The Agent Skills Directory

[Command Execution]: The skill uses subprocess.run within scripts/owner_touch_gate.py to execute git commands. While automated command execution warrants review, the script uses a resolved, cached path for the git executable and disables shell execution to prevent common injection vectors. Contextually, this is necessary for the skill's purpose of managing repository state.
[Automated Tool Interaction]: The skill makes extensive use of the gh (GitHub) and git CLI tools to fetch PR reviews, push code, and watch CI status. These operations are governed by 'plan + execute + verify' patterns and hard limits (caps) to ensure predictable behavior and prevent infinite loops.
[Dependency Probing]: The skill includes a preflight phase that checks for the existence of sibling skills (apm-review-panel, pr-description-skill) on the local disk before proceeding. This is a safety measure to prevent execution when the required environment is incomplete.
[Indirect Prompt Injection Surface]: The skill ingests untrusted data from copilot-pull-request-reviewer[bot] comments and pull request bodies. It mitigates potential instruction injection by applying a manual classification step (LEGIT/NOT-LEGIT) and a strict 'fold-vs-defer' rubric, ensuring the agent only acts on content that aligns with the PR's original stated scope.

shepherd-driver