fix(agent): 插话 - 独立运行时路径的送达与退回 (#673)

复现：Agent 运行中发送插话，独立 Agent 运行时（Pi）下既不送达模型也不退回用户，插话一直停在待送队列里。原因与修法：插话送达点原本只有旧 Go 执行循环（advanceCloudAgent 开步前出队）；Pi 运行时接手执行后调度循环不再推进。现把 Pi 路径的送达搬到 /model 模型步：出队、规范历史追加与 delivered 事件同在入队的一次 CAS 写入里，响应携带 steeringMessages 交给运行时 session.steer。完成、失败、取消、审批拒绝与步数耗尽统一退回；最终回答之后的插话在预算富余时续跑一步。完成闸一次带走多条插话时，合并后的续跑提示词也记入指纹，避免被恢复闸当成遗留提示词清空。
2026-10-05 13:39:57 +08:00
parent dd563d3373
commit 5bf838ebbe
6 changed files with 730 additions and 6 deletions
--- a/backend/internal/app/cloud_agent_interjection_pi_test.go
+++ b/backend/internal/app/cloud_agent_interjection_pi_test.go
@@ -0,0 +1,497 @@
+package app
+
+// Pi 运行时路径的插话送达与退回测试。见 cloud_agent_interjection_pi.go。
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"strings"
+	"testing"
+	"time"
+
+	"gorm.io/gorm"
+	"infinite-canvas/backend/internal/model"
+	"infinite-canvas/backend/internal/repository"
+)
+
+// interjectionPiRoot 建立 Pi 模式的运行（root 任务不占活动任务位，插话送达发生在 /model 桥）。
+func interjectionPiRoot(t *testing.T) (*Service, *gorm.DB, *CloudAgentRun) {
+	t.Helper()
+	s, db, _, _ := creationTestService(t)
+	s.legacyCloudAgentRootTask = false
+	if err := db.Create(&model.CanvasProject{ID: "agent-canvas", UserID: "user", PayloadJSON: `{"nodes":[]}`}).Error; err != nil {
+		t.Fatal(err)
+	}
+	root, err := s.CreateCloudAgentRun("user", agentTestRequest(), "")
+	if err != nil {
+		t.Fatal(err)
+	}
+	return s, db, root
+}
+
+// interjectionApprovalRun 把运行摆进等待审批状态：挂起的工具调用与审批一一对应
+// （检查点校验要求 Approval.Call 与 Calls[CallIndex] 一致）。
+func interjectionApprovalRun(t *testing.T, s *Service, db *gorm.DB, run *model.CloudAgentExecution, state *cloudAgentRuntime) {
+	t.Helper()
+	call := cloudAgentCall{ID: "call-1"}
+	call.Function.Name = "generate_media"
+	call.Function.Arguments = "{}"
+	run.Status = "waiting_approval"
+	state.Calls = []cloudAgentCall{call}
+	state.CallIndex = 0
+	state.Approval = &cloudAgentApproval{ID: "ap-1", Call: call}
+	if state.Decisions == nil {
+		state.Decisions = map[string]string{}
+	}
+	interjectionSaveState(t, s, db, run, state)
+}
+
+// interjectionModelPayload 组装运行时风格的 /model 请求（conversation）。
+func interjectionModelPayload(t *testing.T, messages []map[string]any) map[string]json.RawMessage {
+	t.Helper()
+	encoded, err := json.Marshal(messages)
+	if err != nil {
+		t.Fatal(err)
+	}
+	return map[string]json.RawMessage{
+		"modelId":       json.RawMessage(`"text-test"`),
+		"purpose":       json.RawMessage(`"conversation"`),
+		"messages":      encoded,
+		"thinkingLevel": json.RawMessage(`"off"`),
+	}
+}
+
+// interjectionModelStep 桥接一次模型步：等待步任务入队（跳过建轮时的活动任务）、
+// 把任务改成给定结果、等待桥返回。
+func interjectionModelStep(t *testing.T, s *Service, db *gorm.DB, userID, runID string, messages []map[string]any, resultJSON string) map[string]any {
+	t.Helper()
+	stored, err := s.repo.CloudAgent(userID, runID)
+	if err != nil {
+		t.Fatal(err)
+	}
+	initialState, err := cloudAgentDecode(stored)
+	if err != nil {
+		t.Fatal(err)
+	}
+	ctx, cancel := context.WithTimeout(context.Background(), 20*time.Second)
+	defer cancel()
+	type outcome struct {
+		result any
+		err    error
+	}
+	done := make(chan outcome, 1)
+	go func() {
+		result, stepErr := s.cloudAgentPiModel(ctx, userID, runID, interjectionModelPayload(t, messages))
+		done <- outcome{result: result, err: stepErr}
+	}()
+	var taskID string
+	deadline := time.Now().Add(10 * time.Second)
+	for time.Now().Before(deadline) {
+		stored, readErr := s.repo.CloudAgent(userID, runID)
+		if readErr != nil {
+			t.Fatal(readErr)
+		}
+		state, decodeErr := cloudAgentDecode(stored)
+		if decodeErr != nil {
+			t.Fatal(decodeErr)
+		}
+		if state.ActiveTaskID != "" && state.ActiveTaskID != initialState.ActiveTaskID {
+			taskID = state.ActiveTaskID
+			break
+		}
+		select {
+		case got := <-done:
+			t.Fatalf("model bridge returned before scheduling: result=%v err=%v", got.result, got.err)
+		case <-time.After(25 * time.Millisecond):
+		}
+	}
+	if taskID == "" {
+		t.Fatal("model step was never scheduled")
+	}
+	if err := db.Model(&model.Task{}).Where("id = ?", taskID).Updates(map[string]any{
+		"status": model.TaskStatusSucceeded, "result_json": resultJSON,
+	}).Error; err != nil {
+		t.Fatal(err)
+	}
+	select {
+	case got := <-done:
+		if got.err != nil {
+			t.Fatalf("model bridge failed: %v", got.err)
+		}
+		result, _ := got.result.(map[string]any)
+		if result == nil {
+			t.Fatalf("unexpected bridge result: %#v", got.result)
+		}
+		return result
+	case <-ctx.Done():
+		t.Fatal("model bridge did not return")
+	}
+	return nil
+}
+
+// interjectionSteerTexts 取响应里的 steer 正文（Go 内直接组装，是 []string）。
+func interjectionSteerTexts(result map[string]any) []string {
+	texts, _ := result["steeringMessages"].([]string)
+	return texts
+}
+
+// interjectionSaveState 用生产同款 CAS 路径落库改过的运行态（直接 db.Save 会把
+// Transcript 关联行追加而不是替换，污染下一次 decode 的 canonical）。
+func interjectionSaveState(t *testing.T, s *Service, db *gorm.DB, run *model.CloudAgentExecution, state *cloudAgentRuntime) {
+	t.Helper()
+	if err := s.repo.MutateCloudAgent(run.UserID, run.ID, run.Revision, func(current *model.CloudAgentExecution, _ *repository.Repository) error {
+		current.Status = run.Status
+		return cloudAgentSave(current, state)
+	}); err != nil {
+		t.Fatal(err)
+	}
+}
+
+// 运行中插话在下一次 /model 出队：进入本次请求的 canonical（带来源标记）、
+// 发 delivered 事件、响应携带 steeringMessages 交给运行时 session.steer。
+func TestCloudAgentPiInterjectionReachesModelRequest(t *testing.T) {
+	s, db, root := interjectionPiRoot(t)
+	if _, err := s.InterjectCloudAgent("user", root.ID, "msg-1", "改成 16:9"); err != nil {
+		t.Fatal(err)
+	}
+	messages := []map[string]any{{"role": "user", "content": []any{map[string]any{"type": "text", "text": "继续"}}}}
+	result := interjectionModelStep(t, s, db, "user", root.ID, messages, `{"text":"好的","toolCalls":[]}`)
+
+	steered := interjectionSteerTexts(result)
+	if len(steered) != 1 || steered[0] != "【用户插话】改成 16:9" {
+		t.Fatalf("响应没有携带 steer 正文: %#v", result["steeringMessages"])
+	}
+	_, state := agentInterjectionState(t, s, root.ID)
+	if len(state.PendingInterjections) != 0 {
+		t.Fatalf("插话已送达却没有出队: %+v", state.PendingInterjections)
+	}
+	if !agentHasEvent(state, "user_interjection_delivered") {
+		t.Fatal("送达时没有发 user_interjection_delivered 事件")
+	}
+	if len(state.InterjectionIDs) == 0 || state.InterjectionIDs[0] != "msg-1" {
+		t.Fatalf("送达没有记住插话 id: %#v", state.InterjectionIDs)
+	}
+	// 最关键的一条：插话必须出现在真正发给模型的输入里，并带来源标记。
+	var task model.Task
+	last := state.TaskIDs[len(state.TaskIDs)-1]
+	if err := db.First(&task, "id = ?", last).Error; err != nil {
+		t.Fatal(err)
+	}
+	var input struct {
+		AgentRequests struct {
+			Canonical struct {
+				Messages []map[string]any `json:"messages"`
+			} `json:"canonical"`
+		} `json:"agentRequests"`
+	}
+	if err := json.Unmarshal([]byte(task.InputJSON), &input); err != nil {
+		t.Fatal(err)
+	}
+	delivered := false
+	for _, message := range input.AgentRequests.Canonical.Messages {
+		content, _ := message["content"].(string)
+		if message["role"] == "user" && strings.Contains(content, "改成 16:9") {
+			delivered = true
+			if message[cloudAgentContextSourceKey] != "user_interjection" {
+				t.Fatalf("送达进 canonical 的插话缺少来源标记: %#v", message)
+			}
+		}
+	}
+	if !delivered {
+		t.Fatalf("插话没有进入模型请求的输入: %s", task.InputJSON)
+	}
+}
+
+// 等审批期间入队的插话，在恢复后的第一次 /model 送达（模型这一步就能看到，
+// 不用等 steer 在下一轮生效）。
+func TestCloudAgentPiInterjectionQueuedDuringApprovalDeliversOnResume(t *testing.T) {
+	s, db, root := interjectionPiRoot(t)
+	run, state := agentInterjectionState(t, s, root.ID)
+	interjectionApprovalRun(t, s, db, run, &state)
+	if _, err := s.InterjectCloudAgent("user", root.ID, "msg-approval", "批准后按新要求来"); err != nil {
+		t.Fatal(err)
+	}
+	// 审批批准、被暂停的工具执行完毕后的状态：审批已清、回到 running。
+	// （真实链路里 advanceCloudAgentTool 执行工具时清理审批，之后才恢复会话。）
+	run, state = agentInterjectionState(t, s, root.ID)
+	run.Status = "running"
+	state.Approval = nil
+	state.Calls = nil
+	state.CallIndex = 0
+	interjectionSaveState(t, s, db, run, &state)
+	messages := []map[string]any{{"role": "user", "content": []any{map[string]any{"type": "text", "text": "继续"}}}}
+	interjectionModelStep(t, s, db, "user", root.ID, messages, `{"text":"好的","toolCalls":[]}`)
+	_, state = agentInterjectionState(t, s, root.ID)
+	if len(state.PendingInterjections) != 0 {
+		t.Fatal("恢复后的第一步没有送达等审批时的插话")
+	}
+	if !agentHasEvent(state, "user_interjection_delivered") {
+		t.Fatal("恢复送达没有发 user_interjection_delivered 事件")
+	}
+}
+
+// 已送达插话经过运行时投影回写后来源标记被洗掉：下一步模型步按指纹补回，
+// 且插话在两份历史之间不重复（长度相等时采用运行时投影，插话只出现一次）。
+func TestCloudAgentPiInterjectionMarkerSurvivesRuntimeProjection(t *testing.T) {
+	s, db, root := interjectionPiRoot(t)
+	if _, err := s.InterjectCloudAgent("user", root.ID, "msg-1", "改成 16:9"); err != nil {
+		t.Fatal(err)
+	}
+	first := []map[string]any{{"role": "user", "content": []any{map[string]any{"type": "text", "text": "继续"}}}}
+	interjectionModelStep(t, s, db, "user", root.ID, first, `{"text":"好的","toolCalls":[]}`)
+	_, state := agentInterjectionState(t, s, root.ID)
+	if len(state.Canonical.Messages) == 0 || stringField(state.Canonical.Messages[len(state.Canonical.Messages)-1], "role") != "assistant" {
+		t.Fatalf("这一步的助手回复没有落检查点: %+v", state.Canonical.Messages)
+	}
+	// 下一拍的运行时会话：插话（steer 落盘）排在助手回复之后，且不带标记。
+	second := []map[string]any{
+		{"role": "user", "content": []any{map[string]any{"type": "text", "text": "继续"}}},
+		{"role": "assistant", "content": []any{map[string]any{"type": "text", "text": "好的"}}},
+		{"role": "user", "content": []any{map[string]any{"type": "text", "text": "【用户插话】改成 16:9"}}},
+	}
+	result := interjectionModelStep(t, s, db, "user", root.ID, second, `{"text":"收到","toolCalls":[]}`)
+	if len(interjectionSteerTexts(result)) != 0 {
+		t.Fatalf("队列已空却再次返回 steer 正文: %#v", result["steeringMessages"])
+	}
+	_, state = agentInterjectionState(t, s, root.ID)
+	var task model.Task
+	last := state.TaskIDs[len(state.TaskIDs)-1]
+	if err := db.First(&task, "id = ?", last).Error; err != nil {
+		t.Fatal(err)
+	}
+	if got := strings.Count(task.InputJSON, "改成 16:9"); got != 1 {
+		t.Fatalf("插话在两份历史之间重复或丢失: count=%d input=%s", got, task.InputJSON)
+	}
+	marked := false
+	for _, message := range state.Canonical.Messages {
+		content, _ := message["content"].(string)
+		if strings.Contains(content, "改成 16:9") && message[cloudAgentContextSourceKey] == "user_interjection" {
+			marked = true
+		}
+	}
+	if !marked {
+		t.Fatalf("运行时投影回写后插话标记没有被补回: %+v", state.Canonical.Messages)
+	}
+}
+
+// 运行自然结束但队列里还有插话：预算够就作为续跑提示词再走一步（不标记完成）。
+func TestCloudAgentPiInterjectionResumesAfterFinalAnswer(t *testing.T) {
+	s, db, root := interjectionPiRoot(t)
+	run, state := agentInterjectionState(t, s, root.ID)
+	run.Status = "running"
+	state.Request.Budget.MaxSteps = 3
+	state.Step = 1
+	state.PiAssistantResponses = 1
+	state.Canonical.Messages = []map[string]any{
+		{"role": "user", "content": "分析剧情"},
+		{"role": "assistant", "content": "好的"},
+	}
+	interjectionSaveState(t, s, db, run, &state)
+	if _, err := s.InterjectCloudAgent("user", root.ID, "msg-late", "补一句：加字幕"); err != nil {
+		t.Fatal(err)
+	}
+	if err := s.completeCloudAgentPiRun("user", root.ID); err != nil {
+		t.Fatal(err)
+	}
+	run, state = agentInterjectionState(t, s, root.ID)
+	if run.Status != "running" {
+		t.Fatalf("还有未送达插话且预算富余时不应完成，实际 status=%s", run.Status)
+	}
+	if len(state.PendingInterjections) != 0 {
+		t.Fatal("完成闸没有出队插话")
+	}
+	if state.PiResumePrompt != "【用户插话】补一句：加字幕" {
+		t.Fatalf("插话没有被写进续跑提示词: %q", state.PiResumePrompt)
+	}
+	if !agentHasEvent(state, "user_interjection_delivered") {
+		t.Fatal("完成闸送达没有发 user_interjection_delivered 事件")
+	}
+	// 续跑提示词必须被识别为插话投递：恢复会话遇到定格快照时照常跑出去。
+	if !cloudAgentPiResumePromptDeliversInterjection(&state) {
+		t.Fatal("恢复闸没有识别插话续跑提示词")
+	}
+}
+
+// 步数已尽时插话退回、运行照常完成。
+func TestCloudAgentPiInterjectionDroppedWhenStepBudgetExhausted(t *testing.T) {
+	s, db, root := interjectionPiRoot(t)
+	run, state := agentInterjectionState(t, s, root.ID)
+	run.Status = "running"
+	state.Request.Budget.MaxSteps = 1
+	state.Step = 1
+	state.PiAssistantResponses = 1
+	state.Canonical.Messages = []map[string]any{
+		{"role": "user", "content": "分析剧情"},
+		{"role": "assistant", "content": "好的"},
+	}
+	interjectionSaveState(t, s, db, run, &state)
+	if _, err := s.InterjectCloudAgent("user", root.ID, "msg-late", "补一句：加字幕"); err != nil {
+		t.Fatal(err)
+	}
+	if err := s.completeCloudAgentPiRun("user", root.ID); err != nil {
+		t.Fatal(err)
+	}
+	run, state = agentInterjectionState(t, s, root.ID)
+	if run.Status != "completed" {
+		t.Fatalf("步数已尽时应完成收尾，实际 status=%s", run.Status)
+	}
+	if len(state.PendingInterjections) != 0 {
+		t.Fatal("步数已尽时插话还挂在队列里")
+	}
+	if !agentHasEvent(state, "user_interjection_dropped") {
+		t.Fatal("步数已尽时没有发 user_interjection_dropped 事件")
+	}
+}
+
+// 插话 steer 续步撞上步数上限：模型已带着插话给出最终回答，收尾为 completed
+// 并退回剩余插话，而不是把整轮判失败；停在工具结果上的普通步数耗尽不受影响。
+func TestCloudAgentPiSteeredContinuationCompletesAtFinalAnswer(t *testing.T) {
+	s, db, root := interjectionPiRoot(t)
+	run, state := agentInterjectionState(t, s, root.ID)
+	run.Status = "running"
+	state.Request.Budget.MaxSteps = 1
+	state.Step = 1
+	state.PiAssistantResponses = 1
+	state.Canonical.Messages = []map[string]any{
+		{"role": "user", "content": "分析剧情"},
+		{"role": "assistant", "content": "好的"},
+	}
+	interjectionSaveState(t, s, db, run, &state)
+	if _, err := s.InterjectCloudAgent("user", root.ID, "msg-late", "补一句"); err != nil {
+		t.Fatal(err)
+	}
+	s.settleCloudAgentPiFinalAnswer("user", root.ID)
+	run, state = agentInterjectionState(t, s, root.ID)
+	if run.Status != "completed" {
+		t.Fatalf("最终回答边界上应完成收尾，实际 status=%s", run.Status)
+	}
+	if len(state.PendingInterjections) != 0 || !agentHasEvent(state, "user_interjection_dropped") {
+		t.Fatal("收尾时没有退回插话")
+	}
+
+	// 对照：canonical 停在工具结果上（工具批次之后想再开一步）不是最终回答边界，
+	// 保持原有失败路径。
+	run, state = agentInterjectionState(t, s, root.ID)
+	run.Status = "running"
+	state.Canonical.Messages = []map[string]any{
+		{"role": "user", "content": "分析剧情"},
+		{"role": "assistant", "content": "画一张", "tool_calls": []any{map[string]any{"id": "call-1"}}},
+		{"role": "tool", "tool_call_id": "call-1", "content": "{}"},
+	}
+	interjectionSaveState(t, s, db, run, &state)
+	if _, err := s.InterjectCloudAgent("user", root.ID, "msg-late-2", "补一句"); err != nil {
+		t.Fatal(err)
+	}
+	s.settleCloudAgentPiFinalAnswer("user", root.ID)
+	run, _ = agentInterjectionState(t, s, root.ID)
+	if run.Status != "running" {
+		t.Fatalf("工具批次后的步数耗尽不属于最终回答边界，实际 status=%s", run.Status)
+	}
+}
+
+// 运行失败、被取消、审批被拒时，队列里的插话全部退回。
+func TestCloudAgentPiInterjectionsDroppedOnTerminalPaths(t *testing.T) {
+	t.Run("failed", func(t *testing.T) {
+		s, _, root := interjectionPiRoot(t)
+		if _, err := s.InterjectCloudAgent("user", root.ID, "msg-f", "还在吗"); err != nil {
+			t.Fatal(err)
+		}
+		s.failPiRunner(root.ID, "user", errors.New("模型调用失败"))
+		run, state := agentInterjectionState(t, s, root.ID)
+		if run.Status != "failed" || len(state.PendingInterjections) != 0 || !agentHasEvent(state, "user_interjection_dropped") {
+			t.Fatalf("失败收尾没有退回插话: status=%s pending=%+v", run.Status, state.PendingInterjections)
+		}
+	})
+	t.Run("cancelled", func(t *testing.T) {
+		s, _, root := interjectionPiRoot(t)
+		if _, err := s.InterjectCloudAgent("user", root.ID, "msg-c", "还在吗"); err != nil {
+			t.Fatal(err)
+		}
+		if err := s.CancelCloudAgent(context.Background(), "user", root.ID); err != nil {
+			t.Fatal(err)
+		}
+		run, state := agentInterjectionState(t, s, root.ID)
+		if run.Status != "cancelled" || len(state.PendingInterjections) != 0 || !agentHasEvent(state, "user_interjection_dropped") {
+			t.Fatalf("取消没有退回插话: status=%s pending=%+v", run.Status, state.PendingInterjections)
+		}
+	})
+	t.Run("rejected", func(t *testing.T) {
+		s, db, root := interjectionPiRoot(t)
+		run, state := agentInterjectionState(t, s, root.ID)
+		interjectionApprovalRun(t, s, db, run, &state)
+		if _, err := s.InterjectCloudAgent("user", root.ID, "msg-r", "还在吗"); err != nil {
+			t.Fatal(err)
+		}
+		if err := s.DecideCloudAgentApproval("user", root.ID, "ap-1", "reject", "不要生成"); err != nil {
+			t.Fatal(err)
+		}
+		run, state = agentInterjectionState(t, s, root.ID)
+		if run.Status != "rejected" || len(state.PendingInterjections) != 0 || !agentHasEvent(state, "user_interjection_dropped") {
+			t.Fatalf("拒绝审批没有退回插话: status=%s pending=%+v", run.Status, state.PendingInterjections)
+		}
+	})
+}
+
+// 同一 messageId 重复提交不会送达两次：已送达的 id 被记住，重发直接短路。
+func TestCloudAgentPiInterjectionRedeliveryIsIdempotent(t *testing.T) {
+	s, db, root := interjectionPiRoot(t)
+	if _, err := s.InterjectCloudAgent("user", root.ID, "msg-1", "改成 16:9"); err != nil {
+		t.Fatal(err)
+	}
+	messages := []map[string]any{{"role": "user", "content": []any{map[string]any{"type": "text", "text": "继续"}}}}
+	interjectionModelStep(t, s, db, "user", root.ID, messages, `{"text":"好的","toolCalls":[]}`)
+	count, err := s.InterjectCloudAgent("user", root.ID, "msg-1", "改成 16:9")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if count != 0 {
+		t.Fatalf("已送达的插话被重新入队: pending=%d", count)
+	}
+	_, state := agentInterjectionState(t, s, root.ID)
+	if len(state.PendingInterjections) != 0 {
+		t.Fatal("重复提交把插话又放回了队列")
+	}
+	delivered := 0
+	for _, event := range state.Events {
+		if event.Type == "user_interjection_delivered" {
+			delivered++
+		}
+	}
+	if delivered != 1 {
+		t.Fatalf("delivered 事件出现 %d 次，应当只有 1 次", delivered)
+	}
+}
+
+// 完成闸一次带走多条插话时，合并后的续跑提示词也必须被恢复闸识别；
+// 否则它会被当成遗留提示词清空，已标记送达的插话到不了模型。
+func TestCloudAgentPiInterjectionResumePromptRecognisedForSeveralInterjections(t *testing.T) {
+	s, db, root := interjectionPiRoot(t)
+	run, state := agentInterjectionState(t, s, root.ID)
+	run.Status = "running"
+	state.Request.Budget.MaxSteps = 3
+	state.Step = 1
+	state.PiAssistantResponses = 1
+	state.Canonical.Messages = []map[string]any{
+		{"role": "user", "content": "分析剧情"},
+		{"role": "assistant", "content": "好的"},
+	}
+	interjectionSaveState(t, s, db, run, &state)
+	for id, text := range map[string]string{"msg-late-1": "加字幕", "msg-late-2": "再配个背景音乐"} {
+		if _, err := s.InterjectCloudAgent("user", root.ID, id, text); err != nil {
+			t.Fatal(err)
+		}
+	}
+	if err := s.completeCloudAgentPiRun("user", root.ID); err != nil {
+		t.Fatal(err)
+	}
+	run, state = agentInterjectionState(t, s, root.ID)
+	if run.Status != "running" || len(state.PendingInterjections) != 0 {
+		t.Fatalf("status=%s pending=%d", run.Status, len(state.PendingInterjections))
+	}
+	if !strings.Contains(state.PiResumePrompt, "加字幕") || !strings.Contains(state.PiResumePrompt, "再配个背景音乐") {
+		t.Fatalf("续跑提示词缺少插话: %q", state.PiResumePrompt)
+	}
+	if !cloudAgentPiResumePromptDeliversInterjection(&state) {
+		t.Fatal("恢复闸没有识别多条插话合并后的续跑提示词")
+	}
+}