## Usage: one meter, and spend what expires - **Before any big delegated job, run `meter.py`.** It's about 14 lines, no model call, no network. Never scan transcripts or logs to work out usage; that costs the tokens it measures. - When I read `/usage` or an account page, record it: `meter.py --read `. A human reading is the only ground truth for plans that hide their percentage. - **Unused capacity expires.** When a window is near its reset with room left, spend it on purpose: the queue of problems the cheaper agents already failed goes first. - **Don't hoard an expiring bucket.** Send it the work it does best, and move everything else to another agent. - **Size burn-off jobs for the budget:** two problems per agent, and write the first file early. An agent that reads everything first can be cut off with nothing written. - **Respect the stop lines** in the meter. A self-imposed stop is a decision, not an error. - **After a failed worker call, assume less headroom than the meter shows.** Failures are billed but often not logged.