Conversation
|
@codex reveiw |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: f10e576f08
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".
| fn is_instruction_text(text: &str) -> bool { | ||
| text.starts_with("author: ") | ||
| && text.contains("\nrecipient: ") | ||
| && text.contains("\nother_recipients: [") | ||
| && text.contains("]\nContent: ") |
There was a problem hiding this comment.
Tighten inter-agent instruction detection to avoid false boundaries
Instruction detection is substring-based, so any assistant message containing these labels is treated as an inter-agent instruction. That feeds into is_user_turn_boundary, which drives rollback and token/turn accounting. A normal assistant response that prints this template can therefore be miscounted as a turn boundary and distort history behavior.
Useful? React with 👍 / 👎.
Send input now sends messages as assistant message and with this format: