dsh-nce

English · 简体中文

A New Concept English Book 1 study panel that lives in the DeepSeek Harness sidebar: 144 lessons on one scrollable page, Anki-style SRS flashcards, a mistake notebook, an IPA / phonics reference, and fillable exercise papers graded by your own model. The host half serves /dsh-nce/api; the browser half renders the panel.

The textbook data is not bundled — this repository is code only. The lesson dataset is compiled from copyrighted material and is excluded from git; see Textbook data is not bundled.

Features

Course browser

• Split view: a persistent 144-lesson list on the left and the selected lesson on the right, with a progress line (已学 x / 144), search, four lesson groups, a status dropdown (not started / learning / done) and current-lesson highlight. Three step dots per lesson (words / grammar / exercise) show where you stopped; the panel opens on your last lesson or the next unfinished one and collapses to one column when narrow.

One-page lesson

• Header (prev / next, number, English + Chinese title, counts, textbook pages, status) plus a sticky anchor bar that smooth-scrolls to Video / Text / Words / Grammar / Practice. Video is collapsed by default; expanding loads an embedded Bilibili player mapped to the right part, and a custom Bilibili / YouTube URL can be saved per lesson with the built-in one restorable at any time.
• Text + words — two columns: question, dialogue lines (click a line or the speaker to read it aloud) and collapsible reference translation on the left; a vocabulary table with IPA, part of speech, meaning, read-aloud, and star-to-flashcards on the right. Stacks automatically on narrow containers.
• Grammar — textbook notes plus classroom-notes blocks; Practice — a per-lesson homework checklist, a classroom recall card, and the fillable worksheet. The three study steps are toggled from the status dropdown (choosing "done" checks all three) or each section's "mark complete" switch.

Fillable exercise papers

• Rendered from the textbook layout: blanks become inline inputs, matching questions become real Column A / B tables graded from the answer key, tabular questions become tables, word boxes become bordered cards, and printed worked examples stay static. Enter / Shift+Enter move between blanks; drafts autosave to localStorage per lesson and survive lesson switches and a DSH restart.
• One-click grading = AI teacher grades the whole paper. Every answered question is sent to the profile's default model with the answer-key reference; the reply is per-question correct / close / wrong plus a comment and reference answer, the score is AI-correct / total, and every verdict can be overridden by hand. Without a default model (or on failure) it degrades to deterministic system grading with a header note.
• Local PDF — "Preview PDF" shows the real A4 pages and "Download PDF" generates an image-based PDF entirely in the browser (no network, no server, no print dialog).

Mistake notebook

• Questions graded wrong or close are collected automatically and fixed ones drop out; grouped by lesson with your answer, reference answer, AI comment, verdict, and timestamp, filtered by all / wrong / close. Locally saved graded papers are scanned once so earlier mistakes are backfilled.
• AI explanation per question explains the correct answer, the mistake (word order / tense / agreement…) and a memory hint, cached on the question.

Flashcards and spaced repetition

• Cards come from the de-duplicated 866-word word bank; search and filter by lesson. Space flips, 1–4 grade — Anki-aligned Again (1 min) / Hard (10 min) / Good (tomorrow) / Easy (4 days) — with new / learning / due counters and daily limit.
• New cards follow the course: only words from started lessons enter the queue (default 10/day); due cards are never locked, so reviews always come first. Sentence cards are imported automatically from the translation plugin's history (~/.dsh/dsh-translate/history.json, default 3/day).
• Page, selected lesson, scroll position, and expanded video are remembered per view.

IPA, phonics and read-aloud

• 48 IPA symbols (12 monophthongs / 8 diphthongs / 28 consonants) with pronunciation tips, common letter combinations, and clickable example words; 38 phonics rules, 22 minimal pairs, and syllable / stress rules.
• Read-aloud reuses the translation plugin's speech pipeline when installed (Microsoft Edge neural voices, no API key, shared cache), then macOS say, then browser speech; the voice menu sets voice, speed, auto-read, and preview.

Install

From GitHub (recommended)

dsh plugin --profile <profile> add github:ChenYichener/dsh-nce

From a local clone

git clone https://github.com/ChenYichener/dsh-nce.git
dsh plugin --profile <profile> add /path/to/dsh-nce

DSH desktop app

Open the Plugins page, choose "install from local directory", and pick the repo folder.

Restart dsh web (or the desktop app) after installing so the host half is loaded; after that, client-only changes need a page refresh, while host-half or dataset changes need a DSH restart because both are cached in the host process.

Requirements

Requirement	Notes
DeepSeek Harness	Plugin manifest (dsh.bundle patch + dsh.client); developed and tested with DSH 0.2.2-rc.1 and its dsh web / desktop web surface
Node.js	≥ 20 on the machine running DSH; the host uses node:sqlite and falls back to JSON when unavailable
macOS (optional)	/usr/bin/say is the offline read-aloud fallback; other platforms use browser speech
dsh-translate (optional)	Adds the Edge neural voices and imports translation history as sentence cards
Textbook dataset	Not bundled — build it yourself, see below

Textbook data is not bundled

New Concept English is copyrighted. This repository contains only plugin code; the lesson data in the author's local install is compiled from their own copies of the books and is excluded from git (.gitignore has data/). A fresh clone installs and mounts, but lesson, word-bank and phonetics views show a load error with a Retry button until you supply data. To study for real, build a dataset from material you own:

1. Extract the content into structured JSON that follows docs/DATA-CONTRACT.md (nce1/, notes/, gwb/).
2. Run NCE_CONTENT_DIR=/path/to/nce-content npm run build:data — it writes the gitignored data/ directory.
3. Run npm run build:phonetics for the IPA / phonics reference (hand-written in the script, so it needs no textbook data).
4. Restart DSH to load the dataset.

The compiled data is for personal use only — do not publish it. See docs/DATA-QUALITY.md for how it is extracted and verified.

Usage

Open NCE in the DSH sidebar. Four destinations: 课程 (Course) / 记忆卡 (Flashcards) / 错题本 (Mistakes) / 音标&自然拼读 (IPA & phonics), plus a 声音 (Voice) menu in the top right.

Action	How
Open a lesson	Pick it in the left list; the right pane switches instantly
Follow the lesson	Sticky anchor bar jumps between video / text / words / grammar / practice
Expand the video	Click the video section header or the "Video" anchor (collapsed by default)
Read aloud / change voice	Click a dialogue line or press Enter / Space; voice and speed live in the voice menu
Mark progress	Lesson status in the list, or a section's "mark complete" switch
Flashcards / paper	Space flips a card and 1–4 grade it; Enter / Shift+Enter move through paper blanks

Screenshots

Course list + one-page lesson	Fillable paper with grading	Flashcards

SRS review (four grades)	IPA & phonics	Local A4 PDF preview

How it works

• Host half. cordis.patch.yml inserts one plugin row; index.js is a thin composition root that wires src/host/ (clean architecture: domain rules → application use cases with injected ports → infrastructure SQLite/JSON/data/ speech/LLM adapters → http transport) and mounts POST /dsh-nce/api/<op> on the host web server.
• Browser half. client.js is generated from src/client/NN-*.js fragments by scripts/build-client.mjs; fragments are concatenated in ascending two-digit layer order and may only reference earlier layers. React and the host theme come from the DSH module table.
• API and storage. JSON in / JSON out ({ ok: true, ... } or { ok: false, error }) behind a loopback Host fence. Lesson status, SRS cards, streak, and saved video URLs live in SQLite (node:sqlite, WAL) with a JSON fallback; paper drafts, homework, mistakes, and UI preferences live in browser localStorage. AI grading uses the profile's default model through the host LLM gateway; no API key is stored.

Deeper contracts: PLUGIN-ARCH · API-CONTRACT · PERSISTENCE · LEARNING-FLOW · EXERCISE-PAPER · DATA-CONTRACT · DATA-QUALITY.

Data & privacy

What	Where
Learning progress (status, SRS, streak, video URLs)	~/.dsh/dsh-nce/nce.sqlite (+ -wal / -shm), legacy ~/.dsh/dsh-nce/progress.json
Paper drafts / homework ticks / mistakes / UI state	localStorage → dsh-nce/paper/v1/lesson-<n>, dsh-nce/homework/v1/lesson-<n>, dsh-nce/mistakes/v1, `dsh-nce/nav
Read-aloud cache / translation history (read-only)	~/.dsh/dsh-translate/tts-cache/ (or ~/.dsh/dsh-nce/tts-cache/), ~/.dsh/dsh-translate/history.json
Course content	The plugin's data/ directory, read-only; not published in this repo

Nothing is uploaded by default: progress, drafts, and the mistake notebook stay on the local machine. Data leaves it only for features that need the network — the Bilibili video embed, Microsoft Edge voices, and AI grading (sent to the model configured in your DSH profile). The repository ships no lesson dataset and no personal data.

Configuration

There is no config file: daily new-card limits, the translation-card limit, voice, speed, and auto-read live in the panel UI and are stored in localStorage; host routes, state directory, and body limits are fixed in src/host/config.js. NCE_CONTENT_DIR is a build-time variable for npm run build:data only — the running plugin never reads it.

Development

npm run build              # regenerate client.js from src/client fragments
npm run build:check        # fail if client.js is stale (writes nothing)
npm run check:architecture # layering / fragment order / dead code / doc paths guard
npm test                   # full suite, rebuilds client.js first

Command	What it covers
npm run test:data	data hygiene, IPA validity, OCR-junk guards
npm run test:paper, test:paper-ui	exercise-paper parser/grading engine against the real 144-lesson data, plus paper UI states
npm run test:ai-grade, test:explain	AI-grading / explanation input sanitising, prompts, model-output parsing
npm run test:host-domain, test:lesson-page	SRS, progress and unlock rules, one-page lesson layout
npm run test:mistakes, npm run smoke:api	mistake notebook store/page; every /dsh-nce/api op with an isolated HOME

npm test runs all of the above (12 steps) and boots the host API against a temporary HOME, so your real learning data is never touched. Build the dataset first (npm run build:data) — data-dependent steps fail without it, which is expected for a fresh clone.

Project layout

dsh-nce/
├── index.js                 # host composition root; mounts /dsh-nce/api
├── src/host/                # domain / application / infrastructure / http layers
├── client.js                # browser bundle (generated by scripts/build-client.mjs — do not edit)
├── src/client/              # layered fragments, concatenated in NN order
│   └── generated/           # generated fragments (Bilibili part map)
├── scripts/                 # data/client builds, architecture guard, data checks, tests
├── locale/{en,zh}.json      # plugin manager title and description
├── docs/                    # architecture, API, data, persistence and paper contracts
└── data/                    # local dataset build output — gitignored, never published

FAQ / Troubleshooting

Symptom	Fix
Course list shows a load error	The data/ dataset is missing; build it (see above) and restart DSH
Code changes do nothing	Host half (index.js, src/host/) is cached — restart dsh web; client changes need npm run build + page refresh
No sound	Check the voice menu; non-macOS needs browser speech, and online Edge voices need the translation plugin plus network
AI grading says unavailable	The profile has no default model (or the call failed); the paper falls back to system grading and verdicts can be overridden manually
Video area is blank	The embed needs network access to Bilibili; save a custom URL if the built-in one is unavailable
Where is my progress / how do I reset it?	~/.dsh/dsh-nce/nce.sqlite; backup and reset commands are in PERSISTENCE

Compatibility

• macOS, Windows, and Linux for the panel itself (a browser surface inside DSH); the say read-aloud fallback is macOS-only, other platforms fall back to browser speech.
• Built against the DSH plugin manifest used by dsh 0.2.x; a change to that manifest or to the client module table could require a plugin update. The textbook dataset is bring-your-own and platform-independent once built.

Related DSH plugins

Plugin	What it does
dsh-book	Markdown draft library + AI copywriting workspace for Xiaohongshu / WeChat / generic posts
dsh-notify	Native desktop notification when a session finishes while DSH is in the background
dsh-pomodoro	Distraction-free Pomodoro timer in the sidebar
dsh-translate	Chinese/English translation workbench with an always-on AI tutor and neural read-aloud
dsh-nce	New Concept English Book 1 study panel (lessons, SRS flashcards, mistake notebook)

License

MIT © 2026 chenyichen. The plugin code is MIT-licensed and does not include any textbook content: the lesson dataset is compiled locally from material the user owns. New Concept English and the related books, exercises, and audio are the property of their respective authors and publishers; the embedded video player streams third-party videos from Bilibili. Use the plugin only with content you are entitled to use.