Sequential Tool Calling and Response

src/cobolt-backend/chat_history.ts

@@ -56,15 +56,44 @@ class ChatHistory {
 
   /**
    * Convert chat history to format used by Ollama LLM
-   * @returns Array of Message objects in Ollama format
+   * Removes execution event metadata from AI context
+   * @returns Array of Message objects in Ollama format with clean content
    */
   toOllamaMessages(): Message[] {
     return this.messages.map(message => ({
       role: message.role,
-      content: message.content
+      content: this.cleanContentForAI(message.content)
     }));
   }
 
+  /**
+   * Clean content for AI context - removes ALL execution metadata
+   * This prevents AI from learning how to fake execution patterns
+   */
+  private cleanContentForAI(content: string): string {
+    let cleaned = content;
+
+    // AGGRESSIVE CLEANING: Remove ALL XML-like tags that start with these patterns
+    cleaned = cleaned.replace(/<execution_event\b[^>]*>.*?<\/execution_event>/gs, '');
+    cleaned = cleaned.replace(/<tool_call_position\b[^>]*>/g, '');
+    cleaned = cleaned.replace(/<tool_calls_update\b[^>]*>.*?<\/tool_calls_update>/gs, '');
+    cleaned = cleaned.replace(/<tool_calls_complete\b[^>]*>.*?<\/tool_calls_complete>/gs, '');
+    cleaned = cleaned.replace(/<tool_calls\b[^>]*>.*?<\/tool_calls>/gs, '');
+
+    // Remove standalone closing tags that might be left behind
+    cleaned = cleaned.replace(/<\/(?:execution_event|tool_calls_update|tool_calls_complete|tool_calls|tool_call_position)>/g, '');
+
+    // Remove any remaining XML-like execution metadata
+    cleaned = cleaned.replace(/<[^>]*(?:execution|tool_call|metadata|event)[^>]*>.*?<\/[^>]+>/gs, '');
+
+    // Clean up any excessive whitespace left behind
+    cleaned = cleaned.replace(/\n\s*\n\s*\n/g, '\n\n');
+    cleaned = cleaned.replace(/^\s+|\s+$/gm, ''); // Trim each line
+    cleaned = cleaned.trim();
+
+    return cleaned;
+  }
+
   /**
    * Clear the chat history
    */

src/cobolt-backend/memory.ts

@@ -64,12 +64,27 @@ function initMemory(): void {
  * Adds messages to the memory store
  * @param messages Array of messages to add to memory
  */
+/* og
 async function addToMemory(messages: Message[]): Promise<void> {
     if (!memoryEnabled) {
       return;
     }
     await memory?.add(messages, {userId: "userid"});
 }
+*/
+async function addToMemory(messages: Message[]): Promise<void> {
+    if (!memoryEnabled) {
+      return;
+    }
+
+    try {
+      const result = await memory?.add(messages, {userId: "userid"});
+      console.log('[Memory] SUCCESS memory.add() result:', result);
+    } catch (error) {
+      console.error('[Memory] Memory storage FAILED:', error);
+      throw error;
+    }
+}
 
 /**
  * Searches the memory store for relevant memories
@@ -128,4 +143,8 @@ if (require.main === module) {
   }
 }
 
-export { addToMemory, searchMemories, clearMemory, listMemories, updateMemoryEnabled };
+function isMemoryEnabled(): boolean {
+  return memoryEnabled;
+}
+
+export { addToMemory, searchMemories, clearMemory, listMemories, updateMemoryEnabled, isMemoryEnabled };

src/cobolt-backend/ollama_client.ts

@@ -1,15 +1,12 @@
-import { Ollama, Message, ChatResponse } from 'ollama';
+import { Ollama, Message, ChatResponse} from 'ollama';
 import { exec, spawn } from 'child_process';
 import log from 'electron-log/main';
-import { FunctionTool } from './ollama_tools';
 import * as os from 'os';
 import configStore from './data_models/config_store';
 import { addToMemory } from './memory';
 import { RequestContext, TraceLogger } from './logger';
-import { formatDateTime } from './datetime_parser';
-import { createQueryWithToolsPrompt } from './prompt_templates';
-import  { ChatHistory } from './chat_history';
-import { MODELS } from './model_manager'
+import { MODELS } from './model_manager';
+import { FunctionTool } from './ollama_tools';
 import { BrowserWindow } from 'electron';
 
 let progressWindow: BrowserWindow | null = null;
@@ -461,9 +458,12 @@ async function* simpleChatOllamaStream(requestContext: RequestContext,
 
 /**
  * Send a simple query to ollama with the specified tools.
- * @param messages - a slice of messages objects
- * @param toolCalls - the list of FunctionTools to pass with the query
- * @returns - The response from the LLM
+ * Uses the dedicated TOOLS_MODEL for efficient tool calling.
+ * @param requestContext - Request context with user query
+ * @param systemPrompt - System prompt for tool selection
+ * @param toolCalls - List of available tools
+ * @param memories - User memories for context
+ * @returns The response from the LLM with tool calls
  */
 async function queryOllamaWithTools(requestContext: RequestContext,
   systemPrompt: string,
@@ -515,27 +515,4 @@ function logExecOutput(platform: string) {
   };
 }
 
-if (require.main === module) {
-  (async () => {
-    await initOllama();
-    const toolCalls: FunctionTool[] = [];
-    const requestContext = {
-      requestId: '123',
-      currentDatetime: new Date(),
-      question: 'Give me all of my calender events since last week from friends',
-      chatHistory: new ChatHistory(),
-    };
-    const toolUserMessage = createQueryWithToolsPrompt(formatDateTime(new Date()).toString())
-    const response = await queryOllamaWithTools(requestContext, toolUserMessage, toolCalls);
-    console.log(response)
-    if (!response.message.tool_calls) {
-      console.log('No tool calls');
-      return;
-    }
-    for (const toolCall of response.message.tool_calls) {
-      console.log('Tool call:', toolCall);
-    }
-  })();
-}
-
 export { initOllama, getOllamaClient, queryOllamaWithTools, simpleChatOllamaStream, stopOllama, setProgressWindow };