Skill

adversarial-workflow

This skill should be used when the user asks about "adversarial review", "codex review loop", "plan review with codex", "dual-agent review", or discusses using multiple AI agents for code review.

From adversarial-review

Install

Run in your terminal

npx claudepluginhub anthropics/claude-plugins-community --plugin adversarial-review

Tool Access

This skill uses the workspace's default tool permissions.

Skill Content

Similar Skills

agent-harness-construction

Designs and optimizes AI agent action spaces, tool definitions, observation formats, error recovery, and context for higher task completion rates.

ecc

140.3k

agent-payment-x402

Enables AI agents to execute x402 payments with per-task budgets, spending controls, and non-custodial wallets via MCP tools. Use when agents pay for APIs, services, or other agents.

ecc

140.3k

agent-eval

Compares coding agents like Claude Code and Aider on custom YAML-defined codebase tasks using git worktrees, measuring pass rate, cost, time, and consistency.

ecc

140.3k

Stats

Stars1

Forks0

Last CommitMar 25, 2026

Actions

View Source View Plugin View on GitHub View README

Adversarial Review Workflow

This project has an adversarial review plugin that orchestrates iterative review loops between Claude Code and OpenAI Codex CLI.

Available Commands

`/adversarial-plan <task description>`

Creates an implementation plan with adversarial review:

Claude analyzes the codebase and writes a detailed implementation plan
Codex reviews the plan critically as a senior staff engineer
Claude revises based on feedback
Loop continues until Codex approves or max iterations (20) reached
Pauses for user approval before any implementation

`/adversarial-review`

Runs adversarial code review on current changes:

Captures git diff of all changes
Codex reviews the code for bugs, security issues, performance, and style
Claude autonomously fixes any critical issues found
Loop continues until Codex approves or max iterations (20) reached
No pause — fixes are applied automatically

Convergence Protocol

Codex ends every review with exactly one of:

APPROVED — no blocking issues, work is acceptable
NEEDS_REVISION — blocking issues listed that must be addressed

Convergence typically happens in 2-3 iterations. The 20-iteration limit is a safety backstop.

Artifact Storage

All artifacts are saved to .adversarial-review/<session-slug>/ in the project root. Each run of /adversarial-plan or /adversarial-review creates a new session directory with a unique slug, so previous sessions are preserved.

A latest symlink at .adversarial-review/latest always points to the most recent session. This is used by /adversarial-review to find plan_final.md from a prior planning session.

.adversarial-review/
  latest -> abcd-efgh-ijkl/
  abcd-efgh-ijkl/           # planning session
    plan_v1.md, plan_v2.md, ...
    review_v1.md, review_v2.md, ...
    plan_final.md
  mnop-qrst-uvwx/           # code review session
    diff_for_review.md
    code_review_v1.md, code_review_v2.md, ...

You may want to add .adversarial-review/ to your project's .gitignore.

Customization

Plugin files are installed to the Claude Code plugin cache. To customize, fork the repo and point your marketplace at your fork, or edit files in the cache directly (changes will be overwritten on update).

Iteration limit: Currently hardcoded to 20 in the command files (commands/adversarial-plan.md, commands/adversarial-review.md)
Review prompts: Customize what Codex evaluates by editing scripts/codex-review.sh
Review types: The script supports plan (architectural review) and code (code change review) types