dsh-vision-tools

Pure text LLMs like DeepSeek inherently lack the ability to perceive images, creating a barrier when users need to handle visual content directly within a chat interface. dsh-vision-tools, as a comprehensive vision suite for DeepSeek Harness, aims to seamlessly grant 'sight' to these text-only models. It achieves this by registering a global tool called vision_understand, which invokes a vision-capable large model API (defaulting to the free Zhipu GLM-4.6V-Flash) compatible with OpenAI's protocol to perform image description, text recognition, or visual question answering.

The plugin's core strength lies in its smooth interaction integration and robust backend resilience. Users can easily trigger image recognition via three methods: pasting a screenshot from the clipboard (Cmd/Ctrl+V), dragging an image onto the dedicated '📷 Identify' button next to the input box, or clicking that button to select a file. The image is saved automatically and recognition is initiated. On the backend, it supports switching between four major vision model providers—including Zhipu, Alibaba's DashScope, SiliconFlow, and OpenAI—and features an intelligent rate-limit fallback mechanism. If the primary model hits rate limits, it automatically retries with a fallback model, ensuring high availability for recognition tasks.

Consequently, dsh-vision-tools is ideal for any user looking to seamlessly integrate image understanding capabilities into their DeepSeek Harness conversations. Whether for rapid screenshot analysis, chart interpretation, or building AI workflows that combine text and visuals, it offers a low-barrier, highly flexible solution. For developers and technical users, its configuration-based provider switching (which doesn't require restarts) and project-level overrides also provide excellent customizability.