The Agent Skills Directory

[SAFE]: The skill implements strong security boundaries, explicitly forbidding administrative operations such as cluster configuration, node taints, or server-side OSMO configuration changes. All potentially destructive commands, such as workflow cancellation or data deletion, require explicit user confirmation before execution.
[SAFE]: The skill handles sensitive credentials securely. It includes specific instructions to never print secret values back to the user and to prioritize using existing environment variables over manual input. It manages various credential types (registry, generic, and data) through the official OSMO CLI interface.
[SAFE]: The orchestration pattern involves spawning specialized subagents for submission and log reading. While a deterministic detector flagged instructions for the subagent to return control as a concealment attempt, this is a standard design for multi-agent delegation that ensures the primary user interface maintains visibility of the workflow status.
[COMMAND_EXECUTION]: The skill is designed to interact with the local OSMO CLI to manage workflows and resources. It includes explicit rules to never fabricate command output and to use only the CLI's results rather than inferring state from files, ensuring data integrity.
[EXTERNAL_DOWNLOADS]: The skill fetches workflow examples from official repositories (e.g., NVIDIA's OSMO cookbook on GitHub). These are trusted resources, and the agent is instructed to adapt these templates rather than execute them blindly.
[INDIRECT_PROMPT_INJECTION]: The skill processes external data, such as workflow logs, Kubernetes events, and documentation, to perform troubleshooting and submission tasks. The instructions include guidelines for summarizing and digesting this data, which reduces the risk of the agent following malicious instructions embedded in technical logs. This is a standard operational characteristic for a technical assistant skill.

osmo-user