fix(prompt): omit tool_choice when prompt manages built-in tools #2464
Conversation
![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: 8e93bffd17
ℹ️ About Codex in GitHub
Your team has set up Codex to 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 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| should_omit_model = prompt is not None and not self._model_is_explicit | ||
| model_param: str | ChatModel | Omit = self.model if not should_omit_model else omit | ||
| should_omit_tools = prompt is not None and len(converted_tools_payload) == 0 | ||
| should_omit_tool_choice = should_omit_tools and tool_choice is not omit |
There was a problem hiding this comment.
none tool_choice with prompt-managed tools
This condition now omits every non-default tool_choice whenever prompt is set and local tools are empty, which also drops valid literals like "none"/"required" rather than only problematic named choices. In prompt-template flows where tools are defined server-side, ModelSettings(tool_choice="none") will be silently ignored, so callers can no longer reliably disable tool use per request and behavior can change unexpectedly.
Useful? React with 👍 / 👎.
1 participant
Summary
Fixes an issue where prompt-managed built-in tools can fail with:
Tool choice 'web_search_preview' not found in 'tools' parameter.This happens when a prompt is supplied, local tools are empty, and a
tool_choicevalue is still forwarded in the request.What changed
OpenAIResponsesModel._fetch_response(...), when prompt mode is active and localtoolsare omitted,tool_choiceis also omitted.tool_choiceis omitted in this prompt-managed-tools path.Why this is safe
Validation
uv run --with ruff ruff check src/agents/models/openai_responses.py tests/test_openai_responses.pyuv run --with mypy --with socksio mypy src/agents/models/openai_responses.py tests/test_openai_responses.pyuv run --with pytest --with socksio pytest -q tests/test_openai_responses.py