We finally swapped the NoopAudioMediaPlane default for a real PipecatAudioMediaPlane in PR #316, turning the abstract /voice on command into a live Whisper STT and Kokoro TTS stream over LiveKit. It's a small semantic change on the surface, but under the hood, we were moving the entire LiveKit bridge out of the main public surface into a dedicated mcp-server-voice/ server. This isn't just a refactor; it's an architectural bet that failure isolation matters more than convenience--when a voice session crashes, we want it to take down only the voice server, leaving create_post and health checks untouched. PR #315 While the audio work felt like a creative breakthrough, the observability side was a quieter, darker fix. We had been booting workers with a broken signature for setup_sentry that silently left the SDK uninitiated, causing errors to vanish into the void.…