Introduction¶

In DeepSeek Harness (DSH) for idea organization, the common approach is back-and-forth questioning in the chat window: first clarify the problem, then let the model expand point by point, and finally manually organize the conclusions into a document. As ideas pile up, pure text Q&A makes it hard to see branching relationships, and dragging, merging, and long-term review are impossible. dsh-brainstorming attempts to move this to a canvas: start with a vague idea on an infinite canvas, have the plugin ask questions first to align, then generate a research outline and results. Below introduces the positioning, installation, and typical operations of this plugin.

What is this¶

m430/dsh-brainstorming is a DSH client plugin, maintained by m430. It is positioned as “the AI brainstorming plugin for DeepSeek Harness Web UI”. It starts from a vague idea on an infinite canvas, defaults to aligning questions via “Interview”, then “Confirm Research” to generate outline and results; it can also choose “Diverge” to quickly expand directions. The project code uses MIT License; see THIRD_PARTY_NOTICES.md in the repository for packaged third-party dependencies and icon licenses.

It belongs to the client category in the community directory skillhub.cn. The directory site has no official affiliation with DeepSeek / Hypi; DSH’s design philosophy is “everything is a plugin”, and such plugins belong to community-side extensions, not applications in the official app store.

Core Features¶

• Infinite Canvas: Dot matrix background with arrow-connected lines, supports pan, zoom, subtree dragging, “Organize Layout”, and auto-save.
• Interview: Ask one question at a time, Q&A falls into the interview group; after ending, it can lead to a research outline. New ideas default to entering “Interview”.
• Diverge: Quickly expand several directional intents, no need to go through a full interview first.
• Confirm Research: Can supplement research requirements, confirm the outline first, then generate research results; results support comments and secondary rewriting.
• Non-overlapping Layout: Sibling reordering and column collision avoidance; automatically pushes apart when content becomes taller.
• Node Editing: Double-click to edit body text, rendered in Markdown; select and press Delete / Backspace to delete.
• History Canvas: List to manage multiple canvases; new defaults to “Untitled”, AI auto-summarizes the title after writing the first valid node.
• Unified UI: shadcn-style components plus Base UI primitives, design conventions see DESIGN.md in the repository.

Installation and Enablement¶

The environment requirements are DeepSeek Harness Web UI, and a usable DSH model provider.

Install from npm (recommended as marked in the README): install in DSH’s Plugin Manager / Plugin Management, the name is the package name.

dsh-brainstorming

You can also specify a version; the current README version is 0.0.5.

dsh-brainstorming@0.0.5

After installation, refresh the page, and the “Brainstorming” panel will appear on the left sidebar.

The README also provides the path for installing from source, suitable for situations where code needs to be modified: clone the repository, install dependencies and build, then select this local directory for installation in the DSH Plugin Manager.

git clone https://github.com/m430/dsh-brainstorming.git
cd dsh-brainstorming
pnpm install
pnpm build

Typical Usage¶

After starting, follow these steps to get a complete generation pipeline.

1. Click “Brainstorming” in the left sidebar to enter the canvas list.
2. Click “+ New Brainstorming” to open a blank canvas, title defaults to “Untitled”.
3. Double-click in the empty space to add an idea and submit, the root node will default to “Interview”.
4. After selecting a node, the toolbar provides three main operations:
   - Interview: Ask one question at a time, answers are written to the interview group; after the interview ends, you can “Confirm Research”.
   - Diverge: Quickly generate directional intents.
   - Confirm Research: Optionally supplement requirements, confirm the outline then generate research results; results can be commented on and rewritten based on comments.
5. Dragging a node will move the entire subtree; you can click “Organize Layout” if needed. Select a node and press Delete / Backspace to delete it.

Configuration¶

By default, it follows the model currently selected in DSH, regardless of whether you use the official API Key or other providers configured in Harness. The plugin does not read or require a separate API Key; authentication and keys are managed by DSH’s credentials layer.

Generally, no configuration change is needed. If you want to specify a model for the plugin separately, or adjust timeout and output length, you can write it in the profile’s cordis.patch.yml, for example:

- id: dsh-brainstorming
  config:
    # provider / model must match an entry you have configured and available in Harness
    provider: your-provider-id
    model: your-model-id
    timeoutMs: 120000
    maxOutputTokens: 1800

Note that provider and model must be filled in together to override the default selection; filling in only one of them will be ignored.

Applicable Scenarios and Notes¶

Suitable for users already working in DSH who want to move the idea research process out of the chat window and need to keep multiple drafts long-term. Auto-save, history canvas list, and Markdown rendering make this closer to a reviewable workspace rather than a one-time conversation record.

Regarding security and privacy, the README clarifies a few things: the plugin will not package the API Key into the frontend or submit it to the repo; AI requests are sent via DSH’s ctx.llm service, with the Key staying in the Host’s credentials layer; all plugin routes go through DSH’s /api authentication channel, unauthenticated requests return 401; canvas content is persisted in the brainstorm domain of the Host’s storageDomain, accessed via authenticated /api. It is worth noting that node content and prompts are sent to the currently configured model provider; when handling sensitive content, the README suggests using a local model.

Such plugins run with the permissions of the current DSH process and can access content accessible by that process. Before installation, it is recommended to read the source code and license (MIT License) first to ensure dependencies and configuration items match your deployment environment.

Conclusion¶

dsh-brainstorming connects DSH’s plugin extension to a specific scenario: instead of opening another chat window, it uses a canvas to carry the entire process from questioning, diverging, to research results. If you are already using DSH, it is worth trying as a low-cost experiment.

Directory Page: https://www.skillhub.cn/plugins/m430/dsh-brainstorming

GitHub: https://github.com/m430/dsh-brainstorming