@shop-os/modelkit 0.3.5 → 0.3.7

package/dist/index.d.cts

@@ -1,4 +1,4 @@
-import { ImageModel, experimental_generateVideo, generateText as generateText$1, LanguageModel, FlexibleSchema, InferSchema, ToolSet, Output, streamText as streamText$1, JSONValue, generateObject as generateObject$1, streamObject as streamObject$1, LanguageModelUsage } from 'ai';
+import { ImageModel, experimental_generateVideo, generateText as generateText$1, LanguageModel, FlexibleSchema, InferSchema, ToolSet, Output, streamText as streamText$1, JSONValue, generateObject as generateObject$1, streamObject as streamObject$1, LanguageModelUsage, ProviderMetadata } from 'ai';
 export * from 'ai';
 export { Warning as AiWarning } from 'ai';
 import { z } from 'zod';
@@ -1302,7 +1302,7 @@ declare function mergeTextProviderOptions(catalog: CatalogModel, caller: AiTextP
  */
 declare function passthroughTextModel(model: TextModelInstance): CatalogModel;
 /** Canonical token usage for a completed text call. Quantity is total tokens. */
-declare function textUsageEvent(model: CatalogModel, provider: string, usage: LanguageModelUsage): UsageEvent;
+declare function textUsageEvent(model: CatalogModel, provider: string, usage: LanguageModelUsage, providerMetadata?: ProviderMetadata): UsageEvent;
 
 /** The SDK's streamText option union; `experimental_prepareStep` is a generateText-only alias. */
 type StreamTextCallOptions<TOOLS extends ToolSet, OUTPUT extends Output.Output = Output.Output<string, string, never>> = ModelkitStreamTextOptions<TOOLS, OUTPUT> & {

package/dist/index.d.ts

@@ -1,4 +1,4 @@
-import { ImageModel, experimental_generateVideo, generateText as generateText$1, LanguageModel, FlexibleSchema, InferSchema, ToolSet, Output, streamText as streamText$1, JSONValue, generateObject as generateObject$1, streamObject as streamObject$1, LanguageModelUsage } from 'ai';
+import { ImageModel, experimental_generateVideo, generateText as generateText$1, LanguageModel, FlexibleSchema, InferSchema, ToolSet, Output, streamText as streamText$1, JSONValue, generateObject as generateObject$1, streamObject as streamObject$1, LanguageModelUsage, ProviderMetadata } from 'ai';
 export * from 'ai';
 export { Warning as AiWarning } from 'ai';
 import { z } from 'zod';
@@ -1302,7 +1302,7 @@ declare function mergeTextProviderOptions(catalog: CatalogModel, caller: AiTextP
  */
 declare function passthroughTextModel(model: TextModelInstance): CatalogModel;
 /** Canonical token usage for a completed text call. Quantity is total tokens. */
-declare function textUsageEvent(model: CatalogModel, provider: string, usage: LanguageModelUsage): UsageEvent;
+declare function textUsageEvent(model: CatalogModel, provider: string, usage: LanguageModelUsage, providerMetadata?: ProviderMetadata): UsageEvent;
 
 /** The SDK's streamText option union; `experimental_prepareStep` is a generateText-only alias. */
 type StreamTextCallOptions<TOOLS extends ToolSet, OUTPUT extends Output.Output = Output.Output<string, string, never>> = ModelkitStreamTextOptions<TOOLS, OUTPUT> & {

package/dist/index.js

@@ -1191,12 +1191,18 @@ var TEXT_MODELS = [
       contextWindow: 1048576
     }
   }),
+  // Perplexity stopped serving Sonar Chat Completions on 2026-09-27; @ai-sdk/perplexity 3.1
+  // speaks the Agent API, which rejects the bare `sonar*` ids. Its migration guide maps Sonar
+  // and Sonar Pro to the `fast` preset and Sonar Reasoning Pro to `low`
+  // (docs.perplexity.ai/docs/agent-api/migrate-from-sonar/overview).
   text("sonar", "Sonar", "sonar", {
+    lifecycle: "deprecated",
+    replacedBy: "perplexity/sonar",
     exposure: [],
     display: {
       provider: "perplexity",
       costTier: 1,
-      description: "Perplexity Sonar lightweight web-grounded search model returning cited answers; 128K-token context."
+      description: "Perplexity Sonar on Chat Completions, which Perplexity stopped serving on 2026-09-27; calls now fail."
     },
     capabilities: {
       vision: true,
@@ -1204,11 +1210,13 @@ var TEXT_MODELS = [
     }
   }),
   text("sonar-pro", "Sonar Pro", "sonar-pro", {
+    lifecycle: "deprecated",
+    replacedBy: "perplexity/fast",
     exposure: [],
     display: {
       provider: "perplexity",
       costTier: 2,
-      description: "Perplexity Sonar Pro advanced web-grounded search model for complex multi-step queries; 200K-token context."
+      description: "Perplexity Sonar Pro on Chat Completions, which Perplexity stopped serving on 2026-09-27; calls now fail."
     },
     capabilities: {
       vision: true,
@@ -1216,12 +1224,49 @@ var TEXT_MODELS = [
     }
   }),
   text("sonar-reasoning-pro", "Sonar Reasoning Pro", "sonar-reasoning-pro", {
+    lifecycle: "deprecated",
+    replacedBy: "perplexity/fast",
     exposure: [],
     display: {
       provider: "perplexity",
       costTier: 2,
-      description: "Perplexity Sonar Reasoning Pro chain-of-thought web-grounded search model for complex problem-solving; 128K-token context."
+      description: "Perplexity Sonar Reasoning Pro on Chat Completions, which Perplexity stopped serving on 2026-09-27; calls now fail."
+    },
+    capabilities: {
+      vision: true,
+      contextWindow: 128e3
+    }
+  }),
+  // Sonar as an Agent API model. A bare model searches only when given the web_search
+  // tool, so the row sends it on every call; the answer's sources come from its results.
+  text("perplexity/sonar", "Perplexity Sonar", "perplexity/sonar", {
+    exposure: [],
+    providerOptions: {
+      modelkit: { provider: "perplexity" },
+      perplexity: { tools: [{ type: "web_search" }] }
+    },
+    display: {
+      provider: "perplexity",
+      costTier: 1,
+      description: "Perplexity Sonar on the Agent API with web search, returning answers grounded in cited sources."
+    },
+    // Perplexity publishes no Agent API context window; Sonar's own is 128K.
+    capabilities: {
+      vision: true,
+      contextWindow: 128e3
+    }
+  }),
+  // The Agent API `fast` preset: openai/gpt-5.6-luna with one web search, Perplexity's
+  // recommended replacement for Sonar and Sonar Pro.
+  text("perplexity/fast", "Perplexity Fast", "fast", {
+    exposure: [],
+    providerOptions: { modelkit: { provider: "perplexity" } },
+    display: {
+      provider: "perplexity",
+      costTier: 1,
+      description: "Perplexity Agent API fast preset: a quick web-grounded answer from one search."
     },
+    // No published preset window; Sonar's 128K, the size the preset replaces.
     capabilities: {
       vision: true,
       contextWindow: 128e3
@@ -2923,7 +2968,7 @@ function detectTextProvider(modelId) {
   if (id.startsWith("claude") || id.includes("anthropic.")) return "anthropic";
   if (id.startsWith("gemini") || id.startsWith("gemma")) return "google";
   if (id.startsWith("grok")) return "xai";
-  if (id.startsWith("sonar")) return "perplexity";
+  if (id.startsWith("sonar") || id.startsWith("perplexity/")) return "perplexity";
   if (id.startsWith("deepseek")) return "deepseek";
   if (id.startsWith("kimi") || id.startsWith("moonshot")) return "moonshot";
   if (id.startsWith("minimax")) return "minimax";
@@ -6706,6 +6751,7 @@ var openaiImageAdapter = {
 
 // src/adapters/prodia.ts
 var PRODIA_JOB_URL = "https://inference.prodia.com/v2/job";
+var FLUX_2_PRO_T2I_JOB = "inference.flux-2.pro.txt2img.v1";
 var FLUX_2_PRO_JOB = "inference.flux-2.pro.img2img.v1";
 var FLUX_KLEIN_T2I_JOB = "inference.flux-2.klein.txt2img.v1";
 var FLUX_KLEIN_I2I_JOB = "inference.flux-2.klein.img2img.v1";
@@ -6878,7 +6924,7 @@ async function generateProdiaImage(req, ctx) {
       safety_tolerance: 5
     };
     if (req.seed !== void 0) config.seed = req.seed;
-    job = { type: FLUX_2_PRO_JOB, config };
+    job = { type: refs.length > 0 ? FLUX_2_PRO_JOB : FLUX_2_PRO_T2I_JOB, config };
     inputs = refs;
     accept = "image/png";
     mime = "image/png";
@@ -8769,6 +8815,7 @@ function keysFromEnv(env, prefix = "") {
 
 // src/text.ts
 import "ai";
+import { z as z22 } from "zod";
 function mergeTextProviderOptions(catalog, caller) {
   const merged = {};
   const catalogOptions = catalog.providerOptions ?? {};
@@ -8807,11 +8854,18 @@ function passthroughTextModel(model) {
     pricingKind: "token"
   };
 }
-function textUsageEvent(model, provider, usage2) {
+var perplexityCostSchema = z22.object({
+  perplexity: z22.object({ cost: z22.object({ totalCost: z22.number().finite().nonnegative() }) })
+});
+function reportedCostUsd(metadata) {
+  const parsed = perplexityCostSchema.safeParse(metadata);
+  return parsed.success ? parsed.data.perplexity.cost.totalCost : void 0;
+}
+function textUsageEvent(model, provider, usage2, providerMetadata) {
   const inputTokens = usage2.inputTokens ?? 0;
   const outputTokens = usage2.outputTokens ?? 0;
   const total = usage2.totalTokens ?? inputTokens + outputTokens;
-  return {
+  const event = {
     model: model.id,
     provider,
     modality: "TEXT",
@@ -8820,6 +8874,9 @@ function textUsageEvent(model, provider, usage2) {
     inputTokens,
     outputTokens
   };
+  const cost = reportedCostUsd(providerMetadata);
+  if (cost !== void 0) event.reportedCostUsd = cost;
+  return event;
 }
 
 // src/text-steps.ts
@@ -8846,12 +8903,20 @@ function createTextSteps(initial, resolve, callerProviderOptions) {
     const grouped = /* @__PURE__ */ new Map();
     for (const [index, step] of results.entries()) {
       const target = steps.get(index) ?? initial;
-      const event = textUsageEvent(target.entry, target.provider, step.usage);
+      const event = textUsageEvent(
+        target.entry,
+        target.provider,
+        step.usage,
+        step.providerMetadata
+      );
       const previous = grouped.get(target.entry.id);
       if (previous) {
         previous.quantity += event.quantity;
         previous.inputTokens = (previous.inputTokens ?? 0) + (event.inputTokens ?? 0);
         previous.outputTokens = (previous.outputTokens ?? 0) + (event.outputTokens ?? 0);
+        if (event.reportedCostUsd !== void 0) {
+          previous.reportedCostUsd = (previous.reportedCostUsd ?? 0) + event.reportedCostUsd;
+        }
       } else {
         grouped.set(target.entry.id, event);
       }
@@ -8870,7 +8935,7 @@ import {
   streamObject as aiStreamObject2,
   streamText as aiStreamText2
 } from "ai";
-import { z as z22 } from "zod";
+import { z as z23 } from "zod";
 
 // src/generated-image.ts
 function toGeneratedImage(image) {
@@ -8952,10 +9017,10 @@ function textAbortOptions(sdk, context) {
   if (abortSignal) return { abortSignal };
   return {};
 }
-var catalogId = z22.string();
-var exactSize = z22.templateLiteral([z22.number(), "x", z22.number()]);
-var sdkRatio = z22.templateLiteral([z22.number(), ":", z22.number()]);
-var sizeTier = z22.enum(["512", "1K", "2K", "4K"]);
+var catalogId = z23.string();
+var exactSize = z23.templateLiteral([z23.number(), "x", z23.number()]);
+var sdkRatio = z23.templateLiteral([z23.number(), ":", z23.number()]);
+var sizeTier = z23.enum(["512", "1K", "2K", "4K"]);
 function isCatalogId(model) {
   return catalogId.safeParse(model).success;
 }
@@ -9524,7 +9589,9 @@ function createModelKit(config, adapters) {
         providerOptions: mergeTextProviderOptions(entry, options.providerOptions)
       };
       const result = await aiGenerateObject2(sdkOptions);
-      await emit(withRowUnit(textUsageEvent(entry, provider, result.usage)));
+      await emit(
+        withRowUnit(textUsageEvent(entry, provider, result.usage, result.providerMetadata))
+      );
       return result;
     },
     streamObject(options, call = {}) {
@@ -9536,7 +9603,11 @@ function createModelKit(config, adapters) {
         model: languageModel2,
         providerOptions: mergeTextProviderOptions(entry, options.providerOptions),
         onFinish: async (event) => {
-          await emit(withRowUnit(textUsageEvent(entry, provider, event.usage)));
+          await emit(
+            withRowUnit(
+              textUsageEvent(entry, provider, event.usage, event.providerMetadata)
+            )
+          );
           await callerOnFinish?.(event);
         }
       };