dsh-novels

dsh-novels

English | 简体中文

Read .txt novels inside DeepSeek Harness: import a file, get chapters split automatically, and read it in a panel next to the conversation.

Status: 0.1.3 — upload, shelf, reader, progress restore, rename, delete, reader prefs (font size, line height, page width), upload progress, long-book table-of-contents grouping, empty/error/loading states, and show or open the folder that holds your uploaded files.

Features:

• Import .txt by button or drag & drop; 20 MB cap, UTF-8/BOM/GB18030/GBK/Big5/UTF-16.
• Automatic chapter split with a 4000-character block fallback.
• Shelf cards show chapter/character counts, encoding, reading progress bar and "读到第 N 章"; rename in place; delete with confirmation.
• Storage row above the shelf shows the absolute data folder and opens it; each card has a Location button that reveals (and selects) that book's .txt in the OS file manager. The browser cannot do this itself, so the Host spawns explorer.exe /select,… (Windows), open -R (macOS) or xdg-open (Linux). The client only posts the book id — never a path.
• Uploaded bodies keep the book's name: books/<title>.txt (+ .chapters.json). Renaming a book renames both files, a second book with the same title becomes <title> (2).txt, and characters Windows forbids (\ / : * ? " < > |) are replaced before the name touches the disk. Older <id>.txt layouts migrate on the next start.
• Reader: table of contents, previous/next chapter, font size, line height, page width, keyboard (← → PageUp PageDown Space j k Esc Ctrl±), chapter counter and a bottom progress line.
• Reading position (chapter + scroll offset) and reader prefs persist on the Host, so they survive a reload and a restart. /api ports change every launch, so the client keeps no state of its own.
• Long books (>300 chapters) get a grouped table of contents (100-chapter blocks, only the active block expanded).

Screenshots

Reader — table of contents, font size / line height / page width controls, chapter counter, keyboard navigation.

Grouped table of contents — books over 300 chapters are split into blocks of 100, with only the current block expanded.

Panel styling

The stylesheet is installed once into document.head when the plugin activates (ensureStyles(), tag style[data-plugin-css="dsh-novels/panel.css"]).

It used to be a <style> element rendered inside the shelf's own tree. Every rule lives in the same string, so the shelf looked right — but opening a book unmounted the shelf, the style element went with it, and the reader lost the whole sheet: no height:100%, no flex column, no overflow:auto. The symptom was a reader that could not scroll and a chapter footer pushed out of the panel. Keeping the sheet in head makes it independent of whichever view is mounted. render.test.mjs asserts both halves of that contract (sheet in head, no <style> inside the panel tree).

Install

# clone once, then install from the DSH profile that should carry the plugin
git clone https://github.com/darlingsakurua/dsh-novels.git
plugin_manager install_bundle target="<path-to-your-clone>"

The bundle is linked (not copied), so edits under client/ and lib/ apply after the plugin is reloaded (plugin_manager set_plugin target="include:dsh-novels" enabled=false then true).

Layout

dsh-novels/
  package.json         dsh.bundle.patch + dsh.client (platform: web)
  cordis.patch.yml     inserts the bundle into the profile
  icon.svg             sidebar glyph (currentColor)
  locale/{zh,en}.json  panel title for the sidebar tooltip
  lib/index.mjs        Host half: HTTP routes under /dsh-novels/api/v1
  lib/store.mjs        Host half: library index, atomic writes, path checks
  client/index.mjs     Web half: shelf + reader, plus the text pipeline

Data lives in <profileDir>/plugin-data/dsh-novels/:

library.json            {version, prefs, books[]}
books/<id>.txt          normalized UTF-8 body
books/<id>.chapters.json  [{i, title, start}]

API

All routes are gated exactly like the shipped /api channel (the request must carry the browser session the shell already forwards).

Method	Path	Purpose
GET	/ping	diagnostics (dataDir, bookCount, version)
GET	/library	books + reader prefs
POST	/books	{name, text, encoding, chapters} → book
GET	/books/{id}	book + chapter offsets
GET	/books/{id}/text	full body as text/plain
PATCH	/books/{id}	rename ({title, author})
PATCH	/books/{id}/progress	{chapterIndex, offsetChars, percent}
DELETE	/books/{id}	remove book and its files
POST	/reveal	open the books/ folder in the OS file manager
POST	/books/{id}/reveal	reveal + select that book's .txt
PATCH	/prefs	{fontSize, lineHeight, measure, fontFamily}

Envelopes are {ok: true, ...} or {ok: false, error, message}; request bodies are capped at 24 MB for uploads and 64 KB for everything else. The client rejects files over 20 MB before reading them; the Host cap leaves room for the JSON envelope (body + chapter offsets).

Text pipeline

client/index.mjs decodes bytes in this order: BOM (UTF-8/UTF-16) → strict UTF-8 → GB18030 → GBK → Big5. Line endings are normalized to \n before chapter offsets are computed, because every offset is a character index into that normalized text.

Chapter headings are matched with ^第…[章节回卷篇部集] and ^Chapter/Part/Book N. A heading with less than 200 characters before the next one is treated as a false positive and dropped (the previous chapter absorbs it); the final chapter is always kept. Files with no heading at all are split into 4000-character blocks.

Tables of contents above 300 chapters are rendered in blocks of 100 (1–100, 101–200, …) with only the block holding the current chapter expanded. Grouping by the first character of the title was tried first and abandoned: nearly every Chinese chapter heading starts with 第, so it produced a single bucket.

Tests

No test runner needed — plain Node scripts, each printing ok/FAIL lines:

node _test/store.test.mjs        # Host store: index, atomic writes, clamping
node _test/client-parse.test.mjs # decode/normalize/split (bundled client code)
node _test/host-api.test.mjs     # routes, envelopes, 401/403 gate, limits
node _test/perf.test.mjs         # 5 MB GB18030 sample end to end
node _test/render.test.mjs       # shelf → upload → read → delete, headless React

render.test.mjs loads the real client bundle with a miniature React (hooks + static tree) and points fetch at a server wired to the plugin's own Host routes, so the browser flow is exercised without a browser. _test/tmp/ holds its throwaway profiles.

_test/make-sample.mjs regenerates the GB18030 samples in samples/ (the 5 MB one is handy for checking upload time by hand).

Manual acceptance

These cannot be automated from here (the sandbox cannot start a browser):

1. Import samples/sample-gb18030-small.txt and samples/sample-gb18030-5mb.txt.
2. Read past chapter 10 of the 5 MB book, scroll mid-chapter, go back to the shelf — the card should read "读到第 N 章" with a filled bar.
3. Reopen it: same chapter, same scroll offset, same font size/line height/width.
4. Rename a book on the shelf; the reader header follows the new title.
5. Toggle the app theme (light/dark): all panel surfaces and text follow the --dsw-alias-* tokens.
6. Disable the plugin (set_plugin target="include:dsh-novels" enabled=false): the 书架 icon disappears and /dsh-novels/api/v1/ping returns 404; re-enable and refresh to get it back with the same books.

Rules this plugin follows

• Never require @deepseek-ai/dsh-client-*; only react and the builtins.
• Never poke document.body or use an iframe; the panel is a React tree in a slot. (The stylesheet is a <style> in head, installed imperatively — like the shell's own style service does — not DOM inside the panel.)
• Styles use --dsw-alias-* tokens only — no literal colors.
• Keyboard shortcuts live on the reader container, never on window.
• The API prefix is /dsh-novels/api/v1; /api belongs to the connection channel.

License

MIT © 2026 darlingsakurua