Codex Review

Codex Review

Use Codex's dedicated review command as an independent fallback check. This is codex review, not OpenClaw autoreview, not Guardian approval auto_review, and not a generic "ask a subagent to review" pass.

For plain "autoreview" or normal closeout review requests, use the autoreview skill instead. Use this skill only when the user explicitly asks for built-in Codex review, when comparing review tools, or when autoreview is unavailable.

Use when:

• the user asks specifically for built-in Codex review or codex review
• comparing codex review against autoreview
• autoreview fails or is unavailable and a fallback review is still useful
• reviewing a local dirty tree, branch, commit, or PR branch after fixes

Contract

• Treat review output as advisory. Verify every finding against the real code path before changing code.
• Prefer small fixes at the right ownership boundary. Do not add broad refactors just to satisfy a speculative finding.
• Reject findings that depend on unstated assumptions, unrealistic edge cases, intentional behavior, or over-complicated fixes.
• If a review finding causes a code change, rerun the focused proof/tests and rerun Codex review.
• Keep looping until the final review has no accepted/actionable findings, or until a remaining finding is explicitly rejected with a short reason.
• Do not push only to run review. Push only when the user asked for push, ship, or PR update.

Pick The Target

Dirty local work:

codex review --uncommitted

Use this only when the working tree actually has staged, unstaged, or untracked changes. A clean --uncommitted run only proves there is no local patch.

Branch or PR work:

git fetch origin
codex review --base origin/main

If an open PR exists, review against its actual base:

base=$(gh pr view --json baseRefName --jq .baseRefName)
codex review --base "origin/$base"

Committed single change:

codex review --commit HEAD

Custom review prompt:

codex review "Review the error handling in the new auth flow."

Current Codex CLI targets are mutually exclusive. Do not combine --base, --commit, or --uncommitted with a positional custom prompt.

Helper

This skill includes a helper that chooses a sensible target and can run focused tests in parallel:

~/.agents/skills/codex-review/scripts/codex-review --help

After sync, Claude can call the same helper here:

~/.claude/skills/codex-review/scripts/codex-review --help

The helper:

• chooses dirty --uncommitted first in --mode auto
• otherwise uses the current PR base if gh pr view works
• otherwise uses origin/main for non-main branches
• supports --mode local, --mode branch, --base, --commit, --dry-run, and --parallel-tests
• supports --full-access for nested Codex review runs that need the local command to bypass Codex sandbox prompts
• prints codex-review clean: no accepted/actionable findings reported when the selected review command exits 0 and no priority findings are detected

Format first if formatting can move line numbers, then it is OK to run tests and review together:

~/.agents/skills/codex-review/scripts/codex-review --parallel-tests "npm test -- --runInBand"

If tests or review lead to edits, rerun the affected tests and rerun review. Once the final review/helper run exits cleanly, stop. Do not run an extra review just to get nicer wording.

Subagent Filter

codex review is already a dedicated review harness with a review-specific rubric, constrained tools, fresh context, and structured findings. A normal subagent reviewer is more flexible but easier to bias with the main agent's framing.

For larger diffs, use a subagent as a filter over codex review when available. Ask it to run the review and return only:

• findings it accepts as actionable
• findings it rejects, with one-line reasons
• exact files and tests to rerun

For tiny changes, run the review inline.

Final Report

Include:

• review command used
• tests or proof run
• accepted findings and fixes
• rejected findings with short reasons
• final clean review result, or why any remaining finding was consciously rejected

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: minghinmatthewlam
• Source: minghinmatthewlam/agent-guards
• License: MIT

Install and usage instructions live in the source repository linked above.