AI Skill Hub 强烈推荐:Codex对抗评审 是一款优质的Agent工作流。AI 综合评分 8.0 分,在同类工具中表现稳健。如果你正在寻找可靠的Agent工作流解决方案,这是一个值得深入了解的选择。
Codex对抗评审 是一套完整的 AI Agent 自动化工作流方案。通过可视化的节点编排,将复杂的多步骤任务拆解为清晰的自动化流程,实现全程无人值守的智能处理。支持与数百种外部服务和 API 无缝集成,适合构建数据处理管线、业务自动化和 AI 辅助决策系统。
Codex对抗评审 是一套完整的 AI Agent 自动化工作流方案。通过可视化的节点编排,将复杂的多步骤任务拆解为清晰的自动化流程,实现全程无人值守的智能处理。支持与数百种外部服务和 API 无缝集成,适合构建数据处理管线、业务自动化和 AI 辅助决策系统。
# 方式一:npm 全局安装 npm install -g codex-adversary # 方式二:npx 直接运行(无需安装) npx codex-adversary --help # 方式三:项目依赖安装 npm install codex-adversary # 方式四:从源码运行 git clone https://github.com/todd866/codex-adversary cd codex-adversary npm install npm start
# 命令行使用
codex-adversary --help
# 基本用法
codex-adversary [options] <input>
# Node.js 代码中使用
const codex_adversary = require('codex-adversary');
const result = await codex_adversary.run(options);
console.log(result);
# codex-adversary 配置说明 # 查看配置选项 codex-adversary --config-example > config.yml # 常见配置项 # output_dir: ./output # log_level: info # workers: 4 # 环境变量(覆盖配置文件) export CODEX_ADVERSARY_CONFIG="/path/to/config.yml"
Recruit Codex as an automatic adversarial reviewer inside Claude Code.
When Claude runs a review or red-team pass, it usually reviews with itself. This setup makes Claude also recruit Codex — a second, independently-prompted model from a different vendor — on the same artifact, then reconcile the two. A second model catches mistakes a single reviewer's blind spots would miss: agreement raises confidence, and disagreement is where the lead looks harder.
It works on code (local git diffs and branch diffs) and on prose (a manuscript, a research claim, an argument, a design doc).
Codex plays two roles: an adversarial reviewer (retrospective — tear apart finished work) and an advisor (prospective — a second opinion on a decision before you act).
One honest caveat up front: Codex and Claude are both transformer LLMs trained on overlapping data, so this is correlated diversity, not independence. Agreement reduces stochastic misses; it does not clear shared blind spots. Treat agreement as raised confidence, never as verification — anything correctness/safety/money-critical still needs a non-LLM check (run the code, check the source, do the arithmetic). See Limitations.
- Claude Code - Codex CLI, installed and authenticated (codex login). It runs on whatever account/model your Codex CLI is configured with. Tested against Codex CLI 0.139.0. - bash 3.2+ on macOS or Linux (or WSL on Windows). Native Windows shells are not supported.
git clone https://github.com/todd866/codex-adversary
cd codex-adversary
./install.sh
The installer copies the wrapper, skill, and command into ~/.claude/, and appends the auto-trigger directive to ~/.claude/CLAUDE.md (idempotently — re-running upgrades the block in place). Start a new Claude Code session to pick up the skill. Remove everything later with ./uninstall.sh.
To install into a non-default location: CLAUDE_HOME=/path/to/.claude ./install.sh.
./uninstall.sh
Removes the installed files (only those carrying this project's stamp) and strips the directive block from ~/.claude/CLAUDE.md.
- Automatic: when Claude runs a substantive review/red-team pass, it recruits Codex and synthesizes. Nothing to type. - Explicit: /adversarial-review <file, diff, or claim>. - Standalone (bash):
# prose
cat draft.md | ~/.claude/bin/codex-adversary.sh --mode prose --effort xhigh \
--focus "scrutinise the statistical claims"
# code (uncommitted changes, or vs a base branch)
~/.claude/bin/codex-adversary.sh --mode diff
~/.claude/bin/codex-adversary.sh --mode diff --base main
# advise — a second opinion on a decision before acting
echo "Queue vs direct calls between A and B? Leaning queue. What am I missing?" \
| ~/.claude/bin/codex-adversary.sh --mode advise --repo .
- First, run codex-adversary.sh --doctor — it reports your Codex version, whether codex exec works, and which required flags are present. - "Codex exited non-zero / no output" (exit 4): usually not logged in — run codex login. - "timed out" (exit 5): raise --timeout (default 600s), or the artifact is too large. - "incompatible" (exit 6): your Codex lacks codex exec or --output-last-message — update Codex (see MAINTENANCE.md).
Exit codes: 0 ran · 2 usage · 3 not installed · 4 failed/empty · 5 timed out · 6 incompatible.
高质量的自动化代码评审工具
AI Skill Hub 为第三方内容聚合平台,本页面信息基于公开数据整理,不对工具功能和质量作任何法律背书。
建议在沙箱或测试环境中充分验证后,再部署至生产环境,并做好必要的安全评估。
✅ MIT 协议 — 最宽松的开源协议之一,可自由商用、修改、分发,仅需保留版权声明。
总体来看,Codex对抗评审 是一款质量优秀的Agent工作流,在同类工具中具备一定竞争力。AI Skill Hub 将持续追踪其更新动态,建议收藏备用,结合自身场景选择合适时机引入使用。
| 原始名称 | codex-adversary |
| 原始描述 | 开源AI工作流:Recruit Codex (GPT) as an automatic adversarial reviewer inside Claude Code — di。⭐23 · JavaScript |
| Topics | ai-agentsai-code-reviewcodexjavascript |
| GitHub | https://github.com/todd866/codex-adversary |
| License | MIT |
| 语言 | JavaScript |
收录时间:2026-06-24 · 更新时间:2026-06-24 · License:MIT · AI Skill Hub 不对第三方内容的准确性作法律背书。
选择 Agent 类型,复制安装指令后粘贴到对应客户端