The Agent Skills Directory

Indirect Prompt Injection Surface: The skill fetches review comments and data from external entities, specifically the copilot-pull-request-reviewer[bot], using the GitHub API (gh api). This information is then processed by the agent to classify findings and implement code changes. Because this data originates from outside the immediate instruction set, it serves as an ingestion point for potentially untrusted content. Since the agent has the capability to modify source code and push updates to the repository, instructions embedded within these external comments could theoretically influence the agent's actions. Currently, no explicit boundary markers or sanitization steps are defined for this ingestion process.
Command Execution: To fulfill its role in managing code lifecycle, the skill executes a variety of shell commands using tools like git, gh, and uv. These operations include checking out branches, performing rebases, committing changes, and pushing to remote repositories. These commands are a core part of the automated convergence loop described in the instructions.
Dynamic Agent Execution: The skill employs a modular design where it dynamically invokes the apm-review-panel for code analysis. This is achieved by either using the platform's native tool-calling mechanism or by loading the instructions and schemas of sibling skills directly from the file system. This method of task fan-out is a standard pattern within the agent's composition framework.

shepherd-driver