TL;DR: Claude Code injects dynamic telemetry headers and git status updates into the system prompt on every single request. If you are using a local inference backend like llama.cpp downstream llama-server or LM Studio, this dynamic injection instantly breaks prefix matching, flushes your entire KV
