In auto mode, Claude Code makes a separate model call, the classifier, to judge actions before they run. After each of those calls, a new helper now tracks prompt-cache reads and writes for each conversation and each place the call is made from. The prompt cache is a stored copy of the start of a prompt that the server can reuse instead of processing it again.
The helper sends a tengu_classifier_cache_break telemetry event and writes a line to the debug log in these cases:
- The cache is written for the first time
- The amount read from the cache drops
- At least half of the prompt is written again
The event records what changed since the previous call: the model, the system prompt, the cache lifetime, betas, extra request fields and thinking settings. It also records how far the previous prompt still matched and whether a compaction happened in between. A compaction is when Claude Code shrinks a long conversation to fit.
A broken cache means more of the prompt is processed from scratch, which costs time and usage. This event and debug line show when that happens to the classifier and point to the likely cause.