pridesong/orchestration-skill

README 摘要

orchestration-skill Long-task orchestration engine for AI agents: reliability comes from a materialized state machine + mechanical execution + two-layer audit , not from the model's long-context capability. The agent condenses intent → an orchestrator subagent materializes a three-file contract ( steps / op-table / minds ) → a mechanical executor enforces the state machine → each step is dispatched via the T3 protocol (fill/rules/schema/data/write/forbidden, zero lead-in) → every artifact lands on disk (folder = external memory, resumable) → two-layer audit catches drift. Why Long tasks fail not because the model is weak, but because unreliability compounds across a long chain: - steps blur together, errors propagate silently - the orchestrator's prose instructions get re-interpreted ("protocol wrapped in prose") - state lives only in context and is lost on resume This skill pushes all of it into structure : JSON contracts with JSON-Schema constraints, a state machine with hard transition tables, T3 protocol dispatch, and mechanical gates at every boundary. Never fight drift with more prose — fight it with structure. Features - Three-file contract — steps.json (state layer: what to…

分类