What problem does it solve? When orchestrating multi-agent work, worker subagent outputs can ship with undetected errors, safety violations, or scope creep. This Skill enforces an independent adversarial review of every worker output before it ships, with a hard cap of 2 review rounds. ## Core Features & Use Cases - Adversarial Review Loop: Dispatches a fresh Reviewer subagent with an adversarial mandate to probe worker output for dietary safety, correctness, security, repo rule compliance, copyright, test honesty, and documentation drift. - Structured Verdict Schema: Reviewers return JSON findings with severity levels (blocker/major/minor), verdicts (CONFIRMED/PLAUSIBLE/REFUTED), and an overall SHIP or CHANGES_REQUESTED decision. - Two Execution Paths: Batch work delegates to the deterministic .claude/workflows/adversarial-review.mjs workflow; single interactive tasks run the loop manually via Agent-tool calls. - Use Case: You orchestrate three worker subagents drafting meal-plan content. The Skill routes each output through a fresh adversarial Reviewer that flags any dietary-safety violation as an automatic blocker, then sends confirmed findings back to the original Worker to fix or rebut. ## Quick Start Ask the orchestrator to run the adversarial review loop on the worker subagent's output before shipping it.
