The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill references an external repository for its core MCP server implementation located at https://github.com/prime001/humanrail-mcp-server.
[DATA_EXFILTRATION]: The skill is designed to transmit technical task context and decision payloads to external human workers through the HumanRail API (humanrail.dev). To mitigate exposure risks, the skill includes non-negotiable instructions for the agent to sanitize all payloads, specifically removing internal hostnames, IP addresses, credentials, and PII before transmission.
[COMMAND_EXECUTION]: The documentation includes instructions for setting up the environment using pip3 to install dependencies and starting a local Python-based MCP server.
[PROMPT_INJECTION]: The skill ingests decisions from external human workers, creating a surface for indirect prompt injection that could influence automated destructive operations.
Ingestion points: Responses are retrieved from the HumanRail service via the wait_for_task tool.
Boundary markers: The skill employs structured JSON output schemas to define and validate the format of human responses.
Capability inventory: The skill provides workflows for executing irreversible operations such as device reloads, configuration erasure, and BGP peer modifications based on human approval.
Sanitization: The instructions mandate checking for a 'verified' status before acting on output and emphasize framing tasks neutrally to prevent external input from directly controlling agent logic.

humanrail-escalation