Codex Commit Review
Purpose
Use this skill to debate commit message quality before or after committing. Claude and Codex are equal analytical peers — Claude orchestrates the debate loop and final synthesis. No commit messages are modified.
Prerequisites
- Draft mode: user provides draft commit message text. Staged changes available for alignment check.
- Last mode: recent commits exist (
git log -n N). Repository has commit history. codexCLI is installed and authenticated.codex-reviewskill pack is installed (npx github:lploc94/codex_skill).
Runner
RUNNER="{{RUNNER_PATH}}"Workflow
- Ask user to choose review effort level:
low,medium,high, orxhigh(default:medium). Ask input source:draft(user provides message text) orlast(review last N commits, default 1). SetEFFORTandMODE. - Run pre-flight checks (see
references/workflow.md§1.5). - Build Codex prompt + Claude analysis prompt from
references/prompts.md, following the Placeholder Injection Guide. Start Codex (background) withnode "$RUNNER" start. - Claude Independent Analysis (BEFORE reading Codex output): Claude analyzes commit message(s) independently using format from
references/claude-analysis-template.md. INFORMATION BARRIER — do NOT read$STATE_DIR/review.mduntil analysis is complete. Seereferences/workflow.mdStep 2.5. - Poll Codex with adaptive intervals (Round 1: 60s/60s/30s/15s..., Round 2+: 30s/15s...). After each poll, report specific activities from poll output. See
references/workflow.mdfor parsing guide. NEVER report generic "Codex is running" — always extract concrete details. - Cross-Analysis: Compare Claude's FINDING-{N} with Codex's ISSUE-{N}. Identify genuine agreements, genuine disagreements, and unique findings from each side. See
references/workflow.mdStep 4. - Resume debate via
--thread-iduntil consensus, stalemate, or hard cap (5 rounds). - Present final consensus report with agreements, disagreements, and both sides' overall assessments. NEVER propose revised commit messages.
- Cleanup:
node "$RUNNER" stop "$STATE_DIR".
Effort Level Guide
| Level | Depth | Best for | Typical time |
|---|---|---|---|
low | Surface check | Quick sanity check | ~1-2 min |
medium | Standard review | Most day-to-day work | ~3-5 min |
high | Deep analysis | Important features | ~5-10 min |
xhigh | Exhaustive | Critical/security-sensitive | ~10-15 min |
Required References
- Detailed execution:
references/workflow.md - Prompt templates:
references/prompts.md - Output contract:
references/output-format.md - Claude analysis format:
references/claude-analysis-template.md
Rules
- Safety: NEVER run
git commit --amend,git rebase, or any command that modifies commit history. This skill is debate-only. - Both Claude and Codex are equal peers — no reviewer/implementer framing.
- Information barrier: Claude MUST complete independent analysis (Step 2.5) before reading Codex output. This prevents anchoring bias.
- NEVER propose revised commit messages — only debate quality. The final output is a consensus report, not a fix.
- Codex reviews message quality only; it does not review code.
- Discover project conventions before reviewing (see
references/workflow.md§1.6). - For
lastmode with N > 1: findings must reference specific commit SHA/subject in Evidence. - If stalemate persists (same unresolved points for 2 consecutive rounds), present both sides and defer to user.