The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/run_training.py executes training commands provided via the --command argument using the run_persistent_command utility. While intended for deep learning tasks, this allows for the execution of arbitrary shell commands within the environment.\n- [DYNAMIC_EXECUTION]: The skill employs dynamic code loading in two places:\n
scripts/run_training.py modifies sys.path at runtime to import shared modules from computed relative paths.\n
scripts/write_outputs.py uses importlib.util to load and execute the write_run_bundle.py module from computed relative file system paths.\n- [INDIRECT_PROMPT_INJECTION]: The skill parses training logs (STDOUT/STDERR) to extract epochs, steps, and metrics. A malicious training process could output specific strings to manipulate the generated training summary and status reports.\n
Ingestion points: scripts/run_training.py captures STDOUT and STDERR from the training command.\n
Boundary markers: No explicit boundary markers or 'ignore' instructions are used when parsing the logs.\n
Capability inventory: The skill can execute subprocesses and write files to the train_outputs/ directory.\n
Sanitization: Log parsing relies on regex matches for specific keywords (e.g., 'epoch', 'step', 'accuracy'), which limits the scope but does not sanitize the input against adversarial strings that match these patterns.

run-train