Casita

Casita scrapes rental listings from Zillow, Craigslist, Zumper and Redfin, normalizes them into SQLite, enriches and ranks them with Gemini, computes walking times to curated San Francisco and Marin anchors, and publishes the result for review.

This fork replaces the frontend. The pipeline, the ranker and the detail pages are unchanged.

My reasoning for every decision below is in WRITEUP.md.

What I changed

The old frontend was a single static page. Every listing stacked in one scrolling document, with each listing's walking times printed as rows of text.

The new frontend is a browser app in app/, three pages over one state store, served at the same address the old page was.

Page	What it does
Data	pick sources and places, set price, bedroom and walk limits, see what enters the deck
Review	one listing at a time, with a map of the geography the ranker scores on
Settings	the anchors and priorities that hold between runs

Ranking is geographic. rank.py weights walk to trail as its primary signal and walk to beach as its secondary, walk.py carries 19 anchors, and the fixture ships 5,553 cached route results. The app draws that on a map instead of printing it as a table.

A few other things changed.

• visual_summary was populated on all 137 listings but wasn't rendered anywhere for users to look at. It leads the back of the card.
• The demo server now serves HTTP range requests, which the map and the video both need.
• 49 tests were added. 42 browser tests driving the real UI, plus 7 covering the dataset builder. The suite went from 18 to 67.

Run it

uv sync
uv run playwright install chromium
uv run casita demo

Open http://127.0.0.1:8765/.

The demo is credentials-free and reads a sanitized SQLite fixture with cached route times and precomputed enrichment. It does not scrape, call Vertex, read GCS, deploy to Firebase, or call the Google Maps Routes API. It does use Playwright's local Chromium to render Open Graph preview images.

Four routes are served there.

Route	What
/	the review app
/listing/<slug>	the generated detail page, unchanged
/og/*	Open Graph preview images, unchanged
/classic.html	the single page frontend this replaced

Nothing about the app needs a build step. It is plain modules, vendored MapLibre and PMTiles, and one generated JSON file.

Tests

uv run pytest

67 tests. The browser tests drive the real page in Chromium and assert on computed values, not on markup.

make check

Compiles the modules, runs the suite, runs the public leak validator, builds the docs, builds the package, and checks the CLI imports.

Live runs

Unchanged from upstream, still optional and environment driven. Copy .env.example to .env first.

uv run casita search --headed --local
uv run casita enrich --local
CASITA_FIREBASE_PROJECT=your-project uv run casita publish --local

search and enrich can call the paid Google Maps Routes API. publish requires Firebase Hosting and deploys to it.

Layout

app/                     the frontend
  index.html             one shell, three pages
  js/state.js            membership, one rebuild(), one subscribe
  js/card.js             one renderer, both faces of a card
  js/map.js              pins, anchors, camera, legend
  js/dogs.js             the reject animation
  js/{scrape,review,settings}.js   the three pages
  js/{data,share}.js     the extract, and the vote payload format
  map/                   MapLibre style and the basemap extract
  vendor/                MapLibre and PMTiles, vendored
src/casita/app_data.py   builds the app's dataset from the listings DB
src/casita/__init__.py   CLI, render, and the app install step

Upstream project docs are at https://matin.github.io/casita/ and in docs/. AGENTS.md describes the repo for coding agents.

License

MIT, same as upstream.