I'm a little surprised that preserve_thinking would matter here for cache purposes. for actual capabilities/intelligence, yes, I'd imagine it helps to have past reasoning traces in multi-turn setups.
but for caching, all you are doing is leaving off a fraction of the most recent assistant message generation, which will have little/no impact on cache hit rate.
but for caching, all you are doing is leaving off a fraction of the most recent assistant message generation, which will have little/no impact on cache hit rate.