@galda/cli 0.10.98 → 0.10.100

package/app/index.html

@@ -1389,6 +1389,9 @@
      Was a pulsing dot (fsrundot). Centre WITHOUT transform (spin's rotate would override a centring
      translate → drift bottom-right) via top/left:50% + negative-half margin; concentric with the 4px dot. */
   #fsRoot .fsproj.run i{background:none;box-shadow:none;position:relative}
+  /* .fsproj.on i and .fsproj.run i have equal specificity. Keep the running
+     ring visible when the currently selected project is also running. */
+  #fsRoot .fsproj.on.run i{background:none;box-shadow:none}
   /* Review-blue comparison baseline: the Doing cue is a 7px / 1.3px broken
      ring, while idle and Review remain 4px dots. */
   #fsRoot .fsproj.run i::after{content:"";position:absolute;top:50%;left:50%;width:7px;height:7px;margin:-3.5px 0 0 -3.5px;box-sizing:border-box;border-radius:50%;border:1.3px solid var(--green);border-top-color:transparent;animation:spin 1.8s linear infinite}
@@ -1920,7 +1923,7 @@
     box-shadow:inset 0 0 0 1px rgba(var(--green-rgb),.30);font:600 12px/1 var(--ui);cursor:pointer;display:inline-flex;align-items:center;gap:7px}
   .sa-app:hover{box-shadow:inset 0 0 0 1px var(--green);background:rgba(var(--green-rgb),.12)}
   .sa-app svg{width:12px;height:12px;fill:none;stroke:currentColor;stroke-width:2.4;stroke-linecap:round;stroke-linejoin:round}
-  .sa-approve-wrap{display:flex;flex-direction:column;align-items:flex-start;gap:5px;flex:0 0 auto}
+  .sa-approve-wrap{display:flex;flex-direction:column;align-items:center;gap:5px;flex:0 0 auto}
   .sa-approve-wrap + .sa-dis{align-self:flex-start}
   .sa-delivery{font:500 9.5px/1.15 var(--mono);color:var(--ink3);text-decoration:none;white-space:nowrap}
   a.sa-delivery{color:var(--green)}
@@ -3472,6 +3475,28 @@ function restoreFocusedAnswerInput(container, attr, saved){
   el.focus();
   try { el.setSelectionRange(saved.start, saved.end); } catch { /* input type may not support it */ }
 }
+// The capture/restore pair above still loses an in-progress Japanese/Chinese/Korean
+// IME conversion: rebuilding the DOM node while the browser has an open composition
+// session aborts that session outright, so the half-typed conversion is dropped even
+// though the plain .value is restored. Track composition state at the document level
+// (delegated, since the actual input node gets destroyed and recreated every rebuild)
+// and let callers skip the destructive innerHTML rebuild while composing, retrying
+// once composition ends.
+let composingAnswerInput = null; // { attr, id } while a qansinput/fsansinput IME session is open
+document.addEventListener('compositionstart', (e) => {
+  const attr = e.target?.getAttribute?.('data-qansinput') != null ? 'data-qansinput'
+    : e.target?.getAttribute?.('data-fsansinput') != null ? 'data-fsansinput' : null;
+  if (attr) composingAnswerInput = { attr, id: e.target.getAttribute(attr) };
+});
+document.addEventListener('compositionend', (e) => {
+  if (!composingAnswerInput) return;
+  if (e.target?.getAttribute?.(composingAnswerInput.attr) !== composingAnswerInput.id) return;
+  composingAnswerInput = null;
+  // Catch up on whatever rebuild(s) were skipped mid-composition.
+  if (typeof renderTasks === 'function') renderTasks();
+  if (typeof renderFsFeed === 'function') renderFsFeed();
+});
+function answerInputComposing(attr){ return composingAnswerInput?.attr === attr; }
 
 const STATUS_LABEL = {
   sending: 'Sending…', stacked: 'Queued', planning: 'Planning…', queued: 'To do', running: 'Working…', needsInput: 'Needs input', needsApproval: 'Needs you',
@@ -4804,6 +4829,7 @@ function renderTasks(){
      <div class="cond">shelved — press Start to queue it</div></div>
      <button type="button" class="prmini pstart" data-pstart="${g.id}">Start</button>
      <button type="button" class="pdel" data-pdel="${g.id}" title="Delete">✕</button></div>`).join('');
+  if (answerInputComposing('data-qansinput')) return; // mid-IME conversion — rebuilding now would abort it; compositionend re-runs this
   const savedQAns = captureFocusedAnswerInput($('tasklist'), 'data-qansinput');
   $('tasklist').innerHTML =
     sec('Needs you', needsYou, 'needsyou')
@@ -6532,8 +6558,11 @@ function renderFsRail(){
     // project the To Do/Kanban view showed as Doing kept a static dot in the rail
     // (Masa report 2026-07-27: "全部doingなのに緑のアイコンが回ってない"). Match the
     // same bucket the board/list already use, minus Paused (intentionally static).
-    const running = pgoals.some((g) => fsBoardBucket(g, ptasks) === 'doing'
-      && !(g.status === 'blocked' && g.blocked?.kind === 'paused'));
+    // A live task is the strongest running signal. Keep the goal-bucket fallback
+    // for Doing goals that have not created their first task yet.
+    const running = ptasks.some((t) => t.status === 'running')
+      || pgoals.some((g) => fsBoardBucket(g, ptasks) === 'doing'
+        && !(g.status === 'blocked' && g.blocked?.kind === 'paused'));
     const td = ptasks.filter((t) => ['running', 'queued'].includes(t.status)).length;
     // The rail's eye icon is the review/attention count. Needs-you goals
     // (needsInput/needsApproval/blocked, minus Paused) block on the person the
@@ -6762,6 +6791,7 @@ function renderFsFeed(){
   $('fsRoot').classList.toggle('clawdhero', !hasWork);
   const sc = feed.parentNode;
   const nearBottom = sc.scrollHeight - sc.clientHeight - sc.scrollTop <= 80;
+  if (answerInputComposing('data-fsansinput')) return; // mid-IME conversion — rebuilding now would abort it; compositionend re-runs this
   const savedFsAns = captureFocusedAnswerInput(feed, 'data-fsansinput');
   feed.innerHTML = hasWork ? out : '';
   if (nearBottom) sc.scrollTop = sc.scrollHeight;
@@ -8511,15 +8541,40 @@ window.addEventListener('focus', () => { if (signinPending) refreshAfterSignin()
 
 // When the served UI is newer than the one this tab is running (hot-copy
 // or release), reload in place — drafts survive via localStorage.
+//
+// A new task/goal running elsewhere routinely touches app/index.html's mtime
+// (its own edit, or a hot-copy after another lane's merge) — that alone used
+// to force this reload on every open tab, mid-review, wiping the open Review
+// dialog with no way back in (Masa 2026-07-29). Two fixes: (1) never reload
+// while the Review dialog (SA.open) is up — remember a pending version and
+// apply it the moment the dialog closes instead; (2) whichever card was open
+// is saved to localStorage first, so if a reload does happen while reviewing,
+// boot() reopens the same card instead of dropping the user back at nothing.
+let _pendingUiVersion = null;
 function checkUiVersion(v){
   if (!v) return;
   if (window.__uiv && window.__uiv !== v) {
-    try { localStorage.setItem('draft', $('input').value ?? ''); } catch {}
+    if (typeof SA !== 'undefined' && SA.open) { _pendingUiVersion = v; return; }
+    try {
+      localStorage.setItem('draft', $('input').value ?? '');
+      if (typeof SA !== 'undefined' && SA.open) {
+        localStorage.setItem('reopenReview', JSON.stringify({ goalId: SA.solo || SA.focusGid || SA.cur, solo: !!SA.solo }));
+      } else {
+        localStorage.removeItem('reopenReview');
+      }
+    } catch {}
     location.reload();
     return;
   }
   window.__uiv = v;
 }
+// Called from saClose() so a version that arrived mid-review gets applied
+// (reloaded) right after the user is done, instead of never at all.
+function applyPendingUiVersionIfAny(){
+  if (_pendingUiVersion == null) return;
+  const v = _pendingUiVersion; _pendingUiVersion = null;
+  checkUiVersion(v);
+}
 
 async function refresh(){
   try {
@@ -8817,6 +8872,13 @@ async function boot(){
   if (window.__GALDA_DISCONNECTED) return bootDisconnected(window.__GALDA_DISCONNECTED);
   $('input').value = localStorage.getItem('draft') ?? '';
   await refresh();
+  // Reopen the Review dialog a version-triggered reload had to close mid-review
+  // (see checkUiVersion) so the user lands back where they were, not on the board.
+  try {
+    const reopen = JSON.parse(localStorage.getItem('reopenReview') ?? 'null');
+    localStorage.removeItem('reopenReview');
+    if (reopen && typeof saOpen === 'function') saOpen(reopen.goalId, reopen.solo);
+  } catch {}
   // The folder sits above the composer and is read on every screen, so its list is
   // fetched once at boot rather than on first open: without HOME the bar can only
   // print the absolute path (measured — it read "/var/folders/…/agent-manager"
@@ -9020,32 +9082,15 @@ function resolveComposer(raw){
     review: !!(msg.review || parsed.review),
   };
 }
-// Feature A (ONE-CONVERSATION §1.2/§1.4 "1本の会話"): the composer CONTINUES the
-// goal you're focused on (state.selectedGoal) instead of always making a new goal
-// — like Claude Code, where you keep talking to the same session. It binds only
-// when that goal is in the ACTIVE project AND in a repliable status: a review goal
-// routes to the AI-decide reply (continue/rescope/split/answer), an in-flight or
-// finished-rework goal takes a plain follow-up. Anything else (done, needsInput,
-// planning/stacked, or nothing selected) → new-goal mode, exactly as before.
-const REPLY_AUTO = new Set(['review']);
-const REPLY_PLAIN = new Set(['running', 'partial', 'failed', 'interrupted', 'blocked']);
+// The centre composer is always a NEW task. `selectedGoal` is viewing state: using
+// it as an implicit reply target made a request typed while reading Review #856 get
+// appended to that unfinished goal. Replying is intentionally available only from
+// the goal/review card's dedicated Reply / Fix controls, whose destination is shown
+// next to the control. Never infer a delivery destination from what happens to be
+// open in the centre lane.
 function replyBinding(){
-  const id = state.selectedGoal;
-  if (id == null) return null;
-  // Focus and reply-target are separate concepts. A Review reply keeps this goal
-  // selected so its live work remains in the centre feed, while the composer goes
-  // back to new-goal mode and shows no stale “Replying to” chip.
-  if (state.composerDetachedGoal === id) return null;
-  const g = state.goals.find((x) => x.id === id);
-  if (!g || g.projectId !== state.active) return null;   // stale after a project switch → new-goal mode
-  // needsInput has NO worker session yet (planGoal() sets it before any task/worker
-  // exists — see engine/server.mjs) — there's no "conversation" to resume via /reply
-  // (that endpoint 409s on needsInput; engine/lib.mjs:3010). Binding here routes to
-  // the SAME /answer endpoint the question card's own free-text box already uses
-  // (fsAnswerQuestion), so the composer becomes a second way to answer, not a new path.
-  if (g.status === 'needsInput' && g.question) return { goal: g, mode: 'answer' };
-  if (REPLY_AUTO.has(g.status)) return { goal: g, mode: 'auto' };
-  if (REPLY_PLAIN.has(g.status)) return { goal: g, mode: 'plain' };
+  // Deliberately do not bind the main composer. Keep this one routing boundary so
+  // a future UI change cannot accidentally reintroduce selectedGoal -> /reply.
   return null;
 }
 async function send(){
@@ -10331,6 +10376,24 @@ function saRender(){
   const previousBody = S.querySelector('.sa-body');
   const previousScrollTop = previousBody?.scrollTop ?? 0;
   const restoreScroll = Boolean(previousBody) && SA.focusGid == null;
+  // A reply changes the height of the card being replied to (new thread line, status
+  // change) — restoring a raw scrollTop pixel then leaves the viewport at the same
+  // number but showing DIFFERENT content, a visible jolt on whatever card the person
+  // was actually reading (2026-07-29). Anchor on that card's own goalId instead: find
+  // the topmost visible card before the rebuild, and after the rebuild scroll so that
+  // SAME card sits at the SAME on-screen position, regardless of how much other cards
+  // above it grew or shrank.
+  let anchorGid = null, anchorOffset = 0;
+  if (restoreScroll) {
+    const containerTop = previousBody.getBoundingClientRect().top;
+    // The card actually being read sits near the middle of the viewport, not whatever
+    // sliver of a previous card's bottom edge still peeks in at the very top.
+    const midY = containerTop + previousBody.clientHeight / 2;
+    for (const el of previousBody.querySelectorAll('.sa-item')) {
+      const r = el.getBoundingClientRect();
+      if (r.top <= midY && r.bottom >= midY) { anchorGid = el.dataset.gid; anchorOffset = r.top - containerTop; break; }
+    }
+  }
   const focusedReplyGoal = document.activeElement?.matches?.('#seeall .sa-chat')
     ? document.activeElement.dataset.goal
     : null;
@@ -10355,7 +10418,6 @@ function saRender(){
   SA.lastMarkup = markup;
   S.innerHTML = markup;
   const body = S.querySelector('.sa-body');
-  if (body && restoreScroll) body.scrollTop = previousScrollTop;
   // Once the person scrolls, their position wins over the short opening-focus
   // animation. Programmatic opening scroll fires this too, which is fine: the
   // first placement happened and later SSE ticks should preserve it.
@@ -10367,7 +10429,18 @@ function saRender(){
   // Native <details> would collapse on every rebuild (SSE tick, approve, etc.) —
   // track what's open per goal+section in SA.detOpen (persists across re-renders)
   // and wire each one's toggle to keep that set current.
-  paintReplyNote();   // the answer/question under a card's reply box survives its rebuild
+  paintReplyNote();   // the answer/question under a card's reply box survives its rebuild — this is
+                       // also what actually grows a replied-to card's height, so the scroll anchor
+                       // below must run AFTER it, not before (2026-07-29 jolt fix).
+  if (body && restoreScroll) {
+    const anchorEl = anchorGid != null ? body.querySelector(`.sa-item[data-gid="${anchorGid}"]`) : null;
+    if (anchorEl) {
+      const newOffset = anchorEl.getBoundingClientRect().top - body.getBoundingClientRect().top;
+      body.scrollTop += (newOffset - anchorOffset);
+    } else {
+      body.scrollTop = previousScrollTop;
+    }
+  }
   for (const det of S.querySelectorAll('[data-detkey]')) {
     det.addEventListener('toggle', () => {
       const k = det.dataset.detkey;
@@ -10495,7 +10568,7 @@ function saOpen(focusGoalId, solo = false){
   saFetchDiffs();
   saEnhanceShotBoxes();
 }
-function saClose(){ SA.open = false; SA.solo = null; SA.focusGid = null; SA.lastMarkup = null; if (SA.focusTimer) { clearTimeout(SA.focusTimer); SA.focusTimer = null; } $('seeall').classList.remove('open'); }
+function saClose(){ SA.open = false; SA.solo = null; SA.focusGid = null; SA.lastMarkup = null; if (SA.focusTimer) { clearTimeout(SA.focusTimer); SA.focusTimer = null; } $('seeall').classList.remove('open'); applyPendingUiVersionIfAny(); }
 // Continuous review (Masa 2026-07-08): move a persistent "current" card with
 // ↑↓ and auto-advance to the next pending after each verdict — same screen, same
 // Ledger, so approving flows without hunting for the next one.

package/engine/lib.mjs

@@ -2148,16 +2148,34 @@ export function wantsPullRequest(text) {
   return /プルリク|pull\s*request|(^|[^A-Za-z])PR([^A-Za-z]|$)/i.test(String(text ?? ''));
 }
 
+// Auto delivery must distinguish a request to CHANGE the connected project
+// from a request to THINK about it.  A project-level PR preference remains the
+// user's strongest default, and an explicit Deliverable choice below always
+// wins.  This only prevents the silent "implemented, locally previewed, but no
+// PR" outcome for ordinary implementation requests in projects that have not
+// configured that preference yet.
+export function isImplementationRequest(text) {
+  const value = String(text ?? '');
+  const implementation = /実装|修正|直して|直す|不具合|バグ|追加|変更|作って|できるように|直したい|fix\b|implement\b|build\b|add\b|change\b|bug\b/i;
+  const explorationOnly = /リサーチ|調査|分析|アイデア|案(?:を|が|は|ください)|提案|比較|相談|教えて|説明|要約|デザイン|モック|ワイヤーフレーム|資料|スライド|レポート|ドキュメント|research\b|analysis\b|brainstorm\b|ideas?\b|suggest(?:ion)?\b|design\b|mockup\b|wireframe\b|slides?\b|report\b/i;
+  // "実装案を出して" is exploration; "案3を実装して" is implementation.
+  return implementation.test(value) && !(explorationOnly.test(value) && !/実装|修正|直して|直す|追加|変更|fix\b|implement\b|build\b|add\b|change\b/i.test(value));
+}
+
 // Task 46: the "Deliverable" toggle (pr | none | auto) on task creation and
 // editing. An explicit 'pr'/'none' always wins over the text heuristic, so
 // a user's choice survives editing the goal's text afterwards. `projectDefault`
 // (task 45: a project's reviewDefinition.defaultWantsPR, edited from the same
 // settings panel as the review definition) is only consulted for 'auto' when
 // the text itself gives no signal either way — an explicit "PRを作って" in the
-// text still wins even if the project's default is off, matching the
-// pre-task-45 behavior that new callers who omit this argument keep exactly.
-export function resolveWantsPR(pr, text, projectDefault = false) {
-  return pr === 'pr' ? true : pr === 'none' ? false : (wantsPullRequest(text) || !!projectDefault);
+// text still wins even if the project's default is off. `preferenceKnown`
+// distinguishes an intentional per-project "No PR" from an older project
+// which has never made a delivery choice. Only the latter gets the helpful
+// implementation-request fallback.
+export function resolveWantsPR(pr, text, projectDefault = false, preferenceKnown = false) {
+  if (pr === 'pr') return true;
+  if (pr === 'none') return false;
+  return wantsPullRequest(text) || !!projectDefault || (!preferenceKnown && isImplementationRequest(text));
 }
 
 // Task 90: which entry point created a goal (Web chat / MCP / future API),

package/engine/requirement-model.mjs

@@ -63,16 +63,26 @@ function normalizeEvidence(value) {
   const requirementId = cleanText(value?.requirementId);
   const attemptId = cleanText(value?.attemptId);
   if (!requirementId || !attemptId) return null;
-  const claim = CLAIM_STATUSES.has(value.claim) ? value.claim : 'not-addressed';
+  const claim = CLAIM_STATUSES.has(value.claim) ? value.claim : null;
   const coverage = COVERAGE_STATUSES.has(value.coverage) ? value.coverage : 'unverified';
+  const workerEvidence = cleanText(value.workerEvidence ?? value.evidence);
+  const workerFiles = [...new Set([
+    ...(Array.isArray(value.workerFiles) ? value.workerFiles : []),
+    ...(Array.isArray(value.files) ? value.files : []),
+  ].filter((item) => typeof item === 'string' && item))];
+  const observedFiles = [...new Set((Array.isArray(value.observedFiles) ? value.observedFiles : []).filter((item) => typeof item === 'string' && item))];
   return {
     requirementId,
     attemptId,
     taskId: value.taskId ?? null,
     claim,
     coverage,
-    evidence: cleanText(value.evidence),
-    files: [...new Set(Array.isArray(value.files) ? value.files.filter((item) => typeof item === 'string' && item) : [])],
+    evidence: workerEvidence,
+    files: [...new Set([...workerFiles, ...observedFiles])],
+    workerEvidence,
+    workerFiles,
+    observedFiles,
+    checks: Array.isArray(value.checks) ? value.checks : [],
   };
 }
 

package/engine/requirement-verification.mjs

@@ -66,3 +66,56 @@ export function mergeRequirementVerificationEvidence(existing = [], incoming = [
   }
   return [...rows.values()];
 }
+
+/**
+ * Translate Manager-owned observations into explicitly requirement-scoped
+ * checks. Merely seeing files or successfully starting a preview is useful
+ * evidence, but it is not proof that the requested behavior is correct.
+ */
+export function buildManagerVerificationChecks({
+  requirementIds = [], testResult = null, testsFailing = false, proof = null,
+  run = null, previewVerification = null, changedFiles = [],
+} = {}) {
+  const ids = cleanList(requirementIds);
+  const files = cleanList(changedFiles);
+  const checks = [];
+
+  if (files.length) {
+    checks.push({
+      kind: 'changed-files', status: 'observed', requirementIds: ids, files,
+      detail: `${files.length} changed file${files.length === 1 ? '' : 's'} observed`,
+    });
+  }
+
+  if (testResult) {
+    let status = 'observed';
+    if (!testResult.ran || testResult.skipped) status = 'not-applicable';
+    else if (testsFailing) status = 'failed';
+    else if (testResult.ok === true && !testResult.inconclusive && !testResult.infraOnly && !testResult.preExistingOnly) status = 'verified';
+    checks.push({
+      kind: 'test', status, requirementIds: ids,
+      detail: testResult.detail || testResult.skippedReason || `${testResult.passed ?? 0} passed, ${testResult.failed ?? 0} failed`,
+    });
+  }
+
+  if (proof) {
+    checks.push({
+      kind: 'proof',
+      status: proof.verified === true ? (proof.pass === true ? 'verified' : 'failed') : 'observed',
+      requirementIds: ids,
+      detail: proof.detail || 'Proof artifact observed',
+    });
+  }
+
+  if (run || previewVerification) {
+    checks.push({
+      kind: 'open-it', status: 'observed', requirementIds: ids,
+      detail: previewVerification?.detail || (previewVerification?.status === 'started'
+        ? 'Local preview started; visual correctness still requires review'
+        : 'Worker declared a local preview'),
+      ...(previewVerification?.url || run?.url ? { url: previewVerification?.url || run.url } : {}),
+    });
+  }
+
+  return checks;
+}

package/engine/server.mjs

@@ -23,7 +23,9 @@ import { pathToFileURL } from 'node:url';
 import { needsProjectFolder, folderLabel, chooseFolderScript, parseChosenFolder, buildFolderQuestion, resolveFolderAnswer, classifyFolderTarget, buildFolderInitConfirm, isFolderInitConfirmed, needsReviewPreference, buildReviewPreferenceQuestion, resolveReviewPreferenceAnswer, routeQuestionAnswer, notUnderstoodNote, parseRelocalizedQuestion, classifyAuthProbe, authBannerText, makeSigninChallenge, parseClaimResponse, isCommandOnPath, parseRequestedWorkspaceDir } from './lib.mjs';
 import { parseStreamEvents, parseCodexEvents, buildCodexArgs, parsePlan, refinePlanTasks, canEditGoal, canDeleteGoal, shouldAskClarification, workerExitReason, classifyWorkerFailure, isForcedStop, taskStatusAfterVerify, workerResultText, taskCountsAsComplete, resolveWantsPR, resolveGoalSource, buildGoalSummary, buildReviewAttemptLedger, latestGoalOutcomeTask, buildDirtyWorkspacePrBlock, buildGoalProofMd, buildGoalPrBody, buildReviewSummaryPrompt, parseReviewSummary, buildReviewRuleSection, nextGoalStatus, isPrMerged, isPrApproved, reviewToDoneStatus, advanceReviewGoal, revertReviewGoal, approveGoal, dismissGoal, rejectGoal, reopenGoal, permissionModeFor, isPlanReview, nextAfterPlanApprove, GOAL_MODES, parseSkillFrontmatter, collectSkills, retestGoal, cancelRetestGoal, computeRetestOutcome, revertGoal, planRevertActions, archiveGoal, unarchiveGoal, undismissGoal, parseNumstat, truncateDiffText, sumUsage, resolveEntryUrl, rebasePreviewUrl, shouldCreateGoalPR, normalizePullRequestUrl, parseGitLog, diffNewCommits, replayQueueLog, reconcileOrphanGoals, reorderQueue, sortQueueByPriority, TASK_PRIORITIES, validateWorkflowColumns, DEFAULT_WORKFLOW_COLUMNS, validateReviewDefinition, DEFAULT_REVIEW_DEFINITION, resolveTestGate, buildEphemeralSeedLog, buildEphemeralSeedProjects, trimTaskActivityForState, buildAskContext, WORKER_TOOLS, workerPrompt, verifyCfAccessJwt, resolveIdentity, goalVisibleTo, clampParallelLimit, canStartMore, nextRunnableTasks, goalsConflict, detectConflicts, pickFoldTarget, autoFoldReviewsEnabled, parseGitUnifiedDiffLocations, hasTestRelevantChanges, parseFailingTestNames, classifyTestGate, classifyPrSafety, isInconclusiveTestRun, countInfraFlakes, classifyRunFailures, classifyVerifyGate, buildVerificationInfraError, verificationInfraErrorFrom, buildExecutionPlan, buildContextHandoffSummary, buildFailurePostmortem, appendFailureMemory, latestFailurePolicy, classifyChangeRisk, checkRunBudget, usageBudgetTokens, isRateLimited, nextResumeDelay, checkFreeTierLimit, resolveEntitlement, resolveCachedEntitlement, FREE_TIER_LIMITS, QUEUED_GOAL_STATUSES, verifyLicenseToken, licenseTokenPayload, shouldRefreshLicense, shouldEmitSetupCompleted, pickAnalyticsUid, shouldEmitFreeExhausted, detectRequestLanguage, testFailureReason, manualTestRetryPrompt, isNothingVerifiable, isNothingVerifiableForPR, isSuspiciouslyIncompleteDone, classifyComposerIntentHeuristic, detectPauseIntent, buildIntentPrompt, parseIntentResponse, buildBoardSnapshot, validateBoardSnapshot, boardIsEmpty, decideBoardPull, resolveConnectedAgents, pickAvailableAgent, pickUtilityAgent, utilityModel, liveTakeoverDecision, shouldOpenBrowserOnBoot, goalTaskTitle, buildGoalTask, shouldUsePlannerForGoal, RUN_DECL_FILE, REQUIREMENT_EVIDENCE_FILE, parseRequirementEvidenceDocument, parseRunDeclaration, SHOT_DIR, collectShotNames, parseGoalAddress, REPLY_OUTCOMES, buildGoalMessage, serializeGoalMessage, parseGoalMessages, computeInternalQualityMetrics, classifyInternalQualityTaskError, resolveWorkspaceMode, goalSessionFor, setGoalAgentSession, switchGoalAgentSession, parseApprovalRequest } from './lib.mjs';
 import { migrateRequirementModel, validateRequirementModel } from './requirement-model.mjs';
+import { buildManagerVerificationChecks, buildRequirementVerificationEvidence, mergeRequirementVerificationEvidence } from './requirement-verification.mjs';
 import { createSerialQueue } from './lib.mjs';
+import { wantsPullRequest } from './lib.mjs';
 import { parseRequirementEvidence, unmappedChangedFiles, previewOpensSomethingElse } from './lib.mjs';
 import { isPrClosed } from './lib.mjs';
 import { collectProjectRules } from './lib.mjs';
@@ -2446,7 +2448,7 @@ function createSplitGoal(parent, text, identity, targetProjectId = parent.projec
   if (!targetProject) throw new Error(`unknown target project: ${targetProjectId}`);
   const goal = {
     id: nextId++, projectId: targetProjectId, text: text.slice(0, 8000), status: 'planning',
-    wantsPR: resolveWantsPR('auto', text, getReviewDefinition(targetProjectId).defaultWantsPR),
+    wantsPR: resolveWantsPR('auto', text, getReviewDefinition(targetProjectId).defaultWantsPR, !!targetProject.reviewPrefAsked),
     model: parent.model, effort: parent.effort, agent: parent.agent, mode: 'auto',
     skill: null, images: [], source: resolveGoalSource('app'), sourceRef: null,
     executionPlan: buildExecutionPlan({ projectId: targetProjectId, text, previousSession: false }),
@@ -3854,6 +3856,26 @@ function prSafetyIssue(goal, project, files) {
   return classifyPrSafety({ baseline, files, currentFiles: gitStatusPaths(lines), currentBranch });
 }
 
+function syncGoalRequirementVerification(goal, siblings, { testsFailing = false } = {}) {
+  const incoming = siblings.flatMap((task) => buildRequirementVerificationEvidence({
+    requirementIds: task.activeRequirementIds ?? [],
+    attemptId: task.attemptId || `task-${task.id}`,
+    taskId: task.id,
+    workerEvidence: task.requirementEvidence,
+    changedFiles: task.changedFiles ?? [],
+    checks: buildManagerVerificationChecks({
+      requirementIds: task.activeRequirementIds ?? [],
+      testResult: goal.testResult,
+      testsFailing,
+      proof: task.proof,
+      run: task.run,
+      previewVerification: task.previewVerification,
+      changedFiles: task.changedFiles ?? [],
+    }),
+  }));
+  goal.requirementEvidence = mergeRequirementVerificationEvidence(goal.requirementEvidence ?? [], incoming);
+}
+
 // The Manager-verified review gate (shared by goal completion and /reverify):
 // run OUR OWN tests, generate a plain-language headline, and require proof for a
 // UI change — capturing it (3 tries) if missing. Any failure → 'blocked' (shown
@@ -3975,6 +3997,10 @@ async function verifyGate(goal, prProject, siblings, baseProject, activityTask =
   // Count that drives the block message: the NEW failures when we could diff
   // against the base, else the raw total (fail-closed when the base is unknown).
   const blockingFailCount = goal.testResult.newFailures?.length ?? goal.testResult.failed;
+  // Persist the Manager's own requirement-scoped observations before any
+  // automatic rework return. A worker claim alone never becomes verified.
+  syncGoalRequirementVerification(goal, siblings, { testsFailing });
+  saveGoal(goal);
   // §4補足: requireVerifyPass (opt-in, off by default) makes the gate strict —
   // EVERY task must carry a passing proof, not just UI tasks. Only ever tightens;
   // never conflicts with §6 (still can't reach Done without a human).
@@ -4807,7 +4833,7 @@ const server = createServer(async (req, res) => {
           approvalDecision: null,
           reviewOnly: review ? true : undefined,
           note: review && typeof note === 'string' && note.trim() ? note.trim().slice(0, 8000) : undefined,
-          wantsPR: (review || approval) ? Boolean(importedPr) : resolveWantsPR(pr, text, getReviewDefinition(projectId).defaultWantsPR),
+          wantsPR: (review || approval) ? Boolean(importedPr) : resolveWantsPR(pr, text, getReviewDefinition(projectId).defaultWantsPR, !!project.reviewPrefAsked),
           model: workerModel(goalAgent, model),
           effort: workerEffort(goalAgent, effort),
           agent: goalAgent,
@@ -4925,7 +4951,8 @@ const server = createServer(async (req, res) => {
         if (promote) goal.prio = ++prioCounter;
         if (text?.trim()) {
           goal.text = text.trim().slice(0, 8000);
-          goal.wantsPR = resolveWantsPR(pr, goal.text, getReviewDefinition(goal.projectId).defaultWantsPR);
+          const project = projects.find((p) => p.id === goal.projectId);
+          goal.wantsPR = resolveWantsPR(pr, goal.text, getReviewDefinition(goal.projectId).defaultWantsPR, !!project?.reviewPrefAsked);
         }
         saveGoal(goal);
         json(res, 200, goal);
@@ -5036,14 +5063,27 @@ const server = createServer(async (req, res) => {
     const goal = goals.find((g) => g.id === task.goalId);
     startPreview(task)
       .then((p) => {
+        task.previewVerification = {
+          status: 'started', url: p.url, at: new Date().toISOString(),
+          detail: 'Local preview started; visual correctness still requires review',
+        };
+        saveTask(task);
         if (goal?.verificationInfraError?.source === 'preview') {
           goal.verificationInfraError = null;
+        }
+        if (goal) {
+          syncGoalRequirementVerification(goal, tasks.filter((item) => item.goalId === goal.id), {
+            testsFailing: Boolean(goal.testResult?.ran && goal.testResult.failed > 0
+              && !goal.testResult.inconclusive && !goal.testResult.infraOnly && !goal.testResult.preExistingOnly),
+          });
           saveGoal(goal);
         }
         json(res, 200, p);
       })
       .catch((e) => {
         const detail = String(e.message ?? e).slice(0, 200);
+        task.previewVerification = { status: 'infra-error', detail, at: new Date().toISOString() };
+        saveTask(task);
         if (goal && goal.status === 'review') {
           goal.verificationInfraError = buildVerificationInfraError({
             source: 'preview',
@@ -5483,7 +5523,7 @@ const server = createServer(async (req, res) => {
             send({ ev: 'review-definition', projectId: project.id, reviewDefinition: merged.definition });
           }
         }
-        goal.wantsPR = resolveWantsPR(undefined, goal.text, patch.defaultWantsPR);
+        goal.wantsPR = resolveWantsPR(undefined, goal.text, patch.defaultWantsPR, true);
         goal.question = null;
         goal.clarified = true;
         goal.status = 'stacked';
@@ -5549,7 +5589,7 @@ const server = createServer(async (req, res) => {
         // both asked twice and got no PR either time.
         // Upgrade only: this can turn the deliverable on, never off, so a later
         // unrelated reply cannot silently withdraw a PR the person already asked for.
-        if (resolveWantsPR('auto', text, false) && goal.wantsPR !== true) {
+        if (wantsPullRequest(text) && goal.wantsPR !== true) {
           goal.wantsPR = true;
           saveGoal(goal);
         }
@@ -5958,7 +5998,7 @@ async function answerGoalQuestion(goal, question) {
       // (rescope) or while asking a question (answer) is still a PR they asked
       // for, and it must survive until close-out rather than being dropped
       // because this particular reply happened not to wake a worker.
-      if (resolveWantsPR('auto', text, false)) goal.wantsPR = true;
+      if (wantsPullRequest(text)) goal.wantsPR = true;
       goal.status = result.status;
       saveGoal(goal);
       // rescope はワーカーを起こさない——「いったん止めて考え直す」なので、走り出したら

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@galda/cli",
-  "version": "0.10.98",
+  "version": "0.10.100",
   "type": "module",
   "description": "Galda - hand off work to Claude Code or Codex, get proof back. Runs on your existing subscription, no extra API cost.",
   "scripts": {