_2_17_i686.manylinux_2_28_i686.whl", hash = "sha256:845a511b60ca43b9880dcc84a9784c891d6a2098c829130b320846c69c9c0c68"},
+    {file = "librt-0.16.0-cp39-cp39-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:d608f0bf3b8cbddd0067fe02cb8cab7d13e8eb9386b23a1843d4363044fd9e22"},
+    {file = "librt-0.16.0-cp39-cp39-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:df183721229ae51eef90108c115b43e98cb169b6155d34480338e5fc6616df00"},
+    {file = "librt-0.16.0-cp39-cp39-manylinux_2_34_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:4b6183e2e2e0ee00aac2ec07c7f7d151c97e85666b304f574b71cff0f9fccc4e"},
+    {file = "librt-0.16.0-cp39-cp39-musllinux_1_2_aarch64.whl", hash = "sha256:273d00be33792a15189331df10f1f1331621b043881e66c6c4377f7f776e1291"},
+    {file = "librt-0.16.0-cp39-cp39-musllinux_1_2_i686.whl", hash = "sha256:aa9357a1b4d4fc787bb718a59cb1112c28c8a976d6bfa268b71cc0a4ab8f3a94"},
+    {file = "librt-0.16.0-cp39-cp39-musllinux_1_2_ppc64le.whl", hash = "sha256:e9ce0bc440e7fd09b5f51f372f0f6640658f854b8fb05260f819cc669c93c42d"},
+    {file = "librt-0.16.0-cp39-cp39-musllinux_1_2_riscv64.whl", hash = "sha256:6c8893eae2fd13c5488d94056f3e6e5cf3142bfb1c4acaf136cb33d760c5964b"},
+    {file = "librt-0.16.0-cp39-cp39-musllinux_1_2_x86_64.whl", hash = "sha256:7393c9a48dcce4817dbd4b0d8ff6237efe9b0a0609f5b0adaef315f8541726b5"},
+    {file = "librt-0.16.0-cp39-cp39-win32.whl", hash = "sha256:71b93b42784e25b975079573c642a8fedb049a7bb31d70a51721b1666b3b2ced"},
+    {file = "librt-0.16.0-cp39-cp39-win_amd64.whl", hash = "sha256:5750a105b42a416f930edc59054927a406effb2550cd5bab92ad7a5842ed5d05"},
+    {file = "librt-0.16.0.tar.gz", hash = "sha256:ac38d6d8d66bf3d744148dbbc0b8e193e195a51e364ed55e224631f5721891fc"},
+]
+
+[[package]]
+name = "markdown-it-py"
+version = "4.2.0"
+description = "Python port of markdown-it. Markdown parsing, done right!"
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "markdown_it_py-4.2.0-py3-none-any.whl", hash = "sha256:9f7ebbcd14fe59494226453aed97c1070d83f8d24b6fc3a3bcf9a38092641c4a"},
+    {file = "markdown_it_py-4.2.0.tar.gz", hash = "sha256:04a21681d6fbb623de53f6f364d352309d4094dd4194040a10fd51833e418d49"},
+]
+
+[package.dependencies]
+mdurl = ">=0.1,<1.0"
+
+[package.extras]
+benchmarking = ["psutil", "pytest", "pytest-benchmark"]
+compare = ["commonmark (>=0.9,<1.0)", "markdown (>=3.4,<4.0)", "markdown-it-pyrs", "mistletoe (>=1.0,<2.0)", "mistune (>=3.0,<4.0)", "panflute (>=2.3,<3.0)"]
+linkify = ["linkify-it-py (>=1,<3)"]
+plugins = ["mdit-py-plugins (>=0.5.0)"]
+profiling = ["gprof2dot"]
+rtd = ["ipykernel", "jupyter_sphinx", "mdit-py-plugins (>=0.5.0)", "myst-parser", "pyyaml", "sphinx", "sphinx-book-theme (>=1.0,<2.0)", "sphinx-copybutton", "sphinx-design"]
+testing = ["coverage", "pytest", "pytest-cov", "pytest-regressions", "pytest-timeout", "requests"]
+
+[[package]]
+name = "markupsafe"
+version = "3.0.4"
+description = "Safely add untrusted strings to HTML/XML markup."
+optional = false
+python-versions = ">=3.9"
+groups = ["main"]
+files = [
+    {file = "markupsafe-3.0.4-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:dd8ea6ebee7aedbf7c749fa80521d9ccf1ba473e0d1e14805caafbaad281c889"},
+    {file = "markupsafe-3.0.4-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:dff05cb7016dff1e9fd68f4122c127b65dfc59de5306cfb7ad92f956f230bee2"},
+    {file = "markupsafe-3.0.4-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:cf63c214fe879a65e69a386f915e36104fc84254ab141240f8854602d8e0be2a"},
+    {file = "markupsafe-3.0.4-cp310-cp310-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:2a6ef68ae94aed8721934072b27a3b654ea2100b97e4ab864cf1489c90926fbc"},
+    {file = "markupsafe-3.0.4-cp310-cp310-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:fd9f8797427910198f95bced71ddfed61130d7e349213bfb8466c9c99e2c46a8"},
+    {file = "markupsafe-3.0.4-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:d1aca03ede943eb80ab3d63bb082c84b7aab85ea83bd0fd0c200260945fb49d9"},
+    {file = "markupsafe-3.0.4-cp310-cp310-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:0764a13d34cae40db7bbf3a09b7e9b491bf4603e20b263a7a9d6b8e324975d0a"},
+    {file = "markupsafe-3.0.4-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:9388003072b95f2f1e3fd908604194d653ba21330d811961a78b7da1a77e9e36"},
+    {file = "markupsafe-3.0.4-cp310-cp310-musllinux_1_2_armv7l.whl", hash = "sha256:8698d70a8081ee8c090dbb394768b5789a1da8b131b5499f89d071dd3cfaf6be"},
+    {file = "markupsafe-3.0.4-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:bf053da3c97a4bc5ecfbb218cdd2983febd91c617be8367d139882aa11e490aa"},
+    {file = "markupsafe-3.0.4-cp310-cp310-musllinux_1_2_riscv64.whl", hash = "sha256:9438a2648b2195980cb2dd8e53ed7b8df91319e2d0b70ae61a9e1d1bc8d3bec9"},
+    {file = "markupsafe-3.0.4-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:88d59b473bfb03259722600839af9bbd7fa13a2eb514beefeedb95997882f69a"},
+    {file = "markupsafe-3.0.4-cp310-cp310-win32.whl", hash = "sha256:4a540e2d3192792fc84eced57bef37851ccb2b41f73291bb17408eea77bcd278"},
+    {file = "markupsafe-3.0.4-cp310-cp310-win_amd64.whl", hash = "sha256:5c22873ad1f0532ba40fa1727f3c0fc1bbbaab6d373d4cbe3f0dc74b2e2521c7"},
+    {file = "markupsafe-3.0.4-cp310-cp310-win_arm64.whl", hash = "sha256:3d23795802fc8bd72534836d64489bbf0f67c088959091bdb22e10735a5107bf"},
+    {file = "markupsafe-3.0.4-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:9e25feb9e330b63edb0278a0acdf85e50d0cb0fbf49c3084abbe4e24ae195346"},
+    {file = "markupsafe-3.0.4-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:7d3391b2188d18737cb2fa147028b1096236eaa7e156446c650a489fa2cadc91"},
+    {file = "markupsafe-3.0.4-cp311-cp311-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:849dd2bb0e5e4ab2b71c7191726a4a8d5aa8a610daa584728cbee0b710ddc4ef"},
+    {file = "markupsafe-3.0.4-cp311-cp311-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:befb4158af32106b9a93db8d6d1d1cbbd418c0d5aca0cabb7b1780abf0c89169"},
+    {file = "markupsafe-3.0.4-cp311-cp311-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:71f88e749ea29f67f21f3b36433c1dc54c7729ed2a6d9e2da2e0d9e0d7b224eb"},
+    {file = "markupsafe-3.0.4-cp311-cp311-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:6da83a088f8ef93b2d483a8232a4dbf4d69d3d8496b568a03c56becac43e1808"},
+    {file = "markupsafe-3.0.4-cp311-cp311-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:8f0fac8b13d14bb06c68195f849371924ae53dd7b1c00fed24650f704383b692"},
+    {file = "markupsafe-3.0.4-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:4a7cdc2a420ca01058182da4253329764d4bfa055564d1eced90e6ba1e8b1d3d"},
+    {file = "markupsafe-3.0.4-cp311-cp311-musllinux_1_2_armv7l.whl", hash = "sha256:83b3944fea42a8400edf92fd1770fb8d0d4f7de651353bd2d8525a92dba69a21"},
+    {file = "markupsafe-3.0.4-cp311-cp311-musllinux_1_2_ppc64le.whl", hash = "sha256:8138eb83940ec7299024d92d4dee45f601b9e6c5ffde9d25f4e35e326203c707"},
+    {file = "markupsafe-3.0.4-cp311-cp311-musllinux_1_2_riscv64.whl", hash = "sha256:811d02d5122171c1941357efd8f9bf4ffe907b7f0a1a4e729a880e4be3f46e3e"},
+    {file = "markupsafe-3.0.4-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:50b5bedc9ed8a94fc8857a42ef4f84a81ea88f8d4f05dc8705fb23ee6d8dcca7"},
+    {file = "markupsafe-3.0.4-cp311-cp311-win32.whl", hash = "sha256:2e5a7cd7fdd14fcb1ae5d7d8bf23d24fbd1daefd1fbca2580132e1ea75f098b5"},
+    {file = "markupsafe-3.0.4-cp311-cp311-win_amd64.whl", hash = "sha256:fdb4ca07ab75ffadab4a8b135ad59cdbb3156b99310f3d565370da74a15d6bd3"},
+    {file = "markupsafe-3.0.4-cp311-cp311-win_arm64.whl", hash = "sha256:569d65055d367e3dcdf30c3f41119467b73d9ee9faf332bdf40402644f5ac08e"},
+    {file = "markupsafe-3.0.4-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:61631e08084be9e21a8967ec3139c7616ed7c5e9368e05c86d1b39562c8a57b6"},
+    {file = "markupsafe-3.0.4-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:0930db9bdc62d22944e10b066448bb65dc9abe9112880c7cab8da54db4284d5f"},
+    {file = "markupsafe-3.0.4-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:6a45c3d514f2436064db00d7fc8778d888f0236ebfed649b53d13a59e69ad51b"},
+    {file = "markupsafe-3.0.4-cp312-cp312-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:1e1451fab512d1bcc3dc26988ec1edb0b82c2db909132872cd9356070a6b63df"},
+    {file = "markupsafe-3.0.4-cp312-cp312-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:bd3ce56ae2cbae3ba82b683bc425cd7e48d2ed8b10f3e818186b6f5646d9271c"},
+    {file = "markupsafe-3.0.4-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:8e124f974786f831d6043728e38296969d3579db8896fe004682f5758e613581"},
+    {file = "markupsafe-3.0.4-cp312-cp312-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:c02e8f18bdedba082cef725942ac823b9b60656db07f7e265cb31618dfd00d77"},
+    {file = "markupsafe-3.0.4-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:9f098115c247e11d138ab83a28fa0323c77015007ea2df73ba5fd714dfefd67c"},
+    {file = "markupsafe-3.0.4-cp312-cp312-musllinux_1_2_armv7l.whl", hash = "sha256:d5f93ebbeb8032d47e349328ec8662d973d9b05a70b3c35df1f91fe419b84749"},
+    {file = "markupsafe-3.0.4-cp312-cp312-musllinux_1_2_ppc64le.whl", hash = "sha256:64511c54db4e4987aef4c41923235927428729e8174c5dba488429be70a998ed"},
+    {file = "markupsafe-3.0.4-cp312-cp312-musllinux_1_2_riscv64.whl", hash = "sha256:e1a622f13970d81f95d0c72f9dc090dce9085fccfa4c9f2174377ee32bd15786"},
+    {file = "markupsafe-3.0.4-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:c9a7f43c0b202b334cc9184af09bb8f21d3a209e038efaf106936fb69e6b026e"},
+    {file = "markupsafe-3.0.4-cp312-cp312-win32.whl", hash = "sha256:f0ec3b750b59375eab5b0fb2b9254810c00a3375be6d789899f1055a1d556237"},
+    {file = "markupsafe-3.0.4-cp312-cp312-win_amd64.whl", hash = "sha256:11935df9bf455ed0c04eb87bcd720f02b1fe5e02128a9430f23aed6f93336fc7"},
+    {file = "markupsafe-3.0.4-cp312-cp312-win_arm64.whl", hash = "sha256:a4bbd2d87dd233b9fc5812160c3d0ffbe42edc22a26ce0469f58479ede633fe9"},
+    {file = "markupsafe-3.0.4-cp313-cp313-android_24_arm64_v8a.whl", hash = "sha256:de8b364c423ef0a4bad9069657d617f9a5d2b2062457a89b1fa16ee199c399c1"},
+    {file = "markupsafe-3.0.4-cp313-cp313-android_24_x86_64.whl", hash = "sha256:34bdde374c5932765d7dc685c4a1d191a3207852d67e8e0a9eb6ea85156181f1"},
+    {file = "markupsafe-3.0.4-cp313-cp313-ios_13_0_arm64_iphoneos.whl", hash = "sha256:6bd9e1788e15bfcf6a9082de42e30387e7b85d211ab21e57a939bb8cfaaf8d96"},
+    {file = "markupsafe-3.0.4-cp313-cp313-ios_13_0_arm64_iphonesimulator.whl", hash = "sha256:5066b244f576f91afc8ee3ba029a89f99d39c79b1853fe9d39bea9f0afbec148"},
+    {file = "markupsafe-3.0.4-cp313-cp313-ios_13_0_x86_64_iphonesimulator.whl", hash = "sha256:7a83aa6e4805df46fed18e989d3d16f86ef60cb50bbc8d9ce3a6be89165fbf6e"},
+    {file = "markupsafe-3.0.4-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:2d1b7d9308288661f56672b1b157d75fc536714d3638487bbea17b6318a78248"},
+    {file = "markupsafe-3.0.4-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:73e77980c7207854f00fc4e71fb1626868d5740ab4012623d55c7a99ad122a72"},
+    {file = "markupsafe-3.0.4-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:7018d4af1cd272e847aa5917983ab5e83e4f6579f9dbfecd4a79c0ca80b144c2"},
+    {file = "markupsafe-3.0.4-cp313-cp313-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:c90d5b3d4e944e065a301d741b3c1d784f6bd1f503aa68b4967e32b2ba313d85"},
+    {file = "markupsafe-3.0.4-cp313-cp313-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:18a801868a884f216e784d7d14db2a4077143ce7610440aee2ce8f734e7cfcde"},
+    {file = "markupsafe-3.0.4-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:434139499bb20b502ed3baa1f169e618f924a97e7a777fea1a49446d80106cf6"},
+    {file = "markupsafe-3.0.4-cp313-cp313-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:9e227f3dbe6bde7491cf0a9965d00b88c6b1a4a95d11480ddf88bb96d397c19f"},
+    {file = "markupsafe-3.0.4-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:b8cd1f918b26fd7b1832ece557cc18f2d8747309ff8b3f0ef9d4250c5ad67a39"},
+    {file = "markupsafe-3.0.4-cp313-cp313-musllinux_1_2_armv7l.whl", hash = "sha256:a5fcffb37e602b0b3c1638a97746b9b96125caa9bcf6fa41d337a9261de231ee"},
+    {file = "markupsafe-3.0.4-cp313-cp313-musllinux_1_2_ppc64le.whl", hash = "sha256:5989cb26b2e1efc6a42216a9f6b5ee495ce5ace2e5b352a9af489976b32d1ee2"},
+    {file = "markupsafe-3.0.4-cp313-cp313-musllinux_1_2_riscv64.whl", hash = "sha256:add96447a86d205ab616665d53b2950ee81083757f56e6ea833c8b2917646b46"},
+    {file = "markupsafe-3.0.4-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:2628d3a8cb648ecebb3c5d6b0a1052d400e4d8b7ac0fb786be8d285b50040d17"},
+    {file = "markupsafe-3.0.4-cp313-cp313-win32.whl", hash = "sha256:672d207103e6b16ca098611b0f9efad6bc00afd47c03d6ef62186495ca677dc0"},
+    {file = "markupsafe-3.0.4-cp313-cp313-win_amd64.whl", hash = "sha256:1f1f9477e174582b0a1b583d60b66e1f2cf5d3fe12cee985e4aedf44766600e5"},
+    {file = "markupsafe-3.0.4-cp313-cp313-win_arm64.whl", hash = "sha256:06de8ef6331f6e822c28d577dc8bf43fe398800477c49498f38fc38b67ff33fc"},
+    {file = "markupsafe-3.0.4-cp314-cp314-android_24_arm64_v8a.whl", hash = "sha256:4ed644d75aa94a2baf7ec3a96eaa160ea58c742eb9d27c6506053c5c40fc84ed"},
+    {file = "markupsafe-3.0.4-cp314-cp314-android_24_x86_64.whl", hash = "sha256:6d2a9efe686f9de00d0d1ea32a4a5a86d558a2277501bd78d964214eab625e59"},
+    {file = "markupsafe-3.0.4-cp314-cp314-ios_13_0_arm64_iphoneos.whl", hash = "sha256:8781a792a070cf2bd1b86d3aa943894115faaba6e88122a7bf32d62072742453"},
+    {file = "markupsafe-3.0.4-cp314-cp314-ios_13_0_arm64_iphonesimulator.whl", hash = "sha256:971a3bbb75d97ae4e2e8f7d4834236f86f85f0c85e04ab2e191db1123b04f80b"},
+    {file = "markupsafe-3.0.4-cp314-cp314-ios_13_0_x86_64_iphonesimulator.whl", hash = "sha256:8909c2f1c6dd65e054ac4b573a91c8384d1492281e55d82d159d653f7a13adf6"},
+    {file = "markupsafe-3.0.4-cp314-cp314-macosx_10_15_x86_64.whl", hash = "sha256:4cf3468d5ec187ffffcaca8e61929a37448f215dafc1386a12c750a72fe53634"},
+    {file = "markupsafe-3.0.4-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:52704c5d36eb6dda8866493decd61111fff86244c9b1ad225ca01b9e91e5970f"},
+    {file = "markupsafe-3.0.4-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:1caa2fa5a6184fb233153b35f654e6687bd555476f6170f29d8ee9be1a8b0af9"},
+    {file = "markupsafe-3.0.4-cp314-cp314-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:387d8cd30e69b3f0a72877b9ae717033396404e19095b17fe89753a981fda44f"},
+    {file = "markupsafe-3.0.4-cp314-cp314-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:051417f74bcaaefa316276e0ff723f541616ca51043d070da00249d9bddd3e3c"},
+    {file = "markupsafe-3.0.4-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:a8e9f292fcda89b324f2f5c91d13f1424a153e40fc2756f38ee23b15835ff300"},
+    {file = "markupsafe-3.0.4-cp314-cp314-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:df1ae86ff54725a01fa1a0510b914ca53a161b7050be74f6204e24aded5971d0"},
+    {file = "markupsafe-3.0.4-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:8965520ac587c94a4ac48b729be3d8b8de00af39699b17585dfb599babe77977"},
+    {file = "markupsafe-3.0.4-cp314-cp314-musllinux_1_2_armv7l.whl", hash = "sha256:340cbb1957ba99929cbf19a75626d36ba1ae21d1730b287d1cf7f824a20c4fc7"},
+    {file = "markupsafe-3.0.4-cp314-cp314-musllinux_1_2_ppc64le.whl", hash = "sha256:3a93d9616ddecfb393727a0041a562cf0b15a244e20f2bd25efc7949be4c4f17"},
+    {file = "markupsafe-3.0.4-cp314-cp314-musllinux_1_2_riscv64.whl", hash = "sha256:d2e56fd3b00222722abfb3f5f0759ddbae4b90811b5ad4343c64030ad1bde70c"},
+    {file = "markupsafe-3.0.4-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:0d9c47709875fdb321452056622e930c52afbc07a7d780762fbb8b4d91ce6fa4"},
+    {file = "markupsafe-3.0.4-cp314-cp314-win32.whl", hash = "sha256:38fc55594dab834470b6733dead2ee9e3f657fb0608c769dcafa0ba5ab52f45c"},
+    {file = "markupsafe-3.0.4-cp314-cp314-win_amd64.whl", hash = "sha256:c1bc67752d5f21013cfe430df4062441714eab79f65a6a05e01505957e9c35fe"},
+    {file = "markupsafe-3.0.4-cp314-cp314-win_arm64.whl", hash = "sha256:7e1636da3d8dfc220b6dd10264db5f2b165e4888c4518594898fbe381049af8a"},
+    {file = "markupsafe-3.0.4-cp314-cp314t-macosx_10_15_x86_64.whl", hash = "sha256:805c8b84534fa10891890f0e4be39f3a99e94615d93e8836bf9fa1fdca2feeb2"},
+    {file = "markupsafe-3.0.4-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:fa95848c929b6a75f6848d3c9793e59db365ee436776e57db835cdbfa79ba977"},
+    {file = "markupsafe-3.0.4-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:e916035e3e9930cbdfdd10abf48861340221857f45509565898e012263f7b289"},
+    {file = "markupsafe-3.0.4-cp314-cp314t-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:b4d12837e0203bbace818ff4a7461afdcd78bcd782351cea148139180d7bcffe"},
+    {file = "markupsafe-3.0.4-cp314-cp314t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:5086f9975abb1ab531ee6afca1761e4b59a19b446f3f6522ed776963228cfe5a"},
+    {file = "markupsafe-3.0.4-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:b4a635a0487774f841cb1fb62e907e7195cc95bc761e053184b8acc3ceb20733"},
+    {file = "markupsafe-3.0.4-cp314-cp314t-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:cb96e6e088d6cf71c1ea977510948320234824cf226e32f6f6e044f7a9c82b34"},
+    {file = "markupsafe-3.0.4-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:8b5d563170ff8ba3181caa967c99a3c804d1dedb702c7cb93a6a7c32247da978"},
+    {file = "markupsafe-3.0.4-cp314-cp314t-musllinux_1_2_armv7l.whl", hash = "sha256:396ec4e65cc889f69786b3b89478b471cee5a3bcf468b9d9bb03e1a30fb291fc"},
+    {file = "markupsafe-3.0.4-cp314-cp314t-musllinux_1_2_ppc64le.whl", hash = "sha256:15ba9e28640feef770374b116a6f019c21f52404aeabe516aa7f800587b98cfc"},
+    {file = "markupsafe-3.0.4-cp314-cp314t-musllinux_1_2_riscv64.whl", hash = "sha256:d920abdfa61279ba1a2ef9484aab07bf03331f8c08a10120fa332353d06e6932"},
+    {file = "markupsafe-3.0.4-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:a9f54054101545a9a9cccefddf54316aa6e4491611fcbef9e91b3b6bebec04f6"},
+    {file = "markupsafe-3.0.4-cp314-cp314t-win32.whl", hash = "sha256:12a606a492de952afcb43b59a14aaaaad120e708d3663dd0fdf2d738d427a691"},
+    {file = "markupsafe-3.0.4-cp314-cp314t-win_amd64.whl", hash = "sha256:a18f38cafc329bac5e3c2b96c765b4c96d3d103421ed22ab7988c1e3fce27464"},
+    {file = "markupsafe-3.0.4-cp314-cp314t-win_arm64.whl", hash = "sha256:eba154571c16e032112afac0dc2dfe9e63c2ceb7aedd07bb7eecf2ce26d4dd4c"},
+    {file = "markupsafe-3.0.4-cp315-cp315-android_24_arm64_v8a.whl", hash = "sha256:737c9c3981998eba27f11786f84fddcbabc74068b72a4a1f454ea02094b57b65"},
+    {file = "markupsafe-3.0.4-cp315-cp315-android_24_x86_64.whl", hash = "sha256:489505b03f692c3f376394e49194fa7a7f9e8558d6e293a7056a0032b0c38163"},
+    {file = "markupsafe-3.0.4-cp315-cp315-ios_13_0_arm64_iphoneos.whl", hash = "sha256:077293e425f28ec737dbcad442a71752e28f8ae27cde3d68acd1fb212091cd92"},
+    {file = "markupsafe-3.0.4-cp315-cp315-ios_13_0_arm64_iphonesimulator.whl", hash = "sha256:9348cbb300d224fe3b89793262cb093504d4ae927004468463f745188a193e4a"},
+    {file = "markupsafe-3.0.4-cp315-cp315-ios_13_0_x86_64_iphonesimulator.whl", hash = "sha256:b807e598953730f82e4eae3bd30f6a122cf6b31c398c6b504c0e04c13c170429"},
+    {file = "markupsafe-3.0.4-cp315-cp315-macosx_10_15_x86_64.whl", hash = "sha256:799c39bdf5e2f1292fedd3009f7b3c9e760f10b2420cb9638d56920840ff6db8"},
+    {file = "markupsafe-3.0.4-cp315-cp315-macosx_11_0_arm64.whl", hash = "sha256:ae9dcb8fbe244cb82f8a6458b455b927a03685e383d9bacf1ea5ce180b96dc97"},
+    {file = "markupsafe-3.0.4-cp315-cp315-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:4bced6e2a6dba6a28f7dd3c6ce14df1b2dd495923f16ea484cad03decd463b2b"},
+    {file = "markupsafe-3.0.4-cp315-cp315-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:3882fb412298575bae3b9c46868251f15cc69307359f87bb1b382e53d6e5a2c9"},
+    {file = "markupsafe-3.0.4-cp315-cp315-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:04e7902ba80ee4bac1d50a549606527a1dcf0476cd81403db41099d3b60ec653"},
+    {file = "markupsafe-3.0.4-cp315-cp315-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:925f929d6b59a8b3f8b8c6ac363cd0af7eecc81efb3071770b3c6717c450a369"},
+    {file = "markupsafe-3.0.4-cp315-cp315-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:f68edfc67aabac33708941f26f22a7b8e9f81429bc0cf249fcf7d66b23af8d19"},
+    {file = "markupsafe-3.0.4-cp315-cp315-musllinux_1_2_aarch64.whl", hash = "sha256:e5c802729725bd07e2bc3ab7b76dc7e0bbfc53129d8f1eb1c002c24cf774717e"},
+    {file = "markupsafe-3.0.4-cp315-cp315-musllinux_1_2_armv7l.whl", hash = "sha256:55ffd6ce583d97dc71dc92e930324c8c0d25aea7e3ade6ae54ef77cedb096811"},
+    {file = "markupsafe-3.0.4-cp315-cp315-musllinux_1_2_ppc64le.whl", hash = "sha256:2cb3dd71fc6be918ad4264346a8ed69485f9b7ed7bf35495d8e22807cd6b8bea"},
+    {file = "markupsafe-3.0.4-cp315-cp315-musllinux_1_2_riscv64.whl", hash = "sha256:94f5407f7bc64fa6463906b896f9904beeeb7dd8dc116ee8e9056c8714ff9916"},
+    {file = "markupsafe-3.0.4-cp315-cp315-musllinux_1_2_x86_64.whl", hash = "sha256:2dad610540cb2e6272855c178f08ae9a1c7ac258a7fb71660553a5f104b42741"},
+    {file = "markupsafe-3.0.4-cp315-cp315-win32.whl", hash = "sha256:03470d1a8268e692ecf79ecd565593e59d44219377a7ead61f1f1b94c1f7ff6b"},
+    {file = "markupsafe-3.0.4-cp315-cp315-win_amd64.whl", hash = "sha256:d882a373d8093c2941e01291b7ced96e9cbe4781da9a7751ca7e6c70385e5214"},
+    {file = "markupsafe-3.0.4-cp315-cp315-win_arm64.whl", hash = "sha256:353bd63081912ab8cfa6a0c7d185934cdf8426f04c618bba6bc4b394f2069b67"},
+    {file = "markupsafe-3.0.4-cp315-cp315t-macosx_10_15_x86_64.whl", hash = "sha256:c61750fadcd119d0825bcb7d7d675dd264dcc89cc05292aab5be68ebdbb374ad"},
+    {file = "markupsafe-3.0.4-cp315-cp315t-macosx_11_0_arm64.whl", hash = "sha256:1c0df495a977d10460a94941799c72d5b5ab03d3858d949b55b5a66c8f371c99"},
+    {file = "markupsafe-3.0.4-cp315-cp315t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:02fa4acbc6a3fc5c693c34d4dd8c1130b7fe99cc915181b0ddd6f72aeb296002"},
+    {file = "markupsafe-3.0.4-cp315-cp315t-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:05295589e619b9bed252a86b532b8e27350abc372d18ba89b59375325e91ec1e"},
+    {file = "markupsafe-3.0.4-cp315-cp315t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:be6cb0c799abb0e2ba3e618e6d28ddddf7e485f6c2ce938dfa237daf3905072c"},
+    {file = "markupsafe-3.0.4-cp315-cp315t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:26e9867520db70d37f7fb421a7f0d8adb40171011fb84ce869afa1a83370dfa8"},
+    {file = "markupsafe-3.0.4-cp315-cp315t-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:f03460ff076f70ab595bb45a0205ccea1971443575b6920c52e755dec2b3fbfe"},
+    {file = "markupsafe-3.0.4-cp315-cp315t-musllinux_1_2_aarch64.whl", hash = "sha256:436e3ffc6310d3c41878c601db29098102fe5d8a467c49da4a4125254e0980f2"},
+    {file = "markupsafe-3.0.4-cp315-cp315t-musllinux_1_2_armv7l.whl", hash = "sha256:4e2c4809c14559aa7ef426f27fb35afbb38104c349a903bf8f3600456764bb38"},
+    {file = "markupsafe-3.0.4-cp315-cp315t-musllinux_1_2_ppc64le.whl", hash = "sha256:da2af0d7aebfc2074080d72efa6ab8317c62481ef1f896f65d9999c1c01f4494"},
+    {file = "markupsafe-3.0.4-cp315-cp315t-musllinux_1_2_riscv64.whl", hash = "sha256:aa2c838cc024642cc04c6854232f32b43e5e22833dd11119c1766c7873b8370d"},
+    {file = "markupsafe-3.0.4-cp315-cp315t-musllinux_1_2_x86_64.whl", hash = "sha256:b91cc9d336957239ff200f30097e6fea2dc6d6fb3c81e853eaa09eac904fd894"},
+    {file = "markupsafe-3.0.4-cp315-cp315t-win32.whl", hash = "sha256:e49fb0d1ce92cfa0cb198cc5b1b11cdf9d0638658e2a2db2687e39db7c87fc78"},
+    {file = "markupsafe-3.0.4-cp315-cp315t-win_amd64.whl", hash = "sha256:4f6e0852a0283b1b1fd776eeb7b766a5f440b3e2bd31ab51af3b400585f3965c"},
+    {file = "markupsafe-3.0.4-cp315-cp315t-win_arm64.whl", hash = "sha256:39dbacefc411633db5b4378b066a9aca70a3d7e2922c9e578d825f844026eeba"},
+    {file = "markupsafe-3.0.4-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:f291bcf42ae98eb5107edb162c3c998b4a89648fd8e99ed4cbd12705292788cd"},
+    {file = "markupsafe-3.0.4-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:ac0c7c9f1609b0c4c114feb1d7a3409564c7fb77e360bed9e97e5d25dfeaf868"},
+    {file = "markupsafe-3.0.4-cp39-cp39-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:6768d67d1bce64270e0fdc2e69309d68b9b18ae56ddf6c711d168e9d051c2cac"},
+    {file = "markupsafe-3.0.4-cp39-cp39-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:14bd2d845d62ab678eaf81da89d7b621b51756c72346745c1a594c09d49207a2"},
+    {file = "markupsafe-3.0.4-cp39-cp39-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:007e1ffd9bf65bb6ee96df7b258fc632a4868dd5566037986c64781f35a36e98"},
+    {file = "markupsafe-3.0.4-cp39-cp39-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:5e8b3d0b18fd623afa12ecb2ce8d8becef69f9b5440c6330c7972200e0bb84b0"},
+    {file = "markupsafe-3.0.4-cp39-cp39-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:57f9947a7e57a081c1e3e0a2dd0d2dcf290a4531450e6f611e30084c222a7295"},
+    {file = "markupsafe-3.0.4-cp39-cp39-musllinux_1_2_aarch64.whl", hash = "sha256:b61687d0828e72bf5cda24a2690188f37170bd31c9359ac97e4e66569f120a16"},
+    {file = "markupsafe-3.0.4-cp39-cp39-musllinux_1_2_armv7l.whl", hash = "sha256:0cee7cb0f9a1b6892ea482237d9403b3d1b4603aee057d0ff01f0fac2d019a97"},
+    {file = "markupsafe-3.0.4-cp39-cp39-musllinux_1_2_ppc64le.whl", hash = "sha256:94e4c421742086aeee4c32a506eec8859d7634aad943f7e6aacf70f813478768"},
+    {file = "markupsafe-3.0.4-cp39-cp39-musllinux_1_2_riscv64.whl", hash = "sha256:9240187afb63d2f9ddc3e032c670356fe941f6e20662ea168a5dc3f1f317e1b3"},
+    {file = "markupsafe-3.0.4-cp39-cp39-musllinux_1_2_x86_64.whl", hash = "sha256:e841068dc0be4cb6dfb5c890eb88cbdcff2f4a332393c7ec94e8e618bd32c1a8"},
+    {file = "markupsafe-3.0.4-cp39-cp39-win32.whl", hash = "sha256:f61efe1d2fe0de16158a5fe1d1cf3c14bdb6aecd54d8938fd26512c525c1f624"},
+    {file = "markupsafe-3.0.4-cp39-cp39-win_amd64.whl", hash = "sha256:2b2b1e18af909b448bb3cf9e3433366f7a8726271fc214e8b10e0f62a78c724b"},
+    {file = "markupsafe-3.0.4-cp39-cp39-win_arm64.whl", hash = "sha256:6669c1bf34080161ce49c589cc512ef24d4c704ac9d2b2d3667f519c60418378"},
+    {file = "markupsafe-3.0.4.tar.gz", hash = "sha256:2e9ad7dd851bf45fab9f75cbff4cb493fee9979e8d8c7c9c3ee119022518edd6"},
+]
+
+[[package]]
+name = "mdurl"
+version = "0.1.2"
+description = "Markdown URL utilities"
+optional = false
+python-versions = ">=3.7"
+groups = ["main"]
+files = [
+    {file = "mdurl-0.1.2-py3-none-any.whl", hash = "sha256:84008a41e51615a49fc9966191ff91509e3c40b939176e643fd50a5c2196b8f8"},
+    {file = "mdurl-0.1.2.tar.gz", hash = "sha256:bb413d29f5eea38f31dd4754dd7377d4465116fb207585f97bf925588687c1ba"},
+]
+
+[[package]]
+name = "mpmath"
+version = "1.3.0"
+description = "Python library for arbitrary-precision floating-point arithmetic"
+optional = false
+python-versions = "*"
+groups = ["main"]
+files = [
+    {file = "mpmath-1.3.0-py3-none-any.whl", hash = "sha256:a0b2b9fe80bbcd81a6647ff13108738cfb482d481d826cc0e02f5b35e5c88d2c"},
+    {file = "mpmath-1.3.0.tar.gz", hash = "sha256:7a28eb2a9774d00c7bc92411c19a89209d5da7c4c9a9e227be8330a23a25b91f"},
+]
+
+[package.extras]
+develop = ["codecov", "pycodestyle", "pytest (>=4.6)", "pytest-cov", "wheel"]
+docs = ["sphinx"]
+gmpy = ["gmpy2 (>=2.1.0a4) ; platform_python_implementation != \"PyPy\""]
+tests = ["pytest (>=4.6)"]
+
+[[package]]
+name = "multidict"
+version = "6.9.1"
+description = "multidict implementation"
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "multidict-6.9.1-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:aef74e9beabbd6c4aafc091dabff86d046ccf013ce1e4396c0fbb01b4cad9de8"},
+    {file = "multidict-6.9.1-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:1ca5ebe6d454f1e5496cf052386a559f1080bf2de75bf327ebca0a6003b79f19"},
+    {file = "multidict-6.9.1-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:02f6d0c4b70f783305e73f9944d8efe6be1022550f0974ba0ae9d8893c0350fa"},
+    {file = "multidict-6.9.1-cp310-cp310-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:8050e75af7e4c6e2d5260b84eeedb618f3e452e66473432e085b5d1b81429299"},
+    {file = "multidict-6.9.1-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:4736d350371337825cac1793c9f7c40701c32a03912548c2a7608e51679cbb96"},
+    {file = "multidict-6.9.1-cp310-cp310-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:02fe09dc197b8ae7e355371e51e5dce2f39060cd5a8badf94904527e23a3f188"},
+    {file = "multidict-6.9.1-cp310-cp310-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:ad4528cdce058b684f75fad1faf4a6a6c992fe2f08376370ca66e4ce5916a84a"},
+    {file = "multidict-6.9.1-cp310-cp310-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:8885e3808aedbd6725b921fb67dacaae0678933561ddd47c2b01315198c70e2e"},
+    {file = "multidict-6.9.1-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:d3b6e6840421c83ccb60398e333b44f910b0907bb409597685ab2eedd1e22eab"},
+    {file = "multidict-6.9.1-cp310-cp310-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:c9648ed33dc8179e4ec04bbc73bd7f0038e1e81217a69467f61f02a78bf07e88"},
+    {file = "multidict-6.9.1-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:d9d6544790ba50438a9c1a903c3c4afb1ec8a7832db5518549275b40ef0dd4b1"},
+    {file = "multidict-6.9.1-cp310-cp310-musllinux_1_2_armv7l.whl", hash = "sha256:1c9f3c25df6c9d3bbae4f6fd3f514b1c5c740a2110f56ccf36069c85417e289c"},
+    {file = "multidict-6.9.1-cp310-cp310-musllinux_1_2_i686.whl", hash = "sha256:cf22b43f35b7dbb9f71e8ee2041b5c00029cdfc28ede3a2f31cf8906f9a6c126"},
+    {file = "multidict-6.9.1-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:8ed78ccad4c7b421804f5d524b7740946a7ef75d89dc0dac52e9d7c24c472410"},
+    {file = "multidict-6.9.1-cp310-cp310-musllinux_1_2_riscv64.whl", hash = "sha256:e3b1aa25f01238886a6baed9e13b9a9240ed344346c79ae280ae65e8603b21d5"},
+    {file = "multidict-6.9.1-cp310-cp310-musllinux_1_2_s390x.whl", hash = "sha256:bc94a68ea5e18f8e85dc6b522bcb53093f692c8eb62b4837ac047da73956cbe4"},
+    {file = "multidict-6.9.1-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:73533644e1f69ea1164d56cc6505f564c6c44072b646b66d67df2d31fed0348c"},
+    {file = "multidict-6.9.1-cp310-cp310-win32.whl", hash = "sha256:66987aa68b0f7c2a1cc5f388ca962b8ed92b10f79de38d6d0d8c716154f519d9"},
+    {file = "multidict-6.9.1-cp310-cp310-win_amd64.whl", hash = "sha256:a32b78c1e52ebd8e247bb68300b90b233300d8816faa008ed0713bc539fb6af0"},
+    {file = "multidict-6.9.1-cp310-cp310-win_arm64.whl", hash = "sha256:ab64ace1a68682d191d9bedd9d4c939406ad86b1d9f628180410644249ad46c2"},
+    {file = "multidict-6.9.1-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:910d4260512660484c0dc1588a316fbb35a40c081c36fc51d1225351af17cfe4"},
+    {file = "multidict-6.9.1-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:33fa55b990f81c2927e01399ace0d18926c69d69baa8cdaa819424132fb97987"},
+    {file = "multidict-6.9.1-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:369b5aa01b241cd3fea6890bdbb11a1425d87bf1515831500d518f4223e9d72c"},
+    {file = "multidict-6.9.1-cp311-cp311-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:803f8b575a71b1b299d677c28db0653459c79b5308874efec813f17b7457c7f1"},
+    {file = "multidict-6.9.1-cp311-cp311-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:b8a65621b98984a62e59403009591b8a5a7736273aefe1cab64cfb85b365cc07"},
+    {file = "multidict-6.9.1-cp311-cp311-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:13849a1d4f54c3809ae721e9e83ab28f5ea602f33660cb84eb6ef261eac706c1"},
+    {file = "multidict-6.9.1-cp311-cp311-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:dd9a137a4a9becda3094f3831cd026380f75f6855e051eefe4c73ade524f1cc3"},
+    {file = "multidict-6.9.1-cp311-cp311-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:9ae9614c317c50836689ce2dfde07c05fe0b16378562e2221746c3913ede3c80"},
+    {file = "multidict-6.9.1-cp311-cp311-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:e8f1e362c9352b50ed120f001046fdbb80810c9d56580f4c3fc13bbe30823387"},
+    {file = "multidict-6.9.1-cp311-cp311-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:0dd655518f136febd96c05131a76a863e32fc2a1d7acd4e3c959e3ceb77d8345"},
+    {file = "multidict-6.9.1-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:2c5d675da8f1cb5650271c8ad5e95c0a3e5a183c105e72d953b12877b1c8d0fd"},
+    {file = "multidict-6.9.1-cp311-cp311-musllinux_1_2_armv7l.whl", hash = "sha256:658f90f49cf5af2441cad0a2b801c3ef520471989a1ec55bcb25b255b2ca8d2f"},
+    {file = "multidict-6.9.1-cp311-cp311-musllinux_1_2_i686.whl", hash = "sha256:43124fe172ada86d03ac3c8dc8179091341f6724d5e5d5b160e1587e4cd3761b"},
+    {file = "multidict-6.9.1-cp311-cp311-musllinux_1_2_ppc64le.whl", hash = "sha256:b2f0adc22a4eb31e545221d93fc73a0f6a8cc2379d0f4309f71d1d17ba938b82"},
+    {file = "multidict-6.9.1-cp311-cp311-musllinux_1_2_riscv64.whl", hash = "sha256:7fc59e9ba821b220944ccfe0f89c9dc4745f6d097569992356eb03869f21e953"},
+    {file = "multidict-6.9.1-cp311-cp311-musllinux_1_2_s390x.whl", hash = "sha256:87cc632c88ee5dc80e12681047839304d98ee5c9a708d686505767001c9b8b9a"},
+    {file = "multidict-6.9.1-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:b828cf64d62dc09ac183f03c1aeedd164ade96a2ce4934109452edf29de1dd13"},
+    {file = "multidict-6.9.1-cp311-cp311-win32.whl", hash = "sha256:2c1aeb92eea59d824f004341b26d5e4b47a8a441cf9726769b9a90abf9d0e08f"},
+    {file = "multidict-6.9.1-cp311-cp311-win_amd64.whl", hash = "sha256:5f89dad732280e7a10b74d40b91364f88e13c3f2c08c2ef83a8cd42f7a61af2e"},
+    {file = "multidict-6.9.1-cp311-cp311-win_arm64.whl", hash = "sha256:5800368526647146978389dfaa46da3356291e9f0fff9a4ef12e8c2bef964a0d"},
+    {file = "multidict-6.9.1-cp312-cp312-macosx_10_13_universal2.whl", hash = "sha256:29138fef49828542e859828107e42e50d0e587c513b7eb4b2d92bade2b0860fe"},
+    {file = "multidict-6.9.1-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:19e31815d41cefc365489e591d105d2baceb2f65aa75d29471fbdbda8651e006"},
+    {file = "multidict-6.9.1-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:6ed30be8918e18c8bed0a2e8b70639ecf02feb61ed00ca2e41cfcb2a50fa3f42"},
+    {file = "multidict-6.9.1-cp312-cp312-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:637f4ae36264bd7b8d9a60193acddc1d735ad52e8ed53a19931ea6d921fea8e5"},
+    {file = "multidict-6.9.1-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:35fc236507fb1b3138f0af5ecd5f94ed752d4d6d826248eae425f86204013eea"},
+    {file = "multidict-6.9.1-cp312-cp312-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:e58952f04772f59f11c6e007471449809a30165188669bca8fdb19dde40a8f24"},
+    {file = "multidict-6.9.1-cp312-cp312-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:d35a4f1c63f07fbb8c8f9946dea98b21eddf6c57421585f71d91864be3ba2a24"},
+    {file = "multidict-6.9.1-cp312-cp312-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:b5f8771aaaed7f80e84a4e471d2f29ab6721e4595075e54d03ae1ed951b2000a"},
+    {file = "multidict-6.9.1-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:976fd7689d69ec78d67d31d38d396d8adb562f7e8368279f76aed4aa451fa06d"},
+    {file = "multidict-6.9.1-cp312-cp312-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:95052e8777a86bae87c0bd0b5ab22d809e3d1d02bf69e3e66ddda5ba75a05805"},
+    {file = "multidict-6.9.1-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:1a8adfcaf96f587ab138476eaddef95f29b8a2a8a9afbfea8d2fd62180995d02"},
+    {file = "multidict-6.9.1-cp312-cp312-musllinux_1_2_armv7l.whl", hash = "sha256:1a53de2772cfb74559df2eb4456ec4eeb908435ec55a84b69370d9d745d62aa8"},
+    {file = "multidict-6.9.1-cp312-cp312-musllinux_1_2_i686.whl", hash = "sha256:0f2ce963299d42fa3f22a90adc0fdf174792ffef5ff4c7ffb68260548fb05580"},
+    {file = "multidict-6.9.1-cp312-cp312-musllinux_1_2_ppc64le.whl", hash = "sha256:5b30ddf7234e611ca877575b62840e6af5977f92f1f9d532eedbb05a44ff8004"},
+    {file = "multidict-6.9.1-cp312-cp312-musllinux_1_2_riscv64.whl", hash = "sha256:63ada7ee2e9345695f9e9bc4c65d72222253f07b1ac94fd0e37555cc6f3c7f60"},
+    {file = "multidict-6.9.1-cp312-cp312-musllinux_1_2_s390x.whl", hash = "sha256:3c95601ed98fad3f6e2f8fe809c3b526b0fab31ef525e00a155e227f3d17f58a"},
+    {file = "multidict-6.9.1-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:c148e8b596000dd3e4bfe206e70f3e666be18d72032e0012555f2373c52e35d6"},
+    {file = "multidict-6.9.1-cp312-cp312-win32.whl", hash = "sha256:f9dad513626a33670f17cddc6078e30e311f444c957e8dbfc5b2b4603c8b4edb"},
+    {file = "multidict-6.9.1-cp312-cp312-win_amd64.whl", hash = "sha256:a16a1dc8529f9e734a41c3b856f3eae7ebacdc061dde3f8a844e0c7889c97203"},
+    {file = "multidict-6.9.1-cp312-cp312-win_arm64.whl", hash = "sha256:361f7206cf341ba94fb015688f5c8b480f8e63bd58a4c14a48aeca7851a241cc"},
+    {file = "multidict-6.9.1-cp313-cp313-android_24_x86_64.whl", hash = "sha256:d7bf9e43282d69561618e8a0ea33368d532ebef42f15c096f427090521dd74f3"},
+    {file = "multidict-6.9.1-cp313-cp313-ios_13_0_arm64_iphoneos.whl", hash = "sha256:03d47df72f084f757c1cb771188d5f4e3a805e4abc4d67e32509272343ae9382"},
+    {file = "multidict-6.9.1-cp313-cp313-ios_13_0_arm64_iphonesimulator.whl", hash = "sha256:6bc94fe17c3c56e5418f79515b786b101845f70609b0d19d0c1ba13448e5633a"},
+    {file = "multidict-6.9.1-cp313-cp313-macosx_10_13_universal2.whl", hash = "sha256:8f2973bbd2bebd9d2e0cd6394c1292a1a19ccd56bdcbe1e174059f1a39be5b40"},
+    {file = "multidict-6.9.1-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:de7738b8c0bb74c4cc16bbd7fb49fc2bcf6430dba11b3432cd52768ae40933e8"},
+    {file = "multidict-6.9.1-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:e5ccad4b7bac125722f48d6f862bed3b514d8526deea06316bb72f152cd30a7c"},
+    {file = "multidict-6.9.1-cp313-cp313-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:a49ff5cdb33654cb7d6a3c377aa2a83ddefaa1db31eb10bcf3c180aa84f9af8a"},
+    {file = "multidict-6.9.1-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:b22ff30006a2f28f8bff878fb93413cbe3a4d1fd517c28081d848c90e9cfd2c8"},
+    {file = "multidict-6.9.1-cp313-cp313-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:3adf06c66041aa21eeb8a71e82379b74773298c8e6d3d839b151aae441a99b94"},
+    {file = "multidict-6.9.1-cp313-cp313-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:963a8d8f97057082679523d0fd4c53a38f86bc58cabe4556faef682ae53fa2fa"},
+    {file = "multidict-6.9.1-cp313-cp313-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:b66ccc5c2cdd26e74fa5d4c29ffae424cc6148bf93ce574821783fb3b6d452c5"},
+    {file = "multidict-6.9.1-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:7fd79c521f6290c69125fa2b85fa65d9e657e6a8ffaf722dc881b926bef4aa5c"},
+    {file = "multidict-6.9.1-cp313-cp313-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:1d804e4caf5d5da37d6dac1325da5629ebef1e27a294c2b568b295814aa36c7b"},
+    {file = "multidict-6.9.1-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:bb0d664505f4b112f384cffeee82e91e3f6448d8e574989479db4439b68cba05"},
+    {file = "multidict-6.9.1-cp313-cp313-musllinux_1_2_armv7l.whl", hash = "sha256:9e14d17773b1b3c758ff153659a1824608a0cb562c45f484b5ed8a433428444a"},
+    {file = "multidict-6.9.1-cp313-cp313-musllinux_1_2_i686.whl", hash = "sha256:083735b7f395894e43adb278d5dae901448883a835ff8f1977e285fefdb10418"},
+    {file = "multidict-6.9.1-cp313-cp313-musllinux_1_2_ppc64le.whl", hash = "sha256:b65121091567847a8cb520d364ab22ba90e00d3cc55fa9eb34bb439f0684bcd1"},
+    {file = "multidict-6.9.1-cp313-cp313-musllinux_1_2_riscv64.whl", hash = "sha256:ea999ae6e80e66ad5eea287860951b033d0104ca34d6d87c7b5125ebe0e12721"},
+    {file = "multidict-6.9.1-cp313-cp313-musllinux_1_2_s390x.whl", hash = "sha256:095d900c242e00fbe5f321ee072e7278b4153e78c5ce9c1efde167d62c1e4771"},
+    {file = "multidict-6.9.1-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:6441cc837aea58be7d9baef1b2383eb8311ab9303f500f99ac90b584cd78bb14"},
+    {file = "multidict-6.9.1-cp313-cp313-win32.whl", hash = "sha256:9c4880d017555d70dea367dd49271830842d48e3891c2da97da7ce8c4abcee40"},
+    {file = "multidict-6.9.1-cp313-cp313-win_amd64.whl", hash = "sha256:ac51cd64bae51c462ea58ad2492c9b8209667a4ef60c45c4a304518b67598d5d"},
+    {file = "multidict-6.9.1-cp313-cp313-win_arm64.whl", hash = "sha256:37a9ebe00c698279213d56e6c64e1962ab1e092918270649b397cac3dc196ca4"},
+    {file = "multidict-6.9.1-cp314-cp314-android_24_x86_64.whl", hash = "sha256:fc0dcb22fa9aeabfe3fa4e0430099acff985ec5d77a851382f76cc6146e780e5"},
+    {file = "multidict-6.9.1-cp314-cp314-ios_13_0_arm64_iphoneos.whl", hash = "sha256:024123f0ab402ab33828e24eb80fa8f25167d0d3783ba5f287e39ed741e6abf9"},
+    {file = "multidict-6.9.1-cp314-cp314-ios_13_0_arm64_iphonesimulator.whl", hash = "sha256:af1b5a92315048c3e36bbebfa7d4760a9c3e910bc4166f11b20d77d20d6bcfca"},
+    {file = "multidict-6.9.1-cp314-cp314-macosx_10_15_universal2.whl", hash = "sha256:b2483da477932ad1983d1d33c18bc3771c6fb00cfbaaed70a875fd547ef8e840"},
+    {file = "multidict-6.9.1-cp314-cp314-macosx_10_15_x86_64.whl", hash = "sha256:db4d697b18b6ef5528b1f36bfa25072cd2a421869f5963bc0e92c8a34b9e2800"},
+    {file = "multidict-6.9.1-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:854fd2f1bc6e8a56b89910b5cd7261a8b40f13ebb31572da985ce59c7da0886d"},
+    {file = "multidict-6.9.1-cp314-cp314-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:03731f6fc036180700c9dc2308205a48e5ca6f3ff03087739ab746f294022201"},
+    {file = "multidict-6.9.1-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:989261c5f1735a165f2e4e87cf6d5f17ab734fa18f9ad0383d5adfcdaefce701"},
+    {file = "multidict-6.9.1-cp314-cp314-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:b4c9e5d05b126b267ac048a89a0e2d9b48b1b648add62d4872906304a8590610"},
+    {file = "multidict-6.9.1-cp314-cp314-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:9d0a21cf76153de8f2d96a877991d6bc59b9ab5180b949e50db73cc193b4a694"},
+    {file = "multidict-6.9.1-cp314-cp314-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:e7386aa18d98d6b8af44b92173654ec469237fda35f8e8523e43b581a86f476a"},
+    {file = "multidict-6.9.1-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:b69651732c64afb691e50cdc3387cae305e0eeff8804fe3e3ce203876494932a"},
+    {file = "multidict-6.9.1-cp314-cp314-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:1fee9a16d88a1c4865610de31ef5c666671020d7a81d1f510eaf3c97d00ebeba"},
+    {file = "multidict-6.9.1-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:7814bbee202acd3bd240204c17d8b87a4c48c81064fd8674dbe527d94d5a4290"},
+    {file = "multidict-6.9.1-cp314-cp314-musllinux_1_2_armv7l.whl", hash = "sha256:95f273bea318a194f656527ee2ed19494327bc500b8e87d9358e3222579aa28d"},
+    {file = "multidict-6.9.1-cp314-cp314-musllinux_1_2_i686.whl", hash = "sha256:73bcafa21a78d0776b3ee7cd2a63c66f968eb7db8e8d594e32d7329950f6e828"},
+    {file = "multidict-6.9.1-cp314-cp314-musllinux_1_2_ppc64le.whl", hash = "sha256:1a938761c77e0e6edb0c93d02f4e988d44a69e5195e5a3b893e5553311347132"},
+    {file = "multidict-6.9.1-cp314-cp314-musllinux_1_2_riscv64.whl", hash = "sha256:37245ca4105386194dd1d292a6f2aae09bfe1bd7ac6f9ec25093e3cf8e9b143b"},
+    {file = "multidict-6.9.1-cp314-cp314-musllinux_1_2_s390x.whl", hash = "sha256:f0700527dd5bfa8b7204b08330542f4f388899d3c14d885d8a368992e0eb562d"},
+    {file = "multidict-6.9.1-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:6b7e54fd883671d1a8810704851c044b7173287c552b9f5c3d9e0eb9f00ae194"},
+    {file = "multidict-6.9.1-cp314-cp314-win32.whl", hash = "sha256:e81ae656b9935ac4528a71f96bb7a14d949778ed1897c573d3e7ebb9187f8841"},
+    {file = "multidict-6.9.1-cp314-cp314-win_amd64.whl", hash = "sha256:acddcac38adc8342ba48aba98896faa7928854bebb62542362138655b5367ee3"},
+    {file = "multidict-6.9.1-cp314-cp314-win_arm64.whl", hash = "sha256:32217133dddc58c927805cf6c0731d8144584176b768042ee886d51e71860bc9"},
+    {file = "multidict-6.9.1-cp314-cp314t-macosx_10_15_universal2.whl", hash = "sha256:0b2fb8c349d1103863750b5d8cb5ace766917f4b35f3d883c8f778853eaa9f76"},
+    {file = "multidict-6.9.1-cp314-cp314t-macosx_10_15_x86_64.whl", hash = "sha256:56d834b74c993a7d7cb2b8ab33a0d55c3e0d4a3d2f2da2808a4ad3d79189711b"},
+    {file = "multidict-6.9.1-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:783ba7d845d79ce976afd9c1e91a4e5714671defa198ee789e8b23316083a485"},
+    {file = "multidict-6.9.1-cp314-cp314t-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:4d718fa1b5f0d0dd75e86fbbc5b0c93ea3a5d65216c85c61cd5d7cdddfe08455"},
+    {file = "multidict-6.9.1-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:a313bad717dde740959d50850c315b75fd4eb0c5e6b4dc8535db0f1c369be125"},
+    {file = "multidict-6.9.1-cp314-cp314t-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:ef01d29fca550ab871fd99154f82c6472aacf8e7def272dfb07b460123850390"},
+    {file = "multidict-6.9.1-cp314-cp314t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:0f3bd290711c6e9486173a6ee7cd4e7f00c3971c7908c1ec1b6e5437c5e4c6f9"},
+    {file = "multidict-6.9.1-cp314-cp314t-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:81cdc537e0a42e3c0170752fcadbe450246d5c3b4b7231d6eb9672456605ac94"},
+    {file = "multidict-6.9.1-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:8b193bf7a443c97d81c47052f60c486071a4bdef4a573fa2514f920089414d45"},
+    {file = "multidict-6.9.1-cp314-cp314t-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:7a0c98f6a636adf0d7edd60c61589eaf52d149239af754d0bfeb0effedba53a8"},
+    {file = "multidict-6.9.1-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:6e6b7e3a1520c39a2772f414cd9ddf5995f77e4e823dfa1522af383addd465a3"},
+    {file = "multidict-6.9.1-cp314-cp314t-musllinux_1_2_armv7l.whl", hash = "sha256:ca65cced0d67a9039e93bcd98a369920e499bf98296844ff56ead01c9085321f"},
+    {file = "multidict-6.9.1-cp314-cp314t-musllinux_1_2_i686.whl", hash = "sha256:372c063f37480f62c1ae32dc3a1a0a5942b180c883f99789075a8a8ec3c4e709"},
+    {file = "multidict-6.9.1-cp314-cp314t-musllinux_1_2_ppc64le.whl", hash = "sha256:827c92145b3b976b39430129c89d213b250dacbe5fce678e9a03940e6e848983"},
+    {file = "multidict-6.9.1-cp314-cp314t-musllinux_1_2_riscv64.whl", hash = "sha256:9b0124b9c17e9890f0819b2e7a5f65ec9a2f5aabd6c8e7ad090c1675cf70dee6"},
+    {file = "multidict-6.9.1-cp314-cp314t-musllinux_1_2_s390x.whl", hash = "sha256:23136f5a564654eb61061ec6d5620a4c1ea32c8f552b65e9982a12b72bff601b"},
+    {file = "multidict-6.9.1-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:f2524ec55b3e65cbe235a8b3c36e2af3b635be02ed05e20c94e70c5c943c009e"},
+    {file = "multidict-6.9.1-cp314-cp314t-win32.whl", hash = "sha256:35ba0263bae5dd3ad5aad767cc9afc01a8598c7dae30f1b3b2de98b1b32c28bd"},
+    {file = "multidict-6.9.1-cp314-cp314t-win_amd64.whl", hash = "sha256:7c8d5882ba25ac0282258be435d8a05aa0cbcacfce15799154a338f847f159b9"},
+    {file = "multidict-6.9.1-cp314-cp314t-win_arm64.whl", hash = "sha256:ac348379cf4de5538a0a213be1532d289aa801ec5d267c5909446b9ea2f8e2c3"},
+    {file = "multidict-6.9.1-cp315-cp315-android_24_x86_64.whl", hash = "sha256:f04551dce5a7db8c9659f2e4245494c182d0663b83661803e08d46bfcae5eda1"},
+    {file = "multidict-6.9.1-cp315-cp315-ios_13_0_arm64_iphoneos.whl", hash = "sha256:c0c88085affd35c33e124e36930c5ad96aff9294ce195eaa0fd9cec962b64a82"},
+    {file = "multidict-6.9.1-cp315-cp315-ios_13_0_arm64_iphonesimulator.whl", hash = "sha256:a8b75dd3d3638d9a19f23e84af4ffab3b8940422c0df2da2a77005ef5aa3d7ea"},
+    {file = "multidict-6.9.1-cp315-cp315-macosx_10_15_universal2.whl", hash = "sha256:006c4478de0a1876f4834e14255776286f09b9846b505fe63f67f9d173a9487c"},
+    {file = "multidict-6.9.1-cp315-cp315-macosx_10_15_x86_64.whl", hash = "sha256:2784090c30a586d5b45197bd9c32f87fb927216cde302f6cfd76d76577e90f08"},
+    {file = "multidict-6.9.1-cp315-cp315-macosx_11_0_arm64.whl", hash = "sha256:7c708566da8014b120a64b1eb6d200c6c0c8cb36296383723cdb6fc82038270b"},
+    {file = "multidict-6.9.1-cp315-cp315-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:042fb0196047e786936a730bd302de83143950da45f2c16078da8f35e1cf7919"},
+    {file = "multidict-6.9.1-cp315-cp315-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:a2212a0c842c723d919ea4a22a9296cb6b244b386e4e6ea92adfc7fbf3095519"},
+    {file = "multidict-6.9.1-cp315-cp315-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:fd882aa29bf402b62bf1fd7c19fd5df4b6528cf468a908864b39368572b662a9"},
+    {file = "multidict-6.9.1-cp315-cp315-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:c57541034d12b215ab0a2bfa371d1a8a198da18176d0426c27105b9161a6862d"},
+    {file = "multidict-6.9.1-cp315-cp315-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:d73fed4e37158ff00cd271871170b79e40138db51e625fd352fa17c6acb34f67"},
+    {file = "multidict-6.9.1-cp315-cp315-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:79da2491348b30810728050b4a8ec0416f85884125c2fd44655b3d01150d9c3e"},
+    {file = "multidict-6.9.1-cp315-cp315-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:b4da208a63434d21a3df64d29758e650fc4aa8cb05848554b76949c296539cca"},
+    {file = "multidict-6.9.1-cp315-cp315-musllinux_1_2_aarch64.whl", hash = "sha256:877ca17fdcfdf5c397493a71e5ff97a87bb181417fe717fdadc77c08c09301ac"},
+    {file = "multidict-6.9.1-cp315-cp315-musllinux_1_2_armv7l.whl", hash = "sha256:40aec5299e1ed71fbb988389059da381c3c1a60e0c649acceb2a35d9b128848e"},
+    {file = "multidict-6.9.1-cp315-cp315-musllinux_1_2_i686.whl", hash = "sha256:4a409ccc42aefec904038695d5d7fd6d8f3af2721b7b6401d55135ec7d6d298e"},
+    {file = "multidict-6.9.1-cp315-cp315-musllinux_1_2_ppc64le.whl", hash = "sha256:0a5769559e3312dd96731fbe15b4abb6033368ac1cad5a98dadd21946a4c7d6c"},
+    {file = "multidict-6.9.1-cp315-cp315-musllinux_1_2_riscv64.whl", hash = "sha256:03fac50ddfd8302175b77863a015eccfae76767cda5506eef86df559ba861e1f"},
+    {file = "multidict-6.9.1-cp315-cp315-musllinux_1_2_s390x.whl", hash = "sha256:5ecace251ccfa705bf3d7d35c5032cf750f5c629809740405697bce5c118c4a4"},
+    {file = "multidict-6.9.1-cp315-cp315-musllinux_1_2_x86_64.whl", hash = "sha256:ecc89dbd4155b2f8a47f4bbd89242a35ed15e8e0ec581cab5ac65fa38407329d"},
+    {file = "multidict-6.9.1-cp315-cp315-win32.whl", hash = "sha256:a3ffe881246d28a862f1985824f484cb7361f44d6b99c4c620436ef56462f38d"},
+    {file = "multidict-6.9.1-cp315-cp315-win_amd64.whl", hash = "sha256:59c123d0e948d760a5f930f316cfefa07e8d632ab84327c0693ee6a88171154f"},
+    {file = "multidict-6.9.1-cp315-cp315-win_arm64.whl", hash = "sha256:31199204b3ced121ff5407a2c342326a5d27e3870abbf94bd80dbd2451b7bc8f"},
+    {file = "multidict-6.9.1-cp315-cp315t-macosx_10_15_universal2.whl", hash = "sha256:8879510a76940670517ea1cb589978da44b86e286ec3e50d664ef330817afce7"},
+    {file = "multidict-6.9.1-cp315-cp315t-macosx_10_15_x86_64.whl", hash = "sha256:aec65b53a07f580606593f877eefbb29a45939bfc0d3fe6e6d9f42b41b749f68"},
+    {file = "multidict-6.9.1-cp315-cp315t-macosx_11_0_arm64.whl", hash = "sha256:14c56f73e78faa1f68bbb826197cd5871994e70e841b8590829c35912ece5c64"},
+    {file = "multidict-6.9.1-cp315-cp315t-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:38c9986f9ce50c459b10de216a05f4bd7ed5ac63887d56e500a52bb464b861ce"},
+    {file = "multidict-6.9.1-cp315-cp315t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:15db6a102cbaf1949cf028ecf080aac76d20bcd29ad4e092574db6c6b7af78a5"},
+    {file = "multidict-6.9.1-cp315-cp315t-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:bd82c4977196681a499bb6ca9e462afbc5c91c1c15b6a991dfbd72733fea4dd2"},
+    {file = "multidict-6.9.1-cp315-cp315t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:bb58ba73a3f96f9a3e46b1fab69929d7edbbddb3133ee74b5c3c54074a53c4f1"},
+    {file = "multidict-6.9.1-cp315-cp315t-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:7c6eecfab7ce4cd9487ff8ba936fe38cdfd68c04faf3d5c710363c6a8e695659"},
+    {file = "multidict-6.9.1-cp315-cp315t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:9c62e71e6289d78c0108d8aeb495f8bd3cad4bc1632fedddc9297ddf287ecc20"},
+    {file = "multidict-6.9.1-cp315-cp315t-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:539c2cd5fed0947c135cd7eabaaac55f48300dfa1de0f3ca4edb5efa6606f471"},
+    {file = "multidict-6.9.1-cp315-cp315t-musllinux_1_2_aarch64.whl", hash = "sha256:3100d169ceb7bc8f05f89a6db11d1b21f119975fc26f29dc45a472e3569f0879"},
+    {file = "multidict-6.9.1-cp315-cp315t-musllinux_1_2_armv7l.whl", hash = "sha256:696477ad71385c4795e3b8e4cf10b0d2c28c2a1ca6a955e031cb1e62993e9ee3"},
+    {file = "multidict-6.9.1-cp315-cp315t-musllinux_1_2_i686.whl", hash = "sha256:f5844e7befc707367807586f550fa97e23dcfef0728f02b98c7bc498a961a5df"},
+    {file = "multidict-6.9.1-cp315-cp315t-musllinux_1_2_ppc64le.whl", hash = "sha256:abc7c2e4b47bfe6a9aea434d3fdebb9597ee636e914e92ba352f2068b9142f4c"},
+    {file = "multidict-6.9.1-cp315-cp315t-musllinux_1_2_riscv64.whl", hash = "sha256:7ab379f95caee071a37cbd8be86d4c65accc651d391f9f97748fef006f38769c"},
+    {file = "multidict-6.9.1-cp315-cp315t-musllinux_1_2_s390x.whl", hash = "sha256:557a4e1708df428ebe6c3081c83a273d275dad2446ce0d81fc648ac71afdb18d"},
+    {file = "multidict-6.9.1-cp315-cp315t-musllinux_1_2_x86_64.whl", hash = "sha256:475d04d5192eba487a3e2f935976340baa24529046e9c1c9c7a3b7bf80445ae1"},
+    {file = "multidict-6.9.1-cp315-cp315t-win32.whl", hash = "sha256:10083a8a0f4e1b26b599889e90b9802504ce5d3f7722f925bbb7ca47dd22a7c1"},
+    {file = "multidict-6.9.1-cp315-cp315t-win_amd64.whl", hash = "sha256:e96ca64383efa107262ee3949f047af5ee4f1845ba09463466c04d35a83bd3ec"},
+    {file = "multidict-6.9.1-cp315-cp315t-win_arm64.whl", hash = "sha256:501ed8b02a5990c67a91c732843609d43a6be1f7576fcdfc867331239f37fbd3"},
+    {file = "multidict-6.9.1-py3-none-any.whl", hash = "sha256:7bf6478188f4e47bf5686e8a33da4ae28bf43b1b2528d9ee144d28492bfac60b"},
+    {file = "multidict-6.9.1.tar.gz", hash = "sha256:0f06e60fa190aa7abd0914c2a766736fdc8e9f34878c4346338534b73d1b20e2"},
+]
+
+[[package]]
+name = "multiprocess"
+version = "0.70.19"
+description = "better multiprocessing and multithreading in Python"
+optional = false
+python-versions = ">=3.9"
+groups = ["main"]
+files = [
+    {file = "multiprocess-0.70.19-pp310-pypy310_pp73-macosx_10_15_x86_64.whl", hash = "sha256:02e5c35d7d6cd2bdc89c1858867f7bde4012837411023a4696c148c1bdd7c80e"},
+    {file = "multiprocess-0.70.19-pp310-pypy310_pp73-macosx_11_0_arm64.whl", hash = "sha256:79576c02d1207ec405b00cabf2c643c36070800cca433860e14539df7818b2aa"},
+    {file = "multiprocess-0.70.19-pp310-pypy310_pp73-manylinux_2_28_x86_64.whl", hash = "sha256:c6b6d78d43a03b68014ca1f0b7937d965393a670c5de7c29026beb2258f2f896"},
+    {file = "multiprocess-0.70.19-pp311-pypy311_pp73-macosx_10_15_x86_64.whl", hash = "sha256:1bbf1b69af1cf64cd05f65337d9215b88079ec819cd0ea7bac4dab84e162efe7"},
+    {file = "multiprocess-0.70.19-pp311-pypy311_pp73-macosx_11_0_arm64.whl", hash = "sha256:5be9ec7f0c1c49a4f4a6fd20d5dda4aeabc2d39a50f4ad53720f1cd02b3a7c2e"},
+    {file = "multiprocess-0.70.19-pp311-pypy311_pp73-manylinux_2_28_x86_64.whl", hash = "sha256:1c3dce098845a0db43b32a0b76a228ca059a668071cfeaa0f40c36c0b1585d45"},
+    {file = "multiprocess-0.70.19-pp39-pypy39_pp73-macosx_10_13_arm64.whl", hash = "sha256:e5e7dc3e3e1732e88c07aaec17eeb9917f9ed1107d9e60d5ab985cdc14bac43a"},
+    {file = "multiprocess-0.70.19-pp39-pypy39_pp73-macosx_10_13_x86_64.whl", hash = "sha256:e6c0674d34b8adac22533f6786576b3de4e396aaeda9e0c15378af9b8ada2702"},
+    {file = "multiprocess-0.70.19-pp39-pypy39_pp73-manylinux_2_28_x86_64.whl", hash = "sha256:d6db91ca6391eebc139c352f34578cea382df6bfa03d3b4146ed12b18b01cc14"},
+    {file = "multiprocess-0.70.19-py310-none-any.whl", hash = "sha256:97404393419dcb2a8385910864eedf47a3cadf82c66345b44f036420eb0b5d87"},
+    {file = "multiprocess-0.70.19-py311-none-any.whl", hash = "sha256:928851ae7973aea4ce0eaf330bbdafb2e01398a91518d5c8818802845564f45c"},
+    {file = "multiprocess-0.70.19-py312-none-any.whl", hash = "sha256:3a56c0e85dd5025161bac5ce138dcac1e49174c7d8e74596537e729fd5c53c28"},
+    {file = "multiprocess-0.70.19-py313-none-any.whl", hash = "sha256:8d5eb4ec5017ba2fab4e34a747c6d2c2b6fecfe9e7236e77988db91580ada952"},
+    {file = "multiprocess-0.70.19-py314-none-any.whl", hash = "sha256:e8cc7fbdff15c0613f0a1f1f8744bef961b0a164c0ca29bdff53e9d2d93c5e5f"},
+    {file = "multiprocess-0.70.19-py39-none-any.whl", hash = "sha256:0d4b4397ed669d371c81dcd1ef33fd384a44d6c3de1bd0ca7ac06d837720d3c5"},
+    {file = "multiprocess-0.70.19.tar.gz", hash = "sha256:952021e0e6c55a4a9fe4cd787895b86e239a40e76802a789d6305398d3975897"},
+]
+
+[package.dependencies]
+dill = ">=0.4.1"
+
+[[package]]
+name = "mypy"
+version = "2.4.0"
+description = "Optional static typing for Python"
+optional = false
+python-versions = ">=3.10"
+groups = ["dev"]
+files = [
+    {file = "mypy-2.4.0-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:ffda5244fd1ad71a1e54405e35f50d09b80c3978efa120f58bd1252ae32c62d2"},
+    {file = "mypy-2.4.0-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:615b03922d40e186fd1df73156473db0ba525c639bb4fd1cf28e1694879c9b23"},
+    {file = "mypy-2.4.0-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:13fa24f439c0e48a290a3922fa14ccd22f2762bae99d2142931b3e40a9055080"},
+    {file = "mypy-2.4.0-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:3adef556a19eb3b630bf86a79c29d7da3d61e541472d0d897ca01b171c0abf8f"},
+    {file = "mypy-2.4.0-cp310-cp310-win_amd64.whl", hash = "sha256:1dc0f64b0a92ae27a49d2175f0bacfa56e15bdc5b420cf92e0c1f79292219cbb"},
+    {file = "mypy-2.4.0-cp310-cp310-win_arm64.whl", hash = "sha256:4a378fc15fb33e321f04652c166ce73eeb8833a97c3d218132844e938cd93220"},
+    {file = "mypy-2.4.0-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:5d20e6c7c35fcbf2a0ebdd0eaeacfbc243009dfd33ab7822d54e213912e6dbbd"},
+    {file = "mypy-2.4.0-cp311-cp311-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:3011537be6cf1de4511c0255a324362a812b58184bbe61e15f59c8b31033bd74"},
+    {file = "mypy-2.4.0-cp311-cp311-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:a96b07a49b7b1d025ce59c1b3acbcf24bead9a83da4523c4a6bde1bb94e7a0e1"},
+    {file = "mypy-2.4.0-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:7c4f8f8d1d1c0e2832d8ee7113dd08f6df6c7aad9e863fcbed9f25832be0b8c4"},
+    {file = "mypy-2.4.0-cp311-cp311-win_amd64.whl", hash = "sha256:ba05652540bf12828e52abae807b024b09ca144ff4f75e2450a81d69c376425b"},
+    {file = "mypy-2.4.0-cp311-cp311-win_arm64.whl", hash = "sha256:6306086b87cf7f8a29aa618d9fd9bffb56c59247166b9660fdb54d86d7714ecd"},
+    {file = "mypy-2.4.0-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:e05ff2925d8b37ad26c80c1b9dc43ae5d455da2df1e23c24c095a6425917c57e"},
+    {file = "mypy-2.4.0-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:29243242cf72582b65f9582ad9e56e8cb281566ed3519f4cd70bb8b9f2977e90"},
+    {file = "mypy-2.4.0-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:29eb0b9427a6b11b992e452f6cceb8af724f4dceb47e779d0b35e405e996ea5e"},
+    {file = "mypy-2.4.0-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:e3ebe2f72a2a1156065a9851570ffbf50c0a93cdccadef9c6e05c508a4fd10b1"},
+    {file = "mypy-2.4.0-cp312-cp312-win_amd64.whl", hash = "sha256:236e0d68f6941992b0811128e652590f590db444ab29ad8f1324765b9298b946"},
+    {file = "mypy-2.4.0-cp312-cp312-win_arm64.whl", hash = "sha256:82d0f94c8587ccb472622ee7795280aaa38a06640d5f45b3f16909d6dd86a989"},
+    {file = "mypy-2.4.0-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:7da85fbcff6dac1abcc636707bed38b45598131fb7a605d9719c70b5cc733af8"},
+    {file = "mypy-2.4.0-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:4209da39d85cf240f762af622d8180fcdfcb4727d021f44ade62d613a1a43324"},
+    {file = "mypy-2.4.0-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:6be721bd4bd57576193653b75b4af3461c9d0bf7dd8b528f782e9be210dc75bb"},
+    {file = "mypy-2.4.0-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:e1fde197ae65be856a034a91b70ed747a16562ca69577785f06c661548424bf1"},
+    {file = "mypy-2.4.0-cp313-cp313-win_amd64.whl", hash = "sha256:295ecf2e57542cd836ca537486951289678c8c7d1ee6ad74ebe29b2168a003cf"},
+    {file = "mypy-2.4.0-cp313-cp313-win_arm64.whl", hash = "sha256:bc378bdad4e9f12b5bd96466083d1e71acf00594ec9c7b2bdb5e02816f77f303"},
+    {file = "mypy-2.4.0-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:058165f564ccf559c68c70fec2091fca5891110480210c22594635e3f6683437"},
+    {file = "mypy-2.4.0-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:9fa247e02b505a45a2775f69df38d360d197e3790bc60f717595db9eda358b6e"},
+    {file = "mypy-2.4.0-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:20e9a5cd875837520c43db98dea0b6d0c2197833d95c30127d8f570fb9b1f00b"},
+    {file = "mypy-2.4.0-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:9f03a7828cca2b0adcd6662aee8f2711ff8830e1027641fdea3ab0b787483566"},
+    {file = "mypy-2.4.0-cp314-cp314-pyemscripten_2026_0_wasm32.whl", hash = "sha256:9279488933040b638c0ab739084c0ca100efeea6db581bf5d7628d8e89de53fe"},
+    {file = "mypy-2.4.0-cp314-cp314-win_amd64.whl", hash = "sha256:2106b55105ba5ea9be4f53a24517fc5fa927ff1585edc9bc1a975abb72caef89"},
+    {file = "mypy-2.4.0-cp314-cp314-win_arm64.whl", hash = "sha256:528c8744b8b5e3ecb8774f86af38d2376216816e9908317ad055f3c9c2d74799"},
+    {file = "mypy-2.4.0-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:0bb95cf34899e4619c61ab0a8667804e139e580b30d5df12af2102dfe44d0c97"},
+    {file = "mypy-2.4.0-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:86d616fe84c6eab8026f8c50ab5bcb90db780d2ccd233d971e34e92bede9b359"},
+    {file = "mypy-2.4.0-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:a3f86fd1313dd69d013e265f1fdcd12ea7a9d606f9875b2a3db946cd334555f3"},
+    {file = "mypy-2.4.0-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:720434d48542ecfe84d32d287b727569d3fc8f5769acd39051130e490a5c295c"},
+    {file = "mypy-2.4.0-cp314-cp314t-win_amd64.whl", hash = "sha256:a6e851b82c0661f69f1630fc16172c68787a6a9cf0991e7c6437d60976cdcd76"},
+    {file = "mypy-2.4.0-cp314-cp314t-win_arm64.whl", hash = "sha256:3bd0e340f0ebe65c548210f53be3fd8192e83964760caf0c28bef368e68b0d37"},
+    {file = "mypy-2.4.0-cp315-cp315-macosx_11_0_arm64.whl", hash = "sha256:afa89837d9be67e0cadfa33bca3bb7efdda98c3b07e74dc3b635ebfb1c8a926a"},
+    {file = "mypy-2.4.0-cp315-cp315-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:fb443e81057896132d3642d6be219e6efd158691ac7883e3ba8fcb469865f05d"},
+    {file = "mypy-2.4.0-cp315-cp315-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:7f38f57d344f8b6accb40e01c3d83cfc590498231724d16c07ffb7940f157818"},
+    {file = "mypy-2.4.0-cp315-cp315-musllinux_1_2_x86_64.whl", hash = "sha256:e76172710bd4e5eeae061abfd68347e5264632e02778be61784671ae3a2132f5"},
+    {file = "mypy-2.4.0-cp315-cp315-pyemscripten_2026_5_wasm32.whl", hash = "sha256:f83353e47ab520bf6fd4df8f5897d9fe081211f2fbc4b7d37736a3e3c166cbcf"},
+    {file = "mypy-2.4.0-cp315-cp315-win_amd64.whl", hash = "sha256:970b221ed5842213d98e3c480c08f795ace4b1f81fb21e1b126bd0476bce1c34"},
+    {file = "mypy-2.4.0-cp315-cp315-win_arm64.whl", hash = "sha256:502b94b0b331f7dafe32fd6b151797ddbb4f32385b362e722c783a025e5954a3"},
+    {file = "mypy-2.4.0-cp315-cp315t-macosx_11_0_arm64.whl", hash = "sha256:c9de622fd397495695d0598ddc789222bfcfec9d7c9ec3a1e385c855e3bc5e01"},
+    {file = "mypy-2.4.0-cp315-cp315t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:9f459f0b4f0596d9d51fe7716b404b35287b99e77da98a7af90a65dd5fd61141"},
+    {file = "mypy-2.4.0-cp315-cp315t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:f9b028548b3af480e2b1ed8df14ccaac86f99c9f600d1580770ab7ba3dcd40f0"},
+    {file = "mypy-2.4.0-cp315-cp315t-musllinux_1_2_x86_64.whl", hash = "sha256:cb734b2668c1f40d07ce093bbeb4407e9527c67901627b0e1679825be3f09975"},
+    {file = "mypy-2.4.0-cp315-cp315t-win_amd64.whl", hash = "sha256:172e30b8fea631fe310f0c665477f52d9ea40bb4e99e0c81dc30118563b13710"},
+    {file = "mypy-2.4.0-cp315-cp315t-win_arm64.whl", hash = "sha256:5786ef987b3767e51aaa53f20aec104c0252b42ecda7aef8e8b4cbae279b05c5"},
+    {file = "mypy-2.4.0-py3-none-any.whl", hash = "sha256:d01c5d26a352acc6d5cf3128225477e1e8465e8d3029d4c345807fbf7f3cf093"},
+    {file = "mypy-2.4.0.tar.gz", hash = "sha256:77bdaebd452f43fcfc4cc3ba94352a3ea537cd01e3f2d0879f48673d2ec00d6e"},
+]
+
+[package.dependencies]
+ast-serialize = ">=0.11.2,<1.0.0"
+librt = {version = ">=0.16.0", markers = "platform_python_implementation != \"PyPy\""}
+mypy_extensions = ">=1.0.0"
+pathspec = ">=1.0.0"
+typing_extensions = {version = ">=4.6.0", markers = "python_version < \"3.15\""}
+
+[package.extras]
+dmypy = ["psutil (>=4.0)"]
+faster-cache = ["orjson"]
+install-types = ["pip"]
+mypyc = ["setuptools (>=50)"]
+reports = ["lxml"]
+
+[[package]]
+name = "mypy-extensions"
+version = "1.1.0"
+description = "Type system extensions for programs checked with the mypy type checker."
+optional = false
+python-versions = ">=3.8"
+groups = ["dev"]
+files = [
+    {file = "mypy_extensions-1.1.0-py3-none-any.whl", hash = "sha256:1be4cccdb0f2482337c4743e60421de3a356cd97508abadd57d47403e94f5505"},
+    {file = "mypy_extensions-1.1.0.tar.gz", hash = "sha256:52e68efc3284861e772bbcd66823fde5ae21fd2fdb51c62a211403730b916558"},
+]
+
+[[package]]
+name = "narwhals"
+version = "2.26.0"
+description = "Extremely lightweight compatibility layer between dataframe libraries"
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "narwhals-2.26.0-py3-none-any.whl", hash = "sha256:29326d74f107c347fd1009bd58e38d9f7c7c5b51e6de97bc93dbc325d9038b54"},
+    {file = "narwhals-2.26.0.tar.gz", hash = "sha256:6b9cadca82f375c7e4cf584fdc86ca25da54827307a9c58f94547ee6104b82dd"},
+]
+
+[package.extras]
+cudf = ["cudf-cu12 (>=24.10.0) ; sys_platform == \"linux\""]
+dask = ["dask[dataframe] (>=2024.8)"]
+duckdb = ["duckdb (>=1.1)"]
+ibis = ["ibis-framework (>=6.0.0)", "packaging (>=21.3)", "pyarrow-hotfix (>=0.7)"]
+modin = ["modin (>=0.22.0)"]
+pandas = ["pandas (>=1.3.4)"]
+polars = ["polars (>=0.20.4)"]
+pyarrow = ["pyarrow (>=13.0.0)"]
+pyspark = ["pyspark (>=3.5.0)"]
+pyspark-connect = ["pyspark[connect] (>=3.5.0)"]
+sql = ["narwhals[duckdb]", "sqlparse (>=0.5.5)"]
+sqlframe = ["sqlframe (>=3.22.0,!=3.39.3)"]
+
+[[package]]
+name = "networkx"
+version = "3.7"
+description = "Python package for creating and manipulating graphs and networks"
+optional = false
+python-versions = "!=3.14.1,>=3.12"
+groups = ["main"]
+files = [
+    {file = "networkx-3.7-py3-none-any.whl", hash = "sha256:e3fd2c13a7814cee3746340d8d7f8598a67f16a58bf47fb7f8793fab6efca1b0"},
+    {file = "networkx-3.7.tar.gz", hash = "sha256:fd77a511bd90f39f3d016351345b52cf5319b813bdca01de3f755d3cca62e96a"},
+]
+
+[package.extras]
+benchmarking = ["asv (>=0.6.5)", "py-rattler (>=0.21,<0.22)", "virtualenv"]
+default = ["matplotlib (>=3.10)", "numpy (>=2.2)", "pandas (>=2.3)", "scipy (>=1.15)"]
+developer = ["mypy (>=1.15)", "pre-commit (>=4.1)"]
+doc = ["intersphinx-registry", "myst-nb (>=1.1)", "numpydoc (>=1.8.0)", "pillow (>=10)", "pydata-sphinx-theme (>=0.16)", "sphinx (>=8.0)", "sphinx-gallery (>=0.18)", "texext (>=0.6.7)"]
+example = ["cairocffi (>=1.7)", "contextily (>=1.6)", "hiveplotlib (>=0.28)", "igraph (>=0.11)", "iplotx (>=0.9.0)", "momepy (>=0.7.2)", "osmnx (>=2.0.0)", "scikit-learn (>=1.6)", "seaborn (>=0.13)"]
+extra = ["lxml (>=4.6)", "pydot (>=3.0.1)", "pygraphviz (>=2.0)", "sympy (>=1.10)"]
+release = ["build (>=0.10)", "changelist (==0.5)", "twine (>=4.0)", "wheel (>=0.40)"]
+test = ["pytest (>=7.2)", "pytest-cov (>=4.0)", "pytest-xdist (>=3.0)"]
+test-extras = ["pytest-mpl", "pytest-randomly"]
+
+[[package]]
+name = "numpy"
+version = "2.5.4"
+description = "Fundamental package for array computing in Python"
+optional = false
+python-versions = ">=3.12"
+groups = ["main"]
+files = [
+    {file = "numpy-2.5.4-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:c6342f54c67093cae5c0227eb0eb772fdb79f2a2c37a6eb278b9909ee06aa356"},
+    {file = "numpy-2.5.4-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:b11e8fda06a7d69f15ebf542660b74466c2e51094800c1fb794f47ad4faeef17"},
+    {file = "numpy-2.5.4-cp312-cp312-macosx_14_0_arm64.whl", hash = "sha256:9cb18a327b49c5c337f972b03682f6a49855525faaf3c0d3e9c96cd0fd8880a8"},
+    {file = "numpy-2.5.4-cp312-cp312-macosx_14_0_x86_64.whl", hash = "sha256:aec3fc4b32ff82421274f5d205c559c51c840c8df66a78efd7f3612dd005a26a"},
+    {file = "numpy-2.5.4-cp312-cp312-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:fe4d21ab149f15e4e6043dfb0de87e6e5f34ac176cde83060e9802981fca2ac2"},
+    {file = "numpy-2.5.4-cp312-cp312-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:fbde6962867ee75b48b0ee29b2b9372ec5d617799dbaf38e82dc0596f2f7738a"},
+    {file = "numpy-2.5.4-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:381a7a3d2e65e64c0ec302795ab9dc12bb1e73f150904699c153716177eebdaf"},
+    {file = "numpy-2.5.4-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:b89d0aaae2fe498c648f4c4795c084db535af5bd98ef942b2a3681fb74ce8645"},
+    {file = "numpy-2.5.4-cp312-cp312-win32.whl", hash = "sha256:9968ab7e49b93ac6e1c3b2239732183152c9150f16308d30b66a372cffe3483c"},
+    {file = "numpy-2.5.4-cp312-cp312-win_amd64.whl", hash = "sha256:a7b1b6353e36a7e50de2973a38d705c88ee93adcf120673cee7f45a4a3fa223a"},
+    {file = "numpy-2.5.4-cp312-cp312-win_arm64.whl", hash = "sha256:aa1cce2ff3f8d953de38b76bf44602caeb69f101430208f64a10067f7cb4b1d3"},
+    {file = "numpy-2.5.4-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:2377da2dd3ba2c1200956acbab2a358c83b8e1f8531191672d1cd6ad83250d53"},
+    {file = "numpy-2.5.4-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:7415db95818b39ec475a5eea54d9e3b6bc83e3912158e46da3438cdce399804d"},
+    {file = "numpy-2.5.4-cp313-cp313-macosx_14_0_arm64.whl", hash = "sha256:6d6a71b9d9a97c03633aa12565ef2825ffa036cc1d99cfd50dacf0f128af4fe2"},
+    {file = "numpy-2.5.4-cp313-cp313-macosx_14_0_x86_64.whl", hash = "sha256:d8200f16437b289a5bb927c6e184eccc3e8389bc0070fea4cd5b9e13c1757959"},
+    {file = "numpy-2.5.4-cp313-cp313-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:1c2e71b04c6cad90026e544501bbe0ab9290fa8a4d845e7e8c0d124fb429c988"},
+    {file = "numpy-2.5.4-cp313-cp313-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:6ffa07666f8da0eef81d149934a626d0d95fbd6838432a33e66245423a9062c0"},
+    {file = "numpy-2.5.4-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:2fa3328f784fc8277fc48026f6cad516f5c561c5d8e2e39b3c9e0c8f23223b34"},
+    {file = "numpy-2.5.4-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:b86966fbe4ad7de710422175572bcdc75fdedadfb54bc6fab7deabccddd7780b"},
+    {file = "numpy-2.5.4-cp313-cp313-win32.whl", hash = "sha256:5258bc06526964be5face2fc6f756857a3f24f21ec3e72ca131337a75b165d6c"},
+    {file = "numpy-2.5.4-cp313-cp313-win_amd64.whl", hash = "sha256:8b4d2fd2d34e5f8c9235ee787de5631a37a28402b15cb80814df973d2be54129"},
+    {file = "numpy-2.5.4-cp313-cp313-win_arm64.whl", hash = "sha256:bc39ac66a7a9a3fbd6134fda43136b60ffde99c8f4501e64e0d2b24da137babf"},
+    {file = "numpy-2.5.4-cp314-cp314-macosx_10_15_x86_64.whl", hash = "sha256:c668b2f0d651605b58892644b0e302c7157f7159544227758c896982ef384b18"},
+    {file = "numpy-2.5.4-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:ffa6ce09a1c6a08e9667dd9c97aa0b14184e8d18f2a14b78b2a2328c9147f076"},
+    {file = "numpy-2.5.4-cp314-cp314-macosx_14_0_arm64.whl", hash = "sha256:956555e0603a4d38019ae6925711cb9dc43195c076a928accf7ea5d50bddfe53"},
+    {file = "numpy-2.5.4-cp314-cp314-macosx_14_0_x86_64.whl", hash = "sha256:2c2c4afffdeb7920e445028dd71eb932cac3e704792e964bc2a232426d4f1255"},
+    {file = "numpy-2.5.4-cp314-cp314-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:4054173604cd8658796053f1f3bc0befb68ec1c0762c57fdad61e199256a8617"},
+    {file = "numpy-2.5.4-cp314-cp314-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:d549420b8858885cea8838a727842249218b9c1da24dd517e25c9c7a948310a3"},
+    {file = "numpy-2.5.4-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:823874a507a84af050493b622affde94b6f7c3a0dc22cb2801381bc03b871c00"},
+    {file = "numpy-2.5.4-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:4e263278bfb5ee6409db8aedbc4cc32973b1b82bc1e8d3c668551d04d83a7e37"},
+    {file = "numpy-2.5.4-cp314-cp314-win32.whl", hash = "sha256:cfd73180400042a7c532d30c5e287bdd03c59ff9ee1b4c0316af0539e29dfe23"},
+    {file = "numpy-2.5.4-cp314-cp314-win_amd64.whl", hash = "sha256:2ca144f15135b6212a5c47b1e2aeca6e412f102f95a2d5d88d8aec77eb255de3"},
+    {file = "numpy-2.5.4-cp314-cp314-win_arm64.whl", hash = "sha256:468397ba3c64427474706e5c9123fe266395496714dc684294eac75cd4930d1e"},
+    {file = "numpy-2.5.4-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:1ef3aa6d7e29bb13677323114280b05acc57607fa2300e66432d665d5418a162"},
+    {file = "numpy-2.5.4-cp314-cp314t-macosx_14_0_arm64.whl", hash = "sha256:98b053943e5a0474ec0da309d2cb9d3f18ea57f8a2067c2ab7b5f763d1068380"},
+    {file = "numpy-2.5.4-cp314-cp314t-macosx_14_0_x86_64.whl", hash = "sha256:b64a85f40e154983960a4167d4c1d57a50c7f109b3d3264a3a984154e90a8454"},
+    {file = "numpy-2.5.4-cp314-cp314t-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:a813ed7719bf45463c51779e6a98d0385fe905e48447526938a4b8337333d551"},
+    {file = "numpy-2.5.4-cp314-cp314t-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:c9b80cdf5cedba0e90d93fa5f9a333c4d65bd545cd669b71bb97ce2b703c9d73"},
+    {file = "numpy-2.5.4-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:2199ed071f460487c8db2c0e5c0b564494190edb4772fe80f9aad88b2604def5"},
+    {file = "numpy-2.5.4-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:64f9c9878c1938476365e11ccfb6b770f3b9e5f045ccddc514235041e6959365"},
+    {file = "numpy-2.5.4-cp314-cp314t-win32.whl", hash = "sha256:64d1c8ac28a4077cf987e0a71a7a0ef7e2df70722f07f0baa42dbb7eb6938647"},
+    {file = "numpy-2.5.4-cp314-cp314t-win_amd64.whl", hash = "sha256:067374eb538c34c745436365cf7b0112595c1d326f21ce4ff340f61230239fbb"},
+    {file = "numpy-2.5.4-cp314-cp314t-win_arm64.whl", hash = "sha256:e94aef2c639da4a960ad0db8e06471208d8589974953d78b61d345b4eb99e394"},
+    {file = "numpy-2.5.4-cp315-cp315-macosx_10_15_x86_64.whl", hash = "sha256:8dddfbee2e68d26d0d7d7d9cb247b1fd4409241cce32d815a11d97ec2cfde179"},
+    {file = "numpy-2.5.4-cp315-cp315-macosx_11_0_arm64.whl", hash = "sha256:81e3420b27048b65eb14c3acf0c174a8cb0e023277716110347d2dcb26026dad"},
+    {file = "numpy-2.5.4-cp315-cp315-macosx_14_0_arm64.whl", hash = "sha256:0b4724a19de67bea8cfc4970798efa78bcbbe2ac2613cfac16721a42d44de2a5"},
+    {file = "numpy-2.5.4-cp315-cp315-macosx_14_0_x86_64.whl", hash = "sha256:2132418bf8dd124a427ca9e6a1daf9ee1a87185344c95119ceae868b99466da1"},
+    {file = "numpy-2.5.4-cp315-cp315-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:325518d4245b9e331387702aa58c2ce1dc4cdcbb41dfb4ccd5dcbc7e08db1266"},
+    {file = "numpy-2.5.4-cp315-cp315-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:56733449d2544178beaa4545cee357370440cf056c197f9c7bfb19dbfdd0e86d"},
+    {file = "numpy-2.5.4-cp315-cp315-musllinux_1_2_aarch64.whl", hash = "sha256:5ec3753760c1a6d8bb91200666e545c3a9728e6269dfb5d6ce02340996698aa3"},
+    {file = "numpy-2.5.4-cp315-cp315-musllinux_1_2_x86_64.whl", hash = "sha256:b1185012870173de7ae33d370bd45b1cf5baee747ea4b97036b65f4e93016877"},
+    {file = "numpy-2.5.4-cp315-cp315-win32.whl", hash = "sha256:298eca75243f2cbbfdb460560b9fb2a1792a33cf2ab4286efd43d92e8d3df508"},
+    {file = "numpy-2.5.4-cp315-cp315-win_amd64.whl", hash = "sha256:332f3378fe077dd850e677ec01bdcc4f22368fb5d50ef10b2c79230b1bf5a592"},
+    {file = "numpy-2.5.4-cp315-cp315-win_arm64.whl", hash = "sha256:d4cccbbc78717966f764cd3af4fb70276fa01fc7a2688af11c78901fa5c04f05"},
+    {file = "numpy-2.5.4-cp315-cp315t-macosx_10_15_x86_64.whl", hash = "sha256:950ea81d57ef070665581b6e1b5f6a029306423cd1739c5b95fe78aa30db6b9d"},
+    {file = "numpy-2.5.4-cp315-cp315t-macosx_11_0_arm64.whl", hash = "sha256:c05ede731b03fb1b7591faca9389ade3267d2bddf1ad8882bb3f2cc5e101694f"},
+    {file = "numpy-2.5.4-cp315-cp315t-macosx_14_0_arm64.whl", hash = "sha256:5fbf7141bbfd63aea22f435c9062a032b9ea0082fe9845dad7f021d3f1234e71"},
+    {file = "numpy-2.5.4-cp315-cp315t-macosx_14_0_x86_64.whl", hash = "sha256:3573cd22564692a5b899ec344e5d5b9cc4576f2985b96f22af3564ed54f2710f"},
+    {file = "numpy-2.5.4-cp315-cp315t-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:6c109eac9cd439193678f69d70733c1108487546ca8eafc107b510ae10c1aecd"},
+    {file = "numpy-2.5.4-cp315-cp315t-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:80d6ef6e8620eb2c2b4c4caad50b5935d6db3cde2d51581b55dcc79e14016d1d"},
+    {file = "numpy-2.5.4-cp315-cp315t-musllinux_1_2_aarch64.whl", hash = "sha256:77045a4b175bbf5316ec08003880804336c78f92281a1b72222b274ea85ec5ac"},
+    {file = "numpy-2.5.4-cp315-cp315t-musllinux_1_2_x86_64.whl", hash = "sha256:0f02a46e49cfb6c73bdb7aea1c0d3461dbae9aba613542b65f657cd3d17b9fab"},
+    {file = "numpy-2.5.4-cp315-cp315t-win32.whl", hash = "sha256:ad62a416ddcf863bf44bba76fbf6b53366ab0692e294f51cae4b5fbe0d246788"},
+    {file = "numpy-2.5.4-cp315-cp315t-win_amd64.whl", hash = "sha256:38f47be9f74ab870d2633b5456ae519c43758a8d1fd05342f0ce4ecc034396ee"},
+    {file = "numpy-2.5.4-cp315-cp315t-win_arm64.whl", hash = "sha256:7a14a461d9340f1b46b8648578aed9cdb8b3b018a8fac6c1dde2c9192a01a87f"},
+    {file = "numpy-2.5.4.tar.gz", hash = "sha256:9a94cf751c9ad8ebaa835bcd3d40dacf8534ad086b88c38029b65123c7999d2a"},
+]
+
+[[package]]
+name = "nvidia-cublas-cu12"
+version = "12.8.4.1"
+description = "CUBLAS native runtime libraries"
+optional = false
+python-versions = ">=3"
+groups = ["main"]
+markers = "platform_system == \"Linux\" and platform_machine == \"x86_64\" and sys_platform != \"linux\""
+files = [
+    {file = "nvidia_cublas_cu12-12.8.4.1-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:b86f6dd8935884615a0683b663891d43781b819ac4f2ba2b0c9604676af346d0"},
+    {file = "nvidia_cublas_cu12-12.8.4.1-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:8ac4e771d5a348c551b2a426eda6193c19aa630236b418086020df5ba9667142"},
+    {file = "nvidia_cublas_cu12-12.8.4.1-py3-none-win_amd64.whl", hash = "sha256:47e9b82132fa8d2b4944e708049229601448aaad7e6f296f630f2d1a32de35af"},
+]
+
+[[package]]
+name = "nvidia-cuda-cupti-cu12"
+version = "12.8.90"
+description = "CUDA profiling tools runtime libs."
+optional = false
+python-versions = ">=3"
+groups = ["main"]
+markers = "platform_system == \"Linux\" and platform_machine == \"x86_64\" and sys_platform != \"linux\""
+files = [
+    {file = "nvidia_cuda_cupti_cu12-12.8.90-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:4412396548808ddfed3f17a467b104ba7751e6b58678a4b840675c56d21cf7ed"},
+    {file = "nvidia_cuda_cupti_cu12-12.8.90-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:ea0cb07ebda26bb9b29ba82cda34849e73c166c18162d3913575b0c9db9a6182"},
+    {file = "nvidia_cuda_cupti_cu12-12.8.90-py3-none-win_amd64.whl", hash = "sha256:bb479dcdf7e6d4f8b0b01b115260399bf34154a1a2e9fe11c85c517d87efd98e"},
+]
+
+[[package]]
+name = "nvidia-cuda-nvrtc-cu12"
+version = "12.8.93"
+description = "NVRTC native runtime libraries"
+optional = false
+python-versions = ">=3"
+groups = ["main"]
+markers = "platform_system == \"Linux\" and platform_machine == \"x86_64\" and sys_platform != \"linux\""
+files = [
+    {file = "nvidia_cuda_nvrtc_cu12-12.8.93-py3-none-manylinux2010_x86_64.manylinux_2_12_x86_64.whl", hash = "sha256:a7756528852ef889772a84c6cd89d41dfa74667e24cca16bb31f8f061e3e9994"},
+    {file = "nvidia_cuda_nvrtc_cu12-12.8.93-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:fc1fec1e1637854b4c0a65fb9a8346b51dd9ee69e61ebaccc82058441f15bce8"},
+    {file = "nvidia_cuda_nvrtc_cu12-12.8.93-py3-none-win_amd64.whl", hash = "sha256:7a4b6b2904850fe78e0bd179c4b655c404d4bb799ef03ddc60804247099ae909"},
+]
+
+[[package]]
+name = "nvidia-cuda-runtime-cu12"
+version = "12.8.90"
+description = "CUDA Runtime native Libraries"
+optional = false
+python-versions = ">=3"
+groups = ["main"]
+markers = "platform_system == \"Linux\" and platform_machine == \"x86_64\" and sys_platform != \"linux\""
+files = [
+    {file = "nvidia_cuda_runtime_cu12-12.8.90-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:52bf7bbee900262ffefe5e9d5a2a69a30d97e2bc5bb6cc866688caa976966e3d"},
+    {file = "nvidia_cuda_runtime_cu12-12.8.90-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:adade8dcbd0edf427b7204d480d6066d33902cab2a4707dcfc48a2d0fd44ab90"},
+    {file = "nvidia_cuda_runtime_cu12-12.8.90-py3-none-win_amd64.whl", hash = "sha256:c0c6027f01505bfed6c3b21ec546f69c687689aad5f1a377554bc6ca4aa993a8"},
+]
+
+[[package]]
+name = "nvidia-cudnn-cu12"
+version = "9.10.2.21"
+description = "cuDNN runtime libraries"
+optional = false
+python-versions = ">=3"
+groups = ["main"]
+markers = "platform_system == \"Linux\" and platform_machine == \"x86_64\" and sys_platform != \"linux\""
+files = [
+    {file = "nvidia_cudnn_cu12-9.10.2.21-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:c9132cc3f8958447b4910a1720036d9eff5928cc3179b0a51fb6d167c6cc87d8"},
+    {file = "nvidia_cudnn_cu12-9.10.2.21-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:949452be657fa16687d0930933f032835951ef0892b37d2d53824d1a84dc97a8"},
+    {file = "nvidia_cudnn_cu12-9.10.2.21-py3-none-win_amd64.whl", hash = "sha256:c6288de7d63e6cf62988f0923f96dc339cea362decb1bf5b3141883392a7d65e"},
+]
+
+[package.dependencies]
+nvidia-cublas-cu12 = "*"
+
+[[package]]
+name = "nvidia-cufft-cu12"
+version = "11.3.3.83"
+description = "CUFFT native runtime libraries"
+optional = false
+python-versions = ">=3"
+groups = ["main"]
+markers = "platform_system == \"Linux\" and platform_machine == \"x86_64\" and sys_platform != \"linux\""
+files = [
+    {file = "nvidia_cufft_cu12-11.3.3.83-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:848ef7224d6305cdb2a4df928759dca7b1201874787083b6e7550dd6765ce69a"},
+    {file = "nvidia_cufft_cu12-11.3.3.83-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:4d2dd21ec0b88cf61b62e6b43564355e5222e4a3fb394cac0db101f2dd0d4f74"},
+    {file = "nvidia_cufft_cu12-11.3.3.83-py3-none-win_amd64.whl", hash = "sha256:7a64a98ef2a7c47f905aaf8931b69a3a43f27c55530c698bb2ed7c75c0b42cb7"},
+]
+
+[package.dependencies]
+nvidia-nvjitlink-cu12 = "*"
+
+[[package]]
+name = "nvidia-cufile-cu12"
+version = "1.13.1.3"
+description = "cuFile GPUDirect libraries"
+optional = false
+python-versions = ">=3"
+groups = ["main"]
+markers = "platform_system == \"Linux\" and platform_machine == \"x86_64\" and sys_platform != \"linux\""
+files = [
+    {file = "nvidia_cufile_cu12-1.13.1.3-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:1d069003be650e131b21c932ec3d8969c1715379251f8d23a1860554b1cb24fc"},
+    {file = "nvidia_cufile_cu12-1.13.1.3-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:4beb6d4cce47c1a0f1013d72e02b0994730359e17801d395bdcbf20cfb3bb00a"},
+]
+
+[[package]]
+name = "nvidia-curand-cu12"
+version = "10.3.9.90"
+description = "CURAND native runtime libraries"
+optional = false
+python-versions = ">=3"
+groups = ["main"]
+markers = "platform_system == \"Linux\" and platform_machine == \"x86_64\" and sys_platform != \"linux\""
+files = [
+    {file = "nvidia_curand_cu12-10.3.9.90-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:dfab99248034673b779bc6decafdc3404a8a6f502462201f2f31f11354204acd"},
+    {file = "nvidia_curand_cu12-10.3.9.90-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:b32331d4f4df5d6eefa0554c565b626c7216f87a06a4f56fab27c3b68a830ec9"},
+    {file = "nvidia_curand_cu12-10.3.9.90-py3-none-win_amd64.whl", hash = "sha256:f149a8ca457277da854f89cf282d6ef43176861926c7ac85b2a0fbd237c587ec"},
+]
+
+[[package]]
+name = "nvidia-cusolver-cu12"
+version = "11.7.3.90"
+description = "CUDA solver native runtime libraries"
+optional = false
+python-versions = ">=3"
+groups = ["main"]
+markers = "platform_system == \"Linux\" and platform_machine == \"x86_64\" and sys_platform != \"linux\""
+files = [
+    {file = "nvidia_cusolver_cu12-11.7.3.90-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:db9ed69dbef9715071232caa9b69c52ac7de3a95773c2db65bdba85916e4e5c0"},
+    {file = "nvidia_cusolver_cu12-11.7.3.90-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:4376c11ad263152bd50ea295c05370360776f8c3427b30991df774f9fb26c450"},
+    {file = "nvidia_cusolver_cu12-11.7.3.90-py3-none-win_amd64.whl", hash = "sha256:4a550db115fcabc4d495eb7d39ac8b58d4ab5d8e63274d3754df1c0ad6a22d34"},
+]
+
+[package.dependencies]
+nvidia-cublas-cu12 = "*"
+nvidia-cusparse-cu12 = "*"
+nvidia-nvjitlink-cu12 = "*"
+
+[[package]]
+name = "nvidia-cusparse-cu12"
+version = "12.5.8.93"
+description = "CUSPARSE native runtime libraries"
+optional = false
+python-versions = ">=3"
+groups = ["main"]
+markers = "platform_system == \"Linux\" and platform_machine == \"x86_64\" and sys_platform != \"linux\""
+files = [
+    {file = "nvidia_cusparse_cu12-12.5.8.93-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:9b6c161cb130be1a07a27ea6923df8141f3c295852f4b260c65f18f3e0a091dc"},
+    {file = "nvidia_cusparse_cu12-12.5.8.93-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:1ec05d76bbbd8b61b06a80e1eaf8cf4959c3d4ce8e711b65ebd0443bb0ebb13b"},
+    {file = "nvidia_cusparse_cu12-12.5.8.93-py3-none-win_amd64.whl", hash = "sha256:9a33604331cb2cac199f2e7f5104dfbb8a5a898c367a53dfda9ff2acb6b6b4dd"},
+]
+
+[package.dependencies]
+nvidia-nvjitlink-cu12 = "*"
+
+[[package]]
+name = "nvidia-cusparselt-cu12"
+version = "0.7.1"
+description = "NVIDIA cuSPARSELt"
+optional = false
+python-versions = "*"
+groups = ["main"]
+markers = "platform_system == \"Linux\" and platform_machine == \"x86_64\" and sys_platform != \"linux\""
+files = [
+    {file = "nvidia_cusparselt_cu12-0.7.1-py3-none-manylinux2014_aarch64.whl", hash = "sha256:8878dce784d0fac90131b6817b607e803c36e629ba34dc5b433471382196b6a5"},
+    {file = "nvidia_cusparselt_cu12-0.7.1-py3-none-manylinux2014_x86_64.whl", hash = "sha256:f1bb701d6b930d5a7cea44c19ceb973311500847f81b634d802b7b539dc55623"},
+    {file = "nvidia_cusparselt_cu12-0.7.1-py3-none-win_amd64.whl", hash = "sha256:f67fbb5831940ec829c9117b7f33807db9f9678dc2a617fbe781cac17b4e1075"},
+]
+
+[[package]]
+name = "nvidia-nccl-cu12"
+version = "2.27.3"
+description = "NVIDIA Collective Communication Library (NCCL) Runtime"
+optional = false
+python-versions = ">=3"
+groups = ["main"]
+markers = "platform_system == \"Linux\" and platform_machine == \"x86_64\" and sys_platform != \"linux\""
+files = [
+    {file = "nvidia_nccl_cu12-2.27.3-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:9ddf1a245abc36c550870f26d537a9b6087fb2e2e3d6e0ef03374c6fd19d984f"},
+    {file = "nvidia_nccl_cu12-2.27.3-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:adf27ccf4238253e0b826bce3ff5fa532d65fc42322c8bfdfaf28024c0fbe039"},
+]
+
+[[package]]
+name = "nvidia-nvjitlink-cu12"
+version = "12.8.93"
+description = "Nvidia JIT LTO Library"
+optional = false
+python-versions = ">=3"
+groups = ["main"]
+markers = "platform_system == \"Linux\" and platform_machine == \"x86_64\" and sys_platform != \"linux\""
+files = [
+    {file = "nvidia_nvjitlink_cu12-12.8.93-py3-none-manylinux2010_x86_64.manylinux_2_12_x86_64.whl", hash = "sha256:81ff63371a7ebd6e6451970684f916be2eab07321b73c9d244dc2b4da7f73b88"},
+    {file = "nvidia_nvjitlink_cu12-12.8.93-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:adccd7161ace7261e01bb91e44e88da350895c270d23f744f0820c818b7229e7"},
+    {file = "nvidia_nvjitlink_cu12-12.8.93-py3-none-win_amd64.whl", hash = "sha256:bd93fbeeee850917903583587f4fc3a4eafa022e34572251368238ab5e6bd67f"},
+]
+
+[[package]]
+name = "nvidia-nvtx-cu12"
+version = "12.8.90"
+description = "NVIDIA Tools Extension"
+optional = false
+python-versions = ">=3"
+groups = ["main"]
+markers = "platform_system == \"Linux\" and platform_machine == \"x86_64\" and sys_platform != \"linux\""
+files = [
+    {file = "nvidia_nvtx_cu12-12.8.90-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:d7ad891da111ebafbf7e015d34879f7112832fc239ff0d7d776b6cb685274615"},
+    {file = "nvidia_nvtx_cu12-12.8.90-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:5b17e2001cc0d751a5bc2c6ec6d26ad95913324a4adb86788c944f8ce9ba441f"},
+    {file = "nvidia_nvtx_cu12-12.8.90-py3-none-win_amd64.whl", hash = "sha256:619c8304aedc69f02ea82dd244541a83c3d9d40993381b3b590f1adaed3db41e"},
+]
+
+[[package]]
+name = "packaging"
+version = "26.3"
+description = "Core utilities for Python packages"
+optional = false
+python-versions = ">=3.9"
+groups = ["main", "dev"]
+files = [
+    {file = "packaging-26.3-py3-none-any.whl", hash = "sha256:d7193f7c8e4e93f444fde0262bf90af30e16fa0ad0ad44cb553c87339b23cd1c"},
+    {file = "packaging-26.3.tar.gz", hash = "sha256:94edc256424af38762eb31306eed28beb9f0efc50a8837492c9d6fd6004aed79"},
+]
+
+[[package]]
+name = "pandas"
+version = "3.0.6"
+description = "Powerful data structures for data analysis, time series, and statistics"
+optional = false
+python-versions = ">=3.11"
+groups = ["main"]
+files = [
+    {file = "pandas-3.0.6-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:085e3786ae6b2e82b406266bce36690f72b9dc1421903ba9296b2981a9fcf586"},
+    {file = "pandas-3.0.6-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:d7564d86a94c2eb8ab290b07f63ddaae5c032fa53897c29a2ff2197d43aee8af"},
+    {file = "pandas-3.0.6-cp311-cp311-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:1e7c0afdcaf6661d795fcefc2f647ddd1136f62cdc153fba177c685d97a87808"},
+    {file = "pandas-3.0.6-cp311-cp311-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:47121f9571503f724c9b93e297ab6254ac99c77adf5e9ed085ea419fd585c258"},
+    {file = "pandas-3.0.6-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:994a79608263fe1c14cc48ffa7300e2b834b7d1cb406ffe96a08828cb0cdd79b"},
+    {file = "pandas-3.0.6-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:a3a22e07fe75347eaacc75b0e85297947af4fba6b4aae23916bd8b6828d0bba3"},
+    {file = "pandas-3.0.6-cp311-cp311-win_amd64.whl", hash = "sha256:2e5fa32ff162dfdbc280157d664f44d23049ae414725af9676df339c501d82cd"},
+    {file = "pandas-3.0.6-cp311-cp311-win_arm64.whl", hash = "sha256:5e75072773c1b2f7cb63faa3a6f562aede11f3976f68ed34cb538bc091a28171"},
+    {file = "pandas-3.0.6-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:7dac2d65e9087e8e7b5a45fe15c4920911a221df061ab629943ce016489145c7"},
+    {file = "pandas-3.0.6-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:9dab635a549e58a053c7b0fa054dc0bd7be22f0ed9a720f4a85d5fb993276172"},
+    {file = "pandas-3.0.6-cp312-cp312-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:e3dccb584123b399c07562ac4d62543e90ede49ddf8ce3c13ffc64cbe828c281"},
+    {file = "pandas-3.0.6-cp312-cp312-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:0704044b676496b8350e023b09f174a26772456c974a2b11c36bebb558c9490d"},
+    {file = "pandas-3.0.6-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:e7c1905ef02c3d6d43d9dbd5b6ccb4da4870a0b0c821bbc103fbdb6f3ad2707b"},
+    {file = "pandas-3.0.6-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:569e114072b24fc4970c12e2b4bab252671668a40b324318903380cab0254c0c"},
+    {file = "pandas-3.0.6-cp312-cp312-pyemscripten_2024_0_wasm32.whl", hash = "sha256:2a8fc94be2ee5f1d86f97aacd8cc566f81680b6498e76f3007421bb5d98151bf"},
+    {file = "pandas-3.0.6-cp312-cp312-win_amd64.whl", hash = "sha256:3ef908d28590b3f42d7070e7ad8f9b34b442b260b7f3c1afb57e0040c58cdb1b"},
+    {file = "pandas-3.0.6-cp312-cp312-win_arm64.whl", hash = "sha256:f4e7c52eb108d752e7592268108fd3e98efd76d83a3125cdd06c621c2e44359b"},
+    {file = "pandas-3.0.6-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:9ae8073aed8e21d1a7fe263dcdc6840743549722a6738198a0a46000fa9476f2"},
+    {file = "pandas-3.0.6-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:60d81f9e1799b36f3739e7fff44d1fbb2e8fd5a271b3863e03de9715fccda0fa"},
+    {file = "pandas-3.0.6-cp313-cp313-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:097090508a1dd335013d39106fc10b20f4fd4a171638e47b77d55798ed9dab6c"},
+    {file = "pandas-3.0.6-cp313-cp313-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:1e92d9fa834c7d877130027cddc0cad8dcff97c1f6cca26bd6310f847228b658"},
+    {file = "pandas-3.0.6-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:b27c8d890e4aa2171437ae2a39de1d215e674158e4865c4023a8b31c932513b2"},
+    {file = "pandas-3.0.6-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:f8029ec0f1f89e4f985929ce1f6626dabf3140d61a4e9c1215afdab34eaf9a5d"},
+    {file = "pandas-3.0.6-cp313-cp313-win_amd64.whl", hash = "sha256:f3ce8a6968045481e91a3990e797e348ce13db45ee164a7095bbc824e26c09dd"},
+    {file = "pandas-3.0.6-cp313-cp313-win_arm64.whl", hash = "sha256:cc39303913e2ea129915670de5d1c9fbd647f543bb72e5543bac8baa94e9e42f"},
+    {file = "pandas-3.0.6-cp314-cp314-macosx_10_15_x86_64.whl", hash = "sha256:ee913a91669056c1de1a6b733fbfeab711de9e54e3bee2dfa5fe79d9457247d1"},
+    {file = "pandas-3.0.6-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:ff51a4459ed036e93d1eb1bb5e6e7b28685d3cb6b7c12b91c05b31024e234729"},
+    {file = "pandas-3.0.6-cp314-cp314-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:654aae059295dbba6ecd2328ca12712a2cf1676214c8699f1c29213f7ccf9c34"},
+    {file = "pandas-3.0.6-cp314-cp314-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:62f51d7f651c8054c5e82a69265c98082e795d1442df7ca6edc3a545d61214b1"},
+    {file = "pandas-3.0.6-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:22172a92e7ee678ec0140c7af4fc9366b55413834a1cd86af78b3caa0b0574de"},
+    {file = "pandas-3.0.6-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:583be68728a31d0d750d5b8d9e00f02b153df0d4655f858bde93cb84cfc4227c"},
+    {file = "pandas-3.0.6-cp314-cp314-win_amd64.whl", hash = "sha256:77ccbe5057aece6fc172b9b77f19c04335af6882bc2e10c8f3ee4e6bfb3da553"},
+    {file = "pandas-3.0.6-cp314-cp314-win_arm64.whl", hash = "sha256:fb625f426b375bcc96e3a04c5d5d266cd7be6ae5d6866e0e703382ab5164068c"},
+    {file = "pandas-3.0.6-cp314-cp314t-macosx_10_15_x86_64.whl", hash = "sha256:9e492cd4bdba6778de4fe0df7f4590c012161ebcf9902dce01b01dc683105514"},
+    {file = "pandas-3.0.6-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:d7dcd21238cbb4828ff148481ba01cac8946dc5121457b5aeba28636f8f99a60"},
+    {file = "pandas-3.0.6-cp314-cp314t-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:6ff482fa91fa2bafd92e8fe66ce3645c851824310f295c1f0a2f96e928fc4541"},
+    {file = "pandas-3.0.6-cp314-cp314t-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:db7ec631f26223beee8e5c9e0b8f23c24d8197bbd1d982421d4e3188bea51965"},
+    {file = "pandas-3.0.6-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:bd75ed0c840f709fc2ae26ddd9534ac77ca1a48ac0cce521a74acaa85f3340a7"},
+    {file = "pandas-3.0.6-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:ef738d71d1059245b6bb03e312be06d8b3821326a83486c1ad03b9aba3710e44"},
+    {file = "pandas-3.0.6-cp314-cp314t-win_amd64.whl", hash = "sha256:429d9df32731ab01383ed98f2baa7a60368090d1a94fc06019a12062510e8630"},
+    {file = "pandas-3.0.6-cp314-cp314t-win_arm64.whl", hash = "sha256:a4dbd4dc65cbe645b92b8785d0f96dd7311010dc6606cf620e51b07b8788a12a"},
+    {file = "pandas-3.0.6-cp315-cp315-macosx_10_15_x86_64.whl", hash = "sha256:50c44cbf5820b6b91a5f74aae04972472aefadd3cd9fbd1010409d85528bd570"},
+    {file = "pandas-3.0.6-cp315-cp315-macosx_11_0_arm64.whl", hash = "sha256:eb6900de08ac85f93ac4948aa6b80842eba555875337b8359035ac9c43e92d34"},
+    {file = "pandas-3.0.6-cp315-cp315-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:4e25e2e1adee99ddfada6f7206a79ae8e9c8a8861b0e3eaaba165006d3eef18e"},
+    {file = "pandas-3.0.6-cp315-cp315-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:4ff44b2cb51cbd691c91f92c4ea6c71e34003f239ebd67c2e857dc898466b49c"},
+    {file = "pandas-3.0.6-cp315-cp315-musllinux_1_2_aarch64.whl", hash = "sha256:5edd0a7abb0986ecce1ac81f56d99b6763f86aa6946dceb6c661224f90af5a19"},
+    {file = "pandas-3.0.6-cp315-cp315-musllinux_1_2_x86_64.whl", hash = "sha256:1bcb3e9ed29e74a7439cedff9e2aefd3ea65de84d7de9ccb6c194192541bd60e"},
+    {file = "pandas-3.0.6-cp315-cp315-win_amd64.whl", hash = "sha256:253e12cb9081b0afbac607920f6142975966bc315135e09de275fdbaa415d2de"},
+    {file = "pandas-3.0.6-cp315-cp315-win_arm64.whl", hash = "sha256:97274c9adf6255bb48c620cd6959805efa7f09ea2167f0e0ae006a448cd2fca7"},
+    {file = "pandas-3.0.6-cp315-cp315t-macosx_10_15_x86_64.whl", hash = "sha256:265f562fdd1079f69f3de96dd425c3405224038c0af4f920c54bd240ee2c4640"},
+    {file = "pandas-3.0.6-cp315-cp315t-macosx_11_0_arm64.whl", hash = "sha256:c6e4aae3e9bea26c6c9a20d88d96c86ec4a99b4db5fd516bcb4e829ab2c0ee36"},
+    {file = "pandas-3.0.6-cp315-cp315t-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:a77a1a44e4d88f1c6a2a64d3eb12efec8420875722e14279800b173a7c7c2804"},
+    {file = "pandas-3.0.6-cp315-cp315t-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:86fa853a12e0b70927e2b1ee00d56d2224ec9cbb4b9d58348b5ad52d2f21150e"},
+    {file = "pandas-3.0.6-cp315-cp315t-musllinux_1_2_aarch64.whl", hash = "sha256:c826e9babb7790142c399f58599d8de679bea059d7b39c5b6efa2096fac37266"},
+    {file = "pandas-3.0.6-cp315-cp315t-musllinux_1_2_x86_64.whl", hash = "sha256:8fe77b408d82e2615674dfed62533b95e18a03610573877422aada4f625d4947"},
+    {file = "pandas-3.0.6-cp315-cp315t-win_amd64.whl", hash = "sha256:83e91d15738d7783c050197cef2f2cf82fc6353dae9865aa87ed1fa16aa4d55a"},
+    {file = "pandas-3.0.6-cp315-cp315t-win_arm64.whl", hash = "sha256:963ca21199097a84c7827c4678b04e30833084fbf8ef44fde3fa7180a29f8fa0"},
+    {file = "pandas-3.0.6.tar.gz", hash = "sha256:66b07ef7315a31bfe1089cd3d71a7de781c9dca986762d0b4fe7c0ef17465d10"},
+]
+
+[package.dependencies]
+numpy = {version = ">=1.26.0", markers = "python_version < \"3.14\""}
+python-dateutil = ">=2.8.2"
+tzdata = {version = "*", markers = "sys_platform == \"win32\" or sys_platform == \"emscripten\""}
+
+[package.extras]
+all = ["PyQt5 (>=5.15.9)", "SQLAlchemy (>=2.0.36)", "adbc-driver-postgresql (>=1.2.0)", "adbc-driver-sqlite (>=1.2.0)", "beautifulsoup4 (>=4.12.3)", "bottleneck (>=1.4.2)", "fastparquet (>=2024.11.0)", "fsspec (>=2024.10.0)", "gcsfs (>=2024.10.0)", "html5lib (>=1.1)", "hypothesis (>=6.116.0)", "jinja2 (>=3.1.5)", "lxml (>=5.3.0)", "matplotlib (>=3.9.3)", "numba (>=0.60.0)", "numexpr (>=2.10.2)", "odfpy (>=1.4.1)", "openpyxl (>=3.1.5)", "psycopg2 (>=2.9.10)", "pyarrow (>=13.0.0)", "pyiceberg (>=0.8.1)", "pymysql (>=1.1.1)", "pyreadstat (>=1.2.8)", "pytest (>=8.3.4)", "pytest-xdist (>=3.6.1)", "python-calamine (>=0.3.0)", "pytz (>=2020.1)", "pyxlsb (>=1.0.10)", "qtpy (>=2.4.2)", "s3fs (>=2024.10.0)", "scipy (>=1.14.1)", "tables (>=3.10.1)", "tabulate (>=0.9.0)", "xarray (>=2024.10.0)", "xlrd (>=2.0.1)", "xlsxwriter (>=3.2.0)", "zstandard (>=0.23.0)"]
+aws = ["s3fs (>=2024.10.0)"]
+clipboard = ["PyQt5 (>=5.15.9)", "qtpy (>=2.4.2)"]
+compression = ["zstandard (>=0.23.0)"]
+computation = ["scipy (>=1.14.1)", "xarray (>=2024.10.0)"]
+excel = ["odfpy (>=1.4.1)", "openpyxl (>=3.1.5)", "python-calamine (>=0.3.0)", "pyxlsb (>=1.0.10)", "xlrd (>=2.0.1)", "xlsxwriter (>=3.2.0)"]
+feather = ["pyarrow (>=13.0.0)"]
+fss = ["fsspec (>=2024.10.0)"]
+gcp = ["gcsfs (>=2024.10.0)"]
+hdf5 = ["tables (>=3.10.1)"]
+html = ["beautifulsoup4 (>=4.12.3)", "html5lib (>=1.1)", "lxml (>=5.3.0)"]
+iceberg = ["pyiceberg (>=0.8.1)"]
+mysql = ["SQLAlchemy (>=2.0.36)", "pymysql (>=1.1.1)"]
+output-formatting = ["jinja2 (>=3.1.5)", "tabulate (>=0.9.0)"]
+parquet = ["pyarrow (>=13.0.0)"]
+performance = ["bottleneck (>=1.4.2)", "numba (>=0.60.0)", "numexpr (>=2.10.2)"]
+plot = ["matplotlib (>=3.9.3)"]
+postgresql = ["SQLAlchemy (>=2.0.36)", "adbc-driver-postgresql (>=1.2.0)", "psycopg2 (>=2.9.10)"]
+pyarrow = ["pyarrow (>=13.0.0)"]
+spss = ["pyreadstat (>=1.2.8)"]
+sql-other = ["SQLAlchemy (>=2.0.36)", "adbc-driver-postgresql (>=1.2.0)", "adbc-driver-sqlite (>=1.2.0)"]
+test = ["hypothesis (>=6.116.0)", "pytest (>=8.3.4,<9.1)", "pytest-xdist (>=3.6.1)"]
+timezone = ["pytz (>=2020.1)"]
+xml = ["lxml (>=5.3.0)"]
+
+[[package]]
+name = "pathspec"
+version = "1.1.1"
+description = "Utility library for gitignore style pattern matching of file paths."
+optional = false
+python-versions = ">=3.9"
+groups = ["dev"]
+files = [
+    {file = "pathspec-1.1.1-py3-none-any.whl", hash = "sha256:a00ce642f577bf7f473932318056212bc4f8bfdf53128c78bbd5af0b9b20b189"},
+    {file = "pathspec-1.1.1.tar.gz", hash = "sha256:17db5ecd524104a120e173814c90367a96a98d07c45b2e10c2f3919fff91bf5a"},
+]
+
+[package.extras]
+hyperscan = ["hyperscan (>=0.7)"]
+optional = ["typing-extensions (>=4)"]
+re2 = ["google-re2 (>=1.1)"]
+
+[[package]]
+name = "peft"
+version = "0.21.2"
+description = "Parameter-Efficient Fine-Tuning (PEFT)"
+optional = false
+python-versions = ">=3.10.0"
+groups = ["main"]
+files = [
+    {file = "peft-0.21.2-py3-none-any.whl", hash = "sha256:106ab6077ff72c54d21577f9af5970e34bac7582cb14209f7b2b511e322a4eae"},
+    {file = "peft-0.21.2.tar.gz", hash = "sha256:b803ccfb3f3f316004d850284306687833a2235ea278fb56abc856203456142e"},
+]
+
+[package.dependencies]
+accelerate = ">=0.21.0"
+huggingface_hub = ">=0.25.0"
+numpy = ">=1.17"
+packaging = ">=20.0"
+psutil = "*"
+pyyaml = "*"
+safetensors = "*"
+torch = ">=1.13.0"
+tqdm = "*"
+transformers = "*"
+
+[package.extras]
+dev = ["black", "black", "griffe", "hf-doc-builder", "hf-doc-builder", "requests", "ruff (==0.16.4)"]
+docs-specific = ["black", "hf-doc-builder", "requests"]
+quality = ["black", "griffe", "hf-doc-builder", "ruff (==0.16.4)"]
+test = ["black", "black", "datasets", "diffusers", "griffe", "hf-doc-builder", "hf-doc-builder", "parameterized", "protobuf", "pytest", "pytest-cov", "pytest-xdist", "requests", "ruff (==0.16.4)", "scikit-learn", "scipy", "sentencepiece", "torchvision"]
+
+[[package]]
+name = "pluggy"
+version = "1.6.0"
+description = "plugin and hook calling mechanisms for python"
+optional = false
+python-versions = ">=3.9"
+groups = ["dev"]
+files = [
+    {file = "pluggy-1.6.0-py3-none-any.whl", hash = "sha256:e920276dd6813095e9377c0bc5566d94c932c33b27a3e3945d8389c374dd4746"},
+    {file = "pluggy-1.6.0.tar.gz", hash = "sha256:7dcc130b76258d33b90f61b658791dede3486c3e6bfb003ee5c9bfb396dd22f3"},
+]
+
+[package.extras]
+dev = ["pre-commit", "tox"]
+testing = ["coverage", "pytest", "pytest-benchmark"]
+
+[[package]]
+name = "propcache"
+version = "0.5.4"
+description = "Accelerated property cache"
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "propcache-0.5.4-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:b77c313314524ca9c38fbd70f73515d04597ac58c40c939bc0e71eeb4abff680"},
+    {file = "propcache-0.5.4-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:8f911c395cef73c510bac566da9507bb6a43e7763d0c79138dc60ee53f11207e"},
+    {file = "propcache-0.5.4-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:d83b12902eb8bce151259c86c03ba746600b2d994543de46e370cecf96c452f2"},
+    {file = "propcache-0.5.4-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:c9281e922c072158c91974d4589f1dbe0fee6d467f284c28e463f9f5a4d933f4"},
+    {file = "propcache-0.5.4-cp310-cp310-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:9f3551b8a35c1df3e7ea4d2d86edee15f0dde1bddd434a71744048683544d0ef"},
+    {file = "propcache-0.5.4-cp310-cp310-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:ec6a85f424afa8d23e0d9a094e5dbb6eda01da91c92b9183cd433768247ffc97"},
+    {file = "propcache-0.5.4-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:f574e460d1c8a08384a016fdb09ccf3543433263ed6b2f97104f979e64ea57c2"},
+    {file = "propcache-0.5.4-cp310-cp310-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:d8e017eeb7482bed34cdb0d61cf2bcfc88d104bbab296a17cd16a6af8aabc70e"},
+    {file = "propcache-0.5.4-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:f273dcf7149a50527c4fd1f55cfe9eac0f60753f5af544b4c9352578e20c0874"},
+    {file = "propcache-0.5.4-cp310-cp310-musllinux_1_2_armv7l.whl", hash = "sha256:fc2461ecc45f17893f8207e73b46ea8ba93e33630e51cf4af3fbc21d47462b1a"},
+    {file = "propcache-0.5.4-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:279655a16973f1ee2bd2fe79973137681642fd9ae0d89215bba263726eb0dc3a"},
+    {file = "propcache-0.5.4-cp310-cp310-musllinux_1_2_riscv64.whl", hash = "sha256:e9f165403b81fea7e89c932d89046a1e3d9a3a60e8d7ef2f249dccdcb0982bf5"},
+    {file = "propcache-0.5.4-cp310-cp310-musllinux_1_2_s390x.whl", hash = "sha256:1783582065a1f07f9d9ee1e992e13f15d7dc8fb1eb3a7476d43eb3f2e69d26bb"},
+    {file = "propcache-0.5.4-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:3d605bb239b796e82a81c6709548b2bd460ab73b4590cb0c83de8a2dd9694d0f"},
+    {file = "propcache-0.5.4-cp310-cp310-win32.whl", hash = "sha256:141fdbd73748db0cf7636035030aaac383d2efde8f34e7bc24594cc776d225b8"},
+    {file = "propcache-0.5.4-cp310-cp310-win_amd64.whl", hash = "sha256:146f48a9e4812611a7581003b1a39de56c34967046310c4171a68ef908c9a745"},
+    {file = "propcache-0.5.4-cp310-cp310-win_arm64.whl", hash = "sha256:6c7599df2b57ebeea8de011b5f2f7b85de95e76037d43d34b95e328430275487"},
+    {file = "propcache-0.5.4-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:897d1ddf6716e8f47200f7aad9a0efa6cc7586df66c6defa572f9eab379c078e"},
+    {file = "propcache-0.5.4-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:9cbfff4423eef4cc6cafc021469641a2b835f610b2647a6c5281903e21b8670d"},
+    {file = "propcache-0.5.4-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:3fc24f209c1b7f7f688b66b98293954f5504279760999b58920ee12dd8471c1d"},
+    {file = "propcache-0.5.4-cp311-cp311-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:62530ca89187827e4a4fe733f971abe81a7542eeea48ff61995f19b64d7199c8"},
+    {file = "propcache-0.5.4-cp311-cp311-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:56fc3f7599528db40b1efa0889a620116e2704144495273d66066e8164e45838"},
+    {file = "propcache-0.5.4-cp311-cp311-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:4f2d880ff60f45898f4acfa152aac8d04e3ee627d90ff4003491bf92239d5757"},
+    {file = "propcache-0.5.4-cp311-cp311-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:6e9368e87a3efc285e559131092c5db643eb8e56de4ee42064d5baec22ef2bb5"},
+    {file = "propcache-0.5.4-cp311-cp311-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:004e685b315646c410771836e72a44f143bbe624f29653a42687815069a303d5"},
+    {file = "propcache-0.5.4-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:594eb4c6ec35e7179b058481f4e9f02521b56de16fa577c4b85c76fb1bf8a9f8"},
+    {file = "propcache-0.5.4-cp311-cp311-musllinux_1_2_armv7l.whl", hash = "sha256:2dba2f02d2d5c09ef8a0e6c1a42aeaa451f4be9898cb00b04fe98717da2eb23b"},
+    {file = "propcache-0.5.4-cp311-cp311-musllinux_1_2_ppc64le.whl", hash = "sha256:c3ef2818d63bc86071e9d2989ae75a1bc32b8f7059cfd9f5abbbee70c32e2ed6"},
+    {file = "propcache-0.5.4-cp311-cp311-musllinux_1_2_riscv64.whl", hash = "sha256:dd2ac8f5b643454c2cc6b6118b13da16e88f4a6434fc3ba61aca384029f04f36"},
+    {file = "propcache-0.5.4-cp311-cp311-musllinux_1_2_s390x.whl", hash = "sha256:4054acf80d40456a0537f2913b349718649d8d6458a14ab7f48d0ce28c30869d"},
+    {file = "propcache-0.5.4-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:40e94adb1e7d39ff28a8bd8d8b8fbd1df6b9f40976dbe379134f1ce058e532dd"},
+    {file = "propcache-0.5.4-cp311-cp311-win32.whl", hash = "sha256:9f86f7259efe2c951f43e57d471c9b41daa5bfc7db9f67189059cf1ae6d77fd9"},
+    {file = "propcache-0.5.4-cp311-cp311-win_amd64.whl", hash = "sha256:e904d4d01f36bd6e197590be1533c44e06058771e0746dd073a8ebb3ef880858"},
+    {file = "propcache-0.5.4-cp311-cp311-win_arm64.whl", hash = "sha256:d42a9a856a4a6e2f6c10f1318c07e7daa498d6593abe745c71dae4521a26ca39"},
+    {file = "propcache-0.5.4-cp312-cp312-macosx_10_13_universal2.whl", hash = "sha256:b28f41fa3b8c6900457f858ec5b03998f3a6d535fbc1bb2edec5961ea05ec429"},
+    {file = "propcache-0.5.4-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:dcbf346a318a5e30063f547630b02bb787ce2f45b6368d5da143660b6a3835d8"},
+    {file = "propcache-0.5.4-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:87a3caecf8095e48dc72f84bfa42e23a848cf410cc9cc13031fba4869b706a21"},
+    {file = "propcache-0.5.4-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:60a64cbccaa11b7760ce705a14ada17ba459e7ca9f23ba587eb013821032d7ef"},
+    {file = "propcache-0.5.4-cp312-cp312-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:a74bfa37147cc08fb29df10bd9c16f40fa7f860cd3a6d2fff853323a94f6e17f"},
+    {file = "propcache-0.5.4-cp312-cp312-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:a4d7a54719b67338a305dca2ce6aafe366817df94ddfd4b5514374356f5ca546"},
+    {file = "propcache-0.5.4-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:2814ecd8e818f487bee4b0f921bc4d1c176cc5fc71ac0f072d0fa67eda4ac14b"},
+    {file = "propcache-0.5.4-cp312-cp312-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:6af4693716bfb03f1752ef1b30faa593db2c01d5272e9b8564a1549452a979ab"},
+    {file = "propcache-0.5.4-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:4fbc1a15dc8cd1689508758d626b372b1f09d28d9577667feaf9e6bfcd8efcbc"},
+    {file = "propcache-0.5.4-cp312-cp312-musllinux_1_2_armv7l.whl", hash = "sha256:cdee8205a44d0be91bbac4c41b95d86641b72dfc7aef1279400e4fda3f26a937"},
+    {file = "propcache-0.5.4-cp312-cp312-musllinux_1_2_ppc64le.whl", hash = "sha256:9a2a8a50a93dee0268a860a07fa3b4bd968f8ce4dbd794957da772f395368526"},
+    {file = "propcache-0.5.4-cp312-cp312-musllinux_1_2_riscv64.whl", hash = "sha256:7ffafcbfc7b549ab940047e505c831eabac5e67de53e1bc174adbc5285c55944"},
+    {file = "propcache-0.5.4-cp312-cp312-musllinux_1_2_s390x.whl", hash = "sha256:d1f5a500bfcbb2c0ab85e98a0dcd70f5899d34efe365a0187700369a79603031"},
+    {file = "propcache-0.5.4-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:8a235f73d6e020855dc29dff012d920c02ee0feab8d73a24185a7569f4be1161"},
+    {file = "propcache-0.5.4-cp312-cp312-win32.whl", hash = "sha256:b3083bfe87f95c756e610bd8025f26cbd1cd4aaa03a422f2d65efb7a97cd53d8"},
+    {file = "propcache-0.5.4-cp312-cp312-win_amd64.whl", hash = "sha256:98914de2c4d7f0f9f4a8c6ea4bf05841f4175796941e3ef7d47eb718f22311fb"},
+    {file = "propcache-0.5.4-cp312-cp312-win_arm64.whl", hash = "sha256:8876b39961e33d912afe3c1bee18ee564fdad0206f873cc15d522756b7f50737"},
+    {file = "propcache-0.5.4-cp313-cp313-macosx_10_13_universal2.whl", hash = "sha256:36c0d9db44b523ef93d03341b1c42d69ff01d673c053d1b1c6c3a363bcaa39ba"},
+    {file = "propcache-0.5.4-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:e1d52a05dc417279f7e5c7618c5dfbbc29923aaf9bc0a5c1802ddcebf54c61a0"},
+    {file = "propcache-0.5.4-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:44149f46500a0a41b95b4d99c2e586a77319539730607b9892974a092788b111"},
+    {file = "propcache-0.5.4-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:dbab5f5ff6897c81f355d079010cdae85b02e5a0b518b5251523b8ad8ae9ac3c"},
+    {file = "propcache-0.5.4-cp313-cp313-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:c3e98c55bde2bcf7db3c70d1aed7ae9aa8aebbf19a250c66645cde44cdb8b867"},
+    {file = "propcache-0.5.4-cp313-cp313-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:db3ae52ccc150dbc84704e9d642743897f3e1c54742ff34cacb661e52e3818a9"},
+    {file = "propcache-0.5.4-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:f85915e00dcb1cd9f2f890ead064ed40a27df06f0db65be427b29482ae357572"},
+    {file = "propcache-0.5.4-cp313-cp313-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:c2ba30a89035b57b73e00475de948521602f543d79ce01db10b04b36c4c76fc8"},
+    {file = "propcache-0.5.4-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:ae58f361bd5dae942717c65d3413b478c70aea9c462599e7b9adad3731db3894"},
+    {file = "propcache-0.5.4-cp313-cp313-musllinux_1_2_armv7l.whl", hash = "sha256:96f7c5c15656040ddcbc51e56dc59b58aa25999d743c126abd425b9766ab43e9"},
+    {file = "propcache-0.5.4-cp313-cp313-musllinux_1_2_ppc64le.whl", hash = "sha256:7cc528e760a8af06f2b13e9b9f362cd90c7c718ea61228a96dbd31ba16ed7f47"},
+    {file = "propcache-0.5.4-cp313-cp313-musllinux_1_2_riscv64.whl", hash = "sha256:425f8cc86ab5018b4b8d4a23bc8e74d964bd3d757c3702e301aa79be76c53f6c"},
+    {file = "propcache-0.5.4-cp313-cp313-musllinux_1_2_s390x.whl", hash = "sha256:a5793c7698a53f56f4a1889a4737c7eeb1b7ad0842fa6b1abca22913ff79c8c1"},
+    {file = "propcache-0.5.4-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:c02c0e570c5c7e077b0181a9f3cdb7d4c3617d1cda6b5c95bd5d34022923d82c"},
+    {file = "propcache-0.5.4-cp313-cp313-win32.whl", hash = "sha256:3e413d7a4a9b4866b7a761d6060d434b64d23cd35122eda3b026a0bbe8196b25"},
+    {file = "propcache-0.5.4-cp313-cp313-win_amd64.whl", hash = "sha256:0c889f6fa84957bc7e8b4eab71fd16a0455068d5045e3aa40c733071d2b2fd77"},
+    {file = "propcache-0.5.4-cp313-cp313-win_arm64.whl", hash = "sha256:69fc35c0779522da366c563e5faf203ffc1f8ff0021d5b1337fa4efa5be73177"},
+    {file = "propcache-0.5.4-cp314-cp314-macosx_10_15_universal2.whl", hash = "sha256:e6720ba44ad7e72174314d0e1fb0172494cff5c73a3a8a2159c3d2402ff15565"},
+    {file = "propcache-0.5.4-cp314-cp314-macosx_10_15_x86_64.whl", hash = "sha256:4cfe0a92ae30151869e67a4b5f5e105e4e03ad30b3f38e5211b5bf77d0881993"},
+    {file = "propcache-0.5.4-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:1d759d05634f1b038fb625a66662a8c85e5a8fec912da381b5149ddac107482b"},
+    {file = "propcache-0.5.4-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:251c63dd46a0659bb875cb254dc4c1e79ee91a847c737cd62373295afc2235dc"},
+    {file = "propcache-0.5.4-cp314-cp314-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:7a8d5ff04eb1f85698a78d20c62a14676e7b960dcafde09a388d60ad377d355d"},
+    {file = "propcache-0.5.4-cp314-cp314-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:7b9100a93b372418d8688f3f2a3e5b45c64d70ca4d6176e121aca1e3bfc1e32f"},
+    {file = "propcache-0.5.4-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:cc07876cfb079b6f6f36d21ce75784ad6c2c6b563eeac0ed26c2fa2669b85df9"},
+    {file = "propcache-0.5.4-cp314-cp314-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:0951315a6b3142ee2167404d707743f0157c110091342b1aa0accac5cf0e4acf"},
+    {file = "propcache-0.5.4-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:bee7d3aed13d56f54e681df38c3a23031bc9e3863f687d9d598825c9146acd7d"},
+    {file = "propcache-0.5.4-cp314-cp314-musllinux_1_2_armv7l.whl", hash = "sha256:4e985382be6d15da8d0c2710a6fa7b9070fc9ecdeefb7f580e88373984ec8be3"},
+    {file = "propcache-0.5.4-cp314-cp314-musllinux_1_2_ppc64le.whl", hash = "sha256:9e9ab13760aa8b6d0881ae7cb04fd891d8d490cd2554ea8e79bb278399169bcc"},
+    {file = "propcache-0.5.4-cp314-cp314-musllinux_1_2_riscv64.whl", hash = "sha256:1b2f3bec4261a94019575481c726c29850f72e27907773c75b1de421e20e9f9d"},
+    {file = "propcache-0.5.4-cp314-cp314-musllinux_1_2_s390x.whl", hash = "sha256:720cf832eb2d0b0dfee129cb3335a26f6ce3cc45ee1187e8f0731758caa16792"},
+    {file = "propcache-0.5.4-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:9fb0a5be8d9aa213150e8d8148a42aca4984b285bcad1e69587dc4298edd929b"},
+    {file = "propcache-0.5.4-cp314-cp314-win32.whl", hash = "sha256:30cc1cebaf9aef49db06357a50398323ae04d70460c0491837d026ab7d6452ea"},
+    {file = "propcache-0.5.4-cp314-cp314-win_amd64.whl", hash = "sha256:0a095db8e15a6020db149ecbed6461939fe74f6acaa3ae8b702a1fe8c38cd983"},
+    {file = "propcache-0.5.4-cp314-cp314-win_arm64.whl", hash = "sha256:45488d1a5f9ab5bd90aaa1ca20f50fe1922b8ffad71a2009d2adf41355897aac"},
+    {file = "propcache-0.5.4-cp314-cp314t-macosx_10_15_universal2.whl", hash = "sha256:53eaa697c4d0422ff4cb714d00231b43352064d97b944033b30c1d57cc506ec0"},
+    {file = "propcache-0.5.4-cp314-cp314t-macosx_10_15_x86_64.whl", hash = "sha256:886b59c4d28ca97dd23b025fdfc50a0356be934efbbbca89ad26230067f86fe5"},
+    {file = "propcache-0.5.4-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:3fa15757fea1dfcd5b7745cad9f4638929605531bd4018ab2adff7955f1a403d"},
+    {file = "propcache-0.5.4-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:6f0093ac3e9daada202c2082439d414a625c57184727a46e112a3fb2a81cb788"},
+    {file = "propcache-0.5.4-cp314-cp314t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:3cd3a7edb6b95b9b33998135ebfa18d709da82290fb8f27c858970b5a12c8b56"},
+    {file = "propcache-0.5.4-cp314-cp314t-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:c174bfd1c48a1b51a3078e95586dde718374bac79719ab3541ec9e74aec40574"},
+    {file = "propcache-0.5.4-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:a219f0ac59817a9114dd2aa57c13180f993e819ba658c7ddab4b66ed1ee0d370"},
+    {file = "propcache-0.5.4-cp314-cp314t-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:17a7400cec0256f0a71ae71f9da398f9894c956ff6668a1c9d317b3367316320"},
+    {file = "propcache-0.5.4-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:978f28401afbc76cdc3df9e1717b4229a06b626a1dcc75db4e1f2beb3884c3e9"},
+    {file = "propcache-0.5.4-cp314-cp314t-musllinux_1_2_armv7l.whl", hash = "sha256:4a1f4f5ffa55dce6307631f3cb2948e117e665966ea512e0d502b16c24f567e7"},
+    {file = "propcache-0.5.4-cp314-cp314t-musllinux_1_2_ppc64le.whl", hash = "sha256:213bb68d9ced5cf2bf717b1071bf2b09b4b04c426256f9fe6d054c60318424c4"},
+    {file = "propcache-0.5.4-cp314-cp314t-musllinux_1_2_riscv64.whl", hash = "sha256:286867fb156488c251a3721766e380ac4495e4fd6b51aaa1403d89ce7f4359d9"},
+    {file = "propcache-0.5.4-cp314-cp314t-musllinux_1_2_s390x.whl", hash = "sha256:445ee3bfb46e85838387fb3c536a73cc0b994dc192b004e40e170adc54aa2a7e"},
+    {file = "propcache-0.5.4-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:48cb48c5346a97de792254af77715aa2529c2a1ebc5f586aa0aae44a02f1fe57"},
+    {file = "propcache-0.5.4-cp314-cp314t-win32.whl", hash = "sha256:03b229037d25b801e7af53fd52b9fc49d9439b036fca1e087e02780631adfa97"},
+    {file = "propcache-0.5.4-cp314-cp314t-win_amd64.whl", hash = "sha256:8a1fc236528c457cd739c88abe823da851b7ab645d72792f88658114cc340c12"},
+    {file = "propcache-0.5.4-cp314-cp314t-win_arm64.whl", hash = "sha256:135036c5cfc93864affb0f9af9a27e5d7a71cb7bd745e7b6dbfc2d56cc30e827"},
+    {file = "propcache-0.5.4-cp315-cp315-macosx_10_15_universal2.whl", hash = "sha256:45bf2e730ab8905d0527fe05a86500f406e64305c34cc81ebe64b4617cab9760"},
+    {file = "propcache-0.5.4-cp315-cp315-macosx_10_15_x86_64.whl", hash = "sha256:31eb43ba2edc704ab2ec27815315dd8a19def0fb16215be4cfe8d32fe78ffd51"},
+    {file = "propcache-0.5.4-cp315-cp315-macosx_11_0_arm64.whl", hash = "sha256:174507f82d3594622acb1dd2dafecf2d899d6d506335494e7107767bf05f3aae"},
+    {file = "propcache-0.5.4-cp315-cp315-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:50e337653721d20ead710da33bf44487fbe8a0db8782714b60306481e9f95b51"},
+    {file = "propcache-0.5.4-cp315-cp315-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:0d21d0d2c82bbfeb1677a9711f38df968f9837576102bb4add1bd449d28d88f1"},
+    {file = "propcache-0.5.4-cp315-cp315-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:ccf4f7a79e26bb7efb06ecd50c177833b71df05cbc748701372325e6bcc17f6f"},
+    {file = "propcache-0.5.4-cp315-cp315-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:23278f808cd81d5ada7184a76606b925fb3389c60e1077b2cd7da7b1fcf0553c"},
+    {file = "propcache-0.5.4-cp315-cp315-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:e738ab81179510ce79b2eac9a6ecf47feffd9e76d1c72e403005dddb6e36c06c"},
+    {file = "propcache-0.5.4-cp315-cp315-musllinux_1_2_aarch64.whl", hash = "sha256:a419ee85e654927baabda3929c03c0cc1112bf472ff0dfd6142f4e3a81ca4162"},
+    {file = "propcache-0.5.4-cp315-cp315-musllinux_1_2_armv7l.whl", hash = "sha256:b61805357d966680acf68b3b6d49772631ed9df44ebece10ff1460e117a7da8a"},
+    {file = "propcache-0.5.4-cp315-cp315-musllinux_1_2_ppc64le.whl", hash = "sha256:58134228927cee6c047d626c08e60a81be604a20578a12ce752cc5c9a84d4826"},
+    {file = "propcache-0.5.4-cp315-cp315-musllinux_1_2_riscv64.whl", hash = "sha256:350b272b2279f4135a64fc0c304a5d08e28a137c9573442c606152446638a831"},
+    {file = "propcache-0.5.4-cp315-cp315-musllinux_1_2_s390x.whl", hash = "sha256:45bebbe252550fec975ba3b62bc6f931643cfd3b5464ef47619cf3fef154e01c"},
+    {file = "propcache-0.5.4-cp315-cp315-musllinux_1_2_x86_64.whl", hash = "sha256:ada748108a43d29b7c328ba7db3755327cd94f028bcc1a7ee3f0addcfacd9c38"},
+    {file = "propcache-0.5.4-cp315-cp315-win32.whl", hash = "sha256:ee19113bce2f3acd46432050688b70f61acd6857d75abb9ec96341b7e9ced123"},
+    {file = "propcache-0.5.4-cp315-cp315-win_amd64.whl", hash = "sha256:ceb3e879afac028f93d272c957814695dc5569e4904262dbee92f6c41bd5e4a3"},
+    {file = "propcache-0.5.4-cp315-cp315-win_arm64.whl", hash = "sha256:c83acbce9f2b5e3f5f5eda9e53d2001fed22fcdfef81274a9e02d8fd53b70a30"},
+    {file = "propcache-0.5.4-cp315-cp315t-macosx_10_15_universal2.whl", hash = "sha256:a5e8ef588c109725dc713ba69aadcac00a1ef90c2ce9c0a8c7075128f569f47f"},
+    {file = "propcache-0.5.4-cp315-cp315t-macosx_10_15_x86_64.whl", hash = "sha256:4d86476a935c88963d9b8e1a9a0d38188790e9622169bfbafa173046846709d3"},
+    {file = "propcache-0.5.4-cp315-cp315t-macosx_11_0_arm64.whl", hash = "sha256:f5470694918830da62fac9e69133b53d23b736d7070e587b27a4a2be37e08e68"},
+    {file = "propcache-0.5.4-cp315-cp315t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:10ef33a68a61ce317e095fd2e202a592ea92392b90944a78c993f0d9a73ab06c"},
+    {file = "propcache-0.5.4-cp315-cp315t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:5cacf3c9efd09df409dc33654dd077e1c245ba8fb747b0f0236ef41b7c49b589"},
+    {file = "propcache-0.5.4-cp315-cp315t-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:770e8209d018175fc0063936fa9583b6d27e88c5ad31543f3383d66080efdd62"},
+    {file = "propcache-0.5.4-cp315-cp315t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:03969626faf0783a592dfa17e28eac06018bd0b44dafae6943d53b92421a7f72"},
+    {file = "propcache-0.5.4-cp315-cp315t-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:ef3b928d9c984322b5c44e6964d8dbc653da87d2d8ee1647fa6da43072e650a9"},
+    {file = "propcache-0.5.4-cp315-cp315t-musllinux_1_2_aarch64.whl", hash = "sha256:7177c43eddf10a0893c4fec52ebb408fdcd7f7d63962caace9180d8f81b14ece"},
+    {file = "propcache-0.5.4-cp315-cp315t-musllinux_1_2_armv7l.whl", hash = "sha256:420162a77f94eb1cf5ef7893f500016dabd548e73de956785a1dd899cc73006a"},
+    {file = "propcache-0.5.4-cp315-cp315t-musllinux_1_2_ppc64le.whl", hash = "sha256:3eb2e820e8e2101407da93f17c57cbb7d225461955fc60105daaba14cd421ee2"},
+    {file = "propcache-0.5.4-cp315-cp315t-musllinux_1_2_riscv64.whl", hash = "sha256:13e52b6e0bde97dee98ab66552dbff2931649c96f1ac432eac299fe689ec373b"},
+    {file = "propcache-0.5.4-cp315-cp315t-musllinux_1_2_s390x.whl", hash = "sha256:12682126712ddc19b70ff819debbd279e58adf1f0c8f8f8138c18ade2044b284"},
+    {file = "propcache-0.5.4-cp315-cp315t-musllinux_1_2_x86_64.whl", hash = "sha256:3af0c8642b2da4815d86e631232ac8286e17644fad907c19508aa8e7cb4ba8ad"},
+    {file = "propcache-0.5.4-cp315-cp315t-win32.whl", hash = "sha256:1df8d8561b21465c5dd56110a01caf897e026d065b4b84e98a488209094272ec"},
+    {file = "propcache-0.5.4-cp315-cp315t-win_amd64.whl", hash = "sha256:02c0a34f16889cf800f10f0247a564d8ce6eeab6ffcd7c87198f769067eb8432"},
+    {file = "propcache-0.5.4-cp315-cp315t-win_arm64.whl", hash = "sha256:dc4242ca653c9b30ab51c5f8193323e7bc0928f897ee9103201e59a43abcb72e"},
+    {file = "propcache-0.5.4-py3-none-any.whl", hash = "sha256:62c60aec739ed00124573cce1178138fd690c7676352d67a37328c1cf51d7468"},
+    {file = "propcache-0.5.4.tar.gz", hash = "sha256:ff6b113f50bc066a698db5d944d2c6dc7507168dd3341e255a8892fd0715a558"},
+]
+
+[[package]]
+name = "psutil"
+version = "7.2.2"
+description = "Cross-platform lib for process and system monitoring."
+optional = false
+python-versions = ">=3.6"
+groups = ["main"]
+files = [
+    {file = "psutil-7.2.2-cp313-cp313t-macosx_10_13_x86_64.whl", hash = "sha256:2edccc433cbfa046b980b0df0171cd25bcaeb3a68fe9022db0979e7aa74a826b"},
+    {file = "psutil-7.2.2-cp313-cp313t-macosx_11_0_arm64.whl", hash = "sha256:e78c8603dcd9a04c7364f1a3e670cea95d51ee865e4efb3556a3a63adef958ea"},
+    {file = "psutil-7.2.2-cp313-cp313t-manylinux2010_x86_64.manylinux_2_12_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:1a571f2330c966c62aeda00dd24620425d4b0cc86881c89861fbc04549e5dc63"},
+    {file = "psutil-7.2.2-cp313-cp313t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:917e891983ca3c1887b4ef36447b1e0873e70c933afc831c6b6da078ba474312"},
+    {file = "psutil-7.2.2-cp313-cp313t-win_amd64.whl", hash = "sha256:ab486563df44c17f5173621c7b198955bd6b613fb87c71c161f827d3fb149a9b"},
+    {file = "psutil-7.2.2-cp313-cp313t-win_arm64.whl", hash = "sha256:ae0aefdd8796a7737eccea863f80f81e468a1e4cf14d926bd9b6f5f2d5f90ca9"},
+    {file = "psutil-7.2.2-cp314-cp314t-macosx_10_15_x86_64.whl", hash = "sha256:eed63d3b4d62449571547b60578c5b2c4bcccc5387148db46e0c2313dad0ee00"},
+    {file = "psutil-7.2.2-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:7b6d09433a10592ce39b13d7be5a54fbac1d1228ed29abc880fb23df7cb694c9"},
+    {file = "psutil-7.2.2-cp314-cp314t-manylinux2010_x86_64.manylinux_2_12_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:1fa4ecf83bcdf6e6c8f4449aff98eefb5d0604bf88cb883d7da3d8d2d909546a"},
+    {file = "psutil-7.2.2-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:e452c464a02e7dc7822a05d25db4cde564444a67e58539a00f929c51eddda0cf"},
+    {file = "psutil-7.2.2-cp314-cp314t-win_amd64.whl", hash = "sha256:c7663d4e37f13e884d13994247449e9f8f574bc4655d509c3b95e9ec9e2b9dc1"},
+    {file = "psutil-7.2.2-cp314-cp314t-win_arm64.whl", hash = "sha256:11fe5a4f613759764e79c65cf11ebdf26e33d6dd34336f8a337aa2996d71c841"},
+    {file = "psutil-7.2.2-cp36-abi3-macosx_10_9_x86_64.whl", hash = "sha256:ed0cace939114f62738d808fdcecd4c869222507e266e574799e9c0faa17d486"},
+    {file = "psutil-7.2.2-cp36-abi3-macosx_11_0_arm64.whl", hash = "sha256:1a7b04c10f32cc88ab39cbf606e117fd74721c831c98a27dc04578deb0c16979"},
+    {file = "psutil-7.2.2-cp36-abi3-manylinux2010_x86_64.manylinux_2_12_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:076a2d2f923fd4821644f5ba89f059523da90dc9014e85f8e45a5774ca5bc6f9"},
+    {file = "psutil-7.2.2-cp36-abi3-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:b0726cecd84f9474419d67252add4ac0cd9811b04d61123054b9fb6f57df6e9e"},
+    {file = "psutil-7.2.2-cp36-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:fd04ef36b4a6d599bbdb225dd1d3f51e00105f6d48a28f006da7f9822f2606d8"},
+    {file = "psutil-7.2.2-cp36-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:b58fabe35e80b264a4e3bb23e6b96f9e45a3df7fb7eed419ac0e5947c61e47cc"},
+    {file = "psutil-7.2.2-cp37-abi3-win_amd64.whl", hash = "sha256:eb7e81434c8d223ec4a219b5fc1c47d0417b12be7ea866e24fb5ad6e84b3d988"},
+    {file = "psutil-7.2.2-cp37-abi3-win_arm64.whl", hash = "sha256:8c233660f575a5a89e6d4cb65d9f938126312bca76d8fe087b947b3a1aaac9ee"},
+    {file = "psutil-7.2.2.tar.gz", hash = "sha256:0746f5f8d406af344fd547f1c8daa5f5c33dbc293bb8d6a16d80b4bb88f59372"},
+]
+
+[package.extras]
+dev = ["abi3audit", "black", "check-manifest", "colorama ; os_name == \"nt\"", "coverage", "packaging", "psleak", "pylint", "pyperf", "pypinfo", "pyreadline3 ; os_name == \"nt\"", "pytest", "pytest-cov", "pytest-instafail", "pytest-xdist", "pywin32 ; os_name == \"nt\" and implementation_name != \"pypy\"", "requests", "rstcheck", "ruff", "setuptools", "sphinx", "sphinx_rtd_theme", "toml-sort", "twine", "validate-pyproject[all]", "virtualenv", "vulture", "wheel", "wheel ; os_name == \"nt\" and implementation_name != \"pypy\"", "wmi ; os_name == \"nt\" and implementation_name != \"pypy\""]
+test = ["psleak", "pytest", "pytest-instafail", "pytest-xdist", "pywin32 ; os_name == \"nt\" and implementation_name != \"pypy\"", "setuptools", "wheel ; os_name == \"nt\" and implementation_name != \"pypy\"", "wmi ; os_name == \"nt\" and implementation_name != \"pypy\""]
+
+[[package]]
+name = "pyarrow"
+version = "25.0.1"
+description = "Python library for Apache Arrow"
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "pyarrow-25.0.1-cp310-cp310-macosx_12_0_arm64.whl", hash = "sha256:0b1edbb2f385a6a65e9711b62ba86ac54a7816a3f8d17bb3e8a5929d65fb2485"},
+    {file = "pyarrow-25.0.1-cp310-cp310-macosx_12_0_x86_64.whl", hash = "sha256:a4dd8bf99a8fac133efc0ed6a92f5fddbe2adba0d0f6dd720e39ba9855cea85c"},
+    {file = "pyarrow-25.0.1-cp310-cp310-manylinux_2_28_aarch64.whl", hash = "sha256:bddd0c4f7630c2a3ddf6347c1bdaa79d97bcf6bd445f9e60c816b7d77c85a5ae"},
+    {file = "pyarrow-25.0.1-cp310-cp310-manylinux_2_28_x86_64.whl", hash = "sha256:a4d6d5e9a3d1879a97c08ded0c797579b7965eafd0f0c26c30b45ccc06db939b"},
+    {file = "pyarrow-25.0.1-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:514ddb60285631af068875550c90eddc181db3e8e63a032b1559be189e82f056"},
+    {file = "pyarrow-25.0.1-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:cab40b1edfef0262e0e5251aa2c58d75630f24d06dd7794480243acc001a1d7d"},
+    {file = "pyarrow-25.0.1-cp310-cp310-win_amd64.whl", hash = "sha256:60e89d8f13861a1f7f8d950fa54aebb8023b30734d0ac51ffa80beabe2df4bba"},
+    {file = "pyarrow-25.0.1-cp311-cp311-macosx_12_0_arm64.whl", hash = "sha256:51093dd9e10325fbdb3c10a2ae7c4806e5c822d94e74ae4938b26524a3323fee"},
+    {file = "pyarrow-25.0.1-cp311-cp311-macosx_12_0_x86_64.whl", hash = "sha256:eb6203482ff3746a5632303a7279ae0b5a304c46985b49ed1378cb350ea6728d"},
+    {file = "pyarrow-25.0.1-cp311-cp311-manylinux_2_28_aarch64.whl", hash = "sha256:880523be3d29efcf83d3998835d206118ccf35e3871dbd2fb60408cf6b007a80"},
+    {file = "pyarrow-25.0.1-cp311-cp311-manylinux_2_28_x86_64.whl", hash = "sha256:25f8720bf6387d5dc2ebd2622112de630760419e4b66134405dd24110d15f37e"},
+    {file = "pyarrow-25.0.1-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:4facd65742a024a4a366328a1d2292062d72d6e023c1b7dda8d4c37544933a25"},
+    {file = "pyarrow-25.0.1-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:aa0559502e1cd6254d6814614085dd9c5a3dd0419362978a936a3f68a9e5c3df"},
+    {file = "pyarrow-25.0.1-cp311-cp311-win_amd64.whl", hash = "sha256:62cd0d785b8aa6675ee355f9fc02252a340f4441257c42674937826fd7594325"},
+    {file = "pyarrow-25.0.1-cp312-cp312-macosx_12_0_arm64.whl", hash = "sha256:df961f2e7ae9cf496459259d798652c70625f6c080650d6952f8c04053c58ee9"},
+    {file = "pyarrow-25.0.1-cp312-cp312-macosx_12_0_x86_64.whl", hash = "sha256:cc4aa407fde9fc660be3939e49ea31f50f3e9fec17c0ec63159f7711edd3efc9"},
+    {file = "pyarrow-25.0.1-cp312-cp312-manylinux_2_28_aarch64.whl", hash = "sha256:4340f0ba6c1d2e13f21658de1d7c662ca2545018568d0030a1e9afca159d87e3"},
+    {file = "pyarrow-25.0.1-cp312-cp312-manylinux_2_28_x86_64.whl", hash = "sha256:5389cdf79447ed1515c9e31620e6e1e2302249564d603f2ad727d4f6d313e4c3"},
+    {file = "pyarrow-25.0.1-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:d51592cb7561e87877c506113e7adbf1342ab579e6c21f0ef44b8ba41cb74c80"},
+    {file = "pyarrow-25.0.1-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:6109c94d8b9f3b17a041daca16cacb2f651ad8f1ef70a4232c2c0f37a23da2a8"},
+    {file = "pyarrow-25.0.1-cp312-cp312-win_amd64.whl", hash = "sha256:8858d7bfc22e3f51529aeaa4077225029724623e4595dc9eff8c793935c34140"},
+    {file = "pyarrow-25.0.1-cp313-cp313-macosx_12_0_arm64.whl", hash = "sha256:c7c534ec03c358a76ea3e505e74c1b6aef290af90c444dfd092dbfe23e755b85"},
+    {file = "pyarrow-25.0.1-cp313-cp313-macosx_12_0_x86_64.whl", hash = "sha256:dda9470024204d7bbf2042b47c6e8a0e47a3eeb8e34405882dfaea6577e0c153"},
+    {file = "pyarrow-25.0.1-cp313-cp313-manylinux_2_28_aarch64.whl", hash = "sha256:44a9120ce5bd81936b8ab9a88076e3fd47c2c6838e0e43630fed83626aca81d9"},
+    {file = "pyarrow-25.0.1-cp313-cp313-manylinux_2_28_x86_64.whl", hash = "sha256:0befcf816e45a1af33ac775a9970b749e4868a230c7372f0ae5e932bee27039f"},
+    {file = "pyarrow-25.0.1-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:3f89685964f46e4216103c75483aac0c0692a5f72212d7ca835adba5ede56ce3"},
+    {file = "pyarrow-25.0.1-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:6943e2fe7954d29d84de45d29d34c8dc36ce96570e67d89aa9976e650a4a9138"},
+    {file = "pyarrow-25.0.1-cp313-cp313-win_amd64.whl", hash = "sha256:31e49a7888fcdf3a835da33ae777f6bb9a866334e5a789282fc26dcf426f7f15"},
+    {file = "pyarrow-25.0.1-cp314-cp314-macosx_12_0_arm64.whl", hash = "sha256:bf0b672390cdcb640d7288f96b826d71ff4e9abb254a86c89890baf51a29cee6"},
+    {file = "pyarrow-25.0.1-cp314-cp314-macosx_12_0_x86_64.whl", hash = "sha256:38a9a4b4b9613380e200641891495a56c3d5a98a092db4a870af9975e220471d"},
+    {file = "pyarrow-25.0.1-cp314-cp314-manylinux_2_28_aarch64.whl", hash = "sha256:0b726ad7e7b669be982b0c71c07fe4b037d654354130da79a7902a669e93a66b"},
+    {file = "pyarrow-25.0.1-cp314-cp314-manylinux_2_28_x86_64.whl", hash = "sha256:9171748cdf796972d85a4b60157c279913e242992e350c90c7450182a9838b2a"},
+    {file = "pyarrow-25.0.1-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:b7a296aac7a71fa0886c08e155ddb6c636a50013f801f6178daafa0f9e726188"},
+    {file = "pyarrow-25.0.1-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:0fe7c8b6c03969b49c8c66182e4a18e3819ab92d07cfab5d8370c531b9369ef0"},
+    {file = "pyarrow-25.0.1-cp314-cp314-win_amd64.whl", hash = "sha256:f729cfdbd36fd99d543b67a914d2de044c84ebe45be8b34902b299b608c15c8f"},
+    {file = "pyarrow-25.0.1-cp314-cp314t-macosx_12_0_arm64.whl", hash = "sha256:59a2de54c0cbd954da861eee4d1d330f8e909c45b53455baef696380f2c55033"},
+    {file = "pyarrow-25.0.1-cp314-cp314t-macosx_12_0_x86_64.whl", hash = "sha256:35935cd5de130aa5cf4dea052a63e6bf2e17006c35c3a468194242b9b2bf5956"},
+    {file = "pyarrow-25.0.1-cp314-cp314t-manylinux_2_28_aarch64.whl", hash = "sha256:f3831aaa25c67a99f99dc8b05873cb9d64560390372e2aa197ce9dd4a3f06a44"},
+    {file = "pyarrow-25.0.1-cp314-cp314t-manylinux_2_28_x86_64.whl", hash = "sha256:6a1fdfc6659b6b19022f2e50627fb5cf7156a66c46bf4299379955cbe742382a"},
+    {file = "pyarrow-25.0.1-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:169d3429d5be7c752125890620f75a60776d38b0035eddae939651640822332e"},
+    {file = "pyarrow-25.0.1-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:119297a6dc197e45d9c6d4415f7814a67ffa36c180d26f68c154c58067ae782d"},
+    {file = "pyarrow-25.0.1-cp314-cp314t-win_amd64.whl", hash = "sha256:4288f27577352d608ca08553b0865e4a9b3aa14820c5d95b53337218d609835b"},
+    {file = "pyarrow-25.0.1.tar.gz", hash = "sha256:9150a83248bfed9813ea3c3af74c3856c1984d444aa28e58bf7733b9750ddf6a"},
+]
+
+[[package]]
+name = "pydantic"
+version = "2.13.5"
+description = "Data validation using Python type hints"
+optional = false
+python-versions = ">=3.9"
+groups = ["main"]
+files = [
+    {file = "pydantic-2.13.5-py3-none-any.whl", hash = "sha256:346a034f080da3755d8e9cb5e00e8b07de1d39e4f6e2c87d8ab7cafa0b269a73"},
+    {file = "pydantic-2.13.5.tar.gz", hash = "sha256:51a9c5f7b2f8e636f04c6cada605d9b6a3bf1348fdf945a3d8869b19bba0ee08"},
+]
+
+[package.dependencies]
+annotated-types = ">=0.6.0"
+pydantic-core = "2.46.5"
+typing-extensions = ">=4.14.1"
+typing-inspection = ">=0.4.2"
+
+[package.extras]
+email = ["email-validator (>=2.0.0)"]
+timezone = ["tzdata ; python_version >= \"3.9\" and platform_system == \"Windows\""]
+
+[[package]]
+name = "pydantic-core"
+version = "2.46.5"
+description = "Core functionality for Pydantic validation and serialization"
+optional = false
+python-versions = ">=3.9"
+groups = ["main"]
+files = [
+    {file = "pydantic_core-2.46.5-cp310-cp310-macosx_10_12_x86_64.whl", hash = "sha256:657b40d6240c0a7b6a64b30f22d1e3aa631c7e846c621b0c0f6d1d75e2e15ea6"},
+    {file = "pydantic_core-2.46.5-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:ecb42011e12ee19cafbc312887cbf3546959fe02fbad44f272d4be5baa997615"},
+    {file = "pydantic_core-2.46.5-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:4dedce55295becb61921e386b99d4f2706045306e7fa52249a33004c837379fb"},
+    {file = "pydantic_core-2.46.5-cp310-cp310-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:9f47b8a949e60f027f0aa0a6f6c7b7e9c55cbf4380d10b344e282fa4e7ab1e1b"},
+    {file = "pydantic_core-2.46.5-cp310-cp310-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:200aa3dc9f8d54f0754f43247c0bad0999fdcfbfd2488384dd44f37279271fe6"},
+    {file = "pydantic_core-2.46.5-cp310-cp310-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:6d30e1a4f138b8951063e9a394752a9179b51da288ffa507b1e659222f4c1793"},
+    {file = "pydantic_core-2.46.5-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:850a08d167dde16db8702c274f320c7be9d7da6f6dff2b58b18f9e815bd94f5b"},
+    {file = "pydantic_core-2.46.5-cp310-cp310-manylinux_2_31_riscv64.whl", hash = "sha256:c3471e5c4a949c26ec00a77f01df59096aa9495877de76fd60a980f8ee6be461"},
+    {file = "pydantic_core-2.46.5-cp310-cp310-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:3a3e26b6a8274211bddee2d0e4d0d42778f17a34510f49d2ec44b58abfc41736"},
+    {file = "pydantic_core-2.46.5-cp310-cp310-musllinux_1_1_aarch64.whl", hash = "sha256:fc5d783bd4a2387e97b8a2d5ec781cfb92b3d893bf82370548e99db5915935d3"},
+    {file = "pydantic_core-2.46.5-cp310-cp310-musllinux_1_1_armv7l.whl", hash = "sha256:356c8368cbc321050b169595683a2e1d63413b1e0e2868b330af9fc14c616d3f"},
+    {file = "pydantic_core-2.46.5-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:eb7d8d0e5886a89a55d2eef490e272fa965a9d57c6b29a5b5088a7997ec2cad1"},
+    {file = "pydantic_core-2.46.5-cp310-cp310-win32.whl", hash = "sha256:4d44cf99ddebf875f9b68cc267aa684c99b7b44fe63ee1cac4ec163807290069"},
+    {file = "pydantic_core-2.46.5-cp310-cp310-win_amd64.whl", hash = "sha256:1e5aad1220a1192c42341c8fd4a8686657e73ab2a920c970bdc4de334fe3193d"},
+    {file = "pydantic_core-2.46.5-cp311-cp311-macosx_10_12_x86_64.whl", hash = "sha256:a1dee1b804ff4d11c663636cf15d2ea47e9f79cd56c033fb1cbf08924842a48f"},
+    {file = "pydantic_core-2.46.5-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:d625a186a65201c23a9e3b8ed9c47e90a026e03256608cc91851c6709096844f"},
+    {file = "pydantic_core-2.46.5-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:4f8507560a9284e1370bb048ed4282012fbef4e8d109875b95e884d228552061"},
+    {file = "pydantic_core-2.46.5-cp311-cp311-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:5f93c5fe914d75fbec9a49209b00da5f08e9e467d69da2b1510c81940cfd10be"},
+    {file = "pydantic_core-2.46.5-cp311-cp311-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:aca6c767f552b21b10f774aeac128e828eafb796adfa1b666a18bf6321453c3a"},
+    {file = "pydantic_core-2.46.5-cp311-cp311-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:701b2e04b560eeb4bddf7a25ab8ca476176e34fdbd9a0e18196f0d12d4685f0b"},
+    {file = "pydantic_core-2.46.5-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:49776eab08766a08dfff7012f8b422dcd7e25e43b316eedf0477c24fcfa84b7c"},
+    {file = "pydantic_core-2.46.5-cp311-cp311-manylinux_2_31_riscv64.whl", hash = "sha256:a2468d93d181667a7abd66e1b64bb9f76f361b0fef8faddf687456453576f5ee"},
+    {file = "pydantic_core-2.46.5-cp311-cp311-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:53feb344243bb9510a9dec7bf3cf1b64d88a98af5dc7872a5160465f8b198c8e"},
+    {file = "pydantic_core-2.46.5-cp311-cp311-musllinux_1_1_aarch64.whl", hash = "sha256:cd5214352ae68f3b5e9af7768bdc5253695ee069675db3480518420b3be881f2"},
+    {file = "pydantic_core-2.46.5-cp311-cp311-musllinux_1_1_armv7l.whl", hash = "sha256:9432f3598db432cb51c5b37fdbf29a60fcccc79e30d37a05022776a6bc4ab689"},
+    {file = "pydantic_core-2.46.5-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:8feeac04b5794e513e710af2f9c87d49f31a6dc47967bb264a1fed61a8989bec"},
+    {file = "pydantic_core-2.46.5-cp311-cp311-win32.whl", hash = "sha256:892a881d5f68c2b9ea304b7a6c2c60d9343df578a311b0f86b94bc8f1ffe8129"},
+    {file = "pydantic_core-2.46.5-cp311-cp311-win_amd64.whl", hash = "sha256:40375c2d05acec10323e45dfe2077ac44bc74659008614af5069034e2cfc781c"},
+    {file = "pydantic_core-2.46.5-cp311-cp311-win_arm64.whl", hash = "sha256:28a6a556cd3b6066bea827857f9d9cce027c96f776e512f544a581f9e42161f8"},
+    {file = "pydantic_core-2.46.5-cp312-cp312-macosx_10_12_x86_64.whl", hash = "sha256:b9fe6fb92520e3fd61f2e49000b6911b188824f089b75973ea06d6267f0b476d"},
+    {file = "pydantic_core-2.46.5-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:a39ac25a9a2fa4072efdb429833c4a4c8009a51ff9eea3eeae131713cd27991e"},
+    {file = "pydantic_core-2.46.5-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:4fdc8b93a41521988916eeaa271173fcca7fa0803d62f87675aac8dcec1c8e29"},
+    {file = "pydantic_core-2.46.5-cp312-cp312-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:b98134087d9de723658d17a42c7d0da8d6e2ef08015dee7dc93889047315f5e4"},
+    {file = "pydantic_core-2.46.5-cp312-cp312-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:e652ab17569c94bff5475520f907b7148b8c24036a8ebbe5cf7cf7493d28579a"},
+    {file = "pydantic_core-2.46.5-cp312-cp312-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:d925f3d9afd05a8c0fb3a1031463a8d59ebe5e2afad297e29c78be19e13b4e62"},
+    {file = "pydantic_core-2.46.5-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:0fc5be0abd4a407e200d844b404e33639a554e7bd0d448e7b9ae181be4789ac2"},
+    {file = "pydantic_core-2.46.5-cp312-cp312-manylinux_2_31_riscv64.whl", hash = "sha256:816ff0a6550ffc06c098ccd2e0698600f9aa7da192a79eaa6f9af504a35db869"},
+    {file = "pydantic_core-2.46.5-cp312-cp312-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:c7ea57fc63aa7da93a1bd2d644e6577befae10c52c4e36377635eea1056a74f5"},
+    {file = "pydantic_core-2.46.5-cp312-cp312-musllinux_1_1_aarch64.whl", hash = "sha256:efd62a42486f1bda5d24cb4f63d15a3c7768375fe83d36f9417b4ad7a2fb20b3"},
+    {file = "pydantic_core-2.46.5-cp312-cp312-musllinux_1_1_armv7l.whl", hash = "sha256:2bc9419666990c06d7397831f2126a1ecc3594aaa3ff7de5bf2d066802f4e07b"},
+    {file = "pydantic_core-2.46.5-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:18a09e1e1011b462f2e32774f25859ef1223d5c2b0546a633cf56654710721e0"},
+    {file = "pydantic_core-2.46.5-cp312-cp312-win32.whl", hash = "sha256:5cb482e9e84c851f4e623fe4acc1ced89168cf1fe18f7089db4548c8f5bbb65b"},
+    {file = "pydantic_core-2.46.5-cp312-cp312-win_amd64.whl", hash = "sha256:5e81740c09e310f5aa5cbd3e434a01c154d4bef93241c7877b39f211d2b78ba8"},
+    {file = "pydantic_core-2.46.5-cp312-cp312-win_arm64.whl", hash = "sha256:f7b0ec93a2893de856652154d73b7ba622f26fa97726487dcac373de5f4c6084"},
+    {file = "pydantic_core-2.46.5-cp313-cp313-macosx_10_12_x86_64.whl", hash = "sha256:b7ca9034437b6022f941f4857459562ee00a560b97e7cce8a0ec5a74fc6766e0"},
+    {file = "pydantic_core-2.46.5-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:f332f0e72a5a0400141f830744e141bf9f97917878dbe968669e8a7fefea78ff"},
+    {file = "pydantic_core-2.46.5-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:193375f3548919d3f0b60936ca113ada3e38f264f91b9b8e0508efaad57be931"},
+    {file = "pydantic_core-2.46.5-cp313-cp313-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:79bdfa52f843137045b2d081cc05c120ba6665d29b7559c2c47690906f39279f"},
+    {file = "pydantic_core-2.46.5-cp313-cp313-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:24922243639cbdac66c75fcb6fd6495a9cb52b213d62f9a0d16f0310b1ff8038"},
+    {file = "pydantic_core-2.46.5-cp313-cp313-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:c76fe65e607be28c7fd4d56fc3c42b1583aa058ce3408b7ad0fd540171d31f9f"},
+    {file = "pydantic_core-2.46.5-cp313-cp313-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:6f7b393a8b3da82f5c1fc0751e6d01ac6c55b93c18226a60bdfba4a724efafd1"},
+    {file = "pydantic_core-2.46.5-cp313-cp313-manylinux_2_31_riscv64.whl", hash = "sha256:7ac031912d54f3d83ef3b3eb98dfabc1608802e2202263d25957eeed40b94761"},
+    {file = "pydantic_core-2.46.5-cp313-cp313-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:837b396ca3d7b74091ca623f6cbd8351bd42d670a79c2683e79fb089f06a2de5"},
+    {file = "pydantic_core-2.46.5-cp313-cp313-musllinux_1_1_aarch64.whl", hash = "sha256:5ee239d575f80b08eca11f6e20f90c4c695de7825c67eefe6091fbf20dda648e"},
+    {file = "pydantic_core-2.46.5-cp313-cp313-musllinux_1_1_armv7l.whl", hash = "sha256:e80675d75ae2cd14372cb65cad5400d9347a3d3f6c13000183f22dfd027283ed"},
+    {file = "pydantic_core-2.46.5-cp313-cp313-musllinux_1_1_x86_64.whl", hash = "sha256:9c4b71f10dd532fb7a5cbc8f58707779e64f03a258c2bf8bfbaecfcd9970b519"},
+    {file = "pydantic_core-2.46.5-cp313-cp313-win32.whl", hash = "sha256:97bf8de4d541598c94a59344eeb988a94c08ff76b5723c41f6567ec18c7892ea"},
+    {file = "pydantic_core-2.46.5-cp313-cp313-win_amd64.whl", hash = "sha256:15f4a94963c95accac15b7b657bb177d3ad82bb90b0d0526d9a9b85079925db5"},
+    {file = "pydantic_core-2.46.5-cp313-cp313-win_arm64.whl", hash = "sha256:d22a945598fb91236b4dd793a6e42e4f3dd7740bb5aace5ebd7d4c08d13bb575"},
+    {file = "pydantic_core-2.46.5-cp314-cp314-macosx_10_12_x86_64.whl", hash = "sha256:c1c43ad4339643d70ebb8124e1305a7dab423001eff58bb41a0f731adbc98355"},
+    {file = "pydantic_core-2.46.5-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:1a353f84de772f423b5ffb11d7ae352fbbef0f446f3c0b0af0f8236d7233606e"},
+    {file = "pydantic_core-2.46.5-cp314-cp314-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:5086029a57366b8cf81b130a43908738095c270c21a8d7f0e8bdfdb89718e2f3"},
+    {file = "pydantic_core-2.46.5-cp314-cp314-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:46c25dda9d092a06c08db76ffe0a197107904d0dfac653f7d5306bbcd6d6119c"},
+    {file = "pydantic_core-2.46.5-cp314-cp314-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:37ea7b83c935e5b0d68c9449b82651accf78a10828b2c02b2f2d9e9496446c21"},
+    {file = "pydantic_core-2.46.5-cp314-cp314-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:e64e88d5585bea9ce95861079de72006c7fa6d3df4e3a3b65ba31eb979c15c9f"},
+    {file = "pydantic_core-2.46.5-cp314-cp314-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:54d510bac3ee52247af28ed4bb18a1e799f040ac60fd2bf5ccd4c92f1fbe786f"},
+    {file = "pydantic_core-2.46.5-cp314-cp314-manylinux_2_31_riscv64.whl", hash = "sha256:a2a5e1d0ff29adddc9f6d6821a66302e4493f8ca898b715b6b1182c2c201ea0a"},
+    {file = "pydantic_core-2.46.5-cp314-cp314-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:03b9666e41e35d8909852ba191a0607520f81b74eaf12ccf8737005dbb313821"},
+    {file = "pydantic_core-2.46.5-cp314-cp314-musllinux_1_1_aarch64.whl", hash = "sha256:a91c17edf6eea2402cb5457b4c89e99bc5ed1004aa34c4adf1d4258c1a5c22c2"},
+    {file = "pydantic_core-2.46.5-cp314-cp314-musllinux_1_1_armv7l.whl", hash = "sha256:b49924c73a235e969511bf2aabdff3beebf9820931f646c80274d5d780010c47"},
+    {file = "pydantic_core-2.46.5-cp314-cp314-musllinux_1_1_x86_64.whl", hash = "sha256:2cbd9a5eff05e51c447c34dfa4632145b26b09120cf04bd0c871e44c1a5e1c9a"},
+    {file = "pydantic_core-2.46.5-cp314-cp314-win32.whl", hash = "sha256:2d5d76654becf5efd62c9e51c3756c67b49498b0c9a40884934c40807adbd074"},
+    {file = "pydantic_core-2.46.5-cp314-cp314-win_amd64.whl", hash = "sha256:fa10ef4112775900e7a0661068635eb67b2ab824fbde764de6e0e21982a93db0"},
+    {file = "pydantic_core-2.46.5-cp314-cp314-win_arm64.whl", hash = "sha256:045ab3b6d308439e32b81cc173bba5b9018bc6ed896afd0c65b3b009b1699af5"},
+    {file = "pydantic_core-2.46.5-cp314-cp314t-macosx_10_12_x86_64.whl", hash = "sha256:8816f3d218beb4b787de5c9759c259b8fa61f9dec42dc7811f320a33771778b7"},
+    {file = "pydantic_core-2.46.5-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:bce57638e08ac148e5778cce7feb968307a727d66f8e2274a543d0cf0c9ad6a3"},
+    {file = "pydantic_core-2.46.5-cp314-cp314t-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:976e1128455aa595ea04c79ccfedff1aaeab96ee013fcc916bed120c4f0ad94f"},
+    {file = "pydantic_core-2.46.5-cp314-cp314t-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:e7b891faeedeafba41b2983e5001a81b6a915b69544c7e7570d1989ce1c36ac7"},
+    {file = "pydantic_core-2.46.5-cp314-cp314t-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:5f194189415698233dd1114a093a9b56e61e2c57e11b469be3b0506f46f0771c"},
+    {file = "pydantic_core-2.46.5-cp314-cp314t-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:82a36973cf8a2ef5406f4fe2edbf8ed0c99629535d959e0b100c76a32535a111"},
+    {file = "pydantic_core-2.46.5-cp314-cp314t-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:cdbb78909f52b981d3b2d56b97328d71eb0b974c36bd77c920123a7ebb192829"},
+    {file = "pydantic_core-2.46.5-cp314-cp314t-manylinux_2_31_riscv64.whl", hash = "sha256:52e24eacdb536cade636aa90fb851835222becff8484b7001fdc78cb0290f2aa"},
+    {file = "pydantic_core-2.46.5-cp314-cp314t-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:37ae34309d7bd8c0d61ab839668058f2a7962ea1fc51d105d2db228fe0618034"},
+    {file = "pydantic_core-2.46.5-cp314-cp314t-musllinux_1_1_aarch64.whl", hash = "sha256:0cdbada856a1c69a7624a64d3d9aefe79300bd6ef827b43a4f265010b9b55184"},
+    {file = "pydantic_core-2.46.5-cp314-cp314t-musllinux_1_1_armv7l.whl", hash = "sha256:545f26c504b27c3758439a5e6d9349931f0a04f855668d5fe323c89e82300a38"},
+    {file = "pydantic_core-2.46.5-cp314-cp314t-musllinux_1_1_x86_64.whl", hash = "sha256:ff218293c9c806138dca139765e3b067621be52bcd93cdc14c7711be7ddc90a9"},
+    {file = "pydantic_core-2.46.5-cp314-cp314t-win32.whl", hash = "sha256:97cf3eb53a8cccacf9d46686a0926186c9bfb5574f2ed66d3639d5fe117cd3a9"},
+    {file = "pydantic_core-2.46.5-cp314-cp314t-win_amd64.whl", hash = "sha256:d2f9fc07a8042a8f95925b35c4f04f469707c981fc33245b6ca187cf5d2dd290"},
+    {file = "pydantic_core-2.46.5-cp314-cp314t-win_arm64.whl", hash = "sha256:acf8a67ba51f4ca9ddbd0e6b3000a65ac51ab734661778b3e7ba64d99a710f2f"},
+    {file = "pydantic_core-2.46.5-cp39-cp39-macosx_10_12_x86_64.whl", hash = "sha256:c583b927a8838dab890706a6fa7573fbb8b70e24000ef9f7238e2d6f6435a5ed"},
+    {file = "pydantic_core-2.46.5-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:cdc8b74ecc48c0cb1e9607a05ec4e9e88db60a19ffcc9a1d5f9088ede40c8dc0"},
+    {file = "pydantic_core-2.46.5-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:8b10e3e8fd7ddc2bd915848a2768e44c15b22936f1cc54c462ad1164deb02655"},
+    {file = "pydantic_core-2.46.5-cp39-cp39-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:f077d0b97ab11fa7dcc633fca53515f290bca8a8a633e966d5b6d1879d9ed01a"},
+    {file = "pydantic_core-2.46.5-cp39-cp39-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:7b0fc826b16c55e561e5d2a0c5c77b051ba1d92808118c4e4b5390f5e0cf191d"},
+    {file = "pydantic_core-2.46.5-cp39-cp39-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:ef3fbbf161dc9351a2fe0422e51b129f9e97e42385bd0320b309c15f7d287dd8"},
+    {file = "pydantic_core-2.46.5-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:978e7b97d4824b5be09c69fb70507cbde3b0323fc147332ca40a94d9a6a0ebbf"},
+    {file = "pydantic_core-2.46.5-cp39-cp39-manylinux_2_31_riscv64.whl", hash = "sha256:9b68938dd5b0c783d88ff8e2dcc69451b5eb936fe212d516b21b9d5567f6d464"},
+    {file = "pydantic_core-2.46.5-cp39-cp39-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:771cf63ae0b1b50dd22e5f3e3549fab5f3f4ff1635d352a9e1a97fe01c7b2e64"},
+    {file = "pydantic_core-2.46.5-cp39-cp39-musllinux_1_1_aarch64.whl", hash = "sha256:7c6be839a5a8312626b32029a415644a0846b420bc8b52b95b28cd92da162168"},
+    {file = "pydantic_core-2.46.5-cp39-cp39-musllinux_1_1_armv7l.whl", hash = "sha256:895395f8918627b04efb1ad2a4cf605387143300ba03304cd1dfa6d03f5e095e"},
+    {file = "pydantic_core-2.46.5-cp39-cp39-musllinux_1_1_x86_64.whl", hash = "sha256:fc8515076c11f3cfdf4fb142dcca0fe384b1230a3b5415458ac84f3e0903ec13"},
+    {file = "pydantic_core-2.46.5-cp39-cp39-win32.whl", hash = "sha256:3d2652072b2d774947ba5cf78a9e59644ac62ee572daf6dd2e1dfe905e15b2b7"},
+    {file = "pydantic_core-2.46.5-cp39-cp39-win_amd64.whl", hash = "sha256:3aa166e99c4f2985407fb8714aebede877ecb5455cf321b606adca926d30d5a0"},
+    {file = "pydantic_core-2.46.5-graalpy311-graalpy242_311_native-macosx_10_12_x86_64.whl", hash = "sha256:c14ad3bdc85ee7f318742c457ca3968a92126d144b15721c759033bfb06296c2"},
+    {file = "pydantic_core-2.46.5-graalpy311-graalpy242_311_native-macosx_11_0_arm64.whl", hash = "sha256:0bddb4020d8f04175865ccd17eff3040874fc11fb593f424edb452653b4b947c"},
+    {file = "pydantic_core-2.46.5-graalpy311-graalpy242_311_native-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:2471fd51c61c610e1dcf7de44d7299283661654d11264ab4802b303368d69c47"},
+    {file = "pydantic_core-2.46.5-graalpy311-graalpy242_311_native-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:b10ec717381bdbfafef34607824db4c91de69ff085e4fca3b2af91b4fa17e68a"},
+    {file = "pydantic_core-2.46.5-graalpy312-graalpy250_312_native-macosx_10_12_x86_64.whl", hash = "sha256:013d6f3483d81e02e7c328831808f336c8596ee33b4bd4026b9ffb1e960b8942"},
+    {file = "pydantic_core-2.46.5-graalpy312-graalpy250_312_native-macosx_11_0_arm64.whl", hash = "sha256:e9c134bb666dd54b778b9fc0d2b50cbb7f979b9e3716f26a88c9ab3b6fc1dd0f"},
+    {file = "pydantic_core-2.46.5-graalpy312-graalpy250_312_native-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:347ec774390c87326a2e4929d58d3f7e8763a104d5d35f4cd595a4c952366433"},
+    {file = "pydantic_core-2.46.5-graalpy312-graalpy250_312_native-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:8e24d8f05fa2d28513d94e877e9c75ad66175376209b3977f916e240e623193c"},
+    {file = "pydantic_core-2.46.5-pp311-pypy311_pp73-macosx_10_12_x86_64.whl", hash = "sha256:ab4b66edffb32d9e951efb3814bd104b8367a7501b81b955cacb5726d897389f"},
+    {file = "pydantic_core-2.46.5-pp311-pypy311_pp73-macosx_11_0_arm64.whl", hash = "sha256:337639ba62a11acde6ef3aeb08c8ea755f8ef1fe5e513356c0f36a2b0d7568b0"},
+    {file = "pydantic_core-2.46.5-pp311-pypy311_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:413a717a410d0c817ef5b786a059415550b3794e1d0c2abffd9efb93a3d9f7b4"},
+    {file = "pydantic_core-2.46.5-pp311-pypy311_pp73-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:1e449def1945a462c464331254e5a44fca7c3b4f9aedf59ec2f50f8066dd8e25"},
+    {file = "pydantic_core-2.46.5-pp311-pypy311_pp73-musllinux_1_1_aarch64.whl", hash = "sha256:a445486499897b88a7d6c310c88ed64dd37b1b59bfd7ae9107490bbb362f47d6"},
+    {file = "pydantic_core-2.46.5-pp311-pypy311_pp73-musllinux_1_1_armv7l.whl", hash = "sha256:2d330aaba8621b1edcec8ae2c4050f63b84ccf6d98723a8f212e9684713abf0e"},
+    {file = "pydantic_core-2.46.5-pp311-pypy311_pp73-musllinux_1_1_x86_64.whl", hash = "sha256:b6acfb46a814762367fb7ba0828b0a17d441b92ce249a0e007474c9072662dda"},
+    {file = "pydantic_core-2.46.5-pp311-pypy311_pp73-win_amd64.whl", hash = "sha256:d0a24b40877af2de4950252be9d21eaf7fb07660f3c2cae1f56c6b599ada5266"},
+    {file = "pydantic_core-2.46.5.tar.gz", hash = "sha256:10416c15b8839ecc4ef4d0885da76da6fd0f67333a0eb8aff6d93c4b8f2910fc"},
+]
+
+[package.dependencies]
+typing-extensions = ">=4.14.1"
+
+[[package]]
+name = "pygments"
+version = "2.21.0"
+description = "Pygments is a syntax highlighting package written in Python."
+optional = false
+python-versions = ">=3.9"
+groups = ["main", "dev"]
+files = [
+    {file = "pygments-2.21.0-py3-none-any.whl", hash = "sha256:2363c69b61c4a97c838da3b130dcd6468f4848992b21a82f2a63ec34377137d9"},
+    {file = "pygments-2.21.0.tar.gz", hash = "sha256:610ca751c9bc2492b38eb9a38a7fbc93edbbb2d7182edaf34e66ae493dee5c8c"},
+]
+
+[package.extras]
+windows-terminal = ["colorama (>=0.4.6)"]
+
+[[package]]
+name = "pytest"
+version = "9.1.1"
+description = "pytest: simple powerful testing with Python"
+optional = false
+python-versions = ">=3.10"
+groups = ["dev"]
+files = [
+    {file = "pytest-9.1.1-py3-none-any.whl", hash = "sha256:37a86b45efb9a47a61a36449063e8e18d0cab3161329fc099eb21783169c4f0c"},
+    {file = "pytest-9.1.1.tar.gz", hash = "sha256:1088fbde8f2b49d95a549a195707afa7a76a3ce9bcadc26b6d71f0ffda5fe313"},
+]
+
+[package.dependencies]
+colorama = {version = ">=0.4", markers = "sys_platform == \"win32\""}
+iniconfig = ">=1.0.1"
+packaging = ">=22"
+pluggy = ">=1.5,<2"
+pygments = ">=2.7.2"
+
+[package.extras]
+dev = ["argcomplete", "attrs (>=19.2)", "hypothesis (>=3.56)", "mock", "requests", "setuptools", "xmlschema"]
+
+[[package]]
+name = "pytest-cov"
+version = "7.1.0"
+description = "Pytest plugin for measuring coverage."
+optional = false
+python-versions = ">=3.9"
+groups = ["dev"]
+files = [
+    {file = "pytest_cov-7.1.0-py3-none-any.whl", hash = "sha256:a0461110b7865f9a271aa1b51e516c9a95de9d696734a2f71e3e78f46e1d4678"},
+    {file = "pytest_cov-7.1.0.tar.gz", hash = "sha256:30674f2b5f6351aa09702a9c8c364f6a01c27aae0c1366ae8016160d1efc56b2"},
+]
+
+[package.dependencies]
+coverage = {version = ">=7.10.6", extras = ["toml"]}
+pluggy = ">=1.2"
+pytest = ">=7"
+
+[package.extras]
+testing = ["process-tests", "pytest-xdist", "virtualenv"]
+
+[[package]]
+name = "python-dateutil"
+version = "2.9.0.post0"
+description = "Extensions to the standard Python datetime module"
+optional = false
+python-versions = "!=3.0.*,!=3.1.*,!=3.2.*,>=2.7"
+groups = ["main"]
+files = [
+    {file = "python-dateutil-2.9.0.post0.tar.gz", hash = "sha256:37dd54208da7e1cd875388217d5e00ebd4179249f90fb72437e91a35459a0ad3"},
+    {file = "python_dateutil-2.9.0.post0-py2.py3-none-any.whl", hash = "sha256:a8b2bc7bffae282281c8140a97d3aa9c14da0b136dfe83f850eea9a5f7470427"},
+]
+
+[package.dependencies]
+six = ">=1.5"
+
+[[package]]
+name = "pyyaml"
+version = "6.0.3"
+description = "YAML parser and emitter for Python"
+optional = false
+python-versions = ">=3.8"
+groups = ["main"]
+files = [
+    {file = "PyYAML-6.0.3-cp38-cp38-macosx_10_13_x86_64.whl", hash = "sha256:c2514fceb77bc5e7a2f7adfaa1feb2fb311607c9cb518dbc378688ec73d8292f"},
+    {file = "PyYAML-6.0.3-cp38-cp38-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:9c57bb8c96f6d1808c030b1687b9b5fb476abaa47f0db9c0101f5e9f394e97f4"},
+    {file = "PyYAML-6.0.3-cp38-cp38-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:efd7b85f94a6f21e4932043973a7ba2613b059c4a000551892ac9f1d11f5baf3"},
+    {file = "PyYAML-6.0.3-cp38-cp38-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:22ba7cfcad58ef3ecddc7ed1db3409af68d023b7f940da23c6c2a1890976eda6"},
+    {file = "PyYAML-6.0.3-cp38-cp38-musllinux_1_2_x86_64.whl", hash = "sha256:6344df0d5755a2c9a276d4473ae6b90647e216ab4757f8426893b5dd2ac3f369"},
+    {file = "PyYAML-6.0.3-cp38-cp38-win32.whl", hash = "sha256:3ff07ec89bae51176c0549bc4c63aa6202991da2d9a6129d7aef7f1407d3f295"},
+    {file = "PyYAML-6.0.3-cp38-cp38-win_amd64.whl", hash = "sha256:5cf4e27da7e3fbed4d6c3d8e797387aaad68102272f8f9752883bc32d61cb87b"},
+    {file = "pyyaml-6.0.3-cp310-cp310-macosx_10_13_x86_64.whl", hash = "sha256:214ed4befebe12df36bcc8bc2b64b396ca31be9304b8f59e25c11cf94a4c033b"},
+    {file = "pyyaml-6.0.3-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:02ea2dfa234451bbb8772601d7b8e426c2bfa197136796224e50e35a78777956"},
+    {file = "pyyaml-6.0.3-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:b30236e45cf30d2b8e7b3e85881719e98507abed1011bf463a8fa23e9c3e98a8"},
+    {file = "pyyaml-6.0.3-cp310-cp310-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:66291b10affd76d76f54fad28e22e51719ef9ba22b29e1d7d03d6777a9174198"},
+    {file = "pyyaml-6.0.3-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:9c7708761fccb9397fe64bbc0395abcae8c4bf7b0eac081e12b809bf47700d0b"},
+    {file = "pyyaml-6.0.3-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:418cf3f2111bc80e0933b2cd8cd04f286338bb88bdc7bc8e6dd775ebde60b5e0"},
+    {file = "pyyaml-6.0.3-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:5e0b74767e5f8c593e8c9b5912019159ed0533c70051e9cce3e8b6aa699fcd69"},
+    {file = "pyyaml-6.0.3-cp310-cp310-win32.whl", hash = "sha256:28c8d926f98f432f88adc23edf2e6d4921ac26fb084b028c733d01868d19007e"},
+    {file = "pyyaml-6.0.3-cp310-cp310-win_amd64.whl", hash = "sha256:bdb2c67c6c1390b63c6ff89f210c8fd09d9a1217a465701eac7316313c915e4c"},
+    {file = "pyyaml-6.0.3-cp311-cp311-macosx_10_13_x86_64.whl", hash = "sha256:44edc647873928551a01e7a563d7452ccdebee747728c1080d881d68af7b997e"},
+    {file = "pyyaml-6.0.3-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:652cb6edd41e718550aad172851962662ff2681490a8a711af6a4d288dd96824"},
+    {file = "pyyaml-6.0.3-cp311-cp311-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:10892704fc220243f5305762e276552a0395f7beb4dbf9b14ec8fd43b57f126c"},
+    {file = "pyyaml-6.0.3-cp311-cp311-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:850774a7879607d3a6f50d36d04f00ee69e7fc816450e5f7e58d7f17f1ae5c00"},
+    {file = "pyyaml-6.0.3-cp311-cp311-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:b8bb0864c5a28024fac8a632c443c87c5aa6f215c0b126c449ae1a150412f31d"},
+    {file = "pyyaml-6.0.3-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:1d37d57ad971609cf3c53ba6a7e365e40660e3be0e5175fa9f2365a379d6095a"},
+    {file = "pyyaml-6.0.3-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:37503bfbfc9d2c40b344d06b2199cf0e96e97957ab1c1b546fd4f87e53e5d3e4"},
+    {file = "pyyaml-6.0.3-cp311-cp311-win32.whl", hash = "sha256:8098f252adfa6c80ab48096053f512f2321f0b998f98150cea9bd23d83e1467b"},
+    {file = "pyyaml-6.0.3-cp311-cp311-win_amd64.whl", hash = "sha256:9f3bfb4965eb874431221a3ff3fdcddc7e74e3b07799e0e84ca4a0f867d449bf"},
+    {file = "pyyaml-6.0.3-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:7f047e29dcae44602496db43be01ad42fc6f1cc0d8cd6c83d342306c32270196"},
+    {file = "pyyaml-6.0.3-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:fc09d0aa354569bc501d4e787133afc08552722d3ab34836a80547331bb5d4a0"},
+    {file = "pyyaml-6.0.3-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:9149cad251584d5fb4981be1ecde53a1ca46c891a79788c0df828d2f166bda28"},
+    {file = "pyyaml-6.0.3-cp312-cp312-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:5fdec68f91a0c6739b380c83b951e2c72ac0197ace422360e6d5a959d8d97b2c"},
+    {file = "pyyaml-6.0.3-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:ba1cc08a7ccde2d2ec775841541641e4548226580ab850948cbfda66a1befcdc"},
+    {file = "pyyaml-6.0.3-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:8dc52c23056b9ddd46818a57b78404882310fb473d63f17b07d5c40421e47f8e"},
+    {file = "pyyaml-6.0.3-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:41715c910c881bc081f1e8872880d3c650acf13dfa8214bad49ed4cede7c34ea"},
+    {file = "pyyaml-6.0.3-cp312-cp312-win32.whl", hash = "sha256:96b533f0e99f6579b3d4d4995707cf36df9100d67e0c8303a0c55b27b5f99bc5"},
+    {file = "pyyaml-6.0.3-cp312-cp312-win_amd64.whl", hash = "sha256:5fcd34e47f6e0b794d17de1b4ff496c00986e1c83f7ab2fb8fcfe9616ff7477b"},
+    {file = "pyyaml-6.0.3-cp312-cp312-win_arm64.whl", hash = "sha256:64386e5e707d03a7e172c0701abfb7e10f0fb753ee1d773128192742712a98fd"},
+    {file = "pyyaml-6.0.3-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:8da9669d359f02c0b91ccc01cac4a67f16afec0dac22c2ad09f46bee0697eba8"},
+    {file = "pyyaml-6.0.3-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:2283a07e2c21a2aa78d9c4442724ec1eb15f5e42a723b99cb3d822d48f5f7ad1"},
+    {file = "pyyaml-6.0.3-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:ee2922902c45ae8ccada2c5b501ab86c36525b883eff4255313a253a3160861c"},
+    {file = "pyyaml-6.0.3-cp313-cp313-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:a33284e20b78bd4a18c8c2282d549d10bc8408a2a7ff57653c0cf0b9be0afce5"},
+    {file = "pyyaml-6.0.3-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:0f29edc409a6392443abf94b9cf89ce99889a1dd5376d94316ae5145dfedd5d6"},
+    {file = "pyyaml-6.0.3-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:f7057c9a337546edc7973c0d3ba84ddcdf0daa14533c2065749c9075001090e6"},
+    {file = "pyyaml-6.0.3-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:eda16858a3cab07b80edaf74336ece1f986ba330fdb8ee0d6c0d68fe82bc96be"},
+    {file = "pyyaml-6.0.3-cp313-cp313-win32.whl", hash = "sha256:d0eae10f8159e8fdad514efdc92d74fd8d682c933a6dd088030f3834bc8e6b26"},
+    {file = "pyyaml-6.0.3-cp313-cp313-win_amd64.whl", hash = "sha256:79005a0d97d5ddabfeeea4cf676af11e647e41d81c9a7722a193022accdb6b7c"},
+    {file = "pyyaml-6.0.3-cp313-cp313-win_arm64.whl", hash = "sha256:5498cd1645aa724a7c71c8f378eb29ebe23da2fc0d7a08071d89469bf1d2defb"},
+    {file = "pyyaml-6.0.3-cp314-cp314-macosx_10_13_x86_64.whl", hash = "sha256:8d1fab6bb153a416f9aeb4b8763bc0f22a5586065f86f7664fc23339fc1c1fac"},
+    {file = "pyyaml-6.0.3-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:34d5fcd24b8445fadc33f9cf348c1047101756fd760b4dacb5c3e99755703310"},
+    {file = "pyyaml-6.0.3-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:501a031947e3a9025ed4405a168e6ef5ae3126c59f90ce0cd6f2bfc477be31b7"},
+    {file = "pyyaml-6.0.3-cp314-cp314-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:b3bc83488de33889877a0f2543ade9f70c67d66d9ebb4ac959502e12de895788"},
+    {file = "pyyaml-6.0.3-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:c458b6d084f9b935061bc36216e8a69a7e293a2f1e68bf956dcd9e6cbcd143f5"},
+    {file = "pyyaml-6.0.3-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:7c6610def4f163542a622a73fb39f534f8c101d690126992300bf3207eab9764"},
+    {file = "pyyaml-6.0.3-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:5190d403f121660ce8d1d2c1bb2ef1bd05b5f68533fc5c2ea899bd15f4399b35"},
+    {file = "pyyaml-6.0.3-cp314-cp314-win_amd64.whl", hash = "sha256:4a2e8cebe2ff6ab7d1050ecd59c25d4c8bd7e6f400f5f82b96557ac0abafd0ac"},
+    {file = "pyyaml-6.0.3-cp314-cp314-win_arm64.whl", hash = "sha256:93dda82c9c22deb0a405ea4dc5f2d0cda384168e466364dec6255b293923b2f3"},
+    {file = "pyyaml-6.0.3-cp314-cp314t-macosx_10_13_x86_64.whl", hash = "sha256:02893d100e99e03eda1c8fd5c441d8c60103fd175728e23e431db1b589cf5ab3"},
+    {file = "pyyaml-6.0.3-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:c1ff362665ae507275af2853520967820d9124984e0f7466736aea23d8611fba"},
+    {file = "pyyaml-6.0.3-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:6adc77889b628398debc7b65c073bcb99c4a0237b248cacaf3fe8a557563ef6c"},
+    {file = "pyyaml-6.0.3-cp314-cp314t-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:a80cb027f6b349846a3bf6d73b5e95e782175e52f22108cfa17876aaeff93702"},
+    {file = "pyyaml-6.0.3-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:00c4bdeba853cc34e7dd471f16b4114f4162dc03e6b7afcc2128711f0eca823c"},
+    {file = "pyyaml-6.0.3-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:66e1674c3ef6f541c35191caae2d429b967b99e02040f5ba928632d9a7f0f065"},
+    {file = "pyyaml-6.0.3-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:16249ee61e95f858e83976573de0f5b2893b3677ba71c9dd36b9cf8be9ac6d65"},
+    {file = "pyyaml-6.0.3-cp314-cp314t-win_amd64.whl", hash = "sha256:4ad1906908f2f5ae4e5a8ddfce73c320c2a1429ec52eafd27138b7f1cbe341c9"},
+    {file = "pyyaml-6.0.3-cp314-cp314t-win_arm64.whl", hash = "sha256:ebc55a14a21cb14062aa4162f906cd962b28e2e9ea38f9b4391244cd8de4ae0b"},
+    {file = "pyyaml-6.0.3-cp39-cp39-macosx_10_13_x86_64.whl", hash = "sha256:b865addae83924361678b652338317d1bd7e79b1f4596f96b96c77a5a34b34da"},
+    {file = "pyyaml-6.0.3-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:c3355370a2c156cffb25e876646f149d5d68f5e0a3ce86a5084dd0b64a994917"},
+    {file = "pyyaml-6.0.3-cp39-cp39-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:3c5677e12444c15717b902a5798264fa7909e41153cdf9ef7ad571b704a63dd9"},
+    {file = "pyyaml-6.0.3-cp39-cp39-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:5ed875a24292240029e4483f9d4a4b8a1ae08843b9c54f43fcc11e404532a8a5"},
+    {file = "pyyaml-6.0.3-cp39-cp39-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:0150219816b6a1fa26fb4699fb7daa9caf09eb1999f3b70fb6e786805e80375a"},
+    {file = "pyyaml-6.0.3-cp39-cp39-musllinux_1_2_aarch64.whl", hash = "sha256:fa160448684b4e94d80416c0fa4aac48967a969efe22931448d853ada8baf926"},
+    {file = "pyyaml-6.0.3-cp39-cp39-musllinux_1_2_x86_64.whl", hash = "sha256:27c0abcb4a5dac13684a37f76e701e054692a9b2d3064b70f5e4eb54810553d7"},
+    {file = "pyyaml-6.0.3-cp39-cp39-win32.whl", hash = "sha256:1ebe39cb5fc479422b83de611d14e2c0d3bb2a18bbcb01f229ab3cfbd8fee7a0"},
+    {file = "pyyaml-6.0.3-cp39-cp39-win_amd64.whl", hash = "sha256:2e71d11abed7344e42a8849600193d15b6def118602c4c176f748e4583246007"},
+    {file = "pyyaml-6.0.3.tar.gz", hash = "sha256:d76623373421df22fb4cf8817020cbb7ef15c725b9d5e45f17e189bfc384190f"},
+]
+
+[[package]]
+name = "regex"
+version = "2026.9.29"
+description = "Alternative regular expression module, to replace re."
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "regex-2026.9.29-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:9916fda742cd4eede63b286f58c06718324265d727ce0856eb1aac86d0d150d6"},
+    {file = "regex-2026.9.29-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:8873c4a11c50b9989168881aeb3f08859f469d809941866aa1feefd8be5431f6"},
+    {file = "regex-2026.9.29-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:1d9fe8091b2e89d470df68a9331111ed008ae8aae6bf1e8e1fba4086a495c84e"},
+    {file = "regex-2026.9.29-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:fb00027a09a8f9f08028b40dce4c933cf73e4833240ed356583fdc9cfa721566"},
+    {file = "regex-2026.9.29-cp310-cp310-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:14e953ff3607c92d7675bf79c4d4509ef6782aa8c08509f179f9b3d6d0679e86"},
+    {file = "regex-2026.9.29-cp310-cp310-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:0476e5bcbe6e1ba3d1c4cc7bbb1c3ba78e3b979b5c8a88d0a6a8cdd4992b8c84"},
+    {file = "regex-2026.9.29-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:4fb41211d2333eb930a51e0546a65999761cf1f572a4da56ef9b8a62966c06f2"},
+    {file = "regex-2026.9.29-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:edf06545875f3efa31560d94121e95c7fd70d98b1dfedc0157097d79b13b52ea"},
+    {file = "regex-2026.9.29-cp310-cp310-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:6398d5145689503412cc1748895242598d8846b8967b851133b20dc2ed1e21e8"},
+    {file = "regex-2026.9.29-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:45010bcfe66df41522d56c9b6114e87ecc597a08970ff6a2ced24415c141ae5f"},
+    {file = "regex-2026.9.29-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:a5758353650079898dc1b2b0e95aa51fa23a30d020e06f62c430dd08ee56cdd8"},
+    {file = "regex-2026.9.29-cp310-cp310-musllinux_1_2_riscv64.whl", hash = "sha256:6f7121a8914ed13fcfe2099f895341bfb789f004d4c5a0bdece8fa667da10849"},
+    {file = "regex-2026.9.29-cp310-cp310-musllinux_1_2_s390x.whl", hash = "sha256:b9d74e4eee9ddb64c2e92d5d61472c59c21684c059eb7b68767be9628e977859"},
+    {file = "regex-2026.9.29-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:143533cc4b6fbc5b95aca0a5b8d541088d374831593def000ec89322c220221d"},
+    {file = "regex-2026.9.29-cp310-cp310-win32.whl", hash = "sha256:b84f186a7f0536fe4ff9a9fa12d06d007b9b71d4b5352ddcc41f59ad6522a312"},
+    {file = "regex-2026.9.29-cp310-cp310-win_amd64.whl", hash = "sha256:23ae6fdad9e63e54038f5ef78aba2933faca61e24d432786589e737bc5522ebb"},
+    {file = "regex-2026.9.29-cp310-cp310-win_arm64.whl", hash = "sha256:c0094897d7d01f184b2d7fe8c56c66d64efe01b31f4b7d34205b391387df1111"},
+    {file = "regex-2026.9.29-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:6abb75ab16bc3281714a5b99548a2225db70dba1f995f6d7f7419b76eb5a8fbe"},
+    {file = "regex-2026.9.29-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:b7b893976e7fe42053da64f2aa27239c24252fd2ec6df471e1be197c0addc3b1"},
+    {file = "regex-2026.9.29-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:066d0e3dbfdd739bce2bf8c2a41dd16f73e3d8adc2eb06dd803a36a307f56075"},
+    {file = "regex-2026.9.29-cp311-cp311-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:7020ed44df30b3aa492c00ee3b52d0548c1f30c2c6c5bb13ae897680900d3413"},
+    {file = "regex-2026.9.29-cp311-cp311-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:ae4613d7d9dda60fcba95f846cc6f808017f1843f392cf9daad14a6534493d71"},
+    {file = "regex-2026.9.29-cp311-cp311-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:bec37990e3d6121f29ecfb594bd8f1bf009e9f7926daba2e50e3b27d3892a783"},
+    {file = "regex-2026.9.29-cp311-cp311-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:612b709381c0355b70d89cdb51b7f670591ed5cbbc0e3b5337488019dc667b65"},
+    {file = "regex-2026.9.29-cp311-cp311-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:a760da040b47767b4b873adfb7c3b691e9ba2fc60f113f9d0b88f1a62f323e85"},
+    {file = "regex-2026.9.29-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:49ee178ca31c94621294bf9b8b676a92a2e6bba8af0529591753719e57edb621"},
+    {file = "regex-2026.9.29-cp311-cp311-musllinux_1_2_ppc64le.whl", hash = "sha256:5eeb8edc6110d9194a4d0d54610f64c37a31c605b5dbb7e407fc6ec7fa34a4a1"},
+    {file = "regex-2026.9.29-cp311-cp311-musllinux_1_2_riscv64.whl", hash = "sha256:ccb64d887a9db1cd76dbc0f92051a1a478a2a67e7f56c62d915cb881d7734704"},
+    {file = "regex-2026.9.29-cp311-cp311-musllinux_1_2_s390x.whl", hash = "sha256:9e4482589065c8ecd761cff522dcd85f2d39e62f551e37e025d1c7d54772def3"},
+    {file = "regex-2026.9.29-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:d60030baaa7bfbb02d650c126cdcddcb6e33dbff14d819434c8fa2fdcaeeeba5"},
+    {file = "regex-2026.9.29-cp311-cp311-win32.whl", hash = "sha256:18ae8eed4526e35bdb754d61562b90bf5c00a67fdcf3cc1380dd59597486631b"},
+    {file = "regex-2026.9.29-cp311-cp311-win_amd64.whl", hash = "sha256:1043aedf5917caa861bcb25a9c11460049656bdf0017a90a309fa8f255467725"},
+    {file = "regex-2026.9.29-cp311-cp311-win_arm64.whl", hash = "sha256:352cf115a810b357caa35193ab656ecf5ef41056855e82f292c99e8514f8d954"},
+    {file = "regex-2026.9.29-cp312-cp312-macosx_10_13_universal2.whl", hash = "sha256:dc79d36d0618752265f0d575915bdc5c5130ecb9c9f6b3bcefeae32e4bdfafcf"},
+    {file = "regex-2026.9.29-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:3a21a9509d0ee88e7a70e1ad228cd2f0e0fd1e187458db132e8a8d18c97daf9d"},
+    {file = "regex-2026.9.29-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:f57dc6b8fef170f105d2cf5cdce254f47b137d7755086cf7050f47e16582abba"},
+    {file = "regex-2026.9.29-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:f93bc1c3486ef3747e07c9d7c1d0a147b8fbaab975f80e348aed6f71309dfaca"},
+    {file = "regex-2026.9.29-cp312-cp312-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:9e1d3a4cb7993b708f0ada8d0c84590efd853f169e7147d2202c9da503180242"},
+    {file = "regex-2026.9.29-cp312-cp312-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:dabee8f4935e731fb46b2a3091bdda0d3d94b3bbfb907d2b4f12eefce4009619"},
+    {file = "regex-2026.9.29-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:39ab5894d971f9ac68baa6eca5c50387db579cfcacf36ae8df3feceb1815e6d0"},
+    {file = "regex-2026.9.29-cp312-cp312-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:c1a9a6651197fbed6f0212591418b9def774fc3f8324f78d1bf0e6a63e5f8aa1"},
+    {file = "regex-2026.9.29-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:87fb80cbe3557e27e7b28b995c2b2eedf689b8886f941ab93e0e288f0976518a"},
+    {file = "regex-2026.9.29-cp312-cp312-musllinux_1_2_ppc64le.whl", hash = "sha256:3c5c2ef13797466aa64170cbb66ad98a32351dd4127694cea7199f80f213750d"},
+    {file = "regex-2026.9.29-cp312-cp312-musllinux_1_2_riscv64.whl", hash = "sha256:59b49507f47479e299a9e1bc41b5cb83a7afda0540625f1dbae886615978acbf"},
+    {file = "regex-2026.9.29-cp312-cp312-musllinux_1_2_s390x.whl", hash = "sha256:0dd8af32e9f7b56b7f95cc1fd79b23054c3bdc172392ae560acc24d57b7ffe71"},
+    {file = "regex-2026.9.29-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:db5e82ba15c142425b8406690032df89e39cca4a2e8afbbb9a3d84edc2373ac3"},
+    {file = "regex-2026.9.29-cp312-cp312-win32.whl", hash = "sha256:d0c3082bf79bcd6a614d55916590ad4b8f93200e10b97f463ea5d9d07c9b5f23"},
+    {file = "regex-2026.9.29-cp312-cp312-win_amd64.whl", hash = "sha256:fdd88ed5e20b1bcdd234421e454962c971aa44b653bdb7f1ea9ef683e90fb649"},
+    {file = "regex-2026.9.29-cp312-cp312-win_arm64.whl", hash = "sha256:4fe97894d1b306c919b4e50def1e6f6c522f4d03a7283811f4d108f1ce5d3ac2"},
+    {file = "regex-2026.9.29-cp313-cp313-macosx_10_13_universal2.whl", hash = "sha256:f1a0d5117230dd46b399a30a38afa44f79c99f3168988fdc4f425c3f928b39df"},
+    {file = "regex-2026.9.29-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:f0fe9834e5aeccaf19a0d8feb296d66a24be1a7c9922002f842a682cd5abb787"},
+    {file = "regex-2026.9.29-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:c90fcf7804ea0a54b896ce0f2b9565350220b8d4890fd0db461a476a4c687963"},
+    {file = "regex-2026.9.29-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:e11edba5bc344a32b029a7af9d4b3173982dd79eeafa0b9dbd787364414b0509"},
+    {file = "regex-2026.9.29-cp313-cp313-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:bb90e7177944b6684738c1fc36aabd2dd00d1de3be7dbe09f91e196f1bc0dc81"},
+    {file = "regex-2026.9.29-cp313-cp313-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:d06fcdecc10fc7954d7c8f27a03c96055fe525274dc84a7b0dbdc3d6b9e03dab"},
+    {file = "regex-2026.9.29-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:d49c18f1ea294cf4adde2e5ac256e98c82ea9d708462ce4bf799dffa7cfe8a2c"},
+    {file = "regex-2026.9.29-cp313-cp313-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:3e778bfccd63075167709136afbc251c1f683758d5bf49c803c60ac3f894ce6b"},
+    {file = "regex-2026.9.29-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:686ac5350fceae63830bb98805fcb8039325bf4c06d9f6f048ff65229d5bffa5"},
+    {file = "regex-2026.9.29-cp313-cp313-musllinux_1_2_ppc64le.whl", hash = "sha256:26ec4ccce55aa533fbd603d08911b01101a8fcfec987845ac3ae2c7087b2bde3"},
+    {file = "regex-2026.9.29-cp313-cp313-musllinux_1_2_riscv64.whl", hash = "sha256:a655d34b2a6943af32401f3d94f72e9d731f6ad16285815550bf2b4ee69d420a"},
+    {file = "regex-2026.9.29-cp313-cp313-musllinux_1_2_s390x.whl", hash = "sha256:0c992c19cd45058a4b92f68f139c93db168b48fb1f322c9a7cd620806afb6b51"},
+    {file = "regex-2026.9.29-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:ebb8912f565b8cdbbf27debfe00df04202c20e2f651b9e32767930c5eace3621"},
+    {file = "regex-2026.9.29-cp313-cp313-win32.whl", hash = "sha256:4d7d93613b01b0199961330e49cfc52d479b3d5776c56c691db31130c0a07d91"},
+    {file = "regex-2026.9.29-cp313-cp313-win_amd64.whl", hash = "sha256:61956f074ecd123f55adca68ee3eab46e6a07ad3f8e64e6db95dfacb444f55c4"},
+    {file = "regex-2026.9.29-cp313-cp313-win_arm64.whl", hash = "sha256:bfc71e6d970419c1309b3640305298643e2a734cad3f7cfb6d2ddee4175ab53d"},
+    {file = "regex-2026.9.29-cp314-cp314-macosx_10_15_universal2.whl", hash = "sha256:957bb708e8057ab1649ba566456429d691ec9b90d1c9ad1af1ba7ffbbeaf05f2"},
+    {file = "regex-2026.9.29-cp314-cp314-macosx_10_15_x86_64.whl", hash = "sha256:c9b602fae1e00b7c035d661ce85575365719192a7b46784bd71cf64c68053aa0"},
+    {file = "regex-2026.9.29-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:0166844493626c5015c6088ee15c9ca2fd060ca15b7641d1657da6a58432ae33"},
+    {file = "regex-2026.9.29-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:b97a38fb4c732b6832db6bf108963adbcd82ef1268ba2025dce390f45af75efa"},
+    {file = "regex-2026.9.29-cp314-cp314-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:a540abfab208e1b7ef2df231c40ef3b6cbb30a0aad6204e9b6a81c10a6794628"},
+    {file = "regex-2026.9.29-cp314-cp314-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:ddfa987262763c3c22a8367d2a49c244b018a74c3a8e3ab1a864119ad45c5633"},
+    {file = "regex-2026.9.29-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:2f7f7aa47b229f2b39a2ae2596d2ad5625d77b5eb9856fac2dab3eb506cdd0a0"},
+    {file = "regex-2026.9.29-cp314-cp314-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:d9b77b25b4f395f92de6099ab08e8ae2bc7e51dfe157f22900902243a5cc90c7"},
+    {file = "regex-2026.9.29-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:34b6925af9853bf461950e6508910f179fd6e9b1a7ec8548e069606b7e51a26b"},
+    {file = "regex-2026.9.29-cp314-cp314-musllinux_1_2_ppc64le.whl", hash = "sha256:addd736a0547d553283adaf4e05d7104e7f2c7b0b092e9b4d28756825f14531f"},
+    {file = "regex-2026.9.29-cp314-cp314-musllinux_1_2_riscv64.whl", hash = "sha256:fe3fa1dd453ed5c7f5ea23a26218329790ed7197a99b90e94330e313959a7f52"},
+    {file = "regex-2026.9.29-cp314-cp314-musllinux_1_2_s390x.whl", hash = "sha256:0cc63b5e47c12a48d90c7e9d7de6a035dd14f62868aaedbb4e0ff8ba2b8bfe7b"},
+    {file = "regex-2026.9.29-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:724184b4aafed865e4f13ca313fdcb43024300c028ec67319cfa16847d84685e"},
+    {file = "regex-2026.9.29-cp314-cp314-win32.whl", hash = "sha256:c6c8fabf1dafc1f1ddcbb67896d3f93efb092e8c4b6322d7389b944e76a484e5"},
+    {file = "regex-2026.9.29-cp314-cp314-win_amd64.whl", hash = "sha256:1c2a0026062abcc321a53db4a185ceba0b59a66b5d37b0808917a88b55a5257f"},
+    {file = "regex-2026.9.29-cp314-cp314-win_arm64.whl", hash = "sha256:121a76a0985db80ceae9e171c337f8c927868e37d01b54e3ce87bc87f9c6a208"},
+    {file = "regex-2026.9.29-cp314-cp314t-macosx_10_15_universal2.whl", hash = "sha256:e31f72490b7c12f7790e1e25c3afffd20503ee1bfb43461d7838b871ff244b19"},
+    {file = "regex-2026.9.29-cp314-cp314t-macosx_10_15_x86_64.whl", hash = "sha256:80ea96f5c1a30bf09007d48466521d9c294bebe197c708c3359096e3e3691632"},
+    {file = "regex-2026.9.29-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:554bffadcbcb6d5f4e5fb10a61cc52084b9a63d1dab5f10bcd2c4343972e8e2c"},
+    {file = "regex-2026.9.29-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:864e9b87ac33c3fb9fb4ad48166d4fdb579c351d5c77deb0d34bccb36a775cd9"},
+    {file = "regex-2026.9.29-cp314-cp314t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:044265d77d94f5e3cb2fd72c76723807c429cb8c533e9d4672d0334a6f14f588"},
+    {file = "regex-2026.9.29-cp314-cp314t-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:2089fe39c406784d90101c726755ffa1497bb74638fd434300d2b88006186de8"},
+    {file = "regex-2026.9.29-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:0def9fb6abac55492d6d51cddb7225d07d6f279e774e0adc08569a54a5fc8d46"},
+    {file = "regex-2026.9.29-cp314-cp314t-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:888d60953908dcf761aa320c3e390ab8556efbdb551ace63921de90f6ae0848d"},
+    {file = "regex-2026.9.29-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:ed511a0708e2297e1d6431e7fb217e3402791e491e02da800658ace4973df1bb"},
+    {file = "regex-2026.9.29-cp314-cp314t-musllinux_1_2_ppc64le.whl", hash = "sha256:e1172147d28d8fbcf8cb8d26c41506169f5ad8fe9ec969cb116835a19d4d8eca"},
+    {file = "regex-2026.9.29-cp314-cp314t-musllinux_1_2_riscv64.whl", hash = "sha256:92f05c9c42bde5785dc48770bc2194d9f7442544156f951e19cd31b096cec562"},
+    {file = "regex-2026.9.29-cp314-cp314t-musllinux_1_2_s390x.whl", hash = "sha256:f37964e4a5e993d2fd45147741e9dff7f34a2d8c00ab94c4ea0514a4677f959e"},
+    {file = "regex-2026.9.29-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:951733b1bbdb71e377cec567b409f1a7881b47cfcad84121aa74cb575fa425ea"},
+    {file = "regex-2026.9.29-cp314-cp314t-win32.whl", hash = "sha256:65b408d8fcb273e3499e7ef2ce796810da1becd208c7fb4373692a242d79d461"},
+    {file = "regex-2026.9.29-cp314-cp314t-win_amd64.whl", hash = "sha256:bf48516e35cf848390ea68850aba53e7c333720d2945b4d2c25b69fc5171723f"},
+    {file = "regex-2026.9.29-cp314-cp314t-win_arm64.whl", hash = "sha256:9173db3be74a35cb6731701094b98120f7ee4876a287882a59cdea1fa7da342f"},
+    {file = "regex-2026.9.29-cp315-cp315-macosx_10_15_universal2.whl", hash = "sha256:c3589f40749acce747510bf5d589d54e376cb0930ea58b35effac97e5312b0c1"},
+    {file = "regex-2026.9.29-cp315-cp315-macosx_10_15_x86_64.whl", hash = "sha256:32ab11df9677ca80bcbb5fe4eb1da9109a5019239a054836efc6fa1c64e683cf"},
+    {file = "regex-2026.9.29-cp315-cp315-macosx_11_0_arm64.whl", hash = "sha256:7c03031610e3e6ed1768a2b7a8fc84637c1257b50c5eacaf094c6e17a84fc563"},
+    {file = "regex-2026.9.29-cp315-cp315-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:42e82e578c904445d4c8a35b8f28052cf567593215fa5db06266fbc6f77aaa2e"},
+    {file = "regex-2026.9.29-cp315-cp315-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:0b65c72739f981377c9c22e0c5c3cd7f42da7bd8a3c9209330fac772c7d893ed"},
+    {file = "regex-2026.9.29-cp315-cp315-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:4408b2b27a95ca8cc48b7411945753773353b5c93b307754781086c99d3a576f"},
+    {file = "regex-2026.9.29-cp315-cp315-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:a714befaacbd10092ffe4cea0d3c5f008fb9efe9bc322c715bcdfdee414b9a3d"},
+    {file = "regex-2026.9.29-cp315-cp315-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:33026515aebc0e70d1c89978e53e8d695d35d9e472f8d5b34465ba3c74028650"},
+    {file = "regex-2026.9.29-cp315-cp315-musllinux_1_2_aarch64.whl", hash = "sha256:31b003f9a070335e2a8233ee9b14a3ca8e6d792012ae011f741bf0aaf11744c5"},
+    {file = "regex-2026.9.29-cp315-cp315-musllinux_1_2_ppc64le.whl", hash = "sha256:c03c6eb6ece86dfdcbb34799efaa339b093132e1aceed491ba5e08fe06cdf699"},
+    {file = "regex-2026.9.29-cp315-cp315-musllinux_1_2_riscv64.whl", hash = "sha256:a5300757f8a68f5b6cc33f57338d72a0e3589c5cc9ad5f8504ea06f028be582a"},
+    {file = "regex-2026.9.29-cp315-cp315-musllinux_1_2_s390x.whl", hash = "sha256:80c7cadd3fd2bfde5df8aa0787e315812cad0c313a753095d02f4c2b6c01677b"},
+    {file = "regex-2026.9.29-cp315-cp315-musllinux_1_2_x86_64.whl", hash = "sha256:3f1e6cb402a89457582cd696f982559217d13484a193202c394015297968c86d"},
+    {file = "regex-2026.9.29-cp315-cp315-win32.whl", hash = "sha256:a64b85a4760337cfefdb27d42da6ed8b58e8cde3f2d57b6ef43e76ef6ea9ef47"},
+    {file = "regex-2026.9.29-cp315-cp315-win_amd64.whl", hash = "sha256:b3e445b66c80b4eb4234e855ce94d9adc183eedbd632816228d89930b91b2c5b"},
+    {file = "regex-2026.9.29-cp315-cp315-win_arm64.whl", hash = "sha256:8f39588af4731c8923c26810eb3b33f76f17633985e40f59c3cd45a33805a895"},
+    {file = "regex-2026.9.29-cp315-cp315t-macosx_10_15_universal2.whl", hash = "sha256:fb99cc9d45f48895d9d67f6a0b8a57f08d39c174d9f25ad97a313e0470267b1c"},
+    {file = "regex-2026.9.29-cp315-cp315t-macosx_10_15_x86_64.whl", hash = "sha256:720537c7ea6f80dc61913184edb0ce2497a306b39ef19f28505b322553d52bdb"},
+    {file = "regex-2026.9.29-cp315-cp315t-macosx_11_0_arm64.whl", hash = "sha256:0fd2c901cc307a745ad4bc87f20060d7a0825a3371d1e93488af22e7a387f78f"},
+    {file = "regex-2026.9.29-cp315-cp315t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:b11b589e00095ec69cf79841a76360f9b079e95b0368a25b5ebb951ab0c157ff"},
+    {file = "regex-2026.9.29-cp315-cp315t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:d7cab119d0df0b9413f106b4d7fc34f2872d3574ed3806fb48959c830b1537da"},
+    {file = "regex-2026.9.29-cp315-cp315t-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:b89efc38431793d28b7cd91227e2f952ad7c48df19132b17f43a5fec3c14143b"},
+    {file = "regex-2026.9.29-cp315-cp315t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:80a5ea3b4fd9d6a5b9a44f7976a9acaaab35aa3c1f6b29e5bd857dfabaded223"},
+    {file = "regex-2026.9.29-cp315-cp315t-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:19959129885356df0e97556856f77eb2888380dac18bed075a7c05c5128c618d"},
+    {file = "regex-2026.9.29-cp315-cp315t-musllinux_1_2_aarch64.whl", hash = "sha256:6a1a824fbed817e0a891103886b68f063b1e83cc51bc97192a90a60195a9291f"},
+    {file = "regex-2026.9.29-cp315-cp315t-musllinux_1_2_ppc64le.whl", hash = "sha256:1ba8c6a416569ce0d37e83e28a254a61dc99a419084dfb6476cea02d997f74fa"},
+    {file = "regex-2026.9.29-cp315-cp315t-musllinux_1_2_riscv64.whl", hash = "sha256:446654b29bfaa30500d80947eda42cef1449dc8a87f4e3cf061cc8485d3a1f0b"},
+    {file = "regex-2026.9.29-cp315-cp315t-musllinux_1_2_s390x.whl", hash = "sha256:bf3c49863c23a1ad6da9c30351aed6cff8d5ddbeb63c5c8420ae54e98c7d0138"},
+    {file = "regex-2026.9.29-cp315-cp315t-musllinux_1_2_x86_64.whl", hash = "sha256:01000ddf0e3ffef97f2413ceb514f6313040106b6d18a03ee00a4fe35c1eb1db"},
+    {file = "regex-2026.9.29-cp315-cp315t-win32.whl", hash = "sha256:c4e38dd8f39c43a91d2410ad2b85610701b0979342c3df1d69eaf8e838c757d8"},
+    {file = "regex-2026.9.29-cp315-cp315t-win_amd64.whl", hash = "sha256:e2c89e9b762c57f59d5e99ee8b20202adb892e35f8d3485741340999ca55058e"},
+    {file = "regex-2026.9.29-cp315-cp315t-win_arm64.whl", hash = "sha256:e8c65ef3862a8ad6e86492b6ed9327805dd66904c012bd3649dc67d822ed6c34"},
+    {file = "regex-2026.9.29.tar.gz", hash = "sha256:8b5fcc4771732191b2b7d1dd68d8f0353f47f8d90b6150f6dce58bf1112442cb"},
+]
+
+[[package]]
+name = "requests"
+version = "2.34.2"
+description = "Python HTTP for Humans."
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "requests-2.34.2-py3-none-any.whl", hash = "sha256:2a0d60c172f83ac6ab31e4554906c0f3b3588d37b5cb939b1c061f4907e278e0"},
+    {file = "requests-2.34.2.tar.gz", hash = "sha256:f288924cae4e29463698d6d60bc6a4da69c89185ad1e0bcc4104f584e960b9ed"},
+]
+
+[package.dependencies]
+certifi = ">=2023.5.7"
+charset_normalizer = ">=2,<4"
+idna = ">=2.5,<4"
+urllib3 = ">=1.26,<3"
+
+[package.extras]
+socks = ["PySocks (>=1.5.6,!=1.5.7)"]
+use-chardet-on-py3 = ["chardet (>=3.0.2,<8)"]
+
+[[package]]
+name = "rich"
+version = "15.0.0"
+description = "Render rich text, tables, progress bars, syntax highlighting, markdown and more to the terminal"
+optional = false
+python-versions = ">=3.9.0"
+groups = ["main"]
+files = [
+    {file = "rich-15.0.0-py3-none-any.whl", hash = "sha256:33bd4ef74232fb73fe9279a257718407f169c09b78a87ad3d296f548e27de0bb"},
+    {file = "rich-15.0.0.tar.gz", hash = "sha256:edd07a4824c6b40189fb7ac9bc4c52536e9780fbbfbddf6f1e2502c31b068c36"},
+]
+
+[package.dependencies]
+markdown-it-py = ">=2.2.0"
+pygments = ">=2.13.0,<3.0.0"
+
+[package.extras]
+jupyter = ["ipywidgets (>=7.5.1,<9)"]
+
+[[package]]
+name = "safetensors"
+version = "0.8.0"
+description = ""
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "safetensors-0.8.0-cp310-abi3-macosx_10_12_x86_64.whl", hash = "sha256:c554f85858e05226d3c2828e32395e677434685d6d94594a41643361c5e837f0"},
+    {file = "safetensors-0.8.0-cp310-abi3-macosx_11_0_arm64.whl", hash = "sha256:c80201d22cbf405b80647a60ada77bba06c8fba2da2743ba1e89cdcc39a81f25"},
+    {file = "safetensors-0.8.0-cp310-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:7a46e5ff292c356d6991e60942ba7f79817682d3a2cef0702136448cb9c4d235"},
+    {file = "safetensors-0.8.0-cp310-abi3-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:4124502b78f03534117c848f87a39b8f31e577b15eff423bf8bfb95f2a8c30d0"},
+    {file = "safetensors-0.8.0-cp310-abi3-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:7bc0a787ba8a35be368ee3574edfa2b1ad389eebd0a72e482ae275490e3f6c98"},
+    {file = "safetensors-0.8.0-cp310-abi3-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:040070828e36dc8e122178bbbd5830ff9e97920affb84cbe0f46442497bed358"},
+    {file = "safetensors-0.8.0-cp310-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:fd6f3f93c9a0a7cc2788ee63fb763353d4bd2e89b0751bc78fcf7dda00bea774"},
+    {file = "safetensors-0.8.0-cp310-abi3-manylinux_2_31_riscv64.whl", hash = "sha256:fcdd41ec4628fee5799f807c73c353629130fbd942aa23d83c623dd6c9d52d78"},
+    {file = "safetensors-0.8.0-cp310-abi3-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:8e9f537aa183a38ace122d27303dcd986b26bd2a7591f9181d7f0c396f4677ca"},
+    {file = "safetensors-0.8.0-cp310-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:87eec7ffed2b809f05a398a8becb7d013f19f7837cd15d9748580d6cf30dbaf4"},
+    {file = "safetensors-0.8.0-cp310-abi3-musllinux_1_2_armv7l.whl", hash = "sha256:4a95ae2b05d7726d751da4ebf626a2ca782b706e101bd894c95bc2450b1cffcc"},
+    {file = "safetensors-0.8.0-cp310-abi3-musllinux_1_2_i686.whl", hash = "sha256:3ae091f16662658bdc019a4ff6cb4c085bb7d725eb5978b183ffd265863b6d2d"},
+    {file = "safetensors-0.8.0-cp310-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:8e080062fcde23be189565e1c3305d16751a218ecf9412c8601e64204eb6f846"},
+    {file = "safetensors-0.8.0-cp310-abi3-win32.whl", hash = "sha256:2ddf52eac562eda224f99acfa7889d02968c1fd59a5b011ae7d8137c37e9c02d"},
+    {file = "safetensors-0.8.0-cp310-abi3-win_amd64.whl", hash = "sha256:096ec1a98435df7beb08853bb5aa9081a84f23d0adc67ed1a0a10550f608373f"},
+    {file = "safetensors-0.8.0-cp310-abi3-win_arm64.whl", hash = "sha256:f7838e5135a406ad3e02efdcb8cf2e5397d368b0154537c4fec682dbc544d452"},
+    {file = "safetensors-0.8.0.tar.gz", hash = "sha256:fabaf3e0f18a6618d9b36560682562157f77c2b71fcffc7b432be2baed9d753d"},
+]
+
+[package.extras]
+all = ["safetensors[convert]", "safetensors[jax]", "safetensors[numpy]", "safetensors[paddlepaddle]", "safetensors[quality]", "safetensors[testing]", "safetensors[torch]"]
+convert = ["huggingface-hub (>=1.4)", "safetensors[torch]"]
+dev = ["safetensors[all]", "safetensors[pinned-tf]"]
+jax = ["flax (>=0.6.3)", "jax (>=0.3.25)", "jaxlib (>=0.3.25)", "safetensors[numpy]"]
+mlx = ["mlx (>=0.0.9)"]
+numpy = ["numpy (>=1.24.6)"]
+paddlepaddle = ["paddlepaddle (>=2.4.1)", "safetensors[numpy]"]
+pinned-tf = ["safetensors[numpy]", "tensorflow (==2.18.0)"]
+quality = ["ruff"]
+tensorflow = ["safetensors[numpy]", "tensorflow (>=2.11.0)"]
+testing = ["fsspec (>=2024.6.0)", "h5py (>=3.7.0)", "hypothesis (>=6.70.2)", "pytest (>=9.0)", "pytest-benchmark (>=5.2)", "s3fs (>=2024.6.0)", "safetensors[numpy]", "setuptools-rust (>=1.12.0)"]
+tf-nightly = ["safetensors[numpy]", "tf-nightly"]
+torch = ["safetensors[numpy]", "torch (>=2.4)"]
+
+[[package]]
+name = "scikit-learn"
+version = "1.9.1"
+description = "A set of python modules for machine learning and data mining"
+optional = false
+python-versions = ">=3.11"
+groups = ["main"]
+files = [
+    {file = "scikit_learn-1.9.1-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:326c188f92084bf58664229f4578eeab6176313b37cd5dfc85abd92b94130c58"},
+    {file = "scikit_learn-1.9.1-cp311-cp311-macosx_12_0_arm64.whl", hash = "sha256:b48b2b5b41d9c5fbafef5f37b042f61110df3318ad2a45baf57287ea5b9ba5a2"},
+    {file = "scikit_learn-1.9.1-cp311-cp311-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:4298fcc01b3d8fa9768d36894e99cce0747b3b2dd73bfd80779e393769d0afab"},
+    {file = "scikit_learn-1.9.1-cp311-cp311-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:52a0703bbc07ad27f560fa63fa68e4c54dd735bfbbf65b4dd3c225dc7547b6df"},
+    {file = "scikit_learn-1.9.1-cp311-cp311-win_amd64.whl", hash = "sha256:220fa18152852a5ce29c49e1eaba9d44ec44631cd2e5cf65f5a40eafa5ab3412"},
+    {file = "scikit_learn-1.9.1-cp311-cp311-win_arm64.whl", hash = "sha256:8218cb8938d3031e0d23842838425490c229ecf3e99f666776e09d12ed902352"},
+    {file = "scikit_learn-1.9.1-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:0c0f8b5d09b44101cea2767f300680bada1ea27f976fe4b48b83950a4f55a49a"},
+    {file = "scikit_learn-1.9.1-cp312-cp312-macosx_12_0_arm64.whl", hash = "sha256:8c14ce41d561f7749f990b41d6703fe02c4669fbc485e598e069e0a1967b488e"},
+    {file = "scikit_learn-1.9.1-cp312-cp312-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:e4c20a6c017d820faa7ac8c783e3d0c6a9a2e297bf9f55332ca17cdf7fd4d04d"},
+    {file = "scikit_learn-1.9.1-cp312-cp312-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:e5d7b18a5b9dca241a74695f3275fa4c895a9dadc72b3d8df5fa9d1083c9b83e"},
+    {file = "scikit_learn-1.9.1-cp312-cp312-win_amd64.whl", hash = "sha256:4b59abb30618121cc46b45972d6bf53a7128b4df4cd346c6ca6f4d5f9031e49c"},
+    {file = "scikit_learn-1.9.1-cp312-cp312-win_arm64.whl", hash = "sha256:d5945a2908be62350e2978344e62b56c1552c2ca4f844ebf6277c94944d647dd"},
+    {file = "scikit_learn-1.9.1-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:2c2b312fd8c02951a364fa120ea08c1cec10d863466bf1701b013152d7537835"},
+    {file = "scikit_learn-1.9.1-cp313-cp313-macosx_12_0_arm64.whl", hash = "sha256:61cd968ab831a76d0ecbaf0347ab2270268716da28f94fd022497e3d6f205f13"},
+    {file = "scikit_learn-1.9.1-cp313-cp313-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:5990f9c69e431bfaddcde1a6d7c5355243e026bc9b9e560c13893b90dab53fb4"},
+    {file = "scikit_learn-1.9.1-cp313-cp313-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:55e79d6e9b0923f1a978179822bd43d7f5543f45e970a00fe861f43486380aba"},
+    {file = "scikit_learn-1.9.1-cp313-cp313-win_amd64.whl", hash = "sha256:2070f271e5375dc42c6bb93b461ab1c0aa5841d4009267e0cfd95a39dca94a43"},
+    {file = "scikit_learn-1.9.1-cp313-cp313-win_arm64.whl", hash = "sha256:613f0a783ca05aa844a4e1ac42d48425058f2c52be73f40f8cd98b7cd111acd6"},
+    {file = "scikit_learn-1.9.1-cp314-cp314-macosx_10_15_x86_64.whl", hash = "sha256:d5d117952769b563067656784e03c75a2d8235a7a05cf7fffa78a311e75aac08"},
+    {file = "scikit_learn-1.9.1-cp314-cp314-macosx_12_0_arm64.whl", hash = "sha256:8893bc6331f60f18d4ac75e12ed356e2dcf6a564bf767918b5b7ca54c8c8be49"},
+    {file = "scikit_learn-1.9.1-cp314-cp314-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:b5492cf2df5226691c32611de8734bcf42148c6547ae53c7f4e6b847793addc0"},
+    {file = "scikit_learn-1.9.1-cp314-cp314-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:993d332ff80e62efae9e39603b7e872297c418d780f01a01855269a3489c950f"},
+    {file = "scikit_learn-1.9.1-cp314-cp314-win_amd64.whl", hash = "sha256:ca9051447455dae341d4d591eece7deb2d8e3d1020298fc87a81fc51e4da8f53"},
+    {file = "scikit_learn-1.9.1-cp314-cp314-win_arm64.whl", hash = "sha256:90de6573f733a9fb79476ff1371af52a397d41c8b35f9146e20923db010d67b6"},
+    {file = "scikit_learn-1.9.1-cp314-cp314t-macosx_10_15_x86_64.whl", hash = "sha256:7b5cad1624de8b75e5b9ccb7b0ce1ff1d01306340a3efc56d5529c5ba92392eb"},
+    {file = "scikit_learn-1.9.1-cp314-cp314t-macosx_12_0_arm64.whl", hash = "sha256:d137ce8a6142029fb5c35bd82f470c40cd9e760e5e2f7694b362c497c4ab3fa2"},
+    {file = "scikit_learn-1.9.1-cp314-cp314t-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:66f852f7325b5070bc28329005aca76055a2def78faac039548ae889aeaa45a6"},
+    {file = "scikit_learn-1.9.1-cp314-cp314t-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:748bcb0a4cc04aec470652c9e5ec68450948e867387e7dfade647107ade68d25"},
+    {file = "scikit_learn-1.9.1-cp314-cp314t-win_amd64.whl", hash = "sha256:38cd925e893e5539be704d5edc64dbe081aacdab6b89d8c2977c1f6a7a453ce5"},
+    {file = "scikit_learn-1.9.1-cp314-cp314t-win_arm64.whl", hash = "sha256:b01e5b01735d38474127ca3f49319b592506225a87793b27559816b5c75cea39"},
+    {file = "scikit_learn-1.9.1-cp315-cp315-macosx_10_15_x86_64.whl", hash = "sha256:8dec64f31a6e0ec826aca6c1b39a51e16d946e400d4c0904316f3ca72ccfb825"},
+    {file = "scikit_learn-1.9.1-cp315-cp315-macosx_12_0_arm64.whl", hash = "sha256:e1b468241f4a7a9a7a0d6479ad3cc47681cc151a4046c530f2777c3d68f08942"},
+    {file = "scikit_learn-1.9.1-cp315-cp315-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:8ca869d0080a5723cde2d5a8b54a2da1ff7e68735a9e9adb3da1243183a0fa01"},
+    {file = "scikit_learn-1.9.1-cp315-cp315-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:6754b7cabfc3df0b1f7b38f7a344f559bbae9d82f0ac5e3d48cccbd19fdcefdf"},
+    {file = "scikit_learn-1.9.1-cp315-cp315-win_amd64.whl", hash = "sha256:52cfdb1fed3a34362dbc0bd96f2e761a66fd5724d6901629f5a558f1f3bd9849"},
+    {file = "scikit_learn-1.9.1-cp315-cp315-win_arm64.whl", hash = "sha256:ae6571a4828c6f5019bcd2b4125e5b18c0af3dbc9c99726c891f45f41335ec8e"},
+    {file = "scikit_learn-1.9.1-cp315-cp315t-macosx_10_15_x86_64.whl", hash = "sha256:48fefd8eb42bd4eec3e2d348149368ccd6d71987e20c30706a56a24eb86a6e73"},
+    {file = "scikit_learn-1.9.1-cp315-cp315t-macosx_12_0_arm64.whl", hash = "sha256:09f4d73049cd63575157f6b1060e06a8c83a4bd3488dbfaeedf35ccba7aad712"},
+    {file = "scikit_learn-1.9.1-cp315-cp315t-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:b3da53831534214322d9cb240fa6f390b36cf69eba727a6d4bd3238677630d70"},
+    {file = "scikit_learn-1.9.1-cp315-cp315t-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:caae15634feceafa2612566b109a3082d3293167fac388eedaf77bff66b51983"},
+    {file = "scikit_learn-1.9.1-cp315-cp315t-win_amd64.whl", hash = "sha256:ffbcbbbb44202fbe9bc64bced25a145759adb9ef010b3d37a8064958ac13df2a"},
+    {file = "scikit_learn-1.9.1-cp315-cp315t-win_arm64.whl", hash = "sha256:800dd22dd87fe97dcea484c24e85dd93cf1734d86bd74e668ad18f7967f4d1b5"},
+    {file = "scikit_learn-1.9.1.tar.gz", hash = "sha256:629cada3e33e2b9bf376cdc7614a47a4140b8aedc1d836579e359736fbd82977"},
+]
+
+[package.dependencies]
+joblib = ">=1.4.0"
+narwhals = ">=2.0.1"
+numpy = ">=1.24.1"
+scipy = ">=1.10.0"
+threadpoolctl = ">=3.5.0"
+
+[[package]]
+name = "scipy"
+version = "1.18.1"
+description = "Fundamental algorithms for scientific computing in Python"
+optional = false
+python-versions = ">=3.12"
+groups = ["main"]
+files = [
+    {file = "scipy-1.18.1-cp312-cp312-macosx_10_15_x86_64.whl", hash = "sha256:457fd7a2a8edeb044ab6ffbc0aa03ff6cd18491356e5e0c834d76ce621b916d1"},
+    {file = "scipy-1.18.1-cp312-cp312-macosx_12_0_arm64.whl", hash = "sha256:e708533e8b2ae2497d65346538a7dcc92814410b25b81432eac66de0f2af8265"},
+    {file = "scipy-1.18.1-cp312-cp312-macosx_14_0_arm64.whl", hash = "sha256:7bbf207c4453ce1ad2e00b17313852b33310b83090c2311bdaf97f93c0380d12"},
+    {file = "scipy-1.18.1-cp312-cp312-macosx_14_0_x86_64.whl", hash = "sha256:78c0665edead396b1abb4897c41a5c1d9bf090c8a637a4c20a61678e0a264e66"},
+    {file = "scipy-1.18.1-cp312-cp312-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:3c085faa2cfa879c5141df483f836f4d691045a078224a670fa570fa01612d89"},
+    {file = "scipy-1.18.1-cp312-cp312-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:f55fa87b6c612ecd6b058f167c53231b1d14e412efe361d3d6e38b3631c73218"},
+    {file = "scipy-1.18.1-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:c35d74ce0e193ff740c2f2be2ac913ddc232fe6c1ff40b26cfecb9c670c63314"},
+    {file = "scipy-1.18.1-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:d2924a03db38dc2e848bca2fe9f077dafb891480b91a00a0963a8cf86dfc31c1"},
+    {file = "scipy-1.18.1-cp312-cp312-win_amd64.whl", hash = "sha256:5e4d44984abc0020154ea81b247adeddcc3ac5527b975ff798bd1ba0adc513c2"},
+    {file = "scipy-1.18.1-cp312-cp312-win_arm64.whl", hash = "sha256:d65d448389b8436493abcf629cc94ad0cf32aecaf06e1acca1de53cc795f2f12"},
+    {file = "scipy-1.18.1-cp313-cp313-macosx_10_15_x86_64.whl", hash = "sha256:3ab3523da44749156e1f68b464dc56af11ae4cbc5c739a49d05f32b982eca9f3"},
+    {file = "scipy-1.18.1-cp313-cp313-macosx_12_0_arm64.whl", hash = "sha256:e6fb6a55cc0ba97b59a1f288fb86dc6fce8bdfc0fffcbfd015e3a954bf2a2d93"},
+    {file = "scipy-1.18.1-cp313-cp313-macosx_14_0_arm64.whl", hash = "sha256:ea324d9dd34c38bfb9bec8ca4d1b407db97dbb74029f566b8e322b1b6fe56fe6"},
+    {file = "scipy-1.18.1-cp313-cp313-macosx_14_0_x86_64.whl", hash = "sha256:75b00eb8fb802090aa903f4ea1c7f5a584779f967361e68b7e98e531cc2d7174"},
+    {file = "scipy-1.18.1-cp313-cp313-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:d416b16cccfd70fbf62400e84d0bb2f4e6af519a45557f1692c749b37f14b315"},
+    {file = "scipy-1.18.1-cp313-cp313-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:fdaf5ea890a6183d0565f51a61799d67081bd5b1cf03c5f4b3fd3732108625c9"},
+    {file = "scipy-1.18.1-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:c825cef2f49e46753726a7181a8e199804a912b29519ada542c6ebc654951899"},
+    {file = "scipy-1.18.1-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:e3b417bf8c2c7c16e8f58ad91db17783ec911ac16e7b50eb6eab6e809b4f5b07"},
+    {file = "scipy-1.18.1-cp313-cp313-win_amd64.whl", hash = "sha256:559ed65f60c1af5a03f3912605a1b5114f522c7c32fb23c3376ae8f03219fe28"},
+    {file = "scipy-1.18.1-cp313-cp313-win_arm64.whl", hash = "sha256:cd479fc04dd9401e3b4f49e76518768ef99c4f517a98c284eb091fd725719adf"},
+    {file = "scipy-1.18.1-cp314-cp314-macosx_10_15_x86_64.whl", hash = "sha256:83de5453a7799afc9048b4616bd085cef126e36412f0ea2f6370c36a2a3a51e7"},
+    {file = "scipy-1.18.1-cp314-cp314-macosx_12_0_arm64.whl", hash = "sha256:9554bcc6d715ee87a633a3cc8e7703c6628b100dd29cb8a2efc4c0533c7ff729"},
+    {file = "scipy-1.18.1-cp314-cp314-macosx_14_0_arm64.whl", hash = "sha256:011413b7426b75012840e35649e00fe0a2c3bae89fed433876e3a99251572efc"},
+    {file = "scipy-1.18.1-cp314-cp314-macosx_14_0_x86_64.whl", hash = "sha256:88f0e784020649f88ea48c9f5ddfa403bf9205820667c0914740b392035afb82"},
+    {file = "scipy-1.18.1-cp314-cp314-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:2d3ab0e8c69a17dd3559eab8cbb88f258e285c94d572c2719033f90f83290c89"},
+    {file = "scipy-1.18.1-cp314-cp314-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:ac0333bdf38309aa3dcbe7e3fa7ea29e7a2c37c6ea306a757b700ded8e4596ad"},
+    {file = "scipy-1.18.1-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:911de823097db8b63f034299d12662db93344e6ffa0b881cbb57748974b70168"},
+    {file = "scipy-1.18.1-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:95298364e251be3e60249facbeeca03631d3bb7584f85879516ec55ac717b81f"},
+    {file = "scipy-1.18.1-cp314-cp314-win_amd64.whl", hash = "sha256:78a0d7c918e74a232394117160e7e3db503377572a45bcef8826e4ab8a35feba"},
+    {file = "scipy-1.18.1-cp314-cp314-win_arm64.whl", hash = "sha256:cbf38d043c1aa4ab306e1ada6ab6eddacc3322a20b7af1b30bc93254b366fe09"},
+    {file = "scipy-1.18.1-cp314-cp314t-macosx_10_15_x86_64.whl", hash = "sha256:0fcb3c93519f27bb4f0c4b0f7802cdcaca7fcf93267b75edda2e9f4e8a55cbd7"},
+    {file = "scipy-1.18.1-cp314-cp314t-macosx_12_0_arm64.whl", hash = "sha256:ddef79fb382df40104a19bb7151b3b23e57c1778fcf857c71ceecd9bd264513f"},
+    {file = "scipy-1.18.1-cp314-cp314t-macosx_14_0_arm64.whl", hash = "sha256:0e82073ecc7acc6436fac4b31674109c7e1d3e596789767eda01258a8c9e8123"},
+    {file = "scipy-1.18.1-cp314-cp314t-macosx_14_0_x86_64.whl", hash = "sha256:8bcf3c1ba5d6456e2effd30fcbd3459b044d683fcdac79a2e6830f0bdf7de487"},
+    {file = "scipy-1.18.1-cp314-cp314t-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:cfbf154f2ba187f2ed6cce2639efff7d105f1140573642c0161615b6d91d6a87"},
+    {file = "scipy-1.18.1-cp314-cp314t-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:a1d33a7836f7ddc1993427966a0823468ec41bcbdb1a9f9942d1d7e57f803ba3"},
+    {file = "scipy-1.18.1-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:7f4b8bc363b6d65ee2152bec57568e3c52639bb34c46057b09857a307ed5e21d"},
+    {file = "scipy-1.18.1-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:11c423f1049c5755ad4409af52a9ada1cff96fe9b50795d4af3619f292901239"},
+    {file = "scipy-1.18.1-cp314-cp314t-win_amd64.whl", hash = "sha256:c24acac1e18912761c4700239bbc1fd32f615af690f1584d49b35859be51324d"},
+    {file = "scipy-1.18.1-cp314-cp314t-win_arm64.whl", hash = "sha256:9f2897bf7737392ad0d5213ea7b6add72a4edf5679b3153106aeb88b6507b3b9"},
+    {file = "scipy-1.18.1-cp315-cp315-macosx_10_15_x86_64.whl", hash = "sha256:eb0dfcf4e28a99c12c999744a2ff67c9b06200e20401c7c88186e33552a46331"},
+    {file = "scipy-1.18.1-cp315-cp315-macosx_12_0_arm64.whl", hash = "sha256:30f464bee641fa8e282577c7dce027308403213c6ca8270bba73285c91024bc5"},
+    {file = "scipy-1.18.1-cp315-cp315-macosx_14_0_arm64.whl", hash = "sha256:1bca3b943fc2567ea49cd02c99abde49da4d5178ec46f624bd8255cda8755beb"},
+    {file = "scipy-1.18.1-cp315-cp315-macosx_14_0_x86_64.whl", hash = "sha256:c9d18a33309122074ea483dd92dd444189166b8b2ec429fe9ed5ac73c7a0aa23"},
+    {file = "scipy-1.18.1-cp315-cp315-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:82f201b4c878551d48558337aab270d3c6cca5507b8737c8d8a608d234cccde0"},
+    {file = "scipy-1.18.1-cp315-cp315-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:0ac49ea97594532dd44b7136094d35f5440fa06e6d9c6384a74c01764df388c5"},
+    {file = "scipy-1.18.1-cp315-cp315-musllinux_1_2_aarch64.whl", hash = "sha256:ceb30a00ce7c92d459819443d29ca486d882b83fb6738bdcbb2a1cce94ac5daa"},
+    {file = "scipy-1.18.1-cp315-cp315-musllinux_1_2_x86_64.whl", hash = "sha256:f29633129f9fa7e88a3f0fca835de2d030bfc9643f7799e1a0c46cee24d38fc7"},
+    {file = "scipy-1.18.1-cp315-cp315-win_amd64.whl", hash = "sha256:92c14f5bdbfb6216315ce33e78080474082de8b3830122ba97809bfbe65f75c0"},
+    {file = "scipy-1.18.1-cp315-cp315-win_arm64.whl", hash = "sha256:e402cf31eb68f453dbb2d36fc6d722b33f24a55d68b2ae1d92fa6305ca71c298"},
+    {file = "scipy-1.18.1-cp315-cp315t-macosx_10_15_x86_64.whl", hash = "sha256:2a0b02f9fc46f8520330c23d45e6560db7e3a0d927232139427637f98943e11d"},
+    {file = "scipy-1.18.1-cp315-cp315t-macosx_12_0_arm64.whl", hash = "sha256:1d73131e358976663dd969e1fb4ed1404b815cd977eaaedc3b3a133ba2d81c35"},
+    {file = "scipy-1.18.1-cp315-cp315t-macosx_14_0_arm64.whl", hash = "sha256:bff0b729edd992766136b34e39cc76bc2fad905aa58897ee72a9cd000a6d8443"},
+    {file = "scipy-1.18.1-cp315-cp315t-macosx_14_0_x86_64.whl", hash = "sha256:10ac20c69d880f77f375db44c22e3e6a644f9fefa291d4cd2fb9790a89fc99fd"},
+    {file = "scipy-1.18.1-cp315-cp315t-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:33a834464fdabc0f26a45508df31b3cc5d028e04dbf6c5ed398541418e0a12fe"},
+    {file = "scipy-1.18.1-cp315-cp315t-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:49023963c193dacee096301452f223ee24d86ec5807f8df93c0f7221d119e305"},
+    {file = "scipy-1.18.1-cp315-cp315t-musllinux_1_2_aarch64.whl", hash = "sha256:d84a09d0dad90ba6525d8ac1c2334b33e64bf3ccfe9e841f02feb867a22681e4"},
+    {file = "scipy-1.18.1-cp315-cp315t-musllinux_1_2_x86_64.whl", hash = "sha256:179ce34a8d0fe273d8883ba59e17e052247d08973dfcb743ca52bb1cce2d60b0"},
+    {file = "scipy-1.18.1-cp315-cp315t-win_amd64.whl", hash = "sha256:5632e3ae3d09197c446310cd5187de63e28448ce22f0f67b2b93d97503c0c230"},
+    {file = "scipy-1.18.1-cp315-cp315t-win_arm64.whl", hash = "sha256:eda632a7981f69730d6281f451db9c1c370993a2c0d7ddb43e2a809a2862b83a"},
+    {file = "scipy-1.18.1.tar.gz", hash = "sha256:52c4b7422442aba924d03ad4019852b08a92e64ea187b933135687bfe2747307"},
+]
+
+[package.dependencies]
+numpy = ">=2.0.0,<2.8"
+
+[package.extras]
+dev = ["click (<8.3.0)", "cython-lint (>=0.12.2)", "mypy (==1.19.1)", "pycodestyle", "pyrefly (==0.63.0)", "ruff (>=0.12.0)", "spin", "types-psutil", "typing_extensions"]
+doc = ["intersphinx_registry", "jupyterlite-pyodide-kernel", "jupyterlite-sphinx (>=0.19.1)", "jupytext", "linkify-it-py", "matplotlib (>=3.5)", "myst-nb (>=1.2.0)", "numpydoc", "pooch", "pydata-sphinx-theme (>=0.15.2)", "sphinx (>=5.0.0,<8.2.0)", "sphinx-copybutton", "sphinx-design (>=0.4.0)", "tabulate"]
+test = ["Cython", "array-api-strict (>=2.3.1)", "asv", "gmpy2", "hypothesis (>=6.30)", "meson", "mpmath", "ninja ; sys_platform != \"emscripten\"", "pooch", "pytest (>=8.0.0)", "pytest-cov", "pytest-timeout", "pytest-xdist", "scikit-umfpack", "scipy-doctest (>=2.0.0)", "threadpoolctl"]
+
+[[package]]
+name = "setuptools"
+version = "84.0.0"
+description = "Most extensible Python build backend with support for C/C++ extension modules"
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "setuptools-84.0.0-py3-none-any.whl", hash = "sha256:51a52592b3b99e102b609654876bd65f19f999935166d1352678931132b0c670"},
+    {file = "setuptools-84.0.0.tar.gz", hash = "sha256:f4695c21257f0d9b537ec2692c941d02ee143b7cc1276941349a546573b2ef73"},
+]
+
+[package.extras]
+check = ["pytest-checkdocs (>=2.14)", "pytest-ruff (>=0.2.1) ; sys_platform != \"cygwin\"", "ruff (>=0.13.0) ; sys_platform != \"cygwin\""]
+core = ["importlib_metadata (>=6) ; python_version < \"3.10\"", "jaraco.functools (>=4)", "jaraco.text (>=3.7)", "more_itertools", "more_itertools (>=8.8)", "packaging (>=24.2)", "tomli (>=2.0.1) ; python_version < \"3.11\"", "wheel (>=0.43.0)"]
+cover = ["pytest-cov"]
+doc = ["furo", "jaraco.packaging (>=9.3)", "jaraco.tidelift (>=1.4)", "pygments-github-lexers (==0.0.5)", "pyproject-hooks (!=1.1)", "rst.linker (>=1.9)", "sphinx (>=3.5)", "sphinx-favicon", "sphinx-inline-tabs", "sphinx-lint", "sphinx-notfound-page (>=1,<2)", "sphinx-reredirects", "sphinxcontrib-towncrier", "towncrier (<24.7)"]
+enabler = ["pytest-enabler (>=3.4)"]
+test = ["build[virtualenv] (>=1.0.3)", "filelock (>=3.4.0)", "ini2toml[lite] (>=0.14)", "jaraco.develop (>=7.21) ; python_version >= \"3.9\" and sys_platform != \"cygwin\"", "jaraco.envs (>=2.2)", "jaraco.path (>=3.7.2)", "jaraco.test (>=5.5)", "packaging (>=24.2)", "pip (>=19.1)", "pyproject-hooks (!=1.1)", "pytest (>=6,!=8.1.*)", "pytest-home (>=0.5)", "pytest-perf ; sys_platform != \"cygwin\"", "pytest-subprocess", "pytest-timeout", "pytest-xdist (>=3)", "tomli-w (>=1.0.0)", "virtualenv (>=13.0.0)", "wheel (>=0.44.0)"]
+type = ["importlib_metadata (>=7.0.2) ; python_version < \"3.10\"", "jaraco.develop (>=7.21) ; sys_platform != \"cygwin\"", "mypy (==1.18.*)", "pytest-mypy (>=1.0.1) ; platform_python_implementation != \"PyPy\""]
+
+[[package]]
+name = "shellingham"
+version = "1.5.4"
+description = "Tool to Detect Surrounding Shell"
+optional = false
+python-versions = ">=3.7"
+groups = ["main"]
+files = [
+    {file = "shellingham-1.5.4-py2.py3-none-any.whl", hash = "sha256:7ecfff8f2fd72616f7481040475a65b2bf8af90a56c89140852d1120324e8686"},
+    {file = "shellingham-1.5.4.tar.gz", hash = "sha256:8dbca0739d487e5bd35ab3ca4b36e11c4078f3a234bfce294b0a0291363404de"},
+]
+
+[[package]]
+name = "six"
+version = "1.17.0"
+description = "Python 2 and 3 compatibility utilities"
+optional = false
+python-versions = "!=3.0.*,!=3.1.*,!=3.2.*,>=2.7"
+groups = ["main"]
+files = [
+    {file = "six-1.17.0-py2.py3-none-any.whl", hash = "sha256:4721f391ed90541fddacab5acf947aa0d3dc7d27b2e1e8eda2be8970586c3274"},
+    {file = "six-1.17.0.tar.gz", hash = "sha256:ff70335d468e7eb6ec65b95b99d3a2836546063f63acc5171de367e834932a81"},
+]
+
+[[package]]
+name = "sympy"
+version = "1.14.0"
+description = "Computer algebra system (CAS) in Python"
+optional = false
+python-versions = ">=3.9"
+groups = ["main"]
+files = [
+    {file = "sympy-1.14.0-py3-none-any.whl", hash = "sha256:e091cc3e99d2141a0ba2847328f5479b05d94a6635cb96148ccb3f34671bd8f5"},
+    {file = "sympy-1.14.0.tar.gz", hash = "sha256:d3d3fe8df1e5a0b42f0e7bdf50541697dbe7d23746e894990c030e2b05e72517"},
+]
+
+[package.dependencies]
+mpmath = ">=1.1.0,<1.4"
+
+[package.extras]
+dev = ["hypothesis (>=6.70.0)", "pytest (>=7.1.0)"]
+
+[[package]]
+name = "threadpoolctl"
+version = "3.7.0"
+description = "threadpoolctl"
+optional = false
+python-versions = ">=3.9"
+groups = ["main"]
+files = [
+    {file = "threadpoolctl-3.7.0-py3-none-any.whl", hash = "sha256:cd8b60b5641b45c67bbf73c64c843235fc2d8a480c87389f52f5dbee893b86be"},
+    {file = "threadpoolctl-3.7.0.tar.gz", hash = "sha256:61348cfb77d53b9242e0017029244b559b810c142ced65b4e21eeca1843959a7"},
+]
+
+[[package]]
+name = "tokenizers"
+version = "0.23.3"
+description = ""
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "tokenizers-0.23.3-cp310-abi3-macosx_10_12_x86_64.whl", hash = "sha256:9d2b5c97daf61688c2ad1803ca851800feaba50fb68d5821779e9ea5880d968c"},
+    {file = "tokenizers-0.23.3-cp310-abi3-macosx_11_0_arm64.whl", hash = "sha256:68649e97d5b43c44c031d8d848874a6eecae8f8fe40ea989aa777a5a83aca716"},
+    {file = "tokenizers-0.23.3-cp310-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:ec82e80e65a862275b97c3d90b7a523df8d9519ee48aeb4e9625b2cc909274e0"},
+    {file = "tokenizers-0.23.3-cp310-abi3-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:c64a0713180ff16829d4e7f39a658b77ea11443af4e1aa46523692943c9b1414"},
+    {file = "tokenizers-0.23.3-cp310-abi3-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:ddedfd4b3b4be6be24ff6ca645c4a37fddfd305f6f3e354c54cf10b715c48215"},
+    {file = "tokenizers-0.23.3-cp310-abi3-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:2a89614730d7b80940a5d2ed9320e1ec8add5a745c6151d8d05071b7215505b6"},
+    {file = "tokenizers-0.23.3-cp310-abi3-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:e88646b8580c5ad7f4361477f1298e9cc01771a1ee9aecfe32c47b8ff614cc38"},
+    {file = "tokenizers-0.23.3-cp310-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:376851d22bcf9d650a5c3090bb83e6cf9e895fbf0595369fa4cd43c1f69b5f87"},
+    {file = "tokenizers-0.23.3-cp310-abi3-manylinux_2_31_riscv64.whl", hash = "sha256:bf501c40b72d2d5c8623620210430e9cac1ce47a46e45b34107b70a1557d46b0"},
+    {file = "tokenizers-0.23.3-cp310-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:114e2b55ed177179d59f4ab98200a4471e11e78f9e4b5a922d146740f96fcf52"},
+    {file = "tokenizers-0.23.3-cp310-abi3-musllinux_1_2_armv7l.whl", hash = "sha256:d3407fb7b9c4d75dd68850ffd7180bc0a5d2dbaf0762d888e612f31fec3f9c6b"},
+    {file = "tokenizers-0.23.3-cp310-abi3-musllinux_1_2_i686.whl", hash = "sha256:84513ef0aeb8bf8f4ea11a2e8a7ac163ec5288aa115e649a59b470ac5c3107df"},
+    {file = "tokenizers-0.23.3-cp310-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:e05ab7baf7f47b406a95fea6f3b0a484b2ddcd9e1d14b68844c457eb755085a3"},
+    {file = "tokenizers-0.23.3-cp310-abi3-win32.whl", hash = "sha256:1ebf28794e7e4954e20a7f70fbea410b2d1f0418f7dbbca97ca384fcfef38c25"},
+    {file = "tokenizers-0.23.3-cp310-abi3-win_amd64.whl", hash = "sha256:1f0823bb00c5fdc98e487354d54dd55a03848d61a1a0bf29a68c77f24f3b26c3"},
+    {file = "tokenizers-0.23.3-cp310-abi3-win_arm64.whl", hash = "sha256:7e48734d2de9260d86f03ab056d2cfeeff3869f61dbd49aaa15a2793b5f3458b"},
+    {file = "tokenizers-0.23.3-cp314-cp314t-macosx_10_12_x86_64.whl", hash = "sha256:efa3d7318406b4d115dce61ad5061953f1f44b128e79c020ce4615d763e23b6e"},
+    {file = "tokenizers-0.23.3-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:a4fbb3662f9f59d199d61338e54b4bcc11d07ebbb1aeb3540dacb2be9c521cb7"},
+    {file = "tokenizers-0.23.3-cp314-cp314t-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:de536665495cb4b409d25bade41963f801aff4225c19a6b804b048f7d14e34c7"},
+    {file = "tokenizers-0.23.3-cp314-cp314t-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:5cc24bb457dd4a8af89c8fcb40074d570129ec473df2a866c276ee55db4749d7"},
+    {file = "tokenizers-0.23.3-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:acd5c57b4bd3e56e246e2731a3a3a6825a7a7d89b7e3b761ba80bc521710f04b"},
+    {file = "tokenizers-0.23.3-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:82eb480f6f1c21cea3349dec32cf1a6384c6c1e775f00f83b0d51197bc013687"},
+    {file = "tokenizers-0.23.3-cp314-cp314t-win_amd64.whl", hash = "sha256:1554a6eed34d9d6a78d23360f4e06df8dffab1ae08c7e8488e0b3e3b36cc266f"},
+    {file = "tokenizers-0.23.3.tar.gz", hash = "sha256:cded33237c77caeef62944d32aa9a7ef42bdce2b3497e18d137e072a8c4be438"},
+]
+
+[package.dependencies]
+huggingface-hub = ">=0.16.4,<3.0"
+
+[package.extras]
+dev = ["tokenizers[testing]"]
+docs = ["setuptools-rust", "sphinx", "sphinx-rtd-theme"]
+testing = ["datasets", "numpy", "pytest", "pytest-asyncio", "requests", "ruff", "ty"]
+
+[[package]]
+name = "torch"
+version = "2.8.0"
+description = "Tensors and Dynamic neural networks in Python with strong GPU acceleration"
+optional = false
+python-versions = ">=3.9.0"
+groups = ["main"]
+markers = "sys_platform != \"linux\""
+files = [
+    {file = "torch-2.8.0-cp310-cp310-manylinux_2_28_aarch64.whl", hash = "sha256:0be92c08b44009d4131d1ff7a8060d10bafdb7ddcb7359ef8d8c5169007ea905"},
+    {file = "torch-2.8.0-cp310-cp310-manylinux_2_28_x86_64.whl", hash = "sha256:89aa9ee820bb39d4d72b794345cccef106b574508dd17dbec457949678c76011"},
+    {file = "torch-2.8.0-cp310-cp310-win_amd64.whl", hash = "sha256:e8e5bf982e87e2b59d932769938b698858c64cc53753894be25629bdf5cf2f46"},
+    {file = "torch-2.8.0-cp310-none-macosx_11_0_arm64.whl", hash = "sha256:a3f16a58a9a800f589b26d47ee15aca3acf065546137fc2af039876135f4c760"},
+    {file = "torch-2.8.0-cp311-cp311-manylinux_2_28_aarch64.whl", hash = "sha256:220a06fd7af8b653c35d359dfe1aaf32f65aa85befa342629f716acb134b9710"},
+    {file = "torch-2.8.0-cp311-cp311-manylinux_2_28_x86_64.whl", hash = "sha256:c12fa219f51a933d5f80eeb3a7a5d0cbe9168c0a14bbb4055f1979431660879b"},
+    {file = "torch-2.8.0-cp311-cp311-win_amd64.whl", hash = "sha256:8c7ef765e27551b2fbfc0f41bcf270e1292d9bf79f8e0724848b1682be6e80aa"},
+    {file = "torch-2.8.0-cp311-none-macosx_11_0_arm64.whl", hash = "sha256:5ae0524688fb6707c57a530c2325e13bb0090b745ba7b4a2cd6a3ce262572916"},
+    {file = "torch-2.8.0-cp312-cp312-manylinux_2_28_aarch64.whl", hash = "sha256:e2fab4153768d433f8ed9279c8133a114a034a61e77a3a104dcdf54388838705"},
+    {file = "torch-2.8.0-cp312-cp312-manylinux_2_28_x86_64.whl", hash = "sha256:b2aca0939fb7e4d842561febbd4ffda67a8e958ff725c1c27e244e85e982173c"},
+    {file = "torch-2.8.0-cp312-cp312-win_amd64.whl", hash = "sha256:2f4ac52f0130275d7517b03a33d2493bab3693c83dcfadf4f81688ea82147d2e"},
+    {file = "torch-2.8.0-cp312-none-macosx_11_0_arm64.whl", hash = "sha256:619c2869db3ada2c0105487ba21b5008defcc472d23f8b80ed91ac4a380283b0"},
+    {file = "torch-2.8.0-cp313-cp313-manylinux_2_28_aarch64.whl", hash = "sha256:2b2f96814e0345f5a5aed9bf9734efa913678ed19caf6dc2cddb7930672d6128"},
+    {file = "torch-2.8.0-cp313-cp313-manylinux_2_28_x86_64.whl", hash = "sha256:65616ca8ec6f43245e1f5f296603e33923f4c30f93d65e103d9e50c25b35150b"},
+    {file = "torch-2.8.0-cp313-cp313-win_amd64.whl", hash = "sha256:659df54119ae03e83a800addc125856effda88b016dfc54d9f65215c3975be16"},
+    {file = "torch-2.8.0-cp313-cp313t-macosx_14_0_arm64.whl", hash = "sha256:1a62a1ec4b0498930e2543535cf70b1bef8c777713de7ceb84cd79115f553767"},
+    {file = "torch-2.8.0-cp313-cp313t-manylinux_2_28_aarch64.whl", hash = "sha256:83c13411a26fac3d101fe8035a6b0476ae606deb8688e904e796a3534c197def"},
+    {file = "torch-2.8.0-cp313-cp313t-manylinux_2_28_x86_64.whl", hash = "sha256:8f0a9d617a66509ded240add3754e462430a6c1fc5589f86c17b433dd808f97a"},
+    {file = "torch-2.8.0-cp313-cp313t-win_amd64.whl", hash = "sha256:a7242b86f42be98ac674b88a4988643b9bc6145437ec8f048fea23f72feb5eca"},
+    {file = "torch-2.8.0-cp313-none-macosx_11_0_arm64.whl", hash = "sha256:7b677e17f5a3e69fdef7eb3b9da72622f8d322692930297e4ccb52fefc6c8211"},
+    {file = "torch-2.8.0-cp39-cp39-manylinux_2_28_aarch64.whl", hash = "sha256:da6afa31c13b669d4ba49d8a2169f0db2c3ec6bec4af898aa714f401d4c38904"},
+    {file = "torch-2.8.0-cp39-cp39-manylinux_2_28_x86_64.whl", hash = "sha256:06fcee8000e5c62a9f3e52a688b9c5abb7c6228d0e56e3452983416025c41381"},
+    {file = "torch-2.8.0-cp39-cp39-win_amd64.whl", hash = "sha256:5128fe752a355d9308e56af1ad28b15266fe2da5948660fad44de9e3a9e36e8c"},
+    {file = "torch-2.8.0-cp39-none-macosx_11_0_arm64.whl", hash = "sha256:e9f071f5b52a9f6970dc8a919694b27a91ae9dc08898b2b988abbef5eddfd1ae"},
+]
+
+[package.dependencies]
+filelock = "*"
+fsspec = "*"
+jinja2 = "*"
+networkx = "*"
+nvidia-cublas-cu12 = {version = "12.8.4.1", markers = "platform_system == \"Linux\" and platform_machine == \"x86_64\""}
+nvidia-cuda-cupti-cu12 = {version = "12.8.90", markers = "platform_system == \"Linux\" and platform_machine == \"x86_64\""}
+nvidia-cuda-nvrtc-cu12 = {version = "12.8.93", markers = "platform_system == \"Linux\" and platform_machine == \"x86_64\""}
+nvidia-cuda-runtime-cu12 = {version = "12.8.90", markers = "platform_system == \"Linux\" and platform_machine == \"x86_64\""}
+nvidia-cudnn-cu12 = {version = "9.10.2.21", markers = "platform_system == \"Linux\" and platform_machine == \"x86_64\""}
+nvidia-cufft-cu12 = {version = "11.3.3.83", markers = "platform_system == \"Linux\" and platform_machine == \"x86_64\""}
+nvidia-cufile-cu12 = {version = "1.13.1.3", markers = "platform_system == \"Linux\" and platform_machine == \"x86_64\""}
+nvidia-curand-cu12 = {version = "10.3.9.90", markers = "platform_system == \"Linux\" and platform_machine == \"x86_64\""}
+nvidia-cusolver-cu12 = {version = "11.7.3.90", markers = "platform_system == \"Linux\" and platform_machine == \"x86_64\""}
+nvidia-cusparse-cu12 = {version = "12.5.8.93", markers = "platform_system == \"Linux\" and platform_machine == \"x86_64\""}
+nvidia-cusparselt-cu12 = {version = "0.7.1", markers = "platform_system == \"Linux\" and platform_machine == \"x86_64\""}
+nvidia-nccl-cu12 = {version = "2.27.3", markers = "platform_system == \"Linux\" and platform_machine == \"x86_64\""}
+nvidia-nvjitlink-cu12 = {version = "12.8.93", markers = "platform_system == \"Linux\" and platform_machine == \"x86_64\""}
+nvidia-nvtx-cu12 = {version = "12.8.90", markers = "platform_system == \"Linux\" and platform_machine == \"x86_64\""}
+setuptools = {version = "*", markers = "python_version >= \"3.12\""}
+sympy = ">=1.13.3"
+triton = {version = "3.4.0", markers = "platform_system == \"Linux\" and platform_machine == \"x86_64\""}
+typing-extensions = ">=4.10.0"
+
+[package.extras]
+opt-einsum = ["opt-einsum (>=3.3)"]
+optree = ["optree (>=0.13.0)"]
+pyyaml = ["pyyaml"]
+
+[[package]]
+name = "torch"
+version = "2.8.0+cpu"
+description = "Tensors and Dynamic neural networks in Python with strong GPU acceleration"
+optional = false
+python-versions = ">=3.9.0"
+groups = ["main"]
+markers = "sys_platform == \"linux\""
+files = [
+    {file = "torch-2.8.0+cpu-cp310-cp310-linux_s390x.whl", hash = "sha256:5d255d259fbc65439b671580e40fdb8faea4644761b64fed90d6904ffe71bbc1"},
+    {file = "torch-2.8.0+cpu-cp310-cp310-manylinux_2_28_aarch64.whl", hash = "sha256:b2149858b8340aeeb1f3056e0bff5b82b96e43b596fe49a9dba3184522261213"},
+    {file = "torch-2.8.0+cpu-cp310-cp310-manylinux_2_28_x86_64.whl", hash = "sha256:16d75fa4e96ea28a785dfd66083ca55eb1058b6d6c5413f01656ca965ee2077e"},
+    {file = "torch-2.8.0+cpu-cp310-cp310-win_amd64.whl", hash = "sha256:7cc4af6ba954f36c2163eab98cf113c137fc25aa8bbf1b06ef155968627beed2"},
+    {file = "torch-2.8.0+cpu-cp311-cp311-linux_s390x.whl", hash = "sha256:2bfc013dd6efdc8f8223a0241d3529af9f315dffefb53ffa3bf14d3f10127da6"},
+    {file = "torch-2.8.0+cpu-cp311-cp311-manylinux_2_28_aarch64.whl", hash = "sha256:680129efdeeec3db5da3f88ee5d28c1b1e103b774aef40f9d638e2cce8f8d8d8"},
+    {file = "torch-2.8.0+cpu-cp311-cp311-manylinux_2_28_x86_64.whl", hash = "sha256:cb06175284673a581dd91fb1965662ae4ecaba6e5c357aa0ea7bb8b84b6b7eeb"},
+    {file = "torch-2.8.0+cpu-cp311-cp311-win_amd64.whl", hash = "sha256:7631ef49fbd38d382909525b83696dc12a55d68492ade4ace3883c62b9fc140f"},
+    {file = "torch-2.8.0+cpu-cp311-cp311-win_arm64.whl", hash = "sha256:41e6fc5ec0914fcdce44ccf338b1d19a441b55cafdd741fd0bf1af3f9e4cfd14"},
+    {file = "torch-2.8.0+cpu-cp312-cp312-linux_s390x.whl", hash = "sha256:0e34e276722ab7dd0dffa9e12fe2135a9b34a0e300c456ed7ad6430229404eb5"},
+    {file = "torch-2.8.0+cpu-cp312-cp312-manylinux_2_28_aarch64.whl", hash = "sha256:610f600c102386e581327d5efc18c0d6edecb9820b4140d26163354a99cd800d"},
+    {file = "torch-2.8.0+cpu-cp312-cp312-manylinux_2_28_x86_64.whl", hash = "sha256:cb9a8ba8137ab24e36bf1742cb79a1294bd374db570f09fc15a5e1318160db4e"},
+    {file = "torch-2.8.0+cpu-cp312-cp312-win_amd64.whl", hash = "sha256:2be20b2c05a0cce10430cc25f32b689259640d273232b2de357c35729132256d"},
+    {file = "torch-2.8.0+cpu-cp312-cp312-win_arm64.whl", hash = "sha256:99fc421a5d234580e45957a7b02effbf3e1c884a5dd077afc85352c77bf41434"},
+    {file = "torch-2.8.0+cpu-cp313-cp313-linux_s390x.whl", hash = "sha256:8b5882276633cf91fe3d2d7246c743b94d44a7e660b27f1308007fdb1bb89f7d"},
+    {file = "torch-2.8.0+cpu-cp313-cp313-manylinux_2_28_aarch64.whl", hash = "sha256:a5064b5e23772c8d164068cc7c12e01a75faf7b948ecd95a0d4007d7487e5f25"},
+    {file = "torch-2.8.0+cpu-cp313-cp313-manylinux_2_28_x86_64.whl", hash = "sha256:8f81dedb4c6076ec325acc3b47525f9c550e5284a18eae1d9061c543f7b6e7de"},
+    {file = "torch-2.8.0+cpu-cp313-cp313-win_amd64.whl", hash = "sha256:e1ee1b2346ade3ea90306dfbec7e8ff17bc220d344109d189ae09078333b0856"},
+    {file = "torch-2.8.0+cpu-cp313-cp313-win_arm64.whl", hash = "sha256:64c187345509f2b1bb334feed4666e2c781ca381874bde589182f81247e61f88"},
+    {file = "torch-2.8.0+cpu-cp313-cp313t-manylinux_2_28_aarch64.whl", hash = "sha256:af81283ac671f434b1b25c95ba295f270e72db1fad48831eb5e4748ff9840041"},
+    {file = "torch-2.8.0+cpu-cp313-cp313t-manylinux_2_28_x86_64.whl", hash = "sha256:a9dbb6f64f63258bc811e2c0c99640a81e5af93c531ad96e95c5ec777ea46dab"},
+    {file = "torch-2.8.0+cpu-cp313-cp313t-win_amd64.whl", hash = "sha256:6d93a7165419bc4b2b907e859ccab0dea5deeab261448ae9a5ec5431f14c0e64"},
+    {file = "torch-2.8.0+cpu-cp39-cp39-linux_s390x.whl", hash = "sha256:5239ef35402000844b676a9b79ed76d5ae6b028a6762bbdfebdf8421a0f4d2aa"},
+    {file = "torch-2.8.0+cpu-cp39-cp39-manylinux_2_28_aarch64.whl", hash = "sha256:eac8b7ef5c7ca106daec5e829dfa8ca56ca47601db13b402d2608861ad3ab926"},
+    {file = "torch-2.8.0+cpu-cp39-cp39-manylinux_2_28_x86_64.whl", hash = "sha256:bda4f93d64dcd9ae5d51844bbccc6fcb7d603522bcc95d256b5fe3bdb9dccca3"},
+    {file = "torch-2.8.0+cpu-cp39-cp39-win_amd64.whl", hash = "sha256:e3c3fce24ebaac954b837d1498e36d484ad0d93e2a1ed5b6b0c55a02ea748fab"},
+]
+
+[package.dependencies]
+filelock = "*"
+fsspec = "*"
+jinja2 = "*"
+networkx = "*"
+setuptools = {version = "*", markers = "python_version >= \"3.12\""}
+sympy = ">=1.13.3"
+typing-extensions = ">=4.10.0"
+
+[package.extras]
+opt-einsum = ["opt-einsum (>=3.3)"]
+optree = ["optree (>=0.13.0)"]
+pyyaml = ["pyyaml"]
+
+[package.source]
+type = "legacy"
+url = "https://download.pytorch.org/whl/cpu"
+reference = "pytorch-cpu"
+
+[[package]]
+name = "tqdm"
+version = "4.70.1"
+description = "Fast, Extensible Progress Meter"
+optional = false
+python-versions = ">=3.8"
+groups = ["main"]
+files = [
+    {file = "tqdm-4.70.1-py3-none-any.whl", hash = "sha256:c293e525e6fef9c20e8728fd4612df02a0aa31bb5fe91ecd93e123b1b7bffa73"},
+    {file = "tqdm-4.70.1.tar.gz", hash = "sha256:cefd0eca11b2a37a3aee776544d4f4ae913f02688135b5556b8788dfa474afc4"},
+]
+
+[package.dependencies]
+colorama = {version = "*", markers = "platform_system == \"Windows\""}
+
+[package.extras]
+discord = ["envwrap", "requests"]
+notebook = ["ipywidgets (>=6)"]
+slack = ["envwrap", "slack-sdk"]
+telegram = ["envwrap", "requests"]
+
+[[package]]
+name = "transformers"
+version = "5.20.2"
+description = "Transformers: the model-definition framework for state-of-the-art machine learning models in text, vision, audio, and multimodal models, for both inference and training."
+optional = false
+python-versions = ">=3.10.0"
+groups = ["main"]
+files = [
+    {file = "transformers-5.20.2-py3-none-any.whl", hash = "sha256:7ae75dbc1fb3922aff1b6c97b9e5479d6cf26fd86a8866bfb71324190479f807"},
+    {file = "transformers-5.20.2.tar.gz", hash = "sha256:080871e061546d720be57f5846fce426f46b44d61574907cc7404c2cba49d180"},
+]
+
+[package.dependencies]
+huggingface-hub = ">=1.31.0,<3.0"
+numpy = ">=1.17"
+packaging = ">=20.0"
+pyyaml = ">=5.1"
+regex = ">=2025.10.22"
+safetensors = ">=0.8.0"
+tokenizers = ">=0.23.1,<0.24.0"
+tqdm = ">=4.60"
+typer = "*"
+
+[package.extras]
+accelerate = ["accelerate (>=1.1.0)"]
+all = ["Pillow (>=10.0.1,<=15.0)", "accelerate (>=1.1.0)", "av", "blobfile", "jinja2 (>=3.1.0)", "kenlm", "kernels (>=0.17.2,<0.18)", "librosa", "mistral-common[image] (>=1.11.7)", "num2words", "phonemizer", "protobuf", "pyctcdecode (>=0.4.0)", "sentencepiece (>=0.1.91,!=0.1.92)", "tiktoken", "timm (>=1.0.23)", "torch (>=2.6)", "torchaudio", "torchvision"]
+audio = ["kenlm", "librosa", "phonemizer", "pyctcdecode (>=0.4.0)", "torchaudio"]
+benchmark = ["optimum-benchmark (>=0.3.0)"]
+chat-template = ["jinja2 (>=3.1.0)"]
+codecarbon = ["codecarbon (>=2.8.1)"]
+deepspeed = ["accelerate (>=1.1.0)", "deepspeed (>=0.9.3)"]
+deepspeed-testing = ["GitPython (<3.1.19)", "accelerate (>=1.1.0)", "accelerate (>=1.1.0)", "beautifulsoup4", "datasets (>=2.15.0)", "datasets (>=2.15.0)", "deepspeed (>=0.9.3)", "diffusers (==0.35.2)", "faiss-cpu", "fastapi", "filelock", "hf-doc-builder", "libcst", "mistral-common[image] (>=1.11.7)", "nltk (<=3.8.1)", "openai (>=1.98.0)", "optuna", "parameterized (>=0.9)", "protobuf", "protobuf", "psutil", "pydantic (>=2)", "pytest (>=7.2.0,<9.0.0)", "pytest-asyncio (>=1.2.0)", "pytest-env", "pytest-order", "pytest-random-order", "pytest-rerunfailures (<16.0)", "pytest-rich", "pytest-timeout", "pytest-xdist", "rich", "rich", "rjieba", "ruff (==0.14.10)", "sacremoses", "sentencepiece (>=0.1.91,!=0.1.92)", "sentencepiece (>=0.1.91,!=0.1.92)", "starlette", "tensorboard", "timeout-decorator", "tomli", "torch (>=2.6)", "transformers-mlinter (==0.1.5)", "ty (==0.0.20)", "urllib3 (<2.0.0)", "uvicorn"]
+dev = ["GitPython (<3.1.19)", "Pillow (>=10.0.1,<=15.0)", "accelerate (>=1.1.0)", "accelerate (>=1.1.0)", "av", "beautifulsoup4", "blobfile", "datasets (>=2.15.0)", "datasets (>=2.15.0)", "diffusers (==0.35.2)", "faiss-cpu", "fastapi", "filelock", "fugashi (>=1.0)", "hf-doc-builder", "ipadic (>=1.0.0,<2.0)", "jinja2 (>=3.1.0)", "kenlm", "kernels (>=0.17.2,<0.18)", "libcst", "librosa", "mistral-common[image] (>=1.11.7)", "mistral-common[image] (>=1.11.7)", "nltk (<=3.8.1)", "num2words", "openai (>=1.98.0)", "parameterized (>=0.9)", "phonemizer", "protobuf", "protobuf", "psutil", "pyctcdecode (>=0.4.0)", "pydantic (>=2)", "pytest (>=7.2.0,<9.0.0)", "pytest-asyncio (>=1.2.0)", "pytest-env", "pytest-order", "pytest-random-order", "pytest-rerunfailures (<16.0)", "pytest-rich", "pytest-timeout", "pytest-xdist", "rhoknp (>=1.1.0,<1.3.1)", "rich", "rich", "rjieba", "ruff (==0.14.10)", "sacremoses", "scikit-learn", "sentencepiece (>=0.1.91,!=0.1.92)", "sentencepiece (>=0.1.91,!=0.1.92)", "starlette", "sudachidict_core (>=20220729)", "sudachipy (>=0.6.6)", "tensorboard", "tiktoken", "timeout-decorator", "timm (>=1.0.23)", "tomli", "torch (>=2.6)", "torch (>=2.6)", "torchaudio", "torchvision", "transformers-mlinter (==0.1.5)", "ty (==0.0.20)", "unidic (>=1.0.2)", "unidic_lite (>=1.0.7)", "urllib3 (<2.0.0)", "uvicorn"]
+docs = ["hf-doc-builder"]
+integrations = ["codecarbon (>=2.8.1)", "kernels (>=0.17.2,<0.18)", "optuna", "ray[tune] (>=2.7.0)"]
+ja = ["fugashi (>=1.0)", "ipadic (>=1.0.0,<2.0)", "rhoknp (>=1.1.0,<1.3.1)", "sudachidict_core (>=20220729)", "sudachipy (>=0.6.6)", "unidic (>=1.0.2)", "unidic_lite (>=1.0.7)"]
+kernels = ["kernels (>=0.17.2,<0.18)"]
+mistral-common = ["mistral-common[image] (>=1.11.7)"]
+num2words = ["num2words"]
+optuna = ["optuna"]
+quality = ["GitPython (<3.1.19)", "datasets (>=2.15.0)", "libcst", "rich", "ruff (==0.14.10)", "tomli", "transformers-mlinter (==0.1.5)", "ty (==0.0.20)", "urllib3 (<2.0.0)"]
+ray = ["ray[tune] (>=2.7.0)"]
+retrieval = ["datasets (>=2.15.0)", "faiss-cpu"]
+sagemaker = ["sagemaker (>=2.31.0)"]
+sentencepiece = ["protobuf", "sentencepiece (>=0.1.91,!=0.1.92)"]
+serving = ["accelerate (>=1.1.0)", "fastapi", "openai (>=1.98.0)", "pydantic (>=2)", "rich", "starlette", "torch (>=2.6)", "uvicorn"]
+sklearn = ["scikit-learn"]
+testing = ["GitPython (<3.1.19)", "accelerate (>=1.1.0)", "beautifulsoup4", "datasets (>=2.15.0)", "datasets (>=2.15.0)", "diffusers (==0.35.2)", "faiss-cpu", "fastapi", "filelock", "hf-doc-builder", "libcst", "mistral-common[image] (>=1.11.7)", "nltk (<=3.8.1)", "openai (>=1.98.0)", "parameterized (>=0.9)", "protobuf", "psutil", "pydantic (>=2)", "pytest (>=7.2.0,<9.0.0)", "pytest-asyncio (>=1.2.0)", "pytest-env", "pytest-order", "pytest-random-order", "pytest-rerunfailures (<16.0)", "pytest-rich", "pytest-timeout", "pytest-xdist", "rich", "rich", "rjieba", "ruff (==0.14.10)", "sacremoses", "sentencepiece (>=0.1.91,!=0.1.92)", "starlette", "tensorboard", "timeout-decorator", "tomli", "torch (>=2.6)", "transformers-mlinter (==0.1.5)", "ty (==0.0.20)", "urllib3 (<2.0.0)", "uvicorn"]
+tiktoken = ["blobfile", "tiktoken"]
+timm = ["timm (>=1.0.23)"]
+torch = ["accelerate (>=1.1.0)", "torch (>=2.6)"]
+video = ["av"]
+vision = ["Pillow (>=10.0.1,<=15.0)", "torchvision"]
+
+[[package]]
+name = "triton"
+version = "3.4.0"
+description = "A language and compiler for custom Deep Learning operations"
+optional = false
+python-versions = "<3.14,>=3.9"
+groups = ["main"]
+markers = "platform_system == \"Linux\" and platform_machine == \"x86_64\" and sys_platform != \"linux\""
+files = [
+    {file = "triton-3.4.0-cp310-cp310-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:7ff2785de9bc02f500e085420273bb5cc9c9bb767584a4aa28d6e360cec70128"},
+    {file = "triton-3.4.0-cp311-cp311-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:7b70f5e6a41e52e48cfc087436c8a28c17ff98db369447bcaff3b887a3ab4467"},
+    {file = "triton-3.4.0-cp312-cp312-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:31c1d84a5c0ec2c0f8e8a072d7fd150cab84a9c239eaddc6706c081bfae4eb04"},
+    {file = "triton-3.4.0-cp313-cp313-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:00be2964616f4c619193cb0d1b29a99bd4b001d7dc333816073f92cf2a8ccdeb"},
+    {file = "triton-3.4.0-cp313-cp313t-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:7936b18a3499ed62059414d7df563e6c163c5e16c3773678a3ee3d417865035d"},
+    {file = "triton-3.4.0-cp39-cp39-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:98e5c1442eaeabae2e2452ae765801bd53cd4ce873cab0d1bdd59a32ab2d9397"},
+]
+
+[package.dependencies]
+setuptools = ">=40.8.0"
+
+[package.extras]
+build = ["cmake (>=3.20,<4.0)", "lit"]
+tests = ["autopep8", "isort", "llnl-hatchet", "numpy", "pytest", "pytest-forked", "pytest-xdist", "scipy (>=1.7.1)"]
+tutorials = ["matplotlib", "pandas", "tabulate"]
+
+[[package]]
+name = "truststore"
+version = "0.10.5"
+description = "Verify certificates using native system trust stores"
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+markers = "sys_platform != \"emscripten\""
+files = [
+    {file = "truststore-0.10.5-py3-none-any.whl", hash = "sha256:9aaaedaefaf06d8b206278cf8b5012bc897f485a874503501e12d776df78951c"},
+    {file = "truststore-0.10.5.tar.gz", hash = "sha256:30d36967ccaded5cbb38d602c433f53600036c79d502f4533a49b60a03bbefcd"},
+]
+
+[[package]]
+name = "typer"
+version = "0.27.2"
+description = "Typer, build great CLIs. Easy to code. Based on Python type hints."
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "typer-0.27.2-py3-none-any.whl", hash = "sha256:b3a5fc4342d5fc8fda8fc3010b1cf117e9249aab7fae800c2eff62fd3842d97d"},
+    {file = "typer-0.27.2.tar.gz", hash = "sha256:269b7eb9d3c202ca84b4bc9618cb04ebb43d3d4d1e567e4c768607232c05f945"},
+]
+
+[package.dependencies]
+annotated-doc = ">=0.0.2"
+colorama = {version = "*", markers = "platform_system == \"Windows\""}
+rich = ">=13.8.0"
+shellingham = ">=1.3.0"
+
+[[package]]
+name = "typing-extensions"
+version = "4.16.0"
+description = "Backported and Experimental Type Hints for Python 3.9+"
+optional = false
+python-versions = ">=3.9"
+groups = ["main", "dev"]
+files = [
+    {file = "typing_extensions-4.16.0-py3-none-any.whl", hash = "sha256:481caa481374e813c1b176ada14e97f1f67a4539ce9cfeb3f350d78d6370c2e8"},
+    {file = "typing_extensions-4.16.0.tar.gz", hash = "sha256:dc983d19a509c94dba722ee6abd33940f7c05a89e243c47e907eb4db6f1a43e5"},
+]
+
+[[package]]
+name = "typing-inspection"
+version = "0.4.4"
+description = "Runtime typing introspection tools"
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "typing_inspection-0.4.4-py3-none-any.whl", hash = "sha256:65b8397ba37ccbce054456aaccddfc91e6e3083c92824df348d96ca832f3f147"},
+    {file = "typing_inspection-0.4.4.tar.gz", hash = "sha256:547274fa6b0a561ccf549cc9524b999a578e737d015d8709d021f9d0d13bea47"},
+]
+
+[package.dependencies]
+typing-extensions = ">=4.15.0"
+
+[[package]]
+name = "tzdata"
+version = "2026.5"
+description = "Provider of IANA time zone data"
+optional = false
+python-versions = ">=2"
+groups = ["main"]
+markers = "sys_platform == \"win32\" or sys_platform == \"emscripten\""
+files = [
+    {file = "tzdata-2026.5-py2.py3-none-any.whl", hash = "sha256:b683bd1b6659ddcd810ff02ad09ba821d4bf1065072805063eb35c49617905ac"},
+    {file = "tzdata-2026.5.tar.gz", hash = "sha256:8cc73c0a0bfca7dbfa59235d60b2eff82231dee33f53d206db1acd9173cfc0a7"},
+]
+
+[[package]]
+name = "urllib3"
+version = "2.8.0"
+description = "HTTP library with thread-safe connection pooling, file post, and more."
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "urllib3-2.8.0-py3-none-any.whl", hash = "sha256:0cf3cae568d36aa9576b28dfb35f11328f1cb974ca7647d9475ebb86c75ac6e3"},
+    {file = "urllib3-2.8.0.tar.gz", hash = "sha256:63bf2ead4c879426ebf22ef2a781eeb4aa3b4ae798a0435506f8687fd5bb9b63"},
+]
+
+[package.extras]
+brotli = ["brotli (>=1.2.0) ; platform_python_implementation == \"CPython\"", "brotlicffi (>=1.2.0.0) ; platform_python_implementation != \"CPython\""]
+h2 = ["h2 (>=4,<5)"]
+socks = ["pysocks (>=1.5.6,!=1.5.7,<2.0)"]
+zstd = ["backports-zstd (>=1.0.0) ; python_version < \"3.14\""]
+
+[[package]]
+name = "xxhash"
+version = "4.0.1"
+description = "Python binding for xxHash"
+optional = false
+python-versions = ">=3.9"
+groups = ["main"]
+files = [
+    {file = "xxhash-4.0.1-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:3f68fe400ceec235f3e4a4b02a28c2fd2d283584a193223c921dd4c48f1d0754"},
+    {file = "xxhash-4.0.1-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:9b1dddc257279417d93c9e59420d49ef90aece90d7a01996db3aade74b0281b1"},
+    {file = "xxhash-4.0.1-cp310-cp310-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:4972332c079d6aad69c4620a68d015a4ecb33141583f70d642cf9edf6a713763"},
+    {file = "xxhash-4.0.1-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:f1b603d0686c99fa0879f104a74e7db58367634c6e50ba827bee9aa095e23205"},
+    {file = "xxhash-4.0.1-cp310-cp310-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:33fd538191f47071deef6b1f676535e2aa770f1fd150ae4cc75a34c9e930be3d"},
+    {file = "xxhash-4.0.1-cp310-cp310-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:33e270d302c95ec426dfa0f5a4e16bff2ab8d7b8a46faa4746affb05e684ac77"},
+    {file = "xxhash-4.0.1-cp310-cp310-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:3fb1d30d4b6d6e2c4a08e5ac6fffdb2b572d2cfcca15a5509cf4e7a1350f955c"},
+    {file = "xxhash-4.0.1-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:7e27dbed5c4ba033919e4b4ed8dc14e029e91d14a93cd9f920d25277c7df6781"},
+    {file = "xxhash-4.0.1-cp310-cp310-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:43bcf2a871f28f16135545415cab3ec43904d4c80425a64598a9e6cebfb2b5ba"},
+    {file = "xxhash-4.0.1-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:5dc434c946012e6d8a72b10f970ea30755b718251dd7591dbfdabafd3bcb21bc"},
+    {file = "xxhash-4.0.1-cp310-cp310-musllinux_1_2_armv7l.whl", hash = "sha256:0b1082fd0f089ce9098ed77aad8b777b5d156f8ac601c69cab73811822b8ef07"},
+    {file = "xxhash-4.0.1-cp310-cp310-musllinux_1_2_i686.whl", hash = "sha256:dd649663ddeafbfd4734eb8abae921dd5baa1242f20bda54e8bc927369ccded4"},
+    {file = "xxhash-4.0.1-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:97b455de3e8b1b0b1e4594cb61a468992563f03ca264062fbb0a66b393c01d90"},
+    {file = "xxhash-4.0.1-cp310-cp310-musllinux_1_2_riscv64.whl", hash = "sha256:2194bf96d5f3d4e0cb65deba370ec83dda3edfba42155f9384190ed5e51ea5e2"},
+    {file = "xxhash-4.0.1-cp310-cp310-musllinux_1_2_s390x.whl", hash = "sha256:84df5f8da574caadbc0cb1b8866ecc2368cc941f0cd05f677756c802f370dafa"},
+    {file = "xxhash-4.0.1-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:c09ada495567c9c9a8156c5ebcfb93be7fece0755062d738c972dcbecd0d84b5"},
+    {file = "xxhash-4.0.1-cp310-cp310-win32.whl", hash = "sha256:85bdd40cb505a11e0ca04191711266c5fd696ed786ae83849955e457774edc96"},
+    {file = "xxhash-4.0.1-cp310-cp310-win_amd64.whl", hash = "sha256:8ec4777d92fd61a5c8fdeddab894fd65bea301a8092fb5419ec6472aa4d458d7"},
+    {file = "xxhash-4.0.1-cp310-cp310-win_arm64.whl", hash = "sha256:03600a8987849b2bef7be795a60a6052b635c63fa98b718b08ca5ee823691cfc"},
+    {file = "xxhash-4.0.1-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:8b4477edc03091f51f5309406d230851c23cf4822029e3bf40b8df53093fff1c"},
+    {file = "xxhash-4.0.1-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:04f9a24de11a6647666d5302fd73d6a5224ce50ddc965fb0bb44cee736e6bd7c"},
+    {file = "xxhash-4.0.1-cp311-cp311-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:6a8c5ce76b94ba49f3be8a8f2611abc6564210702c72ac9e237ca2bebfd17794"},
+    {file = "xxhash-4.0.1-cp311-cp311-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:b4c8842fb19d78b5e8c2a52baf4c8357658cc56c62bc822b86ce0f942f28e286"},
+    {file = "xxhash-4.0.1-cp311-cp311-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:a43418e1a90b4809a9caf64aeb8b0696e3e1f300a323acc1e6ee2f93ae319fcf"},
+    {file = "xxhash-4.0.1-cp311-cp311-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:b3662719007e059abde7eddacf8517142ba076ddc7b30c807260e57d28c3c191"},
+    {file = "xxhash-4.0.1-cp311-cp311-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:06d7fbd609503c3be5e65cdb6bb2f040d6a98574404e2e1d5c60815c97fff4aa"},
+    {file = "xxhash-4.0.1-cp311-cp311-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:101aa300de6ceef3d9c77569706330d8921fc45dd82bceed2084f1e9f2557a24"},
+    {file = "xxhash-4.0.1-cp311-cp311-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:e4296fcc790876a8b0f297edc83d3b088457b774d8f67b4636807f8a2ec69a79"},
+    {file = "xxhash-4.0.1-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:57d7fa8f23908d173001c21a9e82bfc6ad997d1b6c270fb121812b7ed158891c"},
+    {file = "xxhash-4.0.1-cp311-cp311-musllinux_1_2_armv7l.whl", hash = "sha256:85e402dab0f9acd3604539747c6fcc57dc188a18af6ab07eb8189351cd32466c"},
+    {file = "xxhash-4.0.1-cp311-cp311-musllinux_1_2_i686.whl", hash = "sha256:fb59a0dd61fb2ad481c03fda399d78ce57dab6bb62c2c8fdb446a7ba4754b89a"},
+    {file = "xxhash-4.0.1-cp311-cp311-musllinux_1_2_ppc64le.whl", hash = "sha256:0b20a06454b34f1531fc677c54efe2ecdec691ef9224f7fa919bf2c1363f7ff1"},
+    {file = "xxhash-4.0.1-cp311-cp311-musllinux_1_2_riscv64.whl", hash = "sha256:f7db035447a0ac8959aa230c5d36545ecf9f547413eb1711c0ca6f0ba1418925"},
+    {file = "xxhash-4.0.1-cp311-cp311-musllinux_1_2_s390x.whl", hash = "sha256:34ed93e20bfd98d722b902121643791eeb4b1641871e2dc63d0d4c2d93f187df"},
+    {file = "xxhash-4.0.1-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:f6247f5e23ee94f2557ac9dab738a336f607c6ff476fcf66ca70c3aef5eee15a"},
+    {file = "xxhash-4.0.1-cp311-cp311-win32.whl", hash = "sha256:348c8f288dc961d6bbd1985c8152a3ed7a85c95df00e82320f0c5215d922a399"},
+    {file = "xxhash-4.0.1-cp311-cp311-win_amd64.whl", hash = "sha256:ac0f291ab6485bd71f33941f9b92771318332a05d505460b41e893a549caadc0"},
+    {file = "xxhash-4.0.1-cp311-cp311-win_arm64.whl", hash = "sha256:72f34834518157a75e7090f328ee7a16c70c804cfc7c694fa069cc888e9fc03e"},
+    {file = "xxhash-4.0.1-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:1ee523f51718e41753f04f7102bb4dc55a18d2ea5cbaceef8ec7ca08571bd428"},
+    {file = "xxhash-4.0.1-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:515a822c73abbf6a0b7c70976d9662be342835c9d78b8dc7c023411f39c35dbc"},
+    {file = "xxhash-4.0.1-cp312-cp312-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:f5d031f35962e5483a613214e61f09fe24ab523062c3646d592dc16c4a217451"},
+    {file = "xxhash-4.0.1-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:da0264844a09b538c894e5eff25313d941deb4dedec2131b98418a71a3c9944e"},
+    {file = "xxhash-4.0.1-cp312-cp312-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:1642907941ee4b75aacc3db688af52ea02ca2305ab22af7ee686ed726b332684"},
+    {file = "xxhash-4.0.1-cp312-cp312-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:4af350bc3f329970c0e3a59af84a8a30998bf8a9167eb50cd48e59baaa1d7bec"},
+    {file = "xxhash-4.0.1-cp312-cp312-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:8ba782ca3bf1e81492611152b9a0d5264971339e95e34d69de0ac2c926be496d"},
+    {file = "xxhash-4.0.1-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:237b8f63a2a0fcfb1ffc06e21dad23add44e6d354b2b014364a1d41e419a4dee"},
+    {file = "xxhash-4.0.1-cp312-cp312-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:81507a68ba84c55241fb61cce1469f473a5da4205fc8ef6f698e5948eea8dd88"},
+    {file = "xxhash-4.0.1-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:5f1ea31d61bcd2cd2f3ec4ca80a64187bbd7948f490b63cf0dcbc6e717b4c1e9"},
+    {file = "xxhash-4.0.1-cp312-cp312-musllinux_1_2_armv7l.whl", hash = "sha256:06713a5aaf1d0905c5579416c020c02e42b3ceb931e86c7d3b7fb85403dee3f3"},
+    {file = "xxhash-4.0.1-cp312-cp312-musllinux_1_2_i686.whl", hash = "sha256:e8cda075b10bb3917b002c74a04f9e02b7d13b5bf732571404d51c52b11c7329"},
+    {file = "xxhash-4.0.1-cp312-cp312-musllinux_1_2_ppc64le.whl", hash = "sha256:c10b9206753b64aa791b35b201485477525b26fdec5bf86e8364c388a03e2592"},
+    {file = "xxhash-4.0.1-cp312-cp312-musllinux_1_2_riscv64.whl", hash = "sha256:f3e1a44af01b6692de0ec6caba5f0bf93ceb36896e02b7fc00952c6ea7ef39e1"},
+    {file = "xxhash-4.0.1-cp312-cp312-musllinux_1_2_s390x.whl", hash = "sha256:c6fc415b5568bd9accc7187f1729a99707330c0a67a8b9f93c1149ed573ed75d"},
+    {file = "xxhash-4.0.1-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:96d8de55029d42251945531f6aa7590c32b48163c66a43bf29d8657d7446a377"},
+    {file = "xxhash-4.0.1-cp312-cp312-pyemscripten_2024_0_wasm32.whl", hash = "sha256:0163b5d259de23ae9e07b7eabf435ce4704f6f205589a2b154e6af4be985ce1b"},
+    {file = "xxhash-4.0.1-cp312-cp312-win32.whl", hash = "sha256:1216f7ba5683f17a89eb7dcb4bc50a0b743dfe1902278d7b3d0786f538118433"},
+    {file = "xxhash-4.0.1-cp312-cp312-win_amd64.whl", hash = "sha256:5c2d525a3afabcd8e3549d85fc7e111fde6bc302d06a1893fe73adb79823415e"},
+    {file = "xxhash-4.0.1-cp312-cp312-win_arm64.whl", hash = "sha256:86b2b12bec60c678ed8f5cca0258ad93a8928ebddb6ca7732f0875afe1451d1a"},
+    {file = "xxhash-4.0.1-cp313-cp313-android_24_arm64_v8a.whl", hash = "sha256:8c9fe122444e129881afd1d4d1c7ac0d3ce2d91b68c2b40173b6025ff1c31f9a"},
+    {file = "xxhash-4.0.1-cp313-cp313-android_24_x86_64.whl", hash = "sha256:1f3346c5c287ac3c7f38b20380f55e8768230e7252af59fabcf3b87ab21e4256"},
+    {file = "xxhash-4.0.1-cp313-cp313-ios_13_0_arm64_iphoneos.whl", hash = "sha256:4e5141543c7f7fe3087500bbb4ac2845cb528a980aa91f8f1e661e2292ff4a5d"},
+    {file = "xxhash-4.0.1-cp313-cp313-ios_13_0_arm64_iphonesimulator.whl", hash = "sha256:f09ee747e2a5f876cc5ad56947734811828335e13b403dd8ea1e06d77a9dd48d"},
+    {file = "xxhash-4.0.1-cp313-cp313-ios_13_0_x86_64_iphonesimulator.whl", hash = "sha256:acf52474b2494ef66dc7e0fb6d5e2b50c18313039ad4d275fbf9f9907c804bc5"},
+    {file = "xxhash-4.0.1-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:1b3cccf75eeb5b01639b2feadb042a8e07889293b7ca72fa2985e7dcb64763cf"},
+    {file = "xxhash-4.0.1-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:cd878d32f5c6cbce9783f8d6897561fb772211edba9dde49d85672b88ed45276"},
+    {file = "xxhash-4.0.1-cp313-cp313-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:41e579025a6e13a99e6d71e39c9cfc621a0dcdbbf19106325e145fa858f2d794"},
+    {file = "xxhash-4.0.1-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:74379a577a9f3b6afbdedf1b90e5c7764467051977f18a326d7d607336d743bd"},
+    {file = "xxhash-4.0.1-cp313-cp313-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:acb31ecdd1a97fab5cd39a84ee9f515e727d319f796fec48703b8339b9998360"},
+    {file = "xxhash-4.0.1-cp313-cp313-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:5b7875ac1a2edcb691f27642b8b94b904baa6bcecb7d79c72df2228ba8cb5c51"},
+    {file = "xxhash-4.0.1-cp313-cp313-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:4751f1d7eecae6b2d2a773630f1a7248f125c9a92a456694d03c15bceffc9d68"},
+    {file = "xxhash-4.0.1-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:9a51b061d54cda8b83e62c44458bfbf0dabbef9b975dd9649952ba5076b9f349"},
+    {file = "xxhash-4.0.1-cp313-cp313-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:74a164e8b63f1e9cf35c9a7809d082b033d1a00e7375d5d814415436e7867e57"},
+    {file = "xxhash-4.0.1-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:4f5e5c6df4b703afcbe9352d238a51efd97c3b91fdc3a2052e40fdacb1e7505f"},
+    {file = "xxhash-4.0.1-cp313-cp313-musllinux_1_2_armv7l.whl", hash = "sha256:d54b8ae068af532c8cdf56abb9e09a60fbe7b10792444c9c27987bb6d3b450fa"},
+    {file = "xxhash-4.0.1-cp313-cp313-musllinux_1_2_i686.whl", hash = "sha256:1749f0688020209fe0d357ce1e1cd9ec9c6161ed0405ea949d24581c4c43fa91"},
+    {file = "xxhash-4.0.1-cp313-cp313-musllinux_1_2_ppc64le.whl", hash = "sha256:94ac8a6b8c47951173f0b67bf862bcb971bf24e493b9fbbdb0e010cbbc7d9f54"},
+    {file = "xxhash-4.0.1-cp313-cp313-musllinux_1_2_riscv64.whl", hash = "sha256:a33de7633c948ab2dc144af370a66e7e7af29b425dcd0f7e4f59689fb9391b53"},
+    {file = "xxhash-4.0.1-cp313-cp313-musllinux_1_2_s390x.whl", hash = "sha256:247ece770647c0aef080561fa996f9774b4dadce2d0c42eeb98229db7dcf820d"},
+    {file = "xxhash-4.0.1-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:a4553d36cc0b7fce1f35ba8a94dfd775aa3ed12f5eab2dc3b46ac75a0706b0bb"},
+    {file = "xxhash-4.0.1-cp313-cp313-pyemscripten_2025_0_wasm32.whl", hash = "sha256:87aa309a93bd5ec13f14309a305ff4e9bf74c5363fc46c264c0a22edfd5b0670"},
+    {file = "xxhash-4.0.1-cp313-cp313-win32.whl", hash = "sha256:cba763d84b06bda2c38d5185dee76f1b9dfdc0789e96e476d9e10005526d0788"},
+    {file = "xxhash-4.0.1-cp313-cp313-win_amd64.whl", hash = "sha256:97b94fb29abf21f5f0bde15f7dbdd3a4aa2dc59f37026adc7b4bee8563b84375"},
+    {file = "xxhash-4.0.1-cp313-cp313-win_arm64.whl", hash = "sha256:08ed8da18cd4fd0a6a5d6a444852d8fbd0e565388a74a4937085451b5f1a312a"},
+    {file = "xxhash-4.0.1-cp314-cp314-android_24_arm64_v8a.whl", hash = "sha256:af05a3f650220a6c59fa0ad2410249f2d2470a05225807c378fb67458693f8df"},
+    {file = "xxhash-4.0.1-cp314-cp314-android_24_x86_64.whl", hash = "sha256:a6e3653df1a70b8ac4191216324242e4be2bca18c9a7c10934e1bd56dc7ca15e"},
+    {file = "xxhash-4.0.1-cp314-cp314-ios_13_0_arm64_iphoneos.whl", hash = "sha256:4528cf80ebbbf57d40edfb31521ae265daa6dd636d615b1cf0ac86209579e59d"},
+    {file = "xxhash-4.0.1-cp314-cp314-ios_13_0_arm64_iphonesimulator.whl", hash = "sha256:90cb2a1c9cc503a054a19612b48ff6e8e47805f618bdb3224a07568aad03a37e"},
+    {file = "xxhash-4.0.1-cp314-cp314-ios_13_0_x86_64_iphonesimulator.whl", hash = "sha256:a949b072ea59c6eca0811ccd9e95133cc50d2afda8d464b5b077c78f78efa269"},
+    {file = "xxhash-4.0.1-cp314-cp314-macosx_10_15_x86_64.whl", hash = "sha256:79a3203aadf39637869dfea1185227d8452844d78b837e54fb1117b4d34ba5c3"},
+    {file = "xxhash-4.0.1-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:d9f3848ffaf010bdbabdbf4c25641fa258b6227ff27bc74a4d06edef521a4873"},
+    {file = "xxhash-4.0.1-cp314-cp314-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:9283d9dd6b44acad35118e2976fc763a065509e4118debdb61916ec322ed17b9"},
+    {file = "xxhash-4.0.1-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:1c7c642a0f79c3e3cf2965475507574d3d1a50ec71060039d60cb87358667cb2"},
+    {file = "xxhash-4.0.1-cp314-cp314-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:96dedccfb09a73a25751053a183159b88f4ee75f388df8166040c152ac0531c6"},
+    {file = "xxhash-4.0.1-cp314-cp314-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:81664268dba92e037b740ecf37fa02f1cab4a391f93f28e35792b3341c60648f"},
+    {file = "xxhash-4.0.1-cp314-cp314-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:839f58c5bd9989875be0fd28446dbf32cace2c2cd8bf2f6762acdc38a95cd1aa"},
+    {file = "xxhash-4.0.1-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:ffa44b4c7c5d0ffa31356b4428659516c0e47647825c74079a296b3857b6d99d"},
+    {file = "xxhash-4.0.1-cp314-cp314-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:e681a6fc7e4f715252b9b5acfb30536ec7dd1f75033a32dc617e6fa95af1a3fd"},
+    {file = "xxhash-4.0.1-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:c6301d92545c591ad31c3e050aa40a5f8a4c16413f1f9e6f9322c6f0f9d2b736"},
+    {file = "xxhash-4.0.1-cp314-cp314-musllinux_1_2_armv7l.whl", hash = "sha256:6efb8f21cc136c79b3e5bb747c8682d37916fb202cdbbc32182de5c4e47f821f"},
+    {file = "xxhash-4.0.1-cp314-cp314-musllinux_1_2_i686.whl", hash = "sha256:760de77279e9cf9c81d012ce0705cba13afccee9b09c480f17d778c8c5cefae8"},
+    {file = "xxhash-4.0.1-cp314-cp314-musllinux_1_2_ppc64le.whl", hash = "sha256:a16a3fa6936e36bb1414d16a6bd012c9033e5161b68b426805b61d895392437d"},
+    {file = "xxhash-4.0.1-cp314-cp314-musllinux_1_2_riscv64.whl", hash = "sha256:9c3c4b9aa9a27196b921197f7daf9e6c1412739df06a99cfa6e923879362eff6"},
+    {file = "xxhash-4.0.1-cp314-cp314-musllinux_1_2_s390x.whl", hash = "sha256:863f3d3b44110f7243e86cf994aa5c5d88f2348b6e84ab4402fadadfbf9f7da7"},
+    {file = "xxhash-4.0.1-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:63aa52659bc32bb9bd7cb5caf523b4d14429a477762cfac886132d687c1f80fc"},
+    {file = "xxhash-4.0.1-cp314-cp314-pyemscripten_2026_0_wasm32.whl", hash = "sha256:67e57b834e07ed973cee7b6da1548ff28a56458d77696fd2a5f397f340694848"},
+    {file = "xxhash-4.0.1-cp314-cp314-win32.whl", hash = "sha256:b6c1f9c59bbe593f88a0aad30be4150f15bd57bd64efb95feeabcb8e563f1ecd"},
+    {file = "xxhash-4.0.1-cp314-cp314-win_amd64.whl", hash = "sha256:da544672efd9ad76077928a3e6c5d894e52ce82d3bf14002db4a1bf17d1a36a2"},
+    {file = "xxhash-4.0.1-cp314-cp314-win_arm64.whl", hash = "sha256:d0d24a4f3fb63852cd09af46ae4b7a4d00cc8b8615a046dca543786e728d1056"},
+    {file = "xxhash-4.0.1-cp314-cp314t-macosx_10_15_x86_64.whl", hash = "sha256:349775ac30372b344d2338b2a168c0a1312a644194da25b8bec476d55761a128"},
+    {file = "xxhash-4.0.1-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:43e5f9169e73d0f0db33b5f6b8554bcce69ac278c966daf83d5eb4eb2f13829f"},
+    {file = "xxhash-4.0.1-cp314-cp314t-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:4a252fb862b0ae2590587e625f47a0e03da05cf0205e8830b67b6596c06038b1"},
+    {file = "xxhash-4.0.1-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:2df3ca8757dc381e75e90a4d7995a6324f58a923c7145220a7b2c0231f66fddc"},
+    {file = "xxhash-4.0.1-cp314-cp314t-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:bfed61996d618eb90d6eaae0178002e3466a28b06bfc557a7a3a7266378d8c5a"},
+    {file = "xxhash-4.0.1-cp314-cp314t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:9761ff4a0ffa583fe850731ad24fe82c88cccb7a2294727db0955f3279a4cb3f"},
+    {file = "xxhash-4.0.1-cp314-cp314t-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:edccc2ec58435a580f96a48a3ccae8cd0a480824119165dd90108718ad81ae6e"},
+    {file = "xxhash-4.0.1-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:4741d42d59e4e5fa1a86c17ab9c27dc8ea459c700d91b6742fdb9138d9a516cb"},
+    {file = "xxhash-4.0.1-cp314-cp314t-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:440c401e146ce64bdb3beb8ff0c84677b6f21307c28a34779071cecee5d4d70c"},
+    {file = "xxhash-4.0.1-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:5b7979f71d06ae45a769de0699900a246d8cb632db1e8bfdc79ec019063a503c"},
+    {file = "xxhash-4.0.1-cp314-cp314t-musllinux_1_2_armv7l.whl", hash = "sha256:62198213fc3e0c56e567894b318ba45834e007d065f84ba6dc9165d21546fc56"},
+    {file = "xxhash-4.0.1-cp314-cp314t-musllinux_1_2_i686.whl", hash = "sha256:b3bece52127ac20044311ee73567f9f0893b5de64f9028aecc90cc740cfd525a"},
+    {file = "xxhash-4.0.1-cp314-cp314t-musllinux_1_2_ppc64le.whl", hash = "sha256:a865d2d470220e659220fdb59d5b6c4422802d8d6098e1324bc4d12444798914"},
+    {file = "xxhash-4.0.1-cp314-cp314t-musllinux_1_2_riscv64.whl", hash = "sha256:8580aab306888224074c7edeec734de0c3c5ccde65b2da4e6c9a5e28f7c0a1bd"},
+    {file = "xxhash-4.0.1-cp314-cp314t-musllinux_1_2_s390x.whl", hash = "sha256:2d52dc7c33c1b83082b707f6b7814dc76d2faaa2ea62bd9c5fab4b36f83c087f"},
+    {file = "xxhash-4.0.1-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:6a9f98af872355e0c02439e48583958eee00e60b928bb20476460d9d40cb7b4e"},
+    {file = "xxhash-4.0.1-cp314-cp314t-win32.whl", hash = "sha256:a14578102a6081465aec9cf73c76c3cd3f79f0709bdb3b8ae7ab0b54c9d8b089"},
+    {file = "xxhash-4.0.1-cp314-cp314t-win_amd64.whl", hash = "sha256:c57963970d359a72262f7fe6be88f945e2334d4bc41462b7f08c37b0abf35ca6"},
+    {file = "xxhash-4.0.1-cp314-cp314t-win_arm64.whl", hash = "sha256:b659fad79c99b0238c7ad7e9d7dbf4eebfea9097c2dba65fa0a4d18a25b29a2f"},
+    {file = "xxhash-4.0.1-cp315-cp315-android_24_arm64_v8a.whl", hash = "sha256:5adf927dca8c47fde7e683fe69efdd81bc865c4db1fb6bb00b391e2b6185207b"},
+    {file = "xxhash-4.0.1-cp315-cp315-android_24_x86_64.whl", hash = "sha256:c30dd1af66a820820398b26e0d74e7a9aa43cae705924f23ed828cd8e5c26c3d"},
+    {file = "xxhash-4.0.1-cp315-cp315-ios_13_0_arm64_iphoneos.whl", hash = "sha256:1bc591533fc975614f7e13594daee76af96b8e1fbcf8de76c8773858fa9e7cea"},
+    {file = "xxhash-4.0.1-cp315-cp315-ios_13_0_arm64_iphonesimulator.whl", hash = "sha256:567cbc630302a46a8ecfd943b309ccf5372bb3718f1f3762d452df30f033bcf0"},
+    {file = "xxhash-4.0.1-cp315-cp315-ios_13_0_x86_64_iphonesimulator.whl", hash = "sha256:e998cb3685b92101ec5de0fb4d9485cf01e50bc418211955c55d98064664cf4c"},
+    {file = "xxhash-4.0.1-cp315-cp315-macosx_10_15_x86_64.whl", hash = "sha256:c3074db513c81f764053e3da079312ecf85a50d8350c71f4cc0105d9662a9e6c"},
+    {file = "xxhash-4.0.1-cp315-cp315-macosx_11_0_arm64.whl", hash = "sha256:3088dadbffa33c29e0518578430a7dff2e901a212e487aefa5faaa0dc06dad34"},
+    {file = "xxhash-4.0.1-cp315-cp315-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:1b50223d92df94d54e1a31469335a2c74b16692e6c1cb726f1e6949514458706"},
+    {file = "xxhash-4.0.1-cp315-cp315-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:427b62d62d4f967fbb10b82a3813e4875c2a6e7e7634739f17265b650c7f65a6"},
+    {file = "xxhash-4.0.1-cp315-cp315-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:c6370189e8e66b7e608f533b939a9de092ddca6cce084ca0d3d414d2ed5b5d59"},
+    {file = "xxhash-4.0.1-cp315-cp315-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:ec1a470c6db94ac4589c203921e89ac1bc13e796a8b1784d8135e1893559cd3b"},
+    {file = "xxhash-4.0.1-cp315-cp315-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:37f667dee0f867c42894b34e2a6fe26bf195c0ea4683d9d2b713db023f242c3a"},
+    {file = "xxhash-4.0.1-cp315-cp315-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:f18732adcc271741bd651c3e56fa519d8a237d2cccda01fe3afb226bf87f783b"},
+    {file = "xxhash-4.0.1-cp315-cp315-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:0b42a5a26607e4b2409fea174773a66f2dff9dfdbf2c1a851bb7b804e2c97535"},
+    {file = "xxhash-4.0.1-cp315-cp315-musllinux_1_2_aarch64.whl", hash = "sha256:99166cc98637e8bf550cda2aab07f4f1d5f899c45fbd721801aeabcc9d404824"},
+    {file = "xxhash-4.0.1-cp315-cp315-musllinux_1_2_armv7l.whl", hash = "sha256:6cf633df84d80a1668fcf61e330791dae46825e395549e7d34f376411e75088a"},
+    {file = "xxhash-4.0.1-cp315-cp315-musllinux_1_2_i686.whl", hash = "sha256:e259bb7e1e2d8de6b35f430f5c7220b1c0ebf3962d1ba7ec7545980d5931edb8"},
+    {file = "xxhash-4.0.1-cp315-cp315-musllinux_1_2_ppc64le.whl", hash = "sha256:704381264b36a18b9c62ecbabe2e71d0fc58c77c129c15355c989b10bf05b6b0"},
+    {file = "xxhash-4.0.1-cp315-cp315-musllinux_1_2_riscv64.whl", hash = "sha256:e90b4bcf1d9eb1010fdaee7c9209fb667e74c0684f3ba17f9032bd7319da90c9"},
+    {file = "xxhash-4.0.1-cp315-cp315-musllinux_1_2_s390x.whl", hash = "sha256:a65785e653573fcd1e33062760ab4c3c3440e8e910765018e4b6ed4ad07b54a0"},
+    {file = "xxhash-4.0.1-cp315-cp315-musllinux_1_2_x86_64.whl", hash = "sha256:e3996ff9b6f99180357024336bf5749a8ad6476a9a2523e535c5212b995b12a2"},
+    {file = "xxhash-4.0.1-cp315-cp315-pyemscripten_2026_5_wasm32.whl", hash = "sha256:99054b838b74d8d3995ea0d410976ae967c46207ae22d6ddfc535e809197dab9"},
+    {file = "xxhash-4.0.1-cp315-cp315-win32.whl", hash = "sha256:6c45258a37fc22721395c09927cb982d3e7a83607cab15be7e2416501bd3a330"},
+    {file = "xxhash-4.0.1-cp315-cp315-win_amd64.whl", hash = "sha256:0ab851b45c70d4992be7cdeeee16f97a0b677408c758c4b1efb1cfe8030bfd37"},
+    {file = "xxhash-4.0.1-cp315-cp315-win_arm64.whl", hash = "sha256:a5b21b42a01a343096a1c018d35e9b7aec9c7065dda53ae8da071e37478b2cea"},
+    {file = "xxhash-4.0.1-cp315-cp315t-macosx_10_15_x86_64.whl", hash = "sha256:44ab12e8cd17d4f001769f00ad465208b4bcb897ed29e65f058f74466b57a98f"},
+    {file = "xxhash-4.0.1-cp315-cp315t-macosx_11_0_arm64.whl", hash = "sha256:45e88111ebe331de478ef8d4293efbe88f3cf8b863386c9a2357136b838e1af0"},
+    {file = "xxhash-4.0.1-cp315-cp315t-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:bf430c587f447a554c53768ad76b9846fe7c5632180ef6f69c4fce8b0552fbd0"},
+    {file = "xxhash-4.0.1-cp315-cp315t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:adbd48b30e3f82c89fb2b3e6a87cdd28d113b190a5ed0ee2dee286323ee9a621"},
+    {file = "xxhash-4.0.1-cp315-cp315t-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:e71b34978e77868cbf2d18c5206a4603f9c644dd7181bec5643bd40141d3b8c5"},
+    {file = "xxhash-4.0.1-cp315-cp315t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:488ca5c5e28ef56ec4bbb12f835b3f1cbecc5f3510062e70117bc6594851932a"},
+    {file = "xxhash-4.0.1-cp315-cp315t-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:421b94f3ba7067958d02e38960d987756347aa150df06df11aa68ae1af78c619"},
+    {file = "xxhash-4.0.1-cp315-cp315t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:f33cf0baa91eccd2cb7b62bf00f10c2264ef578b71dd33a12962e71a36eb4d32"},
+    {file = "xxhash-4.0.1-cp315-cp315t-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:23a4376b4a3183cb50d4d2a3179f887a7773cc695eb2c908e551bec3221b8c60"},
+    {file = "xxhash-4.0.1-cp315-cp315t-musllinux_1_2_aarch64.whl", hash = "sha256:38c3d22129a6958846a3098d68bc8e661704461c0be4793ae28836e4690c8478"},
+    {file = "xxhash-4.0.1-cp315-cp315t-musllinux_1_2_armv7l.whl", hash = "sha256:87cbdec1a7dd930079671a60b249f3ca4e773e6fbd0676e21e36fdc9dd0f3b00"},
+    {file = "xxhash-4.0.1-cp315-cp315t-musllinux_1_2_i686.whl", hash = "sha256:6cbf4e21ef0890804b5bb9ad25c48f9c127758d7f6c66bef374efcacc63c738a"},
+    {file = "xxhash-4.0.1-cp315-cp315t-musllinux_1_2_ppc64le.whl", hash = "sha256:c101180495cb4ba3617b279a944345c53a5e73b0c150053d1fa8d8af32de9579"},
+    {file = "xxhash-4.0.1-cp315-cp315t-musllinux_1_2_riscv64.whl", hash = "sha256:c0e6ccc2b19ec8a726b2e26062ac71ea63e15500d6bf85910e42481844fdffc1"},
+    {file = "xxhash-4.0.1-cp315-cp315t-musllinux_1_2_s390x.whl", hash = "sha256:8bcba9456242ebf180a04d9443812fd85ffe6bd12bda464dd116fcece8886ff3"},
+    {file = "xxhash-4.0.1-cp315-cp315t-musllinux_1_2_x86_64.whl", hash = "sha256:83b8c2013edb5dc1f9e7268b6496130705bc48d79c86bb8817b3d210b81a5513"},
+    {file = "xxhash-4.0.1-cp315-cp315t-win32.whl", hash = "sha256:aa6ccc7f31018484d652cf52db020003433f3c9fa83189c028bd807d2adde503"},
+    {file = "xxhash-4.0.1-cp315-cp315t-win_amd64.whl", hash = "sha256:daade8936c4deaaf7b01561324ce438ba4f885d717e9adc62b4d67212ad7d7bd"},
+    {file = "xxhash-4.0.1-cp315-cp315t-win_arm64.whl", hash = "sha256:f00330ac7e24769e2032203f2b01794d670916b0c1799fd261340f1af9499875"},
+    {file = "xxhash-4.0.1-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:e9701c073bd062fb6bf6be51b47186ad15f1e87feedf4ea07198e0333ec068dc"},
+    {file = "xxhash-4.0.1-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:c7484fea54964edd417cc3a104d5180562514aa7c4e2a2bc26d776ef0c4cb4a1"},
+    {file = "xxhash-4.0.1-cp39-cp39-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:fc737c05ca2d48e5dcdbbb249314df3fc6c2a0be6da8b0aa28e13d72afaad7cd"},
+    {file = "xxhash-4.0.1-cp39-cp39-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:88d87719fe6bddf117238b341c5db851f8e96ba68ad9832b450e4a43dc60b37f"},
+    {file = "xxhash-4.0.1-cp39-cp39-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:4bbf3ff651e0f1a19beb5d0f48e0874a9bad2482a588c9d214c96ef1fff1cd9c"},
+    {file = "xxhash-4.0.1-cp39-cp39-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:09f9feb118966cc6650e1806205d577eae7ca394aa6acf349a0b62a94bbeb329"},
+    {file = "xxhash-4.0.1-cp39-cp39-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:303121aab4b7f898058582d7962ea79d9e26e2379d7b6d8743f70f2671674481"},
+    {file = "xxhash-4.0.1-cp39-cp39-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:70129ebb8f20e1ac1da58b78ed381624bd689a43a9a7366560bd8fabea145105"},
+    {file = "xxhash-4.0.1-cp39-cp39-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:3891efe3d7a531ce6da0a4a50a99dd41c75b8fd4ca19d73c86431b4db5c305f0"},
+    {file = "xxhash-4.0.1-cp39-cp39-musllinux_1_2_aarch64.whl", hash = "sha256:2696bbac613f6880fed60316c298bf3091d4f8eee3ae2e9466f70bb76204fb0c"},
+    {file = "xxhash-4.0.1-cp39-cp39-musllinux_1_2_armv7l.whl", hash = "sha256:684160b3c0a9b62c6f0de90f44e11dc5d8643dcfa18a5856b45fb1c47478bb71"},
+    {file = "xxhash-4.0.1-cp39-cp39-musllinux_1_2_i686.whl", hash = "sha256:f83295394d34e1287e5b30fcc496c13b92cf886a131f3dae5444e38da8757efb"},
+    {file = "xxhash-4.0.1-cp39-cp39-musllinux_1_2_ppc64le.whl", hash = "sha256:87da13df72c5612771cd905a8b121e0bfea62d7659b1c92198736eb722220e83"},
+    {file = "xxhash-4.0.1-cp39-cp39-musllinux_1_2_riscv64.whl", hash = "sha256:a6671a8f6ea4f2101ce11fab5023a2e59391cff249fc3928cecb69d971525fd5"},
+    {file = "xxhash-4.0.1-cp39-cp39-musllinux_1_2_s390x.whl", hash = "sha256:e3eba72f9bb84fe696516f4cbca68d3d74a376157e68bacddbb7f2516af61523"},
+    {file = "xxhash-4.0.1-cp39-cp39-musllinux_1_2_x86_64.whl", hash = "sha256:649f2682c090cca1ac4037866381f3652eaacbd56e5178030f4ce1325b8f945b"},
+    {file = "xxhash-4.0.1-cp39-cp39-win32.whl", hash = "sha256:168dd6b51725a222abc722832e56624d15a63fc2e8249021509c93f1063913f6"},
+    {file = "xxhash-4.0.1-cp39-cp39-win_amd64.whl", hash = "sha256:a69e8946e4902ea11fc1c557740cdbfe7d75c78fcc5e4324ff89a696a634357d"},
+    {file = "xxhash-4.0.1-cp39-cp39-win_arm64.whl", hash = "sha256:3358097d333d40657569ec1121e21043dd7d0efa10aead1b50e8b4fa83077d7b"},
+    {file = "xxhash-4.0.1-graalpy312-graalpy250_312_native-macosx_10_13_x86_64.whl", hash = "sha256:ff48915bf1871a1f19f74c11834c6329443d306cedc0c05fe7fe617810422a80"},
+    {file = "xxhash-4.0.1-graalpy312-graalpy250_312_native-macosx_11_0_arm64.whl", hash = "sha256:4a76345f5aceb4ec404918edf9c7f2b5507db864dc0d7455982009ac0890b57b"},
+    {file = "xxhash-4.0.1-graalpy312-graalpy250_312_native-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:31d86f9e81f3e84e00131ac7c54caf5119ae4ddd82c09c31cff597c813ce1ee2"},
+    {file = "xxhash-4.0.1-graalpy312-graalpy250_312_native-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:deca2a30d983d240b8375ec2ee0a4288e72042827fc61df2f7671f8467e4cb2f"},
+    {file = "xxhash-4.0.1-graalpy312-graalpy250_312_native-win_amd64.whl", hash = "sha256:7c343ee174d417a44d0c3355602c0cbbfa52a04d1bbbf1723378c7d2c8f60626"},
+    {file = "xxhash-4.0.1-pp310-pypy310_pp73-macosx_10_15_x86_64.whl", hash = "sha256:26fe6238c2d5b11ed5063b9bf4eb290624b004fd074688da6bb079bd564f10d7"},
+    {file = "xxhash-4.0.1-pp310-pypy310_pp73-macosx_11_0_arm64.whl", hash = "sha256:e53926e76131a74e79cc0b39fa712c227875f180afc68646bd1e1d8a17e60313"},
+    {file = "xxhash-4.0.1-pp310-pypy310_pp73-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:0718ad66f4ded2411f8e62bdba549ee71e313a2d26ef5060ca3fdbf29897dd3c"},
+    {file = "xxhash-4.0.1-pp310-pypy310_pp73-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:1c2200b98a805351cb3142ae4e1fdcc9e91b5e20f5d30d4862b0b96f92558f4e"},
+    {file = "xxhash-4.0.1-pp310-pypy310_pp73-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:ea5ecf800b45bdb34afe05a1d0dae1f8ea02a290e50636dccd399063f6b180f8"},
+    {file = "xxhash-4.0.1-pp310-pypy310_pp73-win_amd64.whl", hash = "sha256:ce6d5cc94a50291d080259a126cbf1e9ba4ac861e6429d2f3cdbb1474f51945d"},
+    {file = "xxhash-4.0.1-pp311-pypy311_pp73-macosx_10_15_x86_64.whl", hash = "sha256:554f87034635bcec47c5d72447bf3db7e02da1bf493a0ada010db28a76f891c6"},
+    {file = "xxhash-4.0.1-pp311-pypy311_pp73-macosx_11_0_arm64.whl", hash = "sha256:3c2445edafc300cc40feb6a25a8356a971c30cd0bf47b5349c2ad74c508343b1"},
+    {file = "xxhash-4.0.1-pp311-pypy311_pp73-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:bdd16718b63aa3ebd68aabb79021a40e47c81374852d41a306b9453141bbcbee"},
+    {file = "xxhash-4.0.1-pp311-pypy311_pp73-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:8b99ebaf9e816ac5069423b1367ee7e8078fbcebcf62545506bb0608d2f4f468"},
+    {file = "xxhash-4.0.1-pp311-pypy311_pp73-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:f484ed57bb3e4142f9d6439568658c38be5f94b702ba00a1ff32c69783b6c66d"},
+    {file = "xxhash-4.0.1-pp311-pypy311_pp73-win_amd64.whl", hash = "sha256:fac4832b638000106207bc44e44b9616a6a416aaee56c62b01d61f3705e49f58"},
+    {file = "xxhash-4.0.1-pp39-pypy39_pp73-macosx_10_15_x86_64.whl", hash = "sha256:436e11b4dd966afe5f7f665e4cc4c5485ffe3ceb42f25a22e1701d236abf1853"},
+    {file = "xxhash-4.0.1-pp39-pypy39_pp73-macosx_11_0_arm64.whl", hash = "sha256:7236be540d6be9ce448d98b940dd26ddf70ca41012e8a14a53fd9354cefe4e8d"},
+    {file = "xxhash-4.0.1-pp39-pypy39_pp73-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:ad889d58361a26ba75f5d6a1a0da08ed4950ec4ac8a6da86e1c5ce1b95ccb43f"},
+    {file = "xxhash-4.0.1-pp39-pypy39_pp73-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:e961093277ff9d42addb9dad5614dfb7800ccba07c245c39c8e9b4daa35d160c"},
+    {file = "xxhash-4.0.1-pp39-pypy39_pp73-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:d1e0d1ea6e44f51808a9e8469c8afdebcdf6fa23d1ea524a0303d57d23919712"},
+    {file = "xxhash-4.0.1-pp39-pypy39_pp73-win_amd64.whl", hash = "sha256:764b32d52d15b8b95ac8160e540772fa1adeb611fe40bffaeb42e7bf98279e44"},
+    {file = "xxhash-4.0.1.tar.gz", hash = "sha256:d55bf4ef10eb09b8b6866790e083d26d087d84caa3cc0946ba87c3ca7ecaf7b7"},
+]
+
+[[package]]
+name = "yarl"
+version = "1.25.1"
+description = "Yet another URL library"
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "yarl-1.25.1-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:142c06c4d6a35ee3ec5da08499805e879cb3ca7c1fbfbecb0140fe72403818d6"},
+    {file = "yarl-1.25.1-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:24ce942011a61953e7d313438038f4d32ff21387b775f58a957f7a07dd55ef95"},
+    {file = "yarl-1.25.1-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:9e23c82b63cd7652fc24d33ed6cc17099d607aa3b4fc4ddc75e95062f3d82df4"},
+    {file = "yarl-1.25.1-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:8ee202350cf57abf0e9502a41601841019c25d3db7ff52d980aaf31446254059"},
+    {file = "yarl-1.25.1-cp310-cp310-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:5df89f769cc8ff94c3d7e7603386fba309d25ce5240132d26c15baa8d0e96c4c"},
+    {file = "yarl-1.25.1-cp310-cp310-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:83e9f4a25085bd4b7214701a0794ff1f50fc633ffb8bdfebf07abdd81c2db126"},
+    {file = "yarl-1.25.1-cp310-cp310-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:e636b64d24fd9c38053c5e389a1174c66361fa49dcfd220f4dd35b4abde7cb89"},
+    {file = "yarl-1.25.1-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:e5637ca8d0bd7fb72648a6c7934af4baaccb697657f7438c9d264fc2abb8b0b1"},
+    {file = "yarl-1.25.1-cp310-cp310-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:683e362b8ba453080f7489c66f4ea794e751c35b72e7eab3575ef784c2fbc7fb"},
+    {file = "yarl-1.25.1-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:df23df54b5114a17c2d0ef192433e2e5a9f0c5178c32375e90b7cfc965f349d0"},
+    {file = "yarl-1.25.1-cp310-cp310-musllinux_1_2_armv7l.whl", hash = "sha256:f53dcd26694f148f738edc052b5a69234833e739f10f4c3287bdfd8ec0f7b326"},
+    {file = "yarl-1.25.1-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:b8075fe90bc08e40b8b8a1874fab42ee4c7b56af05c5886e9cc841397f916908"},
+    {file = "yarl-1.25.1-cp310-cp310-musllinux_1_2_riscv64.whl", hash = "sha256:a8c2b841478068440d8b733005d13a5ef535b9928cbc05f17182d410f32ba449"},
+    {file = "yarl-1.25.1-cp310-cp310-musllinux_1_2_s390x.whl", hash = "sha256:ca32926d7d77bcc8838425c4c95e040a3ace1cb7dfdae599013458dcda2607ca"},
+    {file = "yarl-1.25.1-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:192a866877a49993949ef1975864ad8728bea28ee810f6abe1a0729c2b500426"},
+    {file = "yarl-1.25.1-cp310-cp310-win_amd64.whl", hash = "sha256:3f4d48a6112712973e676bd792121fee470e432d749177162d9949d5c9460a1b"},
+    {file = "yarl-1.25.1-cp310-cp310-win_arm64.whl", hash = "sha256:48796ea00a303961507dc6c8437c4b325a6fc3f95f7c36c71b91ea9a8150963c"},
+    {file = "yarl-1.25.1-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:9d693bf4bf534e9ba3ae2780cfd577f5135629f7b5ac653490859d0b77864865"},
+    {file = "yarl-1.25.1-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:ab2054c5531af2a9ba7b69b8ec91e4f884420e83a8c5e579b013084cb57e5e5d"},
+    {file = "yarl-1.25.1-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:564fdc7085d2245ab84f88882fdb1d6ac0723124bff6ded35bfb1c00f812630d"},
+    {file = "yarl-1.25.1-cp311-cp311-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:acae6b45d1ace09b6ba3876da43b88366ef368f73b988c7f57e14231753d4420"},
+    {file = "yarl-1.25.1-cp311-cp311-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:1fb2a01ba8cd9c5d2c5dc1ec35e0fc951d04b4f037541d4ac090c993ce58b3d7"},
+    {file = "yarl-1.25.1-cp311-cp311-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:e92b6bcc741b86d67606c40d3cb9c7cc8e6c737f81e31f4a94efc204456c92e3"},
+    {file = "yarl-1.25.1-cp311-cp311-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:72c34ac7ad4314c19362d5ce27626dcc8429bd30bbf8c179f4234078851f9492"},
+    {file = "yarl-1.25.1-cp311-cp311-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:d5add7b4ca7afeea91d52e4d4e4db3b1fe9885b71f07054560d8c4296b7441a2"},
+    {file = "yarl-1.25.1-cp311-cp311-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:def538065f9e4d4cf1ae164bd59aba00dfa84f03923e0de4c3788f252d6bcd17"},
+    {file = "yarl-1.25.1-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:0a191bfdb30a79b98e5d175d75285f9fcb78bf0e46ba5efda042e1c72071a0de"},
+    {file = "yarl-1.25.1-cp311-cp311-musllinux_1_2_armv7l.whl", hash = "sha256:71f42c5b9a948c113bbdebfa544598321431d064ff959d32e99b1feb61d68345"},
+    {file = "yarl-1.25.1-cp311-cp311-musllinux_1_2_ppc64le.whl", hash = "sha256:72849d892954be4d09e569b8b831ac39ce58417fedc767d4308a0fe542018a40"},
+    {file = "yarl-1.25.1-cp311-cp311-musllinux_1_2_riscv64.whl", hash = "sha256:efb01a106f971cb3752856bca2318bbdf7f01bd8823779c461586cbe5ffd5258"},
+    {file = "yarl-1.25.1-cp311-cp311-musllinux_1_2_s390x.whl", hash = "sha256:a1daf47cd95a7c3a63456336bc5aaa8c86dd3a47d07ed3d0e76132ae4666a5a1"},
+    {file = "yarl-1.25.1-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:9489e6abf47ba37f332075a91444c7cfedb03e6ce99fbb2f116bfe1ce810da3b"},
+    {file = "yarl-1.25.1-cp311-cp311-win_amd64.whl", hash = "sha256:d7306dee25b8a0e737363f347362b875094b4dc4e367311470656ae420fdbf8e"},
+    {file = "yarl-1.25.1-cp311-cp311-win_arm64.whl", hash = "sha256:abb1384477f5901d436b5d2e5465954de46ea6098f59163d243660b5c4461d35"},
+    {file = "yarl-1.25.1-cp312-cp312-macosx_10_13_universal2.whl", hash = "sha256:94d7aa6debf92a1dd14cb5280b083a764169a13cfb23a452111160274ed989f4"},
+    {file = "yarl-1.25.1-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:83d4a37e4b95da4d8bda930d6d35b75b4cdadbacbb4980cae290ea3100b5d51d"},
+    {file = "yarl-1.25.1-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:e029648f9c951db30e98a7d7ec90835db88ec4b32820efe2a9bdc2287e032eb6"},
+    {file = "yarl-1.25.1-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:4d781294bb815ecb5ea57ff6bbf8038e0a31a95fdf3e1788f66e0dc100d64b58"},
+    {file = "yarl-1.25.1-cp312-cp312-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:e12c538e00e7c1b286a07061046b90e8124e6a9793efae2c70db6a4aad07faad"},
+    {file = "yarl-1.25.1-cp312-cp312-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:7e4de3ac4adbad3d0bc7c6f4360a7dbff5de2f15e3b723be3198074e17fd9c40"},
+    {file = "yarl-1.25.1-cp312-cp312-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:419f392a1da624877975709e3864dfe833af6cc7671b39318086d456e288380c"},
+    {file = "yarl-1.25.1-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:c6f117789d22dce188e5754e8bc65b7e6ebf8cb73963b9fa761f672a5883769d"},
+    {file = "yarl-1.25.1-cp312-cp312-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:80e47012e730da131c9f059c80936783f9659aae22dc31c03c0595590d11ed54"},
+    {file = "yarl-1.25.1-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:e80f557716fd765439577131e526b8942ffc2c07bdbc5e39fa62f660ba1e963f"},
+    {file = "yarl-1.25.1-cp312-cp312-musllinux_1_2_armv7l.whl", hash = "sha256:f61964f235a43738bfac50da46fc4254943a7eea3051aeb0b6fc7c992c29fadc"},
+    {file = "yarl-1.25.1-cp312-cp312-musllinux_1_2_ppc64le.whl", hash = "sha256:e546fe1d4a93ebc2910f0d768baff19faa09843ab3f2036a67ed6e69fae4419d"},
+    {file = "yarl-1.25.1-cp312-cp312-musllinux_1_2_riscv64.whl", hash = "sha256:cce0727fd5ac04d372fa9bbfde9febc2bcf209aadfcf0468e45dec72719895d1"},
+    {file = "yarl-1.25.1-cp312-cp312-musllinux_1_2_s390x.whl", hash = "sha256:af4ea5b37403ef4e30f3927eaed540db942bde01d8d3ff083527c0704d1c9c68"},
+    {file = "yarl-1.25.1-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:68782fdb4027b8d1eee25ec35e9a6db05e863b899eb0310b3a33b6c3fef55707"},
+    {file = "yarl-1.25.1-cp312-cp312-win_amd64.whl", hash = "sha256:7d575b54cb3863ef9bc290ea4b009999d55dc237326131e4853cf33e888fee03"},
+    {file = "yarl-1.25.1-cp312-cp312-win_arm64.whl", hash = "sha256:bc3ac7bf569f6b64dad04dd7808c7872dae8a97df657856eac05e9b7e3614a85"},
+    {file = "yarl-1.25.1-cp313-cp313-macosx_10_13_universal2.whl", hash = "sha256:25868beca8b6765f8f7d0e11fe6dd7c66dd4b0793b9500286d20cc92352126a5"},
+    {file = "yarl-1.25.1-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:10b2fd95332f0d716d5eee3c9fb2ce8eada19082de7fee83d32e37992fd75c26"},
+    {file = "yarl-1.25.1-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:0f12afda4eea8c8994a76d4df1875c765194f5fbe8a9d197929ea303caee29ec"},
+    {file = "yarl-1.25.1-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:14b79a30a93a3ce2e8832603fd0ab780ada281b0ba5110b519a634f2d7d7d1fc"},
+    {file = "yarl-1.25.1-cp313-cp313-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:4bd6340d20ae2c7ca719b87b426e808e90743b676d05d4c26c4fb5ca71f41184"},
+    {file = "yarl-1.25.1-cp313-cp313-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:126a2533570c554719ca40a1288fdee1700b6bc82e7131aa69fa85252d92e651"},
+    {file = "yarl-1.25.1-cp313-cp313-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:a3faadac7d812ddac258feb57b9846b60c1b437c4f4b9ad42595c6f6fe4390df"},
+    {file = "yarl-1.25.1-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:be80550d9bfe83d9b62398a37081a90434e6df2d978ec345c3d2820de6beddab"},
+    {file = "yarl-1.25.1-cp313-cp313-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:e07595c7d6f4db270ceede356a1bd1c07a34f1c26f958d1ed0cd7b48e0d2bba3"},
+    {file = "yarl-1.25.1-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:eb96ed1ae6c7d072d60840c0434aef07a2df611812810807fbc54263a6053e9a"},
+    {file = "yarl-1.25.1-cp313-cp313-musllinux_1_2_armv7l.whl", hash = "sha256:3feb99222553a8cbedfa52c2f59dd84c3f50d5b582c728d522caf8d72769a54b"},
+    {file = "yarl-1.25.1-cp313-cp313-musllinux_1_2_ppc64le.whl", hash = "sha256:a2ed0ba415ccdf08f14bf544cb78346d0f76086707ffee24921a2c84dbf1305a"},
+    {file = "yarl-1.25.1-cp313-cp313-musllinux_1_2_riscv64.whl", hash = "sha256:2b49375d22299b0a834c2bca72f39aaecc270d96fb24c30424899676f487b22a"},
+    {file = "yarl-1.25.1-cp313-cp313-musllinux_1_2_s390x.whl", hash = "sha256:ef74070ac553c59eb4f04258722066d6c6135b7baa03b2e9f2da65c096e96d98"},
+    {file = "yarl-1.25.1-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:0a66db89ea473abeac4b70523cafd94db3772380e565f9d28af7a179b7af71fa"},
+    {file = "yarl-1.25.1-cp313-cp313-win_amd64.whl", hash = "sha256:1f51020b2eb8a003c84925638ec63c21a750a4bddd3a22ec8eac6a742dadf1b9"},
+    {file = "yarl-1.25.1-cp313-cp313-win_arm64.whl", hash = "sha256:b10dd0557ba422715b5206b3743192135a6022acca8baec51aa127d0a75db8fe"},
+    {file = "yarl-1.25.1-cp314-cp314-macosx_10_15_universal2.whl", hash = "sha256:a9ca696eb02e5c02a8afd872ada510eba9b7fe6e68b9572c2e9a9b1941e31e2e"},
+    {file = "yarl-1.25.1-cp314-cp314-macosx_10_15_x86_64.whl", hash = "sha256:a5877f2255aab518ebe528289037699201d5dc5f045f2396cb30aa02db22f57f"},
+    {file = "yarl-1.25.1-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:7a5c3115595995779ee21f2567035793911c3802a43c74f3fbb0314929ec67ac"},
+    {file = "yarl-1.25.1-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:77e5099b99b37f3cf79c246998ca9f7313a78054cd1809ec46bc1afad47e1c4c"},
+    {file = "yarl-1.25.1-cp314-cp314-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:6efaf45df6a849cef613a03a94c845647456662f85438c886bb67a9c027c8c2c"},
+    {file = "yarl-1.25.1-cp314-cp314-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:d5f90e44653c4e0f78501ed9bb7d3fce835a8d62b7c6ed0cb16557534087e743"},
+    {file = "yarl-1.25.1-cp314-cp314-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:632da579b2d879f6bad20f2cfa35ded1efe2f4f77f8abb26a6234a5b236acd2f"},
+    {file = "yarl-1.25.1-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:30eec96e8a91bd588ce897c9543f6d5d8d34b28fbcba28a4dedf20ebeae9fe57"},
+    {file = "yarl-1.25.1-cp314-cp314-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:12b6bc4906e11f5e1a1cdcb12296e7afbd366c783cc8073403cd2fb74334e453"},
+    {file = "yarl-1.25.1-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:9d6ed3d17bccce4c05343e1ca8da13bc5c02c812a4e7282ddd05e8769322d3fc"},
+    {file = "yarl-1.25.1-cp314-cp314-musllinux_1_2_armv7l.whl", hash = "sha256:f38a70074041d3b7e138e452799f5174198bae5bd5ab2000917badf403908c5f"},
+    {file = "yarl-1.25.1-cp314-cp314-musllinux_1_2_ppc64le.whl", hash = "sha256:4ca89e4e21854ed27ec753297dde84b16c9f8e53b14a4866fb44457d643c19f8"},
+    {file = "yarl-1.25.1-cp314-cp314-musllinux_1_2_riscv64.whl", hash = "sha256:1ab7618921a93767387a4b83776f751588f5b5ae9bb5bc96620e2e2e00bca868"},
+    {file = "yarl-1.25.1-cp314-cp314-musllinux_1_2_s390x.whl", hash = "sha256:0ae12ff2b805fa02c4dab838005caef735e39986322698c48588d3beacb65c62"},
+    {file = "yarl-1.25.1-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:90c30ed53546da833c700115c0064c22120d1b1560f474699fd31f22dd668233"},
+    {file = "yarl-1.25.1-cp314-cp314-win_amd64.whl", hash = "sha256:acfa7e22aa6c6e7a5996a41d275bfa01efa7ea56ab890590280e9063e2cf5c1b"},
+    {file = "yarl-1.25.1-cp314-cp314-win_arm64.whl", hash = "sha256:8e7d98cdbb6d71e726f7d525952867096053d1f290dd4e3c50d7d313a136f414"},
+    {file = "yarl-1.25.1-cp314-cp314t-macosx_10_15_universal2.whl", hash = "sha256:d21f0fa80a02d05299207eeaafef345d812ace96d5306e4ef265e1d419a615fa"},
+    {file = "yarl-1.25.1-cp314-cp314t-macosx_10_15_x86_64.whl", hash = "sha256:17c9877a89fb6e2bca6f9087eb24cd7fb434653946ef5075e470d23d49b52287"},
+    {file = "yarl-1.25.1-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:29273edf1530e397bd07cb784db1fbe0d2590b77569f2e24679a9c0a2d763b94"},
+    {file = "yarl-1.25.1-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:b7abffdf37af1cec6a2ad69b827aa84320db5894791bc8ed932dc93fb274b7e9"},
+    {file = "yarl-1.25.1-cp314-cp314t-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:2239a02249d9326655419e0168a28ca9008938eaab31dc29fc875c217927a6c0"},
+    {file = "yarl-1.25.1-cp314-cp314t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:664ec6a520b74a1df2810666eb67695fcb77fa663e6ea0a25aaf2e529cb24dfa"},
+    {file = "yarl-1.25.1-cp314-cp314t-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:4f1c91f5a5980a937ff8e238e98e6897e1ad74a4b1e2c0d68c73b5ffbb3f5c0b"},
+    {file = "yarl-1.25.1-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:7c88edaec8c349ad4c5ad4c486a3defcc4b80ceb2f074436ffa0a87caf5e76a6"},
+    {file = "yarl-1.25.1-cp314-cp314t-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:35dcbea443fafb3eece757ad4e514560ddeb6c34cfae1582c620d7b293d7feee"},
+    {file = "yarl-1.25.1-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:882569ff613758cac762a457a5d72d6e211b28d4bcfea89d1d71ea942b02eac0"},
+    {file = "yarl-1.25.1-cp314-cp314t-musllinux_1_2_armv7l.whl", hash = "sha256:d0f1489233a254bb3643d2f05de7d59019254d81daeca6b9162fe9edef57e0c7"},
+    {file = "yarl-1.25.1-cp314-cp314t-musllinux_1_2_ppc64le.whl", hash = "sha256:f41753a76f4f63927d03a0d8ba8f5ce0f2083bec29a8cfaccc55371b1564b96b"},
+    {file = "yarl-1.25.1-cp314-cp314t-musllinux_1_2_riscv64.whl", hash = "sha256:8fb0eb4955adf0579001581f2f71a126e8781ba61bcd120f127b0401163c6c2d"},
+    {file = "yarl-1.25.1-cp314-cp314t-musllinux_1_2_s390x.whl", hash = "sha256:a1e32763e641a1566507d90a8d3b19bfc3cc04a9d4e5ae3e32189874ed4b58a3"},
+    {file = "yarl-1.25.1-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:65b5b2066651b7432d389e9799d979c703bcc6ef44266bb8153ef54e91e4aab3"},
+    {file = "yarl-1.25.1-cp314-cp314t-win_amd64.whl", hash = "sha256:734f6e5400352ac4254456003d462866c684703570929cff7a7bde015d0cb371"},
+    {file = "yarl-1.25.1-cp314-cp314t-win_arm64.whl", hash = "sha256:287e99ff5aa4dc1c7630bfc683ded6f106d756c99dec432a2d7f197a784f51c6"},
+    {file = "yarl-1.25.1-cp315-cp315-macosx_10_15_universal2.whl", hash = "sha256:9b1bdaae98bc016825dd3c9d8ee1832f829b3341f9cc6ebd1a1b0a7fef7367cc"},
+    {file = "yarl-1.25.1-cp315-cp315-macosx_10_15_x86_64.whl", hash = "sha256:e7011b8fb8c4054bf0c12e5edc6cd83778b0028e99ce59b18586ed036f92cfdc"},
+    {file = "yarl-1.25.1-cp315-cp315-macosx_11_0_arm64.whl", hash = "sha256:f074e8d4aa0a5798920ddb6de3d08b228c614ff3724c3e8bd7577f4bafea867b"},
+    {file = "yarl-1.25.1-cp315-cp315-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:7d42e7e3ca399555578b4d617e3a6ecf13371b3743a115995fa010c7bf341459"},
+    {file = "yarl-1.25.1-cp315-cp315-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:aa4ed3dd308548f9e707d9caaf005d2d7f8c1e7868f858dfeb47fe76e16b391d"},
+    {file = "yarl-1.25.1-cp315-cp315-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:42a66563d8cc056ee32e6191e05097a7b2b3bc302e0bc3133daf8710eb18bd26"},
+    {file = "yarl-1.25.1-cp315-cp315-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:98d370568f393215d605304cdb77b3d5539bd192c75b623c7304c42c8d6d8273"},
+    {file = "yarl-1.25.1-cp315-cp315-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:23bf5b403c879a54964e0feac7285688e04bb220074878d737d331522da0a5bf"},
+    {file = "yarl-1.25.1-cp315-cp315-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:d45673badd08456d0340e9364eddafe1c53a9d2896424294de4d7dd71ad3ee57"},
+    {file = "yarl-1.25.1-cp315-cp315-musllinux_1_2_aarch64.whl", hash = "sha256:0136d640dfa9b0523853e411430a99f8a91eca85774c6420285a33b755bc6de3"},
+    {file = "yarl-1.25.1-cp315-cp315-musllinux_1_2_armv7l.whl", hash = "sha256:59ba3a6e1aa8cfe5adf4bd270fd965db21955401b7ca6f1696010c55ed4daec2"},
+    {file = "yarl-1.25.1-cp315-cp315-musllinux_1_2_ppc64le.whl", hash = "sha256:87796fedc3ba97ec14fab55acb48584276e6c1e4c1e89c422bda62c838e754a9"},
+    {file = "yarl-1.25.1-cp315-cp315-musllinux_1_2_riscv64.whl", hash = "sha256:bd0912757081f89b107d6c00b2ff8a194401b0b87eadcf4481de2b865a8fd44f"},
+    {file = "yarl-1.25.1-cp315-cp315-musllinux_1_2_s390x.whl", hash = "sha256:b51c159a9794633f5e0db7ecec7b2b6e3734eca1f5d17dc989ff3552a43ff78b"},
+    {file = "yarl-1.25.1-cp315-cp315-musllinux_1_2_x86_64.whl", hash = "sha256:319e070a01db9920fb63761843f96a104c8e2b9427266731810dc1e22595b17c"},
+    {file = "yarl-1.25.1-cp315-cp315-win_amd64.whl", hash = "sha256:a2059a2d891bd156bc5184e7ab7a56e78a84dfcfdeac8c501b552533ad1c36ee"},
+    {file = "yarl-1.25.1-cp315-cp315-win_arm64.whl", hash = "sha256:a78b50b4f7918a3de71105d5c0b93bbc57bb8339a4d03a9dfd449f9068e76f3d"},
+    {file = "yarl-1.25.1-cp315-cp315t-macosx_10_15_universal2.whl", hash = "sha256:b5402a340723fa7da00b5cff987ddab61276be6d11251ea71ae02bcac54890d8"},
+    {file = "yarl-1.25.1-cp315-cp315t-macosx_10_15_x86_64.whl", hash = "sha256:eda19ea5ee88742f47a2340816e6f2d40b53bed3ab5b69794769f36af9f35bb4"},
+    {file = "yarl-1.25.1-cp315-cp315t-macosx_11_0_arm64.whl", hash = "sha256:75baa6cf9b6d1c52f3e111a130e202fd8cf0a5b3a066c3f73d615e885092e4ec"},
+    {file = "yarl-1.25.1-cp315-cp315t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:dbcef5a9119ef653653132cccaf999b30a0af6f33bb0a4ba80bec30056868487"},
+    {file = "yarl-1.25.1-cp315-cp315t-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:7efc9f082dfed77c316edffa9deb52888e1bc6789171887cc1f68e06d65465c8"},
+    {file = "yarl-1.25.1-cp315-cp315t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:fe01645169a2112aa1d4ebc3e4c5f029c5c8f97adfc32e5d37c993b39a994d75"},
+    {file = "yarl-1.25.1-cp315-cp315t-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:d1c557dfd5e3db046053a0bdc72261ade790ebe8e2c7a41b36b0ca1f14cb95f3"},
+    {file = "yarl-1.25.1-cp315-cp315t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:8ce4d6ccafb33d39bd78444612d14938ead674c25702ded2ee9c54a47735d225"},
+    {file = "yarl-1.25.1-cp315-cp315t-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:80a063f8297fc796296f00f100be520f209b23dc98f93ce8eba6ee7122598209"},
+    {file = "yarl-1.25.1-cp315-cp315t-musllinux_1_2_aarch64.whl", hash = "sha256:7cb414a73e21a7ab58254926073f2930cb22f5b4314ea4260a687e2b3fd4dce3"},
+    {file = "yarl-1.25.1-cp315-cp315t-musllinux_1_2_armv7l.whl", hash = "sha256:85a18376073f8a39aa07be34f9fc77e2869aa72c55c441efdd2cf79a0407504d"},
+    {file = "yarl-1.25.1-cp315-cp315t-musllinux_1_2_ppc64le.whl", hash = "sha256:77716e245c90f058466a05e6a465bb8600f767a8f4b18b4d40f3aff958e5f73c"},
+    {file = "yarl-1.25.1-cp315-cp315t-musllinux_1_2_riscv64.whl", hash = "sha256:1e80dcf1446e1b080b1932b0d103c464a04112f5bc31f0f983ad418172063cde"},
+    {file = "yarl-1.25.1-cp315-cp315t-musllinux_1_2_s390x.whl", hash = "sha256:bdc8d8b8c22e9e43ac68316b5e6cf083dec537f4ec213cb4aa967b583bc3fa64"},
+    {file = "yarl-1.25.1-cp315-cp315t-musllinux_1_2_x86_64.whl", hash = "sha256:dfbf531053a0935f2e871bcd4753f90313688772ff8c017f5ea402e315a78c1f"},
+    {file = "yarl-1.25.1-cp315-cp315t-win_amd64.whl", hash = "sha256:b13b88747769537f3d32e89e3a735da10c0a9e35d7322928c701b5f93d3afffd"},
+    {file = "yarl-1.25.1-cp315-cp315t-win_arm64.whl", hash = "sha256:783dd1467083f4d3f7722ad6a313f24c173e7571372738fcb7a6e6d1ba48df25"},
+    {file = "yarl-1.25.1-py3-none-any.whl", hash = "sha256:681c758b0490f9e96b78e5fa8e8dc6e648e9185bb6eaebe73183c33ea0c445f3"},
+    {file = "yarl-1.25.1.tar.gz", hash = "sha256:03dd38de09bc213e9a8b29761eec33ee1d5318dac0e49d8af36e4d27830e23a7"},
+]
+
+[package.dependencies]
+idna = ">=2.0"
+multidict = ">=4.0"
+propcache = ">=0.2.1"
+
+[metadata]
+lock-version = "2.1"
+python-versions = ">=3.12,<3.14"
+content-hash = "be2aa86ed6b2404dde1d0e583e2c67c4c3b5b4764839ce4c3f03273177e0ad3f"
diff --git a/paddler_kev_converter_python/pyproject.toml b/paddler_kev_converter_python/pyproject.toml
new file mode 100644
index 000000000..4623105be
--- /dev/null
+++ b/paddler_kev_converter_python/pyproject.toml
@@ -0,0 +1,55 @@
+[tool.poetry]
+name = "paddler-kev-converter"
+version = "5.0.0"
+description = "Converts kev decision checkpoints into the files Paddler serves, and writes the reference fixtures Paddler's decision tests use"
+authors = ["Intentee"]
+license = "Apache-2.0"
+packages = [{include = "paddler_kev_converter"}]
+
+[tool.poetry.scripts]
+paddler-kev-converter = "paddler_kev_converter.command_line:main"
+
+[tool.poetry.dependencies]
+python = ">=3.12,<3.14"
+gguf = { git = "https://github.com/ggml-org/llama.cpp.git", rev = "3f03b17724ad8b69040825f45c43c59b092a0244", subdirectory = "gguf-py" }
+huggingface-hub = "==2.3.0"
+kev = { git = "https://github.com/jaredpalmer/kev.git", rev = "2074cc44256200bb37540103fa4572901faa10f3" }
+numpy = "==2.5.4"
+peft = "==0.21.2"
+torch = [
+    { version = "==2.8.0+cpu", source = "pytorch-cpu", markers = "sys_platform == 'linux'" },
+    { version = "==2.8.0", source = "PyPI", markers = "sys_platform == 'darwin'" },
+]
+transformers = "==5.20.2"
+
+[tool.poetry.group.dev.dependencies]
+pytest = "9.1.1"
+pytest-cov = "7.1.0"
+mypy = "2.4.0"
+
+[[tool.poetry.source]]
+name = "pytorch-cpu"
+url = "https://download.pytorch.org/whl/cpu"
+priority = "explicit"
+
+[build-system]
+requires = ["poetry-core"]
+build-backend = "poetry.core.masonry.api"
+
+[tool.mypy]
+python_version = "3.12"
+strict = true
+extra_checks = true
+mypy_path = "stubs"
+
+[tool.coverage.report]
+fail_under = 100
+show_missing = true
+
+[tool.coverage.run]
+branch = true
+source = ["paddler_kev_converter", "tests"]
+
+[tool.pytest.ini_options]
+addopts = "--cov-report=term-missing"
+tmp_path_retention_policy = "failed"
diff --git a/paddler_client_python/paddler_client/py.typed b/paddler_kev_converter_python/stubs/kev/__init__.pyi
similarity index 100%
rename from paddler_client_python/paddler_client/py.typed
rename to paddler_kev_converter_python/stubs/kev/__init__.pyi
diff --git a/paddler_kev_converter_python/stubs/kev/api.pyi b/paddler_kev_converter_python/stubs/kev/api.pyi
new file mode 100644
index 000000000..ebffb27b7
--- /dev/null
+++ b/paddler_kev_converter_python/stubs/kev/api.pyi
@@ -0,0 +1,12 @@
+from typing import Any
+
+from pydantic import BaseModel
+
+class SystemOneRequest(BaseModel): ...
+
+def to_record(
+    req: SystemOneRequest,
+) -> tuple[dict[str, Any], list[dict[str, Any]]]: ...
+def to_answers(
+    probs: list[list[float]], meta: list[dict[str, Any]]
+) -> dict[str, Any]: ...
diff --git a/paddler_kev_converter_python/stubs/kev/checkpoint.pyi b/paddler_kev_converter_python/stubs/kev/checkpoint.pyi
new file mode 100644
index 000000000..0001857f0
--- /dev/null
+++ b/paddler_kev_converter_python/stubs/kev/checkpoint.pyi
@@ -0,0 +1,28 @@
+import torch
+from kev.model import DecisionModel
+from transformers import PreTrainedTokenizerBase
+
+class Meta:
+    base: str
+    base_revision: str | None
+    head: dict[str, torch.Tensor]
+    option_isolation: bool
+    temperature: float
+    weights: str
+    weights_dtype: str
+
+class LoadOptions:
+    def __init__(self, *, dtype: torch.dtype | None = None) -> None: ...
+
+def read_meta(run: str) -> Meta: ...
+def write_meta(run: str, meta: Meta) -> None: ...
+
+class Checkpoint:
+    meta: Meta
+    path: str
+    def __init__(self, run: str) -> None: ...
+    @property
+    def full(self) -> bool: ...
+    def load(
+        self, device: str, opts: LoadOptions = ...
+    ) -> tuple[PreTrainedTokenizerBase, DecisionModel]: ...
diff --git a/paddler_kev_converter_python/stubs/kev/model.pyi b/paddler_kev_converter_python/stubs/kev/model.pyi
new file mode 100644
index 000000000..04c61a44e
--- /dev/null
+++ b/paddler_kev_converter_python/stubs/kev/model.pyi
@@ -0,0 +1,18 @@
+from typing import Any
+
+import torch
+from transformers import PreTrainedModel, PreTrainedTokenizerBase
+
+SPECIAL: list[str]
+
+def load_tokenizer(
+    name: str, revision: str | None = None
+) -> PreTrainedTokenizerBase: ...
+def user_tokens(tok: PreTrainedTokenizerBase, text: str) -> list[int]: ...
+
+class DecisionModel:
+    lm: PreTrainedModel
+    def encode(
+        self, tok: PreTrainedTokenizerBase, rec: dict[str, Any], *, strict: bool
+    ) -> dict[str, Any]: ...
+    def probs(self, enc: dict[str, Any]) -> list[torch.Tensor]: ...
diff --git a/paddler_client_python/tests/cluster/__init__.py b/paddler_kev_converter_python/tests/__init__.py
similarity index 100%
rename from paddler_client_python/tests/cluster/__init__.py
rename to paddler_kev_converter_python/tests/__init__.py
diff --git a/paddler_kev_converter_python/tests/conftest.py b/paddler_kev_converter_python/tests/conftest.py
new file mode 100644
index 000000000..eeceeaaed
--- /dev/null
+++ b/paddler_kev_converter_python/tests/conftest.py
@@ -0,0 +1,101 @@
+from pathlib import Path
+from shutil import copyfile
+
+import pytest
+import torch
+from kev.checkpoint import Checkpoint, LoadOptions, read_meta, write_meta
+from transformers import AutoModel
+
+from paddler_kev_converter.command_line import main
+from tests.kev_checkpoint import KEV_CHECKPOINT
+
+ATTENTION_ONLY_BASE = "Qwen/Qwen3-0.6B-Base"
+ATTENTION_ONLY_BASE_REVISION = "da87bfb608c14b7cf20ba1ce41287e8de496c0cd"
+
+
+@pytest.fixture(scope="session")
+def kev_checkpoint() -> Checkpoint:
+    return Checkpoint(KEV_CHECKPOINT)
+
+
+@pytest.fixture(scope="session")
+def lora_kev_run() -> str:
+    return KEV_CHECKPOINT
+
+
+@pytest.fixture(scope="session")
+def full_weight_kev_run(
+    kev_checkpoint: Checkpoint, tmp_path_factory: pytest.TempPathFactory
+) -> str:
+    full_weight_directory = tmp_path_factory.mktemp("full_weight")
+    _, decision_model = kev_checkpoint.load("cpu", LoadOptions(dtype=torch.bfloat16))
+    backbone = decision_model.lm
+    backbone.config.tie_word_embeddings = False
+    backbone.save_pretrained(full_weight_directory)
+    meta = read_meta(kev_checkpoint.path)
+    meta.weights = "full"
+    meta.weights_dtype = "bf16"
+    write_meta(str(full_weight_directory), meta)
+
+    return str(full_weight_directory)
+
+
+@pytest.fixture(scope="session")
+def attention_only_full_weight_kev_run(
+    kev_checkpoint: Checkpoint, tmp_path_factory: pytest.TempPathFactory
+) -> str:
+    attention_only_directory = tmp_path_factory.mktemp("attention_only_full_weight")
+    AutoModel.from_pretrained(
+        ATTENTION_ONLY_BASE, revision=ATTENTION_ONLY_BASE_REVISION, dtype=torch.bfloat16
+    ).save_pretrained(attention_only_directory)
+    meta = read_meta(kev_checkpoint.path)
+    meta.base = ATTENTION_ONLY_BASE
+    meta.base_revision = ATTENTION_ONLY_BASE_REVISION
+    meta.weights = "full"
+    meta.weights_dtype = "bf16"
+    write_meta(str(attention_only_directory), meta)
+
+    return str(attention_only_directory)
+
+
+@pytest.fixture(
+    scope="session",
+    params=["lora_kev_run", "full_weight_kev_run", "attention_only_full_weight_kev_run"],
+)
+def convertible_kev_run(request: pytest.FixtureRequest) -> str:
+    run: str = request.getfixturevalue(request.param)
+
+    return run
+
+
+@pytest.fixture(scope="session")
+def converted_backbone_directory(
+    convertible_kev_run: str, tmp_path_factory: pytest.TempPathFactory
+) -> Path:
+    conversion_directory = tmp_path_factory.mktemp("converted")
+    main(
+        [
+            "convert",
+            "--checkpoint",
+            convertible_kev_run,
+            "--backbone-directory",
+            str(conversion_directory / "backbone"),
+            "--pointer-head",
+            str(conversion_directory / "pointer_head.gguf"),
+        ]
+    )
+
+    return conversion_directory / "backbone"
+
+
+@pytest.fixture(scope="session")
+def converted_pointer_head(converted_backbone_directory: Path) -> Path:
+    return converted_backbone_directory.parent / "pointer_head.gguf"
+
+
+@pytest.fixture
+def kev_checkpoint_copy(kev_checkpoint: Checkpoint, tmp_path: Path) -> Path:
+    for checkpoint_file in Path(kev_checkpoint.path).iterdir():
+        copyfile(checkpoint_file, tmp_path / checkpoint_file.name)
+
+    return tmp_path
diff --git a/paddler_kev_converter_python/tests/kev_checkpoint.py b/paddler_kev_converter_python/tests/kev_checkpoint.py
new file mode 100644
index 000000000..bbe25b330
--- /dev/null
+++ b/paddler_kev_converter_python/tests/kev_checkpoint.py
@@ -0,0 +1,3 @@
+from os import environ
+
+KEV_CHECKPOINT = environ["PADDLER_KEV_CHECKPOINT"]
diff --git a/paddler_kev_converter_python/tests/rewrite_adapter_config.py b/paddler_kev_converter_python/tests/rewrite_adapter_config.py
new file mode 100644
index 000000000..3a214e72b
--- /dev/null
+++ b/paddler_kev_converter_python/tests/rewrite_adapter_config.py
@@ -0,0 +1,10 @@
+import json
+from pathlib import Path
+
+
+def rewrite_adapter_config(checkpoint_directory: Path, **changes: object) -> None:
+    adapter_config_path = checkpoint_directory / "adapter_config.json"
+    adapter_config = json.loads(adapter_config_path.read_text(encoding="utf-8"))
+    adapter_config_path.write_text(
+        json.dumps({**adapter_config, **changes}), encoding="utf-8"
+    )
diff --git a/paddler_kev_converter_python/tests/tensor_shapes.py b/paddler_kev_converter_python/tests/tensor_shapes.py
new file mode 100644
index 000000000..11535bf26
--- /dev/null
+++ b/paddler_kev_converter_python/tests/tensor_shapes.py
@@ -0,0 +1,8 @@
+from gguf import GGUFReader
+
+
+def tensor_shapes(reader: GGUFReader) -> dict[str, tuple[int, ...]]:
+    return {
+        tensor.name: tuple(int(dimension) for dimension in tensor.shape)
+        for tensor in reader.tensors
+    }
diff --git a/paddler_kev_converter_python/tests/test_a_lora_checkpoint_trained_on_a_bf16_backbone_is_refused.py b/paddler_kev_converter_python/tests/test_a_lora_checkpoint_trained_on_a_bf16_backbone_is_refused.py
new file mode 100644
index 000000000..1d3356d15
--- /dev/null
+++ b/paddler_kev_converter_python/tests/test_a_lora_checkpoint_trained_on_a_bf16_backbone_is_refused.py
@@ -0,0 +1,20 @@
+from pathlib import Path
+
+import pytest
+from kev.checkpoint import read_meta, write_meta
+
+from paddler_kev_converter.convertible_kev_checkpoint import ConvertibleKevCheckpoint
+from paddler_kev_converter.error import BackboneDtypeUnsupportedError
+
+
+def test_a_lora_checkpoint_trained_on_a_bf16_backbone_is_refused(
+    kev_checkpoint_copy: Path,
+) -> None:
+    meta = read_meta(str(kev_checkpoint_copy))
+    meta.weights_dtype = "bf16"
+    write_meta(str(kev_checkpoint_copy), meta)
+
+    with pytest.raises(BackboneDtypeUnsupportedError) as refusal:
+        ConvertibleKevCheckpoint.open(str(kev_checkpoint_copy))
+
+    assert refusal.value.weights_dtype == "bf16"
diff --git a/paddler_kev_converter_python/tests/test_a_lora_checkpoint_training_token_embeddings_is_refused.py b/paddler_kev_converter_python/tests/test_a_lora_checkpoint_training_token_embeddings_is_refused.py
new file mode 100644
index 000000000..2d45f73d7
--- /dev/null
+++ b/paddler_kev_converter_python/tests/test_a_lora_checkpoint_training_token_embeddings_is_refused.py
@@ -0,0 +1,18 @@
+from pathlib import Path
+
+import pytest
+
+from paddler_kev_converter.convertible_kev_checkpoint import ConvertibleKevCheckpoint
+from paddler_kev_converter.error import TrainedTokenEmbeddingsUnsupportedError
+from tests.rewrite_adapter_config import rewrite_adapter_config
+
+
+def test_a_lora_checkpoint_training_token_embeddings_is_refused(
+    kev_checkpoint_copy: Path,
+) -> None:
+    rewrite_adapter_config(
+        kev_checkpoint_copy, trainable_token_indices={"embed_tokens": [248044]}
+    )
+
+    with pytest.raises(TrainedTokenEmbeddingsUnsupportedError):
+        ConvertibleKevCheckpoint.open(str(kev_checkpoint_copy))
diff --git a/paddler_kev_converter_python/tests/test_a_malformed_pointer_head_with_an_oversized_head_dim_exceeds_sixteen_bits.py b/paddler_kev_converter_python/tests/test_a_malformed_pointer_head_with_an_oversized_head_dim_exceeds_sixteen_bits.py
new file mode 100644
index 000000000..f7a54e17e
--- /dev/null
+++ b/paddler_kev_converter_python/tests/test_a_malformed_pointer_head_with_an_oversized_head_dim_exceeds_sixteen_bits.py
@@ -0,0 +1,14 @@
+from pathlib import Path
+
+from tests.tensor_shapes import tensor_shapes
+from tests.written_malformed_pointer_head import written_malformed_pointer_head
+
+
+def test_a_malformed_pointer_head_with_an_oversized_head_dim_exceeds_sixteen_bits(
+    tmp_path: Path,
+) -> None:
+    shapes = tensor_shapes(
+        written_malformed_pointer_head(tmp_path, "oversized-head-dim")
+    )
+
+    assert shapes["pointer_head.key.bias"] == (2**16,)
diff --git a/paddler_kev_converter_python/tests/test_a_malformed_pointer_head_with_f16_tensors_stores_every_tensor_as_f16.py b/paddler_kev_converter_python/tests/test_a_malformed_pointer_head_with_f16_tensors_stores_every_tensor_as_f16.py
new file mode 100644
index 000000000..40fcbe2af
--- /dev/null
+++ b/paddler_kev_converter_python/tests/test_a_malformed_pointer_head_with_f16_tensors_stores_every_tensor_as_f16.py
@@ -0,0 +1,14 @@
+from pathlib import Path
+
+from gguf import GGMLQuantizationType
+
+from tests.written_malformed_pointer_head import written_malformed_pointer_head
+
+
+def test_a_malformed_pointer_head_with_f16_tensors_stores_every_tensor_as_f16(
+    tmp_path: Path,
+) -> None:
+    assert {
+        tensor.tensor_type
+        for tensor in written_malformed_pointer_head(tmp_path, "f16-tensors").tensors
+    } == {GGMLQuantizationType.F16}
diff --git a/paddler_kev_converter_python/tests/test_a_malformed_pointer_head_with_mismatched_projections_shapes_query_unlike_key.py b/paddler_kev_converter_python/tests/test_a_malformed_pointer_head_with_mismatched_projections_shapes_query_unlike_key.py
new file mode 100644
index 000000000..50a8e8ba7
--- /dev/null
+++ b/paddler_kev_converter_python/tests/test_a_malformed_pointer_head_with_mismatched_projections_shapes_query_unlike_key.py
@@ -0,0 +1,14 @@
+from pathlib import Path
+
+from tests.tensor_shapes import tensor_shapes
+from tests.written_malformed_pointer_head import written_malformed_pointer_head
+
+
+def test_a_malformed_pointer_head_with_mismatched_projections_shapes_query_unlike_key(
+    tmp_path: Path,
+) -> None:
+    shapes = tensor_shapes(
+        written_malformed_pointer_head(tmp_path, "mismatched-projections")
+    )
+
+    assert shapes["pointer_head.query.weight"] != shapes["pointer_head.key.weight"]
diff --git a/paddler_kev_converter_python/tests/test_a_malformed_pointer_head_with_plain_text_delimiters_names_no_control_tokens.py b/paddler_kev_converter_python/tests/test_a_malformed_pointer_head_with_plain_text_delimiters_names_no_control_tokens.py
new file mode 100644
index 000000000..c89d9f6c1
--- /dev/null
+++ b/paddler_kev_converter_python/tests/test_a_malformed_pointer_head_with_plain_text_delimiters_names_no_control_tokens.py
@@ -0,0 +1,11 @@
+from pathlib import Path
+
+from tests.written_malformed_pointer_head import written_malformed_pointer_head
+
+
+def test_a_malformed_pointer_head_with_plain_text_delimiters_names_no_control_tokens(
+    tmp_path: Path,
+) -> None:
+    reader = written_malformed_pointer_head(tmp_path, "plain-text-delimiters")
+
+    assert reader.fields["pointer_head.delimiter.state"].contents() == "state"
diff --git a/paddler_kev_converter_python/tests/test_a_malformed_pointer_head_with_zero_temperature_stores_a_zero_temperature.py b/paddler_kev_converter_python/tests/test_a_malformed_pointer_head_with_zero_temperature_stores_a_zero_temperature.py
new file mode 100644
index 000000000..6512fe3c9
--- /dev/null
+++ b/paddler_kev_converter_python/tests/test_a_malformed_pointer_head_with_zero_temperature_stores_a_zero_temperature.py
@@ -0,0 +1,11 @@
+from pathlib import Path
+
+from tests.written_malformed_pointer_head import written_malformed_pointer_head
+
+
+def test_a_malformed_pointer_head_with_zero_temperature_stores_a_zero_temperature(
+    tmp_path: Path,
+) -> None:
+    reader = written_malformed_pointer_head(tmp_path, "zero-temperature")
+
+    assert reader.fields["pointer_head.temperature"].contents() == 0.0
diff --git a/paddler_kev_converter_python/tests/test_a_malformed_pointer_head_without_its_delimiters_keeps_its_temperature.py b/paddler_kev_converter_python/tests/test_a_malformed_pointer_head_without_its_delimiters_keeps_its_temperature.py
new file mode 100644
index 000000000..1a8facb35
--- /dev/null
+++ b/paddler_kev_converter_python/tests/test_a_malformed_pointer_head_without_its_delimiters_keeps_its_temperature.py
@@ -0,0 +1,12 @@
+from pathlib import Path
+
+from tests.written_malformed_pointer_head import written_malformed_pointer_head
+
+
+def test_a_malformed_pointer_head_without_its_delimiters_keeps_its_temperature(
+    tmp_path: Path,
+) -> None:
+    reader = written_malformed_pointer_head(tmp_path, "missing-delimiters")
+
+    assert "pointer_head.delimiter.state" not in reader.fields
+    assert "pointer_head.temperature" in reader.fields
diff --git a/paddler_kev_converter_python/tests/test_a_malformed_pointer_head_without_its_input_normalization_keeps_its_temperature.py b/paddler_kev_converter_python/tests/test_a_malformed_pointer_head_without_its_input_normalization_keeps_its_temperature.py
new file mode 100644
index 000000000..fa6eefaf5
--- /dev/null
+++ b/paddler_kev_converter_python/tests/test_a_malformed_pointer_head_without_its_input_normalization_keeps_its_temperature.py
@@ -0,0 +1,12 @@
+from pathlib import Path
+
+from tests.written_malformed_pointer_head import written_malformed_pointer_head
+
+
+def test_a_malformed_pointer_head_without_its_input_normalization_keeps_its_temperature(
+    tmp_path: Path,
+) -> None:
+    reader = written_malformed_pointer_head(tmp_path, "missing-input-normalization")
+
+    assert "pointer_head.input_normalization" not in reader.fields
+    assert "pointer_head.temperature" in reader.fields
diff --git a/paddler_kev_converter_python/tests/test_a_malformed_pointer_head_without_its_temperature_keeps_its_delimiters.py b/paddler_kev_converter_python/tests/test_a_malformed_pointer_head_without_its_temperature_keeps_its_delimiters.py
new file mode 100644
index 000000000..41391d8d6
--- /dev/null
+++ b/paddler_kev_converter_python/tests/test_a_malformed_pointer_head_without_its_temperature_keeps_its_delimiters.py
@@ -0,0 +1,12 @@
+from pathlib import Path
+
+from tests.written_malformed_pointer_head import written_malformed_pointer_head
+
+
+def test_a_malformed_pointer_head_without_its_temperature_keeps_its_delimiters(
+    tmp_path: Path,
+) -> None:
+    reader = written_malformed_pointer_head(tmp_path, "missing-temperature")
+
+    assert "pointer_head.temperature" not in reader.fields
+    assert "pointer_head.delimiter.state" in reader.fields
diff --git a/paddler_kev_converter_python/tests/test_a_slow_tokenizer_cannot_declare_its_normalization.py b/paddler_kev_converter_python/tests/test_a_slow_tokenizer_cannot_declare_its_normalization.py
new file mode 100644
index 000000000..eb638ce30
--- /dev/null
+++ b/paddler_kev_converter_python/tests/test_a_slow_tokenizer_cannot_declare_its_normalization.py
@@ -0,0 +1,14 @@
+import pytest
+from transformers import ByT5Tokenizer
+
+from paddler_kev_converter.error import TokenizerNormalizerUninspectableError
+from paddler_kev_converter.tokenizer_input_normalization import (
+    tokenizer_input_normalization,
+)
+
+
+def test_a_slow_tokenizer_cannot_declare_its_normalization() -> None:
+    with pytest.raises(TokenizerNormalizerUninspectableError) as refusal:
+        tokenizer_input_normalization(ByT5Tokenizer())
+
+    assert refusal.value.tokenizer_class == "ByT5Tokenizer"
diff --git a/paddler_kev_converter_python/tests/test_a_tokenizer_declares_the_unicode_normalization_it_applies.py b/paddler_kev_converter_python/tests/test_a_tokenizer_declares_the_unicode_normalization_it_applies.py
new file mode 100644
index 000000000..8a43510f4
--- /dev/null
+++ b/paddler_kev_converter_python/tests/test_a_tokenizer_declares_the_unicode_normalization_it_applies.py
@@ -0,0 +1,28 @@
+from pathlib import Path
+
+from tokenizers.normalizers import NFC, NFD, NFKC, NFKD, Normalizer
+
+from paddler_kev_converter.input_normalization import InputNormalization
+from paddler_kev_converter.tokenizer_input_normalization import (
+    tokenizer_input_normalization,
+)
+from tests.tokenizer_normalizing_with import tokenizer_normalizing_with
+
+
+def test_a_tokenizer_declares_the_unicode_normalization_it_applies(
+    tmp_path: Path,
+) -> None:
+    normalizers: dict[InputNormalization, Normalizer | None] = {
+        InputNormalization.NFC: NFC(),
+        InputNormalization.NFD: NFD(),
+        InputNormalization.NFKC: NFKC(),
+        InputNormalization.NFKD: NFKD(),
+        InputNormalization.NONE: None,
+    }
+
+    assert {
+        input_normalization: tokenizer_input_normalization(
+            tokenizer_normalizing_with(normalizer, tmp_path)
+        )
+        for input_normalization, normalizer in normalizers.items()
+    } == {input_normalization: input_normalization for input_normalization in normalizers}
diff --git a/paddler_kev_converter_python/tests/test_a_tokenizer_normalizing_beyond_a_unicode_form_is_refused.py b/paddler_kev_converter_python/tests/test_a_tokenizer_normalizing_beyond_a_unicode_form_is_refused.py
new file mode 100644
index 000000000..d66cd341a
--- /dev/null
+++ b/paddler_kev_converter_python/tests/test_a_tokenizer_normalizing_beyond_a_unicode_form_is_refused.py
@@ -0,0 +1,21 @@
+from pathlib import Path
+
+import pytest
+from tokenizers.normalizers import NFC, Lowercase, Sequence
+
+from paddler_kev_converter.error import TokenizerNormalizerUnsupportedError
+from paddler_kev_converter.tokenizer_input_normalization import (
+    tokenizer_input_normalization,
+)
+from tests.tokenizer_normalizing_with import tokenizer_normalizing_with
+
+
+def test_a_tokenizer_normalizing_beyond_a_unicode_form_is_refused(
+    tmp_path: Path,
+) -> None:
+    with pytest.raises(TokenizerNormalizerUnsupportedError) as refusal:
+        tokenizer_input_normalization(
+            tokenizer_normalizing_with(Sequence([NFC(), Lowercase()]), tmp_path)
+        )
+
+    assert refusal.value.normalizer == "Sequence"
diff --git a/paddler_kev_converter_python/tests/test_an_option_isolating_checkpoint_is_refused.py b/paddler_kev_converter_python/tests/test_an_option_isolating_checkpoint_is_refused.py
new file mode 100644
index 000000000..1542cbaaf
--- /dev/null
+++ b/paddler_kev_converter_python/tests/test_an_option_isolating_checkpoint_is_refused.py
@@ -0,0 +1,16 @@
+from pathlib import Path
+
+import pytest
+from kev.checkpoint import read_meta, write_meta
+
+from paddler_kev_converter.convertible_kev_checkpoint import ConvertibleKevCheckpoint
+from paddler_kev_converter.error import OptionIsolationUnsupportedError
+
+
+def test_an_option_isolating_checkpoint_is_refused(kev_checkpoint_copy: Path) -> None:
+    meta = read_meta(str(kev_checkpoint_copy))
+    meta.option_isolation = True
+    write_meta(str(kev_checkpoint_copy), meta)
+
+    with pytest.raises(OptionIsolationUnsupportedError):
+        ConvertibleKevCheckpoint.open(str(kev_checkpoint_copy))
diff --git a/paddler_kev_converter_python/tests/test_converting_saves_the_base_tokenizer.py b/paddler_kev_converter_python/tests/test_converting_saves_the_base_tokenizer.py
new file mode 100644
index 000000000..863e12535
--- /dev/null
+++ b/paddler_kev_converter_python/tests/test_converting_saves_the_base_tokenizer.py
@@ -0,0 +1,18 @@
+from pathlib import Path
+
+from kev.checkpoint import Checkpoint
+from kev.model import SPECIAL, load_tokenizer
+from transformers import AutoTokenizer
+
+TOKENIZED_TEXT = "".join(SPECIAL) + " The café served crème brûlée at 9:30."
+
+
+def test_converting_saves_the_base_tokenizer(
+    converted_backbone_directory: Path, convertible_kev_run: str
+) -> None:
+    meta = Checkpoint(convertible_kev_run).meta
+
+    assert (
+        AutoTokenizer.from_pretrained(converted_backbone_directory)(TOKENIZED_TEXT).input_ids
+        == load_tokenizer(meta.base, revision=meta.base_revision)(TOKENIZED_TEXT).input_ids
+    )
diff --git a/paddler_kev_converter_python/tests/test_converting_writes_the_decision_backbone_as_a_tied_causal_language_model.py b/paddler_kev_converter_python/tests/test_converting_writes_the_decision_backbone_as_a_tied_causal_language_model.py
new file mode 100644
index 000000000..f85f30efa
--- /dev/null
+++ b/paddler_kev_converter_python/tests/test_converting_writes_the_decision_backbone_as_a_tied_causal_language_model.py
@@ -0,0 +1,23 @@
+from pathlib import Path
+
+import torch
+from kev.checkpoint import Checkpoint, LoadOptions
+from transformers import AutoModelForCausalLM
+
+
+def test_converting_writes_the_decision_backbone_as_a_tied_causal_language_model(
+    converted_backbone_directory: Path, convertible_kev_run: str
+) -> None:
+    _, decision_model = Checkpoint(convertible_kev_run).load("cpu", LoadOptions())
+    decision_backbone = decision_model.lm.state_dict()
+    written_causal_language_model = AutoModelForCausalLM.from_pretrained(
+        converted_backbone_directory, dtype="auto"
+    )
+    written_backbone = written_causal_language_model.base_model.state_dict()
+
+    assert written_backbone.keys() == decision_backbone.keys()
+    assert all(
+        torch.equal(written_backbone[name], decision_tensor)
+        for name, decision_tensor in decision_backbone.items()
+    )
+    assert written_causal_language_model.config.tie_word_embeddings
diff --git a/paddler_kev_converter_python/tests/test_converting_writes_the_pointer_head_of_the_checkpoint.py b/paddler_kev_converter_python/tests/test_converting_writes_the_pointer_head_of_the_checkpoint.py
new file mode 100644
index 000000000..e2d535da1
--- /dev/null
+++ b/paddler_kev_converter_python/tests/test_converting_writes_the_pointer_head_of_the_checkpoint.py
@@ -0,0 +1,36 @@
+from pathlib import Path
+
+import numpy as np
+from gguf import GGUFReader
+from kev.checkpoint import Checkpoint
+from kev.model import SPECIAL
+
+
+def test_converting_writes_the_pointer_head_of_the_checkpoint(
+    converted_pointer_head: Path, kev_checkpoint: Checkpoint
+) -> None:
+    reader = GGUFReader(converted_pointer_head)
+    tensors = {tensor.name: tensor.data for tensor in reader.tensors}
+    head = kev_checkpoint.meta.head
+
+    assert reader.fields["general.architecture"].contents() == "pointer_head"
+    assert reader.fields["pointer_head.input_normalization"].contents() == "NFC"
+    assert reader.fields["pointer_head.temperature"].contents() == np.float32(
+        kev_checkpoint.meta.temperature
+    )
+    assert [
+        reader.fields[f"pointer_head.delimiter.{delimiter_name}"].contents()
+        for delimiter_name in [
+            "state",
+            "question",
+            "option_start",
+            "option_end",
+            "decide",
+        ]
+    ] == SPECIAL
+    assert np.array_equal(tensors["pointer_head.key.bias"], head["k.bias"].numpy())
+    assert np.array_equal(tensors["pointer_head.key.weight"], head["k.weight"].numpy())
+    assert np.array_equal(tensors["pointer_head.query.bias"], head["q.bias"].numpy())
+    assert np.array_equal(
+        tensors["pointer_head.query.weight"], head["q.weight"].numpy()
+    )
diff --git a/paddler_kev_converter_python/tests/test_synthetic_pointer_heads_are_reproducible.py b/paddler_kev_converter_python/tests/test_synthetic_pointer_heads_are_reproducible.py
new file mode 100644
index 000000000..bbb13e49b
--- /dev/null
+++ b/paddler_kev_converter_python/tests/test_synthetic_pointer_heads_are_reproducible.py
@@ -0,0 +1,28 @@
+from pathlib import Path
+
+from gguf import GGUFReader
+
+from paddler_kev_converter.command_line import main
+from tests.tensor_shapes import tensor_shapes
+
+
+def test_synthetic_pointer_heads_are_reproducible(tmp_path: Path) -> None:
+    for written_head in ["first.gguf", "second.gguf"]:
+        main(
+            [
+                "synthetic-pointer-head",
+                "--hidden-size",
+                "512",
+                "--input-normalization",
+                "NFC",
+                "--output",
+                str(tmp_path / written_head),
+            ]
+        )
+
+    assert (tmp_path / "first.gguf").read_bytes() == (
+        tmp_path / "second.gguf"
+    ).read_bytes()
+    assert tensor_shapes(GGUFReader(tmp_path / "first.gguf"))[
+        "pointer_head.key.weight"
+    ] == (512, 16)
diff --git a/paddler_kev_converter_python/tests/test_the_parity_reference_pairs_rendered_records_with_kev_probabilities.py b/paddler_kev_converter_python/tests/test_the_parity_reference_pairs_rendered_records_with_kev_probabilities.py
new file mode 100644
index 000000000..1a92d3b5a
--- /dev/null
+++ b/paddler_kev_converter_python/tests/test_the_parity_reference_pairs_rendered_records_with_kev_probabilities.py
@@ -0,0 +1,43 @@
+import json
+from math import isclose
+from pathlib import Path
+
+from paddler_kev_converter.command_line import main
+from tests.kev_checkpoint import KEV_CHECKPOINT
+
+
+def test_the_parity_reference_pairs_rendered_records_with_kev_probabilities(
+    tmp_path: Path,
+) -> None:
+    main(
+        [
+            "parity-reference",
+            "--checkpoint",
+            KEV_CHECKPOINT,
+            "--output",
+            str(tmp_path / "parity.json"),
+        ]
+    )
+    records = json.loads((tmp_path / "parity.json").read_text(encoding="utf-8"))
+
+    assert records[1]["state"] == "I was charged twice. Please help."
+    assert records[1]["questions"] == [
+        {
+            "id": "billing",
+            "instructions": "Is this about billing?",
+            "options": ["no", "yes"],
+        },
+        {
+            "id": "tone",
+            "instructions": "What is the tone?",
+            "options": ["calm", "angry"],
+        },
+    ]
+    assert all(
+        len(question_probabilities) == len(question["options"])
+        and isclose(sum(question_probabilities), 1.0, rel_tol=1e-5)
+        for record in records
+        for question, question_probabilities in zip(
+            record["questions"], record["probabilities"], strict=True
+        )
+    )
diff --git a/paddler_kev_converter_python/tests/test_the_tokenizer_reference_pairs_texts_with_kev_user_token_ids.py b/paddler_kev_converter_python/tests/test_the_tokenizer_reference_pairs_texts_with_kev_user_token_ids.py
new file mode 100644
index 000000000..08a981346
--- /dev/null
+++ b/paddler_kev_converter_python/tests/test_the_tokenizer_reference_pairs_texts_with_kev_user_token_ids.py
@@ -0,0 +1,36 @@
+import json
+from pathlib import Path
+
+from kev.checkpoint import Checkpoint
+from kev.model import load_tokenizer, user_tokens
+
+from paddler_kev_converter.command_line import main
+from paddler_kev_converter.tokenizer_reference import REFERENCE_TEXTS
+
+
+def test_the_tokenizer_reference_pairs_texts_with_kev_user_token_ids(
+    kev_checkpoint: Checkpoint, tmp_path: Path
+) -> None:
+    base = kev_checkpoint.meta.base
+    base_revision = kev_checkpoint.meta.base_revision
+    assert base_revision is not None
+    main(
+        [
+            "tokenizer-reference",
+            "--base",
+            base,
+            "--base-revision",
+            base_revision,
+            "--output",
+            str(tmp_path / "reference.json"),
+        ]
+    )
+    tokenizer = load_tokenizer(base, revision=base_revision)
+
+    assert json.loads((tmp_path / "reference.json").read_text(encoding="utf-8")) == {
+        "entries": [
+            {"text": text, "token_ids": user_tokens(tokenizer, text)}
+            for text in REFERENCE_TEXTS
+        ],
+        "input_normalization": "NFC",
+    }
diff --git a/paddler_kev_converter_python/tests/test_the_typesafe_reference_renders_and_answers_like_kev.py b/paddler_kev_converter_python/tests/test_the_typesafe_reference_renders_and_answers_like_kev.py
new file mode 100644
index 000000000..c0a2d5e29
--- /dev/null
+++ b/paddler_kev_converter_python/tests/test_the_typesafe_reference_renders_and_answers_like_kev.py
@@ -0,0 +1,27 @@
+import json
+from pathlib import Path
+
+from paddler_kev_converter.command_line import main
+
+
+def test_the_typesafe_reference_renders_and_answers_like_kev(tmp_path: Path) -> None:
+    main(["typesafe-reference", "--output", str(tmp_path / "reference.json")])
+    reference = json.loads((tmp_path / "reference.json").read_text(encoding="utf-8"))
+
+    assert reference[0] == {
+        "request": {
+            "model": "test",
+            "state": "The invoice was paid on time.",
+            "questions": {
+                "paid": {"type": "noul", "instructions": "Was the invoice paid?"},
+            },
+        },
+        "rendered": {
+            "state": "The invoice was paid on time.",
+            "questions": [
+                {"instructions": "Was the invoice paid?", "options": ["no", "yes"]}
+            ],
+        },
+        "probabilities": [[0.12345678, 0.87654322]],
+        "answers": {"paid": {"type": "noul", "noul": 0.8765}},
+    }
diff --git a/paddler_kev_converter_python/tests/tokenizer_normalizing_with.py b/paddler_kev_converter_python/tests/tokenizer_normalizing_with.py
new file mode 100644
index 000000000..3dcf91969
--- /dev/null
+++ b/paddler_kev_converter_python/tests/tokenizer_normalizing_with.py
@@ -0,0 +1,21 @@
+from json import dumps
+from pathlib import Path
+
+from tokenizers import Tokenizer, models
+from tokenizers.normalizers import Normalizer
+from transformers import AutoTokenizer, PreTrainedTokenizerBase
+
+TOKENIZER_CONFIG = {"tokenizer_class": "PreTrainedTokenizerFast"}
+
+
+def tokenizer_normalizing_with(
+    normalizer: Normalizer | None, directory: Path
+) -> PreTrainedTokenizerBase:
+    backend_tokenizer = Tokenizer(
+        models.WordLevel(vocab={"[UNK]": 0}, unk_token="[UNK]")
+    )
+    backend_tokenizer.normalizer = normalizer
+    backend_tokenizer.save(str(directory / "tokenizer.json"))
+    (directory / "tokenizer_config.json").write_text(dumps(TOKENIZER_CONFIG))
+
+    return AutoTokenizer.from_pretrained(directory)
diff --git a/paddler_kev_converter_python/tests/written_malformed_pointer_head.py b/paddler_kev_converter_python/tests/written_malformed_pointer_head.py
new file mode 100644
index 000000000..5f9758218
--- /dev/null
+++ b/paddler_kev_converter_python/tests/written_malformed_pointer_head.py
@@ -0,0 +1,23 @@
+from pathlib import Path
+
+from gguf import GGUFReader
+
+from paddler_kev_converter.command_line import main
+
+
+def written_malformed_pointer_head(tmp_path: Path, defect: str) -> GGUFReader:
+    main(
+        [
+            "malformed-pointer-head",
+            "--defect",
+            defect,
+            "--hidden-size",
+            "64",
+            "--input-normalization",
+            "NFC",
+            "--output",
+            str(tmp_path / f"{defect}.gguf"),
+        ]
+    )
+
+    return GGUFReader(tmp_path / f"{defect}.gguf")
diff --git a/paddler_llama_cpp_converter_python/.gitignore b/paddler_llama_cpp_converter_python/.gitignore
new file mode 100644
index 000000000..21d0b898f
--- /dev/null
+++ b/paddler_llama_cpp_converter_python/.gitignore
@@ -0,0 +1 @@
+.venv/
diff --git a/paddler_llama_cpp_converter_python/poetry.lock b/paddler_llama_cpp_converter_python/poetry.lock
new file mode 100644
index 000000000..f6e8de772
--- /dev/null
+++ b/paddler_llama_cpp_converter_python/poetry.lock
@@ -0,0 +1,1763 @@
+# This file is automatically @generated by Poetry 2.4.3 and should not be changed by hand.
+
+[[package]]
+name = "annotated-doc"
+version = "0.0.5"
+description = "Document parameters, class attributes, return types, and variables inline, with Annotated."
+optional = false
+python-versions = ">=3.9"
+groups = ["main"]
+files = [
+    {file = "annotated_doc-0.0.5-py3-none-any.whl", hash = "sha256:117bac03a25ede5df5440e855b32d556049ca169ead221505badf432fed4b101"},
+    {file = "annotated_doc-0.0.5.tar.gz", hash = "sha256:c7e58ce09192557605d8bbd92836d7e1d520ac9580096042c0bfd197efacf1bb"},
+]
+
+[[package]]
+name = "anyio"
+version = "4.15.1"
+description = "High-level concurrency and networking framework on top of asyncio or Trio"
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+markers = "sys_platform != \"emscripten\""
+files = [
+    {file = "anyio-4.15.1-py3-none-any.whl", hash = "sha256:6152fdbbf9a77fdec97731721bebf7c4c44f7c29b424b0065826173efc7ed101"},
+    {file = "anyio-4.15.1.tar.gz", hash = "sha256:9f28306018cbd6d329e64a36d58256edff76dd996fe423bc957326e578b82a94"},
+]
+
+[package.dependencies]
+idna = ">=2.8"
+typing_extensions = {version = ">=4.16.0", markers = "python_version < \"3.15\""}
+
+[package.extras]
+trio = ["trio (>=0.32.0)"]
+
+[[package]]
+name = "certifi"
+version = "2026.7.22"
+description = "Python package for providing Mozilla's CA Bundle."
+optional = false
+python-versions = ">=3.7"
+groups = ["main"]
+files = [
+    {file = "certifi-2026.7.22-py3-none-any.whl", hash = "sha256:62f22742b58a1a33014a2b6b706588a8d7e2a88ae7bd1a6ebe8c992928483775"},
+    {file = "certifi-2026.7.22.tar.gz", hash = "sha256:741e2c3b351ddf169a738da9f2c048608ff7f2c5cc02f1ebc6b118bb090d5d55"},
+]
+
+[[package]]
+name = "charset-normalizer"
+version = "3.5.2"
+description = "The Real First Universal Charset Detector. Open, modern and actively maintained alternative to Chardet."
+optional = false
+python-versions = ">=3.7"
+groups = ["main"]
+files = [
+    {file = "charset_normalizer-3.5.2-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:195c26fb65950f8fce54e26349852b7bdd7c5f120aeefbcc440b8a20faaed4a3"},
+    {file = "charset_normalizer-3.5.2-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:9373ad13ef0d2c0fb761e04e55bfdee5a08b52cef2c882c8fbe9935b1517152e"},
+    {file = "charset_normalizer-3.5.2-cp310-cp310-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:ddf19c062bea7a0cc80f519243d2c01dd091be0cf952a0750d4ad576709559f5"},
+    {file = "charset_normalizer-3.5.2-cp310-cp310-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:3d14b50de6bf4d0edf857a9386836846f982b8f524e188e2e68b96d702bcf4aa"},
+    {file = "charset_normalizer-3.5.2-cp310-cp310-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:28a15fdad492a99b6eccfaaed66ef3f74050680545ea61ec8b2f4c538f1f1320"},
+    {file = "charset_normalizer-3.5.2-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:8a893cc101149f80a653f82062ebc95b34525a2614382e1da5458fe7c6997249"},
+    {file = "charset_normalizer-3.5.2-cp310-cp310-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:619799369eeef6366ed3e8755a5670f4f2f0fb6b30a0fd7264dc0fdc2357058e"},
+    {file = "charset_normalizer-3.5.2-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:447441e76ec720b15e64418d32e092297340387053047c7c694f579efb0ee1d9"},
+    {file = "charset_normalizer-3.5.2-cp310-cp310-musllinux_1_2_armv7l.whl", hash = "sha256:62588a277bfb59def052abd940703fa35107152bf479781a878617d60faf8fb5"},
+    {file = "charset_normalizer-3.5.2-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:44bd4fbb29dfbeba60e7d2bd000c59e4b21ddb3cc53912b14048d37092706d7c"},
+    {file = "charset_normalizer-3.5.2-cp310-cp310-musllinux_1_2_riscv64.whl", hash = "sha256:30fcd120b732aa79317f08dee04d7de0847822e4cf7ee0e9f445bb958832252c"},
+    {file = "charset_normalizer-3.5.2-cp310-cp310-musllinux_1_2_s390x.whl", hash = "sha256:50e3adfb96fc189eb27b1cf62d3b598b89b4bb0420d93a3d3e42e137409011be"},
+    {file = "charset_normalizer-3.5.2-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:b736353c0a625bbd5fcec108576e2385db3496f4f771f785ff32e108d3c3bc45"},
+    {file = "charset_normalizer-3.5.2-cp310-cp310-win32.whl", hash = "sha256:f5833ad231be5eb6553de524a70f48d71b2c8563101750531e0b80184e175cd4"},
+    {file = "charset_normalizer-3.5.2-cp310-cp310-win_amd64.whl", hash = "sha256:1461ac396c4fdb983a675f20aa555624f0ee18ac83d832b9244ffff3d8055275"},
+    {file = "charset_normalizer-3.5.2-cp310-cp310-win_arm64.whl", hash = "sha256:c6708715abcf3c73b99508253e961a9967f02fe536532834149574eda6de0d1c"},
+    {file = "charset_normalizer-3.5.2-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:3d21b8b13c7592db2ac5e544a6d83187b995257472b0c9e8351b6d507ae37ed6"},
+    {file = "charset_normalizer-3.5.2-cp311-cp311-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:d760fe2a4d7c3b226cb9026d6a842868d52a7901bd98420e1baf14e80da85cf5"},
+    {file = "charset_normalizer-3.5.2-cp311-cp311-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:c9790464842f85f437dbbb54417eda1e0e6bfc52dd8d22d6fd1c994b73b2dc74"},
+    {file = "charset_normalizer-3.5.2-cp311-cp311-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:4685902cf26edf013ed7a3da0f426ebba7a00ebb9541386d835afbf002c11cab"},
+    {file = "charset_normalizer-3.5.2-cp311-cp311-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:4495c5002a7b28557e7e222e77e0b661183e432b7d6d2e788101e3f240e05b8c"},
+    {file = "charset_normalizer-3.5.2-cp311-cp311-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:211d5a3eb6af8f513b8d4ca19a8c1b7accab1b5f0d3175f9826b03c1a920dc1f"},
+    {file = "charset_normalizer-3.5.2-cp311-cp311-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:ef4fcbf3327382cd4c9f540babd61248208af7b93eec4de397b4d5f58a09e288"},
+    {file = "charset_normalizer-3.5.2-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:bd16aabe4a02a297c23417aa17ac6299dbd8c49f673bcd645b4929b11f5a4400"},
+    {file = "charset_normalizer-3.5.2-cp311-cp311-musllinux_1_2_armv7l.whl", hash = "sha256:fb9e68df06293761f9fe66ade60a9bc6d0f5e42b8acf2939a9158af86ab0e5bd"},
+    {file = "charset_normalizer-3.5.2-cp311-cp311-musllinux_1_2_ppc64le.whl", hash = "sha256:59f63901b0031c3136cf64704dcb21de0bbae62ce2c9529bc39d27665463de37"},
+    {file = "charset_normalizer-3.5.2-cp311-cp311-musllinux_1_2_riscv64.whl", hash = "sha256:304d5463e65a35d7bb0850550e0780395395f6fcf452f04db7d5ca7cecc425ac"},
+    {file = "charset_normalizer-3.5.2-cp311-cp311-musllinux_1_2_s390x.whl", hash = "sha256:9cf9b1a857e25c4baceeb3624e92a56df3668f398c4acba74e174d81fb4d1d3a"},
+    {file = "charset_normalizer-3.5.2-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:114e4d0c92d618409ed82a99e22b5c5e768fe995f2973f78265f4524f49d4640"},
+    {file = "charset_normalizer-3.5.2-cp311-cp311-win32.whl", hash = "sha256:2625388c6c754520c37abaf3b41eb34d1cc4a373f457898f08606c8e362b891d"},
+    {file = "charset_normalizer-3.5.2-cp311-cp311-win_amd64.whl", hash = "sha256:87e50a3e7cb90af586b6c5faf23e302a970415ac73bd7bd90a515a04b427ef96"},
+    {file = "charset_normalizer-3.5.2-cp311-cp311-win_arm64.whl", hash = "sha256:254eb48b9fa5ee9898a3c445825a1f340fe53712a098904b39b0bddba8ea3cb1"},
+    {file = "charset_normalizer-3.5.2-cp312-cp312-macosx_10_13_universal2.whl", hash = "sha256:ed2a239c0ea213acc1908150a3037257083c7c083128f1a4cec2ec4b97dca491"},
+    {file = "charset_normalizer-3.5.2-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:b91363207bd9dc966a691e959bb47f64b30f7ac4b072be9968b366982f7db77c"},
+    {file = "charset_normalizer-3.5.2-cp312-cp312-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:38a873987f3be698494da8b2e3085e29da02da7b633dce73e79c699a113d7bf0"},
+    {file = "charset_normalizer-3.5.2-cp312-cp312-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:355ad8011081dec5412240c087a9a0c9d4d5039f3ed11a3f13e18c2b29b56c51"},
+    {file = "charset_normalizer-3.5.2-cp312-cp312-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:ee21e28f0430bd6dc9086c6e525d5e818a44a5ad19720c8a0ef766792f3eb5e5"},
+    {file = "charset_normalizer-3.5.2-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:3d31298449090ab8d47b7b1b2a555ff73cac7ed438a08b7ac160980c7ebed649"},
+    {file = "charset_normalizer-3.5.2-cp312-cp312-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:5cde776b7cc66e4f6c99612cea4aa7269aa65863f7a15841b2c264f103822f4e"},
+    {file = "charset_normalizer-3.5.2-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:ae4f5fea5b8b8ccff88238cc8569303e5ee95efae67fa62922a311397a71f346"},
+    {file = "charset_normalizer-3.5.2-cp312-cp312-musllinux_1_2_armv7l.whl", hash = "sha256:f7d486c83842422badd511868fd8a9a20e9407ace71564b6af47ce7e60a336c1"},
+    {file = "charset_normalizer-3.5.2-cp312-cp312-musllinux_1_2_ppc64le.whl", hash = "sha256:11a4d68a6ecda3292cb1e50239e111543ba5d709bb62a6b4ea1afcfa729d8875"},
+    {file = "charset_normalizer-3.5.2-cp312-cp312-musllinux_1_2_riscv64.whl", hash = "sha256:d6734d2ef8a50fbf8445c139477da401f50d62a0606bf00e20ec6d87773fefb1"},
+    {file = "charset_normalizer-3.5.2-cp312-cp312-musllinux_1_2_s390x.whl", hash = "sha256:a815775b6c38d4e0ff7bcffbeba67feded90202bb6a226b8dd35f1c855217413"},
+    {file = "charset_normalizer-3.5.2-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:23851fb4e1b85ed3f6c2a27b777cdfe2e19fb5b38429a8faf38c7542b7665869"},
+    {file = "charset_normalizer-3.5.2-cp312-cp312-win32.whl", hash = "sha256:db19d07e2e0129e974a0e65d0064fc222a446cd5122c2fd4184d2af9fc734a9e"},
+    {file = "charset_normalizer-3.5.2-cp312-cp312-win_amd64.whl", hash = "sha256:780fbe7cab297b81dad9fb8dc5eb003c0468ffb0d9e5f65068c53a34661a96bc"},
+    {file = "charset_normalizer-3.5.2-cp312-cp312-win_arm64.whl", hash = "sha256:e2af3aad578aa6bd1384bcf4750fc285e5a9de53f40b7d41e5a0bf748edeb2b3"},
+    {file = "charset_normalizer-3.5.2-cp313-cp313-android_24_arm64_v8a.whl", hash = "sha256:ed905975ab14056a2e5eb1c376cb2e1ebc5396baf84163939c518556fccde9f5"},
+    {file = "charset_normalizer-3.5.2-cp313-cp313-android_24_x86_64.whl", hash = "sha256:a66c3bc5ab1f0ff2164fc9965ddd611ff0802173f4b9d24554c563f6ab7e1d6e"},
+    {file = "charset_normalizer-3.5.2-cp313-cp313-ios_13_0_arm64_iphoneos.whl", hash = "sha256:d2374b62878abb00cd8309b32af6c0b715cd02dec0ca74ef12e5069bdc64144a"},
+    {file = "charset_normalizer-3.5.2-cp313-cp313-ios_13_0_arm64_iphonesimulator.whl", hash = "sha256:d376bbd28b3a8999db1a103b3b388aee6f1ddeb3e51bc2172993efdcd86e064d"},
+    {file = "charset_normalizer-3.5.2-cp313-cp313-macosx_10_13_universal2.whl", hash = "sha256:6045373d5a89a5ec71afde535db987ca28e76dfa276c2d4c818265b375d4b055"},
+    {file = "charset_normalizer-3.5.2-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:849df64e889b2e17230d58410a03dba311a65b163508fd33679b2b737d4b7858"},
+    {file = "charset_normalizer-3.5.2-cp313-cp313-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:15c44f7edfd477b06f517a5cc317fc1707edb9de2c865f43d4b6513907473234"},
+    {file = "charset_normalizer-3.5.2-cp313-cp313-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:a89012d6d5476ee112d20d998570ed58df2260a852afb1758809cd6900411d21"},
+    {file = "charset_normalizer-3.5.2-cp313-cp313-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:0c951d5e6dd9c2ff60609476752bee49da4206adde960ebc247766937f72e718"},
+    {file = "charset_normalizer-3.5.2-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:7218e8f32b0956cfcd048fd42d9d5779809745ca1d86113ca56f66e7ae1549c4"},
+    {file = "charset_normalizer-3.5.2-cp313-cp313-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:a19a731138fc27d5682277d3b9df22855cea1239bce7fcec5f78f42ef2d1f3c3"},
+    {file = "charset_normalizer-3.5.2-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:62603db9a7caa0802eaa28c1c46fecd7b3a263a774069c24c3c28c302448721c"},
+    {file = "charset_normalizer-3.5.2-cp313-cp313-musllinux_1_2_armv7l.whl", hash = "sha256:b6856554c4f44d79fc2307d5768854310a8f0096e501c75637542c82292b0429"},
+    {file = "charset_normalizer-3.5.2-cp313-cp313-musllinux_1_2_ppc64le.whl", hash = "sha256:1bc0baf5ef96b6ede57d47f4b8fe4d9d84019c3bfcbeb20a41edc6a6ee341f1f"},
+    {file = "charset_normalizer-3.5.2-cp313-cp313-musllinux_1_2_riscv64.whl", hash = "sha256:56bc200a365efb37383b7852e4cc5898d3b2da5987289b543956cf8cad71018a"},
+    {file = "charset_normalizer-3.5.2-cp313-cp313-musllinux_1_2_s390x.whl", hash = "sha256:2c9ad19a6cfcd5ea5c0d41161d22f9df1dcc277e9bef2751391334546a314c00"},
+    {file = "charset_normalizer-3.5.2-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:e243bd13217235fc7290c621941c3f5cc8b66e4872495be821d7436ba2fb838d"},
+    {file = "charset_normalizer-3.5.2-cp313-cp313-pyemscripten_2025_0_wasm32.whl", hash = "sha256:a090bb2c68df85450502e3e20d665e3a5af9c65a84d6508ed477badd49166fd3"},
+    {file = "charset_normalizer-3.5.2-cp313-cp313-win32.whl", hash = "sha256:2b7b3bbfb4fe8ef40600792d762fbaa9057559f9d3fad209525b7a22b99e91fd"},
+    {file = "charset_normalizer-3.5.2-cp313-cp313-win_amd64.whl", hash = "sha256:78456a747de8dc58360ffa581f30a002baf5aa28cb262536545e91f113ed7639"},
+    {file = "charset_normalizer-3.5.2-cp313-cp313-win_arm64.whl", hash = "sha256:11912e4bb14baae7c5d8791aa55ba0a3a03ec6729073307b0f57270abaa713d3"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314-android_24_arm64_v8a.whl", hash = "sha256:1afb975bd5d68d5ce9f6b6d44fdf2f7e34b895a35e95708a7a91b20a3b51d187"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314-android_24_x86_64.whl", hash = "sha256:bbbfc8e28816f19d7c0f1816664980c0a9875d01b27cdf8eedddb639d9e108ad"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314-ios_13_0_arm64_iphoneos.whl", hash = "sha256:7967d08cf06dee78443b874f98c98036f624f3a4e73e11f9f64f5be4d25393cf"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314-ios_13_0_arm64_iphonesimulator.whl", hash = "sha256:4c2b5031f63e331e3839b40aed2dd6f191e9c07edbde303e7876846ea1946995"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314-macosx_10_15_universal2.whl", hash = "sha256:fcff63213e8e6e47770541a4607175404f47cbb3ebea7b6058cc82d524a0e424"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:8d86d6fc60743dc916eb79e2eb1ec4818e21e427731543af40a3021851174a13"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:7a881931aa470808df94a8c380eed2bbbc76cd9dc622310f99665658c821eb6d"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:8024d00c3faf3fc0c16e07a69f4405e8eac7cc0ab15f65fe6cf43827c4cf72b4"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:4d48f2d08b9de5864e2c8744d4461b862fb149a18274abc8b698c45975573438"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:34276fd796040bf0993ab33a369aa572e6979c7aab225a88893667ad8eac8f7a"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:0521c5665880b33d603717defa76c094048900010897909952397feb3039da56"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:eff0ac9dbe711a4aee69bf04a83896aa9b85f19641264053a9f6d48573abb7dd"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314-musllinux_1_2_armv7l.whl", hash = "sha256:1503bccbeb36d5527790c3930327704c39af22de3112f1b1666a9f3ce15ee204"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314-musllinux_1_2_ppc64le.whl", hash = "sha256:52aa6992700996af31f375de0c6bacd402b0097fe40b53c426b9f51a90ebabc7"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314-musllinux_1_2_riscv64.whl", hash = "sha256:e09a3942ecbdee5cce73ea9d42da82b81b72ac1bf031ce069b93b5adf4eac8cd"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314-musllinux_1_2_s390x.whl", hash = "sha256:c7c9ab723cde841fefb34efbad91e87f00a674b1fe1cd0784fde742bf2c154dc"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:ddc7dacc8ece3a182e7f15cb862d1fd616b46d076cb1ae9dd232b2c38b655874"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314-pyemscripten_2026_0_wasm32.whl", hash = "sha256:ee43c17b173d46a3212baa6ead3ae258eeabdae48c263a01ccf0218c366dd655"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314-win32.whl", hash = "sha256:4f87960d57feabfb618e4e0af6e7371645fa26a277860739d6e5d6e0012c92f0"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314-win_amd64.whl", hash = "sha256:e4e81e09c1578b8df602e3db08b0b3ea0a6947ad612f52bf8dc5ea8d47691f0c"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314-win_arm64.whl", hash = "sha256:80d02b6f04e92601a081dd97b23d3128033098bff5d35d392ddcc0476ea11253"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314t-macosx_10_15_universal2.whl", hash = "sha256:dca9ab98072a5a54ebacebdc45f53e645336b320c667410b061be1ca588ae709"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:f0aa869112ef88429ae17820d99c3dd9504c9e9c671d3c246f3d7442cb051084"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314t-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:c0afc6800ba57ccc350374c5bd6150419915d95ce93cdbab2d783d75eaf30ecb"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:7dcd882da75ef9adf94903b1e3b9419e8aa8fb4c7396822b834b9ef7fb96954f"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314t-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:2e06a3a98f916dd41d27f3105e02e7a40181c98c94b9158733d03a6f80506c09"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:6bd128f206a7752ae1f2ab6c61bf8a24ba28913a10df8b14c2637b973ff97a80"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314t-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:c8f3d67aeaf55f017982b73683f0e7342ba2f6635a78f69ce89ebb26aa411e5c"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:fe9753dfee015c570d73df76f899f18444d41388bffcde097deba51c4fadbb9f"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314t-musllinux_1_2_armv7l.whl", hash = "sha256:92888bb3187c5ba50500b00b3b310c9f2c651709d28036077680cb5255450a03"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314t-musllinux_1_2_ppc64le.whl", hash = "sha256:d008d90a7f2471519aef0c90dfbe73b3e6e4d5e66ac48e19154c17e89e98b604"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314t-musllinux_1_2_riscv64.whl", hash = "sha256:31f3930700408d211f13378ccbe1c40845d8da54bd0681fac3a9b5aae81c7aa8"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314t-musllinux_1_2_s390x.whl", hash = "sha256:2a925889534b3748302dae5dead07cc13480de1dac3aea80a941b729b471ef93"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:f5ec61164adcec446f8969a3358ec3f9b26bbda3b9213e5586d219afa8df2915"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314t-win32.whl", hash = "sha256:598a11a2c7ebaa5334bf698bf29568c9c390abac6a154d8170fedecd1cea38c5"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314t-win_amd64.whl", hash = "sha256:7fdde2c9fd9e3eca40631e024664cf2584272cc8f96308cbe5fdfc930f51d8bc"},
+    {file = "charset_normalizer-3.5.2-cp314-cp314t-win_arm64.whl", hash = "sha256:d1befeed746d247c81127bb14de9dc3d30edb6e5976d34f83f86ed262b1d9105"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315-macosx_10_15_universal2.whl", hash = "sha256:87475fabc8d9996fd9c27debb395e642e8c838d78a00b6e932227a0e06b81e26"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:9409a8bf35cf78353942504b24a57de3d75b708997a1e4bd8db71ac8633ce364"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:498dc3188ca05a68231ac3fdbfc7f57eb67e1343c30e0fea17f8218c1599b253"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:e242bb1c5e76e97dfa9e7f209a71e93a01d7f19ffdd5cfbb2e2d55b4f08f8ab0"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:def79fa35ef0cef8d2accec024f4fdc7ead3012ff02f5215c783f39f03ef8cfc"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:3df041de8887954562c9b261cba85ca0e9ded74048daf125f45edcfaa4832229"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:04851f73ae72b8413dddadb16a49dfee95263553741fd42d546f7d66907e6be5"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315-musllinux_1_2_aarch64.whl", hash = "sha256:183b88127acdb4fabe59d951ab424faf1af7b63cdbb5f776186c1ea2ffcaed98"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315-musllinux_1_2_armv7l.whl", hash = "sha256:16fa0eccf81304b79c5cd87f9271c3b85dd9dd99245e4422ae9c0dd45e0f99d3"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315-musllinux_1_2_ppc64le.whl", hash = "sha256:7441d755b7ab94f8d4eb3e43ec05482d760842fd263d003a99102d742cd835e2"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315-musllinux_1_2_riscv64.whl", hash = "sha256:ca403d7e4798f525fdfc78e258820419cbbd0f0ecbab9de7840e3c017cf6b8cf"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315-musllinux_1_2_s390x.whl", hash = "sha256:df29a0a7107f7011e77f4eebdddec4c7331e24d787a0b21a46d63bdf7445da95"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315-musllinux_1_2_x86_64.whl", hash = "sha256:f3c96f633825733f735c5a9cf21d21a257d8e1edf0b1cee0a064b9c424ca0f7d"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315-win32.whl", hash = "sha256:281cb91036248400f4cc957495cccd44c275c2e0c5854f7e45ac5cf7dc193847"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315-win_amd64.whl", hash = "sha256:89b53f3cda69831909888e0494f4fa0bcd3537e3e138dabeb620bd6ad946bae8"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315-win_arm64.whl", hash = "sha256:6be488a102b8cf28d0391d8c4ba7748938ae28b78ad901f8585520fca33ead1a"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315t-macosx_10_15_universal2.whl", hash = "sha256:915563965d418f986e7e145accc592eae9e1a1be3566ff98a05d7a9ec42a76e1"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:65cd72beeeca9d3aaea1201e5923859f308f952f9c71de93f06063c79f0f7a3b"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315t-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:b7fd005a73d9e657273b7a10dc71a9e03c8fb9ee6999798d6918ce095b81ac7f"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:e54da4baf05720032d527874d40b65fa4d7e5c6c6a43d0c3adbeffcaf275a2b3"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315t-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:124fbf1a8ff966d87ae05bb8bd45a71f966055ed8bba320d0c7cf450bc5f4d0e"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:28b4f0d66fb834ff90f28209ac7bce77868c45d8c93e26f906709d9b7c2e1af9"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315t-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:58ca3755ee7ff7f59b57789ec9833c9de9ea275405cdd240eda1f193112e398a"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315t-musllinux_1_2_aarch64.whl", hash = "sha256:443eae2bf318abeaf6f15d785138f71fd6de770e99a92158b8b814265e079115"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315t-musllinux_1_2_armv7l.whl", hash = "sha256:58f361dcbab699cf8f42db3f47c8e7fd1036f138c23a5d08de9fde5f425a730c"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315t-musllinux_1_2_ppc64le.whl", hash = "sha256:1b4cbc7c3491ccb4aa17fcd8165649d01cf39f76de1696da8631b5f71b85401d"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315t-musllinux_1_2_riscv64.whl", hash = "sha256:ba0b1d2620edf869789c3879223f52bf2afc5d31b3cb47cc57b3a12c05e2aa9d"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315t-musllinux_1_2_s390x.whl", hash = "sha256:5e2b6b57e9733d39f0c9fd3185efa6b8e29652c4cd8fe94180272cf6ed9a78c4"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315t-musllinux_1_2_x86_64.whl", hash = "sha256:51cf45226a9b588d0d2b4880c62d686934b63ab0bd79ca23ab0e9762eb27441b"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315t-win32.whl", hash = "sha256:5fb29fb8cd1a46c27a1bf9613ad5ec2599310d46b4025d9556404a6b6a292800"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315t-win_amd64.whl", hash = "sha256:a192e2c40070d92c3ccf777e3a5c4ff515573cd2bb7ed0c537fdadbbec5bbf21"},
+    {file = "charset_normalizer-3.5.2-cp315-cp315t-win_arm64.whl", hash = "sha256:749e97e1b32313717a565abbe321bc2190bc8b35f1a67e4cdbc7c56c8d8ffe58"},
+    {file = "charset_normalizer-3.5.2-cp37-abi3-macosx_10_9_universal2.whl", hash = "sha256:4275811936e2f06feff5e598fb42a1b7ae852da8e39605211892b56b81a34efd"},
+    {file = "charset_normalizer-3.5.2-cp37-abi3-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:1c50fe28bbc2ced33386f298650d91218076c05420e6cbd790b913adc41659e7"},
+    {file = "charset_normalizer-3.5.2-cp37-abi3-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:d19fbd981a488e22cd04883659ca6b08f50b5974f9fd7c95655ef6a043e5893f"},
+    {file = "charset_normalizer-3.5.2-cp37-abi3-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:0fed1d06615f022ee3b13caf5e8b180cfea32bb2c5aded8a9d44277afc040f93"},
+    {file = "charset_normalizer-3.5.2-cp37-abi3-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:838dcc90063569a0448120554591a1d6c4a4ffe11babf048908793154ab86ade"},
+    {file = "charset_normalizer-3.5.2-cp37-abi3-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:2ce45c6627b22c47e390bc91a41c3d13032192e699fa0bea96e9671b373d69b0"},
+    {file = "charset_normalizer-3.5.2-cp37-abi3-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:0774bf9bf620249fee3e0b8b9fd3065de213be30f3aa94ce2494b3b638949e26"},
+    {file = "charset_normalizer-3.5.2-cp37-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:1db38f4c5496827c1a501846d64d14c3b80c7e6714e406cd7dc36a9899fa1011"},
+    {file = "charset_normalizer-3.5.2-cp37-abi3-musllinux_1_2_armv7l.whl", hash = "sha256:304d8e4d493af723536393eee0c689eb7813f4a474c8b479dee63f1fdd98f621"},
+    {file = "charset_normalizer-3.5.2-cp37-abi3-musllinux_1_2_ppc64le.whl", hash = "sha256:9b7f416ff0978e2f2249330527f0ad6fa02f4932e6199692d3b52da2048c19e4"},
+    {file = "charset_normalizer-3.5.2-cp37-abi3-musllinux_1_2_riscv64.whl", hash = "sha256:01077390b03f7988f11d700a2194e69b119741a86b1a638b1db88891e3eced8e"},
+    {file = "charset_normalizer-3.5.2-cp37-abi3-musllinux_1_2_s390x.whl", hash = "sha256:7e841fb9010836c992c9f12fcbd43a831de93a5f726fc1ccd8ca1d0268c5014c"},
+    {file = "charset_normalizer-3.5.2-cp37-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:9cae88599c7219005d879f98e5ed53341e9a122af585e1091200358a3003d2a0"},
+    {file = "charset_normalizer-3.5.2-cp37-abi3-win32.whl", hash = "sha256:01b0c0d2262a9e28e8484a278c7e1b5d650e3ac8cf2683d2967e25899f208bdf"},
+    {file = "charset_normalizer-3.5.2-cp37-abi3-win_amd64.whl", hash = "sha256:9f56f72050826f63dcee7a7f55b0a77168cb3bfc553fd405e7f8f9ece75a4036"},
+    {file = "charset_normalizer-3.5.2-cp37-abi3-win_arm64.whl", hash = "sha256:40ab6bffa02ae10a0581e6c198be7d2d8ca5c2a0c64e4ed3465d766df457573e"},
+    {file = "charset_normalizer-3.5.2-cp39-cp39-macosx_10_9_universal2.whl", hash = "sha256:75a3ceed0724d625d64b86ca20aba182e4df462e04c2414fc941c0f523f06aac"},
+    {file = "charset_normalizer-3.5.2-cp39-cp39-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:0891b9d3903c5571c03771ca669a4b0ec5618ca722a5c957d3d29cd4e5062848"},
+    {file = "charset_normalizer-3.5.2-cp39-cp39-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:fc14a032f813bf5fe624d991960ea83e9715adc27e4c1830a2361eb1d02ac341"},
+    {file = "charset_normalizer-3.5.2-cp39-cp39-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:8b2bfab86aa71ae13aa41a6a26aab338e0db2b8bc75434b05aea89e011ff35a4"},
+    {file = "charset_normalizer-3.5.2-cp39-cp39-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:9bde855991b7e362c146535e3136a50bfaffc0487d38b33ca7e5edefc6e23849"},
+    {file = "charset_normalizer-3.5.2-cp39-cp39-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:55ea99acb17b9325618de155a0cd6a2e8f5d10be008113e1d433bbb58db543b2"},
+    {file = "charset_normalizer-3.5.2-cp39-cp39-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:68eb192d85ab8e5f6ec69c2bc6ac0179fbf04a5ac1569d12fbef74883fe102d0"},
+    {file = "charset_normalizer-3.5.2-cp39-cp39-musllinux_1_2_aarch64.whl", hash = "sha256:d913de495d90407cd859d263bee2e5d1a4ed3eb6573c04e70d9ec619a7cbed7f"},
+    {file = "charset_normalizer-3.5.2-cp39-cp39-musllinux_1_2_armv7l.whl", hash = "sha256:3ddacd27458c45bdacd6bd6db644bfb730efbf9e830310186e3045c9c5be8fb2"},
+    {file = "charset_normalizer-3.5.2-cp39-cp39-musllinux_1_2_ppc64le.whl", hash = "sha256:588461c2e8384d309bd63e5826019b6977bc66d629b99ac8737bb795d7b2cb5a"},
+    {file = "charset_normalizer-3.5.2-cp39-cp39-musllinux_1_2_riscv64.whl", hash = "sha256:e80e6c2f55656b4824d72065abb4ddd6a525c74bd78a0aab5d9fc2cf4fb5af50"},
+    {file = "charset_normalizer-3.5.2-cp39-cp39-musllinux_1_2_s390x.whl", hash = "sha256:d4a7319f304a774bed22115bc891618e45f85065ab44ea6acd07d274e750519a"},
+    {file = "charset_normalizer-3.5.2-cp39-cp39-musllinux_1_2_x86_64.whl", hash = "sha256:fd1fbe0f116b6e55da77aca2c6ddcddcfac2186cbf78bdebf40fc156efca389d"},
+    {file = "charset_normalizer-3.5.2-cp39-cp39-win32.whl", hash = "sha256:93223adc95033dd47133a46ccfc316a0139176fd79085762e27202ec56018f03"},
+    {file = "charset_normalizer-3.5.2-cp39-cp39-win_amd64.whl", hash = "sha256:15bb4005af6320d259dc7593ca84a38d7fe06a421dbcf7b910ae23979101e787"},
+    {file = "charset_normalizer-3.5.2-cp39-cp39-win_arm64.whl", hash = "sha256:2cc961b171b3f3440f410489ab3573e86aea8736134ebbb40ea1338b7f0831bc"},
+    {file = "charset_normalizer-3.5.2-py3-none-any.whl", hash = "sha256:b6b751274acb69d77b3323d6b7dbaa3c7fdfc1eb829b7eb61d262f32e1af9685"},
+    {file = "charset_normalizer-3.5.2.tar.gz", hash = "sha256:39de2a259fc954455c57274dc94c79d5842774e1247a016aff30bc0efed0f4ef"},
+]
+
+[[package]]
+name = "click"
+version = "8.5.0"
+description = "Composable command line interface toolkit"
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "click-8.5.0-py3-none-any.whl", hash = "sha256:255bc9599cf7748b4b1a446ccc735421bd08a2ae529a8b88597d3de5664ee360"},
+    {file = "click-8.5.0.tar.gz", hash = "sha256:ba0d2089de75ea0310e2dde03160e6ca10009947fb95a182f9b54021bb272e34"},
+]
+
+[[package]]
+name = "colorama"
+version = "0.4.6"
+description = "Cross-platform colored terminal text."
+optional = false
+python-versions = "!=3.0.*,!=3.1.*,!=3.2.*,!=3.3.*,!=3.4.*,!=3.5.*,!=3.6.*,>=2.7"
+groups = ["main"]
+markers = "platform_system == \"Windows\""
+files = [
+    {file = "colorama-0.4.6-py2.py3-none-any.whl", hash = "sha256:4f1d9991f5acc0ca119f9d443620b77f9d6b33703e51011c16baf57afb285fc6"},
+    {file = "colorama-0.4.6.tar.gz", hash = "sha256:08695f5cb7ed6e0531a20572697297273c47b8cae5a63ffc6d6ed5c201be6e44"},
+]
+
+[[package]]
+name = "cuda-bindings"
+version = "13.4.3"
+description = "Python bindings for CUDA"
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+markers = "platform_system == \"Linux\" and sys_platform == \"darwin\""
+files = [
+    {file = "cuda_bindings-13.4.3-cp310-cp310-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:bc51990309b416e0780a11a921ac597ef9c0e52e1bdf5ae0b8e8c4766b66442b"},
+    {file = "cuda_bindings-13.4.3-cp310-cp310-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:3d7a6506e625be59cdc119ef4423afa0fc3a4830b129dc010cbd51324b93d66d"},
+    {file = "cuda_bindings-13.4.3-cp310-cp310-win_amd64.whl", hash = "sha256:e52f66340785a51b8b77f4487329de188f3cabbf37c62e68f410a8299e8677ae"},
+    {file = "cuda_bindings-13.4.3-cp311-cp311-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:df8b3767facca8acde216460df684dfc3826d519096c13adfd3030a55870dc79"},
+    {file = "cuda_bindings-13.4.3-cp311-cp311-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:a9ea13b9cfe711515ae83b5efc99101af4d3f8ad882f99724a839c8b5417fb7e"},
+    {file = "cuda_bindings-13.4.3-cp311-cp311-win_amd64.whl", hash = "sha256:52d7f3f5f7f014dddddc66cd802ed0ddf65ae99ad42b5619fae7b82e5ddd6771"},
+    {file = "cuda_bindings-13.4.3-cp311-cp311-win_arm64.whl", hash = "sha256:b89d6e738494b7b95c38e3413f86d32c24682c8e714870531a5a2b193a2fc50a"},
+    {file = "cuda_bindings-13.4.3-cp312-cp312-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:bfbd3f7d4ac04dd41dc49121b9e408c8283992f47124c2290ecb79bbbadcca8e"},
+    {file = "cuda_bindings-13.4.3-cp312-cp312-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:d5f72bcfcdf3be23e1da3c792f68f508586f48d037bca8b10f552c4cca5971f2"},
+    {file = "cuda_bindings-13.4.3-cp312-cp312-win_amd64.whl", hash = "sha256:f8519603001c92bf83e7095df3b8211e3999a9c4ded096b57de0f3ff52b66368"},
+    {file = "cuda_bindings-13.4.3-cp312-cp312-win_arm64.whl", hash = "sha256:130ff1daae550db2cef559ba477f3a63d040756bd135f5deba4b5bdc4e110252"},
+    {file = "cuda_bindings-13.4.3-cp313-cp313-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:d7c6c9f46fca7f3fc61959ef9a2398ac656172145b43f408e0a6492360cf1c0c"},
+    {file = "cuda_bindings-13.4.3-cp313-cp313-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:1fd7d8459b364aedc11f3e59703453ced823135f78a9111ca70feef8d56d4d21"},
+    {file = "cuda_bindings-13.4.3-cp313-cp313-win_amd64.whl", hash = "sha256:81eef62bbb95cb4a705fb423b2ad1c63d716edb352ee2af0c28bf8a29cc8258c"},
+    {file = "cuda_bindings-13.4.3-cp313-cp313-win_arm64.whl", hash = "sha256:c2af7e69d2557fdb4e5fc1169159fd08da7e861aa84fb83307a1a0396b6b0973"},
+    {file = "cuda_bindings-13.4.3-cp314-cp314-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:4796864ce829bd95ef2ef0d23c6ba21bb64e08f7fab0a377302ed1affb6605c7"},
+    {file = "cuda_bindings-13.4.3-cp314-cp314-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:bbacde6f75665b197016b986164cfdaa33b17515e5e635a63ddb75926aaa71c3"},
+    {file = "cuda_bindings-13.4.3-cp314-cp314-win_amd64.whl", hash = "sha256:0deff5b22462bb410859684299b1fc6a621780c43a4729342aa6ec28783485b4"},
+    {file = "cuda_bindings-13.4.3-cp314-cp314-win_arm64.whl", hash = "sha256:145cc9b02dfb7bcc3288701b6f19c69c590e4451c62916ba0a8a3db9a64a91a8"},
+    {file = "cuda_bindings-13.4.3-cp314-cp314t-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:d6eb969920e28f66f8fc3b0b3afcb6e09381cc96bf8e8158d774e9488ae89980"},
+    {file = "cuda_bindings-13.4.3-cp314-cp314t-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:7e11cfe8fec4c85ce79feda18124971c52596f0cbd642a94f5dafc257124a4b3"},
+    {file = "cuda_bindings-13.4.3-cp314-cp314t-win_amd64.whl", hash = "sha256:6c6bb1a4c4f7520e062669c6f3605a7016b2e65ba51b8922f92edd91d5fadf8d"},
+    {file = "cuda_bindings-13.4.3-cp314-cp314t-win_arm64.whl", hash = "sha256:044c03b056dcc5cecfad426a071187dd9e1e6817fbb363bc2f3a7520170b3e67"},
+]
+
+[package.dependencies]
+cuda-pathfinder = ">=1.4.2"
+
+[package.extras]
+all = ["cuda-toolkit (==13.*)", "cuda-toolkit[cufile] (==13.*) ; sys_platform == \"linux\"", "cuda-toolkit[nvfatbin,nvjitlink,nvrtc,nvvm] (==13.*)", "nvidia-cudla (==13.*) ; platform_system == \"Linux\" and platform_machine == \"aarch64\""]
+
+[[package]]
+name = "cuda-pathfinder"
+version = "1.8.3"
+description = "Pathfinder for CUDA components"
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+markers = "platform_system == \"Linux\" and sys_platform == \"darwin\""
+files = [
+    {file = "cuda_pathfinder-1.8.3-py3-none-any.whl", hash = "sha256:e29e59829c297a7a5233bd9cc71094fc5bddbd076951482670178f9eade39b1f"},
+]
+
+[[package]]
+name = "cuda-toolkit"
+version = "13.0.3"
+description = "CUDA Toolkit meta-package"
+optional = false
+python-versions = "*"
+groups = ["main"]
+markers = "sys_platform == \"darwin\" and platform_system == \"Linux\""
+files = [
+    {file = "cuda_toolkit-13.0.3.0-py2.py3-none-any.whl", hash = "sha256:d693caaa261214ddd7dbb60d68e71cbed884e68c2be7509778f3051da0b91c3f"},
+]
+
+[package.extras]
+all = ["nvidia-cublas (==13.1.1.3.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\"", "nvidia-cuda-cccl (==13.0.85.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\"", "nvidia-cuda-crt (==13.0.88.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\"", "nvidia-cuda-culibos (==13.0.85.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\")", "nvidia-cuda-cupti (==13.0.85.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\"", "nvidia-cuda-cuxxfilt (==13.0.85.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\"", "nvidia-cuda-nvcc (==13.0.88.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\"", "nvidia-cuda-nvrtc (==13.0.88.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\"", "nvidia-cuda-opencl (==13.0.85.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\"", "nvidia-cuda-profiler-api (==13.0.85.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\"", "nvidia-cuda-runtime (==13.0.96.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\"", "nvidia-cuda-sanitizer-api (==13.0.85.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\"", "nvidia-cufft (==12.0.0.61.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\"", "nvidia-cufile (==1.15.1.6.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\")", "nvidia-curand (==10.4.0.35.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\"", "nvidia-cusolver (==12.0.4.66.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\"", "nvidia-cusparse (==12.6.3.3.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\"", "nvidia-npp (==13.0.1.2.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\"", "nvidia-nvfatbin (==13.0.85.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\"", "nvidia-nvjitlink (>=13.0.88,<14) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\"", "nvidia-nvjpeg (==13.0.1.86.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\"", "nvidia-nvml-dev (==13.0.87.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\"", "nvidia-nvptxcompiler (==13.0.88.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\"", "nvidia-nvtx (==13.0.85.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\"", "nvidia-nvvm (==13.0.88.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\""]
+cccl = ["nvidia-cuda-cccl (==13.0.85.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\""]
+crt = ["nvidia-cuda-crt (==13.0.88.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\""]
+cublas = ["nvidia-cublas (==13.1.1.3.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\"", "nvidia-cuda-nvrtc (==13.0.88.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\""]
+cudart = ["nvidia-cuda-runtime (==13.0.96.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\""]
+cufft = ["nvidia-cufft (==12.0.0.61.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\"", "nvidia-nvjitlink (>=13.0.88,<14) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\""]
+cufile = ["nvidia-cufile (==1.15.1.6.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\")"]
+culibos = ["nvidia-cuda-culibos (==13.0.85.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\")"]
+cupti = ["nvidia-cuda-cupti (==13.0.85.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\""]
+curand = ["nvidia-curand (==10.4.0.35.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\""]
+cusolver = ["nvidia-cublas (==13.1.1.3.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\"", "nvidia-cusolver (==12.0.4.66.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\"", "nvidia-cusparse (==12.6.3.3.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\"", "nvidia-nvjitlink (>=13.0.88,<14) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\""]
+cusparse = ["nvidia-cusparse (==12.6.3.3.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\"", "nvidia-nvjitlink (>=13.0.88,<14) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\""]
+cuxxfilt = ["nvidia-cuda-cuxxfilt (==13.0.85.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\""]
+npp = ["nvidia-npp (==13.0.1.2.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\""]
+nvcc = ["nvidia-cuda-crt (==13.0.88.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\"", "nvidia-cuda-nvcc (==13.0.88.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\"", "nvidia-cuda-runtime (==13.0.96.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\"", "nvidia-nvvm (==13.0.88.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\""]
+nvfatbin = ["nvidia-nvfatbin (==13.0.85.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\""]
+nvjitlink = ["nvidia-nvjitlink (>=13.0.88,<14) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\""]
+nvjpeg = ["nvidia-nvjpeg (==13.0.1.86.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\""]
+nvml = ["nvidia-nvml-dev (==13.0.87.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\""]
+nvptxcompiler = ["nvidia-nvptxcompiler (==13.0.88.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\""]
+nvrtc = ["nvidia-cuda-nvrtc (==13.0.88.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\""]
+nvtx = ["nvidia-nvtx (==13.0.85.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\""]
+nvvm = ["nvidia-nvvm (==13.0.88.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\""]
+opencl = ["nvidia-cuda-opencl (==13.0.85.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\""]
+profiler = ["nvidia-cuda-profiler-api (==13.0.85.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\""]
+sanitizer = ["nvidia-cuda-sanitizer-api (==13.0.85.*) ; sys_platform == \"linux\" and (platform_machine == \"aarch64\" or platform_machine == \"x86_64\") or (sys_platform == \"linux\" or sys_platform == \"win32\") and platform_machine == \"x86_64\""]
+
+[[package]]
+name = "filelock"
+version = "4.0.12"
+description = "A platform independent file lock."
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "filelock-4.0.12-py3-none-any.whl", hash = "sha256:5f17ee83ecee8a6f3e389c75822fb70a1c2f0506b99438a6dffa1d56793588c8"},
+    {file = "filelock-4.0.12.tar.gz", hash = "sha256:cf42711a7ac791818b299fab0332a088c65aeeefa36290de98db92c434303b0c"},
+]
+
+[[package]]
+name = "fsspec"
+version = "2026.9.0"
+description = "File-system specification"
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "fsspec-2026.9.0-py3-none-any.whl", hash = "sha256:8dd6e646e99ea382bd85f97a45e6b526a442d79423a7dc673f1e2756d05fcb5f"},
+    {file = "fsspec-2026.9.0.tar.gz", hash = "sha256:0f08147951c8cb31d844c3547d631053b127863b60be04cf06e121333ee0e2fe"},
+]
+
+[package.extras]
+abfs = ["adlfs"]
+adl = ["adlfs"]
+arrow = ["pyarrow (>=1)"]
+dask = ["dask", "distributed"]
+dev = ["pre-commit", "ruff (>=0.5)"]
+doc = ["numpydoc", "sphinx", "sphinx-design", "sphinx-rtd-theme", "yarl"]
+dropbox = ["dropbox", "dropboxdrivefs", "requests"]
+full = ["adlfs", "aiohttp (!=4.0.0a0,!=4.0.0a1)", "dask", "distributed", "dropbox", "dropboxdrivefs", "fusepy", "gcsfs (>=2026.4.0)", "libarchive-c", "ocifs", "panel", "paramiko", "pyarrow (>=1)", "pygit2", "requests", "s3fs (>=2026.6.0)", "smbprotocol", "tqdm"]
+fuse = ["fusepy"]
+gcs = ["gcsfs (>=2026.4.0)"]
+git = ["pygit2"]
+github = ["requests"]
+gs = ["gcsfs (>=2026.4.0)"]
+gui = ["panel"]
+hdfs = ["pyarrow (>=1)"]
+http = ["aiohttp (!=4.0.0a0,!=4.0.0a1)"]
+libarchive = ["libarchive-c"]
+oci = ["ocifs"]
+s3 = ["s3fs (>=2026.6.0)"]
+sftp = ["paramiko"]
+smb = ["smbprotocol"]
+ssh = ["paramiko"]
+test = ["aiohttp (!=4.0.0a0,!=4.0.0a1)", "numpy", "pytest", "pytest-asyncio (!=0.22.0)", "pytest-benchmark", "pytest-cov", "pytest-mock", "pytest-recording", "pytest-rerunfailures", "requests"]
+test-downstream = ["aiobotocore (>=2.5.4,<3.0.0)", "dask[dataframe,test]", "moto[server] (>4,<5)", "pytest-timeout", "xarray", "zarr"]
+test-full = ["adlfs", "aiohttp (!=4.0.0a0,!=4.0.0a1)", "backports-zstd ; python_version < \"3.14\"", "cloudpickle", "dask", "distributed", "dropbox", "dropboxdrivefs", "fastparquet", "fusepy", "gcsfs (>=2026.4.0)", "jinja2", "kerchunk", "libarchive-c", "lz4", "notebook", "numpy", "ocifs", "pandas (<3.0.0)", "panel", "paramiko", "pyarrow (>=1)", "pyftpdlib", "pygit2", "pytest", "pytest-asyncio (!=0.22.0)", "pytest-benchmark", "pytest-cov", "pytest-mock", "pytest-recording", "pytest-rerunfailures", "python-snappy", "requests", "s3fs (>=2026.6.0)", "smbprotocol", "tqdm", "urllib3", "zarr (<3.2.0)", "zstandard ; python_version < \"3.14\""]
+tqdm = ["tqdm"]
+
+[[package]]
+name = "h11"
+version = "0.16.0"
+description = "A pure-Python, bring-your-own-I/O implementation of HTTP/1.1"
+optional = false
+python-versions = ">=3.8"
+groups = ["main"]
+markers = "sys_platform != \"emscripten\""
+files = [
+    {file = "h11-0.16.0-py3-none-any.whl", hash = "sha256:63cf8bbe7522de3bf65932fda1d9c2772064ffb3dae62d55932da54b31cb6c86"},
+    {file = "h11-0.16.0.tar.gz", hash = "sha256:4e35b956cf45792e4caa5885e69fba00bdbc6ffafbfa020300e549b208ee5ff1"},
+]
+
+[[package]]
+name = "hf-xet"
+version = "1.6.0"
+description = "Fast transfer of large files with the Hugging Face Hub."
+optional = false
+python-versions = ">=3.8"
+groups = ["main"]
+markers = "platform_machine == \"x86_64\" or platform_machine == \"amd64\" or platform_machine == \"AMD64\" or platform_machine == \"arm64\" or platform_machine == \"ARM64\" or platform_machine == \"aarch64\""
+files = [
+    {file = "hf_xet-1.6.0-cp314-cp314t-macosx_10_12_x86_64.whl", hash = "sha256:70cbb9c896901600128cb9b6f06e132954fbede1db30f31f7c6c63f84cb7c31d"},
+    {file = "hf_xet-1.6.0-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:23379c2f9ec8696d952b16414a2bae72cad86a52df869b050698ba60f538c675"},
+    {file = "hf_xet-1.6.0-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:f2f7278c05c22fd60cb436cda1269649b3e81db65ecdc8496e5e164aa4143e7b"},
+    {file = "hf_xet-1.6.0-cp314-cp314t-manylinux_2_28_aarch64.whl", hash = "sha256:948f15d3a9545cfe5932f6bd8b440f6ae630aee108f14b7bd6c561f7c2dcc522"},
+    {file = "hf_xet-1.6.0-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:5153e6bb103ad49d6ea9f1b2e230db5a2ea32551ad09a706d2f61d7c7c80d80e"},
+    {file = "hf_xet-1.6.0-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:35cec30d75c6f9eb9c16a77cef68e85a103b72e24d4b473714ec9ff06428bab9"},
+    {file = "hf_xet-1.6.0-cp314-cp314t-win_amd64.whl", hash = "sha256:5789835d7c6bc9436962853192082374297fb72d7eff7e7762ec25ceb7e25338"},
+    {file = "hf_xet-1.6.0-cp314-cp314t-win_arm64.whl", hash = "sha256:75765820ce4700db3750c94acc8fe27c5fae4c9ec000a0dbac3ca082acf97765"},
+    {file = "hf_xet-1.6.0-cp38-abi3-macosx_10_12_x86_64.whl", hash = "sha256:633dc0cd71d32da58ab8c03ad38e2fac452c15c2b0a2866ebf6ededfe0a5061d"},
+    {file = "hf_xet-1.6.0-cp38-abi3-macosx_11_0_arm64.whl", hash = "sha256:f0906082d9932ae0c0057fa194041c22b4e2cdb46b2592ef3b91f020d62a081a"},
+    {file = "hf_xet-1.6.0-cp38-abi3-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:d62671bb130879cef0ee4c9ebe47a14af6c66ec53e6d84dc15936e5ffdfac82f"},
+    {file = "hf_xet-1.6.0-cp38-abi3-manylinux_2_28_aarch64.whl", hash = "sha256:0e6e21fa3cdfcdcd76748564bf593870a5e013f47d97cf10aed63aa222cff5b7"},
+    {file = "hf_xet-1.6.0-cp38-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:4fc74352a17015bd0ee90038bc9efe38db894cde45f268b6712b04fce8cd0acb"},
+    {file = "hf_xet-1.6.0-cp38-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:8fb4f71cba6129110c3374a33f919001ff130488fc23553698e34cc1c2a1198c"},
+    {file = "hf_xet-1.6.0-cp38-abi3-win_amd64.whl", hash = "sha256:fb4fadde1b2b70bf4c0c14a6dccbe7194b1c28947fefd5bbe3fed9d940676c3b"},
+    {file = "hf_xet-1.6.0-cp38-abi3-win_arm64.whl", hash = "sha256:3dc3e35441ba395006af5aaacc40ef2e603c51ef46c3530b9156185f00935ea3"},
+    {file = "hf_xet-1.6.0.tar.gz", hash = "sha256:2e58454a340b3556dfa4972d5451aff4fba8dd42a236600ba1a1d2b1514f0fef"},
+]
+
+[package.extras]
+tests = ["pytest"]
+
+[[package]]
+name = "httpcore2"
+version = "2.16.0"
+description = "A minimal low-level HTTP client."
+optional = false
+python-versions = ">=3.11"
+groups = ["main"]
+markers = "sys_platform != \"emscripten\""
+files = [
+    {file = "httpcore2-2.16.0-py3-none-any.whl", hash = "sha256:5c4358319f2602d984ad93182c9fee5160f22e9da7babdda870a5e7c4e2dc34f"},
+    {file = "httpcore2-2.16.0.tar.gz", hash = "sha256:c72cd0413563b17e1cd8354677dff887466f8ec2551679759e93126534428643"},
+]
+
+[package.dependencies]
+h11 = ">=0.16"
+truststore = ">=0.10"
+
+[package.extras]
+asyncio = ["anyio (>=4.5.0,<5.0)"]
+http2 = ["h2 (>=3,<5)"]
+socks = ["socksio (==1.*)"]
+trio = ["trio (>=0.34.0,<1.0)"]
+
+[[package]]
+name = "httpx2"
+version = "2.16.0"
+description = "The next generation HTTP client."
+optional = false
+python-versions = ">=3.11"
+groups = ["main"]
+files = [
+    {file = "httpx2-2.16.0-py3-none-any.whl", hash = "sha256:fd2c5d7909e8d86fe92d53020c310ca0e4e431b838c207c5754b9bbf0308d49a"},
+    {file = "httpx2-2.16.0.tar.gz", hash = "sha256:0d54c98128e8f2ff7fbbde0fca8c359e5f032edeaaaac443cdb7f31f685bcd7f"},
+]
+
+[package.dependencies]
+anyio = {version = ">=4.10", markers = "sys_platform != \"emscripten\""}
+httpcore2 = {version = "2.16.0", markers = "sys_platform != \"emscripten\""}
+httpx2-jsfetch = {version = "*", markers = "sys_platform == \"emscripten\" and python_version >= \"3.12\""}
+idna = ">=3.18"
+truststore = {version = ">=0.10", markers = "sys_platform != \"emscripten\""}
+typing-extensions = {version = ">=4.5.0", markers = "python_version < \"3.13\""}
+
+[package.extras]
+brotli = ["brotli (>=1.2.0) ; platform_python_implementation == \"CPython\"", "brotlicffi (>=1.2.0.2) ; platform_python_implementation != \"CPython\""]
+cli = ["click (>=8.4.2)", "pygments (==2.*)", "rich (>=10,<16)"]
+http2 = ["h2 (>=3,<5)"]
+socks = ["socksio (==1.*)"]
+ws = ["wsproto (>=1.2)"]
+zstd = ["backports-zstd (>=1.0.0) ; python_version <= \"3.13\""]
+
+[[package]]
+name = "httpx2-jsfetch"
+version = "1.0"
+description = "httpx2 transports for Emscripten/Pyodide, backed by the JavaScript fetch API."
+optional = false
+python-versions = ">=3.12"
+groups = ["main"]
+markers = "sys_platform == \"emscripten\""
+files = [
+    {file = "httpx2_jsfetch-1.0-py3-none-any.whl", hash = "sha256:cb916b707601e69a07721aabc8f3f6659be3a6893bc1ff5c6f9e02241df2da32"},
+    {file = "httpx2_jsfetch-1.0.tar.gz", hash = "sha256:70a0e3eabfef7cce5ad9c629f7d01ca05e418f586646f4ddf14782e4c1454c60"},
+]
+
+[[package]]
+name = "huggingface-hub"
+version = "2.3.0"
+description = "Client library to download and publish models, datasets and other repos on the huggingface.co hub"
+optional = false
+python-versions = ">=3.10.0"
+groups = ["main"]
+files = [
+    {file = "huggingface_hub-2.3.0-py3-none-any.whl", hash = "sha256:5da369600233568a126eb5578754935b2add9446b42d9e2fc0927d56840ac1a5"},
+    {file = "huggingface_hub-2.3.0.tar.gz", hash = "sha256:0b783b37f200f1d25bef2344a997c7d410f4437fbe9d8fd965f958e75cec7aef"},
+]
+
+[package.dependencies]
+click = ">=8.4.2,<9.0.0"
+filelock = ">=3.10.0"
+fsspec = ">=2023.5.0"
+hf-xet = {version = ">=1.6.0,<2.0.0", markers = "platform_machine == \"x86_64\" or platform_machine == \"amd64\" or platform_machine == \"AMD64\" or platform_machine == \"arm64\" or platform_machine == \"ARM64\" or platform_machine == \"aarch64\""}
+httpx2 = ">=2.0.0,<3"
+packaging = ">=20.9"
+pyyaml = ">=5.1"
+tqdm = ">=4.42.1"
+typing-extensions = ">=4.1.0"
+
+[package.extras]
+all = ["Jinja2", "Pillow", "authlib (>=1.8.0)", "duckdb ; python_version < \"3.15\"", "fastapi", "fastapi", "itsdangerous", "jedi", "libcst (>=1.4.0)", "mypy (==1.15.0)", "numpy", "pytest (>=8.4.2)", "pytest-asyncio", "pytest-cov", "pytest-env", "pytest-mock", "pytest-rerunfailures (>=16.2)", "pytest-vcr", "pytest-xdist", "ruff (>=0.9.0)", "soundfile", "ty", "types-PyYAML", "types-simplejson", "types-toml", "types-tqdm", "types-urllib3", "typing-extensions (>=4.8.0)", "urllib3 (<2.0)"]
+dev = ["Jinja2", "Pillow", "authlib (>=1.8.0)", "duckdb ; python_version < \"3.15\"", "fastapi", "fastapi", "itsdangerous", "jedi", "libcst (>=1.4.0)", "mypy (==1.15.0)", "numpy", "pytest (>=8.4.2)", "pytest-asyncio", "pytest-cov", "pytest-env", "pytest-mock", "pytest-rerunfailures (>=16.2)", "pytest-vcr", "pytest-xdist", "ruff (>=0.9.0)", "soundfile", "ty", "types-PyYAML", "types-simplejson", "types-toml", "types-tqdm", "types-urllib3", "typing-extensions (>=4.8.0)", "urllib3 (<2.0)"]
+fastai = ["fastai (>=2.4)", "fastcore (>=1.3.27)", "toml"]
+gradio = ["gradio (>=5.0.0)", "requests"]
+hf-xet = ["hf-xet (>=1.6.0,<2.0.0)"]
+mcp = ["mcp (>=1.9.4,<2.0.0)"]
+oauth = ["authlib (>=1.8.0)", "fastapi", "itsdangerous"]
+quality = ["libcst (>=1.4.0)", "mypy (==1.15.0)", "ruff (>=0.9.0)", "ty"]
+testing = ["Jinja2", "Pillow", "authlib (>=1.8.0)", "duckdb ; python_version < \"3.15\"", "fastapi", "fastapi", "itsdangerous", "jedi", "numpy", "pytest (>=8.4.2)", "pytest-asyncio", "pytest-cov", "pytest-env", "pytest-mock", "pytest-rerunfailures (>=16.2)", "pytest-vcr", "pytest-xdist", "soundfile", "urllib3 (<2.0)"]
+torch = ["safetensors[torch]", "torch"]
+typing = ["types-PyYAML", "types-simplejson", "types-toml", "types-tqdm", "types-urllib3", "typing-extensions (>=4.8.0)"]
+
+[[package]]
+name = "idna"
+version = "3.20"
+description = "Internationalized Domain Names in Applications (IDNA)"
+optional = false
+python-versions = ">=3.9"
+groups = ["main"]
+files = [
+    {file = "idna-3.20-py3-none-any.whl", hash = "sha256:ab7ae7122974553370f0bdb919e1a960b2cd1bc1ef0276416d896db81c14582c"},
+    {file = "idna-3.20.tar.gz", hash = "sha256:a7db850025b95ded1eae8a46181a1a6c56c92c96f0e2b005d9ff8dc0210cab44"},
+]
+
+[package.extras]
+all = ["coverage (>=7.10.0)", "hypothesis (>=6.141.1)", "mypy (>=1.11.2)", "pytest (>=8.3.2)", "ruff (>=0.16.0)", "ty (>=0.0.37)"]
+
+[[package]]
+name = "jinja2"
+version = "3.1.6"
+description = "A very fast and expressive template engine."
+optional = false
+python-versions = ">=3.7"
+groups = ["main"]
+markers = "sys_platform == \"linux\" or sys_platform == \"darwin\""
+files = [
+    {file = "jinja2-3.1.6-py3-none-any.whl", hash = "sha256:85ece4451f492d0c13c5dd7c13a64681a86afae63a5f347908daf103ce6d2f67"},
+    {file = "jinja2-3.1.6.tar.gz", hash = "sha256:0137fb05990d35f1275a587e9aee6d56da821fc83491a0fb838183be43f66d6d"},
+]
+
+[package.dependencies]
+MarkupSafe = ">=2.0"
+
+[package.extras]
+i18n = ["Babel (>=2.7)"]
+
+[[package]]
+name = "markdown-it-py"
+version = "4.2.0"
+description = "Python port of markdown-it. Markdown parsing, done right!"
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "markdown_it_py-4.2.0-py3-none-any.whl", hash = "sha256:9f7ebbcd14fe59494226453aed97c1070d83f8d24b6fc3a3bcf9a38092641c4a"},
+    {file = "markdown_it_py-4.2.0.tar.gz", hash = "sha256:04a21681d6fbb623de53f6f364d352309d4094dd4194040a10fd51833e418d49"},
+]
+
+[package.dependencies]
+mdurl = ">=0.1,<1.0"
+
+[package.extras]
+benchmarking = ["psutil", "pytest", "pytest-benchmark"]
+compare = ["commonmark (>=0.9,<1.0)", "markdown (>=3.4,<4.0)", "markdown-it-pyrs", "mistletoe (>=1.0,<2.0)", "mistune (>=3.0,<4.0)", "panflute (>=2.3,<3.0)"]
+linkify = ["linkify-it-py (>=1,<3)"]
+plugins = ["mdit-py-plugins (>=0.5.0)"]
+profiling = ["gprof2dot"]
+rtd = ["ipykernel", "jupyter_sphinx", "mdit-py-plugins (>=0.5.0)", "myst-parser", "pyyaml", "sphinx", "sphinx-book-theme (>=1.0,<2.0)", "sphinx-copybutton", "sphinx-design"]
+testing = ["coverage", "pytest", "pytest-cov", "pytest-regressions", "pytest-timeout", "requests"]
+
+[[package]]
+name = "markupsafe"
+version = "3.0.4"
+description = "Safely add untrusted strings to HTML/XML markup."
+optional = false
+python-versions = ">=3.9"
+groups = ["main"]
+markers = "sys_platform == \"linux\" or sys_platform == \"darwin\""
+files = [
+    {file = "markupsafe-3.0.4-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:dd8ea6ebee7aedbf7c749fa80521d9ccf1ba473e0d1e14805caafbaad281c889"},
+    {file = "markupsafe-3.0.4-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:dff05cb7016dff1e9fd68f4122c127b65dfc59de5306cfb7ad92f956f230bee2"},
+    {file = "markupsafe-3.0.4-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:cf63c214fe879a65e69a386f915e36104fc84254ab141240f8854602d8e0be2a"},
+    {file = "markupsafe-3.0.4-cp310-cp310-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:2a6ef68ae94aed8721934072b27a3b654ea2100b97e4ab864cf1489c90926fbc"},
+    {file = "markupsafe-3.0.4-cp310-cp310-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:fd9f8797427910198f95bced71ddfed61130d7e349213bfb8466c9c99e2c46a8"},
+    {file = "markupsafe-3.0.4-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:d1aca03ede943eb80ab3d63bb082c84b7aab85ea83bd0fd0c200260945fb49d9"},
+    {file = "markupsafe-3.0.4-cp310-cp310-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:0764a13d34cae40db7bbf3a09b7e9b491bf4603e20b263a7a9d6b8e324975d0a"},
+    {file = "markupsafe-3.0.4-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:9388003072b95f2f1e3fd908604194d653ba21330d811961a78b7da1a77e9e36"},
+    {file = "markupsafe-3.0.4-cp310-cp310-musllinux_1_2_armv7l.whl", hash = "sha256:8698d70a8081ee8c090dbb394768b5789a1da8b131b5499f89d071dd3cfaf6be"},
+    {file = "markupsafe-3.0.4-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:bf053da3c97a4bc5ecfbb218cdd2983febd91c617be8367d139882aa11e490aa"},
+    {file = "markupsafe-3.0.4-cp310-cp310-musllinux_1_2_riscv64.whl", hash = "sha256:9438a2648b2195980cb2dd8e53ed7b8df91319e2d0b70ae61a9e1d1bc8d3bec9"},
+    {file = "markupsafe-3.0.4-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:88d59b473bfb03259722600839af9bbd7fa13a2eb514beefeedb95997882f69a"},
+    {file = "markupsafe-3.0.4-cp310-cp310-win32.whl", hash = "sha256:4a540e2d3192792fc84eced57bef37851ccb2b41f73291bb17408eea77bcd278"},
+    {file = "markupsafe-3.0.4-cp310-cp310-win_amd64.whl", hash = "sha256:5c22873ad1f0532ba40fa1727f3c0fc1bbbaab6d373d4cbe3f0dc74b2e2521c7"},
+    {file = "markupsafe-3.0.4-cp310-cp310-win_arm64.whl", hash = "sha256:3d23795802fc8bd72534836d64489bbf0f67c088959091bdb22e10735a5107bf"},
+    {file = "markupsafe-3.0.4-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:9e25feb9e330b63edb0278a0acdf85e50d0cb0fbf49c3084abbe4e24ae195346"},
+    {file = "markupsafe-3.0.4-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:7d3391b2188d18737cb2fa147028b1096236eaa7e156446c650a489fa2cadc91"},
+    {file = "markupsafe-3.0.4-cp311-cp311-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:849dd2bb0e5e4ab2b71c7191726a4a8d5aa8a610daa584728cbee0b710ddc4ef"},
+    {file = "markupsafe-3.0.4-cp311-cp311-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:befb4158af32106b9a93db8d6d1d1cbbd418c0d5aca0cabb7b1780abf0c89169"},
+    {file = "markupsafe-3.0.4-cp311-cp311-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:71f88e749ea29f67f21f3b36433c1dc54c7729ed2a6d9e2da2e0d9e0d7b224eb"},
+    {file = "markupsafe-3.0.4-cp311-cp311-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:6da83a088f8ef93b2d483a8232a4dbf4d69d3d8496b568a03c56becac43e1808"},
+    {file = "markupsafe-3.0.4-cp311-cp311-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:8f0fac8b13d14bb06c68195f849371924ae53dd7b1c00fed24650f704383b692"},
+    {file = "markupsafe-3.0.4-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:4a7cdc2a420ca01058182da4253329764d4bfa055564d1eced90e6ba1e8b1d3d"},
+    {file = "markupsafe-3.0.4-cp311-cp311-musllinux_1_2_armv7l.whl", hash = "sha256:83b3944fea42a8400edf92fd1770fb8d0d4f7de651353bd2d8525a92dba69a21"},
+    {file = "markupsafe-3.0.4-cp311-cp311-musllinux_1_2_ppc64le.whl", hash = "sha256:8138eb83940ec7299024d92d4dee45f601b9e6c5ffde9d25f4e35e326203c707"},
+    {file = "markupsafe-3.0.4-cp311-cp311-musllinux_1_2_riscv64.whl", hash = "sha256:811d02d5122171c1941357efd8f9bf4ffe907b7f0a1a4e729a880e4be3f46e3e"},
+    {file = "markupsafe-3.0.4-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:50b5bedc9ed8a94fc8857a42ef4f84a81ea88f8d4f05dc8705fb23ee6d8dcca7"},
+    {file = "markupsafe-3.0.4-cp311-cp311-win32.whl", hash = "sha256:2e5a7cd7fdd14fcb1ae5d7d8bf23d24fbd1daefd1fbca2580132e1ea75f098b5"},
+    {file = "markupsafe-3.0.4-cp311-cp311-win_amd64.whl", hash = "sha256:fdb4ca07ab75ffadab4a8b135ad59cdbb3156b99310f3d565370da74a15d6bd3"},
+    {file = "markupsafe-3.0.4-cp311-cp311-win_arm64.whl", hash = "sha256:569d65055d367e3dcdf30c3f41119467b73d9ee9faf332bdf40402644f5ac08e"},
+    {file = "markupsafe-3.0.4-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:61631e08084be9e21a8967ec3139c7616ed7c5e9368e05c86d1b39562c8a57b6"},
+    {file = "markupsafe-3.0.4-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:0930db9bdc62d22944e10b066448bb65dc9abe9112880c7cab8da54db4284d5f"},
+    {file = "markupsafe-3.0.4-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:6a45c3d514f2436064db00d7fc8778d888f0236ebfed649b53d13a59e69ad51b"},
+    {file = "markupsafe-3.0.4-cp312-cp312-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:1e1451fab512d1bcc3dc26988ec1edb0b82c2db909132872cd9356070a6b63df"},
+    {file = "markupsafe-3.0.4-cp312-cp312-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:bd3ce56ae2cbae3ba82b683bc425cd7e48d2ed8b10f3e818186b6f5646d9271c"},
+    {file = "markupsafe-3.0.4-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:8e124f974786f831d6043728e38296969d3579db8896fe004682f5758e613581"},
+    {file = "markupsafe-3.0.4-cp312-cp312-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:c02e8f18bdedba082cef725942ac823b9b60656db07f7e265cb31618dfd00d77"},
+    {file = "markupsafe-3.0.4-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:9f098115c247e11d138ab83a28fa0323c77015007ea2df73ba5fd714dfefd67c"},
+    {file = "markupsafe-3.0.4-cp312-cp312-musllinux_1_2_armv7l.whl", hash = "sha256:d5f93ebbeb8032d47e349328ec8662d973d9b05a70b3c35df1f91fe419b84749"},
+    {file = "markupsafe-3.0.4-cp312-cp312-musllinux_1_2_ppc64le.whl", hash = "sha256:64511c54db4e4987aef4c41923235927428729e8174c5dba488429be70a998ed"},
+    {file = "markupsafe-3.0.4-cp312-cp312-musllinux_1_2_riscv64.whl", hash = "sha256:e1a622f13970d81f95d0c72f9dc090dce9085fccfa4c9f2174377ee32bd15786"},
+    {file = "markupsafe-3.0.4-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:c9a7f43c0b202b334cc9184af09bb8f21d3a209e038efaf106936fb69e6b026e"},
+    {file = "markupsafe-3.0.4-cp312-cp312-win32.whl", hash = "sha256:f0ec3b750b59375eab5b0fb2b9254810c00a3375be6d789899f1055a1d556237"},
+    {file = "markupsafe-3.0.4-cp312-cp312-win_amd64.whl", hash = "sha256:11935df9bf455ed0c04eb87bcd720f02b1fe5e02128a9430f23aed6f93336fc7"},
+    {file = "markupsafe-3.0.4-cp312-cp312-win_arm64.whl", hash = "sha256:a4bbd2d87dd233b9fc5812160c3d0ffbe42edc22a26ce0469f58479ede633fe9"},
+    {file = "markupsafe-3.0.4-cp313-cp313-android_24_arm64_v8a.whl", hash = "sha256:de8b364c423ef0a4bad9069657d617f9a5d2b2062457a89b1fa16ee199c399c1"},
+    {file = "markupsafe-3.0.4-cp313-cp313-android_24_x86_64.whl", hash = "sha256:34bdde374c5932765d7dc685c4a1d191a3207852d67e8e0a9eb6ea85156181f1"},
+    {file = "markupsafe-3.0.4-cp313-cp313-ios_13_0_arm64_iphoneos.whl", hash = "sha256:6bd9e1788e15bfcf6a9082de42e30387e7b85d211ab21e57a939bb8cfaaf8d96"},
+    {file = "markupsafe-3.0.4-cp313-cp313-ios_13_0_arm64_iphonesimulator.whl", hash = "sha256:5066b244f576f91afc8ee3ba029a89f99d39c79b1853fe9d39bea9f0afbec148"},
+    {file = "markupsafe-3.0.4-cp313-cp313-ios_13_0_x86_64_iphonesimulator.whl", hash = "sha256:7a83aa6e4805df46fed18e989d3d16f86ef60cb50bbc8d9ce3a6be89165fbf6e"},
+    {file = "markupsafe-3.0.4-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:2d1b7d9308288661f56672b1b157d75fc536714d3638487bbea17b6318a78248"},
+    {file = "markupsafe-3.0.4-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:73e77980c7207854f00fc4e71fb1626868d5740ab4012623d55c7a99ad122a72"},
+    {file = "markupsafe-3.0.4-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:7018d4af1cd272e847aa5917983ab5e83e4f6579f9dbfecd4a79c0ca80b144c2"},
+    {file = "markupsafe-3.0.4-cp313-cp313-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:c90d5b3d4e944e065a301d741b3c1d784f6bd1f503aa68b4967e32b2ba313d85"},
+    {file = "markupsafe-3.0.4-cp313-cp313-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:18a801868a884f216e784d7d14db2a4077143ce7610440aee2ce8f734e7cfcde"},
+    {file = "markupsafe-3.0.4-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:434139499bb20b502ed3baa1f169e618f924a97e7a777fea1a49446d80106cf6"},
+    {file = "markupsafe-3.0.4-cp313-cp313-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:9e227f3dbe6bde7491cf0a9965d00b88c6b1a4a95d11480ddf88bb96d397c19f"},
+    {file = "markupsafe-3.0.4-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:b8cd1f918b26fd7b1832ece557cc18f2d8747309ff8b3f0ef9d4250c5ad67a39"},
+    {file = "markupsafe-3.0.4-cp313-cp313-musllinux_1_2_armv7l.whl", hash = "sha256:a5fcffb37e602b0b3c1638a97746b9b96125caa9bcf6fa41d337a9261de231ee"},
+    {file = "markupsafe-3.0.4-cp313-cp313-musllinux_1_2_ppc64le.whl", hash = "sha256:5989cb26b2e1efc6a42216a9f6b5ee495ce5ace2e5b352a9af489976b32d1ee2"},
+    {file = "markupsafe-3.0.4-cp313-cp313-musllinux_1_2_riscv64.whl", hash = "sha256:add96447a86d205ab616665d53b2950ee81083757f56e6ea833c8b2917646b46"},
+    {file = "markupsafe-3.0.4-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:2628d3a8cb648ecebb3c5d6b0a1052d400e4d8b7ac0fb786be8d285b50040d17"},
+    {file = "markupsafe-3.0.4-cp313-cp313-win32.whl", hash = "sha256:672d207103e6b16ca098611b0f9efad6bc00afd47c03d6ef62186495ca677dc0"},
+    {file = "markupsafe-3.0.4-cp313-cp313-win_amd64.whl", hash = "sha256:1f1f9477e174582b0a1b583d60b66e1f2cf5d3fe12cee985e4aedf44766600e5"},
+    {file = "markupsafe-3.0.4-cp313-cp313-win_arm64.whl", hash = "sha256:06de8ef6331f6e822c28d577dc8bf43fe398800477c49498f38fc38b67ff33fc"},
+    {file = "markupsafe-3.0.4-cp314-cp314-android_24_arm64_v8a.whl", hash = "sha256:4ed644d75aa94a2baf7ec3a96eaa160ea58c742eb9d27c6506053c5c40fc84ed"},
+    {file = "markupsafe-3.0.4-cp314-cp314-android_24_x86_64.whl", hash = "sha256:6d2a9efe686f9de00d0d1ea32a4a5a86d558a2277501bd78d964214eab625e59"},
+    {file = "markupsafe-3.0.4-cp314-cp314-ios_13_0_arm64_iphoneos.whl", hash = "sha256:8781a792a070cf2bd1b86d3aa943894115faaba6e88122a7bf32d62072742453"},
+    {file = "markupsafe-3.0.4-cp314-cp314-ios_13_0_arm64_iphonesimulator.whl", hash = "sha256:971a3bbb75d97ae4e2e8f7d4834236f86f85f0c85e04ab2e191db1123b04f80b"},
+    {file = "markupsafe-3.0.4-cp314-cp314-ios_13_0_x86_64_iphonesimulator.whl", hash = "sha256:8909c2f1c6dd65e054ac4b573a91c8384d1492281e55d82d159d653f7a13adf6"},
+    {file = "markupsafe-3.0.4-cp314-cp314-macosx_10_15_x86_64.whl", hash = "sha256:4cf3468d5ec187ffffcaca8e61929a37448f215dafc1386a12c750a72fe53634"},
+    {file = "markupsafe-3.0.4-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:52704c5d36eb6dda8866493decd61111fff86244c9b1ad225ca01b9e91e5970f"},
+    {file = "markupsafe-3.0.4-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:1caa2fa5a6184fb233153b35f654e6687bd555476f6170f29d8ee9be1a8b0af9"},
+    {file = "markupsafe-3.0.4-cp314-cp314-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:387d8cd30e69b3f0a72877b9ae717033396404e19095b17fe89753a981fda44f"},
+    {file = "markupsafe-3.0.4-cp314-cp314-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:051417f74bcaaefa316276e0ff723f541616ca51043d070da00249d9bddd3e3c"},
+    {file = "markupsafe-3.0.4-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:a8e9f292fcda89b324f2f5c91d13f1424a153e40fc2756f38ee23b15835ff300"},
+    {file = "markupsafe-3.0.4-cp314-cp314-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:df1ae86ff54725a01fa1a0510b914ca53a161b7050be74f6204e24aded5971d0"},
+    {file = "markupsafe-3.0.4-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:8965520ac587c94a4ac48b729be3d8b8de00af39699b17585dfb599babe77977"},
+    {file = "markupsafe-3.0.4-cp314-cp314-musllinux_1_2_armv7l.whl", hash = "sha256:340cbb1957ba99929cbf19a75626d36ba1ae21d1730b287d1cf7f824a20c4fc7"},
+    {file = "markupsafe-3.0.4-cp314-cp314-musllinux_1_2_ppc64le.whl", hash = "sha256:3a93d9616ddecfb393727a0041a562cf0b15a244e20f2bd25efc7949be4c4f17"},
+    {file = "markupsafe-3.0.4-cp314-cp314-musllinux_1_2_riscv64.whl", hash = "sha256:d2e56fd3b00222722abfb3f5f0759ddbae4b90811b5ad4343c64030ad1bde70c"},
+    {file = "markupsafe-3.0.4-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:0d9c47709875fdb321452056622e930c52afbc07a7d780762fbb8b4d91ce6fa4"},
+    {file = "markupsafe-3.0.4-cp314-cp314-win32.whl", hash = "sha256:38fc55594dab834470b6733dead2ee9e3f657fb0608c769dcafa0ba5ab52f45c"},
+    {file = "markupsafe-3.0.4-cp314-cp314-win_amd64.whl", hash = "sha256:c1bc67752d5f21013cfe430df4062441714eab79f65a6a05e01505957e9c35fe"},
+    {file = "markupsafe-3.0.4-cp314-cp314-win_arm64.whl", hash = "sha256:7e1636da3d8dfc220b6dd10264db5f2b165e4888c4518594898fbe381049af8a"},
+    {file = "markupsafe-3.0.4-cp314-cp314t-macosx_10_15_x86_64.whl", hash = "sha256:805c8b84534fa10891890f0e4be39f3a99e94615d93e8836bf9fa1fdca2feeb2"},
+    {file = "markupsafe-3.0.4-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:fa95848c929b6a75f6848d3c9793e59db365ee436776e57db835cdbfa79ba977"},
+    {file = "markupsafe-3.0.4-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:e916035e3e9930cbdfdd10abf48861340221857f45509565898e012263f7b289"},
+    {file = "markupsafe-3.0.4-cp314-cp314t-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:b4d12837e0203bbace818ff4a7461afdcd78bcd782351cea148139180d7bcffe"},
+    {file = "markupsafe-3.0.4-cp314-cp314t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:5086f9975abb1ab531ee6afca1761e4b59a19b446f3f6522ed776963228cfe5a"},
+    {file = "markupsafe-3.0.4-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:b4a635a0487774f841cb1fb62e907e7195cc95bc761e053184b8acc3ceb20733"},
+    {file = "markupsafe-3.0.4-cp314-cp314t-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:cb96e6e088d6cf71c1ea977510948320234824cf226e32f6f6e044f7a9c82b34"},
+    {file = "markupsafe-3.0.4-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:8b5d563170ff8ba3181caa967c99a3c804d1dedb702c7cb93a6a7c32247da978"},
+    {file = "markupsafe-3.0.4-cp314-cp314t-musllinux_1_2_armv7l.whl", hash = "sha256:396ec4e65cc889f69786b3b89478b471cee5a3bcf468b9d9bb03e1a30fb291fc"},
+    {file = "markupsafe-3.0.4-cp314-cp314t-musllinux_1_2_ppc64le.whl", hash = "sha256:15ba9e28640feef770374b116a6f019c21f52404aeabe516aa7f800587b98cfc"},
+    {file = "markupsafe-3.0.4-cp314-cp314t-musllinux_1_2_riscv64.whl", hash = "sha256:d920abdfa61279ba1a2ef9484aab07bf03331f8c08a10120fa332353d06e6932"},
+    {file = "markupsafe-3.0.4-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:a9f54054101545a9a9cccefddf54316aa6e4491611fcbef9e91b3b6bebec04f6"},
+    {file = "markupsafe-3.0.4-cp314-cp314t-win32.whl", hash = "sha256:12a606a492de952afcb43b59a14aaaaad120e708d3663dd0fdf2d738d427a691"},
+    {file = "markupsafe-3.0.4-cp314-cp314t-win_amd64.whl", hash = "sha256:a18f38cafc329bac5e3c2b96c765b4c96d3d103421ed22ab7988c1e3fce27464"},
+    {file = "markupsafe-3.0.4-cp314-cp314t-win_arm64.whl", hash = "sha256:eba154571c16e032112afac0dc2dfe9e63c2ceb7aedd07bb7eecf2ce26d4dd4c"},
+    {file = "markupsafe-3.0.4-cp315-cp315-android_24_arm64_v8a.whl", hash = "sha256:737c9c3981998eba27f11786f84fddcbabc74068b72a4a1f454ea02094b57b65"},
+    {file = "markupsafe-3.0.4-cp315-cp315-android_24_x86_64.whl", hash = "sha256:489505b03f692c3f376394e49194fa7a7f9e8558d6e293a7056a0032b0c38163"},
+    {file = "markupsafe-3.0.4-cp315-cp315-ios_13_0_arm64_iphoneos.whl", hash = "sha256:077293e425f28ec737dbcad442a71752e28f8ae27cde3d68acd1fb212091cd92"},
+    {file = "markupsafe-3.0.4-cp315-cp315-ios_13_0_arm64_iphonesimulator.whl", hash = "sha256:9348cbb300d224fe3b89793262cb093504d4ae927004468463f745188a193e4a"},
+    {file = "markupsafe-3.0.4-cp315-cp315-ios_13_0_x86_64_iphonesimulator.whl", hash = "sha256:b807e598953730f82e4eae3bd30f6a122cf6b31c398c6b504c0e04c13c170429"},
+    {file = "markupsafe-3.0.4-cp315-cp315-macosx_10_15_x86_64.whl", hash = "sha256:799c39bdf5e2f1292fedd3009f7b3c9e760f10b2420cb9638d56920840ff6db8"},
+    {file = "markupsafe-3.0.4-cp315-cp315-macosx_11_0_arm64.whl", hash = "sha256:ae9dcb8fbe244cb82f8a6458b455b927a03685e383d9bacf1ea5ce180b96dc97"},
+    {file = "markupsafe-3.0.4-cp315-cp315-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:4bced6e2a6dba6a28f7dd3c6ce14df1b2dd495923f16ea484cad03decd463b2b"},
+    {file = "markupsafe-3.0.4-cp315-cp315-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:3882fb412298575bae3b9c46868251f15cc69307359f87bb1b382e53d6e5a2c9"},
+    {file = "markupsafe-3.0.4-cp315-cp315-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:04e7902ba80ee4bac1d50a549606527a1dcf0476cd81403db41099d3b60ec653"},
+    {file = "markupsafe-3.0.4-cp315-cp315-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:925f929d6b59a8b3f8b8c6ac363cd0af7eecc81efb3071770b3c6717c450a369"},
+    {file = "markupsafe-3.0.4-cp315-cp315-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:f68edfc67aabac33708941f26f22a7b8e9f81429bc0cf249fcf7d66b23af8d19"},
+    {file = "markupsafe-3.0.4-cp315-cp315-musllinux_1_2_aarch64.whl", hash = "sha256:e5c802729725bd07e2bc3ab7b76dc7e0bbfc53129d8f1eb1c002c24cf774717e"},
+    {file = "markupsafe-3.0.4-cp315-cp315-musllinux_1_2_armv7l.whl", hash = "sha256:55ffd6ce583d97dc71dc92e930324c8c0d25aea7e3ade6ae54ef77cedb096811"},
+    {file = "markupsafe-3.0.4-cp315-cp315-musllinux_1_2_ppc64le.whl", hash = "sha256:2cb3dd71fc6be918ad4264346a8ed69485f9b7ed7bf35495d8e22807cd6b8bea"},
+    {file = "markupsafe-3.0.4-cp315-cp315-musllinux_1_2_riscv64.whl", hash = "sha256:94f5407f7bc64fa6463906b896f9904beeeb7dd8dc116ee8e9056c8714ff9916"},
+    {file = "markupsafe-3.0.4-cp315-cp315-musllinux_1_2_x86_64.whl", hash = "sha256:2dad610540cb2e6272855c178f08ae9a1c7ac258a7fb71660553a5f104b42741"},
+    {file = "markupsafe-3.0.4-cp315-cp315-win32.whl", hash = "sha256:03470d1a8268e692ecf79ecd565593e59d44219377a7ead61f1f1b94c1f7ff6b"},
+    {file = "markupsafe-3.0.4-cp315-cp315-win_amd64.whl", hash = "sha256:d882a373d8093c2941e01291b7ced96e9cbe4781da9a7751ca7e6c70385e5214"},
+    {file = "markupsafe-3.0.4-cp315-cp315-win_arm64.whl", hash = "sha256:353bd63081912ab8cfa6a0c7d185934cdf8426f04c618bba6bc4b394f2069b67"},
+    {file = "markupsafe-3.0.4-cp315-cp315t-macosx_10_15_x86_64.whl", hash = "sha256:c61750fadcd119d0825bcb7d7d675dd264dcc89cc05292aab5be68ebdbb374ad"},
+    {file = "markupsafe-3.0.4-cp315-cp315t-macosx_11_0_arm64.whl", hash = "sha256:1c0df495a977d10460a94941799c72d5b5ab03d3858d949b55b5a66c8f371c99"},
+    {file = "markupsafe-3.0.4-cp315-cp315t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:02fa4acbc6a3fc5c693c34d4dd8c1130b7fe99cc915181b0ddd6f72aeb296002"},
+    {file = "markupsafe-3.0.4-cp315-cp315t-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:05295589e619b9bed252a86b532b8e27350abc372d18ba89b59375325e91ec1e"},
+    {file = "markupsafe-3.0.4-cp315-cp315t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:be6cb0c799abb0e2ba3e618e6d28ddddf7e485f6c2ce938dfa237daf3905072c"},
+    {file = "markupsafe-3.0.4-cp315-cp315t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:26e9867520db70d37f7fb421a7f0d8adb40171011fb84ce869afa1a83370dfa8"},
+    {file = "markupsafe-3.0.4-cp315-cp315t-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:f03460ff076f70ab595bb45a0205ccea1971443575b6920c52e755dec2b3fbfe"},
+    {file = "markupsafe-3.0.4-cp315-cp315t-musllinux_1_2_aarch64.whl", hash = "sha256:436e3ffc6310d3c41878c601db29098102fe5d8a467c49da4a4125254e0980f2"},
+    {file = "markupsafe-3.0.4-cp315-cp315t-musllinux_1_2_armv7l.whl", hash = "sha256:4e2c4809c14559aa7ef426f27fb35afbb38104c349a903bf8f3600456764bb38"},
+    {file = "markupsafe-3.0.4-cp315-cp315t-musllinux_1_2_ppc64le.whl", hash = "sha256:da2af0d7aebfc2074080d72efa6ab8317c62481ef1f896f65d9999c1c01f4494"},
+    {file = "markupsafe-3.0.4-cp315-cp315t-musllinux_1_2_riscv64.whl", hash = "sha256:aa2c838cc024642cc04c6854232f32b43e5e22833dd11119c1766c7873b8370d"},
+    {file = "markupsafe-3.0.4-cp315-cp315t-musllinux_1_2_x86_64.whl", hash = "sha256:b91cc9d336957239ff200f30097e6fea2dc6d6fb3c81e853eaa09eac904fd894"},
+    {file = "markupsafe-3.0.4-cp315-cp315t-win32.whl", hash = "sha256:e49fb0d1ce92cfa0cb198cc5b1b11cdf9d0638658e2a2db2687e39db7c87fc78"},
+    {file = "markupsafe-3.0.4-cp315-cp315t-win_amd64.whl", hash = "sha256:4f6e0852a0283b1b1fd776eeb7b766a5f440b3e2bd31ab51af3b400585f3965c"},
+    {file = "markupsafe-3.0.4-cp315-cp315t-win_arm64.whl", hash = "sha256:39dbacefc411633db5b4378b066a9aca70a3d7e2922c9e578d825f844026eeba"},
+    {file = "markupsafe-3.0.4-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:f291bcf42ae98eb5107edb162c3c998b4a89648fd8e99ed4cbd12705292788cd"},
+    {file = "markupsafe-3.0.4-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:ac0c7c9f1609b0c4c114feb1d7a3409564c7fb77e360bed9e97e5d25dfeaf868"},
+    {file = "markupsafe-3.0.4-cp39-cp39-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:6768d67d1bce64270e0fdc2e69309d68b9b18ae56ddf6c711d168e9d051c2cac"},
+    {file = "markupsafe-3.0.4-cp39-cp39-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:14bd2d845d62ab678eaf81da89d7b621b51756c72346745c1a594c09d49207a2"},
+    {file = "markupsafe-3.0.4-cp39-cp39-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:007e1ffd9bf65bb6ee96df7b258fc632a4868dd5566037986c64781f35a36e98"},
+    {file = "markupsafe-3.0.4-cp39-cp39-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:5e8b3d0b18fd623afa12ecb2ce8d8becef69f9b5440c6330c7972200e0bb84b0"},
+    {file = "markupsafe-3.0.4-cp39-cp39-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:57f9947a7e57a081c1e3e0a2dd0d2dcf290a4531450e6f611e30084c222a7295"},
+    {file = "markupsafe-3.0.4-cp39-cp39-musllinux_1_2_aarch64.whl", hash = "sha256:b61687d0828e72bf5cda24a2690188f37170bd31c9359ac97e4e66569f120a16"},
+    {file = "markupsafe-3.0.4-cp39-cp39-musllinux_1_2_armv7l.whl", hash = "sha256:0cee7cb0f9a1b6892ea482237d9403b3d1b4603aee057d0ff01f0fac2d019a97"},
+    {file = "markupsafe-3.0.4-cp39-cp39-musllinux_1_2_ppc64le.whl", hash = "sha256:94e4c421742086aeee4c32a506eec8859d7634aad943f7e6aacf70f813478768"},
+    {file = "markupsafe-3.0.4-cp39-cp39-musllinux_1_2_riscv64.whl", hash = "sha256:9240187afb63d2f9ddc3e032c670356fe941f6e20662ea168a5dc3f1f317e1b3"},
+    {file = "markupsafe-3.0.4-cp39-cp39-musllinux_1_2_x86_64.whl", hash = "sha256:e841068dc0be4cb6dfb5c890eb88cbdcff2f4a332393c7ec94e8e618bd32c1a8"},
+    {file = "markupsafe-3.0.4-cp39-cp39-win32.whl", hash = "sha256:f61efe1d2fe0de16158a5fe1d1cf3c14bdb6aecd54d8938fd26512c525c1f624"},
+    {file = "markupsafe-3.0.4-cp39-cp39-win_amd64.whl", hash = "sha256:2b2b1e18af909b448bb3cf9e3433366f7a8726271fc214e8b10e0f62a78c724b"},
+    {file = "markupsafe-3.0.4-cp39-cp39-win_arm64.whl", hash = "sha256:6669c1bf34080161ce49c589cc512ef24d4c704ac9d2b2d3667f519c60418378"},
+    {file = "markupsafe-3.0.4.tar.gz", hash = "sha256:2e9ad7dd851bf45fab9f75cbff4cb493fee9979e8d8c7c9c3ee119022518edd6"},
+]
+
+[[package]]
+name = "mdurl"
+version = "0.1.2"
+description = "Markdown URL utilities"
+optional = false
+python-versions = ">=3.7"
+groups = ["main"]
+files = [
+    {file = "mdurl-0.1.2-py3-none-any.whl", hash = "sha256:84008a41e51615a49fc9966191ff91509e3c40b939176e643fd50a5c2196b8f8"},
+    {file = "mdurl-0.1.2.tar.gz", hash = "sha256:bb413d29f5eea38f31dd4754dd7377d4465116fb207585f97bf925588687c1ba"},
+]
+
+[[package]]
+name = "mpmath"
+version = "1.3.0"
+description = "Python library for arbitrary-precision floating-point arithmetic"
+optional = false
+python-versions = "*"
+groups = ["main"]
+markers = "sys_platform == \"linux\" or sys_platform == \"darwin\""
+files = [
+    {file = "mpmath-1.3.0-py3-none-any.whl", hash = "sha256:a0b2b9fe80bbcd81a6647ff13108738cfb482d481d826cc0e02f5b35e5c88d2c"},
+    {file = "mpmath-1.3.0.tar.gz", hash = "sha256:7a28eb2a9774d00c7bc92411c19a89209d5da7c4c9a9e227be8330a23a25b91f"},
+]
+
+[package.extras]
+develop = ["codecov", "pycodestyle", "pytest (>=4.6)", "pytest-cov", "wheel"]
+docs = ["sphinx"]
+gmpy = ["gmpy2 (>=2.1.0a4) ; platform_python_implementation != \"PyPy\""]
+tests = ["pytest (>=4.6)"]
+
+[[package]]
+name = "networkx"
+version = "3.7"
+description = "Python package for creating and manipulating graphs and networks"
+optional = false
+python-versions = "!=3.14.1,>=3.12"
+groups = ["main"]
+markers = "sys_platform == \"linux\" or sys_platform == \"darwin\""
+files = [
+    {file = "networkx-3.7-py3-none-any.whl", hash = "sha256:e3fd2c13a7814cee3746340d8d7f8598a67f16a58bf47fb7f8793fab6efca1b0"},
+    {file = "networkx-3.7.tar.gz", hash = "sha256:fd77a511bd90f39f3d016351345b52cf5319b813bdca01de3f755d3cca62e96a"},
+]
+
+[package.extras]
+benchmarking = ["asv (>=0.6.5)", "py-rattler (>=0.21,<0.22)", "virtualenv"]
+default = ["matplotlib (>=3.10)", "numpy (>=2.2)", "pandas (>=2.3)", "scipy (>=1.15)"]
+developer = ["mypy (>=1.15)", "pre-commit (>=4.1)"]
+doc = ["intersphinx-registry", "myst-nb (>=1.1)", "numpydoc (>=1.8.0)", "pillow (>=10)", "pydata-sphinx-theme (>=0.16)", "sphinx (>=8.0)", "sphinx-gallery (>=0.18)", "texext (>=0.6.7)"]
+example = ["cairocffi (>=1.7)", "contextily (>=1.6)", "hiveplotlib (>=0.28)", "igraph (>=0.11)", "iplotx (>=0.9.0)", "momepy (>=0.7.2)", "osmnx (>=2.0.0)", "scikit-learn (>=1.6)", "seaborn (>=0.13)"]
+extra = ["lxml (>=4.6)", "pydot (>=3.0.1)", "pygraphviz (>=2.0)", "sympy (>=1.10)"]
+release = ["build (>=0.10)", "changelist (==0.5)", "twine (>=4.0)", "wheel (>=0.40)"]
+test = ["pytest (>=7.2)", "pytest-cov (>=4.0)", "pytest-xdist (>=3.0)"]
+test-extras = ["pytest-mpl", "pytest-randomly"]
+
+[[package]]
+name = "numpy"
+version = "2.5.4"
+description = "Fundamental package for array computing in Python"
+optional = false
+python-versions = ">=3.12"
+groups = ["main"]
+files = [
+    {file = "numpy-2.5.4-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:c6342f54c67093cae5c0227eb0eb772fdb79f2a2c37a6eb278b9909ee06aa356"},
+    {file = "numpy-2.5.4-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:b11e8fda06a7d69f15ebf542660b74466c2e51094800c1fb794f47ad4faeef17"},
+    {file = "numpy-2.5.4-cp312-cp312-macosx_14_0_arm64.whl", hash = "sha256:9cb18a327b49c5c337f972b03682f6a49855525faaf3c0d3e9c96cd0fd8880a8"},
+    {file = "numpy-2.5.4-cp312-cp312-macosx_14_0_x86_64.whl", hash = "sha256:aec3fc4b32ff82421274f5d205c559c51c840c8df66a78efd7f3612dd005a26a"},
+    {file = "numpy-2.5.4-cp312-cp312-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:fe4d21ab149f15e4e6043dfb0de87e6e5f34ac176cde83060e9802981fca2ac2"},
+    {file = "numpy-2.5.4-cp312-cp312-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:fbde6962867ee75b48b0ee29b2b9372ec5d617799dbaf38e82dc0596f2f7738a"},
+    {file = "numpy-2.5.4-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:381a7a3d2e65e64c0ec302795ab9dc12bb1e73f150904699c153716177eebdaf"},
+    {file = "numpy-2.5.4-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:b89d0aaae2fe498c648f4c4795c084db535af5bd98ef942b2a3681fb74ce8645"},
+    {file = "numpy-2.5.4-cp312-cp312-win32.whl", hash = "sha256:9968ab7e49b93ac6e1c3b2239732183152c9150f16308d30b66a372cffe3483c"},
+    {file = "numpy-2.5.4-cp312-cp312-win_amd64.whl", hash = "sha256:a7b1b6353e36a7e50de2973a38d705c88ee93adcf120673cee7f45a4a3fa223a"},
+    {file = "numpy-2.5.4-cp312-cp312-win_arm64.whl", hash = "sha256:aa1cce2ff3f8d953de38b76bf44602caeb69f101430208f64a10067f7cb4b1d3"},
+    {file = "numpy-2.5.4-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:2377da2dd3ba2c1200956acbab2a358c83b8e1f8531191672d1cd6ad83250d53"},
+    {file = "numpy-2.5.4-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:7415db95818b39ec475a5eea54d9e3b6bc83e3912158e46da3438cdce399804d"},
+    {file = "numpy-2.5.4-cp313-cp313-macosx_14_0_arm64.whl", hash = "sha256:6d6a71b9d9a97c03633aa12565ef2825ffa036cc1d99cfd50dacf0f128af4fe2"},
+    {file = "numpy-2.5.4-cp313-cp313-macosx_14_0_x86_64.whl", hash = "sha256:d8200f16437b289a5bb927c6e184eccc3e8389bc0070fea4cd5b9e13c1757959"},
+    {file = "numpy-2.5.4-cp313-cp313-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:1c2e71b04c6cad90026e544501bbe0ab9290fa8a4d845e7e8c0d124fb429c988"},
+    {file = "numpy-2.5.4-cp313-cp313-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:6ffa07666f8da0eef81d149934a626d0d95fbd6838432a33e66245423a9062c0"},
+    {file = "numpy-2.5.4-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:2fa3328f784fc8277fc48026f6cad516f5c561c5d8e2e39b3c9e0c8f23223b34"},
+    {file = "numpy-2.5.4-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:b86966fbe4ad7de710422175572bcdc75fdedadfb54bc6fab7deabccddd7780b"},
+    {file = "numpy-2.5.4-cp313-cp313-win32.whl", hash = "sha256:5258bc06526964be5face2fc6f756857a3f24f21ec3e72ca131337a75b165d6c"},
+    {file = "numpy-2.5.4-cp313-cp313-win_amd64.whl", hash = "sha256:8b4d2fd2d34e5f8c9235ee787de5631a37a28402b15cb80814df973d2be54129"},
+    {file = "numpy-2.5.4-cp313-cp313-win_arm64.whl", hash = "sha256:bc39ac66a7a9a3fbd6134fda43136b60ffde99c8f4501e64e0d2b24da137babf"},
+    {file = "numpy-2.5.4-cp314-cp314-macosx_10_15_x86_64.whl", hash = "sha256:c668b2f0d651605b58892644b0e302c7157f7159544227758c896982ef384b18"},
+    {file = "numpy-2.5.4-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:ffa6ce09a1c6a08e9667dd9c97aa0b14184e8d18f2a14b78b2a2328c9147f076"},
+    {file = "numpy-2.5.4-cp314-cp314-macosx_14_0_arm64.whl", hash = "sha256:956555e0603a4d38019ae6925711cb9dc43195c076a928accf7ea5d50bddfe53"},
+    {file = "numpy-2.5.4-cp314-cp314-macosx_14_0_x86_64.whl", hash = "sha256:2c2c4afffdeb7920e445028dd71eb932cac3e704792e964bc2a232426d4f1255"},
+    {file = "numpy-2.5.4-cp314-cp314-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:4054173604cd8658796053f1f3bc0befb68ec1c0762c57fdad61e199256a8617"},
+    {file = "numpy-2.5.4-cp314-cp314-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:d549420b8858885cea8838a727842249218b9c1da24dd517e25c9c7a948310a3"},
+    {file = "numpy-2.5.4-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:823874a507a84af050493b622affde94b6f7c3a0dc22cb2801381bc03b871c00"},
+    {file = "numpy-2.5.4-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:4e263278bfb5ee6409db8aedbc4cc32973b1b82bc1e8d3c668551d04d83a7e37"},
+    {file = "numpy-2.5.4-cp314-cp314-win32.whl", hash = "sha256:cfd73180400042a7c532d30c5e287bdd03c59ff9ee1b4c0316af0539e29dfe23"},
+    {file = "numpy-2.5.4-cp314-cp314-win_amd64.whl", hash = "sha256:2ca144f15135b6212a5c47b1e2aeca6e412f102f95a2d5d88d8aec77eb255de3"},
+    {file = "numpy-2.5.4-cp314-cp314-win_arm64.whl", hash = "sha256:468397ba3c64427474706e5c9123fe266395496714dc684294eac75cd4930d1e"},
+    {file = "numpy-2.5.4-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:1ef3aa6d7e29bb13677323114280b05acc57607fa2300e66432d665d5418a162"},
+    {file = "numpy-2.5.4-cp314-cp314t-macosx_14_0_arm64.whl", hash = "sha256:98b053943e5a0474ec0da309d2cb9d3f18ea57f8a2067c2ab7b5f763d1068380"},
+    {file = "numpy-2.5.4-cp314-cp314t-macosx_14_0_x86_64.whl", hash = "sha256:b64a85f40e154983960a4167d4c1d57a50c7f109b3d3264a3a984154e90a8454"},
+    {file = "numpy-2.5.4-cp314-cp314t-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:a813ed7719bf45463c51779e6a98d0385fe905e48447526938a4b8337333d551"},
+    {file = "numpy-2.5.4-cp314-cp314t-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:c9b80cdf5cedba0e90d93fa5f9a333c4d65bd545cd669b71bb97ce2b703c9d73"},
+    {file = "numpy-2.5.4-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:2199ed071f460487c8db2c0e5c0b564494190edb4772fe80f9aad88b2604def5"},
+    {file = "numpy-2.5.4-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:64f9c9878c1938476365e11ccfb6b770f3b9e5f045ccddc514235041e6959365"},
+    {file = "numpy-2.5.4-cp314-cp314t-win32.whl", hash = "sha256:64d1c8ac28a4077cf987e0a71a7a0ef7e2df70722f07f0baa42dbb7eb6938647"},
+    {file = "numpy-2.5.4-cp314-cp314t-win_amd64.whl", hash = "sha256:067374eb538c34c745436365cf7b0112595c1d326f21ce4ff340f61230239fbb"},
+    {file = "numpy-2.5.4-cp314-cp314t-win_arm64.whl", hash = "sha256:e94aef2c639da4a960ad0db8e06471208d8589974953d78b61d345b4eb99e394"},
+    {file = "numpy-2.5.4-cp315-cp315-macosx_10_15_x86_64.whl", hash = "sha256:8dddfbee2e68d26d0d7d7d9cb247b1fd4409241cce32d815a11d97ec2cfde179"},
+    {file = "numpy-2.5.4-cp315-cp315-macosx_11_0_arm64.whl", hash = "sha256:81e3420b27048b65eb14c3acf0c174a8cb0e023277716110347d2dcb26026dad"},
+    {file = "numpy-2.5.4-cp315-cp315-macosx_14_0_arm64.whl", hash = "sha256:0b4724a19de67bea8cfc4970798efa78bcbbe2ac2613cfac16721a42d44de2a5"},
+    {file = "numpy-2.5.4-cp315-cp315-macosx_14_0_x86_64.whl", hash = "sha256:2132418bf8dd124a427ca9e6a1daf9ee1a87185344c95119ceae868b99466da1"},
+    {file = "numpy-2.5.4-cp315-cp315-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:325518d4245b9e331387702aa58c2ce1dc4cdcbb41dfb4ccd5dcbc7e08db1266"},
+    {file = "numpy-2.5.4-cp315-cp315-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:56733449d2544178beaa4545cee357370440cf056c197f9c7bfb19dbfdd0e86d"},
+    {file = "numpy-2.5.4-cp315-cp315-musllinux_1_2_aarch64.whl", hash = "sha256:5ec3753760c1a6d8bb91200666e545c3a9728e6269dfb5d6ce02340996698aa3"},
+    {file = "numpy-2.5.4-cp315-cp315-musllinux_1_2_x86_64.whl", hash = "sha256:b1185012870173de7ae33d370bd45b1cf5baee747ea4b97036b65f4e93016877"},
+    {file = "numpy-2.5.4-cp315-cp315-win32.whl", hash = "sha256:298eca75243f2cbbfdb460560b9fb2a1792a33cf2ab4286efd43d92e8d3df508"},
+    {file = "numpy-2.5.4-cp315-cp315-win_amd64.whl", hash = "sha256:332f3378fe077dd850e677ec01bdcc4f22368fb5d50ef10b2c79230b1bf5a592"},
+    {file = "numpy-2.5.4-cp315-cp315-win_arm64.whl", hash = "sha256:d4cccbbc78717966f764cd3af4fb70276fa01fc7a2688af11c78901fa5c04f05"},
+    {file = "numpy-2.5.4-cp315-cp315t-macosx_10_15_x86_64.whl", hash = "sha256:950ea81d57ef070665581b6e1b5f6a029306423cd1739c5b95fe78aa30db6b9d"},
+    {file = "numpy-2.5.4-cp315-cp315t-macosx_11_0_arm64.whl", hash = "sha256:c05ede731b03fb1b7591faca9389ade3267d2bddf1ad8882bb3f2cc5e101694f"},
+    {file = "numpy-2.5.4-cp315-cp315t-macosx_14_0_arm64.whl", hash = "sha256:5fbf7141bbfd63aea22f435c9062a032b9ea0082fe9845dad7f021d3f1234e71"},
+    {file = "numpy-2.5.4-cp315-cp315t-macosx_14_0_x86_64.whl", hash = "sha256:3573cd22564692a5b899ec344e5d5b9cc4576f2985b96f22af3564ed54f2710f"},
+    {file = "numpy-2.5.4-cp315-cp315t-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:6c109eac9cd439193678f69d70733c1108487546ca8eafc107b510ae10c1aecd"},
+    {file = "numpy-2.5.4-cp315-cp315t-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:80d6ef6e8620eb2c2b4c4caad50b5935d6db3cde2d51581b55dcc79e14016d1d"},
+    {file = "numpy-2.5.4-cp315-cp315t-musllinux_1_2_aarch64.whl", hash = "sha256:77045a4b175bbf5316ec08003880804336c78f92281a1b72222b274ea85ec5ac"},
+    {file = "numpy-2.5.4-cp315-cp315t-musllinux_1_2_x86_64.whl", hash = "sha256:0f02a46e49cfb6c73bdb7aea1c0d3461dbae9aba613542b65f657cd3d17b9fab"},
+    {file = "numpy-2.5.4-cp315-cp315t-win32.whl", hash = "sha256:ad62a416ddcf863bf44bba76fbf6b53366ab0692e294f51cae4b5fbe0d246788"},
+    {file = "numpy-2.5.4-cp315-cp315t-win_amd64.whl", hash = "sha256:38f47be9f74ab870d2633b5456ae519c43758a8d1fd05342f0ce4ecc034396ee"},
+    {file = "numpy-2.5.4-cp315-cp315t-win_arm64.whl", hash = "sha256:7a14a461d9340f1b46b8648578aed9cdb8b3b018a8fac6c1dde2c9192a01a87f"},
+    {file = "numpy-2.5.4.tar.gz", hash = "sha256:9a94cf751c9ad8ebaa835bcd3d40dacf8534ad086b88c38029b65123c7999d2a"},
+]
+
+[[package]]
+name = "nvidia-cublas"
+version = "13.1.0.3"
+description = "CUBLAS native runtime libraries"
+optional = false
+python-versions = ">=3"
+groups = ["main"]
+markers = "sys_platform == \"darwin\" and platform_system == \"Linux\""
+files = [
+    {file = "nvidia_cublas-13.1.0.3-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:c86fc7f7ae36d7528288c5d88098edcb7b02c633d262e7ddbb86b0ad91be5df2"},
+    {file = "nvidia_cublas-13.1.0.3-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:ee8722c1f0145ab246bccb9e452153b5e0515fd094c3678df50b2a0888b8b171"},
+    {file = "nvidia_cublas-13.1.0.3-py3-none-win_amd64.whl", hash = "sha256:2a3b94a37def342471c59fad7856caee4926809a72dd5270155d6a31b5b277be"},
+]
+
+[[package]]
+name = "nvidia-cudnn-cu13"
+version = "9.24.0.43"
+description = "cuDNN runtime libraries"
+optional = false
+python-versions = ">=3"
+groups = ["main"]
+markers = "sys_platform == \"darwin\" and platform_system == \"Linux\""
+files = [
+    {file = "nvidia_cudnn_cu13-9.24.0.43-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:a6812a554a1ff0413e9c52b84c26c050380649ab9615f9c16bded368ce9f421f"},
+    {file = "nvidia_cudnn_cu13-9.24.0.43-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:71f181cd810e90f9b6023b01186fe82d13d65f0ec098581ee201d39fad769e4b"},
+    {file = "nvidia_cudnn_cu13-9.24.0.43-py3-none-win_amd64.whl", hash = "sha256:67a7273b5cf062f9446fd76cf464351a1c0f66501e6cd78f6675c0d604d8ac87"},
+]
+
+[package.dependencies]
+nvidia-cublas = "*"
+
+[[package]]
+name = "nvidia-cusparselt-cu13"
+version = "0.8.1"
+description = "NVIDIA cuSPARSELt"
+optional = false
+python-versions = "*"
+groups = ["main"]
+markers = "sys_platform == \"darwin\" and platform_system == \"Linux\""
+files = [
+    {file = "nvidia_cusparselt_cu13-0.8.1-py3-none-manylinux2014_aarch64.whl", hash = "sha256:4dca476c50bf4780d46cd0bfbd82e2bc10a08e4fef7950917ce8d7578d22a23f"},
+    {file = "nvidia_cusparselt_cu13-0.8.1-py3-none-manylinux2014_x86_64.whl", hash = "sha256:786ce87568c303fadb5afcc7102d454cd3040d75f6f8626f5db460d1871f4dd0"},
+    {file = "nvidia_cusparselt_cu13-0.8.1-py3-none-win_amd64.whl", hash = "sha256:dccbd362f91a7b9024d1f55ee9f548ac065027ff15d8c8b0db889ab3a8f31215"},
+]
+
+[[package]]
+name = "nvidia-nccl-cu13"
+version = "2.30.7"
+description = "NVIDIA Collective Communication Library (NCCL) Runtime"
+optional = false
+python-versions = ">=3"
+groups = ["main"]
+markers = "sys_platform == \"darwin\" and platform_system == \"Linux\""
+files = [
+    {file = "nvidia_nccl_cu13-2.30.7-py3-none-manylinux_2_18_aarch64.whl", hash = "sha256:ca786ffa5a647c75d4d1f5cc72a6c4f537947e2ba8823d7c8aaf768e7a7b9f77"},
+    {file = "nvidia_nccl_cu13-2.30.7-py3-none-manylinux_2_18_x86_64.whl", hash = "sha256:cefa7fdb9710efd0f39c5f1be1d61ff6fc9a996c451265bd7fbdcf9455ed4b50"},
+]
+
+[[package]]
+name = "nvidia-nvshmem-cu13"
+version = "3.4.5"
+description = "NVSHMEM creates a global address space that provides efficient and scalable communication for NVIDIA GPU clusters."
+optional = false
+python-versions = ">=3"
+groups = ["main"]
+markers = "sys_platform == \"darwin\" and platform_system == \"Linux\""
+files = [
+    {file = "nvidia_nvshmem_cu13-3.4.5-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:6dc2a197f38e5d0376ad52cd1a2a3617d3cdc150fd5966f4aee9bcebb1d68fe9"},
+    {file = "nvidia_nvshmem_cu13-3.4.5-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:290f0a2ee94c9f3687a02502f3b9299a9f9fe826e6d0287ee18482e78d495b80"},
+]
+
+[[package]]
+name = "packaging"
+version = "26.3"
+description = "Core utilities for Python packages"
+optional = false
+python-versions = ">=3.9"
+groups = ["main"]
+files = [
+    {file = "packaging-26.3-py3-none-any.whl", hash = "sha256:d7193f7c8e4e93f444fde0262bf90af30e16fa0ad0ad44cb553c87339b23cd1c"},
+    {file = "packaging-26.3.tar.gz", hash = "sha256:94edc256424af38762eb31306eed28beb9f0efc50a8837492c9d6fd6004aed79"},
+]
+
+[[package]]
+name = "protobuf"
+version = "7.36.2"
+description = ""
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "protobuf-7.36.2-cp310-abi3-macosx_10_9_universal2.whl", hash = "sha256:cbc70b17ee27e28894c7fee8bb04be1abead49e936bc70eb60052531eee2079e"},
+    {file = "protobuf-7.36.2-cp310-abi3-manylinux2014_aarch64.whl", hash = "sha256:e11e1f0180583a2af89db6a2ecd9e8dc40aa6d2988ca175bfd0e6d12ea72d74e"},
+    {file = "protobuf-7.36.2-cp310-abi3-manylinux2014_s390x.whl", hash = "sha256:f4fee11ec330d238b34a05c9b675f693c20415d1c5bd7d5320cc2f8a798eb9cf"},
+    {file = "protobuf-7.36.2-cp310-abi3-manylinux2014_x86_64.whl", hash = "sha256:89f23aa53c24553a2416fd4fd1ec06f74fa42b14b546d8883128813f775bbfd2"},
+    {file = "protobuf-7.36.2-cp310-abi3-win32.whl", hash = "sha256:912c1221170e16c08d1f086762f563dd61ff83c18b5fa6652952dfaded66f728"},
+    {file = "protobuf-7.36.2-cp310-abi3-win_amd64.whl", hash = "sha256:a300819d441e078a5608c0d3c709796bb548136058fda017ae51d425b44fd353"},
+    {file = "protobuf-7.36.2-py3-none-any.whl", hash = "sha256:bdb3a345d48db958e6ce1f18e508beb0cc981d64f24088427549c866cd039f1e"},
+    {file = "protobuf-7.36.2.tar.gz", hash = "sha256:497d0463ff3316681da6c0b9e8d06cb465d61abce00b613ab42226175644d1bb"},
+]
+
+[[package]]
+name = "pygments"
+version = "2.21.0"
+description = "Pygments is a syntax highlighting package written in Python."
+optional = false
+python-versions = ">=3.9"
+groups = ["main"]
+files = [
+    {file = "pygments-2.21.0-py3-none-any.whl", hash = "sha256:2363c69b61c4a97c838da3b130dcd6468f4848992b21a82f2a63ec34377137d9"},
+    {file = "pygments-2.21.0.tar.gz", hash = "sha256:610ca751c9bc2492b38eb9a38a7fbc93edbbb2d7182edaf34e66ae493dee5c8c"},
+]
+
+[package.extras]
+windows-terminal = ["colorama (>=0.4.6)"]
+
+[[package]]
+name = "pyyaml"
+version = "6.0.3"
+description = "YAML parser and emitter for Python"
+optional = false
+python-versions = ">=3.8"
+groups = ["main"]
+files = [
+    {file = "PyYAML-6.0.3-cp38-cp38-macosx_10_13_x86_64.whl", hash = "sha256:c2514fceb77bc5e7a2f7adfaa1feb2fb311607c9cb518dbc378688ec73d8292f"},
+    {file = "PyYAML-6.0.3-cp38-cp38-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:9c57bb8c96f6d1808c030b1687b9b5fb476abaa47f0db9c0101f5e9f394e97f4"},
+    {file = "PyYAML-6.0.3-cp38-cp38-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:efd7b85f94a6f21e4932043973a7ba2613b059c4a000551892ac9f1d11f5baf3"},
+    {file = "PyYAML-6.0.3-cp38-cp38-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:22ba7cfcad58ef3ecddc7ed1db3409af68d023b7f940da23c6c2a1890976eda6"},
+    {file = "PyYAML-6.0.3-cp38-cp38-musllinux_1_2_x86_64.whl", hash = "sha256:6344df0d5755a2c9a276d4473ae6b90647e216ab4757f8426893b5dd2ac3f369"},
+    {file = "PyYAML-6.0.3-cp38-cp38-win32.whl", hash = "sha256:3ff07ec89bae51176c0549bc4c63aa6202991da2d9a6129d7aef7f1407d3f295"},
+    {file = "PyYAML-6.0.3-cp38-cp38-win_amd64.whl", hash = "sha256:5cf4e27da7e3fbed4d6c3d8e797387aaad68102272f8f9752883bc32d61cb87b"},
+    {file = "pyyaml-6.0.3-cp310-cp310-macosx_10_13_x86_64.whl", hash = "sha256:214ed4befebe12df36bcc8bc2b64b396ca31be9304b8f59e25c11cf94a4c033b"},
+    {file = "pyyaml-6.0.3-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:02ea2dfa234451bbb8772601d7b8e426c2bfa197136796224e50e35a78777956"},
+    {file = "pyyaml-6.0.3-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:b30236e45cf30d2b8e7b3e85881719e98507abed1011bf463a8fa23e9c3e98a8"},
+    {file = "pyyaml-6.0.3-cp310-cp310-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:66291b10affd76d76f54fad28e22e51719ef9ba22b29e1d7d03d6777a9174198"},
+    {file = "pyyaml-6.0.3-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:9c7708761fccb9397fe64bbc0395abcae8c4bf7b0eac081e12b809bf47700d0b"},
+    {file = "pyyaml-6.0.3-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:418cf3f2111bc80e0933b2cd8cd04f286338bb88bdc7bc8e6dd775ebde60b5e0"},
+    {file = "pyyaml-6.0.3-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:5e0b74767e5f8c593e8c9b5912019159ed0533c70051e9cce3e8b6aa699fcd69"},
+    {file = "pyyaml-6.0.3-cp310-cp310-win32.whl", hash = "sha256:28c8d926f98f432f88adc23edf2e6d4921ac26fb084b028c733d01868d19007e"},
+    {file = "pyyaml-6.0.3-cp310-cp310-win_amd64.whl", hash = "sha256:bdb2c67c6c1390b63c6ff89f210c8fd09d9a1217a465701eac7316313c915e4c"},
+    {file = "pyyaml-6.0.3-cp311-cp311-macosx_10_13_x86_64.whl", hash = "sha256:44edc647873928551a01e7a563d7452ccdebee747728c1080d881d68af7b997e"},
+    {file = "pyyaml-6.0.3-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:652cb6edd41e718550aad172851962662ff2681490a8a711af6a4d288dd96824"},
+    {file = "pyyaml-6.0.3-cp311-cp311-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:10892704fc220243f5305762e276552a0395f7beb4dbf9b14ec8fd43b57f126c"},
+    {file = "pyyaml-6.0.3-cp311-cp311-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:850774a7879607d3a6f50d36d04f00ee69e7fc816450e5f7e58d7f17f1ae5c00"},
+    {file = "pyyaml-6.0.3-cp311-cp311-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:b8bb0864c5a28024fac8a632c443c87c5aa6f215c0b126c449ae1a150412f31d"},
+    {file = "pyyaml-6.0.3-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:1d37d57ad971609cf3c53ba6a7e365e40660e3be0e5175fa9f2365a379d6095a"},
+    {file = "pyyaml-6.0.3-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:37503bfbfc9d2c40b344d06b2199cf0e96e97957ab1c1b546fd4f87e53e5d3e4"},
+    {file = "pyyaml-6.0.3-cp311-cp311-win32.whl", hash = "sha256:8098f252adfa6c80ab48096053f512f2321f0b998f98150cea9bd23d83e1467b"},
+    {file = "pyyaml-6.0.3-cp311-cp311-win_amd64.whl", hash = "sha256:9f3bfb4965eb874431221a3ff3fdcddc7e74e3b07799e0e84ca4a0f867d449bf"},
+    {file = "pyyaml-6.0.3-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:7f047e29dcae44602496db43be01ad42fc6f1cc0d8cd6c83d342306c32270196"},
+    {file = "pyyaml-6.0.3-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:fc09d0aa354569bc501d4e787133afc08552722d3ab34836a80547331bb5d4a0"},
+    {file = "pyyaml-6.0.3-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:9149cad251584d5fb4981be1ecde53a1ca46c891a79788c0df828d2f166bda28"},
+    {file = "pyyaml-6.0.3-cp312-cp312-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:5fdec68f91a0c6739b380c83b951e2c72ac0197ace422360e6d5a959d8d97b2c"},
+    {file = "pyyaml-6.0.3-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:ba1cc08a7ccde2d2ec775841541641e4548226580ab850948cbfda66a1befcdc"},
+    {file = "pyyaml-6.0.3-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:8dc52c23056b9ddd46818a57b78404882310fb473d63f17b07d5c40421e47f8e"},
+    {file = "pyyaml-6.0.3-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:41715c910c881bc081f1e8872880d3c650acf13dfa8214bad49ed4cede7c34ea"},
+    {file = "pyyaml-6.0.3-cp312-cp312-win32.whl", hash = "sha256:96b533f0e99f6579b3d4d4995707cf36df9100d67e0c8303a0c55b27b5f99bc5"},
+    {file = "pyyaml-6.0.3-cp312-cp312-win_amd64.whl", hash = "sha256:5fcd34e47f6e0b794d17de1b4ff496c00986e1c83f7ab2fb8fcfe9616ff7477b"},
+    {file = "pyyaml-6.0.3-cp312-cp312-win_arm64.whl", hash = "sha256:64386e5e707d03a7e172c0701abfb7e10f0fb753ee1d773128192742712a98fd"},
+    {file = "pyyaml-6.0.3-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:8da9669d359f02c0b91ccc01cac4a67f16afec0dac22c2ad09f46bee0697eba8"},
+    {file = "pyyaml-6.0.3-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:2283a07e2c21a2aa78d9c4442724ec1eb15f5e42a723b99cb3d822d48f5f7ad1"},
+    {file = "pyyaml-6.0.3-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:ee2922902c45ae8ccada2c5b501ab86c36525b883eff4255313a253a3160861c"},
+    {file = "pyyaml-6.0.3-cp313-cp313-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:a33284e20b78bd4a18c8c2282d549d10bc8408a2a7ff57653c0cf0b9be0afce5"},
+    {file = "pyyaml-6.0.3-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:0f29edc409a6392443abf94b9cf89ce99889a1dd5376d94316ae5145dfedd5d6"},
+    {file = "pyyaml-6.0.3-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:f7057c9a337546edc7973c0d3ba84ddcdf0daa14533c2065749c9075001090e6"},
+    {file = "pyyaml-6.0.3-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:eda16858a3cab07b80edaf74336ece1f986ba330fdb8ee0d6c0d68fe82bc96be"},
+    {file = "pyyaml-6.0.3-cp313-cp313-win32.whl", hash = "sha256:d0eae10f8159e8fdad514efdc92d74fd8d682c933a6dd088030f3834bc8e6b26"},
+    {file = "pyyaml-6.0.3-cp313-cp313-win_amd64.whl", hash = "sha256:79005a0d97d5ddabfeeea4cf676af11e647e41d81c9a7722a193022accdb6b7c"},
+    {file = "pyyaml-6.0.3-cp313-cp313-win_arm64.whl", hash = "sha256:5498cd1645aa724a7c71c8f378eb29ebe23da2fc0d7a08071d89469bf1d2defb"},
+    {file = "pyyaml-6.0.3-cp314-cp314-macosx_10_13_x86_64.whl", hash = "sha256:8d1fab6bb153a416f9aeb4b8763bc0f22a5586065f86f7664fc23339fc1c1fac"},
+    {file = "pyyaml-6.0.3-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:34d5fcd24b8445fadc33f9cf348c1047101756fd760b4dacb5c3e99755703310"},
+    {file = "pyyaml-6.0.3-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:501a031947e3a9025ed4405a168e6ef5ae3126c59f90ce0cd6f2bfc477be31b7"},
+    {file = "pyyaml-6.0.3-cp314-cp314-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:b3bc83488de33889877a0f2543ade9f70c67d66d9ebb4ac959502e12de895788"},
+    {file = "pyyaml-6.0.3-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:c458b6d084f9b935061bc36216e8a69a7e293a2f1e68bf956dcd9e6cbcd143f5"},
+    {file = "pyyaml-6.0.3-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:7c6610def4f163542a622a73fb39f534f8c101d690126992300bf3207eab9764"},
+    {file = "pyyaml-6.0.3-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:5190d403f121660ce8d1d2c1bb2ef1bd05b5f68533fc5c2ea899bd15f4399b35"},
+    {file = "pyyaml-6.0.3-cp314-cp314-win_amd64.whl", hash = "sha256:4a2e8cebe2ff6ab7d1050ecd59c25d4c8bd7e6f400f5f82b96557ac0abafd0ac"},
+    {file = "pyyaml-6.0.3-cp314-cp314-win_arm64.whl", hash = "sha256:93dda82c9c22deb0a405ea4dc5f2d0cda384168e466364dec6255b293923b2f3"},
+    {file = "pyyaml-6.0.3-cp314-cp314t-macosx_10_13_x86_64.whl", hash = "sha256:02893d100e99e03eda1c8fd5c441d8c60103fd175728e23e431db1b589cf5ab3"},
+    {file = "pyyaml-6.0.3-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:c1ff362665ae507275af2853520967820d9124984e0f7466736aea23d8611fba"},
+    {file = "pyyaml-6.0.3-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:6adc77889b628398debc7b65c073bcb99c4a0237b248cacaf3fe8a557563ef6c"},
+    {file = "pyyaml-6.0.3-cp314-cp314t-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:a80cb027f6b349846a3bf6d73b5e95e782175e52f22108cfa17876aaeff93702"},
+    {file = "pyyaml-6.0.3-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:00c4bdeba853cc34e7dd471f16b4114f4162dc03e6b7afcc2128711f0eca823c"},
+    {file = "pyyaml-6.0.3-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:66e1674c3ef6f541c35191caae2d429b967b99e02040f5ba928632d9a7f0f065"},
+    {file = "pyyaml-6.0.3-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:16249ee61e95f858e83976573de0f5b2893b3677ba71c9dd36b9cf8be9ac6d65"},
+    {file = "pyyaml-6.0.3-cp314-cp314t-win_amd64.whl", hash = "sha256:4ad1906908f2f5ae4e5a8ddfce73c320c2a1429ec52eafd27138b7f1cbe341c9"},
+    {file = "pyyaml-6.0.3-cp314-cp314t-win_arm64.whl", hash = "sha256:ebc55a14a21cb14062aa4162f906cd962b28e2e9ea38f9b4391244cd8de4ae0b"},
+    {file = "pyyaml-6.0.3-cp39-cp39-macosx_10_13_x86_64.whl", hash = "sha256:b865addae83924361678b652338317d1bd7e79b1f4596f96b96c77a5a34b34da"},
+    {file = "pyyaml-6.0.3-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:c3355370a2c156cffb25e876646f149d5d68f5e0a3ce86a5084dd0b64a994917"},
+    {file = "pyyaml-6.0.3-cp39-cp39-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:3c5677e12444c15717b902a5798264fa7909e41153cdf9ef7ad571b704a63dd9"},
+    {file = "pyyaml-6.0.3-cp39-cp39-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:5ed875a24292240029e4483f9d4a4b8a1ae08843b9c54f43fcc11e404532a8a5"},
+    {file = "pyyaml-6.0.3-cp39-cp39-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:0150219816b6a1fa26fb4699fb7daa9caf09eb1999f3b70fb6e786805e80375a"},
+    {file = "pyyaml-6.0.3-cp39-cp39-musllinux_1_2_aarch64.whl", hash = "sha256:fa160448684b4e94d80416c0fa4aac48967a969efe22931448d853ada8baf926"},
+    {file = "pyyaml-6.0.3-cp39-cp39-musllinux_1_2_x86_64.whl", hash = "sha256:27c0abcb4a5dac13684a37f76e701e054692a9b2d3064b70f5e4eb54810553d7"},
+    {file = "pyyaml-6.0.3-cp39-cp39-win32.whl", hash = "sha256:1ebe39cb5fc479422b83de611d14e2c0d3bb2a18bbcb01f229ab3cfbd8fee7a0"},
+    {file = "pyyaml-6.0.3-cp39-cp39-win_amd64.whl", hash = "sha256:2e71d11abed7344e42a8849600193d15b6def118602c4c176f748e4583246007"},
+    {file = "pyyaml-6.0.3.tar.gz", hash = "sha256:d76623373421df22fb4cf8817020cbb7ef15c725b9d5e45f17e189bfc384190f"},
+]
+
+[[package]]
+name = "regex"
+version = "2026.9.29"
+description = "Alternative regular expression module, to replace re."
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "regex-2026.9.29-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:9916fda742cd4eede63b286f58c06718324265d727ce0856eb1aac86d0d150d6"},
+    {file = "regex-2026.9.29-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:8873c4a11c50b9989168881aeb3f08859f469d809941866aa1feefd8be5431f6"},
+    {file = "regex-2026.9.29-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:1d9fe8091b2e89d470df68a9331111ed008ae8aae6bf1e8e1fba4086a495c84e"},
+    {file = "regex-2026.9.29-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:fb00027a09a8f9f08028b40dce4c933cf73e4833240ed356583fdc9cfa721566"},
+    {file = "regex-2026.9.29-cp310-cp310-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:14e953ff3607c92d7675bf79c4d4509ef6782aa8c08509f179f9b3d6d0679e86"},
+    {file = "regex-2026.9.29-cp310-cp310-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:0476e5bcbe6e1ba3d1c4cc7bbb1c3ba78e3b979b5c8a88d0a6a8cdd4992b8c84"},
+    {file = "regex-2026.9.29-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:4fb41211d2333eb930a51e0546a65999761cf1f572a4da56ef9b8a62966c06f2"},
+    {file = "regex-2026.9.29-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:edf06545875f3efa31560d94121e95c7fd70d98b1dfedc0157097d79b13b52ea"},
+    {file = "regex-2026.9.29-cp310-cp310-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:6398d5145689503412cc1748895242598d8846b8967b851133b20dc2ed1e21e8"},
+    {file = "regex-2026.9.29-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:45010bcfe66df41522d56c9b6114e87ecc597a08970ff6a2ced24415c141ae5f"},
+    {file = "regex-2026.9.29-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:a5758353650079898dc1b2b0e95aa51fa23a30d020e06f62c430dd08ee56cdd8"},
+    {file = "regex-2026.9.29-cp310-cp310-musllinux_1_2_riscv64.whl", hash = "sha256:6f7121a8914ed13fcfe2099f895341bfb789f004d4c5a0bdece8fa667da10849"},
+    {file = "regex-2026.9.29-cp310-cp310-musllinux_1_2_s390x.whl", hash = "sha256:b9d74e4eee9ddb64c2e92d5d61472c59c21684c059eb7b68767be9628e977859"},
+    {file = "regex-2026.9.29-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:143533cc4b6fbc5b95aca0a5b8d541088d374831593def000ec89322c220221d"},
+    {file = "regex-2026.9.29-cp310-cp310-win32.whl", hash = "sha256:b84f186a7f0536fe4ff9a9fa12d06d007b9b71d4b5352ddcc41f59ad6522a312"},
+    {file = "regex-2026.9.29-cp310-cp310-win_amd64.whl", hash = "sha256:23ae6fdad9e63e54038f5ef78aba2933faca61e24d432786589e737bc5522ebb"},
+    {file = "regex-2026.9.29-cp310-cp310-win_arm64.whl", hash = "sha256:c0094897d7d01f184b2d7fe8c56c66d64efe01b31f4b7d34205b391387df1111"},
+    {file = "regex-2026.9.29-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:6abb75ab16bc3281714a5b99548a2225db70dba1f995f6d7f7419b76eb5a8fbe"},
+    {file = "regex-2026.9.29-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:b7b893976e7fe42053da64f2aa27239c24252fd2ec6df471e1be197c0addc3b1"},
+    {file = "regex-2026.9.29-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:066d0e3dbfdd739bce2bf8c2a41dd16f73e3d8adc2eb06dd803a36a307f56075"},
+    {file = "regex-2026.9.29-cp311-cp311-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:7020ed44df30b3aa492c00ee3b52d0548c1f30c2c6c5bb13ae897680900d3413"},
+    {file = "regex-2026.9.29-cp311-cp311-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:ae4613d7d9dda60fcba95f846cc6f808017f1843f392cf9daad14a6534493d71"},
+    {file = "regex-2026.9.29-cp311-cp311-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:bec37990e3d6121f29ecfb594bd8f1bf009e9f7926daba2e50e3b27d3892a783"},
+    {file = "regex-2026.9.29-cp311-cp311-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:612b709381c0355b70d89cdb51b7f670591ed5cbbc0e3b5337488019dc667b65"},
+    {file = "regex-2026.9.29-cp311-cp311-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:a760da040b47767b4b873adfb7c3b691e9ba2fc60f113f9d0b88f1a62f323e85"},
+    {file = "regex-2026.9.29-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:49ee178ca31c94621294bf9b8b676a92a2e6bba8af0529591753719e57edb621"},
+    {file = "regex-2026.9.29-cp311-cp311-musllinux_1_2_ppc64le.whl", hash = "sha256:5eeb8edc6110d9194a4d0d54610f64c37a31c605b5dbb7e407fc6ec7fa34a4a1"},
+    {file = "regex-2026.9.29-cp311-cp311-musllinux_1_2_riscv64.whl", hash = "sha256:ccb64d887a9db1cd76dbc0f92051a1a478a2a67e7f56c62d915cb881d7734704"},
+    {file = "regex-2026.9.29-cp311-cp311-musllinux_1_2_s390x.whl", hash = "sha256:9e4482589065c8ecd761cff522dcd85f2d39e62f551e37e025d1c7d54772def3"},
+    {file = "regex-2026.9.29-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:d60030baaa7bfbb02d650c126cdcddcb6e33dbff14d819434c8fa2fdcaeeeba5"},
+    {file = "regex-2026.9.29-cp311-cp311-win32.whl", hash = "sha256:18ae8eed4526e35bdb754d61562b90bf5c00a67fdcf3cc1380dd59597486631b"},
+    {file = "regex-2026.9.29-cp311-cp311-win_amd64.whl", hash = "sha256:1043aedf5917caa861bcb25a9c11460049656bdf0017a90a309fa8f255467725"},
+    {file = "regex-2026.9.29-cp311-cp311-win_arm64.whl", hash = "sha256:352cf115a810b357caa35193ab656ecf5ef41056855e82f292c99e8514f8d954"},
+    {file = "regex-2026.9.29-cp312-cp312-macosx_10_13_universal2.whl", hash = "sha256:dc79d36d0618752265f0d575915bdc5c5130ecb9c9f6b3bcefeae32e4bdfafcf"},
+    {file = "regex-2026.9.29-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:3a21a9509d0ee88e7a70e1ad228cd2f0e0fd1e187458db132e8a8d18c97daf9d"},
+    {file = "regex-2026.9.29-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:f57dc6b8fef170f105d2cf5cdce254f47b137d7755086cf7050f47e16582abba"},
+    {file = "regex-2026.9.29-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:f93bc1c3486ef3747e07c9d7c1d0a147b8fbaab975f80e348aed6f71309dfaca"},
+    {file = "regex-2026.9.29-cp312-cp312-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:9e1d3a4cb7993b708f0ada8d0c84590efd853f169e7147d2202c9da503180242"},
+    {file = "regex-2026.9.29-cp312-cp312-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:dabee8f4935e731fb46b2a3091bdda0d3d94b3bbfb907d2b4f12eefce4009619"},
+    {file = "regex-2026.9.29-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:39ab5894d971f9ac68baa6eca5c50387db579cfcacf36ae8df3feceb1815e6d0"},
+    {file = "regex-2026.9.29-cp312-cp312-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:c1a9a6651197fbed6f0212591418b9def774fc3f8324f78d1bf0e6a63e5f8aa1"},
+    {file = "regex-2026.9.29-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:87fb80cbe3557e27e7b28b995c2b2eedf689b8886f941ab93e0e288f0976518a"},
+    {file = "regex-2026.9.29-cp312-cp312-musllinux_1_2_ppc64le.whl", hash = "sha256:3c5c2ef13797466aa64170cbb66ad98a32351dd4127694cea7199f80f213750d"},
+    {file = "regex-2026.9.29-cp312-cp312-musllinux_1_2_riscv64.whl", hash = "sha256:59b49507f47479e299a9e1bc41b5cb83a7afda0540625f1dbae886615978acbf"},
+    {file = "regex-2026.9.29-cp312-cp312-musllinux_1_2_s390x.whl", hash = "sha256:0dd8af32e9f7b56b7f95cc1fd79b23054c3bdc172392ae560acc24d57b7ffe71"},
+    {file = "regex-2026.9.29-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:db5e82ba15c142425b8406690032df89e39cca4a2e8afbbb9a3d84edc2373ac3"},
+    {file = "regex-2026.9.29-cp312-cp312-win32.whl", hash = "sha256:d0c3082bf79bcd6a614d55916590ad4b8f93200e10b97f463ea5d9d07c9b5f23"},
+    {file = "regex-2026.9.29-cp312-cp312-win_amd64.whl", hash = "sha256:fdd88ed5e20b1bcdd234421e454962c971aa44b653bdb7f1ea9ef683e90fb649"},
+    {file = "regex-2026.9.29-cp312-cp312-win_arm64.whl", hash = "sha256:4fe97894d1b306c919b4e50def1e6f6c522f4d03a7283811f4d108f1ce5d3ac2"},
+    {file = "regex-2026.9.29-cp313-cp313-macosx_10_13_universal2.whl", hash = "sha256:f1a0d5117230dd46b399a30a38afa44f79c99f3168988fdc4f425c3f928b39df"},
+    {file = "regex-2026.9.29-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:f0fe9834e5aeccaf19a0d8feb296d66a24be1a7c9922002f842a682cd5abb787"},
+    {file = "regex-2026.9.29-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:c90fcf7804ea0a54b896ce0f2b9565350220b8d4890fd0db461a476a4c687963"},
+    {file = "regex-2026.9.29-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:e11edba5bc344a32b029a7af9d4b3173982dd79eeafa0b9dbd787364414b0509"},
+    {file = "regex-2026.9.29-cp313-cp313-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:bb90e7177944b6684738c1fc36aabd2dd00d1de3be7dbe09f91e196f1bc0dc81"},
+    {file = "regex-2026.9.29-cp313-cp313-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:d06fcdecc10fc7954d7c8f27a03c96055fe525274dc84a7b0dbdc3d6b9e03dab"},
+    {file = "regex-2026.9.29-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:d49c18f1ea294cf4adde2e5ac256e98c82ea9d708462ce4bf799dffa7cfe8a2c"},
+    {file = "regex-2026.9.29-cp313-cp313-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:3e778bfccd63075167709136afbc251c1f683758d5bf49c803c60ac3f894ce6b"},
+    {file = "regex-2026.9.29-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:686ac5350fceae63830bb98805fcb8039325bf4c06d9f6f048ff65229d5bffa5"},
+    {file = "regex-2026.9.29-cp313-cp313-musllinux_1_2_ppc64le.whl", hash = "sha256:26ec4ccce55aa533fbd603d08911b01101a8fcfec987845ac3ae2c7087b2bde3"},
+    {file = "regex-2026.9.29-cp313-cp313-musllinux_1_2_riscv64.whl", hash = "sha256:a655d34b2a6943af32401f3d94f72e9d731f6ad16285815550bf2b4ee69d420a"},
+    {file = "regex-2026.9.29-cp313-cp313-musllinux_1_2_s390x.whl", hash = "sha256:0c992c19cd45058a4b92f68f139c93db168b48fb1f322c9a7cd620806afb6b51"},
+    {file = "regex-2026.9.29-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:ebb8912f565b8cdbbf27debfe00df04202c20e2f651b9e32767930c5eace3621"},
+    {file = "regex-2026.9.29-cp313-cp313-win32.whl", hash = "sha256:4d7d93613b01b0199961330e49cfc52d479b3d5776c56c691db31130c0a07d91"},
+    {file = "regex-2026.9.29-cp313-cp313-win_amd64.whl", hash = "sha256:61956f074ecd123f55adca68ee3eab46e6a07ad3f8e64e6db95dfacb444f55c4"},
+    {file = "regex-2026.9.29-cp313-cp313-win_arm64.whl", hash = "sha256:bfc71e6d970419c1309b3640305298643e2a734cad3f7cfb6d2ddee4175ab53d"},
+    {file = "regex-2026.9.29-cp314-cp314-macosx_10_15_universal2.whl", hash = "sha256:957bb708e8057ab1649ba566456429d691ec9b90d1c9ad1af1ba7ffbbeaf05f2"},
+    {file = "regex-2026.9.29-cp314-cp314-macosx_10_15_x86_64.whl", hash = "sha256:c9b602fae1e00b7c035d661ce85575365719192a7b46784bd71cf64c68053aa0"},
+    {file = "regex-2026.9.29-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:0166844493626c5015c6088ee15c9ca2fd060ca15b7641d1657da6a58432ae33"},
+    {file = "regex-2026.9.29-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:b97a38fb4c732b6832db6bf108963adbcd82ef1268ba2025dce390f45af75efa"},
+    {file = "regex-2026.9.29-cp314-cp314-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:a540abfab208e1b7ef2df231c40ef3b6cbb30a0aad6204e9b6a81c10a6794628"},
+    {file = "regex-2026.9.29-cp314-cp314-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:ddfa987262763c3c22a8367d2a49c244b018a74c3a8e3ab1a864119ad45c5633"},
+    {file = "regex-2026.9.29-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:2f7f7aa47b229f2b39a2ae2596d2ad5625d77b5eb9856fac2dab3eb506cdd0a0"},
+    {file = "regex-2026.9.29-cp314-cp314-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:d9b77b25b4f395f92de6099ab08e8ae2bc7e51dfe157f22900902243a5cc90c7"},
+    {file = "regex-2026.9.29-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:34b6925af9853bf461950e6508910f179fd6e9b1a7ec8548e069606b7e51a26b"},
+    {file = "regex-2026.9.29-cp314-cp314-musllinux_1_2_ppc64le.whl", hash = "sha256:addd736a0547d553283adaf4e05d7104e7f2c7b0b092e9b4d28756825f14531f"},
+    {file = "regex-2026.9.29-cp314-cp314-musllinux_1_2_riscv64.whl", hash = "sha256:fe3fa1dd453ed5c7f5ea23a26218329790ed7197a99b90e94330e313959a7f52"},
+    {file = "regex-2026.9.29-cp314-cp314-musllinux_1_2_s390x.whl", hash = "sha256:0cc63b5e47c12a48d90c7e9d7de6a035dd14f62868aaedbb4e0ff8ba2b8bfe7b"},
+    {file = "regex-2026.9.29-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:724184b4aafed865e4f13ca313fdcb43024300c028ec67319cfa16847d84685e"},
+    {file = "regex-2026.9.29-cp314-cp314-win32.whl", hash = "sha256:c6c8fabf1dafc1f1ddcbb67896d3f93efb092e8c4b6322d7389b944e76a484e5"},
+    {file = "regex-2026.9.29-cp314-cp314-win_amd64.whl", hash = "sha256:1c2a0026062abcc321a53db4a185ceba0b59a66b5d37b0808917a88b55a5257f"},
+    {file = "regex-2026.9.29-cp314-cp314-win_arm64.whl", hash = "sha256:121a76a0985db80ceae9e171c337f8c927868e37d01b54e3ce87bc87f9c6a208"},
+    {file = "regex-2026.9.29-cp314-cp314t-macosx_10_15_universal2.whl", hash = "sha256:e31f72490b7c12f7790e1e25c3afffd20503ee1bfb43461d7838b871ff244b19"},
+    {file = "regex-2026.9.29-cp314-cp314t-macosx_10_15_x86_64.whl", hash = "sha256:80ea96f5c1a30bf09007d48466521d9c294bebe197c708c3359096e3e3691632"},
+    {file = "regex-2026.9.29-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:554bffadcbcb6d5f4e5fb10a61cc52084b9a63d1dab5f10bcd2c4343972e8e2c"},
+    {file = "regex-2026.9.29-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:864e9b87ac33c3fb9fb4ad48166d4fdb579c351d5c77deb0d34bccb36a775cd9"},
+    {file = "regex-2026.9.29-cp314-cp314t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:044265d77d94f5e3cb2fd72c76723807c429cb8c533e9d4672d0334a6f14f588"},
+    {file = "regex-2026.9.29-cp314-cp314t-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:2089fe39c406784d90101c726755ffa1497bb74638fd434300d2b88006186de8"},
+    {file = "regex-2026.9.29-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:0def9fb6abac55492d6d51cddb7225d07d6f279e774e0adc08569a54a5fc8d46"},
+    {file = "regex-2026.9.29-cp314-cp314t-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:888d60953908dcf761aa320c3e390ab8556efbdb551ace63921de90f6ae0848d"},
+    {file = "regex-2026.9.29-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:ed511a0708e2297e1d6431e7fb217e3402791e491e02da800658ace4973df1bb"},
+    {file = "regex-2026.9.29-cp314-cp314t-musllinux_1_2_ppc64le.whl", hash = "sha256:e1172147d28d8fbcf8cb8d26c41506169f5ad8fe9ec969cb116835a19d4d8eca"},
+    {file = "regex-2026.9.29-cp314-cp314t-musllinux_1_2_riscv64.whl", hash = "sha256:92f05c9c42bde5785dc48770bc2194d9f7442544156f951e19cd31b096cec562"},
+    {file = "regex-2026.9.29-cp314-cp314t-musllinux_1_2_s390x.whl", hash = "sha256:f37964e4a5e993d2fd45147741e9dff7f34a2d8c00ab94c4ea0514a4677f959e"},
+    {file = "regex-2026.9.29-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:951733b1bbdb71e377cec567b409f1a7881b47cfcad84121aa74cb575fa425ea"},
+    {file = "regex-2026.9.29-cp314-cp314t-win32.whl", hash = "sha256:65b408d8fcb273e3499e7ef2ce796810da1becd208c7fb4373692a242d79d461"},
+    {file = "regex-2026.9.29-cp314-cp314t-win_amd64.whl", hash = "sha256:bf48516e35cf848390ea68850aba53e7c333720d2945b4d2c25b69fc5171723f"},
+    {file = "regex-2026.9.29-cp314-cp314t-win_arm64.whl", hash = "sha256:9173db3be74a35cb6731701094b98120f7ee4876a287882a59cdea1fa7da342f"},
+    {file = "regex-2026.9.29-cp315-cp315-macosx_10_15_universal2.whl", hash = "sha256:c3589f40749acce747510bf5d589d54e376cb0930ea58b35effac97e5312b0c1"},
+    {file = "regex-2026.9.29-cp315-cp315-macosx_10_15_x86_64.whl", hash = "sha256:32ab11df9677ca80bcbb5fe4eb1da9109a5019239a054836efc6fa1c64e683cf"},
+    {file = "regex-2026.9.29-cp315-cp315-macosx_11_0_arm64.whl", hash = "sha256:7c03031610e3e6ed1768a2b7a8fc84637c1257b50c5eacaf094c6e17a84fc563"},
+    {file = "regex-2026.9.29-cp315-cp315-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:42e82e578c904445d4c8a35b8f28052cf567593215fa5db06266fbc6f77aaa2e"},
+    {file = "regex-2026.9.29-cp315-cp315-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:0b65c72739f981377c9c22e0c5c3cd7f42da7bd8a3c9209330fac772c7d893ed"},
+    {file = "regex-2026.9.29-cp315-cp315-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:4408b2b27a95ca8cc48b7411945753773353b5c93b307754781086c99d3a576f"},
+    {file = "regex-2026.9.29-cp315-cp315-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:a714befaacbd10092ffe4cea0d3c5f008fb9efe9bc322c715bcdfdee414b9a3d"},
+    {file = "regex-2026.9.29-cp315-cp315-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:33026515aebc0e70d1c89978e53e8d695d35d9e472f8d5b34465ba3c74028650"},
+    {file = "regex-2026.9.29-cp315-cp315-musllinux_1_2_aarch64.whl", hash = "sha256:31b003f9a070335e2a8233ee9b14a3ca8e6d792012ae011f741bf0aaf11744c5"},
+    {file = "regex-2026.9.29-cp315-cp315-musllinux_1_2_ppc64le.whl", hash = "sha256:c03c6eb6ece86dfdcbb34799efaa339b093132e1aceed491ba5e08fe06cdf699"},
+    {file = "regex-2026.9.29-cp315-cp315-musllinux_1_2_riscv64.whl", hash = "sha256:a5300757f8a68f5b6cc33f57338d72a0e3589c5cc9ad5f8504ea06f028be582a"},
+    {file = "regex-2026.9.29-cp315-cp315-musllinux_1_2_s390x.whl", hash = "sha256:80c7cadd3fd2bfde5df8aa0787e315812cad0c313a753095d02f4c2b6c01677b"},
+    {file = "regex-2026.9.29-cp315-cp315-musllinux_1_2_x86_64.whl", hash = "sha256:3f1e6cb402a89457582cd696f982559217d13484a193202c394015297968c86d"},
+    {file = "regex-2026.9.29-cp315-cp315-win32.whl", hash = "sha256:a64b85a4760337cfefdb27d42da6ed8b58e8cde3f2d57b6ef43e76ef6ea9ef47"},
+    {file = "regex-2026.9.29-cp315-cp315-win_amd64.whl", hash = "sha256:b3e445b66c80b4eb4234e855ce94d9adc183eedbd632816228d89930b91b2c5b"},
+    {file = "regex-2026.9.29-cp315-cp315-win_arm64.whl", hash = "sha256:8f39588af4731c8923c26810eb3b33f76f17633985e40f59c3cd45a33805a895"},
+    {file = "regex-2026.9.29-cp315-cp315t-macosx_10_15_universal2.whl", hash = "sha256:fb99cc9d45f48895d9d67f6a0b8a57f08d39c174d9f25ad97a313e0470267b1c"},
+    {file = "regex-2026.9.29-cp315-cp315t-macosx_10_15_x86_64.whl", hash = "sha256:720537c7ea6f80dc61913184edb0ce2497a306b39ef19f28505b322553d52bdb"},
+    {file = "regex-2026.9.29-cp315-cp315t-macosx_11_0_arm64.whl", hash = "sha256:0fd2c901cc307a745ad4bc87f20060d7a0825a3371d1e93488af22e7a387f78f"},
+    {file = "regex-2026.9.29-cp315-cp315t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:b11b589e00095ec69cf79841a76360f9b079e95b0368a25b5ebb951ab0c157ff"},
+    {file = "regex-2026.9.29-cp315-cp315t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:d7cab119d0df0b9413f106b4d7fc34f2872d3574ed3806fb48959c830b1537da"},
+    {file = "regex-2026.9.29-cp315-cp315t-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:b89efc38431793d28b7cd91227e2f952ad7c48df19132b17f43a5fec3c14143b"},
+    {file = "regex-2026.9.29-cp315-cp315t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:80a5ea3b4fd9d6a5b9a44f7976a9acaaab35aa3c1f6b29e5bd857dfabaded223"},
+    {file = "regex-2026.9.29-cp315-cp315t-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:19959129885356df0e97556856f77eb2888380dac18bed075a7c05c5128c618d"},
+    {file = "regex-2026.9.29-cp315-cp315t-musllinux_1_2_aarch64.whl", hash = "sha256:6a1a824fbed817e0a891103886b68f063b1e83cc51bc97192a90a60195a9291f"},
+    {file = "regex-2026.9.29-cp315-cp315t-musllinux_1_2_ppc64le.whl", hash = "sha256:1ba8c6a416569ce0d37e83e28a254a61dc99a419084dfb6476cea02d997f74fa"},
+    {file = "regex-2026.9.29-cp315-cp315t-musllinux_1_2_riscv64.whl", hash = "sha256:446654b29bfaa30500d80947eda42cef1449dc8a87f4e3cf061cc8485d3a1f0b"},
+    {file = "regex-2026.9.29-cp315-cp315t-musllinux_1_2_s390x.whl", hash = "sha256:bf3c49863c23a1ad6da9c30351aed6cff8d5ddbeb63c5c8420ae54e98c7d0138"},
+    {file = "regex-2026.9.29-cp315-cp315t-musllinux_1_2_x86_64.whl", hash = "sha256:01000ddf0e3ffef97f2413ceb514f6313040106b6d18a03ee00a4fe35c1eb1db"},
+    {file = "regex-2026.9.29-cp315-cp315t-win32.whl", hash = "sha256:c4e38dd8f39c43a91d2410ad2b85610701b0979342c3df1d69eaf8e838c757d8"},
+    {file = "regex-2026.9.29-cp315-cp315t-win_amd64.whl", hash = "sha256:e2c89e9b762c57f59d5e99ee8b20202adb892e35f8d3485741340999ca55058e"},
+    {file = "regex-2026.9.29-cp315-cp315t-win_arm64.whl", hash = "sha256:e8c65ef3862a8ad6e86492b6ed9327805dd66904c012bd3649dc67d822ed6c34"},
+    {file = "regex-2026.9.29.tar.gz", hash = "sha256:8b5fcc4771732191b2b7d1dd68d8f0353f47f8d90b6150f6dce58bf1112442cb"},
+]
+
+[[package]]
+name = "requests"
+version = "2.34.2"
+description = "Python HTTP for Humans."
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "requests-2.34.2-py3-none-any.whl", hash = "sha256:2a0d60c172f83ac6ab31e4554906c0f3b3588d37b5cb939b1c061f4907e278e0"},
+    {file = "requests-2.34.2.tar.gz", hash = "sha256:f288924cae4e29463698d6d60bc6a4da69c89185ad1e0bcc4104f584e960b9ed"},
+]
+
+[package.dependencies]
+certifi = ">=2023.5.7"
+charset_normalizer = ">=2,<4"
+idna = ">=2.5,<4"
+urllib3 = ">=1.26,<3"
+
+[package.extras]
+socks = ["PySocks (>=1.5.6,!=1.5.7)"]
+use-chardet-on-py3 = ["chardet (>=3.0.2,<8)"]
+
+[[package]]
+name = "rich"
+version = "15.0.0"
+description = "Render rich text, tables, progress bars, syntax highlighting, markdown and more to the terminal"
+optional = false
+python-versions = ">=3.9.0"
+groups = ["main"]
+files = [
+    {file = "rich-15.0.0-py3-none-any.whl", hash = "sha256:33bd4ef74232fb73fe9279a257718407f169c09b78a87ad3d296f548e27de0bb"},
+    {file = "rich-15.0.0.tar.gz", hash = "sha256:edd07a4824c6b40189fb7ac9bc4c52536e9780fbbfbddf6f1e2502c31b068c36"},
+]
+
+[package.dependencies]
+markdown-it-py = ">=2.2.0"
+pygments = ">=2.13.0,<3.0.0"
+
+[package.extras]
+jupyter = ["ipywidgets (>=7.5.1,<9)"]
+
+[[package]]
+name = "safetensors"
+version = "0.8.0"
+description = ""
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "safetensors-0.8.0-cp310-abi3-macosx_10_12_x86_64.whl", hash = "sha256:c554f85858e05226d3c2828e32395e677434685d6d94594a41643361c5e837f0"},
+    {file = "safetensors-0.8.0-cp310-abi3-macosx_11_0_arm64.whl", hash = "sha256:c80201d22cbf405b80647a60ada77bba06c8fba2da2743ba1e89cdcc39a81f25"},
+    {file = "safetensors-0.8.0-cp310-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:7a46e5ff292c356d6991e60942ba7f79817682d3a2cef0702136448cb9c4d235"},
+    {file = "safetensors-0.8.0-cp310-abi3-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:4124502b78f03534117c848f87a39b8f31e577b15eff423bf8bfb95f2a8c30d0"},
+    {file = "safetensors-0.8.0-cp310-abi3-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:7bc0a787ba8a35be368ee3574edfa2b1ad389eebd0a72e482ae275490e3f6c98"},
+    {file = "safetensors-0.8.0-cp310-abi3-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:040070828e36dc8e122178bbbd5830ff9e97920affb84cbe0f46442497bed358"},
+    {file = "safetensors-0.8.0-cp310-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:fd6f3f93c9a0a7cc2788ee63fb763353d4bd2e89b0751bc78fcf7dda00bea774"},
+    {file = "safetensors-0.8.0-cp310-abi3-manylinux_2_31_riscv64.whl", hash = "sha256:fcdd41ec4628fee5799f807c73c353629130fbd942aa23d83c623dd6c9d52d78"},
+    {file = "safetensors-0.8.0-cp310-abi3-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:8e9f537aa183a38ace122d27303dcd986b26bd2a7591f9181d7f0c396f4677ca"},
+    {file = "safetensors-0.8.0-cp310-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:87eec7ffed2b809f05a398a8becb7d013f19f7837cd15d9748580d6cf30dbaf4"},
+    {file = "safetensors-0.8.0-cp310-abi3-musllinux_1_2_armv7l.whl", hash = "sha256:4a95ae2b05d7726d751da4ebf626a2ca782b706e101bd894c95bc2450b1cffcc"},
+    {file = "safetensors-0.8.0-cp310-abi3-musllinux_1_2_i686.whl", hash = "sha256:3ae091f16662658bdc019a4ff6cb4c085bb7d725eb5978b183ffd265863b6d2d"},
+    {file = "safetensors-0.8.0-cp310-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:8e080062fcde23be189565e1c3305d16751a218ecf9412c8601e64204eb6f846"},
+    {file = "safetensors-0.8.0-cp310-abi3-win32.whl", hash = "sha256:2ddf52eac562eda224f99acfa7889d02968c1fd59a5b011ae7d8137c37e9c02d"},
+    {file = "safetensors-0.8.0-cp310-abi3-win_amd64.whl", hash = "sha256:096ec1a98435df7beb08853bb5aa9081a84f23d0adc67ed1a0a10550f608373f"},
+    {file = "safetensors-0.8.0-cp310-abi3-win_arm64.whl", hash = "sha256:f7838e5135a406ad3e02efdcb8cf2e5397d368b0154537c4fec682dbc544d452"},
+    {file = "safetensors-0.8.0.tar.gz", hash = "sha256:fabaf3e0f18a6618d9b36560682562157f77c2b71fcffc7b432be2baed9d753d"},
+]
+
+[package.extras]
+all = ["safetensors[convert]", "safetensors[jax]", "safetensors[numpy]", "safetensors[paddlepaddle]", "safetensors[quality]", "safetensors[testing]", "safetensors[torch]"]
+convert = ["huggingface-hub (>=1.4)", "safetensors[torch]"]
+dev = ["safetensors[all]", "safetensors[pinned-tf]"]
+jax = ["flax (>=0.6.3)", "jax (>=0.3.25)", "jaxlib (>=0.3.25)", "safetensors[numpy]"]
+mlx = ["mlx (>=0.0.9)"]
+numpy = ["numpy (>=1.24.6)"]
+paddlepaddle = ["paddlepaddle (>=2.4.1)", "safetensors[numpy]"]
+pinned-tf = ["safetensors[numpy]", "tensorflow (==2.18.0)"]
+quality = ["ruff"]
+tensorflow = ["safetensors[numpy]", "tensorflow (>=2.11.0)"]
+testing = ["fsspec (>=2024.6.0)", "h5py (>=3.7.0)", "hypothesis (>=6.70.2)", "pytest (>=9.0)", "pytest-benchmark (>=5.2)", "s3fs (>=2024.6.0)", "safetensors[numpy]", "setuptools-rust (>=1.12.0)"]
+tf-nightly = ["safetensors[numpy]", "tf-nightly"]
+torch = ["safetensors[numpy]", "torch (>=2.4)"]
+
+[[package]]
+name = "sentencepiece"
+version = "0.2.2"
+description = "Unsupervised text tokenizer and detokenizer."
+optional = false
+python-versions = ">=3.9"
+groups = ["main"]
+files = [
+    {file = "sentencepiece-0.2.2-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:bc7b0b1da20f856bfac5f84b2673fe534b167e41980b27442ca8f78c2b7eb77e"},
+    {file = "sentencepiece-0.2.2-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:8b2db2056c97224e122054fd794543cde5d24b7cae28424f6e3eb79bbe08e42b"},
+    {file = "sentencepiece-0.2.2-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:8f1f61592e7cabd45d49ce8cc0ef42ca655c091e037153754fb3fa59725b5914"},
+    {file = "sentencepiece-0.2.2-cp310-cp310-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:c798f0b327bac10dc95cdac77b9a197ab2bd7dd1e60ebd7586a12d918d4be711"},
+    {file = "sentencepiece-0.2.2-cp310-cp310-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:44284adc6fbe9d5bdd480541431a3d93f674fa44736714d3ad4bcee8283ace7d"},
+    {file = "sentencepiece-0.2.2-cp310-cp310-win_amd64.whl", hash = "sha256:1120e0791540615e650b2e9bea835bf38a7362455d8ab62dee7968219c2d79a0"},
+    {file = "sentencepiece-0.2.2-cp310-cp310-win_arm64.whl", hash = "sha256:524e2a85c028a0d2f9935191fa751e5ef9d9bcc39616f70ab14b28d0369c9936"},
+    {file = "sentencepiece-0.2.2-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:69e9dc8078e128286ed3b975e37c837ba96e215a50c3ef9f3f8b7ab9e5a832a0"},
+    {file = "sentencepiece-0.2.2-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:6dd76f3e5c8b2eb8a3a3efee787bbf5b9a66e52a048fe09cab85eca33fec6790"},
+    {file = "sentencepiece-0.2.2-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:443ac618c7a2a1377cf5c82581fbb849591d14e656d5e5a3e4682d4e36a34e4e"},
+    {file = "sentencepiece-0.2.2-cp311-cp311-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:0e2aae42960392d6dcb9a72d8e1e65a97294c965071b43c7b3429a42f350250e"},
+    {file = "sentencepiece-0.2.2-cp311-cp311-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:1416b92f2f010333786fe6306ed2631121d5ea492219b0841e967b6765e64107"},
+    {file = "sentencepiece-0.2.2-cp311-cp311-win_amd64.whl", hash = "sha256:70d4ca6f4d06df7f0ccab6fe4f49c8a712c8c8b6847b4f0af9a0e1dbb0e0337e"},
+    {file = "sentencepiece-0.2.2-cp311-cp311-win_arm64.whl", hash = "sha256:252908153eeec06c3ca3a32077e64a49d572e3d89881475b4e0f02d99d9fcc7c"},
+    {file = "sentencepiece-0.2.2-cp312-cp312-macosx_10_13_universal2.whl", hash = "sha256:1edb10e520e4bddf74d85b0f5ae74cc2d60c2b448885080bfb618bc2b3a49f6b"},
+    {file = "sentencepiece-0.2.2-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:f7c06c751c19d923435a54bff4f7e66e728fad160e8da28254f133abc9725820"},
+    {file = "sentencepiece-0.2.2-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:38111ed1f79268f399c505028023d5eaaf0ab4e5eafceb709468b0d3323e7838"},
+    {file = "sentencepiece-0.2.2-cp312-cp312-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:cbce24284f51f71d10a42b7b9c964dcb9048b28f1c8e5db40bcbcb6f428cba6a"},
+    {file = "sentencepiece-0.2.2-cp312-cp312-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:c8a168b040bc61681293f79a949b5d911c8e25086f4260285b8d97ab5f1195da"},
+    {file = "sentencepiece-0.2.2-cp312-cp312-win_amd64.whl", hash = "sha256:7c6e7bf684dc12145bfa685d3060beaea55139134ba848289bee514ed42e7383"},
+    {file = "sentencepiece-0.2.2-cp312-cp312-win_arm64.whl", hash = "sha256:76ff5814db72e7462dece042d7593cdf102b8ec82c2b1cc201a2add34ee3050d"},
+    {file = "sentencepiece-0.2.2-cp313-cp313-macosx_10_13_universal2.whl", hash = "sha256:77c3ce990b23441e5ecfa5bce181fd6f408b564aeb6d7e1d1e7de9c5612501c8"},
+    {file = "sentencepiece-0.2.2-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:fd523c4992041faa5c2b3cde62253d11a96c30d73a34afe48a486e8e2254cd1c"},
+    {file = "sentencepiece-0.2.2-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:201a8e0f55501a76e08dbf2c54bc45f4642b379271e89c667d517bfbc2191f2a"},
+    {file = "sentencepiece-0.2.2-cp313-cp313-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:8eed98514bffe5ecac37f493f91869c351fbb05629328bfdbc08502c6c094dc0"},
+    {file = "sentencepiece-0.2.2-cp313-cp313-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:64b656f025355cf8c51abe9fbe3848540756c6d7ca5e6791b1afa664bc24c7cb"},
+    {file = "sentencepiece-0.2.2-cp313-cp313-win_amd64.whl", hash = "sha256:74f0ee601047c0c12a783088b51be4e6214a62ecd9e02278c477433cd16e0ed9"},
+    {file = "sentencepiece-0.2.2-cp313-cp313-win_arm64.whl", hash = "sha256:b23fe17779834d3c27aaf2edac9486d04cca1a7deb8f5facda35150ac6263a91"},
+    {file = "sentencepiece-0.2.2-cp313-cp313t-macosx_10_13_universal2.whl", hash = "sha256:72b7825b331b1b7e7c45be2e674b3e3c65af608fa376bad2d851b20aaf0cdc78"},
+    {file = "sentencepiece-0.2.2-cp313-cp313t-macosx_10_13_x86_64.whl", hash = "sha256:d795c4ac689a57f9d4ba2288126ec7901d389ad5827d2f8b8533c883974fe563"},
+    {file = "sentencepiece-0.2.2-cp313-cp313t-macosx_11_0_arm64.whl", hash = "sha256:3ab3f1ae98970b5590e2209341522718900ba19bcc2c207ffaa6bd417ad960c5"},
+    {file = "sentencepiece-0.2.2-cp313-cp313t-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:3ec27c152a1f1b24bc9168b55a5880f3c16e2334e697da6f55a1046a22405a3d"},
+    {file = "sentencepiece-0.2.2-cp313-cp313t-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:59d6588712101ccfcae9b03692be3aaae1514c2078666d7b05f15ba3a702e41b"},
+    {file = "sentencepiece-0.2.2-cp313-cp313t-win_amd64.whl", hash = "sha256:89625fb43765cccaa1443b9adb61f283e5fe4cb1536728205d06bada730caa53"},
+    {file = "sentencepiece-0.2.2-cp313-cp313t-win_arm64.whl", hash = "sha256:4f0603267cd15b92b68c2c0e852a441507614b70dc7773659baa6b8c214a91fd"},
+    {file = "sentencepiece-0.2.2-cp314-cp314-macosx_10_15_universal2.whl", hash = "sha256:c62bd361cec1f5b556eb8210264ecfff37486cd990c3386cc00310f26c54090a"},
+    {file = "sentencepiece-0.2.2-cp314-cp314-macosx_10_15_x86_64.whl", hash = "sha256:46ba07b543add034de0ff47ac5f907e9a06682f91d85121a972764628933be6b"},
+    {file = "sentencepiece-0.2.2-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:79bac5a251f23a7341e28fda9ce0d5319edf45328239ce037c0682936f137906"},
+    {file = "sentencepiece-0.2.2-cp314-cp314-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:1402d8ee36f0d851cea8eee4dbb85fea14643b7503cf4d00d102eec0fe3ca719"},
+    {file = "sentencepiece-0.2.2-cp314-cp314-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:8d44b20234905ff022b7d535f79d1f823ad7670c9851cc4f03cdc34787cdb3ab"},
+    {file = "sentencepiece-0.2.2-cp314-cp314-win_amd64.whl", hash = "sha256:63250cfab8b80a1ef82a614eb2b3cadfec2c405f870cedc139d08e2f063eb708"},
+    {file = "sentencepiece-0.2.2-cp314-cp314-win_arm64.whl", hash = "sha256:65d84ec36888de4a848eee5f910e67fbc79b064685ef1e10a502e14520ead9c9"},
+    {file = "sentencepiece-0.2.2-cp314-cp314t-macosx_10_15_universal2.whl", hash = "sha256:d254c98ca6387655400b3959c33c83efd807f5edeb608e3aca45800ceaa77151"},
+    {file = "sentencepiece-0.2.2-cp314-cp314t-macosx_10_15_x86_64.whl", hash = "sha256:3fd9ce2ab4460c713cfdeb4aca693ca6732a11538e05fb332d5af42e3d7fde25"},
+    {file = "sentencepiece-0.2.2-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:7fc14c1585139fa6b68775e616a6b90cf622ebf219f9558c0aeaf5d253ee6c9b"},
+    {file = "sentencepiece-0.2.2-cp314-cp314t-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:df88b0c34f2fa909d322f7b06b1398e1e81af4b2f42a7b8e3556f928b25d1811"},
+    {file = "sentencepiece-0.2.2-cp314-cp314t-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:3f5851441ab1ef8634963a5100b733a8bbeefe623e0c5c005b1f1f3880e574cf"},
+    {file = "sentencepiece-0.2.2-cp314-cp314t-win_amd64.whl", hash = "sha256:046b15ea22d8042e2e173561d464ec3b64a9c2081324df70ebce7bf7ebb3e497"},
+    {file = "sentencepiece-0.2.2-cp314-cp314t-win_arm64.whl", hash = "sha256:fa9f5ef0e2a82233dd0b8b32ea3f5710e0c44afbc07ed3620219f32601e56090"},
+    {file = "sentencepiece-0.2.2-cp39-cp39-macosx_10_9_universal2.whl", hash = "sha256:16c84ddef8d3084a8af37208acd365b08092ca089080f1a71fbfdd911adda9b3"},
+    {file = "sentencepiece-0.2.2-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:c76c9b3324efd79029eeb0fd2ced1964bdbeca7d45e030b46fa3ef3cf74f8032"},
+    {file = "sentencepiece-0.2.2-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:54a83df9260a89c1734256e620fe1f1a6bfedd7547139d4dc1384efac11a3a85"},
+    {file = "sentencepiece-0.2.2-cp39-cp39-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:741b4b367140e9b5c36b5a14c72179f2c946d991ea9a7c031a2a1ee6ad097b99"},
+    {file = "sentencepiece-0.2.2-cp39-cp39-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:eb8da9d9a9b418422c21a07fd19b9d9228692b7a7468a45eec6b11642d3c808b"},
+    {file = "sentencepiece-0.2.2-cp39-cp39-win_amd64.whl", hash = "sha256:caad9566e2ef0e5640d36032c69b0edc7ac6028277b93d93815898804fac450c"},
+    {file = "sentencepiece-0.2.2-cp39-cp39-win_arm64.whl", hash = "sha256:cd810878180a52950e5a61f25ada5248a453bbdbafe474f89514135fbc1f633d"},
+    {file = "sentencepiece-0.2.2.tar.gz", hash = "sha256:3d2b5e824b5622038dc7b490897efe05ebbbb9e7350fc142f3ecc8789ef9bdf6"},
+]
+
+[package.extras]
+numpy = ["numpy"]
+protobuf = ["protobuf"]
+test = ["numpy", "protobuf", "pytest"]
+
+[[package]]
+name = "setuptools"
+version = "81.0.0"
+description = "Easily download, build, install, upgrade, and uninstall Python packages"
+optional = false
+python-versions = ">=3.9"
+groups = ["main"]
+markers = "sys_platform == \"linux\" or sys_platform == \"darwin\""
+files = [
+    {file = "setuptools-81.0.0-py3-none-any.whl", hash = "sha256:fdd925d5c5d9f62e4b74b30d6dd7828ce236fd6ed998a08d81de62ce5a6310d6"},
+    {file = "setuptools-81.0.0.tar.gz", hash = "sha256:487b53915f52501f0a79ccfd0c02c165ffe06631443a886740b91af4b7a5845a"},
+]
+
+[package.extras]
+check = ["pytest-checkdocs (>=2.4)", "pytest-ruff (>=0.2.1) ; sys_platform != \"cygwin\"", "ruff (>=0.13.0) ; sys_platform != \"cygwin\""]
+core = ["importlib_metadata (>=6) ; python_version < \"3.10\"", "jaraco.functools (>=4)", "jaraco.text (>=3.7)", "more_itertools", "more_itertools (>=8.8)", "packaging (>=24.2)", "platformdirs (>=4.2.2)", "tomli (>=2.0.1) ; python_version < \"3.11\"", "wheel (>=0.43.0)"]
+cover = ["pytest-cov"]
+doc = ["furo", "jaraco.packaging (>=9.3)", "jaraco.tidelift (>=1.4)", "pygments-github-lexers (==0.0.5)", "pyproject-hooks (!=1.1)", "rst.linker (>=1.9)", "sphinx (>=3.5)", "sphinx-favicon", "sphinx-inline-tabs", "sphinx-lint", "sphinx-notfound-page (>=1,<2)", "sphinx-reredirects", "sphinxcontrib-towncrier", "towncrier (<24.7)"]
+enabler = ["pytest-enabler (>=2.2)"]
+test = ["build[virtualenv] (>=1.0.3)", "filelock (>=3.4.0)", "ini2toml[lite] (>=0.14)", "jaraco.develop (>=7.21) ; python_version >= \"3.9\" and sys_platform != \"cygwin\"", "jaraco.envs (>=2.2)", "jaraco.path (>=3.7.2)", "jaraco.test (>=5.5)", "packaging (>=24.2)", "pip (>=19.1)", "pyproject-hooks (!=1.1)", "pytest (>=6,!=8.1.*)", "pytest-home (>=0.5)", "pytest-perf ; sys_platform != \"cygwin\"", "pytest-subprocess", "pytest-timeout", "pytest-xdist (>=3)", "tomli-w (>=1.0.0)", "virtualenv (>=13.0.0)", "wheel (>=0.44.0)"]
+type = ["importlib_metadata (>=7.0.2) ; python_version < \"3.10\"", "jaraco.develop (>=7.21) ; sys_platform != \"cygwin\"", "mypy (==1.18.*)", "pytest-mypy"]
+
+[[package]]
+name = "shellingham"
+version = "1.5.4"
+description = "Tool to Detect Surrounding Shell"
+optional = false
+python-versions = ">=3.7"
+groups = ["main"]
+files = [
+    {file = "shellingham-1.5.4-py2.py3-none-any.whl", hash = "sha256:7ecfff8f2fd72616f7481040475a65b2bf8af90a56c89140852d1120324e8686"},
+    {file = "shellingham-1.5.4.tar.gz", hash = "sha256:8dbca0739d487e5bd35ab3ca4b36e11c4078f3a234bfce294b0a0291363404de"},
+]
+
+[[package]]
+name = "sympy"
+version = "1.14.0"
+description = "Computer algebra system (CAS) in Python"
+optional = false
+python-versions = ">=3.9"
+groups = ["main"]
+markers = "sys_platform == \"linux\" or sys_platform == \"darwin\""
+files = [
+    {file = "sympy-1.14.0-py3-none-any.whl", hash = "sha256:e091cc3e99d2141a0ba2847328f5479b05d94a6635cb96148ccb3f34671bd8f5"},
+    {file = "sympy-1.14.0.tar.gz", hash = "sha256:d3d3fe8df1e5a0b42f0e7bdf50541697dbe7d23746e894990c030e2b05e72517"},
+]
+
+[package.dependencies]
+mpmath = ">=1.1.0,<1.4"
+
+[package.extras]
+dev = ["hypothesis (>=6.70.0)", "pytest (>=7.1.0)"]
+
+[[package]]
+name = "tokenizers"
+version = "0.23.3"
+description = ""
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "tokenizers-0.23.3-cp310-abi3-macosx_10_12_x86_64.whl", hash = "sha256:9d2b5c97daf61688c2ad1803ca851800feaba50fb68d5821779e9ea5880d968c"},
+    {file = "tokenizers-0.23.3-cp310-abi3-macosx_11_0_arm64.whl", hash = "sha256:68649e97d5b43c44c031d8d848874a6eecae8f8fe40ea989aa777a5a83aca716"},
+    {file = "tokenizers-0.23.3-cp310-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:ec82e80e65a862275b97c3d90b7a523df8d9519ee48aeb4e9625b2cc909274e0"},
+    {file = "tokenizers-0.23.3-cp310-abi3-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:c64a0713180ff16829d4e7f39a658b77ea11443af4e1aa46523692943c9b1414"},
+    {file = "tokenizers-0.23.3-cp310-abi3-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:ddedfd4b3b4be6be24ff6ca645c4a37fddfd305f6f3e354c54cf10b715c48215"},
+    {file = "tokenizers-0.23.3-cp310-abi3-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:2a89614730d7b80940a5d2ed9320e1ec8add5a745c6151d8d05071b7215505b6"},
+    {file = "tokenizers-0.23.3-cp310-abi3-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:e88646b8580c5ad7f4361477f1298e9cc01771a1ee9aecfe32c47b8ff614cc38"},
+    {file = "tokenizers-0.23.3-cp310-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:376851d22bcf9d650a5c3090bb83e6cf9e895fbf0595369fa4cd43c1f69b5f87"},
+    {file = "tokenizers-0.23.3-cp310-abi3-manylinux_2_31_riscv64.whl", hash = "sha256:bf501c40b72d2d5c8623620210430e9cac1ce47a46e45b34107b70a1557d46b0"},
+    {file = "tokenizers-0.23.3-cp310-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:114e2b55ed177179d59f4ab98200a4471e11e78f9e4b5a922d146740f96fcf52"},
+    {file = "tokenizers-0.23.3-cp310-abi3-musllinux_1_2_armv7l.whl", hash = "sha256:d3407fb7b9c4d75dd68850ffd7180bc0a5d2dbaf0762d888e612f31fec3f9c6b"},
+    {file = "tokenizers-0.23.3-cp310-abi3-musllinux_1_2_i686.whl", hash = "sha256:84513ef0aeb8bf8f4ea11a2e8a7ac163ec5288aa115e649a59b470ac5c3107df"},
+    {file = "tokenizers-0.23.3-cp310-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:e05ab7baf7f47b406a95fea6f3b0a484b2ddcd9e1d14b68844c457eb755085a3"},
+    {file = "tokenizers-0.23.3-cp310-abi3-win32.whl", hash = "sha256:1ebf28794e7e4954e20a7f70fbea410b2d1f0418f7dbbca97ca384fcfef38c25"},
+    {file = "tokenizers-0.23.3-cp310-abi3-win_amd64.whl", hash = "sha256:1f0823bb00c5fdc98e487354d54dd55a03848d61a1a0bf29a68c77f24f3b26c3"},
+    {file = "tokenizers-0.23.3-cp310-abi3-win_arm64.whl", hash = "sha256:7e48734d2de9260d86f03ab056d2cfeeff3869f61dbd49aaa15a2793b5f3458b"},
+    {file = "tokenizers-0.23.3-cp314-cp314t-macosx_10_12_x86_64.whl", hash = "sha256:efa3d7318406b4d115dce61ad5061953f1f44b128e79c020ce4615d763e23b6e"},
+    {file = "tokenizers-0.23.3-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:a4fbb3662f9f59d199d61338e54b4bcc11d07ebbb1aeb3540dacb2be9c521cb7"},
+    {file = "tokenizers-0.23.3-cp314-cp314t-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:de536665495cb4b409d25bade41963f801aff4225c19a6b804b048f7d14e34c7"},
+    {file = "tokenizers-0.23.3-cp314-cp314t-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:5cc24bb457dd4a8af89c8fcb40074d570129ec473df2a866c276ee55db4749d7"},
+    {file = "tokenizers-0.23.3-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:acd5c57b4bd3e56e246e2731a3a3a6825a7a7d89b7e3b761ba80bc521710f04b"},
+    {file = "tokenizers-0.23.3-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:82eb480f6f1c21cea3349dec32cf1a6384c6c1e775f00f83b0d51197bc013687"},
+    {file = "tokenizers-0.23.3-cp314-cp314t-win_amd64.whl", hash = "sha256:1554a6eed34d9d6a78d23360f4e06df8dffab1ae08c7e8488e0b3e3b36cc266f"},
+    {file = "tokenizers-0.23.3.tar.gz", hash = "sha256:cded33237c77caeef62944d32aa9a7ef42bdce2b3497e18d137e072a8c4be438"},
+]
+
+[package.dependencies]
+huggingface-hub = ">=0.16.4,<3.0"
+
+[package.extras]
+dev = ["tokenizers[testing]"]
+docs = ["setuptools-rust", "sphinx", "sphinx-rtd-theme"]
+testing = ["datasets", "numpy", "pytest", "pytest-asyncio", "requests", "ruff", "ty"]
+
+[[package]]
+name = "torch"
+version = "2.14.1"
+description = "Tensors and Dynamic neural networks in Python with strong GPU acceleration"
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+markers = "sys_platform == \"darwin\""
+files = [
+    {file = "torch-2.14.1-cp310-cp310-macosx_14_0_arm64.whl", hash = "sha256:e24b9815001d18e73e76eae1ea25f8ef908c6103e857378ad20f8f6cf5c954f5"},
+    {file = "torch-2.14.1-cp310-cp310-manylinux_2_28_aarch64.whl", hash = "sha256:957df236525bcd940ccd9ca59074e9821182171253417c0d98e7b82647254799"},
+    {file = "torch-2.14.1-cp310-cp310-manylinux_2_28_x86_64.whl", hash = "sha256:1fee6b20869a90ccec05e35559a706338ac8aab40a9c4fb58f16667e93af36ae"},
+    {file = "torch-2.14.1-cp310-cp310-win_amd64.whl", hash = "sha256:9ee7d50ec80ae5b4258cafc0ea108711988d961f21d0a68cf685650da2812e5f"},
+    {file = "torch-2.14.1-cp311-cp311-macosx_14_0_arm64.whl", hash = "sha256:b6074b130fd26bc50f5d171f07dfed70db22dd0c354ab34767729249f9e0f042"},
+    {file = "torch-2.14.1-cp311-cp311-manylinux_2_28_aarch64.whl", hash = "sha256:ef1276fd07adb8d463ac402a24ea7482e45c9a42df8ebdac211aa757779200ba"},
+    {file = "torch-2.14.1-cp311-cp311-manylinux_2_28_x86_64.whl", hash = "sha256:305a61f61f35f128579f299c5bd33d475f6a01c6307336139632e30856c4854d"},
+    {file = "torch-2.14.1-cp311-cp311-win_amd64.whl", hash = "sha256:0bd2fb0c4098856bc5af59fe5988bf38bdbcdb6550b44273c12ae425cb8ec17e"},
+    {file = "torch-2.14.1-cp312-cp312-macosx_14_0_arm64.whl", hash = "sha256:420dbf314c180ee4b86e9bc00aee5746a7d6e5bacdd7df925af671ed393f0b2e"},
+    {file = "torch-2.14.1-cp312-cp312-manylinux_2_28_aarch64.whl", hash = "sha256:0b13acde294b401509830d85a8906cbf2fc97425a12a66cce1d9fa13c5106a2a"},
+    {file = "torch-2.14.1-cp312-cp312-manylinux_2_28_x86_64.whl", hash = "sha256:23011fe29a99b591eabb3a31a25080c62e2c2f0690a3c798744e5489dae50451"},
+    {file = "torch-2.14.1-cp312-cp312-win_amd64.whl", hash = "sha256:38bee9f2a2ccfc6898172a5075e4ba52522fbe143eeb099674fc67ab390e858d"},
+    {file = "torch-2.14.1-cp313-cp313-macosx_14_0_arm64.whl", hash = "sha256:dbe359d705f4d67236743794c296c6dee93a922fd8117eff8c3e880d7d0fb2b9"},
+    {file = "torch-2.14.1-cp313-cp313-manylinux_2_28_aarch64.whl", hash = "sha256:6d530bd11479fb574567af3a1f384af9a77bf5bb9550cde8080b356ca7220d5a"},
+    {file = "torch-2.14.1-cp313-cp313-manylinux_2_28_x86_64.whl", hash = "sha256:c8f71aabc67bcbfc9373dc131537a5968d04edce73e88add21354a7cd0a76985"},
+    {file = "torch-2.14.1-cp313-cp313-win_amd64.whl", hash = "sha256:711713391d26a1ce5e9fbc6c996d954a8c12e8825374cc77b809a6af29539b8c"},
+    {file = "torch-2.14.1-cp314-cp314-macosx_14_0_arm64.whl", hash = "sha256:cee091caf2a6229e248daf41d18ceebf590ba02d9179205610062c64ee5fef03"},
+    {file = "torch-2.14.1-cp314-cp314-manylinux_2_28_aarch64.whl", hash = "sha256:f68f5476e2bc0e8f60b74f7ca21acda885c4477af6392f8977e4f0d1ea1aa162"},
+    {file = "torch-2.14.1-cp314-cp314-manylinux_2_28_x86_64.whl", hash = "sha256:1d4df07be7338bbcc4d54085adee21363c91314702b9bd4d78ef72ffac9465ba"},
+    {file = "torch-2.14.1-cp314-cp314-win_amd64.whl", hash = "sha256:d02a4c48a2ca5fb7654e36e71f710f74494d83f1f10aaff8e059dd554adca956"},
+    {file = "torch-2.14.1-cp314-cp314t-macosx_14_0_arm64.whl", hash = "sha256:06c3ec25f3b497f9a73dc0c36f293f295cfe8d5584fd446238d9a501b30a66d5"},
+    {file = "torch-2.14.1-cp314-cp314t-manylinux_2_28_aarch64.whl", hash = "sha256:9113f94f70429f9f302bf55b090b5411269083e8a72a5d015ffcc5a7f83f2c69"},
+    {file = "torch-2.14.1-cp314-cp314t-manylinux_2_28_x86_64.whl", hash = "sha256:e65d5fe136e533b23c2d134377f7f126721c7c78dd75b3dbb735af082a8aeb85"},
+    {file = "torch-2.14.1-cp314-cp314t-win_amd64.whl", hash = "sha256:e07306caa1de2a4ac1467e11ecfc92fc44f523dd6a521145039aef46d913963c"},
+]
+
+[package.dependencies]
+cuda-bindings = {version = ">=13.0.3,<14", markers = "platform_system == \"Linux\" and python_version < \"3.15\""}
+cuda-toolkit = {version = "13.0.3", extras = ["cublas", "cudart", "cufft", "cufile", "cupti", "curand", "cusolver", "cusparse", "nvjitlink", "nvrtc", "nvtx"], markers = "platform_system == \"Linux\""}
+filelock = "*"
+fsspec = ">=0.8.5"
+jinja2 = "*"
+networkx = ">=2.5.1"
+nvidia-cudnn-cu13 = {version = "9.24.0.43", markers = "platform_system == \"Linux\""}
+nvidia-cusparselt-cu13 = {version = "0.8.1", markers = "platform_system == \"Linux\""}
+nvidia-nccl-cu13 = {version = "2.30.7", markers = "platform_system == \"Linux\""}
+nvidia-nvshmem-cu13 = {version = "3.4.5", markers = "platform_system == \"Linux\""}
+setuptools = ">=77.0.3"
+sympy = ">=1.13.3"
+triton = {version = ">=3.8.0,<3.9.0", markers = "platform_system == \"Linux\" and python_version < \"3.15\""}
+typing-extensions = ">=4.10.0"
+
+[package.extras]
+opt-einsum = ["opt-einsum (>=3.3)"]
+optree = ["optree (>=0.13.0)"]
+pyyaml = ["pyyaml"]
+
+[[package]]
+name = "torch"
+version = "2.14.1+cpu"
+description = "Tensors and Dynamic neural networks in Python with strong GPU acceleration"
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+markers = "sys_platform == \"linux\""
+files = [
+    {file = "torch-2.14.1+cpu-cp310-cp310-linux_s390x.whl", hash = "sha256:24b279ee181df14dc353f3e3f0e3f088d88a38175989dadb5645c0194830d4de"},
+    {file = "torch-2.14.1+cpu-cp310-cp310-manylinux_2_28_aarch64.whl", hash = "sha256:b869fe9358c90dd82d02db89aef89729df93a89a48b7c3e21a7719abf55586b3"},
+    {file = "torch-2.14.1+cpu-cp310-cp310-manylinux_2_28_x86_64.whl", hash = "sha256:4ce2a6ae7d302bba56ef7c5c92d9f752728896cfe2626a7f3e5293f879e36650"},
+    {file = "torch-2.14.1+cpu-cp310-cp310-win_amd64.whl", hash = "sha256:bda7043ec45539b1b61538af468e19854b72ac6be8d6fde2f51dfd0c91721568"},
+    {file = "torch-2.14.1+cpu-cp311-cp311-linux_s390x.whl", hash = "sha256:cccd0a609fe3f83ab2ff651b627d5bd68c7fb33ee77a994b31c18b7c9a6f9a52"},
+    {file = "torch-2.14.1+cpu-cp311-cp311-manylinux_2_28_aarch64.whl", hash = "sha256:36a03a3f87ec875ca39c8e29eccbaf0fc704605f78a3f7f179f157048c3c395e"},
+    {file = "torch-2.14.1+cpu-cp311-cp311-manylinux_2_28_x86_64.whl", hash = "sha256:5e38154c8896d426a5df58bf2276b603e82184dfae4f54d77dbb212f36e34f37"},
+    {file = "torch-2.14.1+cpu-cp311-cp311-win_amd64.whl", hash = "sha256:d14b0c8a66345b7ba43adfa86fc20d89087a0a92e440004755d49d2468ddc6e0"},
+    {file = "torch-2.14.1+cpu-cp311-cp311-win_arm64.whl", hash = "sha256:100df7e44af5c973606cfef571a14b36b0acfd165d6bbc7aabe3edac95c9ccbd"},
+    {file = "torch-2.14.1+cpu-cp312-cp312-linux_s390x.whl", hash = "sha256:6b06a7f6275c5fdecaf1dbf5546f0820b024a97785575bb99e3d9930d83cd86b"},
+    {file = "torch-2.14.1+cpu-cp312-cp312-manylinux_2_28_aarch64.whl", hash = "sha256:2af649994b36dffc627175c3f306d9e513a309b58d4619ac7107577613336454"},
+    {file = "torch-2.14.1+cpu-cp312-cp312-manylinux_2_28_x86_64.whl", hash = "sha256:5a6363570c753812540a05eb82380e329469cbe668643e88111414c12627711f"},
+    {file = "torch-2.14.1+cpu-cp312-cp312-win_amd64.whl", hash = "sha256:c52adc3bd526ff7fc5bdcc58fc4a76e6214d09951b646ffe5256ebd7ba60cef1"},
+    {file = "torch-2.14.1+cpu-cp312-cp312-win_arm64.whl", hash = "sha256:2f8162c6b265eaf2b0bdd6bef20e96cece430f308e1477a32d0aef51e472124c"},
+    {file = "torch-2.14.1+cpu-cp313-cp313-linux_s390x.whl", hash = "sha256:4b323c71a5ac184485c87d694842157b1eb217c425ad69761dafd379d1590521"},
+    {file = "torch-2.14.1+cpu-cp313-cp313-manylinux_2_28_aarch64.whl", hash = "sha256:783a3dec10d272e40ed3f1a05ec228400300d3229b8f7521a9327645fae88646"},
+    {file = "torch-2.14.1+cpu-cp313-cp313-manylinux_2_28_x86_64.whl", hash = "sha256:331fa474e26e428e2e6af1143b2fdab19cee492d7d21897b8899ce1c14755662"},
+    {file = "torch-2.14.1+cpu-cp313-cp313-win_amd64.whl", hash = "sha256:4eec5a4fadd7a18f9c75b97fdb374fbb579ce4b1604bf66eeb14dc2201d54def"},
+    {file = "torch-2.14.1+cpu-cp313-cp313-win_arm64.whl", hash = "sha256:a46884e8b38cbdf5af8d4fdd39223a5fa97491419c3417c93f9b3e28749fa356"},
+    {file = "torch-2.14.1+cpu-cp314-cp314-linux_s390x.whl", hash = "sha256:a7a580e3d92467516c9247a22bc805ae4f79c7b1b9a2052f92849f34e89720f5"},
+    {file = "torch-2.14.1+cpu-cp314-cp314-manylinux_2_28_aarch64.whl", hash = "sha256:f56cd23a3bf003bfbed3da0645ba09b3ccc66c0a2f2ac726a9b98f34e0e4926f"},
+    {file = "torch-2.14.1+cpu-cp314-cp314-manylinux_2_28_x86_64.whl", hash = "sha256:35103e180214793207f95d7e12e100ecba3ace6f5fb37e1fcefbb07e096ed180"},
+    {file = "torch-2.14.1+cpu-cp314-cp314-win_amd64.whl", hash = "sha256:a93db7947d53caab62c2436205a60ddf9c3150c2e09eb9b24610c1c3f4cbaf99"},
+    {file = "torch-2.14.1+cpu-cp314-cp314t-linux_s390x.whl", hash = "sha256:f67ee1fe9c9f7c6ad3785d150bc3cecda442a5880a6d40816e13fee475d97dd8"},
+    {file = "torch-2.14.1+cpu-cp314-cp314t-manylinux_2_28_aarch64.whl", hash = "sha256:2547c9f32f8fd9c8fced83119269d85519b123d4e8c87eba71660fe3a6996336"},
+    {file = "torch-2.14.1+cpu-cp314-cp314t-manylinux_2_28_x86_64.whl", hash = "sha256:7d2c35e915793f3ae82e8ee4041a960140bcd0bd35f3ec6eb06e1f1f37a89dba"},
+    {file = "torch-2.14.1+cpu-cp314-cp314t-win_amd64.whl", hash = "sha256:5c0d1ef51127c9c65ec6394226b82ef1a5f3cd0ab8ffb10f48cc5c08aa72a5cf"},
+    {file = "torch-2.14.1+cpu-cp315-cp315-manylinux_2_28_aarch64.whl", hash = "sha256:88634fad89f7b953acd011bad0d5c131d269aa71a7de6e0ef2344e3a2c95fabc"},
+    {file = "torch-2.14.1+cpu-cp315-cp315-manylinux_2_28_x86_64.whl", hash = "sha256:94edee897fbd15a73f91aa38e22e2d2b7b60ac3acf6e8afc6b8dad89229c810c"},
+    {file = "torch-2.14.1+cpu-cp315-cp315-win_amd64.whl", hash = "sha256:e799161f94d97485857ec82424e40a83d7648ab02c03e3d6e71c7838a4d7b89e"},
+    {file = "torch-2.14.1+cpu-cp315-cp315t-manylinux_2_28_aarch64.whl", hash = "sha256:930d724490955b822c0b005499128279e54876a318b51fdb9b9da35e0230cc0f"},
+    {file = "torch-2.14.1+cpu-cp315-cp315t-manylinux_2_28_x86_64.whl", hash = "sha256:2948f065d848178a0402791baf87493b7ebf067dae48a8e04c498d6668e94895"},
+    {file = "torch-2.14.1+cpu-cp315-cp315t-win_amd64.whl", hash = "sha256:192745b92c9ac392298e73de950cd02d8a7886117923073d890625622d343ede"},
+]
+
+[package.dependencies]
+filelock = "*"
+fsspec = ">=0.8.5"
+jinja2 = "*"
+networkx = ">=2.5.1"
+setuptools = ">=77.0.3"
+sympy = ">=1.13.3"
+typing-extensions = ">=4.10.0"
+
+[package.extras]
+opt-einsum = ["opt-einsum (>=3.3)"]
+optree = ["optree (>=0.13.0)"]
+pyyaml = ["pyyaml"]
+
+[package.source]
+type = "legacy"
+url = "https://download.pytorch.org/whl/cpu"
+reference = "pytorch-cpu"
+
+[[package]]
+name = "tqdm"
+version = "4.70.1"
+description = "Fast, Extensible Progress Meter"
+optional = false
+python-versions = ">=3.8"
+groups = ["main"]
+files = [
+    {file = "tqdm-4.70.1-py3-none-any.whl", hash = "sha256:c293e525e6fef9c20e8728fd4612df02a0aa31bb5fe91ecd93e123b1b7bffa73"},
+    {file = "tqdm-4.70.1.tar.gz", hash = "sha256:cefd0eca11b2a37a3aee776544d4f4ae913f02688135b5556b8788dfa474afc4"},
+]
+
+[package.dependencies]
+colorama = {version = "*", markers = "platform_system == \"Windows\""}
+
+[package.extras]
+discord = ["envwrap", "requests"]
+notebook = ["ipywidgets (>=6)"]
+slack = ["envwrap", "slack-sdk"]
+telegram = ["envwrap", "requests"]
+
+[[package]]
+name = "transformers"
+version = "5.20.2"
+description = "Transformers: the model-definition framework for state-of-the-art machine learning models in text, vision, audio, and multimodal models, for both inference and training."
+optional = false
+python-versions = ">=3.10.0"
+groups = ["main"]
+files = [
+    {file = "transformers-5.20.2-py3-none-any.whl", hash = "sha256:7ae75dbc1fb3922aff1b6c97b9e5479d6cf26fd86a8866bfb71324190479f807"},
+    {file = "transformers-5.20.2.tar.gz", hash = "sha256:080871e061546d720be57f5846fce426f46b44d61574907cc7404c2cba49d180"},
+]
+
+[package.dependencies]
+huggingface-hub = ">=1.31.0,<3.0"
+numpy = ">=1.17"
+packaging = ">=20.0"
+pyyaml = ">=5.1"
+regex = ">=2025.10.22"
+safetensors = ">=0.8.0"
+tokenizers = ">=0.23.1,<0.24.0"
+tqdm = ">=4.60"
+typer = "*"
+
+[package.extras]
+accelerate = ["accelerate (>=1.1.0)"]
+all = ["Pillow (>=10.0.1,<=15.0)", "accelerate (>=1.1.0)", "av", "blobfile", "jinja2 (>=3.1.0)", "kenlm", "kernels (>=0.17.2,<0.18)", "librosa", "mistral-common[image] (>=1.11.7)", "num2words", "phonemizer", "protobuf", "pyctcdecode (>=0.4.0)", "sentencepiece (>=0.1.91,!=0.1.92)", "tiktoken", "timm (>=1.0.23)", "torch (>=2.6)", "torchaudio", "torchvision"]
+audio = ["kenlm", "librosa", "phonemizer", "pyctcdecode (>=0.4.0)", "torchaudio"]
+benchmark = ["optimum-benchmark (>=0.3.0)"]
+chat-template = ["jinja2 (>=3.1.0)"]
+codecarbon = ["codecarbon (>=2.8.1)"]
+deepspeed = ["accelerate (>=1.1.0)", "deepspeed (>=0.9.3)"]
+deepspeed-testing = ["GitPython (<3.1.19)", "accelerate (>=1.1.0)", "accelerate (>=1.1.0)", "beautifulsoup4", "datasets (>=2.15.0)", "datasets (>=2.15.0)", "deepspeed (>=0.9.3)", "diffusers (==0.35.2)", "faiss-cpu", "fastapi", "filelock", "hf-doc-builder", "libcst", "mistral-common[image] (>=1.11.7)", "nltk (<=3.8.1)", "openai (>=1.98.0)", "optuna", "parameterized (>=0.9)", "protobuf", "protobuf", "psutil", "pydantic (>=2)", "pytest (>=7.2.0,<9.0.0)", "pytest-asyncio (>=1.2.0)", "pytest-env", "pytest-order", "pytest-random-order", "pytest-rerunfailures (<16.0)", "pytest-rich", "pytest-timeout", "pytest-xdist", "rich", "rich", "rjieba", "ruff (==0.14.10)", "sacremoses", "sentencepiece (>=0.1.91,!=0.1.92)", "sentencepiece (>=0.1.91,!=0.1.92)", "starlette", "tensorboard", "timeout-decorator", "tomli", "torch (>=2.6)", "transformers-mlinter (==0.1.5)", "ty (==0.0.20)", "urllib3 (<2.0.0)", "uvicorn"]
+dev = ["GitPython (<3.1.19)", "Pillow (>=10.0.1,<=15.0)", "accelerate (>=1.1.0)", "accelerate (>=1.1.0)", "av", "beautifulsoup4", "blobfile", "datasets (>=2.15.0)", "datasets (>=2.15.0)", "diffusers (==0.35.2)", "faiss-cpu", "fastapi", "filelock", "fugashi (>=1.0)", "hf-doc-builder", "ipadic (>=1.0.0,<2.0)", "jinja2 (>=3.1.0)", "kenlm", "kernels (>=0.17.2,<0.18)", "libcst", "librosa", "mistral-common[image] (>=1.11.7)", "mistral-common[image] (>=1.11.7)", "nltk (<=3.8.1)", "num2words", "openai (>=1.98.0)", "parameterized (>=0.9)", "phonemizer", "protobuf", "protobuf", "psutil", "pyctcdecode (>=0.4.0)", "pydantic (>=2)", "pytest (>=7.2.0,<9.0.0)", "pytest-asyncio (>=1.2.0)", "pytest-env", "pytest-order", "pytest-random-order", "pytest-rerunfailures (<16.0)", "pytest-rich", "pytest-timeout", "pytest-xdist", "rhoknp (>=1.1.0,<1.3.1)", "rich", "rich", "rjieba", "ruff (==0.14.10)", "sacremoses", "scikit-learn", "sentencepiece (>=0.1.91,!=0.1.92)", "sentencepiece (>=0.1.91,!=0.1.92)", "starlette", "sudachidict_core (>=20220729)", "sudachipy (>=0.6.6)", "tensorboard", "tiktoken", "timeout-decorator", "timm (>=1.0.23)", "tomli", "torch (>=2.6)", "torch (>=2.6)", "torchaudio", "torchvision", "transformers-mlinter (==0.1.5)", "ty (==0.0.20)", "unidic (>=1.0.2)", "unidic_lite (>=1.0.7)", "urllib3 (<2.0.0)", "uvicorn"]
+docs = ["hf-doc-builder"]
+integrations = ["codecarbon (>=2.8.1)", "kernels (>=0.17.2,<0.18)", "optuna", "ray[tune] (>=2.7.0)"]
+ja = ["fugashi (>=1.0)", "ipadic (>=1.0.0,<2.0)", "rhoknp (>=1.1.0,<1.3.1)", "sudachidict_core (>=20220729)", "sudachipy (>=0.6.6)", "unidic (>=1.0.2)", "unidic_lite (>=1.0.7)"]
+kernels = ["kernels (>=0.17.2,<0.18)"]
+mistral-common = ["mistral-common[image] (>=1.11.7)"]
+num2words = ["num2words"]
+optuna = ["optuna"]
+quality = ["GitPython (<3.1.19)", "datasets (>=2.15.0)", "libcst", "rich", "ruff (==0.14.10)", "tomli", "transformers-mlinter (==0.1.5)", "ty (==0.0.20)", "urllib3 (<2.0.0)"]
+ray = ["ray[tune] (>=2.7.0)"]
+retrieval = ["datasets (>=2.15.0)", "faiss-cpu"]
+sagemaker = ["sagemaker (>=2.31.0)"]
+sentencepiece = ["protobuf", "sentencepiece (>=0.1.91,!=0.1.92)"]
+serving = ["accelerate (>=1.1.0)", "fastapi", "openai (>=1.98.0)", "pydantic (>=2)", "rich", "starlette", "torch (>=2.6)", "uvicorn"]
+sklearn = ["scikit-learn"]
+testing = ["GitPython (<3.1.19)", "accelerate (>=1.1.0)", "beautifulsoup4", "datasets (>=2.15.0)", "datasets (>=2.15.0)", "diffusers (==0.35.2)", "faiss-cpu", "fastapi", "filelock", "hf-doc-builder", "libcst", "mistral-common[image] (>=1.11.7)", "nltk (<=3.8.1)", "openai (>=1.98.0)", "parameterized (>=0.9)", "protobuf", "psutil", "pydantic (>=2)", "pytest (>=7.2.0,<9.0.0)", "pytest-asyncio (>=1.2.0)", "pytest-env", "pytest-order", "pytest-random-order", "pytest-rerunfailures (<16.0)", "pytest-rich", "pytest-timeout", "pytest-xdist", "rich", "rich", "rjieba", "ruff (==0.14.10)", "sacremoses", "sentencepiece (>=0.1.91,!=0.1.92)", "starlette", "tensorboard", "timeout-decorator", "tomli", "torch (>=2.6)", "transformers-mlinter (==0.1.5)", "ty (==0.0.20)", "urllib3 (<2.0.0)", "uvicorn"]
+tiktoken = ["blobfile", "tiktoken"]
+timm = ["timm (>=1.0.23)"]
+torch = ["accelerate (>=1.1.0)", "torch (>=2.6)"]
+video = ["av"]
+vision = ["Pillow (>=10.0.1,<=15.0)", "torchvision"]
+
+[[package]]
+name = "triton"
+version = "3.8.0"
+description = "A language and compiler for custom Deep Learning operations"
+optional = false
+python-versions = "<3.15,>=3.10"
+groups = ["main"]
+markers = "platform_system == \"Linux\" and sys_platform == \"darwin\""
+files = [
+    {file = "triton-3.8.0-cp310-cp310-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:398f4b009c7ab08ed9aeb1d1282dee822945b53f399a5e12a4fecb643cf2007d"},
+    {file = "triton-3.8.0-cp310-cp310-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:6d914c52f89dc942b1819db959e7c07a5999f678ec760f677625c706b7e0d743"},
+    {file = "triton-3.8.0-cp311-cp311-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:372285307d4c44ee74cee32de0b4f04bd157e071427e38c6f4ee3e3beb2194f4"},
+    {file = "triton-3.8.0-cp311-cp311-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:68988ac85d5e7086baeda0ddc175af9667db7529b3c5e11a5c0601b8bef2200a"},
+    {file = "triton-3.8.0-cp312-cp312-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:a9c404c69ed4a39e8ec632eaf6b9fe058a060bf98979c177f6ef666f06bb8d50"},
+    {file = "triton-3.8.0-cp312-cp312-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:e91ffa46d095b252248297292dd22bcbacd53a125a0c2eefbbbf74925a320bc3"},
+    {file = "triton-3.8.0-cp313-cp313-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:b7004666652f500ed854a86988e4b3d69d247188b5d2092b5df1e44f4a954099"},
+    {file = "triton-3.8.0-cp313-cp313-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:1f0497218e26b7d79773ad9c2a3fa3b539ee69f587a13fac2e552b1d322a8015"},
+    {file = "triton-3.8.0-cp314-cp314-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:1f6b48d0591929a3867973acac3dccd4e058585f91bfb41022de496c9ffab304"},
+    {file = "triton-3.8.0-cp314-cp314-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:387dae4cb0089a7b6ba1a428ae0782b65c4c58f57d94617cb22ca8593d8ccbca"},
+    {file = "triton-3.8.0-cp314-cp314t-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:1b84e7d512490ba529111260fa6f7cad8b254a6bb5fbdf41d5ef9a5e57f52d0a"},
+    {file = "triton-3.8.0-cp314-cp314t-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:74217bb56ed8692759227758e4c4b3bd2d608a209c1a7a081bf361fb4c2c1bf9"},
+]
+
+[[package]]
+name = "truststore"
+version = "0.10.5"
+description = "Verify certificates using native system trust stores"
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+markers = "sys_platform != \"emscripten\""
+files = [
+    {file = "truststore-0.10.5-py3-none-any.whl", hash = "sha256:9aaaedaefaf06d8b206278cf8b5012bc897f485a874503501e12d776df78951c"},
+    {file = "truststore-0.10.5.tar.gz", hash = "sha256:30d36967ccaded5cbb38d602c433f53600036c79d502f4533a49b60a03bbefcd"},
+]
+
+[[package]]
+name = "typer"
+version = "0.27.3"
+description = "Typer, build great CLIs. Easy to code. Based on Python type hints."
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "typer-0.27.3-py3-none-any.whl", hash = "sha256:e50022f28b82a86313e54501317a1db64bf8f8d036ff8cfe5ca7e47675454aff"},
+    {file = "typer-0.27.3.tar.gz", hash = "sha256:d0396f770a560ab1b0a8504e13b5f254b728cedb05c61cf0359e944e50ce8901"},
+]
+
+[package.dependencies]
+annotated-doc = ">=0.0.2"
+colorama = {version = "*", markers = "platform_system == \"Windows\""}
+rich = ">=13.8.0"
+shellingham = ">=1.3.0"
+
+[[package]]
+name = "typing-extensions"
+version = "4.16.0"
+description = "Backported and Experimental Type Hints for Python 3.9+"
+optional = false
+python-versions = ">=3.9"
+groups = ["main"]
+files = [
+    {file = "typing_extensions-4.16.0-py3-none-any.whl", hash = "sha256:481caa481374e813c1b176ada14e97f1f67a4539ce9cfeb3f350d78d6370c2e8"},
+    {file = "typing_extensions-4.16.0.tar.gz", hash = "sha256:dc983d19a509c94dba722ee6abd33940f7c05a89e243c47e907eb4db6f1a43e5"},
+]
+
+[[package]]
+name = "urllib3"
+version = "2.8.0"
+description = "HTTP library with thread-safe connection pooling, file post, and more."
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "urllib3-2.8.0-py3-none-any.whl", hash = "sha256:0cf3cae568d36aa9576b28dfb35f11328f1cb974ca7647d9475ebb86c75ac6e3"},
+    {file = "urllib3-2.8.0.tar.gz", hash = "sha256:63bf2ead4c879426ebf22ef2a781eeb4aa3b4ae798a0435506f8687fd5bb9b63"},
+]
+
+[package.extras]
+brotli = ["brotli (>=1.2.0) ; platform_python_implementation == \"CPython\"", "brotlicffi (>=1.2.0.0) ; platform_python_implementation != \"CPython\""]
+h2 = ["h2 (>=4,<5)"]
+socks = ["pysocks (>=1.5.6,!=1.5.7,<2.0)"]
+zstd = ["backports-zstd (>=1.0.0) ; python_version < \"3.14\""]
+
+[metadata]
+lock-version = "2.1"
+python-versions = ">=3.12,<3.14"
+content-hash = "bf2ec9e50ee1c8d9edabb53f592c3c13f217e1f14da6322466cede513a81599f"
diff --git a/paddler_llama_cpp_converter_python/pyproject.toml b/paddler_llama_cpp_converter_python/pyproject.toml
new file mode 100644
index 000000000..862ccc6bc
--- /dev/null
+++ b/paddler_llama_cpp_converter_python/pyproject.toml
@@ -0,0 +1,30 @@
+[tool.poetry]
+name = "paddler-llama-cpp-converter"
+version = "5.0.0"
+description = "Runs the Hugging Face to GGUF converter shipped with the llama.cpp sources Paddler's bindings build"
+authors = ["Intentee"]
+license = "Apache-2.0"
+package-mode = false
+
+[tool.poetry.dependencies]
+python = ">=3.12,<3.14"
+numpy = "==2.5.4"
+protobuf = "==7.36.2"
+pyyaml = "==6.0.3"
+requests = "==2.34.2"
+sentencepiece = "==0.2.2"
+torch = [
+    { version = "==2.14.1+cpu", source = "pytorch-cpu", markers = "sys_platform == 'linux'" },
+    { version = "==2.14.1", source = "PyPI", markers = "sys_platform == 'darwin'" },
+]
+tqdm = "==4.70.1"
+transformers = "==5.20.2"
+
+[[tool.poetry.source]]
+name = "pytorch-cpu"
+url = "https://download.pytorch.org/whl/cpu"
+priority = "explicit"
+
+[build-system]
+requires = ["poetry-core"]
+build-backend = "poetry.core.masonry.api"
diff --git a/paddler_local_http_fixture/Cargo.toml b/paddler_local_http_fixture/Cargo.toml
index bd6960a83..163e596e6 100644
--- a/paddler_local_http_fixture/Cargo.toml
+++ b/paddler_local_http_fixture/Cargo.toml
@@ -1,5 +1,6 @@
 [package]
 name = "paddler_local_http_fixture"
+autotests = false
 authors.workspace = true
 description = "Local HTTP file host that stands in for third-party model servers in Paddler tests"
 edition.workspace = true
@@ -9,10 +10,18 @@ publish = false
 repository.workspace = true
 version.workspace = true
 
+[[test]]
+name = "integration"
+path = "tests/main.rs"
+
 [dependencies]
 http = { workspace = true }
 httparse = { workspace = true }
 tokio = { workspace = true }
 
+[dev-dependencies]
+nix = { workspace = true }
+paddler_tcp_buffer_limits = { workspace = true }
+
 [lints]
 workspace = true
diff --git a/paddler_local_http_fixture/src/fixture_response.rs b/paddler_local_http_fixture/src/fixture_response.rs
index c3e63cc06..8c595b0a7 100644
--- a/paddler_local_http_fixture/src/fixture_response.rs
+++ b/paddler_local_http_fixture/src/fixture_response.rs
@@ -1,8 +1,17 @@
+use std::time::Duration;
+
 use http::StatusCode;
 
 #[derive(Clone)]
 pub enum FixtureResponse {
-    CloseBeforeHeaders,
+    EndlessTrickle {
+        byte_interval: Duration,
+    },
+    HuggingFaceFile {
+        body: Vec<u8>,
+        commit_hash: String,
+        etag: String,
+    },
     Ok(Vec<u8>),
     PartialContent {
         body: Vec<u8>,
@@ -19,4 +28,10 @@ pub enum FixtureResponse {
         status: StatusCode,
         withheld_byte_count: usize,
     },
+    TruncatedHuggingFaceFile {
+        commit_hash: String,
+        etag: String,
+        sent_body: Vec<u8>,
+        withheld_byte_count: usize,
+    },
 }
diff --git a/paddler_local_http_fixture/src/local_http_fixture.rs b/paddler_local_http_fixture/src/local_http_fixture.rs
index 8d920d45a..c1e8c765a 100644
--- a/paddler_local_http_fixture/src/local_http_fixture.rs
+++ b/paddler_local_http_fixture/src/local_http_fixture.rs
@@ -3,23 +3,40 @@ use std::net::Ipv4Addr;
 use std::net::SocketAddr;
 use std::sync::Arc;
 
+use std::io::ErrorKind;
+
 use tokio::net::TcpListener;
+use tokio::select;
 use tokio::spawn;
+use tokio::sync::oneshot;
+use tokio::sync::watch::error::RecvError;
+use tokio::task::JoinError;
 use tokio::task::JoinHandle;
 
 use crate::fixture_response::FixtureResponse;
 use crate::observed_requests::ObservedRequests;
 use crate::serve_connection::serve_connection;
 
-async fn serve_sequentially(
+async fn serve_sequentially_until_stopped(
     listener: TcpListener,
     response: FixtureResponse,
     observed_requests: Arc<ObservedRequests>,
+    mut serving_stopped: oneshot::Receiver<()>,
 ) -> io::Result<()> {
     loop {
-        let (stream, _peer_addr) = listener.accept().await?;
+        let (stream, _peer_addr) = select! {
+            biased;
+            accepted_connection = listener.accept() => accepted_connection?,
+            Err(oneshot::error::RecvError { .. }) = &mut serving_stopped => return Ok(()),
+        };
 
-        serve_connection(stream, &response, &observed_requests).await?;
+        select! {
+            biased;
+            served_connection = serve_connection(stream, &response, &observed_requests) => {
+                served_connection?;
+            }
+            Err(oneshot::error::RecvError { .. }) = &mut serving_stopped => return Ok(()),
+        }
     }
 }
 
@@ -27,23 +44,33 @@ pub struct LocalHttpFixture {
     pub addr: SocketAddr,
     observed_requests: Arc<ObservedRequests>,
     serving_task: JoinHandle<io::Result<()>>,
+    stop_serving: oneshot::Sender<()>,
 }
 
 impl LocalHttpFixture {
     pub async fn start(response: FixtureResponse) -> io::Result<Self> {
-        let listener = TcpListener::bind(SocketAddr::from((Ipv4Addr::LOCALHOST, 0))).await?;
-        let addr = listener.local_addr()?;
-        let observed_requests = Arc::new(ObservedRequests::new());
-        let serving_task = spawn(serve_sequentially(
-            listener,
-            response,
-            observed_requests.clone(),
-        ));
-
-        Ok(Self {
-            addr,
-            observed_requests,
-            serving_task,
+        TcpListener::bind(SocketAddr::from((Ipv4Addr::LOCALHOST, 0)))
+            .await
+            .and_then(|listener| Self::serving_on(listener, response))
+    }
+
+    fn serving_on(listener: TcpListener, response: FixtureResponse) -> io::Result<Self> {
+        listener.local_addr().map(|addr| {
+            let observed_requests = Arc::new(ObservedRequests::new());
+            let (stop_serving, serving_stopped) = oneshot::channel();
+            let serving_task = spawn(serve_sequentially_until_stopped(
+                listener,
+                response,
+                observed_requests.clone(),
+                serving_stopped,
+            ));
+
+            Self {
+                addr,
+                observed_requests,
+                serving_task,
+                stop_serving,
+            }
         })
     }
 
@@ -61,10 +88,51 @@ impl LocalHttpFixture {
     pub fn last_range_header(&self) -> Option<Vec<u8>> {
         self.observed_requests.last_range_header()
     }
+
+    pub async fn wait_for_requests(&self, expected_request_count: u32) -> Result<(), RecvError> {
+        self.observed_requests
+            .wait_for_count(expected_request_count)
+            .await
+    }
+
+    #[must_use]
+    pub fn abandoned_trickle_departures(&self) -> Vec<ErrorKind> {
+        self.observed_requests.abandoned_trickle_departures()
+    }
+
+    pub async fn wait_for_abandoned_trickles(
+        &self,
+        expected_abandoned_trickle_count: usize,
+    ) -> Result<(), RecvError> {
+        self.observed_requests
+            .wait_for_abandoned_trickles(expected_abandoned_trickle_count)
+            .await
+    }
 }
 
-impl Drop for LocalHttpFixture {
-    fn drop(&mut self) {
-        self.serving_task.abort();
+impl LocalHttpFixture {
+    pub async fn stop(self) -> Result<io::Result<()>, JoinError> {
+        let Self {
+            serving_task,
+            stop_serving,
+            ..
+        } = self;
+
+        drop(stop_serving);
+
+        serving_task.await
+    }
+
+    pub async fn wait_until_serving_fails(self) -> Result<io::Result<()>, JoinError> {
+        let Self {
+            serving_task,
+            stop_serving,
+            ..
+        } = self;
+        let serving_outcome = serving_task.await;
+
+        drop(stop_serving);
+
+        serving_outcome
     }
 }
diff --git a/paddler_local_http_fixture/src/observed_requests.rs b/paddler_local_http_fixture/src/observed_requests.rs
index 1b7ba2392..dd00134b1 100644
--- a/paddler_local_http_fixture/src/observed_requests.rs
+++ b/paddler_local_http_fixture/src/observed_requests.rs
@@ -1,31 +1,69 @@
-use std::sync::atomic::AtomicU32;
-use std::sync::atomic::Ordering;
+use std::io;
+use std::io::ErrorKind;
 
 use tokio::sync::watch;
+use tokio::sync::watch::error::RecvError;
+
+async fn wait_until_reached(
+    counter: &watch::Sender<u32>,
+    expected_count: u32,
+) -> Result<(), RecvError> {
+    counter
+        .subscribe()
+        .wait_for(|count| *count >= expected_count)
+        .await
+        .map(drop)
+}
 
 pub struct ObservedRequests {
-    count: AtomicU32,
+    abandoned_trickle_departures: watch::Sender<Vec<ErrorKind>>,
+    count: watch::Sender<u32>,
     last_range_header: watch::Sender<Option<Vec<u8>>>,
 }
 
 impl ObservedRequests {
     pub fn new() -> Self {
         Self {
-            count: AtomicU32::new(0),
+            abandoned_trickle_departures: watch::Sender::new(Vec::new()),
+            count: watch::Sender::new(0),
             last_range_header: watch::Sender::new(None),
         }
     }
 
     pub fn record(&self, range_header: Option<Vec<u8>>) {
         self.last_range_header.send_replace(range_header);
-        self.count.fetch_add(1, Ordering::SeqCst);
+        self.count.send_modify(|count| *count += 1);
+    }
+
+    pub fn record_abandoned_trickle(&self, client_departure: &io::Error) {
+        self.abandoned_trickle_departures
+            .send_modify(|departures| departures.push(client_departure.kind()));
+    }
+
+    pub fn abandoned_trickle_departures(&self) -> Vec<ErrorKind> {
+        self.abandoned_trickle_departures.borrow().clone()
     }
 
     pub fn count(&self) -> u32 {
-        self.count.load(Ordering::SeqCst)
+        *self.count.borrow()
     }
 
     pub fn last_range_header(&self) -> Option<Vec<u8>> {
         self.last_range_header.borrow().clone()
     }
+
+    pub async fn wait_for_count(&self, expected_count: u32) -> Result<(), RecvError> {
+        wait_until_reached(&self.count, expected_count).await
+    }
+
+    pub async fn wait_for_abandoned_trickles(
+        &self,
+        expected_count: usize,
+    ) -> Result<(), RecvError> {
+        self.abandoned_trickle_departures
+            .subscribe()
+            .wait_for(|departures| departures.len() >= expected_count)
+            .await
+            .map(drop)
+    }
 }
diff --git a/paddler_local_http_fixture/src/serve_connection.rs b/paddler_local_http_fixture/src/serve_connection.rs
index 2152b10a4..606722295 100644
--- a/paddler_local_http_fixture/src/serve_connection.rs
+++ b/paddler_local_http_fixture/src/serve_connection.rs
@@ -1,14 +1,23 @@
 use std::future::pending;
 use std::io;
+use std::time::Duration;
 
 use http::StatusCode;
 use tokio::io::AsyncWriteExt as _;
 use tokio::net::TcpStream;
+use tokio::time::interval;
 
 use crate::fixture_response::FixtureResponse;
 use crate::observed_requests::ObservedRequests;
 use crate::read_range_header::read_range_header;
 
+fn hugging_face_file_headers(file_length: usize, commit_hash: &str, etag: &str) -> String {
+    format!(
+        "Content-Range: bytes 0-{}/{file_length}\r\nETag: \"{etag}\"\r\nX-Repo-Commit: {commit_hash}\r\n",
+        file_length - 1
+    )
+}
+
 fn status_line_and_headers(
     status: StatusCode,
     content_length: usize,
@@ -19,24 +28,54 @@ fn status_line_and_headers(
     )
 }
 
+const ENDLESS_CHUNKED_RESPONSE_HEAD: &str =
+    "HTTP/1.1 200 OK\r\nTransfer-Encoding: chunked\r\nConnection: close\r\n\r\n";
+const ONE_BYTE_CHUNK: &[u8] = b"1\r\n\0\r\n";
+
+async fn trickle_until_the_client_stops_receiving(
+    stream: &mut TcpStream,
+    byte_interval: Duration,
+) -> io::Error {
+    let mut trickle = interval(byte_interval);
+    let mut next_trickled_bytes = ENDLESS_CHUNKED_RESPONSE_HEAD.as_bytes();
+
+    loop {
+        trickle.tick().await;
+
+        if let Err(client_departure) = stream.write_all(next_trickled_bytes).await {
+            return client_departure;
+        }
+
+        next_trickled_bytes = ONE_BYTE_CHUNK;
+    }
+}
+
 async fn write_complete_body(stream: &mut TcpStream, head: &str, body: &[u8]) -> io::Result<()> {
-    stream.write_all(head.as_bytes()).await?;
-    stream.write_all(body).await?;
+    stream.write_all(&[head.as_bytes(), body].concat()).await?;
     stream.shutdown().await
 }
 
 async fn write_body_prefix(
     stream: &mut TcpStream,
     status: StatusCode,
+    extra_headers: &str,
     sent_body: &[u8],
     withheld_byte_count: usize,
 ) -> io::Result<()> {
     stream
         .write_all(
-            status_line_and_headers(status, sent_body.len() + withheld_byte_count, "").as_bytes(),
+            &[
+                status_line_and_headers(
+                    status,
+                    sent_body.len() + withheld_byte_count,
+                    extra_headers,
+                )
+                .as_bytes(),
+                sent_body,
+            ]
+            .concat(),
         )
-        .await?;
-    stream.write_all(sent_body).await
+        .await
 }
 
 pub async fn serve_connection(
@@ -47,7 +86,29 @@ pub async fn serve_connection(
     observed_requests.record(read_range_header(&mut stream).await?);
 
     match response {
-        FixtureResponse::CloseBeforeHeaders => stream.shutdown().await,
+        FixtureResponse::EndlessTrickle { byte_interval } => {
+            observed_requests.record_abandoned_trickle(
+                &trickle_until_the_client_stops_receiving(&mut stream, *byte_interval).await,
+            );
+
+            Ok(())
+        }
+        FixtureResponse::HuggingFaceFile {
+            body,
+            commit_hash,
+            etag,
+        } => {
+            write_complete_body(
+                &mut stream,
+                &status_line_and_headers(
+                    StatusCode::PARTIAL_CONTENT,
+                    body.len(),
+                    &hugging_face_file_headers(body.len(), commit_hash, etag),
+                ),
+                body,
+            )
+            .await
+        }
         FixtureResponse::Ok(body) => {
             write_complete_body(
                 &mut stream,
@@ -76,8 +137,14 @@ pub async fn serve_connection(
             sent_body,
             withheld_byte_count,
         } => {
-            write_body_prefix(&mut stream, StatusCode::OK, sent_body, *withheld_byte_count).await?;
-            stream.flush().await?;
+            write_body_prefix(
+                &mut stream,
+                StatusCode::OK,
+                "",
+                sent_body,
+                *withheld_byte_count,
+            )
+            .await?;
 
             pending().await
         }
@@ -88,6 +155,25 @@ pub async fn serve_connection(
             sent_body,
             status,
             withheld_byte_count,
-        } => write_body_prefix(&mut stream, *status, sent_body, *withheld_byte_count).await,
+        } => write_body_prefix(&mut stream, *status, "", sent_body, *withheld_byte_count).await,
+        FixtureResponse::TruncatedHuggingFaceFile {
+            commit_hash,
+            etag,
+            sent_body,
+            withheld_byte_count,
+        } => {
+            write_body_prefix(
+                &mut stream,
+                StatusCode::PARTIAL_CONTENT,
+                &hugging_face_file_headers(
+                    sent_body.len() + withheld_byte_count,
+                    commit_hash,
+                    etag,
+                ),
+                sent_body,
+                *withheld_byte_count,
+            )
+            .await
+        }
     }
 }
diff --git a/paddler_local_http_fixture/tests/a_client_leaving_before_its_request_ends_serving_with_an_unexpected_eof.rs b/paddler_local_http_fixture/tests/a_client_leaving_before_its_request_ends_serving_with_an_unexpected_eof.rs
new file mode 100644
index 000000000..625bdf8fc
--- /dev/null
+++ b/paddler_local_http_fixture/tests/a_client_leaving_before_its_request_ends_serving_with_an_unexpected_eof.rs
@@ -0,0 +1,31 @@
+use std::io;
+use std::io::ErrorKind;
+
+use tokio::net::TcpStream;
+
+use paddler_local_http_fixture::fixture_response::FixtureResponse;
+use paddler_local_http_fixture::local_http_fixture::LocalHttpFixture;
+
+#[tokio::test]
+async fn a_client_leaving_before_its_request_ends_serving_with_an_unexpected_eof() {
+    let fixture = LocalHttpFixture::start(FixtureResponse::Ok(Vec::new()))
+        .await
+        .expect("the local HTTP fixture must start");
+
+    drop(
+        TcpStream::connect(fixture.addr)
+            .await
+            .expect("the client must connect to the fixture"),
+    );
+
+    assert_eq!(
+        fixture
+            .wait_until_serving_fails()
+            .await
+            .expect("the fixture task must not panic")
+            .err()
+            .as_ref()
+            .map(io::Error::kind),
+        Some(ErrorKind::UnexpectedEof)
+    );
+}
diff --git a/paddler_local_http_fixture/tests/a_client_leaving_with_unread_trickled_bytes_abandons_the_trickle.rs b/paddler_local_http_fixture/tests/a_client_leaving_with_unread_trickled_bytes_abandons_the_trickle.rs
new file mode 100644
index 000000000..962f34bd8
--- /dev/null
+++ b/paddler_local_http_fixture/tests/a_client_leaving_with_unread_trickled_bytes_abandons_the_trickle.rs
@@ -0,0 +1,48 @@
+use std::io::ErrorKind;
+use std::time::Duration;
+
+use tokio::io::AsyncWriteExt as _;
+use tokio::net::TcpStream;
+
+use paddler_local_http_fixture::fixture_response::FixtureResponse;
+use paddler_local_http_fixture::local_http_fixture::LocalHttpFixture;
+
+const BYTE_INTERVAL_AFTER_THE_IMMEDIATE_FIRST_BYTE: Duration = Duration::from_secs(1);
+
+#[tokio::test]
+async fn a_client_leaving_with_unread_trickled_bytes_abandons_the_trickle() {
+    let fixture = LocalHttpFixture::start(FixtureResponse::EndlessTrickle {
+        byte_interval: BYTE_INTERVAL_AFTER_THE_IMMEDIATE_FIRST_BYTE,
+    })
+    .await
+    .expect("the local HTTP fixture must start");
+    let mut client = TcpStream::connect(fixture.addr)
+        .await
+        .expect("the client must connect to the fixture");
+
+    client
+        .write_all(b"GET /endless.gguf HTTP/1.1\r\nHost: fixture\r\n\r\n")
+        .await
+        .expect("the client must send its request");
+    client
+        .peek(&mut [0; 1])
+        .await
+        .expect("the fixture must start trickling the response");
+    drop(client);
+
+    fixture
+        .wait_for_abandoned_trickles(1)
+        .await
+        .expect("the fixture must record the client leaving the trickle");
+
+    assert_eq!(
+        fixture.abandoned_trickle_departures(),
+        vec![ErrorKind::ConnectionReset]
+    );
+
+    fixture
+        .stop()
+        .await
+        .expect("the fixture task must not panic")
+        .expect("the fixture must serve every request");
+}
diff --git a/paddler_local_http_fixture/tests/a_client_resetting_during_a_stalled_body_ends_serving_with_a_connection_reset.rs b/paddler_local_http_fixture/tests/a_client_resetting_during_a_stalled_body_ends_serving_with_a_connection_reset.rs
new file mode 100644
index 000000000..54402df85
--- /dev/null
+++ b/paddler_local_http_fixture/tests/a_client_resetting_during_a_stalled_body_ends_serving_with_a_connection_reset.rs
@@ -0,0 +1,31 @@
+use std::io;
+use std::io::ErrorKind;
+
+use paddler_local_http_fixture::fixture_response::FixtureResponse;
+use paddler_local_http_fixture::local_http_fixture::LocalHttpFixture;
+
+use crate::resetting_client::ResettingClient;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn a_client_resetting_during_a_stalled_body_ends_serving_with_a_connection_reset() {
+    let client = ResettingClient::with_the_smallest_receive_buffer();
+    let fixture = LocalHttpFixture::start(FixtureResponse::StalledBody {
+        sent_body: client.body_outlasting_both_socket_buffers(),
+        withheld_byte_count: 1,
+    })
+    .await
+    .expect("the local HTTP fixture must start");
+
+    client.request_then_reset(&fixture).await;
+
+    assert_eq!(
+        fixture
+            .wait_until_serving_fails()
+            .await
+            .expect("the fixture task must not panic")
+            .err()
+            .as_ref()
+            .map(io::Error::kind),
+        Some(ErrorKind::ConnectionReset)
+    );
+}
diff --git a/paddler_local_http_fixture/tests/a_client_resetting_during_the_body_ends_serving_with_a_connection_reset.rs b/paddler_local_http_fixture/tests/a_client_resetting_during_the_body_ends_serving_with_a_connection_reset.rs
new file mode 100644
index 000000000..413cd8991
--- /dev/null
+++ b/paddler_local_http_fixture/tests/a_client_resetting_during_the_body_ends_serving_with_a_connection_reset.rs
@@ -0,0 +1,30 @@
+use std::io;
+use std::io::ErrorKind;
+
+use paddler_local_http_fixture::fixture_response::FixtureResponse;
+use paddler_local_http_fixture::local_http_fixture::LocalHttpFixture;
+
+use crate::resetting_client::ResettingClient;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn a_client_resetting_during_the_body_ends_serving_with_a_connection_reset() {
+    let client = ResettingClient::with_the_smallest_receive_buffer();
+    let fixture = LocalHttpFixture::start(FixtureResponse::Ok(
+        client.body_outlasting_both_socket_buffers(),
+    ))
+    .await
+    .expect("the local HTTP fixture must start");
+
+    client.request_then_reset(&fixture).await;
+
+    assert_eq!(
+        fixture
+            .wait_until_serving_fails()
+            .await
+            .expect("the fixture task must not panic")
+            .err()
+            .as_ref()
+            .map(io::Error::kind),
+        Some(ErrorKind::ConnectionReset)
+    );
+}
diff --git a/paddler_local_http_fixture/tests/a_client_resetting_its_connection_ends_serving_with_a_connection_reset.rs b/paddler_local_http_fixture/tests/a_client_resetting_its_connection_ends_serving_with_a_connection_reset.rs
new file mode 100644
index 000000000..3f520673b
--- /dev/null
+++ b/paddler_local_http_fixture/tests/a_client_resetting_its_connection_ends_serving_with_a_connection_reset.rs
@@ -0,0 +1,33 @@
+use std::io;
+use std::io::ErrorKind;
+
+use tokio::net::TcpStream;
+
+use paddler_local_http_fixture::fixture_response::FixtureResponse;
+use paddler_local_http_fixture::local_http_fixture::LocalHttpFixture;
+
+#[tokio::test]
+async fn a_client_resetting_its_connection_ends_serving_with_a_connection_reset() {
+    let fixture = LocalHttpFixture::start(FixtureResponse::Ok(Vec::new()))
+        .await
+        .expect("the local HTTP fixture must start");
+    let client = TcpStream::connect(fixture.addr)
+        .await
+        .expect("the client must connect to the fixture");
+
+    client
+        .set_zero_linger()
+        .expect("the client must reset its connection when it closes");
+    drop(client);
+
+    assert_eq!(
+        fixture
+            .wait_until_serving_fails()
+            .await
+            .expect("the fixture task must not panic")
+            .err()
+            .as_ref()
+            .map(io::Error::kind),
+        Some(ErrorKind::ConnectionReset)
+    );
+}
diff --git a/paddler_local_http_fixture/tests/a_connection_the_fixture_cannot_accept_ends_serving.rs b/paddler_local_http_fixture/tests/a_connection_the_fixture_cannot_accept_ends_serving.rs
new file mode 100644
index 000000000..5e6a81302
--- /dev/null
+++ b/paddler_local_http_fixture/tests/a_connection_the_fixture_cannot_accept_ends_serving.rs
@@ -0,0 +1,62 @@
+use std::fs::File;
+use std::io;
+use std::net::Ipv4Addr;
+use std::net::SocketAddrV4;
+use std::os::fd::AsRawFd as _;
+
+use nix::errno::Errno;
+use nix::sys::resource::Resource;
+use nix::sys::resource::getrlimit;
+use nix::sys::resource::setrlimit;
+use nix::sys::socket::AddressFamily;
+use nix::sys::socket::SockFlag;
+use nix::sys::socket::SockType;
+use nix::sys::socket::SockaddrIn;
+use nix::sys::socket::connect;
+use nix::sys::socket::socket;
+
+use paddler_local_http_fixture::fixture_response::FixtureResponse;
+use paddler_local_http_fixture::local_http_fixture::LocalHttpFixture;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn a_connection_the_fixture_cannot_accept_ends_serving() {
+    let fixture = LocalHttpFixture::start(FixtureResponse::Ok(Vec::new()))
+        .await
+        .expect("the local HTTP fixture must start");
+    let client_socket = socket(
+        AddressFamily::Inet,
+        SockType::Stream,
+        SockFlag::empty(),
+        None,
+    )
+    .expect("the client socket must open");
+    let lowest_free_descriptor = File::open("/dev/null")
+        .expect("a probe descriptor must open")
+        .as_raw_fd();
+    let (_soft_descriptor_limit, hard_descriptor_limit) =
+        getrlimit(Resource::RLIMIT_NOFILE).expect("the descriptor limit must be readable");
+
+    setrlimit(
+        Resource::RLIMIT_NOFILE,
+        u64::try_from(lowest_free_descriptor).expect("descriptors are never negative"),
+        hard_descriptor_limit,
+    )
+    .expect("the descriptor limit must be lowered");
+    connect(
+        client_socket.as_raw_fd(),
+        &SockaddrIn::from(SocketAddrV4::new(Ipv4Addr::LOCALHOST, fixture.addr.port())),
+    )
+    .expect("the kernel must complete the handshake without a new descriptor");
+
+    assert_eq!(
+        fixture
+            .wait_until_serving_fails()
+            .await
+            .expect("the fixture task must not panic")
+            .err()
+            .as_ref()
+            .and_then(io::Error::raw_os_error)
+            .map(Errno::from_raw),
+        Some(Errno::EMFILE)
+    );
+}
diff --git a/paddler_local_http_fixture/tests/a_fixture_that_cannot_open_its_listener_reports_the_descriptor_limit.rs b/paddler_local_http_fixture/tests/a_fixture_that_cannot_open_its_listener_reports_the_descriptor_limit.rs
new file mode 100644
index 000000000..5141ab797
--- /dev/null
+++ b/paddler_local_http_fixture/tests/a_fixture_that_cannot_open_its_listener_reports_the_descriptor_limit.rs
@@ -0,0 +1,37 @@
+use std::fs::File;
+use std::io;
+use std::os::fd::AsRawFd as _;
+
+use nix::errno::Errno;
+use nix::sys::resource::Resource;
+use nix::sys::resource::getrlimit;
+use nix::sys::resource::setrlimit;
+
+use paddler_local_http_fixture::fixture_response::FixtureResponse;
+use paddler_local_http_fixture::local_http_fixture::LocalHttpFixture;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn a_fixture_that_cannot_open_its_listener_reports_the_descriptor_limit() {
+    let lowest_free_descriptor = File::open("/dev/null")
+        .expect("a probe descriptor must open")
+        .as_raw_fd();
+    let (_soft_descriptor_limit, hard_descriptor_limit) =
+        getrlimit(Resource::RLIMIT_NOFILE).expect("the descriptor limit must be readable");
+
+    setrlimit(
+        Resource::RLIMIT_NOFILE,
+        u64::try_from(lowest_free_descriptor).expect("descriptors are never negative"),
+        hard_descriptor_limit,
+    )
+    .expect("the descriptor limit must be lowered");
+
+    assert_eq!(
+        LocalHttpFixture::start(FixtureResponse::Ok(Vec::new()))
+            .await
+            .err()
+            .as_ref()
+            .and_then(io::Error::raw_os_error)
+            .map(Errno::from_raw),
+        Some(Errno::EMFILE)
+    );
+}
diff --git a/paddler_local_http_fixture/tests/a_malformed_request_ends_serving_with_invalid_data.rs b/paddler_local_http_fixture/tests/a_malformed_request_ends_serving_with_invalid_data.rs
new file mode 100644
index 000000000..fcdd027bc
--- /dev/null
+++ b/paddler_local_http_fixture/tests/a_malformed_request_ends_serving_with_invalid_data.rs
@@ -0,0 +1,42 @@
+use std::io;
+use std::io::ErrorKind;
+
+use httparse::Error as HttpParseError;
+use tokio::io::AsyncWriteExt as _;
+use tokio::net::TcpStream;
+
+use paddler_local_http_fixture::fixture_response::FixtureResponse;
+use paddler_local_http_fixture::local_http_fixture::LocalHttpFixture;
+
+#[tokio::test]
+async fn a_malformed_request_ends_serving_with_invalid_data() {
+    let fixture = LocalHttpFixture::start(FixtureResponse::Ok(Vec::new()))
+        .await
+        .expect("the local HTTP fixture must start");
+    let mut client = TcpStream::connect(fixture.addr)
+        .await
+        .expect("the client must connect to the fixture");
+
+    client
+        .write_all(b"not an http request\r\n\r\n")
+        .await
+        .expect("the client must send its malformed request");
+
+    let serving_error = fixture
+        .wait_until_serving_fails()
+        .await
+        .expect("the fixture task must not panic")
+        .err();
+
+    assert_eq!(
+        serving_error.as_ref().map(io::Error::kind),
+        Some(ErrorKind::InvalidData)
+    );
+    assert_eq!(
+        serving_error
+            .as_ref()
+            .and_then(io::Error::get_ref)
+            .and_then(|parse_error| parse_error.downcast_ref::<HttpParseError>()),
+        Some(&HttpParseError::Version)
+    );
+}
diff --git a/paddler_local_http_fixture/tests/main.rs b/paddler_local_http_fixture/tests/main.rs
new file mode 100644
index 000000000..9a98882ca
--- /dev/null
+++ b/paddler_local_http_fixture/tests/main.rs
@@ -0,0 +1,10 @@
+mod a_client_leaving_before_its_request_ends_serving_with_an_unexpected_eof;
+mod a_client_leaving_with_unread_trickled_bytes_abandons_the_trickle;
+mod a_client_resetting_during_a_stalled_body_ends_serving_with_a_connection_reset;
+mod a_client_resetting_during_the_body_ends_serving_with_a_connection_reset;
+mod a_client_resetting_its_connection_ends_serving_with_a_connection_reset;
+mod a_connection_the_fixture_cannot_accept_ends_serving;
+mod a_fixture_that_cannot_open_its_listener_reports_the_descriptor_limit;
+mod a_malformed_request_ends_serving_with_invalid_data;
+#[cfg(test)]
+mod resetting_client;
diff --git a/paddler_local_http_fixture/tests/resetting_client.rs b/paddler_local_http_fixture/tests/resetting_client.rs
new file mode 100644
index 000000000..06fd1a28e
--- /dev/null
+++ b/paddler_local_http_fixture/tests/resetting_client.rs
@@ -0,0 +1,66 @@
+use std::net::Ipv4Addr;
+use std::net::SocketAddr;
+
+use tokio::io::AsyncWriteExt as _;
+use tokio::net::TcpSocket;
+
+use paddler_local_http_fixture::local_http_fixture::LocalHttpFixture;
+use paddler_tcp_buffer_limits::tcp_buffer_limits_file::TcpBufferLimitsFile;
+
+const SMALLEST_RECEIVE_BUFFER_REQUEST: u32 = 1;
+
+pub struct ResettingClient {
+    receive_buffer: usize,
+    socket: TcpSocket,
+}
+
+impl ResettingClient {
+    pub fn with_the_smallest_receive_buffer() -> Self {
+        let socket = TcpSocket::new_v4().expect("the client socket must open");
+
+        socket
+            .set_recv_buffer_size(SMALLEST_RECEIVE_BUFFER_REQUEST)
+            .expect("the client must shrink its receive buffer");
+
+        Self {
+            receive_buffer: usize::try_from(
+                socket
+                    .recv_buffer_size()
+                    .expect("the client must read back its receive buffer"),
+            )
+            .expect("the receive buffer size must fit in memory"),
+            socket,
+        }
+    }
+
+    pub fn body_outlasting_both_socket_buffers(&self) -> Vec<u8> {
+        vec![
+            0;
+            TcpBufferLimitsFile::of_kernel_send_buffers()
+                .largest_buffer()
+                .expect("the kernel must report its TCP send buffer limits")
+                + self.receive_buffer
+                + 1
+        ]
+    }
+
+    pub async fn request_then_reset(self, fixture: &LocalHttpFixture) {
+        let mut client = self
+            .socket
+            .connect(SocketAddr::from((Ipv4Addr::LOCALHOST, fixture.addr.port())))
+            .await
+            .expect("the client must connect to the fixture");
+
+        client
+            .write_all(b"GET / HTTP/1.1\r\nHost: fixture\r\n\r\n")
+            .await
+            .expect("the client must send its request");
+        fixture
+            .wait_for_requests(1)
+            .await
+            .expect("the fixture must read the request before the client resets");
+        client
+            .set_zero_linger()
+            .expect("the client must reset its connection when it closes");
+    }
+}
diff --git a/paddler_messaging/Cargo.toml b/paddler_messaging/Cargo.toml
index 4254a29b7..428f652f9 100644
--- a/paddler_messaging/Cargo.toml
+++ b/paddler_messaging/Cargo.toml
@@ -9,11 +9,7 @@ repository.workspace = true
 version.workspace = true
 
 [dependencies]
-base64 = { workspace = true }
-encoding_rs = { workspace = true }
 llama-cpp-bindings-types = { workspace = true }
-log = { workspace = true }
-nanoid = { workspace = true }
 paddler_inference_parameters = { workspace = true }
 serde = { workspace = true }
 serde_json = { workspace = true }
diff --git a/paddler_messaging/src/agent_decisions_serving.rs b/paddler_messaging/src/agent_decisions_serving.rs
new file mode 100644
index 000000000..147c91ffd
--- /dev/null
+++ b/paddler_messaging/src/agent_decisions_serving.rs
@@ -0,0 +1,47 @@
+use serde::Deserialize;
+use serde::Serialize;
+
+use crate::inference_capability::InferenceCapability;
+
+#[derive(Clone, Copy, Debug, Deserialize, Eq, PartialEq, Serialize)]
+#[serde(deny_unknown_fields)]
+pub enum AgentDecisionsServing {
+    NotServed,
+    Served,
+}
+
+impl AgentDecisionsServing {
+    #[must_use]
+    pub const fn serves(&self, inference_capability: InferenceCapability) -> bool {
+        match inference_capability {
+            InferenceCapability::Decisions => match self {
+                Self::NotServed => false,
+                Self::Served => true,
+            },
+            InferenceCapability::Embeddings | InferenceCapability::TextGeneration => true,
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::AgentDecisionsServing;
+    use crate::inference_capability::InferenceCapability;
+
+    #[test]
+    fn only_an_agent_serving_decisions_serves_decisions() {
+        assert!(!AgentDecisionsServing::NotServed.serves(InferenceCapability::Decisions));
+        assert!(AgentDecisionsServing::Served.serves(InferenceCapability::Decisions));
+    }
+
+    #[test]
+    fn every_serving_agent_serves_text_generation_and_embeddings() {
+        for agent_decisions_serving in [
+            AgentDecisionsServing::NotServed,
+            AgentDecisionsServing::Served,
+        ] {
+            assert!(agent_decisions_serving.serves(InferenceCapability::Embeddings));
+            assert!(agent_decisions_serving.serves(InferenceCapability::TextGeneration));
+        }
+    }
+}
diff --git a/paddler_messaging/src/agent_decisions_settings.rs b/paddler_messaging/src/agent_decisions_settings.rs
new file mode 100644
index 000000000..289d19925
--- /dev/null
+++ b/paddler_messaging/src/agent_decisions_settings.rs
@@ -0,0 +1,11 @@
+use serde::Deserialize;
+use serde::Serialize;
+
+use crate::served_decision_settings::ServedDecisionSettings;
+
+#[derive(Clone, Debug, Deserialize, Eq, PartialEq, Serialize)]
+#[serde(deny_unknown_fields)]
+pub enum AgentDecisionsSettings {
+    NotServed,
+    Served(ServedDecisionSettings),
+}
diff --git a/paddler_messaging/src/agent_desired_model.rs b/paddler_messaging/src/agent_desired_model.rs
index 02e6ea73d..fe7b0654c 100644
--- a/paddler_messaging/src/agent_desired_model.rs
+++ b/paddler_messaging/src/agent_desired_model.rs
@@ -1,15 +1,10 @@
 use serde::Deserialize;
 use serde::Serialize;
 
-use crate::huggingface_model_reference::HuggingFaceModelReference;
-use crate::url_model_reference::UrlModelReference;
-
 #[derive(Clone, Debug, Default, Deserialize, Eq, PartialEq, Serialize)]
 #[serde(deny_unknown_fields)]
 pub enum AgentDesiredModel {
-    HuggingFace(HuggingFaceModelReference),
-    LocalToAgent(String),
-    Url(UrlModelReference),
     #[default]
     None,
+    Uri(String),
 }
diff --git a/paddler_messaging/src/agent_desired_state.rs b/paddler_messaging/src/agent_desired_state.rs
index e6b577151..4f258d4e0 100644
--- a/paddler_messaging/src/agent_desired_state.rs
+++ b/paddler_messaging/src/agent_desired_state.rs
@@ -1,16 +1,76 @@
 use serde::Deserialize;
 use serde::Serialize;
 
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
+use paddler_inference_parameters::embedding_parameters::EmbeddingParameters;
+use paddler_inference_parameters::model_runtime_parameters::ModelRuntimeParameters;
+use paddler_inference_parameters::pooling_type::PoolingType;
 
+use crate::agent_decisions_settings::AgentDecisionsSettings;
 use crate::agent_desired_model::AgentDesiredModel;
-use crate::chat_template::ChatTemplate;
+use crate::balancer_desired_state::BalancerDesiredState;
+use crate::balancer_text_generation_settings::BalancerTextGenerationSettings;
 
-#[derive(Clone, Debug, Default, Deserialize, PartialEq, Serialize)]
+#[derive(Clone, Debug, Deserialize, PartialEq, Serialize)]
 #[serde(deny_unknown_fields)]
 pub struct AgentDesiredState {
-    pub chat_template_override: Option<ChatTemplate>,
-    pub inference_parameters: InferenceParameters,
+    pub decisions: AgentDecisionsSettings,
     pub model: AgentDesiredModel,
-    pub multimodal_projection: AgentDesiredModel,
+    pub model_runtime_parameters: ModelRuntimeParameters,
+    pub pooling_type: PoolingType,
+    pub text_generation: BalancerTextGenerationSettings,
+}
+
+impl Default for AgentDesiredState {
+    fn default() -> Self {
+        Self::from(&BalancerDesiredState::default())
+    }
+}
+
+impl From<&BalancerDesiredState> for AgentDesiredState {
+    fn from(
+        BalancerDesiredState {
+            decision,
+            embeddings: EmbeddingParameters { pooling_type, .. },
+            model,
+            model_runtime_parameters,
+            text_generation,
+        }: &BalancerDesiredState,
+    ) -> Self {
+        Self {
+            decisions: decision.agent_decisions(),
+            model: model.clone(),
+            model_runtime_parameters: model_runtime_parameters.clone(),
+            pooling_type: pooling_type.clone(),
+            text_generation: text_generation.clone(),
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::num::NonZeroUsize;
+
+    use paddler_inference_parameters::embedding_parameters::EmbeddingParameters;
+    use paddler_inference_parameters::pooling_type::PoolingType;
+
+    use super::AgentDesiredState;
+    use crate::balancer_desired_state::BalancerDesiredState;
+
+    #[test]
+    fn the_balancer_only_embedding_batch_size_does_not_reach_agents() {
+        let desired_state_batching = |embedding_batch_size| {
+            AgentDesiredState::from(&BalancerDesiredState {
+                embeddings: EmbeddingParameters {
+                    embedding_batch_size,
+                    pooling_type: PoolingType::Mean,
+                },
+                ..BalancerDesiredState::default()
+            })
+        };
+
+        assert_eq!(
+            desired_state_batching(NonZeroUsize::MIN),
+            desired_state_batching(NonZeroUsize::MAX)
+        );
+    }
 }
diff --git a/paddler_messaging/src/agent_issue.rs b/paddler_messaging/src/agent_issue.rs
index bd59193b8..314adb0c2 100644
--- a/paddler_messaging/src/agent_issue.rs
+++ b/paddler_messaging/src/agent_issue.rs
@@ -1,14 +1,20 @@
 use serde::Deserialize;
 use serde::Serialize;
 
+use crate::agent_issue_params::agent_model_uri_not_canonical_params::AgentModelUriNotCanonicalParams;
 use crate::agent_issue_params::chat_template_does_not_compile_params::ChatTemplateDoesNotCompileParams;
+use crate::agent_issue_params::hugging_face_download_failed_params::HuggingFaceDownloadFailedParams;
 use crate::agent_issue_params::hugging_face_download_lock::HuggingFaceDownloadLock;
 use crate::agent_issue_params::model_path::ModelPath;
+use crate::agent_issue_params::pointer_head_incompatible_with_model_params::PointerHeadIncompatibleWithModelParams;
 use crate::agent_issue_params::slot_cannot_start_params::SlotCannotStartParams;
+use crate::agent_issue_params::slots_insufficient_for_decisions_params::SlotsInsufficientForDecisionsParams;
 
 #[derive(Clone, Debug, Deserialize, Eq, Hash, Ord, PartialEq, PartialOrd, Serialize)]
 #[serde(deny_unknown_fields)]
 pub enum AgentIssue {
+    AgentModelUriIsNotAnAbsoluteLocalPath(ModelPath),
+    AgentModelUriIsNotCanonical(AgentModelUriNotCanonicalParams),
     CacheCannotAcquireLock(ModelPath),
     CacheDirectoryIsNotWritable(ModelPath),
     CacheStorageIsFull(ModelPath),
@@ -20,13 +26,20 @@ pub enum AgentIssue {
     DownloadServerRejectedRequest(ModelPath),
     DownloadUrlIsMalformed(ModelPath),
     HuggingFaceCannotAcquireLock(HuggingFaceDownloadLock),
+    HuggingFaceDownloadFailed(HuggingFaceDownloadFailedParams),
     HuggingFaceModelDoesNotExist(ModelPath),
+    HuggingFaceModelUriIsMalformed(ModelPath),
     HuggingFacePermissions(ModelPath),
+    LocalModelFileCannotBeChecked(ModelPath),
     ModelCacheIsCorrupted(ModelPath),
     ModelCannotBeLoaded(ModelPath),
     ModelDoesNotExistAtUrl(ModelPath),
     ModelFileDoesNotExist(ModelPath),
+    ModelUriIsUnparseable(ModelPath),
     MultimodalProjectionCannotBeLoaded(ModelPath),
+    PointerHeadCannotBeLoaded(ModelPath),
+    PointerHeadIncompatibleWithModel(PointerHeadIncompatibleWithModelParams),
     SlotCannotStart(SlotCannotStartParams),
+    SlotsInsufficientForDecisions(SlotsInsufficientForDecisionsParams),
     UnableToFindChatTemplate(ModelPath),
 }
diff --git a/paddler_messaging/src/agent_issue_params/agent_model_uri_not_canonical_params.rs b/paddler_messaging/src/agent_issue_params/agent_model_uri_not_canonical_params.rs
new file mode 100644
index 000000000..b0d4ba521
--- /dev/null
+++ b/paddler_messaging/src/agent_issue_params/agent_model_uri_not_canonical_params.rs
@@ -0,0 +1,11 @@
+use serde::Deserialize;
+use serde::Serialize;
+
+use crate::agent_issue_params::model_path::ModelPath;
+
+#[derive(Clone, Debug, Deserialize, Eq, Hash, Ord, PartialEq, PartialOrd, Serialize)]
+#[serde(deny_unknown_fields)]
+pub struct AgentModelUriNotCanonicalParams {
+    pub canonical_uri: String,
+    pub model_path: ModelPath,
+}
diff --git a/paddler_messaging/src/agent_issue_params/hugging_face_download_failed_params.rs b/paddler_messaging/src/agent_issue_params/hugging_face_download_failed_params.rs
new file mode 100644
index 000000000..99bbce091
--- /dev/null
+++ b/paddler_messaging/src/agent_issue_params/hugging_face_download_failed_params.rs
@@ -0,0 +1,12 @@
+use serde::Deserialize;
+use serde::Serialize;
+
+use crate::agent_issue_params::hugging_face_download_failure::HuggingFaceDownloadFailure;
+use crate::agent_issue_params::model_path::ModelPath;
+
+#[derive(Clone, Debug, Deserialize, Eq, Hash, Ord, PartialEq, PartialOrd, Serialize)]
+#[serde(deny_unknown_fields)]
+pub struct HuggingFaceDownloadFailedParams {
+    pub failure: HuggingFaceDownloadFailure,
+    pub model_path: ModelPath,
+}
diff --git a/paddler_messaging/src/agent_issue_params/hugging_face_download_failure.rs b/paddler_messaging/src/agent_issue_params/hugging_face_download_failure.rs
new file mode 100644
index 000000000..11694b3e1
--- /dev/null
+++ b/paddler_messaging/src/agent_issue_params/hugging_face_download_failure.rs
@@ -0,0 +1,13 @@
+use serde::Deserialize;
+use serde::Serialize;
+
+#[derive(Clone, Debug, Deserialize, Eq, Hash, Ord, PartialEq, PartialOrd, Serialize)]
+#[serde(deny_unknown_fields)]
+pub enum HuggingFaceDownloadFailure {
+    CacheIoFailed,
+    HubConnectionFailed,
+    HubRespondedWithStatus { status: u16 },
+    HubResponseMalformed,
+    UnexpectedHubClientFailure,
+    XetTransferFailed,
+}
diff --git a/paddler_messaging/src/agent_issue_params/mod.rs b/paddler_messaging/src/agent_issue_params/mod.rs
index f507cd1e6..9e21829e0 100644
--- a/paddler_messaging/src/agent_issue_params/mod.rs
+++ b/paddler_messaging/src/agent_issue_params/mod.rs
@@ -1,4 +1,10 @@
+pub mod agent_model_uri_not_canonical_params;
 pub mod chat_template_does_not_compile_params;
+pub mod hugging_face_download_failed_params;
+pub mod hugging_face_download_failure;
 pub mod hugging_face_download_lock;
 pub mod model_path;
+pub mod pointer_head_incompatibility;
+pub mod pointer_head_incompatible_with_model_params;
 pub mod slot_cannot_start_params;
+pub mod slots_insufficient_for_decisions_params;
diff --git a/paddler_messaging/src/agent_issue_params/pointer_head_incompatibility.rs b/paddler_messaging/src/agent_issue_params/pointer_head_incompatibility.rs
new file mode 100644
index 000000000..0f9e01ae1
--- /dev/null
+++ b/paddler_messaging/src/agent_issue_params/pointer_head_incompatibility.rs
@@ -0,0 +1,14 @@
+use serde::Deserialize;
+use serde::Serialize;
+
+#[derive(Clone, Debug, Deserialize, Eq, Hash, Ord, PartialEq, PartialOrd, Serialize)]
+#[serde(deny_unknown_fields)]
+pub enum PointerHeadIncompatibility {
+    DelimiterIsNotAControlToken {
+        delimiter: String,
+    },
+    HiddenSizeMismatch {
+        model_hidden_size: usize,
+        pointer_head_hidden_size: usize,
+    },
+}
diff --git a/paddler_messaging/src/agent_issue_params/pointer_head_incompatible_with_model_params.rs b/paddler_messaging/src/agent_issue_params/pointer_head_incompatible_with_model_params.rs
new file mode 100644
index 000000000..e26ba4928
--- /dev/null
+++ b/paddler_messaging/src/agent_issue_params/pointer_head_incompatible_with_model_params.rs
@@ -0,0 +1,12 @@
+use serde::Deserialize;
+use serde::Serialize;
+
+use crate::agent_issue_params::model_path::ModelPath;
+use crate::agent_issue_params::pointer_head_incompatibility::PointerHeadIncompatibility;
+
+#[derive(Clone, Debug, Deserialize, Eq, Hash, Ord, PartialEq, PartialOrd, Serialize)]
+#[serde(deny_unknown_fields)]
+pub struct PointerHeadIncompatibleWithModelParams {
+    pub incompatibility: PointerHeadIncompatibility,
+    pub pointer_head_path: ModelPath,
+}
diff --git a/paddler_messaging/src/agent_issue_params/slot_cannot_start_params.rs b/paddler_messaging/src/agent_issue_params/slot_cannot_start_params.rs
index 0deee5126..ed6a4bd74 100644
--- a/paddler_messaging/src/agent_issue_params/slot_cannot_start_params.rs
+++ b/paddler_messaging/src/agent_issue_params/slot_cannot_start_params.rs
@@ -1,9 +1,12 @@
 use serde::Deserialize;
 use serde::Serialize;
 
+use crate::inference_capability::InferenceCapability;
+
 #[derive(Clone, Debug, Deserialize, Eq, Hash, Ord, PartialEq, PartialOrd, Serialize)]
 #[serde(deny_unknown_fields)]
 pub struct SlotCannotStartParams {
     pub error: String,
+    pub inference_capability: InferenceCapability,
     pub slot_index: u32,
 }
diff --git a/paddler_messaging/src/agent_issue_params/slots_insufficient_for_decisions_params.rs b/paddler_messaging/src/agent_issue_params/slots_insufficient_for_decisions_params.rs
new file mode 100644
index 000000000..b3fb3960a
--- /dev/null
+++ b/paddler_messaging/src/agent_issue_params/slots_insufficient_for_decisions_params.rs
@@ -0,0 +1,9 @@
+use serde::Deserialize;
+use serde::Serialize;
+
+#[derive(Clone, Debug, Deserialize, Eq, Hash, Ord, PartialEq, PartialOrd, Serialize)]
+#[serde(deny_unknown_fields)]
+pub struct SlotsInsufficientForDecisionsParams {
+    pub desired_slots: u16,
+    pub required_slots: u16,
+}
diff --git a/paddler_messaging/src/agent_runtime_status.rs b/paddler_messaging/src/agent_runtime_status.rs
new file mode 100644
index 000000000..ec5696032
--- /dev/null
+++ b/paddler_messaging/src/agent_runtime_status.rs
@@ -0,0 +1,93 @@
+use serde::Deserialize;
+use serde::Serialize;
+
+use crate::agent_decisions_serving::AgentDecisionsServing;
+use crate::inference_capability::InferenceCapability;
+
+#[derive(Clone, Copy, Debug, Default, Deserialize, Eq, PartialEq, Serialize)]
+#[serde(deny_unknown_fields)]
+pub enum AgentRuntimeStatus {
+    #[default]
+    Idle,
+    Serving {
+        decisions: AgentDecisionsServing,
+        slots_total: u64,
+    },
+}
+
+impl AgentRuntimeStatus {
+    #[must_use]
+    pub const fn slots_total(&self) -> u64 {
+        match self {
+            Self::Idle => 0,
+            Self::Serving { slots_total, .. } => *slots_total,
+        }
+    }
+
+    #[must_use]
+    pub const fn serves(&self, inference_capability: InferenceCapability) -> bool {
+        match self {
+            Self::Idle => false,
+            Self::Serving { decisions, .. } => decisions.serves(inference_capability),
+        }
+    }
+
+    #[must_use]
+    pub const fn slots_serving(&self, inference_capability: InferenceCapability) -> u64 {
+        if self.serves(inference_capability) {
+            self.slots_total()
+        } else {
+            0
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::AgentRuntimeStatus;
+    use crate::agent_decisions_serving::AgentDecisionsServing;
+    use crate::inference_capability::InferenceCapability;
+
+    const SERVING_WITHOUT_DECISIONS: AgentRuntimeStatus = AgentRuntimeStatus::Serving {
+        decisions: AgentDecisionsServing::NotServed,
+        slots_total: 3,
+    };
+
+    #[test]
+    fn an_idle_agent_has_no_slots() {
+        assert_eq!(AgentRuntimeStatus::Idle.slots_total(), 0);
+    }
+
+    #[test]
+    fn a_serving_agent_has_its_slots() {
+        assert_eq!(SERVING_WITHOUT_DECISIONS.slots_total(), 3);
+    }
+
+    #[test]
+    fn a_serving_agent_offers_its_slots_to_every_capability_it_serves() {
+        assert_eq!(
+            SERVING_WITHOUT_DECISIONS.slots_serving(InferenceCapability::Embeddings),
+            3
+        );
+        assert_eq!(
+            SERVING_WITHOUT_DECISIONS.slots_serving(InferenceCapability::TextGeneration),
+            3
+        );
+    }
+
+    #[test]
+    fn a_serving_agent_offers_no_slots_to_a_capability_it_does_not_serve() {
+        assert_eq!(
+            SERVING_WITHOUT_DECISIONS.slots_serving(InferenceCapability::Decisions),
+            0
+        );
+    }
+
+    #[test]
+    fn an_idle_agent_offers_no_slots_to_any_capability() {
+        assert_eq!(
+            AgentRuntimeStatus::Idle.slots_serving(InferenceCapability::TextGeneration),
+            0
+        );
+    }
+}
diff --git a/paddler_messaging/src/agent_state_application_status.rs b/paddler_messaging/src/agent_state_application_status.rs
index eca8bc95e..69a83abb5 100644
--- a/paddler_messaging/src/agent_state_application_status.rs
+++ b/paddler_messaging/src/agent_state_application_status.rs
@@ -4,49 +4,8 @@ use serde::Serialize;
 #[derive(Clone, Copy, Debug, Default, Deserialize, Eq, PartialEq, Serialize)]
 pub enum AgentStateApplicationStatus {
     Applied,
-    AttemptedAndNotAppliable,
     AttemptedAndRetrying,
     #[default]
     Fresh,
     Stuck,
 }
-
-impl AgentStateApplicationStatus {
-    #[must_use]
-    pub const fn should_try_to_apply(&self) -> bool {
-        match self {
-            Self::Applied | Self::AttemptedAndNotAppliable => false,
-            Self::AttemptedAndRetrying | Self::Fresh | Self::Stuck => true,
-        }
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use super::AgentStateApplicationStatus;
-
-    #[test]
-    fn applied_should_not_try_to_apply() {
-        assert!(!AgentStateApplicationStatus::Applied.should_try_to_apply());
-    }
-
-    #[test]
-    fn attempted_and_not_appliable_should_not_try_to_apply() {
-        assert!(!AgentStateApplicationStatus::AttemptedAndNotAppliable.should_try_to_apply());
-    }
-
-    #[test]
-    fn attempted_and_retrying_should_try_to_apply() {
-        assert!(AgentStateApplicationStatus::AttemptedAndRetrying.should_try_to_apply());
-    }
-
-    #[test]
-    fn fresh_should_try_to_apply() {
-        assert!(AgentStateApplicationStatus::Fresh.should_try_to_apply());
-    }
-
-    #[test]
-    fn stuck_should_try_to_apply() {
-        assert!(AgentStateApplicationStatus::Stuck.should_try_to_apply());
-    }
-}
diff --git a/paddler_messaging/src/agent_status.rs b/paddler_messaging/src/agent_status.rs
index 93f2d3cb2..49f42f22a 100644
--- a/paddler_messaging/src/agent_status.rs
+++ b/paddler_messaging/src/agent_status.rs
@@ -4,6 +4,7 @@ use serde::Deserialize;
 use serde::Serialize;
 
 use crate::agent_issue::AgentIssue;
+use crate::agent_runtime_status::AgentRuntimeStatus;
 use crate::agent_state_application_status::AgentStateApplicationStatus;
 use crate::model_download_status::ModelDownloadStatus;
 
@@ -14,7 +15,7 @@ pub struct AgentStatus {
     pub download_status: ModelDownloadStatus,
     pub issues: BTreeSet<AgentIssue>,
     pub model_path: Option<String>,
-    pub slots_total: u64,
+    pub runtime: AgentRuntimeStatus,
     pub state_application_status: AgentStateApplicationStatus,
     pub uses_chat_template_override: bool,
 }
diff --git a/paddler_messaging/src/api_path.rs b/paddler_messaging/src/api_path.rs
index be2ad3db7..15b294846 100644
--- a/paddler_messaging/src/api_path.rs
+++ b/paddler_messaging/src/api_path.rs
@@ -10,6 +10,7 @@ impl ApiPath {
     pub const CONTINUE_FROM_CONVERSATION_HISTORY: &str =
         "/api/v1/continue_from_conversation_history";
     pub const CONTINUE_FROM_RAW_PROMPT: &str = "/api/v1/continue_from_raw_prompt";
+    pub const DECIDE: &str = "/api/v1/decide";
     pub const GENERATE_EMBEDDING_BATCH: &str = "/api/v1/generate_embedding_batch";
     pub const HEALTH: &str = "/health";
     pub const INFERENCE_SOCKET: &str = "/api/v1/inference_socket";
diff --git a/paddler_balancer/src/compatibility/openai_service/assistant_role.rs b/paddler_messaging/src/assistant_role.rs
similarity index 100%
rename from paddler_balancer/src/compatibility/openai_service/assistant_role.rs
rename to paddler_messaging/src/assistant_role.rs
diff --git a/paddler_messaging/src/atomic_value.rs b/paddler_messaging/src/atomic_value.rs
index 57f099e4b..40b2c7752 100644
--- a/paddler_messaging/src/atomic_value.rs
+++ b/paddler_messaging/src/atomic_value.rs
@@ -37,8 +37,8 @@ impl AtomicValue<AtomicU64> {
             .is_ok()
     }
 
-    pub fn decrement(&self) {
-        self.value.fetch_sub(1, Ordering::SeqCst);
+    pub fn subtract(&self, amount: u64) {
+        self.value.fetch_sub(amount, Ordering::SeqCst);
     }
 
     pub fn get(&self) -> u64 {
@@ -51,15 +51,11 @@ impl AtomicValue<AtomicU64> {
 
     pub fn increment_below(&self, limit: u64) -> bool {
         self.value
-            .fetch_update(Ordering::SeqCst, Ordering::SeqCst, |current| {
+            .try_update(Ordering::SeqCst, Ordering::SeqCst, |current| {
                 (current < limit).then_some(current + 1)
             })
             .is_ok()
     }
-
-    pub fn reset(&self) {
-        self.value.store(0, Ordering::SeqCst);
-    }
 }
 
 #[cfg(test)]
diff --git a/paddler_messaging/src/balancer_connection.rs b/paddler_messaging/src/balancer_connection.rs
deleted file mode 100644
index df7b0b441..000000000
--- a/paddler_messaging/src/balancer_connection.rs
+++ /dev/null
@@ -1,5 +0,0 @@
-#[derive(Clone, Copy, Debug, Eq, PartialEq)]
-pub enum BalancerConnection {
-    Connected,
-    Connecting,
-}
diff --git a/paddler_messaging/src/balancer_desired_state.rs b/paddler_messaging/src/balancer_desired_state.rs
index b53691fa8..aec326e6e 100644
--- a/paddler_messaging/src/balancer_desired_state.rs
+++ b/paddler_messaging/src/balancer_desired_state.rs
@@ -1,17 +1,19 @@
 use serde::Deserialize;
 use serde::Serialize;
 
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
+use paddler_inference_parameters::embedding_parameters::EmbeddingParameters;
+use paddler_inference_parameters::model_runtime_parameters::ModelRuntimeParameters;
 
 use crate::agent_desired_model::AgentDesiredModel;
-use crate::chat_template::ChatTemplate;
+use crate::balancer_text_generation_settings::BalancerTextGenerationSettings;
+use crate::decision_settings::DecisionSettings;
 
 #[derive(Clone, Debug, Default, Deserialize, PartialEq, Serialize)]
 #[serde(deny_unknown_fields)]
 pub struct BalancerDesiredState {
-    pub chat_template_override: Option<ChatTemplate>,
-    pub inference_parameters: InferenceParameters,
+    pub decision: DecisionSettings,
+    pub embeddings: EmbeddingParameters,
     pub model: AgentDesiredModel,
-    pub multimodal_projection: AgentDesiredModel,
-    pub use_chat_template_override: bool,
+    pub model_runtime_parameters: ModelRuntimeParameters,
+    pub text_generation: BalancerTextGenerationSettings,
 }
diff --git a/paddler_messaging/src/balancer_text_generation_settings.rs b/paddler_messaging/src/balancer_text_generation_settings.rs
new file mode 100644
index 000000000..1e317e53b
--- /dev/null
+++ b/paddler_messaging/src/balancer_text_generation_settings.rs
@@ -0,0 +1,15 @@
+use serde::Deserialize;
+use serde::Serialize;
+
+use paddler_inference_parameters::sampling_parameters::SamplingParameters;
+
+use crate::chat_template_source::ChatTemplateSource;
+use crate::multimodal_settings::MultimodalSettings;
+
+#[derive(Clone, Debug, Default, Deserialize, PartialEq, Serialize)]
+#[serde(deny_unknown_fields)]
+pub struct BalancerTextGenerationSettings {
+    pub chat_template_source: ChatTemplateSource,
+    pub multimodal: MultimodalSettings,
+    pub sampling_parameters: SamplingParameters,
+}
diff --git a/paddler_messaging/src/chat_template_content_form.rs b/paddler_messaging/src/chat_template_content_form.rs
new file mode 100644
index 000000000..74e956b85
--- /dev/null
+++ b/paddler_messaging/src/chat_template_content_form.rs
@@ -0,0 +1,29 @@
+use crate::chat_template_message_content::ChatTemplateMessageContent;
+use crate::chat_template_message_content_part::ChatTemplateMessageContentPart;
+
+const TEXT_CONTENT_PART_TYPE: &str = "text";
+const TEXT_PARTS_SEPARATOR: &str = "\n";
+
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+pub enum ChatTemplateContentForm {
+    JoinedText,
+    TypedParts,
+}
+
+impl ChatTemplateContentForm {
+    #[must_use]
+    pub fn content_of(self, texts: Vec<String>) -> ChatTemplateMessageContent {
+        match self {
+            Self::JoinedText => ChatTemplateMessageContent::Text(texts.join(TEXT_PARTS_SEPARATOR)),
+            Self::TypedParts => ChatTemplateMessageContent::Parts(
+                texts
+                    .into_iter()
+                    .map(|text| ChatTemplateMessageContentPart {
+                        content_type: TEXT_CONTENT_PART_TYPE,
+                        text,
+                    })
+                    .collect(),
+            ),
+        }
+    }
+}
diff --git a/paddler_messaging/src/chat_template_message.rs b/paddler_messaging/src/chat_template_message.rs
index 3372af9da..8ad201247 100644
--- a/paddler_messaging/src/chat_template_message.rs
+++ b/paddler_messaging/src/chat_template_message.rs
@@ -1,9 +1,27 @@
 use serde::Serialize;
 
 use crate::chat_template_message_content::ChatTemplateMessageContent;
+use crate::chat_template_tool_call::ChatTemplateToolCall;
 
 #[derive(Clone, Debug, Eq, PartialEq, Serialize)]
-pub struct ChatTemplateMessage {
-    pub content: ChatTemplateMessageContent,
-    pub role: String,
+#[serde(untagged)]
+pub enum ChatTemplateMessage {
+    AssistantToolCalls {
+        content: ChatTemplateMessageContent,
+        #[serde(skip_serializing_if = "Option::is_none")]
+        reasoning_content: Option<String>,
+        role: &'static str,
+        tool_calls: Vec<ChatTemplateToolCall>,
+    },
+    Message {
+        content: ChatTemplateMessageContent,
+        #[serde(skip_serializing_if = "Option::is_none")]
+        reasoning_content: Option<String>,
+        role: String,
+    },
+    ToolResult {
+        content: ChatTemplateMessageContent,
+        role: &'static str,
+        tool_call_id: String,
+    },
 }
diff --git a/paddler_messaging/src/chat_template_source.rs b/paddler_messaging/src/chat_template_source.rs
new file mode 100644
index 000000000..1f00e41c3
--- /dev/null
+++ b/paddler_messaging/src/chat_template_source.rs
@@ -0,0 +1,12 @@
+use serde::Deserialize;
+use serde::Serialize;
+
+use crate::chat_template::ChatTemplate;
+
+#[derive(Clone, Debug, Default, Deserialize, Eq, PartialEq, Serialize)]
+#[serde(deny_unknown_fields)]
+pub enum ChatTemplateSource {
+    #[default]
+    EmbeddedInModel,
+    Override(ChatTemplate),
+}
diff --git a/paddler_messaging/src/chat_template_tool_call.rs b/paddler_messaging/src/chat_template_tool_call.rs
new file mode 100644
index 000000000..619c124e4
--- /dev/null
+++ b/paddler_messaging/src/chat_template_tool_call.rs
@@ -0,0 +1,30 @@
+use serde::Serialize;
+
+use crate::chat_template_tool_call_function::ChatTemplateToolCallFunction;
+use crate::conversation_tool_call::ConversationToolCall;
+
+const FUNCTION_TOOL_CALL_TYPE: &str = "function";
+
+#[derive(Clone, Debug, Eq, PartialEq, Serialize)]
+pub struct ChatTemplateToolCall {
+    pub function: ChatTemplateToolCallFunction,
+    pub id: String,
+    #[serde(rename = "type")]
+    pub tool_call_type: &'static str,
+}
+
+impl From<ConversationToolCall> for ChatTemplateToolCall {
+    fn from(
+        ConversationToolCall {
+            arguments,
+            id,
+            name,
+        }: ConversationToolCall,
+    ) -> Self {
+        Self {
+            function: ChatTemplateToolCallFunction { arguments, name },
+            id,
+            tool_call_type: FUNCTION_TOOL_CALL_TYPE,
+        }
+    }
+}
diff --git a/paddler_messaging/src/chat_template_tool_call_function.rs b/paddler_messaging/src/chat_template_tool_call_function.rs
new file mode 100644
index 000000000..f0e2facef
--- /dev/null
+++ b/paddler_messaging/src/chat_template_tool_call_function.rs
@@ -0,0 +1,9 @@
+use serde::Serialize;
+use serde_json::Map;
+use serde_json::Value;
+
+#[derive(Clone, Debug, Eq, PartialEq, Serialize)]
+pub struct ChatTemplateToolCallFunction {
+    pub arguments: Map<String, Value>,
+    pub name: String,
+}
diff --git a/paddler_messaging/src/conversation_history.rs b/paddler_messaging/src/conversation_history.rs
index d41d7a687..96cd84b33 100644
--- a/paddler_messaging/src/conversation_history.rs
+++ b/paddler_messaging/src/conversation_history.rs
@@ -1,13 +1,9 @@
 use serde::Deserialize;
 use serde::Serialize;
 
+use crate::chat_template_content_form::ChatTemplateContentForm;
 use crate::chat_template_conversation::ChatTemplateConversation;
-use crate::chat_template_message::ChatTemplateMessage;
-use crate::chat_template_message_content::ChatTemplateMessageContent;
-use crate::chat_template_message_content_part::ChatTemplateMessageContentPart;
 use crate::conversation_message::ConversationMessage;
-use crate::conversation_message_content::ConversationMessageContent;
-use crate::conversation_message_content_part::ConversationMessageContentPart;
 use crate::media_marker::MediaMarker;
 
 #[derive(Clone, Debug, Deserialize, Eq, PartialEq, Serialize)]
@@ -25,43 +21,40 @@ impl ConversationHistory {
     #[must_use]
     pub fn into_chat_template_conversation(
         self,
+        content_form: ChatTemplateContentForm,
         media_marker: &MediaMarker,
     ) -> ChatTemplateConversation {
         let mut image_urls = Vec::new();
-        let messages = self
-            .messages
-            .into_iter()
-            .map(
-                |ConversationMessage { content, role }| ChatTemplateMessage {
-                    content: match content {
-                        ConversationMessageContent::Text(text) => {
-                            ChatTemplateMessageContent::Text(text)
-                        }
-                        ConversationMessageContent::Parts(parts) => {
-                            ChatTemplateMessageContent::Parts(
-                                parts
-                                    .into_iter()
-                                    .map(|part| ChatTemplateMessageContentPart {
-                                        content_type: "text",
-                                        text: match part {
-                                            ConversationMessageContentPart::Text { text } => text,
-                                            ConversationMessageContentPart::ImageUrl {
-                                                image_url,
-                                            } => {
-                                                image_urls.push(image_url);
-
-                                                media_marker.marker.clone()
-                                            }
-                                        },
-                                    })
-                                    .collect(),
-                            )
-                        }
-                    },
-                    role,
-                },
-            )
-            .collect();
+        let mut messages = Vec::with_capacity(self.messages.len());
+        let mut conversation_messages = self.messages.into_iter().peekable();
+
+        while let Some(conversation_message) = conversation_messages.next() {
+            messages.push(match conversation_message {
+                ConversationMessage::AssistantReasoning { reasoning_content } => {
+                    match conversation_messages.next_if(ConversationMessage::is_assistant_reply) {
+                        Some(assistant_reply) => assistant_reply.into_chat_template_message(
+                            content_form,
+                            media_marker,
+                            &mut image_urls,
+                            Some(reasoning_content),
+                        ),
+                        None => ConversationMessage::AssistantReasoning { reasoning_content }
+                            .into_chat_template_message(
+                                content_form,
+                                media_marker,
+                                &mut image_urls,
+                                None,
+                            ),
+                    }
+                }
+                unreasoned_message => unreasoned_message.into_chat_template_message(
+                    content_form,
+                    media_marker,
+                    &mut image_urls,
+                    None,
+                ),
+            });
+        }
 
         ChatTemplateConversation {
             image_urls,
@@ -72,7 +65,12 @@ impl ConversationHistory {
 
 #[cfg(test)]
 mod tests {
+    use serde_json::Value;
+    use serde_json::json;
+    use serde_json::to_value;
+
     use super::ConversationHistory;
+    use crate::chat_template_content_form::ChatTemplateContentForm;
     use crate::chat_template_conversation::ChatTemplateConversation;
     use crate::chat_template_message::ChatTemplateMessage;
     use crate::chat_template_message_content::ChatTemplateMessageContent;
@@ -80,11 +78,12 @@ mod tests {
     use crate::conversation_message::ConversationMessage;
     use crate::conversation_message_content::ConversationMessageContent;
     use crate::conversation_message_content_part::ConversationMessageContentPart;
+    use crate::conversation_tool_call::ConversationToolCall;
     use crate::image_url::ImageUrl;
     use crate::media_marker::MediaMarker;
 
     fn make_text_message(role: &str, text: &str) -> ConversationMessage {
-        ConversationMessage {
+        ConversationMessage::Message {
             content: ConversationMessageContent::Text(text.to_owned()),
             role: role.to_owned(),
         }
@@ -94,12 +93,24 @@ mod tests {
         role: &str,
         parts: Vec<ConversationMessageContentPart>,
     ) -> ConversationMessage {
-        ConversationMessage {
+        ConversationMessage::Message {
             content: ConversationMessageContent::Parts(parts),
             role: role.to_owned(),
         }
     }
 
+    fn rendered_messages(history: ConversationHistory) -> Value {
+        to_value(
+            history
+                .into_chat_template_conversation(
+                    ChatTemplateContentForm::TypedParts,
+                    &MediaMarker::new("[IMAGE]".to_owned()),
+                )
+                .messages,
+        )
+        .unwrap()
+    }
+
     #[test]
     fn collects_image_urls_and_replaces_images_with_the_media_marker() {
         let history = ConversationHistory::new(vec![
@@ -123,17 +134,21 @@ mod tests {
         ]);
 
         assert_eq!(
-            history.into_chat_template_conversation(&MediaMarker::new("[IMAGE]".to_owned())),
+            history.into_chat_template_conversation(
+                ChatTemplateContentForm::TypedParts,
+                &MediaMarker::new("[IMAGE]".to_owned())
+            ),
             ChatTemplateConversation {
                 image_urls: vec![ImageUrl {
                     url: "http://example.com/img.png".to_owned(),
                 }],
                 messages: vec![
-                    ChatTemplateMessage {
+                    ChatTemplateMessage::Message {
                         content: ChatTemplateMessageContent::Text("hello".to_owned()),
+                        reasoning_content: None,
                         role: "assistant".to_owned(),
                     },
-                    ChatTemplateMessage {
+                    ChatTemplateMessage::Message {
                         content: ChatTemplateMessageContent::Parts(vec![
                             ChatTemplateMessageContentPart {
                                 content_type: "text",
@@ -148,10 +163,127 @@ mod tests {
                                 text: "after".to_owned(),
                             },
                         ]),
+                        reasoning_content: None,
                         role: "user".to_owned(),
                     },
                 ],
             }
         );
     }
+
+    #[test]
+    fn an_assistant_tool_call_turn_reaches_the_template_with_structured_tool_calls() {
+        let history = ConversationHistory::new(vec![ConversationMessage::AssistantToolCalls {
+            content: ConversationMessageContent::Text(String::new()),
+            tool_calls: vec![ConversationToolCall {
+                arguments: json!({ "location": "Paris" }).as_object().unwrap().clone(),
+                id: "call_weather".to_owned(),
+                name: "get_weather".to_owned(),
+            }],
+        }]);
+
+        assert_eq!(
+            rendered_messages(history),
+            json!([{
+                "content": "",
+                "role": "assistant",
+                "tool_calls": [{
+                    "function": { "arguments": { "location": "Paris" }, "name": "get_weather" },
+                    "id": "call_weather",
+                    "type": "function"
+                }]
+            }])
+        );
+    }
+
+    #[test]
+    fn a_tool_result_reaches_the_template_with_its_tool_call_id() {
+        let history = ConversationHistory::new(vec![ConversationMessage::ToolResult {
+            content: ConversationMessageContent::Text("sunny".to_owned()),
+            tool_call_id: "call_weather".to_owned(),
+        }]);
+
+        assert_eq!(
+            rendered_messages(history),
+            json!([{ "content": "sunny", "role": "tool", "tool_call_id": "call_weather" }])
+        );
+    }
+
+    #[test]
+    fn a_reasoning_reaches_the_template_on_the_assistant_reply_after_it() {
+        let history = ConversationHistory::new(vec![
+            ConversationMessage::AssistantReasoning {
+                reasoning_content: "the user wants the weather".to_owned(),
+            },
+            ConversationMessage::AssistantToolCalls {
+                content: ConversationMessageContent::Text(String::new()),
+                tool_calls: vec![ConversationToolCall {
+                    arguments: json!({ "location": "Paris" }).as_object().unwrap().clone(),
+                    id: "call_weather".to_owned(),
+                    name: "get_weather".to_owned(),
+                }],
+            },
+        ]);
+
+        assert_eq!(
+            rendered_messages(history),
+            json!([{
+                "content": "",
+                "reasoning_content": "the user wants the weather",
+                "role": "assistant",
+                "tool_calls": [{
+                    "function": { "arguments": { "location": "Paris" }, "name": "get_weather" },
+                    "id": "call_weather",
+                    "type": "function"
+                }]
+            }])
+        );
+    }
+
+    #[test]
+    fn a_reasoning_no_assistant_reply_follows_reaches_the_template_as_its_own_assistant_turn() {
+        let history = ConversationHistory::new(vec![
+            ConversationMessage::AssistantReasoning {
+                reasoning_content: "nothing to answer yet".to_owned(),
+            },
+            make_text_message("user", "hello"),
+        ]);
+
+        assert_eq!(
+            rendered_messages(history),
+            json!([
+                {
+                    "content": "",
+                    "reasoning_content": "nothing to answer yet",
+                    "role": "assistant"
+                },
+                { "content": "hello", "role": "user" }
+            ])
+        );
+    }
+
+    #[test]
+    fn a_reasoning_followed_by_a_tool_result_is_not_attached_to_that_tool_result() {
+        let history = ConversationHistory::new(vec![
+            ConversationMessage::AssistantReasoning {
+                reasoning_content: "waiting for the weather".to_owned(),
+            },
+            ConversationMessage::ToolResult {
+                content: ConversationMessageContent::Text("sunny".to_owned()),
+                tool_call_id: "call_weather".to_owned(),
+            },
+        ]);
+
+        assert_eq!(
+            rendered_messages(history),
+            json!([
+                {
+                    "content": "",
+                    "reasoning_content": "waiting for the weather",
+                    "role": "assistant"
+                },
+                { "content": "sunny", "role": "tool", "tool_call_id": "call_weather" }
+            ])
+        );
+    }
 }
diff --git a/paddler_messaging/src/conversation_message.rs b/paddler_messaging/src/conversation_message.rs
index 65a826b7e..cf2ffeeab 100644
--- a/paddler_messaging/src/conversation_message.rs
+++ b/paddler_messaging/src/conversation_message.rs
@@ -1,11 +1,81 @@
 use serde::Deserialize;
 use serde::Serialize;
 
+use crate::assistant_role::ASSISTANT_ROLE;
+use crate::chat_template_content_form::ChatTemplateContentForm;
+use crate::chat_template_message::ChatTemplateMessage;
+use crate::chat_template_message_content::ChatTemplateMessageContent;
 use crate::conversation_message_content::ConversationMessageContent;
+use crate::conversation_tool_call::ConversationToolCall;
+use crate::image_url::ImageUrl;
+use crate::media_marker::MediaMarker;
+use crate::tool_role::TOOL_ROLE;
 
 #[derive(Clone, Debug, Deserialize, Eq, PartialEq, Serialize)]
 #[serde(deny_unknown_fields)]
-pub struct ConversationMessage {
-    pub content: ConversationMessageContent,
-    pub role: String,
+pub enum ConversationMessage {
+    AssistantReasoning {
+        reasoning_content: String,
+    },
+    AssistantToolCalls {
+        content: ConversationMessageContent,
+        tool_calls: Vec<ConversationToolCall>,
+    },
+    Message {
+        content: ConversationMessageContent,
+        role: String,
+    },
+    ToolResult {
+        content: ConversationMessageContent,
+        tool_call_id: String,
+    },
+}
+
+impl ConversationMessage {
+    #[must_use]
+    pub fn is_assistant_reply(&self) -> bool {
+        match self {
+            Self::AssistantToolCalls { .. } => true,
+            Self::Message { role, .. } => role == ASSISTANT_ROLE,
+            Self::AssistantReasoning { .. } | Self::ToolResult { .. } => false,
+        }
+    }
+
+    pub fn into_chat_template_message(
+        self,
+        content_form: ChatTemplateContentForm,
+        media_marker: &MediaMarker,
+        image_urls: &mut Vec<ImageUrl>,
+        preceding_reasoning_content: Option<String>,
+    ) -> ChatTemplateMessage {
+        match self {
+            Self::AssistantReasoning { reasoning_content } => ChatTemplateMessage::Message {
+                content: ChatTemplateMessageContent::Text(String::new()),
+                reasoning_content: Some(reasoning_content),
+                role: ASSISTANT_ROLE.to_owned(),
+            },
+            Self::AssistantToolCalls {
+                content,
+                tool_calls,
+            } => ChatTemplateMessage::AssistantToolCalls {
+                content: content.into_chat_template_content(content_form, media_marker, image_urls),
+                reasoning_content: preceding_reasoning_content,
+                role: ASSISTANT_ROLE,
+                tool_calls: tool_calls.into_iter().map(Into::into).collect(),
+            },
+            Self::Message { content, role } => ChatTemplateMessage::Message {
+                content: content.into_chat_template_content(content_form, media_marker, image_urls),
+                reasoning_content: preceding_reasoning_content,
+                role,
+            },
+            Self::ToolResult {
+                content,
+                tool_call_id,
+            } => ChatTemplateMessage::ToolResult {
+                content: content.into_chat_template_content(content_form, media_marker, image_urls),
+                role: TOOL_ROLE,
+                tool_call_id,
+            },
+        }
+    }
 }
diff --git a/paddler_messaging/src/conversation_message_content.rs b/paddler_messaging/src/conversation_message_content.rs
index 823d451f3..8c856cf41 100644
--- a/paddler_messaging/src/conversation_message_content.rs
+++ b/paddler_messaging/src/conversation_message_content.rs
@@ -1,8 +1,11 @@
 use serde::Deserialize;
 use serde::Serialize;
 
+use crate::chat_template_content_form::ChatTemplateContentForm;
+use crate::chat_template_message_content::ChatTemplateMessageContent;
 use crate::conversation_message_content_part::ConversationMessageContentPart;
 use crate::image_url::ImageUrl;
+use crate::media_marker::MediaMarker;
 
 #[derive(Clone, Debug, Deserialize, Eq, PartialEq, Serialize)]
 #[serde(untagged)]
@@ -12,6 +15,30 @@ pub enum ConversationMessageContent {
 }
 
 impl ConversationMessageContent {
+    pub fn into_chat_template_content(
+        self,
+        content_form: ChatTemplateContentForm,
+        media_marker: &MediaMarker,
+        image_urls: &mut Vec<ImageUrl>,
+    ) -> ChatTemplateMessageContent {
+        match self {
+            Self::Text(text) => ChatTemplateMessageContent::Text(text),
+            Self::Parts(parts) => content_form.content_of(
+                parts
+                    .into_iter()
+                    .map(|part| match part {
+                        ConversationMessageContentPart::Text { text } => text,
+                        ConversationMessageContentPart::ImageUrl { image_url } => {
+                            image_urls.push(image_url);
+
+                            media_marker.marker.clone()
+                        }
+                    })
+                    .collect(),
+            ),
+        }
+    }
+
     #[must_use]
     pub fn text_content(&self) -> String {
         match self {
@@ -44,9 +71,33 @@ impl ConversationMessageContent {
 
 #[cfg(test)]
 mod tests {
+    use crate::chat_template_content_form::ChatTemplateContentForm;
+    use crate::chat_template_message_content::ChatTemplateMessageContent;
     use crate::conversation_message_content::ConversationMessageContent;
     use crate::conversation_message_content_part::ConversationMessageContentPart;
     use crate::image_url::ImageUrl;
+    use crate::media_marker::MediaMarker;
+
+    #[test]
+    fn a_template_taking_text_receives_the_text_parts_joined_by_newlines() {
+        let content = ConversationMessageContent::Parts(vec![
+            ConversationMessageContentPart::Text {
+                text: "first".to_owned(),
+            },
+            ConversationMessageContentPart::Text {
+                text: "second".to_owned(),
+            },
+        ]);
+
+        assert_eq!(
+            content.into_chat_template_content(
+                ChatTemplateContentForm::JoinedText,
+                &MediaMarker::new("[IMAGE]".to_owned()),
+                &mut Vec::new(),
+            ),
+            ChatTemplateMessageContent::Text("first\nsecond".to_owned())
+        );
+    }
 
     #[test]
     fn text_content_from_text_variant() {
@@ -78,7 +129,7 @@ mod tests {
     fn image_urls_from_text_variant_is_empty() {
         let content = ConversationMessageContent::Text("hello".to_owned());
 
-        assert!(content.image_urls().is_empty());
+        assert_eq!(content.image_urls(), Vec::<&ImageUrl>::new());
     }
 
     #[test]
diff --git a/paddler_messaging/src/conversation_tool_call.rs b/paddler_messaging/src/conversation_tool_call.rs
new file mode 100644
index 000000000..fdb590048
--- /dev/null
+++ b/paddler_messaging/src/conversation_tool_call.rs
@@ -0,0 +1,12 @@
+use serde::Deserialize;
+use serde::Serialize;
+use serde_json::Map;
+use serde_json::Value;
+
+#[derive(Clone, Debug, Deserialize, Eq, PartialEq, Serialize)]
+#[serde(deny_unknown_fields)]
+pub struct ConversationToolCall {
+    pub arguments: Map<String, Value>,
+    pub id: String,
+    pub name: String,
+}
diff --git a/paddler_messaging/src/decision_answer.rs b/paddler_messaging/src/decision_answer.rs
new file mode 100644
index 000000000..321d941ef
--- /dev/null
+++ b/paddler_messaging/src/decision_answer.rs
@@ -0,0 +1,9 @@
+use serde::Deserialize;
+use serde::Serialize;
+
+#[derive(Clone, Debug, Deserialize, PartialEq, Serialize)]
+#[serde(deny_unknown_fields)]
+pub struct DecisionAnswer {
+    pub id: String,
+    pub probabilities: Vec<f32>,
+}
diff --git a/paddler_messaging/src/decision_final_result.rs b/paddler_messaging/src/decision_final_result.rs
new file mode 100644
index 000000000..a18f55106
--- /dev/null
+++ b/paddler_messaging/src/decision_final_result.rs
@@ -0,0 +1,19 @@
+use crate::decision_summary::DecisionSummary;
+use crate::oversized_decision_details::OversizedDecisionDetails;
+
+#[derive(Debug, Eq, PartialEq)]
+pub enum DecisionFinalResult {
+    AgentRuntimeFailed(String),
+    BatchAssemblyFailed(String),
+    DecisionsNotServed(String),
+    DecodeFailed(String),
+    Done(DecisionSummary),
+    HiddenStateUnavailable(String),
+    InputTokenizationFailed(String),
+    KvCacheCopyFailed(String),
+    KvCacheRemovalFailed(String),
+    ModelNotLoaded(String),
+    RequestExceedsContext(OversizedDecisionDetails),
+    SchedulerUnavailable(String),
+    StopRequested,
+}
diff --git a/paddler_messaging/src/decision_question.rs b/paddler_messaging/src/decision_question.rs
new file mode 100644
index 000000000..7d871b27a
--- /dev/null
+++ b/paddler_messaging/src/decision_question.rs
@@ -0,0 +1,10 @@
+use serde::Deserialize;
+use serde::Serialize;
+
+#[derive(Clone, Debug, Deserialize, Eq, PartialEq, Serialize)]
+#[serde(deny_unknown_fields)]
+pub struct DecisionQuestion {
+    pub id: String,
+    pub instructions: String,
+    pub options: Vec<String>,
+}
diff --git a/paddler_messaging/src/decision_result.rs b/paddler_messaging/src/decision_result.rs
new file mode 100644
index 000000000..4acf9a0bb
--- /dev/null
+++ b/paddler_messaging/src/decision_result.rs
@@ -0,0 +1,217 @@
+use serde::Deserialize;
+use serde::Serialize;
+
+use crate::decision_answer::DecisionAnswer;
+use crate::decision_final_result::DecisionFinalResult;
+use crate::decision_summary::DecisionSummary;
+use crate::oversized_decision_details::OversizedDecisionDetails;
+use crate::result_step::ResultStep;
+use crate::streamable_result::StreamableResult;
+
+#[derive(Clone, Debug, Deserialize, PartialEq, Serialize)]
+#[serde(deny_unknown_fields)]
+pub enum DecisionResult {
+    AgentRuntimeFailed(String),
+    BatchAssemblyFailed(String),
+    DecisionsNotServed(String),
+    DecodeFailed(String),
+    Done(DecisionSummary),
+    HiddenStateUnavailable(String),
+    InputTokenizationFailed(String),
+    KvCacheCopyFailed(String),
+    KvCacheRemovalFailed(String),
+    ModelNotLoaded(String),
+    QuestionAnswered(DecisionAnswer),
+    RequestExceedsContext(OversizedDecisionDetails),
+    SchedulerUnavailable(String),
+    StopRequested,
+}
+
+impl From<DecisionAnswer> for DecisionResult {
+    fn from(answer: DecisionAnswer) -> Self {
+        Self::QuestionAnswered(answer)
+    }
+}
+
+impl From<DecisionFinalResult> for DecisionResult {
+    fn from(final_result: DecisionFinalResult) -> Self {
+        match final_result {
+            DecisionFinalResult::AgentRuntimeFailed(detail) => Self::AgentRuntimeFailed(detail),
+            DecisionFinalResult::BatchAssemblyFailed(detail) => Self::BatchAssemblyFailed(detail),
+            DecisionFinalResult::DecisionsNotServed(detail) => Self::DecisionsNotServed(detail),
+            DecisionFinalResult::DecodeFailed(detail) => Self::DecodeFailed(detail),
+            DecisionFinalResult::Done(summary) => Self::Done(summary),
+            DecisionFinalResult::HiddenStateUnavailable(detail) => {
+                Self::HiddenStateUnavailable(detail)
+            }
+            DecisionFinalResult::InputTokenizationFailed(detail) => {
+                Self::InputTokenizationFailed(detail)
+            }
+            DecisionFinalResult::KvCacheCopyFailed(detail) => Self::KvCacheCopyFailed(detail),
+            DecisionFinalResult::KvCacheRemovalFailed(detail) => Self::KvCacheRemovalFailed(detail),
+            DecisionFinalResult::ModelNotLoaded(detail) => Self::ModelNotLoaded(detail),
+            DecisionFinalResult::RequestExceedsContext(details) => {
+                Self::RequestExceedsContext(details)
+            }
+            DecisionFinalResult::SchedulerUnavailable(detail) => Self::SchedulerUnavailable(detail),
+            DecisionFinalResult::StopRequested => Self::StopRequested,
+        }
+    }
+}
+
+impl StreamableResult for DecisionResult {
+    type Final = DecisionFinalResult;
+    type Partial = DecisionAnswer;
+
+    fn into_step(self) -> ResultStep<DecisionAnswer, DecisionFinalResult> {
+        match self {
+            Self::AgentRuntimeFailed(detail) => {
+                ResultStep::Final(DecisionFinalResult::AgentRuntimeFailed(detail))
+            }
+            Self::BatchAssemblyFailed(detail) => {
+                ResultStep::Final(DecisionFinalResult::BatchAssemblyFailed(detail))
+            }
+            Self::DecisionsNotServed(detail) => {
+                ResultStep::Final(DecisionFinalResult::DecisionsNotServed(detail))
+            }
+            Self::DecodeFailed(detail) => {
+                ResultStep::Final(DecisionFinalResult::DecodeFailed(detail))
+            }
+            Self::Done(summary) => ResultStep::Final(DecisionFinalResult::Done(summary)),
+            Self::HiddenStateUnavailable(detail) => {
+                ResultStep::Final(DecisionFinalResult::HiddenStateUnavailable(detail))
+            }
+            Self::InputTokenizationFailed(detail) => {
+                ResultStep::Final(DecisionFinalResult::InputTokenizationFailed(detail))
+            }
+            Self::KvCacheCopyFailed(detail) => {
+                ResultStep::Final(DecisionFinalResult::KvCacheCopyFailed(detail))
+            }
+            Self::KvCacheRemovalFailed(detail) => {
+                ResultStep::Final(DecisionFinalResult::KvCacheRemovalFailed(detail))
+            }
+            Self::ModelNotLoaded(detail) => {
+                ResultStep::Final(DecisionFinalResult::ModelNotLoaded(detail))
+            }
+            Self::RequestExceedsContext(details) => {
+                ResultStep::Final(DecisionFinalResult::RequestExceedsContext(details))
+            }
+            Self::SchedulerUnavailable(detail) => {
+                ResultStep::Final(DecisionFinalResult::SchedulerUnavailable(detail))
+            }
+            Self::QuestionAnswered(answer) => ResultStep::Partial(answer),
+            Self::StopRequested => ResultStep::Final(DecisionFinalResult::StopRequested),
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::DecisionResult;
+    use crate::decision_answer::DecisionAnswer;
+    use crate::decision_final_result::DecisionFinalResult;
+    use crate::decision_summary::DecisionSummary;
+    use crate::oversized_decision_details::OversizedDecisionDetails;
+    use crate::result_step::ResultStep;
+    use crate::streamable_result::StreamableResult as _;
+
+    fn step_round_trip(decision_result: DecisionResult) -> DecisionResult {
+        match decision_result.into_step() {
+            ResultStep::Final(final_result) => final_result.into(),
+            ResultStep::Partial(answer) => answer.into(),
+        }
+    }
+
+    #[test]
+    fn answering_a_question_keeps_the_stream_open() {
+        let answer = DecisionAnswer {
+            id: "question".to_owned(),
+            probabilities: vec![1.0],
+        };
+
+        assert_eq!(
+            step_round_trip(DecisionResult::QuestionAnswered(answer.clone())),
+            DecisionResult::QuestionAnswered(answer.clone())
+        );
+        assert_eq!(
+            DecisionResult::QuestionAnswered(answer.clone()).into_step(),
+            ResultStep::Partial(answer)
+        );
+    }
+
+    #[test]
+    fn every_other_result_ends_the_stream_as_its_final_result() {
+        for (decision_result, expected_final_result) in [
+            (
+                DecisionResult::AgentRuntimeFailed("runtime".to_owned()),
+                DecisionFinalResult::AgentRuntimeFailed("runtime".to_owned()),
+            ),
+            (
+                DecisionResult::BatchAssemblyFailed("batch".to_owned()),
+                DecisionFinalResult::BatchAssemblyFailed("batch".to_owned()),
+            ),
+            (
+                DecisionResult::DecisionsNotServed("no pointer head".to_owned()),
+                DecisionFinalResult::DecisionsNotServed("no pointer head".to_owned()),
+            ),
+            (
+                DecisionResult::DecodeFailed("decode".to_owned()),
+                DecisionFinalResult::DecodeFailed("decode".to_owned()),
+            ),
+            (
+                DecisionResult::Done(DecisionSummary {
+                    input_tokens: 3,
+                    processing_milliseconds: 5,
+                }),
+                DecisionFinalResult::Done(DecisionSummary {
+                    input_tokens: 3,
+                    processing_milliseconds: 5,
+                }),
+            ),
+            (
+                DecisionResult::HiddenStateUnavailable("hidden".to_owned()),
+                DecisionFinalResult::HiddenStateUnavailable("hidden".to_owned()),
+            ),
+            (
+                DecisionResult::InputTokenizationFailed("tokenization".to_owned()),
+                DecisionFinalResult::InputTokenizationFailed("tokenization".to_owned()),
+            ),
+            (
+                DecisionResult::KvCacheCopyFailed("copy".to_owned()),
+                DecisionFinalResult::KvCacheCopyFailed("copy".to_owned()),
+            ),
+            (
+                DecisionResult::KvCacheRemovalFailed("removal".to_owned()),
+                DecisionFinalResult::KvCacheRemovalFailed("removal".to_owned()),
+            ),
+            (
+                DecisionResult::ModelNotLoaded("no model".to_owned()),
+                DecisionFinalResult::ModelNotLoaded("no model".to_owned()),
+            ),
+            (
+                DecisionResult::RequestExceedsContext(OversizedDecisionDetails {
+                    context_size: 128,
+                    required_tokens: 410,
+                }),
+                DecisionFinalResult::RequestExceedsContext(OversizedDecisionDetails {
+                    context_size: 128,
+                    required_tokens: 410,
+                }),
+            ),
+            (
+                DecisionResult::SchedulerUnavailable("stopped".to_owned()),
+                DecisionFinalResult::SchedulerUnavailable("stopped".to_owned()),
+            ),
+            (
+                DecisionResult::StopRequested,
+                DecisionFinalResult::StopRequested,
+            ),
+        ] {
+            assert_eq!(step_round_trip(decision_result.clone()), decision_result);
+            assert_eq!(
+                decision_result.into_step(),
+                ResultStep::Final(expected_final_result)
+            );
+        }
+    }
+}
diff --git a/paddler_messaging/src/decision_settings.rs b/paddler_messaging/src/decision_settings.rs
new file mode 100644
index 000000000..268e09041
--- /dev/null
+++ b/paddler_messaging/src/decision_settings.rs
@@ -0,0 +1,55 @@
+use serde::Deserialize;
+use serde::Serialize;
+
+use crate::agent_decisions_settings::AgentDecisionsSettings;
+use crate::agent_desired_model::AgentDesiredModel;
+use crate::served_decision_settings::ServedDecisionSettings;
+
+#[derive(Clone, Debug, Default, Deserialize, Eq, PartialEq, Serialize)]
+#[serde(deny_unknown_fields)]
+pub struct DecisionSettings {
+    pub pointer_head: AgentDesiredModel,
+}
+
+impl DecisionSettings {
+    #[must_use]
+    pub fn agent_decisions(&self) -> AgentDecisionsSettings {
+        match &self.pointer_head {
+            AgentDesiredModel::None => AgentDecisionsSettings::NotServed,
+            AgentDesiredModel::Uri(pointer_head_uri) => {
+                AgentDecisionsSettings::Served(ServedDecisionSettings {
+                    pointer_head_uri: pointer_head_uri.clone(),
+                })
+            }
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::DecisionSettings;
+    use crate::agent_decisions_settings::AgentDecisionsSettings;
+    use crate::agent_desired_model::AgentDesiredModel;
+    use crate::served_decision_settings::ServedDecisionSettings;
+
+    #[test]
+    fn agents_serve_no_decisions_while_no_pointer_head_is_configured() {
+        assert_eq!(
+            DecisionSettings::default().agent_decisions(),
+            AgentDecisionsSettings::NotServed
+        );
+    }
+
+    #[test]
+    fn agents_serve_decisions_with_the_configured_pointer_head() {
+        assert_eq!(
+            DecisionSettings {
+                pointer_head: AgentDesiredModel::Uri("pointer_head.gguf".to_owned()),
+            }
+            .agent_decisions(),
+            AgentDecisionsSettings::Served(ServedDecisionSettings {
+                pointer_head_uri: "pointer_head.gguf".to_owned(),
+            })
+        );
+    }
+}
diff --git a/paddler_messaging/src/decision_slots_for_leading_questions.rs b/paddler_messaging/src/decision_slots_for_leading_questions.rs
new file mode 100644
index 000000000..2c3ddbc13
--- /dev/null
+++ b/paddler_messaging/src/decision_slots_for_leading_questions.rs
@@ -0,0 +1,33 @@
+use std::num::NonZeroU16;
+
+use crate::decision_slots_minimum::DECISION_SLOTS_MINIMUM;
+
+#[must_use]
+pub const fn decision_slots_for_leading_questions(leading_question_count: usize) -> NonZeroU16 {
+    if leading_question_count == 0 {
+        NonZeroU16::MIN
+    } else {
+        DECISION_SLOTS_MINIMUM
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::num::NonZeroU16;
+
+    use super::decision_slots_for_leading_questions;
+    use crate::decision_slots_minimum::DECISION_SLOTS_MINIMUM;
+
+    #[test]
+    fn a_decision_without_leading_questions_needs_only_its_state_slot() {
+        assert_eq!(decision_slots_for_leading_questions(0), NonZeroU16::MIN);
+    }
+
+    #[test]
+    fn a_decision_with_leading_questions_needs_its_state_and_question_lane_slots() {
+        assert_eq!(
+            decision_slots_for_leading_questions(3),
+            DECISION_SLOTS_MINIMUM
+        );
+    }
+}
diff --git a/paddler_messaging/src/decision_slots_minimum.rs b/paddler_messaging/src/decision_slots_minimum.rs
new file mode 100644
index 000000000..1c17ab287
--- /dev/null
+++ b/paddler_messaging/src/decision_slots_minimum.rs
@@ -0,0 +1,3 @@
+use std::num::NonZeroU16;
+
+pub const DECISION_SLOTS_MINIMUM: NonZeroU16 = NonZeroU16::new(2).unwrap();
diff --git a/paddler_messaging/src/decision_summary.rs b/paddler_messaging/src/decision_summary.rs
new file mode 100644
index 000000000..7241f6c85
--- /dev/null
+++ b/paddler_messaging/src/decision_summary.rs
@@ -0,0 +1,9 @@
+use serde::Deserialize;
+use serde::Serialize;
+
+#[derive(Clone, Debug, Deserialize, Eq, PartialEq, Serialize)]
+#[serde(deny_unknown_fields)]
+pub struct DecisionSummary {
+    pub input_tokens: usize,
+    pub processing_milliseconds: u64,
+}
diff --git a/paddler_messaging/src/embedded_chat_template_key.rs b/paddler_messaging/src/embedded_chat_template_key.rs
new file mode 100644
index 000000000..d16ccea91
--- /dev/null
+++ b/paddler_messaging/src/embedded_chat_template_key.rs
@@ -0,0 +1 @@
+pub const EMBEDDED_CHAT_TEMPLATE_KEY: &str = "tokenizer.chat_template";
diff --git a/paddler_messaging/src/embedding.rs b/paddler_messaging/src/embedding.rs
index 6ee29a065..bbd35ca0f 100644
--- a/paddler_messaging/src/embedding.rs
+++ b/paddler_messaging/src/embedding.rs
@@ -4,11 +4,12 @@ use serde::Serialize;
 use paddler_inference_parameters::pooling_type::PoolingType;
 
 use crate::embedding_normalization_method::EmbeddingNormalizationMethod;
+use crate::embedding_vector::EmbeddingVector;
 
-#[derive(Debug, Deserialize, PartialEq, Serialize)]
+#[derive(Clone, Debug, Deserialize, PartialEq, Serialize)]
 #[serde(deny_unknown_fields)]
 pub struct Embedding {
-    pub embedding: Vec<f32>,
+    pub embedding: EmbeddingVector,
     pub normalization_method: EmbeddingNormalizationMethod,
     pub pooling_type: PoolingType,
     pub source_document_id: String,
diff --git a/paddler_messaging/src/embedding_document_exceeding_sequence_context_details.rs b/paddler_messaging/src/embedding_document_exceeding_sequence_context_details.rs
new file mode 100644
index 000000000..c20469a6d
--- /dev/null
+++ b/paddler_messaging/src/embedding_document_exceeding_sequence_context_details.rs
@@ -0,0 +1,10 @@
+use serde::Deserialize;
+use serde::Serialize;
+
+#[derive(Clone, Debug, Deserialize, Eq, PartialEq, Serialize)]
+#[serde(deny_unknown_fields)]
+pub struct EmbeddingDocumentExceedingSequenceContextDetails {
+    pub document_tokens: usize,
+    pub sequence_context_size: usize,
+    pub source_document_id: String,
+}
diff --git a/paddler_messaging/src/embedding_document_tokenized_to_nothing_details.rs b/paddler_messaging/src/embedding_document_tokenized_to_nothing_details.rs
new file mode 100644
index 000000000..609b4c0d3
--- /dev/null
+++ b/paddler_messaging/src/embedding_document_tokenized_to_nothing_details.rs
@@ -0,0 +1,8 @@
+use serde::Deserialize;
+use serde::Serialize;
+
+#[derive(Clone, Debug, Deserialize, Eq, PartialEq, Serialize)]
+#[serde(deny_unknown_fields)]
+pub struct EmbeddingDocumentTokenizedToNothingDetails {
+    pub source_document_id: String,
+}
diff --git a/paddler_messaging/src/embedding_final_result.rs b/paddler_messaging/src/embedding_final_result.rs
new file mode 100644
index 000000000..3f7704e65
--- /dev/null
+++ b/paddler_messaging/src/embedding_final_result.rs
@@ -0,0 +1,13 @@
+#[derive(Debug, Eq, PartialEq)]
+pub enum EmbeddingFinalResult {
+    AgentRuntimeFailed(String),
+    BatchAssemblyFailed(String),
+    DecodeFailed(String),
+    Done,
+    EmbeddingTooLongForRmsNormalization(String),
+    EmbeddingsUnavailable(String),
+    InputTokenizationFailed(String),
+    ModelNotLoaded(String),
+    NoEmbeddingsProduced,
+    SchedulerUnavailable(String),
+}
diff --git a/paddler_messaging/src/embedding_partial_result.rs b/paddler_messaging/src/embedding_partial_result.rs
new file mode 100644
index 000000000..1ed5ba2c7
--- /dev/null
+++ b/paddler_messaging/src/embedding_partial_result.rs
@@ -0,0 +1,12 @@
+use crate::embedding::Embedding;
+use crate::embedding_document_exceeding_sequence_context_details::EmbeddingDocumentExceedingSequenceContextDetails;
+use crate::embedding_document_tokenized_to_nothing_details::EmbeddingDocumentTokenizedToNothingDetails;
+use crate::oversized_embedding_document_details::OversizedEmbeddingDocumentDetails;
+
+#[derive(Debug, PartialEq)]
+pub enum EmbeddingPartialResult {
+    DocumentExceedsBatchSize(OversizedEmbeddingDocumentDetails),
+    DocumentExceedsSequenceContext(EmbeddingDocumentExceedingSequenceContextDetails),
+    DocumentTokenizedToNothing(EmbeddingDocumentTokenizedToNothingDetails),
+    Embedding(Embedding),
+}
diff --git a/paddler_messaging/src/embedding_result.rs b/paddler_messaging/src/embedding_result.rs
index cf06814a1..323c357d2 100644
--- a/paddler_messaging/src/embedding_result.rs
+++ b/paddler_messaging/src/embedding_result.rs
@@ -2,33 +2,122 @@ use serde::Deserialize;
 use serde::Serialize;
 
 use crate::embedding::Embedding;
+use crate::embedding_document_exceeding_sequence_context_details::EmbeddingDocumentExceedingSequenceContextDetails;
+use crate::embedding_document_tokenized_to_nothing_details::EmbeddingDocumentTokenizedToNothingDetails;
+use crate::embedding_final_result::EmbeddingFinalResult;
+use crate::embedding_partial_result::EmbeddingPartialResult;
 use crate::oversized_embedding_document_details::OversizedEmbeddingDocumentDetails;
+use crate::result_step::ResultStep;
 use crate::streamable_result::StreamableResult;
 
-#[derive(Debug, Deserialize, PartialEq, Serialize)]
+#[derive(Clone, Debug, Deserialize, PartialEq, Serialize)]
 #[serde(deny_unknown_fields)]
 pub enum EmbeddingResult {
+    AgentRuntimeFailed(String),
+    BatchAssemblyFailed(String),
+    DecodeFailed(String),
     DocumentExceedsBatchSize(OversizedEmbeddingDocumentDetails),
+    DocumentExceedsSequenceContext(EmbeddingDocumentExceedingSequenceContextDetails),
+    DocumentTokenizedToNothing(EmbeddingDocumentTokenizedToNothingDetails),
     Done,
     Embedding(Embedding),
-    EmbeddingsDisabled,
-    Error(String),
-    EmbeddingRejectedDueToActiveTokenGeneration,
+    EmbeddingTooLongForRmsNormalization(String),
+    EmbeddingsUnavailable(String),
+    InputTokenizationFailed(String),
     ModelNotLoaded(String),
     NoEmbeddingsProduced,
+    SchedulerUnavailable(String),
+}
+
+impl From<EmbeddingFinalResult> for EmbeddingResult {
+    fn from(final_result: EmbeddingFinalResult) -> Self {
+        match final_result {
+            EmbeddingFinalResult::AgentRuntimeFailed(detail) => Self::AgentRuntimeFailed(detail),
+            EmbeddingFinalResult::BatchAssemblyFailed(detail) => Self::BatchAssemblyFailed(detail),
+            EmbeddingFinalResult::DecodeFailed(detail) => Self::DecodeFailed(detail),
+            EmbeddingFinalResult::Done => Self::Done,
+            EmbeddingFinalResult::EmbeddingTooLongForRmsNormalization(detail) => {
+                Self::EmbeddingTooLongForRmsNormalization(detail)
+            }
+            EmbeddingFinalResult::EmbeddingsUnavailable(detail) => {
+                Self::EmbeddingsUnavailable(detail)
+            }
+            EmbeddingFinalResult::InputTokenizationFailed(detail) => {
+                Self::InputTokenizationFailed(detail)
+            }
+            EmbeddingFinalResult::ModelNotLoaded(detail) => Self::ModelNotLoaded(detail),
+            EmbeddingFinalResult::NoEmbeddingsProduced => Self::NoEmbeddingsProduced,
+            EmbeddingFinalResult::SchedulerUnavailable(detail) => {
+                Self::SchedulerUnavailable(detail)
+            }
+        }
+    }
+}
+
+impl From<EmbeddingPartialResult> for EmbeddingResult {
+    fn from(partial_result: EmbeddingPartialResult) -> Self {
+        match partial_result {
+            EmbeddingPartialResult::DocumentExceedsBatchSize(details) => {
+                Self::DocumentExceedsBatchSize(details)
+            }
+            EmbeddingPartialResult::DocumentExceedsSequenceContext(details) => {
+                Self::DocumentExceedsSequenceContext(details)
+            }
+            EmbeddingPartialResult::DocumentTokenizedToNothing(details) => {
+                Self::DocumentTokenizedToNothing(details)
+            }
+            EmbeddingPartialResult::Embedding(embedding) => Self::Embedding(embedding),
+        }
+    }
 }
 
 impl StreamableResult for EmbeddingResult {
-    fn is_done(&self) -> bool {
-        matches!(
-            self,
-            Self::Done
-                | Self::EmbeddingsDisabled
-                | Self::Error(_)
-                | Self::EmbeddingRejectedDueToActiveTokenGeneration
-                | Self::ModelNotLoaded(_)
-                | Self::NoEmbeddingsProduced,
-        )
+    type Final = EmbeddingFinalResult;
+    type Partial = EmbeddingPartialResult;
+
+    fn into_step(self) -> ResultStep<EmbeddingPartialResult, EmbeddingFinalResult> {
+        match self {
+            Self::AgentRuntimeFailed(detail) => {
+                ResultStep::Final(EmbeddingFinalResult::AgentRuntimeFailed(detail))
+            }
+            Self::BatchAssemblyFailed(detail) => {
+                ResultStep::Final(EmbeddingFinalResult::BatchAssemblyFailed(detail))
+            }
+            Self::DecodeFailed(detail) => {
+                ResultStep::Final(EmbeddingFinalResult::DecodeFailed(detail))
+            }
+            Self::DocumentExceedsBatchSize(details) => {
+                ResultStep::Partial(EmbeddingPartialResult::DocumentExceedsBatchSize(details))
+            }
+            Self::DocumentExceedsSequenceContext(details) => ResultStep::Partial(
+                EmbeddingPartialResult::DocumentExceedsSequenceContext(details),
+            ),
+            Self::DocumentTokenizedToNothing(details) => {
+                ResultStep::Partial(EmbeddingPartialResult::DocumentTokenizedToNothing(details))
+            }
+            Self::Done => ResultStep::Final(EmbeddingFinalResult::Done),
+            Self::Embedding(embedding) => {
+                ResultStep::Partial(EmbeddingPartialResult::Embedding(embedding))
+            }
+            Self::EmbeddingTooLongForRmsNormalization(detail) => ResultStep::Final(
+                EmbeddingFinalResult::EmbeddingTooLongForRmsNormalization(detail),
+            ),
+            Self::EmbeddingsUnavailable(detail) => {
+                ResultStep::Final(EmbeddingFinalResult::EmbeddingsUnavailable(detail))
+            }
+            Self::InputTokenizationFailed(detail) => {
+                ResultStep::Final(EmbeddingFinalResult::InputTokenizationFailed(detail))
+            }
+            Self::ModelNotLoaded(detail) => {
+                ResultStep::Final(EmbeddingFinalResult::ModelNotLoaded(detail))
+            }
+            Self::NoEmbeddingsProduced => {
+                ResultStep::Final(EmbeddingFinalResult::NoEmbeddingsProduced)
+            }
+            Self::SchedulerUnavailable(detail) => {
+                ResultStep::Final(EmbeddingFinalResult::SchedulerUnavailable(detail))
+            }
+        }
     }
 }
 
@@ -38,60 +127,125 @@ mod tests {
 
     use super::EmbeddingResult;
     use crate::embedding::Embedding;
+    use crate::embedding_document_exceeding_sequence_context_details::EmbeddingDocumentExceedingSequenceContextDetails;
+    use crate::embedding_document_tokenized_to_nothing_details::EmbeddingDocumentTokenizedToNothingDetails;
+    use crate::embedding_final_result::EmbeddingFinalResult;
     use crate::embedding_normalization_method::EmbeddingNormalizationMethod;
+    use crate::embedding_partial_result::EmbeddingPartialResult;
+    use crate::embedding_vector::EmbeddingVector;
     use crate::oversized_embedding_document_details::OversizedEmbeddingDocumentDetails;
-    use crate::streamable_result::StreamableResult;
-
-    #[test]
-    fn done_is_done() {
-        assert!(EmbeddingResult::Done.is_done());
-    }
-
-    #[test]
-    fn error_is_done() {
-        assert!(EmbeddingResult::Error("fail".to_owned()).is_done());
-    }
-
-    #[test]
-    fn embeddings_disabled_is_done() {
-        assert!(EmbeddingResult::EmbeddingsDisabled.is_done());
-    }
-
-    #[test]
-    fn embedding_rejected_due_to_active_token_generation_is_done() {
-        assert!(EmbeddingResult::EmbeddingRejectedDueToActiveTokenGeneration.is_done());
-    }
+    use crate::result_step::ResultStep;
+    use crate::streamable_result::StreamableResult as _;
 
-    #[test]
-    fn model_not_loaded_is_done() {
-        assert!(EmbeddingResult::ModelNotLoaded("err".to_owned()).is_done());
+    fn step_round_trip(embedding_result: EmbeddingResult) -> EmbeddingResult {
+        match embedding_result.into_step() {
+            ResultStep::Final(final_result) => final_result.into(),
+            ResultStep::Partial(partial_result) => partial_result.into(),
+        }
     }
 
     #[test]
-    fn no_embeddings_produced_is_done() {
-        assert!(EmbeddingResult::NoEmbeddingsProduced.is_done());
+    fn every_terminal_result_ends_the_stream_as_its_final_result() {
+        for (terminal_result, expected_final_result) in [
+            (
+                EmbeddingResult::AgentRuntimeFailed("runtime".to_owned()),
+                EmbeddingFinalResult::AgentRuntimeFailed("runtime".to_owned()),
+            ),
+            (
+                EmbeddingResult::BatchAssemblyFailed("batch".to_owned()),
+                EmbeddingFinalResult::BatchAssemblyFailed("batch".to_owned()),
+            ),
+            (
+                EmbeddingResult::DecodeFailed("decode".to_owned()),
+                EmbeddingFinalResult::DecodeFailed("decode".to_owned()),
+            ),
+            (EmbeddingResult::Done, EmbeddingFinalResult::Done),
+            (
+                EmbeddingResult::EmbeddingTooLongForRmsNormalization("rms".to_owned()),
+                EmbeddingFinalResult::EmbeddingTooLongForRmsNormalization("rms".to_owned()),
+            ),
+            (
+                EmbeddingResult::EmbeddingsUnavailable("unavailable".to_owned()),
+                EmbeddingFinalResult::EmbeddingsUnavailable("unavailable".to_owned()),
+            ),
+            (
+                EmbeddingResult::InputTokenizationFailed("tokenization".to_owned()),
+                EmbeddingFinalResult::InputTokenizationFailed("tokenization".to_owned()),
+            ),
+            (
+                EmbeddingResult::ModelNotLoaded("no model".to_owned()),
+                EmbeddingFinalResult::ModelNotLoaded("no model".to_owned()),
+            ),
+            (
+                EmbeddingResult::NoEmbeddingsProduced,
+                EmbeddingFinalResult::NoEmbeddingsProduced,
+            ),
+            (
+                EmbeddingResult::SchedulerUnavailable("stopped".to_owned()),
+                EmbeddingFinalResult::SchedulerUnavailable("stopped".to_owned()),
+            ),
+        ] {
+            assert_eq!(step_round_trip(terminal_result.clone()), terminal_result);
+            assert_eq!(
+                terminal_result.into_step(),
+                ResultStep::Final(expected_final_result)
+            );
+        }
     }
 
     #[test]
-    fn document_exceeds_batch_size_is_not_done() {
-        let result = EmbeddingResult::DocumentExceedsBatchSize(OversizedEmbeddingDocumentDetails {
-            document_tokens: 4096,
-            n_batch: 2048,
-            source_document_id: "huge".to_owned(),
-        });
-
-        assert!(!result.is_done());
-    }
-
-    #[test]
-    fn embedding_is_not_done() {
-        let result = EmbeddingResult::Embedding(Embedding {
-            embedding: vec![1.0],
+    fn every_per_document_result_keeps_the_stream_open() {
+        let embedding = Embedding {
+            embedding: EmbeddingVector::from(vec![1.0]),
             normalization_method: EmbeddingNormalizationMethod::None,
             pooling_type: PoolingType::Mean,
             source_document_id: "doc".to_owned(),
-        });
+        };
+        let oversized_document = OversizedEmbeddingDocumentDetails {
+            document_tokens: 4096,
+            n_batch: 2048,
+            source_document_id: "huge".to_owned(),
+        };
+        let document_exceeding_sequence_context =
+            EmbeddingDocumentExceedingSequenceContextDetails {
+                document_tokens: 9000,
+                sequence_context_size: 8192,
+                source_document_id: "long".to_owned(),
+            };
+        let document_tokenized_to_nothing = EmbeddingDocumentTokenizedToNothingDetails {
+            source_document_id: "empty".to_owned(),
+        };
 
-        assert!(!result.is_done());
+        for (per_document_result, expected_partial_result) in [
+            (
+                EmbeddingResult::DocumentExceedsBatchSize(oversized_document.clone()),
+                EmbeddingPartialResult::DocumentExceedsBatchSize(oversized_document),
+            ),
+            (
+                EmbeddingResult::DocumentExceedsSequenceContext(
+                    document_exceeding_sequence_context.clone(),
+                ),
+                EmbeddingPartialResult::DocumentExceedsSequenceContext(
+                    document_exceeding_sequence_context,
+                ),
+            ),
+            (
+                EmbeddingResult::DocumentTokenizedToNothing(document_tokenized_to_nothing.clone()),
+                EmbeddingPartialResult::DocumentTokenizedToNothing(document_tokenized_to_nothing),
+            ),
+            (
+                EmbeddingResult::Embedding(embedding.clone()),
+                EmbeddingPartialResult::Embedding(embedding),
+            ),
+        ] {
+            assert_eq!(
+                step_round_trip(per_document_result.clone()),
+                per_document_result
+            );
+            assert_eq!(
+                per_document_result.into_step(),
+                ResultStep::Partial(expected_partial_result)
+            );
+        }
     }
 }
diff --git a/paddler_messaging/src/embedding_vector.rs b/paddler_messaging/src/embedding_vector.rs
new file mode 100644
index 000000000..21495ce2d
--- /dev/null
+++ b/paddler_messaging/src/embedding_vector.rs
@@ -0,0 +1,112 @@
+use std::ops::Deref;
+use std::sync::Arc;
+
+use serde::Deserialize;
+use serde::Deserializer;
+use serde::Serialize;
+use serde::Serializer;
+use serde::de::Error as _;
+use serde_json::from_str;
+use serde_json::value::RawValue;
+
+#[derive(Clone, Debug)]
+pub enum EmbeddingVector {
+    Computed(Arc<[f32]>),
+    Received {
+        json: Arc<RawValue>,
+        values: Arc<[f32]>,
+    },
+}
+
+impl Deref for EmbeddingVector {
+    type Target = [f32];
+
+    fn deref(&self) -> &[f32] {
+        match self {
+            Self::Computed(values) | Self::Received { values, .. } => values,
+        }
+    }
+}
+
+impl From<Vec<f32>> for EmbeddingVector {
+    fn from(values: Vec<f32>) -> Self {
+        Self::Computed(values.into())
+    }
+}
+
+impl PartialEq for EmbeddingVector {
+    fn eq(&self, other: &Self) -> bool {
+        **self == **other
+    }
+}
+
+impl Serialize for EmbeddingVector {
+    fn serialize<TSerializer>(
+        &self,
+        serializer: TSerializer,
+    ) -> Result<TSerializer::Ok, TSerializer::Error>
+    where
+        TSerializer: Serializer,
+    {
+        match self {
+            Self::Computed(values) => values.serialize(serializer),
+            Self::Received { json, .. } => json.serialize(serializer),
+        }
+    }
+}
+
+impl<'de> Deserialize<'de> for EmbeddingVector {
+    fn deserialize<TDeserializer>(deserializer: TDeserializer) -> Result<Self, TDeserializer::Error>
+    where
+        TDeserializer: Deserializer<'de>,
+    {
+        let json = Box::<RawValue>::deserialize(deserializer)?;
+        let values = from_str::<Vec<f32>>(json.get()).map_err(TDeserializer::Error::custom)?;
+
+        Ok(Self::Received {
+            json: Arc::from(json),
+            values: values.into(),
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use serde_json::from_str;
+    use serde_json::to_string;
+
+    use super::EmbeddingVector;
+
+    #[test]
+    fn a_received_vector_relays_the_json_it_arrived_as() {
+        let received: EmbeddingVector = from_str("[0.25, 0.750]").unwrap();
+
+        assert_eq!(
+            [
+                to_string(&received).unwrap(),
+                to_string(&received.clone()).unwrap()
+            ],
+            ["[0.25, 0.750]", "[0.25, 0.750]"]
+        );
+    }
+
+    #[test]
+    fn refuses_a_vector_that_is_not_json() {
+        assert_eq!(
+            from_str::<EmbeddingVector>("[0.25,")
+                .unwrap_err()
+                .to_string(),
+            "EOF while parsing a value at line 1 column 6"
+        );
+    }
+
+    #[test]
+    fn refuses_a_vector_of_values_that_are_not_numbers() {
+        assert_eq!(
+            from_str::<EmbeddingVector>(r#"["a"]"#)
+                .unwrap_err()
+                .to_string(),
+            "invalid type: string \"a\", expected f32 at line 1 column 4"
+        );
+    }
+}
diff --git a/paddler_messaging/src/generated_token_final_result.rs b/paddler_messaging/src/generated_token_final_result.rs
new file mode 100644
index 000000000..9a51987a8
--- /dev/null
+++ b/paddler_messaging/src/generated_token_final_result.rs
@@ -0,0 +1,33 @@
+use crate::generation_summary::GenerationSummary;
+use crate::oversized_media_details::OversizedMediaDetails;
+use crate::oversized_prompt_details::OversizedPromptDetails;
+
+#[derive(Debug, Eq, PartialEq)]
+pub enum GeneratedTokenFinalResult {
+    BatchAssemblyFailed(String),
+    ChatTemplateError(String),
+    DecodeFailed(String),
+    DetokenizationFailed(String),
+    Done(GenerationSummary),
+    GrammarIncompatibleWithThinking(String),
+    GrammarInitializationFailed(String),
+    GrammarRejectedModelOutput(String),
+    GrammarSyntaxError(String),
+    ImageDecodingFailed(String),
+    KvCacheClearFailed(String),
+    MediaExceedsMicroBatch(OversizedMediaDetails),
+    MediaMicroBatchCheckFailed(String),
+    ModelNotLoaded(String),
+    MultimodalIngestionFailed(String),
+    MultimodalNotSupported(String),
+    MultimodalTokenizationFailed(String),
+    NoSequenceSlotAvailable(String),
+    PromptExceedsContextSize(OversizedPromptDetails),
+    PromptTokenizationFailed(String),
+    PromptTokenizedToNothing(String),
+    SamplerChainCreationFailed(String),
+    SamplerError(String),
+    SamplingCandidatesExhausted(String),
+    SchedulerUnavailable(String),
+    ToolSchemaInvalid(String),
+}
diff --git a/paddler_messaging/src/generated_token_partial_result.rs b/paddler_messaging/src/generated_token_partial_result.rs
new file mode 100644
index 000000000..c9e1be953
--- /dev/null
+++ b/paddler_messaging/src/generated_token_partial_result.rs
@@ -0,0 +1,14 @@
+use crate::generated_tool_call::GeneratedToolCall;
+use crate::raw_tool_call_tokens::RawToolCallTokens;
+
+#[derive(Debug, Eq, PartialEq)]
+pub enum GeneratedTokenPartialResult {
+    ContentToken(String),
+    ReasoningToken(String),
+    ToolCallParseFailed(String),
+    ToolCallParsed(Vec<GeneratedToolCall>),
+    ToolCallToken(String),
+    ToolCallValidationFailed(Vec<String>),
+    UndeterminableToken(String),
+    UnrecognizedToolCallFormat(RawToolCallTokens),
+}
diff --git a/paddler_messaging/src/generated_token_result.rs b/paddler_messaging/src/generated_token_result.rs
index a1a6a6d23..1765c2b60 100644
--- a/paddler_messaging/src/generated_token_result.rs
+++ b/paddler_messaging/src/generated_token_result.rs
@@ -1,11 +1,14 @@
-use llama_cpp_bindings_types::ParsedToolCall;
 use serde::Deserialize;
 use serde::Serialize;
 
+use crate::generated_token_final_result::GeneratedTokenFinalResult;
+use crate::generated_token_partial_result::GeneratedTokenPartialResult;
+use crate::generated_tool_call::GeneratedToolCall;
 use crate::generation_summary::GenerationSummary;
 use crate::oversized_media_details::OversizedMediaDetails;
 use crate::oversized_prompt_details::OversizedPromptDetails;
 use crate::raw_tool_call_tokens::RawToolCallTokens;
+use crate::result_step::ResultStep;
 use crate::streamable_result::StreamableResult;
 
 #[derive(Clone, Debug, Deserialize, Eq, PartialEq, Serialize)]
@@ -32,19 +35,17 @@ pub enum GeneratedTokenResult {
     NoSequenceSlotAvailable(String),
     PromptExceedsContextSize(OversizedPromptDetails),
     PromptTokenizationFailed(String),
+    PromptTokenizedToNothing(String),
     ReasoningToken(String),
     SamplerChainCreationFailed(String),
     SamplerError(String),
     SamplingCandidatesExhausted(String),
     SchedulerUnavailable(String),
-    SequenceIdOutOfRange(String),
-    TokenGenerationDisabled(String),
     ToolCallParseFailed(String),
-    ToolCallParsed(Vec<ParsedToolCall>),
+    ToolCallParsed(Vec<GeneratedToolCall>),
     ToolCallToken(String),
     ToolCallValidationFailed(Vec<String>),
     ToolSchemaInvalid(String),
-    ToolsSerializationFailed(String),
     UndeterminableToken(String),
     UnrecognizedToolCallFormat(RawToolCallTokens),
 }
@@ -86,39 +87,211 @@ impl GeneratedTokenResult {
     }
 }
 
+impl From<GeneratedTokenFinalResult> for GeneratedTokenResult {
+    fn from(final_result: GeneratedTokenFinalResult) -> Self {
+        match final_result {
+            GeneratedTokenFinalResult::BatchAssemblyFailed(detail) => {
+                Self::BatchAssemblyFailed(detail)
+            }
+            GeneratedTokenFinalResult::ChatTemplateError(detail) => Self::ChatTemplateError(detail),
+            GeneratedTokenFinalResult::DecodeFailed(detail) => Self::DecodeFailed(detail),
+            GeneratedTokenFinalResult::DetokenizationFailed(detail) => {
+                Self::DetokenizationFailed(detail)
+            }
+            GeneratedTokenFinalResult::Done(summary) => Self::Done(summary),
+            GeneratedTokenFinalResult::GrammarIncompatibleWithThinking(detail) => {
+                Self::GrammarIncompatibleWithThinking(detail)
+            }
+            GeneratedTokenFinalResult::GrammarInitializationFailed(detail) => {
+                Self::GrammarInitializationFailed(detail)
+            }
+            GeneratedTokenFinalResult::GrammarRejectedModelOutput(detail) => {
+                Self::GrammarRejectedModelOutput(detail)
+            }
+            GeneratedTokenFinalResult::GrammarSyntaxError(detail) => {
+                Self::GrammarSyntaxError(detail)
+            }
+            GeneratedTokenFinalResult::ImageDecodingFailed(detail) => {
+                Self::ImageDecodingFailed(detail)
+            }
+            GeneratedTokenFinalResult::KvCacheClearFailed(detail) => {
+                Self::KvCacheClearFailed(detail)
+            }
+            GeneratedTokenFinalResult::MediaExceedsMicroBatch(details) => {
+                Self::MediaExceedsMicroBatch(details)
+            }
+            GeneratedTokenFinalResult::MediaMicroBatchCheckFailed(detail) => {
+                Self::MediaMicroBatchCheckFailed(detail)
+            }
+            GeneratedTokenFinalResult::ModelNotLoaded(detail) => Self::ModelNotLoaded(detail),
+            GeneratedTokenFinalResult::MultimodalIngestionFailed(detail) => {
+                Self::MultimodalIngestionFailed(detail)
+            }
+            GeneratedTokenFinalResult::MultimodalNotSupported(detail) => {
+                Self::MultimodalNotSupported(detail)
+            }
+            GeneratedTokenFinalResult::MultimodalTokenizationFailed(detail) => {
+                Self::MultimodalTokenizationFailed(detail)
+            }
+            GeneratedTokenFinalResult::NoSequenceSlotAvailable(detail) => {
+                Self::NoSequenceSlotAvailable(detail)
+            }
+            GeneratedTokenFinalResult::PromptExceedsContextSize(details) => {
+                Self::PromptExceedsContextSize(details)
+            }
+            GeneratedTokenFinalResult::PromptTokenizationFailed(detail) => {
+                Self::PromptTokenizationFailed(detail)
+            }
+            GeneratedTokenFinalResult::PromptTokenizedToNothing(detail) => {
+                Self::PromptTokenizedToNothing(detail)
+            }
+            GeneratedTokenFinalResult::SamplerChainCreationFailed(detail) => {
+                Self::SamplerChainCreationFailed(detail)
+            }
+            GeneratedTokenFinalResult::SamplerError(detail) => Self::SamplerError(detail),
+            GeneratedTokenFinalResult::SamplingCandidatesExhausted(detail) => {
+                Self::SamplingCandidatesExhausted(detail)
+            }
+            GeneratedTokenFinalResult::SchedulerUnavailable(detail) => {
+                Self::SchedulerUnavailable(detail)
+            }
+            GeneratedTokenFinalResult::ToolSchemaInvalid(detail) => Self::ToolSchemaInvalid(detail),
+        }
+    }
+}
+
+impl From<GeneratedTokenPartialResult> for GeneratedTokenResult {
+    fn from(partial_result: GeneratedTokenPartialResult) -> Self {
+        match partial_result {
+            GeneratedTokenPartialResult::ContentToken(text) => Self::ContentToken(text),
+            GeneratedTokenPartialResult::ReasoningToken(text) => Self::ReasoningToken(text),
+            GeneratedTokenPartialResult::ToolCallParseFailed(detail) => {
+                Self::ToolCallParseFailed(detail)
+            }
+            GeneratedTokenPartialResult::ToolCallParsed(parsed_calls) => {
+                Self::ToolCallParsed(parsed_calls)
+            }
+            GeneratedTokenPartialResult::ToolCallToken(text) => Self::ToolCallToken(text),
+            GeneratedTokenPartialResult::ToolCallValidationFailed(validation_errors) => {
+                Self::ToolCallValidationFailed(validation_errors)
+            }
+            GeneratedTokenPartialResult::UndeterminableToken(text) => {
+                Self::UndeterminableToken(text)
+            }
+            GeneratedTokenPartialResult::UnrecognizedToolCallFormat(raw_tokens) => {
+                Self::UnrecognizedToolCallFormat(raw_tokens)
+            }
+        }
+    }
+}
+
 impl StreamableResult for GeneratedTokenResult {
-    fn is_done(&self) -> bool {
-        matches!(
-            self,
-            Self::BatchAssemblyFailed(_)
-                | Self::ChatTemplateError(_)
-                | Self::DecodeFailed(_)
-                | Self::DetokenizationFailed(_)
-                | Self::Done(_)
-                | Self::GrammarIncompatibleWithThinking(_)
-                | Self::GrammarInitializationFailed(_)
-                | Self::GrammarRejectedModelOutput(_)
-                | Self::GrammarSyntaxError(_)
-                | Self::ImageDecodingFailed(_)
-                | Self::KvCacheClearFailed(_)
-                | Self::MediaExceedsMicroBatch(_)
-                | Self::MediaMicroBatchCheckFailed(_)
-                | Self::ModelNotLoaded(_)
-                | Self::MultimodalIngestionFailed(_)
-                | Self::MultimodalNotSupported(_)
-                | Self::MultimodalTokenizationFailed(_)
-                | Self::NoSequenceSlotAvailable(_)
-                | Self::PromptExceedsContextSize(_)
-                | Self::PromptTokenizationFailed(_)
-                | Self::SamplerChainCreationFailed(_)
-                | Self::SamplerError(_)
-                | Self::SamplingCandidatesExhausted(_)
-                | Self::SchedulerUnavailable(_)
-                | Self::SequenceIdOutOfRange(_)
-                | Self::TokenGenerationDisabled(_)
-                | Self::ToolSchemaInvalid(_)
-                | Self::ToolsSerializationFailed(_)
-        )
+    type Final = GeneratedTokenFinalResult;
+    type Partial = GeneratedTokenPartialResult;
+
+    fn into_step(self) -> ResultStep<GeneratedTokenPartialResult, GeneratedTokenFinalResult> {
+        match self {
+            Self::BatchAssemblyFailed(detail) => {
+                ResultStep::Final(GeneratedTokenFinalResult::BatchAssemblyFailed(detail))
+            }
+            Self::ChatTemplateError(detail) => {
+                ResultStep::Final(GeneratedTokenFinalResult::ChatTemplateError(detail))
+            }
+            Self::ContentToken(text) => {
+                ResultStep::Partial(GeneratedTokenPartialResult::ContentToken(text))
+            }
+            Self::DecodeFailed(detail) => {
+                ResultStep::Final(GeneratedTokenFinalResult::DecodeFailed(detail))
+            }
+            Self::DetokenizationFailed(detail) => {
+                ResultStep::Final(GeneratedTokenFinalResult::DetokenizationFailed(detail))
+            }
+            Self::Done(summary) => ResultStep::Final(GeneratedTokenFinalResult::Done(summary)),
+            Self::GrammarIncompatibleWithThinking(detail) => ResultStep::Final(
+                GeneratedTokenFinalResult::GrammarIncompatibleWithThinking(detail),
+            ),
+            Self::GrammarInitializationFailed(detail) => ResultStep::Final(
+                GeneratedTokenFinalResult::GrammarInitializationFailed(detail),
+            ),
+            Self::GrammarRejectedModelOutput(detail) => ResultStep::Final(
+                GeneratedTokenFinalResult::GrammarRejectedModelOutput(detail),
+            ),
+            Self::GrammarSyntaxError(detail) => {
+                ResultStep::Final(GeneratedTokenFinalResult::GrammarSyntaxError(detail))
+            }
+            Self::ImageDecodingFailed(detail) => {
+                ResultStep::Final(GeneratedTokenFinalResult::ImageDecodingFailed(detail))
+            }
+            Self::KvCacheClearFailed(detail) => {
+                ResultStep::Final(GeneratedTokenFinalResult::KvCacheClearFailed(detail))
+            }
+            Self::MediaExceedsMicroBatch(details) => {
+                ResultStep::Final(GeneratedTokenFinalResult::MediaExceedsMicroBatch(details))
+            }
+            Self::MediaMicroBatchCheckFailed(detail) => ResultStep::Final(
+                GeneratedTokenFinalResult::MediaMicroBatchCheckFailed(detail),
+            ),
+            Self::ModelNotLoaded(detail) => {
+                ResultStep::Final(GeneratedTokenFinalResult::ModelNotLoaded(detail))
+            }
+            Self::MultimodalIngestionFailed(detail) => {
+                ResultStep::Final(GeneratedTokenFinalResult::MultimodalIngestionFailed(detail))
+            }
+            Self::MultimodalNotSupported(detail) => {
+                ResultStep::Final(GeneratedTokenFinalResult::MultimodalNotSupported(detail))
+            }
+            Self::MultimodalTokenizationFailed(detail) => ResultStep::Final(
+                GeneratedTokenFinalResult::MultimodalTokenizationFailed(detail),
+            ),
+            Self::NoSequenceSlotAvailable(detail) => {
+                ResultStep::Final(GeneratedTokenFinalResult::NoSequenceSlotAvailable(detail))
+            }
+            Self::PromptExceedsContextSize(details) => {
+                ResultStep::Final(GeneratedTokenFinalResult::PromptExceedsContextSize(details))
+            }
+            Self::PromptTokenizationFailed(detail) => {
+                ResultStep::Final(GeneratedTokenFinalResult::PromptTokenizationFailed(detail))
+            }
+            Self::PromptTokenizedToNothing(detail) => {
+                ResultStep::Final(GeneratedTokenFinalResult::PromptTokenizedToNothing(detail))
+            }
+            Self::ReasoningToken(text) => {
+                ResultStep::Partial(GeneratedTokenPartialResult::ReasoningToken(text))
+            }
+            Self::SamplerChainCreationFailed(detail) => ResultStep::Final(
+                GeneratedTokenFinalResult::SamplerChainCreationFailed(detail),
+            ),
+            Self::SamplerError(detail) => {
+                ResultStep::Final(GeneratedTokenFinalResult::SamplerError(detail))
+            }
+            Self::SamplingCandidatesExhausted(detail) => ResultStep::Final(
+                GeneratedTokenFinalResult::SamplingCandidatesExhausted(detail),
+            ),
+            Self::SchedulerUnavailable(detail) => {
+                ResultStep::Final(GeneratedTokenFinalResult::SchedulerUnavailable(detail))
+            }
+            Self::ToolCallParseFailed(detail) => {
+                ResultStep::Partial(GeneratedTokenPartialResult::ToolCallParseFailed(detail))
+            }
+            Self::ToolCallParsed(parsed_calls) => {
+                ResultStep::Partial(GeneratedTokenPartialResult::ToolCallParsed(parsed_calls))
+            }
+            Self::ToolCallToken(text) => {
+                ResultStep::Partial(GeneratedTokenPartialResult::ToolCallToken(text))
+            }
+            Self::ToolCallValidationFailed(validation_errors) => ResultStep::Partial(
+                GeneratedTokenPartialResult::ToolCallValidationFailed(validation_errors),
+            ),
+            Self::ToolSchemaInvalid(detail) => {
+                ResultStep::Final(GeneratedTokenFinalResult::ToolSchemaInvalid(detail))
+            }
+            Self::UndeterminableToken(text) => {
+                ResultStep::Partial(GeneratedTokenPartialResult::UndeterminableToken(text))
+            }
+            Self::UnrecognizedToolCallFormat(raw_tokens) => ResultStep::Partial(
+                GeneratedTokenPartialResult::UnrecognizedToolCallFormat(raw_tokens),
+            ),
+        }
     }
 }
 
@@ -132,7 +305,23 @@ mod tests {
     use crate::oversized_media_details::OversizedMediaDetails;
     use crate::oversized_prompt_details::OversizedPromptDetails;
     use crate::raw_tool_call_tokens::RawToolCallTokens;
-    use crate::streamable_result::StreamableResult;
+    use crate::result_step::ResultStep;
+    use crate::streamable_result::StreamableResult as _;
+
+    fn keeps_the_stream_open(token_result: &GeneratedTokenResult) -> bool {
+        match token_result.clone().into_step() {
+            ResultStep::Final(final_result) => {
+                assert_eq!(&GeneratedTokenResult::from(final_result), token_result);
+
+                false
+            }
+            ResultStep::Partial(partial_result) => {
+                assert_eq!(&GeneratedTokenResult::from(partial_result), token_result);
+
+                true
+            }
+        }
+    }
 
     #[test]
     fn every_terminal_result_ends_the_stream_without_a_token() {
@@ -166,16 +355,14 @@ mod tests {
                 sequence_context_size: 8192,
             }),
             GeneratedTokenResult::PromptTokenizationFailed("failure".to_owned()),
+            GeneratedTokenResult::PromptTokenizedToNothing("failure".to_owned()),
             GeneratedTokenResult::SamplerChainCreationFailed("failure".to_owned()),
             GeneratedTokenResult::SamplerError("failure".to_owned()),
             GeneratedTokenResult::SamplingCandidatesExhausted("failure".to_owned()),
             GeneratedTokenResult::SchedulerUnavailable("failure".to_owned()),
-            GeneratedTokenResult::SequenceIdOutOfRange("failure".to_owned()),
-            GeneratedTokenResult::TokenGenerationDisabled("failure".to_owned()),
             GeneratedTokenResult::ToolSchemaInvalid("failure".to_owned()),
-            GeneratedTokenResult::ToolsSerializationFailed("failure".to_owned()),
         ] {
-            assert!(terminal_result.is_done());
+            assert!(!keeps_the_stream_open(&terminal_result));
             assert!(!terminal_result.is_token());
             assert_eq!(terminal_result.token_text(), None);
         }
@@ -189,7 +376,7 @@ mod tests {
             GeneratedTokenResult::ToolCallToken("piece".to_owned()),
             GeneratedTokenResult::UndeterminableToken("piece".to_owned()),
         ] {
-            assert!(!token_result.is_done());
+            assert!(keeps_the_stream_open(&token_result));
             assert!(token_result.is_token());
             assert_eq!(token_result.token_text(), Some("piece"));
         }
@@ -199,7 +386,7 @@ mod tests {
     fn tool_call_parsed_is_not_done() {
         let event = GeneratedTokenResult::ToolCallParsed(vec![]);
 
-        assert!(!event.is_done());
+        assert!(keeps_the_stream_open(&event));
         assert!(event.is_tool_call_parsed());
         assert!(!event.is_tool_call_failure());
     }
@@ -208,7 +395,7 @@ mod tests {
     fn tool_call_parse_failed_is_failure_but_not_done() {
         let event = GeneratedTokenResult::ToolCallParseFailed("oops".to_owned());
 
-        assert!(!event.is_done());
+        assert!(keeps_the_stream_open(&event));
         assert!(!event.is_tool_call_parsed());
         assert!(event.is_tool_call_failure());
     }
@@ -217,7 +404,7 @@ mod tests {
     fn tool_call_validation_failed_is_failure_but_not_done() {
         let event = GeneratedTokenResult::ToolCallValidationFailed(vec!["missing".to_owned()]);
 
-        assert!(!event.is_done());
+        assert!(keeps_the_stream_open(&event));
         assert!(!event.is_tool_call_parsed());
         assert!(event.is_tool_call_failure());
     }
@@ -229,7 +416,7 @@ mod tests {
             ffi_error_message: "parser bailed".to_owned(),
         });
 
-        assert!(!event.is_done());
+        assert!(keeps_the_stream_open(&event));
         assert!(!event.is_token());
         assert!(event.token_text().is_none());
         assert!(!event.is_tool_call_parsed());
diff --git a/paddler_messaging/src/generated_tool_call.rs b/paddler_messaging/src/generated_tool_call.rs
new file mode 100644
index 000000000..421a02f01
--- /dev/null
+++ b/paddler_messaging/src/generated_tool_call.rs
@@ -0,0 +1,84 @@
+use llama_cpp_bindings_types::ParsedToolCall;
+use llama_cpp_bindings_types::ParsedToolCallId;
+use llama_cpp_bindings_types::ToolCallArguments;
+use serde::Deserialize;
+use serde::Serialize;
+
+#[derive(Clone, Debug, Deserialize, Eq, PartialEq, Serialize)]
+#[serde(deny_unknown_fields)]
+pub struct GeneratedToolCall {
+    pub arguments: ToolCallArguments,
+    pub id: String,
+    pub index: usize,
+    pub name: String,
+}
+
+impl GeneratedToolCall {
+    #[must_use]
+    pub fn at_index(
+        ParsedToolCall {
+            arguments,
+            id,
+            name,
+        }: ParsedToolCall,
+        index: usize,
+    ) -> Self {
+        Self {
+            arguments,
+            id: match id {
+                ParsedToolCallId::ModelProvided(model_provided_id) => {
+                    model_provided_id.as_str().to_owned()
+                }
+                ParsedToolCallId::NotProvidedByModel => format!("call_{index}"),
+            },
+            index,
+            name,
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use llama_cpp_bindings_types::ModelProvidedToolCallId;
+    use llama_cpp_bindings_types::ParsedToolCall;
+    use llama_cpp_bindings_types::ParsedToolCallId;
+    use llama_cpp_bindings_types::ToolCallArguments;
+    use serde_json::json;
+
+    use super::GeneratedToolCall;
+
+    fn weather_call(id: ParsedToolCallId) -> ParsedToolCall {
+        ParsedToolCall::new(
+            id,
+            "get_weather".to_owned(),
+            ToolCallArguments::ValidJson(json!({"location": "Paris"})),
+        )
+    }
+
+    #[test]
+    fn a_tool_call_the_model_gave_an_id_keeps_that_id() {
+        assert_eq!(
+            GeneratedToolCall::at_index(
+                weather_call(ParsedToolCallId::ModelProvided(
+                    ModelProvidedToolCallId::try_from("abc123def".to_owned()).unwrap(),
+                )),
+                3,
+            )
+            .id,
+            "abc123def"
+        );
+    }
+
+    #[test]
+    fn a_tool_call_the_model_gave_no_id_is_named_after_its_index_in_the_generation() {
+        assert_eq!(
+            GeneratedToolCall::at_index(weather_call(ParsedToolCallId::NotProvidedByModel), 3),
+            GeneratedToolCall {
+                arguments: ToolCallArguments::ValidJson(json!({"location": "Paris"})),
+                id: "call_3".to_owned(),
+                index: 3,
+                name: "get_weather".to_owned(),
+            }
+        );
+    }
+}
diff --git a/paddler_messaging/src/huggingface_model_reference.rs b/paddler_messaging/src/huggingface_model_reference.rs
deleted file mode 100644
index 8ac1ba48c..000000000
--- a/paddler_messaging/src/huggingface_model_reference.rs
+++ /dev/null
@@ -1,49 +0,0 @@
-use serde::Deserialize;
-use serde::Serialize;
-
-use crate::agent_issue_params::model_path::ModelPath;
-
-#[derive(Clone, Debug, Deserialize, Eq, PartialEq, Serialize)]
-#[serde(deny_unknown_fields)]
-pub struct HuggingFaceModelReference {
-    pub filename: String,
-    pub repo_id: String,
-    pub revision: String,
-}
-
-impl HuggingFaceModelReference {
-    #[must_use]
-    pub fn model_path(&self) -> ModelPath {
-        let Self {
-            filename,
-            repo_id,
-            revision,
-        } = self;
-
-        ModelPath {
-            model_path: format!("{repo_id}/{revision}/{filename}"),
-        }
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use super::HuggingFaceModelReference;
-    use crate::agent_issue_params::model_path::ModelPath;
-
-    #[test]
-    fn names_the_model_path_after_its_repository_revision_and_filename() {
-        let reference = HuggingFaceModelReference {
-            filename: "model.gguf".to_owned(),
-            repo_id: "owner/repo".to_owned(),
-            revision: "v2".to_owned(),
-        };
-
-        assert_eq!(
-            reference.model_path(),
-            ModelPath {
-                model_path: "owner/repo/v2/model.gguf".to_owned(),
-            }
-        );
-    }
-}
diff --git a/paddler_messaging/src/inference_capability.rs b/paddler_messaging/src/inference_capability.rs
new file mode 100644
index 000000000..97e4e6937
--- /dev/null
+++ b/paddler_messaging/src/inference_capability.rs
@@ -0,0 +1,10 @@
+use serde::Deserialize;
+use serde::Serialize;
+
+#[derive(Clone, Copy, Debug, Deserialize, Eq, Hash, Ord, PartialEq, PartialOrd, Serialize)]
+#[serde(deny_unknown_fields)]
+pub enum InferenceCapability {
+    Decisions,
+    Embeddings,
+    TextGeneration,
+}
diff --git a/paddler_messaging/src/inference_client/message.rs b/paddler_messaging/src/inference_client/message.rs
index 2c801d266..45c6d4243 100644
--- a/paddler_messaging/src/inference_client/message.rs
+++ b/paddler_messaging/src/inference_client/message.rs
@@ -1,18 +1,16 @@
 use serde::Deserialize;
 use serde::Serialize;
 
-use super::notification::Notification;
 use super::response::Response;
 use crate::jsonrpc::error::Error;
 use crate::jsonrpc::error_envelope::ErrorEnvelope;
 use crate::jsonrpc::response_envelope::ResponseEnvelope;
 use crate::rpc_message::RpcMessage;
 
-#[derive(Debug, Deserialize, Serialize)]
+#[derive(Debug, Deserialize, PartialEq, Serialize)]
 #[serde(deny_unknown_fields)]
 pub enum Message {
     Error(ErrorEnvelope<Error>),
-    Notification(Notification),
     Response(ResponseEnvelope<Response>),
 }
 
diff --git a/paddler_messaging/src/inference_client/mod.rs b/paddler_messaging/src/inference_client/mod.rs
index 9c7e6cab5..3479387e0 100644
--- a/paddler_messaging/src/inference_client/mod.rs
+++ b/paddler_messaging/src/inference_client/mod.rs
@@ -1,4 +1,3 @@
 pub mod identified_message;
 pub mod message;
-pub mod notification;
 pub mod response;
diff --git a/paddler_messaging/src/inference_client/notification.rs b/paddler_messaging/src/inference_client/notification.rs
deleted file mode 100644
index 11187b9fe..000000000
--- a/paddler_messaging/src/inference_client/notification.rs
+++ /dev/null
@@ -1,9 +0,0 @@
-use serde::Deserialize;
-use serde::Serialize;
-
-#[derive(Clone, Debug, Deserialize, Eq, PartialEq, Serialize)]
-#[serde(deny_unknown_fields)]
-pub enum Notification {
-    TokenGenerationDisabled,
-    TokenGenerationEnabled,
-}
diff --git a/paddler_messaging/src/inference_client/response.rs b/paddler_messaging/src/inference_client/response.rs
index d4224ee1f..251a706fb 100644
--- a/paddler_messaging/src/inference_client/response.rs
+++ b/paddler_messaging/src/inference_client/response.rs
@@ -1,16 +1,51 @@
 use serde::Deserialize;
 use serde::Serialize;
 
+use crate::decision_result::DecisionResult;
 use crate::embedding_result::EmbeddingResult;
 use crate::generated_token_result::GeneratedTokenResult;
+use crate::result_step::ResultStep;
+use crate::streamable_result::StreamableResult;
 
-#[derive(Debug, Deserialize, Serialize)]
+fn response_step<TResult>(result: TResult) -> ResultStep<Response, Response>
+where
+    TResult: StreamableResult + Into<Response>,
+{
+    match result.into_step() {
+        ResultStep::Final(final_result) => {
+            ResultStep::Final(Into::<TResult>::into(final_result).into())
+        }
+        ResultStep::Partial(partial_result) => {
+            ResultStep::Partial(Into::<TResult>::into(partial_result).into())
+        }
+    }
+}
+
+#[derive(Debug, Deserialize, PartialEq, Serialize)]
 #[serde(deny_unknown_fields)]
 pub enum Response {
+    Decision(DecisionResult),
     Embedding(EmbeddingResult),
     GeneratedToken(GeneratedTokenResult),
 }
 
+impl Response {
+    #[must_use]
+    pub fn into_step(self) -> ResultStep<Self, Self> {
+        match self {
+            Self::Decision(decision_result) => response_step(decision_result),
+            Self::Embedding(embedding_result) => response_step(embedding_result),
+            Self::GeneratedToken(generated_token_result) => response_step(generated_token_result),
+        }
+    }
+}
+
+impl From<DecisionResult> for Response {
+    fn from(result: DecisionResult) -> Self {
+        Self::Decision(result)
+    }
+}
+
 impl From<EmbeddingResult> for Response {
     fn from(result: EmbeddingResult) -> Self {
         Self::Embedding(result)
diff --git a/paddler_messaging/src/jsonrpc/error_envelope.rs b/paddler_messaging/src/jsonrpc/error_envelope.rs
index 601a2aaad..bd45fe4b6 100644
--- a/paddler_messaging/src/jsonrpc/error_envelope.rs
+++ b/paddler_messaging/src/jsonrpc/error_envelope.rs
@@ -1,7 +1,7 @@
 use serde::Deserialize;
 use serde::Serialize;
 
-#[derive(Debug, Deserialize, Serialize)]
+#[derive(Debug, Deserialize, Eq, PartialEq, Serialize)]
 #[serde(deny_unknown_fields)]
 pub struct ErrorEnvelope<TRequest> {
     pub request_id: String,
diff --git a/paddler_messaging/src/jsonrpc/response_envelope.rs b/paddler_messaging/src/jsonrpc/response_envelope.rs
index 00818344b..e4fbbc92b 100644
--- a/paddler_messaging/src/jsonrpc/response_envelope.rs
+++ b/paddler_messaging/src/jsonrpc/response_envelope.rs
@@ -1,7 +1,7 @@
 use serde::Deserialize;
 use serde::Serialize;
 
-#[derive(Debug, Deserialize, Serialize)]
+#[derive(Debug, Deserialize, Eq, PartialEq, Serialize)]
 #[serde(deny_unknown_fields)]
 pub struct ResponseEnvelope<TResponse> {
     pub generated_by: Option<String>,
diff --git a/paddler_messaging/src/lib.rs b/paddler_messaging/src/lib.rs
index e81848a51..ed4b9a516 100644
--- a/paddler_messaging/src/lib.rs
+++ b/paddler_messaging/src/lib.rs
@@ -1,35 +1,61 @@
 pub mod agent_controller_pool_snapshot;
 pub mod agent_controller_snapshot;
+pub mod agent_decisions_serving;
+pub mod agent_decisions_settings;
 pub mod agent_desired_model;
 pub mod agent_desired_state;
 pub mod agent_issue;
 pub mod agent_issue_params;
+pub mod agent_runtime_status;
 pub mod agent_state_application_status;
 pub mod agent_status;
 pub mod api_path;
+pub mod assistant_role;
 pub mod atomic_value;
-pub mod balancer_connection;
 pub mod balancer_desired_state;
+pub mod balancer_text_generation_settings;
 pub mod buffered_request_manager_snapshot;
 pub mod chat_template;
+pub mod chat_template_content_form;
 pub mod chat_template_conversation;
 pub mod chat_template_message;
 pub mod chat_template_message_content;
 pub mod chat_template_message_content_part;
+pub mod chat_template_source;
+pub mod chat_template_tool_call;
+pub mod chat_template_tool_call_function;
 pub mod conversation_history;
 pub mod conversation_message;
 pub mod conversation_message_content;
 pub mod conversation_message_content_part;
+pub mod conversation_tool_call;
+pub mod decision_answer;
+pub mod decision_final_result;
+pub mod decision_question;
+pub mod decision_result;
+pub mod decision_settings;
+pub mod decision_slots_for_leading_questions;
+pub mod decision_slots_minimum;
+pub mod decision_summary;
+pub mod embedded_chat_template_key;
 pub mod embedding;
+pub mod embedding_document_exceeding_sequence_context_details;
+pub mod embedding_document_tokenized_to_nothing_details;
+pub mod embedding_final_result;
 pub mod embedding_input_document;
 pub mod embedding_normalization_method;
+pub mod embedding_partial_result;
 pub mod embedding_result;
+pub mod embedding_vector;
+pub mod generated_token_final_result;
+pub mod generated_token_partial_result;
 pub mod generated_token_result;
+pub mod generated_tool_call;
 pub mod generation_finish;
 pub mod generation_summary;
 pub mod grammar_constraint;
-pub mod huggingface_model_reference;
 pub mod image_url;
+pub mod inference_capability;
 pub mod inference_client;
 pub mod inference_server;
 pub mod jsonrpc;
@@ -37,16 +63,27 @@ pub mod management_socket;
 pub mod media_marker;
 pub mod model_download_status;
 pub mod model_metadata;
+pub mod multimodal_settings;
+pub mod oversized_decision_details;
 pub mod oversized_embedding_document_details;
 pub mod oversized_media_details;
 pub mod oversized_prompt_details;
+pub mod ping_interval;
+pub mod ping_liveness;
+pub mod ping_turn;
 pub mod produces_snapshot;
 pub mod raw_tool_call_tokens;
 pub mod request_params;
 pub mod request_params_validation_error;
+pub mod reserved_request;
+pub mod result_step;
 pub mod rpc_message;
+pub mod served_decision_settings;
 pub mod slot_aggregated_status_snapshot;
 pub mod streamable_result;
 pub mod subscribes_to_updates;
-pub mod url_model_reference;
+pub mod system_role;
+pub mod tool_role;
+pub mod unanswered_pings_tolerated;
+pub mod user_role;
 pub mod validates;
diff --git a/paddler_messaging/src/management_socket/agent/request.rs b/paddler_messaging/src/management_socket/agent/request.rs
index e5829d32a..676cd8a6e 100644
--- a/paddler_messaging/src/management_socket/agent/request.rs
+++ b/paddler_messaging/src/management_socket/agent/request.rs
@@ -1,37 +1,52 @@
 use serde::Deserialize;
 use serde::Serialize;
 
-use crate::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
-use crate::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
 use crate::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
 use crate::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters_schema::validated_parameters_schema::ValidatedParametersSchema;
+use crate::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
+use crate::request_params::decide_params::DecideParams;
+use crate::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
+use crate::reserved_request::ReservedRequest;
 
 #[derive(Deserialize, Serialize)]
 #[serde(deny_unknown_fields)]
 pub enum Request {
     ContinueFromConversationHistory(
-        ContinueFromConversationHistoryParams<ValidatedParametersSchema>,
+        ReservedRequest<ContinueFromConversationHistoryParams<ValidatedParametersSchema>>,
     ),
-    ContinueFromRawPrompt(ContinueFromRawPromptParams),
-    GenerateEmbeddingBatch(GenerateEmbeddingBatchParams),
+    ContinueFromRawPrompt(ReservedRequest<ContinueFromRawPromptParams>),
+    Decide(ReservedRequest<DecideParams>),
+    GenerateEmbeddingBatch(ReservedRequest<GenerateEmbeddingBatchParams>),
     GetChatTemplateOverride,
     GetModelMetadata,
 }
 
-impl From<ContinueFromConversationHistoryParams<ValidatedParametersSchema>> for Request {
-    fn from(params: ContinueFromConversationHistoryParams<ValidatedParametersSchema>) -> Self {
-        Self::ContinueFromConversationHistory(params)
+impl From<ReservedRequest<ContinueFromConversationHistoryParams<ValidatedParametersSchema>>>
+    for Request
+{
+    fn from(
+        reserved_request: ReservedRequest<
+            ContinueFromConversationHistoryParams<ValidatedParametersSchema>,
+        >,
+    ) -> Self {
+        Self::ContinueFromConversationHistory(reserved_request)
+    }
+}
+
+impl From<ReservedRequest<ContinueFromRawPromptParams>> for Request {
+    fn from(reserved_request: ReservedRequest<ContinueFromRawPromptParams>) -> Self {
+        Self::ContinueFromRawPrompt(reserved_request)
     }
 }
 
-impl From<ContinueFromRawPromptParams> for Request {
-    fn from(params: ContinueFromRawPromptParams) -> Self {
-        Self::ContinueFromRawPrompt(params)
+impl From<ReservedRequest<DecideParams>> for Request {
+    fn from(reserved_request: ReservedRequest<DecideParams>) -> Self {
+        Self::Decide(reserved_request)
     }
 }
 
-impl From<GenerateEmbeddingBatchParams> for Request {
-    fn from(params: GenerateEmbeddingBatchParams) -> Self {
-        Self::GenerateEmbeddingBatch(params)
+impl From<ReservedRequest<GenerateEmbeddingBatchParams>> for Request {
+    fn from(reserved_request: ReservedRequest<GenerateEmbeddingBatchParams>) -> Self {
+        Self::GenerateEmbeddingBatch(reserved_request)
     }
 }
diff --git a/paddler_messaging/src/management_socket/agent/response.rs b/paddler_messaging/src/management_socket/agent/response.rs
index 76651c9a6..5625ec71d 100644
--- a/paddler_messaging/src/management_socket/agent/response.rs
+++ b/paddler_messaging/src/management_socket/agent/response.rs
@@ -2,14 +2,16 @@ use serde::Deserialize;
 use serde::Serialize;
 
 use crate::chat_template::ChatTemplate;
+use crate::decision_result::DecisionResult;
 use crate::embedding_result::EmbeddingResult;
 use crate::generated_token_result::GeneratedTokenResult;
 use crate::model_metadata::ModelMetadata;
 
-#[derive(Debug, Deserialize, Serialize)]
+#[derive(Debug, Deserialize, PartialEq, Serialize)]
 #[serde(deny_unknown_fields)]
 pub enum Response {
     ChatTemplateOverride(Option<ChatTemplate>),
+    Decision(DecisionResult),
     Embedding(EmbeddingResult),
     GeneratedToken(GeneratedTokenResult),
     ModelMetadata(Option<ModelMetadata>),
@@ -21,6 +23,12 @@ impl From<Option<ChatTemplate>> for Response {
     }
 }
 
+impl From<DecisionResult> for Response {
+    fn from(decision_result: DecisionResult) -> Self {
+        Self::Decision(decision_result)
+    }
+}
+
 impl From<EmbeddingResult> for Response {
     fn from(embedding_result: EmbeddingResult) -> Self {
         Self::Embedding(embedding_result)
@@ -52,9 +60,10 @@ mod tests {
     fn chat_template_payload(response: &Response) -> Option<&ChatTemplate> {
         match response {
             Response::ChatTemplateOverride(chat_template) => chat_template.as_ref(),
-            Response::Embedding(_) | Response::GeneratedToken(_) | Response::ModelMetadata(_) => {
-                None
-            }
+            Response::Decision(_)
+            | Response::Embedding(_)
+            | Response::GeneratedToken(_)
+            | Response::ModelMetadata(_) => None,
         }
     }
 
@@ -62,6 +71,7 @@ mod tests {
         match response {
             Response::ModelMetadata(model_metadata) => model_metadata.as_ref(),
             Response::ChatTemplateOverride(_)
+            | Response::Decision(_)
             | Response::Embedding(_)
             | Response::GeneratedToken(_) => None,
         }
diff --git a/paddler_messaging/src/management_socket/balancer/message.rs b/paddler_messaging/src/management_socket/balancer/message.rs
index d7a9902c8..64cc580f0 100644
--- a/paddler_messaging/src/management_socket/balancer/message.rs
+++ b/paddler_messaging/src/management_socket/balancer/message.rs
@@ -6,7 +6,7 @@ use crate::jsonrpc::response_envelope::ResponseEnvelope;
 use crate::management_socket::agent::response::Response;
 use crate::rpc_message::RpcMessage;
 
-#[derive(Debug, Deserialize, Serialize)]
+#[derive(Debug, Deserialize, PartialEq, Serialize)]
 #[serde(deny_unknown_fields)]
 pub enum Message {
     Notification(Notification),
diff --git a/paddler_messaging/src/management_socket/balancer/notification.rs b/paddler_messaging/src/management_socket/balancer/notification.rs
index 9a52063a6..45acc5025 100644
--- a/paddler_messaging/src/management_socket/balancer/notification.rs
+++ b/paddler_messaging/src/management_socket/balancer/notification.rs
@@ -4,7 +4,7 @@ use serde::Serialize;
 use super::notification_params::register_agent_params::RegisterAgentParams;
 use super::notification_params::update_agent_status_params::UpdateAgentStatusParams;
 
-#[derive(Debug, Deserialize, Serialize)]
+#[derive(Debug, Deserialize, Eq, PartialEq, Serialize)]
 #[serde(deny_unknown_fields)]
 pub enum Notification {
     DeregisterAgent,
diff --git a/paddler_messaging/src/model_download_status.rs b/paddler_messaging/src/model_download_status.rs
index 0d900a215..b004b74af 100644
--- a/paddler_messaging/src/model_download_status.rs
+++ b/paddler_messaging/src/model_download_status.rs
@@ -29,6 +29,18 @@ impl ModelDownloadStatus {
             Self::NotDownloading => {}
         }
     }
+
+    pub const fn set_downloaded_bytes(&mut self, completed_bytes: u64) {
+        match self {
+            Self::Downloading {
+                downloaded_bytes, ..
+            }
+            | Self::DownloadingWithUnknownSize {
+                downloaded_bytes, ..
+            } => *downloaded_bytes = completed_bytes,
+            Self::NotDownloading => {}
+        }
+    }
 }
 
 #[cfg(test)]
@@ -53,6 +65,26 @@ mod tests {
         );
     }
 
+    #[test]
+    fn adds_downloaded_bytes_to_a_download_of_known_size() {
+        let mut download_status = ModelDownloadStatus::Downloading {
+            downloaded_bytes: 100,
+            model_path: "https://example.com/model.gguf".to_owned(),
+            total_bytes: 400,
+        };
+
+        download_status.add_downloaded_bytes(50);
+
+        assert_eq!(
+            download_status,
+            ModelDownloadStatus::Downloading {
+                downloaded_bytes: 150,
+                model_path: "https://example.com/model.gguf".to_owned(),
+                total_bytes: 400,
+            }
+        );
+    }
+
     #[test]
     fn bytes_arriving_after_the_download_stopped_do_not_restart_it() {
         let mut download_status = ModelDownloadStatus::NotDownloading;
@@ -61,4 +93,51 @@ mod tests {
 
         assert_eq!(download_status, ModelDownloadStatus::NotDownloading);
     }
+
+    #[test]
+    fn replaces_the_downloaded_bytes_of_a_running_download() {
+        let mut download_status = ModelDownloadStatus::Downloading {
+            downloaded_bytes: 100,
+            model_path: "https://example.com/model.gguf".to_owned(),
+            total_bytes: 400,
+        };
+
+        download_status.set_downloaded_bytes(250);
+
+        assert_eq!(
+            download_status,
+            ModelDownloadStatus::Downloading {
+                downloaded_bytes: 250,
+                model_path: "https://example.com/model.gguf".to_owned(),
+                total_bytes: 400,
+            }
+        );
+    }
+
+    #[test]
+    fn replaces_the_downloaded_bytes_of_a_download_of_unknown_size() {
+        let mut download_status = ModelDownloadStatus::DownloadingWithUnknownSize {
+            downloaded_bytes: 100,
+            model_path: "https://example.com/model.gguf".to_owned(),
+        };
+
+        download_status.set_downloaded_bytes(250);
+
+        assert_eq!(
+            download_status,
+            ModelDownloadStatus::DownloadingWithUnknownSize {
+                downloaded_bytes: 250,
+                model_path: "https://example.com/model.gguf".to_owned(),
+            }
+        );
+    }
+
+    #[test]
+    fn a_byte_count_reported_after_the_download_stopped_does_not_restart_it() {
+        let mut download_status = ModelDownloadStatus::NotDownloading;
+
+        download_status.set_downloaded_bytes(50);
+
+        assert_eq!(download_status, ModelDownloadStatus::NotDownloading);
+    }
 }
diff --git a/paddler_messaging/src/model_metadata.rs b/paddler_messaging/src/model_metadata.rs
index 33cc8b125..5b6ae024e 100644
--- a/paddler_messaging/src/model_metadata.rs
+++ b/paddler_messaging/src/model_metadata.rs
@@ -3,13 +3,25 @@ use std::collections::BTreeMap;
 use serde::Deserialize;
 use serde::Serialize;
 
-#[derive(Clone, Debug, Default, Deserialize, Serialize)]
+use crate::chat_template::ChatTemplate;
+use crate::embedded_chat_template_key::EMBEDDED_CHAT_TEMPLATE_KEY;
+
+#[derive(Clone, Debug, Default, Deserialize, Eq, PartialEq, Serialize)]
 #[serde(deny_unknown_fields)]
 pub struct ModelMetadata {
     pub metadata: BTreeMap<String, String>,
 }
 
 impl ModelMetadata {
+    #[must_use]
+    pub fn embedded_chat_template(&self) -> Option<ChatTemplate> {
+        self.metadata
+            .get(EMBEDDED_CHAT_TEMPLATE_KEY)
+            .map(|content| ChatTemplate {
+                content: content.clone(),
+            })
+    }
+
     pub fn set_meta_field(&mut self, key: String, value: String) {
         self.metadata.insert(key, value);
     }
diff --git a/paddler_messaging/src/multimodal_settings.rs b/paddler_messaging/src/multimodal_settings.rs
new file mode 100644
index 000000000..2c211cb02
--- /dev/null
+++ b/paddler_messaging/src/multimodal_settings.rs
@@ -0,0 +1,24 @@
+use std::num::NonZeroU32;
+
+use serde::Deserialize;
+use serde::Serialize;
+
+use crate::agent_desired_model::AgentDesiredModel;
+
+const DEFAULT_IMAGE_RESIZE_TO_FIT: NonZeroU32 = NonZeroU32::new(1024).unwrap();
+
+#[derive(Clone, Debug, Deserialize, Eq, PartialEq, Serialize)]
+#[serde(deny_unknown_fields)]
+pub struct MultimodalSettings {
+    pub image_resize_to_fit: NonZeroU32,
+    pub projection: AgentDesiredModel,
+}
+
+impl Default for MultimodalSettings {
+    fn default() -> Self {
+        Self {
+            image_resize_to_fit: DEFAULT_IMAGE_RESIZE_TO_FIT,
+            projection: AgentDesiredModel::None,
+        }
+    }
+}
diff --git a/paddler_messaging/src/oversized_decision_details.rs b/paddler_messaging/src/oversized_decision_details.rs
new file mode 100644
index 000000000..3d69bb47d
--- /dev/null
+++ b/paddler_messaging/src/oversized_decision_details.rs
@@ -0,0 +1,9 @@
+use serde::Deserialize;
+use serde::Serialize;
+
+#[derive(Clone, Debug, Deserialize, Eq, PartialEq, Serialize)]
+#[serde(deny_unknown_fields)]
+pub struct OversizedDecisionDetails {
+    pub context_size: u32,
+    pub required_tokens: usize,
+}
diff --git a/paddler_messaging/src/oversized_embedding_document_details.rs b/paddler_messaging/src/oversized_embedding_document_details.rs
index 2d68a6a31..085b08922 100644
--- a/paddler_messaging/src/oversized_embedding_document_details.rs
+++ b/paddler_messaging/src/oversized_embedding_document_details.rs
@@ -1,7 +1,7 @@
 use serde::Deserialize;
 use serde::Serialize;
 
-#[derive(Debug, Deserialize, Eq, PartialEq, Serialize)]
+#[derive(Clone, Debug, Deserialize, Eq, PartialEq, Serialize)]
 #[serde(deny_unknown_fields)]
 pub struct OversizedEmbeddingDocumentDetails {
     pub document_tokens: usize,
diff --git a/paddler_messaging/src/ping_interval.rs b/paddler_messaging/src/ping_interval.rs
new file mode 100644
index 000000000..42b8b891d
--- /dev/null
+++ b/paddler_messaging/src/ping_interval.rs
@@ -0,0 +1,3 @@
+use std::time::Duration;
+
+pub const PING_INTERVAL: Duration = Duration::from_secs(3);
diff --git a/paddler_messaging/src/ping_liveness.rs b/paddler_messaging/src/ping_liveness.rs
new file mode 100644
index 000000000..f044c3a83
--- /dev/null
+++ b/paddler_messaging/src/ping_liveness.rs
@@ -0,0 +1,62 @@
+use crate::ping_turn::PingTurn;
+use crate::unanswered_pings_tolerated::UNANSWERED_PINGS_TOLERATED;
+
+#[derive(Default)]
+pub struct PingLiveness {
+    unanswered_pings: u32,
+}
+
+impl PingLiveness {
+    pub const fn record_pong(&mut self) {
+        self.unanswered_pings = 0;
+    }
+
+    #[must_use]
+    pub const fn next_ping(&mut self) -> PingTurn {
+        if self.unanswered_pings > UNANSWERED_PINGS_TOLERATED {
+            PingTurn::PeerStoppedAnswering
+        } else {
+            self.unanswered_pings += 1;
+
+            PingTurn::SendPing
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::PingLiveness;
+    use crate::ping_turn::PingTurn;
+    use crate::unanswered_pings_tolerated::UNANSWERED_PINGS_TOLERATED;
+
+    fn pings_until_the_peer_is_unresponsive(ping_liveness: &mut PingLiveness) -> u32 {
+        let mut sent_pings = 0;
+
+        while ping_liveness.next_ping() == PingTurn::SendPing {
+            sent_pings += 1;
+        }
+
+        sent_pings
+    }
+
+    #[test]
+    fn a_peer_is_unresponsive_once_more_pings_than_tolerated_go_unanswered() {
+        assert_eq!(
+            pings_until_the_peer_is_unresponsive(&mut PingLiveness::default()),
+            UNANSWERED_PINGS_TOLERATED + 1
+        );
+    }
+
+    #[test]
+    fn a_pong_forgives_the_pings_sent_before_it() {
+        let mut ping_liveness = PingLiveness::default();
+
+        assert_eq!(ping_liveness.next_ping(), PingTurn::SendPing);
+        ping_liveness.record_pong();
+
+        assert_eq!(
+            pings_until_the_peer_is_unresponsive(&mut ping_liveness),
+            UNANSWERED_PINGS_TOLERATED + 1
+        );
+    }
+}
diff --git a/paddler_messaging/src/ping_turn.rs b/paddler_messaging/src/ping_turn.rs
new file mode 100644
index 000000000..43c30dcc6
--- /dev/null
+++ b/paddler_messaging/src/ping_turn.rs
@@ -0,0 +1,5 @@
+#[derive(Debug, Eq, PartialEq)]
+pub enum PingTurn {
+    PeerStoppedAnswering,
+    SendPing,
+}
diff --git a/paddler_messaging/src/request_params/continue_from_conversation_history_params/mod.rs b/paddler_messaging/src/request_params/continue_from_conversation_history_params/mod.rs
index d6f3ed0ca..83d36c15b 100644
--- a/paddler_messaging/src/request_params/continue_from_conversation_history_params/mod.rs
+++ b/paddler_messaging/src/request_params/continue_from_conversation_history_params/mod.rs
@@ -74,7 +74,7 @@ use crate::validates::Validates as _;
             from_value(json!({
                 "add_generation_prompt": true,
                 "conversation_history": [
-                    {"content": "Hello", "role": "user"}
+                    {"Message": {"content": "Hello", "role": "user"}}
                 ],
                 "enable_thinking": false,
                 "max_tokens": 10,
@@ -94,7 +94,7 @@ use crate::validates::Validates as _;
         let request_without_grammar = json!({
             "add_generation_prompt": true,
             "conversation_history": [
-                {"content": "Hello", "role": "user"}
+                {"Message": {"content": "Hello", "role": "user"}}
             ],
             "enable_thinking": false,
             "max_tokens": 10,
diff --git a/paddler_messaging/src/request_params/continue_from_conversation_history_params/tool/mod.rs b/paddler_messaging/src/request_params/continue_from_conversation_history_params/tool/mod.rs
index af21419b5..ae141bf36 100644
--- a/paddler_messaging/src/request_params/continue_from_conversation_history_params/tool/mod.rs
+++ b/paddler_messaging/src/request_params/continue_from_conversation_history_params/tool/mod.rs
@@ -2,6 +2,7 @@ pub mod tool_params;
 
 use serde::Deserialize;
 use serde::Serialize;
+use serde_json::Value;
 
 use crate::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters_schema::raw_parameters_schema::RawParametersSchema;
 use crate::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters_schema::validated_parameters_schema::ValidatedParametersSchema;
@@ -24,3 +25,11 @@ impl Validates<Tool<ValidatedParametersSchema>> for Tool<RawParametersSchema> {
         }
     }
 }
+
+impl From<Tool<ValidatedParametersSchema>> for Tool<Value> {
+    fn from(tool: Tool<ValidatedParametersSchema>) -> Self {
+        match tool {
+            Tool::Function(function_call) => Self::Function(FunctionCall::from(function_call)),
+        }
+    }
+}
diff --git a/paddler_messaging/src/request_params/continue_from_conversation_history_params/tool/tool_params/function_call/function.rs b/paddler_messaging/src/request_params/continue_from_conversation_history_params/tool/tool_params/function_call/function.rs
index 2c97532fd..4f7795ccc 100644
--- a/paddler_messaging/src/request_params/continue_from_conversation_history_params/tool/tool_params/function_call/function.rs
+++ b/paddler_messaging/src/request_params/continue_from_conversation_history_params/tool/tool_params/function_call/function.rs
@@ -1,5 +1,6 @@
 use serde::Deserialize;
 use serde::Serialize;
+use serde_json::Value;
 
 use crate::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters_schema::raw_parameters_schema::RawParametersSchema;
 use crate::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters_schema::validated_parameters_schema::ValidatedParametersSchema;
@@ -26,3 +27,19 @@ impl Validates<Function<ValidatedParametersSchema>> for Function<RawParametersSc
         })
     }
 }
+
+impl From<Function<ValidatedParametersSchema>> for Function<Value> {
+    fn from(
+        Function {
+            name,
+            description,
+            parameters,
+        }: Function<ValidatedParametersSchema>,
+    ) -> Self {
+        Self {
+            name,
+            description,
+            parameters: Parameters::from(parameters),
+        }
+    }
+}
diff --git a/paddler_messaging/src/request_params/continue_from_conversation_history_params/tool/tool_params/function_call/mod.rs b/paddler_messaging/src/request_params/continue_from_conversation_history_params/tool/tool_params/function_call/mod.rs
index 5f3415efc..0b99fb464 100644
--- a/paddler_messaging/src/request_params/continue_from_conversation_history_params/tool/tool_params/function_call/mod.rs
+++ b/paddler_messaging/src/request_params/continue_from_conversation_history_params/tool/tool_params/function_call/mod.rs
@@ -4,6 +4,7 @@ pub mod parameters_schema;
 
 use serde::Deserialize;
 use serde::Serialize;
+use serde_json::Value;
 
 use crate::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters_schema::raw_parameters_schema::RawParametersSchema;
 use crate::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters_schema::validated_parameters_schema::ValidatedParametersSchema;
@@ -26,3 +27,11 @@ impl Validates<FunctionCall<ValidatedParametersSchema>> for FunctionCall<RawPara
         })
     }
 }
+
+impl From<FunctionCall<ValidatedParametersSchema>> for FunctionCall<Value> {
+    fn from(FunctionCall { function }: FunctionCall<ValidatedParametersSchema>) -> Self {
+        Self {
+            function: Function::from(function),
+        }
+    }
+}
diff --git a/paddler_messaging/src/request_params/continue_from_conversation_history_params/tool/tool_params/function_call/parameters.rs b/paddler_messaging/src/request_params/continue_from_conversation_history_params/tool/tool_params/function_call/parameters.rs
index 5432f98d8..cb4fb857f 100644
--- a/paddler_messaging/src/request_params/continue_from_conversation_history_params/tool/tool_params/function_call/parameters.rs
+++ b/paddler_messaging/src/request_params/continue_from_conversation_history_params/tool/tool_params/function_call/parameters.rs
@@ -1,5 +1,6 @@
 use serde::Deserialize;
 use serde::Serialize;
+use serde_json::Value;
 
 use crate::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters_schema::raw_parameters_schema::RawParametersSchema;
 use crate::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters_schema::validated_parameters_schema::ValidatedParametersSchema;
@@ -31,6 +32,15 @@ impl Validates<Parameters<ValidatedParametersSchema>> for Parameters<RawParamete
     }
 }
 
+impl From<Parameters<ValidatedParametersSchema>> for Parameters<Value> {
+    fn from(parameters: Parameters<ValidatedParametersSchema>) -> Self {
+        match parameters {
+            Parameters::Empty => Self::Empty,
+            Parameters::Schema(schema) => Self::Schema(Value::from(schema)),
+        }
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use serde_json::Map;
diff --git a/paddler_messaging/src/request_params/continue_from_conversation_history_params/tool/tool_params/function_call/parameters_schema/validated_parameters_schema.rs b/paddler_messaging/src/request_params/continue_from_conversation_history_params/tool/tool_params/function_call/parameters_schema/validated_parameters_schema.rs
index d11df332e..11ddc0044 100644
--- a/paddler_messaging/src/request_params/continue_from_conversation_history_params/tool/tool_params/function_call/parameters_schema/validated_parameters_schema.rs
+++ b/paddler_messaging/src/request_params/continue_from_conversation_history_params/tool/tool_params/function_call/parameters_schema/validated_parameters_schema.rs
@@ -1,67 +1,56 @@
 use serde::Deserialize;
 use serde::Serialize;
-use serde::Serializer;
 use serde_json::Map;
 use serde_json::Value;
 
-#[derive(Clone, Debug, Default, Deserialize, Eq, PartialEq)]
+#[derive(Clone, Debug, Default, Deserialize, Eq, PartialEq, Serialize)]
 #[serde(deny_unknown_fields)]
 pub struct ValidatedParametersSchema {
     #[serde(rename = "type")]
     pub schema_type: String,
+    #[serde(skip_serializing_if = "Option::is_none")]
     pub properties: Option<Map<String, Value>>,
+    #[serde(skip_serializing_if = "Option::is_none")]
     pub required: Option<Vec<String>>,
-    #[serde(rename = "additionalProperties")]
+    #[serde(
+        rename = "additionalProperties",
+        skip_serializing_if = "Option::is_none"
+    )]
     pub additional_properties: Option<Value>,
 }
 
-impl From<&ValidatedParametersSchema> for Value {
+impl From<ValidatedParametersSchema> for Value {
     fn from(
         ValidatedParametersSchema {
             schema_type,
             properties,
             required,
             additional_properties,
-        }: &ValidatedParametersSchema,
+        }: ValidatedParametersSchema,
     ) -> Self {
         let mut keywords = Map::new();
 
-        keywords.insert("type".to_owned(), Self::String(schema_type.clone()));
+        keywords.insert("type".to_owned(), Self::String(schema_type));
 
         if let Some(properties) = properties {
-            keywords.insert("properties".to_owned(), Self::Object(properties.clone()));
+            keywords.insert("properties".to_owned(), Self::Object(properties));
         }
 
         if let Some(required) = required {
             keywords.insert(
                 "required".to_owned(),
-                Self::Array(required.iter().cloned().map(Self::String).collect()),
+                Self::Array(required.into_iter().map(Self::String).collect()),
             );
         }
 
         if let Some(additional_properties) = additional_properties {
-            keywords.insert(
-                "additionalProperties".to_owned(),
-                additional_properties.clone(),
-            );
+            keywords.insert("additionalProperties".to_owned(), additional_properties);
         }
 
         Self::Object(keywords)
     }
 }
 
-impl Serialize for ValidatedParametersSchema {
-    fn serialize<TSerializer>(
-        &self,
-        serializer: TSerializer,
-    ) -> Result<TSerializer::Ok, TSerializer::Error>
-    where
-        TSerializer: Serializer,
-    {
-        Value::from(self).serialize(serializer)
-    }
-}
-
 #[cfg(test)]
 mod tests {
     use serde_json::Map;
diff --git a/paddler_messaging/src/request_params/decide_params/mod.rs b/paddler_messaging/src/request_params/decide_params/mod.rs
new file mode 100644
index 000000000..a2a3adb10
--- /dev/null
+++ b/paddler_messaging/src/request_params/decide_params/mod.rs
@@ -0,0 +1,24 @@
+pub mod raw_decide_params;
+
+use std::num::NonZeroU16;
+
+use serde::Deserialize;
+use serde::Serialize;
+
+use crate::decision_question::DecisionQuestion;
+use crate::decision_slots_for_leading_questions::decision_slots_for_leading_questions;
+
+#[derive(Clone, Debug, Deserialize, Eq, PartialEq, Serialize)]
+#[serde(deny_unknown_fields)]
+pub struct DecideParams {
+    pub last_question: DecisionQuestion,
+    pub leading_questions: Vec<DecisionQuestion>,
+    pub state: String,
+}
+
+impl DecideParams {
+    #[must_use]
+    pub const fn required_slots(&self) -> NonZeroU16 {
+        decision_slots_for_leading_questions(self.leading_questions.len())
+    }
+}
diff --git a/paddler_messaging/src/request_params/decide_params/raw_decide_params.rs b/paddler_messaging/src/request_params/decide_params/raw_decide_params.rs
new file mode 100644
index 000000000..7db905f76
--- /dev/null
+++ b/paddler_messaging/src/request_params/decide_params/raw_decide_params.rs
@@ -0,0 +1,131 @@
+use std::collections::BTreeSet;
+
+use serde::Deserialize;
+use serde::Serialize;
+
+use crate::decision_question::DecisionQuestion;
+use crate::request_params::decide_params::DecideParams;
+use crate::request_params_validation_error::RequestParamsValidationError;
+use crate::validates::Validates;
+
+#[derive(Clone, Debug, Deserialize, Eq, PartialEq, Serialize)]
+#[serde(deny_unknown_fields)]
+pub struct RawDecideParams {
+    pub questions: Vec<DecisionQuestion>,
+    pub state: String,
+}
+
+impl Validates<DecideParams> for RawDecideParams {
+    fn validate(self) -> Result<DecideParams, RequestParamsValidationError> {
+        let mut question_ids = BTreeSet::new();
+
+        for question in &self.questions {
+            if question.options.is_empty() {
+                return Err(
+                    RequestParamsValidationError::DecisionQuestionWithoutOptions {
+                        question_id: question.id.clone(),
+                    },
+                );
+            }
+
+            if !question_ids.insert(question.id.as_str()) {
+                return Err(RequestParamsValidationError::DecisionQuestionIdRepeated {
+                    question_id: question.id.clone(),
+                });
+            }
+        }
+
+        let mut leading_questions = self.questions;
+        let last_question = leading_questions
+            .pop()
+            .ok_or(RequestParamsValidationError::DecisionWithoutQuestions)?;
+
+        Ok(DecideParams {
+            last_question,
+            leading_questions,
+            state: self.state,
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::RawDecideParams;
+    use crate::decision_question::DecisionQuestion;
+    use crate::request_params::decide_params::DecideParams;
+    use crate::request_params_validation_error::RequestParamsValidationError;
+    use crate::validates::Validates as _;
+
+    fn question(id: &str, options: &[&str]) -> DecisionQuestion {
+        DecisionQuestion {
+            id: id.to_owned(),
+            instructions: "Pick one".to_owned(),
+            options: options.iter().map(|option| (*option).to_owned()).collect(),
+        }
+    }
+
+    #[test]
+    fn a_decision_needs_at_least_one_question() {
+        assert_eq!(
+            RawDecideParams {
+                questions: Vec::new(),
+                state: "state".to_owned(),
+            }
+            .validate(),
+            Err(RequestParamsValidationError::DecisionWithoutQuestions)
+        );
+    }
+
+    #[test]
+    fn every_question_needs_at_least_one_option() {
+        assert_eq!(
+            RawDecideParams {
+                questions: vec![question("first", &["yes"]), question("second", &[])],
+                state: "state".to_owned(),
+            }
+            .validate(),
+            Err(
+                RequestParamsValidationError::DecisionQuestionWithoutOptions {
+                    question_id: "second".to_owned(),
+                }
+            )
+        );
+    }
+
+    #[test]
+    fn question_ids_must_be_unique() {
+        assert_eq!(
+            RawDecideParams {
+                questions: vec![question("same", &["yes"]), question("same", &["no"])],
+                state: "state".to_owned(),
+            }
+            .validate(),
+            Err(RequestParamsValidationError::DecisionQuestionIdRepeated {
+                question_id: "same".to_owned(),
+            })
+        );
+    }
+
+    #[test]
+    fn valid_questions_keep_their_order_with_the_last_one_apart() {
+        assert_eq!(
+            RawDecideParams {
+                questions: vec![
+                    question("first", &["yes", "no"]),
+                    question("second", &["a"]),
+                    question("third", &["b"]),
+                ],
+                state: "state".to_owned(),
+            }
+            .validate(),
+            Ok(DecideParams {
+                last_question: question("third", &["b"]),
+                leading_questions: vec![
+                    question("first", &["yes", "no"]),
+                    question("second", &["a"]),
+                ],
+                state: "state".to_owned(),
+            })
+        );
+    }
+}
diff --git a/paddler_messaging/src/request_params/generate_embedding_batch_params/mod.rs b/paddler_messaging/src/request_params/generate_embedding_batch_params/mod.rs
index 81b9eb70f..e4b835465 100644
--- a/paddler_messaging/src/request_params/generate_embedding_batch_params/mod.rs
+++ b/paddler_messaging/src/request_params/generate_embedding_batch_params/mod.rs
@@ -17,7 +17,7 @@ impl GenerateEmbeddingBatchParams {
     #[must_use]
     pub fn chunk_evenly_with_cap(
         self,
-        agent_count: NonZeroUsize,
+        agent_count: usize,
         max_documents_per_chunk: NonZeroUsize,
     ) -> Vec<Self> {
         let Self {
@@ -31,7 +31,7 @@ impl GenerateEmbeddingBatchParams {
         }
 
         let chunks_to_honor_cap = document_count.div_ceil(max_documents_per_chunk.get());
-        let chunk_count = document_count.min(agent_count.get().max(chunks_to_honor_cap));
+        let chunk_count = document_count.min(agent_count.max(chunks_to_honor_cap));
 
         let quotient = document_count / chunk_count;
         let remainder = document_count % chunk_count;
@@ -63,10 +63,6 @@ mod tests {
     use crate::embedding_input_document::EmbeddingInputDocument;
     use crate::embedding_normalization_method::EmbeddingNormalizationMethod;
 
-    fn agents(agent_count: usize) -> NonZeroUsize {
-        NonZeroUsize::new(agent_count).expect("an agent count must be non-zero")
-    }
-
     fn chunk_cap(max_documents_per_chunk: usize) -> NonZeroUsize {
         NonZeroUsize::new(max_documents_per_chunk).expect("a chunk cap must be non-zero")
     }
@@ -95,7 +91,7 @@ mod tests {
     fn chunk_evenly_with_cap_empty_input() {
         let params = make_params(vec![]);
 
-        let sub_batches = params.chunk_evenly_with_cap(agents(4), chunk_cap(256));
+        let sub_batches = params.chunk_evenly_with_cap(4, chunk_cap(256));
 
         assert!(sub_batches.is_empty());
     }
@@ -104,7 +100,7 @@ mod tests {
     fn chunk_evenly_with_cap_single_doc_single_agent() {
         let params = make_params(vec![make_doc("only", "content")]);
 
-        let sub_batches = params.chunk_evenly_with_cap(agents(1), chunk_cap(256));
+        let sub_batches = params.chunk_evenly_with_cap(1, chunk_cap(256));
 
         assert_eq!(sub_batches.len(), 1);
         assert_eq!(sub_batches[0].input_batch.len(), 1);
@@ -115,7 +111,7 @@ mod tests {
     fn chunk_evenly_with_cap_single_doc_many_agents() {
         let params = make_params(vec![make_doc("only", "content")]);
 
-        let sub_batches = params.chunk_evenly_with_cap(agents(5), chunk_cap(256));
+        let sub_batches = params.chunk_evenly_with_cap(5, chunk_cap(256));
 
         assert_eq!(sub_batches.len(), 1);
         assert_eq!(sub_batches[0].input_batch.len(), 1);
@@ -126,7 +122,7 @@ mod tests {
     fn chunk_evenly_with_cap_more_agents_than_docs_uses_n_chunks() {
         let params = make_params(make_docs(3));
 
-        let sub_batches = params.chunk_evenly_with_cap(agents(5), chunk_cap(256));
+        let sub_batches = params.chunk_evenly_with_cap(5, chunk_cap(256));
 
         assert_eq!(sub_batches.len(), 3);
         for sub_batch in &sub_batches {
@@ -138,7 +134,7 @@ mod tests {
     fn chunk_evenly_with_cap_below_cap_splits_per_agent() {
         let params = make_params(make_docs(4));
 
-        let sub_batches = params.chunk_evenly_with_cap(agents(4), chunk_cap(256));
+        let sub_batches = params.chunk_evenly_with_cap(4, chunk_cap(256));
 
         assert_eq!(sub_batches.len(), 4);
         for sub_batch in &sub_batches {
@@ -150,7 +146,7 @@ mod tests {
     fn chunk_evenly_with_cap_below_cap_uneven_split() {
         let params = make_params(make_docs(11));
 
-        let sub_batches = params.chunk_evenly_with_cap(agents(4), chunk_cap(256));
+        let sub_batches = params.chunk_evenly_with_cap(4, chunk_cap(256));
 
         assert_eq!(sub_batches.len(), 4);
         assert_eq!(sub_batches[0].input_batch.len(), 3);
@@ -163,7 +159,7 @@ mod tests {
     fn chunk_evenly_with_cap_splits_a_batch_below_the_cap_across_every_agent() {
         let params = make_params(make_docs(80));
 
-        let sub_batches = params.chunk_evenly_with_cap(agents(4), chunk_cap(100));
+        let sub_batches = params.chunk_evenly_with_cap(4, chunk_cap(100));
 
         assert_eq!(sub_batches.len(), 4);
         for sub_batch in &sub_batches {
@@ -175,7 +171,7 @@ mod tests {
     fn chunk_evenly_with_cap_splits_a_batch_above_the_cap_into_capped_chunks() {
         let params = make_params(make_docs(1000));
 
-        let sub_batches = params.chunk_evenly_with_cap(agents(4), chunk_cap(100));
+        let sub_batches = params.chunk_evenly_with_cap(4, chunk_cap(100));
 
         assert_eq!(sub_batches.len(), 10);
         for sub_batch in &sub_batches {
@@ -187,7 +183,7 @@ mod tests {
     fn chunk_evenly_with_cap_at_cap_boundary_uses_agent_count() {
         let params = make_params(make_docs(1024));
 
-        let sub_batches = params.chunk_evenly_with_cap(agents(4), chunk_cap(256));
+        let sub_batches = params.chunk_evenly_with_cap(4, chunk_cap(256));
 
         assert_eq!(sub_batches.len(), 4);
         for sub_batch in &sub_batches {
@@ -199,7 +195,7 @@ mod tests {
     fn chunk_evenly_with_cap_above_cap_boundary_creates_extra_chunks() {
         let params = make_params(make_docs(2000));
 
-        let sub_batches = params.chunk_evenly_with_cap(agents(4), chunk_cap(256));
+        let sub_batches = params.chunk_evenly_with_cap(4, chunk_cap(256));
 
         assert_eq!(sub_batches.len(), 8);
         for sub_batch in &sub_batches {
@@ -211,7 +207,7 @@ mod tests {
     fn chunk_evenly_with_cap_far_above_cap_distributes_evenly() {
         let params = make_params(make_docs(1100));
 
-        let sub_batches = params.chunk_evenly_with_cap(agents(4), chunk_cap(256));
+        let sub_batches = params.chunk_evenly_with_cap(4, chunk_cap(256));
 
         assert_eq!(sub_batches.len(), 5);
         for sub_batch in &sub_batches {
@@ -223,7 +219,7 @@ mod tests {
     fn chunk_evenly_with_cap_extreme_large_n_small_cap() {
         let params = make_params(make_docs(10_000));
 
-        let sub_batches = params.chunk_evenly_with_cap(agents(4), chunk_cap(1));
+        let sub_batches = params.chunk_evenly_with_cap(4, chunk_cap(1));
 
         assert_eq!(sub_batches.len(), 10_000);
         for sub_batch in &sub_batches {
@@ -235,7 +231,7 @@ mod tests {
     fn chunk_evenly_with_cap_extreme_one_doc_per_chunk() {
         let params = make_params(make_docs(100));
 
-        let sub_batches = params.chunk_evenly_with_cap(agents(100), chunk_cap(256));
+        let sub_batches = params.chunk_evenly_with_cap(100, chunk_cap(256));
 
         assert_eq!(sub_batches.len(), 100);
         for sub_batch in &sub_batches {
@@ -246,7 +242,7 @@ mod tests {
     #[test]
     fn chunk_evenly_with_cap_no_sub_batch_exceeds_cap_sweep() {
         let document_counts: Vec<usize> = (0..=50).chain([256, 257, 1000, 2001]).collect();
-        let agent_counts: Vec<usize> = (1..=8).collect();
+        let agent_counts: Vec<usize> = (0..=8).collect();
         let caps: Vec<usize> = vec![1, 2, 4, 100, 256];
 
         for &document_count in &document_counts {
@@ -254,8 +250,7 @@ mod tests {
                 for &cap in &caps {
                     let params = make_params(make_docs(document_count));
 
-                    let sub_batches =
-                        params.chunk_evenly_with_cap(agents(agent_count), chunk_cap(cap));
+                    let sub_batches = params.chunk_evenly_with_cap(agent_count, chunk_cap(cap));
 
                     let total_documents: usize =
                         sub_batches.iter().map(|sub| sub.input_batch.len()).sum();
@@ -305,6 +300,19 @@ mod tests {
         }
     }
 
+    #[test]
+    fn chunk_evenly_with_cap_without_agents_splits_only_to_honor_the_cap() {
+        let params = make_params(make_docs(600));
+
+        let sub_batches = params.chunk_evenly_with_cap(0, chunk_cap(256));
+
+        let sizes: Vec<usize> = sub_batches
+            .iter()
+            .map(|sub_batch| sub_batch.input_batch.len())
+            .collect();
+        assert_eq!(sizes, vec![200, 200, 200]);
+    }
+
     #[test]
     fn chunk_evenly_with_cap_preserves_normalization_method() {
         let params = GenerateEmbeddingBatchParams {
@@ -312,7 +320,7 @@ mod tests {
             normalization_method: EmbeddingNormalizationMethod::L2,
         };
 
-        let sub_batches = params.chunk_evenly_with_cap(agents(4), chunk_cap(256));
+        let sub_batches = params.chunk_evenly_with_cap(4, chunk_cap(256));
 
         let is_l2 = |normalization_method: &EmbeddingNormalizationMethod| {
             matches!(normalization_method, EmbeddingNormalizationMethod::L2)
@@ -331,7 +339,7 @@ mod tests {
     fn chunk_evenly_with_cap_preserves_document_ids_and_order() {
         let params = make_params(make_docs(12));
 
-        let sub_batches = params.chunk_evenly_with_cap(agents(5), chunk_cap(256));
+        let sub_batches = params.chunk_evenly_with_cap(5, chunk_cap(256));
 
         let collected_ids: Vec<String> = sub_batches
             .iter()
diff --git a/paddler_messaging/src/request_params/mod.rs b/paddler_messaging/src/request_params/mod.rs
index 036cc852b..54c6b798a 100644
--- a/paddler_messaging/src/request_params/mod.rs
+++ b/paddler_messaging/src/request_params/mod.rs
@@ -1,3 +1,4 @@
 pub mod continue_from_conversation_history_params;
 pub mod continue_from_raw_prompt_params;
+pub mod decide_params;
 pub mod generate_embedding_batch_params;
diff --git a/paddler_messaging/src/request_params_validation_error.rs b/paddler_messaging/src/request_params_validation_error.rs
index 1ad197d4e..05d806454 100644
--- a/paddler_messaging/src/request_params_validation_error.rs
+++ b/paddler_messaging/src/request_params_validation_error.rs
@@ -2,6 +2,12 @@ use thiserror::Error;
 
 #[derive(Debug, Eq, Error, PartialEq)]
 pub enum RequestParamsValidationError {
+    #[error("question '{question_id}' appears more than once")]
+    DecisionQuestionIdRepeated { question_id: String },
+    #[error("question '{question_id}' has no options")]
+    DecisionQuestionWithoutOptions { question_id: String },
+    #[error("a decision needs at least one question")]
+    DecisionWithoutQuestions,
     #[error("Required field '{field}' not found in properties")]
     RequiredFieldNotInProperties { field: String },
     #[error("parse_tool_calls requires at least one tool")]
diff --git a/paddler_messaging/src/reserved_request.rs b/paddler_messaging/src/reserved_request.rs
new file mode 100644
index 000000000..50a40bed3
--- /dev/null
+++ b/paddler_messaging/src/reserved_request.rs
@@ -0,0 +1,11 @@
+use std::num::NonZeroU16;
+
+use serde::Deserialize;
+use serde::Serialize;
+
+#[derive(Clone, Debug, Deserialize, Eq, PartialEq, Serialize)]
+#[serde(deny_unknown_fields)]
+pub struct ReservedRequest<TParams> {
+    pub params: TParams,
+    pub reserved_slots: NonZeroU16,
+}
diff --git a/paddler_messaging/src/result_step.rs b/paddler_messaging/src/result_step.rs
new file mode 100644
index 000000000..5cb9e7b9c
--- /dev/null
+++ b/paddler_messaging/src/result_step.rs
@@ -0,0 +1,5 @@
+#[derive(Debug, Eq, PartialEq)]
+pub enum ResultStep<TPartial, TFinal> {
+    Final(TFinal),
+    Partial(TPartial),
+}
diff --git a/paddler_messaging/src/served_decision_settings.rs b/paddler_messaging/src/served_decision_settings.rs
new file mode 100644
index 000000000..54b04bff4
--- /dev/null
+++ b/paddler_messaging/src/served_decision_settings.rs
@@ -0,0 +1,8 @@
+use serde::Deserialize;
+use serde::Serialize;
+
+#[derive(Clone, Debug, Deserialize, Eq, PartialEq, Serialize)]
+#[serde(deny_unknown_fields)]
+pub struct ServedDecisionSettings {
+    pub pointer_head_uri: String,
+}
diff --git a/paddler_messaging/src/streamable_result.rs b/paddler_messaging/src/streamable_result.rs
index aafe7641e..c08a361f7 100644
--- a/paddler_messaging/src/streamable_result.rs
+++ b/paddler_messaging/src/streamable_result.rs
@@ -1,3 +1,8 @@
-pub trait StreamableResult {
-    fn is_done(&self) -> bool;
+use crate::result_step::ResultStep;
+
+pub trait StreamableResult: Sized {
+    type Final: Into<Self>;
+    type Partial: Into<Self>;
+
+    fn into_step(self) -> ResultStep<Self::Partial, Self::Final>;
 }
diff --git a/paddler_messaging/src/system_role.rs b/paddler_messaging/src/system_role.rs
new file mode 100644
index 000000000..e36242624
--- /dev/null
+++ b/paddler_messaging/src/system_role.rs
@@ -0,0 +1 @@
+pub const SYSTEM_ROLE: &str = "system";
diff --git a/paddler_messaging/src/tool_role.rs b/paddler_messaging/src/tool_role.rs
new file mode 100644
index 000000000..1cdc443b4
--- /dev/null
+++ b/paddler_messaging/src/tool_role.rs
@@ -0,0 +1 @@
+pub const TOOL_ROLE: &str = "tool";
diff --git a/paddler_messaging/src/unanswered_pings_tolerated.rs b/paddler_messaging/src/unanswered_pings_tolerated.rs
new file mode 100644
index 000000000..d358c18f5
--- /dev/null
+++ b/paddler_messaging/src/unanswered_pings_tolerated.rs
@@ -0,0 +1 @@
+pub const UNANSWERED_PINGS_TOLERATED: u32 = 1;
diff --git a/paddler_messaging/src/url_model_reference.rs b/paddler_messaging/src/url_model_reference.rs
deleted file mode 100644
index 740decc4c..000000000
--- a/paddler_messaging/src/url_model_reference.rs
+++ /dev/null
@@ -1,8 +0,0 @@
-use serde::Deserialize;
-use serde::Serialize;
-
-#[derive(Clone, Debug, Deserialize, Eq, PartialEq, Serialize)]
-#[serde(deny_unknown_fields)]
-pub struct UrlModelReference {
-    pub url: String,
-}
diff --git a/paddler_messaging/src/user_role.rs b/paddler_messaging/src/user_role.rs
new file mode 100644
index 000000000..459497118
--- /dev/null
+++ b/paddler_messaging/src/user_role.rs
@@ -0,0 +1 @@
+pub const USER_ROLE: &str = "user";
diff --git a/paddler_model_source/Cargo.toml b/paddler_model_source/Cargo.toml
index ecf6edb01..8c11bd75b 100644
--- a/paddler_model_source/Cargo.toml
+++ b/paddler_model_source/Cargo.toml
@@ -9,13 +9,14 @@ repository.workspace = true
 version.workspace = true
 
 [dependencies]
-async-trait = { workspace = true }
 hf-hub = { workspace = true }
 log = { workspace = true }
+percent-encoding = { workspace = true }
 reqwest = { workspace = true }
 thiserror = { workspace = true }
 tokio = { workspace = true }
 tokio-util = { workspace = true }
+url = { workspace = true }
 paddler_agent_status = { workspace = true }
 paddler_cache_dir = { workspace = true }
 paddler_download_manager = { workspace = true }
@@ -24,7 +25,7 @@ paddler_messaging = { workspace = true }
 [dev-dependencies]
 paddler_local_http_fixture = { workspace = true }
 tempfile = { workspace = true }
-url = { workspace = true }
+tokio = { workspace = true, features = ["test-util"] }
 
 [lints]
 workspace = true
diff --git a/paddler_model_source/src/agent_local_model_path.rs b/paddler_model_source/src/agent_local_model_path.rs
new file mode 100644
index 000000000..a96f4dd5c
--- /dev/null
+++ b/paddler_model_source/src/agent_local_model_path.rs
@@ -0,0 +1,274 @@
+use std::os::unix::ffi::OsStrExt as _;
+use std::path::Component;
+use std::path::Path;
+use std::path::PathBuf;
+
+use percent_encoding::percent_encode;
+use url::Url;
+
+use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
+use paddler_messaging::agent_issue::AgentIssue;
+use paddler_messaging::agent_issue_params::agent_model_uri_not_canonical_params::AgentModelUriNotCanonicalParams;
+use paddler_messaging::agent_issue_params::model_path::ModelPath;
+
+use crate::model_source_error::ModelSourceError;
+use crate::url_path_segment_encode_set::URL_PATH_SEGMENT_ENCODE_SET;
+
+fn canonical_agent_uri(normalized_absolute_path: &Path) -> String {
+    let mut encoded_path = String::new();
+
+    for component in normalized_absolute_path.components().skip(1) {
+        encoded_path.push('/');
+        encoded_path.extend(percent_encode(
+            component.as_os_str().as_bytes(),
+            URL_PATH_SEGMENT_ENCODE_SET,
+        ));
+    }
+
+    if encoded_path.is_empty() {
+        format!("{}:///", AgentLocalModelPath::SCHEME)
+    } else {
+        format!("{}://{encoded_path}", AgentLocalModelPath::SCHEME)
+    }
+}
+
+#[derive(Clone, Debug, Eq, PartialEq)]
+pub struct AgentLocalModelPath {
+    canonical_uri: String,
+    path: PathBuf,
+}
+
+impl AgentLocalModelPath {
+    pub const SCHEME: &str = "agent";
+
+    pub fn from_absolute_path(path: &Path) -> Result<Self, ModelSourceError> {
+        if !path.is_absolute() {
+            return Err(ModelSourceError::LocalModelPathNotAbsolute {
+                path: path.to_path_buf(),
+            });
+        }
+
+        if path
+            .components()
+            .any(|component| component == Component::ParentDir)
+        {
+            return Err(ModelSourceError::LocalModelPathClimbsToParentDirectory {
+                path: path.to_path_buf(),
+            });
+        }
+
+        Ok(Self::from_normalized_absolute_path(
+            path.components().collect(),
+        ))
+    }
+
+    pub fn from_agent_url(
+        uri: &str,
+        agent_url: &Url,
+        slot_aggregated_status: &SlotAggregatedStatus,
+    ) -> Result<Self, ModelSourceError> {
+        let model_path = || ModelPath {
+            model_path: uri.to_owned(),
+        };
+        let path = agent_url.to_file_path().map_err(|()| {
+            slot_aggregated_status.register_desired_state_issue(
+                AgentIssue::AgentModelUriIsNotAnAbsoluteLocalPath(model_path()),
+            );
+
+            ModelSourceError::AgentUriNotAnAbsoluteLocalPath {
+                uri: uri.to_owned(),
+            }
+        })?;
+        let agent_local_model_path = Self::from_normalized_absolute_path(path);
+
+        if agent_local_model_path.canonical_uri != uri {
+            slot_aggregated_status.register_desired_state_issue(
+                AgentIssue::AgentModelUriIsNotCanonical(AgentModelUriNotCanonicalParams {
+                    canonical_uri: agent_local_model_path.canonical_uri.clone(),
+                    model_path: model_path(),
+                }),
+            );
+
+            return Err(ModelSourceError::AgentUriNotCanonical {
+                canonical_uri: agent_local_model_path.canonical_uri,
+                uri: uri.to_owned(),
+            });
+        }
+
+        Ok(agent_local_model_path)
+    }
+
+    #[must_use]
+    pub fn canonical_uri(&self) -> &str {
+        &self.canonical_uri
+    }
+
+    #[must_use]
+    pub fn path(&self) -> &Path {
+        &self.path
+    }
+
+    fn from_normalized_absolute_path(path: PathBuf) -> Self {
+        Self {
+            canonical_uri: canonical_agent_uri(&path),
+            path,
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::path::Path;
+
+    use url::Url;
+
+    use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
+    use paddler_messaging::agent_issue::AgentIssue;
+    use paddler_messaging::agent_issue_params::agent_model_uri_not_canonical_params::AgentModelUriNotCanonicalParams;
+    use paddler_messaging::agent_issue_params::model_path::ModelPath;
+
+    use super::AgentLocalModelPath;
+    use crate::model_source_error::ModelSourceError;
+
+    fn parsed(uri: &str, slot_aggregated_status: &SlotAggregatedStatus) -> AgentLocalModelPath {
+        AgentLocalModelPath::from_agent_url(uri, &Url::parse(uri).unwrap(), slot_aggregated_status)
+            .unwrap()
+    }
+
+    fn refusal(uri: &str, slot_aggregated_status: &SlotAggregatedStatus) -> String {
+        AgentLocalModelPath::from_agent_url(uri, &Url::parse(uri).unwrap(), slot_aggregated_status)
+            .unwrap_err()
+            .to_string()
+    }
+
+    #[test]
+    fn formats_a_path_with_reserved_and_non_ascii_characters_as_its_canonical_uri() {
+        assert_eq!(
+            AgentLocalModelPath::from_absolute_path(Path::new("/models/my mod\u{e8}l #1?%.gguf"))
+                .unwrap()
+                .canonical_uri(),
+            "agent:///models/my%20mod%C3%A8l%20%231%3F%25.gguf"
+        );
+    }
+
+    #[test]
+    fn formats_the_root_directory_as_the_canonical_root_uri() {
+        assert_eq!(
+            AgentLocalModelPath::from_absolute_path(Path::new("/"))
+                .unwrap()
+                .canonical_uri(),
+            "agent:///"
+        );
+    }
+
+    #[test]
+    fn decodes_a_canonical_uri_into_the_path_it_names() {
+        assert_eq!(
+            parsed(
+                "agent:///models/my%20mod%C3%A8l%20%231%3F%25.gguf",
+                &SlotAggregatedStatus::new(1)
+            )
+            .path(),
+            Path::new("/models/my mod\u{e8}l #1?%.gguf")
+        );
+    }
+
+    #[test]
+    fn refuses_to_format_a_relative_path() {
+        assert_eq!(
+            AgentLocalModelPath::from_absolute_path(Path::new("models/model.gguf"))
+                .unwrap_err()
+                .to_string(),
+            ModelSourceError::LocalModelPathNotAbsolute {
+                path: "models/model.gguf".into(),
+            }
+            .to_string()
+        );
+    }
+
+    #[test]
+    fn refuses_to_format_a_path_that_climbs_to_a_parent_directory() {
+        assert_eq!(
+            AgentLocalModelPath::from_absolute_path(Path::new("/models/../model.gguf"))
+                .unwrap_err()
+                .to_string(),
+            ModelSourceError::LocalModelPathClimbsToParentDirectory {
+                path: "/models/../model.gguf".into(),
+            }
+            .to_string()
+        );
+    }
+
+    #[test]
+    fn refuses_every_spelling_that_differs_from_the_canonical_uri() {
+        for (uri, canonical_uri) in [
+            (
+                "agent:///models/my model.gguf",
+                "agent:///models/my%20model.gguf",
+            ),
+            (
+                "agent:///models/mod\u{e8}l.gguf",
+                "agent:///models/mod%C3%A8l.gguf",
+            ),
+            ("agent:///models/%6Dodel.gguf", "agent:///models/model.gguf"),
+            ("agent:///models/a%2Fb.gguf", "agent:///models/a/b.gguf"),
+            ("agent:///models/../model.gguf", "agent:///model.gguf"),
+            ("agent:///models//model.gguf", "agent:///models/model.gguf"),
+            (
+                "agent:///models/model.gguf?revision=1",
+                "agent:///models/model.gguf",
+            ),
+            (
+                "agent:///models/model.gguf#fragment",
+                "agent:///models/model.gguf",
+            ),
+            (
+                "agent://localhost/models/model.gguf",
+                "agent:///models/model.gguf",
+            ),
+            ("agent:/models/model.gguf", "agent:///models/model.gguf"),
+            ("AGENT:///models/model.gguf", "agent:///models/model.gguf"),
+        ] {
+            let slot_aggregated_status = SlotAggregatedStatus::new(1);
+
+            assert_eq!(
+                refusal(uri, &slot_aggregated_status),
+                ModelSourceError::AgentUriNotCanonical {
+                    canonical_uri: canonical_uri.to_owned(),
+                    uri: uri.to_owned(),
+                }
+                .to_string()
+            );
+            assert!(
+                slot_aggregated_status.has_issue(&AgentIssue::AgentModelUriIsNotCanonical(
+                    AgentModelUriNotCanonicalParams {
+                        canonical_uri: canonical_uri.to_owned(),
+                        model_path: ModelPath {
+                            model_path: uri.to_owned(),
+                        },
+                    }
+                ))
+            );
+        }
+    }
+
+    #[test]
+    fn refuses_uris_that_do_not_name_an_absolute_local_path() {
+        for uri in ["agent://models/model.gguf", "agent:model.gguf"] {
+            let slot_aggregated_status = SlotAggregatedStatus::new(1);
+
+            assert_eq!(
+                refusal(uri, &slot_aggregated_status),
+                ModelSourceError::AgentUriNotAnAbsoluteLocalPath {
+                    uri: uri.to_owned(),
+                }
+                .to_string()
+            );
+            assert!(slot_aggregated_status.has_issue(
+                &AgentIssue::AgentModelUriIsNotAnAbsoluteLocalPath(ModelPath {
+                    model_path: uri.to_owned(),
+                })
+            ));
+        }
+    }
+}
diff --git a/paddler_model_source/src/desired_model_resolution.rs b/paddler_model_source/src/desired_model_resolution.rs
index 9ccad2c28..1045a1c33 100644
--- a/paddler_model_source/src/desired_model_resolution.rs
+++ b/paddler_model_source/src/desired_model_resolution.rs
@@ -3,7 +3,6 @@ use std::path::PathBuf;
 #[derive(Debug, Eq, PartialEq)]
 pub enum DesiredModelResolution {
     Cancelled,
-    NotConfigured,
     Resolved(PathBuf),
     LocalFileMissing(PathBuf),
 }
diff --git a/paddler_model_source/src/desired_model_resolver.rs b/paddler_model_source/src/desired_model_resolver.rs
new file mode 100644
index 000000000..7415c0cc1
--- /dev/null
+++ b/paddler_model_source/src/desired_model_resolver.rs
@@ -0,0 +1,133 @@
+use std::sync::Arc;
+
+use hf_hub::HFClient;
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
+use paddler_cache_dir::cache_dir::CacheDir;
+use paddler_download_manager::download_manager::DownloadManager;
+
+use crate::desired_model_resolution::DesiredModelResolution;
+use crate::huggingface_cache_location::HuggingFaceCacheLocation;
+use crate::huggingface_hub::HuggingFaceHub;
+use crate::huggingface_model_source::HuggingFaceModelSource;
+use crate::local_model_path::LocalModelPath;
+use crate::model_source::ModelSource;
+use crate::model_source_error::ModelSourceError;
+use crate::resolves_model_source::ResolvesModelSource as _;
+use crate::url_model_source::UrlModelSource;
+
+#[derive(Clone)]
+pub struct DesiredModelResolver {
+    pub cache_dir: CacheDir,
+    pub download_manager: DownloadManager,
+    pub huggingface_client: HFClient,
+}
+
+impl DesiredModelResolver {
+    pub fn from_process_env(huggingface_endpoint: String) -> Result<Self, ModelSourceError> {
+        Ok(Self {
+            cache_dir: CacheDir::from_process_env(),
+            download_manager: DownloadManager::new()
+                .map_err(ModelSourceError::DownloadClientUnavailable)?,
+            huggingface_client: HuggingFaceHub {
+                cache_location: HuggingFaceCacheLocation::HubDefault,
+                endpoint: huggingface_endpoint,
+            }
+            .client()
+            .map_err(ModelSourceError::HuggingFaceClientUnavailable)?,
+        })
+    }
+
+    pub async fn resolve_uri(
+        &self,
+        cancellation_token: &CancellationToken,
+        uri: &str,
+        slot_aggregated_status: Arc<SlotAggregatedStatus>,
+    ) -> Result<DesiredModelResolution, ModelSourceError> {
+        match ModelSource::parse(uri, &slot_aggregated_status)? {
+            ModelSource::HuggingFace(reference) => {
+                HuggingFaceModelSource {
+                    client: self.huggingface_client.clone(),
+                    reference,
+                }
+                .resolve(cancellation_token, slot_aggregated_status)
+                .await
+            }
+            ModelSource::LocalToAgent(agent_local_model_path) => {
+                LocalModelPath {
+                    path: agent_local_model_path.path().to_path_buf(),
+                }
+                .resolve(cancellation_token, slot_aggregated_status)
+                .await
+            }
+            ModelSource::Url(download_url) => {
+                UrlModelSource {
+                    cache_dir: self.cache_dir.clone(),
+                    download_manager: self.download_manager.clone(),
+                    download_url,
+                }
+                .resolve(cancellation_token, slot_aggregated_status)
+                .await
+            }
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::fs::File;
+    use std::sync::Arc;
+
+    use hf_hub::HFError;
+    use tempfile::tempdir;
+    use tokio_util::sync::CancellationToken;
+    use url::ParseError;
+    use url::Url;
+
+    use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
+
+    use super::DesiredModelResolver;
+    use crate::desired_model_resolution::DesiredModelResolution;
+    use crate::huggingface_default_endpoint::HUGGINGFACE_DEFAULT_ENDPOINT;
+    use crate::model_source_error::ModelSourceError;
+
+    #[test]
+    fn a_malformed_hugging_face_endpoint_keeps_the_resolver_from_being_built() {
+        assert_eq!(
+            DesiredModelResolver::from_process_env("not an endpoint".to_owned())
+                .err()
+                .map(|construction_error| construction_error.to_string()),
+            Some(
+                ModelSourceError::HuggingFaceClientUnavailable(HFError::Url(
+                    ParseError::RelativeUrlWithoutBase
+                ))
+                .to_string()
+            )
+        );
+    }
+
+    #[tokio::test]
+    async fn resolves_an_agent_uri_whose_file_name_is_percent_encoded() {
+        let model_directory = tempdir().unwrap();
+        let model_path = model_directory.path().join("my modèl.gguf");
+
+        File::create(&model_path).unwrap();
+
+        let resolution =
+            DesiredModelResolver::from_process_env(HUGGINGFACE_DEFAULT_ENDPOINT.to_owned())
+                .unwrap()
+                .resolve_uri(
+                    &CancellationToken::new(),
+                    &format!(
+                        "agent://{}/my%20mod%C3%A8l.gguf",
+                        Url::from_file_path(model_directory.path()).unwrap().path()
+                    ),
+                    Arc::new(SlotAggregatedStatus::new(1)),
+                )
+                .await
+                .unwrap();
+
+        assert_eq!(resolution, DesiredModelResolution::Resolved(model_path));
+    }
+}
diff --git a/paddler_model_source/src/download_error_agent_issue.rs b/paddler_model_source/src/download_error_agent_issue.rs
new file mode 100644
index 000000000..ba32013eb
--- /dev/null
+++ b/paddler_model_source/src/download_error_agent_issue.rs
@@ -0,0 +1,226 @@
+use paddler_download_manager::download_error::DownloadError;
+use paddler_messaging::agent_issue::AgentIssue;
+use paddler_messaging::agent_issue_params::model_path::ModelPath;
+
+#[must_use]
+pub const fn download_error_agent_issue(
+    error: &DownloadError,
+    model_path: ModelPath,
+) -> AgentIssue {
+    match error {
+        DownloadError::InvalidUrl { .. } | DownloadError::UnsupportedUrlScheme { .. } => {
+            AgentIssue::DownloadUrlIsMalformed(model_path)
+        }
+        DownloadError::NotFound { .. } => AgentIssue::ModelDoesNotExistAtUrl(model_path),
+        DownloadError::PermissionDenied { .. } => {
+            AgentIssue::DownloadServerDeniedAccess(model_path)
+        }
+        DownloadError::DownloadServerIsUnreachable { .. } => {
+            AgentIssue::DownloadServerIsUnreachable(model_path)
+        }
+        DownloadError::DownloadServerErrored { .. } => {
+            AgentIssue::DownloadServerErrored(model_path)
+        }
+        DownloadError::DownloadServerRejectedRequest { .. } => {
+            AgentIssue::DownloadServerRejectedRequest(model_path)
+        }
+        DownloadError::DownloadInterrupted { .. } => AgentIssue::DownloadInterrupted(model_path),
+        DownloadError::CachePermissionDenied { .. } => {
+            AgentIssue::CacheDirectoryIsNotWritable(model_path)
+        }
+        DownloadError::CacheDiskFull { .. } => AgentIssue::CacheStorageIsFull(model_path),
+        DownloadError::PartialFileStale { .. }
+        | DownloadError::PartialPathIsADirectory { .. }
+        | DownloadError::FinalPathHasNoFileName { .. }
+        | DownloadError::Io { .. } => AgentIssue::ModelCacheIsCorrupted(model_path),
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::io;
+    use std::path::PathBuf;
+
+    use reqwest::Client;
+    use reqwest::Error as ReqwestError;
+    use reqwest::StatusCode;
+    use url::Url;
+
+    use paddler_download_manager::download_error::DownloadError;
+    use paddler_messaging::agent_issue::AgentIssue;
+    use paddler_messaging::agent_issue_params::model_path::ModelPath;
+
+    use super::download_error_agent_issue;
+
+    const TEST_URL: &str = "https://example.com/m.gguf";
+
+    fn test_model_path() -> ModelPath {
+        ModelPath {
+            model_path: TEST_URL.to_owned(),
+        }
+    }
+
+    async fn refused_connection_error() -> ReqwestError {
+        Client::new()
+            .get("http://127.0.0.1:1/model.gguf")
+            .send()
+            .await
+            .expect_err("nothing listens on the loopback discard port")
+    }
+
+    #[test]
+    fn invalid_url_maps_to_download_url_is_malformed() {
+        let parse_error = Url::parse("not a url").err().unwrap();
+        let error = DownloadError::InvalidUrl {
+            url: "not a url".to_owned(),
+            source: parse_error,
+        };
+
+        assert_eq!(
+            download_error_agent_issue(&error, test_model_path()),
+            AgentIssue::DownloadUrlIsMalformed(test_model_path())
+        );
+    }
+
+    #[test]
+    fn unsupported_url_scheme_maps_to_download_url_is_malformed() {
+        let error = DownloadError::UnsupportedUrlScheme {
+            url: TEST_URL.to_owned(),
+            scheme: "ftp".to_owned(),
+        };
+
+        assert_eq!(
+            download_error_agent_issue(&error, test_model_path()),
+            AgentIssue::DownloadUrlIsMalformed(test_model_path())
+        );
+    }
+
+    #[test]
+    fn not_found_maps_to_model_does_not_exist_at_url() {
+        let error = DownloadError::NotFound {
+            url: TEST_URL.to_owned(),
+        };
+
+        assert_eq!(
+            download_error_agent_issue(&error, test_model_path()),
+            AgentIssue::ModelDoesNotExistAtUrl(test_model_path())
+        );
+    }
+
+    #[test]
+    fn permission_denied_maps_to_download_server_denied_access() {
+        let error = DownloadError::PermissionDenied {
+            url: TEST_URL.to_owned(),
+            status: StatusCode::FORBIDDEN,
+        };
+
+        assert_eq!(
+            download_error_agent_issue(&error, test_model_path()),
+            AgentIssue::DownloadServerDeniedAccess(test_model_path())
+        );
+    }
+
+    #[test]
+    fn partial_file_stale_maps_to_model_cache_is_corrupted() {
+        let error = DownloadError::PartialFileStale {
+            url: TEST_URL.to_owned(),
+            partial_path: PathBuf::from("/tmp/stale.partial"),
+        };
+
+        assert_eq!(
+            download_error_agent_issue(&error, test_model_path()),
+            AgentIssue::ModelCacheIsCorrupted(test_model_path())
+        );
+    }
+
+    #[tokio::test]
+    async fn download_server_is_unreachable_maps_to_agent_issue() {
+        let error = DownloadError::DownloadServerIsUnreachable {
+            url: TEST_URL.to_owned(),
+            source: refused_connection_error().await,
+        };
+
+        assert_eq!(
+            download_error_agent_issue(&error, test_model_path()),
+            AgentIssue::DownloadServerIsUnreachable(test_model_path())
+        );
+    }
+
+    #[test]
+    fn download_server_errored_maps_to_agent_issue() {
+        let error = DownloadError::DownloadServerErrored {
+            url: TEST_URL.to_owned(),
+            status: StatusCode::INTERNAL_SERVER_ERROR,
+        };
+
+        assert_eq!(
+            download_error_agent_issue(&error, test_model_path()),
+            AgentIssue::DownloadServerErrored(test_model_path())
+        );
+    }
+
+    #[test]
+    fn download_server_rejected_request_maps_to_agent_issue() {
+        let error = DownloadError::DownloadServerRejectedRequest {
+            url: TEST_URL.to_owned(),
+            status: StatusCode::BAD_REQUEST,
+        };
+
+        assert_eq!(
+            download_error_agent_issue(&error, test_model_path()),
+            AgentIssue::DownloadServerRejectedRequest(test_model_path())
+        );
+    }
+
+    #[tokio::test]
+    async fn download_interrupted_maps_to_agent_issue() {
+        let error = DownloadError::DownloadInterrupted {
+            url: TEST_URL.to_owned(),
+            source: refused_connection_error().await,
+        };
+
+        assert_eq!(
+            download_error_agent_issue(&error, test_model_path()),
+            AgentIssue::DownloadInterrupted(test_model_path())
+        );
+    }
+
+    #[test]
+    fn cache_permission_denied_maps_to_cache_directory_is_not_writable() {
+        let error = DownloadError::CachePermissionDenied {
+            path: PathBuf::from("/tmp/locked/model.partial"),
+            source: io::Error::from(io::ErrorKind::PermissionDenied),
+        };
+
+        assert_eq!(
+            download_error_agent_issue(&error, test_model_path()),
+            AgentIssue::CacheDirectoryIsNotWritable(test_model_path())
+        );
+    }
+
+    #[test]
+    fn cache_disk_full_maps_to_cache_storage_is_full() {
+        let error = DownloadError::CacheDiskFull {
+            path: PathBuf::from("/tmp/full/model.partial"),
+            source: io::Error::from(io::ErrorKind::StorageFull),
+        };
+
+        assert_eq!(
+            download_error_agent_issue(&error, test_model_path()),
+            AgentIssue::CacheStorageIsFull(test_model_path())
+        );
+    }
+
+    #[test]
+    fn io_maps_to_model_cache_is_corrupted() {
+        let error = DownloadError::Io {
+            path: PathBuf::from("/tmp/anywhere/model.partial"),
+            source: io::Error::from(io::ErrorKind::NotFound),
+        };
+
+        assert_eq!(
+            download_error_agent_issue(&error, test_model_path()),
+            AgentIssue::ModelCacheIsCorrupted(test_model_path())
+        );
+    }
+}
diff --git a/paddler_model_source/src/huggingface_cache_location.rs b/paddler_model_source/src/huggingface_cache_location.rs
new file mode 100644
index 000000000..99ca15789
--- /dev/null
+++ b/paddler_model_source/src/huggingface_cache_location.rs
@@ -0,0 +1,20 @@
+use std::path::PathBuf;
+
+use hf_hub::HFClient;
+use hf_hub::HFClientBuilder;
+
+#[derive(Clone, Debug)]
+pub enum HuggingFaceCacheLocation {
+    Directory(PathBuf),
+    HubDefault,
+}
+
+impl HuggingFaceCacheLocation {
+    #[must_use]
+    pub fn client_builder(&self) -> HFClientBuilder {
+        match self {
+            Self::Directory(cache_dir) => HFClient::builder().cache_dir(cache_dir.clone()),
+            Self::HubDefault => HFClient::builder(),
+        }
+    }
+}
diff --git a/paddler_model_source/src/huggingface_default_endpoint.rs b/paddler_model_source/src/huggingface_default_endpoint.rs
new file mode 100644
index 000000000..9d30a2efb
--- /dev/null
+++ b/paddler_model_source/src/huggingface_default_endpoint.rs
@@ -0,0 +1 @@
+pub const HUGGINGFACE_DEFAULT_ENDPOINT: &str = "https://huggingface.co";
diff --git a/paddler_model_source/src/huggingface_download_agent_issue.rs b/paddler_model_source/src/huggingface_download_agent_issue.rs
new file mode 100644
index 000000000..1fda01b28
--- /dev/null
+++ b/paddler_model_source/src/huggingface_download_agent_issue.rs
@@ -0,0 +1,147 @@
+use hf_hub::HFError;
+
+use paddler_messaging::agent_issue::AgentIssue;
+use paddler_messaging::agent_issue_params::hugging_face_download_failed_params::HuggingFaceDownloadFailedParams;
+use paddler_messaging::agent_issue_params::hugging_face_download_failure::HuggingFaceDownloadFailure;
+use paddler_messaging::agent_issue_params::hugging_face_download_lock::HuggingFaceDownloadLock;
+use paddler_messaging::agent_issue_params::model_path::ModelPath;
+
+fn hugging_face_download_failure(download_error: &HFError) -> HuggingFaceDownloadFailure {
+    match download_error {
+        HFError::Http { context }
+        | HFError::Conflict { context }
+        | HFError::RateLimited { context, .. } => {
+            HuggingFaceDownloadFailure::HubRespondedWithStatus {
+                status: context.status.as_u16(),
+            }
+        }
+        HFError::Request { .. } => HuggingFaceDownloadFailure::HubConnectionFailed,
+        HFError::Io(_) => HuggingFaceDownloadFailure::CacheIoFailed,
+        HFError::Json(_) | HFError::MalformedResponse { .. } => {
+            HuggingFaceDownloadFailure::HubResponseMalformed
+        }
+        HFError::Xet { .. } => HuggingFaceDownloadFailure::XetTransferFailed,
+        _ => HuggingFaceDownloadFailure::UnexpectedHubClientFailure,
+    }
+}
+
+#[must_use]
+pub fn huggingface_download_agent_issue(
+    download_error: &HFError,
+    model_path: ModelPath,
+) -> AgentIssue {
+    match download_error {
+        HFError::EntryNotFound { .. }
+        | HFError::RepoNotFound { .. }
+        | HFError::RevisionNotFound { .. } => AgentIssue::HuggingFaceModelDoesNotExist(model_path),
+        HFError::AuthRequired { .. } | HFError::Forbidden { .. } => {
+            AgentIssue::HuggingFacePermissions(model_path)
+        }
+        HFError::CacheLockTimeout { path } => {
+            AgentIssue::HuggingFaceCannotAcquireLock(HuggingFaceDownloadLock {
+                lock_path: path.display().to_string(),
+                model_path,
+            })
+        }
+        other_download_error => {
+            AgentIssue::HuggingFaceDownloadFailed(HuggingFaceDownloadFailedParams {
+                failure: hugging_face_download_failure(other_download_error),
+                model_path,
+            })
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::io;
+    use std::io::ErrorKind;
+    use std::path::PathBuf;
+
+    use hf_hub::HFError;
+    use hf_hub::XetOperation;
+    use reqwest::get;
+
+    use paddler_messaging::agent_issue::AgentIssue;
+    use paddler_messaging::agent_issue_params::hugging_face_download_failed_params::HuggingFaceDownloadFailedParams;
+    use paddler_messaging::agent_issue_params::hugging_face_download_failure::HuggingFaceDownloadFailure;
+    use paddler_messaging::agent_issue_params::hugging_face_download_lock::HuggingFaceDownloadLock;
+    use paddler_messaging::agent_issue_params::model_path::ModelPath;
+
+    use super::huggingface_download_agent_issue;
+
+    fn model_path() -> ModelPath {
+        ModelPath {
+            model_path: "owner/repo/model.gguf".to_owned(),
+        }
+    }
+
+    fn download_failure(failure: HuggingFaceDownloadFailure) -> AgentIssue {
+        AgentIssue::HuggingFaceDownloadFailed(HuggingFaceDownloadFailedParams {
+            failure,
+            model_path: model_path(),
+        })
+    }
+
+    #[test]
+    fn a_timed_out_cache_lock_is_reported_with_its_lock_path() {
+        assert_eq!(
+            huggingface_download_agent_issue(
+                &HFError::CacheLockTimeout {
+                    path: PathBuf::from("/cache/.locks/model.lock"),
+                },
+                model_path(),
+            ),
+            AgentIssue::HuggingFaceCannotAcquireLock(HuggingFaceDownloadLock {
+                lock_path: "/cache/.locks/model.lock".to_owned(),
+                model_path: model_path(),
+            })
+        );
+    }
+
+    #[tokio::test]
+    async fn a_hub_the_client_cannot_reach_is_a_failed_connection() {
+        let unreachable_hub_error = get("http://127.0.0.1:1/").await.unwrap_err();
+
+        assert_eq!(
+            huggingface_download_agent_issue(&HFError::from(unreachable_hub_error), model_path()),
+            download_failure(HuggingFaceDownloadFailure::HubConnectionFailed)
+        );
+    }
+
+    #[test]
+    fn a_cache_file_operation_failure_is_a_cache_io_failure() {
+        assert_eq!(
+            huggingface_download_agent_issue(
+                &HFError::Io(io::Error::from(ErrorKind::StorageFull)),
+                model_path(),
+            ),
+            download_failure(HuggingFaceDownloadFailure::CacheIoFailed)
+        );
+    }
+
+    #[test]
+    fn a_failed_xet_transfer_is_reported_as_such() {
+        assert_eq!(
+            huggingface_download_agent_issue(
+                &HFError::xet(
+                    XetOperation::Download,
+                    io::Error::from(ErrorKind::ConnectionReset),
+                ),
+                model_path(),
+            ),
+            download_failure(HuggingFaceDownloadFailure::XetTransferFailed)
+        );
+    }
+
+    #[test]
+    fn a_client_failure_paddler_does_not_expect_is_reported_as_unexpected() {
+        assert_eq!(
+            huggingface_download_agent_issue(
+                &HFError::InvalidParameter("revision".to_owned()),
+                model_path(),
+            ),
+            download_failure(HuggingFaceDownloadFailure::UnexpectedHubClientFailure)
+        );
+    }
+}
diff --git a/paddler_model_source/src/huggingface_download_progress.rs b/paddler_model_source/src/huggingface_download_progress.rs
index 1df1c01b9..bf48ab6c2 100644
--- a/paddler_model_source/src/huggingface_download_progress.rs
+++ b/paddler_model_source/src/huggingface_download_progress.rs
@@ -1,41 +1,69 @@
 use std::sync::Arc;
 
-use hf_hub::api::tokio::Progress;
+use hf_hub::progress::DownloadEvent;
+use hf_hub::progress::FileProgress;
+use hf_hub::progress::ProgressEvent;
+use hf_hub::progress::ProgressHandler;
+use tokio_util::sync::CancellationToken;
 
 use paddler_agent_status::agent_issue_fix::AgentIssueFix;
 use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
 use paddler_messaging::agent_issue_params::model_path::ModelPath;
 use paddler_messaging::model_download_status::ModelDownloadStatus;
 
-#[derive(Clone)]
 pub struct HuggingFaceDownloadProgress {
     pub model_path: ModelPath,
+    pub resolve_ended: CancellationToken,
     pub slot_aggregated_status: Arc<SlotAggregatedStatus>,
 }
 
-impl Progress for HuggingFaceDownloadProgress {
-    async fn init(&mut self, size: usize, _filename_in_repo: &str) {
-        self.slot_aggregated_status
-            .register_fix(&AgentIssueFix::HuggingFaceStartedDownloading(
-                self.model_path.clone(),
-            ));
-
-        self.slot_aggregated_status
-            .set_download_status(ModelDownloadStatus::Downloading {
-                downloaded_bytes: 0,
-                model_path: self.model_path.model_path.clone(),
-                total_bytes: size as u64,
-            });
+impl HuggingFaceDownloadProgress {
+    fn apply(&self, download_event: &DownloadEvent) {
+        match download_event {
+            DownloadEvent::Start { total_bytes, .. } => {
+                self.slot_aggregated_status.register_fix(
+                    &AgentIssueFix::HuggingFaceStartedDownloading(self.model_path.clone()),
+                );
+                self.slot_aggregated_status
+                    .set_download_status(ModelDownloadStatus::Downloading {
+                        downloaded_bytes: 0,
+                        model_path: self.model_path.model_path.clone(),
+                        total_bytes: *total_bytes,
+                    });
+            }
+            DownloadEvent::Progress { files } => {
+                for FileProgress {
+                    bytes_completed, ..
+                } in files
+                {
+                    self.slot_aggregated_status
+                        .set_downloaded_bytes(*bytes_completed);
+                }
+            }
+            DownloadEvent::AggregateProgress {
+                bytes_completed, ..
+            } => {
+                self.slot_aggregated_status
+                    .set_downloaded_bytes(*bytes_completed);
+            }
+            DownloadEvent::Complete => {
+                self.slot_aggregated_status
+                    .set_download_status(ModelDownloadStatus::NotDownloading);
+            }
+        }
     }
+}
 
-    async fn update(&mut self, size: usize) {
-        self.slot_aggregated_status
-            .add_downloaded_bytes(size as u64);
-    }
+impl ProgressHandler for HuggingFaceDownloadProgress {
+    fn on_progress(&self, progress_event: &ProgressEvent) {
+        if self.resolve_ended.is_cancelled() {
+            return;
+        }
 
-    async fn finish(&mut self) {
-        self.slot_aggregated_status
-            .set_download_status(ModelDownloadStatus::NotDownloading);
+        match progress_event {
+            ProgressEvent::Download(download_event) => self.apply(download_event),
+            ProgressEvent::Upload(_upload_event) => {}
+        }
     }
 }
 
@@ -43,7 +71,13 @@ impl Progress for HuggingFaceDownloadProgress {
 mod tests {
     use std::sync::Arc;
 
-    use hf_hub::api::tokio::Progress;
+    use hf_hub::progress::DownloadEvent;
+    use hf_hub::progress::FileProgress;
+    use hf_hub::progress::FileStatus;
+    use hf_hub::progress::ProgressEvent;
+    use hf_hub::progress::ProgressHandler as _;
+    use hf_hub::progress::UploadEvent;
+    use tokio_util::sync::CancellationToken;
 
     use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
     use paddler_messaging::agent_issue::AgentIssue;
@@ -61,74 +95,142 @@ mod tests {
             model_path: ModelPath {
                 model_path: MODEL_PATH.to_owned(),
             },
+            resolve_ended: CancellationToken::new(),
             slot_aggregated_status: Arc::clone(status),
         }
     }
 
-    #[tokio::test]
-    async fn starting_a_download_reports_its_size_and_model_path() {
+    fn download(download_event: DownloadEvent) -> ProgressEvent {
+        ProgressEvent::Download(download_event)
+    }
+
+    fn start_of(total_bytes: u64) -> ProgressEvent {
+        download(DownloadEvent::Start {
+            total_files: 1,
+            total_bytes,
+        })
+    }
+
+    fn file_progress(bytes_completed: u64) -> ProgressEvent {
+        download(DownloadEvent::Progress {
+            files: vec![FileProgress {
+                filename: "model.gguf".to_owned(),
+                bytes_completed,
+                total_bytes: 1000,
+                status: FileStatus::InProgress,
+            }],
+        })
+    }
+
+    fn downloading(downloaded_bytes: u64) -> ModelDownloadStatus {
+        ModelDownloadStatus::Downloading {
+            downloaded_bytes,
+            model_path: MODEL_PATH.to_owned(),
+            total_bytes: 1000,
+        }
+    }
+
+    #[test]
+    fn starting_a_download_reports_its_size_and_model_path() {
         let status = Arc::new(SlotAggregatedStatus::new(2));
-        let mut progress = downloading_progress(&status);
 
-        progress.init(1000, "model.gguf").await;
+        downloading_progress(&status).on_progress(&start_of(1000));
 
         assert_eq!(
             status.make_snapshot().status.download_status,
-            ModelDownloadStatus::Downloading {
-                downloaded_bytes: 0,
-                model_path: MODEL_PATH.to_owned(),
-                total_bytes: 1000,
-            }
+            downloading(0)
         );
     }
 
-    #[tokio::test]
-    async fn starting_a_download_clears_the_lock_issue_of_the_downloaded_model() {
+    #[test]
+    fn starting_a_download_clears_the_lock_issue_of_the_downloaded_model() {
         let status = Arc::new(SlotAggregatedStatus::new(2));
-        let mut progress = downloading_progress(&status);
+        let progress = downloading_progress(&status);
         let lock_issue = AgentIssue::HuggingFaceCannotAcquireLock(HuggingFaceDownloadLock {
             lock_path: "/tmp/lock".to_owned(),
             model_path: progress.model_path.clone(),
         });
 
-        status.register_issue(lock_issue.clone());
+        status.register_desired_state_issue(lock_issue.clone());
 
-        progress.init(1000, "model.gguf").await;
+        progress.on_progress(&start_of(1000));
 
         assert!(!status.has_issue(&lock_issue));
     }
 
-    #[tokio::test]
-    async fn progress_accumulates_the_downloaded_bytes() {
+    #[test]
+    fn file_progress_reports_the_bytes_completed_so_far() {
         let status = Arc::new(SlotAggregatedStatus::new(2));
-        let mut progress = downloading_progress(&status);
+        let progress = downloading_progress(&status);
 
-        progress.init(1000, "model.gguf").await;
-        progress.update(300).await;
-        progress.update(200).await;
+        progress.on_progress(&start_of(1000));
+        progress.on_progress(&file_progress(300));
+        progress.on_progress(&file_progress(500));
 
         assert_eq!(
             status.make_snapshot().status.download_status,
-            ModelDownloadStatus::Downloading {
-                downloaded_bytes: 500,
-                model_path: MODEL_PATH.to_owned(),
-                total_bytes: 1000,
-            }
+            downloading(500)
         );
     }
 
-    #[tokio::test]
-    async fn finishing_a_download_stops_reporting_it() {
+    #[test]
+    fn aggregate_progress_reports_the_bytes_completed_so_far() {
         let status = Arc::new(SlotAggregatedStatus::new(2));
-        let mut progress = downloading_progress(&status);
+        let progress = downloading_progress(&status);
 
-        progress.init(1000, "model.gguf").await;
-        progress.update(1000).await;
-        progress.finish().await;
+        progress.on_progress(&start_of(1000));
+        progress.on_progress(&download(DownloadEvent::AggregateProgress {
+            bytes_completed: 700,
+            total_bytes: 1000,
+            bytes_per_sec: None,
+        }));
+
+        assert_eq!(
+            status.make_snapshot().status.download_status,
+            downloading(700)
+        );
+    }
+
+    #[test]
+    fn completing_a_download_stops_reporting_it() {
+        let status = Arc::new(SlotAggregatedStatus::new(2));
+        let progress = downloading_progress(&status);
+
+        progress.on_progress(&start_of(1000));
+        progress.on_progress(&download(DownloadEvent::Complete));
 
         assert_eq!(
             status.make_snapshot().status.download_status,
             ModelDownloadStatus::NotDownloading
         );
     }
+
+    #[test]
+    fn upload_events_leave_the_download_status_unchanged() {
+        let status = Arc::new(SlotAggregatedStatus::new(2));
+        let progress = downloading_progress(&status);
+
+        progress.on_progress(&start_of(1000));
+        progress.on_progress(&ProgressEvent::Upload(UploadEvent::Committing));
+
+        assert_eq!(
+            status.make_snapshot().status.download_status,
+            downloading(0)
+        );
+    }
+
+    #[test]
+    fn events_after_the_resolve_ended_leave_the_download_status_unchanged() {
+        let status = Arc::new(SlotAggregatedStatus::new(2));
+        let progress = downloading_progress(&status);
+
+        progress.on_progress(&start_of(1000));
+        progress.resolve_ended.cancel();
+        progress.on_progress(&file_progress(300));
+
+        assert_eq!(
+            status.make_snapshot().status.download_status,
+            downloading(0)
+        );
+    }
 }
diff --git a/paddler_model_source/src/huggingface_hub.rs b/paddler_model_source/src/huggingface_hub.rs
new file mode 100644
index 000000000..efaf4e6bb
--- /dev/null
+++ b/paddler_model_source/src/huggingface_hub.rs
@@ -0,0 +1,19 @@
+use hf_hub::HFClient;
+use hf_hub::HFError;
+
+use crate::huggingface_cache_location::HuggingFaceCacheLocation;
+
+#[derive(Clone, Debug)]
+pub struct HuggingFaceHub {
+    pub cache_location: HuggingFaceCacheLocation,
+    pub endpoint: String,
+}
+
+impl HuggingFaceHub {
+    pub fn client(&self) -> Result<HFClient, HFError> {
+        self.cache_location
+            .client_builder()
+            .endpoint(self.endpoint.clone())
+            .build()
+    }
+}
diff --git a/paddler_model_source/src/huggingface_model_reference.rs b/paddler_model_source/src/huggingface_model_reference.rs
new file mode 100644
index 000000000..178aabe13
--- /dev/null
+++ b/paddler_model_source/src/huggingface_model_reference.rs
@@ -0,0 +1,70 @@
+use std::path::PathBuf;
+
+use hf_hub::HFClient;
+use hf_hub::HFRepository;
+use hf_hub::HFResult;
+use hf_hub::RepoTypeModel;
+
+use paddler_messaging::agent_issue_params::model_path::ModelPath;
+
+#[derive(Clone, Debug, Eq, PartialEq)]
+pub struct HuggingFaceModelReference {
+    pub filename: String,
+    pub repository_name: String,
+    pub repository_owner: String,
+    pub revision: String,
+}
+
+impl HuggingFaceModelReference {
+    #[must_use]
+    pub fn model_path(&self) -> ModelPath {
+        let Self {
+            filename,
+            repository_name,
+            repository_owner,
+            revision,
+        } = self;
+
+        ModelPath {
+            model_path: format!("{repository_owner}/{repository_name}/{revision}/{filename}"),
+        }
+    }
+
+    #[must_use]
+    pub fn repository(&self, client: &HFClient) -> HFRepository<RepoTypeModel> {
+        client.model(self.repository_owner.clone(), self.repository_name.clone())
+    }
+
+    pub async fn cached_path(&self, repository: &HFRepository<RepoTypeModel>) -> HFResult<PathBuf> {
+        repository
+            .download_file()
+            .filename(self.filename.clone())
+            .revision(self.revision.clone())
+            .local_files_only(true)
+            .send()
+            .await
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::HuggingFaceModelReference;
+    use paddler_messaging::agent_issue_params::model_path::ModelPath;
+
+    #[test]
+    fn names_the_model_path_after_its_repository_revision_and_filename() {
+        let reference = HuggingFaceModelReference {
+            filename: "model.gguf".to_owned(),
+            repository_name: "repo".to_owned(),
+            repository_owner: "owner".to_owned(),
+            revision: "v2".to_owned(),
+        };
+
+        assert_eq!(
+            reference.model_path(),
+            ModelPath {
+                model_path: "owner/repo/v2/model.gguf".to_owned(),
+            }
+        );
+    }
+}
diff --git a/paddler_model_source/src/huggingface_model_source.rs b/paddler_model_source/src/huggingface_model_source.rs
index 4de504d44..0abe6bdaa 100644
--- a/paddler_model_source/src/huggingface_model_source.rs
+++ b/paddler_model_source/src/huggingface_model_source.rs
@@ -1,13 +1,9 @@
+use std::path::PathBuf;
 use std::sync::Arc;
 
-use async_trait::async_trait;
-use hf_hub::Cache;
-use hf_hub::Repo;
-use hf_hub::RepoType;
-use hf_hub::api::tokio::ApiBuilder;
-use hf_hub::api::tokio::ApiError;
+use hf_hub::HFClient;
+use hf_hub::HFError;
 use log::warn;
-use reqwest::StatusCode;
 use tokio::time::Duration;
 use tokio_util::sync::CancellationToken;
 
@@ -16,35 +12,80 @@ use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
 use paddler_cache_dir::download_lock_wait_outcome::DownloadLockWaitOutcome;
 use paddler_cache_dir::wait_for_download_lock_retry::wait_for_download_lock_retry;
 use paddler_messaging::agent_issue::AgentIssue;
-use paddler_messaging::agent_issue_params::hugging_face_download_lock::HuggingFaceDownloadLock;
-use paddler_messaging::huggingface_model_reference::HuggingFaceModelReference;
+use paddler_messaging::agent_issue_params::model_path::ModelPath;
 use paddler_messaging::model_download_status::ModelDownloadStatus;
 
 use crate::desired_model_resolution::DesiredModelResolution;
+use crate::huggingface_download_agent_issue::huggingface_download_agent_issue;
 use crate::huggingface_download_progress::HuggingFaceDownloadProgress;
+use crate::huggingface_model_reference::HuggingFaceModelReference;
 use crate::model_source_error::ModelSourceError;
 use crate::resolves_model_source::ResolvesModelSource;
 
 const LOCK_RETRY_TIMEOUT: Duration = Duration::from_secs(10);
 
+fn resolved(
+    slot_aggregated_status: &SlotAggregatedStatus,
+    model_path: ModelPath,
+    resolved_path: PathBuf,
+) -> DesiredModelResolution {
+    slot_aggregated_status.set_download_status(ModelDownloadStatus::NotDownloading);
+    slot_aggregated_status.register_fix(&AgentIssueFix::HuggingFaceDownloadedModel(model_path));
+
+    DesiredModelResolution::Resolved(resolved_path)
+}
+
+async fn download_failure_resolution(
+    cancellation_token: &CancellationToken,
+    model_path: ModelPath,
+    download_error: HFError,
+) -> Result<DesiredModelResolution, ModelSourceError> {
+    match download_error {
+        HFError::CacheLockTimeout { path } => {
+            let lock_path = path.display().to_string();
+
+            warn!(
+                "Waiting to acquire download lock for '{lock_path}'. Sleeping for {LOCK_RETRY_TIMEOUT:?}"
+            );
+
+            match wait_for_download_lock_retry(cancellation_token, LOCK_RETRY_TIMEOUT, lock_path)
+                .await
+            {
+                DownloadLockWaitOutcome::Cancelled => Ok(DesiredModelResolution::Cancelled),
+                DownloadLockWaitOutcome::LockStillUnavailable { lock_path } => {
+                    Err(ModelSourceError::HuggingFaceDownloadLockUnavailable { lock_path })
+                }
+            }
+        }
+        HFError::EntryNotFound { .. }
+        | HFError::RepoNotFound { .. }
+        | HFError::RevisionNotFound { .. } => Err(ModelSourceError::HuggingFaceModelMissing {
+            model_path: model_path.model_path,
+        }),
+        HFError::AuthRequired { .. } | HFError::Forbidden { .. } => {
+            Err(ModelSourceError::HuggingFacePermissionDenied {
+                model_path: model_path.model_path,
+            })
+        }
+        source => Err(ModelSourceError::HuggingFaceDownloadFailed {
+            model_path: model_path.model_path,
+            source,
+        }),
+    }
+}
+
 pub struct HuggingFaceModelSource {
-    pub cache: Cache,
+    pub client: HFClient,
     pub reference: HuggingFaceModelReference,
 }
 
-#[async_trait]
 impl ResolvesModelSource for HuggingFaceModelSource {
     async fn resolve(
         &self,
         cancellation_token: &CancellationToken,
         slot_aggregated_status: Arc<SlotAggregatedStatus>,
     ) -> Result<DesiredModelResolution, ModelSourceError> {
-        let Self { cache, reference } = self;
-        let HuggingFaceModelReference {
-            filename,
-            repo_id,
-            revision,
-        } = reference;
+        let reference = &self.reference;
         let model_path = reference.model_path();
 
         if slot_aggregated_status.has_issue(&AgentIssue::HuggingFaceModelDoesNotExist(
@@ -55,93 +96,60 @@ impl ResolvesModelSource for HuggingFaceModelSource {
             });
         }
 
-        let repo = Repo::with_revision(repo_id.to_owned(), RepoType::Model, revision.to_owned());
+        let repository = reference.repository(&self.client);
 
-        if let Some(cached_path) = cache.repo(repo.clone()).get(filename) {
-            slot_aggregated_status.set_download_status(ModelDownloadStatus::NotDownloading);
+        match reference.cached_path(&repository).await {
+            Ok(cached_path) => {
+                return Ok(resolved(&slot_aggregated_status, model_path, cached_path));
+            }
+            Err(HFError::EntryNotFound { .. } | HFError::LocalEntryNotFound { .. }) => {}
+            Err(source) => {
+                slot_aggregated_status.fail_download(huggingface_download_agent_issue(
+                    &source,
+                    model_path.clone(),
+                ));
 
-            return Ok(DesiredModelResolution::Resolved(cached_path));
+                return Err(ModelSourceError::HuggingFaceCacheLookupFailed {
+                    model_path: model_path.model_path,
+                    source,
+                });
+            }
         }
 
-        let hf_repo = ApiBuilder::from_cache(cache.clone())
-            .build()
-            .map_err(ModelSourceError::HuggingFaceClientUnavailable)?
-            .repo(repo);
+        let resolve_ended = CancellationToken::new();
+        let _resolve_ended_on_return = resolve_ended.clone().drop_guard();
         let Some(download_result) = cancellation_token
-            .run_until_cancelled(hf_repo.download_with_progress(
-                filename,
-                HuggingFaceDownloadProgress {
-                    model_path: model_path.clone(),
-                    slot_aggregated_status: slot_aggregated_status.clone(),
-                },
-            ))
+            .run_until_cancelled(
+                repository
+                    .download_file()
+                    .filename(reference.filename.clone())
+                    .revision(reference.revision.clone())
+                    .progress(HuggingFaceDownloadProgress {
+                        model_path: model_path.clone(),
+                        resolve_ended,
+                        slot_aggregated_status: slot_aggregated_status.clone(),
+                    })
+                    .send(),
+            )
             .await
         else {
             return Ok(DesiredModelResolution::Cancelled);
         };
 
         match download_result {
-            Ok(resolved_filename) => {
-                slot_aggregated_status
-                    .register_fix(&AgentIssueFix::HuggingFaceDownloadedModel(model_path));
-
-                Ok(DesiredModelResolution::Resolved(resolved_filename))
-            }
-            Err(ApiError::LockAcquisition(lock_path)) => {
-                let lock_path = lock_path.display().to_string();
-
-                slot_aggregated_status.register_issue(AgentIssue::HuggingFaceCannotAcquireLock(
-                    HuggingFaceDownloadLock {
-                        lock_path: lock_path.clone(),
-                        model_path,
-                    },
+            Ok(downloaded_path) => Ok(resolved(
+                &slot_aggregated_status,
+                model_path,
+                downloaded_path,
+            )),
+            Err(download_error) => {
+                slot_aggregated_status.fail_download(huggingface_download_agent_issue(
+                    &download_error,
+                    model_path.clone(),
                 ));
 
-                warn!(
-                    "Waiting to acquire download lock for '{lock_path}'. Sleeping for {} secs",
-                    LOCK_RETRY_TIMEOUT.as_secs()
-                );
-
-                match wait_for_download_lock_retry(
-                    cancellation_token,
-                    LOCK_RETRY_TIMEOUT,
-                    lock_path,
-                )
-                .await
-                {
-                    DownloadLockWaitOutcome::Cancelled => Ok(DesiredModelResolution::Cancelled),
-                    DownloadLockWaitOutcome::LockStillUnavailable { lock_path } => {
-                        Err(ModelSourceError::HuggingFaceDownloadLockUnavailable { lock_path })
-                    }
-                }
+                download_failure_resolution(cancellation_token, model_path, download_error).await
             }
-            Err(ApiError::RequestError(reqwest_error)) => match reqwest_error.status() {
-                Some(StatusCode::NOT_FOUND) => {
-                    slot_aggregated_status.register_issue(
-                        AgentIssue::HuggingFaceModelDoesNotExist(model_path.clone()),
-                    );
-
-                    Err(ModelSourceError::HuggingFaceModelMissing {
-                        model_path: model_path.model_path,
-                    })
-                }
-                Some(StatusCode::FORBIDDEN | StatusCode::UNAUTHORIZED) => {
-                    slot_aggregated_status
-                        .register_issue(AgentIssue::HuggingFacePermissions(model_path.clone()));
-
-                    Err(ModelSourceError::HuggingFacePermissionDenied {
-                        model_path: model_path.model_path,
-                    })
-                }
-                _other_status => Err(ModelSourceError::HuggingFaceDownloadFailed {
-                    model_path: model_path.model_path,
-                    source: ApiError::RequestError(reqwest_error),
-                }),
-            },
-            Err(source) => Err(ModelSourceError::HuggingFaceDownloadFailed {
-                model_path: model_path.model_path,
-                source,
-            }),
         }
     }
 }
@@ -150,41 +158,136 @@ impl ResolvesModelSource for HuggingFaceModelSource {
 mod tests {
     use std::fs::File;
     use std::fs::create_dir_all;
+    use std::fs::read;
+    use std::fs::write;
+    use std::path::Path;
     use std::path::PathBuf;
     use std::sync::Arc;
 
-    use hf_hub::Cache;
-    use hf_hub::Repo;
-    use hf_hub::RepoType;
+    use hf_hub::HFClient;
+    use hf_hub::HFError;
+    use reqwest::StatusCode;
+    use tempfile::TempDir;
     use tempfile::tempdir;
+    use tokio::join;
     use tokio_util::sync::CancellationToken;
 
     use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
-    use paddler_messaging::huggingface_model_reference::HuggingFaceModelReference;
+    use paddler_local_http_fixture::fixture_response::FixtureResponse;
+    use paddler_local_http_fixture::local_http_fixture::LocalHttpFixture;
+    use paddler_messaging::agent_issue::AgentIssue;
+    use paddler_messaging::agent_issue_params::hugging_face_download_failed_params::HuggingFaceDownloadFailedParams;
+    use paddler_messaging::agent_issue_params::hugging_face_download_failure::HuggingFaceDownloadFailure;
+    use paddler_messaging::agent_issue_params::hugging_face_download_lock::HuggingFaceDownloadLock;
+    use paddler_messaging::model_download_status::ModelDownloadStatus;
+    use paddler_messaging::produces_snapshot::ProducesSnapshot;
+    use paddler_messaging::subscribes_to_updates::SubscribesToUpdates as _;
 
     use super::HuggingFaceModelSource;
     use crate::desired_model_resolution::DesiredModelResolution;
+    use crate::huggingface_cache_location::HuggingFaceCacheLocation;
+    use crate::huggingface_default_endpoint::HUGGINGFACE_DEFAULT_ENDPOINT;
+    use crate::huggingface_hub::HuggingFaceHub;
+    use crate::huggingface_model_reference::HuggingFaceModelReference;
+    use crate::model_source::ModelSource;
+    use crate::model_source_error::ModelSourceError;
     use crate::resolves_model_source::ResolvesModelSource as _;
 
-    fn cache_model_file_at_revision(cache: &Cache, revision: &str, commit_hash: &str) -> PathBuf {
-        let repo = Repo::with_revision(
-            "owner/repo".to_owned(),
-            RepoType::Model,
-            revision.to_owned(),
-        );
+    const HUB_COMMIT_HASH: &str = "hub-commit";
+    const HUB_ETAG: &str = "hub-etag";
+    const HUB_MODEL_BYTES: &[u8] = b"hub model bytes";
+    const REPOSITORY_CACHE_FOLDER: &str = "models--owner--repo";
 
-        cache.repo(repo.clone()).create_ref(commit_hash).unwrap();
+    struct HubCache {
+        cache_dir: TempDir,
+    }
 
-        let snapshot_dir = cache
-            .path()
-            .join(repo.folder_name())
-            .join("snapshots")
-            .join(commit_hash);
+    impl HubCache {
+        fn new() -> Self {
+            Self {
+                cache_dir: tempdir().unwrap(),
+            }
+        }
 
-        create_dir_all(&snapshot_dir).unwrap();
+        fn client_of(&self, hub: &LocalHttpFixture) -> HFClient {
+            HFClient::builder()
+                .cache_dir(self.cache_dir.path())
+                .endpoint(hub.url(""))
+                .build()
+                .unwrap()
+        }
+
+        fn model_source_downloading_from(&self, hub: &LocalHttpFixture) -> HuggingFaceModelSource {
+            HuggingFaceModelSource {
+                client: self.client_of(hub),
+                reference: uncached_model_reference(),
+            }
+        }
+
+        fn blob_lock_path(&self) -> PathBuf {
+            self.cache_dir
+                .path()
+                .join(".locks")
+                .join(REPOSITORY_CACHE_FOLDER)
+                .join(format!("{HUB_ETAG}.lock"))
+        }
+
+        fn hold_blob_lock(&self) -> File {
+            let blob_lock_path = self.blob_lock_path();
+
+            create_dir_all(blob_lock_path.parent().unwrap()).unwrap();
+
+            let held_blob_lock = File::create(blob_lock_path).unwrap();
+
+            held_blob_lock.lock().unwrap();
+
+            held_blob_lock
+        }
+    }
 
-        let model_file = snapshot_dir.join("model.gguf");
+    fn uncached_model_reference() -> HuggingFaceModelReference {
+        HuggingFaceModelReference {
+            filename: "model.gguf".to_owned(),
+            repository_name: "repo".to_owned(),
+            repository_owner: "owner".to_owned(),
+            revision: "main".to_owned(),
+        }
+    }
 
+    fn hub_model_file() -> FixtureResponse {
+        FixtureResponse::HuggingFaceFile {
+            body: HUB_MODEL_BYTES.to_vec(),
+            commit_hash: HUB_COMMIT_HASH.to_owned(),
+            etag: HUB_ETAG.to_owned(),
+        }
+    }
+
+    fn fresh_status() -> Arc<SlotAggregatedStatus> {
+        Arc::new(SlotAggregatedStatus::new(1))
+    }
+
+    fn offline_client_of(cache_dir: &Path) -> HFClient {
+        HuggingFaceHub {
+            cache_location: HuggingFaceCacheLocation::Directory(cache_dir.to_path_buf()),
+            endpoint: HUGGINGFACE_DEFAULT_ENDPOINT.to_owned(),
+        }
+        .client()
+        .unwrap()
+    }
+
+    fn cache_model_file_at_revision(
+        cache_dir: &Path,
+        revision: &str,
+        commit_hash: &str,
+        filename: &str,
+    ) -> PathBuf {
+        let repository_cache = cache_dir.join(REPOSITORY_CACHE_FOLDER);
+        let snapshot_dir = repository_cache.join("snapshots").join(commit_hash);
+        let model_file = snapshot_dir.join(filename);
+
+        create_dir_all(repository_cache.join("refs")).unwrap();
+        write(repository_cache.join("refs").join(revision), commit_hash).unwrap();
+        create_dir_all(&snapshot_dir).unwrap();
         File::create(&model_file).unwrap();
 
         model_file
@@ -193,29 +296,400 @@ mod tests {
     #[tokio::test]
     async fn resolves_the_cached_file_of_the_requested_revision() {
         let cache_dir = tempdir().unwrap();
-        let cache = Cache::new(cache_dir.path().to_path_buf());
 
-        cache_model_file_at_revision(&cache, "main", "main-commit");
+        cache_model_file_at_revision(cache_dir.path(), "main", "main-commit", "model.gguf");
 
-        let requested_revision_file = cache_model_file_at_revision(&cache, "v2", "v2-commit");
+        let requested_revision_file =
+            cache_model_file_at_revision(cache_dir.path(), "v2", "v2-commit", "model.gguf");
         let resolution = HuggingFaceModelSource {
-            cache,
+            client: offline_client_of(cache_dir.path()),
             reference: HuggingFaceModelReference {
-                filename: "model.gguf".to_owned(),
-                repo_id: "owner/repo".to_owned(),
                 revision: "v2".to_owned(),
+                ..uncached_model_reference()
             },
         }
-        .resolve(
-            &CancellationToken::new(),
-            Arc::new(SlotAggregatedStatus::new(1)),
+        .resolve(&CancellationToken::new(), fresh_status())
+        .await
+        .unwrap();
+
+        assert_eq!(
+            resolution,
+            DesiredModelResolution::Resolved(requested_revision_file)
+        );
+    }
+
+    #[tokio::test]
+    async fn resolves_a_cached_file_from_a_uri_that_percent_encodes_its_name() {
+        let cache_dir = tempdir().unwrap();
+        let cached_file =
+            cache_model_file_at_revision(cache_dir.path(), "main", "main-commit", "my modèle.gguf");
+        let reference = HuggingFaceModelReference {
+            filename: "my modèle.gguf".to_owned(),
+            ..uncached_model_reference()
+        };
+
+        assert_eq!(
+            ModelSource::parse(
+                "https://huggingface.co/owner/repo/resolve/main/my%20mod%C3%A8le.gguf",
+                &SlotAggregatedStatus::new(1),
+            )
+            .unwrap(),
+            ModelSource::HuggingFace(reference.clone())
+        );
+        assert_eq!(
+            HuggingFaceModelSource {
+                client: offline_client_of(cache_dir.path()),
+                reference,
+            }
+            .resolve(&CancellationToken::new(), fresh_status())
+            .await
+            .unwrap(),
+            DesiredModelResolution::Resolved(cached_file)
+        );
+    }
+
+    #[tokio::test]
+    async fn a_client_without_a_cache_cannot_look_up_cached_models() {
+        let slot_aggregated_status = fresh_status();
+        let resolution = HuggingFaceModelSource {
+            client: HFClient::builder().cache_enabled(false).build().unwrap(),
+            reference: uncached_model_reference(),
+        }
+        .resolve(&CancellationToken::new(), slot_aggregated_status.clone())
+        .await;
+
+        assert_eq!(
+            resolution
+                .err()
+                .map(|resolution_error| resolution_error.to_string()),
+            Some(
+                ModelSourceError::HuggingFaceCacheLookupFailed {
+                    model_path: uncached_model_reference().model_path().model_path,
+                    source: HFError::CacheNotEnabled,
+                }
+                .to_string()
+            )
+        );
+        assert!(
+            slot_aggregated_status.has_issue(&AgentIssue::HuggingFaceDownloadFailed(
+                HuggingFaceDownloadFailedParams {
+                    failure: HuggingFaceDownloadFailure::UnexpectedHubClientFailure,
+                    model_path: uncached_model_reference().model_path(),
+                }
+            ))
+        );
+    }
+
+    #[tokio::test]
+    async fn downloads_an_uncached_file_from_its_hub_into_the_cache() {
+        let hub_cache = HubCache::new();
+        let hub = LocalHttpFixture::start(hub_model_file()).await.unwrap();
+        let slot_aggregated_status = fresh_status();
+
+        let resolution = hub_cache
+            .model_source_downloading_from(&hub)
+            .resolve(&CancellationToken::new(), slot_aggregated_status.clone())
+            .await
+            .unwrap();
+        let cached_path = uncached_model_reference()
+            .cached_path(&uncached_model_reference().repository(&hub_cache.client_of(&hub)))
+            .await
+            .unwrap();
+
+        assert_eq!(
+            resolution,
+            DesiredModelResolution::Resolved(cached_path.clone())
+        );
+        assert_eq!(read(cached_path).unwrap(), HUB_MODEL_BYTES);
+        assert_eq!(
+            slot_aggregated_status
+                .make_snapshot()
+                .status
+                .download_status,
+            ModelDownloadStatus::NotDownloading
+        );
+    }
+
+    #[tokio::test]
+    async fn a_file_missing_from_its_hub_is_not_requested_again() {
+        let hub_cache = HubCache::new();
+        let hub = LocalHttpFixture::start(FixtureResponse::Status(StatusCode::NOT_FOUND))
+            .await
+            .unwrap();
+        let model_source = hub_cache.model_source_downloading_from(&hub);
+        let slot_aggregated_status = fresh_status();
+        let model_path = uncached_model_reference().model_path().model_path;
+
+        let first_resolution = model_source
+            .resolve(&CancellationToken::new(), slot_aggregated_status.clone())
+            .await;
+        let second_resolution = model_source
+            .resolve(&CancellationToken::new(), slot_aggregated_status)
+            .await;
+
+        assert!(matches!(
+            first_resolution,
+            Err(ModelSourceError::HuggingFaceModelMissing { model_path: missing_model_path })
+                if missing_model_path == model_path
+        ));
+        assert!(matches!(
+            second_resolution,
+            Err(ModelSourceError::HuggingFaceModelKnownToBeMissing { model_path: missing_model_path })
+                if missing_model_path == model_path
+        ));
+        assert_eq!(hub.request_count(), 1);
+    }
+
+    #[tokio::test]
+    async fn a_file_its_hub_refuses_to_share_reports_missing_permissions() {
+        let hub_cache = HubCache::new();
+        let hub = LocalHttpFixture::start(FixtureResponse::Status(StatusCode::FORBIDDEN))
+            .await
+            .unwrap();
+        let slot_aggregated_status = fresh_status();
+
+        let resolution = hub_cache
+            .model_source_downloading_from(&hub)
+            .resolve(&CancellationToken::new(), slot_aggregated_status.clone())
+            .await;
+
+        assert_eq!(
+            resolution
+                .err()
+                .map(|resolution_error| resolution_error.to_string()),
+            Some(
+                ModelSourceError::HuggingFacePermissionDenied {
+                    model_path: uncached_model_reference().model_path().model_path,
+                }
+                .to_string()
+            )
+        );
+        assert!(
+            slot_aggregated_status.has_issue(&AgentIssue::HuggingFacePermissions(
+                uncached_model_reference().model_path()
+            ))
+        );
+    }
+
+    async fn assert_hub_status_fails_the_download(status: StatusCode, hub_failure: &str) {
+        let hub_cache = HubCache::new();
+        let hub = LocalHttpFixture::start(FixtureResponse::Status(status))
+            .await
+            .unwrap();
+        let slot_aggregated_status = fresh_status();
+
+        let resolution = hub_cache
+            .model_source_downloading_from(&hub)
+            .resolve(&CancellationToken::new(), slot_aggregated_status.clone())
+            .await;
+
+        assert_eq!(
+            resolution
+                .err()
+                .map(|resolution_error| resolution_error.to_string()),
+            Some(format!(
+                "Failed to download model '{}' from Hugging Face: {hub_failure} (url={})",
+                uncached_model_reference().model_path().model_path,
+                hub.url("/owner/repo/resolve/main/model.gguf")
+            ))
+        );
+        assert!(
+            slot_aggregated_status.has_issue(&AgentIssue::HuggingFaceDownloadFailed(
+                HuggingFaceDownloadFailedParams {
+                    failure: HuggingFaceDownloadFailure::HubRespondedWithStatus {
+                        status: status.as_u16(),
+                    },
+                    model_path: uncached_model_reference().model_path(),
+                }
+            ))
+        );
+    }
+
+    #[tokio::test]
+    async fn a_hub_failing_with_a_server_error_fails_the_download() {
+        assert_hub_status_fails_the_download(
+            StatusCode::INTERNAL_SERVER_ERROR,
+            "HTTP error: 500 Internal Server Error",
         )
+        .await;
+    }
+
+    #[tokio::test]
+    async fn a_hub_reporting_a_conflict_fails_the_download() {
+        assert_hub_status_fails_the_download(StatusCode::CONFLICT, "Conflict").await;
+    }
+
+    #[tokio::test]
+    async fn a_hub_rate_limiting_the_client_fails_the_download() {
+        assert_hub_status_fails_the_download(StatusCode::TOO_MANY_REQUESTS, "Rate limited").await;
+    }
+
+    #[tokio::test]
+    async fn a_download_failing_after_it_started_stops_reporting_download_progress() {
+        let hub_cache = HubCache::new();
+        let hub = LocalHttpFixture::start(FixtureResponse::TruncatedHuggingFaceFile {
+            commit_hash: HUB_COMMIT_HASH.to_owned(),
+            etag: HUB_ETAG.to_owned(),
+            sent_body: HUB_MODEL_BYTES[..4].to_vec(),
+            withheld_byte_count: HUB_MODEL_BYTES.len() - 4,
+        })
+        .await
+        .unwrap();
+        let slot_aggregated_status = fresh_status();
+
+        let resolution = hub_cache
+            .model_source_downloading_from(&hub)
+            .resolve(&CancellationToken::new(), slot_aggregated_status.clone())
+            .await;
+
+        assert_eq!(
+            resolution
+                .err()
+                .map(|resolution_error| resolution_error.to_string()),
+            Some(format!(
+                "Failed to download model '{}' from Hugging Face: HTTP request error: error decoding response body",
+                uncached_model_reference().model_path().model_path
+            ))
+        );
+        assert_eq!(
+            slot_aggregated_status
+                .make_snapshot()
+                .status
+                .download_status,
+            ModelDownloadStatus::NotDownloading
+        );
+        assert!(
+            slot_aggregated_status.has_issue(&AgentIssue::HuggingFaceDownloadFailed(
+                HuggingFaceDownloadFailedParams {
+                    failure: HuggingFaceDownloadFailure::HubConnectionFailed,
+                    model_path: uncached_model_reference().model_path(),
+                }
+            ))
+        );
+    }
+
+    #[tokio::test]
+    async fn a_cached_model_clears_the_lock_issue_of_an_earlier_attempt() {
+        let cache_dir = tempdir().unwrap();
+        let cached_file =
+            cache_model_file_at_revision(cache_dir.path(), "main", "main-commit", "model.gguf");
+        let slot_aggregated_status = fresh_status();
+        let lock_issue = AgentIssue::HuggingFaceCannotAcquireLock(HuggingFaceDownloadLock {
+            lock_path: "/cache/.locks/model.lock".to_owned(),
+            model_path: uncached_model_reference().model_path(),
+        });
+
+        slot_aggregated_status.register_desired_state_issue(lock_issue.clone());
+
+        let resolution = HuggingFaceModelSource {
+            client: offline_client_of(cache_dir.path()),
+            reference: uncached_model_reference(),
+        }
+        .resolve(&CancellationToken::new(), slot_aggregated_status.clone())
         .await
         .unwrap();
 
+        assert_eq!(resolution, DesiredModelResolution::Resolved(cached_file));
+        assert!(!slot_aggregated_status.has_issue(&lock_issue));
+    }
+
+    #[tokio::test]
+    async fn a_hub_answer_without_file_metadata_fails_the_download() {
+        let hub_cache = HubCache::new();
+        let hub = LocalHttpFixture::start(FixtureResponse::Ok(HUB_MODEL_BYTES.to_vec()))
+            .await
+            .unwrap();
+
+        let resolution = hub_cache
+            .model_source_downloading_from(&hub)
+            .resolve(&CancellationToken::new(), fresh_status())
+            .await;
+
+        assert!(matches!(
+            resolution,
+            Err(ModelSourceError::HuggingFaceDownloadFailed {
+                source: HFError::MalformedResponse { what, .. },
+                model_path,
+            }) if what == "missing ETag header"
+                && model_path == uncached_model_reference().model_path().model_path
+        ));
+    }
+
+    #[tokio::test]
+    async fn a_blob_locked_by_another_process_stays_unavailable_after_the_lock_retry_wait() {
+        let hub_cache = HubCache::new();
+        let hub = LocalHttpFixture::start(hub_model_file()).await.unwrap();
+        let _held_blob_lock = hub_cache.hold_blob_lock();
+        let slot_aggregated_status = fresh_status();
+        let lock_path = hub_cache.blob_lock_path().display().to_string();
+
+        let resolution = hub_cache
+            .model_source_downloading_from(&hub)
+            .resolve(&CancellationToken::new(), slot_aggregated_status.clone())
+            .await;
+
         assert!(matches!(
             resolution,
-            DesiredModelResolution::Resolved(resolved_file) if resolved_file == requested_revision_file
+            Err(ModelSourceError::HuggingFaceDownloadLockUnavailable { lock_path: unavailable_lock_path })
+                if unavailable_lock_path == lock_path
         ));
+        assert!(
+            slot_aggregated_status.has_issue(&AgentIssue::HuggingFaceCannotAcquireLock(
+                HuggingFaceDownloadLock {
+                    lock_path,
+                    model_path: uncached_model_reference().model_path(),
+                }
+            ))
+        );
+    }
+
+    #[tokio::test]
+    async fn a_resolution_cancelled_while_waiting_for_a_locked_blob_is_cancelled() {
+        let hub_cache = HubCache::new();
+        let hub = LocalHttpFixture::start(hub_model_file()).await.unwrap();
+        let model_source = hub_cache.model_source_downloading_from(&hub);
+        let _held_blob_lock = hub_cache.hold_blob_lock();
+        let slot_aggregated_status = fresh_status();
+        let mut status_updates = slot_aggregated_status.subscribe_to_updates();
+        let cancellation_token = CancellationToken::new();
+
+        let (resolution, ()) = join!(
+            model_source.resolve(&cancellation_token, slot_aggregated_status.clone()),
+            async {
+                status_updates
+                    .wait_for(|()| {
+                        slot_aggregated_status.has_issue(&AgentIssue::HuggingFaceCannotAcquireLock(
+                            HuggingFaceDownloadLock {
+                                lock_path: hub_cache.blob_lock_path().display().to_string(),
+                                model_path: uncached_model_reference().model_path(),
+                            },
+                        ))
+                    })
+                    .await
+                    .unwrap();
+                cancellation_token.cancel();
+            }
+        );
+
+        assert_eq!(resolution.unwrap(), DesiredModelResolution::Cancelled);
+    }
+
+    #[tokio::test]
+    async fn a_cancelled_resolution_of_an_uncached_model_requests_nothing_from_its_hub() {
+        let hub_cache = HubCache::new();
+        let hub = LocalHttpFixture::start(hub_model_file()).await.unwrap();
+        let cancellation_token = CancellationToken::new();
+
+        cancellation_token.cancel();
+
+        assert_eq!(
+            hub_cache
+                .model_source_downloading_from(&hub)
+                .resolve(&cancellation_token, fresh_status())
+                .await
+                .unwrap(),
+            DesiredModelResolution::Cancelled
+        );
+        assert_eq!(hub.request_count(), 0);
     }
 }
diff --git a/paddler_model_source/src/lib.rs b/paddler_model_source/src/lib.rs
index 9811dd9ff..c69d0c6a2 100644
--- a/paddler_model_source/src/lib.rs
+++ b/paddler_model_source/src/lib.rs
@@ -1,9 +1,18 @@
+pub mod agent_local_model_path;
 pub mod desired_model_resolution;
+pub mod desired_model_resolver;
+pub mod download_error_agent_issue;
+pub mod huggingface_cache_location;
+pub mod huggingface_default_endpoint;
+pub mod huggingface_download_agent_issue;
 pub mod huggingface_download_progress;
+pub mod huggingface_hub;
+pub mod huggingface_model_reference;
 pub mod huggingface_model_source;
 pub mod local_model_path;
+pub mod model_source;
 pub mod model_source_error;
-pub mod resolve_desired_model;
 pub mod resolves_model_source;
 pub mod url_download_progress;
 pub mod url_model_source;
+pub mod url_path_segment_encode_set;
diff --git a/paddler_model_source/src/local_model_path.rs b/paddler_model_source/src/local_model_path.rs
index c23b30fc8..cf8f80817 100644
--- a/paddler_model_source/src/local_model_path.rs
+++ b/paddler_model_source/src/local_model_path.rs
@@ -1,43 +1,103 @@
 use std::path::PathBuf;
 use std::sync::Arc;
 
-use async_trait::async_trait;
 use tokio::fs::try_exists;
 use tokio_util::sync::CancellationToken;
 
 use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
+use paddler_messaging::agent_issue::AgentIssue;
+use paddler_messaging::agent_issue_params::model_path::ModelPath;
 
 use crate::desired_model_resolution::DesiredModelResolution;
 use crate::model_source_error::ModelSourceError;
 use crate::resolves_model_source::ResolvesModelSource;
 
 pub struct LocalModelPath {
-    pub path: String,
+    pub path: PathBuf,
 }
 
-impl LocalModelPath {
-    #[must_use]
-    pub const fn new(path: String) -> Self {
-        Self { path }
-    }
-}
-
-#[async_trait]
 impl ResolvesModelSource for LocalModelPath {
     async fn resolve(
         &self,
         _cancellation_token: &CancellationToken,
-        _slot_aggregated_status: Arc<SlotAggregatedStatus>,
+        slot_aggregated_status: Arc<SlotAggregatedStatus>,
     ) -> Result<DesiredModelResolution, ModelSourceError> {
-        let local_path = PathBuf::from(&self.path);
-
-        match try_exists(&local_path).await {
-            Ok(true) => Ok(DesiredModelResolution::Resolved(local_path)),
-            Ok(false) => Ok(DesiredModelResolution::LocalFileMissing(local_path)),
-            Err(source) => Err(ModelSourceError::LocalFileUncheckable {
-                path: local_path,
-                source,
-            }),
+        match try_exists(&self.path).await {
+            Ok(true) => Ok(DesiredModelResolution::Resolved(self.path.clone())),
+            Ok(false) => Ok(DesiredModelResolution::LocalFileMissing(self.path.clone())),
+            Err(source) => {
+                slot_aggregated_status.register_desired_state_issue(
+                    AgentIssue::LocalModelFileCannotBeChecked(ModelPath::from(self.path.as_path())),
+                );
+
+                Err(ModelSourceError::LocalFileUncheckable {
+                    path: self.path.clone(),
+                    source,
+                })
+            }
         }
     }
 }
+
+#[cfg(all(test, unix))]
+mod tests {
+    use std::fs::metadata;
+    use std::sync::Arc;
+
+    use tempfile::NamedTempFile;
+    use tokio_util::sync::CancellationToken;
+
+    use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
+    use paddler_messaging::agent_issue::AgentIssue;
+    use paddler_messaging::agent_issue_params::model_path::ModelPath;
+
+    use super::LocalModelPath;
+    use crate::model_source_error::ModelSourceError;
+    use crate::resolves_model_source::ResolvesModelSource as _;
+
+    #[tokio::test]
+    async fn a_model_path_below_a_regular_file_cannot_be_checked() {
+        let regular_file = NamedTempFile::new().unwrap();
+        let path_below_the_regular_file = regular_file.path().join("model.gguf");
+
+        assert_eq!(
+            LocalModelPath {
+                path: path_below_the_regular_file.clone(),
+            }
+            .resolve(
+                &CancellationToken::new(),
+                Arc::new(SlotAggregatedStatus::new(1))
+            )
+            .await
+            .err()
+            .map(|resolution_error| resolution_error.to_string()),
+            Some(
+                ModelSourceError::LocalFileUncheckable {
+                    path: path_below_the_regular_file.clone(),
+                    source: metadata(&path_below_the_regular_file).unwrap_err(),
+                }
+                .to_string()
+            )
+        );
+    }
+
+    #[tokio::test]
+    async fn a_model_path_that_cannot_be_checked_is_reported_as_an_agent_issue() {
+        let regular_file = NamedTempFile::new().unwrap();
+        let path_below_the_regular_file = regular_file.path().join("model.gguf");
+        let slot_aggregated_status = Arc::new(SlotAggregatedStatus::new(1));
+
+        LocalModelPath {
+            path: path_below_the_regular_file.clone(),
+        }
+        .resolve(&CancellationToken::new(), slot_aggregated_status.clone())
+        .await
+        .unwrap_err();
+
+        assert!(
+            slot_aggregated_status.has_issue(&AgentIssue::LocalModelFileCannotBeChecked(
+                ModelPath::from(path_below_the_regular_file.as_path())
+            ))
+        );
+    }
+}
diff --git a/paddler_model_source/src/model_source.rs b/paddler_model_source/src/model_source.rs
new file mode 100644
index 000000000..c3e98524b
--- /dev/null
+++ b/paddler_model_source/src/model_source.rs
@@ -0,0 +1,395 @@
+use percent_encoding::percent_decode_str;
+use percent_encoding::percent_encode;
+use url::Url;
+
+use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
+use paddler_download_manager::download_url::DownloadUrl;
+use paddler_messaging::agent_desired_model::AgentDesiredModel;
+use paddler_messaging::agent_issue::AgentIssue;
+use paddler_messaging::agent_issue_params::model_path::ModelPath;
+
+use crate::agent_local_model_path::AgentLocalModelPath;
+use crate::download_error_agent_issue::download_error_agent_issue;
+use crate::huggingface_model_reference::HuggingFaceModelReference;
+use crate::model_source_error::ModelSourceError;
+use crate::url_path_segment_encode_set::URL_PATH_SEGMENT_ENCODE_SET;
+
+const HUGGING_FACE_BLOB: &str = "blob";
+const HUGGING_FACE_HOST: &str = "huggingface.co";
+const HUGGING_FACE_RESOLVE: &str = "resolve";
+
+fn huggingface_model_reference(
+    uri: &str,
+    url: &Url,
+) -> Result<HuggingFaceModelReference, ModelSourceError> {
+    let path_segments = url
+        .path()
+        .split('/')
+        .filter(|path_segment| !path_segment.is_empty())
+        .map(|path_segment| {
+            percent_decode_str(path_segment)
+                .decode_utf8()
+                .map_err(
+                    |source| ModelSourceError::HuggingFaceModelUriComponentNotUtf8 {
+                        source,
+                        uri: uri.to_owned(),
+                    },
+                )
+        })
+        .collect::<Result<Vec<_>, _>>()?;
+
+    match path_segments.as_slice() {
+        [owner, repo, file_kind, revision, filename_segments @ ..]
+            if (file_kind == HUGGING_FACE_BLOB || file_kind == HUGGING_FACE_RESOLVE)
+                && !filename_segments.is_empty() =>
+        {
+            Ok(HuggingFaceModelReference {
+                filename: filename_segments.join("/"),
+                repository_name: repo.to_string(),
+                repository_owner: owner.to_string(),
+                revision: revision.to_string(),
+            })
+        }
+        _ => Err(ModelSourceError::HuggingFaceModelUriMalformed {
+            uri: uri.to_owned(),
+        }),
+    }
+}
+
+fn percent_encoded_path_segment(path_segment: &str) -> String {
+    percent_encode(path_segment.as_bytes(), URL_PATH_SEGMENT_ENCODE_SET).to_string()
+}
+
+fn percent_encoded_path(path: &str) -> String {
+    path.split('/')
+        .map(percent_encoded_path_segment)
+        .collect::<Vec<_>>()
+        .join("/")
+}
+
+#[derive(Debug, Eq, PartialEq)]
+pub enum ModelSource {
+    HuggingFace(HuggingFaceModelReference),
+    LocalToAgent(AgentLocalModelPath),
+    Url(DownloadUrl),
+}
+
+impl ModelSource {
+    pub fn parse(
+        uri: &str,
+        slot_aggregated_status: &SlotAggregatedStatus,
+    ) -> Result<Self, ModelSourceError> {
+        let model_path = || ModelPath {
+            model_path: uri.to_owned(),
+        };
+        let url = Url::parse(uri).map_err(|source| {
+            slot_aggregated_status
+                .register_desired_state_issue(AgentIssue::ModelUriIsUnparseable(model_path()));
+
+            ModelSourceError::ModelUriUnparseable {
+                source,
+                uri: uri.to_owned(),
+            }
+        })?;
+
+        if url.host_str() == Some(HUGGING_FACE_HOST) {
+            return huggingface_model_reference(uri, &url)
+                .map(Self::HuggingFace)
+                .inspect_err(|_malformed_uri| {
+                    slot_aggregated_status.register_desired_state_issue(
+                        AgentIssue::HuggingFaceModelUriIsMalformed(model_path()),
+                    );
+                });
+        }
+
+        if url.scheme() == AgentLocalModelPath::SCHEME {
+            return AgentLocalModelPath::from_agent_url(uri, &url, slot_aggregated_status)
+                .map(Self::LocalToAgent);
+        }
+
+        DownloadUrl::try_from(url)
+            .map(Self::Url)
+            .map_err(|download_error| {
+                slot_aggregated_status.register_desired_state_issue(download_error_agent_issue(
+                    &download_error,
+                    model_path(),
+                ));
+
+                ModelSourceError::ModelUriNotDownloadable {
+                    source: download_error,
+                    uri: uri.to_owned(),
+                }
+            })
+    }
+
+    #[must_use]
+    pub fn uri(&self) -> String {
+        match self {
+            Self::HuggingFace(HuggingFaceModelReference {
+                filename,
+                repository_name,
+                repository_owner,
+                revision,
+            }) => format!(
+                "https://{HUGGING_FACE_HOST}/{}/{}/{HUGGING_FACE_BLOB}/{}/{}",
+                percent_encoded_path_segment(repository_owner),
+                percent_encoded_path_segment(repository_name),
+                percent_encoded_path_segment(revision),
+                percent_encoded_path(filename)
+            ),
+            Self::LocalToAgent(agent_local_model_path) => {
+                agent_local_model_path.canonical_uri().to_owned()
+            }
+            Self::Url(download_url) => download_url.as_str().to_owned(),
+        }
+    }
+
+    #[must_use]
+    pub fn into_agent_desired_model(self) -> AgentDesiredModel {
+        AgentDesiredModel::Uri(self.uri())
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::error::Error as _;
+    use std::path::Path;
+
+    use url::ParseError;
+
+    use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
+    use paddler_download_manager::download_error::DownloadError;
+    use paddler_download_manager::download_url::DownloadUrl;
+    use paddler_messaging::agent_issue::AgentIssue;
+    use paddler_messaging::agent_issue_params::model_path::ModelPath;
+
+    use super::ModelSource;
+    use crate::agent_local_model_path::AgentLocalModelPath;
+    use crate::huggingface_model_reference::HuggingFaceModelReference;
+    use crate::model_source_error::ModelSourceError;
+
+    fn model_path(uri: &str) -> ModelPath {
+        ModelPath {
+            model_path: uri.to_owned(),
+        }
+    }
+
+    #[test]
+    fn parses_hugging_face_file_uris() {
+        let slot_aggregated_status = SlotAggregatedStatus::new(1);
+
+        for (uri, filename) in [
+            (
+                "https://huggingface.co/unsloth/Qwen3-0.6B-GGUF/blob/main/Qwen3-0.6B-Q4_K_M.gguf",
+                "Qwen3-0.6B-Q4_K_M.gguf",
+            ),
+            (
+                "https://huggingface.co/unsloth/Qwen3-0.6B-GGUF/resolve/main/nested/model.gguf",
+                "nested/model.gguf",
+            ),
+        ] {
+            assert_eq!(
+                ModelSource::parse(uri, &slot_aggregated_status).unwrap(),
+                ModelSource::HuggingFace(HuggingFaceModelReference {
+                    filename: filename.to_owned(),
+                    repository_name: "Qwen3-0.6B-GGUF".to_owned(),
+                    repository_owner: "unsloth".to_owned(),
+                    revision: "main".to_owned(),
+                })
+            );
+        }
+    }
+
+    #[test]
+    fn decodes_the_percent_encoded_components_of_hugging_face_uris() {
+        let slot_aggregated_status = SlotAggregatedStatus::new(1);
+
+        for (uri, filename, revision) in [
+            (
+                "https://huggingface.co/owner/repo/resolve/main/my%20model.gguf",
+                "my model.gguf",
+                "main",
+            ),
+            (
+                "https://huggingface.co/owner/repo/blob/main/nested/mod%C3%A8le.gguf",
+                "nested/modèle.gguf",
+                "main",
+            ),
+            (
+                "https://huggingface.co/owner/repo/resolve/refs%2Fpr%2F1/model.gguf",
+                "model.gguf",
+                "refs/pr/1",
+            ),
+        ] {
+            assert_eq!(
+                ModelSource::parse(uri, &slot_aggregated_status).unwrap(),
+                ModelSource::HuggingFace(HuggingFaceModelReference {
+                    filename: filename.to_owned(),
+                    repository_name: "repo".to_owned(),
+                    repository_owner: "owner".to_owned(),
+                    revision: revision.to_owned(),
+                })
+            );
+        }
+    }
+
+    #[test]
+    fn rejects_hugging_face_uris_whose_components_do_not_decode_to_utf8() {
+        let slot_aggregated_status = SlotAggregatedStatus::new(1);
+        let uri = "https://huggingface.co/owner/repo/blob/main/model%FF.gguf";
+        let parse_error = ModelSource::parse(uri, &slot_aggregated_status)
+            .expect_err("a component that is not utf-8 must be rejected");
+
+        assert_eq!(
+            (
+                parse_error.to_string(),
+                parse_error.source().map(ToString::to_string)
+            ),
+            (
+                format!(
+                    "Hugging Face model URI '{uri}' has a path component that is not valid UTF-8 once percent-decoded: invalid utf-8 sequence of 1 bytes from index 5"
+                ),
+                Some("invalid utf-8 sequence of 1 bytes from index 5".to_owned())
+            )
+        );
+        assert!(
+            slot_aggregated_status
+                .has_issue(&AgentIssue::HuggingFaceModelUriIsMalformed(model_path(uri)))
+        );
+    }
+
+    #[test]
+    fn rejects_hugging_face_uris_that_do_not_point_at_a_file() {
+        for uri in [
+            "https://huggingface.co/owner/repo",
+            "https://huggingface.co/owner/repo/tree/main/model.gguf",
+            "https://huggingface.co/owner/repo/blob/main",
+        ] {
+            let slot_aggregated_status = SlotAggregatedStatus::new(1);
+
+            assert_eq!(
+                ModelSource::parse(uri, &slot_aggregated_status)
+                    .err()
+                    .as_ref()
+                    .map(ToString::to_string),
+                Some(
+                    ModelSourceError::HuggingFaceModelUriMalformed {
+                        uri: uri.to_owned(),
+                    }
+                    .to_string()
+                )
+            );
+            assert!(
+                slot_aggregated_status
+                    .has_issue(&AgentIssue::HuggingFaceModelUriIsMalformed(model_path(uri)))
+            );
+        }
+    }
+
+    #[test]
+    fn parses_an_agent_uri_as_a_path_local_to_the_agent() {
+        assert_eq!(
+            ModelSource::parse("agent:///models/model.gguf", &SlotAggregatedStatus::new(1))
+                .unwrap(),
+            ModelSource::LocalToAgent(
+                AgentLocalModelPath::from_absolute_path(Path::new("/models/model.gguf")).unwrap()
+            )
+        );
+    }
+
+    #[test]
+    fn parses_an_http_uri_as_a_download() {
+        assert_eq!(
+            ModelSource::parse(
+                "https://example.com/models/model.gguf",
+                &SlotAggregatedStatus::new(1)
+            )
+            .unwrap(),
+            ModelSource::Url(DownloadUrl::parse("https://example.com/models/model.gguf").unwrap())
+        );
+    }
+
+    #[test]
+    fn formats_each_source_as_its_canonical_uri() {
+        assert_eq!(
+            [
+                ModelSource::HuggingFace(HuggingFaceModelReference {
+                    filename: "nested/my model.gguf".to_owned(),
+                    repository_name: "repo".to_owned(),
+                    repository_owner: "owner".to_owned(),
+                    revision: "refs/pr/1".to_owned(),
+                }),
+                ModelSource::LocalToAgent(
+                    AgentLocalModelPath::from_absolute_path(Path::new("/models/model.gguf"))
+                        .unwrap(),
+                ),
+                ModelSource::Url(DownloadUrl::parse("https://example.com/model.gguf").unwrap()),
+            ]
+            .iter()
+            .map(ModelSource::uri)
+            .collect::<Vec<_>>(),
+            [
+                "https://huggingface.co/owner/repo/blob/refs%2Fpr%2F1/nested/my%20model.gguf",
+                "agent:///models/model.gguf",
+                "https://example.com/model.gguf",
+            ]
+        );
+    }
+
+    #[test]
+    fn rejects_an_unparseable_uri() {
+        let slot_aggregated_status = SlotAggregatedStatus::new(1);
+        let parse_error = ModelSource::parse("not a uri", &slot_aggregated_status)
+            .expect_err("a uri that is not a url must be rejected");
+
+        assert_eq!(
+            parse_error.to_string(),
+            ModelSourceError::ModelUriUnparseable {
+                source: ParseError::RelativeUrlWithoutBase,
+                uri: "not a uri".to_owned(),
+            }
+            .to_string()
+        );
+        assert_eq!(
+            parse_error
+                .source()
+                .and_then(|source| source.downcast_ref::<ParseError>()),
+            Some(&ParseError::RelativeUrlWithoutBase)
+        );
+        assert!(
+            slot_aggregated_status
+                .has_issue(&AgentIssue::ModelUriIsUnparseable(model_path("not a uri")))
+        );
+    }
+
+    #[test]
+    fn rejects_a_uri_that_cannot_be_downloaded() {
+        let slot_aggregated_status = SlotAggregatedStatus::new(1);
+        let uri = "ftp://example.com/model.gguf";
+        let unsupported_scheme = || DownloadError::UnsupportedUrlScheme {
+            scheme: "ftp".to_owned(),
+            url: uri.to_owned(),
+        };
+        let parse_error = ModelSource::parse(uri, &slot_aggregated_status)
+            .expect_err("a uri with a scheme that cannot be downloaded must be rejected");
+
+        assert_eq!(
+            parse_error.to_string(),
+            ModelSourceError::ModelUriNotDownloadable {
+                source: unsupported_scheme(),
+                uri: uri.to_owned(),
+            }
+            .to_string()
+        );
+        assert_eq!(
+            parse_error
+                .source()
+                .and_then(|source| source.downcast_ref::<DownloadError>())
+                .map(ToString::to_string),
+            Some(unsupported_scheme().to_string())
+        );
+        assert!(
+            slot_aggregated_status.has_issue(&AgentIssue::DownloadUrlIsMalformed(model_path(uri)))
+        );
+    }
+}
diff --git a/paddler_model_source/src/model_source_error.rs b/paddler_model_source/src/model_source_error.rs
index 3ea1aa8e8..160b83b50 100644
--- a/paddler_model_source/src/model_source_error.rs
+++ b/paddler_model_source/src/model_source_error.rs
@@ -1,20 +1,28 @@
 use std::io;
 use std::path::PathBuf;
+use std::str::Utf8Error;
 
-use hf_hub::api::tokio::ApiError;
+use hf_hub::HFError;
 use reqwest::Error as ReqwestError;
 use thiserror::Error;
+use url::ParseError;
 
 use paddler_cache_dir::cache_dir_error::CacheDirError;
 use paddler_download_manager::download_error::DownloadError;
 
 #[derive(Debug, Error)]
 pub enum ModelSourceError {
-    #[error("Unable to resolve the model cache directory")]
+    #[error(
+        "Agent URI '{uri}' does not name an absolute local path: use agent:///absolute/path/to/model.gguf"
+    )]
+    AgentUriNotAnAbsoluteLocalPath { uri: String },
+    #[error("Agent URI '{uri}' is not canonical: use '{canonical_uri}'")]
+    AgentUriNotCanonical { canonical_uri: String, uri: String },
+    #[error("Unable to resolve the model cache directory: {0}")]
     CacheDirectoryUnresolvable(#[source] CacheDirError),
-    #[error("Unable to set up the model downloader")]
-    DownloaderUnavailable(#[source] ReqwestError),
-    #[error("Failed to download '{url}'")]
+    #[error("Unable to set up the model download client: {0}")]
+    DownloadClientUnavailable(#[source] ReqwestError),
+    #[error("Failed to download '{url}': {source}")]
     DownloadFailed {
         url: String,
         #[source]
@@ -22,13 +30,19 @@ pub enum ModelSourceError {
     },
     #[error("Cannot download '{url}': another agent on this host is downloading it")]
     DownloadLockHeldByAnotherProcess { url: String },
-    #[error("Unable to set up the Hugging Face client")]
-    HuggingFaceClientUnavailable(#[source] ApiError),
-    #[error("Failed to download model '{model_path}' from Hugging Face")]
+    #[error("Unable to look up model '{model_path}' in the Hugging Face cache: {source}")]
+    HuggingFaceCacheLookupFailed {
+        model_path: String,
+        #[source]
+        source: HFError,
+    },
+    #[error("Unable to set up the Hugging Face client: {0}")]
+    HuggingFaceClientUnavailable(#[source] HFError),
+    #[error("Failed to download model '{model_path}' from Hugging Face: {source}")]
     HuggingFaceDownloadFailed {
         model_path: String,
         #[source]
-        source: ApiError,
+        source: HFError,
     },
     #[error(
         "Failed to acquire download lock '{lock_path}'. Is more than one agent running on this machine?"
@@ -40,14 +54,44 @@ pub enum ModelSourceError {
     HuggingFaceModelKnownToBeMissing { model_path: String },
     #[error("Model '{model_path}' does not exist on Hugging Face.")]
     HuggingFaceModelMissing { model_path: String },
+    #[error(
+        "Hugging Face model URI '{uri}' must point at a file: https://huggingface.co/<owner>/<repo>/blob/<revision>/<file>"
+    )]
+    HuggingFaceModelUriMalformed { uri: String },
+    #[error(
+        "Hugging Face model URI '{uri}' has a path component that is not valid UTF-8 once percent-decoded: {source}"
+    )]
+    HuggingFaceModelUriComponentNotUtf8 {
+        uri: String,
+        #[source]
+        source: Utf8Error,
+    },
     #[error("You do not have enough permissions to download '{model_path}' from Hugging Face.")]
     HuggingFacePermissionDenied { model_path: String },
     #[error("Local file does not exist: {path_display}", path_display = path.display())]
     LocalFileMissing { path: PathBuf },
-    #[error("Unable to check whether the local model file '{path_display}' exists", path_display = path.display())]
+    #[error("Unable to check whether the local model file '{path_display}' exists: {source}", path_display = path.display())]
     LocalFileUncheckable {
         path: PathBuf,
         #[source]
         source: io::Error,
     },
+    #[error("Local model path '{path_display}' climbs to a parent directory", path_display = path.display())]
+    LocalModelPathClimbsToParentDirectory { path: PathBuf },
+    #[error("Local model path '{path_display}' is not absolute", path_display = path.display())]
+    LocalModelPathNotAbsolute { path: PathBuf },
+    #[error(
+        "Model URI '{uri}' is neither a Hugging Face, an agent:// nor a downloadable URI: {source}"
+    )]
+    ModelUriNotDownloadable {
+        uri: String,
+        #[source]
+        source: DownloadError,
+    },
+    #[error("Model URI '{uri}' cannot be parsed: {source}")]
+    ModelUriUnparseable {
+        uri: String,
+        #[source]
+        source: ParseError,
+    },
 }
diff --git a/paddler_model_source/src/resolve_desired_model.rs b/paddler_model_source/src/resolve_desired_model.rs
deleted file mode 100644
index 2b1a18528..000000000
--- a/paddler_model_source/src/resolve_desired_model.rs
+++ /dev/null
@@ -1,162 +0,0 @@
-use std::sync::Arc;
-
-use hf_hub::Cache;
-use tokio_util::sync::CancellationToken;
-
-use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
-use paddler_messaging::agent_desired_model::AgentDesiredModel;
-
-use crate::desired_model_resolution::DesiredModelResolution;
-use crate::huggingface_model_source::HuggingFaceModelSource;
-use crate::local_model_path::LocalModelPath;
-use crate::model_source_error::ModelSourceError;
-use crate::resolves_model_source::ResolvesModelSource;
-use crate::url_model_source::UrlModelSource;
-
-pub async fn resolve_desired_model(
-    cancellation_token: &CancellationToken,
-    desired: &AgentDesiredModel,
-    slot_aggregated_status: Arc<SlotAggregatedStatus>,
-) -> Result<DesiredModelResolution, ModelSourceError> {
-    match desired {
-        AgentDesiredModel::HuggingFace(reference) => {
-            HuggingFaceModelSource {
-                cache: Cache::from_env(),
-                reference: reference.clone(),
-            }
-            .resolve(cancellation_token, slot_aggregated_status)
-            .await
-        }
-        AgentDesiredModel::LocalToAgent(path) => {
-            LocalModelPath::new(path.clone())
-                .resolve(cancellation_token, slot_aggregated_status)
-                .await
-        }
-        AgentDesiredModel::Url(reference) => {
-            UrlModelSource(reference.clone())
-                .resolve(cancellation_token, slot_aggregated_status)
-                .await
-        }
-        AgentDesiredModel::None => Ok(DesiredModelResolution::NotConfigured),
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use std::mem;
-    use std::sync::Arc;
-
-    use tempfile::NamedTempFile;
-    use tempfile::tempdir;
-    use tokio_util::sync::CancellationToken;
-
-    use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
-    use paddler_messaging::agent_desired_model::AgentDesiredModel;
-    use paddler_messaging::agent_issue::AgentIssue;
-    use paddler_messaging::huggingface_model_reference::HuggingFaceModelReference;
-    use paddler_messaging::produces_snapshot::ProducesSnapshot;
-
-    use crate::desired_model_resolution::DesiredModelResolution;
-    use crate::model_source_error::ModelSourceError;
-    use crate::resolve_desired_model::resolve_desired_model;
-
-    fn fresh_status() -> Arc<SlotAggregatedStatus> {
-        Arc::new(SlotAggregatedStatus::new(1))
-    }
-
-    #[tokio::test]
-    async fn local_existing_file_resolves_to_resolved_with_that_path() {
-        let status = fresh_status();
-        let temp_file = NamedTempFile::new().unwrap();
-        let path = temp_file.path().to_path_buf();
-        let desired = AgentDesiredModel::LocalToAgent(path.display().to_string());
-
-        let resolution = resolve_desired_model(&CancellationToken::new(), &desired, status)
-            .await
-            .unwrap();
-
-        assert!(matches!(
-            resolution,
-            DesiredModelResolution::Resolved(ref resolved) if *resolved == path
-        ));
-    }
-
-    #[tokio::test]
-    async fn local_missing_file_resolves_to_local_file_missing_with_that_path() {
-        let status = fresh_status();
-        let temp_dir = tempdir().unwrap();
-        let path = temp_dir.path().join("missing-desired.gguf");
-        let desired = AgentDesiredModel::LocalToAgent(path.display().to_string());
-
-        let resolution = resolve_desired_model(&CancellationToken::new(), &desired, status)
-            .await
-            .unwrap();
-
-        assert!(matches!(
-            resolution,
-            DesiredModelResolution::LocalFileMissing(ref missing) if *missing == path
-        ));
-    }
-
-    #[tokio::test]
-    async fn huggingface_already_marked_missing_resolves_to_error_without_network() {
-        let status = fresh_status();
-        let reference = HuggingFaceModelReference {
-            filename: "model.gguf".to_owned(),
-            repo_id: "owner/repo".to_owned(),
-            revision: "main".to_owned(),
-        };
-        status.register_issue(AgentIssue::HuggingFaceModelDoesNotExist(
-            reference.model_path(),
-        ));
-        let desired = AgentDesiredModel::HuggingFace(reference);
-
-        let resolution = resolve_desired_model(&CancellationToken::new(), &desired, status).await;
-
-        assert!(matches!(
-            resolution,
-            Err(ModelSourceError::HuggingFaceModelKnownToBeMissing { model_path })
-                if model_path == "owner/repo/main/model.gguf"
-        ));
-    }
-
-    #[tokio::test]
-    async fn a_cancelled_token_aborts_a_huggingface_download_without_registering_an_issue() {
-        let status = fresh_status();
-        let reference = HuggingFaceModelReference {
-            filename: "model.gguf".to_owned(),
-            repo_id: "owner/repo".to_owned(),
-            revision: "main".to_owned(),
-        };
-        let desired = AgentDesiredModel::HuggingFace(reference);
-        let cancellation_token = CancellationToken::new();
-
-        cancellation_token.cancel();
-
-        let resolution = resolve_desired_model(&cancellation_token, &desired, status.clone()).await;
-
-        assert!(
-            resolution.is_ok_and(|resolution| resolution == DesiredModelResolution::Cancelled),
-            "a cancelled Hugging Face download must report cancellation as an outcome, not an error"
-        );
-        assert!(
-            status.make_snapshot().status.issues.is_empty(),
-            "a cancelled Hugging Face download must not register a slot issue"
-        );
-    }
-
-    #[tokio::test]
-    async fn none_variant_resolves_to_not_configured() {
-        let status = fresh_status();
-        let desired = AgentDesiredModel::None;
-
-        let resolution = resolve_desired_model(&CancellationToken::new(), &desired, status)
-            .await
-            .unwrap();
-
-        assert_eq!(
-            mem::discriminant(&resolution),
-            mem::discriminant(&DesiredModelResolution::NotConfigured)
-        );
-    }
-}
diff --git a/paddler_model_source/src/resolves_model_source.rs b/paddler_model_source/src/resolves_model_source.rs
index 6e2759109..2c8df53d3 100644
--- a/paddler_model_source/src/resolves_model_source.rs
+++ b/paddler_model_source/src/resolves_model_source.rs
@@ -1,6 +1,5 @@
 use std::sync::Arc;
 
-use async_trait::async_trait;
 use tokio_util::sync::CancellationToken;
 
 use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
@@ -8,11 +7,10 @@ use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
 use crate::desired_model_resolution::DesiredModelResolution;
 use crate::model_source_error::ModelSourceError;
 
-#[async_trait]
 pub trait ResolvesModelSource {
-    async fn resolve(
+    fn resolve(
         &self,
         cancellation_token: &CancellationToken,
         slot_aggregated_status: Arc<SlotAggregatedStatus>,
-    ) -> Result<DesiredModelResolution, ModelSourceError>;
+    ) -> impl Future<Output = Result<DesiredModelResolution, ModelSourceError>> + Send;
 }
diff --git a/paddler_model_source/src/url_download_progress.rs b/paddler_model_source/src/url_download_progress.rs
index ed63d0779..fcddddfa0 100644
--- a/paddler_model_source/src/url_download_progress.rs
+++ b/paddler_model_source/src/url_download_progress.rs
@@ -93,7 +93,7 @@ mod tests {
     fn starting_a_download_of_known_size_reports_it_and_clears_the_matching_download_issue() {
         let status = Arc::new(SlotAggregatedStatus::new(1));
 
-        status.register_issue(AgentIssue::DownloadInterrupted(test_model_path()));
+        status.register_desired_state_issue(AgentIssue::DownloadInterrupted(test_model_path()));
 
         url_download_progress(&status).apply(DownloadProgress::Started {
             already_downloaded_bytes: 100,
@@ -162,7 +162,7 @@ mod tests {
             already_downloaded_bytes: 200,
             total_bytes: Some(500),
         });
-        status.register_issue(AgentIssue::DownloadInterrupted(test_model_path()));
+        status.register_desired_state_issue(AgentIssue::DownloadInterrupted(test_model_path()));
         download_progress.apply(DownloadProgress::Finished);
 
         let snapshot = status.make_snapshot();
diff --git a/paddler_model_source/src/url_model_source.rs b/paddler_model_source/src/url_model_source.rs
index fdd1a5c29..9c7ab0f2f 100644
--- a/paddler_model_source/src/url_model_source.rs
+++ b/paddler_model_source/src/url_model_source.rs
@@ -2,7 +2,6 @@ use std::io;
 use std::path::Path;
 use std::sync::Arc;
 
-use async_trait::async_trait;
 use tokio_util::sync::CancellationToken;
 
 use paddler_agent_status::agent_issue_fix::AgentIssueFix;
@@ -17,49 +16,23 @@ use paddler_download_manager::download_url::DownloadUrl;
 use paddler_messaging::agent_issue::AgentIssue;
 use paddler_messaging::agent_issue_params::model_path::ModelPath;
 use paddler_messaging::model_download_status::ModelDownloadStatus;
-use paddler_messaging::url_model_reference::UrlModelReference;
 
 use crate::desired_model_resolution::DesiredModelResolution;
+use crate::download_error_agent_issue::download_error_agent_issue;
 use crate::model_source_error::ModelSourceError;
 use crate::resolves_model_source::ResolvesModelSource;
 use crate::url_download_progress::UrlDownloadProgress;
 
-const fn agent_issue_for(error: &DownloadError, model_path: ModelPath) -> AgentIssue {
-    match error {
-        DownloadError::InvalidUrl { .. } | DownloadError::UnsupportedUrlScheme { .. } => {
-            AgentIssue::DownloadUrlIsMalformed(model_path)
-        }
-        DownloadError::NotFound { .. } => AgentIssue::ModelDoesNotExistAtUrl(model_path),
-        DownloadError::PermissionDenied { .. } => {
-            AgentIssue::DownloadServerDeniedAccess(model_path)
-        }
-        DownloadError::DownloadServerIsUnreachable { .. } => {
-            AgentIssue::DownloadServerIsUnreachable(model_path)
-        }
-        DownloadError::DownloadServerErrored { .. } => {
-            AgentIssue::DownloadServerErrored(model_path)
-        }
-        DownloadError::DownloadServerRejectedRequest { .. } => {
-            AgentIssue::DownloadServerRejectedRequest(model_path)
-        }
-        DownloadError::DownloadInterrupted { .. } => AgentIssue::DownloadInterrupted(model_path),
-        DownloadError::CachePermissionDenied { .. } => {
-            AgentIssue::CacheDirectoryIsNotWritable(model_path)
-        }
-        DownloadError::CacheDiskFull { .. } => AgentIssue::CacheStorageIsFull(model_path),
-        DownloadError::PartialFileStale { .. }
-        | DownloadError::PartialPathIsADirectory { .. }
-        | DownloadError::FinalPathHasNoFileName { .. }
-        | DownloadError::Io { .. } => AgentIssue::ModelCacheIsCorrupted(model_path),
-    }
+pub struct UrlModelSource {
+    pub cache_dir: CacheDir,
+    pub download_manager: DownloadManager,
+    pub download_url: DownloadUrl,
 }
 
-pub struct UrlModelSource(pub UrlModelReference);
-
 impl UrlModelSource {
     fn model_path(&self) -> ModelPath {
         ModelPath {
-            model_path: self.0.url.clone(),
+            model_path: self.download_url.as_str().to_owned(),
         }
     }
 
@@ -68,10 +41,13 @@ impl UrlModelSource {
         slot_aggregated_status: &SlotAggregatedStatus,
         download_error: DownloadError,
     ) -> ModelSourceError {
-        slot_aggregated_status.fail_download(agent_issue_for(&download_error, self.model_path()));
+        slot_aggregated_status.fail_download(download_error_agent_issue(
+            &download_error,
+            self.model_path(),
+        ));
 
         ModelSourceError::DownloadFailed {
-            url: self.0.url.clone(),
+            url: self.download_url.as_str().to_owned(),
             source: download_error,
         }
     }
@@ -87,16 +63,20 @@ impl UrlModelSource {
             DownloadError::cache_failure(path.to_path_buf(), io_error),
         )
     }
+}
 
-    async fn resolve_into_cache(
+impl ResolvesModelSource for UrlModelSource {
+    async fn resolve(
         &self,
         cancellation_token: &CancellationToken,
-        cache_dir: &CacheDir,
         slot_aggregated_status: Arc<SlotAggregatedStatus>,
     ) -> Result<DesiredModelResolution, ModelSourceError> {
-        let Self(UrlModelReference { url }) = self;
-        let download_url = DownloadUrl::parse(url)
-            .map_err(|url_error| self.fail_download(&slot_aggregated_status, url_error))?;
+        let Self {
+            cache_dir,
+            download_manager,
+            download_url,
+        } = self;
+        let url = download_url.as_str();
         let cached = CachedDownloadedModel::new(cache_dir, url)
             .map_err(ModelSourceError::CacheDirectoryUnresolvable)?;
         let is_cached = cached.is_cached().await.map_err(|io_error| {
@@ -125,7 +105,7 @@ impl UrlModelSource {
                     .fail_download(AgentIssue::CacheCannotAcquireLock(self.model_path()));
 
                 return Err(ModelSourceError::DownloadLockHeldByAnotherProcess {
-                    url: url.clone(),
+                    url: url.to_owned(),
                 });
             }
             Err(io_error) => {
@@ -141,11 +121,10 @@ impl UrlModelSource {
             model_path: self.model_path(),
             slot_aggregated_status: slot_aggregated_status.clone(),
         };
-        let download_outcome = DownloadManager::new()
-            .map_err(ModelSourceError::DownloaderUnavailable)?
+        let download_outcome = download_manager
             .download(
                 cancellation_token,
-                &download_url,
+                download_url,
                 &cached.cache_file_path,
                 &|progress| download_progress.apply(progress),
             )
@@ -163,66 +142,43 @@ impl UrlModelSource {
     }
 }
 
-#[async_trait]
-impl ResolvesModelSource for UrlModelSource {
-    async fn resolve(
-        &self,
-        cancellation_token: &CancellationToken,
-        slot_aggregated_status: Arc<SlotAggregatedStatus>,
-    ) -> Result<DesiredModelResolution, ModelSourceError> {
-        self.resolve_into_cache(
-            cancellation_token,
-            &CacheDir::from_process_env(),
-            slot_aggregated_status,
-        )
-        .await
-    }
-}
-
 #[cfg(test)]
 mod tests {
-    use std::io;
     use std::mem::discriminant;
-
-    use paddler_cache_dir::cache_dir_error::CacheDirError;
-    #[cfg(unix)]
     use std::os::unix::fs::symlink;
     use std::path::Path;
-    use std::path::PathBuf;
     use std::sync::Arc;
 
-    use reqwest::Client;
-    use reqwest::Error as ReqwestError;
-    use reqwest::StatusCode;
     use tempfile::TempDir;
     use tokio::fs::create_dir;
     use tokio::fs::read;
     use tokio::fs::write;
     use tokio_util::sync::CancellationToken;
-    use url::Url;
 
     use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
     use paddler_cache_dir::cache_dir::CacheDir;
+    use paddler_cache_dir::cache_dir_error::CacheDirError;
     use paddler_cache_dir::cached_downloaded_model::CachedDownloadedModel;
     use paddler_download_manager::download_error::DownloadError;
+    use paddler_download_manager::download_manager::DownloadManager;
+    use paddler_download_manager::download_url::DownloadUrl;
     use paddler_local_http_fixture::fixture_response::FixtureResponse;
     use paddler_local_http_fixture::local_http_fixture::LocalHttpFixture;
     use paddler_messaging::agent_issue::AgentIssue;
     use paddler_messaging::agent_issue_params::model_path::ModelPath;
     use paddler_messaging::produces_snapshot::ProducesSnapshot;
-    use paddler_messaging::url_model_reference::UrlModelReference;
 
     use crate::desired_model_resolution::DesiredModelResolution;
     use crate::model_source_error::ModelSourceError;
+    use crate::resolves_model_source::ResolvesModelSource as _;
     use crate::url_model_source::UrlModelSource;
-    use crate::url_model_source::agent_issue_for;
-
-    const TEST_URL: &str = "https://example.com/m.gguf";
 
-    fn url_model_source(url: &str) -> UrlModelSource {
-        UrlModelSource(UrlModelReference {
-            url: url.to_owned(),
-        })
+    fn url_model_source(cache_dir: CacheDir, url: &str) -> UrlModelSource {
+        UrlModelSource {
+            cache_dir,
+            download_manager: DownloadManager::new().unwrap(),
+            download_url: DownloadUrl::parse(url).unwrap(),
+        }
     }
 
     fn fresh_status() -> Arc<SlotAggregatedStatus> {
@@ -230,21 +186,10 @@ mod tests {
     }
 
     fn cache_dir_at(path: &Path) -> CacheDir {
-        #[cfg(unix)]
-        {
-            CacheDir {
-                explicit: Some(path.to_path_buf()),
-                home: None,
-                xdg: None,
-            }
-        }
-        #[cfg(windows)]
-        {
-            CacheDir {
-                explicit: Some(path.to_path_buf()),
-                localappdata: None,
-                userprofile: None,
-            }
+        CacheDir {
+            explicit: Some(path.to_path_buf()),
+            home: None,
+            xdg: None,
         }
     }
 
@@ -259,8 +204,8 @@ mod tests {
             .await
             .unwrap();
 
-        let resolution = url_model_source(url_string)
-            .resolve_into_cache(&CancellationToken::new(), &cache_dir, fresh_status())
+        let resolution = url_model_source(cache_dir.clone(), url_string)
+            .resolve(&CancellationToken::new(), fresh_status())
             .await
             .unwrap();
 
@@ -270,60 +215,6 @@ mod tests {
         ));
     }
 
-    #[tokio::test]
-    async fn malformed_url_registers_download_url_is_malformed() {
-        let directory = TempDir::new().unwrap();
-        let cache_dir = cache_dir_at(directory.path());
-        let url_string = "not a url";
-
-        let status = fresh_status();
-        let result = url_model_source(url_string)
-            .resolve_into_cache(&CancellationToken::new(), &cache_dir, status.clone())
-            .await;
-
-        assert!(matches!(
-            result,
-            Err(ModelSourceError::DownloadFailed {
-                source: DownloadError::InvalidUrl { .. },
-                url,
-            }) if url == url_string
-        ));
-        assert!(
-            status.has_issue(&AgentIssue::DownloadUrlIsMalformed(ModelPath {
-                model_path: url_string.to_owned(),
-            }))
-        );
-    }
-
-    #[tokio::test]
-    async fn unsupported_scheme_registers_download_url_is_malformed_without_creating_cache_state() {
-        let directory = TempDir::new().unwrap();
-        let cache_dir = cache_dir_at(directory.path());
-        let url_string = "ftp://example.invalid/m.gguf";
-
-        let status = fresh_status();
-        let result = url_model_source(url_string)
-            .resolve_into_cache(&CancellationToken::new(), &cache_dir, status.clone())
-            .await;
-
-        assert!(matches!(
-            result,
-            Err(ModelSourceError::DownloadFailed {
-                source: DownloadError::UnsupportedUrlScheme { scheme, .. },
-                ..
-            }) if scheme == "ftp"
-        ));
-        assert!(
-            status.has_issue(&AgentIssue::DownloadUrlIsMalformed(ModelPath {
-                model_path: url_string.to_owned(),
-            }))
-        );
-        assert!(
-            !directory.path().join("downloaded-models").exists(),
-            "no cache subdirectory must be created for an unsupported scheme"
-        );
-    }
-
     #[tokio::test]
     async fn lock_contention_registers_cache_cannot_acquire_lock() {
         let directory = TempDir::new().unwrap();
@@ -335,8 +226,8 @@ mod tests {
         let _blocker = cached.try_acquire_download_lock().unwrap();
 
         let status = fresh_status();
-        let result = url_model_source(url_string)
-            .resolve_into_cache(&CancellationToken::new(), &cache_dir, status.clone())
+        let result = url_model_source(cache_dir.clone(), url_string)
+            .resolve(&CancellationToken::new(), status.clone())
             .await;
 
         assert!(matches!(
@@ -350,7 +241,6 @@ mod tests {
         );
     }
 
-    #[cfg(unix)]
     #[tokio::test]
     async fn cache_subdir_creation_failure_registers_model_cache_is_corrupted() {
         let directory = TempDir::new().unwrap();
@@ -360,8 +250,8 @@ mod tests {
         symlink(directory.path().join("missing-target"), &subdir_path).unwrap();
 
         let status = fresh_status();
-        let result = url_model_source(url_string)
-            .resolve_into_cache(&CancellationToken::new(), &cache_dir, status.clone())
+        let result = url_model_source(cache_dir.clone(), url_string)
+            .resolve(&CancellationToken::new(), status.clone())
             .await;
 
         assert!(matches!(
@@ -376,7 +266,6 @@ mod tests {
         );
     }
 
-    #[cfg(unix)]
     #[tokio::test]
     async fn lock_open_io_error_registers_model_cache_is_corrupted() {
         let directory = TempDir::new().unwrap();
@@ -387,8 +276,8 @@ mod tests {
         create_dir(&cached.lock_file_path).await.unwrap();
 
         let status = fresh_status();
-        let result = url_model_source(url_string)
-            .resolve_into_cache(&CancellationToken::new(), &cache_dir, status.clone())
+        let result = url_model_source(cache_dir.clone(), url_string)
+            .resolve(&CancellationToken::new(), status.clone())
             .await;
 
         assert!(matches!(
@@ -403,176 +292,6 @@ mod tests {
         );
     }
 
-    async fn refused_connection_error() -> ReqwestError {
-        Client::new()
-            .get("http://127.0.0.1:1/model.gguf")
-            .send()
-            .await
-            .expect_err("nothing listens on the loopback discard port")
-    }
-
-    fn test_model_path() -> ModelPath {
-        ModelPath {
-            model_path: TEST_URL.to_owned(),
-        }
-    }
-
-    #[test]
-    fn invalid_url_maps_to_download_url_is_malformed() {
-        let parse_error = Url::parse("not a url").err().unwrap();
-        let error = DownloadError::InvalidUrl {
-            url: "not a url".to_owned(),
-            source: parse_error,
-        };
-
-        assert_eq!(
-            agent_issue_for(&error, test_model_path()),
-            AgentIssue::DownloadUrlIsMalformed(test_model_path())
-        );
-    }
-
-    #[test]
-    fn unsupported_url_scheme_maps_to_download_url_is_malformed() {
-        let error = DownloadError::UnsupportedUrlScheme {
-            url: TEST_URL.to_owned(),
-            scheme: "ftp".to_owned(),
-        };
-
-        assert_eq!(
-            agent_issue_for(&error, test_model_path()),
-            AgentIssue::DownloadUrlIsMalformed(test_model_path())
-        );
-    }
-
-    #[test]
-    fn not_found_maps_to_model_does_not_exist_at_url() {
-        let error = DownloadError::NotFound {
-            url: TEST_URL.to_owned(),
-        };
-
-        assert_eq!(
-            agent_issue_for(&error, test_model_path()),
-            AgentIssue::ModelDoesNotExistAtUrl(test_model_path())
-        );
-    }
-
-    #[test]
-    fn permission_denied_maps_to_download_server_denied_access() {
-        let error = DownloadError::PermissionDenied {
-            url: TEST_URL.to_owned(),
-            status: StatusCode::FORBIDDEN,
-        };
-
-        assert_eq!(
-            agent_issue_for(&error, test_model_path()),
-            AgentIssue::DownloadServerDeniedAccess(test_model_path())
-        );
-    }
-
-    #[test]
-    fn partial_file_stale_maps_to_model_cache_is_corrupted() {
-        let error = DownloadError::PartialFileStale {
-            url: TEST_URL.to_owned(),
-            partial_path: PathBuf::from("/tmp/stale.partial"),
-        };
-
-        assert_eq!(
-            agent_issue_for(&error, test_model_path()),
-            AgentIssue::ModelCacheIsCorrupted(test_model_path())
-        );
-    }
-
-    #[tokio::test]
-    async fn download_server_is_unreachable_maps_to_agent_issue() {
-        let error = DownloadError::DownloadServerIsUnreachable {
-            url: TEST_URL.to_owned(),
-            source: refused_connection_error().await,
-        };
-
-        assert_eq!(
-            agent_issue_for(&error, test_model_path()),
-            AgentIssue::DownloadServerIsUnreachable(test_model_path())
-        );
-    }
-
-    #[test]
-    fn download_server_errored_maps_to_agent_issue() {
-        let error = DownloadError::DownloadServerErrored {
-            url: TEST_URL.to_owned(),
-            status: StatusCode::INTERNAL_SERVER_ERROR,
-        };
-
-        assert_eq!(
-            agent_issue_for(&error, test_model_path()),
-            AgentIssue::DownloadServerErrored(test_model_path())
-        );
-    }
-
-    #[test]
-    fn download_server_rejected_request_maps_to_agent_issue() {
-        let error = DownloadError::DownloadServerRejectedRequest {
-            url: TEST_URL.to_owned(),
-            status: StatusCode::BAD_REQUEST,
-        };
-
-        assert_eq!(
-            agent_issue_for(&error, test_model_path()),
-            AgentIssue::DownloadServerRejectedRequest(test_model_path())
-        );
-    }
-
-    #[tokio::test]
-    async fn download_interrupted_maps_to_agent_issue() {
-        let error = DownloadError::DownloadInterrupted {
-            url: TEST_URL.to_owned(),
-            source: refused_connection_error().await,
-        };
-
-        assert_eq!(
-            agent_issue_for(&error, test_model_path()),
-            AgentIssue::DownloadInterrupted(test_model_path())
-        );
-    }
-
-    #[test]
-    fn cache_permission_denied_maps_to_cache_directory_is_not_writable() {
-        let error = DownloadError::CachePermissionDenied {
-            path: PathBuf::from("/tmp/locked/model.partial"),
-            source: io::Error::from(io::ErrorKind::PermissionDenied),
-        };
-
-        assert_eq!(
-            agent_issue_for(&error, test_model_path()),
-            AgentIssue::CacheDirectoryIsNotWritable(test_model_path())
-        );
-    }
-
-    #[test]
-    fn cache_disk_full_maps_to_cache_storage_is_full() {
-        let error = DownloadError::CacheDiskFull {
-            path: PathBuf::from("/tmp/full/model.partial"),
-            source: io::Error::from(io::ErrorKind::StorageFull),
-        };
-
-        assert_eq!(
-            agent_issue_for(&error, test_model_path()),
-            AgentIssue::CacheStorageIsFull(test_model_path())
-        );
-    }
-
-    #[test]
-    fn io_maps_to_model_cache_is_corrupted() {
-        let error = DownloadError::Io {
-            path: PathBuf::from("/tmp/anywhere/model.partial"),
-            source: io::Error::from(io::ErrorKind::NotFound),
-        };
-
-        assert_eq!(
-            agent_issue_for(&error, test_model_path()),
-            AgentIssue::ModelCacheIsCorrupted(test_model_path())
-        );
-    }
-
     #[tokio::test]
     async fn a_file_blocking_the_cache_subdir_registers_model_cache_is_corrupted() {
         let directory = TempDir::new().unwrap();
@@ -586,8 +305,8 @@ mod tests {
             .cache_file_path;
 
         let status = fresh_status();
-        let result = url_model_source(url_string)
-            .resolve_into_cache(&CancellationToken::new(), &cache_dir, status.clone())
+        let result = url_model_source(cache_dir.clone(), url_string)
+            .resolve(&CancellationToken::new(), status.clone())
             .await;
 
         assert!(matches!(
@@ -603,21 +322,10 @@ mod tests {
     }
 
     fn unresolvable_cache_dir() -> CacheDir {
-        #[cfg(unix)]
-        {
-            CacheDir {
-                explicit: None,
-                home: None,
-                xdg: None,
-            }
-        }
-        #[cfg(windows)]
-        {
-            CacheDir {
-                explicit: None,
-                localappdata: None,
-                userprofile: None,
-            }
+        CacheDir {
+            explicit: None,
+            home: None,
+            xdg: None,
         }
     }
 
@@ -625,18 +333,14 @@ mod tests {
     async fn cache_path_resolution_failure_propagates_error() {
         let url_string = "https://host.example/unresolvable.gguf";
 
-        let result = url_model_source(url_string)
-            .resolve_into_cache(
-                &CancellationToken::new(),
-                &unresolvable_cache_dir(),
-                fresh_status(),
-            )
+        let result = url_model_source(unresolvable_cache_dir(), url_string)
+            .resolve(&CancellationToken::new(), fresh_status())
             .await;
 
         assert_eq!(
             result.err().as_ref().map(discriminant),
             Some(discriminant(&ModelSourceError::CacheDirectoryUnresolvable(
-                CacheDirError::HomeVariableUnset { variable: "HOME" }
+                CacheDirError::HomeVariableUnset
             )))
         );
     }
@@ -655,8 +359,8 @@ mod tests {
         let cached = CachedDownloadedModel::new(&cache_dir, &url_string).unwrap();
         let expected_path = cached.cache_file_path.clone();
 
-        let resolution = url_model_source(&url_string)
-            .resolve_into_cache(&CancellationToken::new(), &cache_dir, fresh_status())
+        let resolution = url_model_source(cache_dir.clone(), &url_string)
+            .resolve(&CancellationToken::new(), fresh_status())
             .await
             .unwrap();
 
@@ -666,6 +370,12 @@ mod tests {
             DesiredModelResolution::Resolved(resolved_path) if resolved_path == expected_path
         ));
         assert_eq!(read(&expected_path).await.unwrap(), body);
+
+        fixture
+            .stop()
+            .await
+            .expect("the fixture task must not panic")
+            .expect("the fixture must serve every request");
     }
 
     #[tokio::test]
@@ -678,8 +388,8 @@ mod tests {
 
         cancellation_token.cancel();
 
-        let result = url_model_source(&url_string)
-            .resolve_into_cache(&cancellation_token, &cache_dir, status.clone())
+        let result = url_model_source(cache_dir.clone(), &url_string)
+            .resolve(&cancellation_token, status.clone())
             .await;
 
         assert!(
diff --git a/paddler_model_source/src/url_path_segment_encode_set.rs b/paddler_model_source/src/url_path_segment_encode_set.rs
new file mode 100644
index 000000000..6ea855a72
--- /dev/null
+++ b/paddler_model_source/src/url_path_segment_encode_set.rs
@@ -0,0 +1,16 @@
+use percent_encoding::AsciiSet;
+use percent_encoding::CONTROLS;
+
+pub const URL_PATH_SEGMENT_ENCODE_SET: &AsciiSet = &CONTROLS
+    .add(b' ')
+    .add(b'"')
+    .add(b'#')
+    .add(b'%')
+    .add(b'/')
+    .add(b'<')
+    .add(b'>')
+    .add(b'?')
+    .add(b'\\')
+    .add(b'`')
+    .add(b'{')
+    .add(b'}');
diff --git a/paddler_openai_client_python_test/.gitignore b/paddler_openai_client_python_test/.gitignore
index 0d778d0f2..521016d76 100644
--- a/paddler_openai_client_python_test/.gitignore
+++ b/paddler_openai_client_python_test/.gitignore
@@ -1,5 +1,4 @@
 __pycache__/
 .pytest_cache/
 .mypy_cache/
-.ruff_cache/
 .venv/
diff --git a/paddler_openai_client_python_test/poetry.lock b/paddler_openai_client_python_test/poetry.lock
index d4c800327..019819f41 100644
--- a/paddler_openai_client_python_test/poetry.lock
+++ b/paddler_openai_client_python_test/poetry.lock
@@ -6,7 +6,7 @@ version = "0.7.0"
 description = "Reusable constraint types to use with typing.Annotated"
 optional = false
 python-versions = ">=3.8"
-groups = ["main", "dev"]
+groups = ["main"]
 files = [
     {file = "annotated_types-0.7.0-py3-none-any.whl", hash = "sha256:1f02e8b43a8fbbc3f3e0d4f0f4bfc8131bcb4eebe8849b8e5c773f3a1c582a53"},
     {file = "annotated_types-0.7.0.tar.gz", hash = "sha256:aff07c09a53a08bc8cfccb9c85b05f1aa9a2a6f23728d790723543408344ce89"},
@@ -18,7 +18,7 @@ version = "4.15.1"
 description = "High-level concurrency and networking framework on top of asyncio or Trio"
 optional = false
 python-versions = ">=3.10"
-groups = ["main", "dev"]
+groups = ["main"]
 files = [
     {file = "anyio-4.15.1-py3-none-any.whl", hash = "sha256:6152fdbbf9a77fdec97731721bebf7c4c44f7c29b424b0065826173efc7ed101"},
     {file = "anyio-4.15.1.tar.gz", hash = "sha256:9f28306018cbd6d329e64a36d58256edff76dd996fe423bc957326e578b82a94"},
@@ -32,15 +32,70 @@ typing_extensions = {version = ">=4.16.0", markers = "python_version < \"3.15\""
 trio = ["trio (>=0.32.0)"]
 
 [[package]]
-name = "certifi"
-version = "2026.5.20"
-description = "Python package for providing Mozilla's CA Bundle."
+name = "ast-serialize"
+version = "0.13.1"
+description = "Python bindings for mypy AST serialization"
 optional = false
 python-versions = ">=3.7"
-groups = ["main", "dev"]
+groups = ["dev"]
 files = [
-    {file = "certifi-2026.5.20-py3-none-any.whl", hash = "sha256:3c52e209ba0a4ad7aebe60436a4ab349c39e1e602e8c134221e546902ad25897"},
-    {file = "certifi-2026.5.20.tar.gz", hash = "sha256:69dea482ab64caa7b9f6aba1c6bf48bb6a5448d1c0f1b17ab42ad8c763a5344d"},
+    {file = "ast_serialize-0.13.1-cp314-cp314-pyemscripten_2026_0_wasm32.whl", hash = "sha256:3e129d44b0cffbea24a7f40117ce49e30993c3d119e584d4817ebd6f87246456"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-macosx_10_12_x86_64.whl", hash = "sha256:523d45bab16ddaf1399c04c4fe81bbf180f81e33c2f7bcbe76be56b0f7311ab4"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:638333fb8ae72c62753f9e794fda20290883ddc65b0d85e26e08f8bb227567ad"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:cc34c670747d85d4232834665e247009413eaebe47c69eaf60a5c439aaecaed9"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:8c8ebd419811c435812d242fb2a173e57dfede198c9ac9a0a584666ebc9b9331"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:8899c038b1c15ae8b98e379057fb3ef398ceb0b21d1737a8156dd5a0923b7e51"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:e14c4de98f965647a74e6eaf37e20207124e45451ccd36e84a6f6ebab45e0372"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:7f3e8a3eff69af640db985360fef35073fb46956dfc3e5c614cf00920d2746dd"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-manylinux_2_31_riscv64.whl", hash = "sha256:75d7750fe98f6f35a96c43aba8d98a2e623e22c778a0c17ab9ede92eb10f5d82"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:43e8ddd63e80fc280c2f790157386f27f9b9f6ccb9c4f8e803b863cbfb085769"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:9221e09496fb942901993c1812ee87950cc0ecbb514c7c0f1f5fa6d78f701ed8"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-musllinux_1_2_armv7l.whl", hash = "sha256:7a5c38bdb93669c596f94fb7204dd6ebe4fadb19058ac9a2faf2948c2a175f96"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-musllinux_1_2_i686.whl", hash = "sha256:acf4e9f055d75058ab01536da330648746f2f82b7697a85230547f2d5ebe895c"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-musllinux_1_2_ppc64le.whl", hash = "sha256:87427913c325adc811e81e7c9a757b061f47ef1e29114ac479aa0953a2687a98"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-musllinux_1_2_riscv64.whl", hash = "sha256:ad88921f95b92bc144d9ad059a0e16f8c1ae69f5239f2d3c6536fd2774f73dfe"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:f8b17e2921e56621bcb46792bd3e0b82ca3b85de78d5936e7b8174ba1758b0aa"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-win32.whl", hash = "sha256:2f9a9ce8b0994f757b941f7f7fe384a6666b4162b941ebc5c248081baea18e47"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-win_amd64.whl", hash = "sha256:5d99910e6b9a9b9360bcca2ae99aee36cc912ce058df2a57d1c33099b6cb2593"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-win_arm64.whl", hash = "sha256:a789cbe14d0a0e9ec9b32a456e5abc81ad8799479c3cfd88a29e3442262aebe6"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-macosx_10_12_x86_64.whl", hash = "sha256:23d768369b89d551077dc173da6892fad00443338417523b6732f4e8ba4a7ed5"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-macosx_11_0_arm64.whl", hash = "sha256:68f8829d114dab6c9db4a8bbb8d0e342958faee1b5e15a29022678df639960f8"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:4e385b17d5bde892a6184003a9aff3518387908b73f8f5ea539313c07c8c0ede"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:02686b281ec9b8e9614f6f0db27e12e4cac515f336c5d05b678bc9eb73ff2af5"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:a8ab2c2585b2e22e48cd4aff2136fce4eab4ced569b85f150ef9a5a5d6f34d49"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:c652d9f5d52d8b81eff3ff2c95c9dd0f281de45f6ad5d3fa95527550c396454a"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:121e1d08d58b75c15491cc3003609cbad71ac2b0016043c36d30a6b2032b6605"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-manylinux_2_31_riscv64.whl", hash = "sha256:30cf89b6a6844542d116c758725d13de5b97b9410a51ee577715f761270d0c1c"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:732aaae9e498c973fc63fd7f308009b57c4afa6ff59d0b674d894e57d0bca01a"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-musllinux_1_2_aarch64.whl", hash = "sha256:d9c1a920b2aa0c1e016bed7721689717034ca0dba64ac3a3ff243b003d4e8dd8"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-musllinux_1_2_armv7l.whl", hash = "sha256:847cad4c22953823f5e76d0b126c6f9da670cfa45b1df41b9324e22189d77611"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-musllinux_1_2_i686.whl", hash = "sha256:aca07de3ca8e3d9f2b716c3c31f4f384fe4bd2d03e09912417cff7a3e08036f7"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-musllinux_1_2_ppc64le.whl", hash = "sha256:5db12d843d931b9d1f1519ce66b9cbfe72d8563b75c1204d255322254d1964dd"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-musllinux_1_2_riscv64.whl", hash = "sha256:fb8021f4dc3296de18737e3b8b34994411da725374d74a0b9eb32a738f2a23d0"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-musllinux_1_2_x86_64.whl", hash = "sha256:0d764913374ba21df8cd618045b371ed104ce55c6276de7c59e554a4f5ba26fd"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-win32.whl", hash = "sha256:2bad302ae2427454ef9dab71bd5c3ad67188860c39d36daee38f72031edb39c4"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-win_amd64.whl", hash = "sha256:032b09d86c785c3984e975378b7861f2a68ac44e22ba7b07b93a536b2524eeb8"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-win_arm64.whl", hash = "sha256:929eda0c419e622ab49df3a4c45f81c57dcc10874bbad55d0470a4a03324dac3"},
+    {file = "ast_serialize-0.13.1-cp315-cp315-pyemscripten_2026_5_wasm32.whl", hash = "sha256:bf53a3ba1cbe45399fc1e4ccd684b8f2a11cca450f89706b79122f478534c6f5"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-macosx_10_12_x86_64.whl", hash = "sha256:4bcf1d5f4eefecaa325a34ecf079b2741189e7d0821523f92cb343a389a67489"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-macosx_11_0_arm64.whl", hash = "sha256:1b158f1c9e3a0d386913924c57995d6df78224868a4d0f215bdd41587c97b34a"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:77a6201b69ce42e919f2fd9a559138d5410b5d8afff7c3a6b1f500b0481ff7d8"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:6eb0be64c9b8f1ed26637eef345abe75ea90351c92c85a8aec7838b3af971011"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:7f2427bddfca208e84de700044fadf9aef2769a70bae740a1bdabb89ee9e1117"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:88df960e67184a71874e5b2ee014f87d1254a0dbc1f903c7bd7b1ab7733ecaef"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:9f99786e0d62082a5b121b3437f3167562133711df22128de1e9886f57f0b226"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-manylinux_2_31_riscv64.whl", hash = "sha256:a01b515b25a03184eff4acef25b981ae007916379682e908fc8785b342a0bf3e"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:29422d7bcd9e7f39411571c36cec45ce8a7a9b3ad4b2e7aa64497c15a9e397f6"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:6ffb0f548a7dc126067fdfb7b2325ab5dfcdeecc29faf5e0f0461c96d6ebfa3d"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-musllinux_1_2_armv7l.whl", hash = "sha256:2876444a5f87a7c76f371af3c07c2fda563a08bc8af52568df812e8aba4bf568"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-musllinux_1_2_i686.whl", hash = "sha256:885a62ec6481ce58ab24c6357360d7a82023b62dd7e5e7ceccf59e9662a3272b"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-musllinux_1_2_ppc64le.whl", hash = "sha256:2015fd0165e463aac3b1aadc6a1fa1e5820c14ca950821f4fe02f8a3f6a388d6"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-musllinux_1_2_riscv64.whl", hash = "sha256:eb41e1906502ac80d24ff343f45008552a3f5b5158607959fe6a07dbfc6eb037"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:ca9acc67f796aa2554fa3037dccc022a4b08b38b2ad153edfac9c9bb7cb1f3c6"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-win32.whl", hash = "sha256:4f98da562e5901d69c7dd30885594c5438f859edf354e9ac7517d188fb026c3a"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-win_amd64.whl", hash = "sha256:ddf7cbc9c8abf70102c4a088efe154bb86bfb825b0a10e556d2b15972cf99bd7"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-win_arm64.whl", hash = "sha256:48f11d96c9ec71119612f8e79a8c55c12f2b136716d68347433c561d27a25f43"},
+    {file = "ast_serialize-0.13.1.tar.gz", hash = "sha256:401cc84e8888d4d76f2c8d296fa006ed1a5ff993b6aff90dd777c03973efd025"},
 ]
 
 [[package]]
@@ -49,24 +104,12 @@ version = "0.4.6"
 description = "Cross-platform colored terminal text."
 optional = false
 python-versions = "!=3.0.*,!=3.1.*,!=3.2.*,!=3.3.*,!=3.4.*,!=3.5.*,!=3.6.*,>=2.7"
-groups = ["main", "dev"]
+groups = ["dev"]
+markers = "sys_platform == \"win32\""
 files = [
     {file = "colorama-0.4.6-py2.py3-none-any.whl", hash = "sha256:4f1d9991f5acc0ca119f9d443620b77f9d6b33703e51011c16baf57afb285fc6"},
     {file = "colorama-0.4.6.tar.gz", hash = "sha256:08695f5cb7ed6e0531a20572697297273c47b8cae5a63ffc6d6ed5c201be6e44"},
 ]
-markers = {main = "platform_system == \"Windows\"", dev = "sys_platform == \"win32\""}
-
-[[package]]
-name = "distro"
-version = "1.9.0"
-description = "Distro - an OS platform information API"
-optional = false
-python-versions = ">=3.6"
-groups = ["main"]
-files = [
-    {file = "distro-1.9.0-py3-none-any.whl", hash = "sha256:7bffd925d65168f85027d8da9af6bddab658135b840670a223589bc0c8ef02b2"},
-    {file = "distro-1.9.0.tar.gz", hash = "sha256:2fa77c6fd8940f116ee1d6b94a2f90b13b5ea8d019b98bc8bafdcabcdd9bdbed"},
-]
 
 [[package]]
 name = "h11"
@@ -74,58 +117,76 @@ version = "0.16.0"
 description = "A pure-Python, bring-your-own-I/O implementation of HTTP/1.1"
 optional = false
 python-versions = ">=3.8"
-groups = ["main", "dev"]
+groups = ["main"]
+markers = "sys_platform != \"emscripten\""
 files = [
     {file = "h11-0.16.0-py3-none-any.whl", hash = "sha256:63cf8bbe7522de3bf65932fda1d9c2772064ffb3dae62d55932da54b31cb6c86"},
     {file = "h11-0.16.0.tar.gz", hash = "sha256:4e35b956cf45792e4caa5885e69fba00bdbc6ffafbfa020300e549b208ee5ff1"},
 ]
 
 [[package]]
-name = "httpcore"
-version = "1.0.9"
+name = "httpcore2"
+version = "2.16.0"
 description = "A minimal low-level HTTP client."
 optional = false
-python-versions = ">=3.8"
-groups = ["main", "dev"]
+python-versions = ">=3.11"
+groups = ["main"]
+markers = "sys_platform != \"emscripten\""
 files = [
-    {file = "httpcore-1.0.9-py3-none-any.whl", hash = "sha256:2d400746a40668fc9dec9810239072b40b4484b640a8c38fd654a024c7a1bf55"},
-    {file = "httpcore-1.0.9.tar.gz", hash = "sha256:6e34463af53fd2ab5d807f399a9b45ea31c3dfa2276f15a2c3f00afff6e176e8"},
+    {file = "httpcore2-2.16.0-py3-none-any.whl", hash = "sha256:5c4358319f2602d984ad93182c9fee5160f22e9da7babdda870a5e7c4e2dc34f"},
+    {file = "httpcore2-2.16.0.tar.gz", hash = "sha256:c72cd0413563b17e1cd8354677dff887466f8ec2551679759e93126534428643"},
 ]
 
 [package.dependencies]
-certifi = "*"
 h11 = ">=0.16"
+truststore = ">=0.10"
 
 [package.extras]
-asyncio = ["anyio (>=4.0,<5.0)"]
+asyncio = ["anyio (>=4.5.0,<5.0)"]
 http2 = ["h2 (>=3,<5)"]
 socks = ["socksio (==1.*)"]
-trio = ["trio (>=0.22.0,<1.0)"]
+trio = ["trio (>=0.34.0,<1.0)"]
 
 [[package]]
-name = "httpx"
-version = "0.28.1"
+name = "httpx2"
+version = "2.16.0"
 description = "The next generation HTTP client."
 optional = false
-python-versions = ">=3.8"
-groups = ["main", "dev"]
+python-versions = ">=3.11"
+groups = ["main"]
 files = [
-    {file = "httpx-0.28.1-py3-none-any.whl", hash = "sha256:d909fcccc110f8c7faf814ca82a9a4d816bc5a6dbfea25d6591d6985b8ba59ad"},
-    {file = "httpx-0.28.1.tar.gz", hash = "sha256:75e98c5f16b0f35b567856f597f06ff2270a374470a5c2392242528e3e3e42fc"},
+    {file = "httpx2-2.16.0-py3-none-any.whl", hash = "sha256:fd2c5d7909e8d86fe92d53020c310ca0e4e431b838c207c5754b9bbf0308d49a"},
+    {file = "httpx2-2.16.0.tar.gz", hash = "sha256:0d54c98128e8f2ff7fbbde0fca8c359e5f032edeaaaac443cdb7f31f685bcd7f"},
 ]
 
 [package.dependencies]
-anyio = "*"
-certifi = "*"
-httpcore = "==1.*"
-idna = "*"
+anyio = {version = ">=4.10", markers = "sys_platform != \"emscripten\""}
+httpcore2 = {version = "2.16.0", markers = "sys_platform != \"emscripten\""}
+httpx2-jsfetch = {version = "*", markers = "sys_platform == \"emscripten\" and python_version >= \"3.12\""}
+idna = ">=3.18"
+truststore = {version = ">=0.10", markers = "sys_platform != \"emscripten\""}
+typing-extensions = {version = ">=4.5.0", markers = "python_version < \"3.13\""}
 
 [package.extras]
-brotli = ["brotli ; platform_python_implementation == \"CPython\"", "brotlicffi ; platform_python_implementation != \"CPython\""]
-cli = ["click (==8.*)", "pygments (==2.*)", "rich (>=10,<14)"]
+brotli = ["brotli (>=1.2.0) ; platform_python_implementation == \"CPython\"", "brotlicffi (>=1.2.0.2) ; platform_python_implementation != \"CPython\""]
+cli = ["click (>=8.4.2)", "pygments (==2.*)", "rich (>=10,<16)"]
 http2 = ["h2 (>=3,<5)"]
 socks = ["socksio (==1.*)"]
-zstd = ["zstandard (>=0.18.0)"]
+ws = ["wsproto (>=1.2)"]
+zstd = ["backports-zstd (>=1.0.0) ; python_version <= \"3.13\""]
+
+[[package]]
+name = "httpx2-jsfetch"
+version = "1.0"
+description = "httpx2 transports for Emscripten/Pyodide, backed by the JavaScript fetch API."
+optional = false
+python-versions = ">=3.12"
+groups = ["main"]
+markers = "sys_platform == \"emscripten\" and python_version >= \"3.12\""
+files = [
+    {file = "httpx2_jsfetch-1.0-py3-none-any.whl", hash = "sha256:cb916b707601e69a07721aabc8f3f6659be3a6893bc1ff5c6f9e02241df2da32"},
+    {file = "httpx2_jsfetch-1.0.tar.gz", hash = "sha256:70a0e3eabfef7cce5ad9c629f7d01ca05e418f586646f4ddf14782e4c1454c60"},
+]
 
 [[package]]
 name = "idna"
@@ -133,7 +194,7 @@ version = "3.18"
 description = "Internationalized Domain Names in Applications (IDNA)"
 optional = false
 python-versions = ">=3.9"
-groups = ["main", "dev"]
+groups = ["main"]
 files = [
     {file = "idna-3.18-py3-none-any.whl", hash = "sha256:7f952cbe720b688055e3f87de14f5c3e5fdaa8bc3928985c4077ca689de849a2"},
     {file = "idna-3.18.tar.gz", hash = "sha256:ffb385a7e039654cef1ab9ef32c6fafe283c0c0467bba1d9029738ce4a14a848"},
@@ -156,280 +217,346 @@ files = [
 
 [[package]]
 name = "jiter"
-version = "0.15.0"
+version = "0.18.0"
 description = "Fast iterable JSON parser."
 optional = false
-python-versions = ">=3.9"
+python-versions = ">=3.10"
 groups = ["main"]
 files = [
-    {file = "jiter-0.15.0-cp310-cp310-macosx_10_12_x86_64.whl", hash = "sha256:edebcf7d1f601199084bb6e844d7dc67e03e04f6ac786b0332d616635c4ff7a4"},
-    {file = "jiter-0.15.0-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:9f924585cdacf631cd382b657966847bb537bf9ed0a6f9b991da5f05a631480f"},
-    {file = "jiter-0.15.0-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:abbf258599526ad0326fe51e252e24f2bd6f24f1852681b4b78feda3808f1d18"},
-    {file = "jiter-0.15.0-cp310-cp310-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:7c468136b8bd6bb18c8786e4236a1fa27362f24cb23450ba0cb204ab379b8e6f"},
-    {file = "jiter-0.15.0-cp310-cp310-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:05906b93d72f03339e6bb7cf8dc10ebda64a0266126eed6beba79e20abcf5fd4"},
-    {file = "jiter-0.15.0-cp310-cp310-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:30ce785d2adb8e32c3f7741442370a74834ec4c01f3c48f0750227a0b4ef27d6"},
-    {file = "jiter-0.15.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:2fd73e3da91a0a722d67165e849ce2cdc10de0e0d48738c142be8c6c5f310f4c"},
-    {file = "jiter-0.15.0-cp310-cp310-manylinux_2_31_riscv64.whl", hash = "sha256:ceb8fc27d38793f9c97149be8302720c5b22e5c195a37bf2c45dc36c4600a512"},
-    {file = "jiter-0.15.0-cp310-cp310-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:d726e3ceeb337191324b49de298142f27c3ad10886341555d1d5315b5f252c6a"},
-    {file = "jiter-0.15.0-cp310-cp310-musllinux_1_1_aarch64.whl", hash = "sha256:2c8aea7781d2a372227871de4e1a1332aa96f5a89fd76c5e835dafdbad102887"},
-    {file = "jiter-0.15.0-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:cf4bd113a69c0a740e27cb962ce10630c36d2b8f59d759a651b955ee9d18a823"},
-    {file = "jiter-0.15.0-cp310-cp310-win32.whl", hash = "sha256:d92a5cd21fdb083931d546c207aa29633787c5dc5b02daab2d32b843f88a2c53"},
-    {file = "jiter-0.15.0-cp310-cp310-win_amd64.whl", hash = "sha256:e58585a58209d72691ce2d62a9147445f5a87beb0bde97fde284c96ae392a3d1"},
-    {file = "jiter-0.15.0-cp311-cp311-macosx_10_12_x86_64.whl", hash = "sha256:0f862193b8696249d22ec433e85fd2ab0ad9596bc3e45e6c0bc55e8aeba97be2"},
-    {file = "jiter-0.15.0-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:1303d4d68a9b051ea90502402063ecf3807da00ad2affa19ca1ae3b90b3c5f67"},
-    {file = "jiter-0.15.0-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:392b8ab019e5502d08aff85c6272209c24bc2cbe706ea82a56368f524236614a"},
-    {file = "jiter-0.15.0-cp311-cp311-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:773b6eb282ce11ee19f05f6b2d4404fa308e5bbd353b0b80a0262caad6db2cd7"},
-    {file = "jiter-0.15.0-cp311-cp311-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:8d2c0c44d569ce0f2850f5c926f8caeb5f245fbc84475aeb36efccc2103e6dbd"},
-    {file = "jiter-0.15.0-cp311-cp311-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:032396229564bca02440396bd327710719f724f5e7b7e9f7a8eb3faa4a2c2281"},
-    {file = "jiter-0.15.0-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:f3d37768fce7f88dd2a8c6091f2325dea27d30d30d5c6e7a1c0f0af77723b708"},
-    {file = "jiter-0.15.0-cp311-cp311-manylinux_2_31_riscv64.whl", hash = "sha256:2c9cb907439d20bd0c7d7565ca01ee52234203208433749bae5b516907526928"},
-    {file = "jiter-0.15.0-cp311-cp311-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:9100ddbec09741cc66feb0fc6773f8bdbd0e3c345689368f260082ff85dcc0cd"},
-    {file = "jiter-0.15.0-cp311-cp311-musllinux_1_1_aarch64.whl", hash = "sha256:ae1b0d82ac2d987f9ea512b1c9adfcc71a28de3dea3a6039b54d76cffda9901e"},
-    {file = "jiter-0.15.0-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:8020c99ec13a7db2b6f96cbe82ef4721c88b426a4892f27478044af0284615ef"},
-    {file = "jiter-0.15.0-cp311-cp311-win32.whl", hash = "sha256:42bfb257930800cf43e7c62c832402c704ab60797c992faf88d20e903eac8f32"},
-    {file = "jiter-0.15.0-cp311-cp311-win_amd64.whl", hash = "sha256:860a74063284a2ae9bfedd694f299cc2c68e2696c5f3d440cc9d18bb81b9dd04"},
-    {file = "jiter-0.15.0-cp311-cp311-win_arm64.whl", hash = "sha256:37a10c377ce3a4a85f4a67f28b7afe093154cde77eaf248a72e856aa08b4d865"},
-    {file = "jiter-0.15.0-cp312-cp312-macosx_10_12_x86_64.whl", hash = "sha256:0e90a1c315a0226ec822d973817967f9223b7701546c8c2a7913e7ab0926294d"},
-    {file = "jiter-0.15.0-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:8c9004af7c8d67cce7f1aae1026fb55607f4aa600710d08ede3a3ce4aeefe7e0"},
-    {file = "jiter-0.15.0-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:c210f8b35dc6f30aafd4b4365ca89b9d1189f21ab49b8e68fa6322a847aef138"},
-    {file = "jiter-0.15.0-cp312-cp312-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:5f30bae8bc1c2d613e28e5af3e8cceb09b742f1c8a8a5f839fb67afaffc03b61"},
-    {file = "jiter-0.15.0-cp312-cp312-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:c60e71b6d10cfc284c9bf36bd885e8d44c46f688ce50aa91b5edd90181dea687"},
-    {file = "jiter-0.15.0-cp312-cp312-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:0ab068bce62a45aa3e7367eceaffb5dde60b7eb853be8dece45132e3d0ff4879"},
-    {file = "jiter-0.15.0-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:fa248c9eb220197d363f688818dac2fd4b2f0cd7d843ca7105d652034823427d"},
-    {file = "jiter-0.15.0-cp312-cp312-manylinux_2_31_riscv64.whl", hash = "sha256:2a77aadd57cac1682e4401a72724d2796d89a4ba129b1a5812aa94ee480826eb"},
-    {file = "jiter-0.15.0-cp312-cp312-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:2ae901f3a55bfafdde31d289590fa25e3245735a2b1e8c7cc15871710a002871"},
-    {file = "jiter-0.15.0-cp312-cp312-musllinux_1_1_aarch64.whl", hash = "sha256:f0b271b462769543716f92d3a4f90527df6ef5ed05ee95ec4137f513e21e1b77"},
-    {file = "jiter-0.15.0-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:2fb6a5d26af81fc0f00f9360a891e05cf755e149bba391c4d563adc54812973d"},
-    {file = "jiter-0.15.0-cp312-cp312-win32.whl", hash = "sha256:c2f6bb8b5216ab9e7873bc08b5d7bef2b8abbb578a3069bf1cd14a45d71d771d"},
-    {file = "jiter-0.15.0-cp312-cp312-win_amd64.whl", hash = "sha256:40b2c7e92c44a84d748d21706c68dc6ff8161d80b59c99d774721a0d2317d7c7"},
-    {file = "jiter-0.15.0-cp312-cp312-win_arm64.whl", hash = "sha256:cc0bc345cf2df9d1c00ac443f50d543c1ccfa8b0422cb85b1ab70d681c0b255b"},
-    {file = "jiter-0.15.0-cp313-cp313-macosx_10_12_x86_64.whl", hash = "sha256:1c11465f97e2abf45a014b83b730222f8f1c5335e802c7055a67d50de6f1f4e3"},
-    {file = "jiter-0.15.0-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:d1e7b1776f0797956c509e123d0952d10d293a9492dea9f288ab9570ec01d1a5"},
-    {file = "jiter-0.15.0-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:351a341c2105aa430b7047e30f1bf7975f6313b00165d3fc07be2edaf741f279"},
-    {file = "jiter-0.15.0-cp313-cp313-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:4ab395feec8d249ec4044e228e98a7033f043426a265df439dc3698823f0a4e4"},
-    {file = "jiter-0.15.0-cp313-cp313-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:a2a438005b6f22d0273413484d6094d7c2c5d10ec1b3a3bf128e0d1d3ba53258"},
-    {file = "jiter-0.15.0-cp313-cp313-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:f18f85e4218d1b40f000f42a92239a7a61a902cd42c65e6c360dbd17dcb20894"},
-    {file = "jiter-0.15.0-cp313-cp313-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:d1aa62e277fc1cbd80e6deacae6f4d983b41b3d7728e0645c5d741a6149bba45"},
-    {file = "jiter-0.15.0-cp313-cp313-manylinux_2_31_riscv64.whl", hash = "sha256:6550fa135c7deb8ead6af49ed7ff648532ea8334a1447fe34a36315ef79c5c29"},
-    {file = "jiter-0.15.0-cp313-cp313-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:066f8f33f18b2419cd8213b2436fa7fbc9c499f315971cfa3ce1f9820c001b1b"},
-    {file = "jiter-0.15.0-cp313-cp313-musllinux_1_1_aarch64.whl", hash = "sha256:75e8a04e91432dde9f1838373cf93d23726c79d3e908d319acf0e796f85592e7"},
-    {file = "jiter-0.15.0-cp313-cp313-musllinux_1_1_x86_64.whl", hash = "sha256:a97261f1fccb8e50ecd2890a96e46efdc3f57c80a197324c6777827231eca712"},
-    {file = "jiter-0.15.0-cp313-cp313-win32.whl", hash = "sha256:c77496cb10bd7549690fbbab3e5ec05857b83e49276f4a9423a766ddd2afcd4c"},
-    {file = "jiter-0.15.0-cp313-cp313-win_amd64.whl", hash = "sha256:b15741f501469009ae0ae90b7147958a664a7dede40aa7ff174a8a4645f546d0"},
-    {file = "jiter-0.15.0-cp313-cp313-win_arm64.whl", hash = "sha256:5d6a60072b44c3c2b797a7ddcbcbbf2b34ea3cfd4721580fbfd2a09d9d9b84ba"},
-    {file = "jiter-0.15.0-cp313-cp313t-macosx_11_0_arm64.whl", hash = "sha256:ef1fd24d9413f6209e00d3d5a453e67acfe004a25cc6c8e8484faed4311ab9e8"},
-    {file = "jiter-0.15.0-cp313-cp313t-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:144f8e72cb53dab146347b91cceac01f5481237f2b93b4a339a1ee8f8878b67c"},
-    {file = "jiter-0.15.0-cp313-cp313t-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:553fcac2ef2cb990877f9fc0833b8b629a3e6a5670b6b5fd58219b41a653ddc4"},
-    {file = "jiter-0.15.0-cp313-cp313t-win_amd64.whl", hash = "sha256:774f93f65031856bf14ad9f59bdcab8b8cad501e5ceabd51ba3525f76937a25b"},
-    {file = "jiter-0.15.0-cp313-cp313t-win_arm64.whl", hash = "sha256:f1e1754960f38ec40613a07e5e372df67acb3b890fb383b6fb3de3e49ddbf3c7"},
-    {file = "jiter-0.15.0-cp314-cp314-macosx_10_12_x86_64.whl", hash = "sha256:ac0d9ddea4350974be7a221fc25895f251a8fee748c889bdced2141c0fec1a49"},
-    {file = "jiter-0.15.0-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:01a8222cf05ab1128e239421156c207949808acaaea2bdfd33130ae666786e86"},
-    {file = "jiter-0.15.0-cp314-cp314-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:182226cbc930c9fab81bc2e41a4da672f89539906dadb05e75670ac07b94f71f"},
-    {file = "jiter-0.15.0-cp314-cp314-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:71683c38c825452999b5717fcae07ea708e8c93003e808be4319c1b02e3d176e"},
-    {file = "jiter-0.15.0-cp314-cp314-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:30f2218e6a9e5c18bc10fe6d41ac189c442c88eacf11bad9f28ef95a9bef00e6"},
-    {file = "jiter-0.15.0-cp314-cp314-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:5157de9f76eb4bc5ea74a1219366a25f945ad305641d74e04f59c54087091aa9"},
-    {file = "jiter-0.15.0-cp314-cp314-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:90c5db5527c221249a876160663ab891ace358c17f7b9c93ec1478b7f0550e5c"},
-    {file = "jiter-0.15.0-cp314-cp314-manylinux_2_31_riscv64.whl", hash = "sha256:3e4540b8e74e4268811ac05db226a6a128ff572e7e0ce3f1163b693cadb184cd"},
-    {file = "jiter-0.15.0-cp314-cp314-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:62ebd14e47e9aed9df4472afcb2663668ce4d74891cd54f86bf6e44029d6dc89"},
-    {file = "jiter-0.15.0-cp314-cp314-musllinux_1_1_aarch64.whl", hash = "sha256:0be6f5ad41a809f303f416d17cec92a7a725902fb9b4f3de3d19362ac0ef8554"},
-    {file = "jiter-0.15.0-cp314-cp314-musllinux_1_1_x86_64.whl", hash = "sha256:813dfbb17d65328bf86e5f0905dd277ba2265d3ca20556e86c0c7035b7182e5a"},
-    {file = "jiter-0.15.0-cp314-cp314-win32.whl", hash = "sha256:50e51156192722a9c58db112837d3f8ef96fb3c5ecc14e95f409134b08b158ec"},
-    {file = "jiter-0.15.0-cp314-cp314-win_amd64.whl", hash = "sha256:30ce1a5d16b5641dc935d50ef775af6a0871e3d14ab05d6fc54dff371b78e558"},
-    {file = "jiter-0.15.0-cp314-cp314-win_arm64.whl", hash = "sha256:510c8b3c17a0ed9ac69850c0438dada3c9b82d9c4d589fcb62002a5a9cf3a866"},
-    {file = "jiter-0.15.0-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:7553333dd0930c104a5a0db8df72bf7219fe663d731383b576bb6ed6351c984d"},
-    {file = "jiter-0.15.0-cp314-cp314t-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:f2143ab06181d2b029eedcb6af3cebe95f11bbac62441781860f98ee9330a6a6"},
-    {file = "jiter-0.15.0-cp314-cp314t-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:6eac374c5c975709b69c10f09afd199df74150172156ad10c8d4fd785b7da995"},
-    {file = "jiter-0.15.0-cp314-cp314t-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:b3b3b775e33d3bfaec9899edc526ae97b0da0bf9d071a46124ba419149a414f8"},
-    {file = "jiter-0.15.0-cp314-cp314t-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:eda3071db3346334beae1360b46da4606da57bf3528c167b3c38533afaf9f2c5"},
-    {file = "jiter-0.15.0-cp314-cp314t-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:c6694a173ecabc12eb60efbc0b474464ead1951ff65cd8b1e72100715c64512b"},
-    {file = "jiter-0.15.0-cp314-cp314t-manylinux_2_31_riscv64.whl", hash = "sha256:a254e10b593624d230c365b6d616b22ca0ad65e63a16e6631c2b3466022e6ba8"},
-    {file = "jiter-0.15.0-cp314-cp314t-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:d8d2955167274e15d79a7a020afdd9b39c990eb80b2d89fca695d92dcfdd38ec"},
-    {file = "jiter-0.15.0-cp314-cp314t-musllinux_1_1_aarch64.whl", hash = "sha256:acf4ee4d1fc55917239fe72972fb292dd773055d05eb040d36f4326e02cc2c0e"},
-    {file = "jiter-0.15.0-cp314-cp314t-musllinux_1_1_x86_64.whl", hash = "sha256:e7196e56f1cd69af1dbb07dff02dcfb260a50b45a82d409d92a06fedb32473b5"},
-    {file = "jiter-0.15.0-cp314-cp314t-win32.whl", hash = "sha256:7f6163c0f10b055245f814dcc59f4818da60dfe72f3e72ab89fc24b6bd5e9c52"},
-    {file = "jiter-0.15.0-cp314-cp314t-win_amd64.whl", hash = "sha256:980c256edb05b78a111b99c4de3b1d32e31634b867fd1fc2cf726e7b7bba9854"},
-    {file = "jiter-0.15.0-cp314-cp314t-win_arm64.whl", hash = "sha256:66b1880df2d01e206e8339769d1c7c1753bcb653efd6289e203f6f24ebada0c0"},
-    {file = "jiter-0.15.0-cp39-cp39-macosx_10_12_x86_64.whl", hash = "sha256:04b400bbf8c9efb03d9bdd976475c919c1d85593b04b9fff7ae234065daf87ae"},
-    {file = "jiter-0.15.0-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:25ffbe229aa8cd98c28879d8aa1a6e34ae77992ab984a65fba800859dab16269"},
-    {file = "jiter-0.15.0-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:5607e6013ed7e6b0ec9661e467b7ffde0aa7ab36833a04850f26fcf88ed4845b"},
-    {file = "jiter-0.15.0-cp39-cp39-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:50164d7610c00e7cd913a873fce30b6beeebf4b37e53983e33f22de4c900f6b8"},
-    {file = "jiter-0.15.0-cp39-cp39-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:ab596fa3837e91e7e6a31b5f639988bfc6a35d1f915ac3932d946062219d588f"},
-    {file = "jiter-0.15.0-cp39-cp39-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:d72d8af5c1013656a8870c866660627d1a75bc185814ee022c8533caa1de88ae"},
-    {file = "jiter-0.15.0-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:c84c1b7be454b0c16f8499b4ebfbfd82ea5cca6527cceefcbbc06a7557b5ed2e"},
-    {file = "jiter-0.15.0-cp39-cp39-manylinux_2_31_riscv64.whl", hash = "sha256:d636d5095155afd364247f65070fab7beda13498d7ff4de331046e704ab9657f"},
-    {file = "jiter-0.15.0-cp39-cp39-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:7d3d6683288c11cbab50e865f2e2f13950179aa45410e30b2cfbd3fb7b0177bf"},
-    {file = "jiter-0.15.0-cp39-cp39-musllinux_1_1_aarch64.whl", hash = "sha256:7ce8902f939970048b233087082e7bb829db29375811c7ad50687b8624c6fd08"},
-    {file = "jiter-0.15.0-cp39-cp39-musllinux_1_1_x86_64.whl", hash = "sha256:4363818355dbc70ae1a8e9eaba9de350d93ede4ff6992b8f8eb8cbb6e5122d42"},
-    {file = "jiter-0.15.0-cp39-cp39-win32.whl", hash = "sha256:8f7e9bc0f1135039b22ee6eab588d42df1ce55842b30740a352885eb267bd941"},
-    {file = "jiter-0.15.0-cp39-cp39-win_amd64.whl", hash = "sha256:1c15024a3d892223b18f597c86d59387249dc396590844ce6b9f6131d1093bae"},
-    {file = "jiter-0.15.0-graalpy311-graalpy242_311_native-macosx_10_12_x86_64.whl", hash = "sha256:411fa4dfa5a7ae3d11491027ffb9beadec3996010a986862db70d91abba1c750"},
-    {file = "jiter-0.15.0-graalpy311-graalpy242_311_native-macosx_11_0_arm64.whl", hash = "sha256:2b0074e2f56eb2dacca1689760fd2852a068f85a0547a157b82cb4cafeb6768b"},
-    {file = "jiter-0.15.0-graalpy311-graalpy242_311_native-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:913d02d29c9606643418d9ccfc3b72492ab25a6bf7889934e09a3490f8d3438b"},
-    {file = "jiter-0.15.0-graalpy311-graalpy242_311_native-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:b15d3ec9b0449c40e85319bdb4caa8b77ab526e74f5532ed94bec15e2f66822c"},
-    {file = "jiter-0.15.0-graalpy312-graalpy250_312_native-macosx_10_12_x86_64.whl", hash = "sha256:631f13a3d04e97d4e083993b10f4b99530e3a10d953e2eb5e196b7dc7f812ce0"},
-    {file = "jiter-0.15.0-graalpy312-graalpy250_312_native-macosx_11_0_arm64.whl", hash = "sha256:b6c0ffae686c39bf3737be60793783267628783ea42545632c10b291105aee45"},
-    {file = "jiter-0.15.0-graalpy312-graalpy250_312_native-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:1d54fb5b31dea401a41af3f8a7d2512e9b6a6a005491e6166c7e4ffab9639a9c"},
-    {file = "jiter-0.15.0-graalpy312-graalpy250_312_native-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:54d5d6090cdc1b7c9e780dfb04949a990adb1e301a2fc0bbcee7de4638d33f9a"},
-    {file = "jiter-0.15.0.tar.gz", hash = "sha256:4251acc80e2b7c9b7b8823456ea0fceeb0734dac2df7636d3c711b38476b5a76"},
+    {file = "jiter-0.18.0-cp310-cp310-macosx_10_12_x86_64.whl", hash = "sha256:74b8d31a39b9d22293c75bfdf0e6ab54d9dc16d32b540b2cb8f7c162961769ee"},
+    {file = "jiter-0.18.0-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:77d61c53ab1732e84ad20f77c44bfc761183ef323eef84a59ffdd9201bd51386"},
+    {file = "jiter-0.18.0-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:6ea6d553fe513447a7a8ef6175aeacecd9a6dbb7f68f9215c751389c58e1288c"},
+    {file = "jiter-0.18.0-cp310-cp310-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:7779088592f8994e021cd79b9b334f41363cfa10a654af9bd5dd82b99c80497f"},
+    {file = "jiter-0.18.0-cp310-cp310-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:879cf95bb0266643c1e8214be0862f2dd7186c9bb2e279f63d7b5fe099cb9a4b"},
+    {file = "jiter-0.18.0-cp310-cp310-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:1b291352b1f7dcf095b2a9d540777d59588c81685cc23f7c53f4356e07385e62"},
+    {file = "jiter-0.18.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:a4ff2d7905b0f3f5c54b29626c7747cd4af3fcddfdedf0cce75fbd06b66e4d40"},
+    {file = "jiter-0.18.0-cp310-cp310-manylinux_2_31_riscv64.whl", hash = "sha256:4d5ce30facd564d97aceb49013a02661daef1cf9c34ce322c1afbb841f1e5984"},
+    {file = "jiter-0.18.0-cp310-cp310-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:131446a60b54ff9d4db62633a74189056387f6b2ad4e5171d3e7ed7977e31eb4"},
+    {file = "jiter-0.18.0-cp310-cp310-musllinux_1_1_aarch64.whl", hash = "sha256:389f5379f7cc5b09f6586db99bf91c5d828fac88f1a52fdeb4fc3b040c1dacc9"},
+    {file = "jiter-0.18.0-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:db7804a8d3a433a3c5e08cd658ea69a6ee9a2ee43b756dd20048dae5a3d7ac67"},
+    {file = "jiter-0.18.0-cp310-cp310-win32.whl", hash = "sha256:5c2a5dac57aec416140812f6d3824cdaf1da8a4f8f009ece64e50f507b6a709a"},
+    {file = "jiter-0.18.0-cp310-cp310-win_amd64.whl", hash = "sha256:a447ed784b8013b518e0b1effe05424d5dfaa6974fed69aebdd4c1349d82fc54"},
+    {file = "jiter-0.18.0-cp311-cp311-macosx_10_12_x86_64.whl", hash = "sha256:97cdff6fffc79d915e3d2691110e014f3f43c7529a73c0bfe8de63530b6efbd9"},
+    {file = "jiter-0.18.0-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:c938de3832ca88ca879bbf0c24673cbe0176b502eec8d2364f35c9edb9bdf890"},
+    {file = "jiter-0.18.0-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:0a41d8ef9f757fd5c11ce2fa56c9c4d008d857e97b6d0edfccbaacbd50830a75"},
+    {file = "jiter-0.18.0-cp311-cp311-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:b2cad787346aa8f458ed6d5ecbbe3b6dd13d6d40fdbaa44980c9e29a8b536048"},
+    {file = "jiter-0.18.0-cp311-cp311-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:3027e8ec2c86106bfd4b125af27aa6dd5eed2ab63427c35138e039931a98ba7c"},
+    {file = "jiter-0.18.0-cp311-cp311-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:2485807c282b6df64d9135551c2d8e9ac2b098b4c85157979fc7ca8bd884a3b4"},
+    {file = "jiter-0.18.0-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:55f808bce951fae459fc916359cc54097d386bef3e72265ddf02b39ba3275118"},
+    {file = "jiter-0.18.0-cp311-cp311-manylinux_2_31_riscv64.whl", hash = "sha256:4656c4b24fd711d6caa61cb2fe895dcc55d0d6944828df23f2fb05504435204b"},
+    {file = "jiter-0.18.0-cp311-cp311-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:5893dc87a2a5297e271dfc5f8f710c3e61ebeb64c137a0b4b786d5bcd11c8fb7"},
+    {file = "jiter-0.18.0-cp311-cp311-musllinux_1_1_aarch64.whl", hash = "sha256:867a217d4a1857fba97c4cc5dc82a38ea257beee8c2a753b3d9b9310c01d12c0"},
+    {file = "jiter-0.18.0-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:1df7b0ce46d869c49c77670ba31db737d96fd1bd4ae44d790d33bb731962c994"},
+    {file = "jiter-0.18.0-cp311-cp311-win32.whl", hash = "sha256:b8b97658b0914e90a352a9a1d515e597bcc89c88514fe4bcb4e37fa2a2fbaa29"},
+    {file = "jiter-0.18.0-cp311-cp311-win_amd64.whl", hash = "sha256:8ad1a36e3d5ebbff515b85468d8e279b336eaa326953b1b36fa21d5e2b8e8f9d"},
+    {file = "jiter-0.18.0-cp311-cp311-win_arm64.whl", hash = "sha256:908f310576554379becf2e8c7730a8962693ebfba9395a54ef3c05c57b1cd9d8"},
+    {file = "jiter-0.18.0-cp312-cp312-macosx_10_12_x86_64.whl", hash = "sha256:77a01cdf4381a1becbe5b827f682006f47b712581667451c5b1eb259b9ae7041"},
+    {file = "jiter-0.18.0-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:7b41cd64ee6b54db94bea6274e6e37002ed526aa06c114a8be6f36ffe21c184e"},
+    {file = "jiter-0.18.0-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:d7424194e68ef44e94290961f59319850d7ce11a329a3957f91709c15ded6618"},
+    {file = "jiter-0.18.0-cp312-cp312-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:04a484fce032a8b2c21dec0410a805d7b16573eb34297156a3b4d242c507d2b4"},
+    {file = "jiter-0.18.0-cp312-cp312-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:860ac97858116bda0ac02ef5f50c541fc1a84ad7b9368cac68d4de5697a9012d"},
+    {file = "jiter-0.18.0-cp312-cp312-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:39ef656cf371971486ec545ca0ae2da8d037d5479bc906121e3e46de86ee73c4"},
+    {file = "jiter-0.18.0-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:b3c741e6b778c951a6861964ee475118812a33208426c54fb92bf4e318bda834"},
+    {file = "jiter-0.18.0-cp312-cp312-manylinux_2_31_riscv64.whl", hash = "sha256:1277ac083a58d2f49a24ff9a28fd9876a6443a1992d4467a1db80fb45e354b18"},
+    {file = "jiter-0.18.0-cp312-cp312-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:ccf224d2e706be81d464b59c2bc3f98ba65d12c092c1789e625127e64488944e"},
+    {file = "jiter-0.18.0-cp312-cp312-musllinux_1_1_aarch64.whl", hash = "sha256:5c7fc8e5357d1ad082bb177be71e398fcd3d458726439255ca5cb0d706060269"},
+    {file = "jiter-0.18.0-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:a6bba0404b46157a651d1fe512cf2c59c8e42836a9284c35cdff35310758ba4f"},
+    {file = "jiter-0.18.0-cp312-cp312-win32.whl", hash = "sha256:07ae058bc052ead609b49bbae78c01bf4dee8f5f9d21705cbae9bfe59b33efc9"},
+    {file = "jiter-0.18.0-cp312-cp312-win_amd64.whl", hash = "sha256:79d1fe529bd7065c6f3f50279a08c2ab82eaa0631a6f5dfafa9c65105eeaaafa"},
+    {file = "jiter-0.18.0-cp312-cp312-win_arm64.whl", hash = "sha256:1f8b3b3d8761dee3cddcb5d3e99eaccc0155bbefc345affcdedcc758a4933a1f"},
+    {file = "jiter-0.18.0-cp313-cp313-macosx_10_12_x86_64.whl", hash = "sha256:1517648b5f6268513dea206241a99d9991d6ffcbdbc406848e77a7f02ed3ccf1"},
+    {file = "jiter-0.18.0-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:7161217e2190b9b2c8c79b81699be29d67010f6f77e09a53a3b62261792d5489"},
+    {file = "jiter-0.18.0-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:9623fc90f1a4754452a525383ce70476476c16556098944709b8cdbd9ad905f0"},
+    {file = "jiter-0.18.0-cp313-cp313-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:f9083a7c2a4ee6358bc0401c46b4d5090f208f6fadfb6275c7b5fe3217e13506"},
+    {file = "jiter-0.18.0-cp313-cp313-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:4632f5172800da4cd42fd53f4553c772acd08f1c33dd7ac135e29d735080fa0d"},
+    {file = "jiter-0.18.0-cp313-cp313-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:c00b1f8fc1dd06c610305be36778b38148ed1e9a532bdd81189c3c0d9905c888"},
+    {file = "jiter-0.18.0-cp313-cp313-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:6e3350b7431f7b8315b925d9ca0f29a44e8fb6ed45c1fe552392cb6297966bb9"},
+    {file = "jiter-0.18.0-cp313-cp313-manylinux_2_31_riscv64.whl", hash = "sha256:3b49071fb7917cdbc6d80c81a491c6ad146c1a8fd6cb3076cd3c1bfeb6e266f6"},
+    {file = "jiter-0.18.0-cp313-cp313-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:04e210f5ddcccdbde484821040216a112381e6484abc549c20c38e9698ab07ca"},
+    {file = "jiter-0.18.0-cp313-cp313-musllinux_1_1_aarch64.whl", hash = "sha256:5b339f6ae14fe7be92f391cc65cc5c80ff76e8da76d63b18fc14664f09b7cece"},
+    {file = "jiter-0.18.0-cp313-cp313-musllinux_1_1_x86_64.whl", hash = "sha256:5cb43336244e063f9cc11e44730e4b92dea70c6eb9b649d268c5ebebc4bc7685"},
+    {file = "jiter-0.18.0-cp313-cp313-win32.whl", hash = "sha256:9656d0e9001f69c4af6523874ff445ab4e5db7b835f59733cc5ff51e73a9aa57"},
+    {file = "jiter-0.18.0-cp313-cp313-win_amd64.whl", hash = "sha256:8f26198220165eb9627847401fff4d0ca8bebf3ef840a88c6312918d2291c33a"},
+    {file = "jiter-0.18.0-cp313-cp313-win_arm64.whl", hash = "sha256:bef1fe687dc6af17bef9f644a95d465f15f0efa20b743c858e0b6a14e2c0c73e"},
+    {file = "jiter-0.18.0-cp314-cp314-macosx_10_12_x86_64.whl", hash = "sha256:248e939f67a9f8030e2624260a9f417decfe275af37aa96728d256468a9c4086"},
+    {file = "jiter-0.18.0-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:0390fa550fcafbf69fb97ccb287bd71192f90410a1c7dd600ca4000aa9aabcf2"},
+    {file = "jiter-0.18.0-cp314-cp314-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:d8013186f2b9a8c27ededa601050392527aa9d27ffd35d49aae424fbd2062f38"},
+    {file = "jiter-0.18.0-cp314-cp314-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:bd4316df8e2074fcf5173f6001f1fc9596e7c1e8c4bf499b02f9ebe92babdfb1"},
+    {file = "jiter-0.18.0-cp314-cp314-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:a9e474da7a9f65d3b460cca93cc8a0f4ec0220d1fa41f1f7a3709195ead60056"},
+    {file = "jiter-0.18.0-cp314-cp314-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:d55acd326db8ecfc94f21951abd20902c7939d35925444e48d94527e287bead7"},
+    {file = "jiter-0.18.0-cp314-cp314-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:a01642d403b8925f2b1e1bb51b4ba49ece8a2bdc9b9560fc8ed0b1f2b35b567a"},
+    {file = "jiter-0.18.0-cp314-cp314-manylinux_2_31_riscv64.whl", hash = "sha256:89ba6b674d4bd7290f9943bbeba0e1bf35955bbb1e07e41ba94691648cf133d5"},
+    {file = "jiter-0.18.0-cp314-cp314-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:feadebdfee32c61e91066433f4aa74c891752e26e43032c1feef08dda6d9b6ed"},
+    {file = "jiter-0.18.0-cp314-cp314-musllinux_1_1_aarch64.whl", hash = "sha256:dd7e57efd54a3580620f635c48a3226f910e23649b7eb83ffbc6f0f3386293b8"},
+    {file = "jiter-0.18.0-cp314-cp314-musllinux_1_1_x86_64.whl", hash = "sha256:ea69e23481497c2143775a98d273ccc43de5c92a6387df18c903d2d37bc2b181"},
+    {file = "jiter-0.18.0-cp314-cp314-pyemscripten_2026_0_wasm32.whl", hash = "sha256:d9b61af76a3cd90ad9656a331737d97dab8c05facb2d04cb41fb16c9adadc966"},
+    {file = "jiter-0.18.0-cp314-cp314-win32.whl", hash = "sha256:fe42726d7fe171cd8fc1966d4457a6aae03bd408d519d2cc5e7726805460b51e"},
+    {file = "jiter-0.18.0-cp314-cp314-win_amd64.whl", hash = "sha256:f1e738b851019d6bb6ff4a42c6319bdede967dec958ef3ec6feb7ab9d12408f7"},
+    {file = "jiter-0.18.0-cp314-cp314-win_arm64.whl", hash = "sha256:80a961b1c5c7786bc03b332deb7f7ebe818b7baa6e26685e2bf480548ca128b4"},
+    {file = "jiter-0.18.0-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:b066f37dd9ec19fc7931d79dee38f56b6f849217bbe824a7dad4518ef96ee515"},
+    {file = "jiter-0.18.0-cp314-cp314t-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:ccdc09887b04f50f1761eedcfd19da882d9d834096bc1215049af637023f7db7"},
+    {file = "jiter-0.18.0-cp314-cp314t-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:6133abcd0f332964fbb8145adb1ce775cd0312c9eaf6e4bf26b7a5c9b6f7cca7"},
+    {file = "jiter-0.18.0-cp314-cp314t-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:ca381b4a466023d1fdd04c889c759b46f7d9654f622b28bb01e65abf31d07bba"},
+    {file = "jiter-0.18.0-cp314-cp314t-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:9b961fedf589ffa182fe302fc4371acf4606d3ed4a6a48a151fbfd724d811f5a"},
+    {file = "jiter-0.18.0-cp314-cp314t-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:3b77eabfbae1ac47872be0076899e66340bbef39b862971c3d6c9a4b7c91fd10"},
+    {file = "jiter-0.18.0-cp314-cp314t-manylinux_2_31_riscv64.whl", hash = "sha256:92a16933d0fa852133500837d33cd6a4f75e4414d77389c58269e12bafc2ee02"},
+    {file = "jiter-0.18.0-cp314-cp314t-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:2f7ba518893a4ccef5afb38e41bc3e126bf652354417ec4e75e5a5b000fb1ef6"},
+    {file = "jiter-0.18.0-cp314-cp314t-musllinux_1_1_aarch64.whl", hash = "sha256:04d67a44b040a6662a35edc9869030f7aa566bf5baaf3549c99d847aa44213e2"},
+    {file = "jiter-0.18.0-cp314-cp314t-musllinux_1_1_x86_64.whl", hash = "sha256:e4d4819398d5330636bc0b91ccc0048dba782076fb99257038257ba6f121c840"},
+    {file = "jiter-0.18.0-cp314-cp314t-win32.whl", hash = "sha256:c435e795a7ea4b372166a058a5b49a25caeb9c94e469a1414868b14c83ee5c41"},
+    {file = "jiter-0.18.0-cp314-cp314t-win_amd64.whl", hash = "sha256:5a3d5397da7ae46c7a9601ef02d2c9ec47a6c3abe15e1ea940eaf93e675ee20e"},
+    {file = "jiter-0.18.0-cp314-cp314t-win_arm64.whl", hash = "sha256:4b81b474e4a8ec745ab20a6762275a360e1147cb70cf08872ec14e1aa5debc37"},
+    {file = "jiter-0.18.0-cp315-cp315-macosx_10_12_x86_64.whl", hash = "sha256:30f8ae18425e0247ffff4e4a5753ec1952e1b81ed4e7949b84e5ecfaf8315833"},
+    {file = "jiter-0.18.0-cp315-cp315-macosx_11_0_arm64.whl", hash = "sha256:93150a3bb0317d8f5a94d4c3ea68db1cb434b94ac819e08c443cea97b7645f2f"},
+    {file = "jiter-0.18.0-cp315-cp315-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:c9c8a61d6227cdcb0a67c678f30a020837c9f3dd7e0b3f916781defa157e4eb9"},
+    {file = "jiter-0.18.0-cp315-cp315-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:285cc15ff9ffcd4ce7f0b9ce02fcbc65e22fcf7dfbda1a3d71ccac8d7b73f8a1"},
+    {file = "jiter-0.18.0-cp315-cp315-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:c8f69ef2e28f8c1759bc1522385ab3e4ade0b1056053cecf67449a947c4525a8"},
+    {file = "jiter-0.18.0-cp315-cp315-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:eec2da60423287bd60377b0fad1723560dcc51cfb2b99615be8431ad425c4438"},
+    {file = "jiter-0.18.0-cp315-cp315-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:5db2cdae6a2f01e8f52ab2154886dcf5bb1844b4f31245440bf623ec91b65b08"},
+    {file = "jiter-0.18.0-cp315-cp315-manylinux_2_31_riscv64.whl", hash = "sha256:029da1f29852beac2f9001f20209b73b29b4d7bdc748c7f8028895769e2cc0d8"},
+    {file = "jiter-0.18.0-cp315-cp315-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:67c6662b1b7e5728a507db06ad1b8a7ff8ade04c2a74f092d467209ef744d3a4"},
+    {file = "jiter-0.18.0-cp315-cp315-musllinux_1_1_aarch64.whl", hash = "sha256:4db12daef4a231df528a16f70d5ff7f5e3be21f70e8ad904e4c61442366e8a10"},
+    {file = "jiter-0.18.0-cp315-cp315-musllinux_1_1_x86_64.whl", hash = "sha256:4654ea93f525ea96adf289afcf1ead32d773b8325ec326862b4f81f5f4570835"},
+    {file = "jiter-0.18.0-cp315-cp315-win32.whl", hash = "sha256:f035d01b450ac4e141502f1ce142462cee638198ff5a0237c80e6683798b3b87"},
+    {file = "jiter-0.18.0-cp315-cp315-win_amd64.whl", hash = "sha256:4a18ef4fe95685b000d95eca797cc683e0c178a4b364f2e39f6c79ab107a26e6"},
+    {file = "jiter-0.18.0-cp315-cp315-win_arm64.whl", hash = "sha256:c73922e4ff4b1ad2929cd795c98223c6688a43c5fd42e9deb8955e4d43b00abb"},
+    {file = "jiter-0.18.0-cp315-cp315t-macosx_11_0_arm64.whl", hash = "sha256:0ccf55f6fcdcee82fe40ae617e9d64b8e5f433fa6cb5bf7d2969e584e6df0e5d"},
+    {file = "jiter-0.18.0-cp315-cp315t-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:163a55e77f8e92d739b1f4ef212f1d445ddc58ecdf0da1feca44ef424b5331f8"},
+    {file = "jiter-0.18.0-cp315-cp315t-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:9de6c9133be12a51ea2d8c270f39a04ed45ee369802bb6270649b952e6cec256"},
+    {file = "jiter-0.18.0-cp315-cp315t-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:2ea4f2ed18eba7454182d4ff727444788ae51ad11bf5a25ad892f5f0ad3e754b"},
+    {file = "jiter-0.18.0-cp315-cp315t-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:626d5c9fb8b57babc6ce9d03224846ecb094e1d22ff5bff6080398e7f8aee162"},
+    {file = "jiter-0.18.0-cp315-cp315t-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:b3c3318cc6ef857b6554b43479c30e8249fc143189d500d4137cde9c58412e6d"},
+    {file = "jiter-0.18.0-cp315-cp315t-manylinux_2_31_riscv64.whl", hash = "sha256:9a70f25f6de76a01e309e109c9bcf2143c5c126b3488dd17a25c7aa404b28f16"},
+    {file = "jiter-0.18.0-cp315-cp315t-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:22d8dd757e603224e46786d63cb7d7bfbddc8f579a02e7b082840468d18c06f3"},
+    {file = "jiter-0.18.0-cp315-cp315t-musllinux_1_1_aarch64.whl", hash = "sha256:552a8714f26ddcb43115c82da6d045b3b9f2fa5997d8b21b2f8ec3721503bebb"},
+    {file = "jiter-0.18.0-cp315-cp315t-musllinux_1_1_x86_64.whl", hash = "sha256:8060a206e290410d3bfc3bd91715f3ec90a499246f8335d48927e0b3b6a6ccc5"},
+    {file = "jiter-0.18.0-cp315-cp315t-win32.whl", hash = "sha256:91d463858f11ee76ce8e39f391bbcbf1d4c3cc4703a1496fed3a5666dcd5e01d"},
+    {file = "jiter-0.18.0-cp315-cp315t-win_amd64.whl", hash = "sha256:6ee49daab0b826279c152ad8d61171ea17f4197f91a5e14251d0d5d4adf51059"},
+    {file = "jiter-0.18.0-cp315-cp315t-win_arm64.whl", hash = "sha256:bec676a0dc24d4dabd4bdfb3d1f1bf80d4d3dd892fc5c5b3f8f953e2459dbfce"},
+    {file = "jiter-0.18.0-graalpy311-graalpy242_311_native-macosx_10_12_x86_64.whl", hash = "sha256:04aa358fd3217dedd673e9331febe8cccfa4a00382e5372b39efcafe7447ff63"},
+    {file = "jiter-0.18.0-graalpy311-graalpy242_311_native-macosx_11_0_arm64.whl", hash = "sha256:f5e5d58f589412c7774b3a2c4878b8ee54f9804840075822ad2bd2356a5080fd"},
+    {file = "jiter-0.18.0-graalpy311-graalpy242_311_native-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:df6d792b41ad90783cf9c3c8f66b02f06025250533b92997913ec600bcb0d075"},
+    {file = "jiter-0.18.0-graalpy311-graalpy242_311_native-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:8f39e7316ee0f16a1e74f23010a058e2bd59481992c7adec0470db4ea57a38e5"},
+    {file = "jiter-0.18.0-graalpy312-graalpy250_312_native-macosx_10_12_x86_64.whl", hash = "sha256:118c1fd995f349f1cad0557f31d6860f336b4aeef7f002dedbc068027f68e2c3"},
+    {file = "jiter-0.18.0-graalpy312-graalpy250_312_native-macosx_11_0_arm64.whl", hash = "sha256:44d0d719cd7845e66b01a0cf346ec9882073e710a05e89a5b6d87be47a9dc69c"},
+    {file = "jiter-0.18.0-graalpy312-graalpy250_312_native-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:4d1d2dc4d930ebc225f488b6b3790acebb2e51b3a165a4a42cc1d05bb927f83e"},
+    {file = "jiter-0.18.0-graalpy312-graalpy250_312_native-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ebb0c346b6c5d9c033fe9e9b6ed6a1ff6dac0288fe681c3434dde6eb9c264fa1"},
+    {file = "jiter-0.18.0.tar.gz", hash = "sha256:a0ca63fd9450a212e4e658d4cfe0ab4977c4f9b0003c68828b843c65e27662ff"},
 ]
 
 [[package]]
 name = "librt"
-version = "0.11.0"
+version = "0.16.0"
 description = "Mypyc runtime library"
 optional = false
 python-versions = ">=3.9"
 groups = ["dev"]
 markers = "platform_python_implementation != \"PyPy\""
 files = [
-    {file = "librt-0.11.0-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:6e94ebfcfa2d5e9926d6c3b9aa4617ffc42a845b4321fb84021b872358c82a0f"},
-    {file = "librt-0.11.0-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:ae627397a2f351560440d872d6f7c8dbb4072e57868e7b2fc5b8b430fe489d45"},
-    {file = "librt-0.11.0-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:dc329359321b67d24efdf4bc69012b0597001649544db662c001db5a0184794c"},
-    {file = "librt-0.11.0-cp310-cp310-manylinux2014_i686.manylinux_2_17_i686.manylinux_2_28_i686.whl", hash = "sha256:7e82e642ab0f7608ce2fe53d76ca2280a9ee33a1b06556142c7c6fe80a86fc33"},
-    {file = "librt-0.11.0-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:88145c15c67731d54283d135b03244028c750cc9edc334a96a4f5950ebdb2884"},
-    {file = "librt-0.11.0-cp310-cp310-manylinux_2_34_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:9d36a51b3d93320b686588e27123f4995804dbf1bce81df78c02fc3c6eea9280"},
-    {file = "librt-0.11.0-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:d00f3ac06a2a8b246327f11e186a53a100a4d5c7ed52346367e5ec751d51586c"},
-    {file = "librt-0.11.0-cp310-cp310-musllinux_1_2_i686.whl", hash = "sha256:461bbceede621f1ffb8839755f8663e886087ee7af16294cab7fb4d782c62eeb"},
-    {file = "librt-0.11.0-cp310-cp310-musllinux_1_2_riscv64.whl", hash = "sha256:0cad8a4d6a8ff03c9b76f9414caccd78e7cfbc8a2e12fa334d8e1d9932753783"},
-    {file = "librt-0.11.0-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:f37aa505b3cf60701562eddb32df74b12a9e380c207fd8b06dd157a943ac7ea0"},
-    {file = "librt-0.11.0-cp310-cp310-win32.whl", hash = "sha256:94663a21534637f0e787ec2a2a756022df6e5b7b2335a5cdd7d8e33d68a2af89"},
-    {file = "librt-0.11.0-cp310-cp310-win_amd64.whl", hash = "sha256:dec7db73758c2b54953fd8b7fe348c45188fe26b39ee18446196edd08453a5d4"},
-    {file = "librt-0.11.0-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:93d95bd45b7d58343d8b90d904450a545144eec19a002511163426f8ab1fae29"},
-    {file = "librt-0.11.0-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:4ee278c769a713638cdacd4c0436d72156e75df3ebc0166ab2b9dc43acc386c9"},
-    {file = "librt-0.11.0-cp311-cp311-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:f230cb1cbc9faaa616f9a678f530ebcf186e414b6bcbd88b960e4ba1b92428d5"},
-    {file = "librt-0.11.0-cp311-cp311-manylinux2014_i686.manylinux_2_17_i686.manylinux_2_28_i686.whl", hash = "sha256:5d63c855d86938d9de93e265c9bd8c705b51ec494de5738340ee93767a686e4b"},
-    {file = "librt-0.11.0-cp311-cp311-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:993f028be9e96a08d31df3479ac80d99be374d17f3b78e4796b3fd3c913d4e89"},
-    {file = "librt-0.11.0-cp311-cp311-manylinux_2_34_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:258d73a0aa66a055e65b2e4d1b8cdb23b9d132c5bb915d9547d804fcaed116cc"},
-    {file = "librt-0.11.0-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:0827efe7854718f04aaddf6496e96960a956e676fe1d0f04eb41511fd8ad06d5"},
-    {file = "librt-0.11.0-cp311-cp311-musllinux_1_2_i686.whl", hash = "sha256:7753e57d6e12d019c0d8786f1c09c709f4c3fcc57c3887b24e36e6c06ec938b7"},
-    {file = "librt-0.11.0-cp311-cp311-musllinux_1_2_riscv64.whl", hash = "sha256:11bd19822431cc21af9f27374e7ae2e58103c7d98bda823536a6c47f6bb2bb3d"},
-    {file = "librt-0.11.0-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:22bdf239b219d3993761a148ffa134b19e52e9989c84f845d5d7b71d70a17412"},
-    {file = "librt-0.11.0-cp311-cp311-win32.whl", hash = "sha256:46c60b61e308eb535fbd6fa622b1ee1bb2815691c1ad9c98bf7b84952ec3bc8d"},
-    {file = "librt-0.11.0-cp311-cp311-win_amd64.whl", hash = "sha256:902e546ff044f579ff1c953ff5fce97b636fe9e3943996b2177710c6ef076f73"},
-    {file = "librt-0.11.0-cp311-cp311-win_arm64.whl", hash = "sha256:65ac3bc20f78aa0ee5ae84baa68917f89fef4af63e941084dd019a0d0e749f0c"},
-    {file = "librt-0.11.0-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:b87504f1690a23b9a2cca841191a04f83895d4fc2dd04df91d82b1a04ca2ad46"},
-    {file = "librt-0.11.0-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:40071fc5fe0ce8daa6de616702314a01e1250711682b0523d6ab8d4525910cb3"},
-    {file = "librt-0.11.0-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:137e79445c896a0ea7b265f52d23954e05b64222ee1af69e2cb34219067cbb67"},
-    {file = "librt-0.11.0-cp312-cp312-manylinux2014_i686.manylinux_2_17_i686.manylinux_2_28_i686.whl", hash = "sha256:cca6644054e78746d8d4ef238681f9c34ff8b584fe6b988ecebb8db3b15e622a"},
-    {file = "librt-0.11.0-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:d5b0eea49f5562861ee8d757a32ef7d559c1d35be2aaaa1ec28941d74c9ffc8a"},
-    {file = "librt-0.11.0-cp312-cp312-manylinux_2_34_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:0d1029d7e1ae1a7e647ed6fb5df8c4ce2dffefb7a9f5fd1376a4554d96dac09f"},
-    {file = "librt-0.11.0-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:bc3ce6b33c5828d9e80592011a5c584cb2ce86edbc4088405f70da47dc1d1b3b"},
-    {file = "librt-0.11.0-cp312-cp312-musllinux_1_2_i686.whl", hash = "sha256:936c5995f3514a42111f20099397d8177c79b4d7e70961e396c6f5a0a3566766"},
-    {file = "librt-0.11.0-cp312-cp312-musllinux_1_2_riscv64.whl", hash = "sha256:9bc0ca6ad9381cbe8e4aa6e5726e4c80c78115a6e9723c599ed1d73e092bc49d"},
-    {file = "librt-0.11.0-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:070aa8c26c0a74774317a72df8851facc7f0f012a5b406557ac56992d92e1ec8"},
-    {file = "librt-0.11.0-cp312-cp312-win32.whl", hash = "sha256:6bf14feb84b05ae945277395451998c89c54d0def4070eb5c08de544930b245a"},
-    {file = "librt-0.11.0-cp312-cp312-win_amd64.whl", hash = "sha256:75672f0bc524ede266287d532d7923dbce94c7514ad07627bac3d0c6d92cc4d9"},
-    {file = "librt-0.11.0-cp312-cp312-win_arm64.whl", hash = "sha256:2f10cf143e4a9bb0f4f5af568a00df94a2d69ef41c2579584454bb0fe5cc642c"},
-    {file = "librt-0.11.0-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:78dc31f7fdfe9c9d0eb0e8f42d139db230e826415bbcabd9f0e9faaaee909894"},
-    {file = "librt-0.11.0-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:fa475675db22290c3158e1d42326d0f5a65f04f44a0e68c3630a25b53560fb9c"},
-    {file = "librt-0.11.0-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:621db29691044bdeda22e789e482e1b0f3a985d90e3426c9c6d17606416205ea"},
-    {file = "librt-0.11.0-cp313-cp313-manylinux2014_i686.manylinux_2_17_i686.manylinux_2_28_i686.whl", hash = "sha256:a9010e2ed5b3a9e158c5fd966b3ab7e834bb3d3aacc8f66c91dd4b57a3799230"},
-    {file = "librt-0.11.0-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:7c39513d8b7477a2e1ed8c43fc21c524e8d5a0f8d4e8b7b074dbdbe7820a08e2"},
-    {file = "librt-0.11.0-cp313-cp313-manylinux_2_34_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:7aef3cf1d5af86e770ab04bfd993dfc4ae8b8c17f66fb77dd4a7d50de7bbb1a3"},
-    {file = "librt-0.11.0-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:557183ddc36babe46b27dd60facbd5adb4492181a5be887587d57cda6e092f21"},
-    {file = "librt-0.11.0-cp313-cp313-musllinux_1_2_i686.whl", hash = "sha256:83d3e1f72bd42f6c5c0b7daec530c3f829bd02db42c70b8ddf0c2d90a2459930"},
-    {file = "librt-0.11.0-cp313-cp313-musllinux_1_2_riscv64.whl", hash = "sha256:4ce1f21fbe589bc1afd7872dece84fb0e1144f794a288e58a10d2c54a55c43be"},
-    {file = "librt-0.11.0-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:970b09f7044ea2b64c9da42fd3d335666518cfd1c6e8a182c95da73d0214b41e"},
-    {file = "librt-0.11.0-cp313-cp313-win32.whl", hash = "sha256:78fddc31cd4d3caa897ad5d31f856b1faadc9474021ad6cb182b9018793e254e"},
-    {file = "librt-0.11.0-cp313-cp313-win_amd64.whl", hash = "sha256:8ca8aa88751a775870b764e93bad5135385f563cb8dcee399abf034ea4d3cb47"},
-    {file = "librt-0.11.0-cp313-cp313-win_arm64.whl", hash = "sha256:96f044bb325fd9cf1a723015638c219e9143f0dfbc0ca54c565df2b7fc748b44"},
-    {file = "librt-0.11.0-cp314-cp314-macosx_10_13_x86_64.whl", hash = "sha256:4a017a95e5837dc15a8c5661d60e05daa96b90908b1aa6b7acdf443cd25c8ebd"},
-    {file = "librt-0.11.0-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:b1ecbd9819deccc39b7542bf4d2a740d8a620694d39989e58661d3763458f8d4"},
-    {file = "librt-0.11.0-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:7da327dacd7be8f8ec36547373550744a3cc0e536d54665cd83f8bcd961200e8"},
-    {file = "librt-0.11.0-cp314-cp314-manylinux2014_i686.manylinux_2_17_i686.manylinux_2_28_i686.whl", hash = "sha256:0dc56b1f8d06e60db362cc3fdae206681817f86ce4725d34511473487f12a34b"},
-    {file = "librt-0.11.0-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:05fb8fb2ab90e21c8d12ea240d744ad514da9baf381ebfa70d91d20d21713175"},
-    {file = "librt-0.11.0-cp314-cp314-manylinux_2_34_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:cae74872be221df4374d10fec61f93ed1513b9546ea84f2c0bf73ab3e9bd0b03"},
-    {file = "librt-0.11.0-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:32bcc918c0148eb7e3d57385125bac7e5f9e4359d05f07448b09f6f778c2f31c"},
-    {file = "librt-0.11.0-cp314-cp314-musllinux_1_2_i686.whl", hash = "sha256:f9743fc99135d5f78d2454435615f6dec0473ca507c26ce9d92b10b562a280d3"},
-    {file = "librt-0.11.0-cp314-cp314-musllinux_1_2_riscv64.whl", hash = "sha256:5ba067f4aadae8fda802d91d2124c90c42195ff32d9161d3549e6d05cfe26f96"},
-    {file = "librt-0.11.0-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:de3bf945454d032f9e390b85c4072e0a0570bf825421c8be0e71209fa65e1abe"},
-    {file = "librt-0.11.0-cp314-cp314-win32.whl", hash = "sha256:d2277a05f6dcb9fd13db9566aac4fabd68c3ea1ea46ee5567d4eef8efa495a2f"},
-    {file = "librt-0.11.0-cp314-cp314-win_amd64.whl", hash = "sha256:ab73e8db5e3f564d812c1f5c3a175930a5f9bc96ccb5e3b22a34d7858b401cf7"},
-    {file = "librt-0.11.0-cp314-cp314-win_arm64.whl", hash = "sha256:aea3caa317752e3a466fa8af45d91ee0ea8c7fdd96e42b0a8dd9b76a7931eba1"},
-    {file = "librt-0.11.0-cp314-cp314t-macosx_10_13_x86_64.whl", hash = "sha256:d1b36540d7aaf9b9101b3a6f376c8d8e9f7a9aec93ed05918f2c69d493ffef72"},
-    {file = "librt-0.11.0-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:efbb343ab2ce3540f4ecbe6315d677ed70f37cd9a72b1e58066c918ca83acbaa"},
-    {file = "librt-0.11.0-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:aa0dd688aab3f7914d3e6e5e3554978e0383312fb8e771d84be008a35b9ee548"},
-    {file = "librt-0.11.0-cp314-cp314t-manylinux2014_i686.manylinux_2_17_i686.manylinux_2_28_i686.whl", hash = "sha256:f5fb36b8c6c63fdcbb1d526d94c0d1331610d43f4118cc1beb4efef4f3faacb2"},
-    {file = "librt-0.11.0-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:4a9a237d13addb93715b6fee74023d5ee3469b53fce527626c0e088aa585805f"},
-    {file = "librt-0.11.0-cp314-cp314t-manylinux_2_34_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:5ddd17bd87b2c56ddd60e546a7984a2e64c4e8eab92fb4cf3830a48ad5469d51"},
-    {file = "librt-0.11.0-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:bd43992b4473d42f12ff9e68326079f0696d9d4e6000e8f39a0238d482ba6ee2"},
-    {file = "librt-0.11.0-cp314-cp314t-musllinux_1_2_i686.whl", hash = "sha256:f8e3e8056dd674e279741485e2e512d6e9a751c7455809d0114e6ebf8d781085"},
-    {file = "librt-0.11.0-cp314-cp314t-musllinux_1_2_riscv64.whl", hash = "sha256:c1f708d8ae9c56cf38a903c44297243d2ec83fd82b396b977e0144a3e76217e3"},
-    {file = "librt-0.11.0-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:0add982e0e7b9fc14cf4b33789d5f13f66581889b88c2f58099f6ce8f92617bd"},
-    {file = "librt-0.11.0-cp314-cp314t-win32.whl", hash = "sha256:2b481d846ac894c4e8403c5fd0e87c5d11d6499e404b474602508a224ff531c8"},
-    {file = "librt-0.11.0-cp314-cp314t-win_amd64.whl", hash = "sha256:28edb433edde181112a908c78907af28f964eabc15f4dd16c9d66c834302677c"},
-    {file = "librt-0.11.0-cp314-cp314t-win_arm64.whl", hash = "sha256:dee008f20b542e3cd162ba338a7f9ec0f6d23d395f66fe8aeeec3c9d067ea253"},
-    {file = "librt-0.11.0-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:6bd72d903911d995ab666dbd1871f8b1e80925a699af8063fbf50053329fb05f"},
-    {file = "librt-0.11.0-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:0ef69ac715f3cd8e5cd252cb2aebfa72c015492aacc339d5d7bf8fef3c62c677"},
-    {file = "librt-0.11.0-cp39-cp39-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:624a40c4a4ad7773315c287276cd024509b2c66ff5904f504bfc08d2c70293ab"},
-    {file = "librt-0.11.0-cp39-cp39-manylinux2014_i686.manylinux_2_17_i686.manylinux_2_28_i686.whl", hash = "sha256:41dc19fe150b69716c8ece4f76773a9e8813fe3e35e032a58b4d46423fb8d7c0"},
-    {file = "librt-0.11.0-cp39-cp39-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:4e8bd98ea9c47ae90b319a087ab28dac493f1ffbc1ecd1f28fcdbf3b7e1108d1"},
-    {file = "librt-0.11.0-cp39-cp39-manylinux_2_34_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:84308fc49423ce6475d1c5d1985cd69a8ca9f0325fc7d5f81bb690a3f3625d4e"},
-    {file = "librt-0.11.0-cp39-cp39-musllinux_1_2_aarch64.whl", hash = "sha256:ff0fbaf5f44a21beeb0110f2ab64f45135a9536a834b79c0d1ef018f2786bbfa"},
-    {file = "librt-0.11.0-cp39-cp39-musllinux_1_2_i686.whl", hash = "sha256:9c028a9442a18e266955d364ce42259136e79a7ba14d773e0d778d5f70cd56f1"},
-    {file = "librt-0.11.0-cp39-cp39-musllinux_1_2_riscv64.whl", hash = "sha256:9f1692105a02bcf853f355032a5fdc5494358ef83d8fd22d16de375c85cec3f5"},
-    {file = "librt-0.11.0-cp39-cp39-musllinux_1_2_x86_64.whl", hash = "sha256:7a80a71e1fda83cc752a9141e87aae7fef279538597564d670e9ce513f286192"},
-    {file = "librt-0.11.0-cp39-cp39-win32.whl", hash = "sha256:140695816ddf3c86eb972981a26f35efd871c44b0c3aed44c8cd01749386617f"},
-    {file = "librt-0.11.0-cp39-cp39-win_amd64.whl", hash = "sha256:92f7ff819c197fc30473190a12c2856f325ac90aabfccbeb2072d28cc2e234e3"},
-    {file = "librt-0.11.0.tar.gz", hash = "sha256:075dc3ef4458a278e0195cbf6ac9d38808d9b906c5a6c7f7f79c3888276a3fb1"},
+    {file = "librt-0.16.0-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:96f576f2711f8519152ec76d0e599243555c1f07679fa73606ca8c8c868c0be6"},
+    {file = "librt-0.16.0-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:2bec3818c7da7c96ceae0ef5915a3d16c52dd08f3ea913bf1fe8568c447c7978"},
+    {file = "librt-0.16.0-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:64c79520414a3fdfc6aabd7593e6169afa14d5f8d9908d4b498db068868b08dd"},
+    {file = "librt-0.16.0-cp310-cp310-manylinux2014_i686.manylinux_2_17_i686.manylinux_2_28_i686.whl", hash = "sha256:f81b5b19ce748ef68d4746656b7929762eb2fe99269b266e4be07e2ee4de7144"},
+    {file = "librt-0.16.0-cp310-cp310-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:c71d1b76210a36729fedfc5115069b50a3d8619054745f8758fe5d6f19e86671"},
+    {file = "librt-0.16.0-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:473eebc7866bb0a0c8849a292b5e7157c1aba5d14d0f0f610c52158d6d964262"},
+    {file = "librt-0.16.0-cp310-cp310-manylinux_2_34_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:4eb1313a19847089ee81e88742abedf285c60538816640b99742d8534b81d26a"},
+    {file = "librt-0.16.0-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:ebefd60b42e2a82b32d136bb5f7c94eadfcd29f772b547df6f3291d1ed855a1c"},
+    {file = "librt-0.16.0-cp310-cp310-musllinux_1_2_i686.whl", hash = "sha256:6c5da27e8056439f927ea896735da60e616c477a8293feaa3233d4e7781a6726"},
+    {file = "librt-0.16.0-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:67e718c7a43f8db325abbbf1404e2d535f12f8f7a1a82259568385cc5274b82a"},
+    {file = "librt-0.16.0-cp310-cp310-musllinux_1_2_riscv64.whl", hash = "sha256:314e703f0c19320dc8094e7a784b9cf29e1b67402515abf580069a6363c0b4f1"},
+    {file = "librt-0.16.0-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:943c6bbecbdf7fa575a4f2952fcfd848c88ef95507c3fca411e89d4ac3ff8143"},
+    {file = "librt-0.16.0-cp310-cp310-win32.whl", hash = "sha256:7cc365f006891afb006b52d5ee5ee74c09306ffa20e2f8705a32a4450af2f3ba"},
+    {file = "librt-0.16.0-cp310-cp310-win_amd64.whl", hash = "sha256:0314058469f4d2fd279ce7c62ac274ac82c3918ef7db62ef0697c4c359370155"},
+    {file = "librt-0.16.0-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:fe4372c52d4849096c6cc1cda2817d293ec51440c890474ed59ef38d46556f18"},
+    {file = "librt-0.16.0-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:c72c5295a84bd249526da9bdca38f2e176d15c31c13bb0063c5053f4ca023421"},
+    {file = "librt-0.16.0-cp311-cp311-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:be56ba9c884143495b517f23fe794ae367d58cd89ea0fdd6d437e3c024a87f9f"},
+    {file = "librt-0.16.0-cp311-cp311-manylinux2014_i686.manylinux_2_17_i686.manylinux_2_28_i686.whl", hash = "sha256:ef46c1a29ffb8c72e882e22618ec618778eacd0578fb22c6e7cf9c11d15f357b"},
+    {file = "librt-0.16.0-cp311-cp311-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:d3c94211ee0c4f8d649ec06b7c115c0ec4eadb873a0e3154ca15cef3f814b071"},
+    {file = "librt-0.16.0-cp311-cp311-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:94aed6a8308818b91677957d1bd03188869cd7aeb23c5dba7912a6c0402f7602"},
+    {file = "librt-0.16.0-cp311-cp311-manylinux_2_34_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:349c0bcb87ebd07481b6ff781e25cdc699723dbe2212e57dabb27f7a13b7b87d"},
+    {file = "librt-0.16.0-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:001bfd59a7d45b17e3e75f2a8c6405280b35e7b84471792778e718c4f368950e"},
+    {file = "librt-0.16.0-cp311-cp311-musllinux_1_2_i686.whl", hash = "sha256:28e038895b998d7a0c7798922ce8a1dc157675df5cf1c9ef0aca809ed804b7a1"},
+    {file = "librt-0.16.0-cp311-cp311-musllinux_1_2_ppc64le.whl", hash = "sha256:0dbe4096a7ecc00fa835d24510ad8545a4efef738dac96e0e63516783ccde905"},
+    {file = "librt-0.16.0-cp311-cp311-musllinux_1_2_riscv64.whl", hash = "sha256:5cd5b092441053364af968ea12084692cb9d4a22f3ce9524e377880bf028761e"},
+    {file = "librt-0.16.0-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:3ddeb3c9dedb461bb457c6c7d9aa7fbf35329da313d1a7543d00c8d0f3473c96"},
+    {file = "librt-0.16.0-cp311-cp311-win32.whl", hash = "sha256:e05108e0849966f53a8d2d3112a7af881d0efaa479bc735bba91108f9f2350a7"},
+    {file = "librt-0.16.0-cp311-cp311-win_amd64.whl", hash = "sha256:5f49cff01bd608ef7d97104cb035c75455e79c2d70bf4a506cf773338ac1860d"},
+    {file = "librt-0.16.0-cp311-cp311-win_arm64.whl", hash = "sha256:1d28ae980ae2218f9c5b95d191e947296f918c9bf0b400d467a9430275bbe678"},
+    {file = "librt-0.16.0-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:fe52bf4641069e7978a14253b036cb9002def1926317e710f2e249f8a8c47742"},
+    {file = "librt-0.16.0-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:5bcc2c4726ced915b00de0c9856a4eeabfb3fddb93e10e0b8f735b7709358b6d"},
+    {file = "librt-0.16.0-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:ff7baa55f8e7c69851419e50a666015d02a74198716fd45c0125a2112e0a389f"},
+    {file = "librt-0.16.0-cp312-cp312-manylinux2014_i686.manylinux_2_17_i686.manylinux_2_28_i686.whl", hash = "sha256:b95d5d92ab83d39e760a52091bb1baba664f3a2351e39b1e16801e5747c2f0e9"},
+    {file = "librt-0.16.0-cp312-cp312-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:b6d085d70bce51d43c5c7c36d63490770180d8779e71c49305c87b4213918de7"},
+    {file = "librt-0.16.0-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:36e53948e99bbe3ffea257124cfcae1cfb01831555c9a9c903c9f9a72db7fd07"},
+    {file = "librt-0.16.0-cp312-cp312-manylinux_2_34_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:54d11f726aae9df5a6ffbbf0a03a52449bbac84a53ef03669cb41cdfd4ae41bf"},
+    {file = "librt-0.16.0-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:4323193ac0cd025f85af531df8ba91bf24d1973b401697347a6282e8fd3fcf5e"},
+    {file = "librt-0.16.0-cp312-cp312-musllinux_1_2_i686.whl", hash = "sha256:e42f8e098b9c5396fefa05fb1cc7e33b0e08fc51da106b5de4a45fd22aac6743"},
+    {file = "librt-0.16.0-cp312-cp312-musllinux_1_2_ppc64le.whl", hash = "sha256:39ec1d5a14e37baf1450a6cabf03fe552340808bf1ad9d71824ab90117716459"},
+    {file = "librt-0.16.0-cp312-cp312-musllinux_1_2_riscv64.whl", hash = "sha256:d1aabe3925cbb4a08d15b7b20ba4011b53019da0c4173a25155139b7b1baed65"},
+    {file = "librt-0.16.0-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:300c3ffdc459f4a779a8411ecb188e3ac0b1ff3a3a7b099642555dedae06c69b"},
+    {file = "librt-0.16.0-cp312-cp312-win32.whl", hash = "sha256:c17194318e4c0c0348b36f36c2ec7534436fe0a4c15582403162a4f08c80797a"},
+    {file = "librt-0.16.0-cp312-cp312-win_amd64.whl", hash = "sha256:25a58a19ea8d83b68209f04912df765e9260635ef77646542ed4b4abe6bc7940"},
+    {file = "librt-0.16.0-cp312-cp312-win_arm64.whl", hash = "sha256:f7be7cf555bc30ec12622e9447299cc4a9b8ff307548b634794353db0c2065dc"},
+    {file = "librt-0.16.0-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:c5e6144e68b577f157519f2ba88ca20e3ed61c29b00e5cdfa76cd2d45acf059a"},
+    {file = "librt-0.16.0-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:33f41443a1f4e1f099331b3d8120e409fbff84b9760bc1cc9ea496f37ddaa5cc"},
+    {file = "librt-0.16.0-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:7e510b7770bee609617a3374a96548eb114cae048023e3f049ee449e7ff2db32"},
+    {file = "librt-0.16.0-cp313-cp313-manylinux2014_i686.manylinux_2_17_i686.manylinux_2_28_i686.whl", hash = "sha256:efc49c462d4516b8a58b00b490078fa64689fd1fe66970cc190131d7afb8027e"},
+    {file = "librt-0.16.0-cp313-cp313-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:92caf82ebef5e12d21c72242b70d1e92536f1711cf2a727a4c276de4b4469087"},
+    {file = "librt-0.16.0-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:17bac7f7a16b328fff77e440287693eb017abde913595b5827ebccbc21ecd8a6"},
+    {file = "librt-0.16.0-cp313-cp313-manylinux_2_34_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:5b976054553670829985ed767feb78fb6bcede0175327c4844dd5c281c1be659"},
+    {file = "librt-0.16.0-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:0058f9d68721094105917254c72ac0569117bb7b13b9769cf45d26d89f9d21cd"},
+    {file = "librt-0.16.0-cp313-cp313-musllinux_1_2_i686.whl", hash = "sha256:30b7beaf3f4487b7d8adef1f158b49067cb4d5a19fa7a3bf31a4e7a820e435c5"},
+    {file = "librt-0.16.0-cp313-cp313-musllinux_1_2_ppc64le.whl", hash = "sha256:468df902df016a06eb0e40b0747dc8d14e47d7a38b18b63b1fb167d85cb94d63"},
+    {file = "librt-0.16.0-cp313-cp313-musllinux_1_2_riscv64.whl", hash = "sha256:aea7b1f2b125dad5de85f049136651bff256c883c65e6b9209b2da0a1ac3cdef"},
+    {file = "librt-0.16.0-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:a8afb6557920860b7a3a596eb804cf37e09e7cf8a803db2478c202acc72d8c2e"},
+    {file = "librt-0.16.0-cp313-cp313-pyemscripten_2025_0_wasm32.whl", hash = "sha256:77c7a2b4fe2c1369e0d5aa1cade26740a7b14be32fbc9a5535d617d20065c39d"},
+    {file = "librt-0.16.0-cp313-cp313-win32.whl", hash = "sha256:02d89c813d5ff74b17df72d3a34819d132cd168e56b81bf755b809bd9e46b8c4"},
+    {file = "librt-0.16.0-cp313-cp313-win_amd64.whl", hash = "sha256:14ed6ebe3e4f85f326d7920011ad30ff49ed9334e62cf88caef9ba973d9e3a92"},
+    {file = "librt-0.16.0-cp313-cp313-win_arm64.whl", hash = "sha256:83d4041a3d9b2fd053a8a4e1f22878b3e5833e2712956382d5c048d791454e91"},
+    {file = "librt-0.16.0-cp314-cp314-macosx_10_15_x86_64.whl", hash = "sha256:931a0bb0fcac88f263e269e46eb30ba8e21402cd3c62ca40cb97034c0693fab1"},
+    {file = "librt-0.16.0-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:1bc17e54e5305f8d40b7ca203671ff5a9e59c1d0f8ea0f625dcca53a3984de11"},
+    {file = "librt-0.16.0-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:877698bf6bca5721d8be345f2fe09778e40ecadea8b58c73075f2b1a53666bf2"},
+    {file = "librt-0.16.0-cp314-cp314-manylinux2014_i686.manylinux_2_17_i686.manylinux_2_28_i686.whl", hash = "sha256:5981c011b306781ce561e18e14230a14524a3d8109b97553666c942c18f31a96"},
+    {file = "librt-0.16.0-cp314-cp314-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:afced3dfc17cd805ecf7a3d77996a71cf5f2c75aa66eb0c21a9930f4fc992f86"},
+    {file = "librt-0.16.0-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:ca8052401c55d7511dda6760719fda7618067e83535d7d0010096d216c34b667"},
+    {file = "librt-0.16.0-cp314-cp314-manylinux_2_34_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:1e511762a074005bb0aa569166779834e75e438370226930d0ce1866d4b6a33b"},
+    {file = "librt-0.16.0-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:f1e8591bd8a5a628cd7f07954c6a1592359a878bf032957a8e9057a41d644311"},
+    {file = "librt-0.16.0-cp314-cp314-musllinux_1_2_i686.whl", hash = "sha256:a4aaefb4ba6c07e1aeebb2795c8958148f1d6f9af3b555b53d23d766edb6d67a"},
+    {file = "librt-0.16.0-cp314-cp314-musllinux_1_2_ppc64le.whl", hash = "sha256:d92db7a0f6aee44f1baee94750457e8d2d1c6ccea41842de6268d34e8dc7eddd"},
+    {file = "librt-0.16.0-cp314-cp314-musllinux_1_2_riscv64.whl", hash = "sha256:378dfaffb38e59c24a87cde5713cd865d51ff7383fa12947f3907f306ea1ca55"},
+    {file = "librt-0.16.0-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:3e0c39bdc85370422e8b637be76eb1fd07d30967551b03e62267dd156f553152"},
+    {file = "librt-0.16.0-cp314-cp314-pyemscripten_2026_0_wasm32.whl", hash = "sha256:1b384b90ab79a7bc30b566895809a636e0666f21f3cf12b54823d025b7e83839"},
+    {file = "librt-0.16.0-cp314-cp314-win32.whl", hash = "sha256:52327da75a94012e7f932f913d20d3876bed3c102be00e6c3e8600ff7bdd58a7"},
+    {file = "librt-0.16.0-cp314-cp314-win_amd64.whl", hash = "sha256:3f0b8114c44b2ac06ff5dacd08e07e8e807ff4f46083f2a1602685122559be41"},
+    {file = "librt-0.16.0-cp314-cp314-win_arm64.whl", hash = "sha256:8caf96a4ef8fb27d0ac0d1ad8337d26a240acd4a02fe4345d0a8f264753e8f99"},
+    {file = "librt-0.16.0-cp314-cp314t-macosx_10_15_x86_64.whl", hash = "sha256:953107e2f68d0f3512c48f898b0dbf0ce5cc52bba0f318d847c985dc555ee4cc"},
+    {file = "librt-0.16.0-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:ad37d5b9abd49c9a655dcda7ea52a8a752884062ef1ee71ae17c2f2a0f81fe6a"},
+    {file = "librt-0.16.0-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:2c4aa329c17bd1aaea4f6e89335d8ccd494b3a5830b6654462273e50e11023f0"},
+    {file = "librt-0.16.0-cp314-cp314t-manylinux2014_i686.manylinux_2_17_i686.manylinux_2_28_i686.whl", hash = "sha256:0ead24d2562a49473dddd9efef8581f020007eb0054389c3ee3ffad38b1ca4c9"},
+    {file = "librt-0.16.0-cp314-cp314t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:02118f56a9c36ddd07dfd9b919d9ecc117ba20a90987d56aa4c429fa34509188"},
+    {file = "librt-0.16.0-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:4e29522c62e28595ff7e324c6834ade51127707f0e255b18d1c1cf03d39c1048"},
+    {file = "librt-0.16.0-cp314-cp314t-manylinux_2_34_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:3ff4b2367926b69c6215635902cccb04048e73094e9862900d27cb2c6bbff143"},
+    {file = "librt-0.16.0-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:c6f1b27bf1632a7e016af9f145f82be95e1edd7721a646505c21059257cb5a04"},
+    {file = "librt-0.16.0-cp314-cp314t-musllinux_1_2_i686.whl", hash = "sha256:5696d7f52e7b37217cb3a8f92c744fe835942602fdd4c1a8bc4741d3bfdce15e"},
+    {file = "librt-0.16.0-cp314-cp314t-musllinux_1_2_ppc64le.whl", hash = "sha256:6072e92dd876ff6ceeb6cf371e35e51f479349837391341f479b08df4564242b"},
+    {file = "librt-0.16.0-cp314-cp314t-musllinux_1_2_riscv64.whl", hash = "sha256:39ca4f2f2fe05de8e63493da592d84311adabe5bef52b193851981da9816b302"},
+    {file = "librt-0.16.0-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:f9807485a908f00355820f18e91e045ffdcdc5adb68aaec40a1e2b88c5f7bba1"},
+    {file = "librt-0.16.0-cp314-cp314t-win32.whl", hash = "sha256:94be5cb7bca4df6201f4183e9e4fa2086c655283d20b38cd84500a69057575a7"},
+    {file = "librt-0.16.0-cp314-cp314t-win_amd64.whl", hash = "sha256:d46ca272b251d033dd4527b0dec5f261a28a52bd5fa0f99c117b0a1f8588cc2d"},
+    {file = "librt-0.16.0-cp314-cp314t-win_arm64.whl", hash = "sha256:b9d6d4b14e92d876f8026b54c20c445f36425214c1081dc76f74e40db386b82b"},
+    {file = "librt-0.16.0-cp315-cp315-macosx_10_15_x86_64.whl", hash = "sha256:6fe436af2eaf630474f491af5d032cbe45f93fcff5c3b9fe4ab194a7255b20ff"},
+    {file = "librt-0.16.0-cp315-cp315-macosx_11_0_arm64.whl", hash = "sha256:8ff5d26c529336be9bd7ae04483235d77778ee7d6444a95353102b542601ce81"},
+    {file = "librt-0.16.0-cp315-cp315-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:909d8e3c1faee44cb762b1c519ff8613dcc5ceae5c99987a00917b5a31fd1d6a"},
+    {file = "librt-0.16.0-cp315-cp315-manylinux2014_i686.manylinux_2_17_i686.manylinux_2_28_i686.whl", hash = "sha256:6d4a64283ee61824b5790de882bc68e2d9d7a5143537cb7a966f7354f71646d4"},
+    {file = "librt-0.16.0-cp315-cp315-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:5810ba811297fdf37a1531a57667cb8ace0842013ca8606bf9eb7c24cf4be154"},
+    {file = "librt-0.16.0-cp315-cp315-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:e56aaf8c167548dc8e5d6f3bd0f48dcdd299a23c73be3f744aab79d99e9c7f5d"},
+    {file = "librt-0.16.0-cp315-cp315-manylinux_2_34_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:8f36c58e33b304b525c6c9c5076399c6ebf1109e17b9051a05a407b091b9215b"},
+    {file = "librt-0.16.0-cp315-cp315-musllinux_1_2_aarch64.whl", hash = "sha256:242e00b3d4fa37c3d3c1ca5f5c9adb7d909ddb1eac9c41f2787320d00caa0af2"},
+    {file = "librt-0.16.0-cp315-cp315-musllinux_1_2_i686.whl", hash = "sha256:0253721561787b8df8443eb347b7a6461015354e5bdd37ee38a41fef220d2bb0"},
+    {file = "librt-0.16.0-cp315-cp315-musllinux_1_2_ppc64le.whl", hash = "sha256:3e483a8d69ede8067db70c0e83007423b6925de6fd53afed01d66160f2e9398c"},
+    {file = "librt-0.16.0-cp315-cp315-musllinux_1_2_riscv64.whl", hash = "sha256:69ba927445cfaaffb4081003ef5224c55a5c2ab67ef956f416ef744916e44121"},
+    {file = "librt-0.16.0-cp315-cp315-musllinux_1_2_x86_64.whl", hash = "sha256:d6a365f2ab45a984d0e00eee0dd17f599ceab8cadab6ea07b6111c8132fc0e42"},
+    {file = "librt-0.16.0-cp315-cp315-pyemscripten_2026_5_wasm32.whl", hash = "sha256:f01f3805f2dae4781c0c34b440e31740d082950bdaf89a6f601ad589a28af57a"},
+    {file = "librt-0.16.0-cp315-cp315-win32.whl", hash = "sha256:b0e3e721c75d2e79a76d4422c79d7ba705fe1bbafec907037fe7a657a480a0e3"},
+    {file = "librt-0.16.0-cp315-cp315-win_amd64.whl", hash = "sha256:bc02954b1295de798bbdb0b4e2d8a28c2117de8b5c73dcbeb27dc32572dfb971"},
+    {file = "librt-0.16.0-cp315-cp315-win_arm64.whl", hash = "sha256:c5db585d43449a5f54303d4b2774e45e1babd975cfe1630a3d708c0b80c3e560"},
+    {file = "librt-0.16.0-cp315-cp315t-macosx_10_15_x86_64.whl", hash = "sha256:f06c689cb14afd9b612727553a5ec5a40febf113ca41c4413a2b0b334285884b"},
+    {file = "librt-0.16.0-cp315-cp315t-macosx_11_0_arm64.whl", hash = "sha256:13b4e8aba90b0b1c82474e9844aa9ffe7ad3faa484350e1da64cb8188d903134"},
+    {file = "librt-0.16.0-cp315-cp315t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:5a269c46ae327d8e6f8c1f85f7516cb52c0fa48127565a1105a4f4a05ff2a0b4"},
+    {file = "librt-0.16.0-cp315-cp315t-manylinux2014_i686.manylinux_2_17_i686.manylinux_2_28_i686.whl", hash = "sha256:a33e0dae1f8592146a4764d54ce842b278732d21a84e17c3bbe6b1bc158a2248"},
+    {file = "librt-0.16.0-cp315-cp315t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:47ada6ea32636492c61aa8ad27ae3b9404bfe7a97e3ba946d1984236cc741da0"},
+    {file = "librt-0.16.0-cp315-cp315t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:c43bd6e642d8a248c114327f98dd25ac5a7cb5aa168ef02f0559b91874df16b8"},
+    {file = "librt-0.16.0-cp315-cp315t-manylinux_2_34_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:c3d1bb7841a816ace6449bb26d3f9560dbfa20e71c568d23f0f62bf1e68f50b1"},
+    {file = "librt-0.16.0-cp315-cp315t-musllinux_1_2_aarch64.whl", hash = "sha256:3931f7a3db322e7f44e02a280e3949326ce9579ad388ee8d691dc7c76da9fb70"},
+    {file = "librt-0.16.0-cp315-cp315t-musllinux_1_2_i686.whl", hash = "sha256:f4462528b6000afe8f16907b5c7c2553abf1df005ba5140e6eb394541c3624c3"},
+    {file = "librt-0.16.0-cp315-cp315t-musllinux_1_2_ppc64le.whl", hash = "sha256:80039ba9b6a7d5f1a0175a4cca6bbefead87bd854c80abad1cb30afe47a830db"},
+    {file = "librt-0.16.0-cp315-cp315t-musllinux_1_2_riscv64.whl", hash = "sha256:7a1d272724b581bb6bc769dfdafed6da2ecc9886ba2450311de55a4ac2e1e9cd"},
+    {file = "librt-0.16.0-cp315-cp315t-musllinux_1_2_x86_64.whl", hash = "sha256:fbe4fb8c5445f7496d7f7f6bb0807875d09d47e6771ffa175fb2df2895fb86ba"},
+    {file = "librt-0.16.0-cp315-cp315t-win32.whl", hash = "sha256:375bfe6b572a8f6cfc398709356046173bf27e64c4c5edaf5f7062f051fb4bf9"},
+    {file = "librt-0.16.0-cp315-cp315t-win_amd64.whl", hash = "sha256:bd3150023d3dc2bc70f3784e59ffa1140d56ddba3d8125b3d6f9f85221279bfc"},
+    {file = "librt-0.16.0-cp315-cp315t-win_arm64.whl", hash = "sha256:8ceafb70f2a4f0826f11031942e59c0728fd98da112dc346d4352bde1e486866"},
+    {file = "librt-0.16.0-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:e1967e36ac4cae0c7e9615ad32e1a513cdacff79f9e8afb28bedc91caf48b4f3"},
+    {file = "librt-0.16.0-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:20fe0bf9053885e21c62b3e091fb5e73e1c54d1daf2e70eb388d70763bcd4220"},
+    {file = "librt-0.16.0-cp39-cp39-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:6a63610fa76524edfa605b5b259a603915c7a6e10e54f003e5503030506e81de"},
+    {file = "librt-0.16.0-cp39-cp39-manylinux2014_i686.manylinux_2_17_i686.manylinux_2_28_i686.whl", hash = "sha256:845a511b60ca43b9880dcc84a9784c891d6a2098c829130b320846c69c9c0c68"},
+    {file = "librt-0.16.0-cp39-cp39-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:d608f0bf3b8cbddd0067fe02cb8cab7d13e8eb9386b23a1843d4363044fd9e22"},
+    {file = "librt-0.16.0-cp39-cp39-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:df183721229ae51eef90108c115b43e98cb169b6155d34480338e5fc6616df00"},
+    {file = "librt-0.16.0-cp39-cp39-manylinux_2_34_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:4b6183e2e2e0ee00aac2ec07c7f7d151c97e85666b304f574b71cff0f9fccc4e"},
+    {file = "librt-0.16.0-cp39-cp39-musllinux_1_2_aarch64.whl", hash = "sha256:273d00be33792a15189331df10f1f1331621b043881e66c6c4377f7f776e1291"},
+    {file = "librt-0.16.0-cp39-cp39-musllinux_1_2_i686.whl", hash = "sha256:aa9357a1b4d4fc787bb718a59cb1112c28c8a976d6bfa268b71cc0a4ab8f3a94"},
+    {file = "librt-0.16.0-cp39-cp39-musllinux_1_2_ppc64le.whl", hash = "sha256:e9ce0bc440e7fd09b5f51f372f0f6640658f854b8fb05260f819cc669c93c42d"},
+    {file = "librt-0.16.0-cp39-cp39-musllinux_1_2_riscv64.whl", hash = "sha256:6c8893eae2fd13c5488d94056f3e6e5cf3142bfb1c4acaf136cb33d760c5964b"},
+    {file = "librt-0.16.0-cp39-cp39-musllinux_1_2_x86_64.whl", hash = "sha256:7393c9a48dcce4817dbd4b0d8ff6237efe9b0a0609f5b0adaef315f8541726b5"},
+    {file = "librt-0.16.0-cp39-cp39-win32.whl", hash = "sha256:71b93b42784e25b975079573c642a8fedb049a7bb31d70a51721b1666b3b2ced"},
+    {file = "librt-0.16.0-cp39-cp39-win_amd64.whl", hash = "sha256:5750a105b42a416f930edc59054927a406effb2550cd5bab92ad7a5842ed5d05"},
+    {file = "librt-0.16.0.tar.gz", hash = "sha256:ac38d6d8d66bf3d744148dbbc0b8e193e195a51e364ed55e224631f5721891fc"},
 ]
 
 [[package]]
 name = "mypy"
-version = "1.20.2"
+version = "2.4.0"
 description = "Optional static typing for Python"
 optional = false
 python-versions = ">=3.10"
 groups = ["dev"]
 files = [
-    {file = "mypy-1.20.2-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:cf5a4db6dca263010e2c7bff081c89383c72d187ba2cf4c44759aac970e2f0c4"},
-    {file = "mypy-1.20.2-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:7b0e817b518bff7facd7f85ea05b643ad8bdcce684cf29784987b0a7c8e1f997"},
-    {file = "mypy-1.20.2-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:97d7b9a485b40f8ca425460e89bf1da2814625b2da627c0dcc6aa46c92631d14"},
-    {file = "mypy-1.20.2-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:1e1c12f6d2db3d78b909b5f77513c11eb7f2dd2782b96a3ab6dffc7d44575c99"},
-    {file = "mypy-1.20.2-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:89dce27e142d25ffbc154c1819383b69f2e9234dc4ed4766f42e0e8cb264ab5c"},
-    {file = "mypy-1.20.2-cp310-cp310-win_amd64.whl", hash = "sha256:f376e37f9bf2a946872fc5fd1199c99310748e3c26c7a26683f13f8bdb756cbd"},
-    {file = "mypy-1.20.2-cp310-cp310-win_arm64.whl", hash = "sha256:6e2b469efd811707bc530fd1effef0f5d6eebcb7fe376affae69025da4b979a2"},
-    {file = "mypy-1.20.2-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:4077797a273e56e8843d001e9dfe4ba10e33323d6ade647ff260e5cd97d9758c"},
-    {file = "mypy-1.20.2-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:cdecf62abcc4292500d7858aeae87a1f8f1150f4c4dd08fb0b336ee79b2a6df3"},
-    {file = "mypy-1.20.2-cp311-cp311-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:c566c3a88b6ece59b3d70f65bedef17304f48eb52ff040a6a18214e1917b3254"},
-    {file = "mypy-1.20.2-cp311-cp311-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:0deb80d062b2479f2c87ae568f89845afc71d11bc41b04179e58165fd9f31e98"},
-    {file = "mypy-1.20.2-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:bba9ad231e92a3e424b3e56b65aa17704993425bba97e302c832f9466bb85bac"},
-    {file = "mypy-1.20.2-cp311-cp311-win_amd64.whl", hash = "sha256:baf593f2765fa3a6b1ef95807dbaa3d25b594f6a52adcc506a6b9cb115e1be67"},
-    {file = "mypy-1.20.2-cp311-cp311-win_arm64.whl", hash = "sha256:20175a1c0f49863946ec20b7f63255768058ac4f07d2b9ded6a6b46cfb5a9100"},
-    {file = "mypy-1.20.2-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:4dbfcf869f6b0517f70cf0030ba6ea1d6645e132337a7d5204a18d8d5636c02b"},
-    {file = "mypy-1.20.2-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:4b6481b228d072315b053210b01ac320e1be243dc17f9e5887ef167f23f5fae4"},
-    {file = "mypy-1.20.2-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:34397cdced6b90b836e38182076049fdb41424322e0b0728c946b0939ebdf9f6"},
-    {file = "mypy-1.20.2-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:a5da6976f20cae27059ea8d0c86e7cef3de720e04c4bb9ee18e3690fdb792066"},
-    {file = "mypy-1.20.2-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:56908d7e08318d39f85b1f0c6cfd47b0cac1a130da677630dac0de3e0623e102"},
-    {file = "mypy-1.20.2-cp312-cp312-win_amd64.whl", hash = "sha256:d52ad8d78522da1d308789df651ee5379088e77c76cb1994858d40a426b343b9"},
-    {file = "mypy-1.20.2-cp312-cp312-win_arm64.whl", hash = "sha256:785b08db19c9f214dc37d65f7c165d19a30fcecb48abfa30f31b01b5acaabb58"},
-    {file = "mypy-1.20.2-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:edfbfca868cdd6bd8d974a60f8a3682f5565d3f5c99b327640cedd24c4264026"},
-    {file = "mypy-1.20.2-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:e2877a02380adfcdbc69071a0f74d6e9dbbf593c0dc9d174e1f223ffd5281943"},
-    {file = "mypy-1.20.2-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:7488448de6007cd5177c6cea0517ac33b4c0f5ee9b5e9f2be51ce75511a85517"},
-    {file = "mypy-1.20.2-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:bb9c2fa06887e21d6a3a868762acb82aec34e2c6fd0174064f27c93ede68ad15"},
-    {file = "mypy-1.20.2-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:9d56a78b646f2e3daa865bc70cd5ec5a46c50045801ca8ff17a0c43abc97e3ee"},
-    {file = "mypy-1.20.2-cp313-cp313-win_amd64.whl", hash = "sha256:2a4102b03bb7481d9a91a6da8d174740c9c8c4401024684b9ca3b7cc5e49852f"},
-    {file = "mypy-1.20.2-cp313-cp313-win_arm64.whl", hash = "sha256:a95a9248b0c6fd933a442c03c3b113c3b61320086b88e2c444676d3fd1ca3330"},
-    {file = "mypy-1.20.2-cp314-cp314-macosx_10_15_x86_64.whl", hash = "sha256:419413398fe250aae057fd2fe50166b61077083c9b82754c341cf4fd73038f30"},
-    {file = "mypy-1.20.2-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:e73c07f23009962885c197ccb9b41356a30cc0e5a1d0c2ea8fd8fb1362d7f924"},
-    {file = "mypy-1.20.2-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:0c64e5973df366b747646fc98da921f9d6eba9716d57d1db94a83c026a08e0fb"},
-    {file = "mypy-1.20.2-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:5a65aa591af023864fd08a97da9974e919452cfe19cb146c8a5dc692626445dc"},
-    {file = "mypy-1.20.2-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:4fef51b01e638974a6e69885687e9bd40c8d1e09a6cd291cca0619625cf1f558"},
-    {file = "mypy-1.20.2-cp314-cp314-win_amd64.whl", hash = "sha256:913485a03f1bcf5d279409a9d2b9ed565c151f61c09f29991e5faa14033da4c8"},
-    {file = "mypy-1.20.2-cp314-cp314-win_arm64.whl", hash = "sha256:c3bae4f855d965b5453784300c12ffc63a548304ac7f99e55d4dc7c898673aa3"},
-    {file = "mypy-1.20.2-cp314-cp314t-macosx_10_15_x86_64.whl", hash = "sha256:2de3dcea53babc1c3237a19002bc3d228ce1833278f093b8d619e06e7cc79609"},
-    {file = "mypy-1.20.2-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:52b176444e2e5054dfcbcb8c75b0b719865c96247b37407184bbfca5c353f2c2"},
-    {file = "mypy-1.20.2-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:688c3312e5dadb573a2c69c82af3a298d43ecf9e6d264e0f95df960b5f6ac19c"},
-    {file = "mypy-1.20.2-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:29752dbbf8cc53f89f6ac096d363314333045c257c9c75cbd189ca2de0455744"},
-    {file = "mypy-1.20.2-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:803203d2b6ea644982c644895c2f78b28d0e208bba7b27d9b921e0ec5eb207c6"},
-    {file = "mypy-1.20.2-cp314-cp314t-win_amd64.whl", hash = "sha256:9bcb8aa397ff0093c824182fd76a935a9ba7ad097fcbef80ae89bf6c1731d8ec"},
-    {file = "mypy-1.20.2-cp314-cp314t-win_arm64.whl", hash = "sha256:e061b58443f1736f8a37c48978d7ab581636d6ab03e3d4f99e3fa90463bb9382"},
-    {file = "mypy-1.20.2-py3-none-any.whl", hash = "sha256:a94c5a76ab46c5e6257c7972b6c8cff0574201ca7dc05647e33e795d78680563"},
-    {file = "mypy-1.20.2.tar.gz", hash = "sha256:e8222c26daaafd9e8626dec58ae36029f82585890589576f769a650dd20fd665"},
+    {file = "mypy-2.4.0-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:ffda5244fd1ad71a1e54405e35f50d09b80c3978efa120f58bd1252ae32c62d2"},
+    {file = "mypy-2.4.0-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:615b03922d40e186fd1df73156473db0ba525c639bb4fd1cf28e1694879c9b23"},
+    {file = "mypy-2.4.0-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:13fa24f439c0e48a290a3922fa14ccd22f2762bae99d2142931b3e40a9055080"},
+    {file = "mypy-2.4.0-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:3adef556a19eb3b630bf86a79c29d7da3d61e541472d0d897ca01b171c0abf8f"},
+    {file = "mypy-2.4.0-cp310-cp310-win_amd64.whl", hash = "sha256:1dc0f64b0a92ae27a49d2175f0bacfa56e15bdc5b420cf92e0c1f79292219cbb"},
+    {file = "mypy-2.4.0-cp310-cp310-win_arm64.whl", hash = "sha256:4a378fc15fb33e321f04652c166ce73eeb8833a97c3d218132844e938cd93220"},
+    {file = "mypy-2.4.0-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:5d20e6c7c35fcbf2a0ebdd0eaeacfbc243009dfd33ab7822d54e213912e6dbbd"},
+    {file = "mypy-2.4.0-cp311-cp311-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:3011537be6cf1de4511c0255a324362a812b58184bbe61e15f59c8b31033bd74"},
+    {file = "mypy-2.4.0-cp311-cp311-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:a96b07a49b7b1d025ce59c1b3acbcf24bead9a83da4523c4a6bde1bb94e7a0e1"},
+    {file = "mypy-2.4.0-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:7c4f8f8d1d1c0e2832d8ee7113dd08f6df6c7aad9e863fcbed9f25832be0b8c4"},
+    {file = "mypy-2.4.0-cp311-cp311-win_amd64.whl", hash = "sha256:ba05652540bf12828e52abae807b024b09ca144ff4f75e2450a81d69c376425b"},
+    {file = "mypy-2.4.0-cp311-cp311-win_arm64.whl", hash = "sha256:6306086b87cf7f8a29aa618d9fd9bffb56c59247166b9660fdb54d86d7714ecd"},
+    {file = "mypy-2.4.0-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:e05ff2925d8b37ad26c80c1b9dc43ae5d455da2df1e23c24c095a6425917c57e"},
+    {file = "mypy-2.4.0-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:29243242cf72582b65f9582ad9e56e8cb281566ed3519f4cd70bb8b9f2977e90"},
+    {file = "mypy-2.4.0-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:29eb0b9427a6b11b992e452f6cceb8af724f4dceb47e779d0b35e405e996ea5e"},
+    {file = "mypy-2.4.0-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:e3ebe2f72a2a1156065a9851570ffbf50c0a93cdccadef9c6e05c508a4fd10b1"},
+    {file = "mypy-2.4.0-cp312-cp312-win_amd64.whl", hash = "sha256:236e0d68f6941992b0811128e652590f590db444ab29ad8f1324765b9298b946"},
+    {file = "mypy-2.4.0-cp312-cp312-win_arm64.whl", hash = "sha256:82d0f94c8587ccb472622ee7795280aaa38a06640d5f45b3f16909d6dd86a989"},
+    {file = "mypy-2.4.0-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:7da85fbcff6dac1abcc636707bed38b45598131fb7a605d9719c70b5cc733af8"},
+    {file = "mypy-2.4.0-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:4209da39d85cf240f762af622d8180fcdfcb4727d021f44ade62d613a1a43324"},
+    {file = "mypy-2.4.0-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:6be721bd4bd57576193653b75b4af3461c9d0bf7dd8b528f782e9be210dc75bb"},
+    {file = "mypy-2.4.0-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:e1fde197ae65be856a034a91b70ed747a16562ca69577785f06c661548424bf1"},
+    {file = "mypy-2.4.0-cp313-cp313-win_amd64.whl", hash = "sha256:295ecf2e57542cd836ca537486951289678c8c7d1ee6ad74ebe29b2168a003cf"},
+    {file = "mypy-2.4.0-cp313-cp313-win_arm64.whl", hash = "sha256:bc378bdad4e9f12b5bd96466083d1e71acf00594ec9c7b2bdb5e02816f77f303"},
+    {file = "mypy-2.4.0-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:058165f564ccf559c68c70fec2091fca5891110480210c22594635e3f6683437"},
+    {file = "mypy-2.4.0-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:9fa247e02b505a45a2775f69df38d360d197e3790bc60f717595db9eda358b6e"},
+    {file = "mypy-2.4.0-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:20e9a5cd875837520c43db98dea0b6d0c2197833d95c30127d8f570fb9b1f00b"},
+    {file = "mypy-2.4.0-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:9f03a7828cca2b0adcd6662aee8f2711ff8830e1027641fdea3ab0b787483566"},
+    {file = "mypy-2.4.0-cp314-cp314-pyemscripten_2026_0_wasm32.whl", hash = "sha256:9279488933040b638c0ab739084c0ca100efeea6db581bf5d7628d8e89de53fe"},
+    {file = "mypy-2.4.0-cp314-cp314-win_amd64.whl", hash = "sha256:2106b55105ba5ea9be4f53a24517fc5fa927ff1585edc9bc1a975abb72caef89"},
+    {file = "mypy-2.4.0-cp314-cp314-win_arm64.whl", hash = "sha256:528c8744b8b5e3ecb8774f86af38d2376216816e9908317ad055f3c9c2d74799"},
+    {file = "mypy-2.4.0-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:0bb95cf34899e4619c61ab0a8667804e139e580b30d5df12af2102dfe44d0c97"},
+    {file = "mypy-2.4.0-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:86d616fe84c6eab8026f8c50ab5bcb90db780d2ccd233d971e34e92bede9b359"},
+    {file = "mypy-2.4.0-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:a3f86fd1313dd69d013e265f1fdcd12ea7a9d606f9875b2a3db946cd334555f3"},
+    {file = "mypy-2.4.0-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:720434d48542ecfe84d32d287b727569d3fc8f5769acd39051130e490a5c295c"},
+    {file = "mypy-2.4.0-cp314-cp314t-win_amd64.whl", hash = "sha256:a6e851b82c0661f69f1630fc16172c68787a6a9cf0991e7c6437d60976cdcd76"},
+    {file = "mypy-2.4.0-cp314-cp314t-win_arm64.whl", hash = "sha256:3bd0e340f0ebe65c548210f53be3fd8192e83964760caf0c28bef368e68b0d37"},
+    {file = "mypy-2.4.0-cp315-cp315-macosx_11_0_arm64.whl", hash = "sha256:afa89837d9be67e0cadfa33bca3bb7efdda98c3b07e74dc3b635ebfb1c8a926a"},
+    {file = "mypy-2.4.0-cp315-cp315-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:fb443e81057896132d3642d6be219e6efd158691ac7883e3ba8fcb469865f05d"},
+    {file = "mypy-2.4.0-cp315-cp315-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:7f38f57d344f8b6accb40e01c3d83cfc590498231724d16c07ffb7940f157818"},
+    {file = "mypy-2.4.0-cp315-cp315-musllinux_1_2_x86_64.whl", hash = "sha256:e76172710bd4e5eeae061abfd68347e5264632e02778be61784671ae3a2132f5"},
+    {file = "mypy-2.4.0-cp315-cp315-pyemscripten_2026_5_wasm32.whl", hash = "sha256:f83353e47ab520bf6fd4df8f5897d9fe081211f2fbc4b7d37736a3e3c166cbcf"},
+    {file = "mypy-2.4.0-cp315-cp315-win_amd64.whl", hash = "sha256:970b221ed5842213d98e3c480c08f795ace4b1f81fb21e1b126bd0476bce1c34"},
+    {file = "mypy-2.4.0-cp315-cp315-win_arm64.whl", hash = "sha256:502b94b0b331f7dafe32fd6b151797ddbb4f32385b362e722c783a025e5954a3"},
+    {file = "mypy-2.4.0-cp315-cp315t-macosx_11_0_arm64.whl", hash = "sha256:c9de622fd397495695d0598ddc789222bfcfec9d7c9ec3a1e385c855e3bc5e01"},
+    {file = "mypy-2.4.0-cp315-cp315t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:9f459f0b4f0596d9d51fe7716b404b35287b99e77da98a7af90a65dd5fd61141"},
+    {file = "mypy-2.4.0-cp315-cp315t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:f9b028548b3af480e2b1ed8df14ccaac86f99c9f600d1580770ab7ba3dcd40f0"},
+    {file = "mypy-2.4.0-cp315-cp315t-musllinux_1_2_x86_64.whl", hash = "sha256:cb734b2668c1f40d07ce093bbeb4407e9527c67901627b0e1679825be3f09975"},
+    {file = "mypy-2.4.0-cp315-cp315t-win_amd64.whl", hash = "sha256:172e30b8fea631fe310f0c665477f52d9ea40bb4e99e0c81dc30118563b13710"},
+    {file = "mypy-2.4.0-cp315-cp315t-win_arm64.whl", hash = "sha256:5786ef987b3767e51aaa53f20aec104c0252b42ecda7aef8e8b4cbae279b05c5"},
+    {file = "mypy-2.4.0-py3-none-any.whl", hash = "sha256:d01c5d26a352acc6d5cf3128225477e1e8465e8d3029d4c345807fbf7f3cf093"},
+    {file = "mypy-2.4.0.tar.gz", hash = "sha256:77bdaebd452f43fcfc4cc3ba94352a3ea537cd01e3f2d0879f48673d2ec00d6e"},
 ]
 
 [package.dependencies]
-librt = {version = ">=0.8.0", markers = "platform_python_implementation != \"PyPy\""}
+ast-serialize = ">=0.11.2,<1.0.0"
+librt = {version = ">=0.16.0", markers = "platform_python_implementation != \"PyPy\""}
 mypy_extensions = ">=1.0.0"
 pathspec = ">=1.0.0"
 typing_extensions = [
@@ -442,7 +569,6 @@ dmypy = ["psutil (>=4.0)"]
 faster-cache = ["orjson"]
 install-types = ["pip"]
 mypyc = ["setuptools (>=50)"]
-native-parser = ["ast-serialize (>=0.1.1,<1.0.0)"]
 reports = ["lxml"]
 
 [[package]]
@@ -471,30 +597,30 @@ files = [
 
 [[package]]
 name = "openai"
-version = "2.41.0"
+version = "3.37.0"
 description = "The official Python library for the openai API"
 optional = false
-python-versions = ">=3.9"
+python-versions = ">=3.10"
 groups = ["main"]
 files = [
-    {file = "openai-2.41.0-py3-none-any.whl", hash = "sha256:20cc7952e8501c7e5773dd2ef7be437bae9cb549044902e1041a83a54516e375"},
-    {file = "openai-2.41.0.tar.gz", hash = "sha256:db5c362acd6604b84f076abbefa66826ea4b46ecba2954ed866e6a149a1352c0"},
+    {file = "openai-3.37.0-py3-none-any.whl", hash = "sha256:d9f5b000f52397a7f9f946012744c5bc4c962eed3c100385252464c273a63e57"},
+    {file = "openai-3.37.0.tar.gz", hash = "sha256:670bd6f699000d43a2073bcf94374a55bb133bfe5de0d084d47eb9bb9f6a8e5d"},
 ]
 
 [package.dependencies]
-anyio = ">=3.5.0,<5"
-distro = ">=1.7.0,<2"
-httpx = ">=0.23.0,<1"
-jiter = ">=0.10.0,<1"
-pydantic = ">=1.9.0,<3"
+anyio = ">=4.10.0,<5"
+httpx2 = ">=2.12.0,<3"
+jiter = ">=0.16.0,<1"
+pydantic = ">=1.10.13,<2.0.dev0 || >=2.4.dev0,<3"
 sniffio = "*"
-tqdm = ">4"
 typing-extensions = ">=4.14,<5"
 
 [package.extras]
-aiohttp = ["aiohttp", "httpx-aiohttp (>=0.1.9)"]
-datalib = ["numpy (>=1)", "pandas (>=1.2.3)", "pandas-stubs (>=1.1.0.11)"]
+aiohttp = ["aiohttp (>=3.14.3)"]
+bedrock = ["botocore (>=1.40.0,<2)", "urllib3 (>=2.7.0,<3)"]
+datalib = ["numpy (>=1)", "pandas (>=1.2.3)"]
 realtime = ["websockets (>=13,<16)"]
+vertex = ["google-auth[requests] (>=2.36,<3)"]
 voice-helpers = ["numpy (>=2.0.2)", "sounddevice (>=0.5.1)"]
 
 [[package]]
@@ -509,42 +635,6 @@ files = [
     {file = "packaging-26.2.tar.gz", hash = "sha256:ff452ff5a3e828ce110190feff1178bb1f2ea2281fa2075aadb987c2fb221661"},
 ]
 
-[[package]]
-name = "paddler-client"
-version = "4.0.1"
-description = "Python client for Paddler LLM load balancer"
-optional = false
-python-versions = "^3.11"
-groups = ["dev"]
-files = []
-develop = true
-
-[package.dependencies]
-httpx = "^0.28"
-pydantic = "^2"
-websockets = "^15"
-
-[package.source]
-type = "directory"
-url = "../paddler_client_python"
-
-[[package]]
-name = "paddler-test-cluster"
-version = "4.0.1"
-description = "Spawns real Paddler balancers and agents for the Python test suites"
-optional = false
-python-versions = "^3.11"
-groups = ["dev"]
-files = []
-develop = true
-
-[package.dependencies]
-paddler-client = {path = "../paddler_client_python", develop = true}
-
-[package.source]
-type = "directory"
-url = "../paddler_test_cluster_python"
-
 [[package]]
 name = "pathspec"
 version = "1.1.1"
@@ -580,19 +670,19 @@ testing = ["coverage", "pytest", "pytest-benchmark"]
 
 [[package]]
 name = "pydantic"
-version = "2.13.4"
+version = "2.12.5"
 description = "Data validation using Python type hints"
 optional = false
 python-versions = ">=3.9"
-groups = ["main", "dev"]
+groups = ["main"]
 files = [
-    {file = "pydantic-2.13.4-py3-none-any.whl", hash = "sha256:45a282cde31d808236fd7ea9d919b128653c8b38b393d1c4ab335c62924d9aba"},
-    {file = "pydantic-2.13.4.tar.gz", hash = "sha256:c40756b57adaa8b1efeeced5c196f3f3b7c435f90e84ea7f443901bec8099ef6"},
+    {file = "pydantic-2.12.5-py3-none-any.whl", hash = "sha256:e561593fccf61e8a20fc46dfc2dfe075b8be7d0188df33f221ad1f0139180f9d"},
+    {file = "pydantic-2.12.5.tar.gz", hash = "sha256:4d351024c75c0f085a9febbb665ce8c0c6ec5d30e903bdb6394b7ede26aebb49"},
 ]
 
 [package.dependencies]
 annotated-types = ">=0.6.0"
-pydantic-core = "2.46.4"
+pydantic-core = "2.41.5"
 typing-extensions = ">=4.14.1"
 typing-inspection = ">=0.4.2"
 
@@ -602,132 +692,133 @@ timezone = ["tzdata ; python_version >= \"3.9\" and platform_system == \"Windows
 
 [[package]]
 name = "pydantic-core"
-version = "2.46.4"
+version = "2.41.5"
 description = "Core functionality for Pydantic validation and serialization"
 optional = false
 python-versions = ">=3.9"
-groups = ["main", "dev"]
+groups = ["main"]
 files = [
-    {file = "pydantic_core-2.46.4-cp310-cp310-macosx_10_12_x86_64.whl", hash = "sha256:a396dcc17e5a0b164dbe026896245a4fa9ff402edca1dff0be3d53a517f74de4"},
-    {file = "pydantic_core-2.46.4-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:da4b951fe36dc7c3a1ccb4e3cd1747c3542b8c9ceede8fc86cae054e764485f5"},
-    {file = "pydantic_core-2.46.4-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:bb63e0198ca18aad131c089b9204c23079c3afa95487e561f4c522d519e55aba"},
-    {file = "pydantic_core-2.46.4-cp310-cp310-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:f47286a97f0bc9b8859519809077b91b2cefe4ae47fcbf5e466a009c1c5d742b"},
-    {file = "pydantic_core-2.46.4-cp310-cp310-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:905a0ed8ea6f2d61c1738835f99b699348d7857379083e5fc497fa0c967a407c"},
-    {file = "pydantic_core-2.46.4-cp310-cp310-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:ea793e075b70290d89d8142074262885d3f7da19634845135751bd6344f73b50"},
-    {file = "pydantic_core-2.46.4-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:395aebd9183f9d112f569aeb5b2214d1a10a33bec8456447f7fbdfa51d38d4cd"},
-    {file = "pydantic_core-2.46.4-cp310-cp310-manylinux_2_31_riscv64.whl", hash = "sha256:b078afbc25f3a1436c7a1d2cd3e322497ee99615ba97c563566fdf46aff1ee01"},
-    {file = "pydantic_core-2.46.4-cp310-cp310-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:f747929cf940cddb5b3668a390056ddd5ba2e5010615ea2dcf4f9c4f3ab8791d"},
-    {file = "pydantic_core-2.46.4-cp310-cp310-musllinux_1_1_aarch64.whl", hash = "sha256:daa27d92c36f24388fe3ad306b174781c747627f134452e4f128ea00ce1fe8c4"},
-    {file = "pydantic_core-2.46.4-cp310-cp310-musllinux_1_1_armv7l.whl", hash = "sha256:19e51f073cd3df251856a8a4189fbdf1de4012c3ebacfb1884f94f1eb406079f"},
-    {file = "pydantic_core-2.46.4-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:c1747f85cee84c26985853c6f3d9bd3e75da5212912443fa111c113b9c246f39"},
-    {file = "pydantic_core-2.46.4-cp310-cp310-win32.whl", hash = "sha256:2f84c03c8607173d16b5a854ec68a2f9079ae03237a54fb506d13af47e1d018d"},
-    {file = "pydantic_core-2.46.4-cp310-cp310-win_amd64.whl", hash = "sha256:8358a950c8909158e3df31538a7e4edc2d7265a7c54b47f0864d9e5bae9dcebf"},
-    {file = "pydantic_core-2.46.4-cp311-cp311-macosx_10_12_x86_64.whl", hash = "sha256:0e96592440881c74a213e5ad528e2b24d3d4f940de2766bed9010ab1d9e51594"},
-    {file = "pydantic_core-2.46.4-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:e0d65b8c354be7fb5f720c3caa8bc940bc2d20ce749c8e06135f07f8ed95dd7c"},
-    {file = "pydantic_core-2.46.4-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:7bfb192b3f4b9e8a89b6277b6ce787564f62cfd272055f6e685726b111dc7826"},
-    {file = "pydantic_core-2.46.4-cp311-cp311-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:9037063db01f09b09e237c282b6792bd4da634b5402c4e7f0c61effed7701a04"},
-    {file = "pydantic_core-2.46.4-cp311-cp311-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:fc010ab034c8c7452522748bf937df58020d256ccae0874463d1f4d01758af8e"},
-    {file = "pydantic_core-2.46.4-cp311-cp311-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:8c5dac79fa1614d1e06ca695109c6105923bd9c7d1d6c918d4e637b7e6b32fd3"},
-    {file = "pydantic_core-2.46.4-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:f9fa868638bf362d3d138ea55829cefb3d5f4b0d7f142234382a15e2485dbec4"},
-    {file = "pydantic_core-2.46.4-cp311-cp311-manylinux_2_31_riscv64.whl", hash = "sha256:17299feefe090f2caa5b8e37222bb5f663e4935a8bfa6931d4102e5df1a9f398"},
-    {file = "pydantic_core-2.46.4-cp311-cp311-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:4c63ebc82684aa89d9a3bcbd13d515b3be44250dc68dd3bd81526c1cb31286c3"},
-    {file = "pydantic_core-2.46.4-cp311-cp311-musllinux_1_1_aarch64.whl", hash = "sha256:aaa2a54443eff1950ba5ddc6b6ccda0d9c84a364276a62f969bdf2a390650848"},
-    {file = "pydantic_core-2.46.4-cp311-cp311-musllinux_1_1_armv7l.whl", hash = "sha256:18e5ceec2ab67e6d5f1a9085e5a24c9c4e2ac4545730bfe668680bca05e555f3"},
-    {file = "pydantic_core-2.46.4-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:a0f62d0a58f4e7da165457e995725421e0064f2255d8eccebc49f41bbc23b109"},
-    {file = "pydantic_core-2.46.4-cp311-cp311-win32.whl", hash = "sha256:041bde0a48fd37cf71cab1c9d56d3e8625a3793fef1f7dd232b3ff37e978ecda"},
-    {file = "pydantic_core-2.46.4-cp311-cp311-win_amd64.whl", hash = "sha256:6f2eeda33a839975441c86a4119e1383c50b47faf0cbb5176985565c6bb02c33"},
-    {file = "pydantic_core-2.46.4-cp311-cp311-win_arm64.whl", hash = "sha256:14f4c5d6db102bd796a627bbb3a17b4cf4574b9ae861d8b7c9a9661c6dd3362d"},
-    {file = "pydantic_core-2.46.4-cp312-cp312-macosx_10_12_x86_64.whl", hash = "sha256:3245406455a5d98187ec35530fd772b1d799b26667980872c8d4614991e2c4a2"},
-    {file = "pydantic_core-2.46.4-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:962ccbab7b642487b1d8b7df90ef677e03134cf1fd8880bf698649b22a69371f"},
-    {file = "pydantic_core-2.46.4-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:8233f2947cf85404441fd7e0085f53b10c93e0ee78611099b5c7237e36aacbf7"},
-    {file = "pydantic_core-2.46.4-cp312-cp312-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:3a233125ac121aa3ffba9a2b59edfc4a985a76092dc8279586ab4b71390875e7"},
-    {file = "pydantic_core-2.46.4-cp312-cp312-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:5b712b53160b79a5850310b912a5ef8e57e56947c8ad690c227f5c9d7e561712"},
-    {file = "pydantic_core-2.46.4-cp312-cp312-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:9401557acd873c3a7f3eb9383edef8ac4968f9510e340f4808d427e75667e7b4"},
-    {file = "pydantic_core-2.46.4-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:926c9541b14b12b1681dca8a0b75feb510b06c6341b70a8e500c2fdcff837cce"},
-    {file = "pydantic_core-2.46.4-cp312-cp312-manylinux_2_31_riscv64.whl", hash = "sha256:56cb4851bcaf3d117eddcef4fe66afd750a50274b0da8e22be256d10e5611987"},
-    {file = "pydantic_core-2.46.4-cp312-cp312-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:c68fcd102d71ea85c5b2dfac3f4f8476eff42a9e078fd5faefff6d145063536b"},
-    {file = "pydantic_core-2.46.4-cp312-cp312-musllinux_1_1_aarch64.whl", hash = "sha256:b2f69dec1725e79a012d920df1707de5caf7ed5e08f3be4435e25803efc47458"},
-    {file = "pydantic_core-2.46.4-cp312-cp312-musllinux_1_1_armv7l.whl", hash = "sha256:8d0820e8192167f80d88d64038e609c31452eeca865b4e1d9950a27a4609b00b"},
-    {file = "pydantic_core-2.46.4-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:fbdb89b3e1c94a30cc5edfce477c6e6a5dc4d8f84665b455c27582f211a1c72c"},
-    {file = "pydantic_core-2.46.4-cp312-cp312-win32.whl", hash = "sha256:9aa768456404a8bf48a4406685ac2bec8e72b62c69313734fa3b73cf33b3a894"},
-    {file = "pydantic_core-2.46.4-cp312-cp312-win_amd64.whl", hash = "sha256:e9c26f834c65f5752f3f06cb08cb86a913ceb7274d0db6e267808a708b46bc89"},
-    {file = "pydantic_core-2.46.4-cp312-cp312-win_arm64.whl", hash = "sha256:4fc73cb559bdb54b1134a706a2802a4cddd27a0633f5abb7e53056268751ac6a"},
-    {file = "pydantic_core-2.46.4-cp313-cp313-macosx_10_12_x86_64.whl", hash = "sha256:5d5902252db0d3cedf8d4a1bc68f70eeb430f7e4c7104c8c476753519b423008"},
-    {file = "pydantic_core-2.46.4-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:c94f0688e7b8d0a67abf40e57a7eaaecd17cc9586706a31b76c031f63df052b4"},
-    {file = "pydantic_core-2.46.4-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:f027324c56cd5406ca49c124b0db10e56c69064fec039acc571c29020cc87c76"},
-    {file = "pydantic_core-2.46.4-cp313-cp313-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:e739fee756ba1010f8bcccb534252e85a35fe45ae92c295a06059ce58b74ccd3"},
-    {file = "pydantic_core-2.46.4-cp313-cp313-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:9d56801be94b86a9da183e5f3766e6310752b99ff647e38b09a9500d88e46e76"},
-    {file = "pydantic_core-2.46.4-cp313-cp313-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:2412e734dcb48da14d4e4006b82b46b74f2518b8a26ee7e58c6844a6cd6d03c4"},
-    {file = "pydantic_core-2.46.4-cp313-cp313-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:9551187363ffc0de2a00b2e47c25aeaeb1020b69b668762966df15fc5659dd5a"},
-    {file = "pydantic_core-2.46.4-cp313-cp313-manylinux_2_31_riscv64.whl", hash = "sha256:0186750b482eefa11d7f435892b09c5c606193ef3375bcf94aa00ae6bfb66262"},
-    {file = "pydantic_core-2.46.4-cp313-cp313-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:5855698a4856556d86e8e6cd8434bc3ac0314ee8e12089ae0e143f64c6256e4e"},
-    {file = "pydantic_core-2.46.4-cp313-cp313-musllinux_1_1_aarch64.whl", hash = "sha256:cbaf13819775b7f769bf4a1f066cb6df7a28d4480081a589828ef190226881cd"},
-    {file = "pydantic_core-2.46.4-cp313-cp313-musllinux_1_1_armv7l.whl", hash = "sha256:633147d34cf4550417f12e2b1a0383973bdf5cdfde212cb09e9a581cf10820be"},
-    {file = "pydantic_core-2.46.4-cp313-cp313-musllinux_1_1_x86_64.whl", hash = "sha256:82cf5301172168103724d49a1444d3378cb20cdee30b116a1bd6031236298a5d"},
-    {file = "pydantic_core-2.46.4-cp313-cp313-win32.whl", hash = "sha256:9fa8ae11da9e2b3126c6426f147e0fba88d96d65921799bb30c6abd1cb2c97fb"},
-    {file = "pydantic_core-2.46.4-cp313-cp313-win_amd64.whl", hash = "sha256:6b3ace8194b0e5204818c92802dcdca7fc6d88aabbb799d7c795540d9cd6d292"},
-    {file = "pydantic_core-2.46.4-cp313-cp313-win_arm64.whl", hash = "sha256:184c081504d17f1c1066e430e117142b2c77d9448a97f7b65c6ac9fd9aee238d"},
-    {file = "pydantic_core-2.46.4-cp314-cp314-macosx_10_12_x86_64.whl", hash = "sha256:428e04521a40150c85216fc8b85e8d39fece235a9cf5e383761238c7fa9b96fb"},
-    {file = "pydantic_core-2.46.4-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:23ace664830ee0bfe014a0c7bc248b1f7f25ed7ad103852c317624a1083af462"},
-    {file = "pydantic_core-2.46.4-cp314-cp314-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:ce5c1d2a8b27468f433ca974829c44060b8097eedc39933e3c206a90ee49c4a9"},
-    {file = "pydantic_core-2.46.4-cp314-cp314-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:7283d57845ecf5a163403eb0702dfc220cc4fbdd18919cb5ccea4f95ee1cdab4"},
-    {file = "pydantic_core-2.46.4-cp314-cp314-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:8daafc69c93ee8a0204506a3b6b30f586ef54028f52aeeeb5c4cfc5184fd5914"},
-    {file = "pydantic_core-2.46.4-cp314-cp314-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:cd2213145bcc2ba85884d0ac63d222fece9209678f77b9b4d76f054c561adb28"},
-    {file = "pydantic_core-2.46.4-cp314-cp314-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:7a5f930472650a82629163023e630d160863fce524c616f4e5186e5de9d9a49b"},
-    {file = "pydantic_core-2.46.4-cp314-cp314-manylinux_2_31_riscv64.whl", hash = "sha256:c1b3f518abeca3aa13c712fd202306e145abf59a18b094a6bafb2d2bbf59192c"},
-    {file = "pydantic_core-2.46.4-cp314-cp314-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:1a7dd0b3ee80d90150e3495a3a13ac34dbcbfd4f012996a6a1d8900e91b5c0fb"},
-    {file = "pydantic_core-2.46.4-cp314-cp314-musllinux_1_1_aarch64.whl", hash = "sha256:3fb702cd90b0446a3a1c5e470bfa0dd23c0233b676a9099ddcc964fa6ca13898"},
-    {file = "pydantic_core-2.46.4-cp314-cp314-musllinux_1_1_armv7l.whl", hash = "sha256:b8458003118a712e66286df6a707db01c52c0f52f7db8e4a38f0da1d3b94fc4e"},
-    {file = "pydantic_core-2.46.4-cp314-cp314-musllinux_1_1_x86_64.whl", hash = "sha256:372429a130e469c9cd698925ce5fc50940b7a1336b0d82038e63d5bbc4edc519"},
-    {file = "pydantic_core-2.46.4-cp314-cp314-win32.whl", hash = "sha256:85bb3611ff1802f3ee7fdd7dbff26b56f343fb432d57a4728fdd49b6ef35e2f4"},
-    {file = "pydantic_core-2.46.4-cp314-cp314-win_amd64.whl", hash = "sha256:811ff8e9c313ab425368bcbb36e5c4ebd7108c2bbf4e4089cfbb0b01eff63fac"},
-    {file = "pydantic_core-2.46.4-cp314-cp314-win_arm64.whl", hash = "sha256:bfec22eab3c8cc2ceec0248aec886624116dc079afa027ecc8ad4a7e62010f8a"},
-    {file = "pydantic_core-2.46.4-cp314-cp314t-macosx_10_12_x86_64.whl", hash = "sha256:af8244b2bef6aaad6d92cda81372de7f8c8d36c9f0c3ea36e827c60e7d9467a0"},
-    {file = "pydantic_core-2.46.4-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:5a4330cdbc57162e4b3aa303f588ba752257694c9c9be3e7ebb11b4aca659b5d"},
-    {file = "pydantic_core-2.46.4-cp314-cp314t-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:29c61fc04a3d840155ff08e475a04809278972fe6aef51e2720554e96367e34b"},
-    {file = "pydantic_core-2.46.4-cp314-cp314t-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:c50f2528cf200c5eed56faf3f4e22fcd5f38c157a8b78576e6ba3168ec35f000"},
-    {file = "pydantic_core-2.46.4-cp314-cp314t-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:0cbe8b01f948de4286c74cdd6c667aceb38f5c1e26f0693b3983d9d74887c65e"},
-    {file = "pydantic_core-2.46.4-cp314-cp314t-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:617d7e2ca7dcb8c5cf6bcb8c59b8832c94b36196bbf1cbd1bfb56ed341905edd"},
-    {file = "pydantic_core-2.46.4-cp314-cp314t-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:7027560ee92211647d0d34e3f7cd6f50da56399d26a9c8ad0da286d3869a53f3"},
-    {file = "pydantic_core-2.46.4-cp314-cp314t-manylinux_2_31_riscv64.whl", hash = "sha256:f99626688942fb746e545232e7726926f3be91b5975f8b55327665fafda991c7"},
-    {file = "pydantic_core-2.46.4-cp314-cp314t-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:fc3e9034a63de20e15e8ade85358bc6efc614008cab72898b4b4952bea0509ff"},
-    {file = "pydantic_core-2.46.4-cp314-cp314t-musllinux_1_1_aarch64.whl", hash = "sha256:97e7cf2be5c77b7d1a9713a05605d49460d02c6078d38d8bef3cbe323c548424"},
-    {file = "pydantic_core-2.46.4-cp314-cp314t-musllinux_1_1_armv7l.whl", hash = "sha256:3bf92c5d0e00fefaab325a4d27828fe6b6e2a21848686b5b60d2d9eeb09d76c6"},
-    {file = "pydantic_core-2.46.4-cp314-cp314t-musllinux_1_1_x86_64.whl", hash = "sha256:3ecbc122d18468d06ca279dc26a8c2e2d5acb10943bb35e36ae92096dc3b5565"},
-    {file = "pydantic_core-2.46.4-cp314-cp314t-win32.whl", hash = "sha256:e846ae7835bf0703ae43f534ab79a867146dadd59dc9ca5c8b53d5c8f7c9ef02"},
-    {file = "pydantic_core-2.46.4-cp314-cp314t-win_amd64.whl", hash = "sha256:2108ba5c1c1eca18030634489dc544844144ee36357f2f9f780b93e7ddbb44b5"},
-    {file = "pydantic_core-2.46.4-cp314-cp314t-win_arm64.whl", hash = "sha256:4fcbe087dbc2068af7eda3aa87634eba216dbda64d1ae73c8684b621d33f6596"},
-    {file = "pydantic_core-2.46.4-cp39-cp39-macosx_10_12_x86_64.whl", hash = "sha256:fd8b3d9fd264be37976686c7f65cd52a83f5e84f4bfd2adf9c1d469676bbb6ae"},
-    {file = "pydantic_core-2.46.4-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:9f444c499b3eefd3a92e348059471ea0c3a6e303d9c1cec09fa748fd9f895201"},
-    {file = "pydantic_core-2.46.4-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:3447661d99f75a3683a4cf5c87da72f2161964611864dbbeac7fbb118bb4bfc0"},
-    {file = "pydantic_core-2.46.4-cp39-cp39-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:8b9bab013d1c7a79d3501ff86d0bc9c31bf587db4551677b96bec07df78c6b15"},
-    {file = "pydantic_core-2.46.4-cp39-cp39-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:d995260fdf4e1db774581b4900e0f832abe3c7c84996726bbc161b19c8f29e76"},
-    {file = "pydantic_core-2.46.4-cp39-cp39-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:f13a646d65d09fbf1bc6b3a9635d30095c8e7e5cc419ff35ecc563c5fd04cd49"},
-    {file = "pydantic_core-2.46.4-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:432c179df7874eeb73307aad2df0755e1ae0efa61ff0ea89b93e194411ae3928"},
-    {file = "pydantic_core-2.46.4-cp39-cp39-manylinux_2_31_riscv64.whl", hash = "sha256:e68b7a074f65a2fd746c52a7ce6142ab7006074ac269ace0c25cd8ba171f8066"},
-    {file = "pydantic_core-2.46.4-cp39-cp39-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:4a05d69cba51d852c5c3e92758653245a50c0b646ced0cf05bd793ed592839d6"},
-    {file = "pydantic_core-2.46.4-cp39-cp39-musllinux_1_1_aarch64.whl", hash = "sha256:228ee9bae8bef5b1e97ec58302f80357c37199e0d0a99174e138d28e6957b9d9"},
-    {file = "pydantic_core-2.46.4-cp39-cp39-musllinux_1_1_armv7l.whl", hash = "sha256:10e17cbb10a330363733efc4d7c4d0dd827ac0909b8f6a6542298fed1ea62f29"},
-    {file = "pydantic_core-2.46.4-cp39-cp39-musllinux_1_1_x86_64.whl", hash = "sha256:91a06d2e259ecfbd8c901d70c3c507900458498142b3026a296b7de4d1322cc9"},
-    {file = "pydantic_core-2.46.4-cp39-cp39-win32.whl", hash = "sha256:d80ee3d731373b24cebbc10d689ca4ee1875caf0d5703a245db18efd4dd37fc1"},
-    {file = "pydantic_core-2.46.4-cp39-cp39-win_amd64.whl", hash = "sha256:3be77f45df024d789a672ae34f8b06fb346c4f9f46ea714956660ea4862e89ac"},
-    {file = "pydantic_core-2.46.4-graalpy311-graalpy242_311_native-macosx_10_12_x86_64.whl", hash = "sha256:14d4edf427bdcf950a8a02d7cb44a08614388dd6e1bdcbf4f67504fa7887da9c"},
-    {file = "pydantic_core-2.46.4-graalpy311-graalpy242_311_native-macosx_11_0_arm64.whl", hash = "sha256:0ce40cd7b21210e99342afafbd4d0f76d784eb5b1d60f3bdc566be4983c6c73b"},
-    {file = "pydantic_core-2.46.4-graalpy311-graalpy242_311_native-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:90884113d8b48f760e9587002789ddd741e76ab9f89518cd1e43b1f1a52ec44b"},
-    {file = "pydantic_core-2.46.4-graalpy311-graalpy242_311_native-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:66ce7632c22d837c95301830e111ad0128a32b8207533b60896a96c4915192ea"},
-    {file = "pydantic_core-2.46.4-graalpy312-graalpy250_312_native-macosx_10_12_x86_64.whl", hash = "sha256:1d8ba486450b14f3b1d63bc521d410ec7565e52f887b9fb671791886436a42f7"},
-    {file = "pydantic_core-2.46.4-graalpy312-graalpy250_312_native-macosx_11_0_arm64.whl", hash = "sha256:3009f12e4e90b7f88b4f9adb1b0c4a3d58fe7820f3238c190047209d148026df"},
-    {file = "pydantic_core-2.46.4-graalpy312-graalpy250_312_native-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:ad785e92e6dc634c21555edc8bd6b64957ab844541bcb96a1366c202951ae526"},
-    {file = "pydantic_core-2.46.4-graalpy312-graalpy250_312_native-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:00c603d540afdd6b80eb39f078f33ebd46211f02f33e34a32d9f053bba711de0"},
-    {file = "pydantic_core-2.46.4-pp311-pypy311_pp73-macosx_10_12_x86_64.whl", hash = "sha256:0c563b08bca408dc7f65f700633d8442fffb2421fc47b8101377e9fd65051ff0"},
-    {file = "pydantic_core-2.46.4-pp311-pypy311_pp73-macosx_11_0_arm64.whl", hash = "sha256:db06ffe51636ffe9ca531fe9023dd64bdd794be8754cb5df57c5498ae5b518a7"},
-    {file = "pydantic_core-2.46.4-pp311-pypy311_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:133878133d271ade3d41d1bfb2a45ec38dbdbda40bc065921c6b04e4630127e2"},
-    {file = "pydantic_core-2.46.4-pp311-pypy311_pp73-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:9bc519fbf2b7578398853d815009ae5e4d4603d12f4e3f91da8c06852d3da3e9"},
-    {file = "pydantic_core-2.46.4-pp311-pypy311_pp73-musllinux_1_1_aarch64.whl", hash = "sha256:c7a7bd4e39e8e4c12c39cd480356842b6a8a06e41b23a55a5e3e191718838ddf"},
-    {file = "pydantic_core-2.46.4-pp311-pypy311_pp73-musllinux_1_1_armv7l.whl", hash = "sha256:d396ec2b979760aaf3218e76c24e65bd0aca24983298653b3a9d7a45f9e47b30"},
-    {file = "pydantic_core-2.46.4-pp311-pypy311_pp73-musllinux_1_1_x86_64.whl", hash = "sha256:86e1a4418c6cd97d60c95c71164158eaf7324fae7b0923264016baa993eba6fc"},
-    {file = "pydantic_core-2.46.4-pp311-pypy311_pp73-win_amd64.whl", hash = "sha256:d51026d73fcfd93610abc7b27789c26b313920fcfb20e27462d74a7f8b06e983"},
-    {file = "pydantic_core-2.46.4.tar.gz", hash = "sha256:62f875393d7f270851f20523dd2e29f082bcc82292d66db2b64ea71f64b6e1c1"},
+    {file = "pydantic_core-2.41.5-cp310-cp310-macosx_10_12_x86_64.whl", hash = "sha256:77b63866ca88d804225eaa4af3e664c5faf3568cea95360d21f4725ab6e07146"},
+    {file = "pydantic_core-2.41.5-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:dfa8a0c812ac681395907e71e1274819dec685fec28273a28905df579ef137e2"},
+    {file = "pydantic_core-2.41.5-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:5921a4d3ca3aee735d9fd163808f5e8dd6c6972101e4adbda9a4667908849b97"},
+    {file = "pydantic_core-2.41.5-cp310-cp310-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:e25c479382d26a2a41b7ebea1043564a937db462816ea07afa8a44c0866d52f9"},
+    {file = "pydantic_core-2.41.5-cp310-cp310-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:f547144f2966e1e16ae626d8ce72b4cfa0caedc7fa28052001c94fb2fcaa1c52"},
+    {file = "pydantic_core-2.41.5-cp310-cp310-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:6f52298fbd394f9ed112d56f3d11aabd0d5bd27beb3084cc3d8ad069483b8941"},
+    {file = "pydantic_core-2.41.5-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:100baa204bb412b74fe285fb0f3a385256dad1d1879f0a5cb1499ed2e83d132a"},
+    {file = "pydantic_core-2.41.5-cp310-cp310-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:05a2c8852530ad2812cb7914dc61a1125dc4e06252ee98e5638a12da6cc6fb6c"},
+    {file = "pydantic_core-2.41.5-cp310-cp310-musllinux_1_1_aarch64.whl", hash = "sha256:29452c56df2ed968d18d7e21f4ab0ac55e71dc59524872f6fc57dcf4a3249ed2"},
+    {file = "pydantic_core-2.41.5-cp310-cp310-musllinux_1_1_armv7l.whl", hash = "sha256:d5160812ea7a8a2ffbe233d8da666880cad0cbaf5d4de74ae15c313213d62556"},
+    {file = "pydantic_core-2.41.5-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:df3959765b553b9440adfd3c795617c352154e497a4eaf3752555cfb5da8fc49"},
+    {file = "pydantic_core-2.41.5-cp310-cp310-win32.whl", hash = "sha256:1f8d33a7f4d5a7889e60dc39856d76d09333d8a6ed0f5f1190635cbec70ec4ba"},
+    {file = "pydantic_core-2.41.5-cp310-cp310-win_amd64.whl", hash = "sha256:62de39db01b8d593e45871af2af9e497295db8d73b085f6bfd0b18c83c70a8f9"},
+    {file = "pydantic_core-2.41.5-cp311-cp311-macosx_10_12_x86_64.whl", hash = "sha256:a3a52f6156e73e7ccb0f8cced536adccb7042be67cb45f9562e12b319c119da6"},
+    {file = "pydantic_core-2.41.5-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:7f3bf998340c6d4b0c9a2f02d6a400e51f123b59565d74dc60d252ce888c260b"},
+    {file = "pydantic_core-2.41.5-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:378bec5c66998815d224c9ca994f1e14c0c21cb95d2f52b6021cc0b2a58f2a5a"},
+    {file = "pydantic_core-2.41.5-cp311-cp311-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:e7b576130c69225432866fe2f4a469a85a54ade141d96fd396dffcf607b558f8"},
+    {file = "pydantic_core-2.41.5-cp311-cp311-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:6cb58b9c66f7e4179a2d5e0f849c48eff5c1fca560994d6eb6543abf955a149e"},
+    {file = "pydantic_core-2.41.5-cp311-cp311-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:88942d3a3dff3afc8288c21e565e476fc278902ae4d6d134f1eeda118cc830b1"},
+    {file = "pydantic_core-2.41.5-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:f31d95a179f8d64d90f6831d71fa93290893a33148d890ba15de25642c5d075b"},
+    {file = "pydantic_core-2.41.5-cp311-cp311-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:c1df3d34aced70add6f867a8cf413e299177e0c22660cc767218373d0779487b"},
+    {file = "pydantic_core-2.41.5-cp311-cp311-musllinux_1_1_aarch64.whl", hash = "sha256:4009935984bd36bd2c774e13f9a09563ce8de4abaa7226f5108262fa3e637284"},
+    {file = "pydantic_core-2.41.5-cp311-cp311-musllinux_1_1_armv7l.whl", hash = "sha256:34a64bc3441dc1213096a20fe27e8e128bd3ff89921706e83c0b1ac971276594"},
+    {file = "pydantic_core-2.41.5-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:c9e19dd6e28fdcaa5a1de679aec4141f691023916427ef9bae8584f9c2fb3b0e"},
+    {file = "pydantic_core-2.41.5-cp311-cp311-win32.whl", hash = "sha256:2c010c6ded393148374c0f6f0bf89d206bf3217f201faa0635dcd56bd1520f6b"},
+    {file = "pydantic_core-2.41.5-cp311-cp311-win_amd64.whl", hash = "sha256:76ee27c6e9c7f16f47db7a94157112a2f3a00e958bc626e2f4ee8bec5c328fbe"},
+    {file = "pydantic_core-2.41.5-cp311-cp311-win_arm64.whl", hash = "sha256:4bc36bbc0b7584de96561184ad7f012478987882ebf9f9c389b23f432ea3d90f"},
+    {file = "pydantic_core-2.41.5-cp312-cp312-macosx_10_12_x86_64.whl", hash = "sha256:f41a7489d32336dbf2199c8c0a215390a751c5b014c2c1c5366e817202e9cdf7"},
+    {file = "pydantic_core-2.41.5-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:070259a8818988b9a84a449a2a7337c7f430a22acc0859c6b110aa7212a6d9c0"},
+    {file = "pydantic_core-2.41.5-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:e96cea19e34778f8d59fe40775a7a574d95816eb150850a85a7a4c8f4b94ac69"},
+    {file = "pydantic_core-2.41.5-cp312-cp312-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:ed2e99c456e3fadd05c991f8f437ef902e00eedf34320ba2b0842bd1c3ca3a75"},
+    {file = "pydantic_core-2.41.5-cp312-cp312-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:65840751b72fbfd82c3c640cff9284545342a4f1eb1586ad0636955b261b0b05"},
+    {file = "pydantic_core-2.41.5-cp312-cp312-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:e536c98a7626a98feb2d3eaf75944ef6f3dbee447e1f841eae16f2f0a72d8ddc"},
+    {file = "pydantic_core-2.41.5-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:eceb81a8d74f9267ef4081e246ffd6d129da5d87e37a77c9bde550cb04870c1c"},
+    {file = "pydantic_core-2.41.5-cp312-cp312-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:d38548150c39b74aeeb0ce8ee1d8e82696f4a4e16ddc6de7b1d8823f7de4b9b5"},
+    {file = "pydantic_core-2.41.5-cp312-cp312-musllinux_1_1_aarch64.whl", hash = "sha256:c23e27686783f60290e36827f9c626e63154b82b116d7fe9adba1fda36da706c"},
+    {file = "pydantic_core-2.41.5-cp312-cp312-musllinux_1_1_armv7l.whl", hash = "sha256:482c982f814460eabe1d3bb0adfdc583387bd4691ef00b90575ca0d2b6fe2294"},
+    {file = "pydantic_core-2.41.5-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:bfea2a5f0b4d8d43adf9d7b8bf019fb46fdd10a2e5cde477fbcb9d1fa08c68e1"},
+    {file = "pydantic_core-2.41.5-cp312-cp312-win32.whl", hash = "sha256:b74557b16e390ec12dca509bce9264c3bbd128f8a2c376eaa68003d7f327276d"},
+    {file = "pydantic_core-2.41.5-cp312-cp312-win_amd64.whl", hash = "sha256:1962293292865bca8e54702b08a4f26da73adc83dd1fcf26fbc875b35d81c815"},
+    {file = "pydantic_core-2.41.5-cp312-cp312-win_arm64.whl", hash = "sha256:1746d4a3d9a794cacae06a5eaaccb4b8643a131d45fbc9af23e353dc0a5ba5c3"},
+    {file = "pydantic_core-2.41.5-cp313-cp313-macosx_10_12_x86_64.whl", hash = "sha256:941103c9be18ac8daf7b7adca8228f8ed6bb7a1849020f643b3a14d15b1924d9"},
+    {file = "pydantic_core-2.41.5-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:112e305c3314f40c93998e567879e887a3160bb8689ef3d2c04b6cc62c33ac34"},
+    {file = "pydantic_core-2.41.5-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:0cbaad15cb0c90aa221d43c00e77bb33c93e8d36e0bf74760cd00e732d10a6a0"},
+    {file = "pydantic_core-2.41.5-cp313-cp313-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:03ca43e12fab6023fc79d28ca6b39b05f794ad08ec2feccc59a339b02f2b3d33"},
+    {file = "pydantic_core-2.41.5-cp313-cp313-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:dc799088c08fa04e43144b164feb0c13f9a0bc40503f8df3e9fde58a3c0c101e"},
+    {file = "pydantic_core-2.41.5-cp313-cp313-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:97aeba56665b4c3235a0e52b2c2f5ae9cd071b8a8310ad27bddb3f7fb30e9aa2"},
+    {file = "pydantic_core-2.41.5-cp313-cp313-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:406bf18d345822d6c21366031003612b9c77b3e29ffdb0f612367352aab7d586"},
+    {file = "pydantic_core-2.41.5-cp313-cp313-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:b93590ae81f7010dbe380cdeab6f515902ebcbefe0b9327cc4804d74e93ae69d"},
+    {file = "pydantic_core-2.41.5-cp313-cp313-musllinux_1_1_aarch64.whl", hash = "sha256:01a3d0ab748ee531f4ea6c3e48ad9dac84ddba4b0d82291f87248f2f9de8d740"},
+    {file = "pydantic_core-2.41.5-cp313-cp313-musllinux_1_1_armv7l.whl", hash = "sha256:6561e94ba9dacc9c61bce40e2d6bdc3bfaa0259d3ff36ace3b1e6901936d2e3e"},
+    {file = "pydantic_core-2.41.5-cp313-cp313-musllinux_1_1_x86_64.whl", hash = "sha256:915c3d10f81bec3a74fbd4faebe8391013ba61e5a1a8d48c4455b923bdda7858"},
+    {file = "pydantic_core-2.41.5-cp313-cp313-win32.whl", hash = "sha256:650ae77860b45cfa6e2cdafc42618ceafab3a2d9a3811fcfbd3bbf8ac3c40d36"},
+    {file = "pydantic_core-2.41.5-cp313-cp313-win_amd64.whl", hash = "sha256:79ec52ec461e99e13791ec6508c722742ad745571f234ea6255bed38c6480f11"},
+    {file = "pydantic_core-2.41.5-cp313-cp313-win_arm64.whl", hash = "sha256:3f84d5c1b4ab906093bdc1ff10484838aca54ef08de4afa9de0f5f14d69639cd"},
+    {file = "pydantic_core-2.41.5-cp314-cp314-macosx_10_12_x86_64.whl", hash = "sha256:3f37a19d7ebcdd20b96485056ba9e8b304e27d9904d233d7b1015db320e51f0a"},
+    {file = "pydantic_core-2.41.5-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:1d1d9764366c73f996edd17abb6d9d7649a7eb690006ab6adbda117717099b14"},
+    {file = "pydantic_core-2.41.5-cp314-cp314-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:25e1c2af0fce638d5f1988b686f3b3ea8cd7de5f244ca147c777769e798a9cd1"},
+    {file = "pydantic_core-2.41.5-cp314-cp314-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:506d766a8727beef16b7adaeb8ee6217c64fc813646b424d0804d67c16eddb66"},
+    {file = "pydantic_core-2.41.5-cp314-cp314-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:4819fa52133c9aa3c387b3328f25c1facc356491e6135b459f1de698ff64d869"},
+    {file = "pydantic_core-2.41.5-cp314-cp314-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:2b761d210c9ea91feda40d25b4efe82a1707da2ef62901466a42492c028553a2"},
+    {file = "pydantic_core-2.41.5-cp314-cp314-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:22f0fb8c1c583a3b6f24df2470833b40207e907b90c928cc8d3594b76f874375"},
+    {file = "pydantic_core-2.41.5-cp314-cp314-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:2782c870e99878c634505236d81e5443092fba820f0373997ff75f90f68cd553"},
+    {file = "pydantic_core-2.41.5-cp314-cp314-musllinux_1_1_aarch64.whl", hash = "sha256:0177272f88ab8312479336e1d777f6b124537d47f2123f89cb37e0accea97f90"},
+    {file = "pydantic_core-2.41.5-cp314-cp314-musllinux_1_1_armv7l.whl", hash = "sha256:63510af5e38f8955b8ee5687740d6ebf7c2a0886d15a6d65c32814613681bc07"},
+    {file = "pydantic_core-2.41.5-cp314-cp314-musllinux_1_1_x86_64.whl", hash = "sha256:e56ba91f47764cc14f1daacd723e3e82d1a89d783f0f5afe9c364b8bb491ccdb"},
+    {file = "pydantic_core-2.41.5-cp314-cp314-win32.whl", hash = "sha256:aec5cf2fd867b4ff45b9959f8b20ea3993fc93e63c7363fe6851424c8a7e7c23"},
+    {file = "pydantic_core-2.41.5-cp314-cp314-win_amd64.whl", hash = "sha256:8e7c86f27c585ef37c35e56a96363ab8de4e549a95512445b85c96d3e2f7c1bf"},
+    {file = "pydantic_core-2.41.5-cp314-cp314-win_arm64.whl", hash = "sha256:e672ba74fbc2dc8eea59fb6d4aed6845e6905fc2a8afe93175d94a83ba2a01a0"},
+    {file = "pydantic_core-2.41.5-cp314-cp314t-macosx_10_12_x86_64.whl", hash = "sha256:8566def80554c3faa0e65ac30ab0932b9e3a5cd7f8323764303d468e5c37595a"},
+    {file = "pydantic_core-2.41.5-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:b80aa5095cd3109962a298ce14110ae16b8c1aece8b72f9dafe81cf597ad80b3"},
+    {file = "pydantic_core-2.41.5-cp314-cp314t-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:3006c3dd9ba34b0c094c544c6006cc79e87d8612999f1a5d43b769b89181f23c"},
+    {file = "pydantic_core-2.41.5-cp314-cp314t-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:72f6c8b11857a856bcfa48c86f5368439f74453563f951e473514579d44aa612"},
+    {file = "pydantic_core-2.41.5-cp314-cp314t-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:5cb1b2f9742240e4bb26b652a5aeb840aa4b417c7748b6f8387927bc6e45e40d"},
+    {file = "pydantic_core-2.41.5-cp314-cp314t-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:bd3d54f38609ff308209bd43acea66061494157703364ae40c951f83ba99a1a9"},
+    {file = "pydantic_core-2.41.5-cp314-cp314t-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:2ff4321e56e879ee8d2a879501c8e469414d948f4aba74a2d4593184eb326660"},
+    {file = "pydantic_core-2.41.5-cp314-cp314t-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:d0d2568a8c11bf8225044aa94409e21da0cb09dcdafe9ecd10250b2baad531a9"},
+    {file = "pydantic_core-2.41.5-cp314-cp314t-musllinux_1_1_aarch64.whl", hash = "sha256:a39455728aabd58ceabb03c90e12f71fd30fa69615760a075b9fec596456ccc3"},
+    {file = "pydantic_core-2.41.5-cp314-cp314t-musllinux_1_1_armv7l.whl", hash = "sha256:239edca560d05757817c13dc17c50766136d21f7cd0fac50295499ae24f90fdf"},
+    {file = "pydantic_core-2.41.5-cp314-cp314t-musllinux_1_1_x86_64.whl", hash = "sha256:2a5e06546e19f24c6a96a129142a75cee553cc018ffee48a460059b1185f4470"},
+    {file = "pydantic_core-2.41.5-cp314-cp314t-win32.whl", hash = "sha256:b4ececa40ac28afa90871c2cc2b9ffd2ff0bf749380fbdf57d165fd23da353aa"},
+    {file = "pydantic_core-2.41.5-cp314-cp314t-win_amd64.whl", hash = "sha256:80aa89cad80b32a912a65332f64a4450ed00966111b6615ca6816153d3585a8c"},
+    {file = "pydantic_core-2.41.5-cp314-cp314t-win_arm64.whl", hash = "sha256:35b44f37a3199f771c3eaa53051bc8a70cd7b54f333531c59e29fd4db5d15008"},
+    {file = "pydantic_core-2.41.5-cp39-cp39-macosx_10_12_x86_64.whl", hash = "sha256:8bfeaf8735be79f225f3fefab7f941c712aaca36f1128c9d7e2352ee1aa87bdf"},
+    {file = "pydantic_core-2.41.5-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:346285d28e4c8017da95144c7f3acd42740d637ff41946af5ce6e5e420502dd5"},
+    {file = "pydantic_core-2.41.5-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:a75dafbf87d6276ddc5b2bf6fae5254e3d0876b626eb24969a574fff9149ee5d"},
+    {file = "pydantic_core-2.41.5-cp39-cp39-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:7b93a4d08587e2b7e7882de461e82b6ed76d9026ce91ca7915e740ecc7855f60"},
+    {file = "pydantic_core-2.41.5-cp39-cp39-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:e8465ab91a4bd96d36dde3263f06caa6a8a6019e4113f24dc753d79a8b3a3f82"},
+    {file = "pydantic_core-2.41.5-cp39-cp39-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:299e0a22e7ae2b85c1a57f104538b2656e8ab1873511fd718a1c1c6f149b77b5"},
+    {file = "pydantic_core-2.41.5-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:707625ef0983fcfb461acfaf14de2067c5942c6bb0f3b4c99158bed6fedd3cf3"},
+    {file = "pydantic_core-2.41.5-cp39-cp39-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:f41eb9797986d6ebac5e8edff36d5cef9de40def462311b3eb3eeded1431e425"},
+    {file = "pydantic_core-2.41.5-cp39-cp39-musllinux_1_1_aarch64.whl", hash = "sha256:0384e2e1021894b1ff5a786dbf94771e2986ebe2869533874d7e43bc79c6f504"},
+    {file = "pydantic_core-2.41.5-cp39-cp39-musllinux_1_1_armv7l.whl", hash = "sha256:f0cd744688278965817fd0839c4a4116add48d23890d468bc436f78beb28abf5"},
+    {file = "pydantic_core-2.41.5-cp39-cp39-musllinux_1_1_x86_64.whl", hash = "sha256:753e230374206729bf0a807954bcc6c150d3743928a73faffee51ac6557a03c3"},
+    {file = "pydantic_core-2.41.5-cp39-cp39-win32.whl", hash = "sha256:873e0d5b4fb9b89ef7c2d2a963ea7d02879d9da0da8d9d4933dee8ee86a8b460"},
+    {file = "pydantic_core-2.41.5-cp39-cp39-win_amd64.whl", hash = "sha256:e4f4a984405e91527a0d62649ee21138f8e3d0ef103be488c1dc11a80d7f184b"},
+    {file = "pydantic_core-2.41.5-graalpy311-graalpy242_311_native-macosx_10_12_x86_64.whl", hash = "sha256:b96d5f26b05d03cc60f11a7761a5ded1741da411e7fe0909e27a5e6a0cb7b034"},
+    {file = "pydantic_core-2.41.5-graalpy311-graalpy242_311_native-macosx_11_0_arm64.whl", hash = "sha256:634e8609e89ceecea15e2d61bc9ac3718caaaa71963717bf3c8f38bfde64242c"},
+    {file = "pydantic_core-2.41.5-graalpy311-graalpy242_311_native-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:93e8740d7503eb008aa2df04d3b9735f845d43ae845e6dcd2be0b55a2da43cd2"},
+    {file = "pydantic_core-2.41.5-graalpy311-graalpy242_311_native-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:f15489ba13d61f670dcc96772e733aad1a6f9c429cc27574c6cdaed82d0146ad"},
+    {file = "pydantic_core-2.41.5-graalpy312-graalpy250_312_native-macosx_10_12_x86_64.whl", hash = "sha256:7da7087d756b19037bc2c06edc6c170eeef3c3bafcb8f532ff17d64dc427adfd"},
+    {file = "pydantic_core-2.41.5-graalpy312-graalpy250_312_native-macosx_11_0_arm64.whl", hash = "sha256:aabf5777b5c8ca26f7824cb4a120a740c9588ed58df9b2d196ce92fba42ff8dc"},
+    {file = "pydantic_core-2.41.5-graalpy312-graalpy250_312_native-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:c007fe8a43d43b3969e8469004e9845944f1a80e6acd47c150856bb87f230c56"},
+    {file = "pydantic_core-2.41.5-graalpy312-graalpy250_312_native-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:76d0819de158cd855d1cbb8fcafdf6f5cf1eb8e470abe056d5d161106e38062b"},
+    {file = "pydantic_core-2.41.5-pp310-pypy310_pp73-macosx_10_12_x86_64.whl", hash = "sha256:b5819cd790dbf0c5eb9f82c73c16b39a65dd6dd4d1439dcdea7816ec9adddab8"},
+    {file = "pydantic_core-2.41.5-pp310-pypy310_pp73-macosx_11_0_arm64.whl", hash = "sha256:5a4e67afbc95fa5c34cf27d9089bca7fcab4e51e57278d710320a70b956d1b9a"},
+    {file = "pydantic_core-2.41.5-pp310-pypy310_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ece5c59f0ce7d001e017643d8d24da587ea1f74f6993467d85ae8a5ef9d4f42b"},
+    {file = "pydantic_core-2.41.5-pp310-pypy310_pp73-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:16f80f7abe3351f8ea6858914ddc8c77e02578544a0ebc15b4c2e1a0e813b0b2"},
+    {file = "pydantic_core-2.41.5-pp310-pypy310_pp73-musllinux_1_1_aarch64.whl", hash = "sha256:33cb885e759a705b426baada1fe68cbb0a2e68e34c5d0d0289a364cf01709093"},
+    {file = "pydantic_core-2.41.5-pp310-pypy310_pp73-musllinux_1_1_armv7l.whl", hash = "sha256:c8d8b4eb992936023be7dee581270af5c6e0697a8559895f527f5b7105ecd36a"},
+    {file = "pydantic_core-2.41.5-pp310-pypy310_pp73-musllinux_1_1_x86_64.whl", hash = "sha256:242a206cd0318f95cd21bdacff3fcc3aab23e79bba5cac3db5a841c9ef9c6963"},
+    {file = "pydantic_core-2.41.5-pp310-pypy310_pp73-win_amd64.whl", hash = "sha256:d3a978c4f57a597908b7e697229d996d77a6d3c94901e9edee593adada95ce1a"},
+    {file = "pydantic_core-2.41.5-pp311-pypy311_pp73-macosx_10_12_x86_64.whl", hash = "sha256:b2379fa7ed44ddecb5bfe4e48577d752db9fc10be00a6b7446e9663ba143de26"},
+    {file = "pydantic_core-2.41.5-pp311-pypy311_pp73-macosx_11_0_arm64.whl", hash = "sha256:266fb4cbf5e3cbd0b53669a6d1b039c45e3ce651fd5442eff4d07c2cc8d66808"},
+    {file = "pydantic_core-2.41.5-pp311-pypy311_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:58133647260ea01e4d0500089a8c4f07bd7aa6ce109682b1426394988d8aaacc"},
+    {file = "pydantic_core-2.41.5-pp311-pypy311_pp73-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:287dad91cfb551c363dc62899a80e9e14da1f0e2b6ebde82c806612ca2a13ef1"},
+    {file = "pydantic_core-2.41.5-pp311-pypy311_pp73-musllinux_1_1_aarch64.whl", hash = "sha256:03b77d184b9eb40240ae9fd676ca364ce1085f203e1b1256f8ab9984dca80a84"},
+    {file = "pydantic_core-2.41.5-pp311-pypy311_pp73-musllinux_1_1_armv7l.whl", hash = "sha256:a668ce24de96165bb239160b3d854943128f4334822900534f2fe947930e5770"},
+    {file = "pydantic_core-2.41.5-pp311-pypy311_pp73-musllinux_1_1_x86_64.whl", hash = "sha256:f14f8f046c14563f8eb3f45f499cc658ab8d10072961e07225e507adb700e93f"},
+    {file = "pydantic_core-2.41.5-pp311-pypy311_pp73-win_amd64.whl", hash = "sha256:56121965f7a4dc965bff783d70b907ddf3d57f6eba29b6d2e5dabfaf07799c51"},
+    {file = "pydantic_core-2.41.5.tar.gz", hash = "sha256:08daa51ea16ad373ffd5e7606252cc32f07bc72b28284b6bc9c6df804816476e"},
 ]
 
 [package.dependencies]
@@ -750,14 +841,14 @@ windows-terminal = ["colorama (>=0.4.6)"]
 
 [[package]]
 name = "pyright"
-version = "1.1.410"
+version = "1.1.414"
 description = "Command line wrapper for pyright"
 optional = false
 python-versions = ">=3.7"
 groups = ["dev"]
 files = [
-    {file = "pyright-1.1.410-py3-none-any.whl", hash = "sha256:5e961bed37cacf96b3f7cd7b1da39b350a9239aa2e69138d0e88f728cfaf296c"},
-    {file = "pyright-1.1.410.tar.gz", hash = "sha256:07a073b8ba6749826773c1269773efa11b93440d9a6aa60419d9a3172d6dc488"},
+    {file = "pyright-1.1.414-py3-none-any.whl", hash = "sha256:2a6b4b3298c9eec174c5ed83bd338de6eee82df2992f3e1930e6199d381be36f"},
+    {file = "pyright-1.1.414.tar.gz", hash = "sha256:523c0a97c60da6333234955c277730c9cf4f5bd6d5399e7b7d2b0fc5d3599524"},
 ]
 
 [package.dependencies]
@@ -771,14 +862,14 @@ nodejs = ["nodejs-wheel-binaries"]
 
 [[package]]
 name = "pytest"
-version = "9.0.3"
+version = "9.1.1"
 description = "pytest: simple powerful testing with Python"
 optional = false
 python-versions = ">=3.10"
 groups = ["dev"]
 files = [
-    {file = "pytest-9.0.3-py3-none-any.whl", hash = "sha256:2c5efc453d45394fdd706ade797c0a81091eccd1d6e4bccfcd476e2b8e0ab5d9"},
-    {file = "pytest-9.0.3.tar.gz", hash = "sha256:b86ada508af81d19edeb213c681b1d48246c1a91d304c6c81a427674c17eb91c"},
+    {file = "pytest-9.1.1-py3-none-any.whl", hash = "sha256:37a86b45efb9a47a61a36449063e8e18d0cab3161329fc099eb21783169c4f0c"},
+    {file = "pytest-9.1.1.tar.gz", hash = "sha256:1088fbde8f2b49d95a549a195707afa7a76a3ce9bcadc26b6d71f0ffda5fe313"},
 ]
 
 [package.dependencies]
@@ -791,54 +882,6 @@ pygments = ">=2.7.2"
 [package.extras]
 dev = ["argcomplete", "attrs (>=19.2)", "hypothesis (>=3.56)", "mock", "requests", "setuptools", "xmlschema"]
 
-[[package]]
-name = "pytest-asyncio"
-version = "1.3.0"
-description = "Pytest support for asyncio"
-optional = false
-python-versions = ">=3.10"
-groups = ["dev"]
-files = [
-    {file = "pytest_asyncio-1.3.0-py3-none-any.whl", hash = "sha256:611e26147c7f77640e6d0a92a38ed17c3e9848063698d5c93d5aa7aa11cebff5"},
-    {file = "pytest_asyncio-1.3.0.tar.gz", hash = "sha256:d7f52f36d231b80ee124cd216ffb19369aa168fc10095013c6b014a34d3ee9e5"},
-]
-
-[package.dependencies]
-pytest = ">=8.2,<10"
-typing-extensions = {version = ">=4.12", markers = "python_version < \"3.13\""}
-
-[package.extras]
-docs = ["sphinx (>=5.3)", "sphinx-rtd-theme (>=1)"]
-testing = ["coverage (>=6.2)", "hypothesis (>=5.7.1)"]
-
-[[package]]
-name = "ruff"
-version = "0.15.12"
-description = "An extremely fast Python linter and code formatter, written in Rust."
-optional = false
-python-versions = ">=3.7"
-groups = ["dev"]
-files = [
-    {file = "ruff-0.15.12-py3-none-linux_armv6l.whl", hash = "sha256:f86f176e188e94d6bdbc09f09bfd9dc729059ad93d0e7390b5a73efe19f8861c"},
-    {file = "ruff-0.15.12-py3-none-macosx_10_12_x86_64.whl", hash = "sha256:e3bcd123364c3770b8e1b7baaf343cc99a35f197c5c6e8af79015c666c423a6c"},
-    {file = "ruff-0.15.12-py3-none-macosx_11_0_arm64.whl", hash = "sha256:fe87510d000220aa1ed530d4448a7c696a0cae1213e5ec30e5874287b66557b5"},
-    {file = "ruff-0.15.12-py3-none-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:84a1630093121375a3e2a95b4a6dc7b59e2b4ee76216e32d81aae550a832d002"},
-    {file = "ruff-0.15.12-py3-none-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:fb129f40f114f089ebe0ca56c0d251cf2061b17651d464bb6478dc01e69f11f5"},
-    {file = "ruff-0.15.12-py3-none-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:b0c862b172d695db7598426b8af465e7e9ac00a3ea2a3630ee67eb82e366aaa6"},
-    {file = "ruff-0.15.12-py3-none-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:2849ea9f3484c3aca43a82f484210370319e7170df4dfe4843395ddf6c57bc33"},
-    {file = "ruff-0.15.12-py3-none-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:9e77c7e51c07fe396826d5969a5b846d9cd4c402535835fb6e21ce8b28fef847"},
-    {file = "ruff-0.15.12-py3-none-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:83b2f4f2f3b1026b5fb449b467d9264bf22067b600f7b6f41fc5958909f449d0"},
-    {file = "ruff-0.15.12-py3-none-manylinux_2_31_riscv64.whl", hash = "sha256:9ba3b8f1afd7e2e43d8943e55f249e13f9682fde09711644a6e7290eb4f3e339"},
-    {file = "ruff-0.15.12-py3-none-musllinux_1_2_aarch64.whl", hash = "sha256:e852ba9fdc890655e1d78f2df1499efbe0e54126bd405362154a75e2bde159c5"},
-    {file = "ruff-0.15.12-py3-none-musllinux_1_2_armv7l.whl", hash = "sha256:dd8aed930da53780d22fc70bdf84452c843cf64f8cb4eb38984319c24c5cd5fd"},
-    {file = "ruff-0.15.12-py3-none-musllinux_1_2_i686.whl", hash = "sha256:01da3988d225628b709493d7dc67c3b9b12c0210016b08690ef9bd27970b262b"},
-    {file = "ruff-0.15.12-py3-none-musllinux_1_2_x86_64.whl", hash = "sha256:9cae0f92bd5700d1213188b31cd3bdd2b315361296d10b96b8e2337d3d11f53e"},
-    {file = "ruff-0.15.12-py3-none-win32.whl", hash = "sha256:d0185894e038d7043ba8fd6aee7499ece6462dc0ea9f1e260c7451807c714c20"},
-    {file = "ruff-0.15.12-py3-none-win_amd64.whl", hash = "sha256:c87a162d61ab3adca47c03f7f717c68672edec7d1b5499e652331780fe74950d"},
-    {file = "ruff-0.15.12-py3-none-win_arm64.whl", hash = "sha256:a538f7a82d061cee7be55542aca1d86d1393d55d81d4fcc314370f4340930d4f"},
-    {file = "ruff-0.15.12.tar.gz", hash = "sha256:ecea26adb26b4232c0c2ca19ccbc0083a68344180bba2a600605538ce51a40a6"},
-]
-
 [[package]]
 name = "sniffio"
 version = "1.3.1"
@@ -852,27 +895,18 @@ files = [
 ]
 
 [[package]]
-name = "tqdm"
-version = "4.67.3"
-description = "Fast, Extensible Progress Meter"
+name = "truststore"
+version = "0.10.5"
+description = "Verify certificates using native system trust stores"
 optional = false
-python-versions = ">=3.7"
+python-versions = ">=3.10"
 groups = ["main"]
+markers = "sys_platform != \"emscripten\""
 files = [
-    {file = "tqdm-4.67.3-py3-none-any.whl", hash = "sha256:ee1e4c0e59148062281c49d80b25b67771a127c85fc9676d3be5f243206826bf"},
-    {file = "tqdm-4.67.3.tar.gz", hash = "sha256:7d825f03f89244ef73f1d4ce193cb1774a8179fd96f31d7e1dcde62092b960bb"},
+    {file = "truststore-0.10.5-py3-none-any.whl", hash = "sha256:9aaaedaefaf06d8b206278cf8b5012bc897f485a874503501e12d776df78951c"},
+    {file = "truststore-0.10.5.tar.gz", hash = "sha256:30d36967ccaded5cbb38d602c433f53600036c79d502f4533a49b60a03bbefcd"},
 ]
 
-[package.dependencies]
-colorama = {version = "*", markers = "platform_system == \"Windows\""}
-
-[package.extras]
-dev = ["nbval", "pytest (>=6)", "pytest-asyncio (>=0.24)", "pytest-cov", "pytest-timeout"]
-discord = ["requests"]
-notebook = ["ipywidgets (>=6)"]
-slack = ["slack-sdk"]
-telegram = ["requests"]
-
 [[package]]
 name = "typing-extensions"
 version = "4.16.0"
@@ -891,7 +925,7 @@ version = "0.4.2"
 description = "Runtime typing introspection tools"
 optional = false
 python-versions = ">=3.9"
-groups = ["main", "dev"]
+groups = ["main"]
 files = [
     {file = "typing_inspection-0.4.2-py3-none-any.whl", hash = "sha256:4ed1cacbdc298c220f1bd249ed5287caa16f34d44ef4e9c3d0cbad5b521545e7"},
     {file = "typing_inspection-0.4.2.tar.gz", hash = "sha256:ba561c48a67c5958007083d386c3295464928b01faa735ab8547c5692e87f464"},
@@ -900,86 +934,7 @@ files = [
 [package.dependencies]
 typing-extensions = ">=4.12.0"
 
-[[package]]
-name = "websockets"
-version = "15.0.1"
-description = "An implementation of the WebSocket Protocol (RFC 6455 & 7692)"
-optional = false
-python-versions = ">=3.9"
-groups = ["dev"]
-files = [
-    {file = "websockets-15.0.1-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:d63efaa0cd96cf0c5fe4d581521d9fa87744540d4bc999ae6e08595a1014b45b"},
-    {file = "websockets-15.0.1-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:ac60e3b188ec7574cb761b08d50fcedf9d77f1530352db4eef1707fe9dee7205"},
-    {file = "websockets-15.0.1-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:5756779642579d902eed757b21b0164cd6fe338506a8083eb58af5c372e39d9a"},
-    {file = "websockets-15.0.1-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:0fdfe3e2a29e4db3659dbd5bbf04560cea53dd9610273917799f1cde46aa725e"},
-    {file = "websockets-15.0.1-cp310-cp310-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:4c2529b320eb9e35af0fa3016c187dffb84a3ecc572bcee7c3ce302bfeba52bf"},
-    {file = "websockets-15.0.1-cp310-cp310-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ac1e5c9054fe23226fb11e05a6e630837f074174c4c2f0fe442996112a6de4fb"},
-    {file = "websockets-15.0.1-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:5df592cd503496351d6dc14f7cdad49f268d8e618f80dce0cd5a36b93c3fc08d"},
-    {file = "websockets-15.0.1-cp310-cp310-musllinux_1_2_i686.whl", hash = "sha256:0a34631031a8f05657e8e90903e656959234f3a04552259458aac0b0f9ae6fd9"},
-    {file = "websockets-15.0.1-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:3d00075aa65772e7ce9e990cab3ff1de702aa09be3940d1dc88d5abf1ab8a09c"},
-    {file = "websockets-15.0.1-cp310-cp310-win32.whl", hash = "sha256:1234d4ef35db82f5446dca8e35a7da7964d02c127b095e172e54397fb6a6c256"},
-    {file = "websockets-15.0.1-cp310-cp310-win_amd64.whl", hash = "sha256:39c1fec2c11dc8d89bba6b2bf1556af381611a173ac2b511cf7231622058af41"},
-    {file = "websockets-15.0.1-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:823c248b690b2fd9303ba00c4f66cd5e2d8c3ba4aa968b2779be9532a4dad431"},
-    {file = "websockets-15.0.1-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:678999709e68425ae2593acf2e3ebcbcf2e69885a5ee78f9eb80e6e371f1bf57"},
-    {file = "websockets-15.0.1-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:d50fd1ee42388dcfb2b3676132c78116490976f1300da28eb629272d5d93e905"},
-    {file = "websockets-15.0.1-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:d99e5546bf73dbad5bf3547174cd6cb8ba7273062a23808ffea025ecb1cf8562"},
-    {file = "websockets-15.0.1-cp311-cp311-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:66dd88c918e3287efc22409d426c8f729688d89a0c587c88971a0faa2c2f3792"},
-    {file = "websockets-15.0.1-cp311-cp311-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:8dd8327c795b3e3f219760fa603dcae1dcc148172290a8ab15158cf85a953413"},
-    {file = "websockets-15.0.1-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:8fdc51055e6ff4adeb88d58a11042ec9a5eae317a0a53d12c062c8a8865909e8"},
-    {file = "websockets-15.0.1-cp311-cp311-musllinux_1_2_i686.whl", hash = "sha256:693f0192126df6c2327cce3baa7c06f2a117575e32ab2308f7f8216c29d9e2e3"},
-    {file = "websockets-15.0.1-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:54479983bd5fb469c38f2f5c7e3a24f9a4e70594cd68cd1fa6b9340dadaff7cf"},
-    {file = "websockets-15.0.1-cp311-cp311-win32.whl", hash = "sha256:16b6c1b3e57799b9d38427dda63edcbe4926352c47cf88588c0be4ace18dac85"},
-    {file = "websockets-15.0.1-cp311-cp311-win_amd64.whl", hash = "sha256:27ccee0071a0e75d22cb35849b1db43f2ecd3e161041ac1ee9d2352ddf72f065"},
-    {file = "websockets-15.0.1-cp312-cp312-macosx_10_13_universal2.whl", hash = "sha256:3e90baa811a5d73f3ca0bcbf32064d663ed81318ab225ee4f427ad4e26e5aff3"},
-    {file = "websockets-15.0.1-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:592f1a9fe869c778694f0aa806ba0374e97648ab57936f092fd9d87f8bc03665"},
-    {file = "websockets-15.0.1-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:0701bc3cfcb9164d04a14b149fd74be7347a530ad3bbf15ab2c678a2cd3dd9a2"},
-    {file = "websockets-15.0.1-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:e8b56bdcdb4505c8078cb6c7157d9811a85790f2f2b3632c7d1462ab5783d215"},
-    {file = "websockets-15.0.1-cp312-cp312-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:0af68c55afbd5f07986df82831c7bff04846928ea8d1fd7f30052638788bc9b5"},
-    {file = "websockets-15.0.1-cp312-cp312-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:64dee438fed052b52e4f98f76c5790513235efaa1ef7f3f2192c392cd7c91b65"},
-    {file = "websockets-15.0.1-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:d5f6b181bb38171a8ad1d6aa58a67a6aa9d4b38d0f8c5f496b9e42561dfc62fe"},
-    {file = "websockets-15.0.1-cp312-cp312-musllinux_1_2_i686.whl", hash = "sha256:5d54b09eba2bada6011aea5375542a157637b91029687eb4fdb2dab11059c1b4"},
-    {file = "websockets-15.0.1-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:3be571a8b5afed347da347bfcf27ba12b069d9d7f42cb8c7028b5e98bbb12597"},
-    {file = "websockets-15.0.1-cp312-cp312-win32.whl", hash = "sha256:c338ffa0520bdb12fbc527265235639fb76e7bc7faafbb93f6ba80d9c06578a9"},
-    {file = "websockets-15.0.1-cp312-cp312-win_amd64.whl", hash = "sha256:fcd5cf9e305d7b8338754470cf69cf81f420459dbae8a3b40cee57417f4614a7"},
-    {file = "websockets-15.0.1-cp313-cp313-macosx_10_13_universal2.whl", hash = "sha256:ee443ef070bb3b6ed74514f5efaa37a252af57c90eb33b956d35c8e9c10a1931"},
-    {file = "websockets-15.0.1-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:5a939de6b7b4e18ca683218320fc67ea886038265fd1ed30173f5ce3f8e85675"},
-    {file = "websockets-15.0.1-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:746ee8dba912cd6fc889a8147168991d50ed70447bf18bcda7039f7d2e3d9151"},
-    {file = "websockets-15.0.1-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:595b6c3969023ecf9041b2936ac3827e4623bfa3ccf007575f04c5a6aa318c22"},
-    {file = "websockets-15.0.1-cp313-cp313-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:3c714d2fc58b5ca3e285461a4cc0c9a66bd0e24c5da9911e30158286c9b5be7f"},
-    {file = "websockets-15.0.1-cp313-cp313-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:0f3c1e2ab208db911594ae5b4f79addeb3501604a165019dd221c0bdcabe4db8"},
-    {file = "websockets-15.0.1-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:229cf1d3ca6c1804400b0a9790dc66528e08a6a1feec0d5040e8b9eb14422375"},
-    {file = "websockets-15.0.1-cp313-cp313-musllinux_1_2_i686.whl", hash = "sha256:756c56e867a90fb00177d530dca4b097dd753cde348448a1012ed6c5131f8b7d"},
-    {file = "websockets-15.0.1-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:558d023b3df0bffe50a04e710bc87742de35060580a293c2a984299ed83bc4e4"},
-    {file = "websockets-15.0.1-cp313-cp313-win32.whl", hash = "sha256:ba9e56e8ceeeedb2e080147ba85ffcd5cd0711b89576b83784d8605a7df455fa"},
-    {file = "websockets-15.0.1-cp313-cp313-win_amd64.whl", hash = "sha256:e09473f095a819042ecb2ab9465aee615bd9c2028e4ef7d933600a8401c79561"},
-    {file = "websockets-15.0.1-cp39-cp39-macosx_10_9_universal2.whl", hash = "sha256:5f4c04ead5aed67c8a1a20491d54cdfba5884507a48dd798ecaf13c74c4489f5"},
-    {file = "websockets-15.0.1-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:abdc0c6c8c648b4805c5eacd131910d2a7f6455dfd3becab248ef108e89ab16a"},
-    {file = "websockets-15.0.1-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:a625e06551975f4b7ea7102bc43895b90742746797e2e14b70ed61c43a90f09b"},
-    {file = "websockets-15.0.1-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:d591f8de75824cbb7acad4e05d2d710484f15f29d4a915092675ad3456f11770"},
-    {file = "websockets-15.0.1-cp39-cp39-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:47819cea040f31d670cc8d324bb6435c6f133b8c7a19ec3d61634e62f8d8f9eb"},
-    {file = "websockets-15.0.1-cp39-cp39-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ac017dd64572e5c3bd01939121e4d16cf30e5d7e110a119399cf3133b63ad054"},
-    {file = "websockets-15.0.1-cp39-cp39-musllinux_1_2_aarch64.whl", hash = "sha256:4a9fac8e469d04ce6c25bb2610dc535235bd4aa14996b4e6dbebf5e007eba5ee"},
-    {file = "websockets-15.0.1-cp39-cp39-musllinux_1_2_i686.whl", hash = "sha256:363c6f671b761efcb30608d24925a382497c12c506b51661883c3e22337265ed"},
-    {file = "websockets-15.0.1-cp39-cp39-musllinux_1_2_x86_64.whl", hash = "sha256:2034693ad3097d5355bfdacfffcbd3ef5694f9718ab7f29c29689a9eae841880"},
-    {file = "websockets-15.0.1-cp39-cp39-win32.whl", hash = "sha256:3b1ac0d3e594bf121308112697cf4b32be538fb1444468fb0a6ae4feebc83411"},
-    {file = "websockets-15.0.1-cp39-cp39-win_amd64.whl", hash = "sha256:b7643a03db5c95c799b89b31c036d5f27eeb4d259c798e878d6937d71832b1e4"},
-    {file = "websockets-15.0.1-pp310-pypy310_pp73-macosx_10_15_x86_64.whl", hash = "sha256:0c9e74d766f2818bb95f84c25be4dea09841ac0f734d1966f415e4edfc4ef1c3"},
-    {file = "websockets-15.0.1-pp310-pypy310_pp73-macosx_11_0_arm64.whl", hash = "sha256:1009ee0c7739c08a0cd59de430d6de452a55e42d6b522de7aa15e6f67db0b8e1"},
-    {file = "websockets-15.0.1-pp310-pypy310_pp73-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:76d1f20b1c7a2fa82367e04982e708723ba0e7b8d43aa643d3dcd404d74f1475"},
-    {file = "websockets-15.0.1-pp310-pypy310_pp73-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:f29d80eb9a9263b8d109135351caf568cc3f80b9928bccde535c235de55c22d9"},
-    {file = "websockets-15.0.1-pp310-pypy310_pp73-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:b359ed09954d7c18bbc1680f380c7301f92c60bf924171629c5db97febb12f04"},
-    {file = "websockets-15.0.1-pp310-pypy310_pp73-win_amd64.whl", hash = "sha256:cad21560da69f4ce7658ca2cb83138fb4cf695a2ba3e475e0559e05991aa8122"},
-    {file = "websockets-15.0.1-pp39-pypy39_pp73-macosx_10_15_x86_64.whl", hash = "sha256:7f493881579c90fc262d9cdbaa05a6b54b3811c2f300766748db79f098db9940"},
-    {file = "websockets-15.0.1-pp39-pypy39_pp73-macosx_11_0_arm64.whl", hash = "sha256:47b099e1f4fbc95b701b6e85768e1fcdaf1630f3cbe4765fa216596f12310e2e"},
-    {file = "websockets-15.0.1-pp39-pypy39_pp73-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:67f2b6de947f8c757db2db9c71527933ad0019737ec374a8a6be9a956786aaf9"},
-    {file = "websockets-15.0.1-pp39-pypy39_pp73-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:d08eb4c2b7d6c41da6ca0600c077e93f5adcfd979cd777d747e9ee624556da4b"},
-    {file = "websockets-15.0.1-pp39-pypy39_pp73-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:4b826973a4a2ae47ba357e4e82fa44a463b8f168e1ca775ac64521442b19e87f"},
-    {file = "websockets-15.0.1-pp39-pypy39_pp73-win_amd64.whl", hash = "sha256:21c1fa28a6a7e3cbdc171c694398b6df4744613ce9b36b1a498e816787e28123"},
-    {file = "websockets-15.0.1-py3-none-any.whl", hash = "sha256:f7a866fbc1e97b5c617ee4116daaa09b722101d4a3c170c787450ba409f9736f"},
-    {file = "websockets-15.0.1.tar.gz", hash = "sha256:82544de02076bafba038ce055ee6412d68da13ab47f0c60cab827346de828dee"},
-]
-
 [metadata]
 lock-version = "2.1"
 python-versions = "^3.11"
-content-hash = "e78fc13c43bf7afb5b8a5bcbcf8539c67b7fa268788bfcfc2653ef5a17144c94"
+content-hash = "ba264ae97bc21055e565fb8dd84f41fa96270432e76056d79c4532b190fcc2fa"
diff --git a/paddler_openai_client_python_test/pyproject.toml b/paddler_openai_client_python_test/pyproject.toml
index c1122fd61..9da51c73d 100644
--- a/paddler_openai_client_python_test/pyproject.toml
+++ b/paddler_openai_client_python_test/pyproject.toml
@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "paddler-openai-client-python-test"
-version = "4.0.1"
+version = "5.0.0"
 description = "Drives Paddler's OpenAI-compatible endpoints with the OpenAI client to verify real-client compatibility"
 authors = ["Intentee"]
 license = "Apache-2.0"
@@ -8,15 +8,12 @@ package-mode = false
 
 [tool.poetry.dependencies]
 python = "^3.11"
-openai = "==2.41.0"
+openai = "==3.37.0"
 
 [tool.poetry.group.dev.dependencies]
-paddler-test-cluster = { path = "../paddler_test_cluster_python", develop = true }
-pytest = "9.0.3"
-pytest-asyncio = "1.3.0"
-pyright = "1.1.410"
-ruff = "0.15.12"
-mypy = "1.20.2"
+pytest = "9.1.1"
+pyright = "1.1.414"
+mypy = "2.4.0"
 
 [tool.mypy]
 python_version = "3.11"
@@ -28,26 +25,6 @@ pythonVersion = "3.11"
 pythonPlatform = "All"
 typeCheckingMode = "strict"
 
-[tool.ruff]
-line-length = 88
-target-version = "py311"
-
-[tool.ruff.lint]
-select = ["ALL"]
-ignore = [
-    "COM812",  # trailing comma - conflicts with formatter
-    "D",       # docstrings - project prefers self-documenting code
-    "ISC001",  # implicit string concatenation - conflicts with formatter
-]
-
-[tool.ruff.lint.per-file-ignores]
-"tests/**/*.py" = [
-    "INP001",  # the tests directory is a pytest suite, not an importable package
-    "PLR2004", # magic values are fine in test assertions
-    "S101",    # assert statements are the point of a test suite
-]
-
 [tool.pytest.ini_options]
-asyncio_mode = "auto"
 testpaths = ["tests"]
 addopts = "-v"
diff --git a/paddler_openai_client_python_test/tests/conftest.py b/paddler_openai_client_python_test/tests/conftest.py
index 1a646fe8d..ff10727ff 100644
--- a/paddler_openai_client_python_test/tests/conftest.py
+++ b/paddler_openai_client_python_test/tests/conftest.py
@@ -1,12 +1,8 @@
-from collections.abc import AsyncIterator
+from collections.abc import Iterator
+from os import environ
 
 import pytest
 from openai import OpenAI
-from paddler_test_cluster.agent_spec import AgentSpec
-from paddler_test_cluster.paddler_cluster import paddler_cluster
-from paddler_test_cluster.qwen3_0_6b_desired_state import QWEN3_0_6B_DESIRED_STATE
-
-QWEN3_AGENT = AgentSpec(name="qwen3-agent", slots=1)
 
 
 @pytest.fixture
@@ -15,9 +11,8 @@ def model() -> str:
 
 
 @pytest.fixture
-async def openai_client() -> AsyncIterator[OpenAI]:
-    async with paddler_cluster(QWEN3_0_6B_DESIRED_STATE, [QWEN3_AGENT]) as addresses:
-        with OpenAI(
-            base_url=f"{addresses.compat_openai_url}/v1", api_key="paddler"
-        ) as openai_client:
-            yield openai_client
+def openai_client() -> Iterator[OpenAI]:
+    with OpenAI(
+        base_url=f"{environ['PADDLER_COMPAT_OPENAI_URL']}/v1", api_key="paddler"
+    ) as openai_client:
+        yield openai_client
diff --git a/paddler_openai_client_python_test/tests/test_chat_completion_returns_message_content_and_usage.py b/paddler_openai_client_python_test/tests/test_chat_completion_returns_message_content_and_usage.py
deleted file mode 100644
index f75453ba2..000000000
--- a/paddler_openai_client_python_test/tests/test_chat_completion_returns_message_content_and_usage.py
+++ /dev/null
@@ -1,18 +0,0 @@
-from openai import OpenAI
-
-
-def test_chat_completion_returns_message_content_and_usage(
-    openai_client: OpenAI,
-    model: str,
-) -> None:
-    completion = openai_client.chat.completions.create(
-        model=model,
-        messages=[{"role": "user", "content": "Say hi briefly."}],
-        max_completion_tokens=600,
-    )
-
-    assert completion.object == "chat.completion"
-    assert completion.choices
-    assert completion.choices[0].message.content
-    assert completion.usage is not None
-    assert completion.usage.total_tokens > 0
diff --git a/paddler_openai_client_python_test/tests/test_chat_completion_stream_accumulates_content_and_reports_usage.py b/paddler_openai_client_python_test/tests/test_chat_completion_stream_accumulates_content_and_reports_usage.py
deleted file mode 100644
index 6616dd5c4..000000000
--- a/paddler_openai_client_python_test/tests/test_chat_completion_stream_accumulates_content_and_reports_usage.py
+++ /dev/null
@@ -1,35 +0,0 @@
-from openai import OpenAI
-
-
-def test_chat_completion_stream_accumulates_content_and_reports_usage(
-    openai_client: OpenAI,
-    model: str,
-) -> None:
-    stream = openai_client.chat.completions.create(
-        model=model,
-        messages=[{"role": "user", "content": "Say hi briefly."}],
-        max_completion_tokens=600,
-        stream=True,
-        stream_options={"include_usage": True},
-    )
-
-    content = ""
-    finish_reason: str | None = None
-    total_tokens = 0
-
-    for chunk in stream:
-        assert chunk.object == "chat.completion.chunk"
-
-        if chunk.choices:
-            choice = chunk.choices[0]
-            if choice.delta.content:
-                content += choice.delta.content
-            if choice.finish_reason:
-                finish_reason = choice.finish_reason
-
-        if chunk.usage is not None:
-            total_tokens = chunk.usage.total_tokens
-
-    assert content
-    assert finish_reason == "stop"
-    assert total_tokens > 0
diff --git a/paddler_openai_client_python_test/tests/test_chat_completions_return_message_content_and_usage.py b/paddler_openai_client_python_test/tests/test_chat_completions_return_message_content_and_usage.py
new file mode 100644
index 000000000..db090bbd1
--- /dev/null
+++ b/paddler_openai_client_python_test/tests/test_chat_completions_return_message_content_and_usage.py
@@ -0,0 +1,18 @@
+from openai import OpenAI
+
+
+def test_chat_completions_return_message_content_and_usage(
+    openai_client: OpenAI,
+    model: str,
+) -> None:
+    completion = openai_client.chat.completions.create(
+        model=model,
+        messages=[{"role": "user", "content": "Say hi briefly."}],
+        max_completion_tokens=600,
+    )
+
+    assert completion.object == "chat.completion"
+    assert completion.choices
+    assert completion.choices[0].message.content
+    assert completion.usage is not None
+    assert completion.usage.total_tokens > 0
diff --git a/paddler_openai_client_python_test/tests/test_chat_completions_stream_accumulates_content_and_reports_usage.py b/paddler_openai_client_python_test/tests/test_chat_completions_stream_accumulates_content_and_reports_usage.py
new file mode 100644
index 000000000..ea4fba316
--- /dev/null
+++ b/paddler_openai_client_python_test/tests/test_chat_completions_stream_accumulates_content_and_reports_usage.py
@@ -0,0 +1,35 @@
+from openai import OpenAI
+
+
+def test_chat_completions_stream_accumulates_content_and_reports_usage(
+    openai_client: OpenAI,
+    model: str,
+) -> None:
+    stream = openai_client.chat.completions.create(
+        model=model,
+        messages=[{"role": "user", "content": "Say hi briefly."}],
+        max_completion_tokens=600,
+        stream=True,
+        stream_options={"include_usage": True},
+    )
+
+    content = ""
+    finish_reason: str | None = None
+    total_tokens = 0
+
+    for chunk in stream:
+        assert chunk.object == "chat.completion.chunk"
+
+        if chunk.choices:
+            choice = chunk.choices[0]
+            if choice.delta.content:
+                content += choice.delta.content
+            if choice.finish_reason:
+                finish_reason = choice.finish_reason
+
+        if chunk.usage is not None:
+            total_tokens = chunk.usage.total_tokens
+
+    assert content
+    assert finish_reason == "stop"
+    assert total_tokens > 0
diff --git a/paddler_openai_client_python_test/tests/test_response_returns_output_text_and_usage.py b/paddler_openai_client_python_test/tests/test_response_returns_output_text_and_usage.py
deleted file mode 100644
index 9047c4fe4..000000000
--- a/paddler_openai_client_python_test/tests/test_response_returns_output_text_and_usage.py
+++ /dev/null
@@ -1,18 +0,0 @@
-from openai import OpenAI
-
-
-def test_response_returns_output_text_and_usage(
-    openai_client: OpenAI,
-    model: str,
-) -> None:
-    response = openai_client.responses.create(
-        model=model,
-        input="Say hi briefly.",
-        max_output_tokens=600,
-    )
-
-    assert response.object == "response"
-    assert response.status == "completed"
-    assert response.output_text
-    assert response.usage is not None
-    assert response.usage.total_tokens > 0
diff --git a/paddler_openai_client_python_test/tests/test_response_stream_reaches_completed_and_accumulates_output_text.py b/paddler_openai_client_python_test/tests/test_response_stream_reaches_completed_and_accumulates_output_text.py
deleted file mode 100644
index 65a9bde45..000000000
--- a/paddler_openai_client_python_test/tests/test_response_stream_reaches_completed_and_accumulates_output_text.py
+++ /dev/null
@@ -1,26 +0,0 @@
-from openai import OpenAI
-
-
-def test_response_stream_reaches_completed_and_accumulates_output_text(
-    openai_client: OpenAI,
-    model: str,
-) -> None:
-    stream = openai_client.responses.create(
-        model=model,
-        input="Say hi briefly.",
-        max_output_tokens=600,
-        stream=True,
-    )
-
-    event_types: list[str] = []
-    output_text = ""
-
-    for event in stream:
-        event_types.append(event.type)
-
-        if event.type == "response.output_text.delta":
-            output_text += event.delta
-
-    assert event_types[0] == "response.created"
-    assert event_types[-1] == "response.completed"
-    assert output_text
diff --git a/paddler_openai_client_python_test/tests/test_responses_return_output_text_and_usage.py b/paddler_openai_client_python_test/tests/test_responses_return_output_text_and_usage.py
new file mode 100644
index 000000000..ef3a2f15e
--- /dev/null
+++ b/paddler_openai_client_python_test/tests/test_responses_return_output_text_and_usage.py
@@ -0,0 +1,18 @@
+from openai import OpenAI
+
+
+def test_responses_return_output_text_and_usage(
+    openai_client: OpenAI,
+    model: str,
+) -> None:
+    response = openai_client.responses.create(
+        model=model,
+        input="Say hi briefly.",
+        max_output_tokens=600,
+    )
+
+    assert response.object == "response"
+    assert response.status == "completed"
+    assert response.output_text
+    assert response.usage is not None
+    assert response.usage.total_tokens > 0
diff --git a/paddler_openai_client_python_test/tests/test_responses_stream_reaches_completed_and_accumulates_output_text.py b/paddler_openai_client_python_test/tests/test_responses_stream_reaches_completed_and_accumulates_output_text.py
new file mode 100644
index 000000000..8adc74cd6
--- /dev/null
+++ b/paddler_openai_client_python_test/tests/test_responses_stream_reaches_completed_and_accumulates_output_text.py
@@ -0,0 +1,26 @@
+from openai import OpenAI
+
+
+def test_responses_stream_reaches_completed_and_accumulates_output_text(
+    openai_client: OpenAI,
+    model: str,
+) -> None:
+    stream = openai_client.responses.create(
+        model=model,
+        input="Say hi briefly.",
+        max_output_tokens=600,
+        stream=True,
+    )
+
+    event_types: list[str] = []
+    output_text = ""
+
+    for event in stream:
+        event_types.append(event.type)
+
+        if event.type == "response.output_text.delta":
+            output_text += event.delta
+
+    assert event_types[0] == "response.created"
+    assert event_types[-1] == "response.completed"
+    assert output_text
diff --git a/paddler_openai_translation/Cargo.toml b/paddler_openai_translation/Cargo.toml
new file mode 100644
index 000000000..b306287f3
--- /dev/null
+++ b/paddler_openai_translation/Cargo.toml
@@ -0,0 +1,29 @@
+[package]
+name = "paddler_openai_translation"
+autotests = false
+authors.workspace = true
+description.workspace = true
+edition.workspace = true
+homepage.workspace = true
+license.workspace = true
+repository.workspace = true
+version.workspace = true
+
+[[test]]
+name = "integration"
+path = "tests/main.rs"
+
+[dependencies]
+llama-cpp-bindings-types = { workspace = true }
+nanoid = { workspace = true }
+paddler_inference_parameters = { workspace = true }
+paddler_messaging = { workspace = true }
+serde = { workspace = true }
+serde_json = { workspace = true }
+thiserror = { workspace = true }
+
+[dev-dependencies]
+paddler_openai_response_format_validator = { workspace = true }
+
+[lints]
+workspace = true
diff --git a/paddler_balancer/src/compatibility/openai_service/arguments_to_tool_call_string.rs b/paddler_openai_translation/src/arguments_to_tool_call_string.rs
similarity index 100%
rename from paddler_balancer/src/compatibility/openai_service/arguments_to_tool_call_string.rs
rename to paddler_openai_translation/src/arguments_to_tool_call_string.rs
diff --git a/paddler_openai_translation/src/chat_completion.rs b/paddler_openai_translation/src/chat_completion.rs
new file mode 100644
index 000000000..1b9f81cdd
--- /dev/null
+++ b/paddler_openai_translation/src/chat_completion.rs
@@ -0,0 +1,112 @@
+use serde::Serialize;
+use serde::Serializer;
+
+use paddler_messaging::assistant_role::ASSISTANT_ROLE;
+use paddler_messaging::generated_tool_call::GeneratedToolCall;
+
+use crate::chat_completion_tool_call::ChatCompletionToolCall;
+use crate::chat_completion_usage::ChatCompletionUsage;
+
+struct SerializedToolCalls<'calls>(&'calls [GeneratedToolCall]);
+
+impl Serialize for SerializedToolCalls<'_> {
+    fn serialize<TSerializer>(
+        &self,
+        serializer: TSerializer,
+    ) -> Result<TSerializer::Ok, TSerializer::Error>
+    where
+        TSerializer: Serializer,
+    {
+        serializer.collect_seq(self.0.iter().map(ChatCompletionToolCall))
+    }
+}
+
+#[derive(Serialize)]
+#[serde(untagged)]
+enum SerializedMessage<'completion> {
+    Content {
+        role: &'static str,
+        content: &'completion str,
+        refusal: (),
+        annotations: [u8; 0],
+    },
+    ToolCalls {
+        role: &'static str,
+        content: Option<&'completion str>,
+        refusal: (),
+        annotations: [u8; 0],
+        tool_calls: SerializedToolCalls<'completion>,
+    },
+}
+
+#[derive(Serialize)]
+struct SerializedChoice<'completion> {
+    index: u8,
+    message: SerializedMessage<'completion>,
+    logprobs: (),
+    finish_reason: &'static str,
+}
+
+#[derive(Serialize)]
+struct SerializedCompletion<'completion> {
+    id: &'completion str,
+    object: &'static str,
+    created: u64,
+    model: &'completion str,
+    choices: [SerializedChoice<'completion>; 1],
+    usage: ChatCompletionUsage,
+    service_tier: &'static str,
+}
+
+pub struct ChatCompletion {
+    pub content: String,
+    pub created: u64,
+    pub finish_reason: &'static str,
+    pub id: String,
+    pub model: String,
+    pub tool_calls: Vec<GeneratedToolCall>,
+    pub usage: ChatCompletionUsage,
+}
+
+impl Serialize for ChatCompletion {
+    fn serialize<TSerializer>(
+        &self,
+        serializer: TSerializer,
+    ) -> Result<TSerializer::Ok, TSerializer::Error>
+    where
+        TSerializer: Serializer,
+    {
+        let message = if self.tool_calls.is_empty() {
+            SerializedMessage::Content {
+                role: ASSISTANT_ROLE,
+                content: &self.content,
+                refusal: (),
+                annotations: [],
+            }
+        } else {
+            SerializedMessage::ToolCalls {
+                role: ASSISTANT_ROLE,
+                content: (!self.content.is_empty()).then_some(self.content.as_str()),
+                refusal: (),
+                annotations: [],
+                tool_calls: SerializedToolCalls(&self.tool_calls),
+            }
+        };
+
+        SerializedCompletion {
+            id: &self.id,
+            object: "chat.completion",
+            created: self.created,
+            model: &self.model,
+            choices: [SerializedChoice {
+                index: 0,
+                message,
+                logprobs: (),
+                finish_reason: self.finish_reason,
+            }],
+            usage: self.usage,
+            service_tier: "default",
+        }
+        .serialize(serializer)
+    }
+}
diff --git a/paddler_openai_translation/src/chat_completion_chunk.rs b/paddler_openai_translation/src/chat_completion_chunk.rs
new file mode 100644
index 000000000..466076b76
--- /dev/null
+++ b/paddler_openai_translation/src/chat_completion_chunk.rs
@@ -0,0 +1,71 @@
+use serde::Serialize;
+use serde::Serializer;
+
+use crate::chat_completion_chunk_choice::ChatCompletionChunkChoice;
+use crate::chat_completion_chunk_payload::ChatCompletionChunkPayload;
+use crate::chat_completion_usage::ChatCompletionUsage;
+
+#[derive(Serialize)]
+struct ChunkEnvelope<'chunk, TBody> {
+    id: &'chunk str,
+    object: &'static str,
+    created: u64,
+    model: &'chunk str,
+    system_fingerprint: &'chunk str,
+    #[serde(flatten)]
+    body: TBody,
+}
+
+#[derive(Serialize)]
+struct ChoiceBody<'body> {
+    choices: [&'body ChatCompletionChunkChoice; 1],
+}
+
+#[derive(Serialize)]
+struct UsageBody {
+    choices: [u8; 0],
+    usage: ChatCompletionUsage,
+}
+
+pub struct ChatCompletionChunk {
+    pub created: u64,
+    pub id: String,
+    pub model: String,
+    pub payload: ChatCompletionChunkPayload,
+    pub system_fingerprint: String,
+}
+
+impl ChatCompletionChunk {
+    fn envelope<TBody>(&self, body: TBody) -> ChunkEnvelope<'_, TBody> {
+        ChunkEnvelope {
+            id: &self.id,
+            object: "chat.completion.chunk",
+            created: self.created,
+            model: &self.model,
+            system_fingerprint: &self.system_fingerprint,
+            body,
+        }
+    }
+}
+
+impl Serialize for ChatCompletionChunk {
+    fn serialize<TSerializer>(
+        &self,
+        serializer: TSerializer,
+    ) -> Result<TSerializer::Ok, TSerializer::Error>
+    where
+        TSerializer: Serializer,
+    {
+        match &self.payload {
+            ChatCompletionChunkPayload::Choice(choice) => self
+                .envelope(ChoiceBody { choices: [choice] })
+                .serialize(serializer),
+            ChatCompletionChunkPayload::Usage(usage) => self
+                .envelope(UsageBody {
+                    choices: [],
+                    usage: ChatCompletionUsage(*usage),
+                })
+                .serialize(serializer),
+        }
+    }
+}
diff --git a/paddler_openai_translation/src/chat_completion_chunk_choice.rs b/paddler_openai_translation/src/chat_completion_chunk_choice.rs
new file mode 100644
index 000000000..083b86082
--- /dev/null
+++ b/paddler_openai_translation/src/chat_completion_chunk_choice.rs
@@ -0,0 +1,97 @@
+use serde::Serialize;
+use serde::Serializer;
+
+use paddler_messaging::assistant_role::ASSISTANT_ROLE;
+use paddler_messaging::generated_tool_call::GeneratedToolCall;
+
+use crate::chat_completion_tool_call::ChatCompletionToolCall;
+
+#[derive(Serialize)]
+struct ToolCallDelta<'call> {
+    index: usize,
+    #[serde(flatten)]
+    tool_call: ChatCompletionToolCall<'call>,
+}
+
+struct ToolCallsDelta<'calls>(&'calls [GeneratedToolCall]);
+
+impl Serialize for ToolCallsDelta<'_> {
+    fn serialize<TSerializer>(
+        &self,
+        serializer: TSerializer,
+    ) -> Result<TSerializer::Ok, TSerializer::Error>
+    where
+        TSerializer: Serializer,
+    {
+        serializer.collect_seq(self.0.iter().map(|call| ToolCallDelta {
+            index: call.index,
+            tool_call: ChatCompletionToolCall(call),
+        }))
+    }
+}
+
+#[derive(Serialize)]
+#[serde(untagged)]
+enum ChoiceDelta<'choice> {
+    Content {
+        role: &'static str,
+        content: &'choice str,
+    },
+    Finish {},
+    ToolCalls {
+        role: &'static str,
+        tool_calls: ToolCallsDelta<'choice>,
+    },
+}
+
+#[derive(Serialize)]
+struct SerializedChoice<'choice> {
+    index: u8,
+    delta: ChoiceDelta<'choice>,
+    logprobs: (),
+    finish_reason: Option<&'static str>,
+}
+
+pub enum ChatCompletionChunkChoice {
+    Content(String),
+    Finish(&'static str),
+    ToolCalls(Vec<GeneratedToolCall>),
+}
+
+impl Serialize for ChatCompletionChunkChoice {
+    fn serialize<TSerializer>(
+        &self,
+        serializer: TSerializer,
+    ) -> Result<TSerializer::Ok, TSerializer::Error>
+    where
+        TSerializer: Serializer,
+    {
+        match self {
+            Self::Content(content) => SerializedChoice {
+                index: 0,
+                delta: ChoiceDelta::Content {
+                    role: ASSISTANT_ROLE,
+                    content,
+                },
+                logprobs: (),
+                finish_reason: None,
+            },
+            Self::Finish(finish_reason) => SerializedChoice {
+                index: 0,
+                delta: ChoiceDelta::Finish {},
+                logprobs: (),
+                finish_reason: Some(*finish_reason),
+            },
+            Self::ToolCalls(parsed_calls) => SerializedChoice {
+                index: 0,
+                delta: ChoiceDelta::ToolCalls {
+                    role: ASSISTANT_ROLE,
+                    tool_calls: ToolCallsDelta(parsed_calls),
+                },
+                logprobs: (),
+                finish_reason: None,
+            },
+        }
+        .serialize(serializer)
+    }
+}
diff --git a/paddler_openai_translation/src/chat_completion_chunk_payload.rs b/paddler_openai_translation/src/chat_completion_chunk_payload.rs
new file mode 100644
index 000000000..7fd425c50
--- /dev/null
+++ b/paddler_openai_translation/src/chat_completion_chunk_payload.rs
@@ -0,0 +1,8 @@
+use llama_cpp_bindings_types::TokenUsage;
+
+use crate::chat_completion_chunk_choice::ChatCompletionChunkChoice;
+
+pub enum ChatCompletionChunkPayload {
+    Choice(ChatCompletionChunkChoice),
+    Usage(TokenUsage),
+}
diff --git a/paddler_openai_translation/src/chat_completion_delivery.rs b/paddler_openai_translation/src/chat_completion_delivery.rs
new file mode 100644
index 000000000..cc1844a00
--- /dev/null
+++ b/paddler_openai_translation/src/chat_completion_delivery.rs
@@ -0,0 +1,7 @@
+use crate::chat_completion_header::ChatCompletionHeader;
+use crate::chat_completion_stream::ChatCompletionStream;
+
+pub enum ChatCompletionDelivery {
+    Buffered(ChatCompletionHeader),
+    Streamed(ChatCompletionStream),
+}
diff --git a/paddler_balancer/src/compatibility/openai_service/chat_completion_finish_reason.rs b/paddler_openai_translation/src/chat_completion_finish_reason.rs
similarity index 100%
rename from paddler_balancer/src/compatibility/openai_service/chat_completion_finish_reason.rs
rename to paddler_openai_translation/src/chat_completion_finish_reason.rs
diff --git a/paddler_openai_translation/src/chat_completion_header.rs b/paddler_openai_translation/src/chat_completion_header.rs
new file mode 100644
index 000000000..7563b5cf2
--- /dev/null
+++ b/paddler_openai_translation/src/chat_completion_header.rs
@@ -0,0 +1,147 @@
+use paddler_messaging::generation_summary::GenerationSummary;
+
+use crate::chat_completion::ChatCompletion;
+use crate::chat_completion_finish_reason::chat_completion_finish_reason;
+use crate::chat_completion_usage::ChatCompletionUsage;
+use crate::completes_generation::CompletesGeneration;
+use crate::generated_output::GeneratedOutput;
+
+#[derive(Clone)]
+pub struct ChatCompletionHeader {
+    pub created: u64,
+    pub model: String,
+}
+
+impl CompletesGeneration for ChatCompletionHeader {
+    type Completion = ChatCompletion;
+
+    fn complete(
+        &self,
+        request_id: &str,
+        GeneratedOutput {
+            content,
+            tool_calls,
+            ..
+        }: GeneratedOutput,
+        GenerationSummary { finish, usage }: &GenerationSummary,
+    ) -> ChatCompletion {
+        ChatCompletion {
+            content,
+            created: self.created,
+            finish_reason: chat_completion_finish_reason(*finish, !tool_calls.is_empty()),
+            id: request_id.to_owned(),
+            model: self.model.clone(),
+            tool_calls,
+            usage: ChatCompletionUsage(*usage),
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use llama_cpp_bindings_types::TokenUsage;
+    use llama_cpp_bindings_types::ToolCallArguments;
+    use serde_json::json;
+    use serde_json::to_string;
+
+    use paddler_messaging::generated_tool_call::GeneratedToolCall;
+    use paddler_messaging::generation_finish::GenerationFinish;
+    use paddler_messaging::generation_summary::GenerationSummary;
+
+    use super::ChatCompletionHeader;
+    use crate::completes_generation::CompletesGeneration as _;
+    use crate::generated_output::GeneratedOutput;
+    use crate::generated_output_part::GeneratedOutputPart;
+
+    fn completion_of(
+        generated_output_parts: Vec<GeneratedOutputPart>,
+        prompt_tokens: u64,
+        content_tokens: u64,
+        reasoning_tokens: u64,
+    ) -> String {
+        let mut generated_output = GeneratedOutput::default();
+
+        for generated_output_part in generated_output_parts {
+            generated_output.append(generated_output_part);
+        }
+
+        to_string(
+            &ChatCompletionHeader {
+                created: 0,
+                model: "test-model".to_owned(),
+            }
+            .complete(
+                "test-request",
+                generated_output,
+                &GenerationSummary {
+                    finish: GenerationFinish::EndOfGeneration,
+                    usage: TokenUsage {
+                        prompt_tokens,
+                        content_tokens,
+                        reasoning_tokens,
+                        ..TokenUsage::default()
+                    },
+                },
+            ),
+        )
+        .unwrap()
+    }
+
+    #[test]
+    fn aggregates_content_into_one_message() {
+        assert_eq!(
+            completion_of(
+                vec![
+                    GeneratedOutputPart::Content("hel".to_owned()),
+                    GeneratedOutputPart::Content("lo".to_owned()),
+                ],
+                4,
+                2,
+                0,
+            ),
+            r#"{"id":"test-request","object":"chat.completion","created":0,"model":"test-model","choices":[{"index":0,"message":{"role":"assistant","content":"hello","refusal":null,"annotations":[]},"logprobs":null,"finish_reason":"stop"}],"usage":{"prompt_tokens":4,"completion_tokens":2,"total_tokens":6,"prompt_tokens_details":{"cached_tokens":0,"audio_tokens":0},"completion_tokens_details":{"reasoning_tokens":0}},"service_tier":"default"}"#
+        );
+    }
+
+    #[test]
+    fn drops_reasoning_but_keeps_its_token_count() {
+        assert_eq!(
+            completion_of(
+                vec![
+                    GeneratedOutputPart::Reasoning("think".to_owned()),
+                    GeneratedOutputPart::Content("answer".to_owned()),
+                ],
+                3,
+                1,
+                1,
+            ),
+            r#"{"id":"test-request","object":"chat.completion","created":0,"model":"test-model","choices":[{"index":0,"message":{"role":"assistant","content":"answer","refusal":null,"annotations":[]},"logprobs":null,"finish_reason":"stop"}],"usage":{"prompt_tokens":3,"completion_tokens":2,"total_tokens":5,"prompt_tokens_details":{"cached_tokens":0,"audio_tokens":0},"completion_tokens_details":{"reasoning_tokens":1}},"service_tier":"default"}"#
+        );
+    }
+
+    #[test]
+    fn reports_parsed_tool_calls_with_their_raw_arguments_and_the_tool_calls_finish_reason() {
+        assert_eq!(
+            completion_of(
+                vec![GeneratedOutputPart::ToolCalls(vec![
+                    GeneratedToolCall {
+                        arguments: ToolCallArguments::ValidJson(json!({ "location": "Paris" })),
+                        id: "call_x".to_owned(),
+                        index: 0,
+                        name: "get_weather".to_owned(),
+                    },
+                    GeneratedToolCall {
+                        arguments: ToolCallArguments::InvalidJson("{not valid json".to_owned()),
+                        id: "call_invalid".to_owned(),
+                        index: 1,
+                        name: "broken_tool".to_owned(),
+                    },
+                ])],
+                4,
+                0,
+                0,
+            ),
+            r#"{"id":"test-request","object":"chat.completion","created":0,"model":"test-model","choices":[{"index":0,"message":{"role":"assistant","content":null,"refusal":null,"annotations":[],"tool_calls":[{"id":"call_x","type":"function","function":{"name":"get_weather","arguments":"{\"location\":\"Paris\"}"}},{"id":"call_invalid","type":"function","function":{"name":"broken_tool","arguments":"{not valid json"}}]},"logprobs":null,"finish_reason":"tool_calls"}],"usage":{"prompt_tokens":4,"completion_tokens":0,"total_tokens":4,"prompt_tokens_details":{"cached_tokens":0,"audio_tokens":0},"completion_tokens_details":{"reasoning_tokens":0}},"service_tier":"default"}"#
+        );
+    }
+}
diff --git a/paddler_openai_translation/src/chat_completion_message.rs b/paddler_openai_translation/src/chat_completion_message.rs
new file mode 100644
index 000000000..86c11a41b
--- /dev/null
+++ b/paddler_openai_translation/src/chat_completion_message.rs
@@ -0,0 +1,333 @@
+use serde::Deserialize;
+
+use paddler_messaging::assistant_role::ASSISTANT_ROLE;
+use paddler_messaging::conversation_message::ConversationMessage;
+use paddler_messaging::conversation_message_content::ConversationMessageContent;
+use paddler_messaging::system_role::SYSTEM_ROLE;
+use paddler_messaging::user_role::USER_ROLE;
+
+use crate::chat_completion_message_role::ChatCompletionMessageRole;
+use crate::chat_completion_requested_tool_call::ChatCompletionRequestedToolCall;
+use crate::openai_translation_error::OpenAITranslationError;
+
+#[derive(Deserialize)]
+pub struct ChatCompletionMessage {
+    pub content: Option<ConversationMessageContent>,
+    pub reasoning_content: Option<String>,
+    pub role: ChatCompletionMessageRole,
+    pub tool_call_id: Option<String>,
+    pub tool_calls: Option<Vec<ChatCompletionRequestedToolCall>>,
+}
+
+impl ChatCompletionMessage {
+    pub fn append_to(
+        self,
+        messages: &mut Vec<ConversationMessage>,
+    ) -> Result<(), OpenAITranslationError> {
+        let Self {
+            content,
+            reasoning_content,
+            role,
+            tool_call_id,
+            tool_calls,
+        } = self;
+        let assistant_reasoning = reasoning_content
+            .map(|reasoning_content| match role {
+                ChatCompletionMessageRole::Assistant => {
+                    Ok(ConversationMessage::AssistantReasoning { reasoning_content })
+                }
+                ChatCompletionMessageRole::Developer
+                | ChatCompletionMessageRole::System
+                | ChatCompletionMessageRole::Tool
+                | ChatCompletionMessageRole::User => {
+                    Err(OpenAITranslationError::ReasoningOnNonAssistantMessage { role })
+                }
+            })
+            .transpose()?;
+        let reply = match tool_calls {
+            Some(tool_calls) if !tool_calls.is_empty() => {
+                if role != ChatCompletionMessageRole::Assistant {
+                    return Err(OpenAITranslationError::ToolCallsOnNonAssistantMessage { role });
+                }
+
+                ConversationMessage::AssistantToolCalls {
+                    content: content
+                        .unwrap_or_else(|| ConversationMessageContent::Text(String::new())),
+                    tool_calls: tool_calls
+                        .into_iter()
+                        .map(ChatCompletionRequestedToolCall::into_conversation_tool_call)
+                        .collect::<Result<_, _>>()?,
+                }
+            }
+            Some(_) | None => {
+                let Some(content) = content else {
+                    return Err(OpenAITranslationError::MessageWithoutContent { role });
+                };
+
+                match role {
+                    ChatCompletionMessageRole::Assistant => ConversationMessage::Message {
+                        content,
+                        role: ASSISTANT_ROLE.to_owned(),
+                    },
+                    ChatCompletionMessageRole::Developer | ChatCompletionMessageRole::System => {
+                        ConversationMessage::Message {
+                            content,
+                            role: SYSTEM_ROLE.to_owned(),
+                        }
+                    }
+                    ChatCompletionMessageRole::Tool => ConversationMessage::ToolResult {
+                        content,
+                        tool_call_id: tool_call_id
+                            .ok_or(OpenAITranslationError::ToolResultWithoutToolCallId)?,
+                    },
+                    ChatCompletionMessageRole::User => ConversationMessage::Message {
+                        content,
+                        role: USER_ROLE.to_owned(),
+                    },
+                }
+            }
+        };
+
+        messages.extend(assistant_reasoning);
+        messages.push(reply);
+
+        Ok(())
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use serde_json::Value;
+    use serde_json::from_value;
+    use serde_json::json;
+
+    use paddler_messaging::conversation_message::ConversationMessage;
+    use paddler_messaging::conversation_message_content::ConversationMessageContent;
+    use paddler_messaging::conversation_message_content_part::ConversationMessageContentPart;
+    use paddler_messaging::conversation_tool_call::ConversationToolCall;
+    use paddler_messaging::image_url::ImageUrl;
+
+    use super::ChatCompletionMessage;
+    use crate::chat_completion_message_role::ChatCompletionMessageRole;
+    use crate::openai_translation_error::OpenAITranslationError;
+
+    fn appended(message: Value) -> Result<Vec<ConversationMessage>, OpenAITranslationError> {
+        let mut messages = Vec::new();
+
+        from_value::<ChatCompletionMessage>(message)
+            .unwrap()
+            .append_to(&mut messages)
+            .map(|()| messages)
+    }
+
+    fn converted(message: Value) -> Result<ConversationMessage, OpenAITranslationError> {
+        appended(message).map(|mut messages| messages.remove(0))
+    }
+
+    fn read_call(id: &str, path: &str) -> ConversationToolCall {
+        ConversationToolCall {
+            arguments: json!({ "path": path }).as_object().unwrap().clone(),
+            id: id.to_owned(),
+            name: "read".to_owned(),
+        }
+    }
+
+    #[test]
+    fn a_user_message_with_parts_keeps_its_role_and_parts() {
+        assert_eq!(
+            converted(json!({
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": "OCR this"},
+                    {"type": "image_url", "image_url": {"url": "data:image/png;base64,abc"}}
+                ]
+            }))
+            .unwrap(),
+            ConversationMessage::Message {
+                content: ConversationMessageContent::Parts(vec![
+                    ConversationMessageContentPart::Text {
+                        text: "OCR this".to_owned(),
+                    },
+                    ConversationMessageContentPart::ImageUrl {
+                        image_url: ImageUrl {
+                            url: "data:image/png;base64,abc".to_owned(),
+                        },
+                    },
+                ]),
+                role: "user".to_owned(),
+            }
+        );
+    }
+
+    #[test]
+    fn parallel_tool_calls_without_content_become_one_assistant_turn_with_empty_text() {
+        assert_eq!(
+            converted(json!({
+                "role": "assistant",
+                "content": null,
+                "tool_calls": [
+                    {"id": "call_0", "type": "function", "function": {"name": "read", "arguments": "{\"path\":\"a.txt\"}"}},
+                    {"id": "call_1", "type": "function", "function": {"name": "read", "arguments": "{\"path\":\"b.txt\"}"}}
+                ]
+            }))
+            .unwrap(),
+            ConversationMessage::AssistantToolCalls {
+                content: ConversationMessageContent::Text(String::new()),
+                tool_calls: vec![read_call("call_0", "a.txt"), read_call("call_1", "b.txt")],
+            }
+        );
+    }
+
+    #[test]
+    fn assistant_text_alongside_tool_calls_stays_in_the_same_turn() {
+        assert_eq!(
+            converted(json!({
+                "role": "assistant",
+                "content": "Let me read it.",
+                "tool_calls": [
+                    {"id": "call_0", "type": "function", "function": {"name": "read", "arguments": "{\"path\":\"a.txt\"}"}}
+                ]
+            }))
+            .unwrap(),
+            ConversationMessage::AssistantToolCalls {
+                content: ConversationMessageContent::Text("Let me read it.".to_owned()),
+                tool_calls: vec![read_call("call_0", "a.txt")],
+            }
+        );
+    }
+
+    #[test]
+    fn a_tool_message_keeps_its_tool_call_id() {
+        assert_eq!(
+            converted(
+                json!({"role": "tool", "tool_call_id": "call_0", "content": "file contents"})
+            )
+            .unwrap(),
+            ConversationMessage::ToolResult {
+                content: ConversationMessageContent::Text("file contents".to_owned()),
+                tool_call_id: "call_0".to_owned(),
+            }
+        );
+    }
+
+    #[test]
+    fn a_message_without_content_or_tool_calls_is_refused_naming_its_role() {
+        assert_eq!(
+            converted(json!({"role": "assistant", "content": null}))
+                .err()
+                .map(|translation_error| translation_error.to_string()),
+            Some(
+                OpenAITranslationError::MessageWithoutContent {
+                    role: ChatCompletionMessageRole::Assistant
+                }
+                .to_string()
+            )
+        );
+    }
+
+    #[test]
+    fn tool_calls_on_a_message_that_is_not_an_assistant_message_are_refused() {
+        assert_eq!(
+            converted(json!({
+                "role": "user",
+                "content": "hi",
+                "tool_calls": [
+                    {"id": "call_0", "type": "function", "function": {"name": "read", "arguments": "{}"}}
+                ]
+            }))
+            .err()
+            .map(|translation_error| translation_error.to_string()),
+            Some(
+                OpenAITranslationError::ToolCallsOnNonAssistantMessage {
+                    role: ChatCompletionMessageRole::User
+                }
+                .to_string()
+            )
+        );
+    }
+
+    #[test]
+    fn a_tool_message_without_a_tool_call_id_is_refused() {
+        assert_eq!(
+            converted(json!({"role": "tool", "content": "file contents"}))
+                .err()
+                .as_ref()
+                .map(ToString::to_string),
+            Some(OpenAITranslationError::ToolResultWithoutToolCallId.to_string())
+        );
+    }
+
+    #[test]
+    fn tool_call_arguments_that_are_not_a_json_object_are_refused() {
+        assert!(matches!(
+            converted(json!({
+                "role": "assistant",
+                "content": null,
+                "tool_calls": [
+                    {"id": "call_0", "type": "function", "function": {"name": "read", "arguments": "[\"a.txt\"]"}}
+                ]
+            })),
+            Err(OpenAITranslationError::ToolCallArgumentsAreNotAJsonObject { tool_call_id, .. })
+                if tool_call_id == "call_0"
+        ));
+    }
+
+    #[test]
+    fn a_developer_message_instructs_the_model_as_a_system_message() {
+        assert_eq!(
+            converted(json!({"role": "developer", "content": "Answer in French"})).unwrap(),
+            ConversationMessage::Message {
+                content: ConversationMessageContent::Text("Answer in French".to_owned()),
+                role: "system".to_owned(),
+            }
+        );
+    }
+
+    #[test]
+    fn a_message_of_a_role_openai_does_not_define_is_refused() {
+        assert!(
+            from_value::<ChatCompletionMessage>(json!({"role": "ipython", "content": "1"}))
+                .is_err()
+        );
+    }
+
+    #[test]
+    fn the_reasoning_of_an_assistant_message_precedes_its_reply() {
+        assert_eq!(
+            appended(json!({
+                "role": "assistant",
+                "content": "It is sunny.",
+                "reasoning_content": "The forecast said sun."
+            }))
+            .unwrap(),
+            vec![
+                ConversationMessage::AssistantReasoning {
+                    reasoning_content: "The forecast said sun.".to_owned(),
+                },
+                ConversationMessage::Message {
+                    content: ConversationMessageContent::Text("It is sunny.".to_owned()),
+                    role: "assistant".to_owned(),
+                },
+            ]
+        );
+    }
+
+    #[test]
+    fn reasoning_on_a_message_that_is_not_an_assistant_message_is_refused() {
+        assert_eq!(
+            appended(json!({
+                "role": "user",
+                "content": "hi",
+                "reasoning_content": "I am thinking"
+            }))
+            .err()
+            .map(|translation_error| translation_error.to_string()),
+            Some(
+                OpenAITranslationError::ReasoningOnNonAssistantMessage {
+                    role: ChatCompletionMessageRole::User
+                }
+                .to_string()
+            )
+        );
+    }
+}
diff --git a/paddler_openai_translation/src/chat_completion_message_role.rs b/paddler_openai_translation/src/chat_completion_message_role.rs
new file mode 100644
index 000000000..5b948bf6e
--- /dev/null
+++ b/paddler_openai_translation/src/chat_completion_message_role.rs
@@ -0,0 +1,51 @@
+use std::fmt::Display;
+use std::fmt::Formatter;
+use std::fmt::Result as FmtResult;
+
+use serde::Deserialize;
+
+#[derive(Clone, Copy, Debug, Deserialize, Eq, PartialEq)]
+#[serde(rename_all = "lowercase")]
+pub enum ChatCompletionMessageRole {
+    Assistant,
+    Developer,
+    System,
+    Tool,
+    User,
+}
+
+impl Display for ChatCompletionMessageRole {
+    fn fmt(&self, formatter: &mut Formatter<'_>) -> FmtResult {
+        formatter.write_str(match self {
+            Self::Assistant => "assistant",
+            Self::Developer => "developer",
+            Self::System => "system",
+            Self::Tool => "tool",
+            Self::User => "user",
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use serde_json::Value;
+    use serde_json::from_value;
+
+    use super::ChatCompletionMessageRole;
+
+    #[test]
+    fn names_every_role_the_way_openai_requests_spell_it() {
+        for role in [
+            ChatCompletionMessageRole::Assistant,
+            ChatCompletionMessageRole::Developer,
+            ChatCompletionMessageRole::System,
+            ChatCompletionMessageRole::Tool,
+            ChatCompletionMessageRole::User,
+        ] {
+            assert_eq!(
+                from_value::<ChatCompletionMessageRole>(Value::String(role.to_string())).unwrap(),
+                role
+            );
+        }
+    }
+}
diff --git a/paddler_openai_translation/src/chat_completion_request.rs b/paddler_openai_translation/src/chat_completion_request.rs
new file mode 100644
index 000000000..58043ef03
--- /dev/null
+++ b/paddler_openai_translation/src/chat_completion_request.rs
@@ -0,0 +1,341 @@
+use std::num::NonZeroU32;
+use std::time::SystemTime;
+
+use nanoid::nanoid;
+use serde::Deserialize;
+
+use paddler_messaging::conversation_history::ConversationHistory;
+use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
+use paddler_messaging::validates::Validates;
+
+use crate::chat_completion_delivery::ChatCompletionDelivery;
+use crate::chat_completion_header::ChatCompletionHeader;
+use crate::chat_completion_message::ChatCompletionMessage;
+use crate::chat_completion_stream::ChatCompletionStream;
+use crate::chat_completion_stream_options::ChatCompletionStreamOptions;
+use crate::chat_completion_stream_params::ChatCompletionStreamParams;
+use crate::chat_completion_tool::ChatCompletionTool;
+use crate::default_max_tokens::DEFAULT_MAX_TOKENS;
+use crate::openai_translation_error::OpenAITranslationError;
+use crate::reasoning_effort::ReasoningEffort;
+use crate::timestamp_from::timestamp_from;
+use crate::translated_chat_completion_request::TranslatedChatCompletionRequest;
+
+#[derive(Deserialize)]
+pub struct ChatCompletionRequest {
+    pub max_completion_tokens: Option<NonZeroU32>,
+    pub max_tokens: Option<NonZeroU32>,
+    pub messages: Vec<ChatCompletionMessage>,
+    pub model: String,
+    pub reasoning_effort: Option<ReasoningEffort>,
+    pub stream: Option<bool>,
+    pub stream_options: Option<ChatCompletionStreamOptions>,
+    #[serde(default)]
+    pub tools: Vec<ChatCompletionTool>,
+}
+
+impl ChatCompletionRequest {
+    #[must_use]
+    pub fn enables_thinking(&self) -> bool {
+        self.reasoning_effort
+            .is_none_or(ReasoningEffort::enables_thinking)
+    }
+
+    #[must_use]
+    pub fn requested_max_tokens(&self) -> Option<NonZeroU32> {
+        self.max_completion_tokens.or(self.max_tokens)
+    }
+
+    pub fn translate(
+        self,
+        now: SystemTime,
+    ) -> Result<TranslatedChatCompletionRequest, OpenAITranslationError> {
+        let enable_thinking = self.enables_thinking();
+        let max_tokens = self.requested_max_tokens().unwrap_or(DEFAULT_MAX_TOKENS);
+        let Self {
+            messages,
+            model,
+            stream,
+            stream_options,
+            tools,
+            ..
+        } = self;
+        let validated_tools = tools
+            .into_iter()
+            .map(ChatCompletionTool::into_tool)
+            .map(Validates::validate)
+            .collect::<Result<Vec<_>, _>>()?;
+        let mut conversation_messages = Vec::with_capacity(messages.len());
+
+        for message in messages {
+            message.append_to(&mut conversation_messages)?;
+        }
+
+        let header = ChatCompletionHeader {
+            created: timestamp_from(now)?,
+            model,
+        };
+
+        Ok(TranslatedChatCompletionRequest {
+            conversation_history_params: ContinueFromConversationHistoryParams {
+                add_generation_prompt: true,
+                conversation_history: ConversationHistory::new(conversation_messages),
+                enable_thinking,
+                grammar: None,
+                max_tokens,
+                parse_tool_calls: !validated_tools.is_empty(),
+                tools: validated_tools,
+            },
+            delivery: if stream.unwrap_or(false) {
+                ChatCompletionDelivery::Streamed(ChatCompletionStream::new(
+                    ChatCompletionStreamParams {
+                        header,
+                        include_usage: stream_options.is_some_and(
+                            |ChatCompletionStreamOptions { include_usage }| include_usage,
+                        ),
+                        system_fingerprint: nanoid!(),
+                    },
+                ))
+            } else {
+                ChatCompletionDelivery::Buffered(header)
+            },
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::num::NonZeroU32;
+    use std::time::UNIX_EPOCH;
+
+    use serde_json::Value;
+    use serde_json::from_value;
+    use serde_json::json;
+
+    use paddler_messaging::conversation_message::ConversationMessage;
+    use paddler_messaging::conversation_message_content::ConversationMessageContent;
+    use paddler_messaging::conversation_message_content_part::ConversationMessageContentPart;
+    use paddler_messaging::image_url::ImageUrl;
+
+    use super::ChatCompletionRequest;
+    use crate::chat_completion_message_role::ChatCompletionMessageRole;
+    use crate::openai_translation_error::OpenAITranslationError;
+
+    fn conversation_messages_of(request: ChatCompletionRequest) -> Vec<ConversationMessage> {
+        let mut conversation_messages = Vec::new();
+
+        for message in request.messages {
+            message.append_to(&mut conversation_messages).unwrap();
+        }
+
+        conversation_messages
+    }
+
+    fn request_with_token_limits(token_limits: &Value) -> ChatCompletionRequest {
+        let mut input = json!({
+            "model": "test-model",
+            "messages": [{"role": "user", "content": "hello"}]
+        });
+
+        input
+            .as_object_mut()
+            .unwrap()
+            .extend(token_limits.as_object().unwrap().clone());
+
+        from_value(input).unwrap()
+    }
+
+    #[test]
+    fn requested_max_tokens_honors_max_tokens() {
+        let params = request_with_token_limits(&json!({"max_tokens": 7}));
+
+        assert_eq!(params.requested_max_tokens(), NonZeroU32::new(7));
+    }
+
+    #[test]
+    fn requested_max_tokens_prefers_max_completion_tokens() {
+        let params =
+            request_with_token_limits(&json!({"max_tokens": 7, "max_completion_tokens": 3}));
+
+        assert_eq!(params.requested_max_tokens(), NonZeroU32::new(3));
+    }
+
+    #[test]
+    fn requested_max_tokens_is_absent_without_limits() {
+        let params = request_with_token_limits(&json!({}));
+
+        assert_eq!(params.requested_max_tokens(), None);
+    }
+
+    #[test]
+    fn deserialize_text_only_request() {
+        let input = json!({
+            "model": "test-model",
+            "messages": [
+                {"role": "user", "content": "hello"}
+            ]
+        });
+
+        let params: ChatCompletionRequest = from_value(input).unwrap();
+
+        assert_eq!(params.model, "test-model");
+        assert_eq!(
+            conversation_messages_of(params),
+            vec![ConversationMessage::Message {
+                content: ConversationMessageContent::Text("hello".to_owned()),
+                role: "user".to_owned(),
+            }]
+        );
+    }
+
+    #[test]
+    fn a_request_with_a_message_it_cannot_translate_is_refused() {
+        let params: ChatCompletionRequest = from_value(json!({
+            "model": "test-model",
+            "messages": [{"role": "user", "content": null}]
+        }))
+        .unwrap();
+
+        assert_eq!(
+            params
+                .translate(UNIX_EPOCH)
+                .err()
+                .map(|translation_error| translation_error.to_string()),
+            Some(
+                OpenAITranslationError::MessageWithoutContent {
+                    role: ChatCompletionMessageRole::User
+                }
+                .to_string()
+            )
+        );
+    }
+
+    #[test]
+    fn deserialize_request_with_stream_options_include_usage_true() {
+        let input = json!({
+            "model": "test-model",
+            "messages": [{"role": "user", "content": "hi"}],
+            "stream": true,
+            "stream_options": {"include_usage": true}
+        });
+
+        let params: ChatCompletionRequest = from_value(input).unwrap();
+
+        let stream_options = params.stream_options.unwrap();
+
+        assert!(stream_options.include_usage);
+    }
+
+    #[test]
+    fn deserialize_request_without_stream_options_defaults_to_none() {
+        let input = json!({
+            "model": "test-model",
+            "messages": [{"role": "user", "content": "hi"}],
+            "stream": true
+        });
+
+        let params: ChatCompletionRequest = from_value(input).unwrap();
+
+        assert!(params.stream_options.is_none());
+    }
+
+    #[test]
+    fn deserialize_multimodal_request_with_image() {
+        let input = json!({
+            "model": "vision-model",
+            "messages": [
+                {
+                    "role": "user",
+                    "content": [
+                        {"type": "text", "text": "describe this image"},
+                        {"type": "image_url", "image_url": {"url": "data:image/jpeg;base64,/9j/4AAQ"}}
+                    ]
+                }
+            ]
+        });
+
+        assert_eq!(
+            conversation_messages_of(from_value::<ChatCompletionRequest>(input).unwrap()),
+            vec![ConversationMessage::Message {
+                content: ConversationMessageContent::Parts(vec![
+                    ConversationMessageContentPart::Text {
+                        text: "describe this image".to_owned(),
+                    },
+                    ConversationMessageContentPart::ImageUrl {
+                        image_url: ImageUrl {
+                            url: "data:image/jpeg;base64,/9j/4AAQ".to_owned(),
+                        },
+                    },
+                ]),
+                role: "user".to_owned(),
+            }]
+        );
+    }
+
+    #[test]
+    fn deserialize_multi_turn_conversation() {
+        let input = json!({
+            "model": "test-model",
+            "messages": [
+                {"role": "system", "content": "You are a helpful assistant."},
+                {"role": "user", "content": "What is 2+2?"},
+                {"role": "assistant", "content": "4"},
+                {"role": "user", "content": "And 3+3?"}
+            ]
+        });
+
+        let params: ChatCompletionRequest = from_value(input).unwrap();
+
+        assert_eq!(params.messages.len(), 4);
+    }
+
+    #[test]
+    fn deserialize_request_with_opencode_style_tools() {
+        let input = json!({
+            "model": "test-model",
+            "messages": [{"role": "user", "content": "hi"}],
+            "tools": [
+                {
+                    "type": "function",
+                    "function": {
+                        "name": "glob",
+                        "description": "Fast file pattern matching tool",
+                        "parameters": {
+                            "$schema": "https://json-schema.org/draft/2020-12/schema",
+                            "type": "object",
+                            "properties": {
+                                "pattern": {"type": "string", "description": "The glob pattern"},
+                                "path": {"type": "string", "description": "The directory to search in"}
+                            },
+                            "required": ["pattern"]
+                        }
+                    }
+                },
+                {
+                    "type": "function",
+                    "function": {
+                        "name": "read",
+                        "description": "Read a file from the local filesystem",
+                        "parameters": {
+                            "$schema": "https://json-schema.org/draft/2020-12/schema",
+                            "type": "object",
+                            "properties": {
+                                "filePath": {"type": "string", "description": "The absolute path"},
+                                "offset": {
+                                    "minimum": 0,
+                                    "type": "integer",
+                                    "description": "The line number to start reading from"
+                                }
+                            },
+                            "required": ["filePath"]
+                        }
+                    }
+                }
+            ]
+        });
+
+        let params: ChatCompletionRequest = from_value(input).unwrap();
+
+        assert_eq!(params.tools.len(), 2);
+    }
+}
diff --git a/paddler_openai_translation/src/chat_completion_requested_function.rs b/paddler_openai_translation/src/chat_completion_requested_function.rs
new file mode 100644
index 000000000..0e8e6de0c
--- /dev/null
+++ b/paddler_openai_translation/src/chat_completion_requested_function.rs
@@ -0,0 +1,7 @@
+use serde::Deserialize;
+
+#[derive(Deserialize)]
+pub struct ChatCompletionRequestedFunction {
+    pub arguments: String,
+    pub name: String,
+}
diff --git a/paddler_openai_translation/src/chat_completion_requested_tool_call.rs b/paddler_openai_translation/src/chat_completion_requested_tool_call.rs
new file mode 100644
index 000000000..d1050bad3
--- /dev/null
+++ b/paddler_openai_translation/src/chat_completion_requested_tool_call.rs
@@ -0,0 +1,26 @@
+use serde::Deserialize;
+
+use paddler_messaging::conversation_tool_call::ConversationToolCall;
+
+use crate::chat_completion_requested_function::ChatCompletionRequestedFunction;
+use crate::conversation_tool_call_from_arguments::conversation_tool_call_from_arguments;
+use crate::openai_translation_error::OpenAITranslationError;
+
+#[derive(Deserialize)]
+pub struct ChatCompletionRequestedToolCall {
+    pub function: ChatCompletionRequestedFunction,
+    pub id: String,
+}
+
+impl ChatCompletionRequestedToolCall {
+    pub fn into_conversation_tool_call(
+        self,
+    ) -> Result<ConversationToolCall, OpenAITranslationError> {
+        let Self {
+            function: ChatCompletionRequestedFunction { arguments, name },
+            id,
+        } = self;
+
+        conversation_tool_call_from_arguments(id, name, &arguments)
+    }
+}
diff --git a/paddler_openai_translation/src/chat_completion_stream.rs b/paddler_openai_translation/src/chat_completion_stream.rs
new file mode 100644
index 000000000..0050f7058
--- /dev/null
+++ b/paddler_openai_translation/src/chat_completion_stream.rs
@@ -0,0 +1,381 @@
+use paddler_messaging::generated_tool_call::GeneratedToolCall;
+use paddler_messaging::generation_summary::GenerationSummary;
+
+use crate::chat_completion_chunk::ChatCompletionChunk;
+use crate::chat_completion_chunk_choice::ChatCompletionChunkChoice;
+use crate::chat_completion_chunk_payload::ChatCompletionChunkPayload;
+use crate::chat_completion_finish_reason::chat_completion_finish_reason;
+use crate::chat_completion_header::ChatCompletionHeader;
+use crate::chat_completion_stream_params::ChatCompletionStreamParams;
+use crate::generated_output_part::GeneratedOutputPart;
+use crate::generation_failure::GenerationFailure;
+use crate::generation_outcome::GenerationOutcome;
+use crate::generation_progress::GenerationProgress;
+
+pub struct ChatCompletionStream {
+    header: ChatCompletionHeader,
+    include_usage: bool,
+    saw_tool_call: bool,
+    system_fingerprint: String,
+}
+
+impl ChatCompletionStream {
+    #[must_use]
+    pub fn new(
+        ChatCompletionStreamParams {
+            header,
+            include_usage,
+            system_fingerprint,
+        }: ChatCompletionStreamParams,
+    ) -> Self {
+        Self {
+            header,
+            include_usage,
+            saw_tool_call: false,
+            system_fingerprint,
+        }
+    }
+
+    pub fn advance(
+        &mut self,
+        request_id: &str,
+        generation_progress: GenerationProgress,
+    ) -> Result<Vec<ChatCompletionChunk>, GenerationFailure> {
+        match generation_progress {
+            GenerationProgress::Produced(GeneratedOutputPart::Content(text)) => {
+                Ok(vec![self.chunk(
+                    request_id,
+                    ChatCompletionChunkPayload::Choice(ChatCompletionChunkChoice::Content(text)),
+                )])
+            }
+            GenerationProgress::Produced(GeneratedOutputPart::Reasoning(_))
+            | GenerationProgress::ToolCallTokenProduced => Ok(Vec::new()),
+            GenerationProgress::Produced(GeneratedOutputPart::ToolCalls(parsed_calls)) => {
+                Ok(self.tool_calls(request_id, parsed_calls))
+            }
+            GenerationProgress::Failed(generation_failure) => Err(generation_failure),
+        }
+    }
+
+    pub fn conclude(
+        &self,
+        request_id: &str,
+        generation_outcome: GenerationOutcome,
+    ) -> Result<Vec<ChatCompletionChunk>, GenerationFailure> {
+        match generation_outcome {
+            GenerationOutcome::Failed(generation_failure) => Err(generation_failure),
+            GenerationOutcome::Finished(generation_summary) => {
+                Ok(self.finish(request_id, &generation_summary))
+            }
+        }
+    }
+
+    fn chunk(&self, request_id: &str, payload: ChatCompletionChunkPayload) -> ChatCompletionChunk {
+        ChatCompletionChunk {
+            created: self.header.created,
+            id: request_id.to_owned(),
+            model: self.header.model.clone(),
+            payload,
+            system_fingerprint: self.system_fingerprint.clone(),
+        }
+    }
+
+    fn finish(
+        &self,
+        request_id: &str,
+        GenerationSummary { finish, usage }: &GenerationSummary,
+    ) -> Vec<ChatCompletionChunk> {
+        let finish_chunk = self.chunk(
+            request_id,
+            ChatCompletionChunkPayload::Choice(ChatCompletionChunkChoice::Finish(
+                chat_completion_finish_reason(*finish, self.saw_tool_call),
+            )),
+        );
+
+        if self.include_usage {
+            vec![
+                finish_chunk,
+                self.chunk(request_id, ChatCompletionChunkPayload::Usage(*usage)),
+            ]
+        } else {
+            vec![finish_chunk]
+        }
+    }
+
+    fn tool_calls(
+        &mut self,
+        request_id: &str,
+        parsed_calls: Vec<GeneratedToolCall>,
+    ) -> Vec<ChatCompletionChunk> {
+        if parsed_calls.is_empty() {
+            return Vec::new();
+        }
+
+        self.saw_tool_call = true;
+
+        vec![self.chunk(
+            request_id,
+            ChatCompletionChunkPayload::Choice(ChatCompletionChunkChoice::ToolCalls(parsed_calls)),
+        )]
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use llama_cpp_bindings_types::TokenUsage;
+    use llama_cpp_bindings_types::ToolCallArguments;
+    use serde_json::json;
+    use serde_json::to_string;
+
+    use paddler_messaging::generated_tool_call::GeneratedToolCall;
+    use paddler_messaging::generation_finish::GenerationFinish;
+    use paddler_messaging::generation_summary::GenerationSummary;
+
+    use super::ChatCompletionStream;
+    use crate::chat_completion_chunk::ChatCompletionChunk;
+    use crate::chat_completion_header::ChatCompletionHeader;
+    use crate::chat_completion_stream_params::ChatCompletionStreamParams;
+    use crate::generated_output_part::GeneratedOutputPart;
+    use crate::generation_failure::GenerationFailure;
+    use crate::generation_failure_cause::GenerationFailureCause;
+    use crate::generation_outcome::GenerationOutcome;
+    use crate::generation_progress::GenerationProgress;
+
+    fn chat_completion_stream(include_usage: bool) -> ChatCompletionStream {
+        ChatCompletionStream::new(ChatCompletionStreamParams {
+            header: ChatCompletionHeader {
+                created: 0,
+                model: "test-model".to_owned(),
+            },
+            include_usage,
+            system_fingerprint: "test-fingerprint".to_owned(),
+        })
+    }
+
+    fn finished(finish: GenerationFinish) -> GenerationOutcome {
+        GenerationOutcome::Finished(GenerationSummary {
+            finish,
+            usage: TokenUsage {
+                prompt_tokens: 7,
+                content_tokens: 4,
+                reasoning_tokens: 1,
+                ..TokenUsage::default()
+            },
+        })
+    }
+
+    fn serialized_chunks(
+        chat_completion_chunks: Result<Vec<ChatCompletionChunk>, GenerationFailure>,
+    ) -> Vec<String> {
+        chat_completion_chunks
+            .unwrap()
+            .iter()
+            .map(|chat_completion_chunk| to_string(chat_completion_chunk).unwrap())
+            .collect()
+    }
+
+    fn tool_call(
+        index: usize,
+        id: &str,
+        name: &str,
+        arguments: ToolCallArguments,
+    ) -> GeneratedToolCall {
+        GeneratedToolCall {
+            arguments,
+            id: id.to_owned(),
+            index,
+            name: name.to_owned(),
+        }
+    }
+
+    #[test]
+    fn emits_a_content_delta_for_produced_content() {
+        assert_eq!(
+            serialized_chunks(chat_completion_stream(false).advance(
+                "test-request",
+                GenerationProgress::Produced(GeneratedOutputPart::Content("hello".to_owned())),
+            )),
+            vec![
+                r#"{"id":"test-request","object":"chat.completion.chunk","created":0,"model":"test-model","system_fingerprint":"test-fingerprint","choices":[{"index":0,"delta":{"role":"assistant","content":"hello"},"logprobs":null,"finish_reason":null}]}"#,
+            ]
+        );
+    }
+
+    #[test]
+    fn drops_reasoning_and_tool_call_tokens() {
+        let mut chat_completion_stream = chat_completion_stream(false);
+
+        for generation_progress in [
+            GenerationProgress::Produced(GeneratedOutputPart::Reasoning("thought".to_owned())),
+            GenerationProgress::ToolCallTokenProduced,
+        ] {
+            assert_eq!(
+                serialized_chunks(
+                    chat_completion_stream.advance("test-request", generation_progress)
+                ),
+                Vec::<String>::new()
+            );
+        }
+    }
+
+    #[test]
+    fn emits_parsed_tool_calls_with_their_raw_arguments() {
+        assert_eq!(
+            serialized_chunks(chat_completion_stream(false).advance(
+                "test-request",
+                GenerationProgress::Produced(GeneratedOutputPart::ToolCalls(vec![
+                    tool_call(
+                        0,
+                        "call_x",
+                        "get_weather",
+                        ToolCallArguments::ValidJson(json!({ "location": "Paris" })),
+                    ),
+                    tool_call(
+                        1,
+                        "call_invalid",
+                        "broken_tool",
+                        ToolCallArguments::InvalidJson("{not valid json".to_owned()),
+                    ),
+                ])),
+            )),
+            vec![
+                r#"{"id":"test-request","object":"chat.completion.chunk","created":0,"model":"test-model","system_fingerprint":"test-fingerprint","choices":[{"index":0,"delta":{"role":"assistant","tool_calls":[{"index":0,"id":"call_x","type":"function","function":{"name":"get_weather","arguments":"{\"location\":\"Paris\"}"}},{"index":1,"id":"call_invalid","type":"function","function":{"name":"broken_tool","arguments":"{not valid json"}}]},"logprobs":null,"finish_reason":null}]}"#,
+            ]
+        );
+    }
+
+    #[test]
+    fn indexes_tool_calls_of_separate_sections_by_their_position_in_the_generation() {
+        let mut chat_completion_stream = chat_completion_stream(false);
+        let tool_call_chunks: Vec<String> = [
+            tool_call(
+                0,
+                "call_0",
+                "get_weather",
+                ToolCallArguments::ValidJson(json!({ "location": "Paris" })),
+            ),
+            tool_call(
+                1,
+                "call_1",
+                "get_time",
+                ToolCallArguments::ValidJson(json!({ "timezone": "CET" })),
+            ),
+        ]
+        .into_iter()
+        .flat_map(|generated_tool_call| {
+            serialized_chunks(chat_completion_stream.advance(
+                "test-request",
+                GenerationProgress::Produced(GeneratedOutputPart::ToolCalls(vec![
+                    generated_tool_call,
+                ])),
+            ))
+        })
+        .collect();
+
+        assert_eq!(
+            tool_call_chunks,
+            vec![
+                r#"{"id":"test-request","object":"chat.completion.chunk","created":0,"model":"test-model","system_fingerprint":"test-fingerprint","choices":[{"index":0,"delta":{"role":"assistant","tool_calls":[{"index":0,"id":"call_0","type":"function","function":{"name":"get_weather","arguments":"{\"location\":\"Paris\"}"}}]},"logprobs":null,"finish_reason":null}]}"#,
+                r#"{"id":"test-request","object":"chat.completion.chunk","created":0,"model":"test-model","system_fingerprint":"test-fingerprint","choices":[{"index":0,"delta":{"role":"assistant","tool_calls":[{"index":1,"id":"call_1","type":"function","function":{"name":"get_time","arguments":"{\"timezone\":\"CET\"}"}}]},"logprobs":null,"finish_reason":null}]}"#,
+            ]
+        );
+    }
+
+    #[test]
+    fn emits_nothing_for_empty_parsed_tool_calls() {
+        let mut chat_completion_stream = chat_completion_stream(false);
+
+        assert_eq!(
+            serialized_chunks(chat_completion_stream.advance(
+                "test-request",
+                GenerationProgress::Produced(GeneratedOutputPart::ToolCalls(Vec::new())),
+            )),
+            Vec::<String>::new()
+        );
+        assert_eq!(
+            serialized_chunks(
+                chat_completion_stream
+                    .conclude("test-request", finished(GenerationFinish::EndOfGeneration))
+            ),
+            vec![
+                r#"{"id":"test-request","object":"chat.completion.chunk","created":0,"model":"test-model","system_fingerprint":"test-fingerprint","choices":[{"index":0,"delta":{},"logprobs":null,"finish_reason":"stop"}]}"#,
+            ]
+        );
+    }
+
+    #[test]
+    fn finishes_with_tool_calls_after_a_tool_call_and_length_at_a_length_limit() {
+        let mut after_tool_call = chat_completion_stream(false);
+
+        serialized_chunks(after_tool_call.advance(
+            "test-request",
+            GenerationProgress::Produced(GeneratedOutputPart::ToolCalls(vec![tool_call(
+                0,
+                "call_x",
+                "get_weather",
+                ToolCallArguments::ValidJson(json!({ "location": "Paris" })),
+            )])),
+        ));
+
+        assert_eq!(
+            serialized_chunks(
+                after_tool_call
+                    .conclude("test-request", finished(GenerationFinish::EndOfGeneration))
+            ),
+            vec![
+                r#"{"id":"test-request","object":"chat.completion.chunk","created":0,"model":"test-model","system_fingerprint":"test-fingerprint","choices":[{"index":0,"delta":{},"logprobs":null,"finish_reason":"tool_calls"}]}"#,
+            ]
+        );
+        assert_eq!(
+            serialized_chunks(
+                chat_completion_stream(false)
+                    .conclude("test-request", finished(GenerationFinish::MaxTokens))
+            ),
+            vec![
+                r#"{"id":"test-request","object":"chat.completion.chunk","created":0,"model":"test-model","system_fingerprint":"test-fingerprint","choices":[{"index":0,"delta":{},"logprobs":null,"finish_reason":"length"}]}"#,
+            ]
+        );
+    }
+
+    #[test]
+    fn appends_a_usage_chunk_when_the_client_asks_for_usage() {
+        assert_eq!(
+            serialized_chunks(
+                chat_completion_stream(true)
+                    .conclude("test-request", finished(GenerationFinish::EndOfGeneration))
+            ),
+            vec![
+                r#"{"id":"test-request","object":"chat.completion.chunk","created":0,"model":"test-model","system_fingerprint":"test-fingerprint","choices":[{"index":0,"delta":{},"logprobs":null,"finish_reason":"stop"}]}"#,
+                r#"{"id":"test-request","object":"chat.completion.chunk","created":0,"model":"test-model","system_fingerprint":"test-fingerprint","choices":[],"usage":{"prompt_tokens":7,"completion_tokens":5,"total_tokens":12,"prompt_tokens_details":{"cached_tokens":0,"audio_tokens":0},"completion_tokens_details":{"reasoning_tokens":1}}}"#,
+            ]
+        );
+    }
+
+    #[test]
+    fn passes_generation_failures_through() {
+        let generation_failure = || GenerationFailure {
+            cause: GenerationFailureCause::AgentFailed,
+            message: "sampler blew up".to_owned(),
+        };
+        let mut chat_completion_stream = chat_completion_stream(false);
+
+        assert_eq!(
+            chat_completion_stream
+                .advance(
+                    "test-request",
+                    GenerationProgress::Failed(generation_failure())
+                )
+                .err(),
+            Some(generation_failure())
+        );
+        assert_eq!(
+            chat_completion_stream
+                .conclude(
+                    "test-request",
+                    GenerationOutcome::Failed(generation_failure())
+                )
+                .err(),
+            Some(generation_failure())
+        );
+    }
+}
diff --git a/paddler_openai_translation/src/chat_completion_stream_options.rs b/paddler_openai_translation/src/chat_completion_stream_options.rs
new file mode 100644
index 000000000..1ac529ec0
--- /dev/null
+++ b/paddler_openai_translation/src/chat_completion_stream_options.rs
@@ -0,0 +1,8 @@
+use serde::Deserialize;
+
+#[derive(Default, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct ChatCompletionStreamOptions {
+    #[serde(default)]
+    pub include_usage: bool,
+}
diff --git a/paddler_openai_translation/src/chat_completion_stream_params.rs b/paddler_openai_translation/src/chat_completion_stream_params.rs
new file mode 100644
index 000000000..75ea984b9
--- /dev/null
+++ b/paddler_openai_translation/src/chat_completion_stream_params.rs
@@ -0,0 +1,7 @@
+use crate::chat_completion_header::ChatCompletionHeader;
+
+pub struct ChatCompletionStreamParams {
+    pub header: ChatCompletionHeader,
+    pub include_usage: bool,
+    pub system_fingerprint: String,
+}
diff --git a/paddler_openai_translation/src/chat_completion_tool.rs b/paddler_openai_translation/src/chat_completion_tool.rs
new file mode 100644
index 000000000..f78d32125
--- /dev/null
+++ b/paddler_openai_translation/src/chat_completion_tool.rs
@@ -0,0 +1,49 @@
+use serde::Deserialize;
+
+use paddler_messaging::request_params::continue_from_conversation_history_params::tool::Tool;
+use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters_schema::raw_parameters_schema::RawParametersSchema;
+
+use crate::function_definition::FunctionDefinition;
+
+#[derive(Deserialize)]
+#[serde(tag = "type")]
+pub enum ChatCompletionTool {
+    #[serde(rename = "function")]
+    Function { function: Box<FunctionDefinition> },
+}
+
+impl ChatCompletionTool {
+    #[must_use]
+    pub fn into_tool(self) -> Tool<RawParametersSchema> {
+        let Self::Function { function } = self;
+
+        (*function).into_tool()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use serde_json::from_value;
+    use serde_json::json;
+
+    use paddler_messaging::request_params::continue_from_conversation_history_params::tool::Tool;
+
+    use super::ChatCompletionTool;
+
+    #[test]
+    fn function_tool_converts_to_internal_tool() {
+        let tool: ChatCompletionTool = from_value(json!({
+            "type": "function",
+            "function": {
+                "name": "get_weather",
+                "description": "fetch weather",
+                "parameters": {"type": "object"}
+            }
+        }))
+        .unwrap();
+
+        let Tool::Function(function_call) = tool.into_tool();
+
+        assert_eq!(function_call.function.name, "get_weather");
+    }
+}
diff --git a/paddler_balancer/src/compatibility/openai_service/chat_completion_tool_call.rs b/paddler_openai_translation/src/chat_completion_tool_call.rs
similarity index 81%
rename from paddler_balancer/src/compatibility/openai_service/chat_completion_tool_call.rs
rename to paddler_openai_translation/src/chat_completion_tool_call.rs
index 1dabd1e04..9c01b53fb 100644
--- a/paddler_balancer/src/compatibility/openai_service/chat_completion_tool_call.rs
+++ b/paddler_openai_translation/src/chat_completion_tool_call.rs
@@ -1,8 +1,9 @@
-use llama_cpp_bindings_types::ParsedToolCall;
 use serde::Serialize;
 use serde::Serializer;
 
-use crate::compatibility::openai_service::arguments_to_tool_call_string::arguments_to_tool_call_string;
+use paddler_messaging::generated_tool_call::GeneratedToolCall;
+
+use crate::arguments_to_tool_call_string::arguments_to_tool_call_string;
 
 #[derive(Serialize)]
 struct SerializedFunction<'call> {
@@ -18,7 +19,7 @@ struct SerializedToolCall<'call> {
     function: SerializedFunction<'call>,
 }
 
-pub struct ChatCompletionToolCall<'call>(pub &'call ParsedToolCall);
+pub struct ChatCompletionToolCall<'call>(pub &'call GeneratedToolCall);
 
 impl Serialize for ChatCompletionToolCall<'_> {
     fn serialize<TSerializer>(
diff --git a/paddler_openai_translation/src/chat_completion_usage.rs b/paddler_openai_translation/src/chat_completion_usage.rs
new file mode 100644
index 000000000..f539c7372
--- /dev/null
+++ b/paddler_openai_translation/src/chat_completion_usage.rs
@@ -0,0 +1,52 @@
+use llama_cpp_bindings_types::TokenUsage;
+use serde::Serialize;
+use serde::Serializer;
+
+#[derive(Serialize)]
+struct PromptTokensDetails {
+    cached_tokens: u64,
+    audio_tokens: u64,
+}
+
+#[derive(Serialize)]
+struct CompletionTokensDetails {
+    reasoning_tokens: u64,
+}
+
+#[derive(Serialize)]
+struct SerializedUsage {
+    prompt_tokens: u64,
+    completion_tokens: u64,
+    total_tokens: u64,
+    prompt_tokens_details: PromptTokensDetails,
+    completion_tokens_details: CompletionTokensDetails,
+}
+
+#[derive(Clone, Copy, Debug)]
+pub struct ChatCompletionUsage(pub TokenUsage);
+
+impl Serialize for ChatCompletionUsage {
+    fn serialize<TSerializer>(
+        &self,
+        serializer: TSerializer,
+    ) -> Result<TSerializer::Ok, TSerializer::Error>
+    where
+        TSerializer: Serializer,
+    {
+        let Self(usage) = self;
+
+        SerializedUsage {
+            prompt_tokens: usage.prompt_tokens,
+            completion_tokens: usage.completion_tokens(),
+            total_tokens: usage.total_tokens(),
+            prompt_tokens_details: PromptTokensDetails {
+                cached_tokens: usage.cached_prompt_tokens,
+                audio_tokens: usage.input_audio_tokens,
+            },
+            completion_tokens_details: CompletionTokensDetails {
+                reasoning_tokens: usage.reasoning_tokens,
+            },
+        }
+        .serialize(serializer)
+    }
+}
diff --git a/paddler_openai_translation/src/completes_generation.rs b/paddler_openai_translation/src/completes_generation.rs
new file mode 100644
index 000000000..0f8aea235
--- /dev/null
+++ b/paddler_openai_translation/src/completes_generation.rs
@@ -0,0 +1,16 @@
+use serde::Serialize;
+
+use paddler_messaging::generation_summary::GenerationSummary;
+
+use crate::generated_output::GeneratedOutput;
+
+pub trait CompletesGeneration {
+    type Completion: Serialize;
+
+    fn complete(
+        &self,
+        request_id: &str,
+        generated_output: GeneratedOutput,
+        generation_summary: &GenerationSummary,
+    ) -> Self::Completion;
+}
diff --git a/paddler_openai_translation/src/content_part_event.rs b/paddler_openai_translation/src/content_part_event.rs
new file mode 100644
index 000000000..1cffabd9f
--- /dev/null
+++ b/paddler_openai_translation/src/content_part_event.rs
@@ -0,0 +1,12 @@
+use serde::Serialize;
+
+use crate::responses_content_part::ResponsesContentPart;
+
+#[derive(Clone, Debug, Serialize)]
+pub struct ContentPartEvent {
+    pub sequence_number: u64,
+    pub item_id: String,
+    pub output_index: usize,
+    pub content_index: usize,
+    pub part: ResponsesContentPart,
+}
diff --git a/paddler_openai_translation/src/conversation_tool_call_from_arguments.rs b/paddler_openai_translation/src/conversation_tool_call_from_arguments.rs
new file mode 100644
index 000000000..08a694fa6
--- /dev/null
+++ b/paddler_openai_translation/src/conversation_tool_call_from_arguments.rs
@@ -0,0 +1,23 @@
+use serde_json::from_str;
+
+use paddler_messaging::conversation_tool_call::ConversationToolCall;
+
+use crate::openai_translation_error::OpenAITranslationError;
+
+pub fn conversation_tool_call_from_arguments(
+    id: String,
+    name: String,
+    arguments: &str,
+) -> Result<ConversationToolCall, OpenAITranslationError> {
+    match from_str(arguments) {
+        Ok(arguments) => Ok(ConversationToolCall {
+            arguments,
+            id,
+            name,
+        }),
+        Err(source) => Err(OpenAITranslationError::ToolCallArgumentsAreNotAJsonObject {
+            tool_call_id: id,
+            source,
+        }),
+    }
+}
diff --git a/paddler_openai_translation/src/default_max_tokens.rs b/paddler_openai_translation/src/default_max_tokens.rs
new file mode 100644
index 000000000..318c75a5d
--- /dev/null
+++ b/paddler_openai_translation/src/default_max_tokens.rs
@@ -0,0 +1,3 @@
+use std::num::NonZeroU32;
+
+pub const DEFAULT_MAX_TOKENS: NonZeroU32 = NonZeroU32::MAX;
diff --git a/paddler_balancer/src/compatibility/openai_service/function_call_arguments_delta_event.rs b/paddler_openai_translation/src/function_call_arguments_delta_event.rs
similarity index 100%
rename from paddler_balancer/src/compatibility/openai_service/function_call_arguments_delta_event.rs
rename to paddler_openai_translation/src/function_call_arguments_delta_event.rs
diff --git a/paddler_balancer/src/compatibility/openai_service/function_call_arguments_done_event.rs b/paddler_openai_translation/src/function_call_arguments_done_event.rs
similarity index 100%
rename from paddler_balancer/src/compatibility/openai_service/function_call_arguments_done_event.rs
rename to paddler_openai_translation/src/function_call_arguments_done_event.rs
diff --git a/paddler_openai_translation/src/function_definition.rs b/paddler_openai_translation/src/function_definition.rs
new file mode 100644
index 000000000..f87b28e6e
--- /dev/null
+++ b/paddler_openai_translation/src/function_definition.rs
@@ -0,0 +1,79 @@
+use serde::Deserialize;
+
+use paddler_messaging::request_params::continue_from_conversation_history_params::tool::Tool;
+use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::FunctionCall;
+use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::function::Function;
+use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters::Parameters;
+use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters_schema::raw_parameters_schema::RawParametersSchema;
+
+use crate::tool_parameters_schema::ToolParametersSchema;
+
+#[derive(Deserialize)]
+pub struct FunctionDefinition {
+    pub name: String,
+    #[serde(default)]
+    pub description: Option<String>,
+    #[serde(default)]
+    pub parameters: Option<ToolParametersSchema>,
+}
+
+impl FunctionDefinition {
+    #[must_use]
+    pub fn into_tool(self) -> Tool<RawParametersSchema> {
+        let Self {
+            name,
+            description,
+            parameters,
+        } = self;
+
+        Tool::Function(FunctionCall {
+            function: Function {
+                name,
+                description: description.unwrap_or_default(),
+                parameters: parameters.map_or(Parameters::Empty, |parameters| {
+                    Parameters::Schema(parameters.into_raw_parameters_schema())
+                }),
+            },
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use serde_json::from_value;
+    use serde_json::json;
+
+    use paddler_messaging::request_params::continue_from_conversation_history_params::tool::Tool;
+
+    use super::FunctionDefinition;
+
+    #[test]
+    fn conversion_ignores_unknown_fields_and_builds_function_with_schema() {
+        let function: FunctionDefinition = from_value(json!({
+            "name": "get_weather",
+            "description": "fetch weather",
+            "parameters": {"type": "object", "properties": {"location": {"type": "string"}}},
+            "strict": true
+        }))
+        .unwrap();
+
+        let Tool::Function(function_call) = function.into_tool();
+
+        assert_eq!(function_call.function.name, "get_weather");
+        assert_eq!(function_call.function.description, "fetch weather");
+        assert!(!function_call.function.parameters.is_empty());
+    }
+
+    #[test]
+    fn conversion_defaults_missing_description_and_parameters() {
+        let function: FunctionDefinition = from_value(json!({
+            "name": "noop"
+        }))
+        .unwrap();
+
+        let Tool::Function(function_call) = function.into_tool();
+
+        assert_eq!(function_call.function.description, "");
+        assert!(function_call.function.parameters.is_empty());
+    }
+}
diff --git a/paddler_openai_translation/src/generated_output.rs b/paddler_openai_translation/src/generated_output.rs
new file mode 100644
index 000000000..ccce1f470
--- /dev/null
+++ b/paddler_openai_translation/src/generated_output.rs
@@ -0,0 +1,20 @@
+use paddler_messaging::generated_tool_call::GeneratedToolCall;
+
+use crate::generated_output_part::GeneratedOutputPart;
+
+#[derive(Default)]
+pub struct GeneratedOutput {
+    pub content: String,
+    pub reasoning: String,
+    pub tool_calls: Vec<GeneratedToolCall>,
+}
+
+impl GeneratedOutput {
+    pub fn append(&mut self, generated_output_part: GeneratedOutputPart) {
+        match generated_output_part {
+            GeneratedOutputPart::Content(text) => self.content.push_str(&text),
+            GeneratedOutputPart::Reasoning(text) => self.reasoning.push_str(&text),
+            GeneratedOutputPart::ToolCalls(parsed_calls) => self.tool_calls.extend(parsed_calls),
+        }
+    }
+}
diff --git a/paddler_openai_translation/src/generated_output_part.rs b/paddler_openai_translation/src/generated_output_part.rs
new file mode 100644
index 000000000..b651d615a
--- /dev/null
+++ b/paddler_openai_translation/src/generated_output_part.rs
@@ -0,0 +1,8 @@
+use paddler_messaging::generated_tool_call::GeneratedToolCall;
+
+#[derive(Debug, Eq, PartialEq)]
+pub enum GeneratedOutputPart {
+    Content(String),
+    Reasoning(String),
+    ToolCalls(Vec<GeneratedToolCall>),
+}
diff --git a/paddler_openai_translation/src/generation_failure.rs b/paddler_openai_translation/src/generation_failure.rs
new file mode 100644
index 000000000..bc0753e18
--- /dev/null
+++ b/paddler_openai_translation/src/generation_failure.rs
@@ -0,0 +1,7 @@
+use crate::generation_failure_cause::GenerationFailureCause;
+
+#[derive(Debug, Eq, PartialEq)]
+pub struct GenerationFailure {
+    pub cause: GenerationFailureCause,
+    pub message: String,
+}
diff --git a/paddler_openai_translation/src/generation_failure_cause.rs b/paddler_openai_translation/src/generation_failure_cause.rs
new file mode 100644
index 000000000..5bbb7cfb3
--- /dev/null
+++ b/paddler_openai_translation/src/generation_failure_cause.rs
@@ -0,0 +1,6 @@
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+pub enum GenerationFailureCause {
+    AgentFailed,
+    InvalidRequest,
+    Unavailable,
+}
diff --git a/paddler_openai_translation/src/generation_outcome.rs b/paddler_openai_translation/src/generation_outcome.rs
new file mode 100644
index 000000000..849238747
--- /dev/null
+++ b/paddler_openai_translation/src/generation_outcome.rs
@@ -0,0 +1,288 @@
+use paddler_messaging::generated_token_final_result::GeneratedTokenFinalResult;
+use paddler_messaging::generation_summary::GenerationSummary;
+use paddler_messaging::oversized_media_details::OversizedMediaDetails;
+use paddler_messaging::oversized_prompt_details::OversizedPromptDetails;
+
+use crate::generation_failure::GenerationFailure;
+use crate::generation_failure_cause::GenerationFailureCause;
+
+fn media_exceeds_micro_batch_message(
+    OversizedMediaDetails {
+        media_tokens,
+        micro_batch_tokens,
+    }: &OversizedMediaDetails,
+) -> String {
+    format!(
+        "media required {media_tokens} tokens but one agent micro batch holds {micro_batch_tokens} tokens"
+    )
+}
+
+fn prompt_exceeds_context_size_message(
+    OversizedPromptDetails {
+        prompt_tokens,
+        sequence_context_size,
+    }: &OversizedPromptDetails,
+) -> String {
+    format!(
+        "prompt has {prompt_tokens} tokens but each agent sequence holds {sequence_context_size} tokens; shorten the prompt or raise context_size"
+    )
+}
+
+#[derive(Debug, Eq, PartialEq)]
+pub enum GenerationOutcome {
+    Failed(GenerationFailure),
+    Finished(GenerationSummary),
+}
+
+impl GenerationOutcome {
+    const fn failed(cause: GenerationFailureCause, message: String) -> Self {
+        Self::Failed(GenerationFailure { cause, message })
+    }
+}
+
+impl From<GeneratedTokenFinalResult> for GenerationOutcome {
+    fn from(generated_token_final_result: GeneratedTokenFinalResult) -> Self {
+        match generated_token_final_result {
+            GeneratedTokenFinalResult::Done(generation_summary) => {
+                Self::Finished(generation_summary)
+            }
+            GeneratedTokenFinalResult::MediaExceedsMicroBatch(oversized_media_details) => {
+                Self::failed(
+                    GenerationFailureCause::InvalidRequest,
+                    media_exceeds_micro_batch_message(&oversized_media_details),
+                )
+            }
+            GeneratedTokenFinalResult::PromptExceedsContextSize(oversized_prompt_details) => {
+                Self::failed(
+                    GenerationFailureCause::InvalidRequest,
+                    prompt_exceeds_context_size_message(&oversized_prompt_details),
+                )
+            }
+            GeneratedTokenFinalResult::GrammarIncompatibleWithThinking(description)
+            | GeneratedTokenFinalResult::GrammarSyntaxError(description)
+            | GeneratedTokenFinalResult::ImageDecodingFailed(description)
+            | GeneratedTokenFinalResult::MultimodalNotSupported(description)
+            | GeneratedTokenFinalResult::MultimodalTokenizationFailed(description)
+            | GeneratedTokenFinalResult::PromptTokenizationFailed(description)
+            | GeneratedTokenFinalResult::PromptTokenizedToNothing(description)
+            | GeneratedTokenFinalResult::ToolSchemaInvalid(description) => {
+                Self::failed(GenerationFailureCause::InvalidRequest, description)
+            }
+            GeneratedTokenFinalResult::ModelNotLoaded(description)
+            | GeneratedTokenFinalResult::NoSequenceSlotAvailable(description)
+            | GeneratedTokenFinalResult::SchedulerUnavailable(description) => {
+                Self::failed(GenerationFailureCause::Unavailable, description)
+            }
+            GeneratedTokenFinalResult::BatchAssemblyFailed(description)
+            | GeneratedTokenFinalResult::ChatTemplateError(description)
+            | GeneratedTokenFinalResult::DecodeFailed(description)
+            | GeneratedTokenFinalResult::DetokenizationFailed(description)
+            | GeneratedTokenFinalResult::GrammarInitializationFailed(description)
+            | GeneratedTokenFinalResult::GrammarRejectedModelOutput(description)
+            | GeneratedTokenFinalResult::KvCacheClearFailed(description)
+            | GeneratedTokenFinalResult::MediaMicroBatchCheckFailed(description)
+            | GeneratedTokenFinalResult::MultimodalIngestionFailed(description)
+            | GeneratedTokenFinalResult::SamplerChainCreationFailed(description)
+            | GeneratedTokenFinalResult::SamplerError(description)
+            | GeneratedTokenFinalResult::SamplingCandidatesExhausted(description) => {
+                Self::failed(GenerationFailureCause::AgentFailed, description)
+            }
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use llama_cpp_bindings_types::TokenUsage;
+
+    use paddler_messaging::generated_token_final_result::GeneratedTokenFinalResult;
+    use paddler_messaging::generation_finish::GenerationFinish;
+    use paddler_messaging::generation_summary::GenerationSummary;
+    use paddler_messaging::oversized_media_details::OversizedMediaDetails;
+    use paddler_messaging::oversized_prompt_details::OversizedPromptDetails;
+
+    use super::GenerationOutcome;
+    use crate::generation_failure::GenerationFailure;
+    use crate::generation_failure_cause::GenerationFailureCause;
+
+    fn assert_failures_classified_as(
+        expected_cause: GenerationFailureCause,
+        failures: Vec<(GeneratedTokenFinalResult, &str)>,
+    ) {
+        for (generated_token_final_result, expected_message) in failures {
+            assert_eq!(
+                GenerationOutcome::from(generated_token_final_result),
+                GenerationOutcome::Failed(GenerationFailure {
+                    cause: expected_cause,
+                    message: expected_message.to_owned(),
+                })
+            );
+        }
+    }
+
+    #[test]
+    fn classifies_the_summary_as_a_finished_generation() {
+        let generation_summary = GenerationSummary {
+            finish: GenerationFinish::EndOfGeneration,
+            usage: TokenUsage::new(),
+        };
+
+        assert_eq!(
+            GenerationOutcome::from(GeneratedTokenFinalResult::Done(generation_summary)),
+            GenerationOutcome::Finished(generation_summary)
+        );
+    }
+
+    #[test]
+    fn classifies_client_mistakes_as_invalid_requests() {
+        assert_failures_classified_as(
+            GenerationFailureCause::InvalidRequest,
+            vec![
+                (
+                    GeneratedTokenFinalResult::GrammarIncompatibleWithThinking(
+                        "client mistake".to_owned(),
+                    ),
+                    "client mistake",
+                ),
+                (
+                    GeneratedTokenFinalResult::GrammarSyntaxError("client mistake".to_owned()),
+                    "client mistake",
+                ),
+                (
+                    GeneratedTokenFinalResult::ImageDecodingFailed("client mistake".to_owned()),
+                    "client mistake",
+                ),
+                (
+                    GeneratedTokenFinalResult::MediaExceedsMicroBatch(OversizedMediaDetails {
+                        media_tokens: 256,
+                        micro_batch_tokens: 128,
+                    }),
+                    "media required 256 tokens but one agent micro batch holds 128 tokens",
+                ),
+                (
+                    GeneratedTokenFinalResult::MultimodalNotSupported("client mistake".to_owned()),
+                    "client mistake",
+                ),
+                (
+                    GeneratedTokenFinalResult::MultimodalTokenizationFailed(
+                        "client mistake".to_owned(),
+                    ),
+                    "client mistake",
+                ),
+                (
+                    GeneratedTokenFinalResult::PromptExceedsContextSize(OversizedPromptDetails {
+                        prompt_tokens: 9895,
+                        sequence_context_size: 8192,
+                    }),
+                    "prompt has 9895 tokens but each agent sequence holds 8192 tokens; shorten the prompt or raise context_size",
+                ),
+                (
+                    GeneratedTokenFinalResult::PromptTokenizationFailed(
+                        "client mistake".to_owned(),
+                    ),
+                    "client mistake",
+                ),
+                (
+                    GeneratedTokenFinalResult::PromptTokenizedToNothing(
+                        "client mistake".to_owned(),
+                    ),
+                    "client mistake",
+                ),
+                (
+                    GeneratedTokenFinalResult::ToolSchemaInvalid("client mistake".to_owned()),
+                    "client mistake",
+                ),
+            ],
+        );
+    }
+
+    #[test]
+    fn classifies_agents_that_cannot_serve_as_unavailable() {
+        assert_failures_classified_as(
+            GenerationFailureCause::Unavailable,
+            vec![
+                (
+                    GeneratedTokenFinalResult::ModelNotLoaded("unavailable".to_owned()),
+                    "unavailable",
+                ),
+                (
+                    GeneratedTokenFinalResult::NoSequenceSlotAvailable("unavailable".to_owned()),
+                    "unavailable",
+                ),
+                (
+                    GeneratedTokenFinalResult::SchedulerUnavailable("unavailable".to_owned()),
+                    "unavailable",
+                ),
+            ],
+        );
+    }
+
+    #[test]
+    fn classifies_agent_failures_as_agent_failed() {
+        assert_failures_classified_as(
+            GenerationFailureCause::AgentFailed,
+            vec![
+                (
+                    GeneratedTokenFinalResult::BatchAssemblyFailed("agent failure".to_owned()),
+                    "agent failure",
+                ),
+                (
+                    GeneratedTokenFinalResult::ChatTemplateError("agent failure".to_owned()),
+                    "agent failure",
+                ),
+                (
+                    GeneratedTokenFinalResult::DecodeFailed("agent failure".to_owned()),
+                    "agent failure",
+                ),
+                (
+                    GeneratedTokenFinalResult::DetokenizationFailed("agent failure".to_owned()),
+                    "agent failure",
+                ),
+                (
+                    GeneratedTokenFinalResult::GrammarInitializationFailed(
+                        "agent failure".to_owned(),
+                    ),
+                    "agent failure",
+                ),
+                (
+                    GeneratedTokenFinalResult::GrammarRejectedModelOutput(
+                        "agent failure".to_owned(),
+                    ),
+                    "agent failure",
+                ),
+                (
+                    GeneratedTokenFinalResult::KvCacheClearFailed("agent failure".to_owned()),
+                    "agent failure",
+                ),
+                (
+                    GeneratedTokenFinalResult::MediaMicroBatchCheckFailed(
+                        "agent failure".to_owned(),
+                    ),
+                    "agent failure",
+                ),
+                (
+                    GeneratedTokenFinalResult::MultimodalIngestionFailed(
+                        "agent failure".to_owned(),
+                    ),
+                    "agent failure",
+                ),
+                (
+                    GeneratedTokenFinalResult::SamplerChainCreationFailed(
+                        "agent failure".to_owned(),
+                    ),
+                    "agent failure",
+                ),
+                (
+                    GeneratedTokenFinalResult::SamplerError("agent failure".to_owned()),
+                    "agent failure",
+                ),
+                (
+                    GeneratedTokenFinalResult::SamplingCandidatesExhausted(
+                        "agent failure".to_owned(),
+                    ),
+                    "agent failure",
+                ),
+            ],
+        );
+    }
+}
diff --git a/paddler_openai_translation/src/generation_progress.rs b/paddler_openai_translation/src/generation_progress.rs
new file mode 100644
index 000000000..fbdfebdf3
--- /dev/null
+++ b/paddler_openai_translation/src/generation_progress.rs
@@ -0,0 +1,154 @@
+use paddler_messaging::generated_token_partial_result::GeneratedTokenPartialResult;
+use paddler_messaging::raw_tool_call_tokens::RawToolCallTokens;
+
+use crate::generated_output_part::GeneratedOutputPart;
+use crate::generation_failure::GenerationFailure;
+use crate::generation_failure_cause::GenerationFailureCause;
+
+fn unrecognized_tool_call_format_message(
+    RawToolCallTokens {
+        ffi_error_message,
+        text,
+    }: &RawToolCallTokens,
+) -> String {
+    format!(
+        "model produced output the parser did not recognise as any registered tool-call format; \
+         FFI error: {ffi_error_message}; raw text: {text}"
+    )
+}
+
+#[derive(Debug, Eq, PartialEq)]
+pub enum GenerationProgress {
+    Failed(GenerationFailure),
+    Produced(GeneratedOutputPart),
+    ToolCallTokenProduced,
+}
+
+impl GenerationProgress {
+    const fn agent_failed(message: String) -> Self {
+        Self::Failed(GenerationFailure {
+            cause: GenerationFailureCause::AgentFailed,
+            message,
+        })
+    }
+}
+
+impl From<GeneratedTokenPartialResult> for GenerationProgress {
+    fn from(generated_token_partial_result: GeneratedTokenPartialResult) -> Self {
+        match generated_token_partial_result {
+            GeneratedTokenPartialResult::ContentToken(text)
+            | GeneratedTokenPartialResult::UndeterminableToken(text) => {
+                Self::Produced(GeneratedOutputPart::Content(text))
+            }
+            GeneratedTokenPartialResult::ReasoningToken(text) => {
+                Self::Produced(GeneratedOutputPart::Reasoning(text))
+            }
+            GeneratedTokenPartialResult::ToolCallParsed(parsed_calls) => {
+                Self::Produced(GeneratedOutputPart::ToolCalls(parsed_calls))
+            }
+            GeneratedTokenPartialResult::ToolCallToken(_) => Self::ToolCallTokenProduced,
+            GeneratedTokenPartialResult::ToolCallParseFailed(description) => {
+                Self::agent_failed(description)
+            }
+            GeneratedTokenPartialResult::ToolCallValidationFailed(validation_errors) => {
+                Self::agent_failed(validation_errors.join("; "))
+            }
+            GeneratedTokenPartialResult::UnrecognizedToolCallFormat(raw_tool_call_tokens) => {
+                Self::agent_failed(unrecognized_tool_call_format_message(&raw_tool_call_tokens))
+            }
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use llama_cpp_bindings_types::ToolCallArguments;
+    use serde_json::json;
+
+    use paddler_messaging::generated_token_partial_result::GeneratedTokenPartialResult;
+    use paddler_messaging::generated_tool_call::GeneratedToolCall;
+    use paddler_messaging::raw_tool_call_tokens::RawToolCallTokens;
+
+    use super::GenerationProgress;
+    use crate::generated_output_part::GeneratedOutputPart;
+    use crate::generation_failure::GenerationFailure;
+    use crate::generation_failure_cause::GenerationFailureCause;
+
+    fn weather_call() -> GeneratedToolCall {
+        GeneratedToolCall {
+            arguments: ToolCallArguments::ValidJson(json!({ "location": "Paris" })),
+            id: "call_x".to_owned(),
+            index: 0,
+            name: "get_weather".to_owned(),
+        }
+    }
+
+    fn agent_failed(message: &str) -> GenerationProgress {
+        GenerationProgress::Failed(GenerationFailure {
+            cause: GenerationFailureCause::AgentFailed,
+            message: message.to_owned(),
+        })
+    }
+
+    #[test]
+    fn classifies_tokens_and_parsed_tool_calls_as_produced_output() {
+        for (generated_token_partial_result, expected_progress) in [
+            (
+                GeneratedTokenPartialResult::ContentToken("hello".to_owned()),
+                GenerationProgress::Produced(GeneratedOutputPart::Content("hello".to_owned())),
+            ),
+            (
+                GeneratedTokenPartialResult::UndeterminableToken("ambig".to_owned()),
+                GenerationProgress::Produced(GeneratedOutputPart::Content("ambig".to_owned())),
+            ),
+            (
+                GeneratedTokenPartialResult::ReasoningToken("thought".to_owned()),
+                GenerationProgress::Produced(GeneratedOutputPart::Reasoning("thought".to_owned())),
+            ),
+            (
+                GeneratedTokenPartialResult::ToolCallParsed(vec![weather_call()]),
+                GenerationProgress::Produced(GeneratedOutputPart::ToolCalls(vec![weather_call()])),
+            ),
+            (
+                GeneratedTokenPartialResult::ToolCallToken("{".to_owned()),
+                GenerationProgress::ToolCallTokenProduced,
+            ),
+        ] {
+            assert_eq!(
+                GenerationProgress::from(generated_token_partial_result),
+                expected_progress
+            );
+        }
+    }
+
+    #[test]
+    fn classifies_tool_call_failures_as_agent_failures() {
+        for (generated_token_partial_result, expected_progress) in [
+            (
+                GeneratedTokenPartialResult::ToolCallParseFailed("agent failure".to_owned()),
+                agent_failed("agent failure"),
+            ),
+            (
+                GeneratedTokenPartialResult::ToolCallValidationFailed(vec![
+                    "first issue".to_owned(),
+                    "second issue".to_owned(),
+                ]),
+                agent_failed("first issue; second issue"),
+            ),
+            (
+                GeneratedTokenPartialResult::UnrecognizedToolCallFormat(RawToolCallTokens {
+                    text: "<unknown_marker>blah</unknown_marker>".to_owned(),
+                    ffi_error_message: "common_chat_parse failed: no parser".to_owned(),
+                }),
+                agent_failed(
+                    "model produced output the parser did not recognise as any registered tool-call format; FFI error: common_chat_parse failed: no parser; raw text: <unknown_marker>blah</unknown_marker>",
+                ),
+            ),
+        ] {
+            assert_eq!(
+                GenerationProgress::from(generated_token_partial_result),
+                expected_progress
+            );
+        }
+    }
+}
diff --git a/paddler_openai_translation/src/lib.rs b/paddler_openai_translation/src/lib.rs
new file mode 100644
index 000000000..e0111e160
--- /dev/null
+++ b/paddler_openai_translation/src/lib.rs
@@ -0,0 +1,79 @@
+pub mod arguments_to_tool_call_string;
+pub mod chat_completion;
+pub mod chat_completion_chunk;
+pub mod chat_completion_chunk_choice;
+pub mod chat_completion_chunk_payload;
+pub mod chat_completion_delivery;
+pub mod chat_completion_finish_reason;
+pub mod chat_completion_header;
+pub mod chat_completion_message;
+pub mod chat_completion_message_role;
+pub mod chat_completion_request;
+pub mod chat_completion_requested_function;
+pub mod chat_completion_requested_tool_call;
+pub mod chat_completion_stream;
+pub mod chat_completion_stream_options;
+pub mod chat_completion_stream_params;
+pub mod chat_completion_tool;
+pub mod chat_completion_tool_call;
+pub mod chat_completion_usage;
+pub mod completes_generation;
+pub mod content_part_event;
+pub mod conversation_tool_call_from_arguments;
+pub mod default_max_tokens;
+pub mod function_call_arguments_delta_event;
+pub mod function_call_arguments_done_event;
+pub mod function_definition;
+pub mod generated_output;
+pub mod generated_output_part;
+pub mod generation_failure;
+pub mod generation_failure_cause;
+pub mod generation_outcome;
+pub mod generation_progress;
+pub mod open_item;
+pub mod openai_translation_error;
+pub mod output_item_event;
+pub mod reasoning_effort;
+pub mod response_snapshot_event;
+pub mod responses_assistant_content_part;
+pub mod responses_assistant_message_content;
+pub mod responses_content_part;
+pub mod responses_delivery;
+pub mod responses_function_call_item;
+pub mod responses_function_call_output_item;
+pub mod responses_function_output;
+pub mod responses_function_output_part;
+pub mod responses_input;
+pub mod responses_input_content_part;
+pub mod responses_input_item;
+pub mod responses_item_status;
+pub mod responses_message_content;
+pub mod responses_message_item;
+pub mod responses_output_item;
+pub mod responses_output_item_kind;
+pub mod responses_reasoning;
+pub mod responses_reasoning_item;
+pub mod responses_reasoning_part;
+pub mod responses_reasoning_text_part;
+pub mod responses_request;
+pub mod responses_response;
+pub mod responses_response_header;
+pub mod responses_response_progress;
+pub mod responses_response_status;
+pub mod responses_stream;
+pub mod responses_stream_event;
+pub mod responses_tagged_item;
+pub mod responses_text_format;
+pub mod responses_text_format_field;
+pub mod responses_text_format_refusal;
+pub mod responses_text_format_schema_occurrences;
+pub mod responses_text_format_type;
+pub mod responses_text_param;
+pub mod responses_tool;
+pub mod responses_usage;
+pub mod text_delta_event;
+pub mod text_done_event;
+pub mod timestamp_from;
+pub mod tool_parameters_schema;
+pub mod translated_chat_completion_request;
+pub mod translated_responses_request;
diff --git a/paddler_balancer/src/compatibility/openai_service/open_item.rs b/paddler_openai_translation/src/open_item.rs
similarity index 100%
rename from paddler_balancer/src/compatibility/openai_service/open_item.rs
rename to paddler_openai_translation/src/open_item.rs
diff --git a/paddler_openai_translation/src/openai_translation_error.rs b/paddler_openai_translation/src/openai_translation_error.rs
new file mode 100644
index 000000000..b99003fcc
--- /dev/null
+++ b/paddler_openai_translation/src/openai_translation_error.rs
@@ -0,0 +1,38 @@
+use std::time::SystemTimeError;
+
+use serde_json::Error as SerdeJsonError;
+
+use paddler_messaging::request_params_validation_error::RequestParamsValidationError;
+
+use crate::chat_completion_message_role::ChatCompletionMessageRole;
+
+#[derive(Debug, thiserror::Error)]
+pub enum OpenAITranslationError {
+    #[error("the system clock reads a time before the Unix epoch")]
+    ClockBeforeUnixEpoch(#[source] SystemTimeError),
+
+    #[error("a '{role}' message must have content")]
+    MessageWithoutContent { role: ChatCompletionMessageRole },
+
+    #[error("only an assistant message can carry reasoning, not a '{role}' message")]
+    ReasoningOnNonAssistantMessage { role: ChatCompletionMessageRole },
+
+    #[error("a reasoning item must carry its reasoning text")]
+    ReasoningItemWithoutContent,
+
+    #[error("the arguments of the tool call '{tool_call_id}' are not a JSON object")]
+    ToolCallArgumentsAreNotAJsonObject {
+        tool_call_id: String,
+        #[source]
+        source: SerdeJsonError,
+    },
+
+    #[error("only an assistant message can request tool calls, not a '{role}' message")]
+    ToolCallsOnNonAssistantMessage { role: ChatCompletionMessageRole },
+
+    #[error(transparent)]
+    ToolRejected(#[from] RequestParamsValidationError),
+
+    #[error("a tool message must name the tool call it answers in 'tool_call_id'")]
+    ToolResultWithoutToolCallId,
+}
diff --git a/paddler_openai_translation/src/output_item_event.rs b/paddler_openai_translation/src/output_item_event.rs
new file mode 100644
index 000000000..6c630a346
--- /dev/null
+++ b/paddler_openai_translation/src/output_item_event.rs
@@ -0,0 +1,10 @@
+use serde::Serialize;
+
+use crate::responses_output_item::ResponsesOutputItem;
+
+#[derive(Clone, Debug, Serialize)]
+pub struct OutputItemEvent {
+    pub sequence_number: u64,
+    pub output_index: usize,
+    pub item: ResponsesOutputItem,
+}
diff --git a/paddler_openai_translation/src/reasoning_effort.rs b/paddler_openai_translation/src/reasoning_effort.rs
new file mode 100644
index 000000000..c80bc3bfa
--- /dev/null
+++ b/paddler_openai_translation/src/reasoning_effort.rs
@@ -0,0 +1,32 @@
+use serde::Deserialize;
+
+#[derive(Clone, Copy, Debug, Deserialize, Eq, PartialEq)]
+#[serde(rename_all = "lowercase")]
+pub enum ReasoningEffort {
+    None,
+    Minimal,
+    Low,
+    Medium,
+    High,
+    Xhigh,
+}
+
+impl ReasoningEffort {
+    #[must_use]
+    pub const fn enables_thinking(self) -> bool {
+        match self {
+            Self::None => false,
+            Self::Minimal | Self::Low | Self::Medium | Self::High | Self::Xhigh => true,
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::ReasoningEffort;
+
+    #[test]
+    fn minimal_effort_still_enables_thinking() {
+        assert!(ReasoningEffort::Minimal.enables_thinking());
+    }
+}
diff --git a/paddler_openai_translation/src/response_snapshot_event.rs b/paddler_openai_translation/src/response_snapshot_event.rs
new file mode 100644
index 000000000..0d7bb14e8
--- /dev/null
+++ b/paddler_openai_translation/src/response_snapshot_event.rs
@@ -0,0 +1,9 @@
+use serde::Serialize;
+
+use crate::responses_response::ResponsesResponse;
+
+#[derive(Clone, Debug, Serialize)]
+pub struct ResponseSnapshotEvent {
+    pub sequence_number: u64,
+    pub response: ResponsesResponse,
+}
diff --git a/paddler_openai_translation/src/responses_assistant_content_part.rs b/paddler_openai_translation/src/responses_assistant_content_part.rs
new file mode 100644
index 000000000..b89f78880
--- /dev/null
+++ b/paddler_openai_translation/src/responses_assistant_content_part.rs
@@ -0,0 +1,23 @@
+use serde::Deserialize;
+
+use paddler_messaging::conversation_message_content_part::ConversationMessageContentPart;
+
+#[derive(Deserialize)]
+#[serde(tag = "type")]
+pub enum ResponsesAssistantContentPart {
+    #[serde(rename = "input_text")]
+    InputText { text: String },
+    #[serde(rename = "output_text")]
+    OutputText { text: String },
+}
+
+impl ResponsesAssistantContentPart {
+    #[must_use]
+    pub fn into_conversation_part(self) -> ConversationMessageContentPart {
+        match self {
+            Self::InputText { text } | Self::OutputText { text } => {
+                ConversationMessageContentPart::Text { text }
+            }
+        }
+    }
+}
diff --git a/paddler_openai_translation/src/responses_assistant_message_content.rs b/paddler_openai_translation/src/responses_assistant_message_content.rs
new file mode 100644
index 000000000..81210a458
--- /dev/null
+++ b/paddler_openai_translation/src/responses_assistant_message_content.rs
@@ -0,0 +1,27 @@
+use serde::Deserialize;
+
+use paddler_messaging::conversation_message_content::ConversationMessageContent;
+
+use crate::responses_assistant_content_part::ResponsesAssistantContentPart;
+
+#[derive(Deserialize)]
+#[serde(untagged)]
+pub enum ResponsesAssistantMessageContent {
+    Text(String),
+    Parts(Vec<ResponsesAssistantContentPart>),
+}
+
+impl ResponsesAssistantMessageContent {
+    #[must_use]
+    pub fn into_conversation_content(self) -> ConversationMessageContent {
+        match self {
+            Self::Text(text) => ConversationMessageContent::Text(text),
+            Self::Parts(parts) => ConversationMessageContent::Parts(
+                parts
+                    .into_iter()
+                    .map(ResponsesAssistantContentPart::into_conversation_part)
+                    .collect(),
+            ),
+        }
+    }
+}
diff --git a/paddler_balancer/src/compatibility/openai_service/responses_content_part.rs b/paddler_openai_translation/src/responses_content_part.rs
similarity index 100%
rename from paddler_balancer/src/compatibility/openai_service/responses_content_part.rs
rename to paddler_openai_translation/src/responses_content_part.rs
diff --git a/paddler_openai_translation/src/responses_delivery.rs b/paddler_openai_translation/src/responses_delivery.rs
new file mode 100644
index 000000000..2fb7b28a0
--- /dev/null
+++ b/paddler_openai_translation/src/responses_delivery.rs
@@ -0,0 +1,7 @@
+use crate::responses_response_header::ResponsesResponseHeader;
+use crate::responses_stream::ResponsesStream;
+
+pub enum ResponsesDelivery {
+    Buffered(ResponsesResponseHeader),
+    Streamed(ResponsesStream),
+}
diff --git a/paddler_openai_translation/src/responses_function_call_item.rs b/paddler_openai_translation/src/responses_function_call_item.rs
new file mode 100644
index 000000000..6f4b3598d
--- /dev/null
+++ b/paddler_openai_translation/src/responses_function_call_item.rs
@@ -0,0 +1,27 @@
+use serde::Deserialize;
+
+use paddler_messaging::conversation_tool_call::ConversationToolCall;
+
+use crate::conversation_tool_call_from_arguments::conversation_tool_call_from_arguments;
+use crate::openai_translation_error::OpenAITranslationError;
+
+#[derive(Deserialize)]
+pub struct ResponsesFunctionCallItem {
+    pub call_id: String,
+    pub name: String,
+    pub arguments: String,
+}
+
+impl ResponsesFunctionCallItem {
+    pub fn into_conversation_tool_call(
+        self,
+    ) -> Result<ConversationToolCall, OpenAITranslationError> {
+        let Self {
+            call_id,
+            name,
+            arguments,
+        } = self;
+
+        conversation_tool_call_from_arguments(call_id, name, &arguments)
+    }
+}
diff --git a/paddler_openai_translation/src/responses_function_call_output_item.rs b/paddler_openai_translation/src/responses_function_call_output_item.rs
new file mode 100644
index 000000000..243b1aec2
--- /dev/null
+++ b/paddler_openai_translation/src/responses_function_call_output_item.rs
@@ -0,0 +1,9 @@
+use serde::Deserialize;
+
+use crate::responses_function_output::ResponsesFunctionOutput;
+
+#[derive(Deserialize)]
+pub struct ResponsesFunctionCallOutputItem {
+    pub call_id: String,
+    pub output: ResponsesFunctionOutput,
+}
diff --git a/paddler_openai_translation/src/responses_function_output.rs b/paddler_openai_translation/src/responses_function_output.rs
new file mode 100644
index 000000000..f8122d9e4
--- /dev/null
+++ b/paddler_openai_translation/src/responses_function_output.rs
@@ -0,0 +1,51 @@
+use serde::Deserialize;
+
+use crate::responses_function_output_part::ResponsesFunctionOutputPart;
+
+#[derive(Deserialize)]
+#[serde(untagged)]
+pub enum ResponsesFunctionOutput {
+    Text(String),
+    Parts(Vec<ResponsesFunctionOutputPart>),
+}
+
+impl ResponsesFunctionOutput {
+    #[must_use]
+    pub fn into_text(self) -> String {
+        match self {
+            Self::Text(text) => text,
+            Self::Parts(parts) => parts
+                .into_iter()
+                .map(|ResponsesFunctionOutputPart::InputText { text }| text)
+                .collect::<String>(),
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::ResponsesFunctionOutput;
+    use crate::responses_function_output_part::ResponsesFunctionOutputPart;
+
+    #[test]
+    fn text_output_returns_its_text() {
+        assert_eq!(
+            ResponsesFunctionOutput::Text("done".to_owned()).into_text(),
+            "done"
+        );
+    }
+
+    #[test]
+    fn parts_output_concatenates_its_text() {
+        let output = ResponsesFunctionOutput::Parts(vec![
+            ResponsesFunctionOutputPart::InputText {
+                text: "foo".to_owned(),
+            },
+            ResponsesFunctionOutputPart::InputText {
+                text: "bar".to_owned(),
+            },
+        ]);
+
+        assert_eq!(output.into_text(), "foobar");
+    }
+}
diff --git a/paddler_openai_translation/src/responses_function_output_part.rs b/paddler_openai_translation/src/responses_function_output_part.rs
new file mode 100644
index 000000000..89bae359f
--- /dev/null
+++ b/paddler_openai_translation/src/responses_function_output_part.rs
@@ -0,0 +1,8 @@
+use serde::Deserialize;
+
+#[derive(Deserialize)]
+#[serde(tag = "type")]
+pub enum ResponsesFunctionOutputPart {
+    #[serde(rename = "input_text")]
+    InputText { text: String },
+}
diff --git a/paddler_openai_translation/src/responses_input.rs b/paddler_openai_translation/src/responses_input.rs
new file mode 100644
index 000000000..ca9d20540
--- /dev/null
+++ b/paddler_openai_translation/src/responses_input.rs
@@ -0,0 +1,34 @@
+use serde::Deserialize;
+use serde::Deserializer;
+
+use crate::responses_input_item::ResponsesInputItem;
+
+pub enum ResponsesInput {
+    Text(String),
+    Items(Vec<ResponsesInputItem>),
+}
+
+impl Default for ResponsesInput {
+    fn default() -> Self {
+        Self::Items(Vec::new())
+    }
+}
+
+impl<'deserializer> Deserialize<'deserializer> for ResponsesInput {
+    fn deserialize<TDeserializer>(deserializer: TDeserializer) -> Result<Self, TDeserializer::Error>
+    where
+        TDeserializer: Deserializer<'deserializer>,
+    {
+        #[derive(Deserialize)]
+        #[serde(untagged)]
+        enum TextOrItems {
+            Text(String),
+            Items(Vec<ResponsesInputItem>),
+        }
+
+        Ok(match TextOrItems::deserialize(deserializer)? {
+            TextOrItems::Text(text) => Self::Text(text),
+            TextOrItems::Items(items) => Self::Items(items),
+        })
+    }
+}
diff --git a/paddler_openai_translation/src/responses_input_content_part.rs b/paddler_openai_translation/src/responses_input_content_part.rs
new file mode 100644
index 000000000..469547e85
--- /dev/null
+++ b/paddler_openai_translation/src/responses_input_content_part.rs
@@ -0,0 +1,55 @@
+use serde::Deserialize;
+
+use paddler_messaging::conversation_message_content_part::ConversationMessageContentPart;
+use paddler_messaging::image_url::ImageUrl;
+
+#[derive(Deserialize)]
+#[serde(tag = "type")]
+pub enum ResponsesInputContentPart {
+    #[serde(rename = "input_text")]
+    InputText { text: String },
+    #[serde(rename = "input_image")]
+    InputImage { image_url: String },
+}
+
+impl ResponsesInputContentPart {
+    #[must_use]
+    pub fn into_conversation_part(self) -> ConversationMessageContentPart {
+        match self {
+            Self::InputText { text } => ConversationMessageContentPart::Text { text },
+            Self::InputImage { image_url } => ConversationMessageContentPart::ImageUrl {
+                image_url: ImageUrl { url: image_url },
+            },
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use paddler_messaging::conversation_message_content_part::ConversationMessageContentPart;
+
+    use super::ResponsesInputContentPart;
+
+    #[test]
+    fn input_text_becomes_text_part() {
+        assert!(matches!(
+            ResponsesInputContentPart::InputText {
+                text: "hello".to_owned(),
+            }
+            .into_conversation_part(),
+            ConversationMessageContentPart::Text { text } if text == "hello"
+        ));
+    }
+
+    #[test]
+    fn input_image_becomes_image_part() {
+        assert!(matches!(
+            ResponsesInputContentPart::InputImage {
+                image_url: "https://example.test/cat.png".to_owned(),
+            }
+            .into_conversation_part(),
+            ConversationMessageContentPart::ImageUrl { image_url }
+                if image_url.url == "https://example.test/cat.png"
+        ));
+    }
+}
diff --git a/paddler_openai_translation/src/responses_input_item.rs b/paddler_openai_translation/src/responses_input_item.rs
new file mode 100644
index 000000000..3fadc9d10
--- /dev/null
+++ b/paddler_openai_translation/src/responses_input_item.rs
@@ -0,0 +1,76 @@
+use serde::Deserialize;
+
+use paddler_messaging::assistant_role::ASSISTANT_ROLE;
+use paddler_messaging::conversation_message::ConversationMessage;
+use paddler_messaging::conversation_message_content::ConversationMessageContent;
+
+use crate::openai_translation_error::OpenAITranslationError;
+use crate::responses_function_call_output_item::ResponsesFunctionCallOutputItem;
+use crate::responses_message_item::ResponsesMessageItem;
+use crate::responses_tagged_item::ResponsesTaggedItem;
+
+#[derive(Deserialize)]
+#[serde(untagged)]
+pub enum ResponsesInputItem {
+    Tagged(ResponsesTaggedItem),
+    Message(ResponsesMessageItem),
+}
+
+impl ResponsesInputItem {
+    pub fn append_to(
+        self,
+        messages: &mut Vec<ConversationMessage>,
+    ) -> Result<(), OpenAITranslationError> {
+        match self {
+            Self::Message(message) | Self::Tagged(ResponsesTaggedItem::Message(message)) => {
+                messages.push(message.into_conversation_message());
+            }
+            Self::Tagged(ResponsesTaggedItem::FunctionCall(function_call)) => {
+                let tool_call = function_call.into_conversation_tool_call()?;
+
+                let assistant_tool_calls_turn = match messages.pop() {
+                    Some(ConversationMessage::AssistantToolCalls {
+                        content,
+                        mut tool_calls,
+                    }) => {
+                        tool_calls.push(tool_call);
+
+                        ConversationMessage::AssistantToolCalls {
+                            content,
+                            tool_calls,
+                        }
+                    }
+                    Some(ConversationMessage::Message { content, role })
+                        if role == ASSISTANT_ROLE =>
+                    {
+                        ConversationMessage::AssistantToolCalls {
+                            content,
+                            tool_calls: vec![tool_call],
+                        }
+                    }
+                    preceding_message => {
+                        messages.extend(preceding_message);
+
+                        ConversationMessage::AssistantToolCalls {
+                            content: ConversationMessageContent::Text(String::new()),
+                            tool_calls: vec![tool_call],
+                        }
+                    }
+                };
+
+                messages.push(assistant_tool_calls_turn);
+            }
+            Self::Tagged(ResponsesTaggedItem::FunctionCallOutput(
+                ResponsesFunctionCallOutputItem { call_id, output },
+            )) => messages.push(ConversationMessage::ToolResult {
+                content: ConversationMessageContent::Text(output.into_text()),
+                tool_call_id: call_id,
+            }),
+            Self::Tagged(ResponsesTaggedItem::Reasoning(reasoning)) => {
+                messages.push(reasoning.into_conversation_message()?);
+            }
+        }
+
+        Ok(())
+    }
+}
diff --git a/paddler_balancer/src/compatibility/openai_service/responses_item_status.rs b/paddler_openai_translation/src/responses_item_status.rs
similarity index 100%
rename from paddler_balancer/src/compatibility/openai_service/responses_item_status.rs
rename to paddler_openai_translation/src/responses_item_status.rs
diff --git a/paddler_openai_translation/src/responses_message_content.rs b/paddler_openai_translation/src/responses_message_content.rs
new file mode 100644
index 000000000..06def88a1
--- /dev/null
+++ b/paddler_openai_translation/src/responses_message_content.rs
@@ -0,0 +1,27 @@
+use serde::Deserialize;
+
+use paddler_messaging::conversation_message_content::ConversationMessageContent;
+
+use crate::responses_input_content_part::ResponsesInputContentPart;
+
+#[derive(Deserialize)]
+#[serde(untagged)]
+pub enum ResponsesMessageContent {
+    Text(String),
+    Parts(Vec<ResponsesInputContentPart>),
+}
+
+impl ResponsesMessageContent {
+    #[must_use]
+    pub fn into_conversation_content(self) -> ConversationMessageContent {
+        match self {
+            Self::Text(text) => ConversationMessageContent::Text(text),
+            Self::Parts(parts) => ConversationMessageContent::Parts(
+                parts
+                    .into_iter()
+                    .map(ResponsesInputContentPart::into_conversation_part)
+                    .collect(),
+            ),
+        }
+    }
+}
diff --git a/paddler_openai_translation/src/responses_message_item.rs b/paddler_openai_translation/src/responses_message_item.rs
new file mode 100644
index 000000000..5dc42b69f
--- /dev/null
+++ b/paddler_openai_translation/src/responses_message_item.rs
@@ -0,0 +1,48 @@
+use serde::Deserialize;
+
+use paddler_messaging::assistant_role::ASSISTANT_ROLE;
+use paddler_messaging::conversation_message::ConversationMessage;
+use paddler_messaging::system_role::SYSTEM_ROLE;
+use paddler_messaging::user_role::USER_ROLE;
+
+use crate::responses_assistant_message_content::ResponsesAssistantMessageContent;
+use crate::responses_message_content::ResponsesMessageContent;
+
+#[derive(Deserialize)]
+#[serde(tag = "role", rename_all = "lowercase")]
+pub enum ResponsesMessageItem {
+    Assistant {
+        content: ResponsesAssistantMessageContent,
+    },
+    Developer {
+        content: ResponsesMessageContent,
+    },
+    System {
+        content: ResponsesMessageContent,
+    },
+    User {
+        content: ResponsesMessageContent,
+    },
+}
+
+impl ResponsesMessageItem {
+    #[must_use]
+    pub fn into_conversation_message(self) -> ConversationMessage {
+        match self {
+            Self::Assistant { content } => ConversationMessage::Message {
+                content: content.into_conversation_content(),
+                role: ASSISTANT_ROLE.to_owned(),
+            },
+            Self::Developer { content } | Self::System { content } => {
+                ConversationMessage::Message {
+                    content: content.into_conversation_content(),
+                    role: SYSTEM_ROLE.to_owned(),
+                }
+            }
+            Self::User { content } => ConversationMessage::Message {
+                content: content.into_conversation_content(),
+                role: USER_ROLE.to_owned(),
+            },
+        }
+    }
+}
diff --git a/paddler_balancer/src/compatibility/openai_service/responses_output_item.rs b/paddler_openai_translation/src/responses_output_item.rs
similarity index 76%
rename from paddler_balancer/src/compatibility/openai_service/responses_output_item.rs
rename to paddler_openai_translation/src/responses_output_item.rs
index 78434a4ca..80279f44d 100644
--- a/paddler_balancer/src/compatibility/openai_service/responses_output_item.rs
+++ b/paddler_openai_translation/src/responses_output_item.rs
@@ -1,11 +1,11 @@
-use llama_cpp_bindings_types::ParsedToolCall;
 use serde::Serialize;
 
-use crate::compatibility::openai_service::arguments_to_tool_call_string::arguments_to_tool_call_string;
-use crate::compatibility::openai_service::assistant_role::ASSISTANT_ROLE;
-use crate::compatibility::openai_service::responses_content_part::ResponsesContentPart;
-use crate::compatibility::openai_service::responses_item_status::ResponsesItemStatus;
-use crate::compatibility::openai_service::responses_reasoning_part::ResponsesReasoningPart;
+use paddler_messaging::assistant_role::ASSISTANT_ROLE;
+use paddler_messaging::generated_tool_call::GeneratedToolCall;
+
+use crate::responses_content_part::ResponsesContentPart;
+use crate::responses_item_status::ResponsesItemStatus;
+use crate::responses_reasoning_part::ResponsesReasoningPart;
 
 #[derive(Clone, Debug, Serialize)]
 #[serde(rename_all = "snake_case", tag = "type")]
@@ -33,9 +33,13 @@ pub enum ResponsesOutputItem {
 
 impl ResponsesOutputItem {
     #[must_use]
-    pub fn completed_function_call(id: String, parsed_tool_call: &ParsedToolCall) -> Self {
+    pub fn completed_function_call(
+        id: String,
+        parsed_tool_call: &GeneratedToolCall,
+        arguments: String,
+    ) -> Self {
         Self::FunctionCall {
-            arguments: arguments_to_tool_call_string(&parsed_tool_call.arguments),
+            arguments,
             call_id: parsed_tool_call.id.clone(),
             id,
             name: parsed_tool_call.name.clone(),
@@ -44,7 +48,7 @@ impl ResponsesOutputItem {
     }
 
     #[must_use]
-    pub fn function_call_in_progress(id: String, parsed_tool_call: &ParsedToolCall) -> Self {
+    pub fn function_call_in_progress(id: String, parsed_tool_call: &GeneratedToolCall) -> Self {
         Self::FunctionCall {
             arguments: String::new(),
             call_id: parsed_tool_call.id.clone(),
diff --git a/paddler_openai_translation/src/responses_output_item_kind.rs b/paddler_openai_translation/src/responses_output_item_kind.rs
new file mode 100644
index 000000000..4cbd9d5a5
--- /dev/null
+++ b/paddler_openai_translation/src/responses_output_item_kind.rs
@@ -0,0 +1,45 @@
+use nanoid::nanoid;
+
+#[derive(Clone, Copy)]
+pub enum ResponsesOutputItemKind {
+    FunctionCall,
+    Message,
+    Reasoning,
+}
+
+impl ResponsesOutputItemKind {
+    #[must_use]
+    pub fn item_id(self) -> String {
+        let prefix = match self {
+            Self::FunctionCall => "fc",
+            Self::Message => "msg",
+            Self::Reasoning => "rs",
+        };
+
+        format!("{prefix}_{}", nanoid!())
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::ResponsesOutputItemKind;
+
+    #[test]
+    fn names_each_output_item_after_its_kind() {
+        assert!(
+            ResponsesOutputItemKind::Reasoning
+                .item_id()
+                .starts_with("rs_")
+        );
+        assert!(
+            ResponsesOutputItemKind::Message
+                .item_id()
+                .starts_with("msg_")
+        );
+        assert!(
+            ResponsesOutputItemKind::FunctionCall
+                .item_id()
+                .starts_with("fc_")
+        );
+    }
+}
diff --git a/paddler_openai_translation/src/responses_reasoning.rs b/paddler_openai_translation/src/responses_reasoning.rs
new file mode 100644
index 000000000..5f18b6954
--- /dev/null
+++ b/paddler_openai_translation/src/responses_reasoning.rs
@@ -0,0 +1,15 @@
+use serde::Deserialize;
+
+use crate::reasoning_effort::ReasoningEffort;
+
+#[derive(Deserialize)]
+pub struct ResponsesReasoning {
+    pub effort: Option<ReasoningEffort>,
+}
+
+impl ResponsesReasoning {
+    #[must_use]
+    pub fn enables_thinking(&self) -> bool {
+        self.effort.is_none_or(ReasoningEffort::enables_thinking)
+    }
+}
diff --git a/paddler_openai_translation/src/responses_reasoning_item.rs b/paddler_openai_translation/src/responses_reasoning_item.rs
new file mode 100644
index 000000000..31cf4e6b8
--- /dev/null
+++ b/paddler_openai_translation/src/responses_reasoning_item.rs
@@ -0,0 +1,27 @@
+use serde::Deserialize;
+
+use paddler_messaging::conversation_message::ConversationMessage;
+
+use crate::openai_translation_error::OpenAITranslationError;
+use crate::responses_reasoning_text_part::ResponsesReasoningTextPart;
+
+#[derive(Deserialize)]
+pub struct ResponsesReasoningItem {
+    pub content: Vec<ResponsesReasoningTextPart>,
+}
+
+impl ResponsesReasoningItem {
+    pub fn into_conversation_message(self) -> Result<ConversationMessage, OpenAITranslationError> {
+        if self.content.is_empty() {
+            return Err(OpenAITranslationError::ReasoningItemWithoutContent);
+        }
+
+        Ok(ConversationMessage::AssistantReasoning {
+            reasoning_content: self
+                .content
+                .into_iter()
+                .map(|ResponsesReasoningTextPart::ReasoningText { text }| text)
+                .collect(),
+        })
+    }
+}
diff --git a/paddler_balancer/src/compatibility/openai_service/responses_reasoning_part.rs b/paddler_openai_translation/src/responses_reasoning_part.rs
similarity index 100%
rename from paddler_balancer/src/compatibility/openai_service/responses_reasoning_part.rs
rename to paddler_openai_translation/src/responses_reasoning_part.rs
diff --git a/paddler_openai_translation/src/responses_reasoning_text_part.rs b/paddler_openai_translation/src/responses_reasoning_text_part.rs
new file mode 100644
index 000000000..0694ef29a
--- /dev/null
+++ b/paddler_openai_translation/src/responses_reasoning_text_part.rs
@@ -0,0 +1,8 @@
+use serde::Deserialize;
+
+#[derive(Deserialize)]
+#[serde(tag = "type")]
+pub enum ResponsesReasoningTextPart {
+    #[serde(rename = "reasoning_text")]
+    ReasoningText { text: String },
+}
diff --git a/paddler_openai_translation/src/responses_request.rs b/paddler_openai_translation/src/responses_request.rs
new file mode 100644
index 000000000..ad487cc3d
--- /dev/null
+++ b/paddler_openai_translation/src/responses_request.rs
@@ -0,0 +1,636 @@
+use std::num::NonZeroU32;
+use std::time::SystemTime;
+
+use nanoid::nanoid;
+use serde::Deserialize;
+
+use paddler_inference_parameters::sampling_parameters::SamplingParameters;
+use paddler_messaging::conversation_history::ConversationHistory;
+use paddler_messaging::conversation_message::ConversationMessage;
+use paddler_messaging::conversation_message_content::ConversationMessageContent;
+use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
+use paddler_messaging::system_role::SYSTEM_ROLE;
+use paddler_messaging::user_role::USER_ROLE;
+use paddler_messaging::validates::Validates;
+
+use crate::default_max_tokens::DEFAULT_MAX_TOKENS;
+use crate::openai_translation_error::OpenAITranslationError;
+use crate::responses_delivery::ResponsesDelivery;
+use crate::responses_input::ResponsesInput;
+use crate::responses_reasoning::ResponsesReasoning;
+use crate::responses_response_header::ResponsesResponseHeader;
+use crate::responses_stream::ResponsesStream;
+use crate::responses_text_param::ResponsesTextParam;
+use crate::responses_tool::ResponsesTool;
+use crate::timestamp_from::timestamp_from;
+use crate::translated_responses_request::TranslatedResponsesRequest;
+
+#[derive(Deserialize)]
+pub struct ResponsesRequest {
+    pub model: String,
+    #[serde(default)]
+    pub input: ResponsesInput,
+    #[serde(default)]
+    pub instructions: Option<String>,
+    #[serde(default)]
+    pub stream: Option<bool>,
+    #[serde(default)]
+    pub max_output_tokens: Option<NonZeroU32>,
+    #[serde(default)]
+    pub tools: Vec<ResponsesTool>,
+    #[serde(default)]
+    pub text: Option<ResponsesTextParam>,
+    #[serde(default)]
+    pub reasoning: Option<ResponsesReasoning>,
+}
+
+impl ResponsesRequest {
+    pub fn translate(
+        self,
+        now: SystemTime,
+        SamplingParameters {
+            temperature, top_p, ..
+        }: &SamplingParameters,
+    ) -> Result<TranslatedResponsesRequest, OpenAITranslationError> {
+        let Self {
+            model,
+            input,
+            instructions,
+            stream,
+            max_output_tokens,
+            tools,
+            text,
+            reasoning,
+        } = self;
+
+        let mut messages: Vec<ConversationMessage> = Vec::new();
+
+        if let Some(instructions) = &instructions
+            && !instructions.is_empty()
+        {
+            messages.push(ConversationMessage::Message {
+                content: ConversationMessageContent::Text(instructions.clone()),
+                role: SYSTEM_ROLE.to_owned(),
+            });
+        }
+
+        match input {
+            ResponsesInput::Text(text) => messages.push(ConversationMessage::Message {
+                content: ConversationMessageContent::Text(text),
+                role: USER_ROLE.to_owned(),
+            }),
+            ResponsesInput::Items(items) => {
+                for item in items {
+                    item.append_to(&mut messages)?;
+                }
+            }
+        }
+
+        let validated_tools = tools
+            .into_iter()
+            .map(|ResponsesTool::Function(function_definition)| function_definition.into_tool())
+            .map(Validates::validate)
+            .collect::<Result<Vec<_>, _>>()?;
+
+        let created_at = timestamp_from(now)?;
+        let header = ResponsesResponseHeader {
+            created_at,
+            id: format!("resp_{}", nanoid!()),
+            instructions,
+            model,
+            temperature: *temperature,
+            top_p: *top_p,
+        };
+
+        Ok(TranslatedResponsesRequest {
+            conversation_history_params: ContinueFromConversationHistoryParams {
+                add_generation_prompt: true,
+                conversation_history: ConversationHistory::new(messages),
+                enable_thinking: reasoning
+                    .as_ref()
+                    .is_none_or(ResponsesReasoning::enables_thinking),
+                grammar: text.and_then(ResponsesTextParam::into_grammar_constraint),
+                max_tokens: max_output_tokens.unwrap_or(DEFAULT_MAX_TOKENS),
+                parse_tool_calls: !validated_tools.is_empty(),
+                tools: validated_tools,
+            },
+            delivery: if stream.unwrap_or(false) {
+                ResponsesDelivery::Streamed(ResponsesStream::new(header))
+            } else {
+                ResponsesDelivery::Buffered(header)
+            },
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::num::NonZeroU32;
+    use std::time::UNIX_EPOCH;
+
+    use serde_json::Value;
+    use serde_json::error::Category;
+    use serde_json::from_value;
+    use serde_json::json;
+
+    use paddler_inference_parameters::sampling_parameters::SamplingParameters;
+    use paddler_messaging::conversation_message::ConversationMessage;
+    use paddler_messaging::conversation_message_content::ConversationMessageContent;
+    use paddler_messaging::conversation_message_content_part::ConversationMessageContentPart;
+    use paddler_messaging::conversation_tool_call::ConversationToolCall;
+    use paddler_messaging::grammar_constraint::GrammarConstraint;
+    use paddler_messaging::image_url::ImageUrl;
+    use paddler_messaging::request_params::continue_from_conversation_history_params::tool::Tool;
+    use paddler_messaging::request_params_validation_error::RequestParamsValidationError;
+
+    use super::ResponsesRequest;
+    use crate::openai_translation_error::OpenAITranslationError;
+    use crate::translated_responses_request::TranslatedResponsesRequest;
+
+    fn text_message(role: &str, text: &str) -> ConversationMessage {
+        ConversationMessage::Message {
+            content: ConversationMessageContent::Text(text.to_owned()),
+            role: role.to_owned(),
+        }
+    }
+
+    fn weather_call(call_id: &str, location: &str) -> ConversationToolCall {
+        ConversationToolCall {
+            arguments: json!({ "location": location }).as_object().unwrap().clone(),
+            id: call_id.to_owned(),
+            name: "get_weather".to_owned(),
+        }
+    }
+
+    fn translated_from(value: Value) -> TranslatedResponsesRequest {
+        let params: ResponsesRequest = from_value(value).unwrap();
+
+        params
+            .translate(UNIX_EPOCH, &SamplingParameters::default())
+            .unwrap()
+    }
+
+    #[test]
+    fn string_input_becomes_a_single_user_message() {
+        let translated = translated_from(json!({ "model": "test", "input": "Say hello" }));
+
+        let messages = &translated
+            .conversation_history_params
+            .conversation_history
+            .messages;
+
+        assert_eq!(*messages, vec![text_message("user", "Say hello")]);
+    }
+
+    #[test]
+    fn instructions_are_prepended_as_a_system_message() {
+        let translated = translated_from(json!({
+            "model": "test",
+            "instructions": "be terse",
+            "input": "hi"
+        }));
+
+        let messages = &translated
+            .conversation_history_params
+            .conversation_history
+            .messages;
+
+        assert_eq!(
+            *messages,
+            vec![
+                text_message("system", "be terse"),
+                text_message("user", "hi")
+            ]
+        );
+    }
+
+    #[test]
+    fn a_function_call_output_keeps_its_call_id() {
+        let translated = translated_from(json!({
+            "model": "test",
+            "input": [
+                { "type": "function_call_output", "call_id": "call_1", "output": "sunny" }
+            ]
+        }));
+
+        let messages = &translated
+            .conversation_history_params
+            .conversation_history
+            .messages;
+
+        assert_eq!(
+            *messages,
+            vec![ConversationMessage::ToolResult {
+                content: ConversationMessageContent::Text("sunny".to_owned()),
+                tool_call_id: "call_1".to_owned(),
+            }]
+        );
+    }
+
+    #[test]
+    fn a_function_call_item_becomes_an_assistant_tool_call_turn() {
+        let translated = translated_from(json!({
+            "model": "test",
+            "input": [
+                { "type": "function_call", "call_id": "call_1", "name": "get_weather", "arguments": "{\"location\":\"Paris\"}" }
+            ]
+        }));
+
+        let messages = &translated
+            .conversation_history_params
+            .conversation_history
+            .messages;
+
+        assert_eq!(
+            *messages,
+            vec![ConversationMessage::AssistantToolCalls {
+                content: ConversationMessageContent::Text(String::new()),
+                tool_calls: vec![weather_call("call_1", "Paris")],
+            }]
+        );
+    }
+
+    #[test]
+    fn consecutive_function_calls_join_the_assistant_message_before_them_in_one_turn() {
+        let translated = translated_from(json!({
+            "model": "test",
+            "input": [
+                { "type": "message", "role": "user", "content": "weather in Paris and Rome?" },
+                { "type": "message", "role": "assistant", "content": "Checking both." },
+                { "type": "function_call", "call_id": "call_1", "name": "get_weather", "arguments": "{\"location\":\"Paris\"}" },
+                { "type": "function_call", "call_id": "call_2", "name": "get_weather", "arguments": "{\"location\":\"Rome\"}" }
+            ]
+        }));
+
+        assert_eq!(
+            translated
+                .conversation_history_params
+                .conversation_history
+                .messages,
+            vec![
+                text_message("user", "weather in Paris and Rome?"),
+                ConversationMessage::AssistantToolCalls {
+                    content: ConversationMessageContent::Text("Checking both.".to_owned()),
+                    tool_calls: vec![
+                        weather_call("call_1", "Paris"),
+                        weather_call("call_2", "Rome")
+                    ],
+                },
+            ]
+        );
+    }
+
+    #[test]
+    fn function_call_arguments_that_are_not_a_json_object_are_refused() {
+        let params: ResponsesRequest = from_value(json!({
+            "model": "test",
+            "input": [
+                { "type": "function_call", "call_id": "call_1", "name": "get_weather", "arguments": "not json" }
+            ]
+        }))
+        .unwrap();
+
+        assert!(matches!(
+            params.translate(UNIX_EPOCH, &SamplingParameters::default()),
+            Err(OpenAITranslationError::ToolCallArgumentsAreNotAJsonObject { tool_call_id, .. })
+                if tool_call_id == "call_1"
+        ));
+    }
+
+    #[test]
+    fn message_content_parts_become_conversation_parts() {
+        let translated = translated_from(json!({
+            "model": "test",
+            "input": [
+                {
+                    "type": "message",
+                    "role": "user",
+                    "content": [
+                        { "type": "input_text", "text": "What is this?" },
+                        { "type": "input_image", "image_url": "https://example.test/cat.png" }
+                    ]
+                }
+            ]
+        }));
+
+        assert_eq!(
+            translated
+                .conversation_history_params
+                .conversation_history
+                .messages,
+            vec![ConversationMessage::Message {
+                content: ConversationMessageContent::Parts(vec![
+                    ConversationMessageContentPart::Text {
+                        text: "What is this?".to_owned(),
+                    },
+                    ConversationMessageContentPart::ImageUrl {
+                        image_url: ImageUrl {
+                            url: "https://example.test/cat.png".to_owned(),
+                        },
+                    },
+                ]),
+                role: "user".to_owned(),
+            }]
+        );
+    }
+
+    #[test]
+    fn text_format_text_leaves_the_output_unconstrained() {
+        let translated = translated_from(json!({
+            "model": "test",
+            "input": "hi",
+            "text": { "format": { "type": "text" } }
+        }));
+
+        assert_eq!(translated.conversation_history_params.grammar, None);
+    }
+
+    #[test]
+    fn developer_and_system_roles_become_system_messages() {
+        let translated = translated_from(json!({
+            "model": "test",
+            "input": [
+                { "type": "message", "role": "developer", "content": "rules" },
+                { "type": "message", "role": "system", "content": "more rules" }
+            ]
+        }));
+
+        assert_eq!(
+            translated
+                .conversation_history_params
+                .conversation_history
+                .messages,
+            vec![
+                text_message("system", "rules"),
+                text_message("system", "more rules")
+            ]
+        );
+    }
+
+    #[test]
+    fn an_assistant_message_written_as_input_text_becomes_assistant_text() {
+        let translated = translated_from(json!({
+            "model": "test",
+            "input": [
+                {
+                    "type": "message",
+                    "role": "assistant",
+                    "content": [ { "type": "input_text", "text": "Hello again." } ]
+                }
+            ]
+        }));
+
+        assert_eq!(
+            translated
+                .conversation_history_params
+                .conversation_history
+                .messages,
+            vec![ConversationMessage::Message {
+                content: ConversationMessageContent::Parts(vec![
+                    ConversationMessageContentPart::Text {
+                        text: "Hello again.".to_owned(),
+                    },
+                ]),
+                role: "assistant".to_owned(),
+            }]
+        );
+    }
+
+    #[test]
+    fn flat_function_tool_maps_to_an_internal_tool_with_default_description() {
+        let translated = translated_from(json!({
+            "model": "test",
+            "input": "hi",
+            "tools": [
+                { "type": "function", "name": "get_weather", "parameters": { "type": "object" } }
+            ]
+        }));
+
+        assert!(translated.conversation_history_params.parse_tool_calls);
+
+        let Tool::Function(function_call) = &translated.conversation_history_params.tools[0];
+
+        assert_eq!(function_call.function.name, "get_weather");
+        assert_eq!(function_call.function.description, "");
+    }
+
+    #[test]
+    fn text_format_json_schema_becomes_a_grammar_constraint() {
+        let translated = translated_from(json!({
+            "model": "test",
+            "input": "hi",
+            "text": { "format": { "type": "json_schema", "name": "out", "schema": { "type": "object" } } }
+        }));
+
+        assert_eq!(
+            translated.conversation_history_params.grammar,
+            Some(GrammarConstraint::JsonSchema {
+                schema: r#"{"type":"object"}"#.to_owned(),
+            })
+        );
+    }
+
+    #[test]
+    fn a_request_without_an_output_limit_generates_until_its_context_is_full() {
+        let translated = translated_from(json!({ "model": "test", "input": "hi" }));
+
+        assert_eq!(
+            translated.conversation_history_params.max_tokens,
+            NonZeroU32::MAX
+        );
+    }
+
+    #[test]
+    fn reasoning_effort_none_disables_thinking() {
+        let translated = translated_from(json!({
+            "model": "test",
+            "input": "hi",
+            "reasoning": { "effort": "none" }
+        }));
+
+        assert!(!translated.conversation_history_params.enable_thinking);
+    }
+
+    #[test]
+    fn rejects_input_it_cannot_represent() {
+        let unrepresentable_requests = [
+            json!({ "model": "test", "input": "hi", "tools": [ { "type": "web_search" } ] }),
+            json!({ "model": "test", "input": [ { "type": "reasoning", "summary": [] } ] }),
+            json!({ "model": "test", "input": [ { "type": "message", "role": "user", "content": [ { "type": "input_file", "file_id": "file_1" } ] } ] }),
+            json!({ "model": "test", "input": [ { "type": "message", "role": "user", "content": [ { "type": "input_image", "file_id": "file_1" } ] } ] }),
+            json!({ "model": "test", "input": [ { "type": "function_call_output", "call_id": "call_1", "output": [ { "type": "input_image", "image_url": "https://example.test/cat.png" } ] } ] }),
+            json!({ "model": "test", "input": "hi", "text": { "format": { "type": "json_object" } } }),
+            json!({ "model": "test", "input": [ { "type": "message", "role": "user", "content": [ { "type": "output_text", "text": "hi" } ] } ] }),
+            json!({ "model": "test", "input": [ { "type": "message", "role": "tool", "content": "hi" } ] }),
+        ];
+
+        for unrepresentable_request in unrepresentable_requests {
+            let rejection = from_value::<ResponsesRequest>(unrepresentable_request)
+                .err()
+                .map(|deserialization_error| deserialization_error.classify());
+
+            assert_eq!(rejection, Some(Category::Data));
+        }
+    }
+
+    #[test]
+    fn unsupported_and_stateful_fields_are_ignored() {
+        let translated = translated_from(json!({
+            "model": "test",
+            "input": "hi",
+            "store": true,
+            "previous_response_id": "resp_prev",
+            "conversation": "conv_1",
+            "temperature": 0.5,
+            "tool_choice": "required"
+        }));
+
+        assert_eq!(
+            translated
+                .conversation_history_params
+                .conversation_history
+                .messages
+                .len(),
+            1
+        );
+    }
+
+    #[test]
+    fn instructions_alone_become_the_whole_conversation() {
+        let translated = translated_from(json!({ "model": "test", "instructions": "be terse" }));
+
+        let messages = &translated
+            .conversation_history_params
+            .conversation_history
+            .messages;
+
+        assert_eq!(messages.len(), 1);
+        assert_eq!(*messages, vec![text_message("system", "be terse")]);
+    }
+
+    #[test]
+    fn message_item_without_a_type_becomes_a_conversation_message() {
+        let translated = translated_from(json!({
+            "model": "test",
+            "input": [ { "role": "user", "content": "hi" } ]
+        }));
+
+        let messages = &translated
+            .conversation_history_params
+            .conversation_history
+            .messages;
+
+        assert_eq!(*messages, vec![text_message("user", "hi")]);
+    }
+
+    #[test]
+    fn rejects_a_function_tool_that_requires_an_undeclared_property() {
+        let params: ResponsesRequest = from_value(json!({
+            "model": "test",
+            "input": "hi",
+            "tools": [
+                {
+                    "type": "function",
+                    "name": "broken",
+                    "parameters": {
+                        "type": "object",
+                        "properties": { "present": { "type": "string" } },
+                        "required": ["absent"]
+                    }
+                }
+            ]
+        }))
+        .unwrap();
+
+        assert_eq!(
+            params
+                .translate(UNIX_EPOCH, &SamplingParameters::default())
+                .err()
+                .as_ref()
+                .map(ToString::to_string),
+            Some(
+                OpenAITranslationError::ToolRejected(
+                    RequestParamsValidationError::RequiredFieldNotInProperties {
+                        field: "absent".to_owned(),
+                    }
+                )
+                .to_string()
+            )
+        );
+    }
+
+    #[test]
+    fn its_own_output_replayed_as_input_becomes_the_previous_assistant_turn() {
+        let translated = translated_from(json!({
+            "model": "test",
+            "input": [
+                { "type": "message", "role": "user", "content": "What is the weather in Paris?" },
+                {
+                    "type": "reasoning",
+                    "id": "rs_1",
+                    "status": "completed",
+                    "summary": [],
+                    "content": [ { "type": "reasoning_text", "text": "I should check the weather." } ]
+                },
+                {
+                    "type": "message",
+                    "id": "msg_1",
+                    "role": "assistant",
+                    "status": "completed",
+                    "content": [ { "type": "output_text", "annotations": [], "logprobs": [], "text": "Checking." } ]
+                },
+                {
+                    "type": "function_call",
+                    "id": "fc_1",
+                    "status": "completed",
+                    "call_id": "call_0",
+                    "name": "get_weather",
+                    "arguments": "{\"location\":\"Paris\"}"
+                },
+                { "type": "function_call_output", "call_id": "call_0", "output": "sunny" }
+            ]
+        }));
+
+        assert_eq!(
+            translated
+                .conversation_history_params
+                .conversation_history
+                .messages,
+            vec![
+                text_message("user", "What is the weather in Paris?"),
+                ConversationMessage::AssistantReasoning {
+                    reasoning_content: "I should check the weather.".to_owned(),
+                },
+                ConversationMessage::AssistantToolCalls {
+                    content: ConversationMessageContent::Parts(vec![
+                        ConversationMessageContentPart::Text {
+                            text: "Checking.".to_owned(),
+                        },
+                    ]),
+                    tool_calls: vec![weather_call("call_0", "Paris")],
+                },
+                ConversationMessage::ToolResult {
+                    content: ConversationMessageContent::Text("sunny".to_owned()),
+                    tool_call_id: "call_0".to_owned(),
+                },
+            ]
+        );
+    }
+
+    #[test]
+    fn a_reasoning_item_without_reasoning_text_is_refused() {
+        let params: ResponsesRequest = from_value(json!({
+            "model": "test",
+            "input": [ { "type": "reasoning", "summary": [], "content": [] } ]
+        }))
+        .unwrap();
+
+        assert_eq!(
+            params
+                .translate(UNIX_EPOCH, &SamplingParameters::default())
+                .err()
+                .map(|translation_error| translation_error.to_string()),
+            Some(OpenAITranslationError::ReasoningItemWithoutContent.to_string())
+        );
+    }
+}
diff --git a/paddler_balancer/src/compatibility/openai_service/responses_response.rs b/paddler_openai_translation/src/responses_response.rs
similarity index 89%
rename from paddler_balancer/src/compatibility/openai_service/responses_response.rs
rename to paddler_openai_translation/src/responses_response.rs
index 815b7c31a..f7aa07cd0 100644
--- a/paddler_balancer/src/compatibility/openai_service/responses_response.rs
+++ b/paddler_openai_translation/src/responses_response.rs
@@ -1,11 +1,11 @@
 use serde::Serialize;
 use serde::Serializer;
 
-use crate::compatibility::openai_service::responses_output_item::ResponsesOutputItem;
-use crate::compatibility::openai_service::responses_response_header::ResponsesResponseHeader;
-use crate::compatibility::openai_service::responses_response_progress::ResponsesResponseProgress;
-use crate::compatibility::openai_service::responses_response_status::ResponsesResponseStatus;
-use crate::compatibility::openai_service::responses_usage::ResponsesUsage;
+use crate::responses_output_item::ResponsesOutputItem;
+use crate::responses_response_header::ResponsesResponseHeader;
+use crate::responses_response_progress::ResponsesResponseProgress;
+use crate::responses_response_status::ResponsesResponseStatus;
+use crate::responses_usage::ResponsesUsage;
 
 fn serialized_progress(progress: &ResponsesResponseProgress) -> SerializedProgress<'_> {
     match progress {
diff --git a/paddler_openai_translation/src/responses_response_header.rs b/paddler_openai_translation/src/responses_response_header.rs
new file mode 100644
index 000000000..42222e2f5
--- /dev/null
+++ b/paddler_openai_translation/src/responses_response_header.rs
@@ -0,0 +1,269 @@
+use llama_cpp_bindings_types::TokenUsage;
+
+use paddler_messaging::generation_finish::GenerationFinish;
+use paddler_messaging::generation_summary::GenerationSummary;
+
+use crate::arguments_to_tool_call_string::arguments_to_tool_call_string;
+use crate::completes_generation::CompletesGeneration;
+use crate::generated_output::GeneratedOutput;
+use crate::responses_output_item::ResponsesOutputItem;
+use crate::responses_output_item_kind::ResponsesOutputItemKind;
+use crate::responses_response::ResponsesResponse;
+use crate::responses_response_progress::ResponsesResponseProgress;
+use crate::responses_usage::ResponsesUsage;
+
+#[derive(Clone, Debug)]
+pub struct ResponsesResponseHeader {
+    pub created_at: u64,
+    pub id: String,
+    pub instructions: Option<String>,
+    pub model: String,
+    pub temperature: f32,
+    pub top_p: f32,
+}
+
+impl ResponsesResponseHeader {
+    #[must_use]
+    pub fn in_progress(&self) -> ResponsesResponse {
+        self.with_progress(ResponsesResponseProgress::InProgress)
+    }
+
+    #[must_use]
+    pub fn finished(
+        &self,
+        output: Vec<ResponsesOutputItem>,
+        usage: &TokenUsage,
+        finish: GenerationFinish,
+    ) -> ResponsesResponse {
+        let usage = ResponsesUsage(*usage);
+
+        self.with_progress(if finish.reached_a_length_limit() {
+            ResponsesResponseProgress::Incomplete { output, usage }
+        } else {
+            ResponsesResponseProgress::Completed { output, usage }
+        })
+    }
+
+    #[must_use]
+    pub fn failed(&self, error_message: String) -> ResponsesResponse {
+        self.with_progress(ResponsesResponseProgress::Failed { error_message })
+    }
+
+    fn with_progress(&self, progress: ResponsesResponseProgress) -> ResponsesResponse {
+        ResponsesResponse {
+            header: self.clone(),
+            progress,
+        }
+    }
+}
+
+impl CompletesGeneration for ResponsesResponseHeader {
+    type Completion = ResponsesResponse;
+
+    fn complete(
+        &self,
+        _request_id: &str,
+        GeneratedOutput {
+            content,
+            reasoning,
+            tool_calls,
+        }: GeneratedOutput,
+        GenerationSummary { finish, usage }: &GenerationSummary,
+    ) -> ResponsesResponse {
+        let mut output: Vec<ResponsesOutputItem> = Vec::new();
+
+        if !reasoning.is_empty() {
+            output.push(ResponsesOutputItem::completed_reasoning(
+                ResponsesOutputItemKind::Reasoning.item_id(),
+                reasoning,
+            ));
+        }
+
+        if !content.is_empty() || tool_calls.is_empty() {
+            output.push(ResponsesOutputItem::completed_message(
+                ResponsesOutputItemKind::Message.item_id(),
+                content,
+            ));
+        }
+
+        for parsed_call in &tool_calls {
+            output.push(ResponsesOutputItem::completed_function_call(
+                ResponsesOutputItemKind::FunctionCall.item_id(),
+                parsed_call,
+                arguments_to_tool_call_string(&parsed_call.arguments),
+            ));
+        }
+
+        self.finished(output, usage, *finish)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use llama_cpp_bindings_types::TokenUsage;
+    use llama_cpp_bindings_types::ToolCallArguments;
+    use serde_json::Value;
+    use serde_json::json;
+    use serde_json::to_value;
+
+    use paddler_messaging::generated_tool_call::GeneratedToolCall;
+    use paddler_messaging::generation_finish::GenerationFinish;
+    use paddler_messaging::generation_summary::GenerationSummary;
+    use paddler_openai_response_format_validator::openai_validator::OpenAIValidator;
+
+    use super::ResponsesResponseHeader;
+    use crate::completes_generation::CompletesGeneration as _;
+    use crate::generated_output::GeneratedOutput;
+    use crate::generated_output_part::GeneratedOutputPart;
+    use crate::responses_output_item::ResponsesOutputItem;
+
+    fn header() -> ResponsesResponseHeader {
+        ResponsesResponseHeader {
+            id: "resp_test".to_owned(),
+            created_at: 1_234,
+            model: "test-model".to_owned(),
+            instructions: None,
+            temperature: 0.25,
+            top_p: 0.5,
+        }
+    }
+
+    #[test]
+    fn in_progress_reports_in_progress_status_with_no_output() {
+        let response = to_value(header().in_progress()).unwrap();
+
+        assert_eq!(response["status"], "in_progress");
+        assert_eq!(response["object"], "response");
+        assert_eq!(response["output"], json!([]));
+        assert!(response.get("usage").is_none());
+    }
+
+    #[test]
+    fn reports_the_sampling_parameters_in_effect() {
+        let response = to_value(header().in_progress()).unwrap();
+
+        assert_eq!(response["temperature"], 0.25);
+        assert_eq!(response["top_p"], 0.5);
+    }
+
+    #[test]
+    fn failed_carries_the_error_message() {
+        let response = to_value(header().failed("boom".to_owned())).unwrap();
+
+        assert_eq!(response["status"], "failed");
+        assert_eq!(response["error"]["message"], "boom");
+        assert!(response.get("usage").is_none());
+    }
+
+    #[test]
+    fn completed_includes_usage_and_output() {
+        let response = to_value(header().finished(
+            vec![ResponsesOutputItem::completed_message(
+                "msg_0".to_owned(),
+                "hello".to_owned(),
+            )],
+            &TokenUsage {
+                prompt_tokens: 3,
+                cached_prompt_tokens: 0,
+                input_image_tokens: 0,
+                input_audio_tokens: 0,
+                content_tokens: 5,
+                reasoning_tokens: 2,
+                tool_call_tokens: 0,
+                undeterminable_tokens: 0,
+            },
+            GenerationFinish::EndOfGeneration,
+        ))
+        .unwrap();
+
+        assert_eq!(response["status"], "completed");
+        assert_eq!(response["usage"]["input_tokens"], 3);
+        assert_eq!(response["output"][0]["type"], "message");
+    }
+
+    fn completed_response_of(generated_output_parts: Vec<GeneratedOutputPart>) -> Value {
+        let mut generated_output = GeneratedOutput::default();
+
+        for generated_output_part in generated_output_parts {
+            generated_output.append(generated_output_part);
+        }
+
+        to_value(header().complete(
+            "test-request",
+            generated_output,
+            &GenerationSummary {
+                finish: GenerationFinish::EndOfGeneration,
+                usage: TokenUsage {
+                    prompt_tokens: 5,
+                    content_tokens: 3,
+                    reasoning_tokens: 2,
+                    ..TokenUsage::default()
+                },
+            },
+        ))
+        .unwrap()
+    }
+
+    fn weather_call() -> GeneratedToolCall {
+        GeneratedToolCall {
+            arguments: ToolCallArguments::ValidJson(json!({ "location": "Paris" })),
+            id: "call_x".to_owned(),
+            index: 0,
+            name: "get_weather".to_owned(),
+        }
+    }
+
+    #[test]
+    fn completes_content_into_a_message_item() {
+        let response = completed_response_of(vec![
+            GeneratedOutputPart::Content("hel".to_owned()),
+            GeneratedOutputPart::Content("lo".to_owned()),
+        ]);
+
+        assert_eq!(response["id"], "resp_test");
+        assert_eq!(response["status"], "completed");
+        assert_eq!(response["output"][0]["type"], "message");
+        assert_eq!(response["output"][0]["content"][0]["text"], "hello");
+    }
+
+    #[test]
+    fn two_responses_never_share_an_output_item_id() {
+        let first_response =
+            completed_response_of(vec![GeneratedOutputPart::Content("hello".to_owned())]);
+        let second_response =
+            completed_response_of(vec![GeneratedOutputPart::Content("hello".to_owned())]);
+
+        assert_ne!(
+            first_response["output"][0]["id"],
+            second_response["output"][0]["id"]
+        );
+    }
+
+    #[test]
+    fn completes_reasoning_and_tool_calls_without_an_empty_message() {
+        let response = completed_response_of(vec![
+            GeneratedOutputPart::Reasoning("ponder".to_owned()),
+            GeneratedOutputPart::ToolCalls(vec![weather_call()]),
+        ]);
+
+        assert_eq!(response["output"][0]["type"], "reasoning");
+        assert_eq!(response["output"][1]["type"], "function_call");
+        assert_eq!(response["output"][1]["name"], "get_weather");
+        assert_eq!(
+            response["usage"]["output_tokens_details"]["reasoning_tokens"],
+            2
+        );
+    }
+
+    #[test]
+    fn the_completed_response_conforms_to_the_schema() {
+        OpenAIValidator::new()
+            .unwrap()
+            .validate_responses_response(&completed_response_of(vec![
+                GeneratedOutputPart::Reasoning("p".to_owned()),
+                GeneratedOutputPart::Content("hello".to_owned()),
+                GeneratedOutputPart::ToolCalls(vec![weather_call()]),
+            ]))
+            .unwrap();
+    }
+}
diff --git a/paddler_openai_translation/src/responses_response_progress.rs b/paddler_openai_translation/src/responses_response_progress.rs
new file mode 100644
index 000000000..03dea62cc
--- /dev/null
+++ b/paddler_openai_translation/src/responses_response_progress.rs
@@ -0,0 +1,18 @@
+use crate::responses_output_item::ResponsesOutputItem;
+use crate::responses_usage::ResponsesUsage;
+
+#[derive(Clone, Debug)]
+pub enum ResponsesResponseProgress {
+    Completed {
+        output: Vec<ResponsesOutputItem>,
+        usage: ResponsesUsage,
+    },
+    Failed {
+        error_message: String,
+    },
+    Incomplete {
+        output: Vec<ResponsesOutputItem>,
+        usage: ResponsesUsage,
+    },
+    InProgress,
+}
diff --git a/paddler_balancer/src/compatibility/openai_service/responses_response_status.rs b/paddler_openai_translation/src/responses_response_status.rs
similarity index 100%
rename from paddler_balancer/src/compatibility/openai_service/responses_response_status.rs
rename to paddler_openai_translation/src/responses_response_status.rs
diff --git a/paddler_openai_translation/src/responses_stream.rs b/paddler_openai_translation/src/responses_stream.rs
new file mode 100644
index 000000000..c21cc09a0
--- /dev/null
+++ b/paddler_openai_translation/src/responses_stream.rs
@@ -0,0 +1,633 @@
+use std::mem::take;
+
+use paddler_messaging::generated_tool_call::GeneratedToolCall;
+use paddler_messaging::generation_summary::GenerationSummary;
+
+use crate::arguments_to_tool_call_string::arguments_to_tool_call_string;
+use crate::content_part_event::ContentPartEvent;
+use crate::function_call_arguments_delta_event::FunctionCallArgumentsDeltaEvent;
+use crate::function_call_arguments_done_event::FunctionCallArgumentsDoneEvent;
+use crate::generated_output_part::GeneratedOutputPart;
+use crate::generation_failure::GenerationFailure;
+use crate::generation_outcome::GenerationOutcome;
+use crate::generation_progress::GenerationProgress;
+use crate::open_item::OpenItem;
+use crate::output_item_event::OutputItemEvent;
+use crate::response_snapshot_event::ResponseSnapshotEvent;
+use crate::responses_content_part::ResponsesContentPart;
+use crate::responses_output_item::ResponsesOutputItem;
+use crate::responses_output_item_kind::ResponsesOutputItemKind;
+use crate::responses_response_header::ResponsesResponseHeader;
+use crate::responses_stream_event::ResponsesStreamEvent;
+use crate::text_delta_event::TextDeltaEvent;
+use crate::text_done_event::TextDoneEvent;
+
+pub struct ResponsesStream {
+    finalized_output: Vec<ResponsesOutputItem>,
+    header: ResponsesResponseHeader,
+    open: OpenItem,
+    output_index: usize,
+    sequence_number: u64,
+    started: bool,
+}
+
+impl ResponsesStream {
+    #[must_use]
+    pub const fn new(header: ResponsesResponseHeader) -> Self {
+        Self {
+            finalized_output: Vec::new(),
+            header,
+            open: OpenItem::Nothing,
+            output_index: 0,
+            sequence_number: 0,
+            started: false,
+        }
+    }
+
+    pub fn advance(
+        &mut self,
+        generation_progress: GenerationProgress,
+    ) -> Vec<ResponsesStreamEvent> {
+        let mut events: Vec<ResponsesStreamEvent> = Vec::new();
+
+        match generation_progress {
+            GenerationProgress::Produced(GeneratedOutputPart::Content(text)) => {
+                self.ensure_preamble(&mut events);
+                self.handle_content(&mut events, &text);
+            }
+            GenerationProgress::Produced(GeneratedOutputPart::Reasoning(text)) => {
+                self.ensure_preamble(&mut events);
+                self.handle_reasoning(&mut events, &text);
+            }
+            GenerationProgress::Produced(GeneratedOutputPart::ToolCalls(parsed_calls)) => {
+                self.ensure_preamble(&mut events);
+                self.handle_tool_calls(&mut events, &parsed_calls);
+            }
+            GenerationProgress::ToolCallTokenProduced => {}
+            GenerationProgress::Failed(GenerationFailure { message, .. }) => {
+                return self.fail(message);
+            }
+        }
+
+        events
+    }
+
+    pub fn conclude(&mut self, generation_outcome: GenerationOutcome) -> Vec<ResponsesStreamEvent> {
+        match generation_outcome {
+            GenerationOutcome::Failed(GenerationFailure { message, .. }) => self.fail(message),
+            GenerationOutcome::Finished(generation_summary) => {
+                let mut events: Vec<ResponsesStreamEvent> = Vec::new();
+
+                self.ensure_preamble(&mut events);
+                self.handle_done(&mut events, &generation_summary);
+
+                events
+            }
+        }
+    }
+
+    pub fn fail(&mut self, error_message: String) -> Vec<ResponsesStreamEvent> {
+        let mut events: Vec<ResponsesStreamEvent> = Vec::new();
+
+        self.ensure_preamble(&mut events);
+
+        let failed_sequence_number = self.next_sequence_number();
+
+        events.push(ResponsesStreamEvent::Failed(ResponseSnapshotEvent {
+            sequence_number: failed_sequence_number,
+            response: self.header.failed(error_message),
+        }));
+
+        events
+    }
+
+    fn ensure_preamble(&mut self, events: &mut Vec<ResponsesStreamEvent>) {
+        if self.started {
+            return;
+        }
+
+        self.started = true;
+
+        let in_progress_response = self.header.in_progress();
+        let created_sequence_number = self.next_sequence_number();
+
+        events.push(ResponsesStreamEvent::Created(ResponseSnapshotEvent {
+            sequence_number: created_sequence_number,
+            response: in_progress_response.clone(),
+        }));
+
+        let in_progress_sequence_number = self.next_sequence_number();
+
+        events.push(ResponsesStreamEvent::InProgress(ResponseSnapshotEvent {
+            sequence_number: in_progress_sequence_number,
+            response: in_progress_response,
+        }));
+    }
+
+    fn handle_done(
+        &mut self,
+        events: &mut Vec<ResponsesStreamEvent>,
+        GenerationSummary { finish, usage }: &GenerationSummary,
+    ) {
+        self.close_open_item(events);
+
+        let output = take(&mut self.finalized_output);
+        let snapshot = ResponseSnapshotEvent {
+            sequence_number: self.next_sequence_number(),
+            response: self.header.finished(output, usage, *finish),
+        };
+
+        events.push(if finish.reached_a_length_limit() {
+            ResponsesStreamEvent::Incomplete(snapshot)
+        } else {
+            ResponsesStreamEvent::Completed(snapshot)
+        });
+    }
+
+    const fn next_sequence_number(&mut self) -> u64 {
+        let sequence_number = self.sequence_number;
+        self.sequence_number += 1;
+
+        sequence_number
+    }
+
+    fn close_open_item(&mut self, events: &mut Vec<ResponsesStreamEvent>) {
+        let output_index = self.output_index;
+        let item = match take(&mut self.open) {
+            OpenItem::Nothing => return,
+            OpenItem::Reasoning { item_id, text } => {
+                let text_done_sequence_number = self.next_sequence_number();
+                events.push(ResponsesStreamEvent::ReasoningTextDone(TextDoneEvent {
+                    sequence_number: text_done_sequence_number,
+                    item_id: item_id.clone(),
+                    output_index,
+                    content_index: 0,
+                    text: text.clone(),
+                }));
+
+                ResponsesOutputItem::completed_reasoning(item_id, text)
+            }
+            OpenItem::Message { item_id, text } => {
+                let text_done_sequence_number = self.next_sequence_number();
+                events.push(ResponsesStreamEvent::OutputTextDone(TextDoneEvent {
+                    sequence_number: text_done_sequence_number,
+                    item_id: item_id.clone(),
+                    output_index,
+                    content_index: 0,
+                    text: text.clone(),
+                }));
+
+                let part_done_sequence_number = self.next_sequence_number();
+                events.push(ResponsesStreamEvent::ContentPartDone(ContentPartEvent {
+                    sequence_number: part_done_sequence_number,
+                    item_id: item_id.clone(),
+                    output_index,
+                    content_index: 0,
+                    part: ResponsesContentPart::output_text(text.clone()),
+                }));
+
+                ResponsesOutputItem::completed_message(item_id, text)
+            }
+        };
+
+        self.finish_item(events, item);
+    }
+
+    fn handle_reasoning(&mut self, events: &mut Vec<ResponsesStreamEvent>, text: &str) {
+        let item_id = if let OpenItem::Reasoning {
+            item_id,
+            text: reasoning_text,
+        } = &mut self.open
+        {
+            reasoning_text.push_str(text);
+
+            item_id.clone()
+        } else {
+            self.close_open_item(events);
+
+            let item_id = ResponsesOutputItemKind::Reasoning.item_id();
+
+            self.add_item(
+                events,
+                ResponsesOutputItem::reasoning_in_progress(item_id.clone()),
+            );
+            self.open = OpenItem::Reasoning {
+                item_id: item_id.clone(),
+                text: text.to_owned(),
+            };
+
+            item_id
+        };
+
+        let delta_sequence_number = self.next_sequence_number();
+        events.push(ResponsesStreamEvent::ReasoningTextDelta(TextDeltaEvent {
+            sequence_number: delta_sequence_number,
+            item_id,
+            output_index: self.output_index,
+            content_index: 0,
+            delta: text.to_owned(),
+        }));
+    }
+
+    fn handle_content(&mut self, events: &mut Vec<ResponsesStreamEvent>, text: &str) {
+        let item_id = if let OpenItem::Message {
+            item_id,
+            text: message_text,
+        } = &mut self.open
+        {
+            message_text.push_str(text);
+
+            item_id.clone()
+        } else {
+            self.close_open_item(events);
+
+            let item_id = ResponsesOutputItemKind::Message.item_id();
+
+            self.add_item(
+                events,
+                ResponsesOutputItem::message_in_progress(item_id.clone()),
+            );
+
+            let part_added_sequence_number = self.next_sequence_number();
+            events.push(ResponsesStreamEvent::ContentPartAdded(ContentPartEvent {
+                sequence_number: part_added_sequence_number,
+                item_id: item_id.clone(),
+                output_index: self.output_index,
+                content_index: 0,
+                part: ResponsesContentPart::output_text(String::new()),
+            }));
+            self.open = OpenItem::Message {
+                item_id: item_id.clone(),
+                text: text.to_owned(),
+            };
+
+            item_id
+        };
+
+        let delta_sequence_number = self.next_sequence_number();
+        events.push(ResponsesStreamEvent::OutputTextDelta(TextDeltaEvent {
+            sequence_number: delta_sequence_number,
+            item_id,
+            output_index: self.output_index,
+            content_index: 0,
+            delta: text.to_owned(),
+        }));
+    }
+
+    fn handle_tool_calls(
+        &mut self,
+        events: &mut Vec<ResponsesStreamEvent>,
+        parsed_calls: &[GeneratedToolCall],
+    ) {
+        self.close_open_item(events);
+
+        for call in parsed_calls {
+            let output_index = self.output_index;
+            let item_id = ResponsesOutputItemKind::FunctionCall.item_id();
+            let arguments = arguments_to_tool_call_string(&call.arguments);
+
+            self.add_item(
+                events,
+                ResponsesOutputItem::function_call_in_progress(item_id.clone(), call),
+            );
+
+            let delta_sequence_number = self.next_sequence_number();
+            events.push(ResponsesStreamEvent::FunctionCallArgumentsDelta(
+                FunctionCallArgumentsDeltaEvent {
+                    sequence_number: delta_sequence_number,
+                    item_id: item_id.clone(),
+                    output_index,
+                    delta: arguments.clone(),
+                },
+            ));
+
+            let done_sequence_number = self.next_sequence_number();
+            events.push(ResponsesStreamEvent::FunctionCallArgumentsDone(
+                FunctionCallArgumentsDoneEvent {
+                    sequence_number: done_sequence_number,
+                    item_id: item_id.clone(),
+                    output_index,
+                    name: call.name.clone(),
+                    arguments: arguments.clone(),
+                },
+            ));
+
+            self.finish_item(
+                events,
+                ResponsesOutputItem::completed_function_call(item_id, call, arguments),
+            );
+        }
+    }
+
+    fn add_item(&mut self, events: &mut Vec<ResponsesStreamEvent>, item: ResponsesOutputItem) {
+        let added_sequence_number = self.next_sequence_number();
+
+        events.push(ResponsesStreamEvent::OutputItemAdded(OutputItemEvent {
+            sequence_number: added_sequence_number,
+            output_index: self.output_index,
+            item,
+        }));
+    }
+
+    fn finish_item(&mut self, events: &mut Vec<ResponsesStreamEvent>, item: ResponsesOutputItem) {
+        let item_done_sequence_number = self.next_sequence_number();
+
+        events.push(ResponsesStreamEvent::OutputItemDone(OutputItemEvent {
+            sequence_number: item_done_sequence_number,
+            output_index: self.output_index,
+            item: item.clone(),
+        }));
+
+        self.finalized_output.push(item);
+        self.output_index += 1;
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use llama_cpp_bindings_types::TokenUsage;
+    use llama_cpp_bindings_types::ToolCallArguments;
+    use serde_json::Value;
+    use serde_json::json;
+    use serde_json::to_value;
+
+    use paddler_messaging::generated_tool_call::GeneratedToolCall;
+    use paddler_messaging::generation_finish::GenerationFinish;
+    use paddler_messaging::generation_summary::GenerationSummary;
+    use paddler_openai_response_format_validator::openai_validator::OpenAIValidator;
+
+    use super::ResponsesStream;
+    use crate::generated_output_part::GeneratedOutputPart;
+    use crate::generation_failure::GenerationFailure;
+    use crate::generation_failure_cause::GenerationFailureCause;
+    use crate::generation_outcome::GenerationOutcome;
+    use crate::generation_progress::GenerationProgress;
+    use crate::responses_response_header::ResponsesResponseHeader;
+    use crate::responses_stream_event::ResponsesStreamEvent;
+
+    fn responses_stream() -> ResponsesStream {
+        ResponsesStream::new(ResponsesResponseHeader {
+            id: "resp_test".to_owned(),
+            created_at: 0,
+            model: "test-model".to_owned(),
+            instructions: None,
+            temperature: 0.25,
+            top_p: 0.5,
+        })
+    }
+
+    fn content(text: &str) -> GenerationProgress {
+        GenerationProgress::Produced(GeneratedOutputPart::Content(text.to_owned()))
+    }
+
+    fn finished(finish: GenerationFinish) -> GenerationOutcome {
+        GenerationOutcome::Finished(GenerationSummary {
+            finish,
+            usage: TokenUsage {
+                prompt_tokens: 7,
+                content_tokens: 4,
+                reasoning_tokens: 1,
+                ..TokenUsage::default()
+            },
+        })
+    }
+
+    fn weather_call() -> GeneratedToolCall {
+        GeneratedToolCall {
+            arguments: ToolCallArguments::ValidJson(json!({ "location": "Paris" })),
+            id: "call_x".to_owned(),
+            index: 0,
+            name: "get_weather".to_owned(),
+        }
+    }
+
+    fn serialized(event: &ResponsesStreamEvent) -> Value {
+        to_value(event).unwrap()
+    }
+
+    fn names(events: &[ResponsesStreamEvent]) -> Vec<&'static str> {
+        events
+            .iter()
+            .map(ResponsesStreamEvent::event_name)
+            .collect()
+    }
+
+    #[test]
+    fn the_first_content_emits_the_preamble_then_a_text_delta() {
+        let events = responses_stream().advance(content("hi"));
+
+        assert_eq!(
+            names(&events),
+            vec![
+                "response.created",
+                "response.in_progress",
+                "response.output_item.added",
+                "response.content_part.added",
+                "response.output_text.delta",
+            ]
+        );
+        assert_eq!(serialized(&events[0])["response"]["status"], "in_progress");
+        assert_eq!(serialized(&events[4])["delta"], "hi");
+    }
+
+    #[test]
+    fn every_event_of_a_streamed_item_carries_the_same_id() {
+        let mut responses_stream = responses_stream();
+        let mut events = responses_stream.advance(content("hel"));
+
+        events.extend(responses_stream.advance(content("lo")));
+        events.extend(responses_stream.conclude(finished(GenerationFinish::EndOfGeneration)));
+
+        let item_ids = events
+            .iter()
+            .map(serialized)
+            .filter_map(|event| match event["type"].as_str() {
+                Some("response.output_item.added" | "response.output_item.done") => {
+                    Some(event["item"]["id"].clone())
+                }
+                Some("response.completed") => Some(event["response"]["output"][0]["id"].clone()),
+                Some("response.created" | "response.in_progress") => None,
+                Some(_) | None => Some(event["item_id"].clone()),
+            })
+            .collect::<Vec<_>>();
+
+        assert_eq!(item_ids.len(), 8);
+        assert!(item_ids.iter().all(|item_id| *item_id == item_ids[0]));
+    }
+
+    #[test]
+    fn the_preamble_is_emitted_only_once() {
+        let mut responses_stream = responses_stream();
+
+        responses_stream.advance(content("a"));
+
+        assert_eq!(
+            names(&responses_stream.advance(content("b"))),
+            vec!["response.output_text.delta"]
+        );
+    }
+
+    #[test]
+    fn tool_call_tokens_do_not_start_the_response() {
+        assert!(
+            responses_stream()
+                .advance(GenerationProgress::ToolCallTokenProduced)
+                .is_empty()
+        );
+    }
+
+    #[test]
+    fn finishing_closes_the_message_and_completes_with_usage() {
+        let mut responses_stream = responses_stream();
+
+        responses_stream.advance(content("hello"));
+
+        let events = responses_stream.conclude(finished(GenerationFinish::EndOfGeneration));
+
+        assert_eq!(
+            names(&events),
+            vec![
+                "response.output_text.done",
+                "response.content_part.done",
+                "response.output_item.done",
+                "response.completed",
+            ]
+        );
+
+        let completed = serialized(&events[3]);
+
+        assert_eq!(completed["response"]["status"], "completed");
+        assert_eq!(completed["response"]["usage"]["input_tokens"], 7);
+        assert_eq!(completed["response"]["usage"]["total_tokens"], 12);
+        assert_eq!(
+            completed["response"]["output"][0]["content"][0]["text"],
+            "hello"
+        );
+        assert_eq!(
+            completed["response"]["output"][0]["content"][0]["logprobs"],
+            json!([])
+        );
+    }
+
+    #[test]
+    fn finishing_at_a_length_limit_reports_an_incomplete_response() {
+        let events = responses_stream().conclude(finished(GenerationFinish::MaxTokens));
+
+        assert_eq!(
+            names(&events),
+            vec![
+                "response.created",
+                "response.in_progress",
+                "response.incomplete"
+            ]
+        );
+        assert_eq!(
+            serialized(&events[2])["response"]["incomplete_details"]["reason"],
+            "max_output_tokens"
+        );
+    }
+
+    #[test]
+    fn reasoning_then_content_closes_the_reasoning_item_first() {
+        let mut responses_stream = responses_stream();
+
+        responses_stream.advance(GenerationProgress::Produced(
+            GeneratedOutputPart::Reasoning("think".to_owned()),
+        ));
+
+        let events = responses_stream.advance(content("answer"));
+
+        assert_eq!(
+            names(&events),
+            vec![
+                "response.reasoning_text.done",
+                "response.output_item.done",
+                "response.output_item.added",
+                "response.content_part.added",
+                "response.output_text.delta",
+            ]
+        );
+        assert_eq!(serialized(&events[1])["output_index"], 0);
+        assert_eq!(serialized(&events[2])["output_index"], 1);
+        assert_eq!(serialized(&events[1])["item"]["type"], "reasoning");
+    }
+
+    #[test]
+    fn tool_calls_emit_function_call_argument_events_without_a_content_index() {
+        let events = responses_stream().advance(GenerationProgress::Produced(
+            GeneratedOutputPart::ToolCalls(vec![weather_call()]),
+        ));
+
+        assert_eq!(
+            names(&events),
+            vec![
+                "response.created",
+                "response.in_progress",
+                "response.output_item.added",
+                "response.function_call_arguments.delta",
+                "response.function_call_arguments.done",
+                "response.output_item.done",
+            ]
+        );
+
+        let delta_event = serialized(&events[3]);
+
+        assert_eq!(delta_event["delta"], "{\"location\":\"Paris\"}");
+        assert!(delta_event.get("content_index").is_none());
+        assert_eq!(serialized(&events[4])["name"], "get_weather");
+        assert_eq!(serialized(&events[5])["item"]["call_id"], "call_x");
+    }
+
+    #[test]
+    fn a_generation_failure_emits_the_preamble_then_response_failed() {
+        let generation_failure = || GenerationFailure {
+            cause: GenerationFailureCause::AgentFailed,
+            message: "boom".to_owned(),
+        };
+
+        for events in [
+            responses_stream().advance(GenerationProgress::Failed(generation_failure())),
+            responses_stream().conclude(GenerationOutcome::Failed(generation_failure())),
+        ] {
+            assert_eq!(
+                names(&events),
+                vec![
+                    "response.created",
+                    "response.in_progress",
+                    "response.failed"
+                ]
+            );
+
+            let failed = serialized(&events[2]);
+
+            assert_eq!(failed["response"]["status"], "failed");
+            assert_eq!(failed["response"]["error"]["code"], "server_error");
+            assert_eq!(failed["response"]["error"]["message"], "boom");
+        }
+    }
+
+    #[test]
+    fn every_emitted_event_conforms_to_the_schema() {
+        let validator = OpenAIValidator::new().unwrap();
+        let mut completed_stream = responses_stream();
+        let mut emitted: Vec<ResponsesStreamEvent> = Vec::new();
+
+        for generation_progress in [
+            GenerationProgress::Produced(GeneratedOutputPart::Reasoning("ponder".to_owned())),
+            content("hello"),
+            GenerationProgress::Produced(GeneratedOutputPart::ToolCalls(vec![weather_call()])),
+        ] {
+            emitted.extend(completed_stream.advance(generation_progress));
+        }
+
+        emitted.extend(completed_stream.conclude(finished(GenerationFinish::EndOfGeneration)));
+
+        emitted.extend(responses_stream().fail("boom".to_owned()));
+
+        for event in &emitted {
+            validator
+                .validate_responses_stream_event(&serialized(event))
+                .unwrap();
+        }
+    }
+}
diff --git a/paddler_balancer/src/compatibility/openai_service/responses_stream_event.rs b/paddler_openai_translation/src/responses_stream_event.rs
similarity index 90%
rename from paddler_balancer/src/compatibility/openai_service/responses_stream_event.rs
rename to paddler_openai_translation/src/responses_stream_event.rs
index dd392d909..682b72e54 100644
--- a/paddler_balancer/src/compatibility/openai_service/responses_stream_event.rs
+++ b/paddler_openai_translation/src/responses_stream_event.rs
@@ -1,13 +1,13 @@
 use serde::Serialize;
 use serde::Serializer;
 
-use crate::compatibility::openai_service::content_part_event::ContentPartEvent;
-use crate::compatibility::openai_service::function_call_arguments_delta_event::FunctionCallArgumentsDeltaEvent;
-use crate::compatibility::openai_service::function_call_arguments_done_event::FunctionCallArgumentsDoneEvent;
-use crate::compatibility::openai_service::output_item_event::OutputItemEvent;
-use crate::compatibility::openai_service::response_snapshot_event::ResponseSnapshotEvent;
-use crate::compatibility::openai_service::text_delta_event::TextDeltaEvent;
-use crate::compatibility::openai_service::text_done_event::TextDoneEvent;
+use crate::content_part_event::ContentPartEvent;
+use crate::function_call_arguments_delta_event::FunctionCallArgumentsDeltaEvent;
+use crate::function_call_arguments_done_event::FunctionCallArgumentsDoneEvent;
+use crate::output_item_event::OutputItemEvent;
+use crate::response_snapshot_event::ResponseSnapshotEvent;
+use crate::text_delta_event::TextDeltaEvent;
+use crate::text_done_event::TextDoneEvent;
 
 fn serialize_typed<TPayload, TSerializer>(
     event_type: &'static str,
@@ -144,7 +144,7 @@ mod tests {
     use super::ResponseSnapshotEvent;
     use super::ResponsesStreamEvent;
     use super::TextDeltaEvent;
-    use crate::compatibility::openai_service::responses_response_header::ResponsesResponseHeader;
+    use crate::responses_response_header::ResponsesResponseHeader;
 
     #[test]
     fn reasoning_and_text_delta_carry_their_distinct_event_names_with_the_same_payload_shape() {
diff --git a/paddler_openai_translation/src/responses_tagged_item.rs b/paddler_openai_translation/src/responses_tagged_item.rs
new file mode 100644
index 000000000..679a503a6
--- /dev/null
+++ b/paddler_openai_translation/src/responses_tagged_item.rs
@@ -0,0 +1,19 @@
+use serde::Deserialize;
+
+use crate::responses_function_call_item::ResponsesFunctionCallItem;
+use crate::responses_function_call_output_item::ResponsesFunctionCallOutputItem;
+use crate::responses_message_item::ResponsesMessageItem;
+use crate::responses_reasoning_item::ResponsesReasoningItem;
+
+#[derive(Deserialize)]
+#[serde(tag = "type")]
+pub enum ResponsesTaggedItem {
+    #[serde(rename = "message")]
+    Message(ResponsesMessageItem),
+    #[serde(rename = "function_call")]
+    FunctionCall(ResponsesFunctionCallItem),
+    #[serde(rename = "function_call_output")]
+    FunctionCallOutput(ResponsesFunctionCallOutputItem),
+    #[serde(rename = "reasoning")]
+    Reasoning(ResponsesReasoningItem),
+}
diff --git a/paddler_openai_translation/src/responses_text_format.rs b/paddler_openai_translation/src/responses_text_format.rs
new file mode 100644
index 000000000..fce0ce1c4
--- /dev/null
+++ b/paddler_openai_translation/src/responses_text_format.rs
@@ -0,0 +1,306 @@
+use std::fmt;
+use std::fmt::Formatter;
+
+use serde::Deserialize;
+use serde::Deserializer;
+use serde::de::Error as _;
+use serde::de::IgnoredAny;
+use serde::de::MapAccess;
+use serde::de::SeqAccess;
+use serde::de::Visitor;
+use serde_json::value::RawValue;
+
+use crate::responses_text_format_field::ResponsesTextFormatField;
+use crate::responses_text_format_refusal::ResponsesTextFormatRefusal;
+use crate::responses_text_format_schema_occurrences::ResponsesTextFormatSchemaOccurrences;
+use crate::responses_text_format_type::ResponsesTextFormatType;
+
+fn count_leftover_elements<'de, TSeqAccess>(
+    mut format_elements: TSeqAccess,
+) -> Result<usize, TSeqAccess::Error>
+where
+    TSeqAccess: SeqAccess<'de>,
+{
+    let mut leftover_elements = 0;
+
+    while format_elements.next_element::<IgnoredAny>()?.is_some() {
+        leftover_elements += 1;
+    }
+
+    Ok(leftover_elements)
+}
+
+fn without_leftover_elements(
+    format: ResponsesTextFormat,
+    consumed_elements: usize,
+    leftover_elements: usize,
+) -> Result<ResponsesTextFormat, ResponsesTextFormatRefusal> {
+    if leftover_elements == 0 {
+        Ok(format)
+    } else {
+        Err(ResponsesTextFormatRefusal::ElementsLeftOver {
+            consumed_elements,
+            leftover_elements,
+        })
+    }
+}
+
+struct ResponsesTextFormatVisitor;
+
+impl<'de> Visitor<'de> for ResponsesTextFormatVisitor {
+    type Value = Result<ResponsesTextFormat, ResponsesTextFormatRefusal>;
+
+    fn expecting(&self, formatter: &mut Formatter) -> fmt::Result {
+        formatter.write_str("internally tagged enum ResponsesTextFormat")
+    }
+
+    fn visit_seq<TSeqAccess>(
+        self,
+        mut format_elements: TSeqAccess,
+    ) -> Result<Self::Value, TSeqAccess::Error>
+    where
+        TSeqAccess: SeqAccess<'de>,
+    {
+        match format_elements.next_element::<ResponsesTextFormatType>()? {
+            None => Err(TSeqAccess::Error::missing_field("type")),
+            Some(ResponsesTextFormatType::Text) => Ok(without_leftover_elements(
+                ResponsesTextFormat::Text,
+                0,
+                count_leftover_elements(format_elements)?,
+            )),
+            Some(ResponsesTextFormatType::JsonSchema) => {
+                match format_elements.next_element::<Box<RawValue>>()? {
+                    None => Ok(Err(ResponsesTextFormatRefusal::SchemaElementMissing)),
+                    Some(schema) => Ok(without_leftover_elements(
+                        ResponsesTextFormat::JsonSchema { schema },
+                        1,
+                        count_leftover_elements(format_elements)?,
+                    )),
+                }
+            }
+        }
+    }
+
+    fn visit_map<TMapAccess>(
+        self,
+        mut format_fields: TMapAccess,
+    ) -> Result<Self::Value, TMapAccess::Error>
+    where
+        TMapAccess: MapAccess<'de>,
+    {
+        let mut format_type = None;
+        let mut schema_occurrences = ResponsesTextFormatSchemaOccurrences::Absent;
+
+        while let Some(format_field) = format_fields.next_key()? {
+            match format_field {
+                ResponsesTextFormatField::Schema => {
+                    schema_occurrences = schema_occurrences
+                        .with_another(format_fields.next_value::<Box<RawValue>>()?);
+                }
+                ResponsesTextFormatField::Type => {
+                    if format_type.is_some() {
+                        return Err(TMapAccess::Error::duplicate_field("type"));
+                    }
+
+                    format_type = Some(format_fields.next_value::<ResponsesTextFormatType>()?);
+                }
+                ResponsesTextFormatField::Unrecognized => {
+                    format_fields.next_value::<IgnoredAny>()?;
+                }
+            }
+        }
+
+        match format_type {
+            None => Err(TMapAccess::Error::missing_field("type")),
+            Some(ResponsesTextFormatType::Text) => Ok(Ok(ResponsesTextFormat::Text)),
+            Some(ResponsesTextFormatType::JsonSchema) => Ok(match schema_occurrences {
+                ResponsesTextFormatSchemaOccurrences::Absent => {
+                    Err(ResponsesTextFormatRefusal::SchemaMissing)
+                }
+                ResponsesTextFormatSchemaOccurrences::Once(schema) => {
+                    Ok(ResponsesTextFormat::JsonSchema { schema })
+                }
+                ResponsesTextFormatSchemaOccurrences::Repeated => {
+                    Err(ResponsesTextFormatRefusal::SchemaRepeated)
+                }
+            }),
+        }
+    }
+}
+
+pub enum ResponsesTextFormat {
+    Text,
+    JsonSchema { schema: Box<RawValue> },
+}
+
+impl<'de> Deserialize<'de> for ResponsesTextFormat {
+    fn deserialize<TDeserializer>(deserializer: TDeserializer) -> Result<Self, TDeserializer::Error>
+    where
+        TDeserializer: Deserializer<'de>,
+    {
+        deserializer
+            .deserialize_any(ResponsesTextFormatVisitor)?
+            .map_err(ResponsesTextFormatRefusal::into_deserialization_error)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use serde_json::from_str;
+
+    use super::ResponsesTextFormat;
+
+    #[derive(Debug, Eq, PartialEq)]
+    enum DeserializedFormat {
+        JsonSchema(String),
+        Text,
+    }
+
+    fn deserialized(format: &str) -> Result<DeserializedFormat, String> {
+        from_str::<ResponsesTextFormat>(format)
+            .map(|deserialized_format| match deserialized_format {
+                ResponsesTextFormat::JsonSchema { schema } => {
+                    DeserializedFormat::JsonSchema(schema.get().to_owned())
+                }
+                ResponsesTextFormat::Text => DeserializedFormat::Text,
+            })
+            .map_err(|deserialization_error| deserialization_error.to_string())
+    }
+
+    #[test]
+    fn keeps_the_schema_text_exactly_as_sent() {
+        let schema = r#"{ "type": "integer", "maximum": 1e3 }"#;
+
+        assert_eq!(
+            deserialized(&format!(
+                r#"{{"type": "json_schema", "name": "out", "schema": {schema}}}"#
+            )),
+            Ok(DeserializedFormat::JsonSchema(schema.to_owned()))
+        );
+    }
+
+    #[test]
+    fn accepts_the_array_form_of_every_format_type() {
+        assert_eq!(
+            [
+                deserialized(r#"["text"]"#),
+                deserialized(r#"["json_schema", {"type": "object"}]"#),
+            ],
+            [
+                Ok(DeserializedFormat::Text),
+                Ok(DeserializedFormat::JsonSchema(
+                    r#"{"type": "object"}"#.to_owned()
+                )),
+            ]
+        );
+    }
+
+    #[test]
+    fn ignores_repeated_schemas_in_a_text_format() {
+        assert_eq!(
+            deserialized(r#"{"type": "text", "schema": {}, "schema": {}}"#),
+            Ok(DeserializedFormat::Text)
+        );
+    }
+
+    #[test]
+    fn refuses_a_format_that_is_neither_an_object_nor_an_array() {
+        assert_eq!(
+            deserialized(r#""json_schema""#),
+            Err(
+                r#"invalid type: string "json_schema", expected internally tagged enum ResponsesTextFormat at line 1 column 13"#
+                    .to_owned()
+            )
+        );
+    }
+
+    #[test]
+    fn refuses_a_format_without_a_type() {
+        assert_eq!(
+            [deserialized(r#"{"schema": {}}"#), deserialized("[]")],
+            [
+                Err("missing field `type` at line 1 column 14".to_owned()),
+                Err("missing field `type` at line 1 column 2".to_owned()),
+            ]
+        );
+    }
+
+    #[test]
+    fn refuses_a_json_schema_format_without_a_schema() {
+        assert_eq!(
+            [
+                deserialized(r#"{"type": "json_schema"}"#),
+                deserialized(r#"["json_schema"]"#),
+            ],
+            [
+                Err("missing field `schema`".to_owned()),
+                Err(
+                    "invalid length 0, expected struct variant ResponsesTextFormat::JsonSchema with 1 element"
+                        .to_owned()
+                ),
+            ]
+        );
+    }
+
+    #[test]
+    fn refuses_a_format_that_repeats_its_type_or_its_json_schema() {
+        assert_eq!(
+            [
+                deserialized(r#"{"type": "text", "type": "text"}"#),
+                deserialized(r#"{"type": "json_schema", "schema": {}, "schema": {}}"#),
+            ],
+            [
+                Err("duplicate field `type` at line 1 column 23".to_owned()),
+                Err("duplicate field `schema`".to_owned()),
+            ]
+        );
+    }
+
+    #[test]
+    fn refuses_an_array_format_with_elements_left_over() {
+        assert_eq!(
+            [
+                deserialized(r#"["text", 1]"#),
+                deserialized(r#"["json_schema", {}, 1]"#),
+            ],
+            [
+                Err("invalid length 1, expected 0 elements in sequence".to_owned()),
+                Err("invalid length 2, expected 1 element in sequence".to_owned()),
+            ]
+        );
+    }
+
+    #[test]
+    fn refuses_a_format_whose_fields_are_malformed() {
+        assert_eq!(
+            [
+                deserialized(r#"{"type": "text", 1: 2}"#),
+                deserialized(r#"{"type": "json_schema", "schema": nul}"#),
+                deserialized(r#"{"type": "text", "name": nul}"#),
+            ],
+            [
+                Err("key must be a string at line 1 column 18".to_owned()),
+                Err("expected ident at line 1 column 38".to_owned()),
+                Err("expected ident at line 1 column 29".to_owned()),
+            ]
+        );
+    }
+
+    #[test]
+    fn refuses_an_array_format_whose_elements_are_malformed() {
+        assert_eq!(
+            [
+                deserialized("[nul]"),
+                deserialized(r#"["json_schema", nul]"#),
+                deserialized(r#"["text", nul]"#),
+                deserialized(r#"["json_schema", {}, nul]"#),
+            ],
+            [
+                Err("expected ident at line 1 column 5".to_owned()),
+                Err("expected ident at line 1 column 20".to_owned()),
+                Err("expected ident at line 1 column 13".to_owned()),
+                Err("expected ident at line 1 column 24".to_owned()),
+            ]
+        );
+    }
+}
diff --git a/paddler_openai_translation/src/responses_text_format_field.rs b/paddler_openai_translation/src/responses_text_format_field.rs
new file mode 100644
index 000000000..cf83bbd16
--- /dev/null
+++ b/paddler_openai_translation/src/responses_text_format_field.rs
@@ -0,0 +1,10 @@
+use serde::Deserialize;
+
+#[derive(Deserialize)]
+#[serde(field_identifier, rename_all = "snake_case")]
+pub enum ResponsesTextFormatField {
+    Schema,
+    Type,
+    #[serde(other)]
+    Unrecognized,
+}
diff --git a/paddler_openai_translation/src/responses_text_format_refusal.rs b/paddler_openai_translation/src/responses_text_format_refusal.rs
new file mode 100644
index 000000000..92106fca2
--- /dev/null
+++ b/paddler_openai_translation/src/responses_text_format_refusal.rs
@@ -0,0 +1,55 @@
+use std::fmt;
+use std::fmt::Formatter;
+
+use serde::de::Error;
+use serde::de::Expected;
+
+const JSON_SCHEMA_FORMAT_ELEMENTS: &str =
+    "struct variant ResponsesTextFormat::JsonSchema with 1 element";
+
+struct ElementsInSequence {
+    count: usize,
+}
+
+impl Expected for ElementsInSequence {
+    fn fmt(&self, formatter: &mut Formatter) -> fmt::Result {
+        if self.count == 1 {
+            formatter.write_str("1 element in sequence")
+        } else {
+            write!(formatter, "{} elements in sequence", self.count)
+        }
+    }
+}
+
+pub enum ResponsesTextFormatRefusal {
+    ElementsLeftOver {
+        consumed_elements: usize,
+        leftover_elements: usize,
+    },
+    SchemaElementMissing,
+    SchemaMissing,
+    SchemaRepeated,
+}
+
+impl ResponsesTextFormatRefusal {
+    #[must_use]
+    pub fn into_deserialization_error<TError>(self) -> TError
+    where
+        TError: Error,
+    {
+        match self {
+            Self::ElementsLeftOver {
+                consumed_elements,
+                leftover_elements,
+            } => TError::invalid_length(
+                consumed_elements + leftover_elements,
+                &ElementsInSequence {
+                    count: consumed_elements,
+                },
+            ),
+            Self::SchemaElementMissing => TError::invalid_length(0, &JSON_SCHEMA_FORMAT_ELEMENTS),
+            Self::SchemaMissing => TError::missing_field("schema"),
+            Self::SchemaRepeated => TError::duplicate_field("schema"),
+        }
+    }
+}
diff --git a/paddler_openai_translation/src/responses_text_format_schema_occurrences.rs b/paddler_openai_translation/src/responses_text_format_schema_occurrences.rs
new file mode 100644
index 000000000..93a067412
--- /dev/null
+++ b/paddler_openai_translation/src/responses_text_format_schema_occurrences.rs
@@ -0,0 +1,17 @@
+use serde_json::value::RawValue;
+
+pub enum ResponsesTextFormatSchemaOccurrences {
+    Absent,
+    Once(Box<RawValue>),
+    Repeated,
+}
+
+impl ResponsesTextFormatSchemaOccurrences {
+    #[must_use]
+    pub fn with_another(self, schema: Box<RawValue>) -> Self {
+        match self {
+            Self::Absent => Self::Once(schema),
+            Self::Once(_) | Self::Repeated => Self::Repeated,
+        }
+    }
+}
diff --git a/paddler_openai_translation/src/responses_text_format_type.rs b/paddler_openai_translation/src/responses_text_format_type.rs
new file mode 100644
index 000000000..755af5bf7
--- /dev/null
+++ b/paddler_openai_translation/src/responses_text_format_type.rs
@@ -0,0 +1,10 @@
+use serde::Deserialize;
+
+#[derive(Deserialize)]
+#[serde(variant_identifier)]
+pub enum ResponsesTextFormatType {
+    #[serde(rename = "text")]
+    Text,
+    #[serde(rename = "json_schema")]
+    JsonSchema,
+}
diff --git a/paddler_openai_translation/src/responses_text_param.rs b/paddler_openai_translation/src/responses_text_param.rs
new file mode 100644
index 000000000..afa13dfa3
--- /dev/null
+++ b/paddler_openai_translation/src/responses_text_param.rs
@@ -0,0 +1,25 @@
+use serde::Deserialize;
+
+use paddler_messaging::grammar_constraint::GrammarConstraint;
+
+use crate::responses_text_format::ResponsesTextFormat;
+
+#[derive(Deserialize)]
+pub struct ResponsesTextParam {
+    #[serde(default)]
+    pub format: Option<ResponsesTextFormat>,
+}
+
+impl ResponsesTextParam {
+    #[must_use]
+    pub fn into_grammar_constraint(self) -> Option<GrammarConstraint> {
+        match self.format {
+            Some(ResponsesTextFormat::JsonSchema { schema }) => {
+                Some(GrammarConstraint::JsonSchema {
+                    schema: Box::<str>::from(schema).into_string(),
+                })
+            }
+            Some(ResponsesTextFormat::Text) | None => None,
+        }
+    }
+}
diff --git a/paddler_openai_translation/src/responses_tool.rs b/paddler_openai_translation/src/responses_tool.rs
new file mode 100644
index 000000000..4c11e8eb2
--- /dev/null
+++ b/paddler_openai_translation/src/responses_tool.rs
@@ -0,0 +1,10 @@
+use serde::Deserialize;
+
+use crate::function_definition::FunctionDefinition;
+
+#[derive(Deserialize)]
+#[serde(tag = "type")]
+pub enum ResponsesTool {
+    #[serde(rename = "function")]
+    Function(Box<FunctionDefinition>),
+}
diff --git a/paddler_balancer/src/compatibility/openai_service/responses_usage.rs b/paddler_openai_translation/src/responses_usage.rs
similarity index 100%
rename from paddler_balancer/src/compatibility/openai_service/responses_usage.rs
rename to paddler_openai_translation/src/responses_usage.rs
diff --git a/paddler_balancer/src/compatibility/openai_service/text_delta_event.rs b/paddler_openai_translation/src/text_delta_event.rs
similarity index 100%
rename from paddler_balancer/src/compatibility/openai_service/text_delta_event.rs
rename to paddler_openai_translation/src/text_delta_event.rs
diff --git a/paddler_balancer/src/compatibility/openai_service/text_done_event.rs b/paddler_openai_translation/src/text_done_event.rs
similarity index 100%
rename from paddler_balancer/src/compatibility/openai_service/text_done_event.rs
rename to paddler_openai_translation/src/text_done_event.rs
diff --git a/paddler_openai_translation/src/timestamp_from.rs b/paddler_openai_translation/src/timestamp_from.rs
new file mode 100644
index 000000000..75119a92d
--- /dev/null
+++ b/paddler_openai_translation/src/timestamp_from.rs
@@ -0,0 +1,52 @@
+use std::time::SystemTime;
+use std::time::UNIX_EPOCH;
+
+use crate::openai_translation_error::OpenAITranslationError;
+
+pub fn timestamp_from(now: SystemTime) -> Result<u64, OpenAITranslationError> {
+    now.duration_since(UNIX_EPOCH)
+        .map(|since_unix_epoch| since_unix_epoch.as_secs())
+        .map_err(OpenAITranslationError::ClockBeforeUnixEpoch)
+}
+
+#[cfg(test)]
+mod tests {
+    use std::error::Error as _;
+    use std::time::Duration;
+    use std::time::SystemTimeError;
+    use std::time::UNIX_EPOCH;
+
+    use super::timestamp_from;
+    use crate::openai_translation_error::OpenAITranslationError;
+
+    #[test]
+    fn counts_whole_seconds_since_the_unix_epoch() {
+        assert_eq!(
+            timestamp_from(UNIX_EPOCH + Duration::from_millis(42_500)).ok(),
+            Some(42)
+        );
+    }
+
+    #[test]
+    fn refuses_a_time_before_the_unix_epoch() {
+        let translation_error = timestamp_from(UNIX_EPOCH - Duration::from_secs(1))
+            .expect_err("a time before the Unix epoch must be refused");
+
+        assert_eq!(
+            translation_error.to_string(),
+            OpenAITranslationError::ClockBeforeUnixEpoch(
+                UNIX_EPOCH
+                    .duration_since(UNIX_EPOCH + Duration::from_secs(1))
+                    .expect_err("the Unix epoch must precede the second after it")
+            )
+            .to_string()
+        );
+        assert_eq!(
+            translation_error
+                .source()
+                .and_then(|source| source.downcast_ref::<SystemTimeError>())
+                .map(SystemTimeError::duration),
+            Some(Duration::from_secs(1))
+        );
+    }
+}
diff --git a/paddler_openai_translation/src/tool_parameters_schema.rs b/paddler_openai_translation/src/tool_parameters_schema.rs
new file mode 100644
index 000000000..ba8f1d9a0
--- /dev/null
+++ b/paddler_openai_translation/src/tool_parameters_schema.rs
@@ -0,0 +1,79 @@
+use serde::Deserialize;
+use serde_json::Map;
+use serde_json::Value;
+
+use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters_schema::raw_parameters_schema::RawParametersSchema;
+
+#[derive(Deserialize)]
+pub struct ToolParametersSchema {
+    #[serde(default, rename = "type")]
+    pub schema_type: Option<String>,
+    #[serde(default)]
+    pub properties: Option<Map<String, Value>>,
+    #[serde(default)]
+    pub required: Option<Vec<String>>,
+    #[serde(default, rename = "additionalProperties")]
+    pub additional_properties: Option<Value>,
+}
+
+impl ToolParametersSchema {
+    #[must_use]
+    pub fn into_raw_parameters_schema(self) -> RawParametersSchema {
+        let Self {
+            schema_type,
+            properties,
+            required,
+            additional_properties,
+        } = self;
+
+        RawParametersSchema {
+            schema_type: schema_type.unwrap_or_else(|| "object".to_owned()),
+            properties,
+            required,
+            additional_properties,
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use serde_json::from_value;
+    use serde_json::json;
+
+    use super::ToolParametersSchema;
+
+    #[test]
+    fn conversion_ignores_unknown_keywords_and_keeps_recognized_fields() {
+        let schema: ToolParametersSchema = from_value(json!({
+            "type": "object",
+            "properties": {"location": {"type": "string"}},
+            "required": ["location"],
+            "additionalProperties": false,
+            "$schema": "http://json-schema.org/draft-07/schema#",
+            "title": "Weather",
+            "$defs": {"Unit": {"type": "string"}}
+        }))
+        .unwrap();
+
+        let raw = schema.into_raw_parameters_schema();
+
+        assert_eq!(raw.schema_type, "object");
+        assert_eq!(raw.properties.unwrap().len(), 1);
+        assert_eq!(raw.required, Some(vec!["location".to_owned()]));
+        assert_eq!(raw.additional_properties, Some(json!(false)));
+    }
+
+    #[test]
+    fn conversion_defaults_missing_type_to_object() {
+        let schema: ToolParametersSchema = from_value(json!({
+            "properties": {"location": {"type": "string"}}
+        }))
+        .unwrap();
+
+        let raw = schema.into_raw_parameters_schema();
+
+        assert_eq!(raw.schema_type, "object");
+        assert!(raw.required.is_none());
+        assert!(raw.additional_properties.is_none());
+    }
+}
diff --git a/paddler_openai_translation/src/translated_chat_completion_request.rs b/paddler_openai_translation/src/translated_chat_completion_request.rs
new file mode 100644
index 000000000..efe0647b3
--- /dev/null
+++ b/paddler_openai_translation/src/translated_chat_completion_request.rs
@@ -0,0 +1,10 @@
+use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
+use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters_schema::validated_parameters_schema::ValidatedParametersSchema;
+
+use crate::chat_completion_delivery::ChatCompletionDelivery;
+
+pub struct TranslatedChatCompletionRequest {
+    pub conversation_history_params:
+        ContinueFromConversationHistoryParams<ValidatedParametersSchema>,
+    pub delivery: ChatCompletionDelivery,
+}
diff --git a/paddler_openai_translation/src/translated_responses_request.rs b/paddler_openai_translation/src/translated_responses_request.rs
new file mode 100644
index 000000000..cd2ef01a0
--- /dev/null
+++ b/paddler_openai_translation/src/translated_responses_request.rs
@@ -0,0 +1,10 @@
+use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
+use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters_schema::validated_parameters_schema::ValidatedParametersSchema;
+
+use crate::responses_delivery::ResponsesDelivery;
+
+pub struct TranslatedResponsesRequest {
+    pub conversation_history_params:
+        ContinueFromConversationHistoryParams<ValidatedParametersSchema>,
+    pub delivery: ResponsesDelivery,
+}
diff --git a/paddler_openai_translation/tests/a_buffered_chat_completion_carries_its_header_fields.rs b/paddler_openai_translation/tests/a_buffered_chat_completion_carries_its_header_fields.rs
new file mode 100644
index 000000000..b370709ba
--- /dev/null
+++ b/paddler_openai_translation/tests/a_buffered_chat_completion_carries_its_header_fields.rs
@@ -0,0 +1,16 @@
+use serde_json::json;
+
+use crate::finishing_json_of::finishing_json_of;
+
+#[test]
+fn a_buffered_chat_completion_carries_its_header_fields() {
+    let buffered = finishing_json_of(json!({
+        "model": "test-model",
+        "messages": [{"role": "user", "content": "hi"}]
+    }));
+
+    assert_eq!(buffered.len(), 1);
+    assert_eq!(buffered[0]["object"], "chat.completion");
+    assert_eq!(buffered[0]["created"], 1_234);
+    assert_eq!(buffered[0]["model"], "test-model");
+}
diff --git a/paddler_openai_translation/tests/a_chat_completion_request_with_an_unsatisfiable_tool_is_refused.rs b/paddler_openai_translation/tests/a_chat_completion_request_with_an_unsatisfiable_tool_is_refused.rs
new file mode 100644
index 000000000..7a3209dd1
--- /dev/null
+++ b/paddler_openai_translation/tests/a_chat_completion_request_with_an_unsatisfiable_tool_is_refused.rs
@@ -0,0 +1,45 @@
+use std::time::UNIX_EPOCH;
+
+use serde_json::from_value;
+use serde_json::json;
+
+use paddler_messaging::request_params_validation_error::RequestParamsValidationError;
+use paddler_openai_translation::chat_completion_request::ChatCompletionRequest;
+use paddler_openai_translation::openai_translation_error::OpenAITranslationError;
+
+#[test]
+fn a_chat_completion_request_with_an_unsatisfiable_tool_is_refused() {
+    let chat_completion_request: ChatCompletionRequest = from_value(json!({
+        "model": "test-model",
+        "messages": [{"role": "user", "content": "hi"}],
+        "tools": [{
+            "type": "function",
+            "function": {
+                "name": "broken",
+                "description": "tool with an unsatisfiable required field",
+                "parameters": {
+                    "type": "object",
+                    "properties": {"present": {"type": "string"}},
+                    "required": ["absent"]
+                }
+            }
+        }]
+    }))
+    .expect("the request must deserialize");
+
+    assert_eq!(
+        chat_completion_request
+            .translate(UNIX_EPOCH)
+            .err()
+            .as_ref()
+            .map(ToString::to_string),
+        Some(
+            OpenAITranslationError::ToolRejected(
+                RequestParamsValidationError::RequiredFieldNotInProperties {
+                    field: "absent".to_owned(),
+                }
+            )
+            .to_string()
+        )
+    );
+}
diff --git a/paddler_openai_translation/tests/a_request_translated_before_the_unix_epoch_is_refused.rs b/paddler_openai_translation/tests/a_request_translated_before_the_unix_epoch_is_refused.rs
new file mode 100644
index 000000000..c2ce8f034
--- /dev/null
+++ b/paddler_openai_translation/tests/a_request_translated_before_the_unix_epoch_is_refused.rs
@@ -0,0 +1,61 @@
+use std::error::Error as _;
+use std::time::Duration;
+use std::time::SystemTimeError;
+use std::time::UNIX_EPOCH;
+
+use serde_json::from_value;
+use serde_json::json;
+
+use paddler_inference_parameters::sampling_parameters::SamplingParameters;
+use paddler_openai_translation::chat_completion_request::ChatCompletionRequest;
+use paddler_openai_translation::openai_translation_error::OpenAITranslationError;
+use paddler_openai_translation::responses_request::ResponsesRequest;
+
+fn clock_error_duration(translation_error: &OpenAITranslationError) -> Option<Duration> {
+    translation_error
+        .source()
+        .and_then(|source| source.downcast_ref::<SystemTimeError>())
+        .map(SystemTimeError::duration)
+}
+
+#[test]
+fn a_request_translated_before_the_unix_epoch_is_refused() {
+    let before_unix_epoch = UNIX_EPOCH - Duration::from_secs(1);
+    let chat_completion_request: ChatCompletionRequest = from_value(json!({
+        "model": "test-model",
+        "messages": [{"role": "user", "content": "hi"}]
+    }))
+    .expect("the chat completion request must deserialize");
+    let responses_request: ResponsesRequest =
+        from_value(json!({"model": "test-model", "input": "hi"}))
+            .expect("the responses request must deserialize");
+    let clock_before_unix_epoch = OpenAITranslationError::ClockBeforeUnixEpoch(
+        UNIX_EPOCH
+            .duration_since(UNIX_EPOCH + Duration::from_secs(1))
+            .expect_err("the Unix epoch must precede the second after it"),
+    );
+
+    let chat_completion_error = chat_completion_request.translate(before_unix_epoch).err();
+    let responses_error = responses_request
+        .translate(before_unix_epoch, &SamplingParameters::default())
+        .err();
+
+    assert_eq!(
+        chat_completion_error.as_ref().map(ToString::to_string),
+        Some(clock_before_unix_epoch.to_string())
+    );
+    assert_eq!(
+        chat_completion_error
+            .as_ref()
+            .and_then(clock_error_duration),
+        Some(Duration::from_secs(1))
+    );
+    assert_eq!(
+        responses_error.as_ref().map(ToString::to_string),
+        Some(clock_before_unix_epoch.to_string())
+    );
+    assert_eq!(
+        responses_error.as_ref().and_then(clock_error_duration),
+        Some(Duration::from_secs(1))
+    );
+}
diff --git a/paddler_openai_translation/tests/a_streamed_chat_completion_including_usage_adds_a_usage_chunk.rs b/paddler_openai_translation/tests/a_streamed_chat_completion_including_usage_adds_a_usage_chunk.rs
new file mode 100644
index 000000000..9ca45d88b
--- /dev/null
+++ b/paddler_openai_translation/tests/a_streamed_chat_completion_including_usage_adds_a_usage_chunk.rs
@@ -0,0 +1,24 @@
+use serde_json::json;
+
+use crate::finishing_json_of::finishing_json_of;
+
+#[test]
+fn a_streamed_chat_completion_including_usage_adds_a_usage_chunk() {
+    let streamed_with_usage = finishing_json_of(json!({
+        "model": "test-model",
+        "messages": [{"role": "user", "content": "hi"}],
+        "stream": true,
+        "stream_options": {"include_usage": true}
+    }));
+
+    assert_eq!(
+        streamed_with_usage
+            .iter()
+            .map(|chunk| (chunk["object"].clone(), chunk.get("usage").is_some()))
+            .collect::<Vec<_>>(),
+        vec![
+            (json!("chat.completion.chunk"), false),
+            (json!("chat.completion.chunk"), true)
+        ]
+    );
+}
diff --git a/paddler_openai_translation/tests/a_streamed_chat_completion_without_usage_sends_a_single_chunk.rs b/paddler_openai_translation/tests/a_streamed_chat_completion_without_usage_sends_a_single_chunk.rs
new file mode 100644
index 000000000..19d36ba91
--- /dev/null
+++ b/paddler_openai_translation/tests/a_streamed_chat_completion_without_usage_sends_a_single_chunk.rs
@@ -0,0 +1,14 @@
+use serde_json::json;
+
+use crate::finishing_json_of::finishing_json_of;
+
+#[test]
+fn a_streamed_chat_completion_without_usage_sends_a_single_chunk() {
+    let streamed_without_usage = finishing_json_of(json!({
+        "model": "test-model",
+        "messages": [{"role": "user", "content": "hi"}],
+        "stream": true
+    }));
+
+    assert_eq!(streamed_without_usage.len(), 1);
+}
diff --git a/paddler_openai_translation/tests/finishing_json_of.rs b/paddler_openai_translation/tests/finishing_json_of.rs
new file mode 100644
index 000000000..df8c25a1b
--- /dev/null
+++ b/paddler_openai_translation/tests/finishing_json_of.rs
@@ -0,0 +1,49 @@
+use std::time::Duration;
+use std::time::UNIX_EPOCH;
+
+use llama_cpp_bindings_types::TokenUsage;
+use serde_json::Value;
+use serde_json::from_value;
+use serde_json::to_value;
+
+use paddler_messaging::generation_finish::GenerationFinish;
+use paddler_messaging::generation_summary::GenerationSummary;
+use paddler_openai_translation::chat_completion_delivery::ChatCompletionDelivery;
+use paddler_openai_translation::chat_completion_request::ChatCompletionRequest;
+use paddler_openai_translation::completes_generation::CompletesGeneration as _;
+use paddler_openai_translation::generated_output::GeneratedOutput;
+use paddler_openai_translation::generation_outcome::GenerationOutcome;
+
+pub fn finishing_json_of(request: Value) -> Vec<Value> {
+    let generation_summary = GenerationSummary {
+        finish: GenerationFinish::EndOfGeneration,
+        usage: TokenUsage::new(),
+    };
+
+    match from_value::<ChatCompletionRequest>(request)
+        .expect("the request must deserialize")
+        .translate(UNIX_EPOCH + Duration::from_secs(1_234))
+        .expect("the request must translate")
+        .delivery
+    {
+        ChatCompletionDelivery::Buffered(chat_completion_header) => vec![
+            to_value(chat_completion_header.complete(
+                "test-request",
+                GeneratedOutput::default(),
+                &generation_summary,
+            ))
+            .expect("a completion must serialize"),
+        ],
+        ChatCompletionDelivery::Streamed(chat_completion_stream) => chat_completion_stream
+            .conclude(
+                "test-request",
+                GenerationOutcome::Finished(generation_summary),
+            )
+            .expect("a finished generation must produce chunks")
+            .iter()
+            .map(|chat_completion_chunk| {
+                to_value(chat_completion_chunk).expect("a chunk must serialize")
+            })
+            .collect(),
+    }
+}
diff --git a/paddler_openai_translation/tests/first_response_snapshots.rs b/paddler_openai_translation/tests/first_response_snapshots.rs
new file mode 100644
index 000000000..0feb774f2
--- /dev/null
+++ b/paddler_openai_translation/tests/first_response_snapshots.rs
@@ -0,0 +1,67 @@
+use std::time::UNIX_EPOCH;
+
+use llama_cpp_bindings_types::TokenUsage;
+use serde_json::Value;
+use serde_json::from_value;
+use serde_json::json;
+use serde_json::to_value;
+
+use paddler_inference_parameters::sampling_parameters::SamplingParameters;
+use paddler_messaging::generation_finish::GenerationFinish;
+use paddler_messaging::generation_summary::GenerationSummary;
+use paddler_openai_translation::completes_generation::CompletesGeneration as _;
+use paddler_openai_translation::generated_output::GeneratedOutput;
+use paddler_openai_translation::generated_output_part::GeneratedOutputPart;
+use paddler_openai_translation::generation_progress::GenerationProgress;
+use paddler_openai_translation::responses_delivery::ResponsesDelivery;
+use paddler_openai_translation::responses_request::ResponsesRequest;
+
+fn first_response_snapshot_of(request: Value) -> Value {
+    let sampling_parameters = SamplingParameters {
+        temperature: 0.25,
+        top_p: 0.5,
+        ..SamplingParameters::default()
+    };
+
+    match from_value::<ResponsesRequest>(request)
+        .expect("the request must deserialize")
+        .translate(UNIX_EPOCH, &sampling_parameters)
+        .expect("the request must translate")
+        .delivery
+    {
+        ResponsesDelivery::Buffered(responses_response_header) => {
+            to_value(responses_response_header.complete(
+                "test-request",
+                GeneratedOutput::default(),
+                &GenerationSummary {
+                    finish: GenerationFinish::EndOfGeneration,
+                    usage: TokenUsage::new(),
+                },
+            ))
+            .expect("a response must serialize")
+        }
+        ResponsesDelivery::Streamed(mut responses_stream) => to_value(
+            &responses_stream.advance(GenerationProgress::Produced(GeneratedOutputPart::Content(
+                "hello".to_owned(),
+            )))[0],
+        )
+        .expect("an event must serialize")["response"]
+            .clone(),
+    }
+}
+
+pub struct FirstResponseSnapshots {
+    pub buffered: Value,
+    pub streamed: Value,
+}
+
+impl FirstResponseSnapshots {
+    pub fn of_buffered_and_streamed_requests() -> Self {
+        Self {
+            buffered: first_response_snapshot_of(json!({"model": "test-model", "input": "hi"})),
+            streamed: first_response_snapshot_of(
+                json!({"model": "test-model", "input": "hi", "stream": true}),
+            ),
+        }
+    }
+}
diff --git a/paddler_openai_translation/tests/main.rs b/paddler_openai_translation/tests/main.rs
new file mode 100644
index 000000000..7d8174e38
--- /dev/null
+++ b/paddler_openai_translation/tests/main.rs
@@ -0,0 +1,11 @@
+mod a_buffered_chat_completion_carries_its_header_fields;
+mod a_chat_completion_request_with_an_unsatisfiable_tool_is_refused;
+mod a_request_translated_before_the_unix_epoch_is_refused;
+mod a_streamed_chat_completion_including_usage_adds_a_usage_chunk;
+mod a_streamed_chat_completion_without_usage_sends_a_single_chunk;
+#[cfg(test)]
+mod finishing_json_of;
+#[cfg(test)]
+mod first_response_snapshots;
+mod responses_requests_report_the_sampling_parameters_in_effect;
+mod responses_requests_report_their_status_and_a_resp_prefixed_id;
diff --git a/paddler_openai_translation/tests/responses_requests_report_the_sampling_parameters_in_effect.rs b/paddler_openai_translation/tests/responses_requests_report_the_sampling_parameters_in_effect.rs
new file mode 100644
index 000000000..146bc71a4
--- /dev/null
+++ b/paddler_openai_translation/tests/responses_requests_report_the_sampling_parameters_in_effect.rs
@@ -0,0 +1,12 @@
+use crate::first_response_snapshots::FirstResponseSnapshots;
+
+#[test]
+fn responses_requests_report_the_sampling_parameters_in_effect() {
+    let FirstResponseSnapshots { buffered, streamed } =
+        FirstResponseSnapshots::of_buffered_and_streamed_requests();
+
+    for response_snapshot in [buffered, streamed] {
+        assert_eq!(response_snapshot["temperature"], 0.25);
+        assert_eq!(response_snapshot["top_p"], 0.5);
+    }
+}
diff --git a/paddler_openai_translation/tests/responses_requests_report_their_status_and_a_resp_prefixed_id.rs b/paddler_openai_translation/tests/responses_requests_report_their_status_and_a_resp_prefixed_id.rs
new file mode 100644
index 000000000..383dceaf8
--- /dev/null
+++ b/paddler_openai_translation/tests/responses_requests_report_their_status_and_a_resp_prefixed_id.rs
@@ -0,0 +1,18 @@
+use crate::first_response_snapshots::FirstResponseSnapshots;
+
+#[test]
+fn responses_requests_report_their_status_and_a_resp_prefixed_id() {
+    let FirstResponseSnapshots { buffered, streamed } =
+        FirstResponseSnapshots::of_buffered_and_streamed_requests();
+
+    assert_eq!(buffered["status"], "completed");
+    assert_eq!(streamed["status"], "in_progress");
+
+    for response_snapshot in [buffered, streamed] {
+        assert!(
+            response_snapshot["id"]
+                .as_str()
+                .is_some_and(|response_id| response_id.starts_with("resp_"))
+        );
+    }
+}
diff --git a/paddler_opencode_tests/Cargo.toml b/paddler_opencode_tests/Cargo.toml
index 5d763761e..60c80bc57 100644
--- a/paddler_opencode_tests/Cargo.toml
+++ b/paddler_opencode_tests/Cargo.toml
@@ -34,9 +34,5 @@ thiserror = { workspace = true }
 tokio = { workspace = true }
 url = { workspace = true }
 
-[dev-dependencies]
-anyhow = { workspace = true }
-paddler_test_cluster_harness = { workspace = true }
-
 [lints]
 workspace = true
diff --git a/paddler_opencode_tests/src/bin/opencode_stub.rs b/paddler_opencode_tests/src/bin/opencode_stub.rs
index affc072ac..907c7e34b 100644
--- a/paddler_opencode_tests/src/bin/opencode_stub.rs
+++ b/paddler_opencode_tests/src/bin/opencode_stub.rs
@@ -1,4 +1,6 @@
 use std::env::args;
+use std::env::current_dir;
+use std::env::var_os;
 use std::io::Result;
 use std::io::Write as _;
 use std::io::stdout;
@@ -7,7 +9,17 @@ fn main() -> Result<()> {
     let forwarded_arguments: Vec<String> = args().skip(1).collect();
     let mut output = stdout().lock();
 
-    writeln!(output, "{}", forwarded_arguments.join(" "))?;
-
-    output.flush()
+    current_dir()
+        .and_then(|working_directory| {
+            writeln!(
+                output,
+                "{}\n{:?}\n{:?}\n{:?}\n{}",
+                working_directory.display(),
+                var_os("PWD"),
+                var_os("HOME"),
+                var_os("XDG_CONFIG_HOME"),
+                forwarded_arguments.join(" ")
+            )
+        })
+        .and_then(|()| output.flush())
 }
diff --git a/paddler_opencode_tests/src/opencode_binary_path.rs b/paddler_opencode_tests/src/opencode_binary_path.rs
index a110f2430..72dfa7f6f 100644
--- a/paddler_opencode_tests/src/opencode_binary_path.rs
+++ b/paddler_opencode_tests/src/opencode_binary_path.rs
@@ -31,6 +31,7 @@ pub fn opencode_binary_path() -> Result<PathBuf, OpenCodeTestError> {
 #[cfg(test)]
 mod tests {
     use std::env::VarError;
+    use std::path::PathBuf;
 
     use tempfile::NamedTempFile;
 
@@ -50,13 +51,13 @@ mod tests {
 
     #[test]
     fn nonexistent_path_reports_missing_binary() {
-        let error =
-            resolve_binary_path(Ok("/paddler/definitely/missing/opencode".to_owned())).unwrap_err();
+        let missing_path = PathBuf::from("/paddler/definitely/missing/opencode");
+        let error = resolve_binary_path(Ok(missing_path.display().to_string())).unwrap_err();
 
-        assert!(matches!(
-            error,
-            OpenCodeTestError::BinaryDoesNotExist { .. }
-        ));
+        assert_eq!(
+            error.to_string(),
+            OpenCodeTestError::BinaryDoesNotExist { path: missing_path }.to_string()
+        );
     }
 
     #[test]
diff --git a/paddler_opencode_tests/src/opencode_test_error.rs b/paddler_opencode_tests/src/opencode_test_error.rs
index 5e3650353..967c9d42d 100644
--- a/paddler_opencode_tests/src/opencode_test_error.rs
+++ b/paddler_opencode_tests/src/opencode_test_error.rs
@@ -2,7 +2,6 @@ use std::env::VarError;
 use std::io::Error as IoError;
 use std::path::PathBuf;
 
-use serde_json::Error as SerdeJsonError;
 use thiserror::Error;
 
 #[derive(Debug, Error)]
@@ -18,16 +17,7 @@ pub enum OpenCodeTestError {
     BinaryDoesNotExist { path: PathBuf },
 
     #[error("failed to set up the OpenCode test project")]
-    ProjectSetupFailed {
-        #[source]
-        source: IoError,
-    },
-
-    #[error("failed to serialize the OpenCode config")]
-    ConfigSerializationFailed {
-        #[source]
-        source: SerdeJsonError,
-    },
+    ProjectSetupFailed(#[source] IoError),
 
     #[error("failed to spawn the OpenCode binary at {path}")]
     SpawnFailed {
@@ -37,8 +27,5 @@ pub enum OpenCodeTestError {
     },
 
     #[error("failed to wait for the OpenCode process to finish")]
-    ProcessWaitFailed {
-        #[source]
-        source: IoError,
-    },
+    ProcessWaitFailed(#[source] IoError),
 }
diff --git a/paddler_opencode_tests/src/opencode_test_project.rs b/paddler_opencode_tests/src/opencode_test_project.rs
index 597890064..c7d2a8f7c 100644
--- a/paddler_opencode_tests/src/opencode_test_project.rs
+++ b/paddler_opencode_tests/src/opencode_test_project.rs
@@ -4,7 +4,6 @@ use std::path::Path;
 use serde_json::Map;
 use serde_json::Value;
 use serde_json::json;
-use serde_json::to_string_pretty;
 use tempfile::TempDir;
 use tempfile::tempdir;
 use url::Url;
@@ -20,19 +19,19 @@ fn provider_config(api_base_url: &Url) -> Value {
 
     models.insert(
         MODEL_ID.to_owned(),
-        json!({ "id": MODEL_ID, "name": "Paddler OpenCode test" }),
+        json!({ "modelID": MODEL_ID, "name": "Paddler OpenCode test" }),
     );
 
     json!({
         "$schema": "https://opencode.ai/config.json",
         "model": format!("paddler/{MODEL_ID}"),
-        "provider": {
+        "providers": {
             "paddler": {
-                "npm": "@ai-sdk/openai-compatible",
                 "name": "Paddler",
-                "options": {
-                    "baseURL": api_base_url.as_str(),
-                    "apiKey": "paddler"
+                "package": "@opencode/ai/providers/openai-compatible",
+                "settings": {
+                    "apiKey": "paddler",
+                    "baseURL": api_base_url.as_str()
                 },
                 "models": models
             }
@@ -42,25 +41,29 @@ fn provider_config(api_base_url: &Url) -> Value {
 
 pub struct OpenCodeTestProject {
     directory: TempDir,
+    home: TempDir,
     marker_contents: String,
 }
 
 impl OpenCodeTestProject {
     pub fn create(api_base_url: &Url, marker_contents: String) -> Result<Self, OpenCodeTestError> {
-        let directory =
-            tempdir().map_err(|source| OpenCodeTestError::ProjectSetupFailed { source })?;
-        let config_contents = to_string_pretty(&provider_config(api_base_url))
-            .map_err(|source| OpenCodeTestError::ConfigSerializationFailed { source })?;
-
-        fs::write(directory.path().join(CONFIG_FILE_NAME), config_contents)
-            .map_err(|source| OpenCodeTestError::ProjectSetupFailed { source })?;
-        fs::write(directory.path().join(MARKER_FILE_NAME), &marker_contents)
-            .map_err(|source| OpenCodeTestError::ProjectSetupFailed { source })?;
-
-        Ok(Self {
-            directory,
-            marker_contents,
-        })
+        tempdir()
+            .and_then(|directory| {
+                fs::write(
+                    directory.path().join(CONFIG_FILE_NAME),
+                    format!("{:#}", provider_config(api_base_url)),
+                )
+                .and_then(|()| fs::write(directory.path().join(MARKER_FILE_NAME), &marker_contents))
+                .map(|()| directory)
+            })
+            .and_then(|directory| {
+                tempdir().map(|home| Self {
+                    directory,
+                    home,
+                    marker_contents,
+                })
+            })
+            .map_err(OpenCodeTestError::ProjectSetupFailed)
     }
 
     #[must_use]
@@ -68,6 +71,11 @@ impl OpenCodeTestProject {
         self.directory.path()
     }
 
+    #[must_use]
+    pub fn home_path(&self) -> &Path {
+        self.home.path()
+    }
+
     #[must_use]
     pub const fn marker_file_name(&self) -> &str {
         MARKER_FILE_NAME
@@ -90,6 +98,7 @@ mod tests {
 
     use serde_json::Value;
     use serde_json::from_str;
+    use serde_json::json;
     use url::Url;
 
     use super::CONFIG_FILE_NAME;
@@ -110,11 +119,25 @@ mod tests {
                 .unwrap();
 
         assert_eq!(
-            config["provider"]["paddler"]["options"]["baseURL"],
-            "http://127.0.0.1:9/v1"
+            config,
+            json!({
+                "$schema": "https://opencode.ai/config.json",
+                "model": "paddler/paddler-opencode-test",
+                "providers": {
+                    "paddler": {
+                        "name": "Paddler",
+                        "package": "@opencode/ai/providers/openai-compatible",
+                        "settings": {
+                            "apiKey": "paddler",
+                            "baseURL": "http://127.0.0.1:9/v1"
+                        },
+                        "models": {
+                            MODEL_ID: { "modelID": MODEL_ID, "name": "Paddler OpenCode test" }
+                        }
+                    }
+                }
+            })
         );
-        assert_eq!(config["model"], "paddler/paddler-opencode-test");
-        assert!(config["provider"]["paddler"]["models"][MODEL_ID].is_object());
     }
 
     #[test]
diff --git a/paddler_opencode_tests/src/run_opencode.rs b/paddler_opencode_tests/src/run_opencode.rs
index 7d14bbacb..ccf200b85 100644
--- a/paddler_opencode_tests/src/run_opencode.rs
+++ b/paddler_opencode_tests/src/run_opencode.rs
@@ -13,17 +13,24 @@ pub async fn run_opencode(
     prompt: &str,
 ) -> Result<OpenCodeRunOutcome, OpenCodeTestError> {
     let mut command = Command::new(binary_path);
+    let home = project.home_path();
 
     command
         .arg("run")
         .arg(prompt)
-        .arg("--dir")
-        .arg(project.directory_path())
-        .arg("--pure")
+        .arg("--standalone")
         .arg("--auto")
         .arg("--model")
         .arg(project.model_reference())
+        .current_dir(project.directory_path())
+        .env("PWD", project.directory_path())
+        .env("HOME", home)
+        .env("XDG_CACHE_HOME", home.join(".cache"))
+        .env("XDG_CONFIG_HOME", home.join(".config"))
+        .env("XDG_DATA_HOME", home.join(".local").join("share"))
+        .env("XDG_STATE_HOME", home.join(".local").join("state"))
         .kill_on_drop(true)
+        .stdin(Stdio::null())
         .stdout(Stdio::piped())
         .stderr(Stdio::piped());
 
@@ -34,14 +41,13 @@ pub async fn run_opencode(
             source,
         })?;
 
-    let output = child
+    child
         .wait_with_output()
         .await
-        .map_err(|source| OpenCodeTestError::ProcessWaitFailed { source })?;
-
-    Ok(OpenCodeRunOutcome {
-        exit_success: output.status.success(),
-        stderr: String::from_utf8_lossy(&output.stderr).into_owned(),
-        stdout: String::from_utf8_lossy(&output.stdout).into_owned(),
-    })
+        .map(|output| OpenCodeRunOutcome {
+            exit_success: output.status.success(),
+            stderr: String::from_utf8_lossy(&output.stderr).into_owned(),
+            stdout: String::from_utf8_lossy(&output.stdout).into_owned(),
+        })
+        .map_err(OpenCodeTestError::ProcessWaitFailed)
 }
diff --git a/paddler_opencode_tests/tests/main.rs b/paddler_opencode_tests/tests/main.rs
index 6b6f57a10..91ce10118 100644
--- a/paddler_opencode_tests/tests/main.rs
+++ b/paddler_opencode_tests/tests/main.rs
@@ -1,4 +1,5 @@
 mod opencode_completes_tool_enabled_chat_request;
 mod run_opencode_captures_output_of_a_finished_process;
 mod run_opencode_reports_spawn_failure_for_a_missing_binary;
+#[cfg(test)]
 mod unreachable_api_opencode_test_project;
diff --git a/paddler_opencode_tests/tests/opencode_completes_tool_enabled_chat_request.rs b/paddler_opencode_tests/tests/opencode_completes_tool_enabled_chat_request.rs
index 2e3537f0b..92ee9f4e3 100644
--- a/paddler_opencode_tests/tests/opencode_completes_tool_enabled_chat_request.rs
+++ b/paddler_opencode_tests/tests/opencode_completes_tool_enabled_chat_request.rs
@@ -3,22 +3,16 @@
 use paddler_opencode_tests::opencode_binary_path::opencode_binary_path;
 use paddler_opencode_tests::opencode_test_project::OpenCodeTestProject;
 use paddler_opencode_tests::run_opencode::run_opencode;
-use paddler_test_cluster_harness::agent_config::AgentConfig;
-use paddler_tests::start_cluster_with_qwen3_5_and_context_size::start_cluster_with_qwen3_5_and_context_size;
-
-const OPENCODE_CONTEXT_SIZE: u32 = 16384;
+use paddler_tests::ministral_3_cluster_params::Ministral3ClusterParams;
+use paddler_tests::start_cluster_with_ministral_3::start_cluster_with_ministral_3;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn opencode_completes_tool_enabled_chat_request() {
     let binary_path = opencode_binary_path().expect("the opencode binary must be configured");
 
-    let cluster = start_cluster_with_qwen3_5_and_context_size(
-        vec![AgentConfig::single(1)],
-        false,
-        OPENCODE_CONTEXT_SIZE,
-    )
-    .await
-    .expect("the cluster must start");
+    let cluster = start_cluster_with_ministral_3(Ministral3ClusterParams::default())
+        .await
+        .expect("the cluster must start");
 
     let api_base_url = cluster
         .balancer
diff --git a/paddler_opencode_tests/tests/run_opencode_captures_output_of_a_finished_process.rs b/paddler_opencode_tests/tests/run_opencode_captures_output_of_a_finished_process.rs
index f5d48dee7..4c10b31c4 100644
--- a/paddler_opencode_tests/tests/run_opencode_captures_output_of_a_finished_process.rs
+++ b/paddler_opencode_tests/tests/run_opencode_captures_output_of_a_finished_process.rs
@@ -6,8 +6,7 @@ use crate::unreachable_api_opencode_test_project::unreachable_api_opencode_test_
 
 #[tokio::test]
 async fn run_opencode_captures_output_of_a_finished_process() {
-    let project = unreachable_api_opencode_test_project()
-        .expect("an OpenCode test project must be creatable");
+    let project = unreachable_api_opencode_test_project();
     let outcome = run_opencode(
         Path::new(env!("CARGO_BIN_EXE_opencode_stub")),
         &project,
@@ -20,8 +19,15 @@ async fn run_opencode_captures_output_of_a_finished_process() {
     assert_eq!(
         outcome.stdout,
         format!(
-            "run paddler-probe-argument --dir {} --pure --auto --model {}\n",
-            project.directory_path().display(),
+            "{}\n{:?}\n{:?}\n{:?}\nrun paddler-probe-argument --standalone --auto --model {}\n",
+            project
+                .directory_path()
+                .canonicalize()
+                .expect("the project directory must resolve")
+                .display(),
+            Some(project.directory_path().as_os_str()),
+            Some(project.home_path().as_os_str()),
+            Some(project.home_path().join(".config").as_os_str()),
             project.model_reference()
         )
     );
diff --git a/paddler_opencode_tests/tests/run_opencode_reports_spawn_failure_for_a_missing_binary.rs b/paddler_opencode_tests/tests/run_opencode_reports_spawn_failure_for_a_missing_binary.rs
index 4d92d47f2..127067c63 100644
--- a/paddler_opencode_tests/tests/run_opencode_reports_spawn_failure_for_a_missing_binary.rs
+++ b/paddler_opencode_tests/tests/run_opencode_reports_spawn_failure_for_a_missing_binary.rs
@@ -1,3 +1,5 @@
+use std::error::Error as _;
+use std::io;
 use std::path::Path;
 
 use paddler_opencode_tests::opencode_test_error::OpenCodeTestError;
@@ -7,14 +9,28 @@ use crate::unreachable_api_opencode_test_project::unreachable_api_opencode_test_
 
 #[tokio::test]
 async fn run_opencode_reports_spawn_failure_for_a_missing_binary() {
+    let missing_binary = Path::new("/paddler/definitely/missing/opencode");
     let run_error = run_opencode(
-        Path::new("/paddler/definitely/missing/opencode"),
-        &unreachable_api_opencode_test_project()
-            .expect("an OpenCode test project must be creatable"),
+        missing_binary,
+        &unreachable_api_opencode_test_project(),
         "hello",
     )
     .await
     .expect_err("a missing binary must fail to spawn");
 
-    assert!(matches!(run_error, OpenCodeTestError::SpawnFailed { .. }));
+    assert_eq!(
+        run_error.to_string(),
+        OpenCodeTestError::SpawnFailed {
+            path: missing_binary.to_path_buf(),
+            source: io::Error::from(io::ErrorKind::NotFound),
+        }
+        .to_string()
+    );
+    assert_eq!(
+        run_error
+            .source()
+            .and_then(|source| source.downcast_ref::<io::Error>())
+            .map(io::Error::kind),
+        Some(io::ErrorKind::NotFound)
+    );
 }
diff --git a/paddler_opencode_tests/tests/unreachable_api_opencode_test_project.rs b/paddler_opencode_tests/tests/unreachable_api_opencode_test_project.rs
index acf08fd21..b5c846388 100644
--- a/paddler_opencode_tests/tests/unreachable_api_opencode_test_project.rs
+++ b/paddler_opencode_tests/tests/unreachable_api_opencode_test_project.rs
@@ -1,13 +1,11 @@
-use anyhow::Result;
 use url::Url;
 
 use paddler_opencode_tests::opencode_test_project::OpenCodeTestProject;
 
-pub fn unreachable_api_opencode_test_project() -> Result<OpenCodeTestProject> {
-    let api_base_url = Url::parse("http://127.0.0.1:1/v1")?;
-
-    Ok(OpenCodeTestProject::create(
-        &api_base_url,
+pub fn unreachable_api_opencode_test_project() -> OpenCodeTestProject {
+    OpenCodeTestProject::create(
+        &Url::parse("http://127.0.0.1:1/v1").expect("the unreachable API URL must be valid"),
         "marker".to_owned(),
-    )?)
+    )
+    .expect("an OpenCode test project must be creatable")
 }
diff --git a/paddler_outbox/Cargo.toml b/paddler_outbox/Cargo.toml
new file mode 100644
index 000000000..870aa1d5f
--- /dev/null
+++ b/paddler_outbox/Cargo.toml
@@ -0,0 +1,16 @@
+[package]
+name = "paddler_outbox"
+authors.workspace = true
+description.workspace = true
+edition.workspace = true
+homepage.workspace = true
+license.workspace = true
+repository.workspace = true
+version.workspace = true
+
+[dependencies]
+parking_lot = { workspace = true }
+tokio = { workspace = true }
+
+[lints]
+workspace = true
diff --git a/paddler_outbox/src/lib.rs b/paddler_outbox/src/lib.rs
new file mode 100644
index 000000000..0450ddccf
--- /dev/null
+++ b/paddler_outbox/src/lib.rs
@@ -0,0 +1 @@
+pub mod outbox;
diff --git a/paddler_outbox/src/outbox.rs b/paddler_outbox/src/outbox.rs
new file mode 100644
index 000000000..bff758174
--- /dev/null
+++ b/paddler_outbox/src/outbox.rs
@@ -0,0 +1,72 @@
+use std::collections::VecDeque;
+
+use parking_lot::Mutex;
+use tokio::sync::Notify;
+
+pub struct Outbox<TMessage> {
+    message_arrived: Notify,
+    messages: Mutex<VecDeque<TMessage>>,
+}
+
+impl<TMessage> Outbox<TMessage> {
+    pub fn push(&self, message: TMessage) {
+        self.messages.lock().push_back(message);
+        self.message_arrived.notify_one();
+    }
+
+    pub async fn next(&self) -> TMessage {
+        loop {
+            let next_message = self.messages.lock().pop_front();
+
+            if let Some(message) = next_message {
+                return message;
+            }
+
+            self.message_arrived.notified().await;
+        }
+    }
+}
+
+impl<TMessage> Default for Outbox<TMessage> {
+    fn default() -> Self {
+        Self {
+            message_arrived: Notify::new(),
+            messages: Mutex::new(VecDeque::new()),
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::sync::Arc;
+
+    use tokio::spawn;
+    use tokio::task::yield_now;
+
+    use super::Outbox;
+
+    #[tokio::test]
+    async fn delivers_messages_in_the_order_they_were_pushed() {
+        let outbox = Outbox::default();
+
+        outbox.push(1);
+        outbox.push(2);
+
+        assert_eq!([outbox.next().await, outbox.next().await], [1, 2]);
+    }
+
+    #[tokio::test]
+    async fn wakes_a_consumer_waiting_for_a_message_pushed_later() {
+        let outbox = Arc::new(Outbox::default());
+        let waiting_consumer = spawn({
+            let outbox = outbox.clone();
+
+            async move { outbox.next().await }
+        });
+
+        yield_now().await;
+        outbox.push(7);
+
+        assert_eq!(waiting_consumer.await.unwrap(), 7);
+    }
+}
diff --git a/paddler_request_registry/Cargo.toml b/paddler_request_registry/Cargo.toml
index b7ec0ad5d..7ef9c0145 100644
--- a/paddler_request_registry/Cargo.toml
+++ b/paddler_request_registry/Cargo.toml
@@ -10,6 +10,7 @@ version.workspace = true
 
 [dependencies]
 dashmap = { workspace = true }
+nanoid = { workspace = true }
 tokio = { workspace = true }
 
 [lints]
diff --git a/paddler_request_registry/src/request_registry.rs b/paddler_request_registry/src/request_registry.rs
index aeaffed10..448ef98eb 100644
--- a/paddler_request_registry/src/request_registry.rs
+++ b/paddler_request_registry/src/request_registry.rs
@@ -1,14 +1,32 @@
+use std::sync::atomic::AtomicU64;
+use std::sync::atomic::Ordering;
+
 use dashmap::DashMap;
 use dashmap::mapref::entry::Entry;
+use nanoid::nanoid;
 use tokio::sync::mpsc;
 
 use crate::request_delivery::RequestDelivery;
 
 pub struct RequestRegistry<TValue> {
     entries: DashMap<String, TValue>,
+    fresh_request_id_prefix: String,
+    next_fresh_request_id: AtomicU64,
 }
 
 impl<TValue> RequestRegistry<TValue> {
+    pub fn insert_fresh(&self, value: TValue) -> String {
+        let request_id = format!(
+            "{}-{}",
+            self.fresh_request_id_prefix,
+            self.next_fresh_request_id.fetch_add(1, Ordering::Relaxed)
+        );
+
+        self.entries.insert(request_id.clone(), value);
+
+        request_id
+    }
+
     #[must_use]
     pub fn insert_if_vacant(&self, request_id: String, value: TValue) -> bool {
         match self.entries.entry(request_id) {
@@ -53,6 +71,8 @@ impl<TValue> Default for RequestRegistry<TValue> {
     fn default() -> Self {
         Self {
             entries: DashMap::new(),
+            fresh_request_id_prefix: nanoid!(),
+            next_fresh_request_id: AtomicU64::new(0),
         }
     }
 }
@@ -73,6 +93,30 @@ mod tests {
         assert_eq!(registry.with_registered("request", |value| *value), Some(1));
     }
 
+    #[test]
+    fn registers_every_fresh_value_under_a_distinct_request_id() {
+        let registry = RequestRegistry::default();
+        let first_request_id = registry.insert_fresh(1);
+        let second_request_id = registry.insert_fresh(2);
+
+        assert_ne!(first_request_id, second_request_id);
+        assert_eq!(
+            [
+                registry.with_registered(&first_request_id, |value| *value),
+                registry.with_registered(&second_request_id, |value| *value)
+            ],
+            [Some(1), Some(2)]
+        );
+    }
+
+    #[test]
+    fn two_registries_hand_out_distinct_fresh_request_ids() {
+        assert_ne!(
+            RequestRegistry::default().insert_fresh(1),
+            RequestRegistry::default().insert_fresh(1)
+        );
+    }
+
     #[test]
     fn a_removed_request_is_no_longer_registered() {
         let registry = RequestRegistry::default();
diff --git a/paddler_request_registry/src/request_registry_guard.rs b/paddler_request_registry/src/request_registry_guard.rs
index d9dbcd4e7..64cf82550 100644
--- a/paddler_request_registry/src/request_registry_guard.rs
+++ b/paddler_request_registry/src/request_registry_guard.rs
@@ -8,6 +8,14 @@ pub struct RequestRegistryGuard<TValue> {
 }
 
 impl<TValue> RequestRegistryGuard<TValue> {
+    #[must_use]
+    pub fn register_fresh(registry: &Arc<RequestRegistry<TValue>>, value: TValue) -> Self {
+        Self {
+            request_id: registry.insert_fresh(value),
+            registry: registry.clone(),
+        }
+    }
+
     #[must_use]
     pub fn register(
         registry: &Arc<RequestRegistry<TValue>>,
@@ -48,4 +56,15 @@ mod tests {
 
         assert_eq!(registry.with_registered("request", |value| *value), None);
     }
+
+    #[test]
+    fn dropping_a_fresh_guard_frees_its_request_id() {
+        let registry = Arc::new(RequestRegistry::default());
+        let guard = RequestRegistryGuard::register_fresh(&registry, 1);
+        let request_id = guard.request_id.clone();
+
+        drop(guard);
+
+        assert_eq!(registry.with_registered(&request_id, |value| *value), None);
+    }
 }
diff --git a/paddler_sha256_hex/Cargo.toml b/paddler_sha256_hex/Cargo.toml
new file mode 100644
index 000000000..789978aa5
--- /dev/null
+++ b/paddler_sha256_hex/Cargo.toml
@@ -0,0 +1,15 @@
+[package]
+name = "paddler_sha256_hex"
+authors.workspace = true
+description.workspace = true
+edition.workspace = true
+homepage.workspace = true
+license.workspace = true
+repository.workspace = true
+version.workspace = true
+
+[dependencies]
+sha2 = { workspace = true }
+
+[lints]
+workspace = true
diff --git a/paddler_sha256_hex/src/lib.rs b/paddler_sha256_hex/src/lib.rs
new file mode 100644
index 000000000..a34bc0a04
--- /dev/null
+++ b/paddler_sha256_hex/src/lib.rs
@@ -0,0 +1 @@
+pub mod sha256_lowercase_hex;
diff --git a/paddler_sha256_hex/src/sha256_lowercase_hex.rs b/paddler_sha256_hex/src/sha256_lowercase_hex.rs
new file mode 100644
index 000000000..ffb7733cb
--- /dev/null
+++ b/paddler_sha256_hex/src/sha256_lowercase_hex.rs
@@ -0,0 +1,31 @@
+use sha2::Digest;
+use sha2::Sha256;
+
+const LOWERCASE_HEX_DIGITS: &[u8; 16] = b"0123456789abcdef";
+
+#[must_use]
+pub fn sha256_lowercase_hex(input: &[u8]) -> String {
+    Sha256::digest(input)
+        .iter()
+        .flat_map(|digest_byte| {
+            [
+                LOWERCASE_HEX_DIGITS[usize::from(digest_byte >> 4)],
+                LOWERCASE_HEX_DIGITS[usize::from(digest_byte & 0x0f)],
+            ]
+        })
+        .map(char::from)
+        .collect()
+}
+
+#[cfg(test)]
+mod tests {
+    use super::sha256_lowercase_hex;
+
+    #[test]
+    fn encodes_the_digest_high_nibble_first_in_lowercase_hex() {
+        assert_eq!(
+            sha256_lowercase_hex(b"abc"),
+            "ba7816bf8f01cfea414140de5dae2223b00361a396177a9cb410ff61f20015ad"
+        );
+    }
+}
diff --git a/paddler_spawn_resuming_panic/Cargo.toml b/paddler_spawn_resuming_panic/Cargo.toml
new file mode 100644
index 000000000..ba8c12c13
--- /dev/null
+++ b/paddler_spawn_resuming_panic/Cargo.toml
@@ -0,0 +1,15 @@
+[package]
+name = "paddler_spawn_resuming_panic"
+authors.workspace = true
+description.workspace = true
+edition.workspace = true
+homepage.workspace = true
+license.workspace = true
+repository.workspace = true
+version.workspace = true
+
+[dependencies]
+tokio = { workspace = true }
+
+[lints]
+workspace = true
diff --git a/paddler_spawn_resuming_panic/src/lib.rs b/paddler_spawn_resuming_panic/src/lib.rs
new file mode 100644
index 000000000..fc91e05fe
--- /dev/null
+++ b/paddler_spawn_resuming_panic/src/lib.rs
@@ -0,0 +1,2 @@
+pub mod spawn_blocking_resuming_panic;
+pub mod spawn_resuming_panic;
diff --git a/paddler_spawn_resuming_panic/src/spawn_blocking_resuming_panic.rs b/paddler_spawn_resuming_panic/src/spawn_blocking_resuming_panic.rs
new file mode 100644
index 000000000..261256ea1
--- /dev/null
+++ b/paddler_spawn_resuming_panic/src/spawn_blocking_resuming_panic.rs
@@ -0,0 +1,37 @@
+use std::panic::resume_unwind;
+
+use tokio::task::spawn_blocking;
+
+pub async fn spawn_blocking_resuming_panic<TBlockingWork, TOutput>(
+    blocking_work: TBlockingWork,
+) -> TOutput
+where
+    TBlockingWork: FnOnce() -> TOutput + Send + 'static,
+    TOutput: Send + 'static,
+{
+    spawn_blocking(blocking_work)
+        .await
+        .unwrap_or_else(|join_error| resume_unwind(join_error.into_panic()))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::spawn_blocking_resuming_panic;
+
+    #[tokio::test]
+    async fn the_awaiting_task_receives_what_the_blocking_work_returns() {
+        assert_eq!(
+            spawn_blocking_resuming_panic(|| "returned").await,
+            "returned"
+        );
+    }
+
+    #[tokio::test]
+    #[should_panic(expected = "the blocking work crashed")]
+    async fn the_awaiting_task_resumes_the_panic_of_the_blocking_work() {
+        spawn_blocking_resuming_panic(|| {
+            panic!("the blocking work crashed");
+        })
+        .await;
+    }
+}
diff --git a/paddler_spawn_resuming_panic/src/spawn_resuming_panic.rs b/paddler_spawn_resuming_panic/src/spawn_resuming_panic.rs
new file mode 100644
index 000000000..b7c34be56
--- /dev/null
+++ b/paddler_spawn_resuming_panic/src/spawn_resuming_panic.rs
@@ -0,0 +1,32 @@
+use std::panic::resume_unwind;
+
+use tokio::spawn;
+
+pub async fn spawn_resuming_panic<TWork>(work: TWork) -> TWork::Output
+where
+    TWork: Future + Send + 'static,
+    TWork::Output: Send + 'static,
+{
+    spawn(work)
+        .await
+        .unwrap_or_else(|join_error| resume_unwind(join_error.into_panic()))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::spawn_resuming_panic;
+
+    #[tokio::test]
+    async fn the_awaiting_task_receives_what_the_spawned_task_returns() {
+        assert_eq!(spawn_resuming_panic(async { "returned" }).await, "returned");
+    }
+
+    #[tokio::test]
+    #[should_panic(expected = "the spawned task crashed")]
+    async fn the_awaiting_task_resumes_the_panic_of_the_spawned_task() {
+        spawn_resuming_panic(async {
+            panic!("the spawned task crashed");
+        })
+        .await;
+    }
+}
diff --git a/paddler_state_database/Cargo.toml b/paddler_state_database/Cargo.toml
index c18ec4bc7..c1c06e74b 100644
--- a/paddler_state_database/Cargo.toml
+++ b/paddler_state_database/Cargo.toml
@@ -9,20 +9,16 @@ repository.workspace = true
 version.workspace = true
 
 [dependencies]
-async-trait = { workspace = true }
-log = { workspace = true }
 parking_lot = { workspace = true }
 serde = { workspace = true }
 serde_json = { workspace = true }
 shellexpand = { workspace = true }
+tempfile = { workspace = true }
 thiserror = { workspace = true }
 tokio = { workspace = true }
 url = { workspace = true }
 paddler_messaging = { workspace = true }
-
-[dev-dependencies]
-paddler_inference_parameters = { workspace = true }
-tempfile = { workspace = true }
+paddler_spawn_resuming_panic = { workspace = true }
 
 [lints]
 workspace = true
diff --git a/paddler_state_database/src/file.rs b/paddler_state_database/src/file.rs
deleted file mode 100644
index 6b63ceeb0..000000000
--- a/paddler_state_database/src/file.rs
+++ /dev/null
@@ -1,393 +0,0 @@
-use std::io::ErrorKind;
-use std::path::PathBuf;
-
-use async_trait::async_trait;
-use log::warn;
-use serde_json::from_str;
-use serde_json::to_string_pretty;
-use tokio::fs::File as TokioFile;
-use tokio::fs::read_to_string;
-use tokio::io::AsyncWriteExt;
-use tokio::sync::RwLock;
-use tokio::sync::watch;
-
-use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-
-use crate::schema::Schema;
-use crate::state_database::StateDatabase;
-use crate::state_database_error::StateDatabaseError;
-
-pub struct File {
-    balancer_desired_state_notify_tx: watch::Sender<BalancerDesiredState>,
-    path: PathBuf,
-    write_lock: RwLock<()>,
-}
-
-impl File {
-    #[must_use]
-    pub fn new(
-        balancer_desired_state_notify_tx: watch::Sender<BalancerDesiredState>,
-        path: PathBuf,
-    ) -> Self {
-        Self {
-            balancer_desired_state_notify_tx,
-            path,
-            write_lock: RwLock::new(()),
-        }
-    }
-
-    async fn read_schema_from_file(&self) -> Result<Schema, StateDatabaseError> {
-        match read_to_string(&self.path).await {
-            Ok(content) => {
-                from_str(&content).map_err(|source| StateDatabaseError::FileContentsInvalid {
-                    path: self.path.clone(),
-                    source,
-                })
-            }
-            Err(read_error) if read_error.kind() == ErrorKind::NotFound => {
-                warn!(
-                    "State database file not found; trying to store the default state: '{}'",
-                    self.path.display()
-                );
-
-                self.store_default_schema().await
-            }
-            Err(source) => Err(StateDatabaseError::FileReadFailed {
-                path: self.path.clone(),
-                source,
-            }),
-        }
-    }
-
-    async fn store_default_schema(&self) -> Result<Schema, StateDatabaseError> {
-        let schema = Schema::default();
-
-        self.store_schema(&schema).await?;
-
-        Ok(schema)
-    }
-
-    async fn store_schema(&self, schema: &Schema) -> Result<(), StateDatabaseError> {
-        let balancer_desired_state = schema.balancer_desired_state.clone();
-        let _lock = self.write_lock.write().await;
-
-        let serialized_schema =
-            to_string_pretty(schema).map_err(StateDatabaseError::SchemaUnserializable)?;
-        let mut file = TokioFile::create(&self.path).await.map_err(|source| {
-            StateDatabaseError::FileCreationFailed {
-                path: self.path.clone(),
-                source,
-            }
-        })?;
-
-        file.write_all(serialized_schema.as_bytes())
-            .await
-            .map_err(|source| StateDatabaseError::FileWriteFailed {
-                path: self.path.clone(),
-                source,
-            })?;
-        file.sync_all()
-            .await
-            .map_err(|source| StateDatabaseError::FileSyncFailed {
-                path: self.path.clone(),
-                source,
-            })?;
-
-        self.balancer_desired_state_notify_tx
-            .send_replace(balancer_desired_state);
-
-        Ok(())
-    }
-}
-
-#[async_trait]
-impl StateDatabase for File {
-    async fn read_balancer_desired_state(
-        &self,
-    ) -> Result<BalancerDesiredState, StateDatabaseError> {
-        self.read_schema_from_file()
-            .await
-            .map(|schema| schema.balancer_desired_state)
-    }
-
-    async fn store_balancer_desired_state(
-        &self,
-        balancer_desired_state: &BalancerDesiredState,
-    ) -> Result<(), StateDatabaseError> {
-        let mut schema = self.read_schema_from_file().await?;
-
-        schema.balancer_desired_state = balancer_desired_state.clone();
-
-        self.store_schema(&schema).await
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use std::io::ErrorKind;
-    use std::path::PathBuf;
-
-    use tempfile::NamedTempFile;
-    use tempfile::TempDir;
-    use tokio::fs::metadata;
-    use tokio::fs::read_to_string;
-    use tokio::fs::write;
-    use tokio::sync::watch;
-
-    use paddler_inference_parameters::inference_parameters::InferenceParameters;
-    use paddler_messaging::agent_desired_model::AgentDesiredModel;
-    use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-    use paddler_messaging::chat_template::ChatTemplate;
-
-    use super::File;
-    use crate::schema::Schema;
-    use crate::state_database::StateDatabase;
-    use crate::state_database_error::StateDatabaseError;
-
-    #[tokio::test]
-    async fn reads_back_the_stored_desired_state_from_the_file() {
-        let (balancer_desired_state_notify_tx, _balancer_desired_state_notify_rx) =
-            watch::channel(BalancerDesiredState::default());
-        let temp_dir = TempDir::new().unwrap();
-        let path = temp_dir.path().join("state.json");
-        let database = File::new(balancer_desired_state_notify_tx, path.clone());
-
-        let desired_state = BalancerDesiredState {
-            chat_template_override: None,
-            inference_parameters: BalancerDesiredState::default().inference_parameters,
-            model: AgentDesiredModel::LocalToAgent("stored_model_path".to_owned()),
-            multimodal_projection: AgentDesiredModel::None,
-            use_chat_template_override: false,
-        };
-
-        database
-            .store_balancer_desired_state(&desired_state)
-            .await
-            .unwrap();
-
-        let read_back = database.read_balancer_desired_state().await.unwrap();
-
-        assert_eq!(read_back.model, desired_state.model);
-        assert!(metadata(&path).await.unwrap().is_file());
-    }
-
-    #[tokio::test]
-    async fn reading_missing_file_stores_and_returns_default_state() {
-        let (balancer_desired_state_notify_tx, _balancer_desired_state_notify_rx) =
-            watch::channel(BalancerDesiredState::default());
-        let temp_dir = TempDir::new().unwrap();
-        let path = temp_dir.path().join("not_yet_created.json");
-        let database = File::new(balancer_desired_state_notify_tx, path.clone());
-
-        let read_state = database.read_balancer_desired_state().await.unwrap();
-
-        assert_eq!(read_state, BalancerDesiredState::default());
-        assert!(metadata(&path).await.unwrap().is_file());
-    }
-
-    #[tokio::test]
-    async fn reading_invalid_json_returns_parse_error() {
-        let (balancer_desired_state_notify_tx, _balancer_desired_state_notify_rx) =
-            watch::channel(BalancerDesiredState::default());
-        let temp_file = NamedTempFile::new().unwrap();
-        let path = temp_file.path().to_path_buf();
-        write(&path, b"this is not valid json").await.unwrap();
-        let database = File::new(balancer_desired_state_notify_tx, path.clone());
-
-        let read_result = database.read_balancer_desired_state().await;
-
-        assert!(matches!(
-            read_result,
-            Err(StateDatabaseError::FileContentsInvalid { path: invalid_path, .. }) if invalid_path == path
-        ));
-    }
-
-    #[tokio::test]
-    async fn reading_an_empty_file_fails_and_leaves_the_file_untouched() {
-        let (balancer_desired_state_notify_tx, _balancer_desired_state_notify_rx) =
-            watch::channel(BalancerDesiredState::default());
-        let temp_file = NamedTempFile::new().unwrap();
-        let path = temp_file.path().to_path_buf();
-        let database = File::new(balancer_desired_state_notify_tx, path.clone());
-
-        let read_result = database.read_balancer_desired_state().await;
-
-        assert!(matches!(
-            read_result,
-            Err(StateDatabaseError::FileContentsInvalid { path: invalid_path, .. }) if invalid_path == path
-        ));
-        assert_eq!(read_to_string(&path).await.unwrap(), "");
-    }
-
-    #[tokio::test]
-    async fn reading_a_directory_path_returns_non_not_found_error() {
-        let (balancer_desired_state_notify_tx, _balancer_desired_state_notify_rx) =
-            watch::channel(BalancerDesiredState::default());
-        let temp_dir = TempDir::new().unwrap();
-        let database = File::new(
-            balancer_desired_state_notify_tx,
-            temp_dir.path().to_path_buf(),
-        );
-
-        let read_result = database.read_balancer_desired_state().await;
-
-        assert!(matches!(
-            read_result,
-            Err(StateDatabaseError::FileReadFailed { source, .. })
-                if source.kind() == ErrorKind::IsADirectory
-        ));
-    }
-
-    #[tokio::test]
-    async fn storing_default_state_fails_when_parent_directory_is_missing() {
-        let (balancer_desired_state_notify_tx, _balancer_desired_state_notify_rx) =
-            watch::channel(BalancerDesiredState::default());
-        let temp_dir = TempDir::new().unwrap();
-        let path: PathBuf = temp_dir.path().join("missing_directory").join("state.json");
-        let database = File::new(balancer_desired_state_notify_tx, path.clone());
-
-        let read_result = database.read_balancer_desired_state().await;
-
-        assert!(matches!(
-            read_result,
-            Err(StateDatabaseError::FileCreationFailed { path: uncreated_path, source })
-                if uncreated_path == path && source.kind() == ErrorKind::NotFound
-        ));
-    }
-
-    #[tokio::test]
-    async fn updating_schema_fails_when_path_is_a_directory() {
-        let (balancer_desired_state_notify_tx, _balancer_desired_state_notify_rx) =
-            watch::channel(BalancerDesiredState::default());
-        let temp_dir = TempDir::new().unwrap();
-        let database = File::new(
-            balancer_desired_state_notify_tx,
-            temp_dir.path().to_path_buf(),
-        );
-
-        let store_result = database
-            .store_balancer_desired_state(&BalancerDesiredState::default())
-            .await;
-
-        assert!(matches!(
-            store_result,
-            Err(StateDatabaseError::FileReadFailed { source, .. })
-                if source.kind() == ErrorKind::IsADirectory
-        ));
-    }
-
-    #[tokio::test]
-    async fn storing_persists_the_state_while_nobody_is_listening() {
-        let (balancer_desired_state_notify_tx, balancer_desired_state_notify_rx) =
-            watch::channel(BalancerDesiredState::default());
-        drop(balancer_desired_state_notify_rx);
-        let temp_dir = TempDir::new().unwrap();
-        let path = temp_dir.path().join("state.json");
-        let database = File::new(balancer_desired_state_notify_tx, path);
-        let stored_state = BalancerDesiredState {
-            model: AgentDesiredModel::LocalToAgent("stored-model".to_owned()),
-            ..BalancerDesiredState::default()
-        };
-
-        database
-            .store_balancer_desired_state(&stored_state)
-            .await
-            .unwrap();
-
-        assert_eq!(
-            database.read_balancer_desired_state().await.unwrap(),
-            stored_state
-        );
-    }
-
-    #[cfg(target_os = "linux")]
-    #[tokio::test]
-    async fn storing_a_small_schema_on_a_character_device_fails_to_sync() {
-        let (balancer_desired_state_notify_tx, _balancer_desired_state_notify_rx) =
-            watch::channel(BalancerDesiredState::default());
-        let database = File::new(balancer_desired_state_notify_tx, PathBuf::from("/dev/full"));
-
-        let store_result = database.store_schema(&Schema::default()).await;
-
-        assert!(matches!(
-            store_result,
-            Err(StateDatabaseError::FileSyncFailed { source, .. })
-                if source.kind() == ErrorKind::InvalidInput
-        ));
-    }
-
-    #[cfg(target_os = "linux")]
-    #[tokio::test]
-    async fn storing_a_large_schema_surfaces_the_write_error_during_write_all() {
-        const TOKIO_FILE_BUFFER_BYTES: usize = 2 * 1024 * 1024;
-
-        let (balancer_desired_state_notify_tx, _balancer_desired_state_notify_rx) =
-            watch::channel(BalancerDesiredState::default());
-        let database = File::new(balancer_desired_state_notify_tx, PathBuf::from("/dev/full"));
-
-        let mut schema = Schema::default();
-        schema.balancer_desired_state.model =
-            AgentDesiredModel::LocalToAgent("x".repeat(TOKIO_FILE_BUFFER_BYTES * 2));
-
-        let store_result = database.store_schema(&schema).await;
-
-        assert!(matches!(
-            store_result,
-            Err(StateDatabaseError::FileWriteFailed { source, .. })
-                if source.kind() == ErrorKind::StorageFull
-        ));
-    }
-
-    #[cfg(target_os = "macos")]
-    #[tokio::test]
-    async fn storing_to_dev_full_surfaces_permission_denied() {
-        let (balancer_desired_state_notify_tx, _balancer_desired_state_notify_rx) =
-            watch::channel(BalancerDesiredState::default());
-        let database = File::new(balancer_desired_state_notify_tx, PathBuf::from("/dev/full"));
-
-        let store_result = database.store_schema(&Schema::default()).await;
-
-        assert!(matches!(
-            store_result,
-            Err(StateDatabaseError::FileCreationFailed { source, .. })
-                if source.kind() == ErrorKind::PermissionDenied
-        ));
-    }
-
-    #[tokio::test]
-    async fn persists_the_chat_template_override_across_instances() {
-        let temp_dir = TempDir::new().unwrap();
-        let path = temp_dir.path().join("state.json");
-
-        let chat_template = ChatTemplate {
-            content: "{% for message in messages %}{{ message.content }}{% endfor %}".to_owned(),
-        };
-        let desired_state = BalancerDesiredState {
-            chat_template_override: Some(chat_template.clone()),
-            inference_parameters: InferenceParameters::default(),
-            model: AgentDesiredModel::LocalToAgent("test_model_path".to_owned()),
-            multimodal_projection: AgentDesiredModel::None,
-            use_chat_template_override: true,
-        };
-
-        {
-            let (balancer_desired_state_tx, _balancer_desired_state_rx) =
-                watch::channel(BalancerDesiredState::default());
-            let database = File::new(balancer_desired_state_tx, path.clone());
-
-            database
-                .store_balancer_desired_state(&desired_state)
-                .await
-                .unwrap();
-        }
-
-        let (balancer_desired_state_tx, _balancer_desired_state_rx) =
-            watch::channel(BalancerDesiredState::default());
-        let database = File::new(balancer_desired_state_tx, path);
-        let read_back = database.read_balancer_desired_state().await.unwrap();
-
-        assert_eq!(read_back.chat_template_override, Some(chat_template));
-        assert!(read_back.use_chat_template_override);
-        assert_eq!(read_back.model, desired_state.model);
-    }
-}
diff --git a/paddler_state_database/src/lib.rs b/paddler_state_database/src/lib.rs
index 3f92099a9..06ab54cae 100644
--- a/paddler_state_database/src/lib.rs
+++ b/paddler_state_database/src/lib.rs
@@ -1,7 +1,9 @@
-pub mod file;
-pub mod memory;
 mod schema;
+mod schema_version_header;
 pub mod state_database;
 pub mod state_database_error;
+pub mod state_database_file;
+pub mod state_database_memory;
 mod state_database_schema_version;
 pub mod state_database_type;
+pub mod stored_balancer_desired_state;
diff --git a/paddler_state_database/src/memory.rs b/paddler_state_database/src/memory.rs
deleted file mode 100644
index d034a9b6b..000000000
--- a/paddler_state_database/src/memory.rs
+++ /dev/null
@@ -1,83 +0,0 @@
-use async_trait::async_trait;
-use parking_lot::RwLock;
-use tokio::sync::watch;
-
-use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-
-use crate::state_database::StateDatabase;
-use crate::state_database_error::StateDatabaseError;
-
-pub struct Memory {
-    balancer_desired_state: RwLock<BalancerDesiredState>,
-    balancer_desired_state_notify_tx: watch::Sender<BalancerDesiredState>,
-}
-
-impl Memory {
-    #[must_use]
-    pub const fn new(
-        balancer_desired_state_notify_tx: watch::Sender<BalancerDesiredState>,
-        initial_desired_state: BalancerDesiredState,
-    ) -> Self {
-        Self {
-            balancer_desired_state: RwLock::new(initial_desired_state),
-            balancer_desired_state_notify_tx,
-        }
-    }
-}
-
-#[async_trait]
-impl StateDatabase for Memory {
-    async fn read_balancer_desired_state(
-        &self,
-    ) -> Result<BalancerDesiredState, StateDatabaseError> {
-        Ok(self.balancer_desired_state.read().clone())
-    }
-
-    async fn store_balancer_desired_state(
-        &self,
-        state: &BalancerDesiredState,
-    ) -> Result<(), StateDatabaseError> {
-        {
-            let mut balancer_desired_state = self.balancer_desired_state.write();
-
-            *balancer_desired_state = state.clone();
-        }
-
-        self.balancer_desired_state_notify_tx
-            .send_replace(state.clone());
-
-        Ok(())
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use tokio::sync::watch;
-
-    use paddler_messaging::agent_desired_model::AgentDesiredModel;
-    use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-
-    use super::Memory;
-    use crate::state_database::StateDatabase;
-
-    #[tokio::test]
-    async fn reads_back_the_stored_desired_state() {
-        let (balancer_desired_state_tx, _balancer_desired_state_rx) =
-            watch::channel(BalancerDesiredState::default());
-        let database = Memory::new(balancer_desired_state_tx, BalancerDesiredState::default());
-        let desired_state = BalancerDesiredState {
-            model: AgentDesiredModel::LocalToAgent("test_model_path".to_owned()),
-            ..BalancerDesiredState::default()
-        };
-
-        database
-            .store_balancer_desired_state(&desired_state)
-            .await
-            .unwrap();
-
-        assert_eq!(
-            database.read_balancer_desired_state().await.unwrap(),
-            desired_state
-        );
-    }
-}
diff --git a/paddler_state_database/src/schema.rs b/paddler_state_database/src/schema.rs
index d16a028e0..c9e327539 100644
--- a/paddler_state_database/src/schema.rs
+++ b/paddler_state_database/src/schema.rs
@@ -5,33 +5,9 @@ use paddler_messaging::balancer_desired_state::BalancerDesiredState;
 
 use crate::state_database_schema_version::StateDatabaseSchemaVersion;
 
-#[derive(Default, Deserialize, Serialize)]
+#[derive(Deserialize, Serialize)]
 #[serde(deny_unknown_fields)]
 pub struct Schema {
     pub balancer_desired_state: BalancerDesiredState,
     pub version: StateDatabaseSchemaVersion,
 }
-
-#[cfg(test)]
-mod tests {
-    use serde_json::from_value;
-    use serde_json::json;
-    use serde_json::to_value;
-
-    use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-
-    use super::Schema;
-
-    #[test]
-    fn rejects_a_schema_without_a_known_version() {
-        let balancer_desired_state = to_value(BalancerDesiredState::default()).unwrap();
-
-        assert!(
-            from_value::<Schema>(json!({
-                "balancer_desired_state": balancer_desired_state,
-                "version": "2",
-            }))
-            .is_err()
-        );
-    }
-}
diff --git a/paddler_state_database/src/schema_version_header.rs b/paddler_state_database/src/schema_version_header.rs
new file mode 100644
index 000000000..b949a86f2
--- /dev/null
+++ b/paddler_state_database/src/schema_version_header.rs
@@ -0,0 +1,7 @@
+use serde::Deserialize;
+use serde_json::Value;
+
+#[derive(Deserialize)]
+pub struct SchemaVersionHeader {
+    pub version: Value,
+}
diff --git a/paddler_state_database/src/state_database.rs b/paddler_state_database/src/state_database.rs
index db1a586ee..f146b2ea7 100644
--- a/paddler_state_database/src/state_database.rs
+++ b/paddler_state_database/src/state_database.rs
@@ -1,16 +1,56 @@
-use async_trait::async_trait;
-
 use paddler_messaging::balancer_desired_state::BalancerDesiredState;
 
 use crate::state_database_error::StateDatabaseError;
+use crate::state_database_file::StateDatabaseFile;
+use crate::state_database_memory::StateDatabaseMemory;
+use crate::state_database_type::StateDatabaseType;
+use crate::stored_balancer_desired_state::StoredBalancerDesiredState;
+
+pub enum StateDatabase {
+    File(StateDatabaseFile),
+    Memory(StateDatabaseMemory),
+}
+
+impl StateDatabase {
+    #[must_use]
+    pub const fn in_memory(stored_balancer_desired_state: StoredBalancerDesiredState) -> Self {
+        Self::Memory(StateDatabaseMemory::new(stored_balancer_desired_state))
+    }
 
-#[async_trait]
-pub trait StateDatabase: Send + Sync {
-    async fn read_balancer_desired_state(&self)
-    -> Result<BalancerDesiredState, StateDatabaseError>;
+    pub async fn read_balancer_desired_state(
+        &self,
+    ) -> Result<StoredBalancerDesiredState, StateDatabaseError> {
+        match self {
+            Self::File(file) => file.read_balancer_desired_state().await,
+            Self::Memory(memory) => Ok(memory.read_balancer_desired_state()),
+        }
+    }
 
-    async fn store_balancer_desired_state(
+    pub async fn store_balancer_desired_state(
         &self,
-        state: &BalancerDesiredState,
-    ) -> Result<(), StateDatabaseError>;
+        balancer_desired_state: &BalancerDesiredState,
+    ) -> Result<(), StateDatabaseError> {
+        match self {
+            Self::File(file) => {
+                file.store_balancer_desired_state(balancer_desired_state)
+                    .await
+            }
+            Self::Memory(memory) => {
+                memory.store_balancer_desired_state(balancer_desired_state);
+
+                Ok(())
+            }
+        }
+    }
+}
+
+impl From<StateDatabaseType> for StateDatabase {
+    fn from(state_database_type: StateDatabaseType) -> Self {
+        match state_database_type {
+            StateDatabaseType::File(path) => Self::File(StateDatabaseFile::new(path)),
+            StateDatabaseType::Memory(stored_balancer_desired_state) => {
+                Self::in_memory(stored_balancer_desired_state)
+            }
+        }
+    }
 }
diff --git a/paddler_state_database/src/state_database_error.rs b/paddler_state_database/src/state_database_error.rs
index fe9caea20..ca350980e 100644
--- a/paddler_state_database/src/state_database_error.rs
+++ b/paddler_state_database/src/state_database_error.rs
@@ -13,12 +13,8 @@ pub enum StateDatabaseError {
         #[source]
         source: SerdeJsonError,
     },
-    #[error("Unable to create the state database file '{path_display}'", path_display = path.display())]
-    FileCreationFailed {
-        path: PathBuf,
-        #[source]
-        source: io::Error,
-    },
+    #[error("The state database file '{path_display}' has no directory to create its replacement in", path_display = path.display())]
+    FileHasNoParentDirectory { path: PathBuf },
     #[error("State database file path cannot be empty")]
     FilePathEmpty,
     #[error("To avoid ambiguity, needing to guess the full file path (and to stay safe overall), Paddler requires absolute paths.\nThe path you wanted is *probably* '{suggested_path_display}'. If that is so, pass it as '--state-database file://{suggested_path_display}'.", suggested_path_display = suggested_path.display())]
@@ -31,12 +27,31 @@ pub enum StateDatabaseError {
         #[source]
         source: io::Error,
     },
+    #[error("Unable to read the access mode of the state database file '{path_display}'", path_display = path.display())]
+    FilePermissionsUnreadable {
+        path: PathBuf,
+        #[source]
+        source: io::Error,
+    },
     #[error("Unable to read the state database file '{path_display}'", path_display = path.display())]
     FileReadFailed {
         path: PathBuf,
         #[source]
         source: io::Error,
     },
+    #[error("Unable to create a replacement of the state database file in '{directory_display}'", directory_display = directory.display())]
+    FileReplacementCreationFailed {
+        directory: PathBuf,
+        #[source]
+        source: io::Error,
+    },
+    #[error("Unable to replace the state database file '{path_display}' with '{replacement_path_display}'", path_display = path.display(), replacement_path_display = replacement_path.display())]
+    FileReplaceFailed {
+        path: PathBuf,
+        replacement_path: PathBuf,
+        #[source]
+        source: io::Error,
+    },
     #[error("Unable to sync the state database file '{path_display}'", path_display = path.display())]
     FileSyncFailed {
         path: PathBuf,
@@ -49,10 +64,15 @@ pub enum StateDatabaseError {
     FileWriteFailed {
         path: PathBuf,
         #[source]
-        source: io::Error,
+        source: SerdeJsonError,
+    },
+    #[error("The state database file '{path_display}' uses schema version {found}, but this version of Paddler only reads version 2. Start from a new state database file.", path_display = path.display())]
+    SchemaVersionUnsupported {
+        found: String,
+        path: PathBuf,
+        #[source]
+        source: SerdeJsonError,
     },
-    #[error("Failed to serialize the state database schema")]
-    SchemaUnserializable(#[source] SerdeJsonError),
     #[error("Unsupported state database scheme '{scheme}'")]
     SchemeUnsupported { scheme: String },
     #[error("Invalid state database URL '{input}'")]
diff --git a/paddler_state_database/src/state_database_file.rs b/paddler_state_database/src/state_database_file.rs
new file mode 100644
index 000000000..2f1d1e71b
--- /dev/null
+++ b/paddler_state_database/src/state_database_file.rs
@@ -0,0 +1,504 @@
+use std::fs::File;
+use std::fs::Permissions;
+use std::fs::metadata;
+use std::io::ErrorKind;
+use std::os::unix::fs::PermissionsExt as _;
+use std::path::Path;
+use std::path::PathBuf;
+use std::sync::Arc;
+
+use serde_json::from_str;
+use serde_json::from_value;
+use serde_json::to_writer_pretty;
+use tempfile::NamedTempFile;
+use tempfile::PersistError;
+use tokio::fs::read_to_string;
+use tokio::sync::Mutex;
+
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_spawn_resuming_panic::spawn_blocking_resuming_panic::spawn_blocking_resuming_panic;
+
+use crate::schema::Schema;
+use crate::schema_version_header::SchemaVersionHeader;
+use crate::state_database_error::StateDatabaseError;
+use crate::state_database_schema_version::StateDatabaseSchemaVersion;
+use crate::stored_balancer_desired_state::StoredBalancerDesiredState;
+
+const NEW_STATE_FILE_ACCESS_MODE: u32 = 0o600;
+
+fn replacement_permissions(state_path: &Path) -> Result<Permissions, StateDatabaseError> {
+    match metadata(state_path) {
+        Ok(state_metadata) => Ok(state_metadata.permissions()),
+        Err(missing_state) if missing_state.kind() == ErrorKind::NotFound => {
+            Ok(Permissions::from_mode(NEW_STATE_FILE_ACCESS_MODE))
+        }
+        Err(source) => Err(StateDatabaseError::FilePermissionsUnreadable {
+            path: state_path.to_path_buf(),
+            source,
+        }),
+    }
+}
+
+fn write_replacement(
+    replacement: &File,
+    replacement_path: &Path,
+    schema: &Schema,
+) -> Result<(), StateDatabaseError> {
+    to_writer_pretty(replacement, schema).map_err(|source| {
+        StateDatabaseError::FileWriteFailed {
+            path: replacement_path.to_path_buf(),
+            source,
+        }
+    })?;
+
+    replacement
+        .sync_all()
+        .map_err(|source| StateDatabaseError::FileSyncFailed {
+            path: replacement_path.to_path_buf(),
+            source,
+        })
+}
+
+fn replace_state_file(state_path: &Path, schema: &Schema) -> Result<(), StateDatabaseError> {
+    let state_directory =
+        state_path
+            .parent()
+            .ok_or_else(|| StateDatabaseError::FileHasNoParentDirectory {
+                path: state_path.to_path_buf(),
+            })?;
+    let permissions = replacement_permissions(state_path)?;
+    let replacement = NamedTempFile::new_in(state_directory)
+        .and_then(|replacement| {
+            replacement
+                .as_file()
+                .set_permissions(permissions)
+                .map(|()| replacement)
+        })
+        .map_err(|source| StateDatabaseError::FileReplacementCreationFailed {
+            directory: state_directory.to_path_buf(),
+            source,
+        })?;
+    let replacement_path = replacement.path().to_path_buf();
+
+    write_replacement(replacement.as_file(), &replacement_path, schema).and_then(|()| {
+        replacement
+            .persist(state_path)
+            .map(drop)
+            .map_err(
+                |PersistError { error, .. }| StateDatabaseError::FileReplaceFailed {
+                    path: state_path.to_path_buf(),
+                    replacement_path,
+                    source: error,
+                },
+            )
+    })
+}
+
+pub struct StateDatabaseFile {
+    path: PathBuf,
+    store_lock: Arc<Mutex<()>>,
+}
+
+impl StateDatabaseFile {
+    #[must_use]
+    pub fn new(path: PathBuf) -> Self {
+        Self {
+            path,
+            store_lock: Arc::new(Mutex::new(())),
+        }
+    }
+
+    fn parse_schema(&self, content: &str) -> Result<Schema, StateDatabaseError> {
+        let SchemaVersionHeader { version } =
+            from_str(content).map_err(|source| StateDatabaseError::FileContentsInvalid {
+                path: self.path.clone(),
+                source,
+            })?;
+
+        from_value::<StateDatabaseSchemaVersion>(version.clone()).map_err(|source| {
+            StateDatabaseError::SchemaVersionUnsupported {
+                found: version.to_string(),
+                path: self.path.clone(),
+                source,
+            }
+        })?;
+
+        from_str(content).map_err(|source| StateDatabaseError::FileContentsInvalid {
+            path: self.path.clone(),
+            source,
+        })
+    }
+
+    pub async fn read_balancer_desired_state(
+        &self,
+    ) -> Result<StoredBalancerDesiredState, StateDatabaseError> {
+        match read_to_string(&self.path).await {
+            Ok(content) => Ok(StoredBalancerDesiredState::Stored(Box::new(
+                self.parse_schema(&content)?.balancer_desired_state,
+            ))),
+            Err(read_error) if read_error.kind() == ErrorKind::NotFound => {
+                Ok(StoredBalancerDesiredState::NothingStored)
+            }
+            Err(source) => Err(StateDatabaseError::FileReadFailed {
+                path: self.path.clone(),
+                source,
+            }),
+        }
+    }
+
+    pub async fn store_balancer_desired_state(
+        &self,
+        balancer_desired_state: &BalancerDesiredState,
+    ) -> Result<(), StateDatabaseError> {
+        let schema = Schema {
+            balancer_desired_state: balancer_desired_state.clone(),
+            version: StateDatabaseSchemaVersion::default(),
+        };
+        let store_guard = self.store_lock.clone().lock_owned().await;
+        let state_path = self.path.clone();
+
+        spawn_blocking_resuming_panic(move || {
+            let replacement_outcome = replace_state_file(&state_path, &schema);
+
+            drop(store_guard);
+
+            replacement_outcome
+        })
+        .await
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::error::Error as _;
+    #[cfg(target_os = "linux")]
+    use std::fs::OpenOptions;
+    use std::fs::Permissions;
+    use std::io::ErrorKind;
+    use std::os::unix::fs::PermissionsExt as _;
+    #[cfg(target_os = "linux")]
+    use std::os::unix::fs::symlink;
+    use std::path::Path;
+    use std::path::PathBuf;
+
+    use tempfile::NamedTempFile;
+    use tempfile::TempDir;
+    use tokio::fs::metadata;
+    use tokio::fs::read_to_string;
+    use tokio::fs::set_permissions;
+    use tokio::fs::write;
+
+    use paddler_messaging::agent_desired_model::AgentDesiredModel;
+    use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+
+    use super::StateDatabaseFile;
+    #[cfg(target_os = "linux")]
+    use super::write_replacement;
+    use crate::schema::Schema;
+    use crate::state_database_error::StateDatabaseError;
+    use crate::state_database_schema_version::StateDatabaseSchemaVersion;
+    use crate::stored_balancer_desired_state::StoredBalancerDesiredState;
+
+    fn state_with_model(model_uri: &str) -> BalancerDesiredState {
+        BalancerDesiredState {
+            model: AgentDesiredModel::Uri(model_uri.to_owned()),
+            ..BalancerDesiredState::default()
+        }
+    }
+
+    #[tokio::test]
+    async fn reads_back_the_stored_desired_state_from_the_file() {
+        let temp_dir = TempDir::new().unwrap();
+        let path = temp_dir.path().join("state.json");
+        let database = StateDatabaseFile::new(path.clone());
+        let desired_state = state_with_model("stored_model_uri");
+
+        database
+            .store_balancer_desired_state(&desired_state)
+            .await
+            .unwrap();
+
+        assert_eq!(
+            database.read_balancer_desired_state().await.unwrap(),
+            StoredBalancerDesiredState::Stored(Box::new(desired_state))
+        );
+        assert!(metadata(&path).await.unwrap().is_file());
+    }
+
+    #[tokio::test]
+    async fn reading_a_missing_file_reports_nothing_stored_and_creates_no_file() {
+        let temp_dir = TempDir::new().unwrap();
+        let path = temp_dir.path().join("not_yet_created.json");
+
+        assert_eq!(
+            StateDatabaseFile::new(path.clone())
+                .read_balancer_desired_state()
+                .await
+                .unwrap(),
+            StoredBalancerDesiredState::NothingStored
+        );
+        assert!(
+            metadata(&path)
+                .await
+                .is_err_and(|missing| missing.kind() == ErrorKind::NotFound)
+        );
+    }
+
+    #[tokio::test]
+    async fn reading_invalid_json_returns_parse_error() {
+        let temp_file = NamedTempFile::new().unwrap();
+        let path = temp_file.path().to_path_buf();
+
+        write(&path, b"this is not valid json").await.unwrap();
+
+        assert!(matches!(
+            StateDatabaseFile::new(path.clone()).read_balancer_desired_state().await,
+            Err(StateDatabaseError::FileContentsInvalid { path: invalid_path, .. }) if invalid_path == path
+        ));
+    }
+
+    #[tokio::test]
+    async fn reading_an_empty_file_fails_and_leaves_the_file_untouched() {
+        let temp_file = NamedTempFile::new().unwrap();
+        let path = temp_file.path().to_path_buf();
+
+        assert!(matches!(
+            StateDatabaseFile::new(path.clone()).read_balancer_desired_state().await,
+            Err(StateDatabaseError::FileContentsInvalid { path: invalid_path, .. }) if invalid_path == path
+        ));
+        assert_eq!(read_to_string(&path).await.unwrap(), "");
+    }
+
+    #[tokio::test]
+    async fn reading_a_directory_path_returns_non_not_found_error() {
+        let temp_dir = TempDir::new().unwrap();
+
+        assert!(matches!(
+            StateDatabaseFile::new(temp_dir.path().to_path_buf()).read_balancer_desired_state().await,
+            Err(StateDatabaseError::FileReadFailed { source, .. })
+                if source.kind() == ErrorKind::IsADirectory
+        ));
+    }
+
+    #[tokio::test]
+    async fn storing_fails_when_the_parent_directory_is_missing() {
+        let temp_dir = TempDir::new().unwrap();
+        let missing_directory = temp_dir.path().join("missing_directory");
+        let database = StateDatabaseFile::new(missing_directory.join("state.json"));
+
+        assert!(matches!(
+            database
+                .store_balancer_desired_state(&BalancerDesiredState::default())
+                .await,
+            Err(StateDatabaseError::FileReplacementCreationFailed { directory, source })
+                if directory == missing_directory && source.kind() == ErrorKind::NotFound
+        ));
+    }
+
+    #[tokio::test]
+    async fn storing_at_the_filesystem_root_has_no_directory_for_its_replacement() {
+        assert!(matches!(
+            StateDatabaseFile::new(PathBuf::from("/"))
+                .store_balancer_desired_state(&BalancerDesiredState::default())
+                .await,
+            Err(StateDatabaseError::FileHasNoParentDirectory { path }) if path == Path::new("/")
+        ));
+    }
+
+    #[tokio::test]
+    async fn storing_under_a_regular_file_cannot_read_the_access_mode_of_its_state() {
+        let regular_file = NamedTempFile::new().unwrap();
+        let path = regular_file.path().join("state.json");
+
+        assert!(matches!(
+            StateDatabaseFile::new(path.clone())
+                .store_balancer_desired_state(&BalancerDesiredState::default())
+                .await,
+            Err(StateDatabaseError::FilePermissionsUnreadable { path: unreadable_path, source })
+                if unreadable_path == path && source.kind() == ErrorKind::NotADirectory
+        ));
+    }
+
+    #[tokio::test]
+    async fn storing_fails_when_a_directory_occupies_the_state_path() {
+        let temp_dir = TempDir::new().unwrap();
+        let database = StateDatabaseFile::new(temp_dir.path().to_path_buf());
+
+        assert!(matches!(
+            database
+                .store_balancer_desired_state(&BalancerDesiredState::default())
+                .await,
+            Err(StateDatabaseError::FileReplaceFailed { path, .. }) if path == temp_dir.path()
+        ));
+    }
+
+    #[tokio::test]
+    async fn a_store_that_fails_leaves_the_previous_state_readable() {
+        let temp_dir = TempDir::new().unwrap();
+        let database = StateDatabaseFile::new(temp_dir.path().join("state.json"));
+        let previous_state = state_with_model("previous_model_uri");
+
+        database
+            .store_balancer_desired_state(&previous_state)
+            .await
+            .unwrap();
+        set_permissions(temp_dir.path(), Permissions::from_mode(0o500))
+            .await
+            .unwrap();
+
+        let failed_store = database
+            .store_balancer_desired_state(&state_with_model("next_model_uri"))
+            .await;
+        let state_after_the_failed_store = database.read_balancer_desired_state().await.unwrap();
+
+        set_permissions(temp_dir.path(), Permissions::from_mode(0o700))
+            .await
+            .unwrap();
+
+        assert!(matches!(
+            failed_store,
+            Err(StateDatabaseError::FileReplacementCreationFailed { source, .. })
+                if source.kind() == ErrorKind::PermissionDenied
+        ));
+        assert_eq!(
+            state_after_the_failed_store,
+            StoredBalancerDesiredState::Stored(Box::new(previous_state))
+        );
+    }
+
+    #[cfg(target_os = "linux")]
+    fn write_replacement_to_device(device_path: &Path) -> Result<(), StateDatabaseError> {
+        write_replacement(
+            &OpenOptions::new().write(true).open(device_path).unwrap(),
+            device_path,
+            &Schema {
+                balancer_desired_state: BalancerDesiredState::default(),
+                version: StateDatabaseSchemaVersion::default(),
+            },
+        )
+    }
+
+    #[cfg(target_os = "linux")]
+    #[test]
+    fn a_replacement_on_a_device_that_refuses_to_sync_fails_to_sync() {
+        assert!(matches!(
+            write_replacement_to_device(Path::new("/dev/null")),
+            Err(StateDatabaseError::FileSyncFailed { path, source })
+                if path == Path::new("/dev/null") && source.kind() == ErrorKind::InvalidInput
+        ));
+    }
+
+    #[cfg(target_os = "linux")]
+    #[test]
+    fn a_replacement_on_a_full_device_fails_to_write_its_schema() {
+        assert!(matches!(
+            write_replacement_to_device(Path::new("/dev/full")),
+            Err(StateDatabaseError::FileWriteFailed { path, source })
+                if path == Path::new("/dev/full")
+                    && source.io_error_kind() == Some(ErrorKind::StorageFull)
+        ));
+    }
+
+    #[cfg(target_os = "linux")]
+    #[tokio::test]
+    async fn storing_never_writes_through_a_link_planted_at_the_old_replacement_path() {
+        let temp_dir = TempDir::new().unwrap();
+        let unrelated_file = temp_dir.path().join("unrelated.txt");
+
+        write(&unrelated_file, "unrelated contents").await.unwrap();
+        symlink(&unrelated_file, temp_dir.path().join("state.json.next")).unwrap();
+
+        StateDatabaseFile::new(temp_dir.path().join("state.json"))
+            .store_balancer_desired_state(&state_with_model("stored_model_uri"))
+            .await
+            .unwrap();
+
+        assert_eq!(
+            read_to_string(&unrelated_file).await.unwrap(),
+            "unrelated contents"
+        );
+    }
+
+    #[tokio::test]
+    async fn storing_keeps_the_access_mode_of_the_replaced_state_file() {
+        let temp_dir = TempDir::new().unwrap();
+        let path = temp_dir.path().join("state.json");
+        let database = StateDatabaseFile::new(path.clone());
+
+        database
+            .store_balancer_desired_state(&state_with_model("previous_model_uri"))
+            .await
+            .unwrap();
+        set_permissions(&path, Permissions::from_mode(0o640))
+            .await
+            .unwrap();
+        database
+            .store_balancer_desired_state(&state_with_model("next_model_uri"))
+            .await
+            .unwrap();
+
+        assert_eq!(
+            metadata(&path).await.unwrap().permissions().mode() & 0o777,
+            0o640
+        );
+    }
+
+    #[tokio::test]
+    async fn a_new_state_file_is_readable_only_by_its_owner() {
+        let temp_dir = TempDir::new().unwrap();
+        let path = temp_dir.path().join("state.json");
+
+        StateDatabaseFile::new(path.clone())
+            .store_balancer_desired_state(&state_with_model("stored_model_uri"))
+            .await
+            .unwrap();
+
+        assert_eq!(
+            metadata(&path).await.unwrap().permissions().mode() & 0o777,
+            0o600
+        );
+    }
+
+    #[tokio::test]
+    async fn reading_a_file_of_an_older_schema_version_names_the_version_it_found() {
+        let temp_file = NamedTempFile::new().unwrap();
+        let path = temp_file.path().to_path_buf();
+
+        write(&path, br#"{"balancer_desired_state": {}, "version": "1"}"#)
+            .await
+            .unwrap();
+
+        assert_eq!(
+            StateDatabaseFile::new(path.clone())
+                .read_balancer_desired_state()
+                .await
+                .err()
+                .as_ref()
+                .map(ToString::to_string),
+            Some(format!(
+                r#"The state database file '{}' uses schema version "1", but this version of Paddler only reads version 2. Start from a new state database file."#,
+                path.display()
+            ))
+        );
+    }
+
+    #[tokio::test]
+    async fn reading_a_file_of_an_older_schema_version_keeps_why_the_version_was_refused() {
+        let temp_file = NamedTempFile::new().unwrap();
+        let path = temp_file.path().to_path_buf();
+
+        write(&path, br#"{"balancer_desired_state": {}, "version": "1"}"#)
+            .await
+            .unwrap();
+
+        let refusal = StateDatabaseFile::new(path)
+            .read_balancer_desired_state()
+            .await
+            .unwrap_err();
+
+        assert_eq!(
+            refusal.source().map(ToString::to_string),
+            Some("unknown variant `1`, expected `2`".to_owned())
+        );
+    }
+}
diff --git a/paddler_state_database/src/state_database_memory.rs b/paddler_state_database/src/state_database_memory.rs
new file mode 100644
index 000000000..f72646180
--- /dev/null
+++ b/paddler_state_database/src/state_database_memory.rs
@@ -0,0 +1,57 @@
+use parking_lot::RwLock;
+
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+
+use crate::stored_balancer_desired_state::StoredBalancerDesiredState;
+
+pub struct StateDatabaseMemory {
+    stored_balancer_desired_state: RwLock<StoredBalancerDesiredState>,
+}
+
+impl StateDatabaseMemory {
+    #[must_use]
+    pub const fn new(stored_balancer_desired_state: StoredBalancerDesiredState) -> Self {
+        Self {
+            stored_balancer_desired_state: RwLock::new(stored_balancer_desired_state),
+        }
+    }
+
+    #[must_use]
+    pub fn read_balancer_desired_state(&self) -> StoredBalancerDesiredState {
+        self.stored_balancer_desired_state.read().clone()
+    }
+
+    pub fn store_balancer_desired_state(&self, balancer_desired_state: &BalancerDesiredState) {
+        *self.stored_balancer_desired_state.write() =
+            StoredBalancerDesiredState::Stored(Box::new(balancer_desired_state.clone()));
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use paddler_messaging::agent_desired_model::AgentDesiredModel;
+    use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+    use paddler_messaging::decision_settings::DecisionSettings;
+
+    use super::StateDatabaseMemory;
+    use crate::stored_balancer_desired_state::StoredBalancerDesiredState;
+
+    #[test]
+    fn reads_back_the_stored_desired_state() {
+        let database = StateDatabaseMemory::new(StoredBalancerDesiredState::NothingStored);
+        let desired_state = BalancerDesiredState {
+            decision: DecisionSettings {
+                pointer_head: AgentDesiredModel::Uri("test_pointer_head_path".to_owned()),
+            },
+            model: AgentDesiredModel::Uri("test_model_path".to_owned()),
+            ..BalancerDesiredState::default()
+        };
+
+        database.store_balancer_desired_state(&desired_state);
+
+        assert_eq!(
+            database.read_balancer_desired_state(),
+            StoredBalancerDesiredState::Stored(Box::new(desired_state))
+        );
+    }
+}
diff --git a/paddler_state_database/src/state_database_schema_version.rs b/paddler_state_database/src/state_database_schema_version.rs
index 1f142a020..51c97f0c1 100644
--- a/paddler_state_database/src/state_database_schema_version.rs
+++ b/paddler_state_database/src/state_database_schema_version.rs
@@ -4,6 +4,6 @@ use serde::Serialize;
 #[derive(Clone, Copy, Debug, Default, Deserialize, Eq, PartialEq, Serialize)]
 pub enum StateDatabaseSchemaVersion {
     #[default]
-    #[serde(rename = "1")]
-    V1,
+    #[serde(rename = "2")]
+    V2,
 }
diff --git a/paddler_state_database/src/state_database_type.rs b/paddler_state_database/src/state_database_type.rs
index 5cc01f3ba..31696f85a 100644
--- a/paddler_state_database/src/state_database_type.rs
+++ b/paddler_state_database/src/state_database_type.rs
@@ -6,9 +6,8 @@ use std::str::FromStr;
 use shellexpand::tilde;
 use url::Url;
 
-use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-
 use crate::state_database_error::StateDatabaseError;
+use crate::stored_balancer_desired_state::StoredBalancerDesiredState;
 
 fn relative_path_error(path: &str) -> StateDatabaseError {
     match absolute(tilde(path).to_string()) {
@@ -20,10 +19,10 @@ fn relative_path_error(path: &str) -> StateDatabaseError {
     }
 }
 
-#[derive(Clone)]
+#[derive(Clone, Debug, PartialEq)]
 pub enum StateDatabaseType {
     File(PathBuf),
-    Memory(Box<BalancerDesiredState>),
+    Memory(StoredBalancerDesiredState),
 }
 
 impl FromStr for StateDatabaseType {
@@ -54,7 +53,7 @@ impl FromStr for StateDatabaseType {
 
                 Ok(Self::File(PathBuf::from(path)))
             }
-            "memory" => Ok(Self::Memory(Box::default())),
+            "memory" => Ok(Self::Memory(StoredBalancerDesiredState::NothingStored)),
             scheme => Err(StateDatabaseError::SchemeUnsupported {
                 scheme: scheme.to_owned(),
             }),
@@ -73,14 +72,13 @@ mod tests {
 
     use super::StateDatabaseType;
     use crate::state_database_error::StateDatabaseError;
+    use crate::stored_balancer_desired_state::StoredBalancerDesiredState;
 
     #[test]
-    fn parses_a_memory_url() {
-        let result = StateDatabaseType::from_str("memory://").unwrap();
-
+    fn parses_a_memory_url_into_an_empty_database() {
         assert_eq!(
-            discriminant(&result),
-            discriminant(&StateDatabaseType::Memory(Box::default())),
+            StateDatabaseType::from_str("memory://").unwrap(),
+            StateDatabaseType::Memory(StoredBalancerDesiredState::NothingStored)
         );
     }
 
@@ -120,19 +118,9 @@ mod tests {
 
     #[test]
     fn parses_a_file_url_with_an_absolute_path() {
-        #[cfg(unix)]
-        let expected_path = "/absolute/path";
-        #[cfg(unix)]
-        let url = "file:///absolute/path";
-        #[cfg(windows)]
-        let expected_path = "C:/absolute/path";
-        #[cfg(windows)]
-        let url = "file://C:/absolute/path";
-
-        let result = StateDatabaseType::from_str(url).unwrap();
-
-        assert!(
-            matches!(&result, StateDatabaseType::File(path) if path == &PathBuf::from(expected_path))
+        assert_eq!(
+            StateDatabaseType::from_str("file:///absolute/path").unwrap(),
+            StateDatabaseType::File(PathBuf::from("/absolute/path"))
         );
     }
 
diff --git a/paddler_state_database/src/stored_balancer_desired_state.rs b/paddler_state_database/src/stored_balancer_desired_state.rs
new file mode 100644
index 000000000..f8d14acbd
--- /dev/null
+++ b/paddler_state_database/src/stored_balancer_desired_state.rs
@@ -0,0 +1,7 @@
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+
+#[derive(Clone, Debug, PartialEq)]
+pub enum StoredBalancerDesiredState {
+    NothingStored,
+    Stored(Box<BalancerDesiredState>),
+}
diff --git a/paddler_tcp_buffer_limits/Cargo.toml b/paddler_tcp_buffer_limits/Cargo.toml
new file mode 100644
index 000000000..9d9092eb5
--- /dev/null
+++ b/paddler_tcp_buffer_limits/Cargo.toml
@@ -0,0 +1,19 @@
+[package]
+name = "paddler_tcp_buffer_limits"
+authors.workspace = true
+description = "Reads the kernel TCP buffer limits so Paddler tests can outlast both socket buffers of a loopback connection"
+edition.workspace = true
+homepage.workspace = true
+license.workspace = true
+publish = false
+repository.workspace = true
+version.workspace = true
+
+[dependencies]
+thiserror = { workspace = true }
+
+[dev-dependencies]
+tempfile = { workspace = true }
+
+[lints]
+workspace = true
diff --git a/paddler_tcp_buffer_limits/src/lib.rs b/paddler_tcp_buffer_limits/src/lib.rs
new file mode 100644
index 000000000..0e780f8b0
--- /dev/null
+++ b/paddler_tcp_buffer_limits/src/lib.rs
@@ -0,0 +1,2 @@
+pub mod tcp_buffer_limits_error;
+pub mod tcp_buffer_limits_file;
diff --git a/paddler_tcp_buffer_limits/src/tcp_buffer_limits_error.rs b/paddler_tcp_buffer_limits/src/tcp_buffer_limits_error.rs
new file mode 100644
index 000000000..da7c2a1fe
--- /dev/null
+++ b/paddler_tcp_buffer_limits/src/tcp_buffer_limits_error.rs
@@ -0,0 +1,24 @@
+use std::io;
+use std::num::ParseIntError;
+use std::path::PathBuf;
+
+use thiserror::Error;
+
+#[derive(Debug, Error)]
+pub enum TcpBufferLimitsError {
+    #[error("The TCP buffer limits in '{path_display}' are not exactly a minimum, a default, and a maximum: '{limits}'", path_display = path.display())]
+    LimitsAreNotThreeByteCounts { limits: String, path: PathBuf },
+    #[error("Unable to read the TCP buffer limits from '{path_display}'", path_display = path.display())]
+    LimitsUnreadable {
+        path: PathBuf,
+        #[source]
+        source: io::Error,
+    },
+    #[error("The largest TCP buffer limit '{largest_limit}' in '{path_display}' is not a byte count", path_display = path.display())]
+    LargestLimitIsNotAByteCount {
+        largest_limit: String,
+        path: PathBuf,
+        #[source]
+        source: ParseIntError,
+    },
+}
diff --git a/paddler_tcp_buffer_limits/src/tcp_buffer_limits_file.rs b/paddler_tcp_buffer_limits/src/tcp_buffer_limits_file.rs
new file mode 100644
index 000000000..03e825b34
--- /dev/null
+++ b/paddler_tcp_buffer_limits/src/tcp_buffer_limits_file.rs
@@ -0,0 +1,112 @@
+use std::fs::read_to_string;
+use std::path::PathBuf;
+
+use crate::tcp_buffer_limits_error::TcpBufferLimitsError;
+
+const KERNEL_TCP_RECEIVE_BUFFER_LIMITS: &str = "/proc/sys/net/ipv4/tcp_rmem";
+const KERNEL_TCP_SEND_BUFFER_LIMITS: &str = "/proc/sys/net/ipv4/tcp_wmem";
+
+pub struct TcpBufferLimitsFile {
+    pub path: PathBuf,
+}
+
+impl TcpBufferLimitsFile {
+    #[must_use]
+    pub fn of_kernel_receive_buffers() -> Self {
+        Self {
+            path: PathBuf::from(KERNEL_TCP_RECEIVE_BUFFER_LIMITS),
+        }
+    }
+
+    #[must_use]
+    pub fn of_kernel_send_buffers() -> Self {
+        Self {
+            path: PathBuf::from(KERNEL_TCP_SEND_BUFFER_LIMITS),
+        }
+    }
+
+    pub fn largest_buffer(&self) -> Result<usize, TcpBufferLimitsError> {
+        let limits = read_to_string(&self.path).map_err(|source| {
+            TcpBufferLimitsError::LimitsUnreadable {
+                path: self.path.clone(),
+                source,
+            }
+        })?;
+        let byte_counts = limits.split_whitespace().collect::<Vec<_>>();
+        let [_smallest_limit, _default_limit, largest_limit] = byte_counts.as_slice() else {
+            return Err(TcpBufferLimitsError::LimitsAreNotThreeByteCounts {
+                limits: limits.clone(),
+                path: self.path.clone(),
+            });
+        };
+
+        largest_limit
+            .parse()
+            .map_err(|source| TcpBufferLimitsError::LargestLimitIsNotAByteCount {
+                largest_limit: (*largest_limit).to_owned(),
+                path: self.path.clone(),
+                source,
+            })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::fs::write;
+
+    use tempfile::TempDir;
+
+    use super::TcpBufferLimitsFile;
+    use crate::tcp_buffer_limits_error::TcpBufferLimitsError;
+
+    fn limits_file_containing(directory: &TempDir, limits: &str) -> TcpBufferLimitsFile {
+        let path = directory.path().join("tcp_wmem");
+
+        write(&path, limits).unwrap();
+
+        TcpBufferLimitsFile { path }
+    }
+
+    #[test]
+    fn the_largest_buffer_is_the_last_of_the_three_limits() {
+        let directory = TempDir::new().unwrap();
+        let limits_file = limits_file_containing(&directory, "4096\t16384\t4194304\n");
+
+        assert_eq!(limits_file.largest_buffer().unwrap(), 4_194_304);
+    }
+
+    #[test]
+    fn a_missing_limits_file_is_unreadable() {
+        let directory = TempDir::new().unwrap();
+        let limits_file = TcpBufferLimitsFile {
+            path: directory.path().join("missing"),
+        };
+
+        assert!(matches!(
+            limits_file.largest_buffer(),
+            Err(TcpBufferLimitsError::LimitsUnreadable { path, .. }) if path == limits_file.path
+        ));
+    }
+
+    #[test]
+    fn limits_other_than_three_values_are_rejected() {
+        let directory = TempDir::new().unwrap();
+        let limits_file = limits_file_containing(&directory, "4096\t16384\n");
+
+        assert!(matches!(
+            limits_file.largest_buffer(),
+            Err(TcpBufferLimitsError::LimitsAreNotThreeByteCounts { limits, .. }) if limits == "4096\t16384\n"
+        ));
+    }
+
+    #[test]
+    fn a_largest_limit_that_is_not_a_byte_count_is_rejected() {
+        let directory = TempDir::new().unwrap();
+        let limits_file = limits_file_containing(&directory, "4096\t16384\tunbounded\n");
+
+        assert!(matches!(
+            limits_file.largest_buffer(),
+            Err(TcpBufferLimitsError::LargestLimitIsNotAByteCount { largest_limit, .. }) if largest_limit == "unbounded"
+        ));
+    }
+}
diff --git a/paddler_test_cluster_harness/Cargo.toml b/paddler_test_cluster_harness/Cargo.toml
index aa6fc0b3c..8c4e27bf5 100644
--- a/paddler_test_cluster_harness/Cargo.toml
+++ b/paddler_test_cluster_harness/Cargo.toml
@@ -14,13 +14,16 @@ anyhow = { workspace = true }
 async-openai = { workspace = true }
 async-trait = { workspace = true }
 base64 = { workspace = true }
+futures = { workspace = true }
 futures-util = { workspace = true }
+hf-hub = { workspace = true }
 paddler_balancer = { workspace = true }
 paddler_client = { workspace = true }
 paddler_inference_parameters = { workspace = true }
 paddler_messaging = { workspace = true }
+paddler_model_source = { workspace = true }
 quick-xml = { workspace = true }
-serde = { workspace = true }
+reqwest = { workspace = true }
 serde_json = { workspace = true }
 tempfile = { workspace = true }
 thiserror = { workspace = true }
@@ -32,7 +35,7 @@ url = { workspace = true }
 
 [dev-dependencies]
 data-url = { workspace = true }
-hf-hub = { workspace = true }
+paddler_agent_status = { workspace = true }
 
 [features]
 tests_that_use_llms = []
diff --git a/paddler_client_python/tests/llm/__init__.py b/paddler_test_cluster_harness/fixtures/ca_bundle_without_certificates.pem
similarity index 100%
rename from paddler_client_python/tests/llm/__init__.py
rename to paddler_test_cluster_harness/fixtures/ca_bundle_without_certificates.pem
diff --git a/paddler_test_cluster_harness/src/agent_desired_model_at.rs b/paddler_test_cluster_harness/src/agent_desired_model_at.rs
new file mode 100644
index 000000000..057492d64
--- /dev/null
+++ b/paddler_test_cluster_harness/src/agent_desired_model_at.rs
@@ -0,0 +1,35 @@
+use std::path::Path;
+
+use paddler_messaging::agent_desired_model::AgentDesiredModel;
+use paddler_model_source::agent_local_model_path::AgentLocalModelPath;
+use paddler_model_source::model_source::ModelSource;
+use paddler_model_source::model_source_error::ModelSourceError;
+
+pub fn agent_desired_model_at(path: &Path) -> Result<AgentDesiredModel, ModelSourceError> {
+    Ok(
+        ModelSource::LocalToAgent(AgentLocalModelPath::from_absolute_path(path)?)
+            .into_agent_desired_model(),
+    )
+}
+
+#[cfg(test)]
+mod tests {
+    use std::path::Path;
+
+    use paddler_model_source::model_source_error::ModelSourceError;
+
+    use super::agent_desired_model_at;
+
+    #[test]
+    fn a_model_at_a_relative_path_cannot_be_desired() {
+        assert_eq!(
+            agent_desired_model_at(Path::new("model.gguf"))
+                .unwrap_err()
+                .to_string(),
+            ModelSourceError::LocalModelPathNotAbsolute {
+                path: "model.gguf".into(),
+            }
+            .to_string()
+        );
+    }
+}
diff --git a/paddler_test_cluster_harness/src/agent_readiness.rs b/paddler_test_cluster_harness/src/agent_readiness.rs
new file mode 100644
index 000000000..b3201e1da
--- /dev/null
+++ b/paddler_test_cluster_harness/src/agent_readiness.rs
@@ -0,0 +1,42 @@
+use paddler_messaging::agent_controller_snapshot::AgentControllerSnapshot;
+
+use crate::cluster_harness_error::ClusterHarnessError;
+
+pub enum AgentReadiness {
+    Registered,
+    SlotsReady(u16),
+}
+
+impl AgentReadiness {
+    #[must_use]
+    pub const fn of_registration(slot_count: u16, wait_for_slots_ready: bool) -> Self {
+        if wait_for_slots_ready {
+            Self::SlotsReady(slot_count)
+        } else {
+            Self::Registered
+        }
+    }
+
+    pub fn is_met_by(
+        &self,
+        agent_name: &str,
+        registered_agent: &AgentControllerSnapshot,
+    ) -> Result<bool, ClusterHarnessError> {
+        match self {
+            Self::Registered => Ok(true),
+            Self::SlotsReady(expected_slot_count)
+                if registered_agent.status.runtime.slots_total()
+                    == u64::from(*expected_slot_count) =>
+            {
+                Ok(true)
+            }
+            Self::SlotsReady(_) if !registered_agent.status.issues.is_empty() => {
+                Err(ClusterHarnessError::AgentReportedIssues {
+                    agent_name: agent_name.to_owned(),
+                    issues: registered_agent.status.issues.clone(),
+                })
+            }
+            Self::SlotsReady(_) => Ok(false),
+        }
+    }
+}
diff --git a/paddler_test_cluster_harness/src/ca_bundle_without_certificates.rs b/paddler_test_cluster_harness/src/ca_bundle_without_certificates.rs
new file mode 100644
index 000000000..3fc2fef79
--- /dev/null
+++ b/paddler_test_cluster_harness/src/ca_bundle_without_certificates.rs
@@ -0,0 +1,4 @@
+pub const CA_BUNDLE_WITHOUT_CERTIFICATES: &str = concat!(
+    env!("CARGO_MANIFEST_DIR"),
+    "/fixtures/ca_bundle_without_certificates.pem"
+);
diff --git a/paddler_test_cluster_harness/src/cluster.rs b/paddler_test_cluster_harness/src/cluster.rs
index 70974b1bf..a2c0daecd 100644
--- a/paddler_test_cluster_harness/src/cluster.rs
+++ b/paddler_test_cluster_harness/src/cluster.rs
@@ -3,6 +3,8 @@ use std::num::NonZeroUsize;
 
 use anyhow::Context as _;
 use anyhow::Result;
+use futures_util::TryFutureExt as _;
+use futures_util::future::ready;
 use serde_json::Value;
 use tokio_util::sync::CancellationToken;
 
@@ -12,23 +14,31 @@ use paddler_client::client_inference_params::ClientInferenceParams;
 use paddler_client::client_management::ClientManagement;
 use paddler_client::inference_message_stream::InferenceMessageStream;
 use paddler_client::reports_health::ReportsHealth as _;
+use paddler_messaging::agent_decisions_serving::AgentDecisionsServing;
 use paddler_messaging::agent_controller_pool_snapshot::AgentControllerPoolSnapshot;
+use paddler_messaging::agent_controller_snapshot::AgentControllerSnapshot;
 use paddler_messaging::agent_issue::AgentIssue;
-use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_messaging::agent_runtime_status::AgentRuntimeStatus;
 use paddler_messaging::buffered_request_manager_snapshot::BufferedRequestManagerSnapshot;
 use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
 use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters_schema::validated_parameters_schema::ValidatedParametersSchema;
 use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
+use paddler_messaging::request_params::decide_params::raw_decide_params::RawDecideParams;
 use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
 
 use crate::agent_config::AgentConfig;
 use crate::agent_count_is::agent_count_is;
+use crate::agent_readiness::AgentReadiness;
 use crate::agent_slots_processing_is::agent_slots_processing_is;
 use crate::agent_spawner::AgentSpawner;
 use crate::buffered_request_count_is::buffered_request_count_is;
+use crate::cluster_desired_state::ClusterDesiredState;
 use crate::cluster_harness_error::ClusterHarnessError;
+use crate::cluster_snapshots_watchers::ClusterSnapshotsWatchers;
+use crate::collect_decision_results::collect_decision_results;
 use crate::collect_embedding_results::collect_embedding_results;
 use crate::collect_generated_tokens::collect_generated_tokens;
+use crate::collected_decision_results::CollectedDecisionResults;
 use crate::collected_embedding_results::CollectedEmbeddingResults;
 use crate::collected_generated_tokens::CollectedGeneratedTokens;
 use crate::concurrent_connection_budget::ConcurrentConnectionBudget;
@@ -36,6 +46,8 @@ use crate::openai_api_client::OpenAIApiClient;
 use crate::running_agent::RunningAgent;
 use crate::running_balancer::RunningBalancer;
 use crate::snapshots_watcher::SnapshotsWatcher;
+use crate::typesafe_api_client::TypeSafeApiClient;
+use crate::typesafe_api_response::TypeSafeApiResponse;
 
 const INFERENCE_SOCKET_POOL_SIZE: NonZeroUsize = NonZeroUsize::MIN;
 
@@ -45,12 +57,10 @@ pub struct Cluster {
     pub agents_watcher: SnapshotsWatcher<AgentControllerPoolSnapshot>,
     pub balancer: RunningBalancer,
     pub buffered_requests_watcher: SnapshotsWatcher<BufferedRequestManagerSnapshot>,
-    pub client_compat_openai_health: ClientHealth,
     pub client_inference: ClientInference,
     pub client_management: ClientManagement,
     agent_spawner: Box<dyn AgentSpawner>,
     connection_budget: ConcurrentConnectionBudget,
-    openai_api_client: OpenAIApiClient,
 }
 
 impl Cluster {
@@ -58,61 +68,44 @@ impl Cluster {
         cancellation_token: CancellationToken,
         balancer: RunningBalancer,
         agent_spawner: Box<dyn AgentSpawner>,
-        desired_state: Option<&BalancerDesiredState>,
+        desired_state: &ClusterDesiredState,
     ) -> Result<Self> {
-        let management_base_url = balancer.management_base_url()?;
-        let inference_base_url = balancer.inference_base_url()?;
-        let openai_base_url = balancer.compat_openai_base_url()?;
-
-        let client_management = ClientManagement::new(management_base_url);
+        let client_management = ClientManagement::new(balancer.management_base_url.clone());
         let client_inference = ClientInference::new(ClientInferenceParams {
             inference_socket_pool_size: INFERENCE_SOCKET_POOL_SIZE,
-            url: inference_base_url,
+            url: balancer.inference_base_url.clone(),
         });
-        let client_compat_openai_health = ClientHealth::new(openai_base_url.clone());
 
         client_management
             .wait_until_healthy(cancellation_token.clone())
             .await
             .context("balancer did not become healthy")?;
 
-        if let Some(desired_state) = desired_state {
+        if let ClusterDesiredState::Apply(desired_state) = desired_state {
             client_management
                 .put_balancer_desired_state(cancellation_token.clone(), desired_state)
                 .await
                 .context("failed to PUT balancer desired state")?;
         }
 
-        let agents_watcher = SnapshotsWatcher::of_agents(
-            client_management
-                .get_agents_stream(cancellation_token.clone())
-                .await
-                .context("failed to open the agents stream")?,
-        )
-        .await?;
-        let buffered_requests_watcher = SnapshotsWatcher::of_buffered_requests(
-            client_management
-                .get_buffered_requests_stream(cancellation_token)
-                .await
-                .context("failed to open the buffered requests stream")?,
-        )
-        .await?;
-
-        let openai_api_client = OpenAIApiClient::new(&openai_base_url)?;
-
-        Ok(Self {
-            agent_ids: Vec::new(),
-            agents: Vec::new(),
-            agents_watcher,
-            balancer,
-            buffered_requests_watcher,
-            client_compat_openai_health,
-            client_inference,
-            client_management,
-            agent_spawner,
-            connection_budget: ConcurrentConnectionBudget::default(),
-            openai_api_client,
-        })
+        ClusterSnapshotsWatchers::open(&client_management, cancellation_token)
+            .await
+            .map(
+                |ClusterSnapshotsWatchers {
+                     agents_watcher,
+                     buffered_requests_watcher,
+                 }| Self {
+                    agent_ids: Vec::new(),
+                    agents: Vec::new(),
+                    agents_watcher,
+                    balancer,
+                    buffered_requests_watcher,
+                    client_inference,
+                    client_management,
+                    agent_spawner,
+                    connection_budget: ConcurrentConnectionBudget::default(),
+                },
+            )
     }
 
     pub fn continue_from_raw_prompt(
@@ -183,6 +176,24 @@ impl Cluster {
         })
     }
 
+    pub fn decide(
+        &self,
+        cancellation_token: CancellationToken,
+        params: &RawDecideParams,
+    ) -> impl Future<Output = Result<CollectedDecisionResults>> + Send + use<> {
+        let client_inference = self.client_inference.clone();
+        let params = params.clone();
+
+        self.connection_budget.hold_during(async move {
+            collect_decision_results(
+                client_inference
+                    .post_decide(cancellation_token, &params)
+                    .await?,
+            )
+            .await
+        })
+    }
+
     pub fn generate_embedding_batch(
         &self,
         cancellation_token: CancellationToken,
@@ -201,64 +212,117 @@ impl Cluster {
         })
     }
 
+    pub fn compat_openai_health_client(&self) -> Result<ClientHealth, ClusterHarnessError> {
+        self.balancer
+            .compat_openai_base_url()
+            .map(ClientHealth::new)
+    }
+
+    pub fn compat_typesafe_health_client(&self) -> Result<ClientHealth, ClusterHarnessError> {
+        self.balancer
+            .compat_typesafe_base_url()
+            .map(ClientHealth::new)
+    }
+
     pub fn openai_chat_completion_streaming(
         &self,
         body: &Value,
     ) -> impl Future<Output = Result<Vec<Value>, ClusterHarnessError>> + Send + use<> {
-        let openai_api_client = self.openai_api_client.clone();
         let body = body.clone();
 
-        self.connection_budget.hold_during(async move {
-            openai_api_client
-                .chat_completion_streaming(&body)
-                .await
-                .map_err(ClusterHarnessError::OpenAIRequestFailed)
-        })
+        self.connection_budget
+            .hold_during(
+                ready(self.openai_api_client()).and_then(|openai_api_client| async move {
+                    openai_api_client
+                        .chat_completion_streaming(&body)
+                        .await
+                        .map_err(ClusterHarnessError::OpenAIRequestFailed)
+                }),
+            )
     }
 
     pub fn openai_chat_completion_non_streaming(
         &self,
         body: &Value,
     ) -> impl Future<Output = Result<Value, ClusterHarnessError>> + Send + use<> {
-        let openai_api_client = self.openai_api_client.clone();
         let body = body.clone();
 
-        self.connection_budget.hold_during(async move {
-            openai_api_client
-                .chat_completion_non_streaming(&body)
-                .await
-                .map_err(ClusterHarnessError::OpenAIRequestFailed)
-        })
+        self.connection_budget
+            .hold_during(
+                ready(self.openai_api_client()).and_then(|openai_api_client| async move {
+                    openai_api_client
+                        .chat_completion_non_streaming(&body)
+                        .await
+                        .map_err(ClusterHarnessError::OpenAIRequestFailed)
+                }),
+            )
     }
 
     pub fn openai_responses_streaming(
         &self,
         body: &Value,
     ) -> impl Future<Output = Result<Vec<Value>, ClusterHarnessError>> + Send + use<> {
-        let openai_api_client = self.openai_api_client.clone();
         let body = body.clone();
 
-        self.connection_budget.hold_during(async move {
-            openai_api_client
-                .responses_streaming(&body)
-                .await
-                .map_err(ClusterHarnessError::OpenAIRequestFailed)
-        })
+        self.connection_budget
+            .hold_during(
+                ready(self.openai_api_client()).and_then(|openai_api_client| async move {
+                    openai_api_client
+                        .responses_streaming(&body)
+                        .await
+                        .map_err(ClusterHarnessError::OpenAIRequestFailed)
+                }),
+            )
     }
 
     pub fn openai_responses_non_streaming(
         &self,
         body: &Value,
     ) -> impl Future<Output = Result<Value, ClusterHarnessError>> + Send + use<> {
-        let openai_api_client = self.openai_api_client.clone();
         let body = body.clone();
 
-        self.connection_budget.hold_during(async move {
-            openai_api_client
-                .responses_non_streaming(&body)
-                .await
-                .map_err(ClusterHarnessError::OpenAIRequestFailed)
-        })
+        self.connection_budget
+            .hold_during(
+                ready(self.openai_api_client()).and_then(|openai_api_client| async move {
+                    openai_api_client
+                        .responses_non_streaming(&body)
+                        .await
+                        .map_err(ClusterHarnessError::OpenAIRequestFailed)
+                }),
+            )
+    }
+
+    pub fn typesafe_models(
+        &self,
+    ) -> impl Future<Output = Result<TypeSafeApiResponse, ClusterHarnessError>> + Send + use<> {
+        self.connection_budget
+            .hold_during(ready(self.typesafe_api_client()).and_then(
+                |typesafe_api_client| async move {
+                    typesafe_api_client
+                        .models()
+                        .await
+                        .map_err(ClusterHarnessError::TypeSafeRequestFailed)
+                },
+            ))
+    }
+
+    pub fn typesafe_system_one(
+        &self,
+        request_id: &str,
+        body: &Value,
+    ) -> impl Future<Output = Result<TypeSafeApiResponse, ClusterHarnessError>> + Send + use<> {
+        let request_id = request_id.to_owned();
+        let body = body.clone();
+
+        self.connection_budget
+            .hold_during(ready(self.typesafe_api_client()).and_then(
+                |typesafe_api_client| async move {
+                    typesafe_api_client
+                        .system_one(&request_id, &body)
+                        .await
+                        .map_err(ClusterHarnessError::TypeSafeRequestFailed)
+                },
+            ))
     }
 
     pub async fn wait_for_agent_count(
@@ -270,16 +334,6 @@ impl Cluster {
             .await
     }
 
-    pub async fn wait_for_agent_ready(
-        &mut self,
-        agent_name: &str,
-        expected_slot_count: u16,
-    ) -> Result<AgentControllerPoolSnapshot, ClusterHarnessError> {
-        self.agents_watcher
-            .wait_for_agent_ready(agent_name, expected_slot_count)
-            .await
-    }
-
     pub async fn wait_for_slots_processing(
         &mut self,
         agent_id: &str,
@@ -319,24 +373,51 @@ impl Cluster {
     pub async fn wait_for_first_agent_issue<TIssueMatcher>(
         &mut self,
         issue_matcher: TIssueMatcher,
-    ) -> Result<AgentControllerPoolSnapshot>
+    ) -> Result<AgentControllerPoolSnapshot, ClusterHarnessError>
     where
         TIssueMatcher: Fn(&AgentIssue) -> bool,
+    {
+        self.wait_for_first_agent(|agent| agent.status.issues.iter().any(&issue_matcher))
+            .await
+    }
+
+    pub async fn wait_for_first_agent_to_serve(
+        &mut self,
+        decisions: AgentDecisionsServing,
+    ) -> Result<AgentControllerPoolSnapshot, ClusterHarnessError> {
+        self.wait_for_first_agent(|agent| {
+            matches!(
+                agent.status.runtime,
+                AgentRuntimeStatus::Serving {
+                    decisions: served_decisions,
+                    ..
+                } if served_decisions == decisions
+            )
+        })
+        .await
+    }
+
+    async fn wait_for_first_agent<TAgentMatcher>(
+        &mut self,
+        agent_matcher: TAgentMatcher,
+    ) -> Result<AgentControllerPoolSnapshot, ClusterHarnessError>
+    where
+        TAgentMatcher: Fn(&AgentControllerSnapshot) -> bool,
     {
         let agent_id = self
             .agent_ids
             .first()
-            .context("the cluster must have a registered agent")?
+            .ok_or(ClusterHarnessError::ClusterHasNoRegisteredAgent)?
             .clone();
 
-        Ok(self
-            .agents_watcher
+        self.agents_watcher
             .until_agent(&agent_id, |snapshot| {
-                snapshot.agents.iter().any(|agent| {
-                    agent.id == agent_id && agent.status.issues.iter().any(&issue_matcher)
-                })
+                snapshot
+                    .agents
+                    .iter()
+                    .any(|agent| agent.id == agent_id && agent_matcher(agent))
             })
-            .await?)
+            .await
     }
 
     pub async fn register_agents(
@@ -344,33 +425,22 @@ impl Cluster {
         agents: &[AgentConfig],
         wait_for_slots_ready: bool,
     ) -> Result<()> {
-        let mut last_ready_snapshot = None;
-
-        for agent in agents {
-            self.spawn_additional_agent(agent)?;
-
-            if wait_for_slots_ready {
-                last_ready_snapshot = Some(
-                    self.wait_for_agent_ready(&agent.name, agent.slot_count)
-                        .await?,
-                );
-            }
+        for agent_config in agents {
+            let mut running_agent = RunningAgent::new(
+                agent_config.clone(),
+                self.agent_spawner.spawn(agent_config)?,
+            );
+            let ready_agent = running_agent
+                .wait_until_ready(
+                    &mut self.agents_watcher,
+                    AgentReadiness::of_registration(agent_config.slot_count, wait_for_slots_ready),
+                )
+                .await?;
+
+            self.agent_ids.push(ready_agent.id);
+            self.agents.push(running_agent);
         }
 
-        let registered_snapshot = match last_ready_snapshot {
-            Some(snapshot) => snapshot,
-            None => self
-                .wait_for_agent_count(agents.len())
-                .await
-                .context("not all agents registered")?,
-        };
-
-        self.agent_ids = registered_snapshot
-            .agents
-            .iter()
-            .map(|registered_agent| registered_agent.id.clone())
-            .collect();
-
         Ok(())
     }
 
@@ -389,4 +459,16 @@ impl Cluster {
 
         self.balancer.shutdown().await
     }
+
+    fn openai_api_client(&self) -> Result<OpenAIApiClient, ClusterHarnessError> {
+        self.balancer
+            .compat_openai_base_url()
+            .map(OpenAIApiClient::new)
+    }
+
+    fn typesafe_api_client(&self) -> Result<TypeSafeApiClient, ClusterHarnessError> {
+        self.balancer
+            .compat_typesafe_base_url()
+            .map(TypeSafeApiClient::new)
+    }
 }
diff --git a/paddler_test_cluster_harness/src/cluster_desired_state.rs b/paddler_test_cluster_harness/src/cluster_desired_state.rs
new file mode 100644
index 000000000..934d3afa2
--- /dev/null
+++ b/paddler_test_cluster_harness/src/cluster_desired_state.rs
@@ -0,0 +1,6 @@
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+
+pub enum ClusterDesiredState {
+    Apply(Box<BalancerDesiredState>),
+    KeepStored,
+}
diff --git a/paddler_test_cluster_harness/src/cluster_harness_error.rs b/paddler_test_cluster_harness/src/cluster_harness_error.rs
index 594a0e30f..6e623d142 100644
--- a/paddler_test_cluster_harness/src/cluster_harness_error.rs
+++ b/paddler_test_cluster_harness/src/cluster_harness_error.rs
@@ -1,14 +1,15 @@
 use std::collections::BTreeSet;
 use std::io::Error as IoError;
 use std::net::SocketAddr;
-use std::str::Utf8Error;
 
 use async_openai::error::OpenAIError;
 use quick_xml::Error as XmlError;
 use quick_xml::events::attributes::AttrError;
+use reqwest::Error as ReqwestError;
 use serde_json::Error as SerdeJsonError;
 use thiserror::Error;
 use tokio_tungstenite::tungstenite::Error as WebSocketError;
+use tokio_tungstenite::tungstenite::protocol::CloseFrame;
 use url::ParseError;
 
 use paddler_client::error::Error as ClientError;
@@ -17,51 +18,75 @@ use paddler_messaging::generated_token_result::GeneratedTokenResult;
 use paddler_messaging::inference_client::message::Message as InferenceMessage;
 use paddler_messaging::jsonrpc::error::Error as JsonRpcError;
 
+use crate::process_end::ProcessEnd;
 use crate::snapshots_stream::SnapshotsStream;
 
 #[derive(Debug, Error)]
 pub enum ClusterHarnessError {
+    #[error("The decision stream ended before it reported its outcome")]
+    DecisionStreamEndedWithoutTerminalResult,
+    #[error("The decision stream yielded an error")]
+    DecisionStreamFailed(#[source] ClientError),
+    #[error("The decision stream carried a message that is not a decision result: {message:?}")]
+    DecisionStreamMessageUnexpected { message: Box<InferenceMessage> },
     #[error("The embedding stream carried a message that is not an embedding result: {message:?}")]
     EmbeddingStreamMessageUnexpected { message: Box<InferenceMessage> },
+    #[error("The embedding stream yielded an error")]
+    EmbeddingStreamFailed(#[source] ClientError),
     #[error("The token generation stream returned JSON-RPC error code {} ({})", error.code, error.description)]
     TokenStreamReturnedError { error: JsonRpcError },
+    #[error("The token generation stream yielded an error")]
+    TokenStreamFailed(#[source] ClientError),
     #[error(
         "The token generation stream carried a message that is not a generated token: {message:?}"
     )]
     TokenStreamMessageUnexpected { message: Box<InferenceMessage> },
+    #[error("The agent hung up on the untrusted balancer socket before it responded")]
+    AgentHungUpBeforeResponding,
+    #[error("The agent hung up on the untrusted balancer socket before it served its model")]
+    AgentHungUpBeforeServing,
     #[error("Agent {agent_id} left the balancer's pool while it was being observed")]
     AgentLeftThePool { agent_id: String },
-    #[error("The raw agent socket could not serialize a notification")]
-    AgentSocketNotificationUnserializable(#[source] SerdeJsonError),
-    #[error("The raw agent socket ended without a close frame")]
+    #[error(
+        "The balancer closed the untrusted agent socket before sending it text: {close_frame:?}"
+    )]
+    AgentSocketClosedBeforeText { close_frame: Option<CloseFrame> },
+    #[error("The untrusted agent socket client ended without a close frame")]
     AgentSocketEndedWithoutClosing,
-    #[error("The raw agent socket ended before the balancer sent a message")]
-    AgentSocketEndedWithoutMessage,
-    #[error("The raw agent socket could not deserialize a balancer message")]
-    AgentSocketMessageUndeserializable(#[source] SerdeJsonError),
-    #[error("The raw agent socket could not receive a frame")]
+    #[error("The untrusted agent socket client ended before the balancer sent it text")]
+    AgentSocketEndedWithoutText,
+    #[error("The untrusted agent socket client could not receive a frame")]
     AgentSocketReceiveFailed(#[source] WebSocketError),
-    #[error("The raw agent socket could not serialize a response")]
-    AgentSocketResponseUnserializable(#[source] SerdeJsonError),
-    #[error("The raw agent socket could not send a frame")]
+    #[error("The untrusted agent socket client could not send a frame")]
     AgentSocketSendFailed(#[source] WebSocketError),
+    #[error("Agent {agent_name:?} exited before it was ready: {process_end:?}")]
+    AgentExitedBeforeReady {
+        agent_name: String,
+        process_end: ProcessEnd,
+    },
     #[error("Agent {agent_name:?} reported issues while starting: {issues:?}")]
     AgentReportedIssues {
         agent_name: String,
         issues: BTreeSet<AgentIssue>,
     },
-    #[error("Unable to build a base URL for {addr}")]
-    BaseUrlInvalid {
-        addr: SocketAddr,
-        #[source]
-        source: ParseError,
-    },
+    #[error("The untrusted balancer socket could not accept an agent connection")]
+    BalancerSocketAcceptFailed(#[source] IoError),
+    #[error("The untrusted balancer socket could not complete the agent's WebSocket handshake")]
+    BalancerSocketHandshakeFailed(#[source] WebSocketError),
+    #[error("The untrusted balancer socket could not send a frame")]
+    BalancerSocketSendFailed(#[source] WebSocketError),
+    #[error("The cluster has no registered agent to wait for")]
+    ClusterHasNoRegisteredAgent,
     #[error("The balancer does not serve the OpenAI compatibility service")]
     CompatOpenAIServiceNotServed,
+    #[error("Unable to build the base URL of the OpenAI compatibility service")]
+    CompatOpenAIBaseUrlInvalid(#[source] ParseError),
+    #[error("The balancer does not serve the TypeSafe compatibility service")]
+    CompatTypeSafeServiceNotServed,
+    #[error("Unable to build the base URL of the typesafe compatibility service")]
+    CompatTypeSafeBaseUrlInvalid(#[source] ParseError),
     #[error("The web admin panel page carries a malformed attribute")]
     DashboardAttributeMalformed(#[source] AttrError),
-    #[error("The web admin panel page carries an attribute whose name is not UTF-8")]
-    DashboardAttributeNameNotUtf8(#[source] Utf8Error),
     #[error("The web admin panel page carries an attribute value that cannot be unescaped")]
     DashboardAttributeValueUnreadable(#[source] XmlError),
     #[error("The web admin panel page has no dashboard element")]
@@ -87,6 +112,8 @@ pub enum ClusterHarnessError {
     },
     #[error("The inference request failed")]
     InferenceRequestFailed(#[source] ClientError),
+    #[error("Unable to build the base URL of the inference service")]
+    InferenceBaseUrlInvalid(#[source] ParseError),
     #[error("The raw inference socket closed before answering request {request_id}")]
     InferenceSocketClosedBeforeAnswer { request_id: String },
     #[error("The raw inference socket closed before answering any request")]
@@ -101,6 +128,8 @@ pub enum ClusterHarnessError {
     InferenceSocketRequestUnserializable(#[source] SerdeJsonError),
     #[error("The raw inference socket could not send a frame")]
     InferenceSocketSendFailed(#[source] WebSocketError),
+    #[error("Unable to build the base URL of the management service")]
+    ManagementBaseUrlInvalid(#[source] ParseError),
     #[error("The OpenAI compatibility request failed")]
     OpenAIRequestFailed(#[source] OpenAIError),
     #[error("Unable to list the open file descriptors of the current process")]
@@ -113,4 +142,6 @@ pub enum ClusterHarnessError {
         #[source]
         source: ClientError,
     },
+    #[error("The TypeSafe compatibility request failed")]
+    TypeSafeRequestFailed(#[source] ReqwestError),
 }
diff --git a/paddler_test_cluster_harness/src/cluster_params.rs b/paddler_test_cluster_harness/src/cluster_params.rs
index 5fd64f1e5..b9c611aab 100644
--- a/paddler_test_cluster_harness/src/cluster_params.rs
+++ b/paddler_test_cluster_harness/src/cluster_params.rs
@@ -1,14 +1,13 @@
 use std::time::Duration;
 
-use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-
 use crate::agent_config::AgentConfig;
+use crate::cluster_desired_state::ClusterDesiredState;
 use crate::longer_than_any_test_run::LONGER_THAN_ANY_TEST_RUN;
 
 pub struct ClusterParams {
     pub agents: Vec<AgentConfig>,
     pub buffered_request_timeout: Duration,
-    pub desired_state: Option<BalancerDesiredState>,
+    pub desired_state: ClusterDesiredState,
     pub inference_cors_allowed_hosts: Vec<String>,
     pub inference_item_timeout: Duration,
     pub management_cors_allowed_hosts: Vec<String>,
@@ -17,12 +16,23 @@ pub struct ClusterParams {
     pub wait_for_slots_ready: bool,
 }
 
+impl ClusterParams {
+    #[must_use]
+    pub fn without_agents() -> Self {
+        Self {
+            agents: Vec::new(),
+            wait_for_slots_ready: false,
+            ..Self::default()
+        }
+    }
+}
+
 impl Default for ClusterParams {
     fn default() -> Self {
         Self {
             agents: AgentConfig::uniform(1, 4),
             buffered_request_timeout: LONGER_THAN_ANY_TEST_RUN,
-            desired_state: Some(BalancerDesiredState::default()),
+            desired_state: ClusterDesiredState::Apply(Box::default()),
             inference_cors_allowed_hosts: Vec::new(),
             inference_item_timeout: LONGER_THAN_ANY_TEST_RUN,
             management_cors_allowed_hosts: Vec::new(),
diff --git a/paddler_test_cluster_harness/src/cluster_snapshots_watchers.rs b/paddler_test_cluster_harness/src/cluster_snapshots_watchers.rs
new file mode 100644
index 000000000..b56c77b64
--- /dev/null
+++ b/paddler_test_cluster_harness/src/cluster_snapshots_watchers.rs
@@ -0,0 +1,44 @@
+use anyhow::Context as _;
+use anyhow::Result;
+use futures_util::FutureExt as _;
+use futures_util::TryFutureExt as _;
+use tokio_util::sync::CancellationToken;
+
+use paddler_client::client_management::ClientManagement;
+use paddler_messaging::agent_controller_pool_snapshot::AgentControllerPoolSnapshot;
+use paddler_messaging::buffered_request_manager_snapshot::BufferedRequestManagerSnapshot;
+
+use crate::snapshots_watcher::SnapshotsWatcher;
+
+pub struct ClusterSnapshotsWatchers {
+    pub agents_watcher: SnapshotsWatcher<AgentControllerPoolSnapshot>,
+    pub buffered_requests_watcher: SnapshotsWatcher<BufferedRequestManagerSnapshot>,
+}
+
+impl ClusterSnapshotsWatchers {
+    pub async fn open(
+        client_management: &ClientManagement,
+        cancellation_token: CancellationToken,
+    ) -> Result<Self> {
+        client_management
+            .get_agents_stream(cancellation_token.clone())
+            .map(|agents_stream| agents_stream.context("failed to open the agents stream"))
+            .and_then(|agents_stream| SnapshotsWatcher::of_agents(agents_stream).err_into())
+            .and_then(|agents_watcher| {
+                client_management
+                    .get_buffered_requests_stream(cancellation_token)
+                    .map(|buffered_requests_stream| {
+                        buffered_requests_stream
+                            .context("failed to open the buffered requests stream")
+                    })
+                    .and_then(|buffered_requests_stream| {
+                        SnapshotsWatcher::of_buffered_requests(buffered_requests_stream).err_into()
+                    })
+                    .map_ok(|buffered_requests_watcher| Self {
+                        agents_watcher,
+                        buffered_requests_watcher,
+                    })
+            })
+            .await
+    }
+}
diff --git a/paddler_test_cluster_harness/src/collect_decision_results.rs b/paddler_test_cluster_harness/src/collect_decision_results.rs
new file mode 100644
index 000000000..bb976b1dd
--- /dev/null
+++ b/paddler_test_cluster_harness/src/collect_decision_results.rs
@@ -0,0 +1,118 @@
+use anyhow::Result;
+use futures_util::StreamExt as _;
+
+use paddler_client::inference_message_stream::InferenceMessageStream;
+use paddler_messaging::decision_result::DecisionResult;
+use paddler_messaging::inference_client::message::Message as InferenceMessage;
+use paddler_messaging::inference_client::response::Response as InferenceResponse;
+use paddler_messaging::jsonrpc::response_envelope::ResponseEnvelope;
+
+use crate::cluster_harness_error::ClusterHarnessError;
+use crate::collected_decision_results::CollectedDecisionResults;
+
+pub async fn collect_decision_results(
+    mut stream: InferenceMessageStream,
+) -> Result<CollectedDecisionResults> {
+    let mut answers = Vec::new();
+
+    while let Some(item) = stream.next().await {
+        match item.map_err(ClusterHarnessError::DecisionStreamFailed)? {
+            InferenceMessage::Response(ResponseEnvelope {
+                response: InferenceResponse::Decision(DecisionResult::QuestionAnswered(answer)),
+                ..
+            }) => answers.push(answer),
+            InferenceMessage::Response(ResponseEnvelope {
+                response: InferenceResponse::Decision(terminal_result),
+                ..
+            }) => {
+                return Ok(CollectedDecisionResults {
+                    answers,
+                    terminal_result,
+                });
+            }
+            unexpected_message => {
+                return Err(ClusterHarnessError::DecisionStreamMessageUnexpected {
+                    message: Box::new(unexpected_message),
+                }
+                .into());
+            }
+        }
+    }
+
+    Err(ClusterHarnessError::DecisionStreamEndedWithoutTerminalResult.into())
+}
+
+#[cfg(test)]
+mod tests {
+    use std::mem::discriminant;
+
+    use futures_util::stream::iter;
+
+    use paddler_client::error::Error as ClientError;
+    use paddler_messaging::embedding_result::EmbeddingResult;
+    use paddler_messaging::inference_client::message::Message as InferenceMessage;
+    use paddler_messaging::inference_client::response::Response as InferenceResponse;
+    use paddler_messaging::jsonrpc::response_envelope::ResponseEnvelope;
+
+    use super::collect_decision_results;
+    use crate::cluster_harness_error::ClusterHarnessError;
+
+    #[tokio::test]
+    async fn a_decision_stream_ending_without_its_outcome_is_an_error() {
+        let collection_error = collect_decision_results(Box::pin(iter(Vec::new())))
+            .await
+            .expect_err("a stream without an outcome must not be collected");
+
+        assert_eq!(
+            collection_error
+                .downcast_ref::<ClusterHarnessError>()
+                .map(discriminant),
+            Some(discriminant(
+                &ClusterHarnessError::DecisionStreamEndedWithoutTerminalResult
+            ))
+        );
+    }
+
+    #[tokio::test]
+    async fn a_decision_stream_carrying_another_result_is_an_error() {
+        let embedding_done = || {
+            InferenceMessage::Response(ResponseEnvelope {
+                generated_by: None,
+                request_id: "request-1".to_owned(),
+                response: InferenceResponse::Embedding(EmbeddingResult::Done),
+            })
+        };
+        let collection_error = collect_decision_results(Box::pin(iter([Ok(embedding_done())])))
+            .await
+            .expect_err("a stream carrying an embedding result must not be collected");
+
+        assert_eq!(
+            collection_error
+                .downcast_ref::<ClusterHarnessError>()
+                .map(ToString::to_string),
+            Some(
+                ClusterHarnessError::DecisionStreamMessageUnexpected {
+                    message: Box::new(embedding_done()),
+                }
+                .to_string()
+            )
+        );
+    }
+
+    #[tokio::test]
+    async fn a_decision_stream_yielding_an_error_is_an_error() {
+        let collection_error =
+            collect_decision_results(Box::pin(iter([Err(ClientError::ConnectionDropped {
+                request_id: "request-1".to_owned(),
+            })])))
+            .await
+            .expect_err("a stream yielding an error must not be collected");
+
+        assert!(matches!(
+            collection_error.downcast_ref::<ClusterHarnessError>(),
+            Some(ClusterHarnessError::DecisionStreamFailed(ClientError::ConnectionDropped {
+                request_id
+            })) if request_id == "request-1"
+        ));
+    }
+}
diff --git a/paddler_test_cluster_harness/src/collect_embedding_results.rs b/paddler_test_cluster_harness/src/collect_embedding_results.rs
index ccd8f991c..19e6db2d5 100644
--- a/paddler_test_cluster_harness/src/collect_embedding_results.rs
+++ b/paddler_test_cluster_harness/src/collect_embedding_results.rs
@@ -1,6 +1,5 @@
 use std::collections::BTreeSet;
 
-use anyhow::Context as _;
 use anyhow::Result;
 use futures_util::StreamExt as _;
 
@@ -19,18 +18,17 @@ pub async fn collect_embedding_results(
     mut stream: InferenceMessageStream,
 ) -> Result<CollectedEmbeddingResults> {
     let mut embeddings: Vec<EmbeddingWithProducer> = Vec::new();
-    let mut embeddings_disabled = false;
-    let mut errors: Vec<String> = Vec::new();
-    let mut embedding_rejected_due_to_active_token_generation_count: usize = 0;
-    let mut model_not_loaded_count: usize = 0;
+    let mut failures: Vec<EmbeddingResult> = Vec::new();
     let mut no_embeddings_produced_count: usize = 0;
+    let mut documents_exceeding_sequence_context = Vec::new();
+    let mut documents_tokenized_to_nothing = Vec::new();
     let mut oversized_documents = Vec::new();
     let mut request_ids = BTreeSet::new();
     let mut saw_done = false;
     let mut wire_errors = Vec::new();
 
     while let Some(item) = stream.next().await {
-        match item.context("embedding stream yielded an error")? {
+        match item.map_err(ClusterHarnessError::EmbeddingStreamFailed)? {
             InferenceMessage::Response(ResponseEnvelope {
                 generated_by,
                 request_id,
@@ -53,17 +51,21 @@ pub async fn collect_embedding_results(
                     EmbeddingResult::DocumentExceedsBatchSize(details) => {
                         oversized_documents.push(details);
                     }
-                    EmbeddingResult::EmbeddingsDisabled => {
-                        embeddings_disabled = true;
+                    EmbeddingResult::DocumentExceedsSequenceContext(details) => {
+                        documents_exceeding_sequence_context.push(details);
                     }
-                    EmbeddingResult::Error(message) => {
-                        errors.push(message);
+                    EmbeddingResult::DocumentTokenizedToNothing(details) => {
+                        documents_tokenized_to_nothing.push(details);
                     }
-                    EmbeddingResult::EmbeddingRejectedDueToActiveTokenGeneration => {
-                        embedding_rejected_due_to_active_token_generation_count += 1;
-                    }
-                    EmbeddingResult::ModelNotLoaded(_) => {
-                        model_not_loaded_count += 1;
+                    failure @ (EmbeddingResult::AgentRuntimeFailed(_)
+                    | EmbeddingResult::BatchAssemblyFailed(_)
+                    | EmbeddingResult::DecodeFailed(_)
+                    | EmbeddingResult::EmbeddingTooLongForRmsNormalization(_)
+                    | EmbeddingResult::EmbeddingsUnavailable(_)
+                    | EmbeddingResult::InputTokenizationFailed(_)
+                    | EmbeddingResult::ModelNotLoaded(_)
+                    | EmbeddingResult::SchedulerUnavailable(_)) => {
+                        failures.push(failure);
                     }
                     EmbeddingResult::NoEmbeddingsProduced => {
                         no_embeddings_produced_count += 1;
@@ -74,7 +76,7 @@ pub async fn collect_embedding_results(
                 request_ids.insert(request_id);
                 wire_errors.push(error);
             }
-            unexpected_message => {
+            unexpected_message @ InferenceMessage::Response(_) => {
                 return Err(ClusterHarnessError::EmbeddingStreamMessageUnexpected {
                     message: Box::new(unexpected_message),
                 }
@@ -84,11 +86,10 @@ pub async fn collect_embedding_results(
     }
 
     Ok(CollectedEmbeddingResults {
+        documents_exceeding_sequence_context,
+        documents_tokenized_to_nothing,
         embeddings,
-        embeddings_disabled,
-        errors,
-        embedding_rejected_due_to_active_token_generation_count,
-        model_not_loaded_count,
+        failures,
         no_embeddings_produced_count,
         oversized_documents,
         request_ids,
@@ -96,3 +97,65 @@ pub async fn collect_embedding_results(
         wire_errors,
     })
 }
+
+#[cfg(test)]
+mod tests {
+    use futures_util::stream::iter;
+
+    use paddler_client::error::Error as ClientError;
+    use paddler_messaging::generated_token_result::GeneratedTokenResult;
+    use paddler_messaging::inference_client::message::Message as InferenceMessage;
+    use paddler_messaging::inference_client::response::Response as InferenceResponse;
+    use paddler_messaging::jsonrpc::response_envelope::ResponseEnvelope;
+
+    use super::collect_embedding_results;
+    use crate::cluster_harness_error::ClusterHarnessError;
+
+    #[tokio::test]
+    async fn refuses_a_message_that_is_not_an_embedding_result() {
+        let generated_content_token = || {
+            InferenceMessage::Response(ResponseEnvelope {
+                generated_by: None,
+                request_id: "generation-request".to_owned(),
+                response: InferenceResponse::GeneratedToken(GeneratedTokenResult::ContentToken(
+                    "piece".to_owned(),
+                )),
+            })
+        };
+        let collection_error =
+            collect_embedding_results(Box::pin(iter([Ok(generated_content_token())])))
+                .await
+                .err()
+                .expect("a stream of another kind of message must be refused");
+
+        assert_eq!(
+            collection_error
+                .downcast_ref::<ClusterHarnessError>()
+                .map(ToString::to_string),
+            Some(
+                ClusterHarnessError::EmbeddingStreamMessageUnexpected {
+                    message: Box::new(generated_content_token()),
+                }
+                .to_string()
+            )
+        );
+    }
+
+    #[tokio::test]
+    async fn an_embedding_stream_yielding_an_error_is_an_error() {
+        let collection_error =
+            collect_embedding_results(Box::pin(iter([Err(ClientError::ConnectionDropped {
+                request_id: "request-1".to_owned(),
+            })])))
+            .await
+            .err()
+            .expect("a stream yielding an error must not be collected");
+
+        assert!(matches!(
+            collection_error.downcast_ref::<ClusterHarnessError>(),
+            Some(ClusterHarnessError::EmbeddingStreamFailed(ClientError::ConnectionDropped {
+                request_id
+            })) if request_id == "request-1"
+        ));
+    }
+}
diff --git a/paddler_test_cluster_harness/src/collect_generated_tokens.rs b/paddler_test_cluster_harness/src/collect_generated_tokens.rs
index dbc23f3f6..c5b6b1e56 100644
--- a/paddler_test_cluster_harness/src/collect_generated_tokens.rs
+++ b/paddler_test_cluster_harness/src/collect_generated_tokens.rs
@@ -1,4 +1,3 @@
-use anyhow::Context as _;
 use anyhow::Result;
 use futures_util::StreamExt as _;
 
@@ -6,6 +5,7 @@ use paddler_client::inference_message_stream::InferenceMessageStream;
 use paddler_messaging::inference_client::message::Message as InferenceMessage;
 use paddler_messaging::inference_client::response::Response as InferenceResponse;
 use paddler_messaging::jsonrpc::response_envelope::ResponseEnvelope;
+use paddler_messaging::result_step::ResultStep;
 use paddler_messaging::streamable_result::StreamableResult as _;
 
 use crate::cluster_harness_error::ClusterHarnessError;
@@ -19,7 +19,7 @@ pub async fn collect_generated_tokens(
     let mut token_results: Vec<TokenResultWithProducer> = Vec::new();
 
     while let Some(item) = stream.next().await {
-        match item.context("inference stream yielded an error")? {
+        match item.map_err(ClusterHarnessError::TokenStreamFailed)? {
             InferenceMessage::Response(ResponseEnvelope {
                 generated_by,
                 response: InferenceResponse::GeneratedToken(token_result),
@@ -29,15 +29,21 @@ pub async fn collect_generated_tokens(
                     text.push_str(token_text);
                 }
 
-                let is_done = token_result.is_done();
+                match token_result.into_step() {
+                    ResultStep::Final(final_result) => {
+                        token_results.push(TokenResultWithProducer {
+                            token_result: final_result.into(),
+                            generated_by,
+                        });
 
-                token_results.push(TokenResultWithProducer {
-                    token_result,
-                    generated_by,
-                });
-
-                if is_done {
-                    break;
+                        break;
+                    }
+                    ResultStep::Partial(partial_result) => {
+                        token_results.push(TokenResultWithProducer {
+                            token_result: partial_result.into(),
+                            generated_by,
+                        });
+                    }
                 }
             }
             InferenceMessage::Error(error_envelope) => {
@@ -46,7 +52,7 @@ pub async fn collect_generated_tokens(
                 }
                 .into());
             }
-            unexpected_message => {
+            unexpected_message @ InferenceMessage::Response(_) => {
                 return Err(ClusterHarnessError::TokenStreamMessageUnexpected {
                     message: Box::new(unexpected_message),
                 }
@@ -60,3 +66,62 @@ pub async fn collect_generated_tokens(
         token_results,
     })
 }
+
+#[cfg(test)]
+mod tests {
+    use futures_util::stream::iter;
+
+    use paddler_client::error::Error as ClientError;
+    use paddler_messaging::embedding_result::EmbeddingResult;
+    use paddler_messaging::inference_client::message::Message as InferenceMessage;
+    use paddler_messaging::inference_client::response::Response as InferenceResponse;
+    use paddler_messaging::jsonrpc::response_envelope::ResponseEnvelope;
+
+    use super::collect_generated_tokens;
+    use crate::cluster_harness_error::ClusterHarnessError;
+
+    #[tokio::test]
+    async fn refuses_a_message_that_is_not_a_generated_token() {
+        let embedding_done = || {
+            InferenceMessage::Response(ResponseEnvelope {
+                generated_by: None,
+                request_id: "embedding-request".to_owned(),
+                response: InferenceResponse::Embedding(EmbeddingResult::Done),
+            })
+        };
+        let collection_error = collect_generated_tokens(Box::pin(iter([Ok(embedding_done())])))
+            .await
+            .err()
+            .expect("a stream of another kind of message must be refused");
+
+        assert_eq!(
+            collection_error
+                .downcast_ref::<ClusterHarnessError>()
+                .map(ToString::to_string),
+            Some(
+                ClusterHarnessError::TokenStreamMessageUnexpected {
+                    message: Box::new(embedding_done()),
+                }
+                .to_string()
+            )
+        );
+    }
+
+    #[tokio::test]
+    async fn a_token_stream_yielding_an_error_is_an_error() {
+        let collection_error =
+            collect_generated_tokens(Box::pin(iter([Err(ClientError::ConnectionDropped {
+                request_id: "request-1".to_owned(),
+            })])))
+            .await
+            .err()
+            .expect("a stream yielding an error must not be collected");
+
+        assert!(matches!(
+            collection_error.downcast_ref::<ClusterHarnessError>(),
+            Some(ClusterHarnessError::TokenStreamFailed(ClientError::ConnectionDropped {
+                request_id
+            })) if request_id == "request-1"
+        ));
+    }
+}
diff --git a/paddler_test_cluster_harness/src/collected_decision_results.rs b/paddler_test_cluster_harness/src/collected_decision_results.rs
new file mode 100644
index 000000000..d4f6d2c89
--- /dev/null
+++ b/paddler_test_cluster_harness/src/collected_decision_results.rs
@@ -0,0 +1,8 @@
+use paddler_messaging::decision_answer::DecisionAnswer;
+use paddler_messaging::decision_result::DecisionResult;
+
+#[derive(Debug)]
+pub struct CollectedDecisionResults {
+    pub answers: Vec<DecisionAnswer>,
+    pub terminal_result: DecisionResult,
+}
diff --git a/paddler_test_cluster_harness/src/collected_embedding_results.rs b/paddler_test_cluster_harness/src/collected_embedding_results.rs
index 35af897c1..5bb34563c 100644
--- a/paddler_test_cluster_harness/src/collected_embedding_results.rs
+++ b/paddler_test_cluster_harness/src/collected_embedding_results.rs
@@ -1,16 +1,18 @@
 use std::collections::BTreeSet;
 
+use paddler_messaging::embedding_document_exceeding_sequence_context_details::EmbeddingDocumentExceedingSequenceContextDetails;
+use paddler_messaging::embedding_document_tokenized_to_nothing_details::EmbeddingDocumentTokenizedToNothingDetails;
+use paddler_messaging::embedding_result::EmbeddingResult;
 use paddler_messaging::jsonrpc::error::Error as JsonRpcError;
 use paddler_messaging::oversized_embedding_document_details::OversizedEmbeddingDocumentDetails;
 
 use crate::embedding_with_producer::EmbeddingWithProducer;
 
 pub struct CollectedEmbeddingResults {
+    pub documents_exceeding_sequence_context: Vec<EmbeddingDocumentExceedingSequenceContextDetails>,
+    pub documents_tokenized_to_nothing: Vec<EmbeddingDocumentTokenizedToNothingDetails>,
     pub embeddings: Vec<EmbeddingWithProducer>,
-    pub embeddings_disabled: bool,
-    pub errors: Vec<String>,
-    pub embedding_rejected_due_to_active_token_generation_count: usize,
-    pub model_not_loaded_count: usize,
+    pub failures: Vec<EmbeddingResult>,
     pub no_embeddings_produced_count: usize,
     pub oversized_documents: Vec<OversizedEmbeddingDocumentDetails>,
     pub request_ids: BTreeSet<String>,
diff --git a/paddler_test_cluster_harness/src/collected_generated_tokens.rs b/paddler_test_cluster_harness/src/collected_generated_tokens.rs
index 76f4899d3..005f7f56b 100644
--- a/paddler_test_cluster_harness/src/collected_generated_tokens.rs
+++ b/paddler_test_cluster_harness/src/collected_generated_tokens.rs
@@ -18,6 +18,32 @@ impl CollectedGeneratedTokens {
             .collect()
     }
 
+    #[must_use]
+    pub fn content_text(&self) -> String {
+        self.token_results
+            .iter()
+            .filter_map(|token_result_with_producer| {
+                match &token_result_with_producer.token_result {
+                    GeneratedTokenResult::ContentToken(piece) => Some(piece.as_str()),
+                    _ => None,
+                }
+            })
+            .collect()
+    }
+
+    #[must_use]
+    pub fn reasoning_text(&self) -> String {
+        self.token_results
+            .iter()
+            .filter_map(|token_result_with_producer| {
+                match &token_result_with_producer.token_result {
+                    GeneratedTokenResult::ReasoningToken(piece) => Some(piece.as_str()),
+                    _ => None,
+                }
+            })
+            .collect()
+    }
+
     pub fn summary(&self) -> Result<GenerationSummary, ClusterHarnessError> {
         match self
             .token_results
@@ -50,11 +76,37 @@ mod tests {
             }],
         };
 
-        assert!(matches!(
-            collected.summary(),
-            Err(ClusterHarnessError::GenerationEndedWithoutSummary {
-                last_token_result: Some(GeneratedTokenResult::SamplerError(_)),
-            })
-        ));
+        assert_eq!(
+            collected.summary().err().map(|refusal| refusal.to_string()),
+            Some(
+                ClusterHarnessError::GenerationEndedWithoutSummary {
+                    last_token_result: Some(GeneratedTokenResult::SamplerError(
+                        "failed".to_owned()
+                    )),
+                }
+                .to_string()
+            )
+        );
+    }
+
+    #[test]
+    fn content_and_reasoning_text_keep_only_their_own_pieces() {
+        let produced = |token_result| TokenResultWithProducer {
+            token_result,
+            generated_by: None,
+        };
+        let collected = CollectedGeneratedTokens {
+            text: String::new(),
+            token_results: vec![
+                produced(GeneratedTokenResult::ReasoningToken("think".to_owned())),
+                produced(GeneratedTokenResult::ContentToken("say".to_owned())),
+                produced(GeneratedTokenResult::ReasoningToken("ing".to_owned())),
+            ],
+        };
+
+        assert_eq!(
+            [collected.reasoning_text(), collected.content_text()],
+            ["thinking".to_owned(), "say".to_owned()]
+        );
     }
 }
diff --git a/paddler_test_cluster_harness/src/concurrent_connection_budget.rs b/paddler_test_cluster_harness/src/concurrent_connection_budget.rs
index 3c5dc1d6c..2a29af113 100644
--- a/paddler_test_cluster_harness/src/concurrent_connection_budget.rs
+++ b/paddler_test_cluster_harness/src/concurrent_connection_budget.rs
@@ -58,7 +58,7 @@ impl ConcurrentConnectionBudget {
 
     pub fn reserve(&self) -> Result<ConnectionReservation, ClusterHarnessError> {
         self.held_connections
-            .fetch_update(Ordering::AcqRel, Ordering::Acquire, |held_connections| {
+            .try_update(Ordering::AcqRel, Ordering::Acquire, |held_connections| {
                 (held_connections < PORTABLE_LISTEN_BACKLOG).then_some(held_connections + 1)
             })
             .map_err(|held_connections| {
@@ -76,30 +76,42 @@ impl ConcurrentConnectionBudget {
 
 #[cfg(test)]
 mod tests {
+    use std::future::pending;
     use std::iter::repeat_with;
     use std::sync::atomic::Ordering;
 
+    use futures_util::FutureExt as _;
+
+    use paddler_client::error::Result as ClientResult;
+    use paddler_client::inference_message_stream::InferenceMessageStream;
+
     use super::ConcurrentConnectionBudget;
     use super::PORTABLE_LISTEN_BACKLOG;
     use crate::cluster_harness_error::ClusterHarnessError;
+    use crate::connection_reservation::ConnectionReservation;
+
+    fn hold_every_connection(budget: &ConcurrentConnectionBudget) -> Vec<ConnectionReservation> {
+        repeat_with(|| budget.reserve())
+            .take(PORTABLE_LISTEN_BACKLOG)
+            .collect::<Result<_, _>>()
+            .unwrap()
+    }
 
     #[test]
     fn refuses_a_connection_beyond_the_portable_listen_backlog() {
         let budget = ConcurrentConnectionBudget::default();
-        let _held_reservations: Vec<_> = repeat_with(|| budget.reserve())
-            .take(PORTABLE_LISTEN_BACKLOG)
-            .collect::<Result<_, _>>()
-            .unwrap();
+        let _held_reservations = hold_every_connection(&budget);
 
-        assert!(matches!(
-            budget.reserve(),
-            Err(
+        assert_eq!(
+            budget.reserve().err().map(|refusal| refusal.to_string()),
+            Some(
                 ClusterHarnessError::ConcurrentConnectionsExceedPortableListenBacklog {
                     held_connections: PORTABLE_LISTEN_BACKLOG,
                     limit: PORTABLE_LISTEN_BACKLOG,
                 }
+                .to_string()
             )
-        ));
+        );
     }
 
     #[test]
@@ -112,4 +124,44 @@ mod tests {
 
         assert_eq!(reservation.held_connections.load(Ordering::Acquire), 1);
     }
+
+    #[test]
+    fn a_request_beyond_the_budget_is_refused_before_it_runs() {
+        let budget = ConcurrentConnectionBudget::default();
+        let _held_reservations = hold_every_connection(&budget);
+
+        assert_eq!(
+            budget
+                .hold_during(pending::<Result<(), ClusterHarnessError>>())
+                .now_or_never()
+                .map(|refused_request| refused_request.map_err(|refusal| refusal.to_string())),
+            Some(Err(
+                ClusterHarnessError::ConcurrentConnectionsExceedPortableListenBacklog {
+                    held_connections: PORTABLE_LISTEN_BACKLOG,
+                    limit: PORTABLE_LISTEN_BACKLOG,
+                }
+                .to_string()
+            ))
+        );
+    }
+
+    #[test]
+    fn a_stream_beyond_the_budget_is_refused_before_it_runs() {
+        let budget = ConcurrentConnectionBudget::default();
+        let _held_reservations = hold_every_connection(&budget);
+
+        assert_eq!(
+            budget
+                .hold_while_streaming(pending::<ClientResult<InferenceMessageStream>>())
+                .now_or_never()
+                .map(|refused_stream| refused_stream.err().map(|refusal| refusal.to_string())),
+            Some(Some(
+                ClusterHarnessError::ConcurrentConnectionsExceedPortableListenBacklog {
+                    held_connections: PORTABLE_LISTEN_BACKLOG,
+                    limit: PORTABLE_LISTEN_BACKLOG,
+                }
+                .to_string()
+            ))
+        );
+    }
 }
diff --git a/paddler_test_cluster_harness/src/dashboard_attributes.rs b/paddler_test_cluster_harness/src/dashboard_attributes.rs
index 362ad04cb..25f7f852d 100644
--- a/paddler_test_cluster_harness/src/dashboard_attributes.rs
+++ b/paddler_test_cluster_harness/src/dashboard_attributes.rs
@@ -1,7 +1,7 @@
 use std::collections::BTreeMap;
-use std::str::from_utf8;
 
 use quick_xml::Reader;
+use quick_xml::XmlVersion;
 use quick_xml::events::BytesStart;
 use quick_xml::events::Event;
 
@@ -17,13 +17,11 @@ fn attributes_of(
         .map(|attribute_result| {
             let attribute =
                 attribute_result.map_err(ClusterHarnessError::DashboardAttributeMalformed)?;
-            let name = from_utf8(attribute.key.as_ref())
-                .map_err(ClusterHarnessError::DashboardAttributeNameNotUtf8)?;
             let value = attribute
-                .unescape_value()
+                .normalized_value(XmlVersion::Implicit1_0)
                 .map_err(ClusterHarnessError::DashboardAttributeValueUnreadable)?;
 
-            Ok((name.to_owned(), value.into_owned()))
+            Ok((attribute.key.as_ref().to_owned(), value.into_owned()))
         })
         .collect()
 }
@@ -53,3 +51,90 @@ pub fn dashboard_attributes(page: &str) -> Result<BTreeMap<String, String>, Clus
         }
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use std::collections::BTreeMap;
+    use std::error::Error as _;
+    use std::mem::discriminant;
+
+    use quick_xml::errors::Error as XmlError;
+    use quick_xml::errors::SyntaxError;
+    use quick_xml::escape::EscapeError;
+    use quick_xml::events::attributes::AttrError;
+
+    use super::dashboard_attributes;
+    use crate::cluster_harness_error::ClusterHarnessError;
+
+    #[test]
+    fn a_self_closing_dashboard_element_carries_its_attributes() {
+        assert_eq!(
+            dashboard_attributes(
+                r#"<html><body><div id="paddler-dashboard" data-buffered-requests="7"/></body></html>"#
+            )
+            .unwrap(),
+            BTreeMap::from([
+                ("data-buffered-requests".to_owned(), "7".to_owned()),
+                ("id".to_owned(), "paddler-dashboard".to_owned()),
+            ])
+        );
+    }
+
+    #[test]
+    fn markup_ending_inside_a_tag_is_unreadable() {
+        let unreadable_markup =
+            dashboard_attributes(r#"<html><body><div id="paddler-dashboard""#).unwrap_err();
+
+        assert_eq!(
+            discriminant(&unreadable_markup),
+            discriminant(&ClusterHarnessError::DashboardMarkupUnreadable(
+                XmlError::Syntax(SyntaxError::UnclosedTag)
+            ))
+        );
+        assert_eq!(
+            unreadable_markup
+                .source()
+                .and_then(|xml_error| xml_error.source())
+                .and_then(|syntax_error| syntax_error.downcast_ref::<SyntaxError>()),
+            Some(&SyntaxError::UnclosedTag)
+        );
+    }
+
+    #[test]
+    fn a_repeated_attribute_is_malformed() {
+        let malformed_attribute =
+            dashboard_attributes(r#"<div id="paddler-dashboard" id="another-dashboard">"#)
+                .unwrap_err();
+
+        assert_eq!(
+            discriminant(&malformed_attribute),
+            discriminant(&ClusterHarnessError::DashboardAttributeMalformed(
+                AttrError::Duplicated(27, 4)
+            ))
+        );
+        assert_eq!(
+            malformed_attribute
+                .source()
+                .and_then(|attribute_error| attribute_error.downcast_ref::<AttrError>()),
+            Some(&AttrError::Duplicated(27, 4))
+        );
+    }
+
+    #[test]
+    fn an_attribute_referencing_an_unknown_entity_is_unreadable() {
+        assert!(matches!(
+            dashboard_attributes(r#"<div id="paddler-dashboard" data-version="&unknown;">"#),
+            Err(ClusterHarnessError::DashboardAttributeValueUnreadable(XmlError::Escape(
+                EscapeError::UnrecognizedEntity(_, entity)
+            ))) if entity == "unknown"
+        ));
+    }
+
+    #[test]
+    fn a_page_without_the_dashboard_reports_it_missing() {
+        assert_eq!(
+            discriminant(&dashboard_attributes("<html><body></body></html>").unwrap_err()),
+            discriminant(&ClusterHarnessError::DashboardElementMissing)
+        );
+    }
+}
diff --git a/paddler_test_cluster_harness/src/embedding_cluster_params.rs b/paddler_test_cluster_harness/src/embedding_cluster_params.rs
index a48d05fce..13d2a5ae3 100644
--- a/paddler_test_cluster_harness/src/embedding_cluster_params.rs
+++ b/paddler_test_cluster_harness/src/embedding_cluster_params.rs
@@ -1,19 +1,24 @@
 use std::time::Duration;
 
 use paddler_inference_parameters::all_gpu_layers::ALL_GPU_LAYERS;
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
+use paddler_inference_parameters::embedding_parameters::EmbeddingParameters;
+use paddler_inference_parameters::model_runtime_parameters::ModelRuntimeParameters;
 use paddler_messaging::balancer_desired_state::BalancerDesiredState;
 
 use crate::agent_config::AgentConfig;
+use crate::cluster_desired_state::ClusterDesiredState;
 use crate::cluster_params::ClusterParams;
 use crate::longer_than_any_test_run::LONGER_THAN_ANY_TEST_RUN;
+use crate::model_card::ModelCard;
 use crate::model_card::nomic_embed_text_v1_5::nomic_embed_text_v1_5;
 
 pub struct EmbeddingClusterParams {
     pub agents: Vec<AgentConfig>,
     pub buffered_request_timeout: Duration,
-    pub inference_parameters: InferenceParameters,
+    pub embedding_parameters: EmbeddingParameters,
     pub max_buffered_requests: u64,
+    pub model_card: ModelCard,
+    pub model_runtime_parameters: ModelRuntimeParameters,
 }
 
 impl EmbeddingClusterParams {
@@ -22,20 +27,22 @@ impl EmbeddingClusterParams {
         let Self {
             agents,
             buffered_request_timeout,
-            inference_parameters,
+            embedding_parameters,
             max_buffered_requests,
+            model_card,
+            model_runtime_parameters,
         } = self;
 
         ClusterParams {
             agents,
             buffered_request_timeout,
-            desired_state: Some(BalancerDesiredState {
-                inference_parameters: InferenceParameters {
+            desired_state: ClusterDesiredState::Apply(Box::new(BalancerDesiredState {
+                model_runtime_parameters: ModelRuntimeParameters {
                     n_gpu_layers: ALL_GPU_LAYERS,
-                    ..inference_parameters
+                    ..model_runtime_parameters
                 },
-                ..nomic_embed_text_v1_5().into_desired_state()
-            }),
+                ..model_card.into_desired_state_with_embedding_parameters(embedding_parameters)
+            })),
             max_buffered_requests,
             wait_for_slots_ready: true,
             ..ClusterParams::default()
@@ -48,8 +55,10 @@ impl Default for EmbeddingClusterParams {
         Self {
             agents: AgentConfig::uniform(1, 4),
             buffered_request_timeout: LONGER_THAN_ANY_TEST_RUN,
-            inference_parameters: InferenceParameters::default(),
+            embedding_parameters: EmbeddingParameters::default(),
             max_buffered_requests: 10,
+            model_card: nomic_embed_text_v1_5(),
+            model_runtime_parameters: ModelRuntimeParameters::default(),
         }
     }
 }
diff --git a/paddler_test_cluster_harness/src/failed_run_exit_code.rs b/paddler_test_cluster_harness/src/failed_run_exit_code.rs
new file mode 100644
index 000000000..65d76fa91
--- /dev/null
+++ b/paddler_test_cluster_harness/src/failed_run_exit_code.rs
@@ -0,0 +1 @@
+pub const FAILED_RUN_EXIT_CODE: i32 = 1;
diff --git a/paddler_test_cluster_harness/src/fixture_path.rs b/paddler_test_cluster_harness/src/fixture_path.rs
new file mode 100644
index 000000000..7b5119b33
--- /dev/null
+++ b/paddler_test_cluster_harness/src/fixture_path.rs
@@ -0,0 +1,8 @@
+use std::path::PathBuf;
+
+#[must_use]
+pub fn fixture_path(fixture_file_name: &str) -> PathBuf {
+    PathBuf::from(env!("CARGO_MANIFEST_DIR"))
+        .with_file_name("fixtures")
+        .join(fixture_file_name)
+}
diff --git a/paddler_test_cluster_harness/src/half_closed_client.rs b/paddler_test_cluster_harness/src/half_closed_client.rs
index 7dd2435aa..0e82f7ed2 100644
--- a/paddler_test_cluster_harness/src/half_closed_client.rs
+++ b/paddler_test_cluster_harness/src/half_closed_client.rs
@@ -2,8 +2,7 @@ use std::net::SocketAddr;
 
 use anyhow::Context as _;
 use anyhow::Result;
-use serde::Serialize;
-use serde_json::to_string;
+use serde_json::Value;
 use tokio::io::AsyncWriteExt as _;
 use tokio::net::TcpStream;
 
@@ -14,15 +13,12 @@ pub struct HalfClosedClient {
 }
 
 impl HalfClosedClient {
-    pub async fn post_json_then_half_close<TBody>(
+    pub async fn post_json_then_half_close(
         addr: SocketAddr,
         path: &str,
-        body: &TBody,
-    ) -> Result<Self>
-    where
-        TBody: Serialize,
-    {
-        let serialized_body = to_string(body)?;
+        body: &Value,
+    ) -> Result<Self> {
+        let serialized_body = body.to_string();
         let request = format!(
             "POST {path} HTTP/1.1\r\n\
              Host: {addr}\r\n\
@@ -37,9 +33,11 @@ impl HalfClosedClient {
             .await
             .map_err(|source| ClusterHarnessError::HalfClosedClientUnreachable { addr, source })?;
 
-        socket.write_all(request.as_bytes()).await?;
-
-        Ok(Self { socket })
+        socket
+            .write_all(request.as_bytes())
+            .await
+            .map(|()| Self { socket })
+            .context("half-closed client must send its whole request")
     }
 
     pub async fn half_close(&mut self) -> Result<()> {
diff --git a/paddler_test_cluster_harness/src/inference_stream_error_codes.rs b/paddler_test_cluster_harness/src/inference_stream_error_codes.rs
new file mode 100644
index 000000000..312e78f82
--- /dev/null
+++ b/paddler_test_cluster_harness/src/inference_stream_error_codes.rs
@@ -0,0 +1,93 @@
+use futures_util::StreamExt as _;
+
+use paddler_client::inference_message_stream::InferenceMessageStream;
+use paddler_messaging::inference_client::message::Message as InferenceMessage;
+
+use crate::cluster_harness_error::ClusterHarnessError;
+
+pub async fn inference_stream_error_codes(
+    mut stream: InferenceMessageStream,
+) -> Result<Vec<i32>, ClusterHarnessError> {
+    let mut error_codes = Vec::new();
+
+    while let Some(message) = stream.next().await {
+        if let InferenceMessage::Error(error_envelope) =
+            message.map_err(ClusterHarnessError::TokenStreamFailed)?
+        {
+            error_codes.push(error_envelope.error.code);
+        }
+    }
+
+    Ok(error_codes)
+}
+
+#[cfg(test)]
+mod tests {
+    use std::error::Error as _;
+    use std::mem::discriminant;
+
+    use futures_util::stream::iter;
+
+    use paddler_client::error::Error as ClientError;
+    use paddler_messaging::generated_token_result::GeneratedTokenResult;
+    use paddler_messaging::inference_client::message::Message as InferenceMessage;
+    use paddler_messaging::inference_client::response::Response as InferenceResponse;
+    use paddler_messaging::jsonrpc::error::Error as JsonRpcError;
+    use paddler_messaging::jsonrpc::error_envelope::ErrorEnvelope;
+    use paddler_messaging::jsonrpc::response_envelope::ResponseEnvelope;
+
+    use super::inference_stream_error_codes;
+    use crate::cluster_harness_error::ClusterHarnessError;
+
+    #[tokio::test]
+    async fn collects_the_code_of_every_error_and_skips_the_other_messages() {
+        let error_codes = inference_stream_error_codes(Box::pin(iter([
+            Ok(InferenceMessage::Response(ResponseEnvelope {
+                generated_by: None,
+                request_id: "request-1".to_owned(),
+                response: InferenceResponse::GeneratedToken(GeneratedTokenResult::ContentToken(
+                    "hi".to_owned(),
+                )),
+            })),
+            Ok(InferenceMessage::Error(ErrorEnvelope {
+                request_id: "request-1".to_owned(),
+                error: JsonRpcError {
+                    code: 503,
+                    description: "unavailable".to_owned(),
+                },
+            })),
+        ])))
+        .await
+        .unwrap();
+
+        assert_eq!(error_codes, vec![503]);
+    }
+
+    #[tokio::test]
+    async fn reports_a_stream_that_yields_an_error() {
+        let stream_error =
+            inference_stream_error_codes(Box::pin(iter([Err(ClientError::ConnectionDropped {
+                request_id: "request-1".to_owned(),
+            })])))
+            .await
+            .unwrap_err();
+
+        assert_eq!(
+            discriminant(&stream_error),
+            discriminant(&ClusterHarnessError::TokenStreamFailed(
+                ClientError::ConnectionDropped {
+                    request_id: String::new(),
+                }
+            ))
+        );
+        assert_eq!(
+            stream_error.source().map(ToString::to_string),
+            Some(
+                ClientError::ConnectionDropped {
+                    request_id: "request-1".to_owned(),
+                }
+                .to_string()
+            )
+        );
+    }
+}
diff --git a/paddler_test_cluster_harness/src/lib.rs b/paddler_test_cluster_harness/src/lib.rs
index cb691880e..79441b71f 100644
--- a/paddler_test_cluster_harness/src/lib.rs
+++ b/paddler_test_cluster_harness/src/lib.rs
@@ -1,14 +1,21 @@
 pub mod agent_config;
 mod agent_count_is;
+pub mod agent_desired_model_at;
+pub mod agent_readiness;
 mod agent_slots_processing_is;
 pub mod agent_spawner;
 pub mod budgeted_inference_message_stream;
 mod buffered_request_count_is;
+pub mod ca_bundle_without_certificates;
 pub mod cluster;
+pub mod cluster_desired_state;
 pub mod cluster_harness_error;
 pub mod cluster_params;
+pub mod cluster_snapshots_watchers;
+pub mod collect_decision_results;
 pub mod collect_embedding_results;
 pub mod collect_generated_tokens;
+pub mod collected_decision_results;
 pub mod collected_embedding_results;
 pub mod collected_generated_tokens;
 pub mod concurrent_connection_budget;
@@ -17,26 +24,38 @@ pub mod dashboard_attributes;
 pub mod embedding_cluster_params;
 pub mod embedding_with_producer;
 pub mod ephemeral_loopback_addr;
+pub mod failed_run_exit_code;
+pub mod fixture_path;
 pub mod half_closed_client;
+pub mod inference_stream_error_codes;
 pub mod is_unending_generation_text;
 pub mod load_fixture_data_uri;
 pub mod load_test_image_data_uri;
 pub mod longer_than_any_test_run;
 pub mod managed_process;
 pub mod model_card;
-mod openai_api_client;
-pub mod raw_agent_socket;
+pub mod noul_system_one_request;
+pub mod noul_system_one_request_about;
+pub mod openai_api_client;
+pub mod pointer_head_fixture;
+pub mod process_end;
 pub mod raw_inference_socket;
-#[cfg(any(target_os = "macos", target_os = "linux"))]
 pub mod resource_snapshot;
-#[cfg(any(target_os = "macos", target_os = "linux"))]
 pub mod resource_snapshot_diff;
 pub mod running_agent;
 pub mod running_balancer;
 pub mod snapshots_stream;
 pub mod snapshots_watcher;
-pub mod state_database_file;
+pub mod synthetic_pointer_head_fixture;
+pub mod system_one_request;
+pub mod temporary_state_database_file;
 pub mod token_result_with_producer;
+pub mod typesafe_api_client;
+pub mod typesafe_api_response;
+pub mod typesafe_request_model;
 pub mod unending_generation;
 pub mod unending_grammar;
 pub mod unexpiring_shutdown_options;
+pub mod untrusted_agent_socket_client;
+pub mod untrusted_balancer_socket;
+pub mod untrusted_balancer_socket_server;
diff --git a/paddler_test_cluster_harness/src/load_fixture_data_uri.rs b/paddler_test_cluster_harness/src/load_fixture_data_uri.rs
index 623a33c8f..2c19b1fb4 100644
--- a/paddler_test_cluster_harness/src/load_fixture_data_uri.rs
+++ b/paddler_test_cluster_harness/src/load_fixture_data_uri.rs
@@ -5,12 +5,33 @@ use anyhow::Result;
 use base64::Engine as _;
 use base64::engine::general_purpose::STANDARD as BASE64_STANDARD;
 
+use crate::fixture_path::fixture_path;
+
 pub fn load_fixture_data_uri(fixture_name: &str, mime_type: &str) -> Result<String> {
-    let fixture_path = format!("{}/../fixtures/{fixture_name}", env!("CARGO_MANIFEST_DIR"));
+    let fixture_path = fixture_path(fixture_name);
     let fixture_bytes = read(&fixture_path)
-        .with_context(|| format!("failed to read test fixture {fixture_path}"))?;
+        .with_context(|| format!("failed to read test fixture {}", fixture_path.display()))?;
 
     let encoded = BASE64_STANDARD.encode(&fixture_bytes);
 
     Ok(format!("data:{mime_type};base64,{encoded}"))
 }
+
+#[cfg(test)]
+mod tests {
+    use std::io;
+    use std::io::ErrorKind;
+
+    use super::load_fixture_data_uri;
+
+    #[test]
+    fn a_missing_fixture_is_reported_as_not_found() {
+        let error = load_fixture_data_uri("missing_fixture.bin", "application/octet-stream")
+            .expect_err("a missing fixture must not load");
+
+        assert_eq!(
+            error.downcast_ref::<io::Error>().map(io::Error::kind),
+            Some(ErrorKind::NotFound)
+        );
+    }
+}
diff --git a/paddler_test_cluster_harness/src/load_test_image_data_uri.rs b/paddler_test_cluster_harness/src/load_test_image_data_uri.rs
index d137452f3..10898a5d9 100644
--- a/paddler_test_cluster_harness/src/load_test_image_data_uri.rs
+++ b/paddler_test_cluster_harness/src/load_test_image_data_uri.rs
@@ -10,6 +10,8 @@ pub fn load_test_image_data_uri() -> Result<String> {
 mod tests {
     use std::fs::read;
 
+    use crate::fixture_path::fixture_path;
+
     use data_url::DataUrl;
 
     use super::load_test_image_data_uri;
@@ -22,13 +24,6 @@ mod tests {
 
         assert_eq!(parsed_fixture.mime_type().type_, "image");
         assert_eq!(parsed_fixture.mime_type().subtype, "jpeg");
-        assert_eq!(
-            body,
-            read(concat!(
-                env!("CARGO_MANIFEST_DIR"),
-                "/../fixtures/llamas.jpg"
-            ))
-            .unwrap()
-        );
+        assert_eq!(body, read(fixture_path("llamas.jpg")).unwrap());
     }
 }
diff --git a/paddler_test_cluster_harness/src/managed_process.rs b/paddler_test_cluster_harness/src/managed_process.rs
index ed203161b..5ea6e1568 100644
--- a/paddler_test_cluster_harness/src/managed_process.rs
+++ b/paddler_test_cluster_harness/src/managed_process.rs
@@ -1,7 +1,18 @@
 use anyhow::Result;
 use async_trait::async_trait;
 
+use crate::process_end::ProcessEnd;
+
 #[async_trait]
 pub trait ManagedProcess: Send {
-    async fn shutdown(self: Box<Self>) -> Result<()>;
+    fn terminate(&mut self) -> Result<()>;
+
+    async fn exited(&mut self) -> ProcessEnd;
+}
+
+impl dyn ManagedProcess {
+    pub async fn shutdown(mut self: Box<Self>) -> Result<()> {
+        self.terminate()?;
+        self.exited().await.into_result()
+    }
 }
diff --git a/paddler_test_cluster_harness/src/model_card/deepseek_r1_distill_llama_8b.rs b/paddler_test_cluster_harness/src/model_card/deepseek_r1_distill_llama_8b.rs
index 4956fcbe3..4c61321c5 100644
--- a/paddler_test_cluster_harness/src/model_card/deepseek_r1_distill_llama_8b.rs
+++ b/paddler_test_cluster_harness/src/model_card/deepseek_r1_distill_llama_8b.rs
@@ -1,4 +1,4 @@
-use paddler_messaging::huggingface_model_reference::HuggingFaceModelReference;
+use paddler_model_source::huggingface_model_reference::HuggingFaceModelReference;
 
 use crate::model_card::ModelCard;
 
@@ -7,7 +7,8 @@ pub fn deepseek_r1_distill_llama_8b() -> ModelCard {
     ModelCard {
         reference: HuggingFaceModelReference {
             filename: "DeepSeek-R1-Distill-Llama-8B-Q4_K_M.gguf".to_owned(),
-            repo_id: "unsloth/DeepSeek-R1-Distill-Llama-8B-GGUF".to_owned(),
+            repository_name: "DeepSeek-R1-Distill-Llama-8B-GGUF".to_owned(),
+            repository_owner: "unsloth".to_owned(),
             revision: "main".to_owned(),
         },
     }
diff --git a/paddler_test_cluster_harness/src/model_card/every_model_card.rs b/paddler_test_cluster_harness/src/model_card/every_model_card.rs
index db94e0082..ff7ea5abf 100644
--- a/paddler_test_cluster_harness/src/model_card/every_model_card.rs
+++ b/paddler_test_cluster_harness/src/model_card/every_model_card.rs
@@ -4,6 +4,7 @@ use crate::model_card::gemma_3_4b_it::gemma_3_4b_it;
 use crate::model_card::gemma_3_4b_it_mmproj::gemma_3_4b_it_mmproj;
 use crate::model_card::gemma_4_e2b_it::gemma_4_e2b_it;
 use crate::model_card::gemma_4_e2b_it_mmproj::gemma_4_e2b_it_mmproj;
+use crate::model_card::lfm2_colbert_350m::lfm2_colbert_350m;
 use crate::model_card::ministral_3_3b_reasoning::ministral_3_3b_reasoning;
 use crate::model_card::ministral_3_8b_reasoning::ministral_3_8b_reasoning;
 use crate::model_card::ministral_3_8b_reasoning_mmproj::ministral_3_8b_reasoning_mmproj;
@@ -13,8 +14,11 @@ use crate::model_card::qwen2_5_vl_3b_mmproj::qwen2_5_vl_3b_mmproj;
 use crate::model_card::qwen3_0_6b::qwen3_0_6b;
 use crate::model_card::qwen3_5_0_8b::qwen3_5_0_8b;
 use crate::model_card::qwen3_5_0_8b_mmproj::qwen3_5_0_8b_mmproj;
+use crate::model_card::qwen3_8_27b::qwen3_8_27b;
 use crate::model_card::smolvlm2_256m::smolvlm2_256m;
 use crate::model_card::smolvlm2_256m_mmproj::smolvlm2_256m_mmproj;
+use crate::model_card::ultravox_v0_5_llama_3_2_1b::ultravox_v0_5_llama_3_2_1b;
+use crate::model_card::ultravox_v0_5_llama_3_2_1b_mmproj::ultravox_v0_5_llama_3_2_1b_mmproj;
 
 #[must_use]
 pub fn every_model_card() -> Vec<ModelCard> {
@@ -24,6 +28,7 @@ pub fn every_model_card() -> Vec<ModelCard> {
         gemma_3_4b_it_mmproj(),
         gemma_4_e2b_it(),
         gemma_4_e2b_it_mmproj(),
+        lfm2_colbert_350m(),
         ministral_3_3b_reasoning(),
         ministral_3_8b_reasoning(),
         ministral_3_8b_reasoning_mmproj(),
@@ -33,7 +38,10 @@ pub fn every_model_card() -> Vec<ModelCard> {
         qwen3_0_6b(),
         qwen3_5_0_8b(),
         qwen3_5_0_8b_mmproj(),
+        qwen3_8_27b(),
         smolvlm2_256m(),
         smolvlm2_256m_mmproj(),
+        ultravox_v0_5_llama_3_2_1b(),
+        ultravox_v0_5_llama_3_2_1b_mmproj(),
     ]
 }
diff --git a/paddler_test_cluster_harness/src/model_card/gemma_3_4b_it.rs b/paddler_test_cluster_harness/src/model_card/gemma_3_4b_it.rs
index f5f775f7c..d51101f7c 100644
--- a/paddler_test_cluster_harness/src/model_card/gemma_3_4b_it.rs
+++ b/paddler_test_cluster_harness/src/model_card/gemma_3_4b_it.rs
@@ -1,4 +1,4 @@
-use paddler_messaging::huggingface_model_reference::HuggingFaceModelReference;
+use paddler_model_source::huggingface_model_reference::HuggingFaceModelReference;
 
 use crate::model_card::ModelCard;
 
@@ -7,7 +7,8 @@ pub fn gemma_3_4b_it() -> ModelCard {
     ModelCard {
         reference: HuggingFaceModelReference {
             filename: "gemma-3-4b-it-Q4_K_M.gguf".to_owned(),
-            repo_id: "unsloth/gemma-3-4b-it-GGUF".to_owned(),
+            repository_name: "gemma-3-4b-it-GGUF".to_owned(),
+            repository_owner: "unsloth".to_owned(),
             revision: "main".to_owned(),
         },
     }
diff --git a/paddler_test_cluster_harness/src/model_card/gemma_3_4b_it_mmproj.rs b/paddler_test_cluster_harness/src/model_card/gemma_3_4b_it_mmproj.rs
index a0769bca8..a1b8ba3b5 100644
--- a/paddler_test_cluster_harness/src/model_card/gemma_3_4b_it_mmproj.rs
+++ b/paddler_test_cluster_harness/src/model_card/gemma_3_4b_it_mmproj.rs
@@ -1,4 +1,4 @@
-use paddler_messaging::huggingface_model_reference::HuggingFaceModelReference;
+use paddler_model_source::huggingface_model_reference::HuggingFaceModelReference;
 
 use crate::model_card::ModelCard;
 
@@ -7,7 +7,8 @@ pub fn gemma_3_4b_it_mmproj() -> ModelCard {
     ModelCard {
         reference: HuggingFaceModelReference {
             filename: "mmproj-F16.gguf".to_owned(),
-            repo_id: "unsloth/gemma-3-4b-it-GGUF".to_owned(),
+            repository_name: "gemma-3-4b-it-GGUF".to_owned(),
+            repository_owner: "unsloth".to_owned(),
             revision: "main".to_owned(),
         },
     }
diff --git a/paddler_test_cluster_harness/src/model_card/gemma_4_e2b_it.rs b/paddler_test_cluster_harness/src/model_card/gemma_4_e2b_it.rs
index f351c2a6e..e95103ea9 100644
--- a/paddler_test_cluster_harness/src/model_card/gemma_4_e2b_it.rs
+++ b/paddler_test_cluster_harness/src/model_card/gemma_4_e2b_it.rs
@@ -1,4 +1,4 @@
-use paddler_messaging::huggingface_model_reference::HuggingFaceModelReference;
+use paddler_model_source::huggingface_model_reference::HuggingFaceModelReference;
 
 use crate::model_card::ModelCard;
 
@@ -7,7 +7,8 @@ pub fn gemma_4_e2b_it() -> ModelCard {
     ModelCard {
         reference: HuggingFaceModelReference {
             filename: "gemma-4-E2B-it-Q4_K_M.gguf".to_owned(),
-            repo_id: "unsloth/gemma-4-E2B-it-GGUF".to_owned(),
+            repository_name: "gemma-4-E2B-it-GGUF".to_owned(),
+            repository_owner: "unsloth".to_owned(),
             revision: "main".to_owned(),
         },
     }
diff --git a/paddler_test_cluster_harness/src/model_card/gemma_4_e2b_it_mmproj.rs b/paddler_test_cluster_harness/src/model_card/gemma_4_e2b_it_mmproj.rs
index 0f51ed26d..7e3bb0542 100644
--- a/paddler_test_cluster_harness/src/model_card/gemma_4_e2b_it_mmproj.rs
+++ b/paddler_test_cluster_harness/src/model_card/gemma_4_e2b_it_mmproj.rs
@@ -1,4 +1,4 @@
-use paddler_messaging::huggingface_model_reference::HuggingFaceModelReference;
+use paddler_model_source::huggingface_model_reference::HuggingFaceModelReference;
 
 use crate::model_card::ModelCard;
 
@@ -7,7 +7,8 @@ pub fn gemma_4_e2b_it_mmproj() -> ModelCard {
     ModelCard {
         reference: HuggingFaceModelReference {
             filename: "mmproj-F16.gguf".to_owned(),
-            repo_id: "unsloth/gemma-4-E2B-it-GGUF".to_owned(),
+            repository_name: "gemma-4-E2B-it-GGUF".to_owned(),
+            repository_owner: "unsloth".to_owned(),
             revision: "main".to_owned(),
         },
     }
diff --git a/paddler_test_cluster_harness/src/model_card/lfm2_colbert_350m.rs b/paddler_test_cluster_harness/src/model_card/lfm2_colbert_350m.rs
new file mode 100644
index 000000000..c15edb032
--- /dev/null
+++ b/paddler_test_cluster_harness/src/model_card/lfm2_colbert_350m.rs
@@ -0,0 +1,15 @@
+use paddler_model_source::huggingface_model_reference::HuggingFaceModelReference;
+
+use crate::model_card::ModelCard;
+
+#[must_use]
+pub fn lfm2_colbert_350m() -> ModelCard {
+    ModelCard {
+        reference: HuggingFaceModelReference {
+            filename: "LFM2-ColBERT-350M-Q4_K_M.gguf".to_owned(),
+            repository_name: "LFM2-ColBERT-350M-GGUF".to_owned(),
+            repository_owner: "LiquidAI".to_owned(),
+            revision: "main".to_owned(),
+        },
+    }
+}
diff --git a/paddler_test_cluster_harness/src/model_card/ministral_3_3b_reasoning.rs b/paddler_test_cluster_harness/src/model_card/ministral_3_3b_reasoning.rs
index f641966d9..b67766ddb 100644
--- a/paddler_test_cluster_harness/src/model_card/ministral_3_3b_reasoning.rs
+++ b/paddler_test_cluster_harness/src/model_card/ministral_3_3b_reasoning.rs
@@ -1,4 +1,4 @@
-use paddler_messaging::huggingface_model_reference::HuggingFaceModelReference;
+use paddler_model_source::huggingface_model_reference::HuggingFaceModelReference;
 
 use crate::model_card::ModelCard;
 
@@ -7,7 +7,8 @@ pub fn ministral_3_3b_reasoning() -> ModelCard {
     ModelCard {
         reference: HuggingFaceModelReference {
             filename: "Ministral-3-3B-Reasoning-2512-Q4_K_M.gguf".to_owned(),
-            repo_id: "unsloth/Ministral-3-3B-Reasoning-2512-GGUF".to_owned(),
+            repository_name: "Ministral-3-3B-Reasoning-2512-GGUF".to_owned(),
+            repository_owner: "unsloth".to_owned(),
             revision: "main".to_owned(),
         },
     }
diff --git a/paddler_test_cluster_harness/src/model_card/ministral_3_8b_reasoning.rs b/paddler_test_cluster_harness/src/model_card/ministral_3_8b_reasoning.rs
index a356de995..32bd288dc 100644
--- a/paddler_test_cluster_harness/src/model_card/ministral_3_8b_reasoning.rs
+++ b/paddler_test_cluster_harness/src/model_card/ministral_3_8b_reasoning.rs
@@ -1,4 +1,4 @@
-use paddler_messaging::huggingface_model_reference::HuggingFaceModelReference;
+use paddler_model_source::huggingface_model_reference::HuggingFaceModelReference;
 
 use crate::model_card::ModelCard;
 
@@ -7,7 +7,8 @@ pub fn ministral_3_8b_reasoning() -> ModelCard {
     ModelCard {
         reference: HuggingFaceModelReference {
             filename: "Ministral-3-8B-Reasoning-2512-Q4_K_M.gguf".to_owned(),
-            repo_id: "unsloth/Ministral-3-8B-Reasoning-2512-GGUF".to_owned(),
+            repository_name: "Ministral-3-8B-Reasoning-2512-GGUF".to_owned(),
+            repository_owner: "unsloth".to_owned(),
             revision: "main".to_owned(),
         },
     }
diff --git a/paddler_test_cluster_harness/src/model_card/ministral_3_8b_reasoning_mmproj.rs b/paddler_test_cluster_harness/src/model_card/ministral_3_8b_reasoning_mmproj.rs
index 5b28255f3..466974778 100644
--- a/paddler_test_cluster_harness/src/model_card/ministral_3_8b_reasoning_mmproj.rs
+++ b/paddler_test_cluster_harness/src/model_card/ministral_3_8b_reasoning_mmproj.rs
@@ -1,4 +1,4 @@
-use paddler_messaging::huggingface_model_reference::HuggingFaceModelReference;
+use paddler_model_source::huggingface_model_reference::HuggingFaceModelReference;
 
 use crate::model_card::ModelCard;
 
@@ -7,7 +7,8 @@ pub fn ministral_3_8b_reasoning_mmproj() -> ModelCard {
     ModelCard {
         reference: HuggingFaceModelReference {
             filename: "mmproj-F16.gguf".to_owned(),
-            repo_id: "unsloth/Ministral-3-8B-Reasoning-2512-GGUF".to_owned(),
+            repository_name: "Ministral-3-8B-Reasoning-2512-GGUF".to_owned(),
+            repository_owner: "unsloth".to_owned(),
             revision: "main".to_owned(),
         },
     }
diff --git a/paddler_test_cluster_harness/src/model_card/mod.rs b/paddler_test_cluster_harness/src/model_card/mod.rs
index 1918f9e2a..8e3ac5b21 100644
--- a/paddler_test_cluster_harness/src/model_card/mod.rs
+++ b/paddler_test_cluster_harness/src/model_card/mod.rs
@@ -4,6 +4,7 @@ pub mod gemma_3_4b_it;
 pub mod gemma_3_4b_it_mmproj;
 pub mod gemma_4_e2b_it;
 pub mod gemma_4_e2b_it_mmproj;
+pub mod lfm2_colbert_350m;
 pub mod ministral_3_3b_reasoning;
 pub mod ministral_3_8b_reasoning;
 pub mod ministral_3_8b_reasoning_mmproj;
@@ -13,36 +14,65 @@ pub mod qwen2_5_vl_3b_mmproj;
 pub mod qwen3_0_6b;
 pub mod qwen3_5_0_8b;
 pub mod qwen3_5_0_8b_mmproj;
+pub mod qwen3_8_27b;
 pub mod smolvlm2_256m;
 pub mod smolvlm2_256m_mmproj;
+pub mod ultravox_v0_5_llama_3_2_1b;
+pub mod ultravox_v0_5_llama_3_2_1b_mmproj;
+
+use std::path::PathBuf;
+
+use futures::executor::block_on;
+use hf_hub::HFResult;
 
 use paddler_inference_parameters::all_gpu_layers::ALL_GPU_LAYERS;
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
+use paddler_inference_parameters::embedding_parameters::EmbeddingParameters;
+use paddler_inference_parameters::model_runtime_parameters::ModelRuntimeParameters;
+use paddler_inference_parameters::sampling_parameters::SamplingParameters;
 use paddler_messaging::agent_desired_model::AgentDesiredModel;
 use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-use paddler_messaging::huggingface_model_reference::HuggingFaceModelReference;
+use paddler_messaging::balancer_text_generation_settings::BalancerTextGenerationSettings;
+use paddler_messaging::decision_settings::DecisionSettings;
+use paddler_messaging::multimodal_settings::MultimodalSettings;
+use paddler_model_source::huggingface_cache_location::HuggingFaceCacheLocation;
+use paddler_model_source::huggingface_model_reference::HuggingFaceModelReference;
+use paddler_model_source::model_source::ModelSource;
 
 pub struct ModelCard {
     pub reference: HuggingFaceModelReference,
 }
 
 impl ModelCard {
+    pub fn cached_path(&self) -> HFResult<PathBuf> {
+        HuggingFaceCacheLocation::HubDefault
+            .client_builder()
+            .build()
+            .and_then(|client| {
+                block_on(
+                    self.reference
+                        .cached_path(&self.reference.repository(&client)),
+                )
+            })
+    }
+
     #[must_use]
     pub fn into_agent_desired_model(self) -> AgentDesiredModel {
-        AgentDesiredModel::HuggingFace(self.reference)
+        ModelSource::HuggingFace(self.reference).into_agent_desired_model()
     }
 
     #[must_use]
     pub fn into_desired_state(self) -> BalancerDesiredState {
         BalancerDesiredState {
-            chat_template_override: None,
-            inference_parameters: InferenceParameters {
+            model: self.into_agent_desired_model(),
+            model_runtime_parameters: ModelRuntimeParameters {
                 n_gpu_layers: ALL_GPU_LAYERS,
-                ..InferenceParameters::deterministic()
+                ..ModelRuntimeParameters::default()
             },
-            model: self.into_agent_desired_model(),
-            multimodal_projection: AgentDesiredModel::None,
-            use_chat_template_override: false,
+            text_generation: BalancerTextGenerationSettings {
+                sampling_parameters: SamplingParameters::deterministic(),
+                ..BalancerTextGenerationSettings::default()
+            },
+            ..BalancerDesiredState::default()
         }
     }
 
@@ -50,9 +80,39 @@ impl ModelCard {
     pub fn into_desired_state_with_multimodal_projection(
         self,
         multimodal_projection: Self,
+    ) -> BalancerDesiredState {
+        let desired_state = self.into_desired_state();
+
+        BalancerDesiredState {
+            text_generation: BalancerTextGenerationSettings {
+                multimodal: MultimodalSettings {
+                    projection: multimodal_projection.into_agent_desired_model(),
+                    ..MultimodalSettings::default()
+                },
+                ..desired_state.text_generation
+            },
+            ..desired_state
+        }
+    }
+
+    #[must_use]
+    pub fn into_decision_desired_state(
+        self,
+        pointer_head: AgentDesiredModel,
+    ) -> BalancerDesiredState {
+        BalancerDesiredState {
+            decision: DecisionSettings { pointer_head },
+            ..self.into_desired_state()
+        }
+    }
+
+    #[must_use]
+    pub fn into_desired_state_with_embedding_parameters(
+        self,
+        embedding_parameters: EmbeddingParameters,
     ) -> BalancerDesiredState {
         BalancerDesiredState {
-            multimodal_projection: multimodal_projection.into_agent_desired_model(),
+            embeddings: embedding_parameters,
             ..self.into_desired_state()
         }
     }
diff --git a/paddler_test_cluster_harness/src/model_card/nomic_embed_text_v1_5.rs b/paddler_test_cluster_harness/src/model_card/nomic_embed_text_v1_5.rs
index fb2ac687c..5589ffd29 100644
--- a/paddler_test_cluster_harness/src/model_card/nomic_embed_text_v1_5.rs
+++ b/paddler_test_cluster_harness/src/model_card/nomic_embed_text_v1_5.rs
@@ -1,4 +1,4 @@
-use paddler_messaging::huggingface_model_reference::HuggingFaceModelReference;
+use paddler_model_source::huggingface_model_reference::HuggingFaceModelReference;
 
 use crate::model_card::ModelCard;
 
@@ -7,7 +7,8 @@ pub fn nomic_embed_text_v1_5() -> ModelCard {
     ModelCard {
         reference: HuggingFaceModelReference {
             filename: "nomic-embed-text-v1.5.Q2_K.gguf".to_owned(),
-            repo_id: "nomic-ai/nomic-embed-text-v1.5-GGUF".to_owned(),
+            repository_name: "nomic-embed-text-v1.5-GGUF".to_owned(),
+            repository_owner: "nomic-ai".to_owned(),
             revision: "main".to_owned(),
         },
     }
diff --git a/paddler_test_cluster_harness/src/model_card/qwen2_5_vl_3b.rs b/paddler_test_cluster_harness/src/model_card/qwen2_5_vl_3b.rs
index e23b14777..874a87dc8 100644
--- a/paddler_test_cluster_harness/src/model_card/qwen2_5_vl_3b.rs
+++ b/paddler_test_cluster_harness/src/model_card/qwen2_5_vl_3b.rs
@@ -1,4 +1,4 @@
-use paddler_messaging::huggingface_model_reference::HuggingFaceModelReference;
+use paddler_model_source::huggingface_model_reference::HuggingFaceModelReference;
 
 use crate::model_card::ModelCard;
 
@@ -7,7 +7,8 @@ pub fn qwen2_5_vl_3b() -> ModelCard {
     ModelCard {
         reference: HuggingFaceModelReference {
             filename: "Qwen2.5-VL-3B-Instruct-Q4_K_M.gguf".to_owned(),
-            repo_id: "ggml-org/Qwen2.5-VL-3B-Instruct-GGUF".to_owned(),
+            repository_name: "Qwen2.5-VL-3B-Instruct-GGUF".to_owned(),
+            repository_owner: "ggml-org".to_owned(),
             revision: "main".to_owned(),
         },
     }
diff --git a/paddler_test_cluster_harness/src/model_card/qwen2_5_vl_3b_mmproj.rs b/paddler_test_cluster_harness/src/model_card/qwen2_5_vl_3b_mmproj.rs
index c6684a41d..ef0d1e18b 100644
--- a/paddler_test_cluster_harness/src/model_card/qwen2_5_vl_3b_mmproj.rs
+++ b/paddler_test_cluster_harness/src/model_card/qwen2_5_vl_3b_mmproj.rs
@@ -1,4 +1,4 @@
-use paddler_messaging::huggingface_model_reference::HuggingFaceModelReference;
+use paddler_model_source::huggingface_model_reference::HuggingFaceModelReference;
 
 use crate::model_card::ModelCard;
 
@@ -7,7 +7,8 @@ pub fn qwen2_5_vl_3b_mmproj() -> ModelCard {
     ModelCard {
         reference: HuggingFaceModelReference {
             filename: "mmproj-Qwen2.5-VL-3B-Instruct-Q8_0.gguf".to_owned(),
-            repo_id: "ggml-org/Qwen2.5-VL-3B-Instruct-GGUF".to_owned(),
+            repository_name: "Qwen2.5-VL-3B-Instruct-GGUF".to_owned(),
+            repository_owner: "ggml-org".to_owned(),
             revision: "main".to_owned(),
         },
     }
diff --git a/paddler_test_cluster_harness/src/model_card/qwen3_0_6b.rs b/paddler_test_cluster_harness/src/model_card/qwen3_0_6b.rs
index ea802f54d..e796e5538 100644
--- a/paddler_test_cluster_harness/src/model_card/qwen3_0_6b.rs
+++ b/paddler_test_cluster_harness/src/model_card/qwen3_0_6b.rs
@@ -1,4 +1,4 @@
-use paddler_messaging::huggingface_model_reference::HuggingFaceModelReference;
+use paddler_model_source::huggingface_model_reference::HuggingFaceModelReference;
 
 use crate::model_card::ModelCard;
 
@@ -7,7 +7,8 @@ pub fn qwen3_0_6b() -> ModelCard {
     ModelCard {
         reference: HuggingFaceModelReference {
             filename: "Qwen3-0.6B-Q8_0.gguf".to_owned(),
-            repo_id: "Qwen/Qwen3-0.6B-GGUF".to_owned(),
+            repository_name: "Qwen3-0.6B-GGUF".to_owned(),
+            repository_owner: "Qwen".to_owned(),
             revision: "main".to_owned(),
         },
     }
diff --git a/paddler_test_cluster_harness/src/model_card/qwen3_5_0_8b.rs b/paddler_test_cluster_harness/src/model_card/qwen3_5_0_8b.rs
index 9bbb2f300..c4704bd01 100644
--- a/paddler_test_cluster_harness/src/model_card/qwen3_5_0_8b.rs
+++ b/paddler_test_cluster_harness/src/model_card/qwen3_5_0_8b.rs
@@ -1,4 +1,4 @@
-use paddler_messaging::huggingface_model_reference::HuggingFaceModelReference;
+use paddler_model_source::huggingface_model_reference::HuggingFaceModelReference;
 
 use crate::model_card::ModelCard;
 
@@ -7,7 +7,8 @@ pub fn qwen3_5_0_8b() -> ModelCard {
     ModelCard {
         reference: HuggingFaceModelReference {
             filename: "Qwen3.5-0.8B-Q4_K_M.gguf".to_owned(),
-            repo_id: "unsloth/Qwen3.5-0.8B-GGUF".to_owned(),
+            repository_name: "Qwen3.5-0.8B-GGUF".to_owned(),
+            repository_owner: "unsloth".to_owned(),
             revision: "main".to_owned(),
         },
     }
diff --git a/paddler_test_cluster_harness/src/model_card/qwen3_5_0_8b_mmproj.rs b/paddler_test_cluster_harness/src/model_card/qwen3_5_0_8b_mmproj.rs
index 563fa96bb..d060dd959 100644
--- a/paddler_test_cluster_harness/src/model_card/qwen3_5_0_8b_mmproj.rs
+++ b/paddler_test_cluster_harness/src/model_card/qwen3_5_0_8b_mmproj.rs
@@ -1,4 +1,4 @@
-use paddler_messaging::huggingface_model_reference::HuggingFaceModelReference;
+use paddler_model_source::huggingface_model_reference::HuggingFaceModelReference;
 
 use crate::model_card::ModelCard;
 
@@ -7,7 +7,8 @@ pub fn qwen3_5_0_8b_mmproj() -> ModelCard {
     ModelCard {
         reference: HuggingFaceModelReference {
             filename: "mmproj-F16.gguf".to_owned(),
-            repo_id: "unsloth/Qwen3.5-0.8B-GGUF".to_owned(),
+            repository_name: "Qwen3.5-0.8B-GGUF".to_owned(),
+            repository_owner: "unsloth".to_owned(),
             revision: "main".to_owned(),
         },
     }
diff --git a/paddler_test_cluster_harness/src/model_card/qwen3_8_27b.rs b/paddler_test_cluster_harness/src/model_card/qwen3_8_27b.rs
new file mode 100644
index 000000000..8bed63a97
--- /dev/null
+++ b/paddler_test_cluster_harness/src/model_card/qwen3_8_27b.rs
@@ -0,0 +1,15 @@
+use paddler_model_source::huggingface_model_reference::HuggingFaceModelReference;
+
+use crate::model_card::ModelCard;
+
+#[must_use]
+pub fn qwen3_8_27b() -> ModelCard {
+    ModelCard {
+        reference: HuggingFaceModelReference {
+            filename: "Qwen3.8-27B-UD-IQ1_S.gguf".to_owned(),
+            repository_name: "Qwen3.8-27B-GGUF".to_owned(),
+            repository_owner: "unsloth".to_owned(),
+            revision: "4ca720788d1e01f1bff70c033e0d0028fd02e502".to_owned(),
+        },
+    }
+}
diff --git a/paddler_test_cluster_harness/src/model_card/smolvlm2_256m.rs b/paddler_test_cluster_harness/src/model_card/smolvlm2_256m.rs
index 08f99c982..5c8bd88bf 100644
--- a/paddler_test_cluster_harness/src/model_card/smolvlm2_256m.rs
+++ b/paddler_test_cluster_harness/src/model_card/smolvlm2_256m.rs
@@ -1,4 +1,4 @@
-use paddler_messaging::huggingface_model_reference::HuggingFaceModelReference;
+use paddler_model_source::huggingface_model_reference::HuggingFaceModelReference;
 
 use crate::model_card::ModelCard;
 
@@ -7,7 +7,8 @@ pub fn smolvlm2_256m() -> ModelCard {
     ModelCard {
         reference: HuggingFaceModelReference {
             filename: "SmolVLM2-256M-Video-Instruct-Q8_0.gguf".to_owned(),
-            repo_id: "ggml-org/SmolVLM2-256M-Video-Instruct-GGUF".to_owned(),
+            repository_name: "SmolVLM2-256M-Video-Instruct-GGUF".to_owned(),
+            repository_owner: "ggml-org".to_owned(),
             revision: "main".to_owned(),
         },
     }
diff --git a/paddler_test_cluster_harness/src/model_card/smolvlm2_256m_mmproj.rs b/paddler_test_cluster_harness/src/model_card/smolvlm2_256m_mmproj.rs
index f08c1bdc1..f47ade9c1 100644
--- a/paddler_test_cluster_harness/src/model_card/smolvlm2_256m_mmproj.rs
+++ b/paddler_test_cluster_harness/src/model_card/smolvlm2_256m_mmproj.rs
@@ -1,4 +1,4 @@
-use paddler_messaging::huggingface_model_reference::HuggingFaceModelReference;
+use paddler_model_source::huggingface_model_reference::HuggingFaceModelReference;
 
 use crate::model_card::ModelCard;
 
@@ -7,7 +7,8 @@ pub fn smolvlm2_256m_mmproj() -> ModelCard {
     ModelCard {
         reference: HuggingFaceModelReference {
             filename: "mmproj-SmolVLM2-256M-Video-Instruct-Q8_0.gguf".to_owned(),
-            repo_id: "ggml-org/SmolVLM2-256M-Video-Instruct-GGUF".to_owned(),
+            repository_name: "SmolVLM2-256M-Video-Instruct-GGUF".to_owned(),
+            repository_owner: "ggml-org".to_owned(),
             revision: "main".to_owned(),
         },
     }
diff --git a/paddler_test_cluster_harness/src/model_card/ultravox_v0_5_llama_3_2_1b.rs b/paddler_test_cluster_harness/src/model_card/ultravox_v0_5_llama_3_2_1b.rs
new file mode 100644
index 000000000..d97fe0ae6
--- /dev/null
+++ b/paddler_test_cluster_harness/src/model_card/ultravox_v0_5_llama_3_2_1b.rs
@@ -0,0 +1,15 @@
+use paddler_model_source::huggingface_model_reference::HuggingFaceModelReference;
+
+use crate::model_card::ModelCard;
+
+#[must_use]
+pub fn ultravox_v0_5_llama_3_2_1b() -> ModelCard {
+    ModelCard {
+        reference: HuggingFaceModelReference {
+            filename: "Llama-3.2-1B-Instruct-Q4_K_M.gguf".to_owned(),
+            repository_name: "ultravox-v0_5-llama-3_2-1b-GGUF".to_owned(),
+            repository_owner: "ggml-org".to_owned(),
+            revision: "main".to_owned(),
+        },
+    }
+}
diff --git a/paddler_test_cluster_harness/src/model_card/ultravox_v0_5_llama_3_2_1b_mmproj.rs b/paddler_test_cluster_harness/src/model_card/ultravox_v0_5_llama_3_2_1b_mmproj.rs
new file mode 100644
index 000000000..883ec5b4b
--- /dev/null
+++ b/paddler_test_cluster_harness/src/model_card/ultravox_v0_5_llama_3_2_1b_mmproj.rs
@@ -0,0 +1,15 @@
+use paddler_model_source::huggingface_model_reference::HuggingFaceModelReference;
+
+use crate::model_card::ModelCard;
+
+#[must_use]
+pub fn ultravox_v0_5_llama_3_2_1b_mmproj() -> ModelCard {
+    ModelCard {
+        reference: HuggingFaceModelReference {
+            filename: "mmproj-ultravox-v0_5-llama-3_2-1b-f16.gguf".to_owned(),
+            repository_name: "ultravox-v0_5-llama-3_2-1b-GGUF".to_owned(),
+            repository_owner: "ggml-org".to_owned(),
+            revision: "main".to_owned(),
+        },
+    }
+}
diff --git a/paddler_test_cluster_harness/src/noul_system_one_request.rs b/paddler_test_cluster_harness/src/noul_system_one_request.rs
new file mode 100644
index 000000000..d689e2c8b
--- /dev/null
+++ b/paddler_test_cluster_harness/src/noul_system_one_request.rs
@@ -0,0 +1,9 @@
+use serde_json::Value;
+use serde_json::json;
+
+use crate::noul_system_one_request_about::noul_system_one_request_about;
+
+#[must_use]
+pub fn noul_system_one_request() -> Value {
+    noul_system_one_request_about(&json!("state"))
+}
diff --git a/paddler_test_cluster_harness/src/noul_system_one_request_about.rs b/paddler_test_cluster_harness/src/noul_system_one_request_about.rs
new file mode 100644
index 000000000..a76ab636a
--- /dev/null
+++ b/paddler_test_cluster_harness/src/noul_system_one_request_about.rs
@@ -0,0 +1,9 @@
+use serde_json::Value;
+use serde_json::json;
+
+use crate::system_one_request::system_one_request;
+
+#[must_use]
+pub fn noul_system_one_request_about(state: &Value) -> Value {
+    system_one_request(state, &json!({"paid": {"type": "noul"}}))
+}
diff --git a/paddler_test_cluster_harness/src/openai_api_client.rs b/paddler_test_cluster_harness/src/openai_api_client.rs
index b11872d8a..ea63b845e 100644
--- a/paddler_test_cluster_harness/src/openai_api_client.rs
+++ b/paddler_test_cluster_harness/src/openai_api_client.rs
@@ -3,23 +3,27 @@ use async_openai::config::OpenAIConfig;
 use async_openai::error::OpenAIError;
 use futures_util::TryStreamExt as _;
 use serde_json::Value;
-use url::ParseError;
 use url::Url;
 
+const OPENAI_API_BASE_PATH: &str = "/v1";
+
 #[derive(Clone)]
 pub struct OpenAIApiClient {
     client: Client<OpenAIConfig>,
 }
 
 impl OpenAIApiClient {
-    pub fn new(openai_base_url: &Url) -> Result<Self, ParseError> {
-        Ok(Self {
+    #[must_use]
+    pub fn new(mut openai_base_url: Url) -> Self {
+        openai_base_url.set_path(OPENAI_API_BASE_PATH);
+
+        Self {
             client: Client::with_config(
                 OpenAIConfig::default()
-                    .with_api_base(openai_base_url.join("v1")?)
+                    .with_api_base(openai_base_url)
                     .with_api_key("paddler"),
             ),
-        })
+        }
     }
 
     pub async fn chat_completion_non_streaming(&self, body: &Value) -> Result<Value, OpenAIError> {
diff --git a/paddler_test_cluster_harness/src/pointer_head_fixture.rs b/paddler_test_cluster_harness/src/pointer_head_fixture.rs
new file mode 100644
index 000000000..a5983ad55
--- /dev/null
+++ b/paddler_test_cluster_harness/src/pointer_head_fixture.rs
@@ -0,0 +1,11 @@
+use paddler_messaging::agent_desired_model::AgentDesiredModel;
+use paddler_model_source::model_source_error::ModelSourceError;
+
+use crate::agent_desired_model_at::agent_desired_model_at;
+use crate::fixture_path::fixture_path;
+
+pub fn pointer_head_fixture(
+    fixture_file_name: &str,
+) -> Result<AgentDesiredModel, ModelSourceError> {
+    agent_desired_model_at(&fixture_path(fixture_file_name))
+}
diff --git a/paddler_test_cluster_harness/src/process_end.rs b/paddler_test_cluster_harness/src/process_end.rs
new file mode 100644
index 000000000..26fcfa61a
--- /dev/null
+++ b/paddler_test_cluster_harness/src/process_end.rs
@@ -0,0 +1,23 @@
+use anyhow::Error;
+
+#[derive(Debug)]
+pub enum ProcessEnd {
+    Clean,
+    Failed(Error),
+}
+
+impl ProcessEnd {
+    pub fn failed<TFailure>(failure: TFailure) -> Self
+    where
+        TFailure: Into<Error>,
+    {
+        Self::Failed(failure.into())
+    }
+
+    pub fn into_result(self) -> Result<(), Error> {
+        match self {
+            Self::Clean => Ok(()),
+            Self::Failed(exit_failure) => Err(exit_failure),
+        }
+    }
+}
diff --git a/paddler_test_cluster_harness/src/raw_agent_socket.rs b/paddler_test_cluster_harness/src/raw_agent_socket.rs
deleted file mode 100644
index fd6ec42ef..000000000
--- a/paddler_test_cluster_harness/src/raw_agent_socket.rs
+++ /dev/null
@@ -1,126 +0,0 @@
-use std::net::SocketAddr;
-
-use futures_util::SinkExt as _;
-use futures_util::StreamExt as _;
-use serde_json::from_str;
-use serde_json::to_string;
-use tokio::net::TcpStream;
-use tokio_tungstenite::MaybeTlsStream;
-use tokio_tungstenite::WebSocketStream;
-use tokio_tungstenite::connect_async;
-use tokio_tungstenite::tungstenite::Error as WebSocketError;
-use tokio_tungstenite::tungstenite::Message;
-use tokio_tungstenite::tungstenite::protocol::CloseFrame;
-
-use paddler_messaging::api_path::ApiPath;
-use paddler_messaging::jsonrpc::request_envelope::RequestEnvelope;
-use paddler_messaging::jsonrpc::response_envelope::ResponseEnvelope;
-use paddler_messaging::management_socket::agent::message::Message as AgentJsonRpcMessage;
-use paddler_messaging::management_socket::agent::notification::Notification as AgentJsonRpcNotification;
-use paddler_messaging::management_socket::agent::request::Request as AgentJsonRpcRequest;
-use paddler_messaging::management_socket::agent::response::Response as AgentJsonRpcResponse;
-use paddler_messaging::management_socket::balancer::message::Message as ManagementJsonRpcMessage;
-use paddler_messaging::management_socket::balancer::notification::Notification as ManagementJsonRpcNotification;
-use paddler_messaging::management_socket::balancer::notification_params::register_agent_params::RegisterAgentParams;
-use paddler_messaging::slot_aggregated_status_snapshot::SlotAggregatedStatusSnapshot;
-
-use crate::cluster_harness_error::ClusterHarnessError;
-
-pub struct RawAgentSocket {
-    websocket: WebSocketStream<MaybeTlsStream<TcpStream>>,
-}
-
-impl RawAgentSocket {
-    pub async fn connect(
-        management_addr: SocketAddr,
-        agent_id: &str,
-    ) -> Result<Self, WebSocketError> {
-        connect_async(format!(
-            "ws://{management_addr}{}",
-            ApiPath::agent_socket(agent_id)
-        ))
-        .await
-        .map(|(websocket, _handshake_response)| Self { websocket })
-    }
-
-    async fn next_message(&mut self) -> Result<AgentJsonRpcMessage, ClusterHarnessError> {
-        while let Some(frame) = self.websocket.next().await {
-            if let Message::Text(text) =
-                frame.map_err(ClusterHarnessError::AgentSocketReceiveFailed)?
-            {
-                return from_str(&text)
-                    .map_err(ClusterHarnessError::AgentSocketMessageUndeserializable);
-            }
-        }
-
-        Err(ClusterHarnessError::AgentSocketEndedWithoutMessage)
-    }
-
-    pub async fn next_close_frame(&mut self) -> Result<Option<CloseFrame>, ClusterHarnessError> {
-        while let Some(frame) = self.websocket.next().await {
-            if let Message::Close(close_frame) =
-                frame.map_err(ClusterHarnessError::AgentSocketReceiveFailed)?
-            {
-                return Ok(close_frame);
-            }
-        }
-
-        Err(ClusterHarnessError::AgentSocketEndedWithoutClosing)
-    }
-
-    pub async fn next_request(
-        &mut self,
-    ) -> Result<RequestEnvelope<AgentJsonRpcRequest>, ClusterHarnessError> {
-        loop {
-            if let AgentJsonRpcMessage::Request(request_envelope) = self.next_message().await? {
-                return Ok(request_envelope);
-            }
-        }
-    }
-
-    pub async fn register(&mut self) -> Result<(), ClusterHarnessError> {
-        self.send_notification(ManagementJsonRpcNotification::RegisterAgent(
-            RegisterAgentParams {
-                name: None,
-                slot_aggregated_status_snapshot: SlotAggregatedStatusSnapshot::default(),
-            },
-        ))
-        .await?;
-
-        loop {
-            if let AgentJsonRpcMessage::Notification(AgentJsonRpcNotification::SetState(_)) =
-                self.next_message().await?
-            {
-                return Ok(());
-            }
-        }
-    }
-
-    pub async fn send(&mut self, message: Message) -> Result<(), ClusterHarnessError> {
-        self.websocket
-            .send(message)
-            .await
-            .map_err(ClusterHarnessError::AgentSocketSendFailed)
-    }
-
-    pub async fn send_notification(
-        &mut self,
-        notification: ManagementJsonRpcNotification,
-    ) -> Result<(), ClusterHarnessError> {
-        let serialized_notification =
-            to_string(&ManagementJsonRpcMessage::Notification(notification))
-                .map_err(ClusterHarnessError::AgentSocketNotificationUnserializable)?;
-
-        self.send(Message::text(serialized_notification)).await
-    }
-
-    pub async fn send_response(
-        &mut self,
-        response_envelope: ResponseEnvelope<AgentJsonRpcResponse>,
-    ) -> Result<(), ClusterHarnessError> {
-        let serialized_response = to_string(&ManagementJsonRpcMessage::Response(response_envelope))
-            .map_err(ClusterHarnessError::AgentSocketResponseUnserializable)?;
-
-        self.send(Message::text(serialized_response)).await
-    }
-}
diff --git a/paddler_test_cluster_harness/src/raw_inference_socket.rs b/paddler_test_cluster_harness/src/raw_inference_socket.rs
index 9eabbf885..f22203e9d 100644
--- a/paddler_test_cluster_harness/src/raw_inference_socket.rs
+++ b/paddler_test_cluster_harness/src/raw_inference_socket.rs
@@ -3,6 +3,8 @@ use std::num::NonZeroU32;
 
 use futures_util::SinkExt as _;
 use futures_util::StreamExt as _;
+use futures_util::TryFutureExt as _;
+use futures_util::future::ready;
 use serde_json::from_str;
 use serde_json::to_string;
 use tokio::net::TcpStream;
@@ -25,13 +27,32 @@ use paddler_messaging::request_params::continue_from_raw_prompt_params::Continue
 
 use crate::cluster_harness_error::ClusterHarnessError;
 
-const fn answered_request_id(message: &InferenceClientMessage) -> Option<&str> {
+const fn answered_request_id(message: &InferenceClientMessage) -> &str {
     match message {
         InferenceClientMessage::Error(ErrorEnvelope { request_id, .. })
         | InferenceClientMessage::Response(ResponseEnvelope { request_id, .. }) => {
-            Some(request_id.as_str())
+            request_id.as_str()
         }
-        InferenceClientMessage::Notification(_) => None,
+    }
+}
+
+fn awaited_answer(
+    frame: Message,
+    is_awaited: impl Fn(&str) -> bool,
+) -> Option<Result<InferenceClientMessage, ClusterHarnessError>> {
+    match frame {
+        Message::Text(text) => from_str::<InferenceClientMessage>(&text)
+            .map_err(ClusterHarnessError::InferenceSocketMessageUnreadable)
+            .map(|message| is_awaited(answered_request_id(&message)).then_some(message))
+            .transpose(),
+        _other_frame => None,
+    }
+}
+
+fn pong_payload(frame: Message) -> Option<Result<Bytes, ClusterHarnessError>> {
+    match frame {
+        Message::Pong(payload) => Some(Ok(payload)),
+        _other_frame => None,
     }
 }
 
@@ -50,52 +71,47 @@ impl RawInferenceSocket {
     }
 
     pub async fn next_answer(&mut self) -> Result<InferenceClientMessage, ClusterHarnessError> {
-        while let Some(message) = self.next_message().await? {
-            if answered_request_id(&message).is_some() {
-                return Ok(message);
-            }
-        }
-
-        Err(ClusterHarnessError::InferenceSocketClosedBeforeAnyAnswer)
+        self.next_selected_frame(|frame| awaited_answer(frame, |_answered_request_id| true))
+            .await
+            .and_then(|answer| {
+                answer.ok_or(ClusterHarnessError::InferenceSocketClosedBeforeAnyAnswer)
+            })
     }
 
     pub async fn next_answer_to(
         &mut self,
         request_id: &str,
     ) -> Result<InferenceClientMessage, ClusterHarnessError> {
-        while let Some(message) = self.next_message().await? {
-            if answered_request_id(&message) == Some(request_id) {
-                return Ok(message);
-            }
-        }
-
-        Err(ClusterHarnessError::InferenceSocketClosedBeforeAnswer {
-            request_id: request_id.to_owned(),
+        self.next_selected_frame(|frame| {
+            awaited_answer(frame, |answered_request_id| {
+                answered_request_id == request_id
+            })
+        })
+        .await
+        .and_then(|answer| {
+            answer.ok_or_else(|| ClusterHarnessError::InferenceSocketClosedBeforeAnswer {
+                request_id: request_id.to_owned(),
+            })
         })
     }
 
-    async fn next_message(
-        &mut self,
-    ) -> Result<Option<InferenceClientMessage>, ClusterHarnessError> {
-        while let Some(frame) = self.next_frame_before_close().await? {
-            if let Message::Text(text) = frame {
-                return from_str::<InferenceClientMessage>(&text)
-                    .map(Some)
-                    .map_err(ClusterHarnessError::InferenceSocketMessageUnreadable);
-            }
-        }
-
-        Ok(None)
+    pub async fn next_pong(&mut self) -> Result<Bytes, ClusterHarnessError> {
+        self.next_selected_frame(pong_payload)
+            .await
+            .and_then(|pong| pong.ok_or(ClusterHarnessError::InferenceSocketClosedBeforePong))
     }
 
-    pub async fn next_pong(&mut self) -> Result<Bytes, ClusterHarnessError> {
+    async fn next_selected_frame<TSelected>(
+        &mut self,
+        select: impl Fn(Message) -> Option<Result<TSelected, ClusterHarnessError>>,
+    ) -> Result<Option<TSelected>, ClusterHarnessError> {
         while let Some(frame) = self.next_frame_before_close().await? {
-            if let Message::Pong(payload) = frame {
-                return Ok(payload);
+            if let Some(selected) = select(frame) {
+                return selected.map(Some);
             }
         }
 
-        Err(ClusterHarnessError::InferenceSocketClosedBeforePong)
+        Ok(None)
     }
 
     async fn next_frame_before_close(&mut self) -> Result<Option<Message>, ClusterHarnessError> {
@@ -128,11 +144,15 @@ impl RawInferenceSocket {
                 ),
             });
 
-        let serialized_request = to_string(&request)
-            .map_err(ClusterHarnessError::InferenceSocketRequestUnserializable)?;
-
-        self.send(Message::text(serialized_request))
-            .await
-            .map_err(ClusterHarnessError::InferenceSocketSendFailed)
+        ready(
+            to_string(&request)
+                .map(Message::text)
+                .map_err(ClusterHarnessError::InferenceSocketRequestUnserializable),
+        )
+        .and_then(|serialized_request| {
+            self.send(serialized_request)
+                .map_err(ClusterHarnessError::InferenceSocketSendFailed)
+        })
+        .await
     }
 }
diff --git a/paddler_test_cluster_harness/src/resource_snapshot.rs b/paddler_test_cluster_harness/src/resource_snapshot.rs
index 26ed7a70d..8615089a7 100644
--- a/paddler_test_cluster_harness/src/resource_snapshot.rs
+++ b/paddler_test_cluster_harness/src/resource_snapshot.rs
@@ -21,13 +21,12 @@ pub struct ResourceSnapshot {
 
 impl ResourceSnapshot {
     pub fn try_from_self() -> Result<Self, ClusterHarnessError> {
-        let open_file_descriptors = read_dir(open_descriptors_directory_path())
+        read_dir(open_descriptors_directory_path())
             .and_then(Iterator::collect::<Result<Vec<DirEntry>, io::Error>>)
-            .map_err(ClusterHarnessError::OpenFileDescriptorsUnreadable)?;
-
-        Ok(Self {
-            open_file_descriptor_count: open_file_descriptors.len(),
-        })
+            .map(|open_file_descriptors| Self {
+                open_file_descriptor_count: open_file_descriptors.len(),
+            })
+            .map_err(ClusterHarnessError::OpenFileDescriptorsUnreadable)
     }
 
     #[must_use]
diff --git a/paddler_test_cluster_harness/src/running_agent.rs b/paddler_test_cluster_harness/src/running_agent.rs
index 847a1b9ac..1f70ea743 100644
--- a/paddler_test_cluster_harness/src/running_agent.rs
+++ b/paddler_test_cluster_harness/src/running_agent.rs
@@ -1,7 +1,14 @@
 use anyhow::Result;
+use tokio::select;
+
+use paddler_messaging::agent_controller_pool_snapshot::AgentControllerPoolSnapshot;
+use paddler_messaging::agent_controller_snapshot::AgentControllerSnapshot;
 
 use crate::agent_config::AgentConfig;
+use crate::agent_readiness::AgentReadiness;
+use crate::cluster_harness_error::ClusterHarnessError;
 use crate::managed_process::ManagedProcess;
+use crate::snapshots_watcher::SnapshotsWatcher;
 
 pub struct RunningAgent {
     pub config: AgentConfig,
@@ -14,6 +21,20 @@ impl RunningAgent {
         Self { config, process }
     }
 
+    pub async fn wait_until_ready(
+        &mut self,
+        agents_watcher: &mut SnapshotsWatcher<AgentControllerPoolSnapshot>,
+        readiness: AgentReadiness,
+    ) -> Result<AgentControllerSnapshot, ClusterHarnessError> {
+        select! {
+            ready_agent = agents_watcher.wait_for_agent(&self.config.name, &readiness) => ready_agent,
+            process_end = self.process.exited() => Err(ClusterHarnessError::AgentExitedBeforeReady {
+                agent_name: self.config.name.clone(),
+                process_end,
+            }),
+        }
+    }
+
     pub async fn shutdown(self) -> Result<()> {
         self.process.shutdown().await
     }
diff --git a/paddler_test_cluster_harness/src/running_balancer.rs b/paddler_test_cluster_harness/src/running_balancer.rs
index 9bf286489..05520a0ea 100644
--- a/paddler_test_cluster_harness/src/running_balancer.rs
+++ b/paddler_test_cluster_harness/src/running_balancer.rs
@@ -1,6 +1,7 @@
 use std::net::SocketAddr;
 
 use anyhow::Result;
+use url::ParseError;
 use url::Url;
 
 use paddler_balancer::balancer_addresses::BalancerAddresses;
@@ -8,20 +9,34 @@ use paddler_balancer::balancer_addresses::BalancerAddresses;
 use crate::cluster_harness_error::ClusterHarnessError;
 use crate::managed_process::ManagedProcess;
 
-fn base_url_for(addr: SocketAddr) -> Result<Url, ClusterHarnessError> {
+fn base_url_for(addr: SocketAddr) -> Result<Url, ParseError> {
     Url::parse(&format!("http://{addr}/"))
-        .map_err(|source| ClusterHarnessError::BaseUrlInvalid { addr, source })
 }
 
 pub struct RunningBalancer {
     pub addresses: BalancerAddresses,
+    pub inference_base_url: Url,
+    pub management_base_url: Url,
     process: Box<dyn ManagedProcess>,
 }
 
 impl RunningBalancer {
-    #[must_use]
-    pub const fn new(addresses: BalancerAddresses, process: Box<dyn ManagedProcess>) -> Self {
-        Self { addresses, process }
+    pub fn new(
+        addresses: BalancerAddresses,
+        process: Box<dyn ManagedProcess>,
+    ) -> Result<Self, ClusterHarnessError> {
+        base_url_for(addresses.inference)
+            .map_err(ClusterHarnessError::InferenceBaseUrlInvalid)
+            .and_then(|inference_base_url| {
+                base_url_for(addresses.management)
+                    .map_err(ClusterHarnessError::ManagementBaseUrlInvalid)
+                    .map(|management_base_url| Self {
+                        addresses,
+                        inference_base_url,
+                        management_base_url,
+                        process,
+                    })
+            })
     }
 
     pub fn compat_openai_addr(&self) -> Result<SocketAddr, ClusterHarnessError> {
@@ -31,15 +46,37 @@ impl RunningBalancer {
     }
 
     pub fn compat_openai_base_url(&self) -> Result<Url, ClusterHarnessError> {
-        self.compat_openai_addr().and_then(base_url_for)
+        self.compat_openai_addr().and_then(|addr| {
+            base_url_for(addr).map_err(ClusterHarnessError::CompatOpenAIBaseUrlInvalid)
+        })
     }
 
-    pub fn inference_base_url(&self) -> Result<Url, ClusterHarnessError> {
-        base_url_for(self.addresses.inference)
+    pub fn compat_openai_url(&self, path: &str) -> Result<Url, ClusterHarnessError> {
+        let mut compat_openai_url = self.compat_openai_base_url()?;
+
+        compat_openai_url.set_path(path);
+
+        Ok(compat_openai_url)
+    }
+
+    pub fn compat_typesafe_addr(&self) -> Result<SocketAddr, ClusterHarnessError> {
+        self.addresses
+            .compat_typesafe
+            .ok_or(ClusterHarnessError::CompatTypeSafeServiceNotServed)
     }
 
-    pub fn management_base_url(&self) -> Result<Url, ClusterHarnessError> {
-        base_url_for(self.addresses.management)
+    pub fn compat_typesafe_base_url(&self) -> Result<Url, ClusterHarnessError> {
+        self.compat_typesafe_addr().and_then(|addr| {
+            base_url_for(addr).map_err(ClusterHarnessError::CompatTypeSafeBaseUrlInvalid)
+        })
+    }
+
+    pub fn compat_typesafe_url(&self, path: &str) -> Result<Url, ClusterHarnessError> {
+        let mut compat_typesafe_url = self.compat_typesafe_base_url()?;
+
+        compat_typesafe_url.set_path(path);
+
+        Ok(compat_typesafe_url)
     }
 
     pub async fn shutdown(self) -> Result<()> {
diff --git a/paddler_test_cluster_harness/src/snapshots_watcher.rs b/paddler_test_cluster_harness/src/snapshots_watcher.rs
index a470962a8..6e4dc400a 100644
--- a/paddler_test_cluster_harness/src/snapshots_watcher.rs
+++ b/paddler_test_cluster_harness/src/snapshots_watcher.rs
@@ -8,8 +8,10 @@ use paddler_client::agents_stream::AgentsStream;
 use paddler_client::buffered_requests_stream::BufferedRequestsStream;
 use paddler_client::error::Result as ClientResult;
 use paddler_messaging::agent_controller_pool_snapshot::AgentControllerPoolSnapshot;
+use paddler_messaging::agent_controller_snapshot::AgentControllerSnapshot;
 use paddler_messaging::buffered_request_manager_snapshot::BufferedRequestManagerSnapshot;
 
+use crate::agent_readiness::AgentReadiness;
 use crate::cluster_harness_error::ClusterHarnessError;
 use crate::snapshots_stream::SnapshotsStream;
 
@@ -70,10 +72,21 @@ impl<TSnapshot: Clone> SnapshotsWatcher<TSnapshot> {
     ) -> Result<TSnapshot, ClusterHarnessError>
     where
         TVerdict: FnMut(&TSnapshot) -> Result<bool, ClusterHarnessError>,
+    {
+        self.until_found(|snapshot| Ok(verdict(snapshot)?.then(|| snapshot.clone())))
+            .await
+    }
+
+    async fn until_found<TFound, TFinder>(
+        &mut self,
+        mut finder: TFinder,
+    ) -> Result<TFound, ClusterHarnessError>
+    where
+        TFinder: FnMut(&TSnapshot) -> Result<Option<TFound>, ClusterHarnessError>,
     {
         loop {
-            if verdict(&self.newest_snapshot)? {
-                return Ok(self.newest_snapshot.clone());
+            if let Some(found) = finder(&self.newest_snapshot)? {
+                return Ok(found);
             }
 
             self.newest_snapshot =
@@ -111,27 +124,21 @@ impl SnapshotsWatcher<AgentControllerPoolSnapshot> {
         .await
     }
 
-    pub async fn wait_for_agent_ready(
+    pub async fn wait_for_agent(
         &mut self,
         agent_name: &str,
-        expected_slot_count: u16,
-    ) -> Result<AgentControllerPoolSnapshot, ClusterHarnessError> {
-        self.until_verdict(|snapshot| {
+        readiness: &AgentReadiness,
+    ) -> Result<AgentControllerSnapshot, ClusterHarnessError> {
+        self.until_found(|snapshot| {
             match snapshot
                 .agents
                 .iter()
                 .find(|registered_agent| registered_agent.name.as_deref() == Some(agent_name))
             {
-                Some(registered_agent) if !registered_agent.status.issues.is_empty() => {
-                    Err(ClusterHarnessError::AgentReportedIssues {
-                        agent_name: agent_name.to_owned(),
-                        issues: registered_agent.status.issues.clone(),
-                    })
-                }
-                Some(registered_agent) => {
-                    Ok(registered_agent.status.slots_total == u64::from(expected_slot_count))
-                }
-                None => Ok(false),
+                Some(registered_agent) => Ok(readiness
+                    .is_met_by(agent_name, registered_agent)?
+                    .then(|| registered_agent.clone())),
+                None => Ok(None),
             }
         })
         .await
@@ -148,11 +155,16 @@ impl SnapshotsWatcher<BufferedRequestManagerSnapshot> {
 
 #[cfg(test)]
 mod tests {
+    use std::mem::discriminant;
+
     use futures_util::stream::iter;
 
+    use paddler_client::error::Error as ClientError;
     use paddler_messaging::buffered_request_manager_snapshot::BufferedRequestManagerSnapshot;
 
     use super::SnapshotsWatcher;
+    use crate::cluster_harness_error::ClusterHarnessError;
+    use crate::snapshots_stream::SnapshotsStream;
 
     fn snapshot(buffered_requests_current: u64) -> BufferedRequestManagerSnapshot {
         BufferedRequestManagerSnapshot {
@@ -181,4 +193,26 @@ mod tests {
 
         assert_eq!(satisfied_by.buffered_requests_current, 2);
     }
+
+    #[tokio::test]
+    async fn a_failed_snapshot_is_reported_as_a_stream_failure() {
+        let connection_error = SnapshotsWatcher::of_buffered_requests(Box::pin(iter(vec![Err(
+            ClientError::NonUtf8StreamLine {
+                source: String::from_utf8(vec![0xFF]).unwrap_err(),
+            },
+        )])))
+        .await
+        .err()
+        .expect("a failed snapshot must not start the watcher");
+
+        assert_eq!(
+            discriminant(&connection_error),
+            discriminant(&ClusterHarnessError::SnapshotsStreamFailed {
+                snapshots_stream: SnapshotsStream::BufferedRequests,
+                source: ClientError::NonUtf8StreamLine {
+                    source: String::from_utf8(vec![0xFF]).unwrap_err(),
+                },
+            })
+        );
+    }
 }
diff --git a/paddler_test_cluster_harness/src/state_database_file.rs b/paddler_test_cluster_harness/src/state_database_file.rs
deleted file mode 100644
index e31b626f1..000000000
--- a/paddler_test_cluster_harness/src/state_database_file.rs
+++ /dev/null
@@ -1,41 +0,0 @@
-use std::path::PathBuf;
-
-use anyhow::Context as _;
-use anyhow::Result;
-use tempfile::TempDir;
-
-pub struct StateDatabaseFile {
-    pub path: PathBuf,
-    pub url: String,
-    _directory_guard: TempDir,
-}
-
-impl StateDatabaseFile {
-    pub fn new() -> Result<Self> {
-        let directory = TempDir::new().context("failed to create temp state database directory")?;
-        let path = directory.path().join("state.json");
-        let url = format!(
-            "file://{}",
-            path.to_str()
-                .context("temp state database file path is not valid UTF-8")?
-        );
-
-        Ok(Self {
-            path,
-            url,
-            _directory_guard: directory,
-        })
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use super::StateDatabaseFile;
-
-    #[test]
-    fn names_its_url_after_its_file_path() {
-        let database = StateDatabaseFile::new().unwrap();
-
-        assert_eq!(database.url, format!("file://{}", database.path.display()));
-    }
-}
diff --git a/paddler_test_cluster_harness/src/synthetic_pointer_head_fixture.rs b/paddler_test_cluster_harness/src/synthetic_pointer_head_fixture.rs
new file mode 100644
index 000000000..a5f3806b2
--- /dev/null
+++ b/paddler_test_cluster_harness/src/synthetic_pointer_head_fixture.rs
@@ -0,0 +1,5 @@
+pub const QWEN3_5_0_8B_SYNTHETIC_POINTER_HEAD_FIXTURE: &str =
+    "qwen3_5_0_8b_synthetic_pointer_head.gguf";
+
+pub const QWEN3_8_27B_SYNTHETIC_POINTER_HEAD_FIXTURE: &str =
+    "qwen3_8_27b_synthetic_pointer_head.gguf";
diff --git a/paddler_test_cluster_harness/src/system_one_request.rs b/paddler_test_cluster_harness/src/system_one_request.rs
new file mode 100644
index 000000000..b17eb8c83
--- /dev/null
+++ b/paddler_test_cluster_harness/src/system_one_request.rs
@@ -0,0 +1,9 @@
+use serde_json::Value;
+use serde_json::json;
+
+use crate::typesafe_request_model::TYPESAFE_REQUEST_MODEL;
+
+#[must_use]
+pub fn system_one_request(state: &Value, questions: &Value) -> Value {
+    json!({"model": TYPESAFE_REQUEST_MODEL, "state": state, "questions": questions})
+}
diff --git a/paddler_test_cluster_harness/src/temporary_state_database_file.rs b/paddler_test_cluster_harness/src/temporary_state_database_file.rs
new file mode 100644
index 000000000..e60e99681
--- /dev/null
+++ b/paddler_test_cluster_harness/src/temporary_state_database_file.rs
@@ -0,0 +1,44 @@
+use std::path::PathBuf;
+
+use anyhow::Context as _;
+use anyhow::Result;
+use tempfile::TempDir;
+
+pub struct TemporaryStateDatabaseFile {
+    pub path: PathBuf,
+    pub url: String,
+    _directory_guard: TempDir,
+}
+
+impl TemporaryStateDatabaseFile {
+    pub fn new() -> Result<Self> {
+        TempDir::new()
+            .context("failed to create temp state database directory")
+            .and_then(Self::in_directory)
+    }
+
+    fn in_directory(directory: TempDir) -> Result<Self> {
+        let path = directory.path().join("state.json");
+
+        path.to_str()
+            .map(|utf8_path| format!("file://{utf8_path}"))
+            .context("temp state database file path is not valid UTF-8")
+            .map(|url| Self {
+                path,
+                url,
+                _directory_guard: directory,
+            })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::TemporaryStateDatabaseFile;
+
+    #[test]
+    fn names_its_url_after_its_file_path() {
+        let database = TemporaryStateDatabaseFile::new().unwrap();
+
+        assert_eq!(database.url, format!("file://{}", database.path.display()));
+    }
+}
diff --git a/paddler_test_cluster_harness/src/typesafe_api_client.rs b/paddler_test_cluster_harness/src/typesafe_api_client.rs
new file mode 100644
index 000000000..3a4356fac
--- /dev/null
+++ b/paddler_test_cluster_harness/src/typesafe_api_client.rs
@@ -0,0 +1,68 @@
+use futures_util::TryFutureExt as _;
+use reqwest::Client;
+use reqwest::Error as ReqwestError;
+use reqwest::RequestBuilder;
+use serde_json::Value;
+use url::Url;
+
+use paddler_balancer::compatibility::typesafe_service::typesafe_api_path::TypeSafeApiPath;
+use paddler_balancer::compatibility::typesafe_service::typesafe_header::TypeSafeHeader;
+
+use crate::typesafe_api_response::TypeSafeApiResponse;
+
+fn answered(
+    request: RequestBuilder,
+) -> impl Future<Output = Result<TypeSafeApiResponse, ReqwestError>> {
+    request.send().and_then(|response| {
+        let headers = response.headers().clone();
+        let status = response.status();
+
+        response.json().map_ok(move |body| TypeSafeApiResponse {
+            body,
+            headers,
+            status,
+        })
+    })
+}
+
+#[derive(Clone)]
+pub struct TypeSafeApiClient {
+    client: Client,
+    typesafe_base_url: Url,
+}
+
+impl TypeSafeApiClient {
+    #[must_use]
+    pub fn new(typesafe_base_url: Url) -> Self {
+        Self {
+            client: Client::new(),
+            typesafe_base_url,
+        }
+    }
+
+    pub async fn models(&self) -> Result<TypeSafeApiResponse, ReqwestError> {
+        answered(self.client.get(self.endpoint(TypeSafeApiPath::MODELS))).await
+    }
+
+    pub async fn system_one(
+        &self,
+        request_id: &str,
+        body: &Value,
+    ) -> Result<TypeSafeApiResponse, ReqwestError> {
+        answered(
+            self.client
+                .post(self.endpoint(TypeSafeApiPath::SYSTEM_ONE))
+                .header(TypeSafeHeader::REQUEST_ID, request_id)
+                .json(body),
+        )
+        .await
+    }
+
+    fn endpoint(&self, path: &str) -> Url {
+        let mut endpoint = self.typesafe_base_url.clone();
+
+        endpoint.set_path(path);
+
+        endpoint
+    }
+}
diff --git a/paddler_test_cluster_harness/src/typesafe_api_response.rs b/paddler_test_cluster_harness/src/typesafe_api_response.rs
new file mode 100644
index 000000000..a51922eaa
--- /dev/null
+++ b/paddler_test_cluster_harness/src/typesafe_api_response.rs
@@ -0,0 +1,9 @@
+use reqwest::StatusCode;
+use reqwest::header::HeaderMap;
+use serde_json::Value;
+
+pub struct TypeSafeApiResponse {
+    pub body: Value,
+    pub headers: HeaderMap,
+    pub status: StatusCode,
+}
diff --git a/paddler_test_cluster_harness/src/typesafe_request_model.rs b/paddler_test_cluster_harness/src/typesafe_request_model.rs
new file mode 100644
index 000000000..3644fffe9
--- /dev/null
+++ b/paddler_test_cluster_harness/src/typesafe_request_model.rs
@@ -0,0 +1 @@
+pub const TYPESAFE_REQUEST_MODEL: &str = "test";
diff --git a/paddler_test_cluster_harness/src/untrusted_agent_socket_client.rs b/paddler_test_cluster_harness/src/untrusted_agent_socket_client.rs
new file mode 100644
index 000000000..09ebc3c78
--- /dev/null
+++ b/paddler_test_cluster_harness/src/untrusted_agent_socket_client.rs
@@ -0,0 +1,69 @@
+use std::net::SocketAddr;
+
+use futures_util::SinkExt as _;
+use futures_util::StreamExt as _;
+use tokio::net::TcpStream;
+use tokio_tungstenite::MaybeTlsStream;
+use tokio_tungstenite::WebSocketStream;
+use tokio_tungstenite::connect_async;
+use tokio_tungstenite::tungstenite::Error as WebSocketError;
+use tokio_tungstenite::tungstenite::Message;
+use tokio_tungstenite::tungstenite::Utf8Bytes;
+use tokio_tungstenite::tungstenite::protocol::CloseFrame;
+
+use paddler_messaging::api_path::ApiPath;
+
+use crate::cluster_harness_error::ClusterHarnessError;
+
+pub struct UntrustedAgentSocketClient {
+    websocket: WebSocketStream<MaybeTlsStream<TcpStream>>,
+}
+
+impl UntrustedAgentSocketClient {
+    pub async fn connect(
+        management_addr: SocketAddr,
+        agent_id: &str,
+    ) -> Result<Self, WebSocketError> {
+        connect_async(format!(
+            "ws://{management_addr}{}",
+            ApiPath::agent_socket(agent_id)
+        ))
+        .await
+        .map(|(websocket, _handshake_response)| Self { websocket })
+    }
+
+    pub async fn next_close_frame(&mut self) -> Result<Option<CloseFrame>, ClusterHarnessError> {
+        while let Some(frame) = self.websocket.next().await {
+            if let Message::Close(close_frame) =
+                frame.map_err(ClusterHarnessError::AgentSocketReceiveFailed)?
+            {
+                return Ok(close_frame);
+            }
+        }
+
+        Err(ClusterHarnessError::AgentSocketEndedWithoutClosing)
+    }
+
+    pub async fn next_text(&mut self) -> Result<Utf8Bytes, ClusterHarnessError> {
+        while let Some(frame) = self.websocket.next().await {
+            let message = frame.map_err(ClusterHarnessError::AgentSocketReceiveFailed)?;
+
+            if let Message::Close(close_frame) = message {
+                return Err(ClusterHarnessError::AgentSocketClosedBeforeText { close_frame });
+            }
+
+            if let Message::Text(text) = message {
+                return Ok(text);
+            }
+        }
+
+        Err(ClusterHarnessError::AgentSocketEndedWithoutText)
+    }
+
+    pub async fn send(&mut self, message: Message) -> Result<(), ClusterHarnessError> {
+        self.websocket
+            .send(message)
+            .await
+            .map_err(ClusterHarnessError::AgentSocketSendFailed)
+    }
+}
diff --git a/paddler_test_cluster_harness/src/untrusted_balancer_socket.rs b/paddler_test_cluster_harness/src/untrusted_balancer_socket.rs
new file mode 100644
index 000000000..124a2bb69
--- /dev/null
+++ b/paddler_test_cluster_harness/src/untrusted_balancer_socket.rs
@@ -0,0 +1,94 @@
+use futures_util::SinkExt as _;
+use futures_util::StreamExt as _;
+use futures_util::future::ready;
+use serde_json::from_str;
+use tokio::net::TcpStream;
+use tokio_tungstenite::WebSocketStream;
+use tokio_tungstenite::tungstenite::Message;
+
+use paddler_messaging::agent_runtime_status::AgentRuntimeStatus;
+use paddler_messaging::jsonrpc::response_envelope::ResponseEnvelope;
+use paddler_messaging::management_socket::agent::response::Response as AgentJsonRpcResponse;
+use paddler_messaging::management_socket::balancer::message::Message as ManagementJsonRpcMessage;
+use paddler_messaging::management_socket::balancer::notification::Notification as ManagementJsonRpcNotification;
+
+use crate::cluster_harness_error::ClusterHarnessError;
+
+pub struct UntrustedBalancerSocket {
+    websocket: WebSocketStream<TcpStream>,
+}
+
+impl UntrustedBalancerSocket {
+    #[must_use]
+    pub const fn new(websocket: WebSocketStream<TcpStream>) -> Self {
+        Self { websocket }
+    }
+
+    pub async fn next_agent_response(
+        &mut self,
+    ) -> Result<ResponseEnvelope<AgentJsonRpcResponse>, ClusterHarnessError> {
+        self.websocket
+            .by_ref()
+            .filter_map(|frame| {
+                ready(
+                    frame
+                        .ok()
+                        .and_then(|frame| frame.into_text().ok())
+                        .and_then(|text| from_str::<ManagementJsonRpcMessage>(&text).ok())
+                        .and_then(|agent_message| match agent_message {
+                            ManagementJsonRpcMessage::Notification(_) => None,
+                            ManagementJsonRpcMessage::Response(response_envelope) => {
+                                Some(response_envelope)
+                            }
+                        }),
+                )
+            })
+            .next()
+            .await
+            .ok_or(ClusterHarnessError::AgentHungUpBeforeResponding)
+    }
+
+    pub async fn send(&mut self, message: Message) -> Result<(), ClusterHarnessError> {
+        self.websocket
+            .send(message)
+            .await
+            .map_err(ClusterHarnessError::BalancerSocketSendFailed)
+    }
+
+    pub async fn wait_for_the_agent_to_serve(&mut self) -> Result<(), ClusterHarnessError> {
+        self.websocket
+            .by_ref()
+            .filter_map(|frame| {
+                ready(
+                    frame
+                        .ok()
+                        .and_then(|frame| frame.into_text().ok())
+                        .and_then(|text| from_str::<ManagementJsonRpcMessage>(&text).ok())
+                        .filter(|agent_message| {
+                            matches!(
+                                agent_message,
+                                ManagementJsonRpcMessage::Notification(
+                                    ManagementJsonRpcNotification::UpdateAgentStatus(
+                                        update_agent_status_params
+                                    )
+                                ) if matches!(
+                                    update_agent_status_params
+                                        .slot_aggregated_status_snapshot
+                                        .status
+                                        .runtime,
+                                    AgentRuntimeStatus::Serving { .. }
+                                )
+                            )
+                        }),
+                )
+            })
+            .next()
+            .await
+            .map(|_serving_report| ())
+            .ok_or(ClusterHarnessError::AgentHungUpBeforeServing)
+    }
+
+    pub async fn wait_for_the_agent_to_hang_up(&mut self) {
+        self.websocket.by_ref().for_each(|_frame| ready(())).await;
+    }
+}
diff --git a/paddler_test_cluster_harness/src/untrusted_balancer_socket_server.rs b/paddler_test_cluster_harness/src/untrusted_balancer_socket_server.rs
new file mode 100644
index 000000000..3b665fee3
--- /dev/null
+++ b/paddler_test_cluster_harness/src/untrusted_balancer_socket_server.rs
@@ -0,0 +1,40 @@
+use std::io::Error as IoError;
+use std::net::SocketAddr;
+
+use futures_util::TryFutureExt as _;
+use tokio::net::TcpListener;
+use tokio_tungstenite::accept_async;
+
+use crate::cluster_harness_error::ClusterHarnessError;
+use crate::ephemeral_loopback_addr::EPHEMERAL_LOOPBACK_ADDR;
+use crate::untrusted_balancer_socket::UntrustedBalancerSocket;
+
+pub struct UntrustedBalancerSocketServer {
+    pub management_addr: SocketAddr,
+    tcp_listener: TcpListener,
+}
+
+impl UntrustedBalancerSocketServer {
+    pub async fn bind() -> Result<Self, IoError> {
+        TcpListener::bind(EPHEMERAL_LOOPBACK_ADDR)
+            .await
+            .and_then(|tcp_listener| {
+                tcp_listener.local_addr().map(|management_addr| Self {
+                    management_addr,
+                    tcp_listener,
+                })
+            })
+    }
+
+    pub async fn accept_agent(&self) -> Result<UntrustedBalancerSocket, ClusterHarnessError> {
+        self.tcp_listener
+            .accept()
+            .map_err(ClusterHarnessError::BalancerSocketAcceptFailed)
+            .and_then(|(tcp_stream, _agent_addr)| {
+                accept_async(tcp_stream)
+                    .map_ok(UntrustedBalancerSocket::new)
+                    .map_err(ClusterHarnessError::BalancerSocketHandshakeFailed)
+            })
+            .await
+    }
+}
diff --git a/paddler_test_cluster_harness/tests/every_model_card_resolves_from_the_hugging_face_cache.rs b/paddler_test_cluster_harness/tests/every_model_card_resolves_from_the_hugging_face_cache.rs
index f7db6b0b6..840c35e74 100644
--- a/paddler_test_cluster_harness/tests/every_model_card_resolves_from_the_hugging_face_cache.rs
+++ b/paddler_test_cluster_harness/tests/every_model_card_resolves_from_the_hugging_face_cache.rs
@@ -1,37 +1,44 @@
 #![cfg(feature = "tests_that_use_llms")]
 
-use anyhow::Result;
-use hf_hub::Cache;
-use hf_hub::Repo;
-use hf_hub::RepoType;
-use hf_hub::api::tokio::ApiBuilder;
+use std::sync::Arc;
 
-use paddler_messaging::huggingface_model_reference::HuggingFaceModelReference;
-use paddler_test_cluster_harness::model_card::ModelCard;
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
+use paddler_model_source::desired_model_resolution::DesiredModelResolution;
+use paddler_model_source::huggingface_cache_location::HuggingFaceCacheLocation;
+use paddler_model_source::huggingface_default_endpoint::HUGGINGFACE_DEFAULT_ENDPOINT;
+use paddler_model_source::huggingface_hub::HuggingFaceHub;
+use paddler_model_source::huggingface_model_source::HuggingFaceModelSource;
+use paddler_model_source::resolves_model_source::ResolvesModelSource as _;
 use paddler_test_cluster_harness::model_card::every_model_card::every_model_card;
 
 #[tokio::test]
-async fn every_model_card_resolves_from_the_hugging_face_cache() -> Result<()> {
-    let hugging_face_cache = Cache::from_env();
-    let hugging_face_api = ApiBuilder::from_cache(hugging_face_cache.clone()).build()?;
-
-    for ModelCard {
-        reference:
-            HuggingFaceModelReference {
-                filename,
-                repo_id,
-                revision,
-            },
-    } in every_model_card()
-    {
-        let repo = Repo::with_revision(repo_id, RepoType::Model, revision);
-        let fetched_path = hugging_face_api.repo(repo.clone()).get(&filename).await?;
+async fn every_model_card_resolves_from_the_hugging_face_cache() {
+    for model_card in every_model_card() {
+        let resolution = HuggingFaceModelSource {
+            client: HuggingFaceHub {
+                cache_location: HuggingFaceCacheLocation::HubDefault,
+                endpoint: HUGGINGFACE_DEFAULT_ENDPOINT.to_owned(),
+            }
+            .client()
+            .expect("the Hugging Face client must start"),
+            reference: model_card.reference.clone(),
+        }
+        .resolve(
+            &CancellationToken::new(),
+            Arc::new(SlotAggregatedStatus::new(1)),
+        )
+        .await
+        .expect("every model card must resolve through the Hugging Face hub");
 
         assert_eq!(
-            hugging_face_cache.repo(repo).get(&filename),
-            Some(fetched_path)
+            resolution,
+            DesiredModelResolution::Resolved(
+                model_card
+                    .cached_path()
+                    .expect("every resolved model card must be in the Hugging Face cache")
+            )
         );
     }
-
-    Ok(())
 }
diff --git a/paddler_test_cluster_python/.gitignore b/paddler_test_cluster_python/.gitignore
deleted file mode 100644
index 4cf161822..000000000
--- a/paddler_test_cluster_python/.gitignore
+++ /dev/null
@@ -1,21 +0,0 @@
-__pycache__/
-*.py[cod]
-*$py.class
-*.so
-dist/
-build/
-*.egg-info/
-*.egg
-.eggs/
-*.whl
-.venv
-.venv/
-venv/
-.mypy_cache/
-.pytest_cache/
-.ruff_cache/
-*.cover
-*.log
-.coverage
-htmlcov/
-target/
diff --git a/paddler_test_cluster_python/paddler_test_cluster/__init__.py b/paddler_test_cluster_python/paddler_test_cluster/__init__.py
deleted file mode 100644
index e69de29bb..000000000
diff --git a/paddler_test_cluster_python/paddler_test_cluster/agent_spec.py b/paddler_test_cluster_python/paddler_test_cluster/agent_spec.py
deleted file mode 100644
index 3987134a0..000000000
--- a/paddler_test_cluster_python/paddler_test_cluster/agent_spec.py
+++ /dev/null
@@ -1,7 +0,0 @@
-from dataclasses import dataclass
-
-
-@dataclass(frozen=True)
-class AgentSpec:
-    name: str
-    slots: int
diff --git a/paddler_test_cluster_python/paddler_test_cluster/all_gpu_layers.py b/paddler_test_cluster_python/paddler_test_cluster/all_gpu_layers.py
deleted file mode 100644
index e3fee6d2c..000000000
--- a/paddler_test_cluster_python/paddler_test_cluster/all_gpu_layers.py
+++ /dev/null
@@ -1 +0,0 @@
-ALL_GPU_LAYERS = -1
diff --git a/paddler_test_cluster_python/paddler_test_cluster/balancer_addresses.py b/paddler_test_cluster_python/paddler_test_cluster/balancer_addresses.py
deleted file mode 100644
index 35292011e..000000000
--- a/paddler_test_cluster_python/paddler_test_cluster/balancer_addresses.py
+++ /dev/null
@@ -1,27 +0,0 @@
-from pydantic import BaseModel, ConfigDict
-
-from paddler_test_cluster.error import CompatOpenAIServiceNotServedError
-
-
-class BalancerAddresses(BaseModel):
-    model_config = ConfigDict(extra="forbid", frozen=True)
-
-    compat_openai: str | None
-    inference: str
-    management: str
-    web_admin_panel: str | None
-
-    @property
-    def compat_openai_url(self) -> str:
-        if self.compat_openai is None:
-            raise CompatOpenAIServiceNotServedError
-
-        return f"http://{self.compat_openai}"
-
-    @property
-    def inference_url(self) -> str:
-        return f"http://{self.inference}"
-
-    @property
-    def management_url(self) -> str:
-        return f"http://{self.management}"
diff --git a/paddler_test_cluster_python/paddler_test_cluster/error.py b/paddler_test_cluster_python/paddler_test_cluster/error.py
deleted file mode 100644
index 5ce703510..000000000
--- a/paddler_test_cluster_python/paddler_test_cluster/error.py
+++ /dev/null
@@ -1,47 +0,0 @@
-from paddler_client.agent_issue import AgentIssue
-
-
-class PaddlerTestClusterError(Exception):
-    pass
-
-
-class AgentReportedIssuesError(PaddlerTestClusterError):
-    def __init__(self, agent_name: str, issues: list[AgentIssue]) -> None:
-        self.agent_name = agent_name
-        self.issues = issues
-        super().__init__(f"Agent {agent_name!r} reported issues: {issues!r}")
-
-
-class AgentsStreamClosedError(PaddlerTestClusterError):
-    def __init__(self, agent_name: str) -> None:
-        self.agent_name = agent_name
-        super().__init__(
-            f"The agents stream closed before agent {agent_name!r} became ready"
-        )
-
-
-class BalancerExitedBeforeAnnouncingError(PaddlerTestClusterError):
-    def __init__(self, returncode: int) -> None:
-        self.returncode = returncode
-        super().__init__(
-            f"The balancer exited with status {returncode} before announcing "
-            "its addresses"
-        )
-
-
-class CompatOpenAIServiceNotServedError(PaddlerTestClusterError):
-    def __init__(self) -> None:
-        super().__init__("The balancer does not serve the OpenAI compatibility service")
-
-
-class ProcessDidNotExitError(PaddlerTestClusterError):
-    def __init__(self, pid: int) -> None:
-        self.pid = pid
-        super().__init__(f"Process {pid} did not exit within Paddler's shutdown bound")
-
-
-class ProcessExitedWithFailureError(PaddlerTestClusterError):
-    def __init__(self, pid: int, returncode: int) -> None:
-        self.pid = pid
-        self.returncode = returncode
-        super().__init__(f"Process {pid} exited with status {returncode}")
diff --git a/paddler_test_cluster_python/paddler_test_cluster/nomic_embed_text_v1_5_desired_state.py b/paddler_test_cluster_python/paddler_test_cluster/nomic_embed_text_v1_5_desired_state.py
deleted file mode 100644
index 0ceccca55..000000000
--- a/paddler_test_cluster_python/paddler_test_cluster/nomic_embed_text_v1_5_desired_state.py
+++ /dev/null
@@ -1,21 +0,0 @@
-from paddler_client.agent_desired_model import AgentDesiredModel
-from paddler_client.balancer_desired_state import BalancerDesiredState
-from paddler_client.huggingface_model_reference import HuggingFaceModelReference
-from paddler_client.inference_parameters import InferenceParameters
-
-from paddler_test_cluster.all_gpu_layers import ALL_GPU_LAYERS
-
-NOMIC_EMBED_TEXT_V1_5_DESIRED_STATE = BalancerDesiredState(
-    inference_parameters=InferenceParameters(
-        context_size=2048,
-        enable_embeddings=True,
-        n_gpu_layers=ALL_GPU_LAYERS,
-    ),
-    model=AgentDesiredModel.from_huggingface(
-        HuggingFaceModelReference(
-            filename="nomic-embed-text-v1.5.Q2_K.gguf",
-            repo_id="nomic-ai/nomic-embed-text-v1.5-GGUF",
-            revision="main",
-        )
-    ),
-)
diff --git a/paddler_test_cluster_python/paddler_test_cluster/paddler_binary_path.py b/paddler_test_cluster_python/paddler_test_cluster/paddler_binary_path.py
deleted file mode 100644
index 3f64d5d82..000000000
--- a/paddler_test_cluster_python/paddler_test_cluster/paddler_binary_path.py
+++ /dev/null
@@ -1,7 +0,0 @@
-import os
-
-PADDLER_BINARY_ENV = "PADDLER_BINARY"
-
-
-def paddler_binary_path() -> str:
-    return os.environ[PADDLER_BINARY_ENV]
diff --git a/paddler_test_cluster_python/paddler_test_cluster/paddler_cluster.py b/paddler_test_cluster_python/paddler_test_cluster/paddler_cluster.py
deleted file mode 100644
index 8f451f708..000000000
--- a/paddler_test_cluster_python/paddler_test_cluster/paddler_cluster.py
+++ /dev/null
@@ -1,57 +0,0 @@
-from collections.abc import AsyncIterator, Sequence
-from contextlib import aclosing, asynccontextmanager
-from datetime import timedelta
-from typing import TYPE_CHECKING
-
-from paddler_client.agent_desired_model import AgentDesiredModel
-from paddler_client.balancer_desired_state import BalancerDesiredState
-from paddler_client.client_management import ClientManagement
-
-from paddler_test_cluster.agent_spec import AgentSpec
-from paddler_test_cluster.balancer_addresses import BalancerAddresses
-from paddler_test_cluster.spawned_agent import spawn_agent
-from paddler_test_cluster.spawned_balancer import (
-    PADDLER_BUFFERED_REQUEST_TIMEOUT,
-    SpawnedBalancer,
-)
-from paddler_test_cluster.wait_for_agent_ready import wait_for_agent_ready
-
-if TYPE_CHECKING:
-    from paddler_test_cluster.spawned_process import SpawnedProcess
-
-
-@asynccontextmanager
-async def paddler_cluster(
-    desired_state: BalancerDesiredState,
-    agents: Sequence[AgentSpec],
-    *,
-    buffered_request_timeout: timedelta = PADDLER_BUFFERED_REQUEST_TIMEOUT,
-) -> AsyncIterator[BalancerAddresses]:
-    balancer = await SpawnedBalancer.spawn(
-        buffered_request_timeout=buffered_request_timeout,
-    )
-    agent_processes: list[SpawnedProcess] = []
-    expects_loaded_slots = desired_state.model != AgentDesiredModel.none()
-
-    try:
-        async with ClientManagement(url=balancer.addresses.management_url) as client:
-            await client.put_balancer_desired_state(desired_state)
-
-            async with aclosing(client.agents_stream()) as snapshots:
-                for agent in agents:
-                    agent_processes.append(
-                        await spawn_agent(balancer.addresses.management, agent)
-                    )
-
-                    await wait_for_agent_ready(
-                        snapshots,
-                        agent.name,
-                        agent.slots if expects_loaded_slots else 0,
-                    )
-
-        yield balancer.addresses
-    finally:
-        for agent_process in agent_processes:
-            await agent_process.terminate()
-
-        await balancer.process.terminate()
diff --git a/paddler_test_cluster_python/paddler_test_cluster/py.typed b/paddler_test_cluster_python/paddler_test_cluster/py.typed
deleted file mode 100644
index e69de29bb..000000000
diff --git a/paddler_test_cluster_python/paddler_test_cluster/qwen3_0_6b_desired_state.py b/paddler_test_cluster_python/paddler_test_cluster/qwen3_0_6b_desired_state.py
deleted file mode 100644
index bc238570a..000000000
--- a/paddler_test_cluster_python/paddler_test_cluster/qwen3_0_6b_desired_state.py
+++ /dev/null
@@ -1,26 +0,0 @@
-from paddler_client.agent_desired_model import AgentDesiredModel
-from paddler_client.balancer_desired_state import BalancerDesiredState
-from paddler_client.huggingface_model_reference import HuggingFaceModelReference
-from paddler_client.inference_parameters import InferenceParameters
-
-from paddler_test_cluster.all_gpu_layers import ALL_GPU_LAYERS
-
-QWEN3_0_6B_DESIRED_STATE = BalancerDesiredState(
-    inference_parameters=InferenceParameters(
-        min_p=0.0,
-        n_gpu_layers=ALL_GPU_LAYERS,
-        penalty_frequency=0.0,
-        penalty_presence=0.0,
-        penalty_repeat=1.0,
-        temperature=0.0,
-        top_k=1,
-        top_p=1.0,
-    ),
-    model=AgentDesiredModel.from_huggingface(
-        HuggingFaceModelReference(
-            filename="Qwen3-0.6B-Q8_0.gguf",
-            repo_id="Qwen/Qwen3-0.6B-GGUF",
-            revision="main",
-        )
-    ),
-)
diff --git a/paddler_test_cluster_python/paddler_test_cluster/spawned_agent.py b/paddler_test_cluster_python/paddler_test_cluster/spawned_agent.py
deleted file mode 100644
index c9bf696ea..000000000
--- a/paddler_test_cluster_python/paddler_test_cluster/spawned_agent.py
+++ /dev/null
@@ -1,21 +0,0 @@
-import asyncio
-
-from paddler_test_cluster.agent_spec import AgentSpec
-from paddler_test_cluster.paddler_binary_path import paddler_binary_path
-from paddler_test_cluster.spawned_process import SpawnedProcess
-
-
-async def spawn_agent(management_address: str, agent: AgentSpec) -> SpawnedProcess:
-    process = await asyncio.create_subprocess_exec(
-        paddler_binary_path(),
-        "agent",
-        "--management-addr",
-        management_address,
-        "--name",
-        agent.name,
-        "--slots",
-        str(agent.slots),
-        stdout=asyncio.subprocess.DEVNULL,
-    )
-
-    return SpawnedProcess(process)
diff --git a/paddler_test_cluster_python/paddler_test_cluster/spawned_balancer.py b/paddler_test_cluster_python/paddler_test_cluster/spawned_balancer.py
deleted file mode 100644
index 80e1fb283..000000000
--- a/paddler_test_cluster_python/paddler_test_cluster/spawned_balancer.py
+++ /dev/null
@@ -1,50 +0,0 @@
-import asyncio
-from dataclasses import dataclass
-from datetime import timedelta
-
-from paddler_test_cluster.balancer_addresses import BalancerAddresses
-from paddler_test_cluster.error import BalancerExitedBeforeAnnouncingError
-from paddler_test_cluster.paddler_binary_path import paddler_binary_path
-from paddler_test_cluster.spawned_process import SpawnedProcess
-
-EPHEMERAL_LOOPBACK_ADDRESS = "127.0.0.1:0"
-MEMORY_STATE_DATABASE = "memory://"
-PADDLER_BUFFERED_REQUEST_TIMEOUT = timedelta(seconds=10)
-
-
-@dataclass(frozen=True)
-class SpawnedBalancer:
-    addresses: BalancerAddresses
-    process: SpawnedProcess
-
-    @classmethod
-    async def spawn(
-        cls,
-        *,
-        buffered_request_timeout: timedelta = PADDLER_BUFFERED_REQUEST_TIMEOUT,
-        state_database_url: str = MEMORY_STATE_DATABASE,
-    ) -> "SpawnedBalancer":
-        process = await asyncio.create_subprocess_exec(
-            paddler_binary_path(),
-            "balancer",
-            "--inference-addr",
-            EPHEMERAL_LOOPBACK_ADDRESS,
-            "--management-addr",
-            EPHEMERAL_LOOPBACK_ADDRESS,
-            "--compat-openai-addr",
-            EPHEMERAL_LOOPBACK_ADDRESS,
-            "--state-database",
-            state_database_url,
-            "--buffered-request-timeout",
-            str(buffered_request_timeout // timedelta(milliseconds=1)),
-            stdout=asyncio.subprocess.PIPE,
-        )
-        announcement = await process.stdout.readline() if process.stdout else b""
-
-        if not announcement:
-            raise BalancerExitedBeforeAnnouncingError(await process.wait())
-
-        return cls(
-            addresses=BalancerAddresses.model_validate_json(announcement),
-            process=SpawnedProcess(process),
-        )
diff --git a/paddler_test_cluster_python/paddler_test_cluster/spawned_process.py b/paddler_test_cluster_python/paddler_test_cluster/spawned_process.py
deleted file mode 100644
index a80ec30ed..000000000
--- a/paddler_test_cluster_python/paddler_test_cluster/spawned_process.py
+++ /dev/null
@@ -1,40 +0,0 @@
-import asyncio
-from datetime import timedelta
-
-from paddler_test_cluster.error import (
-    ProcessDidNotExitError,
-    ProcessExitedWithFailureError,
-)
-
-TRZCINA_DEFAULT_COOPERATIVE_DEADLINE = timedelta(seconds=10)
-TRZCINA_DEFAULT_ABORT_DEADLINE = timedelta(seconds=10)
-PADDLER_SHUTDOWN_BOUND = (
-    TRZCINA_DEFAULT_COOPERATIVE_DEADLINE + TRZCINA_DEFAULT_ABORT_DEADLINE
-)
-
-
-class SpawnedProcess:
-    def __init__(self, process: asyncio.subprocess.Process) -> None:
-        self._process = process
-
-    @property
-    def pid(self) -> int:
-        return self._process.pid
-
-    async def terminate(self) -> None:
-        if self._process.returncode is None:
-            self._process.terminate()
-
-        try:
-            returncode = await asyncio.wait_for(
-                self._process.wait(),
-                PADDLER_SHUTDOWN_BOUND.total_seconds(),
-            )
-        except TimeoutError as timeout_error:
-            self._process.kill()
-            await self._process.wait()
-
-            raise ProcessDidNotExitError(self._process.pid) from timeout_error
-
-        if returncode != 0:
-            raise ProcessExitedWithFailureError(self._process.pid, returncode)
diff --git a/paddler_test_cluster_python/paddler_test_cluster/wait_for_agent_ready.py b/paddler_test_cluster_python/paddler_test_cluster/wait_for_agent_ready.py
deleted file mode 100644
index 7864c638d..000000000
--- a/paddler_test_cluster_python/paddler_test_cluster/wait_for_agent_ready.py
+++ /dev/null
@@ -1,33 +0,0 @@
-from collections.abc import AsyncIterator
-
-from paddler_client.agent_controller_pool_snapshot import (
-    AgentControllerPoolSnapshot,
-)
-from paddler_client.agent_state_application_status import (
-    AgentStateApplicationStatus,
-)
-
-from paddler_test_cluster.error import (
-    AgentReportedIssuesError,
-    AgentsStreamClosedError,
-)
-
-
-async def wait_for_agent_ready(
-    snapshots: AsyncIterator[AgentControllerPoolSnapshot],
-    agent_name: str,
-    expected_slots_total: int,
-) -> None:
-    async for snapshot in snapshots:
-        for agent in (agent for agent in snapshot.agents if agent.name == agent_name):
-            if agent.status.issues:
-                raise AgentReportedIssuesError(agent_name, agent.status.issues)
-
-            if (
-                agent.status.state_application_status
-                == AgentStateApplicationStatus.APPLIED
-                and agent.status.slots_total == expected_slots_total
-            ):
-                return
-
-    raise AgentsStreamClosedError(agent_name)
diff --git a/paddler_test_cluster_python/poetry.lock b/paddler_test_cluster_python/poetry.lock
deleted file mode 100644
index e99ec5999..000000000
--- a/paddler_test_cluster_python/poetry.lock
+++ /dev/null
@@ -1,970 +0,0 @@
-# This file is automatically @generated by Poetry 2.4.3 and should not be changed by hand.
-
-[[package]]
-name = "annotated-types"
-version = "0.8.0"
-description = "Reusable constraint types to use with typing.Annotated"
-optional = false
-python-versions = ">=3.10"
-groups = ["main"]
-files = [
-    {file = "annotated_types-0.8.0-py3-none-any.whl", hash = "sha256:f072f4d804ea359e4eaf198b1af7a8b0943881a87f31bb764f8bf219bb9419e0"},
-    {file = "annotated_types-0.8.0.tar.gz", hash = "sha256:13b2beaad985e05e2d6407ee4c4f35590b11f8d693a258a561055cac8f64cab7"},
-]
-
-[[package]]
-name = "anyio"
-version = "4.15.1"
-description = "High-level concurrency and networking framework on top of asyncio or Trio"
-optional = false
-python-versions = ">=3.10"
-groups = ["main"]
-files = [
-    {file = "anyio-4.15.1-py3-none-any.whl", hash = "sha256:6152fdbbf9a77fdec97731721bebf7c4c44f7c29b424b0065826173efc7ed101"},
-    {file = "anyio-4.15.1.tar.gz", hash = "sha256:9f28306018cbd6d329e64a36d58256edff76dd996fe423bc957326e578b82a94"},
-]
-
-[package.dependencies]
-idna = ">=2.8"
-typing_extensions = {version = ">=4.16.0", markers = "python_version < \"3.15\""}
-
-[package.extras]
-trio = ["trio (>=0.32.0)"]
-
-[[package]]
-name = "certifi"
-version = "2026.7.22"
-description = "Python package for providing Mozilla's CA Bundle."
-optional = false
-python-versions = ">=3.7"
-groups = ["main"]
-files = [
-    {file = "certifi-2026.7.22-py3-none-any.whl", hash = "sha256:62f22742b58a1a33014a2b6b706588a8d7e2a88ae7bd1a6ebe8c992928483775"},
-    {file = "certifi-2026.7.22.tar.gz", hash = "sha256:741e2c3b351ddf169a738da9f2c048608ff7f2c5cc02f1ebc6b118bb090d5d55"},
-]
-
-[[package]]
-name = "colorama"
-version = "0.4.6"
-description = "Cross-platform colored terminal text."
-optional = false
-python-versions = "!=3.0.*,!=3.1.*,!=3.2.*,!=3.3.*,!=3.4.*,!=3.5.*,!=3.6.*,>=2.7"
-groups = ["dev"]
-markers = "sys_platform == \"win32\""
-files = [
-    {file = "colorama-0.4.6-py2.py3-none-any.whl", hash = "sha256:4f1d9991f5acc0ca119f9d443620b77f9d6b33703e51011c16baf57afb285fc6"},
-    {file = "colorama-0.4.6.tar.gz", hash = "sha256:08695f5cb7ed6e0531a20572697297273c47b8cae5a63ffc6d6ed5c201be6e44"},
-]
-
-[[package]]
-name = "coverage"
-version = "7.16.2"
-description = "Code coverage measurement for Python"
-optional = false
-python-versions = ">=3.10"
-groups = ["dev"]
-files = [
-    {file = "coverage-7.16.2-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:23219888477edd736b6fcaec1272d47d93b926e999641ffea7e53a1738e70b2b"},
-    {file = "coverage-7.16.2-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:40c0f00899fe6181ae7f434ceb200e51f5ee4b8ed10e3b5f0b605f0cae15da87"},
-    {file = "coverage-7.16.2-cp310-cp310-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:a4624f80732f6b427ac58f1f59c577a0994a12e8174b5af6a027b4b58795d4c3"},
-    {file = "coverage-7.16.2-cp310-cp310-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:191803c4996b499fcd78c2ad5e5f767dcc53cb4dc6de6d6a741b443a1821ef02"},
-    {file = "coverage-7.16.2-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:9fd670ac43b709c575aefc25bf52d8a598a3bc5017bddfd0a179152ab06a2deb"},
-    {file = "coverage-7.16.2-cp310-cp310-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:705e5af11d34647efdc170c7840b6857c81cf74be96419a553f237e68e62cb72"},
-    {file = "coverage-7.16.2-cp310-cp310-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:8afd9bf35cc6a1f22eb3634808fa8e0b91902459c5721ef2e4461dfe771d7f08"},
-    {file = "coverage-7.16.2-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:3f43bac1856ba269b905302778d4df433d6006489a192174ad77ac528e395032"},
-    {file = "coverage-7.16.2-cp310-cp310-musllinux_1_2_i686.whl", hash = "sha256:f8475460aa33ee28ac896ab1156d0bb3b6c639f7f8383c2677d3359eb35f8205"},
-    {file = "coverage-7.16.2-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:d6276d78f6fca7d0ac066d5da4165c5acd07829e8305c2cb900b738fb3a75a72"},
-    {file = "coverage-7.16.2-cp310-cp310-musllinux_1_2_riscv64.whl", hash = "sha256:736fde09ea39646d11f8e3b76bd3425c075aa4dd45f24891970bb77c14ff20f5"},
-    {file = "coverage-7.16.2-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:c85d54e7e8a2ca932fe8399301af9b8d5907ea2a455ffaff6e7d1208db83b943"},
-    {file = "coverage-7.16.2-cp310-cp310-win32.whl", hash = "sha256:5139009b5efd2194fc168ee9362f0e191ba612ef5d29242f9269c22f9b8f80c7"},
-    {file = "coverage-7.16.2-cp310-cp310-win_amd64.whl", hash = "sha256:c3305c38a2fa21a4254f2ace7dd9ef5fc569c9a558b66e7017650b3d637fb95e"},
-    {file = "coverage-7.16.2-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:732d950e51f3ba4fb6209c73250f3e8924fefca42953ee04a9e65d8c02414d7d"},
-    {file = "coverage-7.16.2-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:5dca0bb66b4c3d624ba047887bf70270030c150692d543cb501293dc38a9f4b5"},
-    {file = "coverage-7.16.2-cp311-cp311-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:af2a2a8c7c74de0559e0c368d94c8def9e16c58faaee33a0bf081057c4227e3b"},
-    {file = "coverage-7.16.2-cp311-cp311-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:db5f8394e17f877a625b257f2ba0ce8e728a499c2c1579ad66220272cd3df510"},
-    {file = "coverage-7.16.2-cp311-cp311-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:5b3146d2317c75f70df2509066d979dadd941f7021cdf9b5db4bcd8568258e25"},
-    {file = "coverage-7.16.2-cp311-cp311-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:9e1d0ced76318bab499693ff25f64faa343415187cb2e4d7befdfdd391a1cf6a"},
-    {file = "coverage-7.16.2-cp311-cp311-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:af98ad5ed9d6daaca956201e00bb429a7eb2b080426686f70a20353e0f9839f5"},
-    {file = "coverage-7.16.2-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:1d56e4d21c56d2046447733f8b118409597db48c01efe898ee9ac24e858ec2d6"},
-    {file = "coverage-7.16.2-cp311-cp311-musllinux_1_2_i686.whl", hash = "sha256:1d5d0e3b660506fb84f995814e3118a21efdc0c8eb80127da1be627d90093c17"},
-    {file = "coverage-7.16.2-cp311-cp311-musllinux_1_2_ppc64le.whl", hash = "sha256:17228fbca0f22976f797be94e975dcd237799c657d49551c7de1e0654d1202e9"},
-    {file = "coverage-7.16.2-cp311-cp311-musllinux_1_2_riscv64.whl", hash = "sha256:bc0b0ac781d489304b741269857f1f8338b7a26b1b89c06c0344658001ec0035"},
-    {file = "coverage-7.16.2-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:bf1bd822ec4e387ed245bed0d71151582cf7be9e5309bc4145eefe36083d5878"},
-    {file = "coverage-7.16.2-cp311-cp311-win32.whl", hash = "sha256:7ed238d227e23cc300c3d464babdaf9f6ddc740aa1b15a77ae96136e6a7c4516"},
-    {file = "coverage-7.16.2-cp311-cp311-win_amd64.whl", hash = "sha256:a90700f743e29aa3d75a6ff5f01953176a889c00e526194bc4d281731b88d99d"},
-    {file = "coverage-7.16.2-cp311-cp311-win_arm64.whl", hash = "sha256:a336eec40e3520d369b8a6cdabb4f596e69a8b42927ca074aa1452fed943238a"},
-    {file = "coverage-7.16.2-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:218d742afca2b5ad5ca759e93eddedfbcc6eadf8322f080dcefc40b7bd4e2d48"},
-    {file = "coverage-7.16.2-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:a9a638be322a8d76a41cdb17781c7f82aaee6a66493d8ffb7e2c09ee22423d99"},
-    {file = "coverage-7.16.2-cp312-cp312-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:724bd0f1e81856b35e59fc98cf7b4e544a3cb662e4e0864dca73d4326ee9d808"},
-    {file = "coverage-7.16.2-cp312-cp312-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:5375ebd99038021b35e99dc88255022912c06565d316212f4a576e4b08d30f5d"},
-    {file = "coverage-7.16.2-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:7a076277ca9f5750cc230f0f578ebd2620cec60255b25707361699fef6fb465c"},
-    {file = "coverage-7.16.2-cp312-cp312-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:58d4a54c6ea672afef66d49be922a2c69826c5ae1a42a9cd94f0c9c2bacdf800"},
-    {file = "coverage-7.16.2-cp312-cp312-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:0dcbcfcc059117284c603ff8cb61a65872512882f84a8cf0339241f7f7c2f148"},
-    {file = "coverage-7.16.2-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:afdf43b72ef3876c1fe66423b91466e37877c9e81e8cec70542b7e8525b9d1b7"},
-    {file = "coverage-7.16.2-cp312-cp312-musllinux_1_2_i686.whl", hash = "sha256:9acc7f7ec4a1b5f89bd929fde5b8a714f6fafdc6cc18725413d510aa082b47ad"},
-    {file = "coverage-7.16.2-cp312-cp312-musllinux_1_2_ppc64le.whl", hash = "sha256:80d3f7b48d43ee8fc5e8707a8adb43d743a5a1a85256c25a24f9d6d0e2238fa6"},
-    {file = "coverage-7.16.2-cp312-cp312-musllinux_1_2_riscv64.whl", hash = "sha256:126d1af8804d7224421fe991ff65d3ce649081560df7a98b1a5ffff07f9923bd"},
-    {file = "coverage-7.16.2-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:c19cd6d025c1673f22afcd22c7df8a662d779e05d8e3fa6820c22afb895b0206"},
-    {file = "coverage-7.16.2-cp312-cp312-win32.whl", hash = "sha256:152877cdc8a07264882cfcd503ba56a3ef6cba56a70e8c70f6eb8ffd7384789a"},
-    {file = "coverage-7.16.2-cp312-cp312-win_amd64.whl", hash = "sha256:e6c52d3307824ff93b39efd99e4185d557db40bd841452abfb32e5d9151ca162"},
-    {file = "coverage-7.16.2-cp312-cp312-win_arm64.whl", hash = "sha256:a678c0b6b22086ec2427359d22e37445d4a792f5fdbbc744112c7dade65cad02"},
-    {file = "coverage-7.16.2-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:1a37c6e478cf687e1aa30a593d19c92c02fad9d122b51ab73f51b8dc7a0c0fc9"},
-    {file = "coverage-7.16.2-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:0993d0e90858c03943d3cb152e068a20dd4707924deec84dd2230261baae3b1b"},
-    {file = "coverage-7.16.2-cp313-cp313-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:bb2fc905bbf4e6b7f40806ea79e31515abf6349594cdf0adf27c4215f0463204"},
-    {file = "coverage-7.16.2-cp313-cp313-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:4358b9c8c0125b460407f3017c6cce8156e904b32772c5630d27112f52bdbfe5"},
-    {file = "coverage-7.16.2-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:1f15254427c9b33eedac4f198eaf9e356eb4f6214551afb43da6194a2c088ad7"},
-    {file = "coverage-7.16.2-cp313-cp313-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:9a75a4704ff640e46170042eec1f984385a121227c505d5a16ad8e495f452541"},
-    {file = "coverage-7.16.2-cp313-cp313-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:14253fc7bb15749b849795a06f5d3b6d8bc3fb8a4b5ddc341faf7a89dce205fc"},
-    {file = "coverage-7.16.2-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:921415102a90637fcc2e3f169f61dad7699ecf690e8639fc21b813acbedc0967"},
-    {file = "coverage-7.16.2-cp313-cp313-musllinux_1_2_i686.whl", hash = "sha256:cce2bc991293f15cc4084ca116827b5900c5f34e1a54dfe83f10ab5c43162eb7"},
-    {file = "coverage-7.16.2-cp313-cp313-musllinux_1_2_ppc64le.whl", hash = "sha256:e1fa594c887365b69745f25a416806e61085dd07b94c9eae68a6e20730629b23"},
-    {file = "coverage-7.16.2-cp313-cp313-musllinux_1_2_riscv64.whl", hash = "sha256:11e597173af1dc33d5f8a7332ada544199269a223af1ee1770ddd5e245ad0fe8"},
-    {file = "coverage-7.16.2-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:3e7f99698ba3a7d13988bdd984b7ebf13af4dbe2166dc8502eef90d77603b0a4"},
-    {file = "coverage-7.16.2-cp313-cp313-win32.whl", hash = "sha256:f80bd9f9633eafc73d0a913ba2645c96ba58bba1befc30590f7c0fbfde59d865"},
-    {file = "coverage-7.16.2-cp313-cp313-win_amd64.whl", hash = "sha256:8be099e979fc42559328a21828281b4578304191ae46ed4e80a407048a82eee6"},
-    {file = "coverage-7.16.2-cp313-cp313-win_arm64.whl", hash = "sha256:28ff850182a67d117990fa2ce5ea1032836d8c9630dae867e8bdd3bff4533b79"},
-    {file = "coverage-7.16.2-cp314-cp314-macosx_10_15_x86_64.whl", hash = "sha256:4ee546b9e4872ffa194bf07ac87bfa1202ebb824d0795dc1ef22f175545ca90a"},
-    {file = "coverage-7.16.2-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:a2fac6895eb299a2e52d7bbb8fb3903502b9da8d3f5309ceb16ec40c646b58ee"},
-    {file = "coverage-7.16.2-cp314-cp314-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:57ff3783f99d75a1e81dd56a9737eb5665e6736a5d93258ba596b6dcad8fd05b"},
-    {file = "coverage-7.16.2-cp314-cp314-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:35f37886699cb9abd29958247d718628d5bc6f39e623dff66a09e546c42a7e03"},
-    {file = "coverage-7.16.2-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:0fd7a86fdda7cb6d616d178654bd0ad6bc0f3f33c2e478aa598500a1a9e34eda"},
-    {file = "coverage-7.16.2-cp314-cp314-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:ac0f3b379c94acc2f7dce5f5f0b24d44fa1cc6a509717ef83dfee07450c2117c"},
-    {file = "coverage-7.16.2-cp314-cp314-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:7d0732c83746bc24123c581a85d9dd96b70ddb538c9076020aa1a041790361e9"},
-    {file = "coverage-7.16.2-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:7b451c68218c150f616bc9649783ec8de76a59792c759b43aa0c9c0466a465e4"},
-    {file = "coverage-7.16.2-cp314-cp314-musllinux_1_2_i686.whl", hash = "sha256:a56ac4fa5a75c7e182e8f62600cfb4aff43c5ed7356a034f3557659c3bec1d90"},
-    {file = "coverage-7.16.2-cp314-cp314-musllinux_1_2_ppc64le.whl", hash = "sha256:4cc4f73aa3fabc36e32046d6cd2971405948d8a903636508a3d3b2f9128b3a95"},
-    {file = "coverage-7.16.2-cp314-cp314-musllinux_1_2_riscv64.whl", hash = "sha256:723dcdab91357159b722935b500ee8abc0a66c8c432e1e9fabf4cc7598952de8"},
-    {file = "coverage-7.16.2-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:5397e21a90dde0e9c6896b77ded8f0be26b66f8b22b33aed41f6043ed95d55e6"},
-    {file = "coverage-7.16.2-cp314-cp314-win32.whl", hash = "sha256:848893e1d361448c113dc2f0913503522a6f7be231d0e38333d2a22d9698a011"},
-    {file = "coverage-7.16.2-cp314-cp314-win_amd64.whl", hash = "sha256:5a27b731c171e43dc8b5f32b76a5051dde2ec9b9366c87028f08a7088ebc2c7b"},
-    {file = "coverage-7.16.2-cp314-cp314-win_arm64.whl", hash = "sha256:1c569a9fd25505f1cd6bea90588818f90373ce90e2632e2cacf19ddbd6e14fdb"},
-    {file = "coverage-7.16.2-cp314-cp314t-macosx_10_15_x86_64.whl", hash = "sha256:d93db87adb6b1c1b408dce4763314b55d76a9f589e96783a84ac9e7689e48bdf"},
-    {file = "coverage-7.16.2-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:aa62c85046473959c13ba9edca9dc90a77d5c1095b1ba313556314d77fe5b036"},
-    {file = "coverage-7.16.2-cp314-cp314t-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:db76506aa5416081f3e8974ae0f7965c58ada0bb0ef7339ac86099588dbb20d3"},
-    {file = "coverage-7.16.2-cp314-cp314t-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:a0f2285329dac10ab08f79cb11f5692c497018e6c7c511f95e6fd63a70b8f831"},
-    {file = "coverage-7.16.2-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:382d3346d56b0eec1b793d53a4c88799c8053f516aa3a8d7c44315696954bacf"},
-    {file = "coverage-7.16.2-cp314-cp314t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:648352b94507179d82637292e7ae8802508d95f78e2f00a705a50b6c48011681"},
-    {file = "coverage-7.16.2-cp314-cp314t-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:fb2bde05838fffae1a1bf75e5d411a6cac3e4e9bb97e6640fed8cd47888b33f0"},
-    {file = "coverage-7.16.2-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:6a75180829efb8ae62b4aded25be6ddca1c888d138d2d82e21d93bfbd88f41cb"},
-    {file = "coverage-7.16.2-cp314-cp314t-musllinux_1_2_i686.whl", hash = "sha256:99704f73721e23859112072d522076e11c31744fc96b5652e5dd2018aa4359f7"},
-    {file = "coverage-7.16.2-cp314-cp314t-musllinux_1_2_ppc64le.whl", hash = "sha256:29309ccc86b7f33df7db12813c299f215bbbc470ed6292d0bedd63ffae1ebf64"},
-    {file = "coverage-7.16.2-cp314-cp314t-musllinux_1_2_riscv64.whl", hash = "sha256:30c1b65d529e46569899fadca59e4a87c1faf2886923f1307ba61e654d4f3c20"},
-    {file = "coverage-7.16.2-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:dcf4bc2aab4e16b1c4c0c2005918f23a7dd5d7821ddae82caed9e3342dc2fcce"},
-    {file = "coverage-7.16.2-cp314-cp314t-win32.whl", hash = "sha256:a9cd3de0a5bfe7b0e21ee10e1a14e3d61bf52efc88217ab1d95d6ace6970bd46"},
-    {file = "coverage-7.16.2-cp314-cp314t-win_amd64.whl", hash = "sha256:611a44e5229a59d7483ce830160e1a0e85f700562c7a5651c7c63fb8f4eb528c"},
-    {file = "coverage-7.16.2-cp314-cp314t-win_arm64.whl", hash = "sha256:22957cef43ce038641de78ba995de7568d2d6a37c6ddbf7fa0fd7d1ae2344d91"},
-    {file = "coverage-7.16.2-cp315-cp315-macosx_10_15_x86_64.whl", hash = "sha256:414c26dfdb96aac2d570a54e03008f001e32eb2d413705365503648c6bd361d8"},
-    {file = "coverage-7.16.2-cp315-cp315-macosx_11_0_arm64.whl", hash = "sha256:00d3eb96e9988c45f50cccd1f1496571ac5c1f91386ac02c4d55516eeda19a24"},
-    {file = "coverage-7.16.2-cp315-cp315-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:4dbbd1155ca46e6e0b6b89d204428c56ef6a459af21333f365d135a2820e5a09"},
-    {file = "coverage-7.16.2-cp315-cp315-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:8fc15cc8d0d06e873c00ef18e1372d605f9aaf3de27d8c24e50782e75bc8b843"},
-    {file = "coverage-7.16.2-cp315-cp315-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:9c6afdd69218202bc1758c9a14b86b8cf1084f37ed2ca143e567a103772b16d1"},
-    {file = "coverage-7.16.2-cp315-cp315-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:aba5c63b7afdc749cc9eae943d5b868cba2b261a176378fa1c5a30bc8bc89982"},
-    {file = "coverage-7.16.2-cp315-cp315-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:9174f0af24e5eff248b9dbfe76ec5275a3d19d37edbc2810543f12cf97347a34"},
-    {file = "coverage-7.16.2-cp315-cp315-musllinux_1_2_aarch64.whl", hash = "sha256:80e9fdb4c3d926b6ba721d4bf7435bdb869c3527ae7803290361d0ab73db13b6"},
-    {file = "coverage-7.16.2-cp315-cp315-musllinux_1_2_i686.whl", hash = "sha256:7b3bce4a0d05401d70b7d0d5ca783e686bc9d30e81dbd7d980d532609bf809e4"},
-    {file = "coverage-7.16.2-cp315-cp315-musllinux_1_2_ppc64le.whl", hash = "sha256:44f21e407b278efdfc1ee5e481e00518bd1d500310a30a5fbf2bcbedfef4aaf0"},
-    {file = "coverage-7.16.2-cp315-cp315-musllinux_1_2_riscv64.whl", hash = "sha256:59c3926585e1cd1f2190f4b2ac9014de1bbeaf0d5d0587b0dc6b0aa90d17896a"},
-    {file = "coverage-7.16.2-cp315-cp315-musllinux_1_2_x86_64.whl", hash = "sha256:066429634299e14dd2d511e1e85f8f9cecc500781f6b41907c0dd6f1baea7e63"},
-    {file = "coverage-7.16.2-cp315-cp315-win32.whl", hash = "sha256:893ea9cf86cb8d2546812ac93d973aaf2ee1fb45110a873b014214fd23e3725e"},
-    {file = "coverage-7.16.2-cp315-cp315-win_amd64.whl", hash = "sha256:01c6908bc613b420c26c818fe948e1b97dfd041a53c98b01c63bd8321f5c9aae"},
-    {file = "coverage-7.16.2-cp315-cp315-win_arm64.whl", hash = "sha256:967d72c835d7a8cf0af99ec813a2d06e3db6df706402f1fe85b31b437645f495"},
-    {file = "coverage-7.16.2-cp315-cp315t-macosx_10_15_x86_64.whl", hash = "sha256:98d9c97f51b334b0adce7b964442a9af33c1a00c6ac856984cc5dc8d18f81c75"},
-    {file = "coverage-7.16.2-cp315-cp315t-macosx_11_0_arm64.whl", hash = "sha256:3e861f1071dcc2fec1e88bef0920f6b1eaa66a143555b4f8ab79ba2b0f30ef55"},
-    {file = "coverage-7.16.2-cp315-cp315t-manylinux1_i686.manylinux_2_28_i686.manylinux_2_5_i686.whl", hash = "sha256:fb9d92ecfe2d5b494367c67f7446f8b75b68d8d0c8cf3bc3e6997478be25d9e2"},
-    {file = "coverage-7.16.2-cp315-cp315t-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:eb57acff4a74246ae513c142d4b36e18c389c3aed8661914a53f7cd0071031b2"},
-    {file = "coverage-7.16.2-cp315-cp315t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:444889f7f66b74e4455c0a97e0e166dd41177f1dca8c0239a47cff25e05ba7e1"},
-    {file = "coverage-7.16.2-cp315-cp315t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:a740ea6f083c6db7b926534d159508f80ba275ab35e722522de0d18d0f56e55f"},
-    {file = "coverage-7.16.2-cp315-cp315t-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:8e209591f7c41ae4a9171335cf6156afda0b21de73b02f73f5aa95b2d5fbb08d"},
-    {file = "coverage-7.16.2-cp315-cp315t-musllinux_1_2_aarch64.whl", hash = "sha256:396bb16e04ce04efbb3df91456ae4e3da918e69ecdf67fb711b0a0fdf35ccce0"},
-    {file = "coverage-7.16.2-cp315-cp315t-musllinux_1_2_i686.whl", hash = "sha256:9cdf19874e0d247f32f03609200370343c3c7aa260b191d8c2bb251d36198283"},
-    {file = "coverage-7.16.2-cp315-cp315t-musllinux_1_2_ppc64le.whl", hash = "sha256:fd3d72233eb8b48acc94fa57d44e2d32ce8e7abed02882ccb6d855ccc4ed33ec"},
-    {file = "coverage-7.16.2-cp315-cp315t-musllinux_1_2_riscv64.whl", hash = "sha256:bb4ffe96aa663cee727659db5a2afeb38c95f8677b747d447b90d6d4874ea2c5"},
-    {file = "coverage-7.16.2-cp315-cp315t-musllinux_1_2_x86_64.whl", hash = "sha256:dba2edfb054f6d4a08df9d1637c39a5aa3865bca6617c13c86be21e45658a59c"},
-    {file = "coverage-7.16.2-cp315-cp315t-win32.whl", hash = "sha256:251aed777c47c77aba047096d4542889db089227655711dfc2b9c54ef0e15e35"},
-    {file = "coverage-7.16.2-cp315-cp315t-win_amd64.whl", hash = "sha256:2aca0bdfa9e91621d5b09d815357bf63def4fc0e9cb66da67bf2cf93f3b1a6f5"},
-    {file = "coverage-7.16.2-cp315-cp315t-win_arm64.whl", hash = "sha256:b88841e654f09732804809e435b3e005a929ffd9998b872b7b213957b8759cb8"},
-    {file = "coverage-7.16.2-py3-none-any.whl", hash = "sha256:11d28e9123a9156cb405d8d27b44256c9a58fb5decc2073a8f17862057e3aa0f"},
-    {file = "coverage-7.16.2.tar.gz", hash = "sha256:ca64d9f1f384f151b9511bec01126072acd2f313439f8ed015a22d8790aab6fa"},
-]
-
-[package.extras]
-toml = ["tomli ; python_full_version <= \"3.11.0a6\""]
-
-[[package]]
-name = "h11"
-version = "0.16.0"
-description = "A pure-Python, bring-your-own-I/O implementation of HTTP/1.1"
-optional = false
-python-versions = ">=3.8"
-groups = ["main"]
-files = [
-    {file = "h11-0.16.0-py3-none-any.whl", hash = "sha256:63cf8bbe7522de3bf65932fda1d9c2772064ffb3dae62d55932da54b31cb6c86"},
-    {file = "h11-0.16.0.tar.gz", hash = "sha256:4e35b956cf45792e4caa5885e69fba00bdbc6ffafbfa020300e549b208ee5ff1"},
-]
-
-[[package]]
-name = "httpcore"
-version = "1.0.9"
-description = "A minimal low-level HTTP client."
-optional = false
-python-versions = ">=3.8"
-groups = ["main"]
-files = [
-    {file = "httpcore-1.0.9-py3-none-any.whl", hash = "sha256:2d400746a40668fc9dec9810239072b40b4484b640a8c38fd654a024c7a1bf55"},
-    {file = "httpcore-1.0.9.tar.gz", hash = "sha256:6e34463af53fd2ab5d807f399a9b45ea31c3dfa2276f15a2c3f00afff6e176e8"},
-]
-
-[package.dependencies]
-certifi = "*"
-h11 = ">=0.16"
-
-[package.extras]
-asyncio = ["anyio (>=4.0,<5.0)"]
-http2 = ["h2 (>=3,<5)"]
-socks = ["socksio (==1.*)"]
-trio = ["trio (>=0.22.0,<1.0)"]
-
-[[package]]
-name = "httpx"
-version = "0.28.1"
-description = "The next generation HTTP client."
-optional = false
-python-versions = ">=3.8"
-groups = ["main"]
-files = [
-    {file = "httpx-0.28.1-py3-none-any.whl", hash = "sha256:d909fcccc110f8c7faf814ca82a9a4d816bc5a6dbfea25d6591d6985b8ba59ad"},
-    {file = "httpx-0.28.1.tar.gz", hash = "sha256:75e98c5f16b0f35b567856f597f06ff2270a374470a5c2392242528e3e3e42fc"},
-]
-
-[package.dependencies]
-anyio = "*"
-certifi = "*"
-httpcore = "==1.*"
-idna = "*"
-
-[package.extras]
-brotli = ["brotli ; platform_python_implementation == \"CPython\"", "brotlicffi ; platform_python_implementation != \"CPython\""]
-cli = ["click (==8.*)", "pygments (==2.*)", "rich (>=10,<14)"]
-http2 = ["h2 (>=3,<5)"]
-socks = ["socksio (==1.*)"]
-zstd = ["zstandard (>=0.18.0)"]
-
-[[package]]
-name = "idna"
-version = "3.20"
-description = "Internationalized Domain Names in Applications (IDNA)"
-optional = false
-python-versions = ">=3.9"
-groups = ["main"]
-files = [
-    {file = "idna-3.20-py3-none-any.whl", hash = "sha256:ab7ae7122974553370f0bdb919e1a960b2cd1bc1ef0276416d896db81c14582c"},
-    {file = "idna-3.20.tar.gz", hash = "sha256:a7db850025b95ded1eae8a46181a1a6c56c92c96f0e2b005d9ff8dc0210cab44"},
-]
-
-[package.extras]
-all = ["coverage (>=7.10.0)", "hypothesis (>=6.141.1)", "mypy (>=1.11.2)", "pytest (>=8.3.2)", "ruff (>=0.16.0)", "ty (>=0.0.37)"]
-
-[[package]]
-name = "iniconfig"
-version = "2.3.0"
-description = "brain-dead simple config-ini parsing"
-optional = false
-python-versions = ">=3.10"
-groups = ["dev"]
-files = [
-    {file = "iniconfig-2.3.0-py3-none-any.whl", hash = "sha256:f631c04d2c48c52b84d0d0549c99ff3859c98df65b3101406327ecc7d53fbf12"},
-    {file = "iniconfig-2.3.0.tar.gz", hash = "sha256:c76315c77db068650d49c5b56314774a7804df16fee4402c1f19d6d15d8c4730"},
-]
-
-[[package]]
-name = "librt"
-version = "0.16.0"
-description = "Mypyc runtime library"
-optional = false
-python-versions = ">=3.9"
-groups = ["dev"]
-markers = "platform_python_implementation != \"PyPy\""
-files = [
-    {file = "librt-0.16.0-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:96f576f2711f8519152ec76d0e599243555c1f07679fa73606ca8c8c868c0be6"},
-    {file = "librt-0.16.0-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:2bec3818c7da7c96ceae0ef5915a3d16c52dd08f3ea913bf1fe8568c447c7978"},
-    {file = "librt-0.16.0-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:64c79520414a3fdfc6aabd7593e6169afa14d5f8d9908d4b498db068868b08dd"},
-    {file = "librt-0.16.0-cp310-cp310-manylinux2014_i686.manylinux_2_17_i686.manylinux_2_28_i686.whl", hash = "sha256:f81b5b19ce748ef68d4746656b7929762eb2fe99269b266e4be07e2ee4de7144"},
-    {file = "librt-0.16.0-cp310-cp310-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:c71d1b76210a36729fedfc5115069b50a3d8619054745f8758fe5d6f19e86671"},
-    {file = "librt-0.16.0-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:473eebc7866bb0a0c8849a292b5e7157c1aba5d14d0f0f610c52158d6d964262"},
-    {file = "librt-0.16.0-cp310-cp310-manylinux_2_34_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:4eb1313a19847089ee81e88742abedf285c60538816640b99742d8534b81d26a"},
-    {file = "librt-0.16.0-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:ebefd60b42e2a82b32d136bb5f7c94eadfcd29f772b547df6f3291d1ed855a1c"},
-    {file = "librt-0.16.0-cp310-cp310-musllinux_1_2_i686.whl", hash = "sha256:6c5da27e8056439f927ea896735da60e616c477a8293feaa3233d4e7781a6726"},
-    {file = "librt-0.16.0-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:67e718c7a43f8db325abbbf1404e2d535f12f8f7a1a82259568385cc5274b82a"},
-    {file = "librt-0.16.0-cp310-cp310-musllinux_1_2_riscv64.whl", hash = "sha256:314e703f0c19320dc8094e7a784b9cf29e1b67402515abf580069a6363c0b4f1"},
-    {file = "librt-0.16.0-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:943c6bbecbdf7fa575a4f2952fcfd848c88ef95507c3fca411e89d4ac3ff8143"},
-    {file = "librt-0.16.0-cp310-cp310-win32.whl", hash = "sha256:7cc365f006891afb006b52d5ee5ee74c09306ffa20e2f8705a32a4450af2f3ba"},
-    {file = "librt-0.16.0-cp310-cp310-win_amd64.whl", hash = "sha256:0314058469f4d2fd279ce7c62ac274ac82c3918ef7db62ef0697c4c359370155"},
-    {file = "librt-0.16.0-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:fe4372c52d4849096c6cc1cda2817d293ec51440c890474ed59ef38d46556f18"},
-    {file = "librt-0.16.0-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:c72c5295a84bd249526da9bdca38f2e176d15c31c13bb0063c5053f4ca023421"},
-    {file = "librt-0.16.0-cp311-cp311-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:be56ba9c884143495b517f23fe794ae367d58cd89ea0fdd6d437e3c024a87f9f"},
-    {file = "librt-0.16.0-cp311-cp311-manylinux2014_i686.manylinux_2_17_i686.manylinux_2_28_i686.whl", hash = "sha256:ef46c1a29ffb8c72e882e22618ec618778eacd0578fb22c6e7cf9c11d15f357b"},
-    {file = "librt-0.16.0-cp311-cp311-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:d3c94211ee0c4f8d649ec06b7c115c0ec4eadb873a0e3154ca15cef3f814b071"},
-    {file = "librt-0.16.0-cp311-cp311-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:94aed6a8308818b91677957d1bd03188869cd7aeb23c5dba7912a6c0402f7602"},
-    {file = "librt-0.16.0-cp311-cp311-manylinux_2_34_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:349c0bcb87ebd07481b6ff781e25cdc699723dbe2212e57dabb27f7a13b7b87d"},
-    {file = "librt-0.16.0-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:001bfd59a7d45b17e3e75f2a8c6405280b35e7b84471792778e718c4f368950e"},
-    {file = "librt-0.16.0-cp311-cp311-musllinux_1_2_i686.whl", hash = "sha256:28e038895b998d7a0c7798922ce8a1dc157675df5cf1c9ef0aca809ed804b7a1"},
-    {file = "librt-0.16.0-cp311-cp311-musllinux_1_2_ppc64le.whl", hash = "sha256:0dbe4096a7ecc00fa835d24510ad8545a4efef738dac96e0e63516783ccde905"},
-    {file = "librt-0.16.0-cp311-cp311-musllinux_1_2_riscv64.whl", hash = "sha256:5cd5b092441053364af968ea12084692cb9d4a22f3ce9524e377880bf028761e"},
-    {file = "librt-0.16.0-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:3ddeb3c9dedb461bb457c6c7d9aa7fbf35329da313d1a7543d00c8d0f3473c96"},
-    {file = "librt-0.16.0-cp311-cp311-win32.whl", hash = "sha256:e05108e0849966f53a8d2d3112a7af881d0efaa479bc735bba91108f9f2350a7"},
-    {file = "librt-0.16.0-cp311-cp311-win_amd64.whl", hash = "sha256:5f49cff01bd608ef7d97104cb035c75455e79c2d70bf4a506cf773338ac1860d"},
-    {file = "librt-0.16.0-cp311-cp311-win_arm64.whl", hash = "sha256:1d28ae980ae2218f9c5b95d191e947296f918c9bf0b400d467a9430275bbe678"},
-    {file = "librt-0.16.0-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:fe52bf4641069e7978a14253b036cb9002def1926317e710f2e249f8a8c47742"},
-    {file = "librt-0.16.0-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:5bcc2c4726ced915b00de0c9856a4eeabfb3fddb93e10e0b8f735b7709358b6d"},
-    {file = "librt-0.16.0-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:ff7baa55f8e7c69851419e50a666015d02a74198716fd45c0125a2112e0a389f"},
-    {file = "librt-0.16.0-cp312-cp312-manylinux2014_i686.manylinux_2_17_i686.manylinux_2_28_i686.whl", hash = "sha256:b95d5d92ab83d39e760a52091bb1baba664f3a2351e39b1e16801e5747c2f0e9"},
-    {file = "librt-0.16.0-cp312-cp312-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:b6d085d70bce51d43c5c7c36d63490770180d8779e71c49305c87b4213918de7"},
-    {file = "librt-0.16.0-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:36e53948e99bbe3ffea257124cfcae1cfb01831555c9a9c903c9f9a72db7fd07"},
-    {file = "librt-0.16.0-cp312-cp312-manylinux_2_34_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:54d11f726aae9df5a6ffbbf0a03a52449bbac84a53ef03669cb41cdfd4ae41bf"},
-    {file = "librt-0.16.0-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:4323193ac0cd025f85af531df8ba91bf24d1973b401697347a6282e8fd3fcf5e"},
-    {file = "librt-0.16.0-cp312-cp312-musllinux_1_2_i686.whl", hash = "sha256:e42f8e098b9c5396fefa05fb1cc7e33b0e08fc51da106b5de4a45fd22aac6743"},
-    {file = "librt-0.16.0-cp312-cp312-musllinux_1_2_ppc64le.whl", hash = "sha256:39ec1d5a14e37baf1450a6cabf03fe552340808bf1ad9d71824ab90117716459"},
-    {file = "librt-0.16.0-cp312-cp312-musllinux_1_2_riscv64.whl", hash = "sha256:d1aabe3925cbb4a08d15b7b20ba4011b53019da0c4173a25155139b7b1baed65"},
-    {file = "librt-0.16.0-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:300c3ffdc459f4a779a8411ecb188e3ac0b1ff3a3a7b099642555dedae06c69b"},
-    {file = "librt-0.16.0-cp312-cp312-win32.whl", hash = "sha256:c17194318e4c0c0348b36f36c2ec7534436fe0a4c15582403162a4f08c80797a"},
-    {file = "librt-0.16.0-cp312-cp312-win_amd64.whl", hash = "sha256:25a58a19ea8d83b68209f04912df765e9260635ef77646542ed4b4abe6bc7940"},
-    {file = "librt-0.16.0-cp312-cp312-win_arm64.whl", hash = "sha256:f7be7cf555bc30ec12622e9447299cc4a9b8ff307548b634794353db0c2065dc"},
-    {file = "librt-0.16.0-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:c5e6144e68b577f157519f2ba88ca20e3ed61c29b00e5cdfa76cd2d45acf059a"},
-    {file = "librt-0.16.0-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:33f41443a1f4e1f099331b3d8120e409fbff84b9760bc1cc9ea496f37ddaa5cc"},
-    {file = "librt-0.16.0-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:7e510b7770bee609617a3374a96548eb114cae048023e3f049ee449e7ff2db32"},
-    {file = "librt-0.16.0-cp313-cp313-manylinux2014_i686.manylinux_2_17_i686.manylinux_2_28_i686.whl", hash = "sha256:efc49c462d4516b8a58b00b490078fa64689fd1fe66970cc190131d7afb8027e"},
-    {file = "librt-0.16.0-cp313-cp313-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:92caf82ebef5e12d21c72242b70d1e92536f1711cf2a727a4c276de4b4469087"},
-    {file = "librt-0.16.0-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:17bac7f7a16b328fff77e440287693eb017abde913595b5827ebccbc21ecd8a6"},
-    {file = "librt-0.16.0-cp313-cp313-manylinux_2_34_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:5b976054553670829985ed767feb78fb6bcede0175327c4844dd5c281c1be659"},
-    {file = "librt-0.16.0-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:0058f9d68721094105917254c72ac0569117bb7b13b9769cf45d26d89f9d21cd"},
-    {file = "librt-0.16.0-cp313-cp313-musllinux_1_2_i686.whl", hash = "sha256:30b7beaf3f4487b7d8adef1f158b49067cb4d5a19fa7a3bf31a4e7a820e435c5"},
-    {file = "librt-0.16.0-cp313-cp313-musllinux_1_2_ppc64le.whl", hash = "sha256:468df902df016a06eb0e40b0747dc8d14e47d7a38b18b63b1fb167d85cb94d63"},
-    {file = "librt-0.16.0-cp313-cp313-musllinux_1_2_riscv64.whl", hash = "sha256:aea7b1f2b125dad5de85f049136651bff256c883c65e6b9209b2da0a1ac3cdef"},
-    {file = "librt-0.16.0-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:a8afb6557920860b7a3a596eb804cf37e09e7cf8a803db2478c202acc72d8c2e"},
-    {file = "librt-0.16.0-cp313-cp313-pyemscripten_2025_0_wasm32.whl", hash = "sha256:77c7a2b4fe2c1369e0d5aa1cade26740a7b14be32fbc9a5535d617d20065c39d"},
-    {file = "librt-0.16.0-cp313-cp313-win32.whl", hash = "sha256:02d89c813d5ff74b17df72d3a34819d132cd168e56b81bf755b809bd9e46b8c4"},
-    {file = "librt-0.16.0-cp313-cp313-win_amd64.whl", hash = "sha256:14ed6ebe3e4f85f326d7920011ad30ff49ed9334e62cf88caef9ba973d9e3a92"},
-    {file = "librt-0.16.0-cp313-cp313-win_arm64.whl", hash = "sha256:83d4041a3d9b2fd053a8a4e1f22878b3e5833e2712956382d5c048d791454e91"},
-    {file = "librt-0.16.0-cp314-cp314-macosx_10_15_x86_64.whl", hash = "sha256:931a0bb0fcac88f263e269e46eb30ba8e21402cd3c62ca40cb97034c0693fab1"},
-    {file = "librt-0.16.0-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:1bc17e54e5305f8d40b7ca203671ff5a9e59c1d0f8ea0f625dcca53a3984de11"},
-    {file = "librt-0.16.0-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:877698bf6bca5721d8be345f2fe09778e40ecadea8b58c73075f2b1a53666bf2"},
-    {file = "librt-0.16.0-cp314-cp314-manylinux2014_i686.manylinux_2_17_i686.manylinux_2_28_i686.whl", hash = "sha256:5981c011b306781ce561e18e14230a14524a3d8109b97553666c942c18f31a96"},
-    {file = "librt-0.16.0-cp314-cp314-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:afced3dfc17cd805ecf7a3d77996a71cf5f2c75aa66eb0c21a9930f4fc992f86"},
-    {file = "librt-0.16.0-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:ca8052401c55d7511dda6760719fda7618067e83535d7d0010096d216c34b667"},
-    {file = "librt-0.16.0-cp314-cp314-manylinux_2_34_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:1e511762a074005bb0aa569166779834e75e438370226930d0ce1866d4b6a33b"},
-    {file = "librt-0.16.0-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:f1e8591bd8a5a628cd7f07954c6a1592359a878bf032957a8e9057a41d644311"},
-    {file = "librt-0.16.0-cp314-cp314-musllinux_1_2_i686.whl", hash = "sha256:a4aaefb4ba6c07e1aeebb2795c8958148f1d6f9af3b555b53d23d766edb6d67a"},
-    {file = "librt-0.16.0-cp314-cp314-musllinux_1_2_ppc64le.whl", hash = "sha256:d92db7a0f6aee44f1baee94750457e8d2d1c6ccea41842de6268d34e8dc7eddd"},
-    {file = "librt-0.16.0-cp314-cp314-musllinux_1_2_riscv64.whl", hash = "sha256:378dfaffb38e59c24a87cde5713cd865d51ff7383fa12947f3907f306ea1ca55"},
-    {file = "librt-0.16.0-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:3e0c39bdc85370422e8b637be76eb1fd07d30967551b03e62267dd156f553152"},
-    {file = "librt-0.16.0-cp314-cp314-pyemscripten_2026_0_wasm32.whl", hash = "sha256:1b384b90ab79a7bc30b566895809a636e0666f21f3cf12b54823d025b7e83839"},
-    {file = "librt-0.16.0-cp314-cp314-win32.whl", hash = "sha256:52327da75a94012e7f932f913d20d3876bed3c102be00e6c3e8600ff7bdd58a7"},
-    {file = "librt-0.16.0-cp314-cp314-win_amd64.whl", hash = "sha256:3f0b8114c44b2ac06ff5dacd08e07e8e807ff4f46083f2a1602685122559be41"},
-    {file = "librt-0.16.0-cp314-cp314-win_arm64.whl", hash = "sha256:8caf96a4ef8fb27d0ac0d1ad8337d26a240acd4a02fe4345d0a8f264753e8f99"},
-    {file = "librt-0.16.0-cp314-cp314t-macosx_10_15_x86_64.whl", hash = "sha256:953107e2f68d0f3512c48f898b0dbf0ce5cc52bba0f318d847c985dc555ee4cc"},
-    {file = "librt-0.16.0-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:ad37d5b9abd49c9a655dcda7ea52a8a752884062ef1ee71ae17c2f2a0f81fe6a"},
-    {file = "librt-0.16.0-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:2c4aa329c17bd1aaea4f6e89335d8ccd494b3a5830b6654462273e50e11023f0"},
-    {file = "librt-0.16.0-cp314-cp314t-manylinux2014_i686.manylinux_2_17_i686.manylinux_2_28_i686.whl", hash = "sha256:0ead24d2562a49473dddd9efef8581f020007eb0054389c3ee3ffad38b1ca4c9"},
-    {file = "librt-0.16.0-cp314-cp314t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:02118f56a9c36ddd07dfd9b919d9ecc117ba20a90987d56aa4c429fa34509188"},
-    {file = "librt-0.16.0-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:4e29522c62e28595ff7e324c6834ade51127707f0e255b18d1c1cf03d39c1048"},
-    {file = "librt-0.16.0-cp314-cp314t-manylinux_2_34_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:3ff4b2367926b69c6215635902cccb04048e73094e9862900d27cb2c6bbff143"},
-    {file = "librt-0.16.0-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:c6f1b27bf1632a7e016af9f145f82be95e1edd7721a646505c21059257cb5a04"},
-    {file = "librt-0.16.0-cp314-cp314t-musllinux_1_2_i686.whl", hash = "sha256:5696d7f52e7b37217cb3a8f92c744fe835942602fdd4c1a8bc4741d3bfdce15e"},
-    {file = "librt-0.16.0-cp314-cp314t-musllinux_1_2_ppc64le.whl", hash = "sha256:6072e92dd876ff6ceeb6cf371e35e51f479349837391341f479b08df4564242b"},
-    {file = "librt-0.16.0-cp314-cp314t-musllinux_1_2_riscv64.whl", hash = "sha256:39ca4f2f2fe05de8e63493da592d84311adabe5bef52b193851981da9816b302"},
-    {file = "librt-0.16.0-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:f9807485a908f00355820f18e91e045ffdcdc5adb68aaec40a1e2b88c5f7bba1"},
-    {file = "librt-0.16.0-cp314-cp314t-win32.whl", hash = "sha256:94be5cb7bca4df6201f4183e9e4fa2086c655283d20b38cd84500a69057575a7"},
-    {file = "librt-0.16.0-cp314-cp314t-win_amd64.whl", hash = "sha256:d46ca272b251d033dd4527b0dec5f261a28a52bd5fa0f99c117b0a1f8588cc2d"},
-    {file = "librt-0.16.0-cp314-cp314t-win_arm64.whl", hash = "sha256:b9d6d4b14e92d876f8026b54c20c445f36425214c1081dc76f74e40db386b82b"},
-    {file = "librt-0.16.0-cp315-cp315-macosx_10_15_x86_64.whl", hash = "sha256:6fe436af2eaf630474f491af5d032cbe45f93fcff5c3b9fe4ab194a7255b20ff"},
-    {file = "librt-0.16.0-cp315-cp315-macosx_11_0_arm64.whl", hash = "sha256:8ff5d26c529336be9bd7ae04483235d77778ee7d6444a95353102b542601ce81"},
-    {file = "librt-0.16.0-cp315-cp315-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:909d8e3c1faee44cb762b1c519ff8613dcc5ceae5c99987a00917b5a31fd1d6a"},
-    {file = "librt-0.16.0-cp315-cp315-manylinux2014_i686.manylinux_2_17_i686.manylinux_2_28_i686.whl", hash = "sha256:6d4a64283ee61824b5790de882bc68e2d9d7a5143537cb7a966f7354f71646d4"},
-    {file = "librt-0.16.0-cp315-cp315-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:5810ba811297fdf37a1531a57667cb8ace0842013ca8606bf9eb7c24cf4be154"},
-    {file = "librt-0.16.0-cp315-cp315-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:e56aaf8c167548dc8e5d6f3bd0f48dcdd299a23c73be3f744aab79d99e9c7f5d"},
-    {file = "librt-0.16.0-cp315-cp315-manylinux_2_34_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:8f36c58e33b304b525c6c9c5076399c6ebf1109e17b9051a05a407b091b9215b"},
-    {file = "librt-0.16.0-cp315-cp315-musllinux_1_2_aarch64.whl", hash = "sha256:242e00b3d4fa37c3d3c1ca5f5c9adb7d909ddb1eac9c41f2787320d00caa0af2"},
-    {file = "librt-0.16.0-cp315-cp315-musllinux_1_2_i686.whl", hash = "sha256:0253721561787b8df8443eb347b7a6461015354e5bdd37ee38a41fef220d2bb0"},
-    {file = "librt-0.16.0-cp315-cp315-musllinux_1_2_ppc64le.whl", hash = "sha256:3e483a8d69ede8067db70c0e83007423b6925de6fd53afed01d66160f2e9398c"},
-    {file = "librt-0.16.0-cp315-cp315-musllinux_1_2_riscv64.whl", hash = "sha256:69ba927445cfaaffb4081003ef5224c55a5c2ab67ef956f416ef744916e44121"},
-    {file = "librt-0.16.0-cp315-cp315-musllinux_1_2_x86_64.whl", hash = "sha256:d6a365f2ab45a984d0e00eee0dd17f599ceab8cadab6ea07b6111c8132fc0e42"},
-    {file = "librt-0.16.0-cp315-cp315-pyemscripten_2026_5_wasm32.whl", hash = "sha256:f01f3805f2dae4781c0c34b440e31740d082950bdaf89a6f601ad589a28af57a"},
-    {file = "librt-0.16.0-cp315-cp315-win32.whl", hash = "sha256:b0e3e721c75d2e79a76d4422c79d7ba705fe1bbafec907037fe7a657a480a0e3"},
-    {file = "librt-0.16.0-cp315-cp315-win_amd64.whl", hash = "sha256:bc02954b1295de798bbdb0b4e2d8a28c2117de8b5c73dcbeb27dc32572dfb971"},
-    {file = "librt-0.16.0-cp315-cp315-win_arm64.whl", hash = "sha256:c5db585d43449a5f54303d4b2774e45e1babd975cfe1630a3d708c0b80c3e560"},
-    {file = "librt-0.16.0-cp315-cp315t-macosx_10_15_x86_64.whl", hash = "sha256:f06c689cb14afd9b612727553a5ec5a40febf113ca41c4413a2b0b334285884b"},
-    {file = "librt-0.16.0-cp315-cp315t-macosx_11_0_arm64.whl", hash = "sha256:13b4e8aba90b0b1c82474e9844aa9ffe7ad3faa484350e1da64cb8188d903134"},
-    {file = "librt-0.16.0-cp315-cp315t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:5a269c46ae327d8e6f8c1f85f7516cb52c0fa48127565a1105a4f4a05ff2a0b4"},
-    {file = "librt-0.16.0-cp315-cp315t-manylinux2014_i686.manylinux_2_17_i686.manylinux_2_28_i686.whl", hash = "sha256:a33e0dae1f8592146a4764d54ce842b278732d21a84e17c3bbe6b1bc158a2248"},
-    {file = "librt-0.16.0-cp315-cp315t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:47ada6ea32636492c61aa8ad27ae3b9404bfe7a97e3ba946d1984236cc741da0"},
-    {file = "librt-0.16.0-cp315-cp315t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:c43bd6e642d8a248c114327f98dd25ac5a7cb5aa168ef02f0559b91874df16b8"},
-    {file = "librt-0.16.0-cp315-cp315t-manylinux_2_34_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:c3d1bb7841a816ace6449bb26d3f9560dbfa20e71c568d23f0f62bf1e68f50b1"},
-    {file = "librt-0.16.0-cp315-cp315t-musllinux_1_2_aarch64.whl", hash = "sha256:3931f7a3db322e7f44e02a280e3949326ce9579ad388ee8d691dc7c76da9fb70"},
-    {file = "librt-0.16.0-cp315-cp315t-musllinux_1_2_i686.whl", hash = "sha256:f4462528b6000afe8f16907b5c7c2553abf1df005ba5140e6eb394541c3624c3"},
-    {file = "librt-0.16.0-cp315-cp315t-musllinux_1_2_ppc64le.whl", hash = "sha256:80039ba9b6a7d5f1a0175a4cca6bbefead87bd854c80abad1cb30afe47a830db"},
-    {file = "librt-0.16.0-cp315-cp315t-musllinux_1_2_riscv64.whl", hash = "sha256:7a1d272724b581bb6bc769dfdafed6da2ecc9886ba2450311de55a4ac2e1e9cd"},
-    {file = "librt-0.16.0-cp315-cp315t-musllinux_1_2_x86_64.whl", hash = "sha256:fbe4fb8c5445f7496d7f7f6bb0807875d09d47e6771ffa175fb2df2895fb86ba"},
-    {file = "librt-0.16.0-cp315-cp315t-win32.whl", hash = "sha256:375bfe6b572a8f6cfc398709356046173bf27e64c4c5edaf5f7062f051fb4bf9"},
-    {file = "librt-0.16.0-cp315-cp315t-win_amd64.whl", hash = "sha256:bd3150023d3dc2bc70f3784e59ffa1140d56ddba3d8125b3d6f9f85221279bfc"},
-    {file = "librt-0.16.0-cp315-cp315t-win_arm64.whl", hash = "sha256:8ceafb70f2a4f0826f11031942e59c0728fd98da112dc346d4352bde1e486866"},
-    {file = "librt-0.16.0-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:e1967e36ac4cae0c7e9615ad32e1a513cdacff79f9e8afb28bedc91caf48b4f3"},
-    {file = "librt-0.16.0-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:20fe0bf9053885e21c62b3e091fb5e73e1c54d1daf2e70eb388d70763bcd4220"},
-    {file = "librt-0.16.0-cp39-cp39-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:6a63610fa76524edfa605b5b259a603915c7a6e10e54f003e5503030506e81de"},
-    {file = "librt-0.16.0-cp39-cp39-manylinux2014_i686.manylinux_2_17_i686.manylinux_2_28_i686.whl", hash = "sha256:845a511b60ca43b9880dcc84a9784c891d6a2098c829130b320846c69c9c0c68"},
-    {file = "librt-0.16.0-cp39-cp39-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:d608f0bf3b8cbddd0067fe02cb8cab7d13e8eb9386b23a1843d4363044fd9e22"},
-    {file = "librt-0.16.0-cp39-cp39-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:df183721229ae51eef90108c115b43e98cb169b6155d34480338e5fc6616df00"},
-    {file = "librt-0.16.0-cp39-cp39-manylinux_2_34_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:4b6183e2e2e0ee00aac2ec07c7f7d151c97e85666b304f574b71cff0f9fccc4e"},
-    {file = "librt-0.16.0-cp39-cp39-musllinux_1_2_aarch64.whl", hash = "sha256:273d00be33792a15189331df10f1f1331621b043881e66c6c4377f7f776e1291"},
-    {file = "librt-0.16.0-cp39-cp39-musllinux_1_2_i686.whl", hash = "sha256:aa9357a1b4d4fc787bb718a59cb1112c28c8a976d6bfa268b71cc0a4ab8f3a94"},
-    {file = "librt-0.16.0-cp39-cp39-musllinux_1_2_ppc64le.whl", hash = "sha256:e9ce0bc440e7fd09b5f51f372f0f6640658f854b8fb05260f819cc669c93c42d"},
-    {file = "librt-0.16.0-cp39-cp39-musllinux_1_2_riscv64.whl", hash = "sha256:6c8893eae2fd13c5488d94056f3e6e5cf3142bfb1c4acaf136cb33d760c5964b"},
-    {file = "librt-0.16.0-cp39-cp39-musllinux_1_2_x86_64.whl", hash = "sha256:7393c9a48dcce4817dbd4b0d8ff6237efe9b0a0609f5b0adaef315f8541726b5"},
-    {file = "librt-0.16.0-cp39-cp39-win32.whl", hash = "sha256:71b93b42784e25b975079573c642a8fedb049a7bb31d70a51721b1666b3b2ced"},
-    {file = "librt-0.16.0-cp39-cp39-win_amd64.whl", hash = "sha256:5750a105b42a416f930edc59054927a406effb2550cd5bab92ad7a5842ed5d05"},
-    {file = "librt-0.16.0.tar.gz", hash = "sha256:ac38d6d8d66bf3d744148dbbc0b8e193e195a51e364ed55e224631f5721891fc"},
-]
-
-[[package]]
-name = "mypy"
-version = "1.19.1"
-description = "Optional static typing for Python"
-optional = false
-python-versions = ">=3.9"
-groups = ["dev"]
-files = [
-    {file = "mypy-1.19.1-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:5f05aa3d375b385734388e844bc01733bd33c644ab48e9684faa54e5389775ec"},
-    {file = "mypy-1.19.1-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:022ea7279374af1a5d78dfcab853fe6a536eebfda4b59deab53cd21f6cd9f00b"},
-    {file = "mypy-1.19.1-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:ee4c11e460685c3e0c64a4c5de82ae143622410950d6be863303a1c4ba0e36d6"},
-    {file = "mypy-1.19.1-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:de759aafbae8763283b2ee5869c7255391fbc4de3ff171f8f030b5ec48381b74"},
-    {file = "mypy-1.19.1-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:ab43590f9cd5108f41aacf9fca31841142c786827a74ab7cc8a2eacb634e09a1"},
-    {file = "mypy-1.19.1-cp310-cp310-win_amd64.whl", hash = "sha256:2899753e2f61e571b3971747e302d5f420c3fd09650e1951e99f823bc3089dac"},
-    {file = "mypy-1.19.1-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:d8dfc6ab58ca7dda47d9237349157500468e404b17213d44fc1cb77bce532288"},
-    {file = "mypy-1.19.1-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:e3f276d8493c3c97930e354b2595a44a21348b320d859fb4a2b9f66da9ed27ab"},
-    {file = "mypy-1.19.1-cp311-cp311-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:2abb24cf3f17864770d18d673c85235ba52456b36a06b6afc1e07c1fdcd3d0e6"},
-    {file = "mypy-1.19.1-cp311-cp311-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:a009ffa5a621762d0c926a078c2d639104becab69e79538a494bcccb62cc0331"},
-    {file = "mypy-1.19.1-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:f7cee03c9a2e2ee26ec07479f38ea9c884e301d42c6d43a19d20fb014e3ba925"},
-    {file = "mypy-1.19.1-cp311-cp311-win_amd64.whl", hash = "sha256:4b84a7a18f41e167f7995200a1d07a4a6810e89d29859df936f1c3923d263042"},
-    {file = "mypy-1.19.1-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:a8174a03289288c1f6c46d55cef02379b478bfbc8e358e02047487cad44c6ca1"},
-    {file = "mypy-1.19.1-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:ffcebe56eb09ff0c0885e750036a095e23793ba6c2e894e7e63f6d89ad51f22e"},
-    {file = "mypy-1.19.1-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:b64d987153888790bcdb03a6473d321820597ab8dd9243b27a92153c4fa50fd2"},
-    {file = "mypy-1.19.1-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:c35d298c2c4bba75feb2195655dfea8124d855dfd7343bf8b8c055421eaf0cf8"},
-    {file = "mypy-1.19.1-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:34c81968774648ab5ac09c29a375fdede03ba253f8f8287847bd480782f73a6a"},
-    {file = "mypy-1.19.1-cp312-cp312-win_amd64.whl", hash = "sha256:b10e7c2cd7870ba4ad9b2d8a6102eb5ffc1f16ca35e3de6bfa390c1113029d13"},
-    {file = "mypy-1.19.1-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:e3157c7594ff2ef1634ee058aafc56a82db665c9438fd41b390f3bde1ab12250"},
-    {file = "mypy-1.19.1-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:bdb12f69bcc02700c2b47e070238f42cb87f18c0bc1fc4cdb4fb2bc5fd7a3b8b"},
-    {file = "mypy-1.19.1-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:f859fb09d9583a985be9a493d5cfc5515b56b08f7447759a0c5deaf68d80506e"},
-    {file = "mypy-1.19.1-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:c9a6538e0415310aad77cb94004ca6482330fece18036b5f360b62c45814c4ef"},
-    {file = "mypy-1.19.1-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:da4869fc5e7f62a88f3fe0b5c919d1d9f7ea3cef92d3689de2823fd27e40aa75"},
-    {file = "mypy-1.19.1-cp313-cp313-win_amd64.whl", hash = "sha256:016f2246209095e8eda7538944daa1d60e1e8134d98983b9fc1e92c1fc0cb8dd"},
-    {file = "mypy-1.19.1-cp314-cp314-macosx_10_15_x86_64.whl", hash = "sha256:06e6170bd5836770e8104c8fdd58e5e725cfeb309f0a6c681a811f557e97eac1"},
-    {file = "mypy-1.19.1-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:804bd67b8054a85447c8954215a906d6eff9cabeabe493fb6334b24f4bfff718"},
-    {file = "mypy-1.19.1-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:21761006a7f497cb0d4de3d8ef4ca70532256688b0523eee02baf9eec895e27b"},
-    {file = "mypy-1.19.1-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:28902ee51f12e0f19e1e16fbe2f8f06b6637f482c459dd393efddd0ec7f82045"},
-    {file = "mypy-1.19.1-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:481daf36a4c443332e2ae9c137dfee878fcea781a2e3f895d54bd3002a900957"},
-    {file = "mypy-1.19.1-cp314-cp314-win_amd64.whl", hash = "sha256:8bb5c6f6d043655e055be9b542aa5f3bdd30e4f3589163e85f93f3640060509f"},
-    {file = "mypy-1.19.1-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:7bcfc336a03a1aaa26dfce9fff3e287a3ba99872a157561cbfcebe67c13308e3"},
-    {file = "mypy-1.19.1-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:b7951a701c07ea584c4fe327834b92a30825514c868b1f69c30445093fdd9d5a"},
-    {file = "mypy-1.19.1-cp39-cp39-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:b13cfdd6c87fc3efb69ea4ec18ef79c74c3f98b4e5498ca9b85ab3b2c2329a67"},
-    {file = "mypy-1.19.1-cp39-cp39-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:4f28f99c824ecebcdaa2e55d82953e38ff60ee5ec938476796636b86afa3956e"},
-    {file = "mypy-1.19.1-cp39-cp39-musllinux_1_2_x86_64.whl", hash = "sha256:c608937067d2fc5a4dd1a5ce92fd9e1398691b8c5d012d66e1ddd430e9244376"},
-    {file = "mypy-1.19.1-cp39-cp39-win_amd64.whl", hash = "sha256:409088884802d511ee52ca067707b90c883426bd95514e8cfda8281dc2effe24"},
-    {file = "mypy-1.19.1-py3-none-any.whl", hash = "sha256:f1235f5ea01b7db5468d53ece6aaddf1ad0b88d9e7462b86ef96fe04995d7247"},
-    {file = "mypy-1.19.1.tar.gz", hash = "sha256:19d88bb05303fe63f71dd2c6270daca27cb9401c4ca8255fe50d1d920e0eb9ba"},
-]
-
-[package.dependencies]
-librt = {version = ">=0.6.2", markers = "platform_python_implementation != \"PyPy\""}
-mypy_extensions = ">=1.0.0"
-pathspec = ">=0.9.0"
-typing_extensions = ">=4.6.0"
-
-[package.extras]
-dmypy = ["psutil (>=4.0)"]
-faster-cache = ["orjson"]
-install-types = ["pip"]
-mypyc = ["setuptools (>=50)"]
-reports = ["lxml"]
-
-[[package]]
-name = "mypy-extensions"
-version = "1.1.0"
-description = "Type system extensions for programs checked with the mypy type checker."
-optional = false
-python-versions = ">=3.8"
-groups = ["dev"]
-files = [
-    {file = "mypy_extensions-1.1.0-py3-none-any.whl", hash = "sha256:1be4cccdb0f2482337c4743e60421de3a356cd97508abadd57d47403e94f5505"},
-    {file = "mypy_extensions-1.1.0.tar.gz", hash = "sha256:52e68efc3284861e772bbcd66823fde5ae21fd2fdb51c62a211403730b916558"},
-]
-
-[[package]]
-name = "nodeenv"
-version = "1.11.0"
-description = "Node.js virtual environment builder"
-optional = false
-python-versions = "!=3.0.*,!=3.1.*,!=3.2.*,!=3.3.*,!=3.4.*,!=3.5.*,!=3.6.*,>=2.7"
-groups = ["dev"]
-files = [
-    {file = "nodeenv-1.11.0-py2.py3-none-any.whl", hash = "sha256:edaa16e6c14d7cf395d75d4bbd5a26390f4dc06501a33b4e76282b02cc688a25"},
-    {file = "nodeenv-1.11.0.tar.gz", hash = "sha256:3ce8fe5b71d16e8af7039ca65257354100bc772965d6bc549070649e53b1b146"},
-]
-
-[package.extras]
-certifi = ["certifi"]
-
-[[package]]
-name = "packaging"
-version = "26.3"
-description = "Core utilities for Python packages"
-optional = false
-python-versions = ">=3.9"
-groups = ["dev"]
-files = [
-    {file = "packaging-26.3-py3-none-any.whl", hash = "sha256:d7193f7c8e4e93f444fde0262bf90af30e16fa0ad0ad44cb553c87339b23cd1c"},
-    {file = "packaging-26.3.tar.gz", hash = "sha256:94edc256424af38762eb31306eed28beb9f0efc50a8837492c9d6fd6004aed79"},
-]
-
-[[package]]
-name = "paddler-client"
-version = "4.0.1"
-description = "Python client for Paddler LLM load balancer"
-optional = false
-python-versions = "^3.11"
-groups = ["main"]
-files = []
-develop = true
-
-[package.dependencies]
-httpx = "==0.28.1"
-pydantic = "==2.12.5"
-websockets = "==15.0.1"
-
-[package.source]
-type = "directory"
-url = "../paddler_client_python"
-
-[[package]]
-name = "pathspec"
-version = "1.1.1"
-description = "Utility library for gitignore style pattern matching of file paths."
-optional = false
-python-versions = ">=3.9"
-groups = ["dev"]
-files = [
-    {file = "pathspec-1.1.1-py3-none-any.whl", hash = "sha256:a00ce642f577bf7f473932318056212bc4f8bfdf53128c78bbd5af0b9b20b189"},
-    {file = "pathspec-1.1.1.tar.gz", hash = "sha256:17db5ecd524104a120e173814c90367a96a98d07c45b2e10c2f3919fff91bf5a"},
-]
-
-[package.extras]
-hyperscan = ["hyperscan (>=0.7)"]
-optional = ["typing-extensions (>=4)"]
-re2 = ["google-re2 (>=1.1)"]
-
-[[package]]
-name = "pluggy"
-version = "1.6.0"
-description = "plugin and hook calling mechanisms for python"
-optional = false
-python-versions = ">=3.9"
-groups = ["dev"]
-files = [
-    {file = "pluggy-1.6.0-py3-none-any.whl", hash = "sha256:e920276dd6813095e9377c0bc5566d94c932c33b27a3e3945d8389c374dd4746"},
-    {file = "pluggy-1.6.0.tar.gz", hash = "sha256:7dcc130b76258d33b90f61b658791dede3486c3e6bfb003ee5c9bfb396dd22f3"},
-]
-
-[package.extras]
-dev = ["pre-commit", "tox"]
-testing = ["coverage", "pytest", "pytest-benchmark"]
-
-[[package]]
-name = "pydantic"
-version = "2.12.5"
-description = "Data validation using Python type hints"
-optional = false
-python-versions = ">=3.9"
-groups = ["main"]
-files = [
-    {file = "pydantic-2.12.5-py3-none-any.whl", hash = "sha256:e561593fccf61e8a20fc46dfc2dfe075b8be7d0188df33f221ad1f0139180f9d"},
-    {file = "pydantic-2.12.5.tar.gz", hash = "sha256:4d351024c75c0f085a9febbb665ce8c0c6ec5d30e903bdb6394b7ede26aebb49"},
-]
-
-[package.dependencies]
-annotated-types = ">=0.6.0"
-pydantic-core = "2.41.5"
-typing-extensions = ">=4.14.1"
-typing-inspection = ">=0.4.2"
-
-[package.extras]
-email = ["email-validator (>=2.0.0)"]
-timezone = ["tzdata ; python_version >= \"3.9\" and platform_system == \"Windows\""]
-
-[[package]]
-name = "pydantic-core"
-version = "2.41.5"
-description = "Core functionality for Pydantic validation and serialization"
-optional = false
-python-versions = ">=3.9"
-groups = ["main"]
-files = [
-    {file = "pydantic_core-2.41.5-cp310-cp310-macosx_10_12_x86_64.whl", hash = "sha256:77b63866ca88d804225eaa4af3e664c5faf3568cea95360d21f4725ab6e07146"},
-    {file = "pydantic_core-2.41.5-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:dfa8a0c812ac681395907e71e1274819dec685fec28273a28905df579ef137e2"},
-    {file = "pydantic_core-2.41.5-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:5921a4d3ca3aee735d9fd163808f5e8dd6c6972101e4adbda9a4667908849b97"},
-    {file = "pydantic_core-2.41.5-cp310-cp310-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:e25c479382d26a2a41b7ebea1043564a937db462816ea07afa8a44c0866d52f9"},
-    {file = "pydantic_core-2.41.5-cp310-cp310-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:f547144f2966e1e16ae626d8ce72b4cfa0caedc7fa28052001c94fb2fcaa1c52"},
-    {file = "pydantic_core-2.41.5-cp310-cp310-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:6f52298fbd394f9ed112d56f3d11aabd0d5bd27beb3084cc3d8ad069483b8941"},
-    {file = "pydantic_core-2.41.5-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:100baa204bb412b74fe285fb0f3a385256dad1d1879f0a5cb1499ed2e83d132a"},
-    {file = "pydantic_core-2.41.5-cp310-cp310-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:05a2c8852530ad2812cb7914dc61a1125dc4e06252ee98e5638a12da6cc6fb6c"},
-    {file = "pydantic_core-2.41.5-cp310-cp310-musllinux_1_1_aarch64.whl", hash = "sha256:29452c56df2ed968d18d7e21f4ab0ac55e71dc59524872f6fc57dcf4a3249ed2"},
-    {file = "pydantic_core-2.41.5-cp310-cp310-musllinux_1_1_armv7l.whl", hash = "sha256:d5160812ea7a8a2ffbe233d8da666880cad0cbaf5d4de74ae15c313213d62556"},
-    {file = "pydantic_core-2.41.5-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:df3959765b553b9440adfd3c795617c352154e497a4eaf3752555cfb5da8fc49"},
-    {file = "pydantic_core-2.41.5-cp310-cp310-win32.whl", hash = "sha256:1f8d33a7f4d5a7889e60dc39856d76d09333d8a6ed0f5f1190635cbec70ec4ba"},
-    {file = "pydantic_core-2.41.5-cp310-cp310-win_amd64.whl", hash = "sha256:62de39db01b8d593e45871af2af9e497295db8d73b085f6bfd0b18c83c70a8f9"},
-    {file = "pydantic_core-2.41.5-cp311-cp311-macosx_10_12_x86_64.whl", hash = "sha256:a3a52f6156e73e7ccb0f8cced536adccb7042be67cb45f9562e12b319c119da6"},
-    {file = "pydantic_core-2.41.5-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:7f3bf998340c6d4b0c9a2f02d6a400e51f123b59565d74dc60d252ce888c260b"},
-    {file = "pydantic_core-2.41.5-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:378bec5c66998815d224c9ca994f1e14c0c21cb95d2f52b6021cc0b2a58f2a5a"},
-    {file = "pydantic_core-2.41.5-cp311-cp311-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:e7b576130c69225432866fe2f4a469a85a54ade141d96fd396dffcf607b558f8"},
-    {file = "pydantic_core-2.41.5-cp311-cp311-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:6cb58b9c66f7e4179a2d5e0f849c48eff5c1fca560994d6eb6543abf955a149e"},
-    {file = "pydantic_core-2.41.5-cp311-cp311-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:88942d3a3dff3afc8288c21e565e476fc278902ae4d6d134f1eeda118cc830b1"},
-    {file = "pydantic_core-2.41.5-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:f31d95a179f8d64d90f6831d71fa93290893a33148d890ba15de25642c5d075b"},
-    {file = "pydantic_core-2.41.5-cp311-cp311-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:c1df3d34aced70add6f867a8cf413e299177e0c22660cc767218373d0779487b"},
-    {file = "pydantic_core-2.41.5-cp311-cp311-musllinux_1_1_aarch64.whl", hash = "sha256:4009935984bd36bd2c774e13f9a09563ce8de4abaa7226f5108262fa3e637284"},
-    {file = "pydantic_core-2.41.5-cp311-cp311-musllinux_1_1_armv7l.whl", hash = "sha256:34a64bc3441dc1213096a20fe27e8e128bd3ff89921706e83c0b1ac971276594"},
-    {file = "pydantic_core-2.41.5-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:c9e19dd6e28fdcaa5a1de679aec4141f691023916427ef9bae8584f9c2fb3b0e"},
-    {file = "pydantic_core-2.41.5-cp311-cp311-win32.whl", hash = "sha256:2c010c6ded393148374c0f6f0bf89d206bf3217f201faa0635dcd56bd1520f6b"},
-    {file = "pydantic_core-2.41.5-cp311-cp311-win_amd64.whl", hash = "sha256:76ee27c6e9c7f16f47db7a94157112a2f3a00e958bc626e2f4ee8bec5c328fbe"},
-    {file = "pydantic_core-2.41.5-cp311-cp311-win_arm64.whl", hash = "sha256:4bc36bbc0b7584de96561184ad7f012478987882ebf9f9c389b23f432ea3d90f"},
-    {file = "pydantic_core-2.41.5-cp312-cp312-macosx_10_12_x86_64.whl", hash = "sha256:f41a7489d32336dbf2199c8c0a215390a751c5b014c2c1c5366e817202e9cdf7"},
-    {file = "pydantic_core-2.41.5-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:070259a8818988b9a84a449a2a7337c7f430a22acc0859c6b110aa7212a6d9c0"},
-    {file = "pydantic_core-2.41.5-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:e96cea19e34778f8d59fe40775a7a574d95816eb150850a85a7a4c8f4b94ac69"},
-    {file = "pydantic_core-2.41.5-cp312-cp312-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:ed2e99c456e3fadd05c991f8f437ef902e00eedf34320ba2b0842bd1c3ca3a75"},
-    {file = "pydantic_core-2.41.5-cp312-cp312-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:65840751b72fbfd82c3c640cff9284545342a4f1eb1586ad0636955b261b0b05"},
-    {file = "pydantic_core-2.41.5-cp312-cp312-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:e536c98a7626a98feb2d3eaf75944ef6f3dbee447e1f841eae16f2f0a72d8ddc"},
-    {file = "pydantic_core-2.41.5-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:eceb81a8d74f9267ef4081e246ffd6d129da5d87e37a77c9bde550cb04870c1c"},
-    {file = "pydantic_core-2.41.5-cp312-cp312-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:d38548150c39b74aeeb0ce8ee1d8e82696f4a4e16ddc6de7b1d8823f7de4b9b5"},
-    {file = "pydantic_core-2.41.5-cp312-cp312-musllinux_1_1_aarch64.whl", hash = "sha256:c23e27686783f60290e36827f9c626e63154b82b116d7fe9adba1fda36da706c"},
-    {file = "pydantic_core-2.41.5-cp312-cp312-musllinux_1_1_armv7l.whl", hash = "sha256:482c982f814460eabe1d3bb0adfdc583387bd4691ef00b90575ca0d2b6fe2294"},
-    {file = "pydantic_core-2.41.5-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:bfea2a5f0b4d8d43adf9d7b8bf019fb46fdd10a2e5cde477fbcb9d1fa08c68e1"},
-    {file = "pydantic_core-2.41.5-cp312-cp312-win32.whl", hash = "sha256:b74557b16e390ec12dca509bce9264c3bbd128f8a2c376eaa68003d7f327276d"},
-    {file = "pydantic_core-2.41.5-cp312-cp312-win_amd64.whl", hash = "sha256:1962293292865bca8e54702b08a4f26da73adc83dd1fcf26fbc875b35d81c815"},
-    {file = "pydantic_core-2.41.5-cp312-cp312-win_arm64.whl", hash = "sha256:1746d4a3d9a794cacae06a5eaaccb4b8643a131d45fbc9af23e353dc0a5ba5c3"},
-    {file = "pydantic_core-2.41.5-cp313-cp313-macosx_10_12_x86_64.whl", hash = "sha256:941103c9be18ac8daf7b7adca8228f8ed6bb7a1849020f643b3a14d15b1924d9"},
-    {file = "pydantic_core-2.41.5-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:112e305c3314f40c93998e567879e887a3160bb8689ef3d2c04b6cc62c33ac34"},
-    {file = "pydantic_core-2.41.5-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:0cbaad15cb0c90aa221d43c00e77bb33c93e8d36e0bf74760cd00e732d10a6a0"},
-    {file = "pydantic_core-2.41.5-cp313-cp313-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:03ca43e12fab6023fc79d28ca6b39b05f794ad08ec2feccc59a339b02f2b3d33"},
-    {file = "pydantic_core-2.41.5-cp313-cp313-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:dc799088c08fa04e43144b164feb0c13f9a0bc40503f8df3e9fde58a3c0c101e"},
-    {file = "pydantic_core-2.41.5-cp313-cp313-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:97aeba56665b4c3235a0e52b2c2f5ae9cd071b8a8310ad27bddb3f7fb30e9aa2"},
-    {file = "pydantic_core-2.41.5-cp313-cp313-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:406bf18d345822d6c21366031003612b9c77b3e29ffdb0f612367352aab7d586"},
-    {file = "pydantic_core-2.41.5-cp313-cp313-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:b93590ae81f7010dbe380cdeab6f515902ebcbefe0b9327cc4804d74e93ae69d"},
-    {file = "pydantic_core-2.41.5-cp313-cp313-musllinux_1_1_aarch64.whl", hash = "sha256:01a3d0ab748ee531f4ea6c3e48ad9dac84ddba4b0d82291f87248f2f9de8d740"},
-    {file = "pydantic_core-2.41.5-cp313-cp313-musllinux_1_1_armv7l.whl", hash = "sha256:6561e94ba9dacc9c61bce40e2d6bdc3bfaa0259d3ff36ace3b1e6901936d2e3e"},
-    {file = "pydantic_core-2.41.5-cp313-cp313-musllinux_1_1_x86_64.whl", hash = "sha256:915c3d10f81bec3a74fbd4faebe8391013ba61e5a1a8d48c4455b923bdda7858"},
-    {file = "pydantic_core-2.41.5-cp313-cp313-win32.whl", hash = "sha256:650ae77860b45cfa6e2cdafc42618ceafab3a2d9a3811fcfbd3bbf8ac3c40d36"},
-    {file = "pydantic_core-2.41.5-cp313-cp313-win_amd64.whl", hash = "sha256:79ec52ec461e99e13791ec6508c722742ad745571f234ea6255bed38c6480f11"},
-    {file = "pydantic_core-2.41.5-cp313-cp313-win_arm64.whl", hash = "sha256:3f84d5c1b4ab906093bdc1ff10484838aca54ef08de4afa9de0f5f14d69639cd"},
-    {file = "pydantic_core-2.41.5-cp314-cp314-macosx_10_12_x86_64.whl", hash = "sha256:3f37a19d7ebcdd20b96485056ba9e8b304e27d9904d233d7b1015db320e51f0a"},
-    {file = "pydantic_core-2.41.5-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:1d1d9764366c73f996edd17abb6d9d7649a7eb690006ab6adbda117717099b14"},
-    {file = "pydantic_core-2.41.5-cp314-cp314-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:25e1c2af0fce638d5f1988b686f3b3ea8cd7de5f244ca147c777769e798a9cd1"},
-    {file = "pydantic_core-2.41.5-cp314-cp314-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:506d766a8727beef16b7adaeb8ee6217c64fc813646b424d0804d67c16eddb66"},
-    {file = "pydantic_core-2.41.5-cp314-cp314-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:4819fa52133c9aa3c387b3328f25c1facc356491e6135b459f1de698ff64d869"},
-    {file = "pydantic_core-2.41.5-cp314-cp314-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:2b761d210c9ea91feda40d25b4efe82a1707da2ef62901466a42492c028553a2"},
-    {file = "pydantic_core-2.41.5-cp314-cp314-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:22f0fb8c1c583a3b6f24df2470833b40207e907b90c928cc8d3594b76f874375"},
-    {file = "pydantic_core-2.41.5-cp314-cp314-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:2782c870e99878c634505236d81e5443092fba820f0373997ff75f90f68cd553"},
-    {file = "pydantic_core-2.41.5-cp314-cp314-musllinux_1_1_aarch64.whl", hash = "sha256:0177272f88ab8312479336e1d777f6b124537d47f2123f89cb37e0accea97f90"},
-    {file = "pydantic_core-2.41.5-cp314-cp314-musllinux_1_1_armv7l.whl", hash = "sha256:63510af5e38f8955b8ee5687740d6ebf7c2a0886d15a6d65c32814613681bc07"},
-    {file = "pydantic_core-2.41.5-cp314-cp314-musllinux_1_1_x86_64.whl", hash = "sha256:e56ba91f47764cc14f1daacd723e3e82d1a89d783f0f5afe9c364b8bb491ccdb"},
-    {file = "pydantic_core-2.41.5-cp314-cp314-win32.whl", hash = "sha256:aec5cf2fd867b4ff45b9959f8b20ea3993fc93e63c7363fe6851424c8a7e7c23"},
-    {file = "pydantic_core-2.41.5-cp314-cp314-win_amd64.whl", hash = "sha256:8e7c86f27c585ef37c35e56a96363ab8de4e549a95512445b85c96d3e2f7c1bf"},
-    {file = "pydantic_core-2.41.5-cp314-cp314-win_arm64.whl", hash = "sha256:e672ba74fbc2dc8eea59fb6d4aed6845e6905fc2a8afe93175d94a83ba2a01a0"},
-    {file = "pydantic_core-2.41.5-cp314-cp314t-macosx_10_12_x86_64.whl", hash = "sha256:8566def80554c3faa0e65ac30ab0932b9e3a5cd7f8323764303d468e5c37595a"},
-    {file = "pydantic_core-2.41.5-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:b80aa5095cd3109962a298ce14110ae16b8c1aece8b72f9dafe81cf597ad80b3"},
-    {file = "pydantic_core-2.41.5-cp314-cp314t-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:3006c3dd9ba34b0c094c544c6006cc79e87d8612999f1a5d43b769b89181f23c"},
-    {file = "pydantic_core-2.41.5-cp314-cp314t-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:72f6c8b11857a856bcfa48c86f5368439f74453563f951e473514579d44aa612"},
-    {file = "pydantic_core-2.41.5-cp314-cp314t-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:5cb1b2f9742240e4bb26b652a5aeb840aa4b417c7748b6f8387927bc6e45e40d"},
-    {file = "pydantic_core-2.41.5-cp314-cp314t-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:bd3d54f38609ff308209bd43acea66061494157703364ae40c951f83ba99a1a9"},
-    {file = "pydantic_core-2.41.5-cp314-cp314t-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:2ff4321e56e879ee8d2a879501c8e469414d948f4aba74a2d4593184eb326660"},
-    {file = "pydantic_core-2.41.5-cp314-cp314t-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:d0d2568a8c11bf8225044aa94409e21da0cb09dcdafe9ecd10250b2baad531a9"},
-    {file = "pydantic_core-2.41.5-cp314-cp314t-musllinux_1_1_aarch64.whl", hash = "sha256:a39455728aabd58ceabb03c90e12f71fd30fa69615760a075b9fec596456ccc3"},
-    {file = "pydantic_core-2.41.5-cp314-cp314t-musllinux_1_1_armv7l.whl", hash = "sha256:239edca560d05757817c13dc17c50766136d21f7cd0fac50295499ae24f90fdf"},
-    {file = "pydantic_core-2.41.5-cp314-cp314t-musllinux_1_1_x86_64.whl", hash = "sha256:2a5e06546e19f24c6a96a129142a75cee553cc018ffee48a460059b1185f4470"},
-    {file = "pydantic_core-2.41.5-cp314-cp314t-win32.whl", hash = "sha256:b4ececa40ac28afa90871c2cc2b9ffd2ff0bf749380fbdf57d165fd23da353aa"},
-    {file = "pydantic_core-2.41.5-cp314-cp314t-win_amd64.whl", hash = "sha256:80aa89cad80b32a912a65332f64a4450ed00966111b6615ca6816153d3585a8c"},
-    {file = "pydantic_core-2.41.5-cp314-cp314t-win_arm64.whl", hash = "sha256:35b44f37a3199f771c3eaa53051bc8a70cd7b54f333531c59e29fd4db5d15008"},
-    {file = "pydantic_core-2.41.5-cp39-cp39-macosx_10_12_x86_64.whl", hash = "sha256:8bfeaf8735be79f225f3fefab7f941c712aaca36f1128c9d7e2352ee1aa87bdf"},
-    {file = "pydantic_core-2.41.5-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:346285d28e4c8017da95144c7f3acd42740d637ff41946af5ce6e5e420502dd5"},
-    {file = "pydantic_core-2.41.5-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:a75dafbf87d6276ddc5b2bf6fae5254e3d0876b626eb24969a574fff9149ee5d"},
-    {file = "pydantic_core-2.41.5-cp39-cp39-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:7b93a4d08587e2b7e7882de461e82b6ed76d9026ce91ca7915e740ecc7855f60"},
-    {file = "pydantic_core-2.41.5-cp39-cp39-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:e8465ab91a4bd96d36dde3263f06caa6a8a6019e4113f24dc753d79a8b3a3f82"},
-    {file = "pydantic_core-2.41.5-cp39-cp39-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:299e0a22e7ae2b85c1a57f104538b2656e8ab1873511fd718a1c1c6f149b77b5"},
-    {file = "pydantic_core-2.41.5-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:707625ef0983fcfb461acfaf14de2067c5942c6bb0f3b4c99158bed6fedd3cf3"},
-    {file = "pydantic_core-2.41.5-cp39-cp39-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:f41eb9797986d6ebac5e8edff36d5cef9de40def462311b3eb3eeded1431e425"},
-    {file = "pydantic_core-2.41.5-cp39-cp39-musllinux_1_1_aarch64.whl", hash = "sha256:0384e2e1021894b1ff5a786dbf94771e2986ebe2869533874d7e43bc79c6f504"},
-    {file = "pydantic_core-2.41.5-cp39-cp39-musllinux_1_1_armv7l.whl", hash = "sha256:f0cd744688278965817fd0839c4a4116add48d23890d468bc436f78beb28abf5"},
-    {file = "pydantic_core-2.41.5-cp39-cp39-musllinux_1_1_x86_64.whl", hash = "sha256:753e230374206729bf0a807954bcc6c150d3743928a73faffee51ac6557a03c3"},
-    {file = "pydantic_core-2.41.5-cp39-cp39-win32.whl", hash = "sha256:873e0d5b4fb9b89ef7c2d2a963ea7d02879d9da0da8d9d4933dee8ee86a8b460"},
-    {file = "pydantic_core-2.41.5-cp39-cp39-win_amd64.whl", hash = "sha256:e4f4a984405e91527a0d62649ee21138f8e3d0ef103be488c1dc11a80d7f184b"},
-    {file = "pydantic_core-2.41.5-graalpy311-graalpy242_311_native-macosx_10_12_x86_64.whl", hash = "sha256:b96d5f26b05d03cc60f11a7761a5ded1741da411e7fe0909e27a5e6a0cb7b034"},
-    {file = "pydantic_core-2.41.5-graalpy311-graalpy242_311_native-macosx_11_0_arm64.whl", hash = "sha256:634e8609e89ceecea15e2d61bc9ac3718caaaa71963717bf3c8f38bfde64242c"},
-    {file = "pydantic_core-2.41.5-graalpy311-graalpy242_311_native-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:93e8740d7503eb008aa2df04d3b9735f845d43ae845e6dcd2be0b55a2da43cd2"},
-    {file = "pydantic_core-2.41.5-graalpy311-graalpy242_311_native-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:f15489ba13d61f670dcc96772e733aad1a6f9c429cc27574c6cdaed82d0146ad"},
-    {file = "pydantic_core-2.41.5-graalpy312-graalpy250_312_native-macosx_10_12_x86_64.whl", hash = "sha256:7da7087d756b19037bc2c06edc6c170eeef3c3bafcb8f532ff17d64dc427adfd"},
-    {file = "pydantic_core-2.41.5-graalpy312-graalpy250_312_native-macosx_11_0_arm64.whl", hash = "sha256:aabf5777b5c8ca26f7824cb4a120a740c9588ed58df9b2d196ce92fba42ff8dc"},
-    {file = "pydantic_core-2.41.5-graalpy312-graalpy250_312_native-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:c007fe8a43d43b3969e8469004e9845944f1a80e6acd47c150856bb87f230c56"},
-    {file = "pydantic_core-2.41.5-graalpy312-graalpy250_312_native-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:76d0819de158cd855d1cbb8fcafdf6f5cf1eb8e470abe056d5d161106e38062b"},
-    {file = "pydantic_core-2.41.5-pp310-pypy310_pp73-macosx_10_12_x86_64.whl", hash = "sha256:b5819cd790dbf0c5eb9f82c73c16b39a65dd6dd4d1439dcdea7816ec9adddab8"},
-    {file = "pydantic_core-2.41.5-pp310-pypy310_pp73-macosx_11_0_arm64.whl", hash = "sha256:5a4e67afbc95fa5c34cf27d9089bca7fcab4e51e57278d710320a70b956d1b9a"},
-    {file = "pydantic_core-2.41.5-pp310-pypy310_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ece5c59f0ce7d001e017643d8d24da587ea1f74f6993467d85ae8a5ef9d4f42b"},
-    {file = "pydantic_core-2.41.5-pp310-pypy310_pp73-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:16f80f7abe3351f8ea6858914ddc8c77e02578544a0ebc15b4c2e1a0e813b0b2"},
-    {file = "pydantic_core-2.41.5-pp310-pypy310_pp73-musllinux_1_1_aarch64.whl", hash = "sha256:33cb885e759a705b426baada1fe68cbb0a2e68e34c5d0d0289a364cf01709093"},
-    {file = "pydantic_core-2.41.5-pp310-pypy310_pp73-musllinux_1_1_armv7l.whl", hash = "sha256:c8d8b4eb992936023be7dee581270af5c6e0697a8559895f527f5b7105ecd36a"},
-    {file = "pydantic_core-2.41.5-pp310-pypy310_pp73-musllinux_1_1_x86_64.whl", hash = "sha256:242a206cd0318f95cd21bdacff3fcc3aab23e79bba5cac3db5a841c9ef9c6963"},
-    {file = "pydantic_core-2.41.5-pp310-pypy310_pp73-win_amd64.whl", hash = "sha256:d3a978c4f57a597908b7e697229d996d77a6d3c94901e9edee593adada95ce1a"},
-    {file = "pydantic_core-2.41.5-pp311-pypy311_pp73-macosx_10_12_x86_64.whl", hash = "sha256:b2379fa7ed44ddecb5bfe4e48577d752db9fc10be00a6b7446e9663ba143de26"},
-    {file = "pydantic_core-2.41.5-pp311-pypy311_pp73-macosx_11_0_arm64.whl", hash = "sha256:266fb4cbf5e3cbd0b53669a6d1b039c45e3ce651fd5442eff4d07c2cc8d66808"},
-    {file = "pydantic_core-2.41.5-pp311-pypy311_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:58133647260ea01e4d0500089a8c4f07bd7aa6ce109682b1426394988d8aaacc"},
-    {file = "pydantic_core-2.41.5-pp311-pypy311_pp73-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:287dad91cfb551c363dc62899a80e9e14da1f0e2b6ebde82c806612ca2a13ef1"},
-    {file = "pydantic_core-2.41.5-pp311-pypy311_pp73-musllinux_1_1_aarch64.whl", hash = "sha256:03b77d184b9eb40240ae9fd676ca364ce1085f203e1b1256f8ab9984dca80a84"},
-    {file = "pydantic_core-2.41.5-pp311-pypy311_pp73-musllinux_1_1_armv7l.whl", hash = "sha256:a668ce24de96165bb239160b3d854943128f4334822900534f2fe947930e5770"},
-    {file = "pydantic_core-2.41.5-pp311-pypy311_pp73-musllinux_1_1_x86_64.whl", hash = "sha256:f14f8f046c14563f8eb3f45f499cc658ab8d10072961e07225e507adb700e93f"},
-    {file = "pydantic_core-2.41.5-pp311-pypy311_pp73-win_amd64.whl", hash = "sha256:56121965f7a4dc965bff783d70b907ddf3d57f6eba29b6d2e5dabfaf07799c51"},
-    {file = "pydantic_core-2.41.5.tar.gz", hash = "sha256:08daa51ea16ad373ffd5e7606252cc32f07bc72b28284b6bc9c6df804816476e"},
-]
-
-[package.dependencies]
-typing-extensions = ">=4.14.1"
-
-[[package]]
-name = "pygments"
-version = "2.21.0"
-description = "Pygments is a syntax highlighting package written in Python."
-optional = false
-python-versions = ">=3.9"
-groups = ["dev"]
-files = [
-    {file = "pygments-2.21.0-py3-none-any.whl", hash = "sha256:2363c69b61c4a97c838da3b130dcd6468f4848992b21a82f2a63ec34377137d9"},
-    {file = "pygments-2.21.0.tar.gz", hash = "sha256:610ca751c9bc2492b38eb9a38a7fbc93edbbb2d7182edaf34e66ae493dee5c8c"},
-]
-
-[package.extras]
-windows-terminal = ["colorama (>=0.4.6)"]
-
-[[package]]
-name = "pyright"
-version = "1.1.408"
-description = "Command line wrapper for pyright"
-optional = false
-python-versions = ">=3.7"
-groups = ["dev"]
-files = [
-    {file = "pyright-1.1.408-py3-none-any.whl", hash = "sha256:090b32865f4fdb1e0e6cd82bf5618480d48eecd2eb2e70f960982a3d9a4c17c1"},
-    {file = "pyright-1.1.408.tar.gz", hash = "sha256:f28f2321f96852fa50b5829ea492f6adb0e6954568d1caa3f3af3a5f555eb684"},
-]
-
-[package.dependencies]
-nodeenv = ">=1.6.0"
-typing-extensions = ">=4.1"
-
-[package.extras]
-all = ["nodejs-wheel-binaries", "twine (>=3.4.1)"]
-dev = ["twine (>=3.4.1)"]
-nodejs = ["nodejs-wheel-binaries"]
-
-[[package]]
-name = "pytest"
-version = "9.0.3"
-description = "pytest: simple powerful testing with Python"
-optional = false
-python-versions = ">=3.10"
-groups = ["dev"]
-files = [
-    {file = "pytest-9.0.3-py3-none-any.whl", hash = "sha256:2c5efc453d45394fdd706ade797c0a81091eccd1d6e4bccfcd476e2b8e0ab5d9"},
-    {file = "pytest-9.0.3.tar.gz", hash = "sha256:b86ada508af81d19edeb213c681b1d48246c1a91d304c6c81a427674c17eb91c"},
-]
-
-[package.dependencies]
-colorama = {version = ">=0.4", markers = "sys_platform == \"win32\""}
-iniconfig = ">=1.0.1"
-packaging = ">=22"
-pluggy = ">=1.5,<2"
-pygments = ">=2.7.2"
-
-[package.extras]
-dev = ["argcomplete", "attrs (>=19.2)", "hypothesis (>=3.56)", "mock", "requests", "setuptools", "xmlschema"]
-
-[[package]]
-name = "pytest-asyncio"
-version = "1.3.0"
-description = "Pytest support for asyncio"
-optional = false
-python-versions = ">=3.10"
-groups = ["dev"]
-files = [
-    {file = "pytest_asyncio-1.3.0-py3-none-any.whl", hash = "sha256:611e26147c7f77640e6d0a92a38ed17c3e9848063698d5c93d5aa7aa11cebff5"},
-    {file = "pytest_asyncio-1.3.0.tar.gz", hash = "sha256:d7f52f36d231b80ee124cd216ffb19369aa168fc10095013c6b014a34d3ee9e5"},
-]
-
-[package.dependencies]
-pytest = ">=8.2,<10"
-typing-extensions = {version = ">=4.12", markers = "python_version < \"3.13\""}
-
-[package.extras]
-docs = ["sphinx (>=5.3)", "sphinx-rtd-theme (>=1)"]
-testing = ["coverage (>=6.2)", "hypothesis (>=5.7.1)"]
-
-[[package]]
-name = "pytest-cov"
-version = "7.1.0"
-description = "Pytest plugin for measuring coverage."
-optional = false
-python-versions = ">=3.9"
-groups = ["dev"]
-files = [
-    {file = "pytest_cov-7.1.0-py3-none-any.whl", hash = "sha256:a0461110b7865f9a271aa1b51e516c9a95de9d696734a2f71e3e78f46e1d4678"},
-    {file = "pytest_cov-7.1.0.tar.gz", hash = "sha256:30674f2b5f6351aa09702a9c8c364f6a01c27aae0c1366ae8016160d1efc56b2"},
-]
-
-[package.dependencies]
-coverage = {version = ">=7.10.6", extras = ["toml"]}
-pluggy = ">=1.2"
-pytest = ">=7"
-
-[package.extras]
-testing = ["process-tests", "pytest-xdist", "virtualenv"]
-
-[[package]]
-name = "ruff"
-version = "0.15.12"
-description = "An extremely fast Python linter and code formatter, written in Rust."
-optional = false
-python-versions = ">=3.7"
-groups = ["dev"]
-files = [
-    {file = "ruff-0.15.12-py3-none-linux_armv6l.whl", hash = "sha256:f86f176e188e94d6bdbc09f09bfd9dc729059ad93d0e7390b5a73efe19f8861c"},
-    {file = "ruff-0.15.12-py3-none-macosx_10_12_x86_64.whl", hash = "sha256:e3bcd123364c3770b8e1b7baaf343cc99a35f197c5c6e8af79015c666c423a6c"},
-    {file = "ruff-0.15.12-py3-none-macosx_11_0_arm64.whl", hash = "sha256:fe87510d000220aa1ed530d4448a7c696a0cae1213e5ec30e5874287b66557b5"},
-    {file = "ruff-0.15.12-py3-none-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:84a1630093121375a3e2a95b4a6dc7b59e2b4ee76216e32d81aae550a832d002"},
-    {file = "ruff-0.15.12-py3-none-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:fb129f40f114f089ebe0ca56c0d251cf2061b17651d464bb6478dc01e69f11f5"},
-    {file = "ruff-0.15.12-py3-none-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:b0c862b172d695db7598426b8af465e7e9ac00a3ea2a3630ee67eb82e366aaa6"},
-    {file = "ruff-0.15.12-py3-none-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:2849ea9f3484c3aca43a82f484210370319e7170df4dfe4843395ddf6c57bc33"},
-    {file = "ruff-0.15.12-py3-none-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:9e77c7e51c07fe396826d5969a5b846d9cd4c402535835fb6e21ce8b28fef847"},
-    {file = "ruff-0.15.12-py3-none-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:83b2f4f2f3b1026b5fb449b467d9264bf22067b600f7b6f41fc5958909f449d0"},
-    {file = "ruff-0.15.12-py3-none-manylinux_2_31_riscv64.whl", hash = "sha256:9ba3b8f1afd7e2e43d8943e55f249e13f9682fde09711644a6e7290eb4f3e339"},
-    {file = "ruff-0.15.12-py3-none-musllinux_1_2_aarch64.whl", hash = "sha256:e852ba9fdc890655e1d78f2df1499efbe0e54126bd405362154a75e2bde159c5"},
-    {file = "ruff-0.15.12-py3-none-musllinux_1_2_armv7l.whl", hash = "sha256:dd8aed930da53780d22fc70bdf84452c843cf64f8cb4eb38984319c24c5cd5fd"},
-    {file = "ruff-0.15.12-py3-none-musllinux_1_2_i686.whl", hash = "sha256:01da3988d225628b709493d7dc67c3b9b12c0210016b08690ef9bd27970b262b"},
-    {file = "ruff-0.15.12-py3-none-musllinux_1_2_x86_64.whl", hash = "sha256:9cae0f92bd5700d1213188b31cd3bdd2b315361296d10b96b8e2337d3d11f53e"},
-    {file = "ruff-0.15.12-py3-none-win32.whl", hash = "sha256:d0185894e038d7043ba8fd6aee7499ece6462dc0ea9f1e260c7451807c714c20"},
-    {file = "ruff-0.15.12-py3-none-win_amd64.whl", hash = "sha256:c87a162d61ab3adca47c03f7f717c68672edec7d1b5499e652331780fe74950d"},
-    {file = "ruff-0.15.12-py3-none-win_arm64.whl", hash = "sha256:a538f7a82d061cee7be55542aca1d86d1393d55d81d4fcc314370f4340930d4f"},
-    {file = "ruff-0.15.12.tar.gz", hash = "sha256:ecea26adb26b4232c0c2ca19ccbc0083a68344180bba2a600605538ce51a40a6"},
-]
-
-[[package]]
-name = "typing-extensions"
-version = "4.16.0"
-description = "Backported and Experimental Type Hints for Python 3.9+"
-optional = false
-python-versions = ">=3.9"
-groups = ["main", "dev"]
-files = [
-    {file = "typing_extensions-4.16.0-py3-none-any.whl", hash = "sha256:481caa481374e813c1b176ada14e97f1f67a4539ce9cfeb3f350d78d6370c2e8"},
-    {file = "typing_extensions-4.16.0.tar.gz", hash = "sha256:dc983d19a509c94dba722ee6abd33940f7c05a89e243c47e907eb4db6f1a43e5"},
-]
-
-[[package]]
-name = "typing-inspection"
-version = "0.4.4"
-description = "Runtime typing introspection tools"
-optional = false
-python-versions = ">=3.10"
-groups = ["main"]
-files = [
-    {file = "typing_inspection-0.4.4-py3-none-any.whl", hash = "sha256:65b8397ba37ccbce054456aaccddfc91e6e3083c92824df348d96ca832f3f147"},
-    {file = "typing_inspection-0.4.4.tar.gz", hash = "sha256:547274fa6b0a561ccf549cc9524b999a578e737d015d8709d021f9d0d13bea47"},
-]
-
-[package.dependencies]
-typing-extensions = ">=4.15.0"
-
-[[package]]
-name = "websockets"
-version = "15.0.1"
-description = "An implementation of the WebSocket Protocol (RFC 6455 & 7692)"
-optional = false
-python-versions = ">=3.9"
-groups = ["main"]
-files = [
-    {file = "websockets-15.0.1-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:d63efaa0cd96cf0c5fe4d581521d9fa87744540d4bc999ae6e08595a1014b45b"},
-    {file = "websockets-15.0.1-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:ac60e3b188ec7574cb761b08d50fcedf9d77f1530352db4eef1707fe9dee7205"},
-    {file = "websockets-15.0.1-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:5756779642579d902eed757b21b0164cd6fe338506a8083eb58af5c372e39d9a"},
-    {file = "websockets-15.0.1-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:0fdfe3e2a29e4db3659dbd5bbf04560cea53dd9610273917799f1cde46aa725e"},
-    {file = "websockets-15.0.1-cp310-cp310-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:4c2529b320eb9e35af0fa3016c187dffb84a3ecc572bcee7c3ce302bfeba52bf"},
-    {file = "websockets-15.0.1-cp310-cp310-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ac1e5c9054fe23226fb11e05a6e630837f074174c4c2f0fe442996112a6de4fb"},
-    {file = "websockets-15.0.1-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:5df592cd503496351d6dc14f7cdad49f268d8e618f80dce0cd5a36b93c3fc08d"},
-    {file = "websockets-15.0.1-cp310-cp310-musllinux_1_2_i686.whl", hash = "sha256:0a34631031a8f05657e8e90903e656959234f3a04552259458aac0b0f9ae6fd9"},
-    {file = "websockets-15.0.1-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:3d00075aa65772e7ce9e990cab3ff1de702aa09be3940d1dc88d5abf1ab8a09c"},
-    {file = "websockets-15.0.1-cp310-cp310-win32.whl", hash = "sha256:1234d4ef35db82f5446dca8e35a7da7964d02c127b095e172e54397fb6a6c256"},
-    {file = "websockets-15.0.1-cp310-cp310-win_amd64.whl", hash = "sha256:39c1fec2c11dc8d89bba6b2bf1556af381611a173ac2b511cf7231622058af41"},
-    {file = "websockets-15.0.1-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:823c248b690b2fd9303ba00c4f66cd5e2d8c3ba4aa968b2779be9532a4dad431"},
-    {file = "websockets-15.0.1-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:678999709e68425ae2593acf2e3ebcbcf2e69885a5ee78f9eb80e6e371f1bf57"},
-    {file = "websockets-15.0.1-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:d50fd1ee42388dcfb2b3676132c78116490976f1300da28eb629272d5d93e905"},
-    {file = "websockets-15.0.1-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:d99e5546bf73dbad5bf3547174cd6cb8ba7273062a23808ffea025ecb1cf8562"},
-    {file = "websockets-15.0.1-cp311-cp311-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:66dd88c918e3287efc22409d426c8f729688d89a0c587c88971a0faa2c2f3792"},
-    {file = "websockets-15.0.1-cp311-cp311-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:8dd8327c795b3e3f219760fa603dcae1dcc148172290a8ab15158cf85a953413"},
-    {file = "websockets-15.0.1-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:8fdc51055e6ff4adeb88d58a11042ec9a5eae317a0a53d12c062c8a8865909e8"},
-    {file = "websockets-15.0.1-cp311-cp311-musllinux_1_2_i686.whl", hash = "sha256:693f0192126df6c2327cce3baa7c06f2a117575e32ab2308f7f8216c29d9e2e3"},
-    {file = "websockets-15.0.1-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:54479983bd5fb469c38f2f5c7e3a24f9a4e70594cd68cd1fa6b9340dadaff7cf"},
-    {file = "websockets-15.0.1-cp311-cp311-win32.whl", hash = "sha256:16b6c1b3e57799b9d38427dda63edcbe4926352c47cf88588c0be4ace18dac85"},
-    {file = "websockets-15.0.1-cp311-cp311-win_amd64.whl", hash = "sha256:27ccee0071a0e75d22cb35849b1db43f2ecd3e161041ac1ee9d2352ddf72f065"},
-    {file = "websockets-15.0.1-cp312-cp312-macosx_10_13_universal2.whl", hash = "sha256:3e90baa811a5d73f3ca0bcbf32064d663ed81318ab225ee4f427ad4e26e5aff3"},
-    {file = "websockets-15.0.1-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:592f1a9fe869c778694f0aa806ba0374e97648ab57936f092fd9d87f8bc03665"},
-    {file = "websockets-15.0.1-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:0701bc3cfcb9164d04a14b149fd74be7347a530ad3bbf15ab2c678a2cd3dd9a2"},
-    {file = "websockets-15.0.1-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:e8b56bdcdb4505c8078cb6c7157d9811a85790f2f2b3632c7d1462ab5783d215"},
-    {file = "websockets-15.0.1-cp312-cp312-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:0af68c55afbd5f07986df82831c7bff04846928ea8d1fd7f30052638788bc9b5"},
-    {file = "websockets-15.0.1-cp312-cp312-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:64dee438fed052b52e4f98f76c5790513235efaa1ef7f3f2192c392cd7c91b65"},
-    {file = "websockets-15.0.1-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:d5f6b181bb38171a8ad1d6aa58a67a6aa9d4b38d0f8c5f496b9e42561dfc62fe"},
-    {file = "websockets-15.0.1-cp312-cp312-musllinux_1_2_i686.whl", hash = "sha256:5d54b09eba2bada6011aea5375542a157637b91029687eb4fdb2dab11059c1b4"},
-    {file = "websockets-15.0.1-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:3be571a8b5afed347da347bfcf27ba12b069d9d7f42cb8c7028b5e98bbb12597"},
-    {file = "websockets-15.0.1-cp312-cp312-win32.whl", hash = "sha256:c338ffa0520bdb12fbc527265235639fb76e7bc7faafbb93f6ba80d9c06578a9"},
-    {file = "websockets-15.0.1-cp312-cp312-win_amd64.whl", hash = "sha256:fcd5cf9e305d7b8338754470cf69cf81f420459dbae8a3b40cee57417f4614a7"},
-    {file = "websockets-15.0.1-cp313-cp313-macosx_10_13_universal2.whl", hash = "sha256:ee443ef070bb3b6ed74514f5efaa37a252af57c90eb33b956d35c8e9c10a1931"},
-    {file = "websockets-15.0.1-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:5a939de6b7b4e18ca683218320fc67ea886038265fd1ed30173f5ce3f8e85675"},
-    {file = "websockets-15.0.1-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:746ee8dba912cd6fc889a8147168991d50ed70447bf18bcda7039f7d2e3d9151"},
-    {file = "websockets-15.0.1-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:595b6c3969023ecf9041b2936ac3827e4623bfa3ccf007575f04c5a6aa318c22"},
-    {file = "websockets-15.0.1-cp313-cp313-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:3c714d2fc58b5ca3e285461a4cc0c9a66bd0e24c5da9911e30158286c9b5be7f"},
-    {file = "websockets-15.0.1-cp313-cp313-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:0f3c1e2ab208db911594ae5b4f79addeb3501604a165019dd221c0bdcabe4db8"},
-    {file = "websockets-15.0.1-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:229cf1d3ca6c1804400b0a9790dc66528e08a6a1feec0d5040e8b9eb14422375"},
-    {file = "websockets-15.0.1-cp313-cp313-musllinux_1_2_i686.whl", hash = "sha256:756c56e867a90fb00177d530dca4b097dd753cde348448a1012ed6c5131f8b7d"},
-    {file = "websockets-15.0.1-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:558d023b3df0bffe50a04e710bc87742de35060580a293c2a984299ed83bc4e4"},
-    {file = "websockets-15.0.1-cp313-cp313-win32.whl", hash = "sha256:ba9e56e8ceeeedb2e080147ba85ffcd5cd0711b89576b83784d8605a7df455fa"},
-    {file = "websockets-15.0.1-cp313-cp313-win_amd64.whl", hash = "sha256:e09473f095a819042ecb2ab9465aee615bd9c2028e4ef7d933600a8401c79561"},
-    {file = "websockets-15.0.1-cp39-cp39-macosx_10_9_universal2.whl", hash = "sha256:5f4c04ead5aed67c8a1a20491d54cdfba5884507a48dd798ecaf13c74c4489f5"},
-    {file = "websockets-15.0.1-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:abdc0c6c8c648b4805c5eacd131910d2a7f6455dfd3becab248ef108e89ab16a"},
-    {file = "websockets-15.0.1-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:a625e06551975f4b7ea7102bc43895b90742746797e2e14b70ed61c43a90f09b"},
-    {file = "websockets-15.0.1-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:d591f8de75824cbb7acad4e05d2d710484f15f29d4a915092675ad3456f11770"},
-    {file = "websockets-15.0.1-cp39-cp39-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:47819cea040f31d670cc8d324bb6435c6f133b8c7a19ec3d61634e62f8d8f9eb"},
-    {file = "websockets-15.0.1-cp39-cp39-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ac017dd64572e5c3bd01939121e4d16cf30e5d7e110a119399cf3133b63ad054"},
-    {file = "websockets-15.0.1-cp39-cp39-musllinux_1_2_aarch64.whl", hash = "sha256:4a9fac8e469d04ce6c25bb2610dc535235bd4aa14996b4e6dbebf5e007eba5ee"},
-    {file = "websockets-15.0.1-cp39-cp39-musllinux_1_2_i686.whl", hash = "sha256:363c6f671b761efcb30608d24925a382497c12c506b51661883c3e22337265ed"},
-    {file = "websockets-15.0.1-cp39-cp39-musllinux_1_2_x86_64.whl", hash = "sha256:2034693ad3097d5355bfdacfffcbd3ef5694f9718ab7f29c29689a9eae841880"},
-    {file = "websockets-15.0.1-cp39-cp39-win32.whl", hash = "sha256:3b1ac0d3e594bf121308112697cf4b32be538fb1444468fb0a6ae4feebc83411"},
-    {file = "websockets-15.0.1-cp39-cp39-win_amd64.whl", hash = "sha256:b7643a03db5c95c799b89b31c036d5f27eeb4d259c798e878d6937d71832b1e4"},
-    {file = "websockets-15.0.1-pp310-pypy310_pp73-macosx_10_15_x86_64.whl", hash = "sha256:0c9e74d766f2818bb95f84c25be4dea09841ac0f734d1966f415e4edfc4ef1c3"},
-    {file = "websockets-15.0.1-pp310-pypy310_pp73-macosx_11_0_arm64.whl", hash = "sha256:1009ee0c7739c08a0cd59de430d6de452a55e42d6b522de7aa15e6f67db0b8e1"},
-    {file = "websockets-15.0.1-pp310-pypy310_pp73-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:76d1f20b1c7a2fa82367e04982e708723ba0e7b8d43aa643d3dcd404d74f1475"},
-    {file = "websockets-15.0.1-pp310-pypy310_pp73-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:f29d80eb9a9263b8d109135351caf568cc3f80b9928bccde535c235de55c22d9"},
-    {file = "websockets-15.0.1-pp310-pypy310_pp73-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:b359ed09954d7c18bbc1680f380c7301f92c60bf924171629c5db97febb12f04"},
-    {file = "websockets-15.0.1-pp310-pypy310_pp73-win_amd64.whl", hash = "sha256:cad21560da69f4ce7658ca2cb83138fb4cf695a2ba3e475e0559e05991aa8122"},
-    {file = "websockets-15.0.1-pp39-pypy39_pp73-macosx_10_15_x86_64.whl", hash = "sha256:7f493881579c90fc262d9cdbaa05a6b54b3811c2f300766748db79f098db9940"},
-    {file = "websockets-15.0.1-pp39-pypy39_pp73-macosx_11_0_arm64.whl", hash = "sha256:47b099e1f4fbc95b701b6e85768e1fcdaf1630f3cbe4765fa216596f12310e2e"},
-    {file = "websockets-15.0.1-pp39-pypy39_pp73-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:67f2b6de947f8c757db2db9c71527933ad0019737ec374a8a6be9a956786aaf9"},
-    {file = "websockets-15.0.1-pp39-pypy39_pp73-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:d08eb4c2b7d6c41da6ca0600c077e93f5adcfd979cd777d747e9ee624556da4b"},
-    {file = "websockets-15.0.1-pp39-pypy39_pp73-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:4b826973a4a2ae47ba357e4e82fa44a463b8f168e1ca775ac64521442b19e87f"},
-    {file = "websockets-15.0.1-pp39-pypy39_pp73-win_amd64.whl", hash = "sha256:21c1fa28a6a7e3cbdc171c694398b6df4744613ce9b36b1a498e816787e28123"},
-    {file = "websockets-15.0.1-py3-none-any.whl", hash = "sha256:f7a866fbc1e97b5c617ee4116daaa09b722101d4a3c170c787450ba409f9736f"},
-    {file = "websockets-15.0.1.tar.gz", hash = "sha256:82544de02076bafba038ce055ee6412d68da13ab47f0c60cab827346de828dee"},
-]
-
-[metadata]
-lock-version = "2.1"
-python-versions = "^3.11"
-content-hash = "3773505ce5b1dedfd5eef2ffdd8e33dc695a54e4f6d57e5378076502ce8c8041"
diff --git a/paddler_test_cluster_python/pyproject.toml b/paddler_test_cluster_python/pyproject.toml
deleted file mode 100644
index c647c9e52..000000000
--- a/paddler_test_cluster_python/pyproject.toml
+++ /dev/null
@@ -1,63 +0,0 @@
-[tool.poetry]
-name = "paddler-test-cluster"
-version = "4.0.1"
-description = "Spawns real Paddler balancers and agents for the Python test suites"
-authors = ["Intentee"]
-license = "Apache-2.0"
-packages = [{include = "paddler_test_cluster"}]
-
-[tool.poetry.dependencies]
-python = "^3.11"
-paddler-client = { path = "../paddler_client_python", develop = true }
-
-[tool.poetry.group.dev.dependencies]
-pytest = "9.0.3"
-pytest-asyncio = "1.3.0"
-pytest-cov = "7.1.0"
-pyright = "1.1.408"
-ruff = "0.15.12"
-mypy = "1.19.1"
-
-[build-system]
-requires = ["poetry-core"]
-build-backend = "poetry.core.masonry.api"
-
-[tool.mypy]
-python_version = "3.11"
-strict = true
-extra_checks = true
-plugins = ["pydantic.mypy"]
-
-[tool.pyright]
-pythonVersion = "3.11"
-pythonPlatform = "All"
-typeCheckingMode = "strict"
-
-[tool.ruff]
-line-length = 88
-target-version = "py311"
-
-[tool.ruff.lint]
-select = ["ALL"]
-ignore = [
-    "COM812",  # trailing comma - conflicts with formatter
-    "D",       # docstrings - project prefers self-documenting code
-    "ISC001",  # implicit string concatenation - conflicts with formatter
-]
-
-[tool.ruff.lint.per-file-ignores]
-"tests/**/*.py" = [
-    "S101",    # assert statements
-]
-
-[tool.coverage.report]
-fail_under = 100
-show_missing = true
-
-[tool.coverage.run]
-branch = true
-source = ["paddler_test_cluster", "tests"]
-
-[tool.pytest.ini_options]
-asyncio_mode = "auto"
-addopts = "--cov-report=term-missing"
diff --git a/paddler_test_cluster_python/tests/__init__.py b/paddler_test_cluster_python/tests/__init__.py
deleted file mode 100644
index e69de29bb..000000000
diff --git a/paddler_test_cluster_python/tests/cluster/__init__.py b/paddler_test_cluster_python/tests/cluster/__init__.py
deleted file mode 100644
index e69de29bb..000000000
diff --git a/paddler_test_cluster_python/tests/cluster/test_agent_reporting_issues_fails_the_cluster.py b/paddler_test_cluster_python/tests/cluster/test_agent_reporting_issues_fails_the_cluster.py
deleted file mode 100644
index 0b0ab4953..000000000
--- a/paddler_test_cluster_python/tests/cluster/test_agent_reporting_issues_fails_the_cluster.py
+++ /dev/null
@@ -1,29 +0,0 @@
-from contextlib import AsyncExitStack
-
-import pytest
-from paddler_client.agent_desired_model import AgentDesiredModel
-from paddler_client.balancer_desired_state import BalancerDesiredState
-
-from paddler_test_cluster.agent_spec import AgentSpec
-from paddler_test_cluster.error import AgentReportedIssuesError
-from paddler_test_cluster.paddler_cluster import paddler_cluster
-
-
-async def test_agent_reporting_issues_fails_the_cluster() -> None:
-    async with AsyncExitStack() as exit_stack:
-        with pytest.raises(AgentReportedIssuesError) as reported_issues:
-            await exit_stack.enter_async_context(
-                paddler_cluster(
-                    BalancerDesiredState(
-                        model=AgentDesiredModel.local_to_agent(
-                            "/nonexistent/model.gguf"
-                        ),
-                    ),
-                    [AgentSpec(name="agent-without-its-model", slots=1)],
-                )
-            )
-
-    assert reported_issues.value.agent_name == "agent-without-its-model"
-    assert [issue.variant for issue in reported_issues.value.issues] == [
-        "ModelFileDoesNotExist"
-    ]
diff --git a/paddler_test_cluster_python/tests/cluster/test_agents_stream_closing_before_readiness_is_reported.py b/paddler_test_cluster_python/tests/cluster/test_agents_stream_closing_before_readiness_is_reported.py
deleted file mode 100644
index dd7a59745..000000000
--- a/paddler_test_cluster_python/tests/cluster/test_agents_stream_closing_before_readiness_is_reported.py
+++ /dev/null
@@ -1,21 +0,0 @@
-import pytest
-from paddler_client.client_management import ClientManagement
-
-from paddler_test_cluster.error import AgentsStreamClosedError
-from paddler_test_cluster.spawned_balancer import SpawnedBalancer
-from paddler_test_cluster.wait_for_agent_ready import wait_for_agent_ready
-
-
-async def test_agents_stream_closing_before_readiness_is_reported() -> None:
-    balancer = await SpawnedBalancer.spawn()
-
-    async with ClientManagement(url=balancer.addresses.management_url) as client:
-        snapshots = client.agents_stream()
-
-        await anext(snapshots)
-        await balancer.process.terminate()
-
-        with pytest.raises(AgentsStreamClosedError) as stream_closed:
-            await wait_for_agent_ready(snapshots, "agent-that-never-joined", 1)
-
-    assert stream_closed.value.agent_name == "agent-that-never-joined"
diff --git a/paddler_test_cluster_python/tests/cluster/test_balancer_serves_on_the_addresses_it_announces.py b/paddler_test_cluster_python/tests/cluster/test_balancer_serves_on_the_addresses_it_announces.py
deleted file mode 100644
index 091f49739..000000000
--- a/paddler_test_cluster_python/tests/cluster/test_balancer_serves_on_the_addresses_it_announces.py
+++ /dev/null
@@ -1,25 +0,0 @@
-import httpx
-from paddler_client.client_inference import ClientInference
-from paddler_client.client_management import ClientManagement
-
-from paddler_test_cluster.spawned_balancer import SpawnedBalancer
-
-
-async def test_balancer_serves_on_the_addresses_it_announces() -> None:
-    balancer = await SpawnedBalancer.spawn()
-
-    try:
-        async with ClientManagement(url=balancer.addresses.management_url) as client:
-            assert await client.get_health() == "OK"
-
-        async with ClientInference(url=balancer.addresses.inference_url) as client:
-            assert await client.get_health() == "OK"
-
-        async with httpx.AsyncClient() as http_client:
-            compat_openai_health = await http_client.get(
-                f"{balancer.addresses.compat_openai_url}/health"
-            )
-
-        assert compat_openai_health.text == "OK"
-    finally:
-        await balancer.process.terminate()
diff --git a/paddler_test_cluster_python/tests/cluster/test_balancer_that_exits_before_announcing_is_reported.py b/paddler_test_cluster_python/tests/cluster/test_balancer_that_exits_before_announcing_is_reported.py
deleted file mode 100644
index 82f75262f..000000000
--- a/paddler_test_cluster_python/tests/cluster/test_balancer_that_exits_before_announcing_is_reported.py
+++ /dev/null
@@ -1,13 +0,0 @@
-import pytest
-
-from paddler_test_cluster.error import BalancerExitedBeforeAnnouncingError
-from paddler_test_cluster.spawned_balancer import SpawnedBalancer
-
-CLAP_USAGE_ERROR_STATUS = 2
-
-
-async def test_balancer_that_exits_before_announcing_is_reported() -> None:
-    with pytest.raises(BalancerExitedBeforeAnnouncingError) as exited_early:
-        await SpawnedBalancer.spawn(state_database_url="file://relative/state.json")
-
-    assert exited_early.value.returncode == CLAP_USAGE_ERROR_STATUS
diff --git a/paddler_test_cluster_python/tests/cluster/test_cluster_without_a_model_registers_its_agents.py b/paddler_test_cluster_python/tests/cluster/test_cluster_without_a_model_registers_its_agents.py
deleted file mode 100644
index e90937a69..000000000
--- a/paddler_test_cluster_python/tests/cluster/test_cluster_without_a_model_registers_its_agents.py
+++ /dev/null
@@ -1,18 +0,0 @@
-from paddler_client.balancer_desired_state import BalancerDesiredState
-from paddler_client.client_management import ClientManagement
-
-from paddler_test_cluster.agent_spec import AgentSpec
-from paddler_test_cluster.paddler_cluster import paddler_cluster
-
-
-async def test_cluster_without_a_model_registers_its_agents() -> None:
-    async with (
-        paddler_cluster(
-            BalancerDesiredState(),
-            [AgentSpec(name="first", slots=1), AgentSpec(name="second", slots=2)],
-        ) as addresses,
-        ClientManagement(url=addresses.management_url) as client,
-    ):
-        snapshot = await client.get_agents()
-
-    assert {agent.name for agent in snapshot.agents} == {"first", "second"}
diff --git a/paddler_test_cluster_python/tests/cluster/test_process_that_exited_with_a_failure_is_reported.py b/paddler_test_cluster_python/tests/cluster/test_process_that_exited_with_a_failure_is_reported.py
deleted file mode 100644
index a03004aa8..000000000
--- a/paddler_test_cluster_python/tests/cluster/test_process_that_exited_with_a_failure_is_reported.py
+++ /dev/null
@@ -1,24 +0,0 @@
-import asyncio
-
-import pytest
-
-from paddler_test_cluster.error import ProcessExitedWithFailureError
-from paddler_test_cluster.paddler_binary_path import paddler_binary_path
-from paddler_test_cluster.spawned_process import SpawnedProcess
-
-CLAP_USAGE_ERROR_STATUS = 2
-
-
-async def test_process_that_exited_with_a_failure_is_reported() -> None:
-    process = await asyncio.create_subprocess_exec(
-        paddler_binary_path(),
-        "--flag-paddler-does-not-have",
-        stdout=asyncio.subprocess.DEVNULL,
-        stderr=asyncio.subprocess.DEVNULL,
-    )
-    await process.wait()
-
-    with pytest.raises(ProcessExitedWithFailureError) as exited_with_failure:
-        await SpawnedProcess(process).terminate()
-
-    assert exited_with_failure.value.returncode == CLAP_USAGE_ERROR_STATUS
diff --git a/paddler_test_cluster_python/tests/cluster/test_process_that_ignores_termination_is_reported.py b/paddler_test_cluster_python/tests/cluster/test_process_that_ignores_termination_is_reported.py
deleted file mode 100644
index 25b923b0f..000000000
--- a/paddler_test_cluster_python/tests/cluster/test_process_that_ignores_termination_is_reported.py
+++ /dev/null
@@ -1,18 +0,0 @@
-import os
-import signal
-
-import pytest
-
-from paddler_test_cluster.error import ProcessDidNotExitError
-from paddler_test_cluster.spawned_balancer import SpawnedBalancer
-
-
-async def test_process_that_ignores_termination_is_reported() -> None:
-    balancer = await SpawnedBalancer.spawn()
-
-    os.kill(balancer.process.pid, signal.SIGSTOP)
-
-    with pytest.raises(ProcessDidNotExitError) as did_not_exit:
-        await balancer.process.terminate()
-
-    assert did_not_exit.value.pid == balancer.process.pid
diff --git a/paddler_test_cluster_python/tests/llm/__init__.py b/paddler_test_cluster_python/tests/llm/__init__.py
deleted file mode 100644
index e69de29bb..000000000
diff --git a/paddler_test_cluster_python/tests/llm/test_cluster_with_a_model_waits_until_every_slot_loads_it.py b/paddler_test_cluster_python/tests/llm/test_cluster_with_a_model_waits_until_every_slot_loads_it.py
deleted file mode 100644
index d60c5b232..000000000
--- a/paddler_test_cluster_python/tests/llm/test_cluster_with_a_model_waits_until_every_slot_loads_it.py
+++ /dev/null
@@ -1,29 +0,0 @@
-import pytest
-from paddler_client.balancer_desired_state import BalancerDesiredState
-from paddler_client.client_management import ClientManagement
-
-from paddler_test_cluster.agent_spec import AgentSpec
-from paddler_test_cluster.nomic_embed_text_v1_5_desired_state import (
-    NOMIC_EMBED_TEXT_V1_5_DESIRED_STATE,
-)
-from paddler_test_cluster.paddler_cluster import paddler_cluster
-from paddler_test_cluster.qwen3_0_6b_desired_state import QWEN3_0_6B_DESIRED_STATE
-
-AGENT = AgentSpec(name="loaded", slots=2)
-
-
-@pytest.mark.parametrize(
-    "desired_state",
-    [QWEN3_0_6B_DESIRED_STATE, NOMIC_EMBED_TEXT_V1_5_DESIRED_STATE],
-)
-async def test_cluster_with_a_model_waits_until_every_slot_loads_it(
-    desired_state: BalancerDesiredState,
-) -> None:
-    async with (
-        paddler_cluster(desired_state, [AGENT]) as addresses,
-        ClientManagement(url=addresses.management_url) as client,
-    ):
-        [agent] = (await client.get_agents()).agents
-
-    assert agent.status.model_path is not None
-    assert agent.status.slots_total == AGENT.slots
diff --git a/paddler_test_cluster_python/tests/unit/__init__.py b/paddler_test_cluster_python/tests/unit/__init__.py
deleted file mode 100644
index e69de29bb..000000000
diff --git a/paddler_test_cluster_python/tests/unit/test_compat_openai_url_requires_the_service.py b/paddler_test_cluster_python/tests/unit/test_compat_openai_url_requires_the_service.py
deleted file mode 100644
index 3e8587f82..000000000
--- a/paddler_test_cluster_python/tests/unit/test_compat_openai_url_requires_the_service.py
+++ /dev/null
@@ -1,16 +0,0 @@
-import pytest
-
-from paddler_test_cluster.balancer_addresses import BalancerAddresses
-from paddler_test_cluster.error import CompatOpenAIServiceNotServedError
-
-
-def test_compat_openai_url_requires_the_service() -> None:
-    addresses = BalancerAddresses(
-        compat_openai=None,
-        inference="127.0.0.1:8061",
-        management="127.0.0.1:8060",
-        web_admin_panel=None,
-    )
-
-    with pytest.raises(CompatOpenAIServiceNotServedError):
-        _ = addresses.compat_openai_url
diff --git a/paddler_tests/Cargo.toml b/paddler_tests/Cargo.toml
index d58c60938..2b96ab033 100644
--- a/paddler_tests/Cargo.toml
+++ b/paddler_tests/Cargo.toml
@@ -10,6 +10,10 @@ publish = false
 repository.workspace = true
 version.workspace = true
 
+[[bin]]
+name = "run_with_test_cluster"
+path = "src/bin/run_with_test_cluster.rs"
+
 [[test]]
 name = "integration"
 path = "tests/main.rs"
@@ -24,28 +28,47 @@ web_admin_panel = ["paddler_balancer/web_admin_panel", "paddler_bootstrap/web_ad
 [dependencies]
 anyhow = { workspace = true }
 async-trait = { workspace = true }
+clap = { workspace = true }
 futures-util = { workspace = true }
+llama-cpp-bindings = { workspace = true }
+llama-cpp-bindings-sys = { workspace = true }
 log = { workspace = true }
+paddler_agent = { workspace = true }
+paddler_agent_decision = { workspace = true }
+paddler_agent_embeddings = { workspace = true }
+paddler_agent_pointer_head = { workspace = true }
+paddler_agent_runtime = { workspace = true }
+paddler_agent_status = { workspace = true }
+paddler_agent_text_generation = { workspace = true }
 paddler_balancer = { workspace = true }
 paddler_bootstrap = { workspace = true }
+paddler_cache_dir = { workspace = true }
 paddler_client = { workspace = true }
+paddler_download_manager = { workspace = true }
 paddler_inference_parameters = { workspace = true }
 paddler_messaging = { workspace = true }
+paddler_model_source = { workspace = true }
+paddler_openai_response_format_validator = { workspace = true }
+paddler_openai_translation = { workspace = true }
 paddler_state_database = { workspace = true }
 paddler_test_cluster_harness = { workspace = true }
 reqwest = { workspace = true }
+serde = { workspace = true }
 serde_json = { workspace = true }
 tempfile = { workspace = true }
 tokio = { workspace = true }
 tokio-util = { workspace = true }
-trzcina = { workspace = true }
 
 [dev-dependencies]
 async-openai = { workspace = true }
 http = { workspace = true }
 llama-cpp-bindings-types = { workspace = true }
+paddler_blocking_thread_fixture = { workspace = true }
 paddler_local_http_fixture = { workspace = true }
-paddler_openai_response_format_validator = { workspace = true }
+paddler_tcp_buffer_limits = { workspace = true }
+paddler_tool_call_validator = { workspace = true }
+paddler_typesafe_translation = { workspace = true }
+syn = { workspace = true }
 tokio-tungstenite = { workspace = true }
 
 [lints]
diff --git a/paddler_tests/src/active_embedding_request_fixture.rs b/paddler_tests/src/active_embedding_request_fixture.rs
new file mode 100644
index 000000000..1e1fa76ff
--- /dev/null
+++ b/paddler_tests/src/active_embedding_request_fixture.rs
@@ -0,0 +1,46 @@
+use std::num::NonZeroU16;
+use std::sync::Arc;
+
+use tokio::sync::mpsc;
+
+use paddler_agent_embeddings::active_embedding_request::ActiveEmbeddingRequest;
+use paddler_agent_embeddings::embedding_input_tokenized::EmbeddingInputTokenized;
+use paddler_agent_embeddings::prepared_embedding_batch_request::PreparedEmbeddingBatchRequest;
+use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
+use paddler_agent_status::slot_guard::SlotGuard;
+use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
+use paddler_messaging::embedding_result::EmbeddingResult;
+
+pub struct ActiveEmbeddingRequestFixture {
+    pub admitted_requests: Vec<ActiveEmbeddingRequest>,
+    pub generate_embedding_stop_tx: mpsc::UnboundedSender<()>,
+    pub generated_embedding_rx: mpsc::UnboundedReceiver<EmbeddingResult>,
+}
+
+impl ActiveEmbeddingRequestFixture {
+    #[must_use]
+    pub fn new(
+        inputs: Vec<EmbeddingInputTokenized>,
+        slot_aggregated_status: Arc<SlotAggregatedStatus>,
+        reserved_slots: NonZeroU16,
+    ) -> Self {
+        let (generated_embedding_tx, generated_embedding_rx) = mpsc::unbounded_channel();
+        let (generate_embedding_stop_tx, generate_embedding_stop_rx) = mpsc::unbounded_channel();
+
+        Self {
+            admitted_requests: PreparedEmbeddingBatchRequest {
+                generate_embedding_stop_rx,
+                generated_embedding_tx,
+                inputs,
+                normalization_method: EmbeddingNormalizationMethod::None,
+                unembeddable_documents: Vec::new(),
+                slot_guard: SlotGuard::new(slot_aggregated_status, reserved_slots),
+            }
+            .admit(None)
+            .into_iter()
+            .collect(),
+            generate_embedding_stop_tx,
+            generated_embedding_rx,
+        }
+    }
+}
diff --git a/paddler_tests/src/admitted_decision.rs b/paddler_tests/src/admitted_decision.rs
new file mode 100644
index 000000000..145d252d1
--- /dev/null
+++ b/paddler_tests/src/admitted_decision.rs
@@ -0,0 +1,42 @@
+use std::sync::Arc;
+
+use tokio::sync::mpsc;
+
+use paddler_agent_decision::active_decision_request::ActiveDecisionRequest;
+use paddler_agent_decision::decision_capacity_ledger::DecisionCapacityLedger;
+use paddler_agent_decision::decision_token_layout::DecisionTokenLayout;
+use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
+use paddler_messaging::decision_result::DecisionResult;
+
+use crate::prepared_decision::PreparedDecision;
+
+pub struct AdmittedDecision {
+    pub active_decision_request: ActiveDecisionRequest,
+    pub decision_result_rx: mpsc::UnboundedReceiver<DecisionResult>,
+    pub decision_stop_tx: mpsc::UnboundedSender<()>,
+}
+
+impl AdmittedDecision {
+    #[must_use]
+    pub fn admit(
+        capacity_ledger: &DecisionCapacityLedger,
+        slot_aggregated_status: Arc<SlotAggregatedStatus>,
+        layout: DecisionTokenLayout,
+    ) -> Option<Self> {
+        let PreparedDecision {
+            decision_result_rx,
+            decision_stop_tx,
+            prepared_decision_request,
+        } = PreparedDecision::new(slot_aggregated_status, layout);
+        capacity_ledger
+            .admit(prepared_decision_request.layout.question_lane_demand())
+            .map(|admission| Self {
+                active_decision_request: ActiveDecisionRequest::new(
+                    prepared_decision_request,
+                    admission,
+                ),
+                decision_result_rx,
+                decision_stop_tx,
+            })
+    }
+}
diff --git a/paddler_tests/src/agent_inference_runtime_context.rs b/paddler_tests/src/agent_inference_runtime_context.rs
new file mode 100644
index 000000000..3e4919b62
--- /dev/null
+++ b/paddler_tests/src/agent_inference_runtime_context.rs
@@ -0,0 +1,10 @@
+use paddler_agent_runtime::inference_runtime_context::InferenceRuntimeContext;
+use paddler_agent_runtime::inference_runtime_context_params::InferenceRuntimeContextParams;
+
+#[must_use]
+pub fn agent_inference_runtime_context(slots: u16) -> InferenceRuntimeContext {
+    InferenceRuntimeContext::new(InferenceRuntimeContextParams {
+        agent_name: Some("agent".to_owned()),
+        slots,
+    })
+}
diff --git a/paddler_tests/src/batch_size_within_context.rs b/paddler_tests/src/batch_size_within_context.rs
deleted file mode 100644
index 3fcd8c439..000000000
--- a/paddler_tests/src/batch_size_within_context.rs
+++ /dev/null
@@ -1,11 +0,0 @@
-use std::num::NonZeroU32;
-
-use anyhow::Result;
-
-use paddler_inference_parameters::batch_size::BatchSize;
-
-pub fn batch_size_within_context(context_size: NonZeroU32) -> Result<BatchSize> {
-    Ok(BatchSize::try_from(
-        context_size.get().min(BatchSize::DEFAULT.tokens().get()),
-    )?)
-}
diff --git a/paddler_tests/src/bin/run_with_test_cluster.rs b/paddler_tests/src/bin/run_with_test_cluster.rs
new file mode 100644
index 000000000..897dd7788
--- /dev/null
+++ b/paddler_tests/src/bin/run_with_test_cluster.rs
@@ -0,0 +1,134 @@
+use std::net::SocketAddr;
+use std::process::ExitCode;
+use std::process::ExitStatus;
+
+use anyhow::Error;
+use anyhow::Result;
+use clap::Parser;
+use futures_util::TryFutureExt as _;
+use futures_util::future::ready;
+use tokio::process::Command;
+
+use paddler_balancer::balancer_addresses::BalancerAddresses;
+use paddler_test_cluster_harness::cluster::Cluster;
+use paddler_tests::start_cluster::start_cluster;
+use paddler_tests::test_cluster_preset::TestClusterPreset;
+
+fn http_url(addr: SocketAddr) -> String {
+    format!("http://{addr}")
+}
+
+fn command_against_cluster(
+    program: &str,
+    arguments: &[String],
+    BalancerAddresses {
+        compat_openai,
+        compat_typesafe,
+        inference,
+        management,
+        ..
+    }: &BalancerAddresses,
+) -> Command {
+    let mut command = Command::new(program);
+
+    command
+        .args(arguments)
+        .env("PADDLER_INFERENCE_URL", http_url(*inference))
+        .env("PADDLER_MANAGEMENT_URL", http_url(*management));
+
+    if let Some(compat_openai) = compat_openai {
+        command.env("PADDLER_COMPAT_OPENAI_URL", http_url(*compat_openai));
+    }
+
+    if let Some(compat_typesafe) = compat_typesafe {
+        command.env("PADDLER_COMPAT_TYPESAFE_URL", http_url(*compat_typesafe));
+    }
+
+    command
+}
+
+async fn run_against_cluster(
+    cluster: Cluster,
+    program: String,
+    arguments: Vec<String>,
+) -> Result<ExitStatus> {
+    let command_status = command_against_cluster(&program, &arguments, &cluster.balancer.addresses)
+        .status()
+        .await
+        .map_err(Error::from);
+
+    cluster.shutdown().await.and(command_status)
+}
+
+fn exit_code(command_status: ExitStatus) -> ExitCode {
+    if command_status.success() {
+        ExitCode::SUCCESS
+    } else {
+        ExitCode::FAILURE
+    }
+}
+
+#[derive(Parser)]
+struct RunWithTestClusterArguments {
+    preset: TestClusterPreset,
+    program: String,
+    #[arg(allow_hyphen_values = true, trailing_var_arg = true)]
+    arguments: Vec<String>,
+}
+
+#[tokio::main]
+async fn main() -> Result<ExitCode> {
+    let RunWithTestClusterArguments {
+        preset,
+        program,
+        arguments,
+    } = RunWithTestClusterArguments::parse();
+
+    ready(preset.cluster_params())
+        .err_into()
+        .and_then(start_cluster)
+        .and_then(|cluster| run_against_cluster(cluster, program, arguments))
+        .await
+        .map(exit_code)
+}
+
+#[cfg(test)]
+mod tests {
+    use std::collections::BTreeMap;
+    use std::ffi::OsStr;
+    use std::net::Ipv4Addr;
+    use std::net::SocketAddr;
+
+    use paddler_balancer::balancer_addresses::BalancerAddresses;
+
+    use super::command_against_cluster;
+
+    #[test]
+    fn exports_no_url_for_a_compatibility_service_the_cluster_does_not_serve() {
+        let command = command_against_cluster(
+            "printenv",
+            &[],
+            &BalancerAddresses {
+                compat_openai: None,
+                compat_typesafe: None,
+                inference: SocketAddr::from((Ipv4Addr::LOCALHOST, 8061)),
+                management: SocketAddr::from((Ipv4Addr::LOCALHOST, 8060)),
+                web_admin_panel: None,
+            },
+        );
+
+        assert_eq!(
+            command.as_std().get_envs().collect::<BTreeMap<_, _>>(),
+            BTreeMap::from([
+                (
+                    OsStr::new("PADDLER_INFERENCE_URL"),
+                    Some(OsStr::new("http://127.0.0.1:8061"))
+                ),
+                (
+                    OsStr::new("PADDLER_MANAGEMENT_URL"),
+                    Some(OsStr::new("http://127.0.0.1:8060"))
+                ),
+            ])
+        );
+    }
+}
diff --git a/paddler_tests/src/chat_template_rendering_every_message_after.rs b/paddler_tests/src/chat_template_rendering_every_message_after.rs
new file mode 100644
index 000000000..1353dc307
--- /dev/null
+++ b/paddler_tests/src/chat_template_rendering_every_message_after.rs
@@ -0,0 +1,10 @@
+use paddler_messaging::chat_template::ChatTemplate;
+
+const EVERY_MESSAGE_AND_THE_GENERATION_PROMPT: &str = "{%- for message in messages -%}{{ message['role'] }}: {% if message['content'] is string %}{{ message['content'] }}{% else %}{% for part in message['content'] %}{{ part['text'] }}{% endfor %}{% endif %}\n{% endfor -%}{% if add_generation_prompt %}assistant: {% endif %}";
+
+#[must_use]
+pub fn chat_template_rendering_every_message_after(prefix: &str) -> ChatTemplate {
+    ChatTemplate {
+        content: format!("{prefix}{EVERY_MESSAGE_AND_THE_GENERATION_PROMPT}"),
+    }
+}
diff --git a/paddler_tests/src/cluster_without_agents_serving_decisions.rs b/paddler_tests/src/cluster_without_agents_serving_decisions.rs
new file mode 100644
index 000000000..0f50de6f4
--- /dev/null
+++ b/paddler_tests/src/cluster_without_agents_serving_decisions.rs
@@ -0,0 +1,12 @@
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+
+use crate::desired_state_serving_decisions::desired_state_serving_decisions;
+
+#[must_use]
+pub fn cluster_without_agents_serving_decisions() -> ClusterParams {
+    ClusterParams {
+        desired_state: ClusterDesiredState::Apply(Box::new(desired_state_serving_decisions())),
+        ..ClusterParams::without_agents()
+    }
+}
diff --git a/paddler_tests/src/conversation_generation.rs b/paddler_tests/src/conversation_generation.rs
new file mode 100644
index 000000000..01e73e65e
--- /dev/null
+++ b/paddler_tests/src/conversation_generation.rs
@@ -0,0 +1,42 @@
+use std::num::NonZeroU16;
+use std::sync::Arc;
+
+use tokio::sync::mpsc;
+
+use paddler_agent_runtime::agent_request::AgentRequest;
+use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
+use paddler_agent_status::slot_guard::SlotGuard;
+use paddler_messaging::generated_token_result::GeneratedTokenResult;
+use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
+use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters_schema::validated_parameters_schema::ValidatedParametersSchema;
+
+pub struct ConversationGeneration {
+    pub agent_request: AgentRequest<
+        ContinueFromConversationHistoryParams<ValidatedParametersSchema>,
+        GeneratedTokenResult,
+    >,
+    pub generate_tokens_stop_tx: mpsc::UnboundedSender<()>,
+    pub generated_tokens_rx: mpsc::UnboundedReceiver<GeneratedTokenResult>,
+}
+
+impl ConversationGeneration {
+    #[must_use]
+    pub fn new(
+        params: ContinueFromConversationHistoryParams<ValidatedParametersSchema>,
+        slot_aggregated_status: Arc<SlotAggregatedStatus>,
+    ) -> Self {
+        let (generated_tokens_tx, generated_tokens_rx) = mpsc::unbounded_channel();
+        let (generate_tokens_stop_tx, generate_tokens_stop_rx) = mpsc::unbounded_channel();
+
+        Self {
+            agent_request: AgentRequest {
+                params,
+                response_tx: generated_tokens_tx,
+                slot_guard: SlotGuard::new(slot_aggregated_status, NonZeroU16::MIN),
+                stop_rx: generate_tokens_stop_rx,
+            },
+            generate_tokens_stop_tx,
+            generated_tokens_rx,
+        }
+    }
+}
diff --git a/paddler_tests/src/conversation_with_a_tool_requiring_an_undeclared_property.rs b/paddler_tests/src/conversation_with_a_tool_requiring_an_undeclared_property.rs
index 9e26c335d..ac73149dd 100644
--- a/paddler_tests/src/conversation_with_a_tool_requiring_an_undeclared_property.rs
+++ b/paddler_tests/src/conversation_with_a_tool_requiring_an_undeclared_property.rs
@@ -24,7 +24,7 @@ pub fn conversation_with_a_tool_requiring_an_undeclared_property()
 
     ContinueFromConversationHistoryParams {
         add_generation_prompt: true,
-        conversation_history: ConversationHistory::new(vec![ConversationMessage {
+        conversation_history: ConversationHistory::new(vec![ConversationMessage::Message {
             content: ConversationMessageContent::Text("Say hello".to_owned()),
             role: "user".to_owned(),
         }]),
diff --git a/paddler_tests/src/decision_backbone.rs b/paddler_tests/src/decision_backbone.rs
new file mode 100644
index 000000000..566d679dd
--- /dev/null
+++ b/paddler_tests/src/decision_backbone.rs
@@ -0,0 +1,23 @@
+use std::path::PathBuf;
+
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_model_source::model_source_error::ModelSourceError;
+use paddler_test_cluster_harness::agent_desired_model_at::agent_desired_model_at;
+use paddler_test_cluster_harness::model_card::ModelCard;
+
+pub enum DecisionBackbone {
+    HuggingFace(ModelCard),
+    LocalFile(PathBuf),
+}
+
+impl DecisionBackbone {
+    pub fn into_desired_state(self) -> Result<BalancerDesiredState, ModelSourceError> {
+        match self {
+            Self::HuggingFace(model_card) => Ok(model_card.into_desired_state()),
+            Self::LocalFile(model_path) => Ok(BalancerDesiredState {
+                model: agent_desired_model_at(&model_path)?,
+                ..BalancerDesiredState::default()
+            }),
+        }
+    }
+}
diff --git a/paddler_tests/src/decision_cluster_params.rs b/paddler_tests/src/decision_cluster_params.rs
new file mode 100644
index 000000000..52ac6f3aa
--- /dev/null
+++ b/paddler_tests/src/decision_cluster_params.rs
@@ -0,0 +1,42 @@
+use paddler_messaging::decision_slots_minimum::DECISION_SLOTS_MINIMUM;
+use paddler_model_source::model_source_error::ModelSourceError;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+
+use crate::decision_desired_state::DecisionDesiredState;
+
+pub struct DecisionClusterParams {
+    pub agents: Vec<AgentConfig>,
+    pub desired_state: DecisionDesiredState,
+    pub wait_for_slots_ready: bool,
+}
+
+impl DecisionClusterParams {
+    pub fn into_cluster_params(self) -> Result<ClusterParams, ModelSourceError> {
+        let Self {
+            agents,
+            desired_state,
+            wait_for_slots_ready,
+        } = self;
+
+        Ok(ClusterParams {
+            agents,
+            desired_state: ClusterDesiredState::Apply(Box::new(
+                desired_state.into_balancer_desired_state()?,
+            )),
+            wait_for_slots_ready,
+            ..ClusterParams::default()
+        })
+    }
+}
+
+impl Default for DecisionClusterParams {
+    fn default() -> Self {
+        Self {
+            agents: AgentConfig::uniform(1, DECISION_SLOTS_MINIMUM.get()),
+            desired_state: DecisionDesiredState::default(),
+            wait_for_slots_ready: true,
+        }
+    }
+}
diff --git a/paddler_tests/src/decision_desired_state.rs b/paddler_tests/src/decision_desired_state.rs
new file mode 100644
index 000000000..2f84debfe
--- /dev/null
+++ b/paddler_tests/src/decision_desired_state.rs
@@ -0,0 +1,65 @@
+use std::num::NonZeroU32;
+use std::path::PathBuf;
+
+use paddler_inference_parameters::all_gpu_layers::ALL_GPU_LAYERS;
+use paddler_inference_parameters::batch_size::BatchSize;
+use paddler_inference_parameters::embedding_parameters::EmbeddingParameters;
+use paddler_inference_parameters::model_runtime_parameters::ModelRuntimeParameters;
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_messaging::decision_settings::DecisionSettings;
+use paddler_model_source::model_source_error::ModelSourceError;
+use paddler_test_cluster_harness::agent_desired_model_at::agent_desired_model_at;
+use paddler_test_cluster_harness::fixture_path::fixture_path;
+use paddler_test_cluster_harness::model_card::qwen3_5_0_8b::qwen3_5_0_8b;
+use paddler_test_cluster_harness::synthetic_pointer_head_fixture::QWEN3_5_0_8B_SYNTHETIC_POINTER_HEAD_FIXTURE;
+
+use crate::decision_backbone::DecisionBackbone;
+
+pub struct DecisionDesiredState {
+    pub backbone: DecisionBackbone,
+    pub context_size: NonZeroU32,
+    pub embedding_parameters: EmbeddingParameters,
+    pub n_batch: BatchSize,
+    pub n_gpu_layers: i32,
+    pub pointer_head: PathBuf,
+}
+
+impl DecisionDesiredState {
+    pub fn into_balancer_desired_state(self) -> Result<BalancerDesiredState, ModelSourceError> {
+        let Self {
+            backbone,
+            context_size,
+            embedding_parameters,
+            n_batch,
+            n_gpu_layers,
+            pointer_head: pointer_head_path,
+        } = self;
+        let desired_state = backbone.into_desired_state()?;
+        let pointer_head = agent_desired_model_at(&pointer_head_path)?;
+
+        Ok(BalancerDesiredState {
+            decision: DecisionSettings { pointer_head },
+            embeddings: embedding_parameters,
+            model_runtime_parameters: ModelRuntimeParameters {
+                context_size,
+                n_batch,
+                n_gpu_layers,
+                ..desired_state.model_runtime_parameters
+            },
+            ..desired_state
+        })
+    }
+}
+
+impl Default for DecisionDesiredState {
+    fn default() -> Self {
+        Self {
+            backbone: DecisionBackbone::HuggingFace(qwen3_5_0_8b()),
+            context_size: ModelRuntimeParameters::default().context_size,
+            embedding_parameters: EmbeddingParameters::default(),
+            n_batch: BatchSize::default(),
+            n_gpu_layers: ALL_GPU_LAYERS,
+            pointer_head: fixture_path(QWEN3_5_0_8B_SYNTHETIC_POINTER_HEAD_FIXTURE),
+        }
+    }
+}
diff --git a/paddler_tests/src/decision_done_discriminant.rs b/paddler_tests/src/decision_done_discriminant.rs
new file mode 100644
index 000000000..76e273c1f
--- /dev/null
+++ b/paddler_tests/src/decision_done_discriminant.rs
@@ -0,0 +1,13 @@
+use std::mem::Discriminant;
+use std::mem::discriminant;
+
+use paddler_messaging::decision_result::DecisionResult;
+use paddler_messaging::decision_summary::DecisionSummary;
+
+#[must_use]
+pub const fn decision_done_discriminant() -> Discriminant<DecisionResult> {
+    discriminant(&DecisionResult::Done(DecisionSummary {
+        input_tokens: 0,
+        processing_milliseconds: 0,
+    }))
+}
diff --git a/paddler_tests/src/decision_scheduler_run.rs b/paddler_tests/src/decision_scheduler_run.rs
new file mode 100644
index 000000000..9040322e9
--- /dev/null
+++ b/paddler_tests/src/decision_scheduler_run.rs
@@ -0,0 +1,17 @@
+use std::collections::VecDeque;
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent_decision::active_decision_request::ActiveDecisionRequest;
+use paddler_agent_decision::decision_capacity_ledger::DecisionCapacityLedger;
+use paddler_agent_decision::prepared_decision_request::PreparedDecisionRequest;
+use paddler_agent_runtime::scheduler_inbox::SchedulerInbox;
+
+pub struct DecisionSchedulerRun {
+    pub active_requests: VecDeque<ActiveDecisionRequest>,
+    pub agent_shutdown: CancellationToken,
+    pub capacity_ledger: DecisionCapacityLedger,
+    pub pending_requests: VecDeque<PreparedDecisionRequest>,
+    pub scheduler_inbox: SchedulerInbox<PreparedDecisionRequest>,
+    pub sequences: u16,
+}
diff --git a/paddler_tests/src/desired_state_reconciled_discriminant.rs b/paddler_tests/src/desired_state_reconciled_discriminant.rs
new file mode 100644
index 000000000..11de3bfaf
--- /dev/null
+++ b/paddler_tests/src/desired_state_reconciled_discriminant.rs
@@ -0,0 +1,9 @@
+use std::mem::Discriminant;
+use std::mem::discriminant;
+
+use paddler_agent::desired_state_reconciliation::DesiredStateReconciliation;
+
+#[must_use]
+pub fn desired_state_reconciled_discriminant() -> Discriminant<DesiredStateReconciliation> {
+    discriminant(&DesiredStateReconciliation::Reconciled(Box::default()))
+}
diff --git a/paddler_tests/src/desired_state_serving_decisions.rs b/paddler_tests/src/desired_state_serving_decisions.rs
new file mode 100644
index 000000000..0107b6824
--- /dev/null
+++ b/paddler_tests/src/desired_state_serving_decisions.rs
@@ -0,0 +1,15 @@
+use paddler_messaging::agent_desired_model::AgentDesiredModel;
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_messaging::decision_settings::DecisionSettings;
+
+const POINTER_HEAD_NO_AGENT_LOADS: &str = "agent:///pointer-head-no-agent-loads.gguf";
+
+#[must_use]
+pub fn desired_state_serving_decisions() -> BalancerDesiredState {
+    BalancerDesiredState {
+        decision: DecisionSettings {
+            pointer_head: AgentDesiredModel::Uri(POINTER_HEAD_NO_AGENT_LOADS.to_owned()),
+        },
+        ..BalancerDesiredState::default()
+    }
+}
diff --git a/paddler_tests/src/desired_state_with_chat_template_override.rs b/paddler_tests/src/desired_state_with_chat_template_override.rs
new file mode 100644
index 000000000..2d9225d94
--- /dev/null
+++ b/paddler_tests/src/desired_state_with_chat_template_override.rs
@@ -0,0 +1,18 @@
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_messaging::balancer_text_generation_settings::BalancerTextGenerationSettings;
+use paddler_messaging::chat_template::ChatTemplate;
+use paddler_messaging::chat_template_source::ChatTemplateSource;
+
+#[must_use]
+pub fn desired_state_with_chat_template_override(
+    desired_state: BalancerDesiredState,
+    chat_template: ChatTemplate,
+) -> BalancerDesiredState {
+    BalancerDesiredState {
+        text_generation: BalancerTextGenerationSettings {
+            chat_template_source: ChatTemplateSource::Override(chat_template),
+            ..desired_state.text_generation
+        },
+        ..desired_state
+    }
+}
diff --git a/paddler_tests/src/desired_state_with_halved_image_resize.rs b/paddler_tests/src/desired_state_with_halved_image_resize.rs
index 97a51a11b..c2e5df412 100644
--- a/paddler_tests/src/desired_state_with_halved_image_resize.rs
+++ b/paddler_tests/src/desired_state_with_halved_image_resize.rs
@@ -1,20 +1,27 @@
 use std::num::NonZeroU32;
 
-use anyhow::Result;
-
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
 use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_messaging::balancer_text_generation_settings::BalancerTextGenerationSettings;
+use paddler_messaging::multimodal_settings::MultimodalSettings;
+
+const HALF: NonZeroU32 = NonZeroU32::new(2).unwrap();
 
+#[must_use]
 pub fn desired_state_with_halved_image_resize(
     desired_state: BalancerDesiredState,
-) -> Result<BalancerDesiredState> {
-    Ok(BalancerDesiredState {
-        inference_parameters: InferenceParameters {
-            image_resize_to_fit: NonZeroU32::try_from(
-                desired_state.inference_parameters.image_resize_to_fit.get() / 2,
-            )?,
-            ..desired_state.inference_parameters.clone()
+) -> BalancerDesiredState {
+    BalancerDesiredState {
+        text_generation: BalancerTextGenerationSettings {
+            multimodal: MultimodalSettings {
+                image_resize_to_fit: desired_state
+                    .text_generation
+                    .multimodal
+                    .image_resize_to_fit
+                    .div_ceil(HALF),
+                ..desired_state.text_generation.multimodal
+            },
+            ..desired_state.text_generation
         },
         ..desired_state
-    })
+    }
 }
diff --git a/paddler_tests/src/desired_state_with_multimodal_projection.rs b/paddler_tests/src/desired_state_with_multimodal_projection.rs
new file mode 100644
index 000000000..7c4b26c67
--- /dev/null
+++ b/paddler_tests/src/desired_state_with_multimodal_projection.rs
@@ -0,0 +1,21 @@
+use paddler_messaging::agent_desired_model::AgentDesiredModel;
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_messaging::balancer_text_generation_settings::BalancerTextGenerationSettings;
+use paddler_messaging::multimodal_settings::MultimodalSettings;
+
+#[must_use]
+pub fn desired_state_with_multimodal_projection(
+    desired_state: BalancerDesiredState,
+    projection: AgentDesiredModel,
+) -> BalancerDesiredState {
+    BalancerDesiredState {
+        text_generation: BalancerTextGenerationSettings {
+            multimodal: MultimodalSettings {
+                projection,
+                ..desired_state.text_generation.multimodal
+            },
+            ..desired_state.text_generation
+        },
+        ..desired_state
+    }
+}
diff --git a/paddler_tests/src/first_message_content_chat_template.rs b/paddler_tests/src/first_message_content_chat_template.rs
new file mode 100644
index 000000000..d62640126
--- /dev/null
+++ b/paddler_tests/src/first_message_content_chat_template.rs
@@ -0,0 +1,8 @@
+use paddler_messaging::chat_template::ChatTemplate;
+
+#[must_use]
+pub fn first_message_content_chat_template() -> ChatTemplate {
+    ChatTemplate {
+        content: "{{ messages[0].content }}".to_owned(),
+    }
+}
diff --git a/paddler_tests/src/gbnf_literal.rs b/paddler_tests/src/gbnf_literal.rs
index 041a60a7c..6676b2965 100644
--- a/paddler_tests/src/gbnf_literal.rs
+++ b/paddler_tests/src/gbnf_literal.rs
@@ -2,7 +2,10 @@ use paddler_messaging::grammar_constraint::GrammarConstraint;
 
 #[must_use]
 pub fn gbnf_literal(text: &str) -> GrammarConstraint {
-    let escaped_text = text.replace('\\', "\\\\").replace('"', "\\\"");
+    let escaped_text = text
+        .replace('\\', "\\\\")
+        .replace('"', "\\\"")
+        .replace('\n', "\\n");
 
     GrammarConstraint::Gbnf {
         grammar: format!("root ::= \"{escaped_text}\""),
diff --git a/paddler_tests/src/generation_done_discriminant.rs b/paddler_tests/src/generation_done_discriminant.rs
new file mode 100644
index 000000000..f79fc32a7
--- /dev/null
+++ b/paddler_tests/src/generation_done_discriminant.rs
@@ -0,0 +1,16 @@
+use std::mem::Discriminant;
+use std::mem::discriminant;
+
+use llama_cpp_bindings::TokenUsage;
+
+use paddler_messaging::generated_token_result::GeneratedTokenResult;
+use paddler_messaging::generation_finish::GenerationFinish;
+use paddler_messaging::generation_summary::GenerationSummary;
+
+#[must_use]
+pub const fn generation_done_discriminant() -> Discriminant<GeneratedTokenResult> {
+    discriminant(&GeneratedTokenResult::Done(GenerationSummary {
+        finish: GenerationFinish::EndOfGeneration,
+        usage: TokenUsage::new(),
+    }))
+}
diff --git a/paddler_tests/src/image_description_conversation.rs b/paddler_tests/src/image_description_conversation.rs
index c67f6f46f..766117a94 100644
--- a/paddler_tests/src/image_description_conversation.rs
+++ b/paddler_tests/src/image_description_conversation.rs
@@ -7,19 +7,19 @@ use paddler_messaging::image_url::ImageUrl;
 #[must_use]
 pub fn image_description_conversation(image_data_uri: String) -> ConversationHistory {
     ConversationHistory::new(vec![
-        ConversationMessage {
+        ConversationMessage::Message {
             content: ConversationMessageContent::Text(
                 "You are a helpful assistant. Give engaging, short, precise answers.".to_owned(),
             ),
             role: "system".to_owned(),
         },
-        ConversationMessage {
+        ConversationMessage::Message {
             content: ConversationMessageContent::Text(
                 "Hello! How can I help you today?".to_owned(),
             ),
             role: "assistant".to_owned(),
         },
-        ConversationMessage {
+        ConversationMessage::Message {
             content: ConversationMessageContent::Parts(vec![
                 ConversationMessageContentPart::ImageUrl {
                     image_url: ImageUrl {
diff --git a/paddler_tests/src/image_fixture.rs b/paddler_tests/src/image_fixture.rs
new file mode 100644
index 000000000..a0f088503
--- /dev/null
+++ b/paddler_tests/src/image_fixture.rs
@@ -0,0 +1,14 @@
+use anyhow::Result;
+
+use paddler_test_cluster_harness::load_fixture_data_uri::load_fixture_data_uri;
+
+pub struct ImageFixture {
+    pub fixture_name: &'static str,
+    pub mime_type: &'static str,
+}
+
+impl ImageFixture {
+    pub fn data_uri(&self) -> Result<String> {
+        load_fixture_data_uri(self.fixture_name, self.mime_type)
+    }
+}
diff --git a/paddler_tests/src/in_process_agent.rs b/paddler_tests/src/in_process_agent.rs
index 4fcc927fe..2aedd7854 100644
--- a/paddler_tests/src/in_process_agent.rs
+++ b/paddler_tests/src/in_process_agent.rs
@@ -1,9 +1,9 @@
-use anyhow::Error;
 use anyhow::Result;
 use async_trait::async_trait;
 
 use paddler_bootstrap::agent_runner::AgentRunner;
 use paddler_test_cluster_harness::managed_process::ManagedProcess;
+use paddler_test_cluster_harness::process_end::ProcessEnd;
 
 pub struct InProcessAgent {
     runner: AgentRunner,
@@ -16,10 +16,24 @@ impl InProcessAgent {
     }
 }
 
+impl InProcessAgent {
+    async fn wait_for_exit(&mut self) -> ProcessEnd {
+        self.runner
+            .wait_for_completion()
+            .await
+            .map_or_else(ProcessEnd::failed, |()| ProcessEnd::Clean)
+    }
+}
+
 #[async_trait]
 impl ManagedProcess for InProcessAgent {
-    async fn shutdown(self: Box<Self>) -> Result<()> {
+    fn terminate(&mut self) -> Result<()> {
         self.runner.cancel();
-        self.runner.wait_for_completion().await.map_err(Error::new)
+
+        Ok(())
+    }
+
+    async fn exited(&mut self) -> ProcessEnd {
+        self.wait_for_exit().await
     }
 }
diff --git a/paddler_tests/src/in_process_agent_spawner.rs b/paddler_tests/src/in_process_agent_spawner.rs
index 001af4bab..2edca415e 100644
--- a/paddler_tests/src/in_process_agent_spawner.rs
+++ b/paddler_tests/src/in_process_agent_spawner.rs
@@ -1,9 +1,11 @@
 use anyhow::Result;
+use tempfile::TempDir;
 use tokio_util::sync::CancellationToken;
 
 use paddler_bootstrap::agent_bootstrap_config::AgentBootstrapConfig;
 use paddler_bootstrap::agent_runner::AgentRunner;
 use paddler_bootstrap::agent_runner_params::AgentRunnerParams;
+use paddler_model_source::desired_model_resolver::DesiredModelResolver;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
 use paddler_test_cluster_harness::agent_spawner::AgentSpawner;
 use paddler_test_cluster_harness::managed_process::ManagedProcess;
@@ -12,28 +14,24 @@ use paddler_test_cluster_harness::unexpiring_shutdown_options::unexpiring_shutdo
 use crate::in_process_agent::InProcessAgent;
 
 pub struct InProcessAgentSpawner {
-    management_address: String,
-}
-
-impl InProcessAgentSpawner {
-    #[must_use]
-    pub const fn new(management_address: String) -> Self {
-        Self { management_address }
-    }
+    pub desired_model_resolver: DesiredModelResolver,
+    pub downloaded_models: TempDir,
+    pub management_address: String,
 }
 
 impl AgentSpawner for InProcessAgentSpawner {
     fn spawn(&self, config: &AgentConfig) -> Result<Box<dyn ManagedProcess>> {
-        let runner = AgentRunner::start(AgentRunnerParams {
-            bootstrap_config: AgentBootstrapConfig {
-                agent_name: Some(config.name.clone()),
-                management_address: self.management_address.clone(),
-                slots: config.slot_count,
+        Ok(Box::new(InProcessAgent::new(AgentRunner::start(
+            AgentRunnerParams {
+                bootstrap_config: AgentBootstrapConfig {
+                    agent_name: Some(config.name.clone()),
+                    desired_model_resolver: self.desired_model_resolver.clone(),
+                    management_address: self.management_address.clone(),
+                    slots: config.slot_count,
+                },
+                cancellation_token: CancellationToken::new(),
+                shutdown_options: unexpiring_shutdown_options(),
             },
-            cancellation_token: CancellationToken::new(),
-            shutdown_options: unexpiring_shutdown_options(),
-        });
-
-        Ok(Box::new(InProcessAgent::new(runner)))
+        ))))
     }
 }
diff --git a/paddler_tests/src/in_process_balancer.rs b/paddler_tests/src/in_process_balancer.rs
index 83d6c3814..e812e6eb8 100644
--- a/paddler_tests/src/in_process_balancer.rs
+++ b/paddler_tests/src/in_process_balancer.rs
@@ -1,9 +1,9 @@
-use anyhow::Error;
 use anyhow::Result;
 use async_trait::async_trait;
 
 use paddler_bootstrap::balancer_runner::BalancerRunner;
 use paddler_test_cluster_harness::managed_process::ManagedProcess;
+use paddler_test_cluster_harness::process_end::ProcessEnd;
 
 pub struct InProcessBalancer {
     runner: BalancerRunner,
@@ -16,10 +16,24 @@ impl InProcessBalancer {
     }
 }
 
+impl InProcessBalancer {
+    async fn wait_for_exit(&mut self) -> ProcessEnd {
+        self.runner
+            .wait_for_completion()
+            .await
+            .map_or_else(ProcessEnd::failed, |()| ProcessEnd::Clean)
+    }
+}
+
 #[async_trait]
 impl ManagedProcess for InProcessBalancer {
-    async fn shutdown(self: Box<Self>) -> Result<()> {
+    fn terminate(&mut self) -> Result<()> {
         self.runner.cancel();
-        self.runner.wait_for_completion().await.map_err(Error::new)
+
+        Ok(())
+    }
+
+    async fn exited(&mut self) -> ProcessEnd {
+        self.wait_for_exit().await
     }
 }
diff --git a/paddler_tests/src/in_process_balancer_runner_params.rs b/paddler_tests/src/in_process_balancer_runner_params.rs
new file mode 100644
index 000000000..a5b383967
--- /dev/null
+++ b/paddler_tests/src/in_process_balancer_runner_params.rs
@@ -0,0 +1,58 @@
+use std::str::FromStr as _;
+
+use anyhow::Context as _;
+use anyhow::Result;
+use tokio_util::sync::CancellationToken;
+
+use paddler_balancer::compatibility::compatibility_service_configuration::CompatibilityServiceConfiguration;
+use paddler_balancer::inference_service::configuration::Configuration as InferenceServiceConfiguration;
+use paddler_balancer::management_service::configuration::Configuration as ManagementServiceConfiguration;
+use paddler_balancer::resolved_socket_addr::ResolvedSocketAddr;
+use paddler_bootstrap::balancer_bootstrap_config::BalancerBootstrapConfig;
+use paddler_bootstrap::balancer_runner_params::BalancerRunnerParams;
+use paddler_state_database::state_database_type::StateDatabaseType;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_test_cluster_harness::ephemeral_loopback_addr::EPHEMERAL_LOOPBACK_ADDR;
+use paddler_test_cluster_harness::unexpiring_shutdown_options::unexpiring_shutdown_options;
+
+pub fn in_process_balancer_runner_params(
+    ClusterParams {
+        buffered_request_timeout,
+        inference_cors_allowed_hosts,
+        inference_item_timeout,
+        management_cors_allowed_hosts,
+        max_buffered_requests,
+        state_database_url,
+        ..
+    }: &ClusterParams,
+) -> Result<BalancerRunnerParams> {
+    Ok(BalancerRunnerParams {
+        bootstrap_config: BalancerBootstrapConfig {
+            buffered_request_timeout: *buffered_request_timeout,
+            inference_service_configuration: InferenceServiceConfiguration {
+                addr: ResolvedSocketAddr::from(EPHEMERAL_LOOPBACK_ADDR),
+                cors_allowed_hosts: inference_cors_allowed_hosts.clone(),
+                inference_item_timeout: *inference_item_timeout,
+            },
+            management_service_configuration: ManagementServiceConfiguration {
+                addr: ResolvedSocketAddr::from(EPHEMERAL_LOOPBACK_ADDR),
+                cors_allowed_hosts: management_cors_allowed_hosts.clone(),
+            },
+            max_buffered_requests: *max_buffered_requests,
+            openai_service_configuration: Some(CompatibilityServiceConfiguration {
+                addr: ResolvedSocketAddr::from(EPHEMERAL_LOOPBACK_ADDR),
+            }),
+            state_database_type: StateDatabaseType::from_str(state_database_url)
+                .context("failed to parse state_database_url")?,
+            statsd_prefix: "paddler_tests_".to_owned(),
+            statsd_service_configuration: None,
+            typesafe_service_configuration: Some(CompatibilityServiceConfiguration {
+                addr: ResolvedSocketAddr::from(EPHEMERAL_LOOPBACK_ADDR),
+            }),
+            #[cfg(feature = "web_admin_panel")]
+            web_admin_panel_service_configuration: None,
+        },
+        cancellation_token: CancellationToken::new(),
+        shutdown_options: unexpiring_shutdown_options(),
+    })
+}
diff --git a/paddler_tests/src/inference_runtime_context_with_unknown_parallelism.rs b/paddler_tests/src/inference_runtime_context_with_unknown_parallelism.rs
new file mode 100644
index 000000000..19e378aa7
--- /dev/null
+++ b/paddler_tests/src/inference_runtime_context_with_unknown_parallelism.rs
@@ -0,0 +1,14 @@
+use std::sync::Arc;
+
+use paddler_agent_runtime::inference_runtime_context::InferenceRuntimeContext;
+
+use crate::agent_inference_runtime_context::agent_inference_runtime_context;
+use crate::parallelism_that_is_unknown::ParallelismThatIsUnknown;
+
+#[must_use]
+pub fn inference_runtime_context_with_unknown_parallelism(slots: u16) -> InferenceRuntimeContext {
+    InferenceRuntimeContext {
+        available_parallelism: Arc::new(ParallelismThatIsUnknown),
+        ..agent_inference_runtime_context(slots)
+    }
+}
diff --git a/paddler_tests/src/jpeg_and_webp_image_fixtures.rs b/paddler_tests/src/jpeg_and_webp_image_fixtures.rs
new file mode 100644
index 000000000..eecfba1dc
--- /dev/null
+++ b/paddler_tests/src/jpeg_and_webp_image_fixtures.rs
@@ -0,0 +1,12 @@
+use crate::image_fixture::ImageFixture;
+
+pub const JPEG_AND_WEBP_IMAGE_FIXTURES: [ImageFixture; 2] = [
+    ImageFixture {
+        fixture_name: "sarnow.jpeg",
+        mime_type: "image/jpeg",
+    },
+    ImageFixture {
+        fixture_name: "llamas.webp",
+        mime_type: "image/webp",
+    },
+];
diff --git a/paddler_tests/src/kev_0_8b_pointer_head.rs b/paddler_tests/src/kev_0_8b_pointer_head.rs
new file mode 100644
index 000000000..60faf9f00
--- /dev/null
+++ b/paddler_tests/src/kev_0_8b_pointer_head.rs
@@ -0,0 +1,8 @@
+use paddler_agent_pointer_head::pointer_head::PointerHead;
+use paddler_agent_pointer_head::pointer_head_error::PointerHeadError;
+
+use crate::kev_checkpoint::KevCheckpoint;
+
+pub fn kev_0_8b_pointer_head() -> Result<PointerHead, PointerHeadError> {
+    PointerHead::load(&KevCheckpoint::Kev0_8b.target_path("pointer_head.gguf"))
+}
diff --git a/paddler_tests/src/kev_checkpoint.rs b/paddler_tests/src/kev_checkpoint.rs
new file mode 100644
index 000000000..755022007
--- /dev/null
+++ b/paddler_tests/src/kev_checkpoint.rs
@@ -0,0 +1,57 @@
+use std::path::PathBuf;
+
+use paddler_inference_parameters::all_gpu_layers::ALL_GPU_LAYERS;
+
+use crate::decision_backbone::DecisionBackbone;
+use crate::decision_cluster_params::DecisionClusterParams;
+use crate::decision_desired_state::DecisionDesiredState;
+
+const KEV_27B_GPU_LAYERS_FITTING_BESIDE_EVERY_CONTEXT: i32 = 55;
+
+#[derive(Clone, Copy)]
+pub enum KevCheckpoint {
+    Kev0_8b,
+    Kev27b,
+}
+
+impl KevCheckpoint {
+    #[must_use]
+    pub fn decision_cluster_params(self) -> DecisionClusterParams {
+        DecisionClusterParams {
+            desired_state: self.decision_desired_state(),
+            ..DecisionClusterParams::default()
+        }
+    }
+
+    #[must_use]
+    pub fn decision_desired_state(self) -> DecisionDesiredState {
+        DecisionDesiredState {
+            backbone: DecisionBackbone::LocalFile(self.target_path("model.gguf")),
+            n_gpu_layers: self.gpu_layers(),
+            pointer_head: self.target_path("pointer_head.gguf"),
+            ..DecisionDesiredState::default()
+        }
+    }
+
+    const fn gpu_layers(self) -> i32 {
+        match self {
+            Self::Kev0_8b => ALL_GPU_LAYERS,
+            Self::Kev27b => KEV_27B_GPU_LAYERS_FITTING_BESIDE_EVERY_CONTEXT,
+        }
+    }
+
+    #[must_use]
+    pub fn target_path(self, artifact_file_name: &str) -> PathBuf {
+        PathBuf::from(env!("CARGO_MANIFEST_DIR"))
+            .with_file_name("target")
+            .join(self.target_directory_name())
+            .join(artifact_file_name)
+    }
+
+    const fn target_directory_name(self) -> &'static str {
+        match self {
+            Self::Kev0_8b => "kev_0_8b",
+            Self::Kev27b => "kev_27b",
+        }
+    }
+}
diff --git a/paddler_tests/src/kev_parity_record.rs b/paddler_tests/src/kev_parity_record.rs
new file mode 100644
index 000000000..07559591b
--- /dev/null
+++ b/paddler_tests/src/kev_parity_record.rs
@@ -0,0 +1,11 @@
+use serde::Deserialize;
+
+use paddler_messaging::decision_question::DecisionQuestion;
+
+#[derive(Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct KevParityRecord {
+    pub probabilities: Vec<Vec<f32>>,
+    pub questions: Vec<DecisionQuestion>,
+    pub state: String,
+}
diff --git a/paddler_tests/src/lib.rs b/paddler_tests/src/lib.rs
index b2905e7be..8c7ea4298 100644
--- a/paddler_tests/src/lib.rs
+++ b/paddler_tests/src/lib.rs
@@ -1,19 +1,59 @@
-pub mod batch_size_within_context;
+pub mod active_embedding_request_fixture;
+pub mod admitted_decision;
+pub mod agent_inference_runtime_context;
+pub mod chat_template_rendering_every_message_after;
+pub mod cluster_without_agents_serving_decisions;
+pub mod conversation_generation;
 pub mod conversation_with_a_tool_requiring_an_undeclared_property;
+pub mod decision_backbone;
+pub mod decision_cluster_params;
+pub mod decision_desired_state;
+pub mod decision_done_discriminant;
+pub mod decision_scheduler_run;
+pub mod desired_state_reconciled_discriminant;
+pub mod desired_state_serving_decisions;
+pub mod desired_state_with_chat_template_override;
 pub mod desired_state_with_halved_image_resize;
+pub mod desired_state_with_multimodal_projection;
+pub mod first_message_content_chat_template;
 pub mod gbnf_literal;
+pub mod generation_done_discriminant;
 pub mod get_weather_tool;
 pub mod image_description_conversation;
+pub mod image_fixture;
 pub mod in_process_agent;
 pub mod in_process_agent_spawner;
 pub mod in_process_balancer;
-pub mod load_qwen3_after_request_is_buffered;
+pub mod in_process_balancer_runner_params;
+pub mod inference_runtime_context_with_unknown_parallelism;
+pub mod jpeg_and_webp_image_fixtures;
+pub mod kev_0_8b_pointer_head;
+pub mod kev_checkpoint;
+pub mod kev_parity_record;
+pub mod llama_runtime;
+pub mod many_question_decision;
+pub mod max_tokens_for_forced_text;
+pub mod minimal_openai_chat_completion;
 pub mod ministral_3_cluster_params;
 pub mod nomic_embed_desired_state_with_chat_template_override;
+pub mod normalization_of_an_embedding_too_long_for_rms_norm;
+pub mod numbered_embedding_documents;
+pub mod oversized_text;
+pub mod parallelism_that_is_unknown;
+pub mod point_qwen3_special_token_at_unused_padding;
+pub mod post_openai_chat_completion;
+pub mod prepared_decision;
+pub mod question_answered_discriminant;
+pub mod questions_outlasting_a_disconnect;
+pub mod qwen3_0_6b_newline_token;
+pub mod qwen3_0_6b_unused_padding_token;
 pub mod qwen3_desired_state;
-pub mod qwen3_desired_state_with_embeddings;
-pub mod release_half_closed_openai_request;
+pub mod raw_prompt_generation;
+pub mod responses_message_text;
+pub mod sample_decision;
 pub mod send_cors_preflight;
+pub mod short_buffered_request_timeout;
+pub mod single_image_question;
 pub mod smolvlm2_desired_state;
 pub mod start_cluster;
 pub mod start_cluster_with_deepseek_r1_distill_llama_8b;
@@ -31,8 +71,18 @@ pub mod start_cluster_with_qwen3_and_context_size;
 pub mod start_cluster_with_smolvlm2;
 pub mod start_cluster_with_smolvlm2_and_context_size;
 pub mod start_cluster_with_smolvlm2_and_n_batch;
-pub mod start_cluster_without_model;
+pub mod start_decision_cluster;
 pub mod start_embedding_cluster;
 pub mod start_single_agent_cluster_with_desired_state;
+pub mod test_cluster_preset;
+pub mod text_generation_pipeline_parts;
+pub mod tokenizer_reference;
+pub mod tokenizer_reference_entry;
+pub mod two_option_question;
+pub mod uncompilable_chat_template;
+pub mod uncompilable_chat_template_issue;
 pub mod weather_question_conversation;
 pub mod weather_tool_call_json;
+pub mod weather_tool_parameters_with_an_unsatisfiable_location;
+pub mod weather_tool_question;
+pub mod write_patched_gguf;
diff --git a/paddler_tests/src/llama_runtime.rs b/paddler_tests/src/llama_runtime.rs
new file mode 100644
index 000000000..7788a3c56
--- /dev/null
+++ b/paddler_tests/src/llama_runtime.rs
@@ -0,0 +1,95 @@
+use std::path::Path;
+
+use paddler_agent_decision::decision_error::DecisionError;
+use paddler_agent_decision::decision_llama_context_settings::DecisionLlamaContextSettings;
+use paddler_agent_embeddings::embedding_llama_context_settings::EmbeddingLlamaContextSettings;
+use paddler_agent_runtime::agent_runtime_error::AgentRuntimeError;
+use paddler_agent_runtime::inference_runtime_context::InferenceRuntimeContext;
+use paddler_agent_runtime::llama_context_settings::LlamaContextSettings;
+use paddler_agent_runtime::llama_model_settings::LlamaModelSettings;
+use paddler_agent_runtime::loaded_llama_model::LoadedLlamaModel;
+use paddler_agent_runtime::resident_llama_context::ResidentLlamaContext;
+use paddler_agent_text_generation::sampler_chain_factory::SamplerChainFactory;
+use paddler_inference_parameters::all_gpu_layers::ALL_GPU_LAYERS;
+use paddler_inference_parameters::model_runtime_parameters::ModelRuntimeParameters;
+use paddler_inference_parameters::pooling_type::PoolingType;
+use paddler_inference_parameters::sampling_parameters::SamplingParameters;
+use paddler_messaging::inference_capability::InferenceCapability;
+
+pub struct LlamaRuntime {
+    pub inference_runtime_context: InferenceRuntimeContext,
+    pub loaded_llama_model: LoadedLlamaModel,
+    pub model_runtime_parameters: ModelRuntimeParameters,
+}
+
+impl LlamaRuntime {
+    pub fn load(
+        model_path: &Path,
+        inference_runtime_context: InferenceRuntimeContext,
+    ) -> Result<Self, AgentRuntimeError> {
+        let model_runtime_parameters = ModelRuntimeParameters {
+            n_gpu_layers: ALL_GPU_LAYERS,
+            ..ModelRuntimeParameters::default()
+        };
+        let loaded_llama_model = LoadedLlamaModel::load(
+            &inference_runtime_context,
+            LlamaModelSettings {
+                model_path: model_path.to_path_buf(),
+                n_gpu_layers: model_runtime_parameters.n_gpu_layers,
+            },
+        )?;
+
+        Ok(Self {
+            inference_runtime_context,
+            loaded_llama_model,
+            model_runtime_parameters,
+        })
+    }
+
+    pub fn create_decision_llama_context(
+        &self,
+        n_seq_max: u32,
+    ) -> Result<ResidentLlamaContext<'_>, DecisionError> {
+        DecisionLlamaContextSettings {
+            llama_context_settings: self.llama_context_settings(n_seq_max),
+        }
+        .create_llama_context(&self.loaded_llama_model, &self.inference_runtime_context)
+    }
+
+    pub fn create_embedding_llama_context(
+        &self,
+        n_seq_max: u32,
+        pooling_type: PoolingType,
+    ) -> Result<ResidentLlamaContext<'_>, AgentRuntimeError> {
+        EmbeddingLlamaContextSettings {
+            llama_context_settings: self.llama_context_settings(n_seq_max),
+            pooling_type,
+        }
+        .create_llama_context(&self.loaded_llama_model, &self.inference_runtime_context)
+    }
+
+    pub fn create_text_generation_llama_context(
+        &self,
+        n_seq_max: u32,
+    ) -> Result<ResidentLlamaContext<'_>, AgentRuntimeError> {
+        self.loaded_llama_model.create_llama_context(
+            &self.inference_runtime_context,
+            InferenceCapability::TextGeneration,
+            self.llama_context_settings(n_seq_max)
+                .into_llama_context_params(),
+        )
+    }
+
+    #[must_use]
+    pub fn sampler_chain_factory(&self) -> SamplerChainFactory {
+        SamplerChainFactory {
+            sampling_parameters: SamplingParameters::default(),
+            n_vocab: self.loaded_llama_model.model.n_vocab(),
+        }
+    }
+
+    const fn llama_context_settings(&self, n_seq_max: u32) -> LlamaContextSettings<'_> {
+        self.loaded_llama_model
+            .llama_context_settings(&self.model_runtime_parameters, n_seq_max)
+    }
+}
diff --git a/paddler_tests/src/load_qwen3_after_request_is_buffered.rs b/paddler_tests/src/load_qwen3_after_request_is_buffered.rs
deleted file mode 100644
index a7dd02bc6..000000000
--- a/paddler_tests/src/load_qwen3_after_request_is_buffered.rs
+++ /dev/null
@@ -1,39 +0,0 @@
-use std::future::Future;
-
-use anyhow::Context as _;
-use anyhow::Result;
-use tokio::spawn;
-use tokio_util::sync::CancellationToken;
-
-use paddler_test_cluster_harness::cluster::Cluster;
-
-use crate::qwen3_desired_state_with_embeddings::qwen3_desired_state_with_embeddings;
-
-pub async fn load_qwen3_after_request_is_buffered<TRequest, TOutput>(
-    cluster: &mut Cluster,
-    request: TRequest,
-    enable_embeddings: bool,
-) -> Result<TOutput>
-where
-    TRequest: Future<Output = Result<TOutput>> + Send + 'static,
-    TOutput: Send + 'static,
-{
-    let buffered_request = spawn(request);
-
-    cluster
-        .wait_for_buffered_request_count(1)
-        .await
-        .context("the request must wait in the buffer while the agent has no model")?;
-
-    cluster
-        .client_management
-        .put_balancer_desired_state(
-            CancellationToken::new(),
-            &qwen3_desired_state_with_embeddings(enable_embeddings),
-        )
-        .await?;
-
-    buffered_request
-        .await
-        .context("the buffered request task must not panic")?
-}
diff --git a/paddler_tests/src/many_question_decision.rs b/paddler_tests/src/many_question_decision.rs
new file mode 100644
index 000000000..091e57a14
--- /dev/null
+++ b/paddler_tests/src/many_question_decision.rs
@@ -0,0 +1,23 @@
+use paddler_messaging::decision_question::DecisionQuestion;
+use paddler_messaging::request_params::decide_params::raw_decide_params::RawDecideParams;
+
+use crate::sample_decision::sample_decision;
+
+#[must_use]
+pub fn many_question_decision(question_count: usize) -> RawDecideParams {
+    let RawDecideParams { questions, state } = sample_decision();
+
+    RawDecideParams {
+        questions: questions
+            .iter()
+            .cycle()
+            .take(question_count)
+            .enumerate()
+            .map(|(question_index, question)| DecisionQuestion {
+                id: format!("{}-{question_index}", question.id),
+                ..question.clone()
+            })
+            .collect(),
+        state,
+    }
+}
diff --git a/paddler_tests/src/max_tokens_for_forced_text.rs b/paddler_tests/src/max_tokens_for_forced_text.rs
new file mode 100644
index 000000000..00145cc4f
--- /dev/null
+++ b/paddler_tests/src/max_tokens_for_forced_text.rs
@@ -0,0 +1,11 @@
+use std::num::NonZeroU32;
+
+#[must_use]
+pub fn max_tokens_for_forced_text(forced_text: &str) -> NonZeroU32 {
+    forced_text
+        .bytes()
+        .skip(1)
+        .fold(NonZeroU32::MIN, |max_tokens, _forced_byte| {
+            max_tokens.saturating_add(1)
+        })
+}
diff --git a/paddler_tests/src/minimal_openai_chat_completion.rs b/paddler_tests/src/minimal_openai_chat_completion.rs
new file mode 100644
index 000000000..da71d18da
--- /dev/null
+++ b/paddler_tests/src/minimal_openai_chat_completion.rs
@@ -0,0 +1,20 @@
+use std::future::Future;
+
+use reqwest::Error;
+use reqwest::Response;
+use reqwest::Url;
+use serde_json::json;
+
+use crate::post_openai_chat_completion::post_openai_chat_completion;
+
+pub fn minimal_openai_chat_completion(
+    chat_completions_url: Url,
+) -> impl Future<Output = Result<Response, Error>> + Send {
+    post_openai_chat_completion(
+        chat_completions_url,
+        &json!({
+            "model": "test-model",
+            "messages": [{"role": "user", "content": "hi"}]
+        }),
+    )
+}
diff --git a/paddler_tests/src/nomic_embed_desired_state_with_chat_template_override.rs b/paddler_tests/src/nomic_embed_desired_state_with_chat_template_override.rs
index ef0690b14..b34d93b85 100644
--- a/paddler_tests/src/nomic_embed_desired_state_with_chat_template_override.rs
+++ b/paddler_tests/src/nomic_embed_desired_state_with_chat_template_override.rs
@@ -2,13 +2,14 @@ use paddler_messaging::balancer_desired_state::BalancerDesiredState;
 use paddler_messaging::chat_template::ChatTemplate;
 use paddler_test_cluster_harness::model_card::nomic_embed_text_v1_5::nomic_embed_text_v1_5;
 
+use crate::desired_state_with_chat_template_override::desired_state_with_chat_template_override;
+
 #[must_use]
 pub fn nomic_embed_desired_state_with_chat_template_override(
     chat_template: ChatTemplate,
 ) -> BalancerDesiredState {
-    BalancerDesiredState {
-        chat_template_override: Some(chat_template),
-        use_chat_template_override: true,
-        ..nomic_embed_text_v1_5().into_desired_state()
-    }
+    desired_state_with_chat_template_override(
+        nomic_embed_text_v1_5().into_desired_state(),
+        chat_template,
+    )
 }
diff --git a/paddler_tests/src/normalization_of_an_embedding_too_long_for_rms_norm.rs b/paddler_tests/src/normalization_of_an_embedding_too_long_for_rms_norm.rs
new file mode 100644
index 000000000..2228110f2
--- /dev/null
+++ b/paddler_tests/src/normalization_of_an_embedding_too_long_for_rms_norm.rs
@@ -0,0 +1,21 @@
+use paddler_agent_embeddings::embedding_error::EmbeddingError;
+use paddler_agent_embeddings::normalization::normalize_embedding::normalize_embedding;
+use paddler_agent_embeddings::normalizes_embedding::NormalizesEmbedding;
+use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
+
+pub const DIMENSIONS_TOO_MANY_FOR_RMS_NORM: usize = 1 << 16;
+
+pub struct NormalizationOfAnEmbeddingTooLongForRmsNorm;
+
+impl NormalizesEmbedding for NormalizationOfAnEmbeddingTooLongForRmsNorm {
+    fn normalize(
+        &self,
+        _embedding: Vec<f32>,
+        _normalization_method: &EmbeddingNormalizationMethod,
+    ) -> Result<Vec<f32>, EmbeddingError> {
+        normalize_embedding(
+            vec![0.0; DIMENSIONS_TOO_MANY_FOR_RMS_NORM],
+            &EmbeddingNormalizationMethod::RmsNorm { epsilon: 0.0 },
+        )
+    }
+}
diff --git a/paddler_tests/src/numbered_embedding_documents.rs b/paddler_tests/src/numbered_embedding_documents.rs
new file mode 100644
index 000000000..aaa5f38ce
--- /dev/null
+++ b/paddler_tests/src/numbered_embedding_documents.rs
@@ -0,0 +1,11 @@
+use paddler_messaging::embedding_input_document::EmbeddingInputDocument;
+
+#[must_use]
+pub fn numbered_embedding_documents(document_count: usize) -> Vec<EmbeddingInputDocument> {
+    (0..document_count)
+        .map(|document_index| EmbeddingInputDocument {
+            content: format!("Document number {document_index}."),
+            id: format!("doc-{document_index}"),
+        })
+        .collect()
+}
diff --git a/paddler_tests/src/oversized_text.rs b/paddler_tests/src/oversized_text.rs
new file mode 100644
index 000000000..fa79e94b1
--- /dev/null
+++ b/paddler_tests/src/oversized_text.rs
@@ -0,0 +1,4 @@
+#[must_use]
+pub fn oversized_text() -> String {
+    "The quick brown fox jumps over the lazy dog. ".repeat(40)
+}
diff --git a/paddler_tests/src/parallelism_that_is_unknown.rs b/paddler_tests/src/parallelism_that_is_unknown.rs
new file mode 100644
index 000000000..83a7ed6c9
--- /dev/null
+++ b/paddler_tests/src/parallelism_that_is_unknown.rs
@@ -0,0 +1,14 @@
+use std::io;
+use std::num::NonZeroUsize;
+
+use paddler_agent_runtime::reports_available_parallelism::ReportsAvailableParallelism;
+
+pub struct ParallelismThatIsUnknown;
+
+impl ReportsAvailableParallelism for ParallelismThatIsUnknown {
+    fn available_parallelism(&self) -> io::Result<NonZeroUsize> {
+        Err(io::Error::other(
+            "the machine hides how many threads it runs in parallel",
+        ))
+    }
+}
diff --git a/paddler_tests/src/point_qwen3_special_token_at_unused_padding.rs b/paddler_tests/src/point_qwen3_special_token_at_unused_padding.rs
new file mode 100644
index 000000000..d388d15e5
--- /dev/null
+++ b/paddler_tests/src/point_qwen3_special_token_at_unused_padding.rs
@@ -0,0 +1,11 @@
+use std::ffi::CStr;
+
+use llama_cpp_bindings::gguf_context::GgufContext;
+
+use crate::qwen3_0_6b_unused_padding_token::QWEN3_0_6B_UNUSED_PADDING_TOKEN;
+
+pub fn point_qwen3_special_token_at_unused_padding(
+    special_token_key: &CStr,
+) -> impl FnOnce(&mut GgufContext) {
+    move |gguf_context| gguf_context.set_val_u32(special_token_key, QWEN3_0_6B_UNUSED_PADDING_TOKEN)
+}
diff --git a/paddler_tests/src/post_openai_chat_completion.rs b/paddler_tests/src/post_openai_chat_completion.rs
new file mode 100644
index 000000000..2a47ae0ee
--- /dev/null
+++ b/paddler_tests/src/post_openai_chat_completion.rs
@@ -0,0 +1,14 @@
+use std::future::Future;
+
+use reqwest::Client;
+use reqwest::Error;
+use reqwest::Response;
+use reqwest::Url;
+use serde_json::Value;
+
+pub fn post_openai_chat_completion(
+    chat_completions_url: Url,
+    body: &Value,
+) -> impl Future<Output = Result<Response, Error>> + Send + use<> {
+    Client::new().post(chat_completions_url).json(body).send()
+}
diff --git a/paddler_tests/src/prepared_decision.rs b/paddler_tests/src/prepared_decision.rs
new file mode 100644
index 000000000..e07d3a507
--- /dev/null
+++ b/paddler_tests/src/prepared_decision.rs
@@ -0,0 +1,41 @@
+use std::sync::Arc;
+use std::time::Instant;
+
+use tokio::sync::mpsc;
+
+use paddler_agent_decision::decision_token_layout::DecisionTokenLayout;
+use paddler_agent_decision::prepared_decision_request::PreparedDecisionRequest;
+use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
+use paddler_agent_status::slot_guard::SlotGuard;
+use paddler_messaging::decision_result::DecisionResult;
+use paddler_messaging::decision_slots_for_leading_questions::decision_slots_for_leading_questions;
+
+pub struct PreparedDecision {
+    pub decision_result_rx: mpsc::UnboundedReceiver<DecisionResult>,
+    pub decision_stop_tx: mpsc::UnboundedSender<()>,
+    pub prepared_decision_request: PreparedDecisionRequest,
+}
+
+impl PreparedDecision {
+    #[must_use]
+    pub fn new(
+        slot_aggregated_status: Arc<SlotAggregatedStatus>,
+        layout: DecisionTokenLayout,
+    ) -> Self {
+        let (decision_result_tx, decision_result_rx) = mpsc::unbounded_channel();
+        let (decision_stop_tx, decision_stop_rx) = mpsc::unbounded_channel();
+        let reserved_slots = decision_slots_for_leading_questions(layout.remaining_questions.len());
+
+        Self {
+            decision_result_rx,
+            decision_stop_tx,
+            prepared_decision_request: PreparedDecisionRequest {
+                decision_result_tx,
+                decision_stop_rx,
+                layout,
+                slot_guard: SlotGuard::new(slot_aggregated_status, reserved_slots),
+                started_at: Instant::now(),
+            },
+        }
+    }
+}
diff --git a/paddler_tests/src/question_answered_discriminant.rs b/paddler_tests/src/question_answered_discriminant.rs
new file mode 100644
index 000000000..5b5fcaa84
--- /dev/null
+++ b/paddler_tests/src/question_answered_discriminant.rs
@@ -0,0 +1,13 @@
+use std::mem::Discriminant;
+use std::mem::discriminant;
+
+use paddler_messaging::decision_answer::DecisionAnswer;
+use paddler_messaging::decision_result::DecisionResult;
+
+#[must_use]
+pub fn question_answered_discriminant() -> Discriminant<DecisionResult> {
+    discriminant(&DecisionResult::QuestionAnswered(DecisionAnswer {
+        id: String::new(),
+        probabilities: Vec::new(),
+    }))
+}
diff --git a/paddler_tests/src/questions_outlasting_a_disconnect.rs b/paddler_tests/src/questions_outlasting_a_disconnect.rs
new file mode 100644
index 000000000..2f286bb54
--- /dev/null
+++ b/paddler_tests/src/questions_outlasting_a_disconnect.rs
@@ -0,0 +1 @@
+pub const QUESTIONS_OUTLASTING_A_DISCONNECT: usize = 96;
diff --git a/paddler_tests/src/qwen3_0_6b_newline_token.rs b/paddler_tests/src/qwen3_0_6b_newline_token.rs
new file mode 100644
index 000000000..f5d8258fb
--- /dev/null
+++ b/paddler_tests/src/qwen3_0_6b_newline_token.rs
@@ -0,0 +1 @@
+pub const QWEN3_0_6B_NEWLINE_TOKEN: usize = 198;
diff --git a/paddler_tests/src/qwen3_0_6b_unused_padding_token.rs b/paddler_tests/src/qwen3_0_6b_unused_padding_token.rs
new file mode 100644
index 000000000..5affa4a13
--- /dev/null
+++ b/paddler_tests/src/qwen3_0_6b_unused_padding_token.rs
@@ -0,0 +1 @@
+pub const QWEN3_0_6B_UNUSED_PADDING_TOKEN: u32 = 151_669;
diff --git a/paddler_tests/src/qwen3_desired_state_with_embeddings.rs b/paddler_tests/src/qwen3_desired_state_with_embeddings.rs
deleted file mode 100644
index 26f3bc991..000000000
--- a/paddler_tests/src/qwen3_desired_state_with_embeddings.rs
+++ /dev/null
@@ -1,17 +0,0 @@
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
-use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-
-use crate::qwen3_desired_state::qwen3_desired_state;
-
-#[must_use]
-pub fn qwen3_desired_state_with_embeddings(enable_embeddings: bool) -> BalancerDesiredState {
-    let desired_state = qwen3_desired_state();
-
-    BalancerDesiredState {
-        inference_parameters: InferenceParameters {
-            enable_embeddings,
-            ..desired_state.inference_parameters
-        },
-        ..desired_state
-    }
-}
diff --git a/paddler_tests/src/raw_prompt_generation.rs b/paddler_tests/src/raw_prompt_generation.rs
new file mode 100644
index 000000000..0f2c622de
--- /dev/null
+++ b/paddler_tests/src/raw_prompt_generation.rs
@@ -0,0 +1,40 @@
+use std::num::NonZeroU16;
+use std::num::NonZeroU32;
+use std::sync::Arc;
+
+use tokio::sync::mpsc;
+
+use paddler_agent_runtime::agent_request::AgentRequest;
+use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
+use paddler_agent_status::slot_guard::SlotGuard;
+use paddler_messaging::generated_token_result::GeneratedTokenResult;
+use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
+
+pub struct RawPromptGeneration {
+    pub agent_request: AgentRequest<ContinueFromRawPromptParams, GeneratedTokenResult>,
+    pub generate_tokens_stop_tx: mpsc::UnboundedSender<()>,
+    pub generated_tokens_rx: mpsc::UnboundedReceiver<GeneratedTokenResult>,
+}
+
+impl RawPromptGeneration {
+    #[must_use]
+    pub fn new(raw_prompt: &str, slot_aggregated_status: Arc<SlotAggregatedStatus>) -> Self {
+        let (generated_tokens_tx, generated_tokens_rx) = mpsc::unbounded_channel();
+        let (generate_tokens_stop_tx, generate_tokens_stop_rx) = mpsc::unbounded_channel();
+
+        Self {
+            agent_request: AgentRequest {
+                params: ContinueFromRawPromptParams {
+                    grammar: None,
+                    max_tokens: NonZeroU32::MIN,
+                    raw_prompt: raw_prompt.to_owned(),
+                },
+                response_tx: generated_tokens_tx,
+                slot_guard: SlotGuard::new(slot_aggregated_status, NonZeroU16::MIN),
+                stop_rx: generate_tokens_stop_rx,
+            },
+            generate_tokens_stop_tx,
+            generated_tokens_rx,
+        }
+    }
+}
diff --git a/paddler_tests/src/release_half_closed_openai_request.rs b/paddler_tests/src/release_half_closed_openai_request.rs
deleted file mode 100644
index 5661d98e9..000000000
--- a/paddler_tests/src/release_half_closed_openai_request.rs
+++ /dev/null
@@ -1,37 +0,0 @@
-use anyhow::Context as _;
-use anyhow::Result;
-use serde_json::Value;
-
-use paddler_test_cluster_harness::cluster_params::ClusterParams;
-use paddler_test_cluster_harness::half_closed_client::HalfClosedClient;
-
-use crate::start_cluster::start_cluster;
-
-pub async fn release_half_closed_openai_request(path: &str, body: &Value) -> Result<()> {
-    let mut cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
-    })
-    .await?;
-    let mut client = HalfClosedClient::post_json_then_half_close(
-        cluster.balancer.compat_openai_addr()?,
-        path,
-        body,
-    )
-    .await?;
-
-    cluster
-        .wait_for_buffered_request_count(1)
-        .await
-        .context("the request must be buffered while no agent is available")?;
-    client.half_close().await?;
-    cluster
-        .wait_for_buffered_request_count(0)
-        .await
-        .context("the balancer must release the buffered request once the client half-closes")?;
-
-    drop(client);
-
-    cluster.shutdown().await
-}
diff --git a/paddler_tests/src/responses_message_text.rs b/paddler_tests/src/responses_message_text.rs
new file mode 100644
index 000000000..f4033d34f
--- /dev/null
+++ b/paddler_tests/src/responses_message_text.rs
@@ -0,0 +1,15 @@
+use serde_json::Value;
+
+#[must_use]
+pub fn responses_message_text(response: &Value) -> Option<&str> {
+    response
+        .get("output")
+        .and_then(Value::as_array)
+        .and_then(|output_items| {
+            output_items.iter().find(|output_item| {
+                output_item.get("type").and_then(Value::as_str) == Some("message")
+            })
+        })
+        .and_then(|message| message.pointer("/content/0/text"))
+        .and_then(Value::as_str)
+}
diff --git a/paddler_tests/src/sample_decision.rs b/paddler_tests/src/sample_decision.rs
new file mode 100644
index 000000000..b8abfdd8b
--- /dev/null
+++ b/paddler_tests/src/sample_decision.rs
@@ -0,0 +1,32 @@
+use paddler_messaging::decision_question::DecisionQuestion;
+use paddler_messaging::request_params::decide_params::raw_decide_params::RawDecideParams;
+
+fn question(id: &str, instructions: &str, options: &[&str]) -> DecisionQuestion {
+    DecisionQuestion {
+        id: id.to_owned(),
+        instructions: instructions.to_owned(),
+        options: options.iter().map(|option| (*option).to_owned()).collect(),
+    }
+}
+
+#[must_use]
+pub fn sample_decision() -> RawDecideParams {
+    RawDecideParams {
+        questions: vec![
+            question("paid", "Was the invoice paid?", &["no", "yes"]),
+            question(
+                "sentiment",
+                "How does the customer feel?",
+                &["positive", "neutral", "negative: the customer complains"],
+            ),
+            question(
+                "urgency",
+                "How urgent is the follow-up?",
+                &["none", "low", "medium", "high"],
+            ),
+        ],
+        state: "Ada paid the invoice for order 1042 two days late and wrote that the delivery was \
+                slow, but the support team was friendly."
+            .to_owned(),
+    }
+}
diff --git a/paddler_tests/src/short_buffered_request_timeout.rs b/paddler_tests/src/short_buffered_request_timeout.rs
new file mode 100644
index 000000000..8796c507e
--- /dev/null
+++ b/paddler_tests/src/short_buffered_request_timeout.rs
@@ -0,0 +1,3 @@
+use std::time::Duration;
+
+pub const SHORT_BUFFERED_REQUEST_TIMEOUT: Duration = Duration::from_millis(50);
diff --git a/paddler_tests/src/single_image_question.rs b/paddler_tests/src/single_image_question.rs
new file mode 100644
index 000000000..c7154a475
--- /dev/null
+++ b/paddler_tests/src/single_image_question.rs
@@ -0,0 +1,50 @@
+use std::num::NonZeroU32;
+
+use paddler_messaging::conversation_history::ConversationHistory;
+use paddler_messaging::conversation_message::ConversationMessage;
+use paddler_messaging::conversation_message_content::ConversationMessageContent;
+use paddler_messaging::conversation_message_content_part::ConversationMessageContentPart;
+use paddler_messaging::image_url::ImageUrl;
+use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
+use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters_schema::validated_parameters_schema::ValidatedParametersSchema;
+
+pub struct SingleImageQuestion {
+    pub image_data_uri: String,
+    pub max_tokens: NonZeroU32,
+    pub question: &'static str,
+}
+
+impl SingleImageQuestion {
+    #[must_use]
+    pub fn into_conversation_params(
+        self,
+    ) -> ContinueFromConversationHistoryParams<ValidatedParametersSchema> {
+        let Self {
+            image_data_uri,
+            max_tokens,
+            question,
+        } = self;
+
+        ContinueFromConversationHistoryParams {
+            add_generation_prompt: true,
+            conversation_history: ConversationHistory::new(vec![ConversationMessage::Message {
+                content: ConversationMessageContent::Parts(vec![
+                    ConversationMessageContentPart::ImageUrl {
+                        image_url: ImageUrl {
+                            url: image_data_uri,
+                        },
+                    },
+                    ConversationMessageContentPart::Text {
+                        text: question.to_owned(),
+                    },
+                ]),
+                role: "user".to_owned(),
+            }]),
+            enable_thinking: false,
+            grammar: None,
+            max_tokens,
+            parse_tool_calls: false,
+            tools: vec![],
+        }
+    }
+}
diff --git a/paddler_tests/src/start_cluster.rs b/paddler_tests/src/start_cluster.rs
index b85d6db77..f181d6454 100644
--- a/paddler_tests/src/start_cluster.rs
+++ b/paddler_tests/src/start_cluster.rs
@@ -1,90 +1,69 @@
-use std::str::FromStr as _;
-
 use anyhow::Context as _;
 use anyhow::Result;
 use log::LevelFilter;
 use log::set_max_level;
+use tempfile::TempDir;
 use tokio_util::sync::CancellationToken;
 
-use paddler_balancer::compatibility::openai_service::configuration::Configuration as OpenAIServiceConfiguration;
-use paddler_balancer::inference_service::configuration::Configuration as InferenceServiceConfiguration;
-use paddler_balancer::management_service::configuration::Configuration as ManagementServiceConfiguration;
-use paddler_balancer::resolved_socket_addr::ResolvedSocketAddr;
-use paddler_bootstrap::balancer_bootstrap_config::BalancerBootstrapConfig;
 use paddler_bootstrap::balancer_runner::BalancerRunner;
-use paddler_bootstrap::balancer_runner_params::BalancerRunnerParams;
-use paddler_state_database::state_database_type::StateDatabaseType;
+use paddler_cache_dir::cache_dir::CacheDir;
+use paddler_download_manager::download_manager::DownloadManager;
+use paddler_model_source::desired_model_resolver::DesiredModelResolver;
+use paddler_model_source::huggingface_cache_location::HuggingFaceCacheLocation;
+use paddler_model_source::huggingface_default_endpoint::HUGGINGFACE_DEFAULT_ENDPOINT;
+use paddler_model_source::huggingface_hub::HuggingFaceHub;
 use paddler_test_cluster_harness::cluster::Cluster;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
-use paddler_test_cluster_harness::ephemeral_loopback_addr::EPHEMERAL_LOOPBACK_ADDR;
 use paddler_test_cluster_harness::running_balancer::RunningBalancer;
-use paddler_test_cluster_harness::unexpiring_shutdown_options::unexpiring_shutdown_options;
 
 use crate::in_process_agent_spawner::InProcessAgentSpawner;
 use crate::in_process_balancer::InProcessBalancer;
+use crate::in_process_balancer_runner_params::in_process_balancer_runner_params;
 
-pub async fn start_cluster(
-    ClusterParams {
-        agents,
-        buffered_request_timeout,
-        desired_state,
-        inference_cors_allowed_hosts,
-        inference_item_timeout,
-        management_cors_allowed_hosts,
-        max_buffered_requests,
-        state_database_url,
-        wait_for_slots_ready,
-    }: ClusterParams,
-) -> Result<Cluster> {
+pub async fn start_cluster(cluster_params: ClusterParams) -> Result<Cluster> {
     set_max_level(LevelFilter::Trace);
 
-    let state_database_type = StateDatabaseType::from_str(&state_database_url)
-        .context("failed to parse state_database_url")?;
-
-    let balancer_runner = BalancerRunner::start(BalancerRunnerParams {
-        bootstrap_config: BalancerBootstrapConfig {
-            buffered_request_timeout,
-            inference_service_configuration: InferenceServiceConfiguration {
-                addr: ResolvedSocketAddr::from(EPHEMERAL_LOOPBACK_ADDR),
-                cors_allowed_hosts: inference_cors_allowed_hosts,
-                inference_item_timeout,
-            },
-            management_service_configuration: ManagementServiceConfiguration {
-                addr: ResolvedSocketAddr::from(EPHEMERAL_LOOPBACK_ADDR),
-                cors_allowed_hosts: management_cors_allowed_hosts,
-            },
-            max_buffered_requests,
-            openai_service_configuration: Some(OpenAIServiceConfiguration {
-                addr: ResolvedSocketAddr::from(EPHEMERAL_LOOPBACK_ADDR),
-            }),
-            state_database_type,
-            statsd_prefix: "paddler_tests_".to_owned(),
-            statsd_service_configuration: None,
-            #[cfg(feature = "web_admin_panel")]
-            web_admin_panel_service_configuration: None,
+    let running_balancer =
+        BalancerRunner::start(in_process_balancer_runner_params(&cluster_params)?)
+            .await
+            .context("failed to start in-process BalancerRunner")
+            .and_then(|balancer_runner| {
+                RunningBalancer::new(
+                    balancer_runner.addresses,
+                    Box::new(InProcessBalancer::new(balancer_runner)),
+                )
+                .context("failed to address the in-process balancer")
+            })?;
+    let management_address = running_balancer.addresses.management.to_string();
+    let downloaded_models = TempDir::new()?;
+    let desired_model_resolver = DesiredModelResolver {
+        cache_dir: CacheDir {
+            explicit: Some(downloaded_models.path().to_path_buf()),
+            home: None,
+            xdg: None,
         },
-        cancellation_token: CancellationToken::new(),
-        shutdown_options: unexpiring_shutdown_options(),
-    })
-    .await
-    .context("failed to start in-process BalancerRunner")?;
-
-    let management_address = balancer_runner.addresses.management.to_string();
-    let running_balancer = RunningBalancer::new(
-        balancer_runner.addresses,
-        Box::new(InProcessBalancer::new(balancer_runner)),
-    );
+        download_manager: DownloadManager::new()?,
+        huggingface_client: HuggingFaceHub {
+            cache_location: HuggingFaceCacheLocation::HubDefault,
+            endpoint: HUGGINGFACE_DEFAULT_ENDPOINT.to_owned(),
+        }
+        .client()?,
+    };
 
     let mut cluster = Cluster::connect(
         CancellationToken::new(),
         running_balancer,
-        Box::new(InProcessAgentSpawner::new(management_address)),
-        desired_state.as_ref(),
+        Box::new(InProcessAgentSpawner {
+            desired_model_resolver,
+            downloaded_models,
+            management_address,
+        }),
+        &cluster_params.desired_state,
     )
     .await?;
 
     cluster
-        .register_agents(&agents, wait_for_slots_ready)
+        .register_agents(&cluster_params.agents, cluster_params.wait_for_slots_ready)
         .await?;
 
     Ok(cluster)
diff --git a/paddler_tests/src/start_cluster_with_deepseek_r1_distill_llama_8b.rs b/paddler_tests/src/start_cluster_with_deepseek_r1_distill_llama_8b.rs
index 4e56acea4..a45904eeb 100644
--- a/paddler_tests/src/start_cluster_with_deepseek_r1_distill_llama_8b.rs
+++ b/paddler_tests/src/start_cluster_with_deepseek_r1_distill_llama_8b.rs
@@ -2,6 +2,7 @@ use anyhow::Result;
 
 use paddler_test_cluster_harness::agent_config::AgentConfig;
 use paddler_test_cluster_harness::cluster::Cluster;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
 use paddler_test_cluster_harness::model_card::deepseek_r1_distill_llama_8b::deepseek_r1_distill_llama_8b;
 
@@ -12,7 +13,9 @@ pub async fn start_cluster_with_deepseek_r1_distill_llama_8b(
 ) -> Result<Cluster> {
     start_cluster(ClusterParams {
         agents,
-        desired_state: Some(deepseek_r1_distill_llama_8b().into_desired_state()),
+        desired_state: ClusterDesiredState::Apply(Box::new(
+            deepseek_r1_distill_llama_8b().into_desired_state(),
+        )),
         wait_for_slots_ready: true,
         ..ClusterParams::default()
     })
diff --git a/paddler_tests/src/start_cluster_with_gemma_3.rs b/paddler_tests/src/start_cluster_with_gemma_3.rs
index 2db775c45..1032bd6c3 100644
--- a/paddler_tests/src/start_cluster_with_gemma_3.rs
+++ b/paddler_tests/src/start_cluster_with_gemma_3.rs
@@ -2,6 +2,7 @@ use anyhow::Result;
 
 use paddler_test_cluster_harness::agent_config::AgentConfig;
 use paddler_test_cluster_harness::cluster::Cluster;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
 use paddler_test_cluster_harness::model_card::gemma_3_4b_it::gemma_3_4b_it;
 
@@ -10,7 +11,7 @@ use crate::start_cluster::start_cluster;
 pub async fn start_cluster_with_gemma_3(agents: Vec<AgentConfig>) -> Result<Cluster> {
     start_cluster(ClusterParams {
         agents,
-        desired_state: Some(gemma_3_4b_it().into_desired_state()),
+        desired_state: ClusterDesiredState::Apply(Box::new(gemma_3_4b_it().into_desired_state())),
         wait_for_slots_ready: true,
         ..ClusterParams::default()
     })
diff --git a/paddler_tests/src/start_cluster_with_gemma_3_and_mmproj_and_n_batch.rs b/paddler_tests/src/start_cluster_with_gemma_3_and_mmproj_and_n_batch.rs
index 4fd813d74..05c92b231 100644
--- a/paddler_tests/src/start_cluster_with_gemma_3_and_mmproj_and_n_batch.rs
+++ b/paddler_tests/src/start_cluster_with_gemma_3_and_mmproj_and_n_batch.rs
@@ -1,10 +1,11 @@
 use anyhow::Result;
 
 use paddler_inference_parameters::batch_size::BatchSize;
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
+use paddler_inference_parameters::model_runtime_parameters::ModelRuntimeParameters;
 use paddler_messaging::balancer_desired_state::BalancerDesiredState;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
 use paddler_test_cluster_harness::cluster::Cluster;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
 use paddler_test_cluster_harness::model_card::gemma_3_4b_it::gemma_3_4b_it;
 use paddler_test_cluster_harness::model_card::gemma_3_4b_it_mmproj::gemma_3_4b_it_mmproj;
@@ -13,20 +14,20 @@ use crate::start_cluster::start_cluster;
 
 pub async fn start_cluster_with_gemma_3_and_mmproj_and_n_batch(
     agents: Vec<AgentConfig>,
-    n_batch: u32,
+    n_batch: BatchSize,
 ) -> Result<Cluster> {
     let base_desired_state =
         gemma_3_4b_it().into_desired_state_with_multimodal_projection(gemma_3_4b_it_mmproj());
 
     start_cluster(ClusterParams {
         agents,
-        desired_state: Some(BalancerDesiredState {
-            inference_parameters: InferenceParameters {
-                n_batch: BatchSize::try_from(n_batch)?,
-                ..base_desired_state.inference_parameters
+        desired_state: ClusterDesiredState::Apply(Box::new(BalancerDesiredState {
+            model_runtime_parameters: ModelRuntimeParameters {
+                n_batch,
+                ..base_desired_state.model_runtime_parameters
             },
             ..base_desired_state
-        }),
+        })),
         wait_for_slots_ready: true,
         ..ClusterParams::default()
     })
diff --git a/paddler_tests/src/start_cluster_with_gemma_4.rs b/paddler_tests/src/start_cluster_with_gemma_4.rs
index e46e1b6cf..80fbef930 100644
--- a/paddler_tests/src/start_cluster_with_gemma_4.rs
+++ b/paddler_tests/src/start_cluster_with_gemma_4.rs
@@ -2,6 +2,7 @@ use anyhow::Result;
 
 use paddler_test_cluster_harness::agent_config::AgentConfig;
 use paddler_test_cluster_harness::cluster::Cluster;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
 use paddler_test_cluster_harness::model_card::gemma_4_e2b_it::gemma_4_e2b_it;
 
@@ -10,7 +11,7 @@ use crate::start_cluster::start_cluster;
 pub async fn start_cluster_with_gemma_4(agents: Vec<AgentConfig>) -> Result<Cluster> {
     start_cluster(ClusterParams {
         agents,
-        desired_state: Some(gemma_4_e2b_it().into_desired_state()),
+        desired_state: ClusterDesiredState::Apply(Box::new(gemma_4_e2b_it().into_desired_state())),
         wait_for_slots_ready: true,
         ..ClusterParams::default()
     })
diff --git a/paddler_tests/src/start_cluster_with_gemma_4_and_mmproj.rs b/paddler_tests/src/start_cluster_with_gemma_4_and_mmproj.rs
index b1075806c..c91fdbeee 100644
--- a/paddler_tests/src/start_cluster_with_gemma_4_and_mmproj.rs
+++ b/paddler_tests/src/start_cluster_with_gemma_4_and_mmproj.rs
@@ -2,6 +2,7 @@ use anyhow::Result;
 
 use paddler_test_cluster_harness::agent_config::AgentConfig;
 use paddler_test_cluster_harness::cluster::Cluster;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
 use paddler_test_cluster_harness::model_card::gemma_4_e2b_it::gemma_4_e2b_it;
 use paddler_test_cluster_harness::model_card::gemma_4_e2b_it_mmproj::gemma_4_e2b_it_mmproj;
@@ -11,9 +12,9 @@ use crate::start_cluster::start_cluster;
 pub async fn start_cluster_with_gemma_4_and_mmproj(agents: Vec<AgentConfig>) -> Result<Cluster> {
     start_cluster(ClusterParams {
         agents,
-        desired_state: Some(
+        desired_state: ClusterDesiredState::Apply(Box::new(
             gemma_4_e2b_it().into_desired_state_with_multimodal_projection(gemma_4_e2b_it_mmproj()),
-        ),
+        )),
         wait_for_slots_ready: true,
         ..ClusterParams::default()
     })
diff --git a/paddler_tests/src/start_cluster_with_ministral_3.rs b/paddler_tests/src/start_cluster_with_ministral_3.rs
index a7db5b496..7c8b02c0a 100644
--- a/paddler_tests/src/start_cluster_with_ministral_3.rs
+++ b/paddler_tests/src/start_cluster_with_ministral_3.rs
@@ -1,6 +1,7 @@
 use anyhow::Result;
 
 use paddler_test_cluster_harness::cluster::Cluster;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
 use paddler_test_cluster_harness::model_card::ministral_3_3b_reasoning::ministral_3_3b_reasoning;
 
@@ -12,7 +13,9 @@ pub async fn start_cluster_with_ministral_3(
 ) -> Result<Cluster> {
     start_cluster(ClusterParams {
         agents,
-        desired_state: Some(ministral_3_3b_reasoning().into_desired_state()),
+        desired_state: ClusterDesiredState::Apply(Box::new(
+            ministral_3_3b_reasoning().into_desired_state(),
+        )),
         wait_for_slots_ready: true,
         ..ClusterParams::default()
     })
diff --git a/paddler_tests/src/start_cluster_with_ministral_3_and_mmproj.rs b/paddler_tests/src/start_cluster_with_ministral_3_and_mmproj.rs
index e465ee16a..00c7aab1e 100644
--- a/paddler_tests/src/start_cluster_with_ministral_3_and_mmproj.rs
+++ b/paddler_tests/src/start_cluster_with_ministral_3_and_mmproj.rs
@@ -2,6 +2,7 @@ use anyhow::Result;
 
 use paddler_test_cluster_harness::agent_config::AgentConfig;
 use paddler_test_cluster_harness::cluster::Cluster;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
 use paddler_test_cluster_harness::model_card::ministral_3_8b_reasoning::ministral_3_8b_reasoning;
 use paddler_test_cluster_harness::model_card::ministral_3_8b_reasoning_mmproj::ministral_3_8b_reasoning_mmproj;
@@ -13,10 +14,10 @@ pub async fn start_cluster_with_ministral_3_and_mmproj(
 ) -> Result<Cluster> {
     start_cluster(ClusterParams {
         agents,
-        desired_state: Some(
+        desired_state: ClusterDesiredState::Apply(Box::new(
             ministral_3_8b_reasoning()
                 .into_desired_state_with_multimodal_projection(ministral_3_8b_reasoning_mmproj()),
-        ),
+        )),
         wait_for_slots_ready: true,
         ..ClusterParams::default()
     })
diff --git a/paddler_tests/src/start_cluster_with_qwen2_5_vl.rs b/paddler_tests/src/start_cluster_with_qwen2_5_vl.rs
index ae0df45fb..8124878c6 100644
--- a/paddler_tests/src/start_cluster_with_qwen2_5_vl.rs
+++ b/paddler_tests/src/start_cluster_with_qwen2_5_vl.rs
@@ -2,6 +2,7 @@ use anyhow::Result;
 
 use paddler_test_cluster_harness::agent_config::AgentConfig;
 use paddler_test_cluster_harness::cluster::Cluster;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
 use paddler_test_cluster_harness::model_card::qwen2_5_vl_3b::qwen2_5_vl_3b;
 use paddler_test_cluster_harness::model_card::qwen2_5_vl_3b_mmproj::qwen2_5_vl_3b_mmproj;
@@ -11,9 +12,9 @@ use crate::start_cluster::start_cluster;
 pub async fn start_cluster_with_qwen2_5_vl(agents: Vec<AgentConfig>) -> Result<Cluster> {
     start_cluster(ClusterParams {
         agents,
-        desired_state: Some(
+        desired_state: ClusterDesiredState::Apply(Box::new(
             qwen2_5_vl_3b().into_desired_state_with_multimodal_projection(qwen2_5_vl_3b_mmproj()),
-        ),
+        )),
         wait_for_slots_ready: true,
         ..ClusterParams::default()
     })
diff --git a/paddler_tests/src/start_cluster_with_qwen3.rs b/paddler_tests/src/start_cluster_with_qwen3.rs
index 1c0e3357e..c0fa5bde1 100644
--- a/paddler_tests/src/start_cluster_with_qwen3.rs
+++ b/paddler_tests/src/start_cluster_with_qwen3.rs
@@ -2,6 +2,7 @@ use anyhow::Result;
 
 use paddler_test_cluster_harness::agent_config::AgentConfig;
 use paddler_test_cluster_harness::cluster::Cluster;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
 
 use crate::qwen3_desired_state::qwen3_desired_state;
@@ -10,7 +11,7 @@ use crate::start_cluster::start_cluster;
 pub async fn start_cluster_with_qwen3(agents: Vec<AgentConfig>) -> Result<Cluster> {
     start_cluster(ClusterParams {
         agents,
-        desired_state: Some(qwen3_desired_state()),
+        desired_state: ClusterDesiredState::Apply(Box::new(qwen3_desired_state())),
         wait_for_slots_ready: true,
         ..ClusterParams::default()
     })
diff --git a/paddler_tests/src/start_cluster_with_qwen3_5.rs b/paddler_tests/src/start_cluster_with_qwen3_5.rs
index 9cda51131..7357be655 100644
--- a/paddler_tests/src/start_cluster_with_qwen3_5.rs
+++ b/paddler_tests/src/start_cluster_with_qwen3_5.rs
@@ -1,6 +1,6 @@
 use anyhow::Result;
 
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
+use paddler_inference_parameters::model_runtime_parameters::ModelRuntimeParameters;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
 use paddler_test_cluster_harness::cluster::Cluster;
 
@@ -13,7 +13,7 @@ pub async fn start_cluster_with_qwen3_5(
     start_cluster_with_qwen3_5_and_context_size(
         agents,
         with_mmproj,
-        InferenceParameters::default().context_size.get(),
+        ModelRuntimeParameters::default().context_size,
     )
     .await
 }
diff --git a/paddler_tests/src/start_cluster_with_qwen3_5_and_context_size.rs b/paddler_tests/src/start_cluster_with_qwen3_5_and_context_size.rs
index 5f815bc2e..897b24e62 100644
--- a/paddler_tests/src/start_cluster_with_qwen3_5_and_context_size.rs
+++ b/paddler_tests/src/start_cluster_with_qwen3_5_and_context_size.rs
@@ -2,42 +2,39 @@ use std::num::NonZeroU32;
 
 use anyhow::Result;
 
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
-use paddler_messaging::agent_desired_model::AgentDesiredModel;
+use paddler_inference_parameters::batch_size::BatchSize;
+use paddler_inference_parameters::model_runtime_parameters::ModelRuntimeParameters;
 use paddler_messaging::balancer_desired_state::BalancerDesiredState;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
 use paddler_test_cluster_harness::cluster::Cluster;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
 use paddler_test_cluster_harness::model_card::qwen3_5_0_8b::qwen3_5_0_8b;
 use paddler_test_cluster_harness::model_card::qwen3_5_0_8b_mmproj::qwen3_5_0_8b_mmproj;
 
-use crate::batch_size_within_context::batch_size_within_context;
 use crate::start_cluster::start_cluster;
 
 pub async fn start_cluster_with_qwen3_5_and_context_size(
     agents: Vec<AgentConfig>,
     with_mmproj: bool,
-    context_size: u32,
+    context_size: NonZeroU32,
 ) -> Result<Cluster> {
-    let context_size = NonZeroU32::try_from(context_size)?;
-    let base_desired_state = qwen3_5_0_8b().into_desired_state();
-    let multimodal_projection = if with_mmproj {
-        qwen3_5_0_8b_mmproj().into_agent_desired_model()
+    let base_desired_state = if with_mmproj {
+        qwen3_5_0_8b().into_desired_state_with_multimodal_projection(qwen3_5_0_8b_mmproj())
     } else {
-        AgentDesiredModel::None
+        qwen3_5_0_8b().into_desired_state()
     };
 
     start_cluster(ClusterParams {
         agents,
-        desired_state: Some(BalancerDesiredState {
-            inference_parameters: InferenceParameters {
+        desired_state: ClusterDesiredState::Apply(Box::new(BalancerDesiredState {
+            model_runtime_parameters: ModelRuntimeParameters {
                 context_size,
-                n_batch: batch_size_within_context(context_size)?,
-                ..base_desired_state.inference_parameters
+                n_batch: BatchSize::default().at_most(context_size),
+                ..base_desired_state.model_runtime_parameters
             },
-            multimodal_projection,
             ..base_desired_state
-        }),
+        })),
         wait_for_slots_ready: true,
         ..ClusterParams::default()
     })
diff --git a/paddler_tests/src/start_cluster_with_qwen3_and_context_size.rs b/paddler_tests/src/start_cluster_with_qwen3_and_context_size.rs
index 6f625efcd..95b323dc8 100644
--- a/paddler_tests/src/start_cluster_with_qwen3_and_context_size.rs
+++ b/paddler_tests/src/start_cluster_with_qwen3_and_context_size.rs
@@ -2,33 +2,33 @@ use std::num::NonZeroU32;
 
 use anyhow::Result;
 
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
+use paddler_inference_parameters::batch_size::BatchSize;
+use paddler_inference_parameters::model_runtime_parameters::ModelRuntimeParameters;
 use paddler_messaging::balancer_desired_state::BalancerDesiredState;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
 use paddler_test_cluster_harness::cluster::Cluster;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
 
-use crate::batch_size_within_context::batch_size_within_context;
 use crate::qwen3_desired_state::qwen3_desired_state;
 use crate::start_cluster::start_cluster;
 
 pub async fn start_cluster_with_qwen3_and_context_size(
     agents: Vec<AgentConfig>,
-    context_size: u32,
+    context_size: NonZeroU32,
 ) -> Result<Cluster> {
-    let context_size = NonZeroU32::try_from(context_size)?;
     let desired_state = qwen3_desired_state();
 
     start_cluster(ClusterParams {
         agents,
-        desired_state: Some(BalancerDesiredState {
-            inference_parameters: InferenceParameters {
+        desired_state: ClusterDesiredState::Apply(Box::new(BalancerDesiredState {
+            model_runtime_parameters: ModelRuntimeParameters {
                 context_size,
-                n_batch: batch_size_within_context(context_size)?,
-                ..desired_state.inference_parameters
+                n_batch: BatchSize::default().at_most(context_size),
+                ..desired_state.model_runtime_parameters
             },
             ..desired_state
-        }),
+        })),
         wait_for_slots_ready: true,
         ..ClusterParams::default()
     })
diff --git a/paddler_tests/src/start_cluster_with_smolvlm2.rs b/paddler_tests/src/start_cluster_with_smolvlm2.rs
index a023f33db..81c906e4d 100644
--- a/paddler_tests/src/start_cluster_with_smolvlm2.rs
+++ b/paddler_tests/src/start_cluster_with_smolvlm2.rs
@@ -2,6 +2,7 @@ use anyhow::Result;
 
 use paddler_test_cluster_harness::agent_config::AgentConfig;
 use paddler_test_cluster_harness::cluster::Cluster;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
 
 use crate::smolvlm2_desired_state::smolvlm2_desired_state;
@@ -10,7 +11,7 @@ use crate::start_cluster::start_cluster;
 pub async fn start_cluster_with_smolvlm2(agents: Vec<AgentConfig>) -> Result<Cluster> {
     start_cluster(ClusterParams {
         agents,
-        desired_state: Some(smolvlm2_desired_state()),
+        desired_state: ClusterDesiredState::Apply(Box::new(smolvlm2_desired_state())),
         wait_for_slots_ready: true,
         ..ClusterParams::default()
     })
diff --git a/paddler_tests/src/start_cluster_with_smolvlm2_and_context_size.rs b/paddler_tests/src/start_cluster_with_smolvlm2_and_context_size.rs
index 857bac263..a024e641b 100644
--- a/paddler_tests/src/start_cluster_with_smolvlm2_and_context_size.rs
+++ b/paddler_tests/src/start_cluster_with_smolvlm2_and_context_size.rs
@@ -2,33 +2,33 @@ use std::num::NonZeroU32;
 
 use anyhow::Result;
 
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
+use paddler_inference_parameters::batch_size::BatchSize;
+use paddler_inference_parameters::model_runtime_parameters::ModelRuntimeParameters;
 use paddler_messaging::balancer_desired_state::BalancerDesiredState;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
 use paddler_test_cluster_harness::cluster::Cluster;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
 
-use crate::batch_size_within_context::batch_size_within_context;
 use crate::smolvlm2_desired_state::smolvlm2_desired_state;
 use crate::start_cluster::start_cluster;
 
 pub async fn start_cluster_with_smolvlm2_and_context_size(
     agents: Vec<AgentConfig>,
-    context_size: u32,
+    context_size: NonZeroU32,
 ) -> Result<Cluster> {
-    let context_size = NonZeroU32::try_from(context_size)?;
     let desired_state = smolvlm2_desired_state();
 
     start_cluster(ClusterParams {
         agents,
-        desired_state: Some(BalancerDesiredState {
-            inference_parameters: InferenceParameters {
+        desired_state: ClusterDesiredState::Apply(Box::new(BalancerDesiredState {
+            model_runtime_parameters: ModelRuntimeParameters {
                 context_size,
-                n_batch: batch_size_within_context(context_size)?,
-                ..desired_state.inference_parameters
+                n_batch: BatchSize::default().at_most(context_size),
+                ..desired_state.model_runtime_parameters
             },
             ..desired_state
-        }),
+        })),
         wait_for_slots_ready: true,
         ..ClusterParams::default()
     })
diff --git a/paddler_tests/src/start_cluster_with_smolvlm2_and_n_batch.rs b/paddler_tests/src/start_cluster_with_smolvlm2_and_n_batch.rs
index e6ed2bfca..5930a0b3c 100644
--- a/paddler_tests/src/start_cluster_with_smolvlm2_and_n_batch.rs
+++ b/paddler_tests/src/start_cluster_with_smolvlm2_and_n_batch.rs
@@ -1,10 +1,11 @@
 use anyhow::Result;
 
 use paddler_inference_parameters::batch_size::BatchSize;
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
+use paddler_inference_parameters::model_runtime_parameters::ModelRuntimeParameters;
 use paddler_messaging::balancer_desired_state::BalancerDesiredState;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
 use paddler_test_cluster_harness::cluster::Cluster;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
 
 use crate::smolvlm2_desired_state::smolvlm2_desired_state;
@@ -12,19 +13,19 @@ use crate::start_cluster::start_cluster;
 
 pub async fn start_cluster_with_smolvlm2_and_n_batch(
     agents: Vec<AgentConfig>,
-    n_batch: u32,
+    n_batch: BatchSize,
 ) -> Result<Cluster> {
     let desired_state = smolvlm2_desired_state();
 
     start_cluster(ClusterParams {
         agents,
-        desired_state: Some(BalancerDesiredState {
-            inference_parameters: InferenceParameters {
-                n_batch: BatchSize::try_from(n_batch)?,
-                ..desired_state.inference_parameters
+        desired_state: ClusterDesiredState::Apply(Box::new(BalancerDesiredState {
+            model_runtime_parameters: ModelRuntimeParameters {
+                n_batch,
+                ..desired_state.model_runtime_parameters
             },
             ..desired_state
-        }),
+        })),
         wait_for_slots_ready: true,
         ..ClusterParams::default()
     })
diff --git a/paddler_tests/src/start_cluster_without_model.rs b/paddler_tests/src/start_cluster_without_model.rs
deleted file mode 100644
index 349a87e2e..000000000
--- a/paddler_tests/src/start_cluster_without_model.rs
+++ /dev/null
@@ -1,29 +0,0 @@
-use anyhow::Result;
-
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
-use paddler_messaging::agent_desired_model::AgentDesiredModel;
-use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-use paddler_test_cluster_harness::agent_config::AgentConfig;
-use paddler_test_cluster_harness::cluster::Cluster;
-use paddler_test_cluster_harness::cluster_params::ClusterParams;
-
-use crate::start_cluster::start_cluster;
-
-pub async fn start_cluster_without_model(
-    agents: Vec<AgentConfig>,
-    inference_parameters: InferenceParameters,
-) -> Result<Cluster> {
-    start_cluster(ClusterParams {
-        agents,
-        desired_state: Some(BalancerDesiredState {
-            chat_template_override: None,
-            inference_parameters,
-            model: AgentDesiredModel::None,
-            multimodal_projection: AgentDesiredModel::None,
-            use_chat_template_override: false,
-        }),
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
-    })
-    .await
-}
diff --git a/paddler_tests/src/start_decision_cluster.rs b/paddler_tests/src/start_decision_cluster.rs
new file mode 100644
index 000000000..0df0a4c21
--- /dev/null
+++ b/paddler_tests/src/start_decision_cluster.rs
@@ -0,0 +1,18 @@
+use std::future::Future;
+
+use anyhow::Result;
+use futures_util::TryFutureExt as _;
+use futures_util::future::ready;
+
+use paddler_test_cluster_harness::cluster::Cluster;
+
+use crate::decision_cluster_params::DecisionClusterParams;
+use crate::start_cluster::start_cluster;
+
+pub fn start_decision_cluster(
+    decision_cluster_params: DecisionClusterParams,
+) -> impl Future<Output = Result<Cluster>> {
+    ready(decision_cluster_params.into_cluster_params())
+        .err_into()
+        .and_then(start_cluster)
+}
diff --git a/paddler_tests/src/start_single_agent_cluster_with_desired_state.rs b/paddler_tests/src/start_single_agent_cluster_with_desired_state.rs
index da30d0f48..1b086949c 100644
--- a/paddler_tests/src/start_single_agent_cluster_with_desired_state.rs
+++ b/paddler_tests/src/start_single_agent_cluster_with_desired_state.rs
@@ -3,6 +3,7 @@ use anyhow::Result;
 use paddler_messaging::balancer_desired_state::BalancerDesiredState;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
 use paddler_test_cluster_harness::cluster::Cluster;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
 
 use crate::start_cluster::start_cluster;
@@ -12,7 +13,7 @@ pub async fn start_single_agent_cluster_with_desired_state(
 ) -> Result<Cluster> {
     start_cluster(ClusterParams {
         agents: AgentConfig::uniform(1, 1),
-        desired_state: Some(desired_state),
+        desired_state: ClusterDesiredState::Apply(Box::new(desired_state)),
         wait_for_slots_ready: false,
         ..ClusterParams::default()
     })
diff --git a/paddler_tests/src/test_cluster_preset.rs b/paddler_tests/src/test_cluster_preset.rs
new file mode 100644
index 000000000..a14656dc3
--- /dev/null
+++ b/paddler_tests/src/test_cluster_preset.rs
@@ -0,0 +1,34 @@
+use clap::ValueEnum;
+
+use paddler_model_source::model_source_error::ModelSourceError;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+
+use crate::kev_checkpoint::KevCheckpoint;
+use crate::qwen3_desired_state::qwen3_desired_state;
+
+const TEXT_GENERATION_SLOTS: u16 = 1;
+
+#[derive(Clone, Copy, ValueEnum)]
+pub enum TestClusterPreset {
+    #[value(name = "kev-0-8b")]
+    Kev0_8b,
+    #[value(name = "qwen3-0-6b")]
+    Qwen3_0_6b,
+}
+
+impl TestClusterPreset {
+    pub fn cluster_params(self) -> Result<ClusterParams, ModelSourceError> {
+        match self {
+            Self::Kev0_8b => KevCheckpoint::Kev0_8b
+                .decision_cluster_params()
+                .into_cluster_params(),
+            Self::Qwen3_0_6b => Ok(ClusterParams {
+                agents: AgentConfig::uniform(1, TEXT_GENERATION_SLOTS),
+                desired_state: ClusterDesiredState::Apply(Box::new(qwen3_desired_state())),
+                ..ClusterParams::default()
+            }),
+        }
+    }
+}
diff --git a/paddler_tests/src/text_generation_pipeline_parts.rs b/paddler_tests/src/text_generation_pipeline_parts.rs
new file mode 100644
index 000000000..4ce371a6c
--- /dev/null
+++ b/paddler_tests/src/text_generation_pipeline_parts.rs
@@ -0,0 +1,53 @@
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent_text_generation::multimodal_projection::MultimodalProjection;
+use paddler_agent_text_generation::text_generation_pipeline::TextGenerationPipeline;
+use paddler_agent_text_generation::text_generation_settings::TextGenerationSettings;
+use paddler_inference_parameters::sampling_parameters::SamplingParameters;
+use paddler_messaging::chat_template_source::ChatTemplateSource;
+use paddler_messaging::multimodal_settings::MultimodalSettings;
+
+use crate::llama_runtime::LlamaRuntime;
+
+pub struct TextGenerationPipelineParts {
+    pub chat_template_source: ChatTemplateSource,
+    pub llama_runtime: LlamaRuntime,
+    pub multimodal_projection: MultimodalProjection,
+}
+
+impl TextGenerationPipelineParts {
+    #[must_use]
+    pub fn into_text_generation_pipeline(self) -> TextGenerationPipeline {
+        let Self {
+            chat_template_source,
+            llama_runtime:
+                LlamaRuntime {
+                    inference_runtime_context,
+                    loaded_llama_model,
+                    model_runtime_parameters,
+                },
+            multimodal_projection,
+        } = self;
+
+        TextGenerationPipeline {
+            inference_runtime_context,
+            loaded_llama_model,
+            model_runtime_parameters,
+            text_generation_settings: TextGenerationSettings {
+                chat_template_source,
+                image_resize_to_fit: MultimodalSettings::default().image_resize_to_fit,
+                multimodal_projection,
+                sampling_parameters: SamplingParameters::default(),
+            },
+        }
+    }
+
+    pub async fn spawn_refusal(self) -> Option<String> {
+        self.into_text_generation_pipeline()
+            .spawn(&CancellationToken::new())
+            .await
+            .err()
+            .as_ref()
+            .map(ToString::to_string)
+    }
+}
diff --git a/paddler_tests/src/tokenizer_reference.rs b/paddler_tests/src/tokenizer_reference.rs
new file mode 100644
index 000000000..8311f1141
--- /dev/null
+++ b/paddler_tests/src/tokenizer_reference.rs
@@ -0,0 +1,10 @@
+use serde::Deserialize;
+
+use crate::tokenizer_reference_entry::TokenizerReferenceEntry;
+
+#[derive(Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct TokenizerReference {
+    pub entries: Vec<TokenizerReferenceEntry>,
+    pub input_normalization: String,
+}
diff --git a/paddler_tests/src/tokenizer_reference_entry.rs b/paddler_tests/src/tokenizer_reference_entry.rs
new file mode 100644
index 000000000..eb3278ab8
--- /dev/null
+++ b/paddler_tests/src/tokenizer_reference_entry.rs
@@ -0,0 +1,8 @@
+use serde::Deserialize;
+
+#[derive(Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct TokenizerReferenceEntry {
+    pub text: String,
+    pub token_ids: Vec<i32>,
+}
diff --git a/paddler_tests/src/two_option_question.rs b/paddler_tests/src/two_option_question.rs
new file mode 100644
index 000000000..af2f65e2f
--- /dev/null
+++ b/paddler_tests/src/two_option_question.rs
@@ -0,0 +1,12 @@
+use llama_cpp_bindings::token::LlamaToken;
+
+use paddler_agent_decision::tokenized_decision_question::TokenizedDecisionQuestion;
+
+#[must_use]
+pub fn two_option_question(id: &str) -> TokenizedDecisionQuestion {
+    TokenizedDecisionQuestion {
+        id: id.to_owned(),
+        instructions: vec![LlamaToken::new(0)],
+        options: vec![vec![LlamaToken::new(1)], vec![LlamaToken::new(2)]],
+    }
+}
diff --git a/paddler_tests/src/uncompilable_chat_template.rs b/paddler_tests/src/uncompilable_chat_template.rs
new file mode 100644
index 000000000..d29c28675
--- /dev/null
+++ b/paddler_tests/src/uncompilable_chat_template.rs
@@ -0,0 +1,10 @@
+use paddler_messaging::chat_template::ChatTemplate;
+
+pub const UNCOMPILABLE_CHAT_TEMPLATE_CONTENT: &str = "{{invalid jinja template";
+
+#[must_use]
+pub fn uncompilable_chat_template() -> ChatTemplate {
+    ChatTemplate {
+        content: UNCOMPILABLE_CHAT_TEMPLATE_CONTENT.to_owned(),
+    }
+}
diff --git a/paddler_tests/src/uncompilable_chat_template_issue.rs b/paddler_tests/src/uncompilable_chat_template_issue.rs
new file mode 100644
index 000000000..b9ad23cd8
--- /dev/null
+++ b/paddler_tests/src/uncompilable_chat_template_issue.rs
@@ -0,0 +1,22 @@
+use std::path::Path;
+
+use paddler_agent_text_generation::chat_template_renderer::ChatTemplateRenderer;
+use paddler_messaging::agent_issue::AgentIssue;
+use paddler_messaging::agent_issue_params::chat_template_does_not_compile_params::ChatTemplateDoesNotCompileParams;
+use paddler_messaging::agent_issue_params::model_path::ModelPath;
+
+use crate::uncompilable_chat_template::UNCOMPILABLE_CHAT_TEMPLATE_CONTENT;
+use crate::uncompilable_chat_template::uncompilable_chat_template;
+
+#[must_use]
+pub fn uncompilable_chat_template_issue(model_path: &Path) -> Option<AgentIssue> {
+    ChatTemplateRenderer::new(uncompilable_chat_template())
+        .err()
+        .map(|compile_error| {
+            AgentIssue::ChatTemplateDoesNotCompile(ChatTemplateDoesNotCompileParams {
+                error: compile_error.to_string(),
+                model_path: ModelPath::from(model_path),
+                template_content: UNCOMPILABLE_CHAT_TEMPLATE_CONTENT.to_owned(),
+            })
+        })
+}
diff --git a/paddler_tests/src/weather_question_conversation.rs b/paddler_tests/src/weather_question_conversation.rs
index dd4af7749..881c49393 100644
--- a/paddler_tests/src/weather_question_conversation.rs
+++ b/paddler_tests/src/weather_question_conversation.rs
@@ -16,7 +16,7 @@ pub fn weather_question_conversation(
 ) -> ContinueFromConversationHistoryParams<ValidatedParametersSchema> {
     ContinueFromConversationHistoryParams {
         add_generation_prompt: true,
-        conversation_history: ConversationHistory::new(vec![ConversationMessage {
+        conversation_history: ConversationHistory::new(vec![ConversationMessage::Message {
             content: ConversationMessageContent::Text("What is the weather in Paris?".to_owned()),
             role: "user".to_owned(),
         }]),
diff --git a/paddler_tests/src/weather_tool_parameters_with_an_unsatisfiable_location.rs b/paddler_tests/src/weather_tool_parameters_with_an_unsatisfiable_location.rs
new file mode 100644
index 000000000..67c250b8d
--- /dev/null
+++ b/paddler_tests/src/weather_tool_parameters_with_an_unsatisfiable_location.rs
@@ -0,0 +1,23 @@
+use serde_json::Map;
+use serde_json::Value;
+use serde_json::json;
+
+use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters_schema::validated_parameters_schema::ValidatedParametersSchema;
+
+#[must_use]
+pub fn weather_tool_parameters_with_an_unsatisfiable_location() -> ValidatedParametersSchema {
+    ValidatedParametersSchema {
+        schema_type: "object".to_owned(),
+        properties: Some(Map::from_iter([(
+            "location".to_owned(),
+            json!({
+                "type": "integer",
+                "minimum": 1,
+                "maximum": 0,
+                "description": "The city name"
+            }),
+        )])),
+        required: Some(vec!["location".to_owned()]),
+        additional_properties: Some(Value::Bool(false)),
+    }
+}
diff --git a/paddler_tests/src/weather_tool_question.rs b/paddler_tests/src/weather_tool_question.rs
new file mode 100644
index 000000000..bee97137e
--- /dev/null
+++ b/paddler_tests/src/weather_tool_question.rs
@@ -0,0 +1,2 @@
+pub const WEATHER_TOOL_QUESTION: &str =
+    "What is the weather in Paris? Use the get_weather tool to find out.";
diff --git a/paddler_tests/src/write_patched_gguf.rs b/paddler_tests/src/write_patched_gguf.rs
new file mode 100644
index 000000000..de069e4a6
--- /dev/null
+++ b/paddler_tests/src/write_patched_gguf.rs
@@ -0,0 +1,19 @@
+use std::path::Path;
+
+use llama_cpp_bindings::gguf_context::GgufContext;
+use llama_cpp_bindings::gguf_context_error::GgufContextError;
+
+pub fn write_patched_gguf<TPatch>(
+    source_path: &Path,
+    destination_path: &Path,
+    patch: TPatch,
+) -> Result<(), GgufContextError>
+where
+    TPatch: FnOnce(&mut GgufContext),
+{
+    let mut gguf_context = GgufContext::from_file(source_path)?;
+
+    patch(&mut gguf_context);
+
+    gguf_context.write_to_file(destination_path)
+}
diff --git a/paddler_tests/tests/a_balancer_starts_from_the_stored_desired_state_through_reconciliation.rs b/paddler_tests/tests/a_balancer_starts_from_the_stored_desired_state_through_reconciliation.rs
new file mode 100644
index 000000000..8c0d9f427
--- /dev/null
+++ b/paddler_tests/tests/a_balancer_starts_from_the_stored_desired_state_through_reconciliation.rs
@@ -0,0 +1,47 @@
+use paddler_messaging::agent_desired_model::AgentDesiredModel;
+use paddler_messaging::agent_issue::AgentIssue;
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_state_database::state_database_file::StateDatabaseFile;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_test_cluster_harness::temporary_state_database_file::TemporaryStateDatabaseFile;
+use paddler_tests::start_cluster::start_cluster;
+
+const STORED_MODEL_URI: &str = "a stored uri that does not parse";
+
+#[tokio::test(flavor = "multi_thread")]
+async fn a_balancer_starts_from_the_stored_desired_state_through_reconciliation() {
+    let database =
+        TemporaryStateDatabaseFile::new().expect("the state database file must be created");
+
+    StateDatabaseFile::new(database.path.clone())
+        .store_balancer_desired_state(&BalancerDesiredState {
+            model: AgentDesiredModel::Uri(STORED_MODEL_URI.to_owned()),
+            ..BalancerDesiredState::default()
+        })
+        .await
+        .expect("the desired state must be stored before the balancer starts");
+
+    let mut cluster = start_cluster(ClusterParams {
+        agents: AgentConfig::uniform(1, 1),
+        desired_state: ClusterDesiredState::KeepStored,
+        state_database_url: database.url.clone(),
+        wait_for_slots_ready: false,
+        ..ClusterParams::default()
+    })
+    .await
+    .expect("the cluster must start from its stored desired state");
+
+    cluster
+        .wait_for_first_agent_issue(|issue| {
+            matches!(issue, AgentIssue::ModelUriIsUnparseable(model_path) if model_path.model_path == STORED_MODEL_URI)
+        })
+        .await
+        .expect("the agent must reconcile the stored desired state");
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/a_client_gone_before_an_unterminated_tool_call_resolves_receives_nothing.rs b/paddler_tests/tests/a_client_gone_before_an_unterminated_tool_call_resolves_receives_nothing.rs
new file mode 100644
index 000000000..c80c4623a
--- /dev/null
+++ b/paddler_tests/tests/a_client_gone_before_an_unterminated_tool_call_resolves_receives_nothing.rs
@@ -0,0 +1,72 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use llama_cpp_bindings::sampling::LlamaSampler;
+use llama_cpp_bindings::token::logit_bias::LlamaLogitBias;
+use llama_cpp_bindings::token_piece::TokenPiece;
+use tokio::sync::mpsc::unbounded_channel;
+
+use paddler_agent_text_generation::continuous_batch_terminal_outcome::ContinuousBatchTerminalOutcome;
+use paddler_agent_text_generation::grammar_sampling::GrammarSampling;
+use paddler_agent_text_generation::prepared_generation_request::PreparedGenerationRequest;
+use paddler_agent_text_generation::token_sampling::TokenSampling;
+use paddler_agent_text_generation::tool_call_handling::ToolCallHandling;
+use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
+use paddler_tests::agent_inference_runtime_context::agent_inference_runtime_context;
+use paddler_tests::llama_runtime::LlamaRuntime;
+
+use crate::decode_prompt_for_sampling::decode_prompt_for_sampling;
+use crate::ready_to_sample_request::ReadyToSampleRequest;
+use crate::tool_call_pipeline_for_tools::tool_call_pipeline_for_tools;
+
+const PROMPT: &str = "Paddler balances";
+
+#[test]
+fn a_client_gone_before_an_unterminated_tool_call_resolves_receives_nothing() {
+    let runtime = LlamaRuntime::load(
+        &qwen3_0_6b()
+            .cached_path()
+            .expect("the model must be in the Hugging Face cache"),
+        agent_inference_runtime_context(1),
+    )
+    .expect("the model must load");
+    let mut llama_context = runtime
+        .create_text_generation_llama_context(1)
+        .expect("the context must be created");
+    let batch_position = decode_prompt_for_sampling(&runtime, &mut llama_context, PROMPT);
+    let model = &runtime.loaded_llama_model.model;
+    let end_of_generation = model
+        .token_eos()
+        .expect("the model must have an end-of-generation token");
+    let mut unterminated_tool_call = tool_call_pipeline_for_tools(model, &[]);
+    let (gone_client_tx, gone_client_rx) = unbounded_channel();
+
+    unterminated_tool_call.feed(&TokenPiece::Marker("<tool_call>".to_owned()));
+    drop(gone_client_rx);
+
+    assert_eq!(
+        ReadyToSampleRequest {
+            batch_position,
+            llama_context: &llama_context,
+            prompt: PROMPT,
+            runtime: &runtime,
+        }
+        .advance(|prepared_request| PreparedGenerationRequest {
+            generated_tokens_tx: gone_client_tx,
+            token_sampling: TokenSampling {
+                chain: LlamaSampler::chain_simple([
+                    LlamaSampler::logit_bias(
+                        model.n_vocab(),
+                        &[LlamaLogitBias::new(end_of_generation, f32::INFINITY)],
+                    )
+                    .expect("the logit bias sampler must initialize"),
+                    LlamaSampler::greedy().expect("the greedy sampler must initialize"),
+                ])
+                .expect("the sampler chain must initialize"),
+                grammar: GrammarSampling::Unconstrained,
+            },
+            tool_call_handling: ToolCallHandling::Parsed(unterminated_tool_call),
+            ..prepared_request
+        }),
+        ContinuousBatchTerminalOutcome::EmitNothing
+    );
+}
diff --git a/paddler_tests/tests/a_cluster_without_a_pointer_head_refuses_decisions_as_unavailable.rs b/paddler_tests/tests/a_cluster_without_a_pointer_head_refuses_decisions_as_unavailable.rs
new file mode 100644
index 000000000..0100be96a
--- /dev/null
+++ b/paddler_tests/tests/a_cluster_without_a_pointer_head_refuses_decisions_as_unavailable.rs
@@ -0,0 +1,51 @@
+use tokio_util::sync::CancellationToken;
+
+use paddler_client::error::Error as ClientError;
+use paddler_messaging::api_path::ApiPath;
+use paddler_messaging::decision_question::DecisionQuestion;
+use paddler_messaging::request_params::decide_params::raw_decide_params::RawDecideParams;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_tests::start_cluster::start_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn a_cluster_without_a_pointer_head_refuses_decisions_as_unavailable() {
+    let cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("a balancer without a pointer head must start");
+
+    let decide_url = cluster
+        .balancer
+        .inference_base_url
+        .join(ApiPath::DECIDE)
+        .expect("the decide path must extend the inference address");
+    let decision_rejection = cluster
+        .client_inference
+        .post_decide(
+            CancellationToken::new(),
+            &RawDecideParams {
+                questions: vec![DecisionQuestion {
+                    id: "paid".to_owned(),
+                    instructions: "Was it paid?".to_owned(),
+                    options: vec!["no".to_owned(), "yes".to_owned()],
+                }],
+                state: "The invoice was paid on time.".to_owned(),
+            },
+        )
+        .await
+        .err()
+        .expect("a cluster without a pointer head must refuse a decision");
+
+    assert_eq!(
+        decision_rejection.to_string(),
+        ClientError::ServiceUnavailable {
+            message: "No pointer head is configured, so the cluster serves no decisions".to_owned(),
+            url: decide_url.to_string(),
+        }
+        .to_string()
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/a_decision_admitted_to_a_sequence_that_cannot_be_cleared_fails_with_kv_cache_removal.rs b/paddler_tests/tests/a_decision_admitted_to_a_sequence_that_cannot_be_cleared_fails_with_kv_cache_removal.rs
new file mode 100644
index 000000000..a593c6fa7
--- /dev/null
+++ b/paddler_tests/tests/a_decision_admitted_to_a_sequence_that_cannot_be_cleared_fails_with_kv_cache_removal.rs
@@ -0,0 +1,72 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::collections::VecDeque;
+
+use llama_cpp_bindings::error::ClearKvCacheSeqError;
+use tokio::sync::mpsc::unbounded_channel;
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent_decision::decision_capacity_ledger::DecisionCapacityLedger;
+use paddler_agent_decision::decision_error::DecisionError;
+use paddler_agent_runtime::scheduler_inbox::SchedulerInbox;
+use paddler_agent_runtime::scheduler_message::SchedulerMessage;
+use paddler_messaging::decision_result::DecisionResult;
+use paddler_tests::decision_scheduler_run::DecisionSchedulerRun;
+use paddler_tests::prepared_decision::PreparedDecision;
+
+use crate::kev_decision_model::KevDecisionModel;
+
+const SEQUENCES_THE_CONTEXT_HOLDS: u16 = 1;
+const SEQUENCES_THE_LEDGER_HANDS_OUT: u16 = 2;
+
+#[test]
+fn a_decision_admitted_to_a_sequence_that_cannot_be_cleared_fails_with_kv_cache_removal() {
+    let kev_decision_model = KevDecisionModel::load(SEQUENCES_THE_LEDGER_HANDS_OUT);
+    let prepared_decision = |question_id: &str| {
+        PreparedDecision::new(
+            kev_decision_model
+                .runtime
+                .inference_runtime_context
+                .slot_aggregated_status
+                .clone(),
+            kev_decision_model.two_option_question_layout(question_id),
+        )
+    };
+    let decision_on_a_clearable_sequence = prepared_decision("clearable");
+    let mut decision_on_a_sequence_beyond_the_context = prepared_decision("beyond");
+    let (scheduler_message_tx, scheduler_message_rx) = unbounded_channel();
+
+    for prepared_decision_request in [
+        decision_on_a_clearable_sequence.prepared_decision_request,
+        decision_on_a_sequence_beyond_the_context.prepared_decision_request,
+    ] {
+        scheduler_message_tx
+            .send(SchedulerMessage::Command(prepared_decision_request))
+            .expect("the scheduler must receive the decision");
+    }
+
+    drop(scheduler_message_tx);
+
+    kev_decision_model.run_scheduler(DecisionSchedulerRun {
+        active_requests: VecDeque::new(),
+        agent_shutdown: CancellationToken::new(),
+        capacity_ledger: DecisionCapacityLedger::new(SEQUENCES_THE_LEDGER_HANDS_OUT),
+        pending_requests: VecDeque::new(),
+        scheduler_inbox: SchedulerInbox::new(scheduler_message_rx),
+        sequences: SEQUENCES_THE_CONTEXT_HOLDS,
+    });
+
+    assert_eq!(
+        decision_on_a_sequence_beyond_the_context
+            .decision_result_rx
+            .try_recv(),
+        Ok(DecisionResult::KvCacheRemovalFailed(
+            DecisionError::KvCacheRemovalFailed(ClearKvCacheSeqError::PartialSequenceNotRemoved {
+                seq_id: 1,
+                p0: -1,
+                p1: -1,
+            })
+            .to_string()
+        ))
+    );
+}
diff --git a/paddler_tests/tests/a_decision_arriving_while_another_is_decided_is_decided_too.rs b/paddler_tests/tests/a_decision_arriving_while_another_is_decided_is_decided_too.rs
new file mode 100644
index 000000000..a125c2714
--- /dev/null
+++ b/paddler_tests/tests/a_decision_arriving_while_another_is_decided_is_decided_too.rs
@@ -0,0 +1,77 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::collections::VecDeque;
+use std::mem::discriminant;
+
+use tokio::sync::mpsc::UnboundedReceiver;
+use tokio::sync::mpsc::unbounded_channel;
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent_decision::decision_capacity_ledger::DecisionCapacityLedger;
+use paddler_agent_runtime::scheduler_inbox::SchedulerInbox;
+use paddler_agent_runtime::scheduler_message::SchedulerMessage;
+use paddler_messaging::decision_answer::DecisionAnswer;
+use paddler_messaging::decision_result::DecisionResult;
+use paddler_tests::decision_done_discriminant::decision_done_discriminant;
+use paddler_tests::decision_scheduler_run::DecisionSchedulerRun;
+use paddler_tests::prepared_decision::PreparedDecision;
+
+use crate::kev_decision_model::KevDecisionModel;
+
+const CONCURRENT_DECISIONS: u16 = 2;
+
+#[test]
+fn a_decision_arriving_while_another_is_decided_is_decided_too() {
+    let kev_decision_model = KevDecisionModel::load(CONCURRENT_DECISIONS);
+    let prepared_decision = |question_id: &str| {
+        PreparedDecision::new(
+            kev_decision_model
+                .runtime
+                .inference_runtime_context
+                .slot_aggregated_status
+                .clone(),
+            kev_decision_model.two_option_question_layout(question_id),
+        )
+    };
+    let first_decision = prepared_decision("first");
+    let second_decision = prepared_decision("second");
+    let (scheduler_message_tx, scheduler_message_rx) = unbounded_channel();
+
+    for prepared_decision_request in [
+        first_decision.prepared_decision_request,
+        second_decision.prepared_decision_request,
+    ] {
+        scheduler_message_tx
+            .send(SchedulerMessage::Command(prepared_decision_request))
+            .expect("the scheduler must receive the decision");
+    }
+
+    drop(scheduler_message_tx);
+
+    kev_decision_model.run_scheduler(DecisionSchedulerRun {
+        active_requests: VecDeque::new(),
+        agent_shutdown: CancellationToken::new(),
+        capacity_ledger: DecisionCapacityLedger::new(CONCURRENT_DECISIONS),
+        pending_requests: VecDeque::new(),
+        scheduler_inbox: SchedulerInbox::new(scheduler_message_rx),
+        sequences: CONCURRENT_DECISIONS,
+    });
+
+    let assert_decided =
+        |question_id: &str, mut decision_result_rx: UnboundedReceiver<DecisionResult>| {
+            assert!(matches!(
+                decision_result_rx.try_recv(),
+                Ok(DecisionResult::QuestionAnswered(DecisionAnswer { id, probabilities }))
+                    if id == question_id && probabilities.len() == 2
+            ));
+            assert_eq!(
+                decision_result_rx
+                    .try_recv()
+                    .map(|decision_result| discriminant(&decision_result)),
+                Ok(decision_done_discriminant())
+            );
+        };
+
+    assert_decided("first", first_decision.decision_result_rx);
+    assert_decided("second", second_decision.decision_result_rx);
+}
diff --git a/paddler_tests/tests/a_decision_cluster_refuses_a_model_path_that_is_not_absolute.rs b/paddler_tests/tests/a_decision_cluster_refuses_a_model_path_that_is_not_absolute.rs
new file mode 100644
index 000000000..2351ceb5e
--- /dev/null
+++ b/paddler_tests/tests/a_decision_cluster_refuses_a_model_path_that_is_not_absolute.rs
@@ -0,0 +1,31 @@
+use std::path::PathBuf;
+
+use paddler_model_source::model_source_error::ModelSourceError;
+use paddler_tests::decision_backbone::DecisionBackbone;
+use paddler_tests::decision_cluster_params::DecisionClusterParams;
+use paddler_tests::decision_desired_state::DecisionDesiredState;
+use paddler_tests::start_decision_cluster::start_decision_cluster;
+
+const RELATIVE_MODEL_PATH: &str = "model.gguf";
+
+#[tokio::test(flavor = "multi_thread")]
+async fn a_decision_cluster_refuses_a_model_path_that_is_not_absolute() {
+    assert_eq!(
+        start_decision_cluster(DecisionClusterParams {
+            desired_state: DecisionDesiredState {
+                backbone: DecisionBackbone::LocalFile(PathBuf::from(RELATIVE_MODEL_PATH)),
+                ..DecisionDesiredState::default()
+            },
+            ..DecisionClusterParams::default()
+        })
+        .await
+        .err()
+        .map(|startup_error| startup_error.to_string()),
+        Some(
+            ModelSourceError::LocalModelPathNotAbsolute {
+                path: PathBuf::from(RELATIVE_MODEL_PATH),
+            }
+            .to_string()
+        )
+    );
+}
diff --git a/paddler_tests/tests/a_decision_cluster_refuses_a_pointer_head_path_that_is_not_absolute.rs b/paddler_tests/tests/a_decision_cluster_refuses_a_pointer_head_path_that_is_not_absolute.rs
new file mode 100644
index 000000000..90a5e5260
--- /dev/null
+++ b/paddler_tests/tests/a_decision_cluster_refuses_a_pointer_head_path_that_is_not_absolute.rs
@@ -0,0 +1,30 @@
+use std::path::PathBuf;
+
+use paddler_model_source::model_source_error::ModelSourceError;
+use paddler_tests::decision_cluster_params::DecisionClusterParams;
+use paddler_tests::decision_desired_state::DecisionDesiredState;
+use paddler_tests::start_decision_cluster::start_decision_cluster;
+
+const RELATIVE_POINTER_HEAD_PATH: &str = "pointer_head.gguf";
+
+#[tokio::test(flavor = "multi_thread")]
+async fn a_decision_cluster_refuses_a_pointer_head_path_that_is_not_absolute() {
+    assert_eq!(
+        start_decision_cluster(DecisionClusterParams {
+            desired_state: DecisionDesiredState {
+                pointer_head: PathBuf::from(RELATIVE_POINTER_HEAD_PATH),
+                ..DecisionDesiredState::default()
+            },
+            ..DecisionClusterParams::default()
+        })
+        .await
+        .err()
+        .map(|startup_error| startup_error.to_string()),
+        Some(
+            ModelSourceError::LocalModelPathNotAbsolute {
+                path: PathBuf::from(RELATIVE_POINTER_HEAD_PATH),
+            }
+            .to_string()
+        )
+    );
+}
diff --git a/paddler_tests/tests/a_decision_delimiter_containing_a_nul_byte_cannot_be_resolved.rs b/paddler_tests/tests/a_decision_delimiter_containing_a_nul_byte_cannot_be_resolved.rs
new file mode 100644
index 000000000..00cfd7228
--- /dev/null
+++ b/paddler_tests/tests/a_decision_delimiter_containing_a_nul_byte_cannot_be_resolved.rs
@@ -0,0 +1,46 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::ffi::CString;
+
+use llama_cpp_bindings::error::StringToTokenError;
+
+use paddler_agent_decision::decision_delimiter_tokens::DecisionDelimiterTokens;
+use paddler_agent_decision::decision_error::DecisionError;
+use paddler_agent_pointer_head::pointer_head_delimiters::PointerHeadDelimiters;
+use paddler_tests::agent_inference_runtime_context::agent_inference_runtime_context;
+use paddler_tests::kev_0_8b_pointer_head::kev_0_8b_pointer_head;
+use paddler_tests::kev_checkpoint::KevCheckpoint;
+use paddler_tests::llama_runtime::LlamaRuntime;
+
+#[test]
+fn a_decision_delimiter_containing_a_nul_byte_cannot_be_resolved() {
+    let runtime = LlamaRuntime::load(
+        &KevCheckpoint::Kev0_8b.target_path("model.gguf"),
+        agent_inference_runtime_context(1),
+    )
+    .expect("the kev decision model must load");
+    let pointer_head = kev_0_8b_pointer_head().expect("the kev pointer head must load");
+    let delimiter_with_nul_byte = format!("{}\0", pointer_head.delimiters.decide);
+    let nul_error = CString::new(delimiter_with_nul_byte.clone())
+        .expect_err("the delimiter must contain a nul byte");
+
+    assert_eq!(
+        DecisionDelimiterTokens::resolve(
+            &runtime.loaded_llama_model.model.llama_model,
+            &PointerHeadDelimiters {
+                decide: delimiter_with_nul_byte.clone(),
+                ..pointer_head.delimiters
+            },
+        )
+        .err()
+        .as_ref()
+        .map(ToString::to_string),
+        Some(
+            DecisionError::DelimiterTokenizationFailed {
+                delimiter: delimiter_with_nul_byte,
+                source: StringToTokenError::NulError(nul_error),
+            }
+            .to_string()
+        )
+    );
+}
diff --git a/paddler_tests/tests/a_decision_evaluates_every_input_token_exactly_once.rs b/paddler_tests/tests/a_decision_evaluates_every_input_token_exactly_once.rs
new file mode 100644
index 000000000..2bf996b91
--- /dev/null
+++ b/paddler_tests/tests/a_decision_evaluates_every_input_token_exactly_once.rs
@@ -0,0 +1,73 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use llama_cpp_bindings::llama_batch::LlamaBatch;
+use llama_cpp_bindings::token::LlamaToken;
+
+use paddler_agent_decision::decision_advance::DecisionAdvance;
+use paddler_agent_decision::decision_capacity_ledger::DecisionCapacityLedger;
+use paddler_agent_decision::decision_decoder::DecisionDecoder;
+use paddler_agent_decision::decision_token_layout::DecisionTokenLayout;
+use paddler_messaging::decision_slots_minimum::DECISION_SLOTS_MINIMUM;
+use paddler_tests::admitted_decision::AdmittedDecision;
+use paddler_tests::two_option_question::two_option_question;
+
+use crate::kev_decision_model::KevDecisionModel;
+
+#[test]
+fn a_decision_evaluates_every_input_token_exactly_once() {
+    let KevDecisionModel {
+        delimiter_tokens,
+        pointer_head,
+        runtime,
+    } = KevDecisionModel::load(DECISION_SLOTS_MINIMUM.get());
+    let mut llama_context = runtime
+        .create_decision_llama_context(u32::from(DECISION_SLOTS_MINIMUM.get()))
+        .expect("the decision context must be created");
+    let n_batch = runtime.model_runtime_parameters.n_batch.tokens_usize();
+    let mut batch = LlamaBatch::new(runtime.model_runtime_parameters.n_batch.capacity(), 1);
+    let layout = DecisionTokenLayout::new(
+        &delimiter_tokens,
+        vec![LlamaToken::new(0); 8],
+        vec![two_option_question("leading")],
+        two_option_question("last"),
+    );
+    let decision_input_tokens = layout.total_tokens();
+    let AdmittedDecision {
+        mut active_decision_request,
+        decision_result_rx: _decision_result_rx,
+        decision_stop_tx: _decision_stop_tx,
+    } = AdmittedDecision::admit(
+        &DecisionCapacityLedger::new(DECISION_SLOTS_MINIMUM.get()),
+        runtime
+            .inference_runtime_context
+            .slot_aggregated_status
+            .clone(),
+        layout,
+    )
+    .expect("the decision must be admitted");
+
+    llama_context.reset_timings();
+
+    while matches!(
+        active_decision_request
+            .advance(
+                &mut DecisionDecoder {
+                    batch: &mut batch,
+                    llama_context: &mut llama_context,
+                    n_batch,
+                },
+                &pointer_head,
+                None,
+            )
+            .expect("the decision must advance"),
+        DecisionAdvance::Continuing
+    ) {}
+
+    llama_context.synchronize();
+
+    assert_eq!(
+        usize::try_from(llama_context.timings().n_p_eval())
+            .expect("the evaluated token count must not be negative"),
+        decision_input_tokens
+    );
+}
diff --git a/paddler_tests/tests/a_decision_leaving_the_waiting_line_lets_the_completion_behind_it_dispatch.rs b/paddler_tests/tests/a_decision_leaving_the_waiting_line_lets_the_completion_behind_it_dispatch.rs
new file mode 100644
index 000000000..727d4aa59
--- /dev/null
+++ b/paddler_tests/tests/a_decision_leaving_the_waiting_line_lets_the_completion_behind_it_dispatch.rs
@@ -0,0 +1,66 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::num::NonZeroU32;
+
+use tokio::spawn;
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
+use paddler_tests::decision_cluster_params::DecisionClusterParams;
+use paddler_tests::sample_decision::sample_decision;
+use paddler_tests::start_decision_cluster::start_decision_cluster;
+
+use crate::slot_held_by_an_unending_generation::SlotHeldByAnUnendingGeneration;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn a_decision_leaving_the_waiting_line_lets_the_completion_behind_it_dispatch() {
+    let mut cluster = start_decision_cluster(DecisionClusterParams::default())
+        .await
+        .expect("a decision cluster must start");
+    let slot_held_by_an_unending_generation =
+        SlotHeldByAnUnendingGeneration::hold(&mut cluster).await;
+    let waiting_decision = cluster
+        .client_inference
+        .post_decide(CancellationToken::new(), &sample_decision())
+        .await
+        .expect("the decision must be accepted");
+
+    cluster
+        .wait_for_buffered_request_count(1)
+        .await
+        .expect("the decision must wait while only one slot is free");
+
+    let completion = spawn(cluster.continue_from_raw_prompt(
+        CancellationToken::new(),
+        &ContinueFromRawPromptParams {
+            grammar: None,
+            max_tokens: NonZeroU32::MIN,
+            raw_prompt: "The capital of France is".to_owned(),
+        },
+    ));
+
+    cluster
+        .wait_for_buffered_request_count(2)
+        .await
+        .expect("the completion must wait behind the decision");
+
+    drop(waiting_decision);
+
+    assert_eq!(
+        completion
+            .await
+            .expect("the completion task must not panic")
+            .expect("the completion must be answered")
+            .summary()
+            .expect("the completion must finish with a summary")
+            .usage
+            .completion_tokens(),
+        1
+    );
+
+    slot_held_by_an_unending_generation.release().await;
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/a_decision_pending_admission_when_the_agent_shuts_down_is_refused_as_unavailable.rs b/paddler_tests/tests/a_decision_pending_admission_when_the_agent_shuts_down_is_refused_as_unavailable.rs
new file mode 100644
index 000000000..170c1d068
--- /dev/null
+++ b/paddler_tests/tests/a_decision_pending_admission_when_the_agent_shuts_down_is_refused_as_unavailable.rs
@@ -0,0 +1,54 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::collections::VecDeque;
+
+use tokio::sync::mpsc::unbounded_channel;
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent_decision::decision_capacity_ledger::DecisionCapacityLedger;
+use paddler_agent_decision::decision_error::DecisionError;
+use paddler_agent_runtime::scheduler_inbox::SchedulerInbox;
+use paddler_messaging::decision_result::DecisionResult;
+use paddler_tests::decision_scheduler_run::DecisionSchedulerRun;
+use paddler_tests::prepared_decision::PreparedDecision;
+
+use crate::kev_decision_model::KevDecisionModel;
+
+const DECISION_SEQUENCES: u16 = 2;
+
+#[test]
+fn a_decision_pending_admission_when_the_agent_shuts_down_is_refused_as_unavailable() {
+    let kev_decision_model = KevDecisionModel::load(DECISION_SEQUENCES);
+    let PreparedDecision {
+        mut decision_result_rx,
+        decision_stop_tx: _decision_stop_tx,
+        prepared_decision_request,
+    } = PreparedDecision::new(
+        kev_decision_model
+            .runtime
+            .inference_runtime_context
+            .slot_aggregated_status
+            .clone(),
+        kev_decision_model.two_option_question_layout("question"),
+    );
+    let agent_shutdown = CancellationToken::new();
+    let (_scheduler_message_tx, scheduler_message_rx) = unbounded_channel();
+
+    agent_shutdown.cancel();
+
+    kev_decision_model.run_scheduler(DecisionSchedulerRun {
+        active_requests: VecDeque::new(),
+        agent_shutdown,
+        capacity_ledger: DecisionCapacityLedger::new(DECISION_SEQUENCES),
+        pending_requests: VecDeque::from([prepared_decision_request]),
+        scheduler_inbox: SchedulerInbox::new(scheduler_message_rx),
+        sequences: DECISION_SEQUENCES,
+    });
+
+    assert_eq!(
+        decision_result_rx.try_recv(),
+        Ok(DecisionResult::SchedulerUnavailable(
+            DecisionError::SchedulerUnavailable.to_string()
+        ))
+    );
+}
diff --git a/paddler_tests/tests/a_decision_pipeline_whose_context_cannot_warm_up_every_slot_refuses_to_start.rs b/paddler_tests/tests/a_decision_pipeline_whose_context_cannot_warm_up_every_slot_refuses_to_start.rs
new file mode 100644
index 000000000..799f1906b
--- /dev/null
+++ b/paddler_tests/tests/a_decision_pipeline_whose_context_cannot_warm_up_every_slot_refuses_to_start.rs
@@ -0,0 +1,58 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::num::NonZeroU32;
+
+use llama_cpp_bindings::error::DecodeError;
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent_decision::decision_error::DecisionError;
+use paddler_agent_decision::decision_pipeline::DecisionPipeline;
+use paddler_agent_decision::loaded_pointer_head::LoadedPointerHead;
+use paddler_agent_runtime::agent_runtime_error::AgentRuntimeError;
+use paddler_inference_parameters::model_runtime_parameters::ModelRuntimeParameters;
+use paddler_tests::agent_inference_runtime_context::agent_inference_runtime_context;
+use paddler_tests::kev_0_8b_pointer_head::kev_0_8b_pointer_head;
+use paddler_tests::kev_checkpoint::KevCheckpoint;
+use paddler_tests::llama_runtime::LlamaRuntime;
+
+const SLOTS: u16 = 200;
+const CACHE_CELLS_FEWER_THAN_TWO_PER_SLOT: NonZeroU32 = NonZeroU32::new(256).unwrap();
+
+#[tokio::test(flavor = "multi_thread")]
+async fn a_decision_pipeline_whose_context_cannot_warm_up_every_slot_refuses_to_start() {
+    let LlamaRuntime {
+        inference_runtime_context,
+        loaded_llama_model,
+        model_runtime_parameters,
+    } = LlamaRuntime::load(
+        &KevCheckpoint::Kev0_8b.target_path("model.gguf"),
+        agent_inference_runtime_context(SLOTS),
+    )
+    .expect("the kev model must load");
+
+    assert_eq!(
+        DecisionPipeline {
+            inference_runtime_context,
+            loaded_llama_model,
+            loaded_pointer_head: LoadedPointerHead {
+                pointer_head: kev_0_8b_pointer_head().expect("the kev pointer head must load"),
+                pointer_head_path: KevCheckpoint::Kev0_8b.target_path("pointer_head.gguf"),
+            },
+            model_runtime_parameters: ModelRuntimeParameters {
+                context_size: CACHE_CELLS_FEWER_THAN_TWO_PER_SLOT,
+                ..model_runtime_parameters
+            },
+        }
+        .spawn(&CancellationToken::new())
+        .await
+        .err()
+        .as_ref()
+        .map(ToString::to_string),
+        Some(
+            DecisionError::AgentRuntime(AgentRuntimeError::WarmUpDecodeFailed(
+                DecodeError::NoKvCacheSlot
+            ))
+            .to_string()
+        )
+    );
+}
diff --git a/paddler_tests/tests/a_decision_running_when_the_agent_shuts_down_is_refused_as_unavailable.rs b/paddler_tests/tests/a_decision_running_when_the_agent_shuts_down_is_refused_as_unavailable.rs
new file mode 100644
index 000000000..a9f10dc18
--- /dev/null
+++ b/paddler_tests/tests/a_decision_running_when_the_agent_shuts_down_is_refused_as_unavailable.rs
@@ -0,0 +1,58 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::collections::VecDeque;
+
+use tokio::sync::mpsc::unbounded_channel;
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent_decision::decision_capacity_ledger::DecisionCapacityLedger;
+use paddler_agent_decision::decision_error::DecisionError;
+use paddler_agent_runtime::scheduler_inbox::SchedulerInbox;
+use paddler_messaging::decision_result::DecisionResult;
+use paddler_tests::admitted_decision::AdmittedDecision;
+use paddler_tests::decision_scheduler_run::DecisionSchedulerRun;
+
+use crate::kev_decision_model::KevDecisionModel;
+
+const DECISION_SEQUENCES: u16 = 2;
+
+#[test]
+fn a_decision_running_when_the_agent_shuts_down_is_refused_as_unavailable() {
+    let kev_decision_model = KevDecisionModel::load(DECISION_SEQUENCES);
+    let capacity_ledger = DecisionCapacityLedger::new(DECISION_SEQUENCES);
+    let layout = kev_decision_model.two_option_question_layout("question");
+    let AdmittedDecision {
+        active_decision_request,
+        mut decision_result_rx,
+        decision_stop_tx: _decision_stop_tx,
+    } = AdmittedDecision::admit(
+        &capacity_ledger,
+        kev_decision_model
+            .runtime
+            .inference_runtime_context
+            .slot_aggregated_status
+            .clone(),
+        layout,
+    )
+    .expect("the decision must be admitted");
+    let agent_shutdown = CancellationToken::new();
+    let (_scheduler_message_tx, scheduler_message_rx) = unbounded_channel();
+
+    agent_shutdown.cancel();
+
+    kev_decision_model.run_scheduler(DecisionSchedulerRun {
+        active_requests: VecDeque::from([active_decision_request]),
+        agent_shutdown,
+        capacity_ledger,
+        pending_requests: VecDeque::new(),
+        scheduler_inbox: SchedulerInbox::new(scheduler_message_rx),
+        sequences: DECISION_SEQUENCES,
+    });
+
+    assert_eq!(
+        decision_result_rx.try_recv(),
+        Ok(DecisionResult::SchedulerUnavailable(
+            DecisionError::SchedulerUnavailable.to_string()
+        ))
+    );
+}
diff --git a/paddler_tests/tests/a_decision_state_llama_cpp_rejects_fails_the_decision.rs b/paddler_tests/tests/a_decision_state_llama_cpp_rejects_fails_the_decision.rs
new file mode 100644
index 000000000..a9553f807
--- /dev/null
+++ b/paddler_tests/tests/a_decision_state_llama_cpp_rejects_fails_the_decision.rs
@@ -0,0 +1,61 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use llama_cpp_bindings::error::DecodeError;
+use llama_cpp_bindings::llama_batch::LlamaBatch;
+use llama_cpp_bindings::token::LlamaToken;
+
+use paddler_agent_decision::decision_capacity_ledger::DecisionCapacityLedger;
+use paddler_agent_decision::decision_decoder::DecisionDecoder;
+use paddler_agent_decision::decision_error::DecisionError;
+use paddler_agent_decision::decision_token_layout::DecisionTokenLayout;
+use paddler_tests::admitted_decision::AdmittedDecision;
+use paddler_tests::two_option_question::two_option_question;
+
+use crate::kev_decision_model::KevDecisionModel;
+
+#[test]
+fn a_decision_state_llama_cpp_rejects_fails_the_decision() {
+    let KevDecisionModel {
+        delimiter_tokens,
+        pointer_head,
+        runtime,
+    } = KevDecisionModel::load(1);
+    let mut llama_context = runtime
+        .create_decision_llama_context(2)
+        .expect("the decision context must be created");
+    let n_batch = runtime.model_runtime_parameters.n_batch.tokens_usize();
+    let mut batch = LlamaBatch::new(runtime.model_runtime_parameters.n_batch.capacity(), 1);
+    let layout = DecisionTokenLayout::new(
+        &delimiter_tokens,
+        vec![LlamaToken::new(runtime.loaded_llama_model.model.n_vocab())],
+        Vec::new(),
+        two_option_question("question"),
+    );
+    let mut admitted_decision = AdmittedDecision::admit(
+        &DecisionCapacityLedger::new(2),
+        runtime
+            .inference_runtime_context
+            .slot_aggregated_status
+            .clone(),
+        layout,
+    )
+    .expect("the decision must be admitted");
+
+    assert_eq!(
+        admitted_decision
+            .active_decision_request
+            .advance(
+                &mut DecisionDecoder {
+                    batch: &mut batch,
+                    llama_context: &mut llama_context,
+                    n_batch,
+                },
+                &pointer_head,
+                None,
+            )
+            .err()
+            .as_ref()
+            .map(ToString::to_string),
+        Some(DecisionError::DecodeFailed(DecodeError::BatchInvalid).to_string())
+    );
+}
diff --git a/paddler_tests/tests/a_decision_state_overflowing_its_batch_fails_as_batch_assembly.rs b/paddler_tests/tests/a_decision_state_overflowing_its_batch_fails_as_batch_assembly.rs
new file mode 100644
index 000000000..e02b6757f
--- /dev/null
+++ b/paddler_tests/tests/a_decision_state_overflowing_its_batch_fails_as_batch_assembly.rs
@@ -0,0 +1,74 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::num::NonZeroU16;
+
+use llama_cpp_bindings::batch_add_error::BatchAddError;
+use llama_cpp_bindings::llama_batch::LlamaBatch;
+use llama_cpp_bindings::token::LlamaToken;
+use llama_cpp_bindings_types::BatchCapacity;
+
+use paddler_agent_decision::decision_capacity_ledger::DecisionCapacityLedger;
+use paddler_agent_decision::decision_decoder::DecisionDecoder;
+use paddler_agent_decision::decision_error::DecisionError;
+use paddler_agent_decision::decision_token_layout::DecisionTokenLayout;
+use paddler_tests::admitted_decision::AdmittedDecision;
+use paddler_tests::two_option_question::two_option_question;
+
+use crate::kev_decision_model::KevDecisionModel;
+
+const TOKENS_THE_BATCH_HOLDS: NonZeroU16 = NonZeroU16::MIN;
+
+#[test]
+fn a_decision_state_overflowing_its_batch_fails_as_batch_assembly() {
+    let KevDecisionModel {
+        delimiter_tokens,
+        pointer_head,
+        runtime,
+    } = KevDecisionModel::load(1);
+    let mut llama_context = runtime
+        .create_decision_llama_context(1)
+        .expect("the decision context must be created");
+    let n_batch = runtime.model_runtime_parameters.n_batch.tokens_usize();
+    let mut batch = LlamaBatch::new(BatchCapacity::from(TOKENS_THE_BATCH_HOLDS), 1);
+    let layout = DecisionTokenLayout::new(
+        &delimiter_tokens,
+        vec![LlamaToken::new(0)],
+        Vec::new(),
+        two_option_question("question"),
+    );
+    let AdmittedDecision {
+        mut active_decision_request,
+        decision_result_rx: _decision_result_rx,
+        decision_stop_tx: _decision_stop_tx,
+    } = AdmittedDecision::admit(
+        &DecisionCapacityLedger::new(1),
+        runtime
+            .inference_runtime_context
+            .slot_aggregated_status
+            .clone(),
+        layout,
+    )
+    .expect("the decision must be admitted");
+
+    assert_eq!(
+        active_decision_request
+            .advance(
+                &mut DecisionDecoder {
+                    batch: &mut batch,
+                    llama_context: &mut llama_context,
+                    n_batch,
+                },
+                &pointer_head,
+                None,
+            )
+            .err()
+            .as_ref()
+            .map(ToString::to_string),
+        Some(
+            DecisionError::BatchAssemblyFailed(BatchAddError::InsufficientSpace(usize::from(
+                TOKENS_THE_BATCH_HOLDS.get()
+            )))
+            .to_string()
+        )
+    );
+}
diff --git a/paddler_tests/tests/a_decision_stopped_before_admission_is_acknowledged.rs b/paddler_tests/tests/a_decision_stopped_before_admission_is_acknowledged.rs
new file mode 100644
index 000000000..3c473a80f
--- /dev/null
+++ b/paddler_tests/tests/a_decision_stopped_before_admission_is_acknowledged.rs
@@ -0,0 +1,29 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use tokio::sync::mpsc;
+
+use paddler_messaging::decision_result::DecisionResult;
+use paddler_tests::sample_decision::sample_decision;
+
+use crate::serving_decision_agent::ServingDecisionAgent;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn a_decision_stopped_before_admission_is_acknowledged() {
+    let serving_decision_agent = ServingDecisionAgent::start().await;
+    let (decision_stop_tx, decision_stop_rx) = mpsc::unbounded_channel();
+
+    decision_stop_tx
+        .send(())
+        .expect("the stop must reach the decision");
+
+    let mut decision_result_rx =
+        serving_decision_agent.forward_decision(sample_decision(), decision_stop_rx);
+
+    assert_eq!(
+        decision_result_rx.recv().await,
+        Some(DecisionResult::StopRequested)
+    );
+    assert_eq!(decision_result_rx.recv().await, None);
+
+    serving_decision_agent.shut_down().await;
+}
diff --git a/paddler_tests/tests/a_decision_stopped_mid_decision_is_acknowledged.rs b/paddler_tests/tests/a_decision_stopped_mid_decision_is_acknowledged.rs
new file mode 100644
index 000000000..ec8c19dff
--- /dev/null
+++ b/paddler_tests/tests/a_decision_stopped_mid_decision_is_acknowledged.rs
@@ -0,0 +1,46 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::mem::discriminant;
+
+use tokio::sync::mpsc;
+
+use paddler_messaging::decision_result::DecisionResult;
+use paddler_tests::many_question_decision::many_question_decision;
+use paddler_tests::question_answered_discriminant::question_answered_discriminant;
+
+use crate::serving_decision_agent::ServingDecisionAgent;
+
+const QUESTIONS_OUTLASTING_A_STOP: usize = 96;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn a_decision_stopped_mid_decision_is_acknowledged() {
+    let serving_decision_agent = ServingDecisionAgent::start().await;
+    let (decision_stop_tx, decision_stop_rx) = mpsc::unbounded_channel();
+    let mut decision_result_rx = serving_decision_agent.forward_decision(
+        many_question_decision(QUESTIONS_OUTLASTING_A_STOP),
+        decision_stop_rx,
+    );
+
+    assert_eq!(
+        decision_result_rx.recv().await.as_ref().map(discriminant),
+        Some(question_answered_discriminant())
+    );
+
+    decision_stop_tx
+        .send(())
+        .expect("the stop must reach the decision");
+
+    let mut results_after_the_stop = Vec::new();
+
+    while let Some(decision_result) = decision_result_rx.recv().await {
+        results_after_the_stop.push(decision_result);
+    }
+
+    assert_eq!(
+        results_after_the_stop.last(),
+        Some(&DecisionResult::StopRequested)
+    );
+    assert!(results_after_the_stop.len() < QUESTIONS_OUTLASTING_A_STOP);
+
+    serving_decision_agent.shut_down().await;
+}
diff --git a/paddler_tests/tests/a_decision_tokenizer_keeps_its_llama_backend_alive.rs b/paddler_tests/tests/a_decision_tokenizer_keeps_its_llama_backend_alive.rs
new file mode 100644
index 000000000..3b9313208
--- /dev/null
+++ b/paddler_tests/tests/a_decision_tokenizer_keeps_its_llama_backend_alive.rs
@@ -0,0 +1,38 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::sync::Arc;
+
+use paddler_agent_decision::decision_text_tokenizer::DecisionTextTokenizer;
+use paddler_agent_pointer_head::pointer_head_input_normalization::PointerHeadInputNormalization;
+use paddler_agent_runtime::llama_model_settings::LlamaModelSettings;
+use paddler_agent_runtime::loaded_llama_model::LoadedLlamaModel;
+use paddler_inference_parameters::all_gpu_layers::ALL_GPU_LAYERS;
+use paddler_tests::agent_inference_runtime_context::agent_inference_runtime_context;
+use paddler_tests::kev_checkpoint::KevCheckpoint;
+
+#[test]
+fn a_decision_tokenizer_keeps_its_llama_backend_alive() {
+    let inference_runtime_context = agent_inference_runtime_context(1);
+    let loaded_llama_model = LoadedLlamaModel::load(
+        &inference_runtime_context,
+        LlamaModelSettings {
+            model_path: KevCheckpoint::Kev0_8b.target_path("model.gguf"),
+            n_gpu_layers: ALL_GPU_LAYERS,
+        },
+    )
+    .expect("the decision model must load");
+    let llama_backend = Arc::downgrade(&loaded_llama_model.model.llama_backend);
+    let text_tokenizer = DecisionTextTokenizer {
+        input_normalization: PointerHeadInputNormalization::Unnormalized,
+        model: loaded_llama_model.model.clone(),
+    };
+
+    drop(loaded_llama_model);
+    drop(inference_runtime_context);
+
+    assert_eq!(llama_backend.strong_count(), 1);
+
+    drop(text_tokenizer);
+
+    assert_eq!(llama_backend.strong_count(), 0);
+}
diff --git a/paddler_tests/tests/a_decision_waiting_when_the_agent_shuts_down_is_refused_as_unavailable.rs b/paddler_tests/tests/a_decision_waiting_when_the_agent_shuts_down_is_refused_as_unavailable.rs
new file mode 100644
index 000000000..70c1bb269
--- /dev/null
+++ b/paddler_tests/tests/a_decision_waiting_when_the_agent_shuts_down_is_refused_as_unavailable.rs
@@ -0,0 +1,58 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::collections::VecDeque;
+
+use tokio::sync::mpsc::unbounded_channel;
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent_decision::decision_capacity_ledger::DecisionCapacityLedger;
+use paddler_agent_decision::decision_error::DecisionError;
+use paddler_agent_runtime::scheduler_inbox::SchedulerInbox;
+use paddler_agent_runtime::scheduler_message::SchedulerMessage;
+use paddler_messaging::decision_result::DecisionResult;
+use paddler_tests::decision_scheduler_run::DecisionSchedulerRun;
+use paddler_tests::prepared_decision::PreparedDecision;
+
+use crate::kev_decision_model::KevDecisionModel;
+
+const DECISION_SEQUENCES: u16 = 2;
+
+#[test]
+fn a_decision_waiting_when_the_agent_shuts_down_is_refused_as_unavailable() {
+    let kev_decision_model = KevDecisionModel::load(DECISION_SEQUENCES);
+    let PreparedDecision {
+        mut decision_result_rx,
+        decision_stop_tx: _decision_stop_tx,
+        prepared_decision_request,
+    } = PreparedDecision::new(
+        kev_decision_model
+            .runtime
+            .inference_runtime_context
+            .slot_aggregated_status
+            .clone(),
+        kev_decision_model.two_option_question_layout("question"),
+    );
+    let agent_shutdown = CancellationToken::new();
+    let (scheduler_message_tx, scheduler_message_rx) = unbounded_channel();
+
+    scheduler_message_tx
+        .send(SchedulerMessage::Command(prepared_decision_request))
+        .expect("the decision must wait for the scheduler");
+    agent_shutdown.cancel();
+
+    kev_decision_model.run_scheduler(DecisionSchedulerRun {
+        active_requests: VecDeque::new(),
+        agent_shutdown,
+        capacity_ledger: DecisionCapacityLedger::new(DECISION_SEQUENCES),
+        pending_requests: VecDeque::new(),
+        scheduler_inbox: SchedulerInbox::new(scheduler_message_rx),
+        sequences: DECISION_SEQUENCES,
+    });
+
+    assert_eq!(
+        decision_result_rx.try_recv(),
+        Ok(DecisionResult::SchedulerUnavailable(
+            DecisionError::SchedulerUnavailable.to_string()
+        ))
+    );
+}
diff --git a/paddler_tests/tests/a_decision_waits_until_an_agent_loads_the_pointer_head.rs b/paddler_tests/tests/a_decision_waits_until_an_agent_loads_the_pointer_head.rs
new file mode 100644
index 000000000..8e5e8c769
--- /dev/null
+++ b/paddler_tests/tests/a_decision_waits_until_an_agent_loads_the_pointer_head.rs
@@ -0,0 +1,101 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::mem::discriminant;
+use std::path::Path;
+
+use tokio::spawn;
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::agent_decisions_serving::AgentDecisionsServing;
+use paddler_messaging::agent_issue::AgentIssue;
+use paddler_messaging::agent_issue_params::model_path::ModelPath;
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_messaging::decision_settings::DecisionSettings;
+use paddler_test_cluster_harness::agent_desired_model_at::agent_desired_model_at;
+use paddler_tests::decision_cluster_params::DecisionClusterParams;
+use paddler_tests::decision_done_discriminant::decision_done_discriminant;
+use paddler_tests::sample_decision::sample_decision;
+use paddler_tests::start_decision_cluster::start_decision_cluster;
+
+const MISSING_POINTER_HEAD_PATH: &str = "/nonexistent/pointer_head.gguf";
+
+#[tokio::test(flavor = "multi_thread")]
+async fn a_decision_waits_until_an_agent_loads_the_pointer_head() {
+    let mut cluster = start_decision_cluster(DecisionClusterParams::default())
+        .await
+        .expect("a decision cluster must start");
+
+    let decisions_desired_state = cluster
+        .client_management
+        .get_balancer_desired_state(CancellationToken::new())
+        .await
+        .expect("the balancer must report its desired state");
+
+    cluster
+        .client_management
+        .put_balancer_desired_state(
+            CancellationToken::new(),
+            &BalancerDesiredState {
+                decision: DecisionSettings::default(),
+                ..decisions_desired_state.clone()
+            },
+        )
+        .await
+        .expect("the balancer must stop serving decisions");
+    cluster
+        .wait_for_first_agent_to_serve(AgentDecisionsServing::NotServed)
+        .await
+        .expect("the agent must stop serving decisions");
+    cluster
+        .client_management
+        .put_balancer_desired_state(
+            CancellationToken::new(),
+            &BalancerDesiredState {
+                decision: DecisionSettings {
+                    pointer_head: agent_desired_model_at(Path::new(MISSING_POINTER_HEAD_PATH))
+                        .expect("the missing pointer head path must be absolute"),
+                },
+                ..decisions_desired_state.clone()
+            },
+        )
+        .await
+        .expect("the balancer must serve decisions again");
+    cluster
+        .wait_for_first_agent_issue(|issue| {
+            *issue
+                == AgentIssue::PointerHeadCannotBeLoaded(ModelPath {
+                    model_path: MISSING_POINTER_HEAD_PATH.to_owned(),
+                })
+        })
+        .await
+        .expect("the agent must report the pointer head it cannot load");
+
+    let decision = spawn(cluster.decide(CancellationToken::new(), &sample_decision()));
+
+    cluster
+        .wait_for_buffered_request_count(1)
+        .await
+        .expect("the decision must wait in the buffer");
+
+    cluster
+        .client_management
+        .put_balancer_desired_state(CancellationToken::new(), &decisions_desired_state)
+        .await
+        .expect("the balancer must accept a pointer head the agent can load");
+
+    assert_eq!(
+        discriminant(
+            &decision
+                .await
+                .expect("the decision task must not panic")
+                .expect("the buffered decision must be answered")
+                .terminal_result
+        ),
+        decision_done_discriminant()
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/a_decision_whose_client_stopped_listening_cannot_send_its_answer.rs b/paddler_tests/tests/a_decision_whose_client_stopped_listening_cannot_send_its_answer.rs
new file mode 100644
index 000000000..562f8ed1b
--- /dev/null
+++ b/paddler_tests/tests/a_decision_whose_client_stopped_listening_cannot_send_its_answer.rs
@@ -0,0 +1,81 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use llama_cpp_bindings::llama_batch::LlamaBatch;
+use llama_cpp_bindings::token::LlamaToken;
+use tokio::sync::mpsc::error::SendError;
+
+use paddler_agent_decision::decision_capacity_ledger::DecisionCapacityLedger;
+use paddler_agent_decision::decision_decoder::DecisionDecoder;
+use paddler_agent_decision::decision_error::DecisionError;
+use paddler_agent_decision::decision_token_layout::DecisionTokenLayout;
+use paddler_messaging::decision_result::DecisionResult;
+use paddler_tests::admitted_decision::AdmittedDecision;
+use paddler_tests::two_option_question::two_option_question;
+
+use crate::kev_decision_model::KevDecisionModel;
+
+#[test]
+fn a_decision_whose_client_stopped_listening_cannot_send_its_answer() {
+    let KevDecisionModel {
+        delimiter_tokens,
+        pointer_head,
+        runtime,
+    } = KevDecisionModel::load(1);
+    let mut llama_context = runtime
+        .create_decision_llama_context(1)
+        .expect("the decision context must be created");
+    let n_batch = runtime.model_runtime_parameters.n_batch.tokens_usize();
+    let mut batch = LlamaBatch::new(runtime.model_runtime_parameters.n_batch.capacity(), 1);
+    let layout = DecisionTokenLayout::new(
+        &delimiter_tokens,
+        vec![LlamaToken::new(0)],
+        Vec::new(),
+        two_option_question("question"),
+    );
+    let AdmittedDecision {
+        mut active_decision_request,
+        decision_result_rx,
+        decision_stop_tx: _decision_stop_tx,
+    } = AdmittedDecision::admit(
+        &DecisionCapacityLedger::new(1),
+        runtime
+            .inference_runtime_context
+            .slot_aggregated_status
+            .clone(),
+        layout,
+    )
+    .expect("the decision must be admitted");
+
+    drop(decision_result_rx);
+
+    active_decision_request
+        .advance(
+            &mut DecisionDecoder {
+                batch: &mut batch,
+                llama_context: &mut llama_context,
+                n_batch,
+            },
+            &pointer_head,
+            None,
+        )
+        .expect("the decision must advance");
+
+    assert_eq!(
+        active_decision_request
+            .advance(
+                &mut DecisionDecoder {
+                    batch: &mut batch,
+                    llama_context: &mut llama_context,
+                    n_batch,
+                },
+                &pointer_head,
+                None,
+            )
+            .err()
+            .as_ref()
+            .map(ToString::to_string),
+        Some(
+            DecisionError::ClientDisconnected(SendError(DecisionResult::StopRequested)).to_string()
+        )
+    );
+}
diff --git a/paddler_tests/tests/a_decision_whose_question_lane_cannot_be_released_still_reports_its_outcome.rs b/paddler_tests/tests/a_decision_whose_question_lane_cannot_be_released_still_reports_its_outcome.rs
new file mode 100644
index 000000000..426640d95
--- /dev/null
+++ b/paddler_tests/tests/a_decision_whose_question_lane_cannot_be_released_still_reports_its_outcome.rs
@@ -0,0 +1,64 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use llama_cpp_bindings::llama_batch::LlamaBatch;
+use llama_cpp_bindings::token::LlamaToken;
+
+use paddler_agent_decision::decision_capacity_ledger::DecisionCapacityLedger;
+use paddler_agent_decision::decision_decoder::DecisionDecoder;
+use paddler_agent_decision::decision_token_layout::DecisionTokenLayout;
+use paddler_messaging::decision_result::DecisionResult;
+use paddler_tests::admitted_decision::AdmittedDecision;
+use paddler_tests::two_option_question::two_option_question;
+
+use crate::kev_decision_model::KevDecisionModel;
+
+const SEQUENCES_THE_CONTEXT_HOLDS: u32 = 1;
+const SEQUENCES_THE_LEDGER_HANDS_OUT: u16 = 2;
+
+#[test]
+fn a_decision_whose_question_lane_cannot_be_released_still_reports_its_outcome() {
+    let KevDecisionModel {
+        delimiter_tokens,
+        runtime,
+        ..
+    } = KevDecisionModel::load(1);
+    let mut llama_context = runtime
+        .create_decision_llama_context(SEQUENCES_THE_CONTEXT_HOLDS)
+        .expect("the decision context must be created");
+    let n_batch = runtime.model_runtime_parameters.n_batch.tokens_usize();
+    let mut batch = LlamaBatch::new(runtime.model_runtime_parameters.n_batch.capacity(), 1);
+    let layout = DecisionTokenLayout::new(
+        &delimiter_tokens,
+        vec![LlamaToken::new(0)],
+        vec![two_option_question("leading")],
+        two_option_question("last"),
+    );
+    let AdmittedDecision {
+        active_decision_request,
+        mut decision_result_rx,
+        decision_stop_tx: _decision_stop_tx,
+    } = AdmittedDecision::admit(
+        &DecisionCapacityLedger::new(SEQUENCES_THE_LEDGER_HANDS_OUT),
+        runtime
+            .inference_runtime_context
+            .slot_aggregated_status
+            .clone(),
+        layout,
+    )
+    .expect("the decision must be admitted");
+
+    active_decision_request.finish_with(
+        &mut DecisionDecoder {
+            batch: &mut batch,
+            llama_context: &mut llama_context,
+            n_batch,
+        },
+        None,
+        DecisionResult::StopRequested,
+    );
+
+    assert_eq!(
+        decision_result_rx.try_recv(),
+        Ok(DecisionResult::StopRequested)
+    );
+}
diff --git a/paddler_tests/tests/a_decision_whose_state_sequence_cannot_be_released_still_reports_its_outcome.rs b/paddler_tests/tests/a_decision_whose_state_sequence_cannot_be_released_still_reports_its_outcome.rs
new file mode 100644
index 000000000..3a20d9eee
--- /dev/null
+++ b/paddler_tests/tests/a_decision_whose_state_sequence_cannot_be_released_still_reports_its_outcome.rs
@@ -0,0 +1,74 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use llama_cpp_bindings::llama_batch::LlamaBatch;
+use llama_cpp_bindings::token::LlamaToken;
+
+use paddler_agent_decision::decision_capacity_ledger::DecisionCapacityLedger;
+use paddler_agent_decision::decision_decoder::DecisionDecoder;
+use paddler_agent_decision::decision_token_layout::DecisionTokenLayout;
+use paddler_messaging::decision_result::DecisionResult;
+use paddler_tests::admitted_decision::AdmittedDecision;
+use paddler_tests::two_option_question::two_option_question;
+
+use crate::kev_decision_model::KevDecisionModel;
+
+const SEQUENCES_THE_CONTEXT_HOLDS: u32 = 1;
+const SEQUENCES_THE_LEDGER_HANDS_OUT: u16 = 4;
+
+#[test]
+fn a_decision_whose_state_sequence_cannot_be_released_still_reports_its_outcome() {
+    let KevDecisionModel {
+        delimiter_tokens,
+        runtime,
+        ..
+    } = KevDecisionModel::load(1);
+    let mut llama_context = runtime
+        .create_decision_llama_context(SEQUENCES_THE_CONTEXT_HOLDS)
+        .expect("the decision context must be created");
+    let n_batch = runtime.model_runtime_parameters.n_batch.tokens_usize();
+    let mut batch = LlamaBatch::new(runtime.model_runtime_parameters.n_batch.capacity(), 1);
+    let layout = DecisionTokenLayout::new(
+        &delimiter_tokens,
+        vec![LlamaToken::new(0)],
+        vec![two_option_question("leading")],
+        two_option_question("last"),
+    );
+    let capacity_ledger = DecisionCapacityLedger::new(SEQUENCES_THE_LEDGER_HANDS_OUT);
+    let _decision_holding_the_sequences_the_context_holds = AdmittedDecision::admit(
+        &capacity_ledger,
+        runtime
+            .inference_runtime_context
+            .slot_aggregated_status
+            .clone(),
+        layout.clone(),
+    )
+    .expect("the first decision must be admitted");
+    let AdmittedDecision {
+        active_decision_request,
+        mut decision_result_rx,
+        decision_stop_tx: _decision_stop_tx,
+    } = AdmittedDecision::admit(
+        &capacity_ledger,
+        runtime
+            .inference_runtime_context
+            .slot_aggregated_status
+            .clone(),
+        layout,
+    )
+    .expect("the second decision must be admitted");
+
+    active_decision_request.finish_with(
+        &mut DecisionDecoder {
+            batch: &mut batch,
+            llama_context: &mut llama_context,
+            n_batch,
+        },
+        None,
+        DecisionResult::StopRequested,
+    );
+
+    assert_eq!(
+        decision_result_rx.try_recv(),
+        Ok(DecisionResult::StopRequested)
+    );
+}
diff --git a/paddler_tests/tests/a_decision_with_a_leading_question_reserving_one_slot_answers_like_one_reserving_two.rs b/paddler_tests/tests/a_decision_with_a_leading_question_reserving_one_slot_answers_like_one_reserving_two.rs
new file mode 100644
index 000000000..58b0203d5
--- /dev/null
+++ b/paddler_tests/tests/a_decision_with_a_leading_question_reserving_one_slot_answers_like_one_reserving_two.rs
@@ -0,0 +1,92 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::collections::VecDeque;
+use std::num::NonZeroU16;
+
+use llama_cpp_bindings::token::LlamaToken;
+use tokio::sync::mpsc::UnboundedReceiver;
+use tokio::sync::mpsc::unbounded_channel;
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent_decision::decision_capacity_ledger::DecisionCapacityLedger;
+use paddler_agent_decision::decision_token_layout::DecisionTokenLayout;
+use paddler_agent_runtime::scheduler_inbox::SchedulerInbox;
+use paddler_agent_runtime::scheduler_message::SchedulerMessage;
+use paddler_agent_status::slot_guard::SlotGuard;
+use paddler_messaging::decision_answer::DecisionAnswer;
+use paddler_messaging::decision_result::DecisionResult;
+use paddler_messaging::decision_slots_minimum::DECISION_SLOTS_MINIMUM;
+use paddler_tests::decision_scheduler_run::DecisionSchedulerRun;
+use paddler_tests::prepared_decision::PreparedDecision;
+use paddler_tests::two_option_question::two_option_question;
+
+use crate::kev_decision_model::KevDecisionModel;
+
+fn answers_of(mut decision_result_rx: UnboundedReceiver<DecisionResult>) -> Vec<DecisionAnswer> {
+    let mut answers = Vec::new();
+
+    while let Ok(decision_result) = decision_result_rx.try_recv() {
+        if let DecisionResult::QuestionAnswered(answer) = decision_result {
+            answers.push(answer);
+        }
+    }
+
+    answers
+}
+
+#[test]
+fn a_decision_with_a_leading_question_reserving_one_slot_answers_like_one_reserving_two() {
+    let kev_decision_model = KevDecisionModel::load(DECISION_SLOTS_MINIMUM.get());
+    let slot_aggregated_status = kev_decision_model
+        .runtime
+        .inference_runtime_context
+        .slot_aggregated_status
+        .clone();
+    let layout_with_a_leading_question = || {
+        DecisionTokenLayout::new(
+            &kev_decision_model.delimiter_tokens,
+            vec![LlamaToken::new(0)],
+            vec![two_option_question("leading")],
+            two_option_question("last"),
+        )
+    };
+    let decision_reserving_two_slots = PreparedDecision::new(
+        slot_aggregated_status.clone(),
+        layout_with_a_leading_question(),
+    );
+    let mut decision_reserving_one_slot = PreparedDecision::new(
+        slot_aggregated_status.clone(),
+        layout_with_a_leading_question(),
+    );
+
+    decision_reserving_one_slot
+        .prepared_decision_request
+        .slot_guard = SlotGuard::new(slot_aggregated_status, NonZeroU16::MIN);
+
+    let (scheduler_message_tx, scheduler_message_rx) = unbounded_channel();
+
+    for prepared_decision_request in [
+        decision_reserving_two_slots.prepared_decision_request,
+        decision_reserving_one_slot.prepared_decision_request,
+    ] {
+        scheduler_message_tx
+            .send(SchedulerMessage::Command(prepared_decision_request))
+            .expect("the scheduler must receive the decision");
+    }
+
+    drop(scheduler_message_tx);
+
+    kev_decision_model.run_scheduler(DecisionSchedulerRun {
+        active_requests: VecDeque::new(),
+        agent_shutdown: CancellationToken::new(),
+        capacity_ledger: DecisionCapacityLedger::new(DECISION_SLOTS_MINIMUM.get()),
+        pending_requests: VecDeque::new(),
+        scheduler_inbox: SchedulerInbox::new(scheduler_message_rx),
+        sequences: DECISION_SLOTS_MINIMUM.get(),
+    });
+
+    assert_eq!(
+        answers_of(decision_reserving_one_slot.decision_result_rx),
+        answers_of(decision_reserving_two_slots.decision_result_rx)
+    );
+}
diff --git a/paddler_tests/tests/a_decision_with_leading_questions_waits_for_two_free_slots.rs b/paddler_tests/tests/a_decision_with_leading_questions_waits_for_two_free_slots.rs
new file mode 100644
index 000000000..2545738e9
--- /dev/null
+++ b/paddler_tests/tests/a_decision_with_leading_questions_waits_for_two_free_slots.rs
@@ -0,0 +1,45 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::mem::discriminant;
+
+use tokio::spawn;
+use tokio_util::sync::CancellationToken;
+
+use paddler_tests::decision_cluster_params::DecisionClusterParams;
+use paddler_tests::decision_done_discriminant::decision_done_discriminant;
+use paddler_tests::sample_decision::sample_decision;
+use paddler_tests::start_decision_cluster::start_decision_cluster;
+
+use crate::slot_held_by_an_unending_generation::SlotHeldByAnUnendingGeneration;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn a_decision_with_leading_questions_waits_for_two_free_slots() {
+    let mut cluster = start_decision_cluster(DecisionClusterParams::default())
+        .await
+        .expect("a decision cluster must start");
+    let slot_held_by_an_unending_generation =
+        SlotHeldByAnUnendingGeneration::hold(&mut cluster).await;
+    let decision = spawn(cluster.decide(CancellationToken::new(), &sample_decision()));
+
+    cluster
+        .wait_for_buffered_request_count(1)
+        .await
+        .expect("the decision must wait while only one slot is free");
+    slot_held_by_an_unending_generation.release().await;
+
+    assert_eq!(
+        discriminant(
+            &decision
+                .await
+                .expect("the decision task must not panic")
+                .expect("the decision must be answered")
+                .terminal_result
+        ),
+        decision_done_discriminant()
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/a_desired_state_with_a_missing_model_is_left_pending_without_starting_a_pipeline.rs b/paddler_tests/tests/a_desired_state_with_a_missing_model_is_left_pending_without_starting_a_pipeline.rs
new file mode 100644
index 000000000..d320faec8
--- /dev/null
+++ b/paddler_tests/tests/a_desired_state_with_a_missing_model_is_left_pending_without_starting_a_pipeline.rs
@@ -0,0 +1,42 @@
+use std::path::Path;
+
+use paddler_agent::desired_state_reconciliation::DesiredStateReconciliation;
+use paddler_messaging::agent_desired_state::AgentDesiredState;
+use paddler_messaging::agent_runtime_status::AgentRuntimeStatus;
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_messaging::produces_snapshot::ProducesSnapshot as _;
+use paddler_model_source::agent_local_model_path::AgentLocalModelPath;
+use paddler_model_source::model_source::ModelSource;
+
+use crate::serving_pipeline_arbiter::ServingPipelineArbiter;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn a_desired_state_with_a_missing_model_is_left_pending_without_starting_a_pipeline() {
+    let mut serving_pipeline_arbiter = ServingPipelineArbiter::new(1);
+    let desired_state = BalancerDesiredState {
+        model: ModelSource::LocalToAgent(
+            AgentLocalModelPath::from_absolute_path(Path::new(
+                "/paddler-nonexistent-model-for-the-serving-pipeline.gguf",
+            ))
+            .expect("the model path must be absolute"),
+        )
+        .into_agent_desired_model(),
+        ..BalancerDesiredState::default()
+    };
+
+    assert_eq!(
+        serving_pipeline_arbiter.apply(desired_state.clone()).await,
+        DesiredStateReconciliation::Pending(Box::new(AgentDesiredState::from(&desired_state)))
+    );
+    assert_eq!(
+        serving_pipeline_arbiter
+            .inference_runtime_context
+            .slot_aggregated_status
+            .make_snapshot()
+            .status
+            .runtime,
+        AgentRuntimeStatus::Idle
+    );
+
+    serving_pipeline_arbiter.shut_down().await;
+}
diff --git a/paddler_tests/tests/a_failed_attempt_retries_without_reloading_its_model.rs b/paddler_tests/tests/a_failed_attempt_retries_without_reloading_its_model.rs
new file mode 100644
index 000000000..f0eeddf3c
--- /dev/null
+++ b/paddler_tests/tests/a_failed_attempt_retries_without_reloading_its_model.rs
@@ -0,0 +1,77 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::collections::BTreeSet;
+use std::mem::discriminant;
+
+use paddler_messaging::agent_issue::AgentIssue;
+use paddler_messaging::agent_issue_params::chat_template_does_not_compile_params::ChatTemplateDoesNotCompileParams;
+use paddler_messaging::agent_issue_params::model_path::ModelPath;
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_messaging::decision_settings::DecisionSettings;
+use paddler_messaging::produces_snapshot::ProducesSnapshot as _;
+use paddler_tests::desired_state_reconciled_discriminant::desired_state_reconciled_discriminant;
+use paddler_tests::desired_state_with_chat_template_override::desired_state_with_chat_template_override;
+use paddler_tests::kev_checkpoint::KevCheckpoint;
+use paddler_tests::uncompilable_chat_template::UNCOMPILABLE_CHAT_TEMPLATE_CONTENT;
+use paddler_tests::uncompilable_chat_template::uncompilable_chat_template;
+
+use crate::serving_pipeline_arbiter::ServingPipelineArbiter;
+use crate::swappable_model_file::SwappableModelFile;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn a_failed_attempt_retries_without_reloading_its_model() {
+    let swappable_model_file =
+        SwappableModelFile::link_to(&KevCheckpoint::Kev0_8b.target_path("model.gguf"));
+    let mut serving_pipeline_arbiter = ServingPipelineArbiter::new(1);
+
+    assert_eq!(
+        discriminant(
+            &serving_pipeline_arbiter
+                .apply(desired_state_with_chat_template_override(
+                    BalancerDesiredState {
+                        decision: DecisionSettings::default(),
+                        model: swappable_model_file.agent_desired_model(),
+                        ..KevCheckpoint::Kev0_8b
+                            .decision_desired_state()
+                            .into_balancer_desired_state()
+                            .expect("the kev desired state must name absolute paths")
+                    },
+                    uncompilable_chat_template(),
+                ))
+                .await
+        ),
+        desired_state_reconciled_discriminant()
+    );
+
+    serving_pipeline_arbiter.conclude_next_progress().await;
+
+    let failed_attempt = serving_pipeline_arbiter
+        .pipeline_arbiter
+        .next_finished_startup()
+        .await;
+
+    swappable_model_file.replace_with_an_unloadable_file();
+    serving_pipeline_arbiter
+        .pipeline_arbiter
+        .conclude(failed_attempt);
+    serving_pipeline_arbiter.conclude_next_progress().await;
+    serving_pipeline_arbiter.conclude_next_progress().await;
+
+    assert_eq!(
+        serving_pipeline_arbiter
+            .inference_runtime_context
+            .slot_aggregated_status
+            .make_snapshot()
+            .status
+            .issues,
+        BTreeSet::from([AgentIssue::ChatTemplateDoesNotCompile(
+            ChatTemplateDoesNotCompileParams {
+                error: "syntax error: unexpected identifier, expected end of variable block (in chat_template:1)".to_owned(),
+                model_path: ModelPath::from(swappable_model_file.model_path().as_path()),
+                template_content: UNCOMPILABLE_CHAT_TEMPLATE_CONTENT.to_owned(),
+            }
+        )])
+    );
+
+    serving_pipeline_arbiter.shut_down().await;
+}
diff --git a/paddler_tests/tests/a_failed_startup_releases_its_model_when_the_agent_shuts_down.rs b/paddler_tests/tests/a_failed_startup_releases_its_model_when_the_agent_shuts_down.rs
new file mode 100644
index 000000000..3213ca31b
--- /dev/null
+++ b/paddler_tests/tests/a_failed_startup_releases_its_model_when_the_agent_shuts_down.rs
@@ -0,0 +1,37 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_messaging::decision_settings::DecisionSettings;
+use paddler_tests::desired_state_with_chat_template_override::desired_state_with_chat_template_override;
+use paddler_tests::kev_checkpoint::KevCheckpoint;
+use paddler_tests::uncompilable_chat_template::uncompilable_chat_template;
+
+use crate::serving_pipeline_arbiter::ServingPipelineArbiter;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn a_failed_startup_releases_its_model_when_the_agent_shuts_down() {
+    let mut serving_pipeline_arbiter = ServingPipelineArbiter::new(1);
+    let model_weights_residency = serving_pipeline_arbiter
+        .inference_runtime_context
+        .vram_residencies
+        .model_weights
+        .clone();
+
+    serving_pipeline_arbiter
+        .serve(desired_state_with_chat_template_override(
+            BalancerDesiredState {
+                decision: DecisionSettings::default(),
+                ..KevCheckpoint::Kev0_8b
+                    .decision_desired_state()
+                    .into_balancer_desired_state()
+                    .expect("the kev desired state must name absolute paths")
+            },
+            uncompilable_chat_template(),
+        ))
+        .await;
+    assert!(model_weights_residency.try_occupy().is_err());
+
+    serving_pipeline_arbiter.shut_down().await;
+
+    assert!(model_weights_residency.try_occupy().is_ok());
+}
diff --git a/paddler_tests/tests/a_generation_admitted_to_a_sequence_that_cannot_be_cleared_fails_with_kv_cache_clear.rs b/paddler_tests/tests/a_generation_admitted_to_a_sequence_that_cannot_be_cleared_fails_with_kv_cache_clear.rs
new file mode 100644
index 000000000..9caec62a6
--- /dev/null
+++ b/paddler_tests/tests/a_generation_admitted_to_a_sequence_that_cannot_be_cleared_fails_with_kv_cache_clear.rs
@@ -0,0 +1,105 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use llama_cpp_bindings::error::ClearKvCacheSeqError;
+use llama_cpp_bindings::llama_batch::LlamaBatch;
+use tokio::sync::mpsc::unbounded_channel;
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent_runtime::scheduler_message::SchedulerMessage;
+use paddler_agent_text_generation::continuous_batch_scheduler::ContinuousBatchScheduler;
+use paddler_agent_text_generation::continuous_batch_scheduler_context::ContinuousBatchSchedulerContext;
+use paddler_agent_text_generation::continuous_batch_scheduler_params::ContinuousBatchSchedulerParams;
+use paddler_agent_text_generation::text_generation_error::TextGenerationError;
+use paddler_messaging::generated_token_result::GeneratedTokenResult;
+use paddler_tests::agent_inference_runtime_context::agent_inference_runtime_context;
+use paddler_tests::kev_checkpoint::KevCheckpoint;
+use paddler_tests::llama_runtime::LlamaRuntime;
+use paddler_tests::raw_prompt_generation::RawPromptGeneration;
+
+use crate::first_message_content_request_preparer::FirstMessageContentRequestPreparer;
+
+const SEQUENCES_THE_CONTEXT_HOLDS: u32 = 1;
+const SEQUENCES_THE_SCHEDULER_HANDS_OUT: u16 = 2;
+
+#[test]
+fn a_generation_admitted_to_a_sequence_that_cannot_be_cleared_fails_with_kv_cache_clear() {
+    let runtime = LlamaRuntime::load(
+        &KevCheckpoint::Kev0_8b.target_path("model.gguf"),
+        agent_inference_runtime_context(SEQUENCES_THE_SCHEDULER_HANDS_OUT),
+    )
+    .expect("the kev model must load");
+    let llama_context = runtime
+        .create_text_generation_llama_context(SEQUENCES_THE_CONTEXT_HOLDS)
+        .expect("the context must be created");
+    let sequence_context_size = llama_context.n_ctx_seq();
+    let request_preparer = FirstMessageContentRequestPreparer {
+        llama_runtime: &runtime,
+        sequence_context_size,
+    }
+    .into_generation_request_preparer();
+    let raw_prompt_generation = || {
+        RawPromptGeneration::new(
+            "Paddler",
+            runtime
+                .inference_runtime_context
+                .slot_aggregated_status
+                .clone(),
+        )
+    };
+    let RawPromptGeneration {
+        agent_request: generation_on_a_clearable_sequence,
+        generate_tokens_stop_tx: _clearable_generate_tokens_stop_tx,
+        generated_tokens_rx: _clearable_generated_tokens_rx,
+    } = raw_prompt_generation();
+    let RawPromptGeneration {
+        agent_request: generation_on_a_sequence_beyond_the_context,
+        generate_tokens_stop_tx: _beyond_generate_tokens_stop_tx,
+        mut generated_tokens_rx,
+    } = raw_prompt_generation();
+    let (scheduler_message_tx, scheduler_message_rx) = unbounded_channel();
+
+    for agent_request in [
+        generation_on_a_clearable_sequence,
+        generation_on_a_sequence_beyond_the_context,
+    ] {
+        scheduler_message_tx
+            .send(SchedulerMessage::Command(
+                request_preparer
+                    .prepare_raw_prompt(agent_request)
+                    .expect("the generation must be prepared"),
+            ))
+            .expect("the scheduler must accept the generation");
+    }
+    scheduler_message_tx
+        .send(SchedulerMessage::Shutdown)
+        .expect("the scheduler must accept the shutdown");
+
+    ContinuousBatchScheduler::new(ContinuousBatchSchedulerParams {
+        agent_shutdown: CancellationToken::new(),
+        batch: LlamaBatch::new(runtime.model_runtime_parameters.n_batch.capacity(), 1),
+        llama_context,
+        scheduler_context: ContinuousBatchSchedulerContext {
+            agent_name: None,
+            desired_slots_total: SEQUENCES_THE_SCHEDULER_HANDS_OUT,
+            model: &runtime.loaded_llama_model.model,
+            n_batch: runtime.model_runtime_parameters.n_batch,
+            sequence_context_size,
+        },
+        scheduler_message_rx,
+    })
+    .run();
+
+    assert_eq!(
+        generated_tokens_rx.try_recv(),
+        Ok(GeneratedTokenResult::KvCacheClearFailed(
+            TextGenerationError::KvCacheClearFailed(
+                ClearKvCacheSeqError::PartialSequenceNotRemoved {
+                    seq_id: 1,
+                    p0: -1,
+                    p1: -1,
+                }
+            )
+            .to_string()
+        ))
+    );
+}
diff --git a/paddler_tests/tests/a_generation_arriving_while_every_sequence_is_taken_is_refused.rs b/paddler_tests/tests/a_generation_arriving_while_every_sequence_is_taken_is_refused.rs
new file mode 100644
index 000000000..0670e3ba8
--- /dev/null
+++ b/paddler_tests/tests/a_generation_arriving_while_every_sequence_is_taken_is_refused.rs
@@ -0,0 +1,97 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use llama_cpp_bindings::llama_batch::LlamaBatch;
+use tokio::sync::mpsc::unbounded_channel;
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent_runtime::scheduler_message::SchedulerMessage;
+use paddler_agent_text_generation::continuous_batch_scheduler::ContinuousBatchScheduler;
+use paddler_agent_text_generation::continuous_batch_scheduler_context::ContinuousBatchSchedulerContext;
+use paddler_agent_text_generation::continuous_batch_scheduler_params::ContinuousBatchSchedulerParams;
+use paddler_agent_text_generation::text_generation_error::TextGenerationError;
+use paddler_messaging::generated_token_result::GeneratedTokenResult;
+use paddler_tests::agent_inference_runtime_context::agent_inference_runtime_context;
+use paddler_tests::kev_checkpoint::KevCheckpoint;
+use paddler_tests::llama_runtime::LlamaRuntime;
+use paddler_tests::raw_prompt_generation::RawPromptGeneration;
+
+use crate::first_message_content_request_preparer::FirstMessageContentRequestPreparer;
+
+#[test]
+fn a_generation_arriving_while_every_sequence_is_taken_is_refused() {
+    let runtime = LlamaRuntime::load(
+        &KevCheckpoint::Kev0_8b.target_path("model.gguf"),
+        agent_inference_runtime_context(1),
+    )
+    .expect("the kev model must load");
+    let llama_context = runtime
+        .create_text_generation_llama_context(1)
+        .expect("the context must be created");
+    let sequence_context_size = llama_context.n_ctx_seq();
+    let request_preparer = FirstMessageContentRequestPreparer {
+        llama_runtime: &runtime,
+        sequence_context_size,
+    }
+    .into_generation_request_preparer();
+    let RawPromptGeneration {
+        agent_request: generation_taking_the_only_sequence,
+        generate_tokens_stop_tx: _first_generate_tokens_stop_tx,
+        generated_tokens_rx: _first_generated_tokens_rx,
+    } = RawPromptGeneration::new(
+        "Paddler",
+        runtime
+            .inference_runtime_context
+            .slot_aggregated_status
+            .clone(),
+    );
+    let RawPromptGeneration {
+        agent_request: generation_without_a_sequence,
+        generate_tokens_stop_tx: _second_generate_tokens_stop_tx,
+        mut generated_tokens_rx,
+    } = RawPromptGeneration::new(
+        "Paddler",
+        runtime
+            .inference_runtime_context
+            .slot_aggregated_status
+            .clone(),
+    );
+    let (scheduler_message_tx, scheduler_message_rx) = unbounded_channel();
+
+    for agent_request in [
+        generation_taking_the_only_sequence,
+        generation_without_a_sequence,
+    ] {
+        scheduler_message_tx
+            .send(SchedulerMessage::Command(
+                request_preparer
+                    .prepare_raw_prompt(agent_request)
+                    .expect("the generation must be prepared"),
+            ))
+            .expect("the scheduler must accept the generation");
+    }
+    scheduler_message_tx
+        .send(SchedulerMessage::Shutdown)
+        .expect("the scheduler must accept the shutdown");
+
+    ContinuousBatchScheduler::new(ContinuousBatchSchedulerParams {
+        agent_shutdown: CancellationToken::new(),
+        batch: LlamaBatch::new(runtime.model_runtime_parameters.n_batch.capacity(), 1),
+        llama_context,
+        scheduler_context: ContinuousBatchSchedulerContext {
+            agent_name: None,
+            desired_slots_total: 1,
+            model: &runtime.loaded_llama_model.model,
+            n_batch: runtime.model_runtime_parameters.n_batch,
+            sequence_context_size,
+        },
+        scheduler_message_rx,
+    })
+    .run();
+
+    assert_eq!(
+        generated_tokens_rx.try_recv(),
+        Ok(GeneratedTokenResult::NoSequenceSlotAvailable(
+            TextGenerationError::NoSequenceSlotAvailable.to_string()
+        ))
+    );
+}
diff --git a/paddler_tests/tests/a_generation_evaluates_its_prompt_and_each_generated_token_exactly_once.rs b/paddler_tests/tests/a_generation_evaluates_its_prompt_and_each_generated_token_exactly_once.rs
new file mode 100644
index 000000000..fc52dcd40
--- /dev/null
+++ b/paddler_tests/tests/a_generation_evaluates_its_prompt_and_each_generated_token_exactly_once.rs
@@ -0,0 +1,112 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::mem::discriminant;
+use std::num::NonZeroU32;
+
+use llama_cpp_bindings::llama_batch::LlamaBatch;
+use tokio::sync::mpsc::unbounded_channel;
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent_runtime::scheduler_message::SchedulerMessage;
+use paddler_agent_text_generation::continuous_batch_scheduler::ContinuousBatchScheduler;
+use paddler_agent_text_generation::continuous_batch_scheduler_context::ContinuousBatchSchedulerContext;
+use paddler_agent_text_generation::continuous_batch_scheduler_params::ContinuousBatchSchedulerParams;
+use paddler_tests::agent_inference_runtime_context::agent_inference_runtime_context;
+use paddler_tests::generation_done_discriminant::generation_done_discriminant;
+use paddler_tests::kev_checkpoint::KevCheckpoint;
+use paddler_tests::llama_runtime::LlamaRuntime;
+use paddler_tests::raw_prompt_generation::RawPromptGeneration;
+
+use crate::first_message_content_request_preparer::FirstMessageContentRequestPreparer;
+
+const PROMPT: &str = "Paddler balances inference across agents, and";
+const MAX_TOKENS: NonZeroU32 = NonZeroU32::new(8).unwrap();
+
+#[test]
+fn a_generation_evaluates_its_prompt_and_each_generated_token_exactly_once() {
+    let runtime = LlamaRuntime::load(
+        &KevCheckpoint::Kev0_8b.target_path("model.gguf"),
+        agent_inference_runtime_context(1),
+    )
+    .expect("the kev model must load");
+    let mut llama_context = runtime
+        .create_text_generation_llama_context(1)
+        .expect("the context must be created");
+    let sequence_context_size = llama_context.n_ctx_seq();
+    let n_batch = runtime.model_runtime_parameters.n_batch;
+    let request_preparer = FirstMessageContentRequestPreparer {
+        llama_runtime: &runtime,
+        sequence_context_size,
+    }
+    .into_generation_request_preparer();
+    let prompt_tokens = request_preparer
+        .prompt_tokenizer
+        .tokenize(PROMPT)
+        .expect("the prompt must tokenize")
+        .len();
+    let RawPromptGeneration {
+        mut agent_request,
+        generate_tokens_stop_tx: _generate_tokens_stop_tx,
+        mut generated_tokens_rx,
+    } = RawPromptGeneration::new(
+        PROMPT,
+        runtime
+            .inference_runtime_context
+            .slot_aggregated_status
+            .clone(),
+    );
+
+    agent_request.params.max_tokens = MAX_TOKENS;
+
+    let (scheduler_message_tx, scheduler_message_rx) = unbounded_channel();
+
+    scheduler_message_tx
+        .send(SchedulerMessage::Command(
+            request_preparer
+                .prepare_raw_prompt(agent_request)
+                .expect("the generation must be prepared"),
+        ))
+        .expect("the scheduler must accept the generation");
+    scheduler_message_tx
+        .send(SchedulerMessage::Shutdown)
+        .expect("the scheduler must accept the shutdown");
+
+    llama_context.reset_timings();
+
+    let mut scheduler = ContinuousBatchScheduler::new(ContinuousBatchSchedulerParams {
+        agent_shutdown: CancellationToken::new(),
+        batch: LlamaBatch::new(n_batch.capacity(), 1),
+        llama_context,
+        scheduler_context: ContinuousBatchSchedulerContext {
+            agent_name: None,
+            desired_slots_total: 1,
+            model: &runtime.loaded_llama_model.model,
+            n_batch,
+            sequence_context_size,
+        },
+        scheduler_message_rx,
+    });
+
+    scheduler.run();
+
+    let mut generated_token_results = Vec::new();
+
+    while let Ok(generated_token_result) = generated_tokens_rx.try_recv() {
+        generated_token_results.push(discriminant(&generated_token_result));
+    }
+
+    let timings = scheduler.llama_context.timings();
+
+    assert_eq!(
+        generated_token_results.last(),
+        Some(&generation_done_discriminant())
+    );
+    assert_eq!(
+        usize::try_from(timings.n_p_eval()).expect("the evaluated prompt tokens must count up"),
+        prompt_tokens
+    );
+    assert_eq!(
+        usize::try_from(timings.n_eval()).expect("the evaluated generated tokens must count up"),
+        generated_token_results.len() - 2
+    );
+}
diff --git a/paddler_tests/tests/a_generation_overflowing_its_batch_fails_as_batch_assembly.rs b/paddler_tests/tests/a_generation_overflowing_its_batch_fails_as_batch_assembly.rs
new file mode 100644
index 000000000..8c68fd95f
--- /dev/null
+++ b/paddler_tests/tests/a_generation_overflowing_its_batch_fails_as_batch_assembly.rs
@@ -0,0 +1,88 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::num::NonZeroU16;
+
+use llama_cpp_bindings::batch_add_error::BatchAddError;
+use llama_cpp_bindings::llama_batch::LlamaBatch;
+use llama_cpp_bindings_types::BatchCapacity;
+use tokio::sync::mpsc::unbounded_channel;
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent_runtime::scheduler_message::SchedulerMessage;
+use paddler_agent_text_generation::continuous_batch_scheduler::ContinuousBatchScheduler;
+use paddler_agent_text_generation::continuous_batch_scheduler_context::ContinuousBatchSchedulerContext;
+use paddler_agent_text_generation::continuous_batch_scheduler_params::ContinuousBatchSchedulerParams;
+use paddler_agent_text_generation::text_generation_error::TextGenerationError;
+use paddler_messaging::generated_token_result::GeneratedTokenResult;
+use paddler_tests::agent_inference_runtime_context::agent_inference_runtime_context;
+use paddler_tests::kev_checkpoint::KevCheckpoint;
+use paddler_tests::llama_runtime::LlamaRuntime;
+use paddler_tests::raw_prompt_generation::RawPromptGeneration;
+
+use crate::first_message_content_request_preparer::FirstMessageContentRequestPreparer;
+
+const TOKENS_THE_BATCH_HOLDS: NonZeroU16 = NonZeroU16::MIN;
+
+#[test]
+fn a_generation_overflowing_its_batch_fails_as_batch_assembly() {
+    let runtime = LlamaRuntime::load(
+        &KevCheckpoint::Kev0_8b.target_path("model.gguf"),
+        agent_inference_runtime_context(1),
+    )
+    .expect("the kev model must load");
+    let llama_context = runtime
+        .create_text_generation_llama_context(1)
+        .expect("the context must be created");
+    let sequence_context_size = llama_context.n_ctx_seq();
+    let RawPromptGeneration {
+        agent_request,
+        generate_tokens_stop_tx: _generate_tokens_stop_tx,
+        mut generated_tokens_rx,
+    } = RawPromptGeneration::new(
+        "Paddler balances requests across agents",
+        runtime
+            .inference_runtime_context
+            .slot_aggregated_status
+            .clone(),
+    );
+    let request = FirstMessageContentRequestPreparer {
+        llama_runtime: &runtime,
+        sequence_context_size,
+    }
+    .into_generation_request_preparer()
+    .prepare_raw_prompt(agent_request)
+    .expect("the generation must be prepared");
+    let (scheduler_message_tx, scheduler_message_rx) = unbounded_channel();
+
+    scheduler_message_tx
+        .send(SchedulerMessage::Command(request))
+        .expect("the scheduler must accept the generation");
+    scheduler_message_tx
+        .send(SchedulerMessage::Shutdown)
+        .expect("the scheduler must accept the shutdown");
+
+    ContinuousBatchScheduler::new(ContinuousBatchSchedulerParams {
+        agent_shutdown: CancellationToken::new(),
+        batch: LlamaBatch::new(BatchCapacity::from(TOKENS_THE_BATCH_HOLDS), 1),
+        llama_context,
+        scheduler_context: ContinuousBatchSchedulerContext {
+            agent_name: None,
+            desired_slots_total: 1,
+            model: &runtime.loaded_llama_model.model,
+            n_batch: runtime.model_runtime_parameters.n_batch,
+            sequence_context_size,
+        },
+        scheduler_message_rx,
+    })
+    .run();
+
+    assert_eq!(
+        generated_tokens_rx.try_recv(),
+        Ok(GeneratedTokenResult::BatchAssemblyFailed(
+            TextGenerationError::BatchAssemblyFailed(BatchAddError::InsufficientSpace(
+                usize::from(TOKENS_THE_BATCH_HOLDS.get())
+            ))
+            .to_string()
+        ))
+    );
+}
diff --git a/paddler_tests/tests/a_generation_step_overflowing_its_batch_fails_as_batch_assembly.rs b/paddler_tests/tests/a_generation_step_overflowing_its_batch_fails_as_batch_assembly.rs
new file mode 100644
index 000000000..cb1795b89
--- /dev/null
+++ b/paddler_tests/tests/a_generation_step_overflowing_its_batch_fails_as_batch_assembly.rs
@@ -0,0 +1,170 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::num::NonZeroU16;
+use std::num::NonZeroU32;
+use std::sync::Arc;
+
+use llama_cpp_bindings::batch_add_error::BatchAddError;
+use llama_cpp_bindings::llama_batch::LlamaBatch;
+use llama_cpp_bindings::mtmd::micro_batch_tokens;
+use llama_cpp_bindings_types::BatchCapacity;
+use tokio::sync::mpsc::UnboundedReceiver;
+use tokio::sync::mpsc::unbounded_channel;
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent_runtime::scheduler_message::SchedulerMessage;
+use paddler_agent_text_generation::continuous_batch_scheduler::ContinuousBatchScheduler;
+use paddler_agent_text_generation::continuous_batch_scheduler_context::ContinuousBatchSchedulerContext;
+use paddler_agent_text_generation::continuous_batch_scheduler_params::ContinuousBatchSchedulerParams;
+use paddler_agent_text_generation::model_bos_token::ModelBosToken;
+use paddler_agent_text_generation::multimodal_prompt_support::MultimodalPromptSupport;
+use paddler_agent_text_generation::resident_multimodal_context::ResidentMultimodalContext;
+use paddler_agent_text_generation::text_generation_error::TextGenerationError;
+use paddler_messaging::generated_token_result::GeneratedTokenResult;
+use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
+use paddler_test_cluster_harness::load_test_image_data_uri::load_test_image_data_uri;
+use paddler_test_cluster_harness::model_card::smolvlm2_256m::smolvlm2_256m;
+use paddler_test_cluster_harness::model_card::smolvlm2_256m_mmproj::smolvlm2_256m_mmproj;
+use paddler_tests::agent_inference_runtime_context::agent_inference_runtime_context;
+use paddler_tests::conversation_generation::ConversationGeneration;
+use paddler_tests::gbnf_literal::gbnf_literal;
+use paddler_tests::llama_runtime::LlamaRuntime;
+use paddler_tests::single_image_question::SingleImageQuestion;
+
+use crate::first_message_content_request_preparer::FirstMessageContentRequestPreparer;
+
+const ANSWER: &str = "two llamas";
+const MAX_TOKENS: NonZeroU32 = NonZeroU32::new(16).unwrap();
+const SLOTS: u16 = 2;
+const TOKENS_THE_BATCH_HOLDS: NonZeroU16 = NonZeroU16::MIN;
+
+fn last_result(
+    mut generated_tokens_rx: UnboundedReceiver<GeneratedTokenResult>,
+) -> Option<GeneratedTokenResult> {
+    let mut last_result = None;
+
+    while let Ok(generated_token_result) = generated_tokens_rx.try_recv() {
+        last_result = Some(generated_token_result);
+    }
+
+    last_result
+}
+
+#[test]
+fn a_generation_step_overflowing_its_batch_fails_as_batch_assembly() {
+    let runtime = LlamaRuntime::load(
+        &smolvlm2_256m()
+            .cached_path()
+            .expect("the vision model must be in the Hugging Face cache"),
+        agent_inference_runtime_context(SLOTS),
+    )
+    .expect("the vision model must load");
+    let llama_context = runtime
+        .create_text_generation_llama_context(u32::from(SLOTS))
+        .expect("the context must be created");
+    let sequence_context_size = llama_context.n_ctx_seq();
+    let model = &runtime.loaded_llama_model.model;
+    let n_batch = runtime.model_runtime_parameters.n_batch;
+    let multimodal_context = ResidentMultimodalContext::load(
+        &smolvlm2_256m_mmproj()
+            .cached_path()
+            .expect("the multimodal projection must be in the Hugging Face cache"),
+        model,
+        runtime
+            .inference_runtime_context
+            .vram_residencies
+            .multimodal_projection
+            .try_occupy()
+            .expect("the multimodal projection residency must be vacant"),
+    )
+    .expect("the multimodal projection must load");
+    let request_preparer = FirstMessageContentRequestPreparer {
+        llama_runtime: &runtime,
+        sequence_context_size,
+    }
+    .into_multimodal_generation_request_preparer(MultimodalPromptSupport {
+        micro_batch_tokens: micro_batch_tokens(&llama_context, n_batch.tokens()),
+        model_bos_token: ModelBosToken::of(model),
+        multimodal_context: Arc::new(multimodal_context),
+        n_batch: n_batch.tokens_i32(),
+        sequence_context_size,
+    });
+    let image_data_uri = load_test_image_data_uri().expect("the test image must load");
+    let image_question = || {
+        ConversationGeneration::new(
+            ContinueFromConversationHistoryParams {
+                grammar: Some(gbnf_literal(ANSWER)),
+                ..SingleImageQuestion {
+                    image_data_uri: image_data_uri.clone(),
+                    max_tokens: MAX_TOKENS,
+                    question: "How many llamas are in the image?",
+                }
+                .into_conversation_params()
+            },
+            runtime
+                .inference_runtime_context
+                .slot_aggregated_status
+                .clone(),
+        )
+    };
+    let ConversationGeneration {
+        agent_request: first_image_question,
+        generate_tokens_stop_tx: _first_generate_tokens_stop_tx,
+        generated_tokens_rx: first_generated_tokens_rx,
+    } = image_question();
+    let ConversationGeneration {
+        agent_request: second_image_question,
+        generate_tokens_stop_tx: _second_generate_tokens_stop_tx,
+        generated_tokens_rx: second_generated_tokens_rx,
+    } = image_question();
+    let (scheduler_message_tx, scheduler_message_rx) = unbounded_channel();
+
+    for agent_request in [first_image_question, second_image_question] {
+        scheduler_message_tx
+            .send(SchedulerMessage::Command(
+                request_preparer
+                    .prepare_conversation_history(agent_request)
+                    .expect("the image question must be prepared"),
+            ))
+            .expect("the scheduler must accept the image question");
+    }
+    scheduler_message_tx
+        .send(SchedulerMessage::Shutdown)
+        .expect("the scheduler must accept the shutdown");
+
+    ContinuousBatchScheduler::new(ContinuousBatchSchedulerParams {
+        agent_shutdown: CancellationToken::new(),
+        batch: LlamaBatch::new(
+            BatchCapacity::from(TOKENS_THE_BATCH_HOLDS),
+            i32::from(SLOTS),
+        ),
+        llama_context,
+        scheduler_context: ContinuousBatchSchedulerContext {
+            agent_name: None,
+            desired_slots_total: SLOTS,
+            model,
+            n_batch,
+            sequence_context_size,
+        },
+        scheduler_message_rx,
+    })
+    .run();
+
+    let batch_assembly_failure = GeneratedTokenResult::BatchAssemblyFailed(
+        TextGenerationError::BatchAssemblyFailed(BatchAddError::InsufficientSpace(usize::from(
+            TOKENS_THE_BATCH_HOLDS.get(),
+        )))
+        .to_string(),
+    );
+
+    assert_eq!(
+        [
+            last_result(first_generated_tokens_rx),
+            last_result(second_generated_tokens_rx),
+        ],
+        [
+            Some(batch_assembly_failure.clone()),
+            Some(batch_assembly_failure)
+        ]
+    );
+}
diff --git a/paddler_tests/tests/a_gguf_patch_refuses_a_source_that_is_not_a_gguf.rs b/paddler_tests/tests/a_gguf_patch_refuses_a_source_that_is_not_a_gguf.rs
new file mode 100644
index 000000000..11d45047e
--- /dev/null
+++ b/paddler_tests/tests/a_gguf_patch_refuses_a_source_that_is_not_a_gguf.rs
@@ -0,0 +1,27 @@
+use std::fs::write;
+
+use llama_cpp_bindings::gguf_context_error::GgufContextError;
+use tempfile::TempDir;
+
+use paddler_tests::point_qwen3_special_token_at_unused_padding::point_qwen3_special_token_at_unused_padding;
+use paddler_tests::write_patched_gguf::write_patched_gguf;
+
+#[test]
+fn a_gguf_patch_refuses_a_source_that_is_not_a_gguf() {
+    let model_directory = TempDir::new().expect("a temporary model directory must be creatable");
+    let source_path = model_directory.path().join("text-that-is-not-a-gguf.gguf");
+
+    write(&source_path, "text that is not a GGUF").expect("the source must be writable");
+
+    assert_eq!(
+        write_patched_gguf(
+            &source_path,
+            &model_directory.path().join("patched.gguf"),
+            point_qwen3_special_token_at_unused_padding(c"tokenizer.ggml.eos_token_id"),
+        )
+        .err()
+        .as_ref()
+        .map(ToString::to_string),
+        Some(GgufContextError::InitFailed(source_path).to_string())
+    );
+}
diff --git a/paddler_tests/tests/a_grammar_refusing_the_token_that_triggers_it_ends_the_generation_as_rejected_model_output.rs b/paddler_tests/tests/a_grammar_refusing_the_token_that_triggers_it_ends_the_generation_as_rejected_model_output.rs
new file mode 100644
index 000000000..cb4736698
--- /dev/null
+++ b/paddler_tests/tests/a_grammar_refusing_the_token_that_triggers_it_ends_the_generation_as_rejected_model_output.rs
@@ -0,0 +1,45 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use llama_cpp_bindings::sampling::LlamaSampler;
+
+use paddler_agent_text_generation::continuous_batch_terminal_outcome::ContinuousBatchTerminalOutcome;
+use paddler_agent_text_generation::grammar_sampling::GrammarSampling;
+use paddler_agent_text_generation::prepared_generation_request::PreparedGenerationRequest;
+use paddler_agent_text_generation::token_sampling::TokenSampling;
+use paddler_messaging::generated_token_result::GeneratedTokenResult;
+
+use crate::grammar_refusing_the_greedy_selection::GrammarRefusingTheGreedySelection;
+use crate::ready_to_sample_request::ReadyToSampleRequest;
+
+#[test]
+fn a_grammar_refusing_the_token_that_triggers_it_ends_the_generation_as_rejected_model_output() {
+    let refusal = GrammarRefusingTheGreedySelection::load();
+    let mut llama_context = refusal
+        .runtime
+        .create_text_generation_llama_context(1)
+        .expect("the context must be created");
+    let batch_position = refusal.decode_the_prompt(&mut llama_context);
+
+    assert_eq!(
+        ReadyToSampleRequest {
+            batch_position,
+            llama_context: &llama_context,
+            prompt: GrammarRefusingTheGreedySelection::PROMPT,
+            runtime: &refusal.runtime,
+        }
+        .advance(|prepared_request| PreparedGenerationRequest {
+            token_sampling: TokenSampling {
+                chain: LlamaSampler::greedy().expect("the greedy sampler must initialize"),
+                grammar: GrammarSampling::Constrained(
+                    refusal.grammar_awaiting_the_greedy_selection()
+                ),
+            },
+            ..prepared_request
+        }),
+        ContinuousBatchTerminalOutcome::EmitToClient(
+            GeneratedTokenResult::GrammarRejectedModelOutput(
+                GrammarRefusingTheGreedySelection::refusal().to_string()
+            )
+        )
+    );
+}
diff --git a/paddler_tests/tests/a_grammar_that_cannot_constrain_the_candidates_fails_the_sampling.rs b/paddler_tests/tests/a_grammar_that_cannot_constrain_the_candidates_fails_the_sampling.rs
new file mode 100644
index 000000000..838f49767
--- /dev/null
+++ b/paddler_tests/tests/a_grammar_that_cannot_constrain_the_candidates_fails_the_sampling.rs
@@ -0,0 +1,40 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use llama_cpp_bindings::error::SamplerApplyError;
+use llama_cpp_bindings::sampling::LlamaSampler;
+
+use paddler_agent_text_generation::grammar_sampling::GrammarSampling;
+use paddler_agent_text_generation::text_generation_error::TextGenerationError;
+use paddler_agent_text_generation::token_sampling::TokenSampling;
+use paddler_test_cluster_harness::model_card::nomic_embed_text_v1_5::nomic_embed_text_v1_5;
+
+use crate::sampling_with_a_smaller_vocabulary_grammar::SamplingWithASmallerVocabularyGrammar;
+
+#[test]
+fn a_grammar_that_cannot_constrain_the_candidates_fails_the_sampling() {
+    let sampling = SamplingWithASmallerVocabularyGrammar::load(
+        &nomic_embed_text_v1_5()
+            .cached_path()
+            .expect("the nomic embedding model must be in the Hugging Face cache"),
+    );
+
+    assert_eq!(
+        sampling
+            .sample_the_prompt_with(|grammar_of_a_smaller_vocabulary| TokenSampling {
+                chain: LlamaSampler::greedy().expect("the greedy sampler must initialize"),
+                grammar: GrammarSampling::Constrained(grammar_of_a_smaller_vocabulary),
+            })
+            .err()
+            .as_ref()
+            .map(ToString::to_string),
+        Some(
+            TextGenerationError::GrammarConstraintFailed(SamplerApplyError::Reported {
+                message: format!(
+                    "vector::_M_range_check: __n (which is {n_vocab}) >= this->size() (which is {n_vocab})",
+                    n_vocab = sampling.smaller_vocabulary_model.n_vocab(),
+                ),
+            })
+            .to_string()
+        )
+    );
+}
diff --git a/paddler_tests/tests/a_last_decision_question_overflowing_its_batch_fails_as_batch_assembly.rs b/paddler_tests/tests/a_last_decision_question_overflowing_its_batch_fails_as_batch_assembly.rs
new file mode 100644
index 000000000..d281b00fb
--- /dev/null
+++ b/paddler_tests/tests/a_last_decision_question_overflowing_its_batch_fails_as_batch_assembly.rs
@@ -0,0 +1,89 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::num::NonZeroU16;
+
+use llama_cpp_bindings::batch_add_error::BatchAddError;
+use llama_cpp_bindings::llama_batch::LlamaBatch;
+use llama_cpp_bindings::token::LlamaToken;
+use llama_cpp_bindings_types::BatchCapacity;
+
+use paddler_agent_decision::decision_capacity_ledger::DecisionCapacityLedger;
+use paddler_agent_decision::decision_decoder::DecisionDecoder;
+use paddler_agent_decision::decision_error::DecisionError;
+use paddler_agent_decision::decision_token_layout::DecisionTokenLayout;
+use paddler_tests::admitted_decision::AdmittedDecision;
+use paddler_tests::two_option_question::two_option_question;
+
+use crate::kev_decision_model::KevDecisionModel;
+
+const TOKENS_THE_BATCH_HOLDS: NonZeroU16 = NonZeroU16::MIN;
+const TOKENS_PER_QUESTION_CHUNK: usize = 2;
+
+#[test]
+fn a_last_decision_question_overflowing_its_batch_fails_as_batch_assembly() {
+    let KevDecisionModel {
+        delimiter_tokens,
+        pointer_head,
+        runtime,
+    } = KevDecisionModel::load(1);
+    let mut llama_context = runtime
+        .create_decision_llama_context(1)
+        .expect("the decision context must be created");
+    let mut batch = LlamaBatch::new(BatchCapacity::from(TOKENS_THE_BATCH_HOLDS), 1);
+    let layout = DecisionTokenLayout::new(
+        &delimiter_tokens,
+        vec![LlamaToken::new(0)],
+        Vec::new(),
+        two_option_question("question"),
+    );
+    let state_tokens = layout.state.len();
+    let AdmittedDecision {
+        mut active_decision_request,
+        decision_result_rx: _decision_result_rx,
+        decision_stop_tx: _decision_stop_tx,
+    } = AdmittedDecision::admit(
+        &DecisionCapacityLedger::new(1),
+        runtime
+            .inference_runtime_context
+            .slot_aggregated_status
+            .clone(),
+        layout,
+    )
+    .expect("the decision must be admitted");
+
+    for _state_token in 0..state_tokens {
+        active_decision_request
+            .advance(
+                &mut DecisionDecoder {
+                    batch: &mut batch,
+                    llama_context: &mut llama_context,
+                    n_batch: TOKENS_THE_BATCH_HOLDS.get().into(),
+                },
+                &pointer_head,
+                None,
+            )
+            .expect("the decision must advance");
+    }
+
+    assert_eq!(
+        active_decision_request
+            .advance(
+                &mut DecisionDecoder {
+                    batch: &mut batch,
+                    llama_context: &mut llama_context,
+                    n_batch: TOKENS_PER_QUESTION_CHUNK,
+                },
+                &pointer_head,
+                None,
+            )
+            .err()
+            .as_ref()
+            .map(ToString::to_string),
+        Some(
+            DecisionError::BatchAssemblyFailed(BatchAddError::InsufficientSpace(usize::from(
+                TOKENS_THE_BATCH_HOLDS.get()
+            )))
+            .to_string()
+        )
+    );
+}
diff --git a/paddler_tests/tests/a_leading_decision_question_llama_cpp_rejects_fails_the_decision.rs b/paddler_tests/tests/a_leading_decision_question_llama_cpp_rejects_fails_the_decision.rs
new file mode 100644
index 000000000..1a685491c
--- /dev/null
+++ b/paddler_tests/tests/a_leading_decision_question_llama_cpp_rejects_fails_the_decision.rs
@@ -0,0 +1,81 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use llama_cpp_bindings::error::DecodeError;
+use llama_cpp_bindings::llama_batch::LlamaBatch;
+use llama_cpp_bindings::token::LlamaToken;
+
+use paddler_agent_decision::decision_capacity_ledger::DecisionCapacityLedger;
+use paddler_agent_decision::decision_decoder::DecisionDecoder;
+use paddler_agent_decision::decision_error::DecisionError;
+use paddler_agent_decision::decision_token_layout::DecisionTokenLayout;
+use paddler_agent_decision::tokenized_decision_question::TokenizedDecisionQuestion;
+use paddler_tests::admitted_decision::AdmittedDecision;
+use paddler_tests::two_option_question::two_option_question;
+
+use crate::kev_decision_model::KevDecisionModel;
+
+#[test]
+fn a_leading_decision_question_llama_cpp_rejects_fails_the_decision() {
+    let KevDecisionModel {
+        delimiter_tokens,
+        pointer_head,
+        runtime,
+    } = KevDecisionModel::load(1);
+    let mut llama_context = runtime
+        .create_decision_llama_context(2)
+        .expect("the decision context must be created");
+    let n_batch = runtime.model_runtime_parameters.n_batch.tokens_usize();
+    let mut batch = LlamaBatch::new(runtime.model_runtime_parameters.n_batch.capacity(), 1);
+    let layout = DecisionTokenLayout::new(
+        &delimiter_tokens,
+        vec![LlamaToken::new(0)],
+        vec![TokenizedDecisionQuestion {
+            id: "leading".to_owned(),
+            instructions: vec![LlamaToken::new(runtime.loaded_llama_model.model.n_vocab())],
+            options: vec![vec![LlamaToken::new(1)], vec![LlamaToken::new(2)]],
+        }],
+        two_option_question("last"),
+    );
+    let AdmittedDecision {
+        mut active_decision_request,
+        decision_result_rx: _decision_result_rx,
+        decision_stop_tx: _decision_stop_tx,
+    } = AdmittedDecision::admit(
+        &DecisionCapacityLedger::new(2),
+        runtime
+            .inference_runtime_context
+            .slot_aggregated_status
+            .clone(),
+        layout,
+    )
+    .expect("the decision must be admitted");
+
+    active_decision_request
+        .advance(
+            &mut DecisionDecoder {
+                batch: &mut batch,
+                llama_context: &mut llama_context,
+                n_batch,
+            },
+            &pointer_head,
+            None,
+        )
+        .expect("the decision must advance");
+
+    assert_eq!(
+        active_decision_request
+            .advance(
+                &mut DecisionDecoder {
+                    batch: &mut batch,
+                    llama_context: &mut llama_context,
+                    n_batch,
+                },
+                &pointer_head,
+                None,
+            )
+            .err()
+            .as_ref()
+            .map(ToString::to_string),
+        Some(DecisionError::DecodeFailed(DecodeError::BatchInvalid).to_string())
+    );
+}
diff --git a/paddler_tests/tests/a_llama_runtime_reports_a_model_file_it_cannot_load.rs b/paddler_tests/tests/a_llama_runtime_reports_a_model_file_it_cannot_load.rs
new file mode 100644
index 000000000..a127e7066
--- /dev/null
+++ b/paddler_tests/tests/a_llama_runtime_reports_a_model_file_it_cannot_load.rs
@@ -0,0 +1,23 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::path::Path;
+
+use paddler_tests::agent_inference_runtime_context::agent_inference_runtime_context;
+use paddler_tests::llama_runtime::LlamaRuntime;
+
+#[test]
+fn a_llama_runtime_reports_a_model_file_it_cannot_load() {
+    assert_eq!(
+        LlamaRuntime::load(
+            Path::new("/paddler/definitely/missing/model.gguf"),
+            agent_inference_runtime_context(1),
+        )
+        .err()
+        .as_ref()
+        .map(ToString::to_string),
+        Some(
+            "unable to load the model from /paddler/definitely/missing/model.gguf: model file not found: /paddler/definitely/missing/model.gguf"
+                .to_owned()
+        )
+    );
+}
diff --git a/paddler_tests/tests/a_llama_runtime_reports_a_parallelism_it_cannot_determine.rs b/paddler_tests/tests/a_llama_runtime_reports_a_parallelism_it_cannot_determine.rs
new file mode 100644
index 000000000..28e773c81
--- /dev/null
+++ b/paddler_tests/tests/a_llama_runtime_reports_a_parallelism_it_cannot_determine.rs
@@ -0,0 +1,22 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use paddler_tests::inference_runtime_context_with_unknown_parallelism::inference_runtime_context_with_unknown_parallelism;
+use paddler_tests::kev_checkpoint::KevCheckpoint;
+use paddler_tests::llama_runtime::LlamaRuntime;
+
+#[test]
+fn a_llama_runtime_reports_a_parallelism_it_cannot_determine() {
+    assert_eq!(
+        LlamaRuntime::load(
+            &KevCheckpoint::Kev0_8b.target_path("model.gguf"),
+            inference_runtime_context_with_unknown_parallelism(1),
+        )
+        .err()
+        .as_ref()
+        .map(ToString::to_string),
+        Some(
+            "unable to determine how many threads this machine runs in parallel: the machine hides how many threads it runs in parallel"
+                .to_owned()
+        )
+    );
+}
diff --git a/paddler_tests/tests/a_model_held_by_a_failed_start_is_released_before_another_model_loads.rs b/paddler_tests/tests/a_model_held_by_a_failed_start_is_released_before_another_model_loads.rs
new file mode 100644
index 000000000..8924b4755
--- /dev/null
+++ b/paddler_tests/tests/a_model_held_by_a_failed_start_is_released_before_another_model_loads.rs
@@ -0,0 +1,66 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::mem::discriminant;
+
+use paddler_messaging::agent_decisions_serving::AgentDecisionsServing;
+use paddler_messaging::agent_runtime_status::AgentRuntimeStatus;
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_messaging::decision_settings::DecisionSettings;
+use paddler_messaging::produces_snapshot::ProducesSnapshot as _;
+use paddler_tests::desired_state_reconciled_discriminant::desired_state_reconciled_discriminant;
+use paddler_tests::desired_state_with_chat_template_override::desired_state_with_chat_template_override;
+use paddler_tests::kev_checkpoint::KevCheckpoint;
+use paddler_tests::qwen3_desired_state::qwen3_desired_state;
+use paddler_tests::uncompilable_chat_template::uncompilable_chat_template;
+
+use crate::serving_pipeline_arbiter::ServingPipelineArbiter;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn a_model_held_by_a_failed_start_is_released_before_another_model_loads() {
+    let mut serving_pipeline_arbiter = ServingPipelineArbiter::new(1);
+
+    assert_eq!(
+        discriminant(
+            &serving_pipeline_arbiter
+                .apply(desired_state_with_chat_template_override(
+                    BalancerDesiredState {
+                        decision: DecisionSettings::default(),
+                        ..KevCheckpoint::Kev0_8b
+                            .decision_desired_state()
+                            .into_balancer_desired_state()
+                            .expect("the kev desired state must name absolute paths")
+                    },
+                    uncompilable_chat_template(),
+                ))
+                .await
+        ),
+        desired_state_reconciled_discriminant()
+    );
+
+    serving_pipeline_arbiter.conclude_next_progress().await;
+    serving_pipeline_arbiter.conclude_next_progress().await;
+
+    assert_eq!(
+        discriminant(&serving_pipeline_arbiter.apply(qwen3_desired_state()).await),
+        desired_state_reconciled_discriminant()
+    );
+
+    serving_pipeline_arbiter.conclude_next_progress().await;
+    serving_pipeline_arbiter.conclude_next_progress().await;
+    serving_pipeline_arbiter.conclude_next_progress().await;
+
+    assert_eq!(
+        serving_pipeline_arbiter
+            .inference_runtime_context
+            .slot_aggregated_status
+            .make_snapshot()
+            .status
+            .runtime,
+        AgentRuntimeStatus::Serving {
+            decisions: AgentDecisionsServing::NotServed,
+            slots_total: 1,
+        }
+    );
+
+    serving_pipeline_arbiter.shut_down().await;
+}
diff --git a/paddler_tests/tests/a_model_keeps_its_vram_occupancy_until_its_last_clone_is_dropped.rs b/paddler_tests/tests/a_model_keeps_its_vram_occupancy_until_its_last_clone_is_dropped.rs
new file mode 100644
index 000000000..aace06fbc
--- /dev/null
+++ b/paddler_tests/tests/a_model_keeps_its_vram_occupancy_until_its_last_clone_is_dropped.rs
@@ -0,0 +1,40 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use paddler_agent_runtime::agent_runtime_error::AgentRuntimeError;
+use paddler_agent_runtime::llama_model_settings::LlamaModelSettings;
+use paddler_agent_runtime::loaded_llama_model::LoadedLlamaModel;
+use paddler_inference_parameters::all_gpu_layers::ALL_GPU_LAYERS;
+use paddler_tests::agent_inference_runtime_context::agent_inference_runtime_context;
+use paddler_tests::kev_checkpoint::KevCheckpoint;
+
+#[test]
+fn a_model_keeps_its_vram_occupancy_until_its_last_clone_is_dropped() {
+    let inference_runtime_context = agent_inference_runtime_context(1);
+    let model_weights_residency = inference_runtime_context
+        .vram_residencies
+        .model_weights
+        .clone();
+    let resident_model = LoadedLlamaModel::load(
+        &inference_runtime_context,
+        LlamaModelSettings {
+            model_path: KevCheckpoint::Kev0_8b.target_path("model.gguf"),
+            n_gpu_layers: ALL_GPU_LAYERS,
+        },
+    )
+    .expect("the model must load");
+    let resident_model_clone = resident_model.clone();
+
+    drop(resident_model);
+
+    assert_eq!(
+        model_weights_residency
+            .try_occupy()
+            .err()
+            .map(|occupancy_error| occupancy_error.to_string()),
+        Some(AgentRuntimeError::ModelWeightsAlreadyResident.to_string())
+    );
+
+    drop(resident_model_clone);
+
+    assert!(model_weights_residency.try_occupy().is_ok());
+}
diff --git a/paddler_tests/tests/a_model_with_a_metadata_key_that_is_not_utf8_reports_unreadable_metadata.rs b/paddler_tests/tests/a_model_with_a_metadata_key_that_is_not_utf8_reports_unreadable_metadata.rs
new file mode 100644
index 000000000..9b14e4c0a
--- /dev/null
+++ b/paddler_tests/tests/a_model_with_a_metadata_key_that_is_not_utf8_reports_unreadable_metadata.rs
@@ -0,0 +1,43 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use llama_cpp_bindings::error::MetaValError;
+use tempfile::TempDir;
+
+use paddler_agent_runtime::agent_runtime_error::AgentRuntimeError;
+use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
+use paddler_tests::agent_inference_runtime_context::agent_inference_runtime_context;
+use paddler_tests::llama_runtime::LlamaRuntime;
+use paddler_tests::write_patched_gguf::write_patched_gguf;
+
+#[test]
+fn a_model_with_a_metadata_key_that_is_not_utf8_reports_unreadable_metadata() {
+    let model_directory = TempDir::new().expect("a temporary model directory must be creatable");
+    let model_path = model_directory
+        .path()
+        .join("qwen3-with-a-key-that-is-not-utf8.gguf");
+
+    write_patched_gguf(
+        &qwen3_0_6b()
+            .cached_path()
+            .expect("Qwen3 must be in the Hugging Face cache"),
+        &model_path,
+        |gguf_context| {
+            gguf_context.set_val_str(c"general.\xff", c"value");
+        },
+    )
+    .expect("the patched model must be writable");
+
+    assert_eq!(
+        LlamaRuntime::load(&model_path, agent_inference_runtime_context(1))
+            .err()
+            .as_ref()
+            .map(ToString::to_string),
+        Some(
+            AgentRuntimeError::ModelMetadataUnreadable(MetaValError::FromUtf8Error(
+                String::from_utf8(b"general.\xff".to_vec())
+                    .expect_err("the bytes must not be UTF-8")
+            ))
+            .to_string()
+        )
+    );
+}
diff --git a/paddler_tests/tests/a_model_with_a_metadata_value_that_is_not_utf8_reports_unreadable_metadata.rs b/paddler_tests/tests/a_model_with_a_metadata_value_that_is_not_utf8_reports_unreadable_metadata.rs
new file mode 100644
index 000000000..850c657c0
--- /dev/null
+++ b/paddler_tests/tests/a_model_with_a_metadata_value_that_is_not_utf8_reports_unreadable_metadata.rs
@@ -0,0 +1,42 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use llama_cpp_bindings::error::MetaValError;
+use tempfile::TempDir;
+
+use paddler_agent_runtime::agent_runtime_error::AgentRuntimeError;
+use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
+use paddler_tests::agent_inference_runtime_context::agent_inference_runtime_context;
+use paddler_tests::llama_runtime::LlamaRuntime;
+use paddler_tests::write_patched_gguf::write_patched_gguf;
+
+#[test]
+fn a_model_with_a_metadata_value_that_is_not_utf8_reports_unreadable_metadata() {
+    let model_directory = TempDir::new().expect("a temporary model directory must be creatable");
+    let model_path = model_directory
+        .path()
+        .join("qwen3-with-a-value-that-is-not-utf8.gguf");
+
+    write_patched_gguf(
+        &qwen3_0_6b()
+            .cached_path()
+            .expect("Qwen3 must be in the Hugging Face cache"),
+        &model_path,
+        |gguf_context| {
+            gguf_context.set_val_str(c"general.name", c"\xff");
+        },
+    )
+    .expect("the patched model must be writable");
+
+    assert_eq!(
+        LlamaRuntime::load(&model_path, agent_inference_runtime_context(1))
+            .err()
+            .as_ref()
+            .map(ToString::to_string),
+        Some(
+            AgentRuntimeError::ModelMetadataUnreadable(MetaValError::FromUtf8Error(
+                String::from_utf8(vec![0xff]).expect_err("the bytes must not be UTF-8")
+            ))
+            .to_string()
+        )
+    );
+}
diff --git a/paddler_tests/tests/a_model_without_a_chat_template_generates_from_raw_prompts.rs b/paddler_tests/tests/a_model_without_a_chat_template_generates_from_raw_prompts.rs
new file mode 100644
index 000000000..be6df1d9b
--- /dev/null
+++ b/paddler_tests/tests/a_model_without_a_chat_template_generates_from_raw_prompts.rs
@@ -0,0 +1,46 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::num::NonZeroU32;
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
+use paddler_tests::start_single_agent_cluster_with_desired_state::start_single_agent_cluster_with_desired_state;
+
+use crate::qwen3_without_a_chat_template::Qwen3WithoutAChatTemplate;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn a_model_without_a_chat_template_generates_from_raw_prompts() {
+    let qwen3_without_a_chat_template = Qwen3WithoutAChatTemplate::write();
+    let cluster = start_single_agent_cluster_with_desired_state(
+        qwen3_without_a_chat_template.desired_state(),
+    )
+    .await
+    .expect("the cluster must start");
+
+    let collected = cluster
+        .continue_from_raw_prompt(
+            CancellationToken::new(),
+            &ContinueFromRawPromptParams {
+                grammar: None,
+                max_tokens: NonZeroU32::MIN,
+                raw_prompt: "The capital of France is".to_owned(),
+            },
+        )
+        .await
+        .expect("the raw prompt must be accepted");
+
+    assert_eq!(
+        collected
+            .summary()
+            .expect("the generation must finish with a summary")
+            .usage
+            .completion_tokens(),
+        1
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/a_model_without_a_chat_template_refuses_conversations.rs b/paddler_tests/tests/a_model_without_a_chat_template_refuses_conversations.rs
new file mode 100644
index 000000000..432447417
--- /dev/null
+++ b/paddler_tests/tests/a_model_without_a_chat_template_refuses_conversations.rs
@@ -0,0 +1,60 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::num::NonZeroU32;
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::conversation_history::ConversationHistory;
+use paddler_messaging::conversation_message::ConversationMessage;
+use paddler_messaging::conversation_message_content::ConversationMessageContent;
+use paddler_messaging::generated_token_result::GeneratedTokenResult;
+use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
+use paddler_tests::start_single_agent_cluster_with_desired_state::start_single_agent_cluster_with_desired_state;
+
+use crate::qwen3_without_a_chat_template::Qwen3WithoutAChatTemplate;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn a_model_without_a_chat_template_refuses_conversations() {
+    let qwen3_without_a_chat_template = Qwen3WithoutAChatTemplate::write();
+    let cluster = start_single_agent_cluster_with_desired_state(
+        qwen3_without_a_chat_template.desired_state(),
+    )
+    .await
+    .expect("the cluster must start");
+
+    let collected = cluster
+        .continue_from_conversation_history(
+            CancellationToken::new(),
+            &ContinueFromConversationHistoryParams {
+                add_generation_prompt: true,
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Text(
+                            "Name a primary color.".to_owned(),
+                        ),
+                        role: "user".to_owned(),
+                    },
+                ]),
+                enable_thinking: false,
+                grammar: None,
+                max_tokens: NonZeroU32::MIN,
+                parse_tool_calls: false,
+                tools: vec![],
+            },
+        )
+        .await
+        .expect("the conversation must be accepted");
+
+    assert_eq!(
+        collected.into_token_results(),
+        vec![GeneratedTokenResult::ChatTemplateError(format!(
+            "test-agent-0: the model {} has no chat template, so it cannot continue a conversation",
+            qwen3_without_a_chat_template.model_path().display()
+        ))]
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/a_multimodal_agent_serves_a_template_that_requires_a_system_message.rs b/paddler_tests/tests/a_multimodal_agent_serves_a_template_that_requires_a_system_message.rs
new file mode 100644
index 000000000..d8add6555
--- /dev/null
+++ b/paddler_tests/tests/a_multimodal_agent_serves_a_template_that_requires_a_system_message.rs
@@ -0,0 +1,63 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::num::NonZeroU32;
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_test_cluster_harness::load_test_image_data_uri::load_test_image_data_uri;
+use paddler_test_cluster_harness::model_card::smolvlm2_256m::smolvlm2_256m;
+use paddler_test_cluster_harness::model_card::smolvlm2_256m_mmproj::smolvlm2_256m_mmproj;
+use paddler_tests::chat_template_rendering_every_message_after::chat_template_rendering_every_message_after;
+use paddler_tests::desired_state_with_chat_template_override::desired_state_with_chat_template_override;
+use paddler_tests::image_description_conversation::image_description_conversation;
+use paddler_tests::start_cluster::start_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn a_multimodal_agent_serves_a_template_that_requires_a_system_message() {
+    let cluster = start_cluster(ClusterParams {
+        agents: vec![AgentConfig::single(1)],
+        desired_state: ClusterDesiredState::Apply(Box::new(
+            desired_state_with_chat_template_override(
+                smolvlm2_256m().into_desired_state_with_multimodal_projection(smolvlm2_256m_mmproj()),
+                chat_template_rendering_every_message_after(
+                    "{% if messages[0]['role'] != 'system' %}{{ raise_exception('the first message must be the system message') }}{% endif %}",
+                ),
+            ),
+        )),
+        wait_for_slots_ready: true,
+        ..ClusterParams::default()
+    })
+    .await
+    .expect("the cluster must start serving");
+
+    let collected = cluster
+        .continue_from_conversation_history(
+            CancellationToken::new(),
+            &ContinueFromConversationHistoryParams {
+                add_generation_prompt: true,
+                conversation_history: image_description_conversation(
+                    load_test_image_data_uri().expect("the test image must load"),
+                ),
+                enable_thinking: false,
+                grammar: None,
+                max_tokens: NonZeroU32::MIN,
+                parse_tool_calls: false,
+                tools: vec![],
+            },
+        )
+        .await
+        .expect("the inference request must be accepted");
+
+    collected
+        .summary()
+        .expect("the conversation opening with a system message must be generated to its end");
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/a_multimodal_agent_whose_chat_template_drops_the_image_rejects_the_prompt.rs b/paddler_tests/tests/a_multimodal_agent_whose_chat_template_drops_the_image_rejects_the_prompt.rs
new file mode 100644
index 000000000..5c3a9705b
--- /dev/null
+++ b/paddler_tests/tests/a_multimodal_agent_whose_chat_template_drops_the_image_rejects_the_prompt.rs
@@ -0,0 +1,69 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::num::NonZeroU32;
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::chat_template::ChatTemplate;
+use paddler_messaging::generated_token_result::GeneratedTokenResult;
+use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_test_cluster_harness::load_test_image_data_uri::load_test_image_data_uri;
+use paddler_tests::desired_state_with_chat_template_override::desired_state_with_chat_template_override;
+use paddler_tests::image_description_conversation::image_description_conversation;
+use paddler_tests::smolvlm2_desired_state::smolvlm2_desired_state;
+use paddler_tests::start_cluster::start_cluster;
+
+const CHAT_TEMPLATE_RENDERING_ONLY_ROLES: &str =
+    "{%- for message in messages -%}{{ message['role'] }}\n{% endfor -%}";
+
+#[tokio::test(flavor = "multi_thread")]
+async fn a_multimodal_agent_whose_chat_template_drops_the_image_rejects_the_prompt() {
+    let cluster = start_cluster(ClusterParams {
+        agents: vec![AgentConfig::single(1)],
+        desired_state: ClusterDesiredState::Apply(Box::new(
+            desired_state_with_chat_template_override(
+                smolvlm2_desired_state(),
+                ChatTemplate {
+                    content: CHAT_TEMPLATE_RENDERING_ONLY_ROLES.to_owned(),
+                },
+            ),
+        )),
+        wait_for_slots_ready: true,
+        ..ClusterParams::default()
+    })
+    .await
+    .expect("the cluster must start serving");
+
+    let collected = cluster
+        .continue_from_conversation_history(
+            CancellationToken::new(),
+            &ContinueFromConversationHistoryParams {
+                add_generation_prompt: true,
+                conversation_history: image_description_conversation(
+                    load_test_image_data_uri().expect("the test image must load"),
+                ),
+                enable_thinking: false,
+                grammar: None,
+                max_tokens: NonZeroU32::MIN,
+                parse_tool_calls: false,
+                tools: vec![],
+            },
+        )
+        .await
+        .expect("the inference request must be accepted");
+
+    assert_eq!(
+        collected.into_token_results(),
+        vec![GeneratedTokenResult::MultimodalTokenizationFailed(
+            "test-agent: the chat template rendered 0 media markers for 1 images".to_owned()
+        )]
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/a_multimodal_prompt_carries_one_bos_whether_or_not_its_template_renders_it.rs b/paddler_tests/tests/a_multimodal_prompt_carries_one_bos_whether_or_not_its_template_renders_it.rs
new file mode 100644
index 000000000..3ddd34cad
--- /dev/null
+++ b/paddler_tests/tests/a_multimodal_prompt_carries_one_bos_whether_or_not_its_template_renders_it.rs
@@ -0,0 +1,93 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::num::NonZeroU32;
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::chat_template::ChatTemplate;
+use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_test_cluster_harness::load_test_image_data_uri::load_test_image_data_uri;
+use paddler_test_cluster_harness::model_card::gemma_3_4b_it::gemma_3_4b_it;
+use paddler_test_cluster_harness::model_card::gemma_3_4b_it_mmproj::gemma_3_4b_it_mmproj;
+use paddler_tests::chat_template_rendering_every_message_after::chat_template_rendering_every_message_after;
+use paddler_tests::desired_state_with_chat_template_override::desired_state_with_chat_template_override;
+use paddler_tests::image_description_conversation::image_description_conversation;
+use paddler_tests::single_image_question::SingleImageQuestion;
+use paddler_tests::start_cluster::start_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn a_multimodal_prompt_carries_one_bos_whether_or_not_its_template_renders_it() {
+    let image_data_uri = load_test_image_data_uri().expect("the test image must load");
+    let conversations = [
+        image_description_conversation(image_data_uri.clone()),
+        SingleImageQuestion {
+            image_data_uri,
+            max_tokens: NonZeroU32::MIN,
+            question: "What do you see in this image?",
+        }
+        .into_conversation_params()
+        .conversation_history,
+    ];
+    let prompt_tokens_rendered_by = async |chat_template: ChatTemplate| -> Vec<u64> {
+        let cluster = start_cluster(ClusterParams {
+            agents: vec![AgentConfig::single(1)],
+            desired_state: ClusterDesiredState::Apply(Box::new(
+                desired_state_with_chat_template_override(
+                    gemma_3_4b_it()
+                        .into_desired_state_with_multimodal_projection(gemma_3_4b_it_mmproj()),
+                    chat_template,
+                ),
+            )),
+            wait_for_slots_ready: true,
+            ..ClusterParams::default()
+        })
+        .await
+        .expect("the multimodal cluster must start");
+        let mut prompt_tokens = Vec::with_capacity(conversations.len());
+
+        for conversation_history in &conversations {
+            prompt_tokens.push(
+                cluster
+                    .continue_from_conversation_history(
+                        CancellationToken::new(),
+                        &ContinueFromConversationHistoryParams {
+                            add_generation_prompt: true,
+                            conversation_history: conversation_history.clone(),
+                            enable_thinking: false,
+                            grammar: None,
+                            max_tokens: NonZeroU32::MIN,
+                            parse_tool_calls: false,
+                            tools: vec![],
+                        },
+                    )
+                    .await
+                    .expect("the conversation must be served")
+                    .summary()
+                    .expect("the conversation must end with a summary")
+                    .usage
+                    .prompt_tokens,
+            );
+        }
+
+        cluster
+            .shutdown()
+            .await
+            .expect("the multimodal cluster must shut down");
+
+        prompt_tokens
+    };
+
+    assert_eq!(
+        prompt_tokens_rendered_by(chat_template_rendering_every_message_after(
+            "{% if messages[0]['role'] == 'system' %}{{ bos_token }}{% endif %}",
+        ))
+        .await,
+        prompt_tokens_rendered_by(chat_template_rendering_every_message_after(
+            "{{ bos_token }}"
+        ))
+        .await
+    );
+}
diff --git a/paddler_tests/tests/a_multimodal_prompt_that_cannot_be_ingested_ends_the_generation.rs b/paddler_tests/tests/a_multimodal_prompt_that_cannot_be_ingested_ends_the_generation.rs
new file mode 100644
index 000000000..03329c3cf
--- /dev/null
+++ b/paddler_tests/tests/a_multimodal_prompt_that_cannot_be_ingested_ends_the_generation.rs
@@ -0,0 +1,150 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::sync::Arc;
+
+use llama_cpp_bindings::llama_batch::LlamaBatch;
+use llama_cpp_bindings::mtmd::micro_batch_tokens;
+use tokio::sync::mpsc::unbounded_channel;
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent_runtime::scheduler_message::SchedulerMessage;
+use paddler_agent_text_generation::continuous_batch_scheduler::ContinuousBatchScheduler;
+use paddler_agent_text_generation::continuous_batch_scheduler_context::ContinuousBatchSchedulerContext;
+use paddler_agent_text_generation::continuous_batch_scheduler_params::ContinuousBatchSchedulerParams;
+use paddler_agent_text_generation::media_marked_prompt::MediaMarkedPrompt;
+use paddler_agent_text_generation::model_bos_token::ModelBosToken;
+use paddler_agent_text_generation::multimodal_prompt_ingestion::MultimodalPromptIngestion;
+use paddler_agent_text_generation::multimodal_prompt_support::MultimodalPromptSupport;
+use paddler_agent_text_generation::prepared_generation_request::PreparedGenerationRequest;
+use paddler_agent_text_generation::prepared_prompt::PreparedPrompt;
+use paddler_agent_text_generation::resident_multimodal_context::ResidentMultimodalContext;
+use paddler_messaging::generated_token_result::GeneratedTokenResult;
+use paddler_messaging::media_marker::MediaMarker;
+use paddler_test_cluster_harness::model_card::smolvlm2_256m::smolvlm2_256m;
+use paddler_test_cluster_harness::model_card::smolvlm2_256m_mmproj::smolvlm2_256m_mmproj;
+use paddler_tests::agent_inference_runtime_context::agent_inference_runtime_context;
+use paddler_tests::llama_runtime::LlamaRuntime;
+use paddler_tests::raw_prompt_generation::RawPromptGeneration;
+
+use crate::first_message_content_request_preparer::FirstMessageContentRequestPreparer;
+
+#[test]
+fn a_multimodal_prompt_that_cannot_be_ingested_ends_the_generation() {
+    let first_result_of_a_multimodal_prompt =
+        |adjust_ingestion: &dyn Fn(&mut MultimodalPromptIngestion)| -> GeneratedTokenResult {
+            let runtime = LlamaRuntime::load(
+                &smolvlm2_256m()
+                    .cached_path()
+                    .expect("the vision model must be in the Hugging Face cache"),
+                agent_inference_runtime_context(1),
+            )
+            .expect("the vision model must load");
+            let llama_context = runtime
+                .create_text_generation_llama_context(1)
+                .expect("the context must be created");
+            let sequence_context_size = llama_context.n_ctx_seq();
+            let n_batch = runtime.model_runtime_parameters.n_batch;
+            let mut ingestion = MultimodalPromptSupport {
+                micro_batch_tokens: micro_batch_tokens(&llama_context, n_batch.tokens()),
+                model_bos_token: ModelBosToken::of(&runtime.loaded_llama_model.model),
+                multimodal_context: Arc::new(
+                    ResidentMultimodalContext::load(
+                        &smolvlm2_256m_mmproj()
+                            .cached_path()
+                            .expect("the multimodal projection must be in the Hugging Face cache"),
+                        &runtime.loaded_llama_model.model,
+                        runtime
+                            .inference_runtime_context
+                            .vram_residencies
+                            .multimodal_projection
+                            .try_occupy()
+                            .expect("no other multimodal projection may be resident"),
+                    )
+                    .expect("the multimodal projection must load"),
+                ),
+                n_batch: n_batch.tokens_i32(),
+                sequence_context_size,
+            }
+            .prepare_prompt(
+                &[],
+                &MediaMarkedPrompt {
+                    media_marker: &MediaMarker::new("<media>".to_owned()),
+                    prompt: "Describe the picture",
+                },
+            )
+            .expect("the multimodal prompt must be prepared");
+
+            adjust_ingestion(&mut ingestion);
+
+            let RawPromptGeneration {
+                agent_request,
+                generate_tokens_stop_tx: _generate_tokens_stop_tx,
+                mut generated_tokens_rx,
+            } = RawPromptGeneration::new(
+                "Describe the picture",
+                runtime
+                    .inference_runtime_context
+                    .slot_aggregated_status
+                    .clone(),
+            );
+            let request = FirstMessageContentRequestPreparer {
+                llama_runtime: &runtime,
+                sequence_context_size,
+            }
+            .into_generation_request_preparer()
+            .prepare_raw_prompt(agent_request)
+            .expect("the generation must be prepared");
+            let (scheduler_message_tx, scheduler_message_rx) = unbounded_channel();
+
+            scheduler_message_tx
+                .send(SchedulerMessage::Command(PreparedGenerationRequest {
+                    prompt: PreparedPrompt::Multimodal(ingestion),
+                    ..request
+                }))
+                .expect("the scheduler must accept the generation");
+            scheduler_message_tx
+                .send(SchedulerMessage::Shutdown)
+                .expect("the scheduler must accept the shutdown");
+
+            ContinuousBatchScheduler::new(ContinuousBatchSchedulerParams {
+                agent_shutdown: CancellationToken::new(),
+                batch: LlamaBatch::new(n_batch.capacity(), 1),
+                llama_context,
+                scheduler_context: ContinuousBatchSchedulerContext {
+                    agent_name: None,
+                    desired_slots_total: 1,
+                    model: &runtime.loaded_llama_model.model,
+                    n_batch,
+                    sequence_context_size,
+                },
+                scheduler_message_rx,
+            })
+            .run();
+
+            generated_tokens_rx
+                .try_recv()
+                .expect("the scheduler must answer the generation")
+        };
+
+    assert_eq!(
+        [
+            first_result_of_a_multimodal_prompt(&|ingestion: &mut MultimodalPromptIngestion| {
+                ingestion.next_chunk_index = ingestion.chunks.len();
+            }),
+            first_result_of_a_multimodal_prompt(&|ingestion: &mut MultimodalPromptIngestion| {
+                ingestion.n_batch = i32::try_from(ingestion.chunks.micro_batch_tokens() - 1)
+                    .expect("the fitted micro batch must fit a batch size");
+            }),
+        ],
+        [
+            GeneratedTokenResult::MultimodalIngestionFailed(
+                "failed to ingest multimodal prompt: chunk index 1 out of bounds during post-eval walk"
+                    .to_owned()
+            ),
+            GeneratedTokenResult::MultimodalIngestionFailed(
+                "failed to ingest multimodal prompt: a chunk fitted to decodes of 512 tokens cannot be evaluated in decodes of 511"
+                    .to_owned()
+            ),
+        ]
+    );
+}
diff --git a/paddler_tests/tests/a_new_desired_state_cancels_the_download_of_the_one_it_replaces.rs b/paddler_tests/tests/a_new_desired_state_cancels_the_download_of_the_one_it_replaces.rs
new file mode 100644
index 000000000..b9f451d5c
--- /dev/null
+++ b/paddler_tests/tests/a_new_desired_state_cancels_the_download_of_the_one_it_replaces.rs
@@ -0,0 +1,55 @@
+use std::time::Duration;
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_download_manager::read_timeout::READ_TIMEOUT;
+use paddler_local_http_fixture::fixture_response::FixtureResponse;
+use paddler_local_http_fixture::local_http_fixture::LocalHttpFixture;
+use paddler_messaging::agent_desired_model::AgentDesiredModel;
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_tests::start_single_agent_cluster_with_desired_state::start_single_agent_cluster_with_desired_state;
+
+const BYTE_INTERVAL_WITHIN_THE_READ_TIMEOUT: Duration =
+    Duration::from_secs(READ_TIMEOUT.as_secs() / 2);
+
+#[tokio::test(flavor = "multi_thread")]
+async fn a_new_desired_state_cancels_the_download_of_the_one_it_replaces() {
+    let fixture = LocalHttpFixture::start(FixtureResponse::EndlessTrickle {
+        byte_interval: BYTE_INTERVAL_WITHIN_THE_READ_TIMEOUT,
+    })
+    .await
+    .expect("the local HTTP fixture must start");
+    let cluster = start_single_agent_cluster_with_desired_state(BalancerDesiredState {
+        model: AgentDesiredModel::Uri(fixture.url("/endless.gguf")),
+        ..BalancerDesiredState::default()
+    })
+    .await
+    .expect("a single-agent cluster must start");
+
+    fixture
+        .wait_for_requests(1)
+        .await
+        .expect("the agent must start downloading the model of the replaced desired state");
+
+    cluster
+        .client_management
+        .put_balancer_desired_state(CancellationToken::new(), &BalancerDesiredState::default())
+        .await
+        .expect("the balancer must apply the replacement desired state");
+
+    fixture
+        .wait_for_abandoned_trickles(1)
+        .await
+        .expect("the agent must cancel the download of the replaced desired state");
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+
+    fixture
+        .stop()
+        .await
+        .expect("the fixture task must not panic")
+        .expect("the fixture must serve every request");
+}
diff --git a/paddler_tests/tests/a_prompt_deferred_by_a_full_batch_is_ingested_in_a_later_pass.rs b/paddler_tests/tests/a_prompt_deferred_by_a_full_batch_is_ingested_in_a_later_pass.rs
new file mode 100644
index 000000000..c6aad7543
--- /dev/null
+++ b/paddler_tests/tests/a_prompt_deferred_by_a_full_batch_is_ingested_in_a_later_pass.rs
@@ -0,0 +1,150 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::mem::Discriminant;
+use std::mem::discriminant;
+
+use llama_cpp_bindings::llama_batch::LlamaBatch;
+use llama_cpp_bindings::model::AddBos;
+use llama_cpp_bindings::model::ParseSpecialTokens;
+use tokio::sync::mpsc::UnboundedReceiver;
+use tokio::sync::mpsc::unbounded_channel;
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent_runtime::scheduler_message::SchedulerMessage;
+use paddler_agent_text_generation::continuous_batch_scheduler::ContinuousBatchScheduler;
+use paddler_agent_text_generation::continuous_batch_scheduler_context::ContinuousBatchSchedulerContext;
+use paddler_agent_text_generation::continuous_batch_scheduler_params::ContinuousBatchSchedulerParams;
+use paddler_messaging::generated_token_result::GeneratedTokenResult;
+use paddler_tests::agent_inference_runtime_context::agent_inference_runtime_context;
+use paddler_tests::generation_done_discriminant::generation_done_discriminant;
+use paddler_tests::kev_checkpoint::KevCheckpoint;
+use paddler_tests::llama_runtime::LlamaRuntime;
+use paddler_tests::raw_prompt_generation::RawPromptGeneration;
+
+use crate::first_message_content_request_preparer::FirstMessageContentRequestPreparer;
+
+const SENTENCE: &str = "Paddler balances inference across agents. ";
+const SLOTS: u16 = 2;
+
+fn last_result_discriminant(
+    mut generated_tokens_rx: UnboundedReceiver<GeneratedTokenResult>,
+) -> Option<Discriminant<GeneratedTokenResult>> {
+    let mut last_result = None;
+
+    while let Ok(generated_token_result) = generated_tokens_rx.try_recv() {
+        last_result = Some(discriminant(&generated_token_result));
+    }
+
+    last_result
+}
+
+#[test]
+fn a_prompt_deferred_by_a_full_batch_is_ingested_in_a_later_pass() {
+    let runtime = LlamaRuntime::load(
+        &KevCheckpoint::Kev0_8b.target_path("model.gguf"),
+        agent_inference_runtime_context(SLOTS),
+    )
+    .expect("the kev model must load");
+    let llama_context = runtime
+        .create_text_generation_llama_context(u32::from(SLOTS))
+        .expect("the context must be created");
+    let sequence_context_size = llama_context.n_ctx_seq();
+    let n_batch = runtime.model_runtime_parameters.n_batch;
+    let model = &runtime.loaded_llama_model.model;
+    let tokens_of_repeated_sentence = |repeats: usize| {
+        model
+            .str_to_token(
+                &SENTENCE.repeat(repeats),
+                AddBos::Never,
+                ParseSpecialTokens::Always,
+            )
+            .expect("the repeated sentence must tokenize")
+            .len()
+    };
+    let tokens_per_repeated_sentence =
+        tokens_of_repeated_sentence(2) - tokens_of_repeated_sentence(1);
+    let batch_filling_prompt =
+        SENTENCE.repeat(n_batch.tokens_usize() / tokens_per_repeated_sentence + 1);
+    let batch_filling_prompt_tokens = model
+        .str_to_token(
+            &batch_filling_prompt,
+            AddBos::Always,
+            ParseSpecialTokens::Always,
+        )
+        .expect("the batch filling prompt must tokenize")
+        .len();
+
+    assert!(
+        batch_filling_prompt_tokens > n_batch.tokens_usize()
+            && batch_filling_prompt_tokens < sequence_context_size as usize,
+        "{batch_filling_prompt_tokens}"
+    );
+
+    let request_preparer = FirstMessageContentRequestPreparer {
+        llama_runtime: &runtime,
+        sequence_context_size,
+    }
+    .into_generation_request_preparer();
+    let RawPromptGeneration {
+        agent_request: batch_filling_generation,
+        generate_tokens_stop_tx: _batch_filling_generate_tokens_stop_tx,
+        generated_tokens_rx: batch_filling_generated_tokens_rx,
+    } = RawPromptGeneration::new(
+        &batch_filling_prompt,
+        runtime
+            .inference_runtime_context
+            .slot_aggregated_status
+            .clone(),
+    );
+    let RawPromptGeneration {
+        agent_request: deferred_generation,
+        generate_tokens_stop_tx: _deferred_generate_tokens_stop_tx,
+        generated_tokens_rx: deferred_generated_tokens_rx,
+    } = RawPromptGeneration::new(
+        "Paddler",
+        runtime
+            .inference_runtime_context
+            .slot_aggregated_status
+            .clone(),
+    );
+    let (scheduler_message_tx, scheduler_message_rx) = unbounded_channel();
+
+    for agent_request in [batch_filling_generation, deferred_generation] {
+        scheduler_message_tx
+            .send(SchedulerMessage::Command(
+                request_preparer
+                    .prepare_raw_prompt(agent_request)
+                    .expect("the generation must be prepared"),
+            ))
+            .expect("the scheduler must accept the generation");
+    }
+    scheduler_message_tx
+        .send(SchedulerMessage::Shutdown)
+        .expect("the scheduler must accept the shutdown");
+
+    ContinuousBatchScheduler::new(ContinuousBatchSchedulerParams {
+        agent_shutdown: CancellationToken::new(),
+        batch: LlamaBatch::new(n_batch.capacity(), i32::from(SLOTS)),
+        llama_context,
+        scheduler_context: ContinuousBatchSchedulerContext {
+            agent_name: None,
+            desired_slots_total: SLOTS,
+            model: &runtime.loaded_llama_model.model,
+            n_batch,
+            sequence_context_size,
+        },
+        scheduler_message_rx,
+    })
+    .run();
+
+    assert_eq!(
+        [
+            last_result_discriminant(batch_filling_generated_tokens_rx),
+            last_result_discriminant(deferred_generated_tokens_rx),
+        ],
+        [
+            Some(generation_done_discriminant()),
+            Some(generation_done_discriminant())
+        ]
+    );
+}
diff --git a/paddler_tests/tests/a_prompt_that_tokenizes_to_nothing_refuses_the_generation.rs b/paddler_tests/tests/a_prompt_that_tokenizes_to_nothing_refuses_the_generation.rs
new file mode 100644
index 000000000..d579c35cd
--- /dev/null
+++ b/paddler_tests/tests/a_prompt_that_tokenizes_to_nothing_refuses_the_generation.rs
@@ -0,0 +1,44 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use paddler_agent_runtime::prepares_scheduler_command::PreparesSchedulerCommand as _;
+use paddler_agent_text_generation::text_generation_request::TextGenerationRequest;
+use paddler_messaging::generated_token_result::GeneratedTokenResult;
+use paddler_tests::agent_inference_runtime_context::agent_inference_runtime_context;
+use paddler_tests::kev_checkpoint::KevCheckpoint;
+use paddler_tests::llama_runtime::LlamaRuntime;
+use paddler_tests::raw_prompt_generation::RawPromptGeneration;
+
+use crate::first_message_content_request_preparer::FirstMessageContentRequestPreparer;
+
+#[test]
+fn a_prompt_that_tokenizes_to_nothing_refuses_the_generation() {
+    let runtime = LlamaRuntime::load(
+        &KevCheckpoint::Kev0_8b.target_path("model.gguf"),
+        agent_inference_runtime_context(1),
+    )
+    .expect("the kev model must load");
+    let preparer = FirstMessageContentRequestPreparer {
+        llama_runtime: &runtime,
+        sequence_context_size: runtime.model_runtime_parameters.context_size.get(),
+    }
+    .into_generation_request_preparer();
+    let RawPromptGeneration {
+        agent_request,
+        generate_tokens_stop_tx: _generate_tokens_stop_tx,
+        mut generated_tokens_rx,
+    } = RawPromptGeneration::new("", runtime.inference_runtime_context.slot_aggregated_status);
+
+    let scheduler_command = preparer.prepare_scheduler_command(
+        Some("test-agent"),
+        TextGenerationRequest::ContinueFromRawPrompt(agent_request),
+    );
+
+    assert!(scheduler_command.is_none());
+    assert_eq!(
+        generated_tokens_rx.try_recv(),
+        Ok(GeneratedTokenResult::PromptTokenizedToNothing(
+            "test-agent: the prompt tokenized to no tokens, so there is nothing to generate from"
+                .to_owned()
+        ))
+    );
+}
diff --git a/paddler_tests/tests/a_put_desired_state_is_answered_once_its_own_version_is_applied.rs b/paddler_tests/tests/a_put_desired_state_is_answered_once_its_own_version_is_applied.rs
new file mode 100644
index 000000000..b159bb8fd
--- /dev/null
+++ b/paddler_tests/tests/a_put_desired_state_is_answered_once_its_own_version_is_applied.rs
@@ -0,0 +1,45 @@
+use reqwest::StatusCode;
+use serde_json::json;
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::agent_desired_model::AgentDesiredModel;
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_tests::desired_state_serving_decisions::desired_state_serving_decisions;
+use paddler_tests::start_cluster::start_cluster;
+
+const DECISION_MODEL_URI: &str = "agent:///decision-model.gguf";
+
+#[tokio::test(flavor = "multi_thread")]
+async fn a_put_desired_state_is_answered_once_its_own_version_is_applied() {
+    let cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("a balancer serving text generation and embeddings must start");
+
+    cluster
+        .client_management
+        .put_balancer_desired_state(
+            CancellationToken::new(),
+            &BalancerDesiredState {
+                model: AgentDesiredModel::Uri(DECISION_MODEL_URI.to_owned()),
+                ..desired_state_serving_decisions()
+            },
+        )
+        .await
+        .expect("the balancer must accept the desired state");
+
+    let response = cluster
+        .typesafe_models()
+        .await
+        .expect("the models request must be answered");
+
+    assert_eq!(
+        (response.status, response.body),
+        (StatusCode::OK, json!({"models": []}))
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/a_question_lane_beyond_the_decision_context_cannot_be_cleared_for_the_state.rs b/paddler_tests/tests/a_question_lane_beyond_the_decision_context_cannot_be_cleared_for_the_state.rs
new file mode 100644
index 000000000..e386e48c0
--- /dev/null
+++ b/paddler_tests/tests/a_question_lane_beyond_the_decision_context_cannot_be_cleared_for_the_state.rs
@@ -0,0 +1,86 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use llama_cpp_bindings::error::ClearKvCacheSeqError;
+use llama_cpp_bindings::llama_batch::LlamaBatch;
+use llama_cpp_bindings::token::LlamaToken;
+
+use paddler_agent_decision::decision_capacity_ledger::DecisionCapacityLedger;
+use paddler_agent_decision::decision_decoder::DecisionDecoder;
+use paddler_agent_decision::decision_error::DecisionError;
+use paddler_agent_decision::decision_token_layout::DecisionTokenLayout;
+use paddler_tests::admitted_decision::AdmittedDecision;
+use paddler_tests::two_option_question::two_option_question;
+
+use crate::kev_decision_model::KevDecisionModel;
+
+const SEQUENCES_THE_CONTEXT_HOLDS: u32 = 1;
+const SEQUENCES_THE_LEDGER_HANDS_OUT: u16 = 2;
+
+#[test]
+fn a_question_lane_beyond_the_decision_context_cannot_be_cleared_for_the_state() {
+    let KevDecisionModel {
+        delimiter_tokens,
+        pointer_head,
+        runtime,
+    } = KevDecisionModel::load(1);
+    let mut llama_context = runtime
+        .create_decision_llama_context(SEQUENCES_THE_CONTEXT_HOLDS)
+        .expect("the decision context must be created");
+    let n_batch = runtime.model_runtime_parameters.n_batch.tokens_usize();
+    let mut batch = LlamaBatch::new(runtime.model_runtime_parameters.n_batch.capacity(), 1);
+    let layout = DecisionTokenLayout::new(
+        &delimiter_tokens,
+        vec![LlamaToken::new(0)],
+        vec![two_option_question("leading")],
+        two_option_question("last"),
+    );
+    let AdmittedDecision {
+        mut active_decision_request,
+        decision_result_rx: _decision_result_rx,
+        decision_stop_tx: _decision_stop_tx,
+    } = AdmittedDecision::admit(
+        &DecisionCapacityLedger::new(SEQUENCES_THE_LEDGER_HANDS_OUT),
+        runtime
+            .inference_runtime_context
+            .slot_aggregated_status
+            .clone(),
+        layout,
+    )
+    .expect("the decision must be admitted");
+
+    active_decision_request
+        .advance(
+            &mut DecisionDecoder {
+                batch: &mut batch,
+                llama_context: &mut llama_context,
+                n_batch,
+            },
+            &pointer_head,
+            None,
+        )
+        .expect("the decision must advance");
+
+    assert_eq!(
+        active_decision_request
+            .advance(
+                &mut DecisionDecoder {
+                    batch: &mut batch,
+                    llama_context: &mut llama_context,
+                    n_batch,
+                },
+                &pointer_head,
+                None,
+            )
+            .err()
+            .as_ref()
+            .map(ToString::to_string),
+        Some(
+            DecisionError::KvCacheRemovalFailed(ClearKvCacheSeqError::PartialSequenceNotRemoved {
+                seq_id: 1,
+                p0: -1,
+                p1: -1,
+            })
+            .to_string()
+        )
+    );
+}
diff --git a/paddler_tests/tests/a_reconfiguration_replaced_before_it_starts_yields_to_its_replacement.rs b/paddler_tests/tests/a_reconfiguration_replaced_before_it_starts_yields_to_its_replacement.rs
new file mode 100644
index 000000000..ddb9dc896
--- /dev/null
+++ b/paddler_tests/tests/a_reconfiguration_replaced_before_it_starts_yields_to_its_replacement.rs
@@ -0,0 +1,93 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::collections::BTreeSet;
+use std::mem::discriminant;
+
+use paddler_inference_parameters::embedding_parameters::EmbeddingParameters;
+use paddler_inference_parameters::pooling_type::PoolingType;
+use paddler_messaging::agent_decisions_serving::AgentDecisionsServing;
+use paddler_messaging::agent_runtime_status::AgentRuntimeStatus;
+use paddler_messaging::agent_state_application_status::AgentStateApplicationStatus;
+use paddler_messaging::agent_status::AgentStatus;
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_messaging::decision_settings::DecisionSettings;
+use paddler_messaging::model_download_status::ModelDownloadStatus;
+use paddler_messaging::produces_snapshot::ProducesSnapshot as _;
+use paddler_tests::desired_state_with_chat_template_override::desired_state_with_chat_template_override;
+use paddler_tests::first_message_content_chat_template::first_message_content_chat_template;
+use paddler_tests::kev_checkpoint::KevCheckpoint;
+
+use crate::serving_pipeline_arbiter::ServingPipelineArbiter;
+use crate::swappable_model_file::SwappableModelFile;
+use paddler_tests::desired_state_reconciled_discriminant::desired_state_reconciled_discriminant;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn a_reconfiguration_replaced_before_it_starts_yields_to_its_replacement() {
+    let swappable_model_file =
+        SwappableModelFile::link_to(&KevCheckpoint::Kev0_8b.target_path("model.gguf"));
+    let served_desired_state = BalancerDesiredState {
+        decision: DecisionSettings::default(),
+        model: swappable_model_file.agent_desired_model(),
+        ..KevCheckpoint::Kev0_8b
+            .decision_desired_state()
+            .into_balancer_desired_state()
+            .expect("the kev desired state must name absolute paths")
+    };
+    let mut serving_pipeline_arbiter = ServingPipelineArbiter::new(1);
+
+    serving_pipeline_arbiter
+        .serve(served_desired_state.clone())
+        .await;
+
+    swappable_model_file.replace_with_an_unloadable_file();
+
+    assert_eq!(
+        discriminant(
+            &serving_pipeline_arbiter
+                .apply(BalancerDesiredState {
+                    embeddings: EmbeddingParameters {
+                        pooling_type: PoolingType::Mean,
+                        ..served_desired_state.embeddings.clone()
+                    },
+                    ..served_desired_state.clone()
+                })
+                .await
+        ),
+        desired_state_reconciled_discriminant()
+    );
+    assert_eq!(
+        discriminant(
+            &serving_pipeline_arbiter
+                .apply(desired_state_with_chat_template_override(
+                    served_desired_state,
+                    first_message_content_chat_template(),
+                ))
+                .await
+        ),
+        desired_state_reconciled_discriminant()
+    );
+    serving_pipeline_arbiter.conclude_next_progress().await;
+    serving_pipeline_arbiter.conclude_next_progress().await;
+    serving_pipeline_arbiter.conclude_next_progress().await;
+    assert_eq!(
+        serving_pipeline_arbiter
+            .inference_runtime_context
+            .slot_aggregated_status
+            .make_snapshot()
+            .status,
+        AgentStatus {
+            desired_slots_total: 1,
+            download_status: ModelDownloadStatus::NotDownloading,
+            issues: BTreeSet::new(),
+            model_path: Some(swappable_model_file.model_path().display().to_string()),
+            runtime: AgentRuntimeStatus::Serving {
+                decisions: AgentDecisionsServing::NotServed,
+                slots_total: 1,
+            },
+            state_application_status: AgentStateApplicationStatus::Applied,
+            uses_chat_template_override: true,
+        }
+    );
+
+    serving_pipeline_arbiter.shut_down().await;
+}
diff --git a/paddler_tests/tests/a_replacement_superseded_by_its_own_state_keeps_reporting_the_model_metadata.rs b/paddler_tests/tests/a_replacement_superseded_by_its_own_state_keeps_reporting_the_model_metadata.rs
new file mode 100644
index 000000000..b7c9772ef
--- /dev/null
+++ b/paddler_tests/tests/a_replacement_superseded_by_its_own_state_keeps_reporting_the_model_metadata.rs
@@ -0,0 +1,49 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_messaging::decision_settings::DecisionSettings;
+use paddler_messaging::decision_slots_minimum::DECISION_SLOTS_MINIMUM;
+use paddler_tests::kev_checkpoint::KevCheckpoint;
+
+use crate::serving_pipeline_arbiter::ServingPipelineArbiter;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn a_replacement_superseded_by_its_own_state_keeps_reporting_the_model_metadata() {
+    let serving_decisions = KevCheckpoint::Kev0_8b
+        .decision_desired_state()
+        .into_balancer_desired_state()
+        .expect("the kev desired state must name absolute paths");
+    let serving_without_decisions = BalancerDesiredState {
+        decision: DecisionSettings::default(),
+        ..serving_decisions.clone()
+    };
+    let mut serving_pipeline_arbiter = ServingPipelineArbiter::new(DECISION_SLOTS_MINIMUM.get());
+
+    serving_pipeline_arbiter.serve(serving_decisions).await;
+
+    let served_model_metadata = serving_pipeline_arbiter
+        .inference_runtime_context
+        .model_metadata_holder
+        .get_model_metadata();
+
+    serving_pipeline_arbiter
+        .apply(serving_without_decisions.clone())
+        .await;
+    serving_pipeline_arbiter.conclude_next_progress().await;
+    serving_pipeline_arbiter
+        .apply(serving_without_decisions)
+        .await;
+    serving_pipeline_arbiter.conclude_next_progress().await;
+    serving_pipeline_arbiter.conclude_next_progress().await;
+
+    assert_ne!(served_model_metadata, None);
+    assert_eq!(
+        serving_pipeline_arbiter
+            .inference_runtime_context
+            .model_metadata_holder
+            .get_model_metadata(),
+        served_model_metadata
+    );
+
+    serving_pipeline_arbiter.shut_down().await;
+}
diff --git a/paddler_tests/tests/a_restart_cancelled_while_starting_a_capability_shuts_down_the_capabilities_it_kept.rs b/paddler_tests/tests/a_restart_cancelled_while_starting_a_capability_shuts_down_the_capabilities_it_kept.rs
new file mode 100644
index 000000000..f65d3ed1e
--- /dev/null
+++ b/paddler_tests/tests/a_restart_cancelled_while_starting_a_capability_shuts_down_the_capabilities_it_kept.rs
@@ -0,0 +1,109 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::mem::discriminant;
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent::agent_applicable_state::AgentApplicableState;
+use paddler_agent::finished_serving_startup::FinishedServingStartup;
+use paddler_agent::inference_capabilities::InferenceCapabilities;
+use paddler_agent::model_serving::ModelServing;
+use paddler_agent::model_serving_settings::ModelServingSettings;
+use paddler_agent::prepared_decisions::PreparedDecisions;
+use paddler_agent::prepared_serving::PreparedServing;
+use paddler_agent::replacement_outcome::ReplacementOutcome;
+use paddler_agent::serving_replacement::ServingReplacement;
+use paddler_agent_text_generation::text_generation_request::TextGenerationRequest;
+use paddler_inference_parameters::pooling_type::PoolingType;
+use paddler_messaging::generated_token_result::GeneratedTokenResult;
+use paddler_tests::llama_runtime::LlamaRuntime;
+use paddler_tests::raw_prompt_generation::RawPromptGeneration;
+
+use crate::kev_model_serving_without_decisions::KevModelServingWithoutDecisions;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn a_restart_cancelled_while_starting_a_capability_shuts_down_the_capabilities_it_kept() {
+    let KevModelServingWithoutDecisions {
+        llama_runtime:
+            LlamaRuntime {
+                inference_runtime_context,
+                loaded_llama_model: kev_model_loaded_for_serving,
+                ..
+            },
+        model_serving_settings,
+        served_model,
+    } = KevModelServingWithoutDecisions::start(&CancellationToken::new()).await;
+
+    drop(kev_model_loaded_for_serving);
+
+    let kept_text_generation_request_preparer = served_model
+        .capabilities
+        .text_generation
+        .request_preparer
+        .clone();
+    let cancelled_restart = CancellationToken::new();
+
+    cancelled_restart.cancel();
+
+    let FinishedServingStartup {
+        outcome: restart_outcome,
+        ..
+    } = ServingReplacement {
+        held: ModelServing::Serving(Box::new(served_model)),
+        prepared_serving: PreparedServing {
+            agent_applicable_state: AgentApplicableState::WithModel(ModelServingSettings {
+                capabilities: InferenceCapabilities {
+                    embeddings: PoolingType::Mean,
+                    ..model_serving_settings.capabilities.clone()
+                },
+                ..model_serving_settings
+            }),
+            prepared_decisions: PreparedDecisions::Disabled,
+        },
+    }
+    .replace(cancelled_restart, inference_runtime_context.clone())
+    .await;
+
+    assert_eq!(
+        discriminant(&restart_outcome),
+        discriminant(&ReplacementOutcome::Cancelled {
+            held: ModelServing::Idle
+        })
+    );
+    assert!(
+        inference_runtime_context
+            .vram_residencies
+            .text_generation_context
+            .try_occupy()
+            .is_ok()
+    );
+
+    let RawPromptGeneration {
+        agent_request,
+        generate_tokens_stop_tx: _generate_tokens_stop_tx,
+        mut generated_tokens_rx,
+    } = RawPromptGeneration::new(
+        "The capital of France is",
+        inference_runtime_context.slot_aggregated_status.clone(),
+    );
+
+    kept_text_generation_request_preparer
+        .prepare(TextGenerationRequest::ContinueFromRawPrompt(agent_request));
+
+    assert_eq!(
+        generated_tokens_rx.recv().await,
+        Some(GeneratedTokenResult::SchedulerUnavailable(
+            "agent: the scheduler is no longer accepting requests".to_owned()
+        ))
+    );
+
+    drop(kept_text_generation_request_preparer);
+
+    let model_weights = &inference_runtime_context.vram_residencies.model_weights;
+
+    assert!(model_weights.try_occupy().is_err());
+
+    drop(restart_outcome);
+
+    assert!(model_weights.try_occupy().is_ok());
+}
diff --git a/paddler_tests/tests/a_sampled_token_the_vocabulary_cannot_render_ends_the_generation.rs b/paddler_tests/tests/a_sampled_token_the_vocabulary_cannot_render_ends_the_generation.rs
new file mode 100644
index 000000000..097da16f7
--- /dev/null
+++ b/paddler_tests/tests/a_sampled_token_the_vocabulary_cannot_render_ends_the_generation.rs
@@ -0,0 +1,66 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use llama_cpp_bindings::error::TokenToStringError;
+use llama_cpp_bindings::sampling::LlamaSampler;
+use llama_cpp_bindings::token::LlamaToken;
+use llama_cpp_bindings::token::logit_bias::LlamaLogitBias;
+
+use paddler_agent_text_generation::continuous_batch_terminal_outcome::ContinuousBatchTerminalOutcome;
+use paddler_agent_text_generation::grammar_sampling::GrammarSampling;
+use paddler_agent_text_generation::prepared_generation_request::PreparedGenerationRequest;
+use paddler_agent_text_generation::token_sampling::TokenSampling;
+use paddler_messaging::generated_token_result::GeneratedTokenResult;
+use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
+use paddler_tests::agent_inference_runtime_context::agent_inference_runtime_context;
+use paddler_tests::llama_runtime::LlamaRuntime;
+use paddler_tests::qwen3_0_6b_unused_padding_token::QWEN3_0_6B_UNUSED_PADDING_TOKEN;
+
+use crate::decode_prompt_for_sampling::decode_prompt_for_sampling;
+use crate::ready_to_sample_request::ReadyToSampleRequest;
+
+const PROMPT: &str = "Paddler balances";
+
+#[test]
+fn a_sampled_token_the_vocabulary_cannot_render_ends_the_generation() {
+    let runtime = LlamaRuntime::load(
+        &qwen3_0_6b()
+            .cached_path()
+            .expect("the model must be in the Hugging Face cache"),
+        agent_inference_runtime_context(1),
+    )
+    .expect("the model must load");
+    let mut llama_context = runtime
+        .create_text_generation_llama_context(1)
+        .expect("the context must be created");
+    let batch_position = decode_prompt_for_sampling(&runtime, &mut llama_context, PROMPT);
+    let unrenderable_token = LlamaToken::new(
+        i32::try_from(QWEN3_0_6B_UNUSED_PADDING_TOKEN).expect("the padding token must fit i32"),
+    );
+
+    assert_eq!(
+        ReadyToSampleRequest {
+            batch_position,
+            llama_context: &llama_context,
+            prompt: PROMPT,
+            runtime: &runtime,
+        }
+        .advance(|prepared_request| PreparedGenerationRequest {
+            token_sampling: TokenSampling {
+                chain: LlamaSampler::chain_simple([
+                    LlamaSampler::logit_bias(
+                        runtime.loaded_llama_model.model.n_vocab(),
+                        &[LlamaLogitBias::new(unrenderable_token, f32::INFINITY)],
+                    )
+                    .expect("the logit bias sampler must initialize"),
+                    LlamaSampler::greedy().expect("the greedy sampler must initialize"),
+                ])
+                .expect("the sampler chain must initialize"),
+                grammar: GrammarSampling::Unconstrained,
+            },
+            ..prepared_request
+        }),
+        ContinuousBatchTerminalOutcome::EmitToClient(GeneratedTokenResult::DetokenizationFailed(
+            TokenToStringError::UnknownTokenType.to_string()
+        ))
+    );
+}
diff --git a/paddler_tests/tests/a_sampler_chain_that_cannot_rank_the_candidates_fails_the_sampling.rs b/paddler_tests/tests/a_sampler_chain_that_cannot_rank_the_candidates_fails_the_sampling.rs
new file mode 100644
index 000000000..6bf988899
--- /dev/null
+++ b/paddler_tests/tests/a_sampler_chain_that_cannot_rank_the_candidates_fails_the_sampling.rs
@@ -0,0 +1,41 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use llama_cpp_bindings::error::SamplerApplyError;
+use llama_cpp_bindings::sampling::LlamaSampler;
+
+use paddler_agent_text_generation::grammar_sampling::GrammarSampling;
+use paddler_agent_text_generation::text_generation_error::TextGenerationError;
+use paddler_agent_text_generation::token_sampling::TokenSampling;
+use paddler_test_cluster_harness::model_card::nomic_embed_text_v1_5::nomic_embed_text_v1_5;
+
+use crate::sampling_with_a_smaller_vocabulary_grammar::SamplingWithASmallerVocabularyGrammar;
+
+#[test]
+fn a_sampler_chain_that_cannot_rank_the_candidates_fails_the_sampling() {
+    let sampling = SamplingWithASmallerVocabularyGrammar::load(
+        &nomic_embed_text_v1_5()
+            .cached_path()
+            .expect("the nomic embedding model must be in the Hugging Face cache"),
+    );
+
+    assert_eq!(
+        sampling
+            .sample_the_prompt_with(|grammar_of_a_smaller_vocabulary| TokenSampling {
+                chain: LlamaSampler::chain_simple([grammar_of_a_smaller_vocabulary])
+                    .expect("the sampler chain must initialize"),
+                grammar: GrammarSampling::Unconstrained,
+            })
+            .err()
+            .as_ref()
+            .map(ToString::to_string),
+        Some(
+            TextGenerationError::SamplerChainApplicationFailed(SamplerApplyError::Reported {
+                message: format!(
+                    "vector::_M_range_check: __n (which is {n_vocab}) >= this->size() (which is {n_vocab})",
+                    n_vocab = sampling.smaller_vocabulary_model.n_vocab(),
+                ),
+            })
+            .to_string()
+        )
+    );
+}
diff --git a/paddler_tests/tests/a_sampler_chain_that_refuses_its_own_selection_fails_the_sampling.rs b/paddler_tests/tests/a_sampler_chain_that_refuses_its_own_selection_fails_the_sampling.rs
new file mode 100644
index 000000000..603cada8a
--- /dev/null
+++ b/paddler_tests/tests/a_sampler_chain_that_refuses_its_own_selection_fails_the_sampling.rs
@@ -0,0 +1,45 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use llama_cpp_bindings::sampling::LlamaSampler;
+use llama_cpp_bindings::token::data_array::LlamaTokenDataArray;
+
+use paddler_agent_text_generation::grammar_sampling::GrammarSampling;
+use paddler_agent_text_generation::text_generation_error::TextGenerationError;
+use paddler_agent_text_generation::token_sampling::TokenSampling;
+
+use crate::grammar_refusing_the_greedy_selection::GrammarRefusingTheGreedySelection;
+
+#[test]
+fn a_sampler_chain_that_refuses_its_own_selection_fails_the_sampling() {
+    let refusal = GrammarRefusingTheGreedySelection::load();
+    let mut llama_context = refusal
+        .runtime
+        .create_text_generation_llama_context(1)
+        .expect("the context must be created");
+    let batch_position = refusal.decode_the_prompt(&mut llama_context);
+
+    assert_eq!(
+        TokenSampling {
+            chain: LlamaSampler::chain_simple([
+                LlamaSampler::greedy().expect("the greedy sampler must initialize"),
+                refusal.grammar(),
+            ])
+            .expect("the sampler chain must initialize"),
+            grammar: GrammarSampling::Unconstrained,
+        }
+        .sample(
+            &llama_context,
+            batch_position,
+            &mut LlamaTokenDataArray::new(Vec::new(), false),
+        )
+        .err()
+        .as_ref()
+        .map(ToString::to_string),
+        Some(
+            TextGenerationError::SamplerChainRejectedToken(
+                GrammarRefusingTheGreedySelection::refusal()
+            )
+            .to_string()
+        )
+    );
+}
diff --git a/paddler_tests/tests/a_sampler_chain_that_selects_no_candidate_ends_the_generation.rs b/paddler_tests/tests/a_sampler_chain_that_selects_no_candidate_ends_the_generation.rs
new file mode 100644
index 000000000..40d243049
--- /dev/null
+++ b/paddler_tests/tests/a_sampler_chain_that_selects_no_candidate_ends_the_generation.rs
@@ -0,0 +1,51 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use llama_cpp_bindings::sampling::LlamaSampler;
+
+use paddler_agent_text_generation::continuous_batch_terminal_outcome::ContinuousBatchTerminalOutcome;
+use paddler_agent_text_generation::grammar_sampling::GrammarSampling;
+use paddler_agent_text_generation::prepared_generation_request::PreparedGenerationRequest;
+use paddler_agent_text_generation::token_sampling::TokenSampling;
+use paddler_messaging::generated_token_result::GeneratedTokenResult;
+use paddler_tests::agent_inference_runtime_context::agent_inference_runtime_context;
+use paddler_tests::kev_checkpoint::KevCheckpoint;
+use paddler_tests::llama_runtime::LlamaRuntime;
+
+use crate::decode_prompt_for_sampling::decode_prompt_for_sampling;
+use crate::ready_to_sample_request::ReadyToSampleRequest;
+
+const PROMPT: &str = "Paddler balances";
+
+#[test]
+fn a_sampler_chain_that_selects_no_candidate_ends_the_generation() {
+    let runtime = LlamaRuntime::load(
+        &KevCheckpoint::Kev0_8b.target_path("model.gguf"),
+        agent_inference_runtime_context(1),
+    )
+    .expect("the kev model must load");
+    let mut llama_context = runtime
+        .create_text_generation_llama_context(1)
+        .expect("the context must be created");
+    let batch_position = decode_prompt_for_sampling(&runtime, &mut llama_context, PROMPT);
+
+    assert_eq!(
+        ReadyToSampleRequest {
+            batch_position,
+            llama_context: &llama_context,
+            prompt: PROMPT,
+            runtime: &runtime,
+        }
+        .advance(|prepared_request| PreparedGenerationRequest {
+            token_sampling: TokenSampling {
+                chain: LlamaSampler::top_k(1).expect("the top-k sampler must initialize"),
+                grammar: GrammarSampling::Unconstrained,
+            },
+            ..prepared_request
+        }),
+        ContinuousBatchTerminalOutcome::EmitToClient(
+            GeneratedTokenResult::SamplingCandidatesExhausted(
+                "all token candidates were eliminated during sampling".to_owned()
+            )
+        )
+    );
+}
diff --git a/paddler_tests/tests/a_second_context_for_a_capability_cannot_be_created_while_one_exists.rs b/paddler_tests/tests/a_second_context_for_a_capability_cannot_be_created_while_one_exists.rs
new file mode 100644
index 000000000..0a6f0eddd
--- /dev/null
+++ b/paddler_tests/tests/a_second_context_for_a_capability_cannot_be_created_while_one_exists.rs
@@ -0,0 +1,32 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use paddler_agent_runtime::agent_runtime_error::AgentRuntimeError;
+use paddler_messaging::inference_capability::InferenceCapability;
+use paddler_tests::agent_inference_runtime_context::agent_inference_runtime_context;
+use paddler_tests::kev_checkpoint::KevCheckpoint;
+use paddler_tests::llama_runtime::LlamaRuntime;
+
+#[test]
+fn a_second_context_for_a_capability_cannot_be_created_while_one_exists() {
+    let llama_runtime = LlamaRuntime::load(
+        &KevCheckpoint::Kev0_8b.target_path("model.gguf"),
+        agent_inference_runtime_context(1),
+    )
+    .expect("the model must load");
+    let _resident_context = llama_runtime
+        .create_text_generation_llama_context(1)
+        .expect("the first context must be created");
+
+    assert_eq!(
+        llama_runtime
+            .create_text_generation_llama_context(1)
+            .err()
+            .map(|context_error| context_error.to_string()),
+        Some(
+            AgentRuntimeError::LlamaContextAlreadyResident {
+                inference_capability: InferenceCapability::TextGeneration,
+            }
+            .to_string()
+        )
+    );
+}
diff --git a/paddler_tests/tests/a_second_model_cannot_load_while_a_model_is_resident.rs b/paddler_tests/tests/a_second_model_cannot_load_while_a_model_is_resident.rs
new file mode 100644
index 000000000..c546bfe64
--- /dev/null
+++ b/paddler_tests/tests/a_second_model_cannot_load_while_a_model_is_resident.rs
@@ -0,0 +1,27 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use paddler_agent_runtime::agent_runtime_error::AgentRuntimeError;
+use paddler_agent_runtime::llama_model_settings::LlamaModelSettings;
+use paddler_agent_runtime::loaded_llama_model::LoadedLlamaModel;
+use paddler_inference_parameters::all_gpu_layers::ALL_GPU_LAYERS;
+use paddler_tests::agent_inference_runtime_context::agent_inference_runtime_context;
+use paddler_tests::kev_checkpoint::KevCheckpoint;
+
+#[test]
+fn a_second_model_cannot_load_while_a_model_is_resident() {
+    let inference_runtime_context = agent_inference_runtime_context(1);
+    let llama_model_settings = || LlamaModelSettings {
+        model_path: KevCheckpoint::Kev0_8b.target_path("model.gguf"),
+        n_gpu_layers: ALL_GPU_LAYERS,
+    };
+    let _resident_model =
+        LoadedLlamaModel::load(&inference_runtime_context, llama_model_settings())
+            .expect("the first model must load");
+
+    assert_eq!(
+        LoadedLlamaModel::load(&inference_runtime_context, llama_model_settings())
+            .err()
+            .map(|load_error| load_error.to_string()),
+        Some(AgentRuntimeError::ModelWeightsAlreadyResident.to_string())
+    );
+}
diff --git a/paddler_tests/tests/a_second_multimodal_projection_cannot_load_while_one_is_resident.rs b/paddler_tests/tests/a_second_multimodal_projection_cannot_load_while_one_is_resident.rs
new file mode 100644
index 000000000..ed76fccf1
--- /dev/null
+++ b/paddler_tests/tests/a_second_multimodal_projection_cannot_load_while_one_is_resident.rs
@@ -0,0 +1,48 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use paddler_agent_runtime::agent_runtime_error::AgentRuntimeError;
+use paddler_agent_text_generation::multimodal_projection::MultimodalProjection;
+use paddler_agent_text_generation::text_generation_error::TextGenerationError;
+use paddler_messaging::chat_template_source::ChatTemplateSource;
+use paddler_test_cluster_harness::model_card::smolvlm2_256m::smolvlm2_256m;
+use paddler_test_cluster_harness::model_card::smolvlm2_256m_mmproj::smolvlm2_256m_mmproj;
+use paddler_tests::agent_inference_runtime_context::agent_inference_runtime_context;
+use paddler_tests::llama_runtime::LlamaRuntime;
+use paddler_tests::text_generation_pipeline_parts::TextGenerationPipelineParts;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn a_second_multimodal_projection_cannot_load_while_one_is_resident() {
+    let llama_runtime = LlamaRuntime::load(
+        &smolvlm2_256m()
+            .cached_path()
+            .expect("the vision model must be in the Hugging Face cache"),
+        agent_inference_runtime_context(1),
+    )
+    .expect("the vision model must load");
+    let _resident_projection = llama_runtime
+        .inference_runtime_context
+        .vram_residencies
+        .multimodal_projection
+        .try_occupy()
+        .expect("the test must occupy the multimodal projection residency");
+
+    assert_eq!(
+        TextGenerationPipelineParts {
+            chat_template_source: ChatTemplateSource::EmbeddedInModel,
+            llama_runtime,
+            multimodal_projection: MultimodalProjection::File(
+                smolvlm2_256m_mmproj()
+                    .cached_path()
+                    .expect("the multimodal projection must be in the Hugging Face cache"),
+            ),
+        }
+        .spawn_refusal()
+        .await,
+        Some(
+            TextGenerationError::AgentRuntime(
+                AgentRuntimeError::MultimodalProjectionAlreadyResident
+            )
+            .to_string()
+        )
+    );
+}
diff --git a/paddler_tests/tests/a_shutdown_while_a_model_loads_cancels_the_load.rs b/paddler_tests/tests/a_shutdown_while_a_model_loads_cancels_the_load.rs
new file mode 100644
index 000000000..a83f7fe01
--- /dev/null
+++ b/paddler_tests/tests/a_shutdown_while_a_model_loads_cancels_the_load.rs
@@ -0,0 +1,51 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::mem::discriminant;
+use std::pin::pin;
+
+use futures_util::FutureExt as _;
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent_runtime::llama_model_settings::LlamaModelSettings;
+use paddler_agent_runtime::loaded_llama_model::LoadedLlamaModel;
+use paddler_agent_runtime::startup_outcome::StartupOutcome;
+use paddler_blocking_thread_fixture::occupied_blocking_thread::OccupiedBlockingThread;
+use paddler_blocking_thread_fixture::single_blocking_thread_runtime::single_blocking_thread_runtime;
+use paddler_inference_parameters::all_gpu_layers::ALL_GPU_LAYERS;
+use paddler_tests::agent_inference_runtime_context::agent_inference_runtime_context;
+use paddler_tests::kev_checkpoint::KevCheckpoint;
+
+#[test]
+fn a_shutdown_while_a_model_loads_cancels_the_load() {
+    single_blocking_thread_runtime()
+        .expect("the test runtime must build")
+        .block_on(async {
+            let agent_shutdown = CancellationToken::new();
+            let occupied_blocking_thread = OccupiedBlockingThread::occupy();
+            let mut model_load = pin!(LoadedLlamaModel::load_unless_shutting_down(
+                &agent_shutdown,
+                agent_inference_runtime_context(1),
+                LlamaModelSettings {
+                    model_path: KevCheckpoint::Kev0_8b.target_path("model.gguf"),
+                    n_gpu_layers: ALL_GPU_LAYERS,
+                },
+            ));
+
+            assert!(model_load.as_mut().now_or_never().is_none());
+
+            agent_shutdown.cancel();
+            occupied_blocking_thread
+                .release()
+                .await
+                .expect("the blocking thread occupant must be released");
+
+            assert_eq!(
+                discriminant(
+                    &model_load
+                        .await
+                        .expect("the model must load before the shutdown is noticed")
+                ),
+                discriminant(&StartupOutcome::Cancelled)
+            );
+        });
+}
diff --git a/paddler_tests/tests/a_single_question_decision_exceeding_its_slots_share_of_the_context_is_rejected.rs b/paddler_tests/tests/a_single_question_decision_exceeding_its_slots_share_of_the_context_is_rejected.rs
new file mode 100644
index 000000000..929a6a7fd
--- /dev/null
+++ b/paddler_tests/tests/a_single_question_decision_exceeding_its_slots_share_of_the_context_is_rejected.rs
@@ -0,0 +1,50 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::decision_answer::DecisionAnswer;
+use paddler_messaging::decision_result::DecisionResult;
+use paddler_messaging::decision_slots_minimum::DECISION_SLOTS_MINIMUM;
+use paddler_messaging::oversized_decision_details::OversizedDecisionDetails;
+use paddler_messaging::request_params::decide_params::raw_decide_params::RawDecideParams;
+use paddler_tests::sample_decision::sample_decision;
+use paddler_tests::start_decision_cluster::start_decision_cluster;
+
+use crate::small_context_decision_cluster_params::SMALL_DECISION_CONTEXT_SIZE;
+use crate::small_context_decision_cluster_params::small_context_decision_cluster_params;
+
+const STATE_REPETITIONS_EXCEEDING_ONE_SLOTS_SHARE: usize = 5;
+const TOKENS_REQUIRED_BY_THE_REPEATED_STATE_QUESTION: usize = 159;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn a_single_question_decision_exceeding_its_slots_share_of_the_context_is_rejected() {
+    let cluster = start_decision_cluster(small_context_decision_cluster_params())
+        .await
+        .expect("the cluster must start");
+    let RawDecideParams { questions, state } = sample_decision();
+
+    let collected = cluster
+        .decide(
+            CancellationToken::new(),
+            &RawDecideParams {
+                questions: questions.into_iter().take(1).collect(),
+                state: state.repeat(STATE_REPETITIONS_EXCEEDING_ONE_SLOTS_SHARE),
+            },
+        )
+        .await
+        .expect("the oversized decision must be answered with a rejection");
+
+    assert_eq!(collected.answers, Vec::<DecisionAnswer>::new());
+    assert_eq!(
+        collected.terminal_result,
+        DecisionResult::RequestExceedsContext(OversizedDecisionDetails {
+            context_size: SMALL_DECISION_CONTEXT_SIZE / u32::from(DECISION_SLOTS_MINIMUM.get()),
+            required_tokens: TOKENS_REQUIRED_BY_THE_REPEATED_STATE_QUESTION,
+        })
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/a_state_changing_only_its_sampling_keeps_its_decisions_without_reloading_the_pointer_head.rs b/paddler_tests/tests/a_state_changing_only_its_sampling_keeps_its_decisions_without_reloading_the_pointer_head.rs
new file mode 100644
index 000000000..e2946ebdf
--- /dev/null
+++ b/paddler_tests/tests/a_state_changing_only_its_sampling_keeps_its_decisions_without_reloading_the_pointer_head.rs
@@ -0,0 +1,58 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use paddler_inference_parameters::sampling_parameters::SamplingParameters;
+use paddler_messaging::agent_state_application_status::AgentStateApplicationStatus;
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_messaging::balancer_text_generation_settings::BalancerTextGenerationSettings;
+use paddler_messaging::decision_settings::DecisionSettings;
+use paddler_messaging::decision_slots_minimum::DECISION_SLOTS_MINIMUM;
+use paddler_messaging::produces_snapshot::ProducesSnapshot as _;
+use paddler_tests::kev_checkpoint::KevCheckpoint;
+
+use crate::serving_pipeline_arbiter::ServingPipelineArbiter;
+use crate::swappable_model_file::SwappableModelFile;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn a_state_changing_only_its_sampling_keeps_its_decisions_without_reloading_the_pointer_head()
+{
+    let swappable_pointer_head =
+        SwappableModelFile::link_to(&KevCheckpoint::Kev0_8b.target_path("pointer_head.gguf"));
+    let serving_decisions = BalancerDesiredState {
+        decision: DecisionSettings {
+            pointer_head: swappable_pointer_head.agent_desired_model(),
+        },
+        ..KevCheckpoint::Kev0_8b
+            .decision_desired_state()
+            .into_balancer_desired_state()
+            .expect("the kev desired state must name absolute paths")
+    };
+    let mut serving_pipeline_arbiter = ServingPipelineArbiter::new(DECISION_SLOTS_MINIMUM.get());
+
+    serving_pipeline_arbiter
+        .serve(serving_decisions.clone())
+        .await;
+    swappable_pointer_head.replace_with_an_unloadable_file();
+    serving_pipeline_arbiter
+        .apply(BalancerDesiredState {
+            text_generation: BalancerTextGenerationSettings {
+                sampling_parameters: SamplingParameters::deterministic(),
+                ..serving_decisions.text_generation.clone()
+            },
+            ..serving_decisions
+        })
+        .await;
+    serving_pipeline_arbiter.conclude_next_progress().await;
+    serving_pipeline_arbiter.conclude_next_progress().await;
+
+    assert_eq!(
+        serving_pipeline_arbiter
+            .inference_runtime_context
+            .slot_aggregated_status
+            .make_snapshot()
+            .status
+            .state_application_status,
+        AgentStateApplicationStatus::Applied
+    );
+
+    serving_pipeline_arbiter.shut_down().await;
+}
diff --git a/paddler_tests/tests/a_superseded_startup_releases_everything_it_held_when_the_agent_shuts_down.rs b/paddler_tests/tests/a_superseded_startup_releases_everything_it_held_when_the_agent_shuts_down.rs
new file mode 100644
index 000000000..cf3acff5b
--- /dev/null
+++ b/paddler_tests/tests/a_superseded_startup_releases_everything_it_held_when_the_agent_shuts_down.rs
@@ -0,0 +1,75 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::mem::discriminant;
+
+use paddler_agent_runtime::agent_vram_residencies::AgentVramResidencies;
+use paddler_inference_parameters::embedding_parameters::EmbeddingParameters;
+use paddler_inference_parameters::pooling_type::PoolingType;
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_messaging::decision_settings::DecisionSettings;
+use paddler_tests::desired_state_reconciled_discriminant::desired_state_reconciled_discriminant;
+use paddler_tests::kev_checkpoint::KevCheckpoint;
+
+use crate::serving_pipeline_arbiter::ServingPipelineArbiter;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn a_superseded_startup_releases_everything_it_held_when_the_agent_shuts_down() {
+    let superseded_desired_state = BalancerDesiredState {
+        decision: DecisionSettings::default(),
+        ..KevCheckpoint::Kev0_8b
+            .decision_desired_state()
+            .into_balancer_desired_state()
+            .expect("the kev desired state must name absolute paths")
+    };
+    let mut serving_pipeline_arbiter = ServingPipelineArbiter::new(1);
+    let AgentVramResidencies {
+        decisions_context,
+        embeddings_context,
+        model_weights,
+        multimodal_projection,
+        text_generation_context,
+    } = serving_pipeline_arbiter
+        .inference_runtime_context
+        .vram_residencies
+        .clone();
+
+    assert_eq!(
+        discriminant(
+            &serving_pipeline_arbiter
+                .apply(superseded_desired_state.clone())
+                .await
+        ),
+        desired_state_reconciled_discriminant()
+    );
+    serving_pipeline_arbiter.conclude_next_progress().await;
+    assert_eq!(
+        discriminant(
+            &serving_pipeline_arbiter
+                .apply(BalancerDesiredState {
+                    embeddings: EmbeddingParameters {
+                        pooling_type: PoolingType::Mean,
+                        ..superseded_desired_state.embeddings.clone()
+                    },
+                    ..superseded_desired_state
+                })
+                .await
+        ),
+        desired_state_reconciled_discriminant()
+    );
+    serving_pipeline_arbiter.conclude_next_progress().await;
+    serving_pipeline_arbiter.conclude_next_progress().await;
+    serving_pipeline_arbiter.conclude_next_progress().await;
+    serving_pipeline_arbiter.shut_down().await;
+
+    assert!(
+        [
+            decisions_context,
+            embeddings_context,
+            model_weights,
+            multimodal_projection,
+            text_generation_context,
+        ]
+        .iter()
+        .all(|vram_residency| vram_residency.try_occupy().is_ok())
+    );
+}
diff --git a/paddler_tests/tests/a_text_generation_pipeline_refuses_a_bos_token_it_cannot_render.rs b/paddler_tests/tests/a_text_generation_pipeline_refuses_a_bos_token_it_cannot_render.rs
new file mode 100644
index 000000000..a1c01cb71
--- /dev/null
+++ b/paddler_tests/tests/a_text_generation_pipeline_refuses_a_bos_token_it_cannot_render.rs
@@ -0,0 +1,46 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use llama_cpp_bindings::error::TokenToStringError;
+use tempfile::TempDir;
+
+use paddler_agent_text_generation::multimodal_projection::MultimodalProjection;
+use paddler_agent_text_generation::text_generation_error::TextGenerationError;
+use paddler_messaging::chat_template_source::ChatTemplateSource;
+use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
+use paddler_tests::agent_inference_runtime_context::agent_inference_runtime_context;
+use paddler_tests::llama_runtime::LlamaRuntime;
+use paddler_tests::point_qwen3_special_token_at_unused_padding::point_qwen3_special_token_at_unused_padding;
+use paddler_tests::text_generation_pipeline_parts::TextGenerationPipelineParts;
+use paddler_tests::write_patched_gguf::write_patched_gguf;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn a_text_generation_pipeline_refuses_a_bos_token_it_cannot_render() {
+    let model_directory = TempDir::new().expect("a temporary model directory must be creatable");
+    let model_path = model_directory
+        .path()
+        .join("qwen3-with-an-unrenderable-bos-token.gguf");
+
+    write_patched_gguf(
+        &qwen3_0_6b()
+            .cached_path()
+            .expect("Qwen3 must be in the Hugging Face cache"),
+        &model_path,
+        point_qwen3_special_token_at_unused_padding(c"tokenizer.ggml.bos_token_id"),
+    )
+    .expect("the model with an unrenderable BOS token must be writable");
+
+    assert_eq!(
+        TextGenerationPipelineParts {
+            chat_template_source: ChatTemplateSource::EmbeddedInModel,
+            llama_runtime: LlamaRuntime::load(&model_path, agent_inference_runtime_context(1))
+                .expect("the model must load"),
+            multimodal_projection: MultimodalProjection::NotConfigured,
+        }
+        .spawn_refusal()
+        .await,
+        Some(
+            TextGenerationError::SpecialTokenUnrenderable(TokenToStringError::UnknownTokenType)
+                .to_string()
+        )
+    );
+}
diff --git a/paddler_tests/tests/a_text_generation_pipeline_refuses_a_model_whose_embedded_template_hides_its_streaming_markers.rs b/paddler_tests/tests/a_text_generation_pipeline_refuses_a_model_whose_embedded_template_hides_its_streaming_markers.rs
new file mode 100644
index 000000000..d5141f36b
--- /dev/null
+++ b/paddler_tests/tests/a_text_generation_pipeline_refuses_a_model_whose_embedded_template_hides_its_streaming_markers.rs
@@ -0,0 +1,55 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use llama_cpp_bindings::error::MarkerDetectionError;
+use tempfile::TempDir;
+
+use paddler_agent_text_generation::multimodal_projection::MultimodalProjection;
+use paddler_agent_text_generation::text_generation_error::TextGenerationError;
+use paddler_messaging::chat_template_source::ChatTemplateSource;
+use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
+use paddler_tests::agent_inference_runtime_context::agent_inference_runtime_context;
+use paddler_tests::first_message_content_chat_template::first_message_content_chat_template;
+use paddler_tests::llama_runtime::LlamaRuntime;
+use paddler_tests::text_generation_pipeline_parts::TextGenerationPipelineParts;
+use paddler_tests::write_patched_gguf::write_patched_gguf;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn a_text_generation_pipeline_refuses_a_model_whose_embedded_template_hides_its_streaming_markers()
+ {
+    let model_directory = TempDir::new().expect("a temporary model directory must be creatable");
+    let model_path = model_directory
+        .path()
+        .join("qwen3-with-an-embedded-template-that-does-not-parse.gguf");
+
+    write_patched_gguf(
+        &qwen3_0_6b()
+            .cached_path()
+            .expect("Qwen3 must be in the Hugging Face cache"),
+        &model_path,
+        |gguf_context| {
+            gguf_context.set_val_str(c"tokenizer.chat_template", c"{% if %}");
+        },
+    )
+    .expect("the model with a template that does not parse must be writable");
+
+    assert_eq!(
+        TextGenerationPipelineParts {
+            chat_template_source: ChatTemplateSource::Override(
+                first_message_content_chat_template()
+            ),
+            llama_runtime: LlamaRuntime::load(&model_path, agent_inference_runtime_context(1))
+                .expect("the model must load"),
+            multimodal_projection: MultimodalProjection::NotConfigured,
+        }
+        .spawn_refusal()
+        .await,
+        Some(
+            TextGenerationError::StreamingMarkersUnavailable(
+                MarkerDetectionError::ReasoningMarkerDetectionFailed {
+                    message: "Unexpected token: %} of type 9".to_owned(),
+                }
+            )
+            .to_string()
+        )
+    );
+}
diff --git a/paddler_tests/tests/a_text_generation_pipeline_refuses_a_newline_token_it_cannot_render.rs b/paddler_tests/tests/a_text_generation_pipeline_refuses_a_newline_token_it_cannot_render.rs
new file mode 100644
index 000000000..0080a0aec
--- /dev/null
+++ b/paddler_tests/tests/a_text_generation_pipeline_refuses_a_newline_token_it_cannot_render.rs
@@ -0,0 +1,59 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use llama_cpp_bindings::error::TokenToStringError;
+use llama_cpp_bindings_sys::LLAMA_TOKEN_TYPE_UNUSED;
+use tempfile::TempDir;
+
+use paddler_agent_text_generation::multimodal_projection::MultimodalProjection;
+use paddler_agent_text_generation::text_generation_error::TextGenerationError;
+use paddler_messaging::chat_template_source::ChatTemplateSource;
+use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
+use paddler_tests::agent_inference_runtime_context::agent_inference_runtime_context;
+use paddler_tests::llama_runtime::LlamaRuntime;
+use paddler_tests::qwen3_0_6b_newline_token::QWEN3_0_6B_NEWLINE_TOKEN;
+use paddler_tests::text_generation_pipeline_parts::TextGenerationPipelineParts;
+use paddler_tests::write_patched_gguf::write_patched_gguf;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn a_text_generation_pipeline_refuses_a_newline_token_it_cannot_render() {
+    let model_directory = TempDir::new().expect("a temporary model directory must be creatable");
+    let model_path = model_directory
+        .path()
+        .join("qwen3-with-an-unrenderable-newline-token.gguf");
+
+    write_patched_gguf(
+        &qwen3_0_6b()
+            .cached_path()
+            .expect("Qwen3 must be in the Hugging Face cache"),
+        &model_path,
+        |gguf_context| {
+            let token_types_key_id = gguf_context
+                .find_key("tokenizer.ggml.token_type")
+                .expect("Qwen3 must declare its token types");
+            let mut token_types = gguf_context
+                .val_arr_i32(token_types_key_id)
+                .expect("Qwen3 token types must be an i32 array")
+                .to_vec();
+
+            token_types[QWEN3_0_6B_NEWLINE_TOKEN] = i32::try_from(LLAMA_TOKEN_TYPE_UNUSED)
+                .expect("the unused token type must fit the GGUF token type array");
+            gguf_context.set_arr_i32(c"tokenizer.ggml.token_type", &token_types);
+        },
+    )
+    .expect("the model with an unrenderable newline token must be writable");
+
+    assert_eq!(
+        TextGenerationPipelineParts {
+            chat_template_source: ChatTemplateSource::EmbeddedInModel,
+            llama_runtime: LlamaRuntime::load(&model_path, agent_inference_runtime_context(1))
+                .expect("the model must load"),
+            multimodal_projection: MultimodalProjection::NotConfigured,
+        }
+        .spawn_refusal()
+        .await,
+        Some(
+            TextGenerationError::SpecialTokenUnrenderable(TokenToStringError::UnknownTokenType)
+                .to_string()
+        )
+    );
+}
diff --git a/paddler_tests/tests/a_text_generation_pipeline_refuses_a_penalty_window_longer_than_its_sequence_context.rs b/paddler_tests/tests/a_text_generation_pipeline_refuses_a_penalty_window_longer_than_its_sequence_context.rs
new file mode 100644
index 000000000..fb91975e5
--- /dev/null
+++ b/paddler_tests/tests/a_text_generation_pipeline_refuses_a_penalty_window_longer_than_its_sequence_context.rs
@@ -0,0 +1,60 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent_text_generation::multimodal_projection::MultimodalProjection;
+use paddler_agent_text_generation::text_generation_error::TextGenerationError;
+use paddler_inference_parameters::invalid_inference_parameters::InvalidInferenceParameters;
+use paddler_inference_parameters::sampling_parameters::SamplingParameters;
+use paddler_messaging::chat_template_source::ChatTemplateSource;
+use paddler_test_cluster_harness::model_card::smolvlm2_256m::smolvlm2_256m;
+use paddler_tests::agent_inference_runtime_context::agent_inference_runtime_context;
+use paddler_tests::llama_runtime::LlamaRuntime;
+use paddler_tests::text_generation_pipeline_parts::TextGenerationPipelineParts;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn a_text_generation_pipeline_refuses_a_penalty_window_longer_than_its_sequence_context() {
+    let llama_runtime = LlamaRuntime::load(
+        &smolvlm2_256m()
+            .cached_path()
+            .expect("the model must be in the Hugging Face cache"),
+        agent_inference_runtime_context(1),
+    )
+    .expect("the model must load");
+    let sequence_context_size = llama_runtime
+        .create_text_generation_llama_context(1)
+        .expect("the context must be created")
+        .n_ctx_seq();
+    let mut text_generation_pipeline = TextGenerationPipelineParts {
+        chat_template_source: ChatTemplateSource::EmbeddedInModel,
+        llama_runtime,
+        multimodal_projection: MultimodalProjection::NotConfigured,
+    }
+    .into_text_generation_pipeline();
+
+    text_generation_pipeline
+        .text_generation_settings
+        .sampling_parameters = SamplingParameters {
+        penalty_last_n: i32::MAX,
+        penalty_repeat: 1.1,
+        ..SamplingParameters::default()
+    };
+
+    assert_eq!(
+        text_generation_pipeline
+            .spawn(&CancellationToken::new())
+            .await
+            .err()
+            .as_ref()
+            .map(ToString::to_string),
+        Some(
+            TextGenerationError::SamplingParametersUnfitForTheSequenceContext(
+                InvalidInferenceParameters::PenaltyWindowExceedsSequenceContext {
+                    penalty_last_n: i32::MAX,
+                    sequence_context_size,
+                }
+            )
+            .to_string()
+        )
+    );
+}
diff --git a/paddler_tests/tests/a_text_generation_pipeline_refuses_an_eos_token_it_cannot_render.rs b/paddler_tests/tests/a_text_generation_pipeline_refuses_an_eos_token_it_cannot_render.rs
new file mode 100644
index 000000000..6f8b0e09e
--- /dev/null
+++ b/paddler_tests/tests/a_text_generation_pipeline_refuses_an_eos_token_it_cannot_render.rs
@@ -0,0 +1,46 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use llama_cpp_bindings::error::TokenToStringError;
+use tempfile::TempDir;
+
+use paddler_agent_text_generation::multimodal_projection::MultimodalProjection;
+use paddler_agent_text_generation::text_generation_error::TextGenerationError;
+use paddler_messaging::chat_template_source::ChatTemplateSource;
+use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
+use paddler_tests::agent_inference_runtime_context::agent_inference_runtime_context;
+use paddler_tests::llama_runtime::LlamaRuntime;
+use paddler_tests::point_qwen3_special_token_at_unused_padding::point_qwen3_special_token_at_unused_padding;
+use paddler_tests::text_generation_pipeline_parts::TextGenerationPipelineParts;
+use paddler_tests::write_patched_gguf::write_patched_gguf;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn a_text_generation_pipeline_refuses_an_eos_token_it_cannot_render() {
+    let model_directory = TempDir::new().expect("a temporary model directory must be creatable");
+    let model_path = model_directory
+        .path()
+        .join("qwen3-with-an-unrenderable-eos-token.gguf");
+
+    write_patched_gguf(
+        &qwen3_0_6b()
+            .cached_path()
+            .expect("Qwen3 must be in the Hugging Face cache"),
+        &model_path,
+        point_qwen3_special_token_at_unused_padding(c"tokenizer.ggml.eos_token_id"),
+    )
+    .expect("the model with an unrenderable EOS token must be writable");
+
+    assert_eq!(
+        TextGenerationPipelineParts {
+            chat_template_source: ChatTemplateSource::EmbeddedInModel,
+            llama_runtime: LlamaRuntime::load(&model_path, agent_inference_runtime_context(1))
+                .expect("the model must load"),
+            multimodal_projection: MultimodalProjection::NotConfigured,
+        }
+        .spawn_refusal()
+        .await,
+        Some(
+            TextGenerationError::SpecialTokenUnrenderable(TokenToStringError::UnknownTokenType)
+                .to_string()
+        )
+    );
+}
diff --git a/paddler_tests/tests/a_text_generation_pipeline_refuses_to_start_while_its_context_is_resident.rs b/paddler_tests/tests/a_text_generation_pipeline_refuses_to_start_while_its_context_is_resident.rs
new file mode 100644
index 000000000..2c5484aea
--- /dev/null
+++ b/paddler_tests/tests/a_text_generation_pipeline_refuses_to_start_while_its_context_is_resident.rs
@@ -0,0 +1,44 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use paddler_agent_runtime::agent_runtime_error::AgentRuntimeError;
+use paddler_agent_text_generation::multimodal_projection::MultimodalProjection;
+use paddler_agent_text_generation::text_generation_error::TextGenerationError;
+use paddler_messaging::chat_template_source::ChatTemplateSource;
+use paddler_messaging::inference_capability::InferenceCapability;
+use paddler_test_cluster_harness::model_card::smolvlm2_256m::smolvlm2_256m;
+use paddler_tests::agent_inference_runtime_context::agent_inference_runtime_context;
+use paddler_tests::llama_runtime::LlamaRuntime;
+use paddler_tests::text_generation_pipeline_parts::TextGenerationPipelineParts;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn a_text_generation_pipeline_refuses_to_start_while_its_context_is_resident() {
+    let llama_runtime = LlamaRuntime::load(
+        &smolvlm2_256m()
+            .cached_path()
+            .expect("the model must be in the Hugging Face cache"),
+        agent_inference_runtime_context(1),
+    )
+    .expect("the model must load");
+    let _resident_text_generation_context = llama_runtime
+        .inference_runtime_context
+        .vram_residencies
+        .llama_context(InferenceCapability::TextGeneration)
+        .try_occupy()
+        .expect("the test must occupy the text generation context residency");
+
+    assert_eq!(
+        TextGenerationPipelineParts {
+            chat_template_source: ChatTemplateSource::EmbeddedInModel,
+            llama_runtime,
+            multimodal_projection: MultimodalProjection::NotConfigured,
+        }
+        .spawn_refusal()
+        .await,
+        Some(
+            TextGenerationError::AgentRuntime(AgentRuntimeError::LlamaContextAlreadyResident {
+                inference_capability: InferenceCapability::TextGeneration,
+            })
+            .to_string()
+        )
+    );
+}
diff --git a/paddler_tests/tests/a_token_the_sampler_cannot_read_ends_the_generation.rs b/paddler_tests/tests/a_token_the_sampler_cannot_read_ends_the_generation.rs
new file mode 100644
index 000000000..d11aceefb
--- /dev/null
+++ b/paddler_tests/tests/a_token_the_sampler_cannot_read_ends_the_generation.rs
@@ -0,0 +1,38 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use llama_cpp_bindings::error::LogitsError;
+
+use paddler_agent_text_generation::continuous_batch_terminal_outcome::ContinuousBatchTerminalOutcome;
+use paddler_agent_text_generation::text_generation_error::TextGenerationError;
+use paddler_messaging::generated_token_result::GeneratedTokenResult;
+use paddler_tests::agent_inference_runtime_context::agent_inference_runtime_context;
+use paddler_tests::kev_checkpoint::KevCheckpoint;
+use paddler_tests::llama_runtime::LlamaRuntime;
+
+use crate::ready_to_sample_request::ReadyToSampleRequest;
+
+#[test]
+fn a_token_the_sampler_cannot_read_ends_the_generation() {
+    let runtime = LlamaRuntime::load(
+        &KevCheckpoint::Kev0_8b.target_path("model.gguf"),
+        agent_inference_runtime_context(1),
+    )
+    .expect("the kev model must load");
+    let llama_context = runtime
+        .create_text_generation_llama_context(1)
+        .expect("the context must be created");
+
+    assert_eq!(
+        ReadyToSampleRequest {
+            batch_position: 0,
+            llama_context: &llama_context,
+            prompt: "Paddler",
+            runtime: &runtime,
+        }
+        .advance(|prepared_request| prepared_request),
+        ContinuousBatchTerminalOutcome::EmitToClient(GeneratedTokenResult::SamplerError(
+            TextGenerationError::TokenDataUnreadable(LogitsError::TokenNotInitialized(0))
+                .to_string()
+        ))
+    );
+}
diff --git a/paddler_tests/tests/a_tool_call_body_with_a_nul_byte_fails_to_parse.rs b/paddler_tests/tests/a_tool_call_body_with_a_nul_byte_fails_to_parse.rs
new file mode 100644
index 000000000..5456ab11f
--- /dev/null
+++ b/paddler_tests/tests/a_tool_call_body_with_a_nul_byte_fails_to_parse.rs
@@ -0,0 +1,43 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::ffi::CString;
+
+use llama_cpp_bindings::error::ParseChatMessageError;
+use llama_cpp_bindings::token_piece::TokenPiece;
+
+use paddler_messaging::generated_token_result::GeneratedTokenResult;
+use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
+use paddler_tests::agent_inference_runtime_context::agent_inference_runtime_context;
+use paddler_tests::llama_runtime::LlamaRuntime;
+
+use crate::tool_call_pipeline_for_tools::tool_call_pipeline_for_tools;
+
+const TOOL_CALL_BODY_WITH_A_NUL_BYTE: &str = "hello\0world";
+
+#[test]
+fn a_tool_call_body_with_a_nul_byte_fails_to_parse() {
+    let runtime = LlamaRuntime::load(
+        &qwen3_0_6b()
+            .cached_path()
+            .expect("the model must be in the Hugging Face cache"),
+        agent_inference_runtime_context(1),
+    )
+    .expect("the model must load");
+    let mut tool_call_pipeline =
+        tool_call_pipeline_for_tools(&runtime.loaded_llama_model.model, &[]);
+
+    tool_call_pipeline.feed(&TokenPiece::Visible(
+        TOOL_CALL_BODY_WITH_A_NUL_BYTE.to_owned(),
+    ));
+
+    assert_eq!(
+        tool_call_pipeline.finalize().into_generated_token_result(),
+        GeneratedTokenResult::ToolCallParseFailed(
+            ParseChatMessageError::InputContainsNulByte(
+                CString::new(TOOL_CALL_BODY_WITH_A_NUL_BYTE)
+                    .expect_err("a body with a NUL byte must not become a C string")
+            )
+            .to_string()
+        )
+    );
+}
diff --git a/paddler_tests/tests/a_tool_call_section_with_text_but_no_tool_calls_is_reported_as_unrecognized.rs b/paddler_tests/tests/a_tool_call_section_with_text_but_no_tool_calls_is_reported_as_unrecognized.rs
new file mode 100644
index 000000000..e8124ebe8
--- /dev/null
+++ b/paddler_tests/tests/a_tool_call_section_with_text_but_no_tool_calls_is_reported_as_unrecognized.rs
@@ -0,0 +1,56 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::generated_token_result::GeneratedTokenResult;
+use paddler_messaging::raw_tool_call_tokens::RawToolCallTokens;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_tests::gbnf_literal::gbnf_literal;
+use paddler_tests::get_weather_tool::get_weather_tool;
+use paddler_tests::max_tokens_for_forced_text::max_tokens_for_forced_text;
+use paddler_tests::start_cluster_with_gemma_4::start_cluster_with_gemma_4;
+use paddler_tests::weather_question_conversation::weather_question_conversation;
+
+const FORCED_OUTPUT: &str = "<|tool_call>call:Sure, let me check.<tool_call|>";
+
+#[tokio::test(flavor = "multi_thread")]
+async fn a_tool_call_section_with_text_but_no_tool_calls_is_reported_as_unrecognized() {
+    let cluster = start_cluster_with_gemma_4(AgentConfig::uniform(1, 1))
+        .await
+        .expect("the cluster must start");
+
+    let token_results = cluster
+        .continue_from_conversation_history(
+            CancellationToken::new(),
+            &weather_question_conversation(
+                gbnf_literal(FORCED_OUTPUT),
+                max_tokens_for_forced_text(FORCED_OUTPUT),
+                vec![get_weather_tool()],
+            ),
+        )
+        .await
+        .expect("the inference request must be accepted")
+        .into_token_results();
+
+    assert_eq!(
+        token_results
+            .iter()
+            .filter(|token_result| token_result.is_tool_call_parsed()
+                || matches!(
+                    token_result,
+                    GeneratedTokenResult::UnrecognizedToolCallFormat(_)
+                ))
+            .collect::<Vec<_>>(),
+        vec![&GeneratedTokenResult::UnrecognizedToolCallFormat(
+            RawToolCallTokens {
+                text: FORCED_OUTPUT.to_owned(),
+                ffi_error_message: String::new(),
+            }
+        )]
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/a_tool_call_the_chat_parser_cannot_match_is_reported_as_unrecognized.rs b/paddler_tests/tests/a_tool_call_the_chat_parser_cannot_match_is_reported_as_unrecognized.rs
new file mode 100644
index 000000000..818c94dfe
--- /dev/null
+++ b/paddler_tests/tests/a_tool_call_the_chat_parser_cannot_match_is_reported_as_unrecognized.rs
@@ -0,0 +1,47 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use llama_cpp_bindings::token_piece::TokenPiece;
+use serde_json::Value;
+
+use paddler_messaging::generated_token_result::GeneratedTokenResult;
+use paddler_messaging::raw_tool_call_tokens::RawToolCallTokens;
+use paddler_messaging::request_params::continue_from_conversation_history_params::tool::Tool;
+use paddler_test_cluster_harness::model_card::ministral_3_3b_reasoning::ministral_3_3b_reasoning;
+use paddler_tests::agent_inference_runtime_context::agent_inference_runtime_context;
+use paddler_tests::get_weather_tool::get_weather_tool;
+use paddler_tests::llama_runtime::LlamaRuntime;
+
+use crate::tool_call_pipeline_for_tools::tool_call_pipeline_for_tools;
+
+const TOOL_CALL_OPEN_MARKER: &str = "[TOOL_CALLS]";
+const TOOL_CALL_WITH_ARGUMENTS_THAT_ARE_NOT_JSON: &str = "get_weather[ARGS]not json";
+
+#[test]
+fn a_tool_call_the_chat_parser_cannot_match_is_reported_as_unrecognized() {
+    let runtime = LlamaRuntime::load(
+        &ministral_3_3b_reasoning()
+            .cached_path()
+            .expect("the model must be in the Hugging Face cache"),
+        agent_inference_runtime_context(1),
+    )
+    .expect("the model must load");
+    let mut tool_call_pipeline = tool_call_pipeline_for_tools(
+        &runtime.loaded_llama_model.model,
+        &[Tool::<Value>::from(get_weather_tool())],
+    );
+
+    tool_call_pipeline.feed(&TokenPiece::Marker(TOOL_CALL_OPEN_MARKER.to_owned()));
+    tool_call_pipeline.feed(&TokenPiece::Visible(
+        TOOL_CALL_WITH_ARGUMENTS_THAT_ARE_NOT_JSON.to_owned(),
+    ));
+
+    assert_eq!(
+        tool_call_pipeline.finalize().into_generated_token_result(),
+        GeneratedTokenResult::UnrecognizedToolCallFormat(RawToolCallTokens {
+            text: format!("{TOOL_CALL_OPEN_MARKER}{TOOL_CALL_WITH_ARGUMENTS_THAT_ARE_NOT_JSON}"),
+            ffi_error_message:
+                "The model produced output that does not match the expected peg-native format"
+                    .to_owned(),
+        })
+    );
+}
diff --git a/paddler_tests/tests/a_waiting_decision_is_not_overtaken_by_later_completions.rs b/paddler_tests/tests/a_waiting_decision_is_not_overtaken_by_later_completions.rs
new file mode 100644
index 000000000..b59c8ad11
--- /dev/null
+++ b/paddler_tests/tests/a_waiting_decision_is_not_overtaken_by_later_completions.rs
@@ -0,0 +1,72 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::mem::discriminant;
+use std::num::NonZeroU32;
+
+use tokio::spawn;
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
+use paddler_tests::decision_cluster_params::DecisionClusterParams;
+use paddler_tests::decision_done_discriminant::decision_done_discriminant;
+use paddler_tests::sample_decision::sample_decision;
+use paddler_tests::start_decision_cluster::start_decision_cluster;
+
+use crate::slot_held_by_an_unending_generation::SlotHeldByAnUnendingGeneration;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn a_waiting_decision_is_not_overtaken_by_later_completions() {
+    let mut cluster = start_decision_cluster(DecisionClusterParams::default())
+        .await
+        .expect("a decision cluster must start");
+    let slot_held_by_an_unending_generation =
+        SlotHeldByAnUnendingGeneration::hold(&mut cluster).await;
+    let decision = spawn(cluster.decide(CancellationToken::new(), &sample_decision()));
+
+    cluster
+        .wait_for_buffered_request_count(1)
+        .await
+        .expect("the decision must wait while only one slot is free");
+
+    let completion = spawn(cluster.continue_from_raw_prompt(
+        CancellationToken::new(),
+        &ContinueFromRawPromptParams {
+            grammar: None,
+            max_tokens: NonZeroU32::MIN,
+            raw_prompt: "The capital of France is".to_owned(),
+        },
+    ));
+
+    cluster
+        .wait_for_buffered_request_count(2)
+        .await
+        .expect("the completion must wait behind the decision although a slot is free");
+    slot_held_by_an_unending_generation.release().await;
+
+    assert_eq!(
+        discriminant(
+            &decision
+                .await
+                .expect("the decision task must not panic")
+                .expect("the decision must be answered")
+                .terminal_result
+        ),
+        decision_done_discriminant()
+    );
+    assert_eq!(
+        completion
+            .await
+            .expect("the completion task must not panic")
+            .expect("the completion must be answered")
+            .summary()
+            .expect("the completion must finish with a summary")
+            .usage
+            .completion_tokens(),
+        1
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/a_warm_up_llama_cpp_rejects_leaves_the_context_decoding.rs b/paddler_tests/tests/a_warm_up_llama_cpp_rejects_leaves_the_context_decoding.rs
new file mode 100644
index 000000000..67ba15386
--- /dev/null
+++ b/paddler_tests/tests/a_warm_up_llama_cpp_rejects_leaves_the_context_decoding.rs
@@ -0,0 +1,51 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use llama_cpp_bindings::SampledToken;
+use llama_cpp_bindings::llama_batch::LlamaBatch;
+use llama_cpp_bindings::token::LlamaToken;
+
+use paddler_tests::agent_inference_runtime_context::agent_inference_runtime_context;
+use paddler_tests::kev_checkpoint::KevCheckpoint;
+use paddler_tests::llama_runtime::LlamaRuntime;
+
+const SEQUENCES_THE_CONTEXT_HOLDS: u32 = 1;
+const SLOTS_THE_WARM_UP_PRIMES: u16 = 2;
+
+#[test]
+fn a_warm_up_llama_cpp_rejects_leaves_the_context_decoding() {
+    let runtime = LlamaRuntime::load(
+        &KevCheckpoint::Kev0_8b.target_path("model.gguf"),
+        agent_inference_runtime_context(1),
+    )
+    .expect("the kev model must load");
+    let mut llama_context = runtime
+        .create_text_generation_llama_context(SEQUENCES_THE_CONTEXT_HOLDS)
+        .expect("the context must be created");
+    let mut batch = LlamaBatch::new(runtime.model_runtime_parameters.n_batch.capacity(), 1);
+
+    let warm_up = runtime.loaded_llama_model.warm_up_llama_context(
+        &mut llama_context,
+        &mut batch,
+        SLOTS_THE_WARM_UP_PRIMES,
+    );
+
+    assert_eq!(
+        warm_up.map_err(|warm_up_error| warm_up_error.to_string()),
+        Err("llama.cpp rejected the warm-up batch: decode batch is invalid (empty, output mismatch, or initialization failure)".to_owned())
+    );
+
+    batch.clear();
+    batch
+        .add(&SampledToken::Content(LlamaToken::new(0)), 0, &[0], true)
+        .expect("a single token must fit the batch");
+    llama_context
+        .decode(&mut batch)
+        .expect("the context must keep decoding after the warm-up was rejected");
+
+    assert_eq!(
+        llama_context
+            .kv_cache_seq_pos_max(0)
+            .expect("the sequence position must be readable"),
+        0
+    );
+}
diff --git a/paddler_tests/tests/a_warm_up_spanning_more_tokens_than_its_batch_holds_primes_every_slot.rs b/paddler_tests/tests/a_warm_up_spanning_more_tokens_than_its_batch_holds_primes_every_slot.rs
new file mode 100644
index 000000000..b20a9ada6
--- /dev/null
+++ b/paddler_tests/tests/a_warm_up_spanning_more_tokens_than_its_batch_holds_primes_every_slot.rs
@@ -0,0 +1,39 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use llama_cpp_bindings::llama_batch::LlamaBatch;
+
+use paddler_inference_parameters::batch_size::BatchSize;
+use paddler_tests::agent_inference_runtime_context::agent_inference_runtime_context;
+use paddler_tests::kev_checkpoint::KevCheckpoint;
+use paddler_tests::llama_runtime::LlamaRuntime;
+
+const SLOTS: u16 = 2;
+
+#[test]
+fn a_warm_up_spanning_more_tokens_than_its_batch_holds_primes_every_slot() {
+    let mut runtime = LlamaRuntime::load(
+        &KevCheckpoint::Kev0_8b.target_path("model.gguf"),
+        agent_inference_runtime_context(SLOTS),
+    )
+    .expect("the kev model must load");
+
+    runtime.model_runtime_parameters.n_batch =
+        BatchSize::try_from(u32::from(SLOTS)).expect("a batch of one token per slot is valid");
+
+    let mut llama_context = runtime
+        .create_text_generation_llama_context(u32::from(SLOTS))
+        .expect("the context must be created");
+    let mut batch = LlamaBatch::new(
+        runtime.model_runtime_parameters.n_batch.capacity(),
+        i32::from(SLOTS),
+    );
+
+    assert_eq!(
+        runtime
+            .loaded_llama_model
+            .warm_up_llama_context(&mut llama_context, &mut batch, SLOTS)
+            .as_ref()
+            .map_err(ToString::to_string),
+        Ok(&())
+    );
+}
diff --git a/paddler_tests/tests/a_warm_up_whose_batch_cannot_hold_every_slot_is_refused.rs b/paddler_tests/tests/a_warm_up_whose_batch_cannot_hold_every_slot_is_refused.rs
new file mode 100644
index 000000000..075db78a0
--- /dev/null
+++ b/paddler_tests/tests/a_warm_up_whose_batch_cannot_hold_every_slot_is_refused.rs
@@ -0,0 +1,34 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::num::NonZeroU16;
+
+use llama_cpp_bindings::llama_batch::LlamaBatch;
+use llama_cpp_bindings_types::BatchCapacity;
+
+use paddler_tests::agent_inference_runtime_context::agent_inference_runtime_context;
+use paddler_tests::kev_checkpoint::KevCheckpoint;
+use paddler_tests::llama_runtime::LlamaRuntime;
+
+const SLOTS: u16 = 2;
+const TOKENS_THE_BATCH_HOLDS: NonZeroU16 = NonZeroU16::MIN;
+
+#[test]
+fn a_warm_up_whose_batch_cannot_hold_every_slot_is_refused() {
+    let runtime = LlamaRuntime::load(
+        &KevCheckpoint::Kev0_8b.target_path("model.gguf"),
+        agent_inference_runtime_context(SLOTS),
+    )
+    .expect("the kev model must load");
+    let mut llama_context = runtime
+        .create_text_generation_llama_context(u32::from(SLOTS))
+        .expect("the context must be created");
+    let mut batch = LlamaBatch::new(BatchCapacity::from(TOKENS_THE_BATCH_HOLDS), 1);
+
+    assert_eq!(
+        runtime
+            .loaded_llama_model
+            .warm_up_llama_context(&mut llama_context, &mut batch, SLOTS)
+            .map_err(|warm_up_error| warm_up_error.to_string()),
+        Err("unable to assemble the warm-up batch: Insufficient Space of 1".to_owned())
+    );
+}
diff --git a/paddler_tests/tests/agent_answer_after_a_notification.rs b/paddler_tests/tests/agent_answer_after_a_notification.rs
new file mode 100644
index 000000000..35a0dfa1e
--- /dev/null
+++ b/paddler_tests/tests/agent_answer_after_a_notification.rs
@@ -0,0 +1,47 @@
+use serde_json::to_string;
+use tokio_tungstenite::tungstenite::Message;
+
+use paddler_messaging::jsonrpc::request_envelope::RequestEnvelope;
+use paddler_messaging::jsonrpc::response_envelope::ResponseEnvelope;
+use paddler_messaging::management_socket::agent::message::Message as AgentJsonRpcMessage;
+use paddler_messaging::management_socket::agent::notification::Notification as AgentJsonRpcNotification;
+use paddler_messaging::management_socket::agent::request::Request as AgentJsonRpcRequest;
+use paddler_messaging::management_socket::agent::response::Response as AgentJsonRpcResponse;
+use paddler_test_cluster_harness::untrusted_balancer_socket_server::UntrustedBalancerSocketServer;
+
+use crate::agent_runner_connected_to::agent_runner_connected_to;
+
+pub const ANSWERED_REQUEST_ID: &str = "model-metadata";
+
+pub async fn agent_answer_after_a_notification(
+    notification: AgentJsonRpcNotification,
+) -> ResponseEnvelope<AgentJsonRpcResponse> {
+    let untrusted_balancer = UntrustedBalancerSocketServer::bind()
+        .await
+        .expect("the untrusted balancer must bind");
+    let _agent_runner = agent_runner_connected_to(untrusted_balancer.management_addr, 1);
+    let mut connection = untrusted_balancer
+        .accept_agent()
+        .await
+        .expect("the agent must connect");
+
+    for agent_message in [
+        AgentJsonRpcMessage::Notification(notification),
+        AgentJsonRpcMessage::Request(RequestEnvelope {
+            id: ANSWERED_REQUEST_ID.to_owned(),
+            request: AgentJsonRpcRequest::GetModelMetadata,
+        }),
+    ] {
+        connection
+            .send(Message::text(
+                to_string(&agent_message).expect("the agent message must serialize"),
+            ))
+            .await
+            .expect("the agent message must be sent");
+    }
+
+    connection
+        .next_agent_response()
+        .await
+        .expect("the agent must answer over the same connection")
+}
diff --git a/paddler_tests/tests/agent_cannot_parse_tool_calls_for_a_model_without_a_chat_template.rs b/paddler_tests/tests/agent_cannot_parse_tool_calls_for_a_model_without_a_chat_template.rs
new file mode 100644
index 000000000..ec96fb9af
--- /dev/null
+++ b/paddler_tests/tests/agent_cannot_parse_tool_calls_for_a_model_without_a_chat_template.rs
@@ -0,0 +1,68 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::num::NonZeroU32;
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::chat_template::ChatTemplate;
+use paddler_messaging::conversation_history::ConversationHistory;
+use paddler_messaging::conversation_message::ConversationMessage;
+use paddler_messaging::conversation_message_content::ConversationMessageContent;
+use paddler_messaging::generated_token_result::GeneratedTokenResult;
+use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_tests::get_weather_tool::get_weather_tool;
+use paddler_tests::nomic_embed_desired_state_with_chat_template_override::nomic_embed_desired_state_with_chat_template_override;
+use paddler_tests::start_cluster::start_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agent_cannot_parse_tool_calls_for_a_model_without_a_chat_template() {
+    let cluster = start_cluster(ClusterParams {
+        agents: AgentConfig::uniform(1, 1),
+        desired_state: ClusterDesiredState::Apply(Box::new(
+            nomic_embed_desired_state_with_chat_template_override(ChatTemplate {
+                content: "{{ messages[0].content }}".to_owned(),
+            }),
+        )),
+        wait_for_slots_ready: true,
+        ..ClusterParams::default()
+    })
+    .await
+    .expect("the cluster must start");
+
+    let collected = cluster
+        .continue_from_conversation_history(
+            CancellationToken::new(),
+            &ContinueFromConversationHistoryParams {
+                add_generation_prompt: true,
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Text("Hello".to_owned()),
+                        role: "user".to_owned(),
+                    },
+                ]),
+                enable_thinking: false,
+                grammar: None,
+                max_tokens: NonZeroU32::new(4).unwrap(),
+                parse_tool_calls: true,
+                tools: vec![get_weather_tool()],
+            },
+        )
+        .await
+        .expect("the inference request must be accepted");
+
+    assert_eq!(
+        collected.into_token_results(),
+        vec![GeneratedTokenResult::ChatTemplateError(
+            "test-agent-0: failed to build this model's tool call parser: model has no chat template"
+                .to_owned()
+        )]
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/agent_chunks_embedding_batch_larger_than_slot_count.rs b/paddler_tests/tests/agent_chunks_embedding_batch_larger_than_slot_count.rs
deleted file mode 100644
index f7e3fedff..000000000
--- a/paddler_tests/tests/agent_chunks_embedding_batch_larger_than_slot_count.rs
+++ /dev/null
@@ -1,69 +0,0 @@
-#![cfg(feature = "tests_that_use_llms")]
-
-use std::collections::BTreeSet;
-
-use tokio_util::sync::CancellationToken;
-
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
-use paddler_messaging::embedding_input_document::EmbeddingInputDocument;
-use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
-use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
-use paddler_test_cluster_harness::agent_config::AgentConfig;
-use paddler_test_cluster_harness::embedding_cluster_params::EmbeddingClusterParams;
-use paddler_tests::start_embedding_cluster::start_embedding_cluster;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn agent_chunks_embedding_batch_larger_than_slot_count() {
-    let cluster = start_embedding_cluster(EmbeddingClusterParams {
-        agents: vec![AgentConfig::single(4)],
-        inference_parameters: InferenceParameters {
-            enable_embeddings: true,
-            ..InferenceParameters::deterministic()
-        },
-        ..EmbeddingClusterParams::default()
-    })
-    .await
-    .expect("the cluster must start");
-
-    let input_batch: Vec<EmbeddingInputDocument> = (0..12)
-        .map(|index| EmbeddingInputDocument {
-            content: format!("Document number {index}."),
-            id: format!("doc-{index}"),
-        })
-        .collect();
-
-    let collected = cluster
-        .generate_embedding_batch(
-            CancellationToken::new(),
-            &GenerateEmbeddingBatchParams {
-                input_batch,
-                normalization_method: EmbeddingNormalizationMethod::None,
-            },
-        )
-        .await
-        .expect("the embedding batch must be accepted");
-
-    assert_eq!(collected.embeddings.len(), 12);
-    assert!(collected.saw_done);
-    assert!(collected.errors.is_empty());
-
-    let returned_ids: BTreeSet<String> = collected
-        .embeddings
-        .iter()
-        .map(|produced| produced.embedding.source_document_id.clone())
-        .collect();
-    let expected_ids: BTreeSet<String> = (0..12).map(|index| format!("doc-{index}")).collect();
-
-    assert_eq!(returned_ids, expected_ids);
-
-    let first_dimension = collected.embeddings[0].embedding.embedding.len();
-
-    for produced in &collected.embeddings {
-        assert_eq!(produced.embedding.embedding.len(), first_dimension);
-    }
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_tests/tests/agent_clears_a_pointer_head_incompatibility_once_the_replaced_file_fits_the_model.rs b/paddler_tests/tests/agent_clears_a_pointer_head_incompatibility_once_the_replaced_file_fits_the_model.rs
new file mode 100644
index 000000000..3beb4fcf4
--- /dev/null
+++ b/paddler_tests/tests/agent_clears_a_pointer_head_incompatibility_once_the_replaced_file_fits_the_model.rs
@@ -0,0 +1,65 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::collections::BTreeSet;
+
+use crate::replaceable_pointer_head::ReplaceablePointerHead;
+use paddler_messaging::agent_decisions_serving::AgentDecisionsServing;
+use paddler_messaging::agent_issue::AgentIssue;
+use paddler_messaging::agent_issue_params::model_path::ModelPath;
+use paddler_messaging::agent_issue_params::pointer_head_incompatibility::PointerHeadIncompatibility;
+use paddler_messaging::agent_issue_params::pointer_head_incompatible_with_model_params::PointerHeadIncompatibleWithModelParams;
+use paddler_test_cluster_harness::synthetic_pointer_head_fixture::QWEN3_5_0_8B_SYNTHETIC_POINTER_HEAD_FIXTURE;
+use paddler_tests::decision_cluster_params::DecisionClusterParams;
+use paddler_tests::decision_desired_state::DecisionDesiredState;
+use paddler_tests::start_decision_cluster::start_decision_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agent_clears_a_pointer_head_incompatibility_once_the_replaced_file_fits_the_model() {
+    let pointer_head = ReplaceablePointerHead::holding_fixture(
+        "qwen3_5_0_8b_synthetic_pointer_head_with_wrong_hidden_size.gguf",
+    );
+    let hidden_size_mismatch =
+        AgentIssue::PointerHeadIncompatibleWithModel(PointerHeadIncompatibleWithModelParams {
+            incompatibility: PointerHeadIncompatibility::HiddenSizeMismatch {
+                model_hidden_size: 1024,
+                pointer_head_hidden_size: 512,
+            },
+            pointer_head_path: ModelPath::from(pointer_head.path.as_path()),
+        });
+    let mut cluster = start_decision_cluster(DecisionClusterParams {
+        desired_state: DecisionDesiredState {
+            pointer_head: pointer_head.path.clone(),
+            ..DecisionDesiredState::default()
+        },
+        wait_for_slots_ready: false,
+        ..DecisionClusterParams::default()
+    })
+    .await
+    .expect("the cluster must start");
+
+    cluster
+        .wait_for_first_agent_issue(|issue| *issue == hidden_size_mismatch)
+        .await
+        .expect("the agent must report the incompatible pointer head");
+
+    pointer_head.replace_with_fixture(QWEN3_5_0_8B_SYNTHETIC_POINTER_HEAD_FIXTURE);
+
+    let serving_snapshot = cluster
+        .wait_for_first_agent_to_serve(AgentDecisionsServing::Served)
+        .await
+        .expect("the agent must serve decisions with the replaced pointer head");
+
+    assert_eq!(
+        serving_snapshot
+            .agents
+            .into_iter()
+            .map(|agent| agent.status.issues)
+            .collect::<Vec<_>>(),
+        vec![BTreeSet::new()]
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/agent_clears_a_pointer_head_load_issue_once_the_repaired_file_loads.rs b/paddler_tests/tests/agent_clears_a_pointer_head_load_issue_once_the_repaired_file_loads.rs
new file mode 100644
index 000000000..c708f74f0
--- /dev/null
+++ b/paddler_tests/tests/agent_clears_a_pointer_head_load_issue_once_the_repaired_file_loads.rs
@@ -0,0 +1,55 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::collections::BTreeSet;
+
+use crate::replaceable_pointer_head::ReplaceablePointerHead;
+use paddler_messaging::agent_decisions_serving::AgentDecisionsServing;
+use paddler_messaging::agent_issue::AgentIssue;
+use paddler_messaging::agent_issue_params::model_path::ModelPath;
+use paddler_test_cluster_harness::synthetic_pointer_head_fixture::QWEN3_5_0_8B_SYNTHETIC_POINTER_HEAD_FIXTURE;
+use paddler_tests::decision_cluster_params::DecisionClusterParams;
+use paddler_tests::decision_desired_state::DecisionDesiredState;
+use paddler_tests::start_decision_cluster::start_decision_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agent_clears_a_pointer_head_load_issue_once_the_repaired_file_loads() {
+    let pointer_head = ReplaceablePointerHead::holding_fixture("invalid.gguf");
+    let unloadable_pointer_head =
+        AgentIssue::PointerHeadCannotBeLoaded(ModelPath::from(pointer_head.path.as_path()));
+    let mut cluster = start_decision_cluster(DecisionClusterParams {
+        desired_state: DecisionDesiredState {
+            pointer_head: pointer_head.path.clone(),
+            ..DecisionDesiredState::default()
+        },
+        wait_for_slots_ready: false,
+        ..DecisionClusterParams::default()
+    })
+    .await
+    .expect("the cluster must start");
+
+    cluster
+        .wait_for_first_agent_issue(|issue| *issue == unloadable_pointer_head)
+        .await
+        .expect("the agent must report the unreadable pointer head");
+
+    pointer_head.replace_with_fixture(QWEN3_5_0_8B_SYNTHETIC_POINTER_HEAD_FIXTURE);
+
+    let serving_snapshot = cluster
+        .wait_for_first_agent_to_serve(AgentDecisionsServing::Served)
+        .await
+        .expect("the agent must serve decisions with the repaired pointer head");
+
+    assert_eq!(
+        serving_snapshot
+            .agents
+            .into_iter()
+            .map(|agent| agent.status.issues)
+            .collect::<Vec<_>>(),
+        vec![BTreeSet::new()]
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/agent_clears_the_chat_template_issue_once_a_valid_template_replaces_it.rs b/paddler_tests/tests/agent_clears_the_chat_template_issue_once_a_valid_template_replaces_it.rs
new file mode 100644
index 000000000..0e4d2970b
--- /dev/null
+++ b/paddler_tests/tests/agent_clears_the_chat_template_issue_once_a_valid_template_replaces_it.rs
@@ -0,0 +1,89 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use tokio_util::sync::CancellationToken;
+
+use std::mem::discriminant;
+
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_messaging::chat_template::ChatTemplate;
+use paddler_model_source::model_source::ModelSource;
+use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
+use paddler_tests::desired_state_with_chat_template_override::desired_state_with_chat_template_override;
+use paddler_tests::start_single_agent_cluster_with_desired_state::start_single_agent_cluster_with_desired_state;
+use paddler_tests::uncompilable_chat_template::uncompilable_chat_template;
+use paddler_tests::uncompilable_chat_template_issue::uncompilable_chat_template_issue;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agent_clears_the_chat_template_issue_once_a_valid_template_replaces_it() {
+    let model_card = qwen3_0_6b();
+    let expected_issue = uncompilable_chat_template_issue(
+        &model_card
+            .cached_path()
+            .expect("the model must be in the Hugging Face cache"),
+    )
+    .expect("the uncompilable chat template must not compile");
+    let mut cluster =
+        start_single_agent_cluster_with_desired_state(desired_state_with_chat_template_override(
+            BalancerDesiredState {
+                model: ModelSource::HuggingFace(model_card.reference.clone())
+                    .into_agent_desired_model(),
+                ..BalancerDesiredState::default()
+            },
+            uncompilable_chat_template(),
+        ))
+        .await
+        .expect("a single-agent cluster must start");
+
+    let snapshot = cluster
+        .wait_for_first_agent_issue(|issue| discriminant(issue) == discriminant(&expected_issue))
+        .await
+        .expect("the agent must report ChatTemplateDoesNotCompile for invalid Jinja syntax");
+    let issues = &snapshot.agents[0].status.issues;
+
+    assert!(issues.contains(&expected_issue), "{issues:?}");
+
+    cluster
+        .client_management
+        .put_balancer_desired_state(
+            CancellationToken::new(),
+            &desired_state_with_chat_template_override(
+                BalancerDesiredState {
+                    model: ModelSource::HuggingFace(model_card.reference)
+                        .into_agent_desired_model(),
+                    ..BalancerDesiredState::default()
+                },
+                ChatTemplate {
+                    content: "{% for message in messages %}{{ message.content }}{% endfor %}"
+                        .to_owned(),
+                },
+            ),
+        )
+        .await
+        .expect("the balancer must accept the desired state with a valid template");
+
+    let agent_id = cluster
+        .agent_ids
+        .first()
+        .expect("the cluster must have a registered agent")
+        .clone();
+
+    cluster
+        .agents_watcher
+        .until_agent(&agent_id, |snapshot| {
+            snapshot.agents.iter().any(|agent| {
+                agent.id == agent_id
+                    && !agent
+                        .status
+                        .issues
+                        .iter()
+                        .any(|issue| discriminant(issue) == discriminant(&expected_issue))
+            })
+        })
+        .await
+        .expect("the agent must stay registered while reconciliation clears its template issue");
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/agent_completes_generation_with_adequate_n_batch.rs b/paddler_tests/tests/agent_completes_generation_with_adequate_n_batch.rs
deleted file mode 100644
index e4a3bd66e..000000000
--- a/paddler_tests/tests/agent_completes_generation_with_adequate_n_batch.rs
+++ /dev/null
@@ -1,96 +0,0 @@
-#![cfg(feature = "tests_that_use_llms")]
-
-use std::future::Future;
-use std::num::NonZeroU32;
-
-use anyhow::Result;
-use tokio_util::sync::CancellationToken;
-
-use paddler_messaging::conversation_history::ConversationHistory;
-use paddler_messaging::conversation_message::ConversationMessage;
-use paddler_messaging::conversation_message_content::ConversationMessageContent;
-use paddler_messaging::conversation_message_content_part::ConversationMessageContentPart;
-use paddler_messaging::image_url::ImageUrl;
-use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
-use paddler_test_cluster_harness::agent_config::AgentConfig;
-use paddler_test_cluster_harness::cluster::Cluster;
-use paddler_test_cluster_harness::load_fixture_data_uri::load_fixture_data_uri;
-use paddler_tests::start_cluster_with_smolvlm2::start_cluster_with_smolvlm2;
-
-const MAX_TOKENS: NonZeroU32 = NonZeroU32::new(20).unwrap();
-
-fn drive_normal_image_fixture(
-    cluster: &Cluster,
-    fixture_name: &str,
-    mime_type: &str,
-) -> Result<impl Future<Output = Result<()>> + Send + use<>> {
-    let image_data_uri = load_fixture_data_uri(fixture_name, mime_type)?;
-    let fixture_name = fixture_name.to_owned();
-
-    let generation = cluster.continue_from_conversation_history(
-        CancellationToken::new(),
-        &ContinueFromConversationHistoryParams {
-            add_generation_prompt: true,
-            conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                content: ConversationMessageContent::Parts(vec![
-                    ConversationMessageContentPart::ImageUrl {
-                        image_url: ImageUrl {
-                            url: image_data_uri,
-                        },
-                    },
-                    ConversationMessageContentPart::Text {
-                        text: "What do you see in this image?".to_owned(),
-                    },
-                ]),
-                role: "user".to_owned(),
-            }]),
-            enable_thinking: false,
-            grammar: None,
-            max_tokens: MAX_TOKENS,
-            parse_tool_calls: false,
-            tools: vec![],
-        },
-    );
-
-    Ok(async move {
-        let collected = generation.await?;
-
-        let saw_token = collected
-            .token_results
-            .iter()
-            .any(|result| result.token_result.is_token());
-
-        assert!(
-            saw_token,
-            "fixture {fixture_name} should produce at least one content/reasoning/tool-call token with adequate n_batch; got {:?}",
-            collected
-                .token_results
-                .iter()
-                .map(|result| &result.token_result)
-                .collect::<Vec<_>>(),
-        );
-
-        Ok(())
-    })
-}
-
-#[tokio::test(flavor = "multi_thread")]
-async fn agent_completes_generation_with_adequate_n_batch() {
-    let cluster = start_cluster_with_smolvlm2(vec![AgentConfig::single(1)])
-        .await
-        .expect("the cluster must start");
-
-    drive_normal_image_fixture(&cluster, "sarnow.jpeg", "image/jpeg")
-        .expect("the image fixture request must complete")
-        .await
-        .expect("the message must be readable");
-    drive_normal_image_fixture(&cluster, "llamas.webp", "image/webp")
-        .expect("the image fixture request must complete")
-        .await
-        .expect("the message must be readable");
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_tests/tests/agent_conversation_accepts_empty_tools_list.rs b/paddler_tests/tests/agent_conversation_accepts_empty_tools_list.rs
index fee48bc30..20ac539f1 100644
--- a/paddler_tests/tests/agent_conversation_accepts_empty_tools_list.rs
+++ b/paddler_tests/tests/agent_conversation_accepts_empty_tools_list.rs
@@ -22,10 +22,12 @@ async fn agent_conversation_accepts_empty_tools_list() {
             CancellationToken::new(),
             &ContinueFromConversationHistoryParams {
                 add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Text("Say hello".to_owned()),
-                    role: "user".to_owned(),
-                }]),
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Text("Say hello".to_owned()),
+                        role: "user".to_owned(),
+                    },
+                ]),
                 enable_thinking: true,
                 grammar: None,
                 max_tokens: NonZeroU32::new(10).unwrap(),
diff --git a/paddler_tests/tests/agent_conversation_history_respects_max_tokens.rs b/paddler_tests/tests/agent_conversation_history_respects_max_tokens.rs
index b40577244..bd7dc47ef 100644
--- a/paddler_tests/tests/agent_conversation_history_respects_max_tokens.rs
+++ b/paddler_tests/tests/agent_conversation_history_respects_max_tokens.rs
@@ -22,10 +22,12 @@ async fn agent_conversation_history_respects_max_tokens() {
             CancellationToken::new(),
             &ContinueFromConversationHistoryParams {
                 add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Text("Say hello".to_owned()),
-                    role: "user".to_owned(),
-                }]),
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Text("Say hello".to_owned()),
+                        role: "user".to_owned(),
+                    },
+                ]),
                 enable_thinking: false,
                 grammar: None,
                 max_tokens: NonZeroU32::new(20).unwrap(),
diff --git a/paddler_tests/tests/agent_conversation_with_function_tool_succeeds.rs b/paddler_tests/tests/agent_conversation_with_function_tool_succeeds.rs
index ae650eee4..4762ad012 100644
--- a/paddler_tests/tests/agent_conversation_with_function_tool_succeeds.rs
+++ b/paddler_tests/tests/agent_conversation_with_function_tool_succeeds.rs
@@ -1,5 +1,6 @@
 #![cfg(feature = "tests_that_use_llms")]
 
+use std::mem::discriminant;
 use std::num::NonZeroU32;
 
 use tokio_util::sync::CancellationToken;
@@ -9,6 +10,7 @@ use paddler_messaging::conversation_message::ConversationMessage;
 use paddler_messaging::conversation_message_content::ConversationMessageContent;
 use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_tests::generation_done_discriminant::generation_done_discriminant;
 use paddler_tests::get_weather_tool::get_weather_tool;
 use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
 
@@ -23,10 +25,12 @@ async fn agent_conversation_with_function_tool_succeeds() {
             CancellationToken::new(),
             &ContinueFromConversationHistoryParams {
                 add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Text("Say hello".to_owned()),
-                    role: "user".to_owned(),
-                }]),
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Text("Say hello".to_owned()),
+                        role: "user".to_owned(),
+                    },
+                ]),
                 enable_thinking: true,
                 grammar: None,
                 max_tokens: NonZeroU32::new(50).unwrap(),
@@ -37,9 +41,14 @@ async fn agent_conversation_with_function_tool_succeeds() {
         .await
         .expect("the inference request must be accepted");
 
-    assert!(
-        !collected.token_results.is_empty(),
-        "should receive a response when a function tool is provided"
+    assert_eq!(
+        collected
+            .token_results
+            .last()
+            .map(|token_result_with_producer| discriminant(
+                &token_result_with_producer.token_result
+            )),
+        Some(generation_done_discriminant())
     );
 
     cluster
diff --git a/paddler_tests/tests/agent_conversation_with_gbnf_grammar_constrains_output.rs b/paddler_tests/tests/agent_conversation_with_gbnf_grammar_constrains_output.rs
index d3d0f29c2..8458001cf 100644
--- a/paddler_tests/tests/agent_conversation_with_gbnf_grammar_constrains_output.rs
+++ b/paddler_tests/tests/agent_conversation_with_gbnf_grammar_constrains_output.rs
@@ -23,12 +23,14 @@ async fn agent_conversation_with_gbnf_grammar_constrains_output() {
             CancellationToken::new(),
             &ContinueFromConversationHistoryParams {
                 add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Text(
-                        "Is the sky blue? Answer yes or no.".to_owned(),
-                    ),
-                    role: "user".to_owned(),
-                }]),
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Text(
+                            "Is the sky blue? Answer yes or no.".to_owned(),
+                        ),
+                        role: "user".to_owned(),
+                    },
+                ]),
                 enable_thinking: false,
                 grammar: Some(GrammarConstraint::Gbnf {
                     grammar: r"root ::= [Yy][Ee][Ss] | [Nn][Oo]".to_owned(),
@@ -45,9 +47,8 @@ async fn agent_conversation_with_gbnf_grammar_constrains_output() {
     let lower = collected.text.to_lowercase();
 
     assert!(
-        lower == "yes" || lower == "no",
-        "GBNF grammar should constrain output to yes/no; got {:?}",
-        collected.text
+        ["yes", "no"].contains(&lower.as_str()),
+        "GBNF grammar should constrain output to yes/no; got {lower:?}"
     );
 
     cluster
diff --git a/paddler_tests/tests/agent_conversation_with_json_schema_grammar_returns_valid_json.rs b/paddler_tests/tests/agent_conversation_with_json_schema_grammar_returns_valid_json.rs
index 2800c1f43..1156afca3 100644
--- a/paddler_tests/tests/agent_conversation_with_json_schema_grammar_returns_valid_json.rs
+++ b/paddler_tests/tests/agent_conversation_with_json_schema_grammar_returns_valid_json.rs
@@ -23,7 +23,7 @@ async fn agent_conversation_with_json_schema_grammar_returns_valid_json() {
     let collected = cluster
         .continue_from_conversation_history(CancellationToken::new(), &ContinueFromConversationHistoryParams {
             add_generation_prompt: true,
-            conversation_history: ConversationHistory::new(vec![ConversationMessage {
+            conversation_history: ConversationHistory::new(vec![ConversationMessage::Message {
                 content: ConversationMessageContent::Text("What is 2+2?".to_owned()),
                 role: "user".to_owned(),
             }]),
diff --git a/paddler_tests/tests/agent_drops_the_issues_of_a_replaced_desired_state.rs b/paddler_tests/tests/agent_drops_the_issues_of_a_replaced_desired_state.rs
new file mode 100644
index 000000000..6b4a9f2f0
--- /dev/null
+++ b/paddler_tests/tests/agent_drops_the_issues_of_a_replaced_desired_state.rs
@@ -0,0 +1,59 @@
+use std::collections::BTreeSet;
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::agent_desired_model::AgentDesiredModel;
+use paddler_messaging::agent_issue::AgentIssue;
+use paddler_messaging::agent_issue_params::model_path::ModelPath;
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_tests::start_single_agent_cluster_with_desired_state::start_single_agent_cluster_with_desired_state;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agent_drops_the_issues_of_a_replaced_desired_state() {
+    let replaced_model_uri = "not a valid uri".to_owned();
+    let replacement_model_uri = "https://huggingface.co/owner/repo".to_owned();
+    let mut cluster = start_single_agent_cluster_with_desired_state(BalancerDesiredState {
+        model: AgentDesiredModel::Uri(replaced_model_uri.clone()),
+        ..BalancerDesiredState::default()
+    })
+    .await
+    .expect("a single-agent cluster must start");
+
+    cluster
+        .wait_for_first_agent_issue(|issue| {
+            matches!(issue, AgentIssue::ModelUriIsUnparseable(model_path) if model_path.model_path == replaced_model_uri)
+        })
+        .await
+        .expect("the agent must report the issue of the replaced desired state");
+
+    cluster
+        .client_management
+        .put_balancer_desired_state(
+            CancellationToken::new(),
+            &BalancerDesiredState {
+                model: AgentDesiredModel::Uri(replacement_model_uri.clone()),
+                ..BalancerDesiredState::default()
+            },
+        )
+        .await
+        .expect("the balancer must apply the replacement desired state");
+
+    let snapshot = cluster
+        .wait_for_first_agent_issue(|issue| {
+            matches!(issue, AgentIssue::HuggingFaceModelUriIsMalformed(model_path) if model_path.model_path == replacement_model_uri)
+        })
+        .await
+        .expect("the agent must report the issue of the replacement desired state");
+
+    assert_eq!(
+        snapshot.agents[0].status.issues,
+        BTreeSet::from([AgentIssue::HuggingFaceModelUriIsMalformed(ModelPath {
+            model_path: replacement_model_uri,
+        })])
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/agent_embedding_batch_distribution_independent_of_context_size.rs b/paddler_tests/tests/agent_embedding_batch_distribution_independent_of_context_size.rs
deleted file mode 100644
index 55529e16f..000000000
--- a/paddler_tests/tests/agent_embedding_batch_distribution_independent_of_context_size.rs
+++ /dev/null
@@ -1,79 +0,0 @@
-#![cfg(feature = "tests_that_use_llms")]
-
-use std::collections::BTreeSet;
-use std::num::NonZeroU32;
-
-use tokio_util::sync::CancellationToken;
-
-use paddler_inference_parameters::batch_size::BatchSize;
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
-use paddler_messaging::embedding_input_document::EmbeddingInputDocument;
-use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
-use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
-use paddler_test_cluster_harness::agent_config::AgentConfig;
-use paddler_test_cluster_harness::embedding_cluster_params::EmbeddingClusterParams;
-use paddler_tests::start_embedding_cluster::start_embedding_cluster;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn agent_embedding_batch_distribution_independent_of_context_size() {
-    let cluster = start_embedding_cluster(EmbeddingClusterParams {
-        agents: vec![AgentConfig::single(4)],
-        inference_parameters: InferenceParameters {
-            n_batch: BatchSize::try_from(64).expect("the value must fit its target type"),
-            context_size: NonZeroU32::try_from(512).expect("the value must fit its target type"),
-            enable_embeddings: true,
-            ..InferenceParameters::deterministic()
-        },
-        ..EmbeddingClusterParams::default()
-    })
-    .await
-    .expect("the cluster must start");
-
-    let collected = cluster
-        .generate_embedding_batch(CancellationToken::new(), &GenerateEmbeddingBatchParams {
-            input_batch: vec![
-                EmbeddingInputDocument {
-                    content: "This is the first document with enough content to contribute meaningfully to the batch size calculation".to_owned(),
-                    id: "doc-chunk-1".to_owned(),
-                },
-                EmbeddingInputDocument {
-                    content: "This is the second document that should be processed in a potentially different batch from the first".to_owned(),
-                    id: "doc-chunk-2".to_owned(),
-                },
-                EmbeddingInputDocument {
-                    content: "This is the third document adding more content to ensure the total exceeds the configured chunk limit".to_owned(),
-                    id: "doc-chunk-3".to_owned(),
-                },
-                EmbeddingInputDocument {
-                    content: "This is the fourth document which should demonstrate that batching distributes across agent requests".to_owned(),
-                    id: "doc-chunk-4".to_owned(),
-                },
-            ],
-            normalization_method: EmbeddingNormalizationMethod::None,
-        })
-        .await.expect("the embedding batch must be accepted");
-
-    assert_eq!(collected.embeddings.len(), 4);
-    assert!(collected.saw_done);
-    assert!(collected.errors.is_empty());
-
-    let returned_ids: BTreeSet<String> = collected
-        .embeddings
-        .iter()
-        .map(|produced| produced.embedding.source_document_id.clone())
-        .collect();
-
-    let expected_ids: BTreeSet<String> = BTreeSet::from([
-        "doc-chunk-1".to_owned(),
-        "doc-chunk-2".to_owned(),
-        "doc-chunk-3".to_owned(),
-        "doc-chunk-4".to_owned(),
-    ]);
-
-    assert_eq!(returned_ids, expected_ids);
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_tests/tests/agent_embedding_batch_returns_one_embedding_per_input_document.rs b/paddler_tests/tests/agent_embedding_batch_returns_one_embedding_per_input_document.rs
index 29d9322ab..8a1751afa 100644
--- a/paddler_tests/tests/agent_embedding_batch_returns_one_embedding_per_input_document.rs
+++ b/paddler_tests/tests/agent_embedding_batch_returns_one_embedding_per_input_document.rs
@@ -4,9 +4,9 @@ use std::collections::BTreeSet;
 
 use tokio_util::sync::CancellationToken;
 
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
 use paddler_messaging::embedding_input_document::EmbeddingInputDocument;
 use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
+use paddler_messaging::embedding_result::EmbeddingResult;
 use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
 use paddler_test_cluster_harness::embedding_cluster_params::EmbeddingClusterParams;
@@ -16,10 +16,7 @@ use paddler_tests::start_embedding_cluster::start_embedding_cluster;
 async fn agent_embedding_batch_returns_one_embedding_per_input_document() {
     let cluster = start_embedding_cluster(EmbeddingClusterParams {
         agents: vec![AgentConfig::single(1)],
-        inference_parameters: InferenceParameters {
-            enable_embeddings: true,
-            ..InferenceParameters::deterministic()
-        },
+
         ..EmbeddingClusterParams::default()
     })
     .await
@@ -48,7 +45,7 @@ async fn agent_embedding_batch_returns_one_embedding_per_input_document() {
 
     assert_eq!(collected.embeddings.len(), 2);
     assert!(collected.saw_done);
-    assert!(collected.errors.is_empty());
+    assert_eq!(collected.failures, Vec::<EmbeddingResult>::new());
 
     let returned_ids: BTreeSet<String> = collected
         .embeddings
diff --git a/paddler_tests/tests/agent_embedding_batch_with_all_oversized_documents_reports_each_and_embeds_nothing.rs b/paddler_tests/tests/agent_embedding_batch_with_all_oversized_documents_reports_each_and_embeds_nothing.rs
new file mode 100644
index 000000000..26013d0a3
--- /dev/null
+++ b/paddler_tests/tests/agent_embedding_batch_with_all_oversized_documents_reports_each_and_embeds_nothing.rs
@@ -0,0 +1,83 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::num::NonZeroU32;
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_inference_parameters::batch_size::BatchSize;
+use paddler_inference_parameters::model_runtime_parameters::ModelRuntimeParameters;
+use paddler_messaging::embedding_input_document::EmbeddingInputDocument;
+use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
+use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::embedding_cluster_params::EmbeddingClusterParams;
+use paddler_tests::oversized_text::oversized_text;
+use paddler_tests::start_embedding_cluster::start_embedding_cluster;
+
+const N_BATCH: u32 = 64;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agent_embedding_batch_with_all_oversized_documents_reports_each_and_embeds_nothing() {
+    let cluster = start_embedding_cluster(EmbeddingClusterParams {
+        agents: vec![AgentConfig::single(1)],
+        model_runtime_parameters: ModelRuntimeParameters {
+            n_batch: BatchSize::try_from(N_BATCH).expect("the value must fit its target type"),
+            context_size: NonZeroU32::try_from(2048).expect("the value must fit its target type"),
+            ..ModelRuntimeParameters::default()
+        },
+        ..EmbeddingClusterParams::default()
+    })
+    .await
+    .expect("the cluster must start");
+
+    let huge_content = oversized_text();
+
+    let collected = cluster
+        .generate_embedding_batch(
+            CancellationToken::new(),
+            &GenerateEmbeddingBatchParams {
+                input_batch: vec![
+                    EmbeddingInputDocument {
+                        content: huge_content.clone(),
+                        id: "huge-1".to_owned(),
+                    },
+                    EmbeddingInputDocument {
+                        content: huge_content,
+                        id: "huge-2".to_owned(),
+                    },
+                ],
+                normalization_method: EmbeddingNormalizationMethod::None,
+            },
+        )
+        .await
+        .expect("the embedding batch must be accepted");
+
+    assert_eq!(
+        collected.embeddings.len(),
+        0,
+        "no embeddings should be produced when all documents are oversized",
+    );
+    assert_eq!(
+        collected
+            .oversized_documents
+            .iter()
+            .map(|details| details.source_document_id.as_str())
+            .collect::<Vec<_>>(),
+        vec!["huge-1", "huge-2"],
+        "both oversized documents should be reported",
+    );
+    assert!(
+        collected.saw_done,
+        "stream must terminate with the balancer's final Done so the client unblocks",
+    );
+    assert_eq!(
+        collected.no_embeddings_produced_count, 1,
+        "the agent must terminate its sub-stream with a single NoEmbeddingsProduced variant when zero embeddings are produced; got errors: {:?}",
+        collected.failures,
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/agent_embedding_batch_with_all_oversized_documents_reports_error.rs b/paddler_tests/tests/agent_embedding_batch_with_all_oversized_documents_reports_error.rs
deleted file mode 100644
index 269937c46..000000000
--- a/paddler_tests/tests/agent_embedding_batch_with_all_oversized_documents_reports_error.rs
+++ /dev/null
@@ -1,85 +0,0 @@
-#![cfg(feature = "tests_that_use_llms")]
-
-use std::num::NonZeroU32;
-
-use tokio_util::sync::CancellationToken;
-
-use paddler_inference_parameters::batch_size::BatchSize;
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
-use paddler_messaging::embedding_input_document::EmbeddingInputDocument;
-use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
-use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
-use paddler_test_cluster_harness::agent_config::AgentConfig;
-use paddler_test_cluster_harness::embedding_cluster_params::EmbeddingClusterParams;
-use paddler_tests::start_embedding_cluster::start_embedding_cluster;
-
-const N_BATCH: u32 = 64;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn agent_embedding_batch_with_all_oversized_documents_reports_error() {
-    let cluster = start_embedding_cluster(EmbeddingClusterParams {
-        agents: vec![AgentConfig::single(1)],
-        inference_parameters: InferenceParameters {
-            n_batch: BatchSize::try_from(N_BATCH).expect("the value must fit its target type"),
-            context_size: NonZeroU32::try_from(2048).expect("the value must fit its target type"),
-            enable_embeddings: true,
-            ..InferenceParameters::deterministic()
-        },
-        ..EmbeddingClusterParams::default()
-    })
-    .await
-    .expect("the cluster must start");
-
-    let huge_content = "The quick brown fox jumps over the lazy dog. ".repeat(40);
-
-    let collected = cluster
-        .generate_embedding_batch(
-            CancellationToken::new(),
-            &GenerateEmbeddingBatchParams {
-                input_batch: vec![
-                    EmbeddingInputDocument {
-                        content: huge_content.clone(),
-                        id: "huge-1".to_owned(),
-                    },
-                    EmbeddingInputDocument {
-                        content: huge_content,
-                        id: "huge-2".to_owned(),
-                    },
-                ],
-                normalization_method: EmbeddingNormalizationMethod::None,
-            },
-        )
-        .await
-        .expect("the embedding batch must be accepted");
-
-    assert_eq!(
-        collected.embeddings.len(),
-        0,
-        "no embeddings should be produced when all documents are oversized",
-    );
-    assert_eq!(
-        collected.oversized_documents.len(),
-        2,
-        "both oversized documents should be reported",
-    );
-    assert!(
-        collected.saw_done,
-        "stream must terminate with the balancer's final Done so the client unblocks",
-    );
-    assert_eq!(
-        collected.no_embeddings_produced_count,
-        1,
-        "the agent must terminate its sub-stream with a single NoEmbeddingsProduced variant when zero embeddings are produced; got oversized_documents: {:?}, errors: {:?}",
-        collected
-            .oversized_documents
-            .iter()
-            .map(|details| &details.source_document_id)
-            .collect::<Vec<_>>(),
-        collected.errors,
-    );
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_tests/tests/agent_embeddings_share_dimension_across_inputs_of_varying_length.rs b/paddler_tests/tests/agent_embeddings_share_dimension_across_inputs_of_varying_length.rs
index 99bb7965a..d893c02cb 100644
--- a/paddler_tests/tests/agent_embeddings_share_dimension_across_inputs_of_varying_length.rs
+++ b/paddler_tests/tests/agent_embeddings_share_dimension_across_inputs_of_varying_length.rs
@@ -2,7 +2,6 @@
 
 use tokio_util::sync::CancellationToken;
 
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
 use paddler_messaging::embedding_input_document::EmbeddingInputDocument;
 use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
 use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
@@ -14,10 +13,7 @@ use paddler_tests::start_embedding_cluster::start_embedding_cluster;
 async fn agent_embeddings_share_dimension_across_inputs_of_varying_length() {
     let cluster = start_embedding_cluster(EmbeddingClusterParams {
         agents: vec![AgentConfig::single(1)],
-        inference_parameters: InferenceParameters {
-            enable_embeddings: true,
-            ..InferenceParameters::deterministic()
-        },
+
         ..EmbeddingClusterParams::default()
     })
     .await
@@ -50,16 +46,15 @@ async fn agent_embeddings_share_dimension_across_inputs_of_varying_length() {
 
     assert!(first_dimension > 0, "embedding dimension must be positive");
 
-    for produced in &collected.embeddings {
-        assert_eq!(
-            produced.embedding.embedding.len(),
-            first_dimension,
-            "all embeddings must share dimension; {} has {} instead of {}",
-            produced.embedding.source_document_id,
-            produced.embedding.embedding.len(),
-            first_dimension
-        );
-    }
+    assert_eq!(
+        collected
+            .embeddings
+            .iter()
+            .map(|produced| produced.embedding.embedding.len())
+            .collect::<Vec<_>>(),
+        vec![first_dimension; collected.embeddings.len()],
+        "all embeddings must share dimension"
+    );
 
     cluster
         .shutdown()
diff --git a/paddler_tests/tests/agent_embeds_a_document_longer_than_a_sequence_context_with_an_encoder.rs b/paddler_tests/tests/agent_embeds_a_document_longer_than_a_sequence_context_with_an_encoder.rs
new file mode 100644
index 000000000..9955a4380
--- /dev/null
+++ b/paddler_tests/tests/agent_embeds_a_document_longer_than_a_sequence_context_with_an_encoder.rs
@@ -0,0 +1,63 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::num::NonZeroU32;
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_inference_parameters::batch_size::BatchSize;
+use paddler_inference_parameters::model_runtime_parameters::ModelRuntimeParameters;
+use paddler_messaging::embedding_input_document::EmbeddingInputDocument;
+use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
+use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::embedding_cluster_params::EmbeddingClusterParams;
+use paddler_tests::start_embedding_cluster::start_embedding_cluster;
+
+const CONTEXT_SIZE: u32 = 512;
+const SLOTS: u16 = 4;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agent_embeds_a_document_longer_than_a_sequence_context_with_an_encoder() {
+    let context_size =
+        NonZeroU32::new(CONTEXT_SIZE).expect("the context size must be greater than zero");
+    let cluster = start_embedding_cluster(EmbeddingClusterParams {
+        agents: AgentConfig::uniform(1, SLOTS),
+        model_runtime_parameters: ModelRuntimeParameters {
+            context_size,
+            n_batch: BatchSize::default().at_most(context_size),
+            ..ModelRuntimeParameters::default()
+        },
+        ..EmbeddingClusterParams::default()
+    })
+    .await
+    .expect("the cluster must start");
+
+    let collected = cluster
+        .generate_embedding_batch(
+            CancellationToken::new(),
+            &GenerateEmbeddingBatchParams {
+                input_batch: vec![EmbeddingInputDocument {
+                    content: "The quick brown fox jumps over the lazy dog. ".repeat(30),
+                    id: "longer-than-a-sequence".to_owned(),
+                }],
+                normalization_method: EmbeddingNormalizationMethod::None,
+            },
+        )
+        .await
+        .expect("the embedding batch must be accepted");
+
+    assert_eq!(collected.failures, vec![]);
+    assert_eq!(
+        collected
+            .embeddings
+            .iter()
+            .map(|produced| produced.embedding.source_document_id.as_str())
+            .collect::<Vec<_>>(),
+        vec!["longer-than-a-sequence"]
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/agent_embeds_a_document_longer_than_the_default_micro_batch.rs b/paddler_tests/tests/agent_embeds_a_document_longer_than_the_default_micro_batch.rs
index 397a0757a..1b99ae074 100644
--- a/paddler_tests/tests/agent_embeds_a_document_longer_than_the_default_micro_batch.rs
+++ b/paddler_tests/tests/agent_embeds_a_document_longer_than_the_default_micro_batch.rs
@@ -4,7 +4,7 @@ use std::num::NonZeroU32;
 
 use tokio_util::sync::CancellationToken;
 
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
+use paddler_inference_parameters::model_runtime_parameters::ModelRuntimeParameters;
 use paddler_messaging::embedding_input_document::EmbeddingInputDocument;
 use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
 use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
@@ -19,11 +19,10 @@ const SENTENCES_EXCEEDING_THE_DEFAULT_MICRO_BATCH: usize = 100;
 async fn agent_embeds_a_document_longer_than_the_default_micro_batch() {
     let cluster = start_embedding_cluster(EmbeddingClusterParams {
         agents: vec![AgentConfig::single(1)],
-        inference_parameters: InferenceParameters {
+        model_runtime_parameters: ModelRuntimeParameters {
             context_size: NonZeroU32::try_from(NOMIC_TRAINING_CONTEXT_SIZE)
                 .expect("the value must fit its target type"),
-            enable_embeddings: true,
-            ..InferenceParameters::deterministic()
+            ..ModelRuntimeParameters::default()
         },
         ..EmbeddingClusterParams::default()
     })
diff --git a/paddler_tests/tests/agent_embeds_every_document_of_a_batch_larger_than_its_slot_count.rs b/paddler_tests/tests/agent_embeds_every_document_of_a_batch_larger_than_its_slot_count.rs
new file mode 100644
index 000000000..55492d737
--- /dev/null
+++ b/paddler_tests/tests/agent_embeds_every_document_of_a_batch_larger_than_its_slot_count.rs
@@ -0,0 +1,63 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::collections::BTreeSet;
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
+use paddler_messaging::embedding_result::EmbeddingResult;
+use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::embedding_cluster_params::EmbeddingClusterParams;
+use paddler_tests::numbered_embedding_documents::numbered_embedding_documents;
+use paddler_tests::start_embedding_cluster::start_embedding_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agent_embeds_every_document_of_a_batch_larger_than_its_slot_count() {
+    let cluster = start_embedding_cluster(EmbeddingClusterParams {
+        agents: vec![AgentConfig::single(4)],
+
+        ..EmbeddingClusterParams::default()
+    })
+    .await
+    .expect("the cluster must start");
+
+    let input_batch = numbered_embedding_documents(12);
+    let expected_ids: BTreeSet<String> = input_batch
+        .iter()
+        .map(|document| document.id.clone())
+        .collect();
+
+    let collected = cluster
+        .generate_embedding_batch(
+            CancellationToken::new(),
+            &GenerateEmbeddingBatchParams {
+                input_batch,
+                normalization_method: EmbeddingNormalizationMethod::None,
+            },
+        )
+        .await
+        .expect("the embedding batch must be accepted");
+
+    assert_eq!(collected.embeddings.len(), 12);
+    assert!(collected.saw_done);
+    assert_eq!(collected.failures, Vec::<EmbeddingResult>::new());
+
+    let returned_ids: BTreeSet<String> = collected
+        .embeddings
+        .iter()
+        .map(|produced| produced.embedding.source_document_id.clone())
+        .collect();
+    assert_eq!(returned_ids, expected_ids);
+
+    let first_dimension = collected.embeddings[0].embedding.embedding.len();
+
+    for produced in &collected.embeddings {
+        assert_eq!(produced.embedding.embedding.len(), first_dimension);
+    }
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/agent_embeds_every_document_with_a_small_batch_and_context.rs b/paddler_tests/tests/agent_embeds_every_document_with_a_small_batch_and_context.rs
new file mode 100644
index 000000000..5149add19
--- /dev/null
+++ b/paddler_tests/tests/agent_embeds_every_document_with_a_small_batch_and_context.rs
@@ -0,0 +1,79 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::collections::BTreeSet;
+use std::num::NonZeroU32;
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_inference_parameters::batch_size::BatchSize;
+use paddler_inference_parameters::model_runtime_parameters::ModelRuntimeParameters;
+use paddler_messaging::embedding_input_document::EmbeddingInputDocument;
+use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
+use paddler_messaging::embedding_result::EmbeddingResult;
+use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::embedding_cluster_params::EmbeddingClusterParams;
+use paddler_tests::start_embedding_cluster::start_embedding_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agent_embeds_every_document_with_a_small_batch_and_context() {
+    let cluster = start_embedding_cluster(EmbeddingClusterParams {
+        agents: vec![AgentConfig::single(4)],
+        model_runtime_parameters: ModelRuntimeParameters {
+            n_batch: BatchSize::try_from(64).expect("the value must fit its target type"),
+            context_size: NonZeroU32::try_from(512).expect("the value must fit its target type"),
+            ..ModelRuntimeParameters::default()
+        },
+        ..EmbeddingClusterParams::default()
+    })
+    .await
+    .expect("the cluster must start");
+
+    let collected = cluster
+        .generate_embedding_batch(CancellationToken::new(), &GenerateEmbeddingBatchParams {
+            input_batch: vec![
+                EmbeddingInputDocument {
+                    content: "This is the first document with enough content to contribute meaningfully to the batch size calculation".to_owned(),
+                    id: "doc-chunk-1".to_owned(),
+                },
+                EmbeddingInputDocument {
+                    content: "This is the second document that should be processed in a potentially different batch from the first".to_owned(),
+                    id: "doc-chunk-2".to_owned(),
+                },
+                EmbeddingInputDocument {
+                    content: "This is the third document adding more content to ensure the total exceeds the configured chunk limit".to_owned(),
+                    id: "doc-chunk-3".to_owned(),
+                },
+                EmbeddingInputDocument {
+                    content: "This is the fourth document which should demonstrate that batching distributes across agent requests".to_owned(),
+                    id: "doc-chunk-4".to_owned(),
+                },
+            ],
+            normalization_method: EmbeddingNormalizationMethod::None,
+        })
+        .await.expect("the embedding batch must be accepted");
+
+    assert_eq!(collected.embeddings.len(), 4);
+    assert!(collected.saw_done);
+    assert_eq!(collected.failures, Vec::<EmbeddingResult>::new());
+
+    let returned_ids: BTreeSet<String> = collected
+        .embeddings
+        .iter()
+        .map(|produced| produced.embedding.source_document_id.clone())
+        .collect();
+
+    let expected_ids: BTreeSet<String> = BTreeSet::from([
+        "doc-chunk-1".to_owned(),
+        "doc-chunk-2".to_owned(),
+        "doc-chunk-3".to_owned(),
+        "doc-chunk-4".to_owned(),
+    ]);
+
+    assert_eq!(returned_ids, expected_ids);
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/agent_emits_tool_call_parsed_event_before_trailing_content.rs b/paddler_tests/tests/agent_emits_tool_call_parsed_event_before_trailing_content.rs
index c809ddc81..6646bf47c 100644
--- a/paddler_tests/tests/agent_emits_tool_call_parsed_event_before_trailing_content.rs
+++ b/paddler_tests/tests/agent_emits_tool_call_parsed_event_before_trailing_content.rs
@@ -1,21 +1,17 @@
 #![cfg(feature = "tests_that_use_llms")]
 
-use std::num::NonZeroU32;
-
 use llama_cpp_bindings_types::ToolCallArguments;
 use serde_json::json;
-use serde_json::to_string;
 use tokio_util::sync::CancellationToken;
 
-use paddler_messaging::conversation_history::ConversationHistory;
-use paddler_messaging::conversation_message::ConversationMessage;
-use paddler_messaging::conversation_message_content::ConversationMessageContent;
 use paddler_messaging::generated_token_result::GeneratedTokenResult;
-use paddler_messaging::grammar_constraint::GrammarConstraint;
-use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
+use paddler_messaging::generated_tool_call::GeneratedToolCall;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_tests::gbnf_literal::gbnf_literal;
 use paddler_tests::get_weather_tool::get_weather_tool;
+use paddler_tests::max_tokens_for_forced_text::max_tokens_for_forced_text;
 use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
+use paddler_tests::weather_question_conversation::weather_question_conversation;
 
 const TOOL_CALL_FOLLOWED_BY_CONTENT: &str = "<tool_call>\n{\"name\": \"get_weather\", \"arguments\": {\"location\": \"Paris\"}}\n</tool_call>\nI asked for the weather in Paris.";
 
@@ -28,30 +24,11 @@ async fn agent_emits_tool_call_parsed_event_before_trailing_content() {
     let collected = cluster
         .continue_from_conversation_history(
             CancellationToken::new(),
-            &ContinueFromConversationHistoryParams {
-                add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Text(
-                        "What is the weather in Paris?".to_owned(),
-                    ),
-                    role: "user".to_owned(),
-                }]),
-                enable_thinking: false,
-                grammar: Some(GrammarConstraint::Gbnf {
-                    grammar: format!(
-                        "root ::= {}",
-                        to_string(TOOL_CALL_FOLLOWED_BY_CONTENT).expect("the value must serialize")
-                    ),
-                    root: "root".to_owned(),
-                }),
-                max_tokens: NonZeroU32::try_from(
-                    u32::try_from(TOOL_CALL_FOLLOWED_BY_CONTENT.len())
-                        .expect("the value must fit its target type"),
-                )
-                .expect("the value must fit its target type"),
-                parse_tool_calls: true,
-                tools: vec![get_weather_tool()],
-            },
+            &weather_question_conversation(
+                gbnf_literal(TOOL_CALL_FOLLOWED_BY_CONTENT),
+                max_tokens_for_forced_text(TOOL_CALL_FOLLOWED_BY_CONTENT),
+                vec![get_weather_tool()],
+            ),
         )
         .await
         .expect("the inference request must be accepted");
@@ -70,11 +47,14 @@ async fn agent_emits_tool_call_parsed_event_before_trailing_content() {
         "expected content after the parsed tool call, got {token_results:?}"
     );
 
-    let Some(GeneratedTokenResult::ToolCallParsed(parsed_tool_calls)) =
-        tool_call_position.map(|position| &token_results[position])
-    else {
-        panic!("expected a parsed tool call, got {token_results:?}");
-    };
+    let parsed_tool_calls: Vec<&GeneratedToolCall> = token_results
+        .iter()
+        .filter_map(|token_result| match token_result {
+            GeneratedTokenResult::ToolCallParsed(parsed_tool_calls) => Some(parsed_tool_calls),
+            _ => None,
+        })
+        .flatten()
+        .collect();
 
     assert_eq!(parsed_tool_calls.len(), 1);
     assert_eq!(parsed_tool_calls[0].name, "get_weather");
diff --git a/paddler_tests/tests/agent_ends_a_management_connection_on_a_binary_frame.rs b/paddler_tests/tests/agent_ends_a_management_connection_on_a_binary_frame.rs
new file mode 100644
index 000000000..6e8196267
--- /dev/null
+++ b/paddler_tests/tests/agent_ends_a_management_connection_on_a_binary_frame.rs
@@ -0,0 +1,8 @@
+use tokio_tungstenite::tungstenite::Message;
+
+use crate::agent_hang_up_on_a_frame::agent_hang_up_on_a_frame;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agent_ends_a_management_connection_on_a_binary_frame() {
+    agent_hang_up_on_a_frame(Message::binary(vec![0_u8, 1])).await;
+}
diff --git a/paddler_tests/tests/agent_ends_a_management_connection_on_text_that_is_not_json.rs b/paddler_tests/tests/agent_ends_a_management_connection_on_text_that_is_not_json.rs
new file mode 100644
index 000000000..011edc7e8
--- /dev/null
+++ b/paddler_tests/tests/agent_ends_a_management_connection_on_text_that_is_not_json.rs
@@ -0,0 +1,8 @@
+use tokio_tungstenite::tungstenite::Message;
+
+use crate::agent_hang_up_on_a_frame::agent_hang_up_on_a_frame;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agent_ends_a_management_connection_on_text_that_is_not_json() {
+    agent_hang_up_on_a_frame(Message::text("not json")).await;
+}
diff --git a/paddler_tests/tests/agent_ends_a_management_connection_whose_balancer_reuses_an_in_flight_request_id.rs b/paddler_tests/tests/agent_ends_a_management_connection_whose_balancer_reuses_an_in_flight_request_id.rs
new file mode 100644
index 000000000..c1503cd3c
--- /dev/null
+++ b/paddler_tests/tests/agent_ends_a_management_connection_whose_balancer_reuses_an_in_flight_request_id.rs
@@ -0,0 +1,75 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::num::NonZeroU16;
+
+use serde_json::to_string;
+use tokio_tungstenite::tungstenite::Message;
+
+use paddler_messaging::agent_desired_state::AgentDesiredState;
+use paddler_messaging::jsonrpc::request_envelope::RequestEnvelope;
+use paddler_messaging::management_socket::agent::message::Message as AgentJsonRpcMessage;
+use paddler_messaging::management_socket::agent::notification::Notification as AgentJsonRpcNotification;
+use paddler_messaging::management_socket::agent::notification_params::set_state_params::SetStateParams;
+use paddler_messaging::management_socket::agent::request::Request as AgentJsonRpcRequest;
+use paddler_messaging::reserved_request::ReservedRequest;
+use paddler_test_cluster_harness::unending_generation::unending_generation;
+use paddler_test_cluster_harness::untrusted_balancer_socket_server::UntrustedBalancerSocketServer;
+use paddler_tests::qwen3_desired_state::qwen3_desired_state;
+
+use crate::agent_runner_connected_to::agent_runner_connected_to;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agent_ends_a_management_connection_whose_balancer_reuses_an_in_flight_request_id() {
+    let untrusted_balancer = UntrustedBalancerSocketServer::bind()
+        .await
+        .expect("the untrusted balancer must bind");
+    let _agent_runner = agent_runner_connected_to(untrusted_balancer.management_addr, 1);
+    let mut first_connection = untrusted_balancer
+        .accept_agent()
+        .await
+        .expect("the agent must connect");
+    let unending_generation_request = Message::text(
+        to_string(&AgentJsonRpcMessage::Request(RequestEnvelope {
+            id: "in-flight".to_owned(),
+            request: AgentJsonRpcRequest::ContinueFromRawPrompt(ReservedRequest {
+                params: unending_generation(),
+                reserved_slots: NonZeroU16::MIN,
+            }),
+        }))
+        .expect("the request must serialize"),
+    );
+
+    first_connection
+        .send(Message::text(
+            to_string(&AgentJsonRpcMessage::Notification(
+                AgentJsonRpcNotification::SetState(Box::new(SetStateParams {
+                    desired_state: AgentDesiredState::from(&qwen3_desired_state()),
+                })),
+            ))
+            .expect("the desired state must serialize"),
+        ))
+        .await
+        .expect("the desired state must be sent");
+    first_connection
+        .wait_for_the_agent_to_serve()
+        .await
+        .expect("the agent must serve its model");
+    first_connection
+        .send(unending_generation_request.clone())
+        .await
+        .expect("the request must be sent");
+    first_connection
+        .next_agent_response()
+        .await
+        .expect("the agent must start generating");
+    first_connection
+        .send(unending_generation_request)
+        .await
+        .expect("the reused request must be sent");
+    first_connection.wait_for_the_agent_to_hang_up().await;
+
+    untrusted_balancer
+        .accept_agent()
+        .await
+        .expect("the agent must reconnect after ending the connection");
+}
diff --git a/paddler_tests/tests/agent_finishes_a_generation_that_fills_its_sequence_context.rs b/paddler_tests/tests/agent_finishes_a_generation_that_fills_its_sequence_context.rs
index 8635c2264..eff6e0b23 100644
--- a/paddler_tests/tests/agent_finishes_a_generation_that_fills_its_sequence_context.rs
+++ b/paddler_tests/tests/agent_finishes_a_generation_that_fills_its_sequence_context.rs
@@ -5,10 +5,11 @@ use std::num::NonZeroU32;
 use tokio_util::sync::CancellationToken;
 
 use paddler_inference_parameters::batch_size::BatchSize;
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
+use paddler_inference_parameters::model_runtime_parameters::ModelRuntimeParameters;
 use paddler_messaging::balancer_desired_state::BalancerDesiredState;
 use paddler_messaging::generation_finish::GenerationFinish;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
 use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
 use paddler_test_cluster_harness::unending_generation::unending_generation;
@@ -20,15 +21,15 @@ async fn agent_finishes_a_generation_that_fills_its_sequence_context() {
 
     let cluster = start_cluster(ClusterParams {
         agents: vec![AgentConfig::single(1)],
-        desired_state: Some(BalancerDesiredState {
-            inference_parameters: InferenceParameters {
+        desired_state: ClusterDesiredState::Apply(Box::new(BalancerDesiredState {
+            model_runtime_parameters: ModelRuntimeParameters {
                 n_batch: BatchSize::try_from(256).expect("the value must fit its target type"),
                 context_size: NonZeroU32::try_from(256)
                     .expect("the value must fit its target type"),
-                ..base_desired_state.inference_parameters
+                ..base_desired_state.model_runtime_parameters
             },
             ..base_desired_state
-        }),
+        })),
         wait_for_slots_ready: true,
         ..ClusterParams::default()
     })
diff --git a/paddler_tests/tests/agent_generates_tokens_from_raw_prompt.rs b/paddler_tests/tests/agent_generates_tokens_from_raw_prompt.rs
index d17d7fc8c..698b02622 100644
--- a/paddler_tests/tests/agent_generates_tokens_from_raw_prompt.rs
+++ b/paddler_tests/tests/agent_generates_tokens_from_raw_prompt.rs
@@ -4,10 +4,8 @@ use std::num::NonZeroU32;
 
 use tokio_util::sync::CancellationToken;
 
-use paddler_messaging::generated_token_result::GeneratedTokenResult;
 use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
-use paddler_test_cluster_harness::token_result_with_producer::TokenResultWithProducer;
 use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
 
 #[tokio::test(flavor = "multi_thread")]
@@ -33,13 +31,9 @@ async fn agent_generates_tokens_from_raw_prompt() {
         .count();
 
     assert!(token_count > 0);
-    assert!(matches!(
-        collected.token_results.last(),
-        Some(TokenResultWithProducer {
-            token_result: GeneratedTokenResult::Done(_),
-            ..
-        })
-    ));
+    collected
+        .summary()
+        .expect("the generation must finish with a summary");
 
     cluster
         .shutdown()
diff --git a/paddler_tests/tests/agent_generates_with_a_model_that_declares_no_bos_token.rs b/paddler_tests/tests/agent_generates_with_a_model_that_declares_no_bos_token.rs
new file mode 100644
index 000000000..9f34c2491
--- /dev/null
+++ b/paddler_tests/tests/agent_generates_with_a_model_that_declares_no_bos_token.rs
@@ -0,0 +1,82 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::num::NonZeroU32;
+
+use tempfile::TempDir;
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_messaging::conversation_history::ConversationHistory;
+use paddler_messaging::conversation_message::ConversationMessage;
+use paddler_messaging::conversation_message_content::ConversationMessageContent;
+use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
+use paddler_test_cluster_harness::agent_desired_model_at::agent_desired_model_at;
+use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
+use paddler_tests::qwen3_desired_state::qwen3_desired_state;
+use paddler_tests::start_single_agent_cluster_with_desired_state::start_single_agent_cluster_with_desired_state;
+use paddler_tests::write_patched_gguf::write_patched_gguf;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agent_generates_with_a_model_that_declares_no_bos_token() {
+    let model_directory = TempDir::new().expect("a temporary model directory must be creatable");
+    let model_without_bos_path = model_directory.path().join("qwen3-without-bos.gguf");
+
+    write_patched_gguf(
+        &qwen3_0_6b()
+            .cached_path()
+            .expect("Qwen3 must be in the Hugging Face cache"),
+        &model_without_bos_path,
+        |gguf_context| {
+            gguf_context.set_val_str(c"tokenizer.ggml.pre", c"chatglm-bpe");
+            gguf_context
+                .remove_key("tokenizer.ggml.bos_token_id")
+                .expect("Qwen3 must declare a BOS token to remove");
+        },
+    )
+    .expect("the model without a BOS token must be writable");
+
+    let cluster = start_single_agent_cluster_with_desired_state(BalancerDesiredState {
+        model: agent_desired_model_at(&model_without_bos_path)
+            .expect("the model without a BOS token must be at an absolute path"),
+        ..qwen3_desired_state()
+    })
+    .await
+    .expect("the cluster must start");
+
+    let collected = cluster
+        .continue_from_conversation_history(
+            CancellationToken::new(),
+            &ContinueFromConversationHistoryParams {
+                add_generation_prompt: true,
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Text(
+                            "Name a primary color.".to_owned(),
+                        ),
+                        role: "user".to_owned(),
+                    },
+                ]),
+                enable_thinking: false,
+                grammar: None,
+                max_tokens: NonZeroU32::MIN,
+                parse_tool_calls: false,
+                tools: vec![],
+            },
+        )
+        .await
+        .expect("the inference request must be accepted");
+
+    assert_eq!(
+        collected
+            .summary()
+            .expect("the generation must finish with a summary")
+            .usage
+            .completion_tokens(),
+        1
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/agent_grammar_with_thinking_returns_incompatible_error.rs b/paddler_tests/tests/agent_grammar_with_thinking_returns_incompatible_error.rs
index 84d53253e..47804b99e 100644
--- a/paddler_tests/tests/agent_grammar_with_thinking_returns_incompatible_error.rs
+++ b/paddler_tests/tests/agent_grammar_with_thinking_returns_incompatible_error.rs
@@ -22,7 +22,7 @@ async fn agent_grammar_with_thinking_returns_incompatible_error() {
     let outcome = cluster
         .continue_from_conversation_history(CancellationToken::new(), &ContinueFromConversationHistoryParams {
             add_generation_prompt: true,
-            conversation_history: ConversationHistory::new(vec![ConversationMessage {
+            conversation_history: ConversationHistory::new(vec![ConversationMessage::Message {
                 content: ConversationMessageContent::Text("What is 2+2?".to_owned()),
                 role: "user".to_owned(),
             }]),
diff --git a/paddler_tests/tests/agent_hang_up_on_a_frame.rs b/paddler_tests/tests/agent_hang_up_on_a_frame.rs
new file mode 100644
index 000000000..8de65e6ef
--- /dev/null
+++ b/paddler_tests/tests/agent_hang_up_on_a_frame.rs
@@ -0,0 +1,27 @@
+use tokio_tungstenite::tungstenite::Message;
+
+use paddler_test_cluster_harness::untrusted_balancer_socket_server::UntrustedBalancerSocketServer;
+
+use crate::agent_runner_connected_to::agent_runner_connected_to;
+
+pub async fn agent_hang_up_on_a_frame(frame: Message) {
+    let untrusted_balancer = UntrustedBalancerSocketServer::bind()
+        .await
+        .expect("the untrusted balancer must bind");
+    let _agent_runner = agent_runner_connected_to(untrusted_balancer.management_addr, 1);
+    let mut first_connection = untrusted_balancer
+        .accept_agent()
+        .await
+        .expect("the agent must connect");
+
+    first_connection
+        .send(frame)
+        .await
+        .expect("the frame must be sent");
+    first_connection.wait_for_the_agent_to_hang_up().await;
+
+    untrusted_balancer
+        .accept_agent()
+        .await
+        .expect("the agent must reconnect after ending the connection");
+}
diff --git a/paddler_tests/tests/agent_isolates_concurrent_embedding_requests_per_client.rs b/paddler_tests/tests/agent_isolates_concurrent_embedding_requests_per_client.rs
index f46d417d1..97bd1f186 100644
--- a/paddler_tests/tests/agent_isolates_concurrent_embedding_requests_per_client.rs
+++ b/paddler_tests/tests/agent_isolates_concurrent_embedding_requests_per_client.rs
@@ -5,7 +5,6 @@ use std::collections::BTreeSet;
 use futures_util::future::join_all;
 use tokio_util::sync::CancellationToken;
 
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
 use paddler_messaging::embedding_input_document::EmbeddingInputDocument;
 use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
 use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
@@ -20,10 +19,7 @@ async fn agent_isolates_concurrent_embedding_requests_per_client() {
 
     let cluster = start_embedding_cluster(EmbeddingClusterParams {
         agents: vec![AgentConfig::single(4)],
-        inference_parameters: InferenceParameters {
-            enable_embeddings: true,
-            ..InferenceParameters::deterministic()
-        },
+
         ..EmbeddingClusterParams::default()
     })
     .await
diff --git a/paddler_tests/tests/agent_keeps_a_management_connection_after_a_stop_for_an_unknown_request.rs b/paddler_tests/tests/agent_keeps_a_management_connection_after_a_stop_for_an_unknown_request.rs
new file mode 100644
index 000000000..1319bb99d
--- /dev/null
+++ b/paddler_tests/tests/agent_keeps_a_management_connection_after_a_stop_for_an_unknown_request.rs
@@ -0,0 +1,21 @@
+use paddler_messaging::jsonrpc::response_envelope::ResponseEnvelope;
+use paddler_messaging::management_socket::agent::notification::Notification as AgentJsonRpcNotification;
+use paddler_messaging::management_socket::agent::response::Response as AgentJsonRpcResponse;
+
+use crate::agent_answer_after_a_notification::ANSWERED_REQUEST_ID;
+use crate::agent_answer_after_a_notification::agent_answer_after_a_notification;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agent_keeps_a_management_connection_after_a_stop_for_an_unknown_request() {
+    assert_eq!(
+        agent_answer_after_a_notification(AgentJsonRpcNotification::StopRespondingTo(
+            "unknown-request".to_owned(),
+        ))
+        .await,
+        ResponseEnvelope {
+            generated_by: None,
+            request_id: ANSWERED_REQUEST_ID.to_owned(),
+            response: AgentJsonRpcResponse::ModelMetadata(None),
+        }
+    );
+}
diff --git a/paddler_tests/tests/agent_keeps_a_management_connection_whose_balancer_reports_another_version.rs b/paddler_tests/tests/agent_keeps_a_management_connection_whose_balancer_reports_another_version.rs
new file mode 100644
index 000000000..883ea95f4
--- /dev/null
+++ b/paddler_tests/tests/agent_keeps_a_management_connection_whose_balancer_reports_another_version.rs
@@ -0,0 +1,22 @@
+use paddler_messaging::jsonrpc::response_envelope::ResponseEnvelope;
+use paddler_messaging::management_socket::agent::notification::Notification as AgentJsonRpcNotification;
+use paddler_messaging::management_socket::agent::notification_params::version_params::VersionParams;
+use paddler_messaging::management_socket::agent::response::Response as AgentJsonRpcResponse;
+
+use crate::agent_answer_after_a_notification::ANSWERED_REQUEST_ID;
+use crate::agent_answer_after_a_notification::agent_answer_after_a_notification;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agent_keeps_a_management_connection_whose_balancer_reports_another_version() {
+    assert_eq!(
+        agent_answer_after_a_notification(AgentJsonRpcNotification::Version(VersionParams {
+            version: "0.0.0-another-version".to_owned(),
+        }))
+        .await,
+        ResponseEnvelope {
+            generated_by: None,
+            request_id: ANSWERED_REQUEST_ID.to_owned(),
+            response: AgentJsonRpcResponse::ModelMetadata(None),
+        }
+    );
+}
diff --git a/paddler_tests/tests/agent_keeps_serving_in_flight_inference_when_the_desired_model_is_missing.rs b/paddler_tests/tests/agent_keeps_serving_in_flight_inference_when_the_desired_model_is_missing.rs
index 82deccc32..b555ac17a 100644
--- a/paddler_tests/tests/agent_keeps_serving_in_flight_inference_when_the_desired_model_is_missing.rs
+++ b/paddler_tests/tests/agent_keeps_serving_in_flight_inference_when_the_desired_model_is_missing.rs
@@ -1,23 +1,23 @@
 #![cfg(feature = "tests_that_use_llms")]
 
+use std::num::NonZeroU32;
+use std::path::Path;
+
 use futures_util::StreamExt as _;
+use futures_util::stream::iter;
 use tokio_util::sync::CancellationToken;
 
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
-use paddler_messaging::agent_desired_model::AgentDesiredModel;
 use paddler_messaging::agent_issue::AgentIssue;
 use paddler_messaging::balancer_desired_state::BalancerDesiredState;
 use paddler_messaging::generation_finish::GenerationFinish;
-use paddler_messaging::inference_client::message::Message;
-use paddler_messaging::inference_client::response::Response;
-use paddler_messaging::jsonrpc::response_envelope::ResponseEnvelope;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::agent_desired_model_at::agent_desired_model_at;
 use paddler_test_cluster_harness::collect_generated_tokens::collect_generated_tokens;
 use paddler_test_cluster_harness::is_unending_generation_text::is_unending_generation_text;
 use paddler_test_cluster_harness::unending_generation::unending_generation;
 use paddler_tests::start_cluster_with_qwen3_and_context_size::start_cluster_with_qwen3_and_context_size;
 
-const CONTEXT_SIZE_THAT_ENDS_THE_IN_FLIGHT_GENERATION: u32 = 256;
+const CONTEXT_SIZE_THAT_ENDS_THE_IN_FLIGHT_GENERATION: NonZeroU32 = NonZeroU32::new(256).unwrap();
 
 #[tokio::test(flavor = "multi_thread")]
 async fn agent_keeps_serving_in_flight_inference_when_the_desired_model_is_missing() {
@@ -32,28 +32,19 @@ async fn agent_keeps_serving_in_flight_inference_when_the_desired_model_is_missi
         .await
         .expect("the inference request must be accepted");
 
-    let Message::Response(ResponseEnvelope {
-        response: Response::GeneratedToken(first_token_result),
-        ..
-    }) = stream
+    let first_streamed_item = stream
         .next()
         .await
-        .expect("the in-flight request must stream a first token")
-        .expect("the message must be readable")
-    else {
-        panic!("the in-flight request must start by streaming a generated token");
-    };
+        .expect("the request must be streaming before the desired model goes missing");
 
     cluster
         .client_management
         .put_balancer_desired_state(
             CancellationToken::new(),
             &BalancerDesiredState {
-                chat_template_override: None,
-                inference_parameters: InferenceParameters::default(),
-                model: AgentDesiredModel::LocalToAgent("/nonexistent/model.gguf".to_owned()),
-                multimodal_projection: AgentDesiredModel::None,
-                use_chat_template_override: false,
+                model: agent_desired_model_at(Path::new("/nonexistent/model.gguf"))
+                    .expect("the missing model path must be absolute"),
+                ..BalancerDesiredState::default()
             },
         )
         .await
@@ -64,14 +55,9 @@ async fn agent_keeps_serving_in_flight_inference_when_the_desired_model_is_missi
         .await
         .expect("the agent must report the missing desired model");
 
-    let collected = collect_generated_tokens(stream)
+    let collected = collect_generated_tokens(Box::pin(iter([first_streamed_item]).chain(stream)))
         .await
         .expect("the generated tokens must be collected");
-    let streamed_text = format!(
-        "{}{}",
-        first_token_result.token_text().unwrap_or_default(),
-        collected.text
-    );
 
     assert_eq!(
         collected
@@ -81,8 +67,9 @@ async fn agent_keeps_serving_in_flight_inference_when_the_desired_model_is_missi
         GenerationFinish::ContextFull
     );
     assert!(
-        is_unending_generation_text(&streamed_text),
-        "{streamed_text:?}"
+        is_unending_generation_text(&collected.text),
+        "{:?}",
+        collected.text
     );
 
     cluster
diff --git a/paddler_tests/tests/agent_l2_normalized_embeddings_have_unit_norm.rs b/paddler_tests/tests/agent_l2_normalized_embeddings_have_unit_norm.rs
index 6a455a698..9f6ecfbcf 100644
--- a/paddler_tests/tests/agent_l2_normalized_embeddings_have_unit_norm.rs
+++ b/paddler_tests/tests/agent_l2_normalized_embeddings_have_unit_norm.rs
@@ -2,7 +2,6 @@
 
 use tokio_util::sync::CancellationToken;
 
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
 use paddler_messaging::embedding_input_document::EmbeddingInputDocument;
 use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
 use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
@@ -14,10 +13,7 @@ use paddler_tests::start_embedding_cluster::start_embedding_cluster;
 async fn agent_l2_normalized_embeddings_have_unit_norm() {
     let cluster = start_embedding_cluster(EmbeddingClusterParams {
         agents: vec![AgentConfig::single(1)],
-        inference_parameters: InferenceParameters {
-            enable_embeddings: true,
-            ..InferenceParameters::deterministic()
-        },
+
         ..EmbeddingClusterParams::default()
     })
     .await
@@ -42,10 +38,10 @@ async fn agent_l2_normalized_embeddings_have_unit_norm() {
 
     let produced = &collected.embeddings[0];
 
-    assert!(matches!(
+    assert_eq!(
         produced.embedding.normalization_method,
         EmbeddingNormalizationMethod::L2
-    ));
+    );
 
     let l2_norm: f32 = produced
         .embedding
diff --git a/paddler_tests/tests/agent_labels_embeddings_as_unnormalized_when_requested.rs b/paddler_tests/tests/agent_labels_embeddings_as_unnormalized_when_requested.rs
new file mode 100644
index 000000000..c0205db21
--- /dev/null
+++ b/paddler_tests/tests/agent_labels_embeddings_as_unnormalized_when_requested.rs
@@ -0,0 +1,47 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::embedding_input_document::EmbeddingInputDocument;
+use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
+use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::embedding_cluster_params::EmbeddingClusterParams;
+use paddler_tests::start_embedding_cluster::start_embedding_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agent_labels_embeddings_as_unnormalized_when_requested() {
+    let cluster = start_embedding_cluster(EmbeddingClusterParams {
+        agents: vec![AgentConfig::single(1)],
+
+        ..EmbeddingClusterParams::default()
+    })
+    .await
+    .expect("the cluster must start");
+
+    let collected = cluster
+        .generate_embedding_batch(
+            CancellationToken::new(),
+            &GenerateEmbeddingBatchParams {
+                input_batch: vec![EmbeddingInputDocument {
+                    content: "Testing no normalization on embeddings".to_owned(),
+                    id: "doc-none".to_owned(),
+                }],
+                normalization_method: EmbeddingNormalizationMethod::None,
+            },
+        )
+        .await
+        .expect("the embedding batch must be accepted");
+
+    assert_eq!(collected.embeddings.len(), 1);
+    assert!(collected.saw_done);
+    assert_eq!(
+        collected.embeddings[0].embedding.normalization_method,
+        EmbeddingNormalizationMethod::None
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/agent_numbers_tool_calls_across_separate_tool_call_sections.rs b/paddler_tests/tests/agent_numbers_tool_calls_across_separate_tool_call_sections.rs
new file mode 100644
index 000000000..42a5f1050
--- /dev/null
+++ b/paddler_tests/tests/agent_numbers_tool_calls_across_separate_tool_call_sections.rs
@@ -0,0 +1,69 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use llama_cpp_bindings_types::ToolCallArguments;
+use serde_json::json;
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::generated_token_result::GeneratedTokenResult;
+use paddler_messaging::generated_tool_call::GeneratedToolCall;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_tests::gbnf_literal::gbnf_literal;
+use paddler_tests::get_weather_tool::get_weather_tool;
+use paddler_tests::max_tokens_for_forced_text::max_tokens_for_forced_text;
+use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
+use paddler_tests::weather_question_conversation::weather_question_conversation;
+
+const TWO_TOOL_CALL_SECTIONS: &str = "<tool_call>\n{\"name\": \"get_weather\", \"arguments\": {\"location\": \"Paris\"}}\n</tool_call>\n<tool_call>\n{\"name\": \"get_weather\", \"arguments\": {\"location\": \"Berlin\"}}\n</tool_call>";
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agent_numbers_tool_calls_across_separate_tool_call_sections() {
+    let cluster = start_cluster_with_qwen3(AgentConfig::uniform(1, 1))
+        .await
+        .expect("the cluster must start");
+
+    let collected = cluster
+        .continue_from_conversation_history(
+            CancellationToken::new(),
+            &weather_question_conversation(
+                gbnf_literal(TWO_TOOL_CALL_SECTIONS),
+                max_tokens_for_forced_text(TWO_TOOL_CALL_SECTIONS),
+                vec![get_weather_tool()],
+            ),
+        )
+        .await
+        .expect("the inference request must be accepted");
+
+    let tool_call_sections: Vec<Vec<GeneratedToolCall>> = collected
+        .into_token_results()
+        .into_iter()
+        .filter_map(|token_result| match token_result {
+            GeneratedTokenResult::ToolCallParsed(generated_tool_calls) => {
+                Some(generated_tool_calls)
+            }
+            _ => None,
+        })
+        .collect();
+
+    assert_eq!(
+        tool_call_sections,
+        vec![
+            vec![GeneratedToolCall {
+                arguments: ToolCallArguments::ValidJson(json!({"location": "Paris"})),
+                id: "call_0".to_owned(),
+                index: 0,
+                name: "get_weather".to_owned(),
+            }],
+            vec![GeneratedToolCall {
+                arguments: ToolCallArguments::ValidJson(json!({"location": "Berlin"})),
+                id: "call_1".to_owned(),
+                index: 1,
+                name: "get_weather".to_owned(),
+            }],
+        ]
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/agent_openai_chat_completions_non_streaming_returns_text.rs b/paddler_tests/tests/agent_openai_chat_completions_non_streaming_returns_text.rs
deleted file mode 100644
index dbafdcf4c..000000000
--- a/paddler_tests/tests/agent_openai_chat_completions_non_streaming_returns_text.rs
+++ /dev/null
@@ -1,38 +0,0 @@
-#![cfg(feature = "tests_that_use_llms")]
-
-use serde_json::json;
-
-use paddler_test_cluster_harness::agent_config::AgentConfig;
-use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn agent_openai_chat_completions_non_streaming_returns_text() {
-    let cluster = start_cluster_with_qwen3(AgentConfig::uniform(1, 2))
-        .await
-        .expect("the cluster must start");
-
-    let body = cluster
-        .openai_chat_completion_non_streaming(&json!({
-            "model": "test",
-            "messages": [{"role": "user", "content": "Say hello"}],
-            "max_completion_tokens": 200,
-            "stream": false,
-        }))
-        .await
-        .expect("the OpenAI chat completion must succeed");
-
-    assert_eq!(body["object"], "chat.completion");
-    assert!(body["choices"].is_array());
-    assert!(
-        !body["choices"][0]["message"]["content"]
-            .as_str()
-            .unwrap_or("")
-            .is_empty(),
-        "response content should not be empty"
-    );
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_tests/tests/agent_openai_chat_completions_streaming_returns_chunks.rs b/paddler_tests/tests/agent_openai_chat_completions_streaming_returns_chunks.rs
deleted file mode 100644
index 0d32cbec1..000000000
--- a/paddler_tests/tests/agent_openai_chat_completions_streaming_returns_chunks.rs
+++ /dev/null
@@ -1,31 +0,0 @@
-#![cfg(feature = "tests_that_use_llms")]
-
-use serde_json::json;
-
-use paddler_test_cluster_harness::agent_config::AgentConfig;
-use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn agent_openai_chat_completions_streaming_returns_chunks() {
-    let cluster = start_cluster_with_qwen3(AgentConfig::uniform(1, 2))
-        .await
-        .expect("the cluster must start");
-
-    let chunks = cluster
-        .openai_chat_completion_streaming(&json!({
-            "model": "test",
-            "messages": [{"role": "user", "content": "Say hello"}],
-            "max_completion_tokens": 10,
-            "stream": true,
-        }))
-        .await
-        .expect("the OpenAI chat completion stream must succeed");
-
-    assert!(!chunks.is_empty(), "should have received streaming chunks");
-    assert_eq!(chunks[0]["object"], "chat.completion.chunk");
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_tests/tests/agent_raw_prompt_with_gbnf_grammar_constrains_output.rs b/paddler_tests/tests/agent_raw_prompt_with_gbnf_grammar_constrains_output.rs
index e2872ffaa..defd460fa 100644
--- a/paddler_tests/tests/agent_raw_prompt_with_gbnf_grammar_constrains_output.rs
+++ b/paddler_tests/tests/agent_raw_prompt_with_gbnf_grammar_constrains_output.rs
@@ -28,10 +28,11 @@ async fn agent_raw_prompt_with_gbnf_grammar_constrains_output() {
         })
         .await.expect("the inference request must be accepted");
 
+    let answer = collected.text.as_str();
+
     assert!(
-        collected.text == "yes" || collected.text == "no",
-        "GBNF grammar should constrain output to yes/no; got {:?}",
-        collected.text
+        ["yes", "no"].contains(&answer),
+        "GBNF grammar should constrain output to yes/no; got {answer:?}"
     );
 
     cluster
diff --git a/paddler_tests/tests/agent_reads_media_marker_text_in_an_image_prompt_as_plain_text.rs b/paddler_tests/tests/agent_reads_media_marker_text_in_an_image_prompt_as_plain_text.rs
new file mode 100644
index 000000000..cf6142cbb
--- /dev/null
+++ b/paddler_tests/tests/agent_reads_media_marker_text_in_an_image_prompt_as_plain_text.rs
@@ -0,0 +1,53 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::generation_finish::GenerationFinish;
+use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::load_test_image_data_uri::load_test_image_data_uri;
+use paddler_tests::gbnf_literal::gbnf_literal;
+use paddler_tests::max_tokens_for_forced_text::max_tokens_for_forced_text;
+use paddler_tests::single_image_question::SingleImageQuestion;
+use paddler_tests::start_cluster_with_smolvlm2::start_cluster_with_smolvlm2;
+
+const ANSWER: &str = "two llamas";
+const QUESTION_QUOTING_THE_LLAMA_CPP_DEFAULT_MEDIA_MARKER: &str =
+    "What does <__media__> mean next to this image?";
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agent_reads_media_marker_text_in_an_image_prompt_as_plain_text() {
+    let cluster = start_cluster_with_smolvlm2(vec![AgentConfig::single(1)])
+        .await
+        .expect("the cluster must start");
+
+    let collected = cluster
+        .continue_from_conversation_history(
+            CancellationToken::new(),
+            &ContinueFromConversationHistoryParams {
+                grammar: Some(gbnf_literal(ANSWER)),
+                ..SingleImageQuestion {
+                    image_data_uri: load_test_image_data_uri().expect("the test image must load"),
+                    max_tokens: max_tokens_for_forced_text(ANSWER),
+                    question: QUESTION_QUOTING_THE_LLAMA_CPP_DEFAULT_MEDIA_MARKER,
+                }
+                .into_conversation_params()
+            },
+        )
+        .await
+        .expect("the inference request must be accepted");
+
+    assert_eq!(collected.text, ANSWER);
+    assert_eq!(
+        collected
+            .summary()
+            .expect("the generation must end with a summary")
+            .finish,
+        GenerationFinish::EndOfGeneration
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/agent_rejects_a_multimodal_conversation_containing_a_nul_byte.rs b/paddler_tests/tests/agent_rejects_a_multimodal_conversation_containing_a_nul_byte.rs
new file mode 100644
index 000000000..9ea8d5ad6
--- /dev/null
+++ b/paddler_tests/tests/agent_rejects_a_multimodal_conversation_containing_a_nul_byte.rs
@@ -0,0 +1,66 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::num::NonZeroU32;
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::conversation_history::ConversationHistory;
+use paddler_messaging::conversation_message::ConversationMessage;
+use paddler_messaging::conversation_message_content::ConversationMessageContent;
+use paddler_messaging::conversation_message_content_part::ConversationMessageContentPart;
+use paddler_messaging::generated_token_result::GeneratedTokenResult;
+use paddler_messaging::image_url::ImageUrl;
+use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::load_test_image_data_uri::load_test_image_data_uri;
+use paddler_tests::start_cluster_with_smolvlm2::start_cluster_with_smolvlm2;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agent_rejects_a_multimodal_conversation_containing_a_nul_byte() {
+    let cluster = start_cluster_with_smolvlm2(vec![AgentConfig::single(1)])
+        .await
+        .expect("the cluster must start");
+
+    let collected = cluster
+        .continue_from_conversation_history(
+            CancellationToken::new(),
+            &ContinueFromConversationHistoryParams {
+                add_generation_prompt: true,
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Parts(vec![
+                            ConversationMessageContentPart::ImageUrl {
+                                image_url: ImageUrl {
+                                    url: load_test_image_data_uri()
+                                        .expect("the test image must load"),
+                                },
+                            },
+                            ConversationMessageContentPart::Text {
+                                text: "before\0after".to_owned(),
+                            },
+                        ]),
+                        role: "user".to_owned(),
+                    },
+                ]),
+                enable_thinking: false,
+                grammar: None,
+                max_tokens: NonZeroU32::MIN,
+                parse_tool_calls: false,
+                tools: vec![],
+            },
+        )
+        .await
+        .expect("the inference request must be accepted");
+
+    assert_eq!(
+        collected.into_token_results(),
+        vec![GeneratedTokenResult::MultimodalTokenizationFailed(
+            "test-agent: failed to tokenize multimodal input: Failed to create CString from input text: nul byte found in provided data at position: 6".to_owned()
+        )]
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/agent_rejects_buffered_conversation_after_switching_to_embeddings.rs b/paddler_tests/tests/agent_rejects_buffered_conversation_after_switching_to_embeddings.rs
deleted file mode 100644
index a5fe28f79..000000000
--- a/paddler_tests/tests/agent_rejects_buffered_conversation_after_switching_to_embeddings.rs
+++ /dev/null
@@ -1,57 +0,0 @@
-#![cfg(feature = "tests_that_use_llms")]
-
-use std::num::NonZeroU32;
-
-use tokio_util::sync::CancellationToken;
-
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
-use paddler_messaging::conversation_history::ConversationHistory;
-use paddler_messaging::conversation_message::ConversationMessage;
-use paddler_messaging::conversation_message_content::ConversationMessageContent;
-use paddler_messaging::generated_token_result::GeneratedTokenResult;
-use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
-use paddler_test_cluster_harness::agent_config::AgentConfig;
-use paddler_tests::load_qwen3_after_request_is_buffered::load_qwen3_after_request_is_buffered;
-use paddler_tests::start_cluster_without_model::start_cluster_without_model;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn agent_rejects_buffered_conversation_after_switching_to_embeddings() {
-    let mut cluster = start_cluster_without_model(
-        AgentConfig::uniform(1, 1),
-        InferenceParameters::deterministic(),
-    )
-    .await
-    .expect("the cluster must start");
-
-    let request = cluster.continue_from_conversation_history(
-        CancellationToken::new(),
-        &ContinueFromConversationHistoryParams {
-            add_generation_prompt: true,
-            conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                content: ConversationMessageContent::Text("Hello".to_owned()),
-                role: "user".to_owned(),
-            }]),
-            enable_thinking: false,
-            grammar: None,
-            max_tokens: NonZeroU32::new(4).unwrap(),
-            parse_tool_calls: false,
-            tools: vec![],
-        },
-    );
-    let collected = load_qwen3_after_request_is_buffered(&mut cluster, request, true)
-        .await
-        .expect("the model must load after the request is buffered");
-
-    assert_eq!(
-        collected.into_token_results(),
-        vec![GeneratedTokenResult::TokenGenerationDisabled(
-            "test-agent-0: token generation is disabled because this agent is running in embeddings-only mode"
-                .to_owned()
-        )]
-    );
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_tests/tests/agent_rejects_buffered_embedding_batch_after_disabling_embeddings.rs b/paddler_tests/tests/agent_rejects_buffered_embedding_batch_after_disabling_embeddings.rs
deleted file mode 100644
index d5f0d369f..000000000
--- a/paddler_tests/tests/agent_rejects_buffered_embedding_batch_after_disabling_embeddings.rs
+++ /dev/null
@@ -1,46 +0,0 @@
-#![cfg(feature = "tests_that_use_llms")]
-
-use tokio_util::sync::CancellationToken;
-
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
-use paddler_messaging::embedding_input_document::EmbeddingInputDocument;
-use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
-use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
-use paddler_test_cluster_harness::agent_config::AgentConfig;
-use paddler_tests::load_qwen3_after_request_is_buffered::load_qwen3_after_request_is_buffered;
-use paddler_tests::start_cluster_without_model::start_cluster_without_model;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn agent_rejects_buffered_embedding_batch_after_disabling_embeddings() {
-    let mut cluster = start_cluster_without_model(
-        AgentConfig::uniform(1, 1),
-        InferenceParameters {
-            enable_embeddings: true,
-            ..InferenceParameters::deterministic()
-        },
-    )
-    .await
-    .expect("the cluster must start");
-
-    let request = cluster.generate_embedding_batch(
-        CancellationToken::new(),
-        &GenerateEmbeddingBatchParams {
-            input_batch: vec![EmbeddingInputDocument {
-                content: "Hello".to_owned(),
-                id: "greeting".to_owned(),
-            }],
-            normalization_method: EmbeddingNormalizationMethod::None,
-        },
-    );
-    let collected = load_qwen3_after_request_is_buffered(&mut cluster, request, false)
-        .await
-        .expect("the model must load after the request is buffered");
-
-    assert!(collected.embeddings_disabled);
-    assert!(collected.embeddings.is_empty());
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_tests/tests/agent_rejects_buffered_raw_prompt_after_switching_to_embeddings.rs b/paddler_tests/tests/agent_rejects_buffered_raw_prompt_after_switching_to_embeddings.rs
deleted file mode 100644
index 5131e1ac9..000000000
--- a/paddler_tests/tests/agent_rejects_buffered_raw_prompt_after_switching_to_embeddings.rs
+++ /dev/null
@@ -1,47 +0,0 @@
-#![cfg(feature = "tests_that_use_llms")]
-
-use std::num::NonZeroU32;
-
-use tokio_util::sync::CancellationToken;
-
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
-use paddler_messaging::generated_token_result::GeneratedTokenResult;
-use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
-use paddler_test_cluster_harness::agent_config::AgentConfig;
-use paddler_tests::load_qwen3_after_request_is_buffered::load_qwen3_after_request_is_buffered;
-use paddler_tests::start_cluster_without_model::start_cluster_without_model;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn agent_rejects_buffered_raw_prompt_after_switching_to_embeddings() {
-    let mut cluster = start_cluster_without_model(
-        AgentConfig::uniform(1, 1),
-        InferenceParameters::deterministic(),
-    )
-    .await
-    .expect("the cluster must start");
-
-    let request = cluster.continue_from_raw_prompt(
-        CancellationToken::new(),
-        &ContinueFromRawPromptParams {
-            grammar: None,
-            max_tokens: NonZeroU32::new(4).unwrap(),
-            raw_prompt: "Hello".to_owned(),
-        },
-    );
-    let collected = load_qwen3_after_request_is_buffered(&mut cluster, request, true)
-        .await
-        .expect("the model must load after the request is buffered");
-
-    assert_eq!(
-        collected.into_token_results(),
-        vec![GeneratedTokenResult::TokenGenerationDisabled(
-            "test-agent-0: token generation is disabled because this agent is running in embeddings-only mode"
-                .to_owned()
-        )]
-    );
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_tests/tests/agent_rejects_conversation_exceeding_sequence_context.rs b/paddler_tests/tests/agent_rejects_conversation_exceeding_sequence_context.rs
index 41d061c18..f063e665a 100644
--- a/paddler_tests/tests/agent_rejects_conversation_exceeding_sequence_context.rs
+++ b/paddler_tests/tests/agent_rejects_conversation_exceeding_sequence_context.rs
@@ -10,9 +10,10 @@ use paddler_messaging::conversation_message_content::ConversationMessageContent;
 use paddler_messaging::generated_token_result::GeneratedTokenResult;
 use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_tests::oversized_text::oversized_text;
 use paddler_tests::start_cluster_with_qwen3_and_context_size::start_cluster_with_qwen3_and_context_size;
 
-const SEQUENCE_CONTEXT_SIZE: u32 = 256;
+const SEQUENCE_CONTEXT_SIZE: NonZeroU32 = NonZeroU32::new(256).unwrap();
 
 #[tokio::test(flavor = "multi_thread")]
 async fn agent_rejects_conversation_exceeding_sequence_context() {
@@ -28,12 +29,12 @@ async fn agent_rejects_conversation_exceeding_sequence_context() {
             CancellationToken::new(),
             &ContinueFromConversationHistoryParams {
                 add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Text(
-                        "The quick brown fox jumps over the lazy dog. ".repeat(40),
-                    ),
-                    role: "user".to_owned(),
-                }]),
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Text(oversized_text()),
+                        role: "user".to_owned(),
+                    },
+                ]),
                 enable_thinking: false,
                 grammar: None,
                 max_tokens: NonZeroU32::new(20).unwrap(),
@@ -47,8 +48,8 @@ async fn agent_rejects_conversation_exceeding_sequence_context() {
     assert!(collected.token_results.iter().any(|result| matches!(
         &result.token_result,
         GeneratedTokenResult::PromptExceedsContextSize(details)
-            if details.sequence_context_size == SEQUENCE_CONTEXT_SIZE
-                && details.prompt_tokens > SEQUENCE_CONTEXT_SIZE as usize
+            if details.sequence_context_size == SEQUENCE_CONTEXT_SIZE.get()
+                && details.prompt_tokens > SEQUENCE_CONTEXT_SIZE.get() as usize
     )));
 
     cluster
diff --git a/paddler_tests/tests/agent_rejects_embedding_document_containing_nul_byte.rs b/paddler_tests/tests/agent_rejects_embedding_document_containing_nul_byte.rs
index 563e8d14b..702bb2431 100644
--- a/paddler_tests/tests/agent_rejects_embedding_document_containing_nul_byte.rs
+++ b/paddler_tests/tests/agent_rejects_embedding_document_containing_nul_byte.rs
@@ -1,23 +1,28 @@
 #![cfg(feature = "tests_that_use_llms")]
 
+use std::ffi::CString;
+
+use llama_cpp_bindings::error::StringToTokenError;
 use tokio_util::sync::CancellationToken;
 
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
+use paddler_agent_embeddings::embedding_error::EmbeddingError;
+use paddler_agent_runtime::rejection_description::rejection_description;
 use paddler_messaging::embedding_input_document::EmbeddingInputDocument;
 use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
+use paddler_messaging::embedding_result::EmbeddingResult;
 use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
 use paddler_test_cluster_harness::embedding_cluster_params::EmbeddingClusterParams;
 use paddler_tests::start_embedding_cluster::start_embedding_cluster;
 
+const DOCUMENT_CONTENT_WITH_NUL_BYTE: &str = "before\0after";
+const DOCUMENT_ID: &str = "nul-document";
+
 #[tokio::test(flavor = "multi_thread")]
 async fn agent_rejects_embedding_document_containing_nul_byte() {
     let cluster = start_embedding_cluster(EmbeddingClusterParams {
         agents: vec![AgentConfig::single(1)],
-        inference_parameters: InferenceParameters {
-            enable_embeddings: true,
-            ..InferenceParameters::deterministic()
-        },
+
         ..EmbeddingClusterParams::default()
     })
     .await
@@ -28,8 +33,8 @@ async fn agent_rejects_embedding_document_containing_nul_byte() {
             CancellationToken::new(),
             &GenerateEmbeddingBatchParams {
                 input_batch: vec![EmbeddingInputDocument {
-                    content: "before\0after".to_owned(),
-                    id: "nul-document".to_owned(),
+                    content: DOCUMENT_CONTENT_WITH_NUL_BYTE.to_owned(),
+                    id: DOCUMENT_ID.to_owned(),
                 }],
                 normalization_method: EmbeddingNormalizationMethod::None,
             },
@@ -37,12 +42,21 @@ async fn agent_rejects_embedding_document_containing_nul_byte() {
         .await
         .expect("the embedding batch must be accepted");
 
+    let nul_error = CString::new(DOCUMENT_CONTENT_WITH_NUL_BYTE)
+        .expect_err("the document content must contain a nul byte");
+
     assert!(collected.embeddings.is_empty());
-    assert_eq!(collected.errors.len(), 1);
-    assert!(
-        collected.errors[0].contains("\"nul-document\""),
-        "the error must name the document that could not be tokenized; got {:?}",
-        collected.errors[0],
+    assert_eq!(
+        collected.failures,
+        vec![EmbeddingResult::InputTokenizationFailed(
+            rejection_description(
+                Some("test-agent"),
+                &EmbeddingError::InputTokenizationFailed {
+                    source_document_id: DOCUMENT_ID.to_owned(),
+                    source: StringToTokenError::NulError(nul_error),
+                }
+            )
+        )]
     );
 
     cluster
diff --git a/paddler_tests/tests/agent_rejects_gemma_3_image_exceeding_the_micro_batch.rs b/paddler_tests/tests/agent_rejects_gemma_3_image_exceeding_the_micro_batch.rs
index cdc1df4f2..94f00ae2d 100644
--- a/paddler_tests/tests/agent_rejects_gemma_3_image_exceeding_the_micro_batch.rs
+++ b/paddler_tests/tests/agent_rejects_gemma_3_image_exceeding_the_micro_batch.rs
@@ -4,6 +4,7 @@ use std::num::NonZeroU32;
 
 use tokio_util::sync::CancellationToken;
 
+use paddler_inference_parameters::batch_size::BatchSize;
 use paddler_messaging::conversation_history::ConversationHistory;
 use paddler_messaging::conversation_message::ConversationMessage;
 use paddler_messaging::conversation_message_content::ConversationMessageContent;
@@ -24,7 +25,8 @@ const N_BATCH_SMALLER_THAN_ONE_IMAGE: u32 = 128;
 async fn agent_rejects_gemma_3_image_exceeding_the_micro_batch() {
     let cluster = start_cluster_with_gemma_3_and_mmproj_and_n_batch(
         vec![AgentConfig::single(1)],
-        N_BATCH_SMALLER_THAN_ONE_IMAGE,
+        BatchSize::try_from(N_BATCH_SMALLER_THAN_ONE_IMAGE)
+            .expect("the batch size must be within the supported range"),
     )
     .await
     .expect("the cluster must start");
@@ -34,20 +36,22 @@ async fn agent_rejects_gemma_3_image_exceeding_the_micro_batch() {
             CancellationToken::new(),
             &ContinueFromConversationHistoryParams {
                 add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Parts(vec![
-                        ConversationMessageContentPart::ImageUrl {
-                            image_url: ImageUrl {
-                                url: load_fixture_data_uri("sarnow.jpeg", "image/jpeg")
-                                    .expect("the test fixture must load"),
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Parts(vec![
+                            ConversationMessageContentPart::ImageUrl {
+                                image_url: ImageUrl {
+                                    url: load_fixture_data_uri("sarnow.jpeg", "image/jpeg")
+                                        .expect("the test fixture must load"),
+                                },
                             },
-                        },
-                        ConversationMessageContentPart::Text {
-                            text: "Describe this image.".to_owned(),
-                        },
-                    ]),
-                    role: "user".to_owned(),
-                }]),
+                            ConversationMessageContentPart::Text {
+                                text: "Describe this image.".to_owned(),
+                            },
+                        ]),
+                        role: "user".to_owned(),
+                    },
+                ]),
                 enable_thinking: false,
                 grammar: None,
                 max_tokens: MAX_TOKENS,
diff --git a/paddler_tests/tests/agent_rejects_image_prompt_containing_media_marker_text.rs b/paddler_tests/tests/agent_rejects_image_prompt_containing_media_marker_text.rs
deleted file mode 100644
index 519de701d..000000000
--- a/paddler_tests/tests/agent_rejects_image_prompt_containing_media_marker_text.rs
+++ /dev/null
@@ -1,66 +0,0 @@
-#![cfg(feature = "tests_that_use_llms")]
-
-use std::num::NonZeroU32;
-
-use tokio_util::sync::CancellationToken;
-
-use paddler_messaging::conversation_history::ConversationHistory;
-use paddler_messaging::conversation_message::ConversationMessage;
-use paddler_messaging::conversation_message_content::ConversationMessageContent;
-use paddler_messaging::conversation_message_content_part::ConversationMessageContentPart;
-use paddler_messaging::generated_token_result::GeneratedTokenResult;
-use paddler_messaging::image_url::ImageUrl;
-use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
-use paddler_test_cluster_harness::agent_config::AgentConfig;
-use paddler_test_cluster_harness::load_test_image_data_uri::load_test_image_data_uri;
-use paddler_tests::start_cluster_with_smolvlm2::start_cluster_with_smolvlm2;
-
-const LLAMA_CPP_DEFAULT_MEDIA_MARKER: &str = "<__media__>";
-
-#[tokio::test(flavor = "multi_thread")]
-async fn agent_rejects_image_prompt_containing_media_marker_text() {
-    let cluster = start_cluster_with_smolvlm2(vec![AgentConfig::single(1)])
-        .await
-        .expect("the cluster must start");
-
-    let collected = cluster
-        .continue_from_conversation_history(
-            CancellationToken::new(),
-            &ContinueFromConversationHistoryParams {
-                add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Parts(vec![
-                        ConversationMessageContentPart::ImageUrl {
-                            image_url: ImageUrl {
-                                url: load_test_image_data_uri().expect("the test image must load"),
-                            },
-                        },
-                        ConversationMessageContentPart::Text {
-                            text: format!("What does {LLAMA_CPP_DEFAULT_MEDIA_MARKER} mean?"),
-                        },
-                    ]),
-                    role: "user".to_owned(),
-                }]),
-                enable_thinking: false,
-                grammar: None,
-                max_tokens: NonZeroU32::new(20).unwrap(),
-                parse_tool_calls: false,
-                tools: vec![],
-            },
-        )
-        .await
-        .expect("the inference request must be accepted");
-
-    assert_eq!(
-        collected.into_token_results(),
-        vec![GeneratedTokenResult::MultimodalTokenizationFailed(
-            "test-agent: failed to tokenize multimodal input: media preprocessing failed (image or audio)"
-                .to_owned(),
-        )]
-    );
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_tests/tests/agent_rejects_image_prompt_exceeding_sequence_context.rs b/paddler_tests/tests/agent_rejects_image_prompt_exceeding_sequence_context.rs
index 62721d11e..1c0465dff 100644
--- a/paddler_tests/tests/agent_rejects_image_prompt_exceeding_sequence_context.rs
+++ b/paddler_tests/tests/agent_rejects_image_prompt_exceeding_sequence_context.rs
@@ -15,7 +15,7 @@ use paddler_test_cluster_harness::agent_config::AgentConfig;
 use paddler_test_cluster_harness::load_test_image_data_uri::load_test_image_data_uri;
 use paddler_tests::start_cluster_with_smolvlm2_and_context_size::start_cluster_with_smolvlm2_and_context_size;
 
-const SMALLEST_PADDED_SEQUENCE_CONTEXT_SIZE: u32 = 256;
+const SMALLEST_PADDED_SEQUENCE_CONTEXT_SIZE: NonZeroU32 = NonZeroU32::new(256).unwrap();
 
 #[tokio::test(flavor = "multi_thread")]
 async fn agent_rejects_image_prompt_exceeding_sequence_context() {
@@ -34,18 +34,20 @@ async fn agent_rejects_image_prompt_exceeding_sequence_context() {
             CancellationToken::new(),
             &ContinueFromConversationHistoryParams {
                 add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Parts(vec![
-                        ConversationMessageContentPart::ImageUrl {
-                            image_url: image_url.clone(),
-                        },
-                        ConversationMessageContentPart::ImageUrl { image_url },
-                        ConversationMessageContentPart::Text {
-                            text: "Compare these images.".to_owned(),
-                        },
-                    ]),
-                    role: "user".to_owned(),
-                }]),
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Parts(vec![
+                            ConversationMessageContentPart::ImageUrl {
+                                image_url: image_url.clone(),
+                            },
+                            ConversationMessageContentPart::ImageUrl { image_url },
+                            ConversationMessageContentPart::Text {
+                                text: "Compare these images.".to_owned(),
+                            },
+                        ]),
+                        role: "user".to_owned(),
+                    },
+                ]),
                 enable_thinking: false,
                 grammar: None,
                 max_tokens: NonZeroU32::new(20).unwrap(),
@@ -59,8 +61,8 @@ async fn agent_rejects_image_prompt_exceeding_sequence_context() {
     assert!(collected.token_results.iter().any(|result| matches!(
         &result.token_result,
         GeneratedTokenResult::PromptExceedsContextSize(details)
-            if details.sequence_context_size == SMALLEST_PADDED_SEQUENCE_CONTEXT_SIZE
-                && details.prompt_tokens > SMALLEST_PADDED_SEQUENCE_CONTEXT_SIZE as usize
+            if details.sequence_context_size == SMALLEST_PADDED_SEQUENCE_CONTEXT_SIZE.get()
+                && details.prompt_tokens > SMALLEST_PADDED_SEQUENCE_CONTEXT_SIZE.get() as usize
     )));
 
     cluster
diff --git a/paddler_tests/tests/agent_rejects_oversized_conversation_before_compiling_its_tools.rs b/paddler_tests/tests/agent_rejects_oversized_conversation_before_compiling_its_tools.rs
index 311da306f..67b8cbed8 100644
--- a/paddler_tests/tests/agent_rejects_oversized_conversation_before_compiling_its_tools.rs
+++ b/paddler_tests/tests/agent_rejects_oversized_conversation_before_compiling_its_tools.rs
@@ -17,9 +17,10 @@ use paddler_messaging::request_params::continue_from_conversation_history_params
 use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters::Parameters;
 use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters_schema::validated_parameters_schema::ValidatedParametersSchema;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_tests::oversized_text::oversized_text;
 use paddler_tests::start_cluster_with_qwen3_and_context_size::start_cluster_with_qwen3_and_context_size;
 
-const SEQUENCE_CONTEXT_SIZE: u32 = 256;
+const SEQUENCE_CONTEXT_SIZE: NonZeroU32 = NonZeroU32::new(256).unwrap();
 
 #[tokio::test(flavor = "multi_thread")]
 async fn agent_rejects_oversized_conversation_before_compiling_its_tools() {
@@ -38,12 +39,12 @@ async fn agent_rejects_oversized_conversation_before_compiling_its_tools() {
             CancellationToken::new(),
             &ContinueFromConversationHistoryParams {
                 add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Text(
-                        "The quick brown fox jumps over the lazy dog. ".repeat(40),
-                    ),
-                    role: "user".to_owned(),
-                }]),
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Text(oversized_text()),
+                        role: "user".to_owned(),
+                    },
+                ]),
                 enable_thinking: false,
                 grammar: None,
                 max_tokens: NonZeroU32::new(20).unwrap(),
@@ -73,7 +74,7 @@ async fn agent_rejects_oversized_conversation_before_compiling_its_tools() {
             .collect::<Vec<_>>()
             .as_slice(),
         [GeneratedTokenResult::PromptExceedsContextSize(details)]
-            if details.sequence_context_size == SEQUENCE_CONTEXT_SIZE
+            if details.sequence_context_size == SEQUENCE_CONTEXT_SIZE.get()
     ));
 
     cluster
diff --git a/paddler_tests/tests/agent_rejects_oversized_conversation_before_converting_its_grammar.rs b/paddler_tests/tests/agent_rejects_oversized_conversation_before_converting_its_grammar.rs
index 2a4b9483d..8f69527bd 100644
--- a/paddler_tests/tests/agent_rejects_oversized_conversation_before_converting_its_grammar.rs
+++ b/paddler_tests/tests/agent_rejects_oversized_conversation_before_converting_its_grammar.rs
@@ -11,9 +11,10 @@ use paddler_messaging::generated_token_result::GeneratedTokenResult;
 use paddler_messaging::grammar_constraint::GrammarConstraint;
 use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_tests::oversized_text::oversized_text;
 use paddler_tests::start_cluster_with_qwen3_and_context_size::start_cluster_with_qwen3_and_context_size;
 
-const SEQUENCE_CONTEXT_SIZE: u32 = 256;
+const SEQUENCE_CONTEXT_SIZE: NonZeroU32 = NonZeroU32::new(256).unwrap();
 
 #[tokio::test(flavor = "multi_thread")]
 async fn agent_rejects_oversized_conversation_before_converting_its_grammar() {
@@ -29,12 +30,12 @@ async fn agent_rejects_oversized_conversation_before_converting_its_grammar() {
             CancellationToken::new(),
             &ContinueFromConversationHistoryParams {
                 add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Text(
-                        "The quick brown fox jumps over the lazy dog. ".repeat(40),
-                    ),
-                    role: "user".to_owned(),
-                }]),
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Text(oversized_text()),
+                        role: "user".to_owned(),
+                    },
+                ]),
                 enable_thinking: false,
                 grammar: Some(GrammarConstraint::JsonSchema {
                     schema: "not a json schema".to_owned(),
@@ -55,7 +56,7 @@ async fn agent_rejects_oversized_conversation_before_converting_its_grammar() {
             .collect::<Vec<_>>()
             .as_slice(),
         [GeneratedTokenResult::PromptExceedsContextSize(details)]
-            if details.sequence_context_size == SEQUENCE_CONTEXT_SIZE
+            if details.sequence_context_size == SEQUENCE_CONTEXT_SIZE.get()
     ));
 
     cluster
diff --git a/paddler_tests/tests/agent_rejects_oversized_raw_prompt_before_converting_its_grammar.rs b/paddler_tests/tests/agent_rejects_oversized_raw_prompt_before_converting_its_grammar.rs
index 98523e444..13a4e8f6a 100644
--- a/paddler_tests/tests/agent_rejects_oversized_raw_prompt_before_converting_its_grammar.rs
+++ b/paddler_tests/tests/agent_rejects_oversized_raw_prompt_before_converting_its_grammar.rs
@@ -8,9 +8,10 @@ use paddler_messaging::generated_token_result::GeneratedTokenResult;
 use paddler_messaging::grammar_constraint::GrammarConstraint;
 use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_tests::oversized_text::oversized_text;
 use paddler_tests::start_cluster_with_qwen3_and_context_size::start_cluster_with_qwen3_and_context_size;
 
-const SEQUENCE_CONTEXT_SIZE: u32 = 256;
+const SEQUENCE_CONTEXT_SIZE: NonZeroU32 = NonZeroU32::new(256).unwrap();
 
 #[tokio::test(flavor = "multi_thread")]
 async fn agent_rejects_oversized_raw_prompt_before_converting_its_grammar() {
@@ -29,7 +30,7 @@ async fn agent_rejects_oversized_raw_prompt_before_converting_its_grammar() {
                     schema: "not a json schema".to_owned(),
                 }),
                 max_tokens: NonZeroU32::new(20).unwrap(),
-                raw_prompt: "The quick brown fox jumps over the lazy dog. ".repeat(40),
+                raw_prompt: oversized_text(),
             },
         )
         .await
@@ -43,7 +44,7 @@ async fn agent_rejects_oversized_raw_prompt_before_converting_its_grammar() {
             .collect::<Vec<_>>()
             .as_slice(),
         [GeneratedTokenResult::PromptExceedsContextSize(details)]
-            if details.sequence_context_size == SEQUENCE_CONTEXT_SIZE
+            if details.sequence_context_size == SEQUENCE_CONTEXT_SIZE.get()
     ));
 
     cluster
diff --git a/paddler_tests/tests/agent_rejects_raw_prompt_exceeding_sequence_context.rs b/paddler_tests/tests/agent_rejects_raw_prompt_exceeding_sequence_context.rs
index 0d53fcf81..b89f74b42 100644
--- a/paddler_tests/tests/agent_rejects_raw_prompt_exceeding_sequence_context.rs
+++ b/paddler_tests/tests/agent_rejects_raw_prompt_exceeding_sequence_context.rs
@@ -4,13 +4,15 @@ use std::num::NonZeroU32;
 
 use tokio_util::sync::CancellationToken;
 
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
+use paddler_inference_parameters::batch_size::BatchSize;
+use paddler_inference_parameters::model_runtime_parameters::ModelRuntimeParameters;
 use paddler_messaging::balancer_desired_state::BalancerDesiredState;
 use paddler_messaging::generated_token_result::GeneratedTokenResult;
 use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
-use paddler_tests::batch_size_within_context::batch_size_within_context;
+use paddler_tests::oversized_text::oversized_text;
 use paddler_tests::qwen3_desired_state::qwen3_desired_state;
 use paddler_tests::start_cluster::start_cluster;
 
@@ -21,19 +23,18 @@ async fn agent_rejects_raw_prompt_exceeding_sequence_context() {
     let desired_state = qwen3_desired_state();
     let cluster = start_cluster(ClusterParams {
         agents: vec![AgentConfig::single(1)],
-        desired_state: Some(BalancerDesiredState {
-            inference_parameters: InferenceParameters {
+        desired_state: ClusterDesiredState::Apply(Box::new(BalancerDesiredState {
+            model_runtime_parameters: ModelRuntimeParameters {
                 context_size: NonZeroU32::try_from(SEQUENCE_CONTEXT_SIZE)
                     .expect("the value must fit its target type"),
-                n_batch: batch_size_within_context(
+                n_batch: BatchSize::default().at_most(
                     NonZeroU32::try_from(SEQUENCE_CONTEXT_SIZE)
                         .expect("the value must fit its target type"),
-                )
-                .expect("the batch size must fit the context"),
-                ..desired_state.inference_parameters
+                ),
+                ..desired_state.model_runtime_parameters
             },
             ..desired_state
-        }),
+        })),
         wait_for_slots_ready: true,
         ..ClusterParams::default()
     })
@@ -46,7 +47,7 @@ async fn agent_rejects_raw_prompt_exceeding_sequence_context() {
             &ContinueFromRawPromptParams {
                 grammar: None,
                 max_tokens: NonZeroU32::new(20).unwrap(),
-                raw_prompt: "The quick brown fox jumps over the lazy dog. ".repeat(40),
+                raw_prompt: oversized_text(),
             },
         )
         .await
diff --git a/paddler_tests/tests/agent_rejects_structurally_invalid_tool_schema.rs b/paddler_tests/tests/agent_rejects_structurally_invalid_tool_schema.rs
index c445bc987..d8250dd7f 100644
--- a/paddler_tests/tests/agent_rejects_structurally_invalid_tool_schema.rs
+++ b/paddler_tests/tests/agent_rejects_structurally_invalid_tool_schema.rs
@@ -6,8 +6,6 @@ use serde_json::Map;
 use serde_json::json;
 use tokio_util::sync::CancellationToken;
 
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
-use paddler_messaging::balancer_desired_state::BalancerDesiredState;
 use paddler_messaging::conversation_history::ConversationHistory;
 use paddler_messaging::conversation_message::ConversationMessage;
 use paddler_messaging::conversation_message_content::ConversationMessageContent;
@@ -22,23 +20,16 @@ use paddler_test_cluster_harness::agent_config::AgentConfig;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
 use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
 use paddler_tests::start_cluster::start_cluster;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn agent_rejects_structurally_invalid_tool_schema() {
-    let base_desired_state = qwen3_0_6b().into_desired_state();
-
     let cluster = start_cluster(ClusterParams {
         agents: vec![AgentConfig {
             name: "test-agent".to_owned(),
             slot_count: 1,
         }],
-        desired_state: Some(BalancerDesiredState {
-            inference_parameters: InferenceParameters {
-                temperature: 0.0,
-                ..base_desired_state.inference_parameters
-            },
-            ..base_desired_state
-        }),
+        desired_state: ClusterDesiredState::Apply(Box::new(qwen3_0_6b().into_desired_state())),
         wait_for_slots_ready: true,
         ..ClusterParams::default()
     })
@@ -53,12 +44,14 @@ async fn agent_rejects_structurally_invalid_tool_schema() {
             CancellationToken::new(),
             &ContinueFromConversationHistoryParams {
                 add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Text(
-                        "What is the weather in Paris?".to_owned(),
-                    ),
-                    role: "user".to_owned(),
-                }]),
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Text(
+                            "What is the weather in Paris?".to_owned(),
+                        ),
+                        role: "user".to_owned(),
+                    },
+                ]),
                 enable_thinking: false,
                 grammar: None,
                 max_tokens: NonZeroU32::new(64).unwrap(),
@@ -80,23 +73,11 @@ async fn agent_rejects_structurally_invalid_tool_schema() {
         .await
         .expect("the inference request must be accepted");
 
-    let schema_invalid_message = collected
-        .token_results
-        .iter()
-        .find_map(|event| match &event.token_result {
-            GeneratedTokenResult::ToolSchemaInvalid(message) => Some(message.clone()),
-            _ => None,
-        })
-        .unwrap_or_else(|| {
-            panic!(
-                "expected a ToolSchemaInvalid event when a tool's JSON Schema is invalid; got:\n{}",
-                collected.text
-            )
-        });
-
-    assert!(
-        schema_invalid_message.contains("get_weather"),
-        "the schema-invalid message should name the offending tool; got: {schema_invalid_message}"
+    assert_eq!(
+        collected.into_token_results(),
+        vec![GeneratedTokenResult::ToolSchemaInvalid(
+            "test-agent: tool \"get_weather\" parameters are not a valid JSON Schema: 123 is not valid under any of the schemas listed in the 'anyOf' keyword".to_owned()
+        )]
     );
 
     cluster
diff --git a/paddler_tests/tests/agent_reports_a_document_exceeding_its_batch_and_embeds_the_rest.rs b/paddler_tests/tests/agent_reports_a_document_exceeding_its_batch_and_embeds_the_rest.rs
new file mode 100644
index 000000000..b4d29f5ed
--- /dev/null
+++ b/paddler_tests/tests/agent_reports_a_document_exceeding_its_batch_and_embeds_the_rest.rs
@@ -0,0 +1,102 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::num::NonZeroU32;
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_inference_parameters::batch_size::BatchSize;
+use paddler_inference_parameters::model_runtime_parameters::ModelRuntimeParameters;
+use paddler_messaging::embedding_input_document::EmbeddingInputDocument;
+use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
+use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::embedding_cluster_params::EmbeddingClusterParams;
+use paddler_tests::oversized_text::oversized_text;
+use paddler_tests::start_embedding_cluster::start_embedding_cluster;
+
+const N_BATCH: usize = 64;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agent_reports_a_document_exceeding_its_batch_and_embeds_the_rest() {
+    let cluster = start_embedding_cluster(EmbeddingClusterParams {
+        agents: vec![AgentConfig::single(1)],
+        model_runtime_parameters: ModelRuntimeParameters {
+            n_batch: BatchSize::try_from(
+                u32::try_from(N_BATCH).expect("the value must fit its target type"),
+            )
+            .expect("the value must fit its target type"),
+            context_size: NonZeroU32::try_from(2048).expect("the value must fit its target type"),
+            ..ModelRuntimeParameters::default()
+        },
+        ..EmbeddingClusterParams::default()
+    })
+    .await
+    .expect("the cluster must start");
+
+    let huge_content = oversized_text();
+
+    let collected = cluster
+        .generate_embedding_batch(
+            CancellationToken::new(),
+            &GenerateEmbeddingBatchParams {
+                input_batch: vec![
+                    EmbeddingInputDocument {
+                        content: "ok".to_owned(),
+                        id: "tiny".to_owned(),
+                    },
+                    EmbeddingInputDocument {
+                        content: huge_content,
+                        id: "huge".to_owned(),
+                    },
+                ],
+                normalization_method: EmbeddingNormalizationMethod::None,
+            },
+        )
+        .await
+        .expect("the embedding batch must be accepted");
+
+    assert!(
+        collected.saw_done,
+        "stream must terminate with Done even when one document is oversized",
+    );
+    assert!(
+        collected.failures.is_empty(),
+        "no generic EmbeddingResult::Error events should be emitted; got {:?}",
+        collected.failures,
+    );
+
+    assert_eq!(
+        collected
+            .oversized_documents
+            .iter()
+            .map(|details| details.source_document_id.as_str())
+            .collect::<Vec<_>>(),
+        vec!["huge"],
+        "exactly one DocumentExceedsBatchSize event expected",
+    );
+
+    let oversized = &collected.oversized_documents[0];
+
+    assert_eq!(oversized.n_batch, N_BATCH);
+    assert!(
+        oversized.document_tokens > oversized.n_batch,
+        "document_tokens ({}) must exceed n_batch ({}) for the assertion to be meaningful",
+        oversized.document_tokens,
+        oversized.n_batch,
+    );
+
+    assert_eq!(
+        collected
+            .embeddings
+            .iter()
+            .map(|produced| produced.embedding.source_document_id.as_str())
+            .collect::<Vec<_>>(),
+        vec!["tiny"],
+        "the small document must still be embedded",
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/agent_reports_a_document_exceeding_its_sequence_context_and_embeds_the_rest.rs b/paddler_tests/tests/agent_reports_a_document_exceeding_its_sequence_context_and_embeds_the_rest.rs
new file mode 100644
index 000000000..43ebcfccd
--- /dev/null
+++ b/paddler_tests/tests/agent_reports_a_document_exceeding_its_sequence_context_and_embeds_the_rest.rs
@@ -0,0 +1,100 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::num::NonZeroU32;
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_inference_parameters::batch_size::BatchSize;
+use paddler_inference_parameters::embedding_parameters::EmbeddingParameters;
+use paddler_inference_parameters::model_runtime_parameters::ModelRuntimeParameters;
+use paddler_inference_parameters::pooling_type::PoolingType;
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_messaging::embedding_input_document::EmbeddingInputDocument;
+use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
+use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
+use paddler_tests::start_cluster::start_cluster;
+
+const CONTEXT_SIZE: u32 = 512;
+const SLOTS: u16 = 4;
+const LLAMA_CPP_SEQUENCE_CONTEXT_SIZE: usize = 256;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agent_reports_a_document_exceeding_its_sequence_context_and_embeds_the_rest() {
+    let context_size =
+        NonZeroU32::new(CONTEXT_SIZE).expect("the context size must be greater than zero");
+    let desired_state =
+        qwen3_0_6b().into_desired_state_with_embedding_parameters(EmbeddingParameters {
+            pooling_type: PoolingType::Last,
+            ..EmbeddingParameters::default()
+        });
+    let cluster = start_cluster(ClusterParams {
+        agents: AgentConfig::uniform(1, SLOTS),
+        desired_state: ClusterDesiredState::Apply(Box::new(BalancerDesiredState {
+            model_runtime_parameters: ModelRuntimeParameters {
+                context_size,
+                n_batch: BatchSize::default().at_most(context_size),
+                ..desired_state.model_runtime_parameters
+            },
+            ..desired_state
+        })),
+        wait_for_slots_ready: true,
+        ..ClusterParams::default()
+    })
+    .await
+    .expect("the cluster must start");
+
+    let collected = cluster
+        .generate_embedding_batch(
+            CancellationToken::new(),
+            &GenerateEmbeddingBatchParams {
+                input_batch: vec![
+                    EmbeddingInputDocument {
+                        content: "ok".to_owned(),
+                        id: "tiny".to_owned(),
+                    },
+                    EmbeddingInputDocument {
+                        content: "The quick brown fox jumps over the lazy dog. ".repeat(30),
+                        id: "longer-than-a-sequence".to_owned(),
+                    },
+                ],
+                normalization_method: EmbeddingNormalizationMethod::None,
+            },
+        )
+        .await
+        .expect("the embedding batch must be accepted");
+
+    assert_eq!(collected.failures, vec![]);
+    assert_eq!(
+        collected
+            .documents_exceeding_sequence_context
+            .iter()
+            .map(|details| details.source_document_id.as_str())
+            .collect::<Vec<_>>(),
+        vec!["longer-than-a-sequence"]
+    );
+
+    let exceeding_document = &collected.documents_exceeding_sequence_context[0];
+
+    assert_eq!(
+        exceeding_document.sequence_context_size,
+        LLAMA_CPP_SEQUENCE_CONTEXT_SIZE
+    );
+    assert!(exceeding_document.document_tokens > exceeding_document.sequence_context_size);
+    assert_eq!(
+        collected
+            .embeddings
+            .iter()
+            .map(|produced| produced.embedding.source_document_id.as_str())
+            .collect::<Vec<_>>(),
+        vec!["tiny"]
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/agent_reports_a_document_that_tokenizes_to_nothing_and_embeds_the_rest.rs b/paddler_tests/tests/agent_reports_a_document_that_tokenizes_to_nothing_and_embeds_the_rest.rs
new file mode 100644
index 000000000..a3fa328e0
--- /dev/null
+++ b/paddler_tests/tests/agent_reports_a_document_that_tokenizes_to_nothing_and_embeds_the_rest.rs
@@ -0,0 +1,69 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_inference_parameters::embedding_parameters::EmbeddingParameters;
+use paddler_inference_parameters::pooling_type::PoolingType;
+use paddler_messaging::embedding_document_tokenized_to_nothing_details::EmbeddingDocumentTokenizedToNothingDetails;
+use paddler_messaging::embedding_input_document::EmbeddingInputDocument;
+use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
+use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::embedding_cluster_params::EmbeddingClusterParams;
+use paddler_test_cluster_harness::model_card::smolvlm2_256m::smolvlm2_256m;
+use paddler_tests::start_embedding_cluster::start_embedding_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agent_reports_a_document_that_tokenizes_to_nothing_and_embeds_the_rest() {
+    let cluster = start_embedding_cluster(EmbeddingClusterParams {
+        agents: vec![AgentConfig::single(1)],
+        embedding_parameters: EmbeddingParameters {
+            pooling_type: PoolingType::Last,
+            ..EmbeddingParameters::default()
+        },
+        model_card: smolvlm2_256m(),
+        ..EmbeddingClusterParams::default()
+    })
+    .await
+    .expect("the cluster must start");
+
+    let collected = cluster
+        .generate_embedding_batch(
+            CancellationToken::new(),
+            &GenerateEmbeddingBatchParams {
+                input_batch: vec![
+                    EmbeddingInputDocument {
+                        content: String::new(),
+                        id: "empty".to_owned(),
+                    },
+                    EmbeddingInputDocument {
+                        content: "ok".to_owned(),
+                        id: "ok".to_owned(),
+                    },
+                ],
+                normalization_method: EmbeddingNormalizationMethod::None,
+            },
+        )
+        .await
+        .expect("the embedding batch must be accepted");
+
+    assert!(collected.saw_done);
+    assert_eq!(collected.failures, Vec::new());
+    assert_eq!(
+        collected.documents_tokenized_to_nothing,
+        vec![EmbeddingDocumentTokenizedToNothingDetails {
+            source_document_id: "empty".to_owned(),
+        }]
+    );
+    assert_eq!(
+        collected
+            .embeddings
+            .iter()
+            .map(|embedding_with_producer| embedding_with_producer
+                .embedding
+                .source_document_id
+                .as_str())
+            .collect::<Vec<_>>(),
+        vec!["ok"]
+    );
+}
diff --git a/paddler_tests/tests/agent_reports_a_grammar_no_token_can_satisfy_as_rejected_model_output.rs b/paddler_tests/tests/agent_reports_a_grammar_no_token_can_satisfy_as_rejected_model_output.rs
new file mode 100644
index 000000000..74f26da4a
--- /dev/null
+++ b/paddler_tests/tests/agent_reports_a_grammar_no_token_can_satisfy_as_rejected_model_output.rs
@@ -0,0 +1,48 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::num::NonZeroU32;
+
+use llama_cpp_bindings::error::SamplerApplyError;
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::generated_token_result::GeneratedTokenResult;
+use paddler_messaging::grammar_constraint::GrammarConstraint;
+use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
+
+const MAX_TOKENS: NonZeroU32 = NonZeroU32::new(10).unwrap();
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agent_reports_a_grammar_no_token_can_satisfy_as_rejected_model_output() {
+    let cluster = start_cluster_with_qwen3(vec![AgentConfig::single(1)])
+        .await
+        .expect("the cluster must start");
+
+    let collected = cluster
+        .continue_from_raw_prompt(
+            CancellationToken::new(),
+            &ContinueFromRawPromptParams {
+                grammar: Some(GrammarConstraint::Gbnf {
+                    grammar: r#"root ::= "\x00""#.to_owned(),
+                    root: "root".to_owned(),
+                }),
+                max_tokens: MAX_TOKENS,
+                raw_prompt: "Paddler".to_owned(),
+            },
+        )
+        .await
+        .expect("the inference request must be accepted");
+
+    assert_eq!(
+        collected.into_token_results(),
+        vec![GeneratedTokenResult::GrammarRejectedModelOutput(
+            SamplerApplyError::EveryCandidateExcluded.to_string()
+        )]
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/agent_reports_a_malformed_hugging_face_model_uri.rs b/paddler_tests/tests/agent_reports_a_malformed_hugging_face_model_uri.rs
new file mode 100644
index 000000000..9c0de1b6c
--- /dev/null
+++ b/paddler_tests/tests/agent_reports_a_malformed_hugging_face_model_uri.rs
@@ -0,0 +1,27 @@
+use paddler_messaging::agent_desired_model::AgentDesiredModel;
+use paddler_messaging::agent_issue::AgentIssue;
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_tests::start_single_agent_cluster_with_desired_state::start_single_agent_cluster_with_desired_state;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agent_reports_a_malformed_hugging_face_model_uri() {
+    let model_uri = "https://huggingface.co/owner/repo".to_owned();
+    let mut cluster = start_single_agent_cluster_with_desired_state(BalancerDesiredState {
+        model: AgentDesiredModel::Uri(model_uri.clone()),
+        ..BalancerDesiredState::default()
+    })
+    .await
+    .expect("a single-agent cluster must start");
+
+    cluster
+        .wait_for_first_agent_issue(|issue| {
+            matches!(issue, AgentIssue::HuggingFaceModelUriIsMalformed(model_path) if model_path.model_path == model_uri)
+        })
+        .await
+        .expect("the agent must report that the Hugging Face model URI does not point at a file");
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/agent_reports_a_pointer_head_it_cannot_load.rs b/paddler_tests/tests/agent_reports_a_pointer_head_it_cannot_load.rs
new file mode 100644
index 000000000..9bb9b2500
--- /dev/null
+++ b/paddler_tests/tests/agent_reports_a_pointer_head_it_cannot_load.rs
@@ -0,0 +1,37 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use paddler_messaging::agent_issue::AgentIssue;
+use paddler_test_cluster_harness::fixture_path::fixture_path;
+use paddler_tests::decision_cluster_params::DecisionClusterParams;
+use paddler_tests::decision_desired_state::DecisionDesiredState;
+use paddler_tests::start_decision_cluster::start_decision_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agent_reports_a_pointer_head_it_cannot_load() {
+    let mut cluster = start_decision_cluster(DecisionClusterParams {
+        desired_state: DecisionDesiredState {
+            pointer_head: fixture_path("invalid.gguf"),
+            ..DecisionDesiredState::default()
+        },
+        wait_for_slots_ready: false,
+        ..DecisionClusterParams::default()
+    })
+    .await
+    .expect("the cluster must start");
+
+    cluster
+        .wait_for_first_agent_issue(|issue| {
+            matches!(
+                issue,
+                AgentIssue::PointerHeadCannotBeLoaded(pointer_head_path)
+                    if pointer_head_path.model_path.ends_with("/fixtures/invalid.gguf")
+            )
+        })
+        .await
+        .expect("the agent must report the unreadable pointer head");
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/agent_reports_a_pointer_head_of_another_hidden_size.rs b/paddler_tests/tests/agent_reports_a_pointer_head_of_another_hidden_size.rs
new file mode 100644
index 000000000..83ca660b2
--- /dev/null
+++ b/paddler_tests/tests/agent_reports_a_pointer_head_of_another_hidden_size.rs
@@ -0,0 +1,43 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use paddler_messaging::agent_issue::AgentIssue;
+use paddler_messaging::agent_issue_params::model_path::ModelPath;
+use paddler_messaging::agent_issue_params::pointer_head_incompatibility::PointerHeadIncompatibility;
+use paddler_messaging::agent_issue_params::pointer_head_incompatible_with_model_params::PointerHeadIncompatibleWithModelParams;
+use paddler_test_cluster_harness::fixture_path::fixture_path;
+use paddler_tests::decision_cluster_params::DecisionClusterParams;
+use paddler_tests::decision_desired_state::DecisionDesiredState;
+use paddler_tests::start_decision_cluster::start_decision_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agent_reports_a_pointer_head_of_another_hidden_size() {
+    let pointer_head_fixture = "qwen3_5_0_8b_synthetic_pointer_head_with_wrong_hidden_size.gguf";
+    let hidden_size_mismatch =
+        AgentIssue::PointerHeadIncompatibleWithModel(PointerHeadIncompatibleWithModelParams {
+            incompatibility: PointerHeadIncompatibility::HiddenSizeMismatch {
+                model_hidden_size: 1024,
+                pointer_head_hidden_size: 512,
+            },
+            pointer_head_path: ModelPath::from(fixture_path(pointer_head_fixture).as_path()),
+        });
+    let mut cluster = start_decision_cluster(DecisionClusterParams {
+        desired_state: DecisionDesiredState {
+            pointer_head: fixture_path(pointer_head_fixture),
+            ..DecisionDesiredState::default()
+        },
+        wait_for_slots_ready: false,
+        ..DecisionClusterParams::default()
+    })
+    .await
+    .expect("the cluster must start");
+
+    cluster
+        .wait_for_first_agent_issue(|issue| *issue == hidden_size_mismatch)
+        .await
+        .expect("the agent must report the hidden size mismatch");
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/agent_reports_a_pointer_head_wider_than_the_hidden_states_the_model_outputs.rs b/paddler_tests/tests/agent_reports_a_pointer_head_wider_than_the_hidden_states_the_model_outputs.rs
new file mode 100644
index 000000000..994243d97
--- /dev/null
+++ b/paddler_tests/tests/agent_reports_a_pointer_head_wider_than_the_hidden_states_the_model_outputs.rs
@@ -0,0 +1,47 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use paddler_messaging::agent_issue::AgentIssue;
+use paddler_messaging::agent_issue_params::model_path::ModelPath;
+use paddler_messaging::agent_issue_params::pointer_head_incompatibility::PointerHeadIncompatibility;
+use paddler_messaging::agent_issue_params::pointer_head_incompatible_with_model_params::PointerHeadIncompatibleWithModelParams;
+use paddler_test_cluster_harness::fixture_path::fixture_path;
+use paddler_test_cluster_harness::model_card::lfm2_colbert_350m::lfm2_colbert_350m;
+use paddler_test_cluster_harness::synthetic_pointer_head_fixture::QWEN3_5_0_8B_SYNTHETIC_POINTER_HEAD_FIXTURE;
+use paddler_tests::decision_backbone::DecisionBackbone;
+use paddler_tests::decision_cluster_params::DecisionClusterParams;
+use paddler_tests::decision_desired_state::DecisionDesiredState;
+use paddler_tests::start_decision_cluster::start_decision_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agent_reports_a_pointer_head_wider_than_the_hidden_states_the_model_outputs() {
+    let hidden_size_mismatch =
+        AgentIssue::PointerHeadIncompatibleWithModel(PointerHeadIncompatibleWithModelParams {
+            incompatibility: PointerHeadIncompatibility::HiddenSizeMismatch {
+                model_hidden_size: 128,
+                pointer_head_hidden_size: 1024,
+            },
+            pointer_head_path: ModelPath::from(
+                fixture_path(QWEN3_5_0_8B_SYNTHETIC_POINTER_HEAD_FIXTURE).as_path(),
+            ),
+        });
+    let mut cluster = start_decision_cluster(DecisionClusterParams {
+        desired_state: DecisionDesiredState {
+            backbone: DecisionBackbone::HuggingFace(lfm2_colbert_350m()),
+            ..DecisionDesiredState::default()
+        },
+        wait_for_slots_ready: false,
+        ..DecisionClusterParams::default()
+    })
+    .await
+    .expect("the cluster must start");
+
+    cluster
+        .wait_for_first_agent_issue(|issue| *issue == hidden_size_mismatch)
+        .await
+        .expect("the agent must report the pointer head as wider than the hidden states it reads");
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/agent_reports_an_agent_uri_that_is_not_canonical.rs b/paddler_tests/tests/agent_reports_an_agent_uri_that_is_not_canonical.rs
new file mode 100644
index 000000000..18aa3ccbb
--- /dev/null
+++ b/paddler_tests/tests/agent_reports_an_agent_uri_that_is_not_canonical.rs
@@ -0,0 +1,33 @@
+use paddler_messaging::agent_desired_model::AgentDesiredModel;
+use paddler_messaging::agent_issue::AgentIssue;
+use paddler_messaging::agent_issue_params::agent_model_uri_not_canonical_params::AgentModelUriNotCanonicalParams;
+use paddler_messaging::agent_issue_params::model_path::ModelPath;
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_tests::start_single_agent_cluster_with_desired_state::start_single_agent_cluster_with_desired_state;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agent_reports_an_agent_uri_that_is_not_canonical() {
+    let model_uri = "agent:///models/my model.gguf".to_owned();
+    let expected_issue = AgentIssue::AgentModelUriIsNotCanonical(AgentModelUriNotCanonicalParams {
+        canonical_uri: "agent:///models/my%20model.gguf".to_owned(),
+        model_path: ModelPath {
+            model_path: model_uri.clone(),
+        },
+    });
+    let mut cluster = start_single_agent_cluster_with_desired_state(BalancerDesiredState {
+        model: AgentDesiredModel::Uri(model_uri),
+        ..BalancerDesiredState::default()
+    })
+    .await
+    .expect("a single-agent cluster must start");
+
+    cluster
+        .wait_for_first_agent_issue(|issue| *issue == expected_issue)
+        .await
+        .expect("the agent must report the canonical form of the agent URI");
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/agent_reports_an_unparseable_model_uri.rs b/paddler_tests/tests/agent_reports_an_unparseable_model_uri.rs
new file mode 100644
index 000000000..b4cb965b3
--- /dev/null
+++ b/paddler_tests/tests/agent_reports_an_unparseable_model_uri.rs
@@ -0,0 +1,27 @@
+use paddler_messaging::agent_desired_model::AgentDesiredModel;
+use paddler_messaging::agent_issue::AgentIssue;
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_tests::start_single_agent_cluster_with_desired_state::start_single_agent_cluster_with_desired_state;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agent_reports_an_unparseable_model_uri() {
+    let model_uri = "not a valid uri".to_owned();
+    let mut cluster = start_single_agent_cluster_with_desired_state(BalancerDesiredState {
+        model: AgentDesiredModel::Uri(model_uri.clone()),
+        ..BalancerDesiredState::default()
+    })
+    .await
+    .expect("a single-agent cluster must start");
+
+    cluster
+        .wait_for_first_agent_issue(|issue| {
+            matches!(issue, AgentIssue::ModelUriIsUnparseable(model_path) if model_path.model_path == model_uri)
+        })
+        .await
+        .expect("the agent must report that the model URI cannot be parsed");
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/agent_reports_chat_template_error_when_rendering_fails.rs b/paddler_tests/tests/agent_reports_chat_template_error_when_rendering_fails.rs
index 271772735..9099d1f96 100644
--- a/paddler_tests/tests/agent_reports_chat_template_error_when_rendering_fails.rs
+++ b/paddler_tests/tests/agent_reports_chat_template_error_when_rendering_fails.rs
@@ -4,7 +4,6 @@ use std::num::NonZeroU32;
 
 use tokio_util::sync::CancellationToken;
 
-use paddler_messaging::balancer_desired_state::BalancerDesiredState;
 use paddler_messaging::chat_template::ChatTemplate;
 use paddler_messaging::conversation_history::ConversationHistory;
 use paddler_messaging::conversation_message::ConversationMessage;
@@ -12,7 +11,9 @@ use paddler_messaging::conversation_message_content::ConversationMessageContent;
 use paddler_messaging::generated_token_result::GeneratedTokenResult;
 use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_tests::desired_state_with_chat_template_override::desired_state_with_chat_template_override;
 use paddler_tests::qwen3_desired_state::qwen3_desired_state;
 use paddler_tests::start_cluster::start_cluster;
 
@@ -20,13 +21,14 @@ use paddler_tests::start_cluster::start_cluster;
 async fn agent_reports_chat_template_error_when_rendering_fails() {
     let cluster = start_cluster(ClusterParams {
         agents: AgentConfig::uniform(1, 1),
-        desired_state: Some(BalancerDesiredState {
-            chat_template_override: Some(ChatTemplate {
-                content: "{{ raise_exception('conversations are not supported') }}".to_owned(),
-            }),
-            use_chat_template_override: true,
-            ..qwen3_desired_state()
-        }),
+        desired_state: ClusterDesiredState::Apply(Box::new(
+            desired_state_with_chat_template_override(
+                qwen3_desired_state(),
+                ChatTemplate {
+                    content: "{{ raise_exception('conversations are not supported') }}".to_owned(),
+                },
+            ),
+        )),
         wait_for_slots_ready: true,
         ..ClusterParams::default()
     })
@@ -38,10 +40,12 @@ async fn agent_reports_chat_template_error_when_rendering_fails() {
             CancellationToken::new(),
             &ContinueFromConversationHistoryParams {
                 add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Text("Hello".to_owned()),
-                    role: "user".to_owned(),
-                }]),
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Text("Hello".to_owned()),
+                        role: "user".to_owned(),
+                    },
+                ]),
                 enable_thinking: false,
                 grammar: None,
                 max_tokens: NonZeroU32::new(4).unwrap(),
diff --git a/paddler_tests/tests/agent_reports_completion_tokens_matching_streamed_token_count.rs b/paddler_tests/tests/agent_reports_completion_tokens_matching_streamed_token_count.rs
index d0a1c8e2d..1b859799e 100644
--- a/paddler_tests/tests/agent_reports_completion_tokens_matching_streamed_token_count.rs
+++ b/paddler_tests/tests/agent_reports_completion_tokens_matching_streamed_token_count.rs
@@ -7,7 +7,6 @@ use tokio_util::sync::CancellationToken;
 use paddler_messaging::conversation_history::ConversationHistory;
 use paddler_messaging::conversation_message::ConversationMessage;
 use paddler_messaging::conversation_message_content::ConversationMessageContent;
-use paddler_messaging::generated_token_result::GeneratedTokenResult;
 use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
 use paddler_test_cluster_harness::unending_grammar::unending_grammar;
@@ -26,10 +25,14 @@ async fn agent_reports_completion_tokens_matching_streamed_token_count() {
             CancellationToken::new(),
             &ContinueFromConversationHistoryParams {
                 add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Text("Tell me a long story.".to_owned()),
-                    role: "user".to_owned(),
-                }]),
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Text(
+                            "Tell me a long story.".to_owned(),
+                        ),
+                        role: "user".to_owned(),
+                    },
+                ]),
                 enable_thinking: false,
                 grammar: Some(unending_grammar()),
                 max_tokens: MAX_TOKENS,
@@ -46,13 +49,9 @@ async fn agent_reports_completion_tokens_matching_streamed_token_count() {
         .filter(|result| result.token_result.is_token())
         .count() as u64;
 
-    let last = collected
-        .token_results
-        .last()
-        .expect("no token results received");
-    let GeneratedTokenResult::Done(summary) = &last.token_result else {
-        panic!("last result was not Done: {last:?}");
-    };
+    let summary = collected
+        .summary()
+        .expect("the generation must finish with a summary");
 
     assert_eq!(streamed_token_count, u64::from(MAX_TOKENS.get()));
     assert_eq!(
diff --git a/paddler_tests/tests/agent_reports_decode_failure_for_a_prompt_exceeding_a_non_causal_micro_batch.rs b/paddler_tests/tests/agent_reports_decode_failure_for_a_prompt_exceeding_a_non_causal_micro_batch.rs
index f15d629da..0fcdda38d 100644
--- a/paddler_tests/tests/agent_reports_decode_failure_for_a_prompt_exceeding_a_non_causal_micro_batch.rs
+++ b/paddler_tests/tests/agent_reports_decode_failure_for_a_prompt_exceeding_a_non_causal_micro_batch.rs
@@ -8,6 +8,7 @@ use paddler_messaging::chat_template::ChatTemplate;
 use paddler_messaging::generated_token_result::GeneratedTokenResult;
 use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
 use paddler_tests::nomic_embed_desired_state_with_chat_template_override::nomic_embed_desired_state_with_chat_template_override;
 use paddler_tests::start_cluster::start_cluster;
@@ -20,10 +21,10 @@ const LLAMA_CPP_DEFAULT_MICRO_BATCH_TOKENS: usize = 512;
 async fn agent_reports_decode_failure_for_a_prompt_exceeding_a_non_causal_micro_batch() {
     let cluster = start_cluster(ClusterParams {
         agents: vec![AgentConfig::single(1)],
-        desired_state: Some(nomic_embed_desired_state_with_chat_template_override(
-            ChatTemplate {
+        desired_state: ClusterDesiredState::Apply(Box::new(
+            nomic_embed_desired_state_with_chat_template_override(ChatTemplate {
                 content: "{{ messages[0].content }}".to_owned(),
-            },
+            }),
         )),
         wait_for_slots_ready: true,
         ..ClusterParams::default()
diff --git a/paddler_tests/tests/agent_reports_grammar_syntax_error_for_malformed_gbnf.rs b/paddler_tests/tests/agent_reports_grammar_syntax_error_for_malformed_gbnf.rs
index 819242ae9..7d624441a 100644
--- a/paddler_tests/tests/agent_reports_grammar_syntax_error_for_malformed_gbnf.rs
+++ b/paddler_tests/tests/agent_reports_grammar_syntax_error_for_malformed_gbnf.rs
@@ -2,13 +2,18 @@
 
 use std::num::NonZeroU32;
 
+use llama_cpp_bindings::GbnfValidationError;
+use llama_cpp_bindings::error::GrammarError;
 use tokio_util::sync::CancellationToken;
 
+use paddler_agent_runtime::rejection_description::rejection_description;
+use paddler_agent_text_generation::text_generation_error::TextGenerationError;
 use paddler_messaging::generated_token_result::GeneratedTokenResult;
 use paddler_messaging::grammar_constraint::GrammarConstraint;
 use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
 use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
+use paddler_tests::weather_question_conversation::weather_question_conversation;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn agent_reports_grammar_syntax_error_for_malformed_gbnf() {
@@ -16,25 +21,43 @@ async fn agent_reports_grammar_syntax_error_for_malformed_gbnf() {
         .await
         .expect("the cluster must start");
 
-    let collected = cluster
+    let malformed_grammar = GrammarConstraint::Gbnf {
+        grammar: r#"root ::= "unterminated"#.to_owned(),
+        root: "root".to_owned(),
+    };
+    let raw_prompt_results = cluster
         .continue_from_raw_prompt(CancellationToken::new(), &ContinueFromRawPromptParams {
-            grammar: Some(GrammarConstraint::Gbnf {
-                grammar: r#"root ::= "unterminated"#.to_owned(),
-                root: "root".to_owned(),
-            }),
+            grammar: Some(malformed_grammar.clone()),
             max_tokens: NonZeroU32::new(10).unwrap(),
             raw_prompt:
                 "<|im_start|>user\nSay hi.<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n"
                     .to_owned(),
         })
-        .await.expect("the inference request must be accepted");
-
-    let token_results = collected.into_token_results();
+        .await
+        .expect("the raw prompt request must be accepted")
+        .into_token_results();
+    let conversation_results = cluster
+        .continue_from_conversation_history(
+            CancellationToken::new(),
+            &weather_question_conversation(malformed_grammar, NonZeroU32::new(10).unwrap(), vec![]),
+        )
+        .await
+        .expect("the conversation request must be accepted")
+        .into_token_results();
 
-    assert!(matches!(
-        token_results.as_slice(),
-        [GeneratedTokenResult::GrammarSyntaxError(_)]
-    ));
+    for token_results in [raw_prompt_results, conversation_results] {
+        assert_eq!(
+            token_results,
+            vec![GeneratedTokenResult::GrammarSyntaxError(
+                rejection_description(
+                    Some("test-agent-0"),
+                    &TextGenerationError::GrammarInvalid(GrammarError::GrammarRejected(
+                        GbnfValidationError::SyntaxError
+                    ))
+                )
+            )]
+        );
+    }
 
     cluster
         .shutdown()
diff --git a/paddler_tests/tests/agent_reports_independent_usage_for_concurrent_requests.rs b/paddler_tests/tests/agent_reports_independent_usage_for_concurrent_requests.rs
index 16d65c45e..ec083c5c0 100644
--- a/paddler_tests/tests/agent_reports_independent_usage_for_concurrent_requests.rs
+++ b/paddler_tests/tests/agent_reports_independent_usage_for_concurrent_requests.rs
@@ -8,7 +8,6 @@ use tokio_util::sync::CancellationToken;
 use paddler_messaging::conversation_history::ConversationHistory;
 use paddler_messaging::conversation_message::ConversationMessage;
 use paddler_messaging::conversation_message_content::ConversationMessageContent;
-use paddler_messaging::generated_token_result::GeneratedTokenResult;
 use paddler_messaging::generation_summary::GenerationSummary;
 use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
@@ -28,10 +27,12 @@ async fn agent_reports_independent_usage_for_concurrent_requests() {
             CancellationToken::new(),
             &ContinueFromConversationHistoryParams {
                 add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Text("Count.".to_owned()),
-                    role: "user".to_owned(),
-                }]),
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Text("Count.".to_owned()),
+                        role: "user".to_owned(),
+                    },
+                ]),
                 enable_thinking: false,
                 grammar: Some(gbnf_literal(forced_reply)),
                 max_tokens: NonZeroU32::new(30).unwrap(),
@@ -43,14 +44,9 @@ async fn agent_reports_independent_usage_for_concurrent_requests() {
         async move {
             let collected = generation.await.expect("the generation must complete");
 
-            let last = collected
-                .token_results
-                .last()
-                .expect("no token results received");
-            match &last.token_result {
-                GeneratedTokenResult::Done(summary) => *summary,
-                other => panic!("last result was not Done: {other:?}"),
-            }
+            collected
+                .summary()
+                .expect("the generation must finish with a summary")
         }
     });
 
diff --git a/paddler_tests/tests/agent_reports_only_content_tokens_in_usage_for_plain_content.rs b/paddler_tests/tests/agent_reports_only_content_tokens_in_usage_for_plain_content.rs
index 7e61d2fc4..68941ce7e 100644
--- a/paddler_tests/tests/agent_reports_only_content_tokens_in_usage_for_plain_content.rs
+++ b/paddler_tests/tests/agent_reports_only_content_tokens_in_usage_for_plain_content.rs
@@ -7,7 +7,6 @@ use tokio_util::sync::CancellationToken;
 use paddler_messaging::conversation_history::ConversationHistory;
 use paddler_messaging::conversation_message::ConversationMessage;
 use paddler_messaging::conversation_message_content::ConversationMessageContent;
-use paddler_messaging::generated_token_result::GeneratedTokenResult;
 use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
 use paddler_tests::gbnf_literal::gbnf_literal;
@@ -24,10 +23,12 @@ async fn agent_reports_only_content_tokens_in_usage_for_plain_content() {
             CancellationToken::new(),
             &ContinueFromConversationHistoryParams {
                 add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Text("Say hello.".to_owned()),
-                    role: "user".to_owned(),
-                }]),
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Text("Say hello.".to_owned()),
+                        role: "user".to_owned(),
+                    },
+                ]),
                 enable_thinking: false,
                 grammar: Some(gbnf_literal("Hello there, nice to meet you.")),
                 max_tokens: NonZeroU32::new(60).unwrap(),
@@ -38,13 +39,9 @@ async fn agent_reports_only_content_tokens_in_usage_for_plain_content() {
         .await
         .expect("the inference request must be accepted");
 
-    let last = collected
-        .token_results
-        .last()
-        .expect("no token results received");
-    let GeneratedTokenResult::Done(summary) = &last.token_result else {
-        panic!("last result was not Done: {last:?}");
-    };
+    let summary = collected
+        .summary()
+        .expect("the generation must finish with a summary");
 
     assert!(summary.usage.prompt_tokens > 0);
     assert!(summary.usage.content_tokens > 0);
diff --git a/paddler_tests/tests/agent_reports_oversized_document_and_embeds_the_rest.rs b/paddler_tests/tests/agent_reports_oversized_document_and_embeds_the_rest.rs
deleted file mode 100644
index fc85a7b34..000000000
--- a/paddler_tests/tests/agent_reports_oversized_document_and_embeds_the_rest.rs
+++ /dev/null
@@ -1,106 +0,0 @@
-#![cfg(feature = "tests_that_use_llms")]
-
-use std::num::NonZeroU32;
-
-use tokio_util::sync::CancellationToken;
-
-use paddler_inference_parameters::batch_size::BatchSize;
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
-use paddler_messaging::embedding_input_document::EmbeddingInputDocument;
-use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
-use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
-use paddler_test_cluster_harness::agent_config::AgentConfig;
-use paddler_test_cluster_harness::embedding_cluster_params::EmbeddingClusterParams;
-use paddler_tests::start_embedding_cluster::start_embedding_cluster;
-
-const N_BATCH: usize = 64;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn agent_reports_oversized_document_and_embeds_the_rest() {
-    let cluster = start_embedding_cluster(EmbeddingClusterParams {
-        agents: vec![AgentConfig::single(1)],
-        inference_parameters: InferenceParameters {
-            n_batch: BatchSize::try_from(
-                u32::try_from(N_BATCH).expect("the value must fit its target type"),
-            )
-            .expect("the value must fit its target type"),
-            context_size: NonZeroU32::try_from(2048).expect("the value must fit its target type"),
-            enable_embeddings: true,
-            ..InferenceParameters::deterministic()
-        },
-        ..EmbeddingClusterParams::default()
-    })
-    .await
-    .expect("the cluster must start");
-
-    let huge_content = "The quick brown fox jumps over the lazy dog. ".repeat(40);
-
-    let collected = cluster
-        .generate_embedding_batch(
-            CancellationToken::new(),
-            &GenerateEmbeddingBatchParams {
-                input_batch: vec![
-                    EmbeddingInputDocument {
-                        content: "ok".to_owned(),
-                        id: "tiny".to_owned(),
-                    },
-                    EmbeddingInputDocument {
-                        content: huge_content,
-                        id: "huge".to_owned(),
-                    },
-                ],
-                normalization_method: EmbeddingNormalizationMethod::None,
-            },
-        )
-        .await
-        .expect("the embedding batch must be accepted");
-
-    assert!(
-        collected.saw_done,
-        "stream must terminate with Done even when one document is oversized",
-    );
-    assert!(
-        collected.errors.is_empty(),
-        "no generic EmbeddingResult::Error events should be emitted; got {:?}",
-        collected.errors,
-    );
-
-    assert_eq!(
-        collected.oversized_documents.len(),
-        1,
-        "exactly one DocumentExceedsBatchSize event expected; got {:?}",
-        collected
-            .oversized_documents
-            .iter()
-            .map(|details| &details.source_document_id)
-            .collect::<Vec<_>>(),
-    );
-
-    let oversized = &collected.oversized_documents[0];
-
-    assert_eq!(oversized.source_document_id, "huge");
-    assert_eq!(oversized.n_batch, N_BATCH);
-    assert!(
-        oversized.document_tokens > oversized.n_batch,
-        "document_tokens ({}) must exceed n_batch ({}) for the assertion to be meaningful",
-        oversized.document_tokens,
-        oversized.n_batch,
-    );
-
-    assert_eq!(
-        collected.embeddings.len(),
-        1,
-        "the small document must still be embedded; got {:?}",
-        collected
-            .embeddings
-            .iter()
-            .map(|produced| &produced.embedding.source_document_id)
-            .collect::<Vec<_>>(),
-    );
-    assert_eq!(collected.embeddings[0].embedding.source_document_id, "tiny",);
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_tests/tests/agent_reports_pointer_head_delimiters_that_are_not_control_tokens.rs b/paddler_tests/tests/agent_reports_pointer_head_delimiters_that_are_not_control_tokens.rs
new file mode 100644
index 000000000..eaa9c30bc
--- /dev/null
+++ b/paddler_tests/tests/agent_reports_pointer_head_delimiters_that_are_not_control_tokens.rs
@@ -0,0 +1,43 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use paddler_messaging::agent_issue::AgentIssue;
+use paddler_messaging::agent_issue_params::pointer_head_incompatibility::PointerHeadIncompatibility;
+use paddler_messaging::agent_issue_params::pointer_head_incompatible_with_model_params::PointerHeadIncompatibleWithModelParams;
+use paddler_test_cluster_harness::fixture_path::fixture_path;
+use paddler_tests::decision_cluster_params::DecisionClusterParams;
+use paddler_tests::decision_desired_state::DecisionDesiredState;
+use paddler_tests::start_decision_cluster::start_decision_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agent_reports_pointer_head_delimiters_that_are_not_control_tokens() {
+    let mut cluster = start_decision_cluster(DecisionClusterParams {
+        desired_state: DecisionDesiredState {
+            pointer_head: fixture_path("qwen3_5_0_8b_pointer_head_with_plain_text_delimiters.gguf"),
+            ..DecisionDesiredState::default()
+        },
+        wait_for_slots_ready: false,
+        ..DecisionClusterParams::default()
+    })
+    .await
+    .expect("the cluster must start");
+
+    cluster
+        .wait_for_first_agent_issue(|issue| {
+            matches!(
+                issue,
+                AgentIssue::PointerHeadIncompatibleWithModel(PointerHeadIncompatibleWithModelParams {
+                    incompatibility: PointerHeadIncompatibility::DelimiterIsNotAControlToken {
+                        delimiter,
+                    },
+                    ..
+                }) if delimiter == "decide"
+            )
+        })
+        .await
+        .expect("the agent must report the plain text delimiter");
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/agent_reports_slot_cannot_start_for_a_batch_smaller_than_its_slots.rs b/paddler_tests/tests/agent_reports_slot_cannot_start_for_a_batch_smaller_than_its_slots.rs
new file mode 100644
index 000000000..6afdbe7c4
--- /dev/null
+++ b/paddler_tests/tests/agent_reports_slot_cannot_start_for_a_batch_smaller_than_its_slots.rs
@@ -0,0 +1,64 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use paddler_inference_parameters::all_gpu_layers::ALL_GPU_LAYERS;
+use paddler_inference_parameters::batch_size::BatchSize;
+use paddler_inference_parameters::embedding_parameters::EmbeddingParameters;
+use paddler_inference_parameters::model_runtime_parameters::ModelRuntimeParameters;
+use paddler_messaging::agent_issue::AgentIssue;
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_test_cluster_harness::model_card::nomic_embed_text_v1_5::nomic_embed_text_v1_5;
+use paddler_test_cluster_harness::model_card::qwen3_5_0_8b::qwen3_5_0_8b;
+use paddler_test_cluster_harness::pointer_head_fixture::pointer_head_fixture;
+use paddler_test_cluster_harness::synthetic_pointer_head_fixture::QWEN3_5_0_8B_SYNTHETIC_POINTER_HEAD_FIXTURE;
+use paddler_tests::start_cluster::start_cluster;
+
+const SLOTS: u16 = 2;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agent_reports_slot_cannot_start_for_a_batch_smaller_than_its_slots() {
+    for desired_state in [
+        qwen3_5_0_8b().into_decision_desired_state(
+            pointer_head_fixture(QWEN3_5_0_8B_SYNTHETIC_POINTER_HEAD_FIXTURE)
+                .expect("the pointer head fixture path must be absolute"),
+        ),
+        nomic_embed_text_v1_5()
+            .into_desired_state_with_embedding_parameters(EmbeddingParameters::default()),
+    ] {
+        let mut cluster = start_cluster(ClusterParams {
+            agents: vec![AgentConfig::single(SLOTS)],
+            desired_state: ClusterDesiredState::Apply(Box::new(BalancerDesiredState {
+                model_runtime_parameters: ModelRuntimeParameters {
+                    n_batch: BatchSize::try_from(u32::from(SLOTS) - 1)
+                        .expect("a batch of one token is valid"),
+                    n_gpu_layers: ALL_GPU_LAYERS,
+                    ..ModelRuntimeParameters::default()
+                },
+                ..desired_state
+            })),
+            wait_for_slots_ready: false,
+            ..ClusterParams::default()
+        })
+        .await
+        .expect("the cluster must start");
+
+        cluster
+            .wait_for_first_agent_issue(|issue| {
+                matches!(
+                    issue,
+                    AgentIssue::SlotCannotStart(slot_cannot_start)
+                        if slot_cannot_start.slot_index == u32::from(SLOTS) - 1
+                            && !slot_cannot_start.error.is_empty()
+                )
+            })
+            .await
+            .expect("the agent must report that its slots cannot start");
+
+        cluster
+            .shutdown()
+            .await
+            .expect("the cluster must shut down cleanly");
+    }
+}
diff --git a/paddler_tests/tests/agent_reports_slot_cannot_start_for_excessive_slots.rs b/paddler_tests/tests/agent_reports_slot_cannot_start_for_excessive_slots.rs
index 2766f74c9..ae18fd830 100644
--- a/paddler_tests/tests/agent_reports_slot_cannot_start_for_excessive_slots.rs
+++ b/paddler_tests/tests/agent_reports_slot_cannot_start_for_excessive_slots.rs
@@ -1,11 +1,19 @@
 #![cfg(feature = "tests_that_use_llms")]
 
+use std::mem::discriminant;
+
+use llama_cpp_bindings::LlamaContextLoadError;
+
+use paddler_agent_runtime::agent_runtime_error::AgentRuntimeError;
 use paddler_inference_parameters::all_gpu_layers::ALL_GPU_LAYERS;
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
-use paddler_messaging::agent_desired_model::AgentDesiredModel;
+use paddler_inference_parameters::model_runtime_parameters::ModelRuntimeParameters;
 use paddler_messaging::agent_issue::AgentIssue;
+use paddler_messaging::agent_issue_params::slot_cannot_start_params::SlotCannotStartParams;
 use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_messaging::inference_capability::InferenceCapability;
+use paddler_model_source::model_source::ModelSource;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
 use paddler_test_cluster_harness::model_card::ModelCard;
 use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
@@ -19,26 +27,33 @@ async fn agent_reports_slot_cannot_start_for_excessive_slots() {
             name: "test-agent".to_owned(),
             slot_count: 257,
         }],
-        desired_state: Some(BalancerDesiredState {
-            inference_parameters: InferenceParameters {
+        desired_state: ClusterDesiredState::Apply(Box::new(BalancerDesiredState {
+            model_runtime_parameters: ModelRuntimeParameters {
                 n_gpu_layers: ALL_GPU_LAYERS,
-                ..InferenceParameters::deterministic()
+                ..ModelRuntimeParameters::default()
             },
-            model: AgentDesiredModel::HuggingFace(reference),
+            model: ModelSource::HuggingFace(reference).into_agent_desired_model(),
             ..BalancerDesiredState::default()
-        }),
+        })),
         wait_for_slots_ready: false,
         ..ClusterParams::default()
     })
     .await
     .expect("a single-agent cluster must start");
 
-    cluster
-        .wait_for_first_agent_issue(|issue| {
-            matches!(issue, AgentIssue::SlotCannotStart(slot_cannot_start) if !slot_cannot_start.error.is_empty())
-        })
+    let expected_issue = AgentIssue::SlotCannotStart(SlotCannotStartParams {
+        error: AgentRuntimeError::ContextCreationFailed(LlamaContextLoadError::Unconstructible)
+            .to_string(),
+        inference_capability: InferenceCapability::Embeddings,
+        slot_index: 0,
+    });
+    let snapshot = cluster
+        .wait_for_first_agent_issue(|issue| discriminant(issue) == discriminant(&expected_issue))
         .await
         .expect("the agent must report SlotCannotStart with the underlying error");
+    let issues = &snapshot.agents[0].status.issues;
+
+    assert!(issues.contains(&expected_issue), "{issues:?}");
 
     cluster
         .shutdown()
diff --git a/paddler_tests/tests/agent_reports_too_few_slots_for_decisions.rs b/paddler_tests/tests/agent_reports_too_few_slots_for_decisions.rs
new file mode 100644
index 000000000..c60c37dc2
--- /dev/null
+++ b/paddler_tests/tests/agent_reports_too_few_slots_for_decisions.rs
@@ -0,0 +1,35 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use paddler_messaging::agent_issue::AgentIssue;
+use paddler_messaging::agent_issue_params::slots_insufficient_for_decisions_params::SlotsInsufficientForDecisionsParams;
+use paddler_messaging::decision_slots_minimum::DECISION_SLOTS_MINIMUM;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_tests::decision_cluster_params::DecisionClusterParams;
+use paddler_tests::start_decision_cluster::start_decision_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agent_reports_too_few_slots_for_decisions() {
+    let mut cluster = start_decision_cluster(DecisionClusterParams {
+        agents: vec![AgentConfig::single(1)],
+        wait_for_slots_ready: false,
+        ..DecisionClusterParams::default()
+    })
+    .await
+    .expect("the cluster must start");
+
+    cluster
+        .wait_for_first_agent_issue(|issue| {
+            *issue
+                == AgentIssue::SlotsInsufficientForDecisions(SlotsInsufficientForDecisionsParams {
+                    desired_slots: 1,
+                    required_slots: DECISION_SLOTS_MINIMUM.get(),
+                })
+        })
+        .await
+        .expect("the agent must report too few slots");
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/agent_reports_tool_call_validation_failure.rs b/paddler_tests/tests/agent_reports_tool_call_validation_failure.rs
index ee8425a74..4bc16c4bb 100644
--- a/paddler_tests/tests/agent_reports_tool_call_validation_failure.rs
+++ b/paddler_tests/tests/agent_reports_tool_call_validation_failure.rs
@@ -2,13 +2,8 @@
 
 use std::num::NonZeroU32;
 
-use serde_json::Map;
-use serde_json::Value;
-use serde_json::json;
 use tokio_util::sync::CancellationToken;
 
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
-use paddler_messaging::balancer_desired_state::BalancerDesiredState;
 use paddler_messaging::conversation_history::ConversationHistory;
 use paddler_messaging::conversation_message::ConversationMessage;
 use paddler_messaging::conversation_message_content::ConversationMessageContent;
@@ -18,52 +13,39 @@ use paddler_messaging::request_params::continue_from_conversation_history_params
 use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::FunctionCall;
 use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::function::Function;
 use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters::Parameters;
-use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters_schema::validated_parameters_schema::ValidatedParametersSchema;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
 use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
 use paddler_tests::start_cluster::start_cluster;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
+use paddler_tests::weather_tool_parameters_with_an_unsatisfiable_location::weather_tool_parameters_with_an_unsatisfiable_location;
+use paddler_tests::weather_tool_question::WEATHER_TOOL_QUESTION;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn agent_reports_tool_call_validation_failure() {
-    let base_desired_state = qwen3_0_6b().into_desired_state();
-
     let cluster = start_cluster(ClusterParams {
         agents: vec![AgentConfig {
             name: "test-agent".to_owned(),
             slot_count: 1,
         }],
-        desired_state: Some(BalancerDesiredState {
-            inference_parameters: InferenceParameters {
-                temperature: 0.0,
-                ..base_desired_state.inference_parameters
-            },
-            ..base_desired_state
-        }),
+        desired_state: ClusterDesiredState::Apply(Box::new(qwen3_0_6b().into_desired_state())),
         wait_for_slots_ready: true,
         ..ClusterParams::default()
     })
     .await
     .expect("the cluster must start");
 
-    let mut location_properties = Map::new();
-    location_properties.insert(
-        "location".to_owned(),
-        json!({"type": "integer", "description": "The city name"}),
-    );
-
     let collected = cluster
         .continue_from_conversation_history(
             CancellationToken::new(),
             &ContinueFromConversationHistoryParams {
                 add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Text(
-                        "What is the weather in Paris? Use the get_weather tool to find out."
-                            .to_owned(),
-                    ),
-                    role: "user".to_owned(),
-                }]),
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Text(WEATHER_TOOL_QUESTION.to_owned()),
+                        role: "user".to_owned(),
+                    },
+                ]),
                 enable_thinking: false,
                 grammar: None,
                 max_tokens: NonZeroU32::new(400).unwrap(),
@@ -72,12 +54,9 @@ async fn agent_reports_tool_call_validation_failure() {
                     function: Function {
                         name: "get_weather".to_owned(),
                         description: "Get the current weather for a location".to_owned(),
-                        parameters: Parameters::Schema(ValidatedParametersSchema {
-                            schema_type: "object".to_owned(),
-                            properties: Some(location_properties),
-                            required: Some(vec!["location".to_owned()]),
-                            additional_properties: Some(Value::Bool(false)),
-                        }),
+                        parameters: Parameters::Schema(
+                            weather_tool_parameters_with_an_unsatisfiable_location(),
+                        ),
                     },
                 })],
             },
@@ -96,8 +75,8 @@ async fn agent_reports_tool_call_validation_failure() {
 
     assert!(
         !validation_failures.is_empty(),
-        "expected at least one ToolCallValidationFailed event when the model emits a string \
-         location against an integer-typed schema; got tokens:\n{}",
+        "expected at least one ToolCallValidationFailed event when the model calls get_weather \
+         with a location that no value can satisfy; got tokens:\n{}",
         collected.text
     );
 
diff --git a/paddler_tests/tests/agent_returns_identical_embeddings_for_identical_documents.rs b/paddler_tests/tests/agent_returns_identical_embeddings_for_identical_documents.rs
index e2f169dd1..9a5724b0c 100644
--- a/paddler_tests/tests/agent_returns_identical_embeddings_for_identical_documents.rs
+++ b/paddler_tests/tests/agent_returns_identical_embeddings_for_identical_documents.rs
@@ -2,7 +2,6 @@
 
 use tokio_util::sync::CancellationToken;
 
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
 use paddler_messaging::embedding_input_document::EmbeddingInputDocument;
 use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
 use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
@@ -14,10 +13,7 @@ use paddler_tests::start_embedding_cluster::start_embedding_cluster;
 async fn agent_returns_identical_embeddings_for_identical_documents() {
     let cluster = start_embedding_cluster(EmbeddingClusterParams {
         agents: vec![AgentConfig::single(1)],
-        inference_parameters: InferenceParameters {
-            enable_embeddings: true,
-            ..InferenceParameters::deterministic()
-        },
+
         ..EmbeddingClusterParams::default()
     })
     .await
diff --git a/paddler_tests/tests/agent_returns_image_decoding_error_for_invalid_base64.rs b/paddler_tests/tests/agent_returns_image_decoding_error_for_invalid_base64.rs
index 9db8096a0..def73fed5 100644
--- a/paddler_tests/tests/agent_returns_image_decoding_error_for_invalid_base64.rs
+++ b/paddler_tests/tests/agent_returns_image_decoding_error_for_invalid_base64.rs
@@ -25,19 +25,21 @@ async fn agent_returns_image_decoding_error_for_invalid_base64() {
             CancellationToken::new(),
             &ContinueFromConversationHistoryParams {
                 add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Parts(vec![
-                        ConversationMessageContentPart::ImageUrl {
-                            image_url: ImageUrl {
-                                url: "data:image/jpeg;base64,!!!not-valid-base64!!!".to_owned(),
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Parts(vec![
+                            ConversationMessageContentPart::ImageUrl {
+                                image_url: ImageUrl {
+                                    url: "data:image/jpeg;base64,!!!not-valid-base64!!!".to_owned(),
+                                },
                             },
-                        },
-                        ConversationMessageContentPart::Text {
-                            text: "Describe this image".to_owned(),
-                        },
-                    ]),
-                    role: "user".to_owned(),
-                }]),
+                            ConversationMessageContentPart::Text {
+                                text: "Describe this image".to_owned(),
+                            },
+                        ]),
+                        role: "user".to_owned(),
+                    },
+                ]),
                 enable_thinking: false,
                 grammar: None,
                 max_tokens: NonZeroU32::new(20).unwrap(),
@@ -48,16 +50,11 @@ async fn agent_returns_image_decoding_error_for_invalid_base64() {
         .await;
 
     let collected = outcome.expect("the request must complete");
-    let saw_decoding_error = collected.token_results.iter().any(|result| {
-        matches!(
-            result.token_result,
-            GeneratedTokenResult::ImageDecodingFailed(_)
-        )
-    });
-
-    assert!(
-        saw_decoding_error,
-        "invalid base64 must produce ImageDecodingFailed"
+    assert_eq!(
+        collected.into_token_results(),
+        vec![GeneratedTokenResult::ImageDecodingFailed(
+            "test-agent-0: failed to decode images: Invalid base64 payload: Invalid symbol 33, offset 0.".to_owned()
+        )]
     );
 
     cluster
diff --git a/paddler_tests/tests/agent_returns_image_decoding_error_for_malformed_data_uri.rs b/paddler_tests/tests/agent_returns_image_decoding_error_for_malformed_data_uri.rs
index f7f938c8e..9c6ab9496 100644
--- a/paddler_tests/tests/agent_returns_image_decoding_error_for_malformed_data_uri.rs
+++ b/paddler_tests/tests/agent_returns_image_decoding_error_for_malformed_data_uri.rs
@@ -25,19 +25,21 @@ async fn agent_returns_image_decoding_error_for_malformed_data_uri() {
             CancellationToken::new(),
             &ContinueFromConversationHistoryParams {
                 add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Parts(vec![
-                        ConversationMessageContentPart::ImageUrl {
-                            image_url: ImageUrl {
-                                url: "data:image/jpegbase64,abc123".to_owned(),
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Parts(vec![
+                            ConversationMessageContentPart::ImageUrl {
+                                image_url: ImageUrl {
+                                    url: "data:image/jpegbase64,abc123".to_owned(),
+                                },
                             },
-                        },
-                        ConversationMessageContentPart::Text {
-                            text: "Describe this image".to_owned(),
-                        },
-                    ]),
-                    role: "user".to_owned(),
-                }]),
+                            ConversationMessageContentPart::Text {
+                                text: "Describe this image".to_owned(),
+                            },
+                        ]),
+                        role: "user".to_owned(),
+                    },
+                ]),
                 enable_thinking: false,
                 grammar: None,
                 max_tokens: NonZeroU32::new(20).unwrap(),
@@ -48,16 +50,12 @@ async fn agent_returns_image_decoding_error_for_malformed_data_uri() {
         .await;
 
     let collected = outcome.expect("the request must complete");
-    let saw_decoding_error = collected.token_results.iter().any(|result| {
-        matches!(
-            result.token_result,
-            GeneratedTokenResult::ImageDecodingFailed(_)
-        )
-    });
-
-    assert!(
-        saw_decoding_error,
-        "malformed data URI must produce ImageDecodingFailed"
+    assert_eq!(
+        collected.into_token_results(),
+        vec![GeneratedTokenResult::ImageDecodingFailed(
+            "test-agent-0: failed to decode images: Invalid base64 payload: Invalid padding"
+                .to_owned()
+        )]
     );
 
     cluster
diff --git a/paddler_tests/tests/agent_returns_image_decoding_error_for_remote_url.rs b/paddler_tests/tests/agent_returns_image_decoding_error_for_remote_url.rs
index 5573b9c9b..0a951fe85 100644
--- a/paddler_tests/tests/agent_returns_image_decoding_error_for_remote_url.rs
+++ b/paddler_tests/tests/agent_returns_image_decoding_error_for_remote_url.rs
@@ -25,19 +25,21 @@ async fn agent_returns_image_decoding_error_for_remote_url() {
             CancellationToken::new(),
             &ContinueFromConversationHistoryParams {
                 add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Parts(vec![
-                        ConversationMessageContentPart::ImageUrl {
-                            image_url: ImageUrl {
-                                url: "https://example.com/image.jpg".to_owned(),
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Parts(vec![
+                            ConversationMessageContentPart::ImageUrl {
+                                image_url: ImageUrl {
+                                    url: "https://example.com/image.jpg".to_owned(),
+                                },
                             },
-                        },
-                        ConversationMessageContentPart::Text {
-                            text: "Describe this image".to_owned(),
-                        },
-                    ]),
-                    role: "user".to_owned(),
-                }]),
+                            ConversationMessageContentPart::Text {
+                                text: "Describe this image".to_owned(),
+                            },
+                        ]),
+                        role: "user".to_owned(),
+                    },
+                ]),
                 enable_thinking: false,
                 grammar: None,
                 max_tokens: NonZeroU32::new(20).unwrap(),
@@ -48,16 +50,11 @@ async fn agent_returns_image_decoding_error_for_remote_url() {
         .await;
 
     let collected = outcome.expect("the request must complete");
-    let saw_decoding_error = collected.token_results.iter().any(|result| {
-        matches!(
-            result.token_result,
-            GeneratedTokenResult::ImageDecodingFailed(_)
-        )
-    });
-
-    assert!(
-        saw_decoding_error,
-        "remote URL must produce ImageDecodingFailed (only data URIs supported)"
+    assert_eq!(
+        collected.into_token_results(),
+        vec![GeneratedTokenResult::ImageDecodingFailed(
+            "test-agent-0: failed to decode images: Remote image URLs are not supported. Use base64 data URIs (data:image/...;base64,...) instead.".to_owned()
+        )]
     );
 
     cluster
diff --git a/paddler_tests/tests/agent_returns_rms_normalized_embeddings_when_requested.rs b/paddler_tests/tests/agent_returns_rms_normalized_embeddings_when_requested.rs
index 9687d9a45..5ff1dff70 100644
--- a/paddler_tests/tests/agent_returns_rms_normalized_embeddings_when_requested.rs
+++ b/paddler_tests/tests/agent_returns_rms_normalized_embeddings_when_requested.rs
@@ -2,47 +2,72 @@
 
 use tokio_util::sync::CancellationToken;
 
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
+use paddler_agent_embeddings::normalization::normalize_embedding::normalize_embedding;
+use paddler_messaging::embedding::Embedding;
 use paddler_messaging::embedding_input_document::EmbeddingInputDocument;
 use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
+use paddler_messaging::embedding_vector::EmbeddingVector;
 use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
 use paddler_test_cluster_harness::embedding_cluster_params::EmbeddingClusterParams;
 use paddler_tests::start_embedding_cluster::start_embedding_cluster;
 
+const RMS_NORMALIZATION: EmbeddingNormalizationMethod =
+    EmbeddingNormalizationMethod::RmsNorm { epsilon: 1e-6 };
+
 #[tokio::test(flavor = "multi_thread")]
 async fn agent_returns_rms_normalized_embeddings_when_requested() {
     let cluster = start_embedding_cluster(EmbeddingClusterParams {
         agents: vec![AgentConfig::single(1)],
-        inference_parameters: InferenceParameters {
-            enable_embeddings: true,
-            ..InferenceParameters::deterministic()
-        },
+
         ..EmbeddingClusterParams::default()
     })
     .await
     .expect("the cluster must start");
-
-    let collected = cluster
-        .generate_embedding_batch(
+    let embedding_of = |normalization_method| {
+        cluster.generate_embedding_batch(
             CancellationToken::new(),
             &GenerateEmbeddingBatchParams {
                 input_batch: vec![EmbeddingInputDocument {
                     content: "Testing RMS normalization on embeddings".to_owned(),
                     id: "doc-rms".to_owned(),
                 }],
-                normalization_method: EmbeddingNormalizationMethod::RmsNorm { epsilon: 1e-6 },
+                normalization_method,
             },
         )
+    };
+
+    let unnormalized_embedding = embedding_of(EmbeddingNormalizationMethod::None)
+        .await
+        .expect("the unnormalized embedding batch must be accepted")
+        .embeddings
+        .into_iter()
+        .map(|produced| produced.embedding)
+        .next()
+        .expect("the document must be embedded without normalization");
+    let rms_normalized = embedding_of(RMS_NORMALIZATION)
         .await
-        .expect("the embedding batch must be accepted");
-
-    assert_eq!(collected.embeddings.len(), 1);
-    assert!(collected.saw_done);
-    assert!(matches!(
-        collected.embeddings[0].embedding.normalization_method,
-        EmbeddingNormalizationMethod::RmsNorm { .. }
-    ));
+        .expect("the RMS-normalized embedding batch must be accepted");
+
+    assert!(rms_normalized.saw_done);
+    assert_eq!(
+        rms_normalized
+            .embeddings
+            .into_iter()
+            .map(|produced| produced.embedding)
+            .collect::<Vec<_>>(),
+        vec![Embedding {
+            embedding: EmbeddingVector::from(
+                normalize_embedding(
+                    unnormalized_embedding.embedding.to_vec(),
+                    &RMS_NORMALIZATION
+                )
+                .expect("an embedding of the served model must fit RMS normalization")
+            ),
+            normalization_method: RMS_NORMALIZATION,
+            ..unnormalized_embedding
+        }]
+    );
 
     cluster
         .shutdown()
diff --git a/paddler_tests/tests/agent_returns_unnormalized_embeddings_when_requested.rs b/paddler_tests/tests/agent_returns_unnormalized_embeddings_when_requested.rs
deleted file mode 100644
index caae6cae3..000000000
--- a/paddler_tests/tests/agent_returns_unnormalized_embeddings_when_requested.rs
+++ /dev/null
@@ -1,51 +0,0 @@
-#![cfg(feature = "tests_that_use_llms")]
-
-use tokio_util::sync::CancellationToken;
-
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
-use paddler_messaging::embedding_input_document::EmbeddingInputDocument;
-use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
-use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
-use paddler_test_cluster_harness::agent_config::AgentConfig;
-use paddler_test_cluster_harness::embedding_cluster_params::EmbeddingClusterParams;
-use paddler_tests::start_embedding_cluster::start_embedding_cluster;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn agent_returns_unnormalized_embeddings_when_requested() {
-    let cluster = start_embedding_cluster(EmbeddingClusterParams {
-        agents: vec![AgentConfig::single(1)],
-        inference_parameters: InferenceParameters {
-            enable_embeddings: true,
-            ..InferenceParameters::deterministic()
-        },
-        ..EmbeddingClusterParams::default()
-    })
-    .await
-    .expect("the cluster must start");
-
-    let collected = cluster
-        .generate_embedding_batch(
-            CancellationToken::new(),
-            &GenerateEmbeddingBatchParams {
-                input_batch: vec![EmbeddingInputDocument {
-                    content: "Testing no normalization on embeddings".to_owned(),
-                    id: "doc-none".to_owned(),
-                }],
-                normalization_method: EmbeddingNormalizationMethod::None,
-            },
-        )
-        .await
-        .expect("the embedding batch must be accepted");
-
-    assert_eq!(collected.embeddings.len(), 1);
-    assert!(collected.saw_done);
-    assert!(matches!(
-        collected.embeddings[0].embedding.normalization_method,
-        EmbeddingNormalizationMethod::None
-    ));
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_tests/tests/agent_runner_connected_to.rs b/paddler_tests/tests/agent_runner_connected_to.rs
new file mode 100644
index 000000000..f0c9584ec
--- /dev/null
+++ b/paddler_tests/tests/agent_runner_connected_to.rs
@@ -0,0 +1,27 @@
+use std::net::SocketAddr;
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_bootstrap::agent_bootstrap_config::AgentBootstrapConfig;
+use paddler_bootstrap::agent_runner::AgentRunner;
+use paddler_bootstrap::agent_runner_params::AgentRunnerParams;
+use paddler_model_source::desired_model_resolver::DesiredModelResolver;
+use paddler_model_source::huggingface_default_endpoint::HUGGINGFACE_DEFAULT_ENDPOINT;
+use paddler_test_cluster_harness::unexpiring_shutdown_options::unexpiring_shutdown_options;
+
+#[must_use]
+pub fn agent_runner_connected_to(management_addr: SocketAddr, slots: u16) -> AgentRunner {
+    AgentRunner::start(AgentRunnerParams {
+        bootstrap_config: AgentBootstrapConfig {
+            agent_name: Some("test-agent".to_owned()),
+            desired_model_resolver: DesiredModelResolver::from_process_env(
+                HUGGINGFACE_DEFAULT_ENDPOINT.to_owned(),
+            )
+            .expect("the model downloader must start"),
+            management_address: management_addr.to_string(),
+            slots,
+        },
+        cancellation_token: CancellationToken::new(),
+        shutdown_options: unexpiring_shutdown_options(),
+    })
+}
diff --git a/paddler_tests/tests/agent_serves_decisions_with_a_batch_too_small_to_warm_up_every_slot.rs b/paddler_tests/tests/agent_serves_decisions_with_a_batch_too_small_to_warm_up_every_slot.rs
new file mode 100644
index 000000000..92b12aaa8
--- /dev/null
+++ b/paddler_tests/tests/agent_serves_decisions_with_a_batch_too_small_to_warm_up_every_slot.rs
@@ -0,0 +1,46 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::mem::discriminant;
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_inference_parameters::batch_size::BatchSize;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_tests::decision_cluster_params::DecisionClusterParams;
+use paddler_tests::decision_desired_state::DecisionDesiredState;
+use paddler_tests::decision_done_discriminant::decision_done_discriminant;
+use paddler_tests::sample_decision::sample_decision;
+use paddler_tests::start_decision_cluster::start_decision_cluster;
+
+const SLOTS: u16 = 2;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agent_serves_decisions_with_a_batch_too_small_to_warm_up_every_slot() {
+    let cluster = start_decision_cluster(DecisionClusterParams {
+        agents: vec![AgentConfig::single(SLOTS)],
+        desired_state: DecisionDesiredState {
+            n_batch: BatchSize::try_from(u32::from(SLOTS))
+                .expect("a batch with one token per slot is valid"),
+            ..DecisionDesiredState::default()
+        },
+        ..DecisionClusterParams::default()
+    })
+    .await
+    .expect("the cluster must start");
+
+    let collected = cluster
+        .decide(CancellationToken::new(), &sample_decision())
+        .await
+        .expect("the decision must be accepted");
+
+    assert_eq!(collected.answers.len(), sample_decision().questions.len());
+    assert_eq!(
+        discriminant(&collected.terminal_result),
+        decision_done_discriminant()
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/agent_serves_embeddings_without_a_chat_template.rs b/paddler_tests/tests/agent_serves_embeddings_without_a_chat_template.rs
index e3d9bdb83..ae5fdc50a 100644
--- a/paddler_tests/tests/agent_serves_embeddings_without_a_chat_template.rs
+++ b/paddler_tests/tests/agent_serves_embeddings_without_a_chat_template.rs
@@ -2,35 +2,26 @@
 
 use tokio_util::sync::CancellationToken;
 
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
-use paddler_messaging::agent_desired_model::AgentDesiredModel;
-use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_inference_parameters::embedding_parameters::EmbeddingParameters;
 use paddler_messaging::embedding_input_document::EmbeddingInputDocument;
 use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
+use paddler_messaging::embedding_result::EmbeddingResult;
 use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
-use paddler_test_cluster_harness::model_card::ModelCard;
 use paddler_test_cluster_harness::model_card::nomic_embed_text_v1_5::nomic_embed_text_v1_5;
 use paddler_tests::start_cluster::start_cluster;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn agent_serves_embeddings_without_a_chat_template() {
-    let ModelCard { reference } = nomic_embed_text_v1_5();
-
     let cluster = start_cluster(ClusterParams {
         agents: AgentConfig::uniform(1, 1),
         wait_for_slots_ready: true,
-        desired_state: Some(BalancerDesiredState {
-            chat_template_override: None,
-            inference_parameters: InferenceParameters {
-                enable_embeddings: true,
-                ..InferenceParameters::deterministic()
-            },
-            model: AgentDesiredModel::HuggingFace(reference),
-            multimodal_projection: AgentDesiredModel::None,
-            use_chat_template_override: false,
-        }),
+        desired_state: ClusterDesiredState::Apply(Box::new(
+            nomic_embed_text_v1_5()
+                .into_desired_state_with_embedding_parameters(EmbeddingParameters::default()),
+        )),
         ..ClusterParams::default()
     })
     .await
@@ -52,7 +43,7 @@ async fn agent_serves_embeddings_without_a_chat_template() {
 
     assert_eq!(collected.embeddings.len(), 1);
     assert!(collected.saw_done);
-    assert!(collected.errors.is_empty());
+    assert_eq!(collected.failures, Vec::<EmbeddingResult>::new());
 
     cluster
         .shutdown()
diff --git a/paddler_tests/tests/agent_serves_request_being_prepared_when_state_change_begins.rs b/paddler_tests/tests/agent_serves_request_being_prepared_when_state_change_begins.rs
index 71d79fecb..295da089e 100644
--- a/paddler_tests/tests/agent_serves_request_being_prepared_when_state_change_begins.rs
+++ b/paddler_tests/tests/agent_serves_request_being_prepared_when_state_change_begins.rs
@@ -11,7 +11,6 @@ use paddler_messaging::conversation_history::ConversationHistory;
 use paddler_messaging::conversation_message::ConversationMessage;
 use paddler_messaging::conversation_message_content::ConversationMessageContent;
 use paddler_messaging::conversation_message_content_part::ConversationMessageContentPart;
-use paddler_messaging::generated_token_result::GeneratedTokenResult;
 use paddler_messaging::image_url::ImageUrl;
 use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
@@ -38,7 +37,7 @@ async fn agent_serves_request_being_prepared_when_state_change_begins() {
         CancellationToken::new(),
         &ContinueFromConversationHistoryParams {
             add_generation_prompt: true,
-            conversation_history: ConversationHistory::new(vec![ConversationMessage {
+            conversation_history: ConversationHistory::new(vec![ConversationMessage::Message {
                 content: ConversationMessageContent::Parts(vec![
                         ConversationMessageContentPart::ImageUrl {
                             image_url: ImageUrl {
@@ -86,8 +85,7 @@ async fn agent_serves_request_being_prepared_when_state_change_begins() {
         .client_management
         .put_balancer_desired_state(
             CancellationToken::new(),
-            &desired_state_with_halved_image_resize(initial_desired_state)
-                .expect("the desired state must be derivable"),
+            &desired_state_with_halved_image_resize(initial_desired_state),
         )
         .await
         .expect("the balancer must accept the desired state");
@@ -110,15 +108,8 @@ async fn agent_serves_request_being_prepared_when_state_change_begins() {
         .expect("the image request task must not panic")
         .expect("the message must be readable");
 
-    assert!(
-        matches!(
-            collected
-                .token_results
-                .last()
-                .map(|token_result_with_producer| &token_result_with_producer.token_result),
-            Some(GeneratedTokenResult::Done(_))
-        ),
-        "the request that was being prepared when the state change began must be served to completion"
+    collected.summary().expect(
+        "the request that was being prepared when the state change began must be served to completion",
     );
 
     cluster
diff --git a/paddler_tests/tests/agent_serves_request_sent_while_state_change_waits_for_in_flight_request.rs b/paddler_tests/tests/agent_serves_request_sent_while_state_change_waits_for_in_flight_request.rs
index 013842c7f..66260a4f5 100644
--- a/paddler_tests/tests/agent_serves_request_sent_while_state_change_waits_for_in_flight_request.rs
+++ b/paddler_tests/tests/agent_serves_request_sent_while_state_change_waits_for_in_flight_request.rs
@@ -15,7 +15,7 @@ use paddler_test_cluster_harness::unending_generation::unending_generation;
 use paddler_tests::desired_state_with_halved_image_resize::desired_state_with_halved_image_resize;
 use paddler_tests::start_cluster_with_qwen3_and_context_size::start_cluster_with_qwen3_and_context_size;
 
-const CONTEXT_SIZE_THAT_ENDS_THE_IN_FLIGHT_GENERATION: u32 = 512;
+const CONTEXT_SIZE_THAT_ENDS_THE_IN_FLIGHT_GENERATION: NonZeroU32 = NonZeroU32::new(512).unwrap();
 
 #[tokio::test(flavor = "multi_thread")]
 async fn agent_serves_request_sent_while_state_change_waits_for_in_flight_request() {
@@ -51,8 +51,7 @@ async fn agent_serves_request_sent_while_state_change_waits_for_in_flight_reques
         .client_management
         .put_balancer_desired_state(
             CancellationToken::new(),
-            &desired_state_with_halved_image_resize(initial_desired_state)
-                .expect("the desired state must be derivable"),
+            &desired_state_with_halved_image_resize(initial_desired_state),
         )
         .await
         .expect("the balancer must accept the desired state");
diff --git a/paddler_tests/tests/agent_shuts_down_cleanly_while_downloading_its_model.rs b/paddler_tests/tests/agent_shuts_down_cleanly_while_downloading_its_model.rs
new file mode 100644
index 000000000..16d3ae149
--- /dev/null
+++ b/paddler_tests/tests/agent_shuts_down_cleanly_while_downloading_its_model.rs
@@ -0,0 +1,34 @@
+use paddler_local_http_fixture::fixture_response::FixtureResponse;
+use paddler_local_http_fixture::local_http_fixture::LocalHttpFixture;
+use paddler_messaging::agent_desired_model::AgentDesiredModel;
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_tests::start_single_agent_cluster_with_desired_state::start_single_agent_cluster_with_desired_state;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agent_shuts_down_cleanly_while_downloading_its_model() {
+    let fixture = LocalHttpFixture::start(FixtureResponse::StallBeforeHeaders)
+        .await
+        .expect("the local HTTP fixture must start");
+    let cluster = start_single_agent_cluster_with_desired_state(BalancerDesiredState {
+        model: AgentDesiredModel::Uri(fixture.url("/stalled.gguf")),
+        ..BalancerDesiredState::default()
+    })
+    .await
+    .expect("a single-agent cluster must start");
+
+    fixture
+        .wait_for_requests(1)
+        .await
+        .expect("the agent must start downloading its model");
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly while its agent downloads a model");
+
+    fixture
+        .stop()
+        .await
+        .expect("the fixture task must not panic")
+        .expect("the fixture must serve every request");
+}
diff --git a/paddler_tests/tests/agent_shuts_down_cleanly_while_loading_its_model.rs b/paddler_tests/tests/agent_shuts_down_cleanly_while_loading_its_model.rs
new file mode 100644
index 000000000..fdb68e658
--- /dev/null
+++ b/paddler_tests/tests/agent_shuts_down_cleanly_while_loading_its_model.rs
@@ -0,0 +1,17 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_tests::qwen3_desired_state::qwen3_desired_state;
+
+use crate::shutdown_during_state_transition::ShutdownDuringStateTransition;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agent_shuts_down_cleanly_while_loading_its_model() {
+    ShutdownDuringStateTransition {
+        initial_desired_state: BalancerDesiredState::default(),
+        next_desired_state: qwen3_desired_state(),
+        slot_count: 1,
+    }
+    .shut_down_while_the_agent_applies_the_next_state()
+    .await;
+}
diff --git a/paddler_tests/tests/agent_shuts_down_cleanly_while_starting_to_serve_decisions.rs b/paddler_tests/tests/agent_shuts_down_cleanly_while_starting_to_serve_decisions.rs
new file mode 100644
index 000000000..896296a95
--- /dev/null
+++ b/paddler_tests/tests/agent_shuts_down_cleanly_while_starting_to_serve_decisions.rs
@@ -0,0 +1,21 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_messaging::decision_slots_minimum::DECISION_SLOTS_MINIMUM;
+use paddler_tests::kev_checkpoint::KevCheckpoint;
+
+use crate::shutdown_during_state_transition::ShutdownDuringStateTransition;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agent_shuts_down_cleanly_while_starting_to_serve_decisions() {
+    ShutdownDuringStateTransition {
+        initial_desired_state: BalancerDesiredState::default(),
+        next_desired_state: KevCheckpoint::Kev0_8b
+            .decision_desired_state()
+            .into_balancer_desired_state()
+            .expect("the kev desired state must name absolute paths"),
+        slot_count: DECISION_SLOTS_MINIMUM.get(),
+    }
+    .shut_down_while_the_agent_applies_the_next_state()
+    .await;
+}
diff --git a/paddler_tests/tests/agent_shuts_down_cleanly_while_switching_to_another_model.rs b/paddler_tests/tests/agent_shuts_down_cleanly_while_switching_to_another_model.rs
new file mode 100644
index 000000000..20546fa64
--- /dev/null
+++ b/paddler_tests/tests/agent_shuts_down_cleanly_while_switching_to_another_model.rs
@@ -0,0 +1,19 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use paddler_inference_parameters::embedding_parameters::EmbeddingParameters;
+use paddler_test_cluster_harness::model_card::nomic_embed_text_v1_5::nomic_embed_text_v1_5;
+use paddler_tests::qwen3_desired_state::qwen3_desired_state;
+
+use crate::shutdown_during_state_transition::ShutdownDuringStateTransition;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agent_shuts_down_cleanly_while_switching_to_another_model() {
+    ShutdownDuringStateTransition {
+        initial_desired_state: qwen3_desired_state(),
+        next_desired_state: nomic_embed_text_v1_5()
+            .into_desired_state_with_embedding_parameters(EmbeddingParameters::default()),
+        slot_count: 1,
+    }
+    .shut_down_while_the_agent_applies_the_next_state()
+    .await;
+}
diff --git a/paddler_tests/tests/agent_shuts_down_cleanly_with_an_embedding_batch_in_flight.rs b/paddler_tests/tests/agent_shuts_down_cleanly_with_an_embedding_batch_in_flight.rs
new file mode 100644
index 000000000..3fd8c4434
--- /dev/null
+++ b/paddler_tests/tests/agent_shuts_down_cleanly_with_an_embedding_batch_in_flight.rs
@@ -0,0 +1,55 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use futures_util::StreamExt as _;
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
+use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::embedding_cluster_params::EmbeddingClusterParams;
+use paddler_tests::numbered_embedding_documents::numbered_embedding_documents;
+use paddler_tests::start_embedding_cluster::start_embedding_cluster;
+
+const DOCUMENTS_OUTLASTING_A_SHUTDOWN: usize = 512;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agent_shuts_down_cleanly_with_an_embedding_batch_in_flight() {
+    let mut cluster = start_embedding_cluster(EmbeddingClusterParams {
+        agents: vec![AgentConfig::single(1)],
+        ..EmbeddingClusterParams::default()
+    })
+    .await
+    .expect("the cluster must start");
+    let mut embedding_stream = cluster
+        .client_inference
+        .post_generate_embedding_batch(
+            CancellationToken::new(),
+            &GenerateEmbeddingBatchParams {
+                input_batch: numbered_embedding_documents(DOCUMENTS_OUTLASTING_A_SHUTDOWN),
+                normalization_method: EmbeddingNormalizationMethod::None,
+            },
+        )
+        .await
+        .expect("the embedding batch must be accepted");
+
+    embedding_stream
+        .next()
+        .await
+        .expect("the embedding batch must produce its first embedding")
+        .expect("the first embedding must be readable");
+
+    cluster
+        .agents
+        .pop()
+        .expect("the cluster must have an agent")
+        .shutdown()
+        .await
+        .expect("the agent must shut down cleanly with an embedding batch in flight");
+
+    drop(embedding_stream);
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/agent_shuts_down_while_state_change_waits_for_in_flight_request.rs b/paddler_tests/tests/agent_shuts_down_while_state_change_waits_for_in_flight_request.rs
index 9f9c87b83..a6b4f2eee 100644
--- a/paddler_tests/tests/agent_shuts_down_while_state_change_waits_for_in_flight_request.rs
+++ b/paddler_tests/tests/agent_shuts_down_while_state_change_waits_for_in_flight_request.rs
@@ -38,8 +38,7 @@ async fn agent_shuts_down_while_state_change_waits_for_in_flight_request() {
         .client_management
         .put_balancer_desired_state(
             CancellationToken::new(),
-            &desired_state_with_halved_image_resize(initial_desired_state)
-                .expect("the desired state must be derivable"),
+            &desired_state_with_halved_image_resize(initial_desired_state),
         )
         .await
         .expect("the balancer must accept the desired state");
diff --git a/paddler_tests/tests/agent_stops_an_embedding_batch_when_the_client_disconnects.rs b/paddler_tests/tests/agent_stops_an_embedding_batch_when_the_client_disconnects.rs
new file mode 100644
index 000000000..dd1ff8393
--- /dev/null
+++ b/paddler_tests/tests/agent_stops_an_embedding_batch_when_the_client_disconnects.rs
@@ -0,0 +1,75 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use futures_util::StreamExt as _;
+use tokio_util::sync::CancellationToken;
+
+use paddler_inference_parameters::embedding_parameters::EmbeddingParameters;
+use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
+use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::embedding_cluster_params::EmbeddingClusterParams;
+use paddler_tests::numbered_embedding_documents::numbered_embedding_documents;
+use paddler_tests::start_embedding_cluster::start_embedding_cluster;
+
+const SLOTS: u16 = 4;
+
+fn embedding_batch(document_count: usize) -> GenerateEmbeddingBatchParams {
+    GenerateEmbeddingBatchParams {
+        input_batch: numbered_embedding_documents(document_count),
+        normalization_method: EmbeddingNormalizationMethod::None,
+    }
+}
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agent_stops_an_embedding_batch_when_the_client_disconnects() {
+    let embedding_parameters = EmbeddingParameters::default();
+    let documents_in_one_chunk = embedding_parameters.embedding_batch_size.get();
+    let mut cluster = start_embedding_cluster(EmbeddingClusterParams {
+        agents: vec![AgentConfig::single(SLOTS)],
+        embedding_parameters,
+        ..EmbeddingClusterParams::default()
+    })
+    .await
+    .expect("the cluster must start");
+    let agent_id = cluster
+        .agent_ids
+        .first()
+        .expect("the cluster must have an agent")
+        .clone();
+    let mut embedding_stream = cluster
+        .client_inference
+        .post_generate_embedding_batch(
+            CancellationToken::new(),
+            &embedding_batch(documents_in_one_chunk),
+        )
+        .await
+        .expect("the embedding batch must be accepted");
+
+    embedding_stream
+        .next()
+        .await
+        .expect("the agent must produce a first embedding")
+        .expect("the first embedding must be readable");
+    drop(embedding_stream);
+
+    cluster
+        .wait_for_slots_processing(&agent_id, 0)
+        .await
+        .expect("the agent must stop the abandoned embedding batch");
+
+    let collected = cluster
+        .generate_embedding_batch(
+            CancellationToken::new(),
+            &embedding_batch(usize::from(SLOTS)),
+        )
+        .await
+        .expect("an embedding batch after the abandoned one must be accepted");
+
+    assert_eq!(collected.embeddings.len(), usize::from(SLOTS));
+    assert!(collected.saw_done);
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/agent_streams_tokens_for_a_causal_image_larger_than_n_batch.rs b/paddler_tests/tests/agent_streams_tokens_for_a_causal_image_larger_than_n_batch.rs
index 0a6c1d425..8dc7ca29e 100644
--- a/paddler_tests/tests/agent_streams_tokens_for_a_causal_image_larger_than_n_batch.rs
+++ b/paddler_tests/tests/agent_streams_tokens_for_a_causal_image_larger_than_n_batch.rs
@@ -2,85 +2,63 @@
 
 use std::num::NonZeroU32;
 
-use anyhow::Result;
 use tokio_util::sync::CancellationToken;
 
-use paddler_messaging::conversation_history::ConversationHistory;
-use paddler_messaging::conversation_message::ConversationMessage;
-use paddler_messaging::conversation_message_content::ConversationMessageContent;
-use paddler_messaging::conversation_message_content_part::ConversationMessageContentPart;
-use paddler_messaging::image_url::ImageUrl;
-use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
+use paddler_inference_parameters::batch_size::BatchSize;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
-use paddler_test_cluster_harness::cluster::Cluster;
-use paddler_test_cluster_harness::load_fixture_data_uri::load_fixture_data_uri;
+use paddler_tests::jpeg_and_webp_image_fixtures::JPEG_AND_WEBP_IMAGE_FIXTURES;
+use paddler_tests::single_image_question::SingleImageQuestion;
 use paddler_tests::start_cluster_with_smolvlm2_and_n_batch::start_cluster_with_smolvlm2_and_n_batch;
 
 const MAX_TOKENS: NonZeroU32 = NonZeroU32::new(20).unwrap();
 const N_BATCH_SMALLER_THAN_ONE_IMAGE: u32 = 32;
 
-async fn assert_streams_tokens_for_image(
-    cluster: &Cluster,
-    fixture_name: &str,
-    mime_type: &str,
-) -> Result<()> {
-    let collected = cluster
-        .continue_from_conversation_history(
-            CancellationToken::new(),
-            &ContinueFromConversationHistoryParams {
-                add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Parts(vec![
-                        ConversationMessageContentPart::ImageUrl {
-                            image_url: ImageUrl {
-                                url: load_fixture_data_uri(fixture_name, mime_type)?,
-                            },
-                        },
-                        ConversationMessageContentPart::Text {
-                            text: "Describe this image.".to_owned(),
-                        },
-                    ]),
-                    role: "user".to_owned(),
-                }]),
-                enable_thinking: false,
-                grammar: None,
-                max_tokens: MAX_TOKENS,
-                parse_tool_calls: false,
-                tools: vec![],
-            },
-        )
-        .await?;
-
-    assert!(
-        collected
-            .token_results
-            .iter()
-            .any(|token_result_with_producer| token_result_with_producer.token_result.is_token()),
-        "fixture {fixture_name} must stream tokens: {:?}",
-        collected.token_results
-    );
-    assert!(
-        collected.summary()?.usage.input_image_tokens > u64::from(N_BATCH_SMALLER_THAN_ONE_IMAGE)
-    );
-
-    Ok(())
-}
-
 #[tokio::test(flavor = "multi_thread")]
 async fn agent_streams_tokens_for_a_causal_image_larger_than_n_batch() {
     let cluster = start_cluster_with_smolvlm2_and_n_batch(
         vec![AgentConfig::single(1)],
-        N_BATCH_SMALLER_THAN_ONE_IMAGE,
+        BatchSize::try_from(N_BATCH_SMALLER_THAN_ONE_IMAGE)
+            .expect("the batch size must be within the supported range"),
     )
     .await
     .expect("the cluster must start");
 
-    assert_streams_tokens_for_image(&cluster, "sarnow.jpeg", "image/jpeg")
-        .await
-        .expect("the image request must stream tokens");
-    assert_streams_tokens_for_image(&cluster, "llamas.webp", "image/webp")
-        .await
-        .expect("the image request must stream tokens");
+    for image_fixture in JPEG_AND_WEBP_IMAGE_FIXTURES {
+        let collected = cluster
+            .continue_from_conversation_history(
+                CancellationToken::new(),
+                &SingleImageQuestion {
+                    image_data_uri: image_fixture
+                        .data_uri()
+                        .expect("the image fixture must load"),
+                    max_tokens: MAX_TOKENS,
+                    question: "Describe this image.",
+                }
+                .into_conversation_params(),
+            )
+            .await
+            .expect("the image request must stream tokens");
+
+        assert!(
+            collected
+                .token_results
+                .iter()
+                .any(|token_result_with_producer| token_result_with_producer
+                    .token_result
+                    .is_token()),
+            "fixture {} must stream tokens: {:?}",
+            image_fixture.fixture_name,
+            collected.token_results
+        );
+        assert!(
+            collected
+                .summary()
+                .expect("the generation must report its summary")
+                .usage
+                .input_image_tokens
+                > u64::from(N_BATCH_SMALLER_THAN_ONE_IMAGE)
+        );
+    }
 
     cluster
         .shutdown()
diff --git a/paddler_tests/tests/agent_streams_tokens_from_conversation_history_over_http.rs b/paddler_tests/tests/agent_streams_tokens_from_conversation_history_over_http.rs
index 24ac69776..56219e0ce 100644
--- a/paddler_tests/tests/agent_streams_tokens_from_conversation_history_over_http.rs
+++ b/paddler_tests/tests/agent_streams_tokens_from_conversation_history_over_http.rs
@@ -22,10 +22,12 @@ async fn agent_streams_tokens_from_conversation_history_over_http() {
             CancellationToken::new(),
             &ContinueFromConversationHistoryParams {
                 add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Text("Say hello".to_owned()),
-                    role: "user".to_owned(),
-                }]),
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Text("Say hello".to_owned()),
+                        role: "user".to_owned(),
+                    },
+                ]),
                 enable_thinking: true,
                 grammar: None,
                 max_tokens: NonZeroU32::new(50).unwrap(),
diff --git a/paddler_tests/tests/agent_streams_tokens_from_image_data_uri.rs b/paddler_tests/tests/agent_streams_tokens_from_image_data_uri.rs
index 37a1dc161..b25a7b0d5 100644
--- a/paddler_tests/tests/agent_streams_tokens_from_image_data_uri.rs
+++ b/paddler_tests/tests/agent_streams_tokens_from_image_data_uri.rs
@@ -27,19 +27,21 @@ async fn agent_streams_tokens_from_image_data_uri() {
             CancellationToken::new(),
             &ContinueFromConversationHistoryParams {
                 add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Parts(vec![
-                        ConversationMessageContentPart::ImageUrl {
-                            image_url: ImageUrl {
-                                url: image_data_uri,
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Parts(vec![
+                            ConversationMessageContentPart::ImageUrl {
+                                image_url: ImageUrl {
+                                    url: image_data_uri,
+                                },
                             },
-                        },
-                        ConversationMessageContentPart::Text {
-                            text: "What do you see in this image?".to_owned(),
-                        },
-                    ]),
-                    role: "user".to_owned(),
-                }]),
+                            ConversationMessageContentPart::Text {
+                                text: "What do you see in this image?".to_owned(),
+                            },
+                        ]),
+                        role: "user".to_owned(),
+                    },
+                ]),
                 enable_thinking: true,
                 grammar: None,
                 max_tokens: NonZeroU32::new(100).unwrap(),
diff --git a/paddler_tests/tests/agent_text_only_model_rejects_image_input.rs b/paddler_tests/tests/agent_text_only_model_rejects_image_input.rs
index 53fcdf231..a2a7e4ff6 100644
--- a/paddler_tests/tests/agent_text_only_model_rejects_image_input.rs
+++ b/paddler_tests/tests/agent_text_only_model_rejects_image_input.rs
@@ -28,19 +28,21 @@ async fn agent_text_only_model_rejects_image_input() {
             CancellationToken::new(),
             &ContinueFromConversationHistoryParams {
                 add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Parts(vec![
-                        ConversationMessageContentPart::ImageUrl {
-                            image_url: ImageUrl {
-                                url: image_data_uri,
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Parts(vec![
+                            ConversationMessageContentPart::ImageUrl {
+                                image_url: ImageUrl {
+                                    url: image_data_uri,
+                                },
                             },
-                        },
-                        ConversationMessageContentPart::Text {
-                            text: "Describe this image".to_owned(),
-                        },
-                    ]),
-                    role: "user".to_owned(),
-                }]),
+                            ConversationMessageContentPart::Text {
+                                text: "Describe this image".to_owned(),
+                            },
+                        ]),
+                        role: "user".to_owned(),
+                    },
+                ]),
                 enable_thinking: false,
                 grammar: None,
                 max_tokens: NonZeroU32::new(20).unwrap(),
diff --git a/paddler_tests/tests/agent_tokenizes_gemma_3_chat_prompt_with_a_single_bos_token.rs b/paddler_tests/tests/agent_tokenizes_gemma_3_chat_prompt_with_a_single_bos_token.rs
index b4ebbeaba..439b396b7 100644
--- a/paddler_tests/tests/agent_tokenizes_gemma_3_chat_prompt_with_a_single_bos_token.rs
+++ b/paddler_tests/tests/agent_tokenizes_gemma_3_chat_prompt_with_a_single_bos_token.rs
@@ -25,10 +25,12 @@ async fn agent_tokenizes_gemma_3_chat_prompt_with_a_single_bos_token() {
             CancellationToken::new(),
             &ContinueFromConversationHistoryParams {
                 add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Text("hi".to_owned()),
-                    role: "user".to_owned(),
-                }]),
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Text("hi".to_owned()),
+                        role: "user".to_owned(),
+                    },
+                ]),
                 enable_thinking: false,
                 grammar: None,
                 max_tokens: NonZeroU32::new(1).unwrap(),
diff --git a/paddler_tests/tests/agent_without_a_model_refuses_every_request_as_model_not_loaded.rs b/paddler_tests/tests/agent_without_a_model_refuses_every_request_as_model_not_loaded.rs
new file mode 100644
index 000000000..4d67284e4
--- /dev/null
+++ b/paddler_tests/tests/agent_without_a_model_refuses_every_request_as_model_not_loaded.rs
@@ -0,0 +1,166 @@
+use std::num::NonZeroU16;
+use std::num::NonZeroU32;
+
+use serde_json::to_string;
+use tokio_tungstenite::tungstenite::Message;
+
+use paddler_agent_decision::decision_error::DecisionError;
+use paddler_agent_embeddings::embedding_error::EmbeddingError;
+use paddler_agent_runtime::rejection_description::rejection_description;
+use paddler_agent_text_generation::text_generation_error::TextGenerationError;
+use paddler_messaging::conversation_history::ConversationHistory;
+use paddler_messaging::decision_question::DecisionQuestion;
+use paddler_messaging::decision_result::DecisionResult;
+use paddler_messaging::embedding_input_document::EmbeddingInputDocument;
+use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
+use paddler_messaging::embedding_result::EmbeddingResult;
+use paddler_messaging::generated_token_result::GeneratedTokenResult;
+use paddler_messaging::jsonrpc::request_envelope::RequestEnvelope;
+use paddler_messaging::jsonrpc::response_envelope::ResponseEnvelope;
+use paddler_messaging::management_socket::agent::message::Message as AgentJsonRpcMessage;
+use paddler_messaging::management_socket::agent::request::Request as AgentJsonRpcRequest;
+use paddler_messaging::management_socket::agent::response::Response as AgentJsonRpcResponse;
+use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
+use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
+use paddler_messaging::request_params::decide_params::DecideParams;
+use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
+use paddler_messaging::reserved_request::ReservedRequest;
+use paddler_test_cluster_harness::untrusted_balancer_socket_server::UntrustedBalancerSocketServer;
+
+use crate::agent_runner_connected_to::agent_runner_connected_to;
+
+const AGENT_NAME: Option<&str> = Some("test-agent");
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agent_without_a_model_refuses_every_request_as_model_not_loaded() {
+    let untrusted_balancer = UntrustedBalancerSocketServer::bind()
+        .await
+        .expect("the untrusted balancer must bind");
+    let _agent_runner = agent_runner_connected_to(untrusted_balancer.management_addr, 1);
+    let mut connection = untrusted_balancer
+        .accept_agent()
+        .await
+        .expect("the agent must connect");
+    let requests = [
+        RequestEnvelope {
+            id: "conversation".to_owned(),
+            request: AgentJsonRpcRequest::ContinueFromConversationHistory(ReservedRequest {
+                params: ContinueFromConversationHistoryParams {
+                    add_generation_prompt: true,
+                    conversation_history: ConversationHistory::new(Vec::new()),
+                    enable_thinking: false,
+                    grammar: None,
+                    max_tokens: NonZeroU32::MIN,
+                    parse_tool_calls: false,
+                    tools: Vec::new(),
+                },
+                reserved_slots: NonZeroU16::MIN,
+            }),
+        },
+        RequestEnvelope {
+            id: "decision".to_owned(),
+            request: AgentJsonRpcRequest::Decide(ReservedRequest {
+                params: DecideParams {
+                    last_question: DecisionQuestion {
+                        id: "paid".to_owned(),
+                        instructions: "Was it paid?".to_owned(),
+                        options: vec!["no".to_owned(), "yes".to_owned()],
+                    },
+                    leading_questions: Vec::new(),
+                    state: "The invoice was paid on time.".to_owned(),
+                },
+                reserved_slots: NonZeroU16::MIN,
+            }),
+        },
+        RequestEnvelope {
+            id: "embedding-batch".to_owned(),
+            request: AgentJsonRpcRequest::GenerateEmbeddingBatch(ReservedRequest {
+                params: GenerateEmbeddingBatchParams {
+                    input_batch: vec![EmbeddingInputDocument {
+                        content: "document".to_owned(),
+                        id: "document".to_owned(),
+                    }],
+                    normalization_method: EmbeddingNormalizationMethod::None,
+                },
+                reserved_slots: NonZeroU16::MIN,
+            }),
+        },
+        RequestEnvelope {
+            id: "raw-prompt".to_owned(),
+            request: AgentJsonRpcRequest::ContinueFromRawPrompt(ReservedRequest {
+                params: ContinueFromRawPromptParams {
+                    grammar: None,
+                    max_tokens: NonZeroU32::MIN,
+                    raw_prompt: "hello".to_owned(),
+                },
+                reserved_slots: NonZeroU16::MIN,
+            }),
+        },
+    ];
+    let request_count = requests.len();
+
+    for request in requests {
+        connection
+            .send(Message::text(
+                to_string(&AgentJsonRpcMessage::Request(request))
+                    .expect("the request must serialize"),
+            ))
+            .await
+            .expect("the request must be sent");
+    }
+
+    let mut responses = Vec::with_capacity(request_count);
+
+    for _request in 0..request_count {
+        responses.push(
+            connection
+                .next_agent_response()
+                .await
+                .expect("the agent must answer every request"),
+        );
+    }
+
+    responses.sort_by(|first_response, second_response| {
+        first_response.request_id.cmp(&second_response.request_id)
+    });
+
+    assert_eq!(
+        responses,
+        vec![
+            ResponseEnvelope {
+                generated_by: None,
+                request_id: "conversation".to_owned(),
+                response: AgentJsonRpcResponse::GeneratedToken(
+                    GeneratedTokenResult::ModelNotLoaded(rejection_description(
+                        AGENT_NAME,
+                        &TextGenerationError::ModelNotLoaded
+                    ))
+                ),
+            },
+            ResponseEnvelope {
+                generated_by: None,
+                request_id: "decision".to_owned(),
+                response: AgentJsonRpcResponse::Decision(DecisionResult::ModelNotLoaded(
+                    rejection_description(AGENT_NAME, &DecisionError::ModelNotLoaded)
+                )),
+            },
+            ResponseEnvelope {
+                generated_by: None,
+                request_id: "embedding-batch".to_owned(),
+                response: AgentJsonRpcResponse::Embedding(EmbeddingResult::ModelNotLoaded(
+                    rejection_description(AGENT_NAME, &EmbeddingError::ModelNotLoaded)
+                )),
+            },
+            ResponseEnvelope {
+                generated_by: None,
+                request_id: "raw-prompt".to_owned(),
+                response: AgentJsonRpcResponse::GeneratedToken(
+                    GeneratedTokenResult::ModelNotLoaded(rejection_description(
+                        AGENT_NAME,
+                        &TextGenerationError::ModelNotLoaded
+                    ))
+                ),
+            },
+        ]
+    );
+}
diff --git a/paddler_tests/tests/agents_stop_serving_decisions_once_the_pointer_head_is_removed.rs b/paddler_tests/tests/agents_stop_serving_decisions_once_the_pointer_head_is_removed.rs
new file mode 100644
index 000000000..d2b16b89e
--- /dev/null
+++ b/paddler_tests/tests/agents_stop_serving_decisions_once_the_pointer_head_is_removed.rs
@@ -0,0 +1,78 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::mem::discriminant;
+use std::num::NonZeroU32;
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::agent_decisions_serving::AgentDecisionsServing;
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_messaging::decision_settings::DecisionSettings;
+use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
+use paddler_tests::decision_cluster_params::DecisionClusterParams;
+use paddler_tests::decision_done_discriminant::decision_done_discriminant;
+use paddler_tests::sample_decision::sample_decision;
+use paddler_tests::start_decision_cluster::start_decision_cluster;
+
+const MAX_TOKENS: NonZeroU32 = NonZeroU32::new(16).unwrap();
+
+#[tokio::test(flavor = "multi_thread")]
+async fn agents_stop_serving_decisions_once_the_pointer_head_is_removed() {
+    let mut cluster = start_decision_cluster(DecisionClusterParams::default())
+        .await
+        .expect("a decision cluster must start");
+
+    let decided = cluster
+        .decide(CancellationToken::new(), &sample_decision())
+        .await
+        .expect("the decision cluster must answer a decision");
+
+    assert_eq!(
+        discriminant(&decided.terminal_result),
+        decision_done_discriminant()
+    );
+
+    let decision_desired_state = cluster
+        .client_management
+        .get_balancer_desired_state(CancellationToken::new())
+        .await
+        .expect("the balancer must report its desired state");
+
+    cluster
+        .client_management
+        .put_balancer_desired_state(
+            CancellationToken::new(),
+            &BalancerDesiredState {
+                decision: DecisionSettings::default(),
+                ..decision_desired_state
+            },
+        )
+        .await
+        .expect("the balancer must stop serving decisions");
+    cluster
+        .wait_for_first_agent_to_serve(AgentDecisionsServing::NotServed)
+        .await
+        .expect("the agent must stop serving decisions");
+
+    let generated = cluster
+        .continue_from_raw_prompt(
+            CancellationToken::new(),
+            &ContinueFromRawPromptParams {
+                grammar: None,
+                max_tokens: MAX_TOKENS,
+                raw_prompt: "The capital of France is".to_owned(),
+            },
+        )
+        .await
+        .expect("the cluster without a pointer head must generate tokens");
+
+    assert_ne!(generated.text, "");
+    generated
+        .summary()
+        .expect("the generation must finish with a summary");
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/an_agent_cannot_answer_a_request_dispatched_to_another_agent.rs b/paddler_tests/tests/an_agent_cannot_answer_a_request_dispatched_to_another_agent.rs
new file mode 100644
index 000000000..ade90f1ca
--- /dev/null
+++ b/paddler_tests/tests/an_agent_cannot_answer_a_request_dispatched_to_another_agent.rs
@@ -0,0 +1,111 @@
+use std::collections::BTreeMap;
+
+use tokio::spawn;
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::agent_status::AgentStatus;
+use paddler_messaging::management_socket::agent::response::Response as AgentJsonRpcResponse;
+use paddler_messaging::model_metadata::ModelMetadata;
+use paddler_messaging::slot_aggregated_status_snapshot::SlotAggregatedStatusSnapshot;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_tests::start_cluster::start_cluster;
+
+use crate::untrusted_management_agent::UntrustedManagementAgent;
+
+const ANSWERING_AGENT_ID: &str = "answering";
+const INJECTING_AGENT_ID: &str = "injecting";
+const INJECTING_AGENT_DESIRED_SLOTS_AFTER_INJECTION: u16 = 7;
+
+fn model_metadata_named(name: &str) -> ModelMetadata {
+    ModelMetadata {
+        metadata: BTreeMap::from([("general.name".to_owned(), name.to_owned())]),
+    }
+}
+
+#[tokio::test(flavor = "multi_thread")]
+async fn an_agent_cannot_answer_a_request_dispatched_to_another_agent() {
+    let mut cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("a cluster without agents must start");
+    let management_addr = cluster.balancer.addresses.management;
+    let mut answering_agent = UntrustedManagementAgent::connect_registered(
+        management_addr,
+        ANSWERING_AGENT_ID,
+        SlotAggregatedStatusSnapshot::default(),
+    )
+    .await
+    .expect("the answering agent must register");
+    let mut injecting_agent = UntrustedManagementAgent::connect_registered(
+        management_addr,
+        INJECTING_AGENT_ID,
+        SlotAggregatedStatusSnapshot::default(),
+    )
+    .await
+    .expect("the injecting agent must register");
+
+    cluster
+        .wait_for_agent_count(2)
+        .await
+        .expect("the balancer must register both agents");
+
+    let client_management = cluster.client_management.clone();
+    let model_metadata_answer = spawn(async move {
+        client_management
+            .get_model_metadata(CancellationToken::new(), ANSWERING_AGENT_ID)
+            .await
+    });
+    let model_metadata_request_id = answering_agent
+        .next_request()
+        .await
+        .expect("the balancer must ask the answering agent for its model metadata")
+        .id;
+
+    injecting_agent
+        .respond(
+            model_metadata_request_id.clone(),
+            AgentJsonRpcResponse::ModelMetadata(Some(model_metadata_named("injected"))),
+        )
+        .await
+        .expect("the injected response must be sent");
+    injecting_agent
+        .report_status(SlotAggregatedStatusSnapshot {
+            status: AgentStatus {
+                desired_slots_total: INJECTING_AGENT_DESIRED_SLOTS_AFTER_INJECTION,
+                ..AgentStatus::default()
+            },
+            version: 1,
+        })
+        .await
+        .expect("the status report following the injection must be sent");
+    cluster
+        .agents_watcher
+        .until_agent(INJECTING_AGENT_ID, |snapshot| {
+            snapshot.agents.iter().any(|agent| {
+                agent.id == INJECTING_AGENT_ID
+                    && agent.status.desired_slots_total
+                        == INJECTING_AGENT_DESIRED_SLOTS_AFTER_INJECTION
+            })
+        })
+        .await
+        .expect("the balancer must handle everything the injecting agent sent");
+    answering_agent
+        .respond(
+            model_metadata_request_id,
+            AgentJsonRpcResponse::ModelMetadata(Some(model_metadata_named("answered"))),
+        )
+        .await
+        .expect("the genuine response must be sent");
+
+    assert_eq!(
+        model_metadata_answer
+            .await
+            .expect("the model metadata request task must not panic")
+            .expect("the balancer must answer the model metadata request"),
+        Some(model_metadata_named("answered"))
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/an_agent_releases_its_old_model_before_loading_the_new_one.rs b/paddler_tests/tests/an_agent_releases_its_old_model_before_loading_the_new_one.rs
new file mode 100644
index 000000000..27ad59cf2
--- /dev/null
+++ b/paddler_tests/tests/an_agent_releases_its_old_model_before_loading_the_new_one.rs
@@ -0,0 +1,58 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::agent_runtime_status::AgentRuntimeStatus;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::model_card::smolvlm2_256m::smolvlm2_256m;
+use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn an_agent_releases_its_old_model_before_loading_the_new_one() {
+    let new_model = smolvlm2_256m();
+    let new_model_path = new_model
+        .cached_path()
+        .expect("the new model must be in the Hugging Face cache")
+        .to_str()
+        .expect("the new model path must be valid UTF-8")
+        .to_owned();
+    let mut cluster = start_cluster_with_qwen3(vec![AgentConfig::single(1)])
+        .await
+        .expect("the cluster must start");
+    let agent_id = cluster
+        .agent_ids
+        .first()
+        .expect("the cluster must have one registered agent")
+        .clone();
+    let mut observed_the_old_model_released = false;
+
+    cluster
+        .client_management
+        .put_balancer_desired_state(CancellationToken::new(), &new_model.into_desired_state())
+        .await
+        .expect("the balancer must accept the desired state");
+    cluster
+        .agents_watcher
+        .until_agent(&agent_id, |snapshot| {
+            snapshot
+                .agents
+                .iter()
+                .filter(|registered_agent| registered_agent.id == agent_id)
+                .any(|registered_agent| {
+                    observed_the_old_model_released |= registered_agent.status.runtime
+                        == AgentRuntimeStatus::Idle
+                        && registered_agent.status.model_path.is_none();
+
+                    registered_agent.status.model_path.as_deref() == Some(new_model_path.as_str())
+                })
+        })
+        .await
+        .expect("the agent must serve the new model");
+
+    assert!(observed_the_old_model_released);
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/an_agent_serving_decisions_also_serves_completions_and_embeddings.rs b/paddler_tests/tests/an_agent_serving_decisions_also_serves_completions_and_embeddings.rs
new file mode 100644
index 000000000..d80c96d3d
--- /dev/null
+++ b/paddler_tests/tests/an_agent_serving_decisions_also_serves_completions_and_embeddings.rs
@@ -0,0 +1,92 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::mem::discriminant;
+use std::num::NonZeroU32;
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_inference_parameters::embedding_parameters::EmbeddingParameters;
+use paddler_inference_parameters::pooling_type::PoolingType;
+use paddler_messaging::embedding_input_document::EmbeddingInputDocument;
+use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
+use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
+use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
+use paddler_tests::decision_cluster_params::DecisionClusterParams;
+use paddler_tests::decision_desired_state::DecisionDesiredState;
+use paddler_tests::decision_done_discriminant::decision_done_discriminant;
+use paddler_tests::sample_decision::sample_decision;
+use paddler_tests::start_decision_cluster::start_decision_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn an_agent_serving_decisions_also_serves_completions_and_embeddings() {
+    let cluster = start_decision_cluster(DecisionClusterParams {
+        desired_state: DecisionDesiredState {
+            embedding_parameters: EmbeddingParameters {
+                pooling_type: PoolingType::Last,
+                ..EmbeddingParameters::default()
+            },
+            ..DecisionDesiredState::default()
+        },
+        ..DecisionClusterParams::default()
+    })
+    .await
+    .expect("a decision cluster must start");
+
+    let decided = cluster
+        .decide(CancellationToken::new(), &sample_decision())
+        .await
+        .expect("the decision must be accepted");
+    let generated = cluster
+        .continue_from_raw_prompt(
+            CancellationToken::new(),
+            &ContinueFromRawPromptParams {
+                grammar: None,
+                max_tokens: NonZeroU32::MIN,
+                raw_prompt: "The capital of France is".to_owned(),
+            },
+        )
+        .await
+        .expect("the raw prompt must be accepted");
+    let embedded = cluster
+        .generate_embedding_batch(
+            CancellationToken::new(),
+            &GenerateEmbeddingBatchParams {
+                input_batch: vec![EmbeddingInputDocument {
+                    content: "the quick brown fox jumps over the lazy dog".to_owned(),
+                    id: "doc-1".to_owned(),
+                }],
+                normalization_method: EmbeddingNormalizationMethod::None,
+            },
+        )
+        .await
+        .expect("the embedding batch must be accepted");
+
+    assert_eq!(
+        discriminant(&decided.terminal_result),
+        decision_done_discriminant()
+    );
+    assert_eq!(
+        generated
+            .summary()
+            .expect("the generation must finish with a summary")
+            .usage
+            .completion_tokens(),
+        1
+    );
+    assert_eq!(
+        embedded
+            .embeddings
+            .iter()
+            .map(|embedding_with_producer| embedding_with_producer
+                .embedding
+                .source_document_id
+                .clone())
+            .collect::<Vec<_>>(),
+        vec!["doc-1".to_owned()]
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/an_agent_whose_balancer_stops_reading_reconnects.rs b/paddler_tests/tests/an_agent_whose_balancer_stops_reading_reconnects.rs
new file mode 100644
index 000000000..15db03e75
--- /dev/null
+++ b/paddler_tests/tests/an_agent_whose_balancer_stops_reading_reconnects.rs
@@ -0,0 +1,46 @@
+use serde_json::to_string;
+use tokio_tungstenite::tungstenite::Message;
+
+use paddler_messaging::jsonrpc::request_envelope::RequestEnvelope;
+use paddler_messaging::management_socket::agent::message::Message as AgentJsonRpcMessage;
+use paddler_messaging::management_socket::agent::request::Request as AgentJsonRpcRequest;
+use paddler_tcp_buffer_limits::tcp_buffer_limits_file::TcpBufferLimitsFile;
+use paddler_test_cluster_harness::untrusted_balancer_socket_server::UntrustedBalancerSocketServer;
+
+use crate::agent_runner_connected_to::agent_runner_connected_to;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn an_agent_whose_balancer_stops_reading_reconnects() {
+    let largest_send_buffer = TcpBufferLimitsFile::of_kernel_send_buffers()
+        .largest_buffer()
+        .expect("the kernel must report its TCP send buffer limits");
+    let largest_receive_buffer = TcpBufferLimitsFile::of_kernel_receive_buffers()
+        .largest_buffer()
+        .expect("the kernel must report its TCP receive buffer limits");
+    let untrusted_balancer = UntrustedBalancerSocketServer::bind()
+        .await
+        .expect("the untrusted balancer must bind");
+    let _agent_runner = agent_runner_connected_to(untrusted_balancer.management_addr, 1);
+    let mut unread_connection = untrusted_balancer
+        .accept_agent()
+        .await
+        .expect("the agent must connect");
+
+    for request_index in 0..=(largest_receive_buffer / largest_send_buffer + 1) {
+        unread_connection
+            .send(Message::text(
+                to_string(&AgentJsonRpcMessage::Request(RequestEnvelope {
+                    id: format!("{request_index}-{}", "x".repeat(largest_send_buffer)),
+                    request: AgentJsonRpcRequest::GetModelMetadata,
+                }))
+                .expect("the agent request must serialize"),
+            ))
+            .await
+            .expect("the agent must read every request whose answer it cannot deliver");
+    }
+
+    untrusted_balancer
+        .accept_agent()
+        .await
+        .expect("the agent must reconnect once its balancer stops reading its answers");
+}
diff --git a/paddler_tests/tests/an_agent_whose_model_cannot_start_reports_that_it_is_stuck.rs b/paddler_tests/tests/an_agent_whose_model_cannot_start_reports_that_it_is_stuck.rs
new file mode 100644
index 000000000..323015d70
--- /dev/null
+++ b/paddler_tests/tests/an_agent_whose_model_cannot_start_reports_that_it_is_stuck.rs
@@ -0,0 +1,48 @@
+use std::io::Write as _;
+
+use tempfile::NamedTempFile;
+
+use paddler_messaging::agent_state_application_status::AgentStateApplicationStatus;
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_test_cluster_harness::agent_desired_model_at::agent_desired_model_at;
+use paddler_tests::start_single_agent_cluster_with_desired_state::start_single_agent_cluster_with_desired_state;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn an_agent_whose_model_cannot_start_reports_that_it_is_stuck() {
+    let mut unloadable_model =
+        NamedTempFile::new().expect("a temporary model file must be creatable");
+
+    unloadable_model
+        .write_all(b"this is not a valid gguf model file")
+        .expect("the unloadable model contents must be writable");
+
+    let mut cluster = start_single_agent_cluster_with_desired_state(BalancerDesiredState {
+        model: agent_desired_model_at(unloadable_model.path())
+            .expect("the temporary model path must be absolute"),
+        ..BalancerDesiredState::default()
+    })
+    .await
+    .expect("a single-agent cluster must start");
+    let agent_id = cluster
+        .agent_ids
+        .first()
+        .expect("the cluster must have one registered agent")
+        .clone();
+
+    cluster
+        .agents_watcher
+        .until_agent(&agent_id, |snapshot| {
+            snapshot.agents.iter().any(|registered_agent| {
+                registered_agent.id == agent_id
+                    && registered_agent.status.state_application_status
+                        == AgentStateApplicationStatus::Stuck
+            })
+        })
+        .await
+        .expect("the agent must report that its pipeline is stuck after retrying");
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/an_agent_whose_new_model_cannot_load_serves_nothing.rs b/paddler_tests/tests/an_agent_whose_new_model_cannot_load_serves_nothing.rs
new file mode 100644
index 000000000..340be333c
--- /dev/null
+++ b/paddler_tests/tests/an_agent_whose_new_model_cannot_load_serves_nothing.rs
@@ -0,0 +1,77 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::io::Write as _;
+
+use tempfile::NamedTempFile;
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::agent_issue::AgentIssue;
+use paddler_messaging::agent_issue_params::model_path::ModelPath;
+use paddler_messaging::agent_runtime_status::AgentRuntimeStatus;
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::agent_desired_model_at::agent_desired_model_at;
+use paddler_tests::qwen3_desired_state::qwen3_desired_state;
+use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn an_agent_whose_new_model_cannot_load_serves_nothing() {
+    let mut unloadable_model =
+        NamedTempFile::new().expect("a temporary model file must be creatable");
+
+    unloadable_model
+        .write_all(b"this is not a valid gguf model file")
+        .expect("the unloadable model contents must be writable");
+
+    let unloadable_model_path = unloadable_model
+        .path()
+        .to_str()
+        .expect("the temporary model path must be valid UTF-8")
+        .to_owned();
+    let mut cluster = start_cluster_with_qwen3(vec![AgentConfig::single(1)])
+        .await
+        .expect("the cluster must start");
+    let agent_id = cluster
+        .agent_ids
+        .first()
+        .expect("the cluster must have one registered agent")
+        .clone();
+
+    cluster
+        .client_management
+        .put_balancer_desired_state(
+            CancellationToken::new(),
+            &BalancerDesiredState {
+                model: agent_desired_model_at(unloadable_model.path())
+                    .expect("the temporary model path must be absolute"),
+                ..qwen3_desired_state()
+            },
+        )
+        .await
+        .expect("the balancer must accept the desired state");
+    cluster
+        .wait_for_first_agent_issue(|issue| {
+            *issue
+                == AgentIssue::ModelCannotBeLoaded(ModelPath {
+                    model_path: unloadable_model_path.clone(),
+                })
+        })
+        .await
+        .expect("the agent must report that the new model cannot be loaded");
+    cluster
+        .agents_watcher
+        .until_agent(&agent_id, |snapshot| {
+            snapshot.agents.iter().any(|registered_agent| {
+                registered_agent.id == agent_id
+                    && registered_agent.status.runtime == AgentRuntimeStatus::Idle
+                    && registered_agent.status.model_path.is_none()
+            })
+        })
+        .await
+        .expect("the agent must serve nothing once the old model was released");
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/an_agent_whose_projection_cannot_see_images_rejects_an_image_prompt.rs b/paddler_tests/tests/an_agent_whose_projection_cannot_see_images_rejects_an_image_prompt.rs
new file mode 100644
index 000000000..6c0adc655
--- /dev/null
+++ b/paddler_tests/tests/an_agent_whose_projection_cannot_see_images_rejects_an_image_prompt.rs
@@ -0,0 +1,78 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::num::NonZeroU32;
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::conversation_history::ConversationHistory;
+use paddler_messaging::conversation_message::ConversationMessage;
+use paddler_messaging::conversation_message_content::ConversationMessageContent;
+use paddler_messaging::conversation_message_content_part::ConversationMessageContentPart;
+use paddler_messaging::generated_token_result::GeneratedTokenResult;
+use paddler_messaging::image_url::ImageUrl;
+use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_test_cluster_harness::load_test_image_data_uri::load_test_image_data_uri;
+use paddler_test_cluster_harness::model_card::ultravox_v0_5_llama_3_2_1b::ultravox_v0_5_llama_3_2_1b;
+use paddler_test_cluster_harness::model_card::ultravox_v0_5_llama_3_2_1b_mmproj::ultravox_v0_5_llama_3_2_1b_mmproj;
+use paddler_tests::start_cluster::start_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn an_agent_whose_projection_cannot_see_images_rejects_an_image_prompt() {
+    let cluster = start_cluster(ClusterParams {
+        agents: vec![AgentConfig::single(1)],
+        desired_state: ClusterDesiredState::Apply(Box::new(
+            ultravox_v0_5_llama_3_2_1b()
+                .into_desired_state_with_multimodal_projection(ultravox_v0_5_llama_3_2_1b_mmproj()),
+        )),
+        wait_for_slots_ready: true,
+        ..ClusterParams::default()
+    })
+    .await
+    .expect("an agent with an audio projection must start");
+
+    let collected = cluster
+        .continue_from_conversation_history(
+            CancellationToken::new(),
+            &ContinueFromConversationHistoryParams {
+                add_generation_prompt: true,
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Parts(vec![
+                            ConversationMessageContentPart::ImageUrl {
+                                image_url: ImageUrl {
+                                    url: load_test_image_data_uri()
+                                        .expect("the test image must load"),
+                                },
+                            },
+                            ConversationMessageContentPart::Text {
+                                text: "Describe the image".to_owned(),
+                            },
+                        ]),
+                        role: "user".to_owned(),
+                    },
+                ]),
+                enable_thinking: false,
+                grammar: None,
+                max_tokens: NonZeroU32::MIN,
+                parse_tool_calls: false,
+                tools: vec![],
+            },
+        )
+        .await
+        .expect("the inference request must be accepted");
+
+    assert_eq!(
+        collected.into_token_results(),
+        vec![GeneratedTokenResult::MultimodalTokenizationFailed(
+            "test-agent: failed to tokenize multimodal input: llama.cpp returned 2: media preprocessing failed, or its tokenizer threw (text too large, missing vocab, media marker and bitmap count mismatch, or allocation failure)".to_owned()
+        )]
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/an_agent_without_a_pointer_head_rejects_decisions.rs b/paddler_tests/tests/an_agent_without_a_pointer_head_rejects_decisions.rs
new file mode 100644
index 000000000..159556fe3
--- /dev/null
+++ b/paddler_tests/tests/an_agent_without_a_pointer_head_rejects_decisions.rs
@@ -0,0 +1,52 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::num::NonZeroU16;
+
+use tokio::sync::mpsc;
+
+use paddler_agent::pipeline_request::PipelineRequest;
+use paddler_agent_runtime::agent_request::AgentRequest;
+use paddler_messaging::decision_question::DecisionQuestion;
+use paddler_messaging::decision_result::DecisionResult;
+use paddler_messaging::request_params::decide_params::DecideParams;
+use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
+
+use crate::serving_pipeline_arbiter::ServingPipelineArbiter;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn an_agent_without_a_pointer_head_rejects_decisions() {
+    let mut serving_pipeline_arbiter = ServingPipelineArbiter::new(1);
+
+    serving_pipeline_arbiter
+        .serve(qwen3_0_6b().into_desired_state())
+        .await;
+
+    let (response_tx, mut response_rx) = mpsc::unbounded_channel();
+    let (_stop_tx, stop_rx) = mpsc::unbounded_channel();
+
+    serving_pipeline_arbiter
+        .pipeline_arbiter
+        .forward(PipelineRequest::Decide(AgentRequest {
+            params: DecideParams {
+                last_question: DecisionQuestion {
+                    id: "paid".to_owned(),
+                    instructions: "Was it paid?".to_owned(),
+                    options: vec!["no".to_owned(), "yes".to_owned()],
+                },
+                leading_questions: Vec::new(),
+                state: "The invoice was paid on time.".to_owned(),
+            },
+            response_tx,
+            slot_guard: serving_pipeline_arbiter.slot_guard(NonZeroU16::MIN),
+            stop_rx,
+        }));
+
+    assert_eq!(
+        response_rx.recv().await,
+        Some(DecisionResult::DecisionsNotServed(
+            "agent: no pointer head is configured, so the agent serves no decisions".to_owned()
+        ))
+    );
+
+    serving_pipeline_arbiter.shut_down().await;
+}
diff --git a/paddler_tests/tests/an_embedding_batch_evaluates_every_document_token_exactly_once.rs b/paddler_tests/tests/an_embedding_batch_evaluates_every_document_token_exactly_once.rs
new file mode 100644
index 000000000..5dd001143
--- /dev/null
+++ b/paddler_tests/tests/an_embedding_batch_evaluates_every_document_token_exactly_once.rs
@@ -0,0 +1,66 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::num::NonZeroU16;
+
+use llama_cpp_bindings::token::LlamaToken;
+
+use paddler_agent_embeddings::embedding_input_tokenized::EmbeddingInputTokenized;
+use paddler_agent_embeddings::embedding_normalization::EmbeddingNormalization;
+use paddler_inference_parameters::pooling_type::PoolingType;
+use paddler_tests::active_embedding_request_fixture::ActiveEmbeddingRequestFixture;
+
+use crate::nomic_embedding_batch_processing::NomicEmbeddingBatchProcessing;
+
+const REQUEST_SEQUENCES: NonZeroU16 = NonZeroU16::new(2).unwrap();
+
+#[test]
+fn an_embedding_batch_evaluates_every_document_token_exactly_once() {
+    let nomic_embedding_batch_processing =
+        NomicEmbeddingBatchProcessing::load(PoolingType::Mean, REQUEST_SEQUENCES);
+    let documents = vec![
+        EmbeddingInputTokenized {
+            id: "short".to_owned(),
+            tokens: vec![
+                LlamaToken::new(101),
+                LlamaToken::new(2023),
+                LlamaToken::new(102),
+            ],
+        },
+        EmbeddingInputTokenized {
+            id: "long".to_owned(),
+            tokens: vec![
+                LlamaToken::new(101),
+                LlamaToken::new(2023),
+                LlamaToken::new(2003),
+                LlamaToken::new(1037),
+                LlamaToken::new(102),
+            ],
+        },
+    ];
+    let document_tokens: usize = documents.iter().map(|document| document.tokens.len()).sum();
+    let ActiveEmbeddingRequestFixture {
+        mut admitted_requests,
+        generate_embedding_stop_tx: _generate_embedding_stop_tx,
+        generated_embedding_rx: _generated_embedding_rx,
+    } = nomic_embedding_batch_processing.admit(documents);
+    let mut llama_context = nomic_embedding_batch_processing.embedding_llama_context();
+
+    llama_context.reset_timings();
+
+    while !admitted_requests.is_empty() {
+        admitted_requests = nomic_embedding_batch_processing.embed_next_batch_in(
+            &mut llama_context,
+            nomic_embedding_batch_processing.batch_capacity(),
+            EmbeddingNormalization,
+            admitted_requests,
+        );
+    }
+
+    llama_context.synchronize();
+
+    assert_eq!(
+        usize::try_from(llama_context.timings().n_p_eval())
+            .expect("the evaluated token count must not be negative"),
+        document_tokens
+    );
+}
diff --git a/paddler_tests/tests/an_embedding_batch_runs_on_the_slots_its_agent_has_free.rs b/paddler_tests/tests/an_embedding_batch_runs_on_the_slots_its_agent_has_free.rs
new file mode 100644
index 000000000..0ea8e98e4
--- /dev/null
+++ b/paddler_tests/tests/an_embedding_batch_runs_on_the_slots_its_agent_has_free.rs
@@ -0,0 +1,68 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_inference_parameters::embedding_parameters::EmbeddingParameters;
+use paddler_inference_parameters::pooling_type::PoolingType;
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
+use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_tests::numbered_embedding_documents::numbered_embedding_documents;
+use paddler_tests::qwen3_desired_state::qwen3_desired_state;
+use paddler_tests::start_cluster::start_cluster;
+
+use crate::slot_held_by_an_unending_generation::SlotHeldByAnUnendingGeneration;
+
+const SLOT_COUNT: u16 = 2;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn an_embedding_batch_runs_on_the_slots_its_agent_has_free() {
+    let mut cluster = start_cluster(ClusterParams {
+        agents: vec![AgentConfig::single(SLOT_COUNT)],
+        desired_state: ClusterDesiredState::Apply(Box::new(BalancerDesiredState {
+            embeddings: EmbeddingParameters {
+                pooling_type: PoolingType::Last,
+                ..EmbeddingParameters::default()
+            },
+            ..qwen3_desired_state()
+        })),
+        wait_for_slots_ready: true,
+        ..ClusterParams::default()
+    })
+    .await
+    .expect("the cluster must start");
+    let slot_held_by_an_unending_generation =
+        SlotHeldByAnUnendingGeneration::hold(&mut cluster).await;
+    let embedded = cluster
+        .generate_embedding_batch(
+            CancellationToken::new(),
+            &GenerateEmbeddingBatchParams {
+                input_batch: numbered_embedding_documents(usize::from(SLOT_COUNT)),
+                normalization_method: EmbeddingNormalizationMethod::None,
+            },
+        )
+        .await
+        .expect("the embedding batch must be accepted");
+
+    assert_eq!(
+        embedded
+            .embeddings
+            .iter()
+            .map(|embedding_with_producer| embedding_with_producer
+                .embedding
+                .source_document_id
+                .clone())
+            .collect::<Vec<_>>(),
+        vec!["doc-0".to_owned(), "doc-1".to_owned()]
+    );
+
+    slot_held_by_an_unending_generation.release().await;
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/an_embedding_batch_stopped_before_its_first_batch_produces_no_embeddings.rs b/paddler_tests/tests/an_embedding_batch_stopped_before_its_first_batch_produces_no_embeddings.rs
new file mode 100644
index 000000000..26d842326
--- /dev/null
+++ b/paddler_tests/tests/an_embedding_batch_stopped_before_its_first_batch_produces_no_embeddings.rs
@@ -0,0 +1,48 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::num::NonZeroU16;
+
+use llama_cpp_bindings::token::LlamaToken;
+use tokio::sync::mpsc::error::TryRecvError;
+
+use paddler_agent_embeddings::embedding_input_tokenized::EmbeddingInputTokenized;
+use paddler_agent_embeddings::embedding_normalization::EmbeddingNormalization;
+use paddler_inference_parameters::pooling_type::PoolingType;
+use paddler_messaging::embedding_result::EmbeddingResult;
+use paddler_tests::active_embedding_request_fixture::ActiveEmbeddingRequestFixture;
+
+use crate::nomic_embedding_batch_processing::NomicEmbeddingBatchProcessing;
+
+#[test]
+fn an_embedding_batch_stopped_before_its_first_batch_produces_no_embeddings() {
+    let nomic_embedding_batch_processing =
+        NomicEmbeddingBatchProcessing::load(PoolingType::Mean, NonZeroU16::MIN);
+    let ActiveEmbeddingRequestFixture {
+        admitted_requests,
+        generate_embedding_stop_tx,
+        mut generated_embedding_rx,
+    } = nomic_embedding_batch_processing.admit(vec![EmbeddingInputTokenized {
+        id: "document".to_owned(),
+        tokens: vec![LlamaToken::new(0)],
+    }]);
+
+    generate_embedding_stop_tx
+        .send(())
+        .expect("the stop request must reach the embedding batch");
+
+    let continuing_requests = nomic_embedding_batch_processing.embed_next_batch(
+        nomic_embedding_batch_processing.batch_capacity(),
+        EmbeddingNormalization,
+        admitted_requests,
+    );
+
+    assert_eq!(continuing_requests.len(), 0);
+    assert_eq!(
+        generated_embedding_rx.try_recv(),
+        Ok(EmbeddingResult::NoEmbeddingsProduced)
+    );
+    assert_eq!(
+        generated_embedding_rx.try_recv(),
+        Err(TryRecvError::Disconnected)
+    );
+}
diff --git a/paddler_tests/tests/an_embedding_context_without_pooling_cannot_read_sequence_embeddings.rs b/paddler_tests/tests/an_embedding_context_without_pooling_cannot_read_sequence_embeddings.rs
new file mode 100644
index 000000000..7c8e302ac
--- /dev/null
+++ b/paddler_tests/tests/an_embedding_context_without_pooling_cannot_read_sequence_embeddings.rs
@@ -0,0 +1,43 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::num::NonZeroU16;
+
+use llama_cpp_bindings::error::EmbeddingsError;
+use llama_cpp_bindings::token::LlamaToken;
+
+use paddler_agent_embeddings::embedding_error::EmbeddingError;
+use paddler_agent_embeddings::embedding_input_tokenized::EmbeddingInputTokenized;
+use paddler_agent_embeddings::embedding_normalization::EmbeddingNormalization;
+use paddler_inference_parameters::pooling_type::PoolingType;
+use paddler_messaging::embedding_result::EmbeddingResult;
+use paddler_tests::active_embedding_request_fixture::ActiveEmbeddingRequestFixture;
+
+use crate::nomic_embedding_batch_processing::NomicEmbeddingBatchProcessing;
+
+#[test]
+fn an_embedding_context_without_pooling_cannot_read_sequence_embeddings() {
+    let nomic_embedding_batch_processing =
+        NomicEmbeddingBatchProcessing::load(PoolingType::None, NonZeroU16::MIN);
+    let ActiveEmbeddingRequestFixture {
+        admitted_requests,
+        generate_embedding_stop_tx: _generate_embedding_stop_tx,
+        mut generated_embedding_rx,
+    } = nomic_embedding_batch_processing.admit(vec![EmbeddingInputTokenized {
+        id: "document".to_owned(),
+        tokens: vec![LlamaToken::new(101), LlamaToken::new(102)],
+    }]);
+
+    let continuing_requests = nomic_embedding_batch_processing.embed_next_batch(
+        nomic_embedding_batch_processing.batch_capacity(),
+        EmbeddingNormalization,
+        admitted_requests,
+    );
+
+    assert_eq!(continuing_requests.len(), 0);
+    assert_eq!(
+        generated_embedding_rx.try_recv(),
+        Ok(EmbeddingResult::EmbeddingsUnavailable(
+            EmbeddingError::EmbeddingsUnavailable(EmbeddingsError::NonePoolType).to_string()
+        ))
+    );
+}
diff --git a/paddler_tests/tests/an_embedding_input_llama_cpp_rejects_fails_the_embedding_batch.rs b/paddler_tests/tests/an_embedding_input_llama_cpp_rejects_fails_the_embedding_batch.rs
new file mode 100644
index 000000000..2ca0f8c1f
--- /dev/null
+++ b/paddler_tests/tests/an_embedding_input_llama_cpp_rejects_fails_the_embedding_batch.rs
@@ -0,0 +1,49 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::num::NonZeroU16;
+
+use llama_cpp_bindings::error::DecodeError;
+use llama_cpp_bindings::token::LlamaToken;
+
+use paddler_agent_embeddings::embedding_error::EmbeddingError;
+use paddler_agent_embeddings::embedding_input_tokenized::EmbeddingInputTokenized;
+use paddler_agent_embeddings::embedding_normalization::EmbeddingNormalization;
+use paddler_inference_parameters::pooling_type::PoolingType;
+use paddler_messaging::embedding_result::EmbeddingResult;
+use paddler_tests::active_embedding_request_fixture::ActiveEmbeddingRequestFixture;
+
+use crate::nomic_embedding_batch_processing::NomicEmbeddingBatchProcessing;
+
+#[test]
+fn an_embedding_input_llama_cpp_rejects_fails_the_embedding_batch() {
+    let nomic_embedding_batch_processing =
+        NomicEmbeddingBatchProcessing::load(PoolingType::Mean, NonZeroU16::MIN);
+    let ActiveEmbeddingRequestFixture {
+        admitted_requests,
+        generate_embedding_stop_tx: _generate_embedding_stop_tx,
+        mut generated_embedding_rx,
+    } = nomic_embedding_batch_processing.admit(vec![EmbeddingInputTokenized {
+        id: "document".to_owned(),
+        tokens: vec![LlamaToken::new(
+            nomic_embedding_batch_processing
+                .llama_runtime
+                .loaded_llama_model
+                .model
+                .n_vocab(),
+        )],
+    }]);
+
+    let continuing_requests = nomic_embedding_batch_processing.embed_next_batch(
+        nomic_embedding_batch_processing.batch_capacity(),
+        EmbeddingNormalization,
+        admitted_requests,
+    );
+
+    assert_eq!(continuing_requests.len(), 0);
+    assert_eq!(
+        generated_embedding_rx.try_recv(),
+        Ok(EmbeddingResult::DecodeFailed(
+            EmbeddingError::DecodeFailed(DecodeError::BatchInvalid).to_string()
+        ))
+    );
+}
diff --git a/paddler_tests/tests/an_embedding_input_overflowing_its_batch_fails_as_batch_assembly.rs b/paddler_tests/tests/an_embedding_input_overflowing_its_batch_fails_as_batch_assembly.rs
new file mode 100644
index 000000000..6fab991a6
--- /dev/null
+++ b/paddler_tests/tests/an_embedding_input_overflowing_its_batch_fails_as_batch_assembly.rs
@@ -0,0 +1,49 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::num::NonZeroU16;
+
+use llama_cpp_bindings::batch_add_error::BatchAddError;
+use llama_cpp_bindings::token::LlamaToken;
+use llama_cpp_bindings_types::BatchCapacity;
+
+use paddler_agent_embeddings::embedding_error::EmbeddingError;
+use paddler_agent_embeddings::embedding_input_tokenized::EmbeddingInputTokenized;
+use paddler_agent_embeddings::embedding_normalization::EmbeddingNormalization;
+use paddler_inference_parameters::pooling_type::PoolingType;
+use paddler_messaging::embedding_result::EmbeddingResult;
+use paddler_tests::active_embedding_request_fixture::ActiveEmbeddingRequestFixture;
+
+use crate::nomic_embedding_batch_processing::NomicEmbeddingBatchProcessing;
+
+const TOKENS_THE_BATCH_HOLDS: NonZeroU16 = NonZeroU16::MIN;
+
+#[test]
+fn an_embedding_input_overflowing_its_batch_fails_as_batch_assembly() {
+    let nomic_embedding_batch_processing =
+        NomicEmbeddingBatchProcessing::load(PoolingType::Mean, NonZeroU16::MIN);
+    let ActiveEmbeddingRequestFixture {
+        admitted_requests,
+        generate_embedding_stop_tx: _generate_embedding_stop_tx,
+        mut generated_embedding_rx,
+    } = nomic_embedding_batch_processing.admit(vec![EmbeddingInputTokenized {
+        id: "document".to_owned(),
+        tokens: vec![LlamaToken::new(101), LlamaToken::new(102)],
+    }]);
+
+    let continuing_requests = nomic_embedding_batch_processing.embed_next_batch(
+        BatchCapacity::from(TOKENS_THE_BATCH_HOLDS),
+        EmbeddingNormalization,
+        admitted_requests,
+    );
+
+    assert_eq!(continuing_requests.len(), 0);
+    assert_eq!(
+        generated_embedding_rx.try_recv(),
+        Ok(EmbeddingResult::BatchAssemblyFailed(
+            EmbeddingError::BatchAssemblyFailed(BatchAddError::InsufficientSpace(usize::from(
+                TOKENS_THE_BATCH_HOLDS.get()
+            )))
+            .to_string()
+        ))
+    );
+}
diff --git a/paddler_tests/tests/an_embedding_request_accepted_before_its_scheduler_stops_is_finished.rs b/paddler_tests/tests/an_embedding_request_accepted_before_its_scheduler_stops_is_finished.rs
new file mode 100644
index 000000000..59d73fd8d
--- /dev/null
+++ b/paddler_tests/tests/an_embedding_request_accepted_before_its_scheduler_stops_is_finished.rs
@@ -0,0 +1,44 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent_runtime::shuts_down::ShutsDown as _;
+use paddler_messaging::embedding_result::EmbeddingResult;
+
+use crate::embedding_documents_fixture::EmbeddingDocumentsFixture;
+use crate::kev_model_serving_without_decisions::KevModelServingWithoutDecisions;
+
+const DOCUMENT_COUNT: usize = 64;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn an_embedding_request_accepted_before_its_scheduler_stops_is_finished() {
+    let KevModelServingWithoutDecisions {
+        llama_runtime,
+        served_model,
+        ..
+    } = KevModelServingWithoutDecisions::start(&CancellationToken::new()).await;
+    let EmbeddingDocumentsFixture {
+        agent_request,
+        generate_embedding_stop_tx: _generate_embedding_stop_tx,
+        generated_embedding_rx,
+    } = EmbeddingDocumentsFixture::embedded_one_at_a_time(
+        DOCUMENT_COUNT,
+        llama_runtime
+            .inference_runtime_context
+            .slot_aggregated_status
+            .clone(),
+    );
+
+    served_model
+        .capabilities
+        .embeddings
+        .request_preparer
+        .prepare(agent_request);
+    served_model.shut_down().await;
+
+    let embedding_results =
+        EmbeddingDocumentsFixture::collect_results(generated_embedding_rx).await;
+
+    assert_eq!(embedding_results.len(), DOCUMENT_COUNT + 1);
+    assert_eq!(embedding_results.last(), Some(&EmbeddingResult::Done));
+}
diff --git a/paddler_tests/tests/an_embedding_request_running_when_the_agent_shuts_down_is_refused_as_unavailable.rs b/paddler_tests/tests/an_embedding_request_running_when_the_agent_shuts_down_is_refused_as_unavailable.rs
new file mode 100644
index 000000000..ede077348
--- /dev/null
+++ b/paddler_tests/tests/an_embedding_request_running_when_the_agent_shuts_down_is_refused_as_unavailable.rs
@@ -0,0 +1,50 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent_runtime::shuts_down::ShutsDown as _;
+use paddler_messaging::embedding_result::EmbeddingResult;
+
+use crate::embedding_documents_fixture::EmbeddingDocumentsFixture;
+use crate::kev_model_serving_without_decisions::KevModelServingWithoutDecisions;
+
+const DOCUMENT_COUNT: usize = 64;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn an_embedding_request_running_when_the_agent_shuts_down_is_refused_as_unavailable() {
+    let agent_shutdown = CancellationToken::new();
+    let KevModelServingWithoutDecisions {
+        llama_runtime,
+        served_model,
+        ..
+    } = KevModelServingWithoutDecisions::start(&agent_shutdown).await;
+    let EmbeddingDocumentsFixture {
+        agent_request,
+        generate_embedding_stop_tx: _generate_embedding_stop_tx,
+        generated_embedding_rx,
+    } = EmbeddingDocumentsFixture::embedded_one_at_a_time(
+        DOCUMENT_COUNT,
+        llama_runtime
+            .inference_runtime_context
+            .slot_aggregated_status
+            .clone(),
+    );
+
+    served_model
+        .capabilities
+        .embeddings
+        .request_preparer
+        .prepare(agent_request);
+    agent_shutdown.cancel();
+
+    assert_eq!(
+        EmbeddingDocumentsFixture::collect_results(generated_embedding_rx)
+            .await
+            .last(),
+        Some(&EmbeddingResult::SchedulerUnavailable(
+            "agent: the scheduler is no longer accepting requests".to_owned()
+        ))
+    );
+
+    served_model.shut_down().await;
+}
diff --git a/paddler_tests/tests/an_embedding_that_cannot_be_normalized_fails_the_embedding_batch.rs b/paddler_tests/tests/an_embedding_that_cannot_be_normalized_fails_the_embedding_batch.rs
new file mode 100644
index 000000000..55cc7bc9a
--- /dev/null
+++ b/paddler_tests/tests/an_embedding_that_cannot_be_normalized_fails_the_embedding_batch.rs
@@ -0,0 +1,61 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::num::NonZeroU16;
+
+use llama_cpp_bindings::token::LlamaToken;
+use tokio::sync::mpsc::error::TryRecvError;
+
+use paddler_agent_embeddings::embedding_error::EmbeddingError;
+use paddler_agent_embeddings::embedding_input_tokenized::EmbeddingInputTokenized;
+use paddler_inference_parameters::pooling_type::PoolingType;
+use paddler_messaging::embedding_result::EmbeddingResult;
+use paddler_tests::active_embedding_request_fixture::ActiveEmbeddingRequestFixture;
+use paddler_tests::normalization_of_an_embedding_too_long_for_rms_norm::DIMENSIONS_TOO_MANY_FOR_RMS_NORM;
+use paddler_tests::normalization_of_an_embedding_too_long_for_rms_norm::NormalizationOfAnEmbeddingTooLongForRmsNorm;
+
+use crate::nomic_embedding_batch_processing::NomicEmbeddingBatchProcessing;
+
+const REQUEST_SEQUENCES: NonZeroU16 = NonZeroU16::new(2).unwrap();
+
+#[test]
+fn an_embedding_that_cannot_be_normalized_fails_the_embedding_batch() {
+    let nomic_embedding_batch_processing =
+        NomicEmbeddingBatchProcessing::load(PoolingType::Mean, REQUEST_SEQUENCES);
+    let ActiveEmbeddingRequestFixture {
+        admitted_requests,
+        generate_embedding_stop_tx: _generate_embedding_stop_tx,
+        mut generated_embedding_rx,
+    } = nomic_embedding_batch_processing.admit(vec![
+        EmbeddingInputTokenized {
+            id: "first".to_owned(),
+            tokens: vec![LlamaToken::new(101), LlamaToken::new(102)],
+        },
+        EmbeddingInputTokenized {
+            id: "second".to_owned(),
+            tokens: vec![LlamaToken::new(101), LlamaToken::new(102)],
+        },
+    ]);
+
+    let continuing_requests = nomic_embedding_batch_processing.embed_next_batch(
+        nomic_embedding_batch_processing.batch_capacity(),
+        NormalizationOfAnEmbeddingTooLongForRmsNorm,
+        admitted_requests,
+    );
+
+    assert_eq!(continuing_requests.len(), 0);
+    assert_eq!(
+        generated_embedding_rx.try_recv(),
+        Ok(EmbeddingResult::EmbeddingTooLongForRmsNormalization(
+            EmbeddingError::EmbeddingTooLongForRmsNormalization {
+                dimensions: DIMENSIONS_TOO_MANY_FOR_RMS_NORM,
+                source: u16::try_from(DIMENSIONS_TOO_MANY_FOR_RMS_NORM)
+                    .expect_err("the dimensions must exceed sixteen bits"),
+            }
+            .to_string()
+        ))
+    );
+    assert_eq!(
+        generated_embedding_rx.try_recv(),
+        Err(TryRecvError::Disconnected)
+    );
+}
diff --git a/paddler_tests/tests/an_empty_tool_call_section_resolves_to_no_tool_calls.rs b/paddler_tests/tests/an_empty_tool_call_section_resolves_to_no_tool_calls.rs
new file mode 100644
index 000000000..b15decdb6
--- /dev/null
+++ b/paddler_tests/tests/an_empty_tool_call_section_resolves_to_no_tool_calls.rs
@@ -0,0 +1,46 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::generated_token_result::GeneratedTokenResult;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_tests::gbnf_literal::gbnf_literal;
+use paddler_tests::get_weather_tool::get_weather_tool;
+use paddler_tests::max_tokens_for_forced_text::max_tokens_for_forced_text;
+use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
+use paddler_tests::weather_question_conversation::weather_question_conversation;
+
+const FORCED_OUTPUT: &str = "<tool_call></tool_call>";
+
+#[tokio::test(flavor = "multi_thread")]
+async fn an_empty_tool_call_section_resolves_to_no_tool_calls() {
+    let cluster = start_cluster_with_qwen3(AgentConfig::uniform(1, 1))
+        .await
+        .expect("the cluster must start");
+
+    let token_results = cluster
+        .continue_from_conversation_history(
+            CancellationToken::new(),
+            &weather_question_conversation(
+                gbnf_literal(FORCED_OUTPUT),
+                max_tokens_for_forced_text(FORCED_OUTPUT),
+                vec![get_weather_tool()],
+            ),
+        )
+        .await
+        .expect("the inference request must be accepted")
+        .into_token_results();
+
+    assert_eq!(
+        token_results
+            .iter()
+            .filter(|token_result| token_result.is_tool_call_parsed())
+            .collect::<Vec<_>>(),
+        vec![&GeneratedTokenResult::ToolCallParsed(Vec::new())]
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/an_untrusted_agent_waiting_for_a_request_reports_the_balancer_closing_its_socket.rs b/paddler_tests/tests/an_untrusted_agent_waiting_for_a_request_reports_the_balancer_closing_its_socket.rs
new file mode 100644
index 000000000..614fec70a
--- /dev/null
+++ b/paddler_tests/tests/an_untrusted_agent_waiting_for_a_request_reports_the_balancer_closing_its_socket.rs
@@ -0,0 +1,51 @@
+use tokio_tungstenite::tungstenite::Message;
+use tokio_tungstenite::tungstenite::protocol::CloseFrame;
+use tokio_tungstenite::tungstenite::protocol::frame::coding::CloseCode;
+
+use paddler_test_cluster_harness::cluster_harness_error::ClusterHarnessError;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_tests::start_cluster::start_cluster;
+
+use crate::untrusted_management_agent::UntrustedManagementAgent;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn an_untrusted_agent_waiting_for_a_request_reports_the_balancer_closing_its_socket() {
+    let cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("a cluster without agents must start");
+    let mut closed_agent = UntrustedManagementAgent::connect(
+        cluster.balancer.addresses.management,
+        "closed-before-a-request",
+    )
+    .await
+    .expect("the untrusted agent must reach the agent socket");
+
+    closed_agent
+        .socket_client
+        .send(Message::text("not json"))
+        .await
+        .expect("the undeserializable message must be sent");
+
+    assert_eq!(
+        closed_agent
+            .next_request()
+            .await
+            .err()
+            .as_ref()
+            .map(ToString::to_string),
+        Some(
+            ClusterHarnessError::AgentSocketClosedBeforeText {
+                close_frame: Some(CloseFrame {
+                    code: CloseCode::Invalid,
+                    reason: "Agent message could not be deserialized".into(),
+                })
+            }
+            .to_string()
+        )
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/an_untrusted_agent_waiting_for_a_request_skips_the_pong_answering_its_ping.rs b/paddler_tests/tests/an_untrusted_agent_waiting_for_a_request_skips_the_pong_answering_its_ping.rs
new file mode 100644
index 000000000..93e6e4620
--- /dev/null
+++ b/paddler_tests/tests/an_untrusted_agent_waiting_for_a_request_skips_the_pong_answering_its_ping.rs
@@ -0,0 +1,78 @@
+use std::mem::discriminant;
+
+use tokio::spawn;
+use tokio_tungstenite::tungstenite::Bytes;
+use tokio_tungstenite::tungstenite::Message;
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::management_socket::agent::request::Request as AgentJsonRpcRequest;
+use paddler_messaging::management_socket::agent::response::Response as AgentJsonRpcResponse;
+use paddler_messaging::management_socket::balancer::notification_params::register_agent_params::RegisterAgentParams;
+use paddler_messaging::slot_aggregated_status_snapshot::SlotAggregatedStatusSnapshot;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_tests::start_cluster::start_cluster;
+
+use crate::untrusted_management_agent::UntrustedManagementAgent;
+
+const PINGING_AGENT_ID: &str = "pinging";
+
+#[tokio::test(flavor = "multi_thread")]
+async fn an_untrusted_agent_waiting_for_a_request_skips_the_pong_answering_its_ping() {
+    let mut cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("a cluster without agents must start");
+    let mut pinging_agent =
+        UntrustedManagementAgent::connect(cluster.balancer.addresses.management, PINGING_AGENT_ID)
+            .await
+            .expect("the pinging agent must reach the agent socket");
+
+    pinging_agent
+        .socket_client
+        .send(Message::Ping(Bytes::new()))
+        .await
+        .expect("the ping must be sent");
+    pinging_agent
+        .register(RegisterAgentParams {
+            name: Some(PINGING_AGENT_ID.to_owned()),
+            slot_aggregated_status_snapshot: SlotAggregatedStatusSnapshot::default(),
+        })
+        .await
+        .expect("the pinging agent must register");
+    cluster
+        .wait_for_agent_count(1)
+        .await
+        .expect("the balancer must register the pinging agent after answering its ping");
+
+    let client_management = cluster.client_management.clone();
+    let model_metadata_answer = spawn(async move {
+        client_management
+            .get_model_metadata(CancellationToken::new(), PINGING_AGENT_ID)
+            .await
+    });
+    let model_metadata_request = pinging_agent
+        .next_request()
+        .await
+        .expect("the balancer must ask the pinging agent for its model metadata");
+
+    assert_eq!(
+        discriminant(&model_metadata_request.request),
+        discriminant(&AgentJsonRpcRequest::GetModelMetadata)
+    );
+
+    pinging_agent
+        .respond(
+            model_metadata_request.id,
+            AgentJsonRpcResponse::ModelMetadata(None),
+        )
+        .await
+        .expect("the model metadata answer must be sent");
+    model_metadata_answer
+        .await
+        .expect("the model metadata request must finish")
+        .expect("the balancer must relay the model metadata answer");
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/answered_decision_probabilities.rs b/paddler_tests/tests/answered_decision_probabilities.rs
new file mode 100644
index 000000000..0eb5667d7
--- /dev/null
+++ b/paddler_tests/tests/answered_decision_probabilities.rs
@@ -0,0 +1,20 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::request_params::decide_params::raw_decide_params::RawDecideParams;
+use paddler_test_cluster_harness::cluster::Cluster;
+
+pub async fn answered_decision_probabilities(
+    cluster: &Cluster,
+    decision: &RawDecideParams,
+) -> Vec<Vec<f32>> {
+    cluster
+        .decide(CancellationToken::new(), decision)
+        .await
+        .expect("the decision must be answered")
+        .answers
+        .into_iter()
+        .map(|answer| answer.probabilities)
+        .collect()
+}
diff --git a/paddler_tests/tests/balancer_answers_bad_gateway_when_an_agent_disconnects_before_answering.rs b/paddler_tests/tests/balancer_answers_bad_gateway_when_an_agent_disconnects_before_answering.rs
deleted file mode 100644
index f0daa3ceb..000000000
--- a/paddler_tests/tests/balancer_answers_bad_gateway_when_an_agent_disconnects_before_answering.rs
+++ /dev/null
@@ -1,62 +0,0 @@
-use std::mem::discriminant;
-
-use http::StatusCode;
-use tokio::join;
-use tokio_util::sync::CancellationToken;
-
-use paddler_client::error::Error as ClientError;
-use paddler_messaging::management_socket::agent::request::Request as AgentJsonRpcRequest;
-use paddler_test_cluster_harness::cluster_params::ClusterParams;
-use paddler_test_cluster_harness::raw_agent_socket::RawAgentSocket;
-use paddler_tests::start_cluster::start_cluster;
-
-const RAW_AGENT_ID: &str = "disconnecting-agent";
-
-#[tokio::test(flavor = "multi_thread")]
-async fn balancer_answers_bad_gateway_when_an_agent_disconnects_before_answering() {
-    let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("a balancer without agents must start");
-    let mut raw_agent_socket =
-        RawAgentSocket::connect(cluster.balancer.addresses.management, RAW_AGENT_ID)
-            .await
-            .expect("the raw agent must reach the agent socket");
-
-    raw_agent_socket
-        .register()
-        .await
-        .expect("the raw agent must register");
-
-    let (metadata_result, ()) = join!(
-        cluster
-            .client_management
-            .get_model_metadata(CancellationToken::new(), RAW_AGENT_ID),
-        async move {
-            let request_envelope = raw_agent_socket
-                .next_request()
-                .await
-                .expect("the balancer must forward the metadata request to the agent");
-
-            assert_eq!(
-                discriminant(&request_envelope.request),
-                discriminant(&AgentJsonRpcRequest::GetModelMetadata)
-            );
-
-            drop(raw_agent_socket);
-        }
-    );
-
-    assert!(matches!(
-        metadata_result,
-        Err(ClientError::UnexpectedResponseStatus { status, .. }) if status == StatusCode::BAD_GATEWAY
-    ));
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_tests/tests/balancer_answers_gateway_timeout_when_an_agent_never_answers.rs b/paddler_tests/tests/balancer_answers_gateway_timeout_when_an_agent_never_answers.rs
deleted file mode 100644
index 668a0cc99..000000000
--- a/paddler_tests/tests/balancer_answers_gateway_timeout_when_an_agent_never_answers.rs
+++ /dev/null
@@ -1,58 +0,0 @@
-use std::mem::discriminant;
-
-use http::StatusCode;
-use tokio::join;
-use tokio_util::sync::CancellationToken;
-
-use paddler_client::error::Error as ClientError;
-use paddler_messaging::management_socket::agent::request::Request as AgentJsonRpcRequest;
-use paddler_test_cluster_harness::cluster_params::ClusterParams;
-use paddler_test_cluster_harness::raw_agent_socket::RawAgentSocket;
-use paddler_tests::start_cluster::start_cluster;
-
-const RAW_AGENT_ID: &str = "silent-agent";
-
-#[tokio::test(flavor = "multi_thread")]
-async fn balancer_answers_gateway_timeout_when_an_agent_never_answers() {
-    let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("a balancer without agents must start");
-    let mut raw_agent_socket =
-        RawAgentSocket::connect(cluster.balancer.addresses.management, RAW_AGENT_ID)
-            .await
-            .expect("the raw agent must reach the agent socket");
-
-    raw_agent_socket
-        .register()
-        .await
-        .expect("the raw agent must register");
-
-    let (metadata_result, unanswered_request) = join!(
-        cluster
-            .client_management
-            .get_model_metadata(CancellationToken::new(), RAW_AGENT_ID),
-        raw_agent_socket.next_request()
-    );
-
-    assert_eq!(
-        discriminant(
-            &unanswered_request
-                .expect("the balancer must forward the metadata request to the agent")
-                .request
-        ),
-        discriminant(&AgentJsonRpcRequest::GetModelMetadata)
-    );
-    assert!(matches!(
-        metadata_result,
-        Err(ClientError::UnexpectedResponseStatus { status, .. }) if status == StatusCode::GATEWAY_TIMEOUT
-    ));
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_tests/tests/balancer_closes_an_agent_socket_that_answers_before_it_registers.rs b/paddler_tests/tests/balancer_closes_an_agent_socket_that_answers_before_it_registers.rs
new file mode 100644
index 000000000..f9abc3c06
--- /dev/null
+++ b/paddler_tests/tests/balancer_closes_an_agent_socket_that_answers_before_it_registers.rs
@@ -0,0 +1,44 @@
+use tokio_tungstenite::tungstenite::protocol::frame::coding::CloseCode;
+
+use paddler_messaging::management_socket::agent::response::Response as AgentJsonRpcResponse;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_tests::start_cluster::start_cluster;
+
+use crate::untrusted_management_agent::UntrustedManagementAgent;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn balancer_closes_an_agent_socket_that_answers_before_it_registers() {
+    let cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("a cluster without agents must start");
+    let mut unregistered_agent =
+        UntrustedManagementAgent::connect(cluster.balancer.addresses.management, "unregistered")
+            .await
+            .expect("the agent must reach the agent socket");
+
+    unregistered_agent
+        .respond(
+            "never-dispatched".to_owned(),
+            AgentJsonRpcResponse::ModelMetadata(None),
+        )
+        .await
+        .expect("the response must be sent");
+
+    assert_eq!(
+        unregistered_agent
+            .socket_client
+            .next_close_frame()
+            .await
+            .expect("the balancer must close the socket")
+            .map(|close_frame| (close_frame.code, close_frame.reason.to_string())),
+        Some((
+            CloseCode::Policy,
+            "Agent sent a response before registering".to_owned()
+        ))
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/balancer_closes_an_agent_socket_that_sends_a_reserved_opcode_frame_as_a_protocol_violation.rs b/paddler_tests/tests/balancer_closes_an_agent_socket_that_sends_a_reserved_opcode_frame_as_a_protocol_violation.rs
index dd8347fd7..9b5be5a99 100644
--- a/paddler_tests/tests/balancer_closes_an_agent_socket_that_sends_a_reserved_opcode_frame_as_a_protocol_violation.rs
+++ b/paddler_tests/tests/balancer_closes_an_agent_socket_that_sends_a_reserved_opcode_frame_as_a_protocol_violation.rs
@@ -5,29 +5,27 @@ use tokio_tungstenite::tungstenite::protocol::frame::coding::Data;
 use tokio_tungstenite::tungstenite::protocol::frame::coding::OpCode;
 
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
-use paddler_test_cluster_harness::raw_agent_socket::RawAgentSocket;
 use paddler_tests::start_cluster::start_cluster;
 
+use crate::untrusted_management_agent::UntrustedManagementAgent;
+
 const FIRST_RESERVED_DATA_OPCODE: u8 = 3;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn balancer_closes_an_agent_socket_that_sends_a_reserved_opcode_frame_as_a_protocol_violation()
  {
-    let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("the cluster must start");
-    let mut agent_socket = RawAgentSocket::connect(
+    let cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("the cluster must start");
+    let mut protocol_violating_agent = UntrustedManagementAgent::connect(
         cluster.balancer.addresses.management,
         "protocol-violating-agent",
     )
     .await
     .expect("the connection must be established");
 
-    agent_socket
+    protocol_violating_agent
+        .socket_client
         .send(Message::Frame(Frame::message(
             "payload",
             OpCode::Data(Data::Reserved(FIRST_RESERVED_DATA_OPCODE)),
@@ -37,7 +35,8 @@ async fn balancer_closes_an_agent_socket_that_sends_a_reserved_opcode_frame_as_a
         .expect("the frame must be sent");
 
     assert_eq!(
-        agent_socket
+        protocol_violating_agent
+            .socket_client
             .next_close_frame()
             .await
             .expect("the balancer must close the socket with a valid close frame")
diff --git a/paddler_tests/tests/balancer_closes_an_agent_socket_that_sends_an_undeserializable_message.rs b/paddler_tests/tests/balancer_closes_an_agent_socket_that_sends_an_undeserializable_message.rs
index 261b9e497..8c0ce47a7 100644
--- a/paddler_tests/tests/balancer_closes_an_agent_socket_that_sends_an_undeserializable_message.rs
+++ b/paddler_tests/tests/balancer_closes_an_agent_socket_that_sends_an_undeserializable_message.rs
@@ -2,32 +2,31 @@ use tokio_tungstenite::tungstenite::Message;
 use tokio_tungstenite::tungstenite::protocol::frame::coding::CloseCode;
 
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
-use paddler_test_cluster_harness::raw_agent_socket::RawAgentSocket;
 use paddler_tests::start_cluster::start_cluster;
 
+use crate::untrusted_management_agent::UntrustedManagementAgent;
+
 #[tokio::test(flavor = "multi_thread")]
 async fn balancer_closes_an_agent_socket_that_sends_an_undeserializable_message() {
-    let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("the cluster must start");
-    let mut agent_socket = RawAgentSocket::connect(
+    let cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("the cluster must start");
+    let mut undeserializable_agent = UntrustedManagementAgent::connect(
         cluster.balancer.addresses.management,
         "undeserializable-agent",
     )
     .await
     .expect("the connection must be established");
 
-    agent_socket
+    undeserializable_agent
+        .socket_client
         .send(Message::text("not json"))
         .await
         .expect("the message must be sent");
 
     assert_eq!(
-        agent_socket
+        undeserializable_agent
+            .socket_client
             .next_close_frame()
             .await
             .expect("the balancer must close the socket")
diff --git a/paddler_tests/tests/balancer_closes_an_agent_socket_with_a_valid_close_frame_after_an_unknown_token_result.rs b/paddler_tests/tests/balancer_closes_an_agent_socket_with_a_valid_close_frame_after_an_unknown_token_result.rs
deleted file mode 100644
index 64f0de0f3..000000000
--- a/paddler_tests/tests/balancer_closes_an_agent_socket_with_a_valid_close_frame_after_an_unknown_token_result.rs
+++ /dev/null
@@ -1,54 +0,0 @@
-use serde_json::json;
-use tokio_tungstenite::tungstenite::Message;
-use tokio_tungstenite::tungstenite::protocol::frame::coding::CloseCode;
-
-use paddler_test_cluster_harness::cluster_params::ClusterParams;
-use paddler_test_cluster_harness::raw_agent_socket::RawAgentSocket;
-use paddler_tests::start_cluster::start_cluster;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn balancer_closes_an_agent_socket_with_a_valid_close_frame_after_an_unknown_token_result() {
-    let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("the cluster must start");
-    let mut agent_socket =
-        RawAgentSocket::connect(cluster.balancer.addresses.management, "newer-agent")
-            .await
-            .expect("the connection must be established");
-
-    agent_socket
-        .send(Message::text(
-            json!({
-                "Response": {
-                    "generated_by": null,
-                    "request_id": "request-answered-by-a-newer-agent",
-                    "response": {
-                        "GeneratedToken": {
-                            "TokenResultFromANewerAgent": "piece"
-                        }
-                    }
-                }
-            })
-            .to_string(),
-        ))
-        .await
-        .expect("the message must be sent");
-
-    assert_eq!(
-        agent_socket
-            .next_close_frame()
-            .await
-            .expect("the balancer must close the socket with a valid close frame")
-            .map(|close_frame| close_frame.code),
-        Some(CloseCode::Invalid)
-    );
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_tests/tests/balancer_completes_buffered_request_after_agent_joins.rs b/paddler_tests/tests/balancer_completes_buffered_request_after_agent_joins.rs
index e27063895..71c15ad2e 100644
--- a/paddler_tests/tests/balancer_completes_buffered_request_after_agent_joins.rs
+++ b/paddler_tests/tests/balancer_completes_buffered_request_after_agent_joins.rs
@@ -2,29 +2,27 @@
 
 use std::num::NonZeroU32;
 
-use futures_util::StreamExt as _;
 use tokio_util::sync::CancellationToken;
 
-use paddler_messaging::inference_client::message::Message;
 use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_test_cluster_harness::collect_generated_tokens::collect_generated_tokens;
 use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
 use paddler_tests::start_cluster::start_cluster;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn balancer_completes_buffered_request_after_agent_joins() {
     let mut cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
         max_buffered_requests: 1,
-        desired_state: Some(qwen3_0_6b().into_desired_state()),
-        ..ClusterParams::default()
+        desired_state: ClusterDesiredState::Apply(Box::new(qwen3_0_6b().into_desired_state())),
+        ..ClusterParams::without_agents()
     })
     .await
     .expect("the cluster must start");
 
-    let mut stream = cluster
+    let stream = cluster
         .continue_from_raw_prompt_stream(
             CancellationToken::new(),
             &ContinueFromRawPromptParams {
@@ -48,24 +46,11 @@ async fn balancer_completes_buffered_request_after_agent_joins() {
         })
         .expect("the additional agent must start");
 
-    let message = stream
-        .next()
+    collect_generated_tokens(stream)
         .await
-        .expect("inference stream must yield a message after agent joins")
-        .expect("the message must be readable");
-
-    match message {
-        Message::Response(_) => {}
-        Message::Error(envelope) => {
-            panic!(
-                "expected a successful response, got error code {}: {}",
-                envelope.error.code, envelope.error.description
-            );
-        }
-        Message::Notification(_) => {
-            panic!("unexpected token-generation-mode notification");
-        }
-    }
+        .expect("the buffered request must be served once the agent joins")
+        .summary()
+        .expect("the buffered request must finish with a summary");
 
     cluster
         .shutdown()
diff --git a/paddler_tests/tests/balancer_ends_an_in_flight_generation_with_a_shutdown_error.rs b/paddler_tests/tests/balancer_ends_an_in_flight_generation_with_a_shutdown_error.rs
index 534c23492..cb51f6423 100644
--- a/paddler_tests/tests/balancer_ends_an_in_flight_generation_with_a_shutdown_error.rs
+++ b/paddler_tests/tests/balancer_ends_an_in_flight_generation_with_a_shutdown_error.rs
@@ -4,9 +4,9 @@ use futures_util::StreamExt as _;
 use tokio::spawn;
 use tokio_util::sync::CancellationToken;
 
-use paddler_messaging::inference_client::message::Message as InferenceClientMessage;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
 use paddler_test_cluster_harness::cluster::Cluster;
+use paddler_test_cluster_harness::inference_stream_error_codes::inference_stream_error_codes;
 use paddler_test_cluster_harness::unending_generation::unending_generation;
 use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
 
@@ -31,17 +31,13 @@ async fn balancer_ends_an_in_flight_generation_with_a_shutdown_error() {
         agents, balancer, ..
     } = cluster;
     let balancer_shutdown = spawn(balancer.shutdown());
-    let mut shutdown_error_codes = Vec::new();
-
-    while let Some(message) = stream.next().await {
-        if let InferenceClientMessage::Error(envelope) =
-            message.expect("every streamed message must be readable")
-        {
-            shutdown_error_codes.push(envelope.error.code);
-        }
-    }
 
-    assert_eq!(shutdown_error_codes, vec![503]);
+    assert_eq!(
+        inference_stream_error_codes(stream)
+            .await
+            .expect("every streamed message must be readable"),
+        vec![503]
+    );
 
     balancer_shutdown
         .await
diff --git a/paddler_tests/tests/balancer_inference_health_returns_ok.rs b/paddler_tests/tests/balancer_inference_health_returns_ok.rs
index 0f0815080..56c7d8bfa 100644
--- a/paddler_tests/tests/balancer_inference_health_returns_ok.rs
+++ b/paddler_tests/tests/balancer_inference_health_returns_ok.rs
@@ -6,13 +6,9 @@ use paddler_tests::start_cluster::start_cluster;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn balancer_inference_health_returns_ok() {
-    let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("the cluster must start");
+    let cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("the cluster must start");
 
     let health = cluster
         .client_inference
diff --git a/paddler_tests/tests/balancer_inference_service_replies_with_configured_cors_origin.rs b/paddler_tests/tests/balancer_inference_service_replies_with_configured_cors_origin.rs
index 7ad38e485..18b2981cc 100644
--- a/paddler_tests/tests/balancer_inference_service_replies_with_configured_cors_origin.rs
+++ b/paddler_tests/tests/balancer_inference_service_replies_with_configured_cors_origin.rs
@@ -7,17 +7,14 @@ const ALLOWED_ORIGIN: &str = "http://example.com";
 #[tokio::test(flavor = "multi_thread")]
 async fn balancer_inference_service_replies_with_configured_cors_origin() {
     let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
         inference_cors_allowed_hosts: vec![ALLOWED_ORIGIN.to_owned()],
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
+        ..ClusterParams::without_agents()
     })
     .await
     .expect("a cluster without agents must start");
     let health_url = cluster
         .balancer
-        .inference_base_url()
-        .expect("the inference service must have a base URL")
+        .inference_base_url
         .join("health")
         .expect("the health path must join onto the base URL");
 
diff --git a/paddler_tests/tests/balancer_inference_socket_recovers_from_embeddings_mode.rs b/paddler_tests/tests/balancer_inference_socket_recovers_from_embeddings_mode.rs
deleted file mode 100644
index 785cf45d2..000000000
--- a/paddler_tests/tests/balancer_inference_socket_recovers_from_embeddings_mode.rs
+++ /dev/null
@@ -1,134 +0,0 @@
-#![cfg(feature = "tests_that_use_llms")]
-
-use std::num::NonZeroU32;
-
-use futures_util::StreamExt as _;
-use tokio_util::sync::CancellationToken;
-
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
-use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-use paddler_messaging::generated_token_result::GeneratedTokenResult;
-use paddler_messaging::inference_client::message::Message as InferenceMessage;
-use paddler_messaging::inference_client::notification::Notification;
-use paddler_messaging::inference_client::response::Response as InferenceResponse;
-use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
-use paddler_test_cluster_harness::agent_config::AgentConfig;
-use paddler_test_cluster_harness::cluster_params::ClusterParams;
-use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
-use paddler_tests::start_cluster::start_cluster;
-
-const MAX_TOKENS: NonZeroU32 = NonZeroU32::new(16).unwrap();
-
-fn capital_of_france_prompt() -> ContinueFromRawPromptParams {
-    ContinueFromRawPromptParams {
-        grammar: None,
-        max_tokens: MAX_TOKENS,
-        raw_prompt: "The capital of France is".to_owned(),
-    }
-}
-
-#[tokio::test(flavor = "multi_thread")]
-async fn balancer_inference_socket_recovers_from_embeddings_mode() {
-    let generation_state = qwen3_0_6b().into_desired_state();
-    let embeddings_state = BalancerDesiredState {
-        inference_parameters: InferenceParameters {
-            enable_embeddings: true,
-            ..generation_state.inference_parameters.clone()
-        },
-        ..generation_state.clone()
-    };
-
-    let cluster = start_cluster(ClusterParams {
-        agents: AgentConfig::uniform(1, 1),
-        desired_state: Some(embeddings_state),
-        wait_for_slots_ready: true,
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("a cluster serving embeddings must start");
-
-    let inference = &cluster.client_inference;
-    let mut token_generation_mode_rx = inference.subscribe_to_token_generation_mode();
-
-    let mut disabled_stream = inference
-        .continue_from_raw_prompt(CancellationToken::new(), capital_of_france_prompt())
-        .await
-        .expect("the inference socket must accept a request in embeddings mode");
-
-    let disabled_message = disabled_stream
-        .next()
-        .await
-        .expect("inference socket must answer instead of rejecting in embeddings mode")
-        .expect("the embeddings-mode answer must be readable");
-
-    match disabled_message {
-        InferenceMessage::Response(envelope) => match envelope.response {
-            InferenceResponse::GeneratedToken(GeneratedTokenResult::TokenGenerationDisabled(_)) => {
-            }
-            other => panic!("expected a token-generation-disabled reply, got {other:?}"),
-        },
-        other => panic!("expected a token-generation-disabled reply, got {other:?}"),
-    }
-
-    let connect_notification = token_generation_mode_rx
-        .recv()
-        .await
-        .expect("client must be told on connect that token generation is disabled");
-
-    assert!(matches!(
-        connect_notification,
-        Notification::TokenGenerationDisabled
-    ));
-
-    cluster
-        .client_management
-        .put_balancer_desired_state(CancellationToken::new(), &generation_state)
-        .await
-        .expect("the balancer must accept the token generation state");
-
-    let recovery_notification = token_generation_mode_rx.recv().await.expect(
-        "client must be told over the open connection that token generation is enabled again",
-    );
-
-    assert!(matches!(
-        recovery_notification,
-        Notification::TokenGenerationEnabled
-    ));
-
-    let mut recovered_stream = inference
-        .continue_from_raw_prompt(CancellationToken::new(), capital_of_france_prompt())
-        .await
-        .expect("the recovered inference socket must accept a request");
-
-    let mut generated_token_count: usize = 0;
-
-    while let Some(message_result) = recovered_stream.next().await {
-        match message_result.expect("the recovered stream must stay readable") {
-            InferenceMessage::Response(envelope) => match envelope.response {
-                InferenceResponse::GeneratedToken(token_result) => {
-                    if token_result.is_token() {
-                        generated_token_count += 1;
-                    }
-                }
-                other @ InferenceResponse::Embedding(_) => {
-                    panic!("unexpected response after recovery: {other:?}")
-                }
-            },
-            InferenceMessage::Error(envelope) => panic!(
-                "recovered inference failed: code {}, description {:?}",
-                envelope.error.code, envelope.error.description
-            ),
-            InferenceMessage::Notification(_) => {}
-        }
-    }
-
-    assert!(
-        generated_token_count > 0,
-        "the recovered connection must stream tokens once token generation is enabled again"
-    );
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_tests/tests/balancer_keeps_an_agent_connected_after_it_answers_an_unknown_request.rs b/paddler_tests/tests/balancer_keeps_an_agent_connected_after_it_answers_an_unknown_request.rs
deleted file mode 100644
index 043d80f3d..000000000
--- a/paddler_tests/tests/balancer_keeps_an_agent_connected_after_it_answers_an_unknown_request.rs
+++ /dev/null
@@ -1,78 +0,0 @@
-use std::collections::BTreeMap;
-
-use tokio::join;
-use tokio_util::sync::CancellationToken;
-
-use paddler_messaging::jsonrpc::response_envelope::ResponseEnvelope;
-use paddler_messaging::management_socket::agent::response::Response as AgentJsonRpcResponse;
-use paddler_messaging::model_metadata::ModelMetadata;
-use paddler_test_cluster_harness::cluster_params::ClusterParams;
-use paddler_test_cluster_harness::raw_agent_socket::RawAgentSocket;
-use paddler_tests::start_cluster::start_cluster;
-
-const RAW_AGENT_ID: &str = "confused-agent";
-
-#[tokio::test(flavor = "multi_thread")]
-async fn balancer_keeps_an_agent_connected_after_it_answers_an_unknown_request() {
-    let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("a balancer without agents must start");
-    let mut raw_agent_socket =
-        RawAgentSocket::connect(cluster.balancer.addresses.management, RAW_AGENT_ID)
-            .await
-            .expect("the raw agent must reach the agent socket");
-    let answered_metadata =
-        BTreeMap::from([("general.architecture".to_owned(), "qwen3".to_owned())]);
-
-    raw_agent_socket
-        .register()
-        .await
-        .expect("the raw agent must register");
-    raw_agent_socket
-        .send_response(ResponseEnvelope {
-            generated_by: None,
-            request_id: "a-request-the-balancer-never-sent".to_owned(),
-            response: AgentJsonRpcResponse::ModelMetadata(None),
-        })
-        .await
-        .expect("the raw agent must send its answer to an unknown request");
-
-    let (metadata_result, ()) = join!(
-        cluster
-            .client_management
-            .get_model_metadata(CancellationToken::new(), RAW_AGENT_ID),
-        async {
-            let request_envelope = raw_agent_socket
-                .next_request()
-                .await
-                .expect("the balancer must forward the metadata request to the agent");
-
-            raw_agent_socket
-                .send_response(ResponseEnvelope {
-                    generated_by: None,
-                    request_id: request_envelope.id,
-                    response: AgentJsonRpcResponse::ModelMetadata(Some(ModelMetadata {
-                        metadata: answered_metadata.clone(),
-                    })),
-                })
-                .await
-                .expect("the raw agent must answer the metadata request");
-        }
-    );
-
-    assert_eq!(
-        metadata_result
-            .expect("the balancer must relay the agent's answer")
-            .map(|model_metadata| model_metadata.metadata),
-        Some(answered_metadata)
-    );
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_tests/tests/balancer_keeps_an_agent_that_repeats_its_status_registered.rs b/paddler_tests/tests/balancer_keeps_an_agent_that_repeats_its_status_registered.rs
new file mode 100644
index 000000000..3de765641
--- /dev/null
+++ b/paddler_tests/tests/balancer_keeps_an_agent_that_repeats_its_status_registered.rs
@@ -0,0 +1,52 @@
+use paddler_messaging::agent_status::AgentStatus;
+use paddler_messaging::slot_aggregated_status_snapshot::SlotAggregatedStatusSnapshot;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_tests::start_cluster::start_cluster;
+
+use crate::untrusted_management_agent::UntrustedManagementAgent;
+
+const REPEATING_AGENT_ID: &str = "repeating-agent";
+
+#[tokio::test(flavor = "multi_thread")]
+async fn balancer_keeps_an_agent_that_repeats_its_status_registered() {
+    let mut cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("a balancer without agents must start");
+    let status_report = |desired_slots_total: u16, version: u64| SlotAggregatedStatusSnapshot {
+        status: AgentStatus {
+            desired_slots_total,
+            ..AgentStatus::default()
+        },
+        version,
+    };
+    let mut repeating_agent = UntrustedManagementAgent::connect_registered(
+        cluster.balancer.addresses.management,
+        REPEATING_AGENT_ID,
+        status_report(1, 1),
+    )
+    .await
+    .expect("the repeating agent must register");
+    repeating_agent
+        .report_status(status_report(1, 2))
+        .await
+        .expect("the repeated status must be sent");
+    repeating_agent
+        .report_status(status_report(2, 3))
+        .await
+        .expect("the changed status must be sent");
+
+    cluster
+        .agents_watcher
+        .until(|snapshot| {
+            snapshot.agents.iter().any(|agent| {
+                agent.id == REPEATING_AGENT_ID && agent.status.desired_slots_total == 2
+            })
+        })
+        .await
+        .expect("the balancer must keep the agent and take its changed status");
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/balancer_management_service_replies_with_configured_cors_origin.rs b/paddler_tests/tests/balancer_management_service_replies_with_configured_cors_origin.rs
index 8f2a75147..582c87fd3 100644
--- a/paddler_tests/tests/balancer_management_service_replies_with_configured_cors_origin.rs
+++ b/paddler_tests/tests/balancer_management_service_replies_with_configured_cors_origin.rs
@@ -7,17 +7,14 @@ const ALLOWED_ORIGIN: &str = "http://example.com";
 #[tokio::test(flavor = "multi_thread")]
 async fn balancer_management_service_replies_with_configured_cors_origin() {
     let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
         management_cors_allowed_hosts: vec![ALLOWED_ORIGIN.to_owned()],
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
+        ..ClusterParams::without_agents()
     })
     .await
     .expect("a cluster without agents must start");
     let health_url = cluster
         .balancer
-        .management_base_url()
-        .expect("the management service must have a base URL")
+        .management_base_url
         .join("health")
         .expect("the health path must join onto the base URL");
 
diff --git a/paddler_tests/tests/balancer_memory_storage_persists_desired_state.rs b/paddler_tests/tests/balancer_memory_storage_persists_desired_state.rs
deleted file mode 100644
index d2284f4d0..000000000
--- a/paddler_tests/tests/balancer_memory_storage_persists_desired_state.rs
+++ /dev/null
@@ -1,45 +0,0 @@
-use tokio_util::sync::CancellationToken;
-
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
-use paddler_messaging::agent_desired_model::AgentDesiredModel;
-use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-use paddler_test_cluster_harness::cluster_params::ClusterParams;
-use paddler_test_cluster_harness::model_card::ModelCard;
-use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
-use paddler_tests::start_cluster::start_cluster;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn balancer_memory_storage_persists_desired_state() {
-    let ModelCard { reference } = qwen3_0_6b();
-
-    let desired_state = BalancerDesiredState {
-        chat_template_override: None,
-        inference_parameters: InferenceParameters::default(),
-        model: AgentDesiredModel::HuggingFace(reference),
-        multimodal_projection: AgentDesiredModel::None,
-        use_chat_template_override: false,
-    };
-
-    let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
-        state_database_url: "memory://".to_owned(),
-        desired_state: Some(desired_state.clone()),
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("the cluster must start");
-
-    let observed = cluster
-        .client_management
-        .get_balancer_desired_state(CancellationToken::new())
-        .await
-        .expect("failed to read desired state from memory backend");
-
-    assert_eq!(observed.model, desired_state.model);
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_tests/tests/balancer_memory_storage_reports_the_desired_state_it_started_with.rs b/paddler_tests/tests/balancer_memory_storage_reports_the_desired_state_it_started_with.rs
new file mode 100644
index 000000000..fe0ce81f6
--- /dev/null
+++ b/paddler_tests/tests/balancer_memory_storage_reports_the_desired_state_it_started_with.rs
@@ -0,0 +1,40 @@
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_model_source::model_source::ModelSource;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_test_cluster_harness::model_card::ModelCard;
+use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
+use paddler_tests::start_cluster::start_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn balancer_memory_storage_reports_the_desired_state_it_started_with() {
+    let ModelCard { reference } = qwen3_0_6b();
+
+    let desired_state = BalancerDesiredState {
+        model: ModelSource::HuggingFace(reference).into_agent_desired_model(),
+        ..BalancerDesiredState::default()
+    };
+
+    let cluster = start_cluster(ClusterParams {
+        state_database_url: "memory://".to_owned(),
+        desired_state: ClusterDesiredState::Apply(Box::new(desired_state.clone())),
+        ..ClusterParams::without_agents()
+    })
+    .await
+    .expect("the cluster must start");
+
+    let observed = cluster
+        .client_management
+        .get_balancer_desired_state(CancellationToken::new())
+        .await
+        .expect("failed to read desired state from memory backend");
+
+    assert_eq!(observed.model, desired_state.model);
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/balancer_openai_compat_health_returns_ok.rs b/paddler_tests/tests/balancer_openai_compat_health_returns_ok.rs
deleted file mode 100644
index 008c76246..000000000
--- a/paddler_tests/tests/balancer_openai_compat_health_returns_ok.rs
+++ /dev/null
@@ -1,29 +0,0 @@
-use tokio_util::sync::CancellationToken;
-
-use paddler_client::reports_health::ReportsHealth as _;
-use paddler_test_cluster_harness::cluster_params::ClusterParams;
-use paddler_tests::start_cluster::start_cluster;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn balancer_openai_compat_health_returns_ok() {
-    let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("the cluster must start");
-
-    let health = cluster
-        .client_compat_openai_health
-        .get_health(CancellationToken::new())
-        .await
-        .expect("failed to GET OpenAI compat /health");
-
-    assert_eq!(health, "OK");
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_tests/tests/balancer_persists_chat_template_override_across_restart.rs b/paddler_tests/tests/balancer_persists_chat_template_override_across_restart.rs
index 08df03927..6206d9ebf 100644
--- a/paddler_tests/tests/balancer_persists_chat_template_override_across_restart.rs
+++ b/paddler_tests/tests/balancer_persists_chat_template_override_across_restart.rs
@@ -1,39 +1,37 @@
 use tokio_util::sync::CancellationToken;
 
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
-use paddler_messaging::agent_desired_model::AgentDesiredModel;
 use paddler_messaging::balancer_desired_state::BalancerDesiredState;
 use paddler_messaging::chat_template::ChatTemplate;
+use paddler_model_source::model_source::ModelSource;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
 use paddler_test_cluster_harness::model_card::ModelCard;
 use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
-use paddler_test_cluster_harness::state_database_file::StateDatabaseFile;
+use paddler_test_cluster_harness::temporary_state_database_file::TemporaryStateDatabaseFile;
+use paddler_tests::desired_state_with_chat_template_override::desired_state_with_chat_template_override;
 use paddler_tests::start_cluster::start_cluster;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn balancer_persists_chat_template_override_across_restart() {
-    let database = StateDatabaseFile::new().expect("the state database file must be created");
+    let database =
+        TemporaryStateDatabaseFile::new().expect("the state database file must be created");
 
     let ModelCard { reference } = qwen3_0_6b();
 
-    let template_content = "{{ messages | tojson }}".to_owned();
-
-    let desired_state = BalancerDesiredState {
-        chat_template_override: Some(ChatTemplate {
-            content: template_content.clone(),
-        }),
-        inference_parameters: InferenceParameters::default(),
-        model: AgentDesiredModel::HuggingFace(reference),
-        multimodal_projection: AgentDesiredModel::None,
-        use_chat_template_override: true,
-    };
+    let desired_state = desired_state_with_chat_template_override(
+        BalancerDesiredState {
+            model: ModelSource::HuggingFace(reference).into_agent_desired_model(),
+            ..BalancerDesiredState::default()
+        },
+        ChatTemplate {
+            content: "{{ messages | tojson }}".to_owned(),
+        },
+    );
 
     let first_cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
         state_database_url: database.url.clone(),
-        desired_state: Some(desired_state.clone()),
-        ..ClusterParams::default()
+        desired_state: ClusterDesiredState::Apply(Box::new(desired_state.clone())),
+        ..ClusterParams::without_agents()
     })
     .await
     .expect("the cluster must start");
@@ -44,11 +42,9 @@ async fn balancer_persists_chat_template_override_across_restart() {
         .expect("the cluster must shut down cleanly");
 
     let second_cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
         state_database_url: database.url.clone(),
-        desired_state: None,
-        ..ClusterParams::default()
+        desired_state: ClusterDesiredState::KeepStored,
+        ..ClusterParams::without_agents()
     })
     .await
     .expect("the cluster must start");
@@ -59,14 +55,7 @@ async fn balancer_persists_chat_template_override_across_restart() {
         .await
         .expect("failed to read restored desired state");
 
-    assert!(restored_state.use_chat_template_override);
-    assert_eq!(
-        restored_state
-            .chat_template_override
-            .as_ref()
-            .map(|template| template.content.as_str()),
-        Some(template_content.as_str())
-    );
+    assert_eq!(restored_state, desired_state);
 
     second_cluster
         .shutdown()
diff --git a/paddler_tests/tests/balancer_persists_desired_state_across_restart.rs b/paddler_tests/tests/balancer_persists_desired_state_across_restart.rs
index 4f1903399..98a3e6588 100644
--- a/paddler_tests/tests/balancer_persists_desired_state_across_restart.rs
+++ b/paddler_tests/tests/balancer_persists_desired_state_across_restart.rs
@@ -1,34 +1,30 @@
 use tokio_util::sync::CancellationToken;
 
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
-use paddler_messaging::agent_desired_model::AgentDesiredModel;
 use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_model_source::model_source::ModelSource;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
 use paddler_test_cluster_harness::model_card::ModelCard;
 use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
-use paddler_test_cluster_harness::state_database_file::StateDatabaseFile;
+use paddler_test_cluster_harness::temporary_state_database_file::TemporaryStateDatabaseFile;
 use paddler_tests::start_cluster::start_cluster;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn balancer_persists_desired_state_across_restart() {
-    let database = StateDatabaseFile::new().expect("the state database file must be created");
+    let database =
+        TemporaryStateDatabaseFile::new().expect("the state database file must be created");
 
     let ModelCard { reference } = qwen3_0_6b();
 
     let desired_state = BalancerDesiredState {
-        chat_template_override: None,
-        inference_parameters: InferenceParameters::default(),
-        model: AgentDesiredModel::HuggingFace(reference),
-        multimodal_projection: AgentDesiredModel::None,
-        use_chat_template_override: false,
+        model: ModelSource::HuggingFace(reference).into_agent_desired_model(),
+        ..BalancerDesiredState::default()
     };
 
     let first_cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
         state_database_url: database.url.clone(),
-        desired_state: Some(desired_state.clone()),
-        ..ClusterParams::default()
+        desired_state: ClusterDesiredState::Apply(Box::new(desired_state.clone())),
+        ..ClusterParams::without_agents()
     })
     .await
     .expect("the cluster must start");
@@ -39,11 +35,9 @@ async fn balancer_persists_desired_state_across_restart() {
         .expect("the cluster must shut down cleanly");
 
     let second_cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
         state_database_url: database.url.clone(),
-        desired_state: None,
-        ..ClusterParams::default()
+        desired_state: ClusterDesiredState::KeepStored,
+        ..ClusterParams::without_agents()
     })
     .await
     .expect("the cluster must start");
diff --git a/paddler_tests/tests/balancer_persists_huggingface_mmproj_in_desired_state.rs b/paddler_tests/tests/balancer_persists_huggingface_mmproj_in_desired_state.rs
deleted file mode 100644
index 3a5a21e89..000000000
--- a/paddler_tests/tests/balancer_persists_huggingface_mmproj_in_desired_state.rs
+++ /dev/null
@@ -1,51 +0,0 @@
-use tokio_util::sync::CancellationToken;
-
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
-use paddler_messaging::agent_desired_model::AgentDesiredModel;
-use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-use paddler_test_cluster_harness::cluster_params::ClusterParams;
-use paddler_test_cluster_harness::model_card::ModelCard;
-use paddler_test_cluster_harness::model_card::smolvlm2_256m::smolvlm2_256m;
-use paddler_test_cluster_harness::model_card::smolvlm2_256m_mmproj::smolvlm2_256m_mmproj;
-use paddler_tests::start_cluster::start_cluster;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn balancer_persists_huggingface_mmproj_in_desired_state() {
-    let ModelCard {
-        reference: primary_reference,
-    } = smolvlm2_256m();
-    let ModelCard {
-        reference: mmproj_reference,
-    } = smolvlm2_256m_mmproj();
-
-    let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
-        desired_state: Some(BalancerDesiredState {
-            chat_template_override: None,
-            inference_parameters: InferenceParameters::default(),
-            model: AgentDesiredModel::HuggingFace(primary_reference),
-            multimodal_projection: AgentDesiredModel::HuggingFace(mmproj_reference.clone()),
-            use_chat_template_override: false,
-        }),
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("the cluster must start");
-
-    let retrieved = cluster
-        .client_management
-        .get_balancer_desired_state(CancellationToken::new())
-        .await
-        .expect("failed to read balancer desired state");
-
-    assert_eq!(
-        retrieved.multimodal_projection,
-        AgentDesiredModel::HuggingFace(mmproj_reference)
-    );
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_tests/tests/balancer_persists_local_mmproj_path_in_desired_state.rs b/paddler_tests/tests/balancer_persists_local_mmproj_path_in_desired_state.rs
deleted file mode 100644
index 9caaca611..000000000
--- a/paddler_tests/tests/balancer_persists_local_mmproj_path_in_desired_state.rs
+++ /dev/null
@@ -1,47 +0,0 @@
-use tokio_util::sync::CancellationToken;
-
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
-use paddler_messaging::agent_desired_model::AgentDesiredModel;
-use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-use paddler_test_cluster_harness::cluster_params::ClusterParams;
-use paddler_test_cluster_harness::model_card::ModelCard;
-use paddler_test_cluster_harness::model_card::smolvlm2_256m::smolvlm2_256m;
-use paddler_tests::start_cluster::start_cluster;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn balancer_persists_local_mmproj_path_in_desired_state() {
-    let ModelCard { reference } = smolvlm2_256m();
-
-    let local_mmproj_path = "/tmp/test-mmproj.gguf".to_owned();
-
-    let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
-        desired_state: Some(BalancerDesiredState {
-            chat_template_override: None,
-            inference_parameters: InferenceParameters::default(),
-            model: AgentDesiredModel::HuggingFace(reference),
-            multimodal_projection: AgentDesiredModel::LocalToAgent(local_mmproj_path.clone()),
-            use_chat_template_override: false,
-        }),
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("the cluster must start");
-
-    let retrieved = cluster
-        .client_management
-        .get_balancer_desired_state(CancellationToken::new())
-        .await
-        .expect("failed to read balancer desired state");
-
-    assert_eq!(
-        retrieved.multimodal_projection,
-        AgentDesiredModel::LocalToAgent(local_mmproj_path)
-    );
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_tests/tests/balancer_persists_model_switch_in_storage.rs b/paddler_tests/tests/balancer_persists_model_switch_in_storage.rs
index e054c9bce..ffec49fed 100644
--- a/paddler_tests/tests/balancer_persists_model_switch_in_storage.rs
+++ b/paddler_tests/tests/balancer_persists_model_switch_in_storage.rs
@@ -1,69 +1,72 @@
+use std::path::Path;
+
 use tokio_util::sync::CancellationToken;
 
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
-use paddler_messaging::agent_desired_model::AgentDesiredModel;
 use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_model_source::model_source::ModelSource;
+use paddler_test_cluster_harness::agent_desired_model_at::agent_desired_model_at;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
 use paddler_test_cluster_harness::model_card::ModelCard;
 use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
-use paddler_test_cluster_harness::state_database_file::StateDatabaseFile;
+use paddler_test_cluster_harness::temporary_state_database_file::TemporaryStateDatabaseFile;
 use paddler_tests::start_cluster::start_cluster;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn balancer_persists_model_switch_in_storage() {
-    let database = StateDatabaseFile::new().expect("the state database file must be created");
+    let database =
+        TemporaryStateDatabaseFile::new().expect("the state database file must be created");
 
     let ModelCard { reference } = qwen3_0_6b();
 
     let initial_state = BalancerDesiredState {
-        chat_template_override: None,
-        inference_parameters: InferenceParameters::default(),
-        model: AgentDesiredModel::HuggingFace(reference),
-        multimodal_projection: AgentDesiredModel::None,
-        use_chat_template_override: false,
+        model: ModelSource::HuggingFace(reference).into_agent_desired_model(),
+        ..BalancerDesiredState::default()
     };
 
-    let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
+    let first_cluster = start_cluster(ClusterParams {
         state_database_url: database.url.clone(),
-        desired_state: Some(initial_state.clone()),
-        ..ClusterParams::default()
+        desired_state: ClusterDesiredState::Apply(Box::new(initial_state)),
+        ..ClusterParams::without_agents()
     })
     .await
     .expect("the cluster must start");
 
-    let observed_initial = cluster
-        .client_management
-        .get_balancer_desired_state(CancellationToken::new())
-        .await
-        .expect("failed to read initial desired state");
-
-    assert_eq!(observed_initial.model, initial_state.model);
-
     let switched_state = BalancerDesiredState {
-        chat_template_override: None,
-        inference_parameters: InferenceParameters::default(),
-        model: AgentDesiredModel::LocalToAgent("/tmp/alternative-model.gguf".to_owned()),
-        multimodal_projection: AgentDesiredModel::None,
-        use_chat_template_override: false,
+        model: agent_desired_model_at(Path::new("/tmp/alternative-model.gguf"))
+            .expect("the alternative model path must be absolute"),
+        ..BalancerDesiredState::default()
     };
 
-    cluster
+    first_cluster
         .client_management
         .put_balancer_desired_state(CancellationToken::new(), &switched_state)
         .await
         .expect("failed to switch desired model");
-
-    let observed_switched = cluster
-        .client_management
-        .get_balancer_desired_state(CancellationToken::new())
+    first_cluster
+        .shutdown()
         .await
-        .expect("failed to read switched desired state");
+        .expect("the cluster must shut down cleanly");
+
+    let second_cluster = start_cluster(ClusterParams {
+        state_database_url: database.url.clone(),
+        desired_state: ClusterDesiredState::KeepStored,
+        ..ClusterParams::without_agents()
+    })
+    .await
+    .expect("the cluster must restart from the stored desired state");
 
-    assert_eq!(observed_switched.model, switched_state.model);
+    assert_eq!(
+        second_cluster
+            .client_management
+            .get_balancer_desired_state(CancellationToken::new())
+            .await
+            .expect("failed to read the restored desired state")
+            .model,
+        switched_state.model
+    );
 
-    cluster
+    second_cluster
         .shutdown()
         .await
         .expect("the cluster must shut down cleanly");
diff --git a/paddler_tests/tests/balancer_persists_url_model_in_desired_state.rs b/paddler_tests/tests/balancer_persists_url_model_in_desired_state.rs
deleted file mode 100644
index 168e77258..000000000
--- a/paddler_tests/tests/balancer_persists_url_model_in_desired_state.rs
+++ /dev/null
@@ -1,48 +0,0 @@
-use tokio_util::sync::CancellationToken;
-
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
-use paddler_messaging::agent_desired_model::AgentDesiredModel;
-use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-use paddler_messaging::url_model_reference::UrlModelReference;
-use paddler_test_cluster_harness::cluster_params::ClusterParams;
-use paddler_tests::start_cluster::start_cluster;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn balancer_persists_url_model_in_desired_state() {
-    let configured_url = "https://example.invalid/persisted-model.gguf".to_owned();
-
-    let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
-        desired_state: Some(BalancerDesiredState {
-            chat_template_override: None,
-            inference_parameters: InferenceParameters::default(),
-            model: AgentDesiredModel::Url(UrlModelReference {
-                url: configured_url.clone(),
-            }),
-            multimodal_projection: AgentDesiredModel::None,
-            use_chat_template_override: false,
-        }),
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("the cluster must start");
-
-    let retrieved = cluster
-        .client_management
-        .get_balancer_desired_state(CancellationToken::new())
-        .await
-        .expect("failed to read balancer desired state");
-
-    assert_eq!(
-        retrieved.model,
-        AgentDesiredModel::Url(UrlModelReference {
-            url: configured_url,
-        })
-    );
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_tests/tests/balancer_refuses_a_stored_desired_state_with_an_ambiguous_parameter.rs b/paddler_tests/tests/balancer_refuses_a_stored_desired_state_with_an_ambiguous_parameter.rs
index dd10004bf..0d29b9d4e 100644
--- a/paddler_tests/tests/balancer_refuses_a_stored_desired_state_with_an_ambiguous_parameter.rs
+++ b/paddler_tests/tests/balancer_refuses_a_stored_desired_state_with_an_ambiguous_parameter.rs
@@ -1,3 +1,8 @@
+use std::error::Error as _;
+use std::mem::discriminant;
+
+use serde_json::Value;
+use serde_json::from_str;
 use serde_json::json;
 use serde_json::to_string;
 use serde_json::to_value;
@@ -5,23 +10,26 @@ use tokio::fs::write;
 
 use paddler_bootstrap::bootstrap_error::BootstrapError;
 use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_state_database::state_database_error::StateDatabaseError;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
-use paddler_test_cluster_harness::state_database_file::StateDatabaseFile;
+use paddler_test_cluster_harness::temporary_state_database_file::TemporaryStateDatabaseFile;
 use paddler_tests::start_cluster::start_cluster;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn balancer_refuses_a_stored_desired_state_with_an_ambiguous_parameter() {
-    let database = StateDatabaseFile::new().expect("the state database file must be created");
+    let database =
+        TemporaryStateDatabaseFile::new().expect("the state database file must be created");
     let mut stored_desired_state =
         to_value(BalancerDesiredState::default()).expect("the value must serialize");
 
-    stored_desired_state["inference_parameters"]["image_resize_to_fit"] = json!(0);
+    stored_desired_state["text_generation"]["multimodal"]["image_resize_to_fit"] = json!(0);
 
     write(
         &database.path,
         to_string(&json!({
             "balancer_desired_state": stored_desired_state,
-            "version": "1",
+            "version": "2",
         }))
         .expect("the value must serialize"),
     )
@@ -29,22 +37,35 @@ async fn balancer_refuses_a_stored_desired_state_with_an_ambiguous_parameter() {
     .expect("the file must be written");
 
     let start_error = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
         state_database_url: database.url.clone(),
-        desired_state: None,
-        ..ClusterParams::default()
+        desired_state: ClusterDesiredState::KeepStored,
+        ..ClusterParams::without_agents()
     })
     .await
     .err();
 
-    assert!(
-        matches!(
-            start_error
-                .as_ref()
-                .and_then(|error| error.downcast_ref::<BootstrapError>()),
-            Some(BootstrapError::StateDatabaseReadFailed { .. })
-        ),
+    let bootstrap_error = start_error
+        .as_ref()
+        .and_then(|error| error.downcast_ref::<BootstrapError>());
+
+    assert_eq!(
+        bootstrap_error.map(discriminant),
+        Some(discriminant(&BootstrapError::StateDatabaseReadFailed {
+            source: StateDatabaseError::FilePathEmpty
+        })),
         "{start_error:?}"
     );
+    assert_eq!(
+        bootstrap_error
+            .and_then(|bootstrap_error| bootstrap_error.source())
+            .map(ToString::to_string),
+        Some(
+            StateDatabaseError::FileContentsInvalid {
+                path: database.path.clone(),
+                source: from_str::<Value>("")
+                    .expect_err("an empty document must not parse as JSON"),
+            }
+            .to_string()
+        )
+    );
 }
diff --git a/paddler_tests/tests/balancer_rejects_a_second_registration_under_a_registered_agent_id.rs b/paddler_tests/tests/balancer_rejects_a_second_registration_under_a_registered_agent_id.rs
index 8eeed311c..c84731fc6 100644
--- a/paddler_tests/tests/balancer_rejects_a_second_registration_under_a_registered_agent_id.rs
+++ b/paddler_tests/tests/balancer_rejects_a_second_registration_under_a_registered_agent_id.rs
@@ -1,20 +1,20 @@
 use tokio_tungstenite::tungstenite::protocol::frame::coding::CloseCode;
 use tokio_util::sync::CancellationToken;
 
-use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-use paddler_messaging::management_socket::balancer::notification::Notification as ManagementJsonRpcNotification;
 use paddler_messaging::management_socket::balancer::notification_params::register_agent_params::RegisterAgentParams;
 use paddler_messaging::slot_aggregated_status_snapshot::SlotAggregatedStatusSnapshot;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
-use paddler_test_cluster_harness::raw_agent_socket::RawAgentSocket;
 use paddler_tests::start_cluster::start_cluster;
 
+use crate::untrusted_management_agent::UntrustedManagementAgent;
+
 #[tokio::test(flavor = "multi_thread")]
 async fn balancer_rejects_a_second_registration_under_a_registered_agent_id() {
     let cluster = start_cluster(ClusterParams {
         agents: AgentConfig::uniform(1, 1),
-        desired_state: Some(BalancerDesiredState::default()),
+        desired_state: ClusterDesiredState::Apply(Box::default()),
         wait_for_slots_ready: false,
         ..ClusterParams::default()
     })
@@ -25,36 +25,37 @@ async fn balancer_rejects_a_second_registration_under_a_registered_agent_id() {
         .first()
         .expect("the cluster must have a registered agent")
         .clone();
-    let mut impostor_socket =
-        RawAgentSocket::connect(cluster.balancer.addresses.management, &registered_agent_id)
-            .await
-            .expect("the impostor must reach the agent socket");
+    let mut impostor_agent = UntrustedManagementAgent::connect(
+        cluster.balancer.addresses.management,
+        &registered_agent_id,
+    )
+    .await
+    .expect("the impostor must reach the agent socket");
 
-    impostor_socket
-        .send_notification(ManagementJsonRpcNotification::RegisterAgent(
-            RegisterAgentParams {
-                name: Some("impostor".to_owned()),
-                slot_aggregated_status_snapshot: SlotAggregatedStatusSnapshot::default(),
-            },
-        ))
+    impostor_agent
+        .register(RegisterAgentParams {
+            name: Some("impostor".to_owned()),
+            slot_aggregated_status_snapshot: SlotAggregatedStatusSnapshot::default(),
+        })
         .await
         .expect("the impostor registration must be sent");
 
     assert_eq!(
-        impostor_socket
+        impostor_agent
+            .socket_client
             .next_close_frame()
             .await
             .expect("the balancer must close the impostor socket")
             .map(|close_frame| close_frame.code),
         Some(CloseCode::Policy)
     );
-    assert!(
+    assert_eq!(
         cluster
             .client_management
             .get_model_metadata(CancellationToken::new(), &registered_agent_id)
             .await
-            .is_ok(),
-        "the registered agent must keep serving requests"
+            .expect("the registered agent must keep serving requests"),
+        None
     );
 
     cluster
diff --git a/paddler_tests/tests/balancer_rejects_a_status_update_from_a_socket_that_did_not_register.rs b/paddler_tests/tests/balancer_rejects_a_status_update_from_a_socket_that_did_not_register.rs
index 000d7ea52..97b4c8715 100644
--- a/paddler_tests/tests/balancer_rejects_a_status_update_from_a_socket_that_did_not_register.rs
+++ b/paddler_tests/tests/balancer_rejects_a_status_update_from_a_socket_that_did_not_register.rs
@@ -2,20 +2,19 @@ use tokio_tungstenite::tungstenite::protocol::frame::coding::CloseCode;
 use tokio_util::sync::CancellationToken;
 
 use paddler_messaging::agent_status::AgentStatus;
-use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-use paddler_messaging::management_socket::balancer::notification::Notification as ManagementJsonRpcNotification;
-use paddler_messaging::management_socket::balancer::notification_params::update_agent_status_params::UpdateAgentStatusParams;
 use paddler_messaging::slot_aggregated_status_snapshot::SlotAggregatedStatusSnapshot;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
-use paddler_test_cluster_harness::raw_agent_socket::RawAgentSocket;
 use paddler_tests::start_cluster::start_cluster;
 
+use crate::untrusted_management_agent::UntrustedManagementAgent;
+
 #[tokio::test(flavor = "multi_thread")]
 async fn balancer_rejects_a_status_update_from_a_socket_that_did_not_register() {
     let cluster = start_cluster(ClusterParams {
         agents: AgentConfig::uniform(1, 1),
-        desired_state: Some(BalancerDesiredState::default()),
+        desired_state: ClusterDesiredState::Apply(Box::default()),
         wait_for_slots_ready: false,
         ..ClusterParams::default()
     })
@@ -26,28 +25,27 @@ async fn balancer_rejects_a_status_update_from_a_socket_that_did_not_register()
         .first()
         .expect("the cluster must have a registered agent")
         .clone();
-    let mut impostor_socket =
-        RawAgentSocket::connect(cluster.balancer.addresses.management, &registered_agent_id)
-            .await
-            .expect("the impostor must reach the agent socket");
+    let mut impostor_agent = UntrustedManagementAgent::connect(
+        cluster.balancer.addresses.management,
+        &registered_agent_id,
+    )
+    .await
+    .expect("the impostor must reach the agent socket");
 
-    impostor_socket
-        .send_notification(ManagementJsonRpcNotification::UpdateAgentStatus(
-            UpdateAgentStatusParams {
-                slot_aggregated_status_snapshot: SlotAggregatedStatusSnapshot {
-                    status: AgentStatus {
-                        desired_slots_total: 99,
-                        ..AgentStatus::default()
-                    },
-                    version: u64::MAX,
-                },
+    impostor_agent
+        .report_status(SlotAggregatedStatusSnapshot {
+            status: AgentStatus {
+                desired_slots_total: 99,
+                ..AgentStatus::default()
             },
-        ))
+            version: u64::MAX,
+        })
         .await
         .expect("the impostor status update must be sent");
 
     assert_eq!(
-        impostor_socket
+        impostor_agent
+            .socket_client
             .next_close_frame()
             .await
             .expect("the balancer must close the impostor socket")
diff --git a/paddler_tests/tests/balancer_rejects_embedding_batch_with_501_when_embeddings_are_disabled.rs b/paddler_tests/tests/balancer_rejects_embedding_batch_with_501_when_embeddings_are_disabled.rs
deleted file mode 100644
index e28986d02..000000000
--- a/paddler_tests/tests/balancer_rejects_embedding_batch_with_501_when_embeddings_are_disabled.rs
+++ /dev/null
@@ -1,49 +0,0 @@
-use tokio_util::sync::CancellationToken;
-
-use paddler_client::error::Error as ClientError;
-use paddler_messaging::embedding_input_document::EmbeddingInputDocument;
-use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
-use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
-use paddler_test_cluster_harness::cluster_params::ClusterParams;
-use paddler_tests::start_cluster::start_cluster;
-
-const NOT_IMPLEMENTED: u16 = 501;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn balancer_rejects_embedding_batch_with_501_when_embeddings_are_disabled() {
-    let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("a cluster without agents must start");
-
-    let rejection = cluster
-        .client_inference
-        .post_generate_embedding_batch(
-            CancellationToken::new(),
-            &GenerateEmbeddingBatchParams {
-                input_batch: vec![EmbeddingInputDocument {
-                    content: "Hello world".to_owned(),
-                    id: "doc-1".to_owned(),
-                }],
-                normalization_method: EmbeddingNormalizationMethod::None,
-            },
-        )
-        .await
-        .err()
-        .expect("the balancer must reject embedding batches while embeddings are disabled");
-
-    assert!(matches!(
-        rejection,
-        ClientError::UnexpectedResponseStatus { message, status, .. }
-            if status.as_u16() == NOT_IMPLEMENTED
-                && message == "Embedding generation is not enabled in the inference parameters"
-    ));
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_tests/tests/balancer_rejects_http_token_generation_with_501_while_the_cluster_serves_embeddings.rs b/paddler_tests/tests/balancer_rejects_http_token_generation_with_501_while_the_cluster_serves_embeddings.rs
deleted file mode 100644
index 15aa87f26..000000000
--- a/paddler_tests/tests/balancer_rejects_http_token_generation_with_501_while_the_cluster_serves_embeddings.rs
+++ /dev/null
@@ -1,89 +0,0 @@
-use std::num::NonZeroU32;
-
-use http::StatusCode;
-use tokio_util::sync::CancellationToken;
-
-use paddler_balancer::cluster_token_generation_mode::TOKEN_GENERATION_DISABLED_MESSAGE;
-use paddler_client::error::Error as ClientError;
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
-use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-use paddler_messaging::conversation_history::ConversationHistory;
-use paddler_messaging::conversation_message::ConversationMessage;
-use paddler_messaging::conversation_message_content::ConversationMessageContent;
-use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
-use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
-use paddler_test_cluster_harness::cluster_params::ClusterParams;
-use paddler_tests::start_cluster::start_cluster;
-
-fn is_token_generation_disabled_rejection(rejection: &ClientError) -> bool {
-    matches!(
-        rejection,
-        ClientError::UnexpectedResponseStatus { message, status, .. }
-            if *status == StatusCode::NOT_IMPLEMENTED
-                && message == TOKEN_GENERATION_DISABLED_MESSAGE
-    )
-}
-
-#[tokio::test(flavor = "multi_thread")]
-async fn balancer_rejects_http_token_generation_with_501_while_the_cluster_serves_embeddings() {
-    let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        desired_state: Some(BalancerDesiredState {
-            inference_parameters: InferenceParameters {
-                enable_embeddings: true,
-                ..InferenceParameters::default()
-            },
-            ..BalancerDesiredState::default()
-        }),
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("a balancer serving embeddings must start");
-
-    let raw_prompt_rejection = cluster
-        .client_inference
-        .post_continue_from_raw_prompt(
-            CancellationToken::new(),
-            &ContinueFromRawPromptParams {
-                grammar: None,
-                max_tokens: NonZeroU32::MIN,
-                raw_prompt: "hello".to_owned(),
-            },
-        )
-        .await
-        .err()
-        .expect("the raw prompt route must reject token generation in embeddings mode");
-    let conversation_history_rejection = cluster
-        .client_inference
-        .post_continue_from_conversation_history(
-            CancellationToken::new(),
-            &ContinueFromConversationHistoryParams {
-                add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Text("hello".to_owned()),
-                    role: "user".to_owned(),
-                }]),
-                enable_thinking: false,
-                grammar: None,
-                max_tokens: NonZeroU32::MIN,
-                parse_tool_calls: false,
-                tools: Vec::new(),
-            },
-        )
-        .await
-        .err()
-        .expect("the conversation history route must reject token generation in embeddings mode");
-
-    assert!(is_token_generation_disabled_rejection(
-        &raw_prompt_rejection
-    ));
-    assert!(is_token_generation_disabled_rejection(
-        &conversation_history_rejection
-    ));
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_tests/tests/balancer_rejects_tool_call_parsing_without_tools.rs b/paddler_tests/tests/balancer_rejects_tool_call_parsing_without_tools.rs
index 00a5fa4cc..3e9d2337b 100644
--- a/paddler_tests/tests/balancer_rejects_tool_call_parsing_without_tools.rs
+++ b/paddler_tests/tests/balancer_rejects_tool_call_parsing_without_tools.rs
@@ -14,13 +14,9 @@ const BAD_REQUEST: u16 = 400;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn balancer_rejects_tool_call_parsing_without_tools() {
-    let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("the cluster must start");
+    let cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("the cluster must start");
 
     let rejection = cluster
         .client_inference
@@ -28,10 +24,14 @@ async fn balancer_rejects_tool_call_parsing_without_tools() {
             CancellationToken::new(),
             &ContinueFromConversationHistoryParams {
                 add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Text("What is the weather?".to_owned()),
-                    role: "user".to_owned(),
-                }]),
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Text(
+                            "What is the weather?".to_owned(),
+                        ),
+                        role: "user".to_owned(),
+                    },
+                ]),
                 enable_thinking: false,
                 grammar: None,
                 max_tokens: NonZeroU32::MIN,
diff --git a/paddler_tests/tests/balancer_rejects_tool_whose_required_field_is_not_a_property.rs b/paddler_tests/tests/balancer_rejects_tool_whose_required_field_is_not_a_property.rs
index 8017227e6..91b49ee99 100644
--- a/paddler_tests/tests/balancer_rejects_tool_whose_required_field_is_not_a_property.rs
+++ b/paddler_tests/tests/balancer_rejects_tool_whose_required_field_is_not_a_property.rs
@@ -9,13 +9,9 @@ const BAD_REQUEST: u16 = 400;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn balancer_rejects_tool_whose_required_field_is_not_a_property() {
-    let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("a cluster without agents must start");
+    let cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("a cluster without agents must start");
 
     let rejection = cluster
         .client_inference
diff --git a/paddler_tests/tests/balancer_reports_a_hugging_face_lookup_issue_for_a_nonexistent_model.rs b/paddler_tests/tests/balancer_reports_a_hugging_face_lookup_issue_for_a_nonexistent_model.rs
new file mode 100644
index 000000000..d4824d2cf
--- /dev/null
+++ b/paddler_tests/tests/balancer_reports_a_hugging_face_lookup_issue_for_a_nonexistent_model.rs
@@ -0,0 +1,37 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use paddler_messaging::agent_issue::AgentIssue;
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_model_source::huggingface_model_reference::HuggingFaceModelReference;
+use paddler_model_source::model_source::ModelSource;
+use paddler_tests::start_single_agent_cluster_with_desired_state::start_single_agent_cluster_with_desired_state;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn balancer_reports_a_hugging_face_lookup_issue_for_a_nonexistent_model() {
+    let reference = HuggingFaceModelReference {
+        filename: "nonexistent.gguf".to_owned(),
+        repository_name: "nonexistent-model-gguf".to_owned(),
+        repository_owner: "nonexistent-org".to_owned(),
+        revision: "main".to_owned(),
+    };
+    let lookup_issues = [
+        AgentIssue::HuggingFaceModelDoesNotExist(reference.model_path()),
+        AgentIssue::HuggingFacePermissions(reference.model_path()),
+    ];
+    let mut cluster = start_single_agent_cluster_with_desired_state(BalancerDesiredState {
+        model: ModelSource::HuggingFace(reference).into_agent_desired_model(),
+        ..BalancerDesiredState::default()
+    })
+    .await
+    .expect("a single-agent cluster must start");
+
+    cluster
+        .wait_for_first_agent_issue(|issue| lookup_issues.contains(issue))
+        .await
+        .expect("the agent must report a Hugging Face lookup issue for a nonexistent repository");
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/balancer_reports_bad_gateway_when_agent_disconnects_mid_generation.rs b/paddler_tests/tests/balancer_reports_bad_gateway_when_agent_disconnects_mid_generation.rs
index 932efc066..6b345bf3c 100644
--- a/paddler_tests/tests/balancer_reports_bad_gateway_when_agent_disconnects_mid_generation.rs
+++ b/paddler_tests/tests/balancer_reports_bad_gateway_when_agent_disconnects_mid_generation.rs
@@ -3,8 +3,8 @@
 use futures_util::StreamExt as _;
 use tokio_util::sync::CancellationToken;
 
-use paddler_messaging::inference_client::message::Message;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::inference_stream_error_codes::inference_stream_error_codes;
 use paddler_test_cluster_harness::unending_generation::unending_generation;
 use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
 
@@ -35,18 +35,15 @@ async fn balancer_reports_bad_gateway_when_agent_disconnects_mid_generation() {
         .await
         .expect("the cluster must shut down cleanly");
 
-    while let Some(message) = stream.next().await {
-        if let Message::Error(error_envelope) = message.expect("the message must be readable") {
-            assert_eq!(error_envelope.error.code, BAD_GATEWAY);
+    assert_eq!(
+        inference_stream_error_codes(stream)
+            .await
+            .expect("every streamed message must be readable"),
+        vec![BAD_GATEWAY]
+    );
 
-            cluster
-                .shutdown()
-                .await
-                .expect("the cluster must shut down cleanly");
-
-            return;
-        }
-    }
-
-    panic!("the stream ended without reporting the lost agent")
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
 }
diff --git a/paddler_tests/tests/balancer_reports_chat_template_does_not_compile_for_invalid_jinja.rs b/paddler_tests/tests/balancer_reports_chat_template_does_not_compile_for_invalid_jinja.rs
index a23d35d7b..7915a424a 100644
--- a/paddler_tests/tests/balancer_reports_chat_template_does_not_compile_for_invalid_jinja.rs
+++ b/paddler_tests/tests/balancer_reports_chat_template_does_not_compile_for_invalid_jinja.rs
@@ -1,33 +1,42 @@
 #![cfg(feature = "tests_that_use_llms")]
 
-use paddler_messaging::agent_desired_model::AgentDesiredModel;
-use paddler_messaging::agent_issue::AgentIssue;
+use std::mem::discriminant;
+
 use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-use paddler_messaging::chat_template::ChatTemplate;
-use paddler_test_cluster_harness::model_card::ModelCard;
+use paddler_model_source::model_source::ModelSource;
 use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
+use paddler_tests::desired_state_with_chat_template_override::desired_state_with_chat_template_override;
 use paddler_tests::start_single_agent_cluster_with_desired_state::start_single_agent_cluster_with_desired_state;
+use paddler_tests::uncompilable_chat_template::uncompilable_chat_template;
+use paddler_tests::uncompilable_chat_template_issue::uncompilable_chat_template_issue;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn balancer_reports_chat_template_does_not_compile_for_invalid_jinja() {
-    let ModelCard { reference } = qwen3_0_6b();
-    let mut cluster = start_single_agent_cluster_with_desired_state(BalancerDesiredState {
-        chat_template_override: Some(ChatTemplate {
-            content: "{{invalid jinja template".to_owned(),
-        }),
-        model: AgentDesiredModel::HuggingFace(reference),
-        use_chat_template_override: true,
-        ..BalancerDesiredState::default()
-    })
-    .await
-    .expect("a single-agent cluster must start");
+    let model_card = qwen3_0_6b();
+    let expected_issue = uncompilable_chat_template_issue(
+        &model_card
+            .cached_path()
+            .expect("the model must be in the Hugging Face cache"),
+    )
+    .expect("the uncompilable chat template must not compile");
+    let mut cluster =
+        start_single_agent_cluster_with_desired_state(desired_state_with_chat_template_override(
+            BalancerDesiredState {
+                model: ModelSource::HuggingFace(model_card.reference).into_agent_desired_model(),
+                ..BalancerDesiredState::default()
+            },
+            uncompilable_chat_template(),
+        ))
+        .await
+        .expect("a single-agent cluster must start");
 
-    cluster
-        .wait_for_first_agent_issue(|issue| {
-            matches!(issue, AgentIssue::ChatTemplateDoesNotCompile(_))
-        })
+    let snapshot = cluster
+        .wait_for_first_agent_issue(|issue| discriminant(issue) == discriminant(&expected_issue))
         .await
         .expect("the agent must report ChatTemplateDoesNotCompile for invalid Jinja syntax");
+    let issues = &snapshot.agents[0].status.issues;
+
+    assert!(issues.contains(&expected_issue), "{issues:?}");
 
     cluster
         .shutdown()
diff --git a/paddler_tests/tests/balancer_reports_chat_template_does_not_compile_recovers_when_template_replaced.rs b/paddler_tests/tests/balancer_reports_chat_template_does_not_compile_recovers_when_template_replaced.rs
deleted file mode 100644
index e8df77935..000000000
--- a/paddler_tests/tests/balancer_reports_chat_template_does_not_compile_recovers_when_template_replaced.rs
+++ /dev/null
@@ -1,76 +0,0 @@
-#![cfg(feature = "tests_that_use_llms")]
-
-use tokio_util::sync::CancellationToken;
-
-use paddler_messaging::agent_desired_model::AgentDesiredModel;
-use paddler_messaging::agent_issue::AgentIssue;
-use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-use paddler_messaging::chat_template::ChatTemplate;
-use paddler_test_cluster_harness::model_card::ModelCard;
-use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
-use paddler_tests::start_single_agent_cluster_with_desired_state::start_single_agent_cluster_with_desired_state;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn balancer_reports_chat_template_does_not_compile_recovers_when_template_replaced() {
-    let ModelCard { reference } = qwen3_0_6b();
-    let mut cluster = start_single_agent_cluster_with_desired_state(BalancerDesiredState {
-        chat_template_override: Some(ChatTemplate {
-            content: "{{invalid jinja template".to_owned(),
-        }),
-        model: AgentDesiredModel::HuggingFace(reference.clone()),
-        use_chat_template_override: true,
-        ..BalancerDesiredState::default()
-    })
-    .await
-    .expect("a single-agent cluster must start");
-
-    cluster
-        .wait_for_first_agent_issue(|issue| {
-            matches!(issue, AgentIssue::ChatTemplateDoesNotCompile(_))
-        })
-        .await
-        .expect("the agent must report ChatTemplateDoesNotCompile for invalid Jinja syntax");
-
-    cluster
-        .client_management
-        .put_balancer_desired_state(
-            CancellationToken::new(),
-            &BalancerDesiredState {
-                chat_template_override: Some(ChatTemplate {
-                    content: "{% for message in messages %}{{ message.content }}{% endfor %}"
-                        .to_owned(),
-                }),
-                model: AgentDesiredModel::HuggingFace(reference),
-                use_chat_template_override: true,
-                ..BalancerDesiredState::default()
-            },
-        )
-        .await
-        .expect("the balancer must accept the desired state with a valid template");
-
-    let agent_id = cluster
-        .agent_ids
-        .first()
-        .expect("the cluster must have a registered agent")
-        .clone();
-
-    cluster
-        .agents_watcher
-        .until_agent(&agent_id, |snapshot| {
-            snapshot.agents.iter().any(|agent| {
-                agent.id == agent_id
-                    && !agent
-                        .status
-                        .issues
-                        .iter()
-                        .any(|issue| matches!(issue, AgentIssue::ChatTemplateDoesNotCompile(_)))
-            })
-        })
-        .await
-        .expect("the agent must stay registered while reconciliation clears its template issue");
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_tests/tests/balancer_reports_download_server_denied_access.rs b/paddler_tests/tests/balancer_reports_download_server_denied_access.rs
index bc52e276e..8421d4145 100644
--- a/paddler_tests/tests/balancer_reports_download_server_denied_access.rs
+++ b/paddler_tests/tests/balancer_reports_download_server_denied_access.rs
@@ -5,7 +5,6 @@ use paddler_local_http_fixture::local_http_fixture::LocalHttpFixture;
 use paddler_messaging::agent_desired_model::AgentDesiredModel;
 use paddler_messaging::agent_issue::AgentIssue;
 use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-use paddler_messaging::url_model_reference::UrlModelReference;
 use paddler_tests::start_single_agent_cluster_with_desired_state::start_single_agent_cluster_with_desired_state;
 
 #[tokio::test(flavor = "multi_thread")]
@@ -15,9 +14,7 @@ async fn balancer_reports_download_server_denied_access() {
         .expect("the local HTTP fixture must start");
     let model_url = fixture.url("/private.gguf");
     let mut cluster = start_single_agent_cluster_with_desired_state(BalancerDesiredState {
-        model: AgentDesiredModel::Url(UrlModelReference {
-            url: model_url.clone(),
-        }),
+        model: AgentDesiredModel::Uri(model_url.clone()),
         ..BalancerDesiredState::default()
     })
     .await
@@ -34,4 +31,10 @@ async fn balancer_reports_download_server_denied_access() {
         .shutdown()
         .await
         .expect("the cluster must shut down cleanly");
+
+    fixture
+        .stop()
+        .await
+        .expect("the fixture task must not panic")
+        .expect("the fixture must serve every request");
 }
diff --git a/paddler_tests/tests/balancer_reports_download_server_errored.rs b/paddler_tests/tests/balancer_reports_download_server_errored.rs
index dafb0f100..5b72109f4 100644
--- a/paddler_tests/tests/balancer_reports_download_server_errored.rs
+++ b/paddler_tests/tests/balancer_reports_download_server_errored.rs
@@ -5,7 +5,6 @@ use paddler_local_http_fixture::local_http_fixture::LocalHttpFixture;
 use paddler_messaging::agent_desired_model::AgentDesiredModel;
 use paddler_messaging::agent_issue::AgentIssue;
 use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-use paddler_messaging::url_model_reference::UrlModelReference;
 use paddler_tests::start_single_agent_cluster_with_desired_state::start_single_agent_cluster_with_desired_state;
 
 #[tokio::test(flavor = "multi_thread")]
@@ -16,9 +15,7 @@ async fn balancer_reports_download_server_errored() {
             .expect("the local HTTP fixture must start");
     let model_url = fixture.url("/broken.gguf");
     let mut cluster = start_single_agent_cluster_with_desired_state(BalancerDesiredState {
-        model: AgentDesiredModel::Url(UrlModelReference {
-            url: model_url.clone(),
-        }),
+        model: AgentDesiredModel::Uri(model_url.clone()),
         ..BalancerDesiredState::default()
     })
     .await
@@ -35,4 +32,10 @@ async fn balancer_reports_download_server_errored() {
         .shutdown()
         .await
         .expect("the cluster must shut down cleanly");
+
+    fixture
+        .stop()
+        .await
+        .expect("the fixture task must not panic")
+        .expect("the fixture must serve every request");
 }
diff --git a/paddler_tests/tests/balancer_reports_download_server_is_unreachable.rs b/paddler_tests/tests/balancer_reports_download_server_is_unreachable.rs
index c76684d70..909dc6822 100644
--- a/paddler_tests/tests/balancer_reports_download_server_is_unreachable.rs
+++ b/paddler_tests/tests/balancer_reports_download_server_is_unreachable.rs
@@ -1,16 +1,13 @@
 use paddler_messaging::agent_desired_model::AgentDesiredModel;
 use paddler_messaging::agent_issue::AgentIssue;
 use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-use paddler_messaging::url_model_reference::UrlModelReference;
 use paddler_tests::start_single_agent_cluster_with_desired_state::start_single_agent_cluster_with_desired_state;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn balancer_reports_download_server_is_unreachable() {
     let model_url = "http://127.0.0.1:1/model.gguf".to_owned();
     let mut cluster = start_single_agent_cluster_with_desired_state(BalancerDesiredState {
-        model: AgentDesiredModel::Url(UrlModelReference {
-            url: model_url.clone(),
-        }),
+        model: AgentDesiredModel::Uri(model_url.clone()),
         ..BalancerDesiredState::default()
     })
     .await
diff --git a/paddler_tests/tests/balancer_reports_download_url_is_malformed.rs b/paddler_tests/tests/balancer_reports_download_url_is_malformed.rs
index 9383a8389..647328383 100644
--- a/paddler_tests/tests/balancer_reports_download_url_is_malformed.rs
+++ b/paddler_tests/tests/balancer_reports_download_url_is_malformed.rs
@@ -1,16 +1,13 @@
 use paddler_messaging::agent_desired_model::AgentDesiredModel;
 use paddler_messaging::agent_issue::AgentIssue;
 use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-use paddler_messaging::url_model_reference::UrlModelReference;
 use paddler_tests::start_single_agent_cluster_with_desired_state::start_single_agent_cluster_with_desired_state;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn balancer_reports_download_url_is_malformed() {
-    let model_url = "not a valid url".to_owned();
+    let model_url = "ftp://example.com/model.gguf".to_owned();
     let mut cluster = start_single_agent_cluster_with_desired_state(BalancerDesiredState {
-        model: AgentDesiredModel::Url(UrlModelReference {
-            url: model_url.clone(),
-        }),
+        model: AgentDesiredModel::Uri(model_url.clone()),
         ..BalancerDesiredState::default()
     })
     .await
diff --git a/paddler_tests/tests/balancer_reports_huggingface_model_does_not_exist.rs b/paddler_tests/tests/balancer_reports_huggingface_model_does_not_exist.rs
deleted file mode 100644
index 9e8bb4632..000000000
--- a/paddler_tests/tests/balancer_reports_huggingface_model_does_not_exist.rs
+++ /dev/null
@@ -1,36 +0,0 @@
-#![cfg(feature = "tests_that_use_llms")]
-
-use paddler_messaging::agent_desired_model::AgentDesiredModel;
-use paddler_messaging::agent_issue::AgentIssue;
-use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-use paddler_messaging::huggingface_model_reference::HuggingFaceModelReference;
-use paddler_tests::start_single_agent_cluster_with_desired_state::start_single_agent_cluster_with_desired_state;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn balancer_reports_huggingface_model_does_not_exist() {
-    let mut cluster = start_single_agent_cluster_with_desired_state(BalancerDesiredState {
-        model: AgentDesiredModel::HuggingFace(HuggingFaceModelReference {
-            filename: "nonexistent.gguf".to_owned(),
-            repo_id: "nonexistent-org/nonexistent-model-gguf".to_owned(),
-            revision: "main".to_owned(),
-        }),
-        ..BalancerDesiredState::default()
-    })
-    .await
-    .expect("a single-agent cluster must start");
-
-    cluster
-        .wait_for_first_agent_issue(|issue| {
-            matches!(
-                issue,
-                AgentIssue::HuggingFaceModelDoesNotExist(_) | AgentIssue::HuggingFacePermissions(_)
-            )
-        })
-        .await
-        .expect("the agent must report a Hugging Face lookup issue for a nonexistent repository");
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_tests/tests/balancer_reports_model_cannot_be_loaded_for_corrupt_file.rs b/paddler_tests/tests/balancer_reports_model_cannot_be_loaded_for_corrupt_file.rs
index fd6c82ead..d5f3c4933 100644
--- a/paddler_tests/tests/balancer_reports_model_cannot_be_loaded_for_corrupt_file.rs
+++ b/paddler_tests/tests/balancer_reports_model_cannot_be_loaded_for_corrupt_file.rs
@@ -2,9 +2,9 @@ use std::io::Write as _;
 
 use tempfile::NamedTempFile;
 
-use paddler_messaging::agent_desired_model::AgentDesiredModel;
 use paddler_messaging::agent_issue::AgentIssue;
 use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_test_cluster_harness::agent_desired_model_at::agent_desired_model_at;
 use paddler_tests::start_single_agent_cluster_with_desired_state::start_single_agent_cluster_with_desired_state;
 
 #[tokio::test(flavor = "multi_thread")]
@@ -21,7 +21,8 @@ async fn balancer_reports_model_cannot_be_loaded_for_corrupt_file() {
         .expect("the temporary model path must be valid UTF-8")
         .to_owned();
     let mut cluster = start_single_agent_cluster_with_desired_state(BalancerDesiredState {
-        model: AgentDesiredModel::LocalToAgent(corrupt_model_path.clone()),
+        model: agent_desired_model_at(corrupt_model.path())
+            .expect("the corrupt model path must be absolute"),
         ..BalancerDesiredState::default()
     })
     .await
diff --git a/paddler_tests/tests/balancer_reports_model_cannot_be_loaded_for_invalid_gguf.rs b/paddler_tests/tests/balancer_reports_model_cannot_be_loaded_for_invalid_gguf.rs
index d32f445d0..224109bc9 100644
--- a/paddler_tests/tests/balancer_reports_model_cannot_be_loaded_for_invalid_gguf.rs
+++ b/paddler_tests/tests/balancer_reports_model_cannot_be_loaded_for_invalid_gguf.rs
@@ -1,28 +1,51 @@
-use paddler_messaging::agent_desired_model::AgentDesiredModel;
 use paddler_messaging::agent_issue::AgentIssue;
 use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-use paddler_tests::start_single_agent_cluster_with_desired_state::start_single_agent_cluster_with_desired_state;
+use paddler_messaging::decision_settings::DecisionSettings;
+use paddler_messaging::decision_slots_minimum::DECISION_SLOTS_MINIMUM;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::agent_desired_model_at::agent_desired_model_at;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_test_cluster_harness::fixture_path::fixture_path;
+use paddler_test_cluster_harness::pointer_head_fixture::pointer_head_fixture;
+use paddler_test_cluster_harness::synthetic_pointer_head_fixture::QWEN3_5_0_8B_SYNTHETIC_POINTER_HEAD_FIXTURE;
+use paddler_tests::start_cluster::start_cluster;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn balancer_reports_model_cannot_be_loaded_for_invalid_gguf() {
-    let model_path_on_agent =
-        concat!(env!("CARGO_MANIFEST_DIR"), "/../fixtures/invalid.gguf").to_owned();
-    let mut cluster = start_single_agent_cluster_with_desired_state(BalancerDesiredState {
-        model: AgentDesiredModel::LocalToAgent(model_path_on_agent.clone()),
-        ..BalancerDesiredState::default()
-    })
-    .await
-    .expect("a single-agent cluster must start");
+    let model_path_on_agent = fixture_path("invalid.gguf").display().to_string();
 
-    cluster
-        .wait_for_first_agent_issue(|issue| {
-            matches!(issue, AgentIssue::ModelCannotBeLoaded(model_path) if model_path.model_path == model_path_on_agent)
+    for decision in [
+        DecisionSettings::default(),
+        DecisionSettings {
+            pointer_head: pointer_head_fixture(QWEN3_5_0_8B_SYNTHETIC_POINTER_HEAD_FIXTURE)
+                .expect("the pointer head fixture path must be absolute"),
+        },
+    ] {
+        let mut cluster = start_cluster(ClusterParams {
+            agents: AgentConfig::uniform(1, DECISION_SLOTS_MINIMUM.get()),
+            desired_state: ClusterDesiredState::Apply(Box::new(BalancerDesiredState {
+                decision,
+                model: agent_desired_model_at(&fixture_path("invalid.gguf"))
+                    .expect("the invalid model fixture path must be absolute"),
+                ..BalancerDesiredState::default()
+            })),
+            wait_for_slots_ready: false,
+            ..ClusterParams::default()
         })
         .await
-        .expect("the agent must report ModelCannotBeLoaded for the configured path");
+        .expect("a single-agent cluster must start");
 
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
+        cluster
+            .wait_for_first_agent_issue(|issue| {
+                matches!(issue, AgentIssue::ModelCannotBeLoaded(model_path) if model_path.model_path == model_path_on_agent)
+            })
+            .await
+            .expect("the agent must report ModelCannotBeLoaded for the configured path");
+
+        cluster
+            .shutdown()
+            .await
+            .expect("the cluster must shut down cleanly");
+    }
 }
diff --git a/paddler_tests/tests/balancer_reports_model_does_not_exist_at_url.rs b/paddler_tests/tests/balancer_reports_model_does_not_exist_at_url.rs
index 0eed090f8..aeecf281f 100644
--- a/paddler_tests/tests/balancer_reports_model_does_not_exist_at_url.rs
+++ b/paddler_tests/tests/balancer_reports_model_does_not_exist_at_url.rs
@@ -5,7 +5,6 @@ use paddler_local_http_fixture::local_http_fixture::LocalHttpFixture;
 use paddler_messaging::agent_desired_model::AgentDesiredModel;
 use paddler_messaging::agent_issue::AgentIssue;
 use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-use paddler_messaging::url_model_reference::UrlModelReference;
 use paddler_tests::start_single_agent_cluster_with_desired_state::start_single_agent_cluster_with_desired_state;
 
 #[tokio::test(flavor = "multi_thread")]
@@ -15,9 +14,7 @@ async fn balancer_reports_model_does_not_exist_at_url() {
         .expect("the local HTTP fixture must start");
     let model_url = fixture.url("/missing.gguf");
     let mut cluster = start_single_agent_cluster_with_desired_state(BalancerDesiredState {
-        model: AgentDesiredModel::Url(UrlModelReference {
-            url: model_url.clone(),
-        }),
+        model: AgentDesiredModel::Uri(model_url.clone()),
         ..BalancerDesiredState::default()
     })
     .await
@@ -34,4 +31,10 @@ async fn balancer_reports_model_does_not_exist_at_url() {
         .shutdown()
         .await
         .expect("the cluster must shut down cleanly");
+
+    fixture
+        .stop()
+        .await
+        .expect("the fixture task must not panic")
+        .expect("the fixture must serve every request");
 }
diff --git a/paddler_tests/tests/balancer_reports_model_file_does_not_exist.rs b/paddler_tests/tests/balancer_reports_model_file_does_not_exist.rs
index fa7c6a52c..173db2c17 100644
--- a/paddler_tests/tests/balancer_reports_model_file_does_not_exist.rs
+++ b/paddler_tests/tests/balancer_reports_model_file_does_not_exist.rs
@@ -1,13 +1,16 @@
-use paddler_messaging::agent_desired_model::AgentDesiredModel;
+use std::path::Path;
+
 use paddler_messaging::agent_issue::AgentIssue;
 use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_test_cluster_harness::agent_desired_model_at::agent_desired_model_at;
 use paddler_tests::start_single_agent_cluster_with_desired_state::start_single_agent_cluster_with_desired_state;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn balancer_reports_model_file_does_not_exist() {
     let model_path_on_agent = "/nonexistent/model.gguf".to_owned();
     let mut cluster = start_single_agent_cluster_with_desired_state(BalancerDesiredState {
-        model: AgentDesiredModel::LocalToAgent(model_path_on_agent.clone()),
+        model: agent_desired_model_at(Path::new(&model_path_on_agent))
+            .expect("the missing model path must be absolute"),
         ..BalancerDesiredState::default()
     })
     .await
diff --git a/paddler_tests/tests/balancer_reports_multimodal_projection_cannot_be_loaded_for_invalid_file.rs b/paddler_tests/tests/balancer_reports_multimodal_projection_cannot_be_loaded_for_invalid_file.rs
index d03706e93..221c9bd04 100644
--- a/paddler_tests/tests/balancer_reports_multimodal_projection_cannot_be_loaded_for_invalid_file.rs
+++ b/paddler_tests/tests/balancer_reports_multimodal_projection_cannot_be_loaded_for_invalid_file.rs
@@ -1,27 +1,30 @@
 #![cfg(feature = "tests_that_use_llms")]
 
-use paddler_messaging::agent_desired_model::AgentDesiredModel;
 use paddler_messaging::agent_issue::AgentIssue;
 use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_model_source::model_source::ModelSource;
+use paddler_test_cluster_harness::agent_desired_model_at::agent_desired_model_at;
+use paddler_test_cluster_harness::fixture_path::fixture_path;
 use paddler_test_cluster_harness::model_card::ModelCard;
 use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
+use paddler_tests::desired_state_with_multimodal_projection::desired_state_with_multimodal_projection;
 use paddler_tests::start_single_agent_cluster_with_desired_state::start_single_agent_cluster_with_desired_state;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn balancer_reports_multimodal_projection_cannot_be_loaded_for_invalid_file() {
     let ModelCard { reference } = qwen3_0_6b();
-    let projection_path = concat!(
-        env!("CARGO_MANIFEST_DIR"),
-        "/../fixtures/invalid_mmproj.gguf"
-    )
-    .to_owned();
-    let mut cluster = start_single_agent_cluster_with_desired_state(BalancerDesiredState {
-        model: AgentDesiredModel::HuggingFace(reference),
-        multimodal_projection: AgentDesiredModel::LocalToAgent(projection_path.clone()),
-        ..BalancerDesiredState::default()
-    })
-    .await
-    .expect("a single-agent cluster must start");
+    let projection_path = fixture_path("invalid_mmproj.gguf").display().to_string();
+    let mut cluster =
+        start_single_agent_cluster_with_desired_state(desired_state_with_multimodal_projection(
+            BalancerDesiredState {
+                model: ModelSource::HuggingFace(reference).into_agent_desired_model(),
+                ..BalancerDesiredState::default()
+            },
+            agent_desired_model_at(&fixture_path("invalid_mmproj.gguf"))
+                .expect("the invalid projection fixture path must be absolute"),
+        ))
+        .await
+        .expect("a single-agent cluster must start");
 
     cluster
         .wait_for_first_agent_issue(|issue| {
diff --git a/paddler_tests/tests/balancer_reports_multimodal_projection_cannot_be_loaded_for_missing_file.rs b/paddler_tests/tests/balancer_reports_multimodal_projection_cannot_be_loaded_for_missing_file.rs
index 7dcf6ba67..bf717e4ef 100644
--- a/paddler_tests/tests/balancer_reports_multimodal_projection_cannot_be_loaded_for_missing_file.rs
+++ b/paddler_tests/tests/balancer_reports_multimodal_projection_cannot_be_loaded_for_missing_file.rs
@@ -1,23 +1,31 @@
 #![cfg(feature = "tests_that_use_llms")]
 
-use paddler_messaging::agent_desired_model::AgentDesiredModel;
+use std::path::Path;
+
 use paddler_messaging::agent_issue::AgentIssue;
 use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_model_source::model_source::ModelSource;
+use paddler_test_cluster_harness::agent_desired_model_at::agent_desired_model_at;
 use paddler_test_cluster_harness::model_card::ModelCard;
 use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
+use paddler_tests::desired_state_with_multimodal_projection::desired_state_with_multimodal_projection;
 use paddler_tests::start_single_agent_cluster_with_desired_state::start_single_agent_cluster_with_desired_state;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn balancer_reports_multimodal_projection_cannot_be_loaded_for_missing_file() {
     let ModelCard { reference } = qwen3_0_6b();
     let projection_path = "/nonexistent/projection.bin".to_owned();
-    let mut cluster = start_single_agent_cluster_with_desired_state(BalancerDesiredState {
-        model: AgentDesiredModel::HuggingFace(reference),
-        multimodal_projection: AgentDesiredModel::LocalToAgent(projection_path.clone()),
-        ..BalancerDesiredState::default()
-    })
-    .await
-    .expect("a single-agent cluster must start");
+    let mut cluster =
+        start_single_agent_cluster_with_desired_state(desired_state_with_multimodal_projection(
+            BalancerDesiredState {
+                model: ModelSource::HuggingFace(reference).into_agent_desired_model(),
+                ..BalancerDesiredState::default()
+            },
+            agent_desired_model_at(Path::new(&projection_path))
+                .expect("the missing projection path must be absolute"),
+        ))
+        .await
+        .expect("a single-agent cluster must start");
 
     cluster
         .wait_for_first_agent_issue(|issue| {
diff --git a/paddler_tests/tests/balancer_reports_one_request_id_across_a_chunked_embedding_batch.rs b/paddler_tests/tests/balancer_reports_one_request_id_across_a_chunked_embedding_batch.rs
index 9e9c6325a..3770b7fea 100644
--- a/paddler_tests/tests/balancer_reports_one_request_id_across_a_chunked_embedding_batch.rs
+++ b/paddler_tests/tests/balancer_reports_one_request_id_across_a_chunked_embedding_batch.rs
@@ -4,22 +4,21 @@ use std::num::NonZeroUsize;
 
 use tokio_util::sync::CancellationToken;
 
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
-use paddler_messaging::embedding_input_document::EmbeddingInputDocument;
+use paddler_inference_parameters::embedding_parameters::EmbeddingParameters;
 use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
 use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
 use paddler_test_cluster_harness::embedding_cluster_params::EmbeddingClusterParams;
+use paddler_tests::numbered_embedding_documents::numbered_embedding_documents;
 use paddler_tests::start_embedding_cluster::start_embedding_cluster;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn balancer_reports_one_request_id_across_a_chunked_embedding_batch() {
     let cluster = start_embedding_cluster(EmbeddingClusterParams {
         agents: vec![AgentConfig::single(2)],
-        inference_parameters: InferenceParameters {
+        embedding_parameters: EmbeddingParameters {
             embedding_batch_size: NonZeroUsize::MIN,
-            enable_embeddings: true,
-            ..InferenceParameters::deterministic()
+            ..EmbeddingParameters::default()
         },
         ..EmbeddingClusterParams::default()
     })
@@ -30,12 +29,7 @@ async fn balancer_reports_one_request_id_across_a_chunked_embedding_batch() {
         .generate_embedding_batch(
             CancellationToken::new(),
             &GenerateEmbeddingBatchParams {
-                input_batch: (0..2)
-                    .map(|index| EmbeddingInputDocument {
-                        content: format!("Document number {index}."),
-                        id: format!("doc-{index}"),
-                    })
-                    .collect(),
+                input_batch: numbered_embedding_documents(2),
                 normalization_method: EmbeddingNormalizationMethod::None,
             },
         )
diff --git a/paddler_tests/tests/balancer_reports_the_huggingface_mmproj_of_its_desired_state.rs b/paddler_tests/tests/balancer_reports_the_huggingface_mmproj_of_its_desired_state.rs
new file mode 100644
index 000000000..abcea7791
--- /dev/null
+++ b/paddler_tests/tests/balancer_reports_the_huggingface_mmproj_of_its_desired_state.rs
@@ -0,0 +1,49 @@
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_model_source::model_source::ModelSource;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_test_cluster_harness::model_card::ModelCard;
+use paddler_test_cluster_harness::model_card::smolvlm2_256m::smolvlm2_256m;
+use paddler_test_cluster_harness::model_card::smolvlm2_256m_mmproj::smolvlm2_256m_mmproj;
+use paddler_tests::desired_state_with_multimodal_projection::desired_state_with_multimodal_projection;
+use paddler_tests::start_cluster::start_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn balancer_reports_the_huggingface_mmproj_of_its_desired_state() {
+    let ModelCard {
+        reference: primary_reference,
+    } = smolvlm2_256m();
+    let ModelCard {
+        reference: mmproj_reference,
+    } = smolvlm2_256m_mmproj();
+
+    let desired_state = desired_state_with_multimodal_projection(
+        BalancerDesiredState {
+            model: ModelSource::HuggingFace(primary_reference).into_agent_desired_model(),
+            ..BalancerDesiredState::default()
+        },
+        ModelSource::HuggingFace(mmproj_reference).into_agent_desired_model(),
+    );
+
+    let cluster = start_cluster(ClusterParams {
+        desired_state: ClusterDesiredState::Apply(Box::new(desired_state.clone())),
+        ..ClusterParams::without_agents()
+    })
+    .await
+    .expect("the cluster must start");
+
+    let retrieved = cluster
+        .client_management
+        .get_balancer_desired_state(CancellationToken::new())
+        .await
+        .expect("failed to read balancer desired state");
+
+    assert_eq!(retrieved, desired_state);
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/balancer_reports_the_local_mmproj_path_of_its_desired_state.rs b/paddler_tests/tests/balancer_reports_the_local_mmproj_path_of_its_desired_state.rs
new file mode 100644
index 000000000..524546d6d
--- /dev/null
+++ b/paddler_tests/tests/balancer_reports_the_local_mmproj_path_of_its_desired_state.rs
@@ -0,0 +1,49 @@
+use std::path::Path;
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_model_source::model_source::ModelSource;
+use paddler_test_cluster_harness::agent_desired_model_at::agent_desired_model_at;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_test_cluster_harness::model_card::ModelCard;
+use paddler_test_cluster_harness::model_card::smolvlm2_256m::smolvlm2_256m;
+use paddler_tests::desired_state_with_multimodal_projection::desired_state_with_multimodal_projection;
+use paddler_tests::start_cluster::start_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn balancer_reports_the_local_mmproj_path_of_its_desired_state() {
+    let ModelCard { reference } = smolvlm2_256m();
+
+    let local_mmproj_path = "/tmp/test-mmproj.gguf".to_owned();
+
+    let desired_state = desired_state_with_multimodal_projection(
+        BalancerDesiredState {
+            model: ModelSource::HuggingFace(reference).into_agent_desired_model(),
+            ..BalancerDesiredState::default()
+        },
+        agent_desired_model_at(Path::new(&local_mmproj_path))
+            .expect("the multimodal projection path must be absolute"),
+    );
+
+    let cluster = start_cluster(ClusterParams {
+        desired_state: ClusterDesiredState::Apply(Box::new(desired_state.clone())),
+        ..ClusterParams::without_agents()
+    })
+    .await
+    .expect("the cluster must start");
+
+    let retrieved = cluster
+        .client_management
+        .get_balancer_desired_state(CancellationToken::new())
+        .await
+        .expect("failed to read balancer desired state");
+
+    assert_eq!(retrieved, desired_state);
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/balancer_reports_the_url_model_of_its_desired_state.rs b/paddler_tests/tests/balancer_reports_the_url_model_of_its_desired_state.rs
new file mode 100644
index 000000000..e07bf38b6
--- /dev/null
+++ b/paddler_tests/tests/balancer_reports_the_url_model_of_its_desired_state.rs
@@ -0,0 +1,35 @@
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::agent_desired_model::AgentDesiredModel;
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_tests::start_cluster::start_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn balancer_reports_the_url_model_of_its_desired_state() {
+    let configured_url = "https://example.invalid/persisted-model.gguf".to_owned();
+
+    let cluster = start_cluster(ClusterParams {
+        desired_state: ClusterDesiredState::Apply(Box::new(BalancerDesiredState {
+            model: AgentDesiredModel::Uri(configured_url.clone()),
+            ..BalancerDesiredState::default()
+        })),
+        ..ClusterParams::without_agents()
+    })
+    .await
+    .expect("the cluster must start");
+
+    let retrieved = cluster
+        .client_management
+        .get_balancer_desired_state(CancellationToken::new())
+        .await
+        .expect("failed to read balancer desired state");
+
+    assert_eq!(retrieved.model, AgentDesiredModel::Uri(configured_url));
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/balancer_reports_unable_to_find_chat_template_for_embedding_model.rs b/paddler_tests/tests/balancer_reports_unable_to_find_chat_template_for_embedding_model.rs
index 6f6e24d7d..ed82bb251 100644
--- a/paddler_tests/tests/balancer_reports_unable_to_find_chat_template_for_embedding_model.rs
+++ b/paddler_tests/tests/balancer_reports_unable_to_find_chat_template_for_embedding_model.rs
@@ -1,28 +1,37 @@
 #![cfg(feature = "tests_that_use_llms")]
 
-use paddler_messaging::agent_desired_model::AgentDesiredModel;
+use std::mem::discriminant;
+
 use paddler_messaging::agent_issue::AgentIssue;
+use paddler_messaging::agent_issue_params::model_path::ModelPath;
 use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-use paddler_test_cluster_harness::model_card::ModelCard;
+use paddler_model_source::model_source::ModelSource;
 use paddler_test_cluster_harness::model_card::nomic_embed_text_v1_5::nomic_embed_text_v1_5;
 use paddler_tests::start_single_agent_cluster_with_desired_state::start_single_agent_cluster_with_desired_state;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn balancer_reports_unable_to_find_chat_template_for_embedding_model() {
-    let ModelCard { reference } = nomic_embed_text_v1_5();
+    let model_card = nomic_embed_text_v1_5();
+    let expected_issue = AgentIssue::UnableToFindChatTemplate(ModelPath::from(
+        model_card
+            .cached_path()
+            .expect("the model must be in the Hugging Face cache")
+            .as_path(),
+    ));
     let mut cluster = start_single_agent_cluster_with_desired_state(BalancerDesiredState {
-        model: AgentDesiredModel::HuggingFace(reference),
+        model: ModelSource::HuggingFace(model_card.reference).into_agent_desired_model(),
         ..BalancerDesiredState::default()
     })
     .await
     .expect("a single-agent cluster must start");
 
-    cluster
-        .wait_for_first_agent_issue(|issue| {
-            matches!(issue, AgentIssue::UnableToFindChatTemplate(_))
-        })
+    let snapshot = cluster
+        .wait_for_first_agent_issue(|issue| discriminant(issue) == discriminant(&expected_issue))
         .await
         .expect("the agent must report UnableToFindChatTemplate for an embedding-only model");
+    let issues = &snapshot.agents[0].status.issues;
+
+    assert!(issues.contains(&expected_issue), "{issues:?}");
 
     cluster
         .shutdown()
diff --git a/paddler_tests/tests/balancer_returns_503_when_request_buffering_disabled.rs b/paddler_tests/tests/balancer_returns_503_when_request_buffering_disabled.rs
index 00ac5ffc5..390cf3dd4 100644
--- a/paddler_tests/tests/balancer_returns_503_when_request_buffering_disabled.rs
+++ b/paddler_tests/tests/balancer_returns_503_when_request_buffering_disabled.rs
@@ -1,23 +1,23 @@
 use std::num::NonZeroU32;
-use std::time::Duration;
 
 use futures_util::StreamExt as _;
 use tokio_util::sync::CancellationToken;
 
 use paddler_messaging::inference_client::message::Message;
+use paddler_messaging::jsonrpc::error::Error as JsonRpcError;
+use paddler_messaging::jsonrpc::error_envelope::ErrorEnvelope;
 use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_tests::short_buffered_request_timeout::SHORT_BUFFERED_REQUEST_TIMEOUT;
 use paddler_tests::start_cluster::start_cluster;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn balancer_returns_503_when_request_buffering_disabled() {
     let mut cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
-        buffered_request_timeout: Duration::from_millis(50),
+        buffered_request_timeout: SHORT_BUFFERED_REQUEST_TIMEOUT,
         max_buffered_requests: 0,
-        ..ClusterParams::default()
+        ..ClusterParams::without_agents()
     })
     .await
     .expect("the cluster must start");
@@ -47,17 +47,15 @@ async fn balancer_returns_503_when_request_buffering_disabled() {
         .expect("inference stream must yield a message")
         .expect("the message must be readable");
 
-    match message {
-        Message::Error(envelope) => {
-            assert_eq!(envelope.error.code, 503);
-        }
-        Message::Response(_) => {
-            panic!("expected buffer overflow error, got success");
-        }
-        Message::Notification(_) => {
-            panic!("unexpected token-generation-mode notification");
-        }
-    }
+    let expected_error = JsonRpcError {
+        code: 503,
+        description: "Buffered requests overflow".to_owned(),
+    };
+
+    assert!(matches!(
+        message,
+        Message::Error(ErrorEnvelope { error, .. }) if error == expected_error
+    ));
 
     cluster
         .shutdown()
diff --git a/paddler_tests/tests/balancer_returns_504_when_inference_item_timeout_is_zero.rs b/paddler_tests/tests/balancer_returns_504_when_inference_item_timeout_is_zero.rs
index fe37999c6..16aafacb8 100644
--- a/paddler_tests/tests/balancer_returns_504_when_inference_item_timeout_is_zero.rs
+++ b/paddler_tests/tests/balancer_returns_504_when_inference_item_timeout_is_zero.rs
@@ -7,8 +7,11 @@ use futures_util::StreamExt as _;
 use tokio_util::sync::CancellationToken;
 
 use paddler_messaging::inference_client::message::Message;
+use paddler_messaging::jsonrpc::error::Error as JsonRpcError;
+use paddler_messaging::jsonrpc::error_envelope::ErrorEnvelope;
 use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
 use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
 use paddler_tests::start_cluster::start_cluster;
@@ -19,7 +22,7 @@ async fn balancer_returns_504_when_inference_item_timeout_is_zero() {
         agents: AgentConfig::uniform(1, 2),
         inference_item_timeout: Duration::ZERO,
         wait_for_slots_ready: true,
-        desired_state: Some(qwen3_0_6b().into_desired_state()),
+        desired_state: ClusterDesiredState::Apply(Box::new(qwen3_0_6b().into_desired_state())),
         ..ClusterParams::default()
     })
     .await
@@ -43,17 +46,17 @@ async fn balancer_returns_504_when_inference_item_timeout_is_zero() {
         .expect("inference stream must yield a message")
         .expect("the message must be readable");
 
-    match message {
-        Message::Error(envelope) => {
-            assert_eq!(envelope.error.code, 504);
-        }
-        Message::Response(_) => {
-            panic!("expected timeout error, got success");
-        }
-        Message::Notification(_) => {
-            panic!("unexpected token-generation-mode notification");
-        }
-    }
+    let expected_error = JsonRpcError {
+        code: 504,
+        description: "Inference timed out after 0ms waiting for next token. \
+                    Increase --inference-item-timeout if the prompt requires longer processing."
+            .to_owned(),
+    };
+
+    assert!(matches!(
+        message,
+        Message::Error(ErrorEnvelope { error, .. }) if error == expected_error
+    ));
 
     cluster
         .shutdown()
diff --git a/paddler_tests/tests/balancer_returns_504_when_no_agents_registered.rs b/paddler_tests/tests/balancer_returns_504_when_no_agents_registered.rs
index c8225632e..cd4b26fe9 100644
--- a/paddler_tests/tests/balancer_returns_504_when_no_agents_registered.rs
+++ b/paddler_tests/tests/balancer_returns_504_when_no_agents_registered.rs
@@ -1,22 +1,22 @@
 use std::num::NonZeroU32;
-use std::time::Duration;
 
 use futures_util::StreamExt as _;
 use tokio_util::sync::CancellationToken;
 
 use paddler_messaging::inference_client::message::Message;
+use paddler_messaging::jsonrpc::error::Error as JsonRpcError;
+use paddler_messaging::jsonrpc::error_envelope::ErrorEnvelope;
 use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_tests::short_buffered_request_timeout::SHORT_BUFFERED_REQUEST_TIMEOUT;
 use paddler_tests::start_cluster::start_cluster;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn balancer_returns_504_when_no_agents_registered() {
     let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        buffered_request_timeout: Duration::from_millis(50),
+        buffered_request_timeout: SHORT_BUFFERED_REQUEST_TIMEOUT,
         max_buffered_requests: 1,
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
+        ..ClusterParams::without_agents()
     })
     .await
     .expect("a cluster without agents must start");
@@ -37,7 +37,13 @@ async fn balancer_returns_504_when_no_agents_registered() {
         .expect("the inference stream must yield a message")
         .expect("the inference stream message must be readable");
 
-    assert!(matches!(message, Message::Error(envelope) if envelope.error.code == 504));
+    assert!(matches!(
+        message,
+        Message::Error(ErrorEnvelope { error, .. }) if error == JsonRpcError {
+            code: 504,
+            description: "Waiting for available slot timed out".to_owned(),
+        }
+    ));
 
     cluster
         .shutdown()
diff --git a/paddler_tests/tests/balancer_serves_a_conversation_over_websocket.rs b/paddler_tests/tests/balancer_serves_a_conversation_over_websocket.rs
new file mode 100644
index 000000000..1531720bb
--- /dev/null
+++ b/paddler_tests/tests/balancer_serves_a_conversation_over_websocket.rs
@@ -0,0 +1,64 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::num::NonZeroU32;
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::conversation_history::ConversationHistory;
+use paddler_messaging::conversation_message::ConversationMessage;
+use paddler_messaging::conversation_message_content::ConversationMessageContent;
+use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::collect_generated_tokens::collect_generated_tokens;
+use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
+
+const MAX_TOKENS: u32 = 16;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn balancer_serves_a_conversation_over_websocket() {
+    let cluster = start_cluster_with_qwen3(AgentConfig::uniform(1, 1))
+        .await
+        .expect("the cluster must start");
+
+    let collected = collect_generated_tokens(
+        cluster
+            .client_inference
+            .continue_from_conversation_history(
+                CancellationToken::new(),
+                ContinueFromConversationHistoryParams {
+                    add_generation_prompt: true,
+                    conversation_history: ConversationHistory::new(vec![
+                        ConversationMessage::Message {
+                            content: ConversationMessageContent::Text(
+                                "What is the capital of France?".to_owned(),
+                            ),
+                            role: "user".to_owned(),
+                        },
+                    ]),
+                    enable_thinking: false,
+                    grammar: None,
+                    max_tokens: NonZeroU32::new(MAX_TOKENS).expect("the token budget is non-zero"),
+                    parse_tool_calls: false,
+                    tools: Vec::new(),
+                },
+            )
+            .await
+            .expect("the conversation must be accepted over the inference socket"),
+    )
+    .await
+    .expect("the conversation tokens must be collected");
+
+    assert!(
+        collected
+            .summary()
+            .expect("the conversation must finish with a summary")
+            .usage
+            .completion_tokens()
+            > 0
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/balancer_serves_every_concurrent_request_on_a_single_slot_agent.rs b/paddler_tests/tests/balancer_serves_every_concurrent_request_on_a_single_slot_agent.rs
index 7f00eaeeb..799bb439d 100644
--- a/paddler_tests/tests/balancer_serves_every_concurrent_request_on_a_single_slot_agent.rs
+++ b/paddler_tests/tests/balancer_serves_every_concurrent_request_on_a_single_slot_agent.rs
@@ -6,9 +6,9 @@ use std::num::NonZeroU32;
 use futures_util::future::try_join_all;
 use tokio_util::sync::CancellationToken;
 
-use paddler_messaging::generated_token_result::GeneratedTokenResult;
 use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
 use paddler_test_cluster_harness::collect_generated_tokens::collect_generated_tokens;
 use paddler_tests::qwen3_desired_state::qwen3_desired_state;
@@ -20,7 +20,7 @@ const CONCURRENT_REQUESTS: usize = 256;
 async fn balancer_serves_every_concurrent_request_on_a_single_slot_agent() {
     let cluster = start_cluster(ClusterParams {
         agents: AgentConfig::uniform(1, 1),
-        desired_state: Some(qwen3_desired_state()),
+        desired_state: ClusterDesiredState::Apply(Box::new(qwen3_desired_state())),
         max_buffered_requests: u64::try_from(CONCURRENT_REQUESTS)
             .expect("the value must fit its target type"),
         wait_for_slots_ready: true,
@@ -53,13 +53,9 @@ async fn balancer_serves_every_concurrent_request_on_a_single_slot_agent() {
     .expect("every concurrent request must succeed");
 
     for collected in collected_requests {
-        assert!(matches!(
-            collected
-                .token_results
-                .last()
-                .map(|token_result_with_producer| &token_result_with_producer.token_result),
-            Some(GeneratedTokenResult::Done(_))
-        ));
+        collected
+            .summary()
+            .expect("the generation must finish with a summary");
     }
 
     cluster
diff --git a/paddler_tests/tests/balancer_serves_inference_over_websocket.rs b/paddler_tests/tests/balancer_serves_inference_over_websocket.rs
index 15baa79a8..7c26e907b 100644
--- a/paddler_tests/tests/balancer_serves_inference_over_websocket.rs
+++ b/paddler_tests/tests/balancer_serves_inference_over_websocket.rs
@@ -1,5 +1,6 @@
 #![cfg(feature = "tests_that_use_llms")]
 
+use std::mem::discriminant;
 use std::num::NonZeroU32;
 
 use futures_util::StreamExt as _;
@@ -7,8 +8,10 @@ use tokio_util::sync::CancellationToken;
 
 use paddler_messaging::inference_client::message::Message as InferenceMessage;
 use paddler_messaging::inference_client::response::Response;
+use paddler_messaging::jsonrpc::response_envelope::ResponseEnvelope;
 use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_tests::generation_done_discriminant::generation_done_discriminant;
 use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
 
 #[tokio::test(flavor = "multi_thread")]
@@ -17,7 +20,7 @@ async fn balancer_serves_inference_over_websocket() {
         .await
         .expect("the cluster must start");
 
-    let mut stream = cluster
+    let stream = cluster
         .client_inference
         .continue_from_raw_prompt(
             CancellationToken::new(),
@@ -30,33 +33,33 @@ async fn balancer_serves_inference_over_websocket() {
         .await
         .expect("the inference request must be accepted");
 
-    let mut token_count: usize = 0;
+    let messages: Vec<InferenceMessage> = stream
+        .map(|message_result| message_result.expect("the message must be readable"))
+        .collect()
+        .await;
 
-    while let Some(message_result) = stream.next().await {
-        match message_result.expect("the message must be readable") {
-            InferenceMessage::Response(envelope) => match envelope.response {
-                Response::GeneratedToken(generated_token_result) => {
-                    if generated_token_result.is_token() {
-                        token_count += 1;
-                    }
-                }
-                Response::Embedding(_) => {
-                    panic!("inference over websocket produced an unexpected response variant")
-                }
-            },
-            InferenceMessage::Error(envelope) => {
-                panic!(
-                    "inference over websocket failed: code {}, description {:?}",
-                    envelope.error.code, envelope.error.description
-                )
-            }
-            InferenceMessage::Notification(_) => {
-                panic!("inference over websocket produced an unexpected notification")
-            }
-        }
-    }
-
-    assert!(token_count > 0);
+    assert!(messages.iter().all(|message| matches!(
+        message,
+        InferenceMessage::Response(ResponseEnvelope {
+            generated_by: Some(agent_name),
+            response: Response::GeneratedToken(_),
+            ..
+        }) if agent_name == "test-agent-0"
+    )));
+    assert!(messages.iter().any(|message| matches!(
+        message,
+        InferenceMessage::Response(ResponseEnvelope {
+            response: Response::GeneratedToken(token_result),
+            ..
+        }) if token_result.is_token()
+    )));
+    assert!(matches!(
+        messages.last(),
+        Some(InferenceMessage::Response(ResponseEnvelope {
+            response: Response::GeneratedToken(token_result),
+            ..
+        })) if discriminant(token_result) == generation_done_discriminant()
+    ));
 
     cluster
         .shutdown()
diff --git a/paddler_tests/tests/balancer_serves_request_after_agent_with_capacity_registers.rs b/paddler_tests/tests/balancer_serves_request_after_agent_with_capacity_registers.rs
index 6601cd85c..ed0f862c4 100644
--- a/paddler_tests/tests/balancer_serves_request_after_agent_with_capacity_registers.rs
+++ b/paddler_tests/tests/balancer_serves_request_after_agent_with_capacity_registers.rs
@@ -1,61 +1,27 @@
 #![cfg(feature = "tests_that_use_llms")]
 
 use std::num::NonZeroU32;
-use std::time::Duration;
 
-use futures_util::StreamExt as _;
 use tokio_util::sync::CancellationToken;
 
-use paddler_messaging::inference_client::message::Message;
 use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_test_cluster_harness::collect_generated_tokens::collect_generated_tokens;
 use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
 use paddler_tests::start_cluster::start_cluster;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn balancer_serves_request_after_agent_with_capacity_registers() {
     let mut cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
-        buffered_request_timeout: Duration::from_millis(50),
         max_buffered_requests: 10,
-        desired_state: Some(qwen3_0_6b().into_desired_state()),
-        ..ClusterParams::default()
+        desired_state: ClusterDesiredState::Apply(Box::new(qwen3_0_6b().into_desired_state())),
+        ..ClusterParams::without_agents()
     })
     .await
     .expect("the cluster must start");
 
-    let mut early_stream = cluster
-        .continue_from_raw_prompt_stream(
-            CancellationToken::new(),
-            &ContinueFromRawPromptParams {
-                grammar: None,
-                max_tokens: NonZeroU32::new(10).unwrap(),
-                raw_prompt: "Hello".to_owned(),
-            },
-        )
-        .await
-        .expect("the inference request must be accepted");
-
-    let early_message = early_stream
-        .next()
-        .await
-        .expect("inference stream must yield a message")
-        .expect("the message must be readable");
-
-    match early_message {
-        Message::Error(envelope) => {
-            assert_eq!(envelope.error.code, 504);
-        }
-        Message::Response(_) => {
-            panic!("expected timeout before agent registered");
-        }
-        Message::Notification(_) => {
-            panic!("unexpected token-generation-mode notification");
-        }
-    }
-
     cluster
         .spawn_additional_agent(&AgentConfig {
             name: "capacity-agent".to_owned(),
@@ -70,12 +36,12 @@ async fn balancer_serves_request_after_agent_with_capacity_registers() {
                 && snapshot
                     .agents
                     .iter()
-                    .any(|agent| agent.status.slots_total >= 4)
+                    .any(|agent| agent.status.runtime.slots_total() >= 4)
         })
         .await
         .expect("agent should register with 4 slots");
 
-    let mut later_stream = cluster
+    let later_stream = cluster
         .continue_from_raw_prompt_stream(
             CancellationToken::new(),
             &ContinueFromRawPromptParams {
@@ -87,23 +53,11 @@ async fn balancer_serves_request_after_agent_with_capacity_registers() {
         .await
         .expect("the inference request must be accepted");
 
-    let later_message = later_stream
-        .next()
+    collect_generated_tokens(later_stream)
         .await
-        .expect("inference stream must yield a message")
-        .expect("the message must be readable");
-
-    match later_message {
-        Message::Error(envelope) => {
-            panic!(
-                "a request must be served once an agent with capacity registered, got {envelope:?}"
-            );
-        }
-        Message::Response(_) => {}
-        Message::Notification(_) => {
-            panic!("unexpected token-generation-mode notification");
-        }
-    }
+        .expect("a request must be served once an agent with capacity registered")
+        .summary()
+        .expect("the served request must finish with a summary");
 
     cluster
         .shutdown()
diff --git a/paddler_tests/tests/balancer_serves_requests_from_different_connections_that_share_a_request_id.rs b/paddler_tests/tests/balancer_serves_requests_from_different_connections_that_share_a_request_id.rs
index bff3c46b9..9abda66bb 100644
--- a/paddler_tests/tests/balancer_serves_requests_from_different_connections_that_share_a_request_id.rs
+++ b/paddler_tests/tests/balancer_serves_requests_from_different_connections_that_share_a_request_id.rs
@@ -6,6 +6,7 @@ use paddler_messaging::inference_client::message::Message as InferenceClientMess
 use paddler_messaging::inference_client::response::Response;
 use paddler_messaging::jsonrpc::response_envelope::ResponseEnvelope;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
 use paddler_test_cluster_harness::raw_inference_socket::RawInferenceSocket;
 use paddler_tests::qwen3_desired_state::qwen3_desired_state;
@@ -29,7 +30,7 @@ const fn is_generated_token(message: &InferenceClientMessage) -> bool {
 async fn balancer_serves_requests_from_different_connections_that_share_a_request_id() {
     let cluster = start_cluster(ClusterParams {
         agents: AgentConfig::uniform(1, 2),
-        desired_state: Some(qwen3_desired_state()),
+        desired_state: ClusterDesiredState::Apply(Box::new(qwen3_desired_state())),
         wait_for_slots_ready: true,
         ..ClusterParams::default()
     })
diff --git a/paddler_tests/tests/balancer_shutdown_does_not_wait_for_an_open_sse_subscriber.rs b/paddler_tests/tests/balancer_shutdown_does_not_wait_for_an_open_sse_subscriber.rs
index f12fe23ab..9927fecbf 100644
--- a/paddler_tests/tests/balancer_shutdown_does_not_wait_for_an_open_sse_subscriber.rs
+++ b/paddler_tests/tests/balancer_shutdown_does_not_wait_for_an_open_sse_subscriber.rs
@@ -6,13 +6,9 @@ use paddler_tests::start_cluster::start_cluster;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn balancer_shutdown_does_not_wait_for_an_open_sse_subscriber() {
-    let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("the cluster must start");
+    let cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("the cluster must start");
     let mut sse_stream = cluster
         .client_management
         .get_buffered_requests_stream(CancellationToken::new())
diff --git a/paddler_tests/tests/changing_the_context_size_restarts_serving_on_the_loaded_model.rs b/paddler_tests/tests/changing_the_context_size_restarts_serving_on_the_loaded_model.rs
new file mode 100644
index 000000000..b704cbfc4
--- /dev/null
+++ b/paddler_tests/tests/changing_the_context_size_restarts_serving_on_the_loaded_model.rs
@@ -0,0 +1,59 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::num::NonZeroU32;
+
+use paddler_inference_parameters::model_runtime_parameters::ModelRuntimeParameters;
+use paddler_messaging::agent_state_application_status::AgentStateApplicationStatus;
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_messaging::decision_settings::DecisionSettings;
+use paddler_tests::kev_checkpoint::KevCheckpoint;
+
+use crate::serving_pipeline_arbiter::ServingPipelineArbiter;
+use crate::swappable_model_file::SwappableModelFile;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn changing_the_context_size_restarts_serving_on_the_loaded_model() {
+    let swappable_model_file =
+        SwappableModelFile::link_to(&KevCheckpoint::Kev0_8b.target_path("model.gguf"));
+    let initial_desired_state = BalancerDesiredState {
+        decision: DecisionSettings::default(),
+        model: swappable_model_file.agent_desired_model(),
+        ..KevCheckpoint::Kev0_8b
+            .decision_desired_state()
+            .into_balancer_desired_state()
+            .expect("the kev desired state must name absolute paths")
+    };
+    let mut serving_pipeline_arbiter = ServingPipelineArbiter::new(1);
+
+    serving_pipeline_arbiter
+        .serve(initial_desired_state.clone())
+        .await;
+
+    swappable_model_file.replace_with_an_unloadable_file();
+
+    serving_pipeline_arbiter
+        .serve(BalancerDesiredState {
+            model_runtime_parameters: ModelRuntimeParameters {
+                context_size: NonZeroU32::new(
+                    initial_desired_state
+                        .model_runtime_parameters
+                        .context_size
+                        .get()
+                        / 2,
+                )
+                .expect("half of the initial context must hold tokens"),
+                ..initial_desired_state.model_runtime_parameters.clone()
+            },
+            ..initial_desired_state
+        })
+        .await;
+    assert_eq!(
+        serving_pipeline_arbiter
+            .inference_runtime_context
+            .slot_aggregated_status
+            .get_state_application_status(),
+        AgentStateApplicationStatus::Applied
+    );
+
+    serving_pipeline_arbiter.shut_down().await;
+}
diff --git a/paddler_tests/tests/chat_template_drains_in_flight_inference_before_swap.rs b/paddler_tests/tests/chat_template_drains_in_flight_inference_before_swap.rs
index 95dc58f3b..1f06b1ffb 100644
--- a/paddler_tests/tests/chat_template_drains_in_flight_inference_before_swap.rs
+++ b/paddler_tests/tests/chat_template_drains_in_flight_inference_before_swap.rs
@@ -4,7 +4,6 @@ use std::num::NonZeroU32;
 
 use tokio_util::sync::CancellationToken;
 
-use paddler_messaging::balancer_desired_state::BalancerDesiredState;
 use paddler_messaging::chat_template::ChatTemplate;
 use paddler_messaging::conversation_history::ConversationHistory;
 use paddler_messaging::conversation_message::ConversationMessage;
@@ -14,9 +13,10 @@ use paddler_messaging::request_params::continue_from_conversation_history_params
 use paddler_test_cluster_harness::agent_config::AgentConfig;
 use paddler_test_cluster_harness::collect_generated_tokens::collect_generated_tokens;
 use paddler_test_cluster_harness::unending_grammar::unending_grammar;
+use paddler_tests::desired_state_with_chat_template_override::desired_state_with_chat_template_override;
 use paddler_tests::start_cluster_with_qwen3_and_context_size::start_cluster_with_qwen3_and_context_size;
 
-const CONTEXT_SIZE_THAT_ENDS_THE_IN_FLIGHT_GENERATION: u32 = 256;
+const CONTEXT_SIZE_THAT_ENDS_THE_IN_FLIGHT_GENERATION: NonZeroU32 = NonZeroU32::new(256).unwrap();
 
 #[tokio::test(flavor = "multi_thread")]
 async fn chat_template_drains_in_flight_inference_before_swap() {
@@ -39,12 +39,14 @@ async fn chat_template_drains_in_flight_inference_before_swap() {
             CancellationToken::new(),
             &ContinueFromConversationHistoryParams {
                 add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Text(
-                        "The capital of France is".to_owned(),
-                    ),
-                    role: "user".to_owned(),
-                }]),
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Text(
+                            "The capital of France is".to_owned(),
+                        ),
+                        role: "user".to_owned(),
+                    },
+                ]),
                 enable_thinking: false,
                 grammar: Some(unending_grammar()),
                 max_tokens: NonZeroU32::MAX,
@@ -64,11 +66,10 @@ async fn chat_template_drains_in_flight_inference_before_swap() {
         .client_management
         .put_balancer_desired_state(
             CancellationToken::new(),
-            &BalancerDesiredState {
-                chat_template_override: Some(swapped_template.clone()),
-                use_chat_template_override: true,
-                ..initial_desired_state
-            },
+            &desired_state_with_chat_template_override(
+                initial_desired_state,
+                swapped_template.clone(),
+            ),
         )
         .await
         .expect("the balancer must accept the desired state");
diff --git a/paddler_tests/tests/chat_template_override_applied_to_embedding_model.rs b/paddler_tests/tests/chat_template_override_applied_to_embedding_model.rs
index 49e773102..83d3e23de 100644
--- a/paddler_tests/tests/chat_template_override_applied_to_embedding_model.rs
+++ b/paddler_tests/tests/chat_template_override_applied_to_embedding_model.rs
@@ -4,6 +4,7 @@ use tokio_util::sync::CancellationToken;
 
 use paddler_messaging::chat_template::ChatTemplate;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
 use paddler_tests::nomic_embed_desired_state_with_chat_template_override::nomic_embed_desired_state_with_chat_template_override;
 use paddler_tests::start_cluster::start_cluster;
@@ -17,8 +18,8 @@ async fn chat_template_override_applied_to_embedding_model() {
     let mut cluster = start_cluster(ClusterParams {
         agents: AgentConfig::uniform(1, 1),
         wait_for_slots_ready: false,
-        desired_state: Some(nomic_embed_desired_state_with_chat_template_override(
-            chat_template.clone(),
+        desired_state: ClusterDesiredState::Apply(Box::new(
+            nomic_embed_desired_state_with_chat_template_override(chat_template.clone()),
         )),
         ..ClusterParams::default()
     })
diff --git a/paddler_tests/tests/chat_template_override_is_reported_and_serves_conversations.rs b/paddler_tests/tests/chat_template_override_is_reported_and_serves_conversations.rs
new file mode 100644
index 000000000..90253df85
--- /dev/null
+++ b/paddler_tests/tests/chat_template_override_is_reported_and_serves_conversations.rs
@@ -0,0 +1,104 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::num::NonZeroU32;
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_inference_parameters::all_gpu_layers::ALL_GPU_LAYERS;
+use paddler_inference_parameters::model_runtime_parameters::ModelRuntimeParameters;
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_messaging::chat_template::ChatTemplate;
+use paddler_messaging::conversation_history::ConversationHistory;
+use paddler_messaging::conversation_message::ConversationMessage;
+use paddler_messaging::conversation_message_content::ConversationMessageContent;
+use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
+use paddler_model_source::model_source::ModelSource;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_test_cluster_harness::model_card::ModelCard;
+use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
+use paddler_tests::desired_state_with_chat_template_override::desired_state_with_chat_template_override;
+use paddler_tests::start_cluster::start_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn chat_template_override_is_reported_and_serves_conversations() {
+    let ModelCard { reference } = qwen3_0_6b();
+
+    let chat_template = ChatTemplate {
+        content: "{{ messages[0].content }}".to_owned(),
+    };
+
+    let cluster = start_cluster(ClusterParams {
+        agents: AgentConfig::uniform(1, 1),
+        wait_for_slots_ready: true,
+        desired_state: ClusterDesiredState::Apply(Box::new(
+            desired_state_with_chat_template_override(
+                BalancerDesiredState {
+                    model_runtime_parameters: ModelRuntimeParameters {
+                        n_gpu_layers: ALL_GPU_LAYERS,
+                        ..ModelRuntimeParameters::default()
+                    },
+                    model: ModelSource::HuggingFace(reference).into_agent_desired_model(),
+                    ..BalancerDesiredState::default()
+                },
+                chat_template.clone(),
+            ),
+        )),
+        ..ClusterParams::default()
+    })
+    .await
+    .expect("the cluster must start");
+
+    let agent_id = cluster
+        .agent_ids
+        .first()
+        .expect("cluster must have one registered agent")
+        .clone();
+
+    let retrieved = cluster
+        .client_management
+        .get_chat_template_override(CancellationToken::new(), &agent_id)
+        .await
+        .expect("failed to read chat template override");
+
+    assert_eq!(retrieved, Some(chat_template));
+
+    let collected = cluster
+        .continue_from_conversation_history(
+            CancellationToken::new(),
+            &ContinueFromConversationHistoryParams {
+                add_generation_prompt: true,
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Text(
+                            "The capital of France is".to_owned(),
+                        ),
+                        role: "user".to_owned(),
+                    },
+                ]),
+                enable_thinking: false,
+                grammar: None,
+                max_tokens: NonZeroU32::new(10).unwrap(),
+                parse_tool_calls: false,
+                tools: vec![],
+            },
+        )
+        .await
+        .expect("the inference request must be accepted");
+
+    let received_tokens = collected
+        .token_results
+        .iter()
+        .any(|result| result.token_result.is_token());
+
+    assert!(
+        received_tokens,
+        "override template should render the prompt and produce tokens"
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/chat_template_override_replaces_model_builtin.rs b/paddler_tests/tests/chat_template_override_replaces_model_builtin.rs
deleted file mode 100644
index 6c518d65e..000000000
--- a/paddler_tests/tests/chat_template_override_replaces_model_builtin.rs
+++ /dev/null
@@ -1,97 +0,0 @@
-#![cfg(feature = "tests_that_use_llms")]
-
-use std::num::NonZeroU32;
-
-use tokio_util::sync::CancellationToken;
-
-use paddler_inference_parameters::all_gpu_layers::ALL_GPU_LAYERS;
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
-use paddler_messaging::agent_desired_model::AgentDesiredModel;
-use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-use paddler_messaging::chat_template::ChatTemplate;
-use paddler_messaging::conversation_history::ConversationHistory;
-use paddler_messaging::conversation_message::ConversationMessage;
-use paddler_messaging::conversation_message_content::ConversationMessageContent;
-use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
-use paddler_test_cluster_harness::agent_config::AgentConfig;
-use paddler_test_cluster_harness::cluster_params::ClusterParams;
-use paddler_test_cluster_harness::model_card::ModelCard;
-use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
-use paddler_tests::start_cluster::start_cluster;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn chat_template_override_replaces_model_builtin() {
-    let ModelCard { reference } = qwen3_0_6b();
-
-    let chat_template = ChatTemplate {
-        content: "{{ messages[0].content }}".to_owned(),
-    };
-
-    let cluster = start_cluster(ClusterParams {
-        agents: AgentConfig::uniform(1, 1),
-        wait_for_slots_ready: true,
-        desired_state: Some(BalancerDesiredState {
-            chat_template_override: Some(chat_template.clone()),
-            inference_parameters: InferenceParameters {
-                n_gpu_layers: ALL_GPU_LAYERS,
-                ..InferenceParameters::deterministic()
-            },
-            model: AgentDesiredModel::HuggingFace(reference),
-            multimodal_projection: AgentDesiredModel::None,
-            use_chat_template_override: true,
-        }),
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("the cluster must start");
-
-    let agent_id = cluster
-        .agent_ids
-        .first()
-        .expect("cluster must have one registered agent")
-        .clone();
-
-    let retrieved = cluster
-        .client_management
-        .get_chat_template_override(CancellationToken::new(), &agent_id)
-        .await
-        .expect("failed to read chat template override");
-
-    assert_eq!(retrieved, Some(chat_template));
-
-    let collected = cluster
-        .continue_from_conversation_history(
-            CancellationToken::new(),
-            &ContinueFromConversationHistoryParams {
-                add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Text(
-                        "The capital of France is".to_owned(),
-                    ),
-                    role: "user".to_owned(),
-                }]),
-                enable_thinking: false,
-                grammar: None,
-                max_tokens: NonZeroU32::new(10).unwrap(),
-                parse_tool_calls: false,
-                tools: vec![],
-            },
-        )
-        .await
-        .expect("the inference request must be accepted");
-
-    let received_tokens = collected
-        .token_results
-        .iter()
-        .any(|result| result.token_result.is_token());
-
-    assert!(
-        received_tokens,
-        "override template should render the prompt and produce tokens"
-    );
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_tests/tests/chat_template_swaps_between_inference_calls.rs b/paddler_tests/tests/chat_template_swaps_between_inference_calls.rs
index 70bf49077..3d95d1ab6 100644
--- a/paddler_tests/tests/chat_template_swaps_between_inference_calls.rs
+++ b/paddler_tests/tests/chat_template_swaps_between_inference_calls.rs
@@ -5,7 +5,6 @@ use std::num::NonZeroU32;
 use anyhow::Result;
 use tokio_util::sync::CancellationToken;
 
-use paddler_messaging::balancer_desired_state::BalancerDesiredState;
 use paddler_messaging::chat_template::ChatTemplate;
 use paddler_messaging::conversation_history::ConversationHistory;
 use paddler_messaging::conversation_message::ConversationMessage;
@@ -13,23 +12,27 @@ use paddler_messaging::conversation_message_content::ConversationMessageContent;
 use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
 use paddler_test_cluster_harness::cluster::Cluster;
+use paddler_test_cluster_harness::collected_generated_tokens::CollectedGeneratedTokens;
+use paddler_tests::desired_state_with_chat_template_override::desired_state_with_chat_template_override;
 use paddler_tests::qwen3_desired_state::qwen3_desired_state;
 use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
 
 const CAPITAL_OF_FRANCE_MAX_TOKENS: NonZeroU32 = NonZeroU32::new(10).unwrap();
 
-async fn capital_of_france(cluster: &Cluster) -> Result<()> {
+async fn capital_of_france(cluster: &Cluster) -> Result<CollectedGeneratedTokens> {
     cluster
         .continue_from_conversation_history(
             CancellationToken::new(),
             &ContinueFromConversationHistoryParams {
                 add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Text(
-                        "The capital of France is".to_owned(),
-                    ),
-                    role: "user".to_owned(),
-                }]),
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Text(
+                            "The capital of France is".to_owned(),
+                        ),
+                        role: "user".to_owned(),
+                    },
+                ]),
                 enable_thinking: false,
                 grammar: None,
                 max_tokens: CAPITAL_OF_FRANCE_MAX_TOKENS,
@@ -37,10 +40,7 @@ async fn capital_of_france(cluster: &Cluster) -> Result<()> {
                 tools: vec![],
             },
         )
-        .await?
-        .summary()?;
-
-    Ok(())
+        .await
 }
 
 #[tokio::test(flavor = "multi_thread")]
@@ -59,17 +59,18 @@ async fn chat_template_swaps_between_inference_calls() {
 
     capital_of_france(&cluster)
         .await
-        .expect("the capital of France request must be answered");
+        .expect("the capital of France request must be answered")
+        .summary()
+        .expect("the capital of France answer must finish with a summary");
 
     cluster
         .client_management
         .put_balancer_desired_state(
             CancellationToken::new(),
-            &BalancerDesiredState {
-                chat_template_override: Some(swapped_template.clone()),
-                use_chat_template_override: true,
-                ..qwen3_desired_state()
-            },
+            &desired_state_with_chat_template_override(
+                qwen3_desired_state(),
+                swapped_template.clone(),
+            ),
         )
         .await
         .expect("the balancer must accept the desired state");
@@ -80,7 +81,9 @@ async fn chat_template_swaps_between_inference_calls() {
 
     capital_of_france(&cluster)
         .await
-        .expect("the capital of France request must be answered");
+        .expect("the capital of France request must be answered")
+        .summary()
+        .expect("the capital of France answer must finish with a summary");
 
     assert_eq!(
         cluster
diff --git a/paddler_tests/tests/client_inference_reconnects_after_the_balancer_closes_its_socket.rs b/paddler_tests/tests/client_inference_reconnects_after_the_balancer_closes_its_socket.rs
index 635c04820..bd223a947 100644
--- a/paddler_tests/tests/client_inference_reconnects_after_the_balancer_closes_its_socket.rs
+++ b/paddler_tests/tests/client_inference_reconnects_after_the_balancer_closes_its_socket.rs
@@ -1,6 +1,6 @@
+use std::mem::discriminant;
 use std::num::NonZeroU32;
 use std::num::NonZeroUsize;
-use std::time::Duration;
 
 use futures_util::StreamExt as _;
 use tokio_util::sync::CancellationToken;
@@ -12,6 +12,7 @@ use paddler_client::error::Error;
 use paddler_messaging::inference_client::message::Message;
 use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_tests::short_buffered_request_timeout::SHORT_BUFFERED_REQUEST_TIMEOUT;
 use paddler_tests::start_cluster::start_cluster;
 
 const MAX_TOKENS: NonZeroU32 = NonZeroU32::new(1).unwrap();
@@ -27,20 +28,15 @@ const fn raw_prompt_params(raw_prompt: String) -> ContinueFromRawPromptParams {
 #[tokio::test(flavor = "multi_thread")]
 async fn client_inference_reconnects_after_the_balancer_closes_its_socket() {
     let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        buffered_request_timeout: Duration::from_millis(50),
+        buffered_request_timeout: SHORT_BUFFERED_REQUEST_TIMEOUT,
         max_buffered_requests: 1,
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
+        ..ClusterParams::without_agents()
     })
     .await
     .expect("a cluster without agents must start");
     let client_inference = ClientInference::new(ClientInferenceParams {
         inference_socket_pool_size: NonZeroUsize::MIN,
-        url: cluster
-            .balancer
-            .inference_base_url()
-            .expect("the inference service must have a base URL"),
+        url: cluster.balancer.inference_base_url.clone(),
     });
 
     let oversized_request_outcome = client_inference
@@ -65,10 +61,12 @@ async fn client_inference_reconnects_after_the_balancer_closes_its_socket() {
         .expect("the follow-up request must be answered")
         .expect("the follow-up answer must be readable");
 
-    assert!(matches!(
-        oversized_request_outcome,
-        Err(Error::ConnectionDropped { .. })
-    ));
+    assert_eq!(
+        oversized_request_outcome.as_ref().err().map(discriminant),
+        Some(discriminant(&Error::ConnectionDropped {
+            request_id: String::new()
+        }))
+    );
     assert!(matches!(
         follow_up_message,
         Message::Error(envelope) if envelope.error.code == 504
diff --git a/paddler_tests/tests/client_reports_service_unavailable_for_an_embedding_batch_without_agents.rs b/paddler_tests/tests/client_reports_service_unavailable_for_an_embedding_batch_without_agents.rs
deleted file mode 100644
index 190ab7da9..000000000
--- a/paddler_tests/tests/client_reports_service_unavailable_for_an_embedding_batch_without_agents.rs
+++ /dev/null
@@ -1,54 +0,0 @@
-use tokio_util::sync::CancellationToken;
-
-use paddler_client::error::Error;
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
-use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-use paddler_messaging::embedding_input_document::EmbeddingInputDocument;
-use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
-use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
-use paddler_test_cluster_harness::cluster_params::ClusterParams;
-use paddler_tests::start_cluster::start_cluster;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn client_reports_service_unavailable_for_an_embedding_batch_without_agents() {
-    let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        desired_state: Some(BalancerDesiredState {
-            inference_parameters: InferenceParameters {
-                enable_embeddings: true,
-                ..InferenceParameters::deterministic()
-            },
-            ..BalancerDesiredState::default()
-        }),
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("a cluster without agents must start");
-
-    let rejection = cluster
-        .client_inference
-        .post_generate_embedding_batch(
-            CancellationToken::new(),
-            &GenerateEmbeddingBatchParams {
-                input_batch: vec![EmbeddingInputDocument {
-                    content: "Hello world".to_owned(),
-                    id: "doc-1".to_owned(),
-                }],
-                normalization_method: EmbeddingNormalizationMethod::None,
-            },
-        )
-        .await
-        .err()
-        .expect("the balancer must reject an embedding batch while no agent can serve it");
-
-    assert!(matches!(
-        rejection,
-        Error::ServiceUnavailable { message, .. } if message == "No agents are currently connected"
-    ));
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_tests/tests/client_reports_unexpected_status_for_model_metadata_of_an_unknown_agent.rs b/paddler_tests/tests/client_reports_unexpected_status_for_model_metadata_of_an_unknown_agent.rs
index 6d2465930..916a2f3f1 100644
--- a/paddler_tests/tests/client_reports_unexpected_status_for_model_metadata_of_an_unknown_agent.rs
+++ b/paddler_tests/tests/client_reports_unexpected_status_for_model_metadata_of_an_unknown_agent.rs
@@ -7,13 +7,9 @@ use paddler_tests::start_cluster::start_cluster;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn client_reports_unexpected_status_for_model_metadata_of_an_unknown_agent() {
-    let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("a cluster without agents must start");
+    let cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("a cluster without agents must start");
 
     let rejection = cluster
         .client_management
diff --git a/paddler_tests/tests/cluster_harness_reports_an_agent_that_reports_issues_while_starting.rs b/paddler_tests/tests/cluster_harness_reports_an_agent_that_reports_issues_while_starting.rs
index 1af1da8b5..d9096e327 100644
--- a/paddler_tests/tests/cluster_harness_reports_an_agent_that_reports_issues_while_starting.rs
+++ b/paddler_tests/tests/cluster_harness_reports_an_agent_that_reports_issues_while_starting.rs
@@ -1,10 +1,12 @@
 use std::collections::BTreeSet;
+use std::path::Path;
 
-use paddler_messaging::agent_desired_model::AgentDesiredModel;
 use paddler_messaging::agent_issue::AgentIssue;
 use paddler_messaging::agent_issue_params::model_path::ModelPath;
 use paddler_messaging::balancer_desired_state::BalancerDesiredState;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::agent_desired_model_at::agent_desired_model_at;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
 use paddler_test_cluster_harness::cluster_harness_error::ClusterHarnessError;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
 use paddler_tests::start_cluster::start_cluster;
@@ -19,10 +21,11 @@ async fn cluster_harness_reports_an_agent_that_reports_issues_while_starting() {
             name: AGENT_NAME.to_owned(),
             slot_count: 1,
         }],
-        desired_state: Some(BalancerDesiredState {
-            model: AgentDesiredModel::LocalToAgent(MISSING_MODEL_PATH.to_owned()),
+        desired_state: ClusterDesiredState::Apply(Box::new(BalancerDesiredState {
+            model: agent_desired_model_at(Path::new(MISSING_MODEL_PATH))
+                .expect("the missing model path must be absolute"),
             ..BalancerDesiredState::default()
-        }),
+        })),
         wait_for_slots_ready: true,
         ..ClusterParams::default()
     })
diff --git a/paddler_tests/tests/cluster_harness_reports_an_inference_socket_closed_before_the_answer.rs b/paddler_tests/tests/cluster_harness_reports_an_inference_socket_closed_before_the_answer.rs
index 615ebc257..31bbb82ee 100644
--- a/paddler_tests/tests/cluster_harness_reports_an_inference_socket_closed_before_the_answer.rs
+++ b/paddler_tests/tests/cluster_harness_reports_an_inference_socket_closed_before_the_answer.rs
@@ -7,13 +7,9 @@ const UNANSWERED_REQUEST_ID: &str = "request-the-balancer-never-received";
 
 #[tokio::test(flavor = "multi_thread")]
 async fn cluster_harness_reports_an_inference_socket_closed_before_the_answer() {
-    let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("a cluster without agents must start");
+    let cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("a cluster without agents must start");
     let mut socket = RawInferenceSocket::connect(cluster.balancer.addresses.inference)
         .await
         .expect("the raw inference socket must connect");
diff --git a/paddler_tests/tests/cluster_harness_reports_requests_cancelled_before_they_are_sent.rs b/paddler_tests/tests/cluster_harness_reports_requests_cancelled_before_they_are_sent.rs
new file mode 100644
index 000000000..46d497de6
--- /dev/null
+++ b/paddler_tests/tests/cluster_harness_reports_requests_cancelled_before_they_are_sent.rs
@@ -0,0 +1,88 @@
+use std::mem::discriminant;
+use std::num::NonZeroU32;
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_client::error::Error as ClientError;
+use paddler_messaging::conversation_history::ConversationHistory;
+use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
+use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
+use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
+use paddler_messaging::request_params::decide_params::raw_decide_params::RawDecideParams;
+use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_tests::start_cluster::start_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn cluster_harness_reports_requests_cancelled_before_they_are_sent() {
+    let cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("a cluster without agents must start");
+    let cancelled_requests = CancellationToken::new();
+
+    cancelled_requests.cancel();
+
+    let request_errors = [
+        cluster
+            .continue_from_raw_prompt(
+                cancelled_requests.clone(),
+                &ContinueFromRawPromptParams {
+                    grammar: None,
+                    max_tokens: NonZeroU32::MIN,
+                    raw_prompt: "Paddler".to_owned(),
+                },
+            )
+            .await
+            .err(),
+        cluster
+            .continue_from_conversation_history(
+                cancelled_requests.clone(),
+                &ContinueFromConversationHistoryParams {
+                    add_generation_prompt: true,
+                    conversation_history: ConversationHistory::new(Vec::new()),
+                    enable_thinking: false,
+                    grammar: None,
+                    max_tokens: NonZeroU32::MIN,
+                    parse_tool_calls: false,
+                    tools: Vec::new(),
+                },
+            )
+            .await
+            .err(),
+        cluster
+            .decide(
+                cancelled_requests.clone(),
+                &RawDecideParams {
+                    questions: Vec::new(),
+                    state: "Paddler".to_owned(),
+                },
+            )
+            .await
+            .err(),
+        cluster
+            .generate_embedding_batch(
+                cancelled_requests,
+                &GenerateEmbeddingBatchParams {
+                    input_batch: Vec::new(),
+                    normalization_method: EmbeddingNormalizationMethod::None,
+                },
+            )
+            .await
+            .err(),
+    ];
+
+    assert_eq!(
+        request_errors.map(|request_error| request_error
+            .as_ref()
+            .and_then(|request_error| request_error.root_cause().downcast_ref::<ClientError>())
+            .map(discriminant)),
+        [Some(discriminant(&ClientError::RequestCancelled {
+            url: String::new()
+        })); 4]
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/cluster_harness_reports_streamed_requests_cancelled_before_they_are_sent.rs b/paddler_tests/tests/cluster_harness_reports_streamed_requests_cancelled_before_they_are_sent.rs
new file mode 100644
index 000000000..3910509f0
--- /dev/null
+++ b/paddler_tests/tests/cluster_harness_reports_streamed_requests_cancelled_before_they_are_sent.rs
@@ -0,0 +1,85 @@
+use std::num::NonZeroU32;
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_client::error::Error as ClientError;
+use paddler_client::format_api_url::format_api_url;
+use paddler_messaging::api_path::ApiPath;
+use paddler_messaging::conversation_history::ConversationHistory;
+use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
+use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
+use paddler_test_cluster_harness::cluster_harness_error::ClusterHarnessError;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_tests::start_cluster::start_cluster;
+
+fn is_cancelled_request_to(
+    request_error: Option<&ClusterHarnessError>,
+    expected_url: &str,
+) -> bool {
+    matches!(
+        request_error,
+        Some(ClusterHarnessError::InferenceRequestFailed(ClientError::RequestCancelled { url }))
+            if url == expected_url
+    )
+}
+
+#[tokio::test(flavor = "multi_thread")]
+async fn cluster_harness_reports_streamed_requests_cancelled_before_they_are_sent() {
+    let cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("a cluster without agents must start");
+    let inference_base_url = &cluster.balancer.inference_base_url;
+    let cancelled_requests = CancellationToken::new();
+
+    cancelled_requests.cancel();
+
+    let raw_prompt_stream_error = cluster
+        .continue_from_raw_prompt_stream(
+            cancelled_requests.clone(),
+            &ContinueFromRawPromptParams {
+                grammar: None,
+                max_tokens: NonZeroU32::MIN,
+                raw_prompt: "Paddler".to_owned(),
+            },
+        )
+        .await
+        .err();
+    let conversation_history_stream_error = cluster
+        .continue_from_conversation_history_stream(
+            cancelled_requests,
+            &ContinueFromConversationHistoryParams {
+                add_generation_prompt: true,
+                conversation_history: ConversationHistory::new(Vec::new()),
+                enable_thinking: false,
+                grammar: None,
+                max_tokens: NonZeroU32::MIN,
+                parse_tool_calls: false,
+                tools: Vec::new(),
+            },
+        )
+        .await
+        .err();
+
+    assert!(
+        is_cancelled_request_to(
+            raw_prompt_stream_error.as_ref(),
+            &format_api_url(inference_base_url, ApiPath::CONTINUE_FROM_RAW_PROMPT)
+        ),
+        "{raw_prompt_stream_error:?}"
+    );
+    assert!(
+        is_cancelled_request_to(
+            conversation_history_stream_error.as_ref(),
+            &format_api_url(
+                inference_base_url,
+                ApiPath::CONTINUE_FROM_CONVERSATION_HISTORY
+            )
+        ),
+        "{conversation_history_stream_error:?}"
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/cluster_harness_reports_the_end_of_a_snapshots_stream.rs b/paddler_tests/tests/cluster_harness_reports_the_end_of_a_snapshots_stream.rs
index ace424b36..e073ce427 100644
--- a/paddler_tests/tests/cluster_harness_reports_the_end_of_a_snapshots_stream.rs
+++ b/paddler_tests/tests/cluster_harness_reports_the_end_of_a_snapshots_stream.rs
@@ -10,13 +10,9 @@ async fn cluster_harness_reports_the_end_of_a_snapshots_stream() {
         mut agents_watcher,
         balancer,
         ..
-    } = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("a cluster without agents must start");
+    } = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("a cluster without agents must start");
 
     balancer
         .shutdown()
@@ -25,10 +21,13 @@ async fn cluster_harness_reports_the_end_of_a_snapshots_stream() {
 
     let observation = agents_watcher.until(|_snapshot| false).await;
 
-    assert!(matches!(
-        observation,
-        Err(ClusterHarnessError::SnapshotsStreamClosed {
-            snapshots_stream: SnapshotsStream::Agents
-        })
-    ));
+    assert_eq!(
+        observation.err().map(|stream_end| stream_end.to_string()),
+        Some(
+            ClusterHarnessError::SnapshotsStreamClosed {
+                snapshots_stream: SnapshotsStream::Agents
+            }
+            .to_string()
+        )
+    );
 }
diff --git a/paddler_tests/tests/cluster_harness_waiting_for_its_first_agent_reports_that_none_registered.rs b/paddler_tests/tests/cluster_harness_waiting_for_its_first_agent_reports_that_none_registered.rs
new file mode 100644
index 000000000..adddf5992
--- /dev/null
+++ b/paddler_tests/tests/cluster_harness_waiting_for_its_first_agent_reports_that_none_registered.rs
@@ -0,0 +1,30 @@
+use std::mem::discriminant;
+
+use paddler_messaging::agent_decisions_serving::AgentDecisionsServing;
+use paddler_test_cluster_harness::cluster_harness_error::ClusterHarnessError;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_tests::start_cluster::start_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn cluster_harness_waiting_for_its_first_agent_reports_that_none_registered() {
+    let mut cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("a cluster without agents must start");
+
+    assert_eq!(
+        cluster
+            .wait_for_first_agent_to_serve(AgentDecisionsServing::NotServed)
+            .await
+            .err()
+            .as_ref()
+            .map(discriminant),
+        Some(discriminant(
+            &ClusterHarnessError::ClusterHasNoRegisteredAgent
+        ))
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/cluster_harness_waiting_for_its_first_agent_to_serve_reports_that_the_agent_left.rs b/paddler_tests/tests/cluster_harness_waiting_for_its_first_agent_to_serve_reports_that_the_agent_left.rs
new file mode 100644
index 000000000..16280d4aa
--- /dev/null
+++ b/paddler_tests/tests/cluster_harness_waiting_for_its_first_agent_to_serve_reports_that_the_agent_left.rs
@@ -0,0 +1,44 @@
+use paddler_messaging::agent_decisions_serving::AgentDecisionsServing;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
+use paddler_test_cluster_harness::cluster_harness_error::ClusterHarnessError;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_tests::start_cluster::start_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn cluster_harness_waiting_for_its_first_agent_to_serve_reports_that_the_agent_left() {
+    let mut cluster = start_cluster(ClusterParams {
+        agents: AgentConfig::uniform(1, 1),
+        desired_state: ClusterDesiredState::Apply(Box::default()),
+        wait_for_slots_ready: false,
+        ..ClusterParams::default()
+    })
+    .await
+    .expect("a single-agent cluster must start");
+    let awaited_agent_id = cluster
+        .agent_ids
+        .first()
+        .expect("the cluster must have a registered agent")
+        .clone();
+
+    cluster
+        .agents
+        .pop()
+        .expect("the cluster must have its registered agent")
+        .shutdown()
+        .await
+        .expect("the agent must leave the pool cleanly");
+
+    assert!(matches!(
+        cluster
+            .wait_for_first_agent_to_serve(AgentDecisionsServing::NotServed)
+            .await
+            .err(),
+        Some(ClusterHarnessError::AgentLeftThePool { agent_id }) if agent_id == awaited_agent_id
+    ));
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/cluster_shutdown_returns_fd_count_to_baseline.rs b/paddler_tests/tests/cluster_shutdown_returns_fd_count_to_baseline.rs
index b04ea09c8..5c17b9d62 100644
--- a/paddler_tests/tests/cluster_shutdown_returns_fd_count_to_baseline.rs
+++ b/paddler_tests/tests/cluster_shutdown_returns_fd_count_to_baseline.rs
@@ -1,5 +1,3 @@
-#![cfg(any(target_os = "macos", target_os = "linux"))]
-
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
 use paddler_test_cluster_harness::resource_snapshot::ResourceSnapshot;
 use paddler_tests::start_cluster::start_cluster;
diff --git a/paddler_tests/tests/cluster_shuts_down_cleanly_after_a_client_abandons_a_generation.rs b/paddler_tests/tests/cluster_shuts_down_cleanly_after_a_client_abandons_a_generation.rs
new file mode 100644
index 000000000..6c0efb860
--- /dev/null
+++ b/paddler_tests/tests/cluster_shuts_down_cleanly_after_a_client_abandons_a_generation.rs
@@ -0,0 +1,41 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::num::NonZeroU32;
+
+use futures_util::StreamExt as _;
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn cluster_shuts_down_cleanly_after_a_client_abandons_a_generation() {
+    let cluster = start_cluster_with_qwen3(vec![AgentConfig::single(1)])
+        .await
+        .expect("the cluster must start");
+
+    let mut stream = cluster
+        .continue_from_raw_prompt_stream(
+            CancellationToken::new(),
+            &ContinueFromRawPromptParams {
+                grammar: None,
+                max_tokens: NonZeroU32::new(500).unwrap(),
+                raw_prompt: "Write a long essay".to_owned(),
+            },
+        )
+        .await
+        .expect("the inference request must be accepted");
+
+    let _first_message = stream
+        .next()
+        .await
+        .expect("inference stream must yield at least one message");
+
+    drop(stream);
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/continuous_batch_concurrent_conversation_history_requests_complete.rs b/paddler_tests/tests/continuous_batch_concurrent_conversation_history_requests_complete.rs
index d20ab11fd..1d5f280ba 100644
--- a/paddler_tests/tests/continuous_batch_concurrent_conversation_history_requests_complete.rs
+++ b/paddler_tests/tests/continuous_batch_concurrent_conversation_history_requests_complete.rs
@@ -8,14 +8,12 @@ use tokio_util::sync::CancellationToken;
 use paddler_messaging::conversation_history::ConversationHistory;
 use paddler_messaging::conversation_message::ConversationMessage;
 use paddler_messaging::conversation_message_content::ConversationMessageContent;
-use paddler_messaging::generated_token_result::GeneratedTokenResult;
 use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
-use paddler_test_cluster_harness::token_result_with_producer::TokenResultWithProducer;
 use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
 
 fn user_message(text: &str) -> ConversationMessage {
-    ConversationMessage {
+    ConversationMessage::Message {
         content: ConversationMessageContent::Text(text.to_owned()),
         role: "user".to_owned(),
     }
@@ -66,20 +64,12 @@ async fn continuous_batch_concurrent_conversation_history_requests_complete() {
 
     assert!(tokens_a > 0);
     assert!(tokens_b > 0);
-    assert!(matches!(
-        collected_a.token_results.last(),
-        Some(TokenResultWithProducer {
-            token_result: GeneratedTokenResult::Done(_),
-            ..
-        })
-    ));
-    assert!(matches!(
-        collected_b.token_results.last(),
-        Some(TokenResultWithProducer {
-            token_result: GeneratedTokenResult::Done(_),
-            ..
-        })
-    ));
+    collected_a
+        .summary()
+        .expect("the generation must finish with a summary");
+    collected_b
+        .summary()
+        .expect("the generation must finish with a summary");
 
     cluster
         .shutdown()
diff --git a/paddler_tests/tests/continuous_batch_finishes_a_sequence_that_fills_its_share_of_the_context.rs b/paddler_tests/tests/continuous_batch_finishes_a_sequence_that_fills_its_share_of_the_context.rs
index 223ccd51e..a0b0402d9 100644
--- a/paddler_tests/tests/continuous_batch_finishes_a_sequence_that_fills_its_share_of_the_context.rs
+++ b/paddler_tests/tests/continuous_batch_finishes_a_sequence_that_fills_its_share_of_the_context.rs
@@ -6,11 +6,12 @@ use tokio::join;
 use tokio_util::sync::CancellationToken;
 
 use paddler_inference_parameters::batch_size::BatchSize;
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
+use paddler_inference_parameters::model_runtime_parameters::ModelRuntimeParameters;
 use paddler_messaging::balancer_desired_state::BalancerDesiredState;
 use paddler_messaging::generation_finish::GenerationFinish;
 use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
 use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
 use paddler_test_cluster_harness::unending_generation::unending_generation;
@@ -22,15 +23,15 @@ async fn continuous_batch_finishes_a_sequence_that_fills_its_share_of_the_contex
 
     let cluster = start_cluster(ClusterParams {
         agents: vec![AgentConfig::single(2)],
-        desired_state: Some(BalancerDesiredState {
-            inference_parameters: InferenceParameters {
+        desired_state: ClusterDesiredState::Apply(Box::new(BalancerDesiredState {
+            model_runtime_parameters: ModelRuntimeParameters {
                 n_batch: BatchSize::try_from(256).expect("the value must fit its target type"),
                 context_size: NonZeroU32::try_from(256)
                     .expect("the value must fit its target type"),
-                ..base_desired_state.inference_parameters
+                ..base_desired_state.model_runtime_parameters
             },
             ..base_desired_state
-        }),
+        })),
         wait_for_slots_ready: true,
         ..ClusterParams::default()
     })
diff --git a/paddler_tests/tests/continuous_batch_generates_tokens_with_distinct_k_and_v_cache_dtypes.rs b/paddler_tests/tests/continuous_batch_generates_tokens_with_distinct_k_and_v_cache_dtypes.rs
index 74cdefb85..b760f30eb 100644
--- a/paddler_tests/tests/continuous_batch_generates_tokens_with_distinct_k_and_v_cache_dtypes.rs
+++ b/paddler_tests/tests/continuous_batch_generates_tokens_with_distinct_k_and_v_cache_dtypes.rs
@@ -5,43 +5,35 @@ use std::num::NonZeroU32;
 use tokio_util::sync::CancellationToken;
 
 use paddler_inference_parameters::all_gpu_layers::ALL_GPU_LAYERS;
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
 use paddler_inference_parameters::kv_cache_dtype::KvCacheDtype;
-use paddler_messaging::agent_desired_model::AgentDesiredModel;
+use paddler_inference_parameters::model_runtime_parameters::ModelRuntimeParameters;
 use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-use paddler_messaging::generated_token_result::GeneratedTokenResult;
 use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
-use paddler_test_cluster_harness::model_card::ModelCard;
 use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
-use paddler_test_cluster_harness::token_result_with_producer::TokenResultWithProducer;
 use paddler_tests::start_cluster::start_cluster;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn continuous_batch_generates_tokens_with_distinct_k_and_v_cache_dtypes() {
-    let ModelCard { reference } = qwen3_0_6b();
-
-    let mut inference_parameters = InferenceParameters {
+    let mut model_runtime_parameters = ModelRuntimeParameters {
         n_gpu_layers: ALL_GPU_LAYERS,
-        ..InferenceParameters::deterministic()
+        ..ModelRuntimeParameters::default()
     };
 
-    inference_parameters.k_cache_dtype = KvCacheDtype::Q80;
-    inference_parameters.v_cache_dtype = KvCacheDtype::F16;
+    model_runtime_parameters.k_cache_dtype = KvCacheDtype::Q80;
+    model_runtime_parameters.v_cache_dtype = KvCacheDtype::F16;
 
     let cluster = start_cluster(ClusterParams {
         agents: vec![AgentConfig {
             name: "test-agent".to_owned(),
             slot_count: 1,
         }],
-        desired_state: Some(BalancerDesiredState {
-            chat_template_override: None,
-            inference_parameters,
-            model: AgentDesiredModel::HuggingFace(reference),
-            multimodal_projection: AgentDesiredModel::None,
-            use_chat_template_override: false,
-        }),
+        desired_state: ClusterDesiredState::Apply(Box::new(BalancerDesiredState {
+            model_runtime_parameters,
+            ..qwen3_0_6b().into_desired_state()
+        })),
         wait_for_slots_ready: true,
         ..ClusterParams::default()
     })
@@ -67,13 +59,9 @@ async fn continuous_batch_generates_tokens_with_distinct_k_and_v_cache_dtypes()
         .count();
 
     assert!(token_count > 0);
-    assert!(matches!(
-        collected.token_results.last(),
-        Some(TokenResultWithProducer {
-            token_result: GeneratedTokenResult::Done(_),
-            ..
-        })
-    ));
+    collected
+        .summary()
+        .expect("the generation must finish with a summary");
 
     cluster
         .shutdown()
diff --git a/paddler_tests/tests/continuous_batch_generates_tokens_with_partial_layer_offload.rs b/paddler_tests/tests/continuous_batch_generates_tokens_with_partial_layer_offload.rs
index 7a15e9ee8..92a1e54d1 100644
--- a/paddler_tests/tests/continuous_batch_generates_tokens_with_partial_layer_offload.rs
+++ b/paddler_tests/tests/continuous_batch_generates_tokens_with_partial_layer_offload.rs
@@ -4,27 +4,22 @@ use std::num::NonZeroU32;
 
 use tokio_util::sync::CancellationToken;
 
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
-use paddler_messaging::agent_desired_model::AgentDesiredModel;
+use paddler_inference_parameters::model_runtime_parameters::ModelRuntimeParameters;
 use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-use paddler_messaging::generated_token_result::GeneratedTokenResult;
 use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
-use paddler_test_cluster_harness::model_card::ModelCard;
 use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
-use paddler_test_cluster_harness::token_result_with_producer::TokenResultWithProducer;
 use paddler_tests::start_cluster::start_cluster;
 
 const PARTIAL_GPU_LAYER_COUNT: i32 = 14;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn continuous_batch_generates_tokens_with_partial_layer_offload() {
-    let ModelCard { reference } = qwen3_0_6b();
-
-    let inference_parameters = InferenceParameters {
+    let model_runtime_parameters = ModelRuntimeParameters {
         n_gpu_layers: PARTIAL_GPU_LAYER_COUNT,
-        ..InferenceParameters::deterministic()
+        ..ModelRuntimeParameters::default()
     };
 
     let cluster = start_cluster(ClusterParams {
@@ -32,13 +27,10 @@ async fn continuous_batch_generates_tokens_with_partial_layer_offload() {
             name: "test-agent".to_owned(),
             slot_count: 1,
         }],
-        desired_state: Some(BalancerDesiredState {
-            chat_template_override: None,
-            inference_parameters,
-            model: AgentDesiredModel::HuggingFace(reference),
-            multimodal_projection: AgentDesiredModel::None,
-            use_chat_template_override: false,
-        }),
+        desired_state: ClusterDesiredState::Apply(Box::new(BalancerDesiredState {
+            model_runtime_parameters,
+            ..qwen3_0_6b().into_desired_state()
+        })),
         wait_for_slots_ready: true,
         ..ClusterParams::default()
     })
@@ -64,13 +56,9 @@ async fn continuous_batch_generates_tokens_with_partial_layer_offload() {
         .count();
 
     assert!(token_count > 0);
-    assert!(matches!(
-        collected.token_results.last(),
-        Some(TokenResultWithProducer {
-            token_result: GeneratedTokenResult::Done(_),
-            ..
-        })
-    ));
+    collected
+        .summary()
+        .expect("the generation must finish with a summary");
 
     cluster
         .shutdown()
diff --git a/paddler_tests/tests/continuous_batch_generates_tokens_with_quantized_distinct_k_and_v_cache_dtypes_on_metal.rs b/paddler_tests/tests/continuous_batch_generates_tokens_with_quantized_distinct_k_and_v_cache_dtypes_on_metal.rs
index eb59c73cf..b95c29192 100644
--- a/paddler_tests/tests/continuous_batch_generates_tokens_with_quantized_distinct_k_and_v_cache_dtypes_on_metal.rs
+++ b/paddler_tests/tests/continuous_batch_generates_tokens_with_quantized_distinct_k_and_v_cache_dtypes_on_metal.rs
@@ -2,47 +2,38 @@
 
 use std::num::NonZeroU32;
 
-use anyhow::Result;
 use tokio_util::sync::CancellationToken;
 
 use paddler_inference_parameters::all_gpu_layers::ALL_GPU_LAYERS;
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
 use paddler_inference_parameters::kv_cache_dtype::KvCacheDtype;
-use paddler_messaging::agent_desired_model::AgentDesiredModel;
+use paddler_inference_parameters::model_runtime_parameters::ModelRuntimeParameters;
 use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-use paddler_messaging::generated_token_result::GeneratedTokenResult;
 use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
-use paddler_test_cluster_harness::model_card::ModelCard;
 use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
-use paddler_test_cluster_harness::token_result_with_producer::TokenResultWithProducer;
 use paddler_tests::start_cluster::start_cluster;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn continuous_batch_generates_tokens_with_quantized_distinct_k_and_v_cache_dtypes_on_metal() {
-    let ModelCard { reference } = qwen3_0_6b();
-
-    let mut inference_parameters = InferenceParameters {
+    let mut model_runtime_parameters = ModelRuntimeParameters {
         n_gpu_layers: ALL_GPU_LAYERS,
-        ..InferenceParameters::deterministic()
+        ..ModelRuntimeParameters::default()
     };
 
-    inference_parameters.k_cache_dtype = KvCacheDtype::Q80;
-    inference_parameters.v_cache_dtype = KvCacheDtype::Q40;
+    model_runtime_parameters.k_cache_dtype = KvCacheDtype::Q80;
+    model_runtime_parameters.v_cache_dtype = KvCacheDtype::Q40;
 
     let cluster = start_cluster(ClusterParams {
         agents: vec![AgentConfig {
             name: "test-agent".to_owned(),
             slot_count: 1,
         }],
-        desired_state: Some(BalancerDesiredState {
-            chat_template_override: None,
-            inference_parameters,
-            model: AgentDesiredModel::HuggingFace(reference),
-            multimodal_projection: AgentDesiredModel::None,
-            use_chat_template_override: false,
-        }),
+        desired_state: ClusterDesiredState::Apply(Box::new(BalancerDesiredState {
+            model_runtime_parameters,
+            ..qwen3_0_6b().into_desired_state()
+        })),
         wait_for_slots_ready: true,
         ..ClusterParams::default()
     })
@@ -68,13 +59,9 @@ async fn continuous_batch_generates_tokens_with_quantized_distinct_k_and_v_cache
         .count();
 
     assert!(token_count > 0);
-    assert!(matches!(
-        collected.token_results.last(),
-        Some(TokenResultWithProducer {
-            token_result: GeneratedTokenResult::Done(_),
-            ..
-        })
-    ));
+    collected
+        .summary()
+        .expect("the generation must finish with a summary");
 
     cluster
         .shutdown()
diff --git a/paddler_tests/tests/continuous_batch_long_and_short_prompts_complete_concurrently.rs b/paddler_tests/tests/continuous_batch_long_and_short_prompts_complete_concurrently.rs
index a941f10f2..c985f9ff8 100644
--- a/paddler_tests/tests/continuous_batch_long_and_short_prompts_complete_concurrently.rs
+++ b/paddler_tests/tests/continuous_batch_long_and_short_prompts_complete_concurrently.rs
@@ -5,10 +5,8 @@ use std::num::NonZeroU32;
 use tokio::join;
 use tokio_util::sync::CancellationToken;
 
-use paddler_messaging::generated_token_result::GeneratedTokenResult;
 use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
-use paddler_test_cluster_harness::token_result_with_producer::TokenResultWithProducer;
 use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
 
 #[tokio::test(flavor = "multi_thread")]
@@ -50,20 +48,12 @@ async fn continuous_batch_long_and_short_prompts_complete_concurrently() {
 
     assert!(long_tokens > 0);
     assert!(short_tokens > 0);
-    assert!(matches!(
-        long_collected.token_results.last(),
-        Some(TokenResultWithProducer {
-            token_result: GeneratedTokenResult::Done(_),
-            ..
-        })
-    ));
-    assert!(matches!(
-        short_collected.token_results.last(),
-        Some(TokenResultWithProducer {
-            token_result: GeneratedTokenResult::Done(_),
-            ..
-        })
-    ));
+    long_collected
+        .summary()
+        .expect("the generation must finish with a summary");
+    short_collected
+        .summary()
+        .expect("the generation must finish with a summary");
 
     cluster
         .shutdown()
diff --git a/paddler_tests/tests/continuous_batch_produces_distinct_outputs_for_concurrent_prompts.rs b/paddler_tests/tests/continuous_batch_produces_distinct_outputs_for_concurrent_prompts.rs
index da496c6fc..fa045b514 100644
--- a/paddler_tests/tests/continuous_batch_produces_distinct_outputs_for_concurrent_prompts.rs
+++ b/paddler_tests/tests/continuous_batch_produces_distinct_outputs_for_concurrent_prompts.rs
@@ -7,6 +7,7 @@ use tokio_util::sync::CancellationToken;
 
 use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
 use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
 use paddler_tests::start_cluster::start_cluster;
@@ -20,7 +21,7 @@ async fn continuous_batch_produces_distinct_outputs_for_concurrent_prompts() {
             name: "test-agent".to_owned(),
             slot_count: 2,
         }],
-        desired_state: Some(desired_state),
+        desired_state: ClusterDesiredState::Apply(Box::new(desired_state)),
         wait_for_slots_ready: true,
         ..ClusterParams::default()
     })
diff --git a/paddler_tests/tests/continuous_batch_rejects_embedding_during_active_generation.rs b/paddler_tests/tests/continuous_batch_rejects_embedding_during_active_generation.rs
deleted file mode 100644
index 0eeedfe34..000000000
--- a/paddler_tests/tests/continuous_batch_rejects_embedding_during_active_generation.rs
+++ /dev/null
@@ -1,86 +0,0 @@
-#![cfg(feature = "tests_that_use_llms")]
-
-use futures_util::StreamExt as _;
-use tokio_util::sync::CancellationToken;
-
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
-use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-use paddler_messaging::chat_template::ChatTemplate;
-use paddler_messaging::embedding_input_document::EmbeddingInputDocument;
-use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
-use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
-use paddler_test_cluster_harness::agent_config::AgentConfig;
-use paddler_test_cluster_harness::unending_generation::unending_generation;
-use paddler_tests::qwen3_desired_state_with_embeddings::qwen3_desired_state_with_embeddings;
-use paddler_tests::start_cluster_without_model::start_cluster_without_model;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn continuous_batch_rejects_embedding_during_active_generation() {
-    let mut cluster = start_cluster_without_model(
-        AgentConfig::uniform(1, 2),
-        InferenceParameters::deterministic(),
-    )
-    .await
-    .expect("the cluster must start");
-
-    let generation_cancellation = CancellationToken::new();
-    let mut generation_stream = cluster
-        .continue_from_raw_prompt_stream(generation_cancellation.clone(), &unending_generation())
-        .await
-        .expect("the inference request must be accepted");
-
-    cluster
-        .wait_for_buffered_request_count(1)
-        .await
-        .expect("the generation request must wait in the buffer while the agent has no model");
-
-    cluster
-        .client_management
-        .put_balancer_desired_state(
-            CancellationToken::new(),
-            &BalancerDesiredState {
-                chat_template_override: Some(ChatTemplate {
-                    content: "{% for message in messages %}{{ message.content }}{% endfor %}"
-                        .to_owned(),
-                }),
-                use_chat_template_override: true,
-                ..qwen3_desired_state_with_embeddings(true)
-            },
-        )
-        .await
-        .expect("the balancer must accept the desired state");
-
-    generation_stream
-        .next()
-        .await
-        .expect("the buffered generation request must start streaming tokens")
-        .expect("the message must be readable");
-
-    let collected = cluster
-        .generate_embedding_batch(
-            CancellationToken::new(),
-            &GenerateEmbeddingBatchParams {
-                input_batch: vec![EmbeddingInputDocument {
-                    content: "test".to_owned(),
-                    id: "doc1".to_owned(),
-                }],
-                normalization_method: EmbeddingNormalizationMethod::None,
-            },
-        )
-        .await
-        .expect("the embedding batch must be accepted");
-
-    assert_eq!(
-        collected.embedding_rejected_due_to_active_token_generation_count,
-        1
-    );
-    assert!(collected.embeddings.is_empty());
-
-    generation_cancellation.cancel();
-    drop(generation_stream);
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_tests/tests/continuous_batch_rejects_second_request_when_only_slot_busy.rs b/paddler_tests/tests/continuous_batch_rejects_second_request_when_only_slot_busy.rs
index 679bd434d..8e7ccb861 100644
--- a/paddler_tests/tests/continuous_batch_rejects_second_request_when_only_slot_busy.rs
+++ b/paddler_tests/tests/continuous_batch_rejects_second_request_when_only_slot_busy.rs
@@ -7,6 +7,7 @@ use tokio_util::sync::CancellationToken;
 
 use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
 use paddler_test_cluster_harness::unending_generation::unending_generation;
 use paddler_tests::qwen3_desired_state::qwen3_desired_state;
@@ -17,7 +18,7 @@ async fn continuous_batch_rejects_second_request_when_only_slot_busy() {
     let mut cluster = start_cluster(ClusterParams {
         agents: vec![AgentConfig::single(1)],
         max_buffered_requests: 0,
-        desired_state: Some(qwen3_desired_state()),
+        desired_state: ClusterDesiredState::Apply(Box::new(qwen3_desired_state())),
         wait_for_slots_ready: true,
         ..ClusterParams::default()
     })
diff --git a/paddler_tests/tests/continuous_batch_releases_slots_on_shutdown_with_active_request.rs b/paddler_tests/tests/continuous_batch_releases_slots_on_shutdown_with_active_request.rs
deleted file mode 100644
index 29213204b..000000000
--- a/paddler_tests/tests/continuous_batch_releases_slots_on_shutdown_with_active_request.rs
+++ /dev/null
@@ -1,41 +0,0 @@
-#![cfg(feature = "tests_that_use_llms")]
-
-use std::num::NonZeroU32;
-
-use futures_util::StreamExt as _;
-use tokio_util::sync::CancellationToken;
-
-use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
-use paddler_test_cluster_harness::agent_config::AgentConfig;
-use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn continuous_batch_releases_slots_on_shutdown_with_active_request() {
-    let cluster = start_cluster_with_qwen3(vec![AgentConfig::single(1)])
-        .await
-        .expect("the cluster must start");
-
-    let mut stream = cluster
-        .continue_from_raw_prompt_stream(
-            CancellationToken::new(),
-            &ContinueFromRawPromptParams {
-                grammar: None,
-                max_tokens: NonZeroU32::new(500).unwrap(),
-                raw_prompt: "Write a long essay".to_owned(),
-            },
-        )
-        .await
-        .expect("the inference request must be accepted");
-
-    let _first_message = stream
-        .next()
-        .await
-        .expect("inference stream must yield at least one message");
-
-    drop(stream);
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_tests/tests/continuous_batch_reuses_slot_after_request_completes.rs b/paddler_tests/tests/continuous_batch_reuses_slot_after_request_completes.rs
index c55281673..7f6dc6d06 100644
--- a/paddler_tests/tests/continuous_batch_reuses_slot_after_request_completes.rs
+++ b/paddler_tests/tests/continuous_batch_reuses_slot_after_request_completes.rs
@@ -4,10 +4,8 @@ use std::num::NonZeroU32;
 
 use tokio_util::sync::CancellationToken;
 
-use paddler_messaging::generated_token_result::GeneratedTokenResult;
 use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
-use paddler_test_cluster_harness::token_result_with_producer::TokenResultWithProducer;
 use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
 
 #[tokio::test(flavor = "multi_thread")]
@@ -28,13 +26,9 @@ async fn continuous_batch_reuses_slot_after_request_completes() {
         .await
         .expect("the inference request must be accepted");
 
-    assert!(matches!(
-        first_collected.token_results.last(),
-        Some(TokenResultWithProducer {
-            token_result: GeneratedTokenResult::Done(_),
-            ..
-        })
-    ));
+    first_collected
+        .summary()
+        .expect("the generation must finish with a summary");
 
     let second_collected = cluster
         .continue_from_raw_prompt(
@@ -48,13 +42,9 @@ async fn continuous_batch_reuses_slot_after_request_completes() {
         .await
         .expect("the inference request must be accepted");
 
-    assert!(matches!(
-        second_collected.token_results.last(),
-        Some(TokenResultWithProducer {
-            token_result: GeneratedTokenResult::Done(_),
-            ..
-        })
-    ));
+    second_collected
+        .summary()
+        .expect("the generation must finish with a summary");
 
     let second_token_count = second_collected
         .token_results
diff --git a/paddler_tests/tests/continuous_batch_serves_four_concurrent_requests.rs b/paddler_tests/tests/continuous_batch_serves_four_concurrent_requests.rs
index de2ab5ef2..6152d08f9 100644
--- a/paddler_tests/tests/continuous_batch_serves_four_concurrent_requests.rs
+++ b/paddler_tests/tests/continuous_batch_serves_four_concurrent_requests.rs
@@ -5,10 +5,8 @@ use std::num::NonZeroU32;
 use futures_util::future::try_join_all;
 use tokio_util::sync::CancellationToken;
 
-use paddler_messaging::generated_token_result::GeneratedTokenResult;
 use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
-use paddler_test_cluster_harness::token_result_with_producer::TokenResultWithProducer;
 use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
 
 #[tokio::test(flavor = "multi_thread")]
@@ -45,13 +43,9 @@ async fn continuous_batch_serves_four_concurrent_requests() {
             token_count > 0,
             "every concurrent request must produce tokens"
         );
-        assert!(matches!(
-            collected.token_results.last(),
-            Some(TokenResultWithProducer {
-                token_result: GeneratedTokenResult::Done(_),
-                ..
-            })
-        ));
+        collected
+            .summary()
+            .expect("the generation must finish with a summary");
     }
 
     cluster
diff --git a/paddler_tests/tests/continuous_batch_stops_at_max_tokens_boundary.rs b/paddler_tests/tests/continuous_batch_stops_at_max_tokens_boundary.rs
index d27fbb92f..2c3476129 100644
--- a/paddler_tests/tests/continuous_batch_stops_at_max_tokens_boundary.rs
+++ b/paddler_tests/tests/continuous_batch_stops_at_max_tokens_boundary.rs
@@ -4,10 +4,9 @@ use std::num::NonZeroU32;
 
 use tokio_util::sync::CancellationToken;
 
-use paddler_messaging::generated_token_result::GeneratedTokenResult;
+use paddler_messaging::generation_finish::GenerationFinish;
 use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
-use paddler_test_cluster_harness::token_result_with_producer::TokenResultWithProducer;
 use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
 
 #[tokio::test(flavor = "multi_thread")]
@@ -38,13 +37,13 @@ async fn continuous_batch_stops_at_max_tokens_boundary() {
         token_count, 5,
         "expected exactly 5 tokens with max_tokens=5"
     );
-    assert!(matches!(
-        collected.token_results.last(),
-        Some(TokenResultWithProducer {
-            token_result: GeneratedTokenResult::Done(_),
-            ..
-        })
-    ));
+    assert_eq!(
+        collected
+            .summary()
+            .expect("the generation must finish with a summary")
+            .finish,
+        GenerationFinish::MaxTokens
+    );
 
     cluster
         .shutdown()
diff --git a/paddler_tests/tests/continuous_batch_stops_generation_when_stop_sender_dropped.rs b/paddler_tests/tests/continuous_batch_stops_generation_when_stop_sender_dropped.rs
index 97a0651d1..08d7aefef 100644
--- a/paddler_tests/tests/continuous_batch_stops_generation_when_stop_sender_dropped.rs
+++ b/paddler_tests/tests/continuous_batch_stops_generation_when_stop_sender_dropped.rs
@@ -5,15 +5,13 @@ use std::num::NonZeroU32;
 use futures_util::StreamExt as _;
 use tokio_util::sync::CancellationToken;
 
-use paddler_messaging::generated_token_result::GeneratedTokenResult;
 use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
-use paddler_test_cluster_harness::token_result_with_producer::TokenResultWithProducer;
 use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn continuous_batch_stops_generation_when_stop_sender_dropped() {
-    let cluster = start_cluster_with_qwen3(vec![AgentConfig::single(2)])
+    let cluster = start_cluster_with_qwen3(vec![AgentConfig::single(1)])
         .await
         .expect("the cluster must start");
 
@@ -48,13 +46,9 @@ async fn continuous_batch_stops_generation_when_stop_sender_dropped() {
         .await
         .expect("the inference request must be accepted");
 
-    assert!(matches!(
-        second_collected.token_results.last(),
-        Some(TokenResultWithProducer {
-            token_result: GeneratedTokenResult::Done(_),
-            ..
-        })
-    ));
+    second_collected
+        .summary()
+        .expect("the generation must finish with a summary");
 
     let second_token_count = second_collected
         .token_results
diff --git a/paddler_tests/tests/decision_agent_shuts_down_cleanly_with_decisions_in_flight.rs b/paddler_tests/tests/decision_agent_shuts_down_cleanly_with_decisions_in_flight.rs
new file mode 100644
index 000000000..2f3f35213
--- /dev/null
+++ b/paddler_tests/tests/decision_agent_shuts_down_cleanly_with_decisions_in_flight.rs
@@ -0,0 +1,69 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use futures_util::StreamExt as _;
+use tokio_util::sync::CancellationToken;
+
+use paddler_tests::decision_cluster_params::DecisionClusterParams;
+use paddler_tests::many_question_decision::many_question_decision;
+use paddler_tests::sample_decision::sample_decision;
+use paddler_tests::start_decision_cluster::start_decision_cluster;
+
+const QUESTIONS_OUTLASTING_A_SHUTDOWN: usize = 96;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn decision_agent_shuts_down_cleanly_with_decisions_in_flight() {
+    let mut cluster = start_decision_cluster(DecisionClusterParams::default())
+        .await
+        .expect("the cluster must start");
+    let agent_id = cluster
+        .agent_ids
+        .first()
+        .expect("the cluster must have an agent")
+        .clone();
+    let mut answering_stream = cluster
+        .client_inference
+        .post_decide(
+            CancellationToken::new(),
+            &many_question_decision(QUESTIONS_OUTLASTING_A_SHUTDOWN),
+        )
+        .await
+        .expect("the long decision must be accepted");
+
+    answering_stream
+        .next()
+        .await
+        .expect("the long decision must answer its first question")
+        .expect("the first answer must be readable");
+
+    let waiting_stream = cluster
+        .client_inference
+        .post_decide(CancellationToken::new(), &sample_decision())
+        .await
+        .expect("the waiting decision must be accepted");
+
+    cluster
+        .wait_for_slots_processing(&agent_id, 2)
+        .await
+        .expect("both decisions must reach the agent");
+    answering_stream
+        .next()
+        .await
+        .expect("the long decision must keep answering")
+        .expect("the next answer must be readable");
+
+    cluster
+        .agents
+        .pop()
+        .expect("the cluster must have an agent")
+        .shutdown()
+        .await
+        .expect("the agent must shut down cleanly with decisions in flight");
+
+    drop(answering_stream);
+    drop(waiting_stream);
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/decision_answers_equal_separate_single_question_decisions.rs b/paddler_tests/tests/decision_answers_equal_separate_single_question_decisions.rs
new file mode 100644
index 000000000..f7b41e634
--- /dev/null
+++ b/paddler_tests/tests/decision_answers_equal_separate_single_question_decisions.rs
@@ -0,0 +1,46 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::request_params::decide_params::raw_decide_params::RawDecideParams;
+use paddler_tests::decision_cluster_params::DecisionClusterParams;
+use paddler_tests::sample_decision::sample_decision;
+use paddler_tests::start_decision_cluster::start_decision_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn decision_answers_equal_separate_single_question_decisions() {
+    let cluster = start_decision_cluster(DecisionClusterParams::default())
+        .await
+        .expect("the cluster must start");
+    let decision = sample_decision();
+
+    let joint_answers = cluster
+        .decide(CancellationToken::new(), &decision)
+        .await
+        .expect("the joint decision must be accepted")
+        .answers;
+    let mut separate_answers = Vec::new();
+
+    for question in &decision.questions {
+        separate_answers.extend(
+            cluster
+                .decide(
+                    CancellationToken::new(),
+                    &RawDecideParams {
+                        questions: vec![question.clone()],
+                        state: decision.state.clone(),
+                    },
+                )
+                .await
+                .expect("a single question decision must be accepted")
+                .answers,
+        );
+    }
+
+    assert_eq!(joint_answers, separate_answers);
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/decision_client_disconnecting_mid_decision_releases_its_sequences.rs b/paddler_tests/tests/decision_client_disconnecting_mid_decision_releases_its_sequences.rs
new file mode 100644
index 000000000..03b3ff9af
--- /dev/null
+++ b/paddler_tests/tests/decision_client_disconnecting_mid_decision_releases_its_sequences.rs
@@ -0,0 +1,61 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::mem::discriminant;
+
+use futures_util::StreamExt as _;
+use tokio_util::sync::CancellationToken;
+
+use paddler_tests::decision_cluster_params::DecisionClusterParams;
+use paddler_tests::decision_done_discriminant::decision_done_discriminant;
+use paddler_tests::many_question_decision::many_question_decision;
+use paddler_tests::questions_outlasting_a_disconnect::QUESTIONS_OUTLASTING_A_DISCONNECT;
+use paddler_tests::sample_decision::sample_decision;
+use paddler_tests::start_decision_cluster::start_decision_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn decision_client_disconnecting_mid_decision_releases_its_sequences() {
+    let mut cluster = start_decision_cluster(DecisionClusterParams::default())
+        .await
+        .expect("the cluster must start");
+    let agent_id = cluster
+        .agent_ids
+        .first()
+        .expect("the cluster must have an agent")
+        .clone();
+    let mut decision_stream = cluster
+        .client_inference
+        .post_decide(
+            CancellationToken::new(),
+            &many_question_decision(QUESTIONS_OUTLASTING_A_DISCONNECT),
+        )
+        .await
+        .expect("the decision must be accepted");
+
+    decision_stream
+        .next()
+        .await
+        .expect("the decision must answer its first question")
+        .expect("the first answer must be readable");
+    drop(decision_stream);
+
+    cluster
+        .wait_for_slots_processing(&agent_id, 0)
+        .await
+        .expect("the agent must release the abandoned decision");
+
+    let collected = cluster
+        .decide(CancellationToken::new(), &sample_decision())
+        .await
+        .expect("a decision after the abandoned one must be accepted");
+
+    assert_eq!(collected.answers.len(), sample_decision().questions.len());
+    assert_eq!(
+        discriminant(&collected.terminal_result),
+        decision_done_discriminant()
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/decision_exceeding_the_context_is_rejected.rs b/paddler_tests/tests/decision_exceeding_the_context_is_rejected.rs
new file mode 100644
index 000000000..f91bff0b5
--- /dev/null
+++ b/paddler_tests/tests/decision_exceeding_the_context_is_rejected.rs
@@ -0,0 +1,50 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::decision_answer::DecisionAnswer;
+use paddler_messaging::decision_result::DecisionResult;
+use paddler_messaging::oversized_decision_details::OversizedDecisionDetails;
+use paddler_messaging::request_params::decide_params::raw_decide_params::RawDecideParams;
+use paddler_tests::sample_decision::sample_decision;
+use paddler_tests::start_decision_cluster::start_decision_cluster;
+
+use crate::small_context_decision_cluster_params::SMALL_DECISION_CONTEXT_SIZE;
+use crate::small_context_decision_cluster_params::small_context_decision_cluster_params;
+
+const STATE_REPETITIONS_EXCEEDING_THE_CONTEXT: usize = 20;
+const TOKENS_REQUIRED_BY_THE_REPEATED_STATE_DECISION: usize = 602;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn decision_exceeding_the_context_is_rejected() {
+    let cluster = start_decision_cluster(small_context_decision_cluster_params())
+        .await
+        .expect("the cluster must start");
+
+    let collected = cluster
+        .decide(
+            CancellationToken::new(),
+            &RawDecideParams {
+                state: sample_decision()
+                    .state
+                    .repeat(STATE_REPETITIONS_EXCEEDING_THE_CONTEXT),
+                ..sample_decision()
+            },
+        )
+        .await
+        .expect("the oversized decision must be answered with a rejection");
+
+    assert_eq!(collected.answers, Vec::<DecisionAnswer>::new());
+    assert_eq!(
+        collected.terminal_result,
+        DecisionResult::RequestExceedsContext(OversizedDecisionDetails {
+            context_size: SMALL_DECISION_CONTEXT_SIZE,
+            required_tokens: TOKENS_REQUIRED_BY_THE_REPEATED_STATE_DECISION,
+        })
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/decision_in_flight_is_answered_in_full_when_the_desired_state_changes.rs b/paddler_tests/tests/decision_in_flight_is_answered_in_full_when_the_desired_state_changes.rs
new file mode 100644
index 000000000..903c549f6
--- /dev/null
+++ b/paddler_tests/tests/decision_in_flight_is_answered_in_full_when_the_desired_state_changes.rs
@@ -0,0 +1,79 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::mem::discriminant;
+use std::num::NonZeroU32;
+
+use futures_util::StreamExt as _;
+use tokio_util::sync::CancellationToken;
+
+use paddler_inference_parameters::model_runtime_parameters::ModelRuntimeParameters;
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_test_cluster_harness::collect_decision_results::collect_decision_results;
+use paddler_tests::decision_cluster_params::DecisionClusterParams;
+use paddler_tests::decision_done_discriminant::decision_done_discriminant;
+use paddler_tests::many_question_decision::many_question_decision;
+use paddler_tests::start_decision_cluster::start_decision_cluster;
+
+const QUESTIONS_OUTLASTING_A_STATE_CHANGE: usize = 96;
+const CHANGED_CONTEXT_SIZE: u32 = 4096;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn decision_in_flight_is_answered_in_full_when_the_desired_state_changes() {
+    let cluster = start_decision_cluster(DecisionClusterParams::default())
+        .await
+        .expect("the cluster must start");
+    let mut decision_stream = cluster
+        .client_inference
+        .post_decide(
+            CancellationToken::new(),
+            &many_question_decision(QUESTIONS_OUTLASTING_A_STATE_CHANGE),
+        )
+        .await
+        .expect("the decision must be accepted");
+
+    decision_stream
+        .next()
+        .await
+        .expect("the decision must answer its first question")
+        .expect("the first answer must be readable");
+
+    let desired_state = cluster
+        .client_management
+        .get_balancer_desired_state(CancellationToken::new())
+        .await
+        .expect("the desired state must be readable");
+
+    cluster
+        .client_management
+        .put_balancer_desired_state(
+            CancellationToken::new(),
+            &BalancerDesiredState {
+                model_runtime_parameters: ModelRuntimeParameters {
+                    context_size: NonZeroU32::new(CHANGED_CONTEXT_SIZE)
+                        .expect("the changed context size is not zero"),
+                    ..desired_state.model_runtime_parameters
+                },
+                ..desired_state
+            },
+        )
+        .await
+        .expect("the changed desired state must be accepted");
+
+    let collected = collect_decision_results(decision_stream)
+        .await
+        .expect("the decision in flight must finish");
+
+    assert_eq!(
+        collected.answers.len(),
+        QUESTIONS_OUTLASTING_A_STATE_CHANGE - 1
+    );
+    assert_eq!(
+        discriminant(&collected.terminal_result),
+        decision_done_discriminant()
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/decision_input_with_a_nul_byte_is_rejected.rs b/paddler_tests/tests/decision_input_with_a_nul_byte_is_rejected.rs
new file mode 100644
index 000000000..68408b35a
--- /dev/null
+++ b/paddler_tests/tests/decision_input_with_a_nul_byte_is_rejected.rs
@@ -0,0 +1,92 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::ffi::CString;
+
+use llama_cpp_bindings::error::StringToTokenError;
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent_decision::decision_error::DecisionError;
+use paddler_agent_runtime::rejection_description::rejection_description;
+use paddler_messaging::decision_answer::DecisionAnswer;
+use paddler_messaging::decision_result::DecisionResult;
+use paddler_messaging::request_params::decide_params::raw_decide_params::RawDecideParams;
+use paddler_tests::decision_cluster_params::DecisionClusterParams;
+use paddler_tests::sample_decision::sample_decision;
+use paddler_tests::start_decision_cluster::start_decision_cluster;
+
+use crate::with_nul_byte::with_nul_byte;
+
+struct DecisionRejectedForANulByte {
+    decision: RawDecideParams,
+    text_with_nul_byte: String,
+}
+
+#[tokio::test(flavor = "multi_thread")]
+async fn decision_input_with_a_nul_byte_is_rejected() {
+    let cluster = start_decision_cluster(DecisionClusterParams::default())
+        .await
+        .expect("the cluster must start");
+    let RawDecideParams { questions, state } = sample_decision();
+    let mut nul_in_leading_instructions = questions.clone();
+    let mut nul_in_last_option = questions.clone();
+
+    nul_in_leading_instructions[0].instructions =
+        with_nul_byte(&nul_in_leading_instructions[0].instructions);
+
+    let last_question = nul_in_last_option
+        .last_mut()
+        .expect("the sample decision has questions");
+
+    last_question.options[0] = with_nul_byte(&last_question.options[0]);
+
+    let leading_instructions_with_nul_byte = nul_in_leading_instructions[0].instructions.clone();
+    let last_option_with_nul_byte = last_question.options[0].clone();
+
+    for DecisionRejectedForANulByte {
+        decision,
+        text_with_nul_byte,
+    } in [
+        DecisionRejectedForANulByte {
+            decision: RawDecideParams {
+                questions: questions.clone(),
+                state: with_nul_byte(&state),
+            },
+            text_with_nul_byte: with_nul_byte(&state),
+        },
+        DecisionRejectedForANulByte {
+            decision: RawDecideParams {
+                questions: nul_in_leading_instructions,
+                state: state.clone(),
+            },
+            text_with_nul_byte: leading_instructions_with_nul_byte,
+        },
+        DecisionRejectedForANulByte {
+            decision: RawDecideParams {
+                questions: nul_in_last_option,
+                state: state.clone(),
+            },
+            text_with_nul_byte: last_option_with_nul_byte,
+        },
+    ] {
+        let collected = cluster
+            .decide(CancellationToken::new(), &decision)
+            .await
+            .expect("the decision must be answered with a rejection");
+        let nul_error =
+            CString::new(text_with_nul_byte).expect_err("the text must contain a nul byte");
+
+        assert_eq!(collected.answers, Vec::<DecisionAnswer>::new());
+        assert_eq!(
+            collected.terminal_result,
+            DecisionResult::InputTokenizationFailed(rejection_description(
+                Some("test-agent-0"),
+                &DecisionError::InputTokenizationFailed(StringToTokenError::NulError(nul_error))
+            ))
+        );
+    }
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/decision_rows_longer_than_the_batch_are_answered.rs b/paddler_tests/tests/decision_rows_longer_than_the_batch_are_answered.rs
new file mode 100644
index 000000000..842e5020a
--- /dev/null
+++ b/paddler_tests/tests/decision_rows_longer_than_the_batch_are_answered.rs
@@ -0,0 +1,75 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::mem::discriminant;
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_inference_parameters::batch_size::BatchSize;
+use paddler_messaging::decision_question::DecisionQuestion;
+use paddler_messaging::request_params::decide_params::raw_decide_params::RawDecideParams;
+use paddler_tests::decision_cluster_params::DecisionClusterParams;
+use paddler_tests::decision_desired_state::DecisionDesiredState;
+use paddler_tests::decision_done_discriminant::decision_done_discriminant;
+use paddler_tests::sample_decision::sample_decision;
+use paddler_tests::start_decision_cluster::start_decision_cluster;
+
+const SMALL_BATCH_TOKENS: u32 = 16;
+const PROBABILITY_SUM_TOLERANCE: f32 = 1e-5;
+
+fn long_question(id: &str) -> DecisionQuestion {
+    DecisionQuestion {
+        id: id.to_owned(),
+        instructions: "Read the whole ticket carefully and decide how the customer feels about \
+                       the delivery, the payment and the support they received."
+            .to_owned(),
+        options: vec![
+            "positive, because everything went well".to_owned(),
+            "negative, because the delivery was slow and the payment was late".to_owned(),
+        ],
+    }
+}
+
+#[tokio::test(flavor = "multi_thread")]
+async fn decision_rows_longer_than_the_batch_are_answered() {
+    let cluster = start_decision_cluster(DecisionClusterParams {
+        desired_state: DecisionDesiredState {
+            n_batch: BatchSize::try_from(SMALL_BATCH_TOKENS).expect("the batch size is valid"),
+            ..DecisionDesiredState::default()
+        },
+        ..DecisionClusterParams::default()
+    })
+    .await
+    .expect("the cluster must start");
+
+    let collected = cluster
+        .decide(
+            CancellationToken::new(),
+            &RawDecideParams {
+                questions: vec![long_question("first"), long_question("second")],
+                state: sample_decision().state,
+            },
+        )
+        .await
+        .expect("the decision must be accepted");
+
+    assert_eq!(
+        collected
+            .answers
+            .iter()
+            .map(|answer| answer.id.as_str())
+            .collect::<Vec<_>>(),
+        vec!["first", "second"]
+    );
+    assert!(collected.answers.iter().all(|answer| {
+        (answer.probabilities.iter().sum::<f32>() - 1.0).abs() < PROBABILITY_SUM_TOLERANCE
+    }));
+    assert_eq!(
+        discriminant(&collected.terminal_result),
+        decision_done_discriminant()
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/decision_without_questions_is_rejected_with_bad_request.rs b/paddler_tests/tests/decision_without_questions_is_rejected_with_bad_request.rs
new file mode 100644
index 000000000..c5d95283f
--- /dev/null
+++ b/paddler_tests/tests/decision_without_questions_is_rejected_with_bad_request.rs
@@ -0,0 +1,51 @@
+use reqwest::StatusCode;
+use tokio_util::sync::CancellationToken;
+
+use paddler_client::error::Error as ClientError;
+use paddler_messaging::api_path::ApiPath;
+use paddler_messaging::request_params::decide_params::raw_decide_params::RawDecideParams;
+use paddler_tests::cluster_without_agents_serving_decisions::cluster_without_agents_serving_decisions;
+use paddler_tests::start_cluster::start_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn decision_without_questions_is_rejected_with_bad_request() {
+    let cluster = start_cluster(cluster_without_agents_serving_decisions())
+        .await
+        .expect("the cluster must start");
+
+    let decide_url = cluster
+        .balancer
+        .inference_base_url
+        .join(ApiPath::DECIDE)
+        .expect("the decide path must extend the inference address");
+    let rejection = cluster
+        .decide(
+            CancellationToken::new(),
+            &RawDecideParams {
+                questions: Vec::new(),
+                state: "state".to_owned(),
+            },
+        )
+        .await
+        .expect_err("a decision without questions must be rejected");
+
+    assert_eq!(
+        rejection
+            .downcast_ref::<ClientError>()
+            .map(ToString::to_string),
+        Some(
+            ClientError::UnexpectedResponseStatus {
+                message: "Invalid request parameters: a decision needs at least one question"
+                    .to_owned(),
+                status: StatusCode::BAD_REQUEST,
+                url: decide_url.to_string(),
+            }
+            .to_string()
+        )
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/decisions_are_identical_alone_and_concurrently.rs b/paddler_tests/tests/decisions_are_identical_alone_and_concurrently.rs
new file mode 100644
index 000000000..3cd75bf25
--- /dev/null
+++ b/paddler_tests/tests/decisions_are_identical_alone_and_concurrently.rs
@@ -0,0 +1,60 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use futures_util::future::join_all;
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::request_params::decide_params::raw_decide_params::RawDecideParams;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_tests::decision_cluster_params::DecisionClusterParams;
+use paddler_tests::sample_decision::sample_decision;
+use paddler_tests::start_decision_cluster::start_decision_cluster;
+
+const CONCURRENT_DECISIONS: usize = 5;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn decisions_are_identical_alone_and_concurrently() {
+    let cluster = start_decision_cluster(DecisionClusterParams {
+        agents: vec![AgentConfig::single(2)],
+        ..DecisionClusterParams::default()
+    })
+    .await
+    .expect("the cluster must start");
+    let decisions: Vec<RawDecideParams> = (0..CONCURRENT_DECISIONS)
+        .map(|decision_index| RawDecideParams {
+            state: format!("Ticket {decision_index}: {}", sample_decision().state),
+            ..sample_decision()
+        })
+        .collect();
+    let mut answers_alone = Vec::new();
+
+    for decision in &decisions {
+        answers_alone.push(
+            cluster
+                .decide(CancellationToken::new(), decision)
+                .await
+                .expect("a decision alone must be accepted")
+                .answers,
+        );
+    }
+
+    let answers_concurrently: Vec<_> = join_all(
+        decisions
+            .iter()
+            .map(|decision| cluster.decide(CancellationToken::new(), decision)),
+    )
+    .await
+    .into_iter()
+    .map(|collected| {
+        collected
+            .expect("a concurrent decision must be accepted")
+            .answers
+    })
+    .collect();
+
+    assert_eq!(answers_concurrently, answers_alone);
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/decisions_on_an_attention_only_model_read_the_hidden_states_of_their_own_input.rs b/paddler_tests/tests/decisions_on_an_attention_only_model_read_the_hidden_states_of_their_own_input.rs
new file mode 100644
index 000000000..2ef7c5b6b
--- /dev/null
+++ b/paddler_tests/tests/decisions_on_an_attention_only_model_read_the_hidden_states_of_their_own_input.rs
@@ -0,0 +1,77 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use paddler_messaging::decision_question::DecisionQuestion;
+use paddler_messaging::request_params::decide_params::raw_decide_params::RawDecideParams;
+use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
+use paddler_tests::decision_backbone::DecisionBackbone;
+use paddler_tests::decision_cluster_params::DecisionClusterParams;
+use paddler_tests::decision_desired_state::DecisionDesiredState;
+use paddler_tests::sample_decision::sample_decision;
+use paddler_tests::start_decision_cluster::start_decision_cluster;
+
+use crate::answered_decision_probabilities::answered_decision_probabilities;
+
+fn decision_with_many_options() -> RawDecideParams {
+    RawDecideParams {
+        questions: vec![DecisionQuestion {
+            id: "weekday".to_owned(),
+            instructions: "Which day was the parcel delivered?".to_owned(),
+            options: [
+                "monday",
+                "tuesday",
+                "wednesday",
+                "thursday",
+                "friday",
+                "saturday",
+                "sunday",
+            ]
+            .iter()
+            .map(|weekday| (*weekday).to_owned())
+            .collect(),
+        }],
+        state: "The parcel arrived on the last working day of the week.".to_owned(),
+    }
+}
+
+#[tokio::test(flavor = "multi_thread")]
+async fn decisions_on_an_attention_only_model_read_the_hidden_states_of_their_own_input() {
+    let cluster = start_decision_cluster(DecisionClusterParams {
+        desired_state: DecisionDesiredState {
+            backbone: DecisionBackbone::HuggingFace(qwen3_0_6b()),
+            ..DecisionDesiredState::default()
+        },
+        ..DecisionClusterParams::default()
+    })
+    .await
+    .expect("a decision cluster on an attention-only model must start");
+
+    let invoice_before_other_decisions =
+        answered_decision_probabilities(&cluster, &sample_decision()).await;
+
+    answered_decision_probabilities(&cluster, &decision_with_many_options()).await;
+
+    let weather_with_the_same_questions = answered_decision_probabilities(
+        &cluster,
+        &RawDecideParams {
+            state: "The weather in Oslo is cold, dark and windy all week.".to_owned(),
+            ..sample_decision()
+        },
+    )
+    .await;
+    let invoice_after_other_decisions =
+        answered_decision_probabilities(&cluster, &sample_decision()).await;
+
+    assert_eq!(
+        invoice_before_other_decisions,
+        invoice_after_other_decisions
+    );
+    assert_ne!(
+        invoice_before_other_decisions,
+        weather_with_the_same_questions
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/decode_prompt_for_sampling.rs b/paddler_tests/tests/decode_prompt_for_sampling.rs
new file mode 100644
index 000000000..9dee1a343
--- /dev/null
+++ b/paddler_tests/tests/decode_prompt_for_sampling.rs
@@ -0,0 +1,30 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use llama_cpp_bindings::context::LlamaContext;
+use llama_cpp_bindings::llama_batch::LlamaBatch;
+use llama_cpp_bindings::model::AddBos;
+use llama_cpp_bindings::model::ParseSpecialTokens;
+
+use paddler_tests::llama_runtime::LlamaRuntime;
+
+pub fn decode_prompt_for_sampling(
+    runtime: &LlamaRuntime,
+    llama_context: &mut LlamaContext<'_>,
+    prompt: &str,
+) -> i32 {
+    let prompt_tokens = runtime
+        .loaded_llama_model
+        .model
+        .str_to_token(prompt, AddBos::Always, ParseSpecialTokens::Always)
+        .expect("the prompt must tokenize");
+    let mut batch = LlamaBatch::new(runtime.model_runtime_parameters.n_batch.capacity(), 1);
+
+    batch
+        .add_sequence(&prompt_tokens, 0, false)
+        .expect("the prompt must fit the batch");
+    llama_context
+        .decode(&mut batch)
+        .expect("the prompt must decode");
+
+    batch.n_tokens() - 1
+}
diff --git a/paddler_tests/tests/deepseek_r1_distill_llama_8b_internal_endpoint_emits_reasoning_tokens.rs b/paddler_tests/tests/deepseek_r1_distill_llama_8b_internal_endpoint_emits_reasoning_tokens.rs
index bbb73bfa6..87ea8c9c8 100644
--- a/paddler_tests/tests/deepseek_r1_distill_llama_8b_internal_endpoint_emits_reasoning_tokens.rs
+++ b/paddler_tests/tests/deepseek_r1_distill_llama_8b_internal_endpoint_emits_reasoning_tokens.rs
@@ -23,12 +23,14 @@ async fn deepseek_r1_distill_llama_8b_internal_endpoint_emits_reasoning_tokens()
             CancellationToken::new(),
             &ContinueFromConversationHistoryParams {
                 add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Text(
-                        "What is two plus two? Think step by step.".to_owned(),
-                    ),
-                    role: "user".to_owned(),
-                }]),
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Text(
+                            "What is two plus two? Think step by step.".to_owned(),
+                        ),
+                        role: "user".to_owned(),
+                    },
+                ]),
                 enable_thinking: true,
                 grammar: None,
                 max_tokens: NonZeroU32::new(400).unwrap(),
@@ -50,13 +52,9 @@ async fn deepseek_r1_distill_llama_8b_internal_endpoint_emits_reasoning_tokens()
         "DeepSeek-R1-8B: expected at least one reasoning token from a `<think>` block (got {reasoning_count})"
     );
 
-    let last = collected
-        .token_results
-        .last()
-        .expect("no token results received");
-    let GeneratedTokenResult::Done(summary) = &last.token_result else {
-        panic!("last result was not Done: {last:?}");
-    };
+    let summary = collected
+        .summary()
+        .expect("the generation must finish with a summary");
 
     assert!(summary.usage.prompt_tokens > 0);
     assert!(summary.usage.reasoning_tokens > 0);
@@ -67,22 +65,8 @@ async fn deepseek_r1_distill_llama_8b_internal_endpoint_emits_reasoning_tokens()
             + summary.usage.undeterminable_tokens
     );
 
-    let reasoning_stream: String = collected
-        .token_results
-        .iter()
-        .filter_map(|result| match &result.token_result {
-            GeneratedTokenResult::ReasoningToken(piece) => Some(piece.as_str()),
-            _ => None,
-        })
-        .collect();
-    let content_stream: String = collected
-        .token_results
-        .iter()
-        .filter_map(|result| match &result.token_result {
-            GeneratedTokenResult::ContentToken(piece) => Some(piece.as_str()),
-            _ => None,
-        })
-        .collect();
+    let reasoning_stream = collected.reasoning_text();
+    let content_stream = collected.content_text();
 
     for forbidden in ["<think>", "</think>"] {
         assert!(
diff --git a/paddler_tests/tests/embedding_documents_fixture.rs b/paddler_tests/tests/embedding_documents_fixture.rs
new file mode 100644
index 000000000..6e954bd43
--- /dev/null
+++ b/paddler_tests/tests/embedding_documents_fixture.rs
@@ -0,0 +1,56 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::num::NonZeroU16;
+use std::sync::Arc;
+
+use tokio::sync::mpsc;
+
+use paddler_agent_runtime::agent_request::AgentRequest;
+use paddler_agent_status::slot_aggregated_status::SlotAggregatedStatus;
+use paddler_agent_status::slot_guard::SlotGuard;
+use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
+use paddler_messaging::embedding_result::EmbeddingResult;
+use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
+use paddler_tests::numbered_embedding_documents::numbered_embedding_documents;
+
+pub struct EmbeddingDocumentsFixture {
+    pub agent_request: AgentRequest<GenerateEmbeddingBatchParams, EmbeddingResult>,
+    pub generate_embedding_stop_tx: mpsc::UnboundedSender<()>,
+    pub generated_embedding_rx: mpsc::UnboundedReceiver<EmbeddingResult>,
+}
+
+impl EmbeddingDocumentsFixture {
+    pub fn embedded_one_at_a_time(
+        document_count: usize,
+        slot_aggregated_status: Arc<SlotAggregatedStatus>,
+    ) -> Self {
+        let (generated_embedding_tx, generated_embedding_rx) = mpsc::unbounded_channel();
+        let (generate_embedding_stop_tx, generate_embedding_stop_rx) = mpsc::unbounded_channel();
+
+        Self {
+            agent_request: AgentRequest {
+                params: GenerateEmbeddingBatchParams {
+                    input_batch: numbered_embedding_documents(document_count),
+                    normalization_method: EmbeddingNormalizationMethod::None,
+                },
+                response_tx: generated_embedding_tx,
+                slot_guard: SlotGuard::new(slot_aggregated_status, NonZeroU16::MIN),
+                stop_rx: generate_embedding_stop_rx,
+            },
+            generate_embedding_stop_tx,
+            generated_embedding_rx,
+        }
+    }
+
+    pub async fn collect_results(
+        mut generated_embedding_rx: mpsc::UnboundedReceiver<EmbeddingResult>,
+    ) -> Vec<EmbeddingResult> {
+        let mut embedding_results = Vec::new();
+
+        while let Some(embedding_result) = generated_embedding_rx.recv().await {
+            embedding_results.push(embedding_result);
+        }
+
+        embedding_results
+    }
+}
diff --git a/paddler_tests/tests/embeddings_use_the_pooling_the_model_declares_by_default.rs b/paddler_tests/tests/embeddings_use_the_pooling_the_model_declares_by_default.rs
new file mode 100644
index 000000000..89ea12dd3
--- /dev/null
+++ b/paddler_tests/tests/embeddings_use_the_pooling_the_model_declares_by_default.rs
@@ -0,0 +1,24 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::num::NonZeroUsize;
+
+use paddler_inference_parameters::embedding_parameters::EmbeddingParameters;
+use paddler_inference_parameters::pooling_type::PoolingType;
+
+use crate::nomic_document_embedding_with::nomic_document_embedding_with;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn embeddings_use_the_pooling_the_model_declares_by_default() {
+    assert_eq!(
+        nomic_document_embedding_with(EmbeddingParameters {
+            embedding_batch_size: NonZeroUsize::MIN,
+            ..EmbeddingParameters::default()
+        })
+        .await,
+        nomic_document_embedding_with(EmbeddingParameters {
+            embedding_batch_size: NonZeroUsize::MIN,
+            pooling_type: PoolingType::Mean,
+        })
+        .await
+    );
+}
diff --git a/paddler_tests/tests/enabling_decisions_reuses_the_loaded_model.rs b/paddler_tests/tests/enabling_decisions_reuses_the_loaded_model.rs
new file mode 100644
index 000000000..a4bc3bc42
--- /dev/null
+++ b/paddler_tests/tests/enabling_decisions_reuses_the_loaded_model.rs
@@ -0,0 +1,53 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use paddler_messaging::agent_decisions_serving::AgentDecisionsServing;
+use paddler_messaging::agent_runtime_status::AgentRuntimeStatus;
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_messaging::decision_settings::DecisionSettings;
+use paddler_messaging::decision_slots_minimum::DECISION_SLOTS_MINIMUM;
+use paddler_messaging::produces_snapshot::ProducesSnapshot as _;
+use paddler_tests::kev_checkpoint::KevCheckpoint;
+
+use crate::serving_pipeline_arbiter::ServingPipelineArbiter;
+use crate::swappable_model_file::SwappableModelFile;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn enabling_decisions_reuses_the_loaded_model() {
+    let swappable_model_file =
+        SwappableModelFile::link_to(&KevCheckpoint::Kev0_8b.target_path("model.gguf"));
+    let decisions_desired_state = BalancerDesiredState {
+        model: swappable_model_file.agent_desired_model(),
+        ..KevCheckpoint::Kev0_8b
+            .decision_desired_state()
+            .into_balancer_desired_state()
+            .expect("the kev desired state must name absolute paths")
+    };
+    let mut serving_pipeline_arbiter = ServingPipelineArbiter::new(DECISION_SLOTS_MINIMUM.get());
+
+    serving_pipeline_arbiter
+        .serve(BalancerDesiredState {
+            decision: DecisionSettings::default(),
+            ..decisions_desired_state.clone()
+        })
+        .await;
+
+    swappable_model_file.replace_with_an_unloadable_file();
+
+    serving_pipeline_arbiter
+        .serve(decisions_desired_state)
+        .await;
+    assert_eq!(
+        serving_pipeline_arbiter
+            .inference_runtime_context
+            .slot_aggregated_status
+            .make_snapshot()
+            .status
+            .runtime,
+        AgentRuntimeStatus::Serving {
+            decisions: AgentDecisionsServing::Served,
+            slots_total: u64::from(DECISION_SLOTS_MINIMUM.get()),
+        }
+    );
+
+    serving_pipeline_arbiter.shut_down().await;
+}
diff --git a/paddler_tests/tests/enabling_decisions_with_too_few_slots_keeps_serving_the_loaded_model.rs b/paddler_tests/tests/enabling_decisions_with_too_few_slots_keeps_serving_the_loaded_model.rs
new file mode 100644
index 000000000..8b8bd34bb
--- /dev/null
+++ b/paddler_tests/tests/enabling_decisions_with_too_few_slots_keeps_serving_the_loaded_model.rs
@@ -0,0 +1,82 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::collections::BTreeSet;
+use std::mem::discriminant;
+
+use paddler_messaging::agent_decisions_serving::AgentDecisionsServing;
+use paddler_messaging::agent_issue::AgentIssue;
+use paddler_messaging::agent_issue_params::slots_insufficient_for_decisions_params::SlotsInsufficientForDecisionsParams;
+use paddler_messaging::agent_runtime_status::AgentRuntimeStatus;
+use paddler_messaging::agent_state_application_status::AgentStateApplicationStatus;
+use paddler_messaging::agent_status::AgentStatus;
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_messaging::decision_settings::DecisionSettings;
+use paddler_messaging::decision_slots_minimum::DECISION_SLOTS_MINIMUM;
+use paddler_messaging::model_download_status::ModelDownloadStatus;
+use paddler_messaging::produces_snapshot::ProducesSnapshot as _;
+use paddler_tests::desired_state_reconciled_discriminant::desired_state_reconciled_discriminant;
+use paddler_tests::kev_checkpoint::KevCheckpoint;
+
+use crate::serving_pipeline_arbiter::ServingPipelineArbiter;
+
+const SLOTS_TOO_FEW_FOR_DECISIONS: u16 = 1;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn enabling_decisions_with_too_few_slots_keeps_serving_the_loaded_model() {
+    let mut serving_pipeline_arbiter = ServingPipelineArbiter::new(SLOTS_TOO_FEW_FOR_DECISIONS);
+
+    serving_pipeline_arbiter
+        .serve(BalancerDesiredState {
+            decision: DecisionSettings::default(),
+            ..KevCheckpoint::Kev0_8b
+                .decision_desired_state()
+                .into_balancer_desired_state()
+                .expect("the kev desired state must name absolute paths")
+        })
+        .await;
+    assert_eq!(
+        discriminant(
+            &serving_pipeline_arbiter
+                .apply(
+                    KevCheckpoint::Kev0_8b
+                        .decision_desired_state()
+                        .into_balancer_desired_state()
+                        .expect("the kev desired state must name absolute paths")
+                )
+                .await
+        ),
+        desired_state_reconciled_discriminant()
+    );
+    serving_pipeline_arbiter.conclude_next_progress().await;
+    assert_eq!(
+        serving_pipeline_arbiter
+            .inference_runtime_context
+            .slot_aggregated_status
+            .make_snapshot()
+            .status,
+        AgentStatus {
+            desired_slots_total: SLOTS_TOO_FEW_FOR_DECISIONS,
+            download_status: ModelDownloadStatus::NotDownloading,
+            issues: BTreeSet::from([AgentIssue::SlotsInsufficientForDecisions(
+                SlotsInsufficientForDecisionsParams {
+                    desired_slots: SLOTS_TOO_FEW_FOR_DECISIONS,
+                    required_slots: DECISION_SLOTS_MINIMUM.get(),
+                }
+            )]),
+            model_path: Some(
+                KevCheckpoint::Kev0_8b
+                    .target_path("model.gguf")
+                    .display()
+                    .to_string()
+            ),
+            runtime: AgentRuntimeStatus::Serving {
+                decisions: AgentDecisionsServing::NotServed,
+                slots_total: u64::from(SLOTS_TOO_FEW_FOR_DECISIONS),
+            },
+            state_application_status: AgentStateApplicationStatus::AttemptedAndRetrying,
+            uses_chat_template_override: false,
+        }
+    );
+
+    serving_pipeline_arbiter.shut_down().await;
+}
diff --git a/paddler_tests/tests/finalizing_a_tool_call_with_nothing_buffered_resolves_to_no_tool_calls.rs b/paddler_tests/tests/finalizing_a_tool_call_with_nothing_buffered_resolves_to_no_tool_calls.rs
new file mode 100644
index 000000000..1c2904894
--- /dev/null
+++ b/paddler_tests/tests/finalizing_a_tool_call_with_nothing_buffered_resolves_to_no_tool_calls.rs
@@ -0,0 +1,26 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use paddler_messaging::generated_token_result::GeneratedTokenResult;
+use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
+use paddler_tests::agent_inference_runtime_context::agent_inference_runtime_context;
+use paddler_tests::llama_runtime::LlamaRuntime;
+
+use crate::tool_call_pipeline_for_tools::tool_call_pipeline_for_tools;
+
+#[test]
+fn finalizing_a_tool_call_with_nothing_buffered_resolves_to_no_tool_calls() {
+    let runtime = LlamaRuntime::load(
+        &qwen3_0_6b()
+            .cached_path()
+            .expect("the model must be in the Hugging Face cache"),
+        agent_inference_runtime_context(1),
+    )
+    .expect("the model must load");
+
+    assert_eq!(
+        tool_call_pipeline_for_tools(&runtime.loaded_llama_model.model, &[])
+            .finalize()
+            .into_generated_token_result(),
+        GeneratedTokenResult::ToolCallParsed(Vec::new())
+    );
+}
diff --git a/paddler_tests/tests/first_message_content_request_preparer.rs b/paddler_tests/tests/first_message_content_request_preparer.rs
new file mode 100644
index 000000000..1457637fc
--- /dev/null
+++ b/paddler_tests/tests/first_message_content_request_preparer.rs
@@ -0,0 +1,65 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use paddler_agent_text_generation::chat_prompt_renderer::ChatPromptRenderer;
+use paddler_agent_text_generation::chat_template_renderer::ChatTemplateRenderer;
+use paddler_agent_text_generation::conversation_prompt_support::ConversationPromptSupport;
+use paddler_agent_text_generation::conversation_support::ConversationSupport;
+use paddler_agent_text_generation::generation_request_preparer::GenerationRequestPreparer;
+use paddler_agent_text_generation::image_input::ImageInput;
+use paddler_agent_text_generation::model_bos_token::ModelBosToken;
+use paddler_agent_text_generation::multimodal_prompt_support::MultimodalPromptSupport;
+use paddler_agent_text_generation::prompt_tokenizer::PromptTokenizer;
+use paddler_messaging::multimodal_settings::MultimodalSettings;
+use paddler_tests::first_message_content_chat_template::first_message_content_chat_template;
+use paddler_tests::llama_runtime::LlamaRuntime;
+
+pub struct FirstMessageContentRequestPreparer<'runtime> {
+    pub llama_runtime: &'runtime LlamaRuntime,
+    pub sequence_context_size: u32,
+}
+
+impl FirstMessageContentRequestPreparer<'_> {
+    pub fn into_generation_request_preparer(self) -> GenerationRequestPreparer {
+        self.into_preparer_with_image_input(ImageInput::Unsupported)
+    }
+
+    pub fn into_multimodal_generation_request_preparer(
+        self,
+        multimodal_prompt_support: MultimodalPromptSupport,
+    ) -> GenerationRequestPreparer {
+        self.into_preparer_with_image_input(ImageInput::Supported(multimodal_prompt_support))
+    }
+
+    fn into_preparer_with_image_input(self, image_input: ImageInput) -> GenerationRequestPreparer {
+        let loaded_llama_model = &self.llama_runtime.loaded_llama_model;
+        let model = &loaded_llama_model.model;
+
+        GenerationRequestPreparer {
+            conversation_support: ConversationSupport::Available(Box::new(
+                ConversationPromptSupport {
+                    chat_prompt_renderer: ChatPromptRenderer {
+                        chat_template_renderer: ChatTemplateRenderer::new(
+                            first_message_content_chat_template(),
+                        )
+                        .expect("the first message content chat template must compile"),
+                        token_bos_str: String::new(),
+                        token_eos_str: String::new(),
+                        token_nl_str: String::new(),
+                    },
+                    image_input,
+                },
+            )),
+            image_resize_to_fit: MultimodalSettings::default().image_resize_to_fit,
+            loaded_llama_model: loaded_llama_model.clone(),
+            prompt_tokenizer: PromptTokenizer {
+                model: model.clone(),
+                model_bos_token: ModelBosToken::of(model),
+                sequence_context_size: self.sequence_context_size,
+            },
+            sampler_chain_factory: self.llama_runtime.sampler_chain_factory(),
+            streaming_markers: model
+                .streaming_markers()
+                .expect("the model must report its streaming markers"),
+        }
+    }
+}
diff --git a/paddler_tests/tests/gemma4_internal_endpoint_emits_reasoning_tokens.rs b/paddler_tests/tests/gemma4_internal_endpoint_emits_reasoning_tokens.rs
index 0ab7b8943..895ebdfda 100644
--- a/paddler_tests/tests/gemma4_internal_endpoint_emits_reasoning_tokens.rs
+++ b/paddler_tests/tests/gemma4_internal_endpoint_emits_reasoning_tokens.rs
@@ -23,12 +23,14 @@ async fn gemma4_internal_endpoint_emits_reasoning_tokens() {
             CancellationToken::new(),
             &ContinueFromConversationHistoryParams {
                 add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Text(
-                        "What is two plus two? Think step by step.".to_owned(),
-                    ),
-                    role: "user".to_owned(),
-                }]),
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Text(
+                            "What is two plus two? Think step by step.".to_owned(),
+                        ),
+                        role: "user".to_owned(),
+                    },
+                ]),
                 enable_thinking: true,
                 grammar: None,
                 max_tokens: NonZeroU32::new(200).unwrap(),
@@ -50,13 +52,9 @@ async fn gemma4_internal_endpoint_emits_reasoning_tokens() {
         "Gemma 4: expected at least one reasoning token from a `<|channel>thought` block (got {reasoning_count})"
     );
 
-    let last = collected
-        .token_results
-        .last()
-        .expect("no token results received");
-    let GeneratedTokenResult::Done(summary) = &last.token_result else {
-        panic!("last result was not Done: {last:?}");
-    };
+    let summary = collected
+        .summary()
+        .expect("the generation must finish with a summary");
 
     assert!(summary.usage.prompt_tokens > 0);
     assert!(summary.usage.reasoning_tokens > 0);
@@ -67,22 +65,8 @@ async fn gemma4_internal_endpoint_emits_reasoning_tokens() {
             + summary.usage.undeterminable_tokens
     );
 
-    let reasoning_stream: String = collected
-        .token_results
-        .iter()
-        .filter_map(|result| match &result.token_result {
-            GeneratedTokenResult::ReasoningToken(piece) => Some(piece.as_str()),
-            _ => None,
-        })
-        .collect();
-    let content_stream: String = collected
-        .token_results
-        .iter()
-        .filter_map(|result| match &result.token_result {
-            GeneratedTokenResult::ContentToken(piece) => Some(piece.as_str()),
-            _ => None,
-        })
-        .collect();
+    let reasoning_stream = collected.reasoning_text();
+    let content_stream = collected.content_text();
 
     for forbidden in ["<|channel>thought", "<channel|>"] {
         assert!(
diff --git a/paddler_tests/tests/gemma4_internal_endpoint_emits_reasoning_tokens_for_image_request.rs b/paddler_tests/tests/gemma4_internal_endpoint_emits_reasoning_tokens_for_image_request.rs
index 9af621890..b8158cf5f 100644
--- a/paddler_tests/tests/gemma4_internal_endpoint_emits_reasoning_tokens_for_image_request.rs
+++ b/paddler_tests/tests/gemma4_internal_endpoint_emits_reasoning_tokens_for_image_request.rs
@@ -23,7 +23,7 @@ async fn gemma4_internal_endpoint_emits_reasoning_tokens_for_image_request() {
 
     let image_data_uri = load_test_image_data_uri().expect("the test image must load");
 
-    let conversation_history = ConversationHistory::new(vec![ConversationMessage {
+    let conversation_history = ConversationHistory::new(vec![ConversationMessage::Message {
         content: ConversationMessageContent::Parts(vec![
             ConversationMessageContentPart::ImageUrl {
                 image_url: ImageUrl {
@@ -64,13 +64,9 @@ async fn gemma4_internal_endpoint_emits_reasoning_tokens_for_image_request() {
         "Gemma 4: expected at least one reasoning token from a `<|channel>thought` block when an image is attached (got {reasoning_count})"
     );
 
-    let last = collected
-        .token_results
-        .last()
-        .expect("no token results received");
-    let GeneratedTokenResult::Done(summary) = &last.token_result else {
-        panic!("last result was not Done: {last:?}");
-    };
+    let summary = collected
+        .summary()
+        .expect("the generation must finish with a summary");
 
     assert!(summary.usage.reasoning_tokens > 0);
     assert!(summary.usage.input_image_tokens > 0);
diff --git a/paddler_tests/tests/gemma4_internal_endpoint_emits_tool_call_parsed_event.rs b/paddler_tests/tests/gemma4_internal_endpoint_emits_tool_call_parsed_event.rs
index 35360f1aa..dbc1a70d6 100644
--- a/paddler_tests/tests/gemma4_internal_endpoint_emits_tool_call_parsed_event.rs
+++ b/paddler_tests/tests/gemma4_internal_endpoint_emits_tool_call_parsed_event.rs
@@ -2,17 +2,18 @@
 
 use std::num::NonZeroU32;
 
-use llama_cpp_bindings_types::ParsedToolCall;
 use tokio_util::sync::CancellationToken;
 
 use paddler_messaging::conversation_history::ConversationHistory;
 use paddler_messaging::conversation_message::ConversationMessage;
 use paddler_messaging::conversation_message_content::ConversationMessageContent;
 use paddler_messaging::generated_token_result::GeneratedTokenResult;
+use paddler_messaging::generated_tool_call::GeneratedToolCall;
 use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
 use paddler_tests::get_weather_tool::get_weather_tool;
 use paddler_tests::start_cluster_with_gemma_4::start_cluster_with_gemma_4;
+use paddler_tests::weather_tool_question::WEATHER_TOOL_QUESTION;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn gemma4_internal_endpoint_emits_tool_call_parsed_event() {
@@ -25,13 +26,12 @@ async fn gemma4_internal_endpoint_emits_tool_call_parsed_event() {
             CancellationToken::new(),
             &ContinueFromConversationHistoryParams {
                 add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Text(
-                        "What is the weather in Paris? Use the get_weather tool to find out."
-                            .to_owned(),
-                    ),
-                    role: "user".to_owned(),
-                }]),
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Text(WEATHER_TOOL_QUESTION.to_owned()),
+                        role: "user".to_owned(),
+                    },
+                ]),
                 enable_thinking: false,
                 grammar: None,
                 max_tokens: NonZeroU32::new(400).unwrap(),
@@ -42,7 +42,7 @@ async fn gemma4_internal_endpoint_emits_tool_call_parsed_event() {
         .await
         .expect("the inference request must be accepted");
 
-    let parsed_events: Vec<&Vec<ParsedToolCall>> = collected
+    let parsed_events: Vec<&Vec<GeneratedToolCall>> = collected
         .token_results
         .iter()
         .filter_map(|event| match &event.token_result {
diff --git a/paddler_tests/tests/generating_request_keeps_every_token_when_multimodal_request_is_admitted.rs b/paddler_tests/tests/generating_request_keeps_every_token_when_multimodal_request_is_admitted.rs
index 454b02137..b0aa91ab7 100644
--- a/paddler_tests/tests/generating_request_keeps_every_token_when_multimodal_request_is_admitted.rs
+++ b/paddler_tests/tests/generating_request_keeps_every_token_when_multimodal_request_is_admitted.rs
@@ -3,6 +3,7 @@
 use std::num::NonZeroU32;
 
 use futures_util::StreamExt as _;
+use futures_util::stream::iter;
 use tokio::spawn;
 use tokio_util::sync::CancellationToken;
 
@@ -11,9 +12,6 @@ use paddler_messaging::conversation_message::ConversationMessage;
 use paddler_messaging::conversation_message_content::ConversationMessageContent;
 use paddler_messaging::conversation_message_content_part::ConversationMessageContentPart;
 use paddler_messaging::image_url::ImageUrl;
-use paddler_messaging::inference_client::message::Message;
-use paddler_messaging::inference_client::response::Response;
-use paddler_messaging::jsonrpc::response_envelope::ResponseEnvelope;
 use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
 use paddler_test_cluster_harness::collect_generated_tokens::collect_generated_tokens;
@@ -33,38 +31,35 @@ async fn generating_request_keeps_every_token_when_multimodal_request_is_admitte
         .await
         .expect("the inference request must be accepted");
 
-    let Message::Response(ResponseEnvelope {
-        response: Response::GeneratedToken(first_token_result),
-        ..
-    }) = text_stream
+    let first_streamed_item = text_stream
         .next()
         .await
-        .expect("the text request must stream a first token")
-        .expect("the message must be readable")
-    else {
-        panic!("the text request must start by streaming a generated token");
-    };
-
-    let remaining_text = spawn(collect_generated_tokens(text_stream));
+        .expect("the text request must stream a first token");
+    let streamed_text = spawn(collect_generated_tokens(Box::pin(
+        iter([first_streamed_item]).chain(text_stream),
+    )));
 
     cluster
         .continue_from_conversation_history(
             CancellationToken::new(),
             &ContinueFromConversationHistoryParams {
                 add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Parts(vec![
-                        ConversationMessageContentPart::ImageUrl {
-                            image_url: ImageUrl {
-                                url: load_test_image_data_uri().expect("the test image must load"),
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Parts(vec![
+                            ConversationMessageContentPart::ImageUrl {
+                                image_url: ImageUrl {
+                                    url: load_test_image_data_uri()
+                                        .expect("the test image must load"),
+                                },
+                            },
+                            ConversationMessageContentPart::Text {
+                                text: "Describe this image.".to_owned(),
                             },
-                        },
-                        ConversationMessageContentPart::Text {
-                            text: "Describe this image.".to_owned(),
-                        },
-                    ]),
-                    role: "user".to_owned(),
-                }]),
+                        ]),
+                        role: "user".to_owned(),
+                    },
+                ]),
                 enable_thinking: false,
                 grammar: None,
                 max_tokens: NonZeroU32::new(4).unwrap(),
@@ -79,15 +74,11 @@ async fn generating_request_keeps_every_token_when_multimodal_request_is_admitte
 
     text_cancellation.cancel();
 
-    let streamed_text = format!(
-        "{}{}",
-        first_token_result.token_text().unwrap_or_default(),
-        remaining_text
-            .await
-            .expect("the text collection must not panic")
-            .expect("the message must be readable")
-            .text
-    );
+    let streamed_text = streamed_text
+        .await
+        .expect("the text collection must not panic")
+        .expect("the message must be readable")
+        .text;
 
     assert!(
         is_unending_generation_text(&streamed_text),
diff --git a/paddler_tests/tests/grammar_refusing_the_greedy_selection.rs b/paddler_tests/tests/grammar_refusing_the_greedy_selection.rs
new file mode 100644
index 000000000..84a86d275
--- /dev/null
+++ b/paddler_tests/tests/grammar_refusing_the_greedy_selection.rs
@@ -0,0 +1,76 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use llama_cpp_bindings::context::LlamaContext;
+use llama_cpp_bindings::error::SamplerAcceptError;
+use llama_cpp_bindings::sampled_token::SampledToken;
+use llama_cpp_bindings::sampling::LlamaSampler;
+use llama_cpp_bindings::token::LlamaToken;
+
+use paddler_tests::agent_inference_runtime_context::agent_inference_runtime_context;
+use paddler_tests::kev_checkpoint::KevCheckpoint;
+use paddler_tests::llama_runtime::LlamaRuntime;
+
+use crate::decode_prompt_for_sampling::decode_prompt_for_sampling;
+
+const GREEDY_SELECTION_AFTER_THE_PROMPT: LlamaToken = LlamaToken::new(264);
+const GRAMMAR_REFUSING_EVERY_PROSE_TOKEN: &str = r#"root ::= "{{{{{{""#;
+const GRAMMAR_ROOT: &str = "root";
+
+pub struct GrammarRefusingTheGreedySelection {
+    pub runtime: LlamaRuntime,
+}
+
+impl GrammarRefusingTheGreedySelection {
+    pub const PROMPT: &'static str = "Paddler balances";
+
+    pub fn load() -> Self {
+        let runtime = LlamaRuntime::load(
+            &KevCheckpoint::Kev0_8b.target_path("model.gguf"),
+            agent_inference_runtime_context(1),
+        )
+        .expect("the kev model must load");
+
+        assert!(
+            !runtime
+                .loaded_llama_model
+                .model
+                .is_eog_token(&SampledToken::Content(GREEDY_SELECTION_AFTER_THE_PROMPT)),
+            "llama.cpp aborts instead of reporting a grammar that refuses an end-of-generation token"
+        );
+
+        Self { runtime }
+    }
+
+    pub fn refusal() -> SamplerAcceptError {
+        SamplerAcceptError::GrammarStateCorrupted {
+            message: format!(
+                "Unexpected empty grammar stack after accepting piece:  a ({})",
+                GREEDY_SELECTION_AFTER_THE_PROMPT.0
+            ),
+        }
+    }
+
+    pub fn decode_the_prompt(&self, llama_context: &mut LlamaContext<'_>) -> i32 {
+        decode_prompt_for_sampling(&self.runtime, llama_context, Self::PROMPT)
+    }
+
+    pub fn grammar(&self) -> LlamaSampler {
+        LlamaSampler::grammar(
+            &self.runtime.loaded_llama_model.model,
+            GRAMMAR_REFUSING_EVERY_PROSE_TOKEN,
+            GRAMMAR_ROOT,
+        )
+        .expect("the grammar must compile")
+    }
+
+    pub fn grammar_awaiting_the_greedy_selection(&self) -> LlamaSampler {
+        LlamaSampler::grammar_lazy(
+            &self.runtime.loaded_llama_model.model,
+            GRAMMAR_REFUSING_EVERY_PROSE_TOKEN,
+            GRAMMAR_ROOT,
+            &[],
+            &[GREEDY_SELECTION_AFTER_THE_PROMPT],
+        )
+        .expect("the lazy grammar must compile")
+    }
+}
diff --git a/paddler_tests/tests/half_closed_client_stops_conversation_history_generation_before_max_tokens.rs b/paddler_tests/tests/half_closed_client_stops_conversation_history_generation_before_max_tokens.rs
index 6c6662646..c446f8733 100644
--- a/paddler_tests/tests/half_closed_client_stops_conversation_history_generation_before_max_tokens.rs
+++ b/paddler_tests/tests/half_closed_client_stops_conversation_history_generation_before_max_tokens.rs
@@ -2,6 +2,8 @@
 
 use std::num::NonZeroU32;
 
+use serde_json::to_value;
+
 use paddler_messaging::api_path::ApiPath;
 use paddler_messaging::conversation_history::ConversationHistory;
 use paddler_messaging::conversation_message::ConversationMessage;
@@ -14,13 +16,14 @@ use paddler_test_cluster_harness::half_closed_client::HalfClosedClient;
 use paddler_test_cluster_harness::unending_grammar::unending_grammar;
 use paddler_tests::qwen3_desired_state::qwen3_desired_state;
 use paddler_tests::start_cluster::start_cluster;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn half_closed_client_stops_conversation_history_generation_before_max_tokens() {
     let mut cluster = start_cluster(ClusterParams {
         agents: vec![AgentConfig::single(1)],
         wait_for_slots_ready: true,
-        desired_state: Some(qwen3_desired_state()),
+        desired_state: ClusterDesiredState::Apply(Box::new(qwen3_desired_state())),
         ..ClusterParams::default()
     })
     .await
@@ -35,7 +38,7 @@ async fn half_closed_client_stops_conversation_history_generation_before_max_tok
     let params: ContinueFromConversationHistoryParams<RawParametersSchema> =
         ContinueFromConversationHistoryParams {
             add_generation_prompt: true,
-            conversation_history: ConversationHistory::new(vec![ConversationMessage {
+            conversation_history: ConversationHistory::new(vec![ConversationMessage::Message {
                 content: ConversationMessageContent::Text(
                     "Write a very long story about a dragon".to_owned(),
                 ),
@@ -51,7 +54,7 @@ async fn half_closed_client_stops_conversation_history_generation_before_max_tok
     let mut client = HalfClosedClient::post_json_then_half_close(
         cluster.balancer.addresses.inference,
         ApiPath::CONTINUE_FROM_CONVERSATION_HISTORY,
-        &params,
+        &to_value(&params).expect("the request parameters must serialize"),
     )
     .await
     .expect("the half-closed request must be sent");
diff --git a/paddler_tests/tests/half_closed_decision_client_releases_its_sequences.rs b/paddler_tests/tests/half_closed_decision_client_releases_its_sequences.rs
new file mode 100644
index 000000000..4b1726154
--- /dev/null
+++ b/paddler_tests/tests/half_closed_decision_client_releases_its_sequences.rs
@@ -0,0 +1,78 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::mem::discriminant;
+
+use serde_json::to_value;
+use tokio_util::sync::CancellationToken;
+
+use paddler_inference_parameters::batch_size::BatchSize;
+use paddler_messaging::api_path::ApiPath;
+use paddler_messaging::request_params::decide_params::raw_decide_params::RawDecideParams;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::half_closed_client::HalfClosedClient;
+use paddler_tests::decision_cluster_params::DecisionClusterParams;
+use paddler_tests::decision_desired_state::DecisionDesiredState;
+use paddler_tests::decision_done_discriminant::decision_done_discriminant;
+use paddler_tests::sample_decision::sample_decision;
+use paddler_tests::start_decision_cluster::start_decision_cluster;
+
+const SMALL_BATCH_TOKENS: u32 = 32;
+const STATE_REPETITIONS_SPANNING_MANY_DECODES: usize = 120;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn half_closed_decision_client_releases_its_sequences() {
+    let mut cluster = start_decision_cluster(DecisionClusterParams {
+        agents: vec![AgentConfig::single(2)],
+        desired_state: DecisionDesiredState {
+            n_batch: BatchSize::try_from(SMALL_BATCH_TOKENS).expect("the batch size is valid"),
+            ..DecisionDesiredState::default()
+        },
+        ..DecisionClusterParams::default()
+    })
+    .await
+    .expect("the cluster must start");
+    let agent_id = cluster
+        .agent_ids
+        .first()
+        .expect("the cluster must have an agent")
+        .clone();
+    let mut half_closed_client = HalfClosedClient::post_json_then_half_close(
+        cluster.balancer.addresses.inference,
+        ApiPath::DECIDE,
+        &to_value(RawDecideParams {
+            state: sample_decision()
+                .state
+                .repeat(STATE_REPETITIONS_SPANNING_MANY_DECODES),
+            ..sample_decision()
+        })
+        .expect("the decision parameters must serialize"),
+    )
+    .await
+    .expect("the half-closed decision must be sent");
+
+    half_closed_client
+        .half_close()
+        .await
+        .expect("the decision must be half-closed");
+    cluster
+        .wait_for_slots_processing(&agent_id, 0)
+        .await
+        .expect("the agent must release the abandoned decision");
+    drop(half_closed_client);
+
+    let collected = cluster
+        .decide(CancellationToken::new(), &sample_decision())
+        .await
+        .expect("a decision after the abandoned one must be accepted");
+
+    assert_eq!(collected.answers.len(), sample_decision().questions.len());
+    assert_eq!(
+        discriminant(&collected.terminal_result),
+        decision_done_discriminant()
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/half_closed_embedding_batch_client_releases_the_buffered_request.rs b/paddler_tests/tests/half_closed_embedding_batch_client_releases_the_buffered_request.rs
index 796b7af72..3da4af64f 100644
--- a/paddler_tests/tests/half_closed_embedding_batch_client_releases_the_buffered_request.rs
+++ b/paddler_tests/tests/half_closed_embedding_batch_client_releases_the_buffered_request.rs
@@ -1,52 +1,29 @@
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
+use serde_json::to_value;
+
 use paddler_messaging::api_path::ApiPath;
-use paddler_messaging::balancer_desired_state::BalancerDesiredState;
 use paddler_messaging::embedding_input_document::EmbeddingInputDocument;
 use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
 use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
 use paddler_test_cluster_harness::half_closed_client::HalfClosedClient;
-use paddler_test_cluster_harness::raw_agent_socket::RawAgentSocket;
 use paddler_tests::start_cluster::start_cluster;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn half_closed_embedding_batch_client_releases_the_buffered_request() {
-    let mut cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        desired_state: Some(BalancerDesiredState {
-            inference_parameters: InferenceParameters {
-                enable_embeddings: true,
-                ..InferenceParameters::default()
-            },
-            ..BalancerDesiredState::default()
-        }),
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("a balancer serving embeddings must start");
-    let mut agent_without_free_slots = RawAgentSocket::connect(
-        cluster.balancer.addresses.management,
-        "agent-without-free-slots",
-    )
-    .await
-    .expect("the agent connection must be established");
-
-    agent_without_free_slots
-        .register()
+    let mut cluster = start_cluster(ClusterParams::without_agents())
         .await
-        .expect("the agent must register");
-
+        .expect("a balancer serving text generation and embeddings must start");
     let mut client = HalfClosedClient::post_json_then_half_close(
         cluster.balancer.addresses.inference,
         ApiPath::GENERATE_EMBEDDING_BATCH,
-        &GenerateEmbeddingBatchParams {
+        &to_value(GenerateEmbeddingBatchParams {
             input_batch: vec![EmbeddingInputDocument {
                 content: "Hello world".to_owned(),
                 id: "doc-1".to_owned(),
             }],
             normalization_method: EmbeddingNormalizationMethod::None,
-        },
+        })
+        .expect("the embedding batch parameters must serialize"),
     )
     .await
     .expect("the half-closed embedding batch request must be sent");
@@ -54,7 +31,7 @@ async fn half_closed_embedding_batch_client_releases_the_buffered_request() {
     cluster
         .wait_for_buffered_request_count(1)
         .await
-        .expect("the embedding batch must be buffered while no agent slot is free");
+        .expect("the embedding batch must be buffered while no agent serves embeddings");
 
     client
         .half_close()
diff --git a/paddler_tests/tests/half_closed_http_client_releases_the_buffered_request.rs b/paddler_tests/tests/half_closed_http_client_releases_the_buffered_request.rs
index 75cb1a33b..86c73a3b4 100644
--- a/paddler_tests/tests/half_closed_http_client_releases_the_buffered_request.rs
+++ b/paddler_tests/tests/half_closed_http_client_releases_the_buffered_request.rs
@@ -1,5 +1,7 @@
 use std::num::NonZeroU32;
 
+use serde_json::to_value;
+
 use paddler_messaging::api_path::ApiPath;
 use paddler_messaging::conversation_history::ConversationHistory;
 use paddler_messaging::conversation_message::ConversationMessage;
@@ -12,19 +14,15 @@ use paddler_tests::start_cluster::start_cluster;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn half_closed_http_client_releases_the_buffered_request() {
-    let mut cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("the cluster must start");
+    let mut cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("the cluster must start");
 
     let inference_addr = cluster.balancer.addresses.inference;
     let params: ContinueFromConversationHistoryParams<RawParametersSchema> =
         ContinueFromConversationHistoryParams {
             add_generation_prompt: true,
-            conversation_history: ConversationHistory::new(vec![ConversationMessage {
+            conversation_history: ConversationHistory::new(vec![ConversationMessage::Message {
                 content: ConversationMessageContent::Text("hi".to_owned()),
                 role: "user".to_owned(),
             }]),
@@ -37,7 +35,7 @@ async fn half_closed_http_client_releases_the_buffered_request() {
     let mut client = HalfClosedClient::post_json_then_half_close(
         inference_addr,
         ApiPath::CONTINUE_FROM_CONVERSATION_HISTORY,
-        &params,
+        &to_value(&params).expect("the request parameters must serialize"),
     )
     .await
     .expect("the half-closed request must be sent");
diff --git a/paddler_tests/tests/half_closed_openai_non_streaming_client_releases_the_buffered_request.rs b/paddler_tests/tests/half_closed_openai_non_streaming_client_releases_the_buffered_request.rs
deleted file mode 100644
index 2b35d365f..000000000
--- a/paddler_tests/tests/half_closed_openai_non_streaming_client_releases_the_buffered_request.rs
+++ /dev/null
@@ -1,19 +0,0 @@
-use serde_json::json;
-
-use paddler_balancer::compatibility::openai_service::openai_api_path::OpenAIApiPath;
-use paddler_tests::release_half_closed_openai_request::release_half_closed_openai_request;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn half_closed_openai_non_streaming_client_releases_the_buffered_request() {
-    release_half_closed_openai_request(
-        OpenAIApiPath::CHAT_COMPLETIONS,
-        &json!({
-            "max_completion_tokens": 2048,
-            "messages": [{"role": "user", "content": "hi"}],
-            "model": "paddler",
-            "stream": false,
-        }),
-    )
-    .await
-    .expect("a non-streaming OpenAI-compatible request buffers its response inside the handler, so the balancer must still release it when the client half-closes");
-}
diff --git a/paddler_tests/tests/half_closed_openai_responses_client_releases_the_buffered_request.rs b/paddler_tests/tests/half_closed_openai_responses_client_releases_the_buffered_request.rs
deleted file mode 100644
index 3b1621028..000000000
--- a/paddler_tests/tests/half_closed_openai_responses_client_releases_the_buffered_request.rs
+++ /dev/null
@@ -1,19 +0,0 @@
-use serde_json::json;
-
-use paddler_balancer::compatibility::openai_service::openai_api_path::OpenAIApiPath;
-use paddler_tests::release_half_closed_openai_request::release_half_closed_openai_request;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn half_closed_openai_responses_client_releases_the_buffered_request() {
-    release_half_closed_openai_request(
-        OpenAIApiPath::RESPONSES,
-        &json!({
-            "input": "hi",
-            "max_output_tokens": 2048,
-            "model": "paddler",
-            "stream": true,
-        }),
-    )
-    .await
-    .expect("the OpenAI-compatible responses stream must release its buffered request when the client half-closes");
-}
diff --git a/paddler_tests/tests/half_closed_openai_streaming_client_releases_the_buffered_request.rs b/paddler_tests/tests/half_closed_openai_streaming_client_releases_the_buffered_request.rs
deleted file mode 100644
index 09c63c5a7..000000000
--- a/paddler_tests/tests/half_closed_openai_streaming_client_releases_the_buffered_request.rs
+++ /dev/null
@@ -1,19 +0,0 @@
-use serde_json::json;
-
-use paddler_balancer::compatibility::openai_service::openai_api_path::OpenAIApiPath;
-use paddler_tests::release_half_closed_openai_request::release_half_closed_openai_request;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn half_closed_openai_streaming_client_releases_the_buffered_request() {
-    release_half_closed_openai_request(
-        OpenAIApiPath::CHAT_COMPLETIONS,
-        &json!({
-            "max_completion_tokens": 2048,
-            "messages": [{"role": "user", "content": "hi"}],
-            "model": "paddler",
-            "stream": true,
-        }),
-    )
-    .await
-    .expect("the OpenAI-compatible SSE stream must release its buffered request when the client half-closes");
-}
diff --git a/paddler_tests/tests/http_client_reports_a_body_that_is_not_json_as_undecodable.rs b/paddler_tests/tests/http_client_reports_a_body_that_is_not_json_as_undecodable.rs
new file mode 100644
index 000000000..1af4bf438
--- /dev/null
+++ b/paddler_tests/tests/http_client_reports_a_body_that_is_not_json_as_undecodable.rs
@@ -0,0 +1,33 @@
+use reqwest::Url;
+use tokio_util::sync::CancellationToken;
+
+use paddler_client::error::Error;
+use paddler_client::format_api_url::format_api_url;
+use paddler_client::http_client::HttpClient;
+use paddler_messaging::api_path::ApiPath;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_tests::start_cluster::start_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn http_client_reports_a_body_that_is_not_json_as_undecodable() {
+    let cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("a balancer without agents must start");
+    let health_url = format_api_url(&cluster.balancer.management_base_url, ApiPath::HEALTH);
+
+    let health_as_json = HttpClient::new(cluster.balancer.management_base_url.clone())
+        .get_json::<String>(CancellationToken::new(), ApiPath::HEALTH)
+        .await;
+
+    assert!(matches!(
+        health_as_json,
+        Err(Error::Http(decoding_error))
+            if decoding_error.is_decode()
+                && decoding_error.url().map(Url::as_str) == Some(health_url.as_str())
+    ));
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/http_client_reports_a_server_sent_event_of_another_shape_as_undecodable.rs b/paddler_tests/tests/http_client_reports_a_server_sent_event_of_another_shape_as_undecodable.rs
new file mode 100644
index 000000000..460c92a8d
--- /dev/null
+++ b/paddler_tests/tests/http_client_reports_a_server_sent_event_of_another_shape_as_undecodable.rs
@@ -0,0 +1,47 @@
+use futures_util::StreamExt as _;
+use serde_json::from_str;
+use serde_json::to_string;
+use tokio_util::sync::CancellationToken;
+
+use paddler_client::error::Error;
+use paddler_client::http_client::HttpClient;
+use paddler_messaging::agent_controller_pool_snapshot::AgentControllerPoolSnapshot;
+use paddler_messaging::api_path::ApiPath;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_tests::start_cluster::start_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn http_client_reports_a_server_sent_event_of_another_shape_as_undecodable() {
+    let cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("a balancer without agents must start");
+    let mut events = HttpClient::new(cluster.balancer.management_base_url.clone())
+        .get_sse_json::<String>(CancellationToken::new(), ApiPath::AGENTS_STREAM)
+        .await
+        .expect("the agents event stream must open");
+    let empty_pool_snapshot = to_string(&AgentControllerPoolSnapshot { agents: Vec::new() })
+        .expect("an empty agents snapshot must serialize");
+
+    assert_eq!(
+        events
+            .next()
+            .await
+            .and_then(Result::err)
+            .as_ref()
+            .map(ToString::to_string),
+        Some(
+            Error::Json(
+                from_str::<String>(&empty_pool_snapshot)
+                    .expect_err("an agents snapshot must not decode as a string")
+            )
+            .to_string()
+        )
+    );
+
+    drop(events);
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/http_inference_cancellation_ends_the_client_stream.rs b/paddler_tests/tests/http_inference_cancellation_ends_the_client_stream.rs
index cef302033..43a18b69b 100644
--- a/paddler_tests/tests/http_inference_cancellation_ends_the_client_stream.rs
+++ b/paddler_tests/tests/http_inference_cancellation_ends_the_client_stream.rs
@@ -9,13 +9,9 @@ use paddler_tests::start_cluster::start_cluster;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn http_inference_cancellation_ends_the_client_stream() {
-    let mut cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("the cluster must start");
+    let mut cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("the cluster must start");
 
     let cancellation_token = CancellationToken::new();
 
diff --git a/paddler_tests/tests/inference_socket_answers_a_conversation_history_request_with_token_generation_disabled_while_serving_embeddings.rs b/paddler_tests/tests/inference_socket_answers_a_conversation_history_request_with_token_generation_disabled_while_serving_embeddings.rs
deleted file mode 100644
index 6a16a9ea2..000000000
--- a/paddler_tests/tests/inference_socket_answers_a_conversation_history_request_with_token_generation_disabled_while_serving_embeddings.rs
+++ /dev/null
@@ -1,76 +0,0 @@
-use std::num::NonZeroU32;
-
-use futures_util::StreamExt as _;
-use tokio_util::sync::CancellationToken;
-
-use paddler_balancer::cluster_token_generation_mode::TOKEN_GENERATION_DISABLED_MESSAGE;
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
-use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-use paddler_messaging::conversation_history::ConversationHistory;
-use paddler_messaging::conversation_message::ConversationMessage;
-use paddler_messaging::conversation_message_content::ConversationMessageContent;
-use paddler_messaging::generated_token_result::GeneratedTokenResult;
-use paddler_messaging::inference_client::message::Message as InferenceClientMessage;
-use paddler_messaging::inference_client::response::Response as InferenceClientResponse;
-use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
-use paddler_test_cluster_harness::cluster_params::ClusterParams;
-use paddler_tests::start_cluster::start_cluster;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn inference_socket_answers_a_conversation_history_request_with_token_generation_disabled_while_serving_embeddings()
- {
-    let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        desired_state: Some(BalancerDesiredState {
-            inference_parameters: InferenceParameters {
-                enable_embeddings: true,
-                ..InferenceParameters::default()
-            },
-            ..BalancerDesiredState::default()
-        }),
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("a balancer serving embeddings must start");
-
-    let answer = cluster
-        .client_inference
-        .continue_from_conversation_history(
-            CancellationToken::new(),
-            ContinueFromConversationHistoryParams {
-                add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Text("What is the weather?".to_owned()),
-                    role: "user".to_owned(),
-                }]),
-                enable_thinking: false,
-                grammar: None,
-                max_tokens: NonZeroU32::MIN,
-                parse_tool_calls: false,
-                tools: Vec::new(),
-            },
-        )
-        .await
-        .expect("the inference socket must accept a conversation history request")
-        .next()
-        .await
-        .expect("the inference socket must answer the conversation history request")
-        .expect("the answer must be readable");
-
-    assert!(matches!(
-        answer,
-        InferenceClientMessage::Response(envelope)
-            if matches!(
-                &envelope.response,
-                InferenceClientResponse::GeneratedToken(
-                    GeneratedTokenResult::TokenGenerationDisabled(description)
-                ) if description == TOKEN_GENERATION_DISABLED_MESSAGE
-            )
-    ));
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_tests/tests/inference_socket_answers_a_ping_with_a_pong.rs b/paddler_tests/tests/inference_socket_answers_a_ping_with_a_pong.rs
index b5351e163..2cdfe3fb4 100644
--- a/paddler_tests/tests/inference_socket_answers_a_ping_with_a_pong.rs
+++ b/paddler_tests/tests/inference_socket_answers_a_ping_with_a_pong.rs
@@ -7,13 +7,9 @@ use paddler_tests::start_cluster::start_cluster;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn inference_socket_answers_a_ping_with_a_pong() {
-    let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("a cluster without agents must start");
+    let cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("a cluster without agents must start");
     let mut socket = RawInferenceSocket::connect(cluster.balancer.addresses.inference)
         .await
         .expect("the raw inference socket must connect");
diff --git a/paddler_tests/tests/inference_socket_answers_an_undeserializable_request_with_an_error.rs b/paddler_tests/tests/inference_socket_answers_an_undeserializable_request_with_an_error.rs
index 0a61477fc..e3a045b48 100644
--- a/paddler_tests/tests/inference_socket_answers_an_undeserializable_request_with_an_error.rs
+++ b/paddler_tests/tests/inference_socket_answers_an_undeserializable_request_with_an_error.rs
@@ -13,10 +13,8 @@ use paddler_tests::start_cluster::start_cluster;
 #[tokio::test(flavor = "multi_thread")]
 async fn inference_socket_answers_an_undeserializable_request_with_an_error() {
     let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
         max_buffered_requests: 0,
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
+        ..ClusterParams::without_agents()
     })
     .await
     .expect("the cluster must start");
diff --git a/paddler_tests/tests/inference_socket_burst_cancellation_serves_every_buffered_request.rs b/paddler_tests/tests/inference_socket_burst_cancellation_serves_every_buffered_request.rs
index 2d68e9d00..c9af9c155 100644
--- a/paddler_tests/tests/inference_socket_burst_cancellation_serves_every_buffered_request.rs
+++ b/paddler_tests/tests/inference_socket_burst_cancellation_serves_every_buffered_request.rs
@@ -5,11 +5,9 @@ use std::num::NonZeroU32;
 use futures_util::StreamExt as _;
 use tokio_util::sync::CancellationToken;
 
-use paddler_messaging::generated_token_result::GeneratedTokenResult;
 use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
 use paddler_test_cluster_harness::collect_generated_tokens::collect_generated_tokens;
-use paddler_test_cluster_harness::token_result_with_producer::TokenResultWithProducer;
 use paddler_test_cluster_harness::unending_generation::unending_generation;
 use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
 
@@ -97,17 +95,9 @@ async fn inference_socket_burst_cancellation_serves_every_buffered_request() {
             .await
             .expect("the generated tokens must be collected");
 
-        assert!(
-            matches!(
-                collected.token_results.last(),
-                Some(TokenResultWithProducer {
-                    token_result: GeneratedTokenResult::Done(_),
-                    ..
-                })
-            ),
-            "every buffered request must run to completion once a slot frees, not end with an error: {:?}",
-            collected.token_results.last()
-        );
+        collected
+            .summary()
+            .expect("every buffered request must run to completion once a slot frees, not end with an error");
         assert!(
             collected
                 .token_results
diff --git a/paddler_tests/tests/inference_socket_cancellation_releases_a_buffered_request.rs b/paddler_tests/tests/inference_socket_cancellation_releases_a_buffered_request.rs
index 18388fbf9..b7ea9c729 100644
--- a/paddler_tests/tests/inference_socket_cancellation_releases_a_buffered_request.rs
+++ b/paddler_tests/tests/inference_socket_cancellation_releases_a_buffered_request.rs
@@ -9,13 +9,9 @@ use paddler_tests::start_cluster::start_cluster;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn inference_socket_cancellation_releases_a_buffered_request() {
-    let mut cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("the cluster must start");
+    let mut cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("the cluster must start");
 
     let cancellation_token = CancellationToken::new();
 
diff --git a/paddler_tests/tests/inference_socket_connection_rejects_requests_after_the_balancer_closes_it.rs b/paddler_tests/tests/inference_socket_connection_rejects_requests_after_the_balancer_closes_it.rs
index efc6994c2..998fd1ea8 100644
--- a/paddler_tests/tests/inference_socket_connection_rejects_requests_after_the_balancer_closes_it.rs
+++ b/paddler_tests/tests/inference_socket_connection_rejects_requests_after_the_balancer_closes_it.rs
@@ -1,5 +1,5 @@
-use tokio::sync::broadcast;
-use tokio::task::yield_now;
+use serde_json::json;
+use tokio::join;
 
 use paddler_client::error::Error;
 use paddler_client::inference_socket::connection::Connection;
@@ -8,35 +8,22 @@ use paddler_tests::start_cluster::start_cluster;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn inference_socket_connection_rejects_requests_after_the_balancer_closes_it() {
-    let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("a cluster without agents must start");
-    let (notification_tx, _notification_rx) = broadcast::channel(1);
-    let connection = Connection::connect(
-        cluster
-            .balancer
-            .inference_base_url()
-            .expect("the inference service must have a base URL"),
-        notification_tx,
-    )
-    .await
-    .expect("the connection to the balancer must open");
-
-    cluster
-        .shutdown()
+    let cluster = start_cluster(ClusterParams::without_agents())
         .await
-        .expect("the cluster must shut down cleanly");
+        .expect("a cluster without agents must start");
+    let connection = Connection::connect(cluster.balancer.inference_base_url.clone())
+        .await
+        .expect("the connection to the balancer must open");
 
-    while !connection.is_disconnected() {
-        yield_now().await;
-    }
+    join!(connection.disconnected(), async {
+        cluster
+            .shutdown()
+            .await
+            .expect("the cluster must shut down cleanly");
+    });
 
     assert!(matches!(
-        connection.send("request-after-close".to_owned(), "{}".to_owned()),
+        connection.send("request-after-close".to_owned(), &json!({})),
         Err(Error::ConnectionDropped { request_id }) if request_id == "request-after-close"
     ));
 }
diff --git a/paddler_tests/tests/inference_socket_duplicate_request_id_is_answered_with_an_error.rs b/paddler_tests/tests/inference_socket_duplicate_request_id_is_answered_with_an_error.rs
index 0a59c760c..fd51ac072 100644
--- a/paddler_tests/tests/inference_socket_duplicate_request_id_is_answered_with_an_error.rs
+++ b/paddler_tests/tests/inference_socket_duplicate_request_id_is_answered_with_an_error.rs
@@ -9,13 +9,9 @@ const DUPLICATE_REQUEST_ID: &str = "duplicate-request-id";
 
 #[tokio::test(flavor = "multi_thread")]
 async fn inference_socket_duplicate_request_id_is_answered_with_an_error() {
-    let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("a cluster without agents must start");
+    let cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("a cluster without agents must start");
     let mut socket = RawInferenceSocket::connect(cluster.balancer.addresses.inference)
         .await
         .expect("the raw inference socket must connect");
diff --git a/paddler_tests/tests/inference_socket_keeps_serving_after_a_binary_frame.rs b/paddler_tests/tests/inference_socket_keeps_serving_after_a_binary_frame.rs
index 9b956921e..a1c94a7ff 100644
--- a/paddler_tests/tests/inference_socket_keeps_serving_after_a_binary_frame.rs
+++ b/paddler_tests/tests/inference_socket_keeps_serving_after_a_binary_frame.rs
@@ -1,5 +1,4 @@
 use std::num::NonZeroU32;
-use std::time::Duration;
 
 use tokio_tungstenite::tungstenite::Bytes;
 use tokio_tungstenite::tungstenite::Message;
@@ -7,16 +6,15 @@ use tokio_tungstenite::tungstenite::Message;
 use paddler_messaging::inference_client::message::Message as InferenceClientMessage;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
 use paddler_test_cluster_harness::raw_inference_socket::RawInferenceSocket;
+use paddler_tests::short_buffered_request_timeout::SHORT_BUFFERED_REQUEST_TIMEOUT;
 use paddler_tests::start_cluster::start_cluster;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn inference_socket_keeps_serving_after_a_binary_frame() {
     let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        buffered_request_timeout: Duration::from_millis(50),
+        buffered_request_timeout: SHORT_BUFFERED_REQUEST_TIMEOUT,
         max_buffered_requests: 1,
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
+        ..ClusterParams::without_agents()
     })
     .await
     .expect("a cluster without agents must start");
diff --git a/paddler_tests/tests/inference_socket_keeps_serving_after_a_stop_request_for_an_unknown_request.rs b/paddler_tests/tests/inference_socket_keeps_serving_after_a_stop_request_for_an_unknown_request.rs
new file mode 100644
index 000000000..264ae38d5
--- /dev/null
+++ b/paddler_tests/tests/inference_socket_keeps_serving_after_a_stop_request_for_an_unknown_request.rs
@@ -0,0 +1,56 @@
+use std::num::NonZeroU32;
+
+use serde_json::to_string;
+use tokio_tungstenite::tungstenite::Message;
+
+use paddler_messaging::inference_client::message::Message as InferenceClientMessage;
+use paddler_messaging::inference_server::message::Message as InferenceServerMessage;
+use paddler_messaging::inference_server::notification::Notification as InferenceServerNotification;
+use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters_schema::validated_parameters_schema::ValidatedParametersSchema;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_test_cluster_harness::raw_inference_socket::RawInferenceSocket;
+use paddler_tests::short_buffered_request_timeout::SHORT_BUFFERED_REQUEST_TIMEOUT;
+use paddler_tests::start_cluster::start_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn inference_socket_keeps_serving_after_a_stop_request_for_an_unknown_request() {
+    let cluster = start_cluster(ClusterParams {
+        buffered_request_timeout: SHORT_BUFFERED_REQUEST_TIMEOUT,
+        max_buffered_requests: 1,
+        ..ClusterParams::without_agents()
+    })
+    .await
+    .expect("a cluster without agents must start");
+    let mut socket = RawInferenceSocket::connect(cluster.balancer.addresses.inference)
+        .await
+        .expect("the raw inference socket must connect");
+
+    socket
+        .send(Message::text(
+            to_string(
+                &InferenceServerMessage::<ValidatedParametersSchema>::Notification(
+                    InferenceServerNotification::StopRespondingTo("unknown-request".to_owned()),
+                ),
+            )
+            .expect("the stop request must serialize"),
+        ))
+        .await
+        .expect("the stop request must be sent");
+    socket
+        .send_raw_prompt_request("request-after-the-stop-request", NonZeroU32::MIN)
+        .await
+        .expect("the follow-up request must be sent");
+
+    assert!(matches!(
+        socket
+            .next_answer_to("request-after-the-stop-request")
+            .await
+            .expect("the follow-up request must be answered"),
+        InferenceClientMessage::Error(envelope) if envelope.error.code == 504
+    ));
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/inference_socket_keeps_serving_after_malformed_json.rs b/paddler_tests/tests/inference_socket_keeps_serving_after_malformed_json.rs
index 74cf847d1..5b5e8b8f9 100644
--- a/paddler_tests/tests/inference_socket_keeps_serving_after_malformed_json.rs
+++ b/paddler_tests/tests/inference_socket_keeps_serving_after_malformed_json.rs
@@ -1,21 +1,19 @@
 use std::num::NonZeroU32;
-use std::time::Duration;
 
 use tokio_tungstenite::tungstenite::Message;
 
 use paddler_messaging::inference_client::message::Message as InferenceClientMessage;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
 use paddler_test_cluster_harness::raw_inference_socket::RawInferenceSocket;
+use paddler_tests::short_buffered_request_timeout::SHORT_BUFFERED_REQUEST_TIMEOUT;
 use paddler_tests::start_cluster::start_cluster;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn inference_socket_keeps_serving_after_malformed_json() {
     let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        buffered_request_timeout: Duration::from_millis(50),
+        buffered_request_timeout: SHORT_BUFFERED_REQUEST_TIMEOUT,
         max_buffered_requests: 1,
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
+        ..ClusterParams::without_agents()
     })
     .await
     .expect("a cluster without agents must start");
diff --git a/paddler_tests/tests/inference_socket_notifies_clients_only_when_token_generation_mode_changes.rs b/paddler_tests/tests/inference_socket_notifies_clients_only_when_token_generation_mode_changes.rs
deleted file mode 100644
index d6728705a..000000000
--- a/paddler_tests/tests/inference_socket_notifies_clients_only_when_token_generation_mode_changes.rs
+++ /dev/null
@@ -1,99 +0,0 @@
-use std::num::NonZeroU32;
-
-use futures_util::StreamExt as _;
-use tokio_util::sync::CancellationToken;
-
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
-use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-use paddler_messaging::inference_client::notification::Notification;
-use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
-use paddler_test_cluster_harness::cluster_params::ClusterParams;
-use paddler_tests::start_cluster::start_cluster;
-
-fn embeddings_state(temperature: f32) -> BalancerDesiredState {
-    BalancerDesiredState {
-        inference_parameters: InferenceParameters {
-            enable_embeddings: true,
-            temperature,
-            ..InferenceParameters::default()
-        },
-        ..BalancerDesiredState::default()
-    }
-}
-
-#[tokio::test(flavor = "multi_thread")]
-async fn inference_socket_notifies_clients_only_when_token_generation_mode_changes() {
-    let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        desired_state: Some(embeddings_state(0.5)),
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("a balancer serving embeddings must start");
-    let mut token_generation_mode_rx = cluster
-        .client_inference
-        .subscribe_to_token_generation_mode();
-
-    cluster
-        .client_inference
-        .continue_from_raw_prompt(
-            CancellationToken::new(),
-            ContinueFromRawPromptParams {
-                grammar: None,
-                max_tokens: NonZeroU32::MIN,
-                raw_prompt: "open the inference socket".to_owned(),
-            },
-        )
-        .await
-        .expect("the inference socket must accept a request in embeddings mode")
-        .next()
-        .await
-        .expect("the inference socket must answer in embeddings mode")
-        .expect("the embeddings-mode answer must be readable");
-
-    let connect_notification = token_generation_mode_rx
-        .recv()
-        .await
-        .expect("the client must be told on connect that token generation is disabled");
-
-    cluster
-        .client_management
-        .put_balancer_desired_state(CancellationToken::new(), &embeddings_state(0.7))
-        .await
-        .expect("the balancer must apply a state that keeps embeddings enabled");
-
-    let mut notification_after = async |desired_state: BalancerDesiredState| {
-        cluster
-            .client_management
-            .put_balancer_desired_state(CancellationToken::new(), &desired_state)
-            .await
-            .expect("the balancer must apply the desired state");
-
-        token_generation_mode_rx
-            .recv()
-            .await
-            .expect("the client must receive the token generation mode change")
-    };
-
-    let enabling_notification = notification_after(BalancerDesiredState::default()).await;
-    let disabling_notification = notification_after(embeddings_state(0.7)).await;
-
-    assert_eq!(
-        [
-            connect_notification,
-            enabling_notification,
-            disabling_notification
-        ],
-        [
-            Notification::TokenGenerationDisabled,
-            Notification::TokenGenerationEnabled,
-            Notification::TokenGenerationDisabled,
-        ]
-    );
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_tests/tests/inference_socket_partial_cancellation_serves_only_the_freed_slots.rs b/paddler_tests/tests/inference_socket_partial_cancellation_serves_only_the_freed_slots.rs
index 25134329c..9f3d3035b 100644
--- a/paddler_tests/tests/inference_socket_partial_cancellation_serves_only_the_freed_slots.rs
+++ b/paddler_tests/tests/inference_socket_partial_cancellation_serves_only_the_freed_slots.rs
@@ -5,12 +5,11 @@ use std::num::NonZeroU32;
 use futures_util::StreamExt as _;
 use tokio_util::sync::CancellationToken;
 
-use paddler_messaging::generated_token_result::GeneratedTokenResult;
 use paddler_messaging::request_params::continue_from_raw_prompt_params::ContinueFromRawPromptParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
 use paddler_test_cluster_harness::collect_generated_tokens::collect_generated_tokens;
-use paddler_test_cluster_harness::token_result_with_producer::TokenResultWithProducer;
 use paddler_test_cluster_harness::unending_generation::unending_generation;
 use paddler_tests::qwen3_desired_state::qwen3_desired_state;
 use paddler_tests::start_cluster::start_cluster;
@@ -32,7 +31,7 @@ fn waiting_prompt() -> ContinueFromRawPromptParams {
 async fn inference_socket_partial_cancellation_serves_only_the_freed_slots() {
     let mut cluster = start_cluster(ClusterParams {
         agents: vec![AgentConfig::single(SLOT_COUNT)],
-        desired_state: Some(qwen3_desired_state()),
+        desired_state: ClusterDesiredState::Apply(Box::new(qwen3_desired_state())),
         wait_for_slots_ready: true,
         ..ClusterParams::default()
     })
@@ -108,17 +107,9 @@ async fn inference_socket_partial_cancellation_serves_only_the_freed_slots() {
             .await
             .expect("the generated tokens must be collected");
 
-        assert!(
-            matches!(
-                collected.token_results.last(),
-                Some(TokenResultWithProducer {
-                    token_result: GeneratedTokenResult::Done(_),
-                    ..
-                })
-            ),
-            "every buffered request must run to completion through the freed slots, not end with an error: {:?}",
-            collected.token_results.last()
-        );
+        collected
+            .summary()
+            .expect("every buffered request must run to completion through the freed slots, not end with an error");
         assert!(
             collected
                 .token_results
diff --git a/paddler_tests/tests/inference_socket_pool_rejects_a_second_request_under_an_in_flight_id.rs b/paddler_tests/tests/inference_socket_pool_rejects_a_second_request_under_an_in_flight_id.rs
index 221f9d97f..c830f216d 100644
--- a/paddler_tests/tests/inference_socket_pool_rejects_a_second_request_under_an_in_flight_id.rs
+++ b/paddler_tests/tests/inference_socket_pool_rejects_a_second_request_under_an_in_flight_id.rs
@@ -28,18 +28,11 @@ fn raw_prompt_message() -> InferenceServerMessage<ValidatedParametersSchema> {
 
 #[tokio::test(flavor = "multi_thread")]
 async fn inference_socket_pool_rejects_a_second_request_under_an_in_flight_id() {
-    let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("a cluster without agents must start");
+    let cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("a cluster without agents must start");
     let pool = Pool::new(
-        cluster
-            .balancer
-            .inference_base_url()
-            .expect("the inference service must have a base URL"),
+        cluster.balancer.inference_base_url.clone(),
         NonZeroUsize::MIN,
     );
 
diff --git a/paddler_tests/tests/inference_socket_rejects_a_request_without_a_websocket_upgrade.rs b/paddler_tests/tests/inference_socket_rejects_a_request_without_a_websocket_upgrade.rs
index 3e78bd83b..feaff8669 100644
--- a/paddler_tests/tests/inference_socket_rejects_a_request_without_a_websocket_upgrade.rs
+++ b/paddler_tests/tests/inference_socket_rejects_a_request_without_a_websocket_upgrade.rs
@@ -7,13 +7,9 @@ use paddler_tests::start_cluster::start_cluster;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn inference_socket_rejects_a_request_without_a_websocket_upgrade() {
-    let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("a cluster without agents must start");
+    let cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("a cluster without agents must start");
 
     let response = get(format!(
         "http://{}{}",
diff --git a/paddler_tests/tests/inference_socket_rejects_tool_whose_required_field_is_not_a_property.rs b/paddler_tests/tests/inference_socket_rejects_tool_whose_required_field_is_not_a_property.rs
index c1e1584c6..4e243dc0a 100644
--- a/paddler_tests/tests/inference_socket_rejects_tool_whose_required_field_is_not_a_property.rs
+++ b/paddler_tests/tests/inference_socket_rejects_tool_whose_required_field_is_not_a_property.rs
@@ -8,13 +8,9 @@ use paddler_tests::start_cluster::start_cluster;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn inference_socket_rejects_tool_whose_required_field_is_not_a_property() {
-    let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("a cluster without agents must start");
+    let cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("a cluster without agents must start");
 
     let answer = cluster
         .client_inference
diff --git a/paddler_tests/tests/kev_27b_answers_every_decision_question_in_order_with_the_expected_input_tokens.rs b/paddler_tests/tests/kev_27b_answers_every_decision_question_in_order_with_the_expected_input_tokens.rs
new file mode 100644
index 000000000..e879d2ca2
--- /dev/null
+++ b/paddler_tests/tests/kev_27b_answers_every_decision_question_in_order_with_the_expected_input_tokens.rs
@@ -0,0 +1,17 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use paddler_tests::kev_checkpoint::KevCheckpoint;
+
+use crate::sample_decision_answers_check::SampleDecisionAnswersCheck;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn kev_27b_answers_every_decision_question_in_order_with_the_expected_input_tokens() {
+    SampleDecisionAnswersCheck {
+        cluster_params: KevCheckpoint::Kev27b
+            .decision_cluster_params()
+            .into_cluster_params()
+            .expect("the converted kev-27b artifacts must have absolute paths"),
+    }
+    .assert_for_the_sample_decision()
+    .await;
+}
diff --git a/paddler_tests/tests/kev_decision_model.rs b/paddler_tests/tests/kev_decision_model.rs
new file mode 100644
index 000000000..169a5db5f
--- /dev/null
+++ b/paddler_tests/tests/kev_decision_model.rs
@@ -0,0 +1,91 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use llama_cpp_bindings::llama_batch::LlamaBatch;
+use llama_cpp_bindings::token::LlamaToken;
+
+use paddler_agent_decision::decision_delimiter_tokens::DecisionDelimiterTokens;
+use paddler_agent_decision::decision_scheduler::DecisionScheduler;
+use paddler_agent_decision::decision_scheduler_context::DecisionSchedulerContext;
+use paddler_agent_decision::decision_token_layout::DecisionTokenLayout;
+use paddler_agent_pointer_head::pointer_head::PointerHead;
+
+use paddler_tests::agent_inference_runtime_context::agent_inference_runtime_context;
+use paddler_tests::decision_scheduler_run::DecisionSchedulerRun;
+use paddler_tests::kev_0_8b_pointer_head::kev_0_8b_pointer_head;
+use paddler_tests::kev_checkpoint::KevCheckpoint;
+use paddler_tests::llama_runtime::LlamaRuntime;
+use paddler_tests::two_option_question::two_option_question;
+
+pub struct KevDecisionModel {
+    pub delimiter_tokens: DecisionDelimiterTokens,
+    pub pointer_head: PointerHead,
+    pub runtime: LlamaRuntime,
+}
+
+impl KevDecisionModel {
+    pub fn load(slots: u16) -> Self {
+        let runtime = LlamaRuntime::load(
+            &KevCheckpoint::Kev0_8b.target_path("model.gguf"),
+            agent_inference_runtime_context(slots),
+        )
+        .expect("the kev model must load");
+        let pointer_head = kev_0_8b_pointer_head().expect("the kev pointer head must load");
+        let delimiter_tokens = DecisionDelimiterTokens::resolve(
+            &runtime.loaded_llama_model.model.llama_model,
+            &pointer_head.delimiters,
+        )
+        .expect("the kev pointer head delimiters must resolve");
+
+        Self {
+            delimiter_tokens,
+            pointer_head,
+            runtime,
+        }
+    }
+
+    #[must_use]
+    pub fn two_option_question_layout(&self, question_id: &str) -> DecisionTokenLayout {
+        DecisionTokenLayout::new(
+            &self.delimiter_tokens,
+            vec![LlamaToken::new(0)],
+            Vec::new(),
+            two_option_question(question_id),
+        )
+    }
+
+    pub fn run_scheduler(
+        self,
+        DecisionSchedulerRun {
+            active_requests,
+            agent_shutdown,
+            capacity_ledger,
+            pending_requests,
+            scheduler_inbox,
+            sequences,
+        }: DecisionSchedulerRun,
+    ) {
+        let Self {
+            pointer_head,
+            runtime,
+            ..
+        } = self;
+
+        DecisionScheduler {
+            active_requests,
+            agent_shutdown,
+            batch: LlamaBatch::new(runtime.model_runtime_parameters.n_batch.capacity(), 1),
+            capacity_ledger,
+            llama_context: runtime
+                .create_decision_llama_context(u32::from(sequences))
+                .expect("the decision context must be created"),
+            pending_requests,
+            scheduler_context: DecisionSchedulerContext {
+                agent_name: None,
+                n_batch: runtime.model_runtime_parameters.n_batch.tokens_usize(),
+                pointer_head,
+            },
+            scheduler_inbox,
+        }
+        .run();
+    }
+}
diff --git a/paddler_tests/tests/kev_decisions_match_the_fp32_reference.rs b/paddler_tests/tests/kev_decisions_match_the_fp32_reference.rs
new file mode 100644
index 000000000..97651836e
--- /dev/null
+++ b/paddler_tests/tests/kev_decisions_match_the_fp32_reference.rs
@@ -0,0 +1,105 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::mem::discriminant;
+
+use serde_json::from_str;
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::request_params::decide_params::raw_decide_params::RawDecideParams;
+use paddler_tests::decision_done_discriminant::decision_done_discriminant;
+use paddler_tests::kev_parity_record::KevParityRecord;
+use paddler_tests::start_cluster::start_cluster;
+use paddler_tests::test_cluster_preset::TestClusterPreset;
+
+const KEV_DOCUMENTED_GPU_BF16_DEVIATION: f32 = 0.03;
+
+fn most_probable_option(probabilities: &[f32]) -> Option<usize> {
+    probabilities
+        .iter()
+        .enumerate()
+        .max_by(|(_, left), (_, right)| left.total_cmp(right))
+        .map(|(option_index, _)| option_index)
+}
+
+#[tokio::test(flavor = "multi_thread")]
+async fn kev_decisions_match_the_fp32_reference() {
+    let records: Vec<KevParityRecord> = from_str(include_str!(
+        "../../fixtures/kev_0_8b_parity_reference.json"
+    ))
+    .expect("the parity reference must parse");
+    let cluster = start_cluster(
+        TestClusterPreset::Kev0_8b
+            .cluster_params()
+            .expect("the test cluster preset must name absolute paths"),
+    )
+    .await
+    .expect("the Kev cluster must start");
+
+    for KevParityRecord {
+        probabilities,
+        questions,
+        state,
+    } in records
+    {
+        let collected = cluster
+            .decide(
+                CancellationToken::new(),
+                &RawDecideParams {
+                    questions: questions.clone(),
+                    state,
+                },
+            )
+            .await
+            .expect("the decision must be answered");
+
+        assert_eq!(
+            discriminant(&collected.terminal_result),
+            decision_done_discriminant()
+        );
+        assert_eq!(
+            collected
+                .answers
+                .iter()
+                .map(|answer| answer.id.as_str())
+                .collect::<Vec<_>>(),
+            questions
+                .iter()
+                .map(|question| question.id.as_str())
+                .collect::<Vec<_>>()
+        );
+        assert_eq!(
+            collected
+                .answers
+                .iter()
+                .map(|answer| most_probable_option(&answer.probabilities))
+                .collect::<Vec<_>>(),
+            probabilities
+                .iter()
+                .map(|reference_probabilities| most_probable_option(reference_probabilities))
+                .collect::<Vec<_>>()
+        );
+
+        let largest_deviation = collected
+            .answers
+            .iter()
+            .zip(&probabilities)
+            .flat_map(|(answer, reference_probabilities)| {
+                answer
+                    .probabilities
+                    .iter()
+                    .zip(reference_probabilities)
+                    .map(|(served, reference)| (served - reference).abs())
+            })
+            .fold(0.0_f32, f32::max);
+
+        assert!(
+            largest_deviation <= KEV_DOCUMENTED_GPU_BF16_DEVIATION,
+            "{largest_deviation}"
+        );
+    }
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/kev_model_serving_without_decisions.rs b/paddler_tests/tests/kev_model_serving_without_decisions.rs
new file mode 100644
index 000000000..44ab532bd
--- /dev/null
+++ b/paddler_tests/tests/kev_model_serving_without_decisions.rs
@@ -0,0 +1,69 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::num::NonZeroU32;
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent::applicable_decisions::ApplicableDecisions;
+use paddler_agent::capability_starter::CapabilityStarter;
+use paddler_agent::inference_capabilities::InferenceCapabilities;
+use paddler_agent::loaded_model_serving_plan::LoadedModelServingPlan;
+use paddler_agent::model_serving_settings::ModelServingSettings;
+use paddler_agent::prepared_decisions::PreparedDecisions;
+use paddler_agent::served_model::ServedModel;
+use paddler_agent_text_generation::multimodal_projection::MultimodalProjection;
+use paddler_agent_text_generation::text_generation_settings::TextGenerationSettings;
+use paddler_inference_parameters::pooling_type::PoolingType;
+use paddler_inference_parameters::sampling_parameters::SamplingParameters;
+use paddler_messaging::chat_template_source::ChatTemplateSource;
+use paddler_tests::agent_inference_runtime_context::agent_inference_runtime_context;
+use paddler_tests::kev_checkpoint::KevCheckpoint;
+use paddler_tests::llama_runtime::LlamaRuntime;
+
+pub struct KevModelServingWithoutDecisions {
+    pub llama_runtime: LlamaRuntime,
+    pub model_serving_settings: ModelServingSettings,
+    pub served_model: ServedModel,
+}
+
+impl KevModelServingWithoutDecisions {
+    pub async fn start(agent_shutdown: &CancellationToken) -> Self {
+        let model_path = KevCheckpoint::Kev0_8b.target_path("model.gguf");
+        let llama_runtime = LlamaRuntime::load(&model_path, agent_inference_runtime_context(1))
+            .expect("the model must load");
+        let model_serving_settings = ModelServingSettings {
+            capabilities: InferenceCapabilities {
+                decisions: ApplicableDecisions::Disabled,
+                embeddings: PoolingType::Last,
+                text_generation: TextGenerationSettings {
+                    chat_template_source: ChatTemplateSource::EmbeddedInModel,
+                    image_resize_to_fit: NonZeroU32::MIN,
+                    multimodal_projection: MultimodalProjection::NotConfigured,
+                    sampling_parameters: SamplingParameters::default(),
+                },
+            },
+            model_path,
+            model_runtime_parameters: llama_runtime.model_runtime_parameters.clone(),
+        };
+        let served_model = CapabilityStarter {
+            cancellation_token: agent_shutdown,
+            inference_runtime_context: &llama_runtime.inference_runtime_context,
+        }
+        .serve(
+            LoadedModelServingPlan::starting_every_capability(
+                llama_runtime.loaded_llama_model.clone(),
+                model_serving_settings.clone(),
+            ),
+            PreparedDecisions::Disabled,
+        )
+        .await
+        .map_err(|serving_interruption| serving_interruption.interruption)
+        .expect("text generation and embeddings must start");
+
+        Self {
+            llama_runtime,
+            model_serving_settings,
+            served_model,
+        }
+    }
+}
diff --git a/paddler_tests/tests/kev_tokenizer_reference_comparison.rs b/paddler_tests/tests/kev_tokenizer_reference_comparison.rs
new file mode 100644
index 000000000..e924da26d
--- /dev/null
+++ b/paddler_tests/tests/kev_tokenizer_reference_comparison.rs
@@ -0,0 +1,57 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::fs::read_to_string;
+use std::path::PathBuf;
+use std::sync::Arc;
+
+use serde_json::from_str;
+
+use paddler_agent_decision::decision_text_tokenizer::DecisionTextTokenizer;
+use paddler_agent_pointer_head::pointer_head_input_normalization::PointerHeadInputNormalization;
+use paddler_test_cluster_harness::fixture_path::fixture_path;
+use paddler_tests::agent_inference_runtime_context::agent_inference_runtime_context;
+use paddler_tests::tokenizer_reference::TokenizerReference;
+use paddler_tests::tokenizer_reference_entry::TokenizerReferenceEntry;
+
+use crate::vocabulary_only_llama_model::VocabularyOnlyLlamaModel;
+
+pub struct KevTokenizerReferenceComparison {
+    pub model_path: PathBuf,
+    pub reference_fixture: &'static str,
+}
+
+impl KevTokenizerReferenceComparison {
+    pub fn assert_matched_by_the_decision_text_tokenizer(self) {
+        let TokenizerReference {
+            entries,
+            input_normalization,
+        } = from_str(
+            &read_to_string(fixture_path(self.reference_fixture))
+                .expect("the tokenizer reference must be readable"),
+        )
+        .expect("the tokenizer reference must parse");
+        let text_tokenizer = DecisionTextTokenizer {
+            input_normalization: PointerHeadInputNormalization::try_from(
+                input_normalization.as_str(),
+            )
+            .expect("the tokenizer reference must declare a known input normalization"),
+            model: Arc::new(VocabularyOnlyLlamaModel::load(
+                &agent_inference_runtime_context(1).llama_backend,
+                &self.model_path,
+            )),
+        };
+
+        for TokenizerReferenceEntry { text, token_ids } in entries {
+            assert_eq!(
+                text_tokenizer
+                    .tokenize(&text)
+                    .expect("the reference text must tokenize")
+                    .into_iter()
+                    .map(|token| token.0)
+                    .collect::<Vec<i32>>(),
+                token_ids,
+                "{text:?}"
+            );
+        }
+    }
+}
diff --git a/paddler_tests/tests/main.rs b/paddler_tests/tests/main.rs
index 92d301f57..9efbd5923 100644
--- a/paddler_tests/tests/main.rs
+++ b/paddler_tests/tests/main.rs
@@ -1,34 +1,130 @@
-mod agent_chunks_embedding_batch_larger_than_slot_count;
-mod agent_completes_generation_with_adequate_n_batch;
+mod a_balancer_starts_from_the_stored_desired_state_through_reconciliation;
+mod a_client_gone_before_an_unterminated_tool_call_resolves_receives_nothing;
+mod a_cluster_without_a_pointer_head_refuses_decisions_as_unavailable;
+mod a_decision_admitted_to_a_sequence_that_cannot_be_cleared_fails_with_kv_cache_removal;
+mod a_decision_arriving_while_another_is_decided_is_decided_too;
+mod a_decision_cluster_refuses_a_model_path_that_is_not_absolute;
+mod a_decision_cluster_refuses_a_pointer_head_path_that_is_not_absolute;
+mod a_decision_delimiter_containing_a_nul_byte_cannot_be_resolved;
+mod a_decision_evaluates_every_input_token_exactly_once;
+mod a_decision_leaving_the_waiting_line_lets_the_completion_behind_it_dispatch;
+mod a_decision_pending_admission_when_the_agent_shuts_down_is_refused_as_unavailable;
+mod a_decision_pipeline_whose_context_cannot_warm_up_every_slot_refuses_to_start;
+mod a_decision_running_when_the_agent_shuts_down_is_refused_as_unavailable;
+mod a_decision_state_llama_cpp_rejects_fails_the_decision;
+mod a_decision_state_overflowing_its_batch_fails_as_batch_assembly;
+mod a_decision_stopped_before_admission_is_acknowledged;
+mod a_decision_stopped_mid_decision_is_acknowledged;
+mod a_decision_tokenizer_keeps_its_llama_backend_alive;
+mod a_decision_waiting_when_the_agent_shuts_down_is_refused_as_unavailable;
+mod a_decision_waits_until_an_agent_loads_the_pointer_head;
+mod a_decision_whose_client_stopped_listening_cannot_send_its_answer;
+mod a_decision_whose_question_lane_cannot_be_released_still_reports_its_outcome;
+mod a_decision_whose_state_sequence_cannot_be_released_still_reports_its_outcome;
+mod a_decision_with_a_leading_question_reserving_one_slot_answers_like_one_reserving_two;
+mod a_decision_with_leading_questions_waits_for_two_free_slots;
+mod a_desired_state_with_a_missing_model_is_left_pending_without_starting_a_pipeline;
+mod a_failed_attempt_retries_without_reloading_its_model;
+mod a_failed_startup_releases_its_model_when_the_agent_shuts_down;
+mod a_generation_admitted_to_a_sequence_that_cannot_be_cleared_fails_with_kv_cache_clear;
+mod a_generation_arriving_while_every_sequence_is_taken_is_refused;
+mod a_generation_evaluates_its_prompt_and_each_generated_token_exactly_once;
+mod a_generation_overflowing_its_batch_fails_as_batch_assembly;
+mod a_generation_step_overflowing_its_batch_fails_as_batch_assembly;
+mod a_gguf_patch_refuses_a_source_that_is_not_a_gguf;
+mod a_grammar_refusing_the_token_that_triggers_it_ends_the_generation_as_rejected_model_output;
+mod a_grammar_that_cannot_constrain_the_candidates_fails_the_sampling;
+mod a_last_decision_question_overflowing_its_batch_fails_as_batch_assembly;
+mod a_leading_decision_question_llama_cpp_rejects_fails_the_decision;
+mod a_llama_runtime_reports_a_model_file_it_cannot_load;
+mod a_llama_runtime_reports_a_parallelism_it_cannot_determine;
+mod a_model_held_by_a_failed_start_is_released_before_another_model_loads;
+mod a_model_keeps_its_vram_occupancy_until_its_last_clone_is_dropped;
+mod a_model_with_a_metadata_key_that_is_not_utf8_reports_unreadable_metadata;
+mod a_model_with_a_metadata_value_that_is_not_utf8_reports_unreadable_metadata;
+mod a_model_without_a_chat_template_generates_from_raw_prompts;
+mod a_model_without_a_chat_template_refuses_conversations;
+mod a_multimodal_agent_serves_a_template_that_requires_a_system_message;
+mod a_multimodal_agent_whose_chat_template_drops_the_image_rejects_the_prompt;
+mod a_multimodal_prompt_carries_one_bos_whether_or_not_its_template_renders_it;
+mod a_multimodal_prompt_that_cannot_be_ingested_ends_the_generation;
+mod a_new_desired_state_cancels_the_download_of_the_one_it_replaces;
+mod a_prompt_deferred_by_a_full_batch_is_ingested_in_a_later_pass;
+mod a_prompt_that_tokenizes_to_nothing_refuses_the_generation;
+mod a_put_desired_state_is_answered_once_its_own_version_is_applied;
+mod a_question_lane_beyond_the_decision_context_cannot_be_cleared_for_the_state;
+mod a_reconfiguration_replaced_before_it_starts_yields_to_its_replacement;
+mod a_replacement_superseded_by_its_own_state_keeps_reporting_the_model_metadata;
+mod a_restart_cancelled_while_starting_a_capability_shuts_down_the_capabilities_it_kept;
+mod a_sampled_token_the_vocabulary_cannot_render_ends_the_generation;
+mod a_sampler_chain_that_cannot_rank_the_candidates_fails_the_sampling;
+mod a_sampler_chain_that_refuses_its_own_selection_fails_the_sampling;
+mod a_sampler_chain_that_selects_no_candidate_ends_the_generation;
+mod a_second_context_for_a_capability_cannot_be_created_while_one_exists;
+mod a_second_model_cannot_load_while_a_model_is_resident;
+mod a_second_multimodal_projection_cannot_load_while_one_is_resident;
+mod a_shutdown_while_a_model_loads_cancels_the_load;
+mod a_single_question_decision_exceeding_its_slots_share_of_the_context_is_rejected;
+mod a_state_changing_only_its_sampling_keeps_its_decisions_without_reloading_the_pointer_head;
+mod a_superseded_startup_releases_everything_it_held_when_the_agent_shuts_down;
+mod a_text_generation_pipeline_refuses_a_bos_token_it_cannot_render;
+mod a_text_generation_pipeline_refuses_a_model_whose_embedded_template_hides_its_streaming_markers;
+mod a_text_generation_pipeline_refuses_a_newline_token_it_cannot_render;
+mod a_text_generation_pipeline_refuses_a_penalty_window_longer_than_its_sequence_context;
+mod a_text_generation_pipeline_refuses_an_eos_token_it_cannot_render;
+mod a_text_generation_pipeline_refuses_to_start_while_its_context_is_resident;
+mod a_token_the_sampler_cannot_read_ends_the_generation;
+mod a_tool_call_body_with_a_nul_byte_fails_to_parse;
+mod a_tool_call_section_with_text_but_no_tool_calls_is_reported_as_unrecognized;
+mod a_tool_call_the_chat_parser_cannot_match_is_reported_as_unrecognized;
+mod a_waiting_decision_is_not_overtaken_by_later_completions;
+mod a_warm_up_llama_cpp_rejects_leaves_the_context_decoding;
+mod a_warm_up_spanning_more_tokens_than_its_batch_holds_primes_every_slot;
+mod a_warm_up_whose_batch_cannot_hold_every_slot_is_refused;
+#[cfg(test)]
+mod agent_answer_after_a_notification;
+mod agent_cannot_parse_tool_calls_for_a_model_without_a_chat_template;
+mod agent_clears_a_pointer_head_incompatibility_once_the_replaced_file_fits_the_model;
+mod agent_clears_a_pointer_head_load_issue_once_the_repaired_file_loads;
+mod agent_clears_the_chat_template_issue_once_a_valid_template_replaces_it;
 mod agent_conversation_accepts_empty_tools_list;
 mod agent_conversation_history_respects_max_tokens;
 mod agent_conversation_with_function_tool_succeeds;
 mod agent_conversation_with_gbnf_grammar_constrains_output;
 mod agent_conversation_with_json_schema_grammar_returns_valid_json;
-mod agent_embedding_batch_distribution_independent_of_context_size;
+mod agent_drops_the_issues_of_a_replaced_desired_state;
 mod agent_embedding_batch_returns_one_embedding_per_input_document;
-mod agent_embedding_batch_with_all_oversized_documents_reports_error;
+mod agent_embedding_batch_with_all_oversized_documents_reports_each_and_embeds_nothing;
 mod agent_embeddings_share_dimension_across_inputs_of_varying_length;
+mod agent_embeds_a_document_longer_than_a_sequence_context_with_an_encoder;
 mod agent_embeds_a_document_longer_than_the_default_micro_batch;
+mod agent_embeds_every_document_of_a_batch_larger_than_its_slot_count;
+mod agent_embeds_every_document_with_a_small_batch_and_context;
 mod agent_emits_tool_call_parsed_event_before_trailing_content;
+mod agent_ends_a_management_connection_on_a_binary_frame;
+mod agent_ends_a_management_connection_on_text_that_is_not_json;
+mod agent_ends_a_management_connection_whose_balancer_reuses_an_in_flight_request_id;
 mod agent_finishes_a_generation_that_fills_its_sequence_context;
 mod agent_generates_tokens_from_raw_prompt;
+mod agent_generates_with_a_model_that_declares_no_bos_token;
 mod agent_grammar_with_thinking_returns_incompatible_error;
+#[cfg(test)]
+mod agent_hang_up_on_a_frame;
 mod agent_isolates_concurrent_embedding_requests_per_client;
+mod agent_keeps_a_management_connection_after_a_stop_for_an_unknown_request;
+mod agent_keeps_a_management_connection_whose_balancer_reports_another_version;
 mod agent_keeps_serving_in_flight_inference_when_the_desired_model_is_missing;
 mod agent_l2_normalized_embeddings_have_unit_norm;
-mod agent_openai_chat_completions_non_streaming_returns_text;
-mod agent_openai_chat_completions_streaming_returns_chunks;
+mod agent_labels_embeddings_as_unnormalized_when_requested;
+mod agent_numbers_tool_calls_across_separate_tool_call_sections;
 mod agent_raw_prompt_respects_max_tokens;
 mod agent_raw_prompt_with_gbnf_grammar_constrains_output;
 mod agent_raw_prompt_with_json_schema_grammar_returns_valid_json;
-mod agent_rejects_buffered_conversation_after_switching_to_embeddings;
-mod agent_rejects_buffered_embedding_batch_after_disabling_embeddings;
-mod agent_rejects_buffered_raw_prompt_after_switching_to_embeddings;
+mod agent_reads_media_marker_text_in_an_image_prompt_as_plain_text;
+mod agent_rejects_a_multimodal_conversation_containing_a_nul_byte;
 mod agent_rejects_conversation_exceeding_sequence_context;
 mod agent_rejects_embedding_document_containing_nul_byte;
 mod agent_rejects_gemma_3_image_exceeding_the_micro_batch;
-mod agent_rejects_image_prompt_containing_media_marker_text;
 mod agent_rejects_image_prompt_exceeding_sequence_context;
 mod agent_rejects_oversized_conversation_before_compiling_its_tools;
 mod agent_rejects_oversized_conversation_before_converting_its_grammar;
@@ -37,26 +133,46 @@ mod agent_rejects_raw_prompt_containing_nul_byte;
 mod agent_rejects_raw_prompt_exceeding_sequence_context;
 mod agent_rejects_raw_prompt_with_unconvertible_json_schema_grammar;
 mod agent_rejects_structurally_invalid_tool_schema;
+mod agent_reports_a_document_exceeding_its_batch_and_embeds_the_rest;
+mod agent_reports_a_document_exceeding_its_sequence_context_and_embeds_the_rest;
+mod agent_reports_a_document_that_tokenizes_to_nothing_and_embeds_the_rest;
+mod agent_reports_a_grammar_no_token_can_satisfy_as_rejected_model_output;
+mod agent_reports_a_malformed_hugging_face_model_uri;
+mod agent_reports_a_pointer_head_it_cannot_load;
+mod agent_reports_a_pointer_head_of_another_hidden_size;
+mod agent_reports_a_pointer_head_wider_than_the_hidden_states_the_model_outputs;
+mod agent_reports_an_agent_uri_that_is_not_canonical;
+mod agent_reports_an_unparseable_model_uri;
 mod agent_reports_chat_template_error_when_rendering_fails;
 mod agent_reports_completion_tokens_matching_streamed_token_count;
 mod agent_reports_decode_failure_for_a_prompt_exceeding_a_non_causal_micro_batch;
 mod agent_reports_grammar_syntax_error_for_malformed_gbnf;
 mod agent_reports_independent_usage_for_concurrent_requests;
 mod agent_reports_only_content_tokens_in_usage_for_plain_content;
-mod agent_reports_oversized_document_and_embeds_the_rest;
+mod agent_reports_pointer_head_delimiters_that_are_not_control_tokens;
+mod agent_reports_slot_cannot_start_for_a_batch_smaller_than_its_slots;
 mod agent_reports_slot_cannot_start_for_excessive_slots;
+mod agent_reports_too_few_slots_for_decisions;
 mod agent_reports_tool_call_validation_failure;
 mod agent_returns_identical_embeddings_for_identical_documents;
 mod agent_returns_image_decoding_error_for_invalid_base64;
 mod agent_returns_image_decoding_error_for_malformed_data_uri;
 mod agent_returns_image_decoding_error_for_remote_url;
 mod agent_returns_rms_normalized_embeddings_when_requested;
-mod agent_returns_unnormalized_embeddings_when_requested;
+#[cfg(test)]
+mod agent_runner_connected_to;
+mod agent_serves_decisions_with_a_batch_too_small_to_warm_up_every_slot;
 mod agent_serves_embeddings_without_a_chat_template;
 mod agent_serves_four_concurrent_clients_streaming_tokens;
 mod agent_serves_request_being_prepared_when_state_change_begins;
 mod agent_serves_request_sent_while_state_change_waits_for_in_flight_request;
+mod agent_shuts_down_cleanly_while_downloading_its_model;
+mod agent_shuts_down_cleanly_while_loading_its_model;
+mod agent_shuts_down_cleanly_while_starting_to_serve_decisions;
+mod agent_shuts_down_cleanly_while_switching_to_another_model;
+mod agent_shuts_down_cleanly_with_an_embedding_batch_in_flight;
 mod agent_shuts_down_while_state_change_waits_for_in_flight_request;
+mod agent_stops_an_embedding_batch_when_the_client_disconnects;
 mod agent_stops_at_max_tokens_while_the_bare_json_probe_holds_tokens;
 mod agent_streams_json_grammar_output_as_content_when_no_tools_are_declared;
 mod agent_streams_tokens_for_a_causal_image_larger_than_n_batch;
@@ -67,41 +183,55 @@ mod agent_streams_tokens_held_by_the_bare_json_probe_at_end_of_generation;
 mod agent_text_only_model_rejects_image_input;
 mod agent_tokenizes_a_raw_prompt_opening_with_the_bos_token_with_a_single_bos_token;
 mod agent_tokenizes_gemma_3_chat_prompt_with_a_single_bos_token;
-mod balancer_answers_bad_gateway_when_an_agent_disconnects_before_answering;
-mod balancer_answers_gateway_timeout_when_an_agent_never_answers;
+mod agent_without_a_model_refuses_every_request_as_model_not_loaded;
+mod agents_stop_serving_decisions_once_the_pointer_head_is_removed;
+mod an_agent_cannot_answer_a_request_dispatched_to_another_agent;
+mod an_agent_releases_its_old_model_before_loading_the_new_one;
+mod an_agent_serving_decisions_also_serves_completions_and_embeddings;
+mod an_agent_whose_balancer_stops_reading_reconnects;
+mod an_agent_whose_model_cannot_start_reports_that_it_is_stuck;
+mod an_agent_whose_new_model_cannot_load_serves_nothing;
+mod an_agent_whose_projection_cannot_see_images_rejects_an_image_prompt;
+mod an_agent_without_a_pointer_head_rejects_decisions;
+mod an_embedding_batch_evaluates_every_document_token_exactly_once;
+mod an_embedding_batch_runs_on_the_slots_its_agent_has_free;
+mod an_embedding_batch_stopped_before_its_first_batch_produces_no_embeddings;
+mod an_embedding_context_without_pooling_cannot_read_sequence_embeddings;
+mod an_embedding_input_llama_cpp_rejects_fails_the_embedding_batch;
+mod an_embedding_input_overflowing_its_batch_fails_as_batch_assembly;
+mod an_embedding_request_accepted_before_its_scheduler_stops_is_finished;
+mod an_embedding_request_running_when_the_agent_shuts_down_is_refused_as_unavailable;
+mod an_embedding_that_cannot_be_normalized_fails_the_embedding_batch;
+mod an_empty_tool_call_section_resolves_to_no_tool_calls;
+mod an_untrusted_agent_waiting_for_a_request_reports_the_balancer_closing_its_socket;
+mod an_untrusted_agent_waiting_for_a_request_skips_the_pong_answering_its_ping;
+#[cfg(test)]
+mod answered_decision_probabilities;
+mod balancer_closes_an_agent_socket_that_answers_before_it_registers;
 mod balancer_closes_an_agent_socket_that_sends_a_reserved_opcode_frame_as_a_protocol_violation;
 mod balancer_closes_an_agent_socket_that_sends_an_undeserializable_message;
-mod balancer_closes_an_agent_socket_with_a_valid_close_frame_after_an_unknown_token_result;
 mod balancer_completes_buffered_request_after_agent_joins;
 mod balancer_ends_an_in_flight_generation_with_a_shutdown_error;
 mod balancer_inference_health_returns_ok;
 mod balancer_inference_service_replies_with_configured_cors_origin;
-mod balancer_inference_socket_recovers_from_embeddings_mode;
-mod balancer_keeps_an_agent_connected_after_it_answers_an_unknown_request;
+mod balancer_keeps_an_agent_that_repeats_its_status_registered;
 mod balancer_management_service_replies_with_configured_cors_origin;
-mod balancer_memory_storage_persists_desired_state;
-mod balancer_openai_compat_health_returns_ok;
+mod balancer_memory_storage_reports_the_desired_state_it_started_with;
 mod balancer_persists_chat_template_override_across_restart;
 mod balancer_persists_desired_state_across_restart;
-mod balancer_persists_huggingface_mmproj_in_desired_state;
-mod balancer_persists_local_mmproj_path_in_desired_state;
 mod balancer_persists_model_switch_in_storage;
-mod balancer_persists_url_model_in_desired_state;
 mod balancer_refuses_a_stored_desired_state_with_an_ambiguous_parameter;
 mod balancer_rejects_a_second_registration_under_a_registered_agent_id;
 mod balancer_rejects_a_status_update_from_a_socket_that_did_not_register;
-mod balancer_rejects_embedding_batch_with_501_when_embeddings_are_disabled;
-mod balancer_rejects_http_token_generation_with_501_while_the_cluster_serves_embeddings;
 mod balancer_rejects_tool_call_parsing_without_tools;
 mod balancer_rejects_tool_whose_required_field_is_not_a_property;
+mod balancer_reports_a_hugging_face_lookup_issue_for_a_nonexistent_model;
 mod balancer_reports_bad_gateway_when_agent_disconnects_mid_generation;
 mod balancer_reports_chat_template_does_not_compile_for_invalid_jinja;
-mod balancer_reports_chat_template_does_not_compile_recovers_when_template_replaced;
 mod balancer_reports_download_server_denied_access;
 mod balancer_reports_download_server_errored;
 mod balancer_reports_download_server_is_unreachable;
 mod balancer_reports_download_url_is_malformed;
-mod balancer_reports_huggingface_model_does_not_exist;
 mod balancer_reports_model_cannot_be_loaded_for_corrupt_file;
 mod balancer_reports_model_cannot_be_loaded_for_invalid_gguf;
 mod balancer_reports_model_does_not_exist_at_url;
@@ -109,27 +239,36 @@ mod balancer_reports_model_file_does_not_exist;
 mod balancer_reports_multimodal_projection_cannot_be_loaded_for_invalid_file;
 mod balancer_reports_multimodal_projection_cannot_be_loaded_for_missing_file;
 mod balancer_reports_one_request_id_across_a_chunked_embedding_batch;
+mod balancer_reports_the_huggingface_mmproj_of_its_desired_state;
+mod balancer_reports_the_local_mmproj_path_of_its_desired_state;
+mod balancer_reports_the_url_model_of_its_desired_state;
 mod balancer_reports_unable_to_find_chat_template_for_embedding_model;
 mod balancer_returns_503_when_request_buffering_disabled;
 mod balancer_returns_504_when_inference_item_timeout_is_zero;
 mod balancer_returns_504_when_no_agents_registered;
+mod balancer_serves_a_conversation_over_websocket;
 mod balancer_serves_every_concurrent_request_on_a_single_slot_agent;
 mod balancer_serves_inference_over_websocket;
 mod balancer_serves_request_after_agent_with_capacity_registers;
 mod balancer_serves_requests_from_different_connections_that_share_a_request_id;
 mod balancer_shutdown_does_not_wait_for_an_open_sse_subscriber;
+mod changing_the_context_size_restarts_serving_on_the_loaded_model;
 mod chat_template_drains_in_flight_inference_before_swap;
 mod chat_template_override_applied_to_embedding_model;
-mod chat_template_override_replaces_model_builtin;
+mod chat_template_override_is_reported_and_serves_conversations;
 mod chat_template_swaps_between_inference_calls;
 mod client_inference_reconnects_after_the_balancer_closes_its_socket;
-mod client_reports_service_unavailable_for_an_embedding_batch_without_agents;
 mod client_reports_unexpected_status_for_model_metadata_of_an_unknown_agent;
 mod cluster_harness_reports_an_agent_that_reports_issues_while_starting;
 mod cluster_harness_reports_an_awaited_agent_that_leaves_the_pool;
 mod cluster_harness_reports_an_inference_socket_closed_before_the_answer;
+mod cluster_harness_reports_requests_cancelled_before_they_are_sent;
+mod cluster_harness_reports_streamed_requests_cancelled_before_they_are_sent;
 mod cluster_harness_reports_the_end_of_a_snapshots_stream;
+mod cluster_harness_waiting_for_its_first_agent_reports_that_none_registered;
+mod cluster_harness_waiting_for_its_first_agent_to_serve_reports_that_the_agent_left;
 mod cluster_shutdown_returns_fd_count_to_baseline;
+mod cluster_shuts_down_cleanly_after_a_client_abandons_a_generation;
 mod cluster_shuts_down_without_waiting_for_its_shutdown_deadlines;
 mod continuous_batch_concurrent_conversation_history_requests_complete;
 mod continuous_batch_finishes_a_sequence_that_fills_its_share_of_the_context;
@@ -139,30 +278,49 @@ mod continuous_batch_generates_tokens_with_quantized_distinct_k_and_v_cache_dtyp
 mod continuous_batch_long_and_short_prompts_complete_concurrently;
 mod continuous_batch_plain_and_multimodal_run_concurrently;
 mod continuous_batch_produces_distinct_outputs_for_concurrent_prompts;
-mod continuous_batch_rejects_embedding_during_active_generation;
 mod continuous_batch_rejects_second_request_when_only_slot_busy;
 mod continuous_batch_releases_slot_when_client_disconnects;
-mod continuous_batch_releases_slots_on_shutdown_with_active_request;
 mod continuous_batch_reuses_slot_after_request_completes;
 mod continuous_batch_serves_four_concurrent_requests;
 mod continuous_batch_stop_signal_terminates_generation_before_max_tokens;
 mod continuous_batch_stops_at_max_tokens_boundary;
 mod continuous_batch_stops_generation_when_stop_sender_dropped;
 mod continuous_batch_two_concurrent_multimodal_requests_produce_tokens;
+mod decision_agent_shuts_down_cleanly_with_decisions_in_flight;
+mod decision_answers_equal_separate_single_question_decisions;
+mod decision_client_disconnecting_mid_decision_releases_its_sequences;
+mod decision_exceeding_the_context_is_rejected;
+mod decision_in_flight_is_answered_in_full_when_the_desired_state_changes;
+mod decision_input_with_a_nul_byte_is_rejected;
+mod decision_rows_longer_than_the_batch_are_answered;
+mod decision_without_questions_is_rejected_with_bad_request;
+mod decisions_are_identical_alone_and_concurrently;
+mod decisions_on_an_attention_only_model_read_the_hidden_states_of_their_own_input;
+#[cfg(test)]
+mod decode_prompt_for_sampling;
 mod deepseek_r1_distill_llama_8b_internal_endpoint_emits_reasoning_tokens;
+#[cfg(test)]
+mod embedding_documents_fixture;
+mod embeddings_use_the_pooling_the_model_declares_by_default;
+mod enabling_decisions_reuses_the_loaded_model;
+mod enabling_decisions_with_too_few_slots_keeps_serving_the_loaded_model;
+mod finalizing_a_tool_call_with_nothing_buffered_resolves_to_no_tool_calls;
+#[cfg(test)]
+mod first_message_content_request_preparer;
 mod gemma4_internal_endpoint_emits_reasoning_tokens;
 mod gemma4_internal_endpoint_emits_reasoning_tokens_for_image_request;
 mod gemma4_internal_endpoint_emits_tool_call_parsed_event;
 mod generating_request_keeps_every_token_when_multimodal_request_is_admitted;
+#[cfg(test)]
+mod grammar_refusing_the_greedy_selection;
 mod half_closed_client_stops_conversation_history_generation_before_max_tokens;
+mod half_closed_decision_client_releases_its_sequences;
 mod half_closed_embedding_batch_client_releases_the_buffered_request;
 mod half_closed_http_client_releases_the_buffered_request;
-mod half_closed_openai_non_streaming_client_releases_the_buffered_request;
-mod half_closed_openai_responses_client_releases_the_buffered_request;
-mod half_closed_openai_streaming_client_releases_the_buffered_request;
+mod http_client_reports_a_body_that_is_not_json_as_undecodable;
+mod http_client_reports_a_server_sent_event_of_another_shape_as_undecodable;
 mod http_inference_cancellation_ends_the_client_stream;
 mod http_inference_cancellation_releases_the_agent_slot;
-mod inference_socket_answers_a_conversation_history_request_with_token_generation_disabled_while_serving_embeddings;
 mod inference_socket_answers_a_ping_with_a_pong;
 mod inference_socket_answers_an_undeserializable_request_with_an_error;
 mod inference_socket_burst_cancellation_serves_every_buffered_request;
@@ -172,15 +330,25 @@ mod inference_socket_cancelling_one_request_leaves_a_sibling_request_running;
 mod inference_socket_connection_rejects_requests_after_the_balancer_closes_it;
 mod inference_socket_duplicate_request_id_is_answered_with_an_error;
 mod inference_socket_keeps_serving_after_a_binary_frame;
+mod inference_socket_keeps_serving_after_a_stop_request_for_an_unknown_request;
 mod inference_socket_keeps_serving_after_malformed_json;
-mod inference_socket_notifies_clients_only_when_token_generation_mode_changes;
 mod inference_socket_partial_cancellation_serves_only_the_freed_slots;
 mod inference_socket_pool_rejects_a_second_request_under_an_in_flight_id;
 mod inference_socket_rejects_a_request_without_a_websocket_upgrade;
 mod inference_socket_rejects_tool_whose_required_field_is_not_a_property;
+mod kev_27b_answers_every_decision_question_in_order_with_the_expected_input_tokens;
+#[cfg(test)]
+mod kev_decision_model;
+mod kev_decisions_match_the_fp32_reference;
+#[cfg(test)]
+mod kev_model_serving_without_decisions;
+#[cfg(test)]
+mod kev_tokenizer_reference_comparison;
 mod management_agents_endpoint_returns_empty_snapshot_when_no_agents_registered;
 mod management_agents_stream_ends_when_cancelled;
+mod management_agents_stream_reports_a_slot_taken_and_released;
 mod management_agents_stream_yields_initial_snapshot;
+mod management_api_reports_an_agent_hanging_up_before_it_answers_as_a_bad_gateway;
 mod management_buffered_requests_endpoint_returns_empty_snapshot_when_no_requests_are_buffered;
 mod management_buffered_requests_stream_yields_initial_snapshot;
 mod management_health_endpoint_returns_ok;
@@ -189,20 +357,54 @@ mod management_reports_no_download_in_progress_after_load_complete;
 mod management_returns_balancer_desired_state_that_was_put;
 mod management_returns_model_metadata_for_loaded_agent;
 mod management_serves_the_applicable_state_of_the_desired_state;
-mod management_two_agents_stream_subscribers_receive_slot_usage_changes;
 mod mistral3_internal_endpoint_emits_reasoning_tokens;
 mod mistral3_internal_endpoint_emits_reasoning_tokens_for_image_request;
 mod mistral3_internal_endpoint_emits_tool_call_parsed_event;
+#[cfg(test)]
+mod nomic_document_embedding_with;
+#[cfg(test)]
+mod nomic_embedding_batch_processing;
+mod openai_chat_completion_accepts_an_assistant_tool_call_turn_without_content;
 mod openai_chat_completion_non_streaming_conforms_to_schema;
+mod openai_chat_completion_non_streaming_emits_tool_calls_for_function_tool;
+mod openai_chat_completion_non_streaming_half_closed_client_releases_the_buffered_request;
+mod openai_chat_completion_non_streaming_reports_a_conversation_exceeding_the_context_as_a_bad_request;
 mod openai_chat_completion_non_streaming_reports_length_finish_reason_at_max_tokens;
+mod openai_chat_completion_non_streaming_returns_text;
+mod openai_chat_completion_non_streaming_returns_usage;
+mod openai_chat_completion_non_streaming_usage_adds_up_with_tool_calls;
+mod openai_chat_completion_reports_a_buffered_request_timeout_as_a_gateway_timeout;
+mod openai_chat_completion_reports_a_full_buffer_as_service_unavailable;
+mod openai_chat_completion_reports_a_stalled_generation_as_a_gateway_timeout;
+mod openai_chat_completion_reports_a_tool_call_violating_its_schema_as_a_server_error;
 mod openai_chat_completion_streaming_conforms_to_schema;
+mod openai_chat_completion_streaming_emits_tool_calls_for_function_tool;
+mod openai_chat_completion_streaming_emits_usage_when_requested;
+mod openai_chat_completion_streaming_ends_at_a_tool_call_violating_its_schema;
+mod openai_chat_completion_streaming_half_closed_client_releases_the_buffered_request;
+mod openai_chat_completion_streaming_omits_usage_when_not_requested;
+mod openai_chat_completion_streaming_rejects_a_tool_that_requires_an_undeclared_property;
+mod openai_chat_completion_streaming_reports_a_buffered_request_timeout_as_an_error_chunk;
+mod openai_chat_completion_streaming_reports_a_conversation_exceeding_the_context_as_an_error_chunk;
 mod openai_chat_completion_streaming_reports_length_finish_reason_at_max_tokens;
-mod openai_responses_are_rejected_while_the_cluster_serves_embeddings;
+mod openai_chat_completion_streaming_returns_chunks;
+mod openai_chat_completion_streaming_usage_breakdown_with_thinking;
+mod openai_chat_completion_with_reasoning_effort_none_generates_no_reasoning_tokens;
+mod openai_health_returns_ok;
+mod openai_responses_accepts_its_own_output_replayed_as_input;
+mod openai_responses_half_closed_client_releases_the_buffered_request;
 mod openai_responses_non_streaming_conforms_to_schema;
 mod openai_responses_non_streaming_reports_incomplete_status_at_max_output_tokens;
+mod openai_responses_non_streaming_returns_text_and_usage;
 mod openai_responses_rejects_a_tool_that_requires_an_undeclared_property;
 mod openai_responses_streaming_conforms_to_schema;
+mod openai_responses_streaming_ends_at_a_tool_call_violating_its_schema;
+mod openai_responses_streaming_rejects_a_tool_that_requires_an_undeclared_property;
+mod openai_responses_streaming_reports_a_buffered_request_timeout_as_a_failed_response;
 mod openai_responses_streaming_reports_incomplete_status_at_max_output_tokens;
+mod openai_responses_text_format_json_schema_constrains_the_output;
+mod openai_service_lets_browsers_send_and_read_openai_headers;
+mod production_sources_keep_test_code_in_inline_test_modules;
 mod qwen25vl_generates_tokens_from_image_input;
 mod qwen35_generates_tokens_for_long_system_and_user_prompt;
 mod qwen35_generation_stops_at_eog_before_max_tokens;
@@ -213,22 +415,95 @@ mod qwen35_internal_endpoint_with_thinking_enabled_emits_reasoning_tokens;
 mod qwen35_thinking_mode_stops_cleanly_before_max_tokens;
 mod qwen35_thinking_multi_turn_conversation_stops_cleanly;
 mod qwen35_with_mmproj_generates_tokens_from_image;
-mod qwen35_with_system_message_completes_with_thinking;
-mod qwen35_with_system_message_completes_without_thinking;
+mod qwen35_with_system_message_completes_with_thinking_disabled;
+mod qwen35_with_system_message_completes_with_thinking_enabled;
 mod qwen35_without_mmproj_rejects_image_with_multimodal_not_supported;
+mod qwen3_5_answers_every_decision_question_in_order_with_the_expected_input_tokens;
+mod qwen3_5_decision_text_tokenizes_like_kev;
+mod qwen3_8_answers_every_decision_question_in_order_with_the_expected_input_tokens;
+mod qwen3_8_decision_text_tokenizes_like_kev;
+#[cfg(test)]
+mod qwen3_chat_completion_prompt;
+mod qwen3_chat_template_renders_a_tool_call_history_as_tool_call_and_tool_response_blocks;
+mod qwen3_chat_template_renders_the_reasoning_of_an_assistant_turn_in_a_tool_loop;
 mod qwen3_generates_tokens_from_conversation_history;
 mod qwen3_internal_endpoint_emits_tool_call_parsed_event;
 mod qwen3_internal_endpoint_emits_tool_call_tokens;
 mod qwen3_internal_endpoint_tools_without_parse_flag_emit_only_raw_tokens;
 mod qwen3_internal_endpoint_with_thinking_disabled_emits_no_reasoning_tokens;
 mod qwen3_internal_endpoint_with_thinking_enabled_emits_reasoning_tokens;
-mod qwen3_openai_chat_completion_with_reasoning_effort_none_generates_no_reasoning_tokens;
-mod qwen3_openai_non_streaming_emits_tool_calls_for_function_tool;
-mod qwen3_openai_non_streaming_returns_usage;
-mod qwen3_openai_non_streaming_usage_with_tool_calls;
-mod qwen3_openai_streaming_emits_tool_calls_for_function_tool;
-mod qwen3_openai_streaming_emits_usage_when_requested;
-mod qwen3_openai_streaming_omits_usage_when_not_requested;
-mod qwen3_openai_streaming_usage_breakdown_with_thinking;
-mod qwen3_responses_non_streaming_returns_text_and_usage;
+#[cfg(test)]
+mod qwen3_without_a_chat_template;
+mod raw_inference_socket_reports_a_socket_closed_before_a_pong;
+mod raw_inference_socket_reports_a_socket_closed_before_any_answer;
+mod raw_inference_socket_skips_a_pong_while_waiting_for_an_answer;
+mod raw_inference_socket_skips_an_answer_while_waiting_for_a_pong;
+mod reading_a_missing_state_file_changes_nothing;
+#[cfg(test)]
+mod ready_to_sample_request;
+mod reapplying_the_served_state_keeps_serving_it_without_a_restart;
+#[cfg(test)]
+mod release_half_closed_openai_request;
+#[cfg(test)]
+mod replaceable_pointer_head;
+mod retaining_a_served_model_restarts_its_changed_capabilities_at_once;
+mod run_with_test_cluster_exports_the_url_of_every_service;
+mod run_with_test_cluster_fails_when_its_agents_cannot_set_up_their_hugging_face_client;
+mod run_with_test_cluster_fails_when_its_agents_cannot_set_up_their_model_downloader;
+mod run_with_test_cluster_fails_when_its_program_fails;
+mod run_with_test_cluster_fails_without_a_temporary_directory_for_agent_models;
+#[cfg(test)]
+mod sample_decision_answers_check;
+#[cfg(test)]
+mod sampling_with_a_smaller_vocabulary_grammar;
+#[cfg(test)]
+mod serving_decision_agent;
+#[cfg(test)]
+mod serving_pipeline_arbiter;
+#[cfg(test)]
+mod shutdown_during_state_transition;
+mod shutting_down_a_served_model_closes_every_capability_at_once;
+mod shutting_down_a_served_model_stops_its_schedulers_and_frees_their_contexts;
+#[cfg(test)]
+mod slot_held_by_an_unending_generation;
+#[cfg(test)]
+mod small_context_decision_cluster_params;
+mod smolvlm2_generates_tokens_for_jpeg_and_webp_images_with_the_default_batch;
 mod smolvlm2_generates_tokens_from_image_input;
+mod start_cluster_reports_a_desired_state_the_balancer_cannot_store;
+mod start_cluster_reports_a_state_database_url_it_cannot_parse;
+#[cfg(test)]
+mod swappable_model_file;
+#[cfg(test)]
+mod test_only_attribute_counter;
+mod the_test_only_code_detector_counts_only_test_code_outside_inline_test_modules;
+#[cfg(test)]
+mod tool_call_pipeline_for_tools;
+mod typesafe_health_returns_ok;
+mod typesafe_models_is_unavailable_while_the_cluster_serves_no_decisions;
+mod typesafe_models_lists_nothing_while_no_agent_serves_the_desired_model;
+mod typesafe_models_lists_the_model_the_cluster_serves;
+mod typesafe_service_lets_browsers_send_and_read_typesafe_headers;
+mod typesafe_system_one_answers_a_body_that_is_not_json_as_unprocessable;
+mod typesafe_system_one_answers_an_untranslatable_request_as_unprocessable;
+mod typesafe_system_one_answers_questions_in_their_order;
+mod typesafe_system_one_answers_with_the_decision_model;
+mod typesafe_system_one_client_disconnecting_mid_decision_releases_its_sequences;
+mod typesafe_system_one_echoes_the_model_it_was_asked_for;
+mod typesafe_system_one_echoes_the_request_id_it_was_given;
+mod typesafe_system_one_is_unavailable_while_the_cluster_serves_no_decisions;
+mod typesafe_system_one_rejects_a_request_without_a_model;
+mod typesafe_system_one_reports_a_buffered_request_timeout_as_a_gateway_timeout;
+mod typesafe_system_one_reports_a_decision_exceeding_the_context_as_unprocessable;
+mod typesafe_system_one_reports_a_full_buffer_as_service_unavailable;
+mod typesafe_system_one_reports_a_state_with_a_nul_byte_as_unprocessable;
+mod typesafe_system_one_reports_answers_that_do_not_match_its_questions_as_a_bad_gateway;
+mod typesafe_system_one_reports_that_a_decision_produces_no_output_tokens;
+#[cfg(test)]
+mod typesafe_tone_and_payment_decision;
+#[cfg(test)]
+mod untrusted_management_agent;
+#[cfg(test)]
+mod vocabulary_only_llama_model;
+#[cfg(test)]
+mod with_nul_byte;
diff --git a/paddler_tests/tests/management_agents_endpoint_returns_empty_snapshot_when_no_agents_registered.rs b/paddler_tests/tests/management_agents_endpoint_returns_empty_snapshot_when_no_agents_registered.rs
index 3eb584e08..347298e3d 100644
--- a/paddler_tests/tests/management_agents_endpoint_returns_empty_snapshot_when_no_agents_registered.rs
+++ b/paddler_tests/tests/management_agents_endpoint_returns_empty_snapshot_when_no_agents_registered.rs
@@ -5,13 +5,9 @@ use paddler_tests::start_cluster::start_cluster;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn management_agents_endpoint_returns_empty_snapshot_when_no_agents_registered() {
-    let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("the cluster must start");
+    let cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("the cluster must start");
 
     let snapshot = cluster
         .client_management
diff --git a/paddler_tests/tests/management_agents_stream_ends_when_cancelled.rs b/paddler_tests/tests/management_agents_stream_ends_when_cancelled.rs
index c7f538a9c..4c7117b69 100644
--- a/paddler_tests/tests/management_agents_stream_ends_when_cancelled.rs
+++ b/paddler_tests/tests/management_agents_stream_ends_when_cancelled.rs
@@ -6,13 +6,9 @@ use paddler_tests::start_cluster::start_cluster;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn management_agents_stream_ends_when_cancelled() {
-    let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("the cluster must start");
+    let cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("the cluster must start");
 
     let cancellation_token = CancellationToken::new();
 
diff --git a/paddler_tests/tests/management_agents_stream_reports_a_slot_taken_and_released.rs b/paddler_tests/tests/management_agents_stream_reports_a_slot_taken_and_released.rs
new file mode 100644
index 000000000..2f9c05dcd
--- /dev/null
+++ b/paddler_tests/tests/management_agents_stream_reports_a_slot_taken_and_released.rs
@@ -0,0 +1,55 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_test_cluster_harness::collect_generated_tokens::collect_generated_tokens;
+use paddler_test_cluster_harness::unending_generation::unending_generation;
+use paddler_tests::qwen3_desired_state::qwen3_desired_state;
+use paddler_tests::start_cluster::start_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn management_agents_stream_reports_a_slot_taken_and_released() {
+    let mut cluster = start_cluster(ClusterParams {
+        agents: vec![AgentConfig::single(1)],
+        desired_state: ClusterDesiredState::Apply(Box::new(qwen3_desired_state())),
+        wait_for_slots_ready: true,
+        ..ClusterParams::default()
+    })
+    .await
+    .expect("the cluster must start");
+
+    let agent_id = cluster
+        .agent_ids
+        .first()
+        .expect("cluster must have registered one agent")
+        .clone();
+
+    let generation_cancellation = CancellationToken::new();
+    let token_stream = cluster
+        .continue_from_raw_prompt_stream(generation_cancellation.clone(), &unending_generation())
+        .await
+        .expect("the inference request must be accepted");
+
+    cluster
+        .wait_for_slots_processing(&agent_id, 1)
+        .await
+        .expect("agents_stream must emit a snapshot showing slot usage");
+
+    generation_cancellation.cancel();
+    collect_generated_tokens(token_stream)
+        .await
+        .expect("the generated tokens must be collected");
+
+    cluster
+        .wait_for_slots_processing(&agent_id, 0)
+        .await
+        .expect("agents_stream must emit a snapshot showing the slot was released");
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/management_api_reports_an_agent_hanging_up_before_it_answers_as_a_bad_gateway.rs b/paddler_tests/tests/management_api_reports_an_agent_hanging_up_before_it_answers_as_a_bad_gateway.rs
new file mode 100644
index 000000000..b5289fa83
--- /dev/null
+++ b/paddler_tests/tests/management_api_reports_an_agent_hanging_up_before_it_answers_as_a_bad_gateway.rs
@@ -0,0 +1,66 @@
+use reqwest::Client;
+use reqwest::StatusCode;
+use tokio::spawn;
+
+use paddler_messaging::agent_status::AgentStatus;
+use paddler_messaging::api_path::ApiPath;
+use paddler_messaging::slot_aggregated_status_snapshot::SlotAggregatedStatusSnapshot;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_tests::start_cluster::start_cluster;
+
+use crate::untrusted_management_agent::UntrustedManagementAgent;
+
+const AGENT_ID: &str = "hanging-up";
+
+#[tokio::test(flavor = "multi_thread")]
+async fn management_api_reports_an_agent_hanging_up_before_it_answers_as_a_bad_gateway() {
+    let mut cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("a cluster without agents must start");
+    let mut hanging_up_agent = UntrustedManagementAgent::connect_registered(
+        cluster.balancer.addresses.management,
+        AGENT_ID,
+        SlotAggregatedStatusSnapshot {
+            status: AgentStatus::default(),
+            version: 1,
+        },
+    )
+    .await
+    .expect("the agent must register");
+    cluster
+        .wait_for_agent_count(1)
+        .await
+        .expect("the balancer must register the agent");
+
+    let model_metadata_response = spawn(
+        Client::new()
+            .get(
+                cluster
+                    .balancer
+                    .management_base_url
+                    .join(&ApiPath::agent_model_metadata(AGENT_ID))
+                    .expect("the model metadata path must form a URL"),
+            )
+            .send(),
+    );
+
+    hanging_up_agent
+        .next_request()
+        .await
+        .expect("the balancer must ask the agent for its model metadata");
+    drop(hanging_up_agent);
+
+    assert_eq!(
+        model_metadata_response
+            .await
+            .expect("the model metadata request must finish")
+            .expect("the balancer must answer the model metadata request")
+            .status(),
+        StatusCode::BAD_GATEWAY
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/management_buffered_requests_endpoint_returns_empty_snapshot_when_no_requests_are_buffered.rs b/paddler_tests/tests/management_buffered_requests_endpoint_returns_empty_snapshot_when_no_requests_are_buffered.rs
index 77ec7b5e0..8b147dd72 100644
--- a/paddler_tests/tests/management_buffered_requests_endpoint_returns_empty_snapshot_when_no_requests_are_buffered.rs
+++ b/paddler_tests/tests/management_buffered_requests_endpoint_returns_empty_snapshot_when_no_requests_are_buffered.rs
@@ -6,13 +6,9 @@ use paddler_tests::start_cluster::start_cluster;
 #[tokio::test(flavor = "multi_thread")]
 async fn management_buffered_requests_endpoint_returns_empty_snapshot_when_no_requests_are_buffered()
  {
-    let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("the cluster must start");
+    let cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("the cluster must start");
 
     let snapshot = cluster
         .client_management
diff --git a/paddler_tests/tests/management_buffered_requests_stream_yields_initial_snapshot.rs b/paddler_tests/tests/management_buffered_requests_stream_yields_initial_snapshot.rs
index e367414a4..587567e59 100644
--- a/paddler_tests/tests/management_buffered_requests_stream_yields_initial_snapshot.rs
+++ b/paddler_tests/tests/management_buffered_requests_stream_yields_initial_snapshot.rs
@@ -6,13 +6,9 @@ use paddler_tests::start_cluster::start_cluster;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn management_buffered_requests_stream_yields_initial_snapshot() {
-    let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("the cluster must start");
+    let cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("the cluster must start");
 
     let mut stream = cluster
         .client_management
diff --git a/paddler_tests/tests/management_health_endpoint_returns_ok.rs b/paddler_tests/tests/management_health_endpoint_returns_ok.rs
index d83adef41..f7c23bca8 100644
--- a/paddler_tests/tests/management_health_endpoint_returns_ok.rs
+++ b/paddler_tests/tests/management_health_endpoint_returns_ok.rs
@@ -6,13 +6,9 @@ use paddler_tests::start_cluster::start_cluster;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn management_health_endpoint_returns_ok() {
-    let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("failed to start subprocess cluster");
+    let cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("failed to start subprocess cluster");
 
     let health = cluster
         .client_management
diff --git a/paddler_tests/tests/management_metrics_endpoint_exposes_prometheus_gauges.rs b/paddler_tests/tests/management_metrics_endpoint_exposes_prometheus_gauges.rs
index 9c60aa9cf..a1d6880a7 100644
--- a/paddler_tests/tests/management_metrics_endpoint_exposes_prometheus_gauges.rs
+++ b/paddler_tests/tests/management_metrics_endpoint_exposes_prometheus_gauges.rs
@@ -5,13 +5,9 @@ use paddler_tests::start_cluster::start_cluster;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn management_metrics_endpoint_exposes_prometheus_gauges() {
-    let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("the cluster must start");
+    let cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("the cluster must start");
 
     let metrics = cluster
         .client_management
diff --git a/paddler_tests/tests/management_returns_balancer_desired_state_that_was_put.rs b/paddler_tests/tests/management_returns_balancer_desired_state_that_was_put.rs
index 7d961b12f..29a074ee0 100644
--- a/paddler_tests/tests/management_returns_balancer_desired_state_that_was_put.rs
+++ b/paddler_tests/tests/management_returns_balancer_desired_state_that_was_put.rs
@@ -4,9 +4,9 @@ use anyhow::Result;
 use tokio_util::sync::CancellationToken;
 
 use paddler_client::client_management::ClientManagement;
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
-use paddler_messaging::agent_desired_model::AgentDesiredModel;
+use paddler_inference_parameters::sampling_parameters::SamplingParameters;
 use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_messaging::balancer_text_generation_settings::BalancerTextGenerationSettings;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
 use paddler_tests::start_cluster::start_cluster;
 
@@ -22,13 +22,9 @@ async fn get_balancer_desired_state(
 
 #[tokio::test(flavor = "multi_thread")]
 async fn management_returns_balancer_desired_state_that_was_put() {
-    let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("the cluster must start");
+    let cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("the cluster must start");
 
     assert_eq!(
         get_balancer_desired_state(&cluster.client_management)
@@ -38,11 +34,11 @@ async fn management_returns_balancer_desired_state_that_was_put() {
     );
 
     let desired_state = BalancerDesiredState {
-        chat_template_override: None,
-        inference_parameters: InferenceParameters::deterministic(),
-        model: AgentDesiredModel::None,
-        multimodal_projection: AgentDesiredModel::None,
-        use_chat_template_override: false,
+        text_generation: BalancerTextGenerationSettings {
+            sampling_parameters: SamplingParameters::deterministic(),
+            ..BalancerTextGenerationSettings::default()
+        },
+        ..BalancerDesiredState::default()
     };
 
     cluster
diff --git a/paddler_tests/tests/management_serves_the_applicable_state_of_the_desired_state.rs b/paddler_tests/tests/management_serves_the_applicable_state_of_the_desired_state.rs
index 4d4c49716..75686fcd5 100644
--- a/paddler_tests/tests/management_serves_the_applicable_state_of_the_desired_state.rs
+++ b/paddler_tests/tests/management_serves_the_applicable_state_of_the_desired_state.rs
@@ -1,18 +1,15 @@
 use tokio_util::sync::CancellationToken;
 
-use paddler_messaging::agent_desired_model::AgentDesiredModel;
+use paddler_messaging::agent_desired_state::AgentDesiredState;
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
 use paddler_test_cluster_harness::cluster_params::ClusterParams;
 use paddler_tests::start_cluster::start_cluster;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn management_serves_the_applicable_state_of_the_desired_state() {
-    let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("a cluster without agents must start");
+    let cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("a cluster without agents must start");
 
     let agent_desired_state = cluster
         .client_management
@@ -20,8 +17,10 @@ async fn management_serves_the_applicable_state_of_the_desired_state() {
         .await
         .expect("the balancer must serve its applicable state");
 
-    assert_eq!(agent_desired_state.model, AgentDesiredModel::None);
-    assert_eq!(agent_desired_state.chat_template_override, None);
+    assert_eq!(
+        agent_desired_state,
+        AgentDesiredState::from(&BalancerDesiredState::default())
+    );
 
     cluster
         .shutdown()
diff --git a/paddler_tests/tests/management_two_agents_stream_subscribers_receive_slot_usage_changes.rs b/paddler_tests/tests/management_two_agents_stream_subscribers_receive_slot_usage_changes.rs
deleted file mode 100644
index 7f21f3550..000000000
--- a/paddler_tests/tests/management_two_agents_stream_subscribers_receive_slot_usage_changes.rs
+++ /dev/null
@@ -1,54 +0,0 @@
-#![cfg(feature = "tests_that_use_llms")]
-
-use tokio_util::sync::CancellationToken;
-
-use paddler_test_cluster_harness::agent_config::AgentConfig;
-use paddler_test_cluster_harness::cluster_params::ClusterParams;
-use paddler_test_cluster_harness::collect_generated_tokens::collect_generated_tokens;
-use paddler_test_cluster_harness::unending_generation::unending_generation;
-use paddler_tests::qwen3_desired_state::qwen3_desired_state;
-use paddler_tests::start_cluster::start_cluster;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn management_two_agents_stream_subscribers_receive_slot_usage_changes() {
-    let mut cluster = start_cluster(ClusterParams {
-        agents: vec![AgentConfig::single(1)],
-        desired_state: Some(qwen3_desired_state()),
-        wait_for_slots_ready: true,
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("the cluster must start");
-
-    let agent_id = cluster
-        .agent_ids
-        .first()
-        .expect("cluster must have registered one agent")
-        .clone();
-
-    let generation_cancellation = CancellationToken::new();
-    let token_stream = cluster
-        .continue_from_raw_prompt_stream(generation_cancellation.clone(), &unending_generation())
-        .await
-        .expect("the inference request must be accepted");
-
-    cluster
-        .wait_for_slots_processing(&agent_id, 1)
-        .await
-        .expect("agents_stream must emit a snapshot showing slot usage");
-
-    generation_cancellation.cancel();
-    collect_generated_tokens(token_stream)
-        .await
-        .expect("the generated tokens must be collected");
-
-    cluster
-        .wait_for_slots_processing(&agent_id, 0)
-        .await
-        .expect("agents_stream must emit a snapshot showing the slot was released");
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_tests/tests/mistral3_internal_endpoint_emits_reasoning_tokens.rs b/paddler_tests/tests/mistral3_internal_endpoint_emits_reasoning_tokens.rs
index baba74346..b2a3454fa 100644
--- a/paddler_tests/tests/mistral3_internal_endpoint_emits_reasoning_tokens.rs
+++ b/paddler_tests/tests/mistral3_internal_endpoint_emits_reasoning_tokens.rs
@@ -23,12 +23,14 @@ async fn mistral3_internal_endpoint_emits_reasoning_tokens() {
             CancellationToken::new(),
             &ContinueFromConversationHistoryParams {
                 add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Text(
-                        "What is two plus two? Think step by step.".to_owned(),
-                    ),
-                    role: "user".to_owned(),
-                }]),
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Text(
+                            "What is two plus two? Think step by step.".to_owned(),
+                        ),
+                        role: "user".to_owned(),
+                    },
+                ]),
                 enable_thinking: true,
                 grammar: None,
                 max_tokens: NonZeroU32::new(200).unwrap(),
@@ -50,13 +52,9 @@ async fn mistral3_internal_endpoint_emits_reasoning_tokens() {
         "Mistral 3: expected at least one reasoning token from a [THINK]-emitting model (got {reasoning_count})"
     );
 
-    let last = collected
-        .token_results
-        .last()
-        .expect("no token results received");
-    let GeneratedTokenResult::Done(summary) = &last.token_result else {
-        panic!("last result was not Done: {last:?}");
-    };
+    let summary = collected
+        .summary()
+        .expect("the generation must finish with a summary");
 
     assert!(summary.usage.prompt_tokens > 0);
     assert!(summary.usage.reasoning_tokens > 0);
@@ -67,22 +65,8 @@ async fn mistral3_internal_endpoint_emits_reasoning_tokens() {
             + summary.usage.undeterminable_tokens
     );
 
-    let reasoning_stream: String = collected
-        .token_results
-        .iter()
-        .filter_map(|result| match &result.token_result {
-            GeneratedTokenResult::ReasoningToken(piece) => Some(piece.as_str()),
-            _ => None,
-        })
-        .collect();
-    let content_stream: String = collected
-        .token_results
-        .iter()
-        .filter_map(|result| match &result.token_result {
-            GeneratedTokenResult::ContentToken(piece) => Some(piece.as_str()),
-            _ => None,
-        })
-        .collect();
+    let reasoning_stream = collected.reasoning_text();
+    let content_stream = collected.content_text();
 
     for forbidden in ["[THINK]", "[/THINK]"] {
         assert!(
diff --git a/paddler_tests/tests/mistral3_internal_endpoint_emits_reasoning_tokens_for_image_request.rs b/paddler_tests/tests/mistral3_internal_endpoint_emits_reasoning_tokens_for_image_request.rs
index 195214d08..2760a9ed5 100644
--- a/paddler_tests/tests/mistral3_internal_endpoint_emits_reasoning_tokens_for_image_request.rs
+++ b/paddler_tests/tests/mistral3_internal_endpoint_emits_reasoning_tokens_for_image_request.rs
@@ -23,7 +23,7 @@ async fn mistral3_internal_endpoint_emits_reasoning_tokens_for_image_request() {
 
     let image_data_uri = load_test_image_data_uri().expect("the test image must load");
 
-    let conversation_history = ConversationHistory::new(vec![ConversationMessage {
+    let conversation_history = ConversationHistory::new(vec![ConversationMessage::Message {
         content: ConversationMessageContent::Parts(vec![
             ConversationMessageContentPart::ImageUrl {
                 image_url: ImageUrl {
@@ -64,13 +64,9 @@ async fn mistral3_internal_endpoint_emits_reasoning_tokens_for_image_request() {
         "Mistral 3: expected at least one reasoning token from a `[THINK]` block when an image is attached (got {reasoning_count})"
     );
 
-    let last = collected
-        .token_results
-        .last()
-        .expect("no token results received");
-    let GeneratedTokenResult::Done(summary) = &last.token_result else {
-        panic!("last result was not Done: {last:?}");
-    };
+    let summary = collected
+        .summary()
+        .expect("the generation must finish with a summary");
 
     assert!(summary.usage.reasoning_tokens > 0);
     assert!(summary.usage.input_image_tokens > 0);
diff --git a/paddler_tests/tests/mistral3_internal_endpoint_emits_tool_call_parsed_event.rs b/paddler_tests/tests/mistral3_internal_endpoint_emits_tool_call_parsed_event.rs
index e6ac1916b..5de49bd4b 100644
--- a/paddler_tests/tests/mistral3_internal_endpoint_emits_tool_call_parsed_event.rs
+++ b/paddler_tests/tests/mistral3_internal_endpoint_emits_tool_call_parsed_event.rs
@@ -2,17 +2,18 @@
 
 use std::num::NonZeroU32;
 
-use llama_cpp_bindings_types::ParsedToolCall;
 use tokio_util::sync::CancellationToken;
 
 use paddler_messaging::conversation_history::ConversationHistory;
 use paddler_messaging::conversation_message::ConversationMessage;
 use paddler_messaging::conversation_message_content::ConversationMessageContent;
 use paddler_messaging::generated_token_result::GeneratedTokenResult;
+use paddler_messaging::generated_tool_call::GeneratedToolCall;
 use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
 use paddler_tests::get_weather_tool::get_weather_tool;
 use paddler_tests::ministral_3_cluster_params::Ministral3ClusterParams;
 use paddler_tests::start_cluster_with_ministral_3::start_cluster_with_ministral_3;
+use paddler_tests::weather_tool_question::WEATHER_TOOL_QUESTION;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn mistral3_internal_endpoint_emits_tool_call_parsed_event() {
@@ -25,13 +26,12 @@ async fn mistral3_internal_endpoint_emits_tool_call_parsed_event() {
             CancellationToken::new(),
             &ContinueFromConversationHistoryParams {
                 add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Text(
-                        "What is the weather in Paris? Use the get_weather tool to find out."
-                            .to_owned(),
-                    ),
-                    role: "user".to_owned(),
-                }]),
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Text(WEATHER_TOOL_QUESTION.to_owned()),
+                        role: "user".to_owned(),
+                    },
+                ]),
                 enable_thinking: false,
                 grammar: None,
                 max_tokens: NonZeroU32::new(400).unwrap(),
@@ -42,7 +42,7 @@ async fn mistral3_internal_endpoint_emits_tool_call_parsed_event() {
         .await
         .expect("the inference request must be accepted");
 
-    let parsed_events: Vec<&Vec<ParsedToolCall>> = collected
+    let parsed_events: Vec<&Vec<GeneratedToolCall>> = collected
         .token_results
         .iter()
         .filter_map(|event| match &event.token_result {
diff --git a/paddler_tests/tests/nomic_document_embedding_with.rs b/paddler_tests/tests/nomic_document_embedding_with.rs
new file mode 100644
index 000000000..be72e6171
--- /dev/null
+++ b/paddler_tests/tests/nomic_document_embedding_with.rs
@@ -0,0 +1,55 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_inference_parameters::embedding_parameters::EmbeddingParameters;
+use paddler_messaging::embedding_input_document::EmbeddingInputDocument;
+use paddler_messaging::embedding_normalization_method::EmbeddingNormalizationMethod;
+use paddler_messaging::embedding_vector::EmbeddingVector;
+use paddler_messaging::request_params::generate_embedding_batch_params::GenerateEmbeddingBatchParams;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_test_cluster_harness::model_card::nomic_embed_text_v1_5::nomic_embed_text_v1_5;
+use paddler_tests::start_cluster::start_cluster;
+
+pub async fn nomic_document_embedding_with(
+    embedding_parameters: EmbeddingParameters,
+) -> Vec<EmbeddingVector> {
+    let cluster = start_cluster(ClusterParams {
+        agents: AgentConfig::uniform(1, 1),
+        desired_state: ClusterDesiredState::Apply(Box::new(
+            nomic_embed_text_v1_5()
+                .into_desired_state_with_embedding_parameters(embedding_parameters),
+        )),
+        wait_for_slots_ready: true,
+        ..ClusterParams::default()
+    })
+    .await
+    .expect("the nomic embedding cluster must start");
+
+    let collected = cluster
+        .generate_embedding_batch(
+            CancellationToken::new(),
+            &GenerateEmbeddingBatchParams {
+                input_batch: vec![EmbeddingInputDocument {
+                    content: "the quick brown fox jumps over the lazy dog".to_owned(),
+                    id: "doc-1".to_owned(),
+                }],
+                normalization_method: EmbeddingNormalizationMethod::None,
+            },
+        )
+        .await
+        .expect("the document must be embedded");
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+
+    collected
+        .embeddings
+        .into_iter()
+        .map(|embedding_with_producer| embedding_with_producer.embedding.embedding)
+        .collect()
+}
diff --git a/paddler_tests/tests/nomic_embedding_batch_processing.rs b/paddler_tests/tests/nomic_embedding_batch_processing.rs
new file mode 100644
index 000000000..086076412
--- /dev/null
+++ b/paddler_tests/tests/nomic_embedding_batch_processing.rs
@@ -0,0 +1,109 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::num::NonZeroU16;
+
+use llama_cpp_bindings::context::LlamaContext;
+use llama_cpp_bindings::llama_batch::LlamaBatch;
+use llama_cpp_bindings_types::BatchCapacity;
+
+use paddler_agent_embeddings::active_embedding_request::ActiveEmbeddingRequest;
+use paddler_agent_embeddings::embedding_batch_processor::EmbeddingBatchProcessor;
+use paddler_agent_embeddings::embedding_input_tokenized::EmbeddingInputTokenized;
+use paddler_agent_embeddings::embedding_scheduler_context::EmbeddingSchedulerContext;
+use paddler_agent_embeddings::normalizes_embedding::NormalizesEmbedding;
+use paddler_agent_runtime::resident_llama_context::ResidentLlamaContext;
+use paddler_inference_parameters::pooling_type::PoolingType;
+use paddler_test_cluster_harness::model_card::nomic_embed_text_v1_5::nomic_embed_text_v1_5;
+use paddler_tests::active_embedding_request_fixture::ActiveEmbeddingRequestFixture;
+use paddler_tests::agent_inference_runtime_context::agent_inference_runtime_context;
+use paddler_tests::llama_runtime::LlamaRuntime;
+
+pub struct NomicEmbeddingBatchProcessing {
+    pub llama_runtime: LlamaRuntime,
+    pub pooling_type: PoolingType,
+    pub sequences: NonZeroU16,
+}
+
+impl NomicEmbeddingBatchProcessing {
+    pub fn load(pooling_type: PoolingType, sequences: NonZeroU16) -> Self {
+        Self {
+            llama_runtime: LlamaRuntime::load(
+                &nomic_embed_text_v1_5()
+                    .cached_path()
+                    .expect("the nomic embedding model must be in the Hugging Face cache"),
+                agent_inference_runtime_context(sequences.get()),
+            )
+            .expect("the nomic embedding model must load"),
+            pooling_type,
+            sequences,
+        }
+    }
+
+    pub fn admit(&self, inputs: Vec<EmbeddingInputTokenized>) -> ActiveEmbeddingRequestFixture {
+        ActiveEmbeddingRequestFixture::new(
+            inputs,
+            self.llama_runtime
+                .inference_runtime_context
+                .slot_aggregated_status
+                .clone(),
+            self.sequences,
+        )
+    }
+
+    pub fn batch_capacity(&self) -> BatchCapacity {
+        self.llama_runtime
+            .model_runtime_parameters
+            .n_batch
+            .capacity()
+    }
+
+    pub fn embedding_llama_context(&self) -> ResidentLlamaContext<'_> {
+        self.llama_runtime
+            .create_embedding_llama_context(
+                u32::from(self.sequences.get()),
+                self.pooling_type.clone(),
+            )
+            .expect("the embedding context must be created")
+    }
+
+    pub fn embed_next_batch<TNormalization: NormalizesEmbedding>(
+        &self,
+        batch_capacity: BatchCapacity,
+        normalization: TNormalization,
+        active_requests: Vec<ActiveEmbeddingRequest>,
+    ) -> Vec<ActiveEmbeddingRequest> {
+        self.embed_next_batch_in(
+            &mut self.embedding_llama_context(),
+            batch_capacity,
+            normalization,
+            active_requests,
+        )
+    }
+
+    pub fn embed_next_batch_in<TNormalization: NormalizesEmbedding>(
+        &self,
+        llama_context: &mut LlamaContext<'_>,
+        batch_capacity: BatchCapacity,
+        normalization: TNormalization,
+        active_requests: Vec<ActiveEmbeddingRequest>,
+    ) -> Vec<ActiveEmbeddingRequest> {
+        let mut batch = LlamaBatch::new(batch_capacity, i32::from(self.sequences.get()));
+
+        EmbeddingBatchProcessor {
+            batch: &mut batch,
+            llama_context,
+            normalization,
+            scheduler_context: &EmbeddingSchedulerContext {
+                agent_name: None,
+                desired_slots_total: self.sequences.get(),
+                n_batch: self
+                    .llama_runtime
+                    .model_runtime_parameters
+                    .n_batch
+                    .tokens_usize(),
+                pooling_type: self.pooling_type.clone(),
+            },
+        }
+        .embed_next_batch(active_requests)
+    }
+}
diff --git a/paddler_tests/tests/openai_chat_completion_accepts_an_assistant_tool_call_turn_without_content.rs b/paddler_tests/tests/openai_chat_completion_accepts_an_assistant_tool_call_turn_without_content.rs
new file mode 100644
index 000000000..4cd76bd1e
--- /dev/null
+++ b/paddler_tests/tests/openai_chat_completion_accepts_an_assistant_tool_call_turn_without_content.rs
@@ -0,0 +1,54 @@
+use async_openai::error::ApiErrorResponse;
+use async_openai::error::OpenAIError;
+use http::StatusCode;
+use serde_json::json;
+
+use paddler_test_cluster_harness::cluster_harness_error::ClusterHarnessError;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_tests::short_buffered_request_timeout::SHORT_BUFFERED_REQUEST_TIMEOUT;
+use paddler_tests::start_cluster::start_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn openai_chat_completion_accepts_an_assistant_tool_call_turn_without_content() {
+    let cluster = start_cluster(ClusterParams {
+        buffered_request_timeout: SHORT_BUFFERED_REQUEST_TIMEOUT,
+        ..ClusterParams::without_agents()
+    })
+    .await
+    .expect("a balancer without agents must start");
+
+    let outcome = cluster
+        .openai_chat_completion_non_streaming(&json!({
+            "model": "test-model",
+            "messages": [
+                { "role": "user", "content": "What does marker.txt contain?" },
+                {
+                    "role": "assistant",
+                    "content": null,
+                    "tool_calls": [{
+                        "id": "call_0",
+                        "type": "function",
+                        "function": { "name": "read", "arguments": "{\"path\":\"marker.txt\"}" }
+                    }]
+                },
+                { "role": "tool", "tool_call_id": "call_0", "content": "PADDLER-OPENCODE-MARKER" }
+            ]
+        }))
+        .await;
+
+    assert!(
+        matches!(
+            &outcome,
+            Err(ClusterHarnessError::OpenAIRequestFailed(OpenAIError::ApiError(ApiErrorResponse {
+                status_code,
+                ..
+            }))) if *status_code == StatusCode::GATEWAY_TIMEOUT
+        ),
+        "{outcome:?}"
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/openai_chat_completion_non_streaming_emits_tool_calls_for_function_tool.rs b/paddler_tests/tests/openai_chat_completion_non_streaming_emits_tool_calls_for_function_tool.rs
new file mode 100644
index 000000000..865a23a24
--- /dev/null
+++ b/paddler_tests/tests/openai_chat_completion_non_streaming_emits_tool_calls_for_function_tool.rs
@@ -0,0 +1,103 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use serde_json::Value;
+use serde_json::from_str;
+use serde_json::json;
+
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn openai_chat_completion_non_streaming_emits_tool_calls_for_function_tool() {
+    let cluster = start_cluster_with_qwen3(vec![AgentConfig::single(1)])
+        .await
+        .expect("the cluster must start");
+
+    let response = cluster
+        .openai_chat_completion_non_streaming(&json!({
+            "model": "qwen3-test",
+            "messages": [{
+                "role": "user",
+                "content": "What is the weather in Paris? Use the get_weather tool."
+            }],
+            "max_completion_tokens": 400,
+            "tools": [{
+                "type": "function",
+                "function": {
+                    "name": "get_weather",
+                    "description": "Get the current weather for a location",
+                    "parameters": {
+                        "type": "object",
+                        "properties": {
+                            "location": {"type": "string", "description": "The city name"}
+                        },
+                        "required": ["location"],
+                        "additionalProperties": false
+                    }
+                }
+            }]
+        }))
+        .await
+        .expect("the OpenAI chat completion must succeed");
+
+    let tool_calls = response
+        .pointer("/choices/0/message/tool_calls")
+        .and_then(Value::as_array)
+        .expect("response missing message.tool_calls");
+
+    assert_eq!(
+        tool_calls.len(),
+        1,
+        "expected exactly one structured tool call in non-streaming response"
+    );
+
+    let first_call = &tool_calls[0];
+
+    assert_eq!(
+        first_call.pointer("/type").and_then(Value::as_str),
+        Some("function"),
+    );
+
+    let id = first_call
+        .pointer("/id")
+        .and_then(Value::as_str)
+        .expect("tool call missing id");
+    assert!(!id.is_empty(), "tool call id must not be empty");
+
+    let function_name = first_call
+        .pointer("/function/name")
+        .and_then(Value::as_str)
+        .expect("tool call missing function.name");
+
+    assert_eq!(function_name, "get_weather");
+
+    let function_arguments = first_call
+        .pointer("/function/arguments")
+        .and_then(Value::as_str)
+        .expect("tool call missing function.arguments");
+
+    let parsed_arguments: Value =
+        from_str(function_arguments).expect("the output must be valid JSON");
+    assert!(
+        parsed_arguments.get("location").is_some(),
+        "tool-call arguments JSON missing 'location' field: {function_arguments}"
+    );
+
+    let finish_reason = response
+        .pointer("/choices/0/finish_reason")
+        .and_then(Value::as_str)
+        .expect("response missing finish_reason");
+
+    assert_eq!(finish_reason, "tool_calls");
+
+    let completion_tokens = response
+        .pointer("/usage/completion_tokens")
+        .and_then(Value::as_u64)
+        .expect("response missing usage.completion_tokens");
+    assert!(completion_tokens > 0);
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/openai_chat_completion_non_streaming_half_closed_client_releases_the_buffered_request.rs b/paddler_tests/tests/openai_chat_completion_non_streaming_half_closed_client_releases_the_buffered_request.rs
new file mode 100644
index 000000000..8d1067719
--- /dev/null
+++ b/paddler_tests/tests/openai_chat_completion_non_streaming_half_closed_client_releases_the_buffered_request.rs
@@ -0,0 +1,19 @@
+use serde_json::json;
+
+use paddler_balancer::compatibility::openai_service::openai_api_path::OpenAIApiPath;
+
+use crate::release_half_closed_openai_request::release_half_closed_openai_request;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn openai_chat_completion_non_streaming_half_closed_client_releases_the_buffered_request() {
+    release_half_closed_openai_request(
+        OpenAIApiPath::CHAT_COMPLETIONS,
+        &json!({
+            "max_completion_tokens": 2048,
+            "messages": [{"role": "user", "content": "hi"}],
+            "model": "paddler",
+            "stream": false,
+        }),
+    )
+    .await;
+}
diff --git a/paddler_tests/tests/openai_chat_completion_non_streaming_reports_a_conversation_exceeding_the_context_as_a_bad_request.rs b/paddler_tests/tests/openai_chat_completion_non_streaming_reports_a_conversation_exceeding_the_context_as_a_bad_request.rs
new file mode 100644
index 000000000..487be37b8
--- /dev/null
+++ b/paddler_tests/tests/openai_chat_completion_non_streaming_reports_a_conversation_exceeding_the_context_as_a_bad_request.rs
@@ -0,0 +1,58 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use reqwest::StatusCode;
+use serde_json::Value;
+use serde_json::json;
+use std::num::NonZeroU32;
+
+use paddler_balancer::compatibility::openai_service::openai_api_path::OpenAIApiPath;
+use paddler_openai_response_format_validator::openai_validator::OpenAIValidator;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_tests::oversized_text::oversized_text;
+use paddler_tests::post_openai_chat_completion::post_openai_chat_completion;
+use paddler_tests::start_cluster_with_qwen3_and_context_size::start_cluster_with_qwen3_and_context_size;
+
+const SEQUENCE_CONTEXT_SIZE: NonZeroU32 = NonZeroU32::new(256).unwrap();
+
+#[tokio::test(flavor = "multi_thread")]
+async fn openai_chat_completion_non_streaming_reports_a_conversation_exceeding_the_context_as_a_bad_request()
+ {
+    let cluster = start_cluster_with_qwen3_and_context_size(
+        vec![AgentConfig::single(1)],
+        SEQUENCE_CONTEXT_SIZE,
+    )
+    .await
+    .expect("the cluster must start");
+
+    let response = post_openai_chat_completion(
+        cluster
+            .balancer
+            .compat_openai_url(OpenAIApiPath::CHAT_COMPLETIONS)
+            .expect("the balancer must serve the OpenAI compatibility service"),
+        &json!({
+            "model": "test",
+            "messages": [{"role": "user", "content": oversized_text()}]
+        }),
+    )
+    .await
+    .expect("the balancer must answer the chat completion");
+
+    assert_eq!(response.status(), StatusCode::BAD_REQUEST);
+
+    let error_body = response
+        .json::<Value>()
+        .await
+        .expect("the error body must be JSON");
+
+    OpenAIValidator::new()
+        .expect("the OpenAI specification must load")
+        .validate_error_response(&error_body)
+        .expect("the error body must be an OpenAI error");
+
+    assert_eq!(error_body["error"]["type"], "invalid_request_error");
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/openai_chat_completion_non_streaming_returns_text.rs b/paddler_tests/tests/openai_chat_completion_non_streaming_returns_text.rs
new file mode 100644
index 000000000..b304f99e7
--- /dev/null
+++ b/paddler_tests/tests/openai_chat_completion_non_streaming_returns_text.rs
@@ -0,0 +1,38 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use serde_json::json;
+
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn openai_chat_completion_non_streaming_returns_text() {
+    let cluster = start_cluster_with_qwen3(AgentConfig::uniform(1, 2))
+        .await
+        .expect("the cluster must start");
+
+    let body = cluster
+        .openai_chat_completion_non_streaming(&json!({
+            "model": "test",
+            "messages": [{"role": "user", "content": "Say hello"}],
+            "max_completion_tokens": 200,
+            "stream": false,
+        }))
+        .await
+        .expect("the OpenAI chat completion must succeed");
+
+    assert_eq!(body["object"], "chat.completion");
+    assert!(body["choices"].is_array());
+    assert!(
+        !body["choices"][0]["message"]["content"]
+            .as_str()
+            .unwrap_or("")
+            .is_empty(),
+        "response content should not be empty"
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/openai_chat_completion_non_streaming_returns_usage.rs b/paddler_tests/tests/openai_chat_completion_non_streaming_returns_usage.rs
new file mode 100644
index 000000000..ccbdbdb3c
--- /dev/null
+++ b/paddler_tests/tests/openai_chat_completion_non_streaming_returns_usage.rs
@@ -0,0 +1,59 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use serde_json::Value;
+use serde_json::json;
+
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn openai_chat_completion_non_streaming_returns_usage() {
+    let cluster = start_cluster_with_qwen3(vec![AgentConfig::single(1)])
+        .await
+        .expect("the cluster must start");
+
+    let response = cluster
+        .openai_chat_completion_non_streaming(&json!({
+            "model": "qwen3-test",
+            "messages": [{"role": "user", "content": "Say hi briefly."}],
+            "max_completion_tokens": 600
+        }))
+        .await
+        .expect("the OpenAI chat completion must succeed");
+
+    let usage = response
+        .get("usage")
+        .expect("non-streaming response missing usage");
+
+    let prompt_tokens = usage
+        .get("prompt_tokens")
+        .and_then(Value::as_u64)
+        .expect("usage.prompt_tokens missing");
+    let completion_tokens = usage
+        .get("completion_tokens")
+        .and_then(Value::as_u64)
+        .expect("usage.completion_tokens missing");
+    let total_tokens = usage
+        .get("total_tokens")
+        .and_then(Value::as_u64)
+        .expect("usage.total_tokens missing");
+
+    assert!(prompt_tokens > 0);
+    assert!(completion_tokens > 0);
+    assert_eq!(total_tokens, prompt_tokens + completion_tokens);
+
+    let content = response
+        .pointer("/choices/0/message/content")
+        .and_then(Value::as_str)
+        .expect("non-streaming response missing message content");
+
+    assert!(
+        !content.is_empty(),
+        "non-streaming content must not be empty"
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/openai_chat_completion_non_streaming_usage_adds_up_with_tool_calls.rs b/paddler_tests/tests/openai_chat_completion_non_streaming_usage_adds_up_with_tool_calls.rs
new file mode 100644
index 000000000..4d0d3710b
--- /dev/null
+++ b/paddler_tests/tests/openai_chat_completion_non_streaming_usage_adds_up_with_tool_calls.rs
@@ -0,0 +1,74 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use serde_json::Value;
+use serde_json::json;
+
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn openai_chat_completion_non_streaming_usage_adds_up_with_tool_calls() {
+    let cluster = start_cluster_with_qwen3(vec![AgentConfig::single(1)])
+        .await
+        .expect("the cluster must start");
+
+    let response = cluster
+        .openai_chat_completion_non_streaming(&json!({
+            "model": "qwen3-test",
+            "messages": [{
+                "role": "user",
+                "content": "What is the weather in Paris? Use the get_weather tool."
+            }],
+            "max_completion_tokens": 400,
+            "tools": [{
+                "type": "function",
+                "function": {
+                    "name": "get_weather",
+                    "description": "Get the current weather for a location",
+                    "parameters": {
+                        "type": "object",
+                        "properties": {
+                            "location": {"type": "string"}
+                        },
+                        "required": ["location"],
+                        "additionalProperties": false
+                    }
+                }
+            }]
+        }))
+        .await
+        .expect("the OpenAI chat completion must succeed");
+
+    let tool_calls = response
+        .pointer("/choices/0/message/tool_calls")
+        .and_then(Value::as_array)
+        .expect("response missing message.tool_calls");
+    assert_ne!(tool_calls, &Vec::<Value>::new());
+
+    let usage = response.get("usage").expect("response missing usage");
+
+    let prompt_tokens = usage
+        .get("prompt_tokens")
+        .and_then(Value::as_u64)
+        .expect("usage.prompt_tokens missing");
+    let completion_tokens = usage
+        .get("completion_tokens")
+        .and_then(Value::as_u64)
+        .expect("usage.completion_tokens missing");
+    let total_tokens = usage
+        .get("total_tokens")
+        .and_then(Value::as_u64)
+        .expect("usage.total_tokens missing");
+
+    assert!(prompt_tokens > 0);
+    assert!(
+        completion_tokens > 0,
+        "expected non-zero completion_tokens for a tool-call response (got {completion_tokens})"
+    );
+    assert_eq!(total_tokens, prompt_tokens + completion_tokens);
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/openai_chat_completion_reports_a_buffered_request_timeout_as_a_gateway_timeout.rs b/paddler_tests/tests/openai_chat_completion_reports_a_buffered_request_timeout_as_a_gateway_timeout.rs
new file mode 100644
index 000000000..7c358384d
--- /dev/null
+++ b/paddler_tests/tests/openai_chat_completion_reports_a_buffered_request_timeout_as_a_gateway_timeout.rs
@@ -0,0 +1,45 @@
+use reqwest::StatusCode;
+use serde_json::Value;
+
+use paddler_balancer::compatibility::openai_service::openai_api_path::OpenAIApiPath;
+use paddler_openai_response_format_validator::openai_validator::OpenAIValidator;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_tests::minimal_openai_chat_completion::minimal_openai_chat_completion;
+use paddler_tests::short_buffered_request_timeout::SHORT_BUFFERED_REQUEST_TIMEOUT;
+use paddler_tests::start_cluster::start_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn openai_chat_completion_reports_a_buffered_request_timeout_as_a_gateway_timeout() {
+    let cluster = start_cluster(ClusterParams {
+        buffered_request_timeout: SHORT_BUFFERED_REQUEST_TIMEOUT,
+        ..ClusterParams::without_agents()
+    })
+    .await
+    .expect("a balancer without agents must start");
+
+    let response = minimal_openai_chat_completion(
+        cluster
+            .balancer
+            .compat_openai_url(OpenAIApiPath::CHAT_COMPLETIONS)
+            .expect("the balancer must serve the OpenAI compatibility service"),
+    )
+    .await
+    .expect("the balancer must answer the chat completion");
+
+    assert_eq!(response.status(), StatusCode::GATEWAY_TIMEOUT);
+
+    OpenAIValidator::new()
+        .expect("the OpenAI specification must load")
+        .validate_error_response(
+            &response
+                .json::<Value>()
+                .await
+                .expect("the error body must be JSON"),
+        )
+        .expect("the error body must be an OpenAI error");
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/openai_chat_completion_reports_a_full_buffer_as_service_unavailable.rs b/paddler_tests/tests/openai_chat_completion_reports_a_full_buffer_as_service_unavailable.rs
new file mode 100644
index 000000000..f0f90aa1b
--- /dev/null
+++ b/paddler_tests/tests/openai_chat_completion_reports_a_full_buffer_as_service_unavailable.rs
@@ -0,0 +1,44 @@
+use reqwest::StatusCode;
+use serde_json::Value;
+
+use paddler_balancer::compatibility::openai_service::openai_api_path::OpenAIApiPath;
+use paddler_openai_response_format_validator::openai_validator::OpenAIValidator;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_tests::minimal_openai_chat_completion::minimal_openai_chat_completion;
+use paddler_tests::start_cluster::start_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn openai_chat_completion_reports_a_full_buffer_as_service_unavailable() {
+    let cluster = start_cluster(ClusterParams {
+        max_buffered_requests: 0,
+        ..ClusterParams::without_agents()
+    })
+    .await
+    .expect("a balancer without agents must start");
+
+    let response = minimal_openai_chat_completion(
+        cluster
+            .balancer
+            .compat_openai_url(OpenAIApiPath::CHAT_COMPLETIONS)
+            .expect("the balancer must serve the OpenAI compatibility service"),
+    )
+    .await
+    .expect("the balancer must answer the chat completion");
+
+    assert_eq!(response.status(), StatusCode::SERVICE_UNAVAILABLE);
+
+    OpenAIValidator::new()
+        .expect("the OpenAI specification must load")
+        .validate_error_response(
+            &response
+                .json::<Value>()
+                .await
+                .expect("the error body must be JSON"),
+        )
+        .expect("the error body must be an OpenAI error");
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/openai_chat_completion_reports_a_stalled_generation_as_a_gateway_timeout.rs b/paddler_tests/tests/openai_chat_completion_reports_a_stalled_generation_as_a_gateway_timeout.rs
new file mode 100644
index 000000000..cfaede7d6
--- /dev/null
+++ b/paddler_tests/tests/openai_chat_completion_reports_a_stalled_generation_as_a_gateway_timeout.rs
@@ -0,0 +1,53 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::time::Duration;
+
+use reqwest::StatusCode;
+use serde_json::Value;
+
+use paddler_balancer::compatibility::openai_service::openai_api_path::OpenAIApiPath;
+use paddler_openai_response_format_validator::openai_validator::OpenAIValidator;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
+use paddler_tests::minimal_openai_chat_completion::minimal_openai_chat_completion;
+use paddler_tests::start_cluster::start_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn openai_chat_completion_reports_a_stalled_generation_as_a_gateway_timeout() {
+    let cluster = start_cluster(ClusterParams {
+        agents: vec![AgentConfig::single(1)],
+        desired_state: ClusterDesiredState::Apply(Box::new(qwen3_0_6b().into_desired_state())),
+        inference_item_timeout: Duration::ZERO,
+        ..ClusterParams::default()
+    })
+    .await
+    .expect("a text generation cluster must start");
+
+    let response = minimal_openai_chat_completion(
+        cluster
+            .balancer
+            .compat_openai_url(OpenAIApiPath::CHAT_COMPLETIONS)
+            .expect("the balancer must serve the OpenAI compatibility service"),
+    )
+    .await
+    .expect("the balancer must answer the chat completion");
+
+    assert_eq!(response.status(), StatusCode::GATEWAY_TIMEOUT);
+
+    OpenAIValidator::new()
+        .expect("the OpenAI specification must load")
+        .validate_error_response(
+            &response
+                .json::<Value>()
+                .await
+                .expect("the error body must be JSON"),
+        )
+        .expect("the error body must be an OpenAI error");
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/openai_chat_completion_reports_a_tool_call_violating_its_schema_as_a_server_error.rs b/paddler_tests/tests/openai_chat_completion_reports_a_tool_call_violating_its_schema_as_a_server_error.rs
new file mode 100644
index 000000000..f7532e53b
--- /dev/null
+++ b/paddler_tests/tests/openai_chat_completion_reports_a_tool_call_violating_its_schema_as_a_server_error.rs
@@ -0,0 +1,79 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use reqwest::StatusCode;
+use serde_json::Value;
+use serde_json::json;
+
+use paddler_balancer::compatibility::openai_service::openai_api_path::OpenAIApiPath;
+use paddler_openai_response_format_validator::openai_validator::OpenAIValidator;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
+use paddler_tests::post_openai_chat_completion::post_openai_chat_completion;
+use paddler_tests::start_cluster::start_cluster;
+use paddler_tests::weather_tool_parameters_with_an_unsatisfiable_location::weather_tool_parameters_with_an_unsatisfiable_location;
+use paddler_tests::weather_tool_question::WEATHER_TOOL_QUESTION;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn openai_chat_completion_reports_a_tool_call_violating_its_schema_as_a_server_error() {
+    let cluster = start_cluster(ClusterParams {
+        agents: vec![AgentConfig::single(1)],
+        desired_state: ClusterDesiredState::Apply(Box::new(qwen3_0_6b().into_desired_state())),
+        wait_for_slots_ready: true,
+        ..ClusterParams::default()
+    })
+    .await
+    .expect("the cluster must start");
+
+    let response = post_openai_chat_completion(
+        cluster
+            .balancer
+            .compat_openai_url(OpenAIApiPath::CHAT_COMPLETIONS)
+            .expect("the balancer must serve the OpenAI compatibility service"),
+        &json!({
+            "model": "qwen3-test",
+            "messages": [{
+                "role": "user",
+                "content": WEATHER_TOOL_QUESTION
+            }],
+            "max_completion_tokens": 400,
+            "tools": [{
+                "type": "function",
+                "function": {
+                    "name": "get_weather",
+                    "description": "Get the current weather for a location",
+                    "parameters": weather_tool_parameters_with_an_unsatisfiable_location()
+                }
+            }]
+        }),
+    )
+    .await
+    .expect("the balancer must answer the chat completion");
+
+    assert_eq!(response.status(), StatusCode::INTERNAL_SERVER_ERROR);
+
+    let error_body = response
+        .json::<Value>()
+        .await
+        .expect("the error body must be JSON");
+
+    OpenAIValidator::new()
+        .expect("the OpenAI specification must load")
+        .validate_error_response(&error_body)
+        .expect("the error body must be an OpenAI error");
+
+    assert_eq!(error_body["error"]["type"], "server_error");
+    assert!(
+        error_body["error"]["message"]
+            .as_str()
+            .expect("the error must carry a message")
+            .contains("get_weather"),
+        "the error must name the tool whose call broke its schema: {error_body}"
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/openai_chat_completion_streaming_emits_tool_calls_for_function_tool.rs b/paddler_tests/tests/openai_chat_completion_streaming_emits_tool_calls_for_function_tool.rs
new file mode 100644
index 000000000..59ecfd827
--- /dev/null
+++ b/paddler_tests/tests/openai_chat_completion_streaming_emits_tool_calls_for_function_tool.rs
@@ -0,0 +1,86 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use serde_json::Value;
+use serde_json::from_str;
+use serde_json::json;
+
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn openai_chat_completion_streaming_emits_tool_calls_for_function_tool() {
+    let cluster = start_cluster_with_qwen3(vec![AgentConfig::single(1)])
+        .await
+        .expect("the cluster must start");
+
+    let chunks = cluster
+        .openai_chat_completion_streaming(&json!({
+            "model": "qwen3-test",
+            "messages": [{
+                "role": "user",
+                "content": "What is the weather in Paris? Use the get_weather tool."
+            }],
+            "stream": true,
+            "max_completion_tokens": 400,
+            "tools": [{
+                "type": "function",
+                "function": {
+                    "name": "get_weather",
+                    "description": "Get the current weather for a location",
+                    "parameters": {
+                        "type": "object",
+                        "properties": {
+                            "location": {"type": "string", "description": "The city name"}
+                        },
+                        "required": ["location"],
+                        "additionalProperties": false
+                    }
+                }
+            }]
+        }))
+        .await
+        .expect("the OpenAI chat completion stream must succeed");
+
+    let chunks_with_tool_calls: Vec<&Value> = chunks
+        .iter()
+        .filter(|chunk| {
+            chunk
+                .pointer("/choices/0/delta/tool_calls")
+                .and_then(Value::as_array)
+                .is_some_and(|calls| !calls.is_empty())
+        })
+        .collect();
+
+    assert_eq!(
+        chunks_with_tool_calls.len(),
+        1,
+        "expected exactly one structured tool-call chunk per call"
+    );
+
+    let structured_chunk = chunks_with_tool_calls[0];
+
+    assert_eq!(
+        structured_chunk
+            .pointer("/choices/0/delta/tool_calls/0/function/name")
+            .and_then(Value::as_str),
+        Some("get_weather")
+    );
+
+    let function_arguments = structured_chunk
+        .pointer("/choices/0/delta/tool_calls/0/function/arguments")
+        .and_then(Value::as_str)
+        .expect("the structured tool-call chunk must carry the function arguments");
+
+    let parsed_arguments: Value =
+        from_str(function_arguments).expect("the output must be valid JSON");
+
+    assert!(
+        parsed_arguments.get("location").is_some(),
+        "tool-call arguments JSON missing 'location' field: {function_arguments}"
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/openai_chat_completion_streaming_emits_usage_when_requested.rs b/paddler_tests/tests/openai_chat_completion_streaming_emits_usage_when_requested.rs
new file mode 100644
index 000000000..8193b16dd
--- /dev/null
+++ b/paddler_tests/tests/openai_chat_completion_streaming_emits_usage_when_requested.rs
@@ -0,0 +1,65 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use serde_json::Value;
+use serde_json::json;
+
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn openai_chat_completion_streaming_emits_usage_when_requested() {
+    let cluster = start_cluster_with_qwen3(vec![AgentConfig::single(1)])
+        .await
+        .expect("the cluster must start");
+
+    let chunks = cluster
+        .openai_chat_completion_streaming(&json!({
+            "model": "qwen3-test",
+            "messages": [{"role": "user", "content": "Say hi briefly."}],
+            "stream": true,
+            "stream_options": {"include_usage": true},
+            "max_completion_tokens": 80
+        }))
+        .await
+        .expect("the OpenAI chat completion stream must succeed");
+
+    let last_chunk = chunks
+        .last()
+        .expect("no chunks received from streaming endpoint");
+
+    let usage = last_chunk
+        .get("usage")
+        .expect("trailing chunk lacks usage field");
+
+    let prompt_tokens = usage
+        .get("prompt_tokens")
+        .and_then(Value::as_u64)
+        .expect("usage.prompt_tokens missing or not u64");
+    let completion_tokens = usage
+        .get("completion_tokens")
+        .and_then(Value::as_u64)
+        .expect("usage.completion_tokens missing or not u64");
+    let total_tokens = usage
+        .get("total_tokens")
+        .and_then(Value::as_u64)
+        .expect("usage.total_tokens missing or not u64");
+
+    assert!(prompt_tokens > 0);
+    assert!(completion_tokens > 0);
+    assert_eq!(total_tokens, prompt_tokens + completion_tokens);
+
+    let trailing_choices = last_chunk
+        .get("choices")
+        .and_then(Value::as_array)
+        .expect("trailing chunk lacks choices array");
+
+    assert!(
+        trailing_choices.is_empty(),
+        "OpenAI usage chunk must have empty choices array, got: {trailing_choices:?}"
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/openai_chat_completion_streaming_ends_at_a_tool_call_violating_its_schema.rs b/paddler_tests/tests/openai_chat_completion_streaming_ends_at_a_tool_call_violating_its_schema.rs
new file mode 100644
index 000000000..6b8e3ba9f
--- /dev/null
+++ b/paddler_tests/tests/openai_chat_completion_streaming_ends_at_a_tool_call_violating_its_schema.rs
@@ -0,0 +1,48 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use serde_json::Value;
+use serde_json::json;
+
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
+use paddler_tests::weather_tool_parameters_with_an_unsatisfiable_location::weather_tool_parameters_with_an_unsatisfiable_location;
+use paddler_tests::weather_tool_question::WEATHER_TOOL_QUESTION;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn openai_chat_completion_streaming_ends_at_a_tool_call_violating_its_schema() {
+    let cluster = start_cluster_with_qwen3(vec![AgentConfig::single(1)])
+        .await
+        .expect("the cluster must start");
+
+    let chunks = cluster
+        .openai_chat_completion_streaming(&json!({
+            "model": "qwen3-test",
+            "messages": [{"role": "user", "content": WEATHER_TOOL_QUESTION}],
+            "max_completion_tokens": 400,
+            "stream": true,
+            "tools": [{
+                "type": "function",
+                "function": {
+                    "name": "get_weather",
+                    "description": "Get the current weather for a location",
+                    "parameters": weather_tool_parameters_with_an_unsatisfiable_location()
+                }
+            }]
+        }))
+        .await
+        .expect("the OpenAI chat completion stream must be readable");
+
+    assert_eq!(
+        chunks
+            .iter()
+            .skip_while(|chunk| chunk.get("error").is_none())
+            .map(|chunk| chunk["error"]["type"].clone())
+            .collect::<Vec<Value>>(),
+        vec![json!("server_error")]
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/openai_chat_completion_streaming_half_closed_client_releases_the_buffered_request.rs b/paddler_tests/tests/openai_chat_completion_streaming_half_closed_client_releases_the_buffered_request.rs
new file mode 100644
index 000000000..dca34f455
--- /dev/null
+++ b/paddler_tests/tests/openai_chat_completion_streaming_half_closed_client_releases_the_buffered_request.rs
@@ -0,0 +1,19 @@
+use serde_json::json;
+
+use paddler_balancer::compatibility::openai_service::openai_api_path::OpenAIApiPath;
+
+use crate::release_half_closed_openai_request::release_half_closed_openai_request;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn openai_chat_completion_streaming_half_closed_client_releases_the_buffered_request() {
+    release_half_closed_openai_request(
+        OpenAIApiPath::CHAT_COMPLETIONS,
+        &json!({
+            "max_completion_tokens": 2048,
+            "messages": [{"role": "user", "content": "hi"}],
+            "model": "paddler",
+            "stream": true,
+        }),
+    )
+    .await;
+}
diff --git a/paddler_tests/tests/openai_chat_completion_streaming_omits_usage_when_not_requested.rs b/paddler_tests/tests/openai_chat_completion_streaming_omits_usage_when_not_requested.rs
new file mode 100644
index 000000000..96dda12b0
--- /dev/null
+++ b/paddler_tests/tests/openai_chat_completion_streaming_omits_usage_when_not_requested.rs
@@ -0,0 +1,40 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use serde_json::json;
+
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn openai_chat_completion_streaming_omits_usage_when_not_requested() {
+    let cluster = start_cluster_with_qwen3(vec![AgentConfig::single(1)])
+        .await
+        .expect("the cluster must start");
+
+    let chunks = cluster
+        .openai_chat_completion_streaming(&json!({
+            "model": "qwen3-test",
+            "messages": [{"role": "user", "content": "Say hi briefly."}],
+            "stream": true,
+            "max_completion_tokens": 50
+        }))
+        .await
+        .expect("the OpenAI chat completion stream must succeed");
+
+    assert!(!chunks.is_empty(), "expected at least one chunk");
+
+    let chunks_with_usage = chunks
+        .iter()
+        .filter(|chunk| chunk.get("usage").is_some())
+        .count();
+
+    assert_eq!(
+        chunks_with_usage, 0,
+        "expected no usage chunks when stream_options.include_usage is absent, got {chunks_with_usage}"
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/openai_chat_completion_streaming_rejects_a_tool_that_requires_an_undeclared_property.rs b/paddler_tests/tests/openai_chat_completion_streaming_rejects_a_tool_that_requires_an_undeclared_property.rs
new file mode 100644
index 000000000..23af6c572
--- /dev/null
+++ b/paddler_tests/tests/openai_chat_completion_streaming_rejects_a_tool_that_requires_an_undeclared_property.rs
@@ -0,0 +1,54 @@
+use async_openai::error::ApiErrorResponse;
+use async_openai::error::OpenAIError;
+use http::StatusCode;
+use serde_json::json;
+
+use paddler_test_cluster_harness::cluster_harness_error::ClusterHarnessError;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_tests::start_cluster::start_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn openai_chat_completion_streaming_rejects_a_tool_that_requires_an_undeclared_property() {
+    let cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("a balancer without agents must start");
+
+    let rejection = cluster
+        .openai_chat_completion_streaming(&json!({
+            "model": "test-model",
+            "messages": [{ "role": "user", "content": "hi" }],
+            "stream": true,
+            "tools": [
+                {
+                    "type": "function",
+                    "function": {
+                        "name": "broken",
+                        "parameters": {
+                            "type": "object",
+                            "properties": { "present": { "type": "string" } },
+                            "required": ["absent"]
+                        }
+                    }
+                }
+            ]
+        }))
+        .await;
+
+    assert!(
+        matches!(
+            &rejection,
+            Err(ClusterHarnessError::OpenAIRequestFailed(OpenAIError::ApiError(ApiErrorResponse {
+                status_code,
+                api_error,
+            }))) if *status_code == StatusCode::BAD_REQUEST
+                && api_error.r#type.as_deref() == Some("invalid_request_error")
+                && api_error.message == "Required field 'absent' not found in properties"
+        ),
+        "{rejection:?}"
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/openai_chat_completion_streaming_reports_a_buffered_request_timeout_as_an_error_chunk.rs b/paddler_tests/tests/openai_chat_completion_streaming_reports_a_buffered_request_timeout_as_an_error_chunk.rs
new file mode 100644
index 000000000..0ffda64e9
--- /dev/null
+++ b/paddler_tests/tests/openai_chat_completion_streaming_reports_a_buffered_request_timeout_as_an_error_chunk.rs
@@ -0,0 +1,41 @@
+use serde_json::json;
+
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_tests::short_buffered_request_timeout::SHORT_BUFFERED_REQUEST_TIMEOUT;
+use paddler_tests::start_cluster::start_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn openai_chat_completion_streaming_reports_a_buffered_request_timeout_as_an_error_chunk() {
+    let cluster = start_cluster(ClusterParams {
+        buffered_request_timeout: SHORT_BUFFERED_REQUEST_TIMEOUT,
+        ..ClusterParams::without_agents()
+    })
+    .await
+    .expect("a balancer without agents must start");
+
+    let chunks = cluster
+        .openai_chat_completion_streaming(&json!({
+            "model": "test",
+            "messages": [{"role": "user", "content": "hi"}],
+            "stream": true,
+        }))
+        .await
+        .expect("the OpenAI chat completion stream must be readable");
+
+    assert_eq!(
+        chunks,
+        vec![json!({
+            "error": {
+                "code": null,
+                "message": "Waiting for available slot timed out",
+                "param": null,
+                "type": "server_error"
+            }
+        })]
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/openai_chat_completion_streaming_reports_a_conversation_exceeding_the_context_as_an_error_chunk.rs b/paddler_tests/tests/openai_chat_completion_streaming_reports_a_conversation_exceeding_the_context_as_an_error_chunk.rs
new file mode 100644
index 000000000..2ada369c1
--- /dev/null
+++ b/paddler_tests/tests/openai_chat_completion_streaming_reports_a_conversation_exceeding_the_context_as_an_error_chunk.rs
@@ -0,0 +1,43 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use serde_json::json;
+use std::num::NonZeroU32;
+
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_tests::oversized_text::oversized_text;
+use paddler_tests::start_cluster_with_qwen3_and_context_size::start_cluster_with_qwen3_and_context_size;
+
+const SEQUENCE_CONTEXT_SIZE: NonZeroU32 = NonZeroU32::new(256).unwrap();
+
+#[tokio::test(flavor = "multi_thread")]
+async fn openai_chat_completion_streaming_reports_a_conversation_exceeding_the_context_as_an_error_chunk()
+ {
+    let cluster = start_cluster_with_qwen3_and_context_size(
+        vec![AgentConfig::single(1)],
+        SEQUENCE_CONTEXT_SIZE,
+    )
+    .await
+    .expect("the cluster must start");
+
+    let chunks = cluster
+        .openai_chat_completion_streaming(&json!({
+            "model": "test",
+            "messages": [{"role": "user", "content": oversized_text()}],
+            "stream": true,
+        }))
+        .await
+        .expect("the OpenAI chat completion stream must be readable");
+
+    assert_eq!(
+        chunks
+            .iter()
+            .map(|chunk| chunk["error"]["type"].clone())
+            .collect::<Vec<_>>(),
+        vec![json!("invalid_request_error")]
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/openai_chat_completion_streaming_returns_chunks.rs b/paddler_tests/tests/openai_chat_completion_streaming_returns_chunks.rs
new file mode 100644
index 000000000..64a954902
--- /dev/null
+++ b/paddler_tests/tests/openai_chat_completion_streaming_returns_chunks.rs
@@ -0,0 +1,31 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use serde_json::json;
+
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn openai_chat_completion_streaming_returns_chunks() {
+    let cluster = start_cluster_with_qwen3(AgentConfig::uniform(1, 2))
+        .await
+        .expect("the cluster must start");
+
+    let chunks = cluster
+        .openai_chat_completion_streaming(&json!({
+            "model": "test",
+            "messages": [{"role": "user", "content": "Say hello"}],
+            "max_completion_tokens": 10,
+            "stream": true,
+        }))
+        .await
+        .expect("the OpenAI chat completion stream must succeed");
+
+    assert!(!chunks.is_empty(), "should have received streaming chunks");
+    assert_eq!(chunks[0]["object"], "chat.completion.chunk");
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/openai_chat_completion_streaming_usage_breakdown_with_thinking.rs b/paddler_tests/tests/openai_chat_completion_streaming_usage_breakdown_with_thinking.rs
new file mode 100644
index 000000000..fe1d8adea
--- /dev/null
+++ b/paddler_tests/tests/openai_chat_completion_streaming_usage_breakdown_with_thinking.rs
@@ -0,0 +1,62 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use serde_json::Value;
+use serde_json::json;
+
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn openai_chat_completion_streaming_usage_breakdown_with_thinking() {
+    let cluster = start_cluster_with_qwen3(vec![AgentConfig::single(1)])
+        .await
+        .expect("the cluster must start");
+
+    let chunks = cluster
+        .openai_chat_completion_streaming(&json!({
+            "model": "qwen3-test",
+            "messages": [{
+                "role": "user",
+                "content": "What is two plus two? Think briefly before answering."
+            }],
+            "stream": true,
+            "stream_options": {"include_usage": true},
+            "reasoning_effort": "medium",
+            "max_completion_tokens": 200
+        }))
+        .await
+        .expect("the OpenAI chat completion stream must succeed");
+
+    let usage_chunk = chunks
+        .iter()
+        .rev()
+        .find(|chunk| chunk.get("usage").is_some_and(|usage| !usage.is_null()))
+        .expect("no chunk contained usage information");
+
+    let prompt_tokens = usage_chunk
+        .pointer("/usage/prompt_tokens")
+        .and_then(Value::as_u64)
+        .expect("usage chunk missing prompt_tokens");
+    let completion_tokens = usage_chunk
+        .pointer("/usage/completion_tokens")
+        .and_then(Value::as_u64)
+        .expect("usage chunk missing completion_tokens");
+    let total_tokens = usage_chunk
+        .pointer("/usage/total_tokens")
+        .and_then(Value::as_u64)
+        .expect("usage chunk missing total_tokens");
+    let reasoning_tokens = usage_chunk
+        .pointer("/usage/completion_tokens_details/reasoning_tokens")
+        .and_then(Value::as_u64)
+        .expect("usage chunk missing the reasoning token breakdown");
+
+    assert!(prompt_tokens > 0);
+    assert!(reasoning_tokens > 0);
+    assert!(reasoning_tokens < completion_tokens);
+    assert_eq!(total_tokens, prompt_tokens + completion_tokens);
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/openai_chat_completion_with_reasoning_effort_none_generates_no_reasoning_tokens.rs b/paddler_tests/tests/openai_chat_completion_with_reasoning_effort_none_generates_no_reasoning_tokens.rs
new file mode 100644
index 000000000..5d0e3ee2e
--- /dev/null
+++ b/paddler_tests/tests/openai_chat_completion_with_reasoning_effort_none_generates_no_reasoning_tokens.rs
@@ -0,0 +1,36 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use serde_json::Value;
+use serde_json::json;
+
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn openai_chat_completion_with_reasoning_effort_none_generates_no_reasoning_tokens() {
+    let cluster = start_cluster_with_qwen3(vec![AgentConfig::single(1)])
+        .await
+        .expect("the cluster must start");
+
+    let response = cluster
+        .openai_chat_completion_non_streaming(&json!({
+            "model": "qwen3-test",
+            "messages": [{"role": "user", "content": "What is two plus two?"}],
+            "max_completion_tokens": 200,
+            "reasoning_effort": "none"
+        }))
+        .await
+        .expect("the OpenAI chat completion must succeed");
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+
+    assert_eq!(
+        response
+            .pointer("/usage/completion_tokens_details/reasoning_tokens")
+            .and_then(Value::as_u64),
+        Some(0)
+    );
+}
diff --git a/paddler_tests/tests/openai_health_returns_ok.rs b/paddler_tests/tests/openai_health_returns_ok.rs
new file mode 100644
index 000000000..9ac2bc077
--- /dev/null
+++ b/paddler_tests/tests/openai_health_returns_ok.rs
@@ -0,0 +1,26 @@
+use tokio_util::sync::CancellationToken;
+
+use paddler_client::reports_health::ReportsHealth as _;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_tests::start_cluster::start_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn openai_health_returns_ok() {
+    let cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("the cluster must start");
+
+    let health = cluster
+        .compat_openai_health_client()
+        .expect("a text generation cluster must serve OpenAI compatibility")
+        .get_health(CancellationToken::new())
+        .await
+        .expect("failed to GET OpenAI compat /health");
+
+    assert_eq!(health, "OK");
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/openai_responses_accepts_its_own_output_replayed_as_input.rs b/paddler_tests/tests/openai_responses_accepts_its_own_output_replayed_as_input.rs
new file mode 100644
index 000000000..f57ec9d9b
--- /dev/null
+++ b/paddler_tests/tests/openai_responses_accepts_its_own_output_replayed_as_input.rs
@@ -0,0 +1,63 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use serde_json::Value;
+use serde_json::json;
+
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn openai_responses_accepts_its_own_output_replayed_as_input() {
+    let cluster = start_cluster_with_qwen3(vec![AgentConfig::single(1)])
+        .await
+        .expect("the cluster must start");
+    let first_question = json!({ "type": "message", "role": "user", "content": "Say hi briefly." });
+
+    let first_response = cluster
+        .openai_responses_non_streaming(&json!({
+            "model": "qwen3-test",
+            "input": [first_question],
+            "max_output_tokens": 600
+        }))
+        .await
+        .expect("the first OpenAI response must succeed");
+    let first_output = first_response
+        .get("output")
+        .and_then(Value::as_array)
+        .expect("the first response must carry its output items");
+
+    assert_eq!(
+        first_output
+            .iter()
+            .map(|output_item| output_item.get("type").and_then(Value::as_str))
+            .collect::<Vec<_>>(),
+        vec![Some("reasoning"), Some("message")],
+        "{first_response}"
+    );
+
+    let mut replayed_input = vec![first_question];
+
+    replayed_input.extend(first_output.iter().cloned());
+    replayed_input
+        .push(json!({ "type": "message", "role": "user", "content": "Now say bye briefly." }));
+
+    let second_response = cluster
+        .openai_responses_non_streaming(&json!({
+            "model": "qwen3-test",
+            "input": replayed_input,
+            "max_output_tokens": 600
+        }))
+        .await
+        .expect("the response replaying the first output must succeed");
+
+    assert_eq!(
+        second_response.get("status").and_then(Value::as_str),
+        Some("completed"),
+        "{second_response}"
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/openai_responses_are_rejected_while_the_cluster_serves_embeddings.rs b/paddler_tests/tests/openai_responses_are_rejected_while_the_cluster_serves_embeddings.rs
deleted file mode 100644
index 08ab49620..000000000
--- a/paddler_tests/tests/openai_responses_are_rejected_while_the_cluster_serves_embeddings.rs
+++ /dev/null
@@ -1,55 +0,0 @@
-use async_openai::error::ApiErrorResponse;
-use async_openai::error::OpenAIError;
-use http::StatusCode;
-use serde_json::Value;
-use serde_json::from_str;
-use serde_json::json;
-
-use paddler_inference_parameters::inference_parameters::InferenceParameters;
-use paddler_messaging::balancer_desired_state::BalancerDesiredState;
-use paddler_test_cluster_harness::cluster_harness_error::ClusterHarnessError;
-use paddler_test_cluster_harness::cluster_params::ClusterParams;
-use paddler_tests::start_cluster::start_cluster;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn openai_responses_are_rejected_while_the_cluster_serves_embeddings() {
-    let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        desired_state: Some(BalancerDesiredState {
-            inference_parameters: InferenceParameters {
-                enable_embeddings: true,
-                ..InferenceParameters::default()
-            },
-            ..BalancerDesiredState::default()
-        }),
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("a balancer serving embeddings must start");
-
-    let rejection = cluster
-        .openai_responses_non_streaming(&json!({ "model": "test-model", "input": "hi" }))
-        .await;
-
-    assert!(matches!(
-        rejection,
-        Err(ClusterHarnessError::OpenAIRequestFailed(OpenAIError::ApiError(ApiErrorResponse {
-            status_code,
-            api_error,
-        }))) if status_code == StatusCode::NOT_IMPLEMENTED
-            && from_str::<Value>(&api_error.message).is_ok_and(|error_body| error_body == json!({
-                "error": {
-                    "message": "Responses are disabled while the cluster is configured for embeddings",
-                    "type": "server_error",
-                    "param": null,
-                    "code": null
-                }
-            }))
-    ));
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_tests/tests/openai_responses_half_closed_client_releases_the_buffered_request.rs b/paddler_tests/tests/openai_responses_half_closed_client_releases_the_buffered_request.rs
new file mode 100644
index 000000000..c5c14bbf4
--- /dev/null
+++ b/paddler_tests/tests/openai_responses_half_closed_client_releases_the_buffered_request.rs
@@ -0,0 +1,19 @@
+use serde_json::json;
+
+use paddler_balancer::compatibility::openai_service::openai_api_path::OpenAIApiPath;
+
+use crate::release_half_closed_openai_request::release_half_closed_openai_request;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn openai_responses_half_closed_client_releases_the_buffered_request() {
+    release_half_closed_openai_request(
+        OpenAIApiPath::RESPONSES,
+        &json!({
+            "input": "hi",
+            "max_output_tokens": 2048,
+            "model": "paddler",
+            "stream": true,
+        }),
+    )
+    .await;
+}
diff --git a/paddler_tests/tests/openai_responses_non_streaming_returns_text_and_usage.rs b/paddler_tests/tests/openai_responses_non_streaming_returns_text_and_usage.rs
new file mode 100644
index 000000000..86c217c68
--- /dev/null
+++ b/paddler_tests/tests/openai_responses_non_streaming_returns_text_and_usage.rs
@@ -0,0 +1,64 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use serde_json::Value;
+use serde_json::json;
+
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_tests::responses_message_text::responses_message_text;
+use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn openai_responses_non_streaming_returns_text_and_usage() {
+    let cluster = start_cluster_with_qwen3(vec![AgentConfig::single(1)])
+        .await
+        .expect("the cluster must start");
+
+    let response = cluster
+        .openai_responses_non_streaming(&json!({
+            "model": "qwen3-test",
+            "input": "Say hi briefly.",
+            "max_output_tokens": 600
+        }))
+        .await
+        .expect("the OpenAI response must succeed");
+
+    assert_eq!(
+        response.get("status").and_then(Value::as_str),
+        Some("completed"),
+        "responses status must be completed: {response}"
+    );
+
+    let usage = response
+        .get("usage")
+        .expect("the Responses response must report its usage");
+
+    let input_tokens = usage
+        .get("input_tokens")
+        .and_then(Value::as_u64)
+        .expect("usage.input_tokens missing");
+    let output_tokens = usage
+        .get("output_tokens")
+        .and_then(Value::as_u64)
+        .expect("usage.output_tokens missing");
+    let total_tokens = usage
+        .get("total_tokens")
+        .and_then(Value::as_u64)
+        .expect("usage.total_tokens missing");
+
+    assert!(input_tokens > 0);
+    assert!(output_tokens > 0);
+    assert_eq!(total_tokens, input_tokens + output_tokens);
+
+    let message_text =
+        responses_message_text(&response).expect("the Responses output must carry message text");
+
+    assert!(
+        !message_text.is_empty(),
+        "responses message text must not be empty"
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/openai_responses_rejects_a_tool_that_requires_an_undeclared_property.rs b/paddler_tests/tests/openai_responses_rejects_a_tool_that_requires_an_undeclared_property.rs
index 50e560300..45062c0cd 100644
--- a/paddler_tests/tests/openai_responses_rejects_a_tool_that_requires_an_undeclared_property.rs
+++ b/paddler_tests/tests/openai_responses_rejects_a_tool_that_requires_an_undeclared_property.rs
@@ -9,13 +9,9 @@ use paddler_tests::start_cluster::start_cluster;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn openai_responses_rejects_a_tool_that_requires_an_undeclared_property() {
-    let cluster = start_cluster(ClusterParams {
-        agents: Vec::new(),
-        wait_for_slots_ready: false,
-        ..ClusterParams::default()
-    })
-    .await
-    .expect("a balancer without agents must start");
+    let cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("a balancer without agents must start");
 
     let rejection = cluster
         .openai_responses_non_streaming(&json!({
diff --git a/paddler_tests/tests/openai_responses_streaming_ends_at_a_tool_call_violating_its_schema.rs b/paddler_tests/tests/openai_responses_streaming_ends_at_a_tool_call_violating_its_schema.rs
new file mode 100644
index 000000000..e9fb2f9d8
--- /dev/null
+++ b/paddler_tests/tests/openai_responses_streaming_ends_at_a_tool_call_violating_its_schema.rs
@@ -0,0 +1,46 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use serde_json::Value;
+use serde_json::json;
+
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
+use paddler_tests::weather_tool_parameters_with_an_unsatisfiable_location::weather_tool_parameters_with_an_unsatisfiable_location;
+use paddler_tests::weather_tool_question::WEATHER_TOOL_QUESTION;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn openai_responses_streaming_ends_at_a_tool_call_violating_its_schema() {
+    let cluster = start_cluster_with_qwen3(vec![AgentConfig::single(1)])
+        .await
+        .expect("the cluster must start");
+
+    let events = cluster
+        .openai_responses_streaming(&json!({
+            "model": "qwen3-test",
+            "input": WEATHER_TOOL_QUESTION,
+            "max_output_tokens": 400,
+            "stream": true,
+            "tools": [{
+                "type": "function",
+                "name": "get_weather",
+                "description": "Get the current weather for a location",
+                "parameters": weather_tool_parameters_with_an_unsatisfiable_location()
+            }]
+        }))
+        .await
+        .expect("the OpenAI responses stream must be readable");
+
+    assert_eq!(
+        events
+            .iter()
+            .skip_while(|event| event["type"] != "response.failed")
+            .map(|event| event["type"].clone())
+            .collect::<Vec<Value>>(),
+        vec![json!("response.failed")]
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/openai_responses_streaming_rejects_a_tool_that_requires_an_undeclared_property.rs b/paddler_tests/tests/openai_responses_streaming_rejects_a_tool_that_requires_an_undeclared_property.rs
new file mode 100644
index 000000000..7181b1bd4
--- /dev/null
+++ b/paddler_tests/tests/openai_responses_streaming_rejects_a_tool_that_requires_an_undeclared_property.rs
@@ -0,0 +1,52 @@
+use async_openai::error::ApiErrorResponse;
+use async_openai::error::OpenAIError;
+use http::StatusCode;
+use serde_json::json;
+
+use paddler_test_cluster_harness::cluster_harness_error::ClusterHarnessError;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_tests::start_cluster::start_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn openai_responses_streaming_rejects_a_tool_that_requires_an_undeclared_property() {
+    let cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("a balancer without agents must start");
+
+    let rejection = cluster
+        .openai_responses_streaming(&json!({
+            "model": "test-model",
+            "input": "hi",
+            "stream": true,
+            "tools": [
+                {
+                    "type": "function",
+                    "name": "broken",
+                    "parameters": {
+                        "type": "object",
+                        "properties": { "present": { "type": "string" } },
+                        "required": ["absent"]
+                    }
+                }
+            ]
+        }))
+        .await;
+
+    assert!(
+        matches!(
+            &rejection,
+            Err(ClusterHarnessError::OpenAIRequestFailed(OpenAIError::ApiError(ApiErrorResponse {
+                status_code,
+                api_error,
+            }))) if *status_code == StatusCode::BAD_REQUEST
+                && api_error.r#type.as_deref() == Some("invalid_request_error")
+                && api_error.message == "Required field 'absent' not found in properties"
+        ),
+        "{rejection:?}"
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/openai_responses_streaming_reports_a_buffered_request_timeout_as_a_failed_response.rs b/paddler_tests/tests/openai_responses_streaming_reports_a_buffered_request_timeout_as_a_failed_response.rs
new file mode 100644
index 000000000..0dae28240
--- /dev/null
+++ b/paddler_tests/tests/openai_responses_streaming_reports_a_buffered_request_timeout_as_a_failed_response.rs
@@ -0,0 +1,48 @@
+use serde_json::json;
+
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_tests::short_buffered_request_timeout::SHORT_BUFFERED_REQUEST_TIMEOUT;
+use paddler_tests::start_cluster::start_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn openai_responses_streaming_reports_a_buffered_request_timeout_as_a_failed_response() {
+    let cluster = start_cluster(ClusterParams {
+        buffered_request_timeout: SHORT_BUFFERED_REQUEST_TIMEOUT,
+        ..ClusterParams::without_agents()
+    })
+    .await
+    .expect("a balancer without agents must start");
+
+    let events = cluster
+        .openai_responses_streaming(&json!({
+            "model": "test",
+            "input": "hi",
+            "stream": true,
+        }))
+        .await
+        .expect("the OpenAI responses stream must be readable");
+
+    assert_eq!(
+        events
+            .iter()
+            .map(|event| event["type"].clone())
+            .collect::<Vec<_>>(),
+        vec![
+            json!("response.created"),
+            json!("response.in_progress"),
+            json!("response.failed")
+        ]
+    );
+    assert_eq!(
+        events[2]["response"]["error"],
+        json!({
+            "code": "server_error",
+            "message": "Waiting for available slot timed out"
+        })
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/openai_responses_text_format_json_schema_constrains_the_output.rs b/paddler_tests/tests/openai_responses_text_format_json_schema_constrains_the_output.rs
new file mode 100644
index 000000000..110299dab
--- /dev/null
+++ b/paddler_tests/tests/openai_responses_text_format_json_schema_constrains_the_output.rs
@@ -0,0 +1,56 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use serde_json::Value;
+use serde_json::from_str;
+use serde_json::json;
+
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_tests::responses_message_text::responses_message_text;
+use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn openai_responses_text_format_json_schema_constrains_the_output() {
+    let cluster = start_cluster_with_qwen3(vec![AgentConfig::single(1)])
+        .await
+        .expect("the cluster must start");
+
+    let response = cluster
+        .openai_responses_non_streaming(&json!({
+            "model": "qwen3-test",
+            "input": "What is 2+2?",
+            "max_output_tokens": 50,
+            "reasoning": { "effort": "none" },
+            "text": {
+                "format": {
+                    "type": "json_schema",
+                    "name": "sum",
+                    "schema": {
+                        "type": "object",
+                        "properties": { "answer": { "type": "integer" } },
+                        "required": ["answer"],
+                        "additionalProperties": false
+                    }
+                }
+            }
+        }))
+        .await
+        .expect("the OpenAI response must succeed");
+
+    let answer: Value = from_str(
+        responses_message_text(&response).expect("the Responses output must carry message text"),
+    )
+    .expect("the output must be JSON");
+
+    assert_eq!(
+        answer
+            .as_object()
+            .map(|answer_fields| answer_fields.keys().map(String::as_str).collect::<Vec<_>>()),
+        Some(vec!["answer"])
+    );
+    assert!(answer["answer"].is_i64(), "{answer}");
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/openai_service_lets_browsers_send_and_read_openai_headers.rs b/paddler_tests/tests/openai_service_lets_browsers_send_and_read_openai_headers.rs
new file mode 100644
index 000000000..57502a787
--- /dev/null
+++ b/paddler_tests/tests/openai_service_lets_browsers_send_and_read_openai_headers.rs
@@ -0,0 +1,69 @@
+use reqwest::Client;
+use reqwest::Method;
+use reqwest::StatusCode;
+use reqwest::header::HeaderValue;
+
+use paddler_balancer::compatibility::openai_service::openai_api_path::OpenAIApiPath;
+use paddler_balancer::compatibility::openai_service::openai_header::OpenAIHeader;
+use paddler_messaging::api_path::ApiPath;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_tests::start_cluster::start_cluster;
+
+const ALLOWED_ORIGIN: &str = "http://example.com";
+
+#[tokio::test(flavor = "multi_thread")]
+async fn openai_service_lets_browsers_send_and_read_openai_headers() {
+    let cluster = start_cluster(ClusterParams {
+        inference_cors_allowed_hosts: vec![ALLOWED_ORIGIN.to_owned()],
+        ..ClusterParams::without_agents()
+    })
+    .await
+    .expect("a balancer without agents must start");
+
+    let preflight = Client::new()
+        .request(
+            Method::OPTIONS,
+            cluster
+                .balancer
+                .compat_openai_url(OpenAIApiPath::CHAT_COMPLETIONS)
+                .expect("the cluster serves OpenAI"),
+        )
+        .header("Origin", ALLOWED_ORIGIN)
+        .header("Access-Control-Request-Method", "POST")
+        .header(
+            "Access-Control-Request-Headers",
+            ["authorization", "content-type", OpenAIHeader::REQUEST_ID].join(", "),
+        )
+        .send()
+        .await
+        .expect("the preflight request must be answered");
+
+    assert_eq!(preflight.status(), StatusCode::OK);
+
+    let health = Client::new()
+        .get(
+            cluster
+                .balancer
+                .compat_openai_url(ApiPath::HEALTH)
+                .expect("the cluster serves OpenAI"),
+        )
+        .header("Origin", ALLOWED_ORIGIN)
+        .header(OpenAIHeader::REQUEST_ID, "browser-request")
+        .send()
+        .await
+        .expect("the health request must be answered");
+
+    assert_eq!(
+        health.headers().get("access-control-expose-headers"),
+        Some(&HeaderValue::from_static(OpenAIHeader::REQUEST_ID))
+    );
+    assert_eq!(
+        health.headers().get(OpenAIHeader::REQUEST_ID),
+        Some(&HeaderValue::from_static("browser-request"))
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/production_sources_keep_test_code_in_inline_test_modules.rs b/paddler_tests/tests/production_sources_keep_test_code_in_inline_test_modules.rs
new file mode 100644
index 000000000..8412c8f41
--- /dev/null
+++ b/paddler_tests/tests/production_sources_keep_test_code_in_inline_test_modules.rs
@@ -0,0 +1,57 @@
+use std::fs::read_dir;
+use std::fs::read_to_string;
+use std::path::Path;
+use std::path::PathBuf;
+
+use syn::parse_file;
+use syn::visit::Visit as _;
+
+use crate::test_only_attribute_counter::TestOnlyAttributeCounter;
+
+#[test]
+fn production_sources_keep_test_code_in_inline_test_modules() {
+    let workspace_root = Path::new(env!("CARGO_MANIFEST_DIR"))
+        .parent()
+        .expect("the test crate must live inside the workspace");
+    let mut source_directories: Vec<PathBuf> = read_dir(workspace_root)
+        .expect("the workspace root must be listable")
+        .map(|entry| entry.expect("a workspace entry must be readable").path())
+        .filter(|crate_directory| crate_directory.join("Cargo.toml").is_file())
+        .map(|crate_directory| crate_directory.join("src"))
+        .filter(|source_directory| source_directory.is_dir())
+        .collect();
+    let mut rust_sources = Vec::new();
+
+    while let Some(source_directory) = source_directories.pop() {
+        let source_entries: Vec<PathBuf> = read_dir(source_directory)
+            .expect("a source directory must be listable")
+            .map(|entry| entry.expect("a source entry must be readable").path())
+            .collect();
+
+        source_directories.extend(source_entries.iter().filter(|path| path.is_dir()).cloned());
+        rust_sources.extend(
+            source_entries
+                .into_iter()
+                .filter(|path| path.extension().is_some_and(|extension| extension == "rs")),
+        );
+    }
+
+    let sources_with_test_only_code_outside_inline_test_modules: Vec<PathBuf> = rust_sources
+        .into_iter()
+        .filter(|rust_source| {
+            let mut test_only_attribute_counter = TestOnlyAttributeCounter::default();
+
+            test_only_attribute_counter.visit_file(
+                &parse_file(&read_to_string(rust_source).expect("a source file must be readable"))
+                    .expect("a source file must parse"),
+            );
+
+            test_only_attribute_counter.test_only_attributes > 0
+        })
+        .collect();
+
+    assert_eq!(
+        sources_with_test_only_code_outside_inline_test_modules,
+        Vec::<PathBuf>::new()
+    );
+}
diff --git a/paddler_tests/tests/qwen25vl_generates_tokens_from_image_input.rs b/paddler_tests/tests/qwen25vl_generates_tokens_from_image_input.rs
index c334f7684..9585c9537 100644
--- a/paddler_tests/tests/qwen25vl_generates_tokens_from_image_input.rs
+++ b/paddler_tests/tests/qwen25vl_generates_tokens_from_image_input.rs
@@ -8,12 +8,10 @@ use paddler_messaging::conversation_history::ConversationHistory;
 use paddler_messaging::conversation_message::ConversationMessage;
 use paddler_messaging::conversation_message_content::ConversationMessageContent;
 use paddler_messaging::conversation_message_content_part::ConversationMessageContentPart;
-use paddler_messaging::generated_token_result::GeneratedTokenResult;
 use paddler_messaging::image_url::ImageUrl;
 use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
 use paddler_test_cluster_harness::load_test_image_data_uri::load_test_image_data_uri;
-use paddler_test_cluster_harness::token_result_with_producer::TokenResultWithProducer;
 use paddler_tests::start_cluster_with_qwen2_5_vl::start_cluster_with_qwen2_5_vl;
 
 #[tokio::test(flavor = "multi_thread")]
@@ -24,7 +22,7 @@ async fn qwen25vl_generates_tokens_from_image_input() {
 
     let image_data_uri = load_test_image_data_uri().expect("the test image must load");
 
-    let conversation_history = ConversationHistory::new(vec![ConversationMessage {
+    let conversation_history = ConversationHistory::new(vec![ConversationMessage::Message {
         content: ConversationMessageContent::Parts(vec![
             ConversationMessageContentPart::ImageUrl {
                 image_url: ImageUrl {
@@ -61,13 +59,9 @@ async fn qwen25vl_generates_tokens_from_image_input() {
         .count();
 
     assert!(token_count > 0);
-    assert!(matches!(
-        collected.token_results.last(),
-        Some(TokenResultWithProducer {
-            token_result: GeneratedTokenResult::Done(_),
-            ..
-        })
-    ));
+    collected
+        .summary()
+        .expect("the generation must finish with a summary");
 
     cluster
         .shutdown()
diff --git a/paddler_tests/tests/qwen35_generates_tokens_for_long_system_and_user_prompt.rs b/paddler_tests/tests/qwen35_generates_tokens_for_long_system_and_user_prompt.rs
index f0f24076d..69a00f9a9 100644
--- a/paddler_tests/tests/qwen35_generates_tokens_for_long_system_and_user_prompt.rs
+++ b/paddler_tests/tests/qwen35_generates_tokens_for_long_system_and_user_prompt.rs
@@ -7,10 +7,8 @@ use tokio_util::sync::CancellationToken;
 use paddler_messaging::conversation_history::ConversationHistory;
 use paddler_messaging::conversation_message::ConversationMessage;
 use paddler_messaging::conversation_message_content::ConversationMessageContent;
-use paddler_messaging::generated_token_result::GeneratedTokenResult;
 use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
-use paddler_test_cluster_harness::token_result_with_producer::TokenResultWithProducer;
 use paddler_tests::start_cluster_with_qwen3_5::start_cluster_with_qwen3_5;
 
 fn build_long_link_list() -> String {
@@ -64,11 +62,11 @@ async fn qwen35_generates_tokens_for_long_system_and_user_prompt() {
     );
 
     let conversation_history = ConversationHistory::new(vec![
-        ConversationMessage {
+        ConversationMessage::Message {
             content: ConversationMessageContent::Text(system_prompt.to_owned()),
             role: "system".to_owned(),
         },
-        ConversationMessage {
+        ConversationMessage::Message {
             content: ConversationMessageContent::Text(user_prompt),
             role: "user".to_owned(),
         },
@@ -97,13 +95,9 @@ async fn qwen35_generates_tokens_for_long_system_and_user_prompt() {
         .count();
 
     assert!(token_count > 0);
-    assert!(matches!(
-        collected.token_results.last(),
-        Some(TokenResultWithProducer {
-            token_result: GeneratedTokenResult::Done(_),
-            ..
-        })
-    ));
+    collected
+        .summary()
+        .expect("the generation must finish with a summary");
 
     cluster
         .shutdown()
diff --git a/paddler_tests/tests/qwen35_generation_stops_at_eog_before_max_tokens.rs b/paddler_tests/tests/qwen35_generation_stops_at_eog_before_max_tokens.rs
index a9c95e3aa..606ab9d87 100644
--- a/paddler_tests/tests/qwen35_generation_stops_at_eog_before_max_tokens.rs
+++ b/paddler_tests/tests/qwen35_generation_stops_at_eog_before_max_tokens.rs
@@ -7,10 +7,9 @@ use tokio_util::sync::CancellationToken;
 use paddler_messaging::conversation_history::ConversationHistory;
 use paddler_messaging::conversation_message::ConversationMessage;
 use paddler_messaging::conversation_message_content::ConversationMessageContent;
-use paddler_messaging::generated_token_result::GeneratedTokenResult;
+use paddler_messaging::generation_finish::GenerationFinish;
 use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
-use paddler_test_cluster_harness::token_result_with_producer::TokenResultWithProducer;
 use paddler_tests::start_cluster_with_qwen3_5::start_cluster_with_qwen3_5;
 
 #[tokio::test(flavor = "multi_thread")]
@@ -24,10 +23,12 @@ async fn qwen35_generation_stops_at_eog_before_max_tokens() {
             CancellationToken::new(),
             &ContinueFromConversationHistoryParams {
                 add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Text("hi".to_owned()),
-                    role: "user".to_owned(),
-                }]),
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Text("hi".to_owned()),
+                        role: "user".to_owned(),
+                    },
+                ]),
                 enable_thinking: false,
                 grammar: None,
                 max_tokens: NonZeroU32::new(500).unwrap(),
@@ -49,13 +50,13 @@ async fn qwen35_generation_stops_at_eog_before_max_tokens() {
         token_count < 500,
         "expected EOG to stop generation before max_tokens"
     );
-    assert!(matches!(
-        collected.token_results.last(),
-        Some(TokenResultWithProducer {
-            token_result: GeneratedTokenResult::Done(_),
-            ..
-        })
-    ));
+    assert_eq!(
+        collected
+            .summary()
+            .expect("the generation must finish with a summary")
+            .finish,
+        GenerationFinish::EndOfGeneration
+    );
 
     cluster
         .shutdown()
diff --git a/paddler_tests/tests/qwen35_internal_endpoint_emits_reasoning_tokens_for_image_request.rs b/paddler_tests/tests/qwen35_internal_endpoint_emits_reasoning_tokens_for_image_request.rs
index 02123058e..2dfb58d19 100644
--- a/paddler_tests/tests/qwen35_internal_endpoint_emits_reasoning_tokens_for_image_request.rs
+++ b/paddler_tests/tests/qwen35_internal_endpoint_emits_reasoning_tokens_for_image_request.rs
@@ -23,7 +23,7 @@ async fn qwen35_internal_endpoint_emits_reasoning_tokens_for_image_request() {
 
     let image_data_uri = load_test_image_data_uri().expect("the test image must load");
 
-    let conversation_history = ConversationHistory::new(vec![ConversationMessage {
+    let conversation_history = ConversationHistory::new(vec![ConversationMessage::Message {
         content: ConversationMessageContent::Parts(vec![
             ConversationMessageContentPart::ImageUrl {
                 image_url: ImageUrl {
@@ -64,13 +64,9 @@ async fn qwen35_internal_endpoint_emits_reasoning_tokens_for_image_request() {
         "Qwen 3.5: expected at least one reasoning token from a `<think>` block when an image is attached (got {reasoning_count})"
     );
 
-    let last = collected
-        .token_results
-        .last()
-        .expect("no token results received");
-    let GeneratedTokenResult::Done(summary) = &last.token_result else {
-        panic!("last result was not Done: {last:?}");
-    };
+    let summary = collected
+        .summary()
+        .expect("the generation must finish with a summary");
 
     assert!(summary.usage.reasoning_tokens > 0);
     assert!(summary.usage.input_image_tokens > 0);
diff --git a/paddler_tests/tests/qwen35_internal_endpoint_emits_tool_call_parsed_event.rs b/paddler_tests/tests/qwen35_internal_endpoint_emits_tool_call_parsed_event.rs
index eed92efe6..63af63ca4 100644
--- a/paddler_tests/tests/qwen35_internal_endpoint_emits_tool_call_parsed_event.rs
+++ b/paddler_tests/tests/qwen35_internal_endpoint_emits_tool_call_parsed_event.rs
@@ -2,17 +2,18 @@
 
 use std::num::NonZeroU32;
 
-use llama_cpp_bindings_types::ParsedToolCall;
 use tokio_util::sync::CancellationToken;
 
 use paddler_messaging::conversation_history::ConversationHistory;
 use paddler_messaging::conversation_message::ConversationMessage;
 use paddler_messaging::conversation_message_content::ConversationMessageContent;
 use paddler_messaging::generated_token_result::GeneratedTokenResult;
+use paddler_messaging::generated_tool_call::GeneratedToolCall;
 use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
 use paddler_tests::get_weather_tool::get_weather_tool;
 use paddler_tests::start_cluster_with_qwen3_5::start_cluster_with_qwen3_5;
+use paddler_tests::weather_tool_question::WEATHER_TOOL_QUESTION;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn qwen35_internal_endpoint_emits_tool_call_parsed_event() {
@@ -25,13 +26,12 @@ async fn qwen35_internal_endpoint_emits_tool_call_parsed_event() {
             CancellationToken::new(),
             &ContinueFromConversationHistoryParams {
                 add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Text(
-                        "What is the weather in Paris? Use the get_weather tool to find out."
-                            .to_owned(),
-                    ),
-                    role: "user".to_owned(),
-                }]),
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Text(WEATHER_TOOL_QUESTION.to_owned()),
+                        role: "user".to_owned(),
+                    },
+                ]),
                 enable_thinking: false,
                 grammar: None,
                 max_tokens: NonZeroU32::new(400).unwrap(),
@@ -42,7 +42,7 @@ async fn qwen35_internal_endpoint_emits_tool_call_parsed_event() {
         .await
         .expect("the inference request must be accepted");
 
-    let parsed_events: Vec<&Vec<ParsedToolCall>> = collected
+    let parsed_events: Vec<&Vec<GeneratedToolCall>> = collected
         .token_results
         .iter()
         .filter_map(|event| match &event.token_result {
diff --git a/paddler_tests/tests/qwen35_internal_endpoint_with_thinking_disabled_emits_only_content_tokens.rs b/paddler_tests/tests/qwen35_internal_endpoint_with_thinking_disabled_emits_only_content_tokens.rs
index e8df07acf..d0e92c269 100644
--- a/paddler_tests/tests/qwen35_internal_endpoint_with_thinking_disabled_emits_only_content_tokens.rs
+++ b/paddler_tests/tests/qwen35_internal_endpoint_with_thinking_disabled_emits_only_content_tokens.rs
@@ -23,10 +23,14 @@ async fn qwen35_internal_endpoint_with_thinking_disabled_emits_only_content_toke
             CancellationToken::new(),
             &ContinueFromConversationHistoryParams {
                 add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Text("What is two plus two?".to_owned()),
-                    role: "user".to_owned(),
-                }]),
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Text(
+                            "What is two plus two?".to_owned(),
+                        ),
+                        role: "user".to_owned(),
+                    },
+                ]),
                 enable_thinking: false,
                 grammar: None,
                 max_tokens: NonZeroU32::new(200).unwrap(),
@@ -76,13 +80,9 @@ async fn qwen35_internal_endpoint_with_thinking_disabled_emits_only_content_toke
         "Qwen3.5 thinking-disabled: classifier must stream at least one content token"
     );
 
-    let last = collected
-        .token_results
-        .last()
-        .expect("no token results received");
-    let GeneratedTokenResult::Done(summary) = &last.token_result else {
-        panic!("last result was not Done: {last:?}");
-    };
+    let summary = collected
+        .summary()
+        .expect("the generation must finish with a summary");
 
     assert!(summary.usage.prompt_tokens > 0);
     assert_eq!(
@@ -102,22 +102,8 @@ async fn qwen35_internal_endpoint_with_thinking_disabled_emits_only_content_toke
          reasoning and undeterminable are zero"
     );
 
-    let reasoning_stream: String = collected
-        .token_results
-        .iter()
-        .filter_map(|result| match &result.token_result {
-            GeneratedTokenResult::ReasoningToken(piece) => Some(piece.as_str()),
-            _ => None,
-        })
-        .collect();
-    let content_stream: String = collected
-        .token_results
-        .iter()
-        .filter_map(|result| match &result.token_result {
-            GeneratedTokenResult::ContentToken(piece) => Some(piece.as_str()),
-            _ => None,
-        })
-        .collect();
+    let reasoning_stream = collected.reasoning_text();
+    let content_stream = collected.content_text();
 
     for forbidden in ["<think>", "</think>"] {
         assert!(
diff --git a/paddler_tests/tests/qwen35_internal_endpoint_with_thinking_enabled_emits_reasoning_tokens.rs b/paddler_tests/tests/qwen35_internal_endpoint_with_thinking_enabled_emits_reasoning_tokens.rs
index 935861f69..722a10fae 100644
--- a/paddler_tests/tests/qwen35_internal_endpoint_with_thinking_enabled_emits_reasoning_tokens.rs
+++ b/paddler_tests/tests/qwen35_internal_endpoint_with_thinking_enabled_emits_reasoning_tokens.rs
@@ -23,12 +23,14 @@ async fn qwen35_internal_endpoint_with_thinking_enabled_emits_reasoning_tokens()
             CancellationToken::new(),
             &ContinueFromConversationHistoryParams {
                 add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Text(
-                        "What is two plus two? Think step by step.".to_owned(),
-                    ),
-                    role: "user".to_owned(),
-                }]),
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Text(
+                            "What is two plus two? Think step by step.".to_owned(),
+                        ),
+                        role: "user".to_owned(),
+                    },
+                ]),
                 enable_thinking: true,
                 grammar: None,
                 max_tokens: NonZeroU32::new(600).unwrap(),
@@ -50,13 +52,9 @@ async fn qwen35_internal_endpoint_with_thinking_enabled_emits_reasoning_tokens()
         "Qwen3.5: expected at least one reasoning token when thinking is enabled (got {reasoning_count})"
     );
 
-    let last = collected
-        .token_results
-        .last()
-        .expect("no token results received");
-    let GeneratedTokenResult::Done(summary) = &last.token_result else {
-        panic!("last result was not Done: {last:?}");
-    };
+    let summary = collected
+        .summary()
+        .expect("the generation must finish with a summary");
 
     assert!(summary.usage.prompt_tokens > 0);
     assert!(summary.usage.reasoning_tokens > 0);
@@ -67,22 +65,8 @@ async fn qwen35_internal_endpoint_with_thinking_enabled_emits_reasoning_tokens()
             + summary.usage.undeterminable_tokens
     );
 
-    let reasoning_stream: String = collected
-        .token_results
-        .iter()
-        .filter_map(|result| match &result.token_result {
-            GeneratedTokenResult::ReasoningToken(piece) => Some(piece.as_str()),
-            _ => None,
-        })
-        .collect();
-    let content_stream: String = collected
-        .token_results
-        .iter()
-        .filter_map(|result| match &result.token_result {
-            GeneratedTokenResult::ContentToken(piece) => Some(piece.as_str()),
-            _ => None,
-        })
-        .collect();
+    let reasoning_stream = collected.reasoning_text();
+    let content_stream = collected.content_text();
 
     for forbidden in ["<think>", "</think>"] {
         assert!(
diff --git a/paddler_tests/tests/qwen35_thinking_mode_stops_cleanly_before_max_tokens.rs b/paddler_tests/tests/qwen35_thinking_mode_stops_cleanly_before_max_tokens.rs
index f9d40b0d8..1e6830a09 100644
--- a/paddler_tests/tests/qwen35_thinking_mode_stops_cleanly_before_max_tokens.rs
+++ b/paddler_tests/tests/qwen35_thinking_mode_stops_cleanly_before_max_tokens.rs
@@ -7,10 +7,8 @@ use tokio_util::sync::CancellationToken;
 use paddler_messaging::conversation_history::ConversationHistory;
 use paddler_messaging::conversation_message::ConversationMessage;
 use paddler_messaging::conversation_message_content::ConversationMessageContent;
-use paddler_messaging::generated_token_result::GeneratedTokenResult;
 use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
-use paddler_test_cluster_harness::token_result_with_producer::TokenResultWithProducer;
 use paddler_tests::start_cluster_with_qwen3_5::start_cluster_with_qwen3_5;
 
 #[tokio::test(flavor = "multi_thread")]
@@ -24,10 +22,12 @@ async fn qwen35_thinking_mode_stops_cleanly_before_max_tokens() {
             CancellationToken::new(),
             &ContinueFromConversationHistoryParams {
                 add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Text("What is 2+2?".to_owned()),
-                    role: "user".to_owned(),
-                }]),
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Text("What is 2+2?".to_owned()),
+                        role: "user".to_owned(),
+                    },
+                ]),
                 enable_thinking: true,
                 grammar: None,
                 max_tokens: NonZeroU32::new(2000).unwrap(),
@@ -46,13 +46,9 @@ async fn qwen35_thinking_mode_stops_cleanly_before_max_tokens() {
 
     assert!(token_count > 0);
     assert!(token_count <= 2000);
-    assert!(matches!(
-        collected.token_results.last(),
-        Some(TokenResultWithProducer {
-            token_result: GeneratedTokenResult::Done(_),
-            ..
-        })
-    ));
+    collected
+        .summary()
+        .expect("the generation must finish with a summary");
 
     cluster
         .shutdown()
diff --git a/paddler_tests/tests/qwen35_thinking_multi_turn_conversation_stops_cleanly.rs b/paddler_tests/tests/qwen35_thinking_multi_turn_conversation_stops_cleanly.rs
index e1a7fe893..ccfc4110e 100644
--- a/paddler_tests/tests/qwen35_thinking_multi_turn_conversation_stops_cleanly.rs
+++ b/paddler_tests/tests/qwen35_thinking_multi_turn_conversation_stops_cleanly.rs
@@ -7,10 +7,8 @@ use tokio_util::sync::CancellationToken;
 use paddler_messaging::conversation_history::ConversationHistory;
 use paddler_messaging::conversation_message::ConversationMessage;
 use paddler_messaging::conversation_message_content::ConversationMessageContent;
-use paddler_messaging::generated_token_result::GeneratedTokenResult;
 use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
-use paddler_test_cluster_harness::token_result_with_producer::TokenResultWithProducer;
 use paddler_tests::start_cluster_with_qwen3_5::start_cluster_with_qwen3_5;
 
 #[tokio::test(flavor = "multi_thread")]
@@ -20,19 +18,19 @@ async fn qwen35_thinking_multi_turn_conversation_stops_cleanly() {
         .expect("the cluster must start");
 
     let conversation_history = ConversationHistory::new(vec![
-        ConversationMessage {
+        ConversationMessage::Message {
             content: ConversationMessageContent::Text(
                 "You are a helpful assistant. Give engaging, short, precise answers. Be friendly, supportive, use emojis.".to_owned(),
             ),
             role: "system".to_owned(),
         },
-        ConversationMessage {
+        ConversationMessage::Message {
             content: ConversationMessageContent::Text(
                 "Hello! How can I help you today?".to_owned(),
             ),
             role: "assistant".to_owned(),
         },
-        ConversationMessage {
+        ConversationMessage::Message {
             content: ConversationMessageContent::Text("hi".to_owned()),
             role: "user".to_owned(),
         },
@@ -62,13 +60,9 @@ async fn qwen35_thinking_multi_turn_conversation_stops_cleanly() {
 
     assert!(token_count > 0);
     assert!(token_count <= 1000);
-    assert!(matches!(
-        collected.token_results.last(),
-        Some(TokenResultWithProducer {
-            token_result: GeneratedTokenResult::Done(_),
-            ..
-        })
-    ));
+    collected
+        .summary()
+        .expect("the generation must finish with a summary");
 
     cluster
         .shutdown()
diff --git a/paddler_tests/tests/qwen35_with_mmproj_generates_tokens_from_image.rs b/paddler_tests/tests/qwen35_with_mmproj_generates_tokens_from_image.rs
index 00e731db2..cbf79bd68 100644
--- a/paddler_tests/tests/qwen35_with_mmproj_generates_tokens_from_image.rs
+++ b/paddler_tests/tests/qwen35_with_mmproj_generates_tokens_from_image.rs
@@ -8,12 +8,10 @@ use paddler_messaging::conversation_history::ConversationHistory;
 use paddler_messaging::conversation_message::ConversationMessage;
 use paddler_messaging::conversation_message_content::ConversationMessageContent;
 use paddler_messaging::conversation_message_content_part::ConversationMessageContentPart;
-use paddler_messaging::generated_token_result::GeneratedTokenResult;
 use paddler_messaging::image_url::ImageUrl;
 use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
 use paddler_test_cluster_harness::load_test_image_data_uri::load_test_image_data_uri;
-use paddler_test_cluster_harness::token_result_with_producer::TokenResultWithProducer;
 use paddler_tests::start_cluster_with_qwen3_5::start_cluster_with_qwen3_5;
 
 #[tokio::test(flavor = "multi_thread")]
@@ -24,7 +22,7 @@ async fn qwen35_with_mmproj_generates_tokens_from_image() {
 
     let image_data_uri = load_test_image_data_uri().expect("the test image must load");
 
-    let conversation_history = ConversationHistory::new(vec![ConversationMessage {
+    let conversation_history = ConversationHistory::new(vec![ConversationMessage::Message {
         content: ConversationMessageContent::Parts(vec![
             ConversationMessageContentPart::ImageUrl {
                 image_url: ImageUrl {
@@ -61,13 +59,9 @@ async fn qwen35_with_mmproj_generates_tokens_from_image() {
         .count();
 
     assert!(token_count > 0);
-    assert!(matches!(
-        collected.token_results.last(),
-        Some(TokenResultWithProducer {
-            token_result: GeneratedTokenResult::Done(_),
-            ..
-        })
-    ));
+    collected
+        .summary()
+        .expect("the generation must finish with a summary");
 
     cluster
         .shutdown()
diff --git a/paddler_tests/tests/qwen35_with_system_message_completes_with_thinking.rs b/paddler_tests/tests/qwen35_with_system_message_completes_with_thinking.rs
deleted file mode 100644
index 41ad6030a..000000000
--- a/paddler_tests/tests/qwen35_with_system_message_completes_with_thinking.rs
+++ /dev/null
@@ -1,72 +0,0 @@
-#![cfg(feature = "tests_that_use_llms")]
-
-use std::num::NonZeroU32;
-
-use tokio_util::sync::CancellationToken;
-
-use paddler_messaging::conversation_history::ConversationHistory;
-use paddler_messaging::conversation_message::ConversationMessage;
-use paddler_messaging::conversation_message_content::ConversationMessageContent;
-use paddler_messaging::generated_token_result::GeneratedTokenResult;
-use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
-use paddler_test_cluster_harness::agent_config::AgentConfig;
-use paddler_test_cluster_harness::token_result_with_producer::TokenResultWithProducer;
-use paddler_tests::start_cluster_with_qwen3_5::start_cluster_with_qwen3_5;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn qwen35_with_system_message_completes_with_thinking() {
-    let cluster = start_cluster_with_qwen3_5(vec![AgentConfig::single(1)], false)
-        .await
-        .expect("the cluster must start");
-
-    let conversation_history = ConversationHistory::new(vec![
-        ConversationMessage {
-            content: ConversationMessageContent::Text(
-                "You are a focused web crawler assistant. Your only job is to decide which links to follow to discover more relevant pages. Respond with JSON only.".to_owned(),
-            ),
-            role: "system".to_owned(),
-        },
-        ConversationMessage {
-            content: ConversationMessageContent::Text(
-                "Goal: \"find all PDF reports\"\n\nPage: https://example.com/reports\n\nFollowable links:\n[0] [Navigation] \"Home\" → /home\n[1] [PrimaryListing] \"Annual Report 2024\" → /reports/annual-2024.pdf\n[2] [PrimaryListing] \"Q3 Financial Summary\" → /reports/q3-summary.pdf\n[3] [Navigation] \"Next Page\" → /reports?page=2".to_owned(),
-            ),
-            role: "user".to_owned(),
-        },
-    ]);
-
-    let collected = cluster
-        .continue_from_conversation_history(
-            CancellationToken::new(),
-            &ContinueFromConversationHistoryParams {
-                add_generation_prompt: true,
-                conversation_history,
-                enable_thinking: true,
-                grammar: None,
-                max_tokens: NonZeroU32::new(2000).unwrap(),
-                parse_tool_calls: false,
-                tools: vec![],
-            },
-        )
-        .await
-        .expect("the inference request must be accepted");
-
-    let token_count = collected
-        .token_results
-        .iter()
-        .filter(|result| result.token_result.is_token())
-        .count();
-
-    assert!(token_count > 0);
-    assert!(matches!(
-        collected.token_results.last(),
-        Some(TokenResultWithProducer {
-            token_result: GeneratedTokenResult::Done(_),
-            ..
-        })
-    ));
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_tests/tests/qwen35_with_system_message_completes_with_thinking_disabled.rs b/paddler_tests/tests/qwen35_with_system_message_completes_with_thinking_disabled.rs
new file mode 100644
index 000000000..a4c91f03a
--- /dev/null
+++ b/paddler_tests/tests/qwen35_with_system_message_completes_with_thinking_disabled.rs
@@ -0,0 +1,66 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::num::NonZeroU32;
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::conversation_history::ConversationHistory;
+use paddler_messaging::conversation_message::ConversationMessage;
+use paddler_messaging::conversation_message_content::ConversationMessageContent;
+use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_tests::start_cluster_with_qwen3_5::start_cluster_with_qwen3_5;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn qwen35_with_system_message_completes_with_thinking_disabled() {
+    let cluster = start_cluster_with_qwen3_5(vec![AgentConfig::single(1)], false)
+        .await
+        .expect("the cluster must start");
+
+    let conversation_history = ConversationHistory::new(vec![
+        ConversationMessage::Message {
+            content: ConversationMessageContent::Text(
+                "You are a focused web crawler assistant. Your only job is to decide which links to follow to discover more relevant pages. Respond with JSON only.".to_owned(),
+            ),
+            role: "system".to_owned(),
+        },
+        ConversationMessage::Message {
+            content: ConversationMessageContent::Text(
+                "Goal: \"find all PDF reports\"\n\nPage: https://example.com/reports\n\nFollowable links:\n[0] [Navigation] \"Home\" → /home\n[1] [PrimaryListing] \"Annual Report 2024\" → /reports/annual-2024.pdf\n[2] [PrimaryListing] \"Q3 Financial Summary\" → /reports/q3-summary.pdf\n[3] [Navigation] \"Next Page\" → /reports?page=2".to_owned(),
+            ),
+            role: "user".to_owned(),
+        },
+    ]);
+
+    let collected = cluster
+        .continue_from_conversation_history(
+            CancellationToken::new(),
+            &ContinueFromConversationHistoryParams {
+                add_generation_prompt: true,
+                conversation_history,
+                enable_thinking: false,
+                grammar: None,
+                max_tokens: NonZeroU32::new(512).unwrap(),
+                parse_tool_calls: false,
+                tools: vec![],
+            },
+        )
+        .await
+        .expect("the inference request must be accepted");
+
+    let token_count = collected
+        .token_results
+        .iter()
+        .filter(|result| result.token_result.is_token())
+        .count();
+
+    assert!(token_count > 0);
+    collected
+        .summary()
+        .expect("the generation must finish with a summary");
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/qwen35_with_system_message_completes_with_thinking_enabled.rs b/paddler_tests/tests/qwen35_with_system_message_completes_with_thinking_enabled.rs
new file mode 100644
index 000000000..082f640db
--- /dev/null
+++ b/paddler_tests/tests/qwen35_with_system_message_completes_with_thinking_enabled.rs
@@ -0,0 +1,66 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::num::NonZeroU32;
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::conversation_history::ConversationHistory;
+use paddler_messaging::conversation_message::ConversationMessage;
+use paddler_messaging::conversation_message_content::ConversationMessageContent;
+use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_tests::start_cluster_with_qwen3_5::start_cluster_with_qwen3_5;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn qwen35_with_system_message_completes_with_thinking_enabled() {
+    let cluster = start_cluster_with_qwen3_5(vec![AgentConfig::single(1)], false)
+        .await
+        .expect("the cluster must start");
+
+    let conversation_history = ConversationHistory::new(vec![
+        ConversationMessage::Message {
+            content: ConversationMessageContent::Text(
+                "You are a focused web crawler assistant. Your only job is to decide which links to follow to discover more relevant pages. Respond with JSON only.".to_owned(),
+            ),
+            role: "system".to_owned(),
+        },
+        ConversationMessage::Message {
+            content: ConversationMessageContent::Text(
+                "Goal: \"find all PDF reports\"\n\nPage: https://example.com/reports\n\nFollowable links:\n[0] [Navigation] \"Home\" → /home\n[1] [PrimaryListing] \"Annual Report 2024\" → /reports/annual-2024.pdf\n[2] [PrimaryListing] \"Q3 Financial Summary\" → /reports/q3-summary.pdf\n[3] [Navigation] \"Next Page\" → /reports?page=2".to_owned(),
+            ),
+            role: "user".to_owned(),
+        },
+    ]);
+
+    let collected = cluster
+        .continue_from_conversation_history(
+            CancellationToken::new(),
+            &ContinueFromConversationHistoryParams {
+                add_generation_prompt: true,
+                conversation_history,
+                enable_thinking: true,
+                grammar: None,
+                max_tokens: NonZeroU32::new(2000).unwrap(),
+                parse_tool_calls: false,
+                tools: vec![],
+            },
+        )
+        .await
+        .expect("the inference request must be accepted");
+
+    let token_count = collected
+        .token_results
+        .iter()
+        .filter(|result| result.token_result.is_token())
+        .count();
+
+    assert!(token_count > 0);
+    collected
+        .summary()
+        .expect("the generation must finish with a summary");
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/qwen35_with_system_message_completes_without_thinking.rs b/paddler_tests/tests/qwen35_with_system_message_completes_without_thinking.rs
deleted file mode 100644
index 270a3b57d..000000000
--- a/paddler_tests/tests/qwen35_with_system_message_completes_without_thinking.rs
+++ /dev/null
@@ -1,72 +0,0 @@
-#![cfg(feature = "tests_that_use_llms")]
-
-use std::num::NonZeroU32;
-
-use tokio_util::sync::CancellationToken;
-
-use paddler_messaging::conversation_history::ConversationHistory;
-use paddler_messaging::conversation_message::ConversationMessage;
-use paddler_messaging::conversation_message_content::ConversationMessageContent;
-use paddler_messaging::generated_token_result::GeneratedTokenResult;
-use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
-use paddler_test_cluster_harness::agent_config::AgentConfig;
-use paddler_test_cluster_harness::token_result_with_producer::TokenResultWithProducer;
-use paddler_tests::start_cluster_with_qwen3_5::start_cluster_with_qwen3_5;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn qwen35_with_system_message_completes_without_thinking() {
-    let cluster = start_cluster_with_qwen3_5(vec![AgentConfig::single(1)], false)
-        .await
-        .expect("the cluster must start");
-
-    let conversation_history = ConversationHistory::new(vec![
-        ConversationMessage {
-            content: ConversationMessageContent::Text(
-                "You are a focused web crawler assistant. Your only job is to decide which links to follow to discover more relevant pages. Respond with JSON only.".to_owned(),
-            ),
-            role: "system".to_owned(),
-        },
-        ConversationMessage {
-            content: ConversationMessageContent::Text(
-                "Goal: \"find all PDF reports\"\n\nPage: https://example.com/reports\n\nFollowable links:\n[0] [Navigation] \"Home\" → /home\n[1] [PrimaryListing] \"Annual Report 2024\" → /reports/annual-2024.pdf\n[2] [PrimaryListing] \"Q3 Financial Summary\" → /reports/q3-summary.pdf\n[3] [Navigation] \"Next Page\" → /reports?page=2".to_owned(),
-            ),
-            role: "user".to_owned(),
-        },
-    ]);
-
-    let collected = cluster
-        .continue_from_conversation_history(
-            CancellationToken::new(),
-            &ContinueFromConversationHistoryParams {
-                add_generation_prompt: true,
-                conversation_history,
-                enable_thinking: false,
-                grammar: None,
-                max_tokens: NonZeroU32::new(512).unwrap(),
-                parse_tool_calls: false,
-                tools: vec![],
-            },
-        )
-        .await
-        .expect("the inference request must be accepted");
-
-    let token_count = collected
-        .token_results
-        .iter()
-        .filter(|result| result.token_result.is_token())
-        .count();
-
-    assert!(token_count > 0);
-    assert!(matches!(
-        collected.token_results.last(),
-        Some(TokenResultWithProducer {
-            token_result: GeneratedTokenResult::Done(_),
-            ..
-        })
-    ));
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_tests/tests/qwen35_without_mmproj_rejects_image_with_multimodal_not_supported.rs b/paddler_tests/tests/qwen35_without_mmproj_rejects_image_with_multimodal_not_supported.rs
index 5e086a515..b8b276f9f 100644
--- a/paddler_tests/tests/qwen35_without_mmproj_rejects_image_with_multimodal_not_supported.rs
+++ b/paddler_tests/tests/qwen35_without_mmproj_rejects_image_with_multimodal_not_supported.rs
@@ -23,7 +23,7 @@ async fn qwen35_without_mmproj_rejects_image_with_multimodal_not_supported() {
 
     let image_data_uri = load_test_image_data_uri().expect("the test image must load");
 
-    let conversation_history = ConversationHistory::new(vec![ConversationMessage {
+    let conversation_history = ConversationHistory::new(vec![ConversationMessage::Message {
         content: ConversationMessageContent::Parts(vec![
             ConversationMessageContentPart::ImageUrl {
                 image_url: ImageUrl {
diff --git a/paddler_tests/tests/qwen3_5_answers_every_decision_question_in_order_with_the_expected_input_tokens.rs b/paddler_tests/tests/qwen3_5_answers_every_decision_question_in_order_with_the_expected_input_tokens.rs
new file mode 100644
index 000000000..b376179f4
--- /dev/null
+++ b/paddler_tests/tests/qwen3_5_answers_every_decision_question_in_order_with_the_expected_input_tokens.rs
@@ -0,0 +1,16 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use paddler_tests::decision_cluster_params::DecisionClusterParams;
+
+use crate::sample_decision_answers_check::SampleDecisionAnswersCheck;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn qwen3_5_answers_every_decision_question_in_order_with_the_expected_input_tokens() {
+    SampleDecisionAnswersCheck {
+        cluster_params: DecisionClusterParams::default()
+            .into_cluster_params()
+            .expect("the synthetic pointer head path must be absolute"),
+    }
+    .assert_for_the_sample_decision()
+    .await;
+}
diff --git a/paddler_tests/tests/qwen3_5_decision_text_tokenizes_like_kev.rs b/paddler_tests/tests/qwen3_5_decision_text_tokenizes_like_kev.rs
new file mode 100644
index 000000000..7b3e6969a
--- /dev/null
+++ b/paddler_tests/tests/qwen3_5_decision_text_tokenizes_like_kev.rs
@@ -0,0 +1,14 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use paddler_tests::kev_checkpoint::KevCheckpoint;
+
+use crate::kev_tokenizer_reference_comparison::KevTokenizerReferenceComparison;
+
+#[test]
+fn qwen3_5_decision_text_tokenizes_like_kev() {
+    KevTokenizerReferenceComparison {
+        model_path: KevCheckpoint::Kev0_8b.target_path("model.gguf"),
+        reference_fixture: "qwen3_5_tokenizer_reference.json",
+    }
+    .assert_matched_by_the_decision_text_tokenizer();
+}
diff --git a/paddler_tests/tests/qwen3_8_answers_every_decision_question_in_order_with_the_expected_input_tokens.rs b/paddler_tests/tests/qwen3_8_answers_every_decision_question_in_order_with_the_expected_input_tokens.rs
new file mode 100644
index 000000000..88c81ade8
--- /dev/null
+++ b/paddler_tests/tests/qwen3_8_answers_every_decision_question_in_order_with_the_expected_input_tokens.rs
@@ -0,0 +1,28 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use paddler_test_cluster_harness::fixture_path::fixture_path;
+use paddler_test_cluster_harness::model_card::qwen3_8_27b::qwen3_8_27b;
+use paddler_test_cluster_harness::synthetic_pointer_head_fixture::QWEN3_8_27B_SYNTHETIC_POINTER_HEAD_FIXTURE;
+use paddler_tests::decision_backbone::DecisionBackbone;
+use paddler_tests::decision_cluster_params::DecisionClusterParams;
+use paddler_tests::decision_desired_state::DecisionDesiredState;
+
+use crate::sample_decision_answers_check::SampleDecisionAnswersCheck;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn qwen3_8_answers_every_decision_question_in_order_with_the_expected_input_tokens() {
+    SampleDecisionAnswersCheck {
+        cluster_params: DecisionClusterParams {
+            desired_state: DecisionDesiredState {
+                backbone: DecisionBackbone::HuggingFace(qwen3_8_27b()),
+                pointer_head: fixture_path(QWEN3_8_27B_SYNTHETIC_POINTER_HEAD_FIXTURE),
+                ..DecisionDesiredState::default()
+            },
+            ..DecisionClusterParams::default()
+        }
+        .into_cluster_params()
+        .expect("the synthetic pointer head path must be absolute"),
+    }
+    .assert_for_the_sample_decision()
+    .await;
+}
diff --git a/paddler_tests/tests/qwen3_8_decision_text_tokenizes_like_kev.rs b/paddler_tests/tests/qwen3_8_decision_text_tokenizes_like_kev.rs
new file mode 100644
index 000000000..70f81e80e
--- /dev/null
+++ b/paddler_tests/tests/qwen3_8_decision_text_tokenizes_like_kev.rs
@@ -0,0 +1,16 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use paddler_test_cluster_harness::model_card::qwen3_8_27b::qwen3_8_27b;
+
+use crate::kev_tokenizer_reference_comparison::KevTokenizerReferenceComparison;
+
+#[test]
+fn qwen3_8_decision_text_tokenizes_like_kev() {
+    KevTokenizerReferenceComparison {
+        model_path: qwen3_8_27b()
+            .cached_path()
+            .expect("the Qwen3.8 model card must be in the Hugging Face cache"),
+        reference_fixture: "qwen3_8_tokenizer_reference.json",
+    }
+    .assert_matched_by_the_decision_text_tokenizer();
+}
diff --git a/paddler_tests/tests/qwen3_chat_completion_prompt.rs b/paddler_tests/tests/qwen3_chat_completion_prompt.rs
new file mode 100644
index 000000000..5d0fb838a
--- /dev/null
+++ b/paddler_tests/tests/qwen3_chat_completion_prompt.rs
@@ -0,0 +1,61 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::time::SystemTime;
+
+use serde_json::Value;
+use serde_json::from_value;
+
+use paddler_agent_text_generation::chat_prompt_render_request::ChatPromptRenderRequest;
+use paddler_agent_text_generation::chat_prompt_renderer::ChatPromptRenderer;
+use paddler_agent_text_generation::chat_template_renderer::ChatTemplateRenderer;
+use paddler_messaging::chat_template_content_form::ChatTemplateContentForm;
+use paddler_messaging::media_marker::MediaMarker;
+use paddler_openai_translation::chat_completion_request::ChatCompletionRequest;
+use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
+use paddler_tests::agent_inference_runtime_context::agent_inference_runtime_context;
+use paddler_tests::llama_runtime::LlamaRuntime;
+
+const MEDIA_MARKER: &str = "<media>";
+
+pub fn qwen3_chat_completion_prompt(chat_completion_request: Value) -> String {
+    let runtime = LlamaRuntime::load(
+        &qwen3_0_6b()
+            .cached_path()
+            .expect("the Qwen3 model must be in the Hugging Face cache"),
+        agent_inference_runtime_context(1),
+    )
+    .expect("the Qwen3 model must load");
+    let chat_prompt_renderer = ChatPromptRenderer {
+        chat_template_renderer: ChatTemplateRenderer::new(
+            runtime
+                .loaded_llama_model
+                .model_metadata
+                .embedded_chat_template()
+                .expect("the Qwen3 model must embed its chat template"),
+        )
+        .expect("the Qwen3 chat template must compile"),
+        token_bos_str: String::new(),
+        token_eos_str: String::new(),
+        token_nl_str: String::new(),
+    };
+    let messages = from_value::<ChatCompletionRequest>(chat_completion_request)
+        .expect("the chat completion request must deserialize")
+        .translate(SystemTime::now())
+        .expect("the chat completion request must translate")
+        .conversation_history_params
+        .conversation_history
+        .into_chat_template_conversation(
+            ChatTemplateContentForm::JoinedText,
+            &MediaMarker::new(MEDIA_MARKER.to_owned()),
+        )
+        .messages;
+
+    chat_prompt_renderer
+        .render(ChatPromptRenderRequest {
+            add_generation_prompt: true,
+            enable_thinking: false,
+            messages: &messages,
+            tools: &Value::Array(Vec::new()),
+        })
+        .expect("the Qwen3 chat template must render the conversation")
+}
diff --git a/paddler_tests/tests/qwen3_chat_template_renders_a_tool_call_history_as_tool_call_and_tool_response_blocks.rs b/paddler_tests/tests/qwen3_chat_template_renders_a_tool_call_history_as_tool_call_and_tool_response_blocks.rs
new file mode 100644
index 000000000..650af51f8
--- /dev/null
+++ b/paddler_tests/tests/qwen3_chat_template_renders_a_tool_call_history_as_tool_call_and_tool_response_blocks.rs
@@ -0,0 +1,52 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use serde_json::json;
+
+use crate::qwen3_chat_completion_prompt::qwen3_chat_completion_prompt;
+
+#[test]
+fn qwen3_chat_template_renders_a_tool_call_history_as_tool_call_and_tool_response_blocks() {
+    let rendered_prompt = qwen3_chat_completion_prompt(json!({
+        "model": "qwen3",
+        "messages": [
+            { "role": "user", "content": "What is the weather and the time in Paris?" },
+            {
+                "role": "assistant",
+                "content": null,
+                "tool_calls": [
+                    {
+                        "id": "call_weather",
+                        "type": "function",
+                        "function": { "name": "get_weather", "arguments": "{\"location\":\"Paris\"}" }
+                    },
+                    {
+                        "id": "call_time",
+                        "type": "function",
+                        "function": { "name": "get_time", "arguments": "{\"location\":\"Paris\"}" }
+                    }
+                ]
+            },
+            { "role": "tool", "tool_call_id": "call_weather", "content": "sunny" },
+            { "role": "tool", "tool_call_id": "call_time", "content": "noon" }
+        ]
+    }));
+
+    assert_eq!(
+        rendered_prompt,
+        concat!(
+            "<|im_start|>user\n",
+            "What is the weather and the time in Paris?",
+            "<|im_end|>\n",
+            "<|im_start|>assistant\n",
+            "<tool_call>\n{\"name\": \"get_weather\", \"arguments\": {\"location\": \"Paris\"}}\n</tool_call>\n",
+            "<tool_call>\n{\"name\": \"get_time\", \"arguments\": {\"location\": \"Paris\"}}\n</tool_call>",
+            "<|im_end|>\n",
+            "<|im_start|>user\n",
+            "<tool_response>\nsunny\n</tool_response>\n",
+            "<tool_response>\nnoon\n</tool_response>",
+            "<|im_end|>\n",
+            "<|im_start|>assistant\n",
+            "<think>\n\n</think>\n\n",
+        )
+    );
+}
diff --git a/paddler_tests/tests/qwen3_chat_template_renders_the_reasoning_of_an_assistant_turn_in_a_tool_loop.rs b/paddler_tests/tests/qwen3_chat_template_renders_the_reasoning_of_an_assistant_turn_in_a_tool_loop.rs
new file mode 100644
index 000000000..2de11f08d
--- /dev/null
+++ b/paddler_tests/tests/qwen3_chat_template_renders_the_reasoning_of_an_assistant_turn_in_a_tool_loop.rs
@@ -0,0 +1,46 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use serde_json::json;
+
+use crate::qwen3_chat_completion_prompt::qwen3_chat_completion_prompt;
+
+#[test]
+fn qwen3_chat_template_renders_the_reasoning_of_an_assistant_turn_in_a_tool_loop() {
+    let rendered_prompt = qwen3_chat_completion_prompt(json!({
+        "model": "qwen3",
+        "messages": [
+            { "role": "user", "content": "What is the weather in Paris?" },
+            {
+                "role": "assistant",
+                "content": null,
+                "reasoning_content": "I should look the weather up.",
+                "tool_calls": [
+                    {
+                        "id": "call_weather",
+                        "type": "function",
+                        "function": { "name": "get_weather", "arguments": "{\"location\":\"Paris\"}" }
+                    }
+                ]
+            },
+            { "role": "tool", "tool_call_id": "call_weather", "content": "sunny" }
+        ]
+    }));
+
+    assert_eq!(
+        rendered_prompt,
+        concat!(
+            "<|im_start|>user\n",
+            "What is the weather in Paris?",
+            "<|im_end|>\n",
+            "<|im_start|>assistant\n",
+            "<think>\nI should look the weather up.\n</think>\n\n",
+            "<tool_call>\n{\"name\": \"get_weather\", \"arguments\": {\"location\": \"Paris\"}}\n</tool_call>",
+            "<|im_end|>\n",
+            "<|im_start|>user\n",
+            "<tool_response>\nsunny\n</tool_response>",
+            "<|im_end|>\n",
+            "<|im_start|>assistant\n",
+            "<think>\n\n</think>\n\n",
+        )
+    );
+}
diff --git a/paddler_tests/tests/qwen3_generates_tokens_from_conversation_history.rs b/paddler_tests/tests/qwen3_generates_tokens_from_conversation_history.rs
index 67d291e0e..560b36e7c 100644
--- a/paddler_tests/tests/qwen3_generates_tokens_from_conversation_history.rs
+++ b/paddler_tests/tests/qwen3_generates_tokens_from_conversation_history.rs
@@ -7,10 +7,8 @@ use tokio_util::sync::CancellationToken;
 use paddler_messaging::conversation_history::ConversationHistory;
 use paddler_messaging::conversation_message::ConversationMessage;
 use paddler_messaging::conversation_message_content::ConversationMessageContent;
-use paddler_messaging::generated_token_result::GeneratedTokenResult;
 use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
-use paddler_test_cluster_harness::token_result_with_producer::TokenResultWithProducer;
 use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
 
 #[tokio::test(flavor = "multi_thread")]
@@ -24,10 +22,12 @@ async fn qwen3_generates_tokens_from_conversation_history() {
             CancellationToken::new(),
             &ContinueFromConversationHistoryParams {
                 add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Text("hi".to_owned()),
-                    role: "user".to_owned(),
-                }]),
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Text("hi".to_owned()),
+                        role: "user".to_owned(),
+                    },
+                ]),
                 enable_thinking: false,
                 grammar: None,
                 max_tokens: NonZeroU32::new(500).unwrap(),
@@ -46,13 +46,9 @@ async fn qwen3_generates_tokens_from_conversation_history() {
 
     assert!(token_count > 0);
     assert!(token_count < 500, "EOG should stop generation early");
-    assert!(matches!(
-        collected.token_results.last(),
-        Some(TokenResultWithProducer {
-            token_result: GeneratedTokenResult::Done(_),
-            ..
-        })
-    ));
+    collected
+        .summary()
+        .expect("the generation must finish with a summary");
 
     cluster
         .shutdown()
diff --git a/paddler_tests/tests/qwen3_internal_endpoint_emits_tool_call_parsed_event.rs b/paddler_tests/tests/qwen3_internal_endpoint_emits_tool_call_parsed_event.rs
index 61cdd36a8..f685c57e5 100644
--- a/paddler_tests/tests/qwen3_internal_endpoint_emits_tool_call_parsed_event.rs
+++ b/paddler_tests/tests/qwen3_internal_endpoint_emits_tool_call_parsed_event.rs
@@ -2,17 +2,18 @@
 
 use std::num::NonZeroU32;
 
-use llama_cpp_bindings_types::ParsedToolCall;
 use tokio_util::sync::CancellationToken;
 
 use paddler_messaging::conversation_history::ConversationHistory;
 use paddler_messaging::conversation_message::ConversationMessage;
 use paddler_messaging::conversation_message_content::ConversationMessageContent;
 use paddler_messaging::generated_token_result::GeneratedTokenResult;
+use paddler_messaging::generated_tool_call::GeneratedToolCall;
 use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
 use paddler_tests::get_weather_tool::get_weather_tool;
 use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
+use paddler_tests::weather_tool_question::WEATHER_TOOL_QUESTION;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn qwen3_internal_endpoint_emits_tool_call_parsed_event() {
@@ -25,13 +26,12 @@ async fn qwen3_internal_endpoint_emits_tool_call_parsed_event() {
             CancellationToken::new(),
             &ContinueFromConversationHistoryParams {
                 add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Text(
-                        "What is the weather in Paris? Use the get_weather tool to find out."
-                            .to_owned(),
-                    ),
-                    role: "user".to_owned(),
-                }]),
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Text(WEATHER_TOOL_QUESTION.to_owned()),
+                        role: "user".to_owned(),
+                    },
+                ]),
                 enable_thinking: false,
                 grammar: None,
                 max_tokens: NonZeroU32::new(400).unwrap(),
@@ -42,7 +42,7 @@ async fn qwen3_internal_endpoint_emits_tool_call_parsed_event() {
         .await
         .expect("the inference request must be accepted");
 
-    let parsed_events: Vec<&Vec<ParsedToolCall>> = collected
+    let parsed_events: Vec<&Vec<GeneratedToolCall>> = collected
         .token_results
         .iter()
         .filter_map(|event| match &event.token_result {
diff --git a/paddler_tests/tests/qwen3_internal_endpoint_emits_tool_call_tokens.rs b/paddler_tests/tests/qwen3_internal_endpoint_emits_tool_call_tokens.rs
index dc095cbb3..1310a2756 100644
--- a/paddler_tests/tests/qwen3_internal_endpoint_emits_tool_call_tokens.rs
+++ b/paddler_tests/tests/qwen3_internal_endpoint_emits_tool_call_tokens.rs
@@ -12,6 +12,7 @@ use paddler_messaging::request_params::continue_from_conversation_history_params
 use paddler_test_cluster_harness::agent_config::AgentConfig;
 use paddler_tests::get_weather_tool::get_weather_tool;
 use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
+use paddler_tests::weather_tool_question::WEATHER_TOOL_QUESTION;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn qwen3_internal_endpoint_emits_tool_call_tokens() {
@@ -24,13 +25,12 @@ async fn qwen3_internal_endpoint_emits_tool_call_tokens() {
             CancellationToken::new(),
             &ContinueFromConversationHistoryParams {
                 add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Text(
-                        "What is the weather in Paris? Use the get_weather tool to find out."
-                            .to_owned(),
-                    ),
-                    role: "user".to_owned(),
-                }]),
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Text(WEATHER_TOOL_QUESTION.to_owned()),
+                        role: "user".to_owned(),
+                    },
+                ]),
                 enable_thinking: false,
                 grammar: None,
                 max_tokens: NonZeroU32::new(400).unwrap(),
@@ -52,13 +52,9 @@ async fn qwen3_internal_endpoint_emits_tool_call_tokens() {
         .filter(|result| matches!(result.token_result, GeneratedTokenResult::ContentToken(_)))
         .count();
 
-    let last = collected
-        .token_results
-        .last()
-        .expect("no token results received");
-    let GeneratedTokenResult::Done(summary) = &last.token_result else {
-        panic!("last result was not Done: {last:?}");
-    };
+    let summary = collected
+        .summary()
+        .expect("the generation must finish with a summary");
 
     assert!(summary.usage.prompt_tokens > 0);
     assert!(
diff --git a/paddler_tests/tests/qwen3_internal_endpoint_tools_without_parse_flag_emit_only_raw_tokens.rs b/paddler_tests/tests/qwen3_internal_endpoint_tools_without_parse_flag_emit_only_raw_tokens.rs
index a95375bb3..5d07d1211 100644
--- a/paddler_tests/tests/qwen3_internal_endpoint_tools_without_parse_flag_emit_only_raw_tokens.rs
+++ b/paddler_tests/tests/qwen3_internal_endpoint_tools_without_parse_flag_emit_only_raw_tokens.rs
@@ -12,6 +12,7 @@ use paddler_messaging::request_params::continue_from_conversation_history_params
 use paddler_test_cluster_harness::agent_config::AgentConfig;
 use paddler_tests::get_weather_tool::get_weather_tool;
 use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
+use paddler_tests::weather_tool_question::WEATHER_TOOL_QUESTION;
 
 #[tokio::test(flavor = "multi_thread")]
 async fn qwen3_internal_endpoint_tools_without_parse_flag_emit_only_raw_tokens() {
@@ -24,13 +25,12 @@ async fn qwen3_internal_endpoint_tools_without_parse_flag_emit_only_raw_tokens()
             CancellationToken::new(),
             &ContinueFromConversationHistoryParams {
                 add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Text(
-                        "What is the weather in Paris? Use the get_weather tool to find out."
-                            .to_owned(),
-                    ),
-                    role: "user".to_owned(),
-                }]),
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Text(WEATHER_TOOL_QUESTION.to_owned()),
+                        role: "user".to_owned(),
+                    },
+                ]),
                 enable_thinking: false,
                 grammar: None,
                 max_tokens: NonZeroU32::new(400).unwrap(),
@@ -41,27 +41,24 @@ async fn qwen3_internal_endpoint_tools_without_parse_flag_emit_only_raw_tokens()
         .await
         .expect("the inference request must be accepted");
 
-    for event in &collected.token_results {
-        match &event.token_result {
-            GeneratedTokenResult::ToolCallParsed(_)
-            | GeneratedTokenResult::ToolCallParseFailed(_)
-            | GeneratedTokenResult::ToolSchemaInvalid(_)
-            | GeneratedTokenResult::ToolCallValidationFailed(_) => {
-                panic!(
-                    "expected no parsed/parse-failed/schema-invalid/validation-failed events when parse_tool_calls=false, got: {event:?}"
-                );
-            }
-            _ => {}
-        }
-    }
+    assert_eq!(
+        collected
+            .token_results
+            .iter()
+            .map(|event| &event.token_result)
+            .filter(|token_result| {
+                token_result.is_tool_call_parsed()
+                    || token_result.is_tool_call_failure()
+                    || matches!(token_result, GeneratedTokenResult::ToolSchemaInvalid(_))
+            })
+            .collect::<Vec<_>>(),
+        Vec::<&GeneratedTokenResult>::new(),
+        "expected no parsed/parse-failed/schema-invalid/validation-failed events when parse_tool_calls=false"
+    );
 
-    let last = collected
-        .token_results
-        .last()
-        .expect("no token results received");
-    let GeneratedTokenResult::Done(_) = &last.token_result else {
-        panic!("last result was not Done: {last:?}");
-    };
+    collected
+        .summary()
+        .expect("the generation must finish with a summary");
 
     cluster
         .shutdown()
diff --git a/paddler_tests/tests/qwen3_internal_endpoint_with_thinking_disabled_emits_no_reasoning_tokens.rs b/paddler_tests/tests/qwen3_internal_endpoint_with_thinking_disabled_emits_no_reasoning_tokens.rs
index 9b13989cf..07f53c86a 100644
--- a/paddler_tests/tests/qwen3_internal_endpoint_with_thinking_disabled_emits_no_reasoning_tokens.rs
+++ b/paddler_tests/tests/qwen3_internal_endpoint_with_thinking_disabled_emits_no_reasoning_tokens.rs
@@ -23,10 +23,12 @@ async fn qwen3_internal_endpoint_with_thinking_disabled_emits_no_reasoning_token
             CancellationToken::new(),
             &ContinueFromConversationHistoryParams {
                 add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Text("Say hello.".to_owned()),
-                    role: "user".to_owned(),
-                }]),
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Text("Say hello.".to_owned()),
+                        role: "user".to_owned(),
+                    },
+                ]),
                 enable_thinking: false,
                 grammar: None,
                 max_tokens: NonZeroU32::new(100).unwrap(),
@@ -54,13 +56,9 @@ async fn qwen3_internal_endpoint_with_thinking_disabled_emits_no_reasoning_token
     );
     assert!(content_count > 0, "expected content tokens to be produced");
 
-    let last = collected
-        .token_results
-        .last()
-        .expect("no token results received");
-    let GeneratedTokenResult::Done(summary) = &last.token_result else {
-        panic!("last result was not Done: {last:?}");
-    };
+    let summary = collected
+        .summary()
+        .expect("the generation must finish with a summary");
 
     assert!(summary.usage.prompt_tokens > 0);
     assert_eq!(summary.usage.reasoning_tokens, 0);
diff --git a/paddler_tests/tests/qwen3_internal_endpoint_with_thinking_enabled_emits_reasoning_tokens.rs b/paddler_tests/tests/qwen3_internal_endpoint_with_thinking_enabled_emits_reasoning_tokens.rs
index 158fbb8b6..758863eb1 100644
--- a/paddler_tests/tests/qwen3_internal_endpoint_with_thinking_enabled_emits_reasoning_tokens.rs
+++ b/paddler_tests/tests/qwen3_internal_endpoint_with_thinking_enabled_emits_reasoning_tokens.rs
@@ -23,12 +23,14 @@ async fn qwen3_internal_endpoint_with_thinking_enabled_emits_reasoning_tokens()
             CancellationToken::new(),
             &ContinueFromConversationHistoryParams {
                 add_generation_prompt: true,
-                conversation_history: ConversationHistory::new(vec![ConversationMessage {
-                    content: ConversationMessageContent::Text(
-                        "What is two plus two? Think step by step.".to_owned(),
-                    ),
-                    role: "user".to_owned(),
-                }]),
+                conversation_history: ConversationHistory::new(vec![
+                    ConversationMessage::Message {
+                        content: ConversationMessageContent::Text(
+                            "What is two plus two? Think step by step.".to_owned(),
+                        ),
+                        role: "user".to_owned(),
+                    },
+                ]),
                 enable_thinking: true,
                 grammar: None,
                 max_tokens: NonZeroU32::new(600).unwrap(),
@@ -50,13 +52,9 @@ async fn qwen3_internal_endpoint_with_thinking_enabled_emits_reasoning_tokens()
         "expected at least one reasoning token when thinking is enabled (got {reasoning_count})"
     );
 
-    let last = collected
-        .token_results
-        .last()
-        .expect("no token results received");
-    let GeneratedTokenResult::Done(summary) = &last.token_result else {
-        panic!("last result was not Done: {last:?}");
-    };
+    let summary = collected
+        .summary()
+        .expect("the generation must finish with a summary");
 
     assert!(summary.usage.prompt_tokens > 0);
     assert!(summary.usage.reasoning_tokens > 0);
diff --git a/paddler_tests/tests/qwen3_openai_chat_completion_with_reasoning_effort_none_generates_no_reasoning_tokens.rs b/paddler_tests/tests/qwen3_openai_chat_completion_with_reasoning_effort_none_generates_no_reasoning_tokens.rs
deleted file mode 100644
index f150f9ffe..000000000
--- a/paddler_tests/tests/qwen3_openai_chat_completion_with_reasoning_effort_none_generates_no_reasoning_tokens.rs
+++ /dev/null
@@ -1,36 +0,0 @@
-#![cfg(feature = "tests_that_use_llms")]
-
-use serde_json::Value;
-use serde_json::json;
-
-use paddler_test_cluster_harness::agent_config::AgentConfig;
-use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn qwen3_openai_chat_completion_with_reasoning_effort_none_generates_no_reasoning_tokens() {
-    let cluster = start_cluster_with_qwen3(vec![AgentConfig::single(1)])
-        .await
-        .expect("the cluster must start");
-
-    let response = cluster
-        .openai_chat_completion_non_streaming(&json!({
-            "model": "qwen3-test",
-            "messages": [{"role": "user", "content": "What is two plus two?"}],
-            "max_completion_tokens": 200,
-            "reasoning_effort": "none"
-        }))
-        .await
-        .expect("the OpenAI chat completion must succeed");
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-
-    assert_eq!(
-        response
-            .pointer("/usage/completion_tokens_details/reasoning_tokens")
-            .and_then(Value::as_u64),
-        Some(0)
-    );
-}
diff --git a/paddler_tests/tests/qwen3_openai_non_streaming_emits_tool_calls_for_function_tool.rs b/paddler_tests/tests/qwen3_openai_non_streaming_emits_tool_calls_for_function_tool.rs
deleted file mode 100644
index d82066881..000000000
--- a/paddler_tests/tests/qwen3_openai_non_streaming_emits_tool_calls_for_function_tool.rs
+++ /dev/null
@@ -1,104 +0,0 @@
-#![cfg(feature = "tests_that_use_llms")]
-
-use serde_json::Value;
-use serde_json::from_str;
-use serde_json::json;
-
-use paddler_test_cluster_harness::agent_config::AgentConfig;
-use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn qwen3_openai_non_streaming_emits_tool_calls_for_function_tool() {
-    let cluster = start_cluster_with_qwen3(vec![AgentConfig::single(1)])
-        .await
-        .expect("the cluster must start");
-
-    let response = cluster
-        .openai_chat_completion_non_streaming(&json!({
-            "model": "qwen3-test",
-            "messages": [{
-                "role": "user",
-                "content": "What is the weather in Paris? Use the get_weather tool."
-            }],
-            "max_completion_tokens": 400,
-            "tools": [{
-                "type": "function",
-                "function": {
-                    "name": "get_weather",
-                    "description": "Get the current weather for a location",
-                    "parameters": {
-                        "type": "object",
-                        "properties": {
-                            "location": {"type": "string", "description": "The city name"}
-                        },
-                        "required": ["location"],
-                        "additionalProperties": false
-                    }
-                }
-            }]
-        }))
-        .await
-        .expect("the OpenAI chat completion must succeed");
-
-    let tool_calls = response
-        .pointer("/choices/0/message/tool_calls")
-        .and_then(Value::as_array)
-        .unwrap_or_else(|| panic!("response missing message.tool_calls: {response}"));
-
-    assert_eq!(
-        tool_calls.len(),
-        1,
-        "expected exactly one structured tool call in non-streaming response (got {})",
-        tool_calls.len()
-    );
-
-    let first_call = &tool_calls[0];
-
-    assert_eq!(
-        first_call.pointer("/type").and_then(Value::as_str),
-        Some("function"),
-    );
-
-    let id = first_call
-        .pointer("/id")
-        .and_then(Value::as_str)
-        .expect("tool call missing id");
-    assert!(!id.is_empty(), "tool call id must not be empty");
-
-    let function_name = first_call
-        .pointer("/function/name")
-        .and_then(Value::as_str)
-        .expect("tool call missing function.name");
-
-    assert_eq!(function_name, "get_weather");
-
-    let function_arguments = first_call
-        .pointer("/function/arguments")
-        .and_then(Value::as_str)
-        .expect("tool call missing function.arguments");
-
-    let parsed_arguments: Value =
-        from_str(function_arguments).expect("the output must be valid JSON");
-    assert!(
-        parsed_arguments.get("location").is_some(),
-        "tool-call arguments JSON missing 'location' field: {function_arguments}"
-    );
-
-    let finish_reason = response
-        .pointer("/choices/0/finish_reason")
-        .and_then(Value::as_str)
-        .expect("response missing finish_reason");
-
-    assert_eq!(finish_reason, "tool_calls");
-
-    let completion_tokens = response
-        .pointer("/usage/completion_tokens")
-        .and_then(Value::as_u64)
-        .expect("response missing usage.completion_tokens");
-    assert!(completion_tokens > 0);
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_tests/tests/qwen3_openai_non_streaming_returns_usage.rs b/paddler_tests/tests/qwen3_openai_non_streaming_returns_usage.rs
deleted file mode 100644
index c9bb76594..000000000
--- a/paddler_tests/tests/qwen3_openai_non_streaming_returns_usage.rs
+++ /dev/null
@@ -1,59 +0,0 @@
-#![cfg(feature = "tests_that_use_llms")]
-
-use serde_json::Value;
-use serde_json::json;
-
-use paddler_test_cluster_harness::agent_config::AgentConfig;
-use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn qwen3_openai_non_streaming_returns_usage() {
-    let cluster = start_cluster_with_qwen3(vec![AgentConfig::single(1)])
-        .await
-        .expect("the cluster must start");
-
-    let response = cluster
-        .openai_chat_completion_non_streaming(&json!({
-            "model": "qwen3-test",
-            "messages": [{"role": "user", "content": "Say hi briefly."}],
-            "max_completion_tokens": 600
-        }))
-        .await
-        .expect("the OpenAI chat completion must succeed");
-
-    let usage = response
-        .get("usage")
-        .unwrap_or_else(|| panic!("non-streaming response missing usage: {response}"));
-
-    let prompt_tokens = usage
-        .get("prompt_tokens")
-        .and_then(Value::as_u64)
-        .expect("usage.prompt_tokens missing");
-    let completion_tokens = usage
-        .get("completion_tokens")
-        .and_then(Value::as_u64)
-        .expect("usage.completion_tokens missing");
-    let total_tokens = usage
-        .get("total_tokens")
-        .and_then(Value::as_u64)
-        .expect("usage.total_tokens missing");
-
-    assert!(prompt_tokens > 0);
-    assert!(completion_tokens > 0);
-    assert_eq!(total_tokens, prompt_tokens + completion_tokens);
-
-    let content = response
-        .pointer("/choices/0/message/content")
-        .and_then(Value::as_str)
-        .expect("non-streaming response missing message content");
-
-    assert!(
-        !content.is_empty(),
-        "non-streaming content must not be empty"
-    );
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_tests/tests/qwen3_openai_non_streaming_usage_with_tool_calls.rs b/paddler_tests/tests/qwen3_openai_non_streaming_usage_with_tool_calls.rs
deleted file mode 100644
index 20d499d02..000000000
--- a/paddler_tests/tests/qwen3_openai_non_streaming_usage_with_tool_calls.rs
+++ /dev/null
@@ -1,76 +0,0 @@
-#![cfg(feature = "tests_that_use_llms")]
-
-use serde_json::Value;
-use serde_json::json;
-
-use paddler_test_cluster_harness::agent_config::AgentConfig;
-use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn qwen3_openai_non_streaming_usage_with_tool_calls() {
-    let cluster = start_cluster_with_qwen3(vec![AgentConfig::single(1)])
-        .await
-        .expect("the cluster must start");
-
-    let response = cluster
-        .openai_chat_completion_non_streaming(&json!({
-            "model": "qwen3-test",
-            "messages": [{
-                "role": "user",
-                "content": "What is the weather in Paris? Use the get_weather tool."
-            }],
-            "max_completion_tokens": 400,
-            "tools": [{
-                "type": "function",
-                "function": {
-                    "name": "get_weather",
-                    "description": "Get the current weather for a location",
-                    "parameters": {
-                        "type": "object",
-                        "properties": {
-                            "location": {"type": "string"}
-                        },
-                        "required": ["location"],
-                        "additionalProperties": false
-                    }
-                }
-            }]
-        }))
-        .await
-        .expect("the OpenAI chat completion must succeed");
-
-    let tool_calls = response
-        .pointer("/choices/0/message/tool_calls")
-        .and_then(Value::as_array)
-        .unwrap_or_else(|| panic!("response missing message.tool_calls: {response}"));
-    assert!(!tool_calls.is_empty());
-
-    let usage = response
-        .get("usage")
-        .unwrap_or_else(|| panic!("response missing usage: {response}"));
-
-    let prompt_tokens = usage
-        .get("prompt_tokens")
-        .and_then(Value::as_u64)
-        .expect("usage.prompt_tokens missing");
-    let completion_tokens = usage
-        .get("completion_tokens")
-        .and_then(Value::as_u64)
-        .expect("usage.completion_tokens missing");
-    let total_tokens = usage
-        .get("total_tokens")
-        .and_then(Value::as_u64)
-        .expect("usage.total_tokens missing");
-
-    assert!(prompt_tokens > 0);
-    assert!(
-        completion_tokens > 0,
-        "expected non-zero completion_tokens for a tool-call response (got {completion_tokens})"
-    );
-    assert_eq!(total_tokens, prompt_tokens + completion_tokens);
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_tests/tests/qwen3_openai_streaming_emits_tool_calls_for_function_tool.rs b/paddler_tests/tests/qwen3_openai_streaming_emits_tool_calls_for_function_tool.rs
deleted file mode 100644
index 2d0c1fa1b..000000000
--- a/paddler_tests/tests/qwen3_openai_streaming_emits_tool_calls_for_function_tool.rs
+++ /dev/null
@@ -1,91 +0,0 @@
-#![cfg(feature = "tests_that_use_llms")]
-
-use serde_json::Value;
-use serde_json::from_str;
-use serde_json::json;
-
-use paddler_test_cluster_harness::agent_config::AgentConfig;
-use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn qwen3_openai_streaming_emits_tool_calls_for_function_tool() {
-    let cluster = start_cluster_with_qwen3(vec![AgentConfig::single(1)])
-        .await
-        .expect("the cluster must start");
-
-    let chunks = cluster
-        .openai_chat_completion_streaming(&json!({
-            "model": "qwen3-test",
-            "messages": [{
-                "role": "user",
-                "content": "What is the weather in Paris? Use the get_weather tool."
-            }],
-            "stream": true,
-            "max_completion_tokens": 400,
-            "tools": [{
-                "type": "function",
-                "function": {
-                    "name": "get_weather",
-                    "description": "Get the current weather for a location",
-                    "parameters": {
-                        "type": "object",
-                        "properties": {
-                            "location": {"type": "string", "description": "The city name"}
-                        },
-                        "required": ["location"],
-                        "additionalProperties": false
-                    }
-                }
-            }]
-        }))
-        .await
-        .expect("the OpenAI chat completion stream must succeed");
-
-    let chunks_with_tool_calls: Vec<&Value> = chunks
-        .iter()
-        .filter(|chunk| {
-            chunk
-                .pointer("/choices/0/delta/tool_calls")
-                .and_then(Value::as_array)
-                .is_some_and(|calls| !calls.is_empty())
-        })
-        .collect();
-
-    assert_eq!(
-        chunks_with_tool_calls.len(),
-        1,
-        "expected exactly one structured tool-call chunk per call (got {})",
-        chunks_with_tool_calls.len()
-    );
-
-    let structured_chunk = chunks_with_tool_calls[0];
-
-    let function_name = structured_chunk
-        .pointer("/choices/0/delta/tool_calls/0/function/name")
-        .and_then(Value::as_str)
-        .unwrap_or_else(|| {
-            panic!("structured tool-call chunk missing function.name: {structured_chunk}")
-        });
-
-    assert_eq!(function_name, "get_weather");
-
-    let function_arguments = structured_chunk
-        .pointer("/choices/0/delta/tool_calls/0/function/arguments")
-        .and_then(Value::as_str)
-        .unwrap_or_else(|| {
-            panic!("structured tool-call chunk missing function.arguments: {structured_chunk}")
-        });
-
-    let parsed_arguments: Value =
-        from_str(function_arguments).expect("the output must be valid JSON");
-
-    assert!(
-        parsed_arguments.get("location").is_some(),
-        "tool-call arguments JSON missing 'location' field: {function_arguments}"
-    );
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_tests/tests/qwen3_openai_streaming_emits_usage_when_requested.rs b/paddler_tests/tests/qwen3_openai_streaming_emits_usage_when_requested.rs
deleted file mode 100644
index 660f23abd..000000000
--- a/paddler_tests/tests/qwen3_openai_streaming_emits_usage_when_requested.rs
+++ /dev/null
@@ -1,65 +0,0 @@
-#![cfg(feature = "tests_that_use_llms")]
-
-use serde_json::Value;
-use serde_json::json;
-
-use paddler_test_cluster_harness::agent_config::AgentConfig;
-use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn qwen3_openai_streaming_emits_usage_when_requested() {
-    let cluster = start_cluster_with_qwen3(vec![AgentConfig::single(1)])
-        .await
-        .expect("the cluster must start");
-
-    let chunks = cluster
-        .openai_chat_completion_streaming(&json!({
-            "model": "qwen3-test",
-            "messages": [{"role": "user", "content": "Say hi briefly."}],
-            "stream": true,
-            "stream_options": {"include_usage": true},
-            "max_completion_tokens": 80
-        }))
-        .await
-        .expect("the OpenAI chat completion stream must succeed");
-
-    let last_chunk = chunks
-        .last()
-        .expect("no chunks received from streaming endpoint");
-
-    let usage = last_chunk
-        .get("usage")
-        .unwrap_or_else(|| panic!("trailing chunk lacks usage field: {last_chunk}"));
-
-    let prompt_tokens = usage
-        .get("prompt_tokens")
-        .and_then(Value::as_u64)
-        .expect("usage.prompt_tokens missing or not u64");
-    let completion_tokens = usage
-        .get("completion_tokens")
-        .and_then(Value::as_u64)
-        .expect("usage.completion_tokens missing or not u64");
-    let total_tokens = usage
-        .get("total_tokens")
-        .and_then(Value::as_u64)
-        .expect("usage.total_tokens missing or not u64");
-
-    assert!(prompt_tokens > 0);
-    assert!(completion_tokens > 0);
-    assert_eq!(total_tokens, prompt_tokens + completion_tokens);
-
-    let trailing_choices = last_chunk
-        .get("choices")
-        .and_then(Value::as_array)
-        .expect("trailing chunk lacks choices array");
-
-    assert!(
-        trailing_choices.is_empty(),
-        "OpenAI usage chunk must have empty choices array, got: {trailing_choices:?}"
-    );
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_tests/tests/qwen3_openai_streaming_omits_usage_when_not_requested.rs b/paddler_tests/tests/qwen3_openai_streaming_omits_usage_when_not_requested.rs
deleted file mode 100644
index 87f6cf179..000000000
--- a/paddler_tests/tests/qwen3_openai_streaming_omits_usage_when_not_requested.rs
+++ /dev/null
@@ -1,40 +0,0 @@
-#![cfg(feature = "tests_that_use_llms")]
-
-use serde_json::json;
-
-use paddler_test_cluster_harness::agent_config::AgentConfig;
-use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn qwen3_openai_streaming_omits_usage_when_not_requested() {
-    let cluster = start_cluster_with_qwen3(vec![AgentConfig::single(1)])
-        .await
-        .expect("the cluster must start");
-
-    let chunks = cluster
-        .openai_chat_completion_streaming(&json!({
-            "model": "qwen3-test",
-            "messages": [{"role": "user", "content": "Say hi briefly."}],
-            "stream": true,
-            "max_completion_tokens": 50
-        }))
-        .await
-        .expect("the OpenAI chat completion stream must succeed");
-
-    assert!(!chunks.is_empty(), "expected at least one chunk");
-
-    let chunks_with_usage = chunks
-        .iter()
-        .filter(|chunk| chunk.get("usage").is_some())
-        .count();
-
-    assert_eq!(
-        chunks_with_usage, 0,
-        "expected no usage chunks when stream_options.include_usage is absent, got {chunks_with_usage}"
-    );
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_tests/tests/qwen3_openai_streaming_usage_breakdown_with_thinking.rs b/paddler_tests/tests/qwen3_openai_streaming_usage_breakdown_with_thinking.rs
deleted file mode 100644
index 09eb89d54..000000000
--- a/paddler_tests/tests/qwen3_openai_streaming_usage_breakdown_with_thinking.rs
+++ /dev/null
@@ -1,56 +0,0 @@
-#![cfg(feature = "tests_that_use_llms")]
-
-use serde_json::Value;
-use serde_json::json;
-
-use paddler_test_cluster_harness::agent_config::AgentConfig;
-use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn qwen3_openai_streaming_usage_breakdown_with_thinking() {
-    let cluster = start_cluster_with_qwen3(vec![AgentConfig::single(1)])
-        .await
-        .expect("the cluster must start");
-
-    let chunks = cluster
-        .openai_chat_completion_streaming(&json!({
-            "model": "qwen3-test",
-            "messages": [{
-                "role": "user",
-                "content": "What is two plus two? Think briefly before answering."
-            }],
-            "stream": true,
-            "stream_options": {"include_usage": true},
-            "max_completion_tokens": 200
-        }))
-        .await
-        .expect("the OpenAI chat completion stream must succeed");
-
-    let usage_chunk = chunks
-        .iter()
-        .rev()
-        .find(|chunk| chunk.get("usage").is_some_and(|usage| !usage.is_null()))
-        .expect("no chunk contained usage information");
-
-    let prompt_tokens = usage_chunk
-        .pointer("/usage/prompt_tokens")
-        .and_then(Value::as_u64)
-        .expect("usage chunk missing prompt_tokens");
-    let completion_tokens = usage_chunk
-        .pointer("/usage/completion_tokens")
-        .and_then(Value::as_u64)
-        .expect("usage chunk missing completion_tokens");
-    let total_tokens = usage_chunk
-        .pointer("/usage/total_tokens")
-        .and_then(Value::as_u64)
-        .expect("usage chunk missing total_tokens");
-
-    assert!(prompt_tokens > 0);
-    assert!(completion_tokens > 0);
-    assert_eq!(total_tokens, prompt_tokens + completion_tokens);
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_tests/tests/qwen3_responses_non_streaming_returns_text_and_usage.rs b/paddler_tests/tests/qwen3_responses_non_streaming_returns_text_and_usage.rs
deleted file mode 100644
index e4d9fe57a..000000000
--- a/paddler_tests/tests/qwen3_responses_non_streaming_returns_text_and_usage.rs
+++ /dev/null
@@ -1,70 +0,0 @@
-#![cfg(feature = "tests_that_use_llms")]
-
-use serde_json::Value;
-use serde_json::json;
-
-use paddler_test_cluster_harness::agent_config::AgentConfig;
-use paddler_tests::start_cluster_with_qwen3::start_cluster_with_qwen3;
-
-#[tokio::test(flavor = "multi_thread")]
-async fn qwen3_responses_non_streaming_returns_text_and_usage() {
-    let cluster = start_cluster_with_qwen3(vec![AgentConfig::single(1)])
-        .await
-        .expect("the cluster must start");
-
-    let response = cluster
-        .openai_responses_non_streaming(&json!({
-            "model": "qwen3-test",
-            "input": "Say hi briefly.",
-            "max_output_tokens": 600
-        }))
-        .await
-        .expect("the OpenAI response must succeed");
-
-    assert_eq!(
-        response.get("status").and_then(Value::as_str),
-        Some("completed"),
-        "responses status must be completed: {response}"
-    );
-
-    let usage = response
-        .get("usage")
-        .unwrap_or_else(|| panic!("responses response missing usage: {response}"));
-
-    let input_tokens = usage
-        .get("input_tokens")
-        .and_then(Value::as_u64)
-        .expect("usage.input_tokens missing");
-    let output_tokens = usage
-        .get("output_tokens")
-        .and_then(Value::as_u64)
-        .expect("usage.output_tokens missing");
-    let total_tokens = usage
-        .get("total_tokens")
-        .and_then(Value::as_u64)
-        .expect("usage.total_tokens missing");
-
-    assert!(input_tokens > 0);
-    assert!(output_tokens > 0);
-    assert_eq!(total_tokens, input_tokens + output_tokens);
-
-    let message_text = response
-        .get("output")
-        .and_then(Value::as_array)
-        .expect("responses response missing output array")
-        .iter()
-        .find(|item| item.get("type").and_then(Value::as_str) == Some("message"))
-        .and_then(|message| message.pointer("/content/0/text"))
-        .and_then(Value::as_str)
-        .unwrap_or_else(|| panic!("responses output has no message text: {response}"));
-
-    assert!(
-        !message_text.is_empty(),
-        "responses message text must not be empty"
-    );
-
-    cluster
-        .shutdown()
-        .await
-        .expect("the cluster must shut down cleanly");
-}
diff --git a/paddler_tests/tests/qwen3_without_a_chat_template.rs b/paddler_tests/tests/qwen3_without_a_chat_template.rs
new file mode 100644
index 000000000..eccdb65af
--- /dev/null
+++ b/paddler_tests/tests/qwen3_without_a_chat_template.rs
@@ -0,0 +1,53 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::path::PathBuf;
+
+use tempfile::TempDir;
+
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_messaging::embedded_chat_template_key::EMBEDDED_CHAT_TEMPLATE_KEY;
+use paddler_test_cluster_harness::agent_desired_model_at::agent_desired_model_at;
+use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
+use paddler_tests::qwen3_desired_state::qwen3_desired_state;
+use paddler_tests::write_patched_gguf::write_patched_gguf;
+
+pub struct Qwen3WithoutAChatTemplate {
+    model_directory: TempDir,
+}
+
+impl Qwen3WithoutAChatTemplate {
+    pub fn write() -> Self {
+        let qwen3_without_a_chat_template = Self {
+            model_directory: TempDir::new().expect("a temporary model directory must be creatable"),
+        };
+
+        write_patched_gguf(
+            &qwen3_0_6b()
+                .cached_path()
+                .expect("Qwen3 must be in the Hugging Face cache"),
+            &qwen3_without_a_chat_template.model_path(),
+            |gguf_context| {
+                gguf_context
+                    .remove_key(EMBEDDED_CHAT_TEMPLATE_KEY)
+                    .expect("Qwen3 must embed a chat template to remove");
+            },
+        )
+        .expect("the model without a chat template must be writable");
+
+        qwen3_without_a_chat_template
+    }
+
+    pub fn desired_state(&self) -> BalancerDesiredState {
+        BalancerDesiredState {
+            model: agent_desired_model_at(&self.model_path())
+                .expect("the model without a chat template must be at an absolute path"),
+            ..qwen3_desired_state()
+        }
+    }
+
+    pub fn model_path(&self) -> PathBuf {
+        self.model_directory
+            .path()
+            .join("qwen3-without-a-chat-template.gguf")
+    }
+}
diff --git a/paddler_tests/tests/raw_inference_socket_reports_a_socket_closed_before_a_pong.rs b/paddler_tests/tests/raw_inference_socket_reports_a_socket_closed_before_a_pong.rs
new file mode 100644
index 000000000..41c8f8384
--- /dev/null
+++ b/paddler_tests/tests/raw_inference_socket_reports_a_socket_closed_before_a_pong.rs
@@ -0,0 +1,34 @@
+use std::mem::discriminant;
+
+use paddler_test_cluster_harness::cluster_harness_error::ClusterHarnessError;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_test_cluster_harness::raw_inference_socket::RawInferenceSocket;
+use paddler_tests::start_cluster::start_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn raw_inference_socket_reports_a_socket_closed_before_a_pong() {
+    let cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("a cluster without agents must start");
+    let mut raw_inference_socket =
+        RawInferenceSocket::connect(cluster.balancer.addresses.inference)
+            .await
+            .expect("the raw socket must reach the inference socket");
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+
+    assert_eq!(
+        raw_inference_socket
+            .next_pong()
+            .await
+            .err()
+            .as_ref()
+            .map(discriminant),
+        Some(discriminant(
+            &ClusterHarnessError::InferenceSocketClosedBeforePong
+        ))
+    );
+}
diff --git a/paddler_tests/tests/raw_inference_socket_reports_a_socket_closed_before_any_answer.rs b/paddler_tests/tests/raw_inference_socket_reports_a_socket_closed_before_any_answer.rs
new file mode 100644
index 000000000..092495316
--- /dev/null
+++ b/paddler_tests/tests/raw_inference_socket_reports_a_socket_closed_before_any_answer.rs
@@ -0,0 +1,34 @@
+use std::mem::discriminant;
+
+use paddler_test_cluster_harness::cluster_harness_error::ClusterHarnessError;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_test_cluster_harness::raw_inference_socket::RawInferenceSocket;
+use paddler_tests::start_cluster::start_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn raw_inference_socket_reports_a_socket_closed_before_any_answer() {
+    let cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("a balancer serving text generation and embeddings must start");
+    let mut raw_inference_socket =
+        RawInferenceSocket::connect(cluster.balancer.addresses.inference)
+            .await
+            .expect("the raw socket must reach the inference socket");
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+
+    assert_eq!(
+        raw_inference_socket
+            .next_answer()
+            .await
+            .err()
+            .as_ref()
+            .map(discriminant),
+        Some(discriminant(
+            &ClusterHarnessError::InferenceSocketClosedBeforeAnyAnswer
+        ))
+    );
+}
diff --git a/paddler_tests/tests/raw_inference_socket_skips_a_pong_while_waiting_for_an_answer.rs b/paddler_tests/tests/raw_inference_socket_skips_a_pong_while_waiting_for_an_answer.rs
new file mode 100644
index 000000000..ab929a688
--- /dev/null
+++ b/paddler_tests/tests/raw_inference_socket_skips_a_pong_while_waiting_for_an_answer.rs
@@ -0,0 +1,53 @@
+use std::num::NonZeroU32;
+
+use tokio_tungstenite::tungstenite::Bytes;
+use tokio_tungstenite::tungstenite::Message;
+
+use paddler_messaging::inference_client::message::Message as InferenceClientMessage;
+use paddler_messaging::jsonrpc::error::Error as JsonRpcError;
+use paddler_messaging::jsonrpc::error_envelope::ErrorEnvelope;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_test_cluster_harness::raw_inference_socket::RawInferenceSocket;
+use paddler_tests::short_buffered_request_timeout::SHORT_BUFFERED_REQUEST_TIMEOUT;
+use paddler_tests::start_cluster::start_cluster;
+
+const AWAITED_REQUEST_ID: &str = "request-after-the-ping";
+
+#[tokio::test(flavor = "multi_thread")]
+async fn raw_inference_socket_skips_a_pong_while_waiting_for_an_answer() {
+    let cluster = start_cluster(ClusterParams {
+        buffered_request_timeout: SHORT_BUFFERED_REQUEST_TIMEOUT,
+        max_buffered_requests: 1,
+        ..ClusterParams::without_agents()
+    })
+    .await
+    .expect("a cluster without agents must start");
+    let mut socket = RawInferenceSocket::connect(cluster.balancer.addresses.inference)
+        .await
+        .expect("the raw inference socket must connect");
+
+    socket
+        .send(Message::Ping(Bytes::from_static(b"probe")))
+        .await
+        .expect("the ping must be sent");
+    socket
+        .send_raw_prompt_request(AWAITED_REQUEST_ID, NonZeroU32::MIN)
+        .await
+        .expect("the request must be sent");
+
+    assert!(matches!(
+        socket
+            .next_answer_to(AWAITED_REQUEST_ID)
+            .await
+            .expect("the request must be answered after the pong"),
+        InferenceClientMessage::Error(ErrorEnvelope {
+            request_id,
+            error: JsonRpcError { code: 504, .. },
+        }) if request_id == AWAITED_REQUEST_ID
+    ));
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/raw_inference_socket_skips_an_answer_while_waiting_for_a_pong.rs b/paddler_tests/tests/raw_inference_socket_skips_an_answer_while_waiting_for_a_pong.rs
new file mode 100644
index 000000000..6d6313e5f
--- /dev/null
+++ b/paddler_tests/tests/raw_inference_socket_skips_an_answer_while_waiting_for_a_pong.rs
@@ -0,0 +1,49 @@
+use serde_json::json;
+use tokio_tungstenite::tungstenite::Bytes;
+use tokio_tungstenite::tungstenite::Message;
+
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_test_cluster_harness::raw_inference_socket::RawInferenceSocket;
+use paddler_tests::start_cluster::start_cluster;
+
+const PING_PAYLOAD: &[u8] = b"probe";
+
+#[tokio::test(flavor = "multi_thread")]
+async fn raw_inference_socket_skips_an_answer_while_waiting_for_a_pong() {
+    let cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("a cluster without agents must start");
+    let mut socket = RawInferenceSocket::connect(cluster.balancer.addresses.inference)
+        .await
+        .expect("the raw inference socket must connect");
+
+    socket
+        .send(Message::text(
+            json!({
+                "Request": {
+                    "id": "request-answered-before-the-pong",
+                    "request": { "NoSuchRequest": {} }
+                }
+            })
+            .to_string(),
+        ))
+        .await
+        .expect("the undeserializable request must be sent");
+    socket
+        .send(Message::Ping(Bytes::from_static(PING_PAYLOAD)))
+        .await
+        .expect("the ping must be sent");
+
+    assert_eq!(
+        socket
+            .next_pong()
+            .await
+            .expect("the pong must arrive after the answer"),
+        Bytes::from_static(PING_PAYLOAD)
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/reading_a_missing_state_file_changes_nothing.rs b/paddler_tests/tests/reading_a_missing_state_file_changes_nothing.rs
new file mode 100644
index 000000000..c2098262c
--- /dev/null
+++ b/paddler_tests/tests/reading_a_missing_state_file_changes_nothing.rs
@@ -0,0 +1,37 @@
+use tokio::fs::try_exists;
+use tokio_util::sync::CancellationToken;
+
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_test_cluster_harness::temporary_state_database_file::TemporaryStateDatabaseFile;
+use paddler_tests::start_cluster::start_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn reading_a_missing_state_file_changes_nothing() {
+    let database =
+        TemporaryStateDatabaseFile::new().expect("the state database file must be created");
+    let cluster = start_cluster(ClusterParams {
+        desired_state: ClusterDesiredState::KeepStored,
+        state_database_url: database.url.clone(),
+        ..ClusterParams::without_agents()
+    })
+    .await
+    .expect("the cluster must start without a state file");
+
+    cluster
+        .client_management
+        .get_balancer_desired_state(CancellationToken::new())
+        .await
+        .expect("the balancer must report its desired state");
+
+    assert!(
+        !try_exists(&database.path)
+            .await
+            .expect("the state file path must be inspectable")
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/ready_to_sample_request.rs b/paddler_tests/tests/ready_to_sample_request.rs
new file mode 100644
index 000000000..f8f1537d9
--- /dev/null
+++ b/paddler_tests/tests/ready_to_sample_request.rs
@@ -0,0 +1,126 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::slice;
+use std::sync::Arc;
+
+use llama_cpp_bindings::SampledTokenClassifier;
+use llama_cpp_bindings::context::LlamaContext;
+use llama_cpp_bindings::model::LlamaModel;
+use llama_cpp_bindings::token::data_array::LlamaTokenDataArray;
+use log::LevelFilter;
+use log::set_max_level;
+
+use paddler_agent_runtime::sequence_id_guard::SequenceIdGuard;
+use paddler_agent_runtime::sequence_id_pool::SequenceIdPool;
+use paddler_agent_text_generation::continuous_batch_active_request::ContinuousBatchActiveRequest;
+use paddler_agent_text_generation::continuous_batch_generation_step::ContinuousBatchGenerationStep;
+use paddler_agent_text_generation::continuous_batch_request_phase::ContinuousBatchRequestPhase;
+use paddler_agent_text_generation::continuous_batch_request_state::ContinuousBatchRequestState;
+use paddler_agent_text_generation::continuous_batch_scheduler::advance_generating_phase::AdvanceGeneratingPhase;
+use paddler_agent_text_generation::continuous_batch_scheduler_context::ContinuousBatchSchedulerContext;
+use paddler_agent_text_generation::continuous_batch_terminal_outcome::ContinuousBatchTerminalOutcome;
+use paddler_agent_text_generation::prepared_generation_request::PreparedGenerationRequest;
+use paddler_agent_text_generation::token_classification::TokenClassification;
+use paddler_tests::llama_runtime::LlamaRuntime;
+use paddler_tests::raw_prompt_generation::RawPromptGeneration;
+
+use crate::first_message_content_request_preparer::FirstMessageContentRequestPreparer;
+
+pub struct ReadyToSampleRequest<'runtime> {
+    pub batch_position: i32,
+    pub llama_context: &'runtime LlamaContext<'runtime>,
+    pub prompt: &'static str,
+    pub runtime: &'runtime LlamaRuntime,
+}
+
+impl ReadyToSampleRequest<'_> {
+    pub fn advance(
+        self,
+        customize_request: impl FnOnce(PreparedGenerationRequest) -> PreparedGenerationRequest,
+    ) -> ContinuousBatchTerminalOutcome {
+        let Self {
+            batch_position,
+            llama_context,
+            prompt,
+            runtime,
+        } = self;
+
+        set_max_level(LevelFilter::Trace);
+
+        let sequence_context_size = llama_context.n_ctx_seq();
+        let classifier_model: &'static LlamaModel =
+            Box::leak(Box::new(Arc::clone(&runtime.loaded_llama_model.model)));
+        let RawPromptGeneration {
+            agent_request,
+            generate_tokens_stop_tx: _generate_tokens_stop_tx,
+            generated_tokens_rx: _generated_tokens_rx,
+        } = RawPromptGeneration::new(
+            prompt,
+            runtime
+                .inference_runtime_context
+                .slot_aggregated_status
+                .clone(),
+        );
+        let PreparedGenerationRequest {
+            generate_tokens_stop_rx,
+            generated_tokens_tx,
+            max_tokens,
+            slot_guard,
+            token_classification:
+                TokenClassification {
+                    bare_json_tool_calls,
+                    streaming_markers,
+                },
+            token_sampling,
+            tool_call_handling,
+            ..
+        } = customize_request(
+            FirstMessageContentRequestPreparer {
+                llama_runtime: runtime,
+                sequence_context_size,
+            }
+            .into_generation_request_preparer()
+            .prepare_raw_prompt(agent_request)
+            .expect("the generation must be prepared"),
+        );
+        let token_classifier =
+            SampledTokenClassifier::new(classifier_model, streaming_markers, bare_json_tool_calls);
+        let mut active_request = ContinuousBatchActiveRequest {
+            state: ContinuousBatchRequestState {
+                current_token_position: batch_position + 1,
+                last_outcome_section: token_classifier.current_section(),
+                max_tokens,
+                phase: ContinuousBatchRequestPhase::Generating(
+                    ContinuousBatchGenerationStep::ReadyToSample { batch_position },
+                ),
+                prompt_tokens: Vec::new(),
+                prompt_tokens_ingested: 0,
+                sampled_tokens: 0,
+            },
+            token_classifier,
+            token_sampling,
+            generated_tokens_tx,
+            generate_tokens_stop_rx,
+            sequence_id_guard: SequenceIdGuard::acquire(&SequenceIdPool::new(1))
+                .expect("a fresh pool must hand out a sequence"),
+            slot_guard,
+            tool_call_handling,
+        };
+
+        AdvanceGeneratingPhase {
+            candidates: &mut LlamaTokenDataArray::new(Vec::new(), false),
+            ingest_outcomes: &mut Vec::new(),
+            llama_context,
+            scheduler_context: &ContinuousBatchSchedulerContext {
+                agent_name: None,
+                desired_slots_total: 1,
+                model: &runtime.loaded_llama_model.model,
+                n_batch: runtime.model_runtime_parameters.n_batch,
+                sequence_context_size,
+            },
+        }
+        .run(slice::from_mut(&mut active_request));
+
+        active_request.state.into_terminal_outcome()
+    }
+}
diff --git a/paddler_tests/tests/reapplying_the_served_state_keeps_serving_it_without_a_restart.rs b/paddler_tests/tests/reapplying_the_served_state_keeps_serving_it_without_a_restart.rs
new file mode 100644
index 000000000..bab694ee0
--- /dev/null
+++ b/paddler_tests/tests/reapplying_the_served_state_keeps_serving_it_without_a_restart.rs
@@ -0,0 +1,41 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::mem::discriminant;
+
+use paddler_messaging::produces_snapshot::ProducesSnapshot as _;
+use paddler_test_cluster_harness::model_card::qwen3_0_6b::qwen3_0_6b;
+use paddler_tests::desired_state_reconciled_discriminant::desired_state_reconciled_discriminant;
+
+use crate::serving_pipeline_arbiter::ServingPipelineArbiter;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn reapplying_the_served_state_keeps_serving_it_without_a_restart() {
+    let desired_state = qwen3_0_6b().into_desired_state();
+    let mut serving_pipeline_arbiter = ServingPipelineArbiter::new(1);
+
+    serving_pipeline_arbiter.serve(desired_state.clone()).await;
+
+    let status_while_serving = serving_pipeline_arbiter
+        .inference_runtime_context
+        .slot_aggregated_status
+        .make_snapshot()
+        .status;
+
+    assert_eq!(
+        discriminant(&serving_pipeline_arbiter.apply(desired_state).await),
+        desired_state_reconciled_discriminant()
+    );
+
+    serving_pipeline_arbiter.conclude_next_progress().await;
+
+    assert_eq!(
+        serving_pipeline_arbiter
+            .inference_runtime_context
+            .slot_aggregated_status
+            .make_snapshot()
+            .status,
+        status_while_serving
+    );
+
+    serving_pipeline_arbiter.shut_down().await;
+}
diff --git a/paddler_tests/tests/release_half_closed_openai_request.rs b/paddler_tests/tests/release_half_closed_openai_request.rs
new file mode 100644
index 000000000..422db6cad
--- /dev/null
+++ b/paddler_tests/tests/release_half_closed_openai_request.rs
@@ -0,0 +1,41 @@
+use serde_json::Value;
+
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_test_cluster_harness::half_closed_client::HalfClosedClient;
+use paddler_tests::start_cluster::start_cluster;
+
+pub async fn release_half_closed_openai_request(path: &str, body: &Value) {
+    let mut cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("a cluster without agents must start");
+    let mut client = HalfClosedClient::post_json_then_half_close(
+        cluster
+            .balancer
+            .compat_openai_addr()
+            .expect("the balancer must serve the OpenAI compatibility service"),
+        path,
+        body,
+    )
+    .await
+    .expect("the client must send its OpenAI-compatible request");
+
+    cluster
+        .wait_for_buffered_request_count(1)
+        .await
+        .expect("the request must be buffered while no agent is available");
+    client
+        .half_close()
+        .await
+        .expect("the client must half-close its connection");
+    cluster
+        .wait_for_buffered_request_count(0)
+        .await
+        .expect("the balancer must release the buffered request once the client half-closes");
+
+    drop(client);
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/replaceable_pointer_head.rs b/paddler_tests/tests/replaceable_pointer_head.rs
new file mode 100644
index 000000000..1128d1d1b
--- /dev/null
+++ b/paddler_tests/tests/replaceable_pointer_head.rs
@@ -0,0 +1,36 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::fs::copy;
+use std::fs::rename;
+use std::path::PathBuf;
+
+use tempfile::TempDir;
+
+use paddler_test_cluster_harness::fixture_path::fixture_path;
+
+pub struct ReplaceablePointerHead {
+    pub path: PathBuf,
+    directory: TempDir,
+}
+
+impl ReplaceablePointerHead {
+    pub fn holding_fixture(fixture_file_name: &str) -> Self {
+        let directory =
+            TempDir::new().expect("a temporary pointer head directory must be creatable");
+        let path = directory.path().join("pointer_head.gguf");
+
+        copy(fixture_path(fixture_file_name), &path)
+            .expect("the pointer head fixture must be copied into place");
+
+        Self { path, directory }
+    }
+
+    pub fn replace_with_fixture(&self, fixture_file_name: &str) {
+        let staged_path = self.directory.path().join("staged_pointer_head.gguf");
+
+        copy(fixture_path(fixture_file_name), &staged_path)
+            .expect("the replacing pointer head fixture must be staged");
+        rename(&staged_path, &self.path)
+            .expect("the staged pointer head must take the place of the pointer head");
+    }
+}
diff --git a/paddler_tests/tests/retaining_a_served_model_restarts_its_changed_capabilities_at_once.rs b/paddler_tests/tests/retaining_a_served_model_restarts_its_changed_capabilities_at_once.rs
new file mode 100644
index 000000000..51d539bb3
--- /dev/null
+++ b/paddler_tests/tests/retaining_a_served_model_restarts_its_changed_capabilities_at_once.rs
@@ -0,0 +1,83 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::future::Future as _;
+use std::pin::pin;
+use std::task::Context;
+use std::task::Waker;
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent::inference_capabilities::InferenceCapabilities;
+use paddler_agent::model_serving_settings::ModelServingSettings;
+use paddler_agent_text_generation::text_generation_settings::TextGenerationSettings;
+use paddler_blocking_thread_fixture::occupied_blocking_thread::OccupiedBlockingThread;
+use paddler_blocking_thread_fixture::single_blocking_thread_runtime::single_blocking_thread_runtime;
+use paddler_inference_parameters::pooling_type::PoolingType;
+use paddler_inference_parameters::sampling_parameters::SamplingParameters;
+
+use crate::kev_model_serving_without_decisions::KevModelServingWithoutDecisions;
+
+#[test]
+fn retaining_a_served_model_restarts_its_changed_capabilities_at_once() {
+    single_blocking_thread_runtime()
+        .expect("the test runtime must build")
+        .block_on(async {
+            let KevModelServingWithoutDecisions {
+                model_serving_settings,
+                served_model,
+                ..
+            } = KevModelServingWithoutDecisions::start(&CancellationToken::new()).await;
+            let embeddings_preparations = served_model
+                .capabilities
+                .embeddings
+                .request_preparer
+                .preparation_tasks
+                .clone();
+            let text_generation_preparations = served_model
+                .capabilities
+                .text_generation
+                .request_preparer
+                .preparation_tasks
+                .clone();
+            let desired_settings = ModelServingSettings {
+                capabilities: InferenceCapabilities {
+                    embeddings: PoolingType::Mean,
+                    text_generation: TextGenerationSettings {
+                        sampling_parameters: SamplingParameters {
+                            temperature: model_serving_settings
+                                .capabilities
+                                .text_generation
+                                .sampling_parameters
+                                .temperature
+                                / 2.0,
+                            ..model_serving_settings
+                                .capabilities
+                                .text_generation
+                                .sampling_parameters
+                                .clone()
+                        },
+                        ..model_serving_settings.capabilities.text_generation.clone()
+                    },
+                    ..model_serving_settings.capabilities.clone()
+                },
+                ..model_serving_settings
+            };
+            let occupied_blocking_thread = OccupiedBlockingThread::occupy();
+            let mut retaining = pin!(served_model.retain_for(desired_settings));
+
+            assert!(
+                retaining
+                    .as_mut()
+                    .poll(&mut Context::from_waker(Waker::noop()))
+                    .is_pending()
+            );
+            assert!(embeddings_preparations.is_closed());
+            assert!(text_generation_preparations.is_closed());
+
+            occupied_blocking_thread
+                .release()
+                .await
+                .expect("the blocking thread occupant must be released");
+            drop(retaining.await);
+        });
+}
diff --git a/paddler_tests/tests/run_with_test_cluster_exports_the_url_of_every_service.rs b/paddler_tests/tests/run_with_test_cluster_exports_the_url_of_every_service.rs
new file mode 100644
index 000000000..44f834697
--- /dev/null
+++ b/paddler_tests/tests/run_with_test_cluster_exports_the_url_of_every_service.rs
@@ -0,0 +1,22 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::process::Command;
+
+#[test]
+fn run_with_test_cluster_exports_the_url_of_every_service() {
+    assert!(
+        Command::new(env!("CARGO_BIN_EXE_run_with_test_cluster"))
+            .args([
+                "qwen3-0-6b",
+                "--",
+                "printenv",
+                "PADDLER_COMPAT_OPENAI_URL",
+                "PADDLER_COMPAT_TYPESAFE_URL",
+                "PADDLER_INFERENCE_URL",
+                "PADDLER_MANAGEMENT_URL",
+            ])
+            .status()
+            .expect("the test cluster runner must run")
+            .success()
+    );
+}
diff --git a/paddler_tests/tests/run_with_test_cluster_fails_when_its_agents_cannot_set_up_their_hugging_face_client.rs b/paddler_tests/tests/run_with_test_cluster_fails_when_its_agents_cannot_set_up_their_hugging_face_client.rs
new file mode 100644
index 000000000..09c33c074
--- /dev/null
+++ b/paddler_tests/tests/run_with_test_cluster_fails_when_its_agents_cannot_set_up_their_hugging_face_client.rs
@@ -0,0 +1,23 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::process::Command;
+
+use paddler_test_cluster_harness::failed_run_exit_code::FAILED_RUN_EXIT_CODE;
+
+const USER_AGENT_ORIGIN_WITH_A_LINE_BREAK: &str = "paddler\ntests";
+
+#[test]
+fn run_with_test_cluster_fails_when_its_agents_cannot_set_up_their_hugging_face_client() {
+    assert_eq!(
+        Command::new(env!("CARGO_BIN_EXE_run_with_test_cluster"))
+            .args(["qwen3-0-6b", "--", "true"])
+            .env(
+                "HF_HUB_USER_AGENT_ORIGIN",
+                USER_AGENT_ORIGIN_WITH_A_LINE_BREAK
+            )
+            .status()
+            .expect("the test cluster runner must run")
+            .code(),
+        Some(FAILED_RUN_EXIT_CODE)
+    );
+}
diff --git a/paddler_tests/tests/run_with_test_cluster_fails_when_its_agents_cannot_set_up_their_model_downloader.rs b/paddler_tests/tests/run_with_test_cluster_fails_when_its_agents_cannot_set_up_their_model_downloader.rs
new file mode 100644
index 000000000..570110278
--- /dev/null
+++ b/paddler_tests/tests/run_with_test_cluster_fails_when_its_agents_cannot_set_up_their_model_downloader.rs
@@ -0,0 +1,20 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::process::Command;
+
+use paddler_test_cluster_harness::ca_bundle_without_certificates::CA_BUNDLE_WITHOUT_CERTIFICATES;
+use paddler_test_cluster_harness::failed_run_exit_code::FAILED_RUN_EXIT_CODE;
+
+#[test]
+fn run_with_test_cluster_fails_when_its_agents_cannot_set_up_their_model_downloader() {
+    assert_eq!(
+        Command::new(env!("CARGO_BIN_EXE_run_with_test_cluster"))
+            .args(["qwen3-0-6b", "--", "true"])
+            .env("SSL_CERT_FILE", CA_BUNDLE_WITHOUT_CERTIFICATES)
+            .env_remove("SSL_CERT_DIR")
+            .status()
+            .expect("the test cluster runner must run")
+            .code(),
+        Some(FAILED_RUN_EXIT_CODE)
+    );
+}
diff --git a/paddler_tests/tests/run_with_test_cluster_fails_when_its_program_fails.rs b/paddler_tests/tests/run_with_test_cluster_fails_when_its_program_fails.rs
new file mode 100644
index 000000000..9e57f5b1f
--- /dev/null
+++ b/paddler_tests/tests/run_with_test_cluster_fails_when_its_program_fails.rs
@@ -0,0 +1,15 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::process::Command;
+
+#[test]
+fn run_with_test_cluster_fails_when_its_program_fails() {
+    assert_eq!(
+        Command::new(env!("CARGO_BIN_EXE_run_with_test_cluster"))
+            .args(["qwen3-0-6b", "--", "false"])
+            .status()
+            .expect("the test cluster runner must run")
+            .code(),
+        Some(1)
+    );
+}
diff --git a/paddler_tests/tests/run_with_test_cluster_fails_without_a_temporary_directory_for_agent_models.rs b/paddler_tests/tests/run_with_test_cluster_fails_without_a_temporary_directory_for_agent_models.rs
new file mode 100644
index 000000000..533056b80
--- /dev/null
+++ b/paddler_tests/tests/run_with_test_cluster_fails_without_a_temporary_directory_for_agent_models.rs
@@ -0,0 +1,18 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::process::Command;
+
+use paddler_test_cluster_harness::failed_run_exit_code::FAILED_RUN_EXIT_CODE;
+
+#[test]
+fn run_with_test_cluster_fails_without_a_temporary_directory_for_agent_models() {
+    assert_eq!(
+        Command::new(env!("CARGO_BIN_EXE_run_with_test_cluster"))
+            .args(["qwen3-0-6b", "--", "true"])
+            .env("TMPDIR", "/nonexistent-temporary-directory")
+            .status()
+            .expect("the test cluster runner must run")
+            .code(),
+        Some(FAILED_RUN_EXIT_CODE)
+    );
+}
diff --git a/paddler_tests/tests/sample_decision_answers_check.rs b/paddler_tests/tests/sample_decision_answers_check.rs
new file mode 100644
index 000000000..488c80455
--- /dev/null
+++ b/paddler_tests/tests/sample_decision_answers_check.rs
@@ -0,0 +1,57 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::decision_result::DecisionResult;
+use paddler_messaging::decision_summary::DecisionSummary;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_tests::sample_decision::sample_decision;
+use paddler_tests::start_cluster::start_cluster;
+
+const PROBABILITY_SUM_TOLERANCE: f32 = 1e-5;
+const SAMPLE_DECISION_INPUT_TOKENS: usize = 85;
+
+pub struct SampleDecisionAnswersCheck {
+    pub cluster_params: ClusterParams,
+}
+
+impl SampleDecisionAnswersCheck {
+    pub async fn assert_for_the_sample_decision(self) {
+        let cluster = start_cluster(self.cluster_params)
+            .await
+            .expect("the cluster must start");
+        let decision = sample_decision();
+
+        let collected = cluster
+            .decide(CancellationToken::new(), &decision)
+            .await
+            .expect("the decision must be accepted");
+
+        assert_eq!(
+            collected
+                .answers
+                .iter()
+                .map(|answer| (answer.id.as_str(), answer.probabilities.len()))
+                .collect::<Vec<_>>(),
+            decision
+                .questions
+                .iter()
+                .map(|question| (question.id.as_str(), question.options.len()))
+                .collect::<Vec<_>>()
+        );
+        assert!(collected.answers.iter().all(|answer| {
+            (answer.probabilities.iter().sum::<f32>() - 1.0).abs() < PROBABILITY_SUM_TOLERANCE
+        }));
+
+        assert!(matches!(
+            collected.terminal_result,
+            DecisionResult::Done(DecisionSummary { input_tokens, .. })
+                if input_tokens == SAMPLE_DECISION_INPUT_TOKENS
+        ));
+
+        cluster
+            .shutdown()
+            .await
+            .expect("the cluster must shut down cleanly");
+    }
+}
diff --git a/paddler_tests/tests/sampling_with_a_smaller_vocabulary_grammar.rs b/paddler_tests/tests/sampling_with_a_smaller_vocabulary_grammar.rs
new file mode 100644
index 000000000..5a6c50406
--- /dev/null
+++ b/paddler_tests/tests/sampling_with_a_smaller_vocabulary_grammar.rs
@@ -0,0 +1,63 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::path::Path;
+
+use llama_cpp_bindings::sampling::LlamaSampler;
+use llama_cpp_bindings::token::data_array::LlamaTokenDataArray;
+
+use paddler_agent_text_generation::sampling_outcome::SamplingOutcome;
+use paddler_agent_text_generation::text_generation_error::TextGenerationError;
+use paddler_agent_text_generation::token_sampling::TokenSampling;
+
+use paddler_tests::agent_inference_runtime_context::agent_inference_runtime_context;
+use paddler_tests::kev_checkpoint::KevCheckpoint;
+use paddler_tests::llama_runtime::LlamaRuntime;
+
+use crate::decode_prompt_for_sampling::decode_prompt_for_sampling;
+use crate::vocabulary_only_llama_model::VocabularyOnlyLlamaModel;
+
+pub struct SamplingWithASmallerVocabularyGrammar {
+    pub runtime: LlamaRuntime,
+    pub smaller_vocabulary_model: VocabularyOnlyLlamaModel,
+}
+
+impl SamplingWithASmallerVocabularyGrammar {
+    pub fn load(smaller_vocabulary_model_path: &Path) -> Self {
+        let runtime = LlamaRuntime::load(
+            &KevCheckpoint::Kev0_8b.target_path("model.gguf"),
+            agent_inference_runtime_context(1),
+        )
+        .expect("the kev model must load");
+        let smaller_vocabulary_model = VocabularyOnlyLlamaModel::load(
+            &runtime.inference_runtime_context.llama_backend,
+            smaller_vocabulary_model_path,
+        );
+
+        Self {
+            runtime,
+            smaller_vocabulary_model,
+        }
+    }
+
+    pub fn sample_the_prompt_with(
+        &self,
+        token_sampling: impl FnOnce(LlamaSampler) -> TokenSampling,
+    ) -> Result<SamplingOutcome, TextGenerationError> {
+        let mut llama_context = self
+            .runtime
+            .create_text_generation_llama_context(1)
+            .expect("the text generation context must be created");
+        let batch_position =
+            decode_prompt_for_sampling(&self.runtime, &mut llama_context, "Paddler balances");
+
+        token_sampling(
+            LlamaSampler::grammar(&self.smaller_vocabulary_model, r#"root ::= "a""#, "root")
+                .expect("the grammar of the smaller vocabulary must initialize"),
+        )
+        .sample(
+            &llama_context,
+            batch_position,
+            &mut LlamaTokenDataArray::new(Vec::new(), false),
+        )
+    }
+}
diff --git a/paddler_tests/tests/serving_decision_agent.rs b/paddler_tests/tests/serving_decision_agent.rs
new file mode 100644
index 000000000..4ad96bac8
--- /dev/null
+++ b/paddler_tests/tests/serving_decision_agent.rs
@@ -0,0 +1,68 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use tokio::sync::mpsc;
+use tokio::sync::mpsc::UnboundedReceiver;
+
+use paddler_agent::pipeline_request::PipelineRequest;
+use paddler_agent_runtime::agent_request::AgentRequest;
+use paddler_messaging::decision_result::DecisionResult;
+use paddler_messaging::decision_slots_minimum::DECISION_SLOTS_MINIMUM;
+use paddler_messaging::request_params::decide_params::raw_decide_params::RawDecideParams;
+use paddler_messaging::validates::Validates as _;
+use paddler_test_cluster_harness::model_card::qwen3_5_0_8b::qwen3_5_0_8b;
+use paddler_test_cluster_harness::pointer_head_fixture::pointer_head_fixture;
+use paddler_test_cluster_harness::synthetic_pointer_head_fixture::QWEN3_5_0_8B_SYNTHETIC_POINTER_HEAD_FIXTURE;
+
+use crate::serving_pipeline_arbiter::ServingPipelineArbiter;
+
+pub struct ServingDecisionAgent {
+    pub serving_pipeline_arbiter: ServingPipelineArbiter,
+}
+
+impl ServingDecisionAgent {
+    pub async fn start() -> Self {
+        let mut serving_pipeline_arbiter =
+            ServingPipelineArbiter::new(DECISION_SLOTS_MINIMUM.get());
+
+        serving_pipeline_arbiter
+            .serve(
+                qwen3_5_0_8b().into_decision_desired_state(
+                    pointer_head_fixture(QWEN3_5_0_8B_SYNTHETIC_POINTER_HEAD_FIXTURE)
+                        .expect("the pointer head fixture path must be absolute"),
+                ),
+            )
+            .await;
+
+        Self {
+            serving_pipeline_arbiter,
+        }
+    }
+
+    pub fn forward_decision(
+        &self,
+        raw_decide_params: RawDecideParams,
+        decision_stop_rx: UnboundedReceiver<()>,
+    ) -> UnboundedReceiver<DecisionResult> {
+        let (decision_result_tx, decision_result_rx) = mpsc::unbounded_channel();
+        let decide_params = raw_decide_params
+            .validate()
+            .expect("the decision parameters must be valid");
+
+        self.serving_pipeline_arbiter
+            .pipeline_arbiter
+            .forward(PipelineRequest::Decide(AgentRequest {
+                slot_guard: self
+                    .serving_pipeline_arbiter
+                    .slot_guard(decide_params.required_slots()),
+                params: decide_params,
+                response_tx: decision_result_tx,
+                stop_rx: decision_stop_rx,
+            }));
+
+        decision_result_rx
+    }
+
+    pub async fn shut_down(self) {
+        self.serving_pipeline_arbiter.shut_down().await;
+    }
+}
diff --git a/paddler_tests/tests/serving_pipeline_arbiter.rs b/paddler_tests/tests/serving_pipeline_arbiter.rs
new file mode 100644
index 000000000..2d0b0a1a9
--- /dev/null
+++ b/paddler_tests/tests/serving_pipeline_arbiter.rs
@@ -0,0 +1,108 @@
+#[cfg(feature = "tests_that_use_llms")]
+use std::mem::discriminant;
+#[cfg(feature = "tests_that_use_llms")]
+use std::num::NonZeroU16;
+use std::sync::Arc;
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent::agent_applicable_state_holder::AgentApplicableStateHolder;
+use paddler_agent::desired_state_reconciler::DesiredStateReconciler;
+use paddler_agent::desired_state_reconciliation::DesiredStateReconciliation;
+use paddler_agent::pipeline_arbiter::PipelineArbiter;
+use paddler_agent::pipeline_arbiter_params::PipelineArbiterParams;
+use paddler_agent_runtime::inference_runtime_context::InferenceRuntimeContext;
+#[cfg(feature = "tests_that_use_llms")]
+use paddler_agent_status::slot_guard::SlotGuard;
+use paddler_messaging::agent_desired_state::AgentDesiredState;
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_model_source::desired_model_resolver::DesiredModelResolver;
+use paddler_model_source::huggingface_default_endpoint::HUGGINGFACE_DEFAULT_ENDPOINT;
+use paddler_tests::agent_inference_runtime_context::agent_inference_runtime_context;
+#[cfg(feature = "tests_that_use_llms")]
+use paddler_tests::desired_state_reconciled_discriminant::desired_state_reconciled_discriminant;
+
+pub struct ServingPipelineArbiter {
+    pub desired_state_reconciler: DesiredStateReconciler,
+    pub inference_runtime_context: InferenceRuntimeContext,
+    pub pipeline_arbiter: PipelineArbiter,
+}
+
+impl ServingPipelineArbiter {
+    #[must_use]
+    pub fn new(slots: u16) -> Self {
+        let inference_runtime_context = agent_inference_runtime_context(slots);
+
+        Self {
+            desired_state_reconciler: DesiredStateReconciler {
+                desired_model_resolver: DesiredModelResolver::from_process_env(
+                    HUGGINGFACE_DEFAULT_ENDPOINT.to_owned(),
+                )
+                .expect("the model downloader must start"),
+                slot_aggregated_status: inference_runtime_context.slot_aggregated_status.clone(),
+            },
+            pipeline_arbiter: PipelineArbiter::new(PipelineArbiterParams {
+                agent_shutdown: CancellationToken::new(),
+                inference_runtime_context: inference_runtime_context.clone(),
+                served_applicable_state_holder: Arc::new(AgentApplicableStateHolder::default()),
+            }),
+            inference_runtime_context,
+        }
+    }
+
+    #[cfg(feature = "tests_that_use_llms")]
+    pub async fn serve(&mut self, desired_state: BalancerDesiredState) {
+        let desired_state_reconciliation = self.apply(desired_state).await;
+
+        assert_eq!(
+            discriminant(&desired_state_reconciliation),
+            desired_state_reconciled_discriminant()
+        );
+
+        self.conclude_next_progress().await;
+        self.conclude_next_progress().await;
+    }
+
+    #[cfg(feature = "tests_that_use_llms")]
+    pub async fn conclude_next_progress(&mut self) {
+        let serving_startup_progress = self.pipeline_arbiter.next_finished_startup().await;
+
+        self.pipeline_arbiter.conclude(serving_startup_progress);
+    }
+
+    pub async fn apply(
+        &mut self,
+        desired_state: BalancerDesiredState,
+    ) -> DesiredStateReconciliation {
+        let desired_state_reconciliation = self
+            .desired_state_reconciler
+            .reconcile(
+                &CancellationToken::new(),
+                AgentDesiredState::from(&desired_state),
+            )
+            .await;
+
+        if let DesiredStateReconciliation::Reconciled(agent_applicable_state) =
+            &desired_state_reconciliation
+        {
+            self.pipeline_arbiter.apply(*agent_applicable_state.clone());
+        }
+
+        desired_state_reconciliation
+    }
+
+    #[cfg(feature = "tests_that_use_llms")]
+    #[must_use]
+    pub fn slot_guard(&self, reserved_slots: NonZeroU16) -> SlotGuard {
+        SlotGuard::new(
+            self.inference_runtime_context
+                .slot_aggregated_status
+                .clone(),
+            reserved_slots,
+        )
+    }
+
+    pub async fn shut_down(self) {
+        self.pipeline_arbiter.shut_down().await;
+    }
+}
diff --git a/paddler_tests/tests/shutdown_during_state_transition.rs b/paddler_tests/tests/shutdown_during_state_transition.rs
new file mode 100644
index 000000000..bc25e9684
--- /dev/null
+++ b/paddler_tests/tests/shutdown_during_state_transition.rs
@@ -0,0 +1,63 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::agent_state_application_status::AgentStateApplicationStatus;
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_tests::start_cluster::start_cluster;
+
+pub struct ShutdownDuringStateTransition {
+    pub initial_desired_state: BalancerDesiredState,
+    pub next_desired_state: BalancerDesiredState,
+    pub slot_count: u16,
+}
+
+impl ShutdownDuringStateTransition {
+    pub async fn shut_down_while_the_agent_applies_the_next_state(self) {
+        let mut cluster = start_cluster(ClusterParams {
+            agents: vec![AgentConfig::single(self.slot_count)],
+            desired_state: ClusterDesiredState::Apply(Box::new(self.initial_desired_state)),
+            wait_for_slots_ready: false,
+            ..ClusterParams::default()
+        })
+        .await
+        .expect("the cluster must start");
+        let agent_id = cluster
+            .agent_ids
+            .first()
+            .expect("the cluster must register its agent")
+            .clone();
+
+        cluster
+            .agents_watcher
+            .until_agent(&agent_id, |snapshot| {
+                snapshot.agents.iter().all(|agent| {
+                    agent.status.state_application_status == AgentStateApplicationStatus::Applied
+                })
+            })
+            .await
+            .expect("the agent must apply the initial state");
+        cluster
+            .client_management
+            .put_balancer_desired_state(CancellationToken::new(), &self.next_desired_state)
+            .await
+            .expect("the balancer must accept the next state");
+        cluster
+            .agents_watcher
+            .until_agent(&agent_id, |snapshot| {
+                snapshot.agents.iter().any(|agent| {
+                    agent.status.state_application_status != AgentStateApplicationStatus::Applied
+                })
+            })
+            .await
+            .expect("the agent must start applying the next state");
+
+        cluster
+            .shutdown()
+            .await
+            .expect("the cluster must shut down cleanly while its agent applies the next state");
+    }
+}
diff --git a/paddler_tests/tests/shutting_down_a_served_model_closes_every_capability_at_once.rs b/paddler_tests/tests/shutting_down_a_served_model_closes_every_capability_at_once.rs
new file mode 100644
index 000000000..3b035226c
--- /dev/null
+++ b/paddler_tests/tests/shutting_down_a_served_model_closes_every_capability_at_once.rs
@@ -0,0 +1,53 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::future::Future as _;
+use std::pin::pin;
+use std::task::Context;
+use std::task::Waker;
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent_runtime::shuts_down::ShutsDown as _;
+use paddler_blocking_thread_fixture::occupied_blocking_thread::OccupiedBlockingThread;
+use paddler_blocking_thread_fixture::single_blocking_thread_runtime::single_blocking_thread_runtime;
+
+use crate::kev_model_serving_without_decisions::KevModelServingWithoutDecisions;
+
+#[test]
+fn shutting_down_a_served_model_closes_every_capability_at_once() {
+    single_blocking_thread_runtime()
+        .expect("the test runtime must build")
+        .block_on(async {
+            let KevModelServingWithoutDecisions { served_model, .. } =
+                KevModelServingWithoutDecisions::start(&CancellationToken::new()).await;
+            let embeddings_preparations = served_model
+                .capabilities
+                .embeddings
+                .request_preparer
+                .preparation_tasks
+                .clone();
+            let text_generation_preparations = served_model
+                .capabilities
+                .text_generation
+                .request_preparer
+                .preparation_tasks
+                .clone();
+            let occupied_blocking_thread = OccupiedBlockingThread::occupy();
+            let mut shutting_down = pin!(served_model.shut_down());
+
+            assert!(
+                shutting_down
+                    .as_mut()
+                    .poll(&mut Context::from_waker(Waker::noop()))
+                    .is_pending()
+            );
+            assert!(embeddings_preparations.is_closed());
+            assert!(text_generation_preparations.is_closed());
+
+            occupied_blocking_thread
+                .release()
+                .await
+                .expect("the blocking thread occupant must be released");
+            shutting_down.await;
+        });
+}
diff --git a/paddler_tests/tests/shutting_down_a_served_model_stops_its_schedulers_and_frees_their_contexts.rs b/paddler_tests/tests/shutting_down_a_served_model_stops_its_schedulers_and_frees_their_contexts.rs
new file mode 100644
index 000000000..f14d18973
--- /dev/null
+++ b/paddler_tests/tests/shutting_down_a_served_model_stops_its_schedulers_and_frees_their_contexts.rs
@@ -0,0 +1,30 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_agent_runtime::shuts_down::ShutsDown as _;
+
+use crate::kev_model_serving_without_decisions::KevModelServingWithoutDecisions;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn shutting_down_a_served_model_stops_its_schedulers_and_frees_their_contexts() {
+    let KevModelServingWithoutDecisions {
+        llama_runtime,
+        served_model,
+        ..
+    } = KevModelServingWithoutDecisions::start(&CancellationToken::new()).await;
+    let vram_residencies = llama_runtime
+        .inference_runtime_context
+        .vram_residencies
+        .clone();
+
+    served_model.shut_down().await;
+
+    assert!(vram_residencies.embeddings_context.try_occupy().is_ok());
+    assert!(
+        vram_residencies
+            .text_generation_context
+            .try_occupy()
+            .is_ok()
+    );
+}
diff --git a/paddler_tests/tests/slot_held_by_an_unending_generation.rs b/paddler_tests/tests/slot_held_by_an_unending_generation.rs
new file mode 100644
index 000000000..39dda943e
--- /dev/null
+++ b/paddler_tests/tests/slot_held_by_an_unending_generation.rs
@@ -0,0 +1,53 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use futures_util::StreamExt as _;
+use tokio_util::sync::CancellationToken;
+
+use paddler_client::inference_message_stream::InferenceMessageStream;
+use paddler_test_cluster_harness::cluster::Cluster;
+use paddler_test_cluster_harness::unending_generation::unending_generation;
+
+pub struct SlotHeldByAnUnendingGeneration {
+    generation_stop: CancellationToken,
+    unending_generation_stream: InferenceMessageStream,
+}
+
+impl SlotHeldByAnUnendingGeneration {
+    pub async fn hold(cluster: &mut Cluster) -> Self {
+        let agent_id = cluster
+            .agent_ids
+            .first()
+            .expect("the cluster must register its agent")
+            .clone();
+        let generation_stop = CancellationToken::new();
+        let mut unending_generation_stream = cluster
+            .client_inference
+            .continue_from_raw_prompt(generation_stop.clone(), unending_generation())
+            .await
+            .expect("the generation must be accepted");
+
+        unending_generation_stream
+            .next()
+            .await
+            .expect("the generation must stream a message while it holds its slot")
+            .expect("the message must be readable");
+        cluster
+            .wait_for_slots_processing(&agent_id, 1)
+            .await
+            .expect("the generation must hold one slot");
+
+        Self {
+            generation_stop,
+            unending_generation_stream,
+        }
+    }
+
+    pub async fn release(mut self) {
+        self.generation_stop.cancel();
+
+        assert!(
+            self.unending_generation_stream.next().await.is_none(),
+            "a cancelled generation must end its stream"
+        );
+    }
+}
diff --git a/paddler_tests/tests/small_context_decision_cluster_params.rs b/paddler_tests/tests/small_context_decision_cluster_params.rs
new file mode 100644
index 000000000..34d68dd14
--- /dev/null
+++ b/paddler_tests/tests/small_context_decision_cluster_params.rs
@@ -0,0 +1,22 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::num::NonZeroU32;
+
+use paddler_inference_parameters::batch_size::BatchSize;
+use paddler_tests::decision_cluster_params::DecisionClusterParams;
+use paddler_tests::decision_desired_state::DecisionDesiredState;
+
+pub const SMALL_DECISION_CONTEXT_SIZE: u32 = 256;
+
+pub fn small_context_decision_cluster_params() -> DecisionClusterParams {
+    DecisionClusterParams {
+        desired_state: DecisionDesiredState {
+            context_size: NonZeroU32::new(SMALL_DECISION_CONTEXT_SIZE)
+                .expect("the context size is non-zero"),
+            n_batch: BatchSize::try_from(SMALL_DECISION_CONTEXT_SIZE)
+                .expect("the batch fits the context"),
+            ..DecisionDesiredState::default()
+        },
+        ..DecisionClusterParams::default()
+    }
+}
diff --git a/paddler_tests/tests/smolvlm2_generates_tokens_for_jpeg_and_webp_images_with_the_default_batch.rs b/paddler_tests/tests/smolvlm2_generates_tokens_for_jpeg_and_webp_images_with_the_default_batch.rs
new file mode 100644
index 000000000..c05a41a44
--- /dev/null
+++ b/paddler_tests/tests/smolvlm2_generates_tokens_for_jpeg_and_webp_images_with_the_default_batch.rs
@@ -0,0 +1,49 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::num::NonZeroU32;
+
+use tokio_util::sync::CancellationToken;
+
+use paddler_messaging::generated_token_result::GeneratedTokenResult;
+use paddler_test_cluster_harness::agent_config::AgentConfig;
+use paddler_tests::jpeg_and_webp_image_fixtures::JPEG_AND_WEBP_IMAGE_FIXTURES;
+use paddler_tests::single_image_question::SingleImageQuestion;
+use paddler_tests::start_cluster_with_smolvlm2::start_cluster_with_smolvlm2;
+
+const MAX_TOKENS: NonZeroU32 = NonZeroU32::new(20).unwrap();
+
+#[tokio::test(flavor = "multi_thread")]
+async fn smolvlm2_generates_tokens_for_jpeg_and_webp_images_with_the_default_batch() {
+    let cluster = start_cluster_with_smolvlm2(vec![AgentConfig::single(1)])
+        .await
+        .expect("the cluster must start");
+
+    for image_fixture in JPEG_AND_WEBP_IMAGE_FIXTURES {
+        let token_results = cluster
+            .continue_from_conversation_history(
+                CancellationToken::new(),
+                &SingleImageQuestion {
+                    image_data_uri: image_fixture
+                        .data_uri()
+                        .expect("the image fixture must load"),
+                    max_tokens: MAX_TOKENS,
+                    question: "What do you see in this image?",
+                }
+                .into_conversation_params(),
+            )
+            .await
+            .expect("the image fixture request must complete")
+            .into_token_results();
+
+        assert!(
+            token_results.iter().any(GeneratedTokenResult::is_token),
+            "fixture {} should produce at least one content/reasoning/tool-call token with adequate n_batch; got {token_results:?}",
+            image_fixture.fixture_name,
+        );
+    }
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/smolvlm2_generates_tokens_from_image_input.rs b/paddler_tests/tests/smolvlm2_generates_tokens_from_image_input.rs
index 33c5f29d8..63c910537 100644
--- a/paddler_tests/tests/smolvlm2_generates_tokens_from_image_input.rs
+++ b/paddler_tests/tests/smolvlm2_generates_tokens_from_image_input.rs
@@ -8,12 +8,10 @@ use paddler_messaging::conversation_history::ConversationHistory;
 use paddler_messaging::conversation_message::ConversationMessage;
 use paddler_messaging::conversation_message_content::ConversationMessageContent;
 use paddler_messaging::conversation_message_content_part::ConversationMessageContentPart;
-use paddler_messaging::generated_token_result::GeneratedTokenResult;
 use paddler_messaging::image_url::ImageUrl;
 use paddler_messaging::request_params::continue_from_conversation_history_params::ContinueFromConversationHistoryParams;
 use paddler_test_cluster_harness::agent_config::AgentConfig;
 use paddler_test_cluster_harness::load_test_image_data_uri::load_test_image_data_uri;
-use paddler_test_cluster_harness::token_result_with_producer::TokenResultWithProducer;
 use paddler_tests::start_cluster_with_smolvlm2::start_cluster_with_smolvlm2;
 
 #[tokio::test(flavor = "multi_thread")]
@@ -24,7 +22,7 @@ async fn smolvlm2_generates_tokens_from_image_input() {
 
     let image_data_uri = load_test_image_data_uri().expect("the test image must load");
 
-    let conversation_history = ConversationHistory::new(vec![ConversationMessage {
+    let conversation_history = ConversationHistory::new(vec![ConversationMessage::Message {
         content: ConversationMessageContent::Parts(vec![
             ConversationMessageContentPart::ImageUrl {
                 image_url: ImageUrl {
@@ -61,13 +59,9 @@ async fn smolvlm2_generates_tokens_from_image_input() {
         .count();
 
     assert!(token_count > 0);
-    assert!(matches!(
-        collected.token_results.last(),
-        Some(TokenResultWithProducer {
-            token_result: GeneratedTokenResult::Done(_),
-            ..
-        })
-    ));
+    collected
+        .summary()
+        .expect("the generation must finish with a summary");
 
     cluster
         .shutdown()
diff --git a/paddler_tests/tests/start_cluster_reports_a_desired_state_the_balancer_cannot_store.rs b/paddler_tests/tests/start_cluster_reports_a_desired_state_the_balancer_cannot_store.rs
new file mode 100644
index 000000000..a56859064
--- /dev/null
+++ b/paddler_tests/tests/start_cluster_reports_a_desired_state_the_balancer_cannot_store.rs
@@ -0,0 +1,57 @@
+use std::fs::Permissions;
+use std::fs::set_permissions;
+use std::os::unix::fs::PermissionsExt as _;
+
+use reqwest::StatusCode;
+use reqwest::Url;
+use tempfile::TempDir;
+
+use paddler_client::error::Error as ClientError;
+use paddler_messaging::agent_desired_model::AgentDesiredModel;
+use paddler_messaging::api_path::ApiPath;
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_tests::start_cluster::start_cluster;
+
+const READ_ONLY_DIRECTORY_MODE: u32 = 0o555;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn start_cluster_reports_a_desired_state_the_balancer_cannot_store() {
+    let state_directory = TempDir::new().expect("a temporary directory must be created");
+
+    set_permissions(
+        state_directory.path(),
+        Permissions::from_mode(READ_ONLY_DIRECTORY_MODE),
+    )
+    .expect("the state directory must become read-only");
+
+    let start_error = start_cluster(ClusterParams {
+        desired_state: ClusterDesiredState::Apply(Box::new(BalancerDesiredState {
+            model: AgentDesiredModel::Uri("agent:///unstorable-model.gguf".to_owned()),
+            ..BalancerDesiredState::default()
+        })),
+        state_database_url: format!(
+            "file://{}",
+            state_directory.path().join("state.json").display()
+        ),
+        ..ClusterParams::without_agents()
+    })
+    .await
+    .err()
+    .expect("a desired state the balancer cannot store must fail the cluster start");
+
+    assert!(matches!(
+        start_error.downcast_ref::<ClientError>(),
+        Some(ClientError::UnexpectedResponseStatus { message, status, url })
+            if *status == StatusCode::INTERNAL_SERVER_ERROR
+                && *message
+                    == format!(
+                        "Unable to create a replacement of the state database file in '{}'",
+                        state_directory.path().display()
+                    )
+                && Url::parse(url).is_ok_and(|desired_state_url| {
+                    desired_state_url.path() == ApiPath::BALANCER_DESIRED_STATE
+                })
+    ));
+}
diff --git a/paddler_tests/tests/start_cluster_reports_a_state_database_url_it_cannot_parse.rs b/paddler_tests/tests/start_cluster_reports_a_state_database_url_it_cannot_parse.rs
new file mode 100644
index 000000000..d8d0fd7f3
--- /dev/null
+++ b/paddler_tests/tests/start_cluster_reports_a_state_database_url_it_cannot_parse.rs
@@ -0,0 +1,19 @@
+use paddler_state_database::state_database_error::StateDatabaseError;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_tests::start_cluster::start_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn start_cluster_reports_a_state_database_url_it_cannot_parse() {
+    let start_error = start_cluster(ClusterParams {
+        state_database_url: "unsupported-scheme://state".to_owned(),
+        ..ClusterParams::without_agents()
+    })
+    .await
+    .err()
+    .expect("a state database URL with an unsupported scheme must fail the cluster start");
+
+    assert!(matches!(
+        start_error.downcast_ref::<StateDatabaseError>(),
+        Some(StateDatabaseError::SchemeUnsupported { scheme }) if scheme == "unsupported-scheme"
+    ));
+}
diff --git a/paddler_tests/tests/swappable_model_file.rs b/paddler_tests/tests/swappable_model_file.rs
new file mode 100644
index 000000000..13c5a71ac
--- /dev/null
+++ b/paddler_tests/tests/swappable_model_file.rs
@@ -0,0 +1,47 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::fs::rename;
+use std::fs::write;
+use std::os::unix::fs::symlink;
+use std::path::Path;
+use std::path::PathBuf;
+
+use tempfile::TempDir;
+
+use paddler_messaging::agent_desired_model::AgentDesiredModel;
+use paddler_test_cluster_harness::agent_desired_model_at::agent_desired_model_at;
+
+pub struct SwappableModelFile {
+    model_directory: TempDir,
+}
+
+impl SwappableModelFile {
+    pub fn link_to(source_model_path: &Path) -> Self {
+        let swappable_model_file = Self {
+            model_directory: TempDir::new().expect("a temporary model directory must be creatable"),
+        };
+
+        symlink(source_model_path, swappable_model_file.model_path())
+            .expect("the model must be linkable");
+
+        swappable_model_file
+    }
+
+    pub fn agent_desired_model(&self) -> AgentDesiredModel {
+        agent_desired_model_at(&self.model_path())
+            .expect("the swappable model file must be at an absolute path")
+    }
+
+    pub fn model_path(&self) -> PathBuf {
+        self.model_directory.path().join("model.gguf")
+    }
+
+    pub fn replace_with_an_unloadable_file(&self) {
+        let unloadable_model_path = self.model_directory.path().join("unloadable.gguf");
+
+        write(&unloadable_model_path, b"not a model")
+            .expect("the unloadable file must be writable");
+        rename(&unloadable_model_path, self.model_path())
+            .expect("the unloadable file must replace the model");
+    }
+}
diff --git a/paddler_tests/tests/test_only_attribute_counter.rs b/paddler_tests/tests/test_only_attribute_counter.rs
new file mode 100644
index 000000000..e88b5ebcc
--- /dev/null
+++ b/paddler_tests/tests/test_only_attribute_counter.rs
@@ -0,0 +1,28 @@
+use syn::Attribute;
+use syn::Ident;
+use syn::ItemMod;
+use syn::visit::Visit;
+use syn::visit::visit_item_mod;
+
+#[derive(Default)]
+pub struct TestOnlyAttributeCounter {
+    pub test_only_attributes: usize,
+}
+
+impl<'ast> Visit<'ast> for TestOnlyAttributeCounter {
+    fn visit_attribute(&mut self, attribute: &'ast Attribute) {
+        if attribute.path().is_ident("cfg")
+            && attribute
+                .parse_args::<Ident>()
+                .is_ok_and(|cfg_predicate| cfg_predicate == "test")
+        {
+            self.test_only_attributes += 1;
+        }
+    }
+
+    fn visit_item_mod(&mut self, item_mod: &'ast ItemMod) {
+        if item_mod.ident != "tests" || item_mod.content.is_none() {
+            visit_item_mod(self, item_mod);
+        }
+    }
+}
diff --git a/paddler_tests/tests/the_test_only_code_detector_counts_only_test_code_outside_inline_test_modules.rs b/paddler_tests/tests/the_test_only_code_detector_counts_only_test_code_outside_inline_test_modules.rs
new file mode 100644
index 000000000..23ccdb7a6
--- /dev/null
+++ b/paddler_tests/tests/the_test_only_code_detector_counts_only_test_code_outside_inline_test_modules.rs
@@ -0,0 +1,30 @@
+use syn::parse_file;
+use syn::visit::Visit as _;
+
+use crate::test_only_attribute_counter::TestOnlyAttributeCounter;
+
+#[test]
+fn the_test_only_code_detector_counts_only_test_code_outside_inline_test_modules() {
+    let mut test_only_attribute_counter = TestOnlyAttributeCounter::default();
+
+    test_only_attribute_counter.visit_file(
+        &parse_file(
+            r#"
+                #[cfg(test)]
+                mod helper_outside_the_inline_test_module;
+
+                #[cfg(feature = "tests_that_use_llms")]
+                fn feature_gated_production_code() {}
+
+                #[cfg(test)]
+                mod tests {
+                    #[cfg(test)]
+                    fn helper_inside_the_inline_test_module() {}
+                }
+            "#,
+        )
+        .expect("the probe source must parse"),
+    );
+
+    assert_eq!(test_only_attribute_counter.test_only_attributes, 1);
+}
diff --git a/paddler_tests/tests/tool_call_pipeline_for_tools.rs b/paddler_tests/tests/tool_call_pipeline_for_tools.rs
new file mode 100644
index 000000000..a8d83359b
--- /dev/null
+++ b/paddler_tests/tests/tool_call_pipeline_for_tools.rs
@@ -0,0 +1,18 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use llama_cpp_bindings::ChatMessageParser;
+use llama_cpp_bindings::model::LlamaModel;
+use serde_json::Value;
+
+use paddler_agent_text_generation::chat_template_tools::ChatTemplateTools;
+use paddler_agent_text_generation::tool_call_pipeline::ToolCallPipeline;
+use paddler_messaging::request_params::continue_from_conversation_history_params::tool::Tool;
+use paddler_tool_call_validator::tool_call_validator::ToolCallValidator;
+
+pub fn tool_call_pipeline_for_tools(model: &LlamaModel, tools: &[Tool<Value>]) -> ToolCallPipeline {
+    ToolCallPipeline::new(
+        ChatMessageParser::new(model, &ChatTemplateTools::new(tools).json)
+            .expect("the chat message parser must be created"),
+        ToolCallValidator::from_tools(tools).expect("the tools must validate"),
+    )
+}
diff --git a/paddler_tests/tests/typesafe_health_returns_ok.rs b/paddler_tests/tests/typesafe_health_returns_ok.rs
new file mode 100644
index 000000000..db1bee4aa
--- /dev/null
+++ b/paddler_tests/tests/typesafe_health_returns_ok.rs
@@ -0,0 +1,26 @@
+use tokio_util::sync::CancellationToken;
+
+use paddler_client::reports_health::ReportsHealth as _;
+use paddler_tests::cluster_without_agents_serving_decisions::cluster_without_agents_serving_decisions;
+use paddler_tests::start_cluster::start_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn typesafe_health_returns_ok() {
+    let cluster = start_cluster(cluster_without_agents_serving_decisions())
+        .await
+        .expect("a decision balancer without agents must start");
+
+    let health = cluster
+        .compat_typesafe_health_client()
+        .expect("a decision cluster must serve TypeSafe compatibility")
+        .get_health(CancellationToken::new())
+        .await
+        .expect("failed to GET TypeSafe compat /health");
+
+    assert_eq!(health, "OK");
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/typesafe_models_is_unavailable_while_the_cluster_serves_no_decisions.rs b/paddler_tests/tests/typesafe_models_is_unavailable_while_the_cluster_serves_no_decisions.rs
new file mode 100644
index 000000000..94787b1f8
--- /dev/null
+++ b/paddler_tests/tests/typesafe_models_is_unavailable_while_the_cluster_serves_no_decisions.rs
@@ -0,0 +1,28 @@
+use reqwest::StatusCode;
+use serde_json::json;
+
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_tests::start_cluster::start_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn typesafe_models_is_unavailable_while_the_cluster_serves_no_decisions() {
+    let cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("a balancer serving text generation and embeddings must start");
+
+    let response = cluster
+        .typesafe_models()
+        .await
+        .expect("the models request must be answered");
+
+    assert_eq!(response.status, StatusCode::SERVICE_UNAVAILABLE);
+    assert_eq!(
+        response.body,
+        json!({"detail": "No pointer head is configured, so the cluster serves no decisions"})
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/typesafe_models_lists_nothing_while_no_agent_serves_the_desired_model.rs b/paddler_tests/tests/typesafe_models_lists_nothing_while_no_agent_serves_the_desired_model.rs
new file mode 100644
index 000000000..4f474cdea
--- /dev/null
+++ b/paddler_tests/tests/typesafe_models_lists_nothing_while_no_agent_serves_the_desired_model.rs
@@ -0,0 +1,44 @@
+use reqwest::StatusCode;
+use serde_json::json;
+
+use paddler_balancer::compatibility::typesafe_service::typesafe_header::TypeSafeHeader;
+use paddler_messaging::agent_desired_model::AgentDesiredModel;
+use paddler_messaging::balancer_desired_state::BalancerDesiredState;
+use paddler_test_cluster_harness::cluster_desired_state::ClusterDesiredState;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_tests::cluster_without_agents_serving_decisions::cluster_without_agents_serving_decisions;
+use paddler_tests::desired_state_serving_decisions::desired_state_serving_decisions;
+use paddler_tests::start_cluster::start_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn typesafe_models_lists_nothing_while_no_agent_serves_the_desired_model() {
+    let cluster = start_cluster(ClusterParams {
+        desired_state: ClusterDesiredState::Apply(Box::new(BalancerDesiredState {
+            model: AgentDesiredModel::Uri(
+                "https://huggingface.co/intentee/kev/blob/main/kev-0.8b.gguf".to_owned(),
+            ),
+            ..desired_state_serving_decisions()
+        })),
+        ..cluster_without_agents_serving_decisions()
+    })
+    .await
+    .expect("the cluster must start");
+    let response = cluster
+        .typesafe_models()
+        .await
+        .expect("the models request must be answered");
+
+    assert_eq!(response.status, StatusCode::OK);
+    assert!(
+        response
+            .headers
+            .get(TypeSafeHeader::REQUEST_ID)
+            .is_some_and(|request_id| !request_id.is_empty())
+    );
+    assert_eq!(response.body, json!({"models": []}));
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/typesafe_models_lists_the_model_the_cluster_serves.rs b/paddler_tests/tests/typesafe_models_lists_the_model_the_cluster_serves.rs
new file mode 100644
index 000000000..20fff53ff
--- /dev/null
+++ b/paddler_tests/tests/typesafe_models_lists_the_model_the_cluster_serves.rs
@@ -0,0 +1,50 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use reqwest::StatusCode;
+use serde_json::Value;
+use serde_json::json;
+
+use paddler_tests::decision_cluster_params::DecisionClusterParams;
+use paddler_tests::start_decision_cluster::start_decision_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn typesafe_models_lists_the_model_the_cluster_serves() {
+    let mut cluster = start_decision_cluster(DecisionClusterParams::default())
+        .await
+        .expect("the decision cluster must start");
+    let served_models = cluster
+        .agents_watcher
+        .until(|snapshot| {
+            snapshot
+                .agents
+                .iter()
+                .all(|agent| agent.status.model_path.is_some())
+        })
+        .await
+        .expect("every agent must report the model it serves")
+        .agents
+        .into_iter()
+        .filter_map(|agent| agent.status.model_path)
+        .map(|model_path| {
+            json!({
+                "name": model_path,
+                "description": format!("Decision model {model_path} served by Paddler"),
+                "release_date": "unknown",
+            })
+        })
+        .collect::<Vec<Value>>();
+    let response = cluster
+        .typesafe_models()
+        .await
+        .expect("the models request must be answered");
+
+    assert_eq!(
+        (response.status, response.body),
+        (StatusCode::OK, json!({ "models": served_models }))
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/typesafe_service_lets_browsers_send_and_read_typesafe_headers.rs b/paddler_tests/tests/typesafe_service_lets_browsers_send_and_read_typesafe_headers.rs
new file mode 100644
index 000000000..d59cf2431
--- /dev/null
+++ b/paddler_tests/tests/typesafe_service_lets_browsers_send_and_read_typesafe_headers.rs
@@ -0,0 +1,72 @@
+use reqwest::Client;
+use reqwest::Method;
+use reqwest::StatusCode;
+use reqwest::header::HeaderValue;
+
+use paddler_balancer::compatibility::typesafe_service::typesafe_api_path::TypeSafeApiPath;
+use paddler_balancer::compatibility::typesafe_service::typesafe_header::TypeSafeHeader;
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_tests::cluster_without_agents_serving_decisions::cluster_without_agents_serving_decisions;
+use paddler_tests::start_cluster::start_cluster;
+
+const ALLOWED_ORIGIN: &str = "http://example.com";
+
+#[tokio::test(flavor = "multi_thread")]
+async fn typesafe_service_lets_browsers_send_and_read_typesafe_headers() {
+    let cluster = start_cluster(ClusterParams {
+        inference_cors_allowed_hosts: vec![ALLOWED_ORIGIN.to_owned()],
+        ..cluster_without_agents_serving_decisions()
+    })
+    .await
+    .expect("a decision balancer without agents must start");
+
+    let preflight = Client::new()
+        .request(
+            Method::OPTIONS,
+            cluster
+                .balancer
+                .compat_typesafe_url(TypeSafeApiPath::SYSTEM_ONE)
+                .expect("the cluster serves TypeSafe"),
+        )
+        .header("Origin", ALLOWED_ORIGIN)
+        .header("Access-Control-Request-Method", "POST")
+        .header(
+            "Access-Control-Request-Headers",
+            [
+                "authorization",
+                "content-type",
+                TypeSafeHeader::REQUEST_ID,
+                TypeSafeHeader::RETRY_COUNT,
+                TypeSafeHeader::RUNTIME,
+                TypeSafeHeader::SDK,
+            ]
+            .join(", "),
+        )
+        .send()
+        .await
+        .expect("the preflight request must be answered");
+
+    assert_eq!(preflight.status(), StatusCode::OK);
+
+    let models = Client::new()
+        .get(
+            cluster
+                .balancer
+                .compat_typesafe_url(TypeSafeApiPath::MODELS)
+                .expect("the cluster serves TypeSafe"),
+        )
+        .header("Origin", ALLOWED_ORIGIN)
+        .send()
+        .await
+        .expect("the models request must be answered");
+
+    assert_eq!(
+        models.headers().get("access-control-expose-headers"),
+        Some(&HeaderValue::from_static(TypeSafeHeader::REQUEST_ID))
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/typesafe_system_one_answers_a_body_that_is_not_json_as_unprocessable.rs b/paddler_tests/tests/typesafe_system_one_answers_a_body_that_is_not_json_as_unprocessable.rs
new file mode 100644
index 000000000..eb59bb65c
--- /dev/null
+++ b/paddler_tests/tests/typesafe_system_one_answers_a_body_that_is_not_json_as_unprocessable.rs
@@ -0,0 +1,33 @@
+use reqwest::Client;
+use reqwest::StatusCode;
+
+use paddler_balancer::compatibility::typesafe_service::typesafe_api_path::TypeSafeApiPath;
+use paddler_tests::cluster_without_agents_serving_decisions::cluster_without_agents_serving_decisions;
+use paddler_tests::start_cluster::start_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn typesafe_system_one_answers_a_body_that_is_not_json_as_unprocessable() {
+    let cluster = start_cluster(cluster_without_agents_serving_decisions())
+        .await
+        .expect("a decision balancer without agents must start");
+
+    let not_json = Client::new()
+        .post(
+            cluster
+                .balancer
+                .compat_typesafe_url(TypeSafeApiPath::SYSTEM_ONE)
+                .expect("the cluster serves TypeSafe"),
+        )
+        .header("content-type", "application/json")
+        .body("not json")
+        .send()
+        .await
+        .expect("the request must be answered");
+
+    assert_eq!(not_json.status(), StatusCode::UNPROCESSABLE_ENTITY);
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/typesafe_system_one_answers_an_untranslatable_request_as_unprocessable.rs b/paddler_tests/tests/typesafe_system_one_answers_an_untranslatable_request_as_unprocessable.rs
new file mode 100644
index 000000000..28bc6a629
--- /dev/null
+++ b/paddler_tests/tests/typesafe_system_one_answers_an_untranslatable_request_as_unprocessable.rs
@@ -0,0 +1,37 @@
+use reqwest::StatusCode;
+use serde_json::json;
+
+use paddler_messaging::request_params_validation_error::RequestParamsValidationError;
+use paddler_test_cluster_harness::system_one_request::system_one_request;
+use paddler_tests::cluster_without_agents_serving_decisions::cluster_without_agents_serving_decisions;
+use paddler_tests::start_cluster::start_cluster;
+use paddler_typesafe_translation::typesafe_translation_error::TypeSafeTranslationError;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn typesafe_system_one_answers_an_untranslatable_request_as_unprocessable() {
+    let cluster = start_cluster(cluster_without_agents_serving_decisions())
+        .await
+        .expect("a decision balancer without agents must start");
+
+    let without_questions = cluster
+        .typesafe_system_one(
+            "untranslatable-request",
+            &system_one_request(&json!("state"), &json!({})),
+        )
+        .await
+        .expect("the request must be answered");
+
+    assert_eq!(without_questions.status, StatusCode::UNPROCESSABLE_ENTITY);
+    assert_eq!(
+        without_questions.body,
+        json!({"detail": TypeSafeTranslationError::DecisionInvalid {
+            source: RequestParamsValidationError::DecisionWithoutQuestions,
+        }
+        .to_string()})
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/typesafe_system_one_answers_questions_in_their_order.rs b/paddler_tests/tests/typesafe_system_one_answers_questions_in_their_order.rs
new file mode 100644
index 000000000..6169c50f1
--- /dev/null
+++ b/paddler_tests/tests/typesafe_system_one_answers_questions_in_their_order.rs
@@ -0,0 +1,28 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use reqwest::StatusCode;
+
+use crate::typesafe_tone_and_payment_decision::typesafe_tone_and_payment_decision;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn typesafe_system_one_answers_questions_in_their_order() {
+    let response = typesafe_tone_and_payment_decision().await;
+
+    assert_eq!(response.status, StatusCode::OK);
+
+    let answers = response.body["answers"]
+        .as_object()
+        .expect("the answers must be an object");
+
+    assert_eq!(answers.keys().collect::<Vec<_>>(), vec!["tone", "paid"]);
+    assert_eq!(answers["tone"]["type"], "choice");
+    assert_eq!(
+        answers["tone"]["probabilities"]
+            .as_object()
+            .expect("the choice probabilities must be an object")
+            .keys()
+            .collect::<Vec<_>>(),
+        vec!["calm", "angry"]
+    );
+    assert_eq!(answers["paid"]["type"], "noul");
+}
diff --git a/paddler_tests/tests/typesafe_system_one_answers_with_the_decision_model.rs b/paddler_tests/tests/typesafe_system_one_answers_with_the_decision_model.rs
new file mode 100644
index 000000000..1084469c6
--- /dev/null
+++ b/paddler_tests/tests/typesafe_system_one_answers_with_the_decision_model.rs
@@ -0,0 +1,74 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use reqwest::StatusCode;
+use serde_json::Value;
+use serde_json::json;
+
+use paddler_test_cluster_harness::system_one_request::system_one_request;
+use paddler_tests::decision_cluster_params::DecisionClusterParams;
+use paddler_tests::start_decision_cluster::start_decision_cluster;
+
+const ROUNDED_PROBABILITY_SUM_TOLERANCE: f64 = 0.002;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn typesafe_system_one_answers_with_the_decision_model() {
+    let cluster = start_decision_cluster(DecisionClusterParams::default())
+        .await
+        .expect("the cluster must start");
+
+    let response = cluster
+        .typesafe_system_one(
+            "model-request",
+            &system_one_request(
+                &json!({"customer": "Ada", "message": "I was charged twice. Please help."}),
+                &json!({
+                    "billing": {"type": "noul", "instructions": "Is this about billing?"},
+                    "tone": {
+                        "type": "choice",
+                        "instructions": "What is the tone?",
+                        "criteria": {"calm": null, "angry": null},
+                    },
+                    "urgency": {
+                        "type": "score",
+                        "criteria": ["Can wait", "Needs attention today"],
+                    },
+                }),
+            ),
+        )
+        .await
+        .expect("the System One request must be answered");
+
+    assert_eq!(response.status, StatusCode::OK);
+
+    let body = &response.body;
+    let answers = &body["answers"];
+
+    assert!(
+        answers["billing"]["noul"]
+            .as_f64()
+            .is_some_and(|noul| (0.0..=1.0).contains(&noul))
+    );
+    assert!(["calm", "angry"].contains(&answers["tone"]["choice"].as_str().unwrap_or_default()));
+    assert!(
+        (answers["urgency"]["probabilities"]
+            .as_object()
+            .expect("the score probabilities must be an object")
+            .values()
+            .filter_map(Value::as_f64)
+            .sum::<f64>()
+            - 1.0)
+            .abs()
+            < ROUNDED_PROBABILITY_SUM_TOLERANCE
+    );
+    assert!(
+        body["usage"]["input_tokens"]
+            .as_u64()
+            .is_some_and(|input_tokens| input_tokens > 0)
+    );
+    assert_eq!(body["usage"]["output_tokens"], json!(0));
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/typesafe_system_one_client_disconnecting_mid_decision_releases_its_sequences.rs b/paddler_tests/tests/typesafe_system_one_client_disconnecting_mid_decision_releases_its_sequences.rs
new file mode 100644
index 000000000..fe1ed980a
--- /dev/null
+++ b/paddler_tests/tests/typesafe_system_one_client_disconnecting_mid_decision_releases_its_sequences.rs
@@ -0,0 +1,67 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use reqwest::StatusCode;
+use serde_json::Map;
+use serde_json::Value;
+use serde_json::json;
+
+use paddler_balancer::compatibility::typesafe_service::typesafe_api_path::TypeSafeApiPath;
+use paddler_messaging::decision_slots_minimum::DECISION_SLOTS_MINIMUM;
+use paddler_test_cluster_harness::half_closed_client::HalfClosedClient;
+use paddler_test_cluster_harness::noul_system_one_request::noul_system_one_request;
+use paddler_test_cluster_harness::system_one_request::system_one_request;
+use paddler_tests::decision_cluster_params::DecisionClusterParams;
+use paddler_tests::questions_outlasting_a_disconnect::QUESTIONS_OUTLASTING_A_DISCONNECT;
+use paddler_tests::sample_decision::sample_decision;
+use paddler_tests::start_decision_cluster::start_decision_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn typesafe_system_one_client_disconnecting_mid_decision_releases_its_sequences() {
+    let mut cluster = start_decision_cluster(DecisionClusterParams::default())
+        .await
+        .expect("the cluster must start");
+    let agent_id = cluster
+        .agent_ids
+        .first()
+        .expect("the cluster must have an agent")
+        .clone();
+    let questions: Map<String, Value> = (0..QUESTIONS_OUTLASTING_A_DISCONNECT)
+        .map(|question_index| (format!("paid-{question_index}"), json!({"type": "noul"})))
+        .collect();
+    let mut half_closed_client = HalfClosedClient::post_json_then_half_close(
+        cluster
+            .balancer
+            .compat_typesafe_addr()
+            .expect("the cluster must serve the TypeSafe API"),
+        TypeSafeApiPath::SYSTEM_ONE,
+        &system_one_request(&json!(sample_decision().state), &json!(questions)),
+    )
+    .await
+    .expect("the System One request must be sent");
+
+    cluster
+        .wait_for_slots_processing(&agent_id, u64::from(DECISION_SLOTS_MINIMUM.get()))
+        .await
+        .expect("the agent must start deciding");
+    half_closed_client
+        .half_close()
+        .await
+        .expect("the System One request must be half-closed");
+    cluster
+        .wait_for_slots_processing(&agent_id, 0)
+        .await
+        .expect("the agent must release the abandoned decision");
+    drop(half_closed_client);
+
+    let response = cluster
+        .typesafe_system_one("after-the-abandoned-decision", &noul_system_one_request())
+        .await
+        .expect("a System One request after the abandoned one must be answered");
+
+    assert_eq!(response.status, StatusCode::OK);
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/typesafe_system_one_echoes_the_model_it_was_asked_for.rs b/paddler_tests/tests/typesafe_system_one_echoes_the_model_it_was_asked_for.rs
new file mode 100644
index 000000000..8aa6b7104
--- /dev/null
+++ b/paddler_tests/tests/typesafe_system_one_echoes_the_model_it_was_asked_for.rs
@@ -0,0 +1,12 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use paddler_test_cluster_harness::typesafe_request_model::TYPESAFE_REQUEST_MODEL;
+
+use crate::typesafe_tone_and_payment_decision::typesafe_tone_and_payment_decision;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn typesafe_system_one_echoes_the_model_it_was_asked_for() {
+    let response = typesafe_tone_and_payment_decision().await;
+
+    assert_eq!(response.body["model"], TYPESAFE_REQUEST_MODEL);
+}
diff --git a/paddler_tests/tests/typesafe_system_one_echoes_the_request_id_it_was_given.rs b/paddler_tests/tests/typesafe_system_one_echoes_the_request_id_it_was_given.rs
new file mode 100644
index 000000000..4fff25c46
--- /dev/null
+++ b/paddler_tests/tests/typesafe_system_one_echoes_the_request_id_it_was_given.rs
@@ -0,0 +1,20 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use paddler_balancer::compatibility::typesafe_service::typesafe_header::TypeSafeHeader;
+
+use crate::typesafe_tone_and_payment_decision::TYPESAFE_TONE_AND_PAYMENT_REQUEST_ID;
+use crate::typesafe_tone_and_payment_decision::typesafe_tone_and_payment_decision;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn typesafe_system_one_echoes_the_request_id_it_was_given() {
+    let response = typesafe_tone_and_payment_decision().await;
+
+    assert_eq!(
+        response.headers.get(TypeSafeHeader::REQUEST_ID),
+        Some(
+            &TYPESAFE_TONE_AND_PAYMENT_REQUEST_ID
+                .parse()
+                .expect("the header value is valid")
+        )
+    );
+}
diff --git a/paddler_tests/tests/typesafe_system_one_is_unavailable_while_the_cluster_serves_no_decisions.rs b/paddler_tests/tests/typesafe_system_one_is_unavailable_while_the_cluster_serves_no_decisions.rs
new file mode 100644
index 000000000..9c07ab978
--- /dev/null
+++ b/paddler_tests/tests/typesafe_system_one_is_unavailable_while_the_cluster_serves_no_decisions.rs
@@ -0,0 +1,32 @@
+use reqwest::StatusCode;
+use serde_json::json;
+
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_test_cluster_harness::noul_system_one_request::noul_system_one_request;
+use paddler_tests::start_cluster::start_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn typesafe_system_one_is_unavailable_while_the_cluster_serves_no_decisions() {
+    let cluster = start_cluster(ClusterParams::without_agents())
+        .await
+        .expect("a balancer serving text generation and embeddings must start");
+
+    let response = cluster
+        .typesafe_system_one(
+            "text-generation-cluster-request",
+            &noul_system_one_request(),
+        )
+        .await
+        .expect("the System One request must be answered");
+
+    assert_eq!(response.status, StatusCode::SERVICE_UNAVAILABLE);
+    assert_eq!(
+        response.body,
+        json!({"detail": "No pointer head is configured, so the cluster serves no decisions"})
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/typesafe_system_one_rejects_a_request_without_a_model.rs b/paddler_tests/tests/typesafe_system_one_rejects_a_request_without_a_model.rs
new file mode 100644
index 000000000..510e5b7ca
--- /dev/null
+++ b/paddler_tests/tests/typesafe_system_one_rejects_a_request_without_a_model.rs
@@ -0,0 +1,38 @@
+use reqwest::StatusCode;
+use serde_json::json;
+
+use paddler_test_cluster_harness::noul_system_one_request::noul_system_one_request;
+use paddler_tests::cluster_without_agents_serving_decisions::cluster_without_agents_serving_decisions;
+use paddler_tests::start_cluster::start_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn typesafe_system_one_rejects_a_request_without_a_model() {
+    let cluster = start_cluster(cluster_without_agents_serving_decisions())
+        .await
+        .expect("a decision balancer without agents must start");
+
+    let mut request_without_a_model = noul_system_one_request();
+
+    request_without_a_model
+        .as_object_mut()
+        .expect("the System One request must be a JSON object")
+        .remove("model");
+
+    let response = cluster
+        .typesafe_system_one("without-a-model", &request_without_a_model)
+        .await
+        .expect("the request must be answered");
+
+    assert_eq!(
+        (response.status, response.body),
+        (
+            StatusCode::UNPROCESSABLE_ENTITY,
+            json!({"detail": "Json deserialize error: missing field `model` at line 1 column 54"})
+        )
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/typesafe_system_one_reports_a_buffered_request_timeout_as_a_gateway_timeout.rs b/paddler_tests/tests/typesafe_system_one_reports_a_buffered_request_timeout_as_a_gateway_timeout.rs
new file mode 100644
index 000000000..1264a74cf
--- /dev/null
+++ b/paddler_tests/tests/typesafe_system_one_reports_a_buffered_request_timeout_as_a_gateway_timeout.rs
@@ -0,0 +1,31 @@
+use reqwest::StatusCode;
+
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_test_cluster_harness::noul_system_one_request::noul_system_one_request;
+use paddler_tests::cluster_without_agents_serving_decisions::cluster_without_agents_serving_decisions;
+use paddler_tests::short_buffered_request_timeout::SHORT_BUFFERED_REQUEST_TIMEOUT;
+use paddler_tests::start_cluster::start_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn typesafe_system_one_reports_a_buffered_request_timeout_as_a_gateway_timeout() {
+    let cluster = start_cluster(ClusterParams {
+        buffered_request_timeout: SHORT_BUFFERED_REQUEST_TIMEOUT,
+        ..cluster_without_agents_serving_decisions()
+    })
+    .await
+    .expect("a decision balancer without agents must start");
+
+    assert_eq!(
+        cluster
+            .typesafe_system_one("waiting-request", &noul_system_one_request())
+            .await
+            .expect("the request must be answered")
+            .status,
+        StatusCode::GATEWAY_TIMEOUT
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/typesafe_system_one_reports_a_decision_exceeding_the_context_as_unprocessable.rs b/paddler_tests/tests/typesafe_system_one_reports_a_decision_exceeding_the_context_as_unprocessable.rs
new file mode 100644
index 000000000..325646d4f
--- /dev/null
+++ b/paddler_tests/tests/typesafe_system_one_reports_a_decision_exceeding_the_context_as_unprocessable.rs
@@ -0,0 +1,47 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use reqwest::StatusCode;
+use serde_json::json;
+
+use paddler_messaging::decision_slots_minimum::DECISION_SLOTS_MINIMUM;
+use paddler_test_cluster_harness::noul_system_one_request_about::noul_system_one_request_about;
+use paddler_tests::oversized_text::oversized_text;
+use paddler_tests::start_decision_cluster::start_decision_cluster;
+
+use crate::small_context_decision_cluster_params::SMALL_DECISION_CONTEXT_SIZE;
+use crate::small_context_decision_cluster_params::small_context_decision_cluster_params;
+
+const TOKENS_REQUIRED_BY_THE_OVERSIZED_STATE_DECISION: usize = 410;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn typesafe_system_one_reports_a_decision_exceeding_the_context_as_unprocessable() {
+    let cluster = start_decision_cluster(small_context_decision_cluster_params())
+        .await
+        .expect("the cluster must start");
+
+    let response = cluster
+        .typesafe_system_one(
+            "oversized-decision",
+            &noul_system_one_request_about(&json!(oversized_text())),
+        )
+        .await
+        .expect("the request must be answered");
+
+    assert_eq!(
+        (response.status, response.body),
+        (
+            StatusCode::UNPROCESSABLE_ENTITY,
+            json!({
+                "detail": format!(
+                    "the request needs {TOKENS_REQUIRED_BY_THE_OVERSIZED_STATE_DECISION} tokens in the cache, more than the {} the context holds",
+                    SMALL_DECISION_CONTEXT_SIZE / u32::from(DECISION_SLOTS_MINIMUM.get())
+                )
+            })
+        )
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/typesafe_system_one_reports_a_full_buffer_as_service_unavailable.rs b/paddler_tests/tests/typesafe_system_one_reports_a_full_buffer_as_service_unavailable.rs
new file mode 100644
index 000000000..dce977e0b
--- /dev/null
+++ b/paddler_tests/tests/typesafe_system_one_reports_a_full_buffer_as_service_unavailable.rs
@@ -0,0 +1,30 @@
+use reqwest::StatusCode;
+
+use paddler_test_cluster_harness::cluster_params::ClusterParams;
+use paddler_test_cluster_harness::noul_system_one_request::noul_system_one_request;
+use paddler_tests::cluster_without_agents_serving_decisions::cluster_without_agents_serving_decisions;
+use paddler_tests::start_cluster::start_cluster;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn typesafe_system_one_reports_a_full_buffer_as_service_unavailable() {
+    let cluster = start_cluster(ClusterParams {
+        max_buffered_requests: 0,
+        ..cluster_without_agents_serving_decisions()
+    })
+    .await
+    .expect("a decision balancer without agents must start");
+
+    assert_eq!(
+        cluster
+            .typesafe_system_one("overflowing-request", &noul_system_one_request())
+            .await
+            .expect("the request must be answered")
+            .status,
+        StatusCode::SERVICE_UNAVAILABLE
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/typesafe_system_one_reports_a_state_with_a_nul_byte_as_unprocessable.rs b/paddler_tests/tests/typesafe_system_one_reports_a_state_with_a_nul_byte_as_unprocessable.rs
new file mode 100644
index 000000000..a3313f5ae
--- /dev/null
+++ b/paddler_tests/tests/typesafe_system_one_reports_a_state_with_a_nul_byte_as_unprocessable.rs
@@ -0,0 +1,52 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::ffi::CString;
+
+use llama_cpp_bindings::error::StringToTokenError;
+use reqwest::StatusCode;
+use serde_json::json;
+
+use paddler_agent_decision::decision_error::DecisionError;
+use paddler_agent_runtime::rejection_description::rejection_description;
+use paddler_test_cluster_harness::noul_system_one_request_about::noul_system_one_request_about;
+use paddler_tests::decision_cluster_params::DecisionClusterParams;
+use paddler_tests::start_decision_cluster::start_decision_cluster;
+
+use crate::with_nul_byte::with_nul_byte;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn typesafe_system_one_reports_a_state_with_a_nul_byte_as_unprocessable() {
+    let cluster = start_decision_cluster(DecisionClusterParams::default())
+        .await
+        .expect("the cluster must start");
+    let state_with_nul_byte = with_nul_byte("state");
+
+    let response = cluster
+        .typesafe_system_one(
+            "state-with-a-nul-byte",
+            &noul_system_one_request_about(&json!(state_with_nul_byte)),
+        )
+        .await
+        .expect("the request must be answered");
+
+    assert_eq!(
+        (response.status, response.body),
+        (
+            StatusCode::UNPROCESSABLE_ENTITY,
+            json!({
+                "detail": rejection_description(
+                    Some("test-agent-0"),
+                    &DecisionError::InputTokenizationFailed(StringToTokenError::NulError(
+                        CString::new(state_with_nul_byte)
+                            .expect_err("the state must contain a nul byte")
+                    ))
+                )
+            })
+        )
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/typesafe_system_one_reports_answers_that_do_not_match_its_questions_as_a_bad_gateway.rs b/paddler_tests/tests/typesafe_system_one_reports_answers_that_do_not_match_its_questions_as_a_bad_gateway.rs
new file mode 100644
index 000000000..0bf2b08bc
--- /dev/null
+++ b/paddler_tests/tests/typesafe_system_one_reports_answers_that_do_not_match_its_questions_as_a_bad_gateway.rs
@@ -0,0 +1,95 @@
+use reqwest::StatusCode;
+use serde_json::json;
+use tokio::spawn;
+
+use paddler_messaging::agent_decisions_serving::AgentDecisionsServing;
+use paddler_messaging::agent_runtime_status::AgentRuntimeStatus;
+use paddler_messaging::agent_status::AgentStatus;
+use paddler_messaging::decision_answer::DecisionAnswer;
+use paddler_messaging::decision_result::DecisionResult;
+use paddler_messaging::decision_summary::DecisionSummary;
+use paddler_messaging::management_socket::agent::response::Response as AgentJsonRpcResponse;
+use paddler_messaging::slot_aggregated_status_snapshot::SlotAggregatedStatusSnapshot;
+use paddler_test_cluster_harness::noul_system_one_request::noul_system_one_request;
+use paddler_tests::cluster_without_agents_serving_decisions::cluster_without_agents_serving_decisions;
+use paddler_tests::start_cluster::start_cluster;
+use paddler_typesafe_translation::typesafe_translation_error::TypeSafeTranslationError;
+
+use crate::untrusted_management_agent::UntrustedManagementAgent;
+
+const DECISION_SLOTS: u16 = 2;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn typesafe_system_one_reports_answers_that_do_not_match_its_questions_as_a_bad_gateway() {
+    let cluster = start_cluster(cluster_without_agents_serving_decisions())
+        .await
+        .expect("a decision balancer without agents must start");
+    let mut untruthful_agent = UntrustedManagementAgent::connect_registered(
+        cluster.balancer.addresses.management,
+        "untruthful",
+        SlotAggregatedStatusSnapshot {
+            status: AgentStatus {
+                desired_slots_total: DECISION_SLOTS,
+                runtime: AgentRuntimeStatus::Serving {
+                    decisions: AgentDecisionsServing::Served,
+                    slots_total: u64::from(DECISION_SLOTS),
+                },
+                ..AgentStatus::default()
+            },
+            version: 1,
+        },
+    )
+    .await
+    .expect("the untruthful agent must register");
+
+    let typesafe_response =
+        spawn(cluster.typesafe_system_one("mismatched-answers", &noul_system_one_request()));
+    let decision_request_id = untruthful_agent
+        .next_request()
+        .await
+        .expect("the balancer must dispatch the decision")
+        .id;
+
+    for decision_result in [
+        DecisionResult::QuestionAnswered(DecisionAnswer {
+            id: "unasked".to_owned(),
+            probabilities: vec![0.5, 0.5],
+        }),
+        DecisionResult::Done(DecisionSummary {
+            input_tokens: 1,
+            processing_milliseconds: 1,
+        }),
+    ] {
+        untruthful_agent
+            .respond(
+                decision_request_id.clone(),
+                AgentJsonRpcResponse::Decision(decision_result),
+            )
+            .await
+            .expect("the decision result must be sent");
+    }
+
+    let response = typesafe_response
+        .await
+        .expect("the TypeSafe request task must not panic")
+        .expect("the request must be answered");
+
+    assert_eq!(
+        (response.status, response.body),
+        (
+            StatusCode::BAD_GATEWAY,
+            json!({
+                "detail": TypeSafeTranslationError::AnswerOutOfOrder {
+                    expected_question_id: "paid".to_owned(),
+                    received_question_id: "unasked".to_owned(),
+                }
+                .to_string()
+            })
+        )
+    );
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+}
diff --git a/paddler_tests/tests/typesafe_system_one_reports_that_a_decision_produces_no_output_tokens.rs b/paddler_tests/tests/typesafe_system_one_reports_that_a_decision_produces_no_output_tokens.rs
new file mode 100644
index 000000000..f74db2647
--- /dev/null
+++ b/paddler_tests/tests/typesafe_system_one_reports_that_a_decision_produces_no_output_tokens.rs
@@ -0,0 +1,10 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use crate::typesafe_tone_and_payment_decision::typesafe_tone_and_payment_decision;
+
+#[tokio::test(flavor = "multi_thread")]
+async fn typesafe_system_one_reports_that_a_decision_produces_no_output_tokens() {
+    let response = typesafe_tone_and_payment_decision().await;
+
+    assert_eq!(response.body["usage"]["output_tokens"], 0);
+}
diff --git a/paddler_tests/tests/typesafe_tone_and_payment_decision.rs b/paddler_tests/tests/typesafe_tone_and_payment_decision.rs
new file mode 100644
index 000000000..af8dc901a
--- /dev/null
+++ b/paddler_tests/tests/typesafe_tone_and_payment_decision.rs
@@ -0,0 +1,36 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use serde_json::json;
+
+use paddler_test_cluster_harness::system_one_request::system_one_request;
+use paddler_test_cluster_harness::typesafe_api_response::TypeSafeApiResponse;
+use paddler_tests::decision_cluster_params::DecisionClusterParams;
+use paddler_tests::start_decision_cluster::start_decision_cluster;
+
+pub const TYPESAFE_TONE_AND_PAYMENT_REQUEST_ID: &str = "typesafe-request";
+
+pub async fn typesafe_tone_and_payment_decision() -> TypeSafeApiResponse {
+    let cluster = start_decision_cluster(DecisionClusterParams::default())
+        .await
+        .expect("the cluster must start");
+    let typesafe_decision = cluster
+        .typesafe_system_one(
+            TYPESAFE_TONE_AND_PAYMENT_REQUEST_ID,
+            &system_one_request(
+                &json!("The invoice was paid on time."),
+                &json!({
+                    "tone": {"type": "choice", "criteria": {"calm": null, "angry": "Shouting"}},
+                    "paid": {"type": "noul", "instructions": "Was it paid?"},
+                }),
+            ),
+        )
+        .await
+        .expect("the System One request must be answered");
+
+    cluster
+        .shutdown()
+        .await
+        .expect("the cluster must shut down cleanly");
+
+    typesafe_decision
+}
diff --git a/paddler_tests/tests/untrusted_management_agent.rs b/paddler_tests/tests/untrusted_management_agent.rs
new file mode 100644
index 000000000..65dcf49ff
--- /dev/null
+++ b/paddler_tests/tests/untrusted_management_agent.rs
@@ -0,0 +1,111 @@
+use std::net::SocketAddr;
+
+use serde_json::from_str;
+use serde_json::to_string;
+use tokio_tungstenite::tungstenite::Error as WebSocketError;
+use tokio_tungstenite::tungstenite::Message;
+
+use paddler_messaging::jsonrpc::request_envelope::RequestEnvelope;
+use paddler_messaging::jsonrpc::response_envelope::ResponseEnvelope;
+use paddler_messaging::management_socket::agent::message::Message as AgentJsonRpcMessage;
+use paddler_messaging::management_socket::agent::request::Request as AgentJsonRpcRequest;
+use paddler_messaging::management_socket::agent::response::Response as AgentJsonRpcResponse;
+use paddler_messaging::management_socket::balancer::message::Message as ManagementJsonRpcMessage;
+use paddler_messaging::management_socket::balancer::notification::Notification as ManagementJsonRpcNotification;
+use paddler_messaging::management_socket::balancer::notification_params::register_agent_params::RegisterAgentParams;
+use paddler_messaging::management_socket::balancer::notification_params::update_agent_status_params::UpdateAgentStatusParams;
+use paddler_messaging::slot_aggregated_status_snapshot::SlotAggregatedStatusSnapshot;
+use paddler_test_cluster_harness::cluster_harness_error::ClusterHarnessError;
+use paddler_test_cluster_harness::untrusted_agent_socket_client::UntrustedAgentSocketClient;
+
+pub struct UntrustedManagementAgent {
+    pub socket_client: UntrustedAgentSocketClient,
+}
+
+impl UntrustedManagementAgent {
+    pub async fn connect(
+        management_addr: SocketAddr,
+        agent_id: &str,
+    ) -> Result<Self, WebSocketError> {
+        UntrustedAgentSocketClient::connect(management_addr, agent_id)
+            .await
+            .map(|socket_client| Self { socket_client })
+    }
+
+    pub async fn connect_registered(
+        management_addr: SocketAddr,
+        agent_id: &str,
+        slot_aggregated_status_snapshot: SlotAggregatedStatusSnapshot,
+    ) -> Result<Self, ClusterHarnessError> {
+        let mut untrusted_management_agent = Self::connect(management_addr, agent_id)
+            .await
+            .expect("the untrusted agent must reach the agent socket");
+        let registration = untrusted_management_agent
+            .register(RegisterAgentParams {
+                name: Some(agent_id.to_owned()),
+                slot_aggregated_status_snapshot,
+            })
+            .await;
+
+        registration.map(|()| untrusted_management_agent)
+    }
+
+    pub async fn next_request(
+        &mut self,
+    ) -> Result<RequestEnvelope<AgentJsonRpcRequest>, ClusterHarnessError> {
+        loop {
+            if let AgentJsonRpcMessage::Request(request_envelope) =
+                from_str(&self.socket_client.next_text().await?)
+                    .expect("the balancer must send a decodable agent message")
+            {
+                return Ok(request_envelope);
+            }
+        }
+    }
+
+    pub async fn register(
+        &mut self,
+        register_agent_params: RegisterAgentParams,
+    ) -> Result<(), ClusterHarnessError> {
+        self.send_management_message(&ManagementJsonRpcMessage::Notification(
+            ManagementJsonRpcNotification::RegisterAgent(register_agent_params),
+        ))
+        .await
+    }
+
+    pub async fn report_status(
+        &mut self,
+        slot_aggregated_status_snapshot: SlotAggregatedStatusSnapshot,
+    ) -> Result<(), ClusterHarnessError> {
+        self.send_management_message(&ManagementJsonRpcMessage::Notification(
+            ManagementJsonRpcNotification::UpdateAgentStatus(UpdateAgentStatusParams {
+                slot_aggregated_status_snapshot,
+            }),
+        ))
+        .await
+    }
+
+    pub async fn respond(
+        &mut self,
+        request_id: String,
+        response: AgentJsonRpcResponse,
+    ) -> Result<(), ClusterHarnessError> {
+        self.send_management_message(&ManagementJsonRpcMessage::Response(ResponseEnvelope {
+            generated_by: None,
+            request_id,
+            response,
+        }))
+        .await
+    }
+
+    async fn send_management_message(
+        &mut self,
+        message: &ManagementJsonRpcMessage,
+    ) -> Result<(), ClusterHarnessError> {
+        self.socket_client
+            .send(Message::text(
+                to_string(message).expect("a management message must serialize"),
+            ))
+            .await
+    }
+}
diff --git a/paddler_tests/tests/vocabulary_only_llama_model.rs b/paddler_tests/tests/vocabulary_only_llama_model.rs
new file mode 100644
index 000000000..94113c9fb
--- /dev/null
+++ b/paddler_tests/tests/vocabulary_only_llama_model.rs
@@ -0,0 +1,48 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+use std::ops::Deref;
+use std::path::Path;
+use std::sync::Arc;
+
+use llama_cpp_bindings::llama_backend::LlamaBackend;
+use llama_cpp_bindings::model::LlamaModel;
+use llama_cpp_bindings::model::params::LlamaModelParams;
+
+use paddler_agent_runtime::shared_llama_backend::SharedLlamaBackend;
+
+pub struct VocabularyOnlyLlamaModel {
+    llama_model: LlamaModel,
+    _llama_backend: Arc<LlamaBackend>,
+}
+
+impl VocabularyOnlyLlamaModel {
+    pub fn load(shared_llama_backend: &SharedLlamaBackend, model_path: &Path) -> Self {
+        let llama_backend = shared_llama_backend
+            .get_or_init()
+            .expect("the llama backend must initialize");
+
+        #[expect(
+            clippy::disallowed_methods,
+            reason = "a vocabulary-only load allocates no tensors, so it holds no VRAM residency"
+        )]
+        let llama_model = LlamaModel::load_from_file(
+            &llama_backend,
+            model_path,
+            &LlamaModelParams::default().with_vocab_only(true),
+        )
+        .expect("the model vocabulary must load");
+
+        Self {
+            llama_model,
+            _llama_backend: llama_backend,
+        }
+    }
+}
+
+impl Deref for VocabularyOnlyLlamaModel {
+    type Target = LlamaModel;
+
+    fn deref(&self) -> &LlamaModel {
+        &self.llama_model
+    }
+}
diff --git a/paddler_tests/tests/with_nul_byte.rs b/paddler_tests/tests/with_nul_byte.rs
new file mode 100644
index 000000000..0376ba32a
--- /dev/null
+++ b/paddler_tests/tests/with_nul_byte.rs
@@ -0,0 +1,5 @@
+#![cfg(feature = "tests_that_use_llms")]
+
+pub fn with_nul_byte(text: &str) -> String {
+    format!("{text} \0 byte")
+}
diff --git a/paddler_tool_call_validator/src/tool_call_validator.rs b/paddler_tool_call_validator/src/tool_call_validator.rs
index 71c77f1de..ed54dfb8d 100644
--- a/paddler_tool_call_validator/src/tool_call_validator.rs
+++ b/paddler_tool_call_validator/src/tool_call_validator.rs
@@ -8,7 +8,6 @@ use serde_json::Value;
 
 use paddler_messaging::request_params::continue_from_conversation_history_params::tool::Tool;
 use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters::Parameters;
-use paddler_messaging::request_params::continue_from_conversation_history_params::tool::tool_params::function_call::parameters_schema::validated_parameters_schema::ValidatedParametersSchema;
 
 use crate::tool_call_validation_error::ToolCallValidationError;
 
@@ -22,9 +21,7 @@ pub struct ToolCallValidator {
 }
 
 impl ToolCallValidator {
-    pub fn from_tools(
-        tools: &[Tool<ValidatedParametersSchema>],
-    ) -> Result<Self, ToolCallValidationError> {
+    pub fn from_tools(tools: &[Tool<Value>]) -> Result<Self, ToolCallValidationError> {
         let mut strategies = HashMap::with_capacity(tools.len());
 
         for tool in tools {
@@ -34,7 +31,7 @@ impl ToolCallValidator {
             let strategy = match &function.parameters {
                 Parameters::Empty => ValidationStrategy::JsonObjectOnly,
                 Parameters::Schema(schema) => {
-                    let compiled = validator_for(&Value::from(schema)).map_err(|source| {
+                    let compiled = validator_for(schema).map_err(|source| {
                         ToolCallValidationError::InvalidSchema {
                             tool_name: function.name.clone(),
                             source: Box::new(source),
@@ -100,7 +97,9 @@ impl ToolCallValidator {
 
 #[cfg(test)]
 mod tests {
+    use llama_cpp_bindings_types::ModelProvidedToolCallId;
     use llama_cpp_bindings_types::ParsedToolCall;
+    use llama_cpp_bindings_types::ParsedToolCallId;
     use llama_cpp_bindings_types::ToolCallArguments;
     use serde_json::Map;
     use serde_json::Value;
@@ -115,18 +114,22 @@ use paddler_messaging::request_params::continue_from_conversation_history_params
     use super::ToolCallValidator;
     use crate::tool_call_validation_error::ToolCallValidationError;
 
+    fn model_provided_id() -> ParsedToolCallId {
+        ParsedToolCallId::ModelProvided(ModelProvidedToolCallId::try_from("id".to_owned()).unwrap())
+    }
+
     fn valid_json_arguments(value: Value) -> ToolCallArguments {
         ToolCallArguments::ValidJson(value)
     }
 
-    fn weather_tool_with_schema() -> Tool<ValidatedParametersSchema> {
+    fn weather_tool_with_schema() -> Tool<Value> {
         let mut properties = Map::new();
         properties.insert(
             "location".to_owned(),
             json!({"type": "string", "description": "city"}),
         );
 
-        Tool::Function(FunctionCall {
+        Tool::from(Tool::Function(FunctionCall {
             function: Function {
                 name: "get_weather".to_owned(),
                 description: "fetch weather".to_owned(),
@@ -137,24 +140,24 @@ use paddler_messaging::request_params::continue_from_conversation_history_params
                     additional_properties: Some(Value::Bool(false)),
                 }),
             },
-        })
+        }))
     }
 
-    fn schemaless_tool() -> Tool<ValidatedParametersSchema> {
-        Tool::Function(FunctionCall {
+    fn schemaless_tool() -> Tool<Value> {
+        Tool::from(Tool::<ValidatedParametersSchema>::Function(FunctionCall {
             function: Function {
                 name: "freeform".to_owned(),
                 description: "tool with no schema".to_owned(),
                 parameters: Parameters::Empty,
             },
-        })
+        }))
     }
 
     #[test]
     fn schema_validator_accepts_matching_arguments() {
         let validator = ToolCallValidator::from_tools(&[weather_tool_with_schema()]).unwrap();
         let parsed = ParsedToolCall::new(
-            "id".to_owned(),
+            model_provided_id(),
             "get_weather".to_owned(),
             valid_json_arguments(json!({"location": "Paris"})),
         );
@@ -166,7 +169,7 @@ use paddler_messaging::request_params::continue_from_conversation_history_params
     fn schema_validator_rejects_missing_required_field() {
         let validator = ToolCallValidator::from_tools(&[weather_tool_with_schema()]).unwrap();
         let parsed = ParsedToolCall::new(
-            "id".to_owned(),
+            model_provided_id(),
             "get_weather".to_owned(),
             valid_json_arguments(json!({})),
         );
@@ -183,7 +186,7 @@ use paddler_messaging::request_params::continue_from_conversation_history_params
     fn schema_validator_rejects_wrong_type() {
         let validator = ToolCallValidator::from_tools(&[weather_tool_with_schema()]).unwrap();
         let parsed = ParsedToolCall::new(
-            "id".to_owned(),
+            model_provided_id(),
             "get_weather".to_owned(),
             valid_json_arguments(json!({"location": 42})),
         );
@@ -200,7 +203,7 @@ use paddler_messaging::request_params::continue_from_conversation_history_params
     fn unknown_tool_name_returns_error() {
         let validator = ToolCallValidator::from_tools(&[weather_tool_with_schema()]).unwrap();
         let parsed = ParsedToolCall::new(
-            "id".to_owned(),
+            model_provided_id(),
             "set_thermostat".to_owned(),
             valid_json_arguments(json!({"value": 21})),
         );
@@ -217,7 +220,7 @@ use paddler_messaging::request_params::continue_from_conversation_history_params
     fn rejects_arguments_that_are_not_json() {
         let validator = ToolCallValidator::from_tools(&[weather_tool_with_schema()]).unwrap();
         let parsed = ParsedToolCall::new(
-            "id".to_owned(),
+            model_provided_id(),
             "get_weather".to_owned(),
             ToolCallArguments::InvalidJson("not json".to_owned()),
         );
@@ -235,7 +238,7 @@ use paddler_messaging::request_params::continue_from_conversation_history_params
     fn schemaless_tool_rejects_arguments_that_are_not_an_object() {
         let validator = ToolCallValidator::from_tools(&[schemaless_tool()]).unwrap();
         let parsed = ParsedToolCall::new(
-            "id".to_owned(),
+            model_provided_id(),
             "freeform".to_owned(),
             valid_json_arguments(json!([1, 2])),
         );
@@ -253,7 +256,7 @@ use paddler_messaging::request_params::continue_from_conversation_history_params
     fn schemaless_tool_accepts_any_object() {
         let validator = ToolCallValidator::from_tools(&[schemaless_tool()]).unwrap();
         let parsed = ParsedToolCall::new(
-            "id".to_owned(),
+            model_provided_id(),
             "freeform".to_owned(),
             valid_json_arguments(json!({"x": 1, "y": 2})),
         );
@@ -277,7 +280,7 @@ use paddler_messaging::request_params::continue_from_conversation_history_params
     fn empty_tools_yields_validator_that_rejects_any_call() {
         let validator = ToolCallValidator::from_tools(&[]).unwrap();
         let parsed = ParsedToolCall::new(
-            "id".to_owned(),
+            model_provided_id(),
             "anything".to_owned(),
             valid_json_arguments(json!({})),
         );
@@ -290,11 +293,11 @@ use paddler_messaging::request_params::continue_from_conversation_history_params
         ));
     }
 
-    fn tool_with_invalid_property_schema() -> Tool<ValidatedParametersSchema> {
+    fn tool_with_invalid_property_schema() -> Tool<Value> {
         let mut properties = Map::new();
         properties.insert("location".to_owned(), json!({"type": 42}));
 
-        Tool::Function(FunctionCall {
+        Tool::from(Tool::Function(FunctionCall {
             function: Function {
                 name: "broken_tool".to_owned(),
                 description: "tool whose property schema is not valid JSON Schema".to_owned(),
@@ -305,7 +308,7 @@ use paddler_messaging::request_params::continue_from_conversation_history_params
                     additional_properties: None,
                 }),
             },
-        })
+        }))
     }
 
     #[test]
@@ -320,8 +323,8 @@ use paddler_messaging::request_params::continue_from_conversation_history_params
         ));
     }
 
-    fn tool_with_invalid_additional_properties_schema() -> Tool<ValidatedParametersSchema> {
-        Tool::Function(FunctionCall {
+    fn tool_with_invalid_additional_properties_schema() -> Tool<Value> {
+        Tool::from(Tool::Function(FunctionCall {
             function: Function {
                 name: "broken_additional".to_owned(),
                 description: "tool whose additionalProperties schema is invalid".to_owned(),
@@ -332,7 +335,7 @@ use paddler_messaging::request_params::continue_from_conversation_history_params
                     additional_properties: Some(json!({"type": "not_a_type"})),
                 }),
             },
-        })
+        }))
     }
 
     #[test]
diff --git a/paddler_typesafe_client_python_test/.gitignore b/paddler_typesafe_client_python_test/.gitignore
new file mode 100644
index 000000000..521016d76
--- /dev/null
+++ b/paddler_typesafe_client_python_test/.gitignore
@@ -0,0 +1,4 @@
+__pycache__/
+.pytest_cache/
+.mypy_cache/
+.venv/
diff --git a/paddler_typesafe_client_python_test/poetry.lock b/paddler_typesafe_client_python_test/poetry.lock
new file mode 100644
index 000000000..e63a92693
--- /dev/null
+++ b/paddler_typesafe_client_python_test/poetry.lock
@@ -0,0 +1,817 @@
+# This file is automatically @generated by Poetry 2.4.3 and should not be changed by hand.
+
+[[package]]
+name = "annotated-types"
+version = "0.8.0"
+description = "Reusable constraint types to use with typing.Annotated"
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "annotated_types-0.8.0-py3-none-any.whl", hash = "sha256:f072f4d804ea359e4eaf198b1af7a8b0943881a87f31bb764f8bf219bb9419e0"},
+    {file = "annotated_types-0.8.0.tar.gz", hash = "sha256:13b2beaad985e05e2d6407ee4c4f35590b11f8d693a258a561055cac8f64cab7"},
+]
+
+[[package]]
+name = "anyio"
+version = "4.15.1"
+description = "High-level concurrency and networking framework on top of asyncio or Trio"
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "anyio-4.15.1-py3-none-any.whl", hash = "sha256:6152fdbbf9a77fdec97731721bebf7c4c44f7c29b424b0065826173efc7ed101"},
+    {file = "anyio-4.15.1.tar.gz", hash = "sha256:9f28306018cbd6d329e64a36d58256edff76dd996fe423bc957326e578b82a94"},
+]
+
+[package.dependencies]
+idna = ">=2.8"
+typing_extensions = {version = ">=4.16.0", markers = "python_version < \"3.15\""}
+
+[package.extras]
+trio = ["trio (>=0.32.0)"]
+
+[[package]]
+name = "ast-serialize"
+version = "0.13.1"
+description = "Python bindings for mypy AST serialization"
+optional = false
+python-versions = ">=3.7"
+groups = ["dev"]
+files = [
+    {file = "ast_serialize-0.13.1-cp314-cp314-pyemscripten_2026_0_wasm32.whl", hash = "sha256:3e129d44b0cffbea24a7f40117ce49e30993c3d119e584d4817ebd6f87246456"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-macosx_10_12_x86_64.whl", hash = "sha256:523d45bab16ddaf1399c04c4fe81bbf180f81e33c2f7bcbe76be56b0f7311ab4"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:638333fb8ae72c62753f9e794fda20290883ddc65b0d85e26e08f8bb227567ad"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:cc34c670747d85d4232834665e247009413eaebe47c69eaf60a5c439aaecaed9"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:8c8ebd419811c435812d242fb2a173e57dfede198c9ac9a0a584666ebc9b9331"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:8899c038b1c15ae8b98e379057fb3ef398ceb0b21d1737a8156dd5a0923b7e51"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:e14c4de98f965647a74e6eaf37e20207124e45451ccd36e84a6f6ebab45e0372"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:7f3e8a3eff69af640db985360fef35073fb46956dfc3e5c614cf00920d2746dd"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-manylinux_2_31_riscv64.whl", hash = "sha256:75d7750fe98f6f35a96c43aba8d98a2e623e22c778a0c17ab9ede92eb10f5d82"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:43e8ddd63e80fc280c2f790157386f27f9b9f6ccb9c4f8e803b863cbfb085769"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:9221e09496fb942901993c1812ee87950cc0ecbb514c7c0f1f5fa6d78f701ed8"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-musllinux_1_2_armv7l.whl", hash = "sha256:7a5c38bdb93669c596f94fb7204dd6ebe4fadb19058ac9a2faf2948c2a175f96"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-musllinux_1_2_i686.whl", hash = "sha256:acf4e9f055d75058ab01536da330648746f2f82b7697a85230547f2d5ebe895c"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-musllinux_1_2_ppc64le.whl", hash = "sha256:87427913c325adc811e81e7c9a757b061f47ef1e29114ac479aa0953a2687a98"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-musllinux_1_2_riscv64.whl", hash = "sha256:ad88921f95b92bc144d9ad059a0e16f8c1ae69f5239f2d3c6536fd2774f73dfe"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:f8b17e2921e56621bcb46792bd3e0b82ca3b85de78d5936e7b8174ba1758b0aa"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-win32.whl", hash = "sha256:2f9a9ce8b0994f757b941f7f7fe384a6666b4162b941ebc5c248081baea18e47"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-win_amd64.whl", hash = "sha256:5d99910e6b9a9b9360bcca2ae99aee36cc912ce058df2a57d1c33099b6cb2593"},
+    {file = "ast_serialize-0.13.1-cp314-cp314t-win_arm64.whl", hash = "sha256:a789cbe14d0a0e9ec9b32a456e5abc81ad8799479c3cfd88a29e3442262aebe6"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-macosx_10_12_x86_64.whl", hash = "sha256:23d768369b89d551077dc173da6892fad00443338417523b6732f4e8ba4a7ed5"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-macosx_11_0_arm64.whl", hash = "sha256:68f8829d114dab6c9db4a8bbb8d0e342958faee1b5e15a29022678df639960f8"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:4e385b17d5bde892a6184003a9aff3518387908b73f8f5ea539313c07c8c0ede"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:02686b281ec9b8e9614f6f0db27e12e4cac515f336c5d05b678bc9eb73ff2af5"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:a8ab2c2585b2e22e48cd4aff2136fce4eab4ced569b85f150ef9a5a5d6f34d49"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:c652d9f5d52d8b81eff3ff2c95c9dd0f281de45f6ad5d3fa95527550c396454a"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:121e1d08d58b75c15491cc3003609cbad71ac2b0016043c36d30a6b2032b6605"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-manylinux_2_31_riscv64.whl", hash = "sha256:30cf89b6a6844542d116c758725d13de5b97b9410a51ee577715f761270d0c1c"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:732aaae9e498c973fc63fd7f308009b57c4afa6ff59d0b674d894e57d0bca01a"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-musllinux_1_2_aarch64.whl", hash = "sha256:d9c1a920b2aa0c1e016bed7721689717034ca0dba64ac3a3ff243b003d4e8dd8"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-musllinux_1_2_armv7l.whl", hash = "sha256:847cad4c22953823f5e76d0b126c6f9da670cfa45b1df41b9324e22189d77611"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-musllinux_1_2_i686.whl", hash = "sha256:aca07de3ca8e3d9f2b716c3c31f4f384fe4bd2d03e09912417cff7a3e08036f7"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-musllinux_1_2_ppc64le.whl", hash = "sha256:5db12d843d931b9d1f1519ce66b9cbfe72d8563b75c1204d255322254d1964dd"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-musllinux_1_2_riscv64.whl", hash = "sha256:fb8021f4dc3296de18737e3b8b34994411da725374d74a0b9eb32a738f2a23d0"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-musllinux_1_2_x86_64.whl", hash = "sha256:0d764913374ba21df8cd618045b371ed104ce55c6276de7c59e554a4f5ba26fd"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-win32.whl", hash = "sha256:2bad302ae2427454ef9dab71bd5c3ad67188860c39d36daee38f72031edb39c4"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-win_amd64.whl", hash = "sha256:032b09d86c785c3984e975378b7861f2a68ac44e22ba7b07b93a536b2524eeb8"},
+    {file = "ast_serialize-0.13.1-cp315-abi3.abi3t-win_arm64.whl", hash = "sha256:929eda0c419e622ab49df3a4c45f81c57dcc10874bbad55d0470a4a03324dac3"},
+    {file = "ast_serialize-0.13.1-cp315-cp315-pyemscripten_2026_5_wasm32.whl", hash = "sha256:bf53a3ba1cbe45399fc1e4ccd684b8f2a11cca450f89706b79122f478534c6f5"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-macosx_10_12_x86_64.whl", hash = "sha256:4bcf1d5f4eefecaa325a34ecf079b2741189e7d0821523f92cb343a389a67489"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-macosx_11_0_arm64.whl", hash = "sha256:1b158f1c9e3a0d386913924c57995d6df78224868a4d0f215bdd41587c97b34a"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:77a6201b69ce42e919f2fd9a559138d5410b5d8afff7c3a6b1f500b0481ff7d8"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:6eb0be64c9b8f1ed26637eef345abe75ea90351c92c85a8aec7838b3af971011"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:7f2427bddfca208e84de700044fadf9aef2769a70bae740a1bdabb89ee9e1117"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:88df960e67184a71874e5b2ee014f87d1254a0dbc1f903c7bd7b1ab7733ecaef"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:9f99786e0d62082a5b121b3437f3167562133711df22128de1e9886f57f0b226"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-manylinux_2_31_riscv64.whl", hash = "sha256:a01b515b25a03184eff4acef25b981ae007916379682e908fc8785b342a0bf3e"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:29422d7bcd9e7f39411571c36cec45ce8a7a9b3ad4b2e7aa64497c15a9e397f6"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:6ffb0f548a7dc126067fdfb7b2325ab5dfcdeecc29faf5e0f0461c96d6ebfa3d"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-musllinux_1_2_armv7l.whl", hash = "sha256:2876444a5f87a7c76f371af3c07c2fda563a08bc8af52568df812e8aba4bf568"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-musllinux_1_2_i686.whl", hash = "sha256:885a62ec6481ce58ab24c6357360d7a82023b62dd7e5e7ceccf59e9662a3272b"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-musllinux_1_2_ppc64le.whl", hash = "sha256:2015fd0165e463aac3b1aadc6a1fa1e5820c14ca950821f4fe02f8a3f6a388d6"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-musllinux_1_2_riscv64.whl", hash = "sha256:eb41e1906502ac80d24ff343f45008552a3f5b5158607959fe6a07dbfc6eb037"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:ca9acc67f796aa2554fa3037dccc022a4b08b38b2ad153edfac9c9bb7cb1f3c6"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-win32.whl", hash = "sha256:4f98da562e5901d69c7dd30885594c5438f859edf354e9ac7517d188fb026c3a"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-win_amd64.whl", hash = "sha256:ddf7cbc9c8abf70102c4a088efe154bb86bfb825b0a10e556d2b15972cf99bd7"},
+    {file = "ast_serialize-0.13.1-cp39-abi3-win_arm64.whl", hash = "sha256:48f11d96c9ec71119612f8e79a8c55c12f2b136716d68347433c561d27a25f43"},
+    {file = "ast_serialize-0.13.1.tar.gz", hash = "sha256:401cc84e8888d4d76f2c8d296fa006ed1a5ff993b6aff90dd777c03973efd025"},
+]
+
+[[package]]
+name = "colorama"
+version = "0.4.6"
+description = "Cross-platform colored terminal text."
+optional = false
+python-versions = "!=3.0.*,!=3.1.*,!=3.2.*,!=3.3.*,!=3.4.*,!=3.5.*,!=3.6.*,>=2.7"
+groups = ["dev"]
+markers = "sys_platform == \"win32\""
+files = [
+    {file = "colorama-0.4.6-py2.py3-none-any.whl", hash = "sha256:4f1d9991f5acc0ca119f9d443620b77f9d6b33703e51011c16baf57afb285fc6"},
+    {file = "colorama-0.4.6.tar.gz", hash = "sha256:08695f5cb7ed6e0531a20572697297273c47b8cae5a63ffc6d6ed5c201be6e44"},
+]
+
+[[package]]
+name = "h11"
+version = "0.16.0"
+description = "A pure-Python, bring-your-own-I/O implementation of HTTP/1.1"
+optional = false
+python-versions = ">=3.8"
+groups = ["main"]
+markers = "sys_platform != \"emscripten\""
+files = [
+    {file = "h11-0.16.0-py3-none-any.whl", hash = "sha256:63cf8bbe7522de3bf65932fda1d9c2772064ffb3dae62d55932da54b31cb6c86"},
+    {file = "h11-0.16.0.tar.gz", hash = "sha256:4e35b956cf45792e4caa5885e69fba00bdbc6ffafbfa020300e549b208ee5ff1"},
+]
+
+[[package]]
+name = "httpcore2"
+version = "2.13.1"
+description = "A minimal low-level HTTP client."
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+markers = "sys_platform != \"emscripten\""
+files = [
+    {file = "httpcore2-2.13.1-py3-none-any.whl", hash = "sha256:e1e05d4f25f7d7d496bfb96748f6f4b67657b03da069b3a68c36069f3db73d0a"},
+    {file = "httpcore2-2.13.1.tar.gz", hash = "sha256:e0aa977abe17e69a3b820a24542a6fa88702676d83880b8d194dcd18408e5103"},
+]
+
+[package.dependencies]
+h11 = ">=0.16"
+truststore = ">=0.10"
+
+[package.extras]
+asyncio = ["anyio (>=4.5.0,<5.0)"]
+http2 = ["h2 (>=3,<5)"]
+socks = ["socksio (==1.*)"]
+trio = ["trio (>=0.34.0,<1.0)"]
+
+[[package]]
+name = "httpx2"
+version = "2.13.1"
+description = "The next generation HTTP client."
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "httpx2-2.13.1-py3-none-any.whl", hash = "sha256:6dff50fabc270ee5fd25d845d0b078ed20564579744d6d962850975996d2f9a4"},
+    {file = "httpx2-2.13.1.tar.gz", hash = "sha256:e48744a19e3af5ee48313d0ce5fe941d5422fae5705ea922a4aabf94d7800dfa"},
+]
+
+[package.dependencies]
+anyio = {version = ">=4.10", markers = "sys_platform != \"emscripten\""}
+httpcore2 = {version = "2.13.1", markers = "sys_platform != \"emscripten\""}
+httpx2-jsfetch = {version = "*", markers = "sys_platform == \"emscripten\" and python_version >= \"3.12\""}
+idna = ">=3.18"
+truststore = {version = ">=0.10", markers = "sys_platform != \"emscripten\""}
+typing-extensions = {version = ">=4.5.0", markers = "python_version < \"3.13\""}
+
+[package.extras]
+brotli = ["brotli (>=1.2.0) ; platform_python_implementation == \"CPython\"", "brotlicffi (>=1.2.0.2) ; platform_python_implementation != \"CPython\""]
+cli = ["click (>=8.4.2)", "pygments (==2.*)", "rich (>=10,<16)"]
+http2 = ["h2 (>=3,<5)"]
+socks = ["socksio (==1.*)"]
+ws = ["wsproto (>=1.2)"]
+zstd = ["backports-zstd (>=1.0.0) ; python_version <= \"3.13\""]
+
+[[package]]
+name = "httpx2-jsfetch"
+version = "1.0"
+description = "httpx2 transports for Emscripten/Pyodide, backed by the JavaScript fetch API."
+optional = false
+python-versions = ">=3.12"
+groups = ["main"]
+markers = "sys_platform == \"emscripten\" and python_version >= \"3.12\""
+files = [
+    {file = "httpx2_jsfetch-1.0-py3-none-any.whl", hash = "sha256:cb916b707601e69a07721aabc8f3f6659be3a6893bc1ff5c6f9e02241df2da32"},
+    {file = "httpx2_jsfetch-1.0.tar.gz", hash = "sha256:70a0e3eabfef7cce5ad9c629f7d01ca05e418f586646f4ddf14782e4c1454c60"},
+]
+
+[[package]]
+name = "idna"
+version = "3.20"
+description = "Internationalized Domain Names in Applications (IDNA)"
+optional = false
+python-versions = ">=3.9"
+groups = ["main"]
+files = [
+    {file = "idna-3.20-py3-none-any.whl", hash = "sha256:ab7ae7122974553370f0bdb919e1a960b2cd1bc1ef0276416d896db81c14582c"},
+    {file = "idna-3.20.tar.gz", hash = "sha256:a7db850025b95ded1eae8a46181a1a6c56c92c96f0e2b005d9ff8dc0210cab44"},
+]
+
+[package.extras]
+all = ["coverage (>=7.10.0)", "hypothesis (>=6.141.1)", "mypy (>=1.11.2)", "pytest (>=8.3.2)", "ruff (>=0.16.0)", "ty (>=0.0.37)"]
+
+[[package]]
+name = "iniconfig"
+version = "2.3.0"
+description = "brain-dead simple config-ini parsing"
+optional = false
+python-versions = ">=3.10"
+groups = ["dev"]
+files = [
+    {file = "iniconfig-2.3.0-py3-none-any.whl", hash = "sha256:f631c04d2c48c52b84d0d0549c99ff3859c98df65b3101406327ecc7d53fbf12"},
+    {file = "iniconfig-2.3.0.tar.gz", hash = "sha256:c76315c77db068650d49c5b56314774a7804df16fee4402c1f19d6d15d8c4730"},
+]
+
+[[package]]
+name = "librt"
+version = "0.16.0"
+description = "Mypyc runtime library"
+optional = false
+python-versions = ">=3.9"
+groups = ["dev"]
+markers = "platform_python_implementation != \"PyPy\""
+files = [
+    {file = "librt-0.16.0-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:96f576f2711f8519152ec76d0e599243555c1f07679fa73606ca8c8c868c0be6"},
+    {file = "librt-0.16.0-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:2bec3818c7da7c96ceae0ef5915a3d16c52dd08f3ea913bf1fe8568c447c7978"},
+    {file = "librt-0.16.0-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:64c79520414a3fdfc6aabd7593e6169afa14d5f8d9908d4b498db068868b08dd"},
+    {file = "librt-0.16.0-cp310-cp310-manylinux2014_i686.manylinux_2_17_i686.manylinux_2_28_i686.whl", hash = "sha256:f81b5b19ce748ef68d4746656b7929762eb2fe99269b266e4be07e2ee4de7144"},
+    {file = "librt-0.16.0-cp310-cp310-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:c71d1b76210a36729fedfc5115069b50a3d8619054745f8758fe5d6f19e86671"},
+    {file = "librt-0.16.0-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:473eebc7866bb0a0c8849a292b5e7157c1aba5d14d0f0f610c52158d6d964262"},
+    {file = "librt-0.16.0-cp310-cp310-manylinux_2_34_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:4eb1313a19847089ee81e88742abedf285c60538816640b99742d8534b81d26a"},
+    {file = "librt-0.16.0-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:ebefd60b42e2a82b32d136bb5f7c94eadfcd29f772b547df6f3291d1ed855a1c"},
+    {file = "librt-0.16.0-cp310-cp310-musllinux_1_2_i686.whl", hash = "sha256:6c5da27e8056439f927ea896735da60e616c477a8293feaa3233d4e7781a6726"},
+    {file = "librt-0.16.0-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:67e718c7a43f8db325abbbf1404e2d535f12f8f7a1a82259568385cc5274b82a"},
+    {file = "librt-0.16.0-cp310-cp310-musllinux_1_2_riscv64.whl", hash = "sha256:314e703f0c19320dc8094e7a784b9cf29e1b67402515abf580069a6363c0b4f1"},
+    {file = "librt-0.16.0-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:943c6bbecbdf7fa575a4f2952fcfd848c88ef95507c3fca411e89d4ac3ff8143"},
+    {file = "librt-0.16.0-cp310-cp310-win32.whl", hash = "sha256:7cc365f006891afb006b52d5ee5ee74c09306ffa20e2f8705a32a4450af2f3ba"},
+    {file = "librt-0.16.0-cp310-cp310-win_amd64.whl", hash = "sha256:0314058469f4d2fd279ce7c62ac274ac82c3918ef7db62ef0697c4c359370155"},
+    {file = "librt-0.16.0-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:fe4372c52d4849096c6cc1cda2817d293ec51440c890474ed59ef38d46556f18"},
+    {file = "librt-0.16.0-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:c72c5295a84bd249526da9bdca38f2e176d15c31c13bb0063c5053f4ca023421"},
+    {file = "librt-0.16.0-cp311-cp311-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:be56ba9c884143495b517f23fe794ae367d58cd89ea0fdd6d437e3c024a87f9f"},
+    {file = "librt-0.16.0-cp311-cp311-manylinux2014_i686.manylinux_2_17_i686.manylinux_2_28_i686.whl", hash = "sha256:ef46c1a29ffb8c72e882e22618ec618778eacd0578fb22c6e7cf9c11d15f357b"},
+    {file = "librt-0.16.0-cp311-cp311-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:d3c94211ee0c4f8d649ec06b7c115c0ec4eadb873a0e3154ca15cef3f814b071"},
+    {file = "librt-0.16.0-cp311-cp311-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:94aed6a8308818b91677957d1bd03188869cd7aeb23c5dba7912a6c0402f7602"},
+    {file = "librt-0.16.0-cp311-cp311-manylinux_2_34_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:349c0bcb87ebd07481b6ff781e25cdc699723dbe2212e57dabb27f7a13b7b87d"},
+    {file = "librt-0.16.0-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:001bfd59a7d45b17e3e75f2a8c6405280b35e7b84471792778e718c4f368950e"},
+    {file = "librt-0.16.0-cp311-cp311-musllinux_1_2_i686.whl", hash = "sha256:28e038895b998d7a0c7798922ce8a1dc157675df5cf1c9ef0aca809ed804b7a1"},
+    {file = "librt-0.16.0-cp311-cp311-musllinux_1_2_ppc64le.whl", hash = "sha256:0dbe4096a7ecc00fa835d24510ad8545a4efef738dac96e0e63516783ccde905"},
+    {file = "librt-0.16.0-cp311-cp311-musllinux_1_2_riscv64.whl", hash = "sha256:5cd5b092441053364af968ea12084692cb9d4a22f3ce9524e377880bf028761e"},
+    {file = "librt-0.16.0-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:3ddeb3c9dedb461bb457c6c7d9aa7fbf35329da313d1a7543d00c8d0f3473c96"},
+    {file = "librt-0.16.0-cp311-cp311-win32.whl", hash = "sha256:e05108e0849966f53a8d2d3112a7af881d0efaa479bc735bba91108f9f2350a7"},
+    {file = "librt-0.16.0-cp311-cp311-win_amd64.whl", hash = "sha256:5f49cff01bd608ef7d97104cb035c75455e79c2d70bf4a506cf773338ac1860d"},
+    {file = "librt-0.16.0-cp311-cp311-win_arm64.whl", hash = "sha256:1d28ae980ae2218f9c5b95d191e947296f918c9bf0b400d467a9430275bbe678"},
+    {file = "librt-0.16.0-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:fe52bf4641069e7978a14253b036cb9002def1926317e710f2e249f8a8c47742"},
+    {file = "librt-0.16.0-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:5bcc2c4726ced915b00de0c9856a4eeabfb3fddb93e10e0b8f735b7709358b6d"},
+    {file = "librt-0.16.0-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:ff7baa55f8e7c69851419e50a666015d02a74198716fd45c0125a2112e0a389f"},
+    {file = "librt-0.16.0-cp312-cp312-manylinux2014_i686.manylinux_2_17_i686.manylinux_2_28_i686.whl", hash = "sha256:b95d5d92ab83d39e760a52091bb1baba664f3a2351e39b1e16801e5747c2f0e9"},
+    {file = "librt-0.16.0-cp312-cp312-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:b6d085d70bce51d43c5c7c36d63490770180d8779e71c49305c87b4213918de7"},
+    {file = "librt-0.16.0-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:36e53948e99bbe3ffea257124cfcae1cfb01831555c9a9c903c9f9a72db7fd07"},
+    {file = "librt-0.16.0-cp312-cp312-manylinux_2_34_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:54d11f726aae9df5a6ffbbf0a03a52449bbac84a53ef03669cb41cdfd4ae41bf"},
+    {file = "librt-0.16.0-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:4323193ac0cd025f85af531df8ba91bf24d1973b401697347a6282e8fd3fcf5e"},
+    {file = "librt-0.16.0-cp312-cp312-musllinux_1_2_i686.whl", hash = "sha256:e42f8e098b9c5396fefa05fb1cc7e33b0e08fc51da106b5de4a45fd22aac6743"},
+    {file = "librt-0.16.0-cp312-cp312-musllinux_1_2_ppc64le.whl", hash = "sha256:39ec1d5a14e37baf1450a6cabf03fe552340808bf1ad9d71824ab90117716459"},
+    {file = "librt-0.16.0-cp312-cp312-musllinux_1_2_riscv64.whl", hash = "sha256:d1aabe3925cbb4a08d15b7b20ba4011b53019da0c4173a25155139b7b1baed65"},
+    {file = "librt-0.16.0-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:300c3ffdc459f4a779a8411ecb188e3ac0b1ff3a3a7b099642555dedae06c69b"},
+    {file = "librt-0.16.0-cp312-cp312-win32.whl", hash = "sha256:c17194318e4c0c0348b36f36c2ec7534436fe0a4c15582403162a4f08c80797a"},
+    {file = "librt-0.16.0-cp312-cp312-win_amd64.whl", hash = "sha256:25a58a19ea8d83b68209f04912df765e9260635ef77646542ed4b4abe6bc7940"},
+    {file = "librt-0.16.0-cp312-cp312-win_arm64.whl", hash = "sha256:f7be7cf555bc30ec12622e9447299cc4a9b8ff307548b634794353db0c2065dc"},
+    {file = "librt-0.16.0-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:c5e6144e68b577f157519f2ba88ca20e3ed61c29b00e5cdfa76cd2d45acf059a"},
+    {file = "librt-0.16.0-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:33f41443a1f4e1f099331b3d8120e409fbff84b9760bc1cc9ea496f37ddaa5cc"},
+    {file = "librt-0.16.0-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:7e510b7770bee609617a3374a96548eb114cae048023e3f049ee449e7ff2db32"},
+    {file = "librt-0.16.0-cp313-cp313-manylinux2014_i686.manylinux_2_17_i686.manylinux_2_28_i686.whl", hash = "sha256:efc49c462d4516b8a58b00b490078fa64689fd1fe66970cc190131d7afb8027e"},
+    {file = "librt-0.16.0-cp313-cp313-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:92caf82ebef5e12d21c72242b70d1e92536f1711cf2a727a4c276de4b4469087"},
+    {file = "librt-0.16.0-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:17bac7f7a16b328fff77e440287693eb017abde913595b5827ebccbc21ecd8a6"},
+    {file = "librt-0.16.0-cp313-cp313-manylinux_2_34_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:5b976054553670829985ed767feb78fb6bcede0175327c4844dd5c281c1be659"},
+    {file = "librt-0.16.0-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:0058f9d68721094105917254c72ac0569117bb7b13b9769cf45d26d89f9d21cd"},
+    {file = "librt-0.16.0-cp313-cp313-musllinux_1_2_i686.whl", hash = "sha256:30b7beaf3f4487b7d8adef1f158b49067cb4d5a19fa7a3bf31a4e7a820e435c5"},
+    {file = "librt-0.16.0-cp313-cp313-musllinux_1_2_ppc64le.whl", hash = "sha256:468df902df016a06eb0e40b0747dc8d14e47d7a38b18b63b1fb167d85cb94d63"},
+    {file = "librt-0.16.0-cp313-cp313-musllinux_1_2_riscv64.whl", hash = "sha256:aea7b1f2b125dad5de85f049136651bff256c883c65e6b9209b2da0a1ac3cdef"},
+    {file = "librt-0.16.0-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:a8afb6557920860b7a3a596eb804cf37e09e7cf8a803db2478c202acc72d8c2e"},
+    {file = "librt-0.16.0-cp313-cp313-pyemscripten_2025_0_wasm32.whl", hash = "sha256:77c7a2b4fe2c1369e0d5aa1cade26740a7b14be32fbc9a5535d617d20065c39d"},
+    {file = "librt-0.16.0-cp313-cp313-win32.whl", hash = "sha256:02d89c813d5ff74b17df72d3a34819d132cd168e56b81bf755b809bd9e46b8c4"},
+    {file = "librt-0.16.0-cp313-cp313-win_amd64.whl", hash = "sha256:14ed6ebe3e4f85f326d7920011ad30ff49ed9334e62cf88caef9ba973d9e3a92"},
+    {file = "librt-0.16.0-cp313-cp313-win_arm64.whl", hash = "sha256:83d4041a3d9b2fd053a8a4e1f22878b3e5833e2712956382d5c048d791454e91"},
+    {file = "librt-0.16.0-cp314-cp314-macosx_10_15_x86_64.whl", hash = "sha256:931a0bb0fcac88f263e269e46eb30ba8e21402cd3c62ca40cb97034c0693fab1"},
+    {file = "librt-0.16.0-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:1bc17e54e5305f8d40b7ca203671ff5a9e59c1d0f8ea0f625dcca53a3984de11"},
+    {file = "librt-0.16.0-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:877698bf6bca5721d8be345f2fe09778e40ecadea8b58c73075f2b1a53666bf2"},
+    {file = "librt-0.16.0-cp314-cp314-manylinux2014_i686.manylinux_2_17_i686.manylinux_2_28_i686.whl", hash = "sha256:5981c011b306781ce561e18e14230a14524a3d8109b97553666c942c18f31a96"},
+    {file = "librt-0.16.0-cp314-cp314-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:afced3dfc17cd805ecf7a3d77996a71cf5f2c75aa66eb0c21a9930f4fc992f86"},
+    {file = "librt-0.16.0-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:ca8052401c55d7511dda6760719fda7618067e83535d7d0010096d216c34b667"},
+    {file = "librt-0.16.0-cp314-cp314-manylinux_2_34_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:1e511762a074005bb0aa569166779834e75e438370226930d0ce1866d4b6a33b"},
+    {file = "librt-0.16.0-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:f1e8591bd8a5a628cd7f07954c6a1592359a878bf032957a8e9057a41d644311"},
+    {file = "librt-0.16.0-cp314-cp314-musllinux_1_2_i686.whl", hash = "sha256:a4aaefb4ba6c07e1aeebb2795c8958148f1d6f9af3b555b53d23d766edb6d67a"},
+    {file = "librt-0.16.0-cp314-cp314-musllinux_1_2_ppc64le.whl", hash = "sha256:d92db7a0f6aee44f1baee94750457e8d2d1c6ccea41842de6268d34e8dc7eddd"},
+    {file = "librt-0.16.0-cp314-cp314-musllinux_1_2_riscv64.whl", hash = "sha256:378dfaffb38e59c24a87cde5713cd865d51ff7383fa12947f3907f306ea1ca55"},
+    {file = "librt-0.16.0-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:3e0c39bdc85370422e8b637be76eb1fd07d30967551b03e62267dd156f553152"},
+    {file = "librt-0.16.0-cp314-cp314-pyemscripten_2026_0_wasm32.whl", hash = "sha256:1b384b90ab79a7bc30b566895809a636e0666f21f3cf12b54823d025b7e83839"},
+    {file = "librt-0.16.0-cp314-cp314-win32.whl", hash = "sha256:52327da75a94012e7f932f913d20d3876bed3c102be00e6c3e8600ff7bdd58a7"},
+    {file = "librt-0.16.0-cp314-cp314-win_amd64.whl", hash = "sha256:3f0b8114c44b2ac06ff5dacd08e07e8e807ff4f46083f2a1602685122559be41"},
+    {file = "librt-0.16.0-cp314-cp314-win_arm64.whl", hash = "sha256:8caf96a4ef8fb27d0ac0d1ad8337d26a240acd4a02fe4345d0a8f264753e8f99"},
+    {file = "librt-0.16.0-cp314-cp314t-macosx_10_15_x86_64.whl", hash = "sha256:953107e2f68d0f3512c48f898b0dbf0ce5cc52bba0f318d847c985dc555ee4cc"},
+    {file = "librt-0.16.0-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:ad37d5b9abd49c9a655dcda7ea52a8a752884062ef1ee71ae17c2f2a0f81fe6a"},
+    {file = "librt-0.16.0-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:2c4aa329c17bd1aaea4f6e89335d8ccd494b3a5830b6654462273e50e11023f0"},
+    {file = "librt-0.16.0-cp314-cp314t-manylinux2014_i686.manylinux_2_17_i686.manylinux_2_28_i686.whl", hash = "sha256:0ead24d2562a49473dddd9efef8581f020007eb0054389c3ee3ffad38b1ca4c9"},
+    {file = "librt-0.16.0-cp314-cp314t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:02118f56a9c36ddd07dfd9b919d9ecc117ba20a90987d56aa4c429fa34509188"},
+    {file = "librt-0.16.0-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:4e29522c62e28595ff7e324c6834ade51127707f0e255b18d1c1cf03d39c1048"},
+    {file = "librt-0.16.0-cp314-cp314t-manylinux_2_34_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:3ff4b2367926b69c6215635902cccb04048e73094e9862900d27cb2c6bbff143"},
+    {file = "librt-0.16.0-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:c6f1b27bf1632a7e016af9f145f82be95e1edd7721a646505c21059257cb5a04"},
+    {file = "librt-0.16.0-cp314-cp314t-musllinux_1_2_i686.whl", hash = "sha256:5696d7f52e7b37217cb3a8f92c744fe835942602fdd4c1a8bc4741d3bfdce15e"},
+    {file = "librt-0.16.0-cp314-cp314t-musllinux_1_2_ppc64le.whl", hash = "sha256:6072e92dd876ff6ceeb6cf371e35e51f479349837391341f479b08df4564242b"},
+    {file = "librt-0.16.0-cp314-cp314t-musllinux_1_2_riscv64.whl", hash = "sha256:39ca4f2f2fe05de8e63493da592d84311adabe5bef52b193851981da9816b302"},
+    {file = "librt-0.16.0-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:f9807485a908f00355820f18e91e045ffdcdc5adb68aaec40a1e2b88c5f7bba1"},
+    {file = "librt-0.16.0-cp314-cp314t-win32.whl", hash = "sha256:94be5cb7bca4df6201f4183e9e4fa2086c655283d20b38cd84500a69057575a7"},
+    {file = "librt-0.16.0-cp314-cp314t-win_amd64.whl", hash = "sha256:d46ca272b251d033dd4527b0dec5f261a28a52bd5fa0f99c117b0a1f8588cc2d"},
+    {file = "librt-0.16.0-cp314-cp314t-win_arm64.whl", hash = "sha256:b9d6d4b14e92d876f8026b54c20c445f36425214c1081dc76f74e40db386b82b"},
+    {file = "librt-0.16.0-cp315-cp315-macosx_10_15_x86_64.whl", hash = "sha256:6fe436af2eaf630474f491af5d032cbe45f93fcff5c3b9fe4ab194a7255b20ff"},
+    {file = "librt-0.16.0-cp315-cp315-macosx_11_0_arm64.whl", hash = "sha256:8ff5d26c529336be9bd7ae04483235d77778ee7d6444a95353102b542601ce81"},
+    {file = "librt-0.16.0-cp315-cp315-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:909d8e3c1faee44cb762b1c519ff8613dcc5ceae5c99987a00917b5a31fd1d6a"},
+    {file = "librt-0.16.0-cp315-cp315-manylinux2014_i686.manylinux_2_17_i686.manylinux_2_28_i686.whl", hash = "sha256:6d4a64283ee61824b5790de882bc68e2d9d7a5143537cb7a966f7354f71646d4"},
+    {file = "librt-0.16.0-cp315-cp315-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:5810ba811297fdf37a1531a57667cb8ace0842013ca8606bf9eb7c24cf4be154"},
+    {file = "librt-0.16.0-cp315-cp315-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:e56aaf8c167548dc8e5d6f3bd0f48dcdd299a23c73be3f744aab79d99e9c7f5d"},
+    {file = "librt-0.16.0-cp315-cp315-manylinux_2_34_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:8f36c58e33b304b525c6c9c5076399c6ebf1109e17b9051a05a407b091b9215b"},
+    {file = "librt-0.16.0-cp315-cp315-musllinux_1_2_aarch64.whl", hash = "sha256:242e00b3d4fa37c3d3c1ca5f5c9adb7d909ddb1eac9c41f2787320d00caa0af2"},
+    {file = "librt-0.16.0-cp315-cp315-musllinux_1_2_i686.whl", hash = "sha256:0253721561787b8df8443eb347b7a6461015354e5bdd37ee38a41fef220d2bb0"},
+    {file = "librt-0.16.0-cp315-cp315-musllinux_1_2_ppc64le.whl", hash = "sha256:3e483a8d69ede8067db70c0e83007423b6925de6fd53afed01d66160f2e9398c"},
+    {file = "librt-0.16.0-cp315-cp315-musllinux_1_2_riscv64.whl", hash = "sha256:69ba927445cfaaffb4081003ef5224c55a5c2ab67ef956f416ef744916e44121"},
+    {file = "librt-0.16.0-cp315-cp315-musllinux_1_2_x86_64.whl", hash = "sha256:d6a365f2ab45a984d0e00eee0dd17f599ceab8cadab6ea07b6111c8132fc0e42"},
+    {file = "librt-0.16.0-cp315-cp315-pyemscripten_2026_5_wasm32.whl", hash = "sha256:f01f3805f2dae4781c0c34b440e31740d082950bdaf89a6f601ad589a28af57a"},
+    {file = "librt-0.16.0-cp315-cp315-win32.whl", hash = "sha256:b0e3e721c75d2e79a76d4422c79d7ba705fe1bbafec907037fe7a657a480a0e3"},
+    {file = "librt-0.16.0-cp315-cp315-win_amd64.whl", hash = "sha256:bc02954b1295de798bbdb0b4e2d8a28c2117de8b5c73dcbeb27dc32572dfb971"},
+    {file = "librt-0.16.0-cp315-cp315-win_arm64.whl", hash = "sha256:c5db585d43449a5f54303d4b2774e45e1babd975cfe1630a3d708c0b80c3e560"},
+    {file = "librt-0.16.0-cp315-cp315t-macosx_10_15_x86_64.whl", hash = "sha256:f06c689cb14afd9b612727553a5ec5a40febf113ca41c4413a2b0b334285884b"},
+    {file = "librt-0.16.0-cp315-cp315t-macosx_11_0_arm64.whl", hash = "sha256:13b4e8aba90b0b1c82474e9844aa9ffe7ad3faa484350e1da64cb8188d903134"},
+    {file = "librt-0.16.0-cp315-cp315t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:5a269c46ae327d8e6f8c1f85f7516cb52c0fa48127565a1105a4f4a05ff2a0b4"},
+    {file = "librt-0.16.0-cp315-cp315t-manylinux2014_i686.manylinux_2_17_i686.manylinux_2_28_i686.whl", hash = "sha256:a33e0dae1f8592146a4764d54ce842b278732d21a84e17c3bbe6b1bc158a2248"},
+    {file = "librt-0.16.0-cp315-cp315t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:47ada6ea32636492c61aa8ad27ae3b9404bfe7a97e3ba946d1984236cc741da0"},
+    {file = "librt-0.16.0-cp315-cp315t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:c43bd6e642d8a248c114327f98dd25ac5a7cb5aa168ef02f0559b91874df16b8"},
+    {file = "librt-0.16.0-cp315-cp315t-manylinux_2_34_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:c3d1bb7841a816ace6449bb26d3f9560dbfa20e71c568d23f0f62bf1e68f50b1"},
+    {file = "librt-0.16.0-cp315-cp315t-musllinux_1_2_aarch64.whl", hash = "sha256:3931f7a3db322e7f44e02a280e3949326ce9579ad388ee8d691dc7c76da9fb70"},
+    {file = "librt-0.16.0-cp315-cp315t-musllinux_1_2_i686.whl", hash = "sha256:f4462528b6000afe8f16907b5c7c2553abf1df005ba5140e6eb394541c3624c3"},
+    {file = "librt-0.16.0-cp315-cp315t-musllinux_1_2_ppc64le.whl", hash = "sha256:80039ba9b6a7d5f1a0175a4cca6bbefead87bd854c80abad1cb30afe47a830db"},
+    {file = "librt-0.16.0-cp315-cp315t-musllinux_1_2_riscv64.whl", hash = "sha256:7a1d272724b581bb6bc769dfdafed6da2ecc9886ba2450311de55a4ac2e1e9cd"},
+    {file = "librt-0.16.0-cp315-cp315t-musllinux_1_2_x86_64.whl", hash = "sha256:fbe4fb8c5445f7496d7f7f6bb0807875d09d47e6771ffa175fb2df2895fb86ba"},
+    {file = "librt-0.16.0-cp315-cp315t-win32.whl", hash = "sha256:375bfe6b572a8f6cfc398709356046173bf27e64c4c5edaf5f7062f051fb4bf9"},
+    {file = "librt-0.16.0-cp315-cp315t-win_amd64.whl", hash = "sha256:bd3150023d3dc2bc70f3784e59ffa1140d56ddba3d8125b3d6f9f85221279bfc"},
+    {file = "librt-0.16.0-cp315-cp315t-win_arm64.whl", hash = "sha256:8ceafb70f2a4f0826f11031942e59c0728fd98da112dc346d4352bde1e486866"},
+    {file = "librt-0.16.0-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:e1967e36ac4cae0c7e9615ad32e1a513cdacff79f9e8afb28bedc91caf48b4f3"},
+    {file = "librt-0.16.0-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:20fe0bf9053885e21c62b3e091fb5e73e1c54d1daf2e70eb388d70763bcd4220"},
+    {file = "librt-0.16.0-cp39-cp39-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:6a63610fa76524edfa605b5b259a603915c7a6e10e54f003e5503030506e81de"},
+    {file = "librt-0.16.0-cp39-cp39-manylinux2014_i686.manylinux_2_17_i686.manylinux_2_28_i686.whl", hash = "sha256:845a511b60ca43b9880dcc84a9784c891d6a2098c829130b320846c69c9c0c68"},
+    {file = "librt-0.16.0-cp39-cp39-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:d608f0bf3b8cbddd0067fe02cb8cab7d13e8eb9386b23a1843d4363044fd9e22"},
+    {file = "librt-0.16.0-cp39-cp39-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:df183721229ae51eef90108c115b43e98cb169b6155d34480338e5fc6616df00"},
+    {file = "librt-0.16.0-cp39-cp39-manylinux_2_34_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:4b6183e2e2e0ee00aac2ec07c7f7d151c97e85666b304f574b71cff0f9fccc4e"},
+    {file = "librt-0.16.0-cp39-cp39-musllinux_1_2_aarch64.whl", hash = "sha256:273d00be33792a15189331df10f1f1331621b043881e66c6c4377f7f776e1291"},
+    {file = "librt-0.16.0-cp39-cp39-musllinux_1_2_i686.whl", hash = "sha256:aa9357a1b4d4fc787bb718a59cb1112c28c8a976d6bfa268b71cc0a4ab8f3a94"},
+    {file = "librt-0.16.0-cp39-cp39-musllinux_1_2_ppc64le.whl", hash = "sha256:e9ce0bc440e7fd09b5f51f372f0f6640658f854b8fb05260f819cc669c93c42d"},
+    {file = "librt-0.16.0-cp39-cp39-musllinux_1_2_riscv64.whl", hash = "sha256:6c8893eae2fd13c5488d94056f3e6e5cf3142bfb1c4acaf136cb33d760c5964b"},
+    {file = "librt-0.16.0-cp39-cp39-musllinux_1_2_x86_64.whl", hash = "sha256:7393c9a48dcce4817dbd4b0d8ff6237efe9b0a0609f5b0adaef315f8541726b5"},
+    {file = "librt-0.16.0-cp39-cp39-win32.whl", hash = "sha256:71b93b42784e25b975079573c642a8fedb049a7bb31d70a51721b1666b3b2ced"},
+    {file = "librt-0.16.0-cp39-cp39-win_amd64.whl", hash = "sha256:5750a105b42a416f930edc59054927a406effb2550cd5bab92ad7a5842ed5d05"},
+    {file = "librt-0.16.0.tar.gz", hash = "sha256:ac38d6d8d66bf3d744148dbbc0b8e193e195a51e364ed55e224631f5721891fc"},
+]
+
+[[package]]
+name = "mypy"
+version = "2.4.0"
+description = "Optional static typing for Python"
+optional = false
+python-versions = ">=3.10"
+groups = ["dev"]
+files = [
+    {file = "mypy-2.4.0-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:ffda5244fd1ad71a1e54405e35f50d09b80c3978efa120f58bd1252ae32c62d2"},
+    {file = "mypy-2.4.0-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:615b03922d40e186fd1df73156473db0ba525c639bb4fd1cf28e1694879c9b23"},
+    {file = "mypy-2.4.0-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:13fa24f439c0e48a290a3922fa14ccd22f2762bae99d2142931b3e40a9055080"},
+    {file = "mypy-2.4.0-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:3adef556a19eb3b630bf86a79c29d7da3d61e541472d0d897ca01b171c0abf8f"},
+    {file = "mypy-2.4.0-cp310-cp310-win_amd64.whl", hash = "sha256:1dc0f64b0a92ae27a49d2175f0bacfa56e15bdc5b420cf92e0c1f79292219cbb"},
+    {file = "mypy-2.4.0-cp310-cp310-win_arm64.whl", hash = "sha256:4a378fc15fb33e321f04652c166ce73eeb8833a97c3d218132844e938cd93220"},
+    {file = "mypy-2.4.0-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:5d20e6c7c35fcbf2a0ebdd0eaeacfbc243009dfd33ab7822d54e213912e6dbbd"},
+    {file = "mypy-2.4.0-cp311-cp311-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:3011537be6cf1de4511c0255a324362a812b58184bbe61e15f59c8b31033bd74"},
+    {file = "mypy-2.4.0-cp311-cp311-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:a96b07a49b7b1d025ce59c1b3acbcf24bead9a83da4523c4a6bde1bb94e7a0e1"},
+    {file = "mypy-2.4.0-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:7c4f8f8d1d1c0e2832d8ee7113dd08f6df6c7aad9e863fcbed9f25832be0b8c4"},
+    {file = "mypy-2.4.0-cp311-cp311-win_amd64.whl", hash = "sha256:ba05652540bf12828e52abae807b024b09ca144ff4f75e2450a81d69c376425b"},
+    {file = "mypy-2.4.0-cp311-cp311-win_arm64.whl", hash = "sha256:6306086b87cf7f8a29aa618d9fd9bffb56c59247166b9660fdb54d86d7714ecd"},
+    {file = "mypy-2.4.0-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:e05ff2925d8b37ad26c80c1b9dc43ae5d455da2df1e23c24c095a6425917c57e"},
+    {file = "mypy-2.4.0-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:29243242cf72582b65f9582ad9e56e8cb281566ed3519f4cd70bb8b9f2977e90"},
+    {file = "mypy-2.4.0-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:29eb0b9427a6b11b992e452f6cceb8af724f4dceb47e779d0b35e405e996ea5e"},
+    {file = "mypy-2.4.0-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:e3ebe2f72a2a1156065a9851570ffbf50c0a93cdccadef9c6e05c508a4fd10b1"},
+    {file = "mypy-2.4.0-cp312-cp312-win_amd64.whl", hash = "sha256:236e0d68f6941992b0811128e652590f590db444ab29ad8f1324765b9298b946"},
+    {file = "mypy-2.4.0-cp312-cp312-win_arm64.whl", hash = "sha256:82d0f94c8587ccb472622ee7795280aaa38a06640d5f45b3f16909d6dd86a989"},
+    {file = "mypy-2.4.0-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:7da85fbcff6dac1abcc636707bed38b45598131fb7a605d9719c70b5cc733af8"},
+    {file = "mypy-2.4.0-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:4209da39d85cf240f762af622d8180fcdfcb4727d021f44ade62d613a1a43324"},
+    {file = "mypy-2.4.0-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:6be721bd4bd57576193653b75b4af3461c9d0bf7dd8b528f782e9be210dc75bb"},
+    {file = "mypy-2.4.0-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:e1fde197ae65be856a034a91b70ed747a16562ca69577785f06c661548424bf1"},
+    {file = "mypy-2.4.0-cp313-cp313-win_amd64.whl", hash = "sha256:295ecf2e57542cd836ca537486951289678c8c7d1ee6ad74ebe29b2168a003cf"},
+    {file = "mypy-2.4.0-cp313-cp313-win_arm64.whl", hash = "sha256:bc378bdad4e9f12b5bd96466083d1e71acf00594ec9c7b2bdb5e02816f77f303"},
+    {file = "mypy-2.4.0-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:058165f564ccf559c68c70fec2091fca5891110480210c22594635e3f6683437"},
+    {file = "mypy-2.4.0-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:9fa247e02b505a45a2775f69df38d360d197e3790bc60f717595db9eda358b6e"},
+    {file = "mypy-2.4.0-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:20e9a5cd875837520c43db98dea0b6d0c2197833d95c30127d8f570fb9b1f00b"},
+    {file = "mypy-2.4.0-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:9f03a7828cca2b0adcd6662aee8f2711ff8830e1027641fdea3ab0b787483566"},
+    {file = "mypy-2.4.0-cp314-cp314-pyemscripten_2026_0_wasm32.whl", hash = "sha256:9279488933040b638c0ab739084c0ca100efeea6db581bf5d7628d8e89de53fe"},
+    {file = "mypy-2.4.0-cp314-cp314-win_amd64.whl", hash = "sha256:2106b55105ba5ea9be4f53a24517fc5fa927ff1585edc9bc1a975abb72caef89"},
+    {file = "mypy-2.4.0-cp314-cp314-win_arm64.whl", hash = "sha256:528c8744b8b5e3ecb8774f86af38d2376216816e9908317ad055f3c9c2d74799"},
+    {file = "mypy-2.4.0-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:0bb95cf34899e4619c61ab0a8667804e139e580b30d5df12af2102dfe44d0c97"},
+    {file = "mypy-2.4.0-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:86d616fe84c6eab8026f8c50ab5bcb90db780d2ccd233d971e34e92bede9b359"},
+    {file = "mypy-2.4.0-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:a3f86fd1313dd69d013e265f1fdcd12ea7a9d606f9875b2a3db946cd334555f3"},
+    {file = "mypy-2.4.0-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:720434d48542ecfe84d32d287b727569d3fc8f5769acd39051130e490a5c295c"},
+    {file = "mypy-2.4.0-cp314-cp314t-win_amd64.whl", hash = "sha256:a6e851b82c0661f69f1630fc16172c68787a6a9cf0991e7c6437d60976cdcd76"},
+    {file = "mypy-2.4.0-cp314-cp314t-win_arm64.whl", hash = "sha256:3bd0e340f0ebe65c548210f53be3fd8192e83964760caf0c28bef368e68b0d37"},
+    {file = "mypy-2.4.0-cp315-cp315-macosx_11_0_arm64.whl", hash = "sha256:afa89837d9be67e0cadfa33bca3bb7efdda98c3b07e74dc3b635ebfb1c8a926a"},
+    {file = "mypy-2.4.0-cp315-cp315-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:fb443e81057896132d3642d6be219e6efd158691ac7883e3ba8fcb469865f05d"},
+    {file = "mypy-2.4.0-cp315-cp315-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:7f38f57d344f8b6accb40e01c3d83cfc590498231724d16c07ffb7940f157818"},
+    {file = "mypy-2.4.0-cp315-cp315-musllinux_1_2_x86_64.whl", hash = "sha256:e76172710bd4e5eeae061abfd68347e5264632e02778be61784671ae3a2132f5"},
+    {file = "mypy-2.4.0-cp315-cp315-pyemscripten_2026_5_wasm32.whl", hash = "sha256:f83353e47ab520bf6fd4df8f5897d9fe081211f2fbc4b7d37736a3e3c166cbcf"},
+    {file = "mypy-2.4.0-cp315-cp315-win_amd64.whl", hash = "sha256:970b221ed5842213d98e3c480c08f795ace4b1f81fb21e1b126bd0476bce1c34"},
+    {file = "mypy-2.4.0-cp315-cp315-win_arm64.whl", hash = "sha256:502b94b0b331f7dafe32fd6b151797ddbb4f32385b362e722c783a025e5954a3"},
+    {file = "mypy-2.4.0-cp315-cp315t-macosx_11_0_arm64.whl", hash = "sha256:c9de622fd397495695d0598ddc789222bfcfec9d7c9ec3a1e385c855e3bc5e01"},
+    {file = "mypy-2.4.0-cp315-cp315t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:9f459f0b4f0596d9d51fe7716b404b35287b99e77da98a7af90a65dd5fd61141"},
+    {file = "mypy-2.4.0-cp315-cp315t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:f9b028548b3af480e2b1ed8df14ccaac86f99c9f600d1580770ab7ba3dcd40f0"},
+    {file = "mypy-2.4.0-cp315-cp315t-musllinux_1_2_x86_64.whl", hash = "sha256:cb734b2668c1f40d07ce093bbeb4407e9527c67901627b0e1679825be3f09975"},
+    {file = "mypy-2.4.0-cp315-cp315t-win_amd64.whl", hash = "sha256:172e30b8fea631fe310f0c665477f52d9ea40bb4e99e0c81dc30118563b13710"},
+    {file = "mypy-2.4.0-cp315-cp315t-win_arm64.whl", hash = "sha256:5786ef987b3767e51aaa53f20aec104c0252b42ecda7aef8e8b4cbae279b05c5"},
+    {file = "mypy-2.4.0-py3-none-any.whl", hash = "sha256:d01c5d26a352acc6d5cf3128225477e1e8465e8d3029d4c345807fbf7f3cf093"},
+    {file = "mypy-2.4.0.tar.gz", hash = "sha256:77bdaebd452f43fcfc4cc3ba94352a3ea537cd01e3f2d0879f48673d2ec00d6e"},
+]
+
+[package.dependencies]
+ast-serialize = ">=0.11.2,<1.0.0"
+librt = {version = ">=0.16.0", markers = "platform_python_implementation != \"PyPy\""}
+mypy_extensions = ">=1.0.0"
+pathspec = ">=1.0.0"
+typing_extensions = [
+    {version = ">=4.6.0", markers = "python_version < \"3.15\""},
+    {version = ">=4.14.0", markers = "python_version >= \"3.15\""},
+]
+
+[package.extras]
+dmypy = ["psutil (>=4.0)"]
+faster-cache = ["orjson"]
+install-types = ["pip"]
+mypyc = ["setuptools (>=50)"]
+reports = ["lxml"]
+
+[[package]]
+name = "mypy-extensions"
+version = "1.1.0"
+description = "Type system extensions for programs checked with the mypy type checker."
+optional = false
+python-versions = ">=3.8"
+groups = ["dev"]
+files = [
+    {file = "mypy_extensions-1.1.0-py3-none-any.whl", hash = "sha256:1be4cccdb0f2482337c4743e60421de3a356cd97508abadd57d47403e94f5505"},
+    {file = "mypy_extensions-1.1.0.tar.gz", hash = "sha256:52e68efc3284861e772bbcd66823fde5ae21fd2fdb51c62a211403730b916558"},
+]
+
+[[package]]
+name = "nodeenv"
+version = "1.11.0"
+description = "Node.js virtual environment builder"
+optional = false
+python-versions = "!=3.0.*,!=3.1.*,!=3.2.*,!=3.3.*,!=3.4.*,!=3.5.*,!=3.6.*,>=2.7"
+groups = ["dev"]
+files = [
+    {file = "nodeenv-1.11.0-py2.py3-none-any.whl", hash = "sha256:edaa16e6c14d7cf395d75d4bbd5a26390f4dc06501a33b4e76282b02cc688a25"},
+    {file = "nodeenv-1.11.0.tar.gz", hash = "sha256:3ce8fe5b71d16e8af7039ca65257354100bc772965d6bc549070649e53b1b146"},
+]
+
+[package.extras]
+certifi = ["certifi"]
+
+[[package]]
+name = "packaging"
+version = "26.3"
+description = "Core utilities for Python packages"
+optional = false
+python-versions = ">=3.9"
+groups = ["main", "dev"]
+files = [
+    {file = "packaging-26.3-py3-none-any.whl", hash = "sha256:d7193f7c8e4e93f444fde0262bf90af30e16fa0ad0ad44cb553c87339b23cd1c"},
+    {file = "packaging-26.3.tar.gz", hash = "sha256:94edc256424af38762eb31306eed28beb9f0efc50a8837492c9d6fd6004aed79"},
+]
+
+[[package]]
+name = "pathspec"
+version = "1.1.1"
+description = "Utility library for gitignore style pattern matching of file paths."
+optional = false
+python-versions = ">=3.9"
+groups = ["dev"]
+files = [
+    {file = "pathspec-1.1.1-py3-none-any.whl", hash = "sha256:a00ce642f577bf7f473932318056212bc4f8bfdf53128c78bbd5af0b9b20b189"},
+    {file = "pathspec-1.1.1.tar.gz", hash = "sha256:17db5ecd524104a120e173814c90367a96a98d07c45b2e10c2f3919fff91bf5a"},
+]
+
+[package.extras]
+hyperscan = ["hyperscan (>=0.7)"]
+optional = ["typing-extensions (>=4)"]
+re2 = ["google-re2 (>=1.1)"]
+
+[[package]]
+name = "pluggy"
+version = "1.6.0"
+description = "plugin and hook calling mechanisms for python"
+optional = false
+python-versions = ">=3.9"
+groups = ["dev"]
+files = [
+    {file = "pluggy-1.6.0-py3-none-any.whl", hash = "sha256:e920276dd6813095e9377c0bc5566d94c932c33b27a3e3945d8389c374dd4746"},
+    {file = "pluggy-1.6.0.tar.gz", hash = "sha256:7dcc130b76258d33b90f61b658791dede3486c3e6bfb003ee5c9bfb396dd22f3"},
+]
+
+[package.extras]
+dev = ["pre-commit", "tox"]
+testing = ["coverage", "pytest", "pytest-benchmark"]
+
+[[package]]
+name = "pydantic"
+version = "2.12.5"
+description = "Data validation using Python type hints"
+optional = false
+python-versions = ">=3.9"
+groups = ["main"]
+files = [
+    {file = "pydantic-2.12.5-py3-none-any.whl", hash = "sha256:e561593fccf61e8a20fc46dfc2dfe075b8be7d0188df33f221ad1f0139180f9d"},
+    {file = "pydantic-2.12.5.tar.gz", hash = "sha256:4d351024c75c0f085a9febbb665ce8c0c6ec5d30e903bdb6394b7ede26aebb49"},
+]
+
+[package.dependencies]
+annotated-types = ">=0.6.0"
+pydantic-core = "2.41.5"
+typing-extensions = ">=4.14.1"
+typing-inspection = ">=0.4.2"
+
+[package.extras]
+email = ["email-validator (>=2.0.0)"]
+timezone = ["tzdata ; python_version >= \"3.9\" and platform_system == \"Windows\""]
+
+[[package]]
+name = "pydantic-core"
+version = "2.41.5"
+description = "Core functionality for Pydantic validation and serialization"
+optional = false
+python-versions = ">=3.9"
+groups = ["main"]
+files = [
+    {file = "pydantic_core-2.41.5-cp310-cp310-macosx_10_12_x86_64.whl", hash = "sha256:77b63866ca88d804225eaa4af3e664c5faf3568cea95360d21f4725ab6e07146"},
+    {file = "pydantic_core-2.41.5-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:dfa8a0c812ac681395907e71e1274819dec685fec28273a28905df579ef137e2"},
+    {file = "pydantic_core-2.41.5-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:5921a4d3ca3aee735d9fd163808f5e8dd6c6972101e4adbda9a4667908849b97"},
+    {file = "pydantic_core-2.41.5-cp310-cp310-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:e25c479382d26a2a41b7ebea1043564a937db462816ea07afa8a44c0866d52f9"},
+    {file = "pydantic_core-2.41.5-cp310-cp310-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:f547144f2966e1e16ae626d8ce72b4cfa0caedc7fa28052001c94fb2fcaa1c52"},
+    {file = "pydantic_core-2.41.5-cp310-cp310-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:6f52298fbd394f9ed112d56f3d11aabd0d5bd27beb3084cc3d8ad069483b8941"},
+    {file = "pydantic_core-2.41.5-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:100baa204bb412b74fe285fb0f3a385256dad1d1879f0a5cb1499ed2e83d132a"},
+    {file = "pydantic_core-2.41.5-cp310-cp310-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:05a2c8852530ad2812cb7914dc61a1125dc4e06252ee98e5638a12da6cc6fb6c"},
+    {file = "pydantic_core-2.41.5-cp310-cp310-musllinux_1_1_aarch64.whl", hash = "sha256:29452c56df2ed968d18d7e21f4ab0ac55e71dc59524872f6fc57dcf4a3249ed2"},
+    {file = "pydantic_core-2.41.5-cp310-cp310-musllinux_1_1_armv7l.whl", hash = "sha256:d5160812ea7a8a2ffbe233d8da666880cad0cbaf5d4de74ae15c313213d62556"},
+    {file = "pydantic_core-2.41.5-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:df3959765b553b9440adfd3c795617c352154e497a4eaf3752555cfb5da8fc49"},
+    {file = "pydantic_core-2.41.5-cp310-cp310-win32.whl", hash = "sha256:1f8d33a7f4d5a7889e60dc39856d76d09333d8a6ed0f5f1190635cbec70ec4ba"},
+    {file = "pydantic_core-2.41.5-cp310-cp310-win_amd64.whl", hash = "sha256:62de39db01b8d593e45871af2af9e497295db8d73b085f6bfd0b18c83c70a8f9"},
+    {file = "pydantic_core-2.41.5-cp311-cp311-macosx_10_12_x86_64.whl", hash = "sha256:a3a52f6156e73e7ccb0f8cced536adccb7042be67cb45f9562e12b319c119da6"},
+    {file = "pydantic_core-2.41.5-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:7f3bf998340c6d4b0c9a2f02d6a400e51f123b59565d74dc60d252ce888c260b"},
+    {file = "pydantic_core-2.41.5-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:378bec5c66998815d224c9ca994f1e14c0c21cb95d2f52b6021cc0b2a58f2a5a"},
+    {file = "pydantic_core-2.41.5-cp311-cp311-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:e7b576130c69225432866fe2f4a469a85a54ade141d96fd396dffcf607b558f8"},
+    {file = "pydantic_core-2.41.5-cp311-cp311-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:6cb58b9c66f7e4179a2d5e0f849c48eff5c1fca560994d6eb6543abf955a149e"},
+    {file = "pydantic_core-2.41.5-cp311-cp311-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:88942d3a3dff3afc8288c21e565e476fc278902ae4d6d134f1eeda118cc830b1"},
+    {file = "pydantic_core-2.41.5-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:f31d95a179f8d64d90f6831d71fa93290893a33148d890ba15de25642c5d075b"},
+    {file = "pydantic_core-2.41.5-cp311-cp311-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:c1df3d34aced70add6f867a8cf413e299177e0c22660cc767218373d0779487b"},
+    {file = "pydantic_core-2.41.5-cp311-cp311-musllinux_1_1_aarch64.whl", hash = "sha256:4009935984bd36bd2c774e13f9a09563ce8de4abaa7226f5108262fa3e637284"},
+    {file = "pydantic_core-2.41.5-cp311-cp311-musllinux_1_1_armv7l.whl", hash = "sha256:34a64bc3441dc1213096a20fe27e8e128bd3ff89921706e83c0b1ac971276594"},
+    {file = "pydantic_core-2.41.5-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:c9e19dd6e28fdcaa5a1de679aec4141f691023916427ef9bae8584f9c2fb3b0e"},
+    {file = "pydantic_core-2.41.5-cp311-cp311-win32.whl", hash = "sha256:2c010c6ded393148374c0f6f0bf89d206bf3217f201faa0635dcd56bd1520f6b"},
+    {file = "pydantic_core-2.41.5-cp311-cp311-win_amd64.whl", hash = "sha256:76ee27c6e9c7f16f47db7a94157112a2f3a00e958bc626e2f4ee8bec5c328fbe"},
+    {file = "pydantic_core-2.41.5-cp311-cp311-win_arm64.whl", hash = "sha256:4bc36bbc0b7584de96561184ad7f012478987882ebf9f9c389b23f432ea3d90f"},
+    {file = "pydantic_core-2.41.5-cp312-cp312-macosx_10_12_x86_64.whl", hash = "sha256:f41a7489d32336dbf2199c8c0a215390a751c5b014c2c1c5366e817202e9cdf7"},
+    {file = "pydantic_core-2.41.5-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:070259a8818988b9a84a449a2a7337c7f430a22acc0859c6b110aa7212a6d9c0"},
+    {file = "pydantic_core-2.41.5-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:e96cea19e34778f8d59fe40775a7a574d95816eb150850a85a7a4c8f4b94ac69"},
+    {file = "pydantic_core-2.41.5-cp312-cp312-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:ed2e99c456e3fadd05c991f8f437ef902e00eedf34320ba2b0842bd1c3ca3a75"},
+    {file = "pydantic_core-2.41.5-cp312-cp312-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:65840751b72fbfd82c3c640cff9284545342a4f1eb1586ad0636955b261b0b05"},
+    {file = "pydantic_core-2.41.5-cp312-cp312-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:e536c98a7626a98feb2d3eaf75944ef6f3dbee447e1f841eae16f2f0a72d8ddc"},
+    {file = "pydantic_core-2.41.5-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:eceb81a8d74f9267ef4081e246ffd6d129da5d87e37a77c9bde550cb04870c1c"},
+    {file = "pydantic_core-2.41.5-cp312-cp312-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:d38548150c39b74aeeb0ce8ee1d8e82696f4a4e16ddc6de7b1d8823f7de4b9b5"},
+    {file = "pydantic_core-2.41.5-cp312-cp312-musllinux_1_1_aarch64.whl", hash = "sha256:c23e27686783f60290e36827f9c626e63154b82b116d7fe9adba1fda36da706c"},
+    {file = "pydantic_core-2.41.5-cp312-cp312-musllinux_1_1_armv7l.whl", hash = "sha256:482c982f814460eabe1d3bb0adfdc583387bd4691ef00b90575ca0d2b6fe2294"},
+    {file = "pydantic_core-2.41.5-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:bfea2a5f0b4d8d43adf9d7b8bf019fb46fdd10a2e5cde477fbcb9d1fa08c68e1"},
+    {file = "pydantic_core-2.41.5-cp312-cp312-win32.whl", hash = "sha256:b74557b16e390ec12dca509bce9264c3bbd128f8a2c376eaa68003d7f327276d"},
+    {file = "pydantic_core-2.41.5-cp312-cp312-win_amd64.whl", hash = "sha256:1962293292865bca8e54702b08a4f26da73adc83dd1fcf26fbc875b35d81c815"},
+    {file = "pydantic_core-2.41.5-cp312-cp312-win_arm64.whl", hash = "sha256:1746d4a3d9a794cacae06a5eaaccb4b8643a131d45fbc9af23e353dc0a5ba5c3"},
+    {file = "pydantic_core-2.41.5-cp313-cp313-macosx_10_12_x86_64.whl", hash = "sha256:941103c9be18ac8daf7b7adca8228f8ed6bb7a1849020f643b3a14d15b1924d9"},
+    {file = "pydantic_core-2.41.5-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:112e305c3314f40c93998e567879e887a3160bb8689ef3d2c04b6cc62c33ac34"},
+    {file = "pydantic_core-2.41.5-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:0cbaad15cb0c90aa221d43c00e77bb33c93e8d36e0bf74760cd00e732d10a6a0"},
+    {file = "pydantic_core-2.41.5-cp313-cp313-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:03ca43e12fab6023fc79d28ca6b39b05f794ad08ec2feccc59a339b02f2b3d33"},
+    {file = "pydantic_core-2.41.5-cp313-cp313-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:dc799088c08fa04e43144b164feb0c13f9a0bc40503f8df3e9fde58a3c0c101e"},
+    {file = "pydantic_core-2.41.5-cp313-cp313-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:97aeba56665b4c3235a0e52b2c2f5ae9cd071b8a8310ad27bddb3f7fb30e9aa2"},
+    {file = "pydantic_core-2.41.5-cp313-cp313-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:406bf18d345822d6c21366031003612b9c77b3e29ffdb0f612367352aab7d586"},
+    {file = "pydantic_core-2.41.5-cp313-cp313-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:b93590ae81f7010dbe380cdeab6f515902ebcbefe0b9327cc4804d74e93ae69d"},
+    {file = "pydantic_core-2.41.5-cp313-cp313-musllinux_1_1_aarch64.whl", hash = "sha256:01a3d0ab748ee531f4ea6c3e48ad9dac84ddba4b0d82291f87248f2f9de8d740"},
+    {file = "pydantic_core-2.41.5-cp313-cp313-musllinux_1_1_armv7l.whl", hash = "sha256:6561e94ba9dacc9c61bce40e2d6bdc3bfaa0259d3ff36ace3b1e6901936d2e3e"},
+    {file = "pydantic_core-2.41.5-cp313-cp313-musllinux_1_1_x86_64.whl", hash = "sha256:915c3d10f81bec3a74fbd4faebe8391013ba61e5a1a8d48c4455b923bdda7858"},
+    {file = "pydantic_core-2.41.5-cp313-cp313-win32.whl", hash = "sha256:650ae77860b45cfa6e2cdafc42618ceafab3a2d9a3811fcfbd3bbf8ac3c40d36"},
+    {file = "pydantic_core-2.41.5-cp313-cp313-win_amd64.whl", hash = "sha256:79ec52ec461e99e13791ec6508c722742ad745571f234ea6255bed38c6480f11"},
+    {file = "pydantic_core-2.41.5-cp313-cp313-win_arm64.whl", hash = "sha256:3f84d5c1b4ab906093bdc1ff10484838aca54ef08de4afa9de0f5f14d69639cd"},
+    {file = "pydantic_core-2.41.5-cp314-cp314-macosx_10_12_x86_64.whl", hash = "sha256:3f37a19d7ebcdd20b96485056ba9e8b304e27d9904d233d7b1015db320e51f0a"},
+    {file = "pydantic_core-2.41.5-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:1d1d9764366c73f996edd17abb6d9d7649a7eb690006ab6adbda117717099b14"},
+    {file = "pydantic_core-2.41.5-cp314-cp314-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:25e1c2af0fce638d5f1988b686f3b3ea8cd7de5f244ca147c777769e798a9cd1"},
+    {file = "pydantic_core-2.41.5-cp314-cp314-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:506d766a8727beef16b7adaeb8ee6217c64fc813646b424d0804d67c16eddb66"},
+    {file = "pydantic_core-2.41.5-cp314-cp314-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:4819fa52133c9aa3c387b3328f25c1facc356491e6135b459f1de698ff64d869"},
+    {file = "pydantic_core-2.41.5-cp314-cp314-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:2b761d210c9ea91feda40d25b4efe82a1707da2ef62901466a42492c028553a2"},
+    {file = "pydantic_core-2.41.5-cp314-cp314-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:22f0fb8c1c583a3b6f24df2470833b40207e907b90c928cc8d3594b76f874375"},
+    {file = "pydantic_core-2.41.5-cp314-cp314-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:2782c870e99878c634505236d81e5443092fba820f0373997ff75f90f68cd553"},
+    {file = "pydantic_core-2.41.5-cp314-cp314-musllinux_1_1_aarch64.whl", hash = "sha256:0177272f88ab8312479336e1d777f6b124537d47f2123f89cb37e0accea97f90"},
+    {file = "pydantic_core-2.41.5-cp314-cp314-musllinux_1_1_armv7l.whl", hash = "sha256:63510af5e38f8955b8ee5687740d6ebf7c2a0886d15a6d65c32814613681bc07"},
+    {file = "pydantic_core-2.41.5-cp314-cp314-musllinux_1_1_x86_64.whl", hash = "sha256:e56ba91f47764cc14f1daacd723e3e82d1a89d783f0f5afe9c364b8bb491ccdb"},
+    {file = "pydantic_core-2.41.5-cp314-cp314-win32.whl", hash = "sha256:aec5cf2fd867b4ff45b9959f8b20ea3993fc93e63c7363fe6851424c8a7e7c23"},
+    {file = "pydantic_core-2.41.5-cp314-cp314-win_amd64.whl", hash = "sha256:8e7c86f27c585ef37c35e56a96363ab8de4e549a95512445b85c96d3e2f7c1bf"},
+    {file = "pydantic_core-2.41.5-cp314-cp314-win_arm64.whl", hash = "sha256:e672ba74fbc2dc8eea59fb6d4aed6845e6905fc2a8afe93175d94a83ba2a01a0"},
+    {file = "pydantic_core-2.41.5-cp314-cp314t-macosx_10_12_x86_64.whl", hash = "sha256:8566def80554c3faa0e65ac30ab0932b9e3a5cd7f8323764303d468e5c37595a"},
+    {file = "pydantic_core-2.41.5-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:b80aa5095cd3109962a298ce14110ae16b8c1aece8b72f9dafe81cf597ad80b3"},
+    {file = "pydantic_core-2.41.5-cp314-cp314t-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:3006c3dd9ba34b0c094c544c6006cc79e87d8612999f1a5d43b769b89181f23c"},
+    {file = "pydantic_core-2.41.5-cp314-cp314t-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:72f6c8b11857a856bcfa48c86f5368439f74453563f951e473514579d44aa612"},
+    {file = "pydantic_core-2.41.5-cp314-cp314t-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:5cb1b2f9742240e4bb26b652a5aeb840aa4b417c7748b6f8387927bc6e45e40d"},
+    {file = "pydantic_core-2.41.5-cp314-cp314t-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:bd3d54f38609ff308209bd43acea66061494157703364ae40c951f83ba99a1a9"},
+    {file = "pydantic_core-2.41.5-cp314-cp314t-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:2ff4321e56e879ee8d2a879501c8e469414d948f4aba74a2d4593184eb326660"},
+    {file = "pydantic_core-2.41.5-cp314-cp314t-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:d0d2568a8c11bf8225044aa94409e21da0cb09dcdafe9ecd10250b2baad531a9"},
+    {file = "pydantic_core-2.41.5-cp314-cp314t-musllinux_1_1_aarch64.whl", hash = "sha256:a39455728aabd58ceabb03c90e12f71fd30fa69615760a075b9fec596456ccc3"},
+    {file = "pydantic_core-2.41.5-cp314-cp314t-musllinux_1_1_armv7l.whl", hash = "sha256:239edca560d05757817c13dc17c50766136d21f7cd0fac50295499ae24f90fdf"},
+    {file = "pydantic_core-2.41.5-cp314-cp314t-musllinux_1_1_x86_64.whl", hash = "sha256:2a5e06546e19f24c6a96a129142a75cee553cc018ffee48a460059b1185f4470"},
+    {file = "pydantic_core-2.41.5-cp314-cp314t-win32.whl", hash = "sha256:b4ececa40ac28afa90871c2cc2b9ffd2ff0bf749380fbdf57d165fd23da353aa"},
+    {file = "pydantic_core-2.41.5-cp314-cp314t-win_amd64.whl", hash = "sha256:80aa89cad80b32a912a65332f64a4450ed00966111b6615ca6816153d3585a8c"},
+    {file = "pydantic_core-2.41.5-cp314-cp314t-win_arm64.whl", hash = "sha256:35b44f37a3199f771c3eaa53051bc8a70cd7b54f333531c59e29fd4db5d15008"},
+    {file = "pydantic_core-2.41.5-cp39-cp39-macosx_10_12_x86_64.whl", hash = "sha256:8bfeaf8735be79f225f3fefab7f941c712aaca36f1128c9d7e2352ee1aa87bdf"},
+    {file = "pydantic_core-2.41.5-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:346285d28e4c8017da95144c7f3acd42740d637ff41946af5ce6e5e420502dd5"},
+    {file = "pydantic_core-2.41.5-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:a75dafbf87d6276ddc5b2bf6fae5254e3d0876b626eb24969a574fff9149ee5d"},
+    {file = "pydantic_core-2.41.5-cp39-cp39-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:7b93a4d08587e2b7e7882de461e82b6ed76d9026ce91ca7915e740ecc7855f60"},
+    {file = "pydantic_core-2.41.5-cp39-cp39-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:e8465ab91a4bd96d36dde3263f06caa6a8a6019e4113f24dc753d79a8b3a3f82"},
+    {file = "pydantic_core-2.41.5-cp39-cp39-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:299e0a22e7ae2b85c1a57f104538b2656e8ab1873511fd718a1c1c6f149b77b5"},
+    {file = "pydantic_core-2.41.5-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:707625ef0983fcfb461acfaf14de2067c5942c6bb0f3b4c99158bed6fedd3cf3"},
+    {file = "pydantic_core-2.41.5-cp39-cp39-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:f41eb9797986d6ebac5e8edff36d5cef9de40def462311b3eb3eeded1431e425"},
+    {file = "pydantic_core-2.41.5-cp39-cp39-musllinux_1_1_aarch64.whl", hash = "sha256:0384e2e1021894b1ff5a786dbf94771e2986ebe2869533874d7e43bc79c6f504"},
+    {file = "pydantic_core-2.41.5-cp39-cp39-musllinux_1_1_armv7l.whl", hash = "sha256:f0cd744688278965817fd0839c4a4116add48d23890d468bc436f78beb28abf5"},
+    {file = "pydantic_core-2.41.5-cp39-cp39-musllinux_1_1_x86_64.whl", hash = "sha256:753e230374206729bf0a807954bcc6c150d3743928a73faffee51ac6557a03c3"},
+    {file = "pydantic_core-2.41.5-cp39-cp39-win32.whl", hash = "sha256:873e0d5b4fb9b89ef7c2d2a963ea7d02879d9da0da8d9d4933dee8ee86a8b460"},
+    {file = "pydantic_core-2.41.5-cp39-cp39-win_amd64.whl", hash = "sha256:e4f4a984405e91527a0d62649ee21138f8e3d0ef103be488c1dc11a80d7f184b"},
+    {file = "pydantic_core-2.41.5-graalpy311-graalpy242_311_native-macosx_10_12_x86_64.whl", hash = "sha256:b96d5f26b05d03cc60f11a7761a5ded1741da411e7fe0909e27a5e6a0cb7b034"},
+    {file = "pydantic_core-2.41.5-graalpy311-graalpy242_311_native-macosx_11_0_arm64.whl", hash = "sha256:634e8609e89ceecea15e2d61bc9ac3718caaaa71963717bf3c8f38bfde64242c"},
+    {file = "pydantic_core-2.41.5-graalpy311-graalpy242_311_native-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:93e8740d7503eb008aa2df04d3b9735f845d43ae845e6dcd2be0b55a2da43cd2"},
+    {file = "pydantic_core-2.41.5-graalpy311-graalpy242_311_native-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:f15489ba13d61f670dcc96772e733aad1a6f9c429cc27574c6cdaed82d0146ad"},
+    {file = "pydantic_core-2.41.5-graalpy312-graalpy250_312_native-macosx_10_12_x86_64.whl", hash = "sha256:7da7087d756b19037bc2c06edc6c170eeef3c3bafcb8f532ff17d64dc427adfd"},
+    {file = "pydantic_core-2.41.5-graalpy312-graalpy250_312_native-macosx_11_0_arm64.whl", hash = "sha256:aabf5777b5c8ca26f7824cb4a120a740c9588ed58df9b2d196ce92fba42ff8dc"},
+    {file = "pydantic_core-2.41.5-graalpy312-graalpy250_312_native-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:c007fe8a43d43b3969e8469004e9845944f1a80e6acd47c150856bb87f230c56"},
+    {file = "pydantic_core-2.41.5-graalpy312-graalpy250_312_native-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:76d0819de158cd855d1cbb8fcafdf6f5cf1eb8e470abe056d5d161106e38062b"},
+    {file = "pydantic_core-2.41.5-pp310-pypy310_pp73-macosx_10_12_x86_64.whl", hash = "sha256:b5819cd790dbf0c5eb9f82c73c16b39a65dd6dd4d1439dcdea7816ec9adddab8"},
+    {file = "pydantic_core-2.41.5-pp310-pypy310_pp73-macosx_11_0_arm64.whl", hash = "sha256:5a4e67afbc95fa5c34cf27d9089bca7fcab4e51e57278d710320a70b956d1b9a"},
+    {file = "pydantic_core-2.41.5-pp310-pypy310_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ece5c59f0ce7d001e017643d8d24da587ea1f74f6993467d85ae8a5ef9d4f42b"},
+    {file = "pydantic_core-2.41.5-pp310-pypy310_pp73-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:16f80f7abe3351f8ea6858914ddc8c77e02578544a0ebc15b4c2e1a0e813b0b2"},
+    {file = "pydantic_core-2.41.5-pp310-pypy310_pp73-musllinux_1_1_aarch64.whl", hash = "sha256:33cb885e759a705b426baada1fe68cbb0a2e68e34c5d0d0289a364cf01709093"},
+    {file = "pydantic_core-2.41.5-pp310-pypy310_pp73-musllinux_1_1_armv7l.whl", hash = "sha256:c8d8b4eb992936023be7dee581270af5c6e0697a8559895f527f5b7105ecd36a"},
+    {file = "pydantic_core-2.41.5-pp310-pypy310_pp73-musllinux_1_1_x86_64.whl", hash = "sha256:242a206cd0318f95cd21bdacff3fcc3aab23e79bba5cac3db5a841c9ef9c6963"},
+    {file = "pydantic_core-2.41.5-pp310-pypy310_pp73-win_amd64.whl", hash = "sha256:d3a978c4f57a597908b7e697229d996d77a6d3c94901e9edee593adada95ce1a"},
+    {file = "pydantic_core-2.41.5-pp311-pypy311_pp73-macosx_10_12_x86_64.whl", hash = "sha256:b2379fa7ed44ddecb5bfe4e48577d752db9fc10be00a6b7446e9663ba143de26"},
+    {file = "pydantic_core-2.41.5-pp311-pypy311_pp73-macosx_11_0_arm64.whl", hash = "sha256:266fb4cbf5e3cbd0b53669a6d1b039c45e3ce651fd5442eff4d07c2cc8d66808"},
+    {file = "pydantic_core-2.41.5-pp311-pypy311_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:58133647260ea01e4d0500089a8c4f07bd7aa6ce109682b1426394988d8aaacc"},
+    {file = "pydantic_core-2.41.5-pp311-pypy311_pp73-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:287dad91cfb551c363dc62899a80e9e14da1f0e2b6ebde82c806612ca2a13ef1"},
+    {file = "pydantic_core-2.41.5-pp311-pypy311_pp73-musllinux_1_1_aarch64.whl", hash = "sha256:03b77d184b9eb40240ae9fd676ca364ce1085f203e1b1256f8ab9984dca80a84"},
+    {file = "pydantic_core-2.41.5-pp311-pypy311_pp73-musllinux_1_1_armv7l.whl", hash = "sha256:a668ce24de96165bb239160b3d854943128f4334822900534f2fe947930e5770"},
+    {file = "pydantic_core-2.41.5-pp311-pypy311_pp73-musllinux_1_1_x86_64.whl", hash = "sha256:f14f8f046c14563f8eb3f45f499cc658ab8d10072961e07225e507adb700e93f"},
+    {file = "pydantic_core-2.41.5-pp311-pypy311_pp73-win_amd64.whl", hash = "sha256:56121965f7a4dc965bff783d70b907ddf3d57f6eba29b6d2e5dabfaf07799c51"},
+    {file = "pydantic_core-2.41.5.tar.gz", hash = "sha256:08daa51ea16ad373ffd5e7606252cc32f07bc72b28284b6bc9c6df804816476e"},
+]
+
+[package.dependencies]
+typing-extensions = ">=4.14.1"
+
+[[package]]
+name = "pygments"
+version = "2.21.0"
+description = "Pygments is a syntax highlighting package written in Python."
+optional = false
+python-versions = ">=3.9"
+groups = ["dev"]
+files = [
+    {file = "pygments-2.21.0-py3-none-any.whl", hash = "sha256:2363c69b61c4a97c838da3b130dcd6468f4848992b21a82f2a63ec34377137d9"},
+    {file = "pygments-2.21.0.tar.gz", hash = "sha256:610ca751c9bc2492b38eb9a38a7fbc93edbbb2d7182edaf34e66ae493dee5c8c"},
+]
+
+[package.extras]
+windows-terminal = ["colorama (>=0.4.6)"]
+
+[[package]]
+name = "pyright"
+version = "1.1.414"
+description = "Command line wrapper for pyright"
+optional = false
+python-versions = ">=3.7"
+groups = ["dev"]
+files = [
+    {file = "pyright-1.1.414-py3-none-any.whl", hash = "sha256:2a6b4b3298c9eec174c5ed83bd338de6eee82df2992f3e1930e6199d381be36f"},
+    {file = "pyright-1.1.414.tar.gz", hash = "sha256:523c0a97c60da6333234955c277730c9cf4f5bd6d5399e7b7d2b0fc5d3599524"},
+]
+
+[package.dependencies]
+nodeenv = ">=1.6.0"
+typing-extensions = ">=4.1"
+
+[package.extras]
+all = ["nodejs-wheel-binaries", "twine (>=3.4.1)"]
+dev = ["twine (>=3.4.1)"]
+nodejs = ["nodejs-wheel-binaries"]
+
+[[package]]
+name = "pytest"
+version = "9.1.1"
+description = "pytest: simple powerful testing with Python"
+optional = false
+python-versions = ">=3.10"
+groups = ["dev"]
+files = [
+    {file = "pytest-9.1.1-py3-none-any.whl", hash = "sha256:37a86b45efb9a47a61a36449063e8e18d0cab3161329fc099eb21783169c4f0c"},
+    {file = "pytest-9.1.1.tar.gz", hash = "sha256:1088fbde8f2b49d95a549a195707afa7a76a3ce9bcadc26b6d71f0ffda5fe313"},
+]
+
+[package.dependencies]
+colorama = {version = ">=0.4", markers = "sys_platform == \"win32\""}
+iniconfig = ">=1.0.1"
+packaging = ">=22"
+pluggy = ">=1.5,<2"
+pygments = ">=2.7.2"
+
+[package.extras]
+dev = ["argcomplete", "attrs (>=19.2)", "hypothesis (>=3.56)", "mock", "requests", "setuptools", "xmlschema"]
+
+[[package]]
+name = "tenacity"
+version = "9.1.4"
+description = "Retry code until it succeeds"
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "tenacity-9.1.4-py3-none-any.whl", hash = "sha256:6095a360c919085f28c6527de529e76a06ad89b23659fa881ae0649b867a9d55"},
+    {file = "tenacity-9.1.4.tar.gz", hash = "sha256:adb31d4c263f2bd041081ab33b498309a57c77f9acf2db65aadf0898179cf93a"},
+]
+
+[package.extras]
+doc = ["reno", "sphinx"]
+test = ["pytest", "tornado (>=4.5)", "typeguard"]
+
+[[package]]
+name = "truststore"
+version = "0.10.4"
+description = "Verify certificates using native system trust stores"
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+markers = "sys_platform != \"emscripten\""
+files = [
+    {file = "truststore-0.10.4-py3-none-any.whl", hash = "sha256:adaeaecf1cbb5f4de3b1959b42d41f6fab57b2b1666adb59e89cb0b53361d981"},
+    {file = "truststore-0.10.4.tar.gz", hash = "sha256:9d91bd436463ad5e4ee4aba766628dd6cd7010cf3e2461756b3303710eebc301"},
+]
+
+[[package]]
+name = "typesafe-sdk"
+version = "0.7.5"
+description = "Python SDK for TypeSafe AI API."
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "typesafe_sdk-0.7.5-py3-none-any.whl", hash = "sha256:967a8b61494d536abb55ee3f3a2b10e81685cb6baad6ccdcab588bdc6c9b86d2"},
+    {file = "typesafe_sdk-0.7.5.tar.gz", hash = "sha256:1eeeb0dc80a7161d84896db42a5016dc0a7c8872fa16540b70163e5357b6d398"},
+]
+
+[package.dependencies]
+anyio = ">=4.0.0"
+httpx2 = ">=2.0.0"
+packaging = ">=24.0"
+pydantic = ">=2.12.0"
+pydantic-core = ">=2.41.1"
+tenacity = ">=9.0.0"
+truststore = {version = ">=0.10.0", markers = "sys_platform != \"emscripten\""}
+typing-extensions = ">=4.13.0"
+
+[package.extras]
+http2 = ["httpx2[http2] (>=2.0.0)"]
+pil = ["pillow (>=11.0.0)"]
+zstd = ["backports-zstd (>=1.0.0) ; python_full_version < \"3.14.0\""]
+
+[[package]]
+name = "typing-extensions"
+version = "4.16.0"
+description = "Backported and Experimental Type Hints for Python 3.9+"
+optional = false
+python-versions = ">=3.9"
+groups = ["main", "dev"]
+files = [
+    {file = "typing_extensions-4.16.0-py3-none-any.whl", hash = "sha256:481caa481374e813c1b176ada14e97f1f67a4539ce9cfeb3f350d78d6370c2e8"},
+    {file = "typing_extensions-4.16.0.tar.gz", hash = "sha256:dc983d19a509c94dba722ee6abd33940f7c05a89e243c47e907eb4db6f1a43e5"},
+]
+
+[[package]]
+name = "typing-inspection"
+version = "0.4.4"
+description = "Runtime typing introspection tools"
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "typing_inspection-0.4.4-py3-none-any.whl", hash = "sha256:65b8397ba37ccbce054456aaccddfc91e6e3083c92824df348d96ca832f3f147"},
+    {file = "typing_inspection-0.4.4.tar.gz", hash = "sha256:547274fa6b0a561ccf549cc9524b999a578e737d015d8709d021f9d0d13bea47"},
+]
+
+[package.dependencies]
+typing-extensions = ">=4.15.0"
+
+[metadata]
+lock-version = "2.1"
+python-versions = "^3.11"
+content-hash = "f3cfbab9af11a94325af11c9940e6f67314c0e01f93a3b70eadb5de39119aea5"
diff --git a/paddler_typesafe_client_python_test/pyproject.toml b/paddler_typesafe_client_python_test/pyproject.toml
new file mode 100644
index 000000000..9fd0344d8
--- /dev/null
+++ b/paddler_typesafe_client_python_test/pyproject.toml
@@ -0,0 +1,30 @@
+[tool.poetry]
+name = "paddler-typesafe-client-python-test"
+version = "5.0.0"
+description = "Drives Paddler's TypeSafe-compatible endpoints with the official TypeSafe SDK to verify real-client compatibility"
+authors = ["Intentee"]
+license = "Apache-2.0"
+package-mode = false
+
+[tool.poetry.dependencies]
+python = "^3.11"
+typesafe-sdk = "==0.7.5"
+
+[tool.poetry.group.dev.dependencies]
+pytest = "9.1.1"
+pyright = "1.1.414"
+mypy = "2.4.0"
+
+[tool.mypy]
+python_version = "3.11"
+strict = true
+extra_checks = true
+
+[tool.pyright]
+pythonVersion = "3.11"
+pythonPlatform = "All"
+typeCheckingMode = "strict"
+
+[tool.pytest.ini_options]
+testpaths = ["tests"]
+addopts = "-v"
diff --git a/paddler_typesafe_client_python_test/tests/conftest.py b/paddler_typesafe_client_python_test/tests/conftest.py
new file mode 100644
index 000000000..52ed76761
--- /dev/null
+++ b/paddler_typesafe_client_python_test/tests/conftest.py
@@ -0,0 +1,13 @@
+from collections.abc import Iterator
+from os import environ
+
+import pytest
+from typesafe_sdk import TypeSafeClient
+
+
+@pytest.fixture
+def typesafe_client() -> Iterator[TypeSafeClient]:
+    with TypeSafeClient(
+        api_key="paddler", base_url=environ["PADDLER_COMPAT_TYPESAFE_URL"]
+    ) as typesafe_client:
+        yield typesafe_client
diff --git a/paddler_typesafe_client_python_test/tests/test_models_list_returns_the_served_model.py b/paddler_typesafe_client_python_test/tests/test_models_list_returns_the_served_model.py
new file mode 100644
index 000000000..2f54c71e6
--- /dev/null
+++ b/paddler_typesafe_client_python_test/tests/test_models_list_returns_the_served_model.py
@@ -0,0 +1,13 @@
+from pathlib import Path
+
+from typesafe_sdk import TypeSafeClient
+
+KEV_0_8B_MODEL_PATH = str(Path(__file__).parents[2] / "target" / "kev_0_8b" / "model.gguf")
+
+
+def test_models_list_returns_the_served_model(
+    typesafe_client: TypeSafeClient,
+) -> None:
+    models = typesafe_client.models.list()
+
+    assert [model.name for model in models.models] == [KEV_0_8B_MODEL_PATH]
diff --git a/paddler_typesafe_client_python_test/tests/test_system_one_answers_every_question_type.py b/paddler_typesafe_client_python_test/tests/test_system_one_answers_every_question_type.py
new file mode 100644
index 000000000..c090a2d02
--- /dev/null
+++ b/paddler_typesafe_client_python_test/tests/test_system_one_answers_every_question_type.py
@@ -0,0 +1,30 @@
+from typesafe_sdk import Choice, Noul, Score, TypeSafeClient
+
+
+def test_system_one_answers_every_question_type(
+    typesafe_client: TypeSafeClient,
+) -> None:
+    result = typesafe_client.system_one(
+        state={"customer": "Ada", "message": "I was charged twice. Please help."},
+        questions={
+            "billing": Noul(instructions="Is this about billing?"),
+            "tone": Choice(
+                instructions="What is the tone?",
+                criteria={"calm": None, "angry": "Shouting or insults"},
+            ),
+            "urgency": Score(criteria=["Can wait", "Needs attention today"]),
+        },
+    )
+
+    assert list(result.answers) == ["billing", "tone", "urgency"]
+    assert 0 <= result.nouls["billing"].noul <= 1
+    assert result.choices["tone"].choice in {"calm", "angry"}
+    assert result.scores["urgency"].legend == {
+        0: "Can wait",
+        1: "Needs attention today",
+    }
+    assert 0 <= result.scores["urgency"].score <= 1
+    assert result.usage.input_tokens is not None
+    assert result.usage.input_tokens > 0
+    assert result.usage.output_tokens == 0
+    assert result.request_id
diff --git a/paddler_typesafe_client_python_test/tests/test_system_one_rejects_a_choice_with_more_criteria_than_paddler_accepts.py b/paddler_typesafe_client_python_test/tests/test_system_one_rejects_a_choice_with_more_criteria_than_paddler_accepts.py
new file mode 100644
index 000000000..0be1bd254
--- /dev/null
+++ b/paddler_typesafe_client_python_test/tests/test_system_one_rejects_a_choice_with_more_criteria_than_paddler_accepts.py
@@ -0,0 +1,25 @@
+import pytest
+from typesafe_sdk import Choice, TypeSafeClient, TypeSafeUnprocessableEntityError
+
+MOST_CRITERIA_PADDLER_ACCEPTS_IN_A_CHOICE = 255
+
+
+def test_system_one_rejects_a_choice_with_more_criteria_than_paddler_accepts(
+    typesafe_client: TypeSafeClient,
+) -> None:
+    with pytest.raises(TypeSafeUnprocessableEntityError) as rejection:
+        typesafe_client.system_one(
+            state="I was charged twice. Please help.",
+            questions={
+                "tone": Choice(
+                    criteria={
+                        f"option_{option_index}": None
+                        for option_index in range(
+                            MOST_CRITERIA_PADDLER_ACCEPTS_IN_A_CHOICE + 1
+                        )
+                    }
+                )
+            },
+        )
+
+    assert rejection.value.request_id
diff --git a/paddler_typesafe_translation/Cargo.toml b/paddler_typesafe_translation/Cargo.toml
new file mode 100644
index 000000000..3012eba62
--- /dev/null
+++ b/paddler_typesafe_translation/Cargo.toml
@@ -0,0 +1,23 @@
+[package]
+name = "paddler_typesafe_translation"
+autotests = false
+authors.workspace = true
+description.workspace = true
+edition.workspace = true
+homepage.workspace = true
+license.workspace = true
+repository.workspace = true
+version.workspace = true
+
+[[test]]
+name = "integration"
+path = "tests/main.rs"
+
+[dependencies]
+serde = { workspace = true }
+serde_json = { workspace = true, features = ["float_roundtrip"] }
+thiserror = { workspace = true }
+paddler_messaging = { workspace = true }
+
+[lints]
+workspace = true
diff --git a/paddler_typesafe_translation/src/choice_confidence.rs b/paddler_typesafe_translation/src/choice_confidence.rs
new file mode 100644
index 000000000..fa8373180
--- /dev/null
+++ b/paddler_typesafe_translation/src/choice_confidence.rs
@@ -0,0 +1,31 @@
+use std::num::NonZeroU8;
+
+#[must_use]
+pub fn choice_confidence(
+    probabilities: &[f64],
+    most_likely_option: usize,
+    option_count: NonZeroU8,
+) -> f64 {
+    if option_count == NonZeroU8::MIN {
+        return 1.0;
+    }
+
+    let uniform_probability = 1.0 / f64::from(option_count.get());
+    let largest_probability = probabilities[most_likely_option];
+
+    (largest_probability - uniform_probability) / (1.0 - uniform_probability)
+}
+
+#[cfg(test)]
+mod tests {
+    use std::num::NonZeroU8;
+
+    use super::choice_confidence;
+
+    #[test]
+    fn measures_the_confidence_on_normalized_probabilities() {
+        let option_count = NonZeroU8::new(2).expect("two options are a valid count");
+
+        assert!((choice_confidence(&[0.25, 0.75], 1, option_count) - 0.5).abs() < f64::EPSILON);
+    }
+}
diff --git a/paddler_typesafe_translation/src/lib.rs b/paddler_typesafe_translation/src/lib.rs
new file mode 100644
index 000000000..4b5000f7b
--- /dev/null
+++ b/paddler_typesafe_translation/src/lib.rs
@@ -0,0 +1,15 @@
+pub mod choice_confidence;
+pub mod python_float_repr;
+pub mod render_json_content;
+pub mod round_probability;
+pub mod score_confidence;
+pub mod system_one_answer;
+pub mod system_one_answer_layout;
+pub mod system_one_question;
+pub mod system_one_request;
+pub mod system_one_response;
+pub mod system_one_response_layout;
+pub mod system_one_usage;
+pub mod translated_system_one_question;
+pub mod translated_system_one_request;
+pub mod typesafe_translation_error;
diff --git a/paddler_typesafe_translation/src/python_float_repr.rs b/paddler_typesafe_translation/src/python_float_repr.rs
new file mode 100644
index 000000000..688180dd4
--- /dev/null
+++ b/paddler_typesafe_translation/src/python_float_repr.rs
@@ -0,0 +1,129 @@
+fn exponential_notation(digits: &str, exponent: i64) -> String {
+    let exponent_sign = if exponent < 0 { '-' } else { '+' };
+    let mantissa = if digits.len() > 1 {
+        format!("{}.{}", &digits[..1], &digits[1..])
+    } else {
+        digits.to_owned()
+    };
+
+    format!("{mantissa}e{exponent_sign}{:02}", exponent.abs())
+}
+
+fn positional_notation(digits: &str, decimal_point_position: i64) -> String {
+    if decimal_point_position <= 0 {
+        return format!(
+            "0.{}{digits}",
+            "0".repeat((-decimal_point_position) as usize)
+        );
+    }
+
+    let integer_digit_count = decimal_point_position as usize;
+
+    if integer_digit_count >= digits.len() {
+        format!(
+            "{digits}{}.0",
+            "0".repeat(integer_digit_count - digits.len())
+        )
+    } else {
+        format!(
+            "{}.{}",
+            &digits[..integer_digit_count],
+            &digits[integer_digit_count..]
+        )
+    }
+}
+
+struct ScientificDigits {
+    digits: String,
+    exponent: i64,
+}
+
+impl ScientificDigits {
+    fn parse(scientific_notation: &str) -> Self {
+        let mut digits = String::new();
+        let mut exponent_magnitude = 0_i64;
+        let mut exponent_is_negative = false;
+        let mut reading_exponent = false;
+
+        for character in scientific_notation.chars() {
+            match (reading_exponent, character) {
+                (false, 'e') => reading_exponent = true,
+                (false, '.') => {}
+                (false, digit) => digits.push(digit),
+                (true, '-') => exponent_is_negative = true,
+                (true, digit) => {
+                    exponent_magnitude =
+                        exponent_magnitude * 10 + i64::from(u32::from(digit) - u32::from('0'));
+                }
+            }
+        }
+
+        Self {
+            digits,
+            exponent: if exponent_is_negative {
+                -exponent_magnitude
+            } else {
+                exponent_magnitude
+            },
+        }
+    }
+
+    fn shortest_with_ties_to_even(magnitude: f64) -> Self {
+        let shortest = format!("{magnitude:e}");
+        let shortest_digits = Self::parse(&shortest);
+        let ties_to_even = format!(
+            "{magnitude:.precision$e}",
+            precision = shortest_digits.digits.len() - 1
+        );
+
+        if ties_to_even.parse::<f64>() == Ok(magnitude) {
+            Self::parse(&ties_to_even)
+        } else {
+            shortest_digits
+        }
+    }
+}
+
+#[must_use]
+pub fn python_float_repr(value: f64) -> String {
+    let ScientificDigits { digits, exponent } =
+        ScientificDigits::shortest_with_ties_to_even(value.abs());
+    let decimal_point_position = exponent + 1;
+    let sign = if value.is_sign_negative() { "-" } else { "" };
+    let unsigned = if decimal_point_position <= -4 || decimal_point_position > 16 {
+        exponential_notation(&digits, exponent)
+    } else {
+        positional_notation(&digits, decimal_point_position)
+    };
+
+    format!("{sign}{unsigned}")
+}
+
+#[cfg(test)]
+mod tests {
+    use super::python_float_repr;
+
+    #[test]
+    fn formats_floats_the_way_python_repr_does() {
+        for (value, python_repr) in [
+            (0.0, "0.0"),
+            (-0.0, "-0.0"),
+            (100.0, "100.0"),
+            (19.99, "19.99"),
+            (0.1, "0.1"),
+            (0.0001, "0.0001"),
+            (1e-05, "1e-05"),
+            (1e-07, "1e-07"),
+            (1e15, "1000000000000000.0"),
+            (1e16, "1e+16"),
+            (1.5e300, "1.5e+300"),
+            (-2.5, "-2.5"),
+            (123_456_789.123, "123456789.123"),
+            (123.456_789_012_345_67, "123.45678901234567"),
+            (1_000_000_000_000.656_2, "1000000000000.6562"),
+            (7.120_236_347_223_045e-307, "7.120236347223045e-307"),
+        ] {
+            assert_eq!(python_float_repr(value), python_repr, "{value}");
+        }
+    }
+}
diff --git a/paddler_typesafe_translation/src/render_json_content.rs b/paddler_typesafe_translation/src/render_json_content.rs
new file mode 100644
index 000000000..3041736ed
--- /dev/null
+++ b/paddler_typesafe_translation/src/render_json_content.rs
@@ -0,0 +1,51 @@
+use serde_json::Value;
+
+use crate::python_float_repr::python_float_repr;
+
+fn python_whitespace(character: char) -> bool {
+    character.is_whitespace() || ('\u{1c}'..='\u{1f}').contains(&character)
+}
+
+fn python_scalar(value: &Value) -> String {
+    match value {
+        Value::Bool(true) => "True".to_owned(),
+        Value::Bool(false) => "False".to_owned(),
+        Value::Number(number) => number
+            .as_f64()
+            .filter(|_| number.is_f64())
+            .map_or_else(|| number.to_string(), python_float_repr),
+        Value::String(text) => text.clone(),
+        Value::Array(_) | Value::Null | Value::Object(_) => String::new(),
+    }
+}
+
+#[must_use]
+pub fn render_json_content(value: &Value, indent: usize) -> String {
+    let pad = "  ".repeat(indent);
+
+    match value {
+        Value::Array(items) => items
+            .iter()
+            .map(|item| {
+                format!(
+                    "{pad}- {}",
+                    render_json_content(item, indent + 1).trim_start_matches(python_whitespace)
+                )
+            })
+            .collect::<Vec<_>>()
+            .join("\n"),
+        Value::Object(fields) => fields
+            .iter()
+            .map(|(key, field)| match field {
+                Value::Array(_) | Value::Object(_) => {
+                    format!("{pad}{key}:\n{}", render_json_content(field, indent + 1))
+                }
+                Value::Bool(_) | Value::Null | Value::Number(_) | Value::String(_) => {
+                    format!("{pad}{key}: {}", python_scalar(field))
+                }
+            })
+            .collect::<Vec<_>>()
+            .join("\n"),
+        Value::Bool(_) | Value::Null | Value::Number(_) | Value::String(_) => python_scalar(value),
+    }
+}
diff --git a/paddler_typesafe_translation/src/round_probability.rs b/paddler_typesafe_translation/src/round_probability.rs
new file mode 100644
index 000000000..16f954caa
--- /dev/null
+++ b/paddler_typesafe_translation/src/round_probability.rs
@@ -0,0 +1,17 @@
+const PROBABILITY_DECIMAL_SCALE: f64 = 10_000.0;
+
+#[must_use]
+pub fn round_probability(value: f64) -> f64 {
+    (value * PROBABILITY_DECIMAL_SCALE).round_ties_even() / PROBABILITY_DECIMAL_SCALE
+}
+
+#[cfg(test)]
+mod tests {
+    use super::round_probability;
+
+    #[test]
+    fn keeps_four_decimals_and_rounds_exact_ties_to_even() {
+        assert!((round_probability(0.876_543_22) - 0.8765).abs() < f64::EPSILON);
+        assert!((round_probability(0.031_25) - 0.0312).abs() < f64::EPSILON);
+    }
+}
diff --git a/paddler_typesafe_translation/src/score_confidence.rs b/paddler_typesafe_translation/src/score_confidence.rs
new file mode 100644
index 000000000..2cbdb14d6
--- /dev/null
+++ b/paddler_typesafe_translation/src/score_confidence.rs
@@ -0,0 +1,30 @@
+use std::num::NonZeroU8;
+
+#[must_use]
+pub fn score_confidence(probabilities: &[f64], level_count: NonZeroU8) -> f64 {
+    if level_count == NonZeroU8::MIN {
+        return 1.0;
+    }
+
+    let level_count_f64 = f64::from(level_count.get());
+    let mut most_likely_level: u8 = 0;
+
+    for (level, probability) in (0..=u8::MAX).zip(probabilities) {
+        if *probability > probabilities[usize::from(most_likely_level)] {
+            most_likely_level = level;
+        }
+    }
+
+    let uniform_deviation = (0..level_count.get())
+        .map(|level| (f64::from(level) - (level_count_f64 - 1.0) / 2.0).abs())
+        .sum::<f64>()
+        / level_count_f64;
+    let expected_deviation: f64 = (0..=u8::MAX)
+        .zip(probabilities)
+        .map(|(level, probability)| {
+            probability * (f64::from(level) - f64::from(most_likely_level)).abs()
+        })
+        .sum();
+
+    (1.0 - expected_deviation / uniform_deviation).max(0.0)
+}
diff --git a/paddler_typesafe_translation/src/system_one_answer.rs b/paddler_typesafe_translation/src/system_one_answer.rs
new file mode 100644
index 000000000..54ea38854
--- /dev/null
+++ b/paddler_typesafe_translation/src/system_one_answer.rs
@@ -0,0 +1,52 @@
+use serde_json::Map;
+use serde_json::Value;
+use serde_json::json;
+
+#[derive(Clone, Debug, PartialEq)]
+pub enum SystemOneAnswer {
+    Choice {
+        choice: String,
+        confidence: f64,
+        probabilities: Map<String, Value>,
+    },
+    Noul {
+        noul: f64,
+    },
+    Score {
+        score: f64,
+        legend: Map<String, Value>,
+        probabilities: Map<String, Value>,
+        confidence: f64,
+    },
+}
+
+impl SystemOneAnswer {
+    #[must_use]
+    pub fn into_value(self) -> Value {
+        match self {
+            Self::Choice {
+                choice,
+                confidence,
+                probabilities,
+            } => json!({
+                "type": "choice",
+                "choice": choice,
+                "confidence": confidence,
+                "probabilities": probabilities,
+            }),
+            Self::Noul { noul } => json!({ "type": "noul", "noul": noul }),
+            Self::Score {
+                score,
+                legend,
+                probabilities,
+                confidence,
+            } => json!({
+                "type": "score",
+                "score": score,
+                "legend": legend,
+                "probabilities": probabilities,
+                "confidence": confidence,
+            }),
+        }
+    }
+}
diff --git a/paddler_typesafe_translation/src/system_one_answer_layout.rs b/paddler_typesafe_translation/src/system_one_answer_layout.rs
new file mode 100644
index 000000000..8d4ed4a55
--- /dev/null
+++ b/paddler_typesafe_translation/src/system_one_answer_layout.rs
@@ -0,0 +1,124 @@
+use std::num::NonZeroU8;
+
+use serde_json::Map;
+use serde_json::Value;
+
+use crate::choice_confidence::choice_confidence;
+use crate::round_probability::round_probability;
+use crate::score_confidence::score_confidence;
+use crate::system_one_answer::SystemOneAnswer;
+use crate::typesafe_translation_error::TypeSafeTranslationError;
+
+const NOUL_OPTIONS: usize = 2;
+
+fn rounded_probabilities<'keys>(
+    keys: impl Iterator<Item = &'keys String>,
+    probabilities: &[f64],
+) -> Map<String, Value> {
+    keys.zip(probabilities)
+        .map(|(key, probability)| (key.clone(), Value::from(round_probability(*probability))))
+        .collect()
+}
+
+fn most_likely_option(probabilities: &[f64]) -> usize {
+    let mut most_likely = 0;
+
+    for (option, probability) in probabilities.iter().enumerate() {
+        if *probability > probabilities[most_likely] {
+            most_likely = option;
+        }
+    }
+
+    most_likely
+}
+
+#[derive(Clone, Debug, Eq, PartialEq)]
+pub enum SystemOneAnswerLayout {
+    Choice {
+        id: String,
+        keys: Vec<String>,
+        option_count: NonZeroU8,
+    },
+    Noul {
+        id: String,
+    },
+    Score {
+        id: String,
+        legend: Vec<String>,
+        level_count: NonZeroU8,
+    },
+}
+
+impl SystemOneAnswerLayout {
+    #[must_use]
+    pub fn id(&self) -> &str {
+        match self {
+            Self::Choice { id, .. } | Self::Noul { id } | Self::Score { id, .. } => id,
+        }
+    }
+
+    pub fn answer(
+        &self,
+        probabilities: &[f32],
+    ) -> Result<SystemOneAnswer, TypeSafeTranslationError> {
+        let probabilities: Vec<f64> = probabilities.iter().copied().map(f64::from).collect();
+        let expected = match self {
+            Self::Choice { option_count, .. } => usize::from(option_count.get()),
+            Self::Noul { .. } => NOUL_OPTIONS,
+            Self::Score { level_count, .. } => usize::from(level_count.get()),
+        };
+
+        if probabilities.len() != expected {
+            return Err(TypeSafeTranslationError::AnswerProbabilitiesMismatch {
+                expected,
+                question_id: self.id().to_owned(),
+                received: probabilities.len(),
+            });
+        }
+
+        Ok(match self {
+            Self::Choice {
+                keys, option_count, ..
+            } => {
+                let most_likely = most_likely_option(&probabilities);
+
+                SystemOneAnswer::Choice {
+                    choice: keys[most_likely].clone(),
+                    confidence: round_probability(choice_confidence(
+                        &probabilities,
+                        most_likely,
+                        *option_count,
+                    )),
+                    probabilities: rounded_probabilities(keys.iter(), &probabilities),
+                }
+            }
+            Self::Noul { .. } => SystemOneAnswer::Noul {
+                noul: round_probability(probabilities[1]),
+            },
+            Self::Score {
+                legend,
+                level_count,
+                ..
+            } => {
+                let level_keys: Vec<String> = (0..level_count.get())
+                    .map(|level| level.to_string())
+                    .collect();
+                let score: f64 = (0..=u8::MAX)
+                    .zip(&probabilities)
+                    .map(|(level, probability)| f64::from(level) * probability)
+                    .sum();
+
+                SystemOneAnswer::Score {
+                    score: round_probability(score),
+                    legend: level_keys
+                        .iter()
+                        .cloned()
+                        .zip(legend.iter().cloned().map(Value::from))
+                        .collect(),
+                    probabilities: rounded_probabilities(level_keys.iter(), &probabilities),
+                    confidence: round_probability(score_confidence(&probabilities, *level_count)),
+                }
+            }
+        })
+    }
+}
diff --git a/paddler_typesafe_translation/src/system_one_question.rs b/paddler_typesafe_translation/src/system_one_question.rs
new file mode 100644
index 000000000..e558873c5
--- /dev/null
+++ b/paddler_typesafe_translation/src/system_one_question.rs
@@ -0,0 +1,34 @@
+use serde::Deserialize;
+use serde::Deserializer;
+use serde_json::Map;
+use serde_json::Value;
+
+fn criteria_or_none_described<'de, TDeserializer>(
+    deserializer: TDeserializer,
+) -> Result<Map<String, Value>, TDeserializer::Error>
+where
+    TDeserializer: Deserializer<'de>,
+{
+    Ok(Option::<Map<String, Value>>::deserialize(deserializer)?.unwrap_or_default())
+}
+
+#[derive(Clone, Debug, Deserialize, Eq, PartialEq)]
+#[serde(deny_unknown_fields, rename_all = "lowercase", tag = "type")]
+pub enum SystemOneQuestion {
+    Choice {
+        criteria: Map<String, Value>,
+        #[serde(default)]
+        instructions: Value,
+    },
+    Noul {
+        #[serde(default, deserialize_with = "criteria_or_none_described")]
+        criteria: Map<String, Value>,
+        #[serde(default)]
+        instructions: Value,
+    },
+    Score {
+        criteria: Vec<Value>,
+        #[serde(default)]
+        instructions: Value,
+    },
+}
diff --git a/paddler_typesafe_translation/src/system_one_request.rs b/paddler_typesafe_translation/src/system_one_request.rs
new file mode 100644
index 000000000..4d950cf63
--- /dev/null
+++ b/paddler_typesafe_translation/src/system_one_request.rs
@@ -0,0 +1,154 @@
+use std::num::NonZeroU8;
+
+use serde::Deserialize;
+use serde_json::Map;
+use serde_json::Value;
+use serde_json::from_value;
+
+use paddler_messaging::decision_question::DecisionQuestion;
+use paddler_messaging::request_params::decide_params::raw_decide_params::RawDecideParams;
+use paddler_messaging::validates::Validates as _;
+
+use crate::render_json_content::render_json_content;
+use crate::system_one_answer_layout::SystemOneAnswerLayout;
+use crate::system_one_question::SystemOneQuestion;
+use crate::system_one_response_layout::SystemOneResponseLayout;
+use crate::translated_system_one_question::TranslatedSystemOneQuestion;
+use crate::translated_system_one_request::TranslatedSystemOneRequest;
+use crate::typesafe_translation_error::TypeSafeTranslationError;
+
+fn option_text(name: &str, description: Option<&Value>) -> String {
+    match description {
+        None | Some(Value::Null) => name.to_owned(),
+        Some(Value::String(text)) if text.is_empty() => name.to_owned(),
+        Some(description) => format!("{name}: {}", render_json_content(description, 0)),
+    }
+}
+
+fn criteria_count(criteria_len: usize) -> Option<NonZeroU8> {
+    NonZeroU8::new(u8::try_from(criteria_len).ok()?)
+}
+
+fn translate_question(
+    question_id: String,
+    question: SystemOneQuestion,
+) -> Result<TranslatedSystemOneQuestion, TypeSafeTranslationError> {
+    match question {
+        SystemOneQuestion::Choice {
+            criteria,
+            instructions,
+        } => {
+            let Some(option_count) = criteria_count(criteria.len()) else {
+                return Err(TypeSafeTranslationError::ChoiceCriteriaOutOfRange {
+                    criteria_count: criteria.len(),
+                    question_id,
+                });
+            };
+
+            Ok(TranslatedSystemOneQuestion {
+                decision_question: DecisionQuestion {
+                    id: question_id.clone(),
+                    instructions: render_json_content(&instructions, 0),
+                    options: criteria
+                        .iter()
+                        .map(|(name, description)| option_text(name, Some(description)))
+                        .collect(),
+                },
+                answer_layout: SystemOneAnswerLayout::Choice {
+                    id: question_id,
+                    keys: criteria.keys().cloned().collect(),
+                    option_count,
+                },
+            })
+        }
+        SystemOneQuestion::Noul {
+            criteria,
+            instructions,
+        } => Ok(TranslatedSystemOneQuestion {
+            decision_question: DecisionQuestion {
+                id: question_id.clone(),
+                instructions: render_json_content(&instructions, 0),
+                options: vec![
+                    option_text("no", criteria.get("false")),
+                    option_text("yes", criteria.get("true")),
+                ],
+            },
+            answer_layout: SystemOneAnswerLayout::Noul { id: question_id },
+        }),
+        SystemOneQuestion::Score {
+            criteria,
+            instructions,
+        } => {
+            let Some(level_count) = criteria_count(criteria.len()) else {
+                return Err(TypeSafeTranslationError::ScoreCriteriaOutOfRange {
+                    criteria_count: criteria.len(),
+                    question_id,
+                });
+            };
+
+            let legend: Vec<String> = criteria
+                .iter()
+                .map(|level| render_json_content(level, 0))
+                .collect();
+
+            Ok(TranslatedSystemOneQuestion {
+                decision_question: DecisionQuestion {
+                    id: question_id.clone(),
+                    instructions: render_json_content(&instructions, 0),
+                    options: legend.clone(),
+                },
+                answer_layout: SystemOneAnswerLayout::Score {
+                    id: question_id,
+                    legend,
+                    level_count,
+                },
+            })
+        }
+    }
+}
+
+#[derive(Clone, Debug, Deserialize, Eq, PartialEq)]
+#[serde(deny_unknown_fields)]
+pub struct SystemOneRequest {
+    pub model: String,
+    pub questions: Map<String, Value>,
+    pub state: Value,
+}
+
+impl SystemOneRequest {
+    pub fn translate(self) -> Result<TranslatedSystemOneRequest, TypeSafeTranslationError> {
+        let mut answer_layouts = Vec::with_capacity(self.questions.len());
+        let mut questions = Vec::with_capacity(self.questions.len());
+
+        for (question_id, question) in self.questions {
+            let question = from_value::<SystemOneQuestion>(question).map_err(|source| {
+                TypeSafeTranslationError::QuestionMalformed {
+                    question_id: question_id.clone(),
+                    source,
+                }
+            })?;
+            let TranslatedSystemOneQuestion {
+                answer_layout,
+                decision_question,
+            } = translate_question(question_id, question)?;
+
+            questions.push(decision_question);
+            answer_layouts.push(answer_layout);
+        }
+
+        let decide_params = RawDecideParams {
+            questions,
+            state: render_json_content(&self.state, 0),
+        }
+        .validate()
+        .map_err(|source| TypeSafeTranslationError::DecisionInvalid { source })?;
+
+        Ok(TranslatedSystemOneRequest {
+            decide_params,
+            response_layout: SystemOneResponseLayout {
+                answer_layouts,
+                model: self.model,
+            },
+        })
+    }
+}
diff --git a/paddler_typesafe_translation/src/system_one_response.rs b/paddler_typesafe_translation/src/system_one_response.rs
new file mode 100644
index 000000000..362e3b631
--- /dev/null
+++ b/paddler_typesafe_translation/src/system_one_response.rs
@@ -0,0 +1,13 @@
+use serde::Serialize;
+use serde_json::Map;
+use serde_json::Value;
+
+use crate::system_one_usage::SystemOneUsage;
+
+#[derive(Clone, Debug, Eq, PartialEq, Serialize)]
+pub struct SystemOneResponse {
+    pub model: String,
+    pub answers: Map<String, Value>,
+    pub usage: SystemOneUsage,
+    pub latency_ms: u64,
+}
diff --git a/paddler_typesafe_translation/src/system_one_response_layout.rs b/paddler_typesafe_translation/src/system_one_response_layout.rs
new file mode 100644
index 000000000..71b32a6c0
--- /dev/null
+++ b/paddler_typesafe_translation/src/system_one_response_layout.rs
@@ -0,0 +1,59 @@
+use serde_json::Map;
+
+use paddler_messaging::decision_answer::DecisionAnswer;
+use paddler_messaging::decision_summary::DecisionSummary;
+
+use crate::system_one_answer_layout::SystemOneAnswerLayout;
+use crate::system_one_response::SystemOneResponse;
+use crate::system_one_usage::SystemOneUsage;
+use crate::typesafe_translation_error::TypeSafeTranslationError;
+
+#[derive(Clone, Debug, Eq, PartialEq)]
+pub struct SystemOneResponseLayout {
+    pub answer_layouts: Vec<SystemOneAnswerLayout>,
+    pub model: String,
+}
+
+impl SystemOneResponseLayout {
+    pub fn respond(
+        &self,
+        decision_answers: &[DecisionAnswer],
+        DecisionSummary {
+            input_tokens,
+            processing_milliseconds,
+        }: &DecisionSummary,
+    ) -> Result<SystemOneResponse, TypeSafeTranslationError> {
+        let mut answers = Map::new();
+        let mut decision_answers = decision_answers.iter();
+
+        for answer_layout in &self.answer_layouts {
+            let Some(DecisionAnswer { id, probabilities }) = decision_answers.next() else {
+                return Err(TypeSafeTranslationError::AnswerMissing {
+                    question_id: answer_layout.id().to_owned(),
+                });
+            };
+
+            if id != answer_layout.id() {
+                return Err(TypeSafeTranslationError::AnswerOutOfOrder {
+                    expected_question_id: answer_layout.id().to_owned(),
+                    received_question_id: id.clone(),
+                });
+            }
+
+            answers.insert(
+                id.clone(),
+                answer_layout.answer(probabilities)?.into_value(),
+            );
+        }
+
+        Ok(SystemOneResponse {
+            model: self.model.clone(),
+            answers,
+            usage: SystemOneUsage {
+                input_tokens: *input_tokens,
+                output_tokens: 0,
+            },
+            latency_ms: *processing_milliseconds,
+        })
+    }
+}
diff --git a/paddler_typesafe_translation/src/system_one_usage.rs b/paddler_typesafe_translation/src/system_one_usage.rs
new file mode 100644
index 000000000..c0602532f
--- /dev/null
+++ b/paddler_typesafe_translation/src/system_one_usage.rs
@@ -0,0 +1,7 @@
+use serde::Serialize;
+
+#[derive(Clone, Debug, Eq, PartialEq, Serialize)]
+pub struct SystemOneUsage {
+    pub input_tokens: usize,
+    pub output_tokens: usize,
+}
diff --git a/paddler_typesafe_translation/src/translated_system_one_question.rs b/paddler_typesafe_translation/src/translated_system_one_question.rs
new file mode 100644
index 000000000..838a8e2c1
--- /dev/null
+++ b/paddler_typesafe_translation/src/translated_system_one_question.rs
@@ -0,0 +1,8 @@
+use paddler_messaging::decision_question::DecisionQuestion;
+
+use crate::system_one_answer_layout::SystemOneAnswerLayout;
+
+pub struct TranslatedSystemOneQuestion {
+    pub answer_layout: SystemOneAnswerLayout,
+    pub decision_question: DecisionQuestion,
+}
diff --git a/paddler_typesafe_translation/src/translated_system_one_request.rs b/paddler_typesafe_translation/src/translated_system_one_request.rs
new file mode 100644
index 000000000..3f95d852c
--- /dev/null
+++ b/paddler_typesafe_translation/src/translated_system_one_request.rs
@@ -0,0 +1,9 @@
+use paddler_messaging::request_params::decide_params::DecideParams;
+
+use crate::system_one_response_layout::SystemOneResponseLayout;
+
+#[derive(Clone, Debug, Eq, PartialEq)]
+pub struct TranslatedSystemOneRequest {
+    pub decide_params: DecideParams,
+    pub response_layout: SystemOneResponseLayout,
+}
diff --git a/paddler_typesafe_translation/src/typesafe_translation_error.rs b/paddler_typesafe_translation/src/typesafe_translation_error.rs
new file mode 100644
index 000000000..93cc42893
--- /dev/null
+++ b/paddler_typesafe_translation/src/typesafe_translation_error.rs
@@ -0,0 +1,52 @@
+use serde_json::Error as JsonError;
+
+use paddler_messaging::request_params_validation_error::RequestParamsValidationError;
+
+#[derive(Debug, thiserror::Error)]
+pub enum TypeSafeTranslationError {
+    #[error(
+        "question {question_id} expected {expected} probabilities, the agent answered with {received}"
+    )]
+    AnswerProbabilitiesMismatch {
+        expected: usize,
+        question_id: String,
+        received: usize,
+    },
+
+    #[error("the agent finished without answering question {question_id}")]
+    AnswerMissing { question_id: String },
+
+    #[error(
+        "the agent answered question {received_question_id} where {expected_question_id} was due"
+    )]
+    AnswerOutOfOrder {
+        expected_question_id: String,
+        received_question_id: String,
+    },
+
+    #[error(
+        "choice question {question_id} has {criteria_count} criteria, it needs between 1 and 255"
+    )]
+    ChoiceCriteriaOutOfRange {
+        criteria_count: usize,
+        question_id: String,
+    },
+
+    #[error("the System One request is not a valid decision: {source}")]
+    DecisionInvalid {
+        #[source]
+        source: RequestParamsValidationError,
+    },
+    #[error("question {question_id} is malformed: {source}")]
+    QuestionMalformed {
+        question_id: String,
+        #[source]
+        source: JsonError,
+    },
+
+    #[error("score question {question_id} has {criteria_count} levels, it needs between 1 and 255")]
+    ScoreCriteriaOutOfRange {
+        criteria_count: usize,
+        question_id: String,
+    },
+}
diff --git a/paddler_typesafe_translation/tests/a_choice_question_with_too_many_criteria_is_refused.rs b/paddler_typesafe_translation/tests/a_choice_question_with_too_many_criteria_is_refused.rs
new file mode 100644
index 000000000..50e8b7d14
--- /dev/null
+++ b/paddler_typesafe_translation/tests/a_choice_question_with_too_many_criteria_is_refused.rs
@@ -0,0 +1,32 @@
+use serde_json::Map;
+use serde_json::Value;
+use serde_json::json;
+
+use paddler_typesafe_translation::typesafe_translation_error::TypeSafeTranslationError;
+
+use crate::system_one_request_with_question::system_one_request_with_question;
+
+const TOO_MANY_CRITERIA: usize = 256;
+
+#[test]
+fn a_choice_question_with_too_many_criteria_is_refused() {
+    let too_many_choices: Map<String, Value> = (0..TOO_MANY_CRITERIA)
+        .map(|option| (format!("option-{option}"), Value::Null))
+        .collect();
+
+    assert_eq!(
+        system_one_request_with_question(json!({"type": "choice", "criteria": too_many_choices}))
+            .expect("the request must parse")
+            .translate()
+            .err()
+            .as_ref()
+            .map(ToString::to_string),
+        Some(
+            TypeSafeTranslationError::ChoiceCriteriaOutOfRange {
+                criteria_count: TOO_MANY_CRITERIA,
+                question_id: "question".to_owned(),
+            }
+            .to_string()
+        )
+    );
+}
diff --git a/paddler_typesafe_translation/tests/a_malformed_question_is_refused_with_its_id.rs b/paddler_typesafe_translation/tests/a_malformed_question_is_refused_with_its_id.rs
new file mode 100644
index 000000000..ff98fceb1
--- /dev/null
+++ b/paddler_typesafe_translation/tests/a_malformed_question_is_refused_with_its_id.rs
@@ -0,0 +1,31 @@
+use serde_json::from_value;
+use serde_json::json;
+
+use paddler_typesafe_translation::system_one_question::SystemOneQuestion;
+use paddler_typesafe_translation::typesafe_translation_error::TypeSafeTranslationError;
+
+use crate::system_one_request_with_questions::system_one_request_with_questions;
+
+#[test]
+fn a_malformed_question_is_refused_with_its_id() {
+    for malformed_question in [
+        json!({"type": "multiple_choice"}),
+        json!({"type": "noul", "criteria": "paid on time"}),
+    ] {
+        let question_error = from_value::<SystemOneQuestion>(malformed_question.clone())
+            .expect_err("the question must be malformed");
+        let request = system_one_request_with_questions(json!({"broken": malformed_question}))
+            .expect("the request must parse");
+
+        assert_eq!(
+            request.translate().err().as_ref().map(ToString::to_string),
+            Some(
+                TypeSafeTranslationError::QuestionMalformed {
+                    question_id: "broken".to_owned(),
+                    source: question_error,
+                }
+                .to_string()
+            )
+        );
+    }
+}
diff --git a/paddler_typesafe_translation/tests/a_noul_question_with_null_criteria_describes_neither_outcome.rs b/paddler_typesafe_translation/tests/a_noul_question_with_null_criteria_describes_neither_outcome.rs
new file mode 100644
index 000000000..9d69446b9
--- /dev/null
+++ b/paddler_typesafe_translation/tests/a_noul_question_with_null_criteria_describes_neither_outcome.rs
@@ -0,0 +1,17 @@
+use serde_json::json;
+
+use crate::system_one_request_with_questions::system_one_request_with_questions;
+
+#[test]
+fn a_noul_question_with_null_criteria_describes_neither_outcome() {
+    let translated =
+        system_one_request_with_questions(json!({"paid": {"type": "noul", "criteria": null}}))
+            .expect("the request must parse")
+            .translate()
+            .expect("the request must translate");
+
+    assert_eq!(
+        translated.decide_params.last_question.options,
+        vec!["no".to_owned(), "yes".to_owned()]
+    );
+}
diff --git a/paddler_typesafe_translation/tests/a_question_left_unanswered_is_refused.rs b/paddler_typesafe_translation/tests/a_question_left_unanswered_is_refused.rs
new file mode 100644
index 000000000..87fb5e587
--- /dev/null
+++ b/paddler_typesafe_translation/tests/a_question_left_unanswered_is_refused.rs
@@ -0,0 +1,16 @@
+use paddler_typesafe_translation::typesafe_translation_error::TypeSafeTranslationError;
+
+use crate::refusal_of_a_lone_answer_to_two_noul_questions::refusal_of_a_lone_answer_to_two_noul_questions;
+
+#[test]
+fn a_question_left_unanswered_is_refused() {
+    assert_eq!(
+        refusal_of_a_lone_answer_to_two_noul_questions("first", vec![0.5, 0.5]),
+        Some(
+            TypeSafeTranslationError::AnswerMissing {
+                question_id: "second".to_owned(),
+            }
+            .to_string()
+        )
+    );
+}
diff --git a/paddler_typesafe_translation/tests/a_score_question_without_criteria_is_refused.rs b/paddler_typesafe_translation/tests/a_score_question_without_criteria_is_refused.rs
new file mode 100644
index 000000000..59e8d1537
--- /dev/null
+++ b/paddler_typesafe_translation/tests/a_score_question_without_criteria_is_refused.rs
@@ -0,0 +1,24 @@
+use serde_json::json;
+
+use paddler_typesafe_translation::typesafe_translation_error::TypeSafeTranslationError;
+
+use crate::system_one_request_with_question::system_one_request_with_question;
+
+#[test]
+fn a_score_question_without_criteria_is_refused() {
+    assert_eq!(
+        system_one_request_with_question(json!({"type": "score", "criteria": []}))
+            .expect("the request must parse")
+            .translate()
+            .err()
+            .as_ref()
+            .map(ToString::to_string),
+        Some(
+            TypeSafeTranslationError::ScoreCriteriaOutOfRange {
+                criteria_count: 0,
+                question_id: "question".to_owned(),
+            }
+            .to_string()
+        )
+    );
+}
diff --git a/paddler_typesafe_translation/tests/a_system_one_request_without_questions_is_refused.rs b/paddler_typesafe_translation/tests/a_system_one_request_without_questions_is_refused.rs
new file mode 100644
index 000000000..1786e8c16
--- /dev/null
+++ b/paddler_typesafe_translation/tests/a_system_one_request_without_questions_is_refused.rs
@@ -0,0 +1,21 @@
+use serde_json::json;
+
+use paddler_messaging::request_params_validation_error::RequestParamsValidationError;
+use paddler_typesafe_translation::typesafe_translation_error::TypeSafeTranslationError;
+
+use crate::system_one_request_with_questions::system_one_request_with_questions;
+
+#[test]
+fn a_system_one_request_without_questions_is_refused() {
+    let request = system_one_request_with_questions(json!({})).expect("the request must parse");
+
+    assert_eq!(
+        request.translate().err().map(|error| error.to_string()),
+        Some(
+            TypeSafeTranslationError::DecisionInvalid {
+                source: RequestParamsValidationError::DecisionWithoutQuestions,
+            }
+            .to_string()
+        )
+    );
+}
diff --git a/paddler_typesafe_translation/tests/an_answer_out_of_question_order_is_refused.rs b/paddler_typesafe_translation/tests/an_answer_out_of_question_order_is_refused.rs
new file mode 100644
index 000000000..3640be3aa
--- /dev/null
+++ b/paddler_typesafe_translation/tests/an_answer_out_of_question_order_is_refused.rs
@@ -0,0 +1,17 @@
+use paddler_typesafe_translation::typesafe_translation_error::TypeSafeTranslationError;
+
+use crate::refusal_of_a_lone_answer_to_two_noul_questions::refusal_of_a_lone_answer_to_two_noul_questions;
+
+#[test]
+fn an_answer_out_of_question_order_is_refused() {
+    assert_eq!(
+        refusal_of_a_lone_answer_to_two_noul_questions("second", vec![0.5, 0.5]),
+        Some(
+            TypeSafeTranslationError::AnswerOutOfOrder {
+                expected_question_id: "first".to_owned(),
+                received_question_id: "second".to_owned(),
+            }
+            .to_string()
+        )
+    );
+}
diff --git a/paddler_typesafe_translation/tests/an_answer_whose_probabilities_do_not_match_its_options_is_refused.rs b/paddler_typesafe_translation/tests/an_answer_whose_probabilities_do_not_match_its_options_is_refused.rs
new file mode 100644
index 000000000..0e6c21575
--- /dev/null
+++ b/paddler_typesafe_translation/tests/an_answer_whose_probabilities_do_not_match_its_options_is_refused.rs
@@ -0,0 +1,18 @@
+use paddler_typesafe_translation::typesafe_translation_error::TypeSafeTranslationError;
+
+use crate::refusal_of_a_lone_answer_to_two_noul_questions::refusal_of_a_lone_answer_to_two_noul_questions;
+
+#[test]
+fn an_answer_whose_probabilities_do_not_match_its_options_is_refused() {
+    assert_eq!(
+        refusal_of_a_lone_answer_to_two_noul_questions("first", vec![1.0]),
+        Some(
+            TypeSafeTranslationError::AnswerProbabilitiesMismatch {
+                expected: 2,
+                question_id: "first".to_owned(),
+                received: 1,
+            }
+            .to_string()
+        )
+    );
+}
diff --git a/paddler_typesafe_translation/tests/main.rs b/paddler_typesafe_translation/tests/main.rs
new file mode 100644
index 000000000..9b2095c89
--- /dev/null
+++ b/paddler_typesafe_translation/tests/main.rs
@@ -0,0 +1,15 @@
+mod a_choice_question_with_too_many_criteria_is_refused;
+mod a_malformed_question_is_refused_with_its_id;
+mod a_noul_question_with_null_criteria_describes_neither_outcome;
+mod a_question_left_unanswered_is_refused;
+mod a_score_question_without_criteria_is_refused;
+mod a_system_one_request_without_questions_is_refused;
+mod an_answer_out_of_question_order_is_refused;
+mod an_answer_whose_probabilities_do_not_match_its_options_is_refused;
+#[cfg(test)]
+mod refusal_of_a_lone_answer_to_two_noul_questions;
+#[cfg(test)]
+mod system_one_request_with_question;
+#[cfg(test)]
+mod system_one_request_with_questions;
+mod translation_matches_the_kev_typesafe_reference;
diff --git a/paddler_typesafe_translation/tests/refusal_of_a_lone_answer_to_two_noul_questions.rs b/paddler_typesafe_translation/tests/refusal_of_a_lone_answer_to_two_noul_questions.rs
new file mode 100644
index 000000000..f9ac4c7c1
--- /dev/null
+++ b/paddler_typesafe_translation/tests/refusal_of_a_lone_answer_to_two_noul_questions.rs
@@ -0,0 +1,34 @@
+use serde_json::json;
+
+use paddler_messaging::decision_answer::DecisionAnswer;
+use paddler_messaging::decision_summary::DecisionSummary;
+
+use crate::system_one_request_with_questions::system_one_request_with_questions;
+
+const SUMMARY: DecisionSummary = DecisionSummary {
+    input_tokens: 1,
+    processing_milliseconds: 1,
+};
+
+pub fn refusal_of_a_lone_answer_to_two_noul_questions(
+    id: &str,
+    probabilities: Vec<f32>,
+) -> Option<String> {
+    system_one_request_with_questions(
+        json!({"first": {"type": "noul"}, "second": {"type": "noul"}}),
+    )
+    .expect("the request must parse")
+    .translate()
+    .expect("the request must translate")
+    .response_layout
+    .respond(
+        &[DecisionAnswer {
+            id: id.to_owned(),
+            probabilities,
+        }],
+        &SUMMARY,
+    )
+    .err()
+    .as_ref()
+    .map(ToString::to_string)
+}
diff --git a/paddler_typesafe_translation/tests/system_one_request_with_question.rs b/paddler_typesafe_translation/tests/system_one_request_with_question.rs
new file mode 100644
index 000000000..46e7ef05d
--- /dev/null
+++ b/paddler_typesafe_translation/tests/system_one_request_with_question.rs
@@ -0,0 +1,11 @@
+use serde_json::Error as JsonError;
+use serde_json::Value;
+use serde_json::json;
+
+use paddler_typesafe_translation::system_one_request::SystemOneRequest;
+
+use crate::system_one_request_with_questions::system_one_request_with_questions;
+
+pub fn system_one_request_with_question(question: Value) -> Result<SystemOneRequest, JsonError> {
+    system_one_request_with_questions(json!({"question": question}))
+}
diff --git a/paddler_typesafe_translation/tests/system_one_request_with_questions.rs b/paddler_typesafe_translation/tests/system_one_request_with_questions.rs
new file mode 100644
index 000000000..178cdb845
--- /dev/null
+++ b/paddler_typesafe_translation/tests/system_one_request_with_questions.rs
@@ -0,0 +1,10 @@
+use serde_json::Error as JsonError;
+use serde_json::Value;
+use serde_json::from_value;
+use serde_json::json;
+
+use paddler_typesafe_translation::system_one_request::SystemOneRequest;
+
+pub fn system_one_request_with_questions(questions: Value) -> Result<SystemOneRequest, JsonError> {
+    from_value(json!({"model": "test", "state": "state", "questions": questions}))
+}
diff --git a/paddler_typesafe_translation/tests/translation_matches_the_kev_typesafe_reference.rs b/paddler_typesafe_translation/tests/translation_matches_the_kev_typesafe_reference.rs
new file mode 100644
index 000000000..7cc7286d8
--- /dev/null
+++ b/paddler_typesafe_translation/tests/translation_matches_the_kev_typesafe_reference.rs
@@ -0,0 +1,100 @@
+use std::fs::read_to_string;
+
+use serde::Deserialize;
+use serde_json::Value;
+use serde_json::from_str;
+use serde_json::from_value;
+use serde_json::to_value;
+
+use paddler_messaging::decision_answer::DecisionAnswer;
+use paddler_messaging::decision_summary::DecisionSummary;
+use paddler_typesafe_translation::system_one_request::SystemOneRequest;
+
+#[derive(Deserialize)]
+struct RenderedQuestion {
+    instructions: String,
+    options: Vec<String>,
+}
+
+#[derive(Deserialize)]
+struct Rendered {
+    questions: Vec<RenderedQuestion>,
+    state: String,
+}
+
+#[derive(Deserialize)]
+struct TypeSafeReference {
+    answers: Value,
+    probabilities: Vec<Vec<f32>>,
+    rendered: Rendered,
+    request: Value,
+}
+
+#[test]
+fn translation_matches_the_kev_typesafe_reference() {
+    let references: Vec<TypeSafeReference> = from_str(
+        &read_to_string(concat!(
+            env!("CARGO_MANIFEST_DIR"),
+            "/../fixtures/typesafe_reference.json"
+        ))
+        .expect("the TypeSafe reference must be readable"),
+    )
+    .expect("the TypeSafe reference must parse");
+
+    for TypeSafeReference {
+        answers,
+        probabilities,
+        rendered,
+        request,
+    } in references
+    {
+        let translated = from_value::<SystemOneRequest>(request)
+            .expect("the reference request must parse")
+            .translate()
+            .expect("the reference request must translate");
+
+        assert_eq!(translated.decide_params.state, rendered.state);
+        let translated_questions: Vec<_> = translated
+            .decide_params
+            .leading_questions
+            .iter()
+            .chain([&translated.decide_params.last_question])
+            .collect();
+
+        assert_eq!(
+            translated_questions
+                .iter()
+                .map(|question| (question.instructions.clone(), question.options.clone()))
+                .collect::<Vec<_>>(),
+            rendered
+                .questions
+                .into_iter()
+                .map(|question| (question.instructions, question.options))
+                .collect::<Vec<_>>()
+        );
+
+        let decision_answers: Vec<DecisionAnswer> = translated_questions
+            .iter()
+            .zip(probabilities)
+            .map(|(question, probabilities)| DecisionAnswer {
+                id: question.id.clone(),
+                probabilities,
+            })
+            .collect();
+        let response = translated
+            .response_layout
+            .respond(
+                &decision_answers,
+                &DecisionSummary {
+                    input_tokens: 1,
+                    processing_milliseconds: 1,
+                },
+            )
+            .expect("the reference answers must translate");
+
+        assert_eq!(
+            to_value(response.answers).expect("the answers must serialize"),
+            answers
+        );
+    }
+}
diff --git a/resources/icons/copy.svg b/resources/icons/copy.svg
deleted file mode 100644
index 513e589c7..000000000
--- a/resources/icons/copy.svg
+++ /dev/null
@@ -1 +0,0 @@
-<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 -960 960 960"><path d="M360-240q-33 0-56.5-23.5T280-320v-480q0-33 23.5-56.5T360-880h360q33 0 56.5 23.5T800-800v480q0 33-23.5 56.5T720-240H360Zm0-80h360v-480H360v480ZM200-80q-33 0-56.5-23.5T120-160v-560h80v560h440v80H200Zm160-240v-480 480Z"/></svg>
diff --git a/resources/icons/open_in_new.svg b/resources/icons/open_in_new.svg
deleted file mode 100644
index 5f699494c..000000000
--- a/resources/icons/open_in_new.svg
+++ /dev/null
@@ -1 +0,0 @@
-<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 -960 960 960"><path d="M200-120q-33 0-56.5-23.5T120-200v-560q0-33 23.5-56.5T200-840h280v80H200v560h560v-280h80v280q0 33-23.5 56.5T760-120H200Zm188-212-56-56 372-372H560v-80h280v280h-80v-144L388-332Z"/></svg>
diff --git a/resources/icons/stop.svg b/resources/icons/stop.svg
deleted file mode 100644
index afaf0f264..000000000
--- a/resources/icons/stop.svg
+++ /dev/null
@@ -1 +0,0 @@
-<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 -960 960 960" fill="white"><path d="M324-111.5Q251-143 197-197t-85.5-127Q80-397 80-480t31.5-156Q143-709 197-763t127-85.5Q397-880 480-880t156 31.5Q709-817 763-763t85.5 127Q880-563 880-480t-31.5 156Q817-251 763-197t-127 85.5Q563-80 480-80t-156-31.5ZM480-160q54 0 104-17.5t92-50.5L228-676q-33 42-50.5 92T160-480q0 134 93 227t227 93Zm252-124q33-42 50.5-92T800-480q0-134-93-227t-227-93q-54 0-104 17.5T284-732l448 448ZM480-480Z"/></svg>
diff --git a/resources/images/beta.png b/resources/images/beta.png
deleted file mode 100644
index 4e2904cbc..000000000
Binary files a/resources/images/beta.png and /dev/null differ
diff --git a/resources/images/create_a_cluster.png b/resources/images/create_a_cluster.png
deleted file mode 100644
index 03a28a53c..000000000
Binary files a/resources/images/create_a_cluster.png and /dev/null differ
diff --git a/resources/images/join_a_cluster.png b/resources/images/join_a_cluster.png
deleted file mode 100644
index 16b1ace86..000000000
Binary files a/resources/images/join_a_cluster.png and /dev/null differ
diff --git a/resources/ts/components/AgentIssues.tsx b/resources/ts/components/AgentIssues.tsx
index e214f6943..77a1b306d 100644
--- a/resources/ts/components/AgentIssues.tsx
+++ b/resources/ts/components/AgentIssues.tsx
@@ -59,6 +59,42 @@ export function AgentIssues({ issues }: { issues: Array<AgentIssue> }) {
           );
         }
 
+        if ("HuggingFaceDownloadFailed" in issue) {
+          const { failure, model_path } = issue.HuggingFaceDownloadFailed;
+
+          return (
+            <li className={agentIssues__issue} key={index}>
+              <strong>
+                HuggingFace download failed: {model_path.model_path}
+              </strong>
+              <strong>Reason:</strong>{" "}
+              <p>
+                {failure === "CacheIoFailed" &&
+                  "The HuggingFace cache could not be written or read."}
+                {failure === "HubConnectionFailed" &&
+                  "The connection to HuggingFace failed or was interrupted."}
+                {failure === "HubResponseMalformed" &&
+                  "HuggingFace answered without the data needed to download the file."}
+                {failure === "UnexpectedHubClientFailure" &&
+                  "The HuggingFace client failed in a way Paddler does not expect."}
+                {failure === "XetTransferFailed" &&
+                  "The Xet transfer of the file failed."}
+                {typeof failure === "object" &&
+                  `HuggingFace responded with HTTP status ${failure.HubRespondedWithStatus.status}.`}
+              </p>
+              <strong>What will Paddler do?</strong>{" "}
+              <p>
+                Paddler will reattempt to download the model every few seconds.
+              </p>
+              <strong>What can you do?</strong>{" "}
+              <p>
+                Check the network connection of the agent and the space left in
+                its HuggingFace cache directory.
+              </p>
+            </li>
+          );
+        }
+
         if ("HuggingFaceModelDoesNotExist" in issue) {
           return (
             <li className={agentIssues__issue} key={index}>
@@ -159,6 +195,28 @@ export function AgentIssues({ issues }: { issues: Array<AgentIssue> }) {
           );
         }
 
+        if ("LocalModelFileCannotBeChecked" in issue) {
+          return (
+            <li className={agentIssues__issue} key={index}>
+              <strong>
+                Local model file cannot be checked:{" "}
+                {issue.LocalModelFileCannotBeChecked.model_path}
+              </strong>
+              <strong>What will Paddler do?</strong>{" "}
+              <p>
+                Paddler will keep checking the file every few seconds until the
+                agent can tell whether it exists.
+              </p>
+              <strong>What can you do?</strong>{" "}
+              <p>
+                Make sure the agent can read every directory on the path, or{" "}
+                <Link href="/model">change the model parameters</Link> to use a
+                file the agent can reach.
+              </p>
+            </li>
+          );
+        }
+
         if ("MultimodalProjectionCannotBeLoaded" in issue) {
           return (
             <li className={agentIssues__issue} key={index}>
@@ -182,12 +240,14 @@ export function AgentIssues({ issues }: { issues: Array<AgentIssue> }) {
         }
 
         if ("SlotCannotStart" in issue) {
-          const { error, slot_index } = issue.SlotCannotStart;
+          const { error, inference_capability, slot_index } =
+            issue.SlotCannotStart;
 
           return (
             <li className={agentIssues__issue} key={index}>
               <strong>
-                Unable to start slot {slot_index}: {error}
+                Unable to start slot {slot_index} for {inference_capability}:{" "}
+                {error}
               </strong>
             </li>
           );
@@ -224,6 +284,95 @@ export function AgentIssues({ issues }: { issues: Array<AgentIssue> }) {
           );
         }
 
+        if ("AgentModelUriIsNotAnAbsoluteLocalPath" in issue) {
+          return (
+            <li className={agentIssues__issue} key={index}>
+              <strong>
+                Agent URI does not name an absolute local path:{" "}
+                {issue.AgentModelUriIsNotAnAbsoluteLocalPath.model_path}
+              </strong>
+              <strong>What will Paddler do?</strong>{" "}
+              <p>
+                Paddler will not load the model until the URI names an absolute
+                path on the agent.
+              </p>
+              <strong>What can you do?</strong>{" "}
+              <p>
+                <Link href="/model">
+                  Edit the model URI on the model configuration page
+                </Link>{" "}
+                to the form agent:///absolute/path/to/model.gguf.
+              </p>
+            </li>
+          );
+        }
+
+        if ("AgentModelUriIsNotCanonical" in issue) {
+          return (
+            <li className={agentIssues__issue} key={index}>
+              <strong>
+                Agent URI is not written in its canonical form:{" "}
+                {issue.AgentModelUriIsNotCanonical.model_path.model_path}
+              </strong>
+              <strong>What will Paddler do?</strong>{" "}
+              <p>
+                Paddler will not load the model until the URI is written in its
+                one canonical form, so that it names exactly one file.
+              </p>
+              <strong>What can you do?</strong>{" "}
+              <p>
+                <Link href="/model">
+                  Edit the model URI on the model configuration page
+                </Link>{" "}
+                to {issue.AgentModelUriIsNotCanonical.canonical_uri}
+              </p>
+            </li>
+          );
+        }
+
+        if ("HuggingFaceModelUriIsMalformed" in issue) {
+          return (
+            <li className={agentIssues__issue} key={index}>
+              <strong>
+                Hugging Face model URI does not point at a file:{" "}
+                {issue.HuggingFaceModelUriIsMalformed.model_path}
+              </strong>
+              <strong>What will Paddler do?</strong>{" "}
+              <p>
+                Paddler will not load the model until the URI points at a file.
+              </p>
+              <strong>What can you do?</strong>{" "}
+              <p>
+                <Link href="/model">
+                  Edit the model URI on the model configuration page
+                </Link>{" "}
+                to the form
+                https://huggingface.co/owner/repository/blob/revision/file.gguf.
+              </p>
+            </li>
+          );
+        }
+
+        if ("ModelUriIsUnparseable" in issue) {
+          return (
+            <li className={agentIssues__issue} key={index}>
+              <strong>
+                Model URI cannot be parsed:{" "}
+                {issue.ModelUriIsUnparseable.model_path}
+              </strong>
+              <strong>What will Paddler do?</strong>{" "}
+              <p>Paddler will not load the model until the URI is valid.</p>
+              <strong>What can you do?</strong>{" "}
+              <p>
+                <Link href="/model">
+                  Edit the model URI on the model configuration page
+                </Link>{" "}
+                to a Hugging Face, agent:// or http(s) URI.
+              </p>
+            </li>
+          );
+        }
+
         if ("DownloadUrlIsMalformed" in issue) {
           return (
             <li className={agentIssues__issue} key={index}>
@@ -417,9 +566,10 @@ export function AgentIssues({ issues }: { issues: Array<AgentIssue> }) {
               <strong>What can you do?</strong>{" "}
               <p>
                 Grant write permission to the cache directory (
-                <code>$XDG_CACHE_HOME/paddler</code> on Linux/macOS,{" "}
-                <code>%LOCALAPPDATA%\paddler</code> on Windows), or set{" "}
-                <code>PADDLER_CACHE_DIR</code> to a writable location.
+                <code>$XDG_CACHE_HOME/paddler</code>, or{" "}
+                <code>~/.cache/paddler</code> when <code>XDG_CACHE_HOME</code>{" "}
+                is unset), or set <code>PADDLER_CACHE_DIR</code> to a writable
+                location.
               </p>
             </li>
           );
@@ -465,6 +615,63 @@ export function AgentIssues({ issues }: { issues: Array<AgentIssue> }) {
           );
         }
 
+        if ("PointerHeadCannotBeLoaded" in issue) {
+          return (
+            <li className={agentIssues__issue} key={index}>
+              <strong>
+                Pointer head cannot be loaded:{" "}
+                {issue.PointerHeadCannotBeLoaded.model_path}
+              </strong>
+              <strong>What can you do?</strong>{" "}
+              <p>
+                Either ensure that the file is a pointer head available to the
+                agent, or <Link href="/model">change the pointer head</Link>.
+              </p>
+            </li>
+          );
+        }
+
+        if ("PointerHeadIncompatibleWithModel" in issue) {
+          const { incompatibility, pointer_head_path } =
+            issue.PointerHeadIncompatibleWithModel;
+
+          return (
+            <li className={agentIssues__issue} key={index}>
+              <strong>
+                Pointer head {pointer_head_path.model_path} does not fit the
+                model:{" "}
+                {"HiddenSizeMismatch" in incompatibility
+                  ? `it reads ${incompatibility.HiddenSizeMismatch.pointer_head_hidden_size}-dimensional hidden states, the model produces ${incompatibility.HiddenSizeMismatch.model_hidden_size}`
+                  : `its delimiter ${incompatibility.DelimiterIsNotAControlToken.delimiter} is not a control token of the model`}
+              </strong>
+              <strong>What can you do?</strong>{" "}
+              <p>
+                <Link href="/model">Use the pointer head</Link> that was
+                converted together with the model.
+              </p>
+            </li>
+          );
+        }
+
+        if ("SlotsInsufficientForDecisions" in issue) {
+          const { desired_slots, required_slots } =
+            issue.SlotsInsufficientForDecisions;
+
+          return (
+            <li className={agentIssues__issue} key={index}>
+              <strong>
+                Decisions need at least {required_slots} slots, the agent runs{" "}
+                {desired_slots}
+              </strong>
+              <strong>What can you do?</strong>{" "}
+              <p>
+                Restart the agent with <code>--slots {required_slots}</code> or
+                more.
+              </p>
+            </li>
+          );
+        }
+
         return (
           <li className={agentIssues__issue} key={index}>
             Unknown issue: {JSON.stringify(issue)}
diff --git a/resources/ts/components/AgentListAgentStatus.tsx b/resources/ts/components/AgentListAgentStatus.tsx
index d26fce510..857053d1f 100644
--- a/resources/ts/components/AgentListAgentStatus.tsx
+++ b/resources/ts/components/AgentListAgentStatus.tsx
@@ -7,11 +7,13 @@ import { agentListAgentStatus__progress } from "./AgentListAgentStatus.module.cs
 export function AgentListAgentStatus({
   agent: {
     slots_processing,
-    status: { desired_slots_total, slots_total, state_application_status },
+    status: { desired_slots_total, runtime, state_application_status },
   },
 }: {
   agent: Agent;
 }) {
+  const slots_total = runtime === "Idle" ? 0 : runtime.Serving.slots_total;
+
   switch (state_application_status) {
     case "Applied":
       return (
@@ -33,14 +35,6 @@ export function AgentListAgentStatus({
           </abbr>
         </div>
       );
-    case "AttemptedAndNotAppliable":
-      return (
-        <div className={agentListAgentStatus__progress}>
-          <div>
-            🏳️ <i>Needs your help</i>
-          </div>
-        </div>
-      );
     case "AttemptedAndRetrying":
       return (
         <div className={agentListAgentStatus__progress}>
diff --git a/resources/ts/components/BalancerDesiredStateContextProvider.tsx b/resources/ts/components/BalancerDesiredStateContextProvider.tsx
new file mode 100644
index 000000000..11fc55c09
--- /dev/null
+++ b/resources/ts/components/BalancerDesiredStateContextProvider.tsx
@@ -0,0 +1,34 @@
+import React, { useMemo, useState, type ReactNode } from "react";
+
+import { type BalancerDesiredState } from "@intentee/paddler-client/schemas/BalancerDesiredState";
+import {
+  BalancerDesiredStateContext,
+  type BalancerDesiredStateContextValue,
+} from "../contexts/BalancerDesiredStateContext";
+
+export function BalancerDesiredStateContextProvider({
+  children,
+  defaultBalancerDesiredState,
+}: {
+  children: ReactNode;
+  defaultBalancerDesiredState: BalancerDesiredState;
+}) {
+  const [balancerDesiredState, setBalancerDesiredState] =
+    useState<BalancerDesiredState>(defaultBalancerDesiredState);
+
+  const value = useMemo<BalancerDesiredStateContextValue>(
+    function () {
+      return Object.freeze({
+        balancerDesiredState,
+        setBalancerDesiredState,
+      });
+    },
+    [balancerDesiredState, setBalancerDesiredState],
+  );
+
+  return (
+    <BalancerDesiredStateContext.Provider value={value}>
+      {children}
+    </BalancerDesiredStateContext.Provider>
+  );
+}
diff --git a/resources/ts/components/ChangeModelForm.tsx b/resources/ts/components/ChangeModelForm.tsx
index 162144a70..3818180e8 100644
--- a/resources/ts/components/ChangeModelForm.tsx
+++ b/resources/ts/components/ChangeModelForm.tsx
@@ -1,22 +1,14 @@
-import React, {
-  useCallback,
-  useContext,
-  useMemo,
-  type FormEvent,
-  type InputEvent,
-} from "react";
+import React, { useCallback, useContext, type SubmitEvent } from "react";
 import { useLocation } from "wouter";
 
-import { type BalancerDesiredState } from "@intentee/paddler-client/schemas/BalancerDesiredState";
-import { ChatTemplateContext } from "../contexts/ChatTemplateContext";
-import { InferenceParametersContext } from "../contexts/InferenceParametersContext";
+import { type AgentDesiredModel } from "@intentee/paddler-client/schemas/AgentDesiredModel";
+import { BalancerDesiredStateContext } from "../contexts/BalancerDesiredStateContext";
 import { PaddlerConfigurationContext } from "../contexts/PaddlerConfigurationContext";
-import { useAgentDesiredModelUrl } from "../hooks/useAgentDesiredModelUrl";
 import { ChatTemplateBehavior } from "./ChatTemplateBehavior";
-import { InferenceParameterCacheDtype } from "./InferenceParameterCacheDtype";
-import { InferenceParameterCheckbox } from "./InferenceParameterCheckbox";
-import { InferenceParameterInput } from "./InferenceParameterInput";
-import { InferenceParameterPoolingType } from "./InferenceParameterPoolingType";
+import { EmbeddingParametersFields } from "./EmbeddingParametersFields";
+import { ModelRuntimeParametersFields } from "./ModelRuntimeParametersFields";
+import { ModelUriInput } from "./ModelUriInput";
+import { TextGenerationSettingsFields } from "./TextGenerationSettingsFields";
 
 import {
   changeModelForm,
@@ -27,96 +19,62 @@ import {
   changeModelForm__formControls,
   changeModelForm__formLabel,
   changeModelForm__formLabel__title,
-  changeModelForm__input,
   changeModelForm__main,
   changeModelForm__parameters,
   changeModelForm__submitButton,
 } from "./ChangeModelForm.module.css";
 
-export function ChangeModelForm({
-  defaultBaseModelUri,
-  defaultMultimodalProjectionUri,
-}: {
-  defaultBaseModelUri: null | string;
-  defaultMultimodalProjectionUri: null | string;
-}) {
+export function ChangeModelForm() {
   const [, navigate] = useLocation();
-  const { chatTemplateOverride, useChatTemplateOverride } =
-    useContext(ChatTemplateContext);
-  const { parameters } = useContext(InferenceParametersContext);
-  const { managementAddr } = useContext(PaddlerConfigurationContext);
-  const {
-    agentDesiredModelState: baseModelAgentDesiredModelState,
-    modelUri: baseModelUri,
-    setModelUri: setBaseModelUri,
-  } = useAgentDesiredModelUrl({
-    defaultModelUri: defaultBaseModelUri,
-  });
   const {
-    agentDesiredModelState: multimodalProjecttionAgentDesiredModelState,
-    modelUri: multimodalProjectionModelUri,
-    setModelUri: setMultimodalProjectionModelUri,
-  } = useAgentDesiredModelUrl({
-    defaultModelUri: defaultMultimodalProjectionUri,
-  });
+    balancerDesiredState: editedDesiredState,
+    setBalancerDesiredState: setEditedDesiredState,
+  } = useContext(BalancerDesiredStateContext);
+  const { managementAddr } = useContext(PaddlerConfigurationContext);
 
-  const onBaseModelUriInput = useCallback(
-    function (event: InputEvent<HTMLInputElement>) {
-      setBaseModelUri(event.currentTarget.value);
+  const onBaseModel = useCallback(
+    function (model: AgentDesiredModel) {
+      setEditedDesiredState({ ...editedDesiredState, model });
     },
-    [setBaseModelUri],
+    [editedDesiredState, setEditedDesiredState],
   );
 
-  const onMultimodalProjectionUriInput = useCallback(
-    function (event: InputEvent<HTMLInputElement>) {
-      setMultimodalProjectionModelUri(event.currentTarget.value);
+  const onMultimodalProjection = useCallback(
+    function (projection: AgentDesiredModel) {
+      setEditedDesiredState({
+        ...editedDesiredState,
+        text_generation: {
+          ...editedDesiredState.text_generation,
+          multimodal: {
+            ...editedDesiredState.text_generation.multimodal,
+            projection,
+          },
+        },
+      });
     },
-    [setMultimodalProjectionModelUri],
+    [editedDesiredState, setEditedDesiredState],
   );
 
-  const balancerDesiredState: null | BalancerDesiredState = useMemo(
-    function () {
-      if (
-        !baseModelAgentDesiredModelState.ok ||
-        !multimodalProjecttionAgentDesiredModelState.ok
-      ) {
-        return null;
-      }
-
-      const desiredState: BalancerDesiredState = Object.freeze({
-        chat_template_override: chatTemplateOverride,
-        inference_parameters: parameters,
-        model: baseModelAgentDesiredModelState.agentDesiredModel,
-        multimodal_projection:
-          multimodalProjecttionAgentDesiredModelState.agentDesiredModel,
-        use_chat_template_override: useChatTemplateOverride,
+  const onPointerHead = useCallback(
+    function (pointer_head: AgentDesiredModel) {
+      setEditedDesiredState({
+        ...editedDesiredState,
+        decision: { ...editedDesiredState.decision, pointer_head },
       });
-
-      return desiredState;
     },
-    [
-      baseModelAgentDesiredModelState,
-      chatTemplateOverride,
-      multimodalProjecttionAgentDesiredModelState,
-      parameters,
-      useChatTemplateOverride,
-    ],
+    [editedDesiredState, setEditedDesiredState],
   );
 
   const onSubmit = useCallback(
-    function (event: FormEvent<HTMLFormElement>) {
+    function (event: SubmitEvent<HTMLFormElement>) {
       event.preventDefault();
 
-      if (!balancerDesiredState) {
-        return;
-      }
-
       fetch(`//${managementAddr}/api/v1/balancer_desired_state`, {
         method: "PUT",
         headers: {
           "Content-Type": "application/json",
         },
-        body: JSON.stringify(balancerDesiredState),
+        body: JSON.stringify(editedDesiredState),
       })
         .then(function (response) {
           if (response.ok) {
@@ -131,7 +89,7 @@ export function ChangeModelForm({
           console.error("Error updating agent desired state:", error);
         });
     },
-    [managementAddr, navigate, balancerDesiredState],
+    [editedDesiredState, managementAddr, navigate],
   );
 
   return (
@@ -171,7 +129,12 @@ export function ChangeModelForm({
           </dd>
           <dt>Local File</dt>
           <dd>
-            <p>File path is relative to the agent's working directory.</p>
+            <p>
+              The file path must be absolute and spelled canonically: percent
+              encode spaces and special characters, and leave out relative
+              segments, repeated slashes, queries, and fragments. An agent
+              reports the canonical spelling when the one you enter differs.
+            </p>
             <p>
               If you want all the agents to use the same model, you need to
               ensure that the file is present in the same path on all agents.
@@ -186,27 +149,33 @@ export function ChangeModelForm({
             <div className={changeModelForm__formLabel__title}>
               Base Model URI
             </div>
-            <input
-              className={changeModelForm__input}
+            <ModelUriInput
+              model={editedDesiredState.model}
               name="model_uri"
-              onInput={onBaseModelUriInput}
-              placeholder="https://huggingface.co/..."
+              onModel={onBaseModel}
               required
-              type="url"
-              value={String(baseModelUri)}
+            />
+          </label>
+          <label className={changeModelForm__formLabel}>
+            <div className={changeModelForm__formLabel__title}>
+              Pointer Head URI (optional, set it to serve decisions)
+            </div>
+            <ModelUriInput
+              model={editedDesiredState.decision.pointer_head}
+              name="pointer_head_uri"
+              onModel={onPointerHead}
+              required={false}
             />
           </label>
           <label className={changeModelForm__formLabel}>
             <div className={changeModelForm__formLabel__title}>
               Multimodal Projection URI (optional)
             </div>
-            <input
-              className={changeModelForm__input}
+            <ModelUriInput
+              model={editedDesiredState.text_generation.multimodal.projection}
               name="multimodal_projection_uri"
-              onInput={onMultimodalProjectionUriInput}
-              placeholder="https://huggingface.co/..."
-              type="url"
-              value={String(multimodalProjectionModelUri)}
+              onModel={onMultimodalProjection}
+              required={false}
             />
           </label>
           <fieldset className={changeModelForm__chatTemplate}>
@@ -214,92 +183,29 @@ export function ChangeModelForm({
             <ChatTemplateBehavior />
           </fieldset>
           <fieldset className={changeModelForm__parameters}>
-            <legend>Inference Parameters</legend>
+            <legend>Model Parameters</legend>
             <details className={changeModelForm__details}>
               <summary>What are these parameters?</summary>
               <p>
-                These parameters control how the model behaves during inference.
-                They can affect the quality, speed, and memory usage of the
-                model.
+                These parameters control how the model is loaded and how much
+                memory it uses. Each slot is one sequence that holds an equal
+                share of the context, whichever capability uses it.
               </p>
               <p>
                 They are usually model-specific and are usually provided by the
                 model authors, although Paddler provides some reasonable
                 defaults.
               </p>
-              <p>
-                Experimenting with these settings is worth exploring to optimize
-                performance for your specific needs.
-              </p>
             </details>
-            <InferenceParameterInput
-              description="Batch Size (higher = more memory usage, lower = less inference speed)"
-              name="n_batch"
-            />
-            <InferenceParameterInput
-              description="Context Size (higher = longer chat history, lower = less memory usage)"
-              name="context_size"
-            />
-            <InferenceParameterInput
-              description="Max image dimension in pixels before resizing"
-              name="image_resize_to_fit"
-            />
-            <InferenceParameterInput
-              description="Minimum token probability to consider for selection"
-              name="min_p"
-            />
-            <InferenceParameterInput
-              description="Number of model layers to offload to GPU (0 = CPU only; -1 = all layers)"
-              name="n_gpu_layers"
-            />
-            <InferenceParameterInput
-              description="Frequency Penalty"
-              name="penalty_frequency"
-            />
-            <InferenceParameterInput
-              description="Number of last tokens to consider for penalty (0 = disabled)"
-              name="penalty_last_n"
-            />
-            <InferenceParameterInput
-              description="Presence Penalty"
-              name="penalty_presence"
-            />
-            <InferenceParameterInput
-              description="Repeated Token Penalty"
-              name="penalty_repeat"
-            />
-            <InferenceParameterInput
-              description="Temperature"
-              name="temperature"
-            />
-            <InferenceParameterInput
-              description="Number of tokens to consider for selection"
-              name="top_k"
-            />
-            <InferenceParameterInput
-              description="Probability threshold for selecting tokens"
-              name="top_p"
-            />
-            <InferenceParameterCheckbox
-              description="You need embeddings for stuff like semantic search, RAG, and more"
-              name="enable_embeddings"
-            />
-            <InferenceParameterInput
-              description="Max documents per embedding sub-batch (cap; balancer fans out larger requests across agents and chunks beyond this size)"
-              name="embedding_batch_size"
-            />
-            <InferenceParameterPoolingType
-              description="How to combine token embeddings"
-              disabled={!parameters.enable_embeddings}
-            />
-            <InferenceParameterCacheDtype
-              name="k_cache_dtype"
-              description="Datatype for K cache tensors"
-            />
-            <InferenceParameterCacheDtype
-              name="v_cache_dtype"
-              description="Datatype for V cache tensors"
-            />
+            <ModelRuntimeParametersFields />
+          </fieldset>
+          <fieldset className={changeModelForm__parameters}>
+            <legend>Text Generation Parameters</legend>
+            <TextGenerationSettingsFields />
+          </fieldset>
+          <fieldset className={changeModelForm__parameters}>
+            <legend>Embedding Parameters</legend>
+            <EmbeddingParametersFields />
           </fieldset>
           <div className={changeModelForm__formControls}>
             <button className={changeModelForm__submitButton}>
diff --git a/resources/ts/components/ChangeModelPage.tsx b/resources/ts/components/ChangeModelPage.tsx
index 57ef4e13d..d33c22194 100644
--- a/resources/ts/components/ChangeModelPage.tsx
+++ b/resources/ts/components/ChangeModelPage.tsx
@@ -1,35 +1,12 @@
 import React, { useContext } from "react";
 
-import { type AgentDesiredModel } from "@intentee/paddler-client/schemas/AgentDesiredModel";
 import { PaddlerConfigurationContext } from "../contexts/PaddlerConfigurationContext";
 import { useBalancerDesiredState } from "../hooks/useBalancerDesiredState";
 import { matchFetchJsonState } from "../matchFetchJsonState";
+import { BalancerDesiredStateContextProvider } from "./BalancerDesiredStateContextProvider";
 import { ChangeModelForm } from "./ChangeModelForm";
 import { ChatTemplateContextProvider } from "./ChatTemplateContextProvider";
 import { FloatingStatus } from "./FloatingStatus";
-import { InferenceParametersContextProvider } from "./InferenceParametersContextProvider";
-
-function modelSchemaToUrl(model: AgentDesiredModel): string {
-  if (model === "None") {
-    return "";
-  }
-
-  if ("HuggingFace" in model) {
-    const { HuggingFace } = model;
-
-    return `https://huggingface.co/${HuggingFace.repo_id}/blob/${HuggingFace.revision}/${HuggingFace.filename}`;
-  }
-
-  if ("LocalToAgent" in model) {
-    return `agent://${model.LocalToAgent}`;
-  }
-
-  if ("Url" in model) {
-    return model.Url.url;
-  }
-
-  throw new Error(`Unsupported model schema: ${JSON.stringify(model)}`);
-}
 
 export function ChangeModelPage() {
   const { managementAddr } = useContext(PaddlerConfigurationContext);
@@ -49,31 +26,15 @@ export function ChangeModelPage() {
     loading() {
       return <FloatingStatus>Loading desired state...</FloatingStatus>;
     },
-    ok({
-      response: {
-        chat_template_override,
-        inference_parameters,
-        model,
-        multimodal_projection,
-        use_chat_template_override,
-      },
-    }) {
+    ok({ response: balancerDesiredState }) {
       return (
-        <ChatTemplateContextProvider
-          defaultChatTemplateOverride={chat_template_override}
-          defaultUseChatTemplateOverride={use_chat_template_override}
+        <BalancerDesiredStateContextProvider
+          defaultBalancerDesiredState={balancerDesiredState}
         >
-          <InferenceParametersContextProvider
-            defaultInferenceParameters={inference_parameters}
-          >
-            <ChangeModelForm
-              defaultBaseModelUri={modelSchemaToUrl(model)}
-              defaultMultimodalProjectionUri={modelSchemaToUrl(
-                multimodal_projection,
-              )}
-            />
-          </InferenceParametersContextProvider>
-        </ChatTemplateContextProvider>
+          <ChatTemplateContextProvider>
+            <ChangeModelForm />
+          </ChatTemplateContextProvider>
+        </BalancerDesiredStateContextProvider>
       );
     },
   });
diff --git a/resources/ts/components/ChatTemplateContextProvider.tsx b/resources/ts/components/ChatTemplateContextProvider.tsx
index 9a258c8f4..dfa491545 100644
--- a/resources/ts/components/ChatTemplateContextProvider.tsx
+++ b/resources/ts/components/ChatTemplateContextProvider.tsx
@@ -1,6 +1,7 @@
-import React, { useCallback, useMemo, useState, type ReactNode } from "react";
+import React, { useContext, useMemo, useState, type ReactNode } from "react";
 
-import { type ChatTemplate } from "@intentee/paddler-client/schemas/ChatTemplate";
+import { type ChatTemplateSource } from "@intentee/paddler-client/schemas/ChatTemplateSource";
+import { BalancerDesiredStateContext } from "../contexts/BalancerDesiredStateContext";
 import {
   ChatTemplateContext,
   type ChatTemplateContextValue,
@@ -8,45 +9,63 @@ import {
 
 export function ChatTemplateContextProvider({
   children,
-  defaultChatTemplateOverride,
-  defaultUseChatTemplateOverride,
 }: {
   children: ReactNode;
-  defaultChatTemplateOverride: null | ChatTemplate;
-  defaultUseChatTemplateOverride: boolean;
 }) {
-  const [chatTemplateOverride, setChatTemplateOverride] =
-    useState<null | ChatTemplate>(defaultChatTemplateOverride);
-  const [useChatTemplateOverride, setUseChatTemplateOverride] =
-    useState<boolean>(defaultUseChatTemplateOverride);
-
-  const setChatTemplateOverrideContent = useCallback(
-    function (content: string) {
-      setChatTemplateOverride(function () {
-        return {
-          content,
-        };
-      });
-    },
-    [setChatTemplateOverride],
+  const { balancerDesiredState, setBalancerDesiredState } = useContext(
+    BalancerDesiredStateContext,
   );
+  const [
+    overrideContentWhileUsingTheModelTemplate,
+    setOverrideContentWhileUsingTheModelTemplate,
+  ] = useState("");
 
   const value = useMemo<ChatTemplateContextValue>(
     function () {
+      const { text_generation: textGenerationSettings } = balancerDesiredState;
+      const { chat_template_source: chatTemplateSource } =
+        textGenerationSettings;
+      const chatTemplateOverrideContent =
+        chatTemplateSource === "EmbeddedInModel"
+          ? overrideContentWhileUsingTheModelTemplate
+          : chatTemplateSource.Override.content;
+
+      function setChatTemplateSource(
+        changedChatTemplateSource: ChatTemplateSource,
+      ) {
+        setBalancerDesiredState({
+          ...balancerDesiredState,
+          text_generation: {
+            ...textGenerationSettings,
+            chat_template_source: changedChatTemplateSource,
+          },
+        });
+      }
+
       return Object.freeze({
-        chatTemplateOverride,
-        setChatTemplateOverride,
-        setChatTemplateOverrideContent,
-        setUseChatTemplateOverride,
-        useChatTemplateOverride,
+        chatTemplateOverrideContent,
+        setChatTemplateOverrideContent(content: string) {
+          setChatTemplateSource({ Override: { content } });
+        },
+        setUseChatTemplateOverride(useChatTemplateOverride: boolean) {
+          if (useChatTemplateOverride) {
+            setChatTemplateSource({
+              Override: { content: chatTemplateOverrideContent },
+            });
+          } else {
+            setOverrideContentWhileUsingTheModelTemplate(
+              chatTemplateOverrideContent,
+            );
+            setChatTemplateSource("EmbeddedInModel");
+          }
+        },
+        useChatTemplateOverride: chatTemplateSource !== "EmbeddedInModel",
       });
     },
     [
-      chatTemplateOverride,
-      setChatTemplateOverride,
-      setChatTemplateOverrideContent,
-      setUseChatTemplateOverride,
-      useChatTemplateOverride,
+      balancerDesiredState,
+      overrideContentWhileUsingTheModelTemplate,
+      setBalancerDesiredState,
     ],
   );
 
diff --git a/resources/ts/components/ChatTemplateEditButton.tsx b/resources/ts/components/ChatTemplateEditButton.tsx
index fdfa7f0fc..db6d2166b 100644
--- a/resources/ts/components/ChatTemplateEditButton.tsx
+++ b/resources/ts/components/ChatTemplateEditButton.tsx
@@ -1,13 +1,11 @@
 import React, {
   useCallback,
   useContext,
-  useEffect,
   useMemo,
   useState,
   type MouseEvent,
 } from "react";
 
-import { type ChatTemplate } from "@intentee/paddler-client/schemas/ChatTemplate";
 import { ChatTemplateContext } from "../contexts/ChatTemplateContext";
 import { CodeEditor } from "./CodeEditor";
 import { ModalWindow } from "./ModalWindow";
@@ -21,28 +19,25 @@ import {
   chatTemplateEditButton__modal__toolbarButton,
 } from "./ChatTemplateEditButton.module.css";
 
-function toDefaultContent(chatTemplate: null | ChatTemplate): string {
-  return chatTemplate?.content || "";
-}
-
 export function ChatTemplateEditButton() {
   const [isEditing, setIsEditing] = useState(false);
   const {
-    chatTemplateOverride,
+    chatTemplateOverrideContent,
     setChatTemplateOverrideContent,
     useChatTemplateOverride,
   } = useContext(ChatTemplateContext);
   const [temporaryContent, setTemporaryContent] = useState(
-    toDefaultContent(chatTemplateOverride),
+    chatTemplateOverrideContent,
   );
-
-  useEffect(
-    function () {
-      setTemporaryContent(toDefaultContent(chatTemplateOverride));
-    },
-    [chatTemplateOverride, setTemporaryContent],
+  const [editedContentSource, setEditedContentSource] = useState(
+    chatTemplateOverrideContent,
   );
 
+  if (editedContentSource !== chatTemplateOverrideContent) {
+    setEditedContentSource(chatTemplateOverrideContent);
+    setTemporaryContent(chatTemplateOverrideContent);
+  }
+
   const onClick = useCallback(
     function (event: MouseEvent<HTMLButtonElement>) {
       event.preventDefault();
@@ -54,10 +49,10 @@ export function ChatTemplateEditButton() {
 
   const onClose = useCallback(
     function () {
-      setTemporaryContent(toDefaultContent(chatTemplateOverride));
+      setTemporaryContent(chatTemplateOverrideContent);
       setIsEditing(false);
     },
-    [chatTemplateOverride, setIsEditing, setTemporaryContent],
+    [chatTemplateOverrideContent, setIsEditing, setTemporaryContent],
   );
 
   const onSaveClick = useCallback(
@@ -72,9 +67,9 @@ export function ChatTemplateEditButton() {
 
   const shouldConfirmClose = useMemo(
     function () {
-      return toDefaultContent(chatTemplateOverride) !== temporaryContent;
+      return chatTemplateOverrideContent !== temporaryContent;
     },
-    [chatTemplateOverride, temporaryContent],
+    [chatTemplateOverrideContent, temporaryContent],
   );
 
   return (
diff --git a/resources/ts/components/ConversationMessagePromptGeneratedTokens.tsx b/resources/ts/components/ConversationMessagePromptGeneratedTokens.tsx
index abe288a50..5cdb7129f 100644
--- a/resources/ts/components/ConversationMessagePromptGeneratedTokens.tsx
+++ b/resources/ts/components/ConversationMessagePromptGeneratedTokens.tsx
@@ -34,23 +34,27 @@ function buildUserMessage(
 ): ConversationMessageType {
   if (submittedImageDataUri) {
     return {
-      role: "user",
-      content: [
-        {
-          type: "image_url",
-          image_url: { url: submittedImageDataUri },
-        },
-        {
-          type: "text",
-          text: submittedPrompt,
-        },
-      ],
+      Message: {
+        role: "user",
+        content: [
+          {
+            type: "image_url",
+            image_url: { url: submittedImageDataUri },
+          },
+          {
+            type: "text",
+            text: submittedPrompt,
+          },
+        ],
+      },
     };
   }
 
   return {
-    role: "user",
-    content: submittedPrompt,
+    Message: {
+      role: "user",
+      content: submittedPrompt,
+    },
   };
 }
 
@@ -83,13 +87,17 @@ export const ConversationMessagePromptGeneratedTokens = memo(
             enableThinking: submittedIsThinkingEnabled,
             messages: [
               {
-                role: "system",
-                content:
-                  "You are a helpful assistant. Give engaging, short, precise answers. Be friendly, supportive, use emojis.",
+                Message: {
+                  role: "system",
+                  content:
+                    "You are a helpful assistant. Give engaging, short, precise answers. Be friendly, supportive, use emojis.",
+                },
               },
               {
-                role: "assistant",
-                content: "Hello! How can I help you today?",
+                Message: {
+                  role: "assistant",
+                  content: "Hello! How can I help you today?",
+                },
               },
               buildUserMessage(submittedPrompt, submittedImageDataUri),
             ],
diff --git a/resources/ts/components/ConversationPromptInput.tsx b/resources/ts/components/ConversationPromptInput.tsx
index 08919b76a..ed4f5ccf5 100644
--- a/resources/ts/components/ConversationPromptInput.tsx
+++ b/resources/ts/components/ConversationPromptInput.tsx
@@ -1,4 +1,9 @@
-import React, { useCallback, useContext, type FormEvent } from "react";
+import React, {
+  useCallback,
+  useContext,
+  type InputEvent,
+  type SubmitEvent,
+} from "react";
 
 import {
   conversationPromptInput,
@@ -14,7 +19,6 @@ import iconArrowUpward from "../../icons/arrow_upward.svg";
 import { PromptContext } from "../contexts/PromptContext";
 import { PromptImageContext } from "../contexts/PromptImageContext";
 import { PromptThinkingContext } from "../contexts/PromptThinkingContext";
-import { TokenGenerationDisabledContext } from "../contexts/TokenGenerationDisabledContext";
 import { ConversationPromptInputImageButton } from "./ConversationPromptInputImageButton";
 import { ConversationPromptInputImagePreview } from "./ConversationPromptInputImagePreview";
 import { ConversationPromptInputThinkingToggle } from "./ConversationPromptInputThinkingToggle";
@@ -36,12 +40,9 @@ export function ConversationPromptInput() {
   const { isThinkingEnabled, setSubmittedIsThinkingEnabled } = useContext(
     PromptThinkingContext,
   );
-  const { isTokenGenerationDisabled } = useContext(
-    TokenGenerationDisabledContext,
-  );
 
   const onSubmit = useCallback(
-    function (event: FormEvent<HTMLFormElement>) {
+    function (event: SubmitEvent<HTMLFormElement>) {
       event.preventDefault();
 
       if (currentPrompt.trim() === "") {
@@ -66,7 +67,7 @@ export function ConversationPromptInput() {
   );
 
   const onTextareaInput = useCallback(
-    function (event: FormEvent<HTMLInputElement>) {
+    function (event: InputEvent<HTMLInputElement>) {
       setCurrentPrompt(event.currentTarget.value);
     },
     [setCurrentPrompt],
@@ -84,7 +85,6 @@ export function ConversationPromptInput() {
       <input
         autoFocus
         className={conversationPromptInput__textarea}
-        disabled={isTokenGenerationDisabled}
         placeholder="Type your prompt here..."
         value={currentPrompt}
         onInput={onTextareaInput}
@@ -97,7 +97,7 @@ export function ConversationPromptInput() {
           <ConversationPromptInputImageButton />
           <button
             className={conversationPromptInput__button}
-            disabled={isCurrentPromptEmpty || isTokenGenerationDisabled}
+            disabled={isCurrentPromptEmpty}
           >
             <img src={iconArrowUpward} alt="Send" />
           </button>
diff --git a/resources/ts/components/ConversationPromptInputImageButton.tsx b/resources/ts/components/ConversationPromptInputImageButton.tsx
index 501a5f2e1..5a86f15be 100644
--- a/resources/ts/components/ConversationPromptInputImageButton.tsx
+++ b/resources/ts/components/ConversationPromptInputImageButton.tsx
@@ -6,7 +6,6 @@ import React, {
 } from "react";
 
 import { PromptImageContext } from "../contexts/PromptImageContext";
-import { TokenGenerationDisabledContext } from "../contexts/TokenGenerationDisabledContext";
 
 import { conversationPromptInput__button } from "./ConversationPromptInput.module.css";
 import { conversationPromptInputImageButton__fileInput } from "./ConversationPromptInputImageButton.module.css";
@@ -35,9 +34,6 @@ function readFileAsDataUri(imageFile: File): Promise<string> {
 
 export function ConversationPromptInputImageButton() {
   const { setCurrentImageDataUri } = useContext(PromptImageContext);
-  const { isTokenGenerationDisabled } = useContext(
-    TokenGenerationDisabledContext,
-  );
   const fileInputRef = useRef<HTMLInputElement>(null);
 
   const onButtonClick = useCallback(function () {
@@ -64,14 +60,12 @@ export function ConversationPromptInputImageButton() {
       <input
         accept="image/*"
         className={conversationPromptInputImageButton__fileInput}
-        disabled={isTokenGenerationDisabled}
         ref={fileInputRef}
         type="file"
         onChange={onFileSelected}
       />
       <button
         className={conversationPromptInput__button}
-        disabled={isTokenGenerationDisabled}
         type="button"
         onClick={onButtonClick}
       >
diff --git a/resources/ts/components/ConversationPromptInputThinkingToggle.tsx b/resources/ts/components/ConversationPromptInputThinkingToggle.tsx
index a600761bc..df2cc8de0 100644
--- a/resources/ts/components/ConversationPromptInputThinkingToggle.tsx
+++ b/resources/ts/components/ConversationPromptInputThinkingToggle.tsx
@@ -2,7 +2,6 @@ import clsx from "clsx";
 import React, { useCallback, useContext } from "react";
 
 import { PromptThinkingContext } from "../contexts/PromptThinkingContext";
-import { TokenGenerationDisabledContext } from "../contexts/TokenGenerationDisabledContext";
 
 import { conversationPromptInput__button } from "./ConversationPromptInput.module.css";
 import { conversationPromptInputThinkingToggleActive } from "./ConversationPromptInputThinkingToggle.module.css";
@@ -14,9 +13,6 @@ export function ConversationPromptInputThinkingToggle() {
   const { isThinkingEnabled, setIsThinkingEnabled } = useContext(
     PromptThinkingContext,
   );
-  const { isTokenGenerationDisabled } = useContext(
-    TokenGenerationDisabledContext,
-  );
 
   const onToggle = useCallback(
     function () {
@@ -30,7 +26,6 @@ export function ConversationPromptInputThinkingToggle() {
       className={clsx(conversationPromptInput__button, {
         [conversationPromptInputThinkingToggleActive]: isThinkingEnabled,
       })}
-      disabled={isTokenGenerationDisabled}
       type="button"
       onClick={onToggle}
     >
diff --git a/resources/ts/components/DashboardPage.tsx b/resources/ts/components/DashboardPage.tsx
index 98f9f4ae8..85b99ee12 100644
--- a/resources/ts/components/DashboardPage.tsx
+++ b/resources/ts/components/DashboardPage.tsx
@@ -20,8 +20,13 @@ import {
 } from "./DashboardPage.module.css";
 
 export function DashboardPage() {
-  const { compatOpenAIAddr, inferenceAddr, managementAddr, statsd } =
-    useContext(PaddlerConfigurationContext);
+  const {
+    compatOpenAIAddr,
+    compatTypeSafeAddr,
+    inferenceAddr,
+    managementAddr,
+    statsd,
+  } = useContext(PaddlerConfigurationContext);
 
   return (
     <div className={dashboardPage}>
@@ -57,6 +62,17 @@ export function DashboardPage() {
                   <p>{compatOpenAIAddr}</p>
                 </div>
               )}
+              {compatTypeSafeAddr !== null && (
+                <div
+                  className={`${dashboardPage__genericAddr} ${dashboardPage__inferenceAddr} ${dashboardPage__compatibilityServiceAddr}`}
+                >
+                  <p>
+                    TypeSafe <abbr title="compatibility service">compat</abbr>{" "}
+                    addr:
+                  </p>
+                  <p>{compatTypeSafeAddr}</p>
+                </div>
+              )}
             </div>
             {statsd !== null && (
               <div
diff --git a/resources/ts/components/EmbeddingParametersFields.tsx b/resources/ts/components/EmbeddingParametersFields.tsx
new file mode 100644
index 000000000..1a09f6a9d
--- /dev/null
+++ b/resources/ts/components/EmbeddingParametersFields.tsx
@@ -0,0 +1,54 @@
+import React, { useCallback, useContext } from "react";
+
+import {
+  poolingTypes,
+  type EmbeddingParameters,
+} from "@intentee/paddler-client/schemas/EmbeddingParameters";
+import { BalancerDesiredStateContext } from "../contexts/BalancerDesiredStateContext";
+import { ParameterNumberInput } from "./ParameterNumberInput";
+import { ParameterSelect } from "./ParameterSelect";
+
+export function EmbeddingParametersFields() {
+  const { balancerDesiredState, setBalancerDesiredState } = useContext(
+    BalancerDesiredStateContext,
+  );
+  const { embeddings: embeddingParameters } = balancerDesiredState;
+
+  const setParameter = useCallback(
+    function <TKey extends keyof EmbeddingParameters>(
+      name: TKey,
+      value: EmbeddingParameters[TKey],
+    ) {
+      setBalancerDesiredState({
+        ...balancerDesiredState,
+        embeddings: {
+          ...embeddingParameters,
+          [name]: value,
+        },
+      });
+    },
+    [balancerDesiredState, embeddingParameters, setBalancerDesiredState],
+  );
+
+  return (
+    <>
+      <ParameterNumberInput
+        description="Max documents per embedding sub-batch (cap; balancer fans out larger requests across agents and chunks beyond this size)"
+        name="embedding_batch_size"
+        onValue={function (value) {
+          setParameter("embedding_batch_size", value);
+        }}
+        value={embeddingParameters.embedding_batch_size}
+      />
+      <ParameterSelect
+        description="How to combine token embeddings"
+        name="pooling_type"
+        onValue={function (value) {
+          setParameter("pooling_type", value);
+        }}
+        options={poolingTypes}
+        value={embeddingParameters.pooling_type}
+      />
+    </>
+  );
+}
diff --git a/resources/ts/components/ExplainButton.tsx b/resources/ts/components/ExplainButton.tsx
deleted file mode 100644
index 869ddda4c..000000000
--- a/resources/ts/components/ExplainButton.tsx
+++ /dev/null
@@ -1,44 +0,0 @@
-import React, {
-  ReactNode,
-  useCallback,
-  useState,
-  type MouseEvent,
-} from "react";
-
-import { ExplainButtonTooltip } from "./ExplainButtonTooltip";
-
-export function ExplainButton({
-  children,
-  inferenceAddr,
-  prompt,
-  systemPrompt,
-}: {
-  children: ReactNode;
-  inferenceAddr: string;
-  prompt: string;
-  systemPrompt: string;
-}) {
-  const [isTooltipOpen, setIsTooltipOpen] = useState(false);
-
-  const onClick = useCallback(
-    function (event: MouseEvent<HTMLButtonElement>) {
-      event.preventDefault();
-
-      setIsTooltipOpen(true);
-    },
-    [setIsTooltipOpen],
-  );
-
-  return (
-    <div>
-      <button onClick={onClick}>{children}</button>
-      {isTooltipOpen && (
-        <ExplainButtonTooltip
-          inferenceAddr={inferenceAddr}
-          prompt={prompt}
-          systemPrompt={systemPrompt}
-        />
-      )}
-    </div>
-  );
-}
diff --git a/resources/ts/components/ExplainButtonTooltip.tsx b/resources/ts/components/ExplainButtonTooltip.tsx
deleted file mode 100644
index abdf856e7..000000000
--- a/resources/ts/components/ExplainButtonTooltip.tsx
+++ /dev/null
@@ -1,21 +0,0 @@
-import React from "react";
-
-import { usePrompt } from "../hooks/usePrompt";
-
-export function ExplainButtonTooltip({
-  inferenceAddr,
-  prompt,
-  systemPrompt,
-}: {
-  inferenceAddr: string;
-  prompt: string;
-  systemPrompt: string;
-}) {
-  const { message } = usePrompt({
-    inferenceAddr,
-    prompt,
-    systemPrompt,
-  });
-
-  return <div>{message}</div>;
-}
diff --git a/resources/ts/components/InferenceParameterCacheDtype.tsx b/resources/ts/components/InferenceParameterCacheDtype.tsx
deleted file mode 100644
index 70155109f..000000000
--- a/resources/ts/components/InferenceParameterCacheDtype.tsx
+++ /dev/null
@@ -1,55 +0,0 @@
-import React, { useCallback, useContext, type ChangeEvent } from "react";
-
-import { cacheDtypes } from "@intentee/paddler-client/schemas/InferenceParameters";
-import { InferenceParametersContext } from "../contexts/InferenceParametersContext";
-import {
-  inferenceParameterInput,
-  inferenceParameterInput__label,
-  inferenceParameterInput__select,
-} from "./inferenceParameterInput.module.css";
-
-function isCacheDtype(value: string): value is (typeof cacheDtypes)[number] {
-  return cacheDtypes.includes(value as (typeof cacheDtypes)[number]);
-}
-
-export function InferenceParameterCacheDtype({
-  description,
-  name,
-}: {
-  description: string;
-  name: "k_cache_dtype" | "v_cache_dtype";
-}) {
-  const { parameters, setParameter } = useContext(InferenceParametersContext);
-
-  const onChange = useCallback(
-    function (event: ChangeEvent<HTMLSelectElement>) {
-      const option = event.currentTarget.value;
-
-      if (!isCacheDtype(option)) {
-        throw new Error(`Invalid cache dtype: ${option}`);
-      }
-
-      setParameter(name, option);
-    },
-    [name, setParameter],
-  );
-
-  return (
-    <label className={inferenceParameterInput}>
-      <abbr className={inferenceParameterInput__label} title={description}>
-        {name}
-      </abbr>
-      <div className={inferenceParameterInput__select}>
-        <select name={name} value={parameters[name]} onChange={onChange}>
-          {cacheDtypes.map(function (option: string) {
-            return (
-              <option key={option} value={option}>
-                {option}
-              </option>
-            );
-          })}
-        </select>
-      </div>
-    </label>
-  );
-}
diff --git a/resources/ts/components/InferenceParameterCheckbox.tsx b/resources/ts/components/InferenceParameterCheckbox.tsx
deleted file mode 100644
index 3e251ef78..000000000
--- a/resources/ts/components/InferenceParameterCheckbox.tsx
+++ /dev/null
@@ -1,42 +0,0 @@
-import React, { useCallback, useContext } from "react";
-
-import { InferenceParametersContext } from "../contexts/InferenceParametersContext";
-import { type InferenceParametersBooleanKeys } from "../inferenceParametersFormKeys";
-import {
-  inferenceParameterInput,
-  inferenceParameterInput__checkbox,
-  inferenceParameterInput__label,
-} from "./inferenceParameterInput.module.css";
-
-export function InferenceParameterCheckbox({
-  description,
-  name,
-}: {
-  description: string;
-  name: InferenceParametersBooleanKeys;
-}) {
-  const { parameters, setParameter } = useContext(InferenceParametersContext);
-
-  const onChange = useCallback(
-    function () {
-      setParameter(name, !parameters[name]);
-    },
-    [name, parameters, setParameter],
-  );
-
-  return (
-    <label className={inferenceParameterInput}>
-      <abbr className={inferenceParameterInput__label} title={description}>
-        {name}
-      </abbr>
-      <div className={inferenceParameterInput__checkbox}>
-        <input
-          checked={parameters[name]}
-          name={name}
-          onChange={onChange}
-          type="checkbox"
-        />
-      </div>
-    </label>
-  );
-}
diff --git a/resources/ts/components/InferenceParameterInput.tsx b/resources/ts/components/InferenceParameterInput.tsx
deleted file mode 100644
index 0f6fd5246..000000000
--- a/resources/ts/components/InferenceParameterInput.tsx
+++ /dev/null
@@ -1,44 +0,0 @@
-import React, { useCallback, useContext, type FormEvent } from "react";
-
-import { InferenceParametersContext } from "../contexts/InferenceParametersContext";
-import { type InferenceParametersNumberKeys } from "../inferenceParametersFormKeys";
-import {
-  inferenceParameterInput,
-  inferenceParameterInput__input,
-  inferenceParameterInput__label,
-} from "./inferenceParameterInput.module.css";
-
-export function InferenceParameterInput({
-  description,
-  name,
-}: {
-  description: string;
-  name: InferenceParametersNumberKeys;
-}) {
-  const { parameters, setParameter } = useContext(InferenceParametersContext);
-
-  const onInput = useCallback(
-    function (event: FormEvent<HTMLInputElement>) {
-      event.preventDefault();
-
-      setParameter(name, parseFloat(event.currentTarget.value));
-    },
-    [name, setParameter],
-  );
-
-  return (
-    <label className={inferenceParameterInput}>
-      <abbr className={inferenceParameterInput__label} title={description}>
-        {name}
-      </abbr>
-      <input
-        className={inferenceParameterInput__input}
-        name={name}
-        onInput={onInput}
-        required
-        type="number"
-        value={parameters[name]}
-      />
-    </label>
-  );
-}
diff --git a/resources/ts/components/InferenceParameterPoolingType.tsx b/resources/ts/components/InferenceParameterPoolingType.tsx
deleted file mode 100644
index 10f280721..000000000
--- a/resources/ts/components/InferenceParameterPoolingType.tsx
+++ /dev/null
@@ -1,71 +0,0 @@
-import React, { useCallback, useContext, type ChangeEvent } from "react";
-
-import { poolingTypes } from "@intentee/paddler-client/schemas/InferenceParameters";
-import { InferenceParametersContext } from "../contexts/InferenceParametersContext";
-import {
-  inferenceParameterInput,
-  inferenceParameterInput__disabledHint,
-  inferenceParameterInput__disabledHint__content,
-  inferenceParameterInput__label,
-  inferenceParameterInput__select,
-} from "./inferenceParameterInput.module.css";
-
-const name = "pooling_type";
-
-function isPoolingType(value: string): value is (typeof poolingTypes)[number] {
-  return poolingTypes.includes(value as (typeof poolingTypes)[number]);
-}
-
-export function InferenceParameterPoolingType({
-  description,
-  disabled,
-}: {
-  description: string;
-  disabled: boolean;
-}) {
-  const { parameters, setParameter } = useContext(InferenceParametersContext);
-
-  const onChange = useCallback(
-    function (event: ChangeEvent<HTMLSelectElement>) {
-      const option = event.currentTarget.value;
-
-      if (!isPoolingType(option)) {
-        throw new Error(`Invalid pooling type: ${option}`);
-      }
-
-      setParameter(name, option);
-    },
-    [setParameter],
-  );
-
-  return (
-    <label className={inferenceParameterInput}>
-      <abbr className={inferenceParameterInput__label} title={description}>
-        {name}
-      </abbr>
-      <div className={inferenceParameterInput__select}>
-        <select
-          disabled={disabled}
-          name={name}
-          value={parameters[name]}
-          onChange={onChange}
-        >
-          {poolingTypes.map(function (option: string) {
-            return (
-              <option key={option} value={option}>
-                {option}
-              </option>
-            );
-          })}
-        </select>
-        {disabled && (
-          <div className={inferenceParameterInput__disabledHint}>
-            <div className={inferenceParameterInput__disabledHint__content}>
-              enable embeddings to turn this on
-            </div>
-          </div>
-        )}
-      </div>
-    </label>
-  );
-}
diff --git a/resources/ts/components/InferenceParametersContextProvider.tsx b/resources/ts/components/InferenceParametersContextProvider.tsx
deleted file mode 100644
index b67e2ed4d..000000000
--- a/resources/ts/components/InferenceParametersContextProvider.tsx
+++ /dev/null
@@ -1,57 +0,0 @@
-import React, { useMemo, useState, type ReactNode } from "react";
-
-import { type InferenceParameters } from "@intentee/paddler-client/schemas/InferenceParameters";
-import {
-  InferenceParametersContext,
-  type InferenceParametersContextValue,
-} from "../contexts/InferenceParametersContext";
-
-export function InferenceParametersContextProvider({
-  children,
-  defaultInferenceParameters,
-}: {
-  children: ReactNode;
-  defaultInferenceParameters: InferenceParameters;
-}) {
-  const [parameters, setParameters] = useState<InferenceParameters>(
-    defaultInferenceParameters,
-  );
-
-  const value = useMemo<InferenceParametersContextValue>(
-    function () {
-      function setPartialParameters(
-        partialParameters: Partial<InferenceParameters>,
-      ) {
-        setParameters({
-          ...parameters,
-          ...partialParameters,
-        });
-      }
-
-      function setParameter<TKey extends keyof InferenceParameters>(
-        name: TKey,
-        value: InferenceParameters[TKey],
-      ) {
-        setPartialParameters({
-          ...parameters,
-          ...{
-            [name]: value,
-          },
-        });
-      }
-
-      return Object.freeze({
-        parameters,
-        setParameter,
-        setPartialParameters,
-      });
-    },
-    [parameters, setParameters],
-  );
-
-  return (
-    <InferenceParametersContext.Provider value={value}>
-      {children}
-    </InferenceParametersContext.Provider>
-  );
-}
diff --git a/resources/ts/components/ModelRuntimeParametersFields.tsx b/resources/ts/components/ModelRuntimeParametersFields.tsx
new file mode 100644
index 000000000..5fc990947
--- /dev/null
+++ b/resources/ts/components/ModelRuntimeParametersFields.tsx
@@ -0,0 +1,79 @@
+import React, { useCallback, useContext } from "react";
+
+import {
+  cacheDtypes,
+  type ModelRuntimeParameters,
+} from "@intentee/paddler-client/schemas/ModelRuntimeParameters";
+import { BalancerDesiredStateContext } from "../contexts/BalancerDesiredStateContext";
+import { ParameterNumberInput } from "./ParameterNumberInput";
+import { ParameterSelect } from "./ParameterSelect";
+
+export function ModelRuntimeParametersFields() {
+  const { balancerDesiredState, setBalancerDesiredState } = useContext(
+    BalancerDesiredStateContext,
+  );
+  const modelRuntimeParameters = balancerDesiredState.model_runtime_parameters;
+
+  const setParameter = useCallback(
+    function <TKey extends keyof ModelRuntimeParameters>(
+      name: TKey,
+      value: ModelRuntimeParameters[TKey],
+    ) {
+      setBalancerDesiredState({
+        ...balancerDesiredState,
+        model_runtime_parameters: {
+          ...balancerDesiredState.model_runtime_parameters,
+          [name]: value,
+        },
+      });
+    },
+    [balancerDesiredState, setBalancerDesiredState],
+  );
+
+  return (
+    <>
+      <ParameterNumberInput
+        description="Batch Size (higher = more memory usage, lower = less inference speed)"
+        name="n_batch"
+        onValue={function (value) {
+          setParameter("n_batch", value);
+        }}
+        value={modelRuntimeParameters.n_batch}
+      />
+      <ParameterNumberInput
+        description="Context Size (higher = longer inputs, lower = less memory usage)"
+        name="context_size"
+        onValue={function (value) {
+          setParameter("context_size", value);
+        }}
+        value={modelRuntimeParameters.context_size}
+      />
+      <ParameterNumberInput
+        description="Number of model layers to offload to GPU (0 = CPU only; -1 = all layers)"
+        name="n_gpu_layers"
+        onValue={function (value) {
+          setParameter("n_gpu_layers", value);
+        }}
+        value={modelRuntimeParameters.n_gpu_layers}
+      />
+      <ParameterSelect
+        description="Datatype for K cache tensors"
+        name="k_cache_dtype"
+        onValue={function (value) {
+          setParameter("k_cache_dtype", value);
+        }}
+        options={cacheDtypes}
+        value={modelRuntimeParameters.k_cache_dtype}
+      />
+      <ParameterSelect
+        description="Datatype for V cache tensors"
+        name="v_cache_dtype"
+        onValue={function (value) {
+          setParameter("v_cache_dtype", value);
+        }}
+        options={cacheDtypes}
+        value={modelRuntimeParameters.v_cache_dtype}
+      />
+    </>
+  );
+}
diff --git a/resources/ts/components/ModelUriInput.tsx b/resources/ts/components/ModelUriInput.tsx
new file mode 100644
index 000000000..91012d9a8
--- /dev/null
+++ b/resources/ts/components/ModelUriInput.tsx
@@ -0,0 +1,38 @@
+import React, { useCallback, type InputEvent } from "react";
+
+import { type AgentDesiredModel } from "@intentee/paddler-client/schemas/AgentDesiredModel";
+
+import { changeModelForm__input } from "./ChangeModelForm.module.css";
+
+export function ModelUriInput({
+  model,
+  name,
+  onModel,
+  required,
+}: {
+  model: AgentDesiredModel;
+  name: string;
+  onModel(this: void, model: AgentDesiredModel): void;
+  required: boolean;
+}) {
+  const onInput = useCallback(
+    function (event: InputEvent<HTMLInputElement>) {
+      const uri = event.currentTarget.value;
+
+      onModel(uri === "" ? "None" : { Uri: uri });
+    },
+    [onModel],
+  );
+
+  return (
+    <input
+      className={changeModelForm__input}
+      name={name}
+      onInput={onInput}
+      placeholder="https://huggingface.co/..."
+      required={required}
+      type="text"
+      value={model === "None" ? "" : model.Uri}
+    />
+  );
+}
diff --git a/resources/ts/components/ParameterNumberInput.tsx b/resources/ts/components/ParameterNumberInput.tsx
new file mode 100644
index 000000000..35acf31b0
--- /dev/null
+++ b/resources/ts/components/ParameterNumberInput.tsx
@@ -0,0 +1,44 @@
+import React, { useCallback, type InputEvent } from "react";
+
+import {
+  inferenceParameterInput,
+  inferenceParameterInput__input,
+  inferenceParameterInput__label,
+} from "./inferenceParameterInput.module.css";
+
+export function ParameterNumberInput({
+  description,
+  name,
+  onValue,
+  value,
+}: {
+  description: string;
+  name: string;
+  onValue(this: void, value: number): void;
+  value: number;
+}) {
+  const onInput = useCallback(
+    function (event: InputEvent<HTMLInputElement>) {
+      event.preventDefault();
+
+      onValue(parseFloat(event.currentTarget.value));
+    },
+    [onValue],
+  );
+
+  return (
+    <label className={inferenceParameterInput}>
+      <abbr className={inferenceParameterInput__label} title={description}>
+        {name}
+      </abbr>
+      <input
+        className={inferenceParameterInput__input}
+        name={name}
+        onInput={onInput}
+        required
+        type="number"
+        value={value}
+      />
+    </label>
+  );
+}
diff --git a/resources/ts/components/ParameterSelect.tsx b/resources/ts/components/ParameterSelect.tsx
new file mode 100644
index 000000000..1be4b2af6
--- /dev/null
+++ b/resources/ts/components/ParameterSelect.tsx
@@ -0,0 +1,55 @@
+import React, { useCallback, type ChangeEvent } from "react";
+
+import {
+  inferenceParameterInput,
+  inferenceParameterInput__label,
+  inferenceParameterInput__select,
+} from "./inferenceParameterInput.module.css";
+
+export function ParameterSelect<TOption extends string>({
+  description,
+  name,
+  onValue,
+  options,
+  value,
+}: {
+  description: string;
+  name: string;
+  onValue(this: void, value: TOption): void;
+  options: ReadonlyArray<TOption>;
+  value: TOption;
+}) {
+  const onChange = useCallback(
+    function (event: ChangeEvent<HTMLSelectElement>) {
+      const selectedOption = options.find(function (option) {
+        return option === event.currentTarget.value;
+      });
+
+      if (selectedOption === undefined) {
+        throw new Error(`Invalid ${name}: ${event.currentTarget.value}`);
+      }
+
+      onValue(selectedOption);
+    },
+    [name, onValue, options],
+  );
+
+  return (
+    <label className={inferenceParameterInput}>
+      <abbr className={inferenceParameterInput__label} title={description}>
+        {name}
+      </abbr>
+      <div className={inferenceParameterInput__select}>
+        <select name={name} value={value} onChange={onChange}>
+          {options.map(function (option) {
+            return (
+              <option key={option} value={option}>
+                {option}
+              </option>
+            );
+          })}
+        </select>
+      </div>
+    </label>
+  );
+}
diff --git a/resources/ts/components/PromptContextProvider.tsx b/resources/ts/components/PromptContextProvider.tsx
index ceffe9031..5cf1bc78f 100644
--- a/resources/ts/components/PromptContextProvider.tsx
+++ b/resources/ts/components/PromptContextProvider.tsx
@@ -1,4 +1,4 @@
-import React, { useEffect, useMemo, useState, type ReactNode } from "react";
+import React, { useMemo, useState, type ReactNode } from "react";
 
 import {
   PromptContext,
@@ -15,13 +15,6 @@ export function PromptContextProvider({ children }: { children: ReactNode }) {
     version: 0,
   });
 
-  useEffect(
-    function () {
-      setCurrentPrompt("");
-    },
-    [submittedPrompt, setCurrentPrompt],
-  );
-
   const value = useMemo<PromptContextValue>(
     function () {
       return Object.freeze({
@@ -29,6 +22,7 @@ export function PromptContextProvider({ children }: { children: ReactNode }) {
         isCurrentPromptEmpty: currentPrompt.trim() === "",
         setCurrentPrompt,
         setSubmittedPrompt(prompt: string | null) {
+          setCurrentPrompt("");
           setSubmittedPrompt(function ({ version }) {
             return Object.freeze({
               prompt,
diff --git a/resources/ts/components/PromptPage.tsx b/resources/ts/components/PromptPage.tsx
index a1f74e97b..e3a33d890 100644
--- a/resources/ts/components/PromptPage.tsx
+++ b/resources/ts/components/PromptPage.tsx
@@ -9,8 +9,6 @@ import { ConversationMessage } from "./ConversationMessage";
 import { ConversationMessagePromptGeneratedTokens } from "./ConversationMessagePromptGeneratedTokens";
 import { ConversationPromptInput } from "./ConversationPromptInput";
 import { FloatingStatus } from "./FloatingStatus";
-import { TokenGenerationDisabledContextProvider } from "./TokenGenerationDisabledContextProvider";
-import { TokenGenerationDisabledNotice } from "./TokenGenerationDisabledNotice";
 
 import {
   promptPage,
@@ -21,61 +19,50 @@ import {
 export function PromptPage() {
   const { inferenceAddr } = useContext(PaddlerConfigurationContext);
   const { submittedPrompt } = useContext(PromptContext);
-  const { socketState, webSocket } = useWebSocket({
+  const { socketState } = useWebSocket({
     endpoint: `${webSocketProtocol(window.location.protocol)}//${inferenceAddr}/api/v1/inference_socket`,
   });
 
-  return (
-    <TokenGenerationDisabledContextProvider webSocket={webSocket}>
-      {matchWebSocketState(socketState, {
-        connected({ webSocket }) {
-          return (
-            <div className={promptPage}>
-              <div className={promptPage__messages}>
-                {submittedPrompt && (
-                  <ConversationMessage
-                    author="You"
-                    errors={[]}
-                    isThinking={false}
-                    response={submittedPrompt}
-                    thoughts=""
-                  />
-                )}
-                <ConversationMessagePromptGeneratedTokens
-                  webSocket={webSocket}
-                />
-              </div>
-              <div className={promptPage__promptForm}>
-                <TokenGenerationDisabledNotice />
-                <ConversationPromptInput />
-              </div>
-            </div>
-          );
-        },
-        connecting() {
-          return (
-            <FloatingStatus>
-              Connecting to the inference server...
-            </FloatingStatus>
-          );
-        },
-        connectionClosed() {
-          return (
-            <FloatingStatus>
-              Connection to the inference server closed. Will try to
-              reconnect...
-            </FloatingStatus>
-          );
-        },
-        connectionError() {
-          return (
-            <FloatingStatus>
-              Cannot connect to the inference server. Will try again in a
-              moment...
-            </FloatingStatus>
-          );
-        },
-      })}
-    </TokenGenerationDisabledContextProvider>
-  );
+  return matchWebSocketState(socketState, {
+    connected({ webSocket }) {
+      return (
+        <div className={promptPage}>
+          <div className={promptPage__messages}>
+            {submittedPrompt && (
+              <ConversationMessage
+                author="You"
+                errors={[]}
+                isThinking={false}
+                response={submittedPrompt}
+                thoughts=""
+              />
+            )}
+            <ConversationMessagePromptGeneratedTokens webSocket={webSocket} />
+          </div>
+          <div className={promptPage__promptForm}>
+            <ConversationPromptInput />
+          </div>
+        </div>
+      );
+    },
+    connecting() {
+      return (
+        <FloatingStatus>Connecting to the inference server...</FloatingStatus>
+      );
+    },
+    connectionClosed() {
+      return (
+        <FloatingStatus>
+          Connection to the inference server closed. Will try to reconnect...
+        </FloatingStatus>
+      );
+    },
+    connectionError() {
+      return (
+        <FloatingStatus>
+          Cannot connect to the inference server. Will try again in a moment...
+        </FloatingStatus>
+      );
+    },
+  });
 }
diff --git a/resources/ts/components/TextGenerationSettingsFields.tsx b/resources/ts/components/TextGenerationSettingsFields.tsx
new file mode 100644
index 000000000..74cc2b42b
--- /dev/null
+++ b/resources/ts/components/TextGenerationSettingsFields.tsx
@@ -0,0 +1,94 @@
+import React, { useCallback, useContext } from "react";
+
+import { type SamplingParameters } from "@intentee/paddler-client/schemas/SamplingParameters";
+import { BalancerDesiredStateContext } from "../contexts/BalancerDesiredStateContext";
+import { ParameterNumberInput } from "./ParameterNumberInput";
+
+const samplingParameterDescriptions: ReadonlyArray<{
+  description: string;
+  name: keyof SamplingParameters;
+}> = [
+  {
+    description: "Minimum token probability to consider for selection",
+    name: "min_p",
+  },
+  { description: "Frequency Penalty", name: "penalty_frequency" },
+  {
+    description: "Number of last tokens to consider for penalty (0 = disabled)",
+    name: "penalty_last_n",
+  },
+  { description: "Presence Penalty", name: "penalty_presence" },
+  { description: "Repeated Token Penalty", name: "penalty_repeat" },
+  { description: "Temperature", name: "temperature" },
+  {
+    description: "Number of tokens to consider for selection",
+    name: "top_k",
+  },
+  {
+    description: "Probability threshold for selecting tokens",
+    name: "top_p",
+  },
+];
+
+export function TextGenerationSettingsFields() {
+  const { balancerDesiredState, setBalancerDesiredState } = useContext(
+    BalancerDesiredStateContext,
+  );
+  const { text_generation: textGenerationSettings } = balancerDesiredState;
+
+  const setSamplingParameter = useCallback(
+    function (name: keyof SamplingParameters, value: number) {
+      setBalancerDesiredState({
+        ...balancerDesiredState,
+        text_generation: {
+          ...textGenerationSettings,
+          sampling_parameters: {
+            ...textGenerationSettings.sampling_parameters,
+            [name]: value,
+          },
+        },
+      });
+    },
+    [balancerDesiredState, setBalancerDesiredState, textGenerationSettings],
+  );
+
+  const setImageResizeToFit = useCallback(
+    function (imageResizeToFit: number) {
+      setBalancerDesiredState({
+        ...balancerDesiredState,
+        text_generation: {
+          ...textGenerationSettings,
+          multimodal: {
+            ...textGenerationSettings.multimodal,
+            image_resize_to_fit: imageResizeToFit,
+          },
+        },
+      });
+    },
+    [balancerDesiredState, setBalancerDesiredState, textGenerationSettings],
+  );
+
+  return (
+    <>
+      <ParameterNumberInput
+        description="Max image dimension in pixels before resizing"
+        name="image_resize_to_fit"
+        onValue={setImageResizeToFit}
+        value={textGenerationSettings.multimodal.image_resize_to_fit}
+      />
+      {samplingParameterDescriptions.map(function ({ description, name }) {
+        return (
+          <ParameterNumberInput
+            description={description}
+            key={name}
+            name={name}
+            onValue={function (value) {
+              setSamplingParameter(name, value);
+            }}
+            value={textGenerationSettings.sampling_parameters[name]}
+          />
+        );
+      })}
+    </>
+  );
+}
diff --git a/resources/ts/components/TokenGenerationDisabledContextProvider.tsx b/resources/ts/components/TokenGenerationDisabledContextProvider.tsx
deleted file mode 100644
index f9d9e097b..000000000
--- a/resources/ts/components/TokenGenerationDisabledContextProvider.tsx
+++ /dev/null
@@ -1,57 +0,0 @@
-import React, { useEffect, useMemo, useState, type ReactNode } from "react";
-
-import { inferenceSocketClient } from "@intentee/paddler-client/inferenceSocketClient";
-import {
-  TokenGenerationDisabledContext,
-  type TokenGenerationDisabledContextValue,
-} from "../contexts/TokenGenerationDisabledContext";
-
-export function TokenGenerationDisabledContextProvider({
-  children,
-  webSocket,
-}: {
-  children: ReactNode;
-  webSocket: null | WebSocket;
-}) {
-  const [isTokenGenerationDisabled, setIsTokenGenerationDisabled] =
-    useState<boolean>(false);
-
-  useEffect(
-    function () {
-      setIsTokenGenerationDisabled(false);
-
-      if (!webSocket) {
-        return;
-      }
-
-      const socketClient = inferenceSocketClient({ webSocket });
-      const subscription = socketClient.clusterTokenGenerationMode$.subscribe(
-        function (notification) {
-          setIsTokenGenerationDisabled(
-            "TokenGenerationDisabled" === notification,
-          );
-        },
-      );
-
-      return function () {
-        subscription.unsubscribe();
-      };
-    },
-    [webSocket],
-  );
-
-  const value = useMemo<TokenGenerationDisabledContextValue>(
-    function () {
-      return Object.freeze({
-        isTokenGenerationDisabled,
-      });
-    },
-    [isTokenGenerationDisabled],
-  );
-
-  return (
-    <TokenGenerationDisabledContext.Provider value={value}>
-      {children}
-    </TokenGenerationDisabledContext.Provider>
-  );
-}
diff --git a/resources/ts/components/TokenGenerationDisabledNotice.module.css b/resources/ts/components/TokenGenerationDisabledNotice.module.css
deleted file mode 100644
index 3abedd742..000000000
--- a/resources/ts/components/TokenGenerationDisabledNotice.module.css
+++ /dev/null
@@ -1,8 +0,0 @@
-.tokenGenerationDisabledNotice {
-  background-color: var(--color-highlight-1);
-  border: 2px solid var(--color-border);
-  border-bottom: none;
-  box-shadow: 4px 4px 0 var(--color-border);
-  margin: 0 var(--spacing-half);
-  padding: var(--spacing-half) var(--spacing-base);
-}
diff --git a/resources/ts/components/TokenGenerationDisabledNotice.tsx b/resources/ts/components/TokenGenerationDisabledNotice.tsx
deleted file mode 100644
index df7cf8fc1..000000000
--- a/resources/ts/components/TokenGenerationDisabledNotice.tsx
+++ /dev/null
@@ -1,29 +0,0 @@
-import React, { useContext } from "react";
-
-import { TokenGenerationDisabledContext } from "../contexts/TokenGenerationDisabledContext";
-
-import { tokenGenerationDisabledNotice } from "./TokenGenerationDisabledNotice.module.css";
-
-export function TokenGenerationDisabledNotice() {
-  const { isTokenGenerationDisabled } = useContext(
-    TokenGenerationDisabledContext,
-  );
-
-  if (!isTokenGenerationDisabled) {
-    return null;
-  }
-
-  return (
-    <div className={tokenGenerationDisabledNotice}>
-      Token generation is disabled while the embeddings are enabled. Learn more
-      in the{" "}
-      <a
-        href="https://paddler.intentee.com/docs/starting-out/generating-tokens-and-embeddings/"
-        target="_blank"
-      >
-        documentation
-      </a>
-      .
-    </div>
-  );
-}
diff --git a/resources/ts/contexts/BalancerDesiredStateContext.ts b/resources/ts/contexts/BalancerDesiredStateContext.ts
new file mode 100644
index 000000000..38f01d400
--- /dev/null
+++ b/resources/ts/contexts/BalancerDesiredStateContext.ts
@@ -0,0 +1,21 @@
+import { createContext } from "react";
+
+import { type BalancerDesiredState } from "@intentee/paddler-client/schemas/BalancerDesiredState";
+
+export type BalancerDesiredStateContextValue = {
+  balancerDesiredState: BalancerDesiredState;
+  setBalancerDesiredState(
+    this: void,
+    balancerDesiredState: BalancerDesiredState,
+  ): void;
+};
+
+export const BalancerDesiredStateContext =
+  createContext<BalancerDesiredStateContextValue>({
+    get balancerDesiredState(): never {
+      throw new Error("BalancerDesiredStateContext not provided");
+    },
+    setBalancerDesiredState(): never {
+      throw new Error("BalancerDesiredStateContext not provided");
+    },
+  });
diff --git a/resources/ts/contexts/ChatTemplateContext.ts b/resources/ts/contexts/ChatTemplateContext.ts
index f330d7fb3..70ffbb4a4 100644
--- a/resources/ts/contexts/ChatTemplateContext.ts
+++ b/resources/ts/contexts/ChatTemplateContext.ts
@@ -1,13 +1,7 @@
 import { createContext } from "react";
 
-import { type ChatTemplate } from "@intentee/paddler-client/schemas/ChatTemplate";
-
 export type ChatTemplateContextValue = {
-  chatTemplateOverride: null | ChatTemplate;
-  setChatTemplateOverride(
-    this: void,
-    chatTemplateOverride: null | ChatTemplate,
-  ): void;
+  chatTemplateOverrideContent: string;
   setChatTemplateOverrideContent(this: void, content: string): void;
   setUseChatTemplateOverride(
     this: void,
@@ -17,15 +11,12 @@ export type ChatTemplateContextValue = {
 };
 
 export const ChatTemplateContext = createContext<ChatTemplateContextValue>({
-  get chatTemplateOverride(): never {
+  get chatTemplateOverrideContent(): never {
     throw new Error("ChatTemplateContext not provided");
   },
   get useChatTemplateOverride(): never {
     throw new Error("ChatTemplateContext not provided");
   },
-  setChatTemplateOverride(): void {
-    throw new Error("ChatTemplateContext not provided");
-  },
   setChatTemplateOverrideContent(): void {
     throw new Error("ChatTemplateContext not provided");
   },
diff --git a/resources/ts/contexts/InferenceParametersContext.ts b/resources/ts/contexts/InferenceParametersContext.ts
deleted file mode 100644
index 23e294590..000000000
--- a/resources/ts/contexts/InferenceParametersContext.ts
+++ /dev/null
@@ -1,29 +0,0 @@
-import { createContext } from "react";
-
-import { type InferenceParameters } from "@intentee/paddler-client/schemas/InferenceParameters";
-
-export type InferenceParametersContextValue = {
-  parameters: InferenceParameters;
-  setParameter<TKey extends keyof InferenceParameters>(
-    this: void,
-    name: TKey,
-    value: InferenceParameters[TKey],
-  ): void;
-  setPartialParameters(
-    this: void,
-    parameters: Partial<InferenceParameters>,
-  ): void;
-};
-
-export const InferenceParametersContext =
-  createContext<InferenceParametersContextValue>({
-    get parameters(): never {
-      throw new Error("InferenceParametersContext not provided");
-    },
-    setParameter(): never {
-      throw new Error("InferenceParametersContext not provided");
-    },
-    setPartialParameters(): never {
-      throw new Error("InferenceParametersContext not provided");
-    },
-  });
diff --git a/resources/ts/contexts/PaddlerConfigurationContext.ts b/resources/ts/contexts/PaddlerConfigurationContext.ts
index 00f12224f..e8e2009cb 100644
--- a/resources/ts/contexts/PaddlerConfigurationContext.ts
+++ b/resources/ts/contexts/PaddlerConfigurationContext.ts
@@ -5,6 +5,7 @@ import { type StatsdConfiguration } from "../StatsdConfiguration";
 export type PaddlerConfigurationContextValue = {
   bufferedRequestTimeoutMillis: number;
   compatOpenAIAddr: string | null;
+  compatTypeSafeAddr: string | null;
   inferenceAddr: string;
   managementAddr: string;
   maxBufferedRequests: number;
@@ -19,6 +20,9 @@ export const PaddlerConfigurationContext =
     get compatOpenAIAddr(): never {
       throw new Error("PaddlerConfigurationContext not provided");
     },
+    get compatTypeSafeAddr(): never {
+      throw new Error("PaddlerConfigurationContext not provided");
+    },
     get inferenceAddr(): never {
       throw new Error("PaddlerConfigurationContext not provided");
     },
diff --git a/resources/ts/contexts/TokenGenerationDisabledContext.ts b/resources/ts/contexts/TokenGenerationDisabledContext.ts
deleted file mode 100644
index d3ef4d415..000000000
--- a/resources/ts/contexts/TokenGenerationDisabledContext.ts
+++ /dev/null
@@ -1,12 +0,0 @@
-import { createContext } from "react";
-
-export type TokenGenerationDisabledContextValue = {
-  isTokenGenerationDisabled: boolean;
-};
-
-export const TokenGenerationDisabledContext =
-  createContext<TokenGenerationDisabledContextValue>({
-    get isTokenGenerationDisabled(): never {
-      throw new Error("TokenGenerationDisabledContext not provided");
-    },
-  });
diff --git a/resources/ts/controller_dashboard.tsx b/resources/ts/controller_dashboard.tsx
index be25ea7fe..bca896cbe 100644
--- a/resources/ts/controller_dashboard.tsx
+++ b/resources/ts/controller_dashboard.tsx
@@ -61,6 +61,8 @@ root.render(
       ),
       compatOpenAIAddr:
         rootNode.getOptionalStringFromDataset("compatOpenaiAddr"),
+      compatTypeSafeAddr:
+        rootNode.getOptionalStringFromDataset("compatTypesafeAddr"),
       inferenceAddr: rootNode.getStringFromDataset("inferenceAddr"),
       managementAddr: rootNode.getStringFromDataset("managementAddr"),
       maxBufferedRequests: rootNode.getIntFromDataset("maxBufferedRequests"),
diff --git a/resources/ts/hooks/useAgentDesiredModelUrl.ts b/resources/ts/hooks/useAgentDesiredModelUrl.ts
deleted file mode 100644
index dc8235926..000000000
--- a/resources/ts/hooks/useAgentDesiredModelUrl.ts
+++ /dev/null
@@ -1,73 +0,0 @@
-import { useMemo, useState } from "react";
-
-import { type AgentDesiredModel } from "@intentee/paddler-client/schemas/AgentDesiredModel";
-import { urlToAgentDesiredModel } from "@intentee/paddler-client/urlToAgentDesiredModel";
-
-type EmptyState = {
-  agentDesiredModel: "None";
-  error: null;
-  ok: true;
-};
-
-type ErrorState = {
-  agentDesiredModel: null;
-  error: string;
-  ok: false;
-};
-
-type SuccessState = {
-  agentDesiredModel: AgentDesiredModel;
-  error: null;
-  ok: true;
-};
-
-export type AgentDesiredModelState = EmptyState | ErrorState | SuccessState;
-
-const emptyState: EmptyState = Object.freeze({
-  agentDesiredModel: "None",
-  error: null,
-  ok: true,
-});
-
-export function useAgentDesiredModelUrl({
-  defaultModelUri,
-}: {
-  defaultModelUri: null | string;
-}): {
-  agentDesiredModelState: AgentDesiredModelState;
-  modelUri: null | string;
-  setModelUri(this: void, modelUri: null | string): void;
-} {
-  const [modelUri, setModelUri] = useState<null | string>(defaultModelUri);
-
-  const agentDesiredModelState: AgentDesiredModelState = useMemo(
-    function () {
-      if (!modelUri) {
-        return emptyState;
-      }
-
-      try {
-        return Object.freeze({
-          agentDesiredModel: urlToAgentDesiredModel(new URL(modelUri)),
-          empty: false,
-          error: null,
-          ok: true,
-        });
-      } catch (error: unknown) {
-        return Object.freeze({
-          agentDesiredModel: null,
-          empty: false,
-          error: String(error),
-          ok: false,
-        });
-      }
-    },
-    [modelUri],
-  );
-
-  return Object.freeze({
-    agentDesiredModelState,
-    modelUri,
-    setModelUri,
-  });
-}
diff --git a/resources/ts/hooks/useFetchJson.ts b/resources/ts/hooks/useFetchJson.ts
index 19589d938..763ab967f 100644
--- a/resources/ts/hooks/useFetchJson.ts
+++ b/resources/ts/hooks/useFetchJson.ts
@@ -1,7 +1,6 @@
 import { useEffect, useState } from "react";
 import type { z } from "zod";
 
-import { fetchJsonEmptyState } from "@intentee/paddler-client/FetchJsonEmptyState";
 import { fetchJsonLoadingState } from "@intentee/paddler-client/FetchJsonLoadingState";
 import { type FetchJsonState } from "@intentee/paddler-client/FetchJsonState";
 
@@ -9,32 +8,30 @@ export function useFetchJson<TResponseSchema extends z.ZodType>({
   produceFetchPromise,
   responseSchema,
 }: {
-  produceFetchPromise(
-    this: void,
-    abortSignal: AbortSignal,
-  ): null | Promise<Response>;
+  produceFetchPromise(this: void, abortSignal: AbortSignal): Promise<Response>;
   responseSchema: TResponseSchema;
 }): FetchJsonState<z.infer<TResponseSchema>> {
   const [fetchState, setFetchState] = useState<
     FetchJsonState<z.infer<TResponseSchema>>
   >(fetchJsonLoadingState);
+  const [requestedFetch, setRequestedFetch] = useState({
+    produceFetchPromise,
+    responseSchema,
+  });
+
+  if (
+    requestedFetch.produceFetchPromise !== produceFetchPromise ||
+    requestedFetch.responseSchema !== responseSchema
+  ) {
+    setRequestedFetch({ produceFetchPromise, responseSchema });
+    setFetchState(fetchJsonLoadingState);
+  }
 
   useEffect(
     function () {
       const abortController = new AbortController();
-      const fetchPromise = produceFetchPromise(abortController.signal);
-
-      if (!fetchPromise) {
-        setFetchState(fetchJsonEmptyState);
-
-        return function () {
-          abortController.abort("Fetch promise was not provided.");
-        };
-      }
-
-      setFetchState(fetchJsonLoadingState);
 
-      fetchPromise
+      produceFetchPromise(abortController.signal)
         .then(function (response) {
           if (!response.ok) {
             throw new Error(`HTTP error status: ${response.status}`);
@@ -46,6 +43,10 @@ export function useFetchJson<TResponseSchema extends z.ZodType>({
           return responseSchema.parse(result);
         })
         .then(function (result: z.infer<TResponseSchema>) {
+          if (abortController.signal.aborted) {
+            return;
+          }
+
           setFetchState({
             empty: false,
             error: null,
@@ -55,6 +56,10 @@ export function useFetchJson<TResponseSchema extends z.ZodType>({
           });
         })
         .catch(function (error: unknown) {
+          if (abortController.signal.aborted) {
+            return;
+          }
+
           setFetchState({
             empty: false,
             error: String(error),
diff --git a/resources/ts/hooks/usePrompt.ts b/resources/ts/hooks/usePrompt.ts
deleted file mode 100644
index e5e68edd2..000000000
--- a/resources/ts/hooks/usePrompt.ts
+++ /dev/null
@@ -1,70 +0,0 @@
-import { useEffect, useState } from "react";
-
-import { InferenceServiceGenerateTokensResponseSchema } from "@intentee/paddler-client/schemas/InferenceServiceGenerateTokensResponse";
-import { streamHttpNdjson } from "@intentee/paddler-client/streamHttpNdjson";
-
-export function usePrompt({
-  inferenceAddr,
-  prompt,
-  systemPrompt,
-}: {
-  inferenceAddr: string;
-  prompt: string;
-  systemPrompt: string;
-}) {
-  const [message, setMessage] = useState<string>("");
-
-  useEffect(
-    function () {
-      const abortController = new AbortController();
-
-      setMessage("");
-
-      const subscription = streamHttpNdjson({
-        url: `//${inferenceAddr}/api/v1/continue_from_conversation_history`,
-        body: {
-          add_generation_prompt: true,
-          conversation_history: [
-            { role: "assistant", content: systemPrompt },
-            { role: "user", content: prompt },
-          ],
-          enable_thinking: false,
-          max_tokens: 300,
-        },
-        signal: abortController.signal,
-        schema: InferenceServiceGenerateTokensResponseSchema,
-      }).subscribe({
-        next(validatedMessage) {
-          if (validatedMessage.done) {
-            return;
-          }
-
-          if (null === validatedMessage.token) {
-            return;
-          }
-
-          if ("content" !== validatedMessage.tokenKind) {
-            return;
-          }
-
-          setMessage(function (prevMessage) {
-            return `${prevMessage}${validatedMessage.token}`;
-          });
-        },
-        error(error: unknown) {
-          console.error("Error during fetch:", error);
-        },
-      });
-
-      return function () {
-        subscription.unsubscribe();
-        abortController.abort();
-      };
-    },
-    [inferenceAddr, prompt, systemPrompt],
-  );
-
-  return {
-    message,
-  };
-}
diff --git a/resources/ts/inferenceParametersFormKeys.ts b/resources/ts/inferenceParametersFormKeys.ts
deleted file mode 100644
index 9ebc086d6..000000000
--- a/resources/ts/inferenceParametersFormKeys.ts
+++ /dev/null
@@ -1,17 +0,0 @@
-import type { InferenceParameters } from "@intentee/paddler-client/schemas/InferenceParameters";
-
-export type InferenceParametersBooleanKeys = {
-  [TKey in keyof InferenceParameters]: TKey extends string
-    ? InferenceParameters[TKey] extends boolean
-      ? TKey
-      : never
-    : never;
-}[keyof InferenceParameters];
-
-export type InferenceParametersNumberKeys = {
-  [TKey in keyof InferenceParameters]: TKey extends string
-    ? InferenceParameters[TKey] extends number
-      ? TKey
-      : never
-    : never;
-}[keyof InferenceParameters];
diff --git a/rust-toolchain.toml b/rust-toolchain.toml
index 1c638ee28..c116ff34d 100644
--- a/rust-toolchain.toml
+++ b/rust-toolchain.toml
@@ -1,3 +1,3 @@
 [toolchain]
-channel = "1.95.0"
+channel = "1.99.1"
 components = ["clippy", "rust-analyzer", "rustfmt"]
diff --git a/tsconfig.json b/tsconfig.json
index 617321044..3d0fcb2b0 100644
--- a/tsconfig.json
+++ b/tsconfig.json
@@ -3,7 +3,6 @@
         "allowJs": true,
         "allowSyntheticDefaultImports": true,
         "alwaysStrict": true,
-        "downlevelIteration": false,
         "emitDecoratorMetadata": true,
         "esModuleInterop": true,
         "experimentalDecorators": true,