Skip to content

armgit / cindy

Public Imported from GitHub
0

Consider it done. The open-source AI agent that works out of the box · 想到,就能做到。开源、开箱即用的 AI Agent。

fix(media): 修复旧插件选型和视频提交边界

nana5d40f17fcd3f
diff --git a/apps/desktop/src/main/cindy-brain/__tests__/customImageChannel.test.ts b/apps/desktop/src/main/cindy-brain/__tests__/customImageChannel.test.ts
index ed7f1effc7..9c4755460e 100644
--- a/apps/desktop/src/main/cindy-brain/__tests__/customImageChannel.test.ts
+++ b/apps/desktop/src/main/cindy-brain/__tests__/customImageChannel.test.ts
@@ -6,6 +6,7 @@ import { buildUserProvider, type Provider } from '@cindy/model-providers';
import { createCustomImageChannel, registerCustomImageChannels } from '../customImageChannel.js';
import { pruneDynamicImageChannels } from '../byokImageChannel.js';
import { ImageChannelRegistry } from '../imageChannelRegistry.js';
+import { filterLegacyCindyMediaConfig } from '../cindyMediaCatalog.js';
const MODEL = 'team/gpt-image-2.5-sunburst';
const PNG = Buffer.from('89504e470d0a1a0a', 'hex');
@@ -83,6 +84,21 @@ function harness(initial = provider()) {
}
describe('standalone custom OpenAI Images', () => {
+ it('keeps URL-capable custom sources in Core but out of legacy image preferences', () => {
+ const h = harness();
+ const registry = new ImageChannelRegistry();
+ registry.register('relay', h.channel);
+ registry.register('existing', { ready: () => true, generateImage: vi.fn(), editImage: vi.fn() });
+ const custom = { id: 'media:relay:image', modelId: MODEL, providerId: 'relay', label: 'Custom', supportsEdit: true };
+ const existing = { ...custom, id: 'media:existing:image', providerId: 'existing', label: 'Existing' };
+ const config = { models: [custom, existing], defaults: { standard: custom.id, draft: custom.id, best: custom.id } };
+ const legacy = filterLegacyCindyMediaConfig(config, model => registry.isProviderLegacyReady(model.providerId, model.modelId));
+ expect(legacy).toEqual({ models: [existing], defaults: { standard: existing.id, draft: existing.id, best: existing.id } });
+ expect(registry.isProviderReady('relay', MODEL)).toBe(true);
+ expect(config.models).toEqual([custom, existing]);
+ expect(h.fetchImplementation).not.toHaveBeenCalled();
+ });
+
it.each(['claude-code', 'codex'] as const)('binds duplicate model IDs to the explicitly enabled Images connection, not %s', async (agent) => {
const source = buildUserProvider({
id: 'relay', name: 'Relay', auth: { method: 'apiKey' },
diff --git a/apps/desktop/src/main/cindy-brain/customImageChannel.ts b/apps/desktop/src/main/cindy-brain/customImageChannel.ts
index 470997fe1c..2cfd767d7f 100644
--- a/apps/desktop/src/main/cindy-brain/customImageChannel.ts
+++ b/apps/desktop/src/main/cindy-brain/customImageChannel.ts
@@ -86,6 +86,7 @@ export function createCustomImageChannel(opts: CustomImageChannelOptions): Image
return {
imageProtocol: 'openai',
+ legacyCompatible: false,
ready,
isModelReady: (model) => connection(model) !== null,
generateImage: (params) => invoke('generate', { ...params, imagePaths: [] }),
diff --git a/apps/desktop/src/main/cindy-brain/imageChannelRegistry.ts b/apps/desktop/src/main/cindy-brain/imageChannelRegistry.ts
index c94627c765..90a804729c 100644
--- a/apps/desktop/src/main/cindy-brain/imageChannelRegistry.ts
+++ b/apps/desktop/src/main/cindy-brain/imageChannelRegistry.ts
@@ -65,6 +65,8 @@ export function decodeImageResponse(res: ImageChannelResult): { buffer: Buffer;
*/
export interface ImageChannel {
imageProtocol?: ImageProtocol;
+ /** False for Core-only channels whose URL results the legacy decoder cannot consume. */
+ legacyCompatible?: boolean;
/** 执行凭证是否就绪。false ⇒ 该来源整段不进 cindy 白名单。 */
ready(): boolean;
/** A multi-runtime connection can be executable for only some of its image models. */
@@ -119,6 +121,12 @@ export class ImageChannelRegistry {
(modelId === undefined || channel.isModelReady?.(modelId) !== false);
}
+ /** The frozen synchronous plugin API only accepts channels with legacy results. */
+ isProviderLegacyReady(providerId: string, modelId?: string): boolean {
+ return this.channels.get(providerId)?.legacyCompatible !== false &&
+ this.isProviderReady(providerId, modelId);
+ }
+
/** 编辑操作清单派生用:就绪且 supportsEdit !== false。 */
isProviderEditReady(providerId: string, modelId?: string): boolean {
const ch = this.channels.get(providerId);
diff --git a/apps/desktop/src/main/cindy-brain/index.ts b/apps/desktop/src/main/cindy-brain/index.ts
index 362fc9be0a..ecc693f72d 100644
--- a/apps/desktop/src/main/cindy-brain/index.ts
+++ b/apps/desktop/src/main/cindy-brain/index.ts
@@ -3621,8 +3621,7 @@ function getCatalogMediaConfig(
// managed URL/base64 handling; keep new sources out of the frozen legacy API.
kind === 'image'
? (providerId) =>
- !registeredCustomImageAccounts.has(providerId) &&
- (imageRegistry?.isProviderReady(providerId) ?? false)
+ imageRegistry?.isProviderLegacyReady(providerId) ?? false
: kind === 'video'
? isVideoCatalogProviderReady
: isXdGatewayProviderReady,
@@ -4672,12 +4671,12 @@ export function getGhostCindySlot(): GhostCindySlot {
const videoRegistry = capability.startsWith('video.') ? getVideoProviderRegistry() : null;
// Core 专用选型不能交给旧执行器。复用存量配置迁移,确保回退后
// 用户保存的 provider + modelId 与实际执行一致,不影响 Core 清单。
- const executableConfig = capability.startsWith('video.')
- ? filterLegacyCindyMediaConfig(
- config,
- (model) => videoRegistry?.hasAlias(model.modelId, model.providerId) ?? false,
- )
- : config;
+ const executableConfig = filterLegacyCindyMediaConfig(
+ config,
+ (model) => capability.startsWith('video.')
+ ? videoRegistry?.hasAlias(model.modelId, model.providerId) ?? false
+ : getImageChannelRegistry().isProviderLegacyReady(model.providerId, model.modelId),
+ );
const selected = resolveAndMigrateGhostMediaPreference(
ghostId,
mediaCapability,
diff --git a/apps/desktop/src/main/cindy-media/__tests__/customVideoProtocol.test.ts b/apps/desktop/src/main/cindy-media/__tests__/customVideoProtocol.test.ts
index 81b0e0fa19..0d5186ae1b 100644
--- a/apps/desktop/src/main/cindy-media/__tests__/customVideoProtocol.test.ts
+++ b/apps/desktop/src/main/cindy-media/__tests__/customVideoProtocol.test.ts
@@ -51,7 +51,7 @@ describe('custom xAI Videos protocol', () => {
]) expect(videoConnectionRevision('relay', model.id, changed)).not.toBe(revision);
});
- it.each([{ fps: 30 }, { audio: false }, { duration: 16 }, { duration: 1.5 }, { resolution: '4k' }, { image: { url: 'https://example.com/p.png' } }])('rejects unsupported generation parameters before payment: %j', extra => {
+ it.each([{ fps: 30 }, { audio: false }, { duration: 16 }, { duration: 1.5 }, { resolution: '4k' }, { resolution: ['480p'] }, { aspect_ratio: ['16:9'] }, { image: { url: 'https://example.com/p.png' } }])('rejects unsupported generation parameters before payment: %j', extra => {
expect(() => validateCustomVideoBody({ prompt: 'cat', ...extra }, false)).toThrow();
});
it('requires exactly one image and leaves implicit aspect ratio to the upstream', () => {
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 0d5ae01e11..f566cc7ae4 100644
--- a/apps/desktop/src/main/cindy-media/__tests__/invocationService.test.ts
+++ b/apps/desktop/src/main/cindy-media/__tests__/invocationService.test.ts
@@ -45,6 +45,7 @@ const mocks = vi.hoisted(() => ({
dbOwnerId: 'media-user-0',
transitionDbs: [] as unknown[],
failTransitionTo: null as string | null,
+ onClaim: vi.fn(),
recover: vi.fn<(owner: string, db: unknown) => Promise<number>>(async () => 0),
prune: vi.fn(async () => undefined),
rows: new Map<string, Record<string, unknown>>(),
@@ -184,6 +185,7 @@ vi.mock('../mediaInvocationStore.js', () => ({
const row = mocks.rows.get(id);
if (!row || row.owner !== owner || row.state !== from) return false;
row.state = to;
+ if (to === 'submitting') mocks.onClaim();
row.updatedAt = Date.now();
if (taskId) row.taskId = taskId;
if (responseJson) row.responseJson = responseJson;
@@ -280,6 +282,7 @@ describe('Cindy Core media invocation state and security boundary', () => {
mocks.providerModel.mockReset();
mocks.providerInvoke.mockReset();
mocks.videoConnection.mockReset();
+ mocks.onClaim.mockReset();
mocks.guide.mockReset();
mocks.outboundFetch.mockReset();
mocks.release.mockClear();
@@ -324,6 +327,49 @@ describe('Cindy Core media invocation state and security boundary', () => {
return connection;
}
+ it.each(['reference-read', 'claim'] as const)('rechecks custom video availability after %s before paying, and preserves prepared', async boundary => {
+ setupCustomVideo();
+ const p = await callCindyMedia({ action: 'prepare', providerId: 'relay', modelId: 'private-video', capability: 'video.image_to_video' });
+ const invocationId = p.invocation_id as string;
+ const stopSource = () => { mocks.providerModel.mockReturnValue(null); };
+ mocks.readBlob.mockImplementation(async () => {
+ if (boundary === 'reference-read') stopSource();
+ return { buffer: PNG, mimeType: 'image/png' };
+ });
+ if (boundary === 'claim') mocks.onClaim.mockImplementationOnce(stopSource);
+ expect(await callCindyMedia({ action: 'request', invocationId, body: {
+ prompt: 'cat waves', image: { url: `cindy-media://blobs/${'a'.repeat(64)}.png` },
+ } })).toMatchObject({ errorCode: 'MODEL_NOT_AVAILABLE', outcomeKnown: true });
+ expect(mocks.outboundFetch).not.toHaveBeenCalled();
+ expect(mocks.rows.get(invocationId)?.state).toBe('prepared');
+ });
+
+ it('can retrieve an already paid custom video after its model is disabled', async () => {
+ setupCustomVideo();
+ const p = await callCindyMedia({ action: 'prepare', providerId: 'relay', modelId: 'private-video', capability: 'video.generate' });
+ const invocationId = p.invocation_id as string;
+ mocks.outboundFetch.mockResolvedValueOnce(Response.json({ request_id: 'paid-task' }))
+ .mockResolvedValueOnce(Response.json({ status: 'done', video: { url: '/videos/paid-task/content' } }))
+ .mockResolvedValueOnce(new Response(MP4, { headers: { 'content-type': 'video/mp4' } }));
+ mocks.ingestMedia.mockResolvedValue({ url: `cindy-media://blobs/${'a'.repeat(64)}.mp4` });
+ await callCindyMedia({ action: 'request', invocationId, body: { prompt: 'cat' } });
+ mocks.providerModel.mockReturnValue(null);
+ expect(await callCindyMedia({ action: 'poll', invocationId })).toMatchObject({ status: 'complete' });
+ expect(mocks.outboundFetch.mock.calls.filter(([, init]) => init.method === 'POST')).toHaveLength(1);
+ });
+
+ it.each(['video/mp4', 'audio/mpeg', 'image/gif'])('rejects %s as an image-to-video reference before claiming or paying', async mimeType => {
+ setupCustomVideo();
+ mocks.readBlob.mockResolvedValue({ buffer: PNG, mimeType });
+ const p = await callCindyMedia({ action: 'prepare', providerId: 'relay', modelId: 'private-video', capability: 'video.image_to_video' });
+ const invocationId = p.invocation_id as string;
+ expect(await callCindyMedia({ action: 'request', invocationId, body: {
+ prompt: 'cat waves', image: { url: `cindy-media://blobs/${'a'.repeat(64)}.png` },
+ } })).toMatchObject({ errorCode: mimeType === 'image/gif' ? 'MEDIA_INPUT_INVALID' : 'REQUEST_INVALID' });
+ expect(mocks.outboundFetch).not.toHaveBeenCalled();
+ expect(mocks.rows.get(invocationId)?.state).toBe('prepared');
+ });
+
it.each([['video.generate', false], ['video.image_to_video', false], ['video.generate', true], ['video.image_to_video', true]] as const)('custom %s persists one paid task and ingests video (private content=%s)', async (capability, privateContent) => {
setupCustomVideo();
mocks.readBlob.mockResolvedValue({ buffer: PNG, mimeType: 'image/png' });
diff --git a/apps/desktop/src/main/cindy-media/customVideoProtocol.ts b/apps/desktop/src/main/cindy-media/customVideoProtocol.ts
index e35c66e0bb..93087f829d 100644
--- a/apps/desktop/src/main/cindy-media/customVideoProtocol.ts
+++ b/apps/desktop/src/main/cindy-media/customVideoProtocol.ts
@@ -65,8 +65,8 @@ export function validateCustomVideoBody(body: Record<string, unknown>, image: bo
if (Object.keys(body).some(key => !allowed.has(key))) throw new Error('视频请求包含当前协议不支持的字段');
if (typeof body.prompt !== 'string' || !body.prompt.trim()) throw new Error('视频请求必须提供 prompt');
if (body.duration !== undefined && (!Number.isInteger(body.duration) || Number(body.duration) < 1 || Number(body.duration) > 15)) throw new Error('视频时长必须为 1–15 秒的整数');
- if (body.resolution !== undefined && !['480p', '720p', '1080p'].includes(String(body.resolution))) throw new Error('视频分辨率不受支持');
- if (body.aspect_ratio !== undefined && !['16:9', '9:16', '1:1', '4:3', '3:4', '3:2', '2:3'].includes(String(body.aspect_ratio))) throw new Error('视频画幅不受支持');
+ if (body.resolution !== undefined && (typeof body.resolution !== 'string' || !['480p', '720p', '1080p'].includes(body.resolution))) throw new Error('视频分辨率不受支持');
+ if (body.aspect_ratio !== undefined && (typeof body.aspect_ratio !== 'string' || !['16:9', '9:16', '1:1', '4:3', '3:4', '3:2', '2:3'].includes(body.aspect_ratio))) throw new Error('视频画幅不受支持');
if (image) {
const ref = body.image as Record<string, unknown> | undefined;
if (!ref || Array.isArray(ref) || Object.keys(ref).length !== 1 || typeof ref.url !== 'string' || !ref.url) throw new Error('图生视频必须提供一张图片: image: {url}');
diff --git a/apps/desktop/src/main/cindy-media/invocationService.ts b/apps/desktop/src/main/cindy-media/invocationService.ts
index 8d14699780..27eb43565f 100644
--- a/apps/desktop/src/main/cindy-media/invocationService.ts
+++ b/apps/desktop/src/main/cindy-media/invocationService.ts
@@ -76,6 +76,7 @@ interface MediaConnection {
pathPrefix?: string;
headers?: Record<string, string>;
assertCurrent?(): void;
+ beforeSubmit?(): void;
}
class MediaInvocationError extends Error {
@@ -275,7 +276,16 @@ function resolveInvocationConnection(invocation: StoredMediaInvocation): MediaCo
catch { throw new MediaInvocationError('CONNECTION_CHANGED', '视频连接或账号已改变,本次操作已停止'); }
};
assertCurrent();
- return { ...connection, apiKey: '', pathPrefix: new URL(connection.baseUrl).pathname.replace(/\/+$/, ''), assertCurrent };
+ return {
+ ...connection, apiKey: '', pathPrefix: new URL(connection.baseUrl).pathname.replace(/\/+$/, ''), assertCurrent,
+ beforeSubmit: () => {
+ // Re-read executable membership after reference reads and the database claim.
+ // Paid task polling/recovery remains independent of the model's enable switch.
+ if (!resolveProviderMediaModel(invocation.guide.connection.providerId, invocation.modelId, invocation.capability)) {
+ throw new MediaInvocationError('MODEL_NOT_AVAILABLE', '该模型或来源已停用或不再支持本次能力,本次生成未发出');
+ }
+ },
+ };
}
function resolveConnection(providerId: string): MediaConnection {
@@ -491,6 +501,7 @@ async function dispatchRequest(input: {
params: mediaRequestParamsForLog(input.body ?? {}),
});
input.connection.assertCurrent?.();
+ if (input.operation === 'submit') input.connection.beforeSubmit?.();
dispatched = true;
const response = await outboundFetch(url, {
method: input.method,
@@ -1436,6 +1447,12 @@ async function submitInvocation(
}
const requestBody = await prepareRequestBody(validatedBody, invocation.guide);
assertAuthScope(scope, invocation.owner);
+ if (isClientVideoInvocation(invocation)) {
+ // Managed refs have now become data URLs with their actual stored media MIME.
+ const { model: _injectedModel, ...videoBody } = requestBody;
+ try { validateCustomVideoBody(videoBody, invocation.capability === 'video.image_to_video'); }
+ catch (error) { throw new MediaInvocationError('REQUEST_INVALID', (error as Error).message); }
+ }
let connection: MediaConnection;
try {
connection = resolveInvocationConnection(invocation);
@@ -1588,9 +1605,9 @@ async function submitInvocation(
};
} catch (error) {
const expected = error instanceof MediaInvocationError ? error : null;
- // 本地组装失败(#5081):请求从未出站、无费用,把 invocation 放回 prepared,
+ // 本地组装失败或提交前模型已停用:请求从未出站、无费用,把 invocation 放回 prepared,
// 让同一 invocation_id 可以直接再次 request;其余失败仍进入 failed / unknown。
- const buildFailed = expected?.code === 'REQUEST_BUILD_FAILED';
+ const buildFailed = expected?.code === 'REQUEST_BUILD_FAILED' || expected?.code === 'MODEL_NOT_AVAILABLE';
const restored = await transitionMediaInvocation(
{
id: invocation.id,
diff --git a/docs/custom-image-api.md b/docs/custom-image-api.md
index bd2b5391a3..dc768b2577 100644
--- a/docs/custom-image-api.md
+++ b/docs/custom-image-api.md
@@ -52,6 +52,7 @@ CPA/CLIProxyAPI、Sub2API 等自建中转只要提供上述契约即可复用此
下载失败后使用原 invocation 再次执行 request,只恢复同一结果,不重新付费生成。
- 上游提交结果不确定时不自动重提,不自动换模型或换来源;拒绝与诊断文本脱敏后返回。
- 新来源使用 Core 媒体能力;冻结的旧 `cindy-request` 同步接口不扩展自定义来源。
+ 旧入口的模型清单和偏好均按通道的旧协议兼容性过滤,不能把仅支持 Core 的来源交给 base64 解码器。
旧插件已有能力保持原状。要使用新来源,插件需要采用 Core 媒体能力。
## 后续边界
diff --git a/docs/custom-video-api.md b/docs/custom-video-api.md
index d71613d69c..4b5f32f9da 100644
--- a/docs/custom-video-api.md
+++ b/docs/custom-video-api.md
@@ -52,6 +52,9 @@ OpenAI 官方 Sora Videos API 已于 2026-09-24 停止服务,本次不新增
地址、Key、来源引擎或业务请求头变化时停止旧任务请求,恢复原连接后可查询原任务。
- 一次付费提交保存任务 ID,之后只查询它。完成响应先持久化,下载或入库恢复
不重新生成;签名地址过期时只刷新原任务。提交结果未知时禁止自动重提。
+- 新付费请求在参考图读取和调用占用完成后,实际发出 POST 前再次检查模型、来源及能力。
+ 若已停用则不提交,保留原 prepared 调用;已付费任务的查询与结果恢复不受展示/停用开关影响。
+ 视频参数严格检查类型;受管参考图展开后再次校验图片 MIME,视频或音频不能作为图片提交。
- 新来源使用 Core 媒体能力;冻结的旧 `cindy-request` 视频入口不扩展。
- 出现在目录或通过接口探测不等于真实出片。验收需分别从 Art 执行两种操作,
保存受管视频、检查实际时长/画面、确认任务退出运行状态,并区分准备与生成耗时。