A multi-model adversarial + consensus pull request review skill for opencode and Claude Code.
Four different model families independently review your PR, then an adjudicator deduplicates, resolves disagreements, validates fixes against the surrounding code, and ranks every finding on a 0–5 severity scale. You get one action plan and one question: which fixes to apply.
Models from the same family share blind spots. A single reviewer (or three same-family variants) will miss the same class of bugs. This skill deliberately dispatches diverse model families — DeepSeek, Kimi, Qwen, and GPT — so their disagreements surface real issues that any one model would miss. The adjudicator then merges their findings into a single, confidence-scored plan.