The Agent Skills Directory

[Command Execution]: The skill utilizes standard kubectl and az CLI commands for cluster and cloud resource diagnostics. Evidence: The skill instructs the agent to run commands such as kubectl get, kubectl describe, and az aks show in several reference files (e.g., references/gpu-observability.md and references/gpu-scheduling.md). The skill includes a Host Capability Gate in SKILL.md to ensure that these commands are only executed if authorized by the host environment.
[Indirect Prompt Injection]: The skill processes potentially untrusted data from cluster logs and resource statuses, which represents a potential surface for indirect prompt injection. (1) Ingestion points: Data enters the context via kubectl logs and kubectl get workspace commands. (2) Boundary markers: No explicit delimiters or warnings for embedded instructions are used in the prompt templates. (3) Capability inventory: The skill utilizes subprocess calls for kubectl and az but restricts their use to read-only diagnostic tasks. (4) Sanitization: No explicit sanitization or filtering of external cluster data is performed before analysis. Context: The risk is minimal as the skill is designed for diagnostics and consistently emphasizes the requirement for explicit authorization for any resource mutations.

aks-gpu-inference