🎚️ feat: Reasoning Parameters for Custom Endpoints (#10297)

api/package.json

@@ -48,7 +48,7 @@
     "@langchain/google-genai": "^0.2.13",
     "@langchain/google-vertexai": "^0.2.13",
     "@langchain/textsplitters": "^0.1.0",
-    "@librechat/agents": "^2.4.89",
+    "@librechat/agents": "^2.4.90",
     "@librechat/api": "*",
     "@librechat/data-schemas": "*",
     "@microsoft/microsoft-graph-client": "^3.0.7",

api/server/services/Endpoints/openAI/initialize.js

@@ -143,7 +143,7 @@ const initializeClient = async ({
     modelOptions.model = modelName;
     clientOptions = Object.assign({ modelOptions }, clientOptions);
     clientOptions.modelOptions.user = req.user.id;
-    const options = getOpenAIConfig(apiKey, clientOptions);
+    const options = getOpenAIConfig(apiKey, clientOptions, endpoint);
     if (options != null && serverless === true) {
       options.useLegacyContent = true;
     }

package-lock.json

@@ -64,7 +64,7 @@
         "@langchain/google-genai": "^0.2.13",
         "@langchain/google-vertexai": "^0.2.13",
         "@langchain/textsplitters": "^0.1.0",
-        "@librechat/agents": "^2.4.89",
+        "@librechat/agents": "^2.4.90",
         "@librechat/api": "*",
         "@librechat/data-schemas": "*",
         "@microsoft/microsoft-graph-client": "^3.0.7",
@@ -21690,9 +21690,9 @@
       }
     },
     "node_modules/@librechat/agents": {
-      "version": "2.4.89",
+      "version": "2.4.90",
-      "resolved": "https://registry.npmjs.org/@librechat/agents/-/agents-2.4.89.tgz",
+      "resolved": "https://registry.npmjs.org/@librechat/agents/-/agents-2.4.90.tgz",
-      "integrity": "sha512-QMqaNkkfcDHI8mpaqpgdb1Zz3KT3uVLaaU4NCeXafNH6JvGdG4ueORQH2dM8xtVm3+5DEXwauTdSAi5gHV5tJQ==",
+      "integrity": "sha512-CZI0K0NjIO1mvw4f4tAAMN8x4fFXFyZyrP+NLTCHDvdrIpyTY2k9qRwQHneCGnMgltuUQ+53eXbS3s9psjsAOA==",
       "license": "MIT",
       "dependencies": {
         "@langchain/anthropic": "^0.3.26",
@@ -51984,7 +51984,7 @@
     },
     "packages/api": {
       "name": "@librechat/api",
-      "version": "1.4.1",
+      "version": "1.5.0",
       "license": "ISC",
       "devDependencies": {
         "@babel/preset-env": "^7.21.5",
@@ -52023,7 +52023,7 @@
         "@azure/storage-blob": "^12.27.0",
         "@keyv/redis": "^4.3.3",
         "@langchain/core": "^0.3.62",
-        "@librechat/agents": "^2.4.89",
+        "@librechat/agents": "^2.4.90",
         "@librechat/data-schemas": "*",
         "@modelcontextprotocol/sdk": "^1.17.1",
         "axios": "^1.12.1",

packages/api/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@librechat/api",
-  "version": "1.4.1",
+  "version": "1.5.0",
   "type": "commonjs",
   "description": "MCP services for LibreChat",
   "main": "dist/index.js",
@@ -80,7 +80,7 @@
     "@azure/storage-blob": "^12.27.0",
     "@keyv/redis": "^4.3.3",
     "@langchain/core": "^0.3.62",
-    "@librechat/agents": "^2.4.89",
+    "@librechat/agents": "^2.4.90",
     "@librechat/data-schemas": "*",
     "@modelcontextprotocol/sdk": "^1.17.1",
     "axios": "^1.12.1",

packages/api/src/endpoints/openai/config.spec.ts

@@ -1,4 +1,9 @@
-import { Verbosity, ReasoningEffort, ReasoningSummary } from 'librechat-data-provider';
+import {
+  Verbosity,
+  EModelEndpoint,
+  ReasoningEffort,
+  ReasoningSummary,
+} from 'librechat-data-provider';
 import type { RequestInit } from 'undici';
 import type { OpenAIParameters, AzureOptions } from '~/types';
 import { getOpenAIConfig } from './config';
@@ -103,12 +108,89 @@ describe('getOpenAIConfig', () => {
 
     const result = getOpenAIConfig(mockApiKey, { modelOptions });
 
+    /** When no endpoint is specified, it's treated as non-openAI/azureOpenAI, so uses reasoning object */
+    expect(result.llmConfig.reasoning).toEqual({
+      effort: ReasoningEffort.high,
+      summary: ReasoningSummary.detailed,
+    });
+    expect((result.llmConfig as Record<string, unknown>).reasoning_effort).toBeUndefined();
+  });
+
+  it('should use reasoning_effort for openAI endpoint without useResponsesApi', () => {
+    const modelOptions = {
+      reasoning_effort: ReasoningEffort.high,
+      reasoning_summary: ReasoningSummary.detailed,
+    };
+
+    const result = getOpenAIConfig(mockApiKey, { modelOptions }, EModelEndpoint.openAI);
+
     expect((result.llmConfig as Record<string, unknown>).reasoning_effort).toBe(
       ReasoningEffort.high,
     );
     expect(result.llmConfig.reasoning).toBeUndefined();
   });
 
+  it('should use reasoning_effort for azureOpenAI endpoint without useResponsesApi', () => {
+    const modelOptions = {
+      reasoning_effort: ReasoningEffort.high,
+      reasoning_summary: ReasoningSummary.detailed,
+    };
+
+    const result = getOpenAIConfig(mockApiKey, { modelOptions }, EModelEndpoint.azureOpenAI);
+
+    expect((result.llmConfig as Record<string, unknown>).reasoning_effort).toBe(
+      ReasoningEffort.high,
+    );
+    expect(result.llmConfig.reasoning).toBeUndefined();
+  });
+
+  it('should use reasoning object for openAI endpoint with useResponsesApi=true', () => {
+    const modelOptions = {
+      reasoning_effort: ReasoningEffort.high,
+      reasoning_summary: ReasoningSummary.detailed,
+      useResponsesApi: true,
+    };
+
+    const result = getOpenAIConfig(mockApiKey, { modelOptions }, EModelEndpoint.openAI);
+
+    expect(result.llmConfig.reasoning).toEqual({
+      effort: ReasoningEffort.high,
+      summary: ReasoningSummary.detailed,
+    });
+    expect((result.llmConfig as Record<string, unknown>).reasoning_effort).toBeUndefined();
+  });
+
+  it('should use reasoning object for azureOpenAI endpoint with useResponsesApi=true', () => {
+    const modelOptions = {
+      reasoning_effort: ReasoningEffort.high,
+      reasoning_summary: ReasoningSummary.detailed,
+      useResponsesApi: true,
+    };
+
+    const result = getOpenAIConfig(mockApiKey, { modelOptions }, EModelEndpoint.azureOpenAI);
+
+    expect(result.llmConfig.reasoning).toEqual({
+      effort: ReasoningEffort.high,
+      summary: ReasoningSummary.detailed,
+    });
+    expect((result.llmConfig as Record<string, unknown>).reasoning_effort).toBeUndefined();
+  });
+
+  it('should use reasoning object for non-openAI/azureOpenAI endpoints', () => {
+    const modelOptions = {
+      reasoning_effort: ReasoningEffort.high,
+      reasoning_summary: ReasoningSummary.detailed,
+    };
+
+    const result = getOpenAIConfig(mockApiKey, { modelOptions }, 'custom-endpoint');
+
+    expect(result.llmConfig.reasoning).toEqual({
+      effort: ReasoningEffort.high,
+      summary: ReasoningSummary.detailed,
+    });
+    expect((result.llmConfig as Record<string, unknown>).reasoning_effort).toBeUndefined();
+  });
+
   it('should handle OpenRouter configuration', () => {
     const reverseProxyUrl = 'https://openrouter.ai/api/v1';
 

packages/api/src/endpoints/openai/config.ts

@@ -68,6 +68,7 @@ export function getOpenAIConfig(
       azure,
       apiKey,
       baseURL,
+      endpoint,
       streaming,
       addParams,
       dropParams,

packages/api/src/endpoints/openai/llm.ts

@@ -1,4 +1,4 @@
-import { removeNullishValues } from 'librechat-data-provider';
+import { EModelEndpoint, removeNullishValues } from 'librechat-data-provider';
 import type { BindToolsInput } from '@langchain/core/language_models/chat_models';
 import type { AzureOpenAIInput } from '@langchain/openai';
 import type { OpenAI } from 'openai';
@@ -79,6 +79,7 @@ export function getOpenAILLMConfig({
   azure,
   apiKey,
   baseURL,
+  endpoint,
   streaming,
   addParams,
   dropParams,
@@ -88,6 +89,7 @@ export function getOpenAILLMConfig({
   apiKey: string;
   streaming: boolean;
   baseURL?: string | null;
+  endpoint?: EModelEndpoint | string | null;
   modelOptions: Partial<t.OpenAIParameters>;
   addParams?: Record<string, unknown>;
   dropParams?: string[];
@@ -155,7 +157,8 @@ export function getOpenAILLMConfig({
 
   if (
     hasReasoningParams({ reasoning_effort, reasoning_summary }) &&
-    (llmConfig.useResponsesApi === true || useOpenRouter)
+    (llmConfig.useResponsesApi === true ||
+      (endpoint !== EModelEndpoint.openAI && endpoint !== EModelEndpoint.azureOpenAI))
   ) {
     llmConfig.reasoning = removeNullishValues(
       {