The Agent Skills Directory

[COMMAND_EXECUTION]: The skill facilitates the execution of diagnostic commands within a sandbox environment using rockcli sandbox <id> exec. Examples include checking system resources (df -h, free -h), listing active processes (ps aux), and verifying environment dependencies (python --version, pip list).
[INDIRECT_PROMPT_INJECTION]: The skill presents an attack surface where the agent processes untrusted data retrieved from sandbox logs and command history, which could contain malicious instructions.
Ingestion points: Diagnostic operations such as rockcli sandbox <id> log search, rockcli sandbox <id> log tail, and rockcli sandbox <id> history ingest content that could contain adversarial prompts.
Boundary markers: The skill does not specify the use of delimiters or instructions to ensure log content is treated strictly as data.
Capability inventory: The skill allows the agent to execute commands in the sandbox, download log files, and search cluster-level event logs.
Sanitization: No automated sanitization is described; the skill relies on high-level operational guidelines like "read-only diagnosis" and the requirement for "gradual confirmation" with the user to manage risks.

rock-debug