dsh-task-board

A cross-project task board for DeepSeek Harness — every DSH session becomes a task card, grouped into three columns: 待处理 / 处理中 / 待审阅 (To do / In progress / Needs review).

English | 中文

The screenshot above is rendered from the real bundle against fictional demo data (npm run preview). Nothing in this repo contains a maintainer's workspace names, paths or session titles.

What it does

The board is a first-class DSH destination, not an overlay: it registers into the shell's own global-panel seats (sidebar.panellist + main), the same mechanism DSH's built-in Schedules page uses. Selecting it swaps the centre column; returning to the Conversation does not change the current session.

• Three columns derived from live host facts — running, completed (finished but not yet opened), and unanswered interactions.
• Priority stars with a 16-colour palette. A colour is assigned once and remembered, so it survives filtering, sorting, unstarring, and even removing the card.
• Stable "In progress" ordering — progress, timestamps, titles and stars never reorder the column; new arrivals append at the end.
• One-page adaptive layout — busy columns gain horizontal lanes instead of shrinking text. Font size is never reduced and no card detail is ever hidden.
• Native turn duration folded from the session's own turn/start → turn/end events.
• Add-session picker, panel search, pause/resume, complete-with-undo.
• Light and dark, following the shell's resolved colour scheme.

Install

Install straight from GitHub — this is the supported path, and the only one that works today (the package is not on npm yet):

dsh plugin --profile <your-profile> add git+https://github.com/cpsGGG/dsh-task-board.git

Then restart DSH and open 任务面板 at the bottom of the sidebar.

git clone https://github.com/cpsGGG/dsh-task-board.git
dsh plugin --profile <your-profile> add ./dsh-task-board

dsh plugin --profile <your-profile> add git+https://github.com/cpsGGG/dsh-task-board.git#v1.0.0

Requirements

DSH is in developer preview and its internal slot surface can change between releases. The plugin reads useSessions from the global standard props and registers into sidebar.panellist + main; if a future release renames either, the board fails loudly with a visible stack rather than silently disappearing.

How it works

DSH has no DOM to patch — it has slots. So this is a native plugin rather than an injection:

src/core.mjs is deliberately framework-free — no DOM, no React, no Cordis — so every rule (status precedence, ordering, layout arithmetic, colour balancing) is unit-testable in plain Node.

Layout of the repo

Development

npm run build     # src/*.mjs + panel.css -> lib/client.js
npm test          # 50 tests
npm run preview   # regenerate preview/board.html

The client half is not an ES module. DSH's client module system loads a classic script that calls window.__ModuleLoader__.load({ id, factory }) and resolves externals through an injected require. tools/build.mjs inlines the two local modules, rewrites the react import into that require, and wraps the result — no bundler, because the graph is two files.

The render tests are worth knowing about: they load lib/client.js, drive the same apply() the shell drives, and expand the returned element tree with a stub React. That catches the class of failure a live check hides behind DSH's slot error boundary, which renders an opaque empty div.

Why the preview exists

A screenshot of a real board leaks whatever the maintainer had open. tools/make-preview.mjs renders the same component against a fixed cast of fictional sessions, so the artifact is reproducible and safe to publish.

Behaviour notes

Two behaviours are worth knowing about, because DSH's session model constrains them:

1. "Resume" sends a real message. DSH offers no "continue without a message" verb: prompt rejects content with no non-whitespace text (the gateway's hasPromptContent guard), and there is no session-level resume. So resuming a paused card sends a minimal real instruction (继续). The card leaves the paused state and starts running; the message is visible in the transcript.

2. Blank sessions stay off the board. DSH reuses one empty "New session" draft per workspace, so it is a scratch slot rather than a task, and the board excludes it.

Contributing

Issues and PRs welcome. Please run npm test before opening a PR; if you touch the panel UI, also run npm run preview and check both themes.

License

MIT.

This project is not affiliated with DeepSeek AI.

中文说明

一个 跨项目任务面板：把每条 DSH 会话变成一张任务卡片，按 待处理 / 处理中 / 待审阅 三栏集中查看与打开。

面板是 DSH 的原生目的地，不是浮层——它注册进外壳自己的全局面板槽位（sidebar.panellist + main），和 DSH 内置的「自动化任务」页用的是同一套机制。点它切换中央面板，返回会话不改变当前会话。

功能

• 三栏由宿主真实状态推导：运行中、已完成但未打开、等待回答。
• 优先级星标：16 色色板，颜色只分配一次并记住——筛选、排序、取消星标、甚至移出面板后重新加入，都恢复原色。
• 「处理中」顺序稳定：进度、时间、标题、星标都不触发重排，新进入的追加到末尾。
• 单页自适应：繁忙的列增加横向分栏，而不是缩小文字；字号永不压缩，卡片信息永不隐藏。
• 原生用时：从会话自身的 turn/start → turn/end 事件折算。
• 添加会话弹窗、面板搜索、暂停/启动、完成并撤销。
• 浅色 / 深色跟随外壳解析出的配色。

安装

从 GitHub 直接安装（当前唯一可用方式，包尚未发布到 npm）：

dsh plugin --profile <你的 profile> add git+https://github.com/cpsGGG/dsh-task-board.git

然后重启 DSH，点侧边栏底部的 任务面板。

也可以先 git clone 再 add ./dsh-task-board，或锁定到某个 tag/commit： add git+https://github.com/cpsGGG/dsh-task-board.git#v1.0.0。

lib/client.js 已随仓库提交，无需构建、也不会执行任何安装脚本——本包刻意不提供 prepare，否则 git 安装会触发 pnpm 的 allowBuilds 提示。改动 src/ 后运行 npm run build 并重新安装。

Windows 提示：pnpm 对 file: 依赖是复制而非软链。若你在本地 checkout 上开发，建议把 profile 的 node_modules/dsh-task-board 用目录联接指向工作目录，否则每次构建后都要重装。

行为说明

有两处行为值得了解，都是 DSH 会话模型决定的：

1. 「启动」发送真实消息。DSH 没有"不追加消息地继续"这个动词：prompt 拒绝没有非空白文本的内容，也没有会话级 resume。因此启动暂停的卡片会发送一条最小真实指令（继续）——卡片离开暂停态并开始运行，该消息在对话记录里可见。

2. 空会话不入面板。DSH 每个工作区复用一个空的「新会话」草稿，它是草稿位而非任务，故面板排除。

开发

npm run build     # src/*.mjs + panel.css -> lib/client.js
npm test          # 50 个测试
npm run preview   # 重新生成 preview/board.html

客户端半边不是 ES module：DSH 的客户端模块系统加载的是调用 window.__ModuleLoader__.load({ id, factory }) 的经典脚本，外部依赖通过注入的 require 解析。tools/build.mjs 内联两个本地模块、把 react 改写成该 require，再包上外壳——没有引入打包器，因为依赖图只有两个文件。

渲染测试值得一提：它加载 lib/client.js，走与外壳相同的 apply()，用桩 React 展开元素树。这能抓住一类被 DSH 槽位错误边界（只渲染一个空的 data-slot-error）掩盖的崩溃。

许可

MIT。本项目与 DeepSeek AI 无隶属关系。