dsh-vision-resident

When working with a pure text model, a screenshot, a UI defect, or a snippet of code captured in an image often becomes a blind spot. dsh-vision-resident embeds a resident vision sub-agent directly into the DeepSeek Harness host process, letting your text-based main model see images in the same conversation—pasted screenshots are auto-translated, follow-up questions reference prior recognition results, and no separate vision session is needed.

The vision sub-agent maintains its own persistent context. The last ten recognition rounds are isolated per DSH conversation, so follow-ups never require re-sending the original image. Recognition history and translation caches are written to disk, surviving restarts without loss or duplicate work. When the main conversation is archived, the corresponding vision records are automatically purged. A status capsule in the input bar shows real-time workload and recognition count; clicking it reveals a collapsible conversation log where each entry expands to show the instruction, reasoning trace, and output.

Ideal for developers who use DeepSeek or other text-only models for coding, UI review, or OCR tasks. There is no need to switch models or open a separate tool—just paste an image and keep asking in the same conversation. Vision capability feels like a native part of your workflow.