feat: Improve conversation condensing by completely replacing history with summary

CHANGELOG.md

@@ -147,6 +147,34 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - Graceful error handling and fallback mechanisms
 - State management for condensed conversations in chat history
 
+## [0.3.4] - 2025-08-17
+
+### Improved
+- **Conversation Condensing**: Complete replacement of conversation history with condensed summary
+  - Set `maxMessagesToKeep` to 0 for complete history replacement
+  - Removed `condense_conversation` tool from available tools list
+  - Completely replace chat history with just the condensed summary
+  - Improved conditional logic in `condenseConversation` utility function
+
+### Added
+- **Comprehensive Test Suite**: Added test script for automatic condensing verification
+  - `test/automatic-condense-test.js`: Full test suite for condensing functionality
+  - Verification of token efficiency and conversation state management
+  - Testing of edge cases and error conditions
+
+### Technical Improvements
+- Enhanced token efficiency by completely removing all previous messages
+- Cleaner state management with simplified conversation history handling
+- Better performance through reduced memory footprint after condensing
+## [0.4.0] - 2025-08-18
+
+### Refactored
+- **LLM Configuration Handling**: Simplified LLM config management by removing complex deriveLLMConfigFromClient method
+- **Condense Function**: Improved condenseConversation to accept LLMClient directly instead of LLMConfig
+- **Code Cleanup**: Removed unused imports (LLMProvider, BatchResult) to clean up dependencies
+
+### Added
+- **Condense Threshold Setting**: Added configurable condense threshold setting to UI with percentage values
 ## [0.1.0] - 2025-01-25
 
 ### Added

package.json

@@ -1,6 +1,6 @@
 {
   "name": "@graphteon/juriko-cli",
-  "version": "0.3.3",
+  "version": "0.4.0",
   "description": "JURIKO - An open-source AI agent that brings the power of AI directly into your terminal.",
   "main": "dist/index.js",
   "bin": {

src/agent/multi-llm-agent.ts

@@ -1,5 +1,5 @@
 import { LLMClient } from "../llm/client";
-import { LLMMessage, LLMToolCall, LLMConfig, LLMProvider } from "../llm/types";
+import { LLMMessage, LLMToolCall, LLMConfig } from "../llm/types";
 import { TextEditorTool, BashTool, TodoTool, ConfirmationTool, CondenseTool } from "../tools";
 import { ToolResult } from "../types";
 import { EventEmitter } from "events";
@@ -15,7 +15,7 @@ import {
 import { parseToolCallArguments, validateArgumentTypes } from "../utils/argument-parser";
 import { SystemPromptBuilder, PromptOptions } from "./prompts/system-prompt-builder";
 import { ResponseFormatter, ResponseStyle } from "../utils/response-formatter";
-import { BatchToolExecutor, BatchResult } from "../tools/batch-executor";
+import { BatchToolExecutor } from "../tools/batch-executor";
 import { getEffectiveSettings } from "../utils/user-settings";
 
 export interface ChatEntry {
@@ -176,23 +176,6 @@ const MULTI_LLM_TOOLS = [
       }
     }
   },
-  {
-    type: "function" as const,
-    function: {
-      name: "condense_conversation",
-      description: "Condense the conversation to reduce token usage while preserving important context",
-      parameters: {
-        type: "object" as const,
-        properties: {
-          context: {
-            type: "string",
-            description: "Optional context for the condensing operation"
-          }
-        },
-        required: []
-      }
-    }
-  }
 ];
 
 export class MultiLLMAgent extends EventEmitter {
@@ -220,8 +203,8 @@ export class MultiLLMAgent extends EventEmitter {
     this.confirmationTool = new ConfirmationTool();
     this.condenseTool = new CondenseTool();
 
-    // Initialize LLM config for condensing - use provided config or derive from current client
-    this.llmConfig = llmConfig || this.deriveLLMConfigFromClient();
+    // Store provided LLM config if any (for backward compatibility)
+    this.llmConfig = llmConfig || { provider: 'openai', model: 'gpt-4', apiKey: '', baseURL: undefined };
 
     // Initialize token counter with the current model for accurate counting
     this.tokenCounter = createTokenCounter(this.llmClient.getCurrentModel());
@@ -408,47 +391,7 @@ export class MultiLLMAgent extends EventEmitter {
     }
   }
 
-  private deriveLLMConfigFromClient(): LLMConfig {
-    // Extract the current configuration from the LLM client
-    const currentModel = this.llmClient.getCurrentModel();
-
-    // Determine provider based on model name
-    let provider: LLMProvider = 'openai'; // default fallback
-    let apiKey = '';
-    let baseURL: string | undefined;
-
-    // Check for Anthropic models
-    if (currentModel.includes('claude')) {
-      provider = 'anthropic';
-      apiKey = process.env.ANTHROPIC_API_KEY || '';
-      baseURL = process.env.ANTHROPIC_BASE_URL;
-    }
-    // Check for Grok models
-    else if (currentModel.includes('grok')) {
-      provider = 'grok';
-      apiKey = process.env.GROK_API_KEY || '';
-      baseURL = process.env.GROK_BASE_URL || 'https://api.x.ai/v1';
-    }
-    // Check for local models
-    else if (currentModel === 'custom-model' || process.env.LOCAL_LLM_BASE_URL) {
-      provider = 'local';
-      apiKey = process.env.LOCAL_LLM_API_KEY || 'local-key';
-      baseURL = process.env.LOCAL_LLM_BASE_URL || 'http://localhost:1234/v1';
-    }
-    // Default to OpenAI
-    else {
-      provider = 'openai';
-      apiKey = process.env.OPENAI_API_KEY || '';
-      baseURL = process.env.OPENAI_BASE_URL;
-    }
-
-    return {
-      provider,
-      model: currentModel,
-      apiKey,
-      baseURL
-    };
-  }
+
 
   private async initializeMCP(): Promise<void> {
     try {
@@ -1156,11 +1099,11 @@ export class MultiLLMAgent extends EventEmitter {
 
     const condenseResult = await condenseConversation(
       jurikoMessages,
-      this.llmConfig,
+      this.llmClient,
       this.tokenCounter,
       currentTokens,
       {
-        maxMessagesToKeep: 3,
+        maxMessagesToKeep: 0, // Don't keep any old messages, replace everything with summary
         isAutomaticTrigger,
         systemPrompt: (() => {
           const systemMsg = this.messages.find(m => m.role === 'system');
@@ -1173,24 +1116,20 @@ export class MultiLLMAgent extends EventEmitter {
     );
 
     if (!condenseResult.error) {
-      // Convert back to LLMMessage[] and update the messages
+      // Completely replace messages with condensed version
       this.messages = condenseResult.messages.map(msg => ({
         role: msg.role as 'system' | 'user' | 'assistant' | 'tool',
         content: typeof msg.content === 'string' ? msg.content : JSON.stringify(msg.content),
         tool_calls: (msg as any).tool_calls,
         tool_call_id: (msg as any).tool_call_id
       }));
 
-      // Update chat history to reflect the condensing
-      const condensedEntry: ChatEntry = {
-        type: "assistant",
-        content: `📝 **Conversation Summary**\n\n${condenseResult.summary}`,
+      // Completely replace chat history with just the condensed summary
+      this.chatHistory = [{
+        type: "user",
+        content: `Previous conversation summary:\n\n${condenseResult.summary}`,
         timestamp: new Date(),
-      };
-
-      // Replace older entries with the summary, keep recent ones
-      const recentEntries = this.chatHistory.slice(-6); // Keep last 6 entries
-      this.chatHistory = [condensedEntry, ...recentEntries];
+      }];
     }
 
     return condenseResult;
@@ -1242,8 +1181,6 @@ export class MultiLLMAgent extends EventEmitter {
         case "update_todo_list":
           return await this.todoTool.updateTodoList(args.updates);
 
-        case "condense_conversation":
-          return await this.condenseTool.condenseConversation(args.context);
 
         default:
           // Check if it's an MCP tool
@@ -1285,8 +1222,8 @@ export class MultiLLMAgent extends EventEmitter {
     // Update token counter for new model
     this.tokenCounter.dispose();
     this.tokenCounter = createTokenCounter(model);
-    // Update LLM config for condensing to match new model
-    this.llmConfig = this.deriveLLMConfigFromClient();
+    // Update LLM config for backward compatibility
+    this.llmConfig = { provider: 'openai', model: model, apiKey: '', baseURL: undefined };
   }
 
   getLLMClient(): LLMClient {
@@ -1295,8 +1232,8 @@ export class MultiLLMAgent extends EventEmitter {
 
   setLLMClient(client: LLMClient): void {
     this.llmClient = client;
-    // Update LLM config for condensing to match new client
-    this.llmConfig = this.deriveLLMConfigFromClient();
+    // Update LLM config for backward compatibility
+    this.llmConfig = { provider: 'openai', model: client.getCurrentModel(), apiKey: '', baseURL: undefined };
   }
 
   abortCurrentOperation(): void {

src/ui/components/settings-menu.tsx

@@ -1,11 +1,12 @@
 import React, { useState, useEffect } from 'react';
 import { Box, Text, Newline, useInput } from 'ink';
-import { 
-  getEffectiveSettings, 
-  saveResponseStyle, 
-  saveBetaFeatures, 
+import {
+  getEffectiveSettings,
+  saveResponseStyle,
+  saveBetaFeatures,
   saveSecurityLevel,
-  resetAllSettings 
+  saveCondenseThreshold,
+  resetAllSettings
 } from '../../utils/user-settings';
 
 interface SettingsMenuProps {
@@ -31,6 +32,7 @@ export const SettingsMenu: React.FC<SettingsMenuProps> = ({ onClose }) => {
     { key: 'enableBatching', label: 'Multi-Tool Batching (BETA)', type: 'toggle' as const },
     { key: 'enableCodeReferences', label: 'Code References (BETA)', type: 'toggle' as const },
     { key: 'securityLevel', label: 'Security Level', type: 'select' as const },
+    { key: 'condenseThreshold', label: 'Auto-Condense Threshold', type: 'number' as const },
     { key: 'reset', label: 'Reset to Defaults', type: 'action' as const },
     { key: 'close', label: 'Close Settings', type: 'action' as const },
   ];
@@ -67,6 +69,9 @@ export const SettingsMenu: React.FC<SettingsMenuProps> = ({ onClose }) => {
         case 'securityLevel':
           await saveSecurityLevel(value);
           break;
+        case 'condenseThreshold':
+          await saveCondenseThreshold(value);
+          break;
       }
 
       setSaveStatus('✅ Settings saved successfully!');
@@ -130,6 +135,16 @@ export const SettingsMenu: React.FC<SettingsMenuProps> = ({ onClose }) => {
           setSettings(newSettings);
           await saveIndividualSetting('securityLevel', levels[nextIndex]);
         }
+      } else if (selectedItem.type === 'number') {
+        if (selectedItem.key === 'condenseThreshold') {
+          // Cycle through common threshold values: 50%, 60%, 70%, 75%, 80%, 85%, 90%
+          const thresholds = [50, 60, 70, 75, 80, 85, 90];
+          const currentIndex = thresholds.indexOf(settings.condenseThreshold);
+          const nextIndex = currentIndex === -1 ? 0 : (currentIndex + 1) % thresholds.length;
+          const newSettings = { ...settings, condenseThreshold: thresholds[nextIndex] };
+          setSettings(newSettings);
+          await saveIndividualSetting('condenseThreshold', thresholds[nextIndex]);
+        }
       }
     } else if (key.escape) {
       onClose();
@@ -163,6 +178,8 @@ export const SettingsMenu: React.FC<SettingsMenuProps> = ({ onClose }) => {
         return settings.enableCodeReferences ? '[ON]' : '[OFF]';
       case 'securityLevel':
         return `[${settings.securityLevel}]`;
+      case 'condenseThreshold':
+        return `[${settings.condenseThreshold}%]`;
       default:
         return '';
     }
@@ -213,9 +230,10 @@ export const SettingsMenu: React.FC<SettingsMenuProps> = ({ onClose }) => {
         <Text color="gray">Navigation:</Text>
         <Text color="gray">  ↑/↓ - Navigate  ENTER - Toggle/Change  ESC - Close</Text>
         <Newline />
-        <Text color="yellow">Beta Features:</Text>
+        <Text color="yellow">Features:</Text>
         <Text color="gray">  • Multi-Tool Batching: Execute multiple tools in parallel</Text>
         <Text color="gray">  • Code References: Clickable file links in VSCode</Text>
+        <Text color="gray">  • Auto-Condense: Automatically condense conversation at threshold</Text>
         <Newline />
         <Text color="cyan">Settings saved to: ~/.juriko/user-settings.json</Text>
       </Box>

src/utils/condense.ts

@@ -1,6 +1,6 @@
 import { JurikoMessage } from "../juriko/client";
 import { LLMClient } from "../llm/client";
-import { LLMConfig, LLMMessage } from "../llm/types";
+import { LLMMessage } from "../llm/types";
 import { TokenCounter } from "./token-counter";
 import { getCondenseThresholdWithEnv } from "./user-settings";
 
@@ -25,7 +25,7 @@ export interface CondenseOptions {
  */
 export async function condenseConversation(
   messages: JurikoMessage[],
-  llmConfig: LLMConfig,
+  llmClient: LLMClient,
   tokenCounter: TokenCounter,
   prevContextTokens: number,
   options: CondenseOptions = {}
@@ -100,8 +100,7 @@ export async function condenseConversation(
         "\n\nOutput only the summary of the conversation so far, without any additional commentary or explanation."
     };
 
-    // Create LLM client for condensing
-    const llmClient = new LLMClient(llmConfig);
+    // Use the provided LLM client directly
 
     // Prepare messages for condensing (exclude system message from the conversation to be summarized)
     const conversationMessages = messages.filter(m => m.role !== 'system');
@@ -160,9 +159,11 @@ export async function condenseConversation(
       content: `Previous conversation summary:\n\n${summary}`
     });
 
-    // Add the most recent messages to maintain context
-    const recentMessages = messages.slice(-maxMessagesToKeep);
-    condensedMessages.push(...recentMessages);
+    // Only add recent messages if maxMessagesToKeep > 0
+    if (maxMessagesToKeep > 0) {
+      const recentMessages = messages.slice(-maxMessagesToKeep);
+      condensedMessages.push(...recentMessages);
+    }
 
     // Calculate new token count
     const newContextTokens = tokenCounter.countMessageTokens(condensedMessages as any);