/review only returns 1-2 items at a time. Leading to you needing to run a loop on it over and over to discover other issues. Should be thorough and return with a complete list the first time

Resolved 💬 4 comments Opened Feb 6, 2026 by virgildotcodes Closed Feb 6, 2026

💡 Likely answer: A maintainer (guidedways, contributor) responded on this thread — see the highlighted reply below.

This has bugged me for the last few months, and it seems to me this must just be a prompting issue in the /review script.

As in the title - when you /review, either your uncommitted changes or a branch you're about to merge against a base branch, the review will come back with 1-2 items highlighted. Great, that's all it could find right? Nope. You fix those issues, run /review again, and it finds 1-2 other legitimate issues that were totally separate. Fix those, run /review again, and the same thing.

Can we please have it not "return early" once it's found 1-2 issues and do a thorough pass through all the requested changes and come back with its best attempt at a full list of issues the first time?

View original on GitHub ↗

4 Comments

guidedways contributor · 6 months ago

Honestly, I feel there's merit in the current approach as it can under certain circumstances be counterintuitive to report more than 1-2 items as some of these may just be interlinked. But yes, there's room for improvement.

For this reason, I've instead created a $code-review skill based on the review prompt used by Codex, which also then allows me to fine-tune for certain other edge cases our project(s) require.

guidedways contributor · 6 months ago

For this reason, I've instead created a $code-review skill based on the review prompt used by Codex, which also then allows me to fine-tune for certain other edge cases our project(s) require.

Obviously - this ends up in the _current_ context, however I've found that at times this also _jogs_ Codex's memory and helps with reasoning. My workflow usually is: make changes → /review → fix highlighted issues → $code-review → fix any left over issues → yeet

virgildotcodes · 6 months ago

Honestly, I feel there's merit in the current approach as it can under certain circumstances be counterintuitive to report more than 1-2 items as some of these may just be interlinked. But yes, there's room for improvement. For this reason, I've instead created a $code-review skill based on the review prompt used by Codex, which also then allows me to fine-tune for certain other edge cases our project(s) require.

Yeah it would make sense if they were excluding dependent issues, but I've found time and time again that the issues brought up on each subsequent pass are more often than not completely independent. Seems like we'd be better off if these were all surfaced in one pass.

etraut-openai contributor · 6 months ago

We're tracking this feature request in https://github.com/openai/codex/issues/5547. Please upvote it if you'd like to see this.