The Agent Skills Directory

[SAFE]: The skill defines a methodical 'Pattern Mining' workflow that uses repository-anchored exemplars and a separate holdout to validate abstractions. It uses standard local tools like ripgrep and a bash script that leverages jq for schema validation. No high-risk behaviors such as network exfiltration, credential harvesting, or remote code execution were identified.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted code implementations from the repository to extract recurring patterns. This creates a surface for indirect prompt injection, but the risk is mitigated by the structured workflow requirements. (1) Ingestion points: Repository code implementations (exemplars) and search results. (2) Boundary markers: No explicit prompt boundaries are defined for exemplar ingestion. (3) Capability inventory: Writing evidence artifacts to local scratch directories and executing a local validation script. (4) Sanitization: Employs a specific bash and jq script to validate the logical consistency and schema of output artifacts before promotion.

pattern-mining