Adversarial Review

Adversarial Review

Run the selected profile’s independent reviewer personas against the same bounded evidence, then verify and synthesize their findings. Seat allocation is declared in device config before the run and frozen before any model call.

Gemini is prohibited as an engine, provider, gateway model, alias, or fallback.

Announce at start: “Using adversarial-review <fast|deep> on .”

Step 1 — Load the device contract

The device source of truth is the ignored council-config.json beside this file. council-config.example.json is the only canonical config shape. scripts/council.py is the thin validation and execution entrypoint; its bundled adapter registry is the command, receipt, and derived-family authority. Do not recreate either contract in prose, a shell wrapper, or another launcher.

The config records only device facts: seat ID, adapter, non-secret auth/setup, absolute direct executable, model, profile bindings, and separate last-observed smoke and review status. It never contains command arguments, receipt selectors, adapter policy, or credential values. Recovery text is derived at read time from the current failure, operation, and seat setup; it is never persisted where it can drift. The registry constructs adapter behavior centrally. Live smoke records the installed CLI version when its benign metadata probe succeeds; version discovery never gates readiness and auto-updates do not require editing config.

Keep the working directory at the repository being reviewed. Resolve this skill's installed directory and invoke the runner there by absolute path; the default config remains beside the skill while artifact paths remain relative to the target repository.

Validate the schema:

python3 /absolute/path/to/adversarial-review/scripts/council.py validate

If its setup_available array is non-empty, report SETUP AVAILABLE and offer one-time configuration before continuing. The runner checks known executable aliases in priority order and reports an installed adapter only when no seat for that adapter is configured. This is advisory; it never auto-adds a seat or changes the review outcome. Also report any configured seat whose last-observed smoke or representative review is not ready, even when the selected profile does not bind it. A passing smoke proves only the benign identity/auth path; it never overwrites a failed representative review. smoke --all refreshes only smoke observations during explicit maintenance. Routine reviews update only review observations. review_qualified means the last representative review succeeded and no newer smoke failure has since disproved basic CLI/auth/model readiness. A later smoke success never turns a failed review into qualified.

Step 2 — Configure missing seats

When council-config.json is missing:

1. Check only codex, Cursor's official agent/cursor-agent aliases, cline, and claude on PATH.
2. Tell the user which known seats are detected and ask which to set up.
3. Copy council-config.example.json to the ignored device config and replace the four built-in templates with the selected seats. Inspect the installed CLI’s own help, auth status, model catalog, and structured output. Do not infer a model or auth route from memory.
4. Record the exact non-secret setup and recovery actions. Prepare required environment variables in the invoking process without writing or printing their values.
5. Record the resolved absolute direct executable. Run validate followed by smoke --all. Repair and re-smoke one seat at a time until every configured seat's live structured identity receipt passes.
6. Deliberately bind a new seat into both profiles and run a representative review before treating it as review-qualified. Keep it bound only if that bounded review succeeds; otherwise leave it configured and visibly unbound with its review failure recorded.
7. Keep fast and deep bound to at least two model families whenever the device has two review-qualified families. Keep failed/unqualified seats in inventory but do not let them make valid profiles impossible. If only one family is qualified, report that limitation instead of inventing diversity.

Record these one-time device steps in the seat itself:

• Codex — install codex, complete first-party login, select the exact GPT model, then smoke and run a representative review.
• Cursor — install the official agent/cursor-agent, provision CURSOR_API_KEY through the device's secret manager, and require the init receipt's auth source to be env. On the tested build, browser login was consumed after one headless call and caused repeated browser prompts, so do not use agent login for council automation.
• Cline — npm install -g cline, approve ClinePass OAuth, select an exact GLM model, and run identity smoke. Leave it unbound when a representative review exceeds the profile bound even though smoke passes.
• Claude Code — install claude, complete first-party browser login, select the exact Claude model, then smoke and run a representative review.

The registry owns each adapter's exact read-only command and structured receipt parser; scripts/council_adapters.py is normative. The parent never retries, resumes, or reassigns an invocation.

Any fallback-named argument, Gemini lineage, shell/interpreter launcher, embedded credential, malformed receipt, model mismatch, or unsafe built-in posture is SETUP REQUIRED. validate also rejects a profile bound to one family when the device has two or more review-qualified families.

Do not configure a custom executable. An opaque binary cannot prove provider lineage strongly enough to enforce the no-Gemini rule.

If a known built-in later appears on PATH but is absent from the config, report SETUP AVAILABLE and offer its one-time setup. Do not silently add it. Do not remove a temporarily failed seat to make a review pass. When a seat is permanently added, revoked, or retired, deliberately update the device config and profile bindings before the next run, then validate and smoke the new frozen allocation. That planned maintenance is not runtime substitution.

Step 3 — Diagnose or maintain seats

review validates the whole config, including family diversity, freezes the selected profile, and launches those persona calls directly. Do not run smoke on the normal review path: its latency is real, and a benign identity prompt is not evidence that a representative review will complete.

When diagnosing setup without starting a review, keep the target repository as the working directory:

python3 /absolute/path/to/adversarial-review/scripts/council.py smoke --profile fast

Use the same absolute runner path with smoke --all only for full device maintenance. Diagnostic smoke probes selected seats concurrently, including failed ones, and updates only observed.smoke. Persona calls update only observed.review. State writes hold a per-config lock and retain the newest observation for each operation, so later smoke success cannot erase a review timeout and overlapping processes cannot erase newer state.

Structured stdout is authoritative. Stderr noise is recorded but does not fail a seat when the process exits zero and every configured result and identity assertion passes.

Step 4 — Run the council

Choose:

• profile — fast by default for iterative/routine reviews; deep for final, high-risk, security-sensitive, or architectural decisions. On adapters with an effort control, fast review calls request low and deep review calls request medium;
• objective — one sentence describing the intended outcome;
• artifact — the smallest readable file set containing the decision and rationale; repeat the flag for multiple files;
• focus — the specific bet to attack, or omit;
• resolved — closed decisions not to re-litigate, or omit;
• evidence — output from a deterministic test, lint, build, typecheck, or factual gate actually run in this task (or an attached command receipt), treated as ground truth; otherwise omit.

Run:

python3 /absolute/path/to/adversarial-review/scripts/council.py review \
  --profile fast \
  --objective "Verify that the proposed design satisfies its stated invariants." \
  --artifact SPEC.md \
  --focus "Failure visibility and model identity" \
  --resolved "Gemini and runtime fallback are prohibited" \
  --evidence "python3 -m unittest discover -s tests -v passed"

fast binds Architect and Adversary to two configured fresh calls and is the default because it makes the speed/reasoning tradeoff explicit. deep binds Architect, Pragmatist, Verifier, and Adversary to four configured fresh calls. When the device has two or more review-qualified model families, both profiles must bind at least two families; unqualified inventory seats do not force impossible bindings. A profile may still reuse a seat for multiple fresh processes. That is planned allocation, not fallback. Report actual engine/model allocation only from the frozen run receipt, never from device-specific prose.

The runner accepts only relative artifacts inside the target repository's working directory. fast requests low effort, caps each fully assembled per-persona prompt at 16 KiB, allows one finding, and caps review calls at 60 seconds. deep requests medium effort, caps each prompt at 24 KiB, allows three findings, and caps review calls at 90 seconds. Explicit diagnostic smoke uses low effort with a 20-second (fast) or 30-second (deep) cap. Prompt limits are measured as UTF-8 bytes after assembly. Version probes are capped at 10 seconds. Both freeze configured bindings and launch distinct seats concurrently in fresh empty temporary working directories; repeated bindings to the same seat serialize to avoid self-inflicted capacity failures. Smoke calls always use low effort regardless of profile. Review calls use the profile setting when the adapter exposes an effort control; otherwise effective effort is carried by the configured model variant. The receipt makes that difference explicit for every call, plus the call-timeout ceiling implied by any serialized bindings. Workers never share outputs or write the receipt; the parent serializes results in profile order and atomically records the same typed seat state after review failures. A failed or invalid persona is never reassigned, retried through another model, or repaired. Every adapter process starts in its own process group. An outer timeout terminates and then kills that whole group. Cline's internal deadline is five seconds shorter than the outer bound, and Cline receives a new empty hooks directory so device-global hooks cannot inject context.

Every adapter returns the same typed JSON persona object. Claude enforces it with native JSON Schema; the runner validates exact keys, identity, verdict/severity consistency, and numbered non-blank, non-whitespace artifact citations for every seat. The runner resolves the exact cited line itself. Missing, invented, or inconsistent fields remain invalid.

Step 5 — Verify and synthesize

When personas ran, use the receipt’s post_run_seats.review for final review status and show smoke status separately when it matters. For SETUP REQUIRED, use the validated config/setup error; INVALID REQUEST has no seat mutation. Do not treat model verdicts as council authority. Against the supplied artifacts, classify every structurally valid finding as:

• accepted;
• already covered;
• closed-decision re-litigation;
• unsupported or invented.

Drop everything except accepted findings. Deduplicate the same defect, re-rank severity from demonstrated consequence, and recompute each effective persona verdict: High → FAIL, else Medium → CONCERNS, else PASS.

Return:

COUNCIL VERDICT: PASS | CONCERNS | FAIL | INCOMPLETE | SETUP REQUIRED | INVALID REQUEST
COVERAGE: profile=<fast|deep> personas=<valid>/<required> families=<used> engines=<used> review_ready=<ready seats> timeout_ceiling=<seconds>
RECEIPT: <path; none only when no persona launched>
SEAT STATUS:
- <seat> (<engine>/<model>): smoke=<status> review=<status> — <review recovery if any>
MUST-FIX:
- [H|M|L] <deduped accepted finding> (raised by: <personas>)
DROPPED:
- <finding> — <already covered|re-litigation|unsupported>
PER-PERSONA:
- <persona> via <engine>/<model> effort=<effective>: <raw verdict> -> <effective verdict|invalid|failed> (<failure class if invalid/failed>)
BIGGEST RISK: <one line>

Outcome order:

1. INVALID REQUEST — the assembled prompt exceeds the profile bound or an artifact/request is invalid; no persona ran.
2. SETUP REQUIRED — device config/setup is invalid before allocation; no persona ran.
3. INCOMPLETE — any frozen profile call failed or was invalid, including auth, model, capacity, or timeout failures discovered by the real call.
4. FAIL — every selected output is valid and an accepted High remains.
5. CONCERNS — every selected output is valid, no accepted High remains, and an accepted Medium remains.
6. PASS — every selected output is valid and no accepted High or Medium remains.

There is no degraded pass and no runtime substitution.