code-review-adversarial

code-review-adversarial

One reviewer isn't enough. Adversarial multi-agent code review — a committee of specialist reviewers that catches what single-pass AI review misses.

English | 中文

Quick start:

git clone https://github.com/yuefanxiao/code-review-adversarial.git
ln -s "$(pwd)/code-review-adversarial" ~/.agents/skills/code-review-adversarial  # ZCode
ln -s "$(pwd)/code-review-adversarial" ~/.claude/skills/code-review-adversarial  # Claude Code
ln -s "$(pwd)/code-review-adversarial" ~/.codex/skills/code-review-adversarial   # Codex

Then just ask your agent: "adversarial review PR #123" or "深度评审这个 PR".

What is this?

An Agent Skill that turns a single AI coding agent into a committee of specialized reviewers. Instead of one agent giving you one opinion, this skill dispatches 2–4 reviewers from distinct stances, converges their findings, and outputs a structured verdict.

Before vs After

Single-agent review:

"The CAS implementation looks correct. The WHERE updated_at=? condition properly guards against concurrent updates. Looks good ✅"

Adversarial committee:

Root-cause auditor (P1): The DECISIONS doc claims "no transaction infrastructure exists" but repository.go already has 3 Transaction() call sites — the fix rationale is based on a factual error.

Regression auditor (P1): The new update_mask field is server-enforced as required. The shipped frontend doesn't send it yet → every existing edit will 400 until the frontend is updated.

Devil's advocate: 9 attack vectors tested, no blocking defects found. CAS token source is correctly client-side (cmd.ExpectedUpdatedAt), not the internal read snapshot — no TOCTOU.

Verdict: ready-with-fixes — fix the DECISIONS doc error + track frontend update as follow-up.

How it works

PR diff
  │
  ▼  Phase 0: Baseline — lock merge-base, prevent diff pollution
  ▼  Phase 1: Classify — detect PR type → select personas
  ▼  Phase 2: Dispatch — 2-4 specialized reviewers in parallel
  ▼  Phase 3: Triage — discard out-of-scope, uncited, layer-violating findings
  ▼  Phase 4: Converge — independent→accept, conflict→arbitrate → verdict
  │
  ▼  Output: markdown / JSON / PR comment (asks first)

PR-type-aware

The review focus adapts to what kind of PR it is:

11 Reviewer personas

Installation

Prerequisites

Any agent runtime that supports the Agent Skills format.

Install via symlink (recommended — share one copy across agents)

git clone https://github.com/yuefanxiao/code-review-adversarial.git
cd code-review-adversarial

# Pick your agent(s):
ln -s "$(pwd)" ~/.agents/skills/code-review-adversarial  # ZCode
ln -s "$(pwd)" ~/.claude/skills/code-review-adversarial  # Claude Code
ln -s "$(pwd)" ~/.codex/skills/code-review-adversarial   # Codex

Install via copy

cp -r code-review-adversarial ~/.agents/skills/  # or ~/.claude/skills/, ~/.codex/skills/

Usage

Just describe what you want — the skill auto-triggers on:

• "adversarial review PR #123"
• "深度评审 PR #123"
• "对抗复核 this PR"
• "does this PR have a better alternative?"
• "契约评审"

Options

File structure

code-review-adversarial/
├── SKILL.md                          # Main workflow (326 lines)
├── DESIGN.md                         # Full design rationale + research
├── README.md                         # This file
├── assets/                           # README images (before-after, pipeline, committee)
├── references/
│   ├── type-playbooks.md             # Per-PR-type checklists
│   ├── reviewer-personas.md          # Persona index + shared constraints
│   ├── convergence-protocol.md       # Convergence rules
│   └── personas/                     # One file per persona (15-73 lines)
│       ├── design.md
│       ├── correctness.md
│       ├── root-cause.md
│       ├── regression.md
│       ├── behavior.md
│       ├── test-coverage.md
│       ├── security.md
│       ├── perf.md
│       ├── devils-advocate.md
│       ├── contract.md
│       └── simplification.md
└── examples/
    └── bugfix-cas-review.md          # Sample review output

Design principles

1. Specialized committee, not pro/con debate — Collaborative > adversarial (NeurIPS 2025)
2. PR type determines focus — Feature → L3; Bugfix → L2 (L3 forbidden); Refactor → invariants
3. Free exploration + finding-type constraints — Explore freely, but out-of-diff findings must be "diffusion-impact" only
4. Independent Triage Filter — Primary precision lever (~51% false-positive reduction)
5. Convergence by finding type — Independent → accept; intersection → high-confidence; conflict → arbitrate
6. Merge-base baseline lock — Prevents diff pollution (the #1 real-world lesson)

See DESIGN.md for full rationale and ROADMAP.md for what's coming.

Compatibility

Platform-agnostic — works with GitHub, GitLab, Gitea, or any git-based PR system.

Examples

See examples/ for real review outputs:

• bugfix-cas-review.md — A concurrency-fix PR reviewed by root-cause + regression auditors + devil's advocate

Contributing

Contributions welcome! See CONTRIBUTING.md.

Areas where help is especially appreciated:

• New persona templates (e.g., accessibility, i18n, database migration)
• Benchmark data (single-agent vs committee detection rates)
• Platform-specific integration guides (Cursor, Gemini CLI, Copilot)

License

MIT