@galda/cli 0.10.57 → 0.10.60

package/app/index.html

@@ -7055,9 +7055,10 @@ function renderFsTodo(){
   for (const b of el.querySelectorAll('[data-fslog]')) b.onclick = () => fsToggleLog(Number(b.dataset.fslog));
   for (const b of el.querySelectorAll('[data-fsglog]')) b.onclick = () => toggleGoalDetail(Number(b.dataset.fsglog));
   for (const b of el.querySelectorAll('[data-fsreverify]')) b.onclick = async () => {
-    b.disabled = true; b.textContent = 'Verifying…';
+    b.disabled = true; b.textContent = 'Starting…';
     const r = await fetch(withKey(`/api/goals/${b.dataset.fsreverify}/reverify`), { method: 'POST' });
-    if (!r.ok) showErr('再検証に失敗'); else await refresh();
+    if (!r.ok) { showErr(document.documentElement.lang === 'ja' ? 'Retryに失敗しました' : 'Retry failed.'); b.disabled = false; b.textContent = 'Retry'; }
+    else await refresh();
   };
   for (const b of el.querySelectorAll('[data-fsarchive]')) b.onclick = async () => {
     const r = await fetch(withKey(`/api/goals/${b.dataset.fsarchive}/archive`), { method: 'POST' });
@@ -9452,6 +9453,11 @@ async function saEnhanceShotBoxes(){
 }
 function saRender(){
   const S = $('seeall'); if (!S) return;
+  // SSE, diff fetches and action responses rebuild the dialog. Remember the
+  // review body's position before replacing it so reading never jumps to top.
+  const previousBody = S.querySelector('.sa-body');
+  const previousScrollTop = previousBody?.scrollTop ?? 0;
+  const restoreScroll = Boolean(previousBody) && SA.focusGid == null;
   const pending = SA.items.filter((x) => x.st === 0);
   const green = pending.filter((x) => !x.testResult || x.testResult.ok !== false).length;
   // Solo mode = a single "Needs you" goal opened as this card. It is not a review batch, so the
@@ -9464,6 +9470,16 @@ function saRender(){
   const hdKeys = solo ? '↑↓ MOVE · ESC CLOSE' : '↑↓ MOVE · A APPROVE · D DISMISS · ESC CLOSE';
   S.innerHTML = `<div class="sa-panel"><div class="sa-hd"><span class="t">${hdTitle}</span><span class="c">${c}</span><span class="sa-count">${posLabel}</span><span class="keys">${hdKeys}</span><button class="sa-x" data-saact="close:0">✕</button></div>
     <div class="sa-body">${SA.items.length ? SA.items.map((it, i) => saItemHtml(it, i)).join('') : '<div class="sa-empty">Nothing waiting for review.</div>'}</div></div>`;
+  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.
+  body?.addEventListener('scroll', () => {
+    if (SA.focusGid == null) return;
+    SA.focusGid = null;
+    if (SA.focusTimer) { clearTimeout(SA.focusTimer); SA.focusTimer = null; }
+  }, { once: true });
   // 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.
@@ -9779,7 +9795,7 @@ $('seeall').addEventListener('click', async (e) => {
     if (link) link.disabled = true;
     if (v) v.textContent = 'Starting it…';
     let r = null, body = {};
-    try { r = await fetch(`/api/tasks/${it.run.taskId}/preview`, { method: 'POST' }); body = await r.json().catch(() => ({})); }
+    try { r = await fetch(withKey(`/api/tasks/${it.run.taskId}/preview`), { method: 'POST' }); body = await r.json().catch(() => ({})); }
     catch (e) { body = { error: String(e.message ?? e) }; }
     if (link) link.disabled = false;
     const noteText = it.run.note || it.run.cmd || 'Starts when you press it.';

package/engine/lib.mjs

@@ -1646,6 +1646,25 @@ export function detectRequestLanguage(text) {
   return /[぀-ヿ㐀-鿿]/.test(text || '') ? 'ja' : 'en';
 }
 
+export function testFailureReason({ text = '', count = 0, attempts = null } = {}) {
+  const n = Number(count) || 0;
+  const tries = attempts == null ? null : Number(attempts) || 0;
+  if (detectRequestLanguage(text) === 'ja') {
+    return tries == null ? `テスト失敗 ${n} 件` : `テスト失敗 ${n} 件（自動修正 ${tries} 回後も失敗）`;
+  }
+  const tests = `${n} test${n === 1 ? '' : 's'} failed`;
+  return tries == null ? tests : `${tests} after ${tries} automatic fix attempt${tries === 1 ? '' : 's'}`;
+}
+
+export function manualTestRetryPrompt({ text = '', count = 0, detail = '' } = {}) {
+  const n = Number(count) || 0;
+  const output = String(detail || '(no detailed output captured)').slice(0, 3200);
+  if (detectRequestLanguage(text) === 'ja') {
+    return `Retryが押されました。前回のテスト失敗 ${n} 件を確認し、元のタスクの作業環境で原因を修正してからテストを再実行してください。関係のないリファクタは行わないでください。\n\n前回のテスト出力:\n${output}`;
+  }
+  return `Retry was requested. Inspect the previous ${n} failing test${n === 1 ? '' : 's'}, fix the cause in the original task worktree, then rerun the tests. Do not make unrelated changes.\n\nPrevious test output:\n${output}`;
+}
+
 // SLICE 2 intent routing (docs/HANDOFF-ONBOARDING-conversational.md): the single
 // composer is conversation-first — a chat/help/settings message ("what can you
 // do?", "reply in English", a greeting) must get a conversational reply, NOT a
@@ -2583,6 +2602,7 @@ export function workerPrompt(task, goal, lastFailure, agentName = 'claude-code')
     // (research, summaries, answers) simply skip this — the worker decides,
     // never us.
     `- 人が実際に動かして結果を見られる依頼なら、最後にプロジェクト直下へ ${RUN_DECL_FILE} を書く: {"cmd": "<起動コマンド>", "url": "<開くURL>", "note": "<何が見えるか1行>", "check": "<人に何を見て判断してほしいか1行>", "title": "<この作業のチケット名。対象が分かる程度に説明的で動作を含む短い名前。例「看板モードの追加」「ヘッダのズレの修正」。依頼文をそのまま写さない>"}。Manager はこれを使って人に「動かして見る」と「何を見ればいいか」を出す。動かして見せるものが無い依頼（調査・要約・質問への回答など）では書かなくてよい。`,
+    `- 「ローカルで見れない」というフィードバックには、ターミナル操作や npx の再実行を案内して終わらないこと。${RUN_DECL_FILE} を実際に開ける内容へ更新し、可能なら自分でURLへ接続確認してから完了すること。`,
     // Links are rendered as links. The product cannot tell a real one from a
     // decorative one, so the only place this can be held is here.
     `- 自分が作っていない場所へのリンクを報告に書かない（存在しないURLや「詳しくはここ」のような飾りのリンクは、そのままリンクとして人に表示される）。`,

package/engine/server.mjs

@@ -21,7 +21,7 @@ import { homedir } from 'node:os';
 import { fileURLToPath } from 'node:url';
 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 } from './lib.mjs';
-import { parseStreamEvents, parseCodexEvents, buildCodexArgs, parsePlan, refinePlanTasks, canEditGoal, canDeleteGoal, shouldAskClarification, workerExitReason, isForcedStop, taskStatusAfterVerify, workerResultText, taskCountsAsComplete, resolveWantsPR, resolveGoalSource, buildGoalSummary, 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, 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, hasTestRelevantChanges, parseFailingTestNames, classifyTestGate, isInconclusiveTestRun, countInfraFlakes, classifyRunFailures, classifyVerifyGate, 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, isNothingVerifiable, classifyComposerIntentHeuristic, detectPauseIntent, buildIntentPrompt, parseIntentResponse, buildBoardSnapshot, validateBoardSnapshot, boardIsEmpty, decideBoardPull, resolveConnectedAgents, pickAvailableAgent, pickUtilityAgent, utilityModel, liveTakeoverDecision, shouldOpenBrowserOnBoot, goalTaskTitle, buildGoalTask, RUN_DECL_FILE, parseRunDeclaration, SHOT_DIR, collectShotNames, parseGoalAddress, REPLY_OUTCOMES, REPLY_INTENTS, buildReplyIntentPrompt, parseReplyIntent, buildGoalMessage, serializeGoalMessage, parseGoalMessages } from './lib.mjs';
+import { parseStreamEvents, parseCodexEvents, buildCodexArgs, parsePlan, refinePlanTasks, canEditGoal, canDeleteGoal, shouldAskClarification, workerExitReason, isForcedStop, taskStatusAfterVerify, workerResultText, taskCountsAsComplete, resolveWantsPR, resolveGoalSource, buildGoalSummary, 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, 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, hasTestRelevantChanges, parseFailingTestNames, classifyTestGate, isInconclusiveTestRun, countInfraFlakes, classifyRunFailures, classifyVerifyGate, 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, classifyComposerIntentHeuristic, detectPauseIntent, buildIntentPrompt, parseIntentResponse, buildBoardSnapshot, validateBoardSnapshot, boardIsEmpty, decideBoardPull, resolveConnectedAgents, pickAvailableAgent, pickUtilityAgent, utilityModel, liveTakeoverDecision, shouldOpenBrowserOnBoot, goalTaskTitle, buildGoalTask, RUN_DECL_FILE, parseRunDeclaration, SHOT_DIR, collectShotNames, parseGoalAddress, REPLY_OUTCOMES, REPLY_INTENTS, buildReplyIntentPrompt, parseReplyIntent, buildGoalMessage, serializeGoalMessage, parseGoalMessages } from './lib.mjs';
 import { createSerialQueue } from './lib.mjs';
 import { collectProjectRules } from './lib.mjs';
 import { FALLBACK_CODEX_MODELS, loadCodexModels } from './codex-models.mjs';
@@ -2129,7 +2129,7 @@ async function startEphemeralApp(projectDir) {
     try {
       const key = readFileSync(join(home, 'secret.key'), 'utf8').trim();
       const r = await fetch(`http://localhost:${port}/api/state?key=${key}`);
-      if (r.ok) return { url: `http://localhost:${port}/?key=${key}`, kill: () => { try { child.kill(); } catch {} } };
+      if (r.ok) return { url: `http://localhost:${port}/?key=${key}`, child, kill: () => { try { child.kill(); } catch {} } };
     } catch { /* not up yet */ }
   }
   try { child.kill(); } catch {}
@@ -2259,6 +2259,19 @@ async function startPreview(task) {
     live.expires = Date.now() + PREVIEW_IDLE_MS;
     return { url: live.url, reused: true };
   }
+  // A Manager app cannot preview itself on its declared production port: that
+  // port is already occupied, and its bare URL requires an access key. Boot the
+  // edited checkout with isolated state and return its real key-bearing URL.
+  let managerPackage = false;
+  try { managerPackage = JSON.parse(readFileSync(join(dir, 'package.json'), 'utf8')).name === '@galda/cli'; } catch {}
+  if (managerPackage && existsSync(join(dir, 'engine', 'server.mjs')) && existsSync(join(dir, 'app', 'index.html'))) {
+    const eph = await startEphemeralApp(dir);
+    if (!eph) throw new Error('Could not start the local preview. Please try again.');
+    const entry = { proc: eph.child, url: eph.url, expires: Date.now() + PREVIEW_IDLE_MS };
+    previews.set(task.id, entry);
+    eph.child.on('close', () => { if (previews.get(task.id) === entry) previews.delete(task.id); });
+    return { url: eph.url, reused: false, started: true };
+  }
   // Declared a URL but no command: the thing is already reachable (an app the
   // worker left running, a hosted page). Nothing to start.
   if (!cmd) return { url, reused: false, started: false };
@@ -3174,7 +3187,7 @@ async function verifyGate(goal, prProject, siblings, baseProject, activityTask =
   if (testsFailing) {
     const failure = recordGoalFailure(goal, {
       kind: 'test',
-      reason: `テスト失敗 ${blockingFailCount} 件`,
+      reason: testFailureReason({ text: goal.text, count: blockingFailCount }),
       changedFiles,
       testResult: goal.testResult,
     });
@@ -3186,7 +3199,7 @@ async function verifyGate(goal, prProject, siblings, baseProject, activityTask =
       goal.autoRework = {
         ...(goal.autoRework ?? {}),
         testFailures: used + 1,
-        lastReason: `テスト失敗 ${blockingFailCount} 件`,
+        lastReason: testFailureReason({ text: goal.text, count: blockingFailCount }),
         lastAt: new Date().toISOString(),
       };
       goal.blocked = null;
@@ -3232,7 +3245,7 @@ async function verifyGate(goal, prProject, siblings, baseProject, activityTask =
     }
     goal.blocked = {
       kind: gate.kind,
-      reason: testsFailing ? `テスト失敗 ${blockingFailCount} 件（自動修正 ${Number(goal.autoRework?.testFailures ?? 0)} 回後も失敗）`
+      reason: testsFailing ? testFailureReason({ text: goal.text, count: blockingFailCount, attempts: goal.autoRework?.testFailures ?? 0 })
         : nothingVerifiable ? '検証条件つきのタスクなのに、ファイル変更も proof も無い（約束した検証が行われていない）'
         : '検証(proof)が全タスク分そろっていない（requireVerifyPass）',
     };
@@ -4525,8 +4538,8 @@ const server = createServer(async (req, res) => {
     return json(res, 200, goal);
   }
 
-  // Re-run the Manager-verified gate for a blocked goal (tests + proof), from
-  // Attention's 再検証 button. Passes → review; still failing → stays blocked.
+  // Retry a test-blocked goal by sending the saved failure back to its worker.
+  // Other blocked states still re-run the verification gate in place.
   const reverifyMatch = url.pathname.match(/^\/api\/goals\/(\d+)\/reverify$/);
   if (reverifyMatch && req.method === 'POST') {
     const goal = goals.find((g) => g.id === Number(reverifyMatch[1]));
@@ -4535,6 +4548,19 @@ const server = createServer(async (req, res) => {
     const project = projects.find((p) => p.id === goal.projectId);
     if (!project) return json(res, 404, { error: 'project not found' });
     const siblings = tasks.filter((t) => t.goalId === goal.id);
+    if (goal.status === 'blocked' && goal.blocked?.kind === 'test') {
+      const count = goal.testResult?.newFailures?.length ?? goal.testResult?.failed ?? 0;
+      const prompt = manualTestRetryPrompt({ text: goal.text, count, detail: goal.testResult?.detail });
+      goal.status = 'running';
+      goal.blocked = null;
+      goal.startedAt = null;
+      // A manual retry is a fresh user decision, independent of the automatic
+      // two-attempt budget. Keep the selected agent/model and same conversation.
+      goal.autoRework = { ...(goal.autoRework ?? {}), testFailures: 0, manualRetryAt: new Date().toISOString() };
+      saveGoal(goal);
+      const task = queueReplyTask(goal, prompt, { from: 'manager', via: 'retry' });
+      return json(res, 200, { goal, task, repairing: true });
+    }
     const wdir = (goalWorkDirs.get(goal.id) && existsSync(goalWorkDirs.get(goal.id))) ? goalWorkDirs.get(goal.id) : project.dir;
     // 再検証 is the explicit human "run the tests now" — force the gate on for this
     // one goal even when the default (automatic) gate is off.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@galda/cli",
-  "version": "0.10.57",
+  "version": "0.10.60",
   "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": {