illuma-agents 1.0.37 → 1.0.38

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

@@ -0,0 +1,277 @@
+/**
+ * Test script for multi-turn handoff behavior.
+ *
+ * This tests the fix for the issue where receiving agents would see transfer messages
+ * and prematurely produce end tokens, thinking the work was already done.
+ *
+ * The fix:
+ * 1. Filters out transfer tool calls and ToolMessages from the receiving agent's context
+ * 2. Injects any passthrough instructions as a HumanMessage to ground the receiving agent
+ */
+import { config } from 'dotenv';
+config();
+
+import { HumanMessage, BaseMessage } from '@langchain/core/messages';
+import type { RunnableConfig } from '@langchain/core/runnables';
+import type * as t from '@/types';
+import { ChatModelStreamHandler, createContentAggregator } from '@/stream';
+import { ToolEndHandler, ModelEndHandler } from '@/events';
+import { getLLMConfig } from '@/utils/llmConfig';
+import { GraphEvents, Providers } from '@/common';
+import { Run } from '@/run';
+
+const conversationHistory: BaseMessage[] = [];
+
+/**
+ * Test multi-turn handoff between a coordinator and a specialist
+ */
+async function testHandoffPreamble(): Promise<void> {
+  console.log('='.repeat(60));
+  console.log('Testing Multi-Turn Handoff with Preamble Injection');
+  console.log('='.repeat(60));
+  console.log('\nThis test verifies that:');
+  console.log('1. Transfer messages are filtered from receiving agent context');
+  console.log('2. Passthrough instructions are injected as a HumanMessage');
+  console.log('3. Multi-turn conversations work correctly after handoffs\n');
+
+  const { contentParts, aggregateContent } = createContentAggregator();
+
+  /** Track which agent is responding */
+  let currentAgent = '';
+
+  const customHandlers = {
+    [GraphEvents.TOOL_END]: new ToolEndHandler(),
+    [GraphEvents.CHAT_MODEL_END]: new ModelEndHandler(),
+    [GraphEvents.CHAT_MODEL_STREAM]: new ChatModelStreamHandler(),
+    [GraphEvents.ON_RUN_STEP]: {
+      handle: (
+        event: GraphEvents.ON_RUN_STEP,
+        data: t.StreamEventData
+      ): void => {
+        const runStep = data as t.RunStep;
+        if (runStep.agentId) {
+          currentAgent = runStep.agentId;
+          console.log(`\n[Agent: ${currentAgent}] Processing...`);
+        }
+        aggregateContent({ event, data: runStep });
+      },
+    },
+    [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_MESSAGE_DELTA]: {
+      handle: (
+        event: GraphEvents.ON_MESSAGE_DELTA,
+        data: t.StreamEventData
+      ): void => {
+        console.log('====== ON_MESSAGE_DELTA ======');
+        console.dir(data, { depth: null });
+        aggregateContent({ event, data: data as t.MessageDeltaEvent });
+      },
+    },
+    [GraphEvents.TOOL_START]: {
+      handle: (
+        _event: string,
+        data: t.StreamEventData,
+        _metadata?: Record<string, unknown>
+      ): void => {
+        const toolData = data as { name?: string };
+        if (toolData?.name?.includes('transfer_to_')) {
+          const specialist = toolData.name.replace('lc_transfer_to_', '');
+          console.log(`\n🔀 Handing off to: ${specialist}`);
+        }
+      },
+    },
+  };
+
+  /**
+   * Create agents:
+   * - coordinator: Decides when to hand off to specialist
+   * - specialist: Handles specific tasks delegated by coordinator
+   */
+  const agents: t.AgentInputs[] = [
+    {
+      agentId: 'coordinator',
+      provider: Providers.OPENAI,
+      clientOptions: {
+        modelName: 'gpt-4.1-mini',
+        apiKey: process.env.OPENAI_API_KEY,
+      },
+      instructions: `You are a Task Coordinator. Your role is to:
+1. Understand user requests
+2. If the request involves technical analysis, use the transfer_to_specialist tool to hand off
+3. When handing off, provide clear instructions about what needs to be done
+
+IMPORTANT: When using the handoff tool, include specific instructions for the specialist.`,
+      maxContextTokens: 8000,
+    },
+    {
+      agentId: 'specialist',
+      provider: Providers.OPENAI,
+      clientOptions: {
+        modelName: 'gpt-4.1-mini',
+        apiKey: process.env.OPENAI_API_KEY,
+      },
+      instructions: `You are a Technical Specialist. When you receive a request:
+1. Carefully read any instructions provided
+2. Provide a detailed technical response
+3. Do NOT just acknowledge - provide substantive help
+
+IMPORTANT: You are the specialist - provide a complete, helpful response to the task.`,
+      maxContextTokens: 8000,
+    },
+  ];
+
+  /** Create handoff edge with passthrough instructions */
+  const edges: t.GraphEdge[] = [
+    {
+      from: 'coordinator',
+      to: 'specialist',
+      description: 'Transfer to technical specialist for analysis',
+      edgeType: 'handoff',
+      prompt: 'Specific instructions for the specialist about what to analyze',
+      promptKey: 'instructions',
+    },
+  ];
+
+  const runConfig: t.RunConfig = {
+    runId: `handoff-test-${Date.now()}`,
+    graphConfig: {
+      type: 'multi-agent',
+      agents,
+      edges,
+    },
+    customHandlers,
+    returnContent: true,
+  };
+
+  const run = await Run.create(runConfig);
+
+  const config: Partial<RunnableConfig> & {
+    version: 'v1' | 'v2';
+    streamMode: string;
+  } = {
+    configurable: {
+      thread_id: 'handoff-test-conversation-1',
+    },
+    streamMode: 'values',
+    version: 'v2' as const,
+  };
+
+  /** TURN 1: Initial request that triggers handoff */
+  console.log('\n' + '─'.repeat(60));
+  console.log('TURN 1: Initial request (should trigger handoff)');
+  console.log('─'.repeat(60));
+
+  const userMessage1 = `
+    Hi! Can you help me understand the time complexity of quicksort?
+    I need a technical explanation.
+  `;
+
+  conversationHistory.push(new HumanMessage(userMessage1));
+  console.log('\nUser:', userMessage1.trim());
+  console.log('\nResponse:');
+
+  let inputs = { messages: conversationHistory };
+  await run.processStream(inputs, config);
+  const messages1 = run.getRunMessages();
+  if (messages1) {
+    conversationHistory.push(...messages1);
+  }
+
+  console.log('\n');
+
+  /** TURN 2: Follow-up question to test multi-turn after handoff */
+  console.log('\n' + '─'.repeat(60));
+  console.log('TURN 2: Follow-up question (tests context after handoff)');
+  console.log('─'.repeat(60));
+
+  const userMessage2 = `
+    Thanks! Can you also explain the space complexity and when quicksort 
+    might not be the best choice?
+  `;
+
+  conversationHistory.push(new HumanMessage(userMessage2));
+  console.log('\nUser:', userMessage2.trim());
+  console.log('\nResponse:');
+
+  inputs = { messages: conversationHistory };
+  await run.processStream(inputs, config);
+  const messages2 = run.getRunMessages();
+  if (messages2) {
+    conversationHistory.push(...messages2);
+  }
+
+  console.log('\n');
+
+  /** TURN 3: Another follow-up to verify sustained conversation */
+  console.log('\n' + '─'.repeat(60));
+  console.log('TURN 3: Third turn (tests sustained multi-turn)');
+  console.log('─'.repeat(60));
+
+  const userMessage3 = `
+    Great explanation! One more question - how does quicksort compare 
+    to mergesort in practice?
+  `;
+
+  conversationHistory.push(new HumanMessage(userMessage3));
+  console.log('\nUser:', userMessage3.trim());
+  console.log('\nResponse:');
+
+  inputs = { messages: conversationHistory };
+  await run.processStream(inputs, config);
+  const messages3 = run.getRunMessages();
+  if (messages3) {
+    conversationHistory.push(...messages3);
+  }
+
+  /** Summary */
+  console.log('\n\n' + '='.repeat(60));
+  console.log('TEST SUMMARY');
+  console.log('='.repeat(60));
+  console.log('\nTotal messages in conversation:', conversationHistory.length);
+  console.log('\nMessage types:');
+
+  for (let i = 0; i < conversationHistory.length; i++) {
+    const msg = conversationHistory[i];
+    const type = msg.getType();
+    const preview =
+      typeof msg.content === 'string'
+        ? msg.content.slice(0, 50).replace(/\n/g, ' ')
+        : '[complex content]';
+    console.log(`  ${i + 1}. [${type}] ${preview}...`);
+  }
+
+  console.log('\n✅ Test completed. Review the output above to verify:');
+  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) => {
+  console.error('Unhandled Rejection at:', promise, 'reason:', reason);
+  console.log('\nConversation history at failure:');
+  console.dir(conversationHistory, { depth: null });
+  process.exit(1);
+});
+
+process.on('uncaughtException', (err) => {
+  console.error('Uncaught Exception:', err);
+});
+
+testHandoffPreamble().catch((err) => {
+  console.error('Test failed:', err);
+  console.log('\nConversation history at failure:');
+  console.dir(conversationHistory, { depth: null });
+  process.exit(1);
+});

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/scripts/test-tools-before-handoff.ts

@@ -28,10 +28,14 @@ async function testToolsBeforeHandoff() {
 
   // Create custom handlers
   const customHandlers = {
-    [GraphEvents.TOOL_END]: new ToolEndHandler(undefined, (name?: string) => {
-      console.log(`\n✅ Tool completed: ${name}`);
-      return true;
-    }),
+    [GraphEvents.TOOL_END]: new ToolEndHandler(
+      undefined,
+      undefined,
+      (name?: string) => {
+        console.log(`\n✅ Tool completed: ${name}`);
+        return true;
+      }
+    ),
     [GraphEvents.CHAT_MODEL_END]: new ModelEndHandler(),
     [GraphEvents.CHAT_MODEL_STREAM]: new ChatModelStreamHandler(),
     [GraphEvents.ON_RUN_STEP]: {