The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill fetches model checkpoints from Hugging Face, an established and well-known service for machine learning models.
[COMMAND_EXECUTION]: The setup process involves installing the numba package using uv pip with the --system and --break-system-packages flags. This is a common requirement in ephemeral container environments to ensure optimized computation libraries are available.
[DYNAMIC_EXECUTION]:
The skill uses Numba's njit (Just-In-Time compilation) to compile Python and NumPy code into machine-executable instructions at runtime for performance.
A custom unpickler implementation in load_checkpoint.py is used to load PyTorch checkpoints (.pth files). This implementation follows security best practices by overriding find_class to strictly whitelist only the necessary tensor and collection types, significantly reducing the risks associated with loading serialized Python objects.
[INDIRECT_PROMPT_INJECTION]:
Ingestion points: The skill ingests user-provided time series data via CSV files, lists, or direct input.
Boundary markers: Explicit boundary markers for numerical data are not used, as the input is treated as numerical tokens rather than natural language instructions.
Capability inventory: The skill has the capability to write files (saving forecast plots as PNG) and install packages via the system shell.
Sanitization: The skill effectively sanitizes input by strictly casting user data to 1-D float arrays using np.asarray(series, dtype=np.float32). Any non-numerical or malicious injection attempts within the time series data would be rejected or rendered inert by this numerical conversion before reaching the model or visualization logic.

forecasting-reverso