The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: Fetches the langsmith SDK and references prompts from the LangChain Hub. Both are official resources from a well-known service.\n- [COMMAND_EXECUTION]: Includes standard environment configuration and package installation commands. No arbitrary execution detected.\n- [PROMPT_INJECTION]: The llm_judge_evaluator example in references/advanced-usage.md demonstrates an indirect prompt injection surface. * Ingestion points: Untrusted model outputs (prediction) and user queries (question) are ingested in references/advanced-usage.md. * Boundary markers: No delimiters or 'ignore' instructions are present in the judge prompt template. * Capability inventory: The judge model's output is used to generate evaluation scores; no dangerous capabilities (like file writing or shell execution) are triggered by this specific workflow. * Sanitization: The example does not include sanitization or escaping of the interpolated inputs.\n- [SAFE]: Recommends the use of environment variables for API key management and highlights standard platform features like file attachment for tracing, following documented SDK practices.

langsmith-observability