loudMore/dsh-drop-to-path

dsh-drop-to-path

Drop or paste images, PDFs, office docs, zips, videos and audio into DeepSeek Harness — they reach your text-only model as workspace file paths. Images keep the native attachment UI (preview / remove), everything else shows as square chips in the attachment rail, and paths are appended automatically when you hit send.

English | 中文

Demo

Drag images / PPT / TXT / ZIP into the composer → square file chips appear next to image thumbnails → hit send → the model receives workspace file paths.

The effect: images and files arrive at the text-only model as workspace paths, ready for vision / document tools.

Why this plugin

I use a text-only model (deepseek) in DeepSeek Harness and frequently need to send screenshots to the agent. But DSH image attachments go through the model's native attachment channel, and text-only models are rejected by a preflight check before the message is even sent — the composer shows a system toast:

The current model does not support images; switch to a model that does.

Backed by the host image-admission preflight (attachment-error):

Model "deepseek-v4-flash" does not support image input.

dsh-vision-toolkit solves the "see the image" half — but its tools only accept workspace file paths, so the image must enter the message as a path. Every screenshot became: save to workspace → manually type the path into the message. Painful. And beyond images, DSH refuses to attach PDFs, office documents, videos or any non-image file at all.

After experiencing "paste and go" in another tool, I decided to bring that experience back to DSH: keep the attachment UI completely native, and convert images and files to paths only at the moment of sending. Hence this plugin.

How it works

flowchart LR
    A["Paste / drag in"] --> B{"File type?"}
    B -->|"Images<br/>png / jpg / webp / gif"| C["Native attachment card<br/>preview / remove"]
    C --> D["Hit send"]
    D --> E["Upload → workspace/.drops/"]
    E --> F["Message becomes plain text paths"]
    F --> G["Text-only model reads them<br/>with vision / document tools"]
    B -->|"Other files<br/>pdf / docx / xlsx / video…"| H["Square chip in the<br/>attachment rail<br/>format icon · ✕ remove"]
    H --> D