dsh-vision-mix

In DeepSeek Harness, many Agents are text-only by default, limiting their ability when users upload images or need to analyze web screenshots. The dsh-vision-mix plugin addresses this gap by combining text, vision, and optional image generation APIs into a cohesive 'Mix' model, empowering Agents to recognize, deeply understand, and even follow up on images. Its core strength lies in intelligent routing: the plugin automatically distinguishes between plain text, newly uploaded images, or ambiguous follow-up questions, dispatching the appropriate model accordingly. For instance, when a user asks 'What else is in this picture?', it re-reads the most recent image; for vague cues like 'Give more details', an intent model can decide if image analysis is needed again. Additionally, the plugin supports independent image generation and editing, letting users create new images or modify existing attachments directly. For developers or users seeking visual enhancement, dsh-vision-mix offers flexible configuration—assign different models and Providers to image recognition, base chat, and image generation, and safely enable image capabilities through testing tools. Whether you're building multimodal Agents, want an Agent to interpret design drafts and generate code, or need a chat assistant with image processing skills, this plugin integrates seamlessly into existing workflows, boosting overall intelligence and interactivity.