Asked about Seeing what the system did
Why record cached token counts?
An answer that works
It is the only way to tell whether prompt caching is working. If cached tokens are consistently zero, something volatile has crept into the prompt prefix and your largest cost saving is silently off.
Reading an answer is not being able to give one. The section that teaches this is Seeing what the system did, 4 hours of Tracing & observability.
Learn this properly