Multi-Model Adversarial Review
Runs claude/codex/gemini as independent reviewers, has them cross-critique, then synthesizes one deduped report
Funguje s nastavením
Co umí
A three-stage driver (parallel review, NxN-1 cross-critique, synthesis) that shells out to whichever coding-agent CLIs you have installed, drops findings a critic proves are fabricated, and escalates severity on disagreement. Triggers on 'MMAR review', 'multi-model review', 'cross-model adversarial', or 'review with all the models' for high-stakes diffs; the project's own guidance says to use its cheaper sibling skill for routine review.
Testovací report
The SKILL.md body calls ${CLAUDE_PLUGIN_ROOT}/scripts/mmar.py, a 645-line driver that lives at the repo root, not inside the skill folder -- a bare copy 404s in real GitHub raw fetches and errors 'No such file or directory' when actually run; but running the real driver in the project's own mock mode against its bundled fixtures worked perfectly (4/4 eval fixtures passed, and it correctly caught and dropped a fabricated Gemini finding that two other reviewers flagged as fiction) -- so the mechanism itself is genuinely good, it just needs the whole repo/plugin, not the skill subfolder alone.
Testováno: 2026-07-15 · Claude Code 2.x (agent harness)
Instalace
git clone https://github.com/prime-radiant-inc/parallel-adversarial-review cd parallel-adversarial-review mkdir -p ~/.claude/skills cp -r skills/multi-model-adversarial-review ~/.claude/skills/multi-model-adversarial-review
Příkazy a ukázkové prompty
/multi-model-adversarial-reviewRuns claude/codex/gemini as independent reviewers, has them cross-critique, then synthesizes one deduped report
Skilly se spouštějí běžnými požadavky — žádné příkazy k zapamatování. Po instalaci ho aktivují prompty jako tyto (anglicky):
Run a full cross-model review on this PR before we merge to mainThis diff is security-critical, get several models to cross-check each otherOne reviewer isn't enough here, have Claude, Codex, and Gemini all weigh in