OpenClaw Optimizer: Context Compaction for Subagents

Context Compaction Feature

Key Capabilities

• Automatic token tracking across sessions
• Auto-summarization at 50,000 token threshold
• Preserves critical context and task intent
• Cost-effective summarization using Haiku model

Usage Example

const { SubagentContextCompactor } = require('./context-compaction');
 
// Initialize compactor
const compactor = new SubagentContextCompactor({
  tokenThreshold: 50000,
  summaryModel: 'claude-3-5-haiku'
});
 
// Before spawning subagent
const preparedContext = await compactor.prepareForSubagent(fullContext, sessionKey);
sessions_spawn({
  context: preparedContext,
  // other spawn parameters
});

How It Works

1. Tracks token usage across main and subagent sessions
2. Detects when context approaches 50,000 tokens
3. Extracts and preserves critical information
4. Generates a concise summary using Haiku
5. Maintains session continuity

Performance Impact

• Minimal overhead
• Reduces API costs by 70-90%
• Prevents runaway context growth

Logging and Telemetry

Compaction events are logged with:

• Session Key
• Original Token Count
• Compacted Token Count
• Detected Key Themes

Customization

• Adjust token threshold
• Modify critical information patterns
• Configure preservation of recent exchanges