Three-tier runtime and bounded reminder reads
The shipped runtime now has three execution tiers. Deterministic local reads stay in Tier 0, the shared Bonsai 1.7B runtime and adapters handle recall, retrieval, and single-step work, and LFM2.5 2.6B handles multi-step agentic requests.
Changed
- The Core ML fast router is bundled with complete weights and is load-checked during packaging.
- Tier 2 receives goal-scoped tool schemas and releases the other heavy MLX runtime before loading on the 8 GB target.
- The current interface, history, settings, and structured result cards reflect the simplified runtime.
Fixed
- Reminder reads for today, tomorrow, this week, next week, and named weekdays now run deterministically in Tier 0 instead of being interpreted as reminder creation.
- Private reasoning and partial tool-call markers are withheld from streamed assistant text.
Known limitations
- Cold startup measured 54.381 seconds on the tested 8 GB M1 MacBook Pro.
- The optional speech model was not present on the test Mac, so typed requests—not voice transcription—are the verified path.
- The DMG remains large because all runtime models are bundled inside the signed application.