Inspect turn hooks before runtime preparation#29945
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: e522746dcf
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".
| .await; | ||
| return Ok(None); | ||
| } | ||
| let (inspected_input, blocked_input) = inspect_inputs(&sess, &turn_context, &input).await; |
There was a problem hiding this comment.
This reorders SessionStart/UserPromptSubmit inspection relative to the rest of turn setup and adds blocked/aborted setup paths, but the commit only changes production code. Because this is agent turn logic, please add an integration test under core/tests/suite that exercises the new ordering and persistence behavior so future changes do not regress it.
AGENTS.md reference: AGENTS.md:L112-L118
Useful? React with 👍 / 👎.
| { | ||
| accepted_user_input = true; | ||
| } | ||
| inspected_input.push((input_item, hook_outcome)); |
There was a problem hiding this comment.
When two user prompts are queued while the model is running, this loop now buffers every hook outcome and record_inspected_inputs is called only after inspect_inputs returns. That means the UserPromptSubmit hook for the second queued prompt runs before the first accepted prompt has been written to history/transcript_path; previously each accepted item was recorded before inspecting the next one. Hooks that read the transcript to decide whether to allow/block later queued prompts can now make decisions against stale history.
Useful? React with 👍 / 👎.
1 participant
Summary
Why
A runtime snapshot must not publish for a turn that hooks stop before sampling. Separating inspection from recording establishes that decision boundary without rewriting history or moving hook context ahead of the normal initial developer context.