The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill is configured to fetch pre-trained Grounding DINO checkpoints from the official NVIDIA GPU Cloud (NGC) registry. This is performed using the standard ngc CLI tool as part of the setup process.
[COMMAND_EXECUTION]: The workflow relies on executing official TAO Toolkit Docker containers to perform GPU-intensive tasks such as model training, inference, and data analytics. Commands are constructed to use the host user's UID/GID to maintain proper file permissions.
[DYNAMIC_EXECUTION]: The skill dynamically generates and modifies Hydra-based YAML configuration files for the TAO Toolkit. It uses a structured approach to apply workflow defaults and user-specific overrides through specialized Python scripts like apply_spec_overrides.py.
[SAFE]: The skill implements a robust 'assume-hostile' state management system. An auditing script (audit_deft_run.py) validates the consistency of the run state and the existence of required artifacts on disk before any new stage is allowed to execute.
[SAFE]: A 'Pre-Flight Summary' provides a mandatory manual checkpoint, ensuring the user reviews the resolved configuration, planned resource usage, and data paths before the autonomous loop begins.
[SAFE]: Internal Python scripts use standard libraries and follow best practices, such as using yaml.safe_load for parsing configuration files.

tao-run-deft-object-detection