The Agent Skills Directory

[PROMPT_INJECTION]: The skill ingests production trace data which presents an indirect prompt injection surface. It contains explicit instructions to mitigate this risk by treating content as data rather than instructions.
Ingestion points: The get_observation_details_with_content and get_evaluation_task_with_content tools in references/mcp.md ingest untrusted prompt and completion text from production logs.
Boundary markers: The skill mandates 'defanging' boundary tokens (e.g., converting <input> to < input >) to prevent the agent from misinterpreting trace data as system control tokens.
Capability inventory: The skill is primarily read-only but includes logic to hand off identified gaps to the generate-eval skill.
Sanitization: The instructions require scrubbing PII (emails, phone numbers, SSNs) and restrict all data queries to a 72-hour window.
[EXTERNAL_DOWNLOADS]: The skill connects to https://mcp.observability.progress.com/mcp for data retrieval. This is a well-known service endpoint belonging to the vendor (Progress Software/Telerik).
[COMMAND_EXECUTION]: The references/mcp.md file provides a curl command template for users to verify connectivity to the MCP server using their API key.

coverage-gaps