| name | ml-execution-preflight |
| description | Checklist before long execution_run or execution_run_background for ML/training scripts. |
| always | true |
ML execution preflight
Run these checks before a long or GPU-heavy execution_run / execution_run_background:
execution_env_info — note max_wall_secs, default timeout, provider id, and artifact caps.
- Interpreter / kernel — for Jupyter or SSH, run a one-line
execution_run that prints Python version and (if relevant) torch.cuda.is_available() or equivalent.
- Timeouts — set
timeout_secs explicitly from model size and hardware; training almost always needs a large value within max_wall_secs.
- Pilot — run the smallest slice that still exercises the code path (few steps / tiny data), then scale.
- Logging — prefer plain-text step logs the model can read from captured stdout (avoid hiding loss inside unreadable progress noise).
- Outputs — decide where checkpoints and metrics will live (
execution_artifact_list, sandbox paths, or copy-out plan). Ephemeral sandboxes lose data when torn down.
- Single path first — do not launch many parallel heavy jobs until one configuration completes successfully.
If any step fails, fix the root cause before scaling.
**Before launching a fine-tuning or dataset generation script, also load long-running-execution with load_skill_instructions.