Navigation
ELOWEN · Documentation

Last updated: October 27, 2026

Browse documentation · Understand prompt cache alerts
Autonomous work and permissions

Understand prompt cache alerts

A provider keeps the beginning of your conversation cached, so the next request reuses it instead of reading everything again. Chetty warns you when a reply missed that cache and paid noticeably more for it.

Why the cache matters

With the default long retention the cache stays warm for about an hour of quiet; with short retention, about five minutes.

A request after a longer gap reads the whole history again. That costs more, and it is normal.

What the alert says

A break inside the warm window means the request reused less of the cached text than it should have. When one step costs at least $0.25 extra, Chetty raises a warning on the Alerts bell in the top bar, at most one per reply.

The alert names the conversation, or the sub-agent for delegated work, and states the likely cause. It also lists Lost from cache, Written again, Processed again, Extra cost, Call cost and Model, and links back to the conversation.

What can cause a drop

Chetty compares the request with the one before it, so it can name the usual reasons:

  • The system prompt changed. Chetty's instructions or your Project instruction files changed.
  • The set of available tools changed. A plugin or a permission change altered which tools the turn could use.
  • Earlier conversation history changed. A message was edited, removed or rewritten.

When the prefix is unchanged, the alert says the cause is unclear and that the provider probably evicted the cache, or that a step may have moved an earlier cache write out of reach. Switching between Plan, Build and Workflow changes neither the instructions nor the tools, so it does not explain a drop.

What to do

Most cache drops need nothing. Read the cause, then decide:

  • If the alert names a real change, no action is needed; your next request starts a new, stable prefix.
  • If you see the same alert repeatedly in one long conversation, compacting the conversation or starting a fresh one often costs less than carrying the history forward.
  • Dismissing the alert does not change the cause; it only clears the notice.

Tip: A working conversation shows its cache share beside its live speed, and the usage panel labels the same figure Share read from cache.

Continue with Make room with context compaction, Usage and cost figures and What you see in a conversation.