dsh-vision-toolkit

In DeepSeek Harness, a text-only model confronted with an image can only produce a generic description, falling short of real visual tasks. DSH Vision Toolkit injects a set of visual tools and a field-tested Skill directly into Harness: paste an image and the model automatically switches to a vision-enhanced variant, extracting task-relevant details instead of generating a one-size-fits-all summary.

The toolbox covers ten capabilities—image Q&A, element grounding, region cropping, vector tracing, pixel-level diffing, sequential long-screenshot OCR, foreground extraction, dominant-color detection, and local page screenshot capture—while the bundled vision-skills Skill orchestrates five reusable playbooks for long-screenshot reading, UI restoration, icon and sketch reconstruction, and GUI operation. Multiple vision providers can be configured with automatic fallback for better availability.

It is built for users who want a multimodal-like experience on top of a text model, whether that means pasting a screenshot into a chat and asking a question, turning a design mockup into editable front-end code, or extracting structured content from a long conversation screenshot.