@galda/cli 0.10.90 → 0.10.91

package/app/index.html

@@ -6236,13 +6236,14 @@ function summarizeAttention(goal, tasks){
     decision = ja
       ? 'worker に失敗したテスト出力を確認させ、回帰を修正して、プロジェクトのテストを再実行するよう指示してください。'
       : 'Ask the worker to inspect the failing test output, fix the regression, and rerun the project tests.';
-  } else if (/interrupted/i.test(goal.status || '') || /interrupted/i.test(compact)) {
+  } else if (failed?.status === 'interrupted' || /interrupted/i.test(goal.status || '') || /interrupted/i.test(compact)) {
+    const noReport = !compact;
     happened = ja
-      ? `必要だったこと: ${need}。${taskName ? `「${taskName}」の作業中に、` : ''}実行が中断されました: ${compact.slice(0, 140) || '詳細ログは取得できませんでした'}。`
-      : `Needed: ${need}. The run was interrupted${taskName ? ` while working on "${taskName}"` : ''}: ${compact.slice(0, 140) || 'no detailed log was captured'}.`;
+      ? `必要だったこと: ${need}。${taskName ? `「${taskName}」の作業中に、` : ''}${noReport ? 'Managerとworkerの接続が完了報告前に切れました。run limitだったとは確認できません' : `実行が中断されました: ${compact.slice(0, 140)}`}。`
+      : `Needed: ${need}. ${noReport ? 'The Manager lost its worker connection before a final report was saved; this is not confirmed as a run-limit stop' : `The run was interrupted${taskName ? ` while working on "${taskName}"` : ''}: ${compact.slice(0, 140)}`}.`;
     decision = ja
-      ? '具体的な次の一手を指定して Resume するか、不要なら Archive してください。'
-      : 'Resume with a concrete next step, or archive it if the work is no longer needed.';
+      ? '依頼を言い直したり範囲を狭めたりせず、そのまま Resume できます。既存の作業ファイルから続行します。'
+      : 'Resume the same request without rewording or narrowing it. Work continues from the existing files.';
   } else if (/blocked/i.test(goal.status || '') || compact) {
     happened = ja
       ? `必要だったこと: ${need}。Review に進む前に停止しました${compact ? `: ${compact.slice(0, 160)}` : ''}。`
@@ -6505,7 +6506,12 @@ function renderFsRail(){
     const running = pgoals.some((g) => fsBoardBucket(g, ptasks) === 'doing'
       && !(g.status === 'blocked' && g.blocked?.kind === 'paused'));
     const td = ptasks.filter((t) => ['running', 'queued'].includes(t.status)).length;
-    const rev = buildReviewDigest({ goals: pgoals, tasks: ptasks }).length;
+    // The rail's eye icon is the review/attention count. Needs-you goals
+    // (needsInput/needsApproval/blocked, minus Paused) block on the person the
+    // same way an open Review does, so they belong in the same count (Masa
+    // task 854: "needs youも目のアイコンのマークにいれれますか？").
+    const needsCount = pgoals.filter((g) => fsBoardBucket(g, ptasks) === 'needs').length;
+    const rev = buildReviewDigest({ goals: pgoals, tasks: ptasks }).length + needsCount;
     const pcount = `<span class="pcount" title="${rev} review${rev === 1 ? '' : 's'} ready · ${td} to-do">`
       + `<span class="pc-item rev${rev ? '' : ' zero'}">${FS_ICO_SCAN}${rev}</span>`
       + `<span class="pc-item td">${FS_ICO_TD}${td}</span></span>`;
@@ -7249,7 +7255,9 @@ function taskErrorMeta(t, goal){
   // Reclassify interrupted tasks from their concrete exit text instead of
   // trusting stale persisted metadata. Older records labelled an external
   // SIGTERM/restart as an Agent failure and permanently carried a Codex switch.
-  const inferred = inferErrorClass(t.result, t.agent || goal?.agent);
+  const inferred = t.status === 'interrupted' && !String(t.result || '').trim()
+    ? { owner: 'galda-infra', reason: 'interrupted', actionIds: ['retry', 'view-log'] }
+    : inferErrorClass(t.result, t.agent || goal?.agent);
   const raw = t.status === 'interrupted' ? inferred : (t.errorClass || inferred);
   const meta = { ...raw, actionIds: (raw.actionIds || []).filter((id) => !['run-with-codex', 'run-with-claude'].includes(id)) };
   const lang = failureLang(), C = FAILURE_COPY[lang];
@@ -9246,7 +9254,7 @@ function closeSummaryPera(){ $('summaryOverlay').classList.remove('show'); }
 // 4 archived / 5 reverted — session-local; the server round-trip is what
 // settles it (see saSync's "returned to review" reset).
 // ============================================================================
-const SA = { open: false, items: [], diffs: {}, actLog: {}, msgs: {}, convOpen: new Set(), shotBox: {}, detOpen: new Set(), diffOpen: new Set(), attachments: {}, focusGid: null, focusScrolled: false, focusTimer: null, cur: null, solo: null };
+const SA = { open: false, items: [], diffs: {}, actLog: {}, msgs: {}, convOpen: new Set(), shotBox: {}, detOpen: new Set(), diffOpen: new Set(), attachments: {}, focusGid: null, focusScrolled: false, focusTimer: null, cur: null, solo: null, lastMarkup: null };
 function saReplyAttachmentsHtml(goalId){
   return (SA.attachments[goalId] || []).map((a, i) => `<span class="sa-replythumb"><img src="${esc(a.url)}" alt="Attached image"><button type="button" data-saarm="${goalId}:${i}" aria-label="Remove image">×</button></span>`).join('');
 }
@@ -9936,7 +9944,9 @@ function saItemHtml(it, i){
     const tests = it.testResult?.ran
       ? (it.testResult.ok ? 'Verification completed before the interruption.' : 'Verification remains incomplete or failed.')
       : 'Verification has not completed yet.';
-    const reason = String(t.result || 'The worker reached its run limit before finishing.').replace(/\s+/g, ' ').slice(0, 240);
+    const reason = String(t.result || (failureLang() === 'ja'
+      ? 'Managerとworkerの接続が完了報告前に切れました。run limitだったとは確認できません。作業ファイルを保持したままResumeできます。'
+      : 'The Manager lost its worker connection before a final report was saved. This is not confirmed as a run-limit stop. Resume keeps the existing work files.')).replace(/\s+/g, ' ').slice(0, 240);
     return `<div class="rinfra"><p class="rbody v"><b>Interrupted — resume available</b><br>${esc(reason)}</p>`
       + (last ? `<p class="rbody v">Last recorded step: ${esc(String(last).slice(0, 220))}</p>` : '')
       + (files.length ? `<p class="rbody v">Changed so far: ${esc(files.join(' · '))}</p>` : '')
@@ -10220,6 +10230,9 @@ function saRender(){
   const previousBody = S.querySelector('.sa-body');
   const previousScrollTop = previousBody?.scrollTop ?? 0;
   const restoreScroll = Boolean(previousBody) && SA.focusGid == null;
+  const focusedReplyGoal = document.activeElement?.matches?.('#seeall .sa-chat')
+    ? document.activeElement.dataset.goal
+    : null;
   const pending = SA.items.filter((x) => x.st === 0);
   const green = pending.filter((x) => !x.testResult || x.testResult.ok !== false).length;
   // Solo mode = a single "Needs you" goal opened as this card. It is not a review batch, so the
@@ -10230,8 +10243,16 @@ function saRender(){
   const curIdx = pending.findIndex((x) => x.goalId === SA.cur);
   const posLabel = solo ? '' : (pending.length && curIdx >= 0 ? `${curIdx + 1} / ${pending.length}` : '');
   const hdKeys = solo ? '↑↓ MOVE · ESC CLOSE' : '↑↓ MOVE · A APPROVE · D DISMISS · ESC CLOSE';
-  S.innerHTML = `<div class="sa-panel"><div class="sa-hd"><span class="t">${hdTitle}</span><span class="c">${c}</span><span class="sa-count">${posLabel}</span><span class="keys">${hdKeys}</span><button class="sa-x" data-saact="close:0">✕</button></div>
+  const markup = `<div class="sa-panel"><div class="sa-hd"><span class="t">${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>`;
+  // Most SSE goal/task events are unrelated to the card the person is reading.
+  // Replacing #seeall with identical markup destroyed and recreated every node,
+  // producing a blank flash and resetting browser-owned state. Keep the current
+  // dialog DOM when its semantic markup did not change; the board behind it can
+  // still refresh normally.
+  if (SA.lastMarkup === markup && S.querySelector('.sa-panel')) return;
+  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
@@ -10285,6 +10306,10 @@ function saRender(){
       if (!SA.focusTimer) SA.focusTimer = setTimeout(() => { SA.focusGid = null; SA.focusTimer = null; }, 1400);
     });
   }
+  // A relevant status change may legitimately rebuild the card. Restore the
+  // empty reply box's focus afterwards; non-empty input is protected earlier
+  // by saSync and is never rebuilt at all.
+  if (focusedReplyGoal != null) S.querySelector(`.sa-chat[data-goal="${focusedReplyGoal}"]`)?.focus({ preventScroll: true });
 }
 // keep the open Ledger in sync with server state (SSE/refresh → render() → here).
 // An item whose goal has RETURNED to 'review' (retest ×3 green, rework done,
@@ -10332,6 +10357,7 @@ async function saFetchDiffs(){
 }
 function saOpen(focusGoalId, solo = false){
   SA.open = true;
+  SA.lastMarkup = null; // a newly opened/focused dialog always gets a clean first paint
   SA.solo = solo ? focusGoalId : null;   // solo = a single "Needs you" goal opened as one card
   SA.items = saLedgerRows().map((r) => ({ ...r, st: 0 }));
   SA.detOpen = new Set();
@@ -10362,7 +10388,7 @@ function saOpen(focusGoalId, solo = false){
   saFetchDiffs();
   saEnhanceShotBoxes();
 }
-function saClose(){ SA.open = false; SA.solo = null; SA.focusGid = null; if (SA.focusTimer) { clearTimeout(SA.focusTimer); SA.focusTimer = null; } $('seeall').classList.remove('open'); }
+function saClose(){ SA.open = false; SA.solo = null; SA.focusGid = null; SA.lastMarkup = null; if (SA.focusTimer) { clearTimeout(SA.focusTimer); SA.focusTimer = null; } $('seeall').classList.remove('open'); }
 // Continuous review (Masa 2026-07-08): move a persistent "current" card with
 // ↑↓ and auto-advance to the next pending after each verdict — same screen, same
 // Ledger, so approving flows without hunting for the next one.

package/engine/lib.mjs

@@ -46,6 +46,19 @@ export function isCommandOnPath(cmd, { pathEnv = '', sep = ':', pathExt = '', ex
   return false;
 }
 
+// A directory mention becomes an execution target only when the person labels
+// it as a working/development directory. Bare paths remain ordinary context so
+// a worker is never silently moved just because a filename appeared in prose.
+export function parseRequestedWorkspaceDir(text) {
+  const source = String(text ?? '');
+  const marker = /(?:working\s*directory|workdir|作業(?:ディレクトリ|場所)|開発場所)\s*(?:を|は|:|：|=|＝)?\s*/i;
+  const found = source.match(marker);
+  if (!found) return null;
+  const rest = source.slice((found.index ?? 0) + found[0].length);
+  const path = rest.match(/\/(?:[A-Za-z0-9._~@%+\-]+\/)*[A-Za-z0-9._~@%+\-]+\/?/);
+  return path ? path[0].replace(/\/+$/, '') || '/' : null;
+}
+
 // Galda makes small internal LLM calls (composer-intent classification, proof-card
 // summaries, board Q&A). These prefer Claude Code + haiku (cheap + fast), but must
 // still work on a Codex-only machine — so they run on whatever agent is available,
@@ -2953,12 +2966,10 @@ export function buildEphemeralSeedProjects(cwd) {
 // goal/task written once (e.g. a memo the user wants to survive a restart)
 // is never lost: later lines for the same id override earlier ones (last
 // write wins), and a goal write with `deleted:true` removes it. Mid-flight
-// work can't resume across a restart, so it's marked retryable rather than
-// silently dropped: a 'running' task becomes 'interrupted', a 'planning'
-// goal goes back to 'stacked', and a 'running' goal with no queued task left
-// settles via nextGoalStatus() — a wantsPR goal lands in 'review' just like
-// the normal completion path, even if the crash happened before the PR got
-// created (the caller re-attempts createGoalPR for those on restart).
+// work gets one bounded automatic recovery: a 'running' task becomes 'queued'
+// once and resumes from the files/session already on disk. A second interrupted
+// boot stays visible for human retry so a crash loop cannot relaunch forever.
+// A 'planning' goal still goes back to 'stacked'.
 // `reviewDefinitions` (projectId -> definition, as saved by
 // validateReviewDefinition) lets that settling honor each project's own
 // review definition (task 45); a project with none configured falls back to
@@ -2975,7 +2986,21 @@ export function replayQueueLog(rawText, reviewDefinitions = {}) {
     } catch { /* skip a corrupt line rather than lose the whole log */ }
   }
   for (const t of tasks) {
-    if (t.status === 'running') t.status = 'interrupted';
+    if (t.status !== 'running') continue;
+    // A Manager restart is an infrastructure interruption, not a request for
+    // human judgment. Resume the same task once from the files/session already
+    // on disk. The bounded counter prevents a crash-looping task from being
+    // relaunched forever; a second interrupted boot remains visibly retryable.
+    const recoveries = Number(t.restartRecoveries ?? 0);
+    if (recoveries < 1) {
+      t.status = 'queued';
+      t.restartRecoveries = recoveries + 1;
+      t.recoveredFromRestart = true;
+      t.startedAt = null;
+      t.finishedAt = null;
+    } else {
+      t.status = 'interrupted';
+    }
   }
   for (const g of goals) {
     if (g.status === 'planning') g.status = 'stacked';

package/engine/requirement-model.mjs

@@ -0,0 +1,196 @@
+export const REQUIREMENT_MODEL_VERSION = 1;
+
+const CLAIM_STATUSES = new Set(['addressed', 'partial', 'not-addressed']);
+const COVERAGE_STATUSES = new Set(['unverified', 'verified', 'failed', 'not-applicable']);
+
+function cleanText(value) {
+  return typeof value === 'string' ? value.trim() : '';
+}
+
+function stablePart(value) {
+  return String(value ?? 'unknown').replace(/[^a-zA-Z0-9_-]+/g, '-');
+}
+
+function requirementId(goalId, messageId) {
+  return `req-${stablePart(goalId)}-${stablePart(messageId)}`;
+}
+
+function attemptId(task) {
+  return cleanText(task?.attemptId) || `attempt-task-${stablePart(task?.id)}`;
+}
+
+function sourceForTask(task) {
+  if (!task?.reply) return 'user';
+  if (task.replyFrom === 'you') return 'user';
+  if (task.replyFrom === 'manager') return 'manager-policy';
+  return 'legacy-unknown';
+}
+
+function messageText(task) {
+  return cleanText(task?.reply ? task.detail : (task?.detail || task?.title));
+}
+
+function normalizeRequirement(value) {
+  const id = cleanText(value?.id);
+  const text = cleanText(value?.text);
+  if (!id || !text) return null;
+  return {
+    id,
+    source: cleanText(value.source) || 'legacy-unknown',
+    messageId: value.messageId == null ? null : String(value.messageId),
+    text,
+    ...(Array.isArray(value.span) && value.span.length === 2 ? { span: value.span } : {}),
+    ...(value.derivedFrom ? { derivedFrom: value.derivedFrom } : {}),
+    ...(value.supersedes ? { supersedes: value.supersedes } : {}),
+  };
+}
+
+function normalizeAttempt(value) {
+  const id = cleanText(value?.id);
+  if (!id) return null;
+  return {
+    id,
+    taskId: value.taskId ?? null,
+    status: cleanText(value.status) || 'unknown',
+    createdAt: value.createdAt ?? null,
+    finishedAt: value.finishedAt ?? null,
+    sessionId: value.sessionId ?? null,
+    worktree: value.worktree ?? null,
+  };
+}
+
+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 coverage = COVERAGE_STATUSES.has(value.coverage) ? value.coverage : 'unverified';
+  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) : [])],
+  };
+}
+
+function taskRequirementIds(task, requirements, legacyMap) {
+  const explicit = Array.isArray(task?.activeRequirementIds)
+    ? task.activeRequirementIds.filter((id) => requirements.some((requirement) => requirement.id === id))
+    : [];
+  if (explicit.length) return [...new Set(explicit)];
+
+  const own = legacyMap[`task-${task?.id}`];
+  if (own) return [own];
+  return requirements.length ? [requirements[0].id] : [];
+}
+
+/**
+ * Convert legacy goal/task records into the canonical requirement model.
+ * This function is pure and idempotent. It never removes legacy fields, so a
+ * rollback can continue reading the original records.
+ */
+export function migrateRequirementModel(goal, goalTasks = []) {
+  const requirements = (Array.isArray(goal?.requirements) ? goal.requirements : [])
+    .map(normalizeRequirement).filter(Boolean);
+  const attempts = (Array.isArray(goal?.attempts) ? goal.attempts : [])
+    .map(normalizeAttempt).filter(Boolean);
+  const evidence = (Array.isArray(goal?.requirementEvidence) ? goal.requirementEvidence : [])
+    .map(normalizeEvidence).filter(Boolean);
+  const legacyRequirementMap = { ...(goal?.legacyRequirementMap || {}) };
+
+  if (!requirements.length && cleanText(goal?.text)) {
+    const id = requirementId(goal.id, 'goal');
+    requirements.push({ id, source: 'user', messageId: `goal-${goal.id}`, text: cleanText(goal.text) });
+    legacyRequirementMap.goal = id;
+  }
+
+  for (const task of goalTasks) {
+    const text = messageText(task);
+    const shouldCreateRequirement = text && (task.reply ? sourceForTask(task) !== 'manager-policy' : !requirements.length);
+    if (shouldCreateRequirement) {
+      const legacyId = `task-${task.id}`;
+      const id = legacyRequirementMap[legacyId] || requirementId(goal?.id, legacyId);
+      if (!requirements.some((requirement) => requirement.id === id)) {
+        requirements.push({ id, source: sourceForTask(task), messageId: legacyId, text });
+      }
+      legacyRequirementMap[legacyId] = id;
+    }
+
+    const id = attemptId(task);
+    if (!attempts.some((attempt) => attempt.id === id)) {
+      attempts.push(normalizeAttempt({
+        id,
+        taskId: task.id,
+        status: task.status,
+        createdAt: task.createdAt,
+        finishedAt: task.finishedAt,
+        sessionId: task.sessionId ?? task.agentSessionId,
+        worktree: task.worktree,
+      }));
+    }
+
+    const requirementIds = taskRequirementIds(task, requirements, legacyRequirementMap);
+    for (const claim of task?.requirementEvidence?.requirements ?? []) {
+      const mappedId = legacyRequirementMap[claim.id] || (requirements.some((requirement) => requirement.id === claim.id) ? claim.id : null);
+      const targetIds = mappedId ? [mappedId] : requirementIds;
+      for (const targetId of targetIds) {
+        const row = normalizeEvidence({
+          requirementId: targetId,
+          attemptId: id,
+          taskId: task.id,
+          claim: claim.status,
+          coverage: 'unverified',
+          evidence: claim.evidence,
+          files: claim.files,
+        });
+        if (row && !evidence.some((item) => item.requirementId === row.requirementId && item.attemptId === row.attemptId && item.taskId === row.taskId)) {
+          evidence.push(row);
+        }
+      }
+    }
+  }
+
+  const tasks = goalTasks.map((task) => ({
+    ...task,
+    activeRequirementIds: taskRequirementIds(task, requirements, legacyRequirementMap),
+  }));
+
+  return {
+    goal: {
+      ...goal,
+      requirementModelVersion: REQUIREMENT_MODEL_VERSION,
+      requirements,
+      attempts,
+      requirementEvidence: evidence,
+      legacyRequirementMap,
+    },
+    tasks,
+  };
+}
+
+export function validateRequirementModel(goal, goalTasks = []) {
+  const errors = [];
+  const requirementIds = new Set();
+  const attemptIds = new Set();
+  for (const requirement of goal?.requirements ?? []) {
+    if (!requirement?.id || requirementIds.has(requirement.id)) errors.push(`invalid requirement id: ${requirement?.id ?? ''}`);
+    requirementIds.add(requirement?.id);
+  }
+  for (const attempt of goal?.attempts ?? []) {
+    if (!attempt?.id || attemptIds.has(attempt.id)) errors.push(`invalid attempt id: ${attempt?.id ?? ''}`);
+    attemptIds.add(attempt?.id);
+  }
+  for (const task of goalTasks) {
+    for (const id of task?.activeRequirementIds ?? []) {
+      if (!requirementIds.has(id)) errors.push(`task ${task.id} references unknown requirement ${id}`);
+    }
+  }
+  for (const row of goal?.requirementEvidence ?? []) {
+    if (!requirementIds.has(row?.requirementId)) errors.push(`evidence references unknown requirement ${row?.requirementId ?? ''}`);
+    if (!attemptIds.has(row?.attemptId)) errors.push(`evidence references unknown attempt ${row?.attemptId ?? ''}`);
+  }
+  return { ok: errors.length === 0, errors };
+}

package/engine/server.mjs

@@ -20,7 +20,7 @@ import { resolve, dirname, join, basename } from 'node:path';
 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 { 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, parseRunDeclaration, SHOT_DIR, collectShotNames, parseGoalAddress, REPLY_OUTCOMES, buildReplyIntentPrompt, parseReplyIntent, buildGoalMessage, serializeGoalMessage, parseGoalMessages, computeInternalQualityMetrics, classifyInternalQualityTaskError, resolveWorkspaceMode, goalSessionFor, setGoalAgentSession, switchGoalAgentSession, parseApprovalRequest } from './lib.mjs';
 import { createSerialQueue } from './lib.mjs';
 import { parseRequirementEvidence, unmappedChangedFiles } from './lib.mjs';
@@ -1524,6 +1524,27 @@ function removeGoalWorkDir(goal, project) {
   try { const repoRoot = gitSync(project.dir, ['rev-parse', '--show-toplevel']); gitSync(repoRoot, ['worktree', 'remove', '--force', wt]); } catch { /* best effort */ }
 }
 
+function projectForGoal(goal, project) {
+  return goal?.workspaceRoot ? { ...project, dir: goal.workspaceRoot } : project;
+}
+
+function rebindGoalWorkspace(goal, project, requestedDir) {
+  const abs = resolve(requestedDir);
+  if (!existsSync(abs)) return { ok: false, error: `Working directory is unavailable: ${abs}. Reconnect or choose the folder before starting work.` };
+  try { if (!statSync(abs).isDirectory()) return { ok: false, error: `Working directory is not a folder: ${abs}` }; } catch { return { ok: false, error: `Working directory cannot be read: ${abs}` }; }
+  if (!isGitRepo(abs)) return { ok: false, error: `Working directory is not a Git repository: ${abs}` };
+  removeGoalWorkDir(goal, project);
+  goal.workspaceRoot = abs;
+  // Claude Code's sandbox is fixed when a session starts. Never resume a
+  // session that was launched from the previous project directory.
+  goal.sessions = {};
+  goal.sessionId = null;
+  goal.sessionAgent = null;
+  goal.contextHandoff = null;
+  saveGoal(goal);
+  return { ok: true, dir: abs };
+}
+
 // External commit sync: pick up commits made outside this Manager (Claude
 // Code sessions run directly in the terminal, manual commits, other tools)
 // so they show up alongside Manager-run tasks for the same project. The
@@ -2697,8 +2718,9 @@ async function runTask(task) {
 
   // The workspace mode is chosen at goal creation and then held through every
   // worker/review/PR step.
-  const workDir = goal ? goalWorkDir(goal, project) : project.dir;
-  const wproject = { ...project, dir: workDir };
+  const effectiveProject = goal ? projectForGoal(goal, project) : project;
+  const workDir = goal ? goalWorkDir(goal, effectiveProject) : effectiveProject.dir;
+  const wproject = { ...effectiveProject, dir: workDir };
   const workspaceMode = resolveWorkspaceMode(task.workspaceMode ?? goal?.workspaceMode ?? goal?.worktree);
   sendProcessAct(task, workspaceMode === 'direct-workspace'
     ? `Using direct workspace: ${workDir}`
@@ -2717,7 +2739,11 @@ async function runTask(task) {
   const planMode = goal?.mode === 'plan' && goal?.planPhase !== 'executing';
   const permissionMode = permissionModeFor(planMode ? 'plan' : (task.mode ?? goal?.mode));
   const entryUrl = (task.passCondition && !planMode) ? entryUrlFor(goal, wproject) : null;
-  const maxAttempts = entryUrl ? 3 : 1;
+  // One continuation is reserved for a worker that reaches the infrastructure
+  // time cap. Long research/report tasks commonly need just over one window;
+  // resume their saved session automatically instead of asking the customer to
+  // discover and press Retry. Normal failures still stop after one attempt.
+  const maxAttempts = entryUrl ? 3 : 2;
   const t0 = Date.now();
 
   // Thread replies resume the goal's Claude Code session — same context,
@@ -2882,6 +2908,10 @@ async function runTask(task) {
     if (r.sessionId && goal) { setGoalAgentSession(goal, runAgent, r.sessionId); saveGoal(goal); }
     approval = parseApprovalRequest(result);
     if (approval) break;
+    if (timedOut && attempt < maxAttempts) {
+      sendProcessAct(task, 'Run limit reached — automatically continuing the same task from its saved session.');
+      continue;
+    }
     if (!entryUrl) break;
     try {
       if (!verify) {
@@ -5234,6 +5264,13 @@ const server = createServer(async (req, res) => {
         if (!text?.trim()) return json(res, 400, { error: 'text required' });
         const worker = applyReplyWorker(goal, requested);
         if (!worker.ok) return json(res, 409, { error: worker.error });
+        const requestedDir = parseRequestedWorkspaceDir(text);
+        if (requestedDir) {
+          const project = projects.find((p) => p.id === goal.projectId);
+          if (!project) return json(res, 404, { error: 'project not found' });
+          const rebound = rebindGoalWorkspace(goal, project, requestedDir);
+          if (!rebound.ok) return json(res, 409, rebound);
+        }
         // Reply picks a parked goal back up into the work queue. 'review' MUST be
         // here: a completed goal awaiting your approve/reject is 'review', and a
         // reply to it means "not done — keep working" (Masa 2026-07-24: replied to
@@ -5793,6 +5830,13 @@ server.on('error', (e) => {
 });
 
 server.listen(PORT, '127.0.0.1', () => {
+  // Persist the bounded recovery counter before pumping. If auth is offline
+  // and the machine restarts again while the task is still queued, the ledger
+  // must remember that one automatic recovery was already attempted.
+  for (const task of tasks.filter((t) => t.recoveredFromRestart)) {
+    delete task.recoveredFromRestart;
+    saveTask(task);
+  }
   PORT = server.address().port; // the OS-assigned one when MANAGER_PORT=0
   try { server6.listen(PORT, '::1'); } catch { /* IPv6 loopback unavailable */ }
   console.log(`[manager] Manager for AI → http://localhost:${PORT}/?key=${ACCESS_KEY}`);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@galda/cli",
-  "version": "0.10.90",
+  "version": "0.10.91",
   "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": {