@ai-sdk/google 2.0.102 → 2.0.103

package/CHANGELOG.md

@@ -1,5 +1,11 @@
 # @ai-sdk/google
 
+## 2.0.103
+
+### Patch Changes
+
+- 7c28815: feat(google): add `gemini-4-argon` model id and long decode support: requests that reach the per-request cap (`CONTINUATION`) are continued automatically with usage summed across requests and the first request's prompt token count exposed as `providerMetadata.google.initialPromptTokenCount`, and the `continuationToken` provider option / provider metadata resume responses that used up their `maxOutputTokens` budget
+
 ## 2.0.102
 
 ### Patch Changes

package/dist/index.d.mts

@@ -47,6 +47,7 @@ declare const googleGenerativeAIProviderOptions: _ai_sdk_provider_utils.LazySche
         } | undefined;
     } | undefined;
     serviceTier?: "standard" | "flex" | "priority" | undefined;
+    continuationToken?: string | undefined;
 }>;
 type GoogleGenerativeAIProviderOptions = InferValidator<typeof googleGenerativeAIProviderOptions>;
 
@@ -80,6 +81,7 @@ declare const responseSchema: _ai_sdk_provider_utils.LazySchema<{
             })[] | null | undefined;
         } | null | undefined;
         finishReason?: string | null | undefined;
+        continuationToken?: string | null | undefined;
         safetyRatings?: {
             category?: string | null | undefined;
             probability?: string | null | undefined;
@@ -184,6 +186,22 @@ interface GoogleGenerativeAIProviderMetadata {
     urlContextMetadata: GoogleGenerativeAIUrlContextMetadata | null;
     safetyRatings: GoogleGenerativeAISafetyRating[] | null;
     serviceTier: string | null;
+    /**
+     * Opaque token returned when a response stops with `MAX_TOKENS` because it
+     * used up its `maxOutputTokens` budget. Pass it back as
+     * `providerOptions.google.continuationToken` together with a higher
+     * `maxOutputTokens` and the original prompt to resume generation.
+     */
+    continuationToken?: string | null;
+    /**
+     * Prompt token count of the first request of a long decode response that
+     * was continued automatically (Gemini 4). Usage is summed across requests
+     * to match billing, and each continued request re-sends the prompt plus all
+     * output so far as input. Use this value instead of the summed input tokens
+     * for context window checks and compaction. Only present when the response
+     * spanned more than one request.
+     */
+    initialPromptTokenCount?: number | null;
 }
 
 type GoogleGenerativeAIImageModelId = 'imagen-4.0-generate-001' | 'imagen-4.0-ultra-generate-001' | 'imagen-4.0-fast-generate-001' | (string & {});

package/dist/index.d.ts

@@ -47,6 +47,7 @@ declare const googleGenerativeAIProviderOptions: _ai_sdk_provider_utils.LazySche
         } | undefined;
     } | undefined;
     serviceTier?: "standard" | "flex" | "priority" | undefined;
+    continuationToken?: string | undefined;
 }>;
 type GoogleGenerativeAIProviderOptions = InferValidator<typeof googleGenerativeAIProviderOptions>;
 
@@ -80,6 +81,7 @@ declare const responseSchema: _ai_sdk_provider_utils.LazySchema<{
             })[] | null | undefined;
         } | null | undefined;
         finishReason?: string | null | undefined;
+        continuationToken?: string | null | undefined;
         safetyRatings?: {
             category?: string | null | undefined;
             probability?: string | null | undefined;
@@ -184,6 +186,22 @@ interface GoogleGenerativeAIProviderMetadata {
     urlContextMetadata: GoogleGenerativeAIUrlContextMetadata | null;
     safetyRatings: GoogleGenerativeAISafetyRating[] | null;
     serviceTier: string | null;
+    /**
+     * Opaque token returned when a response stops with `MAX_TOKENS` because it
+     * used up its `maxOutputTokens` budget. Pass it back as
+     * `providerOptions.google.continuationToken` together with a higher
+     * `maxOutputTokens` and the original prompt to resume generation.
+     */
+    continuationToken?: string | null;
+    /**
+     * Prompt token count of the first request of a long decode response that
+     * was continued automatically (Gemini 4). Usage is summed across requests
+     * to match billing, and each continued request re-sends the prompt plus all
+     * output so far as input. Use this value instead of the summed input tokens
+     * for context window checks and compaction. Only present when the response
+     * spanned more than one request.
+     */
+    initialPromptTokenCount?: number | null;
 }
 
 type GoogleGenerativeAIImageModelId = 'imagen-4.0-generate-001' | 'imagen-4.0-ultra-generate-001' | 'imagen-4.0-fast-generate-001' | (string & {});

package/dist/index.js

@@ -30,7 +30,7 @@ module.exports = __toCommonJS(index_exports);
 var import_provider_utils15 = require("@ai-sdk/provider-utils");
 
 // src/version.ts
-var VERSION = true ? "2.0.102" : "0.0.0-test";
+var VERSION = true ? "2.0.103" : "0.0.0-test";
 
 // src/google-generative-ai-embedding-model.ts
 var import_provider = require("@ai-sdk/provider");
@@ -241,6 +241,100 @@ var googleGenerativeAISingleEmbeddingResponseSchema = (0, import_provider_utils3
 var import_provider_utils6 = require("@ai-sdk/provider-utils");
 var import_v45 = require("zod/v4");
 
+// src/add-google-generative-ai-usage-metadata.ts
+function addGoogleGenerativeAIUsageMetadata(a, b) {
+  if (a == null || b == null) {
+    return b != null ? b : a;
+  }
+  return {
+    ...a,
+    ...b,
+    promptTokenCount: addCounts(a.promptTokenCount, b.promptTokenCount),
+    candidatesTokenCount: addCounts(
+      a.candidatesTokenCount,
+      b.candidatesTokenCount
+    ),
+    totalTokenCount: addCounts(a.totalTokenCount, b.totalTokenCount),
+    cachedContentTokenCount: addCounts(
+      a.cachedContentTokenCount,
+      b.cachedContentTokenCount
+    ),
+    thoughtsTokenCount: addCounts(a.thoughtsTokenCount, b.thoughtsTokenCount),
+    promptTokensDetails: addTokenDetails(
+      a.promptTokensDetails,
+      b.promptTokensDetails
+    ),
+    candidatesTokensDetails: addTokenDetails(
+      a.candidatesTokensDetails,
+      b.candidatesTokensDetails
+    )
+  };
+}
+function addCounts(a, b) {
+  return a == null && b == null ? b != null ? b : a : (a != null ? a : 0) + (b != null ? b : 0);
+}
+function addTokenDetails(a, b) {
+  var _a;
+  if (a == null || b == null) {
+    return b != null ? b : a;
+  }
+  const tokenCounts = /* @__PURE__ */ new Map();
+  for (const { modality, tokenCount } of [...a, ...b]) {
+    tokenCounts.set(modality, ((_a = tokenCounts.get(modality)) != null ? _a : 0) + tokenCount);
+  }
+  return Array.from(tokenCounts, ([modality, tokenCount]) => ({
+    modality,
+    tokenCount
+  }));
+}
+
+// src/continue-google-stream.ts
+function continueGoogleStream({
+  stream,
+  getContinuationToken,
+  continueStream
+}) {
+  let reader = stream.getReader();
+  let pendingContinuationToken;
+  return new ReadableStream({
+    async pull(controller) {
+      while (true) {
+        const { done, value } = await reader.read();
+        if (!done) {
+          if (value.success) {
+            const token = getContinuationToken(value.value);
+            if (token != null) {
+              pendingContinuationToken = token;
+            }
+          }
+          controller.enqueue(value);
+          return;
+        }
+        if (pendingContinuationToken == null) {
+          controller.close();
+          return;
+        }
+        const continuationToken = pendingContinuationToken;
+        pendingContinuationToken = void 0;
+        try {
+          reader = (await continueStream(continuationToken)).getReader();
+        } catch (error) {
+          controller.enqueue({
+            success: false,
+            error,
+            rawValue: void 0
+          });
+          controller.close();
+          return;
+        }
+      }
+    },
+    async cancel(reason) {
+      await reader.cancel(reason);
+    }
+  });
+}
+
 // src/convert-json-schema-to-openapi-schema.ts
 var import_provider2 = require("@ai-sdk/provider");
 var recursiveReferenceFunctionalityPrefix = "recursive JSON Schema reference:";
@@ -896,7 +990,17 @@ var googleGenerativeAIProviderOptions = (0, import_provider_utils5.lazySchema)(
       /**
        * Optional. The service tier to use for the request.
        */
-      serviceTier: import_v44.z.enum(["standard", "flex", "priority"]).optional()
+      serviceTier: import_v44.z.enum(["standard", "flex", "priority"]).optional(),
+      /**
+       * Optional. Resumes a response that stopped with `MAX_TOKENS` because it
+       * used up its `maxOutputTokens` budget (Gemini 4 long decode). Pass the
+       * `continuationToken` from the previous response's provider metadata
+       * together with the original prompt and a higher `maxOutputTokens` (the
+       * budget is cumulative across requests). Requests that reach the
+       * per-request cap while budget remains (`CONTINUATION`) are continued
+       * automatically. Sent as the `continuationToken` body field.
+       */
+      continuationToken: import_v44.z.string().optional()
     })
   )
 );
@@ -1194,6 +1298,9 @@ function mapGoogleGenerativeAIFinishReason({
     case "STOP":
       return hasToolCalls ? "tool-calls" : "stop";
     case "MAX_TOKENS":
+    // Gemini 4 long decode: the per-request cap was reached while budget
+    // remains. Continued automatically; only surfaced if continuing fails.
+    case "CONTINUATION":
       return "length";
     case "IMAGE_SAFETY":
     case "RECITATION":
@@ -1368,43 +1475,70 @@ var GoogleGenerativeAILanguageModel = class {
         } : googleToolConfig,
         cachedContent: googleOptions == null ? void 0 : googleOptions.cachedContent,
         labels: googleOptions == null ? void 0 : googleOptions.labels,
-        serviceTier: sanitizedServiceTier
+        serviceTier: sanitizedServiceTier,
+        ...(googleOptions == null ? void 0 : googleOptions.continuationToken) != null && {
+          continuationToken: googleOptions.continuationToken
+        }
       },
       warnings: [...warnings, ...toolWarnings]
     };
   }
   async doGenerate(options) {
-    var _a, _b, _c, _d, _e, _f, _g, _h, _i, _j, _k, _l, _m, _n, _o, _p, _q;
+    var _a, _b, _c, _d, _e, _f, _g, _h, _i, _j, _k, _l, _m, _n, _o, _p, _q, _r, _s, _t, _u, _v, _w;
     const { args, warnings } = await this.getArgs(options);
     const body = JSON.stringify(args);
     const mergedHeaders = (0, import_provider_utils6.combineHeaders)(
       await (0, import_provider_utils6.resolve)(this.config.headers),
       options.headers
     );
-    const {
-      responseHeaders,
-      value: response,
-      rawValue: rawResponse
-    } = await (0, import_provider_utils6.postJsonToApi)({
+    const generateContent = (body2) => (0, import_provider_utils6.postJsonToApi)({
       url: `${this.config.baseURL}/${getModelPath(
         this.modelId
       )}:generateContent`,
       headers: mergedHeaders,
-      body: args,
+      body: body2,
       failedResponseHandler: googleFailedResponseHandler,
       successfulResponseHandler: (0, import_provider_utils6.createJsonResponseHandler)(responseSchema),
       abortSignal: options.abortSignal,
       fetch: this.config.fetch
     });
-    const candidate = (_a = response.candidates) == null ? void 0 : _a[0];
-    const promptBlockReason = (_b = response.promptFeedback) == null ? void 0 : _b.blockReason;
+    let {
+      responseHeaders,
+      value: response,
+      rawValue: rawResponse
+    } = await generateContent(args);
+    const continuedParts = [];
+    let continuedUsageMetadata;
+    const firstPromptTokenCount = (_a = response.usageMetadata) == null ? void 0 : _a.promptTokenCount;
+    while (true) {
+      const candidate2 = (_b = response.candidates) == null ? void 0 : _b[0];
+      const continuationToken = (candidate2 == null ? void 0 : candidate2.finishReason) === "CONTINUATION" ? candidate2.continuationToken : void 0;
+      if (continuationToken == null) {
+        break;
+      }
+      continuedParts.push(...(_d = (_c = candidate2 == null ? void 0 : candidate2.content) == null ? void 0 : _c.parts) != null ? _d : []);
+      continuedUsageMetadata = addGoogleGenerativeAIUsageMetadata(
+        continuedUsageMetadata,
+        (_e = response.usageMetadata) != null ? _e : void 0
+      );
+      ({
+        responseHeaders,
+        value: response,
+        rawValue: rawResponse
+      } = await generateContent({ ...args, continuationToken }));
+    }
+    const candidate = (_f = response.candidates) == null ? void 0 : _f[0];
+    const promptBlockReason = (_g = response.promptFeedback) == null ? void 0 : _g.blockReason;
     const isPromptBlocked = (candidate == null ? void 0 : candidate.finishReason) == null && promptBlockReason != null;
     const content = [];
-    const parts = (_d = (_c = candidate == null ? void 0 : candidate.content) == null ? void 0 : _c.parts) != null ? _d : [];
-    const usageMetadata = response.usageMetadata;
+    const parts = [...continuedParts, ...(_i = (_h = candidate == null ? void 0 : candidate.content) == null ? void 0 : _h.parts) != null ? _i : []];
+    const usageMetadata = addGoogleGenerativeAIUsageMetadata(
+      continuedUsageMetadata,
+      (_j = response.usageMetadata) != null ? _j : void 0
+    );
     let lastCodeExecutionToolCallId;
     for (const part of parts) {
-      if ("executableCode" in part && ((_e = part.executableCode) == null ? void 0 : _e.code)) {
+      if ("executableCode" in part && ((_k = part.executableCode) == null ? void 0 : _k.code)) {
         const toolCallId = this.config.generateId();
         lastCodeExecutionToolCallId = toolCallId;
         content.push({
@@ -1445,7 +1579,7 @@ var GoogleGenerativeAILanguageModel = class {
       } else if ("functionCall" in part) {
         content.push({
           type: "tool-call",
-          toolCallId: (_f = part.functionCall.id) != null ? _f : this.config.generateId(),
+          toolCallId: (_l = part.functionCall.id) != null ? _l : this.config.generateId(),
           toolName: part.functionCall.name,
           input: JSON.stringify(part.functionCall.args),
           providerMetadata: part.thoughtSignature ? { google: { thoughtSignature: part.thoughtSignature } } : void 0
@@ -1458,10 +1592,10 @@ var GoogleGenerativeAILanguageModel = class {
         });
       }
     }
-    const sources = (_g = extractSources({
+    const sources = (_m = extractSources({
       groundingMetadata: candidate == null ? void 0 : candidate.groundingMetadata,
       generateId: this.config.generateId
-    })) != null ? _g : [];
+    })) != null ? _m : [];
     for (const source of sources) {
       content.push(source);
     }
@@ -1472,20 +1606,22 @@ var GoogleGenerativeAILanguageModel = class {
         hasToolCalls: content.some((part) => part.type === "tool-call")
       }),
       usage: {
-        inputTokens: (_h = usageMetadata == null ? void 0 : usageMetadata.promptTokenCount) != null ? _h : void 0,
-        outputTokens: (_i = usageMetadata == null ? void 0 : usageMetadata.candidatesTokenCount) != null ? _i : void 0,
-        totalTokens: (_j = usageMetadata == null ? void 0 : usageMetadata.totalTokenCount) != null ? _j : void 0,
-        reasoningTokens: (_k = usageMetadata == null ? void 0 : usageMetadata.thoughtsTokenCount) != null ? _k : void 0,
-        cachedInputTokens: (_l = usageMetadata == null ? void 0 : usageMetadata.cachedContentTokenCount) != null ? _l : void 0
+        inputTokens: (_n = usageMetadata == null ? void 0 : usageMetadata.promptTokenCount) != null ? _n : void 0,
+        outputTokens: (_o = usageMetadata == null ? void 0 : usageMetadata.candidatesTokenCount) != null ? _o : void 0,
+        totalTokens: (_p = usageMetadata == null ? void 0 : usageMetadata.totalTokenCount) != null ? _p : void 0,
+        reasoningTokens: (_q = usageMetadata == null ? void 0 : usageMetadata.thoughtsTokenCount) != null ? _q : void 0,
+        cachedInputTokens: (_r = usageMetadata == null ? void 0 : usageMetadata.cachedContentTokenCount) != null ? _r : void 0
       },
       warnings,
       providerMetadata: {
         google: {
-          promptFeedback: (_m = response.promptFeedback) != null ? _m : null,
-          groundingMetadata: (_n = candidate == null ? void 0 : candidate.groundingMetadata) != null ? _n : null,
-          urlContextMetadata: (_o = candidate == null ? void 0 : candidate.urlContextMetadata) != null ? _o : null,
-          safetyRatings: (_p = candidate == null ? void 0 : candidate.safetyRatings) != null ? _p : null,
-          serviceTier: (_q = response.serviceTier) != null ? _q : null,
+          promptFeedback: (_s = response.promptFeedback) != null ? _s : null,
+          groundingMetadata: (_t = candidate == null ? void 0 : candidate.groundingMetadata) != null ? _t : null,
+          urlContextMetadata: (_u = candidate == null ? void 0 : candidate.urlContextMetadata) != null ? _u : null,
+          safetyRatings: (_v = candidate == null ? void 0 : candidate.safetyRatings) != null ? _v : null,
+          serviceTier: (_w = response.serviceTier) != null ? _w : null,
+          ...(candidate == null ? void 0 : candidate.continuationToken) != null ? { continuationToken: candidate.continuationToken } : {},
+          ...continuedUsageMetadata != null && firstPromptTokenCount != null ? { initialPromptTokenCount: firstPromptTokenCount } : {},
           usageMetadata: usageMetadata != null ? usageMetadata : null
         }
       },
@@ -1504,17 +1640,27 @@ var GoogleGenerativeAILanguageModel = class {
       await (0, import_provider_utils6.resolve)(this.config.headers),
       options.headers
     );
-    const { responseHeaders, value: response } = await (0, import_provider_utils6.postJsonToApi)({
+    const streamGenerateContent = (body2) => (0, import_provider_utils6.postJsonToApi)({
       url: `${this.config.baseURL}/${getModelPath(
         this.modelId
       )}:streamGenerateContent?alt=sse`,
       headers,
-      body: args,
+      body: body2,
       failedResponseHandler: googleFailedResponseHandler,
       successfulResponseHandler: (0, import_provider_utils6.createEventSourceResponseHandler)(chunkSchema),
       abortSignal: options.abortSignal,
       fetch: this.config.fetch
     });
+    const { responseHeaders, value: firstResponse } = await streamGenerateContent(args);
+    const response = continueGoogleStream({
+      stream: firstResponse,
+      getContinuationToken: (chunk) => {
+        var _a, _b;
+        const candidate = (_a = chunk.candidates) == null ? void 0 : _a[0];
+        return (candidate == null ? void 0 : candidate.finishReason) === "CONTINUATION" ? (_b = candidate.continuationToken) != null ? _b : void 0 : void 0;
+      },
+      continueStream: async (continuationToken) => (await streamGenerateContent({ ...args, continuationToken })).value
+    });
     let finishReason = "unknown";
     const usage = {
       inputTokens: void 0,
@@ -1525,6 +1671,10 @@ var GoogleGenerativeAILanguageModel = class {
     let lastGroundingMetadata = null;
     let lastUrlContextMetadata = null;
     let serviceTier = null;
+    let requestUsageMetadata;
+    let continuedUsageMetadata;
+    let lastContinuationToken = null;
+    let initialPromptTokenCount = void 0;
     const generateId3 = this.config.generateId;
     let hasToolCalls = false;
     let currentTextBlockId = null;
@@ -1548,7 +1698,13 @@ var GoogleGenerativeAILanguageModel = class {
               return;
             }
             const value = chunk.value;
-            const usageMetadata = value.usageMetadata;
+            if (value.usageMetadata != null) {
+              requestUsageMetadata = value.usageMetadata;
+            }
+            const usageMetadata = value.usageMetadata == null ? void 0 : addGoogleGenerativeAIUsageMetadata(
+              continuedUsageMetadata,
+              requestUsageMetadata
+            );
             if (usageMetadata != null) {
               usage.inputTokens = (_a = usageMetadata.promptTokenCount) != null ? _a : void 0;
               usage.outputTokens = (_b = usageMetadata.candidatesTokenCount) != null ? _b : void 0;
@@ -1560,6 +1716,19 @@ var GoogleGenerativeAILanguageModel = class {
               serviceTier = value.serviceTier;
             }
             const candidate = (_f = value.candidates) == null ? void 0 : _f[0];
+            if ((candidate == null ? void 0 : candidate.continuationToken) != null) {
+              lastContinuationToken = candidate.continuationToken;
+            }
+            if ((candidate == null ? void 0 : candidate.finishReason) === "CONTINUATION") {
+              if (continuedUsageMetadata == null) {
+                initialPromptTokenCount = requestUsageMetadata == null ? void 0 : requestUsageMetadata.promptTokenCount;
+              }
+              continuedUsageMetadata = addGoogleGenerativeAIUsageMetadata(
+                continuedUsageMetadata,
+                requestUsageMetadata
+              );
+              requestUsageMetadata = void 0;
+            }
             if (candidate == null) {
               const promptBlockReason2 = (_g = value.promptFeedback) == null ? void 0 : _g.blockReason;
               if (promptBlockReason2 != null) {
@@ -1742,7 +1911,12 @@ var GoogleGenerativeAILanguageModel = class {
                   groundingMetadata: lastGroundingMetadata,
                   urlContextMetadata: lastUrlContextMetadata,
                   safetyRatings: (_o = candidate.safetyRatings) != null ? _o : null,
-                  serviceTier
+                  serviceTier,
+                  // Only surface the token when the response stopped early
+                  // (budget used up, or continuing failed), not for the
+                  // checkpoint tokens of a stream that finished normally.
+                  ...lastContinuationToken != null && (rawFinishReason === "MAX_TOKENS" || rawFinishReason === "CONTINUATION") ? { continuationToken: lastContinuationToken } : {},
+                  ...initialPromptTokenCount != null ? { initialPromptTokenCount } : {}
                 }
               };
               if (usageMetadata != null) {
@@ -2013,6 +2187,7 @@ var responseSchema = (0, import_provider_utils6.lazySchema)(
         import_v45.z.object({
           content: getContentSchema().nullish().or(import_v45.z.object({}).strict()),
           finishReason: import_v45.z.string().nullish(),
+          continuationToken: import_v45.z.string().nullish(),
           safetyRatings: import_v45.z.array(getSafetyRatingSchema()).nullish(),
           groundingMetadata: getGroundingMetadataSchema().nullish(),
           urlContextMetadata: getUrlContextMetadataSchema().nullish()
@@ -2034,6 +2209,7 @@ var chunkSchema = (0, import_provider_utils6.lazySchema)(
         import_v45.z.object({
           content: getContentSchema().nullish(),
           finishReason: import_v45.z.string().nullish(),
+          continuationToken: import_v45.z.string().nullish(),
           safetyRatings: import_v45.z.array(getSafetyRatingSchema()).nullish(),
           groundingMetadata: getGroundingMetadataSchema().nullish(),
           urlContextMetadata: getUrlContextMetadataSchema().nullish()