landing-ai Verifying ade
Build/run: uv sync, then uv run ade …. Tests are offline (fake
transport); verification means driving the real CLI against a real store.
Seeding a parsed doc without the network
parse needs the API, so seed the store directly — write what parse
would have written (see parse.py::write_artifacts for the exact set):
parse.json (raw ParseResponse), parse.md, elements.json
({"job_id", "elements": elements.project(response)}), meta.json
(state: "parsed", job_id matching parse.json's metadata.job_id —
the generation gate in refs.live_parse rejects mismatches).
A ready-made seeder that also draws a matching invoice PNG (so boxes
visibly align) was used for the view verification — pattern: build
markdown piecewise so every element's range is exact, reuse
tests/parse_fixtures.py shapes.
Point the CLI at the seeded store with ADE_HOME=<dir>.
Driving view.html
The in-app browser refuses file:// — serve the doc dir instead:
python3 -m http.server 8742 --bind 127.0.0.1 in
$ADE_HOME/docs/<doc-id>/, then browse http://127.0.0.1:8742/view.html.
Deep links: append #element=<id>. Selection state is inspectable via
JS: document.querySelectorAll('.sel') (both panes share data-id).
Rebuilds are fingerprint-gated — --json reports built: true|false;
touching the source or changing --dpi forces a rebuild.