The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/collect_baseline.py executes the system utility npu-smi info via subprocess.run to retrieve hardware specifications and NPU status for benchmarking reports. Additionally, the skill workflow involves executing generated shell scripts (infer.sh) to run model inference locally.
[EXTERNAL_DOWNLOADS]: The skill instructions guide the agent to download model weights and configuration files from well-known services, specifically Hugging Face and ModelScope, using standard CLI tools like huggingface-cli or modelscope.
[DATA_EXFILTRATION]: The skill collects environment metadata including NPU models, software versions (CANN, PyTorch), and inference latency metrics. This data is stored locally in baseline_metadata.json for performance tracking and is not transmitted to external servers.
[PROMPT_INJECTION]: The skill includes instructions to the agent to process external model code and dataset files (dataset/default_prompt.json). While these represent potential ingestion points for indirect prompt injection, the risk is mitigated as the skill operates within a controlled framework environment for benchmarking purposes.
[SAFE]: The dynamic generation of Python and Shell scripts follows provided templates in references/templates.md. This behavior is transparent and necessary for adapting diverse model architectures to the ModelRunner framework.

external-cannbot-model-model-infer-migrator