Handling long-running LLM streams in a stateful backend | Not Hacker News!