The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted model outputs and benchmark datasets, creating a surface for indirect prompt injection.\n
Ingestion points: The tool ingests model responses and task data from numerous harnesses (e.g., MMLU, GSM8K) as described in SKILL.md and references/custom-benchmarks.md.\n
Boundary markers: The skill relies on containerization and harness-specific logic, but does not specify universal boundary markers for all processed data.\n
Capability inventory: The SDK executes shell commands for job management and performs network operations for API interaction, as detailed in references/execution-backends.md.\n
Sanitization: While output parsing is defined in references/custom-benchmarks.md, there is no evidence of exhaustive sanitization for model-generated content.\n- [DYNAMIC_EXECUTION]: The system allows for dynamic loading of custom interceptor code from arbitrary directories or modules.\n
Evidence: references/adapter-system.md describes the DiscoveryConfig which enables the loading of interceptors from specified dirs and modules at runtime.\n- [COMMAND_EXECUTION]: The tool executes system commands to manage containers and interface with remote computing clusters.\n
Evidence: As documented in references/execution-backends.md, the skill uses docker run for local execution and sbatch/SSH commands to manage jobs on Slurm HPC clusters.

nemo-evaluator-sdk