feat: add retry mechanism for ModelBehaviorError#3587
Conversation
Adds max_model_retries to RunConfig. When set, ModelBehaviorError triggers automatic retry with error feedback to the model. The error message is fed back as a synthetic user message so the model can self-correct malformed responses (invalid tool call JSON, nonexistent tools, etc.). Supports both non-streaming (run_single_turn) and streaming (run_single_turn_streamed) paths. Closes openai#325 Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 62fe089a74
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| ), | ||
| stream_failed_retry_attempts[0], | ||
| async for event in retry_stream: | ||
| streamed_result._event_queue.put_nowait(RawResponsesStreamEvent(data=event)) |
There was a problem hiding this comment.
When max_model_retries is enabled for streamed runs, a malformed completed response (for example, a nonexistent tool) is only detected later in get_single_step_result_from_response, but this line has already delivered all raw events from the failed attempt to the caller. The retry then emits a second attempt into the same stream, so stream consumers can observe invalid response.output_item.done/tool_called events that are not part of the final accepted turn and may act on a tool call the SDK is about to discard. Consider buffering per-attempt events until the response is accepted, or disabling this retry path for streaming after any events are emitted.
Useful? React with 👍 / 👎.
| tool_use_tracker=tool_use_tracker, | ||
| server_manages_conversation=server_conversation_tracker is not None, | ||
| ) | ||
| except ModelBehaviorError as e: |
There was a problem hiding this comment.
This catches ModelBehaviorError around the entire step resolution, not just response parsing. When the error is raised from tool execution (for example, one malformed shell/apply_patch call while another function/computer/custom tool in the same turn has already run, since _execute_tool_plan runs tool groups concurrently), the retry asks the model again and can execute those side-effecting tools a second time. Restrict the retry to validation before any tool side effects, or only retry when no tool execution has started.
Useful? React with 👍 / 👎.
1 participant
Summary
Adds automatic retry for
ModelBehaviorErrorwith configurablemax_model_retriesonRunConfig.When a model produces a malformed response (invalid tool call JSON, nonexistent tool, etc.), the SDK raises
ModelBehaviorError. Previously there was no built-in retry — the agent just failed. This PR adds an automatic retry mechanism that feeds the error message back to the model so it can self-correct.Changes
RunConfig.max_model_retries: int = 0— new parameter controlling how many automatic retries are attempted whenModelBehaviorErroroccurs (defaults to 0, preserving existing behavior)run_single_turn(non-streaming) — wraps model call + response processing in a retry loop; on error, appends a synthetic user message with the error details and retries the model callrun_single_turn_streamed(streaming) — same retry logic for the streaming path; events from failed attempts are emitted (consistent with existing network-level retry behavior)How it works
get_single_step_result_from_responseraisesModelBehaviorError"Your previous response was invalid: <error message>"Usage
Closes #325