新增worker直接对话

server/engine.go

@@ -75,6 +75,7 @@ const (
 )
 
 var errWorkControlConflict = errors.New("work control conflict")
+var errWorkInterventionConflict = errors.New("work intervention conflict")
 
 // retryableWorkerModelError excludes errors already handled by the task router.
 // In particular, a quota error after partial streaming advances the task cursor
@@ -130,6 +131,17 @@ type Engine struct {
 	// runWorkerStep has stopped writing and committed its final state.
 	workMu sync.Mutex
 	work   map[int64]*workExecution
+	// interventionLocks serializes reservation, cancellation and message hand-off
+	// per intent. PostgreSQL separately serializes the durable transaction steps.
+	interventionLocks sync.Map // intentID -> *sync.Mutex
+
+	// directedClaims keeps intents with a new human message ahead of the normal
+	// frontier without mutating their persisted priority. runWorkerStep installs
+	// the entry before releasing ControlWork, so a single worker slot cannot grab
+	// unrelated work in the paused->open hand-off gap.
+	directedMu     sync.Mutex
+	directedClaims map[string][]int64
+	directedReady  map[int64]bool
 
 	// steerBox queues planner course-corrections for a running work (keyed by intent
 	// id). The worker's PreToolUse hook drains it before its next tool call and hands
@@ -162,7 +174,14 @@ type taskRuntime struct {
 type workExecution struct {
 	cancel context.CancelCauseFunc
 	done   chan error
-	action string // user action: pause | cancel
+	action string // user action: pause | cancel | intervene
+}
+
+type WorkInterventionResult struct {
+	RequestID  string
+	ActivityID int64
+	State      string
+	Duplicate  bool
 }
 
 // nextPlannerRound returns the next planner round number for a task (1-based).
@@ -285,6 +304,12 @@ func (e *Engine) StopTask(taskID string) {
 	e.stamped.Delete(taskID)
 	e.inflight.Delete(taskID)
 	e.coordStarted.Delete(taskID)
+	e.directedMu.Lock()
+	for _, intentID := range e.directedClaims[taskID] {
+		delete(e.directedReady, intentID)
+	}
+	delete(e.directedClaims, taskID)
+	e.directedMu.Unlock()
 	e.deleteMu.Lock()
 	e.deleting.Delete(taskID)
 	e.deleteMu.Unlock()
@@ -390,7 +415,7 @@ func NewEngine(m *Manager) *Engine {
 	return &Engine{m: m, debounce: 800 * time.Millisecond, bc: NewBroadcaster(),
 		execCancel: map[string]context.CancelCauseFunc{}, execCtx: map[string]context.Context{},
 		work: map[int64]*workExecution{}, steerBox: map[int64][]string{},
-		runtimes: map[string]*taskRuntime{}}
+		directedClaims: map[string][]int64{}, directedReady: map[int64]bool{}, runtimes: map[string]*taskRuntime{}}
 }
 
 // registerWork records the cancel for the work currently running intentID.
@@ -425,7 +450,7 @@ func (e *Engine) detachWork(intentID int64) (action string, complete func(error)
 // worker has fully stopped writing. Cancellation cleanup is performed by the API
 // handler after this returns; pause state is committed by runWorkerStep itself.
 func (e *Engine) ControlWork(ctx context.Context, intentID int64, action string) error {
-	if action != "pause" && action != "cancel" {
+	if action != "pause" && action != "cancel" && action != "intervene" {
 		return fmt.Errorf("unsupported work action %q", action)
 	}
 	if ctx == nil {
@@ -444,8 +469,11 @@ func (e *Engine) ControlWork(ctx context.Context, intentID int64, action string)
 	run.action = action
 	done := run.done
 	cause := error(agent.AbortWorkPausedByUser)
-	if action == "cancel" {
+	switch action {
+	case "cancel":
 		cause = agent.AbortWorkCancelledByUser
+	case "intervene":
+		cause = agent.AbortWorkIntervenedByUser
 	}
 	run.cancel(cause)
 	e.workMu.Unlock()
@@ -464,6 +492,287 @@ func (e *Engine) ControlWork(ctx context.Context, intentID int64, action string)
 	}
 }
 
+func (e *Engine) lockWorkIntervention(intentID int64) func() {
+	value, _ := e.interventionLocks.LoadOrStore(intentID, &sync.Mutex{})
+	mu := value.(*sync.Mutex)
+	mu.Lock()
+	return mu.Unlock
+}
+
+func (e *Engine) validateWorkInterventionTask(t *Task, allowAdmissionPause bool) error {
+	if t == nil {
+		return fmt.Errorf("%w: task not found", errWorkInterventionConflict)
+	}
+	lifecycle := t.lifecycleSnapshot()
+	switch {
+	case e.IsDeleting(t.ID):
+		return fmt.Errorf("%w: 任务正在删除，无法向 Worker 发送消息", errWorkInterventionConflict)
+	case lifecycle.Paused || (!allowAdmissionPause && e.IsPaused(t.ID)):
+		return fmt.Errorf("%w: 任务已暂停，无法向 Worker 发送消息", errWorkInterventionConflict)
+	case lifecycle.Queued:
+		return fmt.Errorf("%w: 排队中的任务无法向 Worker 发送消息", errWorkInterventionConflict)
+	case isTerminalStatus(lifecycle.Status):
+		return fmt.Errorf("%w: 终态任务无法向 Worker 发送消息", errWorkInterventionConflict)
+	case e.isSettling(t.ID):
+		return fmt.Errorf("%w: 任务正在收尾，无法向 Worker 发送消息", errWorkInterventionConflict)
+	default:
+		return nil
+	}
+}
+
+func (e *Engine) abandonWorkIntervention(t *Task, intentID int64, requestID string, reason error) {
+	e.removeDirectedClaim(t.ID, intentID)
+	target := "open"
+	lifecycle := t.lifecycleSnapshot()
+	if e.IsDeleting(t.ID) || isTerminalStatus(lifecycle.Status) {
+		target = "stopped"
+	} else if e.isSettling(t.ID) {
+		target = "exhausted"
+	}
+	if changed, err := t.Store.CompareAndSetIntentState(intentID, "paused", target); err != nil {
+		log.Printf("[worker intervention] task %s 意图 #%d 补偿 paused->%s 失败: %v", t.ID, intentID, target, err)
+	} else if changed && target == "open" {
+		t.NotifyWorker()
+	}
+	if err := t.Store.RejectIntentIntervention(intentID, requestID, reason.Error()); err != nil {
+		log.Printf("[worker intervention] task %s 意图 #%d 标记请求 %s rejected 失败: %v", t.ID, intentID, requestID, err)
+	}
+}
+
+// InterveneWork performs the complete Worker message protocol. Once the request
+// has a durable pending reservation it no longer depends on the HTTP request
+// context: the caller supplies a server-owned bounded context so disconnects do
+// not strand the intent. The accepted DB transition is running -> paused ->
+// open, with the visible user message and reopen committed atomically.
+func (e *Engine) InterveneWork(ctx context.Context, t *Task, intentID int64, requestID, message string) (WorkInterventionResult, error) {
+	return e.interveneWork(ctx, t, intentID, requestID, message, false)
+}
+
+func (e *Engine) interveneWork(ctx context.Context, t *Task, intentID int64, requestID, message string, allowAdmissionPause bool) (WorkInterventionResult, error) {
+	out := WorkInterventionResult{RequestID: requestID}
+	if t == nil || t.Store == nil {
+		return out, fmt.Errorf("%w: task not found", errWorkInterventionConflict)
+	}
+	if ctx == nil {
+		ctx = context.Background()
+	}
+	unlock := e.lockWorkIntervention(intentID)
+	defer unlock()
+
+	existing, err := t.Store.IntentInterventionByRequest(intentID, requestID)
+	if err != nil {
+		return out, err
+	}
+	if existing != nil {
+		if existing.Message != message {
+			return out, fmt.Errorf("%w: request_id 已用于不同的 Worker 消息", errWorkInterventionConflict)
+		}
+		if existing.Status == db.IntentInterventionAccepted {
+			out.ActivityID, out.State, out.Duplicate = existing.ActivityID, existing.IntentState, true
+			if out.State == "open" && e.validateWorkInterventionTask(t, allowAdmissionPause) == nil {
+				e.queueDirectedClaim(t.ID, intentID)
+				e.readyDirectedClaim(intentID)
+				t.NotifyWorker()
+			}
+			return out, nil
+		}
+		if existing.Status != db.IntentInterventionPending {
+			return out, fmt.Errorf("%w: Worker 消息请求状态为 %s", errWorkInterventionConflict, existing.Status)
+		}
+	}
+	if err := e.validateWorkInterventionTask(t, allowAdmissionPause); err != nil {
+		lifecycle := t.lifecycleSnapshot()
+		if existing != nil && (e.IsDeleting(t.ID) || e.isSettling(t.ID) || isTerminalStatus(lifecycle.Status)) {
+			e.abandonWorkIntervention(t, intentID, requestID, err)
+		}
+		return out, err
+	}
+	_, worker := e.snapshotFor(t)
+	if worker == nil {
+		return out, fmt.Errorf("%w: worker 尚未就绪", errWorkInterventionConflict)
+	}
+	if existing == nil {
+		existing, err = t.Store.ReserveIntentIntervention(intentID, requestID, message)
+		if err != nil {
+			if errors.Is(err, db.ErrIntentInterventionConflict) {
+				return out, fmt.Errorf("%w: %v", errWorkInterventionConflict, err)
+			}
+			return out, err
+		}
+	}
+	out.ActivityID = existing.ActivityID
+
+	// A pending reservation can be recovered after a process restart. An open
+	// row is fenced back to paused before publishing the user turn; a running row
+	// is stopped through the normal live-work protocol so no late writer races us.
+	for {
+		node, getErr := t.Store.GetNode(intentID)
+		if getErr != nil {
+			return out, getErr
+		}
+		if node == nil {
+			return out, fmt.Errorf("%w: intent not found", errWorkInterventionConflict)
+		}
+		switch node.State {
+		case "running":
+			controlErr := e.ControlWork(ctx, intentID, "intervene")
+			if controlErr != nil {
+				// The per-work 30s wait may expire before the server-owned operation
+				// context. Keep waiting for the named cancellation to settle; timeout
+				// release marks the run no-handoff so a late exit cannot install a
+				// permanent directed-claim barrier behind our back. A freshly claimed
+				// intent can also be running just before registerWork; retrying closes
+				// that small claim-to-registration window.
+				for {
+					settled, readErr := t.Store.GetNode(intentID)
+					if readErr != nil {
+						return out, readErr
+					}
+					if settled != nil && settled.State == "paused" {
+						break
+					}
+					if ctx.Err() != nil || settled == nil || settled.State != "running" {
+						e.removeDirectedClaim(t.ID, intentID)
+						return out, fmt.Errorf("%w: %v", errWorkInterventionConflict, controlErr)
+					}
+					if sleepCtx(ctx, 50*time.Millisecond) {
+						e.removeDirectedClaim(t.ID, intentID)
+						return out, fmt.Errorf("%w: %v", errWorkInterventionConflict, controlErr)
+					}
+					if errors.Is(controlErr, errWorkControlConflict) {
+						controlErr = e.ControlWork(ctx, intentID, "intervene")
+					}
+				}
+			}
+		case "open":
+			changed, setErr := t.Store.CompareAndSetIntentState(intentID, "open", "paused")
+			if setErr != nil {
+				return out, setErr
+			}
+			if !changed {
+				continue
+			}
+			e.queueDirectedClaim(t.ID, intentID)
+		case "paused":
+			e.queueDirectedClaim(t.ID, intentID)
+		default:
+			e.abandonWorkIntervention(t, intentID, requestID,
+				fmt.Errorf("意图已进入 %s，无法发送新消息", node.State))
+			return out, fmt.Errorf("%w: 意图已进入 %s，无法发送新消息", errWorkInterventionConflict, node.State)
+		}
+		break
+	}
+	e.queueDirectedClaim(t.ID, intentID)
+
+	// Close lifecycle races that began while cancellation was settling. A task
+	// pause or queue preserves the pending message and paused intent for a later
+	// admission; terminal, settling and deleting transitions close the request.
+	if err := e.validateWorkInterventionTask(t, allowAdmissionPause); err != nil {
+		// The user message has not entered the transcript yet. Preserve the
+		// reservation and paused intent so the same request can be retried after a
+		// transient task pause; do not reopen it behind a terminal transition.
+		e.removeDirectedClaim(t.ID, intentID)
+		lifecycle := t.lifecycleSnapshot()
+		if e.IsDeleting(t.ID) || e.isSettling(t.ID) || isTerminalStatus(lifecycle.Status) {
+			e.abandonWorkIntervention(t, intentID, requestID, err)
+		}
+		return out, err
+	}
+	activity, duplicate, err := t.Store.AcceptIntentIntervention(intentID, requestID, message)
+	if err != nil {
+		// Keep this intent paused so a retry of the same request can safely finish;
+		// remove only the task-wide claim barrier so unrelated workers keep moving.
+		e.removeDirectedClaim(t.ID, intentID)
+		return out, err
+	}
+	if !duplicate {
+		e.publishStoredActivity(t, activity)
+		t.NotifyWorkerMessage(intentID, activity.ID, message)
+	}
+	e.readyDirectedClaim(intentID)
+	t.NotifyWorker()
+	node, err := t.Store.GetNode(intentID)
+	if err != nil {
+		return out, err
+	}
+	out.Duplicate = duplicate
+	if node != nil {
+		out.State = node.State
+	}
+	return out, nil
+}
+
+// RecoverWorkerMessages reconstructs unfinished Worker chat handoffs before
+// ordinary frontier claiming. Pending messages are fenced in paused until an
+// agent runtime is available; accepted messages regain their directed priority
+// and are consumed by ExecuteWithMessage on the next claim.
+func (e *Engine) RecoverWorkerMessages(ctx context.Context, t *Task) error {
+	return e.recoverWorkerMessages(ctx, t, false)
+}
+
+// recoverWorkerMessagesForAdmission runs while the scheduler still owns the
+// Engine pause barrier but has already committed a runnable task lifecycle. It
+// restores directed hand-offs before Resume wakes any worker, so ordinary
+// frontier work cannot overtake the human message.
+func (e *Engine) recoverWorkerMessagesForAdmission(ctx context.Context, t *Task) error {
+	return e.recoverWorkerMessages(ctx, t, true)
+}
+
+func (e *Engine) recoverWorkerMessages(ctx context.Context, t *Task, allowAdmissionPause bool) error {
+	if t == nil || t.Store == nil {
+		return nil
+	}
+	items, err := t.Store.RecoverableIntentInterventions()
+	if err != nil {
+		return err
+	}
+	lifecycle := t.lifecycleSnapshot()
+	runnable := !e.IsDeleting(t.ID) && !e.isSettling(t.ID) && !isTerminalStatus(lifecycle.Status) &&
+		!lifecycle.Paused && !lifecycle.Queued && (allowAdmissionPause || !e.IsPaused(t.ID))
+	if !runnable {
+		for _, item := range items {
+			e.removeDirectedClaim(t.ID, item.IntentID)
+		}
+		return nil
+	}
+	for _, item := range items {
+		switch item.Status {
+		case db.IntentInterventionPending:
+			if item.IntentState == "open" {
+				if changed, setErr := t.Store.CompareAndSetIntentState(item.IntentID, "open", "paused"); setErr != nil {
+					return setErr
+				} else if changed {
+					item.IntentState = "paused"
+				}
+			}
+			if _, worker := e.snapshotFor(t); worker == nil {
+				continue
+			}
+			if _, recoverErr := e.interveneWork(ctx, t, item.IntentID, item.RequestID, item.Message, allowAdmissionPause); recoverErr != nil {
+				log.Printf("[worker chat] task %s 恢复意图 #%d 待处理消息失败: %v", t.ID, item.IntentID, recoverErr)
+			}
+		case db.IntentInterventionAccepted:
+			// Reconstruct the Planner-facing event while the Worker hand-off is still
+			// pending. The durable graph_overview view remains authoritative even if
+			// this notification was already consumed before a later restart.
+			t.NotifyWorkerMessage(item.IntentID, item.ActivityID, item.Message)
+			if item.IntentState == "paused" {
+				if changed, setErr := t.Store.CompareAndSetIntentState(item.IntentID, "paused", "open"); setErr != nil {
+					return setErr
+				} else if changed {
+					item.IntentState = "open"
+				}
+			}
+			if item.IntentState == "open" {
+				e.queueDirectedClaim(t.ID, item.IntentID)
+				e.readyDirectedClaim(item.IntentID)
+				t.NotifyWorker()
+			}
+		}
+	}
+	return nil
+}
+
 // releaseWorkControl drops only this caller's reservation after its wait is
 // cancelled. The work context stays cancelled; runWorkerStep recognizes the
 // named cancellation cause and settles the intent into the recoverable paused
@@ -471,7 +780,11 @@ func (e *Engine) ControlWork(ctx context.Context, intentID int64, action string)
 func (e *Engine) releaseWorkControl(intentID int64, run *workExecution, action string) {
 	e.workMu.Lock()
 	if current := e.work[intentID]; current == run && current.action == action {
-		current.action = ""
+		if action == "intervene" {
+			current.action = "intervene_no_handoff"
+		} else {
+			current.action = ""
+		}
 	}
 	e.workMu.Unlock()
 }
@@ -610,6 +923,19 @@ func (e *Engine) emitActivity(t *Task, r db.Activity) db.Activity {
 	return r
 }
 
+// publishStoredActivity broadcasts an activity inserted by a larger DB
+// transaction, avoiding a second INSERT through emitActivity.
+func (e *Engine) publishStoredActivity(t *Task, r db.Activity) {
+	if t == nil || r.ID <= 0 {
+		return
+	}
+	if r.CreatedAt.IsZero() {
+		r.CreatedAt = time.Now()
+	}
+	e.bc.Publish(t.ID, r)
+	e.touch(t.ID)
+}
+
 // appendActivity persists one activity row, retrying briefly on write failure.
 // Concurrent planner + worker inserts into the same exploration's activity log
 // occasionally fail; a couple of quick retries recover most. Crucially, every
@@ -876,13 +1202,13 @@ func (e *Engine) workerLoop(ctx context.Context, t *Task, name string) {
 		}
 		_, worker := e.snapshotFor(t)
 		if worker == nil {
-			if sleepCtx(ctx, 1500*time.Millisecond) {
+			if waitWorker(ctx, t, 1500*time.Millisecond) {
 				return
 			}
 			continue
 		}
 		if e.IsPaused(t.ID) {
-			if sleepCtx(ctx, 1000*time.Millisecond) {
+			if waitWorker(ctx, t, 1000*time.Millisecond) {
 				return
 			}
 			continue // user-paused: don't claim/execute intents
@@ -891,13 +1217,13 @@ func (e *Engine) workerLoop(ctx context.Context, t *Task, name string) {
 			return
 		}
 		if e.isSettling(t.ID) {
-			if sleepCtx(ctx, 1000*time.Millisecond) {
+			if waitWorker(ctx, t, 1000*time.Millisecond) {
 				return
 			}
 			continue // 任务超时收尾中:不再领新意图(在跑的自行收尾,协调器等其 drain)
 		}
 		if isTerminalStatus(e.m.TaskStatus(t.ID)) {
-			if sleepCtx(ctx, 1000*time.Millisecond) {
+			if waitWorker(ctx, t, 1000*time.Millisecond) {
 				return
 			}
 			continue // 任务已终态(done/failed/timeout):停止领取遗留意图,别在完成后空跑 frontier
@@ -907,7 +1233,7 @@ func (e *Engine) workerLoop(ctx context.Context, t *Task, name string) {
 		}
 		claimed := e.runWorkerStep(ctx, t, name, worker)
 		e.decInflight(t.ID)
-		if !claimed && sleepCtx(ctx, 800*time.Millisecond) {
+		if !claimed && waitWorker(ctx, t, 800*time.Millisecond) {
 			return
 		}
 	}
@@ -946,9 +1272,31 @@ func (e *Engine) runWorkerStep(ctx context.Context, t *Task, name string, worker
 	workCtx = intercept.WithTaskContext(workCtx, t.ID, fmt.Sprintf("%s · #%d", name, iid), taskEmit)
 	hooks := steerHooks{inner: t.Guard.Hooks(), drain: func() (string, bool) { return e.drainSteer(iid) }}
 	wTaskID, _ := strconv.ParseInt(t.ID, 10, 64)
+	chatRequestID, chatMessage, hasChatMessage, chatErr := t.Store.PendingIntentInterventionMessage(intent.ID)
+	if chatErr != nil {
+		log.Printf("[worker %s] task %s 读取意图 #%d 待处理人工对话失败: %v", name, t.ID, intent.ID, chatErr)
+		if err := transitionIntentState(t.Store, intent.ID, "running", "open"); err != nil {
+			log.Printf("[worker %s] task %s 意图 #%d 人工对话读取失败后回退失败: %v", name, t.ID, intent.ID, err)
+		}
+		_, completeWork := e.detachWork(intent.ID)
+		completeWork(chatErr)
+		return true
+	}
 	e.BeginLLMCall(t.ID)
-	reason, wrote, err := worker.Execute(workCtx, name, wTaskID, e.m.assets, t.Store, intent, hooks, emit, e.m.enrich, t.NotifyFinding)
+	var reason harness.TerminalReason
+	var wrote agent.WriteCounts
+	var err error
+	if hasChatMessage {
+		reason, wrote, err = worker.ExecuteWithMessage(workCtx, name, wTaskID, e.m.assets, t.Store, intent, hooks, emit, e.m.enrich, t.NotifyFinding, chatRequestID, chatMessage)
+	} else {
+		reason, wrote, err = worker.Execute(workCtx, name, wTaskID, e.m.assets, t.Store, intent, hooks, emit, e.m.enrich, t.NotifyFinding)
+	}
 	e.EndLLMCall(t.ID)
+	if hasChatMessage {
+		if markErr := t.Store.MarkIntentInterventionHandoffClaimed(intent.ID); markErr != nil {
+			log.Printf("[worker %s] task %s 标记意图 #%d 人工对话已消费失败: %v", name, t.ID, intent.ID, markErr)
+		}
+	}
 	// model_error 收场 → 额外重跑几次（退避后再试）。仅在意图仍属本 work、任务
 	// 未暂停/未终止/未取消【且未进入收尾】时重试；否则让位给对应分支处理(收尾期不
 	// 再重试,避免退避挤占其他 worker 的优雅收尾窗口)。
@@ -961,7 +1309,11 @@ func (e *Engine) runWorkerStep(ctx context.Context, t *Task, name string, worker
 			break // 退避期间被取消（终止/暂停）→ 交给下方分支处理
 		}
 		e.BeginLLMCall(t.ID)
-		reason, wrote, err = worker.Execute(workCtx, name, wTaskID, e.m.assets, t.Store, intent, hooks, emit, e.m.enrich, t.NotifyFinding)
+		if hasChatMessage {
+			reason, wrote, err = worker.ExecuteWithMessage(workCtx, name, wTaskID, e.m.assets, t.Store, intent, hooks, emit, e.m.enrich, t.NotifyFinding, chatRequestID, chatMessage)
+		} else {
+			reason, wrote, err = worker.Execute(workCtx, name, wTaskID, e.m.assets, t.Store, intent, hooks, emit, e.m.enrich, t.NotifyFinding)
+		}
 		e.EndLLMCall(t.ID)
 	}
 	// Capture kill state before detachWork cancels workCtx. kill = this work's
@@ -979,6 +1331,8 @@ func (e *Engine) runWorkerStep(ctx context.Context, t *Task, name string, worker
 		switch {
 		case errors.Is(workCause, agent.AbortWorkPausedByUser):
 			action = "pause"
+		case errors.Is(workCause, agent.AbortWorkIntervenedByUser):
+			action = "intervene"
 		case errors.Is(workCause, agent.AbortWorkCancelledByUser):
 			action = "cancel"
 		}
@@ -995,6 +1349,22 @@ func (e *Engine) runWorkerStep(ctx context.Context, t *Task, name string, worker
 		e.touch(t.ID)
 		return true
 	}
+	if action == "intervene" || action == "intervene_no_handoff" {
+		controlErr = transitionIntentState(t.Store, intent.ID, "running", "paused")
+		if controlErr != nil {
+			log.Printf("[worker %s] task %s 意图 #%d 对话暂停落库失败: %v", name, t.ID, intent.ID, controlErr)
+			return true
+		}
+		// Install the directed barrier before completeWork releases ControlWork.
+		// With one worker slot this prevents its next loop from taking unrelated
+		// frontier work during the message hand-off.
+		if action == "intervene" {
+			e.queueDirectedClaim(t.ID, intent.ID)
+		}
+		log.Printf("[worker %s] task %s 意图 #%d 已暂停，等待新的对话意图", name, t.ID, intent.ID)
+		e.touch(t.ID)
+		return true
+	}
 	if action == "cancel" {
 		// Park the stopped run in paused before handing cleanup to the API. If the
 		// request disconnects after cancellation, the intent remains recoverable and
@@ -1099,7 +1469,114 @@ func sleepCtx(ctx context.Context, d time.Duration) (done bool) {
 	}
 }
 
+func waitWorker(ctx context.Context, t *Task, d time.Duration) (done bool) {
+	timer := time.NewTimer(d)
+	defer timer.Stop()
+	var wake <-chan struct{}
+	if t != nil {
+		wake = t.workerWake
+	}
+	select {
+	case <-ctx.Done():
+		return true
+	case <-wake:
+		return false
+	case <-timer.C:
+		return false
+	}
+}
+
+func (e *Engine) queueDirectedClaim(taskID string, intentID int64) {
+	if taskID == "" || intentID <= 0 {
+		return
+	}
+	e.directedMu.Lock()
+	defer e.directedMu.Unlock()
+	for _, id := range e.directedClaims[taskID] {
+		if id == intentID {
+			return
+		}
+	}
+	e.directedClaims[taskID] = append(e.directedClaims[taskID], intentID)
+	e.directedReady[intentID] = false
+}
+
+func (e *Engine) readyDirectedClaim(intentID int64) {
+	e.directedMu.Lock()
+	if _, exists := e.directedReady[intentID]; exists {
+		e.directedReady[intentID] = true
+	}
+	e.directedMu.Unlock()
+}
+
+func (e *Engine) removeDirectedClaim(taskID string, intentID int64) {
+	e.directedMu.Lock()
+	defer e.directedMu.Unlock()
+	items := e.directedClaims[taskID]
+	for i, id := range items {
+		if id != intentID {
+			continue
+		}
+		items = append(items[:i], items[i+1:]...)
+		delete(e.directedReady, intentID)
+		if len(items) == 0 {
+			delete(e.directedClaims, taskID)
+		} else {
+			e.directedClaims[taskID] = items
+		}
+		return
+	}
+}
+
+// claimDirected tries every pending human-message hand-off before ordinary work
+// frontier. blocked=true means at least one hand-off is still paused/running;
+// callers must not let a just-released single worker slot overtake it.
+func (e *Engine) claimDirected(t *Task, name string) (intent *db.Node, blocked bool) {
+	e.directedMu.Lock()
+	ids := append([]int64(nil), e.directedClaims[t.ID]...)
+	ready := make(map[int64]bool, len(ids))
+	for _, id := range ids {
+		ready[id] = e.directedReady[id]
+	}
+	e.directedMu.Unlock()
+	for _, id := range ids {
+		node, err := t.Store.GetNode(id)
+		if err != nil {
+			return nil, true
+		}
+		if node == nil {
+			e.removeDirectedClaim(t.ID, id)
+			continue
+		}
+		switch node.State {
+		case "open":
+			if !ready[id] {
+				blocked = true
+				continue
+			}
+			ok, err := t.Store.ClaimIntent(id, name)
+			if err != nil {
+				return nil, true
+			}
+			if ok {
+				e.removeDirectedClaim(t.ID, id)
+				node.State, node.Owner = "running", name
+				return node, false
+			}
+			blocked = true
+		case "paused", "running":
+			blocked = true
+		default:
+			e.removeDirectedClaim(t.ID, id)
+		}
+	}
+	return nil, blocked
+}
+
 func (e *Engine) claimNext(t *Task, name string) *db.Node {
+	if intent, blocked := e.claimDirected(t, name); intent != nil || blocked {
+		return intent
+	}
 	fr, _ := t.Store.Frontier(20)
 	for _, in := range fr {
 		if ok, _ := t.Store.ClaimIntent(in.ID, name); ok {

server/engine_cancelcause_test.go

@@ -45,6 +45,7 @@ func TestControlWorkCarriesUserActionCause(t *testing.T) {
 	}{
 		{action: "pause", code: "work_paused_by_user"},
 		{action: "cancel", code: "work_cancelled_by_user"},
+		{action: "intervene", code: "work_intervened_by_user"},
 	} {
 		t.Run(tc.action, func(t *testing.T) {
 			e := NewEngine(nil)

server/goals.go

@@ -225,6 +225,13 @@ func (s *Server) hasReadyQueuedTask(excludeID string) bool {
 }
 
 func (s *Server) startAdmittedTask(t *Task, mode string) {
+	recoverMessages := s.engine.RecoverWorkerMessages
+	if s.engine.IsPaused(t.ID) {
+		recoverMessages = s.engine.recoverWorkerMessagesForAdmission
+	}
+	if err := recoverMessages(s.ctx, t); err != nil {
+		log.Printf("[worker chat] task %s 恢复待处理人工消息失败: %v", t.ID, err)
+	}
 	if mode == "bootstrap" {
 		if !s.engine.beginTaskOperation(t.ID) {
 			return

server/intent_intervention.go

@@ -0,0 +1,125 @@
+package server
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"net/http"
+	"strconv"
+	"strings"
+	"time"
+
+	"github.com/Autumn-27/artex/db"
+)
+
+const maxWorkerMessageBytes = 64 << 10
+
+func validWorkerMessageRequestID(id string) bool {
+	if id == "" || len(id) > 128 {
+		return false
+	}
+	for _, r := range id {
+		if (r >= 'a' && r <= 'z') || (r >= 'A' && r <= 'Z') ||
+			(r >= '0' && r <= '9') || r == '-' || r == '_' || r == '.' || r == ':' {
+			continue
+		}
+		return false
+	}
+	return true
+}
+
+// sendWorkerMessage pauses a running Worker when necessary, records one user
+// turn for its existing conversation, and immediately schedules that same
+// intent to continue. A disconnected browser cannot strand an accepted message.
+func (s *Server) sendWorkerMessage(w http.ResponseWriter, r *http.Request) {
+	t, ok := s.m.Task(r.PathValue("id"))
+	if !ok {
+		writeErr(w, http.StatusNotFound, "task not found")
+		return
+	}
+	iid, err := strconv.ParseInt(r.PathValue("iid"), 10, 64)
+	if err != nil || iid <= 0 {
+		writeErr(w, http.StatusBadRequest, "bad intent id")
+		return
+	}
+	r.Body = http.MaxBytesReader(w, r.Body, maxWorkerMessageBytes)
+	var req struct {
+		Message   string `json:"message"`
+		RequestID string `json:"request_id"`
+	}
+	if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
+		var tooLarge *http.MaxBytesError
+		if errors.As(err, &tooLarge) {
+			writeErr(w, http.StatusRequestEntityTooLarge, "请求体过大")
+			return
+		}
+		writeErr(w, http.StatusBadRequest, "bad json: "+err.Error())
+		return
+	}
+	message := strings.TrimSpace(req.Message)
+	requestID := strings.TrimSpace(req.RequestID)
+	if message == "" {
+		writeErr(w, http.StatusBadRequest, "消息不能为空")
+		return
+	}
+	if len([]rune(message)) > 4000 {
+		writeErr(w, http.StatusBadRequest, "消息不能超过 4000 个字符")
+		return
+	}
+	if !validWorkerMessageRequestID(requestID) {
+		writeErr(w, http.StatusBadRequest, "request_id 必须是 1-128 位字母、数字、-、_、. 或 :")
+		return
+	}
+
+	if !s.engine.beginTaskOperation(t.ID) {
+		writeErr(w, http.StatusConflict, "任务正在删除，无法向 Worker 发送消息")
+		return
+	}
+	defer s.engine.decInflight(t.ID)
+
+	node, err := t.Store.GetNode(iid)
+	if err != nil {
+		writeErr(w, http.StatusInternalServerError, err.Error())
+		return
+	}
+	if node == nil {
+		if inherited, sourceErr := t.Store.GetNodeWithSources(iid); sourceErr == nil && inherited != nil && inherited.Inherited {
+			writeErr(w, http.StatusConflict, "继承意图为只读，不能发送 Worker 消息")
+			return
+		}
+		writeErr(w, http.StatusNotFound, "intent not found")
+		return
+	}
+	if node.Kind != db.KindIntent {
+		writeErr(w, http.StatusConflict, "node is not an intent")
+		return
+	}
+	base := s.ctx
+	if base == nil {
+		base = context.Background()
+	}
+	opCtx, cancel := context.WithTimeout(base, workControlWaitTimeout+15*time.Second)
+	defer cancel()
+	result, err := s.engine.InterveneWork(opCtx, t, iid, requestID, message)
+	if err != nil {
+		switch {
+		case errors.Is(err, context.DeadlineExceeded), errors.Is(err, context.Canceled):
+			writeErr(w, http.StatusGatewayTimeout, err.Error())
+		case errors.Is(err, errWorkInterventionConflict), errors.Is(err, errWorkControlConflict),
+			errors.Is(err, db.ErrIntentInterventionConflict), errors.Is(err, db.ErrIntentStateConflict):
+			writeErr(w, http.StatusConflict, err.Error())
+		default:
+			writeErr(w, http.StatusInternalServerError, err.Error())
+		}
+		return
+	}
+
+	writeJSON(w, http.StatusOK, map[string]any{
+		"id":           iid,
+		"state":        result.State,
+		"accepted":     true,
+		"activity_seq": result.ActivityID,
+		"request_id":   result.RequestID,
+		"duplicate":    result.Duplicate,
+	})
+}

server/intent_intervention_protocol_test.go

@@ -0,0 +1,378 @@
+package server
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"errors"
+	"iter"
+	"net/http"
+	"net/http/httptest"
+	"strconv"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/Autumn-27/artex/agent"
+	"github.com/Autumn-27/artex/db"
+	"github.com/Autumn-27/norma/llm"
+	"github.com/Autumn-27/norma/transcript"
+)
+
+type workerMessageTestProvider struct{}
+
+func (workerMessageTestProvider) Stream(_ context.Context, _ llm.CompletionRequest) iter.Seq2[llm.StreamEvent, error] {
+	return func(yield func(llm.StreamEvent, error) bool) {
+		yield(llm.StreamEvent{Type: llm.SETextDelta, Text: "已按新的对话意图继续执行"}, nil)
+		yield(llm.StreamEvent{Type: llm.SEMessageDelta, StopReason: "end_turn"}, nil)
+		yield(llm.StreamEvent{Type: llm.SEMessageStop}, nil)
+	}
+}
+
+func (p workerMessageTestProvider) Complete(ctx context.Context, req llm.CompletionRequest) (llm.Message, string, llm.Usage, error) {
+	acc := llm.NewAccumulator()
+	for event, err := range p.Stream(ctx, req) {
+		if err != nil {
+			return llm.Message{}, "", llm.Usage{}, err
+		}
+		acc.Add(event)
+	}
+	return acc.Message(), acc.StopReason, acc.Usage, nil
+}
+
+type interventionProtocolFixture struct {
+	m        *Manager
+	task     *Task
+	engine   *Engine
+	worker   *agent.Worker
+	tx       *transcript.Store
+	intentID int64
+}
+
+func newInterventionProtocolFixture(t *testing.T, targetPriority int) *interventionProtocolFixture {
+	t.Helper()
+	m, err := NewManager(t.TempDir(), "")
+	if err != nil {
+		t.Skipf("postgres unavailable (%v) - skipping", err)
+	}
+	task, err := m.CreateTask("worker message protocol", "verify immediate conversation handoff", nil, 0, 0)
+	if err != nil {
+		m.Close()
+		t.Fatal(err)
+	}
+	intentID, err := task.Store.AddIntent(map[string]any{"summary": "target worker intent"}, targetPriority, nil, "planner")
+	if err != nil {
+		m.Close()
+		t.Fatal(err)
+	}
+	if claimed, err := task.Store.ClaimIntent(intentID, "work#1"); err != nil || !claimed {
+		m.Close()
+		t.Fatalf("claim target intent: claimed=%v err=%v", claimed, err)
+	}
+	tx := transcript.NewStore(m.dir + "/transcripts")
+	worker := agent.NewWorker(workerMessageTestProvider{}, "test-model", m.dir, tx, 0, 1)
+	engine := NewEngine(m)
+	engine.UseLLM(nil, worker)
+	fixture := &interventionProtocolFixture{m: m, task: task, engine: engine, worker: worker, tx: tx, intentID: intentID}
+	t.Cleanup(func() {
+		taskID, _ := strconv.ParseInt(task.ID, 10, 64)
+		_ = m.pg.DeleteTask(taskID)
+		_ = m.Close()
+	})
+	return fixture
+}
+
+// settleRegisteredIntervention emulates the ordering inside runWorkerStep after
+// a cancellation-aware provider returns: detach the live work handle, commit the
+// running->paused final state, install the directed hand-off, then release the
+// controller waiting on workExecution.done.
+func settleRegisteredIntervention(t *testing.T, f *interventionProtocolFixture) string {
+	t.Helper()
+	action, complete := f.engine.detachWork(f.intentID)
+	if action != "intervene" {
+		t.Fatalf("settled action=%q, want intervene", action)
+	}
+	if changed, err := f.task.Store.CompareAndSetIntentState(f.intentID, "running", "paused"); err != nil || !changed {
+		t.Fatalf("settle running->paused: changed=%v err=%v", changed, err)
+	}
+	f.engine.queueDirectedClaim(f.task.ID, f.intentID)
+	complete(nil)
+	return action
+}
+
+func countTranscriptWorkerMessages(t *testing.T, tx *transcript.Store, sessionID, requestID string) (int, []transcript.Record) {
+	t.Helper()
+	records, err := tx.Load(tx.MainPath(sessionID))
+	if err != nil {
+		t.Fatal(err)
+	}
+	marker := "ARTEX_WORKER_CHAT:" + requestID
+	count := 0
+	for _, record := range records {
+		if record.Message != nil && strings.Contains(record.Message.Text(), marker) {
+			count++
+		}
+	}
+	return count, records
+}
+
+func TestWorkerMessageCancelsBeforeAcceptAndContinuesSameSessionFirst(t *testing.T) {
+	f := newInterventionProtocolFixture(t, 1)
+	competitorID, err := f.task.Store.AddIntent(map[string]any{"summary": "higher priority competitor"}, 100, nil, "planner")
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	// Pre-seed the same stable session so the message must continue its UUID
+	// chain rather than creating a worker-slot-specific conversation.
+	sessionID := agent.WorkerSessionID(f.task.ExpID, f.intentID)
+	seed := f.tx.NewWriter(sessionID, "")
+	seed.RecordMessage(llm.UserText("original intent turn"), llm.Usage{})
+	if err := seed.Err(); err != nil {
+		t.Fatal(err)
+	}
+	before, err := f.tx.Load(f.tx.MainPath(sessionID))
+	if err != nil || len(before) != 1 {
+		t.Fatalf("seed transcript records=%d err=%v", len(before), err)
+	}
+
+	workCtx, cancelWork := context.WithCancelCause(context.Background())
+	f.engine.registerWork(f.intentID, cancelWork)
+	live, unsubscribe := f.engine.Broadcaster().Subscribe(f.task.ID)
+	defer unsubscribe()
+
+	const requestID = "worker-message-engine-1"
+	const message = "停止主动枚举，立即验证登录越权"
+	type outcome struct {
+		result WorkInterventionResult
+		err    error
+	}
+	done := make(chan outcome, 1)
+	go func() {
+		result, err := f.engine.InterveneWork(context.Background(), f.task, f.intentID, requestID, message)
+		done <- outcome{result: result, err: err}
+	}()
+
+	select {
+	case <-workCtx.Done():
+		if code, _, _, _ := agent.AbortReason(workCtx); code != "work_intervened_by_user" {
+			t.Fatalf("old execution cancel code=%q, want work_intervened_by_user", code)
+		}
+	case <-time.After(time.Second):
+		t.Fatal("Worker message did not immediately cancel the old execution")
+	}
+	select {
+	case early := <-done:
+		t.Fatalf("Worker message accepted before old worker settled: %+v", early)
+	case <-time.After(25 * time.Millisecond):
+	}
+	pending, err := f.task.Store.IntentInterventionByRequest(f.intentID, requestID)
+	if err != nil || pending == nil || pending.Status != db.IntentInterventionPending {
+		t.Fatalf("durable pending reservation=%+v err=%v", pending, err)
+	}
+	settleRegisteredIntervention(t, f)
+
+	var accepted outcome
+	select {
+	case accepted = <-done:
+	case <-time.After(2 * time.Second):
+		t.Fatal("timed out waiting for accepted Worker message")
+	}
+	if accepted.err != nil || accepted.result.ActivityID <= pending.ActivityID || accepted.result.Duplicate || accepted.result.State != "open" {
+		t.Fatalf("accepted Worker message=%+v control=%+v err=%v", accepted.result, pending, accepted.err)
+	}
+	node, err := f.task.Store.GetNode(f.intentID)
+	if err != nil || node == nil || node.State != "open" {
+		t.Fatalf("reopened target=%+v err=%v", node, err)
+	}
+
+	// The API publishes the visible user activity but never edits transcript files.
+	// agentcore appends the standard user turn only after the same Worker is claimed.
+	count, records := countTranscriptWorkerMessages(t, f.tx, sessionID, requestID)
+	if count != 0 || len(records) != 1 {
+		t.Fatalf("HTTP handler edited Worker transcript: count=%d records=%+v", count, records)
+	}
+	select {
+	case activity := <-live:
+		if activity.ID != accepted.result.ActivityID || activity.NodeID == nil || *activity.NodeID != f.intentID ||
+			activity.Worker != "user" || activity.Kind != "user" || activity.Summary != message {
+			t.Fatalf("broadcast activity=%+v", activity)
+		}
+	case <-time.After(time.Second):
+		t.Fatal("accepted Worker message was not broadcast")
+	}
+	select {
+	case duplicate := <-live:
+		t.Fatalf("accepted Worker message broadcast twice: %+v", duplicate)
+	default:
+	}
+
+	// The directed hand-off must outrank the normal frontier without changing the
+	// persisted priority. This catches the one-worker race where the released slot
+	// used to take an unrelated, higher-priority intent before the target reopened.
+	claimed := f.engine.claimNext(f.task, "work#2")
+	if claimed == nil || claimed.ID != f.intentID {
+		t.Fatalf("directed claim=%+v, want target %d ahead of competitor %d", claimed, f.intentID, competitorID)
+	}
+	competitor, err := f.task.Store.GetNode(competitorID)
+	if err != nil || competitor == nil || competitor.State != "open" {
+		t.Fatalf("competitor was claimed before Worker message target: node=%+v err=%v", competitor, err)
+	}
+	request, nextMessage, hasMessage, err := f.task.Store.PendingIntentInterventionMessage(f.intentID)
+	if err != nil || !hasMessage || request != requestID || nextMessage != message {
+		t.Fatalf("pending Worker message request=%q message=%q present=%v err=%v", request, nextMessage, hasMessage, err)
+	}
+	taskID, _ := strconv.ParseInt(f.task.ID, 10, 64)
+	if _, _, err := f.worker.ExecuteWithMessage(context.Background(), "work#2", taskID, nil, f.task.Store,
+		claimed, nil, nil, nil, nil, request, nextMessage); err != nil {
+		t.Fatalf("continue Worker conversation: %v", err)
+	}
+	if err := f.task.Store.MarkIntentInterventionHandoffClaimed(f.intentID); err != nil {
+		t.Fatal(err)
+	}
+	count, records = countTranscriptWorkerMessages(t, f.tx, sessionID, requestID)
+	if count != 1 || len(records) != 3 || records[1].ParentUUID != records[0].UUID || records[2].ParentUUID != records[1].UUID {
+		t.Fatalf("Worker message did not form one normal chained turn: count=%d records=%+v", count, records)
+	}
+	if records[1].Message == nil || records[1].Message.Role != llm.RoleUser || !strings.Contains(records[1].Message.Text(), message) {
+		t.Fatalf("Worker message was not recorded as the next user turn: %+v", records[1])
+	}
+	if _, _, pendingAfterRun, err := f.task.Store.PendingIntentInterventionMessage(f.intentID); err != nil || pendingAfterRun {
+		t.Fatalf("Worker message handoff remains pending=%v err=%v", pendingAfterRun, err)
+	}
+
+	retry, err := f.engine.InterveneWork(context.Background(), f.task, f.intentID, requestID, message)
+	if err != nil || !retry.Duplicate || retry.ActivityID != accepted.result.ActivityID {
+		t.Fatalf("idempotent engine retry=%+v err=%v", retry, err)
+	}
+	if _, err := f.engine.InterveneWork(context.Background(), f.task, f.intentID, requestID, "同 request_id 的另一条消息"); !errors.Is(err, errWorkInterventionConflict) {
+		t.Fatalf("same request_id different message error=%v, want conflict", err)
+	}
+	count, _ = countTranscriptWorkerMessages(t, f.tx, sessionID, requestID)
+	if count != 1 {
+		t.Fatalf("engine retries appended %d Worker messages, want 1", count)
+	}
+}
+
+func TestSendWorkerMessageContinuesAfterRequestContextCancellation(t *testing.T) {
+	f := newInterventionProtocolFixture(t, 1)
+	workCtx, cancelWork := context.WithCancelCause(context.Background())
+	f.engine.registerWork(f.intentID, cancelWork)
+	s := &Server{m: f.m, engine: f.engine, ctx: context.Background()}
+
+	requestBody, _ := json.Marshal(map[string]string{
+		"message": "中止当前动作，只验证权限边界", "request_id": "worker-message-disconnect-1",
+	})
+	requestCtx, disconnect := context.WithCancel(context.Background())
+	disconnect()
+	req := httptest.NewRequest(http.MethodPost, "/api/tasks/1/intents/1/messages", bytes.NewReader(requestBody)).WithContext(requestCtx)
+	req.SetPathValue("id", f.task.ID)
+	req.SetPathValue("iid", strconv.FormatInt(f.intentID, 10))
+	recorder := httptest.NewRecorder()
+	done := make(chan struct{})
+	go func() {
+		s.sendWorkerMessage(recorder, req)
+		close(done)
+	}()
+
+	select {
+	case <-workCtx.Done():
+	case <-time.After(time.Second):
+		t.Fatal("disconnected request did not continue to cancel the worker")
+	}
+	settleRegisteredIntervention(t, f)
+	select {
+	case <-done:
+	case <-time.After(2 * time.Second):
+		t.Fatal("server-owned Worker message did not finish after browser disconnect")
+	}
+	if recorder.Code != http.StatusOK {
+		t.Fatalf("disconnected Worker message status=%d body=%s", recorder.Code, recorder.Body.String())
+	}
+	accepted, err := f.task.Store.IntentInterventionByRequest(f.intentID, "worker-message-disconnect-1")
+	if err != nil || accepted == nil || accepted.Status != db.IntentInterventionAccepted {
+		t.Fatalf("disconnected request acceptance=%+v err=%v", accepted, err)
+	}
+}
+
+func TestWorkerMessageRejectsUnavailableTaskStatesWithoutReservation(t *testing.T) {
+	cases := []struct {
+		name  string
+		block func(*interventionProtocolFixture)
+	}{
+		{name: "paused", block: func(f *interventionProtocolFixture) {
+			f.task.updateLifecycle(func(state *taskLifecycleState) { state.Paused = true })
+			f.engine.Pause(f.task.ID, agent.AbortPausedByUser)
+		}},
+		{name: "queued", block: func(f *interventionProtocolFixture) {
+			f.task.updateLifecycle(func(state *taskLifecycleState) { state.Queued = true })
+		}},
+		{name: "terminal", block: func(f *interventionProtocolFixture) {
+			f.task.updateLifecycle(func(state *taskLifecycleState) { state.Status = "done" })
+		}},
+		{name: "settling", block: func(f *interventionProtocolFixture) { f.engine.markSettling(f.task.ID) }},
+		{name: "deleting", block: func(f *interventionProtocolFixture) { f.engine.BeginDelete(f.task.ID) }},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			f := newInterventionProtocolFixture(t, 1)
+			tc.block(f)
+			requestID := "worker-message-unavailable-" + tc.name
+			if _, err := f.engine.InterveneWork(context.Background(), f.task, f.intentID, requestID, "must not persist"); !errors.Is(err, errWorkInterventionConflict) {
+				t.Fatalf("unavailable state error=%v, want conflict", err)
+			}
+			reservation, err := f.task.Store.IntentInterventionByRequest(f.intentID, requestID)
+			if err != nil || reservation != nil {
+				t.Fatalf("unavailable state persisted reservation=%+v err=%v", reservation, err)
+			}
+			node, err := f.task.Store.GetNode(f.intentID)
+			if err != nil || node == nil || node.State != "running" {
+				t.Fatalf("unavailable state mutated intent=%+v err=%v", node, err)
+			}
+			count, _ := countTranscriptWorkerMessages(t, f.tx, agent.WorkerSessionID(f.task.ExpID, f.intentID), requestID)
+			if count != 0 {
+				t.Fatalf("unavailable state appended %d transcript interventions", count)
+			}
+		})
+	}
+}
+
+func TestWorkerWakeInterruptsLongIdlePoll(t *testing.T) {
+	task := &Task{workerWake: make(chan struct{}, 1)}
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+	started := make(chan struct{})
+	returned := make(chan bool, 1)
+	go func() {
+		close(started)
+		returned <- waitWorker(ctx, task, 5*time.Second)
+	}()
+	<-started
+	start := time.Now()
+	task.NotifyWorker()
+	select {
+	case stopped := <-returned:
+		if stopped {
+			t.Fatal("worker wake was mistaken for context shutdown")
+		}
+		if elapsed := time.Since(start); elapsed > 250*time.Millisecond {
+			t.Fatalf("worker wake took %s, want immediate wake", elapsed)
+		}
+	case <-time.After(time.Second):
+		t.Fatal("worker wake waited for the idle polling interval")
+	}
+}
+
+func TestSendWorkerMessageRejectsOversizedBodyAsEntityTooLarge(t *testing.T) {
+	f := newInterventionProtocolFixture(t, 1)
+	s := &Server{m: f.m, engine: f.engine, ctx: context.Background()}
+	body := `{"message":"` + strings.Repeat("x", maxWorkerMessageBytes) + `","request_id":"oversized"}`
+	req := httptest.NewRequest(http.MethodPost, "/api/tasks/1/intents/1/messages", strings.NewReader(body))
+	req.SetPathValue("id", f.task.ID)
+	req.SetPathValue("iid", strconv.FormatInt(f.intentID, 10))
+	recorder := httptest.NewRecorder()
+	s.sendWorkerMessage(recorder, req)
+	if recorder.Code != http.StatusRequestEntityTooLarge {
+		t.Fatalf("oversized Worker message status=%d body=%s", recorder.Code, recorder.Body.String())
+	}
+}

server/manager.go

@@ -63,6 +63,7 @@ type Task struct {
 	Store                *pgdb.ExplorationStore `json:"-"`
 	Guard                *guard.Guard           `json:"-"`
 	notify               chan struct{}
+	workerWake           chan struct{}
 	lifecycleMu          sync.RWMutex
 	llmMu                sync.RWMutex
 
@@ -737,7 +738,7 @@ func taskFromPG(pt *pgdb.Task, store *pgdb.ExplorationStore, ic *intercept.Inter
 		TimeoutSeconds: pt.TimeoutSeconds, PlanHeartbeatSeconds: pt.PlanHeartbeatSeconds,
 		CoverageEnabled: pt.CoverageEnabled,
 		FirstRunAt:      unixOrZero(pt.FirstRunAt), DeadlineAt: unixOrZero(pt.DeadlineAt),
-		Store: store, Guard: guard.NewWithInterceptor(ic), notify: make(chan struct{}, 1),
+		Store: store, Guard: guard.NewWithInterceptor(ic), notify: make(chan struct{}, 1), workerWake: make(chan struct{}, 64),
 	}
 }
 
@@ -1638,6 +1639,42 @@ func (t *Task) Notify() {
 	}
 }
 
+// NotifyWorker wakes one idle execution slot without itself scheduling a Planner
+// round. Worker-message acceptance separately records a persistent planning event
+// and calls NotifyWorkerMessage; keeping execution wake-up independent prevents
+// Planner debounce from delaying the directed Worker claim.
+func (t *Task) NotifyWorker() {
+	if t == nil || t.workerWake == nil {
+		return
+	}
+	select {
+	case t.workerWake <- struct{}{}:
+	default:
+	}
+}
+
+// NotifyWorkerMessage immediately schedules a Planner round for an accepted,
+// persistent Worker direction. ActivityID identifies the durable kind=user row
+// and prevents retries in the same debounce window from duplicating the prompt.
+func (t *Task) NotifyWorkerMessage(intentID, activityID int64, message string) {
+	message = strings.TrimSpace(message)
+	if intentID <= 0 || activityID <= 0 || message == "" {
+		return
+	}
+	t.trigMu.Lock()
+	for _, event := range t.pendingTriggers {
+		if event.Kind == agent.TriggerWorkerMessage && event.ActivityID == activityID {
+			t.trigMu.Unlock()
+			return
+		}
+	}
+	t.pendingTriggers = append(t.pendingTriggers, agent.TriggerEvent{
+		Kind: agent.TriggerWorkerMessage, ActivityID: activityID, IntentID: intentID, Detail: message,
+	})
+	t.trigMu.Unlock()
+	t.Notify()
+}
+
 // NotifyDone is Notify plus a hint: a worker just finished intentID and that is
 // what triggered this wake-up. The planner reads the accumulated triggers next
 // round so it can spell out which intent finished (+ its output). Events pile up

server/server.go

@@ -249,6 +249,9 @@ func New(ctx context.Context, m *Manager, skillDir string, dataDir string, keyDi
 		if n, _ := t.Store.ResetRunningIntents(); n > 0 {
 			log.Printf("[engine] task %s 重置 %d 个残留 running 意图为 open", t.ID, n)
 		}
+		if err := s.engine.RecoverWorkerMessages(ctx, t); err != nil {
+			log.Printf("[worker chat] task %s 恢复待处理人工消息失败: %v", t.ID, err)
+		}
 		if lifecycle.Paused {
 			s.engine.Pause(t.ID, agent.AbortPausedOnReload)
 		}
@@ -471,6 +474,11 @@ func (s *Server) applyLLM(cfg agent.Config) error {
 	s.llmOn = true
 	s.cfgMu.Unlock()
 	s.invalidateTaskAgents()
+	for _, task := range s.m.List() {
+		if err := s.engine.RecoverWorkerMessages(s.ctx, task); err != nil {
+			log.Printf("[worker chat] task %s 在 LLM 就绪后恢复待处理人工消息失败: %v", task.ID, err)
+		}
+	}
 
 	// wake the active task so a task created while idle starts exploring.
 	if t := s.m.ActiveTask(); t != nil {
@@ -749,6 +757,7 @@ func (s *Server) Handler() http.Handler {
 	mux.HandleFunc("PUT /api/tasks/{id}/llm", s.updateTaskLLMProfiles)
 	mux.HandleFunc("GET /api/tasks/{id}/llm/resolution", s.taskLLMResolutionHandler)
 	mux.HandleFunc("POST /api/tasks/{id}/intents/{iid}/control", s.controlIntent)
+	mux.HandleFunc("POST /api/tasks/{id}/intents/{iid}/messages", s.sendWorkerMessage)
 	mux.HandleFunc("POST /api/tasks/{id}/intents/{iid}/rerun", s.rerunIntent)    // 重跑单条 blocked/exhausted/stopped 意图
 	mux.HandleFunc("POST /api/tasks/{id}/intents/rerun-blocked", s.rerunBlocked) // 批量重跑本任务全部 blocked 意图
 	mux.HandleFunc("POST /api/active", s.setActive)