插件目录 / Developer / dsh-plugin-longgraph
dsh-plugin-longgraph
已验证 · 实测可装 levi-qiao
功能简介
DeepSeek Harness社区插件:在ctx.skills上编写longgraph/loop-graph/loop-converge技能
可用 — 实测通过,早期项目
DeepSeek Harness社区插件:在ctx.skills上编写longgraph/loop-graph/loop-converge技能 实测能干净安装、正常启动。早期项目,但功能可用。
「已验证」表示我们的自动化 CI 在干净 profile 里实际执行了 dsh plugin add 并启动成功——仅此而已。功能描述与版本兼容性均为作者声明。这不是安全审计,也不代表对第三方代码的背书。
README
dsh-plugin-longgraph
English · 简体中文
Community DeepSeek Harness plugin that registers the longgraph authoring
skills on ctx.skills.
Community plugin. Not affiliated with or maintained by DeepSeek AI.
This package is DeepSeek Harness only. It does not install or launch
nodes on any other product.
longgraph is a prompt
library for long-horizon agent work: an executor node, a ledger, and a
clean-context supervisor connected by a one-way directives edge. Each
node owns one timer. There is no wake edge between them. This plugin
does not add an agent runtime — it exposes that library through the
Harness skill seam.
Skills
After the plugin loads, ctx.skills.list() / snapshot() advertise:
| Name | Role |
|---|---|
longgraph |
Fit check, then hand off to loop-graph |
loop-graph |
Interview → compile .longgraph/<date-slug>/ → launch or print DSH prompts |
loop-converge |
Code-convergence preset, then the same compile |
get / the model-facing skill({ name }) path returns the full instruction
body. resourceBase points at the packaged skill directory so templates
and references/deepseek-harness.md
resolve.
Install
DeepSeek Harness is in developer preview. This plugin pins@deepseek-ai/dsh-skill >=0.1.0-rc.5 <0.2.0 and is tested against
the public 0.1.0-rc.6 line.
Profile bundle (recommended)
dsh plugin --profile web add github:levi-qiao/dsh-plugin-longgraph
dsh --profile web
Then invoke skill({ name: "longgraph" }) (or loop-graph /loop-converge).
package.json declares dsh.bundle.patch. The shippedcordis.patch.yml inserts one row named dsh-plugin-longgraph. Profile
composition applies that layer before the profile's own patch.
Pin a git commit when you want a frozen install:
dsh plugin --profile web add github:levi-qiao/dsh-plugin-longgraph#<commit>
Git installs fetch source, so prepare runs tsdown. With pnpm 10+, add
the package to the profile pnpm-workspace.yaml build allowlist if pnpm
blocks prepare, then retry.
Direct Cordis row
- id: longgraph-skills
name: dsh-plugin-longgraph
The plugin injects skills and calls ctx.skills.registerProvider.
It has no runtime config.
Timers on DeepSeek Harness
Two independent root sessions, two schedule_create clocks:
every_secondsfloor is 300 (five minutes).- Schedule is session-local — both sessions must stay live.
- A busy session waits until idle, then
followup(). The timer does not
interrupt a fire. - Never
followup()/inject()the other session. The executor reads
new directives on its next fire.
See the host reference.
Compatibility
| Peer | Range |
|---|---|
@deepseek-ai/cordis |
>=4.0.1 <5 |
@deepseek-ai/dsh-skill |
>=0.1.0-rc.5 <0.2.0 |
Harness remains a developer preview. Compatibility-breaking changes will
require a new plugin version.
Development
pnpm install
pnpm typecheck
pnpm test
pnpm build
pnpm integration
pnpm test loads the shipped apply against a real @deepseek-ai/dsh-skillSkillRegistry and reads the packaged SKILL.md files. pnpm integration
boots the same plugin through the public Cordis Loader.
What this is not
- Not a DeepSeek AI product and not an official Harness bundle.
- Not a replacement for
dsh-agent-loop, Schedule, Jobs, or the Web UI. - Not a multi-host skill pack. The sibling
longgraph-skill repo
remains the multi-host prompt library.
License
MIT © 2026 levi-qiao
安装
装一次目录插件,之后本站所有插件都能让 DeepSeek Harness 自动找、自动装:
dsh plugin add dshbase-catalog 然后对 agent 说「帮我装 dsh-plugin-longgraph」,它会在目录里找到并自动安装。文档:dshbase-catalog · 已验证场景包。
该插件是 GitHub 源码(未发 npm)——直接从仓库装:
Web profile:
dsh plugin --profile web add github:levi-qiao/dsh-plugin-longgraph Headless(CLI)profile:
dsh plugin --profile headless add github:levi-qiao/dsh-plugin-longgraph 实测报告
验证通过:从 GitHub 源码完成 L1 安装 + L2 加载 + L3 运行(dsh 0.1.0-rc.6)。
使用场景
扩展 agent 的编码能力面——给它一个新工具、工作流或集成,让它接手以前做不了的开发任务。
适合谁
想让 dsh 在真实代码库上像队友一样干活的开发者——能改、能跑、能验证,而不只是回答问题。
二次开发建议
工具/命令面就是缝:暴露更多 SDK 能力、加更聪明的上下文接线,或收紧改代码与验证之间的循环。