@yourgpt/llm-sdk 2.1.10-alpha.0 → 2.5.1-beta.0

package/dist/adapters/index.js

@@ -114,6 +114,109 @@ function normalizeObjectJsonSchema(schema) {
   }
   return normalized;
 }
+function isOpenAIReasoningModel(modelId) {
+  if (!modelId) return false;
+  return /^(o1|o3|o4|gpt-5)/i.test(modelId);
+}
+function buildOpenAITokenParams(modelId, maxTokens, temperature) {
+  if (isOpenAIReasoningModel(modelId)) {
+    return { max_completion_tokens: maxTokens };
+  }
+  return { max_tokens: maxTokens, temperature };
+}
+function stripSchemaKeys(schema, keysToDrop, options = {}) {
+  if (Array.isArray(schema)) {
+    return schema.map((item) => stripSchemaKeys(item, keysToDrop, options));
+  }
+  if (!schema || typeof schema !== "object") return schema;
+  const out = {};
+  for (const [key, value] of Object.entries(
+    schema
+  )) {
+    if (keysToDrop.has(key)) continue;
+    const renamed = options.renameKeys?.[key] ?? key;
+    out[renamed] = stripSchemaKeys(value, keysToDrop, options);
+  }
+  if (options.forceAdditionalPropertiesFalse && out.type === "object") {
+    out.additionalProperties = false;
+  }
+  return out;
+}
+function toOpenAIResponseFormat(rf) {
+  if (!rf) return void 0;
+  if (rf.type === "json_object") return { type: "json_object" };
+  return {
+    type: "json_schema",
+    json_schema: {
+      name: rf.json_schema.name,
+      schema: normalizeObjectJsonSchema(rf.json_schema.schema),
+      strict: rf.json_schema.strict ?? true
+    }
+  };
+}
+function toOpenAIResponsesTextFormat(rf) {
+  if (!rf || rf.type !== "json_schema") return void 0;
+  return {
+    type: "json_schema",
+    name: rf.json_schema.name,
+    schema: normalizeObjectJsonSchema(rf.json_schema.schema),
+    strict: rf.json_schema.strict ?? true
+  };
+}
+var ANTHROPIC_UNSUPPORTED_KEYS = /* @__PURE__ */ new Set([
+  "minimum",
+  "maximum",
+  "exclusiveMinimum",
+  "exclusiveMaximum",
+  "multipleOf",
+  "minLength",
+  "maxLength",
+  "minItems",
+  "maxItems",
+  "minProperties",
+  "maxProperties",
+  "pattern",
+  "$schema"
+]);
+function toAnthropicOutputConfig(rf) {
+  if (!rf || rf.type !== "json_schema") return void 0;
+  const schema = stripSchemaKeys(
+    rf.json_schema.schema,
+    ANTHROPIC_UNSUPPORTED_KEYS,
+    {
+      forceAdditionalPropertiesFalse: true,
+      renameKeys: { oneOf: "anyOf" }
+    }
+  );
+  return {
+    format: {
+      type: "json_schema",
+      schema
+    }
+  };
+}
+var GEMINI_UNSUPPORTED_KEYS = /* @__PURE__ */ new Set([
+  "oneOf",
+  "anyOf",
+  "$ref",
+  "$defs",
+  "definitions",
+  "pattern",
+  "$schema",
+  "additionalProperties"
+]);
+function toGeminiSchema(rf) {
+  if (!rf || rf.type !== "json_schema") return void 0;
+  return stripSchemaKeys(
+    rf.json_schema.schema,
+    GEMINI_UNSUPPORTED_KEYS
+  );
+}
+function toOllamaFormat(rf) {
+  if (!rf) return void 0;
+  if (rf.type === "json_object") return "json";
+  return rf.json_schema.schema;
+}
 function formatTools(actions) {
   return actions.map((action) => ({
     type: "function",
@@ -382,7 +485,6 @@ var OpenAIAdapter = class _OpenAIAdapter {
     if (baseUrl.includes("generativelanguage.googleapis.com")) return "google";
     if (baseUrl.includes("x.ai")) return "xai";
     if (baseUrl.includes("azure")) return "azure";
-    if (baseUrl.includes("openrouter.ai")) return "openrouter";
     return "openai";
   }
   async getClient() {
@@ -482,259 +584,12 @@ var OpenAIAdapter = class _OpenAIAdapter {
       rawResponse: response
     };
   }
-  /**
-   * OpenAI reasoning models on OpenRouter (o1/o3/o4/gpt-5 family) hide their
-   * reasoning content on the chat-completions endpoint. To surface reasoning
-   * SUMMARIES (not raw CoT, which OpenAI never exposes) we have to use the
-   * Responses API, which streams `response.reasoning_summary_text.delta` events.
-   *
-   * Match by prefix on the OpenRouter model id. Excludes openai/gpt-4o,
-   * openai/gpt-4.1, openai/chatgpt-* — those continue on chat-completions.
-   */
-  isOpenAIReasoningModelOnOpenRouter(activeModel) {
-    if (this.provider !== "openrouter") return false;
-    return activeModel.startsWith("openai/o1") || activeModel.startsWith("openai/o3") || activeModel.startsWith("openai/o4") || activeModel.startsWith("openai/gpt-5");
-  }
-  /**
-   * Convert ActionDefinition[] (the chat-completions tool shape used by the
-   * adapter) to the Responses API tool shape.
-   */
-  buildResponsesToolsFromActions(actions) {
-    if (!actions || actions.length === 0) return void 0;
-    const formatted = formatTools(actions);
-    return formatted.map((t) => ({
-      type: "function",
-      name: t.function.name,
-      description: t.function.description,
-      parameters: t.function.parameters
-    }));
-  }
-  /**
-   * Streaming Responses API path for OpenAI reasoning models on OpenRouter.
-   *
-   * Maps Responses API SSE events back to the same StreamEvent shapes the
-   * chat-completions path emits, so downstream consumers (processChunk.ts,
-   * frontend tool handlers, plan approval, specialist delegations) see
-   * identical events regardless of which path produced them.
-   *
-   *   response.reasoning_summary_text.delta  → thinking:start (once) + thinking:delta
-   *   response.output_text.delta             → message:delta
-   *   response.output_item.added (function_call) → action:start (queued buffer)
-   *   response.function_call_arguments.delta → action:args (progressive)
-   *   response.output_item.done (function_call) → final action:args + action:end
-   *   response.completed                     → message:end + done(usage)
-   *   response.error                         → error
-   */
-  async *streamWithResponsesAPI(request, activeModel, messageId) {
-    const client = await this.getClient();
-    const maxTokensValue = request.config?.maxTokens ?? this.config.maxTokens;
-    const payload = {
-      model: activeModel,
-      input: this.buildResponsesInput(request),
-      stream: true,
-      reasoning: {
-        effort: request.config?.reasoningEffort ?? "medium",
-        summary: "auto"
-      }
-    };
-    if (request.systemPrompt) payload.instructions = request.systemPrompt;
-    if (typeof maxTokensValue === "number")
-      payload.max_output_tokens = maxTokensValue;
-    const tools = this.buildResponsesToolsFromActions(request.actions);
-    if (tools && tools.length > 0) payload.tools = tools;
-    logProviderPayload(
-      "openai",
-      "responses-api request payload",
-      payload,
-      request.debug
-    );
-    let stream;
-    try {
-      stream = await client.responses.create(payload);
-    } catch (error) {
-      yield {
-        type: "error",
-        message: error instanceof Error ? error.message : "Unknown error",
-        code: "OPENAI_RESPONSES_ERROR"
-      };
-      return;
-    }
-    const toolBuffers = /* @__PURE__ */ new Map();
-    const itemIdToCallId = /* @__PURE__ */ new Map();
-    let usage;
-    let reasoningStarted = false;
-    let textStarted = false;
-    let finishEmitted = false;
-    const resolveCallId = (evt) => {
-      if (evt?.call_id) return evt.call_id;
-      if (evt?.item_id) return itemIdToCallId.get(evt.item_id) ?? evt.item_id;
-      if (evt?.item?.call_id) return evt.item.call_id;
-      if (evt?.item?.id) return evt.item.id;
-      return "";
-    };
-    try {
-      for await (const evt of stream) {
-        logProviderPayload(
-          "openai",
-          "responses-api stream chunk",
-          evt,
-          request.debug
-        );
-        if (request.signal?.aborted) break;
-        const t = evt?.type ?? "";
-        if (t === "response.reasoning_summary_text.delta") {
-          const delta = evt.delta ?? "";
-          if (!delta) continue;
-          if (!reasoningStarted) {
-            yield { type: "thinking:start" };
-            reasoningStarted = true;
-          }
-          yield { type: "thinking:delta", content: delta };
-          continue;
-        }
-        if (t === "response.reasoning_summary_text.done" || t === "response.reasoning.done") {
-          continue;
-        }
-        if (t === "response.output_text.delta") {
-          const text = evt.delta ?? "";
-          if (!text) continue;
-          if (reasoningStarted && !textStarted) {
-            yield { type: "thinking:end" };
-            textStarted = true;
-          }
-          yield { type: "message:delta", content: text };
-          continue;
-        }
-        if (t === "response.output_item.added") {
-          const item = evt.item;
-          if (item?.type === "function_call") {
-            const callId = item.call_id ?? item.id ?? "";
-            const itemId = item.id ?? callId;
-            if (callId) {
-              if (itemId && itemId !== callId) {
-                itemIdToCallId.set(itemId, callId);
-              }
-              if (!toolBuffers.has(callId)) {
-                toolBuffers.set(callId, {
-                  id: callId,
-                  name: item.name ?? "",
-                  arguments: item.arguments ?? "",
-                  emittedStart: false
-                });
-              }
-              const buf = toolBuffers.get(callId);
-              if (buf.name && !buf.emittedStart) {
-                yield { type: "action:start", id: buf.id, name: buf.name };
-                buf.emittedStart = true;
-              }
-            }
-          }
-          continue;
-        }
-        if (t === "response.function_call_arguments.delta") {
-          const callId = resolveCallId(evt);
-          const delta = evt.delta ?? "";
-          if (!callId || !delta) continue;
-          let buf = toolBuffers.get(callId);
-          if (!buf) {
-            buf = { id: callId, name: "", arguments: "", emittedStart: false };
-            toolBuffers.set(callId, buf);
-          }
-          buf.arguments += delta;
-          if (buf.emittedStart) {
-            yield {
-              type: "action:args",
-              id: buf.id,
-              args: buf.arguments
-            };
-          }
-          continue;
-        }
-        if (t === "response.output_item.done") {
-          const item = evt.item;
-          if (item?.type === "function_call") {
-            const callId = item.call_id ?? item.id ?? "";
-            const buf = toolBuffers.get(callId);
-            const name = buf?.name || item.name || "";
-            const argsStr = buf?.arguments || item.arguments || "{}";
-            if (callId && name) {
-              if (!buf?.emittedStart) {
-                yield { type: "action:start", id: callId, name };
-              }
-              yield {
-                type: "action:args",
-                id: callId,
-                args: argsStr
-              };
-              yield {
-                type: "action:end",
-                id: callId,
-                name
-              };
-            }
-            toolBuffers.delete(callId);
-          }
-          continue;
-        }
-        if (t === "response.completed") {
-          const u = evt.response?.usage;
-          if (u) {
-            usage = {
-              prompt_tokens: u.input_tokens ?? 0,
-              completion_tokens: u.output_tokens ?? 0,
-              total_tokens: u.total_tokens ?? (u.input_tokens ?? 0) + (u.output_tokens ?? 0)
-            };
-          }
-          for (const buf of toolBuffers.values()) {
-            if (!buf.id || !buf.name) continue;
-            if (!buf.emittedStart) {
-              yield { type: "action:start", id: buf.id, name: buf.name };
-            }
-            yield {
-              type: "action:args",
-              id: buf.id,
-              args: buf.arguments || "{}"
-            };
-            yield { type: "action:end", id: buf.id, name: buf.name };
-          }
-          toolBuffers.clear();
-          if (reasoningStarted && !textStarted) {
-            yield { type: "thinking:end" };
-          }
-          yield { type: "message:end" };
-          yield { type: "done", usage };
-          finishEmitted = true;
-          continue;
-        }
-        if (t === "response.error" || t === "error") {
-          const msg = evt.error?.message || evt.message || "Responses API error";
-          yield {
-            type: "error",
-            message: msg,
-            code: "OPENAI_RESPONSES_ERROR"
-          };
-          return;
-        }
-      }
-    } catch (error) {
-      yield {
-        type: "error",
-        message: error instanceof Error ? error.message : "Unknown error",
-        code: "OPENAI_RESPONSES_ERROR"
-      };
-      return;
-    }
-    if (!finishEmitted) {
-      if (reasoningStarted && !textStarted) {
-        yield { type: "thinking:end" };
-      }
-      yield { type: "message:end" };
-      yield { type: "done", usage };
-    }
-  }
   async completeWithResponses(request) {
     const client = await this.getClient();
     const openaiToolOptions = request.providerToolOptions?.openai;
+    const responsesTextFormat = toOpenAIResponsesTextFormat(
+      request.config?.responseFormat
+    );
     const payload = {
       model: request.config?.model || this.model,
       instructions: request.systemPrompt,
@@ -744,6 +599,7 @@ var OpenAIAdapter = class _OpenAIAdapter {
       parallel_tool_calls: openaiToolOptions?.parallelToolCalls,
       temperature: request.config?.temperature ?? this.config.temperature,
       max_output_tokens: request.config?.maxTokens ?? this.config.maxTokens,
+      ...responsesTextFormat ? { text: { format: responsesTextFormat } } : {},
       stream: false
     };
     logProviderPayload("openai", "request payload", payload, request.debug);
@@ -865,37 +721,21 @@ var OpenAIAdapter = class _OpenAIAdapter {
           name: openaiToolOptions.toolChoice.name
         }
       } : openaiToolOptions?.toolChoice;
-      const isOpenRouter = this.provider === "openrouter";
-      const activeModel = request.config?.model || this.model;
-      const modelSlug = activeModel.replace("openai/", "");
-      const isOSeries = /^o[1-9]/.test(modelSlug);
-      const isOpenAIOnOpenRouter = isOpenRouter && activeModel.startsWith("openai/");
-      if (!this.config.disableThinking && this.isOpenAIReasoningModelOnOpenRouter(activeModel)) {
-        yield* this.streamWithResponsesAPI(request, activeModel, messageId);
-        return;
-      }
-      const maxTokensValue = request.config?.maxTokens ?? this.config.maxTokens;
+      const modelIdForPayload = request.config?.model || this.model;
       const payload = {
-        model: activeModel,
+        model: modelIdForPayload,
         messages,
         tools: tools.length > 0 ? tools : void 0,
         tool_choice: tools.length > 0 ? toolChoice : void 0,
         parallel_tool_calls: tools.length > 0 ? openaiToolOptions?.parallelToolCalls : void 0,
+        ...buildOpenAITokenParams(
+          modelIdForPayload,
+          request.config?.maxTokens ?? this.config.maxTokens,
+          request.config?.temperature ?? this.config.temperature
+        ),
+        response_format: toOpenAIResponseFormat(request.config?.responseFormat),
         stream: true,
-        stream_options: { include_usage: true },
-        // o-series: use max_completion_tokens + reasoning_effort, no temperature
-        // regular models: use max_tokens + temperature
-        ...isOSeries ? {
-          max_completion_tokens: maxTokensValue,
-          reasoning_effort: request.config?.reasoningEffort ?? "medium"
-        } : {
-          temperature: request.config?.temperature ?? this.config.temperature,
-          max_tokens: maxTokensValue
-        },
-        // Non-OpenAI OpenRouter models support OR's reasoning/include_reasoning params.
-        // When disableThinking=true we must explicitly send include_reasoning:false because
-        // models like Qwen3 and DeepSeek-R1 reason by default even without the reasoning param.
-        ...isOpenRouter && !isOpenAIOnOpenRouter ? this.config.disableThinking ? { include_reasoning: false } : { reasoning: { max_tokens: 8e3 }, include_reasoning: true } : {}
+        stream_options: { include_usage: true }
       };
       logProviderPayload("openai", "request payload", payload, request.debug);
       const stream = await client.chat.completions.create(payload);
@@ -903,7 +743,6 @@ var OpenAIAdapter = class _OpenAIAdapter {
       const collectedCitations = [];
       let citationIndex = 0;
       let usage;
-      let adapterReasoningStarted = false;
       for await (const chunk of stream) {
         logProviderPayload("openai", "stream chunk", chunk, request.debug);
         if (request.signal?.aborted) {
@@ -914,22 +753,6 @@ var OpenAIAdapter = class _OpenAIAdapter {
         if (delta?.content) {
           yield { type: "message:delta", content: delta.content };
         }
-        if (isOpenRouter) {
-          const rc = delta?.reasoning_content ?? delta?.reasoning ?? null;
-          if (rc) {
-            const rcText = typeof rc === "string" ? rc : Array.isArray(rc) && rc[0]?.text ? rc[0].text : "";
-            if (rcText) {
-              if (!adapterReasoningStarted) {
-                yield { type: "thinking:start" };
-                adapterReasoningStarted = true;
-              }
-              yield { type: "thinking:delta", content: rcText };
-            }
-          } else if (adapterReasoningStarted && (delta?.content || choice?.finish_reason)) {
-            yield { type: "thinking:end" };
-            adapterReasoningStarted = false;
-          }
-        }
         const annotations = delta?.annotations;
         if (annotations && annotations.length > 0) {
           for (const annotation of annotations) {
@@ -977,11 +800,6 @@ var OpenAIAdapter = class _OpenAIAdapter {
               };
             } else if (currentToolCall && toolCall.function?.arguments) {
               currentToolCall.arguments += toolCall.function.arguments;
-              yield {
-                type: "action:args",
-                id: currentToolCall.id,
-                args: currentToolCall.arguments
-              };
             }
           }
         }
@@ -1057,24 +875,20 @@ var OpenAIAdapter = class _OpenAIAdapter {
         name: openaiToolOptions.toolChoice.name
       }
     } : openaiToolOptions?.toolChoice;
-    const activeModel2 = request.config?.model || this.model;
-    const modelSlug2 = activeModel2.replace("openai/", "");
-    const isOSeries2 = /^o[1-9]/.test(modelSlug2);
-    const maxTokensValue2 = request.config?.maxTokens ?? this.config.maxTokens;
+    const modelIdForCompletePayload = request.config?.model || this.model;
     const payload = {
-      model: activeModel2,
+      model: modelIdForCompletePayload,
       messages,
       tools: tools.length > 0 ? tools : void 0,
       tool_choice: tools.length > 0 ? toolChoice : void 0,
       parallel_tool_calls: tools.length > 0 ? openaiToolOptions?.parallelToolCalls : void 0,
-      stream: false,
-      ...isOSeries2 ? {
-        max_completion_tokens: maxTokensValue2,
-        reasoning_effort: request.config?.reasoningEffort ?? "medium"
-      } : {
-        temperature: request.config?.temperature ?? this.config.temperature,
-        max_tokens: maxTokensValue2
-      }
+      ...buildOpenAITokenParams(
+        modelIdForCompletePayload,
+        request.config?.maxTokens ?? this.config.maxTokens,
+        request.config?.temperature ?? this.config.temperature
+      ),
+      response_format: toOpenAIResponseFormat(request.config?.responseFormat),
+      stream: false
     };
     logProviderPayload("openai", "request payload", payload, request.debug);
     const response = await client.chat.completions.create(payload);
@@ -1349,7 +1163,9 @@ var AnthropicAdapter = class {
    * Build common request options for both streaming and non-streaming
    */
   buildRequestOptions(request) {
-    const systemMessage = request.systemPrompt || "";
+    const responseFormat = request.config?.responseFormat;
+    const jsonObjectSuffix = responseFormat?.type === "json_object" ? "\n\nRespond with a single JSON object and no other text." : "";
+    const systemMessage = (request.systemPrompt || "") + jsonObjectSuffix;
     let messages;
     if (request.rawMessages && request.rawMessages.length > 0) {
       messages = this.convertToAnthropicMessages(request.rawMessages);
@@ -1430,6 +1246,10 @@ var AnthropicAdapter = class {
     if (serverToolConfiguration) {
       options.server_tool_configuration = serverToolConfiguration;
     }
+    const outputConfig = toAnthropicOutputConfig(responseFormat);
+    if (outputConfig) {
+      options.output_config = outputConfig;
+    }
     if (this.config.thinking?.type === "enabled") {
       options.thinking = {
         type: "enabled",
@@ -1584,13 +1404,6 @@ var AnthropicAdapter = class {
               yield { type: "thinking:delta", content: event.delta.thinking };
             } else if (event.delta.type === "input_json_delta" && currentToolUse) {
               currentToolUse.input += event.delta.partial_json;
-              if (currentToolUse.name !== "web_search") {
-                yield {
-                  type: "action:args",
-                  id: currentToolUse.id,
-                  args: currentToolUse.input
-                };
-              }
             }
             break;
           case "content_block_stop":
@@ -1797,12 +1610,14 @@ var OllamaAdapter = class {
       if (this.config.options) {
         Object.assign(ollamaOptions, this.config.options);
       }
+      const ollamaFormat = toOllamaFormat(request.config?.responseFormat);
       const payload = {
         model: request.config?.model || this.model,
         messages,
         tools,
         stream: true,
-        options: ollamaOptions
+        options: ollamaOptions,
+        ...ollamaFormat !== void 0 ? { format: ollamaFormat } : {}
       };
       logProviderPayload("ollama", "request payload", payload, request.debug);
       const response = await fetch(`${this.baseUrl}/api/chat`, {
@@ -2093,6 +1908,12 @@ var GoogleAdapter = class {
     }
     const messageId = generateMessageId();
     yield { type: "message:start", id: messageId };
+    const responseFormat = request.config?.responseFormat;
+    const geminiSchema = toGeminiSchema(responseFormat);
+    const responseFormatGenConfig = responseFormat ? {
+      responseMimeType: "application/json",
+      ...geminiSchema ? { responseJsonSchema: geminiSchema } : {}
+    } : {};
     try {
       logProviderPayload(
         "google",
@@ -2104,7 +1925,8 @@ var GoogleAdapter = class {
           tools: toolsArray.length > 0 ? toolsArray : void 0,
           generationConfig: {
             temperature: request.config?.temperature ?? this.config.temperature,
-            maxOutputTokens: request.config?.maxTokens ?? this.config.maxTokens
+            maxOutputTokens: request.config?.maxTokens ?? this.config.maxTokens,
+            ...responseFormatGenConfig
           },
           messageParts: mergedContents[mergedContents.length - 1]?.parts
         },
@@ -2117,7 +1939,8 @@ var GoogleAdapter = class {
         tools: toolsArray.length > 0 ? toolsArray : void 0,
         generationConfig: {
           temperature: request.config?.temperature ?? this.config.temperature,
-          maxOutputTokens: request.config?.maxTokens ?? this.config.maxTokens
+          maxOutputTokens: request.config?.maxTokens ?? this.config.maxTokens,
+          ...responseFormatGenConfig
         }
       });
       const lastMessage = mergedContents[mergedContents.length - 1];
@@ -2284,6 +2107,12 @@ var GoogleAdapter = class {
       }
     }
     const tools = formatToolsForGemini(request.actions);
+    const responseFormat = request.config?.responseFormat;
+    const geminiSchema = toGeminiSchema(responseFormat);
+    const responseFormatGenConfig = responseFormat ? {
+      responseMimeType: "application/json",
+      ...geminiSchema ? { responseJsonSchema: geminiSchema } : {}
+    } : {};
     const payload = {
       model: modelId,
       history: mergedContents.slice(0, -1),
@@ -2291,7 +2120,8 @@ var GoogleAdapter = class {
       tools: tools ? [tools] : void 0,
       generationConfig: {
         temperature: request.config?.temperature ?? this.config.temperature,
-        maxOutputTokens: request.config?.maxTokens ?? this.config.maxTokens
+        maxOutputTokens: request.config?.maxTokens ?? this.config.maxTokens,
+        ...responseFormatGenConfig
       },
       messageParts: mergedContents[mergedContents.length - 1]?.parts
     };
@@ -2302,7 +2132,8 @@ var GoogleAdapter = class {
       tools: tools ? [tools] : void 0,
       generationConfig: {
         temperature: request.config?.temperature ?? this.config.temperature,
-        maxOutputTokens: request.config?.maxTokens ?? this.config.maxTokens
+        maxOutputTokens: request.config?.maxTokens ?? this.config.maxTokens,
+        ...responseFormatGenConfig
       }
     });
     const lastMessage = mergedContents[mergedContents.length - 1];
@@ -2441,6 +2272,7 @@ var AzureAdapter = class {
         tools,
         temperature: request.config?.temperature ?? this.config.temperature,
         max_tokens: request.config?.maxTokens ?? this.config.maxTokens,
+        response_format: toOpenAIResponseFormat(request.config?.responseFormat),
         stream: true
       };
       logProviderPayload("azure", "request payload", payload, request.debug);
@@ -2540,7 +2372,8 @@ var AzureAdapter = class {
       messages,
       tools,
       temperature: request.config?.temperature ?? this.config.temperature,
-      max_tokens: request.config?.maxTokens ?? this.config.maxTokens
+      max_tokens: request.config?.maxTokens ?? this.config.maxTokens,
+      response_format: toOpenAIResponseFormat(request.config?.responseFormat)
     };
     logProviderPayload("azure", "request payload", payload, request.debug);
     const response = await client.chat.completions.create(payload);