The Agent Skills Directory

[COMMAND_EXECUTION]: The skill relies extensively on shell command execution to manage Docker containers, verify GPU status via nvidia-smi, and orchestrate the AutoML stages through the driver.sh harness. It uses docker run with host mounts to execute training and evaluation tasks.
[DYNAMIC_EXECUTION]: The workflow generates Python scripts (e.g., driver.py in cards/20-launch-recs.md) and executes them at runtime using the host Python interpreter. It also uses shell heredocs to execute Python blocks for result interpretation and session management.
[INDIRECT_PROMPT_INJECTION]: The skill processes external model skills, search-space schemas, and dataset manifests which could contain malicious instructions. The autoresearch algorithm also utilizes an LLM to propose modifications to hyperparameters based on experiment history.
Ingestion points: skills/models/<network>/SKILL.md, schemas/*.schema.json, and dataset media/annotation paths.
Boundary markers: AutoMLRunner incorporates internal validation; validate_session_settings ensures session identity durability.
Capability inventory: Docker job submission, file system writes for artifacts, and network operations through platform SDKs.
Sanitization: The AutoMLRunner sanitizes feedback provided to LLM brains; search parameters are validated against packaged dataclass schemas before execution.
[EXTERNAL_DOWNLOADS]: The skill performs package installations using pip by resolving versions through a local script (resolve_versions_key.py) and downloads datasets from S3 buckets as part of the evaluation and training processes. These operations target vendor-controlled and trusted resources.

tao-run-automl