adversarial-audit

name	adversarial-audit
description	Adversarial verification of claimed work: read shipped modules (not memory), grep wiring and approvals, run tests, demand evidence before pass verdict. Use when reviewing agent output, PRs, or security-sensitive changes before merge. Triggers: adversarial audit, refute claims, evidence before verdict, pre-merge review.

adversarial-audit

Load rule: Formal SoT. Distinct from re-running verify cmds alone (agent-orchestrator VERIFY) and from Outer REVIEW_GATE shape (control-graph) — this skill is refute-first evidence pack with independent reads.
CLT: ../rules/clt-dual-load.mdc — review is protected germane load; never collapse to implementer self-summary.

// Role
Reviewer ⊥ ImplementerContext   // fresh session/context; try to refute claims

// Axioms
A1  Default real=false until independent evidence
A2  Read shipped files + run commands — never trust summary alone
A3  Verdict requires: scope check + tests/cmds + explicit gaps
A4  Compose: Orch owns re-run of verify cmds; CG owns when REVIEW_GATE fires; this skill owns attack surface + evidence artifact
A5  Fail closed on missing evidence for security/auth/deps claims
A6  Evaluate(δ) ≔ (C, E, η)

Related: agent-orchestrator · control-graph · fix-dependency-security.

---

Use / skip

Use	Skip
worker/agent claimed "done"	pure design chat, no diff
pre-merge high stakes	typo you wrote yourself 30s ago (still run lint if easy)
security / auth / deps changes
refute marketing claims in PR body

vs Orch VERIFY: Orch re-runs listed cmds and checks scope. This skill also opens files, greps wiring, hunts missing authz/tests, and writes a verdict artifact.
vs CG REVIEW_GATE: CG is phase/budget control. This skill is the review content procedure for the review Role.

---

Procedure

1 Scope inventory → 2 Read shipped code → 3 Run verify cmds → 4 Attack claims → 5 Verdict artifact

1. Scope inventory

git diff <base>...HEAD --stat
git log <base>..HEAD --oneline

List claimed outcomes from brief/PR body.

2. Read shipped code

Open every file that implements a claim; grep wiring:

rg -n "pattern|approval|TODO|FIXME|XXX" --glob '!**/node_modules/**'

3. Run verify commands

From brief / AGENTS.md / CI — capture exit codes:

# examples — use project actuals
pnpm type-check && pnpm lint
cargo test -q

4. Attack claims

• Does diff match the claim?
• Missing error paths, authz, cleanup?
• Tests that don't exercise the change?
• Secrets or overly broad permissions?

5. Verdict artifact

## Adversarial verdict
**Claims reviewed:** …
**Evidence:** files read · cmds + exit
**Confirmed:** …
**Refuted / gaps:** …
**Risk:** low|med|high
**Decision:** pass | fix-required | blocked

---

Done when

• Independent reads + at least one automated check (or documented why impossible)
• Explicit pass/fail with gaps list
• No merge recommendation on testimony alone

Anti-patterns

¬	Do
"LGTM" from PR text	open the files
same context that implemented	new review role / session
ignore failing tests	fail closed
duplicate full control-graph Outer	link CG; stay on evidence pack

Done_when cmds: git diff <base>...HEAD --stat · project verify cmds · verdict markdown present.