dsh-vision-auto

Image drag-and-recognize plugin for DeepSeek Harness. Drag images into the Web GUI — they keep showing as images, image turns are answered by a vision model, and text turns fall back to the session model with image history shielded. Text-only models (such as DeepSeek) never hit UNSUPPORTED_CONTENT.

中文文档 · dsh-anydrop (full: files + folders + images)

This is the image-only subset. If you also want file/folder drag-and-drop, use the full plugin dsh-anydrop.

Attribution — the image-recognition and model-routing module is derived from ysr666/dsh-vision-router (MIT). The image drag-and-drop module is independently implemented by this project; it does not come from dsh-vision-router. See Attribution.

Features

• Drag images — the client intercepts image drops and reuses the official draft-image pipeline, so the GUI shows the original image thumbnail.
• route mode (default) — image turns switch to a harness-registered vision model and read raw pixels; when the turn ends, the session model is restored and the UI badge flips back immediately.
• tool mode — image → address text in history; the model calls the built-in vision_describe tool, which talks directly to any OpenAI-compatible vision model (token only, no harness provider).
• Guards — drops an unsupported reasoningEffort/inherited maxTokens, and shields historical images model-only on text turns.

Install

cd ~/.dsh/profiles/web
pnpm add dsh-vision-auto@github:babutiger/dsh-vision-auto

Add to ~/.dsh/profiles/web/cordis.patch.yml:

- insert:
    - id: vision-auto
      name: 'dsh-vision-auto'
      config:
        mode: route                        # route / tool / off
        visionProvider: qwen-dashscope     # route mode: a harness-registered vision provider
        visionModel: qwen3.7-plus
        apiKeyEnv: QWEN_API_KEY            # tool mode: vision-model credential name
        baseURL: https://dashscope.aliyuncs.com/compatible-mode/v1

Then restart dsh web and refresh the page.

API key

• tool mode needs only a token for any OpenAI-compatible vision model; put it in ~/.dsh/.credentials.yaml under the apiKeyEnv name (or set apiKey directly).
• route mode needs the vision provider registered in the harness (for example via llm-pi-ai).

Modes

Mode	How it works	Requires
route (default)	Image turns route to a harness-registered vision model (GUI shows the original image); other turns restore the session model.	a registered vision provider/model
tool	Image → address text; the model calls vision_describe against an OpenAI-compatible vision model.	only a token (apiKey / apiKeyEnv)
off	Only the sanitize/restore/history-shielding guards; no routing, no vision_describe.	—

Configuration

Field	Default	Description
mode	route	route / tool / off
visionProvider / visionModel	'' / qwen3.7-plus	route-mode vision model
apiKey / apiKeyEnv	''	tool-mode credential (literal value / credential name or env var)
baseURL	https://dashscope.aliyuncs.com/compatible-mode/v1	OpenAI-compatible endpoint
visionMaxTokens	4096	max tokens for the vision call
visionTimeoutMs	120000	vision-call timeout
sanitizeReasoningEffort	true	drop reasoningEffort the target model doesn't support
sanitizeMaxTokens	true	drop inherited maxTokens when switching models
restoreSessionModel	true	switch back to the session model on text turns
rewriteHistoryImages	true	shield historical images model-only on text turns

How it works

• Browser client (lib/client.js): intercepts image drops and calls the official createDraftImages + addImages path; non-image drops are left to the default behavior.
• Plugin host (index.js): agent/pre-step accumulates image turns and shields previous turns' images model-only via the surface-replace seam (the transcript keeps the original image); agent/request routes image turns to the vision provider (route mode) and sanitizes reasoningEffort/maxTokens; agent/turn-stopping republishes the default model context so the UI badge flips back immediately. In tool mode it registers the vision_describe tool.

Attribution

The image-recognition and model-routing module in this plugin is derived from ysr666/dsh-vision-router, licensed under the MIT License. That module reuses the upstream blocksHaveImage, eventHasImage, extractJson, createCache, and toOpenAIContent utilities and its image-turn routing, history shielding, and vision_describe architecture.

The image drag-and-drop module is an independent implementation in this project. Its image-drop event handling, integration with the official Harness draft-image APIs, and related client logic do not come from dsh-vision-router. The project as a whole is distributed under MIT, with the upstream copyright and permission notice retained; see LICENSE.

License

MIT