dsh-vision-toggle · Model Vision Switch Plugin

English | 中文

A model vision toggle plugin for DeepSeek Harness (dsh) desktop: let users control whether each configured model supports image input — flip a switch to enable/disable vision per model, no config file editing and no restart required.

Features

Feature	Description
Model List	Read every model declared in the llm-pi-ai provider routes, grouped by provider
Vision Switch	Per-model toggle to turn image input on (['text','image']) or off (['text'])
Immediate Effect	Writes into the official config; llm-pi-ai re-registers on change — takes effect instantly, no restart
Optimistic UI	Switch flips immediately; rolls back if the host write fails
i18n	Chinese / English, follows the dsh desktop language setting

Background

dsh officially supports multimodal (image understanding) with its own deepseek-v4-vision-exp model. However, the settings page offers no way to declare the input modalities (input) of custom providers or third-party models. If you add your own provider (e.g. an OpenAI-compatible service via LM Studio / vLLM / Ollama) whose model actually supports vision, sending an image still fails with:

当前模型不支持图片  (this model does not support images)

The only workaround today is to hand-edit settings.yaml and add input: [text, image] for the model — which is fragile and easy to get wrong.

This plugin removes that friction: it exposes the official llm-pi-ai settings namespace so each model gets a "Support image" switch in the settings page. Toggling it rewrites the model's input modalities for you.

Installation

Prerequisites

• DeepSeek Harness (dsh) desktop
• Node.js >= 18

Integration in dsh-desktop

1. Install the plugin:

dsh plugin add @lijian-ui/dsh-vision-toggle

2. Restart the desktop app.

Local Development

# Enter the plugin directory
cd extensions/dsh-vision-toggle

# Install dependencies
npm install

# Build
npm run build

# Watch mode
npm run watch

# Type check
npm run typecheck

Build output goes to lib/ and is automatically synced to node_modules/@lijian-ui/dsh-vision-toggle via junction. Restart the desktop app after each build to load the new bundle.

Usage

1. Open dsh desktop
2. Navigate to Settings → Model Vision (modeled after the other settings sections)
3. The section lists every configured model, grouped by provider. Next to each model there is a "Support image" switch:
   • On: the model accepts image input; you can now attach images in chat
   • Off: the model is text-only

Turning the switch on simply sets that model's input to ['text', 'image']; turning it off sets it back to ['text']. This matches exactly how dsh decides whether a model supports images.

Technical Architecture

Directory Structure

extensions/dsh-vision-toggle/
├── src/
│   ├── index.ts                    # Host entry (registers the remote service + typert contract)
│   ├── remote.ts                   # Host RPC: listModels / setVision (reads & writes llm-pi-ai config)
│   └── client/
│       ├── index.ts                # Client entry (SECTION_ID, RPC registration, inject)
│       ├── VisionToggleSection.ts  # Settings component (model list grouped by provider + toggles)
│       └── client-i18n.ts          # Client i18n (zh/en)
├── lib/                            # Build output (index.mjs + client.js)
├── package.json
├── tsdown.config.ts
└── cordis.patch.yml

Host Side (src/remote.ts)

Provides the following RPC methods:

Method	Function
listModels()	List all models declared in providers[*].models, with their current input modalities and whether they support vision
setVision(provider, modelId, enabled)	Rewrite one model's input to ['text', 'image'] (on) or ['text'] (off) and persist

Config Read / Write

The plugin operates on the official llm-pi-ai settings namespace (settingsNamespace('llm-pi-ai')) owned by @deepseek-ai/dsh-llm-pi-ai — the same config that backs model selection and the "this model does not support images" error:

• Read: settings.get('llm-pi-ai') returns the provider routes; each model's effective input falls back to the provider's defaultInput.
• Write: uses settings.update('llm-pi-ai', { providers: ... }) with a rebuilt full models array. The path-based settings.mutate API cannot navigate array indexes (it would replace the whole array), so we rebuild the array instead.

Because llm-pi-ai resolves input per request and re-registers on settings change, a flip takes effect immediately with no restart.

Toggle Mechanism

User clicks the switch
  → Client optimistically updates the row (immediate flip)
  → RPC call to host setVision(provider, modelId, enabled)
  → Host: read config, rebuild models[], settings.update('llm-pi-ai', …)
  → llm-pi-ai detects the change and re-registers the model route
  → Config persisted to settings.yaml; next message uses the new modalities

If the host write fails, the client restores the previous switch state and shows an error.

Internationalization

Supports Chinese and English. Translation files are in src/client/client-i18n.ts. Language follows the dsh desktop language setting.

Tech Stack

• Language: TypeScript
• Build: tsdown (rolldown)
• Frontend: React 18
• RPC: @deepseek-ai/dsh-typert-protocol / @deepseek-ai/dsh-typert-registry (host remote)
• Config: @deepseek-ai/dsh-settings (namespace llm-pi-ai)

License

MIT

Related Links

• DeepSeek Harness (dsh)