fix(core): count replayed program items as parents with client-managed input - #4840
Conversation
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
sylvesterkaczmarek
left a comment
sylvesterkaczmarek
left a comment
There was a problem hiding this comment.
The parent relation should come from durable input regardless of who manages conversation history. Counting replayed program items from input_items while leaving the server-managed-only shortcuts inside _collect_program_parent_state avoids widening parent acceptance. The Runner resume regression pins the actual HITL/replay path.
…d input A `program` item arriving in a run's input was only counted as a parent for a later `program_output` when the server manages the conversation. With client-managed history, a program that pauses across runs (a human-in-the-loop approval of a program-owned call, resumed with the app's own transcript as input) failed with "Model produced program_output item that does not match a parent program item" as soon as the model finished it. Pass the run's input items to `process_model_response` unconditionally and read `program` items from them in both modes, as `RunState` resume already does with durable input. The server-managed only branches, which also accept `program_output` and program caller ids as parents because the `program` itself may not be in the input there, are unchanged. Closes openai#4839
dfd2448 to
a56b8e8
Compare
![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: a56b8e8613
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".
| if server_manages_conversation | ||
| else None | ||
| ), | ||
| input_items=ItemHelpers.input_to_new_input_list(original_input), |
There was a problem hiding this comment.
When RunConfig.call_model_input_filter adds replayed program history—or removes it—the model receives filtered.input, but this line validates against the unfiltered original_input. Consequently, a supported client-managed callback that injects the parent program still causes a matching program_output to fail with ModelBehaviorError; conversely, an output can be accepted after the callback removed its parent. Carry the actual filtered input into response processing for both streaming and non-streaming calls.
AGENTS.md reference: AGENTS.md:L147-L148
Useful? React with 👍 / 👎.
seratch
left a comment
seratch
left a comment
There was a problem hiding this comment.
The current head still validates program parents against the unfiltered run input, while the model receives call_model_input_filter's result. A filter that supplies a replayed program can therefore produce a valid completion that the SDK rejects as parentless. Please base parent validation on the input actually sent to the model and cover this through both Runner paths.
3 participants
Summary
A
programitem that arrives in a run's input was never counted as a parent for a laterprogram_outputunless the server manages the conversation. With client-managed history, a program that pauses across runs (for example a human-in-the-loop approval of a program-owned call, resumed with the app's own transcript as input) fails withModelBehaviorError: Model produced program_output item that does not match a parent program itemas soon as the model finishes it.process_model_responsenow receives the run's input items unconditionally and the parent scan readsprogramitems from them in both modes. The server-managed only branches, which also acceptprogram_outputand programcallerids as parents because theprogramitself may not be in the input there, are unchanged. This matches whatRunStateresume already does with durable input items. The parameter is renamed fromserver_managed_input_itemstoinput_itemssince it no longer means that.Test plan
test_process_model_response_accepts_program_output_for_program_in_input: the scan accepts aprogram_outputwhose parent is only in the input.test_runner_completes_a_program_replayed_as_client_managed_input: aRunner.runwhose input carries a program, its program-owned call and output, finishes the program and answers. Fails onmainwith the error above.make format,make lint,make mypy,make testspass locally, and.agents/skills/code-change-verification/scripts/run.shreports all commands passed.Issue number
Closes #4839
Checks
.agents/skills/code-change-verification/scripts/run.sh/reviewbefore submitting this PR