dsh-ocr-vision

Plugging a text-only model into DeepSeek Hits a hard wall: the moment the model wants to look at an image it calls read_image, but the text-only route rejects image input at the tool-fs layer and the model simply sees an error. dsh-ocr-vision sidesteps the problem entirely by introducing an ocr_image tool that returns plain text. Image bytes never leave the host; RapidOCR does all the extraction locally and hands back a clean text envelope, so the model only ever deals with strings.

Under the hood the plugin bundles a RapidOCR engine (2x upscaling, region cropping, coordinate output, empty-image fallback), a SKILL.md that walks the model through a three-step flow (clarify purpose, extract text, organise around the task), and a system-prompt segment that nudges the model toward ocr_image on text-only routes. Every file read goes through the ctx.fs sandbox policy, so images stay on disk and never touch an external API.

It is built for developers running pure-text routes in DeepSeek Harness who still need their model to make sense of screenshots, diagrams, or document scans. It mirrors the workflow and output format of claude-ocr-vision, making it a zero-cost drop-in equivalent on the DeepSeek Harness side.