agent-browser

name	agent-browser
description	Use when a task needs the real browser — harvesting the logged-in Learn session, rendering SPA pages (course outlines), capturing quiz pages, or any authenticated web automation. (Assignment submission is NOT here — it uses the raw D2L API; see learn.py submit.) This project standardizes on Vercel agent-browser (CDP CLI); it reuses a logged-in browser session, so no second login.

agent-browser — the single browser engine for agentic-learn

This project standardizes on Vercel agent-browser (Apache-2.0 CDP CLI) for ALL browser work, wrapped by .learn/agent_browser.py. It keeps a persistent logged-in browser per session, so we log in once (WatIAM + Duo) and reuse it. Verified: it harvests the HttpOnly d2lSessionVal + XSRF.Token and those cookies authenticate the D2L API.

Install: npm install -g agent-browser && agent-browser install.

Use the project wrapper (.learn/agent_browser.py)

PY=.learn/.venv/bin/python
$PY .learn/agent_browser.py auth                    # harvest the logged-in session -> session.json
$PY .learn/agent_browser.py outlines                # render every course-outline link -> markdown
$PY .learn/agent_browser.py render <url> [--pdf p]   # render a page -> {text, pdf}
$PY .learn/agent_browser.py snapshot <url>           # compact a11y tree with refs (for UI automation)

Python API: harvest_session(base_url, session_path), render(url, pdf_path=), outlines(), snapshot(url), available(). All use one persistent session (SESSION="learn").

Raw CLI (when the wrapper doesn't cover it)

agent-browser --session-name learn open --headed <url>   # headed so the user can log in
agent-browser --session-name learn get url               # where are we
agent-browser --session-name learn cookies get           # ALL cookies incl HttpOnly (CDP-backed)
agent-browser --session-name learn storage local get XSRF.Token
agent-browser --session-name learn snapshot              # refs: button "Submit" [ref=e7]
agent-browser --session-name learn find role button "Submit"   # locate by role/text/label
agent-browser --session-name learn upload <sel> <file>   # attach a file to an <input type=file>
agent-browser --session-name learn click @e7             # click a ref or selector
agent-browser --session-name learn pdf out.pdf
agent-browser skills get core --full                     # built-in, version-matched usage guide

What actually works

• Session harvest: open Learn → cookies get (HttpOnly d2lSessionVal/d2lSecureSessionVal) + storage local get XSRF.Token. Feed those to requests; the D2L JSON API authenticates. This is the foundation — the rest of ingestion is the API, not the browser.
• Render session-gated pages (outlines): they render in the same logged-in learn session.
• Quiz capture / general UI: snapshot to get refs → find role|text → click. Low context (compact refs, not screenshots). NOTE: assignment submission is not done here — it's the raw D2L API in .learn/submit.py (a dropbox submission is one atomic mypost call).
• The persistent session survives across CLI calls (a Rust daemon) — log in once, reuse.

Common mistakes

• Forgetting --session-name learn: you'd drive a different browser without the login. Always use it.
• Expecting eval document.cookie to see d2lSessionVal: it's HttpOnly — use cookies get.
• Running render/outlines before auth: session-gated pages need the logged-in session first.
• Auth wall (ADFS/Duo): the harvest opens --headed so the user can sign in — never type credentials yourself.
• A failed render returns empty: agent_browser.render() RAISES on empty output — surface it, don't ignore.