The Agent Skills Directory

[DYNAMIC_EXECUTION]: The validation scripts (validate.py and validate.ts) dynamically load and execute Python or TypeScript code from local file paths provided by the user. This mechanism is the intended way to load custom grader functions into the validation runner for measurement.- [INDIRECT_PROMPT_INJECTION]: The skill's judge templates interpolate external trace data into prompts for LLM evaluators. The templates include explicit boundary markers ( tags) and system instructions directing the evaluator to treat tagged content as data rather than instructions. Mandatory Evidence Chain: (1) Ingestion points: traces.jsonl, goldens.jsonl. (2) Boundary markers: Present in judge templates. (3) Capability: Network operations (LLM API calls) and local script execution. (4) Sanitization: Structural serialization (json.dumps) is used for complex data, though raw strings are interpolated within delimiters.- [COMMAND_EXECUTION]: The setup_workspace.py script performs local file system operations, including directory creation and writing executable template files (Python and TypeScript) to the local project workspace to initialize the evaluation environment.- [EXTERNAL_DOWNLOADS]: The documentation and generated templates reference official SDKs and documentation from well-known providers including Anthropic, OpenAI, and DeepEval. These represent standard dependencies required for the skill's core functionality.

eval-grade