longgraph plugin
A long-horizon skill for Claude Code, Cursor, Codex and Grok Build: a ledger on disk, a supervisor and verified gates.
- 79 ★ on the repository
- Last commit
Extended access
Why this level
- Nodes run long without supervision, write into the workspace and keep ledger files, and the install script downloads through curl.
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add plugins/longgraphDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- Nodes run long without supervision, write into the workspace and keep ledger files, and the install script downloads through curl.
This is third-party code. Review the repository files before installing.
What it does
longgraph is a library of skills and prompts for agent work that runs for hours and outlives one context. `/longgraph` designs a run, detects the host, inspects the workspace and asks only unresolved owner questions, then compiles a loop graph. The executor takes one verifiable item from the `ledger.md` ledger and records the result, while a supervisor in a separate context re-verifies and corrects course through the `directives.md` file. Ready packs cover common tasks: `/loop-deliver` for a feature or migration in slices with traceable acceptance, `/loop-converge` for multi-round cleanup of unused and duplicate code, and `/loop-research` for comparing approaches with experiments. State lives in `.longgraph/<date-slug>/`, so the host can be switched mid-run: open the same workspace and send the frozen node prompt.
Who it is for. Developers running migrations and refactors over many hours who want progress to survive a context reset and a change of agent.
Good fit when
- A storage migration goes in stages and a separate supervisor verifies each.
- A queue of several unrelated long tasks runs in one loop.
- A workspace moves from Claude Code to Codex by resending the frozen prompt.
- Cleanup of unused and duplicate code runs several rounds until it converges.
Not a fit when
- The goal is simple and fits a normal host task.
- You need an orchestration runtime, while longgraph remains a set of Markdown prompts.
- You expect fully unattended production work, whereas the durations in the examples are project wall-clock rather than continuous model execution.
Example request
Run /longgraph for the file storage migration: ask what needs clarifying, compile the graph and build a ledger with stages.Limitations
longgraph remains a library of Markdown prompts rather than a runtime. Codex and Claude Code can create both nodes directly, while Grok Build, Cursor and shell or cron work in a prompts-only mode where you start nodes by hand. The install script downloads through `curl`, and on Grok Build it has no wake link between nodes. The authors rule: if you do not need the graph, use a normal host task. The ledger has a single writer, and the supervisor shares no context with the executor and does not edit its scoreboard. Invocation syntax depends on the host, and only progress is portable.
How to disable. Run `/plugin uninstall longgraph@longgraph-skill` in Claude Code. For other hosts remove the symlinks for `/longgraph`, `/loop-converge`, `/loop-deliver` and `/loop-research`. Delete the `.longgraph/` run ledger directory by hand.
FAQ
How do I install it in Claude Code?
Run `/plugin marketplace add levi-qiao/longgraph-skill`, then `/plugin install longgraph@longgraph-skill`.
How do I install it in Codex or Cursor?
Run the `install.sh` script through `curl ... | sh` or from a local clone: it installs the library and symlinks `/longgraph` and related commands.
When is longgraph not needed?
When the goal is simple and fits a normal host task. The authors advise skipping the graph if you do not need it.
How do I continue on another host?
Open the workspace with the `.longgraph/` directory and resend the frozen executor prompt, and the supervisor one if used. The node reads the ledger and continues from the next open item.
How do loop-deliver, loop-converge and loop-research differ?
`/loop-deliver` takes a requirement in slices with acceptance proof, `/loop-converge` cleans unused and duplicate code over several rounds, and `/loop-research` compares approaches by evidence and experiments.
Discussion
No questions or reviews yet.
Related
PluginsA Codex skill for an autonomous loop: Claude Code computes, Codex reviews hourly, critiques and sends the next prompt.
A CLI that keeps an existing agent on a long task: plan a step, run it, verify the result in the real environment and checkpoint progress
The Nightshift plugin gives an agent a shift with a work contract: task list, budgets, evidence and a gate until the job is done.
A hundred ready loops for Claude Code, Cursor and Codex: the agent acts, runs a check and fixes the cause until the goal is met.
A Codex plugin of five independent meta-skills and one tldraw skill: improves prompts, goals, research, plans and commits.
An MCP bridge: Claude Code hands heavy tasks to the Antigravity CLI and only the finished answer returns to Claude context
README in short
The README gives a table of verifiable examples, when to use it, how to choose the right tool, a quick start, the executor, supervisor and scout roles, host compatibility and a repository map.