First public release.
A workflow harness your coding agent builds for itself. You describe the job to Claude Code, Codex, or Pi; the agent writes a steps.yaml graph, picks the model for each node, and runs it. The graph then owns the order, gates, retries, and budget — the model does the work but cannot quietly skip a step.
Runner — steps.yaml as the single execution contract: needs: dependencies, shell and model steps, when: routing, and gate: assertions. Per-node QA with an independent judge and a bounded improve/retest loop. Retry policies, content-addressed caching, and per-node evidence for every run.
Scale — piw batch runs one frozen graph across a corpus with isolated per-item workspaces and receipts. The manifest pins SHA-256 digests of workflow and corpus; --resume fails closed if either changed. Token and cost ceilings are enforced over recorded usage, including failed attempts.
Inspection — piw detail for a run or one node, piw compare between runs, piw stats, and piw eval to compare models with judges held fixed. Optional local Studio via piw ui.
Guarantees — a model pin that drifts fails the step rather than returning another model's answer. Deterministic orchestration pins configuration and preserves evidence; it does not make LLM output identical between runs.
Full detail in CHANGELOG.md.