The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill instructions provide examples for executing code at runtime using exec() and eval() functions.
Evidence: references/end-to-end-examples.md and references/task-and-eval-patterns.md demonstrate executing row.code and row.test_expression derived from external dataset rows.
Context: This is designed for evaluating coding problem responses, but it lacks execution sandboxing in the provided examples.
[REMOTE_CODE_EXECUTION]: The skill implements a workflow where data retrieved from a remote service is subsequently executed locally.
Evidence: Datasets are fetched via ze.Dataset.pull() and the contents are passed to exec() in the provided code examples in references/end-to-end-examples.md.
[INDIRECT_PROMPT_INJECTION]: The skill exposes an attack surface for indirect prompt injection by processing external dataset content with high-privilege capabilities.
Ingestion points: Data enters the environment through ze.Dataset.pull() as seen in references/end-to-end-examples.md.
Boundary markers: There are no explicit delimiters or warnings provided to the model to ignore potentially malicious instructions within the dataset content.
Capability inventory: The skill utilizes exec() and eval() for code execution and makes network requests via the openai and zeroeval SDKs.
Sanitization: The examples do not implement sanitization, filtering, or sandboxing of the code being executed beyond basic error handling.
[EXTERNAL_DOWNLOADS]: The skill performs downloads from established vendor and service repositories.
Evidence: Fetches the ZeroEval SDK using pip install zeroeval and pulls datasets from the ZeroEval platform.
Evidence: Includes instructions for cloning repositories using git clone for versioning scripts.
[COMMAND_EXECUTION]: The skill uses system commands for environment setup and CLI-based management.
Evidence: Documents the use of pip install, export, and various zeroeval CLI commands (e.g., zeroeval evals list, zeroeval setup) in SKILL.md.

run-evals