Design a resumable streaming voice reply for an AI chat

Users talk to an AI agent in a chat box over a WebSocket. The service streams the model's output from an LLM API (over SSE) and turns it into audio that it plays back to the user frame by frame. A user can drop off mid-reply, so the client reports the last frame it actually played, and after reconnecting it should pick up right after that frame.

Cover:

• how a message flows from the user to the LLM and back as ordered audio frames
• where reply and session state live so app servers stay stateless and can scale out
• reconnecting to a different server and resuming from the last acknowledged frame
• deploys on Kubernetes: handling SIGTERM, draining or handing off open sockets, and choosing the termination grace period without losing what a reconnect needs