changing-the-scraperlisted

# Changing the scraper ## Overview Ollama exposes **no public usage API**. The numbers come from scraping the logged-in settings page HTML. That makes the parser the most fragile part of the project, and the rule that keeps it usable is: **a page change must degrade, not crash.** Files: `src/ollama_usage/parser.py`, `src/ollama_usage/fetch.py`, fixture `tests/fixtures/settings.html`. ## What the parser reads ``` [data-usage-meter] ← parent; the reset time is its sibling └─ [data-usage-track] aria-label="Session usage 2.2% used" └─ [data-usage-segment] data-model="…" data-requests="12" [data-time="2026-08-04T16:30:00Z"] ← next sibling of the meter ``` - `identifier`: `"session"` in the aria-label → `5h`; `"weekly"` → `weekly`. - `pct_used`: first `([\d.]+)%` match in the aria-label. - `reset_at`: `data-time` parsed as ISO 8601 (`Z` → `+00:00`). - `models`: each segment's `data-model` / `data-requests`. ## The degradation rule Every failure path returns *less data*, never an exception: | Broken input | Behaviour | |---|---| | aria-label unknown or without a `%` | **skip that track**, keep the others | | `data-time` unparseable | `reset_at = None` | | `data-requests` not an integer | `requests = 0` | | segment without `data-model` | segment skipped | | no `[data-usage-meter]` parent | `reset_at = None` | Keep this shape. A parser that raises turns a cosmetic page tweak into a `status` line that says `🦙 Unexpected Error` for every user