Task Workbench · Idea Pool · Reviews for DeepSeek Harness

A local-first task manager, idea pool and daily/weekly/monthly review built as a DeepSeek Harness plugin. Sidebar entry with four tabs — Today, This week, Tasks and Ideas — plus a review area at the bottom of the Today view. Tasks can be single-day or multi-day (start date → end date) with optional times; no recurrence rules, projects, lunar calendar or push reminders. Ideas have four states (todo / doing / done / dropped), three maturity levels (low / medium / high), a multi-line description and an optional category. Only a human can change an idea's state. Reviews are daily / weekly / monthly: the daily review is written for today; the weekly review appears Monday–Sunday and covers the previous calendar week; the monthly review appears on the last day of the month and covers that month (can still be written on the 1st–3rd of the next month). Review prose is human-only; agents can read it, never write it. The source of truth is the plugin's own data root (default $DSH_HOME/task-workbench/data, not in Git): tasks are sharded by month, ideas and reviews live in single long-lived files.

中文 README

Status

Open-sourced on GitHub on 2026-11-06. Current version: dsh-task-workbench@0.5.0 (GitHub Release v0.5.0) — the repository/package were renamed from personal-task-workbench and the agent tools from ptw_* to tw_*. V3.0 is implemented and accepted with real GUI clicks + agent tool integration on DSH desktop (2026-11-05). Tests: 89 passing with no runtime dependencies and no build step. The full suite is a release/CI gate; day-to-day changes run only the targeted domain tests (see docs/05-协作与测试.md).

V3.0 adds the review area at the bottom of the Today view (daily / weekly / monthly, human-written, agent-read-only) and a separate review domain (ReviewService, /tw/api/reviews*, tw_review_* read-only tools). V2.0 added the idea pool (single file, four states, three maturity levels, multi-line descriptions, human-only state changes). The UI is currently Chinese-only; the plugin, its files and this README are language-agnostic.

Install

Requirements: DSH >=0.2.2-rc.1 <0.3.0, Node >=24. The plugin has no build step and no runtime dependencies.

From npm (recommended):

dsh plugin --profile <profile> add dsh-task-workbench

On desktop you can also use the GUI: Plugins → Add plugin and enter the package name dsh-task-workbench.

From GitHub: in the desktop app open Plugins → Add plugin and paste the repository URL, or use the CLI:

dsh plugin --profile <profile> add github:Zevin-Labs/task-workbench

From the release tarball (prebuilt; no build approval needed): download dsh-task-workbench-0.5.0.tgz from the v0.5.0 release, then:

dsh plugin --profile <profile> add /path/to/dsh-task-workbench-0.5.0.tgz

From a local checkout:

git clone https://github.com/Zevin-Labs/task-workbench.git
dsh plugin --profile <profile> add /absolute/path/to/task-workbench

Then open Task Workbench in the sidebar. The first open walks you through initializing the data root.

What you get

1. Today — overdue unfinished tasks, tasks due today (including multi-day tasks currently in range) and tasks completed today (check to undo). The bottom of the page is the review area.
2. This week — a natural week (Monday–Sunday) as seven columns; single-day tasks land on their date; multi-day tasks are collected in one summary bar of chips (title + range + priority dot), and hovering a chip highlights the matching day columns. Today is highlighted.
3. Tasks — the current month's list, filterable by keyword / state / priority; the date column shows the time or range.
4. Ideas — the idea pool. The state counters at the top filter on click; the toolbar searches name/description, filters by maturity/category and switches sorting; the state can be changed directly on a card, and the title opens detail/edit.
5. New task — single-day (date + optional time) or multi-day (start → end, end must be later than start, optional start/end times); four-quadrant priority + one-line note; saving writes straight back to Markdown.
6. New idea — the name is required; multi-line description, maturity (low by default) and category are optional; the state starts at todo and is changed from the list later.
7. Reviews (bottom of Today, human-only):
   • Daily — one input per day for today; after submitting it renders in place and can be edited or cleared.
   • Weekly — appears during the current week and covers the previous calendar week; the block can expand "last week's material" (completed / overdue tasks + that week's daily reviews) and turns into a read view once written.
   • Monthly — appears on the last day of the month (28/29/30/31 as applicable) and covers the month; it can still be written on the 1st–3rd of the next month, with the month's task summary and weekly reviews as material.
   • Reviews outside their writable window are read-only (edit the Markdown directly for history); all three can be edited in place and cleared (second confirmation + backup before writing).
8. Edit / delete / complete — inline task operations; multi-day and single-day can be converted into each other; delete is a hard delete (second confirmation + automatic backup before writing). Idea edit/delete is in a modal; state changes happen on the card pill or in the modal.
9. Agents can read tasks, create/edit/delete tasks (tw_*), read ideas, create/edit idea content and delete ideas (tw_idea_*), and read reviews (tw_review_list / tw_review_get). Completing/uncompleting tasks, changing idea states and writing review prose are human-only — no agent write tool exists. The plugin ships a task-workbench skill that tells the agent which tools to use.

Data layout

data root (default $DSH_HOME/task-workbench/data)
├── 任务清单/YYYY-MM.md    the only writable task shard for the month: ## 任务 / ## 已完成
├── 想法池/想法池.md       single idea file: ## 想法 + idea blocks (metadata line + indented description)
├── 复盘/日复盘.md         single daily-review file: ## 记录 + review blocks (period key + 2-space-indented prose)
├── 复盘/周复盘.md         single weekly-review file (key = that week's Monday)
├── 复盘/月复盘.md         single monthly-review file (key = YYYY-MM)
└── 备份/<YYYY-MM-DD>/     per-write backups (tasks / ideas / daily / weekly / monthly; kept 7 days by default, `backupRetentionDays` 1–90)

• Task line: - [ ] `T-20261105-01`｜Workbench dev｜日期：2026-11-05｜时间：09:30｜结束日期：2026-11-12｜结束时间：18:00｜优先级：重要紧急
• Idea line: - `I-20261105-01`｜Quarterly review notes｜状态：推进中｜成熟度：高｜分类：工作｜创建：2026-11-03｜更新：2026-11-06, with the description on the following lines indented by two spaces.
• Review block: - \2026-11-05`｜更新：2026-11-05 22:05`, with the prose on the following lines indented by two spaces (blank lines separate paragraphs).
• The values written to Markdown and accepted by the agent tools are the Chinese enums: priority 重要紧急 / 重要不紧急 / 不重要紧急 / 不重要不紧急; idea status 待实现 / 推进中 / 完成 / 放弃; maturity 低 / 中 / 高. The English labels above are translations.
• Moving machines / backup: copy the whole data root. Two machines do not share data by default.
• It does not scan other workspaces, note vaults or Daily systems, and imports nothing.
• Plugin state lives in $DSH_HOME/task-workbench/ (settings.json / ids.json / idea-ids.json / writes.log) and never holds task/idea/review data.

Privacy and permission boundary

• No external network requests: the plugin only talks to its own same-origin local API (/tw/api) and reads/writes local Markdown with node:fs inside the Host process.
• Data stays in the local data root and $DSH_HOME/task-workbench/, never inside the plugin install directory, so upgrades and reinstalls cannot lose it.
• Agents and the UI share the same service, write gateway, queue, backup and audit log (writes.log) — there is no second write path.
• Completing/uncompleting tasks, changing idea states and writing reviews are human-only: there is no agent tool for them, and the service layer returns permission-denied for non-user actors.
• Deletes are hard deletes with an automatic pre-write backup kept for 7 days by default (configurable 1–90 via backupRetentionDays).

Code layout

lib/
├── index.js          # Host entry: TaskStore/TaskService + IdeaStore/IdeaService + ReviewStore/ReviewService + routes + agent tools
├── client.js         # Client entry: sidebar icon + native React panel (hand-written lazy-CJS, no build)
├── core/             # task / idea / review contract, grammar, derive, validate, window (pure functions)
└── host/             # settings / writer / gateway / http / store / service / routes / tools (one set per domain)
skills/               # bundled `task-workbench` skill (how agents should use the tools)
tests/                # core / service / routes / tools / wiring / settings / skills — temp data roots only
docs/                 # feature spec + implementation architecture + architecture diagram + collaboration/testing guide (Chinese)

Development

git clone https://github.com/Zevin-Labs/task-workbench.git
cd task-workbench
npm run check                      # syntax only
npm run test:task                  # targeted example: task domain (33 tests)
npm test                           # full 89-test suite — pre-release / CI only

• Targeted by default: run the test file(s) for the domain you touched; the change → test matrix is in docs/05-协作与测试.md. The full suite runs before a release, for cross-cutting changes, and in CI — not for every small edit.
• Automated tests only ever use os.tmpdir() data roots and never touch real data/.
• Host code changes: toggle the "Task Workbench" component off/on in the Plugins page, or restart the app.
• Client code changes: refresh the page (the client bundle is hot-replaced by DSH HMR).
• To keep development data inside the repo, point dataRoot in $DSH_HOME/task-workbench/settings.json at the repo's data/ (gitignored).
• Typical loop: make the change → targeted tests → local git commit (do not push) → publish a batch with docs/04-发布与更新.md.

Scope

• In scope: Today / This week (natural week) / all-tasks views; single-day and multi-day tasks; optional times; view, create, edit, delete, complete/undo; four-quadrant priority; note. Idea pool: list/filter/search, create, edit content, change state, delete, four states, three maturity levels, multi-line description, optional category. Reviews: daily/weekly/monthly, submit/edit/clear inside the writable window, weekly/monthly material, agent read-only.
• Out of scope: knowledge-base/Daily scanning, fixed routine rules, project directions, recurrence rules, lunar calendar, push reminders, agent claiming/acceptance, old-data migration, historical month browsing; idea → task conversion, idea kanban, automatic archiving; review history tab, agent-written reviews, backfilling older periods, review charts.
• Delete is a hard delete (backups, no recycle bin); reviews are read-only once their window closes; external edits are picked up by a 5-second poll, and writes use baseHash to avoid clobbering them.
• Single process, single user; concurrent writes from multiple windows are not handled (baseHash conflicts are rejected).

Docs

• Docs index (Chinese): docs/README.md
• Feature spec: docs/01-功能规格.md
• Implementation architecture: docs/02-实现架构.md
• Architecture diagram: docs/03-整体架构图.html · screenshot
• Collaboration & testing (agent onboarding, data-contract → test map, config layers, targeted-test rules): docs/05-协作与测试.md
• Release & update runbook (Chinese): docs/04-发布与更新.md
• Contribution rules: CONTRIBUTING.md

License

MIT