dsh-youreyes: Provides a model-callable image recognition tool for DeepSeek Harness

dsh-youreyes is a DSH plugin specifically designed for text-only DeepSeek, aiming to solve the challenges of processing visual content. By providing an image recognition tool callable by the model, it enables DeepSeek to invoke the visual channel to obtain textual evidence when necessary. The plugin supports multiple channels including local Ollama, Gemini API, OpenAI-compatible endpoints, and Antigravity IDE, and features an automatic fallback mechanism. Its core capabilities include four task modes: glance, ocr, region, and compare, as well as an "auto two-pass" strategy for handling complex scenarios. Additionally, the plugin supports structured JSON evidence output, visual evidence memory, and content hash caching to minimize repeated recognition and optimize reasoning efficiency. Installation requires Node >=20 and dsh >=0.

Read More