dsh-vision-bridge

DeepSeek Harness ships with v4-flash and v4-pro as text-only models: the API rejects any image bytes, paste is blocked at the composer gate, read_image refuses to fire, and the whole session is effectively blind to visuals. dsh-vision-bridge re-registers those same DeepSeek models as a "vision twin" route that declares image support. Before each request leaves your machine, every image block is intercepted at the adapter layer, fed to a local Ollama vision model (qwen2.5vl by default), and replaced with a rich text description. The end result feels native—paste a screenshot, get a thumbnail and an image block, and the model answers as if it had seen the picture—while not a single image byte ever hits DeepSeek's API.

Beyond basic image understanding, the plugin exposes nine inspection tools: full-image description, OCR with normalized coordinates, structured element scanning, region-focused querying, two-stage object localization, side-by-side image comparison, clipboard capture, and a zero-inference health check. All tools run locally, validate output against a schema with automatic retry, and benefit from a content-hash cache so repeated images cost nothing extra. The paste router consults live model metadata to decide per session whether to keep the native photo flow or fall back to a local file path—no name-based regex, no manual switching.

If you code in Harness with DeepSeek on a daily basis and need to inspect screenshots, read UI mockups, or run visual regression checks without swapping to a different model or hunting for a cloud vision key, this 29 KB plugin is the lightest possible bridge: install it, pick the twin route in the selector, and go.