@librechat/agents 3.4.7 → 3.5.1

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

@@ -1,1966 +0,0 @@
-import { convertMessagesToResponsesInput } from '@langchain/openai';
-import {
-  AIMessage,
-  AIMessageChunk,
-  HumanMessage,
-} from '@langchain/core/messages';
-import type { CallbackManagerForLLMRun } from '@langchain/core/callbacks/manager';
-import type { OpenAIChatInput, OpenAIClient } from '@langchain/openai';
-import type { ChatGenerationChunk } from '@langchain/core/outputs';
-import type {
-  ChatAnthropicToolType,
-  AnthropicMCPServerURLDefinition,
-} from '@/llm/anthropic/types';
-import type { ChatOpenRouterCallOptions } from '@/llm/openrouter';
-import type { CustomAnthropicInput } from '@/llm/anthropic';
-import {
-  ChatXAI,
-  ChatOpenAI,
-  ChatDeepSeek,
-  ChatMoonshot,
-  AzureChatOpenAI,
-  CustomOpenAIClient,
-  CustomAzureOpenAIClient,
-} from '@/llm/openai';
-import {
-  OPENAI_RESPONSES_REPLAY_POSITIONS_KEY,
-  projectOpenAIResponsesToolMessageContent,
-} from '@/messages/core';
-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 StreamingOpenAIResponsesDelegate = OpenAIResponsesDelegate & {
-  completionWithRetry: (
-    request: OpenAIClient.Responses.ResponseCreateParamsStreaming
-  ) => Promise<AsyncIterable<OpenAIClient.Responses.ResponseStreamEvent>>;
-};
-type MockableResponsesDelegate = OpenAIResponsesDelegate & {
-  completionWithRetry: (request: {
-    input?: unknown;
-    stream?: boolean;
-  }) => Promise<unknown>;
-  _generate: (
-    messages: AIMessage[],
-    options: Record<string, unknown>
-  ) => Promise<unknown>;
-  _streamChatModelEvents: (
-    messages: AIMessage[],
-    options: Record<string, unknown>
-  ) => AsyncGenerator<unknown>;
-  _streamResponseChunks: (
-    messages: AIMessage[],
-    options: Record<string, unknown>
-  ) => AsyncGenerator<ChatGenerationChunk>;
-};
-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 OpenAIStreamEvent = {
-  event: string;
-  data?: unknown;
-};
-type OpenAIStreamItem =
-  | OpenAIClient.Chat.Completions.ChatCompletionChunk
-  | OpenAIStreamEvent;
-type FetchOutcome = 'resolved' | 'rejected' | 'pending';
-type AbortableFetchCapture = {
-  fetch: typeof fetch;
-  getSignal: () => AbortSignal | undefined;
-};
-type FetchTimeoutClient = {
-  fetchWithTimeout: (
-    url: RequestInfo,
-    init: RequestInit | undefined,
-    ms: number,
-    controller: AbortController
-  ) => Promise<Response>;
-};
-type MockableCompletionCreate = (
-  request: unknown,
-  options?: unknown
-) => Promise<
-  AsyncIterable<OpenAIStreamItem> | OpenAIClient.Chat.Completions.ChatCompletion
->;
-type MockableCompletionClient = {
-  chat: {
-    completions: {
-      create: MockableCompletionCreate;
-    };
-  };
-};
-type MockableCompletionDelegate = OpenAIResponsesDelegate & {
-  client?: MockableCompletionClient;
-};
-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;
-};
-type OpenAICompatibleReasoningStreamDelta =
-  OpenAIClient.Chat.Completions.ChatCompletionChunk.Choice.Delta & {
-    reasoning?: string;
-    reasoning_details?: Array<{
-      type: 'reasoning.text';
-      text?: string;
-      format?: string;
-      index?: number;
-    }>;
-  };
-type OpenAICompatibleReasoningStreamChoice = Omit<
-  OpenAIClient.Chat.Completions.ChatCompletionChunk.Choice,
-  'delta'
-> & {
-  delta: OpenAICompatibleReasoningStreamDelta;
-};
-type PromptTokensDetailsWithCacheWrite = NonNullable<
-  OpenAIClient.Completions.CompletionUsage['prompt_tokens_details']
-> & {
-  cache_write_tokens?: number;
-};
-type CompletionUsageWithCacheWrite = Omit<
-  OpenAIClient.Completions.CompletionUsage,
-  'prompt_tokens_details'
-> & {
-  prompt_tokens_details?: PromptTokensDetailsWithCacheWrite;
-};
-type OpenAIStreamModel = ChatOpenAI | AzureChatOpenAI;
-
-class CallbackTestChatOpenRouter extends ChatOpenRouter {
-  streamChunksWithCallbacks(
-    runManager?: CallbackManagerForLLMRun
-  ): AsyncGenerator<ChatGenerationChunk> {
-    return this._streamResponseChunks(
-      [new HumanMessage('hi')],
-      {} as this['ParsedCallOptions'],
-      runManager
-    );
-  }
-}
-
-const baseAzureFields = {
-  azureOpenAIApiKey: 'test-azure-key',
-  azureOpenAIApiVersion: '2024-10-21',
-  azureOpenAIApiInstanceName: 'test-instance',
-  azureOpenAIApiDeploymentName: 'test-deployment',
-};
-
-function createPersistedPreemptedAzureMessage(): AIMessage {
-  return new AIMessage({
-    id: 'msg_persisted_azure',
-    content: [{ type: 'text', text: 'Partial Azure answer.' }],
-    additional_kwargs: {
-      reasoning: {
-        id: 'rs_persisted_azure',
-        type: 'reasoning',
-        status: 'in_progress',
-        summary: [],
-      },
-      tool_outputs: [
-        {
-          id: 'local_persisted_azure',
-          type: 'local_shell_call_output',
-          status: 'completed',
-          output: 'AZURE_SERVER_RESULT',
-        },
-      ],
-      [OPENAI_RESPONSES_REPLAY_POSITIONS_KEY]: [
-        {
-          itemId: 'local_persisted_azure',
-          kind: 'output',
-          outputIndex: 0,
-        },
-        {
-          itemId: 'rs_persisted_azure',
-          kind: 'reasoning',
-          outputIndex: 1,
-        },
-        {
-          itemId: 'msg_persisted_azure',
-          kind: 'text',
-          outputIndex: 2,
-          contentIndex: 0,
-        },
-      ],
-    },
-    response_metadata: {
-      id: 'resp_persisted_azure',
-      model_provider: 'openai',
-      preempted: true,
-    },
-  });
-}
-
-function expectNeutralizedAzureResponsesRequest(request: unknown): void {
-  const serialized = JSON.stringify(request);
-  expect(serialized).toContain('AZURE_SERVER_RESULT');
-  expect(serialized.match(/AZURE_SERVER_RESULT/g)).toHaveLength(1);
-  expect(serialized).toContain('serverToolResult');
-  expect(serialized).not.toContain('msg_persisted_azure');
-  expect(serialized).not.toContain('rs_persisted_azure');
-  expect(serialized).not.toContain('local_persisted_azure');
-  expect(serialized).not.toContain('resp_persisted_azure');
-  expect(serialized).not.toContain(OPENAI_RESPONSES_REPLAY_POSITIONS_KEY);
-  expect(serialized).not.toContain('local_shell_call_output');
-}
-
-function createCompletedAzureResponse(): OpenAIClient.Responses.Response {
-  return {
-    id: 'resp_fresh_azure',
-    created_at: 0,
-    error: null,
-    incomplete_details: null,
-    instructions: null,
-    metadata: {},
-    model: 'gpt-5',
-    object: 'response',
-    output: [
-      {
-        id: 'msg_fresh_azure',
-        type: 'message',
-        role: 'assistant',
-        status: 'completed',
-        content: [
-          {
-            type: 'output_text',
-            text: 'Fresh Azure answer.',
-            annotations: [],
-          },
-        ],
-      },
-    ],
-    output_text: 'Fresh Azure answer.',
-    parallel_tool_calls: true,
-    status: 'completed',
-    temperature: null,
-    tool_choice: 'auto',
-    tools: [],
-    top_p: null,
-    usage: {
-      input_tokens: 1,
-      input_tokens_details: { cached_tokens: 0, cache_write_tokens: 0 },
-      output_tokens: 1,
-      output_tokens_details: { reasoning_tokens: 0 },
-      total_tokens: 2,
-    },
-  } as OpenAIClient.Responses.Response;
-}
-
-const waitForFetchOutcome = (
-  promise: Promise<Response>,
-  timeoutMs = 100
-): Promise<FetchOutcome> =>
-  Promise.race([
-    promise.then(
-      () => 'resolved' as const,
-      () => 'rejected' as const
-    ),
-    new Promise<'pending'>((resolve) => {
-      setTimeout(() => resolve('pending'), timeoutMs);
-    }),
-  ]);
-
-const createAbortableFetch = (): AbortableFetchCapture => {
-  let requestSignal: AbortSignal | undefined;
-  return {
-    fetch: async (_url, init): Promise<Response> =>
-      new Promise<Response>((_resolve, reject) => {
-        requestSignal = init?.signal ?? undefined;
-        requestSignal?.addEventListener(
-          'abort',
-          () => reject(new Error('Aborted')),
-          { once: true }
-        );
-      }),
-    getSignal: () => requestSignal,
-  };
-};
-
-const expectFetchTimeoutAbort = async (
-  client: FetchTimeoutClient,
-  capturedFetch: AbortableFetchCapture,
-  url: string
-): Promise<void> => {
-  const controller = new AbortController();
-
-  const response = client.fetchWithTimeout(
-    url,
-    { method: 'post' },
-    10,
-    controller
-  );
-
-  await expect(waitForFetchOutcome(response)).resolves.toBe('rejected');
-  expect(controller.signal.aborted).toBe(true);
-  expect(capturedFetch.getSignal()?.aborted).toBe(true);
-};
-
-const baseBedrockFields = {
-  region: 'us-east-1',
-  credentials: {
-    accessKeyId: 'test-access-key',
-    secretAccessKey: 'test-secret-key',
-  },
-};
-
-const createOpenAIStreamChunk = (
-  content: string,
-  finishReason: OpenAIClient.Chat.Completions.ChatCompletionChunk.Choice['finish_reason'] = null
-): OpenAIClient.Chat.Completions.ChatCompletionChunk => ({
-  id: 'chatcmpl-hermes-test',
-  object: 'chat.completion.chunk',
-  created: 0,
-  model: 'hermes-agent',
-  choices: [
-    {
-      index: 0,
-      delta: { content },
-      finish_reason: finishReason,
-    },
-  ],
-});
-
-async function* createOpenAIStreamWithCustomEvents(): AsyncGenerator<OpenAIStreamItem> {
-  yield createOpenAIStreamChunk('Hello ');
-  yield {
-    event: 'hermes.tool.progress',
-    data: {
-      tool: 'execute_code',
-      toolCallId: 'call_1',
-      status: 'running',
-    },
-  };
-  yield {
-    event: 'hermes.tool.progress',
-    data: null,
-  };
-  yield {
-    event: 'message',
-    data: createOpenAIStreamChunk('world', 'stop'),
-  };
-}
-
-function mockCompletionStream(
-  model: OpenAIStreamModel
-): MockableCompletionCreate {
-  const completions = (
-    model as unknown as { completions: MockableCompletionDelegate }
-  ).completions;
-  completions._getClientOptions(undefined);
-  const client = completions.client;
-  if (client == null) {
-    throw new Error('Expected OpenAI completions client');
-  }
-
-  const createMock = jest.fn(async () =>
-    createOpenAIStreamWithCustomEvents()
-  ) as MockableCompletionCreate;
-  client.chat.completions.create = createMock;
-  return createMock;
-}
-
-function mockCompletion(
-  model: ChatOpenAI,
-  response: OpenAIClient.Chat.Completions.ChatCompletion
-): MockableCompletionCreate {
-  const completions = (
-    model as unknown as { completions: MockableCompletionDelegate }
-  ).completions;
-  completions._getClientOptions(undefined);
-  const client = completions.client;
-  if (client == null) {
-    throw new Error('Expected OpenAI completions client');
-  }
-
-  const createMock = jest.fn(async () => response) as MockableCompletionCreate;
-  client.chat.completions.create = createMock;
-  return createMock;
-}
-
-async function expectCustomSSEEventsSkipped(
-  model: OpenAIStreamModel
-): Promise<void> {
-  const createMock = mockCompletionStream(model);
-  const chunks: AIMessageChunk[] = [];
-  const stream = await model.stream([new HumanMessage('use a tool')]);
-  for await (const chunk of stream) {
-    chunks.push(chunk);
-  }
-
-  const text = chunks
-    .map((chunk) => (typeof chunk.content === 'string' ? chunk.content : ''))
-    .join('');
-  expect(chunks).toHaveLength(2);
-  expect(text).toBe('Hello world');
-  expect(createMock).toHaveBeenCalledWith(
-    expect.objectContaining({ stream: true }),
-    expect.any(Object)
-  );
-}
-
-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('retains replayable Responses output items before a terminal event', async () => {
-    const model = new ChatOpenAI({
-      model: 'gpt-5',
-      apiKey: 'test-key',
-      useResponsesApi: true,
-    });
-    const responses = (
-      model as unknown as { responses: StreamingOpenAIResponsesDelegate }
-    ).responses;
-    const items = [
-      {
-        id: 'local_output_item',
-        type: 'local_shell_call_output',
-        status: 'completed',
-        output: 'local output',
-      },
-      {
-        id: 'shell_output_item',
-        call_id: 'shell_call_id',
-        type: 'shell_call_output',
-        status: 'completed',
-        max_output_length: 1_000,
-        output: [
-          {
-            stdout: 'shell output',
-            stderr: '',
-            outcome: { type: 'exit', exit_code: 0 },
-          },
-        ],
-      },
-      {
-        id: 'patch_output_item',
-        call_id: 'patch_call_id',
-        type: 'apply_patch_call_output',
-        status: 'completed',
-        output: 'patch output',
-      },
-      {
-        id: 'program_output_item',
-        call_id: 'program_call_id',
-        type: 'program_output',
-        status: 'completed',
-        result: 'program output',
-      },
-    ] as const;
-    responses.completionWithRetry = async () =>
-      (async function* () {
-        for (let i = 0; i < items.length; i++) {
-          yield {
-            type: 'response.output_item.done',
-            sequence_number: i,
-            output_index: i,
-            item: items[i],
-          } as unknown as OpenAIClient.Responses.ResponseStreamEvent;
-        }
-        yield {
-          type: 'response.output_item.added',
-          sequence_number: items.length,
-          output_index: items.length,
-          item: {
-            id: 'msg_partial',
-            type: 'message',
-            role: 'assistant',
-            status: 'in_progress',
-            content: [],
-          },
-        } as OpenAIClient.Responses.ResponseStreamEvent;
-        yield {
-          type: 'response.output_text.delta',
-          sequence_number: items.length + 1,
-          output_index: items.length,
-          content_index: 0,
-          item_id: 'msg_partial',
-          delta: 'Partial ',
-          logprobs: [],
-        } as OpenAIClient.Responses.ResponseStreamEvent;
-        yield {
-          type: 'response.output_text.delta',
-          sequence_number: items.length + 2,
-          output_index: items.length,
-          content_index: 0,
-          item_id: 'msg_partial',
-          delta: 'answer.',
-          logprobs: [],
-        } as OpenAIClient.Responses.ResponseStreamEvent;
-      })();
-
-    const chunks: AIMessageChunk[] = [];
-    const stream = await model.stream([new HumanMessage('run tools')]);
-    for await (const chunk of stream) {
-      chunks.push(chunk as AIMessageChunk);
-    }
-    const combined = chunks.reduce<AIMessageChunk | undefined>(
-      (current, chunk) =>
-        current == null ? chunk : (current.concat(chunk) as AIMessageChunk),
-      undefined
-    );
-
-    expect(chunks).toHaveLength(items.length + 3);
-    expect(combined?.text).toBe('Partial answer.');
-    expect(combined?.additional_kwargs.tool_outputs).toEqual(items);
-    expect(
-      combined?.additional_kwargs[OPENAI_RESPONSES_REPLAY_POSITIONS_KEY]
-    ).toEqual([
-      ...items.map((item, outputIndex) => ({
-        itemId: item.id,
-        kind: 'output',
-        outputIndex,
-      })),
-      {
-        itemId: 'msg_partial',
-        kind: 'message',
-        outputIndex: items.length,
-      },
-      {
-        contentIndex: 0,
-        itemId: 'msg_partial',
-        kind: 'text',
-        outputIndex: items.length,
-      },
-    ]);
-  });
-
-  it('rehydrates one-chunk Responses replay ordering from constructor kwargs', async () => {
-    const model = new ChatOpenAI({
-      model: 'gpt-5',
-      apiKey: 'test-key',
-      useResponsesApi: true,
-    });
-    const responses = (
-      model as unknown as { responses: StreamingOpenAIResponsesDelegate }
-    ).responses;
-    responses.completionWithRetry = async () =>
-      (async function* () {
-        yield {
-          type: 'response.output_text.delta',
-          sequence_number: 0,
-          output_index: 3,
-          content_index: 2,
-          item_id: 'msg_one_chunk',
-          delta: 'Partial answer.',
-          logprobs: [],
-        } as OpenAIClient.Responses.ResponseStreamEvent;
-      })();
-
-    const stream = await model.stream([new HumanMessage('answer')]);
-    const result = await stream.next();
-    expect(result.done).toBe(false);
-    const chunk = result.value as AIMessageChunk;
-    const serializedFields = JSON.parse(
-      JSON.stringify(chunk.lc_kwargs)
-    ) as ConstructorParameters<typeof AIMessageChunk>[0];
-    const rehydrated = new AIMessageChunk(serializedFields);
-
-    expect(chunk.content).toEqual([
-      { type: 'text', text: 'Partial answer.', index: 0 },
-    ]);
-    expect(rehydrated.content).toEqual([
-      { type: 'text', text: 'Partial answer.', index: 0 },
-    ]);
-    expect(
-      rehydrated.additional_kwargs[OPENAI_RESPONSES_REPLAY_POSITIONS_KEY]
-    ).toEqual([
-      {
-        contentIndex: 2,
-        itemId: 'msg_one_chunk',
-        kind: 'text',
-        outputIndex: 3,
-      },
-    ]);
-  });
-
-  it.each(['response.completed', 'response.incomplete'] as const)(
-    'removes provisional replay outputs after %s supplies the authoritative output',
-    async (terminalType) => {
-      const model = new ChatOpenAI({
-        model: 'gpt-5',
-        apiKey: 'test-key',
-        useResponsesApi: true,
-      });
-      const responses = (
-        model as unknown as { responses: StreamingOpenAIResponsesDelegate }
-      ).responses;
-      const streamedToolOutput = {
-        id: 'web_search_item',
-        type: 'web_search_call',
-        status: 'completed',
-      } as const;
-      const provisionalReplayOutput = {
-        id: 'local_output_item',
-        type: 'local_shell_call_output',
-        status: 'completed',
-        output: 'stale provisional local output',
-      } as const;
-      const authoritativeReplayOutput = {
-        ...provisionalReplayOutput,
-        output: 'unique authoritative local output',
-      } as const;
-      responses.completionWithRetry = async () =>
-        (async function* () {
-          yield {
-            type: 'response.output_item.done',
-            sequence_number: 0,
-            output_index: 0,
-            item: streamedToolOutput,
-          } as unknown as OpenAIClient.Responses.ResponseStreamEvent;
-          yield {
-            type: 'response.output_item.done',
-            sequence_number: 1,
-            output_index: 1,
-            item: provisionalReplayOutput,
-          } as unknown as OpenAIClient.Responses.ResponseStreamEvent;
-          yield {
-            type: terminalType,
-            sequence_number: 2,
-            response: {
-              id: 'resp_terminal',
-              model: 'gpt-5',
-              output: [streamedToolOutput, authoritativeReplayOutput],
-              status:
-                terminalType === 'response.completed'
-                  ? 'completed'
-                  : 'incomplete',
-              usage: null,
-            },
-          } as unknown as OpenAIClient.Responses.ResponseStreamEvent;
-        })();
-
-      let tracedResult: unknown;
-      const chunks: AIMessageChunk[] = [];
-      const stream = await model.stream([new HumanMessage('run tools')], {
-        callbacks: [
-          {
-            handleLLMEnd(result: unknown): void {
-              tracedResult = result;
-            },
-          },
-        ],
-      });
-      for await (const chunk of stream) {
-        chunks.push(chunk as AIMessageChunk);
-      }
-      const combined = chunks.reduce<AIMessageChunk | undefined>(
-        (current, chunk) =>
-          current == null ? chunk : (current.concat(chunk) as AIMessageChunk),
-        undefined
-      );
-
-      expect(combined?.additional_kwargs.tool_outputs).toEqual([
-        streamedToolOutput,
-      ]);
-      expect(combined?.response_metadata.output).toEqual([
-        streamedToolOutput,
-        authoritativeReplayOutput,
-      ]);
-      expect(combined?.toJSON()).toEqual(
-        expect.objectContaining({
-          id: ['langchain_core', 'messages', 'AIMessageChunk'],
-        })
-      );
-      expect(
-        JSON.stringify(combined).match(/unique authoritative local output/g)
-      ).toHaveLength(1);
-      expect(JSON.stringify(combined)).not.toContain(
-        'stale provisional local output'
-      );
-      expect(JSON.stringify(combined?.lc_kwargs)).not.toContain(
-        'stale provisional local output'
-      );
-      expect(combined?.lc_kwargs.additional_kwargs).toEqual(
-        expect.objectContaining({ tool_outputs: [streamedToolOutput] })
-      );
-      expect(combined?.additional_kwargs).not.toHaveProperty(
-        OPENAI_RESPONSES_REPLAY_POSITIONS_KEY
-      );
-      expect(combined?.lc_kwargs.additional_kwargs).not.toHaveProperty(
-        OPENAI_RESPONSES_REPLAY_POSITIONS_KEY
-      );
-      expect(
-        JSON.stringify(tracedResult).match(/unique authoritative local output/g)
-      ).toHaveLength(1);
-      expect(JSON.stringify(tracedResult)).not.toContain(
-        'stale provisional local output'
-      );
-      expect(JSON.stringify(tracedResult)).not.toContain(
-        OPENAI_RESPONSES_REPLAY_POSITIONS_KEY
-      );
-    }
-  );
-
-  it('removes replay positions on a text-only terminal response', async () => {
-    const model = new ChatOpenAI({
-      model: 'gpt-5',
-      apiKey: 'test-key',
-      useResponsesApi: true,
-    });
-    const responses = (
-      model as unknown as { responses: StreamingOpenAIResponsesDelegate }
-    ).responses;
-    const messageItem = {
-      id: 'msg_terminal',
-      type: 'message',
-      role: 'assistant',
-      status: 'completed',
-      content: [
-        {
-          type: 'output_text',
-          text: 'Complete answer.',
-          annotations: [],
-          logprobs: [],
-        },
-      ],
-    } as const;
-    responses.completionWithRetry = async () =>
-      (async function* () {
-        yield {
-          type: 'response.output_text.delta',
-          sequence_number: 0,
-          output_index: 0,
-          content_index: 0,
-          item_id: messageItem.id,
-          delta: messageItem.content[0].text,
-          logprobs: [],
-        } as OpenAIClient.Responses.ResponseStreamEvent;
-        yield {
-          type: 'response.completed',
-          sequence_number: 1,
-          response: {
-            id: 'resp_text_terminal',
-            model: 'gpt-5',
-            output: [messageItem],
-            status: 'completed',
-            usage: null,
-          },
-        } as unknown as OpenAIClient.Responses.ResponseStreamEvent;
-      })();
-
-    let combined: AIMessageChunk | undefined;
-    const stream = await model.stream([new HumanMessage('answer')]);
-    for await (const chunk of stream) {
-      combined =
-        combined == null
-          ? (chunk as AIMessageChunk)
-          : (combined.concat(chunk as AIMessageChunk) as AIMessageChunk);
-    }
-
-    expect(combined?.text).toBe('Complete answer.');
-    expect(combined?.response_metadata.output).toEqual([messageItem]);
-    expect(combined?.additional_kwargs).not.toHaveProperty(
-      OPENAI_RESPONSES_REPLAY_POSITIONS_KEY
-    );
-    expect(combined?.lc_kwargs.additional_kwargs).not.toHaveProperty(
-      OPENAI_RESPONSES_REPLAY_POSITIONS_KEY
-    );
-  });
-
-  it('keeps a raw server result between distinct Responses output messages', async () => {
-    const model = new ChatOpenAI({
-      model: 'gpt-5',
-      apiKey: 'test-key',
-      useResponsesApi: true,
-    });
-    const responses = (
-      model as unknown as { responses: StreamingOpenAIResponsesDelegate }
-    ).responses;
-    responses.completionWithRetry = async () =>
-      (async function* () {
-        yield {
-          type: 'response.output_text.delta',
-          sequence_number: 0,
-          output_index: 0,
-          content_index: 0,
-          item_id: 'msg_first',
-          delta: 'First narration.',
-          logprobs: [],
-        } as OpenAIClient.Responses.ResponseStreamEvent;
-        yield {
-          type: 'response.output_item.done',
-          sequence_number: 1,
-          output_index: 1,
-          item: {
-            id: 'local_output_item',
-            type: 'local_shell_call_output',
-            status: 'completed',
-            output: 'middle server result',
-          },
-        } as OpenAIClient.Responses.ResponseStreamEvent;
-        yield {
-          type: 'response.output_text.delta',
-          sequence_number: 2,
-          output_index: 2,
-          content_index: 0,
-          item_id: 'msg_second',
-          delta: 'Second narration.',
-          logprobs: [],
-        } as OpenAIClient.Responses.ResponseStreamEvent;
-      })();
-
-    let combined: AIMessageChunk | undefined;
-    const stream = await model.stream([new HumanMessage('run a server tool')]);
-    for await (const chunk of stream) {
-      combined =
-        combined == null
-          ? (chunk as AIMessageChunk)
-          : (combined.concat(chunk as AIMessageChunk) as AIMessageChunk);
-    }
-    expect(combined).toBeDefined();
-    combined!.response_metadata.preempted = true;
-
-    const [projected] = projectOpenAIResponsesToolMessageContent([combined!]);
-    const providerInput = convertMessagesToResponsesInput({
-      messages: [projected],
-      model: 'gpt-5',
-      zdrEnabled: false,
-    });
-    const serialized = JSON.stringify(providerInput);
-    const firstIndex = serialized.indexOf('First narration.');
-    const resultIndex = serialized.indexOf('middle server result');
-    const secondIndex = serialized.indexOf('Second narration.');
-
-    expect(firstIndex).toBeGreaterThanOrEqual(0);
-    expect(firstIndex).toBeLessThan(resultIndex);
-    expect(resultIndex).toBeLessThan(secondIndex);
-  });
-
-  it('keeps streamed generated images and server results in cross-item order', async () => {
-    const model = new ChatOpenAI({
-      model: 'gpt-5',
-      apiKey: 'test-key',
-      useResponsesApi: true,
-    });
-    const responses = (
-      model as unknown as { responses: StreamingOpenAIResponsesDelegate }
-    ).responses;
-    responses.completionWithRetry = async () =>
-      (async function* () {
-        yield {
-          type: 'response.output_text.delta',
-          sequence_number: 0,
-          output_index: 0,
-          content_index: 0,
-          item_id: 'msg_first',
-          delta: 'First narration.',
-          logprobs: [],
-        } as OpenAIClient.Responses.ResponseStreamEvent;
-        yield {
-          type: 'response.output_item.done',
-          sequence_number: 1,
-          output_index: 1,
-          item: {
-            id: 'ig_middle',
-            type: 'image_generation_call',
-            status: 'completed',
-            result: 'AA==',
-          },
-        } as OpenAIClient.Responses.ResponseStreamEvent;
-        yield {
-          type: 'response.output_item.done',
-          sequence_number: 2,
-          output_index: 2,
-          item: {
-            id: 'local_middle',
-            type: 'local_shell_call_output',
-            status: 'completed',
-            output: 'middle server result',
-          },
-        } as OpenAIClient.Responses.ResponseStreamEvent;
-        yield {
-          type: 'response.output_item.added',
-          sequence_number: 3,
-          output_index: 3,
-          item: {
-            id: 'rs_middle',
-            type: 'reasoning',
-            status: 'in_progress',
-            summary: [],
-          },
-        } as OpenAIClient.Responses.ResponseStreamEvent;
-        yield {
-          type: 'response.output_item.done',
-          sequence_number: 4,
-          output_index: 3,
-          item: {
-            id: 'rs_middle',
-            type: 'reasoning',
-            status: 'incomplete',
-            summary: [],
-            encrypted_content: 'opaque-middle-reasoning',
-          },
-        } as OpenAIClient.Responses.ResponseStreamEvent;
-        yield {
-          type: 'response.output_text.delta',
-          sequence_number: 5,
-          output_index: 4,
-          content_index: 0,
-          item_id: 'msg_second',
-          delta: 'Second narration.',
-          logprobs: [],
-        } as OpenAIClient.Responses.ResponseStreamEvent;
-      })();
-
-    let combined: AIMessageChunk | undefined;
-    const stream = await model.stream([new HumanMessage('generate and run')]);
-    for await (const chunk of stream) {
-      combined =
-        combined == null
-          ? (chunk as AIMessageChunk)
-          : (combined.concat(chunk as AIMessageChunk) as AIMessageChunk);
-    }
-    expect(combined).toBeDefined();
-    expect(combined!.additional_kwargs.reasoning).toEqual({
-      id: 'rs_middle',
-      type: 'reasoning',
-      status: 'incomplete',
-      summary: [],
-      encrypted_content: 'opaque-middle-reasoning',
-    });
-    expect(
-      combined!.additional_kwargs[OPENAI_RESPONSES_REPLAY_POSITIONS_KEY]
-    ).toEqual([
-      {
-        contentIndex: 0,
-        itemId: 'msg_first',
-        kind: 'text',
-        outputIndex: 0,
-      },
-      { itemId: 'ig_middle', kind: 'output', outputIndex: 1 },
-      { itemId: 'local_middle', kind: 'output', outputIndex: 2 },
-      { itemId: 'rs_middle', kind: 'reasoning', outputIndex: 3 },
-      {
-        contentIndex: 0,
-        itemId: 'msg_second',
-        kind: 'text',
-        outputIndex: 4,
-      },
-    ]);
-    combined!.response_metadata.preempted = true;
-
-    const [projected] = projectOpenAIResponsesToolMessageContent([combined!]);
-    const providerInput = convertMessagesToResponsesInput({
-      messages: [projected],
-      model: 'gpt-5',
-      zdrEnabled: false,
-    });
-    const serialized = JSON.stringify(providerInput);
-    const firstIndex = serialized.indexOf('First narration.');
-    const imageIndex = serialized.indexOf('data:image/png;base64,AA==');
-    const resultIndex = serialized.indexOf('middle server result');
-    const reasoningIndex = serialized.indexOf('opaque-middle-reasoning');
-    const secondIndex = serialized.indexOf('Second narration.');
-
-    expect(firstIndex).toBeGreaterThanOrEqual(0);
-    expect(firstIndex).toBeLessThan(imageIndex);
-    expect(imageIndex).toBeLessThan(resultIndex);
-    expect(resultIndex).toBeLessThan(reasoningIndex);
-    expect(reasoningIndex).toBeLessThan(secondIndex);
-    expect(serialized).not.toContain('ig_middle');
-    expect(serialized).not.toContain('local_middle');
-    expect(projected.additional_kwargs).not.toHaveProperty(
-      OPENAI_RESPONSES_REPLAY_POSITIONS_KEY
-    );
-  });
-
-  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('sanitizes persisted preempted history on direct Azure Responses streams', async () => {
-    const model = new AzureChatOpenAI({
-      ...baseAzureFields,
-    });
-    model.model = 'gpt-5';
-    const responses = (
-      model as unknown as { responses: MockableResponsesDelegate }
-    ).responses;
-    const requests: unknown[] = [];
-    responses.completionWithRetry = async (request) => {
-      requests.push(request);
-      return (async function* () {})();
-    };
-    const message = createPersistedPreemptedAzureMessage();
-    const originalSerialized = JSON.stringify(message.toJSON());
-
-    for await (const _chunk of responses._streamResponseChunks([message], {})) {
-      // The empty mock stream intentionally yields no chunks.
-    }
-    for await (const _event of responses._streamChatModelEvents(
-      [message],
-      {}
-    )) {
-      // The empty mock stream intentionally yields no events.
-    }
-
-    expect(requests).toHaveLength(2);
-    for (const request of requests) {
-      expectNeutralizedAzureResponsesRequest(request);
-    }
-    expect(JSON.stringify(message.toJSON())).toBe(originalSerialized);
-  });
-
-  it('sanitizes persisted preempted history on direct Azure Responses generation', async () => {
-    const model = new AzureChatOpenAI({
-      ...baseAzureFields,
-    });
-    model.model = 'gpt-5';
-    const responses = (
-      model as unknown as { responses: MockableResponsesDelegate }
-    ).responses;
-    let request: unknown;
-    responses.completionWithRetry = async (nextRequest) => {
-      request = nextRequest;
-      return createCompletedAzureResponse();
-    };
-    const message = createPersistedPreemptedAzureMessage();
-    const originalSerialized = JSON.stringify(message.toJSON());
-
-    await responses._generate([message], {});
-
-    expectNeutralizedAzureResponsesRequest(request);
-    expect(JSON.stringify(message.toJSON())).toBe(originalSerialized);
-  });
-
-  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('skips custom OpenAI-compatible SSE events during OpenAI streaming', async () => {
-    await expectCustomSSEEventsSkipped(
-      new ChatOpenAI({
-        model: 'hermes-agent',
-        apiKey: 'test-key',
-        streaming: true,
-      })
-    );
-  });
-
-  it('preserves OpenAI-compatible reasoning deltas during OpenAI streaming', async () => {
-    const model = new ChatOpenAI({
-      model: 'openai/gpt-oss-120b',
-      apiKey: 'test-key',
-      streaming: true,
-    });
-    const completions = (model as unknown as StreamingCompletionBackedModel)
-      .completions;
-    const createChunk = (
-      choice: OpenAICompatibleReasoningStreamChoice
-    ): OpenAIClient.Chat.Completions.ChatCompletionChunk => ({
-      id: 'chatcmpl-openai-compatible-reasoning',
-      object: 'chat.completion.chunk',
-      created: 0,
-      model: 'openai/gpt-oss-120b',
-      choices: [choice],
-    });
-
-    async function* streamChunks(): AsyncGenerator<OpenAIClient.Chat.Completions.ChatCompletionChunk> {
-      yield createChunk({
-        index: 0,
-        delta: {
-          role: 'assistant',
-          content: '',
-        },
-        finish_reason: null,
-      });
-      yield createChunk({
-        index: 0,
-        delta: {
-          reasoning: 'Think ',
-          reasoning_details: [
-            {
-              type: 'reasoning.text',
-              text: 'Think ',
-              format: 'text',
-              index: 0,
-            },
-          ],
-        },
-        finish_reason: null,
-      });
-      yield createChunk({
-        index: 0,
-        delta: {
-          reasoning: 'hard',
-          reasoning_details: [
-            {
-              type: 'reasoning.text',
-              text: 'hard',
-              format: 'text',
-              index: 0,
-            },
-          ],
-        },
-        finish_reason: null,
-      });
-      yield createChunk({
-        index: 0,
-        delta: { content: 'answer' },
-        finish_reason: 'stop',
-      });
-    }
-
-    completions.completionWithRetry = async (): Promise<
-      AsyncIterable<OpenAIClient.Chat.Completions.ChatCompletionChunk>
-    > => streamChunks();
-
-    const chunks: AIMessageChunk[] = [];
-    const stream = await model.stream([new HumanMessage('think')]);
-    for await (const chunk of stream) {
-      chunks.push(chunk);
-    }
-
-    expect(
-      chunks
-        .map((chunk) => chunk.additional_kwargs.reasoning_content)
-        .filter((reasoningContent) => reasoningContent != null)
-    ).toEqual(['Think ', 'hard']);
-    expect(chunks[1].additional_kwargs.reasoning_details).toEqual([
-      {
-        type: 'reasoning.text',
-        text: 'Think ',
-        format: 'text',
-        index: 0,
-      },
-    ]);
-    expect(chunks[2].additional_kwargs.reasoning_details).toEqual([
-      {
-        type: 'reasoning.text',
-        text: 'hard',
-        format: 'text',
-        index: 0,
-      },
-    ]);
-    expect(chunks.at(-1)?.content).toBe('answer');
-  });
-
-  it('skips custom OpenAI-compatible SSE events during Azure streaming', async () => {
-    await expectCustomSSEEventsSkipped(
-      new AzureChatOpenAI({
-        ...baseAzureFields,
-      })
-    );
-  });
-
-  it('passes non-streaming OpenAI completions through unchanged', async () => {
-    const model = new ChatOpenAI({
-      model: 'hermes-agent',
-      apiKey: 'test-key',
-    });
-    const createMock = mockCompletion(model, {
-      id: 'chatcmpl-nonstream-test',
-      object: 'chat.completion',
-      created: 0,
-      model: 'hermes-agent',
-      choices: [
-        {
-          index: 0,
-          finish_reason: 'stop',
-          logprobs: null,
-          message: {
-            role: 'assistant',
-            content: 'plain response',
-            refusal: null,
-          },
-        },
-      ],
-    });
-
-    const response = await model.invoke([new HumanMessage('no stream')]);
-
-    expect(response.content).toBe('plain response');
-    expect(createMock).toHaveBeenCalledWith(
-      expect.objectContaining({ stream: false }),
-      expect.any(Object)
-    );
-  });
-
-  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('maps OpenRouter cache write usage to cache_creation in streaming responses', async () => {
-    const model = new ChatOpenRouter({
-      model: 'anthropic/claude-sonnet-test',
-      apiKey: 'test-key',
-      streamUsage: true,
-    });
-    const completions = (model as unknown as StreamingCompletionBackedModel)
-      .completions;
-    const usage: CompletionUsageWithCacheWrite = {
-      prompt_tokens: 11,
-      completion_tokens: 7,
-      total_tokens: 18,
-      prompt_tokens_details: {
-        audio_tokens: 2,
-        cached_tokens: 3,
-        cache_write_tokens: 5,
-      },
-      completion_tokens_details: {
-        audio_tokens: 4,
-        reasoning_tokens: 6,
-      },
-    };
-
-    async function* streamChunks(): AsyncGenerator<OpenAIClient.Chat.Completions.ChatCompletionChunk> {
-      yield createOpenAIStreamChunk('answer', 'stop');
-      yield {
-        id: 'chatcmpl-openrouter-usage',
-        object: 'chat.completion.chunk',
-        created: 0,
-        model: 'anthropic/claude-sonnet-test',
-        choices: [],
-        usage,
-      } as OpenAIClient.Chat.Completions.ChatCompletionChunk;
-    }
-
-    completions.completionWithRetry = async (): Promise<
-      AsyncIterable<OpenAIClient.Chat.Completions.ChatCompletionChunk>
-    > => streamChunks();
-
-    const chunks: AIMessageChunk[] = [];
-    const stream = await model.stream([new HumanMessage('hi')]);
-    for await (const chunk of stream) {
-      chunks.push(chunk);
-    }
-
-    const usageChunk = chunks.find(
-      (chunk) => chunk.usage_metadata?.input_token_details?.cache_creation === 5
-    );
-    expect(usageChunk?.usage_metadata).toEqual({
-      input_tokens: 11,
-      output_tokens: 7,
-      total_tokens: 18,
-      input_token_details: {
-        audio: 2,
-        cache_read: 3,
-        cache_creation: 5,
-      },
-      output_token_details: {
-        audio: 4,
-        reasoning: 6,
-      },
-    });
-  });
-
-  it('emits OpenRouter callbacks before an early stream break', async () => {
-    const model = new CallbackTestChatOpenRouter({
-      model: 'openai/gpt-4o-mini',
-      apiKey: 'test-key',
-      _lc_stream_delay: 1,
-    } as OpenRouterFields & { _lc_stream_delay: number });
-    const completions = (model as unknown as StreamingCompletionBackedModel)
-      .completions;
-    const textChunks: string[] = [];
-    const callbackTokens: string[] = [];
-
-    async function* streamChunks(): AsyncGenerator<OpenAIClient.Chat.Completions.ChatCompletionChunk> {
-      yield createOpenAIStreamChunk('alpha beta gamma');
-    }
-
-    completions.completionWithRetry = async (): Promise<
-      AsyncIterable<OpenAIClient.Chat.Completions.ChatCompletionChunk>
-    > => streamChunks();
-
-    const runManager = {
-      handleLLMNewToken(token: string): void {
-        if (token !== '') {
-          callbackTokens.push(token);
-        }
-      },
-    } as unknown as CallbackManagerForLLMRun;
-
-    for await (const chunk of model.streamChunksWithCallbacks(runManager)) {
-      if (chunk.text !== '') {
-        textChunks.push(chunk.text);
-      }
-      break;
-    }
-
-    expect(textChunks).toEqual(['alpha beta gamma']);
-    expect(callbackTokens).toEqual(textChunks);
-  });
-
-  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 defaultModel = new CustomAnthropic({
-      model: 'claude-sonnet-4-5-20250929',
-      apiKey: 'test-key',
-    });
-
-    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(defaultModel._lc_stream_delay).toBe(25);
-    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',
-      _lc_stream_delay: 12,
-    });
-    const defaultStreamDelayModel = new CustomChatBedrockConverse({
-      ...baseBedrockFields,
-      model: 'anthropic.claude-3-haiku-20240307-v1:0',
-    });
-
-    expect(CustomChatBedrockConverse.lc_name()).toBe(
-      'LibreChatBedrockConverse'
-    );
-    expect(model.applicationInferenceProfile).toBe(applicationInferenceProfile);
-    expect(model._lc_stream_delay).toBe(12);
-    expect(defaultStreamDelayModel._lc_stream_delay).toBe(25);
-    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();
-  });
-
-  it('aborts custom OpenAI fetches when the request timeout elapses', async () => {
-    const capturedFetch = createAbortableFetch();
-    const client = new CustomOpenAIClient({
-      apiKey: 'test-key',
-      fetch: capturedFetch.fetch,
-    });
-
-    await expectFetchTimeoutAbort(
-      client,
-      capturedFetch,
-      'https://example.test/v1/chat/completions'
-    );
-  });
-
-  it('aborts custom Azure OpenAI fetches when the request timeout elapses', async () => {
-    const capturedFetch = createAbortableFetch();
-    const client = new CustomAzureOpenAIClient({
-      apiKey: 'test-azure-key',
-      apiVersion: '2024-10-21',
-      baseURL: 'https://example.test/openai/deployments/test-deployment',
-      fetch: capturedFetch.fetch,
-    });
-
-    await expectFetchTimeoutAbort(
-      client,
-      capturedFetch,
-      'https://example.test/openai/deployments/test-deployment/chat/completions'
-    );
-  });
-
-  it('propagates caller abort signals to custom OpenAI fetches', async () => {
-    const capturedFetch = createAbortableFetch();
-    const client = new CustomOpenAIClient({
-      apiKey: 'test-key',
-      fetch: capturedFetch.fetch,
-    });
-    const callerController = new AbortController();
-    const requestController = new AbortController();
-
-    const response = client.fetchWithTimeout(
-      'https://example.test/v1/chat/completions',
-      { method: 'post', signal: callerController.signal },
-      1000,
-      requestController
-    );
-
-    callerController.abort();
-
-    await expect(waitForFetchOutcome(response)).resolves.toBe('rejected');
-    expect(requestController.signal.aborted).toBe(true);
-    expect(capturedFetch.getSignal()?.aborted).toBe(true);
-  });
-});
-
-describe('StreamSmoothingOptions type surface', () => {
-  it('accepts _lc_stream_delay: 0 across the typed provider options', () => {
-    const azure: import('@/types').AzureClientOptions = { _lc_stream_delay: 0 };
-    const openrouter: ChatOpenRouterCallOptions &
-      import('@/types').StreamSmoothingOptions = { _lc_stream_delay: 0 };
-    const mistral: import('@/types').MistralAIClientOptions = {
-      model: 'mistral-large-latest',
-      _lc_stream_delay: 0,
-    };
-    expect(azure._lc_stream_delay).toBe(0);
-    expect(openrouter._lc_stream_delay).toBe(0);
-    expect(mistral._lc_stream_delay).toBe(0);
-  });
-});