What's Changed
- Repin vmlx: capture cache boundaries during prefill (fixes a family that could never populate its cache) (#2335) by @jjang-ai
- Download large model shards as parallel HTTP Range requests (#1960) by @jjang-ai
- AppleScript subagent: tell the caller the repair that actually works (#2334) by @jjang-ai
- Repin vmlx: refuse truncated model shards, and stop the post-answer stall on every warm turn (#2333) by @jjang-ai
- Fix first-turn double prefill (warmup cache-salt mismatch), warmup tool-scope kill, cache-size settings hazards; repin vmlx (#2331) by @jjang-ai
- DSV4 Flash integration: vmlx repin (fastpaths + weight-leak fix), warmup triple-prefill fix, reload memory verdict (#2328) by @jjang-ai
🐛 Bug Fixes
- fix context impact ability count chip wrapping to two lines (#2323) by @RaajeevChandran
- Refresh MCP catalogs and streamline Tools management (#2322) by @tpae
- fix tts drift on long text with per sentence synthesis (#2321) by @RaajeevChandran
Full Changelog: 0.22.18...0.22.19