SkillCODER: Towards Auditing and Attribution of Copyright Infringement in LLM Agent Skills
Abstract
Software artifact for SkillCODER: Towards Auditing and Attribution of Copyright Infringement in LLM Agent Skills, accepted at ACM CCS 2026 (Round 2). This is the September 13, 2026 repackaging of v1.0.0, based on source commit 22eccdb452b2d90e475277ebb34ee584e5f617fe. It includes the current English and Chinese documentation, the paper-aligned title, DOI badges, and Zenodo metadata. The software version remains v1.0.0. The original September 3 archive remains available at https://doi.org/10.5281/zenodo.22277502. SkillCODER implements package-wide semantic parsing, private-key buyer codebooks, model-assisted watermark embedding, matched active/decoy/normal probes, error-correcting decoding, and black-box buyer attribution. It supports direct, LangChain, and CAMEL-compatible runtimes. The ZIP contains the complete repository snapshot, including source code, tests, documentation, examples, and pinned research inputs. Third-party Skills retain their upstream licenses, as documented in the README and dataset manifest. Validation on Python 3.13: 196 tests passed and 3 skipped; type checking passed for 15 source files; Python source distribution and wheel built successfully. GitHub release: https://github.com/EonHao/SkillCODER/releases/tag/v1.0.0
