The Agent Skills Directory

[COMMAND_EXECUTION]: The skill requires executing high-privilege commands to manage the infrastructure, including using ssh to access the Colima VM, sudo for host-level kernel modules and service management, and kubectl exec for container interaction.
[EXTERNAL_DOWNLOADS]: Fetches Kubernetes manifests and Helm charts from external sources, including Rancher's GitHub repository for the local-path-provisioner and official Kubernetes-sigs repositories for metrics-server. It also utilizes Helm repositories from LiveKit and Nerkho.
[PROMPT_INJECTION]: The 'Agent Runner' feature introduces a surface for indirect prompt injection by accepting and executing tasks provided via base64-encoded metadata (TASK_PROMPT_B64).
Ingestion points: Request metadata fields including TASK_PROMPT_B64 and VERIFICATION_COMMANDS_B64 in the SKILL.md Agent Runner section.
Boundary markers: No specific delimiters or safety instructions are defined to separate task data from agent instructions.
Capability inventory: Extensive capabilities including cluster-wide kubectl access, docker operations, and ssh access to the host VM (references/operations.md).
Sanitization: The skill relies on base64 encoding for transport but does not specify validation or sanitization of the decoded task content.

k8s