Prompt caching gets a retention setting and a prompt-only mode

The Prompt Caching capability takes a retention setting: default, 5m, 30m or 1h. On Claude models, 1h keeps cache entries for an hour, which suits agents that pause between turns longer than the five minute default. One hour writes cost twice the input rate, and cost estimates now include that.

{ "ref": "prompt_caching", "config": { "retention": "1h" } }

The explicit strategy now works on Claude too: it caches the tool definitions and system prompt but not the conversation, for agents whose history changes from call to call.