feat(media): 支持按 Provider 选择媒体模型

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'],

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。 */
@@ -147,29 +153,31 @@ 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)。
    *  aspectRatio 语义同 generateImage:不传 = 跟随源图画幅(后端 auto)。 */
   editImage(params: {
     prompt: string;
     model: string;
+    providerId?: string;
     imagePaths: string[];
     aspectRatio?: GhostImageAspectRatio;
   }): Promise<{ buffer: Uint8Array; mimeType: string }>;
   /**
    * 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
    * 生效的画面参数回执(上游上报值优先,缺项回落提交值)。长任务:
    * 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 =
@@ -830,20 +847,29 @@ 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:${GHOST_MODEL_TIERS.join(' / ')})` };
       }
       model = defaults[p.tier as GhostModelTier];
     }
     const capability = `${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' } : {}),
       });
@@ -997,19 +1028,33 @@ export class GhostCindySlot {
           generated = await this.deps.editImage({
             prompt,
             model,
+            ...(providerId ? { providerId } : {}),
             imagePaths,
             ...(aspectRatio !== undefined ? { aspectRatio } : {}),
           });
         } else if (kind === 'gen_image') {
           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 =

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 }),
   };
 }

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：