@galda/cli 0.10.97 → 0.10.99

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)}
@@ -6532,8 +6535,11 @@ 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
@@ -9985,8 +9991,23 @@ function saItemHtml(it, i){
   const runDeclErrorBody = (!((it.run?.cmd || it.run?.url)) && it.runDeclError)
     ? `<div class="ropen rdeclerr"><p class="rbody v">${failureLang() === 'ja' ? 'AIが開けるものを用意しようとしましたが、記録が読み取れませんでした。返信して直させてください。' : 'The AI tried to declare something to open here, but the record was unreadable. Reply to have it rewritten.'}</p></div>`
     : '';
+  // The worker declared a static file AND changed something else (engine sets
+  // run.alsoChanged; goal #871). Both arrive as a /upload link, so the address
+  // above cannot tell you whether you are about to open the product or a page
+  // the agent wrote about the product — #871 was the second, and the request
+  // ("something I can check locally") had to be made again.
+  //
+  // One sentence of fact in the same voice as the note beside it: no badge, no
+  // colour, no warning tone. The card is not saying the worker did the wrong
+  // thing — a write-up beside a code change is often exactly right. It is
+  // saying which of the two this link is. Absent → nothing renders (§1.5-3).
+  const alsoChanged = it.run?.alsoChanged?.length
+    ? `<p class="rbody v">${failureLang() === 'ja'
+      ? `AIが書いたファイルを開きます。このタスクは ${esc(it.run.alsoChanged.join(' · '))} も変更しています。`
+      : `Opens a file the agent wrote. This task also changed ${esc(it.run.alsoChanged.join(' · '))}.`}</p>`
+    : '';
   const openBody = (it.run?.cmd || it.run?.url)
-    ? `<div class="ropen" id="runlocal${i}"><button type="button" class="olink" data-saact="preview:${i}" id="prevlink${i}">${esc(it.run.url || 'Start it')}<span class="ar">↗</span></button><p class="rbody v">${esc(it.run.note || it.run.cmd || 'Starts when you press it.')}${it.interruptedTask ? ' Available now; it does not wait for this run to finish.' : ''}</p></div>`
+    ? `<div class="ropen" id="runlocal${i}"><button type="button" class="olink" data-saact="preview:${i}" id="prevlink${i}">${esc(it.run.url || 'Start it')}<span class="ar">↗</span></button><p class="rbody v">${esc(it.run.note || it.run.cmd || 'Starts when you press it.')}${it.interruptedTask ? ' Available now; it does not wait for this run to finish.' : ''}</p>${alsoChanged}</div>`
     : it.pr
       ? `<div class="ropen"><a class="olink" href="${esc(it.pr)}" target="_blank" rel="noopener">${esc(it.pr)}<span class="ar">↗</span></a><p class="rbody v">Review and merge this pull request to finish the item.</p></div>`
       : runDeclErrorBody;
@@ -10316,6 +10337,24 @@ 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;
@@ -10340,7 +10379,6 @@ 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.
@@ -10352,7 +10390,18 @@ 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;

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

@@ -23,6 +23,7 @@ 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 { parseRequirementEvidence, unmappedChangedFiles, previewOpensSomethingElse } from './lib.mjs';
 import { isPrClosed } from './lib.mjs';
@@ -3854,6 +3855,26 @@ 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 +3996,10 @@ 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).
@@ -5036,14 +5061,27 @@ 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',

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@galda/cli",
-  "version": "0.10.97",
+  "version": "0.10.99",
   "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": {