Codex Code Review Skill (CC Edition)

Battle-tested edition with CLI pitfall avoidance, network fallback strategy, and real-world case patterns.

This skill delegates analysis to the Codex CLI, which is powered by OpenAI's models. The whole point is to get a second opinion from a different AI — if you (Claude) do the analysis yourself, the user gets zero additional value.

Two Modes: Change Review vs Focused Analysis

This skill has two distinct modes. Determine the mode FIRST before doing anything else.

How to determine mode

• If user's request is about changes/diff (e.g. "review 我的改动", "审一下代码", "看看有没有 bug") → 变更审查
• If user's request has a specific analysis focus unrelated to changes (e.g. "梳理核心功能", "review 测试代码找冗余", "分析架构", "看看安全性") → 定向审查
• If unclear, use AskUserQuestion:

AskUserQuestion:
  question: "选择审查模式："
  header: "Review mode"
  options:
    - label: "变更审查"
      description: "审查当前未提交的代码变更，找 bug、回归、代码质量问题 (codex review)"
    - label: "定向审查"
      description: "针对特定角度分析代码库，比如测试质量、架构梳理、安全性 (codex exec)"

If the user's original request already clearly indicates one mode, skip the question and go directly to the corresponding path.

CLI Pitfall Guide (Must Read)

Key rule: codex review is a direct subcommand — no need for exec wrapper.

Path A: 变更审查 (Change Review)

For reviewing uncommitted changes, specific commits, or PRs.

A1. Pre-flight checks

Run in parallel:

git diff --name-only && git status --short

git diff --stat | tail -1

• If there are untracked files (?? in status), stage them first:

  git ls-files --others --exclude-standard -z | while IFS= read -r -d '' f; do git add -- "$f"; done
  

• Clean working directory → use codex review --commit HEAD instead

A2. Optionally update CHANGELOG

If the project has a CHANGELOG.md and the user's changes are significant, check if it's already in the diff:

git diff --name-only | grep -iE "changelog"

If not updated, suggest adding an entry under ## [Unreleased]. Skip this step if the project doesn't use a CHANGELOG or the changes are minor.

A3. Assess difficulty & run

Difficulty rules (ANY triggers xhigh):

• Files >= 10
• Total changes (insertions + deletions) >= 500
• Insertions alone >= 300 OR deletions alone >= 300

codex review --uncommitted --config model_reasoning_effort=high

For clean working directories: codex review --commit HEAD --config ...

Path B: 定向审查 (Focused Analysis)

For codebase analysis with a specific focus. NO pre-flight checks needed — go directly to composing and running the prompt.

B1. Compose prompt

Build the prompt from three parts:

1. Objective — what the user wants analyzed, in their words
2. Scope — which files to focus on (derive from user's request)
3. Output format — structured findings (P1/P2/P3 for reviews, or custom structure for analysis)

Prompt template for code quality review:

Review the codebase with this specific focus:

OBJECTIVE: {user's concern}
SCOPE: {file pattern}

INSTRUCTIONS:
- Read the relevant source files to understand the implementation
- Evaluate against the objective
- Be thorough and specific

Report findings as:
- P1 (Critical): ...
- P2 (Important): ...
- P3 (Minor): ...

For each finding: file path, what's wrong, and suggested fix.

Prompt template for architecture/feature analysis:

Analyze the codebase with this focus:

OBJECTIVE: {user's concern, e.g. "Map out all core features and end-to-end user flows"}
SCOPE: {key files to read}

INSTRUCTIONS:
- Read each file listed in scope
- {specific analysis instructions}

Output in {language}. Use structured format with clear headings.

Adapt the template to fit the user's actual request.

B2. Choose effort level

For focused analysis, effort level is based on scope size (not diff size):

B3. Run

codex exec --full-auto "YOUR_COMPOSED_PROMPT" --config model_reasoning_effort=high

Interpret Output & Act

Priority levels (for review-type output):

Do NOT blindly follow all suggestions. Evaluate each:

• Is this a real bug or a false positive?
• Does the suggested fix introduce new complexity?
• Is this relevant to the current scope?

After fixing P1/P2 items: re-run tests to confirm fixes don't break anything.

Network Stability & Fallback

Codex CLI can fail with Transport error: network error during long sessions.

IF codex hangs or returns network error:
  1. DO NOT retry more than once
  2. Tell the user codex failed and you're falling back to manual analysis
  3. Read the relevant files yourself and perform the analysis inline
  4. Report findings in the same format

Always inform the user which path was taken.

Codex CLI Reference

Review Mode

codex review [OPTIONS]

Exec Mode (for focused analysis)

codex exec --full-auto "your prompt here" --config model_reasoning_effort=high

Constraints:

• --uncommitted, --base, --commit are mutually exclusive
• Must run inside a git repository