ahmad2b/langgraph-voice-call-agent — reverse-engineered prompt

Build me a Python backend for a voice call AI agent that lets someone talk out loud to a LangGraph assistant through LiveKit, like a simple ChatGPT voice mode.

I want it to handle a live conversation with low delay, including detecting when the user is speaking, turning speech into text, sending that text to a LangGraph agent, then speaking the answer back. Please include a basic example agent that manages a todo list, with add, list, complete, and delete actions, and make delete ask for confirmation.

Set it up so it can run locally with Docker for LiveKit, a local LangGraph dev server, and environment variables for LiveKit, OpenAI, Deepgram, and the LangGraph URL. Include clear commands to install, download any needed voice models, start the services, and run the agent. Please keep the code organized and easy to swap in a different LangGraph agent later. Look up current docs online if needed.