Responses API endpoint rejects system messages in input array

[Yinhan-Lu]

Bug Description

The /v1/responses endpoint rejects requests that include role: "system" messages in the input array, returning:

400 - {'detail': 'System messages are not allowed'}

The official OpenAI Responses API accepts system messages in the input array, so ccproxy is stricter than the upstream API.

Root Cause

CodexMessage in ccproxy/plugins/codex/models.py restricts roles to only "user" and "assistant":

class CodexMessage(BaseModel):
    role: Literal["user", "assistant"]  # ← "system" not allowed
    content: str

When a client (e.g., langchain-openai with use_responses_api=True) sends a system message in the input, Pydantic validation fails before the request reaches any conversion logic.

Impact

This forces downstream projects like EvoScientist to fall back from Responses API to Chat Completions API when using ccproxy, which means:

• No reasoning parameter support — langchain-openai ties reasoning to the Responses API; using Chat Completions means reasoning must be disabled
• OAuth users get a degraded experience compared to direct API key users

For reference: EvoScientist/EvoScientist#130, EvoScientist/EvoScientist#135

Note

The Chat Completions → Responses API converter (openai_to_openai/requests.py) already handles system messages correctly — it extracts them from messages and puts them into the instructions parameter. But this converter is only triggered on the /v1/chat/completions path, not on /v1/responses.

Suggested Fix

Option A (minimal): Add "system" to CodexMessage.role, and in the /v1/responses handler, extract system messages from input and merge them into the instructions parameter before forwarding upstream — reusing the same logic already in _build_responses_payload_from_chat_request.

Option B (passthrough): Since the official OpenAI Responses API accepts system messages in input, simply allow them to pass through without conversion.

Reproduction

from langchain.chat_models import init_chat_model

model = init_chat_model(
    "gpt-5.4",
    model_provider="openai",
    base_url="http://127.0.0.1:8000/codex/v1",
    api_key="ccproxy-oauth",
    use_responses_api=True,
)
from langchain_core.messages import SystemMessage, HumanMessage
model.invoke([SystemMessage(content="You are helpful"), HumanMessage(content="Hi")])
# → BadRequestError: 400 - {'detail': 'System messages are not allowed'}

Environment

• ccproxy-api: 0.2.7
• langchain-openai: 1.1.12
• Python: 3.13
• OS: macOS