Keeping vLLM's Prefix Cache Warm Between Agent Turns