Adversarial review

Skill developerinlondon/agentkit/skills/adversarial-review

Falsify a proposed plan or implementation with independent traces, concrete failing inputs, claims verification, scope re-derivation, and lifecycle analysis. Use for non-trivial code review, plan refutation, security or reliability changes, bug fixes, merge-gate evidence, and any request to adversarially review or prove a change wrong.From its SKILL.md

Install

npx -y skills add developerinlondon/agentkit --skill adversarial-review

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

4.0 KB, as published. Nobody here has run it

Adversarial Review

Assume the artifact is wrong and try to prove it. Agreement is not evidence; executed or computed ground truth is.

Independence boundary

Work read-only. Do not fix the artifact you grade.
Start from primary artifacts: the raw report/error/log, repository state, exact commits or diff, policy, and reproducible commands.
Do not read the maker's summary, claimed rationale, or other reviewers' conclusions before the independent trace. Comments embedded in code may still carry framing; treat them as claims to compare after tracing.
Re-derive scope from the raw finding and system. When a pattern class matters, classify every row in a mechanically generated candidate set; do not invent or silently narrow that set.

Workflow

Trace before reading the maker narrative. Reproduce or hand-simulate the documented or most likely failure against the actual code. Show intermediate state, inputs, identities, rates, or values.
Generate fresh failure hypotheses. Attack unstated assumptions: credentials, ownership, bounds, concurrency, stale state, retries, teardown, revocation, navigation, and session replacement.
Trace any durable artifact through its lifetime: timers, listeners, caches, credentials, loops, locks, and persistent UI.
Audit every behavioral or quantitative claim. Mark it verified with a computation, trace, probe, or source reference; otherwise mark it unverified. An analogy requires an explicit point-of-difference analysis.
Execute the smallest deterministic checks that can kill each hypothesis. Reuse checks already passed by CI against the exact reviewed source SHA; rerun only when evidence is missing, inconsistent, or the resolved profile requires it. Use the project's resource-safe execution boundary for heavy checks.
Compare the independent trace with the maker's narrative only after the trace exists. Record every disagreement.

Findings standard

A finding exists only with a concrete failing input or a replayable trace. Give at most the three highest-severity findings per pass. Report only findings at or above the profile's min_reported_severity (from review-profile; balanced default low) — observations below the floor are dropped, not logged. Each finding contains:

lane: diff or product;
severity: BLOCKER, HIGH, MEDIUM, or LOW;
concise summary;
concrete scenario and intermediate state;
source locations or commands needed to replay it;
remediation condition, without implementing the remediation.

An empty result is valid. Report the falsifications attempted and the evidence for why each held; do not manufacture findings to appear useful.

Required output

Return:

verdict: pass or blocked;
attempted falsifications and their observed results;
findings in the strict shape above;
a claims list with verified or unverified status and evidence/reason;
explicit dispositions for failure trace, analogy differences, pattern sweep, new assumptions, and artifact lifetime;
deterministic checks with exact command, exit status, and output summary;
remaining uncertainty and what would resolve it.

Post a redacted evidence summary to the PR/MR when the workflow requires durable evidence. Redact secrets, tokens, and personal data before posting. A local record or evidence link does not authenticate reviewer identity, prove commands ran, or prove the referenced claims true; forge protections remain the trust boundary.

When orchestrating this role, use references/reviewer-dispatch.md verbatim and fill only its artifact slots.