agentic-learn

A self-driving archive + study system for University of Waterloo Learn (Desire2Learn / Brightspace). It reuses your logged-in browser session to pull everything from your courses via the raw D2L JSON API, organizes it into a markdown-twinned Semester / Course / Week vault, and helps you do assignments grounded only in your class material, with citations — all driveable from Claude Code.

Built and verified on term 1265 (Spring 2026) = "2B": 5 courses, 129/129 content topics, all lectures/labs/tutorials/slides/PDFs, assignments, announcement attachments, discussions, grades.

Quick start

python3 -m venv .learn/.venv
.learn/.venv/bin/python -m pip install -r .learn/requirements.txt
# browser engine (one tool): Vercel agent-browser
npm install -g agent-browser && agent-browser install

PY=.learn/.venv/bin/python
$PY .learn/learn.py auth        # log in to Learn once in the agent-browser window (WatIAM + Duo)
$PY .learn/learn.py calibrate   # probe the live API -> .learn/calibration.json
$PY .learn/learn.py sync        # incremental: ingest -> outlines -> convert -> index -> audit
$PY .learn/learn.py ground MSE232 Lab 4   # grounded, cited assignment pack
$PY .learn/learn.py submit MSE232 "Lab 4" solution.pdf       # DRY RUN preview of a Dropbox submission
$PY .learn/learn.py submit MSE232 "Lab 4" solution.pdf --confirm   # actually submit (gated)
$PY .learn/learn.py verify      # health check -> .learn/AUDIT.md

How it works

• Browser (one tool): agent-browser (Vercel, CDP) wrapped by .learn/agent_browser.py harvests the logged-in session (HttpOnly d2lSessionVal + XSRF, validated with an in-page whoami) and renders SPA pages. No second login; it reuses your authenticated session.
• Retrieval (the bulk): the raw D2L JSON API (/d2l/api/le|lp/...) via requests with those cookies — content, assignments, announcements, discussions, quizzes (metadata), grades.
• Vault: each file lands under <SEMESTER>/<COURSE>/content/<week>/… with a .md twin, plus INDEX.md, _meta/*.json, and a content_map.json. Term- and structure-agnostic (no hardcoding).
• Grounded assignments: ground.py assembles the exact class sources + a strict cite-everything policy; drafts are for your review.
• Submitting: submit.py resolves the right Dropbox folder, validates your file, and submits via the D2L API with read-back verification — dry-run by default, real submit only on --confirm, nothing submitted unless you explicitly say so. (The actual submit can't be tested in advance, so the first real submission is done supervised; success is reported only after the API read-back confirms it.)

Docs

• RUNBOOK.md — full operator guide (sync, term selection, remote/phone trigger, troubleshooting).
• docs/SELF_HOSTING.md — design + execution plan for running it fully self-hosted/headless (one always-on box, phone-only login) — for a future build session.
• CLAUDE.md / AGENTS.md — agent instructions; .claude/skills/ — the skills (learn-sync, learn-assignment, learn-study, agent-browser).

This is for archiving and studying your own coursework; it respects UW's academic-integrity policy.