The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads model checkpoints from Hugging Face using the huggingface_hub library. Specifically, scripts/prepare_cosmos_embed_inference.py fetches pre-trained embeddings models, defaulting to the vendor's repository (nvidia/Cosmos-Embed1-224p).
[COMMAND_EXECUTION]: The workflow relies on executing shell commands and Python scripts to manage the iterative pipeline. Bundled helpers use subprocess.run to launch Docker containers for training and inference, and scripts/deft_python.sh provides a mechanism for selecting and executing the host Python interpreter with the necessary dependencies.
[PRIVILEGE_ESCALATION]: The skill includes utilities for managing file ownership issues typical in Docker-based workflows. The scripts/restore_docker_mount_permissions.py script uses a root-level container to perform chown and chmod on host directories to restore access to artifacts created by containers. Additionally, the documentation recommends specific shared memory and memory lock ulimit settings required for GPU-accelerated workloads.
[INDIRECT_PROMPT_INJECTION]: The skill acts as a data consumer, ingesting user-provided KPI and training datasets in LLaVA JSON format. These inputs influence the generation of training TOML and inference YAML files. While the skill validates JSON structure and resolves file paths within a workspace boundary, it represents a potential surface for indirect injection if the processed dataset contains malicious instructions targeting the VLM components.

tao-run-deft-cr-its-mining