The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from pod logs and Kubernetes cluster state. It mitigates injection risks by wrapping all external input in nonce-delimited tags and sanitizing control characters to prevent the agent from executing instructions embedded in logs.
Ingestion points: Data enters via pod logs (kubectl logs), resource descriptions (kubectl describe), and helm values.
Boundary markers: A wrap function (defined in references/untrusted-output.md) uses randomized nonces to delimit data blocks.
Capability inventory: The skill uses kubectl, helm, and observe CLI tools for cluster interactions.
Sanitization: Employs sed and tr to strip ANSI escape sequences and non-printable characters that could be used for terminal redressing.
[DYNAMIC_EXECUTION]: Uses local python3 one-liners for JSON processing and data masking. Specifically, it decodes and truncates Kubernetes secrets to ensure only structural metadata (length, prefixes) is analyzed by the agent, avoiding full credential exposure.
[COMMAND_EXECUTION]: Executes administrative commands (kubectl, helm, observe) to gather diagnostics. These are explicitly designed to be run by the user in their local shell rather than the agent's internal environment to maintain security boundaries.
[DATA_EXPOSURE_AND_EXFILTRATION]: Accesses sensitive Kubernetes secrets and cluster configuration. The skill proactively masks this data, preventing full ingestion tokens or sensitive environmental variables from being exposed to the conversation context.

debug-k8s-collection