dsh-image-subagent

Pure-text main models like DeepSeek-v4-Pro often struggle with image attachments, as they cannot directly parse visual data, limiting their utility in multimodal tasks. The dsh-image-subagent plugin addresses this by enabling pure-text models to receive and process images indirectly.

The plugin's core capability lies in projecting images into text placeholders. When an image is attached, it is converted into an explicit placeholder containing a full attachmentId. The main model then delegates to a vision subagent (e.g., DeepSeek-v4-Flash-Vision-Exp) which uses the read_image tool to fetch pixel-level descriptions from the image path. This delegation ensures the main model can handle images without native multimodal support.

This plugin is ideal for developers, researchers, and users who rely on pure-text models but need image processing capabilities. It suits applications like building smart assistants, conducting image analysis, or extending model functionality, allowing seamless integration of image handling while maintaining the main model's text-only nature.