The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides patterns for implementing code verification servers that use subprocess.run via Ray to compile and execute model-generated code. This behavior is the primary purpose of the benchmarking environment.
[EXTERNAL_DOWNLOADS]: Instructions describe auto-installing external tools using package managers like brew and custom shell scripts, as well as fetching datasets from a GitLab registry using provided CLI tools like ng_prepare_data.
[CREDENTIALS_UNSAFE]: The documentation guides users on configuring env.yaml with MLflow tracking tokens and GitLab API keys for dataset management, which is a standard approach for secret management in development workflows.
[PROMPT_INJECTION]: The skill establishes an indirect prompt injection surface (Category 8) by ingesting model-generated text for verification. It includes mitigation strategies such as stripping reasoning blocks (<think>) and using markdown-aware code extraction to isolate the executable content.

add-benchmark