@assemble-dev/providers 0.3.0 → 0.3.1

package/dist/index.cjs

@@ -72,7 +72,9 @@ __export(index_exports, {
   extractMessageText: () => extractMessageText,
   gemini: () => gemini,
   isAbortError: () => isAbortError,
+  isOpenAiReasoningModel: () => isOpenAiReasoningModel,
   isRetryableProviderAppError: () => isRetryableProviderAppError,
+  isZodV4Schema: () => isZodV4Schema,
   joinSystemMessageText: () => joinSystemMessageText,
   mapAnthropicResponseToChatCompletionResult: () => mapAnthropicResponseToChatCompletionResult,
   mapAnthropicUsageToTokenUsage: () => mapAnthropicUsageToTokenUsage,
@@ -95,15 +97,19 @@ __export(index_exports, {
   parseAnthropicToolInputJson: () => parseAnthropicToolInputJson,
   parseAzureOpenAiStreamEvents: () => parseAzureOpenAiStreamEvents,
   parseAzureOpenAiToolArgumentsJson: () => parseAzureOpenAiToolArgumentsJson,
+  prependStructuredOutputJsonOnlySystemMessage: () => prependStructuredOutputJsonOnlySystemMessage,
   resolveAnthropicPromptCachingSettings: () => resolveAnthropicPromptCachingSettings,
   resolveBedrockCredentials: () => resolveBedrockCredentials,
+  resolveOpenAiStructuredOutputMode: () => resolveOpenAiStructuredOutputMode,
   resolveProviderApiKey: () => resolveProviderApiKey,
   sanitizeJsonSchemaForAnthropicStructuredOutput: () => sanitizeJsonSchemaForAnthropicStructuredOutput,
   sanitizeJsonSchemaForGeminiResponseSchema: () => sanitizeJsonSchemaForGeminiResponseSchema,
   sanitizeJsonSchemaForOpenAiStructuredOutput: () => sanitizeJsonSchemaForOpenAiStructuredOutput,
   signAwsRequestWithSigV4: () => signAwsRequestWithSigV4,
   streamAnthropicChatCompletion: () => streamAnthropicChatCompletion,
-  streamAzureOpenAiChatCompletion: () => streamAzureOpenAiChatCompletion
+  streamAzureOpenAiChatCompletion: () => streamAzureOpenAiChatCompletion,
+  structuredOutputJsonOnlySystemInstruction: () => structuredOutputJsonOnlySystemInstruction,
+  structuredOutputJsonOnlySystemMessage: () => structuredOutputJsonOnlySystemMessage
 });
 module.exports = __toCommonJS(index_exports);
 
@@ -313,6 +319,9 @@ var openAiUnsupportedSchemaKeywords = [
   "uniqueItems",
   "default"
 ];
+function isZodV4Schema(schema) {
+  return "_zod" in schema;
+}
 function convertZodSchemaToJsonSchema(schema) {
   try {
     return import_json2.JsonObjectSchema.parse(import_zod2.z.toJSONSchema(schema, { io: "input" }));
@@ -377,18 +386,72 @@ function listSchemaPropertyNames(node) {
   return Object.keys(properties);
 }
 
+// src/structured-output-fallback.ts
+var structuredOutputJsonOnlySystemInstruction = "Respond with a single JSON object that satisfies the caller's requested structure. Output only valid JSON with no surrounding text.";
+var structuredOutputJsonOnlySystemMessage = {
+  role: "system",
+  content: structuredOutputJsonOnlySystemInstruction
+};
+function prependStructuredOutputJsonOnlySystemMessage(messages) {
+  return [structuredOutputJsonOnlySystemMessage, ...messages];
+}
+
 // src/openai-request-mapping.ts
+var openAiReasoningModelNamePrefixes = [
+  "gpt-5",
+  "o1",
+  "o3",
+  "o4"
+];
+function isOpenAiReasoningModel(model) {
+  return openAiReasoningModelNamePrefixes.some(
+    (prefix) => matchesOpenAiModelNamePrefix(model, prefix)
+  );
+}
+function matchesOpenAiModelNamePrefix(model, prefix) {
+  if (!model.startsWith(prefix)) {
+    return false;
+  }
+  const boundaryCharacter = model.charAt(prefix.length);
+  return boundaryCharacter === "" || boundaryCharacter === "-" || boundaryCharacter === ".";
+}
+function resolveOpenAiStructuredOutputMode(messages, schema) {
+  if (isZodV4Schema(schema)) {
+    return {
+      messages,
+      structuredOutputJsonSchema: buildOpenAiStructuredOutputJsonSchema(schema)
+    };
+  }
+  return {
+    messages: prependStructuredOutputJsonOnlySystemMessage(messages),
+    requireJsonObjectResponse: true
+  };
+}
 function buildOpenAiRequestBody(model, input) {
   const requestBody = {
     model,
     messages: input.messages.map(mapChatMessageToOpenAiRequestMessage)
   };
-  if (input.temperature !== void 0) {
+  applyOpenAiGenerationSettings(requestBody, model, input);
+  applyOpenAiResponseFormat(requestBody, input);
+  applyOpenAiToolFields(requestBody, input);
+  return requestBody;
+}
+function applyOpenAiGenerationSettings(requestBody, model, input) {
+  const reasoningModel = input.reasoningModel ?? isOpenAiReasoningModel(model);
+  if (input.temperature !== void 0 && !reasoningModel) {
     requestBody.temperature = input.temperature;
   }
-  if (input.maxOutputTokens !== void 0) {
-    requestBody.max_tokens = input.maxOutputTokens;
+  if (input.maxOutputTokens === void 0) {
+    return;
+  }
+  if (reasoningModel) {
+    requestBody.max_completion_tokens = input.maxOutputTokens;
+    return;
   }
+  requestBody.max_tokens = input.maxOutputTokens;
+}
+function applyOpenAiResponseFormat(requestBody, input) {
   if (input.structuredOutputJsonSchema !== void 0) {
     requestBody.response_format = {
       type: "json_schema",
@@ -398,9 +461,13 @@ function buildOpenAiRequestBody(model, input) {
         schema: input.structuredOutputJsonSchema
       }
     };
-  } else if (input.requireJsonObjectResponse === true) {
+    return;
+  }
+  if (input.requireJsonObjectResponse === true) {
     requestBody.response_format = { type: "json_object" };
   }
+}
+function applyOpenAiToolFields(requestBody, input) {
   if (input.tools !== void 0 && input.tools.length > 0) {
     requestBody.tools = input.tools.map(mapModelToolDefinitionToOpenAiTool);
   }
@@ -409,7 +476,6 @@ function buildOpenAiRequestBody(model, input) {
       input.toolChoice
     );
   }
-  return requestBody;
 }
 function mapModelToolDefinitionToOpenAiTool(tool) {
   return {
@@ -770,12 +836,9 @@ function openai(options) {
     },
     async generateStructuredOutput(request) {
       const completion = await executeOpenAiAdapterCall(options, {
-        messages: request.messages,
+        ...resolveOpenAiStructuredOutputMode(request.messages, request.schema),
         temperature: request.temperature ?? options.temperature,
         maxOutputTokens: request.maxOutputTokens ?? options.maxOutputTokens,
-        structuredOutputJsonSchema: buildOpenAiStructuredOutputJsonSchema(
-          request.schema
-        ),
         abortSignal: request.abortSignal
       });
       return {
@@ -1543,15 +1606,10 @@ function anthropic(options) {
       );
     },
     async generateStructuredOutput(request) {
-      const completion = await executeAnthropicAdapterCall(options, {
-        messages: request.messages,
-        temperature: request.temperature ?? options.temperature,
-        maxOutputTokens: request.maxOutputTokens ?? options.maxOutputTokens,
-        structuredOutputJsonSchema: buildAnthropicStructuredOutputJsonSchema(
-          request.schema
-        ),
-        abortSignal: request.abortSignal
-      });
+      const completion = await executeAnthropicAdapterCall(
+        options,
+        buildAnthropicStructuredOutputInput(options, request)
+      );
       return {
         value: parseStructuredJsonText(completion.content, request.schema),
         tokenUsage: completion.tokenUsage,
@@ -1565,6 +1623,26 @@ function anthropic(options) {
     }
   };
 }
+function buildAnthropicStructuredOutputInput(options, request) {
+  const baseInput = {
+    messages: request.messages,
+    temperature: request.temperature ?? options.temperature,
+    maxOutputTokens: request.maxOutputTokens ?? options.maxOutputTokens,
+    abortSignal: request.abortSignal
+  };
+  if (isZodV4Schema(request.schema)) {
+    return {
+      ...baseInput,
+      structuredOutputJsonSchema: buildAnthropicStructuredOutputJsonSchema(
+        request.schema
+      )
+    };
+  }
+  return {
+    ...baseInput,
+    messages: prependStructuredOutputJsonOnlySystemMessage(request.messages)
+  };
+}
 function buildAnthropicCallInput(options, request) {
   return {
     messages: request.messages,
@@ -1789,13 +1867,10 @@ function gemini(options) {
       });
     },
     async generateStructuredOutput(request) {
-      const completion = await executeGeminiAdapterCall(options, {
-        messages: request.messages,
-        temperature: request.temperature ?? options.temperature,
-        maxOutputTokens: request.maxOutputTokens ?? options.maxOutputTokens,
-        responseSchema: buildGeminiResponseSchema(request.schema),
-        abortSignal: request.abortSignal
-      });
+      const completion = await executeGeminiAdapterCall(
+        options,
+        buildGeminiStructuredOutputInput(options, request)
+      );
       return {
         value: parseStructuredJsonText(completion.content, request.schema),
         tokenUsage: completion.tokenUsage,
@@ -1804,6 +1879,25 @@ function gemini(options) {
     }
   };
 }
+function buildGeminiStructuredOutputInput(options, request) {
+  const baseInput = {
+    messages: request.messages,
+    temperature: request.temperature ?? options.temperature,
+    maxOutputTokens: request.maxOutputTokens ?? options.maxOutputTokens,
+    abortSignal: request.abortSignal
+  };
+  if (isZodV4Schema(request.schema)) {
+    return {
+      ...baseInput,
+      responseSchema: buildGeminiResponseSchema(request.schema)
+    };
+  }
+  return {
+    ...baseInput,
+    messages: prependStructuredOutputJsonOnlySystemMessage(request.messages),
+    requireJsonResponse: true
+  };
+}
 async function executeGeminiAdapterCall(options, input) {
   return await executeProviderCallWithModelFallback(
     options.fallbackModel,
@@ -1921,6 +2015,8 @@ function buildGeminiGenerationConfig(input) {
   if (input.responseSchema !== void 0) {
     generationConfig.responseMimeType = "application/json";
     generationConfig.responseSchema = input.responseSchema;
+  } else if (input.requireJsonResponse === true) {
+    generationConfig.responseMimeType = "application/json";
   }
   if (Object.keys(generationConfig).length === 0) {
     return void 0;
@@ -2128,7 +2224,7 @@ var import_errors29 = require("@assemble-dev/shared-utils/errors");
 var import_errors26 = require("@assemble-dev/shared-types/errors");
 var import_json7 = require("@assemble-dev/shared-types/json");
 var import_errors27 = require("@assemble-dev/shared-utils/errors");
-var anthropicBatchJsonOnlySystemInstruction = "Respond with a single JSON object that satisfies the caller's requested structure. Output only valid JSON with no surrounding text.";
+var anthropicBatchJsonOnlySystemInstruction = structuredOutputJsonOnlySystemInstruction;
 function parseAnthropicBatchStructuredContent(content, schema) {
   const jsonValue = parseStructuredJsonText(
     content,
@@ -2935,6 +3031,7 @@ function azureOpenai(options) {
         messages: request.messages,
         temperature: request.temperature ?? options.temperature,
         maxOutputTokens: request.maxOutputTokens ?? options.maxOutputTokens,
+        reasoningModel: options.reasoningModel,
         tools: request.tools,
         toolChoice: request.toolChoice,
         abortSignal: request.abortSignal,
@@ -2943,12 +3040,10 @@ function azureOpenai(options) {
     },
     async generateStructuredOutput(request) {
       const completion = await executeAzureOpenAiAdapterCall(options, {
-        messages: request.messages,
+        ...resolveOpenAiStructuredOutputMode(request.messages, request.schema),
         temperature: request.temperature ?? options.temperature,
         maxOutputTokens: request.maxOutputTokens ?? options.maxOutputTokens,
-        structuredOutputJsonSchema: buildOpenAiStructuredOutputJsonSchema(
-          request.schema
-        ),
+        reasoningModel: options.reasoningModel,
         abortSignal: request.abortSignal
       });
       return {
@@ -3047,6 +3142,7 @@ function buildAzureOpenAiStreamInput(options, request) {
       messages: request.messages,
       temperature: request.temperature ?? options.temperature,
       maxOutputTokens: request.maxOutputTokens ?? options.maxOutputTokens,
+      reasoningModel: options.reasoningModel,
       tools: request.tools,
       toolChoice: request.toolChoice,
       abortSignal: request.abortSignal,
@@ -3341,6 +3437,9 @@ function buildBedrockInvokeModelRequestBody(input) {
   };
 }
 function buildBedrockStructuredOutputSystemMessage(schema) {
+  if (!isZodV4Schema(schema)) {
+    return structuredOutputJsonOnlySystemMessage;
+  }
   return {
     role: "system",
     content: `Respond with a single JSON object that satisfies the following JSON Schema. Output only valid JSON with no surrounding text.
@@ -3403,7 +3502,9 @@ ${JSON.stringify(convertZodSchemaToJsonSchema(schema))}`
   extractMessageText,
   gemini,
   isAbortError,
+  isOpenAiReasoningModel,
   isRetryableProviderAppError,
+  isZodV4Schema,
   joinSystemMessageText,
   mapAnthropicResponseToChatCompletionResult,
   mapAnthropicUsageToTokenUsage,
@@ -3426,13 +3527,17 @@ ${JSON.stringify(convertZodSchemaToJsonSchema(schema))}`
   parseAnthropicToolInputJson,
   parseAzureOpenAiStreamEvents,
   parseAzureOpenAiToolArgumentsJson,
+  prependStructuredOutputJsonOnlySystemMessage,
   resolveAnthropicPromptCachingSettings,
   resolveBedrockCredentials,
+  resolveOpenAiStructuredOutputMode,
   resolveProviderApiKey,
   sanitizeJsonSchemaForAnthropicStructuredOutput,
   sanitizeJsonSchemaForGeminiResponseSchema,
   sanitizeJsonSchemaForOpenAiStructuredOutput,
   signAwsRequestWithSigV4,
   streamAnthropicChatCompletion,
-  streamAzureOpenAiChatCompletion
+  streamAzureOpenAiChatCompletion,
+  structuredOutputJsonOnlySystemInstruction,
+  structuredOutputJsonOnlySystemMessage
 });

package/dist/index.d.cts

@@ -357,12 +357,17 @@ declare function buildProviderCanceledError(providerLabel: string): AppError;
 declare function executeAbortableProviderFetch(providerLabel: string, performFetch: () => Promise<Response>): Promise<Response>;
 
 declare const openAiStructuredOutputSchemaName = "structured_output";
+declare function isZodV4Schema<T>(schema: z.ZodType<T>): boolean;
 declare function convertZodSchemaToJsonSchema<T>(schema: z.ZodType<T>): JsonObject;
 declare function buildAnthropicStructuredOutputJsonSchema<T>(schema: z.ZodType<T>): JsonObject;
 declare function sanitizeJsonSchemaForAnthropicStructuredOutput(jsonSchema: JsonObject): JsonObject;
 declare function buildOpenAiStructuredOutputJsonSchema<T>(schema: z.ZodType<T>): JsonObject;
 declare function sanitizeJsonSchemaForOpenAiStructuredOutput(jsonSchema: JsonObject): JsonObject;
 
+declare const structuredOutputJsonOnlySystemInstruction = "Respond with a single JSON object that satisfies the caller's requested structure. Output only valid JSON with no surrounding text.";
+declare const structuredOutputJsonOnlySystemMessage: ChatMessage;
+declare function prependStructuredOutputJsonOnlySystemMessage(messages: ChatMessage[]): ChatMessage[];
+
 declare function buildGeminiResponseSchema<T>(schema: z.ZodType<T>): JsonObject;
 declare function sanitizeJsonSchemaForGeminiResponseSchema(jsonSchema: JsonObject): JsonObject;
 
@@ -539,6 +544,7 @@ type OpenAiRequestBody = {
     messages: OpenAiRequestMessage[];
     temperature?: number;
     max_tokens?: number;
+    max_completion_tokens?: number;
     response_format?: OpenAiResponseFormat;
     tools?: OpenAiToolDefinition[];
     tool_choice?: OpenAiToolChoice;
@@ -547,6 +553,7 @@ type OpenAiCallInput = {
     messages: ChatMessage[];
     temperature?: number;
     maxOutputTokens?: number;
+    reasoningModel?: boolean;
     structuredOutputJsonSchema?: JsonObject;
     requireJsonObjectResponse?: boolean;
     tools?: ModelToolDefinition[];
@@ -554,6 +561,13 @@ type OpenAiCallInput = {
     abortSignal?: AbortSignal;
     extraBody?: Record<string, JsonValue>;
 };
+type OpenAiStructuredOutputMode = {
+    messages: ChatMessage[];
+    structuredOutputJsonSchema?: JsonObject;
+    requireJsonObjectResponse?: boolean;
+};
+declare function isOpenAiReasoningModel(model: string): boolean;
+declare function resolveOpenAiStructuredOutputMode<T>(messages: ChatMessage[], schema: z.ZodType<T>): OpenAiStructuredOutputMode;
 declare function buildOpenAiRequestBody(model: string, input: OpenAiCallInput): OpenAiRequestBody;
 declare function mapModelToolDefinitionToOpenAiTool(tool: ModelToolDefinition): OpenAiToolDefinition;
 declare function mapModelToolChoiceToOpenAiToolChoice(toolChoice: ModelToolChoice): OpenAiToolChoice;
@@ -695,6 +709,7 @@ type AzureOpenAiAdapterOptions = {
     fetchImplementation?: typeof fetch;
     temperature?: number;
     maxOutputTokens?: number;
+    reasoningModel?: boolean;
     retry?: ProviderRetryOptions;
     extraBody?: Record<string, JsonValue>;
 };
@@ -791,4 +806,4 @@ declare function buildBedrockInvokeModelUrl(region: string, modelId: string): st
 declare function resolveBedrockCredentials(options: BedrockAdapterOptions): BedrockCredentials;
 declare function buildBedrockInvokeModelRequestBody(input: AnthropicRequestInput): BedrockInvokeModelRequestBody;
 
-export { type AnthropicAdapterOptions, type AnthropicBatchApiRequest, type AnthropicBatchApiRequestMessage, type AnthropicBatchApiRequestParams, type AnthropicBatchCanceledOutcome, type AnthropicBatchClient, type AnthropicBatchClientOptions, type AnthropicBatchErroredOutcome, type AnthropicBatchExpiredOutcome, type AnthropicBatchMessageRequest, type AnthropicBatchPromptCachingMode, type AnthropicBatchResult, type AnthropicBatchResultOutcome, type AnthropicBatchSubmittedRequestSummary, type AnthropicBatchSucceededOutcome, type AnthropicBatchSystemContentBlock, type AnthropicBatchTraceEventHandler, type AnthropicBatchWaitOptions, type AnthropicCacheControl, type AnthropicCacheTtl, type AnthropicMessageBatch, type AnthropicMessageBatchProcessingStatus, type AnthropicMessageBatchRequestCounts, type AnthropicMessagesResponse, AnthropicMessagesResponseSchema, type AnthropicOutputConfig, type AnthropicPromptCachingMode, type AnthropicPromptCachingOptions, type AnthropicPromptCachingSettings, type AnthropicRequestBody, type AnthropicRequestContentBlock, type AnthropicRequestInput, type AnthropicRequestMessage, type AnthropicRequestSettings, type AnthropicStreamEvent, AnthropicStreamEventSchema, type AnthropicStreamRequestInput, type AnthropicStreamStartUsage, type AnthropicSystemContentBlock, type AnthropicThinkingOptions, type AnthropicToolChoice, type AnthropicToolDefinition, type AnthropicUsage, type AwsCredentials, type AwsSigV4SigningInput, type AzureOpenAiAdapterOptions, type AzureOpenAiChatCompletionResponse, AzureOpenAiChatCompletionResponseSchema, type AzureOpenAiChatRequestBody, type AzureOpenAiStreamRequestBody, type AzureOpenAiStreamRequestInput, type AzureOpenAiUsage, type BedrockAdapterOptions, type BedrockCredentials, type BedrockInvokeModelRequestBody, type ChatCompletionRequest, type ChatCompletionResult, type ChatMessage, type ChatMessageCacheControl, ChatMessageCacheControlSchema, type ChatMessageRole, ChatMessageRoleSchema, ChatMessageSchema, type ChatStopReason, ChatStopReasonSchema, type ChatStreamEvent, type GeminiAdapterOptions, type MessageContentBlock, MessageContentBlockSchema, type MockModelOptions, type ModelAdapter, type ModelPricingUsdPerMillionTokens, type ModelToolCall, ModelToolCallSchema, type ModelToolChoice, type ModelToolDefinition, type OpenAiAdapterOptions, type OpenAiCallInput, type OpenAiContentPart, type OpenAiRequestBody, type OpenAiRequestMessage, type OpenAiRequestToolCall, type OpenAiResponseFormat, type OpenAiToolChoice, type OpenAiToolDefinition, type ProviderExtraBody, type ProviderRetryOptions, type ResolveProviderApiKeyInput, type StructuredOutputRequest, type StructuredOutputResult, anthropic, anthropicBatchJsonOnlySystemInstruction, anthropicBatchTraceWorkflowName, applyCacheControlToLastContentBlock, awsSigV4Algorithm, azureOpenai, bedrock, bedrockAnthropicVersion, buildAnthropicCacheControl, buildAnthropicRequestBody, buildAnthropicStructuredOutputJsonSchema, buildAzureOpenAiChatCompletionsUrl, buildAzureOpenAiChatRequestBody, buildBedrockInvokeModelRequestBody, buildBedrockInvokeModelUrl, buildGeminiResponseSchema, buildOpenAiRequestBody, buildOpenAiStructuredOutputJsonSchema, buildProviderCanceledError, calculateProviderRetryDelayMs, convertZodSchemaToJsonSchema, createAnthropicBatchClient, defaultAnthropicBaseUrl, defaultAnthropicBatchPollIntervalMs, defaultAnthropicBatchTimeoutMs, defaultAnthropicMaxOutputTokens, defaultAzureOpenAiApiVersion, defaultBedrockMaxOutputTokens, defaultGeminiBaseUrl, defaultOpenAiBaseUrl, defaultProviderRetryInitialDelayMs, defaultProviderRetryMaxAttempts, defaultProviderRetryMaxDelayMs, estimateModelCallCostUsd, executeAbortableProviderFetch, executeProviderCallWithModelFallback, executeProviderCallWithRetry, extractAnthropicTextContent, extractAnthropicThinkingText, extractAnthropicToolCalls, extractMessageText, gemini, isAbortError, isRetryableProviderAppError, joinSystemMessageText, mapAnthropicResponseToChatCompletionResult, mapAnthropicUsageToTokenUsage, mapAzureOpenAiResponseToChatCompletionResult, mapAzureOpenAiUsageToTokenUsage, mapChatMessageToAnthropicRequestMessage, mapChatMessageToOpenAiRequestMessage, mapModelToolChoiceToAnthropicToolChoice, mapModelToolChoiceToOpenAiToolChoice, mapModelToolDefinitionToOpenAiTool, mapModelToolDefinitionsToAnthropicTools, mergeProviderRequestExtraBody, mockModel, normalizeAnthropicStopReason, normalizeAzureOpenAiFinishReason, openAiStructuredOutputSchemaName, openai, parseAnthropicBatchStructuredContent, parseAnthropicStreamEvents, parseAnthropicToolInputJson, parseAzureOpenAiStreamEvents, parseAzureOpenAiToolArgumentsJson, resolveAnthropicPromptCachingSettings, resolveBedrockCredentials, resolveProviderApiKey, sanitizeJsonSchemaForAnthropicStructuredOutput, sanitizeJsonSchemaForGeminiResponseSchema, sanitizeJsonSchemaForOpenAiStructuredOutput, signAwsRequestWithSigV4, streamAnthropicChatCompletion, streamAzureOpenAiChatCompletion };
+export { type AnthropicAdapterOptions, type AnthropicBatchApiRequest, type AnthropicBatchApiRequestMessage, type AnthropicBatchApiRequestParams, type AnthropicBatchCanceledOutcome, type AnthropicBatchClient, type AnthropicBatchClientOptions, type AnthropicBatchErroredOutcome, type AnthropicBatchExpiredOutcome, type AnthropicBatchMessageRequest, type AnthropicBatchPromptCachingMode, type AnthropicBatchResult, type AnthropicBatchResultOutcome, type AnthropicBatchSubmittedRequestSummary, type AnthropicBatchSucceededOutcome, type AnthropicBatchSystemContentBlock, type AnthropicBatchTraceEventHandler, type AnthropicBatchWaitOptions, type AnthropicCacheControl, type AnthropicCacheTtl, type AnthropicMessageBatch, type AnthropicMessageBatchProcessingStatus, type AnthropicMessageBatchRequestCounts, type AnthropicMessagesResponse, AnthropicMessagesResponseSchema, type AnthropicOutputConfig, type AnthropicPromptCachingMode, type AnthropicPromptCachingOptions, type AnthropicPromptCachingSettings, type AnthropicRequestBody, type AnthropicRequestContentBlock, type AnthropicRequestInput, type AnthropicRequestMessage, type AnthropicRequestSettings, type AnthropicStreamEvent, AnthropicStreamEventSchema, type AnthropicStreamRequestInput, type AnthropicStreamStartUsage, type AnthropicSystemContentBlock, type AnthropicThinkingOptions, type AnthropicToolChoice, type AnthropicToolDefinition, type AnthropicUsage, type AwsCredentials, type AwsSigV4SigningInput, type AzureOpenAiAdapterOptions, type AzureOpenAiChatCompletionResponse, AzureOpenAiChatCompletionResponseSchema, type AzureOpenAiChatRequestBody, type AzureOpenAiStreamRequestBody, type AzureOpenAiStreamRequestInput, type AzureOpenAiUsage, type BedrockAdapterOptions, type BedrockCredentials, type BedrockInvokeModelRequestBody, type ChatCompletionRequest, type ChatCompletionResult, type ChatMessage, type ChatMessageCacheControl, ChatMessageCacheControlSchema, type ChatMessageRole, ChatMessageRoleSchema, ChatMessageSchema, type ChatStopReason, ChatStopReasonSchema, type ChatStreamEvent, type GeminiAdapterOptions, type MessageContentBlock, MessageContentBlockSchema, type MockModelOptions, type ModelAdapter, type ModelPricingUsdPerMillionTokens, type ModelToolCall, ModelToolCallSchema, type ModelToolChoice, type ModelToolDefinition, type OpenAiAdapterOptions, type OpenAiCallInput, type OpenAiContentPart, type OpenAiRequestBody, type OpenAiRequestMessage, type OpenAiRequestToolCall, type OpenAiResponseFormat, type OpenAiStructuredOutputMode, type OpenAiToolChoice, type OpenAiToolDefinition, type ProviderExtraBody, type ProviderRetryOptions, type ResolveProviderApiKeyInput, type StructuredOutputRequest, type StructuredOutputResult, anthropic, anthropicBatchJsonOnlySystemInstruction, anthropicBatchTraceWorkflowName, applyCacheControlToLastContentBlock, awsSigV4Algorithm, azureOpenai, bedrock, bedrockAnthropicVersion, buildAnthropicCacheControl, buildAnthropicRequestBody, buildAnthropicStructuredOutputJsonSchema, buildAzureOpenAiChatCompletionsUrl, buildAzureOpenAiChatRequestBody, buildBedrockInvokeModelRequestBody, buildBedrockInvokeModelUrl, buildGeminiResponseSchema, buildOpenAiRequestBody, buildOpenAiStructuredOutputJsonSchema, buildProviderCanceledError, calculateProviderRetryDelayMs, convertZodSchemaToJsonSchema, createAnthropicBatchClient, defaultAnthropicBaseUrl, defaultAnthropicBatchPollIntervalMs, defaultAnthropicBatchTimeoutMs, defaultAnthropicMaxOutputTokens, defaultAzureOpenAiApiVersion, defaultBedrockMaxOutputTokens, defaultGeminiBaseUrl, defaultOpenAiBaseUrl, defaultProviderRetryInitialDelayMs, defaultProviderRetryMaxAttempts, defaultProviderRetryMaxDelayMs, estimateModelCallCostUsd, executeAbortableProviderFetch, executeProviderCallWithModelFallback, executeProviderCallWithRetry, extractAnthropicTextContent, extractAnthropicThinkingText, extractAnthropicToolCalls, extractMessageText, gemini, isAbortError, isOpenAiReasoningModel, isRetryableProviderAppError, isZodV4Schema, joinSystemMessageText, mapAnthropicResponseToChatCompletionResult, mapAnthropicUsageToTokenUsage, mapAzureOpenAiResponseToChatCompletionResult, mapAzureOpenAiUsageToTokenUsage, mapChatMessageToAnthropicRequestMessage, mapChatMessageToOpenAiRequestMessage, mapModelToolChoiceToAnthropicToolChoice, mapModelToolChoiceToOpenAiToolChoice, mapModelToolDefinitionToOpenAiTool, mapModelToolDefinitionsToAnthropicTools, mergeProviderRequestExtraBody, mockModel, normalizeAnthropicStopReason, normalizeAzureOpenAiFinishReason, openAiStructuredOutputSchemaName, openai, parseAnthropicBatchStructuredContent, parseAnthropicStreamEvents, parseAnthropicToolInputJson, parseAzureOpenAiStreamEvents, parseAzureOpenAiToolArgumentsJson, prependStructuredOutputJsonOnlySystemMessage, resolveAnthropicPromptCachingSettings, resolveBedrockCredentials, resolveOpenAiStructuredOutputMode, resolveProviderApiKey, sanitizeJsonSchemaForAnthropicStructuredOutput, sanitizeJsonSchemaForGeminiResponseSchema, sanitizeJsonSchemaForOpenAiStructuredOutput, signAwsRequestWithSigV4, streamAnthropicChatCompletion, streamAzureOpenAiChatCompletion, structuredOutputJsonOnlySystemInstruction, structuredOutputJsonOnlySystemMessage };

package/dist/index.d.ts

@@ -357,12 +357,17 @@ declare function buildProviderCanceledError(providerLabel: string): AppError;
 declare function executeAbortableProviderFetch(providerLabel: string, performFetch: () => Promise<Response>): Promise<Response>;
 
 declare const openAiStructuredOutputSchemaName = "structured_output";
+declare function isZodV4Schema<T>(schema: z.ZodType<T>): boolean;
 declare function convertZodSchemaToJsonSchema<T>(schema: z.ZodType<T>): JsonObject;
 declare function buildAnthropicStructuredOutputJsonSchema<T>(schema: z.ZodType<T>): JsonObject;
 declare function sanitizeJsonSchemaForAnthropicStructuredOutput(jsonSchema: JsonObject): JsonObject;
 declare function buildOpenAiStructuredOutputJsonSchema<T>(schema: z.ZodType<T>): JsonObject;
 declare function sanitizeJsonSchemaForOpenAiStructuredOutput(jsonSchema: JsonObject): JsonObject;
 
+declare const structuredOutputJsonOnlySystemInstruction = "Respond with a single JSON object that satisfies the caller's requested structure. Output only valid JSON with no surrounding text.";
+declare const structuredOutputJsonOnlySystemMessage: ChatMessage;
+declare function prependStructuredOutputJsonOnlySystemMessage(messages: ChatMessage[]): ChatMessage[];
+
 declare function buildGeminiResponseSchema<T>(schema: z.ZodType<T>): JsonObject;
 declare function sanitizeJsonSchemaForGeminiResponseSchema(jsonSchema: JsonObject): JsonObject;
 
@@ -539,6 +544,7 @@ type OpenAiRequestBody = {
     messages: OpenAiRequestMessage[];
     temperature?: number;
     max_tokens?: number;
+    max_completion_tokens?: number;
     response_format?: OpenAiResponseFormat;
     tools?: OpenAiToolDefinition[];
     tool_choice?: OpenAiToolChoice;
@@ -547,6 +553,7 @@ type OpenAiCallInput = {
     messages: ChatMessage[];
     temperature?: number;
     maxOutputTokens?: number;
+    reasoningModel?: boolean;
     structuredOutputJsonSchema?: JsonObject;
     requireJsonObjectResponse?: boolean;
     tools?: ModelToolDefinition[];
@@ -554,6 +561,13 @@ type OpenAiCallInput = {
     abortSignal?: AbortSignal;
     extraBody?: Record<string, JsonValue>;
 };
+type OpenAiStructuredOutputMode = {
+    messages: ChatMessage[];
+    structuredOutputJsonSchema?: JsonObject;
+    requireJsonObjectResponse?: boolean;
+};
+declare function isOpenAiReasoningModel(model: string): boolean;
+declare function resolveOpenAiStructuredOutputMode<T>(messages: ChatMessage[], schema: z.ZodType<T>): OpenAiStructuredOutputMode;
 declare function buildOpenAiRequestBody(model: string, input: OpenAiCallInput): OpenAiRequestBody;
 declare function mapModelToolDefinitionToOpenAiTool(tool: ModelToolDefinition): OpenAiToolDefinition;
 declare function mapModelToolChoiceToOpenAiToolChoice(toolChoice: ModelToolChoice): OpenAiToolChoice;
@@ -695,6 +709,7 @@ type AzureOpenAiAdapterOptions = {
     fetchImplementation?: typeof fetch;
     temperature?: number;
     maxOutputTokens?: number;
+    reasoningModel?: boolean;
     retry?: ProviderRetryOptions;
     extraBody?: Record<string, JsonValue>;
 };
@@ -791,4 +806,4 @@ declare function buildBedrockInvokeModelUrl(region: string, modelId: string): st
 declare function resolveBedrockCredentials(options: BedrockAdapterOptions): BedrockCredentials;
 declare function buildBedrockInvokeModelRequestBody(input: AnthropicRequestInput): BedrockInvokeModelRequestBody;
 
-export { type AnthropicAdapterOptions, type AnthropicBatchApiRequest, type AnthropicBatchApiRequestMessage, type AnthropicBatchApiRequestParams, type AnthropicBatchCanceledOutcome, type AnthropicBatchClient, type AnthropicBatchClientOptions, type AnthropicBatchErroredOutcome, type AnthropicBatchExpiredOutcome, type AnthropicBatchMessageRequest, type AnthropicBatchPromptCachingMode, type AnthropicBatchResult, type AnthropicBatchResultOutcome, type AnthropicBatchSubmittedRequestSummary, type AnthropicBatchSucceededOutcome, type AnthropicBatchSystemContentBlock, type AnthropicBatchTraceEventHandler, type AnthropicBatchWaitOptions, type AnthropicCacheControl, type AnthropicCacheTtl, type AnthropicMessageBatch, type AnthropicMessageBatchProcessingStatus, type AnthropicMessageBatchRequestCounts, type AnthropicMessagesResponse, AnthropicMessagesResponseSchema, type AnthropicOutputConfig, type AnthropicPromptCachingMode, type AnthropicPromptCachingOptions, type AnthropicPromptCachingSettings, type AnthropicRequestBody, type AnthropicRequestContentBlock, type AnthropicRequestInput, type AnthropicRequestMessage, type AnthropicRequestSettings, type AnthropicStreamEvent, AnthropicStreamEventSchema, type AnthropicStreamRequestInput, type AnthropicStreamStartUsage, type AnthropicSystemContentBlock, type AnthropicThinkingOptions, type AnthropicToolChoice, type AnthropicToolDefinition, type AnthropicUsage, type AwsCredentials, type AwsSigV4SigningInput, type AzureOpenAiAdapterOptions, type AzureOpenAiChatCompletionResponse, AzureOpenAiChatCompletionResponseSchema, type AzureOpenAiChatRequestBody, type AzureOpenAiStreamRequestBody, type AzureOpenAiStreamRequestInput, type AzureOpenAiUsage, type BedrockAdapterOptions, type BedrockCredentials, type BedrockInvokeModelRequestBody, type ChatCompletionRequest, type ChatCompletionResult, type ChatMessage, type ChatMessageCacheControl, ChatMessageCacheControlSchema, type ChatMessageRole, ChatMessageRoleSchema, ChatMessageSchema, type ChatStopReason, ChatStopReasonSchema, type ChatStreamEvent, type GeminiAdapterOptions, type MessageContentBlock, MessageContentBlockSchema, type MockModelOptions, type ModelAdapter, type ModelPricingUsdPerMillionTokens, type ModelToolCall, ModelToolCallSchema, type ModelToolChoice, type ModelToolDefinition, type OpenAiAdapterOptions, type OpenAiCallInput, type OpenAiContentPart, type OpenAiRequestBody, type OpenAiRequestMessage, type OpenAiRequestToolCall, type OpenAiResponseFormat, type OpenAiToolChoice, type OpenAiToolDefinition, type ProviderExtraBody, type ProviderRetryOptions, type ResolveProviderApiKeyInput, type StructuredOutputRequest, type StructuredOutputResult, anthropic, anthropicBatchJsonOnlySystemInstruction, anthropicBatchTraceWorkflowName, applyCacheControlToLastContentBlock, awsSigV4Algorithm, azureOpenai, bedrock, bedrockAnthropicVersion, buildAnthropicCacheControl, buildAnthropicRequestBody, buildAnthropicStructuredOutputJsonSchema, buildAzureOpenAiChatCompletionsUrl, buildAzureOpenAiChatRequestBody, buildBedrockInvokeModelRequestBody, buildBedrockInvokeModelUrl, buildGeminiResponseSchema, buildOpenAiRequestBody, buildOpenAiStructuredOutputJsonSchema, buildProviderCanceledError, calculateProviderRetryDelayMs, convertZodSchemaToJsonSchema, createAnthropicBatchClient, defaultAnthropicBaseUrl, defaultAnthropicBatchPollIntervalMs, defaultAnthropicBatchTimeoutMs, defaultAnthropicMaxOutputTokens, defaultAzureOpenAiApiVersion, defaultBedrockMaxOutputTokens, defaultGeminiBaseUrl, defaultOpenAiBaseUrl, defaultProviderRetryInitialDelayMs, defaultProviderRetryMaxAttempts, defaultProviderRetryMaxDelayMs, estimateModelCallCostUsd, executeAbortableProviderFetch, executeProviderCallWithModelFallback, executeProviderCallWithRetry, extractAnthropicTextContent, extractAnthropicThinkingText, extractAnthropicToolCalls, extractMessageText, gemini, isAbortError, isRetryableProviderAppError, joinSystemMessageText, mapAnthropicResponseToChatCompletionResult, mapAnthropicUsageToTokenUsage, mapAzureOpenAiResponseToChatCompletionResult, mapAzureOpenAiUsageToTokenUsage, mapChatMessageToAnthropicRequestMessage, mapChatMessageToOpenAiRequestMessage, mapModelToolChoiceToAnthropicToolChoice, mapModelToolChoiceToOpenAiToolChoice, mapModelToolDefinitionToOpenAiTool, mapModelToolDefinitionsToAnthropicTools, mergeProviderRequestExtraBody, mockModel, normalizeAnthropicStopReason, normalizeAzureOpenAiFinishReason, openAiStructuredOutputSchemaName, openai, parseAnthropicBatchStructuredContent, parseAnthropicStreamEvents, parseAnthropicToolInputJson, parseAzureOpenAiStreamEvents, parseAzureOpenAiToolArgumentsJson, resolveAnthropicPromptCachingSettings, resolveBedrockCredentials, resolveProviderApiKey, sanitizeJsonSchemaForAnthropicStructuredOutput, sanitizeJsonSchemaForGeminiResponseSchema, sanitizeJsonSchemaForOpenAiStructuredOutput, signAwsRequestWithSigV4, streamAnthropicChatCompletion, streamAzureOpenAiChatCompletion };
+export { type AnthropicAdapterOptions, type AnthropicBatchApiRequest, type AnthropicBatchApiRequestMessage, type AnthropicBatchApiRequestParams, type AnthropicBatchCanceledOutcome, type AnthropicBatchClient, type AnthropicBatchClientOptions, type AnthropicBatchErroredOutcome, type AnthropicBatchExpiredOutcome, type AnthropicBatchMessageRequest, type AnthropicBatchPromptCachingMode, type AnthropicBatchResult, type AnthropicBatchResultOutcome, type AnthropicBatchSubmittedRequestSummary, type AnthropicBatchSucceededOutcome, type AnthropicBatchSystemContentBlock, type AnthropicBatchTraceEventHandler, type AnthropicBatchWaitOptions, type AnthropicCacheControl, type AnthropicCacheTtl, type AnthropicMessageBatch, type AnthropicMessageBatchProcessingStatus, type AnthropicMessageBatchRequestCounts, type AnthropicMessagesResponse, AnthropicMessagesResponseSchema, type AnthropicOutputConfig, type AnthropicPromptCachingMode, type AnthropicPromptCachingOptions, type AnthropicPromptCachingSettings, type AnthropicRequestBody, type AnthropicRequestContentBlock, type AnthropicRequestInput, type AnthropicRequestMessage, type AnthropicRequestSettings, type AnthropicStreamEvent, AnthropicStreamEventSchema, type AnthropicStreamRequestInput, type AnthropicStreamStartUsage, type AnthropicSystemContentBlock, type AnthropicThinkingOptions, type AnthropicToolChoice, type AnthropicToolDefinition, type AnthropicUsage, type AwsCredentials, type AwsSigV4SigningInput, type AzureOpenAiAdapterOptions, type AzureOpenAiChatCompletionResponse, AzureOpenAiChatCompletionResponseSchema, type AzureOpenAiChatRequestBody, type AzureOpenAiStreamRequestBody, type AzureOpenAiStreamRequestInput, type AzureOpenAiUsage, type BedrockAdapterOptions, type BedrockCredentials, type BedrockInvokeModelRequestBody, type ChatCompletionRequest, type ChatCompletionResult, type ChatMessage, type ChatMessageCacheControl, ChatMessageCacheControlSchema, type ChatMessageRole, ChatMessageRoleSchema, ChatMessageSchema, type ChatStopReason, ChatStopReasonSchema, type ChatStreamEvent, type GeminiAdapterOptions, type MessageContentBlock, MessageContentBlockSchema, type MockModelOptions, type ModelAdapter, type ModelPricingUsdPerMillionTokens, type ModelToolCall, ModelToolCallSchema, type ModelToolChoice, type ModelToolDefinition, type OpenAiAdapterOptions, type OpenAiCallInput, type OpenAiContentPart, type OpenAiRequestBody, type OpenAiRequestMessage, type OpenAiRequestToolCall, type OpenAiResponseFormat, type OpenAiStructuredOutputMode, type OpenAiToolChoice, type OpenAiToolDefinition, type ProviderExtraBody, type ProviderRetryOptions, type ResolveProviderApiKeyInput, type StructuredOutputRequest, type StructuredOutputResult, anthropic, anthropicBatchJsonOnlySystemInstruction, anthropicBatchTraceWorkflowName, applyCacheControlToLastContentBlock, awsSigV4Algorithm, azureOpenai, bedrock, bedrockAnthropicVersion, buildAnthropicCacheControl, buildAnthropicRequestBody, buildAnthropicStructuredOutputJsonSchema, buildAzureOpenAiChatCompletionsUrl, buildAzureOpenAiChatRequestBody, buildBedrockInvokeModelRequestBody, buildBedrockInvokeModelUrl, buildGeminiResponseSchema, buildOpenAiRequestBody, buildOpenAiStructuredOutputJsonSchema, buildProviderCanceledError, calculateProviderRetryDelayMs, convertZodSchemaToJsonSchema, createAnthropicBatchClient, defaultAnthropicBaseUrl, defaultAnthropicBatchPollIntervalMs, defaultAnthropicBatchTimeoutMs, defaultAnthropicMaxOutputTokens, defaultAzureOpenAiApiVersion, defaultBedrockMaxOutputTokens, defaultGeminiBaseUrl, defaultOpenAiBaseUrl, defaultProviderRetryInitialDelayMs, defaultProviderRetryMaxAttempts, defaultProviderRetryMaxDelayMs, estimateModelCallCostUsd, executeAbortableProviderFetch, executeProviderCallWithModelFallback, executeProviderCallWithRetry, extractAnthropicTextContent, extractAnthropicThinkingText, extractAnthropicToolCalls, extractMessageText, gemini, isAbortError, isOpenAiReasoningModel, isRetryableProviderAppError, isZodV4Schema, joinSystemMessageText, mapAnthropicResponseToChatCompletionResult, mapAnthropicUsageToTokenUsage, mapAzureOpenAiResponseToChatCompletionResult, mapAzureOpenAiUsageToTokenUsage, mapChatMessageToAnthropicRequestMessage, mapChatMessageToOpenAiRequestMessage, mapModelToolChoiceToAnthropicToolChoice, mapModelToolChoiceToOpenAiToolChoice, mapModelToolDefinitionToOpenAiTool, mapModelToolDefinitionsToAnthropicTools, mergeProviderRequestExtraBody, mockModel, normalizeAnthropicStopReason, normalizeAzureOpenAiFinishReason, openAiStructuredOutputSchemaName, openai, parseAnthropicBatchStructuredContent, parseAnthropicStreamEvents, parseAnthropicToolInputJson, parseAzureOpenAiStreamEvents, parseAzureOpenAiToolArgumentsJson, prependStructuredOutputJsonOnlySystemMessage, resolveAnthropicPromptCachingSettings, resolveBedrockCredentials, resolveOpenAiStructuredOutputMode, resolveProviderApiKey, sanitizeJsonSchemaForAnthropicStructuredOutput, sanitizeJsonSchemaForGeminiResponseSchema, sanitizeJsonSchemaForOpenAiStructuredOutput, signAwsRequestWithSigV4, streamAnthropicChatCompletion, streamAzureOpenAiChatCompletion, structuredOutputJsonOnlySystemInstruction, structuredOutputJsonOnlySystemMessage };

package/dist/index.js

@@ -204,6 +204,9 @@ var openAiUnsupportedSchemaKeywords = [
   "uniqueItems",
   "default"
 ];
+function isZodV4Schema(schema) {
+  return "_zod" in schema;
+}
 function convertZodSchemaToJsonSchema(schema) {
   try {
     return JsonObjectSchema.parse(z2.toJSONSchema(schema, { io: "input" }));
@@ -268,18 +271,72 @@ function listSchemaPropertyNames(node) {
   return Object.keys(properties);
 }
 
+// src/structured-output-fallback.ts
+var structuredOutputJsonOnlySystemInstruction = "Respond with a single JSON object that satisfies the caller's requested structure. Output only valid JSON with no surrounding text.";
+var structuredOutputJsonOnlySystemMessage = {
+  role: "system",
+  content: structuredOutputJsonOnlySystemInstruction
+};
+function prependStructuredOutputJsonOnlySystemMessage(messages) {
+  return [structuredOutputJsonOnlySystemMessage, ...messages];
+}
+
 // src/openai-request-mapping.ts
+var openAiReasoningModelNamePrefixes = [
+  "gpt-5",
+  "o1",
+  "o3",
+  "o4"
+];
+function isOpenAiReasoningModel(model) {
+  return openAiReasoningModelNamePrefixes.some(
+    (prefix) => matchesOpenAiModelNamePrefix(model, prefix)
+  );
+}
+function matchesOpenAiModelNamePrefix(model, prefix) {
+  if (!model.startsWith(prefix)) {
+    return false;
+  }
+  const boundaryCharacter = model.charAt(prefix.length);
+  return boundaryCharacter === "" || boundaryCharacter === "-" || boundaryCharacter === ".";
+}
+function resolveOpenAiStructuredOutputMode(messages, schema) {
+  if (isZodV4Schema(schema)) {
+    return {
+      messages,
+      structuredOutputJsonSchema: buildOpenAiStructuredOutputJsonSchema(schema)
+    };
+  }
+  return {
+    messages: prependStructuredOutputJsonOnlySystemMessage(messages),
+    requireJsonObjectResponse: true
+  };
+}
 function buildOpenAiRequestBody(model, input) {
   const requestBody = {
     model,
     messages: input.messages.map(mapChatMessageToOpenAiRequestMessage)
   };
-  if (input.temperature !== void 0) {
+  applyOpenAiGenerationSettings(requestBody, model, input);
+  applyOpenAiResponseFormat(requestBody, input);
+  applyOpenAiToolFields(requestBody, input);
+  return requestBody;
+}
+function applyOpenAiGenerationSettings(requestBody, model, input) {
+  const reasoningModel = input.reasoningModel ?? isOpenAiReasoningModel(model);
+  if (input.temperature !== void 0 && !reasoningModel) {
     requestBody.temperature = input.temperature;
   }
-  if (input.maxOutputTokens !== void 0) {
-    requestBody.max_tokens = input.maxOutputTokens;
+  if (input.maxOutputTokens === void 0) {
+    return;
+  }
+  if (reasoningModel) {
+    requestBody.max_completion_tokens = input.maxOutputTokens;
+    return;
   }
+  requestBody.max_tokens = input.maxOutputTokens;
+}
+function applyOpenAiResponseFormat(requestBody, input) {
   if (input.structuredOutputJsonSchema !== void 0) {
     requestBody.response_format = {
       type: "json_schema",
@@ -289,9 +346,13 @@ function buildOpenAiRequestBody(model, input) {
         schema: input.structuredOutputJsonSchema
       }
     };
-  } else if (input.requireJsonObjectResponse === true) {
+    return;
+  }
+  if (input.requireJsonObjectResponse === true) {
     requestBody.response_format = { type: "json_object" };
   }
+}
+function applyOpenAiToolFields(requestBody, input) {
   if (input.tools !== void 0 && input.tools.length > 0) {
     requestBody.tools = input.tools.map(mapModelToolDefinitionToOpenAiTool);
   }
@@ -300,7 +361,6 @@ function buildOpenAiRequestBody(model, input) {
       input.toolChoice
     );
   }
-  return requestBody;
 }
 function mapModelToolDefinitionToOpenAiTool(tool) {
   return {
@@ -665,12 +725,9 @@ function openai(options) {
     },
     async generateStructuredOutput(request) {
       const completion = await executeOpenAiAdapterCall(options, {
-        messages: request.messages,
+        ...resolveOpenAiStructuredOutputMode(request.messages, request.schema),
         temperature: request.temperature ?? options.temperature,
         maxOutputTokens: request.maxOutputTokens ?? options.maxOutputTokens,
-        structuredOutputJsonSchema: buildOpenAiStructuredOutputJsonSchema(
-          request.schema
-        ),
         abortSignal: request.abortSignal
       });
       return {
@@ -1438,15 +1495,10 @@ function anthropic(options) {
       );
     },
     async generateStructuredOutput(request) {
-      const completion = await executeAnthropicAdapterCall(options, {
-        messages: request.messages,
-        temperature: request.temperature ?? options.temperature,
-        maxOutputTokens: request.maxOutputTokens ?? options.maxOutputTokens,
-        structuredOutputJsonSchema: buildAnthropicStructuredOutputJsonSchema(
-          request.schema
-        ),
-        abortSignal: request.abortSignal
-      });
+      const completion = await executeAnthropicAdapterCall(
+        options,
+        buildAnthropicStructuredOutputInput(options, request)
+      );
       return {
         value: parseStructuredJsonText(completion.content, request.schema),
         tokenUsage: completion.tokenUsage,
@@ -1460,6 +1512,26 @@ function anthropic(options) {
     }
   };
 }
+function buildAnthropicStructuredOutputInput(options, request) {
+  const baseInput = {
+    messages: request.messages,
+    temperature: request.temperature ?? options.temperature,
+    maxOutputTokens: request.maxOutputTokens ?? options.maxOutputTokens,
+    abortSignal: request.abortSignal
+  };
+  if (isZodV4Schema(request.schema)) {
+    return {
+      ...baseInput,
+      structuredOutputJsonSchema: buildAnthropicStructuredOutputJsonSchema(
+        request.schema
+      )
+    };
+  }
+  return {
+    ...baseInput,
+    messages: prependStructuredOutputJsonOnlySystemMessage(request.messages)
+  };
+}
 function buildAnthropicCallInput(options, request) {
   return {
     messages: request.messages,
@@ -1684,13 +1756,10 @@ function gemini(options) {
       });
     },
     async generateStructuredOutput(request) {
-      const completion = await executeGeminiAdapterCall(options, {
-        messages: request.messages,
-        temperature: request.temperature ?? options.temperature,
-        maxOutputTokens: request.maxOutputTokens ?? options.maxOutputTokens,
-        responseSchema: buildGeminiResponseSchema(request.schema),
-        abortSignal: request.abortSignal
-      });
+      const completion = await executeGeminiAdapterCall(
+        options,
+        buildGeminiStructuredOutputInput(options, request)
+      );
       return {
         value: parseStructuredJsonText(completion.content, request.schema),
         tokenUsage: completion.tokenUsage,
@@ -1699,6 +1768,25 @@ function gemini(options) {
     }
   };
 }
+function buildGeminiStructuredOutputInput(options, request) {
+  const baseInput = {
+    messages: request.messages,
+    temperature: request.temperature ?? options.temperature,
+    maxOutputTokens: request.maxOutputTokens ?? options.maxOutputTokens,
+    abortSignal: request.abortSignal
+  };
+  if (isZodV4Schema(request.schema)) {
+    return {
+      ...baseInput,
+      responseSchema: buildGeminiResponseSchema(request.schema)
+    };
+  }
+  return {
+    ...baseInput,
+    messages: prependStructuredOutputJsonOnlySystemMessage(request.messages),
+    requireJsonResponse: true
+  };
+}
 async function executeGeminiAdapterCall(options, input) {
   return await executeProviderCallWithModelFallback(
     options.fallbackModel,
@@ -1816,6 +1904,8 @@ function buildGeminiGenerationConfig(input) {
   if (input.responseSchema !== void 0) {
     generationConfig.responseMimeType = "application/json";
     generationConfig.responseSchema = input.responseSchema;
+  } else if (input.requireJsonResponse === true) {
+    generationConfig.responseMimeType = "application/json";
   }
   if (Object.keys(generationConfig).length === 0) {
     return void 0;
@@ -2023,7 +2113,7 @@ import { AppError as AppError15 } from "@assemble-dev/shared-utils/errors";
 import { AppErrorCode as AppErrorCode13 } from "@assemble-dev/shared-types/errors";
 import { JsonValueSchema as JsonValueSchema6 } from "@assemble-dev/shared-types/json";
 import { AppError as AppError14 } from "@assemble-dev/shared-utils/errors";
-var anthropicBatchJsonOnlySystemInstruction = "Respond with a single JSON object that satisfies the caller's requested structure. Output only valid JSON with no surrounding text.";
+var anthropicBatchJsonOnlySystemInstruction = structuredOutputJsonOnlySystemInstruction;
 function parseAnthropicBatchStructuredContent(content, schema) {
   const jsonValue = parseStructuredJsonText(
     content,
@@ -2833,6 +2923,7 @@ function azureOpenai(options) {
         messages: request.messages,
         temperature: request.temperature ?? options.temperature,
         maxOutputTokens: request.maxOutputTokens ?? options.maxOutputTokens,
+        reasoningModel: options.reasoningModel,
         tools: request.tools,
         toolChoice: request.toolChoice,
         abortSignal: request.abortSignal,
@@ -2841,12 +2932,10 @@ function azureOpenai(options) {
     },
     async generateStructuredOutput(request) {
       const completion = await executeAzureOpenAiAdapterCall(options, {
-        messages: request.messages,
+        ...resolveOpenAiStructuredOutputMode(request.messages, request.schema),
         temperature: request.temperature ?? options.temperature,
         maxOutputTokens: request.maxOutputTokens ?? options.maxOutputTokens,
-        structuredOutputJsonSchema: buildOpenAiStructuredOutputJsonSchema(
-          request.schema
-        ),
+        reasoningModel: options.reasoningModel,
         abortSignal: request.abortSignal
       });
       return {
@@ -2945,6 +3034,7 @@ function buildAzureOpenAiStreamInput(options, request) {
       messages: request.messages,
       temperature: request.temperature ?? options.temperature,
       maxOutputTokens: request.maxOutputTokens ?? options.maxOutputTokens,
+      reasoningModel: options.reasoningModel,
       tools: request.tools,
       toolChoice: request.toolChoice,
       abortSignal: request.abortSignal,
@@ -3239,6 +3329,9 @@ function buildBedrockInvokeModelRequestBody(input) {
   };
 }
 function buildBedrockStructuredOutputSystemMessage(schema) {
+  if (!isZodV4Schema(schema)) {
+    return structuredOutputJsonOnlySystemMessage;
+  }
   return {
     role: "system",
     content: `Respond with a single JSON object that satisfies the following JSON Schema. Output only valid JSON with no surrounding text.
@@ -3300,7 +3393,9 @@ export {
   extractMessageText,
   gemini,
   isAbortError,
+  isOpenAiReasoningModel,
   isRetryableProviderAppError,
+  isZodV4Schema,
   joinSystemMessageText,
   mapAnthropicResponseToChatCompletionResult,
   mapAnthropicUsageToTokenUsage,
@@ -3323,13 +3418,17 @@ export {
   parseAnthropicToolInputJson,
   parseAzureOpenAiStreamEvents,
   parseAzureOpenAiToolArgumentsJson,
+  prependStructuredOutputJsonOnlySystemMessage,
   resolveAnthropicPromptCachingSettings,
   resolveBedrockCredentials,
+  resolveOpenAiStructuredOutputMode,
   resolveProviderApiKey,
   sanitizeJsonSchemaForAnthropicStructuredOutput,
   sanitizeJsonSchemaForGeminiResponseSchema,
   sanitizeJsonSchemaForOpenAiStructuredOutput,
   signAwsRequestWithSigV4,
   streamAnthropicChatCompletion,
-  streamAzureOpenAiChatCompletion
+  streamAzureOpenAiChatCompletion,
+  structuredOutputJsonOnlySystemInstruction,
+  structuredOutputJsonOnlySystemMessage
 };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@assemble-dev/providers",
-  "version": "0.3.0",
+  "version": "0.3.1",
   "license": "Elastic-2.0",
   "description": "Model provider adapters for the Assemble SDK",
   "repository": {
@@ -27,8 +27,8 @@
   ],
   "dependencies": {
     "zod": "^4.1.13",
-    "@assemble-dev/shared-types": "0.3.0",
-    "@assemble-dev/shared-utils": "0.3.0"
+    "@assemble-dev/shared-types": "0.3.1",
+    "@assemble-dev/shared-utils": "0.3.1"
   },
   "devDependencies": {
     "@types/node": "^20.19.0",