dsh-vision-toolkit

If you've been using text-only models in DeepSeek Harness, you know the pain: you want the model to look at a screenshot, read a long chat history, or turn a sketch into a page—but you end up copying file paths, manually switching models, and getting a vague "image caption" in return. DSH Vision Toolkit solves exactly this. It gives text-only models real eyes, and not just for generic descriptions—it extracts task-relevant evidence: where the error is, where the button is, how far the implementation diverges from the design.

The plugin packages a battle-tested visual methodology into DeepSeek Harness. It ships 10 tools that can be called independently or composed into workflows: glance, grounding, element detection, cropping, vector tracing, pixel diff, long-screenshot OCR, foreground extraction, dominant-color extraction, and HTML screenshotting. With the bundled vision-skills Skill, the agent learns what to inspect for each task, which tools to call in what order, and how to verify the result. And in DSH Web, you can simply paste an image—the model automatically routes to its vision-capable variant, no manual configuration needed.

It's built for two kinds of users. First, anyone who wants a multimodal-like experience: paste an image, ask a question, let the toolchain handle the rest. Second, those with higher-value visual work: turning a sketch into a front-end page, rebuilding a screenshot as editable HTML, or extracting chat messages from long scrolling captures—not toy "describe this image" demos, but reusable, verifiable workflows.