LLM token budget architect. Allocates context capacity across system prompt, RAG retrieval chunks, few-shots, conversation history, and reasoning tokens to prevent catastrophic truncation.