dsh-file-upload 📎

English | 简体中文

One upload button + drag-and-drop files straight into the conversation — a DeepSeek Harness (DSH) web plugin.

Click the 📎 upload button next to the composer (or just drag any file into the window): the file is saved to the current project's uploads/ directory and the absolute path is inserted into the input box as [上传文件] /path/to/uploads/xxx.png. Press send and the model — or any attached vision tool — reads the file by path.

Vision-plugin agnostic: the message only carries a local absolute path (plain text), so it works with dsh-vision's view_image, any other model/tool that can read local paths, or no vision at all. It bypasses DSH's native "the current model does not support images" rejection because no image block is ever submitted.

Screenshot

The 📎 upload button in the composer tool row; the picked file's path (blurred in the screenshot) is inserted into the input box automatically, ready to send.

Features

Action	Effect
Click 📎 upload	System file picker (multi-select) → save → path into the input box
Drag a file into the window	Images & any file type are taken over (no "unsupported" toast) → save → path into the input box
Send the message	The model / vision tool reads the file by absolute path
Switch sessions	Button follows the current session; files land in that session's project uploads/

• Single-file limit: 25 MB (frontend) / 30 MB (backend)
• Filenames keep Chinese/space characters; a timestamp prefix avoids collisions
• Button shows busy state while uploading; failures surface as Chinese notices

Install

Official bundle install (recommended)

dsh plugin --profile web add "github:a903067276-rgb/dsh-file-upload#main"

Restart dsh web (or use scripts/restart-dsh-web.sh). Requires pnpm on PATH (dsh plugin forwards to pnpm).

Manual mount (fallback, macOS-tested)

See docs/install.md: symlink into ~/.dsh/profiles/web/node_modules/ plus a single entry in ~/.dsh/cordis.patch.yml, then restart.

Usage

1. Click 📎 and pick files, or drag files anywhere into the window.
2. The input box gets [上传文件] <absolute path> lines (your existing draft text is kept).
3. Press send; the model calls its vision tool on the path automatically.

Platform support

Platform	Status
macOS	✅ fully tested (development environment)
Linux	✅ expected to work (pure Node implementation), untested
Windows	⚠️ expected to work (pure Node implementation, Windows-safe filename sanitization, platform separator paths), untested

Requirements

• DSH web (run with dsh web)
• No extra shell needed: the host half is pure Node (node:fs), no system commands required on any platform.

How it works

• Host (lib/index.js): one route POST /api/file-upload/save — validates the session and size, then writes the base64 payload to <session cwd>/uploads/ with pure Node (node:fs, no system command dependency, cross-platform); the returned path is built via node:path and follows the platform separator.
• Client (lib/client.js): registers the 📎 upload button in the conversation.input.left seat (visually distinct from the default "+" command button); a capture-phase document listener takes over file drags before the official InputBar's bubble-phase listener (which would reject images); FileReader reads base64, uploads it, then the path text is appended to the input draft (inputActions.setDraft).
• Error boundary: a render crash degrades to a small "⚠ upload component error" chip instead of unmounting the whole composer.

Notes

• The uploads/ directory only grows; it is never cleaned automatically (we don't delete your files) — remove files manually when needed.
• scripts/restart-dsh-web.sh restarts dsh web gracefully (kills the port listener, waits, relaunches, health-checks); the browser auto-reconnects, no manual refresh needed.