The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/run_training.py uses subprocess.run to execute shell commands provided via the --command argument. This is the core functionality intended for training execution.
[COMMAND_EXECUTION]: The script scripts/write_outputs.py uses importlib.util to dynamically load a shared module from a relative path (../../../shared/scripts/write_run_bundle.py) for standardized output handling.
[PROMPT_INJECTION]: The skill ingests untrusted training logs to extract metrics, posing a surface for indirect prompt injection.
Ingestion points: scripts/run_training.py captures stdout and stderr from the training process and parses them in parse_progress.
Boundary markers: Absent; logs are processed directly without delimiters or instruction-ignore warnings.
Capability inventory: subprocess.run in scripts/run_training.py allows execution of external commands.
Sanitization: Absent; log data is matched using regular expressions without additional validation or sanitization.

run-train