@sylphx/workflow-engine

The reusable, zero-dependency core for governed agent workflows — the SOTA "second mind" that runs plan / analyze / research / discuss / council / review / execute on a model outside the host's own, and returns a governed, agent-ready artifact. Bun/Node stdlib only (ADR-0011).

Two entry points:

• @sylphx/workflow-engine — the workflow SDK. createWorkflowEngine(config, effects?) returns an engine you drive directly (no MCP server). Also re-exports the input/envelope types + validators, the per-workflow tool defs, parseBackendRegistry, and the dispatchToolCall primitive (so any transport can host the same dispatch).
• @sylphx/workflow-engine/review-core — the pure review-gate protocol brain (arg parsing, verdict extraction/validation, the convergence ledger + burn-down, multi-pass discovery aggregation, recall telemetry, discovery lenses). All I/O lives in the consumers, never here.

Consumed by @sylphx/workflow-mcp (the stdio MCP server) and @sylphx/gpt-review (the review gate CLI).

import { createWorkflowEngine, type EngineConfig } from "@sylphx/workflow-engine";