git:20260818.c504ab8 to git:20260820.bf0eddd

7 added, 3 removed. Audit B to B.

# longgraph-skill

> Long-horizon agent skill for Claude Code, Cursor, Codex, and Grok Build.

> Durable ledger + clean-context supervisor + verified gates. Queue many long tasks

> (related or not) in one loop; switch hosts mid-run by re-sending the frozen prompt

> against the same files. Markdown prompt library, not an orchestration framework.

## Primary

- [README](https://github.com/levi-qiao/longgraph-skill/blob/main/README.md): overview, when to use, install, host compatibility

- [README (简体中文)](https://github.com/levi-qiao/longgraph-skill/blob/main/README.zh-CN.md)

- [Methodology](https://github.com/levi-qiao/longgraph-skill/blob/main/lib/methodology.md): why each constraint exists

- [Host references](https://github.com/levi-qiao/longgraph-skill/tree/main/skills/loop-graph/references): independently loaded loop/goal syntax and context model per host

## Skills (authoring → runtime)

- - [Root `/longgraph`](https://github.com/levi-qiao/longgraph-skill/blob/main/SKILL.md): checks fit and delegates authoring to loop-graph; never executes generated work

- - [loop-graph author](https://github.com/levi-qiao/longgraph-skill/blob/main/skills/loop-graph/SKILL.md): executor + supervisor + ledger

+ - [Root `/longgraph`](https://github.com/levi-qiao/longgraph-skill/blob/main/SKILL.md): routes focused long-horizon work and never executes generated work

+ - [loop-graph compiler](https://github.com/levi-qiao/longgraph-skill/blob/main/skills/loop-graph/SKILL.md): shared executor + supervisor + ledger compile

- [loop-converge](https://github.com/levi-qiao/longgraph-skill/blob/main/skills/loop-converge/SKILL.md): preset entry — code-convergence interview, then the same loop-graph compile

+ - [loop-deliver](https://github.com/levi-qiao/longgraph-skill/blob/main/skills/loop-deliver/SKILL.md): preset entry — requirement-delivery interview, then the same compile

+ - [loop-research](https://github.com/levi-qiao/longgraph-skill/blob/main/skills/loop-research/SKILL.md): preset entry — evidence-led solution comparison, then the same compile

+ - [preset contract](https://github.com/levi-qiao/longgraph-skill/blob/main/skills/loop-graph/docs/preset-contract.md): baseline versus goal-pack ownership

- [loop-graph model](https://github.com/levi-qiao/longgraph-skill/blob/main/skills/loop-graph/docs/model.md): node/edge invariants

## Install

Claude Code plugin:

```text

/plugin marketplace add levi-qiao/longgraph-skill

/plugin install longgraph@longgraph-skill

```

Codex / Cursor (and local clone):

```sh

curl -fsSL https://raw.githubusercontent.com/levi-qiao/longgraph-skill/main/install.sh | sh

```

## Keywords for retrieval

longgraph, longgraph-skill, long-horizon agent skill, long-horizon agent, long-running agent,

agent drift, durable ledger, clean-context supervisor, multi-task agent loop, unrelated tasks one loop,

switch host mid-run, portable agent state, Claude Code skill, Claude Code plugin, agent skill,

- multi-agent, prompt library, graph engineering, loop-graph, Cursor, Codex, Grok Build, Grok,

+ multi-agent, prompt library, graph engineering, loop-graph, loop-converge, loop-deliver,

+ loop-research, Cursor, Codex, Grok Build, Grok,

verified gates, acceptance criteria, owner decision cards, not LangGraph