Paper Desk · 论文工作台

A Zotero-shaped paper desk for DeepSeek Harness. It adds a Papers entry to the DSH sidebar, a three-pane PDF reader and a per-project mind map: the normal DSH sidebar on the left, the document or canvas in the middle, and a DeepSeek chat / Markdown notebook on the right.

本项目也适合中文用户：界面、AI 精读笔记和共读对话默认使用中文。

Features

• Projects with a purpose — choose Research (find a problem, build a method, compare benchmarks) or Learning (build concept dependencies, timeline, reading queue and recall exercises). Each project keeps papers, tags, annotations, chats, notes and a mind map together.
• Project mind map — one persistent React Flow canvas per project, with drag-and-drop paper nodes, Markdown/KaTeX nodes, edges, backlinks, automatic tidying and a DeepSeek map assistant. Research and learning projects start from different templates.
• Drop PDFs in — uploads, parses, and files PDFs into a Zotero-style folder (Author et al. - Year - Title). Title, authors and year are read from the PDF metadata and first page.
• Double-click to read — click the Read button, or double-click any non-interactive part of a paper row.
• Zotero-shaped paper rows — expand a paper to see paper.pdf, notes.md and any appendix or supplementary files. Extra files are stored verbatim and never parsed.
• PDF reading — PDF.js viewer with fit-width/zoom, page navigation, selectable text, four highlight colours, paragraph-anchored notes, one-click selection translation, and “ask AI” with the selected passage.
• Math-aware Markdown — $...$, $$...$$, \(...\) and \[...\] render in notes and chat using KaTeX's HTML layout with embedded KaTeX fonts. Code spans and currency are left alone.
• AI deep-read note — on first open DeepSeek writes a structured Markdown note: TL;DR, core problem, prior work pros/cons, how this paper fixes the cons, motivation + key idea + example, how to read the experiments, strengths, limitations and future work. The note is a normal Markdown file you edit in the right pane, with preview, autosave and “AI rewrite”.
• Co-reading chat — a dedicated DeepSeek conversation per paper. It receives the paper text, generated note, highlights and the quoted passage, so follow-up questions have context. The conversation is persisted.
• Tags and search — Zotero-style tag chips; filter by one or more tags and search titles, authors, tags and note excerpts.
• Your data is plain files — no database, no lock-in.

paper-desk/
  projects/
    scaling-inference-for-sbi/
      project.json
      mindmap.json          # research or learning map
      mindmap-chat.json     # project-level AI conversation
      papers/
        kuhmichel-et-al-2026-high-dimensional-simulation-based-inference/
          paper.pdf
          paper.txt          # extracted text (AI context)
          notes.md           # Markdown notebook
          annotations.json   # highlights + paragraph notes
          chat.json          # co-reading conversation
          attachments/       # appendix, supplementary material, ...

Requirements

• DeepSeek Harness 0.2.x with the Web GUI and a DeepSeek provider.
• A modern Chromium-based Web renderer (the DSH Desktop app is supported).
• Node.js 22+ for development builds.

Paper Desk uses the active DSH profile's provider/model. It does not ask for a separate API key. When provider/model are not configured explicitly it selects the first DeepSeek Flash/Chat route available in the profile.

Install from GitHub

In DSH, open Settings → Plugins, install the bundle from GitHub (fexas/dsh-paper-desk), and enable it. Restart DSH if the host asks for a restart after installing a new package.

If your DSH version exposes the plugin command directly, the equivalent local checkout flow is:

git clone https://github.com/fexas/dsh-paper-desk.git
cd dsh-paper-desk
npm ci
npm run build

Then add this directory as a file: dependency of the active DSH profile and select dsh-paper-desk in that profile's bundle list (see the DSH plugin documentation). The host half has no build step; client.js is the pre-built browser half.

Data location

The default data directory depends on how the plugin is installed:

Install	Default projects/ location
Linked local checkout	<plugin-root>/projects/
Managed plugin (under node_modules)	$DSH_HOME/paper-desk/projects/ (usually ~/.dsh/paper-desk/projects/)

Override it in the plugin row:

- id: dsh-paper-desk
  name: dsh-paper-desk
  config:
    provider: deepseek-account
    model: deepseek-flash
    # dataDir: /absolute/path/to/paper-desk-projects

The data directory contains project.json plus one folder per paper. Deleting a project or paper moves it to projects/.trash/ rather than destroying it. PDFs, text, notes, annotations and chat history are local files; nothing is sent anywhere except the model requests made by DeepSeek.

How it works

browser (client.js)                         host (host/service.js)
──────────────────                         ───────────────────────
React slot UI                              Typert Remote `paperDesk`
PDF.js viewer + KaTeX                      projects/, uploads, files
highlight / note state          <──RPC──>  annotations.json, notes.md
DeepSeek chat + note UI                    chat.json, paper.txt
                                           DeepSeek call via ctx.llm

• client.js is a single browser module built from client/ with esbuild. PDF.js, KaTeX fonts and React Flow CSS are bundled, so no CDN or extra asset route is needed.
• The host exposes a small paperDesk Remote namespace. The browser never gets direct filesystem access.
• Every paper is a directory with human-readable JSON/Markdown files; the UI can be removed without losing the research library.

Project mind map

Every project has a single mindmap.json. Open it from the 项目脑图 banner or the header button in a project.

Templates

• Research starts with:
  • Motivating example / problem to solve
  • Our methods: framework, motivation / key idea, and why it is superior
  • How to evaluate performance: datasets/benchmarks, metrics, baselines/ablations
• Learning starts with:
  • Foundations and prerequisites
  • Core concepts, examples and misconceptions
  • Development timeline
  • Paper queue
  • Recall and exercises

Templates are just seeded nodes. Every title, Markdown body and edge is editable, and 整理布局 runs a dagre layout for the current graph.

Linking papers and ideas

Open 论文库, then click or drag a paper onto the canvas. A paper node keeps a stable paperId; any node can also reference papers from the node inspector. The inspector shows incoming backlinks for the selected node, and clicking a paper chip opens its reader. This gives a lightweight Zettelkasten-style graph: papers, concepts, questions and experiments live in one canvas and connect in both directions.

Map chat

The right pane switches between 对话 and 节点:

• 对话 sends the whole graph, the paper library metadata and (optionally) the selected node to DeepSeek. The model is explicitly told this is a research-strategy or learning-coach session, depending on the project type.
• 节点 edits the selected node's title, type, Markdown body (with formulas), linked papers and backlinks.

Performance notes

The reader shell opens as soon as the paper record arrives. PDF bytes are read in eight parallel range requests, page 1 renders before the remaining page boxes and text layer are resolved, and recently opened PDFs stay in a small in-memory LRU cache. PDF.js runs its worker bundle on the page thread, which avoids a Chromium/Electron worker handshake failure that can otherwise hang the viewer. AI note generation runs alongside PDF loading; a compact progress card shows the live stage and disappears when the note is written.

Development

npm ci
npm run build      # rebuild client.js
npm run check      # syntax-check host and built client

Source layout:

• host/service.js — projects, chunked uploads, annotations, notes, LLM calls
• client/App.jsx — project library, paper rows, reader shell, chat and notes
• client/PdfViewer.jsx — PDF.js canvas/text/highlight viewer
• client/MindMapPage.jsx — React Flow canvas, paper tray, node inspector and map chat
• client/markdown.js — Markdown + KaTeX formula rendering
• client/styles.css — DSH-token-based UI styles
• host/mindmap.js — research/learning templates and graph normalisation
• build.mjs — esbuild bundle, PDF.js worker embedding, inline KaTeX fonts and React Flow CSS

Pull requests are welcome. Please keep the UI on DSH theme tokens (--dsw-alias-*) and do not add a separate app shell, CDN dependency or a second filesystem root.

Known limitations

• No OCR: image-only PDFs will not have selectable text or AI full-text context.
• The AI note is generated from extracted PDF text; complex multi-column or scanned layouts may lose structure.
• The co-reading chat is a focused paper assistant, not a general-purpose DSH agent with tool access.
• Data is local to one machine; there is no sync or multi-user layer.

License

MIT. See LICENSE and THIRD_PARTY_NOTICES.md. Bundled third-party code includes PDF.js (Apache-2.0), KaTeX (MIT), React Flow, dagre, marked, DOMPurify and lucide.