跪拜 Guibai
← All articles
DeepSeek · Open Source

DSH Task Board Brings Kanban-Style Agent Workflows to DeepSeek Harness

By ovO ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

DeepSeek Harness users running multiple agent tasks lose visibility the moment a chat list grows beyond a handful of items. A kanban board with persistent state and explicit review gates turns ad-hoc prompting into a workflow that can survive restarts, making the tool usable for batch jobs and multi-step agent runs that need human judgment between stages.

Summary

DSH Task Board installs as a single plugin that surfaces a kanban board inside DeepSeek Harness Web. Tasks move through five states—Initialized, Running, Pending Review, Completed, and Failed—with manual approval gates and retry controls. It does not replace the existing session runtime; execution still runs through standard DSH Sessions, while the board handles task cards, status transitions, review workflows, and persistent storage.

Installation is a single npx command that adds one bundle, one package, and one Loader to the profile. The board supports both manual tasks for to-do tracking and Agent tasks that hand off to a DSH Session when model credentials are configured. Task cards carry a title, working directory, Agent Preset, requirements, acceptance criteria, and optional reference images.

Persistence is the headline feature: task records live in DSH Storage, not just React state. Refreshing the page, closing the browser, or restarting the Host restores the full board from an authoritative snapshot. The v0.1.0 release passed 168 tests across 20 test files with byte-identical builds and zero browser errors, though the public demo did not include live model calls.

Takeaways
DSH Task Board adds a five-column kanban (Initialized, Running, Pending Review, Completed, Failed) to DeepSeek Harness Web.
Installation is a single npx command that adds exactly one bundle, one package, and one Loader to the profile.
Tasks persist in DSH Storage, not in-memory React state, so they survive page refreshes, browser restarts, and Host restarts.
Manual tasks track to-do items without calling a model; Agent tasks hand off to a DSH Session when API credentials are configured.
Rejecting a task during review creates a new revision round that reuses the same DSH Session rather than spawning a fresh one.
Failed tasks sit in the Failed column for manual retry decisions instead of auto-retrying indefinitely.
v0.1.0 targets DeepSeek Harness 0.1.0-rc.6 and requires Node.js ^22.19.0 or >=24 with pnpm available.
The release passed 168 tests, produced byte-identical builds across two consecutive runs, and showed zero browser errors or warnings.
Conclusions

Persisting task state in DSH Storage rather than React state is the architectural decision that makes the board more than a UI skin—it turns ephemeral chat sessions into durable workflow records.

The plugin's single-Loader design avoids the fragmentation common in plugin systems where Host, RPC, and UI components ship as separate entries that must be configured individually.

Explicit review gates and manual retry controls address a real gap in agent tooling: unattended agent loops that either succeed silently or fail without human recourse are not production workflows, they are experiments.

Concepts & terms
DSH Session
The standard execution unit in DeepSeek Harness that runs agent tasks. DSH Task Board does not replace Sessions; it wraps them with task cards, status tracking, and persistence.
DSH Storage
A persistent storage layer tied to a DeepSeek Harness profile that holds task records, rounds, and activity logs, allowing the task board to survive restarts and refreshes.
Agent Preset
A preconfigured set of model parameters and behaviors in DeepSeek Harness that an Agent task can reference when handed off to a DSH Session for execution.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗