@galda/cli 0.10.87 → 0.10.88

package/app/index.html

@@ -3633,10 +3633,7 @@ function reviewCardConfigFor(projectId){ return { ...DEFAULT_REVIEW_DEFINITION.r
 // project to one language regardless of what a given goal was typed in.
 function reviewLangFor(projectId){ const l = reviewDefinitionFor(projectId).language; return (l === 'en' || l === 'ja') ? l : 'auto'; }
 function retryCopy(projectId){
-  const ja = reviewLangFor(projectId) === 'ja';
-  return ja
-    ? { infra: '検証基盤エラー', retry: '再検証', retrying: '再検証中…', failed: '再検証に失敗' }
-    : { infra: 'Verification infrastructure error', retry: 'Retry verification', retrying: 'Retrying…', failed: 'Verification retry failed' };
+  return { infra: 'Verification infrastructure error', retry: 'Retry verification', retrying: 'Retrying…', failed: 'Verification retry failed' };
 }
 
 function esc(s){ const d = document.createElement('div'); d.textContent = s ?? ''; return d.innerHTML; }
@@ -4891,7 +4888,8 @@ function renderTasks(){
   for (const b of $('tasklist').querySelectorAll('[data-goalreverify]')) {
     b.onclick = async (e) => { e.stopPropagation(); const copy = retryCopy(state.active); b.disabled = true; b.textContent = copy.retrying;
       let r; try { r = await fetch(withKey(`/api/goals/${b.dataset.goalreverify}/reverify`), { method: 'POST' }); }
-      catch { showErr(copy.failed); return; }
+      catch { showErr(copy.failed); b.disabled = false; b.textContent = copy.retry; return; }
+      if (!r.ok) { showErr(copy.failed); b.disabled = false; b.textContent = copy.retry; return; }
       qrefresh(r, copy.failed); };
   }
   for (const b of $('tasklist').querySelectorAll('[data-qreply]')) {
@@ -5557,8 +5555,8 @@ function renderReviewChecklist(){
       b.disabled = true;
       b.textContent = retryCopy(goal.projectId).retrying;
       let r; try { r = await fetch(withKey(`/api/goals/${b.dataset.goalreverify}/reverify`), { method: 'POST' }); }
-      catch { showErr(retryCopy(goal.projectId).failed); b.disabled = false; return; }
-      if (r.ok) await refresh(); else showErr(retryCopy(goal.projectId).failed);
+      catch { showErr(retryCopy(goal.projectId).failed); b.disabled = false; b.textContent = retryCopy(goal.projectId).retry; return; }
+      if (r.ok) await refresh(); else { showErr(retryCopy(goal.projectId).failed); b.disabled = false; b.textContent = retryCopy(goal.projectId).retry; }
     };
   }
   for (const b of $('reviewList').querySelectorAll('.reqsend')) {
@@ -8530,7 +8528,7 @@ let sseUp = false; // live engine link — the honest signal behind the MCP tab'
 // Connection status line (queue tray): grace-debounce the offline state ~800ms so
 // sub-second SSE blips don't flash a red line. connOffline is the debounced flag
 // renderQueue() reads; connSetOnline/Offline flip it around real sseUp changes.
-let connOffline = false, _connGraceT = null, _sseReconnectT = null, _es = null, _healthT = null;
+let connOffline = false, _connGraceT = null, _sseReconnectT = null, _es = null, _healthT = null, _healthFailStreak = 0;
 function connSetOnline(){ if (_connGraceT) { clearTimeout(_connGraceT); _connGraceT = null; } if (connOffline) { connOffline = false; renderQueue(); } }
 function connSetOffline(){ if (_connGraceT || connOffline) return; _connGraceT = setTimeout(() => { _connGraceT = null; connOffline = true; renderQueue(); }, 800); }
 // Reconnect button: drop any pending backoff timer and reopen the EventSource now.
@@ -8545,15 +8543,23 @@ async function checkHealth(){
   if (document.hidden) return;
   let ok = false;
   try {
-    const c = new AbortController(); const to = setTimeout(() => c.abort(), 3000);
+    // Mobile browsers and DevTools device emulation can briefly starve this low-priority
+    // probe even while the already-open SSE and POST path are healthy.  Do not turn one
+    // transient probe into a user-visible outage.
+    const c = new AbortController(); const to = setTimeout(() => c.abort(), 5000);
     const r = await fetch(withKey('/'), { method: 'HEAD', cache: 'no-store', signal: c.signal });
     clearTimeout(to); ok = r.ok;
   } catch { ok = false; }
   if (!ok) {
+    _healthFailStreak++;
+    if (_healthFailStreak < 2) return;
     connSetOffline(); // grace-debounced ~800ms
     if (!_sseReconnectT) { try { _es && _es.close(); } catch {} sseUp = false; esRetryMs = Math.min(esRetryMs, 4000); _sseReconnectT = setTimeout(connectSSE, esRetryMs); }
-  } else if (connOffline && !_sseReconnectT) {
+  } else {
+    _healthFailStreak = 0;
+    if (connOffline && !_sseReconnectT) {
     forceReconnect(); // server reachable again but the line is still up → reconnect to clear + flush
+    }
   }
 }
 function startHealthPoll(){ clearInterval(_healthT); _healthT = setInterval(checkHealth, 4000); }
@@ -8773,6 +8779,10 @@ async function boot(){
   flushOutbox();
   connectSSE();
   startHealthPoll();
+  // A Wi-Fi ⇄ cellular switch can leave EventSource waiting for its normal retry
+  // budget.  The browser has already confirmed connectivity here, so reconnect now
+  // instead of making a mobile send wait for the next health poll/backoff cycle.
+  window.addEventListener('online', () => { _healthFailStreak = 0; forceReconnect(); });
   document.addEventListener('visibilitychange', () => { if (!document.hidden) { refresh(); flushOutbox(); checkHealth(); } });
   // belt and suspenders for the never-lose guarantee: even if every event
   // path misfires, unsent memos are retried on a slow heartbeat
@@ -10130,13 +10140,12 @@ function saItemHtml(it, i){
     ${coverageSec}
     ${evidenceSec}
     ${reqVerifySec}
-    ${S('Open it', openBody)}
+    ${S('Open it', openBody + infraBody)}
     ${S('Run state', interruptedBody)}
     ${S('PR', prRepairBody)}
     ${S('Also carries', carriesBody)}
     ${S('What to check', checkBody)}
     ${S('Result', reportBody)}
-    ${S(infra?.title || '検証基盤エラー', infraBody)}
     ${S('Screenshots', shotsBody)}
     ${S('Steps', stepsBody)}
     ${S('Proof', proofInner)}
@@ -10629,11 +10638,12 @@ $('seeall').addEventListener('click', async (e) => {
   // meta-row toggle (H23): tap the token chip to reveal the technique breakdown.
   if (k === 'tok') { b.classList.toggle('open'); $('seeall').querySelector(`#tokdetail${i}`)?.classList.toggle('open'); return; }
   if (k === 'reverify') {
+    const copy = retryCopy(it.goal?.projectId || state.active);
     b.disabled = true;
-    b.textContent = '再検証中…';
+    b.textContent = copy.retrying;
     let r; try { r = await fetch(withKey(`/api/goals/${it.goalId}/reverify`), { method: 'POST' }); }
-    catch { showErr('再検証に失敗'); b.disabled = false; return; }
-    if (r.ok) await refresh(); else showErr('再検証に失敗');
+    catch { showErr(copy.failed); b.disabled = false; b.textContent = copy.retry; return; }
+    if (r.ok) await refresh(); else { showErr(copy.failed); b.disabled = false; b.textContent = copy.retry; }
     return;
   }
   // Open it (P1, Masa 2026-07-21): start the thing the worker said to run, then open it.

package/engine/completion-contract.mjs

@@ -0,0 +1,19 @@
+const CHANGE_REQUEST_RE = /(?:修正|直して|解決して|改善して|一致させ|統一して|実装して|追加して|削除して|変更して|作って|fix\b|solve\b|resolve\b|implement\b|change\b|make\b)/i;
+const REFUSAL_RESULT_RE = /(?:直せません|修正できません|解決できません|実現できません|不可能です|cannot\s+(?:be\s+)?(?:fix(?:ed)?|solv(?:e|ed)|implement(?:ed)?|done)|can't\s+(?:fix|solve|implement|do)|unable\s+to\s+(?:fix|solve|implement))/i;
+
+export const COMPLETION_CONTRACT_PROMPT = [
+  'MANAGER COMPLETION CONTRACT:',
+  '- The latest explicit user instruction can revise an older product decision or code comment. Treat old comments as context, not proof that the requested change is impossible. If the change has consequences, explain them and implement the latest request; only a real safety, permission, legal, or external-system constraint is a blocker.',
+  '- A test must assert the observable outcome the user requested. Never replace the request with a tooltip, explanation, or a test that permanently preserves the opposite behavior.',
+  '- Do not report a change request as complete with “cannot fix”, “直せません”, or equivalent. If a real external blocker remains, state the concrete blocker and leave the work explicitly unfinished instead of presenting it for Review as completed.',
+].join('\n');
+
+export function completionContractIssue({ request = '', followUp = '', result = '', status = '' } = {}) {
+  if (status !== 'done') return null;
+  const requested = `${request}\n${followUp}`;
+  if (!CHANGE_REQUEST_RE.test(requested) || !REFUSAL_RESULT_RE.test(String(result))) return null;
+  return {
+    kind: 'unresolved-refusal',
+    reason: '変更・解決依頼に対して「対応不能」で終了しており、要求した結果が実現されていません。',
+  };
+}

package/engine/server.mjs

@@ -26,6 +26,7 @@ import { createSerialQueue } from './lib.mjs';
 import { parseRequirementEvidence, unmappedChangedFiles } from './lib.mjs';
 import { isPrClosed } from './lib.mjs';
 import { collectProjectRules } from './lib.mjs';
+import { COMPLETION_CONTRACT_PROMPT, completionContractIssue } from './completion-contract.mjs';
 import { FALLBACK_CODEX_MODELS, loadCodexModels } from './codex-models.mjs';
 import { CLAUDE_MODELS } from './claude-models.mjs';
 import { classifyWorkerPreflight, shouldProbeClaudeAuth, workerCanStart } from './worker-availability.mjs';
@@ -2378,8 +2379,13 @@ async function startEphemeralApp(projectDir) {
   const home = join(logDir, `ephemeral-${Date.now()}`);
   mkdirSync(home, { recursive: true });
   seedEphemeralHistory(home, projectDir);
+  // Managed worktrees intentionally do not copy node_modules. An ephemeral
+  // preview is still this Manager package, so let it resolve the already
+  // installed runtime dependencies from the parent rather than crashing before
+  // its readiness endpoint can answer (for example `puppeteer-core`).
+  const nodePath = testNodePath(projectDir);
   const child = spawn('node', [join(projectDir, 'engine', 'server.mjs')], {
-    env: { ...process.env, MANAGER_PORT: String(port), MANAGER_HOME: home, MANAGER_OPEN_BROWSER: '0', MANAGER_SEED_RUNNING_FIXTURE: '1' },
+    env: { ...process.env, NODE_PATH: nodePath, MANAGER_PORT: String(port), MANAGER_HOME: home, MANAGER_OPEN_BROWSER: '0', MANAGER_SEED_RUNNING_FIXTURE: '1' },
     cwd: projectDir, stdio: 'ignore',
   });
   ephemeralChildren.add(child);
@@ -2721,6 +2727,7 @@ async function runTask(task) {
     const prompt = [
       managerHandoffPrompt(goal, plan),
       workerPrompt(task, goal, null, runAgent, runContext),
+      COMPLETION_CONTRACT_PROMPT,
       '(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();
@@ -2765,6 +2772,17 @@ async function runTask(task) {
     task.unmappedChangedFiles = unmappedChangedFiles(task.changedFiles, task.requirementEvidence);
     task.requirementVerification = { proof: task.proof ?? null, run: task.run ?? null, changedFiles: task.changedFiles };
     sendProcessAct(task, `Detected ${task.changedFiles.length} changed file${task.changedFiles.length === 1 ? '' : 's'} from this follow-up.`);
+    const completionIssue = completionContractIssue({
+      request: goal?.text,
+      followUp: task.detail,
+      result: task.result,
+      status: task.status,
+    });
+    if (completionIssue) {
+      task.status = 'interrupted';
+      task.result = `[未解決: ${completionIssue.reason}]\n\n${task.result}`.slice(0, RESULT_MAX);
+      sendProcessAct(task, 'Completion contract rejected an unresolved refusal — kept out of Review for retry.');
+    }
     task.finishedAt = new Date().toISOString();
     task.resultAttemptId = task.attemptId;
     saveTask(task);
@@ -2799,7 +2817,7 @@ async function runTask(task) {
     const runEffort = task.effort ?? goal?.effort;
     const resumeSession = shouldFreshSessionForPlan(plan) ? undefined : (goalSessionFor(goal, runAgent) ?? undefined);
     const runContext = workerRunContext({ goal, task, agent: runAgent, model: runModel, effort: runEffort, permissionMode, cwd: workDir, resumeSession, plan });
-    const prompt = [managerHandoffPrompt(goal, plan), workerPrompt(task, goal, null, runAgent, runContext)].filter(Boolean).join('\n\n');
+    const prompt = [managerHandoffPrompt(goal, plan), workerPrompt(task, goal, null, runAgent, runContext), COMPLETION_CONTRACT_PROMPT].filter(Boolean).join('\n\n');
     sendWorkerContextAct(task, runContext);
     sendProcessAct(task, `Starting ${runAgent === 'codex' ? 'Codex' : 'Claude Code'} worker (attempt ${attempt}/${maxAttempts}).`);
     let r = await runWorkerAgent({
@@ -2833,7 +2851,7 @@ async function runTask(task) {
       sendWorkerContextAct(task, freshContext);
       r = await runWorkerAgent({
         agent: runAgent,
-        prompt: workerPrompt(task, goal, null, runAgent, freshContext), cwd: workDir, tools: WORKER_TOOLS,
+        prompt: [workerPrompt(task, goal, null, runAgent, freshContext), COMPLETION_CONTRACT_PROMPT].join('\n\n'), cwd: workDir, tools: WORKER_TOOLS,
         permissionMode,
         model: runModel, effort: runEffort, onEvent: (line) => sendAct(task, line),
         images: goal?.images,
@@ -2984,6 +3002,17 @@ async function runTask(task) {
     task.result = `[未完了の疑い: 変更ファイル0件かつ応答が途中で切れています] ${task.result}`.slice(0, RESULT_MAX);
     sendProcessAct(task, 'Result looked incomplete (0 changed files, cut-off text) — marked interrupted instead of done for human review.');
   }
+  const completionIssue = completionContractIssue({
+    request: goal?.text,
+    followUp: task.reply ? task.detail : '',
+    result: task.result,
+    status: task.status,
+  });
+  if (completionIssue) {
+    task.status = 'interrupted';
+    task.result = `[未解決: ${completionIssue.reason}]\n\n${task.result}`.slice(0, RESULT_MAX);
+    sendProcessAct(task, 'Completion contract rejected an unresolved refusal — kept out of Review for retry.');
+  }
   task.finishedAt = new Date().toISOString();
   task.resultAttemptId = task.attemptId;
   saveTask(task);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@galda/cli",
-  "version": "0.10.87",
+  "version": "0.10.88",
   "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": {