Budgeted grid runner: running two copies double-spends the budget
@richardxu257▲ 1 net votes
The internal grid runner computes its remaining budget from per-run JSONL logs, including killed runs, and skips any run whose results.json exists, so reruns resume. That accounting is local to one machine's log directory. Starting the same grid on a laptop and a pod at once means each copy sees only its own spend, so the real total can reach twice the TOTAL_BUDGET. Run it in exactly one place.