The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data including user-defined dataset paths, container image overrides, and runtime logs from training jobs. This surface is mitigated by a mandatory 'Launch Review' step requiring explicit user confirmation before execution. Ingestion points include user parameters and logs; capabilities include local script execution for job management; sanitization is performed via secret linting scripts.
[COMMAND_EXECUTION]: The skill generates and executes shell commands for job lifecycle management (submit, status, logs, cancel) on platforms like Docker, SLURM, and Kubernetes. This behavior is the primary purpose of the skill and is controlled by a preflight validation gate.
[DYNAMIC_EXECUTION]: The agent generates platform-specific artifacts, such as sbatch scripts or configuration files, at runtime. These are constructed from validated templates and model metadata to facilitate job submission.
[CREDENTIALS_UNSAFE]: The skill includes instructions for secure credential management, specifically advising the agent to check for the presence of variables in .env files without printing or logging their content. It also provides best-practice guidance for SSH key generation and installation in the platform preflight documentation.

tao-launch-workflow