BROKE / RECOVERED
One night, one agent session, the rest of a weekly quota
A supervising agent was woken 211 times in six hours. Each wake-up re-read a ~240K-token context.
- BROKE
- Codex weekly usage went from 22% to 100% between 21 and 25 Sep
- CAUSE
- Every permission prompt from a Claude Code worker woke the Hermes session that supervised it: 211 wake-ups, 808 API calls, ~240K context each
- FIX
- A supervisor script now answers routine prompts without a model and wakes the agent only for plans, questions, results and denials
- PROOF
- Unit tests for the supervisor; the old wait script kept for rollback
Per-turn budgets never fired, because no single turn was expensive. The cost was spread across hundreds of small wake-ups, each one paying for the whole context again. The mechanism was cross-turn, so the guard had to be too.