momo-code

About

momo-code starts from the opencode harness and adds machinery for the agent to learn from its own sessions. The fast loop observes session signals — test pass/fail, edit acceptance, user corrections — distills them into Tactic cards, selects candidates with Thompson sampling, and injects them into subsequent prompts; high-confidence tactics graduate into training curricula for the slow loop, where Monte Carlo Graph Search over pipeline configurations produces LoRA weight updates gated by a ratchet that rejects regressions, with spend bounded by a budget variable. A graph engine compiles long-horizon tasks into dependency DAGs whose nodes run as parallel child processes with state persisted to disk, so runs resume after restarts, and simulation-typed nodes can mix in physics-simulation agents. Claude Code migration is deliberately frictionless: existing MCP servers, settings, and prompts are inherited wholesale unless disabled. The tool stays local-first with sessions auditable on disk, targets developers of Chinese providers (GLM, Kimi, Doubao, DeepSeek) alongside Western ones, and remains a single-release v1.0.0 project from one maintainer.