uv run python -m pipeline runs the default steps, or one with --step <name>. prepare, extend, differences and tiles skip work whose inputs match outputs/state.json. download --only <source>, prepare --only <source>, extend --source <name> and tiles --only <id> always run what they name.
--step differences keeps its tables in tmp/differences.duckdb, so --from assign, --from overlay or --from review reruns that phase and all after it. Phases read as: land (land and lines), assign (cells, pieces, polygons), overlay (matches, land patches, island gaps), review (items, borders, site files, with an island difference listed only where both sources draw islands in the zone their two cells share). DIFF_BBOX=xmin,ymin,xmax,ymax limits a run to one area and writes partial results, which --step differences replaces on its next run.
The site's tiles and tables are published as a Portolan catalog in catalog/, which the catalog step generates (only its .portolan/ settings are committed). After the tiles step, build and push it with:
uv run python -m pipeline --step catalog --pushPushing needs PORTOLAN_REMOTE and the source.coop keys in .env (see .env.example). --only <id> ... stages a subset of layers or tables. uv run pytest checks the portolan wrapper against a scratch catalog (about a minute).
uv run python -m pipeline --step build builds the build layer in numbered stages under outputs/build/. Stage 07 extends it into water. See docs/reference/build-data.md.