Compress the skill graph, keep every route

SkillZip Pro treats an agent skill as a progressively loaded directory, not a single prompt. The compressor moves repeated guidance across the root, references, and subskills while preserving the routes that make each entry reachable. 12

The product problem is bundle growth. Evolving skills can repeat instructions across files, inflate shipped and per-run context, and break when compression removes a path-specific detail. SkillZip Pro uses typed resource contracts, removal witnesses, and a disk audit before publication. In the paper's controlled evaluation, pooled success was 0.480 for SkillZip Pro versus 0.470 for the uncompressed Evolved Bundle, with a 95% bootstrap interval of [-0.029, +0.059]. Shipped tokens fell 19.7%, average per-run tokens fell 14.3%, and routing fidelity remained 1.000. 2

Start with a shadow path beside the current loader: inventory opened files and branch guards, declare private/public/conditional contracts, require a witness for each removal, and compare success, route fidelity, load recall, tokens, latency, and cost. Keep the original bundle live as rollback. The paper names no public code, dataset, or model release, so the first build is an implementation sketch rather than a package to install. 2