The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted interaction data from LLM responses and tool outputs, creating an ingestion surface for indirect prompt injection findings.
Ingestion points: The messages list in llm_tracing.py, args and kwargs in the traced_tool decorator, and input to run_agent in langfuse_integration.py.
Boundary markers: Documentation explicitly advises omitting raw prompts and credentials from telemetry exports to mitigate data exposure risks.
Capability inventory: Includes network communication via httpx.post in context_propagation.py and potential file-write operations through standard Python logging handlers.
Sanitization: Provides a PIIRedactingSpanProcessor and PIIRedactingFilter utilizing regex patterns to identify and remove sensitive data like SSNs, credit cards, and API keys before transmission.
[DATA_EXFILTRATION]: The skill implements telemetry export to external observability platforms and generic HTTP endpoints.
Evidence: Configures OTLPSpanExporter for trace data, integrates with Langfuse via https://cloud.langfuse.com, and supports Helicone as a logging proxy via https://oai.helicone.ai/v1.
Context: A call_downstream_service function in context_propagation.py transmits trace context and payloads to a configurable URL using the httpx library.

agent-observability