EXACT Coding — Predictive TDD v1 — Isolated Refactor

This is the consumer form of the universal EXACT Coding Predictive-TDD line. It runs in one shared context: Test List once, then one-test Red-Green-Refactor cycles. Refactoring uses the Four Rules of Simple Design in an isolated subagent. Refactoring also applies a domain-responsibility review and a mandatory concrete boundary trial whenever it finds a credible semantic seam. This profile deliberately has no APP calculation or metric-driven end pass.

Preparation

1. Read .claude/skills/test-list/SKILL.md.
2. Read .claude/skills/predictive-tdd/SKILL.md.
3. Determine the project's language and test framework. Read the matching file under .claude/skills/predictive-tdd/stacks/ before changing code.
4. Read the complete specification and establish the applicable baseline gates.

Do not assume TypeScript or Vitest from this orchestration file. Concrete inactive-test syntax, examples, paths, commands, compiler behavior, and quality tools belong only to the selected stack profile.

Sequence

1. Create the complete ordered test list with every future behavior inactive, then predict and verify that the inactive list leaves the full suite green.
2. Apply the Test-List checkpoint from .claude/skills/exact-coding-shared/human-in-the-loop.md.
3. For exactly one behavior at a time, follow the Predictive TDD skill:
   • activate one behavior and reach behavioral Red,
   • state falsifiable predictions before deterministic checks and compare them explicitly with reality,
   • apply the Red checkpoint,
   • reach Green with the smallest production change,
   • after every Green, delegate the Four Rules review through the Agent tool with subagent_type: refactor; invoke it even when Green changed no production code,
   • perform the mandatory domain-boundary trial and retain or narrowly undo it based on semantic and test evidence,
   • apply the Refactor checkpoint.
4. Continue until every listed behavior is executable and all applicable gates from the active stack profile pass.

A test already satisfied by an earlier generalization is legitimate evidence. Confirm it and do not manufacture a failure or production change.

Method boundary

This is Predictive TDD, not TCR. Do not create phase commits or use a hard reset as a phase mechanism. Preserve successful work in the working tree. If a refactoring trial fails a check or does not improve intent, undo only that trial before continuing.

The subagent never manages checkpoints or waits for the user. The main context reads its report, verifies Green, and applies the shared Refactor checkpoint.

Human-in-the-loop

.claude/skills/exact-coding-shared/human-in-the-loop.md is the single source of truth for checkpoints. Its default full-hitl level stops after Test List, Red, and Refactor, and whenever a prediction is wrong. Green has no default stop.

Provenance

Exported from the promoted exact-ptdd-v1-pi workflow. The methodology was validated on GPT-5.6 SOL with pi. Other harness trees are mechanical ports of the same files, not claims of cross-harness validation.