The cache shouldn't affect inference. It is purely an I/O optimization.
I think it should, as you dont need to use the encoder layer on the new tokens, you just read the embedding from the cache. that's why cache reads are cheaper
I meant, it shouldn't affect the resulting LLM output. It's a performance optimization that doesn't change the behavior.
Is that from start of a new conversation per conversation?
It's supposed to be for token optimization (https://code.claude.com/docs/en/prompt-caching), but are people experiencing degraded performance when you let Claude Code sit for hours/days and come back?