@shop-os/modelkit 0.1.5 → 0.1.7

package/dist/index.cjs

@@ -68,6 +68,7 @@ __export(index_exports, {
   prodiaAdapter: () => prodiaAdapter,
   readErrorBody: () => readErrorBody,
   replicateAdapter: () => replicateAdapter,
+  resolveAudioOptions: () => resolveAudioOptions,
   resolveImageOptions: () => resolveImageOptions,
   resolveKeys: () => resolveKeys,
   resolveVideoOptions: () => resolveVideoOptions,
@@ -165,6 +166,7 @@ var Catalog = class _Catalog {
 
 // src/catalog/image.ts
 var IMAGE_USAGE = { unit: "images", modality: "IMAGE" };
+var IMAGE_GENERATE_EDIT = { imageGenerate: true, imageEdit: true };
 var GEMINI_RATIOS = [
   "1:1",
   "2:3",
@@ -178,7 +180,7 @@ var GEMINI_RATIOS = [
   "5:4",
   "adaptive"
 ];
-function geminiImage(id, label, aliases, sizes) {
+function geminiImage(id, label, aliases, sizes, display) {
   const model = {
     id,
     aliases,
@@ -187,25 +189,47 @@ function geminiImage(id, label, aliases, sizes) {
     adapter: "google-image",
     providerModelId: id,
     lifecycle: "active",
-    constraints: { aspectRatios: [...GEMINI_RATIOS], sizes, supports: ["safetySettings"] },
+    constraints: {
+      aspectRatios: [...GEMINI_RATIOS],
+      sizes,
+      supports: ["safetySettings", "temperature"]
+    },
     defaults: { size: "1K", aspectRatio: "1:1" },
     usage: { ...IMAGE_USAGE },
-    pricingKind: "image_size"
+    pricingKind: "image_size",
+    display,
+    capabilities: { ...IMAGE_GENERATE_EDIT }
   };
   return model;
 }
 var IMAGE_MODELS = [
-  geminiImage("gemini-3-pro-image-preview", "Nano Banana Pro", ["nb-pro"], ["1K", "2K", "4K"]),
+  geminiImage("gemini-3-pro-image-preview", "Nano Banana Pro", ["nb-pro"], ["1K", "2K", "4K"], {
+    provider: "google",
+    description: "Create stunning, production-quality visuals with exceptional realism, detail, and typography. Best for campaigns, e-commerce, product photography, and hero creatives.",
+    logo: "https://d31vvfpk2ppdyq.cloudfront.net/assets/image/192485d2-a76a-4657-9c06-930b2c70badd_compressed.png"
+  }),
   geminiImage(
     "gemini-3.1-flash-image-preview",
     "Nano Banana Flash",
     ["nb-2"],
     // Flash 3.1 is the only Gemini image model with the 512 tier.
-    ["512", "1K", "2K", "4K"]
+    ["512", "1K", "2K", "4K"],
+    {
+      provider: "google",
+      description: "Generate high-quality images at speed. Perfect for brainstorming, moodboards, concept testing, creative exploration, and large batch workflows.",
+      logo: "https://d31vvfpk2ppdyq.cloudfront.net/assets/image/192485d2-a76a-4657-9c06-930b2c70badd_compressed.png"
+    }
   ),
-  geminiImage("gemini-3.1-flash-lite-image", "Nano Banana Lite", ["nb-lite"], ["1K"]),
+  geminiImage("gemini-3.1-flash-lite-image", "Nano Banana Lite", ["nb-lite"], ["1K"], {
+    provider: "google",
+    description: "The fastest and most affordable image model. Ideal for rapid drafts, bulk exploration, and iterative concept testing at 1K resolution.",
+    logo: "https://d31vvfpk2ppdyq.cloudfront.net/assets/image/192485d2-a76a-4657-9c06-930b2c70badd_compressed.png"
+  }),
   // trigger-workflows validates 1K only; pricing seeds 2K/4K rows so allow all three.
-  geminiImage("gemini-2.5-flash-image", "Gemini 2.5 Flash Image", [], ["1K", "2K", "4K"]),
+  geminiImage("gemini-2.5-flash-image", "Gemini 2.5 Flash Image", [], ["1K", "2K", "4K"], {
+    provider: "google",
+    description: "Nano Banana (legacy Gemini 2.5 Flash Image) native image generation and editing model for high-volume, low-latency creative workflows."
+  }),
   {
     // Cowork "auto" lets OpenAI pick the size; the canonical equivalent is adaptive.
     id: "gpt-image-2",
@@ -223,7 +247,13 @@ var IMAGE_MODELS = [
     },
     defaults: { size: "1K", aspectRatio: "adaptive", quality: "auto" },
     usage: { ...IMAGE_USAGE },
-    pricingKind: "image_size"
+    pricingKind: "image_size",
+    display: {
+      provider: "openai",
+      description: "The most capable model for image editing and complex creative instructions. Excels at text rendering, visual consistency, design generation, and iterative refinement.",
+      logo: "https://d31vvfpk2ppdyq.cloudfront.net/assets/image/1204b494-8962-4f7f-af5e-699a8bdc98e3_compressed.png"
+    },
+    capabilities: { ...IMAGE_GENERATE_EDIT }
   },
   {
     // Same Images API surface as gpt-image-2 plus the "xhigh" and "max" quality tiers.
@@ -242,7 +272,12 @@ var IMAGE_MODELS = [
     },
     defaults: { size: "1K", aspectRatio: "adaptive", quality: "auto" },
     usage: { ...IMAGE_USAGE },
-    pricingKind: "image_size"
+    pricingKind: "image_size",
+    display: {
+      provider: "openai",
+      description: "OpenAI's most capable GPT Image 2.5 tier, built for editing precision; takes text and image input and outputs images in low, medium, high, xhigh, or max quality."
+    },
+    capabilities: { ...IMAGE_GENERATE_EDIT }
   },
   {
     // Same Images API surface as gpt-image-2 plus the "xhigh" and "max" quality tiers.
@@ -261,7 +296,12 @@ var IMAGE_MODELS = [
     },
     defaults: { size: "1K", aspectRatio: "adaptive", quality: "auto" },
     usage: { ...IMAGE_USAGE },
-    pricingKind: "image_size"
+    pricingKind: "image_size",
+    display: {
+      provider: "openai",
+      description: "OpenAI's fastest GPT Image 2.5 tier for high-quality everyday image generation and editing; takes text and image input and outputs images."
+    },
+    capabilities: { ...IMAGE_GENERATE_EDIT }
   },
   {
     // Cowork allows 7 ratios; trigger-workflows allows 9 (adds 3:2, 9:21, drops 4:5/5:4).
@@ -293,7 +333,13 @@ var IMAGE_MODELS = [
     },
     defaults: { size: "1K", aspectRatio: "1:1", outputFormat: "png", count: 1 },
     usage: { ...IMAGE_USAGE },
-    pricingKind: "image_size"
+    pricingKind: "image_size",
+    display: {
+      provider: "meta",
+      description: "Meta's image model with built-in world knowledge and strong prompt adherence. Great for concept-rich scenes, realistic compositions, and reference-guided edits.",
+      logo: "https://d31vvfpk2ppdyq.cloudfront.net/assets/characters/1787909065794-k9a7zz.png"
+    },
+    capabilities: { ...IMAGE_GENERATE_EDIT }
   },
   {
     id: "seedream-5-0-pro",
@@ -313,7 +359,12 @@ var IMAGE_MODELS = [
     },
     defaults: { size: "1K", aspectRatio: "1:1" },
     usage: { ...IMAGE_USAGE },
-    pricingKind: "image_size"
+    pricingKind: "image_size",
+    display: {
+      provider: "bytedance",
+      description: "ByteDance Seedream 5.0 Pro text-to-image and multi-reference image-to-image (up to 10 inputs) with region editing; outputs 1K, 1.5K, or 2K."
+    },
+    capabilities: { ...IMAGE_GENERATE_EDIT }
   },
   {
     id: "seedream-4.5",
@@ -326,14 +377,31 @@ var IMAGE_MODELS = [
     constraints: {
       // 97:30 has no Ark size spelling; the A+ flow pads inputs to 97:30 and
       // omits the ratio instead. Advertise only sizes the adapter can send.
-      aspectRatios: ["1:1", "2:3", "3:2", "4:3", "3:4", "4:5", "5:4", "9:16", "16:9", "21:9"],
-      sizes: ["1K", "2K"],
+      aspectRatios: [
+        "1:1",
+        "2:3",
+        "3:2",
+        "4:3",
+        "3:4",
+        "4:5",
+        "5:4",
+        "9:16",
+        "16:9",
+        "21:9",
+        "9:21"
+      ],
+      sizes: ["1K", "2K", "4K"],
       references: { images: 10 },
       supports: ["seed"]
     },
     defaults: { size: "1K", aspectRatio: "1:1" },
     usage: { ...IMAGE_USAGE },
-    pricingKind: "image_size"
+    pricingKind: "image_size",
+    display: {
+      provider: "bytedance",
+      description: "ByteDance Seedream 4.5 text-to-image and image-to-image with up to 14 reference images and batch image-set generation; outputs 2K or 4K."
+    },
+    capabilities: { ...IMAGE_GENERATE_EDIT }
   },
   {
     id: "flux-2-pro",
@@ -351,7 +419,12 @@ var IMAGE_MODELS = [
     },
     defaults: { size: "1K", aspectRatio: "1:1" },
     usage: { ...IMAGE_USAGE },
-    pricingKind: "image_size"
+    pricingKind: "image_size",
+    display: {
+      provider: "black-forest-labs",
+      description: "Black Forest Labs FLUX.2 [pro] text-to-image and multi-reference editing model served via Prodia; takes a prompt and up to 8 input images and outputs up to 4096x4096."
+    },
+    capabilities: { ...IMAGE_GENERATE_EDIT }
   },
   {
     // Prodia's gemini-3-pro jobs take aspect_ratio + resolution (1K/2K/4K), not
@@ -372,7 +445,12 @@ var IMAGE_MODELS = [
     },
     defaults: { size: "1K", aspectRatio: "1:1" },
     usage: { ...IMAGE_USAGE },
-    pricingKind: "image_size"
+    pricingKind: "image_size",
+    display: {
+      provider: "google",
+      description: "Google Gemini 3 Pro image generation and editing served via Prodia; takes a prompt plus up to 3 reference images and outputs 1K, 2K, or 4K."
+    },
+    capabilities: { ...IMAGE_GENERATE_EDIT }
   },
   {
     // Nano Banana (img2img.v2) is flat-rate with aspect_ratio only; 3:2, 4:5,
@@ -392,7 +470,12 @@ var IMAGE_MODELS = [
     },
     defaults: { size: "1K", aspectRatio: "1:1" },
     usage: { ...IMAGE_USAGE },
-    pricingKind: "image_size"
+    pricingKind: "image_size",
+    display: {
+      provider: "google",
+      description: "Google Nano Banana image generation and editing served via Prodia; takes a prompt plus 1-3 input images and matches the first input's aspect ratio in roughly 8 seconds."
+    },
+    capabilities: { ...IMAGE_GENERATE_EDIT }
   },
   {
     // WAN 2.7 takes `images` (up to 9) and a `size` enum whose exact values
@@ -413,7 +496,12 @@ var IMAGE_MODELS = [
     },
     defaults: { size: "1K", aspectRatio: "1:1" },
     usage: { unit: "compute_seconds", modality: "IMAGE" },
-    pricingKind: "image_size"
+    pricingKind: "image_size",
+    display: {
+      provider: "alibaba",
+      description: "Alibaba Wan 2.7 Image Pro text-to-image and image-editing model; takes a prompt plus up to 9 reference images and generates up to 4K (4096x4096)."
+    },
+    capabilities: { ...IMAGE_GENERATE_EDIT }
   },
   {
     // Qwen Image 2 Pro has aspect_ratio but no size input; resolution follows
@@ -435,7 +523,12 @@ var IMAGE_MODELS = [
     },
     defaults: { aspectRatio: "1:1" },
     usage: { unit: "compute_seconds", modality: "IMAGE" },
-    pricingKind: "image_size"
+    pricingKind: "image_size",
+    display: {
+      provider: "alibaba",
+      description: "Alibaba Qwen Image 2 Pro unified text-to-image and editing model; takes a prompt plus an optional single reference image and generates native 2K output."
+    },
+    capabilities: { ...IMAGE_GENERATE_EDIT }
   },
   {
     // Qwen Image Edit Plus takes aspect_ratio (or match_input_image) and
@@ -457,7 +550,12 @@ var IMAGE_MODELS = [
     },
     defaults: { outputFormat: "png" },
     usage: { unit: "compute_seconds", modality: "IMAGE" },
-    pricingKind: "image_size"
+    pricingKind: "image_size",
+    display: {
+      provider: "alibaba",
+      description: "Alibaba Qwen Image Edit Plus with Lightning and a skin-retouching LoRA; takes one or more reference images plus an edit prompt and outputs webp, jpg, or png."
+    },
+    capabilities: { ...IMAGE_GENERATE_EDIT }
   }
 ];
 
@@ -478,7 +576,11 @@ var OTHER_MODELS = [
     },
     defaults: {},
     usage: { unit: "images", modality: "AUDIO" },
-    pricingKind: "flat"
+    pricingKind: "flat",
+    display: {
+      provider: "google",
+      description: "Google Lyria 3 Clip short music clips, loops, and previews; takes text and up to 10 image prompts and returns a fixed 30-second MP3 plus lyrics."
+    }
   },
   {
     id: "lyria-3-pro-preview",
@@ -494,7 +596,11 @@ var OTHER_MODELS = [
     },
     defaults: {},
     usage: { unit: "images", modality: "AUDIO" },
-    pricingKind: "flat"
+    pricingKind: "flat",
+    display: {
+      provider: "google",
+      description: "Google Lyria 3 Pro flagship music model for full-length songs with verses, choruses, and bridges; takes text and up to 10 image prompts and returns MP3 plus lyrics."
+    }
   },
   {
     id: "google/lyria-2",
@@ -510,7 +616,11 @@ var OTHER_MODELS = [
     },
     defaults: {},
     usage: { unit: "compute_seconds", modality: "AUDIO" },
-    pricingKind: "flat"
+    pricingKind: "flat",
+    display: {
+      provider: "google",
+      description: "Lyria 2 music generation on Replicate; produces 30 seconds of 48 kHz stereo audio from a text prompt."
+    }
   },
   {
     id: "mmaudio",
@@ -527,7 +637,11 @@ var OTHER_MODELS = [
     },
     defaults: { durationSeconds: 8 },
     usage: { unit: "compute_seconds", modality: "AUDIO" },
-    pricingKind: "flat"
+    pricingKind: "flat",
+    display: {
+      provider: "mmaudio",
+      description: "MMAudio V2 video-to-audio model on Replicate; takes a source video and prompt and synthesizes synchronized environmental, action, and speech-like audio."
+    }
   },
   {
     id: "elevenlabs-music",
@@ -544,7 +658,11 @@ var OTHER_MODELS = [
     defaults: { durationSeconds: 15, instrumental: true },
     // Replicate's run() does not report predict_time; billed by generated length.
     usage: { unit: "seconds", modality: "AUDIO" },
-    pricingKind: "flat"
+    pricingKind: "flat",
+    display: {
+      provider: "elevenlabs",
+      description: "ElevenLabs Music text-to-music on Replicate; generates full songs, instrumental or with vocals, up to 5 minutes long at up to 44.1 kHz."
+    }
   },
   {
     id: "topaz",
@@ -557,11 +675,16 @@ var OTHER_MODELS = [
     constraints: {
       references: { images: 1 },
       requires: ["references"],
-      supports: ["scaleFactor", "outputFormat"]
+      supports: ["scaleFactor", "outputFormat", "upscale"]
     },
     defaults: { scaleFactor: 2, outputFormat: "png" },
     usage: { unit: "compute_seconds", modality: "IMAGE" },
-    pricingKind: "flat"
+    pricingKind: "flat",
+    display: {
+      provider: "topaz",
+      description: "Topaz Labs AI image upscaling on Replicate; takes one input image and upscales up to 6x with selectable enhancement models and optional face enhancement."
+    },
+    capabilities: { imageEdit: true }
   },
   {
     id: "bg-remove",
@@ -578,7 +701,12 @@ var OTHER_MODELS = [
     },
     defaults: {},
     usage: { unit: "images", modality: "IMAGE" },
-    pricingKind: "flat"
+    pricingKind: "flat",
+    display: {
+      provider: "prodia",
+      description: "Prodia background removal; takes one input image and returns a PNG cut-out of the subject plus an optional black-and-white mask."
+    },
+    capabilities: { imageEdit: true }
   }
 ];
 
@@ -601,22 +729,112 @@ function text(id, label, providerModelId, extra) {
   };
 }
 var TEXT_MODELS = [
-  text("gpt-4o", "GPT-4o", "gpt-4o"),
-  text("gpt-4o-mini", "GPT-4o mini", "gpt-4o-mini"),
-  text("gpt-5.1", "GPT-5.1", "gpt-5.1"),
-  text("gpt-5.2", "GPT-5.2", "gpt-5.2"),
-  text("gpt-5.4", "GPT-5.4", "gpt-5.4"),
+  text("gpt-4o", "GPT-4o", "gpt-4o", {
+    display: {
+      provider: "openai",
+      description: "OpenAI GPT-4o flagship multimodal model; text and image input, 128K-token context."
+    },
+    capabilities: { vision: true, tools: true, structuredOutput: true, contextWindow: 128e3 }
+  }),
+  text("gpt-4o-mini", "GPT-4o mini", "gpt-4o-mini", {
+    display: {
+      provider: "openai",
+      description: "OpenAI GPT-4o mini fast, low-cost multimodal model; text and image input, 128K-token context."
+    },
+    capabilities: { vision: true, tools: true, structuredOutput: true, contextWindow: 128e3 }
+  }),
+  text("gpt-5.1", "GPT-5.1", "gpt-5.1", {
+    display: {
+      provider: "openai",
+      description: "OpenAI GPT-5.1 reasoning model for coding and agentic tasks; text and image input, 400K-token context."
+    },
+    capabilities: { vision: true, tools: true, structuredOutput: true, contextWindow: 4e5 }
+  }),
+  text("gpt-5.2", "GPT-5.2", "gpt-5.2", {
+    display: {
+      provider: "openai",
+      description: "OpenAI GPT-5.2 reasoning model for complex professional work; text and image input, 400K-token context."
+    },
+    capabilities: { vision: true, tools: true, structuredOutput: true, contextWindow: 4e5 }
+  }),
+  // 1.05M-token context: https://platform.openai.com/docs/models/gpt-5.4
+  text("gpt-5.4", "GPT-5.4", "gpt-5.4", {
+    display: {
+      provider: "openai",
+      description: "OpenAI GPT-5.4 flagship reasoning model; text and image input, 1.05M-token context."
+    },
+    capabilities: {
+      vision: true,
+      tools: true,
+      structuredOutput: true,
+      contextWindow: 105e4
+    }
+  }),
   // OpenAI never shipped gpt-5.1-mini or gpt-5.2-mini; callers that still send them get 5.4 mini.
   text("gpt-5.4-mini", "GPT-5.4 mini", "gpt-5.4-mini", {
-    aliases: ["gpt-5.1-mini", "gpt-5.2-mini"]
+    aliases: ["gpt-5.1-mini", "gpt-5.2-mini"],
+    display: {
+      provider: "openai",
+      description: "OpenAI GPT-5.4 mini efficient reasoning model for high-volume workloads; text and image input, 400K-token context."
+    },
+    capabilities: { vision: true, tools: true, structuredOutput: true, contextWindow: 4e5 }
+  }),
+  text("gpt-5.5", "GPT-5.5", "gpt-5.5", {
+    display: {
+      provider: "openai",
+      description: "OpenAI GPT-5.5 flagship reasoning model for the most complex professional and coding work; text and image input, 1.05M-token context."
+    },
+    capabilities: {
+      vision: true,
+      tools: true,
+      structuredOutput: true,
+      contextWindow: 105e4
+    }
+  }),
+  text("gpt-5.6-luna", "GPT-5.6 Luna", "gpt-5.6-luna", {
+    display: {
+      provider: "openai",
+      description: "OpenAI GPT-5.6 Luna cost-optimized tier of the GPT-5.6 reasoning family; text and image input, 1.05M-token context."
+    },
+    capabilities: {
+      vision: true,
+      tools: true,
+      structuredOutput: true,
+      contextWindow: 105e4
+    }
+  }),
+  text("gpt-5.6-terra", "GPT-5.6 Terra", "gpt-5.6-terra", {
+    display: {
+      provider: "openai",
+      description: "OpenAI GPT-5.6 Terra balanced tier of the GPT-5.6 reasoning family; text and image input, 1.05M-token context."
+    },
+    capabilities: {
+      vision: true,
+      tools: true,
+      structuredOutput: true,
+      contextWindow: 105e4
+    }
+  }),
+  text("gpt-5.6-sol", "GPT-5.6 Sol", "gpt-5.6-sol", {
+    display: {
+      provider: "openai",
+      description: "OpenAI GPT-5.6 Sol flagship tier of the GPT-5.6 reasoning family; text and image input, 1.05M-token context."
+    },
+    capabilities: {
+      vision: true,
+      tools: true,
+      structuredOutput: true,
+      contextWindow: 105e4
+    }
   }),
-  text("gpt-5.5", "GPT-5.5", "gpt-5.5"),
-  text("gpt-5.6-luna", "GPT-5.6 Luna", "gpt-5.6-luna"),
-  text("gpt-5.6-terra", "GPT-5.6 Terra", "gpt-5.6-terra"),
-  text("gpt-5.6-sol", "GPT-5.6 Sol", "gpt-5.6-sol"),
   // Library key "claude-sonnet-4.5" maps to the SDK id "claude-sonnet-4-5".
   text("claude-sonnet-4-5", "Claude Sonnet 4.5", "claude-sonnet-4-5", {
-    aliases: ["claude-sonnet-4.5", "claude-sonnet-4-5-20250929"]
+    aliases: ["claude-sonnet-4.5", "claude-sonnet-4-5-20250929"],
+    display: {
+      provider: "anthropic",
+      description: "Anthropic Claude Sonnet 4.5 (legacy) balanced model; text and image input, 200K-token context."
+    },
+    capabilities: { vision: true, tools: true, structuredOutput: true, contextWindow: 2e5 }
   }),
   // Anthropic retired the dated 4.0 ids; callers that still send them get 4.6.
   text("claude-sonnet-4-6", "Claude Sonnet 4.6", "claude-sonnet-4-6", {
@@ -625,42 +843,280 @@ var TEXT_MODELS = [
       "claude-sonnet-4",
       "claude-sonnet-4.0",
       "claude-sonnet-4-20250514"
-    ]
+    ],
+    display: {
+      provider: "anthropic",
+      description: "Anthropic Claude Sonnet 4.6 (legacy) balanced model; text and image input, 1M-token context."
+    },
+    capabilities: {
+      vision: true,
+      tools: true,
+      structuredOutput: true,
+      contextWindow: 1e6
+    }
+  }),
+  text("claude-sonnet-5", "Claude Sonnet 5", "claude-sonnet-5", {
+    display: {
+      provider: "anthropic",
+      description: "Anthropic Claude Sonnet 5 current balanced model; text and image input, 1M-token context."
+    },
+    capabilities: {
+      vision: true,
+      tools: true,
+      structuredOutput: true,
+      contextWindow: 1e6
+    }
   }),
-  text("claude-sonnet-5", "Claude Sonnet 5", "claude-sonnet-5"),
   text("claude-opus-4-8", "Claude Opus 4.8", "claude-opus-4-8", {
-    aliases: ["claude-opus-4.8", "claude-opus-4", "claude-opus-4.0", "claude-opus-4-20250514"]
+    aliases: ["claude-opus-4.8", "claude-opus-4", "claude-opus-4.0", "claude-opus-4-20250514"],
+    display: {
+      provider: "anthropic",
+      description: "Anthropic Claude Opus 4.8 (legacy) frontier model; text and image input, 1M-token context."
+    },
+    capabilities: {
+      vision: true,
+      tools: true,
+      structuredOutput: true,
+      contextWindow: 1e6
+    }
   }),
   text("claude-haiku-4-5", "Claude Haiku 4.5", "claude-haiku-4-5", {
-    aliases: ["claude-haiku-4.5", "claude-haiku-4-5-20251001", "claude-haiku-4"]
+    aliases: ["claude-haiku-4.5", "claude-haiku-4-5-20251001", "claude-haiku-4"],
+    display: {
+      provider: "anthropic",
+      description: "Anthropic Claude Haiku 4.5 fast model with near-frontier intelligence; text and image input, 200K-token context."
+    },
+    capabilities: { vision: true, tools: true, structuredOutput: true, contextWindow: 2e5 }
   }),
   text("claude-opus-5", "Claude Opus 5", "claude-opus-5", {
-    aliases: ["claude-opus-5.0"]
+    aliases: ["claude-opus-5.0"],
+    display: {
+      provider: "anthropic",
+      description: "Anthropic Claude Opus 5 current frontier model for agentic coding and enterprise work; text and image input, 1M-token context."
+    },
+    capabilities: {
+      vision: true,
+      tools: true,
+      structuredOutput: true,
+      contextWindow: 1e6
+    }
+  }),
+  text("claude-fable-5", "Claude Fable 5", "claude-fable-5", {
+    display: {
+      provider: "anthropic",
+      description: "Anthropic Claude Fable 5 (legacy) model for demanding reasoning and long-horizon agentic work; text and image input, 1M-token context."
+    },
+    capabilities: {
+      vision: true,
+      tools: true,
+      structuredOutput: true,
+      contextWindow: 1e6
+    }
+  }),
+  text("gemini-2.5-flash", "Gemini 2.5 Flash", "gemini-2.5-flash", {
+    display: {
+      provider: "google",
+      description: "Google Gemini 2.5 Flash low-latency multimodal model for high-volume reasoning over text, images, video, and audio; 1M-token context."
+    },
+    capabilities: {
+      vision: true,
+      tools: true,
+      structuredOutput: true,
+      contextWindow: 1048576
+    }
+  }),
+  text("gemini-2.5-pro", "Gemini 2.5 Pro", "gemini-2.5-pro", {
+    display: {
+      provider: "google",
+      description: "Google Gemini 2.5 Pro thinking model for complex reasoning over code, math, and documents; text, image, video, audio, and PDF input, 1M-token context."
+    },
+    capabilities: {
+      vision: true,
+      tools: true,
+      structuredOutput: true,
+      contextWindow: 1048576
+    }
+  }),
+  text("gemini-3.1-pro-preview", "Gemini 3.1 Pro (preview)", "gemini-3.1-pro-preview", {
+    display: {
+      provider: "google",
+      description: "Google Gemini 3.1 Pro preview reasoning model for software engineering, tool use, and reliable agentic workflows; 1M-token context."
+    },
+    capabilities: {
+      vision: true,
+      tools: true,
+      structuredOutput: true,
+      contextWindow: 1048576
+    }
+  }),
+  text("gemini-3.5-flash", "Gemini 3.5 Flash", "gemini-3.5-flash", {
+    display: {
+      provider: "google",
+      description: "Google Gemini 3.5 Flash high-throughput model for frontier-level agentic tasks; text, image, video, audio, and PDF input, 1M-token context."
+    },
+    capabilities: {
+      vision: true,
+      tools: true,
+      structuredOutput: true,
+      contextWindow: 1048576
+    }
+  }),
+  text("gemini-3.6-flash", "Gemini 3.6 Flash", "gemini-3.6-flash", {
+    display: {
+      provider: "google",
+      description: "Google Gemini 3.6 Flash general-purpose multimodal agentic model balancing speed and capability; 1M-token context."
+    },
+    capabilities: {
+      vision: true,
+      tools: true,
+      structuredOutput: true,
+      contextWindow: 1048576
+    }
+  }),
+  text("gemini-3.7-flash", "Gemini 3.7 Flash", "gemini-3.7-flash", {
+    display: {
+      provider: "google",
+      description: "Google Gemini 3.7 Flash natively multimodal reasoning model for coding, agentic workflows, and multi-step execution; 1M-token context."
+    },
+    capabilities: {
+      vision: true,
+      tools: true,
+      structuredOutput: true,
+      contextWindow: 1048576
+    }
+  }),
+  text("gemini-3.8-flash", "Gemini 3.8 Flash", "gemini-3.8-flash", {
+    display: {
+      provider: "google",
+      description: "Google Gemini 3.8 Flash frontier model for long-horizon software engineering, autonomous agents, and enterprise workflows; 1M-token context."
+    },
+    capabilities: {
+      vision: true,
+      tools: true,
+      structuredOutput: true,
+      contextWindow: 1048576
+    }
   }),
-  text("claude-fable-5", "Claude Fable 5", "claude-fable-5"),
-  text("gemini-2.5-flash", "Gemini 2.5 Flash", "gemini-2.5-flash"),
-  text("gemini-2.5-pro", "Gemini 2.5 Pro", "gemini-2.5-pro"),
-  text("gemini-3.1-pro-preview", "Gemini 3.1 Pro (preview)", "gemini-3.1-pro-preview"),
-  text("gemini-3.5-flash", "Gemini 3.5 Flash", "gemini-3.5-flash"),
-  text("gemini-3.6-flash", "Gemini 3.6 Flash", "gemini-3.6-flash"),
-  text("gemini-3.7-flash", "Gemini 3.7 Flash", "gemini-3.7-flash"),
-  text("gemini-3.8-flash", "Gemini 3.8 Flash", "gemini-3.8-flash"),
   // Renamed to gemini-3.6-flash in vibe-shared model-migration.ts.
   text("gemini-3-flash-preview", "Gemini 3 Flash (preview)", "gemini-3-flash-preview", {
     lifecycle: "deprecated",
     replacedBy: "gemini-3.6-flash"
   }),
-  text("gemini-3.1-flash-lite", "Gemini 3.1 Flash Lite", "gemini-3.1-flash-lite"),
-  text("sonar", "Sonar", "sonar"),
-  text("sonar-pro", "Sonar Pro", "sonar-pro"),
-  text("sonar-reasoning-pro", "Sonar Reasoning Pro", "sonar-reasoning-pro"),
-  text("deepseek-v4-flash", "DeepSeek V4 Flash", "deepseek-v4-flash"),
-  text("deepseek-v4-pro", "DeepSeek V4 Pro", "deepseek-v4-pro"),
-  text("grok-4.3", "Grok 4.3", "grok-4.3"),
-  text("kimi-k2.5", "Kimi K2.5", "kimi-k2.5"),
-  text("kimi-k3", "Kimi K3", "kimi-k3"),
+  text("gemini-3.1-flash-lite", "Gemini 3.1 Flash Lite", "gemini-3.1-flash-lite", {
+    display: {
+      provider: "google",
+      description: "Google Gemini 3.1 Flash-Lite low-latency multimodal model for high-volume agentic, translation, and data-extraction tasks; 1M-token context."
+    },
+    capabilities: {
+      vision: true,
+      tools: true,
+      structuredOutput: true,
+      contextWindow: 1048576
+    }
+  }),
+  text("sonar", "Sonar", "sonar", {
+    display: {
+      provider: "perplexity",
+      description: "Perplexity Sonar lightweight web-grounded search model returning cited answers; 128K-token context."
+    },
+    capabilities: {
+      vision: true,
+      tools: false,
+      structuredOutput: true,
+      contextWindow: 128e3
+    }
+  }),
+  text("sonar-pro", "Sonar Pro", "sonar-pro", {
+    display: {
+      provider: "perplexity",
+      description: "Perplexity Sonar Pro advanced web-grounded search model for complex multi-step queries; 200K-token context."
+    },
+    capabilities: {
+      vision: true,
+      tools: false,
+      structuredOutput: true,
+      contextWindow: 2e5
+    }
+  }),
+  text("sonar-reasoning-pro", "Sonar Reasoning Pro", "sonar-reasoning-pro", {
+    display: {
+      provider: "perplexity",
+      description: "Perplexity Sonar Reasoning Pro chain-of-thought web-grounded search model for complex problem-solving; 128K-token context."
+    },
+    capabilities: {
+      vision: true,
+      tools: false,
+      structuredOutput: true,
+      contextWindow: 128e3
+    }
+  }),
+  text("deepseek-v4-flash", "DeepSeek V4 Flash", "deepseek-v4-flash", {
+    display: {
+      provider: "deepseek",
+      description: "DeepSeek V4 Flash thinking/non-thinking model with vision, tool calls, and JSON output; 1M-token context and up to 384K output tokens."
+    },
+    capabilities: {
+      vision: true,
+      tools: true,
+      structuredOutput: true,
+      contextWindow: 1e6
+    }
+  }),
+  text("deepseek-v4-pro", "DeepSeek V4 Pro", "deepseek-v4-pro", {
+    display: {
+      provider: "deepseek",
+      description: "DeepSeek V4 Pro thinking/non-thinking text model with tool calls and JSON output (no vision); 1M-token context and up to 384K output tokens."
+    },
+    capabilities: {
+      vision: false,
+      tools: true,
+      structuredOutput: true,
+      contextWindow: 1e6
+    }
+  }),
+  // 1M-token context: https://docs.x.ai/docs/models/grok-4.3
+  text("grok-4.3", "Grok 4.3", "grok-4.3", {
+    display: {
+      provider: "xai",
+      description: "xAI Grok 4.3 reasoning model with tool calling and structured outputs; text and image input, 1M-token context."
+    },
+    capabilities: {
+      vision: true,
+      tools: true,
+      structuredOutput: true,
+      contextWindow: 1e6
+    }
+  }),
+  // Moonshot retired kimi-k2.5 on 2026-08-31 and no longer publishes its specs.
+  text("kimi-k2.5", "Kimi K2.5", "kimi-k2.5", {
+    display: {
+      provider: "moonshot",
+      description: "Moonshot Kimi K2.5; the provider retired the model on 2026-08-31 and calls now return a model-not-found error."
+    },
+    capabilities: {}
+  }),
+  text("kimi-k3", "Kimi K3", "kimi-k3", {
+    display: {
+      provider: "moonshot",
+      description: "Moonshot Kimi K3 flagship model with native visual understanding and always-on thinking; text, image, and video input, 1M-token context."
+    },
+    capabilities: {
+      vision: true,
+      tools: true,
+      structuredOutput: true,
+      contextWindow: 1e6
+    }
+  }),
   // Cowork calls the provider "MiniMax-M3"; the pricing key is lowercase.
-  text("minimax-m3", "MiniMax M3", "MiniMax-M3", { aliases: ["MiniMax-M3"] })
+  // 1M-token context: https://platform.minimax.io/docs/api-reference/text-openai-api
+  text("minimax-m3", "MiniMax M3", "MiniMax-M3", {
+    aliases: ["MiniMax-M3"],
+    display: {
+      provider: "minimax",
+      description: "MiniMax M3 frontier multimodal model for agentic reasoning, tool use, and coding; text, image, and video input, 1M-token context."
+    },
+    capabilities: { vision: true, tools: true, contextWindow: 1e6 }
+  })
 ];
 
 // src/catalog/video.ts
@@ -672,6 +1128,8 @@ var GROK_RATIOS = ["1:1", "16:9", "9:16", "4:3", "3:4", "3:2", "2:3"];
 var H3_RATIOS = ["21:9", "16:9", "4:3", "1:1", "3:4", "9:16"];
 var H3_REFERENCE_RATIOS = ["adaptive", ...H3_RATIOS];
 var NEGATIVE_PROMPT = "blur, distort, and low quality";
+var FRAMES_EXCLUDE_REFERENCES = { framesExcludeReferences: true };
+var GOOGLE_LOGO = "https://d31vvfpk2ppdyq.cloudfront.net/assets/image/192485d2-a76a-4657-9c06-930b2c70badd_compressed.png";
 var VIDEO_MODELS = [
   {
     id: "veo-3.1",
@@ -695,7 +1153,13 @@ var VIDEO_MODELS = [
     },
     defaults: { durationSeconds: 8, resolution: "720p", aspectRatio: "16:9" },
     usage: { ...VIDEO_USAGE },
-    pricingKind: "video_per_second"
+    pricingKind: "video_per_second",
+    display: {
+      provider: "google",
+      description: "Google's highest-quality video generation model. Supports up to 4K resolution and 8-second clips with reference image conditioning.",
+      logo: GOOGLE_LOGO
+    },
+    capabilities: { ...FRAMES_EXCLUDE_REFERENCES }
   },
   {
     id: "veo-3.1-fast",
@@ -718,7 +1182,12 @@ var VIDEO_MODELS = [
     },
     defaults: { durationSeconds: 8, resolution: "720p", aspectRatio: "16:9" },
     usage: { ...VIDEO_USAGE },
-    pricingKind: "video_per_second"
+    pricingKind: "video_per_second",
+    display: {
+      provider: "google",
+      description: "Faster variant of Google Veo 3.1; text- and image-to-video with natively generated audio at 4 or 8 seconds."
+    },
+    capabilities: { ...FRAMES_EXCLUDE_REFERENCES }
   },
   {
     id: "veo-3.1-lite",
@@ -741,7 +1210,12 @@ var VIDEO_MODELS = [
     },
     defaults: { durationSeconds: 8, resolution: "720p", aspectRatio: "16:9" },
     usage: { ...VIDEO_USAGE },
-    pricingKind: "video_per_second"
+    pricingKind: "video_per_second",
+    display: {
+      provider: "google",
+      description: "Efficiency-focused Google Veo 3.1 Lite for high-volume video generation and editing; text and image input, with no 4K output or video extension."
+    },
+    capabilities: { ...FRAMES_EXCLUDE_REFERENCES }
   },
   {
     id: "gemini-omni-flash",
@@ -768,7 +1242,13 @@ var VIDEO_MODELS = [
     },
     defaults: {},
     usage: { ...VIDEO_USAGE },
-    pricingKind: "video_per_second"
+    pricingKind: "video_per_second",
+    display: {
+      provider: "google",
+      description: "Google's conversational video model with native audio, human support, and natural language editing. Aspect ratio only \u2014 resolution and duration are model-determined.",
+      logo: GOOGLE_LOGO
+    },
+    capabilities: { acceptsSourceVideo: true }
   },
   {
     id: "gemini-omni-1.1-flash",
@@ -787,7 +1267,12 @@ var VIDEO_MODELS = [
     },
     defaults: {},
     usage: { ...VIDEO_USAGE },
-    pricingKind: "video_per_second"
+    pricingKind: "video_per_second",
+    display: {
+      provider: "google",
+      description: "Google Gemini Omni Flash 1.1 conversational video generation and editing; text, image, and up to 10-second video input, 3-10 second output at up to 4K and 24 FPS."
+    },
+    capabilities: { acceptsSourceVideo: true }
   },
   {
     // vibe-triggers and Space Builder drop the ratio when a frame is wired; Cowork always sends it.
@@ -804,7 +1289,7 @@ var VIDEO_MODELS = [
       durations: [4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15],
       references: { images: 9, videos: 3, audios: 3, audioOnly: false },
       ratioIgnoredWith: ["firstFrame"],
-      supports: ["generateAudio", "seed", "lastFrame"]
+      supports: ["generateAudio", "seed", "lastFrame", "watermark"]
     },
     defaults: {
       durationSeconds: 5,
@@ -813,7 +1298,13 @@ var VIDEO_MODELS = [
       generateAudio: true
     },
     usage: { ...VIDEO_USAGE },
-    pricingKind: "video_per_second"
+    pricingKind: "video_per_second",
+    display: {
+      provider: "bytedance",
+      description: "High-quality AI video generation with reference image conditioning. Supports 4-15 second clips with human preprocessing.",
+      logo: "https://d31vvfpk2ppdyq.cloudfront.net/assets/image/3bba5d81-4e57-4e71-911b-cdfa05acca2c_compressed.png"
+    },
+    capabilities: { acceptsSourceVideo: true, ...FRAMES_EXCLUDE_REFERENCES }
   },
   {
     id: "seedance-2.0-fast",
@@ -829,7 +1320,7 @@ var VIDEO_MODELS = [
       durations: [4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15],
       references: { images: 9, videos: 3, audios: 3, audioOnly: false },
       ratioIgnoredWith: ["firstFrame"],
-      supports: ["generateAudio", "seed", "lastFrame"]
+      supports: ["generateAudio", "seed", "lastFrame", "watermark"]
     },
     defaults: {
       durationSeconds: 5,
@@ -838,7 +1329,12 @@ var VIDEO_MODELS = [
       generateAudio: true
     },
     usage: { ...VIDEO_USAGE },
-    pricingKind: "video_per_second"
+    pricingKind: "video_per_second",
+    display: {
+      provider: "bytedance",
+      description: "Speed- and cost-optimized ByteDance Seedance 2.0 Fast; text plus image, video, and audio references to generate or edit 4-15 second clips."
+    },
+    capabilities: { acceptsSourceVideo: true, ...FRAMES_EXCLUDE_REFERENCES }
   },
   {
     // generateAudio defaults to false here (trigger-workflows); Space Builder defaults it to true.
@@ -855,7 +1351,7 @@ var VIDEO_MODELS = [
       durations: { min: 4, max: 30 },
       references: { images: 30, videos: 10, audios: 10 },
       ratioIgnoredWith: ["firstFrame"],
-      supports: ["generateAudio", "seed", "lastFrame"]
+      supports: ["generateAudio", "seed", "lastFrame", "watermark"]
     },
     defaults: {
       durationSeconds: 5,
@@ -864,7 +1360,12 @@ var VIDEO_MODELS = [
       generateAudio: false
     },
     usage: { ...VIDEO_USAGE },
-    pricingKind: "video_per_second"
+    pricingKind: "video_per_second",
+    display: {
+      provider: "bytedance",
+      description: "ByteDance Seedance 2.5 for long-form storytelling and omni-reference generation/editing; up to 50 reference assets and 4-30 second clips at up to 1080p."
+    },
+    capabilities: { acceptsSourceVideo: true, ...FRAMES_EXCLUDE_REFERENCES }
   },
   {
     // Cowork allows 7 reference images and durations 4/6/8/10; Space Builder allows 5 and 1-15s.
@@ -882,9 +1383,16 @@ var VIDEO_MODELS = [
       references: { images: 7 },
       supports: []
     },
-    defaults: { durationSeconds: 8, resolution: "720p", aspectRatio: "16:9" },
+    // No resolution default: xAI picks its own unless the caller asks.
+    defaults: { durationSeconds: 8, aspectRatio: "16:9" },
     usage: { unit: "reported_cost_usd", modality: "VIDEO" },
-    pricingKind: "video_per_second"
+    pricingKind: "video_per_second",
+    display: {
+      provider: "xai",
+      description: "xAI's reference-to-video model. Generates short clips conditioned on up to 7 reference images with flexible aspect ratios.",
+      logo: "https://d31vvfpk2ppdyq.cloudfront.net/assets/image/858b7fcc-9bd5-411d-aea1-a7b35ecfffab_compressed.png"
+    },
+    capabilities: { ...FRAMES_EXCLUDE_REFERENCES }
   },
   {
     id: "kling-2.5",
@@ -907,7 +1415,12 @@ var VIDEO_MODELS = [
       negativePrompt: NEGATIVE_PROMPT
     },
     usage: { ...VIDEO_USAGE },
-    pricingKind: "video_per_second"
+    pricingKind: "video_per_second",
+    display: {
+      provider: "kuaishou",
+      description: "Kuaishou Kling 2.5 Turbo Pro image-to-video; animates a single input image from a text prompt with an optional end image, 5- or 10-second clips."
+    },
+    capabilities: { ...FRAMES_EXCLUDE_REFERENCES }
   },
   {
     // trigger-workflows validates 5/10 only; Space Builder allows 3-15s (Kling30Input).
@@ -931,7 +1444,12 @@ var VIDEO_MODELS = [
       negativePrompt: NEGATIVE_PROMPT
     },
     usage: { ...VIDEO_USAGE },
-    pricingKind: "video_per_second"
+    pricingKind: "video_per_second",
+    display: {
+      provider: "kuaishou",
+      description: "Kuaishou Kling 3.0 Pro image-to-video; a required start image plus a text prompt (or multi-shot prompts) with optional end image and native audio, 3-15 second clips."
+    },
+    capabilities: { ...FRAMES_EXCLUDE_REFERENCES }
   },
   {
     // mode standard/pro selects 720p/1080p; the endpoint takes no separate resolution.
@@ -962,7 +1480,12 @@ var VIDEO_MODELS = [
     },
     defaults: { durationSeconds: 5, aspectRatio: "16:9", generateAudio: false, mode: "pro" },
     usage: { unit: "compute_seconds", modality: "VIDEO" },
-    pricingKind: "video_per_second"
+    pricingKind: "video_per_second",
+    display: {
+      provider: "kuaishou",
+      description: "Kuaishou Kling Video 3.0 Omni unified generation and editing from text, start/end images, up to 7 reference images, and reference video; up to 15 seconds at up to 1080p."
+    },
+    capabilities: {}
   },
   {
     id: "minimax-h3",
@@ -976,7 +1499,7 @@ var VIDEO_MODELS = [
       aspectRatios: [...H3_RATIOS],
       resolutions: ["480p", "768p", "2k", "4k"],
       durations: { min: 5, max: 15 },
-      supports: ["seed", "promptExpansion"]
+      supports: ["seed", "promptExpansion", "lastFrame", "safetyChecker"]
     },
     defaults: {
       durationSeconds: 5,
@@ -985,7 +1508,12 @@ var VIDEO_MODELS = [
       promptExpansion: "balanced"
     },
     usage: { ...VIDEO_USAGE },
-    pricingKind: "video_per_second"
+    pricingKind: "video_per_second",
+    display: {
+      provider: "minimax",
+      description: "MiniMax H3 multimodal video model for text-to-video, image-to-video (first/last frame), and reference-to-video with native audio, up to 2K."
+    },
+    capabilities: { ...FRAMES_EXCLUDE_REFERENCES }
   },
   {
     id: "minimax-h3-max",
@@ -999,7 +1527,7 @@ var VIDEO_MODELS = [
       aspectRatios: [...H3_RATIOS],
       resolutions: ["480p", "768p"],
       durations: { min: 5, max: 15 },
-      supports: ["seed", "promptExpansion"]
+      supports: ["seed", "promptExpansion", "lastFrame", "safetyChecker"]
     },
     defaults: {
       durationSeconds: 5,
@@ -1008,7 +1536,12 @@ var VIDEO_MODELS = [
       promptExpansion: "balanced"
     },
     usage: { ...VIDEO_USAGE },
-    pricingKind: "video_per_second"
+    pricingKind: "video_per_second",
+    display: {
+      provider: "minimax",
+      description: "High-speed MiniMax H3 Max variant for text-to-video and image-to-video (first/last frame) only, up to 768p."
+    },
+    capabilities: { ...FRAMES_EXCLUDE_REFERENCES }
   },
   {
     id: "minimax-h3-reference",
@@ -1023,7 +1556,7 @@ var VIDEO_MODELS = [
       resolutions: ["480p", "768p", "2k", "4k"],
       durations: { min: 5, max: 15 },
       references: { images: 9, videos: 3, audios: 3, total: 12, audioOnly: false },
-      supports: ["seed", "promptExpansion"]
+      supports: ["seed", "promptExpansion", "safetyChecker"]
     },
     defaults: {
       durationSeconds: 5,
@@ -1032,7 +1565,12 @@ var VIDEO_MODELS = [
       promptExpansion: "balanced"
     },
     usage: { ...VIDEO_USAGE },
-    pricingKind: "video_per_second"
+    pricingKind: "video_per_second",
+    display: {
+      provider: "minimax",
+      description: "MiniMax H3 reference-to-video mode guided by up to 9 reference images, 3 reference videos, and 3 audio clips; reference and image-to-video inputs cannot be mixed."
+    },
+    capabilities: {}
   },
   {
     id: "minimax-h3-max-reference",
@@ -1047,7 +1585,7 @@ var VIDEO_MODELS = [
       resolutions: ["480p", "768p"],
       durations: { min: 5, max: 15 },
       references: { images: 9, videos: 3, audios: 3, total: 12, audioOnly: false },
-      supports: ["seed", "promptExpansion"]
+      supports: ["seed", "promptExpansion", "safetyChecker"]
     },
     defaults: {
       durationSeconds: 5,
@@ -1056,7 +1594,12 @@ var VIDEO_MODELS = [
       promptExpansion: "balanced"
     },
     usage: { ...VIDEO_USAGE },
-    pricingKind: "video_per_second"
+    pricingKind: "video_per_second",
+    display: {
+      provider: "minimax",
+      description: "MiniMax H3 Max reference-to-video on fal; conditions on reference images, videos, and audio. MiniMax's own docs do not list a reference mode for H3 Max."
+    },
+    capabilities: {}
   },
   {
     // Duration 3-10s per trigger-workflows mapDuration; the endpoint takes no resolution.
@@ -1075,7 +1618,12 @@ var VIDEO_MODELS = [
     },
     defaults: { durationSeconds: 5, aspectRatio: "16:9" },
     usage: { ...VIDEO_USAGE },
-    pricingKind: "video_per_second"
+    pricingKind: "video_per_second",
+    display: {
+      provider: "kuaishou",
+      description: "Kuaishou Kling O1 Reference Video-to-Video; a prompt plus a 3-10 second reference video (optional reference images) generates 3-10 second clips preserving motion and camera language."
+    },
+    capabilities: { acceptsSourceVideo: true }
   },
   {
     id: "hailuo-02",
@@ -1092,7 +1640,12 @@ var VIDEO_MODELS = [
     },
     defaults: { durationSeconds: 6, resolution: "768p" },
     usage: { unit: "compute_seconds", modality: "VIDEO" },
-    pricingKind: "video_per_second"
+    pricingKind: "video_per_second",
+    display: {
+      provider: "minimax",
+      description: "MiniMax Hailuo 02 text-to-video and image-to-video model with realistic physics and strong prompt following; outputs 768p at 6 or 10 seconds."
+    },
+    capabilities: {}
   },
   {
     id: "higgsfield-dop-preview",
@@ -1108,7 +1661,12 @@ var VIDEO_MODELS = [
     },
     defaults: {},
     usage: { ...VIDEO_USAGE },
-    pricingKind: "video_per_second"
+    pricingKind: "video_per_second",
+    display: {
+      provider: "higgsfield",
+      description: "Higgsfield DOP image-to-video model; animates a required first-frame image from a text prompt with an optional last-frame image and motion controls."
+    },
+    capabilities: {}
   }
 ];
 
@@ -1175,19 +1733,22 @@ function resolveTier(field, code, requested, supported, order, model, onUnsuppor
   });
   return to;
 }
-function resolveDuration(requested, durations, model, warnings) {
+function resolveDuration(requested, durations, model, onUnsupported, warnings) {
   if (requested === void 0 || !durations) return requested;
-  let to = requested;
-  if (Array.isArray(durations)) {
-    if (durations.includes(requested)) return requested;
-    to = durations.reduce(
-      (best, d) => Math.abs(d - requested) < Math.abs(best - requested) ? d : best,
-      durations[0] ?? requested
-    );
-  } else {
-    to = Math.min(Math.max(requested, durations.min), durations.max);
-    if (to === requested) return requested;
+  const inSet = Array.isArray(durations) ? durations.includes(requested) : requested >= durations.min && requested <= durations.max;
+  if (inSet) return requested;
+  if (onUnsupported === "error") {
+    const supported = Array.isArray(durations) ? `Supported: ${durations.join(", ")}.` : `Supported range: ${durations.min}-${durations.max}.`;
+    throw new ModelKitError({
+      code: "unsupported_option",
+      message: `${model.id} does not support durationSeconds ${requested}. ${supported}`,
+      model: model.id
+    });
   }
+  const to = Array.isArray(durations) ? durations.reduce(
+    (best, d) => Math.abs(d - requested) < Math.abs(best - requested) ? d : best,
+    durations[0] ?? requested
+  ) : Math.min(Math.max(requested, durations.min), durations.max);
   warnings.push({
     code: "duration_coerced",
     field: "durationSeconds",
@@ -1197,6 +1758,26 @@ function resolveDuration(requested, durations, model, warnings) {
   });
   return to;
 }
+function resolveAspectRatio(requested, supported, model, onUnsupported, warnings) {
+  if (requested === void 0 || requested === "adaptive" || !supported) return requested;
+  if (supported.includes(requested)) return requested;
+  if (onUnsupported === "error") {
+    throw new ModelKitError({
+      code: "unsupported_option",
+      message: `${model.id} does not support aspectRatio "${requested}". Supported: ${supported.join(", ")}.`,
+      model: model.id
+    });
+  }
+  const to = nearestAspectRatio(requested, supported);
+  warnings.push({
+    code: "ratio_coerced",
+    field: "aspectRatio",
+    from: requested,
+    to,
+    message: `aspect ratio ${requested} coerced to ${to} for ${model.id}`
+  });
+  return to;
+}
 function compact(refs) {
   return (refs ?? []).filter((r) => Boolean(r));
 }
@@ -1215,17 +1796,14 @@ function resolveImageOptions(model, request, onUnsupported = "error") {
   const warnings = [];
   const c = model.constraints;
   const merged = { ...model.defaults, ...stripUndefined(request), model: model.id };
-  if (merged.aspectRatio !== void 0 && merged.aspectRatio !== "adaptive" && c.aspectRatios !== void 0 && !c.aspectRatios.includes(merged.aspectRatio)) {
-    const to = nearestAspectRatio(merged.aspectRatio, c.aspectRatios);
-    warnings.push({
-      code: "ratio_coerced",
-      field: "aspectRatio",
-      from: merged.aspectRatio,
-      to,
-      message: `aspect ratio ${merged.aspectRatio} coerced to ${to} for ${model.id}`
-    });
-    merged.aspectRatio = to;
-  }
+  const aspectRatio = resolveAspectRatio(
+    merged.aspectRatio,
+    c.aspectRatios,
+    model,
+    onUnsupported,
+    warnings
+  );
+  if (aspectRatio !== void 0) merged.aspectRatio = aspectRatio;
   const size = resolveTier(
     "size",
     "size_coerced",
@@ -1259,7 +1837,9 @@ function resolveImageOptions(model, request, onUnsupported = "error") {
     "count",
     "outputFormat",
     "scaleFactor",
-    "safetySettings"
+    "safetySettings",
+    "temperature",
+    "upscale"
   ]);
   return { request: merged, warnings };
 }
@@ -1289,17 +1869,14 @@ function resolveVideoOptions(model, request, onUnsupported = "error") {
       delete merged.aspectRatio;
     }
   }
-  if (merged.aspectRatio !== void 0 && merged.aspectRatio !== "adaptive" && c.aspectRatios !== void 0 && !c.aspectRatios.includes(merged.aspectRatio)) {
-    const to = nearestAspectRatio(merged.aspectRatio, c.aspectRatios);
-    warnings.push({
-      code: "ratio_coerced",
-      field: "aspectRatio",
-      from: merged.aspectRatio,
-      to,
-      message: `aspect ratio ${merged.aspectRatio} coerced to ${to} for ${model.id}`
-    });
-    merged.aspectRatio = to;
-  }
+  const aspectRatio = resolveAspectRatio(
+    merged.aspectRatio,
+    c.aspectRatios,
+    model,
+    onUnsupported,
+    warnings
+  );
+  if (aspectRatio !== void 0) merged.aspectRatio = aspectRatio;
   const resolution = resolveTier(
     "resolution",
     "resolution_coerced",
@@ -1311,7 +1888,13 @@ function resolveVideoOptions(model, request, onUnsupported = "error") {
     warnings
   );
   if (resolution !== void 0) merged.resolution = resolution;
-  const duration = resolveDuration(merged.durationSeconds, c.durations, model, warnings);
+  const duration = resolveDuration(
+    merged.durationSeconds,
+    c.durations,
+    model,
+    onUnsupported,
+    warnings
+  );
   if (duration !== void 0) merged.durationSeconds = duration;
   const limits = c.references ?? {};
   let images = trim(
@@ -1366,10 +1949,45 @@ function resolveVideoOptions(model, request, onUnsupported = "error") {
     "cfgScale",
     "generateAudio",
     "promptExpansion",
-    "lastFrame"
+    "lastFrame",
+    "watermark",
+    "safetyChecker"
   ]);
   return { request: merged, warnings };
 }
+function resolveAudioOptions(model, request, onUnsupported = "error") {
+  const warnings = [];
+  const c = model.constraints;
+  const merged = { ...model.defaults, ...stripUndefined(request), model: model.id };
+  for (const input of c.requires ?? []) {
+    const present = input === "references" ? compact(merged.references).length > 0 : input === "sourceVideo" && merged.sourceVideo !== void 0;
+    if (!present)
+      throw new ModelKitError({
+        code: "missing_input",
+        message: `${model.id} requires ${input}`,
+        model: model.id
+      });
+  }
+  const duration = resolveDuration(
+    merged.durationSeconds,
+    c.durations,
+    model,
+    onUnsupported,
+    warnings
+  );
+  if (duration !== void 0) merged.durationSeconds = duration;
+  const references = trim(
+    "references",
+    compact(merged.references),
+    c.references?.images ?? c.references?.total,
+    model,
+    warnings
+  );
+  if (references.length > 0) merged.references = references;
+  else delete merged.references;
+  dropUnsupported(merged, model, warnings, ["seed", "negativePrompt", "instrumental"]);
+  return { request: merged, warnings };
+}
 function dropUnsupported(req, model, warnings, fields) {
   const supports = model.constraints.supports ?? [];
   for (const field of fields) {
@@ -1411,6 +2029,9 @@ function hasAll(keys, names) {
     (n) => Boolean(keys[n]) && (n !== "higgsfield" || Boolean(keys.higgsfieldSecret))
   );
 }
+function missingEnterpriseKeys(config, names) {
+  return names.filter((name) => !hasAll(config.keys.enterprise, [name]));
+}
 function resolveKeys(config, names, tier = "pro") {
   if (tier === "enterprise" && hasAll(config.keys.enterprise, names))
     return config.keys.enterprise;
@@ -1443,7 +2064,6 @@ function keysFromEnv(env, prefix = "") {
   set("prodia", pick("PRODIA_API_KEY", "PRODIA_TOKEN"));
   set("higgsfield", pick("HIGGSFIELD_API_KEY"));
   set("higgsfieldSecret", pick("HIGGSFIELD_API_SECRET"));
-  set("openrouter", pick("OPENROUTER_API_KEY"));
   set("cloudflare", pick("CLOUDFLARE_AI_GATEWAY_TOKEN", "CF_AIG_TOKEN"));
   return keys;
 }
@@ -1491,7 +2111,9 @@ var textBody = import_zod2.z.string();
 function classifyHttpError(status, body, provider, model) {
   const asText = textBody.safeParse(body);
   const text2 = asText.success ? asText.data : JSON.stringify(body);
-  const blocked = /content_policy|safety|blockReason|PROHIBITED_CONTENT|moderation/i.test(text2);
+  const blocked = /content_policy_violation|blockReason|PROHIBITED_CONTENT|moderation_blocked|safety(?![_A-Za-z])/i.test(
+    text2
+  );
   let code = "provider_rejected";
   if (blocked) code = "content_blocked";
   else if (status === 401 || status === 403) code = "auth";
@@ -1512,8 +2134,8 @@ function classifyHttpError(status, body, provider, model) {
 }
 function shouldRetry(err) {
   if (err instanceof ModelKitError) return err.retryable;
-  if (err.name === "AbortError" || /ECONNRESET|ETIMEDOUT|fetch failed|socket hang up/i.test(err.message))
-    return true;
+  if (err.name === "AbortError") return false;
+  if (/ECONNRESET|ETIMEDOUT|fetch failed|socket hang up/i.test(err.message)) return true;
   return false;
 }
 async function withRetry(fn, opts = {}) {
@@ -1698,6 +2320,12 @@ var import_zod4 = require("zod");
 function isUrlRef(ref) {
   return "url" in ref;
 }
+function abortedError() {
+  return new ModelKitError({ code: "timeout", message: "aborted", retryable: false });
+}
+function throwIfAborted(signal) {
+  if (signal?.aborted) throw abortedError();
+}
 function mediaRefToBase64(bytes) {
   let binary = "";
   const CHUNK = 32768;
@@ -1712,10 +2340,34 @@ function base64ToBytes(b64) {
   for (let i = 0; i < binary.length; i++) out[i] = binary.charCodeAt(i);
   return out;
 }
-async function mediaRefToBytes(ref, fetch) {
+function mediaRefToDataUrl(ref) {
+  if (isUrlRef(ref)) return ref.url;
+  return `data:${ref.mime};base64,${mediaRefToBase64(ref.bytes)}`;
+}
+var refToUrl = mediaRefToDataUrl;
+async function readBody(res) {
+  const text2 = await res.text().catch(() => "");
+  try {
+    const decoded = JsonValueSchema.safeParse(JSON.parse(text2));
+    return decoded.success ? decoded.data : text2;
+  } catch {
+    return text2;
+  }
+}
+async function mediaRefToBytes(ref, fetch, options = {}) {
   if (!isUrlRef(ref)) return { bytes: ref.bytes, mime: ref.mime };
-  const res = await fetch(ref.url);
-  if (!res.ok) throw new Error(`reference fetch failed (${res.status}): ${ref.url}`);
+  const init = {};
+  if (options.signal !== void 0) init.signal = options.signal;
+  const res = await fetch(ref.url, init);
+  if (!res.ok) {
+    const text2 = await res.text().catch(() => "");
+    throw classifyHttpError(
+      res.status,
+      text2,
+      options.provider ?? "media-reference",
+      options.model
+    );
+  }
   const mime = res.headers.get("content-type")?.split(";")[0]?.trim().toLowerCase() || "image/png";
   return { bytes: new Uint8Array(await res.arrayBuffer()), mime };
 }
@@ -1773,16 +2425,6 @@ var GATEWAY_SLUGS = {
   perplexity: "perplexity-ai",
   deepseek: "deepseek"
 };
-var OPENROUTER_VENDORS = {
-  openai: "openai",
-  anthropic: "anthropic",
-  google: "google",
-  xai: "x-ai",
-  perplexity: "perplexity",
-  deepseek: "deepseek",
-  moonshot: "moonshotai",
-  minimax: "minimax"
-};
 function detectTextProvider(modelId) {
   const id = modelId.toLowerCase();
   if (id.startsWith("claude") || id.includes("anthropic.")) return "anthropic";
@@ -1799,14 +2441,17 @@ function detectTextProvider(modelId) {
     message: `ai-sdk-text cannot detect a provider for "${modelId}". Set providerOptions.provider to one of ${TEXT_PROVIDERS.join(", ")}.`
   });
 }
-function textKeyName(model, routing) {
-  if (routing.text === "openrouter") return "openrouter";
+function textKeyName(model, _routing) {
   return textProviderOf(model);
 }
 function textProviderOf(model) {
   const explicit = model.providerOptions?.["modelkit"]?.["provider"];
   return isTextProvider(explicit) ? explicit : detectTextProvider(model.providerModelId);
 }
+function textUsesCloudflareGateway(model, routing) {
+  if (routing.text !== "cloudflare") return false;
+  return GATEWAY_SLUGS[textProviderOf(model)] !== void 0;
+}
 function nativeClientFor(provider) {
   switch (provider) {
     case "openai":
@@ -1839,16 +2484,6 @@ function resolveTextRoute(ctx) {
       retryable: false
     });
   }
-  if (mode === "openrouter") {
-    const modelId = ctx.model.providerModelId.includes("/") ? ctx.model.providerModelId : `${OPENROUTER_VENDORS[provider]}/${ctx.model.providerModelId}`;
-    return {
-      provider,
-      client: "openai-compatible",
-      modelId,
-      keyName: "openrouter",
-      baseURL: "https://openrouter.ai/api/v1"
-    };
-  }
   const compatBase = OPENAI_COMPATIBLE_BASE_URLS[provider];
   const route2 = {
     provider,
@@ -1916,7 +2551,7 @@ var aiSdkTextAdapter = {
       case "openai-compatible":
         return (0, import_openai.createOpenAI)({
           ...settings,
-          name: route2.provider === "openai" ? "openai" : ctx.routing.text === "openrouter" ? "openrouter" : route2.provider
+          name: route2.provider
         }).chat(route2.modelId);
       case "openai":
         return (0, import_openai.createOpenAI)(settings)(route2.modelId);
@@ -1929,8 +2564,9 @@ function requiredKeys(model, config) {
   if (model.adapter === "fal" && config.routing?.fal === "cloudflare-passthrough")
     return ["cloudflare"];
   if (model.adapter === "ai-sdk-text") {
-    const names = [textKeyName(model, config.routing ?? {})];
-    if (config.routing?.text === "cloudflare") names.push("cloudflare");
+    const routing = config.routing ?? {};
+    const names = [textKeyName(model, routing)];
+    if (textUsesCloudflareGateway(model, routing)) names.push("cloudflare");
     return names;
   }
   return ADAPTER_KEYS[model.adapter];
@@ -2004,6 +2640,11 @@ function pushIgnored(warnings, field, from, modelId) {
     message: `${modelId} provider instance does not support ${field}; ignored`
   });
 }
+function withoutModelkit(options) {
+  if (!options) return {};
+  const { modelkit: _modelkit, ...rest } = options;
+  return rest;
+}
 function logCoercions(logger, modelId, warnings) {
   if (warnings.length === 0) return;
   logger.warn(`[modelkit] ${modelId} request coerced`, { model: modelId, warnings });
@@ -2020,7 +2661,16 @@ function createModelKit(config, adapters) {
         message: `No adapter registered for ${model.adapter}`,
         model: model.id
       });
-    const keys = resolveKeys(config, requiredKeys(model, config), call.tier);
+    const names = requiredKeys(model, config);
+    if (call.tier === "enterprise") {
+      const missing = missingEnterpriseKeys(config, names);
+      if (missing.length > 0)
+        logger.warn(
+          `[modelkit] enterprise keys incomplete for ${model.adapter}; using pro`,
+          { missing }
+        );
+    }
+    const keys = resolveKeys(config, names, call.tier);
     const ctx = {
       model,
       keys,
@@ -2341,8 +2991,15 @@ function createModelKit(config, adapters) {
           model: model.id
         });
       }
-      const warnings = [];
-      const { adapter, ctx } = context(model, call, warnings);
+      const { abortSignal: _signal, ...request } = req;
+      const { request: resolved, warnings } = resolveAudioOptions(
+        model,
+        request,
+        call.onUnsupported ?? "error"
+      );
+      logCoercions(logger, model.id, warnings);
+      const signal = combineSignals(req.abortSignal, call.signal);
+      const { adapter, ctx } = context(model, signal ? { ...call, signal } : call, warnings);
       if (!adapter.generateAudio) {
         throw new ModelKitError({
           code: "unsupported_option",
@@ -2350,11 +3007,7 @@ function createModelKit(config, adapters) {
           model: model.id
         });
       }
-      const result = await adapter.generateAudio(
-        { ...model.defaults, ...req, model: model.id },
-        ctx
-      );
-      logCoercions(logger, model.id, warnings);
+      const result = await adapter.generateAudio(resolved, ctx);
       await emit({ ...result.usage, model: model.id, unit: model.usage.unit });
       return result;
     },
@@ -2375,7 +3028,7 @@ function createModelKit(config, adapters) {
         });
       return {
         model,
-        providerOptions: entry.providerOptions ?? {},
+        providerOptions: withoutModelkit(entry.providerOptions),
         catalog: entry
       };
     },
@@ -2411,7 +3064,9 @@ function createModelKit(config, adapters) {
         ...textAbortOptions(options.abortSignal, call.signal),
         model: target.languageModel,
         providerOptions: {},
-        prepareStep: steps.wrap(options.prepareStep),
+        prepareStep: steps.wrap(
+          options.prepareStep ?? options.experimental_prepareStep
+        ),
         onFinish: async (event) => {
           for (const usage2 of steps.usage(event.steps)) await emit(withRowUnit(usage2));
           await callerOnFinish?.(event);
@@ -2468,6 +3123,7 @@ var SEEDREAM_SIZES = {
     "3:2": "1248x832",
     "2:3": "832x1248",
     "21:9": "1568x672",
+    "9:21": "672x1568",
     "4:5": "864x1080",
     "5:4": "1080x864"
   },
@@ -2480,16 +3136,34 @@ var SEEDREAM_SIZES = {
     "3:2": "2496x1664",
     "2:3": "1664x2496",
     "21:9": "3136x1344",
+    "9:21": "1344x3136",
     "4:5": "1728x2160",
     "5:4": "2160x1728"
+  },
+  // 4K keeps the long edge at Ark's 4096 cap (a doubled 2K entry would exceed
+  // it); sides are multiples of 16.
+  "4K": {
+    "1:1": "4096x4096",
+    "4:3": "4096x3072",
+    "3:4": "3072x4096",
+    "16:9": "4096x2304",
+    "9:16": "2304x4096",
+    "3:2": "4096x2736",
+    "2:3": "2736x4096",
+    "21:9": "4096x1760",
+    "9:21": "1760x4096",
+    "4:5": "3280x4096",
+    "5:4": "4096x3280"
   }
 };
 var arkAdapter = {
   id: "ark",
   async generateImage(req, ctx) {
+    throwIfAborted(ctx.signal);
     return generateSeedreamImage(req, ctx);
   },
   async generateVideo(req, ctx) {
+    throwIfAborted(ctx.signal);
     return generateSeedanceVideo(req, ctx);
   }
 };
@@ -2509,7 +3183,7 @@ function arkKey(ctx) {
 function throwArkError(status, body, model) {
   const parsed = import_zod5.z.string().safeParse(body);
   const text2 = parsed.success ? parsed.data : JSON.stringify(body ?? "");
-  if (/sensitive.?content|inputimage|human.?face/i.test(text2)) {
+  if (/sensitive.?content|human.?face/i.test(text2)) {
     const init = {
       code: "content_blocked",
       message: `ark rejected the request: ${text2.slice(0, 500)}`,
@@ -2534,7 +3208,7 @@ function throwArkError(status, body, model) {
   throw classifyHttpError(status, body, "ark", model);
 }
 function seedreamSize(aspect, size) {
-  const tier = size === "1K" ? "1K" : "2K";
+  const tier = size === "4K" || size === "1K" ? size : "2K";
   const table = SEEDREAM_SIZES[tier];
   return table[aspect ?? ""] ?? table["1:1"] ?? "2048x2048";
 }
@@ -2545,10 +3219,6 @@ function sniffedImageMime(bytes) {
     return "image/webp";
   return "image/jpeg";
 }
-function arkUrl(ref) {
-  if ("url" in ref) return ref.url;
-  return `data:${ref.mime};base64,${Buffer.from(ref.bytes).toString("base64")}`;
-}
 var SeedreamImageSchema = import_zod5.z.object({
   url: import_zod5.z.string().optional(),
   b64_json: import_zod5.z.string().optional()
@@ -2575,7 +3245,7 @@ async function generateSeedreamImage(req, ctx) {
   const key = arkKey(ctx);
   const modelId = ctx.model.id;
   const size = seedreamSize(req.aspectRatio, req.size);
-  const refs = (req.references ?? []).map(arkUrl);
+  const refs = (req.references ?? []).map(refToUrl);
   const body = {
     model: ctx.model.providerModelId,
     prompt: req.prompt,
@@ -2627,10 +3297,18 @@ async function generateSeedreamImage(req, ctx) {
   if (req.quality !== void 0) usage2.imageQuality = req.quality;
   return { images: [{ bytes, mime: sniffedImageMime(bytes) }], usage: usage2, raw: data };
 }
+var SeedanceUsageSchema = import_zod5.z.object({
+  total_tokens: import_zod5.z.number().optional(),
+  completion_tokens: import_zod5.z.number().optional()
+});
 var SeedanceResponseSchema = import_zod5.z.object({
   id: import_zod5.z.string().optional(),
   status: import_zod5.z.string().optional(),
   content: import_zod5.z.object({ video_url: import_zod5.z.string().optional() }).optional(),
+  usage: SeedanceUsageSchema.optional(),
+  duration: import_zod5.z.number().optional(),
+  framespersecond: import_zod5.z.number().optional(),
+  resolution: import_zod5.z.string().optional(),
   error: import_zod5.z.union([import_zod5.z.string(), import_zod5.z.object({ message: import_zod5.z.string().optional() })]).optional()
 });
 var SeedanceErrorMessageSchema = import_zod5.z.object({ message: import_zod5.z.string().optional() });
@@ -2662,42 +3340,42 @@ async function generateSeedanceVideo(req, ctx) {
   if (req.firstFrame) {
     content.push({
       type: "image_url",
-      image_url: { url: arkUrl(req.firstFrame) },
+      image_url: { url: refToUrl(req.firstFrame) },
       role: "first_frame"
     });
   }
   if (req.lastFrame) {
     content.push({
       type: "image_url",
-      image_url: { url: arkUrl(req.lastFrame) },
+      image_url: { url: refToUrl(req.lastFrame) },
       role: "last_frame"
     });
   }
   for (const ref of req.referenceImages ?? []) {
     content.push({
       type: "image_url",
-      image_url: { url: arkUrl(ref) },
+      image_url: { url: refToUrl(ref) },
       role: "reference_image"
     });
   }
   if (req.sourceVideo) {
     content.push({
       type: "video_url",
-      video_url: { url: arkUrl(req.sourceVideo) },
+      video_url: { url: refToUrl(req.sourceVideo) },
       role: "reference_video"
     });
   }
   for (const ref of req.referenceVideos ?? []) {
     content.push({
       type: "video_url",
-      video_url: { url: arkUrl(ref) },
+      video_url: { url: refToUrl(ref) },
       role: "reference_video"
     });
   }
   for (const ref of req.referenceAudios ?? []) {
     content.push({
       type: "audio_url",
-      audio_url: { url: arkUrl(ref) },
+      audio_url: { url: refToUrl(ref) },
       role: "reference_audio"
     });
   }
@@ -2711,6 +3389,7 @@ async function generateSeedanceVideo(req, ctx) {
   if (!hasFrameAnchor && req.aspectRatio !== void 0) body.ratio = req.aspectRatio;
   if (req.generateAudio !== void 0) body.generate_audio = req.generateAudio;
   if (req.seed !== void 0) body.seed = req.seed;
+  if (req.watermark !== void 0) body.watermark = req.watermark;
   const submit = await withRetry(async () => {
     const res = await arkFetch(ctx, `${arkBaseUrl(ctx)}${SUBMIT_PATH}`, {
       method: "POST",
@@ -2774,6 +3453,11 @@ async function generateSeedanceVideo(req, ctx) {
               "Seedance task succeeded but returned no video_url",
               modelId
             );
+          const tokens = polled.usage;
+          if (tokens?.total_tokens !== void 0 && tokens.completion_tokens !== void 0) {
+            usage2.inputTokens = tokens.total_tokens - tokens.completion_tokens;
+            usage2.outputTokens = tokens.completion_tokens;
+          }
           return {
             video: { url, expiresAt: new Date(Date.now() + VIDEO_URL_TTL_MS) },
             usage: usage2,
@@ -2837,24 +3521,17 @@ var defaultDeps = { createFalClient: defaultCreateFalClient };
 function createFalAdapter(deps = defaultDeps) {
   return {
     id: "fal",
-    generateImage: (req, ctx) => generateMuseImage(req, ctx, deps),
-    generateVideo: (req, ctx) => generateFalVideo(req, ctx, deps)
+    generateImage: (req, ctx) => {
+      throwIfAborted(ctx.signal);
+      return generateMuseImage(req, ctx, deps);
+    },
+    generateVideo: (req, ctx) => {
+      throwIfAborted(ctx.signal);
+      return generateFalVideo(req, ctx, deps);
+    }
   };
 }
 var falAdapter = createFalAdapter();
-function refToUrl(ref) {
-  if ("url" in ref) return ref.url;
-  return `data:${ref.mime};base64,${Buffer.from(ref.bytes).toString("base64")}`;
-}
-async function readBody(res) {
-  const text2 = await res.text().catch(() => "");
-  try {
-    const decoded = JsonValueSchema.safeParse(JSON.parse(text2));
-    return decoded.success ? decoded.data : text2;
-  } catch {
-    return text2;
-  }
-}
 var falDetailSchema = import_zod6.z.object({ detail: JsonValueSchema });
 function liftDetail(body) {
   const parsed = falDetailSchema.safeParse(body);
@@ -3091,8 +3768,12 @@ async function generateMuseImage(req, ctx, deps) {
     raw: payload
   };
 }
+var H3_MAX_MODEL_IDS = /* @__PURE__ */ new Set([
+  "minimax/h3-max",
+  "minimax/h3-max/reference-to-video"
+]);
 function minimaxFamily(providerModelId) {
-  return providerModelId.toLowerCase().includes("h3-max") ? "h3-max" : "h3";
+  return H3_MAX_MODEL_IDS.has(providerModelId.toLowerCase()) ? "h3-max" : "h3";
 }
 function minimaxBase(req, family) {
   const resolution = (family === "h3-max" ? req.resolution ?? "768p" : req.resolution ?? "2k").toUpperCase();
@@ -3100,14 +3781,17 @@ function minimaxBase(req, family) {
     prompt: req.prompt.trim(),
     duration: req.durationSeconds ?? 5,
     resolution,
-    prompt_expansion_mode: req.promptExpansion ?? "balanced",
-    enable_safety_checker: true
+    prompt_expansion_mode: req.promptExpansion ?? "balanced"
   };
+  if (req.safetyChecker !== void 0) input.enable_safety_checker = req.safetyChecker;
   if (req.seed !== void 0) input.seed = req.seed;
   return input;
 }
+var KLING_MODEL_IDS = /* @__PURE__ */ new Set([
+  "fal-ai/kling-video/o1/video-to-video/reference"
+]);
 function isKling(providerModelId) {
-  return providerModelId.toLowerCase().includes("kling");
+  return KLING_MODEL_IDS.has(providerModelId.toLowerCase());
 }
 function klingO1Input(req) {
   const source = req.sourceVideo ? refToUrl(req.sourceVideo) : void 0;
@@ -3180,7 +3864,9 @@ async function generateFalVideo(req, ctx, deps) {
     if (lastFrame) base.end_image_url = lastFrame;
   } else {
     endpoint = `minimax/${family}/text-to-video`;
-    base.aspect_ratio = req.aspectRatio ?? "adaptive";
+    if (req.aspectRatio !== void 0 && req.aspectRatio !== "adaptive") {
+      base.aspect_ratio = req.aspectRatio;
+    }
   }
   if (ctx.routing.fal === "cloudflare-passthrough") {
     return await passthroughVideoJob(ctx, req, endpoint, base);
@@ -3666,6 +4352,14 @@ async function generateVeoVideo(req, ctx, createGenAIClient) {
       model: modelId
     });
   }
+  if (req.lastFrame !== void 0 && req.firstFrame === void 0) {
+    throw new ModelKitError({
+      code: "missing_input",
+      message: `${modelId} requires firstFrame when lastFrame is set`,
+      provider: "google-genai",
+      model: modelId
+    });
+  }
   const duration = req.durationSeconds ?? 8;
   const personGeneration = req.firstFrame ?? refs.length > 0 ? "allow_adult" : "allow_all";
   let image;
@@ -3732,7 +4426,7 @@ async function generateVeoVideo(req, ctx, createGenAIClient) {
       throwSdkError(cause, modelId);
     }
   };
-  const jobId = operation.name ?? "veo-operation";
+  const jobId = operation.name ?? crypto.randomUUID();
   const usage2 = videoUsage2(ctx, req, duration);
   return {
     jobId,
@@ -3871,6 +4565,7 @@ async function generateOmniVideo(req, ctx, createGenAIClient) {
       });
     }
     for (const ref of videos) {
+      if (ref.label !== void 0) parts.push({ type: "text", text: ref.label });
       const media = await resolveMedia(ref, ctx.fetch, ctx.signal, "video/mp4", modelId);
       parts.push({
         type: "video",
@@ -3879,6 +4574,8 @@ async function generateOmniVideo(req, ctx, createGenAIClient) {
       });
     }
     if (req.firstFrame) {
+      if (req.firstFrame.label !== void 0)
+        parts.push({ type: "text", text: req.firstFrame.label });
       const media = await resolveMedia(
         req.firstFrame,
         ctx.fetch,
@@ -3897,6 +4594,7 @@ async function generateOmniVideo(req, ctx, createGenAIClient) {
       }
     }
     for (const ref of images) {
+      if (ref.label !== void 0) parts.push({ type: "text", text: ref.label });
       const media = await resolveMedia(ref, ctx.fetch, ctx.signal, "image/png", modelId);
       if (media.bytes.length > OMNI_INLINE_LIMIT_BYTES) {
         parts.push({
@@ -3940,7 +4638,7 @@ async function generateOmniVideo(req, ctx, createGenAIClient) {
   } finally {
     await cleanupUploads();
   }
-  const jobId = interaction.id ?? "gemini-omni-interaction";
+  const jobId = interaction.id ?? crypto.randomUUID();
   const usage2 = videoUsage2(ctx, req, duration);
   const getInteraction = async () => {
     try {
@@ -4166,11 +4864,11 @@ function createGoogleImageAdapter(deps = defaultGoogleImageDeps) {
       const model = google(ctx.model.providerModelId);
       const content = [];
       for (const ref of req.references ?? []) {
+        if (ref.label !== void 0) content.push({ type: "text", text: ref.label });
         if ("url" in ref) content.push({ type: "image", image: ref.url });
         else content.push({ type: "image", image: ref.bytes, mediaType: ref.mime });
       }
       content.push({ type: "text", text: req.prompt });
-      void ctx.call.postCropToAspect;
       const imageConfig = {};
       if (req.aspectRatio !== void 0 && req.aspectRatio !== "adaptive")
         imageConfig.aspectRatio = req.aspectRatio;
@@ -4186,7 +4884,7 @@ function createGoogleImageAdapter(deps = defaultGoogleImageDeps) {
         const timeout = AbortSignal.timeout(GEMINI_IMAGE_TIMEOUT_MS);
         const signal = ctx.signal !== void 0 ? AbortSignal.any([ctx.signal, timeout]) : timeout;
         try {
-          return await deps.runTextGeneration({
+          const genOptions = {
             model,
             messages: [{ role: "user", content }],
             maxRetries: 0,
@@ -4194,7 +4892,9 @@ function createGoogleImageAdapter(deps = defaultGoogleImageDeps) {
               google: googleOptions
             },
             abortSignal: signal
-          });
+          };
+          if (req.temperature !== void 0) genOptions.temperature = req.temperature;
+          return await deps.runTextGeneration(genOptions);
         } catch (cause) {
           throw toProviderError(cause, ctx);
         }
@@ -4398,26 +5098,14 @@ function headers(ctx) {
   const { key, secret } = credentials(ctx);
   return { "Content-Type": "application/json", "hf-api-key": key, "hf-secret": secret };
 }
-function refToUrl2(ref) {
-  if ("url" in ref) return ref.url;
-  return `data:${ref.mime};base64,${Buffer.from(ref.bytes).toString("base64")}`;
-}
-async function readBody2(res) {
-  const text2 = await res.text().catch(() => "");
-  try {
-    const decoded = JsonValueSchema.safeParse(JSON.parse(text2));
-    return decoded.success ? decoded.data : text2;
-  } catch {
-    return text2;
-  }
-}
 function klingEndpoint(providerModelId) {
   if (providerModelId.startsWith("http")) return providerModelId;
   const path = providerModelId.startsWith("/") ? providerModelId.slice(1) : providerModelId;
   return `${PLATFORM}/${path}`;
 }
-function isDop(providerModelId) {
-  return providerModelId.toLowerCase().includes("dop");
+var DOP_MODEL_IDS = /* @__PURE__ */ new Set(["dop", "dop-preview"]);
+function higgsfieldVideoKind(providerModelId) {
+  return DOP_MODEL_IDS.has(providerModelId.toLowerCase()) ? "dop" : "kling";
 }
 function videoUsage3(ctx, req, duration) {
   const event = {
@@ -4461,12 +5149,12 @@ async function createRequest(ctx, url, init) {
     throw classifyThrown("higgsfield", cause, ctx.model.id);
   }
   if (!res.ok) {
-    throw classifyHttpError(res.status, await readBody2(res), "higgsfield", ctx.model.id);
+    throw classifyHttpError(res.status, await readBody(res), "higgsfield", ctx.model.id);
   }
   return res;
 }
 async function generateKlingVideo(req, ctx) {
-  const firstFrame = req.firstFrame ? refToUrl2(req.firstFrame) : void 0;
+  const firstFrame = req.firstFrame ? refToUrl(req.firstFrame) : void 0;
   if (!firstFrame) {
     throw new ModelKitError({
       code: "missing_input",
@@ -4485,7 +5173,7 @@ async function generateKlingVideo(req, ctx) {
   if (req.durationSeconds !== void 0) body.duration = req.durationSeconds;
   if (req.aspectRatio !== void 0 && req.aspectRatio !== "adaptive")
     body.aspect_ratio = req.aspectRatio;
-  if (req.lastFrame) body.last_image_url = refToUrl2(req.lastFrame);
+  if (req.lastFrame) body.last_image_url = refToUrl(req.lastFrame);
   const endpoint = klingEndpoint(ctx.model.providerModelId);
   const retryOptions = {};
   if (ctx.signal !== void 0) retryOptions.signal = ctx.signal;
@@ -4511,9 +5199,9 @@ async function generateKlingVideo(req, ctx) {
     } catch {
       return void 0;
     }
-    if (res.status >= 500) return void 0;
+    if (res.status === 429 || res.status >= 500) return void 0;
     if (!res.ok) {
-      throw classifyHttpError(res.status, await readBody2(res), "higgsfield", ctx.model.id);
+      throw classifyHttpError(res.status, await readBody(res), "higgsfield", ctx.model.id);
     }
     const payload = JsonValueSchema.parse(await res.json());
     const parsed = klingStatusSchema.safeParse(payload);
@@ -4620,7 +5308,7 @@ function dopUrl(body) {
 var dopCreateSchema = import_zod10.z.object({ id: import_zod10.z.string() });
 var dopStatusSchema = import_zod10.z.object({ status: import_zod10.z.string().optional() });
 async function generateDopVideo(req, ctx) {
-  const firstFrame = req.firstFrame ? refToUrl2(req.firstFrame) : void 0;
+  const firstFrame = req.firstFrame ? refToUrl(req.firstFrame) : void 0;
   if (!firstFrame) {
     throw new ModelKitError({
       code: "missing_input",
@@ -4640,7 +5328,7 @@ async function generateDopVideo(req, ctx) {
     webhook: null
   };
   if (req.lastFrame) {
-    params.input_images_end = [{ type: "image_url", image_url: refToUrl2(req.lastFrame) }];
+    params.input_images_end = [{ type: "image_url", image_url: refToUrl(req.lastFrame) }];
   }
   if (req.seed !== void 0) params.seed = req.seed;
   const retryOptions = {};
@@ -4730,7 +5418,10 @@ async function generateDopVideo(req, ctx) {
 }
 var higgsfieldAdapter = {
   id: "higgsfield",
-  generateVideo: (req, ctx) => isDop(ctx.model.providerModelId) ? generateDopVideo(req, ctx) : generateKlingVideo(req, ctx)
+  generateVideo: (req, ctx) => {
+    throwIfAborted(ctx.signal);
+    return higgsfieldVideoKind(ctx.model.providerModelId) === "dop" ? generateDopVideo(req, ctx) : generateKlingVideo(req, ctx);
+  }
 };
 
 // src/adapters/openai-image.ts
@@ -4825,6 +5516,7 @@ var openaiUsageSchema = import_zod11.z.object({
 var openaiImageAdapter = {
   id: "openai-image",
   async generateImage(req, ctx) {
+    throwIfAborted(ctx.signal);
     const apiKey = ctx.keys.openai;
     if (!apiKey)
       throw new ModelKitError({
@@ -4834,6 +5526,13 @@ var openaiImageAdapter = {
         model: ctx.model.id,
         retryable: false
       });
+    if ((req.references ?? []).some((ref) => ref.label !== void 0)) {
+      ctx.warnings.push({
+        code: "option_ignored",
+        field: "references.label",
+        message: "openai-image cannot express reference labels; ignoring them"
+      });
+    }
     const client2 = new import_openai2.default({ apiKey, fetch: ctx.fetch, maxRetries: 0 });
     const size = req.exactSize !== void 0 ? req.exactSize : req.aspectRatio === void 0 || req.aspectRatio === "adaptive" ? void 0 : aspectRatioToSize(req.aspectRatio, req.size ?? "1K");
     const refs = req.references ?? [];
@@ -4847,24 +5546,34 @@ var openaiImageAdapter = {
       try {
         if (!isEdit) {
           if (size !== void 0) {
-            return await client2.images.generate({
+            return await client2.images.generate(
+              {
+                model: ctx.model.providerModelId,
+                prompt: req.prompt,
+                size,
+                quality: sdkQuality,
+                ...count
+              },
+              { signal: ctx.signal }
+            );
+          }
+          return await client2.images.generate(
+            {
               model: ctx.model.providerModelId,
               prompt: req.prompt,
-              size,
               quality: sdkQuality,
               ...count
-            });
-          }
-          return await client2.images.generate({
-            model: ctx.model.providerModelId,
-            prompt: req.prompt,
-            quality: sdkQuality,
-            ...count
-          });
+            },
+            { signal: ctx.signal }
+          );
         }
         const files = await Promise.all(
           refs.slice(0, MAX_EDIT_REFERENCES).map(async (ref, i) => {
-            const { bytes, mime } = await mediaRefToBytes(ref, ctx.fetch);
+            const { bytes, mime } = await mediaRefToBytes(ref, ctx.fetch, {
+              signal: ctx.signal,
+              provider: "openai-image",
+              model: ctx.model.id
+            });
             const ext = mime.includes("jpeg") ? "jpg" : mime.includes("webp") ? "webp" : "png";
             return new File([bytes], `ref-${i}.${ext}`, { type: mime });
           })
@@ -4872,22 +5581,28 @@ var openaiImageAdapter = {
         const single = files.length === 1 ? files[0] : void 0;
         const imageInput = single ?? files;
         if (size !== void 0) {
-          return await client2.images.edit({
+          return await client2.images.edit(
+            {
+              model: ctx.model.providerModelId,
+              image: imageInput,
+              prompt: req.prompt,
+              size,
+              quality: sdkQuality,
+              ...count
+            },
+            { signal: ctx.signal }
+          );
+        }
+        return await client2.images.edit(
+          {
             model: ctx.model.providerModelId,
             image: imageInput,
             prompt: req.prompt,
-            size,
             quality: sdkQuality,
             ...count
-          });
-        }
-        return await client2.images.edit({
-          model: ctx.model.providerModelId,
-          image: imageInput,
-          prompt: req.prompt,
-          quality: sdkQuality,
-          ...count
-        });
+          },
+          { signal: ctx.signal }
+        );
       } catch (cause) {
         throw toProviderError2(cause, ctx);
       }
@@ -4969,7 +5684,7 @@ async function refToBytes(ctx, ref) {
   if ("bytes" in ref) return { bytes: ref.bytes, mime: ref.mime };
   let res;
   try {
-    res = await ctx.fetch(ref.url);
+    res = await ctx.fetch(ref.url, ctx.signal ? { signal: ctx.signal } : {});
   } catch (cause) {
     throw classifyThrown("prodia", cause, ctx.model.id);
   }
@@ -5015,6 +5730,12 @@ function emptyImage() {
     retryable: false
   });
 }
+function inputExtension(mime) {
+  if (mime.includes("jpeg") || mime.includes("jpg")) return "jpg";
+  if (mime.includes("webp")) return "webp";
+  if (mime.includes("gif")) return "gif";
+  return "png";
+}
 function isBgRemove(providerModelId) {
   const id = providerModelId.toLowerCase();
   return id.includes("remove-background") || id.includes("bg-remove");
@@ -5023,6 +5744,7 @@ function isKlein(providerModelId) {
   return providerModelId.toLowerCase().includes("klein");
 }
 async function generateProdiaImage(req, ctx) {
+  throwIfAborted(ctx.signal);
   const key = ctx.keys.prodia;
   if (!key) {
     throw new ModelKitError({
@@ -5121,7 +5843,9 @@ async function generateProdiaImage(req, ctx) {
       "input",
       // Uint8Array<ArrayBufferLike> is not a valid BlobPart under the
       // strict lib types; copy into a fresh ArrayBuffer-backed view.
-      new File([new Uint8Array(input.bytes)], `input-${i}.jpg`, { type: "image/jpeg" })
+      new File([new Uint8Array(input.bytes)], `input-${i}.${inputExtension(input.mime)}`, {
+        type: input.mime
+      })
     );
   });
   const retryOpts = {
@@ -5136,11 +5860,13 @@ async function generateProdiaImage(req, ctx) {
     res = await withRetry(async () => {
       let attemptRes;
       try {
-        attemptRes = await ctx.fetch(PRODIA_JOB_URL, {
+        const init = {
           method: "POST",
           headers: { Authorization: `Bearer ${key}`, Accept: accept },
           body: form
-        });
+        };
+        if (ctx.signal !== void 0) init.signal = ctx.signal;
+        attemptRes = await ctx.fetch(PRODIA_JOB_URL, init);
       } catch (err) {
         throw classifyThrown("prodia", err, ctx.model.id);
       }
@@ -5201,11 +5927,7 @@ function client(ctx, deps) {
       retryable: false
     });
   }
-  return deps.createReplicateClient({ auth: token });
-}
-function refToUrl3(ref) {
-  if ("url" in ref) return ref.url;
-  return `data:${ref.mime};base64,${Buffer.from(ref.bytes).toString("base64")}`;
+  return deps.createReplicateClient({ auth: token, fetch: ctx.fetch });
 }
 function route(providerModelId) {
   if (providerModelId.includes(":")) {
@@ -5225,8 +5947,9 @@ function replicateError(cause, ctx) {
   const status = info.status ?? (nested.success ? nested.data.response?.status : void 0);
   return classifyThrown("replicate", cause, ctx.model.id, { status, body: info.body });
 }
-function isLower(id) {
-  return id.toLowerCase();
+function baseModelId(providerModelId) {
+  const colon = providerModelId.indexOf(":");
+  return colon === -1 ? providerModelId : providerModelId.slice(0, colon);
 }
 function predictionErrorBody(cause) {
   const parsed = JsonValueSchema.safeParse(cause);
@@ -5274,6 +5997,7 @@ function noOutput() {
   });
 }
 async function runPrediction(ctx, input, deps) {
+  throwIfAborted(ctx.signal);
   const api = client(ctx, deps);
   const createOptions = { ...route(ctx.model.providerModelId), input };
   let prediction;
@@ -5281,12 +6005,14 @@ async function runPrediction(ctx, input, deps) {
     prediction = await withRetry(
       async () => {
         try {
-          return await api.predictions.create(createOptions);
+          return await api.predictions.create(
+            ctx.signal ? { ...createOptions, signal: ctx.signal } : createOptions
+          );
         } catch (cause) {
           throw replicateError(cause, ctx);
         }
       },
-      ctx.signal ? { signal: ctx.signal } : {}
+      { attempts: 1 }
     );
   } catch (cause) {
     throw replicateError(cause, ctx);
@@ -5299,7 +6025,10 @@ async function runPrediction(ctx, input, deps) {
   const done = await pollUntil(async () => {
     let current;
     try {
-      current = await api.predictions.get(prediction.id);
+      current = await api.predictions.get(
+        prediction.id,
+        ctx.signal ? { signal: ctx.signal } : void 0
+      );
     } catch (cause) {
       throw replicateError(cause, ctx);
     }
@@ -5366,21 +6095,32 @@ async function download(ctx, url) {
   const mime = res.headers.get("content-type") ?? "image/png";
   return { bytes: new Uint8Array(await res.arrayBuffer()), mime };
 }
-function isUpscaler(providerModelId) {
-  const id = isLower(providerModelId);
-  return id.includes("upscal") || id.includes("crystal") || id.includes("topaz");
-}
-function isTopazUpscaler(providerModelId) {
-  return isLower(providerModelId).includes("topazlabs/image-upscale");
-}
-function isQwen(providerModelId) {
-  return isLower(providerModelId).includes("qwen");
-}
-function isWanImage(providerModelId) {
-  return isLower(providerModelId).includes("wan-2.7");
+var REPLICATE_IMAGE_KINDS = {
+  "topazlabs/image-upscale": "topaz",
+  "philz1337x/crystal-upscaler": "upscaler",
+  "wan-video/wan-2.7-image-pro": "wan",
+  "qwen/qwen-image-2-pro": "qwen-image",
+  "qwen-edit-apps/qwen-image-edit-plus-lora-skin": "qwen-edit"
+};
+function imageKind(providerModelId) {
+  return REPLICATE_IMAGE_KINDS[baseModelId(providerModelId)] ?? "generic";
+}
+var REPLICATE_VIDEO_KINDS = {
+  "kling-video/v2.5-turbo/pro/image-to-video": "kling",
+  "kling-video/v3.0/pro/image-to-video": "kling",
+  "kwaivgi/kling-v3-video": "kling",
+  "kwaivgi/kling-v3-omni-video": "kling-omni",
+  "minimax/hailuo-02": "hailuo"
+};
+function videoKind(providerModelId) {
+  return REPLICATE_VIDEO_KINDS[baseModelId(providerModelId)] ?? "wan";
 }
-function isQwenEditPlus(providerModelId) {
-  return isLower(providerModelId).includes("edit");
+var REPLICATE_AUDIO_KINDS = {
+  "zsxkib/mmaudio": "mmaudio",
+  "elevenlabs/music": "elevenlabs"
+};
+function audioKind(providerModelId) {
+  return REPLICATE_AUDIO_KINDS[baseModelId(providerModelId)] ?? "lyria";
 }
 var QWEN_IMAGE_RATIOS = /* @__PURE__ */ new Set([
   "1:1",
@@ -5423,87 +6163,113 @@ function wanImageSize(aspectRatio, size) {
 async function generateReplicateImage(req, ctx, deps) {
   const id = ctx.model.providerModelId;
   let input;
-  if (isTopazUpscaler(id)) {
-    const first = (req.references ?? [])[0];
-    if (!first) {
-      throw new ModelKitError({
-        code: "missing_input",
-        message: `${ctx.model.id} requires a reference image to upscale`,
-        provider: "replicate",
-        model: ctx.model.id,
-        retryable: false
-      });
+  switch (imageKind(id)) {
+    case "topaz": {
+      const first = (req.references ?? [])[0];
+      if (!first) {
+        throw new ModelKitError({
+          code: "missing_input",
+          message: `${ctx.model.id} requires a reference image to upscale`,
+          provider: "replicate",
+          model: ctx.model.id,
+          retryable: false
+        });
+      }
+      let format = req.outputFormat === "jpeg" ? "jpg" : req.outputFormat ?? "png";
+      if (format !== "png" && format !== "jpg") {
+        ctx.warnings.push({
+          code: "option_ignored",
+          field: "outputFormat",
+          from: format,
+          message: "topazlabs/image-upscale supports only png and jpg; using png"
+        });
+        format = "png";
+      }
+      const upscale = req.upscale;
+      const topaz = {
+        image: refToUrl(first),
+        enhance_model: upscale?.enhanceModel ?? "Standard V2",
+        output_format: format,
+        upscale_factor: `${req.scaleFactor ?? 2}x`,
+        face_enhancement: upscale?.faceEnhancement ?? false
+      };
+      if (upscale?.subjectDetection !== void 0)
+        topaz.subject_detection = upscale.subjectDetection;
+      if (upscale?.faceEnhancementCreativity !== void 0)
+        topaz.face_enhancement_creativity = upscale.faceEnhancementCreativity;
+      if (upscale?.faceEnhancementStrength !== void 0)
+        topaz.face_enhancement_strength = upscale.faceEnhancementStrength;
+      input = topaz;
+      break;
     }
-    let format = req.outputFormat === "jpeg" ? "jpg" : req.outputFormat ?? "png";
-    if (format !== "png" && format !== "jpg") {
-      ctx.warnings.push({
-        code: "option_ignored",
-        field: "outputFormat",
-        from: format,
-        message: "topazlabs/image-upscale supports only png and jpg; using png"
-      });
-      format = "png";
+    case "upscaler": {
+      const first = (req.references ?? [])[0];
+      if (!first) {
+        throw new ModelKitError({
+          code: "missing_input",
+          message: `${ctx.model.id} requires a reference image to upscale`,
+          provider: "replicate",
+          model: ctx.model.id,
+          retryable: false
+        });
+      }
+      const imageInput = { prompt: req.prompt };
+      imageInput.image = refToUrl(first);
+      imageInput.creativity = 0;
+      imageInput.scale_factor = 2;
+      imageInput.output_format = "png";
+      input = imageInput;
+      break;
     }
-    input = {
-      image: refToUrl3(first),
-      enhance_model: "Standard V2",
-      output_format: format,
-      upscale_factor: `${req.scaleFactor ?? 2}x`,
-      face_enhancement: true
-    };
-  } else if (isUpscaler(id)) {
-    const refs = req.references ?? [];
-    const first = refs[0];
-    if (!first) {
-      throw new ModelKitError({
-        code: "missing_input",
-        message: `${ctx.model.id} requires a reference image to upscale`,
-        provider: "replicate",
-        model: ctx.model.id,
-        retryable: false
-      });
+    case "wan": {
+      const refs = (req.references ?? []).map(refToUrl);
+      const imageInput = { prompt: req.prompt };
+      if (refs.length > 0) imageInput.images = refs;
+      imageInput.size = wanImageSize(req.aspectRatio, req.size);
+      input = imageInput;
+      break;
     }
-    const imageInput = { prompt: req.prompt };
-    imageInput.image = refToUrl3(first);
-    imageInput.creativity = 0;
-    imageInput.scale_factor = 2;
-    imageInput.output_format = "png";
-    input = imageInput;
-  } else if (isWanImage(id)) {
-    const refs = (req.references ?? []).map(refToUrl3);
-    const imageInput = { prompt: req.prompt };
-    if (refs.length > 0) imageInput.images = refs;
-    imageInput.size = wanImageSize(req.aspectRatio, req.size);
-    input = imageInput;
-  } else if (isQwen(id)) {
-    const refs = (req.references ?? []).map(refToUrl3);
-    const firstUrl = refs[0];
-    const imageInput = { prompt: req.prompt };
-    if (firstUrl !== void 0) {
-      imageInput.image = firstUrl;
-      if (!isQwenEditPlus(id)) imageInput.match_input_image = true;
+    case "qwen-image": {
+      const refs = (req.references ?? []).map(refToUrl);
+      const firstUrl = refs[0];
+      const imageInput = { prompt: req.prompt };
+      if (firstUrl !== void 0) {
+        imageInput.image = firstUrl;
+        imageInput.match_input_image = true;
+      } else if (req.aspectRatio !== void 0 && QWEN_IMAGE_RATIOS.has(req.aspectRatio)) {
+        imageInput.aspect_ratio = req.aspectRatio;
+      }
+      input = imageInput;
+      break;
     }
-    const ratios = isQwenEditPlus(id) ? QWEN_EDIT_RATIOS : QWEN_IMAGE_RATIOS;
-    if (req.aspectRatio !== void 0) {
-      if (ratios.has(req.aspectRatio)) imageInput.aspect_ratio = req.aspectRatio;
-      else
-        ctx.warnings.push({
-          code: "option_ignored",
-          field: "aspectRatio",
-          from: req.aspectRatio,
-          message: `${id} has no ${req.aspectRatio} aspect_ratio; using the provider default`
-        });
+    case "qwen-edit": {
+      const refs = (req.references ?? []).map(refToUrl);
+      const firstUrl = refs[0];
+      const imageInput = { prompt: req.prompt };
+      if (firstUrl !== void 0) imageInput.image = firstUrl;
+      if (req.aspectRatio !== void 0) {
+        if (QWEN_EDIT_RATIOS.has(req.aspectRatio))
+          imageInput.aspect_ratio = req.aspectRatio;
+        else
+          ctx.warnings.push({
+            code: "option_ignored",
+            field: "aspectRatio",
+            from: req.aspectRatio,
+            message: `${id} has no ${req.aspectRatio} aspect_ratio; using the provider default`
+          });
+      }
+      if (req.outputFormat !== void 0) {
+        imageInput.output_format = req.outputFormat === "jpeg" ? "jpg" : req.outputFormat;
+      }
+      input = imageInput;
+      break;
     }
-    if (isQwenEditPlus(id) && req.outputFormat !== void 0) {
-      imageInput.output_format = req.outputFormat === "jpeg" ? "jpg" : req.outputFormat;
+    default: {
+      const first = (req.references ?? [])[0];
+      const imageInput = { prompt: req.prompt };
+      if (first) imageInput.image = refToUrl(first);
+      input = imageInput;
     }
-    input = imageInput;
-  } else {
-    const refs = req.references ?? [];
-    const first = refs[0];
-    const imageInput = { prompt: req.prompt };
-    if (first) imageInput.image = refToUrl3(first);
-    input = imageInput;
   }
   const { url, predictTime, raw } = await runPrediction(ctx, input, deps);
   const file = await download(ctx, url);
@@ -5520,8 +6286,8 @@ function wanSize(aspectRatio) {
   return w > h ? "2048*1152" : "1536*2048";
 }
 function klingInput(req) {
-  const images = (req.firstFrame ? [req.firstFrame] : []).map(refToUrl3);
-  if (req.lastFrame) images.push(refToUrl3(req.lastFrame));
+  const images = (req.firstFrame ? [req.firstFrame] : []).map(refToUrl);
+  if (req.lastFrame) images.push(refToUrl(req.lastFrame));
   const input = {
     prompt: req.prompt,
     duration: req.durationSeconds ?? 10,
@@ -5539,13 +6305,18 @@ function klingInput(req) {
   if (req.negativePrompt !== void 0) input.negative_prompt = req.negativePrompt;
   return input;
 }
+function klingOmniMode(req) {
+  if (req.resolution === "1080p") return "pro";
+  if (req.resolution === "720p") return "standard";
+  return req.mode ?? "standard";
+}
 function klingOmniInput(req) {
-  const images = (req.firstFrame ? [req.firstFrame] : []).map(refToUrl3);
-  if (req.lastFrame) images.push(refToUrl3(req.lastFrame));
+  const images = (req.firstFrame ? [req.firstFrame] : []).map(refToUrl);
+  if (req.lastFrame) images.push(refToUrl(req.lastFrame));
   const input = {
     prompt: req.prompt,
     duration: req.durationSeconds ?? 10,
-    mode: req.mode ?? "standard",
+    mode: klingOmniMode(req),
     generate_audio: req.generateAudio ?? false
   };
   const startImage = images[0];
@@ -5556,7 +6327,7 @@ function klingOmniInput(req) {
   } else if (req.aspectRatio !== "adaptive") {
     input.aspect_ratio = req.aspectRatio ?? "16:9";
   }
-  const refs = (req.referenceImages ?? []).map(refToUrl3);
+  const refs = (req.referenceImages ?? []).map(refToUrl);
   if (refs.length > 0) input.reference_images = refs;
   return input;
 }
@@ -5567,8 +6338,8 @@ function hailuoInput(req) {
     resolution: req.resolution ?? "768p",
     prompt_optimizer: true
   };
-  if (req.firstFrame) input.first_frame_image = refToUrl3(req.firstFrame);
-  if (req.lastFrame) input.last_frame_image = refToUrl3(req.lastFrame);
+  if (req.firstFrame) input.first_frame_image = refToUrl(req.firstFrame);
+  if (req.lastFrame) input.last_frame_image = refToUrl(req.lastFrame);
   return input;
 }
 function wanInput(req) {
@@ -5577,12 +6348,13 @@ function wanInput(req) {
     duration: req.durationSeconds ?? 5,
     size: wanSize(req.aspectRatio)
   };
-  if (req.firstFrame) input.image = refToUrl3(req.firstFrame);
+  if (req.firstFrame) input.image = refToUrl(req.firstFrame);
   return input;
 }
 async function generateReplicateVideo(req, ctx, deps) {
-  const id = isLower(ctx.model.providerModelId);
-  const input = id.includes("kling") ? id.includes("omni") ? klingOmniInput(req) : klingInput(req) : id.includes("hailuo") ? hailuoInput(req) : wanInput(req);
+  throwIfAborted(ctx.signal);
+  const kind = videoKind(ctx.model.providerModelId);
+  const input = kind === "kling" ? klingInput(req) : kind === "kling-omni" ? klingOmniInput(req) : kind === "hailuo" ? hailuoInput(req) : wanInput(req);
   const api = client(ctx, deps);
   const createOptions = { ...route(ctx.model.providerModelId), input };
   let prediction;
@@ -5590,12 +6362,14 @@ async function generateReplicateVideo(req, ctx, deps) {
     prediction = await withRetry(
       async () => {
         try {
-          return await api.predictions.create(createOptions);
+          return await api.predictions.create(
+            ctx.signal ? { ...createOptions, signal: ctx.signal } : createOptions
+          );
         } catch (cause) {
           throw replicateError(cause, ctx);
         }
       },
-      ctx.signal ? { signal: ctx.signal } : {}
+      { attempts: 1 }
     );
   } catch (cause) {
     throw replicateError(cause, ctx);
@@ -5604,7 +6378,10 @@ async function generateReplicateVideo(req, ctx, deps) {
   const check = async () => {
     let current;
     try {
-      current = await api.predictions.get(jobId);
+      current = await api.predictions.get(
+        jobId,
+        ctx.signal ? { signal: ctx.signal } : void 0
+      );
     } catch (cause) {
       throw replicateError(cause, ctx);
     }
@@ -5705,7 +6482,7 @@ function mmaudioInput(req, ctx) {
     });
   }
   const input = {
-    video: refToUrl3(video),
+    video: refToUrl(video),
     duration: req.durationSeconds ?? 8,
     seed: req.seed ?? -1
   };
@@ -5724,8 +6501,8 @@ function predictionRaw(prediction) {
   return raw;
 }
 async function generateReplicateAudio(req, ctx, deps) {
-  const id = isLower(ctx.model.providerModelId);
-  const input = id.includes("mmaudio") ? mmaudioInput(req, ctx) : id.includes("elevenlabs") ? elevenLabsMusicInput(req, ctx) : lyriaAudioInput(req, ctx);
+  const kind = audioKind(ctx.model.providerModelId);
+  const input = kind === "mmaudio" ? mmaudioInput(req, ctx) : kind === "elevenlabs" ? elevenLabsMusicInput(req, ctx) : lyriaAudioInput(req, ctx);
   const { url, predictTime, raw } = await runPrediction(ctx, input, deps);
   const usage2 = audioUsage(ctx, req);
   if (ctx.model.usage.unit === "seconds") {
@@ -5754,10 +6531,6 @@ function createXaiVideoAdapter(deps = defaultXaiDeps) {
   };
 }
 var xaiVideoAdapter = createXaiVideoAdapter();
-function refToUrl4(ref) {
-  if ("url" in ref) return ref.url;
-  return `data:${ref.mime};base64,${Buffer.from(ref.bytes).toString("base64")}`;
-}
 var xaiResponseStatusSchema = import_zod13.z.object({
   response: import_zod13.z.object({ status: import_zod13.z.number() }).partial()
 });
@@ -5789,13 +6562,14 @@ async function generateXaiVideo(req, ctx, deps) {
       retryable: false
     });
   }
-  const refUrls = (req.referenceImages ?? []).map(refToUrl4).slice(0, 7);
-  const resolution = req.resolution === "480p" ? "480p" : "720p";
+  const refUrls = (req.referenceImages ?? []).map(refToUrl).slice(0, 7);
   const xaiOptions = {
-    resolution,
     pollIntervalMs: POLL_INTERVAL_MS2,
     pollTimeoutMs: POLL_TIMEOUT_MS2
   };
+  if (req.resolution === "480p" || req.resolution === "720p" || req.resolution === "1080p") {
+    xaiOptions.resolution = req.resolution;
+  }
   if (refUrls.length > 0) {
     xaiOptions.mode = "reference-to-video";
     xaiOptions.referenceImageUrls = refUrls;
@@ -5815,7 +6589,7 @@ async function generateXaiVideo(req, ctx, deps) {
     providerOptions: { xai: xaiOptions }
   };
   if (req.firstFrame) {
-    callBody.frameImages = [{ image: refToUrl4(req.firstFrame), frameType: "first_frame" }];
+    callBody.frameImages = [{ image: refToUrl(req.firstFrame), frameType: "first_frame" }];
   }
   if (req.aspectRatio !== void 0 && req.aspectRatio !== "adaptive") {
     callBody.aspectRatio = req.aspectRatio;
@@ -5851,7 +6625,8 @@ async function generateXaiVideo(req, ctx, deps) {
     return settled;
   };
   return {
-    jobId: `${ctx.model.id}-grok`,
+    // xAI exposes no request id on the synchronous handle, so each job gets a fresh one.
+    jobId: crypto.randomUUID(),
     async poll() {
       if (failed !== void 0) return { status: "failed" };
       return settled ? { status: "succeeded" } : { status: "pending" };
@@ -5986,6 +6761,7 @@ var describe = (modelId) => createDefaultModelKit().describe(modelId);
   prodiaAdapter,
   readErrorBody,
   replicateAdapter,
+  resolveAudioOptions,
   resolveImageOptions,
   resolveKeys,
   resolveVideoOptions,