# EmailResponder-MCP Evaluation Tools

Evaluation scripts for analyzing model performance on the EmailResponder-MCP task.

## Scripts

### evaluate_success.py
Calculates success rates for each model based on execution logs.

```bash
python evaluate_success.py
```

Outputs:
- `success-finish_detailed_results.json` - Detailed results per session
- `success-finish_rate.csv` - Summary statistics

### evaluate_trajectory.py
Evaluates trajectory metrics comparing actual execution paths against reference trajectory.

```bash
python evaluate_trajectory.py --config reference_trajectory.yaml --output evaluation_results.csv
```

Metrics:
- Exact match
- In-order match
- Any-order match
- Precision / Recall
- Single-tool use
- Path diversity (unique_path_ratio, path_entropy)

### analyze_retry_patterns.py
Analyzes retry patterns and error distributions across models.

```bash
python analyze_retry_patterns.py
```

Outputs:
- `retry_analysis.json` - Detailed retry analysis
- `retry_summary.csv` - Retry statistics summary
- `error_by_agent.csv` - Error counts by agent

## Configuration

Edit `reference_trajectory.yaml` to customize:
- Reference trajectory (ground truth)
- Target tools for single-tool use metric
- Models to evaluate
- Extraction types (SPAN, Chain, AGENT, LLM, Tool)

## Requirements

- Python 3.7+
- pandas
- pyyaml