@galda/cli 0.10.68 → 0.10.69

package/app/index.html

@@ -5204,7 +5204,13 @@ function wireWfEditButtons(){
 // /api/goals/:id/reply thread. Approvals live client-side only
 // (state.reviewApprovals) since they are a review action, not new work.
 function reviewRequirementsFor(goalId){
-  return state.tasks.filter((t) => t.goalId === goalId && !t.reply).sort((a, b) => (a.num ?? 0) - (b.num ?? 0));
+  const reqs = state.tasks.filter((t) => t.goalId === goalId && !t.reply).sort((a, b) => (a.num ?? 0) - (b.num ?? 0));
+  const outcome = latestGoalOutcomeTask(goalId, state.tasks);
+  if (reqs.length === 1 && outcome?.reply) {
+    const req = reqs[0];
+    return [{ ...req, ...outcome, id: req.id, title: req.title, outcomeTaskId: outcome.id }];
+  }
+  return reqs;
 }
 function openReviewChecklist(goalId){
   // H14 (Masa 2026-07-08): a single review IS the See-all Ledger focused on that
@@ -5356,7 +5362,7 @@ function renderReviewChecklist(){
     }
     // P1: one shot when a promised check actually captured one, otherwise nothing at
     // all — "No capture yet" was us apologising for a photo we no longer go and take.
-    const media = proofMediaHtml(t.id, t.proof, esc(t.title ?? ''));
+    const media = proofMediaHtml(t.outcomeTaskId ?? t.id, t.proof, esc(t.title ?? ''));
     // task 156/234 + 160: show the WORKER'S REPORT (what it actually did/found,
     // incl. any test-only verification it describes) — that's what you read to
     // decide. Falls back to the instruction if the worker left no result.
@@ -5933,16 +5939,22 @@ $('reviewOverlay').addEventListener('click', (e) => { if (e.target === $('review
 // result linked to the goal (What was done), and the existing proof
 // video/screenshot rendering (Proof). Approve/Dismiss and the revision send
 // button are visual-only for now — not wired to any API.
+const REVIEW_DETAIL_STATUSES = new Set(['review', 'done', 'reverted', 'skipped', 'blocked', 'partial', 'failed', 'interrupted']);
 function openGoalDetail(goalId){
+  const g = state.goals.find((x) => x.id === goalId);
+  // Completed and human-attention work always uses the current Review Ledger card.
+  // Keeping this guard at the lowest-level opener prevents Done from falling back to
+  // the legacy overlay when it is opened through a different list/log entry point.
+  if (g && REVIEW_DETAIL_STATUSES.has(g.status) && typeof saOpen === 'function') {
+    saOpen(goalId, true);
+    return;
+  }
   state.goalDetailOpen = goalId;
   renderGoalDetail();
   $('goalDetailOverlay').classList.add('show');
 }
-// "Needs you" (attention) goals open as the SAME review card (Masa 2026-07-24, Option A);
-// any other non-review goal keeps the goal-detail overlay.
+// All detail entry points share the routing in openGoalDetail().
 function openGoalOrCard(goalId){
-  const g = state.goals.find((x) => x.id === goalId);
-  if (g && ['blocked', 'partial', 'failed', 'interrupted'].includes(g.status) && typeof saOpen === 'function') { saOpen(goalId, true); return; }
   openGoalDetail(goalId);
 }
 // List / board 共通の詳細導線。Review は review ledger、Needs you は attention card、
@@ -8816,20 +8828,27 @@ function saRows(){
   const ts = (state.tasks || []).filter((t) => t.projectId === state.active);
   return buildReviewDigest({ goals: gs, tasks: ts }).map((r) => saRowFromDigest(r, gs, ts));
 }
+function latestGoalOutcomeTask(goalId, tasks){
+  return (tasks || []).filter((t) => t.goalId === goalId && t.status === 'done' && String(t.result ?? '').trim())
+    .sort((a, b) => Date.parse(b.finishedAt || b.startedAt || b.createdAt || 0)
+      - Date.parse(a.finishedAt || a.startedAt || a.createdAt || 0))[0] || null;
+}
 // One digest-row → one See-all card item. Extracted from saRows (2026-07-24) so a single
 // "Needs you" goal can be opened as the SAME review card (saLedgerRows solo mode) — the card's
 // whole shape (Request/Result/Proof/Activity, empty sections auto-hidden) is reused verbatim.
 function saRowFromDigest(r, gs, ts){
     const g = gs.find((x) => x.id === r.goalIds[0]);
+    const outcome = latestGoalOutcomeTask(g?.id, ts);
     const pt = ts.find((t) => r.goalIds.includes(t.goalId) && t.proof && (t.proof.png || t.proof.gif));
     // "What we did" (H23 §2, replaces the old Plan/Review-thread folds): one bullet
     // per finished requirement task's own report; falls back to the goal's plan
     // steps when nothing has finished yet (e.g. a plan awaiting approval — the plan
     // IS "what we're going to do" in that case, which is the honest content to show).
-    const doneTasks = ts.filter((t) => r.goalIds.includes(t.goalId) && !t.reply && t.result)
+    const doneTasks = ts.filter((t) => r.goalIds.includes(t.goalId) && !t.reply && t.result && t.status !== 'skipped')
       .sort((a, b) => (a.num ?? 0) - (b.num ?? 0));
     const did = doneTasks.length
       ? doneTasks.map((t) => String(t.result).replace(/\s+/g, ' ').trim().slice(0, 160))
+      : outcome ? [String(outcome.result).replace(/\s+/g, ' ').trim().slice(0, 160)]
       : (Array.isArray(g?.plan) ? g.plan.filter(Boolean).map((s) => String(s).replace(/\s+/g, ' ').trim()) : []);
     // Sibling goals (H23 §2/§4 "did the AI drop part of what I asked"): buildReviewDigest
     // already collapses goals that share one PR into a single row (goalIds.length > 1)
@@ -8884,7 +8903,7 @@ function saRowFromDigest(r, gs, ts){
       // lines are NOT in state — 1000 of them per task would bloat every page load — so
       // this is just the id to ask for when the reviewer opens Activity.
       logTaskId: (() => {
-        const t = ts.find((x) => x.goalId === g?.id && !x.reply && x.activityLog);
+        const t = [...ts].reverse().find((x) => x.goalId === g?.id && x.activityLog);
         return t?.id ?? null;
       })(),
       // P1 (Masa 2026-07-21): the AI's own final report, verbatim and untrimmed — the
@@ -8892,14 +8911,14 @@ function saRowFromDigest(r, gs, ts){
       // every ask that has no running thing to open ("research this", "explain that"),
       // and the explanation for every ask that does.
       report: (() => {
-        const t = ts.find((x) => x.goalId === g?.id && !x.reply && x.result);
+        const t = outcome || ts.find((x) => x.goalId === g?.id && !x.reply && x.result);
         return String(t?.result ?? '').replace(/^検証(PASS|FAIL)[^\n]*\n+/, '').trim();
       })(),
       // …and the worker's own "here's how you see it running" declaration (task.run,
       // written by the worker into .manager-run.json). Absent for asks with nothing to
       // run — then no button, which is the honest state, not a missing feature.
       run: (() => {
-        const t = ts.find((x) => x.goalId === g?.id && !x.reply && x.run);
+        const t = [...ts].reverse().find((x) => x.goalId === g?.id && x.run);
         return t ? { taskId: t.id, ...t.run } : null;
       })(),
       // Explicit review-summary display fields (spec §11: the card's shape follows the

package/engine/chrome-safety.mjs

@@ -0,0 +1,25 @@
+import { execFileSync } from 'node:child_process';
+
+const MAC_CHROME = '/Applications/Google Chrome.app/Contents/MacOS/Google Chrome';
+
+// Chrome updates its app bundle while an older copy can remain open. Launching
+// the newly installed executable during that window aborts inside Chrome and
+// macOS shows a scary "quit unexpectedly" dialog.
+export function chromeUpdatePending(chromePath, {
+  platform = process.platform,
+  readCurrentVersion = () => execFileSync('/usr/libexec/PlistBuddy', [
+    '-c', 'Print:CFBundleShortVersionString',
+    '/Applications/Google Chrome.app/Contents/Info.plist',
+  ], { encoding: 'utf8' }).trim(),
+  readProcesses = () => execFileSync('/bin/ps', ['-axo', 'command='], { encoding: 'utf8' }),
+} = {}) {
+  if (platform !== 'darwin' || chromePath !== MAC_CHROME) return false;
+  try {
+    const current = readCurrentVersion();
+    const running = [...readProcesses().matchAll(/Google Chrome Framework\.framework\/Versions\/([^/\s]+)/g)]
+      .map((match) => match[1]);
+    return running.some((version) => version !== current);
+  } catch {
+    return false;
+  }
+}

package/engine/lib.mjs

@@ -2585,6 +2585,7 @@ export function workerPrompt(task, goal, lastFailure, agentName = 'claude-code')
     // instead of expanding scope"): the request, not our fear of scope, sets
     // the size of the work.
     '- Deliver what the request actually asks for. If it asks for multiple options, N variations, a visual, an artifact, or a full exploration, produce ALL of them — do not reduce it to one, and do not cut it short to finish faster.',
+    '- デザイン案・UI案・レイアウト案を求められた場合は、文章だけで完了にしない。画像を使わない案でも、比較して確認できるHTML等のartifactをプロジェクト内に作り、実際に開ける .manager-run.json を残す。',
     '- Keep the work bounded to the request: inspect only files needed; avoid broad repository scans unless they are necessary.',
     "- You may use the user's installed skills (Skill tool, e.g. /galda-doc) when one clearly fits the request.",
     // Token-frugality + the 10-min run cap: the worker used to run the WHOLE suite
@@ -2972,7 +2973,15 @@ export function visibleTodos(list, showAll, limit = 10) {
 // Japanese lines covering ①何をやったか ②変更点 ③確認方法, even when tasks
 // are mixed done/failed/interrupted or there is only a single task.
 export function buildGoalSummary({ goalText, tasks } = {}) {
-  const work = (tasks ?? []).filter((t) => !t.reply);
+  const all = tasks ?? [];
+  const requirements = all.filter((t) => !t.reply);
+  const latestOutcome = latestGoalOutcomeTask(all);
+  // A follow-up can be the run that finally delivered the work after the original
+  // requirement was skipped/interrupted. Keep the requirement count stable, but
+  // summarize the effective outcome instead of preserving a stale "Skipped.".
+  const work = requirements.length === 1 && latestOutcome?.reply
+    ? [{ ...requirements[0], ...latestOutcome, title: requirements[0].title }]
+    : requirements;
   const done = work.filter((t) => ['done', 'skipped'].includes(t.status));
   const failed = work.filter((t) => ['failed', 'interrupted'].includes(t.status));
   const files = [...new Set(work.flatMap((t) => t.changedFiles ?? []))];
@@ -3009,6 +3018,15 @@ export function buildGoalSummary({ goalText, tasks } = {}) {
   return lines.join('\n');
 }
 
+// The Review card represents the goal's latest completed worker outcome. A reply is
+// a real worker run and may supersede an earlier skipped/failed requirement.
+export function latestGoalOutcomeTask(tasks = []) {
+  return tasks
+    .filter((t) => t && t.status === 'done' && String(t.result ?? '').trim())
+    .sort((a, b) => Date.parse(b.finishedAt || b.startedAt || b.createdAt || 0)
+      - Date.parse(a.finishedAt || a.startedAt || a.createdAt || 0))[0] ?? null;
+}
+
 // Plain-text snapshot of Review/Attention/Done fed to the /api/ask LLM pass —
 // same three buckets app/index.html's renderTasks() renders (goal.status
 // 'review'/'blocked' + task.status failed/interrupted + task.status

package/engine/server.mjs

@@ -2363,12 +2363,15 @@ async function runTask(task) {
   if (task.reply) {
     const plan = goal?.executionPlan;
     const resumeSession = shouldFreshSessionForPlan(plan) ? undefined : (goal?.sessionId ?? undefined);
+    // A follow-up is a full worker run, so it needs the same deliverable contract as
+    // the first attempt (artifacts, run declaration, tests, language). The old short
+    // prompt omitted those rules, which let a reply finish with prose only and left
+    // Review showing the original task's stale "Skipped." result.
     const prompt = [
       managerHandoffPrompt(goal, plan),
-      task.detail,
-      '',
-      '(User follow-up on the goal you worked on. Act on it inside this project directory; reply briefly in Japanese.)',
-    ].filter(Boolean).join('\n');
+      workerPrompt(task, goal, null, workerAgent(task.agent ?? goal?.agent)),
+      '(User follow-up on the goal you worked on. Apply it to the same goal and deliver the requested result.)',
+    ].filter(Boolean).join('\n\n');
     const callT0 = Date.now();
     sendProcessAct(task, `Starting ${workerAgent(task.agent ?? goal?.agent) === 'codex' ? 'Codex' : 'Claude Code'} worker with the existing goal context.`);
     const r = await runWorkerAgent({

package/engine/shot.mjs

@@ -21,6 +21,7 @@ import { existsSync, mkdirSync, statSync } from 'node:fs';
 import { dirname, resolve } from 'node:path';
 import { createRequire } from 'node:module';
 import { parseShotArgs, shotTargetUrl, pickChromePath } from './lib.mjs';
+import { chromeUpdatePending } from './chrome-safety.mjs';
 
 const require = createRequire(import.meta.url);
 
@@ -42,6 +43,9 @@ const url = shotTargetUrl(opt.target, (p) => resolve(process.cwd(), p));
 const out = resolve(process.cwd(), opt.out);
 const chrome = pickChromePath(process.env, existsSync);
 if (!chrome) die('Chrome not found — install Google Chrome or set CHROME_PATH to the binary');
+if (process.env.GALDA_TEST_ALLOW_PENDING_CHROME !== '1' && chromeUpdatePending(chrome)) {
+  die('Chrome just updated while it was open. Restart Chrome, then retry the screenshot.');
+}
 
 let puppeteer;
 try {

package/engine/verify.mjs

@@ -6,10 +6,14 @@
 
 import { spawnSync } from 'node:child_process';
 import { createRequire } from 'node:module';
+import { mkdtempSync, rmSync } from 'node:fs';
+import { tmpdir } from 'node:os';
+import { join } from 'node:path';
 
 const require = createRequire(import.meta.url);
 const puppeteer = require('puppeteer-core');
 import { existsSync } from 'node:fs';
+import { chromeUpdatePending } from './chrome-safety.mjs';
 
 const CHROME_CANDIDATES = [
   process.env.CHROME_PATH,
@@ -109,12 +113,28 @@ export async function exerciseUi(page, ui) {
 // still screenshot is produced) — used for a cheap baseline ("before") shot
 // where a comparison image is enough and a second full video isn't worth
 // the extra time/flakiness.
-export async function runVerification({ entryUrl, verify, outBase, record = true, viewport = null }) {
-  const browser = await puppeteer.launch({
-    executablePath: CHROME, headless: 'new',
-    args: ['--no-first-run', '--hide-scrollbars'],
-  });
+export async function runVerification({
+  entryUrl, verify, outBase, record = true, viewport = null,
+  browserUpdatePending = () => chromeUpdatePending(CHROME),
+}) {
+  if (browserUpdatePending()) {
+    return {
+      pass: false,
+      inconclusive: true,
+      detail: 'Chrome was updated while it was open. Automatic browser verification was safely deferred; restart Chrome to enable it.',
+      shotPath: null,
+      videoPath: null,
+      gifPath: null,
+    };
+  }
+  const userDataDir = mkdtempSync(join(tmpdir(), 'galda-verify-chrome-'));
+  let browser;
   try {
+    browser = await puppeteer.launch({
+      executablePath: CHROME, headless: 'new',
+      userDataDir,
+      args: ['--no-first-run', '--hide-scrollbars', '--disable-crash-reporter'],
+    });
     const page = await browser.newPage();
     await page.setViewport(viewport ?? { width: 900, height: 640, deviceScaleFactor: 2 });
     const consoleErrors = [];
@@ -160,6 +180,7 @@ export async function runVerification({ entryUrl, verify, outBase, record = true
     if (consoleErrors.length) result.detail += ` | page errors: ${consoleErrors.join(' / ').slice(0, 300)}`;
     return { ...result, shotPath, videoPath: finalVideo, gifPath };
   } finally {
-    await browser.close();
+    await browser?.close().catch(() => {});
+    rmSync(userDataDir, { recursive: true, force: true });
   }
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@galda/cli",
-  "version": "0.10.68",
+  "version": "0.10.69",
   "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": {