@librechat/agents 3.3.12 → 3.4.0

package/dist/types/types/llm.d.ts

@@ -28,7 +28,7 @@ export type AzureClientOptions = Partial<OpenAIChatInput> & Partial<AzureOpenAII
     deploymentName?: string;
 } & BaseChatModelParams & {
     configuration?: OAIClientOptions;
-} & ManagedRequestOptions;
+} & ManagedRequestOptions & StreamSmoothingOptions;
 /**
  * Controls whether Claude's reasoning content is returned in adaptive
  * thinking responses. Added for Claude Opus 4.7, which omits thinking by
@@ -58,8 +58,19 @@ export type ManagedRequestOptions = {
     promptCacheExplicit?: boolean;
     safety_identifier?: string;
 };
-export type OpenAIClientOptions = ChatOpenAIFields & ManagedRequestOptions;
-export type AnthropicClientOptions = Omit<AnthropicInput, 'thinking'> & {
+/**
+ * Adaptive stream-smoothing configuration shared by every provider client.
+ */
+export type StreamSmoothingOptions = {
+    /**
+     * Minimum delay in milliseconds between visible streamed content deltas.
+     * Defaults to 25; piece sizes adapt to the backlog so render lag stays
+     * bounded regardless of provider chunk size. Set 0 to disable smoothing.
+     */
+    _lc_stream_delay?: number;
+};
+export type OpenAIClientOptions = ChatOpenAIFields & ManagedRequestOptions & StreamSmoothingOptions;
+export type AnthropicClientOptions = Omit<AnthropicInput, 'thinking'> & StreamSmoothingOptions & {
     thinking?: ThinkingConfig;
     promptCache?: boolean;
     /**
@@ -69,12 +80,12 @@ export type AnthropicClientOptions = Omit<AnthropicInput, 'thinking'> & {
      */
     promptCacheTtl?: PromptCacheTtl;
 };
-export type MistralAIClientOptions = ChatMistralAIInput;
-export type VertexAIClientOptions = ChatVertexAIInput & {
+export type MistralAIClientOptions = ChatMistralAIInput & StreamSmoothingOptions;
+export type VertexAIClientOptions = ChatVertexAIInput & StreamSmoothingOptions & {
     includeThoughts?: boolean;
     thinkingConfig?: GoogleThinkingConfig;
 };
-export type BedrockAnthropicInput = ChatBedrockConverseInput & {
+export type BedrockAnthropicInput = ChatBedrockConverseInput & StreamSmoothingOptions & {
     additionalModelRequestFields?: ChatBedrockConverseInput['additionalModelRequestFields'] & AnthropicReasoning;
     promptCache?: boolean;
     /**
@@ -84,20 +95,16 @@ export type BedrockAnthropicInput = ChatBedrockConverseInput & {
      * server-side, so the default is safe to leave on.
      */
     promptCacheTtl?: PromptCacheTtl;
-    /**
-     * Minimum delay in milliseconds between visible streamed content deltas.
-     */
-    _lc_stream_delay?: number;
 };
 export type BedrockConverseClientOptions = BedrockAnthropicInput;
 export type BedrockAnthropicClientOptions = BedrockAnthropicInput;
-export type GoogleClientOptions = GoogleGenerativeAIChatInput & {
+export type GoogleClientOptions = GoogleGenerativeAIChatInput & StreamSmoothingOptions & {
     customHeaders?: RequestOptions['customHeaders'];
     thinkingConfig?: GoogleThinkingConfig;
     includeServerSideToolInvocations?: boolean;
 };
-export type DeepSeekClientOptions = Partial<ChatDeepSeekInput>;
-export type XAIClientOptions = ChatXAIInput;
+export type DeepSeekClientOptions = Partial<ChatDeepSeekInput> & StreamSmoothingOptions;
+export type XAIClientOptions = ChatXAIInput & StreamSmoothingOptions;
 export type ClientOptions = OpenAIClientOptions | AzureClientOptions | AnthropicClientOptions | MistralAIClientOptions | VertexAIClientOptions | BedrockConverseClientOptions | GoogleClientOptions | DeepSeekClientOptions | XAIClientOptions;
 export type SharedLLMConfig = {
     provider: Providers;
@@ -122,7 +129,7 @@ export type ProviderOptionsMap = {
     [Providers.ANTHROPIC]: AnthropicClientOptions;
     [Providers.MISTRALAI]: MistralAIClientOptions;
     [Providers.MISTRAL]: MistralAIClientOptions;
-    [Providers.OPENROUTER]: ChatOpenRouterCallOptions;
+    [Providers.OPENROUTER]: ChatOpenRouterCallOptions & StreamSmoothingOptions;
     [Providers.BEDROCK]: BedrockAnthropicClientOptions;
     [Providers.XAI]: XAIClientOptions;
     [Providers.MOONSHOT]: OpenAIClientOptions;

package/dist/types/utils/tokens.d.ts

@@ -1,5 +1,15 @@
 import type { BaseMessage } from '@langchain/core/messages';
 export type EncodingName = 'o200k_base' | 'claude';
+export type UnsafeTokenMeasurementReason = 'message_proxy' | 'content_proxy' | 'metadata_proxy' | 'metadata_accessor' | 'invalid_count';
+export declare class UnsafeTokenMeasurementError extends Error {
+    readonly type = "unsafe_token_measurement";
+    readonly reason: UnsafeTokenMeasurementReason;
+    readonly path: string;
+    constructor({ reason, path, }: {
+        reason: UnsafeTokenMeasurementReason;
+        path: string;
+    });
+}
 /** Safety margin for image and document token estimates (5% overestimate). */
 export declare const IMAGE_TOKEN_SAFETY_MARGIN = 1.05;
 /**

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@librechat/agents",
-  "version": "3.3.12",
+  "version": "3.4.0",
   "main": "./dist/cjs/main.cjs",
   "module": "./dist/esm/main.mjs",
   "types": "./dist/types/index.d.ts",
@@ -248,7 +248,7 @@
     "nanoid": "^3.3.7",
     "okapibm25": "^1.4.1",
     "openai": "^6.46.0",
-    "reo-census": "^1.2.9",
+    "reo-census": "^1.2.10",
     "socks-proxy-agent": "^8.0.5",
     "uuid": "^11.1.1"
   },

package/src/graphs/Graph.ts

@@ -998,6 +998,7 @@ export abstract class Graph<
 
 export class StandardGraph extends Graph<t.BaseGraphState, t.GraphNode> {
   overrideModel?: t.ChatModel;
+  private subagentModelOverride?: t.ChatModel;
   /** Optional compile options passed into workflow.compile() */
   compileOptions?: t.CompileOptions | undefined;
   /** Whether the workflow was actually compiled with a checkpointer. */
@@ -1320,6 +1321,7 @@ export class StandardGraph extends Graph<t.BaseGraphState, t.GraphNode> {
     super.clearHeavyState();
     this.messages = [];
     this.overrideModel = undefined;
+    this.subagentModelOverride = undefined;
     /** Stream-limit accounting (argument tallies, event counts, charge
      * credits) deliberately SURVIVES cleanup: this runs in `processStream`'s
      * finally, which an ordinary parallel-branch failure reaches while
@@ -1944,6 +1946,11 @@ export class StandardGraph extends Graph<t.BaseGraphState, t.GraphNode> {
     });
   }
 
+  /** Explicitly overrides the model used by isolated descendant subagent graphs. */
+  setSubagentModelOverride(model: t.ChatModel): void {
+    this.subagentModelOverride = model;
+  }
+
   getUsageMetadata(
     finalMessage?: BaseMessage
   ): Partial<UsageMetadata> | undefined {
@@ -4057,6 +4064,10 @@ export class StandardGraph extends Graph<t.BaseGraphState, t.GraphNode> {
           maxDepth: effectiveSubagentDepth,
           createChildGraph: (input): StandardGraph => {
             const childGraph = new StandardGraph(input);
+            if (this.subagentModelOverride != null) {
+              childGraph.overrideModel = this.subagentModelOverride;
+              childGraph.setSubagentModelOverride(this.subagentModelOverride);
+            }
             const toolHandlerRegistry = createToolHandlerRegistry(
               getParentHandlerRegistry()
             );

package/src/index.ts

@@ -77,6 +77,14 @@ export type {
   ChatOpenRouterCallOptions,
 } from './llm/openrouter';
 export { getChatModelClass } from './llm/providers';
+export { CustomChatMistralAI } from './llm/mistral';
+export {
+  smoothStream,
+  resolveStreamDelay,
+  DEFAULT_STREAM_DELAY,
+  computeAdaptivePieceSize,
+} from './llm/stream/smoother';
+export type { SmoothItem, SmoothPiece } from './llm/stream/smoother';
 export { FakeChatModel, createFakeStreamingLLM } from './llm/fake';
 export { initializeModel } from './llm/init';
 export { attemptInvoke, tryFallbackProviders } from './llm/invoke';

package/src/instrumentation.ts

@@ -84,6 +84,7 @@ function getLangfuseProcessorCacheKey(
 ): string {
   return JSON.stringify({
     destinationKey,
+    mediaUploadEnabled: langfuse?.mediaUploadEnabled,
     toolOutputTracing: langfuse?.toolOutputTracing,
   });
 }

package/src/langfuseSpanRegistry.ts

@@ -72,6 +72,9 @@ export function getLangfuseSpanProcessorParams(
       secretKey: langfuse.secretKey,
       ...(isPresent(langfuse.baseUrl) ? { baseUrl: langfuse.baseUrl } : {}),
       ...(isPresent(environment) ? { environment } : {}),
+      ...(langfuse.mediaUploadEnabled != null
+        ? { mediaUploadEnabled: langfuse.mediaUploadEnabled }
+        : {}),
     };
   }
   if (hasLangfuseEnvConfig()) {
@@ -84,6 +87,9 @@ export function getLangfuseSpanProcessorParams(
       secretKey: process.env.LANGFUSE_SECRET_KEY as string,
       ...(isPresent(baseUrl) ? { baseUrl } : {}),
       ...(isPresent(environment) ? { environment } : {}),
+      ...(langfuse?.mediaUploadEnabled != null
+        ? { mediaUploadEnabled: langfuse.mediaUploadEnabled }
+        : {}),
     };
   }
   if (isPresent(langfuse?.baseUrl) && hasLangfuseEnvCredentials()) {
@@ -92,6 +98,9 @@ export function getLangfuseSpanProcessorParams(
       secretKey: process.env.LANGFUSE_SECRET_KEY as string,
       baseUrl: langfuse.baseUrl,
       ...(isPresent(environment) ? { environment } : {}),
+      ...(langfuse.mediaUploadEnabled != null
+        ? { mediaUploadEnabled: langfuse.mediaUploadEnabled }
+        : {}),
     };
   }
   return undefined;