dsh-image-auto-describe

The workhorse models in DeepSeek Harness — deepseek-v4-pro and its peers — are text-only. Paste a screenshot, a stack trace, or a data table into the conversation and the message bounces straight back. Yet for most users, dropping an image and asking a question is the most natural way to reach out to an AI, and that single gesture is exactly where the experience hits a wall.

dsh-image-auto-describe closes that gap with one fully automatic step. When a message carries an image, the plugin invokes a vision model (by default Qwen3-VL-32B via SiliconFlow or GLM-4V-Flash via Zhipu) to transcribe the image verbatim into text — error codes, logs, tables, chat snippets all captured — and then hands that text to your DeepSeek model so the conversation flows on as if you had typed it. During transcription the chat shows a subtle placeholder; once done it becomes a normal message while the original image stays available to preview. If every vision route fails, the plugin falls back to the original rejection notice rather than hallucinating. Vision routes can be switched from the Web settings page and take effect immediately. The session model never sees raw image bytes; token cost tracks text length, not image size.

This plugin is for anyone running text-only workhorse models in Harness who regularly pastes screenshots, error logs, data tables, or chat records into the conversation. The prerequisite is light: at least one multimodal vision route (SiliconFlow, Zhipu, etc.) must already be registered in the profile with its API key configured. The plugin does not register model providers itself — it intercepts the admission layer, translates images into text, and leaves all downstream reasoning to your session model.