codex-impl-review

Codex Implementation Review

Purpose

Use this skill to run adversarial review on uncommitted changes before commit, or on branch changes before merge.

Prerequisites

Runner

RUNNER="{{RUNNER_PATH}}"

Workflow

Ask user to choose review effort level: low , medium , high , or xhigh (default: high ). Ask review mode: working-tree (default) or branch . If branch mode, ask for base branch name and validate (see workflow.md for base branch discovery). Set EFFORT and MODE .
Run pre-flight checks (see references/workflow.md §1.5).
Build prompt from references/prompts.md (Working Tree or Branch Review Prompt), following the Placeholder Injection Guide.
Start round 1 with node "$RUNNER" start --working-dir "$PWD" --effort "$EFFORT" .
Poll with adaptive intervals (Round 1: 60s/60s/30s/15s..., Round 2+: 30s/15s...). After each poll, report specific activities from poll output (e.g. which files Codex is reading, what topic it is analyzing). See references/workflow.md for parsing guide. NEVER report generic "Codex is running" — always extract concrete details.
Parse issue list with references/output-format.md .
Fix valid issues in code; rebut invalid findings with evidence.
Resume debate via --thread-id until APPROVE , stalemate, or hard cap (5 rounds).
Return final review summary, residual risks, and recommended next steps.

Effort Level Guide

Level Depth Best for Typical time

low

Surface check Quick sanity check ~2-3 min

medium

Standard review Most day-to-day work ~5-8 min

high

Deep analysis Important features ~10-15 min

xhigh

Exhaustive Critical/security-sensitive ~20-30 min

Required References

Rules

If invoked during Claude Code plan mode, exit plan mode first — this skill requires code editing.
Codex reviews only; it does not edit files.
Preserve functional intent unless fix requires behavior change.
Every accepted issue must map to a concrete code diff.
If stalemate persists, present both sides and defer to user.

Source Transparency