adversarial-review

Generates verified code and spec review findings using independent multi-lens adversarial agents.

1|Updated Jul 16, 2026
One-click install
npx skills add https://github.com/LunarCommand/claude-skills --skill adversarial-review-lunarcommand
Or copy as Structured Prompt for Agentโ–ผ
Please help me install this Agent Skill.
Skill: adversarial-review
Source: https://github.com/LunarCommand/claude-skills/tree/main/skills/adversarial-review
Command: npx skills add https://github.com/LunarCommand/claude-skills --skill adversarial-review-lunarcommand

SYSTEM DOCUMENTATION & REQUIREMENTS

๐Ÿ’ก This Skill includes scripts (resource) components.

What problem does it solve? Diff-scoped review bots miss whole-system, failure-mode, and consistency defects, and single-pass self-reviews produce confident but wrong findings. This Skill generates review findings from scratch before a PR is opened or merged, running independent adversarial lenses and verifying each finding by refutation so only real defects surface. ## Core Features & Use Cases - Multi-lens adversarial review: Runs specialized lenses (concurrency, failure modes, data consistency, security, spec consistency, normative language) as independent agents so perspectives do not average together. - Refutation-based verification: Every finding is challenged from multiple angles (reproduce, regress, claim-true) and must survive a majority vote before being reported, ranked by severity. - Working-tree protection: Snapshots the git tree, supports throwaway worktree isolation for committed changes, and verifies the tree is unchanged after the review. - Use Case: Before merging a payments refactor, run the Skill on the branch diff; it assembles callers, invariants, and the full changed files, fans out lens agents in isolated worktrees, and returns blocker/should/nit findings with concrete failure scenarios and fixes. ## Quick Start Ask the agent to adversarially review the current uncommitted changes or a specific PR number and report the verified findings ranked by severity.

Frequently Asked Questions about adversarial-review

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I run an adversarial code review before merging a PR?โ–ผ

Pass the PR number, file path, or leave the scope empty to review the current diff. The Skill assembles the changed files, callers, and repo invariants, then runs independent review lenses and verifies each finding by refutation before reporting it ranked by severity.

What is the difference between inline review and the workflow engine?โ–ผ

Inline review runs the lenses in a single context for a quick, lighter pass on one file or a small diff. The workflow engine runs each lens as a genuinely independent agent with merge and multi-angle verification, giving full-fidelity results for PR-scale or release reviews.

Can it review spec or RFC changes, not just code?โ–ผ

Yes. A dedicated spec-accept-review engine carries lenses for normative prose, RFC 2119/8174 keyword usage, conformance fixtures, and cross-reference accuracy. Set targetKind to spec, or leave it as any for mixed code and spec changes.

Does the review modify my working tree or uncommitted changes?โ–ผ

No. Agents operate under a read-only mandate, and when the reviewed state is committed the workflow runs every agent in a throwaway git worktree. The Skill also snapshots the tree before and after and reports any unexpected mutation.

How does it avoid reporting false or already-fixed findings?โ–ผ

Each finding is verified by refutation from multiple angles and must survive a majority vote. Verifiers re-check claims against the reviewed ref via git show and confirm the defect still exists at the branch tip before it is reported.

When should I use delta mode instead of a full review?โ–ผ

Use delta mode when the change was already reviewed and work continued afterward, such as fixes for prior findings. Set baseRef to the previously reviewed commit and reviewRef to the current one so only the new diff is examined.