adversarial-review

adversarial-review

Claude Code plugin for adversarial multi-model code review.

Mechanical checks first (free), then AI agents scaled to change complexity. Two agents — The Optimizer and The Skeptic — review your code independently, challenge each other's findings, and only consensus issues get auto-fixed. A bounded verification loop catches regressions from fixes.

Install

/plugin marketplace add ng/adversarial-review

/plugin install adversarial-review

To update to the latest version, re-run both commands.

Usage

/adversarial-review:run              # auto-fix (default), auto-detect PR
/adversarial-review:run 405          # auto-fix, specific PR
/adversarial-review:run --no-fix     # review only, no code modifications
/adversarial-review:run --no-fix 405 # review only, specific PR

GitHub Action

Use as a GitHub Action to run adversarial reviews automatically on PRs:

name: Adversarial Code Review

on:
  pull_request:
    types: [opened, ready_for_review, reopened, labeled]

jobs:
  review:
    if: >-
      github.event.action != 'labeled' ||
      github.event.label.name == 'review'
    runs-on: ubuntu-latest
    permissions:
      contents: write
      pull-requests: write
      issues: write
      id-token: write
    steps:
      - uses: actions/checkout@v4
        with:
          fetch-depth: 0
      - uses: ng/adversarial-review@v1
        with:
          claude_code_oauth_token: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}

Inputs

Input	Required	Default	Description
claude_code_oauth_token	Yes	—	Claude Code OAuth token
pr_number	No	Triggering PR	PR number to review
mode	No	auto-fix	auto-fix or no-fix (report only)
allowed_tools	No	—	Additional allowed tools (comma-separated)
model	No	—	Model override for lead agent

Recommended triggers

Avoid synchronize (fires on every push) — the review is slow and expensive. Use labeled with a review label for re-runs after pushing fixes.

How it works

Pipeline overview

flowchart TD
    Start(["/adversarial-review:run"]) --> Context["1. Get Context"]
    Context --> PR{PR exists?}
    PR -->|Yes| Feedback["2. Pull PR/MR Feedback"]
    PR -->|No| Docs
    Feedback --> Triage["3. Triage Feedback"]
    Triage --> Docs["4. Read Convention Docs<br/>REVIEW.md · .claude/docs/"]
    Docs --> Mechanical["5. Mechanical Checks<br/>lint · typecheck · build · test"]
    Mechanical --> Gate{"6. Cost Gate"}
    Gate -->|"Score ≤ 0"| Report
    Gate -->|"Score 1–4"| Standard["Sonnet-only<br/>Optimizer + Skeptic<br/>(2 agents)"]
    Gate -->|"Score ≥ 5"| Full["Dual-model<br/>Optimizer + Skeptic<br/>(4 agents)"]
    Standard & Full --> Synth["7. Synthesize"]
    Synth --> ModeCheck{Auto-fix?}
    ModeCheck -->|"--no-fix"| Report
    ModeCheck -->|"Default"| Apply["Apply consensus<br/>Critical/Major fixes"]
    Apply --> Verify{"Verify fixes<br/>(max 2 iterations)"}
    Verify -->|"Checks pass"| Report
    Verify -->|"Checks fail,<br/>iteration < 2"| Fix["Fix regressions"] --> Verify
    Verify -->|"Still failing<br/>after 2 rounds"| Report
    Report["8. Structured Report<br/>+ PR/MR comments"] --> Issues{"File issues?<br/>(offered after report)"}
    Issues -->|Yes| File["9. File Issues<br/>with full review context"]
    Issues -->|No| Done
    File --> Done([Author reviews & approves])

Adversarial review detail (Step 6)

flowchart TD
    Team["TeamCreate +<br/>TaskCreate with dependencies"] --> Pass1

    subgraph Pass1["Pass 1 — The Optimizer"]
        Optimizer["Find every issue worth fixing"]
        OptSonnet["Sonnet teammate"]
        OptOpus["Opus teammate<br/>(full depth only)"]
        MergeOpt["Lead merges &<br/>deduplicates"]
        Optimizer --> OptSonnet & OptOpus
        OptSonnet & OptOpus --> MergeOpt
    end

    MergeOpt -->|"SendMessage<br/>wake Skeptics"| Skeptic

    subgraph Pass2["Pass 2 — The Skeptic"]
        Skeptic["Challenge findings +<br/>catch missed issues"]
        SkpSonnet["Sonnet teammate"]
        SkpOpus["Opus teammate<br/>(full depth only)"]
        MergeSkp["Lead merges<br/>challenges"]
        Skeptic --> SkpSonnet & SkpOpus
        SkpSonnet & SkpOpus --> MergeSkp
    end

    MergeSkp --> Confidence

    subgraph Synthesis["Synthesize"]
        Confidence{"Cross-model<br/>consensus?"}
        AutoFix["Auto-fix<br/>Critical/Major"]
        Dispute["Present dispute<br/>to author"]
        Note["Note for author<br/>(Minor/Nit)"]
        Confidence -->|"Both models agree"| AutoFix
        Confidence -->|"Models disagree"| Dispute
        Confidence -->|"Low confidence"| Note
    end

    AutoFix & Dispute & Note --> Haiku