Hi Archestra Team 👋 Quick finding from an exported session log (Opus 5.5, 44 calls, ~$5.75): tool results never get a cache_control breakpoint. Only the first and latest user text messages do. So in multi-step tool loops, cache reads stay flat while the growing transcript is re-billed at full input price on every call. New user messages also trigger a full re-cache (0 read).

A rolling breakpoint on the last block of each request would have cut this session to ~$2.00. Is the caching strategy configurable, or could this be added? Happy to share the export