A background service that enforces per-session token budgets on OpenClaw agents, auto-prunes context when limits approach, and reports cost per task

OpenClaw retains all conversation history by default, growing from 5K tokens in round 1 to 150K by round 10. Users routinely hit $50-100/day in API costs running stock configurations. Memory consumption climbs to 28GB by day three. Existing tools show token usage after the fact or add a kill-switch, but nothing actively manages the context window in real time by pruning low-value messages, enforcing per-session budgets, and attributing cost to individual tasks. This service sits between OpenClaw and the LLM provider, intercepts every request, enforces configurable token budgets, and auto-prunes context using a relevance scorer before the request hits the API.