The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill describes an automated mechanism for extracting 'Golden Datasets' from user-flagged traces and corrections to be stored as test cases and 'Automatically run against the agent'. This establishes a potential indirect prompt injection surface.
Ingestion points: User-provided trace corrections and flagged outputs are ingested into the system (SKILL.md).
Boundary markers: The skill does not specify the use of delimiters or 'ignore' instructions for the generated test case context.
Capability inventory: The skill notes that these generated test cases are automatically executed against the agent during system updates (SKILL.md).
Sanitization: While the skill recommends sanitizing PII, it does not explicitly detail measures to escape or validate instructions within user corrections before they are interpolated into test prompts.
[SAFE]: The skill provides high-level architectural and operational guidelines for implementing observability, focusing on traceability, metrics, and loop detection without introducing executable code or dangerous commands.
[SAFE]: External integrations refer to established and reputable services like OpenTelemetry, Langfuse, and Helicone, following secure data propagation patterns.
[SAFE]: The skill explicitly includes a best practice for data privacy, instructing the sanitization of Personally Identifiable Information (PII) before data is sent to external observability backends.

agent-observability