From 903686db362afe47c87876b3b806214114f8b25e Mon Sep 17 00:00:00 2001 From: WuYongJie <549727955@qq.com> Date: Mon, 16 Nov 2026 14:43:24 +0800 Subject: [PATCH] fix: align enabled model parameters with provider contracts --- .../api/chat/_lib/detail-replacement.ts | 6 +- .../agent-workflow/material-search/route.ts | 2 +- app/api/canvas/detail-plan/route.ts | 6 +- app/api/canvas/detail-review/route.ts | 4 +- .../product-planning/analyze-image/route.ts | 2 +- .../canvas/LiblibVideoCanvasReplica.tsx | 5 +- components/image-params-bar.tsx | 20 +-- docs/模型参数核对-2026-11-16.md | 118 ++++++++++++++++++ lib/agents/scene-workflow-prompt-optimizer.ts | 2 +- lib/agents/workflow-material-appearance.ts | 2 +- lib/ai/auto-agent-graph.ts | 12 +- lib/ai/detail-agent-kimi-profile.ts | 4 +- lib/ai/prompt-optimizer.ts | 2 +- lib/ai/providers/vip98k-chat.ts | 14 ++- lib/ai/providers/vip98k.ts | 52 ++++++-- .../runtime/direct-agent-analysis-answer.ts | 2 +- lib/ai/smart-analyzer.ts | 6 +- lib/ai/vip98k-chat-contract.ts | 9 ++ lib/ai/vip98k-models.ts | 31 +++-- lib/ai/vip98k-task-invocation.ts | 53 +++++--- .../storyboard-command-interpreter.ts | 2 +- lib/video-canvas/model-matching.ts | 10 +- .../provider-call-compatibility.test.ts | 14 +++ tests/routes/vip98k-model-expansion.test.ts | 87 +++++++++++++ tests/routes/vip98k-reference-routing.test.ts | 10 +- 25 files changed, 394 insertions(+), 81 deletions(-) create mode 100644 docs/模型参数核对-2026-11-16.md diff --git a/app/(chat)/api/chat/_lib/detail-replacement.ts b/app/(chat)/api/chat/_lib/detail-replacement.ts index 1abf5011e..164964b31 100644 --- a/app/(chat)/api/chat/_lib/detail-replacement.ts +++ b/app/(chat)/api/chat/_lib/detail-replacement.ts @@ -623,7 +623,7 @@ async function resolveReplacementProductIdentityProfile(params: { system: "你是高精度商品结构分析器。多张图片属于同一商品;请跨视角还原唯一结构身份,并明确每个视角最清楚的构件。", temperature: params.planningRuntime.kind === "local_moonshot" ? 1 : 0.1, - maxTokens: 1800, + maxOutputTokens: 1800, abortSignal: AbortSignal.timeout( params.planningRuntime.kind === "local_moonshot" ? 90_000 : 60_000, ), @@ -2376,7 +2376,7 @@ export async function analyzeReplacementReferenceImage(params: { }, ], temperature: 0, - maxTokens: 160, + maxOutputTokens: 160, abortSignal: AbortSignal.timeout(45_000), } as any); @@ -2679,7 +2679,7 @@ export async function analyzeReplacementExecutionPlan(params: { system: "你是高精度商品详情图场景规划器。你只分析当前参考图本身,负责理解这一屏的场景、机位、主体位置和人物需求,不要臆测或依赖其他输入图片。", temperature: params.planningRuntime.kind === "local_moonshot" ? 1 : 0.1, - maxTokens: 1600, + maxOutputTokens: 1600, abortSignal: AbortSignal.timeout(replacementScenePlanningTimeoutMs), } as any); diff --git a/app/api/agent-workflow/material-search/route.ts b/app/api/agent-workflow/material-search/route.ts index 7404a38ed..3faa03b7e 100644 --- a/app/api/agent-workflow/material-search/route.ts +++ b/app/api/agent-workflow/material-search/route.ts @@ -450,7 +450,7 @@ async function analyzeFunctionalShowcaseImage(params: { }, ], temperature: 0, - maxTokens: 120, + maxOutputTokens: 120, abortSignal: AbortSignal.timeout(25_000), } as any); diff --git a/app/api/canvas/detail-plan/route.ts b/app/api/canvas/detail-plan/route.ts index 1faf23c9a..4ba0981b8 100644 --- a/app/api/canvas/detail-plan/route.ts +++ b/app/api/canvas/detail-plan/route.ts @@ -1804,7 +1804,7 @@ async function runDualDirectorPlanPreflight(params: { ? "You are both an art director and a marketing director for high-conversion e-commerce detail pages. Output JSON only." : "你同时扮演高转化电商详情页的平面设计总监与营销总监。仅输出JSON。", temperature: params.isLocalMoonshot ? 1 : 0.1, - maxTokens: 2000, + maxOutputTokens: 2000, ...(params.kimiOptions || {}), }, params.timeoutMs || DETAIL_PLAN_PREFLIGHT_TIMEOUT_MS); @@ -2040,7 +2040,7 @@ export async function POST(request: Request) { kimiDetailProfile: plannerKimiOptions.headers ? { thinking: plannerKimiOptions.headers["x-detail-agent-kimi-thinking"], - maxTokens: plannerKimiOptions.maxTokens, + maxOutputTokens: plannerKimiOptions.maxOutputTokens, } : null, }); @@ -2177,7 +2177,7 @@ export async function POST(request: Request) { `{"summary":"简述","globalStyle":{"designLanguage":"一句话风格方向","typography":{"fontFamily":"字体家族","title":"主标题规范","subtitle":"副标题规范","body":"正文规范"},"layout":{"grid":"栅格","safeArea":"安全区","rhythm":"留白节奏"},"palette":"配色规范"},"screens":[{"index":1,"screenType":"B","objective":"本屏目标","title":"首屏示例:甩甩杆,自拍不手抖(6-20字)","subtitle":"一句副标题","copy":"正文文案","layoutNote":"排版说明","useModel":false,"modelReason":"是否需要人物及原因","bundleShowcase":false,"bundleReason":"是否为搭配展示页及原因","fabe":{"b":"利益","f":"特征","a":"优势","e":"证据"},"visualPrompt":"用于生图的完整提示词,必须包含790x1300与统一字体排版"}],"dslV2Screens":[{"version":"detail-ui-dsl/v2","screenIndex":1,"layoutVariant":"hero-left-panel","region":{"x":0.06,"y":0.05,"width":0.88,"height":0.46},"nodes":[{"kind":"panel","id":"hero-panel","variant":"glass","tone":"light","direction":"column","gap":18,"children":[{"kind":"badge","id":"screen-badge","icon":"sparkle","text":"核心升级","tone":"primary"},{"kind":"heading","id":"hero-title","text":"护腰椅,久坐不腰酸","level":1,"emphasis":"hero"},{"kind":"body","id":"hero-subtitle","text":"在家里加班,也能护腰/舒适","emphasis":"strong"}]}]}]}`, temperature: 0.25, ...(localDetailAgentResolved ? { temperature: 1 } : {}), - maxTokens: 1800, + maxOutputTokens: 1800, ...plannerKimiOptions, }); diff --git a/app/api/canvas/detail-review/route.ts b/app/api/canvas/detail-review/route.ts index 45ddde0af..2d3d4038a 100644 --- a/app/api/canvas/detail-review/route.ts +++ b/app/api/canvas/detail-review/route.ts @@ -323,7 +323,7 @@ export async function POST(request: Request) { kimiDetailProfile: reviewKimiOptions.headers ? { thinking: reviewKimiOptions.headers["x-detail-agent-kimi-thinking"], - maxTokens: reviewKimiOptions.maxTokens, + maxOutputTokens: reviewKimiOptions.maxOutputTokens, } : null, }); @@ -438,7 +438,7 @@ export async function POST(request: Request) { ? "You are both an Art Director and a Marketing Director for high-conversion e-commerce detail pages. Evaluate with strict professional standards and output JSON only." : "你同时是高转化电商详情页的平面设计总监与营销总监。按严格专业标准评估,仅输出JSON。", temperature: localDetailAgentResolved ? 1 : 0.1, - maxTokens: 800, + maxOutputTokens: 800, ...reviewKimiOptions, } as any); text = String(result?.text || ""); diff --git a/app/api/product-planning/analyze-image/route.ts b/app/api/product-planning/analyze-image/route.ts index 984241be4..6d194e3cb 100644 --- a/app/api/product-planning/analyze-image/route.ts +++ b/app/api/product-planning/analyze-image/route.ts @@ -114,7 +114,7 @@ export async function POST(request: NextRequest) { }, ], temperature: 0.2, - maxTokens: 1200, + maxOutputTokens: 1200, } as any); return NextResponse.json({ diff --git a/components/canvas/LiblibVideoCanvasReplica.tsx b/components/canvas/LiblibVideoCanvasReplica.tsx index bb26b031a..884058ec9 100644 --- a/components/canvas/LiblibVideoCanvasReplica.tsx +++ b/components/canvas/LiblibVideoCanvasReplica.tsx @@ -15626,7 +15626,8 @@ function VideoTargetModuleComposer({ const ratio = supported.ratios.length && !supported.ratios.includes(normalizeVideoRatioValue(current.ratio)) ? supported.ratios[0] : current.ratio; const executableResolutions = supported.resolutions.filter(value => ["480p", "720p", "1080p"].includes(value.toLowerCase())); const resolution = executableResolutions.length && !executableResolutions.includes(normalizeVideoResolutionValue(current.resolution)) ? executableResolutions[0].toUpperCase() : current.resolution; - const duration = supported.durations.length && !supported.durations.includes(normalizeVideoDurationValue(current.duration)) ? `${supported.durations[0]}s` : current.duration; + const durations = getCanvasVideoParameters(target, activeMode, resolution || "720P").durations; + const duration = durations.length && !durations.includes(normalizeVideoDurationValue(current.duration)) ? `${durations[0]}s` : current.duration; const audio = supported.supportsAudio ? current.audio : "关闭"; return ratio === current.ratio && resolution === current.resolution && duration === current.duration && audio === current.audio ? current : {...current, ratio, resolution, duration, audio}; }); @@ -17929,7 +17930,7 @@ function VideoParameterEditor({ const candidates = filterCanvasModels(modelOptions, audioRequired ? "audio-video" : "video", mode); const model = candidates.find(item => item.id === settings.model || item.modelId === settings.model) || (!normalizeModelRef(settings.model) || settings.model === "auto" ? candidates[0] : undefined); - const options = getCanvasVideoParameters(model, mode); + const options = getCanvasVideoParameters(model, mode, settings.resolution || "720P"); const executableResolutions = options.resolutions.filter(value => ["480p", "720p", "1080p"].includes(value.toLowerCase())).map(value => value.toUpperCase()); const durationIndex = Math.max(0, options.durations.indexOf(normalizeVideoDurationValue(settings.duration))); return ( diff --git a/components/image-params-bar.tsx b/components/image-params-bar.tsx index 95d1b09ac..b1584fdfd 100644 --- a/components/image-params-bar.tsx +++ b/components/image-params-bar.tsx @@ -5,7 +5,7 @@ import { useEffect, useState } from "react"; import { Button } from "@/components/ui/button"; import { ChevronDown, ChevronUp, Link2, Scan, SlidersHorizontal, Sparkles } from "lucide-react"; import { getGrsaiGptImageOutputSize } from "@/lib/ai/grsai-gpt-image-sizes"; -import { resolveVip98kImageTaskSize } from "@/lib/ai/vip98k-task-invocation"; +import { getVip98kQualityOptions, resolveVip98kImageTaskSize, type Vip98kTaskInvocation } from "@/lib/ai/vip98k-task-invocation"; import { cn } from "@/lib/utils"; import { Input } from "@/components/ui/input"; import { Switch } from "@/components/ui/switch"; @@ -268,6 +268,17 @@ export function ImageParamsBar({ availableAspectRatioOptions[0]?.value || "1:1"; + const declaredQualityOptions = providerType === "vip98k" + ? getVip98kQualityOptions((imageGenConfig?.vip98k?.invocation as Vip98kTaskInvocation | undefined)?.inputSchema) + : undefined; + const qualityOptions = ([ + { label: "自动", value: "auto" }, + { label: "高", value: "high" }, + { label: "中", value: "medium" }, + { label: "低", value: "low" }, + ] as const).filter(option => declaredQualityOptions === undefined || + (declaredQualityOptions.length ? declaredQualityOptions.includes(option.value) : option.value === "auto")); + const clampParamsByProvider = (raw: ImageGenerationParams): ImageGenerationParams => { const options = getImageCountOptions(); const fallbackCount = options[0] ?? 1; @@ -286,6 +297,7 @@ export function ImageParamsBar({ aspectRatio, imageSize, imageCount, + quality: qualityOptions.some(option => option.value === raw.quality) ? raw.quality : qualityOptions[0]?.value || "auto", }; }; @@ -322,12 +334,6 @@ export function ImageParamsBar({ ...buildBaseParams(initialParams), }); const [minimalParamsOpen, setMinimalParamsOpen] = useState(false); - const qualityOptions = [ - { label: "自动", value: "auto" }, - { label: "高", value: "high" }, - { label: "中", value: "medium" }, - { label: "低", value: "low" }, - ] as const; const qualityLabel = qualityOptions.find((option) => option.value === params.quality)?.label || "中"; diff --git a/docs/模型参数核对-2026-11-16.md b/docs/模型参数核对-2026-11-16.md new file mode 100644 index 000000000..e5002a8a8 --- /dev/null +++ b/docs/模型参数核对-2026-11-16.md @@ -0,0 +1,118 @@ +# 生效模型参数核对(2026-11-16) + +核对对象为模型和所属上游同时启用的 76 条配置,包括隐藏但仍可被功能调用的模型。依据生产配置快照、实际请求构建代码、98K 当前账户公布的接口契约以及阿里云、AtlasCloud、Anthropic 官方资料。 + +## 修复范围 + +| 问题 | 影响与处理 | +| --- | --- | +| 图片尺寸与比例 | 删除仅允许正方形的错误约束;有独立比例与清晰度字段时按上游契约组合,固定 2K/4K 模型使用对应档位。 | +| 请求路线 | 显式选定的标准接口优先于通用任务示例,保留路线与参数约束的对应关系。 | +| 参考图数量 | 读取嵌套约束并采用当前实际调用接口的上限,GPT Image 2 的上限由 16 校正为 14。 | +| 图片质量与负面提示词 | 质量和负面提示词只写入一个有效位置;固定质量模型不再发送旧的其他档位,页面只提供接口接受的质量选项。 | +| 可选生成控制 | 修复水印、提示词扩展、随机种子、尾帧和固定镜头的已支持字段漏传,保留 false 和 0 的原值;上游未声明的字段不发送。 | +| 视频时长 | Veo 使用当前接口的 duration 字段;Minimax 的 1080p 时长选项限制在接口允许范围内。 | +| 文字输出长度 | AI SDK 调用统一使用 maxOutputTokens,原生厂商接口保留各自标准字段。 | +| Claude 采样参数 | 对固定采样协议移除 temperature、top_p、top_k,保留工具调用和输出长度。Haiku 5.5 已实测。 | +| 世界生成分类 | world-generation 的类型及任务调用配置按现有世界生成接口校正,避免作为聊天模型使用。 | + +## 验证 + +- 图片、视频、音频及工具参数组成 12,135 个离线请求用例,采用实际请求构建函数,并核对当前接口约束与字段值。 +- 阿里云图片另外覆盖 80 个尺寸、数量、参考图、随机种子和开关组合;7 个 AtlasCloud 三维模型按各自 OpenAPI 请求规范核对。 +- 19 个聊天/文字模型捕获官方 SDK 生成的请求,覆盖输出长度、采样、停止词以及适用的图片和工具调用。13 个 98K 模型同时通过当前路线契约校验;其余 6 个根据接入代码和 SDK 请求核对。 +- 自动回归 118 项通过,TypeScript 检查通过。 +- 独立实测:Nano Banana 2 带参考图生成 16:9 图片,返回 1376×768 PNG;固定 4K 编辑返回 4096×4096 JPEG;Claude Haiku 5.5 返回有效工具调用。两张图片均完成七牛保存和下载验证。每项只有一次提交。 + +## 生效模型清单 + +| 模型 | 上游 | 本次核对 | +| --- | --- | --- | +| `grok-4.7` | 98K | SDK 请求、输出长度、采样、停止词、视觉及工具调用 | +| `gpt-image-2` | 98K | 126 组请求字段与接口约束 | +| `gpt-image-2-vip` | 98K | 360 组请求字段与接口约束 | +| `nano-banana-2` | 98K | 90 组请求字段与接口约束 | +| `nano-banana-pro` | 98K | 90 组请求字段与接口约束 | +| `gpt-image-2.5` | 98K | 30 组请求字段与接口约束 | +| `gpt-image-2.5-flare` | 98K | 288 组请求字段与接口约束 | +| `ocr` | 98K | 3 组请求字段与接口约束 | +| `gpt-image-2.5-sunburst` | 98K | 504 组请求字段与接口约束 | +| `nano-banana-2-2k-cl` | 98K | 12 组请求字段与接口约束 | +| `nano-banana-2-4k-cl` | 98K | 12 组请求字段与接口约束 | +| `nano-banana-2-cl` | 98K | 12 组请求字段与接口约束 | +| `nano-banana-pro-4k-vip` | 98K | 12 组请求字段与接口约束 | +| `nano-banana-pro-cl` | 98K | 12 组请求字段与接口约束 | +| `nano-banana-pro-vip` | 98K | 12 组请求字段与接口约束 | +| `doubao-seedream-5-0-pro` | 98K | 240 组请求字段与接口约束 | +| `flux-2-pro` | 98K | 48 组请求字段与接口约束 | +| `gpt-image-2.6` | 98K | 84 组请求字段与接口约束 | +| `gpt-image-2.6-flare` | 98K | 360 组请求字段与接口约束 | +| `gpt-image-2.6-sunburst` | 98K | 36 组请求字段与接口约束 | +| `nano-banana-2-plus` | 98K | 360 组请求字段与接口约束 | +| `nano-banana-pro-plus` | 98K | 360 组请求字段与接口约束 | +| `qwen-image-3.0` | 98K | 48 组请求字段与接口约束 | +| `qwen-image-3.0-pro` | 98K | 48 组请求字段与接口约束 | +| `minimax-h3` | 98K | 320 组请求字段与接口约束 | +| `veo-3.1-fast` | 98K | 16 组请求字段与接口约束 | +| `seedance-2.5` | 98K | 1296 组请求字段与接口约束 | +| `seedance-2.5-image-to-video` | 98K | 1458 组请求字段与接口约束 | +| `seedance-2.5-reference-to-video` | 98K | 1458 组请求字段与接口约束 | +| `seedance-2.5-text-to-video` | 98K | 1134 组请求字段与接口约束 | +| `gemini-omni-1.1-flash-preview` | 98K | 8 组请求字段与接口约束 | +| `grok-imagine-video` | 98K | 2500 组请求字段与接口约束 | +| `grok-imagine-video-1.5` | 98K | 750 组请求字段与接口约束 | +| `gemini-3.8-flash` | 98K | SDK 请求、输出长度、采样、停止词、视觉及工具调用 | +| `gpt-6-astra` | 98K | SDK 请求、输出长度、采样、停止词 | +| `gpt-6-sol` | 98K | SDK 请求、输出长度、采样、停止词 | +| `gpt-6.1-sol` | 98K | SDK 请求、输出长度、采样、停止词 | +| `world-generation` | 98K | 世界生成类型、任务提交与结果查询配置 | +| `claude-haiku-5-5` | 98K | SDK 请求、输出长度、采样、停止词、视觉及工具调用 | +| `claude-sonnet-4-6` | 98K | SDK 请求、输出长度、采样、停止词、视觉及工具调用 | +| `image-expansion` | 98K | 3 组请求字段与接口约束 | +| `gem-3.1-tts` | 98K | 4 组请求字段与接口约束 | +| `mimo-v2.5-asr` | 98K | 1 组请求字段与接口约束 | +| `mimo-v2.5-tts` | 98K | 1 组请求字段与接口约束 | +| `mimo-v2.5-tts-voiceclone` | 98K | 1 组请求字段与接口约束 | +| `mimo-v2.5-tts-voicedesign` | 98K | 1 组请求字段与接口约束 | +| `music-2.5` | 98K | 3 组请求字段与接口约束 | +| `music-2.5+` | 98K | 3 组请求字段与接口约束 | +| `speech-2.8-hd` | 98K | 6 组请求字段与接口约束 | +| `speech-2.8-turbo` | 98K | 6 组请求字段与接口约束 | +| `suno-v4.5` | 98K | 3 组请求字段与接口约束 | +| `suno-v6` | 98K | 3 组请求字段与接口约束 | +| `suno-v5` | 98K | 3 组请求字段与接口约束 | +| `suno-v5-5` | 98K | 3 组请求字段与接口约束 | +| `claude-fable-5-1` | 98K | SDK 请求、输出长度、采样、停止词、视觉及工具调用 | +| `claude-opus-5-5` | 98K | SDK 请求、输出长度、采样、停止词、视觉及工具调用 | +| `claude-sonnet-5-5` | 98K | SDK 请求、输出长度、采样、停止词、视觉及工具调用 | +| `gemini-4-argon` | 98K | SDK 请求、输出长度、采样、停止词、视觉及工具调用 | +| `grok-4.5` | 98K | SDK 请求、输出长度、采样、停止词、视觉及工具调用 | +| `grok-4.6` | 98K | SDK 请求、输出长度、采样、停止词、视觉及工具调用 | +| `tripo-h3.1/image-to-3d` | AtlasCloud | 提示词/参考图、三维输出参数及官方 OpenAPI | +| `tripo-h3.1/text-to-3d` | AtlasCloud | 提示词/参考图、三维输出参数及官方 OpenAPI | +| `tencent/hunyuan3d-pro/text-to-3d` | AtlasCloud | 提示词/参考图、三维输出参数及官方 OpenAPI | +| `tencent/hunyuan3d-pro/image-to-3d` | AtlasCloud | 提示词/参考图、三维输出参数及官方 OpenAPI | +| `tencent/hunyuan3d-rapid/text-to-3d` | AtlasCloud | 提示词/参考图、三维输出参数及官方 OpenAPI | +| `tencent/hunyuan3d-rapid/image-to-3d` | AtlasCloud | 提示词/参考图、三维输出参数及官方 OpenAPI | +| `bytedance/seed3d-v2.0/image-to-3d` | AtlasCloud | 提示词/参考图、三维输出参数及官方 OpenAPI | +| `qwen-image-3.0` | 阿里云图片 | 80 组图片参数及原生请求字段 | +| `qwen3-vl-plus` | 阿里云 | SDK 请求、输出长度、采样、停止词、视觉及工具调用 | +| `qwen3.8-max` | 阿里云 | SDK 请求、输出长度、采样、停止词、视觉及工具调用 | +| `qwen3.8-omni-flash` | 阿里云 | SDK 请求、输出长度、采样、停止词、视觉及工具调用 | +| `qwen3-tts-instruct-flash` | 阿里云 | 官方语音请求、音色、语言与指令 | +| `qwen3-tts-flash` | 阿里云 | 官方语音请求、音色、语言与指令 | +| `deepseek-flash` | 自定义标准接口 | SDK 请求、输出长度、采样、停止词、视觉及工具调用 | +| `agnes-1.5-flash` | 自定义标准接口 | SDK 请求、输出长度、采样、停止词、视觉及工具调用 | +| `agnes-2.0-flash` | 自定义标准接口 | SDK 请求、输出长度、采样、停止词、视觉及工具调用 | + +## 已知问题与验证边界 + +- 场景融合的首选模型配置指向一个视频模型,当前实现会排除该项并使用图片候选模型,因此实际没有将图片请求提交给视频接口;该首选配置尚未调整。 +- 离线参数核对不等于 76 个模型逐个完成付费生成,也不证明上游持续可用、效果或速度;实际生成验证范围为上述三项。 +- 自动选模时页面尚未确定最终模型,接入层仍须依据实际选中接口规范化固定参数。模型清单和接口约束会随上游调整,本报告对应本次核对时点。 + +## 资料 + +- [Anthropic Haiku 5.5 采样参数迁移说明](https://platform.claude.com/docs/en/models/haiku-5-5/migration-guide#remove-sampling-parameters) +- [阿里云图片生成及编辑接口](https://www.alibabacloud.com/help/en/model-studio/qwen-image-generation-and-editing-api-reference) +- [阿里云语音合成接口](https://www.alibabacloud.com/help/en/model-studio/qwen-tts-api) diff --git a/lib/agents/scene-workflow-prompt-optimizer.ts b/lib/agents/scene-workflow-prompt-optimizer.ts index 4c76efa6f..0c06572cc 100644 --- a/lib/agents/scene-workflow-prompt-optimizer.ts +++ b/lib/agents/scene-workflow-prompt-optimizer.ts @@ -185,7 +185,7 @@ async function optimizeWithConfiguredQwenVisionModel(params: { }, ], temperature: 0, - maxTokens: 320, + maxOutputTokens: 320, abortSignal: AbortSignal.timeout(30_000), } as any); diff --git a/lib/agents/workflow-material-appearance.ts b/lib/agents/workflow-material-appearance.ts index 294a9b4a0..3f6aea42d 100644 --- a/lib/agents/workflow-material-appearance.ts +++ b/lib/agents/workflow-material-appearance.ts @@ -124,7 +124,7 @@ async function analyzeWithConfiguredQwenVisionModel(params: { }, ], temperature: 0, - maxTokens: 220, + maxOutputTokens: 220, abortSignal: AbortSignal.timeout(25_000), } as any); diff --git a/lib/ai/auto-agent-graph.ts b/lib/ai/auto-agent-graph.ts index 87e2cbaad..4931b5257 100644 --- a/lib/ai/auto-agent-graph.ts +++ b/lib/ai/auto-agent-graph.ts @@ -1737,7 +1737,7 @@ async function buildExecutionPlanWithAgent( .filter(Boolean) .join("\n\n"), temperature: getCompatibleTextGenerationTemperature(plannerModelConfig, 0.1), - maxTokens: 420, + maxOutputTokens: 420, abortSignal: AbortSignal.timeout(AGENT_OPERATION_TIMEOUT_MS), } as any); @@ -1858,7 +1858,7 @@ async function generateAgentConversation( }, ], temperature: getCompatibleTextGenerationTemperature(visionModelConfig, 0.2), - maxTokens: 360, + maxOutputTokens: 360, abortSignal: AbortSignal.timeout(AGENT_OPERATION_TIMEOUT_MS), } as any); @@ -1920,7 +1920,7 @@ async function generateAgentConversation( system: systemPrompt, prompt, temperature: getCompatibleTextGenerationTemperature(reasoningModelConfig, 0.35), - maxTokens: 520, + maxOutputTokens: 520, abortSignal: AbortSignal.timeout(AGENT_OPERATION_TIMEOUT_MS), } as any); @@ -2002,7 +2002,7 @@ async function recommendImageParamsWithAgent( - negativePrompt 仅在明显需要避坑时填写。`, prompt: requestText, temperature: getCompatibleTextGenerationTemperature(plannerModelConfig, 0.15), - maxTokens: 220, + maxOutputTokens: 220, abortSignal: AbortSignal.timeout(AGENT_OPERATION_TIMEOUT_MS), } as any); @@ -2095,7 +2095,7 @@ async function recommendVideoPlanWithAgent( - 避免抽象空话,必须可执行。`, prompt, temperature: getCompatibleTextGenerationTemperature(plannerModelConfig, 0.2), - maxTokens: 720, + maxOutputTokens: 720, abortSignal: AbortSignal.timeout(AGENT_OPERATION_TIMEOUT_MS), } as any); @@ -2175,7 +2175,7 @@ async function refineRouteWithAgent( 候选模型: ${candidatesText}`, temperature: getCompatibleTextGenerationTemperature(plannerModelConfig, 0.1), - maxTokens: 120, + maxOutputTokens: 120, abortSignal: AbortSignal.timeout(AGENT_OPERATION_TIMEOUT_MS), } as any); diff --git a/lib/ai/detail-agent-kimi-profile.ts b/lib/ai/detail-agent-kimi-profile.ts index 902c054dd..243692254 100644 --- a/lib/ai/detail-agent-kimi-profile.ts +++ b/lib/ai/detail-agent-kimi-profile.ts @@ -7,7 +7,7 @@ export type DetailKimiPurpose = "plan" | "preflight" | "review" | "repair"; export type DetailKimiGenerateOptions = { temperature?: number; topP?: number; - maxTokens?: number; + maxOutputTokens?: number; headers?: Record; }; @@ -80,7 +80,7 @@ export function buildDetailKimiGenerateOptions(params: { return { temperature: thinking === "enabled" ? 1 : 0.6, topP: 0.95, - maxTokens: getPurposeMaxTokens(params), + maxOutputTokens: getPurposeMaxTokens(params), headers: { [DETAIL_KIMI_THINKING_HEADER]: thinking, [DETAIL_KIMI_PROFILE_HEADER]: `detail-ui-${params.purpose}`, diff --git a/lib/ai/prompt-optimizer.ts b/lib/ai/prompt-optimizer.ts index 52668afb3..b09bdddb5 100644 --- a/lib/ai/prompt-optimizer.ts +++ b/lib/ai/prompt-optimizer.ts @@ -687,7 +687,7 @@ export async function optimizePromptWithLLM(params: { system: optimizerSystemPrompt, prompt: finalUserMessage, temperature: getCompatibleTextGenerationTemperature(selectedOptimizerModel, 0.2), - maxTokens: 200, + maxOutputTokens: 200, abortSignal: AbortSignal.timeout(12_000), } as any); diff --git a/lib/ai/providers/vip98k-chat.ts b/lib/ai/providers/vip98k-chat.ts index 4027d6c50..e807d72dd 100644 --- a/lib/ai/providers/vip98k-chat.ts +++ b/lib/ai/providers/vip98k-chat.ts @@ -153,28 +153,30 @@ export function createVip98kChatModel(params: { (!metadata.chatInvocation && invocation.protocol === "anthropic-messages")); const resolveModel = async (signal?: AbortSignal) => { signal?.throwIfAborted(); - if (!needsDiscovery) return model; + if (!needsDiscovery) return { model, contract: invocation }; const models = await awaitWithAbort(discoverContracts(root, baseFetch), signal); const contract = selectVip98kChatInvocation(models.get(params.modelId)); if (!contract) throw new Error("VIP98K 此模型未公布兼容的对话接口,请核对模型类型和接口配置。"); - return createModel(contract); + return { model: createModel(contract), contract }; }; + const callOptions = (options: LanguageModelV2CallOptions, contract: typeof invocation) => + contract?.omitSampling ? { ...options, temperature: undefined, topP: undefined, topK: undefined } : options; return { specificationVersion: model.specificationVersion, provider: model.provider, modelId: model.modelId, supportedUrls: model.supportedUrls, async doGenerate(options) { - const activeModel = await resolveModel(options.abortSignal); - const result = await activeModel.doGenerate(options); + const active = await resolveModel(options.abortSignal); + const result = await active.model.doGenerate(callOptions(options, active.contract)); if (requiresToolCall(options) && !result.content.some((part) => part.type === "tool-call")) throw missingToolError(); if (!result.content.some((part) => part.type === "tool-call" || (part.type === "text" && part.text.trim()))) throw emptyResponseError(); return result; }, async doStream(options) { - const activeModel = await resolveModel(options.abortSignal); - const result = await activeModel.doStream(options); + const active = await resolveModel(options.abortSignal); + const result = await active.model.doStream(callOptions(options, active.contract)); let usable = false; let toolCalled = false; let failed = false; diff --git a/lib/ai/providers/vip98k.ts b/lib/ai/providers/vip98k.ts index 266e8e33c..3661bcba6 100644 --- a/lib/ai/providers/vip98k.ts +++ b/lib/ai/providers/vip98k.ts @@ -28,6 +28,8 @@ import { normalizeVip98kInvocationHeaders, resolveVip98kImageTaskSize, selectVip98kTaskInvocation, + vip98kReferenceLimit, + getVip98kQualityOptions, type Vip98kTaskInvocation, } from "@/lib/ai/vip98k-task-invocation"; import { VIP98K_BASE_URL, normalizeVip98kBaseUrl } from "@/lib/ai/vip98k-models"; @@ -1574,6 +1576,14 @@ export function buildTaskBody(params: Vip98kTaskParams, invocation: Vip98kInvoca const imageParams = params.imageParams || {}; const promptTitle = params.prompt.split(/\s+/).slice(0, 8).join(" ").slice(0, 80); const body: Record = { ...(invocation?.requestExample || {}) }; + if (toRecord(body.metadata)) body.metadata = { ...toRecord(body.metadata)! }; + const setField = (key: string, value: unknown) => { + const metadata = toRecord(body.metadata); + if (metadata && key in metadata) { + metadata[key] = value; + delete body[key]; + } else body[key] = value; + }; if ("model" in body || !invocation) body.model = params.modelId; if ("displayName" in body || !invocation) { body.displayName = promptTitle || params.modelId; @@ -1652,11 +1662,31 @@ export function buildTaskBody(params: Vip98kTaskParams, invocation: Vip98kInvoca } } if (negativePrompt) { - if ("negativePrompt" in body || !invocation) body.negativePrompt = negativePrompt; - if ("negative_prompt" in body || !invocation) body.negative_prompt = negativePrompt; + const key = ["negative_prompt", "negativePrompt"].find(key => + invocationSupportsBodyField(invocation, key)); + if (key) { + const other = key === "negative_prompt" ? "negativePrompt" : "negative_prompt"; + delete body[other]; + if (toRecord(body.metadata)) delete (body.metadata as Record)[other]; + setField(key, negativePrompt); + } } if (quality && invocationSupportsBodyField(invocation, "quality")) { - body.quality = quality; + const options = getVip98kQualityOptions(invocation?.inputSchema); + const effectiveQuality = options?.length === 1 ? options[0] : quality; + if (options?.length && !options.includes(effectiveQuality)) throw Error("当前模型不支持所选图片质量,请重新选择"); + setField("quality", effectiveQuality); + } + const generationParamAliases = [ + ["seed", "seed"], + ["watermark", "watermark"], + ["promptExtend", "prompt_extend"], + ["cameraFixed", "camera_fixed"], + ["returnLastFrame", "return_last_frame"], + ] as const; + for (const [sourceKey, targetKey] of generationParamAliases) { + const value = params.kind === "video" ? videoParams[sourceKey] : imageParams[sourceKey]; + if (value !== undefined && invocationSupportsBodyField(invocation, targetKey)) setField(targetKey, value); } const imageToolParamAliases = [ ["scaleFactor", "scale_factor"], @@ -1698,9 +1728,8 @@ export function buildTaskBody(params: Vip98kTaskParams, invocation: Vip98kInvoca delete body.size; } const imageBindingKeys = bindVip98kImages(body, params.imageUrls, invocation?.inputSchema); - const imageProperties = toRecord(invocation?.inputSchema?.properties) || {}; for (const key of imageBindingKeys) { - const limit = toRecord(imageProperties[key])?.maxItems; + const limit = vip98kReferenceLimit(invocation?.inputSchema, key); if (typeof limit === "number" && params.imageUrls.length > limit) throw Error(`当前模型最多支持 ${limit} 张参考图`); } if (params.imageUrls.length > 0 && !invocation && imageBindingKeys.length === 0) { @@ -1759,13 +1788,18 @@ export function buildTaskBody(params: Vip98kTaskParams, invocation: Vip98kInvoca }) : []; const selectedRatio = aspectRatio || "1:1"; if (ratios.includes(selectedRatio)) body.size = selectedRatio; + else if (sizes?.includes(resolution) && hasNativeAspectRatio) body.size = resolution; else if (pixels.length && (resolution || aspectRatio)) { const [rw, rh] = selectedRatio.split(":").map(Number); const matching = pixels.filter(size => Math.abs(size.width / size.height - rw / rh) < 0.025); - if (!matching.length) throw Error("当前模型不支持所选图片比例"); - const area = resolution === "4K" ? 4096 ** 2 : resolution === "2K" ? 2048 ** 2 : 1024 ** 2; - matching.sort((a, b) => Math.abs(a.width * a.height - area) - Math.abs(b.width * b.height - area)); - body.size = matching[0].size; + if (!matching.length && sizes?.includes("auto") && hasNativeAspectRatio && hasNativeResolution) { + body.size = "auto"; + } else { + if (!matching.length) throw Error("当前模型不支持所选图片比例"); + const area = resolution === "4K" ? 4096 ** 2 : resolution === "2K" ? 2048 ** 2 : 1024 ** 2; + matching.sort((a, b) => Math.abs(a.width * a.height - area) - Math.abs(b.width * b.height - area)); + body.size = matching[0].size; + } } if (ratios.length || pixels.length) { const suffix = "\n\n[Output constraints]"; diff --git a/lib/ai/runtime/direct-agent-analysis-answer.ts b/lib/ai/runtime/direct-agent-analysis-answer.ts index bbdd8d9c7..bc9e98144 100644 --- a/lib/ai/runtime/direct-agent-analysis-answer.ts +++ b/lib/ai/runtime/direct-agent-analysis-answer.ts @@ -55,7 +55,7 @@ export async function buildDirectAgentImageAnalysisAnswer(params: { "请用中文直接回答,不要复述“根据摘要/根据图片”,也不要输出多余寒暄。", ].join("\n\n"), temperature: 0.2, - maxTokens: 320, + maxOutputTokens: 320, abortSignal: AbortSignal.timeout(20_000), } as any); diff --git a/lib/ai/smart-analyzer.ts b/lib/ai/smart-analyzer.ts index c69e0c949..9dfa9ae18 100644 --- a/lib/ai/smart-analyzer.ts +++ b/lib/ai/smart-analyzer.ts @@ -1624,7 +1624,7 @@ export async function summarizeImageForExecutionReport(params: { ...(compatibleTemperature !== undefined ? { temperature: compatibleTemperature } : {}), - maxTokens: 120, + maxOutputTokens: 120, abortSignal: AbortSignal.timeout(attemptTimeoutMs), } as any); @@ -2251,7 +2251,7 @@ ${historyText} }, ], ...(analyzerTemperature !== undefined ? { temperature: analyzerTemperature } : {}), - maxTokens: 360, + maxOutputTokens: 360, abortSignal: AbortSignal.timeout(ANALYZER_TIMEOUT_MS), } : { @@ -2259,7 +2259,7 @@ ${historyText} system: ANALYZER_SYSTEM_PROMPT, prompt: userMessage, ...(analyzerTemperature !== undefined ? { temperature: analyzerTemperature } : {}), - maxTokens: 300, + maxOutputTokens: 300, abortSignal: AbortSignal.timeout(ANALYZER_TIMEOUT_MS), }; diff --git a/lib/ai/vip98k-chat-contract.ts b/lib/ai/vip98k-chat-contract.ts index d3fefa25d..0e728613d 100644 --- a/lib/ai/vip98k-chat-contract.ts +++ b/lib/ai/vip98k-chat-contract.ts @@ -5,6 +5,7 @@ export type Vip98kChatInvocation = { protocol: "openai-chat" | "openai-responses" | "anthropic-messages"; path: "/v1/chat/completions" | "/v1/responses" | "/v1/messages"; headers: Record; + omitSampling?: boolean; }; function record(value: unknown): RecordValue | null { @@ -56,10 +57,18 @@ export function selectVip98kChatInvocation(value: unknown): Vip98kChatInvocation }); const selected = candidates[0]; if (!selected) return null; + const selectedRoute = getVip98kInvocationHeaders(selected.headers)["X-Relay-Route-Option"]; + const schemaSource = candidates.find(candidate => candidate.path === selected.path && + getVip98kInvocationHeaders(candidate.headers)["X-Relay-Route-Option"] === selectedRoute && + (candidate.input_schema || candidate.inputSchema)); + const properties = record(record(schemaSource?.input_schema || schemaSource?.inputSchema)?.properties); + const fixedSampling = selected.omitSampling === true || ["temperature", "top_p"].some(key => + record(properties?.[key])?.const !== undefined); return { protocol: selected.path === "/v1/messages" ? "anthropic-messages" : selected.path === "/v1/responses" ? "openai-responses" : "openai-chat", path: selected.path as Vip98kChatInvocation["path"], headers: getVip98kInvocationHeaders(selected.headers), + ...(fixedSampling ? { omitSampling: true } : {}), }; } diff --git a/lib/ai/vip98k-models.ts b/lib/ai/vip98k-models.ts index f6f8f121e..2453d7631 100644 --- a/lib/ai/vip98k-models.ts +++ b/lib/ai/vip98k-models.ts @@ -5,6 +5,8 @@ import { hasVerifiedVip98kOpenAiImageArrayContract, normalizeVip98kInvocationHeaders, getVip98kImageBindingKeys, + vip98kReferenceLimit, + selectVip98kTaskInvocation, } from "@/lib/ai/vip98k-task-invocation"; import { selectVip98kChatInvocation, vip98kChatSupportsVision } from "./vip98k-chat-contract"; @@ -568,7 +570,7 @@ function getSupportedImageSizes(model: Vip98kCatalogModel) { } const primary = toRecord(toRecord(model.invocation?.input_schema)?.properties) || {}; const resolutionEnum = parseStringArray(toRecord(primary.resolution)?.enum); - if (resolutionEnum.length) return resolutionEnum.map(normalizeImageSize).filter((size): size is string => Boolean(size)); + if (resolutionEnum.length) return [...new Set(resolutionEnum.map(normalizeImageSize).filter((size): size is string => Boolean(size)))]; const schemaSizes = getInvocationCandidates(model).flatMap((invocation) => { const properties = toRecord(toRecord(invocation.input_schema)?.properties) || {}; @@ -752,7 +754,11 @@ function createImagePreset( "ratio", "aspect-ratio", ].some((key) => invocationKeys.has(key)) || hasVerifiedVip98kOpenAiImageArrayContract(modelId); - const supportsQuality = invocationKeys.has("quality"); + const supportsQuality = getInvocationCandidates(model).some(invocation => { + const field = toRecord(toRecord(invocation.input_schema)?.properties)?.quality; + return field !== false && Boolean(field || invocation.request_example?.quality !== undefined) && + !toRecord(field)?.const; + }); // Flare repeatedly returns 3:4/4:3 for 4:5/5:4, even with aspect_ratio supplied. const declaredRatios = getInvocationCandidates(model).flatMap(invocation => { const props = toRecord(toRecord(invocation.input_schema)?.properties) || {}; @@ -767,14 +773,17 @@ function createImagePreset( : []), ...(supportsReferenceImages ? ["image-to-image", "reference-images"] : []), ]; - const declaredReferenceLimits = getInvocationCandidates(model).flatMap(invocation => { - const properties = toRecord(toRecord(invocation.input_schema)?.properties) || {}; - return ["file_urls", "image", "images", "reference_images"].flatMap(key => { - const limit = toRecord(properties[key])?.maxItems; - return typeof limit === "number" && limit > 0 ? [limit] : []; - }); + const metadata = getVip98kMetadata(model, "image"); + const referenceInvocation = selectVip98kTaskInvocation({ + modelId, kind: "image", imageCount: 1, + configuredInvocation: metadata.invocation, + configuredInvocations: metadata.invocations, }); - const maxReferenceImages = declaredReferenceLimits.length ? Math.max(...declaredReferenceLimits) : supportsMultiImageReference + const declaredReferenceLimits = getVip98kImageBindingKeys(referenceInvocation).flatMap(key => { + const limit = vip98kReferenceLimit(referenceInvocation?.inputSchema, key); + return typeof limit === "number" && limit > 0 ? [limit] : []; + }); + const maxReferenceImages = declaredReferenceLimits.length ? Math.min(...declaredReferenceLimits) : supportsMultiImageReference ? 8 : supportsReferenceImages ? 1 @@ -805,7 +814,7 @@ function createImagePreset( supportedAspectRatios, supportsMultiImageReference, maxReferenceImages, - vip98k: getVip98kMetadata(model, "image"), + vip98k: metadata, }, defaultParams: { provider: "vip98k", @@ -841,7 +850,7 @@ function createImagePreset( }, enableLoadingPlaceholder: true, loadingColor: "yellow", - vip98k: getVip98kMetadata(model, "image"), + vip98k: metadata, }, sortOrder: TYPE_SORT_ORDER.image + index, }; diff --git a/lib/ai/vip98k-task-invocation.ts b/lib/ai/vip98k-task-invocation.ts index 0de845128..3f8608517 100644 --- a/lib/ai/vip98k-task-invocation.ts +++ b/lib/ai/vip98k-task-invocation.ts @@ -316,6 +316,32 @@ function scoreReferenceInvocation(invocation: Vip98kTaskInvocation) { return score; } +export function vip98kReferenceLimit(schema: unknown, key: string): number | undefined { + if (!schema || typeof schema !== "object" || Array.isArray(schema)) return; + const current = schema as Record; + const limits: number[] = []; + const visitField = (field: any) => { + if (!field || typeof field !== "object") return; + if (typeof field.maxItems === "number") limits.push(field.maxItems); + for (const child of Array.isArray(field.allOf) ? field.allOf : []) visitField(child); + }; + visitField(current.properties?.[key]); + for (const child of Array.isArray(current.allOf) ? current.allOf : []) { + const limit = vip98kReferenceLimit(child, key); + if (limit !== undefined) limits.push(limit); + } + return limits.length ? Math.min(...limits) : undefined; +} + +export function getVip98kQualityOptions(schema: Record | null | undefined): string[] | undefined { + const field = (schema?.properties as Record | undefined)?.quality; + if (field === false) return []; + if (!field || typeof field !== "object") return undefined; + const quality = field as Record; + if (typeof quality.const === "string") return [quality.const]; + return Array.isArray(quality.enum) ? quality.enum.filter((value): value is string => typeof value === "string") : undefined; +} + export function encodeVip98kTaskBody(body: Record, contentType?: string | null): { body: string | FormData; headers: Record } { if (contentType?.split(";", 1)[0].trim() !== "multipart/form-data") { return { body: JSON.stringify(body), headers: { "Content-Type": contentType || "application/json" } }; @@ -356,14 +382,22 @@ export function selectVip98kTaskInvocation(params: { ); if ((params.imageCount || 0) > 0) { - const referenceInvocation = candidates - .filter(invocationSupportsVip98kImages) + const references = candidates.filter(invocationSupportsVip98kImages); + const routedReferences = references.filter(candidate => + getVip98kInvocationHeaders(candidate.headers)["X-Relay-Route-Option"]); + const referenceInvocation = (routedReferences.length ? routedReferences : references) .sort((a, b) => scoreReferenceInvocation(b) - scoreReferenceInvocation(a))[0]; if (referenceInvocation) { return resolveVip98kTaskInvocation(params.modelId, referenceInvocation); } } + if (params.configuredInvocation && + getVip98kInvocationHeaders(params.configuredInvocation.headers)["X-Relay-Route-Option"] && + !params.configuredInvocation.path.includes(":generateContent")) { + return resolveVip98kTaskInvocation(params.modelId, params.configuredInvocation); + } + if (params.kind === "image" && !params.imageCount) { const textImageTask = candidates.find(candidate => { const properties = candidate.inputSchema?.properties as Record | undefined; @@ -449,7 +483,6 @@ function createVerifiedOpenAiImageInvocation(modelId: string) { function addVerifiedOpenAiImageFields( invocation: Vip98kTaskInvocation, - modelId: string, ): Vip98kTaskInvocation { const inputSchema = invocation.inputSchema || {}; const properties = @@ -468,16 +501,8 @@ function addVerifiedOpenAiImageFields( ...inputSchema, properties: { ...properties, - image: { items: { type: "string" }, type: "array" }, - aspect_ratio: { type: "string" }, - ...(hasVerifiedVip98kGptImage25ResolutionTiers(modelId) - ? { - size: { - enum: ["1024x1024", "2048x2048", "2880x2880"], - type: "string", - }, - } - : {}), + image: properties.image ?? { items: { type: "string" }, type: "array" }, + aspect_ratio: properties.aspect_ratio ?? { type: "string" }, }, }, }; @@ -499,7 +524,7 @@ export function resolveVip98kTaskInvocation( invocationTargetsModel(configuredInvocation, normalizedModelId) ? configuredInvocation : createVerifiedOpenAiImageInvocation(modelId); - return addVerifiedOpenAiImageFields(invocation, normalizedModelId); + return addVerifiedOpenAiImageFields(invocation); } const processingRequest = KNOWN_IMAGE_PROCESSING_TOOL_REQUESTS[normalizedModelId]; if (processingRequest) { diff --git a/lib/video-agent/storyboard-command-interpreter.ts b/lib/video-agent/storyboard-command-interpreter.ts index 9eb466904..40d194530 100644 --- a/lib/video-agent/storyboard-command-interpreter.ts +++ b/lib/video-agent/storyboard-command-interpreter.ts @@ -259,7 +259,7 @@ async function interpretStoryboardCommandWithLLM(params: { .filter(Boolean) .join("\n\n"), temperature: 0.1, - maxTokens: 200, + maxOutputTokens: 200, abortSignal: AbortSignal.timeout(12000), } as any); diff --git a/lib/video-canvas/model-matching.ts b/lib/video-canvas/model-matching.ts index e5fc35f25..2ffe73c3f 100644 --- a/lib/video-canvas/model-matching.ts +++ b/lib/video-canvas/model-matching.ts @@ -86,7 +86,7 @@ export function filterCanvasModels( }); } -export function getCanvasVideoParameters(model: CanvasModelDescriptor | undefined, mode?: VideoGenerationMode) { +export function getCanvasVideoParameters(model: CanvasModelDescriptor | undefined, mode?: VideoGenerationMode, resolution?: string) { const caps = record(model?.capabilities); const scoped = record(record(caps.modeConstraints || caps.constraints)[mode || ""]); const strings = (value: unknown) => Array.isArray(value) ? [...new Set(value.map(String).filter(Boolean))] : []; @@ -94,11 +94,15 @@ export function getCanvasVideoParameters(model: CanvasModelDescriptor | undefine const min = Number(durationRange.min), max = Number(durationRange.max); const configured = record((model as CanvasModelDescriptor & {imageGenConfig?: unknown})?.imageGenConfig).videoDurationOptions; const durations = Array.isArray(scoped.durationOptions) ? scoped.durationOptions : Array.isArray(configured) && configured.length ? configured : caps.durationOptions; + const durationValues = Array.isArray(durations) ? [...new Set(durations.map(Number).filter(n => Number.isFinite(n) && n > 0))].sort((a,b) => a-b) : + min > 0 && max >= min && max-min <= 120 ? Array.from({length: max-min+1}, (_,i) => min+i) : []; + const bounds = model?.providerType === "vip98k" && resolution + ? vip98kVideoDurationBounds(record(caps.vip98k).invocation?.inputSchema, resolution.toLowerCase()) + : { min: 0, max: Infinity, durations: undefined }; return { ratios: strings(scoped.ratioOptions || caps.supportedRatios), resolutions: strings(scoped.resolutionOptions || caps.supportedResolutions), - durations: Array.isArray(durations) ? [...new Set(durations.map(Number).filter(n => Number.isFinite(n) && n > 0))].sort((a,b) => a-b) : - min > 0 && max >= min && max-min <= 120 ? Array.from({length: max-min+1}, (_,i) => min+i) : [], + durations: durationValues.filter(value => value >= bounds.min && value <= bounds.max && (!bounds.durations || bounds.durations.includes(value))), supportsAudio: caps.supportGenerateAudio === true || caps.supportsAudio === true || caps.generateAudio === true || strings(record(record(model?.imageGenConfig).atlasCloud).propertyNames).includes("generate_audio"), }; } diff --git a/tests/routes/provider-call-compatibility.test.ts b/tests/routes/provider-call-compatibility.test.ts index 1792f140d..d4783d1b9 100644 --- a/tests/routes/provider-call-compatibility.test.ts +++ b/tests/routes/provider-call-compatibility.test.ts @@ -309,3 +309,17 @@ test("a cancelled request can leave shared discovery without submitting a paid P finishCatalog(Response.json({ data: { sections: [{ models: [catalogModel] }] } })); expect(posts).toBe(0); }); + +test("fixed-sampling native contracts omit sampling controls while preserving tools and token limits", async () => { + let sent: any; + const metadata={chatInvocation:{protocol:'anthropic-messages',path:'/v1/messages',headers:{'X-Relay-Route-Option':'native_route'}},invocation:{kind:'chat',path:'/v1/messages',headers:{'X-Relay-Route-Option':'native_route'},inputSchema:{properties:{temperature:{const:1},top_p:{const:.99},top_k:false}}}}; + const selected=selectVip98kChatInvocation(metadata); + expect(selected?.omitSampling).toBe(true); + const nativeModel=createVip98kChatModel({apiKey:'fixture',baseURL:'https://relay.example',modelId:'claude-haiku-5-5',metadata,fetch:async(_url,init)=>{ + sent=JSON.parse(String(init?.body)); + return Response.json({id:'test',type:'message',role:'assistant',model:'claude-haiku-5-5',content:[{type:'text',text:'ok'}],stop_reason:'end_turn',stop_sequence:null,usage:{input_tokens:1,output_tokens:1}}); + }}); + await generateText({model:nativeModel,prompt:'test',temperature:.2,topP:.9,topK:4,maxOutputTokens:321,maxRetries:0}); + expect(sent.max_tokens).toBe(321); + for(const key of ['temperature','top_p','top_k'])expect(sent).not.toHaveProperty(key); +}); diff --git a/tests/routes/vip98k-model-expansion.test.ts b/tests/routes/vip98k-model-expansion.test.ts index 18c380b65..23f895815 100644 --- a/tests/routes/vip98k-model-expansion.test.ts +++ b/tests/routes/vip98k-model-expansion.test.ts @@ -178,3 +178,90 @@ test("completed audio checkpoints retain the original provider and resumable tas const checkpoint={operation:"audio",provider:"vip98k",modelRef:"test",modelId:"gem-2.5-tts",providerId:"source",baseUrl:"https://relay.example",taskId:"original",url:"https://example.com/audio.wav",format:"wav"}; expect(getAudioCheckpoint({request:{kind:"audio",params:{modelRef:"test"}},result:{upstreamAudio:checkpoint}})).toEqual(checkpoint); }); + +test("route-scoped image contracts take precedence over generic task examples", () => { + const selected = {...invocation({size:{type:'string'},image:{type:'array',maxItems:8}}),headers:{'X-Relay-Route-Option':'selected'}}; + const generic = {...invocation({resolution:{enum:['1K','2K']},aspect_ratio:{enum:['1:1','16:9']},file_urls:{type:'array'}},{file_urls:[]}),kind:'task',path:'/extensions/v1/tasks'}; + for(const imageCount of [0,1]) { + const resolved=selectVip98kTaskInvocation({modelId:'image',kind:'image',configuredInvocation:selected,configuredInvocations:[generic,selected],imageCount}); + expect(resolved?.headers).toEqual({'X-Relay-Route-Option':'selected'}); + } +}); + +test("native ratio and resolution contracts can use automatic pixels for non-square images", () => { + const contract=invocation({size:{enum:['auto','1024x1024','2048x2048','4096x4096']},resolution:{enum:['1K','2K','4K']},aspect_ratio:{enum:['1:1','16:9','9:16']}},{size:'1024x1024'}); + for(const aspectRatio of ['16:9','9:16'])for(const imageSize of ['1K','2K','4K']) { + const result=buildTaskBody({...params,imageParams:{imageSize,aspectRatio}},contract); + expect(result.body).toMatchObject({size:'auto',resolution:imageSize,aspect_ratio:aspectRatio}); + } +}); + +test("reference limits inside allOf are enforced before submission and during import", () => { + const schema={properties:{model:{type:'string'},prompt:{type:'string'},images:{allOf:[{type:'array',maxItems:16},{maxItems:14}]}}}; + const contract={...invocation({},{images:[]}),inputSchema:schema}; + const images=Array.from({length:15},(_,i)=>`https://fixture.example/${i}.png`); + expect(()=>buildTaskBody({...params,imageUrls:images},contract)).toThrow('14'); + const [model]=buildVip98kModelPresets([{id:'image',type:'image',invocation:{kind:'tool',path:contract.path,input_schema:schema,request_example:contract.requestExample}}]); + expect(model.imageGenConfig?.maxReferenceImages).toBe(14); +}); + +test("quality and negative prompt have a single effective location", () => { + const metadata={quality:'high'}; + const contract=invocation({quality:{const:'high'},negative_prompt:{type:'string'},negativePrompt:{type:'string'}},{metadata,negativePrompt:'example'}); + const result=buildTaskBody({...params,imageParams:{quality:'auto',negativePrompt:'blurry'}},contract); + expect(result.body.metadata).toEqual({quality:'high'}); + expect(result.body).not.toHaveProperty('quality'); + expect(result.body.negative_prompt).toBe('blurry'); + expect(result.body).not.toHaveProperty('negativePrompt'); + expect(metadata).toEqual({quality:'high'}); +}); + +test("fixed resolution edit contracts determine the imported default", () => { + for(const resolution of ['2K','4K']){ + const [model] = buildVip98kModelPresets([{ + id: 'fixed-image', type: 'image', + invocation: { + kind: 'tool', path: '/v1/images/generations', + request_example: { model: 'fixed-image', prompt: 'test' }, + input_schema: { properties: { size: { type: 'string' } } }, + }, + invocations: [{ + kind: 'tool', path: '/v1/images/edits', + request_example: { model: 'fixed-image', prompt: 'test', file_urls: [] }, + input_schema: { properties: { + file_urls: { type: 'array', maxItems: 16 }, size: { enum: ['auto'] }, + resolution: { enum: [resolution] }, aspect_ratio: { enum: ['1:1', '16:9'] }, + } }, + }], + }]); + expect(model.defaultParams.imageSize).toBe(resolution); + expect(model.imageGenConfig?.supportedImageSizes).toEqual([resolution]); + } +}); + +test("video duration choices follow the selected resolution", () => { + const schema={properties:{duration:{minimum:1,maximum:15}},allOf:[{if:{properties:{resolution:{const:'1080p'}}},then:{properties:{duration:{maximum:10}}}}]}; + const model={providerType:'vip98k',capabilities:{kind:'video',durationOptions:[5,10,15],vip98k:{invocation:{inputSchema:schema}}}}; + expect(getCanvasVideoParameters(model,'text-to-video','1080P').durations).toEqual([5,10]); + expect(getCanvasVideoParameters(model,'text-to-video','720p').durations).toEqual([5,10,15]); +}); + +test("generation controls preserve false and zero in the declared wire fields", () => { + const contract=invocation({seed:{type:'integer'},watermark:{type:'boolean'},prompt_extend:{type:'boolean'},return_last_frame:{type:'boolean'},camera_fixed:{type:'boolean'}},{metadata:{watermark:true}}); + const image=buildTaskBody({...params,imageParams:{seed:0,watermark:false,promptExtend:false}},contract).body; + expect(image).toMatchObject({seed:0,prompt_extend:false,metadata:{watermark:false}}); + expect(image).not.toHaveProperty('watermark'); + const video=buildTaskBody({...params,kind:'video',videoParams:{seed:0,watermark:false,returnLastFrame:true,cameraFixed:false}},contract).body; + expect(video).toMatchObject({seed:0,return_last_frame:true,camera_fixed:false,metadata:{watermark:false}}); + const unsupported=buildTaskBody({...params,imageParams:{seed:0,watermark:false}},invocation({seed:false,watermark:false})).body; + expect(unsupported).not.toHaveProperty('seed'); + expect(unsupported).not.toHaveProperty('watermark'); +}); + +test("fixed quality routes replace stale quality choices before submission", () => { + const contract=invocation({quality:{enum:['auto']}},{quality:'auto'}); + for(const quality of ['auto','low','medium','high']) { + expect(buildTaskBody({...params,imageParams:{quality}},contract).body.quality).toBe('auto'); + } + expect(()=>buildTaskBody({...params,imageParams:{quality:'invalid'}},invocation({quality:{enum:['low','high']}}))).toThrow('图片质量'); +}); diff --git a/tests/routes/vip98k-reference-routing.test.ts b/tests/routes/vip98k-reference-routing.test.ts index c43117a38..ce4bce0bf 100644 --- a/tests/routes/vip98k-reference-routing.test.ts +++ b/tests/routes/vip98k-reference-routing.test.ts @@ -970,13 +970,17 @@ test.describe("VIP98K reference-image routing", () => { properties: { aspect_ratio: { type: "string" }, image: { type: "array" }, - size: { - enum: ["1024x1024", "2048x2048", "2880x2880"], - }, + }, }); }); + test("Flare retains non-square pixel geometry", () => { + const contract = resolveVip98kTaskInvocation('gpt-image-2.5-flare', null); + const wide = buildTaskBody({apiKey:'fixture',modelId:'gpt-image-2.5-flare',kind:'image',prompt:'wide product',imageUrls:[],videoUrls:[],modelConfig:{},imageParams:{imageSize:'2K',aspectRatio:'16:9'}},contract); + expect(wide.body.size).toBe('2720x1536'); + }); + test("does not inherit multi-image support from broad authorized metadata", () => { const presets = buildVip98kModelPresets([ {