Enforce strict token budgets, context limits, and output truncation using LiteLLM to prevent runaway cost spikes in multi-turn autonomous agent loops.
