dsh-file-upload

dsh-file-upload ⬆️

English | 简体中文

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

Unofficial project: independently developed and maintained by a community member, not an official DeepSeek product.

Fork of a903067276-rgb/dsh-file-upload (MIT). v0.2.0 adds uploaded-file management and clipboard path reading (zero-copy) — see What's new.

Screenshot

The upload icon button in the composer tool row (official DSH design tokens, follows dark/light theme); either the file's path is inserted into the input box, or — for supported images — the image lands in the official attachment rail (auto file_id reuse).

Features

• Single-file limit: 64 MB (DeepSeek Files API hard limit; the local attachment store defaults to 20 MB — see Large images below)
• Everything lands in the attachment library: ~/Documents/DSH/Attachments/images/<YYYY-MM-DD>/ (images) or .../files/<YYYY-MM-DD>/ (other files) — configurable in Settings
• Button shows busy state while uploading; failures surface as Chinese notices

What's new in v0.2.0 (this fork)

• Uploaded-file management — browse the attachment library right in the settings card ("文件上传" section): image / file / folder tabs, paged newest-first, original name / size / mtime / relative path per row, one-click copy @path or delete (confirm dialog + host-side containment check against the attachment root).
• Clipboard path reading (zero-copy) — copy a file/folder in the OS file manager, then click the clipboard button next to the upload icon: the host reads the system clipboard's real path list (Windows FileDropList / macOS furl / Linux uri-list) and inserts @path / @"path" references into the draft. No content transfer, no disk copy, no size limit — the model reads the file by address.
• Paste path text → @ references — with "listen clipboard = all files" enabled, pasted plain absolute-path text is converted to @ references automatically.
• Fix — Windows "open folder" no longer reports failure (explorer.exe exits with code 1 even on success).

Settings card

• Attachment directory (~/Documents/DSH/Attachments default, ~ supported) — only used for image archives
• Path prefix ([上传文件] default) — text prepended to paths in the draft; clear it to send bare paths
• Images via official attachment (default on) — off = images follow the old path-text logic
• Archive images to the attachment directory (default on) — off = official attachment only (saves disk; if the official channel is unavailable, the image is still force-archived)
• Allow public uploads (default off) — off = same-origin check keeps allowing localhost only (CSRF-safe); on = any origin may upload, for public/port-forwarding access (e.g. ddnsto). Enable only if you trust everyone who can reach your DSH.
• Clipboard listener (default off, keeping official built-in paste; options: off / images only / all files) — paste interception mode. Off = official built-in paste; images only = the plugin takes over pasted screenshots/images (works even when the model has no vision support — falls back to path text); all files = any pasted file is handled (archived to the library + path text in the draft). Changing the mode takes effect immediately after saving, no page reload.
• Read-only display: current image size limit from the host

Large images (20–64 MB)

DeepSeek accepts up to 64 MB per image. DSH's local attachment store defaults to 20 MB; images above that (and ≤64 MB) are archived + path-text referenced (the model can still read them via read_image, which follows the same store limit).

To let large images through the official attachment path too, add this to ~/.dsh/profiles/web/cordis.patch.yml and restart dsh web:

- id: attachment-local
  config:
    maxImageBytes: 67108864   # 20 MiB → 64 MiB (DeepSeek hard limit)

Note: the whole config row is replaced, so keep every key you need; re-check against the DSH version after upgrades. The model always sees the harness-normalized version (≤2048px / ≤4 MiB, ≤384 tokens per image) regardless of the original size.

Install

Official bundle install (one line):

dsh plugin --profile web add "github:666emmm/dsh-file-upload#main"

Restart dsh web (bundle layers are composed at startup). Requires pnpm on PATH (dsh plugin forwards to pnpm).

Manual mount (fallback): see docs/install.md — symlink into ~/.dsh/profiles/web/node_modules/ plus a single entry in ~/.dsh/cordis.patch.yml (a double entry makes the plugin apply twice and crash on duplicate route registration), then restart.

Loader entry id: this project registers the loader entry id file-upup, not file-upload. DSH's own @deepseek-ai/dsh-web-app bundle already claims file-upload (pointing at @deepseek-ai/dsh-client-file-upload), and loader entry ids live in one flat namespace: two rows with the same id make the loader fail hard at mount time (duplicate loader entry id: "file-upload") and the whole profile refuses to boot — the market's pre-update trial validation rolls back for the same reason. Copy id: file-upup from cordis.patch.yml when mounting manually. The settings namespace and HTTP routes keep the file-upload spelling on purpose (unrelated to the entry id, and renaming them would orphan existing user settings).

Usage

1. Click the upload icon and pick files (multi-select), or drag files/folders anywhere into the window.
2. Images (when the current model accepts images): archived to the attachment directory and shown in the official attachment rail — send, and the model sees the image (DeepSeek Files API file_id, auto-reused).
3. Images when the model does not accept images (or you turned the official path off): archived to images/, then <prefix> <absolute path> lines go into the draft — e.g. [上传文件] /path/to/Attachments/images/xxx.png — and your existing draft text is kept.
4. Other files: archived to files/, path text goes into the draft. Press send; the model reads the file by path.
5. Folders: drag one in and it is rebuilt under the attachment library with its original layout; only the folder path (one line) goes into the draft.

Version compatibility (dsh upgrades)

For dsh v0.1.2-alpha.1 and newer: the @deepseek-ai/dsh-client-runtime package was removed and ctx.slots is now provided by @deepseek-ai/dsh-client-ui-renderer — replace @deepseek-ai/dsh-client-runtime with @deepseek-ai/dsh-client-ui-renderer in the dsh.client.inject list of package.json after upgrading, or the client module (buttons / settings card) will not load. Everything else (bundle patch, cordis patch syntax, webServer.register, settings.register, slots, settingsScope, __ModuleLoader__.load) is compatible. No change needed on the current 0.1.1-rc.2. Full checklist: COMPAT-dsh版本.md.

Platform support

Requirements

• DSH web >= 0.1.0-rc.7 (run with dsh web)
• Version compatibility (best effort — new features verified locally on 0.1.1-rc.2 and 0.1.2-alpha.1 (source checkout); on 0.1.0-rc.7/rc.8 the official image rail can't be fully checked, not guaranteed):
• Maintenance policy: this plugin keeps evolving with the latest DSH releases; compatibility with older DSH versions is best-effort only and not guaranteed going forward.

• Pure Node except clipboard: everything (save / folder rebuild / list / delete / settings) is node:fs only; clipboard path reading shells out to the platform clipboard tool (PowerShell Get-Clipboard -Format FileDropList / osascript / xclip|wl-paste) — a documented platform limitation.

How it works

• Host (lib/index.js): POST /api/file-upload/save — validates the session and size, writes the base64 payload to <library>/images/<YYYY-MM-DD>/ (mode=image) or <library>/files/<YYYY-MM-DD>/ (mode=file) with pure Node; POST /api/file-upload/save-folder — takes a relative-path + base64 list and rebuilds the tree under <library>/files/<date>/<timestamp>-<folder>/ (each segment sanitized, .. rejected to prevent traversal); GET/POST /api/file-upload/config reads/writes the settings (official settings service) and exposes the host image limit plus whether the current session's model accepts images (llm.resolveModel inputModalities — same source the adapter uses).
• File management + clipboard (v0.2.0): GET/DELETE /api/file-upload/files lists (paged, kind-filtered) and deletes (containment-checked) the attachment library; POST /api/file-upload/clipboard-paths reads the OS clipboard's real file paths (Windows FileDropList / macOS furl / Linux uri-list) — inserted as @ references, zero-copy.
• Client (lib/client.js): registers the upload icon in the conversation.input.left seat; capture-phase document listener takes over file drags and reads dragged folders via webkitGetAsEntry; routing: supported image + official on + model supports + within host limit → archive + conversation.createDraftImages + inputActions.addImages (the official InputBar's own mechanism) → official attachment rail (no path text); anything else degrades to archive + path text; >64 MB is refused with a notice.
• Error boundary: a render crash degrades to a small "⚠ upload component error" chip instead of unmounting the whole composer.

Notes

• The attachment library only grows; it is never cleaned automatically (we don't delete your files) — remove files manually when needed.
• After modifying the plugin, restart dsh web for changes to take effect (client-side edits apply on a page refresh; host edits need a restart).

Why this plugin exists

DSH natively rejects dragged-in images when the current model doesn't support them. This plugin routes images through the official attachment path when the model can see them (so they ride the DeepSeek Files API with file_id reuse), keeps a user-accessible archive copy you can find later, and falls back to plain path text otherwise — a plain-text message that passes the model's image check and works with any model or vision plugin (it bypasses DSH's native image rejection because no image block is ever submitted on the fallback path).

License

MIT