@ai-sdk/anthropic 2.0.79 → 2.0.81

package/dist/internal/index.mjs

@@ -41,6 +41,12 @@ import {
   zodSchema as zodSchema2
 } from "@ai-sdk/provider-utils";
 import { z as z2 } from "zod/v4";
+var anthropicStopDetailsSchema = z2.object({
+  type: z2.string(),
+  category: z2.string().nullish(),
+  explanation: z2.string().nullish(),
+  recommended_model: z2.string().nullish()
+});
 var anthropicMessagesResponseSchema = lazySchema2(
   () => zodSchema2(
     z2.object({
@@ -228,11 +234,18 @@ var anthropicMessagesResponseSchema = lazySchema2(
                 old_start: z2.number().nullable()
               })
             ])
+          }),
+          // Server-side fallback marker. Parsed so the response validates, but
+          // dropped from the content output (the AI SDK has no model-hop
+          // primitive). The hop remains observable via usage.iterations.
+          z2.object({
+            type: z2.literal("fallback")
           })
         ])
       ),
       stop_reason: z2.string().nullish(),
       stop_sequence: z2.string().nullish(),
+      stop_details: anthropicStopDetailsSchema.nullish(),
       usage: z2.looseObject({
         input_tokens: z2.number(),
         output_tokens: z2.number(),
@@ -240,9 +253,17 @@ var anthropicMessagesResponseSchema = lazySchema2(
         cache_read_input_tokens: z2.number().nullish(),
         iterations: z2.array(
           z2.object({
-            type: z2.union([z2.literal("compaction"), z2.literal("message")]),
+            type: z2.union([
+              z2.literal("compaction"),
+              z2.literal("message"),
+              z2.literal("advisor_message"),
+              z2.literal("fallback_message")
+            ]),
+            model: z2.string().nullish(),
             input_tokens: z2.number(),
-            output_tokens: z2.number()
+            output_tokens: z2.number(),
+            cache_creation_input_tokens: z2.number().nullish(),
+            cache_read_input_tokens: z2.number().nullish()
           })
         ).nullish()
       }),
@@ -448,6 +469,11 @@ var anthropicMessagesChunkSchema = lazySchema2(
                 old_start: z2.number().nullable()
               })
             ])
+          }),
+          // Server-side fallback marker; dropped from content output (see the
+          // response schema). The hop remains observable via usage.iterations.
+          z2.object({
+            type: z2.literal("fallback")
           })
         ])
       }),
@@ -521,6 +547,7 @@ var anthropicMessagesChunkSchema = lazySchema2(
         delta: z2.object({
           stop_reason: z2.string().nullish(),
           stop_sequence: z2.string().nullish(),
+          stop_details: anthropicStopDetailsSchema.nullish(),
           container: z2.object({
             expires_at: z2.string(),
             id: z2.string(),
@@ -543,9 +570,17 @@ var anthropicMessagesChunkSchema = lazySchema2(
           cache_read_input_tokens: z2.number().nullish(),
           iterations: z2.array(
             z2.object({
-              type: z2.union([z2.literal("compaction"), z2.literal("message")]),
+              type: z2.union([
+                z2.literal("compaction"),
+                z2.literal("message"),
+                z2.literal("advisor_message"),
+                z2.literal("fallback_message")
+              ]),
+              model: z2.string().nullish(),
               input_tokens: z2.number(),
-              output_tokens: z2.number()
+              output_tokens: z2.number(),
+              cache_creation_input_tokens: z2.number().nullish(),
+              cache_read_input_tokens: z2.number().nullish()
             })
           ).nullish()
         }),
@@ -720,6 +755,31 @@ var anthropicProviderOptions = z3.object({
    * See https://platform.claude.com/docs/en/build-with-claude/data-residency
    */
   inferenceGeo: z3.enum(["us", "global"]).optional(),
+  /**
+   * Server-side fallback chain.
+   *
+   * When the primary model's safety classifiers block a turn, the API
+   * automatically retries it on the next model in the chain, server-side. A
+   * `content-filter` finish reason means the entire chain refused.
+   *
+   * Each entry is merged into the request as a direct request to that entry's
+   * model, so it must be formatted accordingly: `model` is required, and an
+   * entry may additionally override `max_tokens`, `thinking`, `output_config`,
+   * and `speed` for that attempt only (`speed` additionally requires the speed
+   * beta). The value is passed through to the API as-is.
+   *
+   * The required `server-side-fallback-2026-06-01` beta is added automatically
+   * when this option is set.
+   */
+  fallbacks: z3.array(
+    z3.object({
+      model: z3.string(),
+      max_tokens: z3.number().int().optional(),
+      thinking: z3.record(z3.string(), z3.unknown()).optional(),
+      output_config: z3.record(z3.string(), z3.unknown()).optional(),
+      speed: z3.enum(["fast", "standard"]).optional()
+    })
+  ).optional(),
   /**
    * Context management configuration for automatic context window management.
    * Enables features like automatic compaction and clearing of tool uses/thinking blocks.
@@ -1418,19 +1478,20 @@ async function convertToAnthropicMessagesPrompt({
     const type = block.type;
     switch (type) {
       case "system": {
-        if (system != null) {
-          throw new UnsupportedFunctionalityError2({
-            functionality: "Multiple system messages that are separated by user/assistant messages"
-          });
-        }
-        system = block.messages.map(({ content, providerOptions }) => ({
+        const content = block.messages.map(({ content: content2, providerOptions }) => ({
           type: "text",
-          text: content,
+          text: content2,
           cache_control: validator.getCacheControl(providerOptions, {
             type: "system message",
             canCache: true
           })
         }));
+        if (system == null) {
+          system = content;
+        } else {
+          messages.push({ role: "system", content });
+          betas.add("mid-conversation-system-2026-04-07");
+        }
         break;
       }
       case "user": {
@@ -2307,6 +2368,9 @@ var AnthropicMessagesLanguageModel = class {
       ...(anthropicOptions == null ? void 0 : anthropicOptions.inferenceGeo) && {
         inference_geo: anthropicOptions.inferenceGeo
       },
+      ...(anthropicOptions == null ? void 0 : anthropicOptions.fallbacks) && anthropicOptions.fallbacks.length > 0 && {
+        fallbacks: anthropicOptions.fallbacks
+      },
       ...(anthropicOptions == null ? void 0 : anthropicOptions.cacheControl) && {
         cache_control: anthropicOptions.cacheControl
       },
@@ -2449,6 +2513,9 @@ var AnthropicMessagesLanguageModel = class {
     if ((anthropicOptions == null ? void 0 : anthropicOptions.speed) === "fast") {
       betas.add("fast-mode-2026-02-01");
     }
+    if ((anthropicOptions == null ? void 0 : anthropicOptions.fallbacks) && anthropicOptions.fallbacks.length > 0) {
+      betas.add("server-side-fallback-2026-06-01");
+    }
     if (useStructuredOutput) {
       betas.add("structured-outputs-2025-11-13");
     }
@@ -2548,7 +2615,7 @@ var AnthropicMessagesLanguageModel = class {
     });
   }
   async doGenerate(options) {
-    var _a, _b, _c, _d, _e, _f, _g, _h;
+    var _a, _b, _c, _d, _e, _f, _g, _h, _i;
     const { args, warnings, betas, usesJsonResponseTool } = await this.getArgs({
       ...options,
       userSuppliedBetas: await this.getBetasFromHeaders(options.headers)
@@ -2788,11 +2855,20 @@ var AnthropicMessagesLanguageModel = class {
           });
           break;
         }
+        // Server-side fallback marker: the AI SDK has no content primitive for
+        // a model hop, so drop it. The hop is still observable via
+        // usage.iterations.
+        case "fallback": {
+          break;
+        }
       }
     }
     let inputTokens;
     let outputTokens;
-    if (response.usage.iterations && response.usage.iterations.length > 0) {
+    const servedByFallback = (_b = response.usage.iterations) == null ? void 0 : _b.some(
+      (iter) => iter.type === "fallback_message"
+    );
+    if (response.usage.iterations && response.usage.iterations.length > 0 && !servedByFallback) {
       const totals = response.usage.iterations.reduce(
         (acc, iter) => ({
           input: acc.input + iter.input_tokens,
@@ -2806,6 +2882,7 @@ var AnthropicMessagesLanguageModel = class {
       inputTokens = response.usage.input_tokens;
       outputTokens = response.usage.output_tokens;
     }
+    const stopDetails = mapAnthropicStopDetails(response.stop_details);
     return {
       content,
       finishReason: mapAnthropicStopReason({
@@ -2816,12 +2893,12 @@ var AnthropicMessagesLanguageModel = class {
         inputTokens,
         outputTokens,
         totalTokens: inputTokens + outputTokens,
-        cachedInputTokens: (_b = response.usage.cache_read_input_tokens) != null ? _b : void 0
+        cachedInputTokens: (_c = response.usage.cache_read_input_tokens) != null ? _c : void 0
       },
       request: { body: args },
       response: {
-        id: (_c = response.id) != null ? _c : void 0,
-        modelId: (_d = response.model) != null ? _d : void 0,
+        id: (_d = response.id) != null ? _d : void 0,
+        modelId: (_e = response.model) != null ? _e : void 0,
         headers: responseHeaders,
         body: rawResponse
       },
@@ -2829,21 +2906,23 @@ var AnthropicMessagesLanguageModel = class {
       providerMetadata: {
         anthropic: {
           usage: response.usage,
-          cacheCreationInputTokens: (_e = response.usage.cache_creation_input_tokens) != null ? _e : null,
-          stopSequence: (_f = response.stop_sequence) != null ? _f : null,
+          cacheCreationInputTokens: (_f = response.usage.cache_creation_input_tokens) != null ? _f : null,
+          stopSequence: (_g = response.stop_sequence) != null ? _g : null,
+          ...stopDetails != null ? { stopDetails } : {},
           iterations: response.usage.iterations ? response.usage.iterations.map((iter) => ({
             type: iter.type,
+            ...iter.model != null ? { model: iter.model } : {},
             inputTokens: iter.input_tokens,
             outputTokens: iter.output_tokens
           })) : null,
           container: response.container ? {
             expiresAt: response.container.expires_at,
             id: response.container.id,
-            skills: (_h = (_g = response.container.skills) == null ? void 0 : _g.map((skill) => ({
+            skills: (_i = (_h = response.container.skills) == null ? void 0 : _h.map((skill) => ({
               type: skill.type,
               skillId: skill.skill_id,
               version: skill.version
-            }))) != null ? _h : null
+            }))) != null ? _i : null
           } : null,
           contextManagement: response.context_management ? {
             appliedEdits: response.context_management.applied_edits.map(
@@ -2904,6 +2983,7 @@ var AnthropicMessagesLanguageModel = class {
     let rawUsage = void 0;
     let cacheCreationInputTokens = null;
     let stopSequence = null;
+    let stopDetails = void 0;
     let container = null;
     let iterations = null;
     let contextManagement = null;
@@ -2915,7 +2995,7 @@ var AnthropicMessagesLanguageModel = class {
           controller.enqueue({ type: "stream-start", warnings });
         },
         transform(chunk, controller) {
-          var _a2, _b2, _c, _d, _e, _f, _g, _h, _i, _j;
+          var _a2, _b2, _c, _d, _e, _f, _g, _h, _i, _j, _k;
           if (options.includeRawChunks) {
             controller.enqueue({ type: "raw", rawValue: chunk.rawValue });
           }
@@ -2930,6 +3010,9 @@ var AnthropicMessagesLanguageModel = class {
             }
             case "content_block_start": {
               const contentBlockType = value.content_block.type;
+              if (contentBlockType === "fallback") {
+                return;
+              }
               blockType = contentBlockType;
               switch (contentBlockType) {
                 case "text": {
@@ -3322,7 +3405,10 @@ var AnthropicMessagesLanguageModel = class {
               if (value.usage.iterations != null) {
                 iterations = value.usage.iterations;
               }
-              if (value.usage.iterations && value.usage.iterations.length > 0) {
+              const servedByFallback = (_f = value.usage.iterations) == null ? void 0 : _f.some(
+                (iter) => iter.type === "fallback_message"
+              );
+              if (value.usage.iterations && value.usage.iterations.length > 0 && !servedByFallback) {
                 const totals = value.usage.iterations.reduce(
                   (acc, iter) => ({
                     input: acc.input + iter.input_tokens,
@@ -3338,20 +3424,21 @@ var AnthropicMessagesLanguageModel = class {
                 }
                 usage.outputTokens = value.usage.output_tokens;
               }
-              usage.totalTokens = ((_f = usage.inputTokens) != null ? _f : 0) + ((_g = usage.outputTokens) != null ? _g : 0);
+              usage.totalTokens = ((_g = usage.inputTokens) != null ? _g : 0) + ((_h = usage.outputTokens) != null ? _h : 0);
               finishReason = mapAnthropicStopReason({
                 finishReason: value.delta.stop_reason,
                 isJsonResponseFromTool: usesJsonResponseTool
               });
-              stopSequence = (_h = value.delta.stop_sequence) != null ? _h : null;
+              stopDetails = mapAnthropicStopDetails(value.delta.stop_details);
+              stopSequence = (_i = value.delta.stop_sequence) != null ? _i : null;
               container = value.delta.container != null ? {
                 expiresAt: value.delta.container.expires_at,
                 id: value.delta.container.id,
-                skills: (_j = (_i = value.delta.container.skills) == null ? void 0 : _i.map((skill) => ({
+                skills: (_k = (_j = value.delta.container.skills) == null ? void 0 : _j.map((skill) => ({
                   type: skill.type,
                   skillId: skill.skill_id,
                   version: skill.version
-                }))) != null ? _j : null
+                }))) != null ? _k : null
               } : null;
               if (value.context_management != null) {
                 contextManagement = {
@@ -3396,8 +3483,10 @@ var AnthropicMessagesLanguageModel = class {
                     usage: rawUsage != null ? rawUsage : null,
                     cacheCreationInputTokens,
                     stopSequence,
+                    ...stopDetails != null ? { stopDetails } : {},
                     iterations: iterations ? iterations.map((iter) => ({
                       type: iter.type,
+                      ...iter.model != null ? { model: iter.model } : {},
                       inputTokens: iter.input_tokens,
                       outputTokens: iter.output_tokens
                     })) : null,
@@ -3453,7 +3542,7 @@ var AnthropicMessagesLanguageModel = class {
   }
 };
 function getModelCapabilities(modelId) {
-  if (modelId.includes("claude-opus-4-7")) {
+  if (modelId.includes("claude-opus-4-8") || modelId.includes("claude-opus-4-7") || modelId.includes("claude-fable-5")) {
     return {
       maxOutputTokens: 128e3,
       supportsStructuredOutput: true,
@@ -3518,6 +3607,17 @@ function getModelCapabilities(modelId) {
     };
   }
 }
+function mapAnthropicStopDetails(stopDetails) {
+  if (stopDetails == null) {
+    return void 0;
+  }
+  return {
+    type: stopDetails.type,
+    ...stopDetails.category != null ? { category: stopDetails.category } : {},
+    ...stopDetails.explanation != null ? { explanation: stopDetails.explanation } : {},
+    ...stopDetails.recommended_model != null ? { recommendedModel: stopDetails.recommended_model } : {}
+  };
+}
 
 // src/tool/bash_20241022.ts
 import {