Someone replayed 68,266 real Claude Code requests and plain LRU beat every clever cache policy

agentic-kv-cache simulates cross-request prefix caching against real traces, not synthetic ones: 68,266 requests across 393 Claude Code sessions at 64-token blocks, plus Mooncake traces at 512-token blocks. It models prefix-contiguous hits, radix eviction constraints and pinned in-flight chains, with physically-modelled recompute costs, then races LRU-leaf against TTL-300s, LFU-leaf and a hazard-prediction policy.

At 40,000 blocks, LRU-leaf wins. TTL-300s is byte-identical to it. The hazard-prediction policy degrades everywhere.

The explanation is the part you can carry to your own gateway. 33.1% of recompute tokens come from sub-10-second gaps inside tool loops. Only 17.5% come from gaps over five minutes. So the dominant cost isn't stale entries you should have predicted dying, it's capacity pressure during a burst of tool calls that all want the same prefix within a few seconds of each other. A policy that predicts liveness is solving the smaller half of the problem, and the machinery it adds costs more than the prediction earns.

Which explains the TTL result too. If almost nothing useful dies of old age in the window you care about, a 300-second TTL and an LRU eviction make the same decisions on the same traces. Byte-identical isn't a coincidence, it's what happens when your discriminator never fires.

I've watched this shape of mistake in other caches. You look at a hit-rate curve, assume the misses are recency mispredictions, write a scoring function, and ship something that's slower and harder to debug for a rounding error of improvement. The measurement that would have saved the work is a histogram of inter-request gaps, which takes an hour.

If you're building a gateway or prefix cache in front of a coding agent: don't write the smart policy. Run LRU, then spend the engineering on capacity and on not evicting a prefix while a concurrent request is mid-flight on it, which is where the 33.1% actually lives. And be suspicious that this is one trace set from one harness. Different agent shapes have different gap distributions, and the honest version of this finding is "measure your own gaps," not "LRU always wins."