The Agent Skills Directory

[COMMAND_EXECUTION]: The skill includes internal utility scripts (validate_skill.py and smoke_test.py) that use subprocess.run to execute other bundled Python scripts for validation and testing. These operations are restricted to the skill's local directory and do not involve shell invocation of user-supplied strings, presenting no risk to the host environment.
[PROMPT_INJECTION]: Static detections regarding action concealment were analyzed and found to be benign. The instructions in references/EDGE_CASES.md and references/SAFETY_SEMANTIC_TRIAGE.md advise users in potentially abusive situations to use safe devices and keep safety plans private from monitors or abusers. This is standard safety protocol for domestic violence support and is not an attempt to hide AI behavior or bypass security controls.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user input for coaching purposes. It implements a robust safety triage system described in references/SAFETY_SEMANTIC_TRIAGE.md which identifies dangerous contexts (e.g., violence, coercion) and triggers hard redirects to crisis resources. This layered defense minimizes the risk of the agent obeying malicious instructions embedded in user messages.

relationship-science-coach