Agentic-Compaction-and-Pipleline-by-GLM-5/download/README.md

# Agent System - Complete Implementation

A comprehensive, open-source implementation of context compaction, agent orchestration, and subagent spawning.

## 📦 Package Contents

```
agent-system/
├── core/
│   ├── token-counter.ts      # Token counting & management
│   ├── summarizer.ts         # Conversation summarization (uses LLM)
│   ├── context-manager.ts    # Context compaction logic
│   ├── orchestrator.ts       # Agent orchestration system
│   └── subagent-spawner.ts   # Subagent spawning mechanism
├── agents/
│   ├── base-agent.ts         # Base agent class
│   └── task-agent.ts         # Task-specific agent
├── storage/
│   └── memory-store.ts       # Persistent file storage
├── utils/
│   └── helpers.ts            # Utility functions
└── index.ts                  # Main exports
```

## 🚀 Quick Start

### Installation

1. Copy the `agent-system` folder to your project
2. Install dependencies:
```bash
bun add z-ai-web-dev-sdk
```

### Basic Usage

```typescript
import { 
  TokenCounter, 
  ContextCompactor,
  AgentOrchestrator,
  SubagentSpawner,
  createAgent 
} from './agent-system';

// Token Counting
const counter = new TokenCounter(128000);
const result = counter.countText("Hello world");
console.log(result.tokens); // ~3 tokens

// Context Compaction
const compactor = new ContextCompactor({
  maxTokens: 100000,
  strategy: 'hybrid'
});

if (compactor.needsCompaction(messages)) {
  const compacted = await compactor.compact(messages);
  console.log(`Saved ${compacted.tokensSaved} tokens`);
}

// Agent Orchestration
const orchestrator = new AgentOrchestrator();
orchestrator.registerAgent({
  id: 'worker-1',
  name: 'Worker',
  type: 'worker',
  capabilities: ['process'],
  maxConcurrentTasks: 3,
  timeout: 60000
});
orchestrator.start();

// Subagent Spawning
const spawner = new SubagentSpawner();
const result = await spawner.executeWithSubagent(
  'researcher',
  'Research AI agents'
);

// Create Custom Agent
const agent = createAgent(
  'MyAgent',
  'You are a helpful assistant.'
);
await agent.initialize();
const response = await agent.act('Hello!');
```

## 📚 Components

### 1. Token Counter
Estimates token counts using character-based approximation (~4 chars/token).

```typescript
const counter = new TokenCounter(128000); // max tokens

// Count text
counter.countText("text"); // { tokens, characters, words }

// Count conversation
counter.countConversation(messages); // { total, breakdown[] }

// Check budget
counter.getBudget(usedTokens); // { used, remaining, total, percentageUsed }
```

### 2. Conversation Summarizer
Creates intelligent summaries using LLM.

```typescript
const summarizer = new ConversationSummarizer();

const result = await summarizer.summarize(messages, {
  maxSummaryTokens: 1000,
  extractKeyPoints: true,
  extractDecisions: true,
  extractActionItems: true
});
// result.summary, result.keyPoints[], result.decisions[], result.actionItems[]
```

### 3. Context Compactor
4 compaction strategies:
- `sliding-window` - Keep recent messages only
- `summarize-old` - Summarize older messages
- `priority-retention` - Keep important messages
- `hybrid` - Combine all strategies

```typescript
const compactor = new ContextCompactor({
  maxTokens: 120000,
  targetTokens: 80000,
  strategy: 'hybrid',
  preserveRecentCount: 6,
  triggerThreshold: 80 // % of maxTokens
});

const result = await compactor.compact(messages);
// result.messages, result.tokensSaved, result.compressionRatio
```

### 4. Agent Orchestrator
Manages agent lifecycle and task routing.

```typescript
const orchestrator = new AgentOrchestrator();

// Register agent
orchestrator.registerAgent({
  id: 'agent-1',
  name: 'Worker',
  type: 'worker',
  capabilities: ['process', 'execute'],
  maxConcurrentTasks: 3,
  timeout: 60000
});

// Create task
orchestrator.createTask('process', 'Process data', { data: [...] });

// Events
orchestrator.on('task_completed', (event) => { ... });

// Start processing
orchestrator.start();
```

### 5. Subagent Spawner
Spawn and manage child agents for parallel execution.

```typescript
const spawner = new SubagentSpawner();

// Single execution
const result = await spawner.executeWithSubagent('explorer', 'Find files');

// Parallel execution
const results = await spawner.executeParallel([
  { type: 'explorer', input: 'Task 1' },
  { type: 'coder', input: 'Task 2' }
]);

// Pipeline execution (context passes between steps)
const pipeline = await spawner.executePipeline([
  { type: 'planner', input: 'Plan task' },
  { type: 'coder', input: (prev) => `Implement: ${prev}` },
  { type: 'reviewer', input: (prev) => `Review: ${prev}` }
]);
```

### 6. Predefined Subagent Types
- `explorer` - Code exploration
- `researcher` - Information gathering
- `coder` - Code generation
- `reviewer` - Code review
- `planner` - Task planning
- `executor` - Task execution
- `custom` - Your custom type

## 🔧 Configuration Options

### CompactionConfig
```typescript
{
  maxTokens: 120000,           // Maximum context tokens
  targetTokens: 80000,         // Target after compaction
  strategy: 'hybrid',          // Compaction strategy
  preserveRecentCount: 6,      // Keep last N messages
  preserveSystemMessage: true, // Keep system message
  priorityKeywords: ['important', 'critical'],
  summaryMaxTokens: 2000,      // Max tokens for summaries
  triggerThreshold: 80         // % to trigger compaction
}
```

### AgentConfig
```typescript
{
  id: 'agent-id',
  name: 'Agent Name',
  type: 'agent-type',
  capabilities: ['task1', 'task2'],
  maxConcurrentTasks: 3,
  timeout: 60000
}
```

## 📄 License

MIT License - Free to use, modify, and distribute.

## 🤝 Contributing

Feel free to extend and customize for your needs!