@librechat/agents 3.1.74 → 3.1.75-dev.0

package/src/llm/custom-chat-models.smoke.test.ts

@@ -0,0 +1,662 @@
+import { AIMessage, AIMessageChunk } from '@langchain/core/messages';
+import type { OpenAIChatInput, OpenAIClient } from '@langchain/openai';
+import type { ChatOpenRouterCallOptions } from '@/llm/openrouter';
+import type { CustomAnthropicInput } from '@/llm/anthropic';
+import type {
+  ChatAnthropicToolType,
+  AnthropicMCPServerURLDefinition,
+} from '@/llm/anthropic/types';
+import {
+  ChatXAI,
+  ChatOpenAI,
+  ChatDeepSeek,
+  ChatMoonshot,
+  AzureChatOpenAI,
+  CustomOpenAIClient,
+  CustomAzureOpenAIClient,
+} from '@/llm/openai';
+import { CustomChatGoogleGenerativeAI } from '@/llm/google';
+import { CustomChatBedrockConverse } from '@/llm/bedrock';
+import { ChatOpenRouter } from '@/llm/openrouter';
+import { CustomAnthropic } from '@/llm/anthropic';
+import { ChatVertexAI } from '@/llm/vertexai';
+
+type OpenAIRequestOptions = Parameters<ChatOpenAI['_getClientOptions']>[0];
+type OpenAIRequestOptionsWithBaseURL = ReturnType<
+  ChatXAI['_getClientOptions']
+> & {
+  baseURL?: string;
+};
+type OpenAIResponsesDelegate = {
+  client?: unknown;
+  _getClientOptions: (options?: OpenAIRequestOptions) => OpenAIRequestOptions;
+};
+type AnthropicCallOptions = Parameters<
+  CustomAnthropic['invocationParams']
+>[0] & {
+  outputConfig?: CustomAnthropicInput['outputConfig'] & {
+    task_budget?: { type: 'token_budget'; value: number };
+  };
+  inferenceGeo?: CustomAnthropicInput['inferenceGeo'];
+  betas?: CustomAnthropicInput['betas'];
+  container?: string;
+  mcp_servers?: AnthropicMCPServerURLDefinition[];
+  tools?: ChatAnthropicToolType[];
+};
+type AzureReasoningModel = AzureChatOpenAI & {
+  reasoning?: { effort: 'low' | 'high' };
+};
+type OpenRouterFields = Partial<
+  ChatOpenRouterCallOptions & Pick<OpenAIChatInput, 'model' | 'apiKey'>
+>;
+type CompletionDelegate = {
+  completionWithRetry: (request: { messages?: unknown }) => Promise<unknown>;
+};
+type CompletionBackedModel = {
+  completions: CompletionDelegate;
+};
+type StreamingCompletionRequest = {
+  messages?: unknown;
+  stream?: boolean;
+};
+type StreamingCompletionDelegate = {
+  completionWithRetry: (
+    request: StreamingCompletionRequest
+  ) => Promise<
+    AsyncIterable<OpenAIClient.Chat.Completions.ChatCompletionChunk>
+  >;
+};
+type StreamingCompletionBackedModel = {
+  completions: StreamingCompletionDelegate;
+};
+type OpenRouterReasoningStreamDelta =
+  OpenAIClient.Chat.Completions.ChatCompletionChunk.Choice.Delta & {
+    reasoning_details?: Array<
+      | {
+          type: 'reasoning.text';
+          text?: string;
+          format?: string;
+          index?: number;
+        }
+      | {
+          type: 'reasoning.encrypted';
+          id?: string;
+          data?: string;
+          format?: string;
+          index?: number;
+        }
+    >;
+  };
+type OpenRouterReasoningStreamChoice = Omit<
+  OpenAIClient.Chat.Completions.ChatCompletionChunk.Choice,
+  'delta'
+> & {
+  delta: OpenRouterReasoningStreamDelta;
+};
+
+const baseAzureFields = {
+  azureOpenAIApiKey: 'test-azure-key',
+  azureOpenAIApiVersion: '2024-10-21',
+  azureOpenAIApiInstanceName: 'test-instance',
+  azureOpenAIApiDeploymentName: 'test-deployment',
+};
+
+const baseBedrockFields = {
+  region: 'us-east-1',
+  credentials: {
+    accessKeyId: 'test-access-key',
+    secretAccessKey: 'test-secret-key',
+  },
+};
+
+describe('custom chat model class smoke tests', () => {
+  it('keeps the custom OpenAI client, stream delay, and reasoning precedence', () => {
+    const model = new ChatOpenAI({
+      model: 'gpt-5',
+      apiKey: 'test-key',
+      reasoning: { effort: 'low' },
+      _lc_stream_delay: 3,
+    });
+
+    const requestOptions = model._getClientOptions({
+      headers: { 'x-smoke': 'openai' },
+    } as OpenAIRequestOptions);
+
+    expect(ChatOpenAI.lc_name()).toBe('LibreChatOpenAI');
+    expect(model._lc_stream_delay).toBe(3);
+    expect(model.exposedClient).toBeInstanceOf(CustomOpenAIClient);
+    expect(requestOptions.headers).toEqual(
+      expect.objectContaining({ 'x-smoke': 'openai' })
+    );
+    expect(model.getReasoningParams({ reasoning: { effort: 'high' } })).toEqual(
+      { effort: 'high' }
+    );
+    const params = model.invocationParams({ reasoningEffort: 'medium' }) as {
+      reasoning?: unknown;
+      reasoning_effort?: unknown;
+    };
+    expect(params.reasoning ?? { effort: params.reasoning_effort }).toEqual({
+      effort: 'low',
+    });
+  });
+
+  it('keeps the custom OpenAI client on the Responses API path', () => {
+    const model = new ChatOpenAI({
+      model: 'gpt-5',
+      apiKey: 'test-key',
+      useResponsesApi: true,
+      maxTokens: 32,
+      reasoning: { effort: 'low' },
+    });
+
+    const params = model.invocationParams({
+      reasoningEffort: 'high',
+    }) as {
+      max_output_tokens?: number;
+      max_completion_tokens?: number;
+      reasoning?: unknown;
+    };
+    const responses = (
+      model as unknown as { responses: OpenAIResponsesDelegate }
+    ).responses;
+    const requestOptions = responses._getClientOptions({
+      headers: { 'x-smoke': 'responses' },
+    } as OpenAIRequestOptions);
+
+    expect(params.max_output_tokens).toBe(32);
+    expect(params.max_completion_tokens).toBeUndefined();
+    expect(params.reasoning).toEqual({ effort: 'low' });
+    expect(responses.client).toBeInstanceOf(CustomOpenAIClient);
+    const responsesClient = responses.client as CustomOpenAIClient;
+    responsesClient.abortHandler = (): void => undefined;
+    expect(model.exposedClient).toBe(responsesClient);
+    expect(requestOptions?.headers).toEqual(
+      expect.objectContaining({ 'x-smoke': 'responses' })
+    );
+  });
+
+  it('keeps Azure client customization and gates reasoning to reasoning models', () => {
+    const model = new AzureChatOpenAI({
+      ...baseAzureFields,
+      _lc_stream_delay: 4,
+    }) as AzureReasoningModel;
+    model.model = 'gpt-5';
+    model.reasoning = { effort: 'low' };
+
+    const requestOptions = model._getClientOptions({
+      headers: { 'x-smoke': 'azure' },
+    });
+
+    expect(AzureChatOpenAI.lc_name()).toBe('LibreChatAzureOpenAI');
+    expect(model._lc_stream_delay).toBe(4);
+    expect(model.exposedClient).toBeInstanceOf(CustomAzureOpenAIClient);
+    const azureResponses = (
+      model as unknown as { responses: OpenAIResponsesDelegate }
+    ).responses;
+    azureResponses._getClientOptions(undefined);
+    expect(azureResponses.client).toBeInstanceOf(CustomAzureOpenAIClient);
+    const azureResponsesClient =
+      azureResponses.client as CustomAzureOpenAIClient;
+    azureResponsesClient.abortHandler = (): void => undefined;
+    expect(model.exposedClient).toBe(azureResponsesClient);
+    expect(requestOptions.headers).toEqual(
+      expect.objectContaining({
+        'api-key': 'test-azure-key',
+        'x-smoke': 'azure',
+      })
+    );
+    expect(requestOptions.query).toEqual(
+      expect.objectContaining({ 'api-version': '2024-10-21' })
+    );
+    expect(model.getReasoningParams()).toEqual({ effort: 'low' });
+
+    const nonReasoningModel = new AzureChatOpenAI({
+      ...baseAzureFields,
+    }) as AzureReasoningModel;
+    nonReasoningModel.model = 'gpt-4o';
+    nonReasoningModel.reasoning = { effort: 'low' };
+    expect(nonReasoningModel.getReasoningParams()).toBeUndefined();
+  });
+
+  it('keeps DeepSeek, Moonshot, and xAI on LibreChat wrapper semantics', () => {
+    const deepSeek = new ChatDeepSeek({
+      model: 'deepseek-chat',
+      apiKey: 'test-key',
+      _lc_stream_delay: 5,
+    });
+    deepSeek._getClientOptions();
+
+    const moonshot = new ChatMoonshot({
+      model: 'moonshot-v1-8k',
+      apiKey: 'test-key',
+      _lc_stream_delay: 6,
+    });
+
+    const xai = new ChatXAI({
+      model: 'grok-3-fast',
+      apiKey: 'test-key',
+      configuration: { baseURL: 'https://xai.test/v1' },
+      _lc_stream_delay: 7,
+    });
+    const xaiRequestOptions =
+      xai._getClientOptions() as OpenAIRequestOptionsWithBaseURL;
+
+    expect(ChatDeepSeek.lc_name()).toBe('LibreChatDeepSeek');
+    expect(deepSeek._lc_stream_delay).toBe(5);
+    expect(deepSeek.exposedClient).toBeInstanceOf(CustomOpenAIClient);
+    expect(ChatMoonshot.lc_name()).toBe('LibreChatMoonshot');
+    expect(moonshot._lc_stream_delay).toBe(6);
+    expect(ChatXAI.lc_name()).toBe('LibreChatXAI');
+    expect(xai._lc_stream_delay).toBe(7);
+    expect(xai.exposedClient).toBeInstanceOf(CustomOpenAIClient);
+    expect(xaiRequestOptions.baseURL).toBe('https://xai.test/v1');
+  });
+
+  it('keeps Moonshot reasoning content in completion requests', async () => {
+    const moonshot = new ChatMoonshot({
+      model: 'moonshot-v1-8k',
+      apiKey: 'test-key',
+      streaming: false,
+    });
+    const completions = (moonshot as unknown as CompletionBackedModel)
+      .completions;
+    let requestMessages: unknown;
+
+    completions.completionWithRetry = async (request): Promise<unknown> => {
+      requestMessages = request.messages;
+      return {
+        id: 'chatcmpl-test',
+        object: 'chat.completion',
+        created: 0,
+        model: 'moonshot-v1-8k',
+        choices: [
+          {
+            index: 0,
+            finish_reason: 'stop',
+            message: {
+              role: 'assistant',
+              content: 'ok',
+            },
+          },
+        ],
+      };
+    };
+
+    await moonshot.invoke([
+      new AIMessage({
+        content: '',
+        additional_kwargs: { reasoning_content: 'kept-thinking' },
+        tool_calls: [
+          {
+            id: 'call_1',
+            name: 'lookup',
+            args: { q: 'test' },
+            type: 'tool_call',
+          },
+        ],
+      }),
+    ]);
+
+    expect(requestMessages).toEqual([
+      expect.objectContaining({
+        role: 'assistant',
+        content: '',
+        reasoning_content: 'kept-thinking',
+        tool_calls: expect.any(Array),
+      }),
+    ]);
+  });
+
+  it('keeps OpenRouter reasoning isolated from OpenAI reasoning_effort', () => {
+    const fields: OpenRouterFields = {
+      model: 'openrouter/test-model',
+      apiKey: 'test-key',
+      reasoning: { effort: 'xhigh', max_tokens: 2048 },
+    };
+    const model = new ChatOpenRouter(fields);
+
+    const params = model.invocationParams();
+
+    expect(ChatOpenRouter.lc_name()).toBe('LibreChatOpenRouter');
+    expect(params.reasoning).toEqual({ effort: 'xhigh', max_tokens: 2048 });
+    expect(params.reasoning_effort).toBeUndefined();
+
+    const callParams = model.invocationParams({
+      reasoning: { effort: 'low', exclude: true },
+    } as ChatOpenRouterCallOptions);
+    expect(callParams.reasoning).toEqual({
+      effort: 'low',
+      max_tokens: 2048,
+      exclude: true,
+    });
+
+    const legacyModel = new ChatOpenRouter({
+      model: 'openrouter/test-model',
+      apiKey: 'test-key',
+      include_reasoning: true,
+    });
+    expect(legacyModel.invocationParams().reasoning).toEqual({
+      enabled: true,
+    });
+  });
+
+  it('keeps OpenRouter streaming reasoning details stable', async () => {
+    const model = new ChatOpenRouter({
+      model: 'anthropic/claude-sonnet-test',
+      apiKey: 'test-key',
+    });
+    const completions = (model as unknown as StreamingCompletionBackedModel)
+      .completions;
+    let requestMessages: unknown;
+    const createChunk = (
+      choice: OpenRouterReasoningStreamChoice
+    ): OpenAIClient.Chat.Completions.ChatCompletionChunk => ({
+      id: 'chatcmpl-openrouter-test',
+      object: 'chat.completion.chunk',
+      created: 0,
+      model: 'anthropic/claude-sonnet-test',
+      choices: [choice],
+    });
+
+    async function* streamChunks(): AsyncGenerator<OpenAIClient.Chat.Completions.ChatCompletionChunk> {
+      yield createChunk({
+        index: 0,
+        delta: {
+          role: 'assistant',
+          content: '',
+          reasoning_details: [
+            {
+              type: 'reasoning.text',
+              text: 'Think ',
+              format: 'text',
+              index: 0,
+            },
+          ],
+        },
+        finish_reason: null,
+      });
+      yield createChunk({
+        index: 0,
+        delta: {
+          content: 'answer',
+          reasoning_details: [
+            { type: 'reasoning.text', text: 'hard', index: 0 },
+            {
+              type: 'reasoning.encrypted',
+              id: 'sig_1',
+              data: 'encrypted',
+              format: 'anthropic',
+              index: 1,
+            },
+          ],
+        },
+        finish_reason: null,
+      });
+      yield createChunk({
+        index: 0,
+        delta: { content: '' },
+        finish_reason: 'stop',
+      });
+    }
+
+    completions.completionWithRetry = async (
+      request
+    ): Promise<
+      AsyncIterable<OpenAIClient.Chat.Completions.ChatCompletionChunk>
+    > => {
+      requestMessages = request.messages;
+      return streamChunks();
+    };
+
+    const chunks: AIMessageChunk[] = [];
+    const stream = await model.stream([
+      new AIMessage({
+        content: '',
+        additional_kwargs: {
+          reasoning_details: [
+            {
+              type: 'reasoning.text',
+              text: 'previous thought',
+              index: 0,
+            },
+            {
+              type: 'reasoning.encrypted',
+              id: 'prev_sig',
+              data: 'previous encrypted',
+              index: 1,
+            },
+          ],
+        },
+        tool_calls: [
+          {
+            id: 'call_1',
+            name: 'lookup',
+            args: { q: 'test' },
+            type: 'tool_call',
+          },
+        ],
+      }),
+    ]);
+    for await (const chunk of stream) {
+      chunks.push(chunk);
+    }
+
+    expect(requestMessages).toEqual([
+      expect.objectContaining({
+        role: 'assistant',
+        tool_calls: expect.any(Array),
+        content: [
+          expect.objectContaining({
+            type: 'thinking',
+            thinking: 'previous thought',
+          }),
+          expect.objectContaining({
+            type: 'redacted_thinking',
+            data: 'previous encrypted',
+            id: 'prev_sig',
+          }),
+        ],
+      }),
+    ]);
+    expect(chunks).toHaveLength(3);
+    expect(chunks[0].additional_kwargs.reasoning).toBe('Think ');
+    expect(chunks[0].additional_kwargs.reasoning_details).toBeUndefined();
+    expect(chunks[1].additional_kwargs.reasoning).toBe('hard');
+    expect(chunks[1].additional_kwargs.reasoning_details).toBeUndefined();
+    expect(chunks[2].additional_kwargs.reasoning_details).toEqual([
+      {
+        type: 'reasoning.text',
+        text: 'Think hard',
+        format: 'text',
+        index: 0,
+      },
+      {
+        type: 'reasoning.encrypted',
+        id: 'sig_1',
+        data: 'encrypted',
+        format: 'anthropic',
+        index: 1,
+      },
+    ]);
+  });
+
+  it('keeps Anthropic output, residency, compaction, and stream-delay options', () => {
+    const contextManagement = {
+      edits: [
+        {
+          type: 'compact_20260112' as const,
+          trigger: { type: 'input_tokens' as const, value: 50000 },
+        },
+      ],
+    };
+    const model = new CustomAnthropic({
+      model: 'claude-sonnet-4-5-20250929',
+      apiKey: 'test-key',
+      maxTokens: 4096,
+      outputConfig: { effort: 'medium' },
+      inferenceGeo: 'us',
+      contextManagement,
+      _lc_stream_delay: 8,
+    });
+
+    const params = model.invocationParams({
+      outputConfig: { effort: 'low' },
+      inferenceGeo: 'eu',
+    } as AnthropicCallOptions);
+
+    expect(CustomAnthropic.lc_name()).toBe('LibreChatAnthropic');
+    expect(model._lc_stream_delay).toBe(8);
+    expect(params.output_config).toEqual({ effort: 'low' });
+    expect(params.inference_geo).toBe('eu');
+    expect(params.context_management).toEqual(contextManagement);
+  });
+
+  it('keeps Anthropic beta, MCP, and container request wiring current', () => {
+    const contextManagement = {
+      edits: [
+        {
+          type: 'compact_20260112' as const,
+          trigger: { type: 'input_tokens' as const, value: 50000 },
+        },
+      ],
+    };
+    const mcpServers: AnthropicMCPServerURLDefinition[] = [
+      {
+        type: 'url',
+        url: 'https://example.com/mcp',
+        name: 'docs',
+      },
+    ];
+    const model = new CustomAnthropic({
+      model: 'claude-opus-4-7-test',
+      apiKey: 'test-key',
+      maxTokens: 4096,
+      contextManagement,
+      betas: ['model-beta'],
+    });
+
+    const params = model.invocationParams({
+      outputConfig: {
+        effort: 'low',
+        task_budget: { type: 'token_budget', value: 1024 },
+      },
+      betas: ['request-beta', 'model-beta'],
+      container: 'container_123',
+      mcp_servers: mcpServers,
+      tools: [
+        {
+          type: 'tool_search_tool_bm25_20251119',
+          name: 'search',
+        } as ChatAnthropicToolType,
+      ],
+    } as AnthropicCallOptions);
+
+    expect(params.betas).toEqual([
+      'model-beta',
+      'request-beta',
+      'advanced-tool-use-2025-11-20',
+      'compact-2026-01-12',
+      'task-budgets-2026-03-13',
+    ]);
+    expect(params.container).toBe('container_123');
+    expect(params.mcp_servers).toBe(mcpServers);
+    expect(params.temperature).toBeUndefined();
+    expect(params.top_k).toBeUndefined();
+    expect(params.top_p).toBeUndefined();
+  });
+
+  it('matches Anthropic Opus 4.7 sampling compatibility checks', () => {
+    const thinkingModel = new CustomAnthropic({
+      model: 'claude-opus-4-7-test',
+      apiKey: 'test-key',
+      maxTokens: 4096,
+      thinking: { type: 'enabled', budget_tokens: 1024 },
+    });
+    const topKModel = new CustomAnthropic({
+      model: 'claude-opus-4-7-test',
+      apiKey: 'test-key',
+      maxTokens: 4096,
+      topK: 5,
+    });
+
+    expect(() => thinkingModel.invocationParams()).toThrow(
+      'thinking.type="enabled" is not supported for claude-opus-4-7'
+    );
+    expect(() => topKModel.invocationParams()).toThrow(
+      'topK is not supported for claude-opus-4-7'
+    );
+  });
+
+  it('keeps Bedrock Converse application profiles and service tier passthroughs', () => {
+    const applicationInferenceProfile =
+      'arn:aws:bedrock:eu-west-1:123456789012:application-inference-profile/test-profile';
+    const model = new CustomChatBedrockConverse({
+      ...baseBedrockFields,
+      model: 'anthropic.claude-3-haiku-20240307-v1:0',
+      applicationInferenceProfile,
+      serviceTier: 'priority',
+    });
+
+    expect(CustomChatBedrockConverse.lc_name()).toBe(
+      'LibreChatBedrockConverse'
+    );
+    expect(model.applicationInferenceProfile).toBe(applicationInferenceProfile);
+    expect(model.invocationParams({}).serviceTier).toEqual({
+      type: 'priority',
+    });
+    expect(model.invocationParams({ serviceTier: 'flex' }).serviceTier).toEqual(
+      { type: 'flex' }
+    );
+  });
+
+  it('keeps Google and Vertex thinking configuration wiring offline', () => {
+    const thinkingConfig = {
+      thinkingLevel: 'HIGH' as const,
+      includeThoughts: true,
+    };
+    const google = new CustomChatGoogleGenerativeAI({
+      model: 'models/gemini-3-pro-preview',
+      apiKey: 'test-key',
+      thinkingConfig,
+    });
+    const vertex = new ChatVertexAI({
+      model: 'gemini-3-pro-preview',
+      location: 'global',
+      thinkingBudget: -1,
+      thinkingConfig,
+    });
+
+    expect(CustomChatGoogleGenerativeAI.lc_name()).toBe(
+      'LibreChatGoogleGenerativeAI'
+    );
+    expect(google.model).toBe('gemini-3-pro-preview');
+    expect(google._isMultimodalModel).toBe(true);
+    expect(google.thinkingConfig).toEqual(thinkingConfig);
+    expect(ChatVertexAI.lc_name()).toBe('LibreChatVertexAI');
+    expect(vertex.dynamicThinkingBudget).toBe(true);
+    expect(vertex.thinkingConfig).toEqual(thinkingConfig);
+    expect(vertex.invocationParams({}).maxReasoningTokens).toBe(-1);
+  });
+
+  it('uppercases custom OpenAI fetch methods before dispatch', async () => {
+    let method: string | undefined;
+    const client = new CustomOpenAIClient({
+      apiKey: 'test-key',
+      fetch: async (_url, init): Promise<Response> => {
+        method = init?.method;
+        return new Response('{}', { status: 200 });
+      },
+    });
+
+    const response = await client.fetchWithTimeout(
+      'https://example.test/v1/chat/completions',
+      { method: 'patch' },
+      1000,
+      new AbortController()
+    );
+
+    expect(response.status).toBe(200);
+    expect(method).toBe('PATCH');
+    expect(client.abortHandler).toBeDefined();
+  });
+});