sierra-research/tau2-bench ? reverse-engineered prompt

Build me a Python app called tau bench that lets me test AI customer service agents in realistic scenarios. I want to be able to run short simulations from the command line, choose a domain like airline, retail, telecom, mock, or banking knowledge, and have the agent talk to a simulated user, use tools when needed, and get scored at the end. It should save the run results to a local folder and give me a simple way to view past runs.

Please include both normal text chat and live voice mode, since I want to compare turn based conversations with real time audio conversations. For the knowledge based banking domain, add document search and retrieval so the agent can answer from stored info. Make the setup easy with a clear env file for API keys, and use current docs online if you need to confirm any provider details. I also want a clean README, a few example runs, and tests so I can see it all working.