What's Changed
- Save model-aware adaptive speculation defaults per agent (#3247) by @jjang-ai
- Redesign agent generation overrides with inherited defaults and stop chips (#3246) by @jjang-ai
- Reuse compiled dependencies in PR CI and save caches only from main (#3234) by @tpae
- serve phone chats' context usage and make local-model phone calls faster (#3209) by @RaajeevChandran
- Archive Release builds without whole-module optimization (#3235) by @tpae
🐛 Bug Fixes
- Initialize custom agent JSON fields and clarify inheritance (#3245) by @jjang-ai
- Preserve captured generation settings in queued chat (#3244) by @jjang-ai
- Consume checked prefill and serialized tokenizer fixes (#3243) by @jjang-ai
- Honor bundle sampling defaults for thinking turns (#3241) by @jjang-ai
- Detect bundled DSpark and honor declared reasoning defaults (#3236) by @jjang-ai
Full Changelog: 0.25.28...0.25.29
Downloads
- Osaurus-0.25.29.dmg — standard build. Pick a model during onboarding.
- Osaurus-0.25.29-full.dmg — 3.6 GB, Raptor 0.6 bundled. Chat offline right after install; no model download.
sha256 0a667a2292f686ce375053fb0c27ffbec4e3bc5dc80db58eb18054a368f2a4bf