The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests data from external polled sources but implements multiple layers of defense to prevent this data from being interpreted as instructions by the AI. It wraps messages in an XML-like envelope, escapes characters in envelope attributes, and specifically sanitizes event bodies to ensure closing tags cannot be used to break out of the message context.\n- [PRIVILEGE_ESCALATION]: The daemon enforces strict file permissions by creating its runtime directory with mode 0700 and setting its sockets and state files to mode 0600. This ensures that the communication channel and event data are only accessible to the user running the session, preventing local unauthorized access.\n- [SAFE]: The system uses session-specific identifiers (combining process ID and a start token) to manage its files, which prevents data collisions or leakage between different sessions or in scenarios where process IDs are reused.\n- [SAFE]: All communication occurs via local Unix Domain Sockets, and the skill follows best practices for background task management and state persistence.

session-events