The Agent Skills Directory

[SAFE]: The skill's primary objective is defensive, providing instructions to install industry-standard security tools (Gitleaks, Semgrep, OSV-Scanner) to prevent common AI-generated security regressions.
[COMMAND_EXECUTION]: Phase 0 utilizes standard shell commands (cat, ls, grep, rg) to conduct a read-only audit of the repository's current configuration, including package.json and CI workflow directories.
[SAFE]: Phase 5 incorporates human-in-the-loop safeguards, explicitly instructing the agent to stop for human review before turning on any features that block merges or affect branch protection.
[SAFE]: The testing instructions in Phase 4 include a clear warning to never commit planted test secrets, ensuring that verification activities do not inadvertently introduce vulnerabilities.
[PROMPT_INJECTION]: The skill processes external repository data (such as package.json and existing CI workflows) which constitutes an indirect prompt injection surface.
- Ingestion points: Reads package.json, .github/workflows/, .cursor/rules/, and AGENTS.md (SKILL.md).
- Boundary markers: Not explicitly defined for the audited data.
- Capability inventory: Capability to write to project configuration files, CI workflows, and agent policy files (SKILL.md).
- Sanitization: The skill relies on the agent's logic to interpret the audited files to identify security gaps.

enhance-agent-guardrails