Token usage leaks happen during the run, not after billing refresh

Most AI cost overruns come from prompt loops, retries, and fallback hops while you’re still debugging.

Three practical moves:

• keep a live token signal visible during active runs
• flag sudden context growth early
• compare local vs API token movement before choosing bigger models

If you can see usage in-session, you can stop waste before it compounds.

Built this into a tiny macOS menu bar app: tokenbar.site