YuLee-314/dsh-vision-bridge

README 摘要

dsh-vision-bridge A DeepSeek Harness plugin that lets text-only models receive and understand images. The vision work is done by a local model on your machine. English · 简体中文 What is this? DeepSeek Harness (dsh) is an open-source AI coding environment built entirely from plugins. Its chat models — deepseek-v4-flash and deepseek-v4-pro — are text-only : their API rejects image data. As a result, in Harness you cannot paste a screenshot into a session that uses them, attach an image to a message, or use the built-in read image tool. This plugin fixes that. It works in three layers: 1. A second model route that accepts images. The same DeepSeek models are registered again as a "twin" provider ( deepseek-vision ). Because the twin declares image support, the normal image features work: pasting produces a thumbnail and an image block, and read image is allowed. Before each request is sent to the DeepSeek API, the plugin converts every image in the conversation into a text description produced by a local vision model. The API only ever receives text; the model answers as if it had seen the image. 2. Nine inspection tools. describe image , extract text , structured scan , query region , d…

分类