@shop-os/modelkit 0.3.0 → 0.3.2

package/dist/index.d.cts

@@ -11,7 +11,7 @@ import { LanguageModelV3, LanguageModelV2 } from '@ai-sdk/provider';
 
 /** The AI SDK video model spelling: a catalog id or a provider video model instance. */
 type VideoModel = Parameters<typeof experimental_generateVideo>[0]["model"];
-type AspectRatio = "1:1" | "16:9" | "9:16" | "4:3" | "3:4" | "3:2" | "2:3" | "4:5" | "5:4" | "21:9" | "9:21" | "97:30" | "adaptive";
+type AspectRatio = "1:1" | "16:9" | "9:16" | "4:3" | "3:4" | "3:2" | "2:3" | "4:5" | "5:4" | "21:9" | "9:21" | "97:30" | "1:4" | "4:1" | "1:8" | "8:1" | "adaptive";
 type ImageSize = "512" | "1K" | "2K" | "4K";
 type VideoResolution = "360p" | "480p" | "720p" | "768p" | "1080p" | "2k" | "4k";
 /** "xhigh" and "max" exist only on gpt-image-2.5; gpt-image-2 rejects them at the provider. */
@@ -716,6 +716,10 @@ declare const CatalogModelSchemaBase: z.ZodObject<{
             "21:9": "21:9";
             "9:21": "9:21";
             "97:30": "97:30";
+            "1:4": "1:4";
+            "4:1": "4:1";
+            "1:8": "1:8";
+            "8:1": "8:1";
             adaptive: "adaptive";
         }>>>;
         sizes: z.ZodOptional<z.ZodArray<z.ZodEnum<{
@@ -851,6 +855,10 @@ declare const CatalogModelSchemaBase: z.ZodObject<{
             "21:9": "21:9";
             "9:21": "9:21";
             "97:30": "97:30";
+            "1:4": "1:4";
+            "4:1": "4:1";
+            "1:8": "1:8";
+            "8:1": "8:1";
             adaptive: "adaptive";
         }>>;
         size: z.ZodOptional<z.ZodEnum<{
@@ -1031,6 +1039,7 @@ declare const InputSpecSchema: z.ZodObject<{
     fields: z.ZodRecord<z.ZodString, z.ZodString>;
     values: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodRecord<z.ZodString, z.ZodUnion<readonly [z.ZodString, z.ZodNumber, z.ZodBoolean]>>>>;
     static: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodUnion<readonly [z.ZodString, z.ZodNumber, z.ZodBoolean]>>>;
+    editEndpoint: z.ZodOptional<z.ZodString>;
 }, z.core.$strict>;
 type InputSpec = z.infer<typeof InputSpecSchema>;
 interface InputSpecIssue {

package/dist/index.d.ts

@@ -11,7 +11,7 @@ import { LanguageModelV3, LanguageModelV2 } from '@ai-sdk/provider';
 
 /** The AI SDK video model spelling: a catalog id or a provider video model instance. */
 type VideoModel = Parameters<typeof experimental_generateVideo>[0]["model"];
-type AspectRatio = "1:1" | "16:9" | "9:16" | "4:3" | "3:4" | "3:2" | "2:3" | "4:5" | "5:4" | "21:9" | "9:21" | "97:30" | "adaptive";
+type AspectRatio = "1:1" | "16:9" | "9:16" | "4:3" | "3:4" | "3:2" | "2:3" | "4:5" | "5:4" | "21:9" | "9:21" | "97:30" | "1:4" | "4:1" | "1:8" | "8:1" | "adaptive";
 type ImageSize = "512" | "1K" | "2K" | "4K";
 type VideoResolution = "360p" | "480p" | "720p" | "768p" | "1080p" | "2k" | "4k";
 /** "xhigh" and "max" exist only on gpt-image-2.5; gpt-image-2 rejects them at the provider. */
@@ -716,6 +716,10 @@ declare const CatalogModelSchemaBase: z.ZodObject<{
             "21:9": "21:9";
             "9:21": "9:21";
             "97:30": "97:30";
+            "1:4": "1:4";
+            "4:1": "4:1";
+            "1:8": "1:8";
+            "8:1": "8:1";
             adaptive: "adaptive";
         }>>>;
         sizes: z.ZodOptional<z.ZodArray<z.ZodEnum<{
@@ -851,6 +855,10 @@ declare const CatalogModelSchemaBase: z.ZodObject<{
             "21:9": "21:9";
             "9:21": "9:21";
             "97:30": "97:30";
+            "1:4": "1:4";
+            "4:1": "4:1";
+            "1:8": "1:8";
+            "8:1": "8:1";
             adaptive: "adaptive";
         }>>;
         size: z.ZodOptional<z.ZodEnum<{
@@ -1031,6 +1039,7 @@ declare const InputSpecSchema: z.ZodObject<{
     fields: z.ZodRecord<z.ZodString, z.ZodString>;
     values: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodRecord<z.ZodString, z.ZodUnion<readonly [z.ZodString, z.ZodNumber, z.ZodBoolean]>>>>;
     static: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodUnion<readonly [z.ZodString, z.ZodNumber, z.ZodBoolean]>>>;
+    editEndpoint: z.ZodOptional<z.ZodString>;
 }, z.core.$strict>;
 type InputSpec = z.infer<typeof InputSpecSchema>;
 interface InputSpecIssue {

package/dist/index.js

@@ -169,13 +169,21 @@ var GEMINI_RATIOS = [
   "5:4",
   "adaptive"
 ];
+var NANO_BANANA_2_1_RATIOS = [
+  ...GEMINI_RATIOS.filter((r) => r !== "adaptive"),
+  "1:4",
+  "4:1",
+  "1:8",
+  "8:1",
+  "adaptive"
+];
 var GEMINI_PROMPT_MAX_CHARS = 5e4;
 var GPT_IMAGE_PROMPT = { referenceSyntax: "numbered", maxChars: 32e3 };
-function geminiImage(id, label, aliases, sizes, display, exposure, references, promptMaxChars) {
+function geminiImage(id, label, aliases, sizes, display, exposure, references, promptMaxChars, aspectRatios = GEMINI_RATIOS) {
   const prompt = { referenceSyntax: "caption" };
   if (promptMaxChars !== void 0) prompt.maxChars = promptMaxChars;
   const constraints = {
-    aspectRatios: [...GEMINI_RATIOS],
+    aspectRatios: [...aspectRatios],
     sizes,
     supports: ["safetySettings", "temperature"],
     prompt
@@ -199,6 +207,21 @@ function geminiImage(id, label, aliases, sizes, display, exposure, references, p
   return model;
 }
 var IMAGE_MODELS = [
+  geminiImage(
+    "gemini-nano-banana-2.1",
+    "Nano Banana 2.1",
+    ["nb-2.1"],
+    ["1K", "2K", "4K"],
+    {
+      provider: "google",
+      description: "Latest high-efficiency Nano Banana. Better realism, prompt adherence, text rendering and character consistency at Flash speed. Great for campaigns, infographics, and iterative edits.",
+      logo: "https://d31vvfpk2ppdyq.cloudfront.net/assets/image/192485d2-a76a-4657-9c06-930b2c70badd_compressed.png"
+    },
+    ["cowork", "space_builder"],
+    { images: 14 },
+    GEMINI_PROMPT_MAX_CHARS,
+    NANO_BANANA_2_1_RATIOS
+  ),
   geminiImage(
     "gemini-3-pro-image-preview",
     "Nano Banana Pro",
@@ -2252,6 +2275,10 @@ var ALL_KEY_SET = /* @__PURE__ */ new Set([
   ...IMAGE_INPUT_SPEC_KEYS,
   ...AUDIO_INPUT_SPEC_KEYS
 ]);
+var TYPES_BY_ADAPTER = /* @__PURE__ */ new Map([
+  ["replicate", ["video", "image", "audio"]],
+  ["fal", ["image"]]
+]);
 var ALLOWED_KEYS_BY_TYPE = /* @__PURE__ */ new Map([
   ["video", VIDEO_KEY_SET],
   ["image", IMAGE_KEY_SET],
@@ -2290,7 +2317,8 @@ var StaticSchema = z2.record(z2.string(), ProviderScalarSchema);
 var InputSpecSchema = z2.strictObject({
   fields: FieldsSchema,
   values: ValuesSchema.optional(),
-  static: StaticSchema.optional()
+  static: StaticSchema.optional(),
+  editEndpoint: z2.string().min(1).optional()
 }).superRefine((spec, ctx) => {
   const mappedNames = new Set(Object.values(spec.fields));
   for (const key of Object.keys(spec.static ?? {})) {
@@ -2318,18 +2346,21 @@ function validateInputSpec(raw, row) {
     };
   }
   const spec = parsed.data;
-  if (row.adapter !== "replicate") {
+  const supportedTypes = TYPES_BY_ADAPTER.get(row.adapter);
+  if (supportedTypes === void 0) {
     return {
       ok: false,
-      issues: [{ path: [], message: "inputSpec is only supported on replicate rows" }]
+      issues: [
+        { path: [], message: "inputSpec is only supported on replicate and fal rows" }
+      ]
     };
   }
   const issues = [];
   const allowed = ALLOWED_KEYS_BY_TYPE.get(row.type);
-  if (allowed === void 0) {
+  if (allowed === void 0 || !supportedTypes.includes(row.type)) {
     issues.push({
       path: [],
-      message: `inputSpec is not supported for "${row.type}" rows`
+      message: row.adapter === "fal" ? `inputSpec on fal rows is only supported for image rows, not "${row.type}"` : `inputSpec is not supported for "${row.type}" rows`
     });
   } else {
     for (const key of Object.keys(spec.fields)) {
@@ -2349,6 +2380,19 @@ function validateInputSpec(raw, row) {
       }
     }
   }
+  if (spec.editEndpoint !== void 0) {
+    if (row.adapter !== "fal") {
+      issues.push({
+        path: ["editEndpoint"],
+        message: "inputSpec.editEndpoint is only supported on fal rows"
+      });
+    } else if (!("references" in spec.fields)) {
+      issues.push({
+        path: ["editEndpoint"],
+        message: 'inputSpec.editEndpoint needs a "references" entry in inputSpec.fields'
+      });
+    }
+  }
   const seenNames = /* @__PURE__ */ new Map();
   for (const [key, name] of Object.entries(spec.fields)) {
     const first = seenNames.get(name);
@@ -2424,6 +2468,10 @@ var AspectRatioSchema = z3.enum([
   "21:9",
   "9:21",
   "97:30",
+  "1:4",
+  "4:1",
+  "1:8",
+  "8:1",
   "adaptive"
 ]);
 var ImageSizeSchema = z3.enum(["512", "1K", "2K", "4K"]);
@@ -4093,6 +4141,70 @@ async function cancelFalVideoTool(sub, ctx) {
   throw falError(res.status, await readBody(res), ctx.model.id);
 }
 
+// src/adapters/input-spec-mapper.ts
+function canonicalValue(req, key) {
+  return req[key];
+}
+function isMediaRef(value) {
+  return value instanceof Object && ("url" in value || "bytes" in value);
+}
+function toProviderValue(value, refToUrl2) {
+  if (Array.isArray(value)) {
+    return value.map(refToUrl2);
+  }
+  if (isMediaRef(value)) {
+    return refToUrl2(value);
+  }
+  return value;
+}
+function applyInputSpec(spec, req, refToUrl2, opts) {
+  const out = {};
+  for (const [canonical, providerField] of Object.entries(spec.fields)) {
+    if (!isInputSpecKey(canonical)) continue;
+    const value = canonicalValue(req, canonical);
+    if (value === void 0) continue;
+    if (opts?.omit?.(canonical, value)) continue;
+    const serialised = toProviderValue(value, refToUrl2);
+    const mapped = spec.values?.[canonical]?.[String(value)];
+    out[providerField] = mapped ?? serialised;
+  }
+  for (const [key, value] of Object.entries(spec.static ?? {})) {
+    out[key] = value;
+  }
+  return out;
+}
+function resolveRowInputSpec(req, ctx) {
+  const raw = ctx.model.providerOptions?.["modelkit"]?.["inputSpec"];
+  if (raw === void 0) return { kind: "absent" };
+  const result = validateInputSpec(raw, ctx.model);
+  if (!result.ok) {
+    ctx.logger.warn(
+      `[modelkit] ${ctx.model.id} has an invalid providerOptions.modelkit.inputSpec; the call fails`
+    );
+    return { kind: "invalid", issues: result.issues };
+  }
+  return {
+    kind: "ok",
+    spec: result.spec,
+    input: applyInputSpec(result.spec, req, refToUrl, {
+      omit: (key, value) => key === "aspectRatio" && value === "adaptive"
+    })
+  };
+}
+function describeInputSpecIssue(issue) {
+  if (!issue) return "invalid inputSpec";
+  return issue.path.length === 0 ? issue.message : `${issue.path.map((part) => String(part)).join(".")} ${issue.message}`;
+}
+function invalidInputSpecError(provider, ctx, issues) {
+  return new ModelKitError({
+    code: "unsupported_option",
+    message: `${provider} row "${ctx.model.id}" has an invalid providerOptions.modelkit.inputSpec: ${describeInputSpecIssue(issues[0])}`,
+    provider,
+    model: ctx.model.id,
+    retryable: false
+  });
+}
+
 // src/adapters/fal.ts
 var MUSE_T2I_ENDPOINT = "meta/muse-image/text-to-image";
 var MUSE_EDIT_ENDPOINT = "meta/muse-image/edit";
@@ -4147,6 +4259,11 @@ function createFalAdapter(deps = defaultDeps) {
     id: "fal",
     generateImage: (req, ctx) => {
       throwIfAborted(ctx.signal);
+      const spec = resolveRowInputSpec(req, ctx);
+      if (spec.kind === "invalid") {
+        return Promise.reject(invalidInputSpecError("fal", ctx, spec.issues));
+      }
+      if (spec.kind === "ok") return generateSpecImage(req, ctx, deps, spec);
       if (isIdeogram(ctx.model.providerModelId)) return generateIdeogramImage(req, ctx, deps);
       return generateMuseImage(req, ctx, deps);
     },
@@ -4326,6 +4443,36 @@ var museImageResultSchema = z12.object({
     })
   ).optional()
 });
+async function collectFalImages(payload, ctx, req, label) {
+  const parsed = museImageResultSchema.safeParse(payload);
+  const images = parsed.success ? parsed.data.images ?? [] : [];
+  const files = await Promise.all(
+    images.flatMap((img) => {
+      if (img.url === void 0) return [];
+      return [decodeImage(ctx, img.url, img.content_type ?? void 0)];
+    })
+  );
+  if (files.length === 0) {
+    throw new ModelKitError({
+      code: "provider_rejected",
+      message: `${label} returned no image`,
+      provider: "fal",
+      model: ctx.model.id,
+      retryable: false
+    });
+  }
+  return {
+    images: files.map((f) => ({ bytes: f.bytes, mime: f.mime })),
+    usage: imageUsage(ctx, req, files.length),
+    raw: payload
+  };
+}
+async function generateSpecImage(req, ctx, deps, spec) {
+  const hasReferences = (req.references ?? []).length > 0;
+  const endpoint = hasReferences && spec.spec.editEndpoint !== void 0 ? spec.spec.editEndpoint : ctx.model.providerModelId;
+  const payload = await runJob(ctx, deps, endpoint, spec.input);
+  return collectFalImages(payload, ctx, req, `fal ${endpoint}`);
+}
 async function generateMuseImage(req, ctx, deps) {
   const refs = (req.references ?? []).map(refToUrl);
   const isEdit = refs.length > 0;
@@ -4342,29 +4489,7 @@ async function generateMuseImage(req, ctx, deps) {
   const custom = id.includes("/") && !id.startsWith("meta/muse-image") ? id : void 0;
   const endpoint = custom ?? (isEdit ? MUSE_EDIT_ENDPOINT : MUSE_T2I_ENDPOINT);
   const payload = await runJob(ctx, deps, endpoint, input);
-  const parsed = museImageResultSchema.safeParse(payload);
-  const images = parsed.success ? parsed.data.images ?? [] : [];
-  const firstUrl = images[0]?.url;
-  if (images.length === 0 || firstUrl === void 0) {
-    throw new ModelKitError({
-      code: "provider_rejected",
-      message: "fal Meta Muse returned no image",
-      provider: "fal",
-      model: ctx.model.id,
-      retryable: false
-    });
-  }
-  const files = await Promise.all(
-    images.flatMap((img) => {
-      if (img.url === void 0) return [];
-      return [decodeImage(ctx, img.url, img.content_type ?? void 0)];
-    })
-  );
-  return {
-    images: files.map((f) => ({ bytes: f.bytes, mime: f.mime })),
-    usage: imageUsage(ctx, req, files.length),
-    raw: payload
-  };
+  return collectFalImages(payload, ctx, req, "fal Meta Muse");
 }
 var IDEOGRAM_MODEL_IDS = /* @__PURE__ */ new Set(["ideogram/v4.5"]);
 function isIdeogram(providerModelId) {
@@ -4409,28 +4534,7 @@ async function generateIdeogramImage(req, ctx, deps) {
     if (references.length > 0) input.reference_image_urls = references;
   }
   const payload = await runJob(ctx, deps, endpoint, input);
-  const parsed = museImageResultSchema.safeParse(payload);
-  const images = parsed.success ? parsed.data.images ?? [] : [];
-  const files = await Promise.all(
-    images.flatMap((img) => {
-      if (img.url === void 0) return [];
-      return [decodeImage(ctx, img.url, img.content_type ?? void 0)];
-    })
-  );
-  if (files.length === 0) {
-    throw new ModelKitError({
-      code: "provider_rejected",
-      message: "fal Ideogram returned no image",
-      provider: "fal",
-      model: ctx.model.id,
-      retryable: false
-    });
-  }
-  return {
-    images: files.map((f) => ({ bytes: f.bytes, mime: f.mime })),
-    usage: imageUsage(ctx, req, files.length),
-    raw: payload
-  };
+  return collectFalImages(payload, ctx, req, "fal Ideogram");
 }
 var H3_MAX_MODEL_IDS = /* @__PURE__ */ new Set([
   "minimax/h3-max",
@@ -6781,41 +6885,6 @@ function createReplicateHttpClient(opts) {
 
 // src/adapters/replicate.ts
 import { z as z19 } from "zod";
-
-// src/adapters/input-spec-mapper.ts
-function canonicalValue(req, key) {
-  return req[key];
-}
-function isMediaRef(value) {
-  return value instanceof Object && ("url" in value || "bytes" in value);
-}
-function toProviderValue(value, refToUrl2) {
-  if (Array.isArray(value)) {
-    return value.map(refToUrl2);
-  }
-  if (isMediaRef(value)) {
-    return refToUrl2(value);
-  }
-  return value;
-}
-function applyInputSpec(spec, req, refToUrl2, opts) {
-  const out = {};
-  for (const [canonical, providerField] of Object.entries(spec.fields)) {
-    if (!isInputSpecKey(canonical)) continue;
-    const value = canonicalValue(req, canonical);
-    if (value === void 0) continue;
-    if (opts?.omit?.(canonical, value)) continue;
-    const serialised = toProviderValue(value, refToUrl2);
-    const mapped = spec.values?.[canonical]?.[String(value)];
-    out[providerField] = mapped ?? serialised;
-  }
-  for (const [key, value] of Object.entries(spec.static ?? {})) {
-    out[key] = value;
-  }
-  return out;
-}
-
-// src/adapters/replicate.ts
 var POLL_INTERVAL_MS = 1e3;
 var POLL_TIMEOUT_MS = 15 * 6e4;
 var defaultReplicateDeps = {
@@ -7058,38 +7127,6 @@ function replicateNoMapper(ctx, section) {
     retryable: false
   });
 }
-function describeInputSpecIssue(issue) {
-  if (!issue) return "invalid inputSpec";
-  return issue.path.length === 0 ? issue.message : `${issue.path.map((part) => String(part)).join(".")} ${issue.message}`;
-}
-function replicateInvalidSpec(ctx, issues) {
-  const detail = describeInputSpecIssue(issues[0]);
-  return new ModelKitError({
-    code: "unsupported_option",
-    message: `replicate row "${ctx.model.id}" has an invalid providerOptions.modelkit.inputSpec: ${detail}`,
-    provider: "replicate",
-    model: ctx.model.id,
-    retryable: false
-  });
-}
-function omitAdaptive(key, value) {
-  return key === "aspectRatio" && value === "adaptive";
-}
-function replicateInputSpec(req, ctx) {
-  const raw = ctx.model.providerOptions?.["modelkit"]?.["inputSpec"];
-  if (raw === void 0) return { kind: "absent" };
-  const result = validateInputSpec(raw, ctx.model);
-  if (!result.ok) {
-    ctx.logger.warn(
-      `[modelkit] ${ctx.model.id} has an invalid providerOptions.modelkit.inputSpec; the call fails`
-    );
-    return { kind: "invalid", issues: result.issues };
-  }
-  return {
-    kind: "ok",
-    input: applyInputSpec(result.spec, req, refToUrl, { omit: omitAdaptive })
-  };
-}
 var QWEN_IMAGE_RATIOS = /* @__PURE__ */ new Set([
   "1:1",
   "16:9",
@@ -7233,8 +7270,9 @@ async function generateReplicateImage(req, ctx, deps) {
       break;
     }
     default: {
-      const specResult = replicateInputSpec(req, ctx);
-      if (specResult.kind === "invalid") throw replicateInvalidSpec(ctx, specResult.issues);
+      const specResult = resolveRowInputSpec(req, ctx);
+      if (specResult.kind === "invalid")
+        throw invalidInputSpecError("replicate", ctx, specResult.issues);
       if (specResult.kind === "ok") {
         input = specResult.input;
         break;
@@ -7314,9 +7352,10 @@ async function generateReplicateVideo(req, ctx, deps) {
   else if (kind === "kling-omni") input = klingOmniInput(req);
   else if (kind === "hailuo") input = hailuoInput(req);
   else {
-    const specResult = replicateInputSpec(req, ctx);
+    const specResult = resolveRowInputSpec(req, ctx);
     if (specResult.kind === "absent") throw replicateNoMapper(ctx, "video");
-    if (specResult.kind === "invalid") throw replicateInvalidSpec(ctx, specResult.issues);
+    if (specResult.kind === "invalid")
+      throw invalidInputSpecError("replicate", ctx, specResult.issues);
     input = specResult.input;
   }
   const api = client(ctx, deps);
@@ -7471,9 +7510,10 @@ async function generateReplicateAudio(req, ctx, deps) {
   else if (kind === "elevenlabs") input = elevenLabsMusicInput(req, ctx);
   else if (kind === "lyria") input = lyriaAudioInput(req, ctx);
   else {
-    const specResult = replicateInputSpec(req, ctx);
+    const specResult = resolveRowInputSpec(req, ctx);
     if (specResult.kind === "absent") throw replicateNoMapper(ctx, "audio");
-    if (specResult.kind === "invalid") throw replicateInvalidSpec(ctx, specResult.issues);
+    if (specResult.kind === "invalid")
+      throw invalidInputSpecError("replicate", ctx, specResult.issues);
     input = specResult.input;
   }
   const { url, predictTime, raw } = await runPrediction(ctx, input, deps);