Idle cache rebuild

Key: idle_cache_rebuild

What it is

Claude’s prompt cache keeps your conversation for 5 minutes after its last use (or 1 hour with the extended cache). Come back later and the next message has to write the whole context into the cache again, at 1.25x the input price, instead of reading it at 0.1x.

On long sessions with large contexts, this is often the biggest single source of waste. On the author’s own logs it was about a third of all priced input.

How it’s detected

A model call is flagged when:

1. it writes at least 20,000 tokens to the cache, and
2. the previous call in the same stream was more than 300 seconds earlier, or more than 3,600 seconds if the call used the 1-hour cache.

Waste = tokens written x (write weight - read weight): the extra cost over reading the same context from a warm cache.

How to fix it

• Run /compact before a break, so the context you come back to is small.
• Use /clear when you switch tasks, instead of continuing a huge session.
• For long sessions with breaks, use the 1-hour cache: writes cost more (2x) but survive the break.
• Finish a thought before stepping away rather than leaving the agent mid-task.

Note

Some rebuilds can’t be avoided, such as the first message after compaction or a model switch. Treat the number as an upper bound.