feat: strengthen agent runtime boundaries

2026-03-16 11:45:00 +08:00 · 2026-03-16 11:45:00 +08:00 · 9ecd91376d
commit 9ecd91376d
parent ea561418c4
9 changed files with 1743 additions and 63 deletions
--- a/app/(chat)/api/chat/route.ts
+++ b/app/(chat)/api/chat/route.ts
@ -123,11 +123,26 @@ import type {
 } from "@/lib/ai/context-memory/types";
 import { ChatSDKError } from "@/lib/errors";
 import { extractImageUrlsFromText } from "@/lib/image-url";
-import { isExplicitPromptOptimizationRequest } from "@/lib/ai/smart-analyzer";
 import {
  resolveClarificationBypass,
  resolveClarificationMessage,
 } from "@/lib/ai/runtime/clarification-interruption";
+import {
+  pruneTaskBoundaryContext,
+  pruneTaskBoundaryHistoryMessages,
+} from "@/lib/ai/runtime/task-boundary-context-pruner";
+import {
+  resolveAgentModelSlotDescriptor,
+  upsertAgentModelSlot,
+  type AgentModelSlots,
+} from "@/lib/ai/runtime/agent-model-slots";
+import { isPureGreetingTurnRequest } from "@/lib/ai/runtime/task-boundary-classifier";
+import {
+  getProviderHealthSnapshot,
+  recordProviderHealthFailure,
+  recordProviderHealthSuccess,
+  shouldAvoidProvider,
+} from "@/lib/ai/runtime/provider-health";
 import { uploadFile } from "@/lib/storage/service";
 import { getVideoStoryboardConfig } from "@/lib/video-agent/config";
 import { runRuxaNineGridPipeline } from "@/lib/video-agent/pipeline/ruxa-nine-grid-pipeline";
@ -1234,15 +1249,6 @@ function normalizeAssistantPartsForPersistence(parts: Array<Record<string, any>>
  });
 }

-function isPureGreetingTurnRequest(text: string): boolean {
-  const normalized = String(text || "").toLowerCase().replace(/\s+/g, "").trim();
-  if (!normalized) return false;
-
-  return /^(你好|您好|hello|hi|嗨|哈喽|在吗|早上好|中午好|下午好|晚上好|深夜好|谢谢|感谢|收到|好的|ok|okay|明白了|了解了)$/.test(
-    normalized,
-  );
-}
-
 function buildAgentToolImageMarkdown(params: {
  originalUrl: string;
  thumbnailUrl?: string | null;
@ -3044,8 +3050,10 @@ export async function POST(request: Request) {
    let reusablePromptMemory = null;
    let confirmedPromptForConversationState: string | undefined;
    let smartAutoHistoryMessagesFromDb: DBMessage[] = [];
+    let smartAutoMediaHistoryMessagesFromDb: DBMessage[] = [];
    let preferredAgentModelId: string | undefined;
    let configuredAgentConversationModelId: string | undefined;
+    let agentModelSlots: AgentModelSlots = {};
    let pendingGridConfirmation:
      | PendingGridConfirmation
      | null = null;
@ -3153,6 +3161,24 @@ export async function POST(request: Request) {
      if (configuredAgentModel?.model?.id) {
        preferredAgentModelId = configuredAgentModel.model.id;
        configuredAgentConversationModelId = configuredAgentModel.model.id;
+        agentModelSlots = upsertAgentModelSlot(agentModelSlots, "planner", {
+          modelId: configuredAgentModel.model.id,
+          displayName: configuredAgentModel.model.displayName || configuredAgentModel.model.modelId,
+          providerId: configuredAgentModel.model.providerId,
+          providerType: configuredAgentModel.provider.type,
+          providerKey: configuredAgentModel.model.providerId,
+          source: "configured_agent_default",
+          reason: "agent_default_model",
+        });
+        agentModelSlots = upsertAgentModelSlot(agentModelSlots, "analyzer", {
+          modelId: configuredAgentModel.model.id,
+          displayName: configuredAgentModel.model.displayName || configuredAgentModel.model.modelId,
+          providerId: configuredAgentModel.model.providerId,
+          providerType: configuredAgentModel.provider.type,
+          providerKey: configuredAgentModel.model.providerId,
+          source: "configured_agent_default",
+          reason: "agent_default_model",
+        });
        console.log(
          `[Smart Auto Mode] 🧠 Loaded configured agent conversation model: ${configuredAgentModel.model.displayName}`,
        );
@ -3239,9 +3265,12 @@ export async function POST(request: Request) {
          workingState: requestWorkingState,
          activeStoryboardTask: activeStoryboardTaskForChat,
          hasImageInputForPromptCarryForward: hasImageInputForConversationAction,
+          hasCanvasSelection: Boolean(contextPayload?.selectedNodeIds?.length),
        });
        console.log("[ConversationStateProtocol] Bootstrap:", {
          action: conversationStateBootstrap.actionResolution.kind,
+          taskType: conversationStateBootstrap.taskBoundaryDecision.taskType,
+          taskMode: conversationStateBootstrap.taskBoundaryDecision.taskMode,
          hasPromptMemory: Boolean(conversationStateBootstrap.promptMemory?.optimizedPrompt),
          hasImageWorkflowMemory: Boolean(conversationStateBootstrap.imageWorkflowMemory),
          hasStoryboardPreferenceMemory: Boolean(conversationStateBootstrap.storyboardPreferenceMemory),
@ -3252,6 +3281,7 @@ export async function POST(request: Request) {
          phase: "bootstrap",
          payload: {
            action: conversationStateBootstrap.actionResolution.kind,
+            taskBoundary: conversationStateBootstrap.taskBoundaryDecision,
            pendingGridConfirmation: conversationStateBootstrap.pendingGridConfirmation,
            hasPromptMemory: Boolean(conversationStateBootstrap.promptMemory?.optimizedPrompt),
            hasImageWorkflowMemory: Boolean(conversationStateBootstrap.imageWorkflowMemory),
@ -3265,6 +3295,17 @@ export async function POST(request: Request) {
        pendingGridConfirmation = conversationStateBootstrap.pendingGridConfirmation;
        confirmedPromptForConversationState = conversationStateBootstrap.confirmedPrompt;
        const resolvedConversationAction = conversationStateBootstrap.actionResolution;
+        const taskBoundaryDecision = conversationStateBootstrap.taskBoundaryDecision;
+        const prunedTaskBoundaryContext = pruneTaskBoundaryContext({
+          decision: taskBoundaryDecision,
+          contextPayload: contextPayload || undefined,
+          autoContextText: autoContextText || undefined,
+          workingState: requestWorkingState || undefined,
+          hasCurrentImageInput: hasImageInputForConversationAction,
+        });
+        const effectiveContextPayload = prunedTaskBoundaryContext.contextPayload;
+        const effectiveAutoContextText = prunedTaskBoundaryContext.autoContextText || "";
+        const effectiveWorkingState = prunedTaskBoundaryContext.workingState;

    if (resolvedConversationAction.forceAgentConversation && !shouldRunAgentConversation) {
      shouldRunAgentConversation = true;
@ -3288,13 +3329,13 @@ export async function POST(request: Request) {
    }

    const shouldFastTrackPromptOnly =
-      shouldRunAgentConversation && isExplicitPromptOptimizationRequest(requestTextContent);
+      shouldRunAgentConversation && taskBoundaryDecision.taskType === "prompt_only";
    if (shouldFastTrackPromptOnly) {
      if (!requestResolvedMediaInput) {
        requestResolvedMediaInput = resolveChatMediaInput({
          messageParts: message.parts as any[],
          historyMessages: [],
-          contextPayload: contextPayload || undefined,
+          contextPayload: effectiveContextPayload,
          requestOrigin,
          maxImages: 4,
        });
@ -3304,8 +3345,8 @@ export async function POST(request: Request) {
        requestTargetResolution = resolveTargetAssets({
          text: requestTextContent,
          mediaInput: requestResolvedMediaInput,
-          contextPayload: contextPayload || undefined,
-          workingState: requestWorkingState || undefined,
+          contextPayload: effectiveContextPayload,
+          workingState: effectiveWorkingState || undefined,
        });
      }

@ -3329,8 +3370,26 @@ export async function POST(request: Request) {
      
      try {
        // 🆕 首先获取历史消息（在 Smart Auto Mode 内部单独获取，避免依赖后面的 messagesFromDb）
-        smartAutoHistoryMessagesFromDb = await getMessagesByChatId({ id });
-        console.log("[Smart Auto Mode] Loaded", smartAutoHistoryMessagesFromDb.length, "messages from DB");
+        const loadedHistoryMessages = await getMessagesByChatId({ id });
+        smartAutoHistoryMessagesFromDb = pruneTaskBoundaryHistoryMessages({
+          decision: taskBoundaryDecision,
+          messages: loadedHistoryMessages,
+          purpose: "model",
+          hasCurrentImageInput: hasImageInputForConversationAction,
+        });
+        smartAutoMediaHistoryMessagesFromDb = pruneTaskBoundaryHistoryMessages({
+          decision: taskBoundaryDecision,
+          messages: loadedHistoryMessages,
+          purpose: "media",
+          hasCurrentImageInput: hasImageInputForConversationAction,
+        });
+        console.log("[Smart Auto Mode] Loaded", loadedHistoryMessages.length, "messages from DB");
+        console.log("[Smart Auto Mode] Pruned history:", {
+          modelMessages: smartAutoHistoryMessagesFromDb.length,
+          mediaMessages: smartAutoMediaHistoryMessagesFromDb.length,
+          taskType: taskBoundaryDecision.taskType,
+          taskMode: taskBoundaryDecision.taskMode,
+        });

        autoTextForFallback = smartAutoInputText;
        if (!requestWorkingState) {
@ -3347,8 +3406,8 @@ export async function POST(request: Request) {
        if (!requestResolvedMediaInput) {
          requestResolvedMediaInput = resolveChatMediaInput({
            messageParts: message.parts as any[],
-            historyMessages: smartAutoHistoryMessagesFromDb,
-            contextPayload: contextPayload || undefined,
+            historyMessages: smartAutoMediaHistoryMessagesFromDb,
+            contextPayload: effectiveContextPayload,
            requestOrigin,
            maxImages: 4,
          });
@ -3357,14 +3416,14 @@ export async function POST(request: Request) {
          requestTargetResolution = resolveTargetAssets({
            text: requestTextContent,
            mediaInput: requestResolvedMediaInput,
-            contextPayload: contextPayload || undefined,
-            workingState: requestWorkingState || undefined,
+            contextPayload: effectiveContextPayload,
+            workingState: effectiveWorkingState || undefined,
          });
        }

        const hasSelectedImageHint =
-          Boolean(contextPayload?.selectedNodeIds?.length) ||
-          /用户已选中\s*\d+\s*张图片/.test(autoContextText);
+          Boolean(effectiveContextPayload?.selectedNodeIds?.length) ||
+          /用户已选中\s*\d+\s*张图片/.test(effectiveAutoContextText);
        const hasExplicitTargetHint =
          requestTargetResolution.resolutionMode === "text_reference" ||
          /第[一二三四五六七八九十\d]+张|左边|左侧|左上|左下|右边|右侧|右上|右下|上面|上方|下面|下方|中间|居中|中央|选中/.test(requestTextContent);
@ -3438,7 +3497,7 @@ export async function POST(request: Request) {
        }
        const composedConversationState = composeConversationStateProtocol({
          bootstrap: conversationStateBootstrap,
-          autoContextText: autoContextText || undefined,
+          autoContextText: effectiveAutoContextText || undefined,
          resolvedMediaInput: requestResolvedMediaInput,
          targetResolution: requestTargetResolution,
        });
@ -3462,16 +3521,16 @@ export async function POST(request: Request) {
            hasImage,
            imageUrls,
            assetIds: Array.from(
-              new Set([
-                ...extractMessageImageAssetIds(message.parts as any[]),
-                ...((contextPayload?.selectedAssetIds || []).filter(Boolean) as string[]),
-              ]),
-            ),
+                new Set([
+                  ...extractMessageImageAssetIds(message.parts as any[]),
+                  ...((effectiveContextPayload?.selectedAssetIds || []).filter(Boolean) as string[]),
+                ]),
+              ),
            imageFromHistory: smartAutoHistoryImages.length > 0,
            targetResolutionMode: requestTargetResolution.resolutionMode,
            mediaSource: requestResolvedMediaInput.source,
            hasCanvasSelection:
-              Boolean(contextPayload?.selectedNodeIds?.length) ||
+              Boolean(effectiveContextPayload?.selectedNodeIds?.length) ||
              requestResolvedMediaInput.selectedImageUrls.length > 0,
            selectedImageCount: requestResolvedMediaInput.selectedImageUrls.length,
            historyImageCount:
@ -3481,7 +3540,7 @@ export async function POST(request: Request) {
            historyMessages,
            context:
              composedConversationState.composedConversationMemory.contextSummary ||
-              autoContextText ||
+              effectiveAutoContextText ||
              undefined,
            preferredModelId: preferredAgentModelId,
            threadState: composedConversationState.threadState,
@ -3838,7 +3897,7 @@ export async function POST(request: Request) {
          const rescueAssetIds = Array.from(
            new Set([
              ...extractMessageImageAssetIds(message.parts as any[]),
-              ...((contextPayload?.selectedAssetIds || []).filter(Boolean) as string[]),
+              ...((effectiveContextPayload?.selectedAssetIds || []).filter(Boolean) as string[]),
            ]),
          );
          const rescueTimeoutMs = getBudgetedTimeout({
@ -4043,7 +4102,7 @@ export async function POST(request: Request) {
              const summaryAssetIds = Array.from(
                new Set([
                  ...extractMessageImageAssetIds(message.parts as any[]),
-                  ...((contextPayload?.selectedAssetIds || []).filter(Boolean) as string[]),
+                  ...((effectiveContextPayload?.selectedAssetIds || []).filter(Boolean) as string[]),
                ]),
              );
              const fallbackSummaryTimeoutMs = getBudgetedTimeout({
@ -4056,7 +4115,7 @@ export async function POST(request: Request) {
                  summarizeImageForExecutionReport({
                    text: autoTextForFallback,
                    imageUrls: summaryImageUrls,
-                    context: typeof autoContext === "string" ? autoContext : undefined,
+                    context: effectiveAutoContextText || undefined,
                    preferredModelId: preferredAgentModelId,
                    assetIds: summaryAssetIds,
                  }),
@ -4335,11 +4394,11 @@ export async function POST(request: Request) {
            .map((part) => part.text)
            .join(" ");
          const summaryAssetIds = Array.from(
-            new Set([
-              ...extractMessageImageAssetIds(message.parts as any[]),
-              ...((contextPayload?.selectedAssetIds || []).filter(Boolean) as string[]),
-            ]),
-          );
+                new Set([
+                  ...extractMessageImageAssetIds(message.parts as any[]),
+                  ...((effectiveContextPayload?.selectedAssetIds || []).filter(Boolean) as string[]),
+                ]),
+              );

          let summary: string | null = null;
          const summaryTimeoutMs = getBudgetedTimeout({
@ -4352,7 +4411,7 @@ export async function POST(request: Request) {
              summarizeImageForExecutionReport({
                text: latestUserText,
                imageUrls: summaryImageUrls,
-                context: typeof autoContext === "string" ? autoContext : undefined,
+                context: effectiveAutoContextText || undefined,
                preferredModelId: preferredAgentModelId,
                assetIds: summaryAssetIds,
              }),
@ -4421,6 +4480,86 @@ export async function POST(request: Request) {
      }
    }

+    if (shouldRunAgentConversation && smartAutoExecution?.shouldStayConversational) {
+      const candidateReplyDescriptor = await resolveAgentModelSlotDescriptor(finalModelId).catch(
+        () => null,
+      );
+      if (candidateReplyDescriptor) {
+        agentModelSlots = upsertAgentModelSlot(agentModelSlots, "reply", {
+          ...candidateReplyDescriptor,
+          source: "candidate_reply_model",
+          reason: "post_agent_routing",
+          locked: shouldKeepSelectedModel,
+        });
+
+        const providerHealth = getProviderHealthSnapshot(candidateReplyDescriptor.providerKey);
+        if (
+          !shouldKeepSelectedModel &&
+          shouldAvoidProvider(providerHealth)
+        ) {
+          let fallbackReplyModelId: string | undefined;
+
+          if (
+            preferredAgentModelId &&
+            preferredAgentModelId !== finalModelId
+          ) {
+            const preferredDescriptor = await resolveAgentModelSlotDescriptor(
+              preferredAgentModelId,
+            ).catch(() => null);
+            if (
+              preferredDescriptor &&
+              !shouldAvoidProvider(getProviderHealthSnapshot(preferredDescriptor.providerKey))
+            ) {
+              fallbackReplyModelId = preferredAgentModelId;
+            }
+          }
+
+          if (!fallbackReplyModelId) {
+            try {
+              const { routeToModel } = await import("@/lib/ai/model-router");
+              const replyFallbackRoute = await routeToModel({
+                intent: "chat",
+                subIntent: undefined,
+                hasImage: false,
+              });
+              const fallbackDescriptor = await resolveAgentModelSlotDescriptor(
+                replyFallbackRoute.id,
+              ).catch(() => null);
+              if (
+                fallbackDescriptor &&
+                !shouldAvoidProvider(getProviderHealthSnapshot(fallbackDescriptor.providerKey))
+              ) {
+                fallbackReplyModelId = replyFallbackRoute.id;
+              }
+            } catch (replyRouteError) {
+              console.warn(
+                "[Smart Auto Mode] Failed to resolve reply fallback route after provider degradation:",
+                replyRouteError,
+              );
+            }
+          }
+
+          if (fallbackReplyModelId && fallbackReplyModelId !== finalModelId) {
+            finalModelId = fallbackReplyModelId;
+            const fallbackDescriptor = await resolveAgentModelSlotDescriptor(finalModelId).catch(
+              () => null,
+            );
+            if (fallbackDescriptor) {
+              agentModelSlots = upsertAgentModelSlot(agentModelSlots, "reply", {
+                ...fallbackDescriptor,
+                source: "provider_health_fallback",
+                reason: providerHealth?.lastEventType || "degraded_provider",
+                locked: false,
+              });
+            }
+            console.warn(
+              `[Smart Auto Mode] ♻️ Switched reply model due provider health: ${candidateReplyDescriptor.displayName || candidateReplyDescriptor.modelId} -> ${fallbackDescriptor?.displayName || fallbackReplyModelId}`,
+            );
+          }
+        }
+      }
+    }
+
    const session = await auth();

    if (!session?.user) {
@ -4474,6 +4613,7 @@ export async function POST(request: Request) {

    let chat = await getChatById({ id });
    let messagesFromDb: DBMessage[] = [];
+    let messagesForMediaInput: DBMessage[] = [];

    if (chat) {
      if (chat.userId !== session.user.id) {
@ -4522,6 +4662,20 @@ export async function POST(request: Request) {
      messagesFromDb = await getMessagesByChatId({ id });
    }

+    const loadedMessagesForBoundary = messagesFromDb;
+    messagesFromDb = pruneTaskBoundaryHistoryMessages({
+      decision: taskBoundaryDecision,
+      messages: loadedMessagesForBoundary,
+      purpose: "model",
+      hasCurrentImageInput: hasImageInputForConversationAction,
+    });
+    messagesForMediaInput = pruneTaskBoundaryHistoryMessages({
+      decision: taskBoundaryDecision,
+      messages: loadedMessagesForBoundary,
+      purpose: "media",
+      hasCurrentImageInput: hasImageInputForConversationAction,
+    });
+
    if (!requestWorkingState) {
      requestWorkingState = await getConversationWorkingStateByChatId({
        chatId: id,
@ -4530,8 +4684,8 @@ export async function POST(request: Request) {
    if (!requestResolvedMediaInput) {
      requestResolvedMediaInput = resolveChatMediaInput({
        messageParts: message.parts as any[],
-        historyMessages: messagesFromDb,
-        contextPayload: contextPayload || undefined,
+        historyMessages: messagesForMediaInput,
+        contextPayload: effectiveContextPayload,
        requestOrigin,
        maxImages: 4,
      });
@ -4540,8 +4694,8 @@ export async function POST(request: Request) {
      requestTargetResolution = resolveTargetAssets({
        text: requestTextContent,
        mediaInput: requestResolvedMediaInput,
-        contextPayload: contextPayload || undefined,
-        workingState: requestWorkingState || undefined,
+        contextPayload: effectiveContextPayload,
+        workingState: effectiveWorkingState || undefined,
      });
    }
    if (
@ -4688,7 +4842,7 @@ export async function POST(request: Request) {
      shouldRunAgentConversation &&
      !smartAutoNeedsClarification &&
      promptOnlySourceImageUrls.length > 0 &&
-      isExplicitPromptOptimizationRequest(requestTextContent);
+      taskBoundaryDecision.taskType === "prompt_only";

    if (wantsPromptOnlyFromImageContext) {
      let promptOnlyImageSummary = "";
@ -4701,16 +4855,16 @@ export async function POST(request: Request) {
        });
        if (summaryTimeoutMs) {
          const summaryAssetIds = Array.from(
-            new Set([
-              ...extractMessageImageAssetIds(message.parts as any[]),
-              ...((contextPayload?.selectedAssetIds || []).filter(Boolean) as string[]),
-            ]),
-          );
+              new Set([
+                ...extractMessageImageAssetIds(message.parts as any[]),
+                ...((effectiveContextPayload?.selectedAssetIds || []).filter(Boolean) as string[]),
+              ]),
+            );
          const summary = await withOperationTimeout(
            summarizeImageForExecutionReport({
              text: requestTextContent,
              imageUrls: promptOnlySourceImageUrls,
-              context: typeof autoContext === "string" ? autoContext : undefined,
+              context: effectiveAutoContextText || undefined,
              preferredModelId: preferredAgentModelId,
              assetIds:
                promptOnlySourceAssetIds.length > 0 ? promptOnlySourceAssetIds : summaryAssetIds,
@ -4752,7 +4906,7 @@ export async function POST(request: Request) {
                };
              })
              .filter(Boolean) as Array<{ role: "user" | "assistant"; content: string }>,
-            context: typeof autoContext === "string" ? autoContext : undefined,
+            context: effectiveAutoContextText || undefined,
          }).catch(() =>
            optimizePrompt({
              originalPrompt: promptSeed,
@ -4832,7 +4986,7 @@ export async function POST(request: Request) {
          part?.type === "file" || part?.type === "image" || part?.type === "image_url"
        ),
        hasHistoryImageInput: smartAutoHistoryImages.length > 0,
-        hasCanvasSelection: Boolean(contextPayload?.selectedNodeIds?.length),
+        hasCanvasSelection: Boolean(effectiveContextPayload?.selectedNodeIds?.length),
        targetResolutionMode: requestTargetResolution?.resolutionMode,
        executionMode: smartAutoExecution?.mode,
        friendlyMessage: smartAutoFriendlyMessage,
@ -6495,6 +6649,7 @@ export async function POST(request: Request) {
    let actualModelId: string;
    let resolvedModelSupportsImageGen = false;
    let resolvedProviderType: string | undefined;
+    let resolvedRuntimeModelLookupId = finalModelId;

    if (isUUID) {
      // 从数据库获取动态配置的模型
@ -6542,6 +6697,7 @@ export async function POST(request: Request) {
        ).toResponse();
      }
      languageModel = dynamicModel;
+      resolvedRuntimeModelLookupId = targetModelId;
      // 对于动态模型，我们需要从数据库获取 modelId
      const finalModelData = await getAiModelById({ id: targetModelId });
      actualModelId = finalModelData?.modelId ?? targetModelId;
@ -6556,6 +6712,38 @@ export async function POST(request: Request) {
      languageModel = myProvider.languageModel(finalModelId);
      actualModelId = languageModel.modelId;
      resolvedProviderType = (languageModel as any)?.provider;
+      resolvedRuntimeModelLookupId = actualModelId;
+    }
+
+    const activeRuntimeModelDescriptor = await resolveAgentModelSlotDescriptor(
+      resolvedRuntimeModelLookupId,
+    ).catch(() => null);
+    if (shouldRunAgentConversation && activeRuntimeModelDescriptor) {
+      const slotName =
+        smartAutoExecution?.shouldApplyImagePipeline ||
+        smartAutoExecution?.shouldApplyVideoPipeline
+          ? "execution"
+          : "reply";
+      agentModelSlots = upsertAgentModelSlot(agentModelSlots, slotName, {
+        ...activeRuntimeModelDescriptor,
+        source: "runtime_selected_model",
+        reason:
+          slotName === "reply"
+            ? smartAutoExecution?.shouldStayConversational
+              ? "conversational_reply_runtime"
+              : "default_reply_runtime"
+            : "execution_runtime_model",
+        locked: slotName === "reply" ? shouldKeepSelectedModel : false,
+      });
+      pendingProtocolDebugEvents.push({
+        phase: "modelSlots",
+        payload: {
+          modelSlots: agentModelSlots,
+          providerHealth: activeRuntimeModelDescriptor.providerKey
+            ? getProviderHealthSnapshot(activeRuntimeModelDescriptor.providerKey)
+            : null,
+        },
+      });
    }

    // ========== 图片生成配额检查 ==========
@ -9301,6 +9489,12 @@ export async function POST(request: Request) {
            functionId: "stream-text",
          },
          onFinish: async ({ usage }) => {
+            if (activeRuntimeModelDescriptor?.providerKey) {
+              recordProviderHealthSuccess({
+                providerKey: activeRuntimeModelDescriptor.providerKey,
+                providerType: activeRuntimeModelDescriptor.providerType,
+              });
+            }
            try {
              const providers = await getTokenlensCatalog();
              if (!actualModelId) {
@ -9343,6 +9537,13 @@ export async function POST(request: Request) {
          result.consumeStream({
            onError: (error) => {
              console.error("[Chat API] streamText consume error:", error);
+              if (activeRuntimeModelDescriptor?.providerKey) {
+                recordProviderHealthFailure({
+                  providerKey: activeRuntimeModelDescriptor.providerKey,
+                  providerType: activeRuntimeModelDescriptor.providerType,
+                  error,
+                });
+              }
            },
          });

@ -9351,6 +9552,13 @@ export async function POST(request: Request) {
              sendReasoning: true,
              onError: (error) => {
                console.error("[Chat API] streamText ui stream error:", error);
+                if (activeRuntimeModelDescriptor?.providerKey) {
+                  recordProviderHealthFailure({
+                    providerKey: activeRuntimeModelDescriptor.providerKey,
+                    providerType: activeRuntimeModelDescriptor.providerType,
+                    error,
+                  });
+                }
                return getChatStreamErrorMessage(error);
              },
            })
--- a/docs/Agent真正变聪明-系统级整改清单-2026-03-16.md
+++ b/docs/Agent真正变聪明-系统级整改清单-2026-03-16.md
@ -0,0 +1,667 @@
+# Agent 真正变聪明：系统级整改清单（2026-03-16）
+
+> 目标：不是继续给 Agent 打零散补丁，而是把它从“很多聪明 patch 的组合体”，升级成“统一、稳定、可解释、可扩展的对话智能系统”。
+>
+> 适用范围：
+> - `app/(chat)/api/chat/route.ts`
+> - `lib/ai/auto-agent-graph.ts`
+> - `lib/ai/smart-analyzer.ts`
+> - `lib/ai/conversation-state-protocol.ts`
+> - `lib/ai/conversation-*`
+> - `lib/ai/runtime/*`
+> - `lib/video-agent/*`
+> - 画布上下文、图片工作流、视频工作流
+
+---
+
+## 0. 先说结论
+
+### 0.1 现在 Agent 为什么“看起来不聪明”
+
+不是单纯模型不够强，而是系统层面还有 5 个根问题：
+
+1. **任务边界不稳定**
+   - 新任务、继续任务、只聊天、只看图、只要提示词、直接执行，经常互相串。
+
+2. **上下文污染严重**
+   - 当前输入、历史图片、画布图片、上一轮 workflow、长期记忆，经常一起灌给模型。
+   - 结果就是用户说一句“你好”，系统也可能把它理解成“继续上一轮修图”。
+
+3. **理解层和执行层没完全解耦**
+   - 一部分逻辑已经进 protocol / subagent / runtime。
+   - 但还有不少判断仍然散在 `chat/route.ts`，导致行为不稳定、修改成本高。
+
+4. **模型链路不够统一**
+   - planner、analyzer、最终回复、图片执行、视频执行并不是同一套稳定规则。
+   - 之前就出现过：Agent 默认模型是豆包，但最终对话回复走了别的供应商。
+
+5. **超时退化不够“保正确”**
+   - 一旦 `runAutoAgentGraph`、图像总结、Provider 流式返回超时，系统容易退化成“还能跑，但不对题”。
+
+### 0.2 我们这次整改的目标
+
+这次不是“继续修 10 个 bug”，而是把 Agent 升级成：
+
+- **先判断任务，再选择上下文**
+- **先选正确路线，再决定模型和工具**
+- **宁可简单，也不跑偏**
+- **即使失败，也要明确告诉用户为什么失败**
+
+---
+
+## 1. 这次整改的总原则
+
+### 1.1 一个原则：先判任务，再喂上下文
+
+以后不能再用“把所有上下文先塞给模型，再让模型自己判断”的方式。
+
+要改成：
+
+1. 先判断这轮是：
+   - 新任务
+   - 连续任务
+   - 只分析
+   - 只要提示词
+   - 直接执行
+   - 聊天寒暄
+2. 再按这个任务类型，裁剪要喂给模型的上下文
+
+### 1.2 一个原则：执行型和对话型必须分开
+
+以后要明确分两类：
+
+- **对话型回合**
+  - 聊天
+  - 看图分析
+  - 提示词整理
+  - 澄清确认
+
+- **执行型回合**
+  - 生图
+  - 修图
+  - 生视频
+  - 分镜确认执行
+
+对话型回合不能轻易被旧工作流拖进执行型链路。
+
+### 1.3 一个原则：失败也要“可解释”
+
+以后不允许再出现这几种体验：
+
+- 用户发了消息，看起来“没反应”
+- 数据库里只存了 keepalive
+- 模型报错了，但前端没显示
+- 回退了，但用户不知道系统为什么回退
+
+---
+
+## 2. 总体架构目标
+
+目标架构：
+
+1. **Task Boundary Layer**
+2. **Context Pruning Layer**
+3. **Conversation State Protocol**
+4. **Planner / Subagent Layer**
+5. **Execution Gate Layer**
+6. **Persistence / Event Layer**
+7. **Frontend Projection Layer**
+
+一句话解释：
+
+- **边界层**决定“你这轮到底要干嘛”
+- **裁剪层**决定“该给模型看哪些上下文”
+- **协议层**负责统一状态读写
+- **子代理层**负责专业领域判断
+- **执行门控层**负责确认、取消、修改、超时回退
+- **事件层**负责可追踪
+- **前端层**负责把系统理解结果明确展示给用户
+
+---
+
+## 3. 五大整改块
+
+---
+
+## 3.1 任务理解层
+
+### 3.1.1 目标
+
+让 Agent 先稳定判断：
+
+- 这是新任务还是继续任务
+- 这是对话型还是执行型
+- 这是只分析、只要提示词、还是直接执行
+
+### 3.1.2 当前问题
+
+当前系统虽然已经有：
+
+- `conversation-action-resolver`
+- `smart-analyzer`
+- `skill-router`
+- `auto-agent-graph`
+
+但这几层对“任务边界”的定义还不完全统一。
+
+典型问题：
+
+1. “你好”被旧图片 workflow 拖成连续修图
+2. “帮我写这张图的提示词”被拖成生图/修图流程
+3. 用户只是问“这个图怎么写提示词”，系统却提前进入执行链
+
+### 3.1.3 必做改造
+
+#### A. 新增统一任务边界分类器
+
+建议新增：
+
+- `lib/ai/runtime/task-boundary-classifier.ts`
+
+输出统一结构：
+
+```ts
+type TaskBoundaryDecision = {
+  taskType:
+    | "greeting"
+    | "chat"
+    | "analysis_only"
+    | "prompt_only"
+    | "execution_image"
+    | "execution_video"
+    | "workflow_followup";
+  taskMode: "new" | "continue" | "revise" | "confirm" | "cancel";
+  shouldUseCanvasContext: boolean;
+  shouldUseHistoryImages: boolean;
+  shouldUseWorkflowMemory: boolean;
+  shouldAllowExecution: boolean;
+  confidence: number;
+  reason: string;
+};
+```
+
+#### B. 明确任务优先级规则
+
+优先级必须写死在系统里，不能全靠模型猜：
+
+1. `prompt_only` > `workflow_followup`
+2. `analysis_only` > `execution_image`
+3. `greeting` > 任意旧 workflow
+4. `confirm/cancel/revise` > 普通分析
+5. 当前上传图 > 历史图 > 画布背景图
+
+#### C. 给“非执行任务”开快路径
+
+以下场景应该直接走快路径：
+
+- 纯打招呼
+- 只要提示词
+- 只看图分析
+- 只问建议
+
+这些回合不应该再先跑完整 `runAutoAgentGraph`。
+
+### 3.1.4 落点文件
+
+- `app/(chat)/api/chat/route.ts`
+- `lib/ai/conversation-action-resolver.ts`
+- `lib/ai/smart-analyzer.ts`
+- `lib/ai/runtime/task-boundary-classifier.ts`（新增）
+
+### 3.1.5 验收标准
+
+- “你好”永远不被识别为连续修图/连续生图
+- “帮我写这张图的提示词”永远不直接触发执行
+- “按照刚才那个提示词，再生成 4 张”会稳定识别为连续生图执行
+
+---
+
+## 3.2 记忆层
+
+### 3.2.1 目标
+
+让 Agent 的记忆从“混在一起”变成“分层可控”。
+
+### 3.2.2 当前问题
+
+现在虽然已经有：
+
+- `promptMemory`
+- `imageWorkflowMemory`
+- `storyboardPreferenceMemory`
+- `threadState`
+- `workingState`
+
+但问题是：
+
+1. 历史 workflow 容易污染当前回合
+2. 画布上下文太大，容易把简单任务拖重
+3. “当前相关记忆”和“长期偏好记忆”还没完全分层
+
+### 3.2.3 必做改造
+
+#### A. 拆成 4 层记忆
+
+1. **Turn Memory**
+   - 当前回合输入
+   - 当前上传图片
+   - 当前画布选中
+
+2. **Task Memory**
+   - 当前任务工作流状态
+   - 当前任务目标图
+   - 当前任务优化提示词
+
+3. **Session Memory**
+   - 当前会话最近 3~5 个高相关任务摘要
+
+4. **Preference Memory**
+   - 用户长期偏好
+   - 常用比例
+   - 常用宫格
+   - 常用确认习惯
+
+#### B. 增加“记忆衰减/降权”机制
+
+如果旧 workflow 满足任一条件，应自动降权：
+
+- 超过 N 轮未继续
+- 当前上传了新图
+- 当前任务类型和旧任务类型冲突
+- 用户显式说“重新来 / 新开一个 / 别按刚才那个”
+
+#### C. 建立 relevance score
+
+建议新增：
+
+- `lib/ai/runtime/memory-relevance.ts`
+
+对候选记忆打分：
+
+- 当前上传图命中：+100
+- 当前选中图命中：+90
+- 显式引用“上一张 / 第三张 / 这个提示词”：+80
+- 同类任务：+40
+- 同会话最近 3 轮：+20
+- 超过 10 轮未使用：-50
+
+以后只把 top-k 记忆喂给 Planner。
+
+### 3.2.4 落点文件
+
+- `lib/ai/conversation-memory-composer.ts`
+- `lib/ai/conversation-thread-state-builder.ts`
+- `lib/ai/runtime/agent-thread-state.ts`
+- `lib/ai/runtime/memory-relevance.ts`（新增）
+
+### 3.2.5 验收标准
+
+- 同一会话里，旧图不再轻易干扰新上传图任务
+- 简单聊天不再带上大段无关画布上下文
+- “继续上一个提示词”仍然能稳定命中
+
+---
+
+## 3.3 路由层
+
+### 3.3.1 目标
+
+让 Planner、Analyzer、最终回复模型、执行模型之间的职责彻底清晰。
+
+### 3.3.2 当前问题
+
+现在的问题不是“没有路由”，而是“路由规则不够统一”：
+
+1. planner 模型和 reply 模型曾经不一致
+2. analyzer 有时会被别的 provider 抢走
+3. 生图/修图/视频执行模型与对话模型混杂
+4. provider 报错或额度不足时，回退不够统一
+
+### 3.3.3 必做改造
+
+#### A. 明确 4 类模型槽位
+
+统一定义：
+
+1. `plannerModel`
+2. `analyzerModel`
+3. `replyModel`
+4. `executionModel`
+
+而不是所有地方都混着叫“当前模型”。
+
+#### B. Agent 模式默认规则
+
+在 `selectedChatModel = auto` 或进入 Agent 模式时：
+
+- `plannerModel` = Agent 默认模型
+- `analyzerModel` = Agent 默认模型优先，多模态能力不足时再回退
+- `replyModel` = Agent 默认模型
+- `executionModel` = 按任务类型路由
+
+#### C. Provider 健康与额度状态缓存
+
+建议新增：
+
+- `lib/ai/runtime/provider-health.ts`
+
+缓存内容：
+
+- 最近错误类型
+- 最近额度不足
+- 最近超时率
+- 最近 5 分钟成功率
+
+如果某 provider 连续报：
+
+- `quota not enough`
+- `403`
+- `timeout`
+
+则在短时间内直接降级，避免用户连续踩坑。
+
+### 3.3.4 落点文件
+
+- `app/(chat)/api/chat/route.ts`
+- `lib/ai/model-router.ts`
+- `lib/ai/smart-analyzer.ts`
+- `lib/ai/auto-agent-graph.ts`
+- `lib/ai/runtime/provider-health.ts`（新增）
+
+### 3.3.5 验收标准
+
+- Agent 模式不会再出现“规划走豆包、回复走别家”的错位
+- 上游 provider 额度不足时，能快速回退或明确报错
+- 对话模型、执行模型职责清晰
+
+---
+
+## 3.4 执行层
+
+### 3.4.1 目标
+
+让执行系统遵守：
+
+- 该确认时必须确认
+- 该取消时必须停
+- 该提示时只提示，不执行
+- 该失败时明确失败
+
+### 3.4.2 当前问题
+
+虽然我们已经做了很多 protocol / workflow 事件：
+
+- clarification
+- storyboard confirm
+- revision
+- cancel
+- execution started/completed
+- video generation started/completed/failed
+
+但执行层还存在 3 个结构问题：
+
+1. 快路径不够多
+2. timeout fallback 仍偏重
+3. “对话型输出”和“执行型输出”没有完全剥离
+
+### 3.4.3 必做改造
+
+#### A. 建立 execution policy gate
+
+建议新增：
+
+- `lib/ai/runtime/execution-policy-gate.ts`
+
+统一做这些判定：
+
+- 当前回合是否允许执行
+- 当前回合是否必须先确认
+- 当前回合是否应该只输出提示词
+- 当前回合是否应该只对话
+
+#### B. 扩展 fast-path 列表
+
+必须内建以下快路径：
+
+1. greeting fast-path
+2. prompt-only fast-path
+3. analysis-only fast-path
+4. confirmation fast-path
+5. cancel fast-path
+
+#### C. timeout 退化策略升级
+
+以后 timeout 退化规则应改成：
+
+- `execute` 超时 → 不自动换任务类型
+- `planner` 超时 → 退到简单规则，但不跨任务边界
+- `provider` 超时 → 直接明确告诉用户该 provider 超时
+
+一句话：
+
+**退化后可以更笨，但不能更错。**
+
+### 3.4.4 落点文件
+
+- `lib/ai/runtime/middlewares/human-confirmation-gate.ts`
+- `lib/ai/runtime/execution-policy-gate.ts`（新增）
+- `app/(chat)/api/chat/route.ts`
+- `lib/video-agent/storyboard-chat-followup.ts`
+
+### 3.4.5 验收标准
+
+- “只要提示词”绝不执行生图
+- “确认开始生成”绝不只回复聊天
+- provider 报错时前端总能看到明确原因
+
+---
+
+## 3.5 前端交互层
+
+### 3.5.1 目标
+
+让用户知道 Agent 当前理解成了什么，而不是靠猜。
+
+### 3.5.2 当前问题
+
+现在前端可见性还不够：
+
+1. 用户不知道系统把这轮理解成聊天、分析还是执行
+2. 用户不知道系统到底引用了哪张图
+3. 用户不知道系统为什么突然继续上一轮任务
+4. 用户不知道系统为什么没执行
+
+### 3.5.3 必做改造
+
+#### A. 增加“本轮理解摘要”
+
+建议在输入框附近或 assistant 第一段里展示：
+
+- 本轮类型：`聊天 / 分析 / 提示词 / 生图 / 修图 / 视频`
+- 目标来源：`当前上传 / 画布选中 / 历史结果 / 上一轮提示词`
+- 是否会执行：`仅分析 / 待确认 / 直接执行`
+
+#### B. 增加“任务切换提示”
+
+当系统从旧 workflow 切到新任务时，前端明确显示：
+
+- “已识别为新任务，不继续沿用上一轮修图流程”
+
+当系统继续旧任务时，显示：
+
+- “已继续上一轮修图流程”
+
+#### C. 协议调试面板继续升级
+
+当前已有 `ProtocolDebugPanel`，建议继续加：
+
+- 当前任务边界摘要
+- 当前记忆裁剪摘要
+- 当前模型槽位摘要（planner / reply / execution）
+- 当前 provider 健康状态
+
+### 3.5.4 落点文件
+
+- `components/chat.tsx`
+- `components/protocol-debug-panel.tsx`
+- `components/canvas/CanvasChatPanel.tsx`
+
+### 3.5.5 验收标准
+
+- 用户能看见“系统理解成什么”
+- 用户能看见“为什么这轮不执行”
+- debug 时能看见模型和 workflow 生命周期
+
+---
+
+## 4. 实施顺序（必须按这个顺序）
+
+### P0：先止血
+
+1. greeting fast-path
+2. prompt-only fast-path
+3. stream provider 错误明确回显
+4. Agent 模式 replyModel 跟随 Agent 默认模型
+
+> 这几项已经部分落地，但要继续收边并补测试。
+
+### P1：统一任务边界
+
+1. `task-boundary-classifier`
+2. 统一 taskType / taskMode / allowExecution
+3. 对 route 中旧 heuristics 收口
+
+### P2：统一记忆裁剪
+
+1. relevance scoring
+2. session/task/preference 记忆分层
+3. 旧 workflow 自动降权
+
+### P3：统一模型链路
+
+1. planner / analyzer / reply / execution 四槽位
+2. provider health cache
+3. 额度不足/超时自动降级
+
+### P4：统一前端解释性
+
+1. 本轮理解摘要
+2. 任务切换提示
+3. 调试面板生命周期摘要
+
+---
+
+## 5. 开发清单（可直接排期）
+
+### 5.1 任务理解层
+
+- [ ] 新增 `lib/ai/runtime/task-boundary-classifier.ts`
+- [ ] 把 `conversation-action-resolver` 和 `smart-analyzer` 结果统一映射到 `taskType`
+- [ ] 增加 `taskMode`：`new/continue/revise/confirm/cancel`
+- [ ] 增加 fast-path 优先级表
+
+### 5.2 记忆层
+
+- [ ] 新增 `memory-relevance.ts`
+- [ ] `composeConversationMemory` 支持 top-k 裁剪
+- [ ] 为 workflow memory 增加过期/降权规则
+- [ ] 为 prompt-only / analysis-only 场景建立独立上下文模板
+
+### 5.3 路由层
+
+- [ ] 统一 planner / analyzer / reply / execution 槽位
+- [ ] Agent replyModel 跟随 Agent 默认模型
+- [ ] provider health 状态缓存
+- [ ] 上游 quota / timeout 短时熔断
+
+### 5.4 执行层
+
+- [ ] 新增 `execution-policy-gate.ts`
+- [ ] clarification / prompt-only / analysis-only 全部走快路径
+- [ ] timeout fallback 改成“保正确优先”
+- [ ] 错误统一映射为用户可理解消息
+
+### 5.5 前端层
+
+- [ ] 增加“本轮理解摘要”
+- [ ] 增加“任务切换提示”
+- [ ] Protocol Debug 增加模型槽位摘要
+- [ ] Protocol Debug 增加上下文裁剪摘要
+
+---
+
+## 6. 这次整改完成后，用户体验应该变成什么样
+
+### 场景 1：用户说“你好”
+
+系统行为：
+
+- 识别为 `greeting`
+- 不读取旧修图 workflow
+- 不做大规模图像分析
+- 1 秒内稳定回复
+
+### 场景 2：用户上传图片，说“帮我写这张图的提示词”
+
+系统行为：
+
+- 识别为 `prompt_only`
+- 只分析当前图片
+- 不进入生图/修图执行链
+- 输出提示词
+- 把提示词写入 prompt memory
+
+### 场景 3：用户说“按刚才那个提示词再来 4 张”
+
+系统行为：
+
+- 识别为 `image_generation_followup`
+- 复用上一轮提示词和参数
+- 明确出图数量为 4
+- 进入执行链
+
+### 场景 4：用户说“修改第 3 镜头，再确认”
+
+系统行为：
+
+- 识别为 `workflow_followup + revise`
+- 只进入 storyboard revision 子链路
+- 不误切普通生图
+
+---
+
+## 7. 最后一句话
+
+我们现在已经有：
+
+- protocol
+- subagent
+- workflow event
+- image/video memory
+- debug stream
+
+所以真正差的不是“没有 Agent”，而是：
+
+> **还没有把任务边界、记忆裁剪、模型链路、执行门控做成统一系统。**
+
+这份整改清单的目标，就是把 Agent 模式从：
+
+**很多聪明 patch**
+
+升级成：
+
+**真正稳定的统一 Agent Runtime**
+
+---
+
+## 8. 建议的下一步执行顺序
+
+如果马上开工，建议按下面顺序推进：
+
+1. 先做 **`task-boundary-classifier`**
+2. 再做 **memory relevance / context pruning**
+3. 再做 **provider health + model slot unification**
+4. 最后做 **前端理解摘要**
+
+因为只有先把任务边界做对，后面的“更聪明”才不会继续建立在错误理解之上。
+
--- a/lib/ai/conversation-memory-composer.ts
+++ b/lib/ai/conversation-memory-composer.ts
@ -3,6 +3,11 @@ import "server-only";
 import type { AgentThreadState } from "@/lib/ai/runtime/agent-thread-state";
 import type { ImageWorkflowMemory } from "@/lib/ai/image-workflow-memory";
 import type { PromptMemoryEntry } from "@/lib/ai/prompt-memory";
+import {
+  applyContextLineLimit,
+  applyMemoryHintLimit,
+  type MemoryRelevancePolicy,
+} from "@/lib/ai/runtime/memory-relevance";
 import type { StoryboardPreferenceMemory } from "@/lib/video-agent/storyboard-preference-memory";
 import type {
  ResolvedChatMediaInput,
@ -164,6 +169,7 @@ export function composeConversationMemory(params: {
  imageWorkflowMemory?: ImageWorkflowMemory | null;
  storyboardPreferenceMemory?: StoryboardPreferenceMemory | null;
  threadState?: AgentThreadState | null;
+  memoryRelevancePolicy?: MemoryRelevancePolicy;
 }) {
  const targetSummary = summarizeTargetResolution(params.targetResolution);
  const mediaSummary = summarizeResolvedMedia(params.resolvedMediaInput);
@ -197,8 +203,12 @@ export function composeConversationMemory(params: {
  };

  return {
-    contextSummary,
-    memoryHints,
+    contextSummary: params.memoryRelevancePolicy
+      ? applyContextLineLimit(contextSummary, params.memoryRelevancePolicy.maxContextLines)
+      : contextSummary,
+    memoryHints: params.memoryRelevancePolicy
+      ? applyMemoryHintLimit(memoryHints, params.memoryRelevancePolicy.maxMemoryHints)
+      : memoryHints,
    metadata,
  } satisfies ComposedConversationMemory;
 }
--- a/lib/ai/conversation-state-protocol.ts
+++ b/lib/ai/conversation-state-protocol.ts
@ -22,6 +22,11 @@ import {
  resolveConversationAction,
  extractPendingGridConfirmationFromWorkingState,
 } from "@/lib/ai/conversation-action-resolver";
+import {
+  applyContextLineLimit,
+  applyMemoryHintLimit,
+  resolveMemoryRelevancePolicy,
+} from "@/lib/ai/runtime/memory-relevance";
 import {
  composeConversationMemory,
  type ComposedConversationMemory,
@ -37,6 +42,10 @@ import {
  buildPromptMemoryUpdate,
 } from "@/lib/ai/prompt-memory";
 import type { AgentThreadState } from "@/lib/ai/runtime/agent-thread-state";
+import {
+  classifyTaskBoundary,
+  type TaskBoundaryDecision,
+} from "@/lib/ai/runtime/task-boundary-classifier";
 import type { StoryboardPreferenceMemory } from "@/lib/video-agent/storyboard-preference-memory";
 import {
  extractStoryboardPreferenceMemory,
@ -56,6 +65,7 @@ export interface ConversationStateProtocolBootstrap {
  pendingGridConfirmation: PendingGridConfirmation | null;
  confirmedPrompt: string | undefined;
  actionResolution: ReturnType<typeof resolveConversationAction>;
+  taskBoundaryDecision: TaskBoundaryDecision;
 }

 export interface ConversationStateProtocolComposed {
@ -181,6 +191,7 @@ export function bootstrapConversationStateProtocol(params: {
  } | null;
  activeStoryboardTask?: StoryboardTaskDetail | null;
  hasImageInputForPromptCarryForward: boolean;
+  hasCanvasSelection?: boolean;
 }) {
  const storyboardPreferenceMemory = extractStoryboardPreferenceMemory(params.workingState);
  const imageWorkflowMemory = extractImageWorkflowMemory(params.workingState);
@ -197,6 +208,14 @@ export function bootstrapConversationStateProtocol(params: {
    confirmedPrompt,
    hasImageInputForPromptCarryForward: params.hasImageInputForPromptCarryForward,
  });
+  const taskBoundaryDecision = classifyTaskBoundary({
+    text: params.text,
+    actionResolution,
+    hasCurrentImageInput: params.hasImageInputForPromptCarryForward,
+    hasCanvasSelection: params.hasCanvasSelection === true,
+    pendingGridConfirmation,
+    hasActiveStoryboardTask: Boolean(params.activeStoryboardTask?.task?.id),
+  });

  return {
    workingState: params.workingState || null,
@ -207,6 +226,7 @@ export function bootstrapConversationStateProtocol(params: {
    pendingGridConfirmation,
    confirmedPrompt,
    actionResolution,
+    taskBoundaryDecision,
  } satisfies ConversationStateProtocolBootstrap;
 }

@ -216,33 +236,62 @@ export function composeConversationStateProtocol(params: {
  resolvedMediaInput: ResolvedChatMediaInput;
  targetResolution: ResolvedTargetAssets;
 }) {
+  const memoryRelevancePolicy = resolveMemoryRelevancePolicy({
+    taskBoundaryDecision: params.bootstrap.taskBoundaryDecision,
+    actionResolution: params.bootstrap.actionResolution,
+    hasActiveStoryboardTask: Boolean(params.bootstrap.activeStoryboardTask?.task?.id),
+    hasPendingGridConfirmation: Boolean(params.bootstrap.pendingGridConfirmation),
+    hasCurrentImageInput: params.resolvedMediaInput.currentMessageImageUrls.length > 0,
+    hasCanvasSelection: params.resolvedMediaInput.selectedImageUrls.length > 0,
+  });
+
  const baseThreadState = buildConversationThreadState({
-    activeStoryboardTask: params.bootstrap.activeStoryboardTask,
-    storyboardPreferenceMemory: params.bootstrap.storyboardPreferenceMemory,
-    promptMemory: params.bootstrap.promptMemory,
-    imageWorkflowMemory: params.bootstrap.imageWorkflowMemory,
+    activeStoryboardTask: memoryRelevancePolicy.includeWorkflowState
+      ? params.bootstrap.activeStoryboardTask
+      : null,
+    storyboardPreferenceMemory: memoryRelevancePolicy.includeStoryboardPreferenceMemory
+      ? params.bootstrap.storyboardPreferenceMemory
+      : null,
+    promptMemory: memoryRelevancePolicy.includePromptMemory
+      ? params.bootstrap.promptMemory
+      : null,
+    imageWorkflowMemory: memoryRelevancePolicy.includeImageWorkflowMemory
+      ? params.bootstrap.imageWorkflowMemory
+      : null,
  });

  const composedConversationMemory = composeConversationMemory({
    autoContextText: params.autoContextText || undefined,
    resolvedMediaInput: params.resolvedMediaInput,
    targetResolution: params.targetResolution,
-    promptMemory: params.bootstrap.promptMemory,
-    imageWorkflowMemory: params.bootstrap.imageWorkflowMemory,
-    storyboardPreferenceMemory: params.bootstrap.storyboardPreferenceMemory,
+    promptMemory: memoryRelevancePolicy.includePromptMemory
+      ? params.bootstrap.promptMemory
+      : null,
+    imageWorkflowMemory: memoryRelevancePolicy.includeImageWorkflowMemory
+      ? params.bootstrap.imageWorkflowMemory
+      : null,
+    storyboardPreferenceMemory: memoryRelevancePolicy.includeStoryboardPreferenceMemory
+      ? params.bootstrap.storyboardPreferenceMemory
+      : null,
    threadState: baseThreadState,
+    memoryRelevancePolicy,
  });

  const threadState: AgentThreadState = {
    workflowState: baseThreadState?.workflowState || null,
-    memoryHints: Array.from(
+    memoryHints: applyMemoryHintLimit(
+      Array.from(
      new Set([
        ...(baseThreadState?.memoryHints || []),
        ...composedConversationMemory.memoryHints,
      ])
    ),
-    contextSummary:
+      memoryRelevancePolicy.maxMemoryHints,
+    ),
+    contextSummary: applyContextLineLimit(
      composedConversationMemory.contextSummary || baseThreadState?.contextSummary,
+      memoryRelevancePolicy.maxContextLines,
+    ),
    metadata: {
      ...(baseThreadState?.metadata || {}),
      ...composedConversationMemory.metadata,
--- a/lib/ai/runtime/agent-model-slots.ts
+++ b/lib/ai/runtime/agent-model-slots.ts
@ -0,0 +1,108 @@
+import "server-only";
+
+import {
+  getAiModelById,
+  getAiProviderById,
+  getEnabledAiModelsWithProvider,
+} from "@/lib/db/queries";
+
+export type AgentModelSlotName = "planner" | "analyzer" | "reply" | "execution";
+
+export interface AgentModelSlotDescriptor {
+  modelId: string;
+  displayName?: string;
+  providerId?: string;
+  providerType?: string;
+  providerKey?: string;
+}
+
+export interface AgentModelSlot extends AgentModelSlotDescriptor {
+  slot: AgentModelSlotName;
+  source: string;
+  reason?: string;
+  locked?: boolean;
+}
+
+export interface AgentModelSlots {
+  planner?: AgentModelSlot;
+  analyzer?: AgentModelSlot;
+  reply?: AgentModelSlot;
+  execution?: AgentModelSlot;
+}
+
+function buildProviderKey(params: {
+  providerId?: string;
+  providerType?: string;
+  modelId: string;
+}) {
+  if (params.providerId) return params.providerId;
+  if (params.providerType) return `${params.providerType}:${params.modelId}`;
+  return `model:${params.modelId}`;
+}
+
+export async function resolveAgentModelSlotDescriptor(modelId: string) {
+  const normalizedModelId = String(modelId || "").trim();
+  if (!normalizedModelId) return null;
+
+  const isUuid =
+    /^[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}$/i.test(normalizedModelId);
+
+  if (isUuid) {
+    const model = await getAiModelById({ id: normalizedModelId }).catch(() => null);
+    if (!model) {
+      return {
+        modelId: normalizedModelId,
+        providerKey: `model:${normalizedModelId}`,
+      } satisfies AgentModelSlotDescriptor;
+    }
+    const provider = await getAiProviderById({ id: model.providerId }).catch(() => null);
+    return {
+      modelId: model.id,
+      displayName: model.displayName || model.modelId,
+      providerId: model.providerId,
+      providerType: provider?.type || undefined,
+      providerKey: buildProviderKey({
+        providerId: model.providerId,
+        providerType: provider?.type || undefined,
+        modelId: model.id,
+      }),
+    } satisfies AgentModelSlotDescriptor;
+  }
+
+  const enabledModels = await getEnabledAiModelsWithProvider(false).catch(() => []);
+  const matched = enabledModels.find(
+    ({ model }) => model.id === normalizedModelId || model.modelId === normalizedModelId,
+  );
+  if (!matched) {
+    return {
+      modelId: normalizedModelId,
+      providerKey: `model:${normalizedModelId}`,
+    } satisfies AgentModelSlotDescriptor;
+  }
+
+  return {
+    modelId: matched.model.id,
+    displayName: matched.model.displayName || matched.model.modelId,
+    providerId: matched.model.providerId,
+    providerType: matched.provider?.type || undefined,
+    providerKey: buildProviderKey({
+      providerId: matched.model.providerId,
+      providerType: matched.provider?.type || undefined,
+      modelId: matched.model.id,
+    }),
+  } satisfies AgentModelSlotDescriptor;
+}
+
+export function upsertAgentModelSlot(
+  slots: AgentModelSlots,
+  slot: AgentModelSlotName,
+  value: Omit<AgentModelSlot, "slot">,
+) {
+  return {
+    ...slots,
+    [slot]: {
+      slot,
+      ...value,
+    },
+  } satisfies AgentModelSlots;
+}
--- a/lib/ai/runtime/memory-relevance.ts
+++ b/lib/ai/runtime/memory-relevance.ts
@ -0,0 +1,134 @@
+import "server-only";
+
+import type { ConversationActionResolution } from "@/lib/ai/conversation-action-resolver";
+import type { TaskBoundaryDecision } from "@/lib/ai/runtime/task-boundary-classifier";
+
+export interface MemoryRelevancePolicy {
+  includePromptMemory: boolean;
+  includeImageWorkflowMemory: boolean;
+  includeStoryboardPreferenceMemory: boolean;
+  includeWorkflowState: boolean;
+  maxMemoryHints: number;
+  maxContextLines: number;
+  reason: string;
+}
+
+export function resolveMemoryRelevancePolicy(params: {
+  taskBoundaryDecision: TaskBoundaryDecision;
+  actionResolution: ConversationActionResolution;
+  hasActiveStoryboardTask: boolean;
+  hasPendingGridConfirmation: boolean;
+  hasCurrentImageInput: boolean;
+  hasCanvasSelection: boolean;
+}) {
+  const { taskBoundaryDecision, actionResolution } = params;
+
+  switch (taskBoundaryDecision.taskType) {
+    case "greeting":
+      return {
+        includePromptMemory: false,
+        includeImageWorkflowMemory: false,
+        includeStoryboardPreferenceMemory: false,
+        includeWorkflowState: false,
+        maxMemoryHints: 0,
+        maxContextLines: 2,
+        reason: "greeting_turn_ignores_task_memory",
+      } satisfies MemoryRelevancePolicy;
+    case "chat":
+      return {
+        includePromptMemory: false,
+        includeImageWorkflowMemory: false,
+        includeStoryboardPreferenceMemory: false,
+        includeWorkflowState: false,
+        maxMemoryHints: 1,
+        maxContextLines: 3,
+        reason: "chat_turn_prefers_clean_context",
+      } satisfies MemoryRelevancePolicy;
+    case "prompt_only":
+      return {
+        includePromptMemory: false,
+        includeImageWorkflowMemory: false,
+        includeStoryboardPreferenceMemory: false,
+        includeWorkflowState: false,
+        maxMemoryHints: 2,
+        maxContextLines: 4,
+        reason: "prompt_only_turn_ignores_execution_memory",
+      } satisfies MemoryRelevancePolicy;
+    case "analysis_only":
+      return {
+        includePromptMemory: false,
+        includeImageWorkflowMemory: false,
+        includeStoryboardPreferenceMemory: false,
+        includeWorkflowState: false,
+        maxMemoryHints: 2,
+        maxContextLines: 4,
+        reason: "analysis_only_turn_ignores_workflow_memory",
+      } satisfies MemoryRelevancePolicy;
+    case "execution_video":
+      return {
+        includePromptMemory: false,
+        includeImageWorkflowMemory: false,
+        includeStoryboardPreferenceMemory: true,
+        includeWorkflowState:
+          params.hasActiveStoryboardTask || params.hasPendingGridConfirmation,
+        maxMemoryHints: 4,
+        maxContextLines: 6,
+        reason: "video_execution_prefers_storyboard_preferences",
+      } satisfies MemoryRelevancePolicy;
+    case "workflow_followup":
+      return {
+        includePromptMemory:
+          actionResolution.kind === "image_generation_followup" ||
+          actionResolution.kind === "prompt_carry_forward",
+        includeImageWorkflowMemory:
+          actionResolution.kind === "image_edit_followup" ||
+          actionResolution.kind === "image_generation_followup" ||
+          actionResolution.kind === "prompt_carry_forward",
+        includeStoryboardPreferenceMemory:
+          actionResolution.kind === "grid_confirmation_followup" ||
+          actionResolution.kind === "active_storyboard_followup",
+        includeWorkflowState: true,
+        maxMemoryHints: 6,
+        maxContextLines: 8,
+        reason: "workflow_followup_requires_workflow_memory",
+      } satisfies MemoryRelevancePolicy;
+    case "execution_image":
+    default:
+      return {
+        includePromptMemory:
+          actionResolution.usePromptCarryForward ||
+          taskBoundaryDecision.taskMode === "continue",
+        includeImageWorkflowMemory:
+          taskBoundaryDecision.taskMode === "continue" ||
+          (!params.hasCurrentImageInput && !params.hasCanvasSelection),
+        includeStoryboardPreferenceMemory: false,
+        includeWorkflowState:
+          taskBoundaryDecision.taskMode === "continue" &&
+          (actionResolution.kind === "image_edit_followup" ||
+            actionResolution.kind === "image_generation_followup" ||
+            actionResolution.kind === "prompt_carry_forward"),
+        maxMemoryHints: taskBoundaryDecision.taskMode === "continue" ? 5 : 3,
+        maxContextLines: taskBoundaryDecision.taskMode === "continue" ? 7 : 5,
+        reason:
+          taskBoundaryDecision.taskMode === "continue"
+            ? "image_execution_continue_reuses_relevant_memory"
+            : "image_execution_new_task_limits_old_memory",
+      } satisfies MemoryRelevancePolicy;
+  }
+}
+
+export function applyMemoryHintLimit(hints: string[], maxHints: number) {
+  if (!Array.isArray(hints) || maxHints <= 0) return [] as string[];
+  return hints.slice(0, maxHints);
+}
+
+export function applyContextLineLimit(summary: string | undefined, maxLines: number) {
+  const normalized = String(summary || "").trim();
+  if (!normalized || maxLines <= 0) return "";
+  return normalized
+    .split("\n")
+    .map((line) => line.trim())
+    .filter(Boolean)
+    .slice(0, maxLines)
+    .join("\n");
+}
--- a/lib/ai/runtime/provider-health.ts
+++ b/lib/ai/runtime/provider-health.ts
@ -0,0 +1,109 @@
+import "server-only";
+
+export type ProviderHealthEventType = "success" | "timeout" | "quota" | "error";
+
+export interface ProviderHealthSnapshot {
+  providerKey: string;
+  providerType?: string;
+  lastEventType: ProviderHealthEventType;
+  lastError?: string;
+  consecutiveFailures: number;
+  lastSuccessAt?: number;
+  updatedAt: number;
+}
+
+const providerHealthMap = new Map<string, ProviderHealthSnapshot>();
+
+function normalizeErrorMessage(error: unknown) {
+  return error instanceof Error ? String(error.message || "").trim() : String(error || "").trim();
+}
+
+export function classifyProviderHealthEvent(error: unknown): ProviderHealthEventType {
+  const normalized = normalizeErrorMessage(error).toLowerCase();
+
+  if (
+    normalized.includes("token quota is not enough") ||
+    normalized.includes("pre_consume_token_quota_failed") ||
+    normalized.includes("quota is not enough")
+  ) {
+    return "quota";
+  }
+
+  if (
+    normalized.includes("timeout") ||
+    normalized.includes("timed out") ||
+    normalized.includes("body timeout") ||
+    normalized.includes("headers timeout")
+  ) {
+    return "timeout";
+  }
+
+  return "error";
+}
+
+export function recordProviderHealthSuccess(params: {
+  providerKey?: string;
+  providerType?: string;
+}) {
+  const providerKey = String(params.providerKey || "").trim();
+  if (!providerKey) return;
+
+  const existing = providerHealthMap.get(providerKey);
+  providerHealthMap.set(providerKey, {
+    providerKey,
+    providerType: params.providerType || existing?.providerType,
+    lastEventType: "success",
+    consecutiveFailures: 0,
+    lastError: undefined,
+    lastSuccessAt: Date.now(),
+    updatedAt: Date.now(),
+  });
+}
+
+export function recordProviderHealthFailure(params: {
+  providerKey?: string;
+  providerType?: string;
+  error: unknown;
+}) {
+  const providerKey = String(params.providerKey || "").trim();
+  if (!providerKey) return;
+
+  const existing = providerHealthMap.get(providerKey);
+  providerHealthMap.set(providerKey, {
+    providerKey,
+    providerType: params.providerType || existing?.providerType,
+    lastEventType: classifyProviderHealthEvent(params.error),
+    lastError: normalizeErrorMessage(params.error).slice(0, 400),
+    consecutiveFailures:
+      existing?.lastEventType === "success"
+        ? 1
+        : Math.max(1, (existing?.consecutiveFailures || 0) + 1),
+    lastSuccessAt: existing?.lastSuccessAt,
+    updatedAt: Date.now(),
+  });
+}
+
+export function getProviderHealthSnapshot(providerKey?: string) {
+  const normalized = String(providerKey || "").trim();
+  if (!normalized) return null;
+  return providerHealthMap.get(normalized) || null;
+}
+
+export function shouldAvoidProvider(snapshot?: ProviderHealthSnapshot | null) {
+  if (!snapshot) return false;
+
+  const ageMs = Date.now() - snapshot.updatedAt;
+  if (snapshot.lastEventType === "quota") {
+    return ageMs <= 15 * 60 * 1000;
+  }
+
+  if (snapshot.lastEventType === "timeout") {
+    return snapshot.consecutiveFailures >= 2 && ageMs <= 5 * 60 * 1000;
+  }
+
+  if (snapshot.lastEventType === "error") {
+    return snapshot.consecutiveFailures >= 3 && ageMs <= 5 * 60 * 1000;
+  }
+
+  return false;
+}
--- a/lib/ai/runtime/task-boundary-classifier.ts
+++ b/lib/ai/runtime/task-boundary-classifier.ts
@ -0,0 +1,226 @@
+import "server-only";
+
+import type {
+  ConversationActionResolution,
+  PendingGridConfirmation,
+} from "@/lib/ai/conversation-action-resolver";
+import { isExplicitPromptOptimizationRequest } from "@/lib/ai/smart-analyzer";
+import { detectStoryboardCommandIntent } from "@/lib/video-agent/storyboard-command-interpreter";
+
+export type TaskBoundaryType =
+  | "greeting"
+  | "chat"
+  | "analysis_only"
+  | "prompt_only"
+  | "execution_image"
+  | "execution_video"
+  | "workflow_followup";
+
+export type TaskBoundaryMode = "new" | "continue" | "revise" | "confirm" | "cancel";
+
+export interface TaskBoundaryDecision {
+  taskType: TaskBoundaryType;
+  taskMode: TaskBoundaryMode;
+  shouldUseCanvasContext: boolean;
+  shouldUseHistoryImages: boolean;
+  shouldUseWorkflowMemory: boolean;
+  shouldAllowExecution: boolean;
+  confidence: number;
+  reason: string;
+}
+
+function normalizeCompactText(text: string) {
+  return String(text || "").toLowerCase().replace(/\s+/g, "").trim();
+}
+
+export function isPureGreetingTurnRequest(text: string): boolean {
+  const normalized = normalizeCompactText(text);
+  if (!normalized) return false;
+
+  return /^(你好|您好|hello|hi|嗨|哈喽|在吗|早上好|中午好|下午好|晚上好|深夜好|谢谢|感谢|收到|好的|ok|okay|明白了|了解了)$/.test(
+    normalized,
+  );
+}
+
+function isLikelyVideoExecutionRequest(text: string): boolean {
+  const normalized = normalizeCompactText(text);
+  if (!normalized) return false;
+
+  return /图生视频|文生视频|视频生成|生成视频|做视频|做成视频|开始生成视频|让.*动起来|变成视频|视频/.test(
+    normalized,
+  );
+}
+
+function isLikelyImageExecutionRequest(text: string): boolean {
+  const normalized = normalizeCompactText(text);
+  if (!normalized) return false;
+
+  return /生图|生成图片|出图|画一张|帮我画|做图|做一张图|重新生成|再来一张|执行生成|开始生成|生成出来/.test(
+    normalized,
+  );
+}
+
+function isLikelyAnalysisOnlyRequest(text: string, hasImageContext: boolean): boolean {
+  if (!hasImageContext) return false;
+  const normalized = normalizeCompactText(text);
+  if (!normalized) return false;
+
+  return /分析|看图|看看|识别|拆解|提炼|整理|归纳|总结|描述|解释|判断|帮我看看|帮我分析|帮我判断|是什么|适合吗|怎么样|评价/.test(
+    normalized,
+  );
+}
+
+function resolveStoryboardTaskMode(params: {
+  text: string;
+  hasPendingGridConfirmation: boolean;
+  hasActiveStoryboardTask: boolean;
+}): TaskBoundaryMode {
+  if (!params.hasPendingGridConfirmation && !params.hasActiveStoryboardTask) {
+    return "continue";
+  }
+
+  const intent = detectStoryboardCommandIntent(params.text);
+  switch (intent) {
+    case "confirm":
+      return "confirm";
+    case "revise":
+      return "revise";
+    case "cancel":
+      return "cancel";
+    case "unclear":
+    default:
+      return "continue";
+  }
+}
+
+export function classifyTaskBoundary(params: {
+  text: string;
+  actionResolution: ConversationActionResolution;
+  hasCurrentImageInput: boolean;
+  hasCanvasSelection: boolean;
+  pendingGridConfirmation?: PendingGridConfirmation | null;
+  hasActiveStoryboardTask: boolean;
+}) {
+  const hasImageContext = params.hasCurrentImageInput || params.hasCanvasSelection;
+
+  if (isPureGreetingTurnRequest(params.text) && !params.hasCurrentImageInput) {
+    return {
+      taskType: "greeting",
+      taskMode: "new",
+      shouldUseCanvasContext: false,
+      shouldUseHistoryImages: false,
+      shouldUseWorkflowMemory: false,
+      shouldAllowExecution: false,
+      confidence: 0.99,
+      reason: "pure_greeting_turn",
+    } satisfies TaskBoundaryDecision;
+  }
+
+  if (isExplicitPromptOptimizationRequest(params.text)) {
+    return {
+      taskType: "prompt_only",
+      taskMode: "new",
+      shouldUseCanvasContext: !params.hasCurrentImageInput && params.hasCanvasSelection,
+      shouldUseHistoryImages: !params.hasCurrentImageInput && !params.hasCanvasSelection,
+      shouldUseWorkflowMemory: false,
+      shouldAllowExecution: false,
+      confidence: 0.97,
+      reason: "explicit_prompt_only_request",
+    } satisfies TaskBoundaryDecision;
+  }
+
+  switch (params.actionResolution.kind) {
+    case "grid_confirmation_followup":
+    case "active_storyboard_followup":
+      return {
+        taskType: "workflow_followup",
+        taskMode: resolveStoryboardTaskMode({
+          text: params.text,
+          hasPendingGridConfirmation: Boolean(params.pendingGridConfirmation),
+          hasActiveStoryboardTask: params.hasActiveStoryboardTask,
+        }),
+        shouldUseCanvasContext: params.hasCanvasSelection,
+        shouldUseHistoryImages: !params.hasCurrentImageInput,
+        shouldUseWorkflowMemory: true,
+        shouldAllowExecution: true,
+        confidence: 0.95,
+        reason: params.actionResolution.kind,
+      } satisfies TaskBoundaryDecision;
+    case "image_edit_followup":
+    case "image_generation_followup":
+      return {
+        taskType: "workflow_followup",
+        taskMode: "continue",
+        shouldUseCanvasContext: params.hasCanvasSelection,
+        shouldUseHistoryImages: !params.hasCurrentImageInput,
+        shouldUseWorkflowMemory: true,
+        shouldAllowExecution: true,
+        confidence: 0.92,
+        reason: params.actionResolution.kind,
+      } satisfies TaskBoundaryDecision;
+    case "prompt_carry_forward":
+      return {
+        taskType: "execution_image",
+        taskMode: "continue",
+        shouldUseCanvasContext: params.hasCanvasSelection,
+        shouldUseHistoryImages: !params.hasCurrentImageInput,
+        shouldUseWorkflowMemory: true,
+        shouldAllowExecution: true,
+        confidence: 0.93,
+        reason: "prompt_carry_forward_execution",
+      } satisfies TaskBoundaryDecision;
+    case "default":
+    default:
+      break;
+  }
+
+  if (isLikelyVideoExecutionRequest(params.text)) {
+    return {
+      taskType: "execution_video",
+      taskMode: "new",
+      shouldUseCanvasContext: params.hasCanvasSelection,
+      shouldUseHistoryImages: false,
+      shouldUseWorkflowMemory: false,
+      shouldAllowExecution: true,
+      confidence: 0.88,
+      reason: "explicit_video_execution_request",
+    } satisfies TaskBoundaryDecision;
+  }
+
+  if (isLikelyImageExecutionRequest(params.text)) {
+    return {
+      taskType: "execution_image",
+      taskMode: "new",
+      shouldUseCanvasContext: params.hasCanvasSelection,
+      shouldUseHistoryImages: false,
+      shouldUseWorkflowMemory: false,
+      shouldAllowExecution: true,
+      confidence: 0.87,
+      reason: "explicit_image_execution_request",
+    } satisfies TaskBoundaryDecision;
+  }
+
+  if (isLikelyAnalysisOnlyRequest(params.text, hasImageContext)) {
+    return {
+      taskType: "analysis_only",
+      taskMode: "new",
+      shouldUseCanvasContext: params.hasCanvasSelection,
+      shouldUseHistoryImages: !params.hasCurrentImageInput && !params.hasCanvasSelection,
+      shouldUseWorkflowMemory: false,
+      shouldAllowExecution: false,
+      confidence: 0.9,
+      reason: "analysis_only_request",
+    } satisfies TaskBoundaryDecision;
+  }
+
+  return {
+    taskType: "chat",
+    taskMode: "new",
+    shouldUseCanvasContext: false,
+    shouldUseHistoryImages: false,
+    shouldUseWorkflowMemory: false,
+    shouldAllowExecution: false,
+    confidence: 0.7,
+    reason: "default_chat_turn",
+  } satisfies TaskBoundaryDecision;
+}
--- a/lib/ai/runtime/task-boundary-context-pruner.ts
+++ b/lib/ai/runtime/task-boundary-context-pruner.ts
@ -0,0 +1,169 @@
+import "server-only";
+
+import type { DBMessage } from "@/lib/db/schema";
+import type {
+  CanvasImageNodeContext,
+  ContextPayload,
+  WorkingStateImageMemory,
+} from "@/lib/ai/context-memory/types";
+import type { TaskBoundaryDecision } from "@/lib/ai/runtime/task-boundary-classifier";
+
+function buildSelectedCanvasSubset(contextPayload?: ContextPayload) {
+  const selectedNodeIds = new Set(contextPayload?.selectedNodeIds || []);
+  const selectedAssetIds = new Set(contextPayload?.selectedAssetIds || []);
+  const selectedNodes = (contextPayload?.canvasImageNodes || []).filter((node) => {
+    if (!node) return false;
+    if (node.isSelected) return true;
+    if (node.nodeId && selectedNodeIds.has(node.nodeId)) return true;
+    if (node.assetId && selectedAssetIds.has(node.assetId)) return true;
+    return false;
+  });
+
+  return {
+    selectedNodes,
+    selectedNodeIds: selectedNodes.map((node) => node.nodeId).filter(Boolean),
+    selectedAssetIds: selectedNodes.map((node) => node.assetId).filter(Boolean) as string[],
+  };
+}
+
+function buildSelectedCanvasSummary(selectedNodes: CanvasImageNodeContext[]) {
+  if (selectedNodes.length === 0) return undefined;
+  const previewLines = selectedNodes.slice(0, 3).map((node, index) => {
+    const title = String(node.title || node.prompt || node.nodeId || `图片${index + 1}`).trim();
+    return `选中图片${index + 1}: ${title}`;
+  });
+  return [`用户已选中 ${selectedNodes.length} 张图片，优先处理这些`, ...previewLines].join("\n");
+}
+
+export function pruneTaskBoundaryContext(params: {
+  decision: TaskBoundaryDecision;
+  contextPayload?: ContextPayload;
+  autoContextText?: string;
+  workingState?: WorkingStateImageMemory | null;
+  hasCurrentImageInput: boolean;
+}) {
+  if (!params.contextPayload && !params.autoContextText) {
+    return {
+      contextPayload: params.contextPayload,
+      autoContextText: params.autoContextText,
+      workingState: params.decision.shouldUseWorkflowMemory ? params.workingState || null : null,
+    };
+  }
+
+  const selectedSubset = buildSelectedCanvasSubset(params.contextPayload);
+  const shouldPreferSelectedOnly =
+    selectedSubset.selectedNodes.length > 0 &&
+    (params.decision.taskType === "prompt_only" ||
+      params.decision.taskType === "analysis_only" ||
+      params.decision.taskType === "execution_image" ||
+      params.decision.taskType === "execution_video" ||
+      params.hasCurrentImageInput);
+
+  let nextContextPayload = params.contextPayload;
+  let nextAutoContextText = params.autoContextText;
+
+  if (!params.decision.shouldUseCanvasContext) {
+    nextContextPayload = params.contextPayload
+      ? {
+          source: params.contextPayload.source,
+          projectId: params.contextPayload.projectId,
+          focusPolicy: params.contextPayload.focusPolicy,
+          selectedNodeIds: [],
+          selectedAssetIds: [],
+          canvasImageNodes: [],
+        }
+      : undefined;
+    nextAutoContextText = undefined;
+  } else if (shouldPreferSelectedOnly) {
+    nextContextPayload = params.contextPayload
+      ? {
+          ...params.contextPayload,
+          focusPolicy: "selected",
+          selectedNodeIds: selectedSubset.selectedNodeIds,
+          selectedAssetIds: selectedSubset.selectedAssetIds,
+          canvasImageNodes: selectedSubset.selectedNodes,
+        }
+      : undefined;
+    nextAutoContextText = buildSelectedCanvasSummary(selectedSubset.selectedNodes);
+  }
+
+  return {
+    contextPayload: nextContextPayload,
+    autoContextText: nextAutoContextText,
+    workingState: params.decision.shouldUseWorkflowMemory ? params.workingState || null : null,
+  };
+}
+
+function sanitizeHistoryText(text: string) {
+  return String(text || "")
+    .replace(/!\[[^\]]*]\([^)]+\)/g, "")
+    .replace(/<video[^>]*>[\s\S]*?<\/video>/gi, "")
+    .replace(/<video[^>]+src="[^"]+"[^>]*>/gi, "")
+    .replace(/<!--[\s\S]*?-->/g, "")
+    .replace(/◈\s*[^\n]+\n*/g, "")
+    .replace(/\n{3,}/g, "\n\n")
+    .trim();
+}
+
+function toTextOnlyHistoryMessages(messages: DBMessage[], maxMessages: number): DBMessage[] {
+  const sliced = messages.slice(-Math.max(1, maxMessages));
+  const result: DBMessage[] = [];
+
+  for (const message of sliced) {
+    const parts = Array.isArray(message.parts) ? (message.parts as Array<Record<string, unknown>>) : [];
+    const textParts = parts
+      .filter((part) => part?.type === "text" && typeof part?.text === "string")
+      .map((part) => ({
+        type: "text",
+        text: sanitizeHistoryText(String(part.text || "")),
+      }))
+      .filter((part) => part.text);
+
+    if (textParts.length === 0) continue;
+
+    result.push({
+      ...message,
+      parts: textParts,
+      attachments: [],
+    } satisfies DBMessage);
+  }
+
+  return result;
+}
+
+export function pruneTaskBoundaryHistoryMessages(params: {
+  decision: TaskBoundaryDecision;
+  messages: DBMessage[];
+  purpose: "media" | "model";
+  hasCurrentImageInput: boolean;
+}): DBMessage[] {
+  const messages = Array.isArray(params.messages) ? params.messages : [];
+  if (messages.length === 0) return [] as DBMessage[];
+
+  switch (params.decision.taskType) {
+    case "greeting":
+      return [];
+    case "prompt_only":
+      return params.purpose === "media"
+        ? []
+        : toTextOnlyHistoryMessages(messages, params.hasCurrentImageInput ? 2 : 4);
+    case "analysis_only":
+      return params.purpose === "media"
+        ? []
+        : toTextOnlyHistoryMessages(messages, 4);
+    case "chat":
+      return params.purpose === "media"
+        ? []
+        : toTextOnlyHistoryMessages(messages, 6);
+    case "execution_image":
+    case "execution_video":
+      if (params.decision.shouldUseHistoryImages) {
+        return messages.slice(-8);
+      }
+      return params.purpose === "media" ? [] : toTextOnlyHistoryMessages(messages, 4);
+    case "workflow_followup":
+      return messages.slice(-(params.purpose === "media" ? 8 : 12));
+    default:
+      return params.purpose === "media" ? [] : toTextOnlyHistoryMessages(messages, 4);
+  }
+}