Gossipcat

Runs multiple AI agents in parallel to review your code, then cross-verifies their findings against actual file:line citations in your source. When agents agree, the bug is real. When one hallucinates, peers catch it and the system penalizes that agent's future dispatch weight. Over time, each agent builds an accuracy profile and the orchestrator routes tasks to whoever performs best in each category. Includes a live dashboard at localhost:63007 showing consensus rounds, per-agent stats, and hallucination catches in real time. The learning loop works by updating markdown skill files in .gossip/agents, not model weights. Reach for this when single-pass AI review ships too many false positives and you want mechanical verification before acting on findings.