The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill provides procedures to download official kubectl and minikube binaries from verified Kubernetes release servers (dl.k8s.io and storage.googleapis.com). These tools are essential for the skill's operation and are sourced from trusted infrastructure providers.
[COMMAND_EXECUTION]: The skill facilitates the management of GPU-accelerated container jobs through standard kubectl interactions. It includes pre-submission checks to verify cluster connectivity and resource availability, ensuring reliable job execution.
[INDIRECT_PROMPT_INJECTION]: The skill processes JSON-formatted action requests to generate Kubernetes manifests, presenting a potential ingestion surface for untrusted data.
Ingestion points: The render_action_job.py script reads external action requests and staging maps from local JSON files to define job specifications.
Boundary markers: The workflow incorporates a secret redaction linting step (redact_secrets.py) to inspect manifests before they are applied to the cluster.
Capability inventory: Manifests generated by the skill enable the deployment of containerized training and inference workloads with GPU scheduling and volume mounting capabilities.
Sanitization: The implementation uses rigorous validation logic, including DNS-compliant name checking, binary quantity validation via regex, and path normalization to prevent traversal attacks. Command arguments are handled using shlex and explicit shell-escaping characters are monitored to prevent command injection.

tao-run-on-kubernetes