The Agent Skills Directory

[COMMAND_EXECUTION] (MEDIUM): The subscribe.py script executes arbitrary local scripts as event handlers via subprocess.run. While this facilitates the skill's primary function as an orchestration engine, it establishes a high-risk capability that could be abused if the configuration files or handler scripts are compromised.
[PROMPT_INJECTION] (LOW): The architecture relies on variable substitution ({{payload.field}}) to pass event data to other agents, creating a surface for Indirect Prompt Injection. \n
Ingestion points: Event queue files at ~/.clawdbot/events/queue/*.json.\n
Boundary markers: Absent; the system uses direct template interpolation.\n
Capability inventory: Subprocess execution of handlers and agents.\n
Sanitization: None; payloads are used raw without escaping or validation.
[DATA_EXFILTRATION] (SAFE): No hardcoded credentials, sensitive file access outside of the skill's own data directory, or unauthorized network activity was detected in the provided code.

agent-protocol