By atompilot
Multi-model code review for Claude Code — Claude writes, Codex reviews, loop until clean, then commit.
npx claudepluginhub atompilot/claude-code-cross-review --plugin claude-code-cross-reviewThe writer doesn't review. The reviewer doesn't write.
A multi-model code review workflow for Claude Code. Claude (Opus) writes the code, Codex (GPT) reviews it, Claude fixes what Codex finds — looping until clean, then commits.
写的不审，审的不写。
Claude Code 的多模型代码审查工作流。Claude (Opus) 写代码，Codex (GPT) 审代码，Claude 修复 Codex 发现的问题，循环直到干净，再提交。
Code written / 代码写完
↓
[Codex] Review uncommitted changes / 审查未提交改动
↓
Issues found? / 有问题？
Yes ──→ [Claude] Fix each issue / 逐项修复 ──→ Back to review / 回到审查
No ↓
lint / build / test
↓
git commit
| Role / 角色 | Model / 模型 | Responsibility / 职责 |
|---|---|---|
| Write / 写代码 | Claude Opus 4.6 | Feature dev, bug fixes, refactoring / 功能开发、Bug 修复、重构 |
| Review / 审代码 | Codex (GPT-5.3) | Edge cases, logic bugs, type safety / 边界条件、逻辑漏洞、类型安全 |
Codex is integrated via MCP protocol — no window switching required. Falls back to codex CLI if MCP is unavailable.
Codex 通过 MCP 协议集成，Claude 无需切换窗口即可调用审查。若 MCP 不可用，自动回退到 codex CLI。
Install OpenAI Codex CLI / 安装 Codex CLI：
npm install -g @openai/codex
# or / 或
brew install codex
Verify / 验证：
codex --version
git clone https://github.com/splazapp/claude-code-cross-review.git
cd claude-code-cross-review
bash install.sh
This installs / 安装内容：
~/.claude/skills/cross-review/SKILL.md — workflow skill / 工作流技能~/.claude/commands/codex-review.md — /codex-review slash command / 快捷命令~/.claude/.mcp.json — registers Codex as an MCP server / 注册 Codex 为 MCP 服务器Restart Claude Code after installation. / 安装完成后重启 Claude Code。
Option 1 — Trigger the skill / 触发技能：
Tell Claude in Claude Code / 在 Claude Code 中告诉 Claude：
Use the cross-review skill to review current changes.
使用 cross-review 技能审查当前改动
Option 2 — Slash command / 使用命令：
/codex-review
bash uninstall.sh
Claude is great at writing code but tends to miss its own blind spots. Codex, as an independent model, reviews without bias toward the original implementation. The result: fewer bugs reach git commit.
Claude 擅长写代码，但容易对自己的盲点视而不见。Codex 作为独立模型，审查时不带对原实现的偏见。结果是：更少的 bug 进入 git commit。
v1.0.0
Codex-powered code review gate. Creates a review checkpoint that triggers Codex CLI review when Claude exits.
Share bugs, ideas, or general feedback.
Double-check layer for the Official Codex plugin — wraps every Codex interaction with Claude's independent critical evaluation. Review, adversarial, rescue, verify, research.
Command-driven Claude/Codex review workflows for multi-round plan and implementation review.
Claude Code skill to delegate prompts to OpenAI Codex CLI for code analysis, refactoring, code review, and automated editing
Code review plugin with a standalone reviewer agent and two skill strategies: disposable subagents for one-shot reviews and persistent team members for iterative reviews
v9.30.0 — Model defaults refreshed: Opus 4.7 for planning/strategy/security-review, GPT-5.4 for code-review/implementation. New GPT-5.4 prompting guide. Set OCTOPUS_LEGACY_ROLES=1 to opt out. Run /octo:setup.