dsh-web-file-uploader

🌐 Language · English | 中文

A file-upload plugin for the DeepSeek Harness web UI. It adds a
DeepSeek-web-style paperclip attach button to the composer input row and
uploads the selected files to the DSH host machine — with model-aware
adaptation so the files are actually usable by the running model, and
content-addressed deduplication so storage never gets flooded by
re-uploads.

• Repository: https://github.com/Mooling0602/dsh-web-file-uploader

Features

• 📎 Paperclip attach button in the composer tool row, wearing the official
  attach button's chrome (28px circle, --dsw-specific-selector fill, 14px
  filled glyph, --dsw-alias-interactive-bg-hover-solid hover) with the
  paperclip rotated 45° to keep this plugin's long-standing diagonal, so the
  two buttons stay tellable apart at a glance. It sits directly to the right of
  the official attach button, ahead of the permission chips — and Settings can
  optionally hide the official button and let this one take its place
• Multi-file selection with preview-style attachment cards (image
  thumbnails) in the dock above the composer: reading → uploading → saved,
  with copy-path and remove buttons
• Files are stored on the DSH host, never only in the browser
• Name sanitization (path separators, .., control characters rejected) and
  automatic -1/-2 collision suffixes for distinct files with the same name
• Content-addressed deduplication — see Deduplication
• UI strings localized through the app's locale service (zh / en; follows
  the dsh web language setting or the system default)

Model-aware adaptation & the attachment-card design

Uploaded files are shown as attachment cards in the dock above the
composer. The cards are the source of truth for injection:

State	Behavior
Card present	The file's absolute path is injected into every user message sent while the card is visible
Card closed (×)	The plugin calls the host remove RPC — the file is dropped from the pending registry, no longer injected, and the file is kept on disk so the model can still re-read it from the uploads folder
Card deleted (🗑)	The plugin calls the host delete RPC — the file is permanently removed from the DSH host and the dedup index is scrubbed
Ctrl+V paste	Pasted images ride the native draft-image pipeline — no card is created and the plugin never touches them

The distinction between × and 🗑 is intentional: × just stops injecting
(you might want the model to find the file later), while 🗑 is a statement that
you are done with the file forever. Deleting the last reference to a file
removes it from disk; if another live session still references the same
deduplicated copy, the file is kept and only this session's reference is
dropped.

This is a deliberate design decision: the cards persist after sending (unlike
paste previews that clear on send), so you decide how long a file stays
"attached" to the conversation. Injection is one-shot per message, and the
injected block lists exactly the files whose cards are currently open.

Injection is model-aware:

Model type	Image files (png/jpeg/webp/gif)	Other files
Multimodal (reported inputModalities includes image)	Native ImageBlock via the attachments service — the image is part of the request, like a normal attached image, with no extra prompt text	Path text block
Text-only (e.g. DeepSeek V4 series)	Path text block — the model can call read/vision tools to inspect them	Path text block

The capability check uses llm.resolveModelInfo().inputModalities (cached for
10 minutes, safe text-only fallback).

The injected prompt uses a readable, English-only format (the model reads it):

-----
[Attached files] Some files have uploaded with this message:
- /path/to/file1.txt
- /path/to/image1.png
Read the files or use tools to analyse (like vision tools), then answer the user.

Compatibility with vision-tools and other plugins

• Zero coupling: the plugin only injects absolute file paths into the
  prompt. It never calls, wraps, or assumes any auxiliary tool — vision-tools
  or any other reader simply receives the path and works independently.
• Native multimodal path: for image-capable models, images are injected as
  native ImageBlocks; the model uses its own multimodal ability and is not
  prompted to reach for external tools.
• Non-invasive: injection targets only real user messages
  (source.kind === 'user'); steering/system messages are never touched, and
  a message that already carries the [Attached files] marker is never
  injected twice (guard against concurrent host instances). Pasting images
  via Ctrl+V is handled entirely by the product's native pipeline.

Deduplication

Re-uploads cannot flood storage:

1. On upload, the plugin computes the SHA-256 of the decoded bytes
   (dynamic mode pipes base64 -d | sha256sum through the shell service;
   the static bundle uses node:crypto).
2. A persisted index at uploads/.dfu-index.json maps hash → stored path.
3. If identical content is uploaded again (same name or different name), the
   existing stored copy is reused — no new file is written, no -1
   suffix copy is created, and the response carries dedup: true.
4. Uploads are serialized through a promise queue so concurrent identical
   uploads cannot race; a stale index entry (file deleted) falls back to a
   fresh store.

Scenario	Result
Same file uploaded N times	One copy on disk; all uploads resolve to the same path
Same content, different filename	Reuses the first stored copy
Same name, different content	Normal -1 collision handling (correct)
Process restart	The index file persists, dedup keeps working

The index entry is hash → { path, at } where at is the upload timestamp
(epoch ms) used by the TTL cleanup below. Legacy hash → path entries are read
transparently and upgraded to the object form on the next write; legacy entries
carrying no timestamp are never reclaimed by TTL (only scrubbed when the file
is gone).

Cleanup & retention (TTL GC + manual delete)

Uploaded files are reclaimed in two complementary ways:

Automatic TTL garbage collection

Every upload, every permanent delete, and plugin start trigger a lazy sweep
(after the success). Files older than the TTL that are not referenced by
any live session's pending set are removed from disk and their index entries
are dropped. Sweeps run on the same serialized queue as uploads, so they never
race in-flight writes; no background task is required.

• Default TTL: 7d (7 days).
• Retention units: 1s (seconds), 1m (minutes), 1h (hours), 1d
  (days); a bare number with no unit is treated as days (legacy behavior).
  0 (or an empty/invalid value) disables automatic GC — files are only
  removed when you explicitly click 🗑.
• Configure in Settings → File uploads (native DSH settings panel; the
  tab carries its own paperclip mark, so it never reads as the gear rows other
  plugins get): the retention field and the replace the official attach button
  switch share one Save button. The static bundle persists both in
  ~/.dsh/dsh-web-file-uploader.json ({ ttl, replaceOfficial }) and merges
  updates field by field, so saving one never clears the other. The environment
  variable DSH_UPLOAD_TTL (e.g. DSH_UPLOAD_TTL=30m) is honored as a
  fallback before the default. In dynamic (session) mode the setting lives in
  memory, overlaid over the optional dsh-web-file-uploader settings namespace.

Files whose timestamp is unknown (legacy index entries) are never auto-removed,
protecting pre-existing uploads from being deleted on migration.

Attach-button placement & replacing the official button

The composer tool row reads [+ menu] [official attach] [permission chips].
This plugin's button always stays visible and is placed directly to the right
of the official attach button, ahead of the permission chips. No node React
owns is moved: the shell's slot wrapper is display: contents, so flex
order re-sequences the row — ours at 1, everything the shell draws after its
own attach button at 2.

Replace the official attach button	Result
Off (default)	Two buttons side by side: the official upright paperclip, then this plugin's diagonal one
On	The official button is hidden and this plugin's button stands in its place — a single upload control in the row

• Saving applies at once, with no refresh, and to every composer mounted later.
• Switching it off and saving — or disabling/unloading the plugin — puts the
  official button back exactly as the shell drew it: hiding writes one inline
  style and never detaches a node, so React's tree stays intact.
• While the official button is hidden, its own pending-attachment rail can no
  longer be triggered; uploads, progress and cleanup all belong to this plugin.
• This plugin's button is never disabled by a running turn — a file picked while
  the machine is busy is injected at the next step.
• If a future shell stops matching, the row is left exactly as drawn, the
  official button stays visible, and one console warning names the reason.

Two card actions (recap)

Action	Host call	File on disk
× close	remove	Kept — model may re-read it
🗑 delete	delete	Removed permanently

Closing a card whose upload is still in flight asks first, because the bytes
are still moving and both answers are real: Cancel upload aborts the transfer
and drops the staged bytes (nothing lands on disk), while Keep uploading
dismisses the question and leaves the card — and, once it finishes, its
reference — intact. Neither answer leaves an orphan copy behind, and the
question only appears while the transfer runs: a finished card still closes in
one click.

The delete host call drops this session's reference, checks that no other
live session references the same deduplicated copy, and only then unlinks the
file and scrubs the index. If another session still references it, the file is
kept (removed: false) but this session's card is still cleared.

Attachment size limits

• No per-file cap. The static bundle writes the request body through as it
  arrives while hashing the same pass; the dynamic plugin has the browser slice
  the File and append one chunk per RPC call. Neither path holds the file in
  memory (a 120 MiB upload grew the process by < 1 MiB), so the only bound left
  is disk space.
• The card shows a percentage while uploading: XHR upload.onprogress in the
  static bundle, sent-chunk accounting in the dynamic plugin.
• The fallback path still has a ceiling. A caller that hands over base64
  instead of a File (the old call shape) stays under the single-shot 64 MiB
  base64 / ≈48 MiB payload cap; that path is kept for compatibility only.
• Images additionally follow the deployment's attachments limits (per-message
  byte/pixel caps) when injected natively for multimodal models.

Installation

A. Dynamic plugin (current session, no install)

cordis_define + cordis_run   # host = src/host.js, client = src/client.js

Approve the run card and the paperclip button appears immediately. The plugin
is process-local: after a restart, define and run it again.

B. Static bundle (persistent, dsh plugin)

The package declares dsh.bundle.patch (see cordis.patch.yml), so
dsh plugin --profile web add recognizes it as a profile layer. Any
pnpm-supported source works:

# Git repository (recommended distribution channel)
dsh plugin --profile web add github:Mooling0602/dsh-web-file-uploader

# Local directory (development)
dsh plugin --profile web add ../dsh-web-file-uploader

# Tarball
dsh plugin --profile web add ./dsh-web-file-uploader-0.2.0.tgz

# npm registry (after publishing)
dsh plugin --profile web add dsh-web-file-uploader

Git spec note: pnpm's git shorthand is github:<owner>/<repo> (e.g.
github:Mooling0602/dsh-web-file-uploader). A bare github.com/<owner>/<repo>
is treated by pnpm as a local directory and will fail with a
"non-existent directory" warning. Other valid forms:
git+https://github.com/Mooling0602/dsh-web-file-uploader.git or
https://github.com/Mooling0602/dsh-web-file-uploader.git.

Restart the dsh web process and refresh the page. See
PUBLISHING.md for distribution details and the optional npm
publish flow (requires your npm credentials).

Update

dsh plugin forwards to pnpm in the profile directory, so update through it
(never edit ~/.dsh/profiles/web/node_modules by hand — the next pnpm
operation rewrites it):

dsh plugin --profile web update dsh-web-file-uploader
# or, if the lockfile-pinned resolution refuses to move:
dsh plugin --profile web remove dsh-web-file-uploader
dsh plugin --profile web add github:Mooling0602/dsh-web-file-uploader

Restart the dsh web process afterwards; the served bundle URL carries a
content-hash revision (?rev=…) so the browser picks up the new build on
refresh. For local-directory installs, run pnpm build in the checkout
before re-adding it. Details: PUBLISHING.md.

Architecture

Browser (Client)                          DSH host (Host)
─────────────                             ─────────────────
conversation.input.left                   harness.handle('upload-begin'/'upload-chunk'/
  └ paperclip ── File (sliced) ┐              'upload-finish'/'upload-abort')  [dynamic]
                              │           webServer route POST /upload        [static]
                              ▼           ┌ sandboxPolicy.resolve() → workspace root
              File slices / body          ├ session cwd / DSH_HOME + /uploads/
                              │           ├ stage .dfu-tmp-*, incremental SHA-256
                              ▼           ├ commit: dedup → name → move → .dfu-index.json
                                          └ base64 -d >> tmp (dynamic) / node:fs stream (static)
conversation.input.dock
  └ attachment cards (persist)            harness.handle('remove', …) / remove route
       └ × closes card → stop injecting     └ pending entry deleted (file kept)
       └ 🗑 deletes card                  harness.handle('delete', …) / delete route
                                              └ pending dropped → unlink + index scrub
                                           lazy TTL sweep (upload / delete / start)
                                              └ remove files past TTL, not in pending
                                          agent/pre-step waterfall
                                             ├ resolveModelInfo → multimodal?
                                             ├ attachments.saveImage → ImageBlock
                                             └ path text block (user messages only)

Why does the dynamic mode go through the shell's base64 -d? The dynamic
sandbox disables require, and the fs service only writes whole UTF-8 text —
so binary has no other way to disk than the shell service's stdin. Appending
per chunk (>>) keeps each command bounded by the chunk size instead of the
file size. The static bundle has no such restriction and streams with
node:fs instead.

Storage location

Mode	Destination
Dynamic plugin	<session workspace>/uploads/ (sandboxed shell/fs cannot leave the workspace)
Static bundle	$DSH_HOME/uploads (default ~/.dsh/uploads) via node:fs — the dsh data directory

The dedup index (uploads/.dfu-index.json) lives next to the stored files and
maps hash → { path, at } (see Deduplication and
Cleanup & retention).

Repository layout

The project follows a single-source-of-truth core + thin seam architecture:
all business logic lives in src/core/*; the dynamic plugin and the static
bundle are thin adapters over it, so changes are made once and both sides pick
them up.

dsh-web-file-uploader/
├── src/core/
│   ├── host-core.js        # canonical host logic (transport-agnostic, DI)
│   └── client-core.js      # canonical client logic (transport-agnostic, DI)
├── src/seams/
│   ├── host-dynamic.template.js     # dynamic host seam (harness + shell/fs)
│   ├── client-dynamic.template.js   # dynamic client seam (host.call + React)
│   └── client-static.template.js    # static client seam (fetch + module react)
├── src/host.js             # GENERATED dynamic host (core inlined) — do not edit
├── src/client.js           # GENERATED dynamic client (core inlined) — do not edit
├── lib/index.js            # static host seam (imports core; node:fs/crypto/webServer)
├── client/src/client.js    # GENERATED static client source — do not edit
├── scripts/
│   ├── build-dynamic.mjs   # inlines cores into seams -> src/*.js + client/src/client.js
│   └── build-client.mjs    # wraps client/src/client.js -> lib/client.js
├── cordis.patch.yml        # dsh.bundle patch (profile layer row)
├── package.json            # publishable manifest (dsh.bundle + dsh.client)
├── PUBLISHING.md           # install & npm publish guide
├── README.md / README_zh_CN.md
└── LICENSE                 # MIT

How to change code: edit src/core/* (or a seam), then run
pnpm build — it regenerates the dynamic sources (src/host.js,
src/client.js) and the static client bundle (lib/client.js). For the
running dynamic plugin, redeploy the regenerated src/host.js /
src/client.js via cordis_define + cordis_run.

Development status

• ✅ Dynamic plugin: implemented and verified in live sessions
• ✅ Card-driven injection, model-aware adaptation, dedup, i18n UI
• ⚠️ Static client module: built by scripts/build-client.mjs, but the
  __ModuleLoader__ wrapper must be verified against the real web toolchain
  before distribution

License

MIT