The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external Matrix chat rooms, creating a surface for indirect prompt injection where chat participants could attempt to override agent instructions.
Ingestion points: Timeline events read via matrix-read-e2ee.py and real-time logs processed by matrix-watchd.py.
Boundary markers: The skill includes explicit instructions in SKILL.md and references/agent-governance.md informing the agent that only its principal governs its functions and to ignore instructions embedded in chat rooms.
Capability inventory: The agent possesses capabilities to manage rooms (invite, power levels), send/edit messages, and execute Bash tools.
Sanitization: The documentation instructs the agent to treat chat logs as an event stream rather than a narrative, reducing the risk of accidental obedience.
[COMMAND_EXECUTION]: The matrix-doctor.py script utilizes subprocess.run() to execute package management commands (pip/uv). These calls are constrained to hardcoded dependency installation tasks necessary for the skill's functionality.
[EXTERNAL_DOWNLOADS]: The matrix-doctor.py script facilitates the download of the matrix-nio library from the official Python Package Index (PyPI). This is a standard practice for managing dependencies and targets a well-known, trusted library for Matrix communication.

matrix-communication