codex-review

name	codex-review
description	Independent adversarial code review via Codex CLI to eliminate Claude self-review bias

Skill: codex-review

Trigger: called by front-agent after reviewer PASS Purpose: Run an independent review using Codex (OpenAI) to catch issues Claude's self-review may have missed.

---

Why This Exists

Claude reviews code it just wrote. Same model, same biases, same blind spots. Codex is a different model with no shared context — it will catch different things.

---

How To Run

Build the review prompt from two sources:

Field	Source	Include
Task goal	plan.md → ## Goal	Always — one-line summary
Previous reviewer notes	reviewer output → ### Notes	Only when present
constraints.md	—	Never — Codex judges independently
Full file contents	—	Never — git diff is sufficient

Scoped diff — only the current task's files

--uncommitted and --base main both review ALL changes in the repo, not just the current task. Instead, generate a diff scoped to changed_files from the handoff and pass it directly to codex:

# 1. Generate diff for only the files changed in this task
#    Route through rtk-wrap so `rtk git diff` compresses output when the mode is on.
DIFF=$(bash hooks/rtk-wrap.sh git diff HEAD -- src/features/login/LoginForm.tsx src/features/login/LoginForm.test.tsx)

# 2. Pass scoped diff inline to codex
codex -m o3 "
Task: [plan.md Goal in one line]
Previous reviewer (Claude opus) PASS notes: [reviewer Notes section, or 'none']

Files changed in this task:
- src/features/login/LoginForm.tsx
- src/features/login/LoginForm.test.tsx

Diff:
\`\`\`diff
$DIFF
\`\`\`

Review independently. Be adversarial. Focus on: correctness, security, type safety, and edge cases. Look for what the previous reviewer missed."

If o3 is unavailable or rate-limited, omit -m to fall back to the user's default model.

No git repository or empty diff: If the working directory is not a git repo, or changed_files is empty, skip codex-review and warn the orchestrator — do not block the workflow.

---

Verdict Rules

FAIL if any of the following:

• Security vulnerability
• Incorrect logic or broken edge case
• Type unsafety that could cause a runtime error
• Behavior change not covered by existing tests

PASS for everything else. Style preferences, minor suggestions, and non-blocking observations do not cause FAIL.

---

On FAIL

Do not auto-fix. Surface the issues to the user with a clear decision prompt:

"Codex review found blocking issues:

1. [file:line] — [issue]

Options: a) Fix — send issues to the implementation agent and re-review b) Override — proceed to git-commit as-is"

If user chooses Fix

1. Send the codex issue list to the implementation agent (developer / ui-builder)
2. Re-run reviewer
3. Re-run codex-review (one attempt only)
4. If still FAIL → create a git-issue and report to user — do not retry again

If user chooses Override

Proceed to git-commit as-is. Record the override decision in the handoff block:

- decisions: codex-review FAIL overridden by user — [issue summary]

---

Output Format

## Codex Review Verdict: PASS / FAIL

### Blocking Issues (if FAIL)
1. [file:line] — [issue] → [suggested fix]

### Notes (non-blocking)
- [observation]

---

Constraints

• Run only after reviewer (opus) has already given PASS
• Always scope diff to changed_files from the handoff — never use --uncommitted or --base main (those include unrelated changes)
• Pass only task goal + previous reviewer notes + scoped diff — never pass constraints.md or full file contents
• If codex is not installed, not authenticated, or the working directory is not a git repo: skip and warn the orchestrator — do not block the workflow
• On FAIL: surface to user — do not auto-fix without user approval
• On PASS or user override: signal orchestrator to proceed to git-commit
• Max one fix-retry cycle per task
• Route git diff through bash hooks/rtk-wrap.sh git diff ... (see CLAUDE.md → RTK Wrapping)