SkillTrace proposes a provenance-auditing framework for reusable LLM-agent skills, where copying may appear across text, code, tools, and workflow structure.
The paper argues that agent skills are not ordinary software packages. They can include metadata, natural-language instructions, code fragments, references, and operational procedures. Existing clone detectors usually focus on source code or whole-package similarity, which can miss reuse that preserves only one part of a skill.
SkillTrace extracts three traces: expression, implementation, and operational structure. The operational trace is represented as a Skill Operational Graph that captures activation, procedure, and resource flow. An LLM helps extract that operational trace once during ingestion, while audit-time comparisons are deterministic and calibrated against strict negatives.
The work matters as agent-skill marketplaces emerge. Provenance disputes may not be about identical files; they may involve reused workflows or instruction patterns. Auditing systems will need to show which part of a skill supports a reuse claim.