Take with you

Three answers, one metric.

WHAT

A hit is reused reading.

KV tensors for an identical prefix are skipped. The answer is still generated. cached_tokens greater than zero is the proof.

WHEN

Exact prefix, already stored, still alive.

From token zero, above the minimum, after the first write, inside the TTL, on the same model path.

HOW

Freeze the left. Measure the ratio.

Stable tools, system, docs, history. Variable question last. If hits drop, diff two serialized prompts from byte 0.

hit rate = cached_tokens / prompt_tokens
After turn one of a long agent chat, this should be most of the prompt. If it is not, the prefix moved.