Pretty sure on advisor it needs to send the entire context to another model. The cost is mitigated a bit because there would be few output tokens from the advisor model normally.

Orchestrator seems to automatically delegate things to subagents as required is my read of it. This should be fine for caching.

I've found hydrafusion to be great, although it sometimes does an unnecessary review pass at the end for simple tasks too.