The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's primary function involves ingesting and evaluating external data (datasets and cases), which inherently creates a surface for indirect prompt injection attacks where malicious instructions could be embedded in the data being tested.
Ingestion points: The skill reads from local Python/Node.js dataset modules and hosted Logfire datasets via the LogfireAPIClient (SKILL.md).
Boundary markers: There is no specific instruction to use delimiters or guardrail prompts to isolate untrusted dataset content from the agent's instructions during evaluation.
Capability inventory: The skill executes network operations to Logfire APIs and manages dependencies. Notably, it explicitly states that the Python evaluator for arbitrary code execution was removed for security reasons, significantly reducing the impact of potential injections.
Sanitization: The instructions do not describe sanitization or validation logic for the content within dataset inputs or outputs.
[EXTERNAL_DOWNLOADS]: The skill facilitates the installation of official libraries from the vendor's ecosystem.
Evidence: Instructions include installing pydantic-evals[logfire] for Python and logfire / @pydantic/logfire-node for Node.js (Step 4).
[COMMAND_EXECUTION]: The skill uses shell commands for configuration and environment setup.
Evidence: Employs export and unset commands to manage Braintrust compatibility environment variables (Step 1). It also uses uv and npm for installing necessary dependencies (Step 4).

logfire-evals