AINativeBench / data /processed /RQ1 /BookWriter-H_A2A /analyze_retry_patterns.py

王子睿

restructure + add files

8c10cf2 11 months ago

21.1 kB

#!/usr/bin/env python3
"""
Analyze RETRY patterns in the BookWriter-H_A2A project.
This script computes:
- error locations
- retry counts
- retry rates
Special focus:
- BUSINESS-RETRY: [BUSINESS-RETRY]
- orchestrator retry: (retry N) / [RETRYN]
"""

import os
import re
from pathlib import Path
from collections import defaultdict
import json
import csv

# Base directory and model list
BASE_DIR = Path("/Users/wzr/TOSEM-2025/RESULTS")
MODELS = [
    "DeepSeek-R1",
    "DeepSeek-V3-1",
    "GPT-4o-mini",
    "GPT-5",
    "Gemini-2.5-flash",
    "Gemini-2.5-flash-nothinking",
    "Qwen3-235b",
]
PROJECT_NAME = "BookWriter-H_A2A"


def extract_error_info(line: str) -> dict:
    """Extract error information from a line containing an error marker.
    Returns:
        {'has_error': bool, 'node_type': str, 'node_name': str, 'error_msg': str}
    """
    # Strip tree drawing characters (including ─)
    clean = re.sub(r"^[│├└─\-\s]+", "", line).strip()

    # Check for the error marker
    if "❌" not in clean:
        return {"has_error": False}

    # Remove the error marker
    clean = clean.split("❌", 1)[1].lstrip()

    # Extract node type and name
    node_match = re.match(
        r"\[(SPAN|Chain|AGENT|Tool|LLM)\]\s+([^\[\]]+?)(?:\s+\[ERROR:(.*))?$",
        clean,
    )
    if not node_match:
        return {
            "has_error": True,
            "node_type": "Unknown",
            "node_name": "Unknown",
            "error_msg": "",
        }

    node_type = node_match.group(1)
    node_name = node_match.group(2).strip()
    error_msg = node_match.group(3).strip() if node_match.group(3) else ""

    # Normalize node names
    if node_type == "AGENT":
        node_name = re.sub(r"\._execute_core$", "", node_name)
    elif node_type == "Tool":
        node_name = re.sub(r"\._use$", "", node_name)
    elif node_type == "Chain":
        node_name = re.sub(r"Crew_[a-f0-9\-]+\.kickoff", "Crew***.kickoff", node_name)

    return {
        "has_error": True,
        "node_type": node_type,
        "node_name": node_name,
        "error_msg": error_msg,
    }


def extract_retry_info(line: str) -> dict:
    """Extract retry information from a line (including BUSINESS-RETRY).
    Returns:
        {
            'is_retry': bool,
            'retry_type': str ('orchestrator' or 'business_logic'),
            'retry_number': int (only for orchestrator retries),
            'node_type': str,
            'node_name': str,
            'batch_info': str (BATCH info, if any)
        }
    """
    # Strip tree drawing characters (including ─)
    clean = re.sub(r"^[│├└─\-\s]+", "", line).strip()

    # Check BUSINESS-RETRY: [BUSINESS-RETRY]
    has_business_retry = "[BUSINESS-RETRY]" in clean

    # Check orchestrator retry: (retry N) or [RETRYN]
    orchestrator_retry_match = re.search(r"\(retry\s+(\d+)\)", clean)
    if not orchestrator_retry_match:
        orchestrator_retry_match = re.search(r"\[RETRY(\d+)\]", clean)

    # If neither retry marker exists, return False
    if not has_business_retry and not orchestrator_retry_match:
        return {"is_retry": False}

    # Determine retry type
    if orchestrator_retry_match:
        retry_type = "orchestrator"
        retry_number = int(orchestrator_retry_match.group(1))
    else:
        retry_type = "business_logic"
        retry_number = 0  # BUSINESS-RETRY has no numeric index

    # Extract node type and name
    node_match = re.match(
        r"(?:❌\s*)?\[(SPAN|Chain|AGENT)\]\s+([^\[\]]+?)(?:\s+\[.*)?$",
        clean,
    )
    if not node_match:
        return {
            "is_retry": True,
            "retry_type": retry_type,
            "retry_number": retry_number,
            "node_type": "Unknown",
            "node_name": "Unknown",
            "batch_info": "",
        }

    node_type = node_match.group(1)
    node_name = node_match.group(2).strip()

    # Extract BATCH info (e.g., 📚BATCH1)
    batch_match = re.search(r"📚BATCH(\d+)", clean)
    batch_info = f"BATCH{batch_match.group(1)}" if batch_match else ""

    # Extract chapter title (content inside the trailing parentheses)
    chapter_match = re.search(r"\(([^)]+)\)\s*$", clean)
    if chapter_match:
        chapter_title = chapter_match.group(1)
        batch_info = f"{batch_info} ({chapter_title})" if batch_info else chapter_title

    return {
        "is_retry": True,
        "retry_type": retry_type,
        "retry_number": retry_number,
        "node_type": node_type,
        "node_name": node_name,
        "batch_info": batch_info,
    }


def analyze_session(md_file: str) -> dict:
    """Analyze a single session's `execution_path.md`.
    Returns:
        {
            'has_error': bool,
            'has_retry': bool,
            'has_orchestrator_retry': bool,
            'has_business_retry': bool,
            'max_retry_number': int,
            'total_orchestrator_retries': int,
            'total_business_retries': int,
            'errors': [{'node_type': str, 'node_name': str, 'error_msg': str}, ...],
            'retries': [{
                'retry_type': str,
                'retry_number': int,
                'node_type': str,
                'node_name': str,
                'batch_info': str
            }, ...]
        }
    """
    if not os.path.exists(md_file):
        return None

    with open(md_file, "r", encoding="utf-8") as f:
        content = f.read()

    # Extract the "Execution Path Tree" section
    tree_match = re.search(
        r"## Execution Path Tree.*?```\n(.*?)```", content, re.DOTALL
    )
    if not tree_match:
        return None

    tree_content = tree_match.group(1)

    errors = []
    retries = []

    for line in tree_content.split("\n"):
        if not line.strip():
            continue

        # Check for errors
        error_info = extract_error_info(line)
        if error_info["has_error"]:
            errors.append(
                {
                    "node_type": error_info.get("node_type", "Unknown"),
                    "node_name": error_info.get("node_name", "Unknown"),
                    "error_msg": error_info.get("error_msg", ""),
                }
            )

        # Check for retries (including BUSINESS-RETRY)
        retry_info = extract_retry_info(line)
        if retry_info["is_retry"]:
            retries.append(
                {
                    "retry_type": retry_info["retry_type"],
                    "retry_number": retry_info["retry_number"],
                    "node_type": retry_info["node_type"],
                    "node_name": retry_info["node_name"],
                    "batch_info": retry_info["batch_info"],
                }
            )

    # Aggregate retry types
    orchestrator_retries = [r for r in retries if r["retry_type"] == "orchestrator"]
    business_retries = [r for r in retries if r["retry_type"] == "business_logic"]

    max_retry = (
        max([r["retry_number"] for r in orchestrator_retries])
        if orchestrator_retries
        else 0
    )

    return {
        "has_error": len(errors) > 0,
        "has_retry": len(retries) > 0,
        "has_orchestrator_retry": len(orchestrator_retries) > 0,
        "has_business_retry": len(business_retries) > 0,
        "max_retry_number": max_retry,
        "total_orchestrator_retries": len(orchestrator_retries),
        "total_business_retries": len(business_retries),
        "errors": errors,
        "retries": retries,
    }


def collect_model_stats(model_name: str) -> dict:
    """Collect retry statistics for a single model.
    Returns:
        {
            'model': str,
            'total_sessions': int,
            'sessions_with_error': int,
            'sessions_with_retry': int,
            'sessions_with_orchestrator_retry': int,
            'sessions_with_business_retry': int,
            'total_orchestrator_retry_attempts': int,
            'total_business_retry_attempts': int,
            'error_by_agent': {agent_name: count},
            'error_types': {error_msg: count},
            'max_retry_number': int,
            'business_retry_chapters': {chapter_info: count},
            'session_details': [...]
        }
    """
    test_results_dir = BASE_DIR / model_name / PROJECT_NAME / "test_results"

    if not test_results_dir.exists():
        return None

    stats = {
        "model": model_name,
        "total_sessions": 0,
        "sessions_with_error": 0,
        "sessions_with_retry": 0,
        "sessions_with_orchestrator_retry": 0,
        "sessions_with_business_retry": 0,
        "sessions_with_only_orchestrator_retry": 0,
        "sessions_with_only_business_retry": 0,
        "sessions_with_both_retries": 0,
        "total_orchestrator_retry_attempts": 0,
        "total_business_retry_attempts": 0,
        "error_by_agent": defaultdict(int),
        "error_by_node_type": defaultdict(int),
        "error_types": defaultdict(int),
        "max_retry_number": 0,
        "business_retry_chapters": defaultdict(int),
        "session_details": [],
    }

    for session_dir in sorted(test_results_dir.iterdir()):
        if not session_dir.is_dir():
            continue

        exec_path_file = session_dir / "execution_path.md"
        if not exec_path_file.exists():
            continue

        stats["total_sessions"] += 1

        analysis = analyze_session(str(exec_path_file))
        if not analysis:
            continue

        # Count errors and retries
        if analysis["has_error"]:
            stats["sessions_with_error"] += 1

        if analysis["has_retry"]:
            stats["sessions_with_retry"] += 1

        if analysis["has_orchestrator_retry"]:
            stats["sessions_with_orchestrator_retry"] += 1
            stats["total_orchestrator_retry_attempts"] += analysis[
                "total_orchestrator_retries"
            ]
            stats["max_retry_number"] = max(
                stats["max_retry_number"], analysis["max_retry_number"]
            )

        if analysis["has_business_retry"]:
            stats["sessions_with_business_retry"] += 1
            stats["total_business_retry_attempts"] += analysis["total_business_retries"]

        # Track cases with only one retry type vs. both
        if analysis["has_orchestrator_retry"] and not analysis["has_business_retry"]:
            stats["sessions_with_only_orchestrator_retry"] += 1
        elif analysis["has_business_retry"] and not analysis["has_orchestrator_retry"]:
            stats["sessions_with_only_business_retry"] += 1
        elif analysis["has_orchestrator_retry"] and analysis["has_business_retry"]:
            stats["sessions_with_both_retries"] += 1

        # Aggregate error locations
        for error in analysis["errors"]:
            if error["node_type"] == "AGENT":
                stats["error_by_agent"][error["node_name"]] += 1
            stats["error_by_node_type"][error["node_type"]] += 1

            # Use a short prefix of the error message as the error type key
            error_msg = error["error_msg"]
            if error_msg:
                # Take the first 100 characters as the identifier
                error_type = (
                    error_msg[:100]
                    if len(error_msg) <= 100
                    else error_msg[:100] + "..."
                )
                stats["error_types"][error_type] += 1

        # Aggregate BUSINESS-RETRY chapters
        for retry in analysis["retries"]:
            if retry["retry_type"] == "business_logic" and retry["batch_info"]:
                stats["business_retry_chapters"][retry["batch_info"]] += 1

        # Store session-level details
        stats["session_details"].append(
            {
                "session": session_dir.name,
                "has_error": analysis["has_error"],
                "has_retry": analysis["has_retry"],
                "has_orchestrator_retry": analysis["has_orchestrator_retry"],
                "has_business_retry": analysis["has_business_retry"],
                "orchestrator_retry_count": analysis["total_orchestrator_retries"],
                "business_retry_count": analysis["total_business_retries"],
                "errors": analysis["errors"],
                "retries": analysis["retries"],
            }
        )

    # Compute rates
    total = stats["total_sessions"]
    stats["retry_rate"] = (
        (stats["sessions_with_retry"] / total * 100) if total > 0 else 0
    )
    stats["orchestrator_retry_rate"] = (
        (stats["sessions_with_orchestrator_retry"] / total * 100) if total > 0 else 0
    )
    stats["business_retry_rate"] = (
        (stats["sessions_with_business_retry"] / total * 100) if total > 0 else 0
    )
    stats["error_rate"] = (
        (stats["sessions_with_error"] / total * 100) if total > 0 else 0
    )

    # Convert defaultdict to dict for JSON serialization
    stats["error_by_agent"] = dict(stats["error_by_agent"])
    stats["error_by_node_type"] = dict(stats["error_by_node_type"])
    stats["error_types"] = dict(stats["error_types"])
    stats["business_retry_chapters"] = dict(stats["business_retry_chapters"])

    return stats


def print_summary(all_stats):
    """Print summary statistics."""
    print("\n" + "=" * 120)
    print(f"RETRY pattern analysis summary - {PROJECT_NAME}")
    print("=" * 120 + "\n")

    # Overall summary table
    print("## Model summary\n")
    header = f"{'Model':<30} {'Total':<10} {'Error(%)':<10} {'Retry(%)':<10} {'OrchRetry(%)':<12} {'Business(%)':<12} {'OrchCnt':<12} {'BusinessCnt':<12}"
    print(header)
    print("-" * 120)

    for stats in all_stats:
        if stats:
            print(
                f"{stats['model']:<30} {stats['total_sessions']:<10} "
                f"{stats['error_rate']:>8.1f}% {stats['retry_rate']:>8.1f}% "
                f"{stats['orchestrator_retry_rate']:>10.1f}% {stats['business_retry_rate']:>10.1f}% "
                f"{stats['total_orchestrator_retry_attempts']:<12} {stats['total_business_retry_attempts']:<12}"
            )

    print("\n" + "=" * 120)

    # Per-model details
    for stats in all_stats:
        if not stats:
            continue

        print(f"\n### 📊 {stats['model']}\n")
        print(f"- **Total sessions**: {stats['total_sessions']}")
        print(
            f"- **Sessions with errors**: {stats['sessions_with_error']} ({stats['error_rate']:.1f}%)"
        )
        print(
            f"- **Sessions with retries**: {stats['sessions_with_retry']} ({stats['retry_rate']:.1f}%)"
        )
        print(
            f"  - Orchestrator retry: {stats['sessions_with_orchestrator_retry']} ({stats['orchestrator_retry_rate']:.1f}%) "
            f"[only-orchestrator: {stats['sessions_with_only_orchestrator_retry']}]"
        )
        print(
            f"  - BUSINESS-RETRY: {stats['sessions_with_business_retry']} ({stats['business_retry_rate']:.1f}%) "
            f"[only-business: {stats['sessions_with_only_business_retry']}]"
        )
        print(f"  - Both retry types: {stats['sessions_with_both_retries']}")
        print("- **Total retry attempts**:")
        print(f"  - Orchestrator retry: {stats['total_orchestrator_retry_attempts']}")
        print(f"  - BUSINESS-RETRY: {stats['total_business_retry_attempts']}")
        print(f"- **Max orchestrator retry index**: {stats['max_retry_number']}")

        if stats["error_by_agent"]:
            print("\n**Top agents with errors**:")
            for agent, count in sorted(
                stats["error_by_agent"].items(), key=lambda x: x[1], reverse=True
            )[:5]:
                print(f"  - {agent}: {count}")

        if stats["business_retry_chapters"]:
            print("\n**BUSINESS-RETRY chapters (top 10)**:")
            for chapter, count in sorted(
                stats["business_retry_chapters"].items(),
                key=lambda x: x[1],
                reverse=True,
            )[:10]:
                print(f"  - [{count}] {chapter}")

        if stats["error_types"]:
            print("\n**Top error types (top 3)**:")
            for error_type, count in sorted(
                stats["error_types"].items(), key=lambda x: x[1], reverse=True
            )[:3]:
                print(f"  - [{count}] {error_type}")

        print("\n" + "-" * 120)


def save_results(all_stats):
    """Save results to files."""
    output_dir = Path(__file__).parent

    # Save detailed JSON
    json_file = output_dir / "retry_analysis.json"
    json_data = []
    for stats in all_stats:
        if stats:
            json_data.append(stats)

    with open(json_file, "w", encoding="utf-8") as f:
        json.dump(json_data, f, indent=2, ensure_ascii=False)
    print(f"\n✅ Detailed JSON saved: {json_file}")

    # Save CSV summary
    csv_file = output_dir / "retry_summary.csv"
    with open(csv_file, "w", newline="", encoding="utf-8") as f:
        writer = csv.writer(f)
        writer.writerow(
            [
                "Model",
                "Total_Sessions",
                "Sessions_With_Error",
                "Error_Rate(%)",
                "Sessions_With_Retry",
                "Retry_Rate(%)",
                "Orchestrator_Retry_Sessions",
                "Orchestrator_Retry_Rate(%)",
                "Only_Orchestrator_Retry_Sessions",
                "Business_Retry_Sessions",
                "Business_Retry_Rate(%)",
                "Only_Business_Retry_Sessions",
                "Both_Retries_Sessions",
                "Total_Orchestrator_Retries",
                "Total_Business_Retries",
                "Max_Retry_Number",
            ]
        )

        for stats in all_stats:
            if stats:
                writer.writerow(
                    [
                        stats["model"],
                        stats["total_sessions"],
                        stats["sessions_with_error"],
                        f"{stats['error_rate']:.2f}",
                        stats["sessions_with_retry"],
                        f"{stats['retry_rate']:.2f}",
                        stats["sessions_with_orchestrator_retry"],
                        f"{stats['orchestrator_retry_rate']:.2f}",
                        stats["sessions_with_only_orchestrator_retry"],
                        stats["sessions_with_business_retry"],
                        f"{stats['business_retry_rate']:.2f}",
                        stats["sessions_with_only_business_retry"],
                        stats["sessions_with_both_retries"],
                        stats["total_orchestrator_retry_attempts"],
                        stats["total_business_retry_attempts"],
                        stats["max_retry_number"],
                    ]
                )

    print(f"✅ CSV summary saved: {csv_file}")

    # Save error location stats (by agent)
    error_csv_file = output_dir / "error_by_agent.csv"
    with open(error_csv_file, "w", newline="", encoding="utf-8") as f:
        writer = csv.writer(f)
        writer.writerow(["Model", "Agent_Name", "Error_Count"])

        for stats in all_stats:
            if stats and stats["error_by_agent"]:
                for agent, count in sorted(
                    stats["error_by_agent"].items(), key=lambda x: x[1], reverse=True
                ):
                    writer.writerow([stats["model"], agent, count])

    print(f"✅ Error location stats (by agent) saved: {error_csv_file}")

    # Save BUSINESS-RETRY chapter stats
    chapter_csv_file = output_dir / "business_retry_chapters.csv"
    with open(chapter_csv_file, "w", newline="", encoding="utf-8") as f:
        writer = csv.writer(f)
        writer.writerow(["Model", "Chapter_Info", "Retry_Count"])

        for stats in all_stats:
            if stats and stats["business_retry_chapters"]:
                for chapter, count in sorted(
                    stats["business_retry_chapters"].items(),
                    key=lambda x: x[1],
                    reverse=True,
                ):
                    writer.writerow([stats["model"], chapter, count])

    print(f"✅ BUSINESS-RETRY chapter stats saved: {chapter_csv_file}")


if __name__ == "__main__":
    print(f"Starting RETRY pattern analysis - {PROJECT_NAME}...")
    print("Retry types:")
    print("  1. Orchestrator retry: (retry N) or [RETRYN]")
    print("  2. BUSINESS-RETRY: [BUSINESS-RETRY]\n")

    all_stats = []
    for model in MODELS:
        print(f"\n📊 Analyzing model: {model}")
        stats = collect_model_stats(model)
        if stats:
            all_stats.append(stats)
            print(
                f"  ✅ Done: {stats['total_sessions']} sessions, "
                f"{stats['sessions_with_error']} errors, "
                f"{stats['sessions_with_orchestrator_retry']} orch-retries, "
                f"{stats['sessions_with_business_retry']} business-retries"
            )
        else:
            print("  ⚠️  Skipped (directory not found)")

    print_summary(all_stats)
    save_results(all_stats)

    print("\n" + "=" * 120)
    print("✅ Done!")
    print("=" * 120)

	#!/usr/bin/env python3
	"""
	Analyze RETRY patterns in the BookWriter-H_A2A project.

	This script computes:
	- error locations
	- retry counts
	- retry rates

	Special focus:
	- BUSINESS-RETRY: [BUSINESS-RETRY]
	- orchestrator retry: (retry N) / [RETRYN]
	"""

	import os
	import re
	from pathlib import Path
	from collections import defaultdict
	import json
	import csv

	# Base directory and model list
	BASE_DIR = Path("/Users/wzr/TOSEM-2025/RESULTS")
	MODELS = [
	"DeepSeek-R1",
	"DeepSeek-V3-1",
	"GPT-4o-mini",
	"GPT-5",
	"Gemini-2.5-flash",
	"Gemini-2.5-flash-nothinking",
	"Qwen3-235b",
	]
	PROJECT_NAME = "BookWriter-H_A2A"


	def extract_error_info(line: str) -> dict:
	"""Extract error information from a line containing an error marker.

	Returns:
	{'has_error': bool, 'node_type': str, 'node_name': str, 'error_msg': str}
	"""
	# Strip tree drawing characters (including ─)
	clean = re.sub(r"^[│├└─\-\s]+", "", line).strip()

	# Check for the error marker
	if "❌" not in clean:
	return {"has_error": False}

	# Remove the error marker
	clean = clean.split("❌", 1)[1].lstrip()

	# Extract node type and name
	node_match = re.match(
	r"\[(SPAN\|Chain\|AGENT\|Tool\|LLM)\]\s+([^\[\]]+?)(?:\s+\[ERROR:(.*))?$",
	clean,
	)
	if not node_match:
	return {
	"has_error": True,
	"node_type": "Unknown",
	"node_name": "Unknown",
	"error_msg": "",
	}

	node_type = node_match.group(1)
	node_name = node_match.group(2).strip()
	error_msg = node_match.group(3).strip() if node_match.group(3) else ""

	# Normalize node names
	if node_type == "AGENT":
	node_name = re.sub(r"\._execute_core$", "", node_name)
	elif node_type == "Tool":
	node_name = re.sub(r"\._use$", "", node_name)
	elif node_type == "Chain":
	node_name = re.sub(r"Crew_[a-f0-9\-]+\.kickoff", "Crew***.kickoff", node_name)

	return {
	"has_error": True,
	"node_type": node_type,
	"node_name": node_name,
	"error_msg": error_msg,
	}


	def extract_retry_info(line: str) -> dict:
	"""Extract retry information from a line (including BUSINESS-RETRY).

	Returns:
	{
	'is_retry': bool,
	'retry_type': str ('orchestrator' or 'business_logic'),
	'retry_number': int (only for orchestrator retries),
	'node_type': str,
	'node_name': str,
	'batch_info': str (BATCH info, if any)
	}
	"""
	# Strip tree drawing characters (including ─)
	clean = re.sub(r"^[│├└─\-\s]+", "", line).strip()

	# Check BUSINESS-RETRY: [BUSINESS-RETRY]
	has_business_retry = "[BUSINESS-RETRY]" in clean

	# Check orchestrator retry: (retry N) or [RETRYN]
	orchestrator_retry_match = re.search(r"\(retry\s+(\d+)\)", clean)
	if not orchestrator_retry_match:
	orchestrator_retry_match = re.search(r"\[RETRY(\d+)\]", clean)

	# If neither retry marker exists, return False
	if not has_business_retry and not orchestrator_retry_match:
	return {"is_retry": False}

	# Determine retry type
	if orchestrator_retry_match:
	retry_type = "orchestrator"
	retry_number = int(orchestrator_retry_match.group(1))
	else:
	retry_type = "business_logic"
	retry_number = 0 # BUSINESS-RETRY has no numeric index

	# Extract node type and name
	node_match = re.match(
	r"(?:❌\s)?\[(SPAN\|Chain\|AGENT)\]\s+([^\[\]]+?)(?:\s+\[.)?$",
	clean,
	)
	if not node_match:
	return {
	"is_retry": True,
	"retry_type": retry_type,
	"retry_number": retry_number,
	"node_type": "Unknown",
	"node_name": "Unknown",
	"batch_info": "",
	}

	node_type = node_match.group(1)
	node_name = node_match.group(2).strip()

	# Extract BATCH info (e.g., 📚BATCH1)
	batch_match = re.search(r"📚BATCH(\d+)", clean)
	batch_info = f"BATCH{batch_match.group(1)}" if batch_match else ""

	# Extract chapter title (content inside the trailing parentheses)
	chapter_match = re.search(r"\(([^)]+)\)\s*$", clean)
	if chapter_match:
	chapter_title = chapter_match.group(1)
	batch_info = f"{batch_info} ({chapter_title})" if batch_info else chapter_title

	return {
	"is_retry": True,
	"retry_type": retry_type,
	"retry_number": retry_number,
	"node_type": node_type,
	"node_name": node_name,
	"batch_info": batch_info,
	}


	def analyze_session(md_file: str) -> dict:
	"""Analyze a single session's `execution_path.md`.

	Returns:
	{
	'has_error': bool,
	'has_retry': bool,
	'has_orchestrator_retry': bool,
	'has_business_retry': bool,
	'max_retry_number': int,
	'total_orchestrator_retries': int,
	'total_business_retries': int,
	'errors': [{'node_type': str, 'node_name': str, 'error_msg': str}, ...],
	'retries': [{
	'retry_type': str,
	'retry_number': int,
	'node_type': str,
	'node_name': str,
	'batch_info': str
	}, ...]
	}
	"""
	if not os.path.exists(md_file):
	return None

	with open(md_file, "r", encoding="utf-8") as f:
	content = f.read()

	# Extract the "Execution Path Tree" section
	tree_match = re.search(
	r"## Execution Path Tree.?```\n(.?)```", content, re.DOTALL
	)
	if not tree_match:
	return None

	tree_content = tree_match.group(1)

	errors = []
	retries = []

	for line in tree_content.split("\n"):
	if not line.strip():
	continue

	# Check for errors
	error_info = extract_error_info(line)
	if error_info["has_error"]:
	errors.append(
	{
	"node_type": error_info.get("node_type", "Unknown"),
	"node_name": error_info.get("node_name", "Unknown"),
	"error_msg": error_info.get("error_msg", ""),
	}
	)

	# Check for retries (including BUSINESS-RETRY)
	retry_info = extract_retry_info(line)
	if retry_info["is_retry"]:
	retries.append(
	{
	"retry_type": retry_info["retry_type"],
	"retry_number": retry_info["retry_number"],
	"node_type": retry_info["node_type"],
	"node_name": retry_info["node_name"],
	"batch_info": retry_info["batch_info"],
	}
	)

	# Aggregate retry types
	orchestrator_retries = [r for r in retries if r["retry_type"] == "orchestrator"]
	business_retries = [r for r in retries if r["retry_type"] == "business_logic"]

	max_retry = (
	max([r["retry_number"] for r in orchestrator_retries])
	if orchestrator_retries
	else 0
	)

	return {
	"has_error": len(errors) > 0,
	"has_retry": len(retries) > 0,
	"has_orchestrator_retry": len(orchestrator_retries) > 0,
	"has_business_retry": len(business_retries) > 0,
	"max_retry_number": max_retry,
	"total_orchestrator_retries": len(orchestrator_retries),
	"total_business_retries": len(business_retries),
	"errors": errors,
	"retries": retries,
	}


	def collect_model_stats(model_name: str) -> dict:
	"""Collect retry statistics for a single model.

	Returns:
	{
	'model': str,
	'total_sessions': int,
	'sessions_with_error': int,
	'sessions_with_retry': int,
	'sessions_with_orchestrator_retry': int,
	'sessions_with_business_retry': int,
	'total_orchestrator_retry_attempts': int,
	'total_business_retry_attempts': int,
	'error_by_agent': {agent_name: count},
	'error_types': {error_msg: count},
	'max_retry_number': int,
	'business_retry_chapters': {chapter_info: count},
	'session_details': [...]
	}
	"""
	test_results_dir = BASE_DIR / model_name / PROJECT_NAME / "test_results"

	if not test_results_dir.exists():
	return None

	stats = {
	"model": model_name,
	"total_sessions": 0,
	"sessions_with_error": 0,
	"sessions_with_retry": 0,
	"sessions_with_orchestrator_retry": 0,
	"sessions_with_business_retry": 0,
	"sessions_with_only_orchestrator_retry": 0,
	"sessions_with_only_business_retry": 0,
	"sessions_with_both_retries": 0,
	"total_orchestrator_retry_attempts": 0,
	"total_business_retry_attempts": 0,
	"error_by_agent": defaultdict(int),
	"error_by_node_type": defaultdict(int),
	"error_types": defaultdict(int),
	"max_retry_number": 0,
	"business_retry_chapters": defaultdict(int),
	"session_details": [],
	}

	for session_dir in sorted(test_results_dir.iterdir()):
	if not session_dir.is_dir():
	continue

	exec_path_file = session_dir / "execution_path.md"
	if not exec_path_file.exists():
	continue

	stats["total_sessions"] += 1

	analysis = analyze_session(str(exec_path_file))
	if not analysis:
	continue

	# Count errors and retries
	if analysis["has_error"]:
	stats["sessions_with_error"] += 1

	if analysis["has_retry"]:
	stats["sessions_with_retry"] += 1

	if analysis["has_orchestrator_retry"]:
	stats["sessions_with_orchestrator_retry"] += 1
	stats["total_orchestrator_retry_attempts"] += analysis[
	"total_orchestrator_retries"
	]
	stats["max_retry_number"] = max(
	stats["max_retry_number"], analysis["max_retry_number"]
	)

	if analysis["has_business_retry"]:
	stats["sessions_with_business_retry"] += 1
	stats["total_business_retry_attempts"] += analysis["total_business_retries"]

	# Track cases with only one retry type vs. both
	if analysis["has_orchestrator_retry"] and not analysis["has_business_retry"]:
	stats["sessions_with_only_orchestrator_retry"] += 1
	elif analysis["has_business_retry"] and not analysis["has_orchestrator_retry"]:
	stats["sessions_with_only_business_retry"] += 1
	elif analysis["has_orchestrator_retry"] and analysis["has_business_retry"]:
	stats["sessions_with_both_retries"] += 1

	# Aggregate error locations
	for error in analysis["errors"]:
	if error["node_type"] == "AGENT":
	stats["error_by_agent"][error["node_name"]] += 1
	stats["error_by_node_type"][error["node_type"]] += 1

	# Use a short prefix of the error message as the error type key
	error_msg = error["error_msg"]
	if error_msg:
	# Take the first 100 characters as the identifier
	error_type = (
	error_msg[:100]
	if len(error_msg) <= 100
	else error_msg[:100] + "..."
	)
	stats["error_types"][error_type] += 1

	# Aggregate BUSINESS-RETRY chapters
	for retry in analysis["retries"]:
	if retry["retry_type"] == "business_logic" and retry["batch_info"]:
	stats["business_retry_chapters"][retry["batch_info"]] += 1

	# Store session-level details
	stats["session_details"].append(
	{
	"session": session_dir.name,
	"has_error": analysis["has_error"],
	"has_retry": analysis["has_retry"],
	"has_orchestrator_retry": analysis["has_orchestrator_retry"],
	"has_business_retry": analysis["has_business_retry"],
	"orchestrator_retry_count": analysis["total_orchestrator_retries"],
	"business_retry_count": analysis["total_business_retries"],
	"errors": analysis["errors"],
	"retries": analysis["retries"],
	}
	)

	# Compute rates
	total = stats["total_sessions"]
	stats["retry_rate"] = (
	(stats["sessions_with_retry"] / total * 100) if total > 0 else 0
	)
	stats["orchestrator_retry_rate"] = (
	(stats["sessions_with_orchestrator_retry"] / total * 100) if total > 0 else 0
	)
	stats["business_retry_rate"] = (
	(stats["sessions_with_business_retry"] / total * 100) if total > 0 else 0
	)
	stats["error_rate"] = (
	(stats["sessions_with_error"] / total * 100) if total > 0 else 0
	)

	# Convert defaultdict to dict for JSON serialization
	stats["error_by_agent"] = dict(stats["error_by_agent"])
	stats["error_by_node_type"] = dict(stats["error_by_node_type"])
	stats["error_types"] = dict(stats["error_types"])
	stats["business_retry_chapters"] = dict(stats["business_retry_chapters"])

	return stats


	def print_summary(all_stats):
	"""Print summary statistics."""
	print("\n" + "=" * 120)
	print(f"RETRY pattern analysis summary - {PROJECT_NAME}")
	print("=" * 120 + "\n")

	# Overall summary table
	print("## Model summary\n")
	header = f"{'Model':<30} {'Total':<10} {'Error(%)':<10} {'Retry(%)':<10} {'OrchRetry(%)':<12} {'Business(%)':<12} {'OrchCnt':<12} {'BusinessCnt':<12}"
	print(header)
	print("-" * 120)

	for stats in all_stats:
	if stats:
	print(
	f"{stats['model']:<30} {stats['total_sessions']:<10} "
	f"{stats['error_rate']:>8.1f}% {stats['retry_rate']:>8.1f}% "
	f"{stats['orchestrator_retry_rate']:>10.1f}% {stats['business_retry_rate']:>10.1f}% "
	f"{stats['total_orchestrator_retry_attempts']:<12} {stats['total_business_retry_attempts']:<12}"
	)

	print("\n" + "=" * 120)

	# Per-model details
	for stats in all_stats:
	if not stats:
	continue

	print(f"\n### 📊 {stats['model']}\n")
	print(f"- Total sessions: {stats['total_sessions']}")
	print(
	f"- Sessions with errors: {stats['sessions_with_error']} ({stats['error_rate']:.1f}%)"
	)
	print(
	f"- Sessions with retries: {stats['sessions_with_retry']} ({stats['retry_rate']:.1f}%)"
	)
	print(
	f" - Orchestrator retry: {stats['sessions_with_orchestrator_retry']} ({stats['orchestrator_retry_rate']:.1f}%) "
	f"[only-orchestrator: {stats['sessions_with_only_orchestrator_retry']}]"
	)
	print(
	f" - BUSINESS-RETRY: {stats['sessions_with_business_retry']} ({stats['business_retry_rate']:.1f}%) "
	f"[only-business: {stats['sessions_with_only_business_retry']}]"
	)
	print(f" - Both retry types: {stats['sessions_with_both_retries']}")
	print("- Total retry attempts:")
	print(f" - Orchestrator retry: {stats['total_orchestrator_retry_attempts']}")
	print(f" - BUSINESS-RETRY: {stats['total_business_retry_attempts']}")
	print(f"- Max orchestrator retry index: {stats['max_retry_number']}")

	if stats["error_by_agent"]:
	print("\nTop agents with errors:")
	for agent, count in sorted(
	stats["error_by_agent"].items(), key=lambda x: x[1], reverse=True
	)[:5]:
	print(f" - {agent}: {count}")

	if stats["business_retry_chapters"]:
	print("\nBUSINESS-RETRY chapters (top 10):")
	for chapter, count in sorted(
	stats["business_retry_chapters"].items(),
	key=lambda x: x[1],
	reverse=True,
	)[:10]:
	print(f" - [{count}] {chapter}")

	if stats["error_types"]:
	print("\nTop error types (top 3):")
	for error_type, count in sorted(
	stats["error_types"].items(), key=lambda x: x[1], reverse=True
	)[:3]:
	print(f" - [{count}] {error_type}")

	print("\n" + "-" * 120)


	def save_results(all_stats):
	"""Save results to files."""
	output_dir = Path(__file__).parent

	# Save detailed JSON
	json_file = output_dir / "retry_analysis.json"
	json_data = []
	for stats in all_stats:
	if stats:
	json_data.append(stats)

	with open(json_file, "w", encoding="utf-8") as f:
	json.dump(json_data, f, indent=2, ensure_ascii=False)
	print(f"\n✅ Detailed JSON saved: {json_file}")

	# Save CSV summary
	csv_file = output_dir / "retry_summary.csv"
	with open(csv_file, "w", newline="", encoding="utf-8") as f:
	writer = csv.writer(f)
	writer.writerow(
	[
	"Model",
	"Total_Sessions",
	"Sessions_With_Error",
	"Error_Rate(%)",
	"Sessions_With_Retry",
	"Retry_Rate(%)",
	"Orchestrator_Retry_Sessions",
	"Orchestrator_Retry_Rate(%)",
	"Only_Orchestrator_Retry_Sessions",
	"Business_Retry_Sessions",
	"Business_Retry_Rate(%)",
	"Only_Business_Retry_Sessions",
	"Both_Retries_Sessions",
	"Total_Orchestrator_Retries",
	"Total_Business_Retries",
	"Max_Retry_Number",
	]
	)

	for stats in all_stats:
	if stats:
	writer.writerow(
	[
	stats["model"],
	stats["total_sessions"],
	stats["sessions_with_error"],
	f"{stats['error_rate']:.2f}",
	stats["sessions_with_retry"],
	f"{stats['retry_rate']:.2f}",
	stats["sessions_with_orchestrator_retry"],
	f"{stats['orchestrator_retry_rate']:.2f}",
	stats["sessions_with_only_orchestrator_retry"],
	stats["sessions_with_business_retry"],
	f"{stats['business_retry_rate']:.2f}",
	stats["sessions_with_only_business_retry"],
	stats["sessions_with_both_retries"],
	stats["total_orchestrator_retry_attempts"],
	stats["total_business_retry_attempts"],
	stats["max_retry_number"],
	]
	)

	print(f"✅ CSV summary saved: {csv_file}")

	# Save error location stats (by agent)
	error_csv_file = output_dir / "error_by_agent.csv"
	with open(error_csv_file, "w", newline="", encoding="utf-8") as f:
	writer = csv.writer(f)
	writer.writerow(["Model", "Agent_Name", "Error_Count"])

	for stats in all_stats:
	if stats and stats["error_by_agent"]:
	for agent, count in sorted(
	stats["error_by_agent"].items(), key=lambda x: x[1], reverse=True
	):
	writer.writerow([stats["model"], agent, count])

	print(f"✅ Error location stats (by agent) saved: {error_csv_file}")

	# Save BUSINESS-RETRY chapter stats
	chapter_csv_file = output_dir / "business_retry_chapters.csv"
	with open(chapter_csv_file, "w", newline="", encoding="utf-8") as f:
	writer = csv.writer(f)
	writer.writerow(["Model", "Chapter_Info", "Retry_Count"])

	for stats in all_stats:
	if stats and stats["business_retry_chapters"]:
	for chapter, count in sorted(
	stats["business_retry_chapters"].items(),
	key=lambda x: x[1],
	reverse=True,
	):
	writer.writerow([stats["model"], chapter, count])

	print(f"✅ BUSINESS-RETRY chapter stats saved: {chapter_csv_file}")


	if __name__ == "__main__":
	print(f"Starting RETRY pattern analysis - {PROJECT_NAME}...")
	print("Retry types:")
	print(" 1. Orchestrator retry: (retry N) or [RETRYN]")
	print(" 2. BUSINESS-RETRY: [BUSINESS-RETRY]\n")

	all_stats = []
	for model in MODELS:
	print(f"\n📊 Analyzing model: {model}")
	stats = collect_model_stats(model)
	if stats:
	all_stats.append(stats)
	print(
	f" ✅ Done: {stats['total_sessions']} sessions, "
	f"{stats['sessions_with_error']} errors, "
	f"{stats['sessions_with_orchestrator_retry']} orch-retries, "
	f"{stats['sessions_with_business_retry']} business-retries"
	)
	else:
	print(" ⚠️ Skipped (directory not found)")

	print_summary(all_stats)
	save_results(all_stats)

	print("\n" + "=" * 120)
	print("✅ Done!")
	print("=" * 120)