subtopic

Monitoring, alerts and provider fall-backs

Track per call and per turn: latency by stage, errors, dropped calls, tool failures, cost. Set alerts (latency p95 above target, error rate up, a provider down). Add fall-backs: a second STT/LLM/TTS provider used automatically when the first fails or is slow; graceful messages and transfer to a human on errors.

Builds on: 2.7 Evaluate, 5.8 Evaluate and debug