The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructions and code snippets process application-level inputs and outputs to record performance traces. If the underlying application handles untrusted user input, this data enters the agent's tracing context.
Ingestion points: Found in SKILL.md within the safeObserve wrapper and calls to updateActiveObservation, which capture function arguments and return values.
Boundary markers: No specific delimiters or "ignore embedded instructions" warnings are provided for the data being captured in the traces.
Capability inventory: The skill environment allows for Read, Write, and Edit operations on the local file system.
Sanitization: The skill provides a truncateForTrace utility function that limits the length of strings and arrays, which provides basic data minimization and reduces the risk of long-form injection payloads.

langfuse-performance-tuning