adversarial-review
Before committing substantial work, at worker completion, before external review, or on an explicit adversarial/red-team request for a diff, branch, or PR, use independent finders and a skeptic to verify defects, disposition findings, and enforce the blocker gate.
timharris707 · 已有静态检查记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add timharris707/skills --skill adversarial-review技能说明
name: adversarial-review description: "Before committing substantial work, at worker completion, before external review, or on an explicit adversarial/red-team request for a diff, branch, or PR, use independent finders and a skeptic to verify defects, disposition findings, and enforce the blocker gate."
Adversarial review
Read the Codex desktop binding when this workflow needs harness mechanics, model routing, or recovery.
Reviewers whose job is to find what is wrong with a change before it ships, not to summarize it, not to appreciate it. This skill is the portable protocol: isolated finders, a skeptic pass that separates proven defects from plausible ones, and a gate that only skeptic-confirmed blockers may hold. Everything a repo does differently (its defect history, its probe policy, its stakes) lives in binding slots the team-workflow setup interview fills.
Read the team-workflow binding doc first. The review runs before external reviewers (CodeRabbit, Copilot, human PR review) see the change: they are the safety net, not the review.
1. The two layers
- The floor: before committing substantial work. The implementing session reviews its own change before the commit. Every repo bound to the pack has this layer.
- The close-out layer: the orchestrator reviews the lane. In repos running orchestration, the orchestrator (or its delegated verifier) runs the review against a lane's branch at close-out, per the orchestrate skill's audit rule; the implementer never has the last word on its own work. The executor's model and effort come from the repo's review-tier policy (the orchestrate skill's verification-executor binding slot), never habit. Bound per repo; a repo without lanes simply has no second layer.
What counts as substantial: the implementing session judges, biased toward reviewing when unsure: a skipped review is a silent decision that the change couldn't bite. The repo binding may pin hard rules ("any migration or money-path change is always substantial"); those are never overridable by the implementer's judgment.
2. Composition: three finders, isolated
Three reviewers that never see each other's reasoning. Give each a fresh context, the raw change/spec, applicable project rules and decisions, and its named lens. Use fork_turns="none" for independent Codex reviewers. Schedule within live capacity, in separate batches if needed. A parent that cannot launch reviewers returns the request to its coordinator. Sequential passes in one context are multiple perspectives, not isolated reviews; report that limitation rather than claiming independent agreement.
-
The correctness finder. Logic, edge cases, error paths, invariants; the defect-class checklist is its opening moves, not its limit.
-
The fitting lens. One perspective chosen to match the change, from the menu:
- security: inputs, authz, secrets, injection, egress; anything reaching a trust boundary.
- compatibility / migration: schemas, serialized formats, public APIs, upgrade paths; anything an old client or old data can disagree with.
- money / ledger: balances, idempotency, rounding, double-entry; anything that moves or records value.
- concurrency: races, ordering, retries, partial failure; anything with two writers or a queue.
- performance: hot paths, N+1, unbounded growth; anything multiplied by scale.
- UI / accessibility: states the pixels can lie about, spoken output, keyboard paths; anything a person operates.
- conventions / standards: the diff against the repo's own documented standards, style guides, ADRs, contribution rules, naming and layout conventions the repo wrote down. Available only where such documents exist (the finder reads the repo's documents, never an imported checklist), and it joins the menu as one more fitting lens beside the adversarial ones, never in place of the correctness finder or the spec axis.
The invoker picks and states the pick in the report; the repo binding may pin mandatory lenses for named paths (a money lens on payment code is not optional there). A change fitting no lens well still gets its best fit: a second pair of eyes with a stated angle beats a second correctness pass.
-
The spec axis. Checks the diff against the originating ticket/spec: requirements missing, half-done, or wrongly done, and scope the spec never asked for. No ticket or spec? The gap itself is a line in the report (in a repo with tracker discipline, untracked substantial work is worth flagging), and the other two finders proceed normally. The spec axis never invents requirements to check against.
Finders read code and run proofs (tests, focused scripts, repro snippets) but modify nothing: no source edits, no commits, no state mutation beyond what a test run inherently does. Live probes against real services, seeded databases, or running apps are a binding slot: the repo names where they are mandatory (and safe), the protocol never assumes them.
3. The skeptic pass
Findings from a finder are hypotheses, not evidence. Every finding above a NIT goes to an independent skeptic whose brief is to kill it: re-read the code, run the disproof, find the guard the finder missed. Only findings the skeptic fails to kill stand in the report; killed ones drop to the dismissed bucket (§4) with the reason they died. A NIT skips the skeptic and reports as style-level advice, with no skeptic outcome and no rung.
BLOCKER rank requires a runnable reproduction the skeptic confirmed. No repro, no blocker: it ranks MAJOR at most, stated as unproven. This keeps the gate honest: nothing blocks a merge on a hunch, and a blocker in the report is a defect you can watch happen.
The evidence ladder
Finders and skeptics grade every claim, finding and clean bill alike, in one language: how far down this ladder the proof got.
- Asserted. The reviewer said so. Wo