The Agent Skills Directory

Remote Resource Procurement: The skill is configured to fetch components from the PyTorch GitHub repository to ensure the agent uses the most current version of the CI metrics tools. This represents a dependency on external content during the loading process.
Command Execution and File Permissions: The loader executes chmod +x on the downloaded gcx-wrapper.sh script. This is a standard setup step required to enable the script's functionality, but it is a security consideration as it involves modifying file execution permissions at runtime.
Indirect Prompt Injection Surface: The skill's primary mechanism involves reading and following instructions from a remote source, which represents a structural entry point for external content to influence agent behavior.
Ingestion points: The skill performs a remote fetch via curl for SKILL.md and related scripts.
Boundary markers: No specific delimiters or warnings are used to isolate the fetched instructions; the skill explicitly directs the agent to treat the external content as authoritative.
Capability inventory: The skill environment supports file downloads, permission modifications, and shell command execution.
Sanitization: The skill does not perform validation or filtering of the remote content before it is adopted into the agent's context.

ci-metrics