taskflow plugin
Declarative DAG workflows for coding agents: isolated phases, graph verification, resume and replay of runs.
- 80 ★ on the repository
- Last commit
- MIT
Extended access
Why this level
- Phases run subagents and scripts, and writes to declared paths change project files.
Install
- Works in
Manual install
npm install --global pi-taskflow@1.0.0The first install command in the README. Codex has `codex-taskflow` and Claude Code has `claude-taskflow`; connect the plugin and MCP following the host guide.
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- Phases run subagents and scripts, and writes to declared paths change project files.
This is third-party code. Review the repository files before installing.
What it does
taskflow describes agent work as a graph, verifies it before a run and executes phases in isolation. Intermediate results stay inside the runtime, and only the final output returns to the main conversation. There are 12 phase types, including `agent`, `parallel`, `map`, `reduce`, `script`, `gate`, `approval`, `loop`, `tournament` and `race`, with budgets, retries and timeouts. Version 1.0 also lets a graph declare permitted file effects: a write to a declared path goes through a transaction with rollback. Claude Code and Codex get a plugin and an MCP server with 20 `taskflow_*` tools, among them `taskflow_plan`, `taskflow_verify`, `taskflow_trace`, `taskflow_replay` and `taskflow_recompute`. Pi, OpenCode, Grok Build and Hermes Agent connect through their own adapters.
Who it is for. Developers who run agents across a whole repository, dozens of files or a multi-step process and want to see what will run and what it will cost.
Good fit when
- A codebase audit is split into parallel subtasks with one combined result.
- You need a plan before running: which phases will go and how many agent calls the worst case needs.
- A failed run resumes from one phase with a different task or model.
- The best of several variants is picked with a `tournament` phase.
Not a fit when
- The task fits into one request to an agent.
- You need an OS-level sandbox, which taskflow does not provide.
- The runtime is older than Node.js 22.19.
Example request
Split the audit of this repository into parallel subtasks by module, show the plan through `taskflow_plan` and run it after my confirmation.Limitations
Node.js 22.19 or newer is required. The `effects[]` field is a declaration and validation surface rather than an OS sandbox: a write to an undeclared path depends on host policy. In Claude Code, phases that write are rejected by default because headless mode has no sandbox, and they are allowed explicitly and only for trusted flows. Plugin and MCP install details live in the host guides on the project site. The web control console runs locally over a Unix socket, and Windows support there is not yet declared. This is version 1.0.0, and the verified scope is listed in the repository acceptance scoreboard.
How to disable. Remove the plugin with `/plugin uninstall` in Claude Code or delete the taskflow MCP entry from the agent configuration. Global packages come off with `npm uninstall --global`. Stored runs live in the project directory and are deleted by hand.
FAQ
How do I connect taskflow to Claude Code?
The `claude-taskflow` package provides the MCP server and subagent runner, while the plugin ships through the repository marketplace. Steps are in the host guide on the project site.
How many agent calls does a run cost?
`taskflow_plan` answers: it shows phase order, argument binding and the worst-case call bound with no tokens and no execution.
Can a failed run be resumed?
Yes. `taskflow_resume` forks a child run from the failed phase, optionally with a different task, model or timeouts.
Can I see what the subagents did?
By default only the final output reaches the main chat. Read intermediate work with `taskflow_peek` and `taskflow_trace`.
What are declared effects?
A phase states in advance which path it may write. The runtime checks labels and overlaps, stages the write and commits or rolls it back. `taskflow_why_effect` explains the decision.
Discussion
No questions or reviews yet.
Related
PluginsA visual workflow builder for agents: draw a flow on a canvas and export it as skills and commands for Claude Code, Codex, Cursor and others
The /codex-workflows skill: Claude Code writes a script, dozens of Codex agents run it, and the run shows on a map
Maestro-Flow builds a chain from analysis to tests for a goal and runs Claude Code, Codex, Gemini and other agents through it
The Roboflow plugin gives an agent computer vision skills: data, training, model evaluation, inference and platform API calls.
Frink runs up to four agents in their own worktrees and turns one sentence into a Flow automation with its own editor skill.
A runtime for JavaScript workflows that hand work to dozens of Codex, Gemini, Qwen, Kimi and Claude Code agents.
README in short
The README sets out the declared-effect contract, 12 phase types, zero-token verification and replay operations, the list of host adapters and states the security boundaries plainly.