The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run() to execute system utilities like neuron-ls to verify the environment and hardware availability (in templates/neuron_device_validator_template.py). These are standard checks for the intended purpose.
[DYNAMIC_EXECUTION]: The skill employs importlib.import_module() and importlib.util.spec_from_file_location() to dynamically load the target model's modeling file and user-defined test files (in scripts/adapters/nxdi.py, scripts/nxdi_compiled_loader.py, and scripts/run_stage2.py). This is a core architectural requirement for analyzing arbitrary model implementations.
[EXTERNAL_DOWNLOADS]: The skill fetches model configurations, tokenizers, and reference weights from HuggingFace using the transformers library (e.g., AutoModelForCausalLM.from_pretrained). These are standard operations targeting well-known services (HuggingFace) for obtaining reference data.
[SAFE]: All detected dangerous patterns (shell execution, dynamic imports, remote model fetching) are strictly scoped to the primary purpose of the skill: benchmarking and verifying machine learning model implementations. No signs of obfuscation, exfiltration, or unauthorized persistence were found.

neuron-framework-equivalence