The Agent Skills Directory

[COMMAND_EXECUTION]: The skill performs several shell operations including temporary directory creation via mktemp, environment setup, and executing validation commands such as pnpm spec and pnpm test. It also uses git and the gh CLI to interact with repositories and pull requests.
[REMOTE_CODE_EXECUTION]: The skill spawns a nested agent using the claude CLI to execute the code and instructions of the skill being tested. This child process is granted high-privilege tools, including Bash, Read, and Write, which are necessary for testing but represent a broad execution surface.
[INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection through several ingestion points:
Ingestion points: Processes user-supplied task descriptions via the --task flag and fetches external tool documentation from the Scenario developer portal.
Boundary markers: Uses a specific "contract" file (skill-tester.md) to instruct the sub-agent to ignore repository context, though this is a soft constraint.
Capability inventory: The sub-agent has full Bash access, Read/Write capabilities, and the ability to execute package scripts.
Sanitization: No explicit sanitization or escaping of the user-supplied task is performed before it is written to the child agent's task file.
[EXTERNAL_DOWNLOADS]: Fetches documentation from the vendor's official domain at https://mcp.scenario.com/docs/tools to ensure the validator uses up-to-date tool definitions.

skills-validate