I Built a Self-Healing Codebase Using NATS, Claude AI, and A/B Code Tournaments

The Problem I run 30+ projects across multiple servers. When something breaks - a test fails after a push, a dependency update causes a regression - I used to find out hours later, context-switch, dig through logs, and fix it manually. That workflow doesn't scale. So I built ChangeBus: a distributed agent system where AI fixes compete against each other in tournaments, and only the winner gets merged.