The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process and analyze telemetry data from the $mcp_tool_call event, which includes potentially untrusted strings such as $mcp_error_message and $mcp_intent submitted by external clients.
Ingestion points: External telemetry data is ingested through HogQL queries executed via the execute-sql tool, as detailed in SKILL.md and references/queries.md.
Boundary markers: The skill contains a mandatory 'Untrusted data' section that explicitly instructs the agent to treat telemetry content strictly as data to be reported and never as instructions to be followed.
Capability inventory: The agent has capabilities to execute SQL analytics (execute-sql) and author internal reports (emit_report, edit_report).
Sanitization: The instructions provide clear sanitization requirements, including truncating messages, dropping payload echoes, and strictly forbidding the inclusion of raw telemetry, intent strings, or user identifiers in generated pull requests or report summaries.

signals-scout-mcp-tool-calls