dsh-vision-bridge

When using text-only models like DeepSeek in DSH web sessions, sending messages with images is outright rejected and the read_image tool is blocked by capability checks. DSH declares these models as text-input only, making image functionality completely unavailable. dsh-vision-bridge makes the experience seamless: drag in images or call tools as usual, and the plugin handles vision-to-text translation transparently in the background. No model switching, no manual commands.

The plugin works on two levels. At the capability layer, it wraps resolveModelInfo to add image input capability to target provider models so all capability checks pass. At the stream layer, it wraps the provider adapter stream method, extracts image blocks before serialization, calls the Alibaba Cloud Bailian vision API for description, and replaces them with text blocks. The model always receives pure text, while session logs and GUI remain untouched. The default primary model is qwen-vl-max with automatic fallback to qwen-vl-plus and flash series; auth and parameter errors do not trigger blind retries.

Beyond auto vision, the plugin supports text-to-image generation (tell the model to draw and it calls Qwen-Image to produce and attach a result), result caching (same image is not re-queried; in-memory plus disk-persisted across sessions), and extended utilities such as URL description, OCR prompting, and multi-image comparison. Configuration is flexible with clear priority across environment variables, plugin .env file, settings page, and claude-vision-skill directory.

Ideal for developers and daily users who run DeepSeek text-only models in DSH web sessions but want to keep the image interaction experience. It is a pure server-side plugin with zero GUI changes, and it coexists cleanly with the desktop EAC built-in dsh-tool-vision (with bridgeTextOnly disabled). Install and go.