The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted pull request data (diffs, descriptions, comments) which could contain malicious instructions. However, it includes explicit safety instructions directing the agent to ignore any embedded commands and uses jq to safely serialize output, mitigating the risk of injection into shell commands. Ingestion points: gh pr view, gh pr diff, gh api comments in SKILL.md. Boundary markers: Includes a dedicated security section instructing the agent to treat PR content as untrusted in SKILL.md. Capability inventory: gh api, jq, mktemp, pi across SKILL.md and scripts/review-train.sh. Sanitization: Employs jq for JSON serialization of the review payload to prevent shell command injection from model output.
[EXTERNAL_DOWNLOADS]: The skill and its companion script reference official tools and repositories including the GitHub CLI (gh), jq, glow, and the pi agent runner. These are well-known, legitimate tools required for the skill's stated purpose.
[COMMAND_EXECUTION]: The skill uses shell commands and a companion bash script to automate the review process. The commands are scoped to the necessary tools (gh, pi, jq) and include validation checks (e.g., verifying model IDs and PR existence) to ensure safe operation.

adversarial-review