The Agent Skills Directory

[OBFUSCATION]: The remote_run function in remote_exec.sh encodes shell commands using Base64 before transmitting them to remote hosts. This is documented as a mechanism to avoid shell quoting and escaping issues during SSH transport.\n- [COMMAND_EXECUTION]: The skill provides a library of bash functions for local and remote command execution, including wrapper functions that use eval and bash to process complex command strings.\n- [REMOTE_CODE_EXECUTION]: The skill enables remote code execution through SSH-based utilities and SLURM job submission tools. It manages persistent SSH sessions and automates the staging and execution of scripts on remote GPU clusters.\n- [CREDENTIALS_UNSAFE]: The skill manages and references sensitive files and authentication tokens, including SSH private keys (~/.ssh/id_rsa), Hugging Face tokens, and registry credentials for NVIDIA NGC and Docker Hub stored in files like ~/.docker/config.json and ~/.config/enroot/.credentials.\n- [EXTERNAL_DOWNLOADS]: Fetches source code and examples from NVIDIA's official Model-Optimizer repository on GitHub during the environment setup phase.\n- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted external data by reading cluster configuration files (clusters.yaml) and tailing remote log files, creating a potential surface for indirect instructions to influence agent behavior.\n- [DYNAMIC_EXECUTION]: The skill automates the creation and execution of temporary shell scripts and SLURM batch scripts on both local and remote filesystems to coordinate model optimization tasks.

common