Fix/responses api harmony channel metadata #28262 #28355
Conversation
baonudesifeizhai
requested review from
DarkLight1337,
NickLucche,
aarnphm,
chaunceyjiang,
robertgshaw2-redhat and
simon-mo
as code owners
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request addresses an issue where channel metadata was being lost during multi-turn conversations in the Responses API. The changes correctly assign 'commentary' or 'analysis' channels to function_call_output and reasoning messages, respectively, based on the conversational context. The addition of tests to verify this behavior is a great step. I have one suggestion to simplify the conditional logic for setting the channel on reasoning messages, which will improve code clarity and maintainability.
alecsolder
left a comment
alecsolder
left a comment
There was a problem hiding this comment.
Thank you for this! Left a few comments.
| Author.new(Role.TOOL, f"functions.{call_response.name}"), | ||
| response_msg["output"], | ||
| ) | ||
| msg = msg.with_channel("commentary") |
There was a problem hiding this comment.
Do you want to add the content_type change from the issue here too? Or are you just tackling one specific part of the issue?
There was a problem hiding this comment.
msg = msg.with_channel("commentary")
msg = msg.with_content_type("json") -- is that enough?
|
pytest tests/entrypoints/openai/test_response_api_with_harmony.py::test_function_call_with_previous_input_messages -v pass |
|
pytest tests/entrypoints/openai/test_response_api_with_harmony.py -v -k "not test_code_interpreter" this pass test_code_interpreter dont know why out of time |
2 participants
Purpose
Fix issue #28262: Restore missing channel metadata when converting Responses API output items back to Harmony Messages for multi-turn conversations.
Changes:
Set channel='commentary' for function_call_output type inputs
Set channel='analysis' or 'commentary' for reasoning type based on the following message (commentary if followed by function_call, analysis otherwise)
Add test to verify channel metadata is correctly preserved across conversation turns
Update parse_response_input() to accept optional next_msg parameter for context-aware channel assignmen
Test Plan
pytest tests/entrypoints/openai/test_response_api_with_harmony.py::test_function_call_with_previous_input_messages -v
Test Result
pass
Essential Elements of an Effective PR Description Checklist
supported_models.mdandexamplesfor a new model.