sernote

Use whenever the user mentions LLM prompt/prefix cache misses, cached_tokens=0, cache_read_input_tokens/cache_creation_input_tokens, prompt_cache_key, cache_control/cachePoint, TTFT/prefill latency, KV reuse, LLM cost or speed regressed on repeated long prompts, or speeding up agents through cache stability. Use for LLM request shape changes: prompt text/order/builders, tools, schemas, response_format, model/router, agent loops, or compaction. Not for generic prompt writing/RAG, token counts, or non-LLM perf.