[msmakouz]

What was changed

cache_ttl sets the prompt cache lifetime, as a call option or as a route default in providers[].options.cache_ttl. Agents set it with agent_options.prompt_cache.ttl.
The Claude driver puts "5m" or "1h" on every cache breakpoint and rejects other values before sending the request; other drivers ignore the option. Cache writes are reported per lifetime in tokens.cache_write_by_ttl and stored in the new token_usage_cache_writes table.

local result = llm.generate(prompt, { model = "claude-sonnet-5-5", cache_ttl = "1h" })                                                                                        
-- result.tokens.cache_write_by_ttl = { ["1h"] = 4800 }