v0.9.3

0.9.3 - 2026-08-11

• Start normally on Windows when the filesystem rejects directory fsync, while keeping file writes atomic and surfacing real I/O errors. Add a Windows Node 24 CI lane.
• Emit each Gemini audio part once. The adapter no longer repeats the SDK's derived top-level data value alongside its source parts.
• Raise the browser playback buffer from 5 seconds to 2 minutes. Keep a hard memory bound and show one detailed warning if a response exceeds it.
• Honor explicit OpenAI audio.end messages in push-to-talk and VAD modes. Clients that stop sending packets at turn end no longer need to synthesize silence.
• Enable OpenAI input transcription by default and forward completed user transcripts without repeating the assistant transcript. The model and optional language hint are configurable.
• Resolve the release audit advisory in the development dependency tree.

What's Changed

• Fix realtime audio reliability on Windows, Gemini, and OpenAI by @bielcarpi in #57

Full Changelog: v0.9.2...v0.9.3