The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides instructions for the agent to use the herdr CLI to execute shell commands in different terminal panes via herdr pane run. This allows the agent to control processes and environments outside its immediate shell context.
[DATA_EXFILTRATION]: The herdr pane read command gives the agent the ability to inspect the scrollback and current output of any other active terminal pane in the workspace. This exposes data from separate processes, potentially including logs, configuration data, or credentials displayed in neighbor panes.
[DYNAMIC_EXECUTION]: The skill documents the use of shell pipes and inline Python scripts (e.g., python3 -c 'import sys,json; ...') to dynamically parse JSON output from the herdr CLI and extract identifiers used for subsequent command execution and agent spawning.
[INDIRECT_PROMPT_INJECTION]: The ability to read output from other terminal panes (herdr pane read) introduces a surface for indirect prompt injection. Instructions or malicious content printed to a terminal by a separate process could be ingested by the agent and interpreted as instructions.
Ingestion points: Terminal scrollback and recent output from neighbor panes via the herdr pane read command (SKILL.md).
Boundary markers: No specific delimiters or "ignore embedded instructions" warnings are provided to help the agent distinguish between its own instructions and the data ingested from other terminal panes.
Capability inventory: The agent has full command execution capabilities, pane management, and the ability to spawn sub-agents across the terminal multiplexer (SKILL.md).
Sanitization: The skill does not specify any sanitization, filtering, or validation of the ingested terminal text before it is processed by the agent.

herdr