Prompt caching for agents: semantic deduplication, cache hit detection, token savings calculation, TTL management, cache