Paddler 5.0: completions, embeddings and decisions from one loaded model, kev decisions and TypeSafe compatibility

.cargo/config.toml

@@ -0,0 +1,2 @@
+[env]
+RUST_TEST_THREADS = "1"

.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

.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.

.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

.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

.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

.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
 

.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
 

.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)'

.dockerignore

@@ -1,7 +1,6 @@
 **/.coverage
 **/.mypy_cache
 **/.pytest_cache
-**/.ruff_cache
 **/.venv
 **/__pycache__
 **/dist

.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

.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

.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