fix: bridge Azure Chat tools through Responses

src/handlers/handlerUtils.ts

@@ -851,6 +851,7 @@ export function constructConfigFromRequestHeaders(
       requestHeaders[`x-${POWERED_BY}-azure-entra-client-secret`],
     azureEntraTenantId: requestHeaders[`x-${POWERED_BY}-azure-entra-tenant-id`],
     azureModelName: requestHeaders[`x-${POWERED_BY}-azure-model-name`],
+    chatCompletionsApi: requestHeaders[`x-${POWERED_BY}-chat-completions-api`],
     openaiBeta:
       requestHeaders[`x-${POWERED_BY}-openai-beta`] ||
       requestHeaders[`openai-beta`],

src/providers/azure-openai/api.ts

@@ -6,6 +6,10 @@ import {
   getAzureWorkloadIdentityToken,
 } from './utils';
 import { getRuntimeKey } from 'hono/adapter';
+import {
+  getAzureResponsesChatEndpoint,
+  shouldUseAzureResponsesForChat,
+} from './chatCompletionsResponses';
 
 const runtime = getRuntimeKey();
 
@@ -100,7 +104,12 @@ const AzureOpenAIAPIConfig: ProviderAPIConfig = {
     }
     return headersObj;
   },
-  getEndpoint: ({ providerOptions, fn, gatewayRequestURL }) => {
+  getEndpoint: ({
+    providerOptions,
+    fn,
+    gatewayRequestBodyJSON,
+    gatewayRequestURL,
+  }) => {
     const { apiVersion, urlToFetch, deploymentId } = providerOptions;
     let mappedFn = fn;
 
@@ -140,6 +149,13 @@ const AzureOpenAIAPIConfig: ProviderAPIConfig = {
       searchParams.delete('api-version');
     }
 
+    if (
+      mappedFn === 'chatComplete' &&
+      shouldUseAzureResponsesForChat(gatewayRequestBodyJSON, providerOptions)
+    ) {
+      return getAzureResponsesChatEndpoint();
+    }
+
     switch (mappedFn) {
       case 'complete': {
         return `${prefix}/completions?${searchParams.toString()}`;

src/providers/azure-openai/chatComplete.ts

@@ -5,6 +5,10 @@ import {
   ErrorResponse,
   ProviderConfig,
 } from '../types';
+import {
+  hasFunctionTools,
+  isAzureResponsesChatCompatibilityEnabled,
+} from './chatCompletionsResponses';
 import { getAzureModelValue } from './utils';
 
 // TODOS: this configuration does not enforce the maximum token limit for the input parameter. If you want to enforce this, you might need to add a custom validation function or a max property to the ParameterConfig interface, and then use it in the input configuration. However, this might be complex because the token count is not a simple length check, but depends on the specific tokenization method used by the model.
@@ -102,6 +106,12 @@ export const AzureOpenAIChatCompleteConfig: ProviderConfig = {
   },
   reasoning_effort: {
     param: 'reasoning_effort',
+    transform: (params, providerOptions) =>
+      params.reasoning_effort === 'none' &&
+      isAzureResponsesChatCompatibilityEnabled(providerOptions) &&
+      hasFunctionTools(params)
+        ? undefined
+        : params.reasoning_effort,
   },
   stream_options: {
     param: 'stream_options',