The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data sources including git diffs, PR bodies, and progress logs. It explicitly identifies these as untrusted and provides defensive instructions to detect and reject adversarial attempts to override agent behavior within that content (e.g., instructions telling the agent to 'approve' or 'ignore').
[COMMAND_EXECUTION]: The skill directs the agent to execute npm run verify. This is a standard development workflow command used here to validate code changes locally. While it runs code from the repository being audited, this behavior is integral to the skill's stated purpose as a code validator.
[CREDENTIALS_UNSAFE]: The skill includes a proactive defensive check for hardcoded secrets. It instructs the agent to scan diffs and metadata for credentials and explicitly forbids quoting those sensitive values in final reports to prevent further exposure.

harness-validator