dsh-vision-assist

When your primary model in DeepSeek Harness is text-only, pasting an image immediately triggers a rejection message and derails the conversation. dsh-vision-assist removes that friction by intercepting the request at the LLM call boundary, rewriting image blocks into an attachment-path plus vision_recognize text hint, and delegating the actual image understanding to whichever multimodal model you pick in settings. Chat bubbles and conversation history keep the original image intact, so the degradation is invisible to the user.

The plugin registers a host-level vision_recognize tool with coarse, default, and fine granularity levels and accepts custom prompts. You can swap the recognition model on the fly from the DSH settings card or the settings file, and changes take effect instantly without a restart. Sessions routed to a vision-capable model pass through untouched.

It is built for DSH users whose main model is purely textual but who still want to paste images mid-conversation for visual grounding, especially those who route between multiple models and would rather not switch the entire session just to look at a single picture.