The Agent Skills Directory

[SAFE]: The skill uses local environment variables and scripts (${CLAUDE_PLUGIN_ROOT}/scripts/argument-graph.py) to manage an argument graph. These operations are performed locally and do not involve remote network calls or sensitive data access.
[SAFE]: The hooks section utilizes a PreToolUse command to initialize the environment (auto-init.sh). This is a standard initialization pattern and targets a local script within the plugin root.
[PROMPT_INJECTION]: The skill processes user-provided data via the $ARGUMENTS variable to populate an argument graph. While the instructions provide structural boundaries by using double quotes for shell command arguments, the ingestion of untrusted user data into agent workflows represents a standard indirect prompt injection surface. The impact is limited to the local graph state and the specific analysis session.
[DATA_EXFILTRATION]: Persistence is explicitly stated to occur in /tmp, which is appropriate for temporary local state sharing between related skills. No evidence of data exfiltration to external domains was found.

steelman