Osaurus 0.24.3

What's Changed

• Advance the vmlx-swift pin to 1315fdcb (DeepseekV3 MoE f32 residual fix) (#2573) by @jjang-ai
• Add Claude Code CLI integration (#2257) by @Yao920127
• Warn when switching models mid-conversation (#2538) by @jjang-ai
• Repin vMLX for 90+ tok/s Ornith 35B decode (#2553) by @jjang-ai
• Repin vMLX for scoped Ornith 35B MTP safety (#2549) by @jjang-ai
• evals(community): OsaurusAI/Muse-Glimmer-30B-JANG_6M on Apple M5 Max (#2544) by @jax-0n-git
• Repin vMLX with Qwen 3.8 legacy-parser correction (#2543) by @jjang-ai

🐛 Bug Fixes

• Stop the memory-safety profile default from clamping the MLX allocator pool (#2572) by @jjang-ai
• Use Raptor for mainstream onboarding defaults (#2574) by @jjang-ai
• Preserve reusable cache checkpoints across tool dispatch (#2571) by @jjang-ai
• Fix embedding discovery in configured models directory (#2570) by @jjang-ai
• Fix stale agent completion and heal misaligned model bundles (#2567) by @jjang-ai
• Hide local memory warnings for cloud models (#2566) by @jjang-ai
• Show and cancel exact live inference work (#2563) by @jjang-ai
• Restore type tags on Responses Lite namespace tools (#2565) by @tpae
• Preserve the admitted allocator ceiling for native MTP (#2561) by @jjang-ai
• Recover media rejections and harden OpenAI Responses (#2562) by @tpae
• Remove hidden HTTP reasoning-budget coercion (#2560) by @jjang-ai
• Report native MTP fallback only when MTP did not run (#2554) by @jjang-ai
• Give loaded skills a directory anchor for bundled resources (#2555) by @jjang-ai
• Repin cancellable speculative prefill runtime (#2557) by @jjang-ai
• Make denied chat tool outcomes visible and bounded (#2539) by @jjang-ai
• Repin serialized MLX host reads for Qwen3.8 server crash (#2556) by @jjang-ai
• Recognize bundle-declared local reasoning channels (#2545) by @jjang-ai
• Evals: score grounded filesystem access across equivalent tools (#2548) by @jjang-ai
• Make batch evals architecture-aware and truthful (#2536) by @jjang-ai
• Evals: preserve agent-loop exit and architecture cache taxonomy (#2546) by @jjang-ai
• Preserve explicit reasoning on the first cold send (#2537) by @jjang-ai
• Remove the hidden Orchestrator tool kill switch (#2541) by @jjang-ai
• Fix sequential spawn capacity after active file-cache growth (#2535) by @jjang-ai

Full Changelog: 0.24.2...0.24.3