dsh-knowledge-graph
No description
- Stars
- 1
- Language
- JavaScript
- Created
- Sep 1, 2026
- Updated
- Sep 1, 2026
Introduction
dsh-knowledge-graph · DSH 知识图谱插件
把「你装了哪些插件 / 技能 / 工具、它们之间是什么关系、以及每轮对话里到底调用了什么、想了什么」全部可视化出来。
Visualize what plugins / skills / tools you have installed, how they relate to each other, and exactly what each conversation turn called and reasoned about.
它解决了什么问题 · Why this plugin
中文
DeepSeek Harness 是「一切皆插件」。用久了之后,你会遇到这些麻烦:
- 装了什么心里没数 —— 插件、技能、工具越装越多,几十上百个,根本记不清自己装了啥、哪个是干嘛的。
- 看不见依赖关系 —— 插件之间有
peerDependencies(依赖)和「全家桶包含」关系,少了某个依赖,功能莫名其妙就坏了,但你不知道哪里断了。 - 不知道 AI 到底干了什么 —— 每轮对话里,助手调用了哪些工具、思考了什么、最后怎么回复的,全是黑盒;想排查问题或理解行为只能靠猜。
- 装了一堆「吃灰」的东西 —— 有些插件/工具装上后从没用过,白白占着,还干扰判断。
这个插件把这四件事一次性解决:一张图谱看懂全局,一条轨迹看懂每一轮。
English
DeepSeek Harness treats everything as a plugin. As your environment grows, you run into these problems:
- No overview of what's installed — plugins, skills and tools pile up to dozens or hundreds; you can't remember what you installed or what each one does.
- Invisible dependencies — plugins have
peerDependenciesand "bundle contains" relations. A missing dependency breaks things silently, and you can't tell where the chain is broken. - A black-box agent — which tools the assistant called, what it reasoned, and how it finally replied each turn is opaque; debugging or understanding its behavior becomes guesswork.
- Dead-weight plugins — some plugins/tools were installed but never used, cluttering the environment and your judgment.
This plugin solves all four at once: one graph to see the whole picture, one trace to understand every turn.
功能特性 · Features
中文
- 力导向图谱:插件 / 技能 / 工具以节点呈现,丝滑物理仿真布局;依赖(
depends)以连线表示,「全家桶包含关系」(contains)用凸包气泡圈成一簇,一眼看清哪些是一家人。 - 清单视图:插件 / 技能 / 工具三个页签,支持按名称搜索、按图层过滤,适合精确查找。
- 每轮思考过程:选中某个对话轮次后,高亮该轮调用的工具(序号徽标 + 调用箭头),并把思考(💭)与回复(💬)画成气泡;点气泡看全文、点空白收起;点工具节点,它对应的思考气泡会一起高亮。
- 使用统计:每个节点显示「用过几次、最近什么时候用的、最近在哪些轮次被调用」;「上一轮调用」一键聚焦最新一轮。
- 健康诊断:加载失败 🔴 / 被禁用 ⚪ 在节点外环标出,一眼定位坏掉的插件。
- 导出报告:一键导出 Markdown 报告(概览 / 问题插件 / 吃灰清单 / 使用排行)。
- 本地持久化:统计与历史保存在本地(
~/.dsh/kg-knowledge-graph.json),重启不丢。
English
- Force-directed graph: plugins / skills / tools as nodes with a smooth physics layout;
dependsedges drawn as lines,contains(bundle) relations drawn as convex-hull bubbles. - List view: three tabs (plugins / skills / tools) with search and layer filtering for precise lookup.
- Per-turn thinking: pick a turn to highlight the tools it called (sequence badges + call arrows), with reasoning (💭) and reply (💬) drawn as bubbles; click a bubble for full text, click empty space to collapse; clicking a tool node also highlights its related bubbles.
- Usage stats: per-node call count, last-used time, and recent turns it was called in; "上一轮调用" focuses the latest turn.
- Health diagnosis: failed 🔴 / disabled ⚪ plugins marked with colored outer rings.
- Export report: one-click Markdown report (overview / problem plugins / dormant list / usage ranking).
- Local persistence: stats and history stored locally (
~/.dsh/kg-knowledge-graph.json), surviving restarts.
使用场景 · Use Cases
中文
- 盘点 / 审计:想知道环境里装了多少插件、技能、工具,哪些加载失败、哪些被禁用 —— 打开图谱一目了然。
- 排查依赖问题:功能不工作,怀疑依赖断了 —— 开「全显依赖」看关系边,或看某插件的依赖气泡。
- 复盘对话:想知道 AI 某一轮「调用了哪些工具、怎么一步步思考的」—— 用轮次选择器回看任意一轮的思考轨迹。
- 清理「吃灰」:找出从没用过的插件/工具,卸载瘦身 —— 用「吃灰清单」或导出报告。
- 学习 DSH 生态:新手快速了解 DSH 有哪些能力、工具长什么样、中文用途说明。
English
- Audit your environment: see how many plugins / skills / tools are installed, which failed to load, which are disabled.
- Debug dependencies: when something breaks, open "全显依赖" to inspect relation edges or a plugin's dependency bubble.
- Review conversations: understand which tools a turn called and how it reasoned, step by step, via the turn selector.
- Clean up dormant items: find never-used plugins/tools and uninstall them.
- Learn the DSH ecosystem: quickly see what capabilities DSH offers, with Chinese descriptions.
安装 · Install
dsh plugin --profile web add dsh-knowledge-graph
快速上手 · Quick Start
中文
- 安装后,在右侧栏打开「插件图谱」页签(无 web-ui 侧栏时自动退化为右下角悬浮按钮)。
- 默认是「图谱」视图:滚轮缩放、拖拽平移、点节点看详情。
- 点底部「上一轮调用」或下拉框选一个轮次,查看该轮思考气泡与调用轨迹。
English
- After installing, open the "插件图谱" tab in the right sidebar (falls back to a floating button without web-ui).
- The default "graph" view: scroll to zoom, drag to pan, click a node for details.
- Click "上一轮调用" or pick a turn from the dropdown to view that turn's thinking bubbles and call trace.
详细使用说明 · Usage
中文
- 顶部工具栏:搜索框(过滤节点)、
上一轮调用(聚焦最新一轮)、全显依赖(显示/隐藏依赖边)、图层切换(插件/技能/工具)、刷新(重新枚举)、导出报告(下载 Markdown)。 - 图谱交互:滚轮缩放、按住空白拖拽平移、点节点选中看详情(用途 / 次数 / 最近使用 / 问题诊断)、
重置视图回到全图。 - 轮次选择器(图谱底部):
‹ 上一轮 / 下拉选轮 / 下一轮 ›;选中轮次后,该轮工具高亮 + 序号 + 箭头,思考/回复以气泡显示,点气泡看全文。 - 图例:图谱顶部图例说明节点颜色、连线类型、问题外环的含义。
English
- Toolbar: search box,
上一轮调用(focus latest turn),全显依赖(toggle dependency edges), layer switch,刷新(re-enumerate),导出报告(download Markdown). - Graph interaction: scroll to zoom, drag empty space to pan, click a node to inspect details,
重置视图to reset view. - Turn selector (bottom):
‹ prev / dropdown / next ›; selecting a turn highlights its tools with sequence badges and arrows, shows reasoning/reply bubbles, and clicking a bubble expands full text. - Legend: top legend explains node colors, edge types, and problem rings.
兼容性与权限 · Compatibility & Permissions
中文
- 平台:Web(
platform: web),适配带侧栏的 web-ui,无 web-ui 时自动退化悬浮面板。 - 宿主侧:只读枚举插件/技能/工具的清单与 schema、监听工具调用与技能变化、读取当前会话事件以重建每轮思考轨迹;向本地
~/.dsh/kg-knowledge-graph.json写入使用统计;注册/kg-api/*本地接口;写~/.dsh/kg-debug.log诊断日志。 - 客户端:注册侧栏页签 / 悬浮面板,通过
fetch访问本机/kg-api/*接口。 - 无遥测:不向任何外部服务发送数据,数据只留在本机。
- 外部服务:无(不联网)。
English
- Platform: Web (
platform: web); adapts to web-ui sidebar, falls back to a floating panel without it. - Host: read-only enumeration of plugin/skill/tool inventories and schemas, listens to tool-call and skill changes, reads session events to rebuild per-turn traces; writes usage stats to local
~/.dsh/kg-knowledge-graph.json; registers local/kg-api/*routes; writes~/.dsh/kg-debug.log. - Client: registers a sidebar tab / floating panel; fetches local
/kg-api/*endpoints. - No telemetry: nothing is sent to external services; all data stays local.
- External services: none (fully offline).
依赖 · Dependencies
- 宿主侧 · Host:
@deepseek-ai/cordis - 客户端 · Client:
react(^18.2.0)
License
MIT