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

[carlushuang]

Summary

Add /v1/messages endpoint to ATOM's OpenAI server, enabling Claude Code and other Anthropic-compatible tools to use ATOM as a backend.

Depends on PR #775 (MiniMax M2.7 reasoning parser fix).

What it does

Translates between Anthropic Messages API format and ATOM's internal OpenAI format:

Claude Code CLI → /v1/messages (Anthropic format)
       ↓
serving_anthropic.py (format translation)
       ↓
ATOM engine (any model, e.g., MiniMax M2.7)
       ↓
GPU inference

Features

• Non-streaming and streaming responses
• Anthropic SSE event format: message_start, content_block_start, content_block_delta, content_block_stop, message_delta, message_stop
• Thinking/reasoning separation: <think> blocks → thinking content blocks (via ReasoningFilter)
• System messages: string or content-block array
• Tool definitions: Anthropic → OpenAI format translation
• Tool use/result messages: bidirectional translation

New files

• atom/entrypoints/openai/serving_anthropic.py — request/response schemas, format converters, SSE helpers

Usage with Claude Code

# 1. Start ATOM
python -m atom.entrypoints.openai_server --model MiniMaxAI/MiniMax-M2.7 \
  --trust-remote-code --kv_cache_dtype fp8 -tp 2 --server-port 8000

# 2. Configure Claude Code (~/.claude/settings.json)
{
  "env": {
    "ANTHROPIC_BASE_URL": "http://localhost:8000",
    "ANTHROPIC_AUTH_TOKEN": "dummy",
    "ANTHROPIC_MODEL": "MiniMax-M2.7",
    "DISABLE_PROMPT_CACHING": "1"
  }
}

# 3. Use Claude Code
claude

Verified on

• MiniMax M2.7 on MI355X (gfx950), TP=2, FP8 KV
• Claude Code --print "Say hello world" → Hello, world!
• Claude Code --print "Write is_prime function" → correct Python code
• Streaming and non-streaming both work
• Thinking content properly separated into thinking blocks

Test plan

• Non-streaming /v1/messages returns correct Anthropic format
• Streaming /v1/messages returns correct SSE events
• Thinking/reasoning separated into thinking content blocks
• Claude Code end-to-end: hello world, code generation, math
• Tool calling (needs model with tool-call support)