The Agent Skills Directory

[COMMAND_EXECUTION]: The skill generates and executes shell commands using docker run to perform deep learning tasks such as training, evaluation, and inference via the re_identification CLI tool.
[INDIRECT_PROMPT_INJECTION]: The skill establishes an ingestion surface for potentially untrusted data through the processing of dataset archives and model checkpoints.
Ingestion points: Dataset directories (dataset.train_dataset_dir, dataset.test_dataset_dir, dataset.query_dataset_dir) and model checkpoint paths defined in SKILL.md and references/skill_info.yaml.
Boundary markers: The skill does not define specific delimiters or instructions to ignore embedded content within the ingested data files.
Capability inventory: The skill possesses the capability to execute shell commands and spawn containerized processes to process the ingested files.
Sanitization: There is no explicit description of content sanitization or archive validation for the user-provided datasets.
[DYNAMIC_EXECUTION]: The skill includes instructions for loading model checkpoints that may involve unsafe deserialization, particularly when using specific environment variable overrides.
Evidence: SKILL.md instructs the user on setting TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1 for checkpoint compatibility, which enables full object loading. The skill provides a corresponding security warning advising users to avoid this setting for untrusted checkpoints.

tao-train-reid