I gave Cypher another upgrade.

It was writing thousands of prompt tokens into OpenAI’s cache without reading them back.

I changed how its context is assembled, separated stable instructions from changing memory and conversation data, and added request-level telemetry so I can see what is actually happening.

Testing confirmed one cache write followed by three successful reads.

I also rebuilt the usage and admin pages so the system is easier to inspect, manage, and verify during real use.

Not magic.

Useful.