Shrink LLM context windows 40-80% (input-dependent, see benchmarks) by removing noise, redundancy, and long-tail detail — zero-dependency Python pipeline with tiktoken counting, an extractive summarizer, and a security-scan brief generator.
A pre-execution approval gate for AI agents: preview the plan, see per-step risk, approve/reject/edit before anything runs. Framework-agnostic, zero core deps.
