adversarial-review-fa8f2b — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited adversarial-review-fa8f2b (Agent Skill) and scored it 100/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 0 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 0 flagged
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
A fresh adversary catches what the author can't: assumptions baked in so early they became invisible. But adversaries have their own bias — Codex tends to recommend over-engineered solutions. So the review is two moves: collect hostile findings, then arbitrate them ruthlessly.
Primary — Codex CLI (read-only; mechanics in references/codex-invocation.md, read it before invoking):
codex exec -s read-only -C "<repo-root>" --ephemeral -o /tmp/codex-review.md "<reviewer brief>"The brief must demand minimal-change findings: state what to review (plan files or diff), the repo context, and the constraint — "Report concrete gaps, bugs, ordering errors, and missing failure handling. Do NOT propose redesigns, abstractions, or scope expansions; findings that require reshaping the plan or codebase will be discarded."
Fallback — fresh Claude (when Codex is unavailable or errors; never block on it): spawn a fresh-context subagent with ONLY the artifact + the same hostile brief — no conversation history, so it can't inherit your assumptions. Tell it to actively try to refute the plan/diff, not to improve it.
For each finding, verify against the real code/plan before acting:
Never accept a finding you couldn't defend by pointing at the artifact, and never discard one without checking it first.
ADVERSARIAL REVIEW (<codex|claude-fresh>, <artifact>)
Integrated: <n> — <one line each: what + where fixed>
Discarded: <n> — <one line each: what + why (over-engineering / out of scope / wrong)>
Escalated: <n> — <one line each>Then apply the integrated fixes (if not already applied) and hand back to the caller.
~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.