dsh-plugin-lookatstudy

Turn any markdown document, local folder, or GitHub learning repository into a guided course inside DeepSeek Harness (dsh) — your dsh agent becomes a full AI tutor with the interaction design of LookatStudy: per-concept knowledge tracking, mastery-driven progression, spaced repetition, mastery proposals, friction awareness, learner memory, a Cornell notebook, and an in-chat proposal card. Learning engine modules are vendored from LookatStudy (MIT).

Install

dsh plugin add dsh-plugin-lookatstudy        # from npm
# or from a tarball:
dsh plugin add ./dsh-plugin-lookatstudy-0.2.1.tgz

Works with any profile. In the web profile the plugin additionally serves the study tab's HTTP API and loads its browser half; headless profiles get the plain tool surface.

The two surfaces

1. The tutor (chat). Talk to the agent: "import https://github.com/microsoft/AI-For-Beginners and teach me lesson 1", "what reviews are due today?". The tutor persona (stable core + one of three souls — guide 引导 / direct 精讲 / practice 实战) drives the full LookatStudy loop:

• Knowledge components (KC) — on first teaching a lesson the tutor derives 2–7 concepts (study_define_concepts); every graded answer is attributed (study_record_answer with concept); per-concept BKT runs and lesson mastery is the weakest concept — quizzes target ⚡weak ones first.
• Mastery-driven progression — ≥50% unlocks the next lesson early; ≥90% graduates and schedules the first SM-2 review; answers also nudge the review schedule.
• Mastery proposals (propose → apply) — at ≥85% plus a convincing Feynman-style explanation the tutor proposes early graduation and waits for the learner's yes/no; only the explicit decision applies it.
• Friction awareness — confusion/blocks/frustration are silently logged (study_report_friction) and surface as ⚡😣 weak spots.
• Learner memory — three slots (global style / per-course pattern / per-lesson gap), read-merge-write (study_remember).
• A dynamic learner snapshot (focus, strategy band, weak concepts, friction, memory, due count, pending proposal) is injected as runtime context every turn.

2. The study tab (dsh.client). The whole plugin lives in ONE conversation view tab — 「学习」 — a simplified LookatStudy in three columns, styled entirely with dsh's --dsw-* tokens; nothing outside the tab modifies dsh chrome:

Column	What you get
左 · 课程	Course picker, progress, due box with one-click review kickoff, lesson tree (gating, mastery bars, ⚡😣 weak spots, clickable focus), one-click demo import when empty
中 · 老师	A read-only mini transcript of the live tutor conversation (assistant replies rendered through the plugin's markdown pipeline, tool calls as chips, streaming included) plus the soul pills (直讲/引导/实战), the focus lesson's starters, and the mastery-proposal banner (接受/再练练). Typing happens in dsh's own composer below the tab — the plugin never ships its own input; every button lands its text in that native composer and submits through the same path as the Send button
右 · 黑板	The focus lesson's 讲解 (server-sanitized markdown) and the Cornell 笔记 three zones

The tutor column is a learning surface, not a generic chat: the tutor's quiz options (A–D) render as clickable answer buttons under the latest reply (clicking sends the answer through the native composer), graded answers show as ✓/✗ chips with the tested concept, and every course-tree glyph, tag, and mastery bar carries a hover tooltip explaining its meaning. All content text runs at the dsh chat transcript's own 16 px.

All study state comes from one shared 3 s poll over /lookatstudy/api/state; input is dsh's native composer from the first turn (no activation step). Columns stack below 1024 px.

Tool surface (19)

Import: study_import_markdown / study_import_folder (9 doc + 30+ code formats; PDF/PPTX unsupported) / study_import_github (jsDelivr CDN, works where github.com is unreachable) Learn: study_courses, study_map, study_lesson Progress: study_define_concepts, study_record_answer, study_complete_lesson Proposals: study_propose_mastery, study_resolve_proposal Reviews: study_due_reviews, study_record_review Awareness: study_report_friction, study_remember, study_notes, study_note_save Misc: study_set_mode, study_delete_course

Configuration (cordis.yml patch layer)

- id: lookatstudy
  name: dsh-plugin-lookatstudy
  config:
    mode: guide          # direct | guide | practice — initial soul; persists in state afterwards
    statePath: ''        # default: $DSH_HOME/lookatstudy-plugin/state.json

What is intentionally not restored

LookatStudy's Electron-native experiences have no host surface in dsh: persistent text highlighting with DOM anchors, celebration particles, streak/XP gamification (its effect depends on that UI), exam mode, and multimodal lesson images. Everything else — engine, contracts, data models — is ported.

Development

pnpm exec tsdown        # build lib/ (host + client entries, peers external)
pnpm test               # 48 node:test cases over the real source (no key needed)

# iterate against a live dsh (this repo lives beside a deepseek-harness checkout):
pnpm dsh web --patch ../dsh-plugin-lookatstudy/cordis.dev.yml   # run from the harness checkout
# then open http://127.0.0.1:3080/ and switch to the 学习 tab

Layout: persona + snapshot context in src/index.ts, tools in src/tools.ts, state transitions in src/state.ts, the study tab's HTTP API in src/dashboard.ts, sanitized markdown in src/markdown.ts (shared by the host routes and the client bundle), the browser half in src/client/ (index.ts tab registration, views.tsx the three-column tab plus the pure transcriptRows fold, data.ts shared poll store, styles.ts injected --dsw-* stylesheet), UI card projections in src/cards.ts, vendored zero-dependency engine in src/vendor/ (see each file's provenance header; the one local modification to the folder scanner's dedup key is documented there).

Publishing note: exports must keep "./package.json": "./package.json" — the web bundle's client-module scanner resolves it to discover the dsh.client browser half. When re-installing a rebuilt tarball into a profile, remove the old one first or bump the version (pnpm reuses same-spec tarballs).

License

MIT. Engine modules vendored from LookatStudy (MIT).