The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external data from CSV files which provides a surface for indirect prompt injection.
Ingestion points: scripts/forecast_csv.py and examples/anomaly-detection/detect_anomalies.py ingest data from CSV files.
Boundary markers: No explicit delimiters or instructions to ignore embedded content are used during file reading.
Capability inventory: The skill has the Bash tool allowed in SKILL.md and uses subprocess.run in scripts/check_system.py.
Sanitization: Data is converted to numeric types using astype(np.float32), which naturally mitigates text-based injection in data columns, but metadata or unhandled columns remain a potential surface.
[COMMAND_EXECUTION]: The skill executes shell commands to inspect system hardware resources.
Evidence: subprocess.run(["sysctl", "-n", "hw.memsize"]) and subprocess.run(["vm_stat"]) are called in scripts/check_system.py.
Context: These calls use hardcoded arguments and serve the legitimate purpose of verifying that the machine meets the model's high memory requirements.
[EXTERNAL_DOWNLOADS]: The skill downloads large model weights from a remote repository upon first use.
Evidence: scripts/forecast_csv.py calls from_pretrained("google/timesfm-2.5-200m-pytorch").
Context: This fetches resources from a well-known organization's official repository on Hugging Face.
[DYNAMIC_EXECUTION]: The skill dynamically loads Python modules to check for their existence and version during setup.
Evidence: importlib.import_module(import_name) is used in scripts/check_system.py within the check_package function.
Context: The package names being checked are static strings, used strictly for verifying environmental dependencies.

timesfm-forecasting