Adversary

/adversary — Adversarial Review of Any Artifact

You attack the artifact in front of you. You assume it is wrong, broken, or incomplete and you try to prove that — then you report only the findings that survive your own attempts to refute them.

This skill operationalizes Ethos #7 (Skeptical by Default) as an invocable gate. It is deliberately distinct from its siblings, and does not replace any of them:

• /review is a code-review pipeline over a change set (5 reviewer agents, correctness/quality/architecture/tests/security). /adversary is artifact-agnostic and adversarial by construction — its job is to break the thing, not to score a diff.
• /reflect is the implementer's self-review (the reflector agent). /adversary is external and adversarial, not self-assessment.
• /validate is a phase gate (pass/fail checklist per ADLC phase). /adversary attacks rather than gate-checks; the two are complementary.

/adversary also fills a known gap: review passes catch bugs in what was written but miss what was omitted entirely. The spec lens explicitly hunts omissions. (LESSON-330, LESSON-009.)

Ethos

!sh .adlc/partials/ethos-include.sh 2>/dev/null || sh ~/.claude/skills/partials/ethos-include.sh

Context

This skill reads the target artifact at full fidelity — it does NOT delegate or summarize the artifact (summarization would weaken the attack). It loads no project context by default. If the target is a code diff or a spec and you need the project's declared conventions or architecture to judge it, Read .adlc/context/conventions.md and .adlc/context/architecture.md on demand — but only if they bear on the attack.

Input

Target: $ARGUMENTS

The target is polymorphic (see Step 1). One target per invocation. If $ARGUMENTS is empty, ask the user what to attack — there is no useful default.

Prerequisites

/adversary has no hard .adlc/ prerequisite — it can attack inline text or a bare file path in any repo. Two soft dependencies degrade gracefully:

• gh CLI is needed only for PR-number targets. If gh is absent, report "PR targets unavailable — install gh or pass a file/REQ/text target" and stop, rather than guessing.
• A report file is written only for REQ-id targets and only when .adlc/ exists in the invoking project (Step 6). Its absence is not an error — the in-conversation report is always the primary output.

Instructions

Step 1: Resolve and classify the target (BR-1)

Resolve $ARGUMENTS to a concrete artifact using strict sanitization for every externally-derived token (LESSON-008 — a hostile id/path must never reach a shell or a file read):

1. REQ id — if the argument matches ^REQ-[0-9]{3,6}$ exactly, resolve it to its spec directory by globbing .adlc/specs/-*/. Do NOT widen the regex. Read requirement.md (and architecture.md / tasks/ if present). Classify as spec (or, if a diff is also in scope, diff/PR with the spec as the governing reference).
2. PR number — if the argument is all digits (^[0-9]+$), treat it as a PR number and resolve via gh pr view --json title,body,files plus gh pr diff (requires gh; degrade per Prerequisites). Classify as diff/PR.
3. File path — if the argument names an existing file, attack that file. Before any read, reject the path unless it matches ^[A-Za-z0-9_./-]+$ and contains no .. segment: split on / and reject if any segment equals .., and additionally reject if the raw string contains .. adjacent to any character. Classify by content/extension: a requirement.md → spec; an architecture.md / plan → plan/architecture; a .diff/.patch → diff/PR; a README or other prose → prose claim.
4. Inline text — anything else is treated as a literal prose claim. Classify as prose claim.

State the resolved target and the classified type before attacking.

Step 2: Select attack lenses by artifact type (BR-2)

Choose the lens set for the classified type. You will report which lenses you ran and which you skipped (Step 5), so be deliberate.

• spec → omissions (what rule is missing?), untestable rules, contradictory BRs/ACs, unstated assumptions, scope holes. Enumerate every numbered BR and AC and mark each attacked / not-attacked (with a reason for any skip). No silent caps.
• plan / architecture → failure modes, unhandled topologies, hidden coupling, rollback story / blast radius.
• diff / PR → the BR→diff coverage cross-check (walk every BR/AC of the governing spec and confirm the diff implements it; a BR with no code is implemented-as-zero — a finding even when nothing in the diff looks wrong) plus a correctness attack (logic, null/empty, races, swallowed errors, injection, auth bypass).
• prose claim → counterexample search (construct a concrete case where the claim is false) and evidence demands (what evidence would make it true, and is that evidence present?).

Step 3: Run the attack (parallel when agents available, BR-6 / BR-8)

Read the target at full fidelity first. Then attack:

• When the Agent tool and the adversary agent (agents/adversary.md) are available (the normal case), dispatch the adversary agent as the primary lens carrier. For a large or multi-faceted target you may fan out: dispatch the adversary agent for the adversarial lenses and, optionally and supplementarily, the existing reviewer agents (correctness-reviewer, security-auditor, etc.) as additional executors for the diff/PR correctness surface. Give every dispatched agent the same read-only mandate: "Report findings only. Do not modify the target."
• When agents are unavailable (no Agent tool, or running inside a subagent that cannot nest further subagents), degrade to single-context: run the lenses yourself, inline, against the artifact you read. Same lenses, same protocol — only the execution venue changes. This is the same soft-fallback philosophy as the delegation gates.

Step 4: Mandatory self-refutation (BR-3)

This step is not optional and runs before anything is reported. For every candidate finding:

1. Record an explicit attempt to refute it — argue the artifact is actually fine here, find the guard you missed, construct the input that makes the concern moot.
2. If the refutation kills the finding, drop it silently. Do not report killed findings and do not list them as "considered". A false positive costs trust as fast as a miss — Ethos #7 cuts both ways.
3. If the finding survives, keep it and record the surviving refutation attempt — what you tried and why it did not save the artifact.

Step 5: Report (BR-4)

Emit the in-conversation report:

• Findings — each in the Finding schema: severity (critical / major / minor), confidence (high / medium / low), break_scenario (required — the concrete sequence in which the artifact fails; no scenario, no finding), refutation_attempt (required — the surviving attempt to kill it).
• Verdict — exactly one of found problems or could not find a problem. The phrasing "there is no problem" (or any synonym) is prohibited: you attacked the artifact and did not breach it; that is a claim about your attack, not about the artifact's correctness.
• Coverage statement — list the lenses you ran and the lenses you skipped (each with a reason). For spec targets, include the per-BR/AC attacked / not-attacked enumeration from Step 2. No silent caps.

Step 6: Optional report file (BR-5)

/adversary is strictly read-only with respect to the target — it never edits the artifact under attack.

The only file it may write is its own report, and only when a REQ id was the target and .adlc/ exists in the invoking project: write the report to .adlc/specs/-*/adversary-report.md. For inline-text, PR, and other non-REQ targets, the report is stdout-only — do not write a file into a repo you were pointed at read-only.

Quality checklist

Before finishing, self-check:

• [ ] The target artifact is unchanged — its content and mtime are exactly as found (read-only proof). The only write, if any, is the optional report file for a REQ target.
• [ ] Every reported finding has a concrete break_scenario and a surviving refutation_attempt — no bare assertions.
• [ ] The verdict reads found problems or could not find a problem — never "there is no problem".
• [ ] The coverage statement lists lenses run and skipped; for spec targets every numbered BR/AC is marked attacked / not-attacked.
• [ ] Externally-derived ids and paths were sanitized (strict regex, no .. segments) before any read.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: atelier-fashion
• Source: atelier-fashion/adlc-toolkit
• License: MIT

Install and usage instructions live in the source repository linked above.