The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to generate and execute shell scripts, sbatch snippets for SLURM, and other platform-specific commands. It relies on a suite of helper Python scripts (e.g., list_tao_platforms.py, resolve_tao_image.py, check_tao_launch_preflight.py) located in the directory defined by the environment variable TAO_SKILL_BANK_PATH.- [CREDENTIALS_UNSAFE]: The workflow involves requesting and managing sensitive credentials such as SSH_KEY_PATH, HF_TOKEN, and cloud provider access keys. The instructions specifically guide the agent to source these from local environment files (e.g., ~~/.tao/secrets.env) while strictly prohibiting the output or logging of secret values.- [DATA_EXFILTRATION]: The skill accesses sensitive local file paths, including SSH keys (~~/.ssh/id_ed25519) and configuration files, to verify platform connectivity and satisfy credential requirements during the launch preflight stage. This is a legitimate requirement for cluster access in the context of the skill's purpose.- [EXTERNAL_DOWNLOADS]: The instructions include a dependency remediation step that prompts the user to install necessary external packages and CLI tools, such as nvidia-tao-sdk variants, aws-cli, or Docker clients, when they are detected as missing from the environment.- [INDIRECT_PROMPT_INJECTION]: The skill ingests and processes external data from datasets (annotations and media) in formats like JSON or JSONL. This creates a surface for indirect prompt injection if the content of these files is malicious, although the skill includes preflight checks to validate schema and content before execution.

tao-launch-workflow