loop-budget

name	loop-budget
description	Enforce AutoResearch loop cadence, attempt, token, time, sub-agent, remote-compute, and kill-switch limits.
user_invocable	true

Loop Budget Guard

Run at the start and end of every loop iteration.

Start

1. Read loop-budget.md, STATE.md, and the last 24 hours of loop-run-log.md.
2. Exit if .loop-paused exists or state is paused.
3. Stay report-only at 80% of the daily cap; exit at 100%.
4. If nothing is actionable, return a no-op in under 5k estimated tokens.
5. Enforce L1: one candidate, zero sub-agents, zero remote compute.

End

Append a structured record to loop-run-log.md with timestamp, trigger, evidence/run id, candidate count, actions, attempts, escalations, duration, token estimate, verdict, and stop reason. Never rewrite previous entries.

For L2, stop after three attempts for a hypothesis or two identical failure signatures, even if time/token budget remains.