call-map · 调用可视化

中文｜ English

中文

一个 DeepSeek Harness（DSH）插件：在会话里多看一页，把「这一轮到底调用了哪些技能、哪些 MCP、派了哪些子代理，以及它们是被提示词和思维链里的哪句话带出来的」画成一张可缩放的分支图。

它不埋点、不上传：所有信息都来自本机会话日志。

它解决什么

DSH 自带两个会话视图：对话（消息流）和 轨迹（时间轴泳道：输入 / 模型 / 工具）。轨迹能回答「什么时候、花了多久」，但回答不了：

这条工具调用是哪个技能带出来的？
我配了 5 个 MCP，本会话用了哪几个、哪些一次没碰？
这次派给子代理的活，子代理内部又调了什么？
一句话里，是哪个词把它引到这一步的？

调用可视化 就是补这一层：归因，而不是时序。

树状图可缩放：一页看完整场对话

全轮次聚合的可缩放架构图

所有轮次按时间顺序连续排在同一张画布上，不再靠标签页切换：

滚轮缩放（30%–240%），按住左键拖拽平移，右下角实时显示当前比例；
进入这个页签时默认定位到最新一轮，并按画布宽度自适应缩放，一眼看到「现在跑到哪了」；
右下角 最新 一键定位最后一轮对话，全览 把整棵树缩进一屏；

能看到详细的子代理 / 技能调用

子代理与技能调用明细

树会一直展开到最细一层，左侧是链路，右侧是同源的调用明细：

技能（绿）：这一轮加载了哪些技能，各命中几次；
派发（紫）：spawn_teammate / subagent / workflow / ralph 本身也是节点，节点下面挂子代理卡片——插件会去读那个子代理会话自己的技能 / MCP 调用并折叠进来，点卡片直达该子会话；
MCP（琥珀）：某个技能激活期间实际调用的 MCP 工具，带次数（如 windows-mcp · Click ×6）；
提示词 / 思维链（蓝 / 紫蓝）：正文里与下层节点同名的词按目标着色，所以「哪个词引出了哪个节点」是肉眼可追的；
右侧明细表：类型 / 名称·工具 / 次数 / 轮次 + 触发线索 / 结果，可筛 全部 技能 MCP 未使用，点行打开宿主原生调用详情。

功能

能力	说明
一页聚合	所有轮次按时间顺序连续排布，不用来回切页
可缩放画布	滚轮缩放（30%–240%）、拖拽平移、`最新` / `全览` 一键归位；进入默认落在最新一轮
归因树	`提示词 → 思维链 → 技能 / MCP / 子代理`，命中的词按目标节点上色（技能绿、MCP 琥珀、子代理紫）
子代理聚合	派发节点独立成类；能拿到子会话时，把子会话内部的技能/MCP 折叠成筹码挂在下面，点卡片直达该子会话
调用明细表	类型 / 名称 / 次数 / 轮次 / 触发线索 / 结果，可筛「全部 / 技能 / MCP / 未使用」，点行打开宿主原生调用详情
未使用清单	本会话可用但一次没调用的技能与 MCP（`未使用` 筛选即可看到）
轨道图	输入 / 模型 / 工具三泳道时间条，用于快速定位「哪一段在干什么」

安装

要求：DSH 已安装并可用（本插件注册的是既有槽位 conversation.view，不改宿主编译产物）。

方式一：DSH 内置插件管理器（推荐）

把本目录作为本地包安装：

plugin_manager install_bundle  target = <本目录的绝对路径>

安装后刷新页面，会话头部会多出第三个视图标签 调用可视化。

也可以直接从 GitHub 安装：本仓库已声明 dsh.bundle manifest，可被 dsh plugin add 识别（纯 JS，无构建步骤）。插件已按 awesome-dsh-plugin 的收录规范备好元数据：assets/icon-180.png 为市场卡片图标，screenshots.json 声明商店详情页截图。

方式二：手动挂载

把本目录放到任意位置；
在 profile 目录执行 pnpm add <本目录绝对路径>（或把它软链进 node_modules/call-map）；
在 profile 的 package.json 的 dsh.profile.bundles 数组里加入 call-map；
重启 / 刷新。

卸载：plugin_manager remove_bundle target = call-map。

使用

打开任意会话，点头部 调用可视化；
画布里滚轮缩放、按住左键拖拽平移（单击不会误触发拖动）；
右下角 最新 回到最新一轮，全览 缩到整张图；
右侧表格点行 → 打开该次调用的宿主原生详情；点子代理卡片 → 跳到对应子会话。

数据来源与口径

所有数据来自本地会话事件（user/message、assistant/message、tool/call、tool/result、request/header、turn/start），插件自身不写会话、不联网：

技能：skill 工具的调用，技能名取自其 arguments.name；
MCP：工具名形如 mcp__<server>__<tool> 的调用；
子代理：spawn_teammate / subagent / subagent_fork / workflow / ralph 的派发调用；
未使用清单：已注册技能目录（取自系统提示词中的 <available_skills>）与已向本会话暴露工具的 MCP server，减去实际调用的部分；
触发线索：提示词 / 思维链里与调用同名的词面片段。

已知限制

触发线索是词面匹配，不是因果证明。文字里没出现同名词就不标，不臆造归因；需要语义级解释应另加一次模型调用。
跨会话（子代理）聚合依赖能拿到子会话 id：subagent / workflow 的结果里带 id，可直接聚合；spawn_teammate 只返回队友名，插件用「工作区会话标题 == 队友名」反查，队友会话被改名或尚未进入快照时会退化为「会话未加载」。
只读当前会话的事件窗口：极长会话若早期事件未加载，早期轮次的提示词可能显示为「不在已加载的事件窗口里」。
本插件的轨道图是自己重绘的，与宿主 轨迹 页的原版图表不是同一实现；宿主的 轨迹 页本身不受影响。
宿主 轨迹 页的工具栏（时长 / 轮次 / 调用那一行）没有对外插槽，所以本插件只能作为并列的第三个页签，无法把入口塞进那一行。

隐私

不采集、不发送任何数据。全部计算在本地前端完成，不新增网络请求。

目录结构

call-map/
├─ package.json        # 包清单（dsh.bundle / dsh.client 声明）
├─ cordis.patch.yml    # bundle 补丁：插入一行插件
├─ index.js            # Host 半边（本插件为空实现）
├─ client.js           # 全部逻辑：折叠会话事件 + 视图渲染
├─ locale/en.json      # 插件管理页显示名（英文）
├─ locale/zh.json      # 插件管理页显示名（中文）
├─ docs/*.png          # README 截图
├─ assets/icon-180.png # 市场卡片图标（180×180 PNG）
├─ screenshots.json    # 市场详情页截图声明（1–8 张）
└─ icon.svg            # 插件图标

English

A DeepSeek Harness (DSH) plugin that adds one more page to a session: it draws which skills, which MCP tools, and which subagents were actually used this turn — and which sentence or word in the prompt / reasoning brought them in — as a zoomable branch diagram.

No instrumentation, no uploads: every number comes from the local session log.

What it solves

DSH ships two session views: Chat (the message stream) and Trajectory (a timing overview with input / model / tool lanes). Trajectory answers "when, and for how long". It cannot answer:

Which skill produced this tool call?
I configured five MCP servers — which ones did this session actually use, and which were never touched?
What did the subagent I dispatched actually do inside?
In that prompt, which word is responsible for this step?

Call map adds exactly that layer: attribution, not chronology.

Zoomable tree: the whole conversation on one page

All turns aggregated on a zoomable canvas

Every turn is laid out continuously on a single canvas, instead of one tab per turn:

Wheel to zoom (30%–240%), hold the left button to pan; the live scale sits in the bottom-right corner;
Opening the tab starts at the latest turn and fits the canvas to the available width, so you immediately see where the work is;
Latest jumps back to the newest turn, Overview shrinks the whole tree into one screen;
A left click never starts a drag: panning engages only after the pointer travels 4px, so clicking a card is just a click.

Detailed subagent / skill calls

Subagent and skill call detail

The tree expands all the way down; the chain is on the left, the matching call table on the right:

Skill (green): which skills this turn loaded, and how often each was hit;
Dispatch (purple): spawn_teammate / subagent / workflow / ralph are nodes too, with subagent cards underneath — the plugin reads that child session's own skills / MCP calls and folds them in; click a card to jump to the child session;
MCP (amber): the MCP tools called while a skill was active, with counts (e.g. windows-mcp · Click ×6);
Prompt / reasoning (blue / indigo): words in the text that match a lower node are coloured by their target, so "which word led to which node" is visible at a glance;
Right-hand table: type / name·tool / count / turn + trigger trail / result, filterable by All Skills MCP Unused; click a row to open the host's native call inspector.

Features

Capability	Description
One-page aggregation	Every turn laid out chronologically on one canvas; no tab switching
Zoomable canvas	Wheel zoom (30%–240%), drag to pan, `Latest` / `Overview` shortcuts; opens on the latest turn
Attribution tree	`Prompt → reasoning → skill / MCP / dispatch`, with matched words coloured per target node
Subagent aggregation	Dispatch is its own node class; when a child session is reachable its skills / MCP are folded in as chips, and the card opens that session
Call table	Type / name / count / turn / trigger trail / result, filter by all / skills / MCP / unused; click a row for the host's native inspector
Unused inventory	Skills and MCP servers available in this session but never called (the `Unused` filter)
Lane chart	Input / model / tool time lanes for "what was happening when"

Install

Requires a working DSH install. The plugin registers into the existing conversation.view slot and does not modify any shipped build output.

Option 1 — DSH plugin manager (recommended)

Install this directory as a local package:

plugin_manager install_bundle  target = <absolute path to this directory>

Then reload the page: a third view tab, Call map, appears next to Chat and Trajectory.

You can also install straight from GitHub: the repo declares the dsh.bundle manifest, so dsh plugin add can pick it up (plain JS, no build step). Marketplace metadata follows the awesome-dsh-plugin conventions: assets/icon-180.png is the storefront tile and screenshots.json declares the store screenshots.

Option 2 — manual mount

Put this directory anywhere;
In the profile directory run pnpm add <absolute path> (or symlink it into node_modules/call-map);
Add call-map to the dsh.profile.bundles array in the profile's package.json;
Restart / reload.

Uninstall: plugin_manager remove_bundle target = call-map.

Usage

Open any session and click Call map in the header;
Wheel to zoom, hold the left button to pan (a click never starts a drag);
Latest returns to the newest turn, Overview fits the whole tree;
Click a table row to open the host's native call inspector; click a subagent card to jump to that child session.

Data sources and definitions

Everything comes from local session events (user/message, assistant/message, tool/call, tool/result, request/header, turn/start). The plugin never writes to a session and never touches the network:

Skill: a skill tool call, named by its arguments.name;
MCP: a tool call whose name matches mcp__<server>__<tool>;
Subagent: a dispatch call to spawn_teammate / subagent / subagent_fork / workflow / ralph;
Unused inventory: the registered skill catalogue (from the <available_skills> block in the system prompt) plus MCP servers that exposed tools to this session, minus what was actually called;
Trigger trail: the lexical fragment in the prompt / reasoning that matches the call.

Known limitations

The trigger trail is lexical matching, not proof of causality. If no matching word appears in the text, nothing is highlighted — the plugin does not invent attribution. Semantic attribution would need an extra model call.
Cross-session (subagent) aggregation depends on getting a child session id: subagent / workflow results carry one; spawn_teammate only returns a teammate name, so the plugin looks it up by "workspace session title == teammate name" and degrades to "session not loaded" when that fails.
Only the loaded event window is read: in a very long session, early turns may show "the prompt of this turn is outside the loaded event window".
The lane chart is re-drawn by this plugin and is not the shipped Trajectory implementation; the host's Trajectory page is untouched.
The host Trajectory toolbar (the duration / turns / calls row) exposes no slot to plugins, so this plugin can only be a parallel third tab — the entry point cannot live inside that row.

Privacy

No collection, no transmission. All computation happens in the local front-end, with no additional network requests.

Layout

call-map/
├─ package.json        # package manifest (dsh.bundle / dsh.client)
├─ cordis.patch.yml    # bundle patch: inserts one plugin row
├─ index.js            # Host half (empty for this plugin)
├─ client.js           # all logic: session-event folding + view rendering
├─ locale/en.json      # plugin-page display name (English)
├─ locale/zh.json      # plugin-page display name (Chinese)
├─ docs/*.png          # README screenshots
├─ assets/icon-180.png # storefront tile (180x180 PNG)
├─ screenshots.json    # store screenshots (1-8 images)
└─ icon.svg            # plugin icon

License

MIT

call-map · 调用可视化

中文｜ English

中文

它不埋点、不上传：所有信息都来自本机会话日志。

它解决什么

DSH 自带两个会话视图：对话（消息流）和 轨迹（时间轴泳道：输入 / 模型 / 工具）。轨迹能回答「什么时候、花了多久」，但回答不了：

这条工具调用是哪个技能带出来的？
我配了 5 个 MCP，本会话用了哪几个、哪些一次没碰？
这次派给子代理的活，子代理内部又调了什么？
一句话里，是哪个词把它引到这一步的？

调用可视化 就是补这一层：归因，而不是时序。

树状图可缩放：一页看完整场对话

全轮次聚合的可缩放架构图

所有轮次按时间顺序连续排在同一张画布上，不再靠标签页切换：

滚轮缩放（30%–240%），按住左键拖拽平移，右下角实时显示当前比例；
进入这个页签时默认定位到最新一轮，并按画布宽度自适应缩放，一眼看到「现在跑到哪了」；
右下角 最新 一键定位最后一轮对话，全览 把整棵树缩进一屏；

能看到详细的子代理 / 技能调用

子代理与技能调用明细

树会一直展开到最细一层，左侧是链路，右侧是同源的调用明细：

技能（绿）：这一轮加载了哪些技能，各命中几次；
派发（紫）：spawn_teammate / subagent / workflow / ralph 本身也是节点，节点下面挂子代理卡片——插件会去读那个子代理会话自己的技能 / MCP 调用并折叠进来，点卡片直达该子会话；
MCP（琥珀）：某个技能激活期间实际调用的 MCP 工具，带次数（如 windows-mcp · Click ×6）；
提示词 / 思维链（蓝 / 紫蓝）：正文里与下层节点同名的词按目标着色，所以「哪个词引出了哪个节点」是肉眼可追的；
右侧明细表：类型 / 名称·工具 / 次数 / 轮次 + 触发线索 / 结果，可筛 全部 技能 MCP 未使用，点行打开宿主原生调用详情。

功能

能力	说明
一页聚合	所有轮次按时间顺序连续排布，不用来回切页
可缩放画布	滚轮缩放（30%–240%）、拖拽平移、`最新` / `全览` 一键归位；进入默认落在最新一轮
归因树	`提示词 → 思维链 → 技能 / MCP / 子代理`，命中的词按目标节点上色（技能绿、MCP 琥珀、子代理紫）
子代理聚合	派发节点独立成类；能拿到子会话时，把子会话内部的技能/MCP 折叠成筹码挂在下面，点卡片直达该子会话
调用明细表	类型 / 名称 / 次数 / 轮次 / 触发线索 / 结果，可筛「全部 / 技能 / MCP / 未使用」，点行打开宿主原生调用详情
未使用清单	本会话可用但一次没调用的技能与 MCP（`未使用` 筛选即可看到）
轨道图	输入 / 模型 / 工具三泳道时间条，用于快速定位「哪一段在干什么」

安装

要求：DSH 已安装并可用（本插件注册的是既有槽位 conversation.view，不改宿主编译产物）。

方式一：DSH 内置插件管理器（推荐）

把本目录作为本地包安装：

plugin_manager install_bundle  target = <本目录的绝对路径>

安装后刷新页面，会话头部会多出第三个视图标签 调用可视化。

方式二：手动挂载

把本目录放到任意位置；
在 profile 目录执行 pnpm add <本目录绝对路径>（或把它软链进 node_modules/call-map）；
在 profile 的 package.json 的 dsh.profile.bundles 数组里加入 call-map；
重启 / 刷新。

卸载：plugin_manager remove_bundle target = call-map。

使用

打开任意会话，点头部 调用可视化；
画布里滚轮缩放、按住左键拖拽平移（单击不会误触发拖动）；
右下角 最新 回到最新一轮，全览 缩到整张图；
右侧表格点行 → 打开该次调用的宿主原生详情；点子代理卡片 → 跳到对应子会话。

数据来源与口径

所有数据来自本地会话事件（user/message、assistant/message、tool/call、tool/result、request/header、turn/start），插件自身不写会话、不联网：

技能：skill 工具的调用，技能名取自其 arguments.name；
MCP：工具名形如 mcp__<server>__<tool> 的调用；
子代理：spawn_teammate / subagent / subagent_fork / workflow / ralph 的派发调用；
未使用清单：已注册技能目录（取自系统提示词中的 <available_skills>）与已向本会话暴露工具的 MCP server，减去实际调用的部分；
触发线索：提示词 / 思维链里与调用同名的词面片段。

已知限制

触发线索是词面匹配，不是因果证明。文字里没出现同名词就不标，不臆造归因；需要语义级解释应另加一次模型调用。
跨会话（子代理）聚合依赖能拿到子会话 id：subagent / workflow 的结果里带 id，可直接聚合；spawn_teammate 只返回队友名，插件用「工作区会话标题 == 队友名」反查，队友会话被改名或尚未进入快照时会退化为「会话未加载」。
只读当前会话的事件窗口：极长会话若早期事件未加载，早期轮次的提示词可能显示为「不在已加载的事件窗口里」。
本插件的轨道图是自己重绘的，与宿主 轨迹 页的原版图表不是同一实现；宿主的 轨迹 页本身不受影响。
宿主 轨迹 页的工具栏（时长 / 轮次 / 调用那一行）没有对外插槽，所以本插件只能作为并列的第三个页签，无法把入口塞进那一行。

隐私

不采集、不发送任何数据。全部计算在本地前端完成，不新增网络请求。

目录结构

call-map/
├─ package.json        # 包清单（dsh.bundle / dsh.client 声明）
├─ cordis.patch.yml    # bundle 补丁：插入一行插件
├─ index.js            # Host 半边（本插件为空实现）
├─ client.js           # 全部逻辑：折叠会话事件 + 视图渲染
├─ locale/en.json      # 插件管理页显示名（英文）
├─ locale/zh.json      # 插件管理页显示名（中文）
├─ docs/*.png          # README 截图
├─ assets/icon-180.png # 市场卡片图标（180×180 PNG）
├─ screenshots.json    # 市场详情页截图声明（1–8 张）
└─ icon.svg            # 插件图标

English

No instrumentation, no uploads: every number comes from the local session log.

What it solves

DSH ships two session views: Chat (the message stream) and Trajectory (a timing overview with input / model / tool lanes). Trajectory answers "when, and for how long". It cannot answer:

Which skill produced this tool call?
I configured five MCP servers — which ones did this session actually use, and which were never touched?
What did the subagent I dispatched actually do inside?
In that prompt, which word is responsible for this step?

Call map adds exactly that layer: attribution, not chronology.

Zoomable tree: the whole conversation on one page

All turns aggregated on a zoomable canvas

Every turn is laid out continuously on a single canvas, instead of one tab per turn:

Wheel to zoom (30%–240%), hold the left button to pan; the live scale sits in the bottom-right corner;
Opening the tab starts at the latest turn and fits the canvas to the available width, so you immediately see where the work is;
Latest jumps back to the newest turn, Overview shrinks the whole tree into one screen;
A left click never starts a drag: panning engages only after the pointer travels 4px, so clicking a card is just a click.

Detailed subagent / skill calls

Subagent and skill call detail

The tree expands all the way down; the chain is on the left, the matching call table on the right:

Skill (green): which skills this turn loaded, and how often each was hit;
Dispatch (purple): spawn_teammate / subagent / workflow / ralph are nodes too, with subagent cards underneath — the plugin reads that child session's own skills / MCP calls and folds them in; click a card to jump to the child session;
MCP (amber): the MCP tools called while a skill was active, with counts (e.g. windows-mcp · Click ×6);
Prompt / reasoning (blue / indigo): words in the text that match a lower node are coloured by their target, so "which word led to which node" is visible at a glance;
Right-hand table: type / name·tool / count / turn + trigger trail / result, filterable by All Skills MCP Unused; click a row to open the host's native call inspector.

Features

Capability	Description
One-page aggregation	Every turn laid out chronologically on one canvas; no tab switching
Zoomable canvas	Wheel zoom (30%–240%), drag to pan, `Latest` / `Overview` shortcuts; opens on the latest turn
Attribution tree	`Prompt → reasoning → skill / MCP / dispatch`, with matched words coloured per target node
Subagent aggregation	Dispatch is its own node class; when a child session is reachable its skills / MCP are folded in as chips, and the card opens that session
Call table	Type / name / count / turn / trigger trail / result, filter by all / skills / MCP / unused; click a row for the host's native inspector
Unused inventory	Skills and MCP servers available in this session but never called (the `Unused` filter)
Lane chart	Input / model / tool time lanes for "what was happening when"

Install

Requires a working DSH install. The plugin registers into the existing conversation.view slot and does not modify any shipped build output.

Option 1 — DSH plugin manager (recommended)

Install this directory as a local package:

plugin_manager install_bundle  target = <absolute path to this directory>

Then reload the page: a third view tab, Call map, appears next to Chat and Trajectory.

Option 2 — manual mount

Put this directory anywhere;
In the profile directory run pnpm add <absolute path> (or symlink it into node_modules/call-map);
Add call-map to the dsh.profile.bundles array in the profile's package.json;
Restart / reload.

Uninstall: plugin_manager remove_bundle target = call-map.

Usage

Open any session and click Call map in the header;
Wheel to zoom, hold the left button to pan (a click never starts a drag);
Latest returns to the newest turn, Overview fits the whole tree;
Click a table row to open the host's native call inspector; click a subagent card to jump to that child session.

Data sources and definitions

Skill: a skill tool call, named by its arguments.name;
MCP: a tool call whose name matches mcp__<server>__<tool>;
Subagent: a dispatch call to spawn_teammate / subagent / subagent_fork / workflow / ralph;
Unused inventory: the registered skill catalogue (from the <available_skills> block in the system prompt) plus MCP servers that exposed tools to this session, minus what was actually called;
Trigger trail: the lexical fragment in the prompt / reasoning that matches the call.

Known limitations

The trigger trail is lexical matching, not proof of causality. If no matching word appears in the text, nothing is highlighted — the plugin does not invent attribution. Semantic attribution would need an extra model call.
Cross-session (subagent) aggregation depends on getting a child session id: subagent / workflow results carry one; spawn_teammate only returns a teammate name, so the plugin looks it up by "workspace session title == teammate name" and degrades to "session not loaded" when that fails.
Only the loaded event window is read: in a very long session, early turns may show "the prompt of this turn is outside the loaded event window".
The lane chart is re-drawn by this plugin and is not the shipped Trajectory implementation; the host's Trajectory page is untouched.
The host Trajectory toolbar (the duration / turns / calls row) exposes no slot to plugins, so this plugin can only be a parallel third tab — the entry point cannot live inside that row.

Privacy

No collection, no transmission. All computation happens in the local front-end, with no additional network requests.

Layout

call-map/
├─ package.json        # package manifest (dsh.bundle / dsh.client)
├─ cordis.patch.yml    # bundle patch: inserts one plugin row
├─ index.js            # Host half (empty for this plugin)
├─ client.js           # all logic: session-event folding + view rendering
├─ locale/en.json      # plugin-page display name (English)
├─ locale/zh.json      # plugin-page display name (Chinese)
├─ docs/*.png          # README screenshots
├─ assets/icon-180.png # storefront tile (180x180 PNG)
├─ screenshots.json    # store screenshots (1-8 images)
└─ icon.svg            # plugin icon

License

MIT

NICEGSP/call-map

call-map · 调用可视化

中文

它解决什么

树状图可缩放：一页看完整场对话

能看到详细的子代理 / 技能调用

功能

安装

使用

数据来源与口径

已知限制

隐私

目录结构

English

What it solves

Zoomable tree: the whole conversation on one page

Detailed subagent / skill calls

Features

Install

Usage

Data sources and definitions

Known limitations

Privacy

Layout

License

Community comments

NICEGSP/call-map

call-map · 调用可视化

中文

它解决什么

树状图可缩放：一页看完整场对话

能看到详细的子代理 / 技能调用

功能

安装

使用

数据来源与口径

已知限制

隐私

目录结构

English

What it solves

Zoomable tree: the whole conversation on one page

Detailed subagent / skill calls

Features

Install

Usage

Data sources and definitions

Known limitations

Privacy

Layout

License

Community comments