BookWeave is a DeepSeek Harness (DSH) Web client plugin. It turns the sidebar into an EPUB reading workspace: a shelf, a distraction-free reader, excerpts, a network of every book mentioned in the text, and an AI-assisted book guide that summarises a chapter and turns its claims and evidence into clickable cards backed by the original paragraphs.

Everything is local-first: EPUBs, excerpts, reading progress and AI results live in your browser's IndexedDB. The only network call is the optional "Analyze chapter" request to your configured model provider.

Features

Shelf	Import .epub by click or drag & drop; 2:3 portrait covers extracted from the EPUB; book title, author and reading percentage only.
Reader	Serif prose, adjustable font size, collapsible TOC, chapter navigation, remembered progress. No fixed header — the reading area and the right rail both start at the top of the window.
Book list	Extracts every book the current book mentions: 《》 titles, English “quoted title” cues and a curated lexicon, with noise filters (self-references, TOC, chapter headings, periodicals). Each entry links back to the paragraph where it appears.
Excerpts	Select text → red underline + excerpt card with an optional note; click an excerpt to jump back to the exact paragraph.
Book guide	A right-rail tab, left of Excerpts: it opens on the chapter you are currently reading and follows TOC / prev-next / locate jumps. Chapter picker, AI-generated chapter summary, and cards ordered by where they appear in the text. Claims render as an argument chain: the conclusion at the top, an up-arrow labelled “because”, then the reasons/examples/data beneath, each tagged with its relation. Every node shows an “original” excerpt and highlights/flashes the source paragraph on hover/click.
AI analysis	One explicit button per chapter. Fiction → plot summary in the book's own voice + plot beats; nonfiction → claims + evidence. Input is compressed to ≤6–8k characters, output ≤700 tokens, lowest reasoning effort, cached per book/chapter, never re-billed unless you press “Analyze again”.
Book search	From a book-list entry, open an in-plugin browser panel pre-filled with a configurable search URL (default: Open Library). Edit the template under Book list → Search source; use {query} for the title, and any search provider works.
Export	Export the whole book list as Markdown.
i18n	Chinese and English UI.

Install

Requirements — a DeepSeek Harness build with the Web UI. The AI half additionally needs the Host llm service (any provider that DSH can already call); without it, reading, excerpts, book lists and the heuristic guide still work.

From GitHub (available now)

dsh plugin add github:adrianwang99/dsh-bookweave

From npm (after the first npm release)

dsh plugin add dsh-bookweave

From source (for development)

git clone https://github.com/adrianwang99/dsh-bookweave.git
cd dsh-bookweave
dsh plugin add .

Or, if you manage the profile manually, symlink the checkout into the profile and add a link: dependency:

ln -s "$(pwd)" ~/.dsh/profiles/desktop/node_modules/dsh-bookweave
# then add "dsh-bookweave": "link:<absolute path>" to the profile package.json

Restart DSH (⌘Q on macOS) after installing or updating the Host half (index.js); client-only changes are picked up by a page refresh.

The package declares the standard manifest:

{
  "dsh": {
    "bundle": { "patch": "./cordis.patch.yml" },
    "client": { "platform": "web", "immediately": true }
  }
}

Usage

1. Open 书脉 from the DSH sidebar.
2. Import an EPUB (click the shelf or drag the file onto the window).
3. Read: use the floating pills (← shelf, TOC, rail toggle, A−/A+, export) — they stay available while scrolling but take no layout space.
4. Open the right rail → Guide (default). Pick a chapter, optionally press Analyze chapter.
5. Hover a guide card to highlight the source paragraph in the text; click it to scroll and flash that paragraph.
6. Select text in the reader to create an excerpt; open Excerpts to review and jump.
7. Book list → expand an entry to locate the mention or search the book in the built-in browser. The search URL is configurable (Book list → Search source, default Open Library).

AI analysis & privacy

• Opt-in only. Opening a book, switching chapters or tabs never calls a model.
• Small prompts. One chapter's compressed excerpt (≤6,000 chars client-side, 8,000 host-side), concise JSON output (≤700 tokens), lowest reasoning effort the model advertises, no retries.
• Cached. Results are stored per book + chapter + content fingerprint in IndexedDB. Reopening a chapter shows “Cached” and costs nothing; “Analyze again” is the only way to spend tokens twice.
• No telemetry. The plugin makes exactly one kind of network request: POST /api/dsh-bookweave/analyze to its own DSH host, which forwards the prompt to your configured model provider.
• Model override. Set DSH_BOOKWEAVE_MODEL=provider/model before launching DSH to pin a cheaper model for analysis.

Data storage

IndexedDB database dsh-bookweave (version 3): books, index, excerpts, insights. Deleting a book removes its text, book list, excerpts and AI results. Nothing leaves the machine except the opt-in model request.

Development

git clone https://github.com/adrianwang99/dsh-bookweave.git
cd dsh-bookweave
npm test          # core + archive + structure + host suites (Node only, no dependencies)
npm run check     # syntax check the client and host halves

The suites run in plain Node (≥20) without a browser, React or npm dependencies. They load client/client.js through the same window.__ModuleLoader__ seam the DSH Web client uses, and exercise the plugin.core exports plus a fake host-side LLM for the AI route. See test/README.md.

Optional real-book smoke tests are skipped unless you point at your own EPUBs:

BOOKWEAVE_SMOKE_ZH=/path/to/zh.epub BOOKWEAVE_SMOKE_EN=/path/to/en.epub npm test

Project structure

.
├── client/client.js        # everything UI + core (EPUB parsing, extraction, guide, AI client)
├── index.js                # Host half: /api/dsh-bookweave/analyze (LLM call, prompt, normalisation)
├── cordis.patch.yml        # bundle patch: registers the plugin row
├── locale/{zh,en}.json     # plugin metadata strings
├── icon.svg                # sidebar icon
├── plugin.json             # marketplace metadata (deepseek.stream upload)
├── test/                   # node-only suites + synthetic EPUB fixtures
├── docs/PUBLISHING.md      # npm / awesome-dsh-plugin / WhaleHub / deepseek.stream guide
└── marketplace/            # ready-to-submit community marketplace entries

Known limitations

• EPUB parsing is a tolerant pure-JS implementation; malformed EPUBs fall back to a lighter archive discovery path, but DRM-protected files are not supported.
• Mention extraction is heuristic: it errs on the side of recall and can include the current book or chapter titles; the book-list UI marks them and the export keeps them out of the “found books” count.
• The AI guide is per chapter and capped for cost; it is not a full-book analysis.
• The built-in browser cannot inject text into a cross-origin page; it opens a search URL instead.
• AI results depend on the configured model; a weak model may produce a poor summary. Cached results are never silently replaced.

License

MIT © 2025 adrianwang99