The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external time-series data from CSV and JSON files, creating a surface for indirect prompt injection.
Ingestion points: The script scripts/forecast_csv.py reads user-provided files via the input argument.
Boundary markers: No explicit delimiters or instructions are used to distinguish data from potential system commands for the agent context.
Capability inventory: The skill has high capabilities, including file writing, bash execution, and the ability to execute subprocesses (detected in scripts/check_system.py).
Sanitization: Data is converted to float32 before forecasting, which provides strong protection for the model inference process, though the raw text in the CSV remains visible to the agent.
[COMMAND_EXECUTION]: The preflight script scripts/check_system.py executes system commands to check hardware compatibility.
Evidence: Uses subprocess.run to call sysctl and vm_stat on macOS to verify available RAM.
Context: These operations are benign and restricted to querying system resource levels to prevent out-of-memory crashes.
[EXTERNAL_DOWNLOADS]: The skill automatically downloads significant binary data (model weights) during first-time use.
Evidence: Model weights (~800MB to 2GB) are fetched from Google Research repositories on Hugging Face (google/timesfm-2.5-200m-pytorch).
Context: Downloads are from a well-known service and a trusted organization, necessary for the skill's primary forecasting functionality.

timesfm-forecasting