dsh-subagent-vision

DeepSeek chat models cannot accept image input, and the harness refuses to switch an image-bearing session onto a text-only route. Wanting to have the agent read a screenshot, a chart, or a UI mock inside your regular DeepSeek conversation meant either swapping models, opening a second session, or copy-pasting OCR text by hand. dsh-subagent-vision removes that friction: you paste or drop an image into the composer exactly as before; the moment you hit send, the browser half uploads each draft image to a private temp file and appends the file paths to your prompt, so the outbound request carries text only and never trips image admission. The text-only main agent then calls the subagent_vision tool with those paths; a fresh subagent routed to a vision-capable model you chose under Settings > Vision Model reads the image with its own read_image tool and returns only the textual reading back into the same session. The primary model never sees the image bytes, and the session model never changes. The vision model is picked from a dropdown that enumerates every adapter-registered model declaring image input, defaults to qwen3.8-max, and is persisted across restarts. Intake stays fully native — thumbnail rail, caret behaviour, remove and undo are untouched — the plugin intercepts only at send time. This is ideal for users who work primarily with a text-only DeepSeek model but need to let the agent examine screenshots, diagrams, handwritten notes, or design mockups mid-conversation: paste, describe the task, keep the dialogue going, all in one session with no model juggling.