@galda/cli 0.10.116 → 0.10.118

package/CLAUDE.md

@@ -5,7 +5,9 @@
 ## テスト（Masa指示 2026-07-03・必須）
 
 - `毎回、実装したものにはテストを書いて、実行する`。
-  - engine/ の変更 → `node --test --test-concurrency=1 'engine/test/*.test.mjs'` が緑になるまで完了と言わない。
+  - engine/ の変更 → **`npm test`**（＝`node --test --test-concurrency=1 --test-global-setup=engine/test/helpers/browser-setup.mjs 'engine/test/*.test.mjs'`）が緑になるまで完了と言わない。
+  - 🔴 **`--test-global-setup=…/browser-setup.mjs` を外さない**（2026-07-31 実測）。38ファイルがChromeを使う。付けると**スイート全体でChromeが1つ**になり、交互A/B実測で **825秒→231秒 / 416秒→261秒**、赤も **2件→1件**（ONで新しく壊れたテストはゼロ）。**外した時の実行時間は825秒と416秒で倍ぶれる**＝「テストが遅い」ではなく「マシンの機嫌を測っている」状態に戻る。素の `node --test` を手打ちすると付き忘れるので `npm test` を使う。
+    - 🔴 **共有ブラウザを `close()` しない**。`getBrowser()` が返すのは**スイート全員のChrome**。閉じると以降の全ファイルが `connect ECONNREFUSED` で落ちる（2026-07-31 に実際にそうなっていて、この仕組みは実装済みなのに一度も動いていなかった）。終わりは必ず **`closeBrowser()`**＝自分で起動した時だけ殺し、共有なら切断するだけ。自前で `puppeteer.connect` する場合も同じ（`later-list-new-ui.test.mjs` が手本）。
   - 🔴 **`--test-concurrency=1` は必須**（Masa報告のカオス調査で発見・2026-07-16）。既定（CPUコア数）だと
     58ファイルが同時に実サーバと Chrome を立てて**マシンを飽和させ、同じコードで赤が毎回入れ替わる**。
     **「テスト緑」がコードの状態でなく運を報告している状態を放置しない**＝規律の土台。

package/app/index.html

@@ -1598,6 +1598,7 @@
   #fsRoot .cc-ask-lb{font-family:var(--mono);font-size:8.5px;letter-spacing:.11em;text-transform:uppercase;
     color:var(--blue);margin-bottom:6px}
   #fsRoot .cc-ask-q{font-size:12.5px;font-weight:500;color:var(--ink);line-height:1.35;margin-bottom:10px}
+  #fsRoot .cc-ask-note{font-size:12px;color:var(--amber,#e0a03a);line-height:1.35;margin-bottom:6px}
   #fsRoot .cc-ask .qchips{margin:0}
   /* Free-text answer, always shown alongside chips (Masa 2026-07-08「最初から出そう」):
      input + the shared rainbow send ring. Restores 質問カードの自由記述回答欄. */
@@ -3530,10 +3531,13 @@ function captureFocusedAnswerInput(container, attr){
   if (!el || !container || !container.contains(el) || !el.hasAttribute(attr)) return null;
   return { id: el.getAttribute(attr), value: el.value, start: el.selectionStart, end: el.selectionEnd };
 }
-function restoreFocusedAnswerInput(container, attr, saved){
+function restoreFocusedAnswerInput(container, attr, saved, shouldRestoreValue){
   if (!saved || !container) return;
   const el = container.querySelector(`[${attr}="${saved.id}"]`);
   if (!el) return;
+  // Focus always comes back; the TEXT only when the caller still wants it there.
+  const keep = typeof shouldRestoreValue === 'function' ? shouldRestoreValue(saved.id) : true;
+  if (!keep) { el.value = ''; el.focus(); return; }
   el.value = saved.value;
   el.focus();
   try { el.setSelectionRange(saved.start, saved.end); } catch { /* input type may not support it */ }
@@ -6699,8 +6703,19 @@ $('gdSend').onclick = () => {
   const text = $('gdRevisionInput').value.trim();
   if (goalId == null || !text) return;
   const goal = state.goals.find((g) => g.id === goalId);
-  if (goal?.status === 'review') sendDismiss(goalId, text);
-  else sendGoalInstruction(goalId, text);
+  if (goal?.status === 'review') { sendDismiss(goalId, text); return; }
+  // A goal that is ASKING gets an ANSWER. This box says "Answer the question…" for
+  // needsInput, and it used to post to /reply, where the server accepted it, returned 200
+  // and changed nothing — the question stayed up and nobody was told (a first-time user
+  // hit exactly this, 2026-07-31). /answer is the route that hands the words back to the
+  // agent that asked.
+  if (goal?.status === 'needsInput') {
+    $('gdRevisionInput').value = '';
+    closeGoalDetail();
+    fsAnswerQuestion(goalId, text);
+    return;
+  }
+  sendGoalInstruction(goalId, text);
 };
 // Feedback 8/9: Enter in the goal-detail revision box sends (like the button) —
 // which is what moves the goal Review→To Do for rework. Was button-only before.
@@ -6985,9 +7000,15 @@ function renderFsFeed(){
     // it deleted the GOAL, not the question, from an unlabelled card — and an
     // irreversible action does not belong beside a text box in a feed you are
     // skim-reading. Deleting still lives on the card and in the right lane.
+    // The server marks an answer it could not use. Say it as its own line: it used to be
+    // concatenated into the question text, so the card re-rendered looking identical and
+    // read as "ignored" rather than "I could not use that" (2026-07-31).
+    const notUsed = g.question.notUnderstood === true;
+    const qText = notUsed ? String(g.question.text).replace(/^[^\n]*\n+/, '') : g.question.text;
     askHtml += `<div class="cc-ask"><div class="cc-ask-lb">Question`
       + `<span class="cc-ask-of">#${esc(String(g.id))}${title ? ` ${esc(title)}` : ''}</span></div>`
-      + `<div class="cc-ask-q">${esc(g.question.text)}</div>`
+      + (notUsed ? `<div class="cc-ask-note">I could not use that answer — give me a full path.</div>` : '')
+      + `<div class="cc-ask-q">${esc(qText)}</div>`
       + (chips ? `<div class="qchips">${chips}</div>` : '')
       + `<div class="answ-inline"><input type="text" data-fsansinput="${g.id}" placeholder="Type your answer…" aria-label="Type your answer">`
       + `<button type="button" class="sa-send" data-fsanssend="${g.id}" title="Send answer" aria-label="Send answer"><svg viewBox="0 0 24 24"><path d="M12 19V5M5 12l7-7 7 7"/></svg></button></div></div>`;
@@ -7058,7 +7079,10 @@ function renderFsFeed(){
   const savedFsAns = captureFocusedAnswerInput(feed, 'data-fsansinput');
   feed.innerHTML = hasWork ? out : '';
   if (nearBottom) sc.scrollTop = sc.scrollHeight;
-  restoreFocusedAnswerInput(feed, 'data-fsansinput', savedFsAns);
+  // Not restored when the server just rejected it: putting the same words back is what
+  // made a rejected answer look like nothing had happened at all.
+  restoreFocusedAnswerInput(feed, 'data-fsansinput', savedFsAns, (id) =>
+    !(state.goals.find((g) => String(g.id) === String(id))?.question?.notUnderstood === true));
   window.__fsCineSync?.(); // Cinema dynamic rule: content growth may reach the band region
   for (const b of feed.querySelectorAll('[data-fsretry]')) b.onclick = () => fsRetryTask(b.dataset.fsretry);
   for (const b of feed.querySelectorAll('[data-fslog]')) b.onclick = () => fsToggleLog(Number(b.dataset.fslog));

package/engine/server.mjs

@@ -975,7 +975,15 @@ const workerOpenedWith = new Set();
 function sayFromWorker(goal, task, text) {
   if (!goal || !task || workerOpenedWith.has(task.id)) return;
   workerOpenedWith.add(task.id);
-  addGoalMessage(goal, { from: 'ai', via: task.reply ? 'thread' : 'composer', text });
+  const msg = addGoalMessage(goal, { from: 'ai', via: task.reply ? 'thread' : 'composer', text });
+  if (!msg) return;
+  // Say it out loud, now. The conversation on a goal is read on demand (a board
+  // with 100 cards must not drag 100 threads along), so a message written to it
+  // is invisible until someone opens that card — and the whole point of this one
+  // is that the person is sitting in front of the composer waiting to hear
+  // something. The text rides on the event: no fetch, and nothing for the client
+  // to work out about which messages it has already shown.
+  sendGoalEvent({ ev: 'say', goalId: goal.id, taskId: task.id, agent: task.agent ?? goal.agent, text: msg.text, at: msg.at }, goal);
 }
 
 function readGoalMessages(goalId) {
@@ -5491,15 +5499,7 @@ const server = createServer(async (req, res) => {
       // A worker-authored question is a transport pause, not a new planning
       // cycle. Preserve the worker session and pass the user's answer straight
       // back to it; Galda must not reinterpret or rewrite the answer.
-      if (goal.question?.kind === 'worker') {
-        goal.question = null;
-        goal.blocked = null;
-        goal.status = 'running';
-        goal.startedAt = null;
-        saveGoal(goal);
-        const replyTask = queueReplyTask(goal, answer, { from: 'you', via: 'question' });
-        return json(res, 200, { ...goal, replyTaskId: replyTask.id });
-      }
+      if (goal.question?.kind === 'worker') return json(res, 200, deliverWorkerAnswer(goal, answer));
       // Bind this goal to a project rooted at `dir` and start planning. Shared by
       // the "picked a repo" and "git-init'd a new folder" paths below.
       const bindAndPlan = (dir) => {
@@ -5559,6 +5559,11 @@ const server = createServer(async (req, res) => {
           goal.question = buildFolderQuestion(detectRepos(HOME), { lang });
         }
         goal.question.text = notUnderstoodNote(lang) + goal.question.text;
+        // Machine-readable, because the note above is concatenated into the question text
+        // and the card re-renders looking identical — which reads as "ignored", not as
+        // "I could not use that" (measured 2026-07-31). The app clears the box and shows
+        // this as its own line instead of relying on someone re-reading the paragraph.
+        goal.question.notUnderstood = true;
         saveGoal(goal);
         return json(res, 200, goal);
       }
@@ -5654,12 +5659,47 @@ const server = createServer(async (req, res) => {
     return;
   }
 
+// Hand a person's answer back to the worker that asked the question, unchanged. The
+// worker's session is preserved, so it can answer, ask again, or carry on — which is the
+// whole point: the question layer is Claude Code's / Codex's, not Galda's (Masa
+// 2026-07-31). Shared by /answer and /reply so that WHICH box the person typed into
+// cannot change whether their words arrive.
+function deliverWorkerAnswer(goal, answer) {
+  goal.question = null;
+  goal.blocked = null;
+  goal.status = 'running';
+  goal.startedAt = null;
+  saveGoal(goal);
+  const replyTask = queueReplyTask(goal, answer, { from: 'you', via: 'question' });
+  return { ...goal, replyTaskId: replyTask.id };
+}
   const replyMatch = url.pathname.match(/^\/api\/goals\/(\d+)\/reply$/);
   if (replyMatch && req.method === 'POST') {
     const goal = goals.find((g) => g.id === Number(replyMatch[1]));
     if (!goal) return json(res, 404, { error: 'goal not found' });
     if (!requireGoalOwnership(req, res, goal)) return;
     if (goal.status === 'needsApproval') return json(res, 409, { error: 'goal is awaiting an approval decision — use /approval' });
+    // A goal that is ASKING something used to accept a reply here, return 200, and change
+    // nothing: the question stayed on screen, the goal stayed in needsInput, and the person
+    // was told nothing (measured 2026-07-31 — a first-time user typed into a box labelled
+    // "Answer the question…" and it silently went nowhere). Two honest outcomes instead:
+    //   • the worker asked  → deliver it, exactly as /answer would. Which box someone typed
+    //     into must not decide whether their words reach the agent.
+    //   • Galda's own card  → say so, so the UI can fail loudly instead of pretending.
+    if (goal.status === 'needsInput') {
+      if (goal.question?.kind === 'worker') {
+        let body = '';
+        req.on('data', (d) => { body += d; });
+        req.on('end', () => {
+          let text = '';
+          try { text = String(JSON.parse(body || '{}').text ?? '').trim(); } catch {}
+          if (!text) return json(res, 400, { error: 'text required' });
+          return json(res, 200, deliverWorkerAnswer(goal, text));
+        });
+        return;
+      }
+      return json(res, 409, { error: 'goal is waiting on an answer — use /answer' });
+    }
     if (['stacked', 'planning'].includes(goal.status)) return json(res, 409, { error: 'goal not started yet — edit it instead' });
     let body = '';
     req.on('data', (d) => { body += d; });

package/package.json

@@ -1,11 +1,11 @@
 {
   "name": "@galda/cli",
-  "version": "0.10.116",
+  "version": "0.10.118",
   "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": {
     "start": "node engine/server.mjs",
-    "test": "node --test --test-concurrency=1 'engine/test/*.test.mjs'"
+    "test": "node --test --test-concurrency=1 --test-global-setup=engine/test/helpers/browser-setup.mjs 'engine/test/*.test.mjs'"
   },
   "dependencies": {
     "puppeteer-core": "^23.11.1",