The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes administrative commands using kubectl, helm, and terraform to view logs, restart deployments, and manage cluster configurations. Tools are constrained via the allowed-tools frontmatter to specific kubectl and curl patterns.
[EXTERNAL_DOWNLOADS]: Fetches diagnostic data and policy information from api.cast.ai and status.cast.ai. It also pulls the curlimages/curl Docker image for in-cluster debugging. These represent well-known services associated with the skill's intended purpose.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a vulnerability surface by ingesting untrusted data from the environment.
Ingestion points: Reads pod logs (kubectl logs) and cluster events (kubectl get events) in SKILL.md sections 1, 2, 7, and 10.
Boundary markers: The skill lacks explicit boundary markers for ingested data but includes instructions to "collect the minimum redacted evidence needed".
Capability inventory: Uses Bash(kubectl:*) and Bash(curl:*) which can be used to modify cluster state or exfiltrate data if the agent is influenced by malicious log content.
Sanitization: Instructions explicitly mandate redacting API keys, raw secrets, and complete workload logs before outputting incident records.

castai-common-errors