v1.8.0

Highlights

• 新增 Token Bridge：preflight 自动从 macOS Keychain 提取 Claude OAuth token，缓存到 ~/.cache/a2a/.claude-token，供 Codex sandbox 调 claude -p 使用。
• 支持双向跨模型 dispatch：Claude→Codex 用 codex exec，Codex→Claude 用 claude -p + token bridge。
• 新增代码作者检测与 Step 1 自动补 token。
• 新增 scripts/a2a-health.sh，一键检查 Skill、CLI、Auth、Token Bridge 和实际 claude -p 调用。
• 新增 CLAUDE.md 作为 AI agent quick-context，并补充 README 深度使用指南。
• 主要 fix：preflight 的 Claude ready 判定支持 token bridge fallback，避免 auth=false 但 token 可用时被误判。

Validation

• bash scripts/preflight.sh --json --refresh 通过
• bash scripts/a2a-health.sh 通过，实际 claude -p 调用成功

v1.6.0

Highlights

• preflight.sh 新增 TTL cache，默认复用 15 分钟本地检查结果。
• 新增 --refresh 与 --ttl-seconds N，支持强制刷新和按次禁用缓存。
• 增加 best-effort 登录态检查，区分 CLI 已安装与登录缺失。
• JSON 输出新增 authenticated、cache、exit_code 字段，便于上层复用。
• 修复 Codex 登录态字符串解析，避免 not logged in 被误判成已登录。
• 文档同步补充：preflight 只做本地环境/登录态检查，不会发起 review 请求。

Validation

• bash scripts/preflight.sh --json --ttl-seconds 0 通过
• bash scripts/preflight.sh --json --refresh 通过
• bash scripts/preflight.sh --json 命中缓存，返回 cache.source = cached

v1.5.0

Highlights

• 恢复并明确 Plan Review：没有代码 diff 但明确指定 plan / 设计文档时，也可以进入审查。
• 收紧 reviewer 派发规则：按变更行数、新增代码量和目录跨度决定 Challenger / Architect / Subtractor 的组合。
• review packet 改为按 lens 裁剪，项目 red-line 约束延后到 Claude 汇总阶段统一扫描。

Included From Previously Untagged v1.4.0

• 新增前置 Gate：常规 code review 流程下，没有可审代码 diff 时直接拒绝执行。
• 精简 review packet，并明确推荐工作流：Plan -> Execute code -> a2a review。

v1.3.0 — Cross-Model Adversarial Code Review

Self-review is self-deception. Let a different model find what you missed.

What is this?

agent-2-agent (a2a) is a Claude Code skill that implements cross-model adversarial code review. Claude-written
code gets reviewed by Codex. Codex-written code gets reviewed by Claude. Different models, different blind spots —
let them fight.

Three Review Lenses

Every review runs three independent perspectives in parallel:

• The Challenger — "Prove to me this won't break." Edge cases, error paths, race conditions.
• The Architect — "Will this design survive change?" Coupling, boundaries, responsibility.
• The Subtractor — "What if this code disappeared?" Over-engineering, premature abstraction.

Highlights in v1.3.0

• Emoji-based progress indicators (replaced ANSI sequences)
• Rebranded from "Agent Clash" to "a2a"
• Portable preflight (macOS + Linux)
• Animated braille spinner for terminal UX

Install

git clone https://github.com/CtriXin/agent-2-agent.git ~/.claude/skills/a2a
~/.claude/skills/a2a/scripts/preflight.sh   # check environment

Then in Claude Code:
/a2a src/path/to/file.js

Verdict Types

┌───────────┬──────────────────────────────────┐
│  Verdict  │             Meaning              │
├───────────┼──────────────────────────────────┤
│ PASS      │ Ship it                          │
├───────────┼──────────────────────────────────┤
│ CONTESTED │ Disagreements — human arbitrates │
├───────────┼──────────────────────────────────┤
│ REJECT    │ Issues found — fix and re-review │
├───────────┼──────────────────────────────────┤
│ BLOCKED   │ Not enough context to judge      │
└───────────┴──────────────────────────────────┘

Related

- moebius — loop orchestrator that wraps a2a into a continuous cycle
- fight-agent — atomic review engine used by moebius