Refactor — one phase, manually driven

This is a manual phase control, not a separate method. It runs step 4 of the predictive cycle and then stops. Use it when you want to steer Red-Green-Refactor yourself instead of invoking the whole exact-coding workflow. The method lives in one place; this skill only selects which part of it to execute now.

For a measured cleanup pass over the whole finished tree rather than one green step, use end-refactor instead.

What to read

1. .pi/skills/predictive-tdd/SKILL.md — section 4. Inspect and refactor, including Domain responsibility review and the Mandatory domain-boundary trial. That section is binding as written, with its before/after records.
2. The matching profile under .pi/skills/predictive-tdd/stacks/ for language- and framework-specific application guidance only.
3. .pi/skills/exact-coding-shared/human-in-the-loop.md for the Refactor checkpoint and the prediction-mismatch rule.

What to do

First verify the behavior tests are green and say so. Refactoring on red is not permitted by the Core rules; if anything fails, stop and offer green instead.

Then execute exactly step 4 of the cycle as written in the Predictive TDD skill: the Four Rules review in order, the domain-responsibility review, and the domain-boundary trial with its recorded Domain responsibility / Independent change axes / Boundary candidate and Boundary outcome / Semantic change / Fewest Elements check blocks. Analysis alone is not the attempt — when a credible seam exists, try the structural change and keep or undo it on the evidence.

Make at most one refactoring at a time and predict and run the smallest relevant check after each.

Both refactor profiles ship in this tree. If the human named the isolated profile, delegate the review through the subagent tool with agent: refactor and agentScope: both and verify Green before applying the checkpoint yourself; otherwise refactor inline in this context.

Where to stop

Stop at the Refactor checkpoint from .pi/skills/exact-coding-shared/human-in-the-loop.md — including when no change improved the code. Report the Four Rules decision, the boundary record, any change made, and the passing gates. Do not close the cycle by activating the next behavior — the human invoked a single phase; that invocation is the checkpoint, and they drive the next behavior with red.

On a prediction mismatch, follow Prediction mismatch in the Predictive TDD skill and stop there.