🤖 Inject mode transition context for flaky test #224 #225
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
- Add 'mode' field to CmuxMetadata to track mode per message - Detect mode switches by comparing with last assistant message - Inject explicit transition note when mode changes mid-conversation - Helps models understand they should follow new mode instructions Addresses #224 - flaky mode-specific instructions test
![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.
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 👍.
Changed approach: inject mode switch as synthetic user message instead of system instruction enhancement. Benefits: - More temporal - transitions happen in conversation flow - Models handle in-message context better than system changes - Simpler - no need to persist mode across stream lifecycle - Avoids metadata persistence issue caught by Codex The synthetic message is inserted before the last user message when mode changes, providing natural context for the model. Implementation: - Added injectModeTransition() to modelMessageTransform.ts - Operates on CmuxMessage[] where metadata is available - Inserts synthetic user message: '[Mode switched from X to Y]' - Called after addInterruptedSentinel, before conversion to ModelMessage Co-authored-by: Codex (review feedback)
- Added 5 unit tests for injectModeTransition() - Fixed edge case: don't inject when no user messages exist - Pass mode to StreamManager so it persists in final history - Updated PR description to be clearer and more concise Co-authored-by: Codex (persistence fix)
ammario
approved these changes
Oct 13, 2025
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Problem
Integration test
should include mode-specific instructions in system messageis flaky (#224). When testing mode switches mid-conversation, models sometimes respond with the wrong mode marker ([PLAN_MODE_ACTIVE]instead of[EXEC_MODE_ACTIVE]).Root cause: Models see conflicting signals:
Models sometimes prioritize conversation consistency over system instructions.
Solution
Inject mode transition as a temporal user message in the conversation flow.
When mode changes, insert a synthetic user message before the final user message:
Benefits:
Implementation
injectModeTransition()tomodelMessageTransform.tsCmuxMessage[]where metadata.mode is availableaddInterruptedSentinel, before converting toModelMessageTesting
Closes #224