How Cache Clock decides: every prompt Diagram

Where the host can ask, the refusal is a question: Send anyway goes through; Don't send, dismissed or no answer in 5 minutes holds the send (0 AI credits).
Where it cannot ask (copilot -p) the old refuse-once message appears, then the abort. Resend to pay.
session.abort() is not awaited: awaiting it inside the hook deadlocks. Awaiting select() there did not, measured live 12 Oct 2026 in Copilot CLI 1.0.95 and VS Code.
Asked once per cold episode (refused.json, 7 days); later sends go through with the note. A queued prompt during a running turn is only priced.

Early eviction: Luna was served cold at 29 of 30 min. The guard trusts cacheExpiresAt and stays quiet; the report counts a prefix break.
The price is an estimate: last input plus answer, not your new message.
With auto, the next prompt may use a different model than the one priced.
copilot -p: the guard works, but slash commands go to the model as text.