ref-sp-agents-adversarial-reviewlisted

# Adversarial Review ## Purpose Define adversarial review as a portable method: a change is verified by a reviewer that is **structurally separated** from the agent that produced it, holding an **opposed mandate** (find fault, not confirm success), and every verdict is **gated on an objective oracle**. The method runs across one or more selectable dimensions — skills consistency, code conformance and scope, security, and end-to-end behavior — and degrades gracefully across agent harnesses. This skill owns the **method and its preconditions**. The runnable step-by-step workflow lives in the driver skill (`tool-sp-run-adversarial-review`); read this to understand *why* the pieces are shaped the way they are, and consult it to actually run a pass. ## When to use this skill - Designing or running a review that a separate agent (not the author) performs. - Deciding whether a change is safe to accept, and what evidence that decision rests on. - Reviewing code against a repo's own skills, guidelines, and stated task scope. - Reviewing a change for introduced security risk, or deciding whether active testing is warranted. - Smoke-testing an application end-to-end after a change. - Reasoning about why an agent reviewing its own work misses defects. ## Scope boundaries - `ref-sp-agents-verification-discipline` — the **same agent checking its own claims** against ground truth (two dials, enumerate/route/prune/abstain). Adversarial review is its *structural* complement: a **di