The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/run_training.py executes arbitrary commands provided via the --command argument using subprocess.run. While the command is parsed with shlex.split, this provides a wide execution surface for any binary available in the environment.
[REMOTE_CODE_EXECUTION]: The script scripts/write_outputs.py performs dynamic code loading and execution by using importlib.util to load a module from a path calculated at runtime (../../../shared/scripts/write_run_bundle.py). This allows the execution of code residing outside the skill's own directory.
[PROMPT_INJECTION]: The skill contains a surface for indirect prompt injection through its log parsing logic in scripts/run_training.py.
Ingestion points: The skill captures and processes stdout and stderr from the training process.
Boundary markers: None; the output is parsed directly by regular expressions without delimiters or instructions to ignore embedded content.
Capability inventory: The skill can execute shell commands and load external Python modules.
Sanitization: No sanitization or validation is performed on the log content before using regex to extract metrics and checkpoint names.

run-train