Codex users report rapid quota drains since May 10

One practical way to debug this is to separate three counters that often get collapsed together: prompt text size, tool/runtime activity, and verified progress. A short prompt can still become expensive if it opens browser/tool loops, retries after failures, or keeps re-reading large context without producing a durable change. The usage UI would be much easier to trust if each Codex run exposed a small receipt: model used, wall time, tool calls, context/read volume, whether files/tests changed, and the stop reason. Then users could tell the difference between a real metering bug, a workspace-specific loop, and a normal higher-cost task.