qybaihe/mu
mu moves dozens of agent-side decisions to a small judge model so the main coding model can stay focused on the work.
The project describes 38 decision points per turn, covering context admission, command risk, prompt-injection filtering, completion checks, memory, and multi-agent routing. Its judge, Jev, answers bounded questions with probabilities and logs each verdict; a local 322M-parameter judge called Laya can be used instead. The repo says mu is in early development, with 0.1.x pre-releases on npm and GitHub Releases, and that names, settings, and formats may change. GitHub · LLM repos' note
The project describes 38 decision points per turn, covering context admission, command risk, prompt-injection filtering, completion checks, memory, and multi-agent routing. Its judge, Jev, answers bounded questions with probabilities and logs each verdict; a local 322M-parameter judge called Laya can be used instead. The repo says mu is in early development, with 0.1.x pre-releases on npm and GitHub Releases, and that names, settings, and formats may change. GitHub · LLM repos' note
score 4