prime-radiant-inc/parallel-adversarial-review

15 stars · Last commit 2026-06-06

Two Claude Code skills for adversarial code review (single-model PAR and multi-model MMAR with cross-critique to catch hallucinations), plus a fixture-based eval suite.

README preview

# parallel-adversarial-review

Two skills for adversarial code review, plus an eval suite.

## What's in here

### `skills/parallel-adversarial-review/`

The original PAR pattern, ported from `iterative-development`. Two same-model reviewer subagents run in parallel under a competitive scoring frame; their findings are aggregated, with the worst severity winning on disagreement.

Use this for routine review.

### `skills/multi-model-adversarial-review/` (MMAR)

A three-stage pipeline that uses multiple installed coding-agent CLIs as independent reviewers, then runs a cross-critique grid where each reviewer evaluates the others' findings (catching hallucinations and severity inflation), then synthesizes a final deduplicated report.

```
Stage 1: parallel reviews   (each CLI reviews independently)
Stage 2: cross-critique     (each CLI verifies other CLIs' findings)
Stage 3: synthesis          (one model merges everything, applies rules)

View full repository on GitHub →