dsh-vision-router

Paste an image and it just works: a free built-in vision chain out of the box — DeepSeek keeps thinking, the vision model only does the seeing. No key, no Python, one command.

What problem it solves

Text-only models can’t see images in DSH, and most vision plugins bridge images as text descriptions — lossy, one-shot, and blind to pixels, which makes GUI location and UI restoration impossible. The usual vision setups also demand an API key or a Python environment.

This plugin takes a different route: it makes looking at an image an ordinary tool call — original pixels stay on the vision model’s side while DeepSeek stays on the reasoning side. The agent drives tools like vision_ground (pixel-level location), vision_crop, and vision_pixel_diff across as many steps as the task needs, so image turns are as grounded, measurable, and repeatable as tool-calling turns. A built-in OVHcloud anonymous fallback chain (five models) needs no account or key, and your own vision models run first with automatic failover.

Key features

• Free vision chain out of the box: a built-in OVHcloud anonymous fallback (five models), no account or key; your own models run first and fail over automatically
• Fourteen deep tools: vision_ground (original-pixel box location), vision_detect, vision_crop, vision_pixel_diff, vision_describe, vision_ocr, vision_trace, vision_extract_foreground, vision_html_screenshot, and more, plus a desktop-screenshot tool that is off by default
• Whole-turn auto-routing: pick a ”+ Auto Vision” model group in the selector, paste an image, and the whole turn routes to vision — the original model group stays untouched
• Image memory: answers are cached by image content hash, so later text turns reuse the record without re-spending vision calls
• Verifiable pixel loop: reference → screenshot → vision_pixel_diff ratio + heatmap → fix → repeat, making UI restoration measurable instead of eyeballed
• Local vision backends: optional Ollama / LM Studio recognition for offline, free, private use
• No Python: sharp / potrace / tesseract / system Chrome handle downscale, grounding, crop, OCR, SVG trace, and screenshots

Quick start

Install with one command, then pick a ”+ Auto Vision” group from the model selector in the chat composer’s lower-right corner and paste or upload an image. For example, have the agent use vision_ground to locate a “send button” and answer with coordinates, use vision_pixel_diff to compare a reference with an implementation screenshot, or run vision_long_screenshot_ocr on a long screenshot. Local vision is enabled in settings or a profile patch via localOllama / localLmStudio.

Who it is for

• DSH users who want to give text-only DeepSeek “eyes” without a key or Python
• Developers and testers who need GUI location, pixel-level operations, and quantifiable UI-restoration acceptance
• Users who want offline, private recognition through local Ollama / LM Studio