The Agent Skills Directory

Command Execution via Manifest Discovery: The skill identifies and may suggest or execute verification commands (e.g., npm test, bazel test) or formatters (e.g., prettier) based on manifests found within the repository (e.g., package.json, Cargo.toml). While this is intended for project documentation and verification, executing commands derived from potentially untrusted repository content is a security consideration that relies on the user's trust in the codebase being analyzed.
Indirect Prompt Injection Surface: The skill processes existing source code, schemas, and documentation files to generate its summaries. There is a potential for content within those files (such as code comments or existing README text) to attempt to influence the agent's behavior or output. The skill mitigates this by instructing subagents to cite exact code symbols and perform a final verification pass against the filesystem.
Subagent Orchestration and Safety Gate: The skill employs a multi-agent architecture to process large directory trees. To prevent unintended resource usage or massive filesystem changes, it includes a mandatory safety gate that pauses execution and requests explicit user approval if more than 15 subdirectories are discovered.

egm-bootstrap