The Agent Skills Directory

[SAFE]: The skill is designed to improve the auditability and security of multi-agent workflows by centralizing communications through a controlled 'switch' (message broker).
[COMMAND_EXECUTION]: The skill executes a local Python utility script agent_switch.py to provide its core functionality. These commands are used for serving the communication endpoint, sending messages, and performing audits of the message history. Evidence:
python <plugin>/scripts/agent_switch.py serve --port 7070 --log .livepowers/comms.jsonl
python <plugin>/scripts/agent_switch.py audit
[DATA_EXPOSURE]: The communication service operates on the local loopback address (127.0.0.1:7070) and logs data to a local file (.livepowers/comms.jsonl). No unauthorized network exfiltration or sensitive data exposure was detected.
[INDIRECT_PROMPT_INJECTION]: The skill manages data ingestion from various agents. It includes built-in mitigations such as body size limits (SWITCH_MAX_BODY) and keyword-based filtering (SWITCH_BLOCK_WORDS) to reduce the risk of malicious payloads being propagated through the communication chain. Evidence:
Ingestion points: The body field in the message envelope and the sidecar-log functionality.
Boundary markers: JSON envelope structure and trust domain definitions.
Capability: Local script execution for logging and auditing.
Sanitization: Rejection of messages based on length and block words.

auditable-agent-comms