dsh-chat-files

English · 中文

Session-bound file attachments for DeepSeek Harness. Drag, paste, or select a file into the chat; the agent reads it on its next step through bounded, session-authorized tools. Supports PDF, Office documents, archives, and plain text — with GB18030/GBK decoding, automatic secret redaction, and zip-bomb defense. One-click install on macOS and Windows.

DeepSeek Harness 会话文件附件插件。文件拖入、粘贴或选择进对话后，模型在下一步经会话授权的受限工具读取。支持 PDF、Office 文档、压缩包与纯文本；含 GB18030/GBK 中文编码、敏感信息自动脱敏与压缩炸弹防护。macOS 与 Windows 一键安装。

内部包名保持 @dsh-external/dsh-file-attachments（安装链契约）。

Features

Intake

Drop a file anywhere in the window, paste it, or use the attach button. Raster images continue on DSH's native image path; everything else enters this plugin. A "release to attach" overlay replaces the native images-only blocker.

Auto-announcement

The message wire carries no visible file marker, so a text-only model can silently ignore uploads. Before each step the plugin appends one plugin-source line announcing every unannounced file with its detected type; the model then calls attachment_info → read_attachment / list_archive and combines the file content with the user's words. Injected through the official agent/pre-step waterfall (the serial chain before request derivation) and rendered by the official UI as an inject line.

Typed cards

Composer cards show a distinct icon and colour per detected family (text / document / spreadsheet / presentation / archive / …).

Delivery receipt

A line under the user's bubble reports real events: Agent received → Agent reading… → Agent read. The final stage appears only after the model calls the read tools.

Readers

Text/config/source files, Markdown, HTML/XML/SVG markup (detected and read as text), PDF, DOCX, XLSX (sheet + A1-range reads), PPTX, legacy .doc/.xls, RTF, ODF, EPUB, and ZIP/7z/RAR/EPUB listing with safe single-entry extraction (CJK entry names supported).

Model tools

attachment_info, read_attachment, list_archive — bounded page/range/cursor/paragraph pagination, archives listed before extraction.

Redaction

Credential keys, private-key PEM blocks, and YAML block scalars are redacted line-wise before any byte reaches the model (aws_secret_access_key, apiKey, set-cookie included; public_key, monkey untouched).

Architecture

composer drop / paste / button (client)
   │  full-window takeover, generic files only — images stay native
   ▼
POST /api/dsh-file-attachments/v1/files   (Origin + session-existence verified, fail-closed)
   ▼
content-addressed store  (sha256 blob, atomic rename publish, immutable)
   ├─ detection (magic bytes + ZIP metadata, worker-isolated)
   ├─ refs + batch indexes (durable metadata, path-injection-validated)
   └─ read path: session ownership check → bounded worker parse → redaction
   ▼
model view:
   ├─ agent/pre-step announcement line (plugin source, type-tagged)
   └─ attachment_info / read_attachment / list_archive tools

Install

One-command install

# macOS / Linux (from the repository root)
node scripts/install.mjs
# or double-click: install.command        (Windows: install.bat)

The installer builds host + client bundles, symlinks the package into the shared profile directory ($DSH_HOME/profiles/node_modules, so every profile — web, CLI, headless — resolves it), and appends one row to $DSH_HOME/cordis.patch.yml (a backup is written first). It is idempotent: running it again is safe. Uninstall: node scripts/install.mjs --uninstall or uninstall.bat / uninstall.command.

Prebuilt bundle: download dsh-file-attachments-<version>.zip from Releases, verify it against SHA256SUMS.txt, extract, and double-click the installer. No git, build step, or package manager required.

Windows users: see docs/WINDOWS-INSTALL.zh.md for the full walkthrough (prerequisites, install.ps1, manual steps, macOS comparison table, and FAQ). Requires Node.js ≥ 20, PowerShell 5.1+, and for archives either the built-in tar.exe (Windows 10 1803+, bsdtar) or Git for Windows; dependencies via npm or pnpm.

Per-end deployment

Configuration: the storage root defaults to $DSH_HOME/file-attachments; override with config.root on the plugin row in cordis.patch.yml.

Safety model

• Session ownership — every read verifies ownerSessionId and that the session still exists (sessionQuery.readSession, fail-closed, 10 s TTL cache).
• Origin — uploads require a trusted Origin (exact match or loopback-equivalent 127.0.0.1/localhost/[::1] on the same port); cross-site reads and deletes are rejected while origin-less local clients stay allowed.
• Parsing isolation — every parse runs in a worker thread under maxOldGenerationSizeMb: 512 and a hard timeout that terminates hung parsers (word-extractor/xlsx cannot be aborted otherwise).
• Archive safety — paths are normalized and whitelisted; extraction writes to stdout only (no zip-slip surface); hostile entries are skipped during listing and rejected during extraction; decompressed output is capped at 256 MB.
• Redaction — line-wise credential/private-key redaction before the model (see Features).
• Limits — 25 MB per file (100 MB archives), 10 files / 50 MB per message, 256 KB / 2 000 lines per read, 15 s parser timeout, 10 s archive timeout.
• Announcement degradation — the pre-step listener never throws into the loop; any failure falls through unchanged, and a strengthened systemPrompt section (order: 70) remains as the fallback trigger.

Official API conformance

Every seam this plugin uses is a public DeepSeek Harness contract — no monkey-patching, no core changes, harness upgrades unaffected:

Compatibility

Development

npm install
npm run build        # host + client bundles
npm run typecheck
npm test             # 168 tests (vitest)
node scripts/smoke.mjs http://127.0.0.1:3080 <file>   # live smoke against a running web

Attribution and license

• MIT — see LICENSE.
• Dependency attributions — THIRD_PARTY_NOTICES.md (@keep-lts/xlsx, cfb, fflate, file-type, pdfjs-dist, word-extractor, react, build/test tooling).
• Derived code — the user-message renderer used by the delivery receipt is ported from DeepSeek Harness's own dsh-client-ui-conversation (MIT); the installer's junction-fallback strategy follows @linxin666/dsh-client-ui-skin-center (Apache-2.0). Both are credited in the respective source files and notices.
• Design reference — the content-addressed storage model follows the official @deepseek-ai/dsh-attachment package (see conformance table above).