The Agent Skills Directory

[COMMAND_EXECUTION]: The skill generates shell functions that interpolate variables like $jid and $host directly into execution strings, creating a command injection surface if these variables are poisoned in the registry.
[REMOTE_CODE_EXECUTION]: The skill uses SSH to execute commands on remote hosts using potentially unsanitized variables like cluster hostnames and job IDs derived from session registries.
[DATA_EXFILTRATION]: The skill performs network operations via SSH to connect to remote clusters and accesses local session-specific job registries in .claude/agents/.
[PROMPT_INJECTION]: The skill processes untrusted data from external command outputs and a local registry without sanitization or boundary markers.
Ingestion points: .claude/agents/<session_id>/active_jobs.json and output from nel status and sacct.
Boundary markers: Absent.
Capability inventory: Remote command execution via SSH, local shell execution, and file system writes.
Sanitization: Absent; input values are used directly in shell commands and reporting logic.

monitor