OpenAI bridge drops Claude Code tool calls

[unohee]

Description

Claude Code requests routed through an openai/* model work for text-only turns, but tool calls do not round-trip through the Messages/Responses bridge. When Codex returns a function_call, the bridge drops it and Claude Code receives an empty response with stop_reason: "end_turn".

Multi-turn tool transcripts also use invalid Responses input shapes: function_call and function_call_output are nested inside message content, and assistant text is encoded as input_text.

Steps to reproduce

1. Configure an OpenAI subscription account and an openai/* model.
2. Ask Claude Code to perform a task that requires a tool, such as reading a file.
3. Let the request continue into the tool-result turn.

Expected behaviour

• Codex function_call output is returned to Claude Code as an Anthropic tool_use block.
• The following tool_result is sent back as a top-level Responses function_call_output item using the same call_id.
• Streaming and non-streaming responses preserve text/tool ordering and return stop_reason: "tool_use" when a tool is requested.

Actual behaviour

The response mapper keeps text items only, so function calls disappear. On the next turn, the request converter nests call/result items inside message content; the Codex backend rejects that shape with HTTP 400.

Environment

• OS: macOS 26.5
• Node.js version: v26.4.0
• cc-router version: 0.6.2 (main)
• Mode: standalone

Logs

No proxy error is emitted for the dropped response item. The invalid follow-up request is rejected by the Codex backend with HTTP 400 invalid_value.

[VictorMinemu]

Thank you for the detailed report and for following up with PR #9. We reviewed the implementation and the core fix is valuable, but found a few safety edge cases around incomplete tool calls, non-function Responses items, and refusals. We requested focused changes and negative-path tests before merging. We will keep this issue open and re-review the PR once those updates are ready. Thanks again for contributing to CC-Router.

[VictorMinemu]

@unohee Thank you again for the detailed report and for contributing the original fix in #9. We have now merged it together with the additional hardening identified during review: handling non-function Responses output items safely, rejecting invalid or incomplete tool calls and SSE streams, preserving refusals, and validating tool call IDs, names, and arguments.

When you have a chance, could you please test the latest main in your original Claude Code setup and confirm that the tool-call round trip works correctly for you? If anything still behaves unexpectedly, please reply here with the details and we will take another look.

Thanks again for helping improve CC-Router!