{model.id}
{capabilityLabels[model.capabilities.kind]}{model.capabilities.kind !== 'unknown' && !model.capabilities.allowedStages.length ? ' · 待适配' : ''} · {provider.name}{!provider.enabled ? ' · 上游已停用' : ''}{!model.present ? ' · 已离开上游目录' : ''}
{model.capabilities.reason}{model.id}
{capabilityLabels[model.capabilities.kind]} · {provider.name}{!provider.enabled ? ' · 上游已停用' : ''}{!model.present ? ' · 已离开上游目录' : ''}
{model.capabilities.reason}{model.capabilities.kind === 'unknown' ? '能力待识别，暂不可测试' : '暂未接入详情页用途'}
} + {!testing && !attempt && !expandedDetails[detailsId] && ['failed', 'stale', 'unknown'].includes(result.status) && result.message &&{result.message}
} + :{blockedReason}
} })} diff --git a/docs/ai-image-reference-acceptance-2026-10-15.md b/docs/ai-image-reference-acceptance-2026-10-15.md new file mode 100644 index 0000000..debaa32 --- /dev/null +++ b/docs/ai-image-reference-acceptance-2026-10-15.md @@ -0,0 +1,53 @@ +# 生图模型参考图适配验收（2026-10-15） + +本次平台适配、测试门禁、权限与异常处理验收通过。真实模型结果为两项通过、三项上游失败，失败型号没有获得开放资格；不能将本报告理解为所有型号都已可用。 + +## 实现范围 + +- 按上游正式目录的输入能力和 JSON Schema 识别参考图数组。明确声明图片输入、图生图能力且允许 Data URI 的接口，传入私有参考图；不按模型名字猜能力，不公开产品文件，不切换到其他路由重试。 +- 支持 `gpt-image-2`、`gpt-image-2-vip`、`gpt-image-2.5`、`gpt-image-2.5-flare`、`gpt-image-2.5-sunburst` 的原生 `images/generations` 参考图请求。复用已有图片归一化、结果解码和参考图验证。 +- 仅文字输入的型号显示“仅支持文生图，不能接收产品参考图”。上游未提供参考图参数、上游停用、接口未适配分别给出原因。失败及未知结果直接显示安全诊断；展开详情不重复显示原因。 +- 明确的上游终止拒绝记为失败；无法确认的 502、已受理任务及待核查状态继续暂停，保留请求编号，不自动重新调用。核查失败后保留核查说明，核查本身不会使模型通过。 +- 目录同步增加可选 `modelKeys` 范围，服务端核对归属后只更新选中型号；现有全量同步仍可使用。目录同步不会自动启用、修改默认模型或插入测试通过记录。 + +## 真实模型验收 + +使用独立 QA 数据库、真实上游密钥和平台原生管理员测试接口。每型号只发起一次新请求，随机参考图颜色和位置由服务端校验；没有用模型自称支持或 HTTP 200 代替图片验证。 + +| 型号 | 真实结果 | 证据与处理 | +| --- | --- | --- | +| gpt-image-2 | 通过 | 请求 `b7ef37d7-34ca-41da-9dd0-ac2f02ee85ca`；1254×1254 图片可解码，两个色块颜色和相对位置匹配 | +| gpt-image-2-vip | 通过 | 请求 `11d3f051-fb61-48fd-b3b3-c1c672ec1774`；1254×1254 图片可解码，参考图验证通过 | +| gpt-image-2.5 | 未通过 | 请求 `e01ecf80-57b8-42a5-adfc-a8a88b6b66c7`；上游最终 `failed_final / 502 / service_rejected`，供应端 403 | +| gpt-image-2.5-flare | 未通过 | 请求 `34ee3b42-f86b-4714-aac6-e82a131d4d71`；同类上游终止拒绝，供应端 403 | +| gpt-image-2.5-sunburst | 未通过 | 请求 `46e4acad-d70a-498e-b29f-c5206ebbe92e`；同类上游终止拒绝，供应端 403 | + +三个失败请求按本次幂等编号只读核查上游最终记录，再通过平台已有核查接口关闭未知状态，均保持失败和禁止开放。没有修改上游服务配置。供应端 403 的进一步原因不在本次平台证据中，不能推断为参考图格式错误。 + +`nano-banana-2-2k-cl`、`nano-banana-2-4k-cl` 仅声明文字输入，测试请求在服务端调用前拒绝，不消耗模型调用。其余型号继续以各自的正式能力和实际测试结果为准；没有宣称已完成全部上游型号的真实出图验收。 + +目录依据：[上游正式模型目录](https://api.vip98k.com/api/v2/catalog/models)（schema 2.2）及其 `format=skill` 调用说明。 + +## 验证结果 + +- 独立发布预览仅包含本次文件，基于 `88ce9e3798fb18accf002acb92f2fbc1fc8b9687`；未包含详情页知识库等其他在途修改，也没有数据库迁移或新依赖。 +- 13 个相关测试套件、258 项测试通过；相关文件 ESLint 无错误；TypeScript 检查和本地生产构建通过。 +- 原生 API 验证通过：五个原幂等请求重放没有新增执行；两个纯文生图型号调用前拒绝；三个失败型号启用被拒绝；通过型号可保存开放后恢复原配置，其他配置及默认值未变。 +- 未登录返回 401；非超级管理员测试请求返回 403；以上负向操作没有新增测试记录或模型执行。 +- 已检查 1440×1000 桌面与 390×844 手机实际页面。桌面内容宽/滚动宽均为 1425，手机均为 375，无横向溢出；通过用途能勾选，失败用途禁用；确认弹窗取消未调用模型；失败详情显示一次。 + +## Impeccable audit（Operate 模式，限定本次模型面板） + +| 项目 | 分数 | 结果 | +| --- | --- | --- | +| 可访问性 | 3/4 | 状态可读，按钮和用途有名称、禁用原因；次要文字、错误、通过颜色对白底对比度分别为 4.69、6.57、6.09；未声称完成全站 WCAG 认证 | +| 性能 | 4/4 | 无新依赖、轮询或额外模型调用；沿用分页 | +| 响应式 | 4/4 | 桌面与手机实测，无横向溢出；测试和详情按钮高 44px | +| 主题一致性 | 3/4 | 保留现有管理面板颜色与布局，无无效 dark 变体；现有颜色常量未做全站迁移 | +| 实现完整性 | 4/4 | 具体原因来自服务端能力或真实记录；bundled detector 对目标组件无发现 | + +总分 18/20。本范围无 P0/P1；原有管理页其他区块未纳入本次修改。原始验证记录、测试日志和桌面/手机截图在本地 `output/ai-image-adapter-20261015`。 + +## 发布约束 + +发布只提交本报告与对应适配、状态、测试文件。生产替换前校验原提交和镜像、执行队列门禁，保留上一镜像。上线后只刷新这五个型号的能力；不复制 QA 测试记录，不自动启用新模型，不改变现有默认模型。生产真实复测若仍失败，必须保持失败或待核查并保留原请求，不能为发布而伪造通过。 diff --git a/lib/__tests__/ai-model-capabilities.test.ts b/lib/__tests__/ai-model-capabilities.test.ts index 23cf4bd..59e7846 100644 --- a/lib/__tests__/ai-model-capabilities.test.ts +++ b/lib/__tests__/ai-model-capabilities.test.ts @@ -21,6 +21,17 @@ test('image understanding and image output require separate interfaces', () => { expect(getAiModelCapabilities(generic, 'generator', metadata(['text'], ['image'], ['/images/generations']))).toMatchObject({ kind: 'image', allowedStages: [] }) expect(getAiModelCapabilities(generic, 'editor', metadata(['text', 'image'], ['image'], ['/images/edits']))).toMatchObject({ kind: 'image', allowedStages: ['image'] }) }) + +test('image models explain missing reference input and unavailable upstream separately', () => { + expect(getAiModelCapabilities(generic, 'text-image', metadata(['text'], ['image'], ['/images/generations']))).toMatchObject({ allowedStages: [], reason: '仅支持文生图，不能接收产品参考图' }) + expect(getAiModelCapabilities(generic, 'reference-image', metadata(['text', 'image'], ['image'], ['/images/generations']))).toMatchObject({ allowedStages: [], reason: '上游尚未提供已适配的参考图接口' }) + const undeclaredReference = metadata(['text'], ['image'], ['/images/generations']) + undeclaredReference.inputDeclared = false + undeclaredReference.protocols = [{ id: 'openai-image-generate', path: '/v1/images/generations', headers: {}, supportsImageInput: false }] + expect(getAiModelCapabilities(generic, 'undeclared-reference', undeclaredReference)).toMatchObject({ allowedStages: [], reason: '上游未提供可用的参考图参数，暂不能接收产品参考图' }) + const unavailable = { ...metadata(['text', 'image'], ['image'], ['/images/edits']), unavailable: true } + expect(getAiModelCapabilities(generic, 'unavailable', unavailable)).toMatchObject({ allowedStages: [], reason: '上游当前不可用，暂不能测试或开放' }) +}) test('audio transcription is not a copy-writing model just because it returns text over chat', () => { expect(getAiModelCapabilities(generic, 'asr', metadata(['audio'], ['text'], ['/chat/completions']))).toMatchObject({ kind: 'asr', allowedStages: [] }) }) diff --git a/lib/__tests__/ai-model-protocol.test.ts b/lib/__tests__/ai-model-protocol.test.ts index e4a2215..b1754d3 100644 --- a/lib/__tests__/ai-model-protocol.test.ts +++ b/lib/__tests__/ai-model-protocol.test.ts @@ -30,6 +30,23 @@ test('does not infer private reference uploads from generic image capability or expect(protocols).toMatchObject([{ id: 'openai-image-generate', supportsImageInput: false }]) }) +test('declared image URL arrays admit private data URIs without a redundant pattern', () => { + const entry = { type: 'image', modality: 'image', capabilities: ['image_to_image'], input_modes: ['text', 'image_url'], invocation: { method: 'POST', path: '/v1/images/generations', content_type: 'application/json', headers: {}, input_schema: { properties: { image: { type: 'array', items: { type: 'string', format: 'uri' }, maxItems: 3 } } } } } + expect(parseAiModelProtocols(entry)).toMatchObject([{ id: 'openai-image-generate', supportsImageInput: true, inputField: 'image', maxImages: 3 }]) + expect(parseAiModelProtocols({ ...entry, input_modes: ['text'] })).toMatchObject([{ supportsImageInput: false }]) + expect(parseAiModelProtocols({ ...entry, invocation: { ...entry.invocation, input_schema: { properties: { image: { type: 'array', items: { type: 'string', pattern: '^https://' } } } } } })).toMatchObject([{ supportsImageInput: false }]) +}) + +test('Images Generations sends declared image array with private bytes and retains its route', async () => { + send.mockResolvedValue(Response.json({ data: [{ b64_json: png.toString('base64') }] })) + const result = await callAiModelProtocol({ ...base, protocol: { id: 'openai-image-generate', path: '/v1/images/generations', headers: { 'X-Relay-Route-Option': 'image-route' }, routeId: 'image-route', supportsImageInput: true, inputField: 'image', maxImages: 2 }, images: [{ buffer: png, mimeType: 'image/png' }] }) + expect(send).toHaveBeenCalledTimes(1) + const body = JSON.parse(send.mock.calls[0][3]) + expect(body).toMatchObject({ image: [expect.stringMatching(/^data:image\/png;base64,/)], n: 1, response_format: 'b64_json' }) + expect(send.mock.calls[0][5]).toEqual({ 'X-Relay-Route-Option': 'image-route' }) + expect((await sharp(result.image).metadata()).width).toBe(32) +}) + test.each(['audio', 'video'])('does not turn %s models into chat or detail generation adapters', modality => { expect(parseAiModelProtocols({ type: modality, modality, invocation: { method: 'POST', path: '/v1/chat/completions', content_type: 'application/json', headers: {} } })).toEqual([]) }) @@ -72,6 +89,22 @@ test('rejected requests that echo credentials stay definite failures without exp try { await callAiModelProtocol({ ...base, protocol: chat }); throw new Error('expected rejection') } catch (error: any) { expect(error.outcomeUnknown).toBe(false); expect(error.message).not.toContain(base.apiKey) } }) +test('explicit final upstream rejection is failed, while an opaque 502 retains its request for investigation', async () => { + send.mockResolvedValueOnce(Response.json({ error: { code: 'service_rejected' } }, { status: 502, headers: { 'x-request-id': 'ended-request' } })) + await expect(callAiModelProtocol({ ...base, protocol: chat })).rejects.toMatchObject({ outcomeUnknown: false, requestId: 'ended-request', message: '上游拒绝模型请求（HTTP 502），测试未通过' }) + send.mockResolvedValueOnce(new Response('Bad Gateway', { status: 502, headers: { 'x-request-id': 'opaque-request' } })) + await expect(callAiModelProtocol({ ...base, protocol: chat })).rejects.toMatchObject({ outcomeUnknown: true, requestId: 'opaque-request', message: '上游返回异常响应（502），请核查请求结果；勿重复提交' }) + expect(send).toHaveBeenCalledTimes(2) +}) + +test('review status and an accepted task override an otherwise definite rejection code', async () => { + send.mockResolvedValueOnce(Response.json({ status: 'pending_review', error: { code: 'service_rejected' } }, { status: 502 })) + await expect(callAiModelProtocol({ ...base, protocol: chat })).rejects.toMatchObject({ outcomeUnknown: true }) + send.mockResolvedValueOnce(Response.json({ error: { code: 'service_rejected' } }, { status: 502 })) + await expect(callAiModelProtocol({ ...base, protocol: asyncImage, resumeTaskId: 'accepted-task', onTaskSubmitted: jest.fn() })).rejects.toMatchObject({ outcomeUnknown: true, taskId: 'accepted-task' }) + expect(send).toHaveBeenCalledTimes(2) +}) + test('persists async id before any polling, sends private inline image, decodes typed output', async () => { const events: string[] = [] send.mockImplementation(async (_base, _key, path) => { diff --git a/lib/__tests__/ai-model-recovery.test.ts b/lib/__tests__/ai-model-recovery.test.ts index ca9a488..490050f 100644 --- a/lib/__tests__/ai-model-recovery.test.ts +++ b/lib/__tests__/ai-model-recovery.test.ts @@ -123,6 +123,19 @@ test('old passed rules require retest; task recovery preserves context and canno expect(aiModelVerificationSummary([row], modelKey, 'analysis', 'changed')).toMatchObject({ status: 'unknown', canResume: false, canReconcile: true }) expect(aiModelVerificationSummary([row], modelKey, 'analysis', row.contextHash, false)).toMatchObject({ status: 'unknown', canResume: false, canReconcile: true }) }) + +test('unknown outcome retains its safe cause; an expired running record uses the interruption warning', () => { + row.message = '上游返回异常响应（502），请核查请求结果；勿重复提交' + expect(aiModelVerificationSummary([row], modelKey, 'analysis', row.contextHash)).toMatchObject({ status: 'unknown', message: row.message, canReconcile: true }) + row.status = 'running'; row.message = '正在执行' + expect(aiModelVerificationSummary([row], modelKey, 'analysis', row.contextHash)).toMatchObject({ status: 'unknown', message: '请求结果尚未确认，已暂停重复测试；请核查上游请求记录' }) +}) + +test('a reconciled failed upstream request exposes the checked cause without passing or enabling it', () => { + row.status = 'failed'; row.evidence = { reconciliation: proof } + expect(aiModelVerificationSummary([row], modelKey, 'analysis', row.contextHash)).toMatchObject({ status: 'failed', reconciled: true, canResume: false, canReconcile: false, message: `已核查上游执行失败：${proof.note}` }) + expect(aiModelVerificationSummary([row], modelKey, 'analysis', 'changed')).toMatchObject({ status: 'stale', message: '连接、路由或测试规则已改变，请重新测试' }) +}) test('randomized reference image validation rejects swapped layout and wrong image interpretation', async () => { const challenge = createVerificationChallenge(), correct = await verificationReferenceImage(challenge) expect(challenge.topLeft).not.toBe(challenge.bottomRight) diff --git a/lib/__tests__/ai-model-service.test.ts b/lib/__tests__/ai-model-service.test.ts index eea27c5..4fdad3a 100644 --- a/lib/__tests__/ai-model-service.test.ts +++ b/lib/__tests__/ai-model-service.test.ts @@ -132,6 +132,25 @@ test('directory synchronization never creates verification evidence or enables n expect(statements.some(sql => sql.includes('SET present=FALSE'))).toBe(true) expect(aiModelAudit).toHaveBeenCalledWith(expect.anything(), 'admin', 'provider.sync', legacyId, expect.objectContaining({ addedCount: 1 })) }) + +test('scoped capability synchronization never rewrites unrelated models or creates verification evidence', async () => { + await syncAiModelProvider('admin', legacyId, { revision: 1, modelKeys: [legacyModelKey] }) + const updates = mockDb.$executeRawUnsafe.mock.calls + expect(updates.find(([sql]) => sql.includes('present=FALSE'))).toEqual([expect.stringContaining('AND id=ANY($2::text[])'), legacyId, [legacyModelKey]]) + const inserted = updates.filter(([sql]) => sql.startsWith('INSERT INTO ai_model_catalog')) + expect(inserted).toHaveLength(1) + expect(inserted[0][3]).toBe('vendor/shared-model') + expect(JSON.stringify(updates)).not.toContain('brand/new-model') + expect(updates.some(([sql]) => sql.includes('INSERT INTO ai_model_verifications'))).toBe(false) + expect(aiModelAudit).toHaveBeenCalledWith(expect.anything(), 'admin', 'provider.sync', legacyId, expect.objectContaining({ modelKeys: [legacyModelKey], addedCount: 0 })) +}) + +test('a scoped update cannot use another provider model or an empty selection', async () => { + await expect(syncAiModelProvider('admin', legacyId, { revision: 1, modelKeys: [secondModelKey] })).rejects.toMatchObject({ code: 'catalog_selection_invalid' }) + await expect(syncAiModelProvider('admin', legacyId, { revision: 1, modelKeys: [] })).rejects.toMatchObject({ code: 'invalid_request' }) + expect(requestAiModel).not.toHaveBeenCalled() + expect(mockDb.$executeRawUnsafe).not.toHaveBeenCalled() +}) test('concurrent provider updates during catalog response reject all catalog writes', async () => { ;(requestAiModel as jest.Mock).mockImplementation(async (_base, _key, _endpoint, _body, _idempotency, beforeRequest) => { await beforeRequest(); providers[0].revision++; return Response.json({ data: [{ id: 'new' }] }) }) await expect(syncAiModelProvider('admin', legacyId, { revision: 1 })).rejects.toMatchObject({ code: 'revision_conflict' }) diff --git a/lib/__tests__/ai-model-verification.test.ts b/lib/__tests__/ai-model-verification.test.ts index fc0ac45..b958a39 100644 --- a/lib/__tests__/ai-model-verification.test.ts +++ b/lib/__tests__/ai-model-verification.test.ts @@ -81,6 +81,16 @@ test.each([false, true])('protocol outcomeUnknown=%s is persisted accurately, ne expect(terminal[2]).toBe(unknown ? 'unknown' : 'failed') expect(JSON.stringify(terminal)).not.toContain('synthetic-test-secret') }) + +test.each([false, true])('safe protocol diagnostics are preserved for outcomeUnknown=%s', async unknown => { + const message = unknown ? '上游返回异常响应（502），请核查请求结果；勿重复提交' : '上游拒绝模型请求（HTTP 403），测试未通过' + ;(callAiModelProtocol as jest.Mock).mockImplementation(async request => { await request.beforeRequest(); throw new AiModelProtocolError(message, unknown, 'safe-request-id') }) + await verifyAiModel('admin', providerId, modelKey, input()) + const terminal = mockDb.$executeRawUnsafe.mock.calls.find(([sql]) => sql.includes('status=$2'))! + expect(terminal).toContain(message) + expect(terminal[2]).toBe(unknown ? 'unknown' : 'failed') + expect(callAiModelProtocol).toHaveBeenCalledTimes(1) +}) test('a persisted task can only be queried through resumeTaskId and its original idempotency key', async () => { row.status = 'unknown'; row.taskId = 'task-original' await refreshAiModelVerification('admin', providerId, row.id) diff --git a/lib/ai-model-capabilities.ts b/lib/ai-model-capabilities.ts index 7e87e63..2925225 100644 --- a/lib/ai-model-capabilities.ts +++ b/lib/ai-model-capabilities.ts @@ -134,10 +134,17 @@ export function getAiModelCapabilities(baseUrl: string, modelId: string, metadat // the analysis protocol; its explicit input flag may differ on legacy catalogs. const transport = (protocol: AiModelProtocol | null) => protocol && JSON.stringify({ ...protocol, supportsImageInput: undefined }) const combinedTextTest = allowedStages.includes('analysis') && allowedStages.includes('copy') && !!analysisProtocol?.supportsImageInput && transport(analysisProtocol) === transport(copyProtocol) + const reason = evidence.unavailable ? '上游当前不可用，暂不能测试或开放' + : allowedStages.length ? '接口已适配；通过对应用途的实际测试后可开放' + : kind === 'image' ? evidence.inputDeclared && !imageInput && !imageEdit ? '仅支持文生图，不能接收产品参考图' + : protocols.some(protocol => protocol.id === 'openai-image-generate') ? '上游未提供可用的参考图参数，暂不能接收产品参考图' + : '上游尚未提供已适配的参考图接口' + : ['text', 'vision'].includes(kind) ? '输入输出或接口不符合详情页分析、文案要求' + : '该类型不适用于详情页的分析、文案或生图' return { kind, allowedStages, combinedTextTest, source: 'upstream', protocols: Object.fromEntries(allowedStages.map(stage => { const protocol = getAiModelStageProtocol(metadata, stage)! return [stage, { id: protocol.id, path: protocol.path, scopeId: protocol.scopeId, routeId: protocol.routeId, operationId: protocol.operationId }] - })), reason: evidence.unavailable ? '上游已停用该模型' : allowedStages.length ? '接口已适配；通过对应用途的实际测试后可开放' : kind === 'image' ? '当前路由尚不支持私有产品参考图' : ['text', 'vision'].includes(kind) ? '该接口或明确输入能力尚不匹配详情页' : '该类型不适用于详情页的三个用途' } + })), reason } } export function getAiModelStageProtocol(metadata: unknown, stage: AiModelStage): AiModelProtocol | null { diff --git a/lib/ai-model-protocol.ts b/lib/ai-model-protocol.ts index b1361a4..a2c9df4 100644 --- a/lib/ai-model-protocol.ts +++ b/lib/ai-model-protocol.ts @@ -56,11 +56,17 @@ function parseInvocation(model: Record