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.