What's Changed
- docs: reposition — Swift in-process engine is the differentiator by @magicnight in #59
- docs: v0.6 roadmap — agent backend release by @magicnight in #60
- feat(mcp): chat-side tool routing core (wave 1) by @magicnight in #61
- feat(mcp): chat-side tool routing GUI (wave 2) — closes v0.5 tool routing by @magicnight in #63
- feat(server): OpenAI tools pass-through (/v1/chat/completions) by @magicnight in #64
- feat(batching): batch-decode RoPE-correct cache infrastructure (A1) by @magicnight in #65
- feat(batching): equal-length batch decode core (A2a) by @magicnight in #66
- feat(batching): BatchKVCache — ragged batch cache port (A2b) by @magicnight in #67
- feat(engine): classic draft-model speculative decoding (D1) by @magicnight in #68
- feat(batching): continuous-batching scheduler (A2c) by @magicnight in #69
- feat(server): batch-generation seam + routing (A2d-1) by @magicnight in #70
- feat(engine): batch serving seam implementation (A2d-2) — Track A complete by @magicnight in #71
- test(models): Qwen3.6 compatibility smoke (Track G #1) by @magicnight in #72
- build(deps): controlled minimal fork of mlx-swift (rope batched fix) by @magicnight in #73
- feat(cache): trie-backed LCP prompt cache (Track B) by @magicnight in #74
- feat(engine): structured output — JSON constrained decoding (Track C) by @magicnight in #75
- feat(server): API compatibility pack (Track E) by @magicnight in #76
- feat(gui): oMLX UX adoption wave + speculative decoding UI (Track F/D2) by @magicnight in #77
- feat(models): Mellum2-12B-A2.5B architecture port (Track G) by @magicnight in #78
- feat(models): theoretical-tier trio — Solar, GLM-5.1/DSA, Kimi-K2.5 (Track G) by @magicnight in #79
- docs(release): v0.6 release prep — CHANGELOG, model support tiers, README refresh by @magicnight in #80
- feat(server): agent tool-loop protocol completion — OpenAI tool history + Anthropic tools by @magicnight in #81
- fix(release): CLI tarball ships mlx-swift metallib bundle by @magicnight in #82
Full Changelog: v0.5.3...v0.6.0