The Agent Skills Directory

[COMMAND_EXECUTION]: The skill makes extensive use of the pixi environment manager and subprocess.run to execute helper scripts (e.g., inspect_hf.py, scaffold.py) and external developer tools like the ruff linter. These operations are restricted to the local development environment and are standard for software engineering tasks.
[DYNAMIC_EXECUTION]: Several components utilize dynamic loading or execution. scripts/check_port.py uses importlib.import_module to load user-generated model packages for verification. scripts/test_scripts.py uses exec(compile(...)) to validate the integrity of generated configuration code during its internal test suite. These uses are scoped to the skill's primary function of scaffolding and verifying new code.
[EXTERNAL_DOWNLOADS]: The skill downloads model configurations, metadata, and potentially weights from the Hugging Face Hub, which is a well-known and trusted service in the machine learning domain. These downloads are initiated by the user providing a specific model ID.
[REMOTE_CODE_EXECUTION]: The skill's instructions and verification scripts (e.g., scripts/compare_layers.py) use the trust_remote_code=True parameter when interacting with the Hugging Face transformers library. While this allows for the execution of code contained within a model repository, it is a standard and often necessary requirement for loading novel model architectures. The skill's configuration (pixi.toml) explicitly pins transformers to a version that addresses known vulnerabilities related to this feature, demonstrating a security-conscious approach by the author.

import-model