illuma-agents 1.0.37 → 1.0.39

package/src/scripts/code_exec_session.ts

@@ -0,0 +1,282 @@
+// src/scripts/code_exec_session.ts
+/**
+ * Test script for automatic session tracking in code execution tools.
+ *
+ * This tests the automatic session_id injection feature where:
+ * 1. First code execution generates files and returns a session_id
+ * 2. Session context is stored in Graph.sessions
+ * 3. Subsequent code executions automatically have access to previous files
+ *    without the LLM needing to explicitly pass session_id
+ *
+ * Run with: npm run code_exec_session
+ */
+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,
+  createMetadataAggregator,
+} from '@/events';
+import { getLLMConfig } from '@/utils/llmConfig';
+import { getArgs } from '@/scripts/args';
+import { Constants, GraphEvents } from '@/common';
+import { Run } from '@/run';
+import { createCodeExecutionTool } from '@/tools/CodeExecutor';
+
+const conversationHistory: BaseMessage[] = [];
+
+/**
+ * Prints a formatted section header for test output
+ */
+function printSection(title: string): void {
+  console.log('\n' + '='.repeat(60));
+  console.log(` ${title}`);
+  console.log('='.repeat(60) + '\n');
+}
+
+/**
+ * Prints session context from the graph for debugging
+ */
+function printSessionContext(run: Run<t.IState>): void {
+  const graph = run.Graph;
+  if (!graph) {
+    console.log('[Session] No graph available');
+    return;
+  }
+
+  const session = graph.sessions.get(Constants.EXECUTE_CODE) as
+    | t.CodeSessionContext
+    | undefined;
+
+  if (!session) {
+    console.log('[Session] No session context stored yet');
+    return;
+  }
+
+  console.log('[Session] Current session context:');
+  console.log(`  - session_id: ${session.session_id}`);
+  console.log(`  - files: ${JSON.stringify(session.files, null, 2)}`);
+  console.log(
+    `  - lastUpdated: ${new Date(session.lastUpdated).toISOString()}`
+  );
+}
+
+async function testAutomaticSessionTracking(): Promise<void> {
+  const { userName, location, provider, currentDate } = await getArgs();
+  const { contentParts, aggregateContent } = createContentAggregator();
+
+  const customHandlers = {
+    [GraphEvents.TOOL_END]: new ToolEndHandler(),
+    [GraphEvents.CHAT_MODEL_END]: new ModelEndHandler(),
+    [GraphEvents.CHAT_MODEL_STREAM]: new ChatModelStreamHandler(),
+    [GraphEvents.ON_RUN_STEP_COMPLETED]: {
+      handle: (
+        event: GraphEvents.ON_RUN_STEP_COMPLETED,
+        data: t.StreamEventData
+      ): void => {
+        console.log('====== ON_RUN_STEP_COMPLETED ======');
+        console.dir(data, { depth: null });
+        aggregateContent({
+          event,
+          data: data as unknown as { result: t.ToolEndEvent },
+        });
+      },
+    },
+    [GraphEvents.ON_RUN_STEP]: {
+      handle: (
+        event: GraphEvents.ON_RUN_STEP,
+        data: t.StreamEventData
+      ): void => {
+        console.log('====== ON_RUN_STEP ======');
+        console.dir(data, { depth: null });
+        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 });
+      },
+    },
+    [GraphEvents.TOOL_START]: {
+      handle: (
+        _event: string,
+        data: t.StreamEventData,
+        _metadata?: Record<string, unknown>
+      ): void => {
+        console.log('====== TOOL_START ======');
+        console.dir(data, { depth: null });
+      },
+    },
+  };
+
+  const llmConfig = getLLMConfig(provider);
+
+  const run = await Run.create<t.IState>({
+    runId: 'session-tracking-test-1',
+    graphConfig: {
+      type: 'standard',
+      llmConfig,
+      tools: [createCodeExecutionTool()],
+      instructions: `You are an AI assistant testing automatic file persistence.
+When writing Python code:
+- Use print() for all outputs
+- Files from previous executions are automatically available in /mnt/data/
+- Files are READ-ONLY; write modifications to NEW filenames
+- IMPORTANT: Do NOT include session_id in your tool calls - it's handled automatically.`,
+      additional_instructions: `User: ${userName}, Location: ${location}, Date: ${currentDate}.`,
+    },
+    returnContent: true,
+    customHandlers,
+  });
+
+  const streamConfig: Partial<RunnableConfig> & {
+    version: 'v1' | 'v2';
+    run_id?: string;
+    streamMode: string;
+  } = {
+    configurable: {
+      provider,
+      thread_id: 'session-tracking-test',
+    },
+    streamMode: 'values',
+    version: 'v2' as const,
+  };
+
+  // =========================================================================
+  // Test 1: Create initial file (establishes session)
+  // =========================================================================
+  printSection('Test 1: Create Initial File');
+  console.log(
+    'This test creates a file, which should establish a session context.\n'
+  );
+
+  const userMessage1 = `
+Create a Python file that writes a simple JSON config file named "app_config.json" with the following content:
+{
+  "app_name": "TestApp",
+  "version": "1.0.0",
+  "debug": true
+}
+
+After writing, print the contents to confirm it was created correctly.
+`;
+
+  conversationHistory.push(new HumanMessage(userMessage1));
+  await run.processStream({ messages: conversationHistory }, streamConfig);
+
+  const finalMessages1 = run.getRunMessages();
+  if (finalMessages1) {
+    conversationHistory.push(...finalMessages1);
+  }
+
+  printSection('Session Context After Test 1');
+  printSessionContext(run);
+
+  // =========================================================================
+  // Test 2: Access previously created file (uses automatic session injection)
+  // =========================================================================
+  printSection('Test 2: Access Previous File (Automatic Session)');
+  console.log('This test reads the file created in Test 1.');
+  console.log(
+    'The LLM does NOT need to provide session_id - it should be injected automatically.\n'
+  );
+
+  const userMessage2 = `
+Now read the app_config.json file that was just created and:
+1. Print its contents
+2. Confirm the version is "1.0.0"
+
+Note: You should be able to access this file from the previous execution automatically.
+`;
+
+  conversationHistory.push(new HumanMessage(userMessage2));
+  await run.processStream({ messages: conversationHistory }, streamConfig);
+
+  const finalMessages2 = run.getRunMessages();
+  if (finalMessages2) {
+    conversationHistory.push(...finalMessages2);
+  }
+
+  printSection('Session Context After Test 2');
+  printSessionContext(run);
+
+  // =========================================================================
+  // Test 3: Modify file (write to new filename)
+  // =========================================================================
+  printSection('Test 3: Modify File (Write to New Filename)');
+  console.log(
+    'This test modifies the config by reading the old file and writing a new one.\n'
+  );
+
+  const userMessage3 = `
+Read app_config.json, update the version to "2.0.0" and debug to false, 
+then save it as "app_config_v2.json". Print both the old and new contents.
+`;
+
+  conversationHistory.push(new HumanMessage(userMessage3));
+  await run.processStream({ messages: conversationHistory }, streamConfig);
+
+  const finalMessages3 = run.getRunMessages();
+  if (finalMessages3) {
+    conversationHistory.push(...finalMessages3);
+  }
+
+  printSection('Session Context After Test 3');
+  printSessionContext(run);
+
+  // =========================================================================
+  // Summary
+  // =========================================================================
+  printSection('Test Summary');
+  console.log('The automatic session tracking feature should have:');
+  console.log('1. Stored the session_id after the first code execution');
+  console.log('2. Automatically injected it into subsequent executions');
+  console.log('3. Accumulated file references across all executions');
+  console.log('\nCheck the session context output above to verify.\n');
+
+  // Generate title
+  const { handleLLMEnd, collected } = createMetadataAggregator();
+  const titleResult = await run.generateTitle({
+    provider,
+    inputText: 'Testing automatic session tracking for code execution',
+    contentParts,
+    chainOptions: {
+      callbacks: [{ handleLLMEnd }],
+    },
+  });
+  console.log('Generated Title:', titleResult);
+  console.log('Collected metadata:', collected);
+}
+
+process.on('unhandledRejection', (reason, promise) => {
+  console.error('Unhandled Rejection at:', promise, 'reason:', reason);
+  console.log('Conversation history:');
+  console.dir(conversationHistory, { depth: null });
+  process.exit(1);
+});
+
+process.on('uncaughtException', (err) => {
+  console.error('Uncaught Exception:', err);
+});
+
+testAutomaticSessionTracking().catch((err) => {
+  console.error(err);
+  console.log('Conversation history:');
+  console.dir(conversationHistory, { depth: null });
+  process.exit(1);
+});

package/src/scripts/multi-agent-conditional.ts

@@ -210,6 +210,7 @@ async function testConditionalMultiAgent() {
       console.log(`\n\nExpert used: ${selectedExpert}`);
       console.log('Content parts:', contentParts.length);
       console.log('---');
+      console.dir(contentParts, { depth: null });
     }
   } catch (error) {
     console.error('Error in conditional multi-agent test:', error);

package/src/scripts/multi-agent-supervisor.ts

@@ -353,6 +353,7 @@ async function testSupervisorMultiAgent() {
       );
       console.log(`                 transfer_to_legal_advisor`);
       console.log('─'.repeat(60));
+      console.dir(contentParts, { depth: null });
     }
     await sleep(3000);
   } catch (error) {

package/src/scripts/programmatic_exec_agent.ts

@@ -23,7 +23,7 @@ import type { RunnableConfig } from '@langchain/core/runnables';
 import type * as t from '@/types';
 import { createCodeExecutionTool } from '@/tools/CodeExecutor';
 import { createProgrammaticToolCallingTool } from '@/tools/ProgrammaticToolCalling';
-import { createToolSearchRegexTool } from '@/tools/ToolSearchRegex';
+import { createToolSearch } from '@/tools/ToolSearch';
 import { getLLMConfig } from '@/utils/llmConfig';
 import { getArgs } from '@/scripts/args';
 import { Run } from '@/run';
@@ -40,7 +40,7 @@ import {
 
 /**
  * Tool registry only needs business logic tools that require filtering.
- * Special tools (execute_code, run_tools_with_code, tool_search_regex)
+ * Special tools (execute_code, run_tools_with_code, tool_search)
  * are always bound directly to the LLM and don't need registry entries.
  */
 function createAgentToolRegistry(): t.LCToolRegistry {
@@ -73,7 +73,7 @@ async function main(): Promise<void> {
   // Create special tools (PTC, code execution, tool search)
   const codeExecTool = createCodeExecutionTool();
   const ptcTool = createProgrammaticToolCallingTool();
-  const toolSearchTool = createToolSearchRegexTool();
+  const toolSearchTool = createToolSearch();
 
   // Build complete tool list and map
   const allTools = [...mockTools, codeExecTool, ptcTool, toolSearchTool];
@@ -199,7 +199,7 @@ Use the run_tools_with_code tool to do this efficiently - don't call each tool s
   console.log('='.repeat(70));
   console.log('\nKey observations:');
   console.log(
-    '1. LLM only sees tools with allowed_callers including "direct" (get_weather, execute_code, run_tools_with_code, tool_search_regex)'
+    '1. LLM only sees tools with allowed_callers including "direct" (get_weather, execute_code, run_tools_with_code, tool_search)'
   );
   console.log(
     '2. When PTC is invoked, ToolNode automatically injects programmatic tools (get_team_members, get_expenses, get_weather)'

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);
+});