Run the daily eval→fix loop on a chatbot: pull the OpenBat digest of recent failures (flags, issues, outcomes + reasonings), drill into representative conversations, map each problem cluster to the right lever (system prompt / tool logic / retrieval / new analysis / new alert), and apply fixes in the customer's repo with confirmation. Triggers on 'optimize my chatbot', 'what went wrong yesterday', 'fix my chatbot from the conversations', 'daily eval', 'why is my bot failing X'.
