heguangyong/scene-capability-engine

Built on a semantic model of `project -> scene -> spec -> task`, SCE enforces spec-first governance and recoverable SQLite-backed state to prevent agent drift, keeping requirements and design decisions attached to work rather than buried in chat history. The engine supports both interactive studio workflows (`plan → generate → apply → verify → release`) and autonomous modes (`close-loop`, `close-loop-program`, `close-loop-controller`) with built-in retry, fallback chains, and governance replay. It integrates with IDEs and AI shells via structured task streams and hierarchical task references, while an errorbook-driven learning system and timeline snapshots enable incident staging and deterministic reruns even before git commits.

Available on npm.