vllm - 💡(How to fix) Fix [Bug]: Claude Code CLI >= 2.1.154 sends ctx/msg/system roles and breaks vLLM Anthropic Messages API validation

Your current environment

vLLM version: 0.20.1
Docker image: vllm/vllm-openai:v0.20.1-cu129
Python: 3.12 inside container
Endpoint: Anthropic-compatible /v1/messages
Client: Claude Code CLI >= 2.1.154
Deployment: Docker
Model: openai/gpt-oss-20b or compatible local model

🐛 Describe the bug

After upgrading Claude Code CLI to version 2.1.154 or later, Claude Code appears to send additional message roles such as ctx, msg, and system inside the Anthropic Messages API messages array.

vLLM's Anthropic protocol currently validates incoming message roles as only user or assistant.

Because of this strict validation, requests from recent Claude Code versions fail before model inference starts.

Example error:

API Error: 400 {
  "error": {
    "message": "1 validation error:\n  {'type': 'literal_error', 'loc': ('body', 'messages', 1, 'role'), 'msg': \"Input should be 'user' or 'assistant'\", 'input': 'system', 'ctx': {'expected': \"'user' or 'assistant'\"}}",
    "type": "Bad Request",
    "param": null,
    "code": 400
  }
}

This seems to have started after Claude Code CLI 2.1.154.

A temporary workaround reported by users is to expand the allowed roles in vllm/entrypoints/anthropic/protocol.py, but a proper upstream fix should probably normalize these roles instead of only expanding the enum.

Possible temporary patch:

- role: Literal["user", "assistant"]
+ role: Literal["user", "assistant", "ctx", "msg", "system"]