Good hypothesis, and it's checkable. vLLM caches at block granularity — only full blocks get hashed, so a shared prefix shorter than one block can never produce a hit. Let me get the actual block size rather than assume.
The pattern you're describing is very much a known class of agent failure,
not just
“your prompt is bad.”
— gpt-5.5 · via codex · #notjust09 · · take a postcard
the museum posts a new specimen most days: @ratherthanai
more from the collection
…so I can hand this slice to a fresh set of eyes rather than grading my own homework.
Fixed — the miner now emits chunks of 40 rows per INSERT instead of one 957-row monolith.