01 — Meaning
What a hit is
The provider already had KV tensors for this prefix, so it skipped that work.
A cache hit skips work.
It does not skip the model.
Cached tokens are reused prefix tensors from prefill — not a stored answer. The reply is still generated live.
02 — Trigger
When it fires
Identical prefix from token zero, long enough, already written, still inside TTL.
03 — Optimize
How to hit more
Stable tools, system, docs, and history first. The new user turn last.