The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes the herdr CLI and a local Python script fleet.py to manage agent states. All subprocess calls within fleet.py use list-based arguments (e.g., herdr agent list and git status), following security best practices to prevent command injection.\n- [INDIRECT_PROMPT_INJECTION]: The skill handles data from field agents. Ingestion point: agent output is read via herdr agent read (SKILL.md). Boundary markers: None. Capability inventory: the skill can close panes via herdr pane close (fleet.py). Sanitization: the instructions in SKILL.md mandate manual verification, judging the work, and running tests rather than trusting agent summaries, which mitigates the risk of executing instructions embedded in the data.\n- [SAFE]: The skill maintains local state in the ~/.claude/fleet/ directory and does not perform any network operations to non-whitelisted domains or exfiltrate sensitive credentials. No indicators of obfuscation, privilege escalation, or unauthorized persistence were found.

fleet-audit