qybaihe/mu

What it solves

Coding agents often waste tokens, increase latency, and lose focus by using large models to make trivial non-coding decisions (e.g., deciding what to keep in context, whether a command is safe, or when a task is finished). mu solves this by offloading these decisions to a "judgment kernel"—a small, fast model that handles bounded yes/no or scoring questions, allowing the primary large model to focus exclusively on the actual coding work.

How it works

mu implements a system of 35 specific "decision points" across input, context, tools, and turn management. At each point, a small judge (such as the hosted Jev, the local 322M-parameter Laya model, or another LLM) answers a limited question to determine the next action.

Key mechanisms include:

• Context Management: The judge decides which tool outputs are relevant to keep in context and which should be archived, and identifies stale results to prune.
• The Hive: A multi-agent system where "bees" (sub-agents) perform specific tasks. A judge acts as a gatekeeper, deciding if a bee's finding is worth sharing on a collective board or if a note on the board is relevant to a specific bee.
• Plain-Language Board: Instead of asking the working model to narrate its progress, a dedicated model summarizes the agent's state into a human-readable board.
• Judgment Ledger: Every decision made by the judge is logged, allowing users to audit the reasoning behind the agent's behavior.

Who it’s for

Developers who want a high-performance coding agent that manages its own context efficiently, reduces token waste, and provides transparent, human-readable progress tracking.

Highlights

• Judgment Kernel: Uses small, fast models to handle 35 different decision points per turn.
• Local Judge Option: Includes Laya, a 322M-parameter local model for privacy and speed.
• Intelligent Context Pruning: Prevents context overflow by judging tool output chunk-by-chunk and dropping stale data.
• The Hive: A coordinated multi-agent system with a shared, append-only board for findings and disputes.
• Native Desktop App: Provides a full GUI with dedicated tabs for the board, judgments, hive, and a built-in browser.