@librechat/agents 3.0.63 → 3.0.65

package/src/scripts/test-handoff-preamble.ts

@@ -254,6 +254,8 @@ IMPORTANT: You are the specialist - provide a complete, helpful response to the
   console.log('   - Specialist received and acted on instructions');
   console.log('   - No premature end tokens after handoff');
   console.log('   - Multi-turn conversation continued smoothly');
+
+  console.dir(contentParts, { depth: null });
 }
 
 process.on('unhandledRejection', (reason, promise) => {

package/src/scripts/test-parallel-handoffs.ts

@@ -0,0 +1,291 @@
+import { config } from 'dotenv';
+config();
+
+import { HumanMessage, BaseMessage } from '@langchain/core/messages';
+import type * as t from '@/types';
+import { ChatModelStreamHandler, createContentAggregator } from '@/stream';
+import { ToolEndHandler } from '@/events';
+import { Providers, GraphEvents } from '@/common';
+import { sleep } from '@/utils/run';
+import { Run } from '@/run';
+
+const conversationHistory: BaseMessage[] = [];
+
+/**
+ * Test parallel handoffs - where an LLM calls multiple transfer tools simultaneously
+ *
+ * Graph structure:
+ * coordinator -> [researcher, writer] (via parallel handoff tools)
+ *
+ * The coordinator agent has two transfer tools:
+ * - transfer_to_researcher
+ * - transfer_to_writer
+ *
+ * When given a task that needs both, it should call both tools in parallel.
+ */
+async function testParallelHandoffs() {
+  console.log(
+    'Testing Parallel Handoffs (LLM calling multiple transfers)...\n'
+  );
+
+  const { contentParts, aggregateContent } = createContentAggregator();
+
+  const agents: t.AgentInputs[] = [
+    {
+      agentId: 'coordinator',
+      provider: Providers.OPENAI,
+      clientOptions: {
+        modelName: 'gpt-4o-mini',
+        apiKey: process.env.OPENAI_API_KEY,
+      },
+      instructions: `You are a COORDINATOR agent. Your job is to delegate tasks to specialized agents.
+
+You have access to two transfer tools:
+- transfer_to_researcher: For research and fact-finding tasks
+- transfer_to_writer: For content creation and writing tasks
+
+IMPORTANT: When a task requires BOTH research AND writing, you MUST call BOTH transfer tools SIMULTANEOUSLY in the same response. Do not call them sequentially.
+
+For example, if asked to "research and write about X", call both transfers at once to enable parallel work.
+
+When delegating, provide clear instructions to each agent about what they should do.`,
+    },
+    {
+      agentId: 'researcher',
+      provider: Providers.ANTHROPIC,
+      clientOptions: {
+        modelName: 'claude-haiku-4-5',
+        apiKey: process.env.ANTHROPIC_API_KEY,
+      },
+      instructions: `You are a RESEARCHER. When you receive a task:
+1. Provide concise research findings (100-150 words)
+2. Start your response with "📚 RESEARCH FINDINGS:"`,
+    },
+    {
+      agentId: 'writer',
+      provider: Providers.ANTHROPIC,
+      clientOptions: {
+        modelName: 'claude-haiku-4-5',
+        apiKey: process.env.ANTHROPIC_API_KEY,
+      },
+      instructions: `You are a WRITER. When you receive a task:
+1. Provide creative content (100-150 words)
+2. Start your response with "✍️ WRITTEN CONTENT:"`,
+    },
+  ];
+
+  /**
+   * Create handoff edges from coordinator to both researcher and writer.
+   * These are separate edges so the LLM sees both transfer tools.
+   */
+  const edges: t.GraphEdge[] = [
+    {
+      from: 'coordinator',
+      to: 'researcher',
+      edgeType: 'handoff',
+      description: 'Transfer to researcher for research and fact-finding tasks',
+      prompt: 'Research task instructions',
+    },
+    {
+      from: 'coordinator',
+      to: 'writer',
+      edgeType: 'handoff',
+      description: 'Transfer to writer for content creation and writing tasks',
+      prompt: 'Writing task instructions',
+    },
+  ];
+
+  /** Track which agents are active and their timing */
+  const activeAgents = new Set<string>();
+  const agentTimings: Record<string, { start?: number; end?: number }> = {};
+  const startTime = Date.now();
+
+  const customHandlers = {
+    [GraphEvents.TOOL_END]: new ToolEndHandler(),
+    [GraphEvents.CHAT_MODEL_END]: {
+      handle: (
+        _event: string,
+        _data: t.StreamEventData,
+        metadata?: Record<string, unknown>
+      ): void => {
+        const nodeName = metadata?.langgraph_node as string;
+        if (nodeName) {
+          const elapsed = Date.now() - startTime;
+          agentTimings[nodeName] = agentTimings[nodeName] || {};
+          agentTimings[nodeName].end = elapsed;
+          activeAgents.delete(nodeName);
+          console.log(`\n⏱️  [${nodeName}] COMPLETED at ${elapsed}ms`);
+        }
+      },
+    },
+    [GraphEvents.CHAT_MODEL_START]: {
+      handle: (
+        _event: string,
+        _data: t.StreamEventData,
+        metadata?: Record<string, unknown>
+      ): void => {
+        const nodeName = metadata?.langgraph_node as string;
+        if (nodeName) {
+          const elapsed = Date.now() - startTime;
+          /** Store first start time for parallel overlap calculation */
+          if (!agentTimings[nodeName]?.start) {
+            agentTimings[nodeName] = agentTimings[nodeName] || {};
+            agentTimings[nodeName].start = elapsed;
+          }
+          activeAgents.add(nodeName);
+          console.log(`\n⏱️  [${nodeName}] STARTED at ${elapsed}ms`);
+          console.log(
+            `   Active agents: ${Array.from(activeAgents).join(', ')}`
+          );
+        }
+      },
+    },
+    [GraphEvents.CHAT_MODEL_STREAM]: new ChatModelStreamHandler(),
+    [GraphEvents.ON_RUN_STEP_COMPLETED]: {
+      handle: (
+        event: GraphEvents.ON_RUN_STEP_COMPLETED,
+        data: t.StreamEventData
+      ): void => {
+        aggregateContent({
+          event,
+          data: data as unknown as { result: t.ToolEndEvent },
+        });
+      },
+    },
+    [GraphEvents.ON_RUN_STEP]: {
+      handle: (
+        event: GraphEvents.ON_RUN_STEP,
+        data: t.StreamEventData
+      ): void => {
+        aggregateContent({ event, data: data as t.RunStep });
+      },
+    },
+    [GraphEvents.ON_RUN_STEP_DELTA]: {
+      handle: (
+        event: GraphEvents.ON_RUN_STEP_DELTA,
+        data: t.StreamEventData
+      ): void => {
+        aggregateContent({ event, data: data as t.RunStepDeltaEvent });
+      },
+    },
+    [GraphEvents.ON_MESSAGE_DELTA]: {
+      handle: (
+        event: GraphEvents.ON_MESSAGE_DELTA,
+        data: t.StreamEventData
+      ): void => {
+        aggregateContent({ event, data: data as t.MessageDeltaEvent });
+      },
+    },
+  };
+
+  const runConfig: t.RunConfig = {
+    runId: `parallel-handoffs-${Date.now()}`,
+    graphConfig: {
+      type: 'multi-agent',
+      agents,
+      edges,
+    },
+    customHandlers,
+    returnContent: true,
+  };
+
+  try {
+    const run = await Run.create(runConfig);
+
+    /** Prompt designed to trigger parallel handoffs without confusing language */
+    const userMessage = `Help me with two topics:
+1. The history of the internet
+2. A short poem about technology
+
+I need information on both topics.`;
+
+    conversationHistory.push(new HumanMessage(userMessage));
+
+    console.log('User message:', userMessage);
+    console.log(
+      '\nInvoking multi-agent graph with parallel handoff request...\n'
+    );
+
+    const config = {
+      configurable: {
+        thread_id: 'parallel-handoffs-test-1',
+      },
+      streamMode: 'values',
+      version: 'v2' as const,
+    };
+
+    const inputs = {
+      messages: conversationHistory,
+    };
+
+    await run.processStream(inputs, config);
+    const finalMessages = run.getRunMessages();
+
+    if (finalMessages) {
+      conversationHistory.push(...finalMessages);
+    }
+
+    /** Analyze parallel execution */
+    console.log('\n\n========== TIMING SUMMARY ==========');
+    console.log('Available timing keys:', Object.keys(agentTimings));
+    for (const [agent, timing] of Object.entries(agentTimings)) {
+      const duration =
+        timing.end && timing.start ? timing.end - timing.start : 'N/A';
+      console.log(
+        `${agent}: started=${timing.start}ms, ended=${timing.end}ms, duration=${duration}ms`
+      );
+    }
+
+    /** Check if researcher and writer ran in parallel (handle key variations) */
+    const researcherKey = Object.keys(agentTimings).find((k) =>
+      k.includes('researcher')
+    );
+    const writerKey = Object.keys(agentTimings).find((k) =>
+      k.includes('writer')
+    );
+    const researcherTiming = researcherKey
+      ? agentTimings[researcherKey]
+      : undefined;
+    const writerTiming = writerKey ? agentTimings[writerKey] : undefined;
+
+    if (researcherTiming && writerTiming) {
+      const bothStarted = researcherTiming.start && writerTiming.start;
+      const bothEnded = researcherTiming.end && writerTiming.end;
+
+      if (bothStarted && bothEnded) {
+        const overlap =
+          Math.min(researcherTiming.end!, writerTiming.end!) -
+          Math.max(researcherTiming.start!, writerTiming.start!);
+
+        if (overlap > 0) {
+          console.log(
+            `\n✅ PARALLEL HANDOFFS SUCCESSFUL: ${overlap}ms overlap between researcher and writer`
+          );
+        } else {
+          console.log(
+            `\n⚠️  SEQUENTIAL EXECUTION: researcher and writer did not overlap`
+          );
+          console.log(
+            `   This may indicate the LLM called transfers sequentially, not in parallel`
+          );
+        }
+      }
+    } else {
+      console.log(
+        '\n⚠️  Not all agents were invoked. Check if handoffs occurred.'
+      );
+      console.log('   researcher timing:', researcherTiming);
+      console.log('   writer timing:', writerTiming);
+    }
+    console.log('====================================\n');
+
+    console.log('Final content parts:', contentParts.length, 'parts');
+    console.dir(contentParts, { depth: null });
+    await sleep(3000);
+  } catch (error) {
+    console.error('Error in parallel handoffs test:', error);
+    throw error;
+  }
+}
+
+testParallelHandoffs();

package/src/tools/ToolNode.ts

@@ -281,6 +281,13 @@ export class ToolNode<T = any> extends RunnableCallable<T, T> {
     )[] = [];
     let parentCommand: Command | null = null;
 
+    /**
+     * Collect handoff commands (Commands with string goto and Command.PARENT)
+     * for potential parallel handoff aggregation
+     */
+    const handoffCommands: Command[] = [];
+    const nonCommandOutputs: BaseMessage[] = [];
+
     for (const output of outputs) {
       if (isCommand(output)) {
         if (
@@ -297,24 +304,66 @@ export class ToolNode<T = any> extends RunnableCallable<T, T> {
               goto: output.goto,
             });
           }
-        } else {
+        } else if (output.graph === Command.PARENT) {
           /**
-           * Non-Send Commands (including handoff Commands with string goto)
-           * are passed through as-is. For single handoffs, this works correctly.
-           *
-           * Note: Parallel handoffs (LLM calling multiple transfer tools simultaneously)
-           * are not yet fully supported. For parallel agent execution, use direct edges
-           * with edgeType: 'direct' instead of handoff edges.
+           * Handoff Command with destination.
+           * Handle both string ('agent') and array (['agent']) formats.
+           * Collect for potential parallel aggregation.
            */
+          const goto = output.goto;
+          const isSingleStringDest = typeof goto === 'string';
+          const isSingleArrayDest =
+            Array.isArray(goto) &&
+            goto.length === 1 &&
+            typeof goto[0] === 'string';
+
+          if (isSingleStringDest || isSingleArrayDest) {
+            handoffCommands.push(output);
+          } else {
+            /** Multi-destination or other command - pass through */
+            combinedOutputs.push(output);
+          }
+        } else {
+          /** Other commands - pass through */
           combinedOutputs.push(output);
         }
       } else {
+        nonCommandOutputs.push(output);
         combinedOutputs.push(
           Array.isArray(input) ? [output] : { messages: [output] }
         );
       }
     }
 
+    /**
+     * Handle handoff commands - convert to Send objects for parallel execution
+     * when multiple handoffs are requested
+     */
+    if (handoffCommands.length > 1) {
+      /**
+       * Multiple parallel handoffs - convert to Send objects.
+       * Each Send carries its own state with the appropriate messages.
+       * This enables LLM-initiated parallel execution when calling multiple
+       * transfer tools simultaneously.
+       */
+      const sends = handoffCommands.map((cmd) => {
+        /** Extract destination - handle both string and array formats */
+        const goto = cmd.goto;
+        const destination =
+          typeof goto === 'string' ? goto : (goto as string[])[0];
+        return new Send(destination, cmd.update);
+      });
+
+      const parallelCommand = new Command({
+        graph: Command.PARENT,
+        goto: sends,
+      });
+      combinedOutputs.push(parallelCommand);
+    } else if (handoffCommands.length === 1) {
+      /** Single handoff - pass through as-is */
+      combinedOutputs.push(handoffCommands[0]);
+    }
+
     if (parentCommand) {
       combinedOutputs.push(parentCommand);
     }

package/src/types/graph.ts

@@ -357,6 +357,8 @@ export type MultiAgentGraphInput = StandardGraphInput & {
 
 export interface AgentInputs {
   agentId: string;
+  /** Human-readable name for the agent (used in handoff context). Defaults to agentId if not provided. */
+  name?: string;
   toolEnd?: boolean;
   toolMap?: ToolMap;
   tools?: GraphTools;