Make `/review` configurable to specify number of issues

Potential duplicates detected. Please review them and close your issue if it is a duplicate.

• #5375

Powered by Codex Action

I agree with everything that has been said here!

In addition to being able to submit multiple suggestions, it would be great to be able to launch a review in non-interactive mode, such as codex review --branch my-feature --base develop, or something like that.

Thanks for the positive feedback! We're glad you like the feature.

And thanks also for the suggestions. I'm going to change the title of the issue to reflect the enhancement request.

When does this land? It will greatly improve using codex as reviewer in github integration

When can we expect this landing? This really makes us think to getting another agent x github integration if there is no timeline

+1, useless without this.

@etraut-openai can you put attention to this please?

+1, I keep fixing 1 or 2 and then have it run another review.

This might help some people. What I do is I paste this prompt into my locally running codex agent. Note: it's important that it have GitHub CLI access to your repo. If you're ok with this level of trust, then ship it. It has worked well for me at least. I go work on other things, and eventually it's done and I can review what it did. YMMV 🤷‍♂️

TL;DR - it runs a loop where it polls for additional review comments and then fixes them until there are no more comments.

Run a closed-loop GitHub Codex review remediation cycle for the current branch PR targeting `dev`.

Context:

- You have terminal access and `gh` CLI.

- Do not ask clarifying questions.

- Execute immediately and keep looping until completion criteria are met.

Completion criteria (all required):

1) PR description has Codex 👍 reaction.

2) PR description no longer has Codex 👀 reaction.

3) No unresolved Codex review threads/comments remain.

Rules:

- Poll every 60 seconds.

- Triage only substantive review items (bugs, regressions, security, correctness, tests, perf).

- For each Codex comment:

- If valid: fix it, run relevant tests/lint, commit (Conventional Commit), push.

- If not valid: reply with concise technical rationale and evidence.

- Resolve addressed threads.

- Do not use partial staging.

- Do not amend commits.

- Do not commit secrets.

- Do not stop after one pass.

Required workflow:

1) Discover PR for current branch against `dev`.

2) Enter a 60s polling loop:

- Refresh PR state (including description reactions).

- Fetch unresolved review threads/comments.

- Filter Codex-authored items (e.g. `chatgpt-codex-connector`).

- Process newly found items immediately (fix or reply).

- Push after each fix batch.

3) Continue loop until completion criteria are met.

Reporting format each loop:

- Timestamp

- 👀 present? (yes/no)

- 👍 present? (yes/no)

- New Codex comments found

- Actions taken (fixes, replies, resolved threads)

- Commits pushed (hashes)

Final output:

- “Codex review loop complete”

- PR URL

- Final unresolved Codex thread count (must be 0)

- Last commits pushed

- `git status --porcelain=v1` output 

@etraut-openai any update on when we can expect this to land? It would be nice to have this configurable code review option.

Yes, running every time to find new P1s is insane. Especially when the first attempt displays P1 and P2 (2 findings), but the second attempt does it again. Why not use P1 + P1 the first time?

Why we can't just get it to find ALL issues in the single run?

1 - 2 findings at a time will never be useful to any of my processes. in the real world, who does a Pr review, finds several P1 findings, and only reports 1-2 of them, saving the rest for after the author fixes one issue? It's problematic for many reasons. I've been forced to remove codex from my pipeline and rely on other tools.

+1

I’ve been stuck in a push/review loop for hours. So frustrating 😭

Does anybody how under what conditions Codex limits the findings it returns? I'm worried about this issue, but I can't reproduce it.