Watch out for cache read costs

Even though models advertise huge supported context sizes, you don't really want to use them unless really necessary. Not only they charge you double after crossing certain thresholds, but also there's this quadratic input token size property.

For this reason in Tau I've implemented quite a few ways to get agent to compact itself at the right time. E.g. one can set a threshold (context size) after which agent gets a short notice injected asking it to call compact tool when it's done with its current high level task. Currently I'm experimenting with status tool, which prevents agents from accidentally stopping themselves as they are working, and I might defer the notice or make to compaction automatically when they set their status to done.