---

name: codex-review-loop description: On a GitHub PR, post '@codex review' via gh, wait for Codex’s review, then iterate fix+push cycles until Codex approves / reports no P0/P1 findings. user-invocable: true context: fork

Codex Review Loop (gh + @codex)

What this Skill does

When invoked, this Skill drives a tight PR iteration loop:

1. Identify the target PR (explicit #123/URL, or the PR for your current branch)
2. Post a PR comment containing @codex review (so the comment is authored by your GitHub account via gh)
3. Poll GitHub until Codex posts its review
4. Convert the findings into an actionable checklist
5. Apply fixes, run tests, commit, push
6. Repeat until Codex returns APPROVED or explicitly reports no findings

Codex is triggered by a PR comment containing @codex review (not by “requesting reviewer”). See OpenAI’s GitHub integration docs.

Preconditions

• You are inside a git repo that has a PR (or you pass a PR number/URL).
• gh auth status works (you are logged in as the account that should author the comment).
• Codex GitHub code review is enabled for the repo (in Codex settings).
• Recommended: add an AGENTS.md with “Review guidelines” so Codex focuses on what you care about.

Quick usage

Run on current branch’s PR

Run:

Copypython3 .claude/skills/codex-review-loop/scripts/codex_review_once.py

Run on a specific PR

Copypython3 .claude/skills/codex-review-loop/scripts/codex_review_once.py 123
# or:
python3 .claude/skills/codex-review-loop/scripts/codex_review_once.py https://github.com/OWNER/REPO/pull/123

How to use this Skill in Claude Code

When I’m active, follow this loop:

1. Fetch Codex feedback:

   • Run codex_review_once.py (optionally with PR ref)
   • Parse its JSON output

2. If clean, stop:

   • Clean = review.state == "APPROVED" OR (inline_comments empty AND review body indicates no findings)

3. If findings exist, fix them:

   • Make minimal, surgical changes addressing each comment
   • Prefer adding/adjusting tests when behavior changes

4. Run tests (auto-detect):

   • If package.json exists: npm test (or pnpm test / yarn test if lockfile suggests)
   • If go.mod: go test ./...
   • If pyproject.toml/requirements.txt: pytest -q
   • If CMakeLists.txt: cmake --build ... then ctest
   • If none detected: run the repo’s documented test command in README / CI config

5. Commit + push:

   • Commit message: Fix: address Codex review findings (iter N)
   • Push to the PR branch

6. Go back to step (1).

Guardrails

• Don’t do large refactors unless Codex finds a real bug.
• Keep iteration count capped (default 5). If still not clean, summarize remaining deltas.

Optional: teach Codex what “issues” mean (AGENTS.md)

Codex reads AGENTS.md and follows “Review guidelines”. Put this at repo root:

Copy## Review guidelines
- Treat security/auth issues as P0.
- Treat data races / memory safety issues as P0.
- Treat flaky tests as P1.
- Treat doc typos as P2 (do not block).

Blockers-only mode (recommended default)

This loop is most useful when it stops on blockers only.

Definition of “clean” in blockers-only mode:

• Clean if there are no P0/P1 inline comments from Codex (even if Codex summarizes P2 suggestions in the review body).
• Not clean if there are any inline comments labeled P0 or P1, or if Codex submits a review with state CHANGES_REQUESTED.

Getting reliable severity labels

Set CODEX_EXTRA_PROMPT so Codex labels findings and keeps P2 non-blocking:

Copyexport CODEX_EXTRA_PROMPT="Please label every finding as P0/P1/P2. Only leave inline review comments for P0/P1 blockers. Summarize P2 (nits) in the review body under a 'Non-blocking suggestions' section."

Run blockers-only (default)

Copypython3 .claude/skills/codex-review-loop/scripts/codex_review_once.py --blockers-only
# (or omit the flag; blockers-only is default)

Run strict (rare)

Copypython3 .claude/skills/codex-review-loop/scripts/codex_review_once.py --strict