dsh-vision-subagent

When running a DeepSeek Harness agent on a text-only model, a common gap is the inability to process images directly—screenshots, error logs, and design mockups all need manual transcription. dsh-vision-subagent fills that gap by giving the main agent a dedicated pair of visual eyes: image understanding is delegated to a one-shot subagent on a separately configured vision route (Kimi, MiniMax, or any OpenAI-compatible provider). Crucially, image bytes and the vision model's intermediate context never enter the main session window; only the final text answer returns to the primary model, eliminating the problem of large screenshots crowding out the context window.

Beyond basic image reading, the plugin supports multi-turn visual reasoning—the child agent can call read_image on additional workspace files before answering. In the Web composer, pasting an image triggers context-aware analysis steered by your draft message (debugging asks focus on error text, styling asks on outfit details), while the chat bubble stays clean with just your words plus thumbnails; the full analysis lives in a lightbox. When you need the original file for pixel-level work, vision_image_fetch materializes a full-fidelity copy under .dsh-vision/.

It is well suited for developers whose primary model is text-only but who regularly need to look at images; cost-sensitive teams that want vision calls billed on a separate route rather than mixed into the main model's token count; and compliance-conscious setups where image data must be kept out of the main session context.