The Agent Skills Directory

[Command Execution]: The skill utilizes npm and npx to execute validation scripts and the @microsoft/vally-cli tool. Executing shell commands is a powerful capability that requires careful management; in this context, it is employed for standard development tasks such as running test suites and grading agent outcomes.\n- [Dynamic Module Loading]: The testing framework supports custom grader plugins by loading TypeScript modules at runtime. Loading code from paths is a security consideration as it can lead to arbitrary code execution if not properly restricted, but here it serves as a documented extension mechanism for developers to define custom grading logic.\n- [Data Ingestion Surface]: The skill processes YAML-based evaluation suites that may contain untrusted prompt data, which introduces a surface for indirect prompt injection.\n
Ingestion points: YAML evaluation files located in the evals/ directory.\n
Boundary markers: Boundary enforcement is not detailed in the skill instructions and relies on the vally runner's internal implementation.\n
Capability inventory: The skill facilitates running prompts against LLMs and writing execution results to the file system.\n
Sanitization: Convention checks are performed by a validate-stimulus script to ensure data follows expected schemas.\n- [External Resource Integration]: The skill references documentation and tools hosted on official vendor platforms. While downloading external resources can pose a risk of supply chain interference, these references target established and verified sources consistent with the skill's purpose.

vally-eval