feat: Anthropic Messages API endpoint (/v1/messages) (#21)

* feat: add Anthropic Messages API endpoint (/v1/messages)

* fix: review fixes for /v1/messages edge handling

- model_not_found catch referenced out-of-scope 'model' (ReferenceError
  when resolveRequestedModel threw without a message); read req.body.model
- streaming idleTimeout fallback duplicated already-streamed text: reuse
  the chat-completions startsWith prefix-trim, and only full-poll on
  __error/noData or a fully empty stream
- terminal upstream failure during streaming returned an empty 200
  message; emit an Anthropic error event + message_stop instead

* fix(anthropic): normalize 5xx error types, preserve tool_result metadata, and flush filter buffers

* fix(anthropic): emit streaming deltas incrementally, force required tool calls, and pass registry to markup stripper

Author: TiaraBasori

CHANGELOG.md

@@ -9,6 +9,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ### Added
 
+- **Anthropic Messages API**: `POST /v1/messages` implements the Anthropic Messages protocol (Claude SDK compatible) on top of the same OpenCode backend — `x-api-key` or Bearer auth, `system` (string or blocks), text/image/tool_use/tool_result content blocks, `tools` with `input_schema`, `tool_choice`, `stop_sequences`, and Anthropic-shaped errors (`{type:'error',error:{type,message}}`). Tool calls surface as `tool_use` content blocks with the original client tool names (no `external__` namespace). Streaming (`stream: true`) emits the Anthropic SSE sequence — `message_start`, `content_block_start/delta/stop` (`thinking_delta`, `text_delta`, `input_json_delta`), `message_delta`, `message_stop` — with keepalives and polling fallback, mirroring the chat-completions pipeline including transient-error retry. Documented in both READMEs; covered by `tests/unit/anthropic.test.js`.
 - **Built-in Web Panel (`WEB_PANEL_ENABLED` / `OPENCODE_WEB_PANEL_ENABLED`, default off)**: An optional read-only dashboard served by the proxy itself. `GET /panel` returns a self-contained dark-themed HTML page (zero external assets, CSP locked to same-origin, inline style/script only) showing runtime status (status, uptime, backend URL), the model list, and backend diagnostics from `/health/details` (tool-mode counters, tool ID cache). The page fetches `GET /panel/data.json` — a JSON feed assembled from the models list, the `/health/details` payload and `process.uptime()` — with a Bearer key kept in the browser's localStorage. Auth matches the diagnostics endpoints: with `API_KEY` set, `data.json` requires the Bearer key (the HTML shell is public but carries no data), and `PUBLIC_DIAGNOSTICS=true` or an empty `API_KEY` lifts auth. When disabled, both routes 404 and cost nothing. The feed never returns `OPENCODE_SERVER_PASSWORD`, `ZEN_API_KEY`, `API_KEY` or `API_KEYS` values, and responses carry `X-Content-Type-Options: nosniff` plus `Cache-Control: no-store`. Added `tests/unit/panel.test.js` (10 tests: disabled 404s, headers, 401 gating, secret-leak assertions, public/no-key modes) and full configuration docs.
 
 ## [2.5.0] - 2026-10-15

README.en.md

@@ -104,6 +104,49 @@ curl -X POST http://127.0.0.1:10000/v1/chat/completions \
 
 When the model calls tools, non-streaming responses return `message.tool_calls` and streaming responses return `delta.tool_calls`.
 
+### Anthropic Messages API (/v1/messages)
+
+An endpoint for Claude SDK / Anthropic clients. Authentication uses the `x-api-key` header (`Authorization: Bearer` also works), and tool calls come back as `tool_use` content blocks.
+
+```bash
+curl -X POST http://127.0.0.1:10000/v1/messages \
+  -H "x-api-key: ***" \
+  -H "anthropic-version: 2023-06-01" \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "opencode/big-pickle",
+    "max_tokens": 1024,
+    "system": "You are concise.",
+    "messages": [{"role": "user", "content": "Hello!"}],
+    "stream": true
+  }'
+```
+
+Tool round-trip (send results back as `tool_result` blocks):
+
+```bash
+curl -X POST http://127.0.0.1:10000/v1/messages \
+  -H "x-api-key: ***" \
+  -H "anthropic-version: 2023-06-01" \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "opencode/big-pickle",
+    "max_tokens": 1024,
+    "messages": [{"role": "user", "content": "Fetch the title of https://example.com"}],
+    "tools": [{
+      "name": "web_fetch",
+      "description": "Fetch a URL and return its content summary",
+      "input_schema": {
+        "type": "object",
+        "properties": {"url": {"type": "string"}},
+        "required": ["url"]
+      }
+    }]
+  }'
+```
+
+The response contains a `{"type":"tool_use","id":"…","name":"web_fetch","input":{…}}` content block (`stop_reason: "tool_use"`); send the result back as a `{"type":"tool_result","tool_use_id":"…","content":"…"}` user block to continue the conversation. With `stream: true` the proxy emits the Anthropic SSE sequence `message_start` → `content_block_*` → `message_delta` → `message_stop`.
+
 ## ⚙️ Configuration
 
 | Environment variable | Default | Description |
@@ -146,6 +189,7 @@ OPENCODE_PROXY_AUTO_CLEANUP_CONVERSATIONS=true
 | `GET` | `/v1/models` | Model list |
 | `POST` | `/v1/chat/completions` | Chat Completions |
 | `POST` | `/v1/responses` | Responses API |
+| `POST` | `/v1/messages` | Anthropic Messages API (Claude SDK compatible) |
 
 Model names: `opencode/big-pickle`, `gpt5-nano` (auto-resolved to `gpt-5-nano`), `opencode/gpt5-nano`.
 

README.md

@@ -104,6 +104,49 @@ curl -X POST http://127.0.0.1:10000/v1/chat/completions \
 
 模型决定调用工具时，非流式响应返回 `message.tool_calls`，流式响应返回 `delta.tool_calls`。
 
+### Anthropic Messages API（/v1/messages）
+
+面向 Claude SDK / Anthropic 客户端的端点。认证使用 `x-api-key` header（`Authorization: Bearer` 亦可用），模型决定调用工具时返回 `tool_use` 内容块。
+
+```bash
+curl -X POST http://127.0.0.1:10000/v1/messages \
+  -H "x-api-key: ***" \
+  -H "anthropic-version: 2023-06-01" \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "opencode/big-pickle",
+    "max_tokens": 1024,
+    "system": "You are concise.",
+    "messages": [{"role": "user", "content": "你好!"}],
+    "stream": true
+  }'
+```
+
+工具往返（工具结果以 `tool_result` 块回传）：
+
+```bash
+curl -X POST http://127.0.0.1:10000/v1/messages \
+  -H "x-api-key: ***" \
+  -H "anthropic-version: 2023-06-01" \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "opencode/big-pickle",
+    "max_tokens": 1024,
+    "messages": [{"role": "user", "content": "帮我获取 https://example.com 的标题"}],
+    "tools": [{
+      "name": "web_fetch",
+      "description": "Fetch a URL and return its content summary",
+      "input_schema": {
+        "type": "object",
+        "properties": {"url": {"type": "string"}},
+        "required": ["url"]
+      }
+    }]
+  }'
+```
+
+响应包含 `{"type":"tool_use","id":"…","name":"web_fetch","input":{…}}` 内容块（`stop_reason: "tool_use"`）；把工具结果作为 `{"type":"tool_result","tool_use_id":"…","content":"…"}` 用户块发回即可继续对话。`stream: true` 时按 Anthropic SSE 事件序输出 `message_start` → `content_block_*` → `message_delta` → `message_stop`。
+
 ## ⚙️ 配置
 
 | 环境变量 | 默认值 | 说明 |
@@ -146,6 +189,7 @@ OPENCODE_PROXY_AUTO_CLEANUP_CONVERSATIONS=true
 | `GET` | `/v1/models` | 模型列表 |
 | `POST` | `/v1/chat/completions` | Chat Completions |
 | `POST` | `/v1/responses` | Responses API |
+| `POST` | `/v1/messages` | Anthropic Messages API（兼容 Claude SDK） |
 
 模型名称写法：`opencode/big-pickle`、`gpt5-nano`（自动解析为 `gpt-5-nano`）、`opencode/gpt5-nano`。