A2a

a2a (Agent-to-Agent) — 跨模型对抗式 code review。Claude 写的代码让 Codex 审,Codex 写的让 Claude 审,互相找茬不留死角。

Install

skills CLI (Claude Code, Cursor, Codex and more)
npx skills add CtriXin/agent-2-agent --skill a2a
Manual install
git clone --depth 1 https://github.com/CtriXin/agent-2-agent.git /tmp/agent-2-agent
mkdir -p ~/.claude/skills
cp -r /tmp/agent-2-agent/. ~/.claude/skills/a2a

Personal skills live in ~/.claude/skills. Use .claude/skills in a project folder to share a skill with your team.

From the SKILL.md

# a2a — 对抗式 Code Review

> 谁写的代码,就让对手来审。Self-review is self-deception.

## 核心机制

**对抗原则**:review 优先在**对立模型**上执行。Claude 产出的代码由 Codex 审查,Codex 产出的代码由 Claude 审查。

**为什么有效**:不同模型有不同的 blind spot。Claude 倾向 over-engineer,Codex 倾向跳过 boundary check。让它们互相找茬,bug 无处藏身。

### 运行模式

| 模式 | 条件 | 说明 |
|------|------|------|
| **adversarial** | Claude + Codex 双向均可用 | 跨模型对抗审查,最高质量 |
| **single-model-multi-lens** | 仅一侧可用 | 同模型多视角审查,质量降低但仍有多视角覆盖 |
| **blocked** | 无法执行 | 无法执行 review,报告标注 `mode: blocked` 并中止 |

**跨模型双向 dispatch**:
- **Claude 写的代码 → Codex 审**:`codex exec`(Codex 有自己的 API key)
- **Codex 写的代码 → Claude 审**:`claude -p` + **token bridge**(见下方说明)

**Token Bridge 机制**(解决 Codex sandbox 无法访问 macOS Keychain 的问题):

preflight 在有 Keychain 访问权限的环境中(如 Claude Code、正常终端)运行时,会自动提取 OAuth access token 并缓存到 `~/.cache/a2a/.claude-token`(权限 600)。Codex sandbox 内调 `claude -p` 时,通过读取此文件设置 `ANTHROPIC_API_KEY` 环境变量,绕过 Keychain 限制。

```bash
# Codex 环境内调 Claude 的标准方式:
TOKEN=$(cat ~/.cache/a2a/.claude-token 2>/dev/null)
ANTHROPIC_API_KEY="$TOKEN" claude -p "review prompt..."
```

> ⚠️ **从 Claude Code 内部不要直接调 `claude -p`**:`CLAUDECODE=1` 环境变量会触发 nested session 保护。如需从 Claude 侧 review,使用 Agent tool spawn sub-agent。

报告中 `Mode` 字段必须如实标注当前模式。

## 触发方式

- 用户说 "review"、"审查"、"code review"、"帮我 review 一下"
- 用户说 "a2a"、"a2a review"、"对抗审查"
- plan-to-codex 的 Phase 3 收口环节
- 手动指定:`/a2a src/path/to/file.js`

## 审查规模判定

先算变更量,再决定派几个 reviewer。**必须严格遵守,不得跳级。**
Read the full skill on GitHub

Details

Skill name
a2a
License
MIT
Repo updated
12 March 2026

FAQ

What does the A2a skill do?

a2a (Agent-to-Agent) — 跨模型对抗式 code review。Claude 写的代码让 Codex 审,Codex 写的让 Claude 审,互相找茬不留死角。

How do I install the A2a skill?

skills CLI (Claude Code, Cursor, Codex and more): npx skills add CtriXin/agent-2-agent --skill a2a. Skills go in ~/.claude/skills (personal) or .claude/skills (project), and Claude loads them automatically when a task matches the description.

Is A2a popular?

It lives in CtriXin/agent-2-agent, which has 9 GitHub stars and 0 forks.

Related skills

AI Collab Bridge

9

owgit/ai-collab-bridge

A collaboration protocol for Claude, Codex, and other CLI-capable AI models to review each other's code via CLI handoffs. Use this skill whenever the user mentions AI-to-AI collaboration, Claude-Codex bridge, peer review between AI models, "let Claude check what Codex did", "have Codex verify Claude's work", multi-agent code review, AI handoff, two AIs reviewing each other, or wants two different AI models to verify each other's output before completing a task. Also use whenever you (the running AI) want a second opinion from another model and need a structured way to request that review, even

AI, LLMs & Agents

Product Decision Agent

6

atdy/maoxuan-product-agent

中文产品决策 Agent。用于中国大陆互联网产品、运营、增长、商业化、数据、项目推进和组织协作场景:产品规划、需求分析、PRD、需求优先级、排期、版本规划、Roadmap、MVP、灰度、上线、迭代、增长停滞、拉新、投放、渠道、裂变、CAC、LTV、ROI、留存、转化、DAU/MAU、GMV、漏斗、社区运营、内容供给、创作者、用户运营、活动运营、私域、会员、定价、指标异常、数据口径、埋点、A/B Test、用户反馈、客服/销售反馈、竞品冲击、资源不足、项目延期、需求反复、老板临时插需求、跨部门协作、团队冲突、OKR/KPI、目标拆解、复盘等。触发时像资深互联网产品负责人一样,先判断真实问题、当前阶段、核心阻塞、关键约束、相关方和证据充分性,再给出最值得执行的下一步。默认中文回答,不讲理论、不引用原文、不解释历史、不暴露后台方法来源。

AI, LLMs & AgentsIn 1 list

AI Prd

9

Caspian98/ai-prd-skill

Generates a complete, structured AI Product Requirements Document (PRD) in Markdown format, covering background, requirements, model selection, prompt engineering, evaluation system, and stability strategy.

AI, LLMs & Agents

Zh Tw Humanizer

7

acchuang/zh-tw-humanizer

把 AI 寫的繁體中文改寫成自然、有人味的台灣風格文字(去 AI 味、本土化)。 Use when editing or reviewing Traditional Chinese text to remove AI-generated writing tells and localize it to Taiwanese Mandarin: Taiwan vocabulary (影片/軟體/網路/資料 instead of 視頻/軟件/網絡/數據; 大數據 and 數據顯示 stay), Taiwanese tone, particles, and natural sentence rhythm. Never produces simplified Chinese or mainland-style phrasing. Supports annotation mode (list problems without rewriting), protected spans, scene-based intensity, and a pre-delivery quality self-check. Based on Wikipedia's "Signs of AI writing" and the zh-TW "AI生成文的特徵" guides. Triggers: 人性化、去AI味、A

AI, LLMs & AgentsIn 1 list

Anthropic Mind

9

bozhouDev/anthropic-mind

用 Anthropic 的判断框架给出一个决策建议——回答「按 Anthropic 会怎么做这个取舍」。基于 428 篇一手材料(Claude's Constitution、工程博客、alignment 研究、两份 postmortem、Dario Amodei / Jack Clark / Amanda Askell / Chris Olah / Boris Cherny 的长文与访谈)蒸馏出的 8 条世界观模型、18 条领域模型和 21 条决策先例。触发场景:(1) 问「Anthropic 会怎么判断这个」「按他们的思路」「Dario / Jack Clark / Amanda 怎么看」;(2) 在做 agent 工程、产品、研究或内容决策时想要一个有出处的判断,而不是通用建议;(3) 要写关于 Anthropic 立场的内容,需要准确观点、英文原文引用和内部分歧。不要用于:Claude API 的用法、定价与模型 ID(用 claude-api)、Claude Code 的功能操作(用 claude-code-guide)、纯事实查询(发布时间、参数、榜单数字)。

AI, LLMs & Agents

Fabling

7

gncdev/fabling

Behavioral profile that makes Claude work like the Fable 5 model tier — deep but proportional code investigation, outcome-first communication, high autonomy without permission-asking, faithful outcome reporting, and disciplined turn endings. Use this skill whenever the user asks Claude to behave "like Fable", mentions "fabling", "fable mode" or "fable gibi", complains that Claude is being lazy or superficial, or asks Claude to be more thorough, more autonomous, less chatty, or to stop asking for permission mid-task. Load it once at the start of the session and follow it for every subsequent tu

AI, LLMs & AgentsIn 1 list

Browse the most popular Claude skills, all ai, llms & agents skills, or write your own with the skills generator.