dsh-vision-suite

DeepSeek is strong at text reasoning, yet when the task becomes reading a screenshot, diffing two UI captures, or checking how a page currently looks, Harness has no native visual tool. dsh-vision-suite adds vision on top of the official deepseek-official text route: images are sent to a configured OpenAI-compatible vision model only when a visual tool is invoked, and nothing leaves the machine unless a provider is explicitly set.

Six tools cover the full loop from understanding to manipulation. vision_describe answers questions over one to four images and supports structured screenshot evidence; vision_ocr works through a remote model or a local Tesseract backend; vision_crop slices by pixel coordinates and saves a persistent attachment; vision_compare reports the change ratio and a magenta diff map between two same-size screenshots; vision_palette extracts approximate dominant colors; vision_browser_capture takes a whitelisted web-page snapshot via headless Edge or Chrome. Every crop, diff, and web capture yields a persistent attachment ID that can be fed back into describe or OCR for iterative visual workflows.

Security is built in: the plugin is off by default, never hijacks routing, keeps API keys exclusively in Harness Credentials, and restricts web capture to an exact host whitelist with a fresh browser context each time. It is aimed at Windows-based DeepSeek Harness users who need the model to actually see screenshots and web pages, particularly in UI regression comparison, document OCR, palette verification, and automated visual-evidence pipelines.