diff --git a/app/src/components/intelligence/IntelligenceTasksTab.tsx b/app/src/components/intelligence/IntelligenceTasksTab.tsx
index 406cab331..9122fb626 100644
--- a/app/src/components/intelligence/IntelligenceTasksTab.tsx
+++ b/app/src/components/intelligence/IntelligenceTasksTab.tsx
@@ -214,6 +214,26 @@ export default function IntelligenceTasksTab() {
     };
   }, [loadAll]);
 
+  // Background board mutations — the dispatcher poller claiming a card,
+  // `update_task` from an autonomous run, `write_back`, triage, stale-reclaim —
+  // update the persisted board but emit no live socket event to this tab (the
+  // progress→socket bridge only fires inside an interactive, client-connected
+  // turn). So the user-tasks + task-sources boards would otherwise look frozen
+  // while work happens in the background. Re-read just those two on a light
+  // interval while the tab is visible; they're local in-process RPC, so it's
+  // cheap. (A push-based fix would need a core DomainEvent + socket broadcast;
+  // this keeps the fix isolated to the tab and catches every mutation source.)
+  useEffect(() => {
+    const POLL_MS = 4000;
+    const tick = () => {
+      if (typeof document !== 'undefined' && document.visibilityState !== 'visible') return;
+      void fetchPersonalBoard();
+      void fetchTaskSourcesBoard();
+    };
+    const handle = window.setInterval(tick, POLL_MS);
+    return () => window.clearInterval(handle);
+  }, [fetchPersonalBoard, fetchTaskSourcesBoard]);
+
   // A task created from the composer lands either on the personal board or
   // on a chosen conversation thread. `add` returns the updated board, so we
   // merge it directly — re-fetching listTurnStates would return a stale
@@ -397,7 +417,7 @@ export default function IntelligenceTasksTab() {
               status: 'todo',
               objective: draft.objective,
               notes: draft.notes,
-              assignedAgent: 'agent_coder',
+              assignedAgent: 'orchestrator',
               approvalMode: 'not_required',
               plan: draft.plan,
               allowedTools: draft.allowedTools,
diff --git a/app/src/components/intelligence/UserTaskComposer.test.tsx b/app/src/components/intelligence/UserTaskComposer.test.tsx
index 78d8b5088..5c0b129e5 100644
--- a/app/src/components/intelligence/UserTaskComposer.test.tsx
+++ b/app/src/components/intelligence/UserTaskComposer.test.tsx
@@ -18,10 +18,11 @@ vi.mock('../../store/hooks', () => ({
 
 vi.mock('../../services/api/todosApi', () => ({
   USER_TASKS_THREAD_ID: 'user-tasks',
-  todosApi: { add: vi.fn() },
+  todosApi: { add: vi.fn(), edit: vi.fn() },
 }));
 
 const mockAdd = vi.mocked(todosApi.add);
+const mockEdit = vi.mocked(todosApi.edit);
 
 function emptyBoard(threadId: string) {
   return { threadId, cards: [], updatedAt: '' };
@@ -83,6 +84,57 @@ describe('UserTaskComposer', () => {
     expect(mockAdd.mock.calls[0][0].threadId).toBe('t-1');
   });
 
+  it('assigns the new card to the orchestrator atomically when "assign to agent" is on', async () => {
+    mockAdd.mockResolvedValueOnce(emptyBoard(USER_TASKS_THREAD_ID));
+    render(<UserTaskComposer onCreated={vi.fn()} onClose={vi.fn()} />);
+
+    fireEvent.change(screen.getByPlaceholderText('What needs to be done?'), {
+      target: { value: 'Ship it' },
+    });
+    fireEvent.click(screen.getByRole('checkbox'));
+    fireEvent.click(screen.getByRole('button', { name: 'Create task' }));
+
+    // Single atomic add carries the assignment — no separate edit (no race).
+    await waitFor(() => expect(mockAdd).toHaveBeenCalledTimes(1));
+    expect(mockAdd).toHaveBeenCalledWith(
+      expect.objectContaining({
+        threadId: USER_TASKS_THREAD_ID,
+        content: 'Ship it',
+        assignedAgent: 'orchestrator',
+        approvalMode: 'not_required',
+      })
+    );
+    expect(mockEdit).not.toHaveBeenCalled();
+  });
+
+  it('does not assign an agent when the toggle is left off', async () => {
+    mockAdd.mockResolvedValueOnce(emptyBoard(USER_TASKS_THREAD_ID));
+    render(<UserTaskComposer onCreated={vi.fn()} onClose={vi.fn()} />);
+
+    fireEvent.change(screen.getByPlaceholderText('What needs to be done?'), {
+      target: { value: 'Buy milk' },
+    });
+    fireEvent.click(screen.getByRole('button', { name: 'Create task' }));
+
+    await waitFor(() => expect(mockAdd).toHaveBeenCalledTimes(1));
+    expect(mockEdit).not.toHaveBeenCalled();
+  });
+
+  it('disables assign-to-agent when the task is attached to a conversation', () => {
+    render(<UserTaskComposer onCreated={vi.fn()} onClose={vi.fn()} />);
+    fireEvent.change(screen.getByPlaceholderText('What needs to be done?'), {
+      target: { value: 'Book hotel' },
+    });
+    const checkbox = screen.getByRole('checkbox');
+    expect(checkbox).toBeEnabled();
+    // Attaching to a thread takes it off the personal board — the poller doesn't
+    // poll conversation threads, so auto-run is disabled there.
+    fireEvent.change(screen.getByDisplayValue('Personal (no conversation)'), {
+      target: { value: 't-1' },
+    });
+    expect(checkbox).toBeDisabled();
+  });
+
   it('surfaces an error and keeps the modal open on failure', async () => {
     mockAdd.mockRejectedValueOnce(new Error('boom'));
     const onClose = vi.fn();
diff --git a/app/src/components/intelligence/UserTaskComposer.tsx b/app/src/components/intelligence/UserTaskComposer.tsx
index 53fb25390..b8247205e 100644
--- a/app/src/components/intelligence/UserTaskComposer.tsx
+++ b/app/src/components/intelligence/UserTaskComposer.tsx
@@ -39,6 +39,11 @@ export function UserTaskComposer({ onCreated, onClose }: UserTaskComposerProps)
   const [objective, setObjective] = useState('');
   const [notes, setNotes] = useState('');
   const [attachThreadId, setAttachThreadId] = useState('');
+  // When on, the new personal-board card is assigned to the orchestrator so the
+  // task dispatcher's poller auto-picks and runs it. Off → a plain manual todo
+  // the poller never touches. Only meaningful on the personal board (the poller
+  // doesn't poll attached conversation threads), so it's disabled when attaching.
+  const [assignToAgent, setAssignToAgent] = useState(false);
   const [submitting, setSubmitting] = useState(false);
   const [error, setError] = useState<string | null>(null);
 
@@ -53,17 +58,25 @@ export function UserTaskComposer({ onCreated, onClose }: UserTaskComposerProps)
     const trimmedTitle = title.trim();
     if (!trimmedTitle || submitting) return;
     const threadId = attachThreadId || USER_TASKS_THREAD_ID;
+    // Auto-pick only works on the personal board (the poller doesn't poll
+    // attached conversation threads), so ignore the toggle when attaching.
+    const assign = assignToAgent && !attachThreadId;
     setSubmitting(true);
     setError(null);
-    log('submit threadId=%s status=%s', threadId, status);
+    log('submit threadId=%s status=%s assign=%s', threadId, status, assign);
     try {
+      // Assigning to the orchestrator + waiving the per-card approval gate so
+      // the dispatcher's poller picks it up and runs it — done atomically in
+      // the single `add` call (no create-then-edit race / partial failure).
       const board = await todosApi.add({
         threadId,
         content: trimmedTitle,
         status,
         objective: objective.trim() || null,
         notes: notes.trim() || null,
+        ...(assign ? { assignedAgent: 'orchestrator', approvalMode: 'not_required' } : {}),
       });
+
       onCreated(threadId, board);
       onClose();
     } catch (err) {
@@ -168,6 +181,24 @@ export function UserTaskComposer({ onCreated, onClose }: UserTaskComposerProps)
             />
           </label>
 
+          <label className="flex items-start gap-2">
+            <input
+              type="checkbox"
+              checked={assignToAgent && !attachThreadId}
+              disabled={attachThreadId !== ''}
+              onChange={e => setAssignToAgent(e.target.checked)}
+              className="mt-0.5 h-4 w-4 flex-none rounded border-stone-300 text-ocean-600 focus:ring-ocean-500 disabled:opacity-50 dark:border-neutral-600 dark:bg-neutral-950"
+            />
+            <span className="text-xs text-stone-600 dark:text-neutral-300">
+              <span className="font-semibold text-stone-700 dark:text-neutral-200">
+                {t('intelligence.tasks.composer.assignAgentLabel')}
+              </span>
+              <span className="mt-0.5 block text-stone-500 dark:text-neutral-400">
+                {t('intelligence.tasks.composer.assignAgentHint')}
+              </span>
+            </span>
+          </label>
+
           {error && (
             <p className="rounded-md border border-coral-200 bg-coral-50 px-3 py-2 text-xs text-coral-700 dark:border-coral-500/30 dark:bg-coral-500/10 dark:text-coral-300">
               {t('intelligence.tasks.composer.createFailed')}: {error}
diff --git a/app/src/components/intelligence/__tests__/IntelligenceTasksTab.test.tsx b/app/src/components/intelligence/__tests__/IntelligenceTasksTab.test.tsx
index cebae12ac..98c1191f3 100644
--- a/app/src/components/intelligence/__tests__/IntelligenceTasksTab.test.tsx
+++ b/app/src/components/intelligence/__tests__/IntelligenceTasksTab.test.tsx
@@ -336,7 +336,7 @@ describe('IntelligenceTasksTab', () => {
       expect.objectContaining({
         threadId: 'user-tasks',
         id: 'agent-task-1',
-        assignedAgent: 'agent_coder',
+        assignedAgent: 'orchestrator',
         approvalMode: 'not_required',
       })
     );
@@ -497,6 +497,35 @@ describe('IntelligenceTasksTab', () => {
     expect(hoisted.todosRemove).toHaveBeenCalledWith('user-tasks', 'card-0');
   });
 
+  test('re-polls the personal + task-source boards on an interval (background runs show live)', async () => {
+    vi.useFakeTimers();
+    try {
+      hoisted.todosList.mockImplementation((threadId: string) =>
+        Promise.resolve(makeBoard(threadId, []))
+      );
+      vi.resetModules();
+      const Tab = await importTab();
+      renderTab(Tab);
+
+      // Flush the mount effect's setTimeout(0) + the initial loadAll fetches.
+      await vi.advanceTimersByTimeAsync(0);
+      const userBefore = hoisted.todosList.mock.calls.filter(c => c[0] === 'user-tasks').length;
+      const sourceBefore = hoisted.todosList.mock.calls.filter(c => c[0] === 'task-sources').length;
+      expect(userBefore).toBeGreaterThan(0);
+      expect(sourceBefore).toBeGreaterThan(0);
+
+      // One poll interval later, both boards are re-read (so a background poller
+      // run's board changes surface without a manual refresh).
+      await vi.advanceTimersByTimeAsync(4000);
+      const userAfter = hoisted.todosList.mock.calls.filter(c => c[0] === 'user-tasks').length;
+      const sourceAfter = hoisted.todosList.mock.calls.filter(c => c[0] === 'task-sources').length;
+      expect(userAfter).toBeGreaterThan(userBefore);
+      expect(sourceAfter).toBeGreaterThan(sourceBefore);
+    } finally {
+      vi.useRealTimers();
+    }
+  });
+
   test('opens the composer and applies the created personal board', async () => {
     vi.resetModules();
     const Tab = await importTab();
diff --git a/app/src/lib/i18n/ar.ts b/app/src/lib/i18n/ar.ts
index ff6066929..f6d7515fc 100644
--- a/app/src/lib/i18n/ar.ts
+++ b/app/src/lib/i18n/ar.ts
@@ -2601,6 +2601,9 @@ const messages: TranslationMap = {
   'intelligence.tasks.composer.create': 'إنشاء مهمة',
   'intelligence.tasks.composer.creating': 'خلق...',
   'intelligence.tasks.composer.createFailed': 'لا يمكن خلق المهمة',
+  'intelligence.tasks.composer.assignAgentLabel': 'دع وكيلًا يعمل على هذه المهمة تلقائيًا',
+  'intelligence.tasks.composer.assignAgentHint':
+    'تلتقطها لوحة المهام وتنفّذها نيابةً عنك. اتركها مغلقة للحصول على مهمة شخصية عادية.',
   'intelligence.tasks.sourceList.subtitle': 'مهام المصادر التي تنتظر التحول إلى عمل للوكيل.',
   'intelligence.tasks.sourceList.empty': 'لا توجد مهام مصادر في الانتظار.',
   'intelligence.tasks.sourceList.queued': 'في قائمة الانتظار',
diff --git a/app/src/lib/i18n/bn.ts b/app/src/lib/i18n/bn.ts
index f730b2131..0a046ea99 100644
--- a/app/src/lib/i18n/bn.ts
+++ b/app/src/lib/i18n/bn.ts
@@ -2650,6 +2650,9 @@ const messages: TranslationMap = {
   'intelligence.tasks.composer.create': 'কাজ তৈরি করুন ( T)',
   'intelligence.tasks.composer.creating': 'তৈরি করা হচ্ছে...',
   'intelligence.tasks.composer.createFailed': 'কাজ তৈরি করতে ব্যর্থ',
+  'intelligence.tasks.composer.assignAgentLabel': 'একজন এজেন্টকে স্বয়ংক্রিয়ভাবে এটি করতে দিন',
+  'intelligence.tasks.composer.assignAgentHint':
+    'টাস্ক বোর্ড এটি তুলে নিয়ে আপনার হয়ে চালাবে। সাধারণ ব্যক্তিগত কাজের জন্য বন্ধ রাখুন।',
   'intelligence.tasks.sourceList.subtitle': 'এজেন্টের কাজে রূপান্তরের অপেক্ষায় থাকা উৎস কাজ।',
   'intelligence.tasks.sourceList.empty': 'অপেক্ষায় কোনো উৎস কাজ নেই।',
   'intelligence.tasks.sourceList.queued': 'সারিবদ্ধ',
diff --git a/app/src/lib/i18n/de.ts b/app/src/lib/i18n/de.ts
index e368cf807..47c825116 100644
--- a/app/src/lib/i18n/de.ts
+++ b/app/src/lib/i18n/de.ts
@@ -2716,6 +2716,9 @@ const messages: TranslationMap = {
   'intelligence.tasks.composer.create': 'Aufgabe erstellen',
   'intelligence.tasks.composer.creating': 'Erstellen…',
   'intelligence.tasks.composer.createFailed': 'Die Aufgabe konnte nicht erstellt werden',
+  'intelligence.tasks.composer.assignAgentLabel': 'Einen Agenten automatisch daran arbeiten lassen',
+  'intelligence.tasks.composer.assignAgentHint':
+    'Das Aufgabenboard übernimmt sie und erledigt sie für dich. Für eine einfache persönliche Aufgabe ausgeschaltet lassen.',
   'intelligence.tasks.sourceList.subtitle':
     'Quellaufgaben, die in Agentenarbeit umgewandelt werden sollen.',
   'intelligence.tasks.sourceList.empty': 'Keine Quellaufgaben in der Warteschlange.',
diff --git a/app/src/lib/i18n/en.ts b/app/src/lib/i18n/en.ts
index cf1525081..7a0ac1ad2 100644
--- a/app/src/lib/i18n/en.ts
+++ b/app/src/lib/i18n/en.ts
@@ -3036,6 +3036,9 @@ const en: TranslationMap = {
   'intelligence.tasks.composer.create': 'Create task',
   'intelligence.tasks.composer.creating': 'Creating…',
   'intelligence.tasks.composer.createFailed': "Couldn't create the task",
+  'intelligence.tasks.composer.assignAgentLabel': 'Let an agent work on this automatically',
+  'intelligence.tasks.composer.assignAgentHint':
+    'The task board picks it up and runs it for you. Leave off for a plain personal to-do.',
   'intelligence.tasks.sourceList.subtitle': 'Source tasks waiting to become agent work.',
   'intelligence.tasks.sourceList.empty': 'No source tasks waiting.',
   'intelligence.tasks.sourceList.queued': 'Queued',
diff --git a/app/src/lib/i18n/es.ts b/app/src/lib/i18n/es.ts
index 3af52424f..64ea9458a 100644
--- a/app/src/lib/i18n/es.ts
+++ b/app/src/lib/i18n/es.ts
@@ -2700,6 +2700,10 @@ const messages: TranslationMap = {
   'intelligence.tasks.composer.create': 'Crear tarea',
   'intelligence.tasks.composer.creating': 'Creando…',
   'intelligence.tasks.composer.createFailed': 'No se pudo crear la tarea',
+  'intelligence.tasks.composer.assignAgentLabel':
+    'Deja que un agente se encargue de esto automáticamente',
+  'intelligence.tasks.composer.assignAgentHint':
+    'El tablero de tareas la toma y la ejecuta por ti. Déjalo desactivado para una tarea personal simple.',
   'intelligence.tasks.sourceList.subtitle':
     'Tareas de fuentes esperando convertirse en trabajo del agente.',
   'intelligence.tasks.sourceList.empty': 'No hay tareas de fuentes en espera.',
diff --git a/app/src/lib/i18n/fr.ts b/app/src/lib/i18n/fr.ts
index 83f7de6ce..47710d080 100644
--- a/app/src/lib/i18n/fr.ts
+++ b/app/src/lib/i18n/fr.ts
@@ -2709,6 +2709,9 @@ const messages: TranslationMap = {
   'intelligence.tasks.composer.create': 'Créer une tâche',
   'intelligence.tasks.composer.creating': 'Création…',
   'intelligence.tasks.composer.createFailed': 'Impossible de créer la tâche',
+  'intelligence.tasks.composer.assignAgentLabel': "Laisser un agent s'en charger automatiquement",
+  'intelligence.tasks.composer.assignAgentHint':
+    "Le tableau des tâches la prend en charge et l'exécute pour vous. Laissez désactivé pour une simple tâche personnelle.",
   'intelligence.tasks.sourceList.subtitle':
     "Tâches de sources en attente de devenir du travail d'agent.",
   'intelligence.tasks.sourceList.empty': 'Aucune tâche de source en attente.',
diff --git a/app/src/lib/i18n/hi.ts b/app/src/lib/i18n/hi.ts
index 0f7877d05..3bd4a1963 100644
--- a/app/src/lib/i18n/hi.ts
+++ b/app/src/lib/i18n/hi.ts
@@ -2655,6 +2655,9 @@ const messages: TranslationMap = {
   'intelligence.tasks.composer.create': 'कार्य',
   'intelligence.tasks.composer.creating': 'बनाना',
   'intelligence.tasks.composer.createFailed': 'काम नहीं कर सका',
+  'intelligence.tasks.composer.assignAgentLabel': 'किसी एजेंट को इसे स्वचालित रूप से करने दें',
+  'intelligence.tasks.composer.assignAgentHint':
+    'टास्क बोर्ड इसे उठाकर आपके लिए चला देगा। सामान्य निजी कार्य के लिए इसे बंद रखें।',
   'intelligence.tasks.sourceList.subtitle': 'एजेंट कार्य बनने की प्रतीक्षा कर रहे स्रोत कार्य।',
   'intelligence.tasks.sourceList.empty': 'कोई स्रोत कार्य प्रतीक्षा में नहीं है।',
   'intelligence.tasks.sourceList.queued': 'कतार में',
diff --git a/app/src/lib/i18n/id.ts b/app/src/lib/i18n/id.ts
index 33dde55c1..b35e8ffe9 100644
--- a/app/src/lib/i18n/id.ts
+++ b/app/src/lib/i18n/id.ts
@@ -2657,6 +2657,9 @@ const messages: TranslationMap = {
   'intelligence.tasks.composer.create': 'Buat tugas',
   'intelligence.tasks.composer.creating': 'Membuat…',
   'intelligence.tasks.composer.createFailed': 'Gagal membuat tugas',
+  'intelligence.tasks.composer.assignAgentLabel': 'Biarkan agen mengerjakan ini secara otomatis',
+  'intelligence.tasks.composer.assignAgentHint':
+    'Papan tugas akan mengambil dan menjalankannya untuk Anda. Biarkan nonaktif untuk tugas pribadi biasa.',
   'intelligence.tasks.sourceList.subtitle': 'Tugas sumber yang menunggu menjadi pekerjaan agen.',
   'intelligence.tasks.sourceList.empty': 'Tidak ada tugas sumber yang menunggu.',
   'intelligence.tasks.sourceList.queued': 'Dalam antrean',
diff --git a/app/src/lib/i18n/it.ts b/app/src/lib/i18n/it.ts
index a3cac5a9b..1806d658d 100644
--- a/app/src/lib/i18n/it.ts
+++ b/app/src/lib/i18n/it.ts
@@ -2690,6 +2690,10 @@ const messages: TranslationMap = {
   'intelligence.tasks.composer.create': 'Crea attività',
   'intelligence.tasks.composer.creating': 'Creando…',
   'intelligence.tasks.composer.createFailed': 'Impossibile creare il task',
+  'intelligence.tasks.composer.assignAgentLabel':
+    'Lascia che un agente se ne occupi automaticamente',
+  'intelligence.tasks.composer.assignAgentHint':
+    'La bacheca delle attività la prende in carico e la esegue per te. Lascia disattivato per una semplice attività personale.',
   'intelligence.tasks.sourceList.subtitle':
     "Attività dalle fonti in attesa di diventare lavoro dell'agente.",
   'intelligence.tasks.sourceList.empty': 'Nessuna attività da fonti in attesa.',
diff --git a/app/src/lib/i18n/ko.ts b/app/src/lib/i18n/ko.ts
index b6e65e04b..bc3feedd9 100644
--- a/app/src/lib/i18n/ko.ts
+++ b/app/src/lib/i18n/ko.ts
@@ -2630,6 +2630,9 @@ const messages: TranslationMap = {
   'intelligence.tasks.composer.create': '작업 만들기',
   'intelligence.tasks.composer.creating': '만드는 중…',
   'intelligence.tasks.composer.createFailed': '작업을 만들 수 없습니다',
+  'intelligence.tasks.composer.assignAgentLabel': '에이전트가 이 작업을 자동으로 처리하도록 하기',
+  'intelligence.tasks.composer.assignAgentHint':
+    '작업 보드가 이를 가져와 대신 실행합니다. 일반 개인 할 일은 꺼 두세요.',
   'intelligence.tasks.sourceList.subtitle': '에이전트 작업으로 전환될 소스 작업입니다.',
   'intelligence.tasks.sourceList.empty': '대기 중인 소스 작업이 없습니다.',
   'intelligence.tasks.sourceList.queued': '대기열에 추가됨',
diff --git a/app/src/lib/i18n/pl.ts b/app/src/lib/i18n/pl.ts
index 6d7e61b37..a21e43a65 100644
--- a/app/src/lib/i18n/pl.ts
+++ b/app/src/lib/i18n/pl.ts
@@ -2687,6 +2687,9 @@ const messages: TranslationMap = {
   'intelligence.tasks.composer.create': 'Utwórz zadanie',
   'intelligence.tasks.composer.creating': 'Tworzenie…',
   'intelligence.tasks.composer.createFailed': 'Nie udało się utworzyć zadania',
+  'intelligence.tasks.composer.assignAgentLabel': 'Pozwól agentowi zająć się tym automatycznie',
+  'intelligence.tasks.composer.assignAgentHint':
+    'Tablica zadań przejmie je i wykona za Ciebie. Wyłącz dla zwykłego zadania osobistego.',
   'intelligence.tasks.sourceList.subtitle':
     'Zadania ze źródeł czekające na przekształcenie w pracę agenta.',
   'intelligence.tasks.sourceList.empty': 'Brak oczekujących zadań ze źródeł.',
diff --git a/app/src/lib/i18n/pt.ts b/app/src/lib/i18n/pt.ts
index 84b4ab252..d53f6c584 100644
--- a/app/src/lib/i18n/pt.ts
+++ b/app/src/lib/i18n/pt.ts
@@ -2698,6 +2698,9 @@ const messages: TranslationMap = {
   'intelligence.tasks.composer.create': 'Criar tarefa',
   'intelligence.tasks.composer.creating': 'Criando…',
   'intelligence.tasks.composer.createFailed': 'Não foi possível criar a tarefa',
+  'intelligence.tasks.composer.assignAgentLabel': 'Deixe um agente cuidar disto automaticamente',
+  'intelligence.tasks.composer.assignAgentHint':
+    'O quadro de tarefas o executa por você. Deixe desativado para uma tarefa pessoal simples.',
   'intelligence.tasks.sourceList.subtitle':
     'Tarefas de fontes aguardando virar trabalho de agente.',
   'intelligence.tasks.sourceList.empty': 'Nenhuma tarefa de fonte aguardando.',
diff --git a/app/src/lib/i18n/ru.ts b/app/src/lib/i18n/ru.ts
index e5de78ece..0a3c44d35 100644
--- a/app/src/lib/i18n/ru.ts
+++ b/app/src/lib/i18n/ru.ts
@@ -2670,6 +2670,9 @@ const messages: TranslationMap = {
   'intelligence.tasks.composer.create': 'Создать задачу',
   'intelligence.tasks.composer.creating': 'Создание…',
   'intelligence.tasks.composer.createFailed': 'Не удалось создать задачу',
+  'intelligence.tasks.composer.assignAgentLabel': 'Поручить агенту выполнить это автоматически',
+  'intelligence.tasks.composer.assignAgentHint':
+    'Доска задач возьмёт её и выполнит за вас. Оставьте выключенным для обычной личной задачи.',
   'intelligence.tasks.sourceList.subtitle':
     'Задачи из источников, ожидающие превращения в работу агента.',
   'intelligence.tasks.sourceList.empty': 'Нет ожидающих задач из источников.',
diff --git a/app/src/lib/i18n/zh-CN.ts b/app/src/lib/i18n/zh-CN.ts
index c6d68f59d..b914e8b58 100644
--- a/app/src/lib/i18n/zh-CN.ts
+++ b/app/src/lib/i18n/zh-CN.ts
@@ -2523,6 +2523,9 @@ const messages: TranslationMap = {
   'intelligence.tasks.composer.create': '创建任务',
   'intelligence.tasks.composer.creating': '正在创建…',
   'intelligence.tasks.composer.createFailed': '无法创建任务',
+  'intelligence.tasks.composer.assignAgentLabel': '让智能体自动处理此任务',
+  'intelligence.tasks.composer.assignAgentHint':
+    '任务看板会接手并替你执行。普通的个人待办请保持关闭。',
   'intelligence.tasks.sourceList.subtitle': '等待转为智能体工作的来源任务。',
   'intelligence.tasks.sourceList.empty': '没有等待处理的来源任务。',
   'intelligence.tasks.sourceList.queued': '已入队',
diff --git a/app/src/services/api/todosApi.ts b/app/src/services/api/todosApi.ts
index d396086a1..af981bf36 100644
--- a/app/src/services/api/todosApi.ts
+++ b/app/src/services/api/todosApi.ts
@@ -49,6 +49,8 @@ export interface AddTodoInput {
   status?: TaskBoardCardStatus;
   objective?: string | null;
   notes?: string | null;
+  assignedAgent?: string | null;
+  approvalMode?: TaskApprovalMode | null;
 }
 
 /** Fields accepted when editing a card. Omitted fields are left unchanged. */
@@ -120,6 +122,8 @@ export const todosApi = {
         status: input.status,
         objective: input.objective,
         notes: input.notes,
+        assignedAgent: input.assignedAgent,
+        approvalMode: input.approvalMode,
       }),
     });
     return snapshotToBoard(snap, input.threadId);
diff --git a/src/openhuman/agent/task_dispatcher.rs b/src/openhuman/agent/task_dispatcher.rs
index e1ee7e492..8df1ea2ec 100644
--- a/src/openhuman/agent/task_dispatcher.rs
+++ b/src/openhuman/agent/task_dispatcher.rs
@@ -26,9 +26,9 @@ use crate::openhuman::agent::harness::definition::{AgentDefinitionRegistry, Prom
 use crate::openhuman::agent::harness::session::Agent;
 use crate::openhuman::agent::harness::subagent_runner::with_autonomous_iter_cap;
 use crate::openhuman::agent::personality_paths::PersonalityContext;
-use crate::openhuman::agent::task_board::{TaskBoardCard, TaskCardStatus};
+use crate::openhuman::agent::task_board::{TaskApprovalMode, TaskBoardCard, TaskCardStatus};
 use crate::openhuman::config::Config;
-use crate::openhuman::todos::ops::{self, BoardLocation, CardPatch};
+use crate::openhuman::todos::ops::{self, BoardLocation, CardPatch, USER_TASKS_THREAD_ID};
 use crate::openhuman::todos::runs::{self, RunLimits, RunOutcome};
 
 /// Max chars of a personality SOUL.md / MEMORY.md or skill guideline block
@@ -129,6 +129,33 @@ pub fn build_task_prompt(card: &TaskBoardCard) -> String {
     lines.join("\n")
 }
 
+/// Instruction appended to the run prompt so the autonomous turn keeps its own
+/// task card current via the `update_task` tool while it works.
+///
+/// The card is already `in_progress` (the dispatcher claimed it before
+/// spawning the run), addressed by the exact card id + board the run owns
+/// (without the explicit `threadId` the tool defaults to the `task-sources`
+/// board and would miss a `user-tasks` card). Two things this asks for:
+/// 1. *progress* updates (notes/evidence) as the run works, and
+/// 2. an explicit `status: blocked` + `blocker` when the run needs a
+///    decision/information from the user or cannot proceed — which
+///    [`write_back`] now preserves rather than force-completing, so the task
+///    pauses for the user instead of being silently marked done.
+fn build_progress_instruction(card_id: &str, thread_id: &str) -> String {
+    format!(
+        "\n\nThis task is tracked as card `{card_id}` on the `{thread_id}` board. As you work, \
+         call the `update_task` tool (id `{card_id}`, threadId `{thread_id}`) to keep the card \
+         current — append `notes`/`evidence` as you make progress.\n\nIf you need a decision or \
+         information from the user, or you genuinely cannot proceed (missing access, ambiguous \
+         requirement, an action that needs the user's confirmation), call `update_task` with \
+         `status: blocked` and a `blocker` that states exactly what you need from the user. The \
+         task will stay paused in that blocked state until the user responds — do NOT guess, \
+         fabricate, or take a risky irreversible action just to avoid blocking. If instead you \
+         finish the work, end with a summary of what you did and the evidence; completion is \
+         recorded automatically."
+    )
+}
+
 /// Outcome of a dispatch attempt.
 #[derive(Debug)]
 pub enum DispatchOutcome {
@@ -161,7 +188,16 @@ pub async fn dispatch_card(
     // approval before it can run. `Ready` (already approved) bypasses. We
     // attempt the AwaitingApproval claim first so the gate is also atomic —
     // two dispatchers racing the same Todo card won't both park it.
-    if config.autonomy.require_task_plan_approval {
+    //
+    // A card explicitly marked `approval_mode = NotRequired` also bypasses the
+    // gate: it has already cleared human review (e.g. a task approved out of
+    // the `task-sources` inbox onto the `user-tasks` board, stamped
+    // `not_required` at approval time). Re-parking it under the global default
+    // would strand it on a board nobody approves from. Per-card opt-out wins.
+    if requires_plan_approval(
+        config.autonomy.require_task_plan_approval,
+        card.approval_mode.as_ref(),
+    ) {
         match ops::claim_card(
             &location,
             &card_id,
@@ -199,7 +235,14 @@ pub async fn dispatch_card(
     )
     .map_err(|e| format!("[task_dispatcher] claim rejected for {card_id}: {e}"))?;
 
-    let prompt = build_task_prompt(&fresh_card);
+    let mut prompt = build_task_prompt(&fresh_card);
+    // Tell the run which card it owns so it can post live progress via the
+    // `update_task` tool (notes/evidence) as it works. The terminal
+    // `done`/`blocked` transition is still stamped deterministically by
+    // `write_back` from the run outcome.
+    if let Some(thread_id) = location.thread_id() {
+        prompt.push_str(&build_progress_instruction(&card_id, thread_id));
+    }
 
     let run_id = uuid::Uuid::new_v4().to_string();
 
@@ -400,48 +443,77 @@ async fn run_autonomous(
 /// Success → `done` + evidence; failure → `blocked` + blocker reason. An
 /// external write failure here is logged, never propagated — the run already
 /// happened.
+/// Current persisted status of a card, or `None` if the board can't be read or
+/// the card is gone. Used by `write_back` to detect a run that blocked itself.
+fn current_card_status(location: &BoardLocation, card_id: &str) -> Option<TaskCardStatus> {
+    ops::list(location)
+        .ok()
+        .and_then(|snap| snap.cards.into_iter().find(|c| c.id == card_id))
+        .map(|c| c.status)
+}
+
 fn write_back(
     location: &BoardLocation,
     card_id: &str,
     run_id: &str,
     outcome: Result<String, String>,
 ) {
-    let patch = match &outcome {
-        Ok(output) => {
-            tracing::info!(
-                card_id = %card_id,
-                run_id = %run_id,
-                output_chars = output.chars().count(),
-                "[task_dispatcher] run complete → done"
-            );
-            CardPatch {
-                status: Some(TaskCardStatus::Done),
-                evidence: Some(vec![truncate_chars(output.trim(), EVIDENCE_MAX_CHARS)]),
-                ..Default::default()
+    // Respect a status the run set for itself: if the agent marked the card
+    // `blocked` via `update_task` (it needs a decision/input from the user, or
+    // genuinely cannot proceed), leave it blocked — do NOT force-complete it.
+    // The task then stays paused in that state until the user responds, instead
+    // of a "clean turn" being silently recorded as done. Otherwise mark done
+    // with evidence; a run error marks blocked with the error as the blocker.
+    let agent_self_blocked =
+        outcome.is_ok() && current_card_status(location, card_id) == Some(TaskCardStatus::Blocked);
+
+    let patch = if agent_self_blocked {
+        tracing::info!(
+            card_id = %card_id,
+            run_id = %run_id,
+            "[task_dispatcher] run ended with card self-blocked → leaving blocked (awaiting user input), not auto-completing"
+        );
+        None
+    } else {
+        match &outcome {
+            Ok(output) => {
+                tracing::info!(
+                    card_id = %card_id,
+                    run_id = %run_id,
+                    output_chars = output.chars().count(),
+                    "[task_dispatcher] run complete → done"
+                );
+                Some(CardPatch {
+                    status: Some(TaskCardStatus::Done),
+                    evidence: Some(vec![truncate_chars(output.trim(), EVIDENCE_MAX_CHARS)]),
+                    ..Default::default()
+                })
             }
-        }
-        Err(err) => {
-            tracing::warn!(
-                card_id = %card_id,
-                run_id = %run_id,
-                error = %err,
-                "[task_dispatcher] run failed → blocked"
-            );
-            CardPatch {
-                status: Some(TaskCardStatus::Blocked),
-                blocker: Some(truncate_chars(err, EVIDENCE_MAX_CHARS)),
-                ..Default::default()
+            Err(err) => {
+                tracing::warn!(
+                    card_id = %card_id,
+                    run_id = %run_id,
+                    error = %err,
+                    "[task_dispatcher] run failed → blocked"
+                );
+                Some(CardPatch {
+                    status: Some(TaskCardStatus::Blocked),
+                    blocker: Some(truncate_chars(err, EVIDENCE_MAX_CHARS)),
+                    ..Default::default()
+                })
             }
         }
     };
 
-    if let Err(e) = ops::edit(location, card_id, patch) {
-        tracing::error!(
-            card_id = %card_id,
-            run_id = %run_id,
-            error = %e,
-            "[task_dispatcher] board write-back failed (run outcome lost from board)"
-        );
+    if let Some(patch) = patch {
+        if let Err(e) = ops::edit(location, card_id, patch) {
+            tracing::error!(
+                card_id = %card_id,
+                run_id = %run_id,
+                error = %e,
+                "[task_dispatcher] board write-back failed (run outcome lost from board)"
+            );
+        }
     }
 
     let (run_outcome, run_error, run_evidence) = match &outcome {
@@ -510,9 +582,19 @@ pub fn start_board_poller() {
     });
 }
 
-/// One poller tick: dispatch the highest-urgency `todo` card on the
-/// task-sources board, if any and if capacity allows. `pub(crate)` so tests can
+/// One poller tick: sweep each executor board and dispatch its highest-urgency
+/// dispatchable card, if any and if capacity allows. `pub(crate)` so tests can
 /// drive a tick without the real interval.
+///
+/// Two boards are swept, each independently (own stale-reclaim + single
+/// `in_progress` cap):
+/// - **`user-tasks`** (the kanban work board) — always swept, but only
+///   **agent-assigned** cards are run, so a human's manually-created todo is
+///   never auto-executed. This is where tasks approved out of the inbox run.
+/// - **`task-sources`** (the proactive inbox) — swept only when ingestion is
+///   enabled. With plan-approval required this only ever parks a `todo` at
+///   `awaiting_approval`; it runs a card directly only when approval is off.
+///   Kept in the sweep so its stale/wedged runs are still reclaimed.
 pub(crate) async fn poll_once() -> Result<(), String> {
     // Gate on background-AI capacity (autonomy / power / pause). Dropping the
     // permit immediately is fine: this is a "may background work start now"
@@ -525,21 +607,50 @@ pub(crate) async fn poll_once() -> Result<(), String> {
     let config = Config::load_or_init()
         .await
         .map_err(|e| format!("load config: {e:#}"))?;
-    if !config.task_sources.enabled {
-        return Ok(());
+
+    // (board location, agent_assigned_only). user-tasks first — it's the real
+    // work board; task-sources is only included for parking + reclaim.
+    let mut boards: Vec<(BoardLocation, bool)> = vec![(
+        BoardLocation::Thread {
+            workspace_dir: config.workspace_dir.clone(),
+            thread_id: USER_TASKS_THREAD_ID.to_string(),
+        },
+        true,
+    )];
+    if config.task_sources.enabled {
+        boards.push((
+            BoardLocation::Thread {
+                workspace_dir: config.workspace_dir.clone(),
+                thread_id: crate::openhuman::task_sources::TASK_SOURCES_THREAD_ID.to_string(),
+            },
+            false,
+        ));
     }
 
-    let location = BoardLocation::Thread {
-        workspace_dir: config.workspace_dir.clone(),
-        thread_id: crate::openhuman::task_sources::TASK_SOURCES_THREAD_ID.to_string(),
-    };
+    for (location, agent_assigned_only) in boards {
+        if let Err(e) = poll_board(&location, agent_assigned_only).await {
+            tracing::warn!(
+                thread_id = ?location.thread_id(),
+                error = %e,
+                "[task_dispatcher:poller] board sweep failed (continuing)"
+            );
+        }
+    }
+    Ok(())
+}
 
+/// Sweep one board: reclaim stale runs, then (unless one is already running)
+/// dispatch its highest-urgency dispatchable card. When `agent_assigned_only`
+/// is set, only cards with an `assigned_agent` are eligible — the guard that
+/// keeps the poller off a human's manual `user-tasks` cards.
+async fn poll_board(location: &BoardLocation, agent_assigned_only: bool) -> Result<(), String> {
     // Reclaim stale/wedged runs before looking for new work. Reclaimed
     // cards move back to `todo` (re-dispatchable) so they appear in the
     // snapshot below and can be picked up in the same tick.
-    match runs::reclaim_stale(&location, &RunLimits::default()) {
+    match runs::reclaim_stale(location, &RunLimits::default()) {
         Ok(result) if result.reclaimed_count > 0 || result.blocked_count > 0 => {
             tracing::info!(
+                thread_id = ?location.thread_id(),
                 reclaimed = result.reclaimed_count,
                 blocked = result.blocked_count,
                 "[task_dispatcher:poller] stale runs reclaimed"
@@ -547,6 +658,7 @@ pub(crate) async fn poll_once() -> Result<(), String> {
         }
         Err(e) => {
             tracing::warn!(
+                thread_id = ?location.thread_id(),
                 error = %e,
                 "[task_dispatcher:poller] stale reclaim failed (continuing)"
             );
@@ -554,7 +666,7 @@ pub(crate) async fn poll_once() -> Result<(), String> {
         _ => {}
     }
 
-    let snapshot = ops::list(&location)?;
+    let snapshot = ops::list(location)?;
 
     // `enforce_single_in_progress` caps the board at one running card, so if
     // one is already in progress there's nothing for this tick to claim.
@@ -566,26 +678,39 @@ pub(crate) async fn poll_once() -> Result<(), String> {
         return Ok(());
     }
 
-    let Some(card) = pick_next_todo(&snapshot.cards) else {
+    let Some(card) = pick_next_todo(&snapshot.cards, agent_assigned_only) else {
         return Ok(());
     };
 
     tracing::info!(
         card_id = %card.id,
+        thread_id = ?location.thread_id(),
         urgency = card_urgency(&card),
-        "[task_dispatcher:poller] dispatching highest-urgency todo card"
+        agent_assigned_only,
+        "[task_dispatcher:poller] dispatching highest-urgency dispatchable card"
     );
-    dispatch_card(location, card).await.map(|_| ())
+    dispatch_card(location.clone(), card).await.map(|_| ())
 }
 
 /// Highest-urgency dispatchable card (`todo` or approved `ready`; urgency from
 /// `source_metadata.urgency`, default 0.0; ties broken toward the lower board
 /// `order`). Returns a clone. `dispatch_card` then either runs a `ready` card
 /// or parks a `todo` one for approval, per the autonomy setting.
-fn pick_next_todo(cards: &[TaskBoardCard]) -> Option<TaskBoardCard> {
+///
+/// When `agent_assigned_only` is set, cards without an `assigned_agent` are
+/// excluded — used on the `user-tasks` board so the poller runs only
+/// agent-generated tasks and never picks up a human's manually-created card.
+fn pick_next_todo(cards: &[TaskBoardCard], agent_assigned_only: bool) -> Option<TaskBoardCard> {
     cards
         .iter()
         .filter(|c| matches!(c.status, TaskCardStatus::Todo | TaskCardStatus::Ready))
+        .filter(|c| {
+            !agent_assigned_only
+                || c.assigned_agent
+                    .as_deref()
+                    .map(|a| !a.trim().is_empty())
+                    .unwrap_or(false)
+        })
         .max_by(|a, b| {
             card_urgency(a)
                 .partial_cmp(&card_urgency(b))
@@ -597,6 +722,18 @@ fn pick_next_todo(cards: &[TaskBoardCard]) -> Option<TaskBoardCard> {
         .cloned()
 }
 
+/// Whether a card must be parked at `awaiting_approval` before it can run.
+///
+/// The global `require_task_plan_approval` setting applies *unless* the card is
+/// explicitly marked `approval_mode = NotRequired` — a per-card opt-out for
+/// tasks that have already cleared human review (e.g. approved out of the
+/// `task-sources` inbox onto `user-tasks`). Per-card opt-out wins over the
+/// global default; without this, an already-approved card would be re-parked
+/// and stranded.
+fn requires_plan_approval(global_required: bool, approval_mode: Option<&TaskApprovalMode>) -> bool {
+    global_required && approval_mode != Some(&TaskApprovalMode::NotRequired)
+}
+
 fn card_urgency(card: &TaskBoardCard) -> f64 {
     card.source_metadata
         .as_ref()
@@ -728,7 +865,7 @@ mod tests {
             card_with("c", TaskCardStatus::Todo, Some(0.8), 2),
             card_with("d", TaskCardStatus::Todo, None, 3),
         ];
-        let picked = pick_next_todo(&cards).expect("a todo card is available");
+        let picked = pick_next_todo(&cards, false).expect("a todo card is available");
         assert_eq!(
             picked.id, "c",
             "highest-urgency todo wins, done card ignored"
@@ -741,13 +878,13 @@ mod tests {
             card_with("late", TaskCardStatus::Todo, Some(0.5), 5),
             card_with("early", TaskCardStatus::Todo, Some(0.5), 2),
         ];
-        assert_eq!(pick_next_todo(&cards).unwrap().id, "early");
+        assert_eq!(pick_next_todo(&cards, false).unwrap().id, "early");
     }
 
     #[test]
     fn poller_returns_none_when_no_todo_cards() {
         let cards = vec![card_with("a", TaskCardStatus::Done, Some(0.9), 0)];
-        assert!(pick_next_todo(&cards).is_none());
+        assert!(pick_next_todo(&cards, false).is_none());
     }
 
     #[test]
@@ -759,7 +896,7 @@ mod tests {
             card_with("rej", TaskCardStatus::Rejected, Some(0.95), 1),
             card_with("ready", TaskCardStatus::Ready, Some(0.5), 2),
         ];
-        assert_eq!(pick_next_todo(&cards).unwrap().id, "ready");
+        assert_eq!(pick_next_todo(&cards, false).unwrap().id, "ready");
     }
 
     #[test]
@@ -768,7 +905,65 @@ mod tests {
             card_with("ready-low", TaskCardStatus::Ready, Some(0.3), 0),
             card_with("todo-high", TaskCardStatus::Todo, Some(0.9), 1),
         ];
-        assert_eq!(pick_next_todo(&cards).unwrap().id, "todo-high");
+        assert_eq!(pick_next_todo(&cards, false).unwrap().id, "todo-high");
+    }
+
+    #[test]
+    fn poller_agent_only_skips_unassigned_cards() {
+        // On the user-tasks board we run only agent-assigned cards. A human's
+        // manual todo (no assigned_agent) must be skipped even at high urgency.
+        let mut human = card_with("human", TaskCardStatus::Todo, Some(0.99), 0);
+        human.assigned_agent = None;
+        let mut agent = card_with("agent", TaskCardStatus::Todo, Some(0.20), 1);
+        agent.assigned_agent = Some("orchestrator".into());
+        let cards = vec![human, agent];
+
+        // Agent-only: the lower-urgency assigned card wins; the human card is invisible.
+        assert_eq!(pick_next_todo(&cards, true).unwrap().id, "agent");
+        // Unfiltered (task-sources behaviour): highest urgency wins regardless.
+        assert_eq!(pick_next_todo(&cards, false).unwrap().id, "human");
+    }
+
+    #[test]
+    fn poller_agent_only_returns_none_when_all_unassigned() {
+        let mut a = card_with("a", TaskCardStatus::Todo, Some(0.9), 0);
+        a.assigned_agent = None;
+        let mut b = card_with("b", TaskCardStatus::Todo, Some(0.5), 1);
+        b.assigned_agent = Some("   ".into()); // blank handle is not "assigned"
+        let cards = vec![a, b];
+        assert!(pick_next_todo(&cards, true).is_none());
+    }
+
+    #[test]
+    fn approval_gate_respects_global_and_per_card_optout() {
+        // Global off → never park.
+        assert!(!requires_plan_approval(false, None));
+        assert!(!requires_plan_approval(
+            false,
+            Some(&TaskApprovalMode::Required)
+        ));
+        // Global on → park, unless the card opts out via NotRequired.
+        assert!(requires_plan_approval(true, None));
+        assert!(requires_plan_approval(
+            true,
+            Some(&TaskApprovalMode::Required)
+        ));
+        assert!(!requires_plan_approval(
+            true,
+            Some(&TaskApprovalMode::NotRequired)
+        ));
+    }
+
+    #[test]
+    fn progress_instruction_names_card_thread_and_tool() {
+        let s = build_progress_instruction("task-42", "user-tasks");
+        assert!(s.contains("task-42"));
+        assert!(s.contains("user-tasks"));
+        assert!(s.contains("update_task"));
+        // It must instruct the agent to self-block (status: blocked + blocker)
+        // when it needs the user, so write_back can preserve that state.
+        assert!(s.contains("status: blocked"));
+        assert!(s.contains("blocker"));
     }
 
     #[test]
@@ -839,6 +1034,58 @@ mod tests {
         assert!(card.evidence.iter().any(|e| e.contains("opened PR #5")));
     }
 
+    #[test]
+    fn write_back_preserves_agent_set_blocked_on_clean_run() {
+        // The run marked its own card `blocked` (needs user input) via
+        // update_task, then ended cleanly. write_back must NOT force it to
+        // `done` — the task stays blocked, with the agent's blocker intact,
+        // awaiting the user.
+        let dir = tempfile::tempdir().unwrap();
+        let loc = board_loc(dir.path());
+        let id = ops::add(&loc, "update alan", CardPatch::default())
+            .unwrap()
+            .cards[0]
+            .id
+            .clone();
+        ops::update_status(&loc, &id, TaskCardStatus::InProgress).unwrap();
+        // Agent self-blocks mid-run, as build_progress_instruction asks it to.
+        ops::edit(
+            &loc,
+            &id,
+            CardPatch {
+                status: Some(TaskCardStatus::Blocked),
+                blocker: Some("Slack isn't connected — confirm how to reach Alan".to_string()),
+                ..Default::default()
+            },
+        )
+        .unwrap();
+
+        // Run returns Ok (the turn finished) — but the card is self-blocked.
+        write_back(
+            &loc,
+            &id,
+            "run-2",
+            Ok("I checked GitHub and memory…".to_string()),
+        );
+
+        let card = ops::list(&loc)
+            .unwrap()
+            .cards
+            .into_iter()
+            .find(|c| c.id == id)
+            .unwrap();
+        assert_eq!(
+            card.status,
+            TaskCardStatus::Blocked,
+            "a clean run over a self-blocked card must stay blocked, not auto-done"
+        );
+        assert_eq!(
+            card.blocker.as_deref(),
+            Some("Slack isn't connected — confirm how to reach Alan"),
+            "the agent's blocker reason is preserved"
+        );
+    }
+
     #[test]
     fn write_back_marks_blocked_with_reason_on_failure() {
         let dir = tempfile::tempdir().unwrap();
diff --git a/src/openhuman/agent/tools.rs b/src/openhuman/agent/tools.rs
index 8dd117f64..c73d52a38 100644
--- a/src/openhuman/agent/tools.rs
+++ b/src/openhuman/agent/tools.rs
@@ -6,6 +6,7 @@ pub mod remember_preference;
 mod run_skill;
 pub mod save_preference;
 mod todo;
+mod update_task;
 pub mod workflow_tools;
 
 pub use ask_clarification::AskClarificationTool;
@@ -16,4 +17,5 @@ pub use remember_preference::RememberPreferenceTool;
 pub use run_skill::{RunSkillTool, RUN_SKILL_TOOL_NAME};
 pub use save_preference::SavePreferenceTool;
 pub use todo::TodoTool;
+pub use update_task::UpdateTaskTool;
 pub use workflow_tools::{WorkflowLoadTool, WorkflowPhaseTool};
diff --git a/src/openhuman/agent/tools/update_task.rs b/src/openhuman/agent/tools/update_task.rs
new file mode 100644
index 000000000..be222c832
--- /dev/null
+++ b/src/openhuman/agent/tools/update_task.rs
@@ -0,0 +1,226 @@
+//! `update_task` — move or update a specific task card on a thread's board.
+//!
+//! `todo` only reaches the *current* thread's board. `update_task` addresses a
+//! card **by id** on a *target* board — defaulting to the proactive
+//! `task-sources` board — so the orchestrator (or an autonomous task run on its
+//! own thread) can advance the task it's working: move it to
+//! `in_progress`/`blocked`/`done`, or update its objective/notes/evidence/blocker.
+//!
+//! It is a thin wrapper over [`crate::openhuman::todos::ops::edit`], which
+//! applies the status move + field updates atomically, enforces the
+//! single-`in_progress` invariant, and emits the board-progress event that the
+//! Tasks board UI listens on.
+
+use crate::openhuman::task_sources::TASK_SOURCES_THREAD_ID;
+use crate::openhuman::todos::ops::{self, BoardLocation, CardPatch};
+use crate::openhuman::tools::traits::{PermissionLevel, Tool, ToolResult};
+use async_trait::async_trait;
+use serde_json::json;
+use std::path::PathBuf;
+
+pub struct UpdateTaskTool;
+
+impl UpdateTaskTool {
+    pub fn new() -> Self {
+        Self
+    }
+}
+
+impl Default for UpdateTaskTool {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+#[async_trait]
+impl Tool for UpdateTaskTool {
+    fn name(&self) -> &str {
+        "update_task"
+    }
+
+    fn description(&self) -> &str {
+        "Move or update a specific task card on a task board, addressed by `id`. \
+         Use this to advance the task you're working on: set `status` \
+         (`todo`/`in_progress`/`blocked`/`done`) to move it between columns, and/or \
+         update `objective`, `notes`, `evidence`, `blocker`, `plan`, \
+         `acceptanceCriteria`. When you finish, set `status: done` with `evidence`; \
+         if you cannot proceed, set `status: blocked` with a `blocker` reason. \
+         Targets the proactive `task-sources` board by default — pass `threadId` to \
+         target another thread's board. Returns the updated card list + a markdown \
+         rendering. At most one card may be `in_progress` at a time."
+    }
+
+    fn parameters_schema(&self) -> serde_json::Value {
+        json!({
+            "type": "object",
+            "properties": {
+                "id": { "type": "string", "description": "Id of the card to move/update (required)." },
+                "status": {
+                    "type": "string",
+                    "enum": ["todo", "in_progress", "blocked", "done"],
+                    "description": "New status — moves the card to that column."
+                },
+                "objective": { "type": "string", "description": "Updated desired outcome for the task." },
+                "notes": { "type": "string", "description": "Progress notes / running summary." },
+                "blocker": { "type": "string", "description": "Why the task is blocked (set with status=blocked)." },
+                "evidence": {
+                    "type": "array",
+                    "description": "Links, output, or files proving the work (set with status=done).",
+                    "items": { "type": "string" }
+                },
+                "plan": {
+                    "type": "array",
+                    "description": "Updated ordered execution steps.",
+                    "items": { "type": "string" }
+                },
+                "acceptanceCriteria": {
+                    "type": "array",
+                    "description": "Updated checklist that must hold before the task is done.",
+                    "items": { "type": "string" }
+                },
+                "threadId": {
+                    "type": "string",
+                    "description": "Board to target; defaults to the `task-sources` board."
+                }
+            },
+            "required": ["id"]
+        })
+    }
+
+    fn permission_level(&self) -> PermissionLevel {
+        PermissionLevel::None
+    }
+
+    async fn execute(&self, args: serde_json::Value) -> anyhow::Result<ToolResult> {
+        let Some(id) = optional_string(&args, "id") else {
+            return Ok(ToolResult::error("missing required field `id`".to_string()));
+        };
+
+        let patch = match build_patch(&args) {
+            Ok(patch) => patch,
+            Err(err) => return Ok(ToolResult::error(err)),
+        };
+        if patch_is_empty(&patch) {
+            return Ok(ToolResult::error(
+                "nothing to update — provide `status` and/or a field \
+                 (objective/notes/evidence/blocker/plan/acceptanceCriteria)"
+                    .to_string(),
+            ));
+        }
+
+        let location = match resolve_location(&args).await {
+            Ok(location) => location,
+            Err(err) => return Ok(ToolResult::error(err)),
+        };
+
+        Ok(apply(&location, &id, patch))
+    }
+}
+
+/// Apply the move/update to the card and render the result. Split out from
+/// `execute` so the edit + response shaping is testable without a fork/thread
+/// context (which `resolve_location` needs).
+fn apply(location: &BoardLocation, id: &str, patch: CardPatch) -> ToolResult {
+    tracing::info!(
+        card_id = %id,
+        thread_id = ?location.thread_id(),
+        status = ?patch.status,
+        "[tool][update_task] move/update task card"
+    );
+    match ops::edit(location, id, patch) {
+        Ok(snap) => {
+            let payload = json!({
+                "threadId": snap.thread_id,
+                "cards": snap.cards,
+                "markdown": snap.markdown,
+            });
+            ToolResult::success(payload.to_string())
+        }
+        Err(err) => ToolResult::error(err),
+    }
+}
+
+/// Resolve the board to act on: the explicit `threadId` arg, else the proactive
+/// `task-sources` board. The workspace root comes from the running agent's fork
+/// context when present, otherwise from the loaded config.
+async fn resolve_location(args: &serde_json::Value) -> Result<BoardLocation, String> {
+    let thread_id =
+        optional_string(args, "threadId").unwrap_or_else(|| TASK_SOURCES_THREAD_ID.to_string());
+    Ok(BoardLocation::Thread {
+        workspace_dir: workspace_dir().await?,
+        thread_id,
+    })
+}
+
+async fn workspace_dir() -> Result<PathBuf, String> {
+    if let Some(parent) = crate::openhuman::agent::harness::fork_context::current_parent() {
+        return Ok(parent.workspace_dir.clone());
+    }
+    crate::openhuman::config::ops::load_config_with_timeout()
+        .await
+        .map(|config| config.workspace_dir)
+        .map_err(|e| format!("update_task: failed to load config for workspace dir: {e}"))
+}
+
+fn build_patch(args: &serde_json::Value) -> Result<CardPatch, String> {
+    let status = match args.get("status").and_then(|v| v.as_str()) {
+        Some(s) => Some(ops::parse_status(s)?),
+        None => None,
+    };
+    Ok(CardPatch {
+        status,
+        objective: optional_string(args, "objective"),
+        plan: optional_string_array(args, "plan")?,
+        acceptance_criteria: optional_string_array(args, "acceptanceCriteria")?,
+        evidence: optional_string_array(args, "evidence")?,
+        notes: optional_string(args, "notes"),
+        blocker: optional_string(args, "blocker"),
+        ..Default::default()
+    })
+}
+
+fn patch_is_empty(patch: &CardPatch) -> bool {
+    patch.status.is_none()
+        && patch.objective.is_none()
+        && patch.plan.is_none()
+        && patch.acceptance_criteria.is_none()
+        && patch.evidence.is_none()
+        && patch.notes.is_none()
+        && patch.blocker.is_none()
+}
+
+fn optional_string(args: &serde_json::Value, key: &str) -> Option<String> {
+    args.get(key)
+        .and_then(|v| v.as_str())
+        .map(str::trim)
+        .filter(|s| !s.is_empty())
+        .map(str::to_string)
+}
+
+fn optional_string_array(
+    args: &serde_json::Value,
+    key: &str,
+) -> Result<Option<Vec<String>>, String> {
+    match args.get(key) {
+        None => Ok(None),
+        Some(serde_json::Value::Null) => Ok(None),
+        Some(serde_json::Value::Array(items)) => {
+            let mut out = Vec::with_capacity(items.len());
+            for item in items {
+                let s = item
+                    .as_str()
+                    .ok_or_else(|| format!("`{key}` must be an array of strings"))?
+                    .trim();
+                if !s.is_empty() {
+                    out.push(s.to_string());
+                }
+            }
+            Ok(Some(out))
+        }
+        Some(_) => Err(format!("`{key}` must be an array of strings")),
+    }
+}
+
+#[cfg(test)]
+#[path = "update_task_tests.rs"]
+mod tests;
diff --git a/src/openhuman/agent/tools/update_task_tests.rs b/src/openhuman/agent/tools/update_task_tests.rs
new file mode 100644
index 000000000..23774fdb1
--- /dev/null
+++ b/src/openhuman/agent/tools/update_task_tests.rs
@@ -0,0 +1,123 @@
+//! Tests for the `update_task` tool.
+
+use super::*;
+use crate::openhuman::agent::task_board::TaskCardStatus;
+use crate::openhuman::tools::traits::Tool;
+use serde_json::json;
+
+// ── build_patch ──────────────────────────────────────────────────────────────
+
+#[test]
+fn build_patch_maps_status_and_all_fields() {
+    let patch = build_patch(&json!({
+        "status": "in_progress",
+        "objective": "ship it",
+        "notes": "halfway",
+        "blocker": "",
+        "evidence": ["https://x/1", "  ", "note"],
+        "plan": ["a", "b"],
+        "acceptanceCriteria": ["covered"]
+    }))
+    .expect("valid patch");
+    assert_eq!(patch.status, Some(TaskCardStatus::InProgress));
+    assert_eq!(patch.objective.as_deref(), Some("ship it"));
+    assert_eq!(patch.notes.as_deref(), Some("halfway"));
+    // empty/blank array entries are dropped; blanks elsewhere become None via edit.
+    assert_eq!(
+        patch.evidence,
+        Some(vec!["https://x/1".to_string(), "note".to_string()])
+    );
+    assert_eq!(patch.plan, Some(vec!["a".to_string(), "b".to_string()]));
+    assert_eq!(patch.acceptance_criteria, Some(vec!["covered".to_string()]));
+}
+
+#[test]
+fn build_patch_empty_args_is_empty() {
+    let patch = build_patch(&json!({})).expect("ok");
+    assert!(patch_is_empty(&patch));
+}
+
+#[test]
+fn build_patch_with_only_status_is_not_empty() {
+    let patch = build_patch(&json!({ "status": "done" })).expect("ok");
+    assert!(!patch_is_empty(&patch));
+    assert_eq!(patch.status, Some(TaskCardStatus::Done));
+}
+
+#[test]
+fn build_patch_rejects_unknown_status() {
+    assert!(build_patch(&json!({ "status": "nonsense" })).is_err());
+}
+
+#[test]
+fn build_patch_rejects_non_string_array_item() {
+    assert!(build_patch(&json!({ "evidence": [1, 2] })).is_err());
+}
+
+// ── execute() guard rails (no board/workspace needed) ────────────────────────
+
+#[tokio::test]
+async fn execute_without_id_is_an_error() {
+    let res = UpdateTaskTool::new()
+        .execute(json!({ "status": "done" }))
+        .await
+        .unwrap();
+    assert!(res.is_error, "missing id must be a tool error");
+}
+
+#[tokio::test]
+async fn execute_with_id_but_no_changes_is_an_error() {
+    // id present but nothing to change → rejected before any board write.
+    let res = UpdateTaskTool::new()
+        .execute(json!({ "id": "task-1" }))
+        .await
+        .unwrap();
+    assert!(res.is_error, "empty update must be a tool error");
+}
+
+// ── the move + update applied through ops::edit (the tool's real effect) ─────
+
+#[test]
+fn apply_moves_and_updates_a_card_and_returns_success() {
+    let dir = tempfile::tempdir().unwrap();
+    let location = BoardLocation::Thread {
+        workspace_dir: dir.path().to_path_buf(),
+        thread_id: TASK_SOURCES_THREAD_ID.to_string(),
+    };
+    let id = ops::add(&location, "Review PR #5", CardPatch::default())
+        .unwrap()
+        .cards[0]
+        .id
+        .clone();
+
+    // Move to done + attach evidence — the exact shape the tool produces.
+    let patch = build_patch(&json!({
+        "status": "done",
+        "evidence": ["posted review on PR #5"]
+    }))
+    .unwrap();
+    let res = apply(&location, &id, patch);
+    assert!(!res.is_error, "successful move/update must not be an error");
+
+    // The board reflects the move + evidence.
+    let card = ops::list(&location)
+        .unwrap()
+        .cards
+        .into_iter()
+        .find(|c| c.id == id)
+        .unwrap();
+    assert_eq!(card.status, TaskCardStatus::Done);
+    assert!(card.evidence.iter().any(|e| e.contains("posted review")));
+}
+
+#[test]
+fn apply_on_unknown_id_returns_error() {
+    let dir = tempfile::tempdir().unwrap();
+    let location = BoardLocation::Thread {
+        workspace_dir: dir.path().to_path_buf(),
+        thread_id: TASK_SOURCES_THREAD_ID.to_string(),
+    };
+    let patch = build_patch(&json!({ "status": "blocked", "blocker": "no channel" })).unwrap();
+    let res = apply(&location, "task-does-not-exist", patch);
+    assert!(res.is_error, "missing card must surface as a tool error");
+}
diff --git a/src/openhuman/agent_registry/agents/orchestrator/agent.toml b/src/openhuman/agent_registry/agents/orchestrator/agent.toml
index 0e354a913..2ee6e58ae 100644
--- a/src/openhuman/agent_registry/agents/orchestrator/agent.toml
+++ b/src/openhuman/agent_registry/agents/orchestrator/agent.toml
@@ -148,6 +148,12 @@ named = [
     # downstream agents — the orchestrator coordinates, it doesn't
     # touch files itself.
     "todowrite",
+    # `update_task` moves/updates a specific task card by id on a target board
+    # (defaults to the proactive `task-sources` board) — so the orchestrator can
+    # advance the task it's working: → in_progress when it starts, → done with
+    # evidence when finished, or → blocked with a reason when stuck. Complements
+    # `todowrite`, which only touches the current thread's board.
+    "update_task",
     "plan_exit",
     # Skill chaining: let an in-flight autonomous skill (e.g.
     # `github-issue-crusher` after the draft PR is open) spawn another
diff --git a/src/openhuman/inference/provider/compatible.rs b/src/openhuman/inference/provider/compatible.rs
index 395e5d56c..36ae2a545 100644
--- a/src/openhuman/inference/provider/compatible.rs
+++ b/src/openhuman/inference/provider/compatible.rs
@@ -41,6 +41,19 @@ use compatible_types::{
     ResponsesRequest, StreamChunkResponse, StreamingToolCall, ToolCall,
 };
 
+/// `frequency_penalty` applied to streaming chat-completions requests.
+///
+/// Autoregressive models have a self-reinforcing bias toward repeating spans
+/// already in their context; with no penalty a momentary repeat can spiral into
+/// the same line emitted until the output-token cap (degenerate decoding). A
+/// small positive penalty damps that loop without harming coherence. Carried on
+/// the streaming path (where those loops occur — long autonomous turns) and
+/// retried without it if a strict provider rejects it; the buffered
+/// non-streaming fallback omits it for maximum compatibility. Skipped in
+/// serialisation when `None` so providers that don't accept the field are
+/// unaffected.
+const CHAT_FREQUENCY_PENALTY: f64 = 0.3;
+
 /// A provider that speaks the OpenAI-compatible chat completions API.
 /// Used by: Venice, Vercel AI Gateway, Cloudflare AI Gateway, Moonshot,
 /// Synthetic, `OpenCode` Zen, `Z.AI`, `GLM`, `MiniMax`, Bedrock, Qianfan, Groq, Mistral, `xAI`, etc.
@@ -929,6 +942,23 @@ impl OpenAiCompatibleProvider {
         Self::is_native_tool_schema_unsupported(reqwest::StatusCode::BAD_REQUEST, error)
     }
 
+    /// Detect a provider rejecting the `frequency_penalty` sampling field. Some
+    /// strict OpenAI-compatible backends 400 on unknown params; when this fires
+    /// the caller retries once with the field omitted (mirrors the no-tools
+    /// retry). String-based because the streamed transport error surfaces the
+    /// API error body.
+    fn err_indicates_frequency_penalty_unsupported(error: &str) -> bool {
+        let lower = error.to_lowercase();
+        lower.contains("frequency_penalty")
+            && (lower.contains("unsupported")
+                || lower.contains("unknown")
+                || lower.contains("unrecognized")
+                || lower.contains("not supported")
+                || lower.contains("does not support")
+                || lower.contains("invalid")
+                || lower.contains("unexpected"))
+    }
+
     /// Detect a 404 whose body says the model is completion-only and cannot be
     /// served from `/v1/chat/completions` (OpenAI: "This is not a chat model
     /// and thus not supported in the v1/chat/completions endpoint. Did you
@@ -1914,6 +1944,7 @@ impl Provider for OpenAiCompatibleProvider {
                     include_usage: true,
                 }),
                 options: self.build_ollama_options(),
+                frequency_penalty: Some(CHAT_FREQUENCY_PENALTY),
             };
             let stream_dump_seq = reserve_dump_seq();
             dump_prompt_if_enabled(&self.name, model, stream_dump_seq, &native_request);
@@ -1956,6 +1987,31 @@ impl Provider for OpenAiCompatibleProvider {
                                 );
                             }
                         }
+                    } else if Self::err_indicates_frequency_penalty_unsupported(&err_str) {
+                        // Symmetric to the no-tools retry: a strict provider that
+                        // 400s on `frequency_penalty` should degrade gracefully
+                        // rather than fail the whole chat path.
+                        log::info!(
+                            "[stream] {} rejected frequency_penalty — retrying streaming without it",
+                            self.name,
+                        );
+                        let retry_request = NativeChatRequest {
+                            frequency_penalty: None,
+                            ..native_request.clone()
+                        };
+                        match self
+                            .stream_native_chat(credential, &retry_request, tx, stream_dump_seq)
+                            .await
+                        {
+                            Ok(resp) => return Ok(resp),
+                            Err(retry_err) => {
+                                log::warn!(
+                                    "[stream] {} retry without frequency_penalty also failed, falling back to non-streaming: {}",
+                                    self.name,
+                                    retry_err
+                                );
+                            }
+                        }
                     } else {
                         log::warn!(
                             "[stream] {} streaming chat failed, falling back to non-streaming: {}",
@@ -1963,7 +2019,9 @@ impl Provider for OpenAiCompatibleProvider {
                             err
                         );
                     }
-                    // Fall through to the non-streaming path below.
+                    // Fall through to the non-streaming path below. The
+                    // non-streaming request below omits `frequency_penalty` so a
+                    // provider that rejected it (streaming or not) still succeeds.
                 }
             }
         }
@@ -1985,6 +2043,12 @@ impl Provider for OpenAiCompatibleProvider {
             thread_id,
             stream_options: None,
             options: self.build_ollama_options(),
+            // The buffered (non-streaming) path is the fallback / non-streaming
+            // provider path — omit `frequency_penalty` here for maximum
+            // compatibility (a provider that rejects it still succeeds). The
+            // streaming path above carries it (where degenerate repetition loops
+            // actually occur) and retries without it on rejection.
+            frequency_penalty: None,
         };
         let dump_seq = reserve_dump_seq();
         dump_prompt_if_enabled(&self.name, model, dump_seq, &native_request);
diff --git a/src/openhuman/inference/provider/compatible_tests.rs b/src/openhuman/inference/provider/compatible_tests.rs
index 591d03993..a1f3de390 100644
--- a/src/openhuman/inference/provider/compatible_tests.rs
+++ b/src/openhuman/inference/provider/compatible_tests.rs
@@ -66,6 +66,7 @@ fn native_request_emits_thread_id_when_present() {
         thread_id: Some("thread-abc".to_string()),
         stream_options: None,
         options: None,
+        frequency_penalty: None,
     };
     let json = serde_json::to_value(&req).unwrap();
     assert_eq!(
@@ -84,6 +85,7 @@ fn native_request_emits_thread_id_when_present() {
         thread_id: None,
         stream_options: None,
         options: None,
+        frequency_penalty: None,
     };
     let json_no_thread = serde_json::to_value(&req_no_thread).unwrap();
     assert!(
@@ -92,6 +94,58 @@ fn native_request_emits_thread_id_when_present() {
     );
 }
 
+#[test]
+fn native_request_serializes_frequency_penalty_only_when_set() {
+    let base = super::NativeChatRequest {
+        model: "kimi".to_string(),
+        messages: Vec::new(),
+        temperature: Some(0.7),
+        stream: Some(false),
+        tools: None,
+        tool_choice: None,
+        thread_id: None,
+        stream_options: None,
+        options: None,
+        frequency_penalty: Some(0.3),
+    };
+    let json = serde_json::to_value(&base).unwrap();
+    assert_eq!(
+        json.get("frequency_penalty")
+            .and_then(serde_json::Value::as_f64),
+        Some(0.3),
+        "a set frequency_penalty must be forwarded to damp repetition loops"
+    );
+
+    let none = super::NativeChatRequest {
+        frequency_penalty: None,
+        ..base
+    };
+    let json_none = serde_json::to_value(&none).unwrap();
+    assert!(
+        json_none.get("frequency_penalty").is_none(),
+        "absent frequency_penalty must be omitted so providers that reject it are unaffected"
+    );
+}
+
+#[test]
+fn detects_frequency_penalty_rejection_for_retry() {
+    use super::OpenAiCompatibleProvider as P;
+    // Strict providers that 400 on the field → retry without it.
+    assert!(P::err_indicates_frequency_penalty_unsupported(
+        "400 Bad Request: unknown parameter 'frequency_penalty'"
+    ));
+    assert!(P::err_indicates_frequency_penalty_unsupported(
+        "frequency_penalty is not supported by this model"
+    ));
+    // Unrelated errors, or the field merely mentioned, must NOT trigger a retry.
+    assert!(!P::err_indicates_frequency_penalty_unsupported(
+        "rate limit exceeded"
+    ));
+    assert!(!P::err_indicates_frequency_penalty_unsupported(
+        "applied frequency_penalty 0.3"
+    ));
+}
+
 /// Streaming responses arrive without `usage` unless the request asks
 /// for `stream_options.include_usage = true` (OpenAI spec). Without it
 /// the OpenHuman backend's `openhuman.billing` block also never lands,
@@ -111,6 +165,7 @@ fn streaming_request_sets_stream_options_include_usage() {
             include_usage: true,
         }),
         options: None,
+        frequency_penalty: None,
     };
     let json = serde_json::to_value(&req).unwrap();
     assert_eq!(
@@ -133,6 +188,7 @@ fn non_streaming_request_omits_stream_options() {
         thread_id: None,
         stream_options: None,
         options: None,
+        frequency_penalty: None,
     };
     let json = serde_json::to_value(&req).unwrap();
     assert!(
@@ -155,6 +211,7 @@ fn ollama_options_num_ctx_serializes_correctly() {
         options: Some(super::compatible_types::OllamaOptions {
             num_ctx: Some(32768),
         }),
+        frequency_penalty: None,
     };
     let json = serde_json::to_value(&req).unwrap();
     assert_eq!(
@@ -176,6 +233,7 @@ fn ollama_options_none_is_omitted() {
         thread_id: None,
         stream_options: None,
         options: None,
+        frequency_penalty: None,
     };
     let json = serde_json::to_value(&req).unwrap();
     assert!(
@@ -756,6 +814,7 @@ async fn streaming_chat_config_rejection_propagates_error_without_sentry_report(
             include_usage: true,
         }),
         options: None,
+        frequency_penalty: None,
     };
     let (delta_tx, _delta_rx) = tokio::sync::mpsc::channel(8);
 
diff --git a/src/openhuman/inference/provider/compatible_types.rs b/src/openhuman/inference/provider/compatible_types.rs
index 5a80aa492..b5cd83e03 100644
--- a/src/openhuman/inference/provider/compatible_types.rs
+++ b/src/openhuman/inference/provider/compatible_types.rs
@@ -187,6 +187,13 @@ pub(crate) struct NativeChatRequest {
     /// `num_ctx` override. Ignored (skipped) for non-Ollama providers.
     #[serde(skip_serializing_if = "Option::is_none")]
     pub(crate) options: Option<OllamaOptions>,
+    /// OpenAI-compatible `frequency_penalty`. Set to a small positive value on
+    /// real requests to damp degenerate repetition loops — without it a model
+    /// that starts repeating a line keeps emitting it until the output-token
+    /// cap (self-reinforcing decoding). Skipped when `None` so providers that
+    /// don't accept it are unaffected.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub(crate) frequency_penalty: Option<f64>,
 }
 
 /// Ollama-specific request options passed in the `options` field.
diff --git a/src/openhuman/todos/ops.rs b/src/openhuman/todos/ops.rs
index 14a4478fc..a6fc9e65f 100644
--- a/src/openhuman/todos/ops.rs
+++ b/src/openhuman/todos/ops.rs
@@ -19,6 +19,15 @@ use std::path::PathBuf;
 use std::sync::{Arc, OnceLock};
 use uuid::Uuid;
 
+/// Thread id backing the personal "user tasks" kanban board.
+///
+/// Must match the frontend constant in `app/src/services/api/todosApi.ts`.
+/// This is the board the kanban UI renders as the user's working lane and that
+/// the task dispatcher's board poller executes **agent-assigned** cards on
+/// (tasks approved out of the proactive `task-sources` inbox). Manually-created
+/// human cards on this board carry no `assigned_agent` and are never auto-run.
+pub const USER_TASKS_THREAD_ID: &str = "user-tasks";
+
 use super::store::{global_scratch_store, ScratchTodoStore};
 
 /// Serialise scratch CRUD so each public op's load → mutate → save
diff --git a/src/openhuman/tools/ops.rs b/src/openhuman/tools/ops.rs
index 68483f9ad..b1ff14009 100644
--- a/src/openhuman/tools/ops.rs
+++ b/src/openhuman/tools/ops.rs
@@ -158,6 +158,11 @@ pub fn all_tools_with_runtime(
         // build-mode pass. The plan→build mode switch itself is a
         // follow-up; the tool emits a stable marker today.
         Box::new(TodoTool::new()),
+        // Move/update a specific task card by id on a target board (defaults to
+        // the proactive `task-sources` board) — lets the agent advance the task
+        // it's working (in_progress / done+evidence / blocked+reason) from any
+        // thread, complementing `todo` which only touches the current thread.
+        Box::new(UpdateTaskTool::new()),
         Box::new(PlanExitTool::new()),
         // Skill chaining: let an in-flight autonomous skill (e.g.
         // `github-issue-crusher`) kick off another bundled skill_run as a