codex-review

name	codex-review
description	Cross-model second opinion from OpenAI Codex on plans or code. Auto-detects plan vs code review, iterates up to 5 rounds. Use for architecture decisions, non-trivial refactors, critical config changes. Triggers: '/codex-review', 'codex review', 'second opinion', 'cross-check', '让 codex 看看', '交叉审查'. Not for trivial fixes or formatting.
user_invocable	true

Codex Review

A cross-model review skill that sends your plan or code diff to OpenAI Codex for independent verification. The value of cross-model review is that different models have different blind spots — Codex may catch issues that Claude misses, and vice versa.

The skill operates in two modes (auto-detected):

• Plan Review: when you have an implementation plan, Codex challenges it for completeness, risks, and alternatives
• Code Review: when you have a git diff, Codex inspects the actual code for bugs, security issues, and edge cases

In both modes, Claude doesn't just relay feedback — it actively revises the plan or proposes fixes based on Codex's input, then re-submits for re-review. This continues until Codex approves or 5 rounds are reached.

Prerequisites

• codex CLI installed and on PATH (npm install -g @openai/codex)
• OpenAI credentials configured (API key or ChatGPT login)

Verify with: codex --version

Proxy Setup

Codex CLI requires proxy to reach OpenAI API. Set these environment variables before every codex exec call:

export https_proxy=http://127.0.0.1:7890 http_proxy=http://127.0.0.1:7890 all_proxy=socks5://127.0.0.1:7890

Prepend this export to all codex exec commands in this skill (use && to chain).

Step 1: Detect Review Mode

Check what's available, in this priority order:

1. Plan in context → Plan Review mode
2. Staged changes (git diff --cached --stat) → Code Review mode
3. Unstaged changes (git diff --stat) → Code Review mode
4. Nothing found → ask the user what they'd like reviewed

Tell the user which mode was detected:

"Detected [staged diff / plan in context]. Running Code Review / Plan Review mode."

If the user passed arguments (e.g., /codex-review gpt-5.5), parse the model name and use it instead of the default.

Step 2: Prepare the Review Package

Generate a session-scoped ID to avoid conflicts if multiple reviews run concurrently:

REVIEW_ID=$(uuidgen | tr '[:upper:]' '[:lower:]' | head -c 8)

Write the review content to /tmp/codex-review-input-${REVIEW_ID}.md:

For Plan Review:

# Plan Review Request

## Project Context
[2-3 lines from CLAUDE.md describing what the project does]

## Implementation Plan
[Full plan content from conversation context]

For Code Review:

# Code Review Request

## Project Context
[2-3 lines from CLAUDE.md describing what the project does]

## Summary
[1-2 sentence summary of what these changes do and why]

## Diff
[Output of git diff --cached, or git diff if nothing staged]

For code review, also note the file count and diff size. If the diff exceeds ~3000 lines, warn the user that Codex may not review everything thoroughly and suggest reviewing in smaller chunks.

Step 3: Send to Codex (Round 1)

The review prompt differs by mode because plans and code need different lenses.

Plan Review Prompt

export https_proxy=http://127.0.0.1:7890 http_proxy=http://127.0.0.1:7890 all_proxy=socks5://127.0.0.1:7890 && \
codex exec \
  -m gpt-5.5 \
  -s read-only \
  -c 'model_reasoning_effort="xhigh"' \
  -o /tmp/codex-review-output-${REVIEW_ID}.md \
  "Review the implementation plan in /tmp/codex-review-input-${REVIEW_ID}.md.

Evaluate along these dimensions:
1. GOAL ALIGNMENT - Does the plan actually solve the stated problem? Are there simpler alternatives?
2. COMPLETENESS - Are any steps missing? What about rollback, error handling, migration?
3. RISK ASSESSMENT - What could go wrong? Data loss? Breaking changes? Race conditions?
4. ORDERING & DEPENDENCIES - Are the steps in the right order? Any implicit dependencies?
5. TESTABILITY - How will we know this worked? What should be tested?

For each issue found, be specific: name the step number, explain what's wrong, and suggest a fix.
Skip anything that looks fine — only flag real problems.

End your review with exactly one of:
VERDICT: APPROVED (if the plan is solid and ready to implement)
VERDICT: REVISE (if changes are needed — list what needs to change)" 2>/dev/null

Code Review Prompt

export https_proxy=http://127.0.0.1:7890 http_proxy=http://127.0.0.1:7890 all_proxy=socks5://127.0.0.1:7890 && \
codex exec \
  -m gpt-5.5 \
  -s read-only \
  -c 'model_reasoning_effort="xhigh"' \
  -o /tmp/codex-review-output-${REVIEW_ID}.md \
  "Review the code changes in /tmp/codex-review-input-${REVIEW_ID}.md.

Evaluate along these dimensions:
1. CORRECTNESS - Does the code do what it claims? Off-by-one errors, wrong comparisons, missing null checks?
2. SECURITY - Injection risks, auth bypasses, secrets in code, unsafe deserialization?
3. EDGE CASES - What inputs or states would break this? Empty collections, concurrent access, large inputs?
4. ERROR HANDLING - Are errors caught and handled appropriately? Can failures cascade?
5. COMPATIBILITY - Does this break existing callers, APIs, or data formats?
6. TEST COVERAGE - Are the changes tested? What test cases are missing?

For each issue, reference the specific file and code snippet. Suggest a concrete fix, not just 'consider handling this'.
Skip anything that looks fine — only flag real problems. No praise, no filler.

End your review with exactly one of:
VERDICT: APPROVED (if the code is solid and ready to merge)
VERDICT: REVISE (if changes are needed — list what needs to change)" 2>/dev/null

Capture the Codex session ID from stdout (the line containing session id: <uuid>). Store as CODEX_SESSION_ID for subsequent rounds.

CLI notes:

• Default model: gpt-5.5 (thorough, good balance of quality and speed).
• Always use -s read-only — Codex should never modify files.
• 2>/dev/null suppresses thinking tokens that would bloat context.

Step 4: Process the Verdict

Read /tmp/codex-review-output-${REVIEW_ID}.md and present to the user:

## Codex Review — Round N [Plan/Code] (gpt-5.5)

[Codex's feedback, preserving its structure]

**Verdict: APPROVED / REVISE**

Then branch:

• APPROVED → Step 7 (done)
• REVISE → Step 5
• No clear verdict but only positive comments → treat as APPROVED
• Round 5 reached → Step 7 with unresolved concerns listed

Step 5: Revise Based on Feedback

This is where the skill earns its value — Claude doesn't just relay Codex's feedback, but thinks about it and acts.

For Plan Review:

• Update the plan to address each issue Codex raised
• Rewrite /tmp/codex-review-input-${REVIEW_ID}.md with the revised plan

For Code Review:

• Don't modify files directly (the user hasn't approved yet)
• Instead, write a proposed fix list into the temp file explaining what you'd change and why
• If Codex flagged a false positive (e.g., the code is correct but Codex misread it), note that and explain why you're skipping it

In both cases, summarize for the user:

### Revisions (Round N)
- [What changed and why, one bullet per issue]
- [Skipped: issue X — reason why it's not applicable]

If a suggested revision contradicts the user's explicit requirements, skip it and explain why.

Step 6: Re-submit to Codex (Rounds 2-5)

Resume the Codex session so it retains context from prior rounds:

export https_proxy=http://127.0.0.1:7890 http_proxy=http://127.0.0.1:7890 all_proxy=socks5://127.0.0.1:7890 && \
codex exec resume ${CODEX_SESSION_ID} \
  "I've addressed your feedback. The updated content is in /tmp/codex-review-input-${REVIEW_ID}.md.

Changes made:
[Bullet list of what was revised]

Skipped (with rationale):
[Any items intentionally not addressed]

Please re-review. End with VERDICT: APPROVED or VERDICT: REVISE" 2>&1 | tail -80

codex exec resume doesn't support -o, so read output from stdout (piped through tail to skip startup noise).

If resume fails (session expired), fall back to a fresh codex exec with a note summarizing prior rounds.

Return to Step 4.

Step 7: Present Final Result

If approved:

## Codex Review — Final [Plan/Code] (gpt-5.5)

**Status:** APPROVED after N round(s)

[Final Codex feedback]

---
Ready for your approval to [implement / merge].

For code review, if there were proposed fixes along the way, list them:

**Proposed fixes from review:**
- [ ] file.py:42 — add null check for `user_input`
- [ ] config.yaml:15 — remove hardcoded timeout value

If max rounds reached:

## Codex Review — Final [Plan/Code] (gpt-5.5)

**Status:** 5 rounds reached — not fully approved

**Remaining concerns:**
[Unresolved issues from last review]

---
Review these remaining items and decide whether to proceed or keep refining.

Step 8: Cleanup

rm -f /tmp/codex-review-input-${REVIEW_ID}.md /tmp/codex-review-output-${REVIEW_ID}.md

Ground Rules

• Claude actively revises based on feedback — not just a messenger between the user and Codex
• Codex runs in read-only mode, always. It reviews but never writes files
• Max 5 rounds prevents infinite loops. If 3+ rounds happen on the same issue, surface it to the user as a judgment call
• Show every round to the user so they can follow the conversation and intervene
• In code review mode, don't apply fixes until the user approves — just propose them
• If Codex CLI isn't installed or fails, tell the user: npm install -g @openai/codex
• Default model: gpt-5.5. User can override via arguments (e.g., /codex-review gpt-5.3-codex)
• UUID-scoped temp files support concurrent sessions safely