Code2Skill is a fully automated pipeline that mines source code repositories to produce implementation-anchored agent skills, verifying each one through source-body-blind reconstruction and source-aware comparison rather than relying on trajectory replay or unverified documentation. Applied to nearly 20,000 GitHub repositories, it produces CodeSkillBank, a bank of over one million verified skill records with provenance and source-evidence metadata. Across 72 evaluations spanning nine model settings and eight benchmarks, models retrieving CodeSkillBank skills improve 11.7% on average over matched baselines and outperform trajectory-derived skill banks on every shared benchmark.
