agent-dag-workflow

Most agent pipelines are stitched together ad hoc: a Tool call, an MCP endpoint, a Skill, linked by human context. Orchestration platforms like Coze or Dify solve the flow problem but typically bundle model providers, credential vaults, and tool marketplaces that many teams do not need. agent-dag-workflow takes a narrower path. It is a lightweight, host-neutral DAG workflow runtime that turns the tools, skills, and MCP endpoints you already run into a single WorkflowTemplate JSON you can save, validate, recover, audit, and replay. Your Host keeps ownership of agents, tools, and permissions; the Runtime handles compilation, scheduling, journaling, and replay.

A few design choices set it apart. One template, many entry points: SDK, CLI, a fixed six-tool MCP Gateway, DSH adapter, Canvas, and Triggers all speak the same JSON and call the same Runtime. Permissions only narrow: template-level requires, node-level fixed dependencies, Authority, and Host policy are intersected; core.script nodes have no network, no file access, and no eval; every external result must pass lossless-JSON, schema, and size checks before it enters the Artifact or Checkpoint store. Recovery semantics are explicit- inspect (read-only), recorded (reuse committed upstream results, recompute deterministic downstream), and live (re-invoke external capabilities)- with atomic Journal and Checkpoint commits and reproducible runs. The compiler performs branch-path dominance checks at publish time, rejecting any template where an activated path would inevitably lack required data.

This fits teams that already operate Codex, DeepSeek Harness, or another Agent runtime with a working set of Tools, Skills, and MCP endpoints and want to compose them into deterministic multi-step flows (batch contract review, weekly report generation, multi-source retrieval with ranked output) without standing up an entire platform. Requires Node.js 22.19+; MIT licensed; single-process by design, with horizontal scaling left to the Host when needed.