fix: support thinking with structured output in AI SDK runner

package.json

@@ -54,6 +54,7 @@
     "@ai-sdk/anthropic": "^2.0.45",
     "@ai-sdk/google": "^2.0.39",
     "@ai-sdk/openai": "^2.0.71",
+    "@ai-sdk/provider": "^2.0.0",
     "@anthropic-ai/sdk": "^0.68.0",
     "@axe-core/puppeteer": "^4.10.2",
     "@genkit-ai/compat-oai": "1.23.0",

runner/codegen/ai-sdk-claude-thinking-patch.ts

@@ -0,0 +1,36 @@
+import type {LanguageModelV2Middleware} from '@ai-sdk/provider';
+
+/**
+ * Middleware for Anthropic AI SDK models that is necessary for enabling
+ * thinking mode + structured responses.
+ *
+ * This is necessary because Anthropic would be used with enforced tool usage
+ * by default with `generateObject()`. This is a workaround that makes the tool
+ * optional: https://github.com/vercel/ai/issues/9351.
+ */
+export const anthropicThinkingWithStructuredResponseMiddleware: LanguageModelV2Middleware = {
+  transformParams: ({params}) => {
+    if (params.responseFormat?.type === 'json' && params.responseFormat.schema) {
+      params.tools = [
+        {
+          type: 'function',
+          description: 'Respond with a JSON object for the structured output/answer.',
+          inputSchema: params.responseFormat.schema,
+          name: 'json',
+        },
+      ];
+      params.toolChoice = {type: 'auto'};
+
+      params.prompt.push({
+        role: 'user',
+        content: [
+          {
+            type: 'text',
+            text: 'Use the `json` tool to provide the structured output/answer. No other text is needed.',
+          },
+        ],
+      });
+    }
+    return Promise.resolve(params);
+  },
+};

runner/codegen/ai-sdk-runner.ts

@@ -16,21 +16,26 @@ import {
   ModelMessage,
   SystemModelMessage,
   TextPart,
+  wrapLanguageModel,
 } from 'ai';
 import {google, GoogleGenerativeAIProviderOptions} from '@ai-sdk/google';
 import {anthropic, AnthropicProviderOptions} from '@ai-sdk/anthropic';
 import {openai, OpenAIResponsesProviderOptions} from '@ai-sdk/openai';
 import z from 'zod';
 import {callWithTimeout} from '../utils/timeout.js';
 import {combineAbortSignals} from '../utils/abort-signal.js';
+import {anthropicThinkingWithStructuredResponseMiddleware} from './ai-sdk-claude-thinking-patch.js';
 
 const SUPPORTED_MODELS = [
   'claude-opus-4.1-no-thinking',
   'claude-opus-4.1-with-thinking-16k',
+  'claude-opus-4.1-with-thinking-32k',
   'claude-sonnet-4.5-no-thinking',
   'claude-sonnet-4.5-with-thinking-16k',
+  'claude-sonnet-4.5-with-thinking-32k',
   'gemini-2.5-flash-lite',
   'gemini-2.5-flash',
+  'gemini-2.5-flash-with-dynamic-thinking',
   'gemini-2.5-pro',
   'gemini-3-pro-preview',
   'gpt-5.1-no-thinking',
@@ -45,6 +50,7 @@ const SUPPORTED_MODELS = [
 const DEFAULT_MAX_RETRIES = 100000;
 
 const claude16kThinkingTokenBudget = 16_000;
+const claude32kThinkingTokenBudget = 32_000;
 export class AiSDKRunner implements LlmRunner {
   displayName = 'AI SDK';
   id = 'ai-sdk';
@@ -159,32 +165,32 @@ export class AiSDKRunner implements LlmRunner {
     const modelName = request.model as (typeof SUPPORTED_MODELS)[number];
     switch (modelName) {
       case 'claude-opus-4.1-no-thinking':
-      case 'claude-opus-4.1-with-thinking-16k': {
-        const thinkingEnabled = modelName.includes('-with-thinking');
-        return {
-          model: anthropic('claude-opus-4-1'),
-          providerOptions: {
-            anthropic: {
-              sendReasoning: thinkingEnabled,
-              thinking: {
-                type: thinkingEnabled ? 'enabled' : 'disabled',
-                budgetTokens: thinkingEnabled ? claude16kThinkingTokenBudget : undefined,
-              },
-            } satisfies AnthropicProviderOptions,
-          },
-        };
-      }
+      case 'claude-opus-4.1-with-thinking-16k':
+      case 'claude-opus-4.1-with-thinking-32k':
       case 'claude-sonnet-4.5-no-thinking':
-      case 'claude-sonnet-4.5-with-thinking-16k': {
+      case 'claude-sonnet-4.5-with-thinking-16k':
+      case 'claude-sonnet-4.5-with-thinking-32k': {
         const thinkingEnabled = modelName.includes('-with-thinking');
+        const thinkingBudget = !thinkingEnabled
+          ? undefined
+          : modelName.endsWith('-32k')
+            ? claude32kThinkingTokenBudget
+            : claude16kThinkingTokenBudget;
+        const isOpus4_1Model = modelName.includes('opus-4.1');
+        const model = anthropic(isOpus4_1Model ? 'claude-opus-4-1' : 'claude-sonnet-4-5');
         return {
-          model: anthropic('claude-sonnet-4-5'),
+          model: thinkingEnabled
+            ? wrapLanguageModel({
+                model,
+                middleware: anthropicThinkingWithStructuredResponseMiddleware,
+              })
+            : model,
           providerOptions: {
             anthropic: {
               sendReasoning: thinkingEnabled,
               thinking: {
                 type: thinkingEnabled ? 'enabled' : 'disabled',
-                budgetTokens: thinkingEnabled ? claude16kThinkingTokenBudget : undefined,
+                budgetTokens: thinkingBudget,
               },
             } satisfies AnthropicProviderOptions,
           },
@@ -204,6 +210,20 @@ export class AiSDKRunner implements LlmRunner {
             } satisfies GoogleGenerativeAIProviderOptions,
           },
         };
+      case 'gemini-2.5-flash-with-dynamic-thinking':
+        return {
+          model: google(modelName),
+          providerOptions: {
+            google: {
+              thinkingConfig: {
+                // -1 means "dynamic thinking budget":
+                // https://ai.google.dev/gemini-api/docs/thinking#set-budget.
+                thinkingBudget: -1,
+                includeThoughts: true,
+              },
+            } satisfies GoogleGenerativeAIProviderOptions,
+          },
+        };
       case 'gpt-5.1-no-thinking':
       case 'gpt-5.1-thinking-low':
       case 'gpt-5.1-thinking-medium':