| name | codex-review |
| disable-model-invocation | true |
| description | Iterative code review using OpenAI Codex CLI and Claude working together. Codex reviews the code, Claude applies fixes, and the loop repeats until no issues remain. |
Codex Review
Iterative code review loop: Codex reviews code → Claude fixes issues → Codex re-reviews → repeat until clean.
Workflow
1. Determine review scope
Ask the user if not obvious from context. Support these modes:
- Staged changes:
--uncommitted
- Branch diff:
--base <branch> (default base: main)
- Specific commit:
--commit <sha>
If the user says "review my changes" without specifics, check git status and git log --oneline -5 to infer the best mode. Prefer --base with the default branch if on a feature branch, --uncommitted if there are uncommitted changes.
2. Run Codex review
Use codex exec review (not codex review) to run non-interactively with clean output:
codex exec review <scope-flags> -o .codex-review-output.md
The -o flag writes only the final review message to the file, excluding all progress/status noise. Read .codex-review-output.md after the command completes.
If the user wants to use a specific model, pass -m <model> (e.g., -m o3).
Delete .codex-review-output.md after reading it.
3. Analyze Codex findings
Read the review output. Categorize each finding:
- Actionable: Clear bugs, security issues, logic errors, missing error handling, performance problems — fix these.
- Stylistic/subjective: Naming preferences, minor formatting — skip unless the user explicitly wants these fixed too.
- False positives: Findings that are incorrect or don't apply — skip.
Present a summary to the user:
Codex found N issues (X actionable, Y skipped).
Fixing: [brief list of what will be fixed]
Skipping: [brief list of what was skipped and why]
Wait for user confirmation before applying fixes. The user may override which findings to fix or skip.
4. Apply fixes
Fix each actionable issue. For each fix:
- Read the relevant source file to understand full context
- Apply the minimal change needed
- Do not refactor surrounding code unless the finding specifically calls for it
Stage all fixes after applying them: git add <changed files>.
5. Re-review
Run Codex review again on the new changes using the same scope/flags as step 2.
Read the output. If Codex finds new issues:
- Summarize them to the user
- After user confirmation, apply fixes and re-review (go to step 4)
If Codex reports no issues or only has minor stylistic notes, the loop is complete.
6. Finish
Report the final status:
Review complete. N rounds, M total fixes applied.
List all files modified. Do not commit — leave that to the user.
Notes
- Maximum 5 review rounds to avoid infinite loops. If issues persist after 5 rounds, report remaining findings and stop.
- If
codex CLI is not installed, tell the user to install it: npm install -g @openai/codex
- If Codex returns an error (auth, network, etc.), report it clearly and stop.
- Preserve the user's code style. Match existing patterns in the codebase.
