Token usage auditor and budget guard for OpenClaw agents — scans traces for waste (duplicate prompts, oversized context, model overkill), writes budgets, and builds a compact context capsule so the next run starts lean.
