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

• The image admission preflight (attachment-error in dsh-host-apiproxy) is bypassed: no image block is ever sent, the model only sees text.
• Nothing is typed into the composer while you add files — paths are appended on send, exactly like the image conversion.
• Upload failure is never silent: a visible notice explains what happened before falling back to the native send path; messages are never swallowed.
• Supports paste and drag-and-drop; images, files and mixed batches are uploaded in order and sent as one text block of paths.

Supported file types

Type	Extensions	Limit	Experience
Images	png jpg jpeg webp gif	30MB	Native attachment card (preview / remove), auto-converted to paths on send
Documents	pdf doc docx xls xlsx ppt pptx txt md csv json	100MB	Square chip in the attachment rail (format icon, full name on hover, ✕ to remove), path appended on send
Archives	zip	100MB	Same
Video	mp4 mov webm mkv avi	100MB	Same
Audio	mp3 wav flac m4a	100MB	Same

Images keep the native attachment experience. DSH itself only accepts these image formats as attachments (see the toast below) and refuses everything else — this plugin turns the rest into plain workspace paths instead, so the agent can process them with PDF / document tools.

Pairing with dsh-vision-toolkit (recommended)

🎯 Strongly recommended to pair with dsh-vision-toolkit — it is the eyes of a text-only model, this plugin is the conveyor belt:

• This plugin solves "how an image or file becomes a path" (automatic conversion on send);
• dsh-vision-toolkit solves "how a path becomes vision capability" (intent-aware image Q&A, long-screenshot OCR, UI restoration, grounding, pixel diff — 10 structured vision tools, with DSH Credentials / managed runtime / Web Settings).

Install: dsh plugin --profile web add <path-to-dsh-vision-toolkit> (see its repo README). Together they give a text-only model an experience close to a native multimodal model.

Scenario	Before	Now
Image Q&A	screenshot → save to workspace → type path	screenshot → Ctrl+V → Enter
Multi-image diff (pixel diff)	same, type every path	just send two images
Long-screenshot OCR / UI restoration	same	just send the image
Analyze a PDF / xlsx / video	copy it into the workspace, type the path	just drag it in and send

Comparison with similar plugins

• dsh-drag-and-drop: dropping a file inserts its raw path into the composer, no file copying. Great for referencing files already in the workspace; images are still sent as attachments (still rejected by text-only models), and paste is not supported.
• This plugin: keeps the native attachment UI (preview / remove), automatically copies images and files to the workspace and converts them to paths on send, supports paste and drag-and-drop, and shows non-image files as removable square chips in the attachment rail. Built for "sending images and files to a text-only model".

They are complementary: use dsh-drag-and-drop to reference existing files, use this plugin to send screenshots and new files.

Installation

Requirements: DeepSeek Harness (Web profile), Node.js, dsh CLI.

# Option 1: install via dsh plugin (recommended)
dsh plugin --profile web add /path/to/dsh-drop-to-path

# Option 2: manual (matches this repo layout)
# 1. add to dependencies in package.json:
#    "@dsh-external/dsh-drop-to-path": "link:/path/to/dsh-drop-to-path"
# 2. add to dsh.profile.bundles: "@dsh-external/dsh-drop-to-path"
# 3. copy the directory (including lib/) into profiles/<name>/node_modules/@dsh-external/dsh-drop-to-path

Restart the Web profile after installing (relaunch the launcher / restart dsh web). No settings to configure.

Usage

1. Paste or drop images as usual — the attachment cards behave natively (preview / remove);
2. Drag or paste any other file (pdf / office / zip / video / audio) — a square chip with a format icon appears in the attachment rail next to the image thumbnails (full name on hover, ✕ to remove);
3. Type your message (optional) and hit send — no attachments are sent; images and files are uploaded and their workspace paths are appended to the message automatically;
4. The agent (paired with dsh-vision-toolkit) reads images with vision_glance / vision_pixel_diff and friends, and handles documents with PDF / spreadsheet tools.

Files are stored under <workspace>/.drops/; clean them up whenever you like.

File structure

dsh-drop-to-path/
├─ package.json        bundle manifest (dsh.bundle.patch / dsh.client)
├─ cordis.patch.yml    mount row (insert drop-to-path)
├─ lib/
│  ├─ index.js         host:POST /_dsh/drop-to-path/import route
│  └─ client.js        browser: interception + sendSession wrapper + chips
├─ assets/             demo GIF, combined effect image, social preview, toasts
├─ README.md
├─ README.zh.md
└─ ADAPTING.md         upgrade adaptation guide (read before upgrading DSH)

Implementation notes

• host side (lib/index.js): registers the webServer route POST /_dsh/drop-to-path/import, accepts { name, dataBase64 }, validates (whitelisted extensions; images ≤30MB, other files ≤100MB; Unicode-safe sanitized file names against path traversal), writes into the session workspace .drops/, returns { ok, value: { path, kind } }.
• Workspace resolution: reads $DSH_HOME/storages/workspace.json (the durable workspace registry) and picks the workspace with the newest updatedAt.
• browser side (lib/client.js):
  • pure-image drops pass through untouched (native rail + overlay handling);
  • drops/pastes that contain non-image files are intercepted: files upload in order, chips are rendered, and a synthetic dragend closes DSH's full-screen drop overlay so the page never gets stuck;
  • conversation.sendSession is wrapped on the ConversationController prototype (survives instance recreation): on send, image attachments and queued files are uploaded and their paths are joined with your text into one message;
  • chips are square tiles sized live from the measured image-thumbnail size (falls back to 62px), with per-format icons/colors and a MutationObserver that re-renders them after React updates;
  • after a successful send the chips clear; upload failures show a visible notice and fall back to the native send path — messages are never lost.

Compatibility

DSH version	Status
0.1.0-rc.6 (verified in this repo's environment)	✅ Works

This plugin depends on several undocumented internal interfaces of DSH (service names, method signatures, storage formats) and may break after a DSH upgrade. Read ADAPTING.md before upgrading — it contains the failure symptoms and the fix steps.

Contributing & adapting

• Issues / ideas: GitHub Issues.
• Plugin broken after a DSH upgrade: follow the symptom table in ADAPTING.md, fix, and append a row to the adaptation log.

License

MIT