cache:true on an ordinary render to reuse a recent result with the same validated, response-affecting options. Cache hits avoid Browser Rendering and reduce latency. Credentials are represented only by an opaque fingerprint in cache identity.
AI design specifications use two account-scoped caches automatically:
- A 24-hour snapshot cache stores the complete design-bearing snapshot and is reusable across AI providers.
- A seven-day specification cache is isolated by account, provider, server-selected model, credential mode, and provider-key fingerprint.
refresh:true on an interpreted blueprint to bypass the interpretation cache and generate a new document, so the non-cache tariff applies. The capture itself always runs.
Cache writes are best-effort. A write failure does not fail an otherwise successful render or specification, and failed AI generations are never cached.