This is a multi-tier routing system that picks the right model complexity for your task before you waste tokens on Opus for something Haiku could handle. It queries past patterns, predicts outcomes, then spawns an agent at the appropriate tier: sub-millisecond WASM transforms for simple stuff like converting var to const, Haiku for straightforward fixes, Sonnet/Opus for architecture work. The interesting part is the feedback loop. You're supposed to report success or failure after each task so the router actually learns. Pass the --why flag and it explains its reasoning via hooks_explain. If you're burning through API costs on overpowered models or your tasks keep timing out because you picked Haiku for refactoring, this addresses that.
npx skills add https://github.com/ruvnet/ruflo --skill intelligence-route