Skip to main content

Responses API

First-class inbound OpenAI Responses protocol (openai_responses). Owner transformer: openai-responses. Messages, Chat Completions, and FIM are separate inbound routes — see API Overview.

CCR accepts OpenAI Responses requests at:

  • POST /v1/responses
  • POST /responses

Configure an OpenAI SDK with:

export OPENAI_BASE_URL=http://127.0.0.1:3456/v1
export OPENAI_API_KEY=your-router-api-key

The route supports string and message-array input, instructions, text/images by URL or data URL, function calls/results, function tools, reasoning effort/summary intent, token limits, usage, JSON responses, and ordered response.* SSE events. Function-call IDs are normalized to the Responses length and character limits while remaining correlated within the turn.

CCR does not own OpenAI conversation state. It rejects store: true, previous_response_id, conversation, and background execution. Provider file IDs are also rejected. Every inbound Responses tool — function, custom (Codex MCP / plugin tools), web_search, and other hosted types — is projected onto a Unified Chat Completions function tool so Chat Completions backends never see Responses-only hosted-tool variants. Responses custom tools accept freeform text rather than JSON arguments, so CCR carries their input through a synthetic required string parameter while talking to Chat Completions providers, then restores the client response as custom_tool_call / custom_tool_call_output. The client therefore sees its original tool protocol while the model calls the Unified function representation. Stream success closes after response.completed; stream failures emit response.failed. Chat Completions' [DONE] marker is not part of the client Responses stream.