@pmelab/gtd 50.0.0 → 51.0.0

package/dist/workflows.mjs

@@ -17241,6 +17241,7 @@ const defaults = {
 	qualityReviews: "correctness, owasp-security, ponytail-review, test-audit, conventions, spec-challenge",
 	refactorQualityReviews: "correctness, ponytail-review, conventions",
 	continuations: "3",
+	prototypeRounds: "3",
 	continuationStall: "2"
 };
 /** Environment settings. */
@@ -17318,6 +17319,11 @@ task in front of you end to end, tests first, and come back to fix
 things when redirected — a failing check, or reviewer feedback. Stay
 strictly inside that task; never touch code outside it. Treat feedback as the
 diagnosis to act on; discard it, with reason, only when simply wrong.`;
+const testerPersona = `You are the hands-on tester in gtd's build pipeline: use what was just
+built the way a person would — run the real program, open the real
+page, call the real endpoint — and report exactly what you did and
+what you saw. Never fix anything yourself, and never pass a criterion
+you did not watch hold.`;
 const finisherPersona = `You are the closing identity in gtd's build pipeline — the last coder
 before it closes. Fix a late-breaking failing check after sign-off,
 focused and minimal. Every turn lands its own commit — nothing here
@@ -17743,8 +17749,11 @@ const acceptanceCriteriaBlock = `
 Whichever lap this is, \`.gtd/REQUIREMENTS.md\` carries a
 \`## Acceptance Criteria\` section: one \`- AC-<n>: <criterion>\` line
 per criterion, numbered from \`AC-1\`. They bind every later step —
-the build is done when each one holds — so each must be checkable: a
-test or a command shows it holds or not. A later lap keeps every ID:
+the build is done when each one holds — so each must be behaviour a
+test can pin: one that fails when it does not hold. A green suite, a
+passing typecheck or lint, a generator or build that runs — the health
+check enforces those on every lap, so none is a criterion; fold what it
+must produce into a behaviour a test can observe. A later lap keeps every ID:
 reword a criterion in place, append a new one with the next unused
 number, delete a dropped one — never renumber, never reuse an ID. When
 the file holds no list (a review loop-back rewrites it), recover the
@@ -17852,7 +17861,7 @@ const clearCriteriaQuestion = {
 	id: "clear",
 	primitive: "noul",
 	instructions: "state.criteria lists the acceptance criteria the human's sketch (state.sketch) stated, as gtd numbered them. Decide whether they can stand as the binding definition of done, with no design conversation.",
-	criteria: "yes: every criterion is clear and checkable — a test or a command can show it holds or not — and together they say what to build. no: any one criterion is vague, subjective or untestable, or the sketch leaves an open product question. Unsure means no."
+	criteria: "yes: every criterion is clear behaviour a test can pin — a test would fail if it did not hold — and together they say what to build. no: any one criterion is vague, subjective or untestable, only a command can check it (a suite, typecheck, lint, build or generator run that passes), or the sketch leaves an open product question. Unsure means no."
 };
 const clearCriteriaMessage = () => `The sketch stated its own acceptance criteria. Run \`gtd judge answer\` and
 pipe \`yes\` when every one is clear and checkable, so design is skipped and
@@ -18771,6 +18780,80 @@ Two ways forward (then run \`gtd land\`):
 - **Overrule** — land a clean tree to finish without them. The slice records
   the IDs it covers and, as \`uncovered\`, the ones it does not
 `;
+const testerSystem = () => `${testerPersona}
+
+${agentConduct}`;
+const PROTOTYPE_FORMAT = `- Write \`.gtd/PROTOTYPE.md\`, replacing any earlier report: one \`## \` chunk
+  per criterion, headed \`## AC-<n>: pass\`, \`## AC-<n>: fail\` or
+  \`## AC-<n>: blocked\`, then one chunk per problem you found outside any
+  criterion, headed \`## Problem <short title>: fail\` (\`: pass\` once a
+  re-test shows it fixed). Under each: the exact steps you ran, what you
+  observed (output, a screenshot path, the error), and for a failure what
+  you expected instead
+- \`blocked\` is for a criterion you could not exercise, or one whose
+  expected behaviour itself looks wrong — say which and why: a person
+  decides it, never a fixer`;
+const prototypeTestingPrompt = (slice, retest) => `${stateFileRules}
+- Read \`.gtd/REQUIREMENTS.md\`: the settled requirements for this change
+${sliceLine(slice)}- The change is built and its suite is green. Now test it by hand: use
+  every acceptance criterion the way a person would and watch what
+  happens. Running the test suite is not this — it already ran
+- Find how this project is exercised for real, in this order: a project
+  skill that launches or drives it (e.g. \`.claude/skills/*\`); a manual
+  testing or running section in \`AGENTS.md\`, \`CLAUDE.md\` or \`README.md\`;
+  failing those, the project's shape — a browser or Playwright for a web
+  project, the built CLI in a scratch directory for a CLI, a request client
+  for an API
+- Found none and recognised no shape: write only a chunk headed
+  \`## No testing instructions\` saying what you looked for, and finish —
+  a person is asked; never guess
+- Test in a disposable place only — a scratch directory outside this
+  repository, a throwaway repository, a dev server on a free port — never
+  against the user's real data or this working tree. Point every tool's
+  output (screenshots, traces, test results) outside the repository too
+${retest ? `- This is a re-test: \`.gtd/PROTOTYPE.md\` holds the last round's report and
+  a fix turn or a person has changed the code since. Re-test every failed or blocked criterion and
+  every problem it lists in full, then re-run the passed ones quickly
+` : ""}${PROTOTYPE_FORMAT}
+- Change nothing but \`.gtd/PROTOTYPE.md\` — a turn that does is refused
+- Leave everything uncommitted and finish your turn — do not commit
+`;
+const prototypeFixingPrompt = (slice) => `${stateFileRules}
+- Read \`.gtd/REQUIREMENTS.md\`: the settled requirements for this change
+${sliceLine(slice)}${criteriaBinding(true)}
+- Read \`.gtd/PROTOTYPE.md\`: a hands-on test of the change. Fix every
+  \`fail\` chunk of this change's criteria${slice === void 0 ? "" : " (this slice's only)"} or a problem found
+  beside them, at its cause, with a test that shows it; leave \`blocked\`
+  ones to the person
+- A \`fail\` the code does not explain (the tester misread what it saw) has
+  nothing to fix: leave it and change nothing for it — the next round re-tests
+- Leave \`.gtd/PROTOTYPE.md\` as it is: the next round re-tests from it
+- Run \`${env.testCommand}\` and finish only once it is green
+- Leave everything uncommitted and finish your turn — do not commit
+`;
+const openItems = (items) => items.map((item) => `- ${item.subject}: ${item.verdict}`).join("\n");
+const prototypeOpenMessage = (items, rounds) => `The hands-on test of this slice still has open items${rounds > 0 ? ` after ${rounds} fix round${rounds === 1 ? "" : "s"}` : ", and none a fixer can take"}:
+${openItems(items)}
+
+\`.gtd/PROTOTYPE.md\` holds the last report: the steps run and what was seen.
+
+Two ways forward (then run \`gtd land\`):
+- **Fix** — change the code, or the testing instructions, yourself (the
+  criteria are settled: an edit to them is refused); the suite runs and the
+  slice is tested again, with a fresh round budget
+- **Overrule** — land a clean tree to go on to the quality lap with these
+  items open
+`;
+const prototypeUnguidedMessage = () => `The hands-on test found no way to exercise this project for real — no
+project skill, no manual testing section in \`AGENTS.md\`, \`CLAUDE.md\` or
+\`README.md\`, and no project shape it recognised. \`.gtd/PROTOTYPE.md\` says
+what it looked for.
+
+Two ways forward (then run \`gtd land\`):
+- **Instruct** — add testing instructions (a project skill, or a manual
+  testing section); the slice is tested again
+- **Skip** — land a clean tree to go on to the quality lap untested
+`;
 //#endregion
 //#region src/workflows/shared/steps.ts
 const FEEDBACK = ".gtd/FEEDBACK.md";
@@ -18782,6 +18865,7 @@ const REFACTOR_NOTE = ".gtd/REFACTOR-NOTE.md";
 const QUALITY = ".gtd/QUALITY.md";
 const SLICES = ".gtd/SLICES.md";
 const HANDOFF = ".gtd/HANDOFF.md";
+const PROTOTYPE = ".gtd/PROTOTYPE.md";
 const planner = () => env.plannerModel ?? "";
 const coder = () => env.coderModel ?? "";
 const fixCheck = (rules) => agentWithSkills("fix", buildFixPrompt(rules), {
@@ -38499,6 +38583,7 @@ const ShippedOptions = Struct({
 	qualityReviews: setting$1,
 	refactorQualityReviews: setting$1,
 	continuations: setting$1,
+	prototypeRounds: setting$1,
 	continuationStall: setting$1,
 	testCommand: setting$1,
 	plannerModel: setting$1,
@@ -39748,7 +39833,9 @@ const reviewedRefactor = async (shape, options = {}) => {
 				feedback,
 				fixAccess: strayAccess(p.strays),
 				lensBase: entry,
-				rules: refactorRules(p.strays, p.pattern)
+				rules: refactorRules(p.strays, p.pattern),
+				walkthrough: true,
+				lap: "refactor"
 			});
 			if (outcome.verdict === "signoff") return true;
 			await reUnwind(outcome);
@@ -39763,7 +39850,7 @@ const reviewedRefactor = async (shape, options = {}) => {
 const COVERED_MIN_P = .7;
 /** Covering turns before the process rests at `uncovered`; a human edit there refills them. */
 const COVERING_RETRIES = 2;
-const frozen = () => requireCriteriaFrozen(REQUIREMENTS);
+const frozen$1 = () => requireCriteriaFrozen(REQUIREMENTS);
 const green = (guard) => healthy(() => agentWithSkills("fix.suite.fixing", implementFixPrompt(), {
 	label: "Fixing the check",
 	file: FEEDBACK,
@@ -39771,7 +39858,7 @@ const green = (guard) => healthy(() => agentWithSkills("fix.suite.fixing", imple
 	system: builderSystem()
 }), {
 	guard: () => {
-		frozen();
+		frozen$1();
 		guard?.();
 	},
 	guardRests: false
@@ -39824,7 +39911,7 @@ const requireCoverage = async (base = start(), slice) => {
 				system: builderSystem(),
 				allowEmpty: true
 			});
-			frozen();
+			frozen$1();
 		} else {
 			await human("uncovered", {
 				message: uncoveredMessage(uncovered),
@@ -39833,13 +39920,136 @@ const requireCoverage = async (base = start(), slice) => {
 				acceptClean: true
 			});
 			if (!changes().some((c) => isCode$1(c.path))) return record(covered, uncovered);
-			frozen();
+			frozen$1();
 			retries = 0;
 		}
 		await green();
 	}
 };
 //#endregion
+//#region src/workflows/feature/prototype.ts
+/** The `prototypeRounds` setting. Anything but a non-negative integer fails rather than guess a cap. */
+const prototypeRounds = () => {
+	const value = vars.prototypeRounds ?? defaults.prototypeRounds;
+	if (!/^(0|[1-9]\d*)$/.test(value)) refuse(`gtd: prototypeRounds must be a non-negative integer, got ${JSON.stringify(value)}`);
+	return Number(value);
+};
+const CHUNK = /^(AC-\d+|Problem\b.*?)\s*:\s*(pass|fail|blocked)$/i;
+const UNGUIDED = /^no testing instructions$/i;
+/** The `## ` chunks of a prototype report; a heading in neither shape is prose, not an item. */
+const reportOf = (text) => {
+	let unguided = false;
+	const items = [];
+	const malformed = [];
+	for (const { line, fenced } of markFenced(text)) {
+		const heading = fenced ? null : HEADING.exec(line);
+		if (heading?.[1] !== "##") continue;
+		if (UNGUIDED.test(heading[2])) unguided = true;
+		const chunk = CHUNK.exec(heading[2]);
+		if (chunk === null && /^(ac-\d|problem)/i.test(heading[2])) malformed.push(heading[2]);
+		if (chunk !== null) items.push({
+			subject: /^ac-/i.test(chunk[1]) ? chunk[1].toUpperCase() : chunk[1],
+			verdict: chunk[2].toLowerCase()
+		});
+	}
+	return {
+		unguided,
+		items,
+		malformed
+	};
+};
+/**
+* The report the testing turn left, refused unless it is the only change and
+* names every criterion once, with its open items: another slice's criteria
+* are not this slice's to fix.
+*/
+const requireReport = (slice, retest) => {
+	const others = changes().map((c) => c.path).filter((path) => path !== PROTOTYPE);
+	if (others.length > 0) refuse(`gtd land: prototype-report: test in a disposable place and change nothing but ${PROTOTYPE} — revert these:\n${others.map((path) => `  - ${path}`).join("\n")}`);
+	const text = read(PROTOTYPE);
+	if (text === void 0) return refuse(`gtd land: prototype-report: write ${PROTOTYPE}`);
+	const report = reportOf(text);
+	if (report.malformed.length > 0) refuse(`gtd land: prototype-report: head each chunk \`## <ID>: pass|fail|blocked\` or \`## Problem <title>: pass|fail|blocked\` — these miss it:\n${report.malformed.map((h) => `  - ${h}`).join("\n")}`);
+	if (report.unguided && report.items.length === 0 && !retest) return {
+		unguided: true,
+		open: []
+	};
+	const requirements = read(".gtd/REQUIREMENTS.md") ?? "";
+	const ids = criterionIds(requirements, slice);
+	const elsewhere = criterionIds(requirements).filter((id) => !ids.includes(id));
+	const count = (id) => report.items.filter((i) => i.subject === id).length;
+	const wrong = ids.filter((id) => count(id) !== 1);
+	if (wrong.length > 0) refuse(`gtd land: prototype-report: ${PROTOTYPE} needs exactly one chunk headed \`## <ID>: pass|fail|blocked\` for ${wrong.join(", ")}`);
+	return {
+		unguided: false,
+		open: report.items.filter((i) => i.verdict !== "pass" && !elsewhere.includes(i.subject))
+	};
+};
+const frozen = () => requireCriteriaFrozen(REQUIREMENTS);
+const testing = async (slice, retest) => {
+	await agentWithSkills("testing", prototypeTestingPrompt(sliceIds(slice), retest), {
+		label: retest ? "Re-testing the slice by hand" : "Testing the slice by hand",
+		file: PROTOTYPE,
+		model: coder(),
+		system: testerSystem(),
+		allowEmpty: true
+	});
+	return requireReport(slice, retest);
+};
+const fixing = async (slice, guard) => {
+	await scope("fix", () => agentWithSkills("fixing", prototypeFixingPrompt(sliceIds(slice)), {
+		label: "Fixing what the hands-on test found",
+		file: PROTOTYPE,
+		model: coder(),
+		system: builderSystem(),
+		allowEmpty: true
+	}));
+	if (changes().some((c) => c.path === ".gtd/PROTOTYPE.md")) refuse(`gtd land: prototype-fix: leave ${PROTOTYPE} as it is — the next round re-tests from it`);
+	frozen();
+	guard?.();
+};
+/** A person's rest; resolves `true` when they changed something to test again, `false` on a clean overrule. */
+const askPerson = async (name, message, label) => {
+	await human(name, {
+		message,
+		label,
+		file: PROTOTYPE,
+		acceptClean: true
+	});
+	if (!changes().some((c) => c.path !== ".gtd/PROTOTYPE.md")) return false;
+	frozen();
+	return true;
+};
+/**
+* Use the slice by hand: a testing turn reports on every criterion, a fix turn
+* takes each failure, and the slice is tested again — until the report is
+* clean, or `prototypeRounds` fixes are spent and a person decides. `guard`
+* holds every fix turn, as it holds the implementation's.
+*/
+const prototypeTest = async (slice, guard) => {
+	const cap = prototypeRounds();
+	if (cap === 0) return;
+	await scope("prototype", async () => {
+		let fixes = 0;
+		let retest = false;
+		for (;;) {
+			const { unguided, open } = await testing(slice, retest);
+			retest = !unguided;
+			if (unguided) {
+				if (!await askPerson("unguided", prototypeUnguidedMessage(), "No way to test by hand")) return;
+			} else if (open.length === 0) return;
+			else if (fixes < cap && open.some((i) => i.verdict === "fail")) {
+				fixes++;
+				await fixing(slice, guard);
+			} else {
+				if (!await askPerson("open", prototypeOpenMessage(open, fixes), "Hands-on test still failing")) return;
+				fixes = 0;
+			}
+			await green(guard);
+		}
+	});
+};
+//#endregion
 //#region src/workflows/feature/seam.ts
 /** Naive-pass retries after a `reject` or a red suite; the next one rests at `seam.steer`. */
 const NAIVE_RETRIES = 2;
@@ -40046,8 +40256,10 @@ const implementing = (options) => scope("implement", async () => {
 		...acceptance,
 		ids: settled(acceptance.ids, options.slice)
 	}, since);
-	await green(built && acceptance !== void 0 ? () => requireAcceptanceKept(acceptance) : void 0);
+	const keepsTests = built && acceptance !== void 0 ? () => requireAcceptanceKept(acceptance) : void 0;
+	await green(keepsTests);
 	await requireCoverage(options.base, options.slice);
+	await prototypeTest(options.slice, keepsTests);
 	const ids = criterionIds(read(".gtd/REQUIREMENTS.md") ?? "");
 	await closing(moreToBuild(options.more));
 	return ids;
@@ -40056,11 +40268,16 @@ const moreToBuild = (more) => typeof more === "function" ? more() : more === tru
 /** The last slice's closing, for a last slice a steer left with no criterion to build. */
 const closeOut = () => scope("implement", () => closing(false));
 /** A slice with `more` after it keeps the criteria and the slicing for the next one. */
-const closing = (more) => run("closing", removeScript(more ? [REFACTOR_NOTE, HANDOFF] : [
+const closing = (more) => run("closing", removeScript(more ? [
+	REFACTOR_NOTE,
+	HANDOFF,
+	PROTOTYPE
+] : [
 	REQUIREMENTS,
 	REFACTOR_NOTE,
 	SLICES,
-	HANDOFF
+	HANDOFF,
+	PROTOTYPE
 ]), { label: "Closing out the implementation" });
 /** Resolves `true` when the human steered (edited), `false` on a clean overrule. */
 const steer = async (retry) => {
@@ -40425,6 +40642,7 @@ const ordinaryStart = async () => {
 	refusePendingHandOff();
 	flagModeOn();
 	sliceCap();
+	prototypeRounds();
 	const stated = statedCriteria(changes(), head$1());
 	await unwind();
 	if (changes(".gtd/FEEDBACK.md").some((c) => c.status !== "deleted")) await human("unwind-failed", {
@@ -40459,6 +40677,10 @@ const skills = (vars) => {
 		seam: ["incremental-implementation"],
 		implement: ["test-driven-development", "incremental-implementation"],
 		"implement.fix.suite": ["debugging-and-error-recovery"],
+		"implement.prototype": [],
+		"implement.prototype.fix": ["test-driven-development", "incremental-implementation"],
+		"implement.prototype.fix.suite": ["debugging-and-error-recovery"],
+		"implement.prototype.health": ["debugging-and-error-recovery"],
 		"implement.health": ["debugging-and-error-recovery"],
 		...lensScopes("build.quality", lensesOf(vars.qualityReviews ?? ""), (lens) => builtInLenses[lens]?.skills ?? [lens]),
 		...lensScopes("refactor.build.quality", lensesOf(vars.refactorQualityReviews ?? ""), (lens) => builtInLenses[lens]?.skills ?? [lens])
@@ -40636,4 +40858,4 @@ var workflows_default = feature;
 */
 const doors = {};
 //#endregion
-export { COMMENT, ESCALATION, FEEDBACK, FIX_CAP, FlagOptions, HANDOFF, QUALITY, REFACTOR, REFACTOR_NOTE, REFACTOR_PACKAGE, REQUIREMENTS, REVIEW, REVIEW_ORIG, SLICES, ShippedOptions, acceptanceFault, access, afterTail, agentWithSkills, answerProductQuestions, answerReviewQuestions, awaitReview, base, baseline, buildFeature, buildTail, builtInLenses, cleanup, cleanupFlow, closeOut, closeReview, coder, collecting, criteriaOf, criterionIds, workflows_default as default, defaults, describeEscalation, design, doors, envDefaults, escalate, escalation, escalationExhausted, feature, fix, fixCheck, fixFlow, fixNits, fixQuality, fixReviewNotes, fixRisks, gate, guarded, handOffReview, handOffTail, healthy, holdsCallSites, implement, implementing, isTestFile, judgeSeam, lensEntriesOf, lensOf, lensesOf, naiveAcceptance, naivePass, ordinaryStart, planAndBuild, planner, qualityLap, reUnwind, recheck, refactorCollecting, refactorFirst, refactorNote, refactorStrays, requireAcceptance, requireAcceptanceKept, requireCriteria, requireCriteriaFrozen, requireCriteriaIfEdited, requireNoNewFlags, requireNoTestFiles, requireSlices, requireStableIds, requireTestCallSitesOnly, requireWithin, retryQuestion, review, reviewFlow, reviewMissing, reviewQuality, reviewedRefactor, reviewing, rollout, rolloutFlow, seamBuild, skills, skipDesign, sliceCap, sliceCriteria, sliceIds, slicesOf, slicing, statedCriteria, steering, summary, tail, threads, triage, unstableIds, unwind };
+export { COMMENT, ESCALATION, FEEDBACK, FIX_CAP, FlagOptions, HANDOFF, PROTOTYPE, QUALITY, REFACTOR, REFACTOR_NOTE, REFACTOR_PACKAGE, REQUIREMENTS, REVIEW, REVIEW_ORIG, SLICES, ShippedOptions, acceptanceFault, access, afterTail, agentWithSkills, answerProductQuestions, answerReviewQuestions, awaitReview, base, baseline, buildFeature, buildTail, builtInLenses, cleanup, cleanupFlow, closeOut, closeReview, coder, collecting, criteriaOf, criterionIds, workflows_default as default, defaults, describeEscalation, design, doors, envDefaults, escalate, escalation, escalationExhausted, feature, fix, fixCheck, fixFlow, fixNits, fixQuality, fixReviewNotes, fixRisks, gate, guarded, handOffReview, handOffTail, healthy, holdsCallSites, implement, implementing, isTestFile, judgeSeam, lensEntriesOf, lensOf, lensesOf, naiveAcceptance, naivePass, ordinaryStart, planAndBuild, planner, qualityLap, reUnwind, recheck, refactorCollecting, refactorFirst, refactorNote, refactorStrays, requireAcceptance, requireAcceptanceKept, requireCriteria, requireCriteriaFrozen, requireCriteriaIfEdited, requireNoNewFlags, requireNoTestFiles, requireSlices, requireStableIds, requireTestCallSitesOnly, requireWithin, retryQuestion, review, reviewFlow, reviewMissing, reviewQuality, reviewedRefactor, reviewing, rollout, rolloutFlow, seamBuild, skills, skipDesign, sliceCap, sliceCriteria, sliceIds, slicesOf, slicing, statedCriteria, steering, summary, tail, threads, triage, unstableIds, unwind };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pmelab/gtd",
-  "version": "50.0.0",
+  "version": "51.0.0",
   "private": false,
   "description": "Git-aware CLI that emits the next prompt for an autonomous coding agent based on the current repository state",
   "bin": {

package/src/workflows/feature/coverage.feature

@@ -9,6 +9,7 @@ Feature: a feature is done only when every acceptance criterion has a passing te
 
   Background:
     Given the shipped workflows, as gtd install scaffolds them
+    And an environment variable "GTD_PROTOTYPEROUNDS" set to "0"
     And an environment variable "GTD_QUALITYREVIEWS" set to ""
     And an environment variable "GTD_FLAGMODE" set to "false"
     And a file "src/calc.ts" with:

package/src/workflows/feature/fast-path.feature

@@ -8,6 +8,7 @@ Feature: a trivial change takes the same few beats at either risk tier
 
   Background:
     Given the shipped workflows, as gtd install scaffolds them
+    And an environment variable "GTD_PROTOTYPEROUNDS" set to "0"
     And an environment variable "GTD_QUALITYREVIEWS" set to "correctness"
     And an environment variable "GTD_FLAG_RISK_TIER" set to "1"
     And a file "src/greeting.ts" with:

package/src/workflows/feature/feature.ts

@@ -28,6 +28,7 @@ import {
 } from "./criteria.js"
 import { design } from "./planning.js"
 import { buildTail, humanReview, isCode, reUnwind, type ReviewOutcome } from "../shared/review.js"
+import { prototypeRounds } from "./prototype.js"
 import { implement, implementing, seamBuild } from "./seam.js"
 import { sliceCap, sliceCriteria, slicing } from "./slicing.js"
 import { recordRisk, riskTierOn, type RiskTier } from "./tier.js"
@@ -36,7 +37,8 @@ import { recordRisk, riskTierOn, type RiskTier } from "./tier.js"
 // `unwind` reverts the sketch (its intent survives in history) → a
 // green-baseline gate → design → per slice of the criteria, a naive pass,
 // judged with `flagMode` on: it lands as the implementation, or (refactoring
-// first when the judge asks) is built fresh from → the health loop → the
+// first when the judge asks) is built fresh from → the health loop → a
+// hands-on test of every criterion, fixed and re-tested → the
 // triaged quality lap, whose leftovers are carried on → after the last slice,
 // a reviewer's walkthrough of the whole change, every slice's findings
 // among its chunks, for the human, which
@@ -227,6 +229,7 @@ export const ordinaryStart = async (): Promise<void> => {
   // Read right after idle, so a bad value refuses the landing that would pin it.
   flagModeOn()
   sliceCap()
+  prototypeRounds()
   const stated = statedCriteria(changes(), head())
   await unwind()
   if (changes(FEEDBACK).some((c) => c.status !== "deleted")) {

package/src/workflows/feature/flag-mode.feature

@@ -11,7 +11,7 @@ Feature: flag mode off lands the naive pass with no release flag, seam judge or
       """
       import { feature } from "@pmelab/gtd/workflows/feature"
       export { summary, steering, skills, access } from "@pmelab/gtd/workflows/feature"
-      export default feature({ flagMode: false })
+      export default feature({ flagMode: false, prototypeRounds: "0" })
       """
     And an environment variable "GTD_QUALITYREVIEWS" set to "correctness"
     And a file "src/calc.ts" with:

package/src/workflows/feature/lens-models.feature

@@ -12,6 +12,7 @@ Feature: each quality lens runs on its own model hint
       export { summary, steering, skills, access } from "@pmelab/gtd/workflows/feature"
       export default feature({
         flagMode: false,
+        prototypeRounds: "0",
         qualityReviews: "correctness, conventions, test-audit=opus",
       })
       """

package/src/workflows/feature/lens-review.feature

@@ -9,6 +9,7 @@ Feature: the quality lap's findings are triaged before the review
 
   Background:
     Given the shipped workflows, as gtd install scaffolds them
+    And an environment variable "GTD_PROTOTYPEROUNDS" set to "0"
     And an environment variable "GTD_QUALITYREVIEWS" set to "correctness"
     And an environment variable "GTD_FLAGMODE" set to "false"
     And a file "src/calc.ts" with:

package/src/workflows/feature/one-step-land.feature

@@ -10,6 +10,7 @@ Feature: a naive pass the judge rules land is itself the implementation
 
   Background:
     Given the shipped workflows, as gtd install scaffolds them
+    And an environment variable "GTD_PROTOTYPEROUNDS" set to "0"
     And an environment variable "GTD_QUALITYREVIEWS" set to "correctness"
     And an environment variable "GTD_FLAG_RISK_TIER" set to "0"
     And a file "src/greeting.ts" with:

package/src/workflows/feature/prototype-rounds.feature

@@ -0,0 +1,152 @@
+@inmem
+Feature: past its round cap the hands-on test rests at a person's gate
+
+  `prototypeRounds` caps the fix turns a slice's hands-on test gets. Once they
+  are spent and the report still has open items, the process rests at
+  implement.prototype.open naming them; a person's fix there is tested again
+  with a fresh budget.
+
+  Background:
+    Given the shipped workflows, as gtd install scaffolds them
+    And an environment variable "GTD_QUALITYREVIEWS" set to "correctness"
+    And an environment variable "GTD_FLAGMODE" set to "false"
+    And an environment variable "GTD_PROTOTYPEROUNDS" set to "1"
+    And a file "src/calc.ts" with:
+      """
+      export const add = (a: number, b: number) => a + b
+      """
+    And the working tree is committed as "feat: add calculator"
+    And a file "TODO.md" with:
+      """
+      # Add subtract
+
+      subtract(a, b) returns the difference.
+      """
+    And step "unwind" changes nothing
+    And step "start-gate.check" changes nothing
+    And step "design.triage" writes ".gtd/REQUIREMENTS.md" with:
+      """
+      # Requirements
+
+      ## Acceptance Criteria
+
+      - AC-1: subtract(a, b) returns a - b.
+      - AC-2: add is unchanged.
+      """
+    And step "design.gate.answer" changes nothing
+    And step "seam.naive" writes "src/calc.ts" with:
+      """
+      export const add = (a: number, b: number) => a + b
+      export const subtract = (a: number, b: number) => a + b
+      """
+    And step "seam.naive" writes "src/calc.test.ts" with:
+      """
+      it("[AC-1] subtracts", () => expect(subtract(3, 1)).toBe(2))
+      it("[AC-2] adds", () => expect(add(1, 1)).toBe(2))
+      """
+    And step "seam.check" changes nothing
+    And step "implement.applying" runs its script
+    And step "implement.health.check" changes nothing
+    And judge step "implement.coverage" answers:
+      """
+      [{"id":"AC-1","answer":"yes","p":0.9},{"id":"AC-2","answer":"yes","p":0.9}]
+      """
+    And step "implement.closing" runs its script
+
+  Scenario: past the round cap a person is asked, naming the open items
+    Given step "implement.prototype.testing", when prompted with "Find how this project", writes ".gtd/PROTOTYPE.md" with:
+      """
+      ## AC-1: fail
+
+      `subtract(3, 1)` printed 4. Expected 2.
+
+      ## AC-2: pass
+
+      Printed 2.
+      """
+    And step "implement.prototype.testing", when prompted with "This is a re-test", writes ".gtd/PROTOTYPE.md" with:
+      """
+      ## AC-1: fail
+
+      Re-tested: `subtract(3, 1)` printed -2. Expected 2.
+
+      ## AC-2: pass
+
+      Printed 2.
+      """
+    And step "implement.prototype.fix.fixing" writes "src/calc.ts" with:
+      """
+      export const add = (a: number, b: number) => a + b
+      export const subtract = (a: number, b: number) => b - a
+      """
+    And step "implement.prototype.health.check" changes nothing
+    When the process is driven
+    Then the process rests at "implement.prototype.open"
+    And the resting step's content contains "- AC-1: fail"
+    And the resting step's content contains "after 1 fix round"
+    And the resting step's content does not contain "AC-2"
+
+  Scenario: a person's change at the gate is tested again
+    Given step "implement.prototype.testing", when prompted with "Find how this project", writes ".gtd/PROTOTYPE.md" with:
+      """
+      ## AC-1: blocked
+
+      `subtract` is not exported from any entry point I could run.
+
+      ## AC-2: pass
+
+      Printed 2.
+      """
+    And step "implement.prototype.open" writes "src/calc.ts" with:
+      """
+      export const add = (a: number, b: number) => a + b
+      export const subtract = (a: number, b: number) => a - b
+      """
+    And step "implement.prototype.health.check" changes nothing
+    And step "implement.prototype.testing", when prompted with "This is a re-test", writes ".gtd/PROTOTYPE.md" with:
+      """
+      ## AC-1: pass
+
+      `subtract(3, 1)` printed 2.
+
+      ## AC-2: pass
+
+      Printed 2.
+      """
+    When the process is driven
+    Then the process rests at "build.quality.correctness.reviewing"
+    And the commit subjects end with:
+      """
+      gtd(agent): implement.prototype.testing → implement.prototype.open
+      gtd(human): implement.prototype.open → implement.prototype.health.check
+      gtd(check): implement.prototype.health.check → implement.prototype.testing
+      gtd(agent): implement.prototype.testing → implement.closing
+      gtd(check): implement.closing → build.quality.correctness.reviewing
+      """
+
+  Scenario: a re-test that sees exactly what the last round saw still ends the round
+    Given step "implement.prototype.testing" writes ".gtd/PROTOTYPE.md" with:
+      """
+      ## AC-1: fail
+
+      `subtract(3, 1)` printed 4. Expected 2.
+
+      ## AC-2: pass
+
+      Printed 2.
+      """
+    And step "implement.prototype.fix.fixing" writes "src/calc.ts" with:
+      """
+      export const add = (a: number, b: number) => a + b
+      export const subtract = (a: number, b: number) => b - a
+      """
+    And step "implement.prototype.health.check" changes nothing
+    When the process is driven
+    Then the process rests at "implement.prototype.open"
+    And the commit subjects end with:
+      """
+      gtd(agent): implement.prototype.testing → implement.prototype.fix.fixing
+      gtd(agent): implement.prototype.fix.fixing → implement.prototype.health.check
+      gtd(check): implement.prototype.health.check → implement.prototype.testing
+      gtd(agent): implement.prototype.testing → implement.prototype.open
+      """