The left edge changes every call. Matching dies at token 0.
Same tools, different key order. The prefix hash is not the same.
Picking 3 of 20 examples per request is diversity for you, a miss for the cache.
Summarizing, redacting, or resorting history invalidates everything after the edit.
tool_choice, web search, effort, and extra images are part of the prefix.
Claude writes a new hash every time. Lookback will not rescue a write that never existed.
A four-minute generation burns a five-minute cache. The follow-up must start before expiry.
The write is not readable until the first response begins. The herd all miss together.