dsh-ext-vision-proxy

When working with text-only large language models such as DeepSeek-V3 or DeepSeek-R1, users often struggle with the inability to directly analyze images, limiting multimodal tasks. The dsh-ext-vision-proxy plugin addresses this by acting as an intelligent bridge, seamlessly injecting vision capabilities into these text-only models. By proxying calls to OpenAI-compatible vision model APIs, it enables LLMs to autonomously invoke the vision_describe tool to inspect, analyze, and understand images, screenshots, or diagrams, thereby expanding the models' application boundaries.

The plugin's core strengths lie in its seamless integration and flexible management. It embeds directly into the DeepSeek Harness chat interface, offering an intuitive toggle and model selector for easy activation or deactivation of the vision proxy. It supports multiple image sources, including local files, URLs, and hash references, and automatically resolves MIME types to ensure proper processing. Additionally, it features a 4-state session matrix that intelligently adjusts behavior based on model type and switch state, allowing text-only models to use the proxy while enabling multimodal models to handle images natively without interference.

dsh-ext-vision-proxy is ideal for developers, researchers, and end-users who work with DeepSeek text-only models and need to quickly add image understanding capabilities without switching models. Whether building interactive applications, conducting data analysis, or exploring multimodal interactions in educational or creative domains, this plugin provides a concise and powerful support system, empowering text-only LLMs to transcend their limits and embrace richer visual tasks.