The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes subprocess.run to execute local helper scripts, formatting tools like ruff, and test suites. For example, scripts/scaffold.py invokes ruff to format generated code, and scripts/test_scripts.py runs various scripts in subprocesses to verify end-to-end functionality within the development environment.
[DYNAMIC_EXECUTION]: The skill employs dynamic code execution to facilitate testing and modular loading of generated code. scripts/test_scripts.py uses exec() and compile() to validate the behavior of rendered configuration classes. Additionally, scripts/check_port.py uses importlib.import_module to dynamically load the newly scaffolded architecture package for parameter validation against model checkpoints.
[EXTERNAL_DOWNLOADS]: The skill connects to the Hugging Face Hub (huggingface.co) to download model configurations, safetensors metadata, and weight files. These operations target a well-known, trusted service and are essential for the model import and inspection workflow. The skill also utilizes the trust_remote_code=True feature in the transformers library, which is a known high-capability requirement for this domain that is appropriately documented in the skill's safety guidelines.
[INDIRECT_PROMPT_INJECTION]: The skill possesses an indirect injection surface by ingesting untrusted model configurations and cards from Hugging Face. While malicious data in these sources could theoretically attempt to influence the code generation process, the skill mitigates this by using AST-based generation (ast.unparse) rather than simple string templates for sensitive code sections.

import-model