v0.2.9

Highlight: one-click integrations for OpenClaw, OpenCode, and Codex + M5 Neural Accelerator support

• you can now set up OpenClaw, OpenCode, and Codex from the oMLX admin dashboard with a single click and paste. no more manual config editing.
• M5 Neural Accelerator support added. if you're on an M5 Mac, download the macos26-tahoe DMG for full performance. the macos15-sequoia build does not include M5 acceleration.

New Features (v0.2.9)

Qwen3-Reranker support (PR #165)

• added CausalLM-based reranker support for Qwen3-Reranker family models. uses yes/no logit scoring, auto-detected by architecture + directory name, loaded via mlx-lm.

Bug Fixes (v0.2.9)

Codex agentic coding stops mid-session

• fixed Responses API input converter creating duplicate assistant messages per tool-call round, causing models to output EOS after several rounds. assistant message and tool_calls are now merged into a single turn.

Tool-call streaming markup leakage (follow-up)

• fixed additional edge cases where tool-call envelope markup ([Tool call: ...]) leaked as literal text during streaming. covers bracket prefix sanitizer for multi-prefix detection, unresolved bracket prefix leakage, and mixed bracket marker scenarios (#172, PR #174)

Other fixes

• fixed menubar app now resolves host/port from settings instead of hardcoding
• fixed built-in chat now uses the configured default model instead of the first model in the list (#168)
• fixed model download stall detection with mtime tracking and sequential queue
• added --macos-target build flag for platform-specific mlx wheel swapping

full changelog: v0.2.8...v0.2.9

New Contributors

• @shyuan made their first contribution in #165
• @jwcrystal made their first contribution in #168

Thanks to @shyuan, @jwcrystal, and @lyonsno for their contributions!