loushang

About

loushang's premise is that agent reliability comes from structure around the model, not from a smarter model. A method defines the work contract for a class of work - roles, phases, constraints, artifacts, acceptance criteria - sessions are durable execution records that can be resumed, forked, exported, and audited, and tools execute under policy governance with project-level extensions adding hooks and custom commands. The loushang.ai SDK routes across GPT, Claude, DeepSeek, Qwen, Kimi, GLM, and MiniMax with a model catalog, streaming, and cost tracking. The V1 product surface, loushang code, delivers this as a CLI and terminal workbench; later roadmap stages extend the same runtime objects to research workflows, team approvals, and a managed runtime. It is an actively developed solo project (roughly 1.3k stars) still in early V1.