pi-workflowslisted

# Deterministic Workflows The failure this prevents: an agent given N steps skips some, reorders them, or declares done early. The structural fix: **the model never owns control flow** — a runner calls the model once per step, gates decide pass/fail, artifacts chain between steps. The graph transition logic is deterministic for the same validated node outputs; model outputs can vary, so pin models, gate them, and retain run evidence instead of promising identical paths across live LLM calls. CLI: `piw`. Runner: `scripts/run_steps.py`. Authoring is direct: write `steps.yaml`, validate it, run it. Run `piw schema` for the concise node/input catalog or `piw schema --json` for the complete machine-readable contract. Before rebuilding a common graph fragment, run `piw actions`; templates expand into ordinary inspectable nodes with `piw create --action` or `piw add`. ## Agent operating contract Use pi workflows when control flow must be remembered by code: required order, repeatable inputs, branches, retries, quality gates, cost ceilings, external effects, or scheduled execution. Do not create a workflow for a one-step task. For ordinary work, follow this loop and do not skip inspection: ```bash piw actions --json piw create work --action ACTION piw validate work/steps.yaml --json piw run work/steps.yaml --input-file input.txt --json # Use the returned run id. Inspect the whole trace, then material nodes. piw detail work/steps.yaml RUN_ID --json piw detail work/steps.yaml RU