v1.3.0

Added

• Experimental: Report model-download progress and replace the fixed per-file timeout with a stall-based one. install-model shows a live percentage/rate/ETA line (plain milestone lines when output is piped), and ad-hoc provisioning at startup logs throttled progress. Downloads now abort only after stalling (no data received for a while) plus a generous absolute ceiling, so large models on slow links finish; a stalled startup download fails with guidance to pre-provision via install-model.
• Add per-service embedding.root for model lookup without changing the CAP project root, and embedding.offline for read-only local-then-home model lookup.
• Export installModel from @cap-js/ai/models.js for validated programmatic provisioning without CAP plugin hooks on import; existing package entrypoints and deep paths remain available.

Changed

• Download local embedding models into ~/.cds/models by default on macOS and Windows, shared across projects. Runtime lookup still prefers project-local .cds/models; explicit directories remain offline and project-local overrides require explicit repair.

Fixed

• Register recommendation handlers only on application services. A READ handler on the messaging service creates an Event Mesh subscription for the topic READ.