The Agent Skills Directory

[COMMAND_EXECUTION]: The skill makes extensive use of the herdr CLI through Python and Bash scripts. These scripts automate tasks like pane splitting, resizing, and starting agent processes. The execution is scoped to the platform's own tool and uses sanitized inputs such as pane IDs and agent names.
[INDIRECT_PROMPT_INJECTION]: The skill processes output from other agents in the fleet. It mitigates injection risks by relaying only requested deltas rather than full transcripts, checking for blocked or unsafe agent states before communication, and providing structured handoff templates that maintain orchestrator context without re-executing untrusted history.
[DYNAMIC_EXECUTION]: The launch_profile.py script dynamically assembles command-line arguments for launching sub-agents. It strictly filters these flags through a per-harness allowlist (e.g., for Claude, Pi, and Codex), ensuring that only recognized setup flags are inherited and preventing arbitrary flag injection.
[DATA_EXFILTRATION]: No unauthorized network operations or access to sensitive local credentials were found. All communication flows through the established herdr CLI and server infrastructure as intended by the skill's design.
[OBFUSCATION]: The skill's code, including its Bash and Python components, is written in plain text with no evidence of Base64 encoding, zero-width character embedding, or other techniques used to hide malicious intent.

herdr-agent