What problem does it solve? Autonomous research loops can run away with unbounded attempts, token spend, and remote compute usage. This Skill enforces hard cadence and budget limits at the start and end of every loop iteration so experiments stay within daily caps and stop conditions. ## Core Features & Use Cases - Budget Enforcement: Reads loop-budget.md, STATE.md, and the last 24 hours of loop-run-log.md to enforce daily caps, switching to report-only at 80% and exiting at 100%. - Tiered Limits: Enforces L1 constraints (one candidate, zero sub-agents, zero remote compute) and L2 stop rules (three attempts per hypothesis or two identical failure signatures). - Structured Audit Logging: Appends timestamped records to loop-run-log.md with trigger, evidence/run id, candidate count, actions, attempts, escalations, duration, token estimate, verdict, and stop reason, never rewriting prior entries. - Use Case: An AutoResearch agent running overnight experiment loops invokes this Skill each iteration to guarantee it pauses when .loop-paused exists, stays under the daily token cap, and halts a failing hypothesis after repeated identical failures. ## Quick Start Run the loop-budget check at the start and end of each AutoResearch loop iteration to enforce budget limits and append a structured run record.
