The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes scanner output fields such as scanner_output_summary, scanner_output_tags, and scanner_output_reasoning. These fields contain content generated by other AI models based on untrusted end-user recording data, creating an indirect prompt injection surface.
Ingestion points: Data enters the agent context via the execute-sql and vision-scanners-observations-list tools, specifically the scanner_output_* properties defined in SKILL.md.
Boundary markers: The skill includes a dedicated section "Untrusted data — scanner output is LLM text over user content" which explicitly instructs the agent to treat this data as untrusted snippets and never as instructions.
Capability inventory: The skill has the ability to query event data (execute-sql), emit/edit reports (emit_report, edit_report), and potentially modify scanner configurations (vision-scanners-update) if write permissions are granted.
Sanitization: The instructions mandate using sanitized/slugified identifiers for memory storage and truncating untrusted prose snippets in reports.
[COMMAND_EXECUTION]: The skill performs SQL queries using the execute-sql tool against the events table to aggregate observation data and check for throughput issues. This is a core functional requirement of the analytics scout for processing telemetry data.

signals-scout-replay-vision