6BNBN/FlowPilot

Implements a file-driven agentic workflow system that decomposes requirements into parallel subtasks executed by multiple AI agents (Claude Code/Codex/Cursor/snow-cli), with four-layer memory architecture (progress.md/task-xxx.md/summary.md/memory.json) ensuring stateful resumption across interruptions and preventing main agent context bloat. Built on a checkpoint-based execution model with automatic dependency analysis, batch task dispatch, and filesystem-persisted state that survives client crashes and compact operations. Integrates native OpenSpec format detection, long-term memory retrieval (BM25 + dense vectors with RRF fusion), and a self-evolution engine (Reflect→Experiment→Review cycle) that automatically tunes retry parameters and protocol hints based on per-workflow metrics.