DSH Design Mode

中文

An image-first Design Mode for DeepSeek Harness, built on the idea that everything is a plugin.

DSH Design Mode connects the native DSH conversation to an infinite canvas, guided intent clarification, focused image tools, and location-aware comments. The goal is not to put another image generator beside chat. The goal is to make the whole image workflow—from an ambiguous idea to a traceable series of edits—part of one Agent session.

Developer preview based on dsh-v0.1.1-rc.2. This independent community project is listed under the GitHub dsh-plugin topic and is not an official DeepSeek release.

Why this exists

Powerful open design environments can expose too many choices before a new user has a clear brief. DSH Design Mode starts from ordinary language, lets the Agent clarify material ambiguity, and opens the canvas only when there is an image to work with. Experienced users still get direct tools, while new users get a guided path instead of a wall of controls.

What is included

• Intent-first entry: Design Mode can be armed and disarmed. Arming it does not open an empty canvas; sending text, an image, or both starts the workflow.
• Adaptive ask_user: new-image requests use 3–5 sequential clarification cards. Each card provides three suggested choices plus a UI-owned “Other” input.
• Infinite canvas: pan, pointer-centered zoom, selection, multi-selection, move, proportional resize, rotate, lock, hide, download, fit-to-view, and local undo/redo.
• Image-only creation: upload, paste, drop, generate, or add an existing chat image. Video generation and video controls are intentionally excluded from v1.
• Task-specific tools: background removal/replacement, erase, upscale, outpaint, OCR-assisted text editing and translation, angle change, recolor, and authorized mark removal.
• Location-aware comments: point, rectangle, and brush comments return to the native conversation. Multiple comments are processed in order so the chat remains the auditable record of every change.
• One Host-owned gateway: the UI and Agent call the same Design service, job queue, persistence layer, and provider routing.
• Private credentials: OPENAI_API_KEY and FAL_KEY are read from the DSH credential vault and never shipped to the browser or written into the workspace.
• Desktop and Web: the same profile and browser surface run in both DSH Desktop and DSH Web.
• Canvas export: PNG, PDF, standalone HTML, ZIP, and PPTX.

Demo chapters

The repository includes four compressed walkthroughs recorded from the working build:

1. Enter Design Mode and start from chat
2. Clarify an image request with ask_user
3. Work on the canvas with object tools
4. Place comments and continue the edit in chat

The interactive product story is in showcase/. Run it with npm install && npm run dev.

The editable investor/client deck is available as DSH Design Mode product introduction (PPTX).

Install from source

Requirements: Node.js ^22.19.0 or >=24.0.0, Corepack, pnpm 11.7.0, and Git.

1. Clone and install dependencies

git clone https://github.com/KaichenCurry/dsh-design-mode.git
cd dsh-design-mode
corepack enable
pnpm install

2. Build and create the isolated profile

pnpm run build
pnpm run design:install

3. Start DSH with Design Mode

pnpm run design:start

The command starts the complete DSH Web application with the design profile and opens its local URL. Use pnpm run design:start --no-open when another shell, including DSH Desktop, should open the URL. The profile adds Design Mode to the standard DSH application; it does not replace another DSH executable or change the default web profile.

4. Add provider credentials only when needed

Open Settings → Design Mode and store credentials through the write-only vault surface.

Task	Required credential
GPT Image generation and generative editing	OPENAI_API_KEY
Background removal and upscale	FAL_KEY

Canvas interaction, comments, composer context, persistence, and keyless contract tests work without provider credentials. A real provider job reports a recoverable missing-credential error until the corresponding key exists.

5. Manage or remove the profile

Useful commands:

pnpm run design:status
pnpm run design:upgrade
pnpm run design:uninstall

design:uninstall moves the profile to a timestamped recovery directory and leaves workspace .dsh/design-mode data intact.

Current v1 scope

V1 is intentionally image-only. It supports image generation, import, canvas editing, contextual tools, and location-aware comments; it does not expose video generation, video editing, or video buttons.

Architecture

Native DSH Composer
  ├─ text / images / removable Design context chips
  └─ ask_user clarification
          ↓
DSH Agent
  ├─ design_read_context
  ├─ design_generate_image
  ├─ design_edit_image
  └─ design_update_comments
          ↓
Design Job Router + per-session queue
  ├─ GPT Image: generation and generative editing
  ├─ FAL providers: background removal and upscale
  └─ OCR provider: region detection for text workflows
          ↓
Local canvas document, assets, comments, versions, and journal

The core packages are:

• packages/design/design-mode: Host service, storage, provider routing, and Agent tools.
• packages/client/ui-design-mode: React/Konva canvas and DSH client integration.
• packages/bundle/design-app: optional bundle layer used by the isolated design profile.

Privacy and safety

• Only assets referenced by the current job are sent to the selected provider.
• Local canvas documents and history stay in the workspace.
• Provider failures never trigger an automatic paid retry.
• Watermark/mark removal requires the user to confirm ownership or authorization.
• Do not commit .env files, workspace data, or provider keys. See SECURITY.md.

Roadmap

• Strengthen real-provider success/failure smoke tests and cost-aware routing.
• Package signed Desktop installers and reproducible upgrade/uninstall flows.
• Add team review, permissions, approvals, and shared asset governance.
• Open an adapter surface for third-party providers, Agent skills, and reusable workflows.

License and attribution

This repository is derived from DeepSeek Harness and retains its MIT license and third-party notices. New Design Mode code is released under the same MIT license. DeepSeek and DeepSeek Harness are trademarks of their respective owners.