Microclimate Forecasting

Free, zero-maintenance hourly temperature and probability-of-precipitation forecasts for a local station, by downscaling Environment Canada's HRDPS. Designed around Lethbridge, Alberta; deployable for any microclimate by config.

• Domain glossary (source of truth for terms): CONTEXT.md
• Architectural decisions (source of truth): docs/adr/
• Data licenses & attribution: DATA_LICENSES.md
• Scaffolding spec: docs/superpowers/specs/2026-05-30-scaffolding-spec.md

docs/adr/ records why the architecture is the way it is, and CONTEXT.md defines the vocabulary the whole project uses. Both are kept current as the project evolves — a change that alters a decision or introduces a concept updates them in the same PR.

The architecture is enforced mechanically: typed boundaries (Pydantic/Pandera), connector ABCs, a single feature-snapshot builder, source-eligibility validation, and an import-linter layer contract — all gated in CI.

Project status

Pre-1.0, built bottom-up. The data backbone is implemented: typed contracts, validated deployment config, the connector framework with live + historical sources (Environment Canada observations, HRDPS from Datamart and CaSPAr), features.build_snapshot — the single as-of feature path shared by training and inference (ADR-0011) — and features.build_features — the read-time transform from a FeatureSnapshot to the feature matrix (derived features + explode-to-per-lead-hour rows, ADR-0012).

Not yet implemented (currently stubs): the temp/PoP models, evaluation + publish gate, forecast-JSON / registry publication, and the inference/training pipeline CLIs.

Develop

uv sync                      # install deps + dev group
uv run ruff check . && uv run ruff format --check .
uv run lint-imports          # enforce the layered architecture
uv run pyright               # strict type check
uv run pytest                # network-marked tests are deselected by default

These mirror the CI gate (.github/workflows/ci.yml). The package lives in src/microclimate, layered low→high (contracts → config → connectors → features → evaluation/models → publication → pipelines); the layer order is enforced by import-linter. Deployment configs live in config/deployments/; the thin-client dashboard in dashboard/.