본문으로 건너뛰기
A

dsh-kanban

alpacachen/dsh-kanban

하나의 DSH 플러그인에서 사람이 주도하는 Kanban과 AI 실행 워크플로를 제공합니다. 끌어 놓는 카드, 검색, 우선순위, 라벨, 기한, Markdown, 이미지, 댓글로 에이전트와 Workspace 작업을 관리합니다. 채팅으로 재사용 가능한 워크플로를 만들고 편집하며 단계별 모델을 설정하고 DSH 도구와 권한으로 AI가 각 단계를 실행하게 합니다. 흐름도에서 진행을 확인하고 단계별 기본 대화를 살펴보며 결과를 승인하거나 피드백과 함께 되돌릴 수 있습니다. 실패한 단계 재시도나 작업 종료도 가능합니다. 작업별 artifact 폴더의 파일을 미리 보고 실행 기록과 리뷰 피드백을 활용해 다음 워크플로를 개선합니다.

설치

dsh plugin --profile web add github:alpacachen/dsh-kanban

README

dsh-kanban

A board for people. Workflows for AI.
Plan, run and review work inside DeepSeek Harness.

npm version CI MIT license Awesome DSH Plugin

简体中文 · English
Get started · Kanban · Workflow · Documentation

One plugin, two ways to work. Use Kanban to organize tasks and decide what happens next. Use Workflow to give AI a repeatable sequence, follow its execution, and step in at review points.

Both live under Kanban · Workflow in the sidebar. Switch modes beside the workspace name; the plugin remembers your last mode and workspace.

Kanban · Work you leadWorkflow · Work AI runs
Start withA task or a plan from a conversationA reusable process designed with your Agent
Move forwardYou or your Agent update cardsAI executes stages in order
Keep controlPriorities, deadlines and review columnsHuman approval, feedback, retry and stop
Keep the contextDescriptions, screenshots, comments and activityStage conversations, reports and task-specific files

Get started

Requires DSH 0.2.2-rc.3, with desktop and Web support. For Workflow, configure an execution model in DSH first.

Desktop — Open Plugins, add @alpacachen/dsh-kanban, and enable it.

Web — Install the plugin, then restart your Web instance to load it:

dsh plugin --profile web add @alpacachen/dsh-kanban

Open Kanban · Workflow in the sidebar and choose a workspace. Stay in Kanban to create your first card, or switch to Workflow, choose Create workflow in chat, and describe a process.

Workflow is included in 1.12.0. Until that release is published, installing from npm gives you the latest published version. To try this branch, build and install it locally.

Kanban — Work you lead

Turn plans into visible tasks. Move cards yourself, ask your Agent to keep them up to date, and take the next step when you are ready.

Kanban board with task search, priority and label filters, and four task columns

Actual plugin UI with isolated example data. Each workspace has its own board.

  • Plan together. Ask your Agent to break down a goal, write acceptance criteria and create cards with 19 built-in board tools.
  • Find the next task. Search titles, descriptions, comments and IDs; combine priorities, labels and due-date filters.
  • Keep evidence close. Add Markdown, screenshots, comments and activity history to a card.
  • Continue in chat. Bring a card and its saved images into a new conversation as an editable draft, ready for you to send.

Try asking your Agent:

Break phone sign-in into testable tasks. Check existing cards first to avoid duplicates. Add implementation steps and acceptance criteria, and mark launch blockers as P0.

Kanban guide → · All 19 board tools →

Workflow — Work AI runs

Describe a process once, then give it new inputs. AI runs its stages using DSH's normal tools and authorization; you can follow the work and decide at human review points.

Example workflow: read materials, draft a report, wait for human review, then finalize; feedback returns to drafting

An example process. Human review pauses execution until you approve, return it with feedback, or stop.

Design it in conversation

Ask your Agent to create or revise a workflow. It saves the stages, instructions, deliverables and review points. New AI stages inherit the model used in that conversation; you can later edit stage instructions and choose a model and reasoning effort in the stage dialog.

Create a reusable research workflow: read the materials, draft a report with sources and open questions, wait for my review, then finalize the report. If I return it, go back to drafting. Save the workflow without running a task.

Run it with a new input

Expand a workflow in the home list and choose Create task. Enter what you need and send it. No separate task name or model setup is required.

  • See the work. The diagram highlights the current stage. Click a stage to inspect its native DSH conversation, available reasoning text, tool calls and replies.
  • Review and revise. Approve to continue or return with feedback. Earlier reports and review decisions remain available.
  • Keep files separate. Every task has its own output directory, shared by its stages. Preview files in place, open the directory, or delete deliverables with confirmation.
  • Resume deliberately. Retry a failed stage after checking the cause. After a crash, reconcile the old execution before retrying; completed upstream stages are retained.

Improve the process from its history

Choose Optimize from history to start a conversation with an analysis prompt. The Agent can inspect previous task results, review feedback, stage conversations and recorded timings. Changes apply to future tasks; existing tasks keep their original workflow version.

Workflow guide → · Agent interface (中文) → · Validation and recovery boundaries (中文) →

What to expect

  • Native DSH execution. Workflow adds orchestration and human review, with no separate tool-permission system or extra model-turn, tool-call or stage-time budgets.
  • Sequential stages. Stages in a task run in order. Tool concurrency and model-request retries follow DSH; a failed workflow stage is retried by you.
  • Small handoffs. Stages receive upstream conclusions, evidence references and a file index. Full reports are read when needed instead of carrying the entire conversation forward.
  • Review scope is explicit. Approval applies to the saved stage reports. It does not snapshot project files or undo changes made by tools.
  • Two interface languages. Controls, owned error messages and dates follow DSH's Chinese or English setting. Your inputs, prompts and saved outputs retain their original language.

If DSH is forcibly killed, previously launched commands can survive on macOS. Check that old commands and external operations have stopped before confirming recovery. The plugin will not silently rerun an interrupted stage.

Documentation

You want to…Read
Use cards, search, images and chat handoffKanban guide
Create, run, review and improve workflowsWorkflow guide
Understand board toolsBoard tool reference
Back up your dataKanban data · Workflow data
Build and validate the pluginDevelopment checks · Workflow acceptance record
Report a problem or suggest a changeGitHub Issues

MIT · Built for DeepSeek Harness

관련 플러그인