fix: keep agent text image routing safe

components/canvas/CanvasChatPanel.tsx

@@ -2394,6 +2394,15 @@ export function CanvasChatPanel({
       return next;
     });
   }, []);
+  const clearSettledAgentProgressStateForChat = useCallback((chatSessionId: string) => {
+    setAgentProgressStateByChatId((prev) => {
+      const current = prev[chatSessionId];
+      if (!current || current.stage === "clarify") return prev;
+      const next = { ...prev };
+      delete next[chatSessionId];
+      return next;
+    });
+  }, []);
 
   const shouldBlockDuplicateSend = useCallback((chatSessionId: string, signature: string) => {
     const now = Date.now();
@@ -6813,6 +6822,7 @@ export function CanvasChatPanel({
       } catch (error) {
         console.error("[CanvasChat] Failed to sync messages after finish:", error);
       } finally {
+        clearSettledAgentProgressStateForChat(finishedChatId);
         if (inFlightChatIdRef.current === finishedChatId) {
           inFlightChatIdRef.current = null;
           inFlightRawChatIdRef.current = null;
@@ -12143,7 +12153,10 @@ export function CanvasChatPanel({
   const isAgentTargetClarificationActive =
     shouldShowAgentTargetLockCard && activeAgentTargetLock?.stage === "clarify";
   const shouldShowAgentProgressCard =
-    activeComposerMode === "agent" && Boolean(activeAgentProgress);
+    activeComposerMode === "agent" &&
+    Boolean(activeAgentProgress) &&
+    activeAgentProgress?.stage !== "completed" &&
+    activeAgentProgress?.stage !== "failed";
   const agentTargetLockPreviewItems = useMemo(() => {
     if (!activeAgentTargetLock) return [] as AgentTargetLockRefPreview[];
     const targets = Array.isArray(activeAgentTargetLock.targets) ? activeAgentTargetLock.targets : [];

lib/ai/model-classification.ts

@@ -49,10 +49,55 @@ const VIDEO_MODE_KEYS = new Set([
   "reference-images",
 ]);
 
+const TEXT_TO_IMAGE_MODE_KEYS = new Set([
+  "text-to-image",
+  "text2image",
+  "txt2img",
+  "image-generation",
+  "image_generate",
+  "generate-image",
+]);
+
+const IMAGE_EDIT_ONLY_KIND_KEYS = new Set([
+  "image-edit",
+  "image-to-image",
+  "img2img",
+  "reference-image",
+  "reference-images",
+]);
+
+const TEXT_TO_IMAGE_KIND_KEYS = new Set([
+  "text-to-image",
+  "text2image",
+  "txt2img",
+  "image",
+  "image-generation",
+  "image_generate",
+]);
+
 function normalizeText(value?: string | null) {
   return String(value || "").trim().toLowerCase();
 }
 
+function toRecord(value: unknown) {
+  if (!value || typeof value !== "object" || Array.isArray(value)) {
+    return null;
+  }
+  return value as Record<string, unknown>;
+}
+
+function getStringArrayField(source: unknown, key: string) {
+  const record = toRecord(source);
+  const value = record?.[key];
+  if (!Array.isArray(value)) return [] as string[];
+  return value.map((item) => normalizeText(String(item || ""))).filter(Boolean);
+}
+
+function getStringField(source: unknown, key: string) {
+  const record = toRecord(source);
+  return normalizeText(String(record?.[key] || ""));
+}
+
 function hasVideoModes(capabilities: unknown) {
   if (!capabilities || typeof capabilities !== "object") return false;
   const supportedModes = (capabilities as { supportedModes?: unknown }).supportedModes;
@@ -131,3 +176,63 @@ export function isImageGenerationModelDescriptor(params: {
 
   return STATIC_IMAGE_PROVIDER_TYPES.has(providerType);
 }
+
+export function modelSupportsTextToImageDescriptor(params: {
+  providerType?: string | null;
+  modelId?: string | null;
+  capabilities?: unknown;
+  defaultParams?: unknown;
+  supportsImageGen?: boolean | null;
+}): boolean {
+  if (
+    !isImageGenerationModelDescriptor({
+      providerType: params.providerType,
+      modelId: params.modelId,
+      capabilities: params.capabilities,
+      supportsImageGen: params.supportsImageGen,
+    })
+  ) {
+    return false;
+  }
+
+  const providerType = normalizeText(params.providerType);
+  const modelId = normalizeText(params.modelId);
+  const capabilityModes = getStringArrayField(params.capabilities, "supportedModes");
+  const defaultModes = getStringArrayField(params.defaultParams, "supportedModes");
+  const modes = new Set([...capabilityModes, ...defaultModes]);
+
+  if ([...modes].some((mode) => TEXT_TO_IMAGE_MODE_KEYS.has(mode))) {
+    return true;
+  }
+
+  const kind =
+    getStringField(params.capabilities, "kind") ||
+    getStringField(params.defaultParams, "kind");
+  if (TEXT_TO_IMAGE_KIND_KEYS.has(kind)) {
+    return true;
+  }
+
+  if (providerType === "stepfun-image") {
+    return true;
+  }
+
+  const runtimeEndpoint =
+    getStringField(params.capabilities, "endpoint") ||
+    getStringField(params.defaultParams, "endpoint");
+  const runtimeModelText = `${modelId} ${runtimeEndpoint}`.trim();
+
+  if (
+    IMAGE_EDIT_ONLY_KIND_KEYS.has(kind) ||
+    [...modes].some((mode) => IMAGE_EDIT_ONLY_KIND_KEYS.has(mode)) ||
+    runtimeModelText.endsWith("/edit") ||
+    runtimeModelText.includes("image-edit")
+  ) {
+    return false;
+  }
+
+  if (providerType === "alibaba-image") {
+    return false;
+  }
+
+  return true;
+}

lib/ai/model-router.ts

@@ -8,6 +8,7 @@
 import { isAiModelMultimodal } from "@/lib/db/ai-model-flags";
 import {
   isImageGenerationModelDescriptor,
+  modelSupportsTextToImageDescriptor,
   isVideoModelDescriptor,
 } from "@/lib/ai/model-classification";
 import { isKnownReferenceImageModel } from "@/lib/ai/reference-image-capability";
@@ -137,12 +138,20 @@ function matchesIntentRequirements(
   if (params.requireVision && !hasVisionCapability(item.model, item.provider.type)) return false;
   if (
     params.requireImageGen &&
-    !isImageGenerationModelDescriptor({
+    !(params.intent === "image_gen"
+      ? modelSupportsTextToImageDescriptor({
+          providerType: item.provider.type,
+          modelId: item.model.modelId,
+          capabilities: item.model.capabilities,
+          defaultParams: item.model.defaultParams,
+          supportsImageGen: item.model.supportsImageGen,
+        })
+      : isImageGenerationModelDescriptor({
       providerType: item.provider.type,
       modelId: item.model.modelId,
       capabilities: item.model.capabilities,
       supportsImageGen: item.model.supportsImageGen,
-    })
+        }))
   ) {
     return false;
   }
@@ -280,9 +289,20 @@ export async function routeToModel(params: {
       const matchesVision =
         !config.requireVision ||
         hasVisionCapability(configuredModel.model, configuredModel.provider.type);
-      const matchesImageGen = !config.requireImageGen || configuredModel.model.supportsImageGen;
+      const configuredModelSupportsRequestedImageGeneration =
+        intent === "image_gen"
+          ? modelSupportsTextToImageDescriptor({
+              providerType: configuredModel.provider.type,
+              modelId: configuredModel.model.modelId,
+              capabilities: configuredModel.model.capabilities,
+              defaultParams: configuredModel.model.defaultParams,
+              supportsImageGen: configuredModel.model.supportsImageGen,
+            })
+          : configuredModel.model.supportsImageGen;
+      const matchesImageGen =
+        !config.requireImageGen || configuredModelSupportsRequestedImageGeneration;
       const allowImageIntentWithImageGenOnly =
-        isImageIntent && configuredModel.model.supportsImageGen;
+        isImageIntent && configuredModelSupportsRequestedImageGeneration;
 
       if ((matchesVision && matchesImageGen) || allowImageIntentWithImageGenOnly) {
         const reason = allowImageIntentWithImageGenOnly && !matchesVision
@@ -355,7 +375,15 @@ export async function routeToModel(params: {
       !item.model.isSubModel &&
       item.model.isEnabled &&
       !isVideoProviderType(String(item.provider.type || ""), item.model) &&
-      (item.model.supportsImageGen ||
+      ((intent === "image_gen"
+        ? modelSupportsTextToImageDescriptor({
+            providerType: item.provider.type,
+            modelId: item.model.modelId,
+            capabilities: item.model.capabilities,
+            defaultParams: item.model.defaultParams,
+            supportsImageGen: item.model.supportsImageGen,
+          })
+        : item.model.supportsImageGen) ||
         hasVisionCapability(item.model, item.provider.type))
     );
   });

lib/canvas/generate-image.ts

@@ -16,6 +16,7 @@ import { createRunway } from "@/lib/ai/providers/runway";
 import { createKrea } from "@/lib/ai/providers/krea";
 import { createVolcanoVision } from "@/lib/ai/providers/volcano-vision";
 import { resolveModelRuntimeId } from "@/lib/ai/model-runtime";
+import { modelSupportsTextToImageDescriptor } from "@/lib/ai/model-classification";
 import { modelLacksReferenceImageInputSupport } from "@/lib/ai/reference-image-capability";
 import { filterReachableRemoteUrls } from "@/lib/net/url-access";
 import {
@@ -1828,6 +1829,7 @@ function findCanvasFailoverModel(params: {
   failedModelRefs: string[];
   preferLowCostSameNameProviders?: boolean;
   requiresImageInput?: boolean;
+  requiresTextToImage?: boolean;
   scaleAdjustModelFamily?: string | null;
 }) {
   const failedRefs = new Set(
@@ -1872,6 +1874,18 @@ function findCanvasFailoverModel(params: {
     ) {
       return false;
     }
+    if (
+      params.requiresTextToImage &&
+      !modelSupportsTextToImageDescriptor({
+        providerType: item.provider.type,
+        modelId: item.model.modelId,
+        capabilities: item.model.capabilities,
+        defaultParams: item.model.defaultParams,
+        supportsImageGen: item.model.supportsImageGen,
+      })
+    ) {
+      return false;
+    }
     if (!SUPPORTED_CANVAS_PROVIDER_TYPES.has(item.provider.type)) return false;
     return getCanvasModelRefKeys(item).every((key) => !failedRefs.has(key));
   };
@@ -2018,6 +2032,7 @@ export async function generateCanvasImage(
     )
   ).slice(0, 12);
   const requiresImageInput = normalizedReferenceImages.length > 0;
+  const requiresTextToImage = normalizedReferenceImages.length === 0;
 
   let provider = selectedModel.provider;
 
@@ -2057,6 +2072,48 @@ export async function generateCanvasImage(
     provider = selectedModel.provider;
   }
 
+  if (
+    requiresTextToImage &&
+    !modelSupportsTextToImageDescriptor({
+      providerType: selectedModel.provider.type,
+      modelId: selectedModel.model.modelId,
+      capabilities: selectedModel.model.capabilities,
+      defaultParams: selectedModel.model.defaultParams,
+      supportsImageGen: selectedModel.model.supportsImageGen,
+    })
+  ) {
+    if (request.allowFailover === false) {
+      throw new Error("当前图片模型不支持纯文生图，请选择支持文生图的生图模型。");
+    }
+
+    const textToImageFallback = findCanvasFailoverModel({
+      selectedModel,
+      imageModels,
+      failedModelRefs: getCanvasModelRefKeys(selectedModel),
+      preferLowCostSameNameProviders: Boolean(
+        debugMeta?.scaleAdjustPreferSameNameProviderFailover
+      ),
+      requiresTextToImage,
+      scaleAdjustModelFamily,
+    });
+
+    if (!textToImageFallback) {
+      throw new Error("当前任务需要纯文生图，但暂无可用的文生图模型。");
+    }
+
+    console.warn("[Canvas Generate Image] Fallback to text-to-image capable model:", {
+      debugLabel,
+      requestedModel: selectedModel.model.modelId || selectedModel.model.id,
+      fallbackModel: textToImageFallback.model.modelId || textToImageFallback.model.id,
+      requestedProviderType: provider.type,
+      fallbackProviderType: textToImageFallback.provider.type,
+      reason: "selected-model-lacks-text-to-image-support",
+    });
+
+    selectedModel = textToImageFallback;
+    provider = selectedModel.provider;
+  }
+
   if (!SUPPORTED_CANVAS_PROVIDER_TYPES.has(provider.type)) {
     if (request.allowFailover === false) {
       throw new Error(
@@ -2075,6 +2132,18 @@ export async function generateCanvasImage(
       ) {
         continue;
       }
+      if (
+        requiresTextToImage &&
+        !modelSupportsTextToImageDescriptor({
+          providerType: candidate.provider.type,
+          modelId: candidate.model.modelId,
+          capabilities: candidate.model.capabilities,
+          defaultParams: candidate.model.defaultParams,
+          supportsImageGen: candidate.model.supportsImageGen,
+        })
+      ) {
+        continue;
+      }
       fallbackModel = candidate;
       fallbackProvider = candidateProvider;
       break;
@@ -2642,6 +2711,7 @@ export async function generateCanvasImage(
           debugMeta?.scaleAdjustPreferSameNameProviderFailover
         ),
         requiresImageInput: normalizedReferenceImages.length > 0,
+        requiresTextToImage: normalizedReferenceImages.length === 0,
         scaleAdjustModelFamily,
       })
     : null;

lib/external-image-api/service.ts

@@ -2,7 +2,10 @@ import { createHash, timingSafeEqual } from "node:crypto";
 import { sql } from "drizzle-orm";
 import sharp from "sharp";
 import { getLanguageModelById } from "@/lib/ai/dynamic-provider";
-import { isImageGenerationModelDescriptor } from "@/lib/ai/model-classification";
+import {
+  isImageGenerationModelDescriptor,
+  modelSupportsTextToImageDescriptor,
+} from "@/lib/ai/model-classification";
 import { db } from "@/lib/db";
 import { getEnabledAiModelsWithProvider, getSystemSetting } from "@/lib/db/queries";
 import { user } from "@/lib/db/schema";
@@ -322,7 +325,13 @@ function supportsImageEditForReferenceCount(
   referenceImageCount: number,
 ) {
   if (referenceImageCount <= 0) {
-    return true;
+    return modelSupportsTextToImageDescriptor({
+      providerType: entry.provider.type,
+      modelId: entry.model.modelId,
+      capabilities: entry.model.capabilities,
+      defaultParams: entry.model.defaultParams,
+      supportsImageGen: entry.model.supportsImageGen,
+    });
   }
 
   const capabilities = toRecord(entry.model.capabilities);