The Agent Skills Directory

[SAFE]: The skill maintains a strict read-only boundary. All instructions and the core analysis script ('scripts/analyze_deploy_evidence.py') are designed to prohibit mutating operations such as 'terraform apply', 'terraform destroy', 'schemachange deploy', or Snowflake DDL/DML.
[SAFE]: The deterministic evidence classifier ('scripts/analyze_deploy_evidence.py') is implemented as a 'pure' script. It uses only Python standard libraries, performs no network operations, and contains no filesystem write or subprocess execution capabilities.
[SAFE]: The skill implements automated data sanitization and secret protection. The '_reject_secrets' function in the analysis script recursively checks for and rejects credential-shaped strings (passwords, tokens, private keys) in the input JSON evidence to prevent accidental exposure.
[SAFE]: The skill enforces cryptographic data integrity. It requires a SHA-256 digest from an independent trusted source to verify that the evidence has not been tampered with or recomputed locally by an untrusted entity.
[SAFE]: High-quality path sanitization is employed in the analysis script to prevent directory traversal attacks, ensuring that input files are relative and remain within the intended execution context.
[SAFE]: The deterministic detector hit in 'eval-spec.yaml' regarding prompt injection is a false positive. The string 'Ignore previous instructions' is contained within an evaluation test case intended to verify the model's resistance to such attacks, rather than being an instruction for the model's actual behavior.

snowflake-deploy-medic