@pmelab/gtd 49.0.0 → 51.0.0

package/.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "gtd",
-  "version": "49.0.0",
+  "version": "51.0.0",
   "description": "Drive the gtd loop inside Claude Code: agent turns run as subagents, gtd decides every beat",
   "author": {
     "name": "Philipp Melab"

package/README.md

@@ -538,9 +538,15 @@ noted in steps 3 and 4 below.
    commit) and what was **dropped as noise**. A review with a chunk pointing at
    no code, or a note left out, is refused (`review-pointers`). A pointer note
    the reviewer opens with `Risk:` is fixed before you see the review
-   (`build.review.fix.risks.fixing`), then the review is written again. On
-   `refactor-first` the refactor slice gets a review of its own first, composed
-   by gtd from its lap's notes, before the feature is built; see
+   (`build.review.fix.risks.fixing`), then the review is written again. **What
+   matters most comes first:** a judge (`build.review.ranking`) rates every
+   chunk on the four severity questions below, and the review lists `critical`
+   chunks first (security, then one-way doors), then `should`, then `could`,
+   each led by its tier and reason (`_critical: security, one-way door_`) — the
+   phone shows it on the chunk's card. A chunk the judge is unsure of, or whose
+   evidence the budget cut, is `critical`. On `refactor-first` the refactor
+   slice gets a walkthrough of its own first, covering the whole refactor on
+   every round, before the feature is built; see
    [the naive pass](#the-naive-pass). **Every note needs exactly one tick before
    `gtd land` accepts the review** (`answer-completeness`; an open `- H:` thread
    waiting on the agent is the one exception). Fix as suggested joins the nit
@@ -687,6 +693,13 @@ an uncovered one goes back to the agent (twice, then to you) instead of ending
 the feature, and each covered ID is recorded as a `Gtd-Mark: covers=<id>`
 trailer. See
 [Acceptance-criteria coverage](docs/configuration.md#acceptance-criteria-coverage).
+Then an agent uses the slice by hand — the real program in a scratch directory,
+the real page, guided by a project skill or a manual testing section in
+`AGENTS.md`, `CLAUDE.md` or `README.md` — and writes `.gtd/PROTOTYPE.md`: per
+criterion pass, fail or blocked, with the steps run and what it saw. Failures
+are fixed and re-tested before the quality lap, up to `prototypeRounds` (3) fix
+rounds; past them, with only blocked items, or with no way to test found, it
+rests for you. See [The hands-on test](docs/configuration.md#the-hands-on-test).
 Feedback at the feature review re-runs design and the implementation, never the
 naive pass; feedback at the refactor review re-runs the refactor, never design;
 a refactor-review change the refactor cannot build is refused, and the only way

package/dist/chunks/{reserved-CYZ1bSBy.mjs → reserved-CUkk9XgL.mjs}

@@ -1,5 +1,5 @@
 import { accessFor, agent, answered, changesSince, check, codeThreads, defineWorkflow, displayPath, env, head, human, judge, mark, numeric, quote, read, refuse, scope, skillsFor, start, vars } from "../flows.mjs";
-import { d as ANSWER_FIX, f as ANSWER_LEAD, m as reviewChunks, p as ANSWER_WONT_FIX, y as FREE_TEXT_PLACEHOLDER } from "./steering-BXwjmocB.mjs";
+import { b as FREE_TEXT_PLACEHOLDER, d as ANSWER_FIX, f as ANSWER_LEAD, m as reviewChunks, p as ANSWER_WONT_FIX, v as severityLine } from "./steering-DtUVUz9l.mjs";
 import { gate as gate$1, healthy as healthy$1, lensScopes, lensesOf, removedAssertions } from "../workflow-kit.mjs";
 import { c as String$, g as standardSchemaV1, h as optionalWith, l as Struct, n as Boolean$, s as Number$ } from "./Schema-C6fLlQzx.mjs";
 //#region src/workflows/shared/registry.ts
@@ -243,6 +243,7 @@ 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. */
@@ -320,6 +321,11 @@ 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
@@ -641,6 +647,56 @@ const withWalkthrough = (review, report) => {
 	const body = lines.join("\n").trimEnd();
 	return summary.length === 0 ? `${body}\n` : `${body}\n\n${summary.join("\n\n")}\n`;
 };
+const chunkBlocks = (review) => {
+	const lines = review.split("\n");
+	const chunks = reviewChunks(review);
+	return {
+		head: lines.slice(0, chunks[0]?.line ?? lines.length),
+		blocks: chunks.map((c, i) => ({
+			title: c.title,
+			lines: lines.slice(c.line, chunks[i + 1]?.line ?? lines.length)
+		}))
+	};
+};
+const isBlank = (line) => line !== void 0 && line.trim() === "";
+/** \`review\` with every chunk's severity line removed, so a re-review is ranked afresh. */
+const unranked = (review) => {
+	const drop = new Set(reviewChunks(review).flatMap((c) => c.severityLine ?? []));
+	if (drop.size === 0) return review;
+	const lines = review.split("\n");
+	return lines.filter((_, i) => !drop.has(i) && !(drop.has(i - 1) && isBlank(lines[i]))).join("\n");
+};
+const chunkTexts = (review) => new Map(chunkBlocks(review).blocks.map((b) => [b.title, b.lines.join("\n").trim()]));
+const trimBlankLines = (lines) => {
+	let start = 0;
+	let end = lines.length;
+	while (start < end && isBlank(lines[start])) start++;
+	while (end > start && isBlank(lines[end - 1])) end--;
+	return lines.slice(start, end);
+};
+/**
+* \`review\` with the \`ranked\` chunks first, in their order, each led by its
+* severity line; every other chunk follows in place.
+*/
+const rankedReview = (review, ranked) => {
+	const { head, blocks } = chunkBlocks(review);
+	const place = new Map(ranked.map((r, k) => [r.title, k]));
+	const labels = new Map(ranked.map((r) => [r.title, r.label]));
+	const order = (i) => place.get(blocks[i].title) ?? ranked.length;
+	const chunk = (i) => {
+		const { title, lines } = blocks[i];
+		const label = labels.get(title);
+		const heading = /^ {0,3}-+\s*$/.test(lines[1] ?? "") ? 2 : 1;
+		const body = trimBlankLines(lines.slice(heading)).join("\n");
+		return [
+			lines.slice(0, heading).join("\n"),
+			...label === void 0 ? [] : [severityLine(label)],
+			...body === "" ? [] : [body]
+		].join("\n\n");
+	};
+	const sorted = blocks.map((_, i) => i).sort((x, y) => order(x) - order(y) || x - y);
+	return `${[head.join("\n").trim(), ...sorted.map(chunk)].filter((p) => p !== "").join("\n\n")}\n`;
+};
 /**
 * \`.gtd/REVIEW.md\` as gtd writes it: one chunk per note, each ending in its
 * answer boxes, then the fixed and noise lists. Chunk titles are made unique —
@@ -695,8 +751,11 @@ 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
@@ -804,7 +863,7 @@ 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
@@ -1195,6 +1254,13 @@ const wrapItem = (item) => {
 	} else line = line === "" ? word : `${line} ${word}`;
 	return [...out, line];
 };
+const reviewRankingMessage = () => `Ranking each chunk of the review before the human reads it. Run
+\`gtd judge answer\` and pipe, per chunk \`C<n>\`: \`C<n>.security\`
+(yes/no), \`C<n>.reversibility\` (one-way/two-way),
+\`C<n>.blastRadius\` (local/module/contract) and \`C<n>.newConcept\`
+(yes/no) — or land without a verdict, which lists every chunk as
+critical.
+`;
 const buildReviewTriageMessage = () => `Judging each note the human added to \`.gtd/REVIEW.md\` (a line note, a
 footnote, or prose under a chunk) with one verdict: \`edit\` (a change
 request), \`question\`, \`nit\` (a small fix needing no re-plan) or
@@ -1716,6 +1782,80 @@ Two ways forward (then run \`gtd land\`):
 - **Overrule** — land a clean tree to finish without them. The slice records
   the IDs it covers and, as \`uncovered\`, the ones it does not
 `;
+const testerSystem = () => `${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";
@@ -1727,6 +1867,7 @@ 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), {
@@ -1971,6 +2112,7 @@ const ShippedOptions = Struct({
 	qualityReviews: setting,
 	refactorQualityReviews: setting,
 	continuations: setting,
+	prototypeRounds: setting,
 	continuationStall: setting,
 	testCommand: setting,
 	plannerModel: setting,
@@ -2050,6 +2192,10 @@ 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])
@@ -2074,4 +2220,4 @@ const steering = {
 	[REVIEW]: "review"
 };
 //#endregion
-export { naivePassPrompt as $, reviewQuality as A, conflictBody as At, coverageQuestion as B, flipProblems as Bt, collecting as C, walkthroughBlock as Ct, fixRisks as D, envDefaults as Dt, fixQuality as E, defaults as Et, builderSystem as F, parseStandards as Ft, designTriagePrompt as G, criterionLine as H, requireFlag as Ht, clearCriteriaMessage as I, renderStandards as It, handOffAwaitReviewMessage as J, finisherSystem as K, clearCriteriaQuestion as L, standardKey as Lt, agentWithSkills as M, conflictTitle as Mt, buildReviewCollectingPrompt as N, isConflictTitle as Nt, planner as O, testFileGlobs as Ot, buildReviewTriageMessage as P, oneLine as Pt, naiveApplyFailure as Q, composedReview as R, flagMentions as Rt, coder as S, unwindFailure as St, fixNits as T, withWalkthrough as Tt, designGateAnswerMessage as U, coveringPrompt as V, registryEntry as Vt, designSystem as W, implementBuildingPrompt as X, idleMessage as Y, implementFixPrompt as Z, REQUIREMENTS as _, slicingPrompt as _t, flagModeOn as a, refactorDroppedRefusal as at, answerReviewQuestions as b, uncoveredMessage as bt, baseline as c, refactorRules as ct, healthy as d, reviewFeedbackFixPrompt as dt, normaliseCriteriaScript as et, FEEDBACK as f, reviewGateBlockedMessage as ft, REFACTOR_NOTE as g, seamSteerMessage as gt, REFACTOR as h, seamByHandMessage as ht, skills as i, refactorBuildingPrompt as it, reviewing as j, conflictFor as jt, reviewMissing as k, STANDARDS as kt, gate as l, reviewEditNotesCapture as lt, QUALITY as m, reviewerSystem as mt, summary as n, qualityFindingsFile as nt, flagShipped as o, refactorFixPrompt as ot, HANDOFF as p, reviewNotesCapture as pt, fixReviewNotesPrompt as q, access as r, qualityTriageMessage as rt, shipped as s, refactorNotePrompt as st, steering as t, oversizedMessage as tt, guarded as u, reviewEditsCapture as ut, REVIEW as v, startGateBlockedMessage as vt, fixCheck as w, withLapFindings as wt, awaitReview as x, unwindFailedMessage as xt, SLICES as y, titledNotes as yt, coverageMessage as z, flagReferences as zt };
+export { implementFixPrompt as $, reviewMissing as A, unranked as At, composedReview as B, conflictBody as Bt, coder as C, seamByHandMessage as Ct, fixQuality as D, testerSystem as Dt, fixNits as E, startGateBlockedMessage as Et, buildReviewTriageMessage as F, withWalkthrough as Ft, designGateAnswerMessage as G, parseStandards as Gt, coverageQuestion as H, conflictTitle as Ht, builderSystem as I, defaults as It, finisherSystem as J, flagMentions as Jt, designSystem as K, renderStandards as Kt, chunkTexts as L, envDefaults as Lt, reviewing as M, unwindFailure as Mt, agentWithSkills as N, walkthroughBlock as Nt, fixRisks as O, titledNotes as Ot, buildReviewCollectingPrompt as P, withLapFindings as Pt, implementBuildingPrompt as Q, requireFlag as Qt, clearCriteriaMessage as R, testFileGlobs as Rt, awaitReview as S, reviewerSystem as St, fixCheck as T, slicingPrompt as Tt, coveringPrompt as U, isConflictTitle as Ut, coverageMessage as V, conflictFor as Vt, criterionLine as W, oneLine as Wt, handOffAwaitReviewMessage as X, flipProblems as Xt, fixReviewNotesPrompt as Y, flagReferences as Yt, idleMessage as Z, registryEntry as Zt, REFACTOR_NOTE as _, reviewEditsCapture as _t, flagModeOn as a, prototypeOpenMessage as at, SLICES as b, reviewNotesCapture as bt, baseline as c, qualityFindingsFile as ct, healthy as d, refactorBuildingPrompt as dt, naiveApplyFailure as et, FEEDBACK as f, refactorDroppedRefusal as ft, REFACTOR as g, reviewEditNotesCapture as gt, QUALITY as h, refactorRules as ht, skills as i, prototypeFixingPrompt as it, reviewQuality as j, unwindFailedMessage as jt, planner as k, uncoveredMessage as kt, gate as l, qualityTriageMessage as lt, PROTOTYPE as m, refactorNotePrompt as mt, summary as n, normaliseCriteriaScript as nt, flagShipped as o, prototypeTestingPrompt as ot, HANDOFF as p, refactorFixPrompt as pt, designTriagePrompt as q, standardKey as qt, access as r, oversizedMessage as rt, shipped as s, prototypeUnguidedMessage as st, steering as t, naivePassPrompt as tt, guarded as u, rankedReview as ut, REQUIREMENTS as v, reviewFeedbackFixPrompt as vt, collecting as w, seamSteerMessage as wt, answerReviewQuestions as x, reviewRankingMessage as xt, REVIEW as y, reviewGateBlockedMessage as yt, clearCriteriaQuestion as z, STANDARDS as zt };

package/dist/chunks/{review-BewyJflS.mjs → review-BaCXlIT2.mjs}

@@ -1,8 +1,8 @@
 import { answered, changesSince, head, judge, numeric, read, refuse, requireRevert, restoreScript, run, showScript, vars, writeScript } from "../flows.mjs";
-import { _ as reviewRisks, m as reviewChunks } from "./steering-BXwjmocB.mjs";
+import { _ as reviewRisks, m as reviewChunks } from "./steering-DtUVUz9l.mjs";
 import { removedAssertions, review } from "../workflow-kit.mjs";
-import { At as conflictBody, C as collecting, Ct as walkthroughBlock, D as fixRisks, Ft as parseStandards, It as renderStandards, Lt as standardKey, Mt as conflictTitle, Nt as isConflictTitle, P as buildReviewTriageMessage, Pt as oneLine, R as composedReview, T as fixNits, Tt as withWalkthrough, b as answerReviewQuestions, d as healthy, j as reviewing, jt as conflictFor, k as reviewMissing, kt as STANDARDS, lt as reviewEditNotesCapture, pt as reviewNotesCapture, u as guarded, ut as reviewEditsCapture, v as REVIEW, w as fixCheck, x as awaitReview, yt as titledNotes } from "./reserved-CYZ1bSBy.mjs";
-import { i as isOptionalObject, n as closeReview, o as tail, s as NO_FINDINGS } from "./tail-C1Oi2YEy.mjs";
+import { A as reviewMissing, At as unranked, B as composedReview, Bt as conflictBody, E as fixNits, F as buildReviewTriageMessage, Ft as withWalkthrough, Gt as parseStandards, Ht as conflictTitle, Kt as renderStandards, L as chunkTexts, M as reviewing, Nt as walkthroughBlock, O as fixRisks, Ot as titledNotes, S as awaitReview, T as fixCheck, Ut as isConflictTitle, Vt as conflictFor, Wt as oneLine, _t as reviewEditsCapture, bt as reviewNotesCapture, d as healthy, gt as reviewEditNotesCapture, qt as standardKey, u as guarded, ut as rankedReview, w as collecting, x as answerReviewQuestions, xt as reviewRankingMessage, y as REVIEW, zt as STANDARDS } from "./reserved-CUkk9XgL.mjs";
+import { i as isOptionalObject, n as closeReview, o as tail, s as NO_FINDINGS } from "./tail-CX6Pdsql.mjs";
 //#region src/workflows/shared/learn.ts
 const SUMMARY_MAX = 200;
 const BANNER = /^_(?:The fixer left this unresolved\.|A fix turn removed this while renaming another finding — check whether it was fixed\.|No detail given\.)_\s*/;
@@ -205,7 +205,8 @@ const severityOf = (note, judgment, minP) => {
 	if (critical.length > 0) return {
 		tier: "critical",
 		reason: `critical: ${critical.join(", ")}`,
-		rank: rank()
+		rank: rank(),
+		...a("security") === "yes" ? { security: true } : {}
 	};
 	const should = [...a("blastRadius") === "contract" ? ["contract"] : [], ...a("newConcept") === "yes" ? ["new concept"] : []];
 	if (should.length > 0) return {
@@ -219,29 +220,30 @@ const severityOf = (note, judgment, minP) => {
 		rank: 2
 	};
 };
-const severityQuestions = (note) => [
+/** `asks` names the change in full, `change` refers back to it. */
+const severityQuestions = (id, asks, change) => [
 	{
-		id: `${note.id}.security`,
+		id: `${id}.security`,
 		primitive: "noul",
-		instructions: `Does the change the human asks for in note "${note.id}" (on "${note.anchor}") touch security — authentication, authorisation, secrets, untrusted input or data exposure?`,
+		instructions: `Does ${asks} touch security — authentication, authorisation, secrets, untrusted input or data exposure?`,
 		criteria: "yes: it changes or could weaken a security property. no: it does not."
 	},
 	{
-		id: `${note.id}.reversibility`,
+		id: `${id}.reversibility`,
 		primitive: "choice",
-		instructions: `Once the change in note "${note.id}" ships, can it be undone without breaking anyone who already depends on it?`,
+		instructions: `Once ${change} ships, can it be undone without breaking anyone who already depends on it?`,
 		criteria: "one-way: a published API, file format, persisted data, protocol or other contract that cannot be taken back. two-way: can be reverted freely."
 	},
 	{
-		id: `${note.id}.blastRadius`,
+		id: `${id}.blastRadius`,
 		primitive: "choice",
-		instructions: `How far does the change in note "${note.id}" reach?`,
+		instructions: `How far does ${change} reach?`,
 		criteria: "local: one function or file. module: a whole module or several files. contract: a public interface other code or users depend on."
 	},
 	{
-		id: `${note.id}.newConcept`,
+		id: `${id}.newConcept`,
 		primitive: "noul",
-		instructions: `Does the change in note "${note.id}" introduce a concept new to this codebase — a new abstraction, entity or term? Judge from the evidence, never a glossary; reusing or extending an existing concept is not new.`,
+		instructions: `Does ${change} introduce a concept new to this codebase — a new abstraction, entity or term? Judge from the evidence, never a glossary; reusing or extending an existing concept is not new.`,
 		criteria: "yes: a genuinely new concept. no: reuses or extends an existing one."
 	}
 ];
@@ -258,7 +260,7 @@ const triage = (severities) => async (notes) => {
 		answers: {},
 		truncated: []
 	} : await judge("review.triage", {
-		questions: asked.flatMap((n) => [verdictQuestion(n), ...severityQuestions(n)]),
+		questions: asked.flatMap((n) => [verdictQuestion(n), ...severityQuestions(n.id, `the change the human asks for in note "${n.id}" (on "${n.anchor}")`, `the change in note "${n.id}"`)]),
 		evidence: Object.fromEntries(asked.map((n) => [n.id, `Anchor: ${n.anchor}\nReviewer's text: ${n.before}\nHuman's text: ${n.text}`])),
 		message: buildReviewTriageMessage(),
 		label: "Judging each note"
@@ -274,6 +276,67 @@ const triage = (severities) => async (notes) => {
 		return [n.id, "edit"];
 	}));
 };
+/** Review order: security, one-way doors, the judge unsure, then should, then could. */
+const chunkKey = (s) => s.tier === "could" ? 4 : s.tier === "should" ? 3 : s.security === true ? 0 : s.rank === 0 ? 1 : 2;
+/** A chunk's text, then the code each of its pointers covers: a bare path the whole file. */
+const chunkEvidence = (text, files) => [text, ...files.map(({ path, line, rangeEnd }) => {
+	const where = `${path}${line === void 0 ? "" : `#${line}${rangeEnd === void 0 ? "" : `-${rangeEnd}`}`}`;
+	const lines = read(path.replace(/^\.\//, ""))?.split("\n");
+	if (lines === void 0) return `${where}: no longer exists`;
+	return `${where}:\n${(line === void 0 ? lines : lines.slice(line - 1, rangeEnd ?? line)).join("\n")}`;
+})].join("\n\n");
+const bytes = (text) => new TextEncoder().encode(text).length;
+/** Fills each batch while the budget's even share still holds its widest evidence, so nothing is cut that fits alone. */
+const batched = (items, budget) => {
+	const batches = [];
+	let batch = [];
+	let widest = 0;
+	for (const item of items) {
+		const size = bytes(item.evidence);
+		if (batch.length > 0 && Math.floor(budget / (batch.length + 1)) < Math.max(widest, size)) {
+			batches.push(batch);
+			batch = [];
+			widest = 0;
+		}
+		batch.push(item);
+		widest = Math.max(widest, size);
+	}
+	return batch.length === 0 ? batches : [...batches, batch];
+};
+/**
+* Rates every chunk the reviewer wrote — a walkthrough chunk or a carried
+* finding — and sorts \`full\` by it, each chunk led by its tier and reason.
+* gtd's summary chunks are not rated and stay last.
+*/
+const ranked = async (review, full) => {
+	const texts = chunkTexts(full);
+	const chunks = reviewChunks(review).map((c, i) => ({
+		id: `C${i + 1}`,
+		title: c.title,
+		evidence: chunkEvidence(texts.get(c.title) ?? "", c.files)
+	}));
+	if (chunks.length === 0) return full;
+	const minP = numeric(vars.reviewSeverityMinP, Infinity);
+	const batches = batched(chunks, numeric(vars.judgeBudgetBytes, 32768));
+	const severities = /* @__PURE__ */ new Map();
+	for (const [i, batch] of batches.entries()) {
+		const judgment = await judge("review.ranking", {
+			questions: batch.flatMap((c) => severityQuestions(c.id, `the change review chunk "${c.id}" ("${c.title}") walks through`, `the change in review chunk "${c.id}"`)),
+			evidence: Object.fromEntries(batch.map((c) => [c.id, c.evidence])),
+			message: reviewRankingMessage(),
+			label: batches.length === 1 ? "Ranking the review" : `Ranking the review (${i + 1}/${batches.length})`
+		});
+		for (const c of batch) severities.set(c.id, severityOf(c, judgment, minP));
+	}
+	return rankedReview(full, chunks.map((c, i) => ({
+		c,
+		i,
+		s: severities.get(c.id)
+	})).sort((x, y) => chunkKey(x.s) - chunkKey(y.s) || x.i - y.i).map(({ c, s }) => ({
+		title: c.title,
+		label: s.reason
+	})));
+};
 const isCode = (path) => path !== ".gtd" && !path.startsWith(".gtd/");
 const composedOf = (base, report, nitFix, answered) => ({
 	...report,
@@ -325,12 +388,13 @@ const requirePointers = (review, notes, codeChanged) => {
 const walkthrough = async (base, report, lap, carry) => {
 	const notes = titledNotes(report);
 	await reviewing(base, carry, lap, walkthroughBlock(notes, report.nitFix));
-	const review = read(REVIEW);
-	if (review === void 0) return;
-	requirePointers(review, notes, changesSince(base).some((c) => isCode(c.path)));
-	const full = withWalkthrough(review, report);
-	if (full !== review) await run("review.carrying", writeScript(REVIEW, full), {
-		label: "Adding the quality lap's findings",
+	const written = read(REVIEW);
+	if (written === void 0) return;
+	const review = unranked(written);
+	requirePointers(written, notes, changesSince(base).some((c) => isCode(c.path)));
+	const full = await ranked(review, withWalkthrough(review, report));
+	if (full !== written) await run("review.carrying", writeScript(REVIEW, full), {
+		label: "Ordering the review and adding the quality lap's findings",
 		file: REVIEW,
 		base
 	});

package/dist/chunks/{steering-BXwjmocB.mjs → steering-DtUVUz9l.mjs}

@@ -15089,6 +15089,9 @@ const qaDescriptor = {
 const ANSWER_LEAD = "Tick exactly one:";
 const ANSWER_FIX = "fix as suggested";
 const ANSWER_WONT_FIX = "won't fix";
+/** A chunk's leading `_critical: security_` paragraph: the tier gtd ranked it with, and why. */
+const SEVERITY_LINE_RE = /^_((?:critical|should|could)(?::[^_\n]*)?)_$/;
+const severityLine = (label) => `_${label}_`;
 /**
 * The canonical sample. Every detail here is load-bearing for the formatter
 * round-trip: both footnote bodies exceed 80 characters so a reflow actually
@@ -15379,12 +15382,26 @@ const answerListOf = (content, body) => {
 		})
 	};
 };
+/** A leading severity line is the chunk's tier, never its prose. */
+const splitSeverity = (content, run) => {
+	const first = run[0];
+	const label = first?.type === "paragraph" ? SEVERITY_LINE_RE.exec(sourceText(content, first).trim())?.[1] : void 0;
+	if (label === void 0) return { leadingRun: run };
+	return {
+		severity: {
+			label,
+			line: toLspPosition(first.position.start).line
+		},
+		leadingRun: run.slice(1)
+	};
+};
 /** Splits one chunk's body nodes into its file pointers (hunk pointers are task items, collected recursively at ANY nesting depth via `taskItems` — a nested hunk is the same kind of hunk as a top-level one) and description prose. The description is the chunk's own leading run and never a node that contains a hunk pointer: it stops at the first node whose `taskItems` include a real pointer (at any depth — a blockquote wrapping a pointer list stops it just as a top-level list would). Every node kind in what's left before that point contributes — a `footnoteDefinition` is the one exception, excluded by `blockNodesOfRun` itself (it's a note, never document content). */
 const parseChunkBody = (content, lines, title, body) => {
 	const answer = answerListOf(content, body);
 	const isAnswer = (n) => answer !== void 0 && (n === answer.list || n === body[answer.lead]);
 	const firstPointerIndex = body.findIndex((n) => !isAnswer(n) && taskItems(n).some((item) => hasPointerToken(content, item)));
-	const leadingRun = (firstPointerIndex === -1 ? body : body.slice(0, firstPointerIndex)).filter((n) => !isAnswer(n));
+	const run = (firstPointerIndex === -1 ? body : body.slice(0, firstPointerIndex)).filter((n) => !isAnswer(n));
+	const { severity, leadingRun } = splitSeverity(content, run);
 	const descriptionNodes = blockNodesOfRun(content, leadingRun);
 	const descriptionInline = collapseImages(blockRunInline(content, leadingRun));
 	const description = leadingRun.filter((n) => n.type !== "footnoteDefinition").map((n) => sourceText(content, n)).join(" ").trim();
@@ -15403,6 +15420,7 @@ const parseChunkBody = (content, lines, title, body) => {
 		files,
 		errors,
 		...answer === void 0 ? {} : { answers: answer.options },
+		...severity === void 0 ? {} : { severity },
 		boxed: body.some((n) => !isAnswer(n) && taskItems(n).length > 0)
 	};
 };
@@ -15439,7 +15457,7 @@ const parseChangesets = (tree, content, lines) => {
 	const errors = [];
 	for (const { heading, headingLine, body } of splitChunks(tree)) {
 		const title = headingText$1(content, heading);
-		const { description, descriptionNodes, descriptionInline, files, answers, boxed, errors: bodyErrors } = parseChunkBody(content, lines, title, body);
+		const { description, descriptionNodes, descriptionInline, files, answers, severity, boxed, errors: bodyErrors } = parseChunkBody(content, lines, title, body);
 		errors.push(...bodyErrors);
 		if (files.length === 0 && boxed) errors.push({
 			message: `Chunk "${title}" has no file pointers`,
@@ -15453,7 +15471,8 @@ const parseChangesets = (tree, content, lines) => {
 			descriptionInline,
 			files,
 			headingLine,
-			...answers === void 0 ? {} : { answers }
+			...answers === void 0 ? {} : { answers },
+			...severity === void 0 ? {} : { severity }
 		});
 	}
 	if (changesets.length === 0) errors.push({ message: "REVIEW.md has no '##' chunks" });
@@ -15754,6 +15773,7 @@ const reviewView = (content) => {
 			detail: chunk.description,
 			detailInline: chunk.descriptionInline,
 			body: chunk.descriptionNodes,
+			...chunk.severity === void 0 ? {} : { status: chunk.severity.label },
 			anchor: {
 				kind: "chunk",
 				index: chunkIndex
@@ -15901,15 +15921,19 @@ const reviewDescriptor = {
 	openThreads: openThreadsOf
 };
 const pointerLabel = (f) => `${f.path}${f.line === void 0 ? "" : `#${f.line}${f.rangeEnd === void 0 ? "" : `-${f.rangeEnd}`}`}`;
-const chunkNote = (chunk, prior) => chunk.description === (prior?.description ?? "") || prior === void 0 && chunk.answers !== void 0 ? [] : [{
-	line: chunk.headingLine,
-	note: {
-		kind: "chunk",
-		anchor: chunk.title,
-		before: prior?.description ?? "",
-		text: chunk.description
-	}
-}];
+const chunkProse = (chunk, retiered) => [chunk?.description ?? "", retiered && chunk?.severity !== void 0 ? severityLine(chunk.severity.label) : ""].filter((p) => p !== "").join("\n\n");
+const chunkNote = (chunk, prior) => {
+	const retiered = chunk.severity !== void 0 && chunk.severity.label !== prior?.severity?.label;
+	return chunkProse(chunk, retiered) === chunkProse(prior, retiered) || prior === void 0 && chunk.answers !== void 0 ? [] : [{
+		line: chunk.headingLine,
+		note: {
+			kind: "chunk",
+			anchor: chunk.title,
+			before: chunkProse(prior, retiered),
+			text: chunkProse(chunk, retiered)
+		}
+	}];
+};
 const pointerNotes = (chunk, prior) => chunk.files.flatMap((file) => {
 	const label = pointerLabel(file);
 	const was = prior?.files.find((f) => pointerLabel(f) === label);
@@ -16023,7 +16047,13 @@ const reviewDecisions = (before, after) => {
 const reviewChunks = (content) => parseReviewDoc(content).changesets.map((c) => ({
 	title: c.title,
 	line: c.headingLine,
-	pointers: c.files.length
+	pointers: c.files.length,
+	...c.severity === void 0 ? {} : { severityLine: c.severity.line },
+	files: c.files.map((f) => ({
+		path: f.path,
+		...f.line === void 0 ? {} : { line: f.line },
+		...f.rangeEnd === void 0 ? {} : { rangeEnd: f.rangeEnd }
+	}))
 }));
 //#endregion
 //#region src/steering/Edits.ts
@@ -16393,4 +16423,4 @@ const viewOf = (format, content) => ({
 /** Every OPEN thread — last entry is the agent's — as `{ name, line, firstMe }` (0-based line). `[]` for a format with no thread concept. */
 const openThreads = (format, content) => format.openThreads?.(content) ?? [];
 //#endregion
-export { formatFinding as A, openThreadFindings as C, getParseCount as D, parseThreadsWithFindings as E, headingSectionBodies as O, THREAD_REPLY_ACTION_TITLE as S, parseThreads as T, reviewRisks as _, unansweredQuestions as a, isAnswered as b, stripCodeThreads as c, ANSWER_FIX as d, ANSWER_LEAD as f, reviewNotes as g, reviewDecisions as h, steeringFormatOrFreeForm as i, headingSections as k, freeFormFormat as l, reviewChunks as m, openThreads as n, viewOf as o, ANSWER_WONT_FIX as p, steeringFormatFor as r, parseCodeThreads as s, BUILT_IN_MODE_NAMES as t, applySteeringEdits as u, unansweredNotes as v, parseFootnotes as w, FOOTNOTE_ACTION_TITLE as x, FREE_TEXT_PLACEHOLDER as y };
+export { headingSections as A, THREAD_REPLY_ACTION_TITLE as C, parseThreadsWithFindings as D, parseThreads as E, getParseCount as O, FOOTNOTE_ACTION_TITLE as S, parseFootnotes as T, reviewRisks as _, unansweredQuestions as a, FREE_TEXT_PLACEHOLDER as b, stripCodeThreads as c, ANSWER_FIX as d, ANSWER_LEAD as f, reviewNotes as g, reviewDecisions as h, steeringFormatOrFreeForm as i, formatFinding as j, headingSectionBodies as k, freeFormFormat as l, reviewChunks as m, openThreads as n, viewOf as o, ANSWER_WONT_FIX as p, steeringFormatFor as r, parseCodeThreads as s, BUILT_IN_MODE_NAMES as t, applySteeringEdits as u, severityLine as v, openThreadFindings as w, isAnswered as x, unansweredNotes as y };

package/dist/chunks/{tail-C1Oi2YEy.mjs → tail-CX6Pdsql.mjs}

@@ -1,6 +1,6 @@
 import { answered, head, judge, mark, numeric, read, refuse, removeScript, run, scope, sectionBodies, showScript, vars, writeScript } from "../flows.mjs";
 import { lensEntriesOf, lensOf, qualityLap, reviewToHandOff } from "../workflow-kit.mjs";
-import { A as reviewQuality, D as fixRisks, E as fixQuality, J as handOffAwaitReviewMessage, d as healthy, j as reviewing, k as reviewMissing, m as QUALITY, nt as qualityFindingsFile, rt as qualityTriageMessage, u as guarded, v as REVIEW, w as fixCheck, wt as withLapFindings, x as awaitReview } from "./reserved-CYZ1bSBy.mjs";
+import { A as reviewMissing, D as fixQuality, M as reviewing, O as fixRisks, Pt as withLapFindings, S as awaitReview, T as fixCheck, X as handOffAwaitReviewMessage, ct as qualityFindingsFile, d as healthy, h as QUALITY, j as reviewQuality, lt as qualityTriageMessage, u as guarded, y as REVIEW } from "./reserved-CUkk9XgL.mjs";
 //#region src/workflows/shared/lenses.ts
 const NO_FINDINGS = {
 	notes: [],