The SDK already caches AI-driven browser steps in local .cache-lock files so later runs replay without an LLM. Those caches can now optionally mirror to Donobu Cloud when a Donobu API key is present and the account is entitled. A local miss consults the cloud and seeds the repo file, so teammates and CI share solved page.ai steps.
Default mode stays local-first. Unentitled or unreachable cloud leaves behavior unchanged. Cloud failures degrade silently to local. Clearing the cache still forces a fresh solve.
- Modes:
local-only,local-first(default),cloud-first,cloud-onlyviaDONOBU_AI_CACHE_MODE - CLI:
donobu cache status,push, andshow - Structured cache entries, not generated code; absolute paths never leave the machine
- Clear-cache paths never read from the cloud (repair loop)