[WIP] feat(messages): add Anthropic Messages API extension

docs/content/docs/references/meta.json

@@ -17,6 +17,7 @@
     "[@xsai/utils-chat](https://doc.deno.land/https://esm.sh/@xsai/utils-chat)",
     "[@xsai/utils-reasoning](https://doc.deno.land/https://esm.sh/@xsai/utils-reasoning)",
     "[@xsai/utils-stream](https://doc.deno.land/https://esm.sh/@xsai/utils-stream)",
+    "[@xsai-ext/messages](https://doc.deno.land/https://esm.sh/@xsai-ext/messages)",
     "[@xsai-ext/providers](https://doc.deno.land/https://esm.sh/@xsai-ext/providers)",
     "[@xsai-ext/responses](https://doc.deno.land/https://esm.sh/@xsai-ext/responses)",
     "[@xsai-ext/telemetry](https://doc.deno.land/https://esm.sh/@xsai-ext/telemetry)",

packages-ext/messages/LIVE_INVESTIGATION.md

@@ -0,0 +1,80 @@
+# Anthropic Messages Live Investigation
+
+Date: 2026-03-15
+
+## xsAI changes made during investigation
+
+- Ignore trailing `data: [DONE]` sentinels in the streaming event parser.
+- Stop defaulting Anthropic tool definitions to `strict: true`.
+- Add a live verification script at `scripts/live.ts`.
+
+## Live results against current gateway
+
+Environment:
+
+- `ANTHROPIC_BASE_URL=https://copilot.api.menci.xyz`
+- `ANTHROPIC_MODEL=claude-opus-4.6-1m`
+
+Actual API base used by the live script:
+
+- `https://copilot.api.menci.xyz/v1/`
+
+Results:
+
+- `countTokens`: failed.
+  - Gateway response: `{"error":{"type":"invalid_request_error","message":"Failed to count tokens: malloc is not a function"}}`
+- `messages` basic stream: passed.
+- `messages` thinking stream: passed.
+- `messages` tool loop: passed when `tool_choice` was omitted and the prompt strongly required tool use.
+
+Additional manual verification:
+
+- Raw streaming output from the gateway ends with `data: [DONE]` after `event: message_stop`.
+- `tool_choice: { type: 'tool', name: 'add' }` and `tool_choice: { type: 'any' }` both failed through the gateway with:
+  - `Thinking may not be enabled when tool_choice forces tool use.`
+
+## What this means for xsAI
+
+- Ignoring `[DONE]` is required for real proxy compatibility.
+- Defaulting Anthropic tools to `strict: true` is too aggressive for compatibility layers.
+- The remaining `countTokens` failure is not explained by xsAI request shape.
+
+## copilot-deno findings
+
+### Endpoint shape
+
+- The app only exposes `POST /v1/messages` and `POST /v1/messages/count_tokens`.
+- It does not expose `/messages` or `/messages/count_tokens`.
+
+### Native messages proxy behavior
+
+- Native `/v1/messages` requests are forwarded through `proxySSE(...)` without filtering `[DONE]`.
+- This is why xsAI must tolerate a terminal `[DONE]` line even though Anthropic-native streams do not require it.
+
+### Request rewriting
+
+- `web_search` tools are stripped before forwarding.
+- The reserved keyword `x-anthropic-billing-header` is stripped from prompts.
+- Invalid thinking blocks are removed before native forwarding.
+- If the selected model advertises `adaptive_thinking`, the route overwrites the request with `payload.thinking = { type: 'adaptive' }` and also injects `output_config.effort = 'high'` when missing.
+
+### Tool schema shape
+
+- `copilot-deno`'s local Anthropic type for tools is:
+  - `name`
+  - `description?`
+  - `input_schema`
+- It does not include `strict` in its local Anthropic route types.
+
+### Why `countTokens` fails on the deployed gateway
+
+- The route tries to use `@anthropic-ai/tokenizer` for Claude models.
+- It only falls back to estimation if loading the tokenizer fails.
+- It does not fall back if calling the tokenizer function throws.
+- The deployed error `malloc is not a function` strongly suggests the tokenizer function is being loaded but then crashing at runtime in the deployed environment.
+
+## Suggested follow-up in copilot-deno
+
+- Add a second fallback around `countFn(extractPayloadText(payload))` so tokenizer invocation failures downgrade to estimation instead of returning 400.
+- Decide whether native `/v1/messages` should strip trailing `[DONE]` when proxying Anthropic-compatible streams.
+- Decide whether forced `tool_choice` should disable adaptive thinking rather than returning an upstream validation error.

packages-ext/messages/README.md

@@ -0,0 +1 @@
+https://xsai.js.org/docs/packages-ext/messages

packages-ext/messages/package.json

@@ -0,0 +1,49 @@
+{
+  "name": "@xsai-ext/messages",
+  "type": "module",
+  "version": "0.4.4",
+  "description": "extra-small AI SDK.",
+  "author": "Moeru AI",
+  "license": "MIT",
+  "homepage": "https://xsai.js.org",
+  "repository": {
+    "type": "git",
+    "url": "git+https://github.com/moeru-ai/xsai.git",
+    "directory": "packages-ext/messages"
+  },
+  "bugs": "https://github.com/moeru-ai/xsai/issues",
+  "keywords": [
+    "xsai",
+    "anthropic",
+    "ai"
+  ],
+  "sideEffects": false,
+  "exports": "./src/index.ts",
+  "files": [
+    "dist"
+  ],
+  "publishConfig": {
+    "exports": {
+      ".": {
+        "types": "./dist/index.d.ts",
+        "default": "./dist/index.js"
+      },
+      "./package.json": "./package.json"
+    },
+    "main": "./dist/index.js",
+    "types": "./dist/index.d.ts"
+  },
+  "scripts": {
+    "build": "pkgroll",
+    "test": "vitest run",
+    "test:live": "tsx scripts/live.ts"
+  },
+  "dependencies": {
+    "@xsai/shared": "workspace:~",
+    "eventsource-parser": "catalog:"
+  },
+  "devDependencies": {
+    "@standard-schema/spec": "catalog:",
+    "zod": "catalog:schema-dev"
+  }
+}