analyzer

analyzer/README.md

@@ -0,0 +1,125 @@
+# SWE-Bench Trajectory Analysis
+
+A comprehensive analysis tool for comparing failed and successful software engineering trajectories across different SWE-bench dataset splits. This tool helps identify patterns in agent failures and provides actionable insights for improving software engineering agents.
+
+## Features
+
+- **Multi-Split Support**: Analyze trajectories across SWE-bench splits (Verified, Lite, Test, Multimodal)
+- **Automated Comparison**: Compare your agent's failed trajectories with successful trajectories from top-performing agents
+- **LLM-Powered Analysis**: Uses advanced language models to generate detailed failure analysis reports
+- **Batch Processing**: Process multiple instances in parallel for balanced analysis
+- **Leaderboard Integration**: Automatically downloads and analyzes top agent results from SWE-bench leaderboards
+
+## Requirements
+
+- Python 3.8+
+- Required packages (see `requirements.txt`):
+  - `datasets>=2.0.0` - For loading SWE-bench datasets
+  - `requests>=2.25.0` - For API calls
+  - `python-dotenv>=0.19.0` - For environment variable management
+  - `PyYAML>=6.0` - For configuration file parsing
+  - `boto3>=1.40.52` - For AWS S3 access
+
+## Installation
+
+1. Clone the repository and navigate to the directory:
+   ```bash
+   cd trajectory_analysis
+   ```
+
+2. Install dependencies:
+   ```bash
+   pip install -r requirements.txt
+   ```
+
+3. Set up environment variables:
+   ```bash
+   cp .env.example .env
+   # Edit .env and add your OPENROUTER_API_KEY
+   ```
+
+## Configuration
+
+Edit `config.yaml` to configure your analysis:
+
+```yaml
+# Dataset configuration
+dataset:
+  split: "Verified"  # Options: Verified, Lite, Test, Multimodal
+
+# Paths
+paths:
+  txt_path: "/path/to/unsolved_instances.txt"
+  intermediate_output_path: "/path/to/intermediate_output/"
+  output_path: "/path/to/final_report.txt"
+  logs_directory: "/path/to/failed_trajectories/"
+  prompts_path: "/path/to/prompts_manager.py"
+  tools_path: "/path/to/tools/"
+
+# Settings
+settings:
+  max_workers: 20 # Maximum number of parallel requests to OpenRouter
+  num_agents: 4 # Number of top SWE-bench leaderboard agents' to download trajectories for
+```
+
+## Usage
+
+### Basic Usage
+
+1. **Prepare your data**:
+   - Create a text file listing unresolved instance IDs (one per line)
+   - Ensure your failed trajectories are in the specified logs directory
+
+2. **Run the analysis**:
+   ```bash
+   python run.py
+   ```
+
+### Input Requirements
+
+- **Failed Trajectories**: Your agent's failed trajectories in one of these formats:
+  - Direct `.log` files named `{instance_id}.log`
+  - Issues directory structure: `issues/{numerical_id}-{instance_id}/logs/{timestamp}/run/{instance_id}/generator/{generator_id}/`
+
+- **Unresolved Instances**: Text file containing instance IDs that your agent failed to solve
+
+### Output
+
+The tool generates:
+- **Individual Analysis**: Detailed failure analysis for each instance in the intermediate output directory
+- **Comprehensive Report**: Categorized analysis of all failures with improvement recommendations
+
+## Architecture
+
+### Core Components
+
+- **`run.py`**: Main entry point and batch processing coordinator
+- **`trajectory_discovery.py`**: Handles trajectory finding and leaderboard integration
+- **`pairwise_analysis.py`**: LLM-powered failure analysis and dataset loading
+- **`utils.py`**: Utility functions for file operations and trajectory parsing
+- **`config.yaml`**: Configuration management
+
+## Supported SWE-bench Splits
+
+| Split | Dataset | Instances | Description |
+|-------|---------|-----------|-------------|
+| Verified | `princeton-nlp/SWE-bench_Verified` | 500 | Verified high-quality instances |
+| Lite | `SWE-bench/SWE-bench_Lite` | 300 | Smaller subset for faster testing |
+| Test | `SWE-bench/SWE-bench` | 2,294 | Full test set |
+| Multimodal | `SWE-bench/SWE-bench_Multimodal` | 510 | Instances with UI elements |
+
+## Error Handling
+
+The tool includes robust error handling for:
+- Missing trajectory files
+- API rate limiting and failures
+- Dataset loading errors
+- Invalid configuration parameters
+
+# Contributor
+
+- [Hengzhi Zhang](https://henryzhang11.github.io)
+
+## License
+
+This project is licensed under the MIT License - see the project LICENSE file for details.

analyzer/config.yaml

@@ -0,0 +1,17 @@
+# Configuration file for trajectory analysis
+paths:
+  txt_path: "/root/henry/trajectory_analysis/unsolved.txt"
+  intermediate_output_path: "/root/henry/trajectory_analysis_test/per_instance_failure_analysis"
+  output_path: "/root/henry/trajectory_analysis_test/failure_analysis_report.txt"
+  logs_directory: "/root/peteWong/project/Code-agent/batch_out/batch-first/issues"
+  prompts_path: "/root/henry/Code-agent/src/managers/prompts/prompts_manager.py"
+  tools_path: "/root/henry/Code-agent/src/tools"
+
+# Dataset configuration
+dataset:
+  split: "Verified"  # Options: Verified, Lite, Test, Multimodal (bash-only not available on Hugging Face)
+
+# Optional: Add other configuration parameters
+settings:
+  max_workers: 20
+  num_agents: 1

analyzer/pairwise_analysis.py

@@ -0,0 +1,179 @@
+import json
+import re
+import os
+import requests
+import time
+from datetime import datetime
+from pathlib import Path
+from typing import List, Optional, Callable
+from datasets import load_dataset
+import random
+from dotenv import load_dotenv
+
+
+# Load environment variables from .env file
+load_dotenv()
+
+def failure_analysis_prompt(problem_statement: str, hints_text: str, working_trajectory: str, failed_trajectory: str, code_agent_code: str) -> str:
+    """
+    Generate a failure analysis prompt for comparing successful and failed trajectories.
+    
+    Creates a structured prompt that asks an AI model to analyze why one trajectory
+    succeeded while another failed, and what the failed trajectory could have done
+    differently to succeed.
+    """
+    return f"""Here's a software engineering issue: {problem_statement}.
+
+Here's the hints: {hints_text}.
+
+Here's the trajectory that fixed the issue: {working_trajectory}.
+
+Here's the trajectory that failed: {failed_trajectory}.
+
+Here's the prompts/tools code for the software engineering agent: {code_agent_code}.
+
+Why does the first SWE-agent resolve the issue but the second failed?
+
+What could the second agent have done differently to resolve the issue?
+
+What improvements could be made to the prompt/tool usage?
+
+Answer in the following format:
+# Why the second agent failed 
+...
+# What could the second agent have done differently
+...
+# What improvements could be made to the prompt/tool usage
+..."""
+
+def openrouter_gemini_client(prompt: str, max_retries: int = 10, base_delay: float = 1.0) -> str:
+    """
+    Send a prompt to OpenRouter API using Gemini model for analysis.
+    
+    Makes API calls to OpenRouter service using Google's Gemini 2.5 Pro model
+    with exponential backoff retry logic. Handles authentication, rate limiting,
+    and error recovery for robust API communication.
+    
+    Args:
+        prompt (str): The analysis prompt to send to the model
+        max_retries (int, optional): Maximum number of retry attempts. Defaults to 10.
+        base_delay (float, optional): Base delay in seconds for exponential backoff. Defaults to 1.0.
+        
+    Returns:
+        str: The model's response content as a string
+        
+    Raises:
+        ValueError: If OPENROUTER_API_KEY environment variable is not set
+        Exception: If API request fails after all retry attempts or returns invalid response
+        
+    Note:
+        Requires OPENROUTER_API_KEY environment variable to be set with valid API key.
+        Uses exponential backoff: 1s, 2s, 4s, 8s, 16s, etc. between retries.
+    """
+    api_key = os.getenv('OPENROUTER_API_KEY')
+    if not api_key:
+        raise ValueError("OPENROUTER_API_KEY environment variable not found")
+    
+    url = "https://openrouter.ai/api/v1/chat/completions"
+    
+    headers = {
+        "Authorization": f"Bearer {api_key}",
+        "Content-Type": "application/json"
+    }
+    
+    data = {
+        "model": "google/gemini-2.5-pro",
+        "messages": [
+            {
+                "role": "user",
+                "content": prompt
+            }
+        ],
+        "max_tokens": 200000,  # Increased limit for comprehensive analysis
+        "temperature": 0.7  # Lower temperature for more consistent summaries
+    }
+    
+    for attempt in range(max_retries):
+        try:
+            print(f"ATTEMPT: API attempt {attempt + 1}/{max_retries}...")
+            response = requests.post(url, headers=headers, json=data, timeout=30)
+            response.raise_for_status()
+            
+            result = response.json()
+                        
+            if 'choices' not in result or not result['choices']:
+                raise Exception(f"Invalid API response: missing or empty choices")
+            
+            content = result['choices'][0]['message']['content']
+            
+            if not content or content.strip() == "":
+                print(f"WARNING: Empty content returned from API on attempt {attempt + 1}/{max_retries}")
+                if attempt < max_retries - 1:
+                    delay = base_delay * (2 ** attempt) / (2 ** random.uniform(0, 1))  # Exponential backoff with jitter
+                    print(f"WAIT: Waiting {delay:.2f} seconds before retry...")
+                    time.sleep(delay)
+                    continue
+                else:
+                    raise Exception(f"Empty content returned from API after {max_retries} attempts")
+            
+            print(f"SUCCESS: API request successful on attempt {attempt + 1}")
+            return content.strip()
+            
+        except requests.exceptions.RequestException as e:
+            print(f"FAILED: Request failed on attempt {attempt + 1}/{max_retries}: {e}")
+            if attempt < max_retries - 1:
+                delay = base_delay * (2 ** attempt) / (2 ** random.uniform(0, 1))  # Exponential backoff with jitter
+                print(f"WAIT: Waiting {delay:.2f} seconds before retry...")
+                time.sleep(delay)
+                continue
+            else:
+                raise Exception(f"FAILED: OpenRouter API request failed after {max_retries} attempts: {e}")
+        
+        except (KeyError, IndexError) as e:
+            raise Exception(f"FAILED: Unexpected response format from OpenRouter API: {e}")
+    
+    raise Exception(f"FAILED: All {max_retries} attempts failed")
+
+def get_swe_bench_problem_and_hints_for_split(instance_id: str, split_name: str) -> tuple[str, str]:
+    """
+    Retrieve problem statement and hints for a specific SWE-Bench instance in a specific split.
+    
+    Loads the appropriate SWE-Bench dataset and searches for the specified instance ID to extract the problem statement and hints text. 
+    This data is used for generating analysis prompts and understanding the context of the software engineering problem.
+    
+    Args:
+        instance_id (str): The unique identifier for the SWE-Bench instance
+                          (e.g., 'scikit-learn__scikit-learn-10908')
+        split_name (str): The split name (e.g., "Verified", "Lite", "Test", "Multimodal", "bash-only")
+        
+    Returns:
+        tuple[str, str]: A tuple containing (problem_statement, hints_text)
+        
+    Raises:
+        ValueError: If the specified instance_id is not found in the dataset
+    Note:
+        This function loads the full SWE-Bench dataset for the specified split, which may take some time on first execution due to dataset download and caching.
+    """
+    dataset_mapping = {
+        "Verified": "princeton-nlp/SWE-bench_Verified",
+        "Lite": "SWE-bench/SWE-bench_Lite", 
+        "Test": "SWE-bench/SWE-bench",
+        "Multimodal": "SWE-bench/SWE-bench_Multimodal",
+        # Note: bash-only dataset not yet available on Hugging Face
+    }
+    
+    dataset_name = dataset_mapping.get(split_name)
+    if not dataset_name:
+        raise ValueError(f"No dataset mapping found for split '{split_name}'")
+    
+    ds = load_dataset(dataset_name, split="test")
+    target_row = None
+    for row in ds:
+        if row['instance_id'] == instance_id:
+            target_row = row
+            break
+    if target_row is None:
+        raise ValueError(f"Instance ID '{instance_id}' not found in the {split_name} dataset")
+    problem_statement = target_row['problem_statement']
+    hints_text = target_row['hints_text']
+    return problem_statement, hints_text

analyzer/requirements.txt

@@ -0,0 +1,5 @@
+datasets>=2.0.0
+requests>=2.25.0
+python-dotenv>=0.19.0
+PyYAML>=6.0
+boto3>=1.40.52

analyzer/run.py

@@ -0,0 +1,222 @@
+import json
+import os
+import yaml
+from typing import List, Dict, Any, Tuple, Optional
+from pathlib import Path
+from concurrent.futures import ThreadPoolExecutor, as_completed
+import threading
+from trajectory_discovery import find_failed_trajectory, find_successful_trajectory_for_split, download_top_agents_logs_for_split, count_top_agents_solved_for_split
+from pairwise_analysis import get_swe_bench_problem_and_hints_for_split, failure_analysis_prompt, openrouter_gemini_client
+
+
+def load_config(config_path: str = "config.yaml") -> Dict[str, Any]:
+    """Load configuration from YAML file."""
+    with open(config_path, 'r', encoding='utf-8') as f:
+        return yaml.safe_load(f)
+
+
+def extract_code_agent_code(prompts_path: str, tools_path: str) -> Dict[str, str]:
+    """
+    Extract content from Code-agent directory including prompts_manager.py and tools/ files.
+    """
+    result = ""
+    prompts_path = Path(prompts_path)
+    if prompts_path.exists():
+        try:
+            with open(prompts_path, 'r', encoding='utf-8') as f:
+                result += f"{prompts_path.name}: \n" + "```python\n" + f.read() + "```\n"
+        except Exception as e:
+            print(f"Error reading prompts_manager.py: {e}")   
+    tools_path = Path(tools_path)
+    if tools_path.exists():
+        for file_path in tools_path.iterdir():
+            if file_path.is_file() and not file_path.name.startswith('test'):
+                try:
+                    with open(file_path, 'r', encoding='utf-8') as f:
+                        result += f"{file_path.name}: \n" + "```python\n" + f.read() + "```\n"
+                except Exception as e:
+                    print(f"Error reading {file_path}: {e}")
+    return result
+
+def analyze_failure_causes(instance_id: str, current_directory: str, code_agent_code: str, split_name: str, num_agents: int = 5) -> Optional[str]:
+    """
+    Compare trajectories alongside test results to analyze specific failure causes.
+    
+    This function finds both failed and successful trajectories for a given instance, retrieves the problem context, and uses an LLM to analyze why the failure occurred.
+    
+    Args:
+        instance_id: The instance ID to analyze
+        current_directory: Directory containing failed trajectories
+        code_agent_code: Code agent code for analysis
+        split_name: The split name (e.g., "Verified", "Lite", "Test", "Multimodal", "bash-only")
+        num_agents: Number of top agents to check for successful trajectories
+    """
+    failed_trajectory_path = find_failed_trajectory(instance_id, current_directory)
+    if not failed_trajectory_path:
+        print(f"No failed trajectory found for {instance_id}")
+        return None    
+    successful_trajectory_path = find_successful_trajectory_for_split(instance_id, split_name, num_agents)
+    if not successful_trajectory_path:
+        print(f"No successful trajectory found for {instance_id} in {split_name} split")
+        return None
+    try:
+        with open(failed_trajectory_path, 'r', encoding='utf-8') as f:
+            failed_content = f.read()
+        with open(successful_trajectory_path, 'r', encoding='utf-8') as f:
+            successful_content = f.read()        
+        problem_statement, hints_text = get_swe_bench_problem_and_hints_for_split(instance_id, split_name)        
+        prompt = failure_analysis_prompt(problem_statement, hints_text, successful_content, failed_content, code_agent_code)        
+        analysis_result = openrouter_gemini_client(prompt)        
+        formatted_result = f"Instance: {instance_id}\nAnalysis: {analysis_result}\n"
+        print(f"Analysis result for {instance_id}: {analysis_result}")
+        return formatted_result
+    except Exception as e:
+        print(f"Error analyzing {instance_id}: {e}")
+        return None
+
+def batch_failure_analysis(unresolved_id_txt_path: str, intermediate_output_path: str, output_path: str, logs_directory: str, prompts_path: str, tools_path: str, split_name: str, max_workers: int = 20, num_agents: int = 5) -> None:
+    """
+    Complete batch failure analysis pipeline for a specific split.    
+    Process failed trajectories, find successful trajectories, analyze failure causes, and categorize them by cause. Generate a comprehensive report with recommendations and save it to output_path.
+    
+    Args:
+        unresolved_id_txt_path: Path to text file containing unresolved instance IDs
+        intermediate_output_path: Path to directory for intermediate output files
+        output_path: Path to final output report file
+        logs_directory: Directory containing failed trajectory logs
+        prompts_path: Path to prompts manager Python file
+        tools_path: Path to tools directory
+        split_name: The split name (e.g., "Verified", "Lite", "Test", "Multimodal", "bash-only")
+        max_workers: Maximum number of worker threads
+        num_agents: Number of top agents to analyze
+    """    
+    print(f"Regenerating top agents performance CSV for {split_name} split...")
+    count_top_agents_solved_for_split(split_name, num_agents)
+    print(f"Ensuring experiments repository and logs are up to date for {split_name} split...")
+    download_top_agents_logs_for_split("experiments", split_name, num_agents)
+    try:
+        with open(unresolved_id_txt_path, 'r', encoding='utf-8') as f:
+            unresolved_ids = [line.strip() for line in f.readlines() if line.strip()]
+        
+        if not unresolved_ids:
+            print("No unresolved instances found in text file")
+            return
+        print(f"Found {len(unresolved_ids)} unresolved instances to analyze")
+    except Exception as e:
+        print(f"Error loading text file: {e}")
+        return
+    
+    failure_analysis_results = []
+    skipped_instances = []
+    code_agent_code = extract_code_agent_code(prompts_path, tools_path)
+    # create the intermediate output directory if it doesn't exist
+    os.makedirs(intermediate_output_path, exist_ok=True)
+    
+    def process_instance(instance_id):
+        print(f"\nProcessing: {instance_id}")
+        try:
+            analysis_result = analyze_failure_causes(instance_id, logs_directory, code_agent_code, split_name, num_agents)
+            if analysis_result:
+                with open(os.path.join(intermediate_output_path, f'{instance_id}.txt'), 'w', encoding='utf-8') as f:
+                    f.write(analysis_result)
+                return (instance_id, analysis_result, None)
+            else:
+                print(f"Skipped {instance_id} - no trajectories found")
+                return (instance_id, None, "no trajectories found")
+        except Exception as e:
+            print(f"Error processing {instance_id}: {e}")
+            return (instance_id, None, str(e))
+    
+    with ThreadPoolExecutor(max_workers=max_workers) as executor:
+        futures = [executor.submit(process_instance, instance_id) for instance_id in unresolved_ids]
+        for future in as_completed(futures):
+            instance_id, analysis_result, error = future.result()
+            if analysis_result:
+                failure_analysis_results.append(analysis_result)
+            else:
+                skipped_instances.append(instance_id)
+    # combine the intermediate output files into a single file
+    with open(output_path, 'w', encoding='utf-8') as f:
+        for file in os.listdir(intermediate_output_path):
+            if file.endswith('.txt'):
+                with open(os.path.join(intermediate_output_path, file), 'r', encoding='utf-8') as f1:
+                    f.write(f1.read() + '\n')
+    # read the combined failure analysis report
+    with open(output_path, 'r', encoding='utf-8') as f:
+        failure_analysis_report = f.read()
+    batch_prompt = f"""Read failure analysis reports and group failure causes into categories.
+
+Here are the failure analysis reports:
+
+{failure_analysis_report}
+
+Here are the prompts/tools code for the software engineering agent:
+
+{code_agent_code}
+
+Please read the above analysis and categorize the failure causes into categories.
+
+For each category, please describe the failure cause in detail, list related instances, and suggest improvements for the prompt/tool usage to address the failure cause. 
+
+Format your response as:
+
+# Problem 1: 
+**Description**: [Your analysis of this problem]
+**Related Instances**: 
+- instance_id_1: [brief description of the failure]
+- instance_id_2: [brief description of the failure]
+**Improvements**: [Suggestions for improving the prompt/tool usage to address the failure cause]
+
+# Problem 2:
+**Description**: [Your analysis of this problem]
+**Related Instances**: 
+- instance_id_3: [brief description of the failure]
+- instance_id_4: [brief description of the failure]
+**Improvements**: [Suggestions for improving the prompt/tool usage to address the failure cause]
+
+Continue for all identified problems."""
+    
+    try:
+        batch_analysis = openrouter_gemini_client(batch_prompt)
+        with open(output_path, 'w', encoding='utf-8') as f:
+            f.write(batch_analysis)
+    except Exception as e:
+        print(f"Error generating batch analysis: {e}")
+        return f"Error generating batch analysis: {e}"
+
+if __name__ == "__main__":
+    config = load_config("config.yaml")
+    
+    # Validate paths exist
+    paths = config["paths"]
+    for path_name in ["txt_path", "logs_directory", "prompts_path", "tools_path"]:
+        if not os.path.exists(paths[path_name]):
+            raise FileNotFoundError(f"Path {path_name}: {paths[path_name]} does not exist")
+    
+    # Validate dataset split
+    split_name = config["dataset"]["split"]
+    valid_splits = ["Verified", "Lite", "Test", "Multimodal"]  # bash-only not yet available on Hugging Face
+    if split_name not in valid_splits:
+        raise ValueError(f"split must be one of {valid_splits}, got: {split_name}")
+    
+    # Validate max_workers
+    max_workers = config["settings"]["max_workers"]
+    if not isinstance(max_workers, int) or max_workers <= 0:
+        raise ValueError(f"max_workers must be positive integer, got: {max_workers}")
+    
+    # Validate num_agents
+    num_agents = config["settings"]["num_agents"]
+    if not isinstance(num_agents, int) or num_agents <= 0:
+        raise ValueError(f"num_agents must be positive integer, got: {num_agents}")
+    
+    batch_failure_analysis(
+        paths["txt_path"], 
+        paths["intermediate_output_path"], 
+        paths["output_path"], 
+        paths["logs_directory"], 
+        paths["prompts_path"], 
+        paths["tools_path"], 
+        split_name,
+        max_workers,
+        num_agents
+    )

analyzer/trajectory_discovery.py

@@ -0,0 +1,386 @@
+from __future__ import annotations
+import os
+import re
+import subprocess
+import sys
+import json
+import pathlib
+import csv
+import urllib.request
+import argparse
+from typing import Optional, List, Tuple, Dict, Any
+from pathlib import Path
+from concurrent.futures import ProcessPoolExecutor, as_completed
+import multiprocessing
+from utils import read_json_url_or_path
+from datasets import load_dataset
+from utils import find_failed_trajectory_in_issues_directory
+
+LEADERBOARD_JSON = "https://raw.githubusercontent.com/swe-bench/swe-bench.github.io/master/data/leaderboards.json"
+EXPERIMENTS_DIR = str(Path(__file__).parent / "experiments")
+EXPERIMENTS_REPO_URL = "https://github.com/SWE-bench/experiments.git"
+
+def top_agents_for_split(leaderboards_json: dict, split_name: str, num_agents: int = 5) -> list[dict]:
+    """
+    Get the top N SWE-bench agents for a specific split from the leaderboard.
+    
+    Args:
+        leaderboards_json: The leaderboard JSON data
+        split_name: The split name (e.g., "Verified", "Lite", "Test", "Multimodal")
+        num_agents: Number of top agents to return (default: 5)
+    
+    Returns:
+        List of top N agents sorted by resolution rate
+    """
+    lbs = leaderboards_json.get("leaderboards") or leaderboards_json
+    split_lb = next((lb for lb in lbs if (lb.get("name") or "").lower() == split_name.lower()), None)
+    if not split_lb:
+        raise KeyError(f"Could not find '{split_name}' leaderboard in leaderboards.json")
+    rows = [r for r in split_lb.get("results", []) if not r.get("warning")]
+    rows.sort(key=lambda r: float(r.get("resolved", 0)), reverse=True)
+    return rows[:num_agents]
+
+def read_submission_resolved_ids(submission_dir: pathlib.Path) -> set[str]:
+    """
+    Read the resolved instance_ids from the results.json file in a SWE-bench verified submission folder.
+    """
+    fp = submission_dir / "results" / "results.json"
+    if not fp.is_file():
+        return set()
+    try:
+        j = json.loads(fp.read_text(encoding="utf-8"))
+        return set(j.get("resolved", []))
+    except Exception:
+        return set()
+
+def load_canonical_ids_for_split(split_name: str) -> list[str]:
+    """
+    Load canonical instance IDs for a specific SWE-bench split.
+    
+    Args:
+        split_name: The split name (e.g., "Verified", "Lite", "Test", "Multimodal")
+    
+    Returns:
+        List of sorted instance IDs for the split
+    """
+    dataset_mapping = {
+        "Verified": "princeton-nlp/SWE-bench_Verified",
+        "Lite": "SWE-bench/SWE-bench_Lite", 
+        "Test": "SWE-bench/SWE-bench",
+        "Multimodal": "SWE-bench/SWE-bench_Multimodal",
+        # Note: bash-only dataset not available on Hugging Face
+    }
+    
+    dataset_name = dataset_mapping.get(split_name)
+    if not dataset_name:
+        print(f"Warning: No dataset mapping found for split '{split_name}'")
+        return []
+    
+    try:
+        ds = load_dataset(dataset_name, split="test")
+        ids = sorted(ds["instance_id"])
+        print(f"Loaded {len(ids)} instance IDs for {split_name} split")
+        return ids
+    except Exception as e:
+        print(f"Error loading dataset for {split_name}: {e}")
+        return []
+
+def ensure_experiments_repo() -> pathlib.Path:
+    """Clone or update the SWE-bench experiments repository."""
+    experiments_path = pathlib.Path(EXPERIMENTS_DIR).expanduser().resolve()
+    
+    if experiments_path.exists() and (experiments_path / ".git").exists():
+        try:
+            subprocess.run(["git", "pull"], cwd=experiments_path, check=True, capture_output=True, text=True)
+        except subprocess.CalledProcessError as e:
+            print(f"Warning: Failed to update repository: {e.stderr}")
+            print("Continuing with existing version...")
+    else:
+        try:
+            subprocess.run(["git", "clone", EXPERIMENTS_REPO_URL, str(experiments_path)], check=True, capture_output=True, text=True)
+            print("Repository cloned successfully")
+        except subprocess.CalledProcessError as e:
+            print(f"Error: Failed to clone repository: {e.stderr}")
+            sys.exit(1)
+    return experiments_path
+
+def count_top_agents_solved_for_split(split_name: str, num_agents: int = 5):
+    """
+    Generate CSV showing which top agents solved which instances for a specific split.
+    
+    Args:
+        split_name: The split name (e.g., "Verified", "Lite", "Test", "Multimodal")
+        num_agents: Number of top agents to analyze
+    """
+    # Ensure we have the latest experiments repository
+    ensure_experiments_repo()
+    split_root = Path(EXPERIMENTS_DIR) / "evaluation" / split_name.lower()
+    if not split_root.is_dir():
+        print(f"ERROR: {split_root} not found. Set EXPERIMENTS_DIR correctly.", file=sys.stderr)
+        sys.exit(2)
+
+    lb = read_json_url_or_path(LEADERBOARD_JSON)
+    top_agents = top_agents_for_split(lb, split_name, num_agents)
+
+    columns = [] 
+    local_union_ids: set[str] = set()
+
+    print(f"Top {num_agents} ({split_name}):")
+    for i, row in enumerate(top_agents, 1):
+        folder = row.get("folder")
+        name = row.get("name") or f"rank{i}"
+        print(f"{i:2d}. {name} | resolved={row.get('resolved')} | folder={folder}")
+        if not folder:
+            continue
+        subdir = split_root / folder
+        if not subdir.is_dir():
+            print(f"[warn] missing folder locally: {subdir}", file=sys.stderr)
+            continue
+        resolved = read_submission_resolved_ids(subdir)
+        columns.append((f"{i:02d}_{folder}", resolved))
+        # also collect skipped ids so rows don't get lost
+        try:
+            j = json.loads((subdir / "results" / "results.json").read_text(encoding="utf-8"))
+            local_union_ids |= set(j.get("resolved", []))
+            local_union_ids |= set(j.get("no_generation", []))
+            local_union_ids |= set(j.get("no_logs", []))
+        except Exception:
+            pass
+
+    # decide row universe (instance_ids)
+    instance_ids = load_canonical_ids_for_split(split_name)
+    if not instance_ids:
+        instance_ids = sorted(local_union_ids)
+        print(f"[warn] using local union of IDs ({len(instance_ids)}). "
+              f"For guaranteed canonical IDs, install `datasets` and internet access.", file=sys.stderr)
+
+    out_csv = pathlib.Path(f"top_agents_performance_{split_name.lower()}.csv")
+
+    headers = ["instance_id"] + [col for col, _ in columns] + ["any_top_agents_solved"]
+    with out_csv.open("w", newline="", encoding="utf-8") as f:
+        w = csv.writer(f)
+        w.writerow(headers)
+        
+        non_resolved_instances = []
+        
+        for iid in instance_ids:
+            row_vals = [iid]
+            solved_by_any = 0
+            
+            for _, resolved in columns:
+                solved = 1 if iid in resolved else 0
+                row_vals.append(solved)
+                if solved == 1:
+                    solved_by_any = 1
+            
+            row_vals.append(solved_by_any)
+            w.writerow(row_vals)
+            
+            if solved_by_any == 0:
+                non_resolved_instances.append(iid)
+    
+    print(f"\nResults saved to: {out_csv}")
+    print(f"\nNon-resolved instances ({len(non_resolved_instances)}):")
+    for instance_id in non_resolved_instances:
+        print(f"  {instance_id}")
+
+def find_failed_trajectory(instance_id: str, experiments_dir: str) -> Optional[str]:
+    """
+    Find a failed trajectory for the given instance within the experiments directory (assume names like '{instance_id}.log').
+    If the file exists, return the path to the file.
+    Otherwise, call find_failed_trajectory_in_issues_directory to find the failed trajectory.
+    """
+    trajectory_path = os.path.join(experiments_dir, instance_id + ".log")
+    if os.path.exists(trajectory_path):
+        return trajectory_path
+    else:
+        return find_failed_trajectory_in_issues_directory(instance_id, experiments_dir)
+
+def find_successful_trajectory_for_split(instance_id: str, split_name: str, num_agents: int = 5) -> Optional[str]:
+    """
+    Find a successful trajectory for the given instance in a specific split.
+    Checks the top agents performance CSV to find which agents solved the instance, then looks in each agent's directory for trajectory files.
+    Tries agents in order of ranking until a trajectory is found.
+    Special handling for 20250915_JoyCode agent where trajectories are in instance-named folders.
+    Automatically downloads logs and generates CSV if missing.
+    
+    Args:
+        instance_id: The instance ID to find trajectory for
+        split_name: The split name (e.g., "Verified", "Lite", "Test", "Multimodal")
+        num_agents: Number of top agents to check
+    """
+    # Check if trajectory folders exist, download if missing
+    split_dir = os.path.join(EXPERIMENTS_DIR, "evaluation", split_name.lower())
+    if not os.path.exists(split_dir):
+        download_top_agents_logs_for_split(EXPERIMENTS_DIR, split_name, num_agents)
+    csv_path = os.path.join(os.path.dirname(__file__), f"top_agents_performance_{split_name.lower()}.csv")
+    if not os.path.exists(csv_path):
+        count_top_agents_solved_for_split(split_name, num_agents)
+    
+    # Get all agents that solved this instance
+    solved_agents = _find_all_agents_for_instance(instance_id, csv_path)
+    if not solved_agents:
+        print(f"No agent found that solved {instance_id} in {split_name} split")
+        return None
+    
+    # Try each agent until we find a trajectory
+    for agent_name in solved_agents:
+        print(f"Trying agent {agent_name} for {instance_id}")
+        # Special handling for JoyCode agent
+        if agent_name == "20250915_JoyCode":
+            trajectory_path = _find_joycode_trajectory_for_split(instance_id, EXPERIMENTS_DIR, split_name)
+        else:
+            trajectory_path = _find_regular_agent_trajectory_for_split(instance_id, agent_name, EXPERIMENTS_DIR, split_name)
+        
+        if trajectory_path:
+            print(f"Found trajectory for {instance_id} using agent {agent_name}")
+            return trajectory_path
+        else:
+            print(f"No trajectory found for {instance_id} using agent {agent_name}")
+    
+    return None
+
+def _find_best_agent_for_instance(instance_id: str, csv_path: str) -> Optional[str]:
+    """
+    Find the best agent that solved the given instance by checking the CSV.
+    """
+    try:
+        with open(csv_path, 'r', encoding='utf-8') as f:
+            reader = csv.DictReader(f)
+            for row in reader:
+                if row['instance_id'] == instance_id:
+                    for col_name in reader.fieldnames:
+                        if col_name.startswith(('01_', '02_', '03_', '04_', '05_', '06_', '07_', '08_', '09_', '10_')) and row.get(col_name) == '1':
+                            agent_name = col_name[3:]
+                            return agent_name
+                    return None
+        return None
+    except Exception as e:
+        print(f"Error reading CSV: {e}")
+        return None
+    return None
+
+def _find_all_agents_for_instance(instance_id: str, csv_path: str) -> List[str]:
+    """
+    Find all agents that solved the given instance by checking the CSV.
+    Returns a list of agent names sorted by their ranking (01_, 02_, etc.).
+    """
+    agents = []
+    try:
+        with open(csv_path, 'r', encoding='utf-8') as f:
+            reader = csv.DictReader(f)
+            for row in reader:
+                if row['instance_id'] == instance_id:
+                    for col_name in reader.fieldnames:
+                        if col_name.startswith(('01_', '02_', '03_', '04_', '05_', '06_', '07_', '08_', '09_', '10_')) and row.get(col_name) == '1':
+                            agent_name = col_name[3:]
+                            agents.append(agent_name)
+                    break
+    except Exception as e:
+        print(f"Error reading CSV: {e}")
+        return []
+    return agents
+
+def _find_joycode_trajectory_for_split(instance_id: str, experiments_dir: str, split_name: str) -> Optional[str]:
+    """
+    Find trajectory for JoyCode agent in a specific split (special case with instance-named folders).
+    """
+    joycode_trajs_dir = os.path.join(experiments_dir, "evaluation", split_name.lower(), "20250915_JoyCode", "trajs")
+    instance_dir = os.path.join(joycode_trajs_dir, instance_id)
+    if not os.path.exists(instance_dir):
+        print(f"JoyCode trajectory directory not found: {instance_dir}")
+        return None
+    completed_logs_path = os.path.join(instance_dir, "completed_logs.txt")
+    if os.path.exists(completed_logs_path):
+        return os.path.abspath(completed_logs_path)
+    else:
+        print(f"completed_logs.txt not found in {instance_dir}")
+        return None
+
+def _find_regular_agent_trajectory_for_split(instance_id: str, agent_name: str, experiments_dir: str, split_name: str) -> Optional[str]:
+    """
+    Find trajectory for regular agents in a specific split (files named with instance_id).
+    """
+    agent_trajs_dir = os.path.join(experiments_dir, "evaluation", split_name.lower(), agent_name, "trajs")
+    if not os.path.exists(agent_trajs_dir):
+        print(f"Agent trajectory directory not found: {agent_trajs_dir}")
+        return None
+    
+    try:
+        for filename in os.listdir(agent_trajs_dir):
+            if instance_id in filename:
+                trajectory_path = os.path.join(agent_trajs_dir, filename)
+                if os.path.isfile(trajectory_path):
+                    return os.path.abspath(trajectory_path)
+    except OSError:
+        pass    
+    return None
+
+def download_single_agent_for_split(agent_info: tuple, experiments_dir: str, split_name: str) -> tuple[str, bool, str]:
+    """
+    Download logs for a single agent in a specific split. Returns (agent_name, success, error_message).
+    """
+    i, agent = agent_info
+    folder = agent.get("folder")
+    name = agent.get("name") or f"rank{i}"
+    
+    if not folder:
+        return (name, False, "No folder specified")
+    
+    # Check if trajectories already exist
+    trajs_path = os.path.join(experiments_dir, "evaluation", split_name.lower(), folder, "trajs")
+    if os.path.exists(trajs_path) and os.listdir(trajs_path):
+        print(f"[{name}] ✓ Trajectories already exist, skipping download")
+        return (name, True, "")
+    
+    download_path = f"evaluation/{split_name.lower()}/{folder}"
+    cmd = [sys.executable, "-m", "analysis.download_logs", "--only_trajs", "--skip_existing", download_path]    
+    try:
+        result = subprocess.run(cmd, cwd=experiments_dir, check=True)
+        print(f"[{name}] ✓ Download completed successfully")
+        return (name, True, "")
+    except Exception as e:
+        error_msg = f"Unexpected error: {e}"
+        print(f"[{name}] ✗ {error_msg}")
+        return (name, False, error_msg)
+
+def download_top_agents_logs_for_split(experiments_dir: str = "experiments", split_name: str = "Verified", num_agents: int = 5) -> None:
+    """
+    Download logs for all top N agents in a specific split using parallel processing.
+    
+    Args:
+        experiments_dir: Directory containing the experiments repository
+        split_name: The split name (e.g., "Verified", "Lite", "Test", "Multimodal")
+        num_agents: Number of top agents to download logs for (default: 5)
+    """
+    ensure_experiments_repo()
+    print("Fetching leaderboard data...")
+    leaderboard_json = read_json_url_or_path(LEADERBOARD_JSON)
+    top_agents = top_agents_for_split(leaderboard_json, split_name, num_agents)
+    
+    print(f"Checking/downloading trajectories for {len(top_agents)} agents in {split_name} split...")
+    
+    agent_infos = [(i, agent) for i, agent in enumerate(top_agents, 1)]
+    
+    with ProcessPoolExecutor(max_workers=10) as executor:
+        future_to_agent = {
+            executor.submit(download_single_agent_for_split, agent_info, experiments_dir, split_name): agent_info[1]
+            for agent_info in agent_infos
+        }
+        
+        successful = 0
+        failed = 0
+        
+        for future in as_completed(future_to_agent):
+            agent = future_to_agent[future]
+            try:
+                name, success, error_msg = future.result()
+                if success:
+                    successful += 1
+                else:
+                    failed += 1
+            except Exception as e:
+                name = agent.get("name", "Unknown")
+                print(f"[{name}] ✗ Exception: {e}")
+                failed += 1
+    print(f"\nDownload completed for {split_name} split. Success: {successful}, Failed: {failed}")

analyzer/utils.py

@@ -0,0 +1,101 @@
+import urllib.request
+import json
+import pathlib
+import os
+import re
+from typing import Optional
+
+
+def read_json_url_or_path(src: str) -> dict:
+    if src.startswith(("http://", "https://")):
+        with urllib.request.urlopen(src, timeout=60) as r:
+            return json.load(r)
+    p = pathlib.Path(src).expanduser().resolve()
+    with p.open("r", encoding="utf-8") as f:
+        return json.load(f)
+
+# an "issues" directory containing "{numerical_id}-{instance_id}/logs/{timestamp}/run/{instance_id}/generator/{generator_id}/" and within it "notice.log" and "debug.log"
+def find_failed_trajectory_in_issues_directory(instance_id: str, experiments_dir: str) -> Optional[str]:
+    """
+    Find a failed trajectory for the given instance within the experiments directory.
+    
+    The function searches for debug.log files that contain successful patches and returns
+    the path to the corresponding notice.log file.
+        
+    The function searches through the directory structure:
+    {experiments_dir}/{numerical_id}-{instance_id}/logs/{timestamp}/run/{instance_id}/generator/{generator_id}/
+    """
+    if not os.path.exists(experiments_dir):
+        return None
+    pattern = re.compile(r'^\d+-' + re.escape(instance_id) + r'$')
+    for item in os.listdir(experiments_dir):
+        if pattern.match(item):
+            instance_dir = os.path.join(experiments_dir, item)
+            if not os.path.isdir(instance_dir):
+                continue
+            logs_dir = os.path.join(instance_dir, 'logs')
+            if not os.path.exists(logs_dir):
+                continue
+            timestamp_dirs = []
+            for timestamp in os.listdir(logs_dir):
+                timestamp_path = os.path.join(logs_dir, timestamp)
+                if os.path.isdir(timestamp_path):
+                    timestamp_dirs.append(timestamp)
+            timestamp_dirs.sort(reverse=True)
+            for timestamp in timestamp_dirs:
+                timestamp_path = os.path.join(logs_dir, timestamp)
+                run_dir = os.path.join(timestamp_path, 'run', instance_id, 'generator')
+                if not os.path.exists(run_dir):
+                    continue
+                for generator_id in range(5):
+                    generator_dir = os.path.join(run_dir, f"{generator_id:03d}")
+                    debug_log_path = os.path.join(generator_dir, 'debug.log')
+                    notice_log_path = os.path.join(generator_dir, 'notice.log')
+                    if not os.path.exists(debug_log_path):
+                        continue                    
+                    if _contains_successful_patch(debug_log_path):
+                        if os.path.exists(notice_log_path):
+                            return os.path.abspath(notice_log_path)
+    return None
+
+def _contains_successful_patch(debug_log_path: str) -> bool:
+    """
+    Check if a debug.log file contains a successful patch.
+    
+    Reads the last non-empty line, splits on 'result_data: ', parses the dict,
+    and checks if the golden_patch contains non-empty patch_content.
+    """
+    try:
+        with open(debug_log_path, 'r', encoding='utf-8') as f:
+            lines = f.readlines()        
+        last_line = None
+        for line in reversed(lines):
+            if line.strip():
+                last_line = line.strip()
+                break
+        if not last_line:
+            return False
+        if 'result_data: ' not in last_line:
+            return False        
+        parts = last_line.split('result_data: ', 1)
+        if len(parts) != 2:
+            return False
+        dict_str = parts[1]        
+        try:
+            result_data = eval(dict_str)
+        except:
+            return False        
+        if not isinstance(result_data, dict):
+            return False        
+        golden_patch = result_data.get('golden_patch')
+        if not isinstance(golden_patch, list) or len(golden_patch) == 0:
+            return False        
+        first_patch = golden_patch[0]
+        if not isinstance(first_patch, dict):
+            return False        
+        patch_content = first_patch.get('patch_content')
+        if not isinstance(patch_content, str) or not patch_content.strip():
+            return False
+        return True
+    except (IOError, UnicodeDecodeError, Exception):
+        return False