reaatech/context-window-planner

These packages provide a deterministic engine for managing LLM context windows by deciding which data to include, summarize, or drop based on token budgets. They solve the problem of prompt overflow in complex agentic workflows where input data frequently exceeds model limits. The system uses a modular builder pattern that allows you to compose custom packing strategies, tokenizer adapters, and typed context primitives into a single, framework-agnostic execution plan.