Claude Parser

A high-performance stream-JSON parser for Claude Code CLI output, designed for benchmarking and training data collection.

Features

• Stream Event Parsing: Supports all Claude event types:

  • message_start, message_stop
  • content_block_start, content_block_delta, content_block_stop
  • tool_use, function_result
  • thinking
  • error
  • usage

• Performance Metrics Extraction:

  • Timing measurements (total duration, time to first output)
  • Token usage tracking (input, output, total)
  • Tool invocation statistics
  • Thinking sequence analysis
  • Error and recovery patterns

• Error Handling: Robust error recovery with detailed tracking of error types and recovery strategies

• Export Functionality: Export parsed data for training in JSON or JSONL format

Usage

use claude_parser::ClaudeStreamParser;
use tokio::fs::File;

#[tokio::main]
async fn main() -> Result<(), Box<dyn std::error::Error>> {
    // Create parser
    let mut parser = ClaudeStreamParser::new();
    
    // Parse stream from file
    let file = File::open("claude_output.jsonl").await?;
    let metrics = parser.parse_stream(file).await?;
    
    // Access performance metrics
    println!("Total duration: {:?}", metrics.total_duration);
    println!("Tool invocations: {}", metrics.tool_invocations.len());
    println!("Total tokens: {}", metrics.token_usage.total_tokens);
    
    // Export training data
    let export = parser.export_training_data();
    export.to_json_file("training_data.json").await?;
    
    Ok(())
}

Metrics Collected

Tool Metrics

• Invocation count per tool
• Average duration
• Success rate
• Parameter sizes

Thinking Metrics

• Total sequences
• Token count per sequence
• Duration analysis
• Pattern recognition

Error Metrics

• Error types and frequencies
• Recovery success rates
• Recovery strategies used

Timeline

• Chronological event sequence
• Relative timing information
• Event correlations

Integration with Benchmarking

This parser is designed to work with the RUV-SWARM benchmarking system for:

• SWE-Bench instance evaluation
• Real-time performance monitoring
• Comparative analysis between baseline and ML-optimized approaches
• Training data collection for model improvement

License

Licensed under MIT OR Apache-2.0