SkillZip is a graph compression framework for agent skill libraries that operates at the section level rather than treating each skill as an indivisible unit. It performs contract-preserving compression by identifying recurring procedural patterns and replacing them with reversible macros, while maintaining execution dependencies, verifier reachability, and source traceability. The approach lets agents retrieve smaller, task-specific executable contexts from large skill libraries without losing the procedural guarantees needed for reliable execution, achieving up to 3.46x compression while preserving 99.2% of dependencies.
