The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data from CSV files which could contain adversarial instructions if the data source is untrusted.
Ingestion points: scripts/forecast_csv.py and examples/global-temperature/run_forecast.py use pandas.read_csv to load data for forecasting.
Boundary markers: Absent. There are no instructions to the agent to ignore text content within the data files.
Capability inventory: The skill has Bash access and utilizes subprocess.run and importlib in its system verification scripts.
Sanitization: Numerical data is explicitly cast to float32, which prevents the execution of code within the model pipeline, though the raw text remains visible to the agent during file reading.
[EXTERNAL_DOWNLOADS]: Fetches pre-trained model weights from a remote repository at runtime.
Evidence: SKILL.md and references/api_reference.md describe the automatic download of weights from google/timesfm-2.5-200m-pytorch on Hugging Face.
Context: The weights are hosted by a well-known technology company (Google) on an established platform (Hugging Face), making this a standard and safe operational dependency.
[COMMAND_EXECUTION]: Executes system-level commands to check hardware compatibility.
Evidence: scripts/check_system.py uses subprocess.run to call sysctl and vm_stat on macOS to verify available memory.
Context: These commands are used for safety to ensure the local machine has enough RAM to load the model without crashing, representing best practice for high-resource foundation models.

timesfm-forecasting