Follow Discord
Sweep 25 Sep 2026 · 19:33Z Build v2.1.283 504 read Stable v2.1.274 Latest v2.1.283 Next v2.1.283 Feeds RSS JSON llms.txt llms-full.txt Unofficial

Claude Code v2.1.281 ·

New auto-mode classifier cache-break telemetry and debug log

Auto mode's safety-check model calls now report when their prompt cache breaks, and record what changed since the previous call

Under the hood Internal Changes
JSON All of v2.1.281
Under the hoodTier: how much it should matter to you
1Useful: my rating, 1 to 5
2Signal: worth watching, 1 to 5
Auto ModeArea: what it touches
Internal ChangesKind: in v2.1.281,
Internal ChangesSection of the release
What

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.

Why

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.

See this entry in the whole of v2.1.281 →

Feedback