v2.2.0

Highlights

• Opt-in API token authentication for the chap-core API, letting deployments require a token without breaking existing unauthenticated setups.
• Pluggable future-weather providers with a look-ahead flag, so forecasts can draw future climate data from different sources.
• Consolidated backtest endpoints with persisted backtest parameters, making evaluation runs reproducible and inspectable after the fact.
• Single consolidated Helm chart replacing the previous multi-chart layout, plus a self-contained compose.ghcr.yml deployment file.
• Python 3.14 compatibility.

Features

• Opt-in API token for the chap-core API (#512)
• Pluggable future-weather providers with look-ahead flag (#568)
• Endpoint suggesting covariate names for model-independent datasets (#575)
• Persist backtest parameters and consolidate the backtest endpoints (#569)
• Ensemble stacking (#534)
• optimization_direction in MetricSpec and extra search-space argument (#552)
• Restart policies and container healthcheck gaps closed (#532)
• Chart: let the registration job be tracked (#516)
• Compose: make compose.ghcr.yml a self-contained deployment file (#550)

Refactors and performance

• Consolidate CHAP Helm charts into a single chart (#489)
• Align prediction horizon fields with chapkit (CLIM-972) (#547)
• Move train-test-generator out of backtest into Evaluation.create (#554)
• Plotting: drop vegafusion and pyarrow, compile Vega with vl-convert (#469)
• Vectorise sample CRPS via order-statistic formula (#467)
• C4 architecture model with multi-renderer diagram export (CLIM-791) (#470)

Bug Fixes

• Make chap-core compatible with Python 3.14 (#540)
• Make service registry heartbeat atomic so pings cannot resurrect or overwrite registrations (CLIM-1073) (#576)
• Bound chapkit client requests, treat cancelled jobs as failures, keep error detail (CLIM-1074) (#577)
• Drain chapkit service output so local services do not block (CLIM-1072) (#574)
• Persist data sources when creating a dataset via make-dataset (#578)
• Fail backtests when a model skips org units or splits (#571)
• Preserve forecast locations for models without covariates (CLIM-1048) (#562)
• Preserve model template and configuration provenance (CLIM-886) (#527)
• Accept provenance fields from chapkit 2.0 services in the service info parser (#564)
• Populate dataset period metadata when loading evaluations (#443)
• Restrict all-locations dataset when last_n is set for explain-lime (#549)
• Report runner command output as text on failure (#536)
• Wait for DHIS2 before registering the chap route (#529)
• Stop service registration opening a second redis connection (#511)
• Stop forwarding X-Service-Key in the v2 service proxy (#513)
• Keep the service proxy out of the OpenAPI schema (#530)
• Use chap_core as coverage target in make test-all (#493)

Documentation

• DAC2026 series with configurable model tutorial (#442, #444, #446)
• Document compose overlays and running your own model (CLIM-504) (#541)
• Clean up last_n docs and fix surrogate_name options for LIME (#553)
• Point local DHIS2 setup at docker-dhis2-core (#538)
• Correct X-Service-Key scope in the API authentication page (#514)

Full Changelog: v2.1.0...v2.2.0