diff --git a/apps/desktop/src/main/cindy-brain/__tests__/cindySlot.test.ts b/apps/desktop/src/main/cindy-brain/__tests__/cindySlot.test.ts
index 99857220b9..1be490d421 100644
--- a/apps/desktop/src/main/cindy-brain/__tests__/cindySlot.test.ts
+++ b/apps/desktop/src/main/cindy-brain/__tests__/cindySlot.test.ts
@@ -748,6 +748,25 @@ describe('目录空清单 = 能力暂不可用', () => {
 });
 
 describe('意识专属后端覆盖(解析表第②层)', () => {
+  it('新版 Provider-aware 覆盖保留精确来源到最终派发', async () => {
+    const { slot, generateImage } = makeSlot({
+      getMediaOverride: vi.fn(() => ({
+        modelId: 'gpt-image-2',
+        providerId: 'openai',
+        label: 'GPT Image 2',
+      })),
+    });
+
+    const result = await slot.handleModelRequest('art', REQ);
+
+    expect(result).toMatchObject({ ok: true, model: 'gpt-image-2', modelLabel: 'GPT Image 2' });
+    expect(generateImage).toHaveBeenCalledWith({
+      prompt: '一只猫',
+      model: 'gpt-image-2',
+      providerId: 'openai',
+    });
+  });
+
   it('覆盖压过档位;调用显式点名仍压过覆盖;下架型号的覆盖静默落回', async () => {
     const pinned = makeSlot({
       getOverride: vi.fn(() => 'gemini-3-pro-image') as unknown as CindySlotDeps['getOverride'],
diff --git a/apps/desktop/src/main/cindy-brain/cindySlot.ts b/apps/desktop/src/main/cindy-brain/cindySlot.ts
index 3e0b3977fb..9cc57ba773 100644
--- a/apps/desktop/src/main/cindy-brain/cindySlot.ts
+++ b/apps/desktop/src/main/cindy-brain/cindySlot.ts
@@ -136,6 +136,12 @@ export interface CindyImageCapabilities {
   maxEditImages?: number;
 }
 
+export interface CindyMediaOverrideSelection {
+  modelId: string;
+  providerId: string;
+  label?: string;
+}
+
 export interface CindySlotDeps {
   getGhost(id: string): InstalledGhost | null;
   /** 当前账号作用域；跨 await 的媒体任务必须捕获并持续复核。 */
@@ -147,6 +153,7 @@ export interface CindySlotDeps {
   generateImage(params: {
     prompt: string;
     model: string;
+    providerId?: string;
     aspectRatio?: GhostImageAspectRatio;
   }): Promise<{ buffer: Uint8Array; mimeType: string }>;
   /** 主机统一图片通道·改图;源图以磁盘路径喂给网关(意识摸不到路径)。
@@ -154,6 +161,7 @@ export interface CindySlotDeps {
   editImage(params: {
     prompt: string;
     model: string;
+    providerId?: string;
     imagePaths: string[];
     aspectRatio?: GhostImageAspectRatio;
   }): Promise<{ buffer: Uint8Array; mimeType: string }>;
@@ -161,7 +169,7 @@ export interface CindySlotDeps {
    * 该图像型号的 provider 实际能力。缺席/查无 = 只执行通用 1–4 图粗筛；
    * provider 上限更低时，slot 在读源图与出网前给出型号级明确拒绝。
    */
-  imageCapabilities?(model: string): CindyImageCapabilities | null;
+  imageCapabilities?(model: string, providerId?: string): CindyImageCapabilities | null;
   /**
    * 主机统一视频通道·文生视频(art 视频 provider 层复用,submit→
    * 轮询→下载一条龙在注入实现里完成);返回视频字节与 mime,外加实际
@@ -169,7 +177,7 @@ export interface CindySlotDeps {
    * 分钟级才 resolve,在途名额在整个等待期占用。
    */
   generateVideo(
-    params: { prompt: string; model: string } & CindyVideoParams,
+    params: { prompt: string; model: string; providerId?: string } & CindyVideoParams,
   ): Promise<{ buffer: Uint8Array; mimeType: string; videoParams?: GhostVideoResultParams }>;
   /**
    * 主机统一视频通道·参考图生视频(源图以磁盘路径注入)。`refMode` 决定这
@@ -181,6 +189,7 @@ export interface CindySlotDeps {
     params: {
       prompt: string;
       model: string;
+      providerId?: string;
       imagePaths: string[];
       refMode: GhostVideoRefMode;
     } & CindyVideoParams,
@@ -190,7 +199,7 @@ export interface CindySlotDeps {
    * 该型号 → null)。可选依赖:不注入 = 跳过按型号校验,只做协议层粗筛
    * (值仍会被 provider 层自己的校验拦下,只是话术不如这里友好)。
    */
-  videoCapabilities?(model: string): CindyVideoCapabilities | null;
+  videoCapabilities?(model: string, providerId?: string): CindyVideoCapabilities | null;
   /**
    * 指纹 → 磁盘路径,且仅当该媒体在此意识名下(出生或画廊,查账本);
    * 不属于它 / 查无此账 / 文件缺失一律 null(不区分,不给探测空间)。
@@ -204,6 +213,14 @@ export interface CindySlotDeps {
    * 白名单校验(型号可能已随主机演进下架)。
    */
   getOverride(ghostId: string, capability: string): string | null;
+  /**
+   * Provider-aware 媒体覆盖。新版 Host 偏好按 providerId + modelId 保存；存在时
+   * 必须贯穿能力校验与最终派发，不能降回裸 modelId 后走 first-wins。
+   */
+  getMediaOverride?(
+    ghostId: string,
+    capability: string,
+  ): CindyMediaOverrideSelection | null;
   /**
    * 当前图像能力配置——真身是 providers.json 运行时目录(与会话模型列表
    * 同一获取来源),每单现读跟随热更。models = 白名单与显示名;defaults =
@@ -830,6 +847,8 @@ export class GhostCindySlot {
     const defaults = cfg.defaults;
     const whitelist = new Set(cfg.models.map((m) => m.id));
     let model = defaults.standard;
+    let providerId: string | undefined;
+    let providerModelLabel: string | undefined;
     if (p.tier !== undefined) {
       if (typeof p.tier !== 'string' || !(GHOST_MODEL_TIERS as readonly string[]).includes(p.tier)) {
         return { ok: false, message: `未知档位(可用:${GHOST_MODEL_TIERS.join(' / ')})` };
@@ -837,13 +856,20 @@ export class GhostCindySlot {
       model = defaults[p.tier as GhostModelTier];
     }
     const capability = `${info.category}.${info.action}`;
-    const override = this.deps.getOverride(ghostId, capability);
-    if (override !== null) {
-      if (whitelist.has(override)) {
-        model = override;
-      } else {
-        // 钉的型号已随白名单演进下架:落回上面的档位/默认,不让老配置卡死能力。
-        this.deps.log?.warn('ghost cindy override no longer whitelisted, ignored', { ghostId, override });
+    const mediaOverride = this.deps.getMediaOverride?.(ghostId, capability) ?? null;
+    if (mediaOverride) {
+      model = mediaOverride.modelId;
+      providerId = mediaOverride.providerId;
+      providerModelLabel = mediaOverride.label;
+    } else {
+      const override = this.deps.getOverride(ghostId, capability);
+      if (override !== null) {
+        if (whitelist.has(override)) {
+          model = override;
+        } else {
+          // 钉的型号已随白名单演进下架:落回上面的档位/默认,不让老配置卡死能力。
+          this.deps.log?.warn('ghost cindy override no longer whitelisted, ignored', { ghostId, override });
+        }
       }
     }
     if (p.model !== undefined) {
@@ -855,6 +881,10 @@ export class GhostCindySlot {
           message: `不支持的模型(不在主机白名单内)。当前可用:${cfg.models.length > 0 ? cfg.models.map((m) => m.id).join(' / ') : '(暂无可用型号)'}`,
         };
       }
+      if (p.model !== model) {
+        providerId = undefined;
+        providerModelLabel = undefined;
+      }
       model = p.model;
     }
 
@@ -864,7 +894,7 @@ export class GhostCindySlot {
     // 明拒并列出该型号的可用值,不做最近似降级——静默改成别的档位会让意识
     // 以为自己的参数生效了。
     if (info.category === 'video' && presentVideoKeys.length > 0) {
-      const caps = this.deps.videoCapabilities?.(model) ?? null;
+      const caps = this.deps.videoCapabilities?.(model, providerId) ?? null;
       if (caps) {
         const unsupported = describeUnsupportedVideoParams(videoParams, caps);
         if (unsupported) {
@@ -897,7 +927,7 @@ export class GhostCindySlot {
     }
 
     if (kind === 'edit_image') {
-      const perModelMax = this.deps.imageCapabilities?.(model)?.maxEditImages;
+      const perModelMax = this.deps.imageCapabilities?.(model, providerId)?.maxEditImages;
       if (perModelMax !== undefined && hashes.length > perModelMax) {
         const label = cfg.models.find((m) => m.id === model)?.label ?? model;
         return {
@@ -912,7 +942,7 @@ export class GhostCindySlot {
     // 张数上限都不一样。不支持即明拒,不降级成另一种用法——降级会出一条
     // 用户没要的片子,还照样计费。
     if (kind === 'edit_video') {
-      const caps = this.deps.videoCapabilities?.(model) ?? null;
+      const caps = this.deps.videoCapabilities?.(model, providerId) ?? null;
       const perModelMax = caps?.maxImagesByRefMode?.[refMode];
       if (caps?.maxImagesByRefMode !== undefined) {
         const label = cfg.models.find((m) => m.id === model)?.label ?? model;
@@ -961,6 +991,7 @@ export class GhostCindySlot {
       this.deps.log?.info(`ghost cindy-request ${kind} start`, {
         ghostId,
         model,
+        ...(providerId ? { providerId } : {}),
         callId,
         ...(p.mode === 'submit' ? { mode: 'submit' } : {}),
       });
@@ -997,6 +1028,7 @@ export class GhostCindySlot {
           generated = await this.deps.editImage({
             prompt,
             model,
+            ...(providerId ? { providerId } : {}),
             imagePaths,
             ...(aspectRatio !== undefined ? { aspectRatio } : {}),
           });
@@ -1004,12 +1036,25 @@ export class GhostCindySlot {
           generated = await this.deps.generateImage({
             prompt,
             model,
+            ...(providerId ? { providerId } : {}),
             ...(aspectRatio !== undefined ? { aspectRatio } : {}),
           });
         } else if (kind === 'edit_video') {
-          generated = await this.deps.editVideo({ prompt, model, imagePaths, refMode, ...videoParams });
+          generated = await this.deps.editVideo({
+            prompt,
+            model,
+            ...(providerId ? { providerId } : {}),
+            imagePaths,
+            refMode,
+            ...videoParams,
+          });
         } else {
-          generated = await this.deps.generateVideo({ prompt, model, ...videoParams });
+          generated = await this.deps.generateVideo({
+            prompt,
+            model,
+            ...(providerId ? { providerId } : {}),
+            ...videoParams,
+          });
         }
         assertOwnerScopeCurrent();
 
@@ -1027,13 +1072,15 @@ export class GhostCindySlot {
         this.deps.log?.info(`ghost cindy-request ${kind} done`, {
           ghostId,
           model,
+          ...(providerId ? { providerId } : {}),
           callId,
           hash: saved.hash,
           bytes: generated.buffer.byteLength,
         });
         // 实际选型随结果回传(主机权威信息):意识交卷 note、会话里的 AI 与
         // 用户由此看得见"这单是谁画的"。
-        const modelLabel = cfg.models.find((m) => m.id === model)?.label ?? model;
+        const modelLabel =
+          providerModelLabel ?? cfg.models.find((m) => m.id === model)?.label ?? model;
         // 图片代办附带像素宽高(字节头解析,best-effort):意识供聊天卡片时
         // 据此精确声明卡高,首帧零跳动;解析不出就缺省,意识回退估计值。
         const dims =
diff --git a/apps/desktop/src/main/cindy-brain/codexImageClient.ts b/apps/desktop/src/main/cindy-brain/codexImageClient.ts
index 527ccc13da..5cf4883e41 100644
--- a/apps/desktop/src/main/cindy-brain/codexImageClient.ts
+++ b/apps/desktop/src/main/cindy-brain/codexImageClient.ts
@@ -7,9 +7,11 @@
  * SSE stream because image-generation events may be newer than SDK typings.
  */
 
+import { randomUUID } from 'node:crypto';
 import fs from 'node:fs/promises';
 
 import type { ImageChannel, ImageChannelResult } from './imageChannelRegistry.js';
+import { mediaRequestParamsForLog, mediaRequestUrlForLog } from '../cindy-media/mediaRequestLog.js';
 import { sniffMediaMime } from '../cindy-media/sniffMediaMime.js';
 import { createLogger } from '../logger.js';
 
@@ -143,6 +145,7 @@ export function createCodexImageChannel(opts: CreateCodexImageChannelOptions): I
     prompt: string;
     imagePaths?: string[];
     aspectRatio?: '1:1' | '3:2' | '2:3';
+    signal?: AbortSignal;
   }): Promise<ImageChannelResult> {
     if (params.model !== `openai/${IMAGE_MODEL}`) {
       throw new Error(`Codex 图像通道不支持模型:${params.model}`);
@@ -159,36 +162,66 @@ export function createCodexImageChannel(opts: CreateCodexImageChannelOptions): I
       { type: 'input_text', text: params.prompt },
       ...images,
     ];
-    const response = await doFetch(CODEX_RESPONSES_URL, {
+    const body = {
+      model: HOST_MODEL,
+      store: false,
+      stream: true,
+      instructions: 'Use the image_generation tool to fulfill this image request.',
+      input: [{ type: 'message', role: 'user', content }],
+      tools: [
+        {
+          type: 'image_generation',
+          model: IMAGE_MODEL,
+          ...(params.aspectRatio ? { size: SIZE_BY_ASPECT[params.aspectRatio] } : {}),
+          quality: 'medium',
+          output_format: 'png',
+          background: 'opaque',
+          partial_images: 1,
+        },
+      ],
+    };
+    const requestId = randomUUID();
+    const startedAt = Date.now();
+    const requestLog = {
+      requestId,
+      providerId: 'openai',
+      modelId: params.model,
       method: 'POST',
-      headers: {
-        Authorization: `Bearer ${auth.accessToken}`,
-        'Content-Type': 'application/json',
-        Accept: 'text/event-stream',
-        'OpenAI-Beta': 'responses=experimental',
-        originator: 'codex_cli_rs',
-        'User-Agent': USER_AGENT,
-        ...(auth.accountId ? { 'ChatGPT-Account-Id': auth.accountId } : {}),
-      },
-      body: JSON.stringify({
-        model: HOST_MODEL,
-        store: false,
-        stream: true,
-        instructions: 'Use the image_generation tool to fulfill this image request.',
-        input: [{ type: 'message', role: 'user', content }],
-        tools: [
-          {
-            type: 'image_generation',
-            model: IMAGE_MODEL,
-            ...(params.aspectRatio ? { size: SIZE_BY_ASPECT[params.aspectRatio] } : {}),
-            quality: 'medium',
-            output_format: 'png',
-            background: 'opaque',
-            partial_images: 1,
-          },
-        ],
-      }),
+      url: mediaRequestUrlForLog(CODEX_RESPONSES_URL),
+    };
+    log.info('media request dispatch', {
+      ...requestLog,
+      params: mediaRequestParamsForLog(body),
     });
+    let response: Response;
+    try {
+      response = await doFetch(CODEX_RESPONSES_URL, {
+        method: 'POST',
+        headers: {
+          Authorization: `Bearer ${auth.accessToken}`,
+          'Content-Type': 'application/json',
+          Accept: 'text/event-stream',
+          'OpenAI-Beta': 'responses=experimental',
+          originator: 'codex_cli_rs',
+          'User-Agent': USER_AGENT,
+          ...(auth.accountId ? { 'ChatGPT-Account-Id': auth.accountId } : {}),
+        },
+        body: JSON.stringify(body),
+        signal: params.signal,
+      });
+      log.info('media request response', {
+        ...requestLog,
+        status: response.status,
+        durationMs: Date.now() - startedAt,
+      });
+    } catch (error) {
+      log.warn('media request failed', {
+        ...requestLog,
+        durationMs: Date.now() - startedAt,
+        error: mediaRequestParamsForLog(error instanceof Error ? error.message : String(error)),
+      });
+      throw error;
+    }
     if (!response.ok) await httpError(response, auth.accessToken, opts.onAuthFailure);
     const b64 = await collectImageB64(response);
     if (!b64) throw new Error('Codex 返回中没有图片,请重试或改用 OpenAI Platform API key');
@@ -197,8 +230,9 @@ export function createCodexImageChannel(opts: CreateCodexImageChannelOptions): I
 
   return {
     ready: opts.hasOAuthLogin,
-    generateImage: ({ model, prompt, aspectRatio }) => generate({ model, prompt, aspectRatio }),
-    editImage: ({ model, prompt, imagePaths, aspectRatio }) =>
-      generate({ model, prompt, imagePaths, aspectRatio }),
+    generateImage: ({ model, prompt, aspectRatio, signal }) =>
+      generate({ model, prompt, aspectRatio, signal }),
+    editImage: ({ model, prompt, imagePaths, aspectRatio, signal }) =>
+      generate({ model, prompt, imagePaths, aspectRatio, signal }),
   };
 }
diff --git a/apps/desktop/src/main/cindy-brain/forge.ts b/apps/desktop/src/main/cindy-brain/forge.ts
index caf1549d43..ba5f20c972 100644
--- a/apps/desktop/src/main/cindy-brain/forge.ts
+++ b/apps/desktop/src/main/cindy-brain/forge.ts
@@ -2242,20 +2242,23 @@ const result = await (await fetch('/media-models?type=image')).json();
 //   models:[{
 //     id,
 //     name,
+//     providerId,
 //     modalities:{ input:['text','image'], output:['image'] }
 //   }],
-//   defaultModelId:string|null
+//   defaultModelId:string|null,
+//   defaultProviderId:string|null
 // }
 \`\`\`
 
-\`type\` 只接受 \`image\` / \`video\`。Host 按 Gateway \`mode\` 切大类，并结合插件在
-\`cindy.image/video\` 声明的动作、Gateway \`modalities\`、Guide operation 与当前客户端
+\`type\` 只接受 \`image\` / \`video\`。Host 按模型 \`mode\` 切大类，并结合插件在
+\`cindy.image/video\` 声明的动作、模型 \`modalities\`、Guide operation 与当前客户端
 协议支持度，只返回当前真正可执行的模型。单个模型的 Guide 缺失、损坏或版本过新只隔离
 该模型，不拖垮整个目录。
 
-响应仍只把 Gateway \`architecture\` 已归一化后的 \`modalities.input/output\` 原样交给插件，
+响应只把已归一化的 \`modalities.input/output\` 与来源 \`providerId\` 交给插件，
 不下发 Guide、endpoint、凭证或 Host 内部兼容判定。插件可把用户选择的模型 id 存进自己的
-\`/kv\`，再通过工具结果或插件说明交给当前 Agent；付费请求前 Core 会再次校验。
+\`/kv\`，但必须同时保存 \`providerId\`，并把这对精确选择交给当前 Agent；同一个模型 id
+可由多个 Provider 提供，不能按 id 去重或自行改换来源。付费请求前 Core 会再次校验。
 
 插件与 Agent 不需要新的媒体协议，继续使用现有工具调用链：
 
diff --git a/apps/desktop/src/main/cindy-brain/imageChannelRegistry.ts b/apps/desktop/src/main/cindy-brain/imageChannelRegistry.ts
index 4f8ad470da..99e566a6fe 100644
--- a/apps/desktop/src/main/cindy-brain/imageChannelRegistry.ts
+++ b/apps/desktop/src/main/cindy-brain/imageChannelRegistry.ts
@@ -68,12 +68,14 @@ export interface ImageChannel {
     model: string;
     prompt: string;
     aspectRatio?: GhostImageAspectRatio;
+    signal?: AbortSignal;
   }): Promise<ImageChannelResult>;
   editImage(params: {
     model: string;
     prompt: string;
     imagePaths: string[];
     aspectRatio?: GhostImageAspectRatio;
+    signal?: AbortSignal;
   }): Promise<ImageChannelResult>;
 }
 
diff --git a/apps/desktop/src/main/cindy-brain/index.ts b/apps/desktop/src/main/cindy-brain/index.ts
index 067056251f..57011956aa 100644
--- a/apps/desktop/src/main/cindy-brain/index.ts
+++ b/apps/desktop/src/main/cindy-brain/index.ts
@@ -361,6 +361,10 @@ import { getCodexImageAuthBinding } from './codexImageAuthBinding.js';
 import { createGatewayImageClient } from '../cindy-proxy-media/api/gatewayImageClient.js';
 import { createXaiVideoProvider } from '../cindy-proxy-media/video/providers/xai.js';
 import * as blobStore from '../cindy-media/blobStore.js';
+import {
+  configureProviderMediaRuntime,
+  listProviderMediaModels,
+} from '../cindy-media/providerMediaRuntime.js';
 import * as ledger from '../cindy-media/ledger.js';
 import { ingestMedia, supportedMime } from '../cindy-media/ingest.js';
 import { captureMediaRefCompensationScope } from '../cindy-media/refCompensationJournal.js';
@@ -374,6 +378,7 @@ import {
   filterEnabledGatewayMediaModels,
   isMediaModelExecutable,
   listExecutableMediaModels,
+  supportsMediaCapability,
 } from '../model-access/mediaModels.js';
 // ⚠️ 下面三个依赖必须保持模块顶层静态 import,禁止改回函数内 await import():
 // 运行时 import() 会被 Rollup 编译成跨 chunk 的 require(尤其 drizzle-orm 会拆独立
@@ -3148,38 +3153,118 @@ const getCatalogVideoConfig = (): ReturnType<typeof getCatalogMediaConfig> =>
 const getCatalogEmbedConfig = (): ReturnType<typeof getCatalogMediaConfig> =>
   getCatalogMediaConfig('embed');
 
-/**
- * Art 等插件的媒体偏好只认 Gateway `/models` 快照，并叠加客户端现有停用准入。
- * 不合并 providers.json 的 OpenAI/Gemini/自定义来源；第三方媒体模型后续单独接入。
- */
-function getGatewayMediaPreferenceConfig(
+const MEDIA_PREFERENCE_PREFIX = 'media:';
+
+function encodeMediaPreference(providerId: string, modelId: string): string {
+  return `${MEDIA_PREFERENCE_PREFIX}${encodeURIComponent(providerId)}:${encodeURIComponent(modelId)}`;
+}
+
+function decodeMediaPreference(value: string): { providerId: string; modelId: string } | null {
+  if (!value.startsWith(MEDIA_PREFERENCE_PREFIX)) return null;
+  const encoded = value.slice(MEDIA_PREFERENCE_PREFIX.length);
+  const separator = encoded.indexOf(':');
+  if (separator <= 0 || separator === encoded.length - 1) return null;
+  try {
+    const providerId = decodeURIComponent(encoded.slice(0, separator));
+    const modelId = decodeURIComponent(encoded.slice(separator + 1));
+    return providerId && modelId ? { providerId, modelId } : null;
+  } catch {
+    return null;
+  }
+}
+
+type CindyMediaPreferenceModel = CindyMediaCatalogConfig['models'][number] & {
+  modelId: string;
+  modelName: string;
+  providerName: string;
+  group: string;
+  routing?: import('@cindy/model-providers').Provider['routing'];
+};
+
+interface CindyMediaPreferenceConfig {
+  models: CindyMediaPreferenceModel[];
+  defaults: { standard: string; draft: string; best: string } | null;
+}
+
+/** Art 等插件的媒体偏好统一合并已就绪 Provider 与 Gateway 可执行模型。 */
+function getMediaPreferenceConfig(
   capability: GhostMediaCapability,
-): CindyMediaCatalogConfig {
+): CindyMediaPreferenceConfig {
   const kind = capability.startsWith('image.') ? 'image' : 'video';
   const coreCapability: MediaCapability =
     capability === 'video.edit' ? 'video.image_to_video' : capability;
-  const models = filterEnabledGatewayMediaModels(
+  const providers = new Map(
+    getActiveCatalog().providers.map((provider) => [provider.id, provider] as const),
+  );
+  const providerModels: CindyMediaPreferenceModel[] = listProviderMediaModels()
+    .filter(
+      (model) =>
+        model.mode === (kind === 'image' ? 'image_generation' : 'video_generation') &&
+        supportsMediaCapability(model.modalities, coreCapability),
+    )
+    .map((model) => {
+      const provider = providers.get(model.providerId);
+      const providerName = provider?.name ?? model.providerId;
+      return {
+        id: encodeMediaPreference(model.providerId, model.id),
+        modelId: model.id,
+        label: model.name,
+        modelName: model.name,
+        providerId: model.providerId,
+        providerName,
+        group: providerName,
+        ...(provider?.routing ? { routing: provider.routing } : {}),
+        supportsEdit: supportsMediaCapability(model.modalities, 'image.edit'),
+      };
+    });
+  const gatewayModels: CindyMediaPreferenceModel[] = filterEnabledGatewayMediaModels(
     getXdGatewayModels(),
     coreCapability,
     readModelDisableOverrides(),
   )
     .filter((model) => isMediaModelExecutable(model.id, coreCapability))
-    .map((model) => ({
-      id: model.id,
-      label: model.name ?? model.id,
-      providerId: 'xd',
-      supportsEdit: isMediaModelExecutable(
-        model.id,
-        kind === 'image' ? 'image.edit' : 'video.image_to_video',
-      ),
-    }));
-  const standard = models[0]?.id;
+    .map((model) => {
+      const provider = providers.get('xd');
+      const providerName = provider?.name ?? 'Cindy AI';
+      const modelName = model.name ?? model.id;
+      return {
+        id: encodeMediaPreference('xd', model.id),
+        modelId: model.id,
+        label: modelName,
+        modelName,
+        providerId: 'xd',
+        providerName,
+        group: providerName,
+        ...(provider?.routing ? { routing: provider.routing } : {}),
+        supportsEdit: isMediaModelExecutable(
+          model.id,
+          kind === 'image' ? 'image.edit' : 'video.image_to_video',
+        ),
+      };
+    });
+  const models = [...gatewayModels, ...providerModels];
+  const standard = gatewayModels[0]?.id ?? providerModels[0]?.id;
   return {
     models,
     defaults: standard ? { standard, draft: standard, best: standard } : null,
   };
 }
 
+function resolveMediaPreferenceModel(
+  config: CindyMediaPreferenceConfig,
+  preference: string | undefined,
+): CindyMediaPreferenceModel | null {
+  if (!preference) return null;
+  const exact = config.models.find((model) => model.id === preference);
+  if (exact) return exact;
+  if (decodeMediaPreference(preference)) return null;
+  return (
+    config.models.find((model) => model.providerId === 'xd' && model.modelId === preference) ??
+    config.models.find((model) => model.modelId === preference) ??
+    null
+  );
+}
+
 /**
  * 插件配置页的模型目录；这里只读目录，不提供任何生成入口。Host 只按 Gateway mode
  * 切图片/视频大类并透传 modalities，具体动作支持度由插件自行解释。
@@ -3229,6 +3314,7 @@ async function getGhostConfigurableMediaModels(
       models: models.map((model) => ({
         id: model.id,
         name: model.name ?? model.id,
+        providerId: model.providerId,
         ...(model.modalities
           ? {
               modalities: {
@@ -3238,7 +3324,11 @@ async function getGhostConfigurableMediaModels(
             }
           : {}),
       })),
-      defaultModelId: models[0]?.id ?? null,
+      defaultModelId: models.find((model) => model.providerId === 'xd')?.id ?? models[0]?.id ?? null,
+      defaultProviderId:
+        models.find((model) => model.providerId === 'xd')?.providerId ??
+        models[0]?.providerId ??
+        null,
     };
   } catch (error) {
     log.warn('read plugin media model catalog failed', {
@@ -3252,7 +3342,8 @@ async function getGhostConfigurableMediaModels(
 
 /**
  * 读取插件在 Host「Cindy 能力」中的现有媒体选型。
- * 这里只投影当前有效值，不复制或迁移配置；override 失效时与设置 UI 一样回落目录默认。
+ * 这里只投影当前有效值，不复制或迁移配置；用户已明确配置的精确来源失效时直接报错，
+ * 不能静默回落到另一个 Provider 或默认模型。
  */
 function getGhostConfiguredMediaModel(
   ghostId: string,
@@ -3279,14 +3370,24 @@ function getGhostConfiguredMediaModel(
     };
   }
 
-  const config = getGatewayMediaPreferenceConfig(mediaCapability);
-  const available = new Set(config.models.map((model) => model.id));
+  const config = getMediaPreferenceConfig(mediaCapability);
   const override = readGhostCindyOverrides(ghostId)[mediaCapability];
-  const modelId = override && available.has(override) ? override : config.defaults?.standard;
-  if (!modelId) {
-    return { ok: false, errorCode: 'NOT_AVAILABLE', message: '当前没有可用的媒体模型' };
+  const selected = override
+    ? resolveMediaPreferenceModel(config, override)
+    : resolveMediaPreferenceModel(config, config.defaults?.standard);
+  if (!selected) {
+    return {
+      ok: false,
+      errorCode: 'NOT_AVAILABLE',
+      message: override ? '已配置的媒体模型当前不可用' : '当前没有可用的媒体模型',
+    };
   }
-  return { ok: true, capability: mediaCapability, modelId };
+  return {
+    ok: true,
+    capability: mediaCapability,
+    modelId: selected.modelId,
+    providerId: selected.providerId,
+  };
 }
 
 /**
@@ -3295,8 +3396,17 @@ function getGhostConfiguredMediaModel(
  * generateImage / editImage / 视频提交边界按**当前** override 重算启用候选再验一次,
  * 不在册即拒,这次付费请求不发出(与 scheduler 派发前重裁决同语义)。
  */
-function assertMediaModelStillEnabled(kind: 'image' | 'video', model: string): void {
-  if (!getCatalogMediaConfig(kind).models.some((m) => m.id === model)) {
+function assertMediaModelStillEnabled(
+  kind: 'image' | 'video',
+  model: string,
+  providerId?: string,
+): void {
+  const available = providerId
+    ? getMediaPreferenceConfig(kind === 'image' ? 'image.generate' : 'video.generate').models.some(
+        (candidate) => candidate.providerId === providerId && candidate.modelId === model,
+      )
+    : getCatalogMediaConfig(kind).models.some((candidate) => candidate.id === model);
+  if (!available) {
     throw new Error(
       kind === 'image'
         ? '图像模型不可用(可能已停用或来源凭证未就绪),本次生成已取消'
@@ -3343,6 +3453,7 @@ async function readImageFileAsDataUri(absPath: string): Promise<string> {
 async function runGhostVideo(
   params: {
     alias: string;
+    providerId?: string;
     prompt: string;
     imageDataUris?: string[];
     /** 参考图用法(仅图生视频有);不传 = 执行器缺省的首尾帧。 */
@@ -3354,7 +3465,7 @@ async function runGhostVideo(
     throw new Error('视频能力不可用:主机未配置视频通道');
   }
   // 提交紧前重查(第二十一轮):参考图 data URI 准备是 await,窗口内被停用即拒。
-  assertMediaModelStillEnabled('video', params.alias);
+  assertMediaModelStillEnabled('video', params.alias, params.providerId);
   const r = await submitAndAwaitVideo(registry, params);
   return {
     buffer: r.buffer,
@@ -3376,8 +3487,17 @@ async function runGhostVideo(
  * 某视频型号的画面参数支持集(cindySlot 按型号二次校验用)。registry 缺席
  * 或 alias 查无 → null,cindySlot 据此跳过按型号校验(值仍会被执行器兜底拦下)。
  */
-function getGhostVideoCapabilities(model: string): CindyVideoCapabilities | null {
+function getGhostVideoCapabilities(
+  model: string,
+  providerId?: string,
+): CindyVideoCapabilities | null {
   try {
+    if (providerId) {
+      const available = getMediaPreferenceConfig('video.generate').models.some(
+        (candidate) => candidate.providerId === providerId && candidate.modelId === model,
+      );
+      if (!available) return null;
+    }
     const registry = getVideoProviderRegistry();
     if (!registry || !registry.hasAny()) return null;
     const caps = registry.resolveByAlias(model).provider.capabilities;
@@ -3395,9 +3515,14 @@ function getGhostVideoCapabilities(model: string): CindyVideoCapabilities | null
 }
 
 /** 图像 provider 的型号级编辑上限；slot 用它在文件 IO / 凭证读取前早拒。 */
-function getGhostImageCapabilities(model: string): CindyImageCapabilities | null {
+function getGhostImageCapabilities(
+  model: string,
+  providerId?: string,
+): CindyImageCapabilities | null {
   try {
-    return { maxEditImages: resolveImageChannelForModel(model, 'edit').maxEditImages };
+    return {
+      maxEditImages: resolveImageChannelForModel(model, 'edit', providerId).maxEditImages,
+    };
   } catch {
     return null;
   }
@@ -3474,6 +3599,7 @@ function getImageChannelRegistry(): ImageChannelRegistry {
     // 通道;目录 id 带 openai/ 前缀,public API 适配层剥前缀。
     const openaiImagesClient = createGatewayImageClient({
       getApiKey: () => getProviderSecretStore().get('openai-images'),
+      logger: log,
       // 境外端点吃系统代理(outboundFetch):main 的裸 fetch 不读系统代理设置,
       // 代理软件非 TUN 模式下会直连失败(2026-07 review;xd 网关通道不注 ——
       // 网关域名境内直连,与现状一致)。
@@ -3485,7 +3611,8 @@ function getImageChannelRegistry(): ImageChannelRegistry {
       },
       brandLabel: 'OpenAI',
       missingKeyMessage: 'OpenAI 图像 API key 未配置,请到「设置 → 模型供应商 → OpenAI」填入后重试',
-      beforeDispatch: (model) => assertMediaModelStillEnabled('image', `openai/${model}`),
+      beforeDispatch: (model) =>
+        assertMediaModelStillEnabled('image', `openai/${model}`, 'openai'),
     });
     const stripOpenaiPrefix = (id: string) =>
       id.startsWith('openai/') ? id.slice('openai/'.length) : id;
@@ -3498,7 +3625,7 @@ function getImageChannelRegistry(): ImageChannelRegistry {
         await getCodexImageAuthBinding().onAuthFailure(failure);
       },
       fetchImplementation: ((url, init) => outboundFetch(url as string, init)) as typeof fetch,
-      beforeDispatch: (model) => assertMediaModelStillEnabled('image', model),
+      beforeDispatch: (model) => assertMediaModelStillEnabled('image', model, 'openai'),
     });
     registry.register('openai', {
       // 用户明确配置 Platform key 时优先走确定性的 public Images API；否则复用
@@ -3506,24 +3633,30 @@ function getImageChannelRegistry(): ImageChannelRegistry {
       ready: () => hasOpenaiPlatformKey() || codexImagesClient.ready(),
       generateImage: (params) =>
         hasOpenaiPlatformKey()
-          ? openaiImagesClient.generateImage({
-              model: stripOpenaiPrefix(params.model),
-              prompt: params.prompt,
-              ...(params.aspectRatio
-                ? { size: GHOST_ASPECT_TO_GATEWAY_SIZE[params.aspectRatio] }
-                : {}),
-            })
+          ? openaiImagesClient.generateImage(
+              {
+                model: stripOpenaiPrefix(params.model),
+                prompt: params.prompt,
+                ...(params.aspectRatio
+                  ? { size: GHOST_ASPECT_TO_GATEWAY_SIZE[params.aspectRatio] }
+                  : {}),
+              },
+              params.signal,
+            )
           : codexImagesClient.generateImage(params),
       editImage: (params) =>
         hasOpenaiPlatformKey()
-          ? openaiImagesClient.editImage({
-              model: stripOpenaiPrefix(params.model),
-              prompt: params.prompt,
-              imagePaths: params.imagePaths,
-              ...(params.aspectRatio
-                ? { size: GHOST_ASPECT_TO_GATEWAY_SIZE[params.aspectRatio] }
-                : {}),
-            })
+          ? openaiImagesClient.editImage(
+              {
+                model: stripOpenaiPrefix(params.model),
+                prompt: params.prompt,
+                imagePaths: params.imagePaths,
+                ...(params.aspectRatio
+                  ? { size: GHOST_ASPECT_TO_GATEWAY_SIZE[params.aspectRatio] }
+                  : {}),
+              },
+              params.signal,
+            )
           : codexImagesClient.editImage(params),
     });
     imageChannelRegistrySingleton = registry;
@@ -3536,7 +3669,22 @@ function getImageChannelRegistry(): ImageChannelRegistry {
  * (cindyMediaCatalog first-wins 定格);白名单查无该模型时视同已停用
  * (assertMediaModelStillEnabled 同窗口语义)。
  */
-function resolveImageChannelForModel(model: string, operation: 'generate' | 'edit' = 'generate') {
+function resolveImageChannelForModel(
+  model: string,
+  operation: 'generate' | 'edit' = 'generate',
+  providerId?: string,
+) {
+  if (providerId) {
+    const capability: MediaCapability = operation === 'edit' ? 'image.edit' : 'image.generate';
+    const exact = getMediaPreferenceConfig(capability).models.find(
+      (candidate) => candidate.providerId === providerId && candidate.modelId === model,
+    );
+    if (!exact) throw new Error('图像模型或来源不可用,本次生成已取消');
+    if (operation === 'edit' && !exact.supportsEdit) {
+      throw new Error(`图像来源 ${providerId} 不支持图像编辑,请在设置中选择支持编辑的来源`);
+    }
+    return getImageChannelRegistry().resolve(providerId);
+  }
   const entry = getCatalogMediaConfig('image').models.find((m) => m.id === model);
   if (!entry) {
     const slash = model.indexOf('/');
@@ -3549,6 +3697,76 @@ function resolveImageChannelForModel(model: string, operation: 'generate' | 'edi
   return getImageChannelRegistry().resolve(entry.providerId);
 }
 
+function listLocalProviderMediaModels() {
+  const access = readModelDisableOverrides();
+  return getActiveCatalog().providers.flatMap((provider) => {
+    if (
+      provider.id === 'xd' ||
+      isProviderDisabled(access, provider.id) ||
+      !getImageChannelRegistry().isProviderReady(provider.id)
+    ) {
+      return [];
+    }
+    const supportsEdit = getImageChannelRegistry().isProviderEditReady(provider.id);
+    return (provider.imageModels ?? []).flatMap((model) => {
+      if (
+        !model.modalities ||
+        isModelDisabled(access, provider.id, model.id) ||
+        !model.modalities.output.includes('image')
+      ) {
+        return [];
+      }
+      const input = supportsEdit
+        ? [...model.modalities.input]
+        : model.modalities.input.filter((modality) => modality !== 'image');
+      return [
+        {
+          id: model.id,
+          name: model.name,
+          providerId: provider.id,
+          mode: 'image_generation' as const,
+          modalities: { input, output: [...model.modalities.output] },
+          ...(model.officialDocs ? { officialDocs: model.officialDocs } : {}),
+        },
+      ];
+    });
+  });
+}
+
+configureProviderMediaRuntime({
+  listModels: listLocalProviderMediaModels,
+  invoke: async (request) => {
+    if (request.capability !== 'image.generate' && request.capability !== 'image.edit') {
+      throw new Error('当前第三方 Provider 执行通道不支持该媒体能力');
+    }
+    const operation = request.capability === 'image.edit' ? 'edit' : 'generate';
+    const channel = resolveImageChannelForModel(request.modelId, operation, request.providerId);
+    if (
+      operation === 'edit' &&
+      channel.maxEditImages !== undefined &&
+      request.imagePaths.length > channel.maxEditImages
+    ) {
+      throw new Error(`当前图像来源最多支持 ${channel.maxEditImages} 张参考图`);
+    }
+    const response =
+      operation === 'edit'
+        ? await channel.editImage({
+            model: request.modelId,
+            prompt: request.prompt,
+            imagePaths: request.imagePaths,
+            ...(request.aspectRatio ? { aspectRatio: request.aspectRatio } : {}),
+            signal: request.signal,
+          })
+        : await channel.generateImage({
+            model: request.modelId,
+            prompt: request.prompt,
+            ...(request.aspectRatio ? { aspectRatio: request.aspectRatio } : {}),
+            signal: request.signal,
+          });
+    return decodeImageResponse(response);
+  },
+});
+
 /**
  * Plugin media and storage use the same process-local AppSession owner
  * boundary as the rest of the owner-scoped runtime.
@@ -3565,10 +3783,10 @@ export function getGhostCindySlot(): GhostCindySlot {
       isOwnerBoundaryPending: () => isGhostBoundaryPending(),
       // model 已在 modelSlot 按白名单校验;归属来源(providerId)按白名单条目
       // 定位,经 imageChannelRegistry 取对应执行通道(2026-07 图像多来源)。
-      generateImage: async ({ prompt, model, aspectRatio }) => {
+      generateImage: async ({ prompt, model, providerId, aspectRatio }) => {
         try {
-          assertMediaModelStillEnabled('image', model);
-          const channel = resolveImageChannelForModel(model);
+          assertMediaModelStillEnabled('image', model, providerId);
+          const channel = resolveImageChannelForModel(model, 'generate', providerId);
           return decodeImageResponse(
             await channel.generateImage({
               model,
@@ -3580,10 +3798,10 @@ export function getGhostCindySlot(): GhostCindySlot {
           humanizeImageChannelError(err);
         }
       },
-      editImage: async ({ prompt, model, imagePaths, aspectRatio }) => {
+      editImage: async ({ prompt, model, providerId, imagePaths, aspectRatio }) => {
         try {
-          assertMediaModelStillEnabled('image', model);
-          const channel = resolveImageChannelForModel(model, 'edit');
+          assertMediaModelStillEnabled('image', model, providerId);
+          const channel = resolveImageChannelForModel(model, 'edit', providerId);
           return decodeImageResponse(
             await channel.editImage({
               model,
@@ -3596,17 +3814,17 @@ export function getGhostCindySlot(): GhostCindySlot {
           humanizeImageChannelError(err);
         }
       },
-      generateVideo: async ({ prompt, model, ...videoParams }) => {
+      generateVideo: async ({ prompt, model, providerId, ...videoParams }) => {
         try {
-          assertMediaModelStillEnabled('video', model);
-          return await runGhostVideo({ alias: model, prompt, ...videoParams });
+          assertMediaModelStillEnabled('video', model, providerId);
+          return await runGhostVideo({ alias: model, providerId, prompt, ...videoParams });
         } catch (err) {
           humanizeImageChannelError(err);
         }
       },
-      editVideo: async ({ prompt, model, imagePaths, refMode, ...videoParams }) => {
+      editVideo: async ({ prompt, model, providerId, imagePaths, refMode, ...videoParams }) => {
         try {
-          assertMediaModelStillEnabled('video', model);
+          assertMediaModelStillEnabled('video', model, providerId);
           // 先算总量再读(闸按 refMode 分档:存量首尾帧不设闸,原样)。闸与
           // 读取绑在一个入口里,顺序是那边的结构保证、不是这里的约定;结果
           // 保序——顺序即语义:首/尾帧,或提示词里 [Image 1]… 的序号。
@@ -3617,6 +3835,7 @@ export function getGhostCindySlot(): GhostCindySlot {
           );
           return await runGhostVideo({
             alias: model,
+            providerId,
             prompt,
             imageDataUris,
             refMode,
@@ -3629,8 +3848,23 @@ export function getGhostCindySlot(): GhostCindySlot {
       // 画面参数按型号二次校验的数据源(registry capabilities)。
       imageCapabilities: getGhostImageCapabilities,
       videoCapabilities: getGhostVideoCapabilities,
+      getMediaOverride: (ghostId, capability) => {
+        const value = readGhostCindyOverrides(ghostId)[capability as CindyCapabilityKey] ?? null;
+        if (!value) return null;
+        const decoded = decodeMediaPreference(value);
+        if (!decoded) return null;
+        const selected = getMediaPreferenceConfig(capability as GhostMediaCapability).models.find(
+          (candidate) =>
+            candidate.providerId === decoded.providerId && candidate.modelId === decoded.modelId,
+        );
+        return {
+          ...decoded,
+          ...(selected ? { label: selected.label } : {}),
+        };
+      },
       getOverride: (ghostId, capability) => {
-        return readGhostCindyOverrides(ghostId)[capability as CindyCapabilityKey] ?? null;
+        const value = readGhostCindyOverrides(ghostId)[capability as CindyCapabilityKey] ?? null;
+        return value && !decodeMediaPreference(value) ? value : null;
       },
       getImageConfig: getCatalogImageConfig,
       getVideoConfig: getCatalogVideoConfig,
@@ -5823,7 +6057,7 @@ export function registerGhostIpc(): void {
     // 能力键的类目取对应清单)。defaultModel:目录默认选型的展示信息
     // ("默认(GPT Image 2)"),让用户看得见"跟随"当下跟的是谁;
     // null = 目录没有该类目的模型(能力暂不可用),渲染层据此显示灰字而非下拉。
-    const byKind = (cfg: CindyMediaCatalogConfig) => {
+    const byKind = (cfg: CindyMediaCatalogConfig | CindyMediaPreferenceConfig) => {
       const standard = cfg.defaults?.standard;
       return {
         options: cfg.models,
@@ -5863,10 +6097,10 @@ export function registerGhostIpc(): void {
       : null;
     event.returnValue = {
       overrides,
-      image: byKind(getGatewayMediaPreferenceConfig('image.generate')),
-      imageEdit: byKind(getGatewayMediaPreferenceConfig('image.edit')),
-      video: byKind(getGatewayMediaPreferenceConfig('video.generate')),
-      videoEdit: byKind(getGatewayMediaPreferenceConfig('video.edit')),
+      image: byKind(getMediaPreferenceConfig('image.generate')),
+      imageEdit: byKind(getMediaPreferenceConfig('image.edit')),
+      video: byKind(getMediaPreferenceConfig('video.generate')),
+      videoEdit: byKind(getMediaPreferenceConfig('video.edit')),
       text: {
         options: textOptions,
         defaultModel:
@@ -5939,10 +6173,10 @@ export function registerGhostIpc(): void {
       // fail-closed 兜底,不在这层把「暂时没配 key」当成非法值。
       if (
         !isCindyOverrideModelAllowed(capability as string, model, {
-          image: getGatewayMediaPreferenceConfig(
+          image: getMediaPreferenceConfig(
             capability === 'image.edit' ? 'image.edit' : 'image.generate',
           ).models,
-          video: getGatewayMediaPreferenceConfig(
+          video: getMediaPreferenceConfig(
             capability === 'video.edit' ? 'video.edit' : 'video.generate',
           ).models,
           embed: getCatalogEmbedConfig().models,
diff --git a/apps/desktop/src/main/cindy-media/__tests__/invocationService.test.ts b/apps/desktop/src/main/cindy-media/__tests__/invocationService.test.ts
index 281feb41e3..5013079979 100644
--- a/apps/desktop/src/main/cindy-media/__tests__/invocationService.test.ts
+++ b/apps/desktop/src/main/cindy-media/__tests__/invocationService.test.ts
@@ -20,6 +20,8 @@ const mocks = vi.hoisted(() => ({
   ownerGeneration: 1,
   models: vi.fn(),
   executableModels: vi.fn(),
+  providerModel: vi.fn(),
+  providerInvoke: vi.fn(),
   guide: vi.fn(),
   outboundFetch: vi.fn(),
   ingestMedia: vi.fn(),
@@ -83,6 +85,10 @@ vi.mock('../../model-access/mediaModels.js', () => ({
     }
   },
 }));
+vi.mock('../providerMediaRuntime.js', () => ({
+  resolveProviderMediaModel: mocks.providerModel,
+  invokeProviderMedia: mocks.providerInvoke,
+}));
 vi.mock('../ingest.js', () => ({ ingestMedia: mocks.ingestMedia }));
 vi.mock('../blobStore.js', () => ({
   readFile: mocks.readBlob,
@@ -216,13 +222,17 @@ describe('Cindy Core media invocation state and security boundary', () => {
     mocks.failTransitionTo = null;
     mocks.rows.clear();
     mocks.models.mockReset().mockResolvedValue([
-      { id: 'image-model', name: 'Image Model', mode: 'image_generation' },
+      { id: 'image-model', name: 'Image Model', providerId: 'xd', mode: 'image_generation' },
     ]);
     mocks.executableModels.mockReset().mockResolvedValue({
-      models: [{ id: 'image-model', name: 'Image Model', mode: 'image_generation' }],
+      models: [
+        { id: 'image-model', name: 'Image Model', providerId: 'xd', mode: 'image_generation' },
+      ],
       unavailable: [],
       candidateCount: 1,
     });
+    mocks.providerModel.mockReset();
+    mocks.providerInvoke.mockReset();
     mocks.guide.mockReset();
     mocks.outboundFetch.mockReset();
     mocks.ingestMedia.mockReset().mockResolvedValue({
@@ -274,6 +284,85 @@ describe('Cindy Core media invocation state and security boundary', () => {
     expect(mocks.ingestMedia).toHaveBeenCalledTimes(1);
   });
 
+  it('同名媒体模型按 providerId 精确准备并调用第三方来源', async () => {
+    const providerModel = {
+      id: 'openai/gpt-image-2',
+      name: 'GPT Image 2',
+      providerId: 'openai',
+      mode: 'image_generation',
+      modalities: { input: ['text', 'image'], output: ['image'] },
+      officialDocs: 'https://platform.openai.com/docs/guides/image-generation',
+    };
+    mocks.models.mockResolvedValue([
+      { ...providerModel, providerId: 'xd' },
+      providerModel,
+    ]);
+    mocks.providerModel.mockReturnValue(providerModel);
+    mocks.providerInvoke.mockResolvedValue({ buffer: PNG, mimeType: 'image/png' });
+
+    const prepared = await callCindyMedia({
+      action: 'prepare',
+      providerId: 'openai',
+      modelId: providerModel.id,
+      capability: 'image.generate',
+    });
+
+    expect(prepared).toMatchObject({
+      ok: true,
+      status: 'prepared',
+      provider_id: 'openai',
+      model_id: providerModel.id,
+    });
+    expect(mocks.guide).not.toHaveBeenCalled();
+
+    await expect(
+      callCindyMedia({
+        action: 'request',
+        invocationId: prepared.invocation_id as string,
+        body: { prompt: 'cat' },
+      }),
+    ).resolves.toMatchObject({
+      ok: true,
+      status: 'complete',
+      xdt_image_urls: [`cindy-media://blobs/${'a'.repeat(64)}.png`],
+    });
+    expect(mocks.providerInvoke).toHaveBeenCalledWith({
+      providerId: 'openai',
+      modelId: providerModel.id,
+      capability: 'image.generate',
+      prompt: 'cat',
+      imagePaths: [],
+      signal: expect.any(AbortSignal),
+    });
+  });
+
+  it('旧调用未传 providerId 时不在同名 Provider 中静默选择来源', async () => {
+    const model = {
+      id: 'openai/gpt-image-2',
+      name: 'GPT Image 2',
+      providerId: 'openai',
+      mode: 'image_generation',
+      modalities: { input: ['text', 'image'], output: ['image'] },
+      officialDocs: 'https://platform.openai.com/docs/guides/image-generation',
+    };
+    mocks.models.mockResolvedValue([{ ...model, providerId: 'xd' }, model]);
+
+    await expect(
+      callCindyMedia({
+        action: 'prepare',
+        modelId: model.id,
+        capability: 'image.generate',
+      }),
+    ).resolves.toMatchObject({
+      ok: false,
+      errorCode: 'MODEL_NOT_AVAILABLE',
+      retryable: false,
+      message: expect.stringContaining('provider_id'),
+    });
+    expect(mocks.guide).not.toHaveBeenCalled();
+    expect(mocks.providerModel).not.toHaveBeenCalled();
+  });
+
   it('终态历史不占用在途 invocation 上限', async () => {
     mocks.guide.mockResolvedValue(
       resolvedGuide(
@@ -626,7 +715,9 @@ describe('Cindy Core media invocation state and security boundary', () => {
         owner: mocks.currentUserId,
         drizzle: { owner: mocks.currentUserId },
       };
-      return [{ id: 'image-model', name: 'Image Model', mode: 'image_generation' }];
+      return [
+        { id: 'image-model', name: 'Image Model', providerId: 'xd', mode: 'image_generation' },
+      ];
     });
 
     await expect(
@@ -844,7 +935,7 @@ describe('Cindy Core media invocation state and security boundary', () => {
       capability: 'video.generate',
     };
     mocks.models.mockResolvedValue([
-      { id: 'image-model', name: 'Video Model', mode: 'video_generation' },
+      { id: 'image-model', name: 'Video Model', providerId: 'xd', mode: 'video_generation' },
     ]);
     mocks.guide.mockResolvedValue(resolvedGuide(asyncOperation));
     mocks.outboundFetch
@@ -905,7 +996,7 @@ describe('Cindy Core media invocation state and security boundary', () => {
       capability: 'video.generate',
     };
     mocks.models.mockResolvedValue([
-      { id: 'image-model', name: 'Video Model', mode: 'video_generation' },
+      { id: 'image-model', name: 'Video Model', providerId: 'xd', mode: 'video_generation' },
     ]);
     mocks.guide.mockResolvedValue(resolvedGuide(asyncOperation));
     const successPayload = {
@@ -978,7 +1069,7 @@ describe('Cindy Core media invocation state and security boundary', () => {
       capability: 'video.generate',
     };
     mocks.models.mockResolvedValue([
-      { id: 'image-model', name: 'Video Model', mode: 'video_generation' },
+      { id: 'image-model', name: 'Video Model', providerId: 'xd', mode: 'video_generation' },
     ]);
     mocks.guide.mockResolvedValue(resolvedGuide(asyncOperation));
     mocks.outboundFetch
@@ -1037,7 +1128,7 @@ describe('Cindy Core media invocation state and security boundary', () => {
       capability: 'video.generate',
     };
     mocks.models.mockResolvedValue([
-      { id: 'image-model', name: 'Video Model', mode: 'video_generation' },
+      { id: 'image-model', name: 'Video Model', providerId: 'xd', mode: 'video_generation' },
     ]);
     mocks.guide.mockResolvedValue(resolvedGuide(asyncOperation));
     mocks.outboundFetch
diff --git a/apps/desktop/src/main/cindy-media/__tests__/mediaRequestLog.test.ts b/apps/desktop/src/main/cindy-media/__tests__/mediaRequestLog.test.ts
new file mode 100644
index 0000000000..69d6e5a0da
--- /dev/null
+++ b/apps/desktop/src/main/cindy-media/__tests__/mediaRequestLog.test.ts
@@ -0,0 +1,33 @@
+import { describe, expect, it } from 'vitest';
+
+import { mediaRequestParamsForLog, mediaRequestUrlForLog } from '../mediaRequestLog.js';
+
+describe('media request log redaction', () => {
+  it('保留实际 URL 并脱敏 query 凭证', () => {
+    expect(
+      mediaRequestUrlForLog(
+        'https://user:pass@example.test/v1/images?model=gpt-image-2&api_key=secret#local',
+      ),
+    ).toBe(
+      'https://%5BREDACTED%5D:%5BREDACTED%5D@example.test/v1/images?model=gpt-image-2&api_key=%5BREDACTED%5D',
+    );
+  });
+
+  it('保留参数结构并收敛凭证和媒体正文', () => {
+    expect(
+      mediaRequestParamsForLog({
+        model: 'openai/gpt-image-2',
+        prompt: '生成一张图',
+        apiKey: 'secret',
+        image: 'data:image/png;base64,aGk=',
+        source: 'https://example.test/input.png?token=secret#frame',
+      }),
+    ).toEqual({
+      model: 'openai/gpt-image-2',
+      prompt: '生成一张图',
+      apiKey: '[REDACTED]',
+      image: '[data URL mime=image/png bytes=2]',
+      source: 'https://example.test/input.png?token=%5BREDACTED%5D#frame',
+    });
+  });
+});
diff --git a/apps/desktop/src/main/cindy-media/invocationService.ts b/apps/desktop/src/main/cindy-media/invocationService.ts
index 6fb4940bf6..cf15f1dc8a 100644
--- a/apps/desktop/src/main/cindy-media/invocationService.ts
+++ b/apps/desktop/src/main/cindy-media/invocationService.ts
@@ -10,6 +10,8 @@ import type {
   MediaResultKind,
   ResolvedMediaInvocationGuide,
 } from '../../shared/mediaInvocation.js';
+import { MODEL_ACCESS_INVOCATION_GUIDE_SCHEMA_VERSION } from '../../shared/mediaInvocation.js';
+import { GHOST_IMAGE_ASPECT_RATIOS, type GhostImageAspectRatio } from '../../shared/ghost.js';
 import { getAppCapabilities } from '../appCapabilities.js';
 import * as authManager from '../authManager.js';
 import * as imageCacheStore from '../imageCacheStore.js';
@@ -29,6 +31,12 @@ import {
 import { getProviderSecretStore } from '../secrets/providerSecretStore.js';
 import * as blobStore from './blobStore.js';
 import { ingestMedia } from './ingest.js';
+import { mediaRequestParamsForLog, mediaRequestUrlForLog } from './mediaRequestLog.js';
+import {
+  invokeProviderMedia,
+  resolveProviderMediaModel,
+  type ProviderMediaRuntimeModel,
+} from './providerMediaRuntime.js';
 import { sniffMediaMime } from './sniffMediaMime.js';
 import {
   countMediaInvocations,
@@ -59,6 +67,7 @@ const TERMINAL_MEDIA_RESULT_ERRORS = new Set([
   'MEDIA_RESULT_TOO_LARGE',
   'RESPONSE_TOO_LARGE',
 ]);
+const CLIENT_PROVIDER_IMAGE_GUIDE_ID = 'cindy-provider-image-v1';
 
 interface MediaConnection {
   baseUrl: string;
@@ -176,6 +185,65 @@ function submissionOutcomeUnknown(message: string): Record<string, unknown> {
   });
 }
 
+function providerImageGuide(
+  model: ProviderMediaRuntimeModel,
+  capability: 'image.generate' | 'image.edit',
+): PreparedMediaInvocationGuide {
+  const edit = capability === 'image.edit';
+  return {
+    schemaVersion: MODEL_ACCESS_INVOCATION_GUIDE_SCHEMA_VERSION,
+    guideId: CLIENT_PROVIDER_IMAGE_GUIDE_ID,
+    modelId: model.id,
+    revision: '1',
+    connection: { providerId: model.providerId },
+    capability,
+    request: {
+      method: 'POST',
+      path: '/client-provider-media',
+      bodyEncoding: 'json',
+      bodyModelPath: ['model'],
+      timeoutMs: 600_000,
+      maxRequestBytes: MAX_LOCAL_MEDIA_INPUT_TOTAL_BYTES + 1024 * 1024,
+      maxResponseBytes: MAX_IMAGE_RESULT_BYTES,
+    },
+    response: {
+      mode: 'sync',
+      media: [{ path: ['image'], encoding: 'base64', kind: 'image' }],
+    },
+    instructions: edit
+      ? '必填 prompt 和 image。image 可传一条 Cindy 本地媒体引用或引用数组；可选 aspect_ratio。'
+      : '必填 prompt；可选 aspect_ratio。model 与凭证由 Cindy 注入。',
+    exampleBody: {
+      prompt: edit ? '描述希望如何修改图片' : '描述希望生成的图片',
+      ...(edit ? { image: 'cindy-media://blobs/<hash>.png' } : {}),
+    },
+    inputSchema: {
+      type: 'object',
+      additionalProperties: false,
+      required: edit ? ['prompt', 'image'] : ['prompt'],
+      properties: {
+        prompt: { type: 'string', minLength: 1 },
+        ...(edit
+          ? {
+              image: {
+                oneOf: [
+                  { type: 'string', minLength: 1 },
+                  { type: 'array', minItems: 1, items: { type: 'string', minLength: 1 } },
+                ],
+              },
+            }
+          : {}),
+        aspect_ratio: { type: 'string', enum: [...GHOST_IMAGE_ASPECT_RATIOS] },
+      },
+    },
+    officialDocs: model.officialDocs ?? 'https://platform.openai.com/docs/guides/images',
+  };
+}
+
+function isClientProviderInvocation(invocation: StoredMediaInvocation): boolean {
+  return invocation.guide.guideId === CLIENT_PROVIDER_IMAGE_GUIDE_ID;
+}
+
 function resolveConnection(providerId: string): MediaConnection {
   if (providerId !== 'xd') {
     throw new MediaInvocationError(
@@ -342,6 +410,10 @@ function multipartRequestBody(
 }
 
 async function dispatchRequest(input: {
+  invocationId: string;
+  providerId: string;
+  modelId: string;
+  capability: MediaCapability;
   connection: MediaConnection;
   method: 'GET' | 'POST';
   path: string;
@@ -356,13 +428,29 @@ async function dispatchRequest(input: {
   const controller = new AbortController();
   const timeout = setTimeout(() => controller.abort(), input.timeoutMs);
   timeout.unref?.();
+  const url = requestUrl(input.connection.baseUrl, input.path);
+  const startedAt = Date.now();
+  const requestLog = {
+    invocationId: input.invocationId,
+    providerId: input.providerId,
+    modelId: input.modelId,
+    capability: input.capability,
+    operation: input.operation,
+    method: input.method,
+    url: mediaRequestUrlForLog(url),
+  };
+  let responseStatus: number | undefined;
   try {
     const requestBody = input.body
       ? input.bodyEncoding === 'multipart'
         ? multipartRequestBody(input.body, input.multipartFiles ?? [])
         : JSON.stringify(input.body)
       : undefined;
-    const response = await outboundFetch(requestUrl(input.connection.baseUrl, input.path), {
+    log.info('media request dispatch', {
+      ...requestLog,
+      params: mediaRequestParamsForLog(input.body ?? {}),
+    });
+    const response = await outboundFetch(url, {
       method: input.method,
       headers: {
         Accept: 'application/json',
@@ -376,6 +464,12 @@ async function dispatchRequest(input: {
       redirect: 'error',
       signal: controller.signal,
     });
+    responseStatus = response.status;
+    log.info('media request response', {
+      ...requestLog,
+      status: response.status,
+      durationMs: Date.now() - startedAt,
+    });
     const buffer = await readBoundedResponse(response, input.maxResponseBytes);
     if (!response.ok) {
       const message = providerErrorMessage(buffer);
@@ -410,6 +504,12 @@ async function dispatchRequest(input: {
       throw new MediaInvocationError('UPSTREAM_RESPONSE_INVALID', '上游成功响应不是合法 JSON');
     }
   } catch (error) {
+    log.warn('media request failed', {
+      ...requestLog,
+      ...(responseStatus !== undefined ? { status: responseStatus } : {}),
+      durationMs: Date.now() - startedAt,
+      error: mediaRequestParamsForLog(error instanceof Error ? error.message : String(error)),
+    });
     if (error instanceof MediaInvocationError) throw error;
     const aborted = error instanceof Error && error.name === 'AbortError';
     if (input.operation === 'poll') {
@@ -562,6 +662,81 @@ async function prepareRequestBody(
   return output;
 }
 
+async function localImagePath(
+  ref: string,
+  state: { localInputs: number; localBytes: number },
+): Promise<string> {
+  let resolved: { absPath: string; mimeType: string };
+  try {
+    if (ref.startsWith('cindy-media://')) resolved = blobStore.resolveSafe(ref);
+    else if (ref.startsWith('xdt-image://')) resolved = imageCacheStore.resolveSafe(ref);
+    else throw new Error('unsupported local media reference');
+  } catch {
+    throw new MediaInvocationError(
+      'MEDIA_INPUT_INVALID',
+      '第三方 Provider 参考图必须是 Cindy 本地媒体引用',
+    );
+  }
+  const stat = await fs.stat(resolved.absPath);
+  state.localInputs += 1;
+  state.localBytes += stat.size;
+  if (
+    state.localInputs > MAX_LOCAL_MEDIA_INPUTS ||
+    state.localBytes > MAX_LOCAL_MEDIA_INPUT_TOTAL_BYTES ||
+    stat.size <= 0 ||
+    stat.size > MAX_LOCAL_MEDIA_INPUT_BYTES ||
+    !resolved.mimeType.startsWith('image/')
+  ) {
+    throw new MediaInvocationError('MEDIA_INPUT_INVALID', '参考图数量、大小或格式不受支持');
+  }
+  return resolved.absPath;
+}
+
+async function providerImageRequest(
+  invocation: StoredMediaInvocation,
+  body: Record<string, unknown>,
+): Promise<{
+  prompt: string;
+  imagePaths: string[];
+  aspectRatio?: GhostImageAspectRatio;
+}> {
+  const prompt = body.prompt;
+  if (typeof prompt !== 'string' || prompt.trim().length === 0 || prompt.length > 100_000) {
+    throw new MediaInvocationError('REQUEST_INVALID', 'prompt 必须是非空字符串');
+  }
+  let aspectRatio: GhostImageAspectRatio | undefined;
+  if (body.aspect_ratio !== undefined) {
+    if (
+      typeof body.aspect_ratio !== 'string' ||
+      !(GHOST_IMAGE_ASPECT_RATIOS as readonly string[]).includes(body.aspect_ratio)
+    ) {
+      throw new MediaInvocationError(
+        'REQUEST_INVALID',
+        `aspect_ratio 只支持 ${GHOST_IMAGE_ASPECT_RATIOS.join(' / ')}`,
+      );
+    }
+    aspectRatio = body.aspect_ratio as GhostImageAspectRatio;
+  }
+  const imagePaths: string[] = [];
+  if (invocation.capability === 'image.edit') {
+    const raw = body.image;
+    const refs = typeof raw === 'string' ? [raw] : Array.isArray(raw) ? raw : [];
+    if (
+      refs.length === 0 ||
+      refs.some((value) => typeof value !== 'string' || value.length === 0)
+    ) {
+      throw new MediaInvocationError('REQUEST_INVALID', 'image.edit 必须提供本地图片引用');
+    }
+    const state = { localInputs: 0, localBytes: 0 };
+    for (const ref of refs as string[]) imagePaths.push(await localImagePath(ref, state));
+  }
+  return {
+    prompt,
+    imagePaths,
+    ...(aspectRatio ? { aspectRatio } : {}),
+  };
+}
+
 function maxResultBytes(kind: MediaResultKind): number {
   if (kind === 'image') return MAX_IMAGE_RESULT_BYTES;
   if (kind === 'audio') return MAX_AUDIO_RESULT_BYTES;
@@ -776,6 +951,122 @@ async function persistCompletedInvocation(
   return failure('INTERNAL', '媒体结果已生成，但本地未能保存最终结果');
 }
 
+async function submitProviderInvocation(
+  invocation: StoredMediaInvocation,
+  body: Record<string, unknown>,
+  scope: MediaAuthScope,
+  db: DbClient,
+): Promise<Record<string, unknown>> {
+  const providerId = invocation.guide.connection.providerId;
+  const providerModel = resolveProviderMediaModel(
+    providerId,
+    invocation.modelId,
+    invocation.capability,
+  );
+  if (!providerModel) {
+    return failure('MODEL_NOT_AVAILABLE', '该第三方媒体模型或执行来源已不可用，本次生成未发出');
+  }
+  const input = await providerImageRequest(invocation, body);
+  assertAuthScope(scope, invocation.owner);
+  const claimed = await transitionMediaInvocation(
+    {
+      id: invocation.id,
+      owner: invocation.owner,
+      from: 'prepared',
+      to: 'submitting',
+    },
+    db,
+  );
+  if (!claimed) {
+    const current = await getMediaInvocation(invocation.id, invocation.owner, db);
+    assertAuthScope(scope, invocation.owner);
+    return failure(
+      'INVOCATION_ALREADY_USED',
+      `该 invocation 当前状态为 ${current?.state ?? 'unknown'}；付费提交不可重复执行`,
+    );
+  }
+  assertAuthScope(scope, invocation.owner);
+  try {
+    const result = await invokeProviderMedia({
+      providerId,
+      modelId: providerModel.id,
+      capability: invocation.capability,
+      ...input,
+      signal: AbortSignal.timeout(invocation.guide.request.timeoutMs),
+    });
+    assertAuthScope(scope, invocation.owner);
+    if (
+      result.buffer.byteLength === 0 ||
+      result.buffer.byteLength > MAX_IMAGE_RESULT_BYTES ||
+      !result.mimeType.startsWith('image/') ||
+      !blobStore.supportedMime(result.mimeType)
+    ) {
+      throw new MediaInvocationError('MEDIA_RESULT_INVALID', '第三方 Provider 返回了无效图片');
+    }
+    const stored = await ingestMedia(
+      {
+        buffer: result.buffer,
+        mimeType: result.mimeType,
+        refs: [],
+        assertStillValid: () => assertAuthScope(scope, invocation.owner),
+      },
+      db.drizzle,
+    );
+    assertAuthScope(scope, invocation.owner);
+    const media = { xdt_image_urls: [stored.url] };
+    const responseJson = JSON.stringify(media);
+    const persisted = await transitionMediaInvocation(
+      {
+        id: invocation.id,
+        owner: invocation.owner,
+        from: 'submitting',
+        to: 'pending',
+        responseJson,
+      },
+      db,
+    );
+    assertAuthScope(scope, invocation.owner);
+    if (!persisted) {
+      await transitionMediaInvocation(
+        {
+          id: invocation.id,
+          owner: invocation.owner,
+          from: 'submitting',
+          to: 'unknown',
+        },
+        db,
+      ).catch(() => false);
+      return submissionOutcomeUnknown('第三方媒体已生成，但本地未能保存调用结果；不要自动重提');
+    }
+    return persistCompletedInvocation(
+      { ...invocation, state: 'pending', responseJson },
+      media,
+      scope,
+      db,
+    );
+  } catch (error) {
+    await transitionMediaInvocation(
+      {
+        id: invocation.id,
+        owner: invocation.owner,
+        from: 'submitting',
+        to: 'unknown',
+      },
+      db,
+    ).catch(() => false);
+    log.warn('provider media submission failed after claim', {
+      providerId,
+      modelId: providerModel.id,
+      error: error instanceof Error ? error.message : String(error),
+    });
+    return submissionOutcomeUnknown(
+      error instanceof Error
+        ? `第三方媒体请求未能确认结果：${error.message}`
+        : '第三方媒体请求未能确认结果；不要自动重提',
+    );
+  }
+}
+
 async function materializeSyncInvocation(
   invocation: StoredMediaInvocation,
   response: unknown,
@@ -825,6 +1116,7 @@ async function materializeSyncInvocation(
 }
 
 async function prepareInvocation(
+  providerId: string | undefined,
   modelId: string,
   capability: MediaCapability,
 ): Promise<Record<string, unknown>> {
@@ -835,61 +1127,82 @@ async function prepareInvocation(
   assertAuthScope(scope);
   const models = await listAvailableMediaModels(capability);
   assertAuthScope(scope);
-  const model = models.find((candidate) => candidate.id === modelId);
+  const matchingModels = models.filter(
+    (candidate) => candidate.id === modelId && (!providerId || candidate.providerId === providerId),
+  );
+  if (!providerId && matchingModels.length > 1) {
+    return failure(
+      'MODEL_NOT_AVAILABLE',
+      '该模型同时来自多个 Provider，请从模型目录选择精确来源并在 prepare 时传入 provider_id',
+    );
+  }
+  const model = matchingModels[0];
   if (!model) {
-    return failure('MODEL_NOT_AVAILABLE', '该模型当前不可见，或不是请求的媒体类型');
+    return failure('MODEL_NOT_AVAILABLE', '该模型或指定 Provider 当前不可见，或不是请求的媒体类型');
   }
-  let resolvedGuide: ResolvedMediaInvocationGuide;
-  try {
-    resolvedGuide = await fetchMediaInvocationGuide(modelId);
-    assertAuthScope(scope);
-  } catch (error) {
-    if (error instanceof ServerApiError && error.code === 'MEDIA_INVOCATION_GUIDE_NOT_FOUND') {
-      return failure('GUIDE_NOT_AVAILABLE', '该模型当前没有可用的调用说明', false, {
-        outcomeKnown: true,
-        allowedActions: GUIDE_FALLBACK_ACTIONS,
-      });
+  let preparedGuide: PreparedMediaInvocationGuide;
+  if (model.providerId !== 'xd') {
+    if (capability !== 'image.generate' && capability !== 'image.edit') {
+      return failure('CAPABILITY_NOT_SUPPORTED', '该第三方 Provider 当前不支持请求的媒体能力');
     }
-    if (error instanceof MediaGuideCompatibilityError) {
-      log.warn('media Guide rejected by current client', {
-        modelId,
-        code: error.code,
-        detail: error.detail,
-      });
-      return failure(error.code, error.message, false, {
-        outcomeKnown: true,
-        allowedActions: GUIDE_FALLBACK_ACTIONS,
-      });
+    const providerModel = resolveProviderMediaModel(model.providerId, modelId, capability);
+    if (!providerModel) {
+      return failure('MODEL_NOT_AVAILABLE', '该第三方媒体模型或执行来源当前不可用');
     }
-    if (error instanceof ServerApiError) {
-      return failure('GUIDE_SERVICE_UNAVAILABLE', '媒体调用说明暂时无法读取，请稍后重试', true, {
-        outcomeKnown: true,
-        allowedActions: GUIDE_FALLBACK_ACTIONS,
-      });
+    preparedGuide = providerImageGuide(providerModel, capability);
+  } else {
+    let resolvedGuide: ResolvedMediaInvocationGuide;
+    try {
+      resolvedGuide = await fetchMediaInvocationGuide(modelId);
+      assertAuthScope(scope);
+    } catch (error) {
+      if (error instanceof ServerApiError && error.code === 'MEDIA_INVOCATION_GUIDE_NOT_FOUND') {
+        return failure('GUIDE_NOT_AVAILABLE', '该模型当前没有可用的调用说明', false, {
+          outcomeKnown: true,
+          allowedActions: GUIDE_FALLBACK_ACTIONS,
+        });
+      }
+      if (error instanceof MediaGuideCompatibilityError) {
+        log.warn('media Guide rejected by current client', {
+          modelId,
+          code: error.code,
+          detail: error.detail,
+        });
+        return failure(error.code, error.message, false, {
+          outcomeKnown: true,
+          allowedActions: GUIDE_FALLBACK_ACTIONS,
+        });
+      }
+      if (error instanceof ServerApiError) {
+        return failure('GUIDE_SERVICE_UNAVAILABLE', '媒体调用说明暂时无法读取，请稍后重试', true, {
+          outcomeKnown: true,
+          allowedActions: GUIDE_FALLBACK_ACTIONS,
+        });
+      }
+      throw error;
     }
-    throw error;
-  }
-  const operation = resolvedGuide.guide.operations.find(
-    (candidate) => candidate.capability === capability,
-  );
-  if (!operation) {
-    return failure(
-      'CAPABILITY_NOT_SUPPORTED',
-      '该模型的调用协议当前不支持请求的媒体能力',
-      false,
-      {
-        outcomeKnown: true,
-        allowedActions: GUIDE_FALLBACK_ACTIONS,
-      },
+    const operation = resolvedGuide.guide.operations.find(
+      (candidate) => candidate.capability === capability,
     );
+    if (!operation) {
+      return failure(
+        'CAPABILITY_NOT_SUPPORTED',
+        '该模型的调用协议当前不支持请求的媒体能力',
+        false,
+        {
+          outcomeKnown: true,
+          allowedActions: GUIDE_FALLBACK_ACTIONS,
+        },
+      );
+    }
+    const { operations: _operations, ...guideProtocol } = resolvedGuide.guide;
+    void _operations;
+    preparedGuide = {
+      modelId: resolvedGuide.modelId,
+      ...guideProtocol,
+      ...operation,
+    };
   }
-  const { operations: _operations, ...guideProtocol } = resolvedGuide.guide;
-  void _operations;
-  const preparedGuide: PreparedMediaInvocationGuide = {
-    modelId: resolvedGuide.modelId,
-    ...guideProtocol,
-    ...operation,
-  };
   await pruneInvocations(owner, db);
   assertAuthScope(scope);
   if ((await countMediaInvocations(owner, db)) >= MAX_INVOCATIONS) {
@@ -912,8 +1225,9 @@ async function prepareInvocation(
     ok: true,
     status: 'prepared',
     invocation_id: id,
+    provider_id: model.providerId,
     model_id: modelId,
-    model_name: model.name,
+    model_name: model.name ?? model.id,
     capability,
     guide_revision: preparedGuide.revision,
     instructions: preparedGuide.instructions,
@@ -949,6 +1263,17 @@ async function submitInvocation(
   if (invocation.state === 'complete') {
     return completedInvocationResult(invocation);
   }
+  if (
+    invocation.state === 'pending' &&
+    isClientProviderInvocation(invocation) &&
+    invocation.responseJson
+  ) {
+    const media = persistedResponse(invocation);
+    if (!media || typeof media !== 'object' || Array.isArray(media)) {
+      return failure('MEDIA_RESULT_INVALID', '第三方媒体调用保存的结果不合法');
+    }
+    return persistCompletedInvocation(invocation, media as Record<string, unknown>, scope, db);
+  }
   if (
     invocation.state === 'pending' &&
     invocation.guide.response.mode === 'sync' &&
@@ -975,11 +1300,14 @@ async function submitInvocation(
     assertAuthScope(scope, invocation.owner);
     return failure('INVOCATION_EXPIRED', '调用准备已超过 5 分钟，请重新查询模型并 prepare');
   }
+  if (isClientProviderInvocation(invocation)) {
+    return submitProviderInvocation(invocation, body, scope, db);
+  }
   // prepare 与实际付费提交之间可能隔着 Agent 组装参数的时间；提交边界重新读取
   // Gateway 清单和客户端停用状态，避免模型/供应商刚被停用后仍发出新请求。
   const models = await listAvailableMediaModels(invocation.capability);
   assertAuthScope(scope, invocation.owner);
-  if (!models.some((model) => model.id === invocation.modelId)) {
+  if (!models.some((model) => model.providerId === 'xd' && model.id === invocation.modelId)) {
     return failure('MODEL_NOT_AVAILABLE', '该模型已下架或被停用，本次生成未发出');
   }
   const requestBody = await prepareRequestBody(body, invocation.guide);
@@ -1017,6 +1345,10 @@ async function submitInvocation(
   assertAuthScope(scope, invocation.owner);
   try {
     const response = await dispatchRequest({
+      invocationId: invocation.id,
+      providerId: invocation.guide.connection.providerId,
+      modelId: invocation.modelId,
+      capability: invocation.capability,
       connection,
       method: invocation.guide.request.method,
       path: invocation.guide.request.path,
@@ -1243,6 +1575,10 @@ async function pollInvocation(invocation: StoredMediaInvocation): Promise<Record
   }
   try {
     const response = await dispatchRequest({
+      invocationId: invocation.id,
+      providerId: invocation.guide.connection.providerId,
+      modelId: invocation.modelId,
+      capability: invocation.capability,
       connection: resolveConnection(invocation.guide.connection.providerId),
       method: guide.method,
       path: pollPath(guide.path, invocation.taskId),
@@ -1362,6 +1698,7 @@ export async function callCindyMedia(
         ok: true,
         models: availability.models.map((model) => ({
           id: model.id,
+          provider_id: model.providerId,
           ...(model.name ? { name: model.name } : {}),
           ...(model.mode ? { mode: model.mode } : {}),
         })),
@@ -1374,7 +1711,11 @@ export async function callCindyMedia(
       };
     }
     if (request.action === 'prepare') {
-      return prepareInvocation(request.modelId, request.capability as MediaCapability);
+      return prepareInvocation(
+        request.providerId,
+        request.modelId,
+        request.capability as MediaCapability,
+      );
     }
     const invocation = await requireInvocation(request.invocationId);
     return await (request.action === 'request'
diff --git a/apps/desktop/src/main/cindy-media/mediaCapabilities.ts b/apps/desktop/src/main/cindy-media/mediaCapabilities.ts
new file mode 100644
index 0000000000..af4caf2d8c
--- /dev/null
+++ b/apps/desktop/src/main/cindy-media/mediaCapabilities.ts
@@ -0,0 +1,21 @@
+import type { MediaCapability } from '@cindy/model-providers';
+
+const MEDIA_CAPABILITY_REQUIREMENTS: Record<MediaCapability, { input: string[]; output: string }> = {
+  'image.generate': { input: ['text'], output: 'image' },
+  'image.edit': { input: ['text', 'image'], output: 'image' },
+  'video.generate': { input: ['text'], output: 'video' },
+  'video.image_to_video': { input: ['text', 'image'], output: 'video' },
+};
+
+export function supportsMediaCapability(
+  modalities: { input: string[]; output: string[] } | undefined,
+  capability: MediaCapability,
+): boolean {
+  if (!modalities) return false;
+  const requirement = MEDIA_CAPABILITY_REQUIREMENTS[capability];
+  const inputs = new Set(modalities.input);
+  return (
+    modalities.output.includes(requirement.output) &&
+    requirement.input.every((input) => inputs.has(input))
+  );
+}
diff --git a/apps/desktop/src/main/cindy-media/mediaRequestLog.ts b/apps/desktop/src/main/cindy-media/mediaRequestLog.ts
new file mode 100644
index 0000000000..0a68c057d4
--- /dev/null
+++ b/apps/desktop/src/main/cindy-media/mediaRequestLog.ts
@@ -0,0 +1,98 @@
+import { redactSensitiveText } from '@cindy/maker-shared/error-redaction';
+
+const MAX_LOG_STRING_CHARS = 20_000;
+const MAX_LOG_DEPTH = 24;
+const SENSITIVE_PARAM_NAME =
+  /(?:^|[-_.])(authorization|proxy[-_]?authorization|api[-_]?key|access[-_]?key(?:[-_]?id)?|private[-_]?key|key|access[-_]?token|refresh[-_]?token|id[-_]?token|token|secret|password|passwd|signature|credential|cookie|session)(?:$|[-_.])/i;
+
+function boundedText(value: string): string {
+  const redacted = redactSensitiveText(value);
+  if (redacted.length <= MAX_LOG_STRING_CHARS) return redacted;
+  return `${redacted.slice(0, MAX_LOG_STRING_CHARS)}...[truncated ${redacted.length - MAX_LOG_STRING_CHARS} chars]`;
+}
+
+function dataUrlSummary(value: string): string | null {
+  const match = /^data:([^;,]+)(;base64)?,([\s\S]*)$/i.exec(value);
+  if (!match) return null;
+  const mimeType = match[1].toLowerCase();
+  const encoded = match[3].replace(/\s/g, '');
+  const bytes = match[2]
+    ? Math.max(
+        0,
+        Math.floor((encoded.length * 3) / 4) -
+          (encoded.endsWith('==') ? 2 : encoded.endsWith('=') ? 1 : 0),
+      )
+    : Buffer.byteLength(encoded, 'utf8');
+  return `[data URL mime=${mimeType} bytes=${bytes}]`;
+}
+
+function urlForLog(rawUrl: string, stripFragment: boolean): string {
+  try {
+    const url = new URL(rawUrl);
+    if (url.username) url.username = '[REDACTED]';
+    if (url.password) url.password = '[REDACTED]';
+    for (const key of [...url.searchParams.keys()]) {
+      if (SENSITIVE_PARAM_NAME.test(key)) url.searchParams.set(key, '[REDACTED]');
+    }
+    if (stripFragment) url.hash = '';
+    else if (url.hash) url.hash = boundedText(url.hash);
+    return url.toString();
+  } catch {
+    return boundedText(rawUrl);
+  }
+}
+
+/** 保留实际请求 URL，只脱敏可能携带凭证的 query 值；fragment 不会发给上游。 */
+export function mediaRequestUrlForLog(rawUrl: string): string {
+  return urlForLog(rawUrl, true);
+}
+
+/**
+ * 请求参数日志保留真实结构和值；凭证、data URL、二进制与本地文件内容只留描述，
+ * 避免日志泄露密钥或被图片/视频 base64 撑爆。
+ */
+export function mediaRequestParamsForLog(value: unknown): unknown {
+  const seen = new WeakSet<object>();
+
+  const visit = (item: unknown, key: string | null, depth: number): unknown => {
+    if (key && SENSITIVE_PARAM_NAME.test(key)) return '[REDACTED]';
+    if (typeof item === 'string') {
+      const dataSummary = dataUrlSummary(item);
+      if (dataSummary) return dataSummary;
+      if (/^https?:\/\/\S+$/i.test(item)) return urlForLog(item, false);
+      return boundedText(item);
+    }
+    if (
+      item === null ||
+      typeof item === 'number' ||
+      typeof item === 'boolean' ||
+      typeof item === 'undefined'
+    ) {
+      return item;
+    }
+    if (typeof item === 'bigint') return item.toString();
+    if (Buffer.isBuffer(item) || item instanceof Uint8Array) {
+      return `[binary bytes=${item.byteLength}]`;
+    }
+    if (typeof Blob !== 'undefined' && item instanceof Blob) {
+      return `[blob mime=${item.type || 'unknown'} bytes=${item.size}]`;
+    }
+    if (typeof item !== 'object') return boundedText(String(item));
+    if (depth >= MAX_LOG_DEPTH) return '[max depth]';
+    if (seen.has(item)) return '[circular]';
+    seen.add(item);
+    if (Array.isArray(item)) {
+      const result = item.map((child) => visit(child, null, depth + 1));
+      seen.delete(item);
+      return result;
+    }
+    const result: Record<string, unknown> = {};
+    for (const [childKey, child] of Object.entries(item as Record<string, unknown>)) {
+      result[childKey] = visit(child, childKey, depth + 1);
+    }
+    seen.delete(item);
+    return result;
+  };
+
+  return visit(value, null, 0);
+}
diff --git a/apps/desktop/src/main/cindy-media/providerMediaRuntime.ts b/apps/desktop/src/main/cindy-media/providerMediaRuntime.ts
new file mode 100644
index 0000000000..4f8de9dc50
--- /dev/null
+++ b/apps/desktop/src/main/cindy-media/providerMediaRuntime.ts
@@ -0,0 +1,69 @@
+import type { MediaCapability } from '@cindy/model-providers';
+import type { GhostImageAspectRatio } from '../../shared/ghost.js';
+import { supportsMediaCapability } from './mediaCapabilities.js';
+
+export interface ProviderMediaRuntimeModel {
+  id: string;
+  name: string;
+  providerId: string;
+  mode: 'image_generation' | 'video_generation';
+  modalities: { input: string[]; output: string[] };
+  officialDocs?: string;
+}
+
+export interface ProviderMediaRuntimeRequest {
+  providerId: string;
+  modelId: string;
+  capability: MediaCapability;
+  prompt: string;
+  imagePaths: string[];
+  aspectRatio?: GhostImageAspectRatio;
+  signal?: AbortSignal;
+}
+
+export interface ProviderMediaRuntimeResult {
+  buffer: Buffer;
+  mimeType: string;
+}
+
+interface ProviderMediaRuntime {
+  listModels(): ProviderMediaRuntimeModel[];
+  invoke(request: ProviderMediaRuntimeRequest): Promise<ProviderMediaRuntimeResult>;
+}
+
+let runtime: ProviderMediaRuntime | null = null;
+
+export function configureProviderMediaRuntime(next: ProviderMediaRuntime): void {
+  runtime = next;
+}
+
+export function listProviderMediaModels(): ProviderMediaRuntimeModel[] {
+  return runtime?.listModels() ?? [];
+}
+
+export function resolveProviderMediaModel(
+  providerId: string,
+  modelId: string,
+  capability: MediaCapability,
+): ProviderMediaRuntimeModel | null {
+  return (
+    listProviderMediaModels().find(
+      (model) =>
+        model.providerId === providerId &&
+        model.id === modelId &&
+        supportsMediaCapability(model.modalities, capability),
+    ) ?? null
+  );
+}
+
+export async function invokeProviderMedia(
+  request: ProviderMediaRuntimeRequest,
+): Promise<ProviderMediaRuntimeResult> {
+  const active = resolveProviderMediaModel(
+    request.providerId,
+    request.modelId,
+    request.capability,
+  );
+  if (!active || !runtime) throw new Error('第三方媒体模型或执行来源当前不可用');
+  return runtime.invoke(request);
+}
diff --git a/apps/desktop/src/main/cindy-proxy-media/api/gatewayImageClient.ts b/apps/desktop/src/main/cindy-proxy-media/api/gatewayImageClient.ts
index 9e6761f3b1..61464ab731 100644
--- a/apps/desktop/src/main/cindy-proxy-media/api/gatewayImageClient.ts
+++ b/apps/desktop/src/main/cindy-proxy-media/api/gatewayImageClient.ts
@@ -1,3 +1,4 @@
+import { randomUUID } from 'node:crypto';
 import fs from 'node:fs/promises';
 import path from 'node:path';
 import type {
@@ -7,6 +8,10 @@ import type {
 } from '../types.js';
 import type { LiziMcpLogger } from '@cindy/mcps';
 import type { CindyProxyMediaMaybePromise, CindyProxyMediaProxyConfig } from '../types.js';
+import {
+  mediaRequestParamsForLog,
+  mediaRequestUrlForLog,
+} from '../../cindy-media/mediaRequestLog.js';
 
 export class GatewayImageError extends Error {
   constructor(
@@ -151,6 +156,43 @@ export function createGatewayImageClient(opts: CreateGatewayImageClientOptions):
   const editUrl = joinProxyUrl(baseUrl, opts.proxy.editPath);
   const doFetch = opts.fetchImplementation ?? fetch;
 
+  async function loggedFetch(
+    url: string,
+    init: RequestInit,
+    model: string,
+    params: unknown,
+  ): Promise<Response> {
+    const requestId = randomUUID();
+    const startedAt = Date.now();
+    const requestLog = {
+      requestId,
+      provider: brandLabel,
+      modelId: model,
+      method: init.method ?? 'GET',
+      url: mediaRequestUrlForLog(url),
+    };
+    opts.logger?.info('media request dispatch', {
+      ...requestLog,
+      params: mediaRequestParamsForLog(params),
+    });
+    try {
+      const response = await doFetch(url, init);
+      opts.logger?.info('media request response', {
+        ...requestLog,
+        status: response.status,
+        durationMs: Date.now() - startedAt,
+      });
+      return response;
+    } catch (error) {
+      opts.logger?.warn('media request failed', {
+        ...requestLog,
+        durationMs: Date.now() - startedAt,
+        error: mediaRequestParamsForLog(error instanceof Error ? error.message : String(error)),
+      });
+      throw error;
+    }
+  }
+
   async function requireApiKey(): Promise<string> {
     const key = await Promise.resolve(opts.getApiKey());
     if (!key) {
@@ -190,15 +232,20 @@ export function createGatewayImageClient(opts: CreateGatewayImageClientOptions):
     // 停用轴派发前重查(PR #744 review 第二十一轮):凭证获取是 await,期间该
     // (供应商, 模型) 可能被用户停用 —— payload 就绪、请求发出的紧前再验一次。
     beforeDispatch?.(params.model);
-    const res = await doFetch(generateUrl, {
-      method: 'POST',
-      headers: {
-        Authorization: `Bearer ${apiKey}`,
-        'Content-Type': 'application/json',
+    const res = await loggedFetch(
+      generateUrl,
+      {
+        method: 'POST',
+        headers: {
+          Authorization: `Bearer ${apiKey}`,
+          'Content-Type': 'application/json',
+        },
+        body: JSON.stringify(body),
+        signal,
       },
-      body: JSON.stringify(body),
-      signal,
-    });
+      params.model,
+      body,
+    );
     return parseResponse(res, params.model, brandLabel);
   }
 
@@ -223,22 +270,37 @@ export function createGatewayImageClient(opts: CreateGatewayImageClientOptions):
     if (allowSizeQuality) form.append('size', params.size ?? 'auto');
     if (params.quality) form.append('quality', params.quality);
 
+    const imageParams: Array<{ filename: string; mimeType: string; bytes: number }> = [];
     for (const p of params.imagePaths) {
       const buf = await fs.readFile(p);
       const filename = path.basename(p);
-      form.append('image[]', new Blob([buf], { type: mimeFromFilename(filename) }), filename);
+      const mimeType = mimeFromFilename(filename);
+      form.append('image[]', new Blob([buf], { type: mimeType }), filename);
+      imageParams.push({ filename, mimeType, bytes: buf.byteLength });
     }
 
     // 同上:凭证获取 + 逐张 fs.readFile 都是 await,提交紧前重查。
     beforeDispatch?.(params.model);
-    const res = await doFetch(editUrl, {
-      method: 'POST',
-      headers: {
-        Authorization: `Bearer ${apiKey}`,
+    const res = await loggedFetch(
+      editUrl,
+      {
+        method: 'POST',
+        headers: {
+          Authorization: `Bearer ${apiKey}`,
+        },
+        body: form as unknown as BodyInit,
+        signal,
       },
-      body: form as unknown as BodyInit,
-      signal,
-    });
+      params.model,
+      {
+        model: params.model,
+        prompt: params.prompt,
+        n: params.n ?? 1,
+        ...(allowSizeQuality ? { size: params.size ?? 'auto' } : {}),
+        ...(params.quality ? { quality: params.quality } : {}),
+        images: imageParams,
+      },
+    );
     return parseResponse(res, params.model, brandLabel);
   }
 
diff --git a/apps/desktop/src/main/model-access/__tests__/mediaModels.test.ts b/apps/desktop/src/main/model-access/__tests__/mediaModels.test.ts
index 623284c125..534cac5f22 100644
--- a/apps/desktop/src/main/model-access/__tests__/mediaModels.test.ts
+++ b/apps/desktop/src/main/model-access/__tests__/mediaModels.test.ts
@@ -2,6 +2,7 @@ import { beforeEach, describe, expect, it, vi } from 'vitest';
 
 const serverApiFetchMock = vi.hoisted(() => vi.fn());
 const readModelDisableOverridesMock = vi.hoisted(() => vi.fn());
+const listProviderMediaModelsMock = vi.hoisted(() => vi.fn());
 
 vi.mock('../../serverApiClient.js', () => ({
   serverApiFetch: serverApiFetchMock,
@@ -21,6 +22,9 @@ vi.mock('../../clientEndpointsService.js', () => ({
 vi.mock('../../maker-host/model-disable-store.js', () => ({
   readModelDisableOverrides: readModelDisableOverridesMock,
 }));
+vi.mock('../../cindy-media/providerMediaRuntime.js', () => ({
+  listProviderMediaModels: listProviderMediaModelsMock,
+}));
 
 import {
   fetchMediaInvocationGuide,
@@ -73,6 +77,7 @@ describe('listAvailableMediaModels', () => {
   beforeEach(() => {
     serverApiFetchMock.mockReset().mockResolvedValue(payload);
     readModelDisableOverridesMock.mockReset().mockReturnValue({});
+    listProviderMediaModelsMock.mockReset().mockReturnValue([]);
   });
 
   it('不带操作筛选时按 Gateway mode 返回图片/视频模型', async () => {
@@ -99,6 +104,25 @@ describe('listAvailableMediaModels', () => {
     ]);
   });
 
+  it('同名媒体模型按 providerId 保留为两个可选来源', async () => {
+    listProviderMediaModelsMock.mockReturnValue([
+      {
+        id: 'image-with-guide',
+        name: 'Provider Image',
+        providerId: 'openai',
+        mode: 'image_generation',
+        modalities: { input: ['text', 'image'], output: ['image'] },
+      },
+    ]);
+
+    const models = await listAvailableMediaModels('image.generate');
+
+    expect(models.filter((model) => model.id === 'image-with-guide')).toMatchObject([
+      { id: 'image-with-guide', providerId: 'xd' },
+      { id: 'image-with-guide', providerId: 'openai' },
+    ]);
+  });
+
   it('叠加客户端现有 XD provider/model 停用准入', async () => {
     readModelDisableOverridesMock.mockReturnValueOnce({
       disabledModels: { 'xd:image-without-guide': true },
diff --git a/apps/desktop/src/main/model-access/mediaModels.ts b/apps/desktop/src/main/model-access/mediaModels.ts
index 5485e450ee..c344487314 100644
--- a/apps/desktop/src/main/model-access/mediaModels.ts
+++ b/apps/desktop/src/main/model-access/mediaModels.ts
@@ -16,6 +16,8 @@ import {
   type ResolvedMediaInvocationGuide,
 } from '../../shared/mediaInvocation.js';
 import { getClientEndpoint } from '../clientEndpointsService.js';
+import { supportsMediaCapability } from '../cindy-media/mediaCapabilities.js';
+import { listProviderMediaModels } from '../cindy-media/providerMediaRuntime.js';
 import { readModelDisableOverrides } from '../maker-host/model-disable-store.js';
 import { serverApiFetch, ServerApiError } from '../serverApiClient.js';
 
@@ -62,11 +64,13 @@ export interface UnavailableMediaModel {
 }
 
 export interface ExecutableMediaModelsResult {
-  models: ModelCatalogEntry[];
+  models: ExecutableMediaModel[];
   unavailable: UnavailableMediaModel[];
   candidateCount: number;
 }
 
+export type ExecutableMediaModel = ModelCatalogEntry & { providerId: string };
+
 interface ExecutableMediaSnapshot {
   models: ModelCatalogEntry[];
   capabilitiesByModel: Map<string, ReadonlySet<MediaCapability>>;
@@ -99,28 +103,38 @@ export function isMediaModelExecutable(
   return executableMediaSnapshot?.capabilitiesByModel.get(modelId)?.has(capability) === true;
 }
 
-const MEDIA_CAPABILITY_REQUIREMENTS: Record<
-  MediaCapability,
-  { input: string[]; inputAny?: string[]; output: string }
-> = {
-  'image.generate': { input: ['text'], output: 'image' },
-  'image.edit': { input: ['text', 'image'], output: 'image' },
-  'video.generate': { input: ['text'], output: 'video' },
-  'video.image_to_video': { input: ['text', 'image'], output: 'video' },
-};
+export { supportsMediaCapability } from '../cindy-media/mediaCapabilities.js';
 
-export function supportsMediaCapability(
-  modalities: { input: string[]; output: string[] } | undefined,
-  capability: MediaCapability,
-): boolean {
-  if (!modalities) return false;
-  const requirement = MEDIA_CAPABILITY_REQUIREMENTS[capability];
-  const inputs = new Set(modalities.input);
-  return (
-    modalities.output.includes(requirement.output) &&
-    requirement.input.every((input) => inputs.has(input)) &&
-    (requirement.inputAny === undefined || requirement.inputAny.some((input) => inputs.has(input)))
-  );
+function availableProviderMediaModels(capability?: MediaCapability): ExecutableMediaModel[] {
+  return listProviderMediaModels()
+    .filter(
+      (model) => capability === undefined || supportsMediaCapability(model.modalities, capability),
+    )
+    .map((model) => ({
+      id: model.id,
+      name: model.name,
+      providerId: model.providerId,
+      mode: model.mode,
+      modalities: {
+        input: [...model.modalities.input],
+        output: [...model.modalities.output],
+      },
+    }));
+}
+
+function mergeMediaModels(
+  preferred: readonly ExecutableMediaModel[],
+  fallback: readonly ExecutableMediaModel[],
+): ExecutableMediaModel[] {
+  const seen = new Set<string>();
+  const models: ExecutableMediaModel[] = [];
+  for (const model of [...preferred, ...fallback]) {
+    const key = `${model.providerId}\u0000${model.id}`;
+    if (seen.has(key)) continue;
+    seen.add(key);
+    models.push(model);
+  }
+  return models;
 }
 
 /**
@@ -184,12 +198,19 @@ async function fetchGatewayMediaModels(): Promise<ModelCatalogEntry[]> {
 
 export async function listAvailableMediaModels(
   capability?: MediaCapability,
-): Promise<ModelCatalogEntry[]> {
-  return filterEnabledGatewayMediaModels(
-    await fetchGatewayMediaModels(),
-    capability,
-    readModelDisableOverrides(),
-  );
+): Promise<ExecutableMediaModel[]> {
+  const providerModels = availableProviderMediaModels(capability);
+  try {
+    const gatewayModels = filterEnabledGatewayMediaModels(
+      await fetchGatewayMediaModels(),
+      capability,
+      readModelDisableOverrides(),
+    ).map((model) => ({ ...model, providerId: CINDY_AI_PROVIDER_ID }));
+    return mergeMediaModels(gatewayModels, providerModels);
+  } catch (error) {
+    if (providerModels.length > 0) return providerModels;
+    throw error;
+  }
 }
 
 export async function fetchMediaInvocationGuide(
@@ -417,7 +438,18 @@ export async function listExecutableMediaModels(
   capabilities: readonly MediaCapability[] = [],
   options: { includeDisabled?: boolean; forceRefresh?: boolean } = {},
 ): Promise<ExecutableMediaModelsResult> {
-  const snapshot = await getExecutableMediaSnapshot(options.forceRefresh === true);
+  const providerModels = availableProviderMediaModels().filter((model) =>
+    capabilities.every((capability) => supportsMediaCapability(model.modalities, capability)),
+  );
+  let snapshot: ExecutableMediaSnapshot;
+  try {
+    snapshot = await getExecutableMediaSnapshot(options.forceRefresh === true);
+  } catch (error) {
+    if (providerModels.length > 0) {
+      return { models: providerModels, unavailable: [], candidateCount: providerModels.length };
+    }
+    throw error;
+  }
   const candidates = filterEnabledGatewayMediaModels(
     snapshot.models,
     undefined,
@@ -425,12 +457,12 @@ export async function listExecutableMediaModels(
   ).filter((model) =>
     capabilities.every((capability) => supportsMediaCapability(model.modalities, capability)),
   );
-  const models: ModelCatalogEntry[] = [];
+  const gatewayModels: ExecutableMediaModel[] = [];
   const unavailable: UnavailableMediaModel[] = [];
   for (const model of candidates) {
     const supported = snapshot.capabilitiesByModel.get(model.id);
     if (supported && capabilities.every((capability) => supported.has(capability))) {
-      models.push(model);
+      gatewayModels.push({ ...model, providerId: CINDY_AI_PROVIDER_ID });
       continue;
     }
     unavailable.push(
@@ -442,5 +474,10 @@ export async function listExecutableMediaModels(
       },
     );
   }
-  return { models, unavailable, candidateCount: candidates.length };
+  const models = mergeMediaModels(gatewayModels, providerModels);
+  return {
+    models,
+    unavailable,
+    candidateCount: candidates.length + providerModels.length,
+  };
 }
diff --git a/apps/desktop/src/preload/preload.ts b/apps/desktop/src/preload/preload.ts
index a6424f7cc9..b6e5e973f1 100644
--- a/apps/desktop/src/preload/preload.ts
+++ b/apps/desktop/src/preload/preload.ts
@@ -846,6 +846,22 @@ const appearanceSettingsInfo = ipcRenderer.sendSync(
   'appearance-settings:get-sync',
 ) as AppearanceSettings | null;
 
+type CindyMediaPreferenceOption = {
+  id: string;
+  label: string;
+  group: string;
+  providerId: string;
+  providerName: string;
+  modelId: string;
+  modelName: string;
+  routing?: import('@cindy/model-providers').Provider['routing'];
+};
+
+type CindyMediaPreferenceKind = {
+  options: CindyMediaPreferenceOption[];
+  defaultModel: CindyMediaPreferenceOption | null;
+};
+
 contextBridge.exposeInMainWorld('electronAPI', {
   platform: process.platform,
   osRelease: ipcRenderer.sendSync('get-os-release') as string,
@@ -1057,22 +1073,10 @@ contextBridge.exposeInMainWorld('electronAPI', {
        * 每类目一份下拉数据(能力键按类目取对应清单)。
        * options 为空或 defaultModel 为 null = 目录没给该类目模型,能力暂不可用。
        */
-      image: {
-        options: Array<{ id: string; label: string }>;
-        defaultModel: { id: string; label: string } | null;
-      };
-      imageEdit: {
-        options: Array<{ id: string; label: string }>;
-        defaultModel: { id: string; label: string } | null;
-      };
-      video: {
-        options: Array<{ id: string; label: string }>;
-        defaultModel: { id: string; label: string } | null;
-      };
-      videoEdit: {
-        options: Array<{ id: string; label: string }>;
-        defaultModel: { id: string; label: string } | null;
-      };
+      image: CindyMediaPreferenceKind;
+      imageEdit: CindyMediaPreferenceKind;
+      video: CindyMediaPreferenceKind;
+      videoEdit: CindyMediaPreferenceKind;
       /** 文本类(快问快答):选项是当前供应商目录的全部文本模型(cat: 编码钉值,
        *  带供应商/模型/徽标等结构化字段供富列表渲染);declaredModel = 身份卡声明
        *  的偏好模型;utilityProfiles = 存量轻量档位钉的展示名表(老钉值回显用)。 */
diff --git a/apps/desktop/src/renderer/cindy-brain/CindyCapabilityPrefs.tsx b/apps/desktop/src/renderer/cindy-brain/CindyCapabilityPrefs.tsx
index 62b2863cca..fe44f0736c 100644
--- a/apps/desktop/src/renderer/cindy-brain/CindyCapabilityPrefs.tsx
+++ b/apps/desktop/src/renderer/cindy-brain/CindyCapabilityPrefs.tsx
@@ -5,18 +5,170 @@
 
 import { useCallback, useState } from 'react';
 import { useTranslation } from 'react-i18next';
-import { Sparkles } from 'lucide-react';
+import { Check, ChevronDown, Sparkles } from 'lucide-react';
 
 import { toast } from '@/lib/toast';
 import { cn } from '@/lib/utils';
-import {
-  OneshotModelPinPicker,
-  type OneshotPinOption,
-} from '@/cindy-brain/OneshotModelPinPicker';
+import { Popover, PopoverContent, PopoverTrigger } from '@/components/ui/popover';
+import { ModelIconMark } from '@/components/new-chat/ModelSelector';
+import { OneshotModelPinPicker, type OneshotPinOption } from '@/cindy-brain/OneshotModelPinPicker';
 
 /** 跟随默认在 select 里的哨兵值(覆盖表里"没有这项"= 跟随默认)。 */
 const FOLLOW_DEFAULT_VALUE = '__default__';
 
+interface MediaModelOption {
+  id: string;
+  modelId: string;
+  label: string;
+  providerId: string;
+  providerName: string;
+  routing?: import('@cindy/model-providers').Provider['routing'];
+}
+
+/** 媒体模型沿用原来的轻量下拉，只在模型名前补来源 Provider 图标。 */
+function MediaModelPicker({
+  value,
+  defaultModel,
+  options,
+  onChange,
+  ariaLabel,
+  dense,
+}: {
+  value?: string;
+  defaultModel: MediaModelOption;
+  options: readonly MediaModelOption[];
+  onChange: (pin: string | null) => void;
+  ariaLabel: string;
+  dense?: boolean;
+}) {
+  const { t } = useTranslation();
+  const [open, setOpen] = useState(false);
+  const current = value
+    ? options.find((option) => option.id === value) ??
+      options.find((option) => option.providerId === 'xd' && option.modelId === value) ??
+      options.find((option) => option.modelId === value)
+    : undefined;
+  const staleValue = value && !current ? value : null;
+  const selected = current ?? defaultModel;
+  const selectable = options.filter((option) => option.id !== defaultModel.id);
+  const triggerLabel =
+    current?.label ??
+    staleValue ??
+    t('settings.ghosts.detail.cindyPrefs.defaultOption', { model: defaultModel.label });
+
+  const select = (pin: string | null): void => {
+    if (pin !== value) onChange(pin);
+    setOpen(false);
+  };
+
+  const optionClass = (active: boolean): string =>
+    cn(
+      'flex w-full items-center gap-2 rounded-lg px-3 py-2 text-left transition-colors',
+      'hover:bg-[var(--model-item-hover)] focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-[var(--focus-ring)]',
+      dense ? 'text-12 leading-5' : 'text-13 leading-5',
+      active && 'bg-[var(--model-item-hover)]',
+    );
+
+  return (
+    <Popover open={open} onOpenChange={setOpen}>
+      <PopoverTrigger asChild>
+        <button
+          type="button"
+          role="combobox"
+          aria-label={ariaLabel}
+          aria-expanded={open}
+          aria-haspopup="listbox"
+          className={cn(
+            'flex h-8 w-[300px] max-w-[60%] min-w-0 shrink items-center gap-2 rounded-full border border-[var(--settings-input-border)] bg-[var(--settings-input-bg)] py-0 pl-3 pr-2.5 text-[var(--settings-input-text)] outline-none focus:ring-2 focus:ring-[var(--focus-ring-soft)]',
+            dense ? 'text-12' : 'text-13',
+          )}
+        >
+          {!staleValue && (
+            <ModelIconMark
+              providerId={selected.providerId}
+              name={selected.providerName}
+              routing={selected.routing}
+              withMargin={false}
+              dense
+            />
+          )}
+          <span className="min-w-0 flex-1 truncate text-left">{triggerLabel}</span>
+          <ChevronDown size={13} className="shrink-0 text-[var(--text-tertiary)]" />
+        </button>
+      </PopoverTrigger>
+      <PopoverContent
+        align="end"
+        sideOffset={6}
+        className="w-[300px] overflow-hidden rounded-xl border border-[var(--model-dropdown-border)] bg-[var(--model-dropdown-bg)] p-1 shadow-[var(--shadow-menu)]"
+      >
+        <div role="listbox" aria-label={ariaLabel} className="max-h-[300px] overflow-y-auto">
+          <button
+            type="button"
+            role="option"
+            aria-selected={value === undefined}
+            className={optionClass(value === undefined)}
+            onClick={() => select(null)}
+          >
+            <ModelIconMark
+              providerId={defaultModel.providerId}
+              name={defaultModel.providerName}
+              routing={defaultModel.routing}
+              withMargin={false}
+              dense
+            />
+            <span className="min-w-0 flex-1 truncate text-[var(--model-item-text)]">
+              {t('settings.ghosts.detail.cindyPrefs.defaultOption', { model: defaultModel.label })}
+            </span>
+            {value === undefined && (
+              <Check size={15} className="shrink-0 text-[var(--model-item-check)]" />
+            )}
+          </button>
+          {selectable.map((option) => {
+            const active = value === option.id;
+            return (
+              <button
+                key={option.id}
+                type="button"
+                role="option"
+                aria-selected={active}
+                title={option.providerName}
+                className={optionClass(active)}
+                onClick={() => select(option.id)}
+              >
+                <ModelIconMark
+                  providerId={option.providerId}
+                  name={option.providerName}
+                  routing={option.routing}
+                  withMargin={false}
+                  dense
+                />
+                <span className="min-w-0 flex-1 truncate text-[var(--model-item-text)]">
+                  {option.label}
+                </span>
+                {active && <Check size={15} className="shrink-0 text-[var(--model-item-check)]" />}
+              </button>
+            );
+          })}
+          {staleValue && (
+            <button
+              type="button"
+              role="option"
+              aria-selected
+              className={optionClass(true)}
+              onClick={() => select(staleValue)}
+            >
+              <span className="min-w-0 flex-1 truncate text-[var(--model-item-text)]">
+                {staleValue}
+              </span>
+              <Check size={15} className="shrink-0 text-[var(--model-item-check)]" />
+            </button>
+          )}
+        </div>
+      </PopoverContent>
+    </Popover>
+  );
+}
+
 /**
  * Ghost 申请的每项 Cindy 能力一行,可钉后端(供应商×模型)。
  * Settings 详情与 Plugin 详情共用这一份实现,避免两个入口产生不同配置口径。
@@ -89,17 +241,18 @@ export function CindyCapabilityPrefs({
         // 按能力键的类目取对应清单。少一个分支的后果不是少个下拉,而是拿
         // **图像模型清单**去填一个文本能力——用户存进去的值链路根本不认。
         const isText = capability.startsWith('text.');
-        const kind = capability === 'video.edit'
-          ? prefs.videoEdit
-          : capability.startsWith('video.')
-            ? prefs.video
-          : isText
-            ? prefs.text
-            : capability.startsWith('embed.')
-              ? prefs.embed
-              : capability === 'image.edit'
-                ? prefs.imageEdit
-                : prefs.image;
+        const kind =
+          capability === 'video.edit'
+            ? prefs.videoEdit
+            : capability.startsWith('video.')
+              ? prefs.video
+              : isText
+                ? prefs.text
+                : capability.startsWith('embed.')
+                  ? prefs.embed
+                  : capability === 'image.edit'
+                    ? prefs.imageEdit
+                    : prefs.image;
         // 目录没给这个类目任何模型 = 能力暂不可用:行照旧显示(插件确实申请了
         // 这项能力),但右侧不给下拉,改一句不可点的灰字,不拿旧型号冒充可选。
         const defaultModel = kind.defaultModel;
@@ -114,10 +267,11 @@ export function CindyCapabilityPrefs({
         // (插件声明优先于系统链;用户在下面的钉档永远最大)。
         const declaredModel = isText ? (prefs.text?.declaredModel ?? null) : null;
         const current = overrides[capability];
-        const selectValue = current && current !== defaultModel?.id ? current : FOLLOW_DEFAULT_VALUE;
+        const selectValue =
+          current && current !== defaultModel?.id ? current : FOLLOW_DEFAULT_VALUE;
         if (isText) {
           // 快问快答:目录全量文本模型的富列表选择器(图标/折扣与订阅徽标/分组/
-          // 搜索,对齐新建对话的模型选择器);image/video 仍是原生 select。
+          // 搜索,对齐新建对话的模型选择器)。
           const textOptions: readonly OneshotPinOption[] = prefs.text?.options ?? [];
           // 存量轻量档位钉(目录扩展前钉下的合法值)回显友好名,不当 stale 露 id。
           const legacyPinLabel = current
@@ -160,7 +314,59 @@ export function CindyCapabilityPrefs({
             </div>
           );
         }
-        // image/video 类目:原生 select(options 无分组信息,平铺)。
+
+        // image/video 保持原有轻量下拉，只用 Provider 图标区分同名模型的不同来源。
+        if (!capability.startsWith('embed.')) {
+          const mediaOptions = kind.options as unknown as readonly MediaModelOption[];
+          const mediaDefault = defaultModel as unknown as MediaModelOption;
+          return (
+            <div
+              key={capability}
+              className="cindy-capability-row flex min-w-0 items-center justify-between gap-4"
+            >
+              <span
+                className={cn(
+                  'min-w-0 text-[var(--text-secondary)]',
+                  appearance === 'plugin' ? 'text-13 leading-5' : 'text-12',
+                )}
+              >
+                {t(`settings.ghosts.detail.cindyPrefs.cap.${capability}`)}
+              </span>
+              {unavailable && current === undefined ? (
+                <span
+                  className={cn(
+                    'cindy-capability-empty min-w-0 truncate text-[var(--text-tertiary)]',
+                    appearance === 'plugin' ? 'text-13 leading-5' : 'text-12',
+                  )}
+                >
+                  {t('settings.ghosts.detail.cindyPrefs.noModels')}
+                </span>
+              ) : unavailable ? (
+                <button
+                  type="button"
+                  className={cn(
+                    'shrink-0 text-[var(--text-secondary)] hover:text-[var(--text-primary)]',
+                    appearance === 'plugin' ? 'text-13 leading-5' : 'text-12',
+                  )}
+                  onClick={() => void handleChange(capability, FOLLOW_DEFAULT_VALUE)}
+                >
+                  {t('settings.defaults.restore')}
+                </button>
+              ) : (
+                <MediaModelPicker
+                  value={selectValue === FOLLOW_DEFAULT_VALUE ? undefined : selectValue}
+                  defaultModel={mediaDefault}
+                  options={mediaOptions}
+                  onChange={(pin) => void handleChange(capability, pin ?? FOLLOW_DEFAULT_VALUE)}
+                  ariaLabel={t(`settings.ghosts.detail.cindyPrefs.cap.${capability}`)}
+                  dense={appearance !== 'plugin'}
+                />
+              )}
+            </div>
+          );
+        }
+
+        // embed 类目继续使用原生 select。
         const options: { id: string; label: string; group?: string }[] = kind.options;
         const selectable = options.filter((o) => o.id !== defaultModel?.id);
         const groupNames: string[] = [];
@@ -170,7 +376,9 @@ export function CindyCapabilityPrefs({
         }
         // 覆盖值已不在当前清单(目录演进/形态更替):如实显示原值,不假装跟随默认。
         const staleOverride =
-          current && selectValue !== FOLLOW_DEFAULT_VALUE && !selectable.some((o) => o.id === current)
+          current &&
+          selectValue !== FOLLOW_DEFAULT_VALUE &&
+          !selectable.some((o) => o.id === current)
             ? current
             : null;
         return (
@@ -206,7 +414,9 @@ export function CindyCapabilityPrefs({
                 )}
               >
                 <option value={FOLLOW_DEFAULT_VALUE}>
-                  {t('settings.ghosts.detail.cindyPrefs.defaultOption', { model: defaultModel.label })}
+                  {t('settings.ghosts.detail.cindyPrefs.defaultOption', {
+                    model: defaultModel.label,
+                  })}
                 </option>
                 {groupNames.map((groupName) =>
                   groupName === '' ? (
diff --git a/apps/desktop/src/renderer/features/plugin/__tests__/GhostPluginDetailSections.test.tsx b/apps/desktop/src/renderer/features/plugin/__tests__/GhostPluginDetailSections.test.tsx
index 90c1b7f517..3df4df5e92 100644
--- a/apps/desktop/src/renderer/features/plugin/__tests__/GhostPluginDetailSections.test.tsx
+++ b/apps/desktop/src/renderer/features/plugin/__tests__/GhostPluginDetailSections.test.tsx
@@ -52,6 +52,7 @@ vi.mock('react-i18next', () => ({
         'settings.ghosts.detail.collapseInfoValue': `Collapse ${String(options?.label ?? '')}`,
         'settings.ghosts.detail.panelNotDocked': 'Not docked',
         'settings.ghosts.detail.cindyPrefs.noModels': 'No models available',
+        'settings.defaults.restore': 'Restore default',
         'settings.ghosts.detail.oauthScopeStale':
           'This authorization does not include newly added permissions. Reconnect to enable them.',
       };
@@ -589,17 +590,47 @@ describe('Ghost plugin detail sections', () => {
             overrides: {},
             image: {
               options: [
-                { id: 'image-default', label: 'Image Default' },
-                { id: 'image-option', label: 'Image Option' },
+                {
+                  id: 'image-default',
+                  label: 'Image Default',
+                  providerId: 'xd',
+                  providerName: 'Cindy AI',
+                },
+                {
+                  id: 'image-option',
+                  label: 'Image Option',
+                  providerId: 'xd',
+                  providerName: 'Cindy AI',
+                },
               ],
-              defaultModel: { id: 'image-default', label: 'Image Default' },
+              defaultModel: {
+                id: 'image-default',
+                label: 'Image Default',
+                providerId: 'xd',
+                providerName: 'Cindy AI',
+              },
             },
             imageEdit: {
               options: [
-                { id: 'image-edit', label: 'Image Edit' },
-                { id: 'image-edit-option', label: 'Image Edit Option' },
+                {
+                  id: 'image-edit',
+                  label: 'Image Edit',
+                  providerId: 'xd',
+                  providerName: 'Cindy AI',
+                },
+                {
+                  id: 'image-edit-option',
+                  label: 'Image Edit Option',
+                  providerId: 'xd',
+                  providerName: 'Cindy AI',
+                },
               ],
-              defaultModel: { id: 'image-edit', label: 'Image Edit' },
+              defaultModel: {
+                id: 'image-edit',
+                label: 'Image Edit',
+                providerId: 'xd',
+                providerName: 'Cindy AI',
+              },
             },
             video: {
               options: [{ id: 'video-default', label: 'Video Default' }],
@@ -625,18 +656,16 @@ describe('Ghost plugin detail sections', () => {
 
     expect(container.querySelector('.cindy-capability-prefs')).toBeTruthy();
     expect(container.querySelector('.cindy-capability-row')).toBeTruthy();
-    const selects = screen.getAllByRole('combobox') as HTMLSelectElement[];
-    expect(selects).toHaveLength(2);
-    expect(Array.from(selects[0]!.options, (option) => option.value)).toEqual([
-      '__default__',
-      'image-option',
-    ]);
-    expect(Array.from(selects[1]!.options, (option) => option.value)).toEqual([
-      '__default__',
-      'image-edit-option',
-    ]);
-    expect(selects[0]!.className).toContain('cindy-capability-select');
-    expect(selects[0]!.className).toContain('max-w-[60%]');
+    const pickers = screen.getAllByRole('combobox');
+    expect(pickers).toHaveLength(2);
+    fireEvent.click(pickers[0]!);
+    expect(within(screen.getByRole('listbox')).getAllByRole('option')).toHaveLength(2);
+    expect(screen.getByText('Image Option')).toBeTruthy();
+    fireEvent.click(pickers[0]!);
+    fireEvent.click(pickers[1]!);
+    expect(within(screen.getByRole('listbox')).getAllByRole('option')).toHaveLength(2);
+    expect(screen.getByText('Image Edit Option')).toBeTruthy();
+    expect(pickers[0]!.className).toContain('max-w-[60%]');
   });
 
   // 2026-08-05:快问快答钉档扩展为目录全量文本模型——富列表选择器(供应商
@@ -851,7 +880,8 @@ describe('Ghost plugin detail sections', () => {
     vi.unstubAllEnvs();
   });
 
-  it('replaces the select with tertiary copy for ability categories the catalog has no models for', () => {
+  it('keeps a reset entry for a stale media override when the catalog has no models', async () => {
+    const setCindyPref = vi.fn(async () => ({ overrides: {} }));
     Object.defineProperty(window, 'electronAPI', {
       configurable: true,
       value: {
@@ -859,8 +889,20 @@ describe('Ghost plugin detail sections', () => {
           cindyPrefsSync: () => ({
             overrides: { 'video.generate': 'retired-video-model' },
             image: {
-              options: [{ id: 'image-default', label: 'Image Default' }],
-              defaultModel: { id: 'image-default', label: 'Image Default' },
+              options: [
+                {
+                  id: 'image-default',
+                  label: 'Image Default',
+                  providerId: 'xd',
+                  providerName: 'Cindy AI',
+                },
+              ],
+              defaultModel: {
+                id: 'image-default',
+                label: 'Image Default',
+                providerId: 'xd',
+                providerName: 'Cindy AI',
+              },
             },
             imageEdit: {
               options: [{ id: 'image-edit', label: 'Image Edit' }],
@@ -870,7 +912,7 @@ describe('Ghost plugin detail sections', () => {
             video: { options: [], defaultModel: null },
             videoEdit: { options: [], defaultModel: null },
           }),
-          setCindyPref: vi.fn(),
+          setCindyPref,
         },
       },
     });
@@ -888,10 +930,13 @@ describe('Ghost plugin detail sections', () => {
     expect(screen.getAllByRole('combobox')).toHaveLength(1);
 
     const empties = container.querySelectorAll('.cindy-capability-empty');
-    expect(empties).toHaveLength(2);
-    empties.forEach((node) => {
-      expect(node.textContent).toBe('No models available');
-      expect(node.className).toContain('text-[var(--text-tertiary)]');
+    expect(empties).toHaveLength(1);
+    expect(empties[0]!.textContent).toBe('No models available');
+    expect(empties[0]!.className).toContain('text-[var(--text-tertiary)]');
+
+    fireEvent.click(screen.getByRole('button', { name: 'Restore default' }));
+    await waitFor(() => {
+      expect(setCindyPref).toHaveBeenCalledWith('builtin.example', 'video.generate', null);
     });
   });
 
diff --git a/apps/desktop/src/renderer/vite-env.d.ts b/apps/desktop/src/renderer/vite-env.d.ts
index f5d114d709..0123e96ba4 100644
--- a/apps/desktop/src/renderer/vite-env.d.ts
+++ b/apps/desktop/src/renderer/vite-env.d.ts
@@ -1075,6 +1075,22 @@ interface GoalStatusPayload {
   lastReason: string | null;
 }
 
+type CindyMediaPreferenceOption = {
+  id: string;
+  label: string;
+  group: string;
+  providerId: string;
+  providerName: string;
+  modelId: string;
+  modelName: string;
+  routing?: import('@cindy/model-providers').Provider['routing'];
+};
+
+type CindyMediaPreferenceKind = {
+  options: CindyMediaPreferenceOption[];
+  defaultModel: CindyMediaPreferenceOption | null;
+};
+
 interface ElectronAPI {
   platform: string;
   osRelease: string;
@@ -1227,22 +1243,10 @@ interface ElectronAPI {
      */
     cindyPrefsSync: (id: string) => {
       overrides: Record<string, string>;
-      image: {
-        options: Array<{ id: string; label: string }>;
-        defaultModel: { id: string; label: string } | null;
-      };
-      imageEdit: {
-        options: Array<{ id: string; label: string }>;
-        defaultModel: { id: string; label: string } | null;
-      };
-      video: {
-        options: Array<{ id: string; label: string }>;
-        defaultModel: { id: string; label: string } | null;
-      };
-      videoEdit: {
-        options: Array<{ id: string; label: string }>;
-        defaultModel: { id: string; label: string } | null;
-      };
+      image: CindyMediaPreferenceKind;
+      imageEdit: CindyMediaPreferenceKind;
+      video: CindyMediaPreferenceKind;
+      videoEdit: CindyMediaPreferenceKind;
       /** 文本类(快问快答):选项是当前供应商目录的全部文本模型(cat: 编码钉值,
        *  带供应商/模型/徽标等结构化字段供富列表渲染);declaredModel = 身份卡声明
        *  的偏好模型(目录里解析得到才给,"跟随默认"行据此如实展示实际路由)。 */
diff --git a/apps/desktop/src/shared/ghost.ts b/apps/desktop/src/shared/ghost.ts
index 82aedc795c..c5ee96a896 100644
--- a/apps/desktop/src/shared/ghost.ts
+++ b/apps/desktop/src/shared/ghost.ts
@@ -6055,7 +6055,7 @@ export type GhostMediaCapability = (typeof GHOST_MEDIA_CAPABILITIES)[number];
 
 /** 插件读取自身某项 Cindy 媒体能力当前实际选型的只读结果。 */
 export type GhostCindyPreferenceResult =
-  | { ok: true; capability: GhostMediaCapability; modelId: string }
+  | { ok: true; capability: GhostMediaCapability; modelId: string; providerId: string }
   | {
       ok: false;
       errorCode: 'INVALID_REQUEST' | 'PERMISSION_DENIED' | 'NOT_AVAILABLE';
@@ -6074,10 +6074,13 @@ export type GhostMediaModelsResult =
       models: Array<{
         id: string;
         name: string;
+        /** 同名模型可由多个来源提供；与 id 一起构成精确选择。 */
+        providerId: string;
         /** Gateway architecture 的归一化投影；缺省表示上游未声明，插件不得猜测。 */
         modalities?: { input: string[]; output: string[] };
       }>;
       defaultModelId: string | null;
+      defaultProviderId: string | null;
     }
   | {
       ok: false;
diff --git a/packages/cindy-tools/src/__tests__/ghostMcp.test.ts b/packages/cindy-tools/src/__tests__/ghostMcp.test.ts
index de05d2bb1e..9de39df692 100644
--- a/packages/cindy-tools/src/__tests__/ghostMcp.test.ts
+++ b/packages/cindy-tools/src/__tests__/ghostMcp.test.ts
@@ -1114,12 +1114,14 @@ describe("cindy · media MCP 边界", () => {
     const result = await handleMedia(fakeDeps({ callMedia }), {
       action: "prepare",
       capability: "image.generate",
+      provider_id: "openai",
       model_id: "vendor/image-model",
     });
 
     expect(callMedia).toHaveBeenCalledWith({
       action: "prepare",
       capability: "image.generate",
+      providerId: "openai",
       modelId: "vendor/image-model",
     });
     expect(parsePayload(result)).toMatchObject({ ok: true, status: "prepared" });
diff --git a/packages/cindy-tools/src/ghost/mcpServer.ts b/packages/cindy-tools/src/ghost/mcpServer.ts
index f06c4e9754..2d7a38cc45 100644
--- a/packages/cindy-tools/src/ghost/mcpServer.ts
+++ b/packages/cindy-tools/src/ghost/mcpServer.ts
@@ -96,7 +96,7 @@ const D_MEDIA = [
   "如果插件需要消费最终结果，Agent 可在本工具成功后再调用插件声明的普通工具，并通过 ghost_call.attachments 显式交接；这不是生成请求的必经步骤。",
   "模型存在性来自当前账号的 Model Access model group；list_models 只投影同时满足 Gateway modalities、Guide operation 与当前客户端协议支持度的可执行模型。",
   "媒体生成必须由当前 Agent 通过本工具发起；插件面板和插件沙箱代码不得直接提交生成请求。",
-  "插件已返回用户配置的 model_id 时可直接用它和目标 capability 走 prepare → request；没有已配置 model_id 时，先用 list_models 查询可用媒体模型。prepare 会由 Server 根据 model_id 返回 Guide，并在 Guide 不存在或不支持该 capability 时明确报错。",
+  "插件已返回用户配置的 model_id/provider_id 时必须原样传给 prepare，再按目标 capability 走 prepare → request；provider_id 用于区分不同 Provider 下的同名模型。没有已配置模型时，先用 list_models 查询。Gateway 模型的 prepare 会由 Server 根据 model_id 返回 Guide，并在 Guide 不存在或不支持该 capability 时明确报错。",
   "异步任务的 request 返回 pending 时，再按 recommended_poll_after_ms 调 poll；同步任务会直接返回 xdt_image_urls / xdt_video_urls。",
   "模型 id、endpoint、Authorization 和 wire model 均由 Host 管理，不要写进 body，不要猜测或覆盖。",
   "Guide 缺失、能力不匹配或当前客户端不支持协议时，结果会带稳定 errorCode、retryable、outcomeKnown 和 allowedActions；可按 allowedActions 换模型、改用其它已授权工具或仍存在的旧链路，不要把 INTERNAL 当成协议能力结论。",
@@ -508,6 +508,7 @@ export async function handleMedia(
   input: {
     action: "list_models" | "prepare" | "request" | "poll";
     capability?: CindyMediaCapability;
+    provider_id?: string;
     model_id?: string;
     invocation_id?: string;
     body?: Record<string, unknown>;
@@ -549,6 +550,7 @@ export async function handleMedia(
       }
       result = await deps.callMedia({
         action: "prepare",
+        ...(input.provider_id ? { providerId: input.provider_id } : {}),
         modelId: input.model_id,
         capability: input.capability,
       });
@@ -1152,6 +1154,11 @@ export function createCindyGhostsMcpServer(
         .max(256)
         .optional()
         .describe("prepare 时必填；使用插件返回的已配置 model_id，或来自本次 list_models"),
+      provider_id: z
+        .string()
+        .max(128)
+        .optional()
+        .describe("prepare 时可选；插件或 list_models 返回 provider_id 时必须原样传入，以区分同名模型的执行来源"),
       invocation_id: z
         .string()
         .max(128)
diff --git a/packages/cindy-tools/src/types.ts b/packages/cindy-tools/src/types.ts
index e1c7086798..c206a53ea3 100644
--- a/packages/cindy-tools/src/types.ts
+++ b/packages/cindy-tools/src/types.ts
@@ -277,7 +277,13 @@ export type CindyMediaCapability =
 /** 当前 Agent 专用的永久 media 工具与 Host 之间的稳定请求面。插件运行时代码不调用。 */
 export type CindyMediaToolRequest =
   | { action: 'list_models'; capability?: CindyMediaCapability }
-  | { action: 'prepare'; modelId: string; capability: CindyMediaCapability }
+  | {
+      action: 'prepare';
+      /** 精确执行来源；插件配置返回 providerId 时必须原样传入。 */
+      providerId?: string;
+      modelId: string;
+      capability: CindyMediaCapability;
+    }
   | { action: 'request'; invocationId: string; body: Record<string, unknown> }
   | { action: 'poll'; invocationId: string };
 
diff --git a/packages/model-providers/src/builtin.ts b/packages/model-providers/src/builtin.ts
index 2d6e0fac5b..95fa6b25e8 100644
--- a/packages/model-providers/src/builtin.ts
+++ b/packages/model-providers/src/builtin.ts
@@ -145,7 +145,14 @@ const OPENAI_PROVIDER: Provider = {
   // image_generation tool;用户另配 `openai-images` Platform key 时优先走 public
   // Images API。id 带 openai/ 前缀(跨供应商数据契约,防 first-wins 归属漂移);
   // 不声明 imageDefaults(xd 默认地位不动)。
-  imageModels: [{ id: 'openai/gpt-image-2', name: 'GPT Image 2' }],
+  imageModels: [
+    {
+      id: 'openai/gpt-image-2',
+      name: 'GPT Image 2',
+      modalities: { input: ['text', 'image'], output: ['image'] },
+      officialDocs: 'https://platform.openai.com/docs/guides/image-generation',
+    },
+  ],
   routing: {
     codex: {
       upstream: 'https://chatgpt.com/backend-api/codex',
diff --git a/packages/model-providers/src/catalog.ts b/packages/model-providers/src/catalog.ts
index 234dbf9272..e42aabdb34 100644
--- a/packages/model-providers/src/catalog.ts
+++ b/packages/model-providers/src/catalog.ts
@@ -392,7 +392,12 @@ function validateProvider(p: Provider): void {
 function validateMediaModels(
   providerId: string,
   modelsField: string,
-  models: { id: string; name: string }[] | undefined,
+  models: Array<{
+    id: string;
+    name: string;
+    modalities?: { input: string[]; output: string[] };
+    officialDocs?: string;
+  }> | undefined,
   defaultsField: string,
   defaults: { standard: string; draft?: string; best?: string } | undefined,
 ): void {
@@ -405,6 +410,44 @@ function validateMediaModels(
       assert(typeof m.name === 'string' && m.name.length > 0, `provider '${providerId}' ${modelsField} '${m.id}' missing name`);
       assert(!seen.has(m.id), `provider '${providerId}' ${modelsField} has duplicate id '${m.id}'`);
       seen.add(m.id);
+      if (m.modalities !== undefined) {
+        assert(
+          m.modalities && typeof m.modalities === 'object' && !Array.isArray(m.modalities),
+          `provider '${providerId}' ${modelsField} '${m.id}' modalities must be an object`,
+        );
+        for (const key of ['input', 'output'] as const) {
+          const values = m.modalities[key];
+          assert(
+            Array.isArray(values) && values.length > 0 && values.length <= 16,
+            `provider '${providerId}' ${modelsField} '${m.id}' modalities.${key} must be a non-empty bounded array`,
+          );
+          assert(
+            values.every(
+              (value) =>
+                typeof value === 'string' &&
+                value.length > 0 &&
+                value.length <= 64 &&
+                value.trim() === value,
+            ) && new Set(values).size === values.length,
+            `provider '${providerId}' ${modelsField} '${m.id}' modalities.${key} contains invalid values`,
+          );
+        }
+      }
+      if (m.officialDocs !== undefined) {
+        let valid = false;
+        if (typeof m.officialDocs === 'string' && m.officialDocs.length <= 2_048) {
+          try {
+            const url = new URL(m.officialDocs);
+            valid = url.protocol === 'https:' && !url.username && !url.password;
+          } catch {
+            valid = false;
+          }
+        }
+        assert(
+          valid,
+          `provider '${providerId}' ${modelsField} '${m.id}' officialDocs must be https`,
+        );
+      }
     }
   }
   if (defaults !== undefined) {
diff --git a/packages/model-providers/src/index.ts b/packages/model-providers/src/index.ts
index c1c291b731..8898adf80c 100644
--- a/packages/model-providers/src/index.ts
+++ b/packages/model-providers/src/index.ts
@@ -19,6 +19,7 @@ export type {
   RoutingDescriptor,
   ModelCost,
   CatalogModel,
+  ProviderMediaModel,
   Provider,
   Catalog,
   CustomProviderConfig,
diff --git a/packages/model-providers/src/types.ts b/packages/model-providers/src/types.ts
index 0f7bf46f03..b980383502 100644
--- a/packages/model-providers/src/types.ts
+++ b/packages/model-providers/src/types.ts
@@ -421,6 +421,15 @@ export interface CatalogModel {
   disabled?: boolean;
 }
 
+/** Provider 自己执行的媒体模型；modalities 是能力判断的唯一依据。 */
+export interface ProviderMediaModel {
+  id: string;
+  name: string;
+  modalities?: { input: string[]; output: string[] };
+  officialDocs?: string;
+  disabled?: boolean;
+}
+
 /** 供应商定义。 */
 export interface Provider {
   /** 'anthropic' | 'openai' | 'xd' | 未来自定义 id。 */
@@ -461,7 +470,7 @@ export interface Provider {
    * `disabled` 是视图层字段(与 CatalogModel.disabled 同语义):buildRegistry 按用户
    * 停用 override 烘焙,设置页据此渲染专属媒体条目的停用状态;目录数据本身不携带。
    */
-  imageModels?: { id: string; name: string; disabled?: boolean }[];
+  imageModels?: ProviderMediaModel[];
   /**
    * 图像能力的默认选型(与 imageModels 配套;值必须是 imageModels 里的 id):
    * - standard:未指定任何偏好时的默认模型(意识 cindy 槽"默认"档的真身);
@@ -476,7 +485,7 @@ export interface Provider {
    * 消费方为意识 cindy 槽(白名单 + 详情页下拉)。可选,additions-only。
    * `disabled` 同 imageModels:视图层停用标志,buildRegistry 烘焙。
    */
-  videoModels?: { id: string; name: string; disabled?: boolean }[];
+  videoModels?: ProviderMediaModel[];
   /**
    * 视频能力的默认选型(与 videoModels 配套;值必须是 videoModels 里的 id;
    * 语义同 imageDefaults:standard 必填,draft/best 缺省回落 standard)。