The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/host_bridge.py uses subprocess.run to call an internal Python controller (control.py). This execution is restricted to the specific script using the system's Python interpreter and does not involve a shell, mitigating risks of command injection.
[DATA_EXPOSURE]: The controller script implements a rigorous safe_path function that prevents directory traversal and unauthorized file access by rejecting absolute paths, parent directory references ('..'), and symbolic links. File content is hashed to ensure integrity throughout the task session.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external repository data. It includes defensive instructions in the introspection.md module, training the agent to evaluate such data as evidence rather than allowing it to override its core task or safety instructions.
[OBFUSCATION]: The skill documents an internal reasoning "shorthand" used for token efficiency in long reasoning chains. This notation is explicitly managed as an internal register, and the skill includes an audit script (jspace.py) to verify that this notation does not leak into the agent's final output to the user.

j-space