AgentSkillsCN

codex-review

以无头代码审查工具的身份运行 OpenAI Codex CLI，并生成持久化的文件输出。当您需要从 Codex 获得第二意见的代码审查、希望以非交互方式运行 Codex exec 进行代码分析，或需要在沙盒环境中借助 GPT 助力完成审查并保存相关工件时，此功能尤为适用。该功能可在 Claude Code 会话内顺畅运行（不会产生嵌套冲突）。触发条件包括：“codex 审阅”、“对某段代码运行 Codex”、“从 Codex 获取第二意见”、“Codex 检查”。

SKILL.md
--- frontmatter
name: codex-review
description: Run OpenAI Codex CLI as a headless code reviewer with durable file output. Use when you need a second-opinion review from Codex, want to run codex exec non-interactively for code analysis, or need sandboxed GPT-powered review with saved artifacts. Works inside Claude Code sessions (no nesting conflict). Triggers on "codex review", "run codex on", "get a second opinion from codex", "codex check".
todo:
  - "Fix example paths in docs — should use full .claude/skills/codex-review/ prefix"
  - "Output location should respect --workdir when set, not just caller pwd"
  - "Add passthrough for --output-schema, --output-last-message, --profile"
  - "Clarify stderr handling rationale — thinking tokens vs error diagnostics"

codex-review

Run codex exec headless as a code reviewer with timestamped artifact output.

Defaults

  • •Sandbox: read-only (safe by default)
  • •Model: inherits from ~/.codex/config.toml
  • •Effort: inherits from config
  • •File mode: stderr captured to output file (contains thinking tokens + errors)
  • •No-file mode: stderr passes through (errors visible)
  • •Output: cache/codex-reviews/CODEX_REVIEW_<tag>_<timestamp>.txt

Command pattern

bash
# Basic review (read-only, default model)
.claude/skills/codex-review/scripts/codex_review.sh \
  --prompt "Review app/wibwob_view.cpp for memory leaks" --tag mem

# High-effort security audit
.claude/skills/codex-review/scripts/codex_review.sh \
  --prompt "Security audit of tools/api_server/main.py" \
  --effort xhigh --tag security

# Allow edits (workspace-write sandbox)
.claude/skills/codex-review/scripts/codex_review.sh \
  --prompt "Refactor error handling in app/api_ipc.cpp" \
  --sandbox workspace-write --tag refactor

# JSON Lines output for machine parsing
.claude/skills/codex-review/scripts/codex_review.sh \
  --prompt "List all TODO comments with severity" \
  --json --tag todos

Flags

  • •--prompt <text> review prompt (required)
  • •--tag <slug> output filename tag (default: review, alphanumeric/dash/underscore/dot only)
  • •--out <path> explicit output file path
  • •--no-file stream to stdout only
  • •--model <name> override codex model
  • •--effort <level> reasoning: low | medium | high | xhigh
  • •--sandbox <mode> read-only | workspace-write | danger-full-access
  • •--confirm-danger required alongside --sandbox danger-full-access
  • •--workdir <path> working directory for codex
  • •--json JSON Lines output format
  • •--skip-git skip git repository check

Sandbox modes

ModeUse caseGate
read-only (default)Code review, analysis, auditsNone
workspace-writeRefactoring, applying fixesNone
danger-full-accessCI/CD onlyRequires --confirm-danger flag

After running

  1. •Read the output file and summarize key findings
  2. •Verify suggestions before applying
  3. •Only use workspace-write when edits are explicitly requested

Guardrails

  • •Default to read-only unless user explicitly requests writes
  • •danger-full-access rejected without --confirm-danger flag
  • •Always write raw output for debugging (default behavior)
  • •Preflight checks: codex CLI must be in PATH, tag must be safe characters
<todo> From codex self-review (cache/codex-reviews/CODEX_REVIEW_self-review_20260215_183658.txt): 4. Example paths in docs should use full .claude/skills/codex-review/ prefix from repo root 5. Output location uses caller pwd, not --workdir — consider aligning these 6. No passthrough for --output-schema, --output-last-message, --add-dir, --profile 7. Stderr handling rationale could be clearer — file mode captures everything, no-file mode passes through </todo>