dsh-vision-bridge

Text-only DeepSeek sessions often face a dilemma when handling image analysis: traditional multimodal approaches require sending the entire conversation history along with images, leading to context bloat and soaring costs. The dsh-vision-bridge plugin precisely addresses this pain point by injecting on-demand vision capability into sessions—when the text model determines it needs to view screenshots, charts, or uploaded images, it automatically invokes the vision_describe tool.

The core of this plugin lies in its elegant 'minimal-touch' design: it intercepts image attachments in the session via hooks, but only sends images and focused questions to an OpenAI-compatible vision model when absolutely necessary, completely bypassing the conversation history that can span hundreds of thousands of tokens. This ensures costs remain controlled while the session log and user interface fully preserve original images for a seamless interaction experience. Users can also freely configure their own vision endpoints, whether cloud services or local models, for effortless integration.

This plugin is particularly suited for two groups: first, developers and knowledge workers who rely on DeepSeek Harness for efficient text-based work but frequently need to temporarily analyze document images, UI screenshots, or data charts; and second, teams looking to expand AI multimodal capabilities at the lowest cost, without switching their entire model ecosystem for occasional image needs. It empowers text sessions to maintain lightweight efficiency while gaining the ability to derive visual insights on demand.