The user requests the ability to configure the Anthropic cache's Time-To-Live (TTL) both globally and on a per-agent basis.
### AI Disclosure - [ ] This issue was written entirely by a human - [x] This issue was written with AI assistance (e.g. Copilot, ChatGPT, Claude) and **reviewed and edited by a human** - [x] I have read the [AI Policy](https://github.com/letta-ai/letta/blob/main/AI_POLICY.md) and agree to its terms ### Human Verification I have read the AI policy and I confirm this issue was reviewed by a human. ### Problem Statement ### Summary: The current Anthropic prompt caching implementation (PR #2684) hardcodes {"type": "ephemeral"} which defaults to a 5-minute TTL. Anthropic now supports a 1-hour TTL option ({"type": "ephemeral", "ttl": "1h"}), but there's no way to configure this in Letta. ### Use Case: I run a multi-agent system where multiple companions take turns in group conversations. Each agent is on a different model (Claude, ChatGPT, Gemini, Grok, Kimi, GLM, etc.) A typical round-robin conversation might involve 6-8 agents, with each turn taking 30-60 seconds of user inte