重构 worker 对话：复用 resume 机制，改专属 goroutine 运行

把「给运行中 Worker 发消息调方向」从自建的介入持久化协议，改为复用
已有的暂停/恢复 + transcript resume 能力，实现与主 Agent chat 一致。

- 拆 runWorkerStep：抽出 runIntent(执行+善后)，池子与消息路径共用一份
  善后逻辑；新增 runDetachedIntent，把意图 paused→running(不经 frontier)
  并在专属 goroutine 里直接跑，用户消息不受 3 槽池限制、随发随跑。
- 精简 sendWorkerMessage handler：校验 + emit 用户活动 + 专属运行；消息经
  ExecuteWithMessage 作为下一轮输入注入，request_id 仅做 transcript 去重。
- 删除介入 DB 协议与 15 处 activity 查询过滤(回退 exploration.go)、
  directedClaims 调度屏障、RecoverWorkerMessages 恢复、workerWake/waitWorker、
  AbortWorkIntervenedByUser。无 DB schema 改动。
- 前端：保留消息框，恢复「直接继续」(resume)按钮；发送改由 SSE 承载，
  去掉乐观插入。
- 取舍(低频可接受)：消息仅存内存，不做崩溃恢复；发消息时该任务可能瞬时
  超并发 1 个 worker。

净改动 9 文件 +583/-46(原 PR 15 文件 +1940/-72)。

server/engine.go

@@ -75,7 +75,6 @@ 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
@@ -131,17 +130,6 @@ 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
@@ -174,14 +162,7 @@ type taskRuntime struct {
 type workExecution struct {
 	cancel context.CancelCauseFunc
 	done   chan error
-	action string // user action: pause | cancel | intervene
-}
-
-type WorkInterventionResult struct {
-	RequestID  string
-	ActivityID int64
-	State      string
-	Duplicate  bool
+	action string // user action: pause | cancel
 }
 
 // nextPlannerRound returns the next planner round number for a task (1-based).
@@ -304,12 +285,6 @@ 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()
@@ -415,7 +390,7 @@ 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{},
-		directedClaims: map[string][]int64{}, directedReady: map[int64]bool{}, runtimes: map[string]*taskRuntime{}}
+		runtimes: map[string]*taskRuntime{}}
 }
 
 // registerWork records the cancel for the work currently running intentID.
@@ -450,7 +425,7 @@ 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" && action != "intervene" {
+	if action != "pause" && action != "cancel" {
 		return fmt.Errorf("unsupported work action %q", action)
 	}
 	if ctx == nil {
@@ -469,11 +444,8 @@ func (e *Engine) ControlWork(ctx context.Context, intentID int64, action string)
 	run.action = action
 	done := run.done
 	cause := error(agent.AbortWorkPausedByUser)
-	switch action {
-	case "cancel":
+	if action == "cancel" {
 		cause = agent.AbortWorkCancelledByUser
-	case "intervene":
-		cause = agent.AbortWorkIntervenedByUser
 	}
 	run.cancel(cause)
 	e.workMu.Unlock()
@@ -492,282 +464,6 @@ 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)
-	}
-	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:
-			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
@@ -775,11 +471,7 @@ func (e *Engine) recoverWorkerMessages(ctx context.Context, t *Task, allowAdmiss
 func (e *Engine) releaseWorkControl(intentID int64, run *workExecution, action string) {
 	e.workMu.Lock()
 	if current := e.work[intentID]; current == run && current.action == action {
-		if action == "intervene" {
-			current.action = "intervene_no_handoff"
-		} else {
-			current.action = ""
-		}
+		current.action = ""
 	}
 	e.workMu.Unlock()
 }
@@ -918,19 +610,6 @@ 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
@@ -1197,13 +876,13 @@ func (e *Engine) workerLoop(ctx context.Context, t *Task, name string) {
 		}
 		_, worker := e.snapshotFor(t)
 		if worker == nil {
-			if waitWorker(ctx, t, 1500*time.Millisecond) {
+			if sleepCtx(ctx, 1500*time.Millisecond) {
 				return
 			}
 			continue
 		}
 		if e.IsPaused(t.ID) {
-			if waitWorker(ctx, t, 1000*time.Millisecond) {
+			if sleepCtx(ctx, 1000*time.Millisecond) {
 				return
 			}
 			continue // user-paused: don't claim/execute intents
@@ -1212,13 +891,13 @@ func (e *Engine) workerLoop(ctx context.Context, t *Task, name string) {
 			return
 		}
 		if e.isSettling(t.ID) {
-			if waitWorker(ctx, t, 1000*time.Millisecond) {
+			if sleepCtx(ctx, 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 waitWorker(ctx, t, 1000*time.Millisecond) {
+			if sleepCtx(ctx, 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
@@ -1228,21 +907,34 @@ func (e *Engine) workerLoop(ctx context.Context, t *Task, name string) {
 		}
 		claimed := e.runWorkerStep(ctx, t, name, worker)
 		e.decInflight(t.ID)
-		if !claimed && waitWorker(ctx, t, 800*time.Millisecond) {
+		if !claimed && sleepCtx(ctx, 800*time.Millisecond) {
 			return
 		}
 	}
 }
 
-// runWorkerStep claims and fully settles at most one intent. Its caller holds one
-// task-operation admission for the whole claim/execute/state-write sequence so a
-// delete cannot observe quiescence between the LLM return and the final DB writes.
+// runWorkerStep claims one intent from the frontier and fully settles it via
+// runIntent. Returns false when nothing was claimable. The pool worker loop is its
+// only caller.
 func (e *Engine) runWorkerStep(ctx context.Context, t *Task, name string, worker *agent.Worker) bool {
 	intent := e.claimNext(t, name)
 	if intent == nil {
 		return false
 	}
 	log.Printf("[worker %s] task %s 领取意图 #%d", name, t.ID, intent.ID)
+	return e.runIntent(ctx, t, name, worker, intent, "", "")
+}
+
+// runIntent executes and fully settles one already-claimed (state=running) intent.
+// Both the pool worker loop (via runWorkerStep) and the human-message handler (via
+// runDetachedIntent, a dedicated goroutine outside the worker pool) call it, so the
+// execute/retry/state-write logic lives in exactly one place. A non-empty message
+// is injected as this turn's input through ExecuteWithMessage; requestID keys the
+// transcript marker that dedups re-injection across model_error retries. The caller
+// must already hold one task-operation admission for the whole sequence so a delete
+// cannot observe quiescence between the LLM return and the final DB writes.
+func (e *Engine) runIntent(ctx context.Context, t *Task, name string, worker *agent.Worker, intent *db.Node, requestID, message string) bool {
+	hasChatMessage := message != ""
 	e.stampFirstRun(t) // 首次真正执行 → 盖 first_run_at + 算 deadline(仅带 timeout 的任务)
 	e.touch(t.ID)
 	emit := func(r db.Activity) { e.emitActivity(t, r) }
@@ -1267,31 +959,16 @@ 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)
 	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)
+		reason, wrote, err = worker.ExecuteWithMessage(workCtx, name, wTaskID, e.m.assets, t.Store, intent, hooks, emit, e.m.enrich, t.NotifyFinding, requestID, message)
 	} 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)。
@@ -1305,7 +982,7 @@ func (e *Engine) runWorkerStep(ctx context.Context, t *Task, name string, worker
 		}
 		e.BeginLLMCall(t.ID)
 		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)
+			reason, wrote, err = worker.ExecuteWithMessage(workCtx, name, wTaskID, e.m.assets, t.Store, intent, hooks, emit, e.m.enrich, t.NotifyFinding, requestID, message)
 		} else {
 			reason, wrote, err = worker.Execute(workCtx, name, wTaskID, e.m.assets, t.Store, intent, hooks, emit, e.m.enrich, t.NotifyFinding)
 		}
@@ -1326,8 +1003,6 @@ 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"
 		}
@@ -1344,22 +1019,6 @@ 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
@@ -1441,6 +1100,61 @@ func (e *Engine) runWorkerStep(ctx context.Context, t *Task, name string, worker
 	return true
 }
 
+// runDetachedIntent runs one paused intent OUTSIDE the worker pool in its own
+// goroutine — the human-message path. It transitions the intent paused->running
+// itself (never through 'open'), so the pool, which only claims 'open', can never
+// race it; the "at most one run per intent" invariant still holds because winning
+// the CAS is the sole entry and work[intentID] was cleared when the pause settled.
+// Because it does not compete for a frontier slot, a user message continues the
+// worker immediately even when all pool slots are busy (mirroring how the
+// main-agent chat handler starts its run directly). The spawned goroutine owns one
+// task-operation admission for the whole run and roots its context at ctx (pass the
+// server root, never the HTTP request, so a disconnect cannot strand the run while
+// task pause/delete/shutdown still stops it). Returns an error if the run could not
+// be started; the intent is left untouched in that case.
+func (e *Engine) runDetachedIntent(ctx context.Context, t *Task, intentID int64, requestID, message string) error {
+	if !e.beginTaskOperation(t.ID) {
+		return fmt.Errorf("task is being deleted")
+	}
+	release := true
+	defer func() {
+		if release {
+			e.decInflight(t.ID)
+		}
+	}()
+	_, worker := e.snapshotFor(t)
+	if worker == nil {
+		return fmt.Errorf("worker 尚未就绪")
+	}
+	node, err := t.Store.GetNode(intentID)
+	if err != nil {
+		return err
+	}
+	if node == nil || node.Kind != db.KindIntent {
+		return fmt.Errorf("intent not found")
+	}
+	changed, err := t.Store.CompareAndSetIntentState(intentID, "paused", "running")
+	if err != nil {
+		return err
+	}
+	if !changed {
+		return fmt.Errorf("%w: 意图不再是 paused 状态", db.ErrIntentStateConflict)
+	}
+	node.State, node.Owner = "running", "chat"
+	// Record the human turn as a visible activity BEFORE the run starts, so it is
+	// ordered ahead of any worker step and never appears without the run happening.
+	// This is the UI copy; ExecuteWithMessage separately writes it into the intent
+	// transcript as the LLM input.
+	uid := intentID
+	e.emitActivity(t, db.Activity{NodeID: &uid, Worker: "user", Kind: "user", Summary: message, Detail: message})
+	release = false // ownership of the admission passes to the goroutine
+	go func() {
+		defer e.decInflight(t.ID)
+		e.runIntent(ctx, t, "chat", worker, node, requestID, message)
+	}()
+	return nil
+}
+
 func taskExecutionPaused(cause error) bool {
 	var abort *agent.AbortCause
 	if !errors.As(cause, &abort) {
@@ -1464,114 +1178,7 @@ 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

@@ -45,7 +45,6 @@ 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

@@ -225,13 +225,6 @@ 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

@@ -1,13 +1,11 @@
 package server
 
 import (
-	"context"
 	"encoding/json"
 	"errors"
 	"net/http"
 	"strconv"
 	"strings"
-	"time"
 
 	"github.com/Autumn-27/artex/db"
 )
@@ -28,9 +26,15 @@ func validWorkerMessageRequestID(id string) bool {
 	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.
+// sendWorkerMessage continues a paused Worker intent with a human-authored message.
+// The message is injected as the next turn's input through the same
+// resume-from-transcript path the worker uses on a normal resume (ExecuteWithMessage),
+// so this reuses the existing pause/resume machinery rather than a bespoke protocol.
+// The run happens in a dedicated goroutine outside the worker pool (runDetachedIntent),
+// so the message is picked up immediately even when every pool slot is busy — the same
+// way the main-agent chat handler starts its run directly. In-memory only: a process
+// restart re-runs the intent from its transcript without the message, which is
+// acceptable for this rare interrupt-then-continue action.
 func (s *Server) sendWorkerMessage(w http.ResponseWriter, r *http.Request) {
 	t, ok := s.m.Task(r.PathValue("id"))
 	if !ok {
@@ -71,11 +75,28 @@ func (s *Server) sendWorkerMessage(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 
-	if !s.engine.beginTaskOperation(t.ID) {
+	// Reject non-runnable task lifecycles up front so the caller gets a clear reason
+	// instead of a silent no-op. The intent itself must be paused: the UI flow is
+	// interrupt (pause) first, then send.
+	if s.engine.IsDeleting(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)
+	lifecycle := t.lifecycleSnapshot()
+	switch {
+	case lifecycle.Paused || s.engine.IsPaused(t.ID):
+		writeErr(w, http.StatusConflict, "任务已暂停，请先恢复任务再向 Worker 发送消息")
+		return
+	case lifecycle.Queued:
+		writeErr(w, http.StatusConflict, "排队中的任务无法向 Worker 发送消息")
+		return
+	case isTerminalStatus(lifecycle.Status):
+		writeErr(w, http.StatusConflict, "终态任务无法向 Worker 发送消息")
+		return
+	case s.engine.isSettling(t.ID):
+		writeErr(w, http.StatusConflict, "任务正在收尾，无法向 Worker 发送消息")
+		return
+	}
 
 	node, err := t.Store.GetNode(iid)
 	if err != nil {
@@ -94,19 +115,17 @@ func (s *Server) sendWorkerMessage(w http.ResponseWriter, r *http.Request) {
 		writeErr(w, http.StatusConflict, "node is not an intent")
 		return
 	}
-	base := s.ctx
-	if base == nil {
-		base = context.Background()
+	if node.State != "paused" {
+		writeErr(w, http.StatusConflict, "仅已暂停的 Worker 可以发送消息，请先暂停")
+		return
 	}
-	opCtx, cancel := context.WithTimeout(base, workControlWaitTimeout+15*time.Second)
-	defer cancel()
-	result, err := s.engine.InterveneWork(opCtx, t, iid, requestID, message)
-	if err != nil {
+
+	// runDetachedIntent transitions paused->running, emits the user turn and starts a
+	// dedicated run. Root the run at s.ctx so a disconnected browser cannot strand it
+	// while task pause/delete/shutdown still stop it.
+	if err := s.engine.runDetachedIntent(s.ctx, t, iid, requestID, message); 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):
+		case errors.Is(err, db.ErrIntentStateConflict):
 			writeErr(w, http.StatusConflict, err.Error())
 		default:
 			writeErr(w, http.StatusInternalServerError, err.Error())
@@ -115,11 +134,9 @@ func (s *Server) sendWorkerMessage(w http.ResponseWriter, r *http.Request) {
 	}
 
 	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,
+		"id":         iid,
+		"state":      "running",
+		"accepted":   true,
+		"request_id": requestID,
 	})
 }

server/intent_intervention_protocol_test.go

@@ -1,378 +0,0 @@
-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

@@ -63,7 +63,6 @@ type Task struct {
 	Store                *pgdb.ExplorationStore `json:"-"`
 	Guard                *guard.Guard           `json:"-"`
 	notify               chan struct{}
-	workerWake           chan struct{}
 	lifecycleMu          sync.RWMutex
 	llmMu                sync.RWMutex
 
@@ -738,7 +737,7 @@ 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), workerWake: make(chan struct{}, 64),
+		Store: store, Guard: guard.NewWithInterceptor(ic), notify: make(chan struct{}, 1),
 	}
 }
 
@@ -1639,20 +1638,6 @@ func (t *Task) Notify() {
 	}
 }
 
-// NotifyWorker wakes one idle execution slot without also scheduling a Planner
-// round. The channel is separate from notify because human intervention needs
-// the reopened intent reclaimed immediately, while graph-change notifications
-// are Planner-facing and debounced. A nil channel is tolerated by test fixtures.
-func (t *Task) NotifyWorker() {
-	if t == nil || t.workerWake == nil {
-		return
-	}
-	select {
-	case t.workerWake <- struct{}{}:
-	default:
-	}
-}
-
 // 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

@@ -249,9 +249,6 @@ 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)
 		}
@@ -474,11 +471,6 @@ 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 {

server/worker_message_test.go

@@ -0,0 +1,54 @@
+package server
+
+import (
+	"context"
+	"errors"
+	"strings"
+	"testing"
+)
+
+// After the worker-message rework, the human-message path no longer goes through
+// ControlWork with an "intervene" action — it runs the intent in a dedicated
+// goroutine (runDetachedIntent) and pauses, if needed, with the ordinary "pause"
+// action. Guard that ControlWork only knows pause|cancel so a reintroduced
+// "intervene" caller fails loudly instead of silently no-op'ing.
+func TestControlWorkRejectsRemovedInterveneAction(t *testing.T) {
+	e := NewEngine(nil)
+	err := e.ControlWork(context.Background(), 1, "intervene")
+	if err == nil || !containsUnsupported(err) {
+		t.Fatalf("intervene should be unsupported, got %v", err)
+	}
+}
+
+// A control request for an intent with no live work is a conflict, not a panic or
+// a silent success — the same guard the message handler relies on when a run has
+// already finished between the UI read and the request.
+func TestControlWorkConflictsWhenNoWorkRegistered(t *testing.T) {
+	e := NewEngine(nil)
+	err := e.ControlWork(context.Background(), 42, "pause")
+	if !errors.Is(err, errWorkControlConflict) {
+		t.Fatalf("pause with no work = %v, want errWorkControlConflict", err)
+	}
+}
+
+// validWorkerMessageRequestID gates the idempotency token; keep its charset stable
+// so a client-generated UUID is always accepted and injection-y ids are rejected.
+func TestValidWorkerMessageRequestID(t *testing.T) {
+	ok := []string{"worker-message-abc123", "a.b:c_d-1", "550e8400-e29b-41d4-a716-446655440000"}
+	for _, id := range ok {
+		if !validWorkerMessageRequestID(id) {
+			t.Errorf("id %q should be valid", id)
+		}
+	}
+	bad := []string{"", "has space", "emoji😀", "slash/here", string(make([]byte, 129))}
+	for _, id := range bad {
+		if validWorkerMessageRequestID(id) {
+			t.Errorf("id %q should be rejected", id)
+		}
+	}
+}
+
+func containsUnsupported(err error) bool {
+	return err != nil && !errors.Is(err, errWorkControlConflict) &&
+		strings.Contains(err.Error(), "unsupported work action")
+}