The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.Popen and subprocess.call within references/virtualenv_runner.py to execute user-provided Python scripts and the Python interpreter associated with a virtual environment. It also executes standard system utilities like ps and pgrep for process monitoring on Linux and macOS hosts.
[DYNAMIC_EXECUTION]: The runner utility generates a local wrapper script, launch_job.py, which is written to the job's results directory. This script acts as a gatekeeper to manage the process lifecycle, ensuring the training script is executed in a new session and that all child processes are cleaned up upon completion.
[SAFE]: The implementation includes several security best practices for local process execution:
Symlink Protection: Uses os.O_EXCL and os.O_NOFOLLOW when creating the job wrapper and metadata files to prevent symlink-based file clobbering or unauthorized execution of pre-planted files.
Process Identity Verification: Employs an opaque process start marker (retrieved from /proc/[pid]/stat or ps lstart) to ensure that status and cancel operations only target the exact process intended, preventing actions against recycled PIDs.
Group Cleanup: Automatically terminates the entire process group upon script completion to prevent resource leaks from background workers like data loaders.

tao-run-on-virtualenv