Augur — browser-free, model-driven site mapper & tester

No Chromium. Plain HTTP + raw HTML. A local fast model (Ollama) is the oracle.

Pipeline

1. affordances.py — raw HTML → every click/submit as a fireable HTTP request (links=GET, forms=POST w/ fields, CSRF tokens lifted automatically). Deterministic, no model.
2. model_layer.py — the model layer: describe_page() classifies a page; infer_expected() predicts a request's expected response (the test oracle). Talks to Ollama; stub fallback offline.
3. crawl.py — FIRST PASS: model parses each page, crawler follows every internal link → site tree.
4. runner.py — fires each affordance and diffs actual vs the oracle's expectation → pass/fail.
5. mockshop.py — a tiny multi-page server-rendered site to run against with no Spree.

Run (live, with your model)

Ollama up + llama3.2:3b pulled. Two terminals:

python mockshop.py                     # terminal 1: target on :3000
python crawl.py http://localhost:3000  # terminal 2: real model builds the tree

Point crawl.py / runner.py at any server-rendered site (e.g. your Spree on :3000).

Config (env)

TESTER_MODEL    default llama3.2:3b   (try qwen2.5:1.5b / llama3.2:1b for more speed)
TESTER_OLLAMA   default http://localhost:11434
TESTER_STUB=1   force the offline deterministic oracle (no model)