What's this about?
Baoyu has now complained twice about a recent Claude Code feature that sends messages between running sessions. His objection has sharpened into a specific, measurable one: it notifies all historical sessions, waking long-finished conversations that no longer have any cache, which makes token consumption spike. It is on by default and he still cannot find where to turn it off.
What exactly is the cost mechanism?
Prompt caching means a provider stores your already-processed context so re-reading it is far cheaper than processing it fresh. Caches expire. A session finished days ago has no warm cache left, so anything that revives it pays full price to reload its entire context — and the notification produces no value in return, because the session is over. Multiply by every historical session and a background feature becomes a recurring bill. This is the flip side of the token economics Anthropic itself explained last week: cache misses are where the money goes.
Why is this worth noticing?
- The complaint is quantitative rather than aesthetic, which makes it actionable — it identifies a specific mechanism producing a specific cost.
- "Enabled by default with no discoverable off switch" is the real problem, and it lands badly in a week where users are already sensitive about limits and capacity.
- It's a useful counterweight to this week's positive Claude Code coverage — the same tool that shipped `/design` and halved its CPU usage also shipped this.