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

[suyuan2022]

Summary

• Add native Anthropic Messages API (/v1/messages) endpoint alongside the existing OpenAI-compatible /v1/chat/completions
• Both streaming (SSE with proper event: types) and non-streaming modes supported
• Preserves Anthropic-native response format: content blocks, usage stats (including cache tokens), stop_reason

Motivation

Many tools (OpenClaw/Clawdbot, etc.) work best with Anthropic's native Messages API format. The current OpenAI-only format causes issues:

• tool_use blocks get lost in OpenAI format conversion (related to Response format incompatible with Clawdbot's OpenAI provider #2)
• cache_creation_input_tokens / cache_read_input_tokens are dropped
• Streaming event format mismatch causes parsing failures

This PR adds a parallel Anthropic endpoint so clients can choose their preferred format.

Changes

File	Description
src/types/anthropic.ts	Anthropic Messages API type definitions
src/adapter/anthropic-to-cli.ts	Request conversion: Anthropic Messages → CLI input
src/adapter/cli-to-anthropic.ts	Response conversion: CLI output → Anthropic Messages
src/server/routes.ts	New handleAnthropicMessages handler
src/server/index.ts	Register /v1/messages route

Streaming format

Follows Anthropic's SSE spec exactly:

event: message_start
data: {"type":"message_start","message":{...}}

event: content_block_start
data: {"type":"content_block_start","index":0,...}

event: content_block_delta
data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"..."}}

event: content_block_stop
data: {"type":"content_block_stop","index":0}

event: message_delta
data: {"type":"message_delta","delta":{"stop_reason":"end_turn",...},"usage":{...}}

event: message_stop
data: {"type":"message_stop"}

Test plan

• TypeScript compiles with zero errors
• /v1/messages returns proper 400 for empty messages
• /v1/messages returns proper 400 for missing max_tokens
• Non-streaming response returns correct Anthropic format with usage stats
• Streaming response produces correct SSE event sequence
• Existing /v1/chat/completions endpoint unaffected

[suyuan2022]

Production Verification ✅

This has been deployed and tested in a real production environment:

Setup:

• Server: Ubuntu 24.04, Node.js v24.13.0
• Claude CLI: v2.1.62, Pro subscription (OAuth)
• Client: OpenClaw gateway with anthropic-messages API mode
• Channel: Telegram group chat → OpenClaw → proxy /v1/messages → Claude CLI → Anthropic API

Results:

• ✅ Non-streaming Anthropic format response with correct content blocks
• ✅ Usage stats including cache_creation_input_tokens and cache_read_input_tokens correctly passed through
• ✅ CLI subprocess spawns, processes, and exits cleanly (code 0)
• ✅ stop_reason: "stop" correctly mapped
• ✅ Running as systemd service with auto-restart, stable over extended period

Note: When configuring baseUrl in OpenClaw, use http://127.0.0.1:3456 instead of http://localhost:3456 — the proxy binds to IPv4 only, and localhost may resolve to ::1 (IPv6) on some systems.

🤖 Generated with Claude Code