The Agent Skills Directory

[COMMAND_EXECUTION]: The skill includes local scripts used for evaluation and testing. The shell script evals/fixtures/check-domain-decisions.sh executes a Python script evals/fixtures/check_domain_decisions.py.
Analysis of the Python script shows it is designed for deterministic final-artifact checks, parsing markdown tables, and validating metrics against hardcoded values.
It does not perform network operations, access sensitive file paths, or use dangerous functions like eval() or exec().
These scripts are part of the standard testing workflow for the skill's own quality assurance.
[INDIRECT_PROMPT_INJECTION]: The skill has an inherent surface for indirect prompt injection as it is designed to ingest and analyze untrusted data, such as "agent goals and instructions", "task traces", and "memory policy" provided by the user.
Ingestion points: Data enters the context via the primary prompt in prompts/ai-agent-testing.md which instructs the agent to process external materials.
Boundary markers: The prompt includes instructions to "Audit the input" and "separate confirmed facts, working assumptions, and open questions," which helps mitigate accidental obedience to instructions embedded within traces.
Capability inventory: The skill primarily performs analysis and reporting. It does not have access to tools that would allow for high-impact actions like network exfiltration or destructive file writes in its default configuration.
Sanitization: While no technical sanitization (escaping) is explicitly implemented, the analytical framework and the requirement to link conclusions to evidence act as a procedural control against following malicious instructions hidden in the data.

ai-agent-testing