dsh-youreyes

DeepSeek, as a powerful pure-text large language model, cannot directly process image inputs, often leaving users frustrated when needing visual understanding. The dsh-youreyes plugin is designed to solve this by acting as a "pair of eyes" for DeepSeek, enabling it to "see" and comprehend image content. Users simply paste images or provide file paths, and the model can automatically recognize and answer related questions without switching to other vision models.

The plugin's core strength lies in its flexible adaptability and intelligent recognition pipeline. It supports multiple backend channels, including local Ollama (zero-config, images never leave your machine), free Gemini API, and any OpenAI-compatible endpoint, catering to diverse user needs. The recognition process uses an "auto" tier that automatically escalates complexity, handling various task modes from simple identification to complex OCR, region focus, and image comparison, while outputting structured JSON evidence for accuracy and traceability.

Additionally, the plugin introduces visual evidence memory and content hash caching mechanisms. Recognition results are written into the session timeline for cross-turn reuse, avoiding redundant identifications and associated delays or costs; in-memory caching ensures each recognition request for the same image is executed efficiently. These features make DeepSeek reliable and practical in image understanding scenarios.

dsh-youreyes is suitable for all users of DeepSeek Harness who need to process images, especially developers, researchers, content creators, or everyday office personnel. Whether you're quickly parsing screenshots, comparing image differences, or having the model answer questions based on visuals, this plugin seamlessly integrates into your workflow, giving the pure-text LLM true vision capabilities.