What WinDAGs is, in code.

Topologically sorts subtasks into waves, runs each wave's nodes as parallel claude -p processes, propagates aborts via SIGTERM → SIGKILL. Zero token overhead per node.

• packages/core/src/core/executor.ts
• packages/core/src/executors/process-executor.ts
• packages/core/src/core/topology.ts

Six stages: BM25 → Tool2Vec → RRF fusion → cross-encoder rerank → local attribution k-NN → cross-user global priors. One scoring path; the DAG executor, MCP server, slash skill, and shortlister all call SkillSearchService.search(). The cascade is what gets the agent's first response to look like the one a senior engineer would have written.

• packages/core/src/retrieval/skill-search-service.ts
• packages/core/src/retrieval/bm25-index.ts
• packages/core/src/retrieval/tool2vec.ts
• packages/core/src/retrieval/cross-encoder.ts
• packages/core/src/retrieval/attribution-db.ts
• packages/core/src/retrieval/global-priors-client.ts

Pre-computes per-skill task-space vectors. For each skill: ask Haiku (or gpt-4o-mini / gemini-2.5-flash, whichever key is present) for 15 diverse task descriptions; embed each via Xenova/all-MiniLM-L6-v2 (384d, local Transformers.js) — same model used at runtime so the cosine is honest; average the 15 vectors; L2-normalize. Provider chain (Anthropic → OpenAI → Gemini) auto-falls-through on quota/auth errors. .env.local autoload up the directory tree.

• windags-skills/scripts/build-tool2vec-cache.mjs

Self-contained MCP server distributed with the windags-skills plugin. As of v2.8.0 it exposes 9 tools: the four retrieval primitives (windags_skill_search/graft/reference/history with the full local cascade — BM25 + MiniLM + RRF + cross-encoder + per-user attribution k-NN + cross-user global priors) plus five planner-grade tools (skill_search_batch and skill_graft_batch for N-in-one round-trips, node_requirements with provider-native model IDs, validate_dag for schema-checking before save, and estimate_cost for planning-time cost surfacing). Reads are local + cheap aggregates from api.windags.ai (no API key); writes are anonymized telemetry, opt-out via WINDAGS_TELEMETRY=off.

• windags-skills/mcp-server/index.js
• packages/mcp-server/src/index.ts

CostCalculator returns a per-node + per-wave estimate before execution. The onCostEstimate callback in DAGExecutorOptions can return false to abort. PhaseOrchestrator forwards the gate to each phase's inner executor.

• packages/core/src/pricing/cost-calculator.ts
• packages/core/src/core/phase-orchestrator.ts

Five orthogonal signals per graft (completion, accept, downstream rating, self-confidence, shipped) keyed by query embedding in ~/.windags/skill-state.db. Stage 1 (deterministic Floor + Wall + Envelope) runs after every node. Stage 2 (Haiku-based eval) is wired but the haikuCallFn isn't configured by default.

• packages/core/src/observability/
• packages/core/src/skills/skill-quality-store.ts

ExecutionEvent → DAGStateEvent via event-mapper. WebSocket server with rooms, buffering, heartbeat, sequence numbers + replay. useDAGStateEvents hook dispatches to a Zustand reducer. Handles node_status_change, wave_transition, cost_update, human_gate_waiting, evaluation, mutation.

• packages/cli/src/ws-server.ts
• packages/ui/src/state/event-mapper.ts

Seven topologies built: DAG, Team Loop, Swarm, Blackboard, Team Builder, Recurring, Workflow. Core executors exist for all seven. /api/execute today only dispatches DAG; rolling out the others is gated on UI controls.

• packages/core/src/topologies/

Cloudflare Worker over D1 FTS5. Implements /v1/graft today (server-side BM25, returns full skill bodies). Tool2Vec via KV + Workers AI cross-encoder is the next-release plan — eliminates the 80 MB local model dependency entirely.

• apps/api/src/index.ts