Wire mic through Whisper-compatible transcriptions on LLM_BASE_URL. Photos upload to MinIO and reach the vision model as base64 image_url parts.