deepdiver_v2/src/tools/mcp_tools.py

# Copyright (c) 2025 Huawei Technologies Co., Ltd. All rights reserved.
import os
import json
import random
import re
import time
import logging
from pathlib import Path
from typing import Dict, Any, List, Optional

import requests
import dateutil.parser
from concurrent.futures import ThreadPoolExecutor, as_completed
from urllib.parse import urlparse
from bs4 import BeautifulSoup
from dataclasses import dataclass

try:
    from config.config import get_config
except ImportError:
    import sys

    sys.path.append(str(Path(__file__).parent.parent.parent))
    from config.config import get_config

# Workspace management is handled directly in the tools

logger = logging.getLogger(__name__)


@dataclass
class MCPToolResult:
    """Standard result format for MCP tools"""
    success: bool
    data: Any = None
    error: str = None
    metadata: Dict[str, Any] = None

    def to_dict(self) -> Dict[str, Any]:
        return {
            "success": self.success,
            "data": self.data,
            "error": self.error,
            "metadata": self.metadata or {}
        }


class MCPTools:
    """Multi Agent System MCP Tools Implementation"""

    def __init__(self, workspace_path: str = None):
        self.config = get_config()
        self.workspace_path = Path(workspace_path) if workspace_path else Path.cwd()
        self.workspace_path.mkdir(exist_ok=True, parents=True)

        # Session context for workspace-aware operations
        self.session_id = None
        self.session_workspace_path = None
        self.full_workspace_path = os.path.realpath(self.workspace_path)
        if not self.full_workspace_path.endswith(os.sep):
            self.full_workspace_path += os.sep

    def set_session_context(self, session_id: str, session_workspace_path: str):
        """Set session context for workspace-aware operations"""
        self.session_id = session_id
        self.session_workspace_path = Path(session_workspace_path)
        # Update workspace path to session-specific path
        self.workspace_path = self.session_workspace_path
        self.full_workspace_path = os.path.realpath(self.workspace_path)
        if not self.full_workspace_path.endswith(os.sep):
            self.full_workspace_path += os.sep
        logger.info(f"Set session context - ID: {session_id}, Workspace: {session_workspace_path}")

    def get_session_context(self) -> Dict[str, Any]:
        """Get current session context"""
        return {
            "session_id": self.session_id,
            "session_workspace_path": str(self.session_workspace_path) if self.session_workspace_path else None,
            "workspace_path": str(self.workspace_path)
        }

    def _safe_join(self, path: str) -> Path:
        if os.path.isabs(path):
            raise Exception(f"Path '{path}' is absolute. Only relative paths are allowed.")
        joined_path = os.path.join(self.workspace_path, path)
        full_joined_path = os.path.realpath(joined_path)
        if not full_joined_path.startswith(self.full_workspace_path):
            raise Exception(f"Path '{path}' is outside workspace directory.")
        return Path(full_joined_path)

    # ================ WEB SEARCH TOOLS ================

    def batch_web_search(
            self,
            queries: List[str],
            max_results_per_query: int = 15,
            max_workers: int = 5
    ) -> MCPToolResult:
        """
        Batch web search using configurable search provider with concurrent processing.
        
        Users need to implement their own search provider. Below is an example available:
        [
            {
                "query": "search query",
                "search_results": [
                    {
                        "title": "Page title",
                        "link": "https://example.com",
                        "snippet": "Description snippet",
                        "date": "Feb 8, 2022",
                    },
                    ...
                ]
            },
            ...
        ]
        
        Args:
            queries: List of search queries
            max_results_per_query: Maximum search results per query
            max_workers: Maximum number of concurrent search requests
        """
        try:
            from config.config import get_search_engine_config
            search_config = get_search_engine_config()

            if not search_config:
                return MCPToolResult(
                    success=False,
                    error="Search engine not configured"
                )

            # Ensure we never return more than 15 results per query
            actual_max_results = min(max_results_per_query, 15)

            def search_single_query(query: str) -> Dict[str, Any]:
                """Search a single query"""
                try:
                    search_results = self._generic_search(query, actual_max_results, search_config)

                    if not search_results.success:
                        return {
                            'query': query,
                            'success': False,
                            'error': search_results.error,
                            'results': []
                        }

                    # Process search results
                    search_data = search_results.data
                    search_data["organic"] = search_data["organic"][:actual_max_results]

                    return {
                        'query': query,
                        'success': True,
                        'results': search_data,
                        'timestamp': time.time()
                    }

                except Exception as e:
                    logger.error(f"Error searching query '{query}': {e}")
                    return {
                        'query': query,
                        'success': False,
                        'error': str(e),
                        'results': []
                    }

            # Execute searches concurrently
            all_results = []
            with ThreadPoolExecutor(max_workers=min(max_workers, len(queries))) as executor:
                # Submit all search tasks
                future_to_query = {executor.submit(search_single_query, query): query for query in queries}

                # Collect results as they complete
                for future in as_completed(future_to_query):
                    try:
                        result = future.result()
                        all_results.append(result)
                    except Exception as e:
                        query = future_to_query[future]
                        logger.error(f"Error processing search for '{query}': {e}")
                        all_results.append({
                            'query': query,
                            'success': False,
                            'error': str(e),
                            'results': []
                        })

            # Sort results to maintain original query order
            query_order = {query: i for i, query in enumerate(queries)}
            all_results.sort(key=lambda x: query_order.get(x['query'], float('inf')))

            return MCPToolResult(
                success=True,
                data=all_results,
                metadata={
                    'total_queries': len(queries),
                    'successful_queries': len([r for r in all_results if r.get('success', False)]),
                    'concurrent_workers': min(max_workers, len(queries))
                }
            )

        except Exception as e:
            logger.error(f"Batch web search failed: {e}")
            return MCPToolResult(success=False, error=str(e))

    def _generic_search(self, query: str, max_results: int, config: Dict[str, Any]) -> MCPToolResult:
        """
        Generic search function that users need to implement.
        
        This function should return results in the standard format and be wrapped in MCPToolResult:
        
        search_res = {
            "organic": [
                {
                    "title": "Page title",
                    "link": "https://example.com",
                    "snippet": "Description snippet",
                    "date": "Feb 8, 2022"
                }
            ]
        }

        return MCPToolResult(success=True, data=search_res)
        
        Users should implement their own search logic here based on their preferred search service.
        
        Notes:
        1. It is recommended to use search engine APIs that comply with relevant safety and regulatory requirements.
        The user assumes full responsibility for any safety issues, legal consequences, or policy violations
        arising from the use of the search engine results.
        
        2. User requests may be indirectly transmitted to the search engine API in the form of search queries.
        It is the user's responsibility to implement appropriate measures to protect personal privacy and
        sensitive information. We assume no liability for privacy-related issues arising from such transmission.
        """
        try:
            # This is a placeholder - users should implement their own search logic
            raise NotImplementedError(
                "Generic search provider not implemented. Please implement your own search logic in _generic_search method. "
                "The return format should match the standard format with 'organic' results containing title, link, snippet, and date fields."
            )
            
            # Example implementation for serper (commented out):
            # url = config.get('base_url', 'https://api.search-provider.com/search')
            # payload = json.dumps({"q": query, "num": max_results})
            # api_keys = config.get('api_keys', [])
            # headers = {
            #     'X-API-KEY': random.choice(api_keys),
            #     'Content-Type': 'application/json'
            # }

        except Exception as e:
            return MCPToolResult(success=False, error=f"Generic search failed: {e}")

    @staticmethod
    def _extract_publication_date_from_html(url: str) -> Optional[str]:
        """Extract publication date directly from webpage HTML"""
        try:
            # Fetch HTML content
            response = requests.get(url, timeout=10, headers={
                'User-Agent': 'Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36'
            })
            response.raise_for_status()

            soup = BeautifulSoup(response.content, 'html.parser')

            # Common meta tags for publication date
            meta_selectors = [
                'meta[property="article:published_time"]',
                'meta[property="article:modified_time"]',
                'meta[name="date"]',
                'meta[name="pubdate"]',
                'meta[name="published"]',
                'meta[name="datePublished"]',
                'meta[name="publication-date"]',
                'meta[property="og:published_time"]',
                'meta[name="DC.date"]',
                'meta[name="DC.date.created"]',
                'meta[itemprop="datePublished"]',
                'meta[itemprop="dateModified"]'
            ]

            for selector in meta_selectors:
                meta_tag = soup.select_one(selector)
                if meta_tag:
                    content = meta_tag.get('content') or meta_tag.get('datetime')
                    if content:
                        try:
                            # Parse and standardize the date
                            parsed_date = dateutil.parser.parse(content)
                            return parsed_date.isoformat()
                        except ValueError:
                            continue

            # Check for time tags with datetime attribute
            time_tags = soup.find_all('time', {'datetime': True})
            for time_tag in time_tags:
                try:
                    parsed_date = dateutil.parser.parse(time_tag['datetime'])
                    return parsed_date.isoformat()
                except ValueError:
                    continue

            # JSON-LD structured data
            json_ld_scripts = soup.find_all('script', {'type': 'application/ld+json'})
            for script in json_ld_scripts:
                try:
                    data = json.loads(script.string)
                    if isinstance(data, dict):
                        date_fields = ['datePublished', 'dateCreated', 'dateModified']
                        for field in date_fields:
                            if field in data:
                                parsed_date = dateutil.parser.parse(data[field])
                                return parsed_date.isoformat()
                    elif isinstance(data, list):
                        for item in data:
                            if isinstance(item, dict):
                                for field in date_fields:
                                    if field in item:
                                        parsed_date = dateutil.parser.parse(item[field])
                                        return parsed_date.isoformat()
                except ValueError:
                    continue

            return None

        except Exception as e:
            logger.debug(f"Error extracting publication date from {url}: {e}")
            return None

    def _content_extractor(self, url: str, max_tokens: int, config: Dict[str, Any]) -> MCPToolResult:
        """Get content using URL content extractor"""
        max_retry_num = 5
        sleep_time = 5
        retry_num = 0
        while True:
            retry_num += 1
            try:
                api_key = random.choice(config.get('api_keys', ['default_key']))
                headers = {
                    'Authorization': f'Bearer {api_key}',
                    'Content-Type': 'application/json'
                }

                # Users need to implement this - placeholder for custom URL crawler
                raise NotImplementedError(
                    "URL crawler not implemented. Please implement your own URL crawling logic. "
                    "The function should extract text content from URLs and return it in a structured format "
                    "with metadata like title, publication date, and word count."
                )
                
                # Example implementation for content extractor (commented out):
                # crawler_url = f"{crawler_config.get('base_url', 'https://api.content-extractor.com')}/{url}"
                # response = requests.get(crawler_url, headers=headers, timeout=crawler_config.get('timeout', 30))
                # response.raise_for_status()

                # content = response.text

                # # Truncate if needed
                # if max_tokens and len(content.split()) > max_tokens:
                #     words = content.split()[:max_tokens]
                #     content = ' '.join(words) + '...'

                # return MCPToolResult(success=True, data=content)

            except Exception as e:
                if retry_num == max_retry_num:
                    return MCPToolResult(success=False, error=f"Content extractor failed: {e}")
                else:
                    time.sleep(sleep_time)

    def url_crawler(
            self,
            documents: List[Dict],
            max_tokens_per_url: int = 100000,
            include_metadata: bool = True,
            max_workers: int = 10
    ) -> MCPToolResult:
        """
        Extract LLM-friendly content from URLs using configurable crawler service.
        Content is saved to specified file paths.

        Users need to implement their own URL crawler. The return format should include:
        - Extracted text content from the URL
        - Metadata like title, publication date, word count
        - Success/error status for each URL

        Args:
            documents: List of document dictionaries containing:
                - url: Web page URL to extract
                - file_path: Local path to save extracted content
                - title: (Optional) Web page title
                - time: (Optional) Web page publication time
            max_tokens_per_url: Maximum tokens per URL result
            include_metadata: Whether to include metadata about extraction
            max_workers: Maximum number of concurrent extraction requests
        """
        try:
            from config.config import get_url_crawler_config
            crawler_config = get_url_crawler_config()

            if not crawler_config:
                return MCPToolResult(
                    success=False,
                    error="URL crawler not configured"
                )

            def process_single_document(doc: Dict) -> Dict[str, Any]:
                """Process a single document: extract content and save to file"""
                url = doc['url']
                file_path = doc['file_path']
                title = doc.get('title')
                doc_time = doc.get('time')

                result_base = {
                    'url': url,
                    'file_path': file_path,
                    'title': title,
                    'time': doc_time,
                    'success': False,
                    'error': None,
                    'content_length': 0,
                    'word_count': 0,
                    'publication_date': None,
                    'extraction_timestamp': time.time(),
                    'write_success': False
                }

                try:
                    # Extract publication date from the webpage
                    publication_date = self._extract_publication_date_from_html(url)
                    result_base["publication_date"] = publication_date

                    # Extract content using content extractor
                    content_result = self._content_extractor(url, max_tokens_per_url, crawler_config)

                    if not content_result.success:
                        result_base['error'] = content_result.error
                        return result_base

                    content = content_result.data
                    if not content:
                        result_base['error'] = "Extracted content is empty"
                        return result_base

                    # Save content to file
                    write_result = self.file_write(
                        file_path=file_path,
                        content=content,
                        create_dirs=True
                    )

                    if not write_result.success:
                        result_base['error'] = f"File write failed: {write_result.error}"
                        return result_base

                    # Build success result
                    result = {
                        **result_base,
                        'success': True,
                        'content_length': len(content),
                        'word_count': len(content.split()),
                        'publication_date': publication_date,
                        'write_success': True
                    }

                    if include_metadata:
                        result['metadata'] = {
                            'truncated': len(content.split()) >= max_tokens_per_url,
                            'has_publication_date': publication_date is not None,
                            'date_extraction_method': 'html_parsing' if publication_date else None,
                            'file_size': len(content.encode('utf-8'))
                        }

                    return result

                except Exception as e:
                    logger.error(f"Error processing document {url}: {e}")
                    # Try to extract publication date even if processing failed
                    try:
                        publication_date = self._extract_publication_date_from_html(url)
                    except:
                        publication_date = None

                    return {
                        **result_base,
                        'error': str(e),
                        'publication_date': publication_date
                    }

            # Execute processing concurrently
            results = []
            with ThreadPoolExecutor(max_workers=min(max_workers, len(documents))) as executor:
                # Submit all processing tasks
                future_to_doc = {executor.submit(process_single_document, doc): doc for doc in documents}

                # Collect results as they complete
                for future in as_completed(future_to_doc):
                    doc = future_to_doc[future]
                    try:
                        result = future.result()
                        results.append(result)
                    except Exception as e:
                        url = doc['url']
                        logger.error(f"Error processing extraction for '{url}': {e}")

                        # Try to extract publication date even if processing failed
                        try:
                            publication_date = self._extract_publication_date_from_html(url)
                        except:
                            publication_date = None

                        results.append({
                            'url': url,
                            'file_path': doc['file_path'],
                            'title': doc.get('title'),
                            'time': doc.get('time'),
                            'success': False,
                            'error': str(e),
                            'publication_date': publication_date,
                            'extraction_timestamp': time.time(),
                            'write_success': False
                        })

            # Sort results to maintain original document order
            url_order = {doc['url']: i for i, doc in enumerate(documents)}
            results.sort(key=lambda x: url_order.get(x['url'], float('inf')))

            successful_extractions = len([r for r in results if r.get('success', False)])
            successful_writes = len([r for r in results if r.get('write_success', False)])

            return MCPToolResult(
                success=True,
                data=results,
                metadata={
                    'total_documents': len(documents),
                    'successful_extractions': successful_extractions,
                    'successful_writes': successful_writes,
                    'failed_processing': len(documents) - successful_extractions,
                    'concurrent_workers': min(max_workers, len(documents))
                }
            )

        except Exception as e:
            logger.error(f"URL crawler batch processing failed: {e}")
            return MCPToolResult(success=False, error=str(e))
        
    def file_read_dq(self, file_path: str, encoding: str = 'utf-8') -> MCPToolResult:
        """Read file content"""
        try:
            full_path = self._safe_join(file_path)
            
            if not full_path.exists():
                return MCPToolResult(
                    success=False,
                    error=f"File does not exist: {file_path}"
                )
            
            content = full_path.read_text(encoding=encoding)
            if len(content) > 40000:
                content = (
                    "Due to the content being too long, only the first 30,000 and last 10,000 characters are returned.\n"
                    "Below is the returned portion of the file content:\n\n"
                    f"First 30,000 characters:\n\n{content[:30000]}\n\n"
                    f"Last 10,000 characters:\n\n{content[-10000:]}"
                )
            
            return MCPToolResult(
                success=True,
                data=content,
                metadata={
                    'file_size': len(content),
                    'line_count': len(content.splitlines()),
                    'encoding': encoding
                }
            )
            
        except Exception as e:
            logger.error(f"File read failed: {e}")
            return MCPToolResult(success=False, error=str(e))
        
    def load_json(self, file_path: str, encoding: str = 'utf-8') -> MCPToolResult:
        """
        Read JSON format file 
        """
        try:
            full_path = self._safe_join(file_path)
            
            if not full_path.exists():
                return MCPToolResult(
                    success=False,
                    error=f"File does not exist: {file_path}"
                )
            
            res = []

            with open(full_path, "r", encoding=encoding, errors='ignore') as f:
                for idx, line in enumerate(f):
                    try:
                        ele = json.loads(line.strip())
                        res.append(ele)
                    except Exception as e:
                        logger.warning(f"Failed to process file: {e}")
                        continue

            return MCPToolResult(
                success=True,
                data=res,
                metadata={
                    'line_count': len(res),
                    'encoding': encoding
                }
            )
            
        except Exception as e:
            logger.error(f"File read failed: {e}")
            return MCPToolResult(success=False, error=str(e))

    def document_qa(
                self,
                tasks: List[Dict],
                model: str = "pangu_auto",
                temperature: Optional[float] = None,
                max_tokens: Optional[int] = None,
                max_workers: int = 5
            ) -> MCPToolResult:
                """
                Answer questions based on content stored in local files. 
                Each task contains a file path and a question to be answered using that file's content.
                
                Args:
                    tasks: List of task dictionaries containing:
                        - file_path: Relative path to the file (relative to workspace root) to read
                        - question: Question to ask about this file
                    model: AI model to use for generating answers
                    temperature: Creativity level for the AI response (0-1)
                    max_tokens: Maximum tokens for the AI response
                    max_workers: Maximum number of concurrent model API requests
                """
                try:
                    # Get configuration
                    from config.config import get_model_config, get_storage_config
                    model_config = get_model_config()
                    storage_config = get_storage_config()
                    
                    # Use config values or defaults
                    if temperature is None:
                        temperature = model_config.get('temperature', 0.3)
                    if max_tokens is None:
                        max_tokens = model_config.get('max_tokens', 8192)

                    logger.debug(f"Starting document QA processing: {tasks}")
                    # 处理单个任务
                    def process_single_task(task: Dict) -> Dict:
                        file_path = task['file_path']
                        question = task['question']
                        
                        # 1. 读取文件内容
                        read_result = self.file_read_dq(file_path)
                        if not read_result.success:
                            return {
                                'file_path': file_path,
                                'question': question,
                                'success': False,
                                'error': f"File read error: {read_result.error}",
                                'answer': None
                            }
                        
                        content = read_result.data
                        
                        # 2. 构建系统提示
                        system_prompt = (
                            "You are an expert document analyst. Answer the user's question "
                            "based ONLY on the provided context. If the answer cannot be found "
                            "in the context, say 'I don't know'. Answer in the same language as the user's question. "
                            "If the document section you reference includes citation sources, such as URLs, you must also include the original citation sources from the document in your response.\n\n"
                            "CONTEXT:\n{context}"
                        ).format(context=content)
                        
                        # 3. 调用大模型API
                        try:
                            # Get model URL and token from config
                            # Prefer config over environment variables, with hardcoded fallback
                            model_url = model_config.get('url') or os.getenv('MODEL_REQUEST_URL', '')
                            model_token = model_config.get('token') or os.getenv('MODEL_REQUEST_TOKEN', '')
                            headers = {'Content-Type': 'application/json', 'csb-token': model_token}

                            response = requests.post(
                                url=model_url,
                                headers=headers,
                                json={
                                    "model": model_config.get('model', 'pangu_auto'),
                                    "chat_template": "{% for message in messages %}{% if loop.first and messages[0]['role'] != 'system' %}{{ '<s>[unused9]系统：[unused10]' }}{% endif %}{% if message['role'] == 'system' %}{{'<s>[unused9]系统：' + message['content'] + '[unused10]'}}{% endif %}{% if message['role'] == 'assistant' %}{{'[unused9]助手：' + message['content'] + '[unused10]'}}{% endif %}{% if message['role'] == 'tool' %}{{'[unused9]工具：' + message['content'] + '[unused10]'}}{% endif %}{% if message['role'] == 'function' %}{{'[unused9]方法：' + message['content'] + '[unused10]'}}{% endif %}{% if message['role'] == 'user' %}{{'[unused9]用户：' + message['content'] + '[unused10]'}}{% endif %}{% endfor %}{% if add_generation_prompt %}{{ '[unused9]助手：' }}{% endif %}",
                                    "messages": [
                                        {"role": "system", "content": system_prompt},
                                        {"role": "user", "content": question+" /no_think"}
                                    ],
                                    "spaces_between_special_tokens": False,
                                    "temperature": temperature,
                                    "max_tokens": max_tokens
                                },
                                timeout=model_config.get("timeout", 180)
                            )
                            response = response.json()
                            answer = response["choices"][0]["message"]["content"].split("[unused17]")[-1]

                            return {
                                'file_path': file_path,
                                'question': question,
                                'success': True,
                                'answer': answer,
                                'metadata': {
                                    'file_size': len(content),
                                    'line_count': len(content.splitlines())
                                }
                            }
                            
                        except Exception as e:
                            logger.error(f"Model API call failed for file '{file_path}': {e}")
                            return {
                                'file_path': file_path,
                                'question': question,
                                'success': False,
                                'error': f"Model API error: {str(e)}"
                            }
                    
                    # 4. 并发处理所有任务
                    results = []
                    with ThreadPoolExecutor(max_workers=min(max_workers, len(tasks))) as executor:
                        future_to_task = {executor.submit(process_single_task, task): task for task in tasks}
                        
                        for future in as_completed(future_to_task):
                            try:
                                result = future.result()
                                results.append(result)
                            except Exception as e:
                                task = future_to_task[future]
                                logger.error(f"Task processing failed for file '{task['file_path']}': {e}")
                                results.append({
                                    'file_path': task['file_path'],
                                    'question': task['question'],
                                    'success': False,
                                    'error': f"Task processing exception: {str(e)}"
                                })
                    
                    # 5. 保持原始任务顺序
                    task_order = {task['file_path']: i for i, task in enumerate(tasks)}
                    results.sort(key=lambda x: task_order.get(x['file_path'], float('inf')))
                    
                    # 6. 统计结果
                    successful_tasks = len([r for r in results if r.get('success', False)])
                    
                    return MCPToolResult(
                        success=True,
                        data=results,
                        metadata={
                            'total_tasks': len(tasks),
                            'successful_tasks': successful_tasks,
                            'failed_tasks': len(tasks) - successful_tasks,
                            'concurrent_workers': min(max_workers, len(tasks))
                        }
                    )
                    
                except Exception as e:
                    logger.error(f"Context-based QA batch processing failed: {e}")
                    return MCPToolResult(success=False, error=str(e))

    def search_result_classifier(
            self,
            # save_analysis_file_path: str,
            outline: str,
            key_files: List[Dict],
            model: str = "pangu_auto",
            temperature: Optional[float] = None,
            max_tokens: Optional[int] = None
    ) -> MCPToolResult:
        """
            Classify and organize search result files according to a structured outline for comprehensive long-form content generation.

            Args:
                outline: Structured outline defining the sections and subsections for organizing the long-form content
                key_files: List of key files to classify
                model: AI model to use for classification and organization
                temperature: Creativity level for the AI classification (0-1)
                max_tokens: Maximum tokens for the AI response
            """
        try:
            # Get configuration
            from config.config import get_model_config, get_storage_config
            model_config = get_model_config()
            storage_config = get_storage_config()
            
            # Use config values or defaults
            if temperature is None:
                temperature = model_config.get('temperature', 0.3)
            if max_tokens is None:
                max_tokens = model_config.get('max_tokens', 2000)
                
            logger.debug(f"Starting search result classification: outline={outline}, key_files={key_files}")

            key_files_dict = {}
            # Create full path relative to workspace using config
            analysis_path = storage_config.get('document_analysis_path', './doc_analysis')
            file_analysis_list = self.load_json(f"{analysis_path}/file_analysis.jsonl").data

            for file_info in file_analysis_list:
                if file_info.get('file_path'):
                    key_files_dict[file_info.get('file_path')] = file_info

            prompt_files = ""
            if key_files:
                prompt_files += f"Key Files with Multi-Dimensional Analysis:\n"
                for i, file_info in enumerate(key_files, 1):
                    if file_info.get('file_path') in key_files_dict:
                        file_info = key_files_dict[file_info.get('file_path')]
                        prompt_files += f"\n{i}. File: {file_info.get('file_path', 'Unknown')}\n"
                        prompt_files += f"   Document Time: {file_info.get('doc_time', 'Not specified')}\n"
                        prompt_files += f"   Source Authority: {file_info.get('source_authority', 'Not specified')}\n"
                        prompt_files += f"   Core Content: {file_info.get('core_content', 'Not specified')}\n"
                        prompt_files += f"   Task Relevance: {file_info.get('task_relevance', 'Not specified')}\n"
                        prompt_files += f"   Information Richness: {file_info.get('information_richness', 'Not specified')}\n"
                        prompt_files += "\n"

                prompt_files += "\n"

            system_prompt = (
                "You are an expert content organizer specializing in multi-dimensional file classification. "
                "Your task is to classify research files according to a given outline by according the five key dimensions.\n\n"

                "FIVE KEY DIMENSIONS:\n"
                "1. DOCUMENT TIME: Consider the temporal relevance and recency of the source\n"
                "2. SOURCE AUTHORITY: Consider the credibility and expertise of the source\n"
                "3. CORE CONTENT: Focus on the main themes and key insights\n"
                "4. TASK RELEVANCE: Assess alignment with the specific research goals\n"
                "5. INFORMATION RICHNESS: Determine the information richness of the document\n"

                "CLASSIFICATION INSTRUCTIONS:\n"
                "1. Read the outline and understand its structure\n"
                "2. Analyze each file across the five dimensions provided\n"
                "3. Accept files with moderate source authority, moderate information richness and reasonable task relevance\n"
                "4. Use core content insights to find connections to outline sections, ensure the relevance of the assignment.\n"
                "5. Files should be assigned to multiple sections when they contain information spanning different topics\n"
                "6. Prioritize comprehensive coverage: Ensure every paragraph/chapter gets file assignments when any relevant content exists\n"
                "7.Note that if the first chapter is an abstract or introduction, a corresponding file must be assigned and there cannot be an empty file.\n"

                "CRITICAL REQUIREMENT - OUTLINE STRUCTURE PRESERVATION:\n"
                "- The number of files assigned to each chapter cannot exceed 11, so if the number of files exceeds 11, you need the most relevant files.\n"
                "- When the outline is divided into paragraphs/sections, you must preserve the original content and structure exactly.\n "
                "- Do NOT modify, rephrase, or alter any paragraph titles, headings, or structural elements from the original outline. \n"
                "- Keep all outline content completely intact in your output, including formatting and wording.\n"
                "- You should split each paragraph according to the granularity of the first-level heading. You are not allowed to separate any second-level headings or smaller headings under the first-level heading into individual paragraphs.\n"
                "- Formatting requirements for split paragraphs: 1. You are not allowed to modify the format of the paragraphs in the original outline, including all bold symbols, etc.; 2. If the given outline includes the title of the article, you must include the article title in the first paragraph, maintaining its original bold symbols.\n"

                "Strictly follow the following format for output:\n"
                "paragraph 1: ...\n"
                "file_path_list: file_path1, file_path2, ...\n"
                "\n"
                "paragraph 2: ...\n"
                "file_path_list: file_path3, file_path4, ...\n"
                "..."
            )

            # 构建用户提示 - 仅包含输入数据
            user_prompt = f"""
    OUTLINE TO ORGANIZE CONTENT:
    {outline}

    {prompt_files}
    """

            # 调用AI模型进行分类
            # Get model URL and token from config
            config = get_config()
            model_config = config.get_custom_llm_config()
            
            model_url = model_config.get('url') or os.getenv('MODEL_REQUEST_URL', '')
            model_token = model_config.get('token') or os.getenv('MODEL_REQUEST_TOKEN', '')
            headers = {'Content-Type': 'application/json', 'csb-token': model_token}

            try:
                # Add retry logic for AI model call
                max_retries = 5
                response = None

                for attempt in range(max_retries):
                    try:
                        response = requests.post(
                            url=model_url,
                            headers=headers,
                            json={
                                "model": model_config.get('model', 'pangu_auto'),
                                "chat_template": "{% for message in messages %}{% if loop.first and messages[0]['role'] != 'system' %}{{ '<s>[unused9]系统：[unused10]' }}{% endif %}{% if message['role'] == 'system' %}{{'<s>[unused9]系统：' + message['content'] + '[unused10]'}}{% endif %}{% if message['role'] == 'assistant' %}{{'[unused9]助手：' + message['content'] + '[unused10]'}}{% endif %}{% if message['role'] == 'tool' %}{{'[unused9]工具：' + message['content'] + '[unused10]'}}{% endif %}{% if message['role'] == 'function' %}{{'[unused9]方法：' + message['content'] + '[unused10]'}}{% endif %}{% if message['role'] == 'user' %}{{'[unused9]用户：' + message['content'] + '[unused10]'}}{% endif %}{% endfor %}{% if add_generation_prompt %}{{ '[unused9]助手：' }}{% endif %}",
                                "messages": [
                                    {"role": "system", "content": system_prompt},
                                    {"role": "user", "content": user_prompt + " /no_think"}
                                ],
                                "spaces_between_special_tokens": False,
                                "max_tokens": max_tokens,
                                "temperature": temperature,
                            },
                            timeout=model_config.get("timeout", 180)
                        )
                        response = response.json()
                        logger.debug(f"API response received")

                        break  # Success, exit retry loop
                    except Exception as e:
                        logger.warning(f"LLM API call attempt {attempt + 1} failed: {e}")
                        if attempt == max_retries - 1:
                            raise e  # Last attempt, re-raise the exception
                        time.sleep(5)  # Simple 1 second delay between retries

                if response is None:
                    raise Exception("Failed to get response after all retries")

                # ai_response = response.choices[0].message.content.strip()
                ai_response = response["choices"][0]["message"]["content"].strip()

                session_context = self.get_session_context()
                session_id = session_context.get("session_id")

                # 切换保存的方式
                conversation_history = [
                    # {"role": "system", "content": "system 1"},
                    {"role": "system", "content": system_prompt},
                    {"role": "user", "content": user_prompt + " /no_think"},
                    {"role": "assistant", "content": ai_response}
                ]

                return MCPToolResult(
                    success=True,
                    data=ai_response.split('think>')[-1].strip(),
                )

            except Exception as e:
                logger.error(f"AI model call failed: {e}")
                return MCPToolResult(
                    success=False,
                    error=f"AI classification failed: {str(e)}"
                )

        except Exception as e:
            logger.error(f"Search result classifier failed: {e}")
            return MCPToolResult(success=False, error=str(e))

        # 实现section writer的具体内容

    @staticmethod
    def _correct_title_format(content: str, overall_outline: str) -> str:
        """
        Correct title formats in content to match those in overall_outline.

        Args:
            content: The generated chapter content
            overall_outline: The overall outline containing correct title formats

        Returns:
            Content with corrected title formats
        """
        # Extract titles from overall_outline
        outline_titles = {}

        for line in overall_outline.split('\n'):
            line = line.strip()
            if line:
                # Extract core title content by removing various formatting symbols
                core_content = line

                # Remove leading symbols like **, -, etc.
                core_content = re.sub(r'^[\*\-\s]+', '', core_content)
                # Remove trailing symbols like **
                core_content = re.sub(r'[\*\s]+$', '', core_content)
                core_content = core_content.strip()

                if core_content:
                    # Store mapping from core content to the formatted line from outline
                    outline_titles[core_content.lower()] = line

        # Process content line by line
        content_lines = content.split('\n')
        corrected_lines = []

        for line in content_lines:
            original_line = line
            line_stripped = line.strip()

            # Check if this line is a title (starts with # and typically has ** formatting)
            if line_stripped and re.match(r'^#+\s*[\*]*.*', line_stripped):
                # Extract core content from the title
                core_content = line_stripped

                # Remove markdown headers (#)
                core_content = re.sub(r'^#+\s*', '', core_content)
                # Remove ** formatting
                core_content = re.sub(r'^\*\*', '', core_content)
                core_content = re.sub(r'\*\*$', '', core_content)
                core_content = core_content.strip()

                # Look for exact matching title in outline_titles
                found_match = False
                core_content_lower = core_content.lower()

                for outline_core, outline_format in outline_titles.items():
                    if outline_core == core_content_lower:
                        # Replace with the correct format from overall_outline
                        corrected_lines.append(outline_format)
                        found_match = True
                        break

                if not found_match:
                    # If no exact match found, keep original line
                    corrected_lines.append(original_line)
            else:
                corrected_lines.append(original_line)

        return '\n'.join(corrected_lines)

    def section_writer(
            self,
            written_chapters_summary: str,
            task_content: str,
            user_query: str,
            current_chapter_outline: str,
            overall_outline: str,
            target_file_path: str,
            key_files: List[Dict],
            model: str = "pangu_auto",
            temperature: Optional[float] = None,
            max_tokens: Optional[int] = None
    ) -> MCPToolResult:
        """
        Write the current chapter content based on given web information and chapter structure; also consider user questions, completed chapters, and overall outline to ensure content relevance while avoiding duplication or contradictions.

        Args:
            user_query: The user query, ensure the drafted content is highly relevant to the user's inquiry.
            current_chapter_outline: This field represents the current chapter structure to be drafted. When composing the chapter content, do not modify content and bold formatting symbols of the existing structure's titles!!!
            overall_outline: This field represents the overall outline of the article. When drafting the chapter content, you should consider the overall outline to ensure the chapter content is consistent with the overall outline.
            target_file_path: The path to save the chapter content
            key_files: These files are the source materials required for drafting the current chapter.
            model: AI model to use for writing the chapter content
            temperature: Creativity level for the AI response (0-1)
            max_tokens: Maximum tokens for the AI response
        """
        try:
            # Get configuration
            from config.config import get_model_config, get_storage_config
            model_config = get_model_config()
            storage_config = get_storage_config()
            
            # Use config values or defaults
            if temperature is None:
                temperature = model_config.get('temperature', 0.3)
            if max_tokens is None:
                max_tokens = model_config.get('max_tokens', 8192)
                

            key_files_dict = {}
            # Create full path relative to workspace using config
            analysis_path = storage_config.get('document_analysis_path', './doc_analysis')
            file_analysis_list = self.load_json(f"{analysis_path}/file_analysis.jsonl").data
            logger.debug("File analysis loaded successfully")

            for i, file_info in enumerate(file_analysis_list, 1):
                file_info['index'] = i  # 在这里给网页进行编号， 从1开始
                if file_info.get('file_path'):
                    key_files_dict[file_info.get('file_path')] = file_info

            prompt_files = ""
            if key_files:
                prompt_files += f"Web Information Source(s) As Follows::\n"
                for i, file_info in enumerate(key_files, 1):
                    if file_info.get('file_path') in key_files_dict:
                        file_info = key_files_dict[file_info.get('file_path')]
                        index = file_info.get('index')
                        # 通过file_path获取原始网页信息
                        file_path = file_info.get('file_path')

                        def get_file_head_content(file_path, max_length=10000):
                            try:
                                full_path = self._safe_join(file_path)
                                if not full_path.exists():
                                    return f"[Error: File does not exist: {file_path}]"
                                with open(full_path, 'r', encoding='utf-8', errors='ignore') as f:
                                    content = f.read(max_length)
                                return content
                            except Exception as e:
                                return f"[Error reading file {file_path}: {str(e)}]"

                        file_content = get_file_head_content(file_path)
                        doc_time = file_info.get('doc_time', 'Not specified')
                        source_authority = file_info.get('source_authority', 'Not specified')
                        task_relevance = file_info.get('task_relevance', 'Not specified')
                        # 开始组装这些字段的值
                        prompt_files += f"\n[webpaeg{index} begin]网页时间: {doc_time}|||网页权威性：{source_authority}|||网页相关性：{task_relevance}|||网页内容：{file_content}[webpaeg{index} end]"

            # 设计system prompt
            system_prompt = """You are a writing master. Next, you will receive web page information, user questions, and the structure of the current chapter. You need to integrate the user's questions with the provided web content and write the chapter based on its given structure. Additionally, an overall outline and summaries of previously completed chapters will be provided for reference to avoid repetition or contradictions and ensure logical consistency within the broader framework. Specific requirements will be detailed below.

When drafting the current chapter content, strictly comply with the following requirements:
- In the web page information I gave you, each result is in the format of [webpage X begin]...[webpage X end], where X represents the numerical index of each article. Please cite the context at the end of the sentence when appropriate. Please cite the context in the corresponding part of the answer in the format of the reference number [citation:X]. If a sentence comes from multiple contexts, please list all relevant reference numbers, such as [citation:3][citation:5]. Remember not to collect the references at the end and return the reference numbers, but list them in the corresponding part of the answer.
- You can only use the provided web page information for writing, don't make up any content, ensure the accuracy of the facts. Note that when there are contradictions between the facts described in the above search results, you should use your internal knowledge to reasonably identify the correct information. If identification is impossible, you may select the most factual result based on the authority of the web pages and a voting mechanism (e.g., the description consistent with the majority of web pages). If judgment remains impossible using these methods, you may appropriately list possible differing statements, but you must not conflate different claims—prioritize ensuring factual accuracy!
- You are only permitted to write content strictly within the provided chapter framework. You are forbidden from creating additional subheadings or bullet points within the framework! However, there is a special exception: You may appropriately use tables for narration when necessary. Furthermore, you are not allowed to use concise or summarizing language for narration! We must strictly ensure the information density of the writing and avoid excessive compression.
- You cannot make any changes to the structure of the chapter you are currently writing, such as the title content and the bold symbols in the title, you are not allowed to make any changes. **Important Note:** When writing Chapter 1, if you find the chapter lacks article title, you must create one based on user query. However, this rule only applies to Chapter 1 - do not add any titles to any other chapters in the work. 
- Be careful to ensure that the narrative content is highly relevant and does not contain any common sense errors, note that although you are asked to ensure the richness of information when writing, you must ensure that the content you write is highly relevant and that the context is logically coherent and readable.
- Proceeding to explain the roles of other specified fields:
    * user_query: The user query, ensure the drafted content is highly relevant to the user's inquiry.
    * written_chapters: Reference written_chapters to avoid large amounts of repetitive or conflicting content
    * overall_outline: The purpose of giving an overall outline is to let you understand the summary of the article and avoid content inconsistent with other parts during your writing. In short, focus on writing the current chapter.
    * task_content: The task_content may provide the requirements for writing the current chapter as well as prompts for what to avoid. You can refer to this content when drafting.

Other points to note::
- If the first chapter is an **Abstract** or **Introduction**, do not include subheadings (level-2 or finer bullet points)—begin the content directly under the level-1 heading.  
- CONTENT LENGTH: Each section should contain approximately 2500 words to ensure comprehensive coverage.
- **CRITICAL TITLE PRESERVATION RULE:** You MUST preserve the exact format, structure, and content of chapter titles as provided in the current_chapter_outline. This includes:
  * DO NOT change any markdown formatting symbols (# ## ### ** etc.)
  * DO NOT add, remove, or rearrange any part of the title structure
  * Copy the title lines EXACTLY as they appear in current_chapter_outline
  * Only write content under the provided title structure - never modify the titles themselves
  * When the title symbols in the current chapter outline are inconsistent with those in the overall outline, use the overall outline's title symbols as the standard and maintain symbol consistency throughout the writing process
- Note that in Chapter 1, omit any mention of research objectives, methodology, or procedural details.
- Be sure to ensure that the language of your output is consistent with the language of the user's question. For example, if the user's question is in Chinese, your reply should also be in Chinese.

Strictly follow the following format for output:
<chapter_content>xxx</chapter_content>
"""

            user_prompt = f"""TASK CONTENT: {task_content}
    WEB PAGE INFORMATION: {prompt_files}
    OVERALL OUTLINE: {overall_outline}
    CURRENT CHAPTER OUTLINE: {current_chapter_outline}
    PREVIOUSLY WRITTEN CHAPTERS SUMMARY: {written_chapters_summary}
    USER QUERY: {user_query}"""

            # 调用AI模型进行分类
            # Get model URL and token from config
            config = get_config()
            model_config = config.get_custom_llm_config()
            
            model_url = model_config.get('url') or os.getenv('MODEL_REQUEST_URL', '')
            model_token = model_config.get('token') or os.getenv('MODEL_REQUEST_TOKEN', '')
            headers = {'Content-Type': 'application/json', 'csb-token': model_token}
            try:
                max_retries = 5
                response = None
                for attempt in range(max_retries):
                    try:
                        response = requests.post(
                            url=model_url,
                            headers=headers,
                            json={
                                "model": model_config.get('model', 'pangu_auto'),
                                "chat_template": "{% for message in messages %}{% if loop.first and messages[0]['role'] != 'system' %}{{ '<s>[unused9]系统：[unused10]' }}{% endif %}{% if message['role'] == 'system' %}{{'<s>[unused9]系统：' + message['content'] + '[unused10]'}}{% endif %}{% if message['role'] == 'assistant' %}{{'[unused9]助手：' + message['content'] + '[unused10]'}}{% endif %}{% if message['role'] == 'tool' %}{{'[unused9]工具：' + message['content'] + '[unused10]'}}{% endif %}{% if message['role'] == 'function' %}{{'[unused9]方法：' + message['content'] + '[unused10]'}}{% endif %}{% if message['role'] == 'user' %}{{'[unused9]用户：' + message['content'] + '[unused10]'}}{% endif %}{% endfor %}{% if add_generation_prompt %}{{ '[unused9]助手：' }}{% endif %}",
                                "messages": [
                                    {"role": "system", "content": system_prompt},
                                    {"role": "user", "content": user_prompt + " /no_think"}
                                ],
                                "spaces_between_special_tokens": False,
                                "max_tokens": max_tokens,
                                "temperature": temperature,
                            },
                            timeout=model_config.get("timeout", 180)
                        )
                        response = response.json()
                        logger.debug(f"API response received")

                        break  # Success, exit retry loop
                    except Exception as e:
                        logger.warning(f"LLM API call attempt {attempt + 1} failed: {e}")
                        if attempt == max_retries - 1:
                            raise e  # Last attempt, re-raise the exception
                        time.sleep(5)  # Simple 1 second delay between retries

                if response is None:
                    raise Exception("Failed to get response after all retries")

                # ai_response = response.choices[0].message.content.strip()
                ai_response = response["choices"][0]["message"]["content"].strip()

                # Extract content from first response
                content = ""
                if "<chapter_content>" in ai_response:
                    content = ai_response.split("<chapter_content>")[1].split("</chapter_content>")[0].strip()
                else:
                    content = ai_response

                logger.debug(f"Content before correction: {content[:200]}...")
                logger.debug(f"Overall outline: {overall_outline[:200]}...")
                content = self._correct_title_format(content, overall_outline)
                logger.debug(f"Content after correction: {content[:200]}...")
                # Second round: Request summary
                summary_prompt = "Please give a brief summary of the output chapter content. Be sure to ensure that the language of the summary is consistent with the language of the output chapter content. For example, if the chapter content is in Chinese, your summary should also be in Chinese."

                summary_response = None
                max_retries = 5
                for attempt in range(max_retries):
                    try:
                        summary_response = requests.post(
                            url=model_url,
                            headers=headers,
                            json={
                                "model": model_config.get('model', 'pangu_auto'),
                                "chat_template":"{% for message in messages %}{% if loop.first and messages[0]['role'] != 'system' %}{{ '<s>[unused9]系统：[unused10]' }}{% endif %}{% if message['role'] == 'system' %}{{'<s>[unused9]系统：' + message['content'] + '[unused10]'}}{% endif %}{% if message['role'] == 'assistant' %}{{'[unused9]助手：' + message['content'] + '[unused10]'}}{% endif %}{% if message['role'] == 'tool' %}{{'[unused9]工具：' + message['content'] + '[unused10]'}}{% endif %}{% if message['role'] == 'function' %}{{'[unused9]方法：' + message['content'] + '[unused10]'}}{% endif %}{% if message['role'] == 'user' %}{{'[unused9]用户：' + message['content'] + '[unused10]'}}{% endif %}{% endfor %}{% if add_generation_prompt %}{{ '[unused9]助手：' }}{% endif %}",
                                "messages": [
                                    {"role": "system", "content": system_prompt},
                                    {"role": "user", "content": user_prompt + " /no_think"},
                                    {"role": "assistant", "content": ai_response},
                                    {"role": "user", "content": summary_prompt + " /no_think"}
                                ],
                                "max_tokens": max_tokens,
                                "spaces_between_special_tokens": False,
                                "temperature": temperature,
                            },
                            timeout=model_config.get("timeout", 180)
                        )
                        summary_response = summary_response.json()
                        logger.debug(f"Summary API response received")

                        break  # Success, exit retry loop
                    except Exception as e:
                        logger.warning(f"Summary LLM API call attempt {attempt + 1} failed: {e}")
                        if attempt == max_retries - 1:
                            raise e  # Last attempt, re-raise the exception
                        time.sleep(5)  # Simple delay between retries

                if summary_response is None:
                    raise Exception("Failed to get summary response after all retries")

                # summary_ai_response = summary_response.choices[0].message.content.strip()
                summary_ai_response = summary_response["choices"][0]["message"]["content"].strip()
                summary = summary_ai_response

                session_context = self.get_session_context()
                session_id = session_context.get("session_id")

                # Save multi-round conversation history
                conversation_history = [
                    # {"role": "system", "content": "system 1"},
                    {"role": "system", "content": system_prompt},
                    {"role": "user", "content": user_prompt + " /no_think"},
                    {"role": "assistant", "content": ai_response},
                    {"role": "user", "content": summary_prompt + " /no_think"},
                    {"role": "assistant", "content": summary_ai_response}
                ]

                # 把当前内容写入到target_file_path中
                write_result = self.file_write(file_path=target_file_path,
                                               content=content,
                                               create_dirs=True)
                if not write_result.success:
                    raise Exception(f"File write failed: {write_result.error}")

                results = []
                return MCPToolResult(
                    success=True,
                    data=results.append({
                        "chapter_summary": summary,
                    }),
                    metadata={
                        'content_length': len(content),
                        'summary_length': len(summary)
                    }
                )

            except Exception as e:
                logger.error(f"AI model call failed: {e}")
                return MCPToolResult(
                    success=False,
                    error=f"section writer failed: {str(e)}"
                )

        except Exception as e:
            logger.error(f"section writer failed: {e}")
            return MCPToolResult(success=False, error=str(e))

    def merge_reports(self, section_contents, output_file):
        report_files = []
        for section_content in section_contents:
            # Handle both dict format (expected) and string format (fallback)
            if isinstance(section_content, dict):
                file_path = section_content.get('file_path')
            else:
                # Fallback for direct string paths
                file_path = section_content

            if file_path:
                full_path = self._safe_join(file_path)
                report_files.append(full_path)

        # 提取文件名中的数字索引并排序
        def extract_index(file_path):
            """从文件名中提取数字索引"""
            filename = os.path.basename(file_path)
            match = re.search(r'part_(\d+)\.md', filename)
            if match:
                return int(match.group(1))
            return None  # 不符合格式的文件

        # 创建(索引, 文件路径)元组列表并排序
        indexed_files = []
        for file_path in report_files:
            idx = extract_index(file_path)
            if idx is not None:
                indexed_files.append((idx, file_path))

        # 按索引排序
        indexed_files.sort(key=lambda x: x[0])

        if not indexed_files:
            logger.warning("No report_*.md files found matching criteria")
            return

        # 合并文件
        try:
            with open(output_file, 'w', encoding='utf-8') as outfile:
                for idx, file_path in indexed_files:
                    filename = os.path.basename(file_path)
                    try:
                        with open(file_path, 'r', encoding='utf-8') as infile:
                            # 写入文件内容
                            outfile.write(infile.read())
                            outfile.write("\n\n")
                            # 添加分隔标记
                            # outfile.write(f"\n\n--- 文件结束: {filename} ---\n\n")
                        logger.info(f"Merged: {filename}")
                    except Exception as e:
                        logger.warning(f"Unable to read file {filename}: {str(e)}")

            logger.info(f"\nMerge completed! The result is saved at: {output_file}")
            logger.info(f"A total of {len(indexed_files)} files have been merged.")

        except Exception as e:
            logger.error(f"Error: Failed to write to output file - {str(e)}")
            # 清理可能创建的不完整输出文件
            if os.path.exists(output_file):
                os.remove(output_file)
            raise

    def concat_section_files(
            self,
            section_files: List[Dict],
            final_file_path: str
    ) -> MCPToolResult:
        """
        Concatenate the content of the saved section files into a single file
        """
        try:
            logger.debug(f"Starting file concatenation: section_files={section_files}, final_file={final_file_path}")

            # Convert relative path to absolute if needed
            final_file_path = self._safe_join(final_file_path)

            output_dir = os.path.dirname(final_file_path)
            if output_dir:
                os.makedirs(output_dir, exist_ok=True)                

            # 读取section_files中的文件内容
            self.merge_reports(section_files, final_file_path)
            return MCPToolResult(
                success=True,
                data={"merged_files": len(section_files), "output_path": str(final_file_path)},
                metadata={
                    'final_file_path': str(final_file_path),
                    'section_count': len(section_files)
                }
            )
        except Exception as e:
            logger.error(f"Concatenate section files failed: {e}")
            return MCPToolResult(success=False, error=str(e))

    def document_extract(
            self,
            # save_analysis_file_path: str,
            tasks: List[Dict],
            model: str = "pangu_auto",
            temperature: Optional[float] = None,
            max_tokens: Optional[int] = None,
            max_workers: int = 5
    ) -> MCPToolResult:
        """
        Multi-dimensional analysis of locally stored files using AI models.
        Evaluates each file across four key dimensions: source authority, core content extraction,
        information richness, and query relevance scoring.

        Args:
            tasks: List of task dictionaries containing:
                - file_path: Relative path to the file (relative to workspace root) to read
                - task: task for relevance assessment
            model: AI model to use for multi-dimensional analysis
            temperature: Creativity level for the AI response (0-1)
            max_tokens: Maximum tokens for the AI response
            max_workers: Maximum number of concurrent model API requests
        """
        try:
            # Get configuration
            from config.config import get_model_config, get_storage_config
            model_config = get_model_config()
            storage_config = get_storage_config()
            
            # Use config values or defaults
            if temperature is None:
                temperature = model_config.get('temperature', 0.3)
            if max_tokens is None:
                max_tokens = model_config.get('max_tokens', 8192)
                
            logger.debug(f"Starting document extraction: tasks={tasks}")

            def process_single_task(task: Dict) -> Dict:
                file_path = task['file_path']
                task_content = task['task']

                # 1. 读取文件内容
                read_result = self.file_read(file_path)
                if not read_result.success:
                    return {
                        'file_path': file_path,
                        'task': task_content,
                        'success': False,
                        'error': f"File read error: {read_result.error}",
                        'answer': None
                    }

                content = read_result.data
                system_prompt = (
                    "You are a text expert. Next, you will be given a document and task content. You need to analyze this document carefully and then provide multiple dimensional information for this document.\n\n"

                    "The following are some dimensional information extracted:\n"
                    "1. Web page time: According to the content of the document, extract the web page time of the document content. If it cannot be judged, it is expressed as \"unable to determine the web page time\"; otherwise, the time of the web page is output, accurate to the month, in the format of \"YYYY year MM month\", such as \"2023 June\";\n"
                    "2. Authority: According to the information of the document, judge the source of the web page to confirm the credibility of the web page.\n"
                    "3. Relevance: According to the current task (task_content) and the given document, judge whether the current document is related to the current task.\n"
                    "4. Core content: Based on this document, you make a core content summary to ensure the richness of information, with a word count of about 200 words.\n"
                    "Information richness: Estimate the total word count of substantive content in the document. Less than 200 words indicates scarcity; over 800 words suggests high richness; between these thresholds denotes moderate richness. Be careful not to just give the word count results, but also give a corresponding text description of how informative the content is.\n"

                    "Note:\n1. Ensure the document's language aligns with the extracted dimensions (e.g., Chinese content requires Chinese extraction).\n2. For **source_authority** and **task_relevance**, first provide a brief description before concluding.  \n"
                    "- **Authority**: Briefly assess the source's credibility (e.g., expertise, reputation). *Conclusion*: [High/Medium/Low].  \n"
                    "- **Relevance**: Summarize content alignment with the topic. *Conclusion*: [High/Medium/Low].\n"

                    "The final output format must be a valid JSON object:\n"
                    "{\n"
                    "  \"doc_time\": \"xxx\",\n"
                    "  \"source_authority\": \"xxx\",\n"
                    "  \"task_relevance\": \"xxx\",\n"
                    "  \"core_content\": \"xxx\",\n"
                    "  \"information_richness\": \"xxx\"\n"
                    "}\n\n"
                    "Important: Return ONLY the JSON object, no additional text or formatting."
                )

                # 构建用户提示
                user_prompt = (
                    f"DOCUMENT CONTENT:\n{content}\n"
                    # f"DOCUMENT LEN: The length of the file content is{len(content)}\n"
                    f"TASK FOR RELEVANCE ASSESSMENT: {task_content}"
                )

                # Get model URL and token from config
                config = get_config()
                model_config = config.get_custom_llm_config()
                
                model_url = model_config.get('url') or os.getenv('MODEL_REQUEST_URL', '')
                model_token = model_config.get('token') or os.getenv('MODEL_REQUEST_TOKEN', '')
                headers = {'Content-Type': 'application/json', 'csb-token': model_token}

                try:
                    # Add retry logic for AI model call
                    max_retries = 5
                    response = None

                    for attempt in range(max_retries):
                        try:
                            response = requests.post(
                                url=model_url,
                                headers=headers,
                                json={
                                    "model": model_config.get('model', 'pangu_auto'),
                                    "chat_template": "{% for message in messages %}{% if loop.first and messages[0]['role'] != 'system' %}{{ '<s>[unused9]系统：[unused10]' }}{% endif %}{% if message['role'] == 'system' %}{{'<s>[unused9]系统：' + message['content'] + '[unused10]'}}{% endif %}{% if message['role'] == 'assistant' %}{{'[unused9]助手：' + message['content'] + '[unused10]'}}{% endif %}{% if message['role'] == 'tool' %}{{'[unused9]工具：' + message['content'] + '[unused10]'}}{% endif %}{% if message['role'] == 'function' %}{{'[unused9]方法：' + message['content'] + '[unused10]'}}{% endif %}{% if message['role'] == 'user' %}{{'[unused9]用户：' + message['content'] + '[unused10]'}}{% endif %}{% endfor %}{% if add_generation_prompt %}{{ '[unused9]助手：' }}{% endif %}",
                                    "messages": [
                                        {"role": "system", "content": system_prompt},
                                        {"role": "user", "content": user_prompt + " /no_think"}
                                    ],
                                    "max_tokens": max_tokens,
                                    "spaces_between_special_tokens": False,
                                    "temperature": temperature,
                                },
                                timeout=model_config.get("timeout", 180)
                            )
                            response = response.json()
                            logger.info(f"LLM API response: {response}")

                            break  # Success, exit retry loop
                        except Exception as e:
                            logger.warning(f"LLM API call attempt {attempt + 1} failed: {e}")
                            if attempt == max_retries - 1:
                                raise e  # Last attempt, re-raise the exception
                            time.sleep(4)  # Simple 1 second delay between retries

                    if response is None:
                        raise Exception("Failed to get response after all retries")

                    # answer = response.choices[0].message.content
                    answer = response["choices"][0]["message"]["content"]

                    session_context = self.get_session_context()
                    session_id = session_context.get("session_id")
                    # 切换保存的方式
                    conversation_history = [
                        # {"role": "system", "content": "system 1"},
                        {"role": "system", "content": system_prompt},
                        {"role": "user", "content": user_prompt + " /no_think"},
                        {"role": "assistant", "content": answer}
                    ]

                    return {
                        'file_path': file_path,
                        'task': task_content,
                        'success': True,
                        'answer': answer,
                        'metadata': {
                            'file_size': len(content),
                            'line_count': len(content.splitlines())
                        }
                    }

                except Exception as e:
                    logger.error(f"Model API call failed for file '{file_path}': {e}")
                    return {
                        'file_path': file_path,
                        'task': task_content,
                        'success': False,
                        'error': f"Model API error: {str(e)}"
                    }

            # 4. 并发处理所有任务
            results = []
            with ThreadPoolExecutor(max_workers=min(max_workers, len(tasks))) as executor:
                future_to_task = {executor.submit(process_single_task, task): task for task in tasks}

                for future in as_completed(future_to_task):
                    try:
                        result = future.result()
                        results.append(result)
                    except Exception as e:
                        task = future_to_task[future]
                        logger.error(f"Task processing failed for file '{task['file_path']}': {e}")
                        results.append({
                            'file_path': task['file_path'],
                            'task': task['task'],
                            'success': False,
                            'error': f"Task processing exception: {str(e)}"
                        })

            # 5. 保持原始任务顺序
            task_order = {task['file_path']: i for i, task in enumerate(tasks)}
            results.sort(key=lambda x: task_order.get(x['file_path'], float('inf')))
            # 保存结果到文件
            def parse_answer_to_structured_data(answer_text: str, file_path: str) -> Dict[str, str]:
                """Parse the AI JSON response into structured data"""
                # Default structure
                structured_data = {
                    "file_path": file_path,
                    "doc_time": "Unknown",
                    "source_authority": "Unknown",
                    "task_relevance": "Unknown",
                    "information_richness": "Unknown",
                    "core_content": "Unknown"
                }

                if not answer_text:
                    return structured_data

                try:
                    # Try to parse as JSON
                    answer_text = answer_text.strip()

                    # Remove any markdown code blocks if present
                    if answer_text.startswith('```'):
                        lines = answer_text.split('\n')
                        # Find the start and end of JSON content
                        start_idx = 0
                        end_idx = len(lines)
                        for i, line in enumerate(lines):
                            if line.strip().startswith('{'):
                                start_idx = i
                                break
                        for i in range(len(lines) - 1, -1, -1):
                            if lines[i].strip().endswith('}'):
                                end_idx = i + 1
                                break
                        answer_text = '\n'.join(lines[start_idx:end_idx])

                    # Parse JSON
                    parsed_data = json.loads(answer_text)

                    # Update structured_data with parsed values
                    if isinstance(parsed_data, dict):
                        structured_data.update({
                            "file_path": file_path,
                            "doc_time": parsed_data.get("doc_time", "Unknown"),
                            "source_authority": parsed_data.get("source_authority", "Unknown"),
                            "task_relevance": parsed_data.get("task_relevance", "Unknown"),
                            "core_content": parsed_data.get("core_content", "Unknown"),
                            "information_richness": parsed_data.get("information_richness", "Unknown")
                        })

                    return structured_data

                except json.JSONDecodeError as e:
                    # If JSON parsing fails, return default with error info
                    structured_data[
                        "core_content"] = f"JSON parsing error: {str(e)}. Raw response: {answer_text[:200]}..."
                    return structured_data
                except Exception as e:
                    # Handle any other parsing errors
                    structured_data["core_content"] = f"Parsing error: {str(e)}"
                    return structured_data

            # Transform results into the desired format
            structured_results = []
            for result in results:
                if result.get('success', False) and result.get('answer'):
                    structured_data = parse_answer_to_structured_data(
                        result['answer'],
                        result['file_path']
                    )
                    structured_results.append(structured_data)
                else:
                    # For failed results, still include basic info
                    structured_results.append({
                        "file_path": result.get('file_path', 'Unknown'),
                        "doc_time": "Processing failed",
                        "source_authority": "Processing failed",
                        "task_relevance": "Processing failed",
                        "information_richness": "Unknown",
                        "core_content": f"Error: {result.get('error', 'Unknown error')}"
                    })

            # Save structured results to JSON file
            # Create full path relative to workspace
            analysis_path = storage_config.get('document_analysis_path', './doc_analysis')
            full_save_path = self.workspace_path / analysis_path / "file_analysis.jsonl"
            full_save_path.parent.mkdir(parents=True, exist_ok=True)

            # with open(full_save_path, "a", encoding='utf-8') as f:
            #     json.dump(structured_results, f, ensure_ascii=False, indent=2)

            with open(full_save_path, mode="a", encoding='utf-8') as f:
                for ii in structured_results:
                    f.write(json.dumps(ii, ensure_ascii=False) + "\n")

            # 6. 统计结果
            successful_tasks = len([r for r in results if r.get('success', False)])

            return MCPToolResult(
                success=True,
                data=results,
                metadata={
                    'total_tasks': len(tasks),
                    'successful_tasks': successful_tasks,
                    'failed_tasks': len(tasks) - successful_tasks,
                    'model': model,
                    'concurrent_workers': min(max_workers, len(tasks))
                }
            )

        except Exception as e:
            logger.error(f"Context-based document extraction failed: {e}")
            return MCPToolResult(success=False, error=str(e))

    # ================ FILE DOWNLOAD TOOLS ================

    def download_files(
            self,
            urls: List[str],
            target_directory: str = None,
            overwrite: bool = False,
            max_file_size_mb: int = 100
    ) -> MCPToolResult:
        """
        Download human-readable research files such as PDFs, documents, and data files.
        
        Use this tool for downloading research papers, documentation, reports, data files (CSV, JSON, XML),
        academic publications, and other human-readable content that you can analyze.
        
        WARNING: Do NOT use this tool for downloading web pages (HTML/HTM files) or other non-readable formats.
        For web page content extraction, use the url_crawler tool instead.
        
        Args:
            urls: List of URLs to download (PDFs, DOCs, research papers, data files, etc.)
            target_directory: Directory to save files (relative to session workspace)
            overwrite: Whether to overwrite existing files
            max_file_size_mb: Maximum file size in MB
        """
        try:
            if target_directory:
                # Ensure target_directory is relative to session workspace for security
                download_dir = self._safe_join(target_directory)
            else:
                download_dir = self.workspace_path / "downloads"

            download_dir.mkdir(parents=True, exist_ok=True)

            def download_single_file(url: str) -> Dict[str, Any]:
                """Download a single file"""
                try:
                    # Parse URL to get filename
                    parsed_url = urlparse(url)
                    filename = os.path.basename(parsed_url.path) or 'downloaded_file'

                    # Ensure filename has extension
                    if '.' not in filename:
                        filename += '.html'  # Default extension

                    if os.path.isabs(filename):
                        raise Exception(f"Path '{filename}' is absolute. Only relative paths are allowed.")
                    file_path = download_dir / filename
                    if not os.path.realpath(file_path).startswith(self.full_workspace_path):
                        raise Exception(f"Path '{filename}' is outside workspace directory.")

                    # Check if file exists
                    if file_path.exists() and not overwrite:
                        return {
                            'url': url,
                            'success': False,
                            'error': 'File already exists',
                            'file_path': str(file_path)
                        }

                    # Download file
                    response = requests.get(url, stream=True, timeout=30)
                    response.raise_for_status()

                    # Check file size
                    content_length = response.headers.get('content-length')
                    if content_length and int(content_length) > max_file_size_mb * 1024 * 1024:
                        return {
                            'url': url,
                            'success': False,
                            'error': f'File too large (>{max_file_size_mb}MB)',
                            'file_path': None
                        }

                    # Save file
                    with open(file_path, 'wb') as f:
                        for chunk in response.iter_content(chunk_size=8192):
                            f.write(chunk)

                    return {
                        'url': url,
                        'success': True,
                        'file_path': str(file_path),
                        'file_size': file_path.stat().st_size
                    }

                except Exception as e:
                    return {
                        'url': url,
                        'success': False,
                        'error': str(e),
                        'file_path': None
                    }

            # Process downloads concurrently
            results = []
            max_concurrent_downloads = min(5, len(urls))  # Limit concurrent downloads to avoid overwhelming servers
            with ThreadPoolExecutor(max_workers=max_concurrent_downloads) as executor:
                # Submit all download tasks
                future_to_url = {executor.submit(download_single_file, url): url for url in urls}

                # Collect results as they complete
                for future in as_completed(future_to_url):
                    try:
                        result = future.result()
                        results.append(result)
                    except Exception as e:
                        url = future_to_url[future]
                        logger.error(f"Download task failed for '{url}': {e}")
                        results.append({
                            'url': url,
                            'success': False,
                            'error': f"Download task exception: {str(e)}",
                            'file_path': None
                        })

            # Sort results to maintain original URL order
            url_order = {urls[i]: i for i in range(len(urls))}
            results.sort(key=lambda x: url_order.get(x['url'], float('inf')))

            # Generate status message following context management philosophy
            successful_downloads = len([r for r in results if r.get('success', False)])
            failed_downloads = len(results) - successful_downloads

            status_msg = f"File download task completed. Processed {len(urls)} URLs with {successful_downloads} successful downloads and {failed_downloads} failures. Files saved to {download_dir.relative_to(self.workspace_path)}. Use file reading tools to examine the downloaded files."

            return MCPToolResult(
                success=True,
                data=status_msg,
                metadata={
                    'download_directory': str(download_dir),
                    'total_urls': len(urls),
                    'successful_downloads': successful_downloads,
                    'failed_downloads': failed_downloads
                }
            )

        except Exception as e:
            logger.error(f"Download files failed: {e}")
            return MCPToolResult(success=False, error=str(e))

    # ================ WORKSPACE TOOLS ================

    def list_workspace(
            self,
            path: str = None,
            recursive: bool = False,
            include_hidden: bool = False,
            max_depth: int = 3
    ) -> MCPToolResult:
        """
        List files and directories in workspace with tree structure visualization
        
        Args:
            path: Specific path to list (relative to session workspace)
            recursive: Whether to list recursively
            include_hidden: Whether to include hidden files
            max_depth: Maximum recursion depth
        """
        try:
            if path:
                # Ensure path is relative to session workspace for security
                target_path = self._safe_join(path)
            else:
                target_path = self.workspace_path

            if not target_path.exists():
                return MCPToolResult(
                    success=False,
                    error=f"Path does not exist: {target_path}"
                )

            if not target_path.is_dir():
                return MCPToolResult(
                    success=False,
                    error=f"Path is not a directory: {target_path}"
                )

            items = []
            tree_structure = []

            def _list_items(current_path: Path, current_depth: int = 0):
                if current_depth > max_depth:
                    return

                try:
                    # Get all items and sort them (directories first, then files, both alphabetically)
                    all_items = list(current_path.iterdir())
                    if not include_hidden:
                        all_items = [item for item in all_items if not item.name.startswith('.')]

                    # Sort: directories first, then files, both alphabetically
                    all_items.sort(key=lambda x: (not x.is_dir(), x.name.lower()))

                    for item in all_items:
                        item_info = {
                            'name': item.name,
                            'path': str(item.relative_to(self.workspace_path)),
                            'type': 'directory' if item.is_dir() else 'file',
                            'size': item.stat().st_size if item.is_file() else None,
                            'modified': item.stat().st_mtime,
                            'depth': current_depth
                        }

                        items.append(item_info)

                        # Recurse into directories
                        if recursive and item.is_dir():
                            _list_items(item, current_depth + 1)

                except PermissionError:
                    pass  # Skip directories we can't read

            def _generate_tree_structure(current_path: Path, prefix: str = "", is_last: bool = True,
                                         current_depth: int = 0):
                """Generate ASCII tree structure recursively"""
                if current_depth > max_depth:
                    return

                try:
                    # Get all items and sort them (directories first, then files, both alphabetically)
                    all_items = list(current_path.iterdir())
                    if not include_hidden:
                        all_items = [item for item in all_items if not item.name.startswith('.')]

                    # Sort: directories first, then files, both alphabetically
                    all_items.sort(key=lambda x: (not x.is_dir(), x.name.lower()))

                    for i, item in enumerate(all_items):
                        is_last_item = i == len(all_items) - 1

                        # Choose the appropriate tree symbols
                        if is_last_item:
                            current_symbol = "└── "
                            extension = "    "
                        else:
                            current_symbol = "├── "
                            extension = "│   "

                        # Add file/directory indicator
                        if item.is_dir():
                            name_with_indicator = f"📁 {item.name}/"
                        else:
                            # Add file size for files
                            try:
                                size = item.stat().st_size
                                if size < 1024:
                                    size_str = f"{size}B"
                                elif size < 1024 * 1024:
                                    size_str = f"{size / 1024:.1f}KB"
                                else:
                                    size_str = f"{size / (1024 * 1024):.1f}MB"
                                name_with_indicator = f"📄 {item.name} ({size_str})"
                            except:
                                name_with_indicator = f"📄 {item.name}"

                        tree_line = prefix + current_symbol + name_with_indicator
                        tree_structure.append(tree_line)

                        # Recurse into directories
                        if recursive and item.is_dir():
                            _generate_tree_structure(
                                item,
                                prefix + extension,
                                is_last_item,
                                current_depth + 1
                            )

                except PermissionError:
                    tree_structure.append(prefix + "└── [Permission Denied]")

            # Generate both flat list and tree structure
            _list_items(target_path)

            # Generate tree structure
            root_name = target_path.name if target_path.name else "workspace"
            tree_structure.append(f"📁 {root_name}/")

            if recursive:
                _generate_tree_structure(target_path)
            else:
                # For non-recursive, just show immediate children
                try:
                    all_items = list(target_path.iterdir())
                    if not include_hidden:
                        all_items = [item for item in all_items if not item.name.startswith('.')]

                    all_items.sort(key=lambda x: (not x.is_dir(), x.name.lower()))

                    for i, item in enumerate(all_items):
                        is_last_item = i == len(all_items) - 1
                        symbol = "└── " if is_last_item else "├── "

                        if item.is_dir():
                            name_with_indicator = f"📁 {item.name}/"
                        else:
                            try:
                                size = item.stat().st_size
                                if size < 1024:
                                    size_str = f"{size}B"
                                elif size < 1024 * 1024:
                                    size_str = f"{size / 1024:.1f}KB"
                                else:
                                    size_str = f"{size / (1024 * 1024):.1f}MB"
                                name_with_indicator = f"📄 {item.name} ({size_str})"
                            except:
                                name_with_indicator = f"📄 {item.name}"

                        tree_structure.append(symbol + name_with_indicator)

                except PermissionError:
                    tree_structure.append("└── [Permission Denied]")

            # Create the tree string
            tree_string = "\n".join(tree_structure)

            return MCPToolResult(
                success=True,
                data={
                    'items': items,
                    'tree_structure': tree_string,
                    'tree_lines': tree_structure
                },
                metadata={
                    'target_path': str(target_path.relative_to(self.workspace_path)) if path else '.',
                    'total_items': len(items),
                    'recursive': recursive,
                    'max_depth': max_depth,
                    'include_hidden': include_hidden
                }
            )

        except Exception as e:
            logger.error(f"List workspace failed: {e}")
            return MCPToolResult(success=False, error=str(e))

    # ================ FILE EDITING TOOLS ================
    def str_replace_based_edit_tool(
            self,
            action: str,
            file_path: str,
            content: str = None,
            old_str: str = None,
            new_str: str = None,
            line_number: int = None,
            max_char_len: int = 10000,
    ) -> MCPToolResult:
        """
        Comprehensive file editing tool
        
        Args:
            action: 'create', 'view', 'str_replace', 'insert', 'append', 'delete'
            file_path: Path to the file
            content: Content for create action
            old_str: String to replace (for str_replace)
            new_str: Replacement string (for str_replace)
            line_number: Line number for insert action
        """
        try:
            full_path = self._safe_join(file_path)

            if action == 'create':
                if full_path.exists():
                    return MCPToolResult(
                        success=False,
                        error=f"File already exists: {file_path}"
                    )

                full_path.parent.mkdir(parents=True, exist_ok=True)
                full_path.write_text(content or '', encoding='utf-8')

                return MCPToolResult(
                    success=True,
                    data=f"File created: {file_path}",
                    metadata={'file_size': full_path.stat().st_size}
                )

            elif action == 'view':
                if not full_path.exists():
                    return MCPToolResult(
                        success=False,
                        error=f"File does not exist: {file_path}"
                    )

                content = full_path.read_text(encoding='utf-8')
                if len(content) > max_char_len:
                    content = ("Due to the content being too long, only the first 10,000 characters are returned. "
                               "It is recommended to use other tools such as `document_qa` to extract the required content from the file. "
                               "Below is the returned portion of the file content: \n\n") + content[:max_char_len]

                return MCPToolResult(
                    success=True,
                    data=content,
                    metadata={
                        'file_size': len(content),
                        'line_count': len(content.splitlines())
                    }
                )

            elif action == 'str_replace':
                if not full_path.exists():
                    return MCPToolResult(
                        success=False,
                        error=f"File does not exist: {file_path}"
                    )

                if not old_str or new_str is None:
                    return MCPToolResult(
                        success=False,
                        error="Both old_str and new_str are required for str_replace"
                    )

                original_content = full_path.read_text(encoding='utf-8')

                if old_str not in original_content:
                    return MCPToolResult(
                        success=False,
                        error=f"String not found: {old_str[:50]}..."
                    )

                new_content = original_content.replace(old_str, new_str)
                full_path.write_text(new_content, encoding='utf-8')

                return MCPToolResult(
                    success=True,
                    data=f"Replaced {original_content.count(old_str)} occurrence(s)",
                    metadata={
                        'old_size': len(original_content),
                        'new_size': len(new_content)
                    }
                )

            elif action == 'insert':
                if not full_path.exists():
                    return MCPToolResult(
                        success=False,
                        error=f"File does not exist: {file_path}"
                    )

                if line_number is None or content is None:
                    return MCPToolResult(
                        success=False,
                        error="Both line_number and content are required for insert"
                    )

                lines = full_path.read_text(encoding='utf-8').splitlines()

                if line_number < 0 or line_number > len(lines):
                    return MCPToolResult(
                        success=False,
                        error=f"Invalid line number: {line_number}"
                    )

                lines.insert(line_number, content)
                full_path.write_text('\n'.join(lines), encoding='utf-8')

                return MCPToolResult(
                    success=True,
                    data=f"Inserted content at line {line_number}",
                    metadata={'new_line_count': len(lines)}
                )

            elif action == 'append':
                if not full_path.exists():
                    full_path.touch()

                with open(full_path, 'a', encoding='utf-8') as f:
                    f.write(content or '')

                return MCPToolResult(
                    success=True,
                    data=f"Appended content to {file_path}",
                    metadata={'file_size': full_path.stat().st_size}
                )

            elif action == 'delete':
                if not full_path.exists():
                    return MCPToolResult(
                        success=False,
                        error=f"File does not exist: {file_path}"
                    )

                full_path.unlink()

                return MCPToolResult(
                    success=True,
                    data=f"Deleted file: {file_path}"
                )

            else:
                return MCPToolResult(
                    success=False,
                    error=f"Unknown action: {action}"
                )

        except Exception as e:
            logger.error(f"File edit failed: {e}")
            return MCPToolResult(success=False, error=str(e))

    # ================ BASIC FILE TOOLS ================

    def file_read(self, file_path: str, encoding: str = 'utf-8', max_char_len: int = 10000) -> MCPToolResult:
        """Read file content"""
        try:
            full_path = self._safe_join(file_path)

            if not full_path.exists():
                return MCPToolResult(
                    success=False,
                    error=f"File does not exist: {file_path}"
                )

            content = full_path.read_text(encoding=encoding)
            if len(content) > max_char_len:
                content = "Due to the content being too long, only the first 10,000 characters are returned. It is recommended to use other tools such as `document_qa` to extract the required content from the file. Below is the returned portion of the file content: \n\n" + content[:max_char_len]

            return MCPToolResult(
                success=True,
                data=content,
                metadata={
                    'file_size': len(content),
                    'line_count': len(content.splitlines()),
                    'encoding': encoding
                }
            )

        except Exception as e:
            logger.error(f"File read failed: {e}")
            return MCPToolResult(success=False, error=str(e))

    # ================ ENHANCED FILE ANALYSIS TOOLS ================

    def file_stats(self, file_path: str) -> MCPToolResult:
        """
        Get comprehensive file statistics without reading full content.
        Perfect for deciding whether to read full file or use targeted extraction.
        
        Args:
            file_path: Path to the file (relative to workspace)
        """
        try:
            full_path = self._safe_join(file_path)

            if not full_path.exists():
                return MCPToolResult(
                    success=False,
                    error=f"File does not exist: {file_path}"
                )

            if not full_path.is_file():
                return MCPToolResult(
                    success=False,
                    error=f"Path is not a file: {file_path}"
                )

            # Get basic file stats
            stat_info = full_path.stat()
            file_size = stat_info.st_size

            # Quick content analysis without loading full file
            encoding = 'utf-8'
            line_count = 0
            word_count = 0
            char_count = 0
            first_lines = []
            last_lines = []

            try:
                with open(full_path, 'r', encoding=encoding, errors='ignore') as f:
                    # Read first few lines for preview
                    for i, line in enumerate(f):
                        line_count += 1
                        if i < 5:  # First 5 lines
                            first_lines.append(line.rstrip())

                        char_count += len(line)
                        word_count += len(line.split())

                        # For efficiency, stop detailed counting after reasonable limit
                        if line_count > 10000:
                            # Estimate remaining based on average
                            remaining_size = file_size - f.tell()
                            if remaining_size > 0:
                                avg_line_size = f.tell() / line_count
                                estimated_remaining_lines = int(remaining_size / avg_line_size)
                                line_count += estimated_remaining_lines

                                # Estimate words and chars
                                avg_chars_per_line = char_count / min(line_count, 10000)
                                avg_words_per_line = word_count / min(line_count, 10000)
                                char_count += int(remaining_size)
                                word_count += int(estimated_remaining_lines * avg_words_per_line)
                            break

                # Get last few lines if file is reasonable size
                if file_size < 1024 * 1024:  # Less than 1MB
                    with open(full_path, 'r', encoding=encoding, errors='ignore') as f:
                        lines = f.readlines()
                        last_lines = [line.rstrip() for line in lines[-5:]]
                        if line_count <= 10000:  # Recalculate if we estimated
                            line_count = len(lines)
                            char_count = sum(len(line) for line in lines)
                            word_count = sum(len(line.split()) for line in lines)

            except Exception as e:
                # Try binary mode to at least get size info
                encoding = 'binary'
                char_count = file_size

            # Determine file type
            file_extension = full_path.suffix.lower()
            file_type = self._detect_file_type(full_path, file_extension)

            # Reading recommendation
            reading_recommendation = self._get_reading_recommendation(
                file_size, line_count, word_count, file_type
            )

            stats = {
                'file_path': file_path,
                'file_size_bytes': file_size,
                'file_size_human': self._format_file_size(file_size),
                'line_count': line_count,
                'word_count': word_count,
                'character_count': char_count,
                'encoding': encoding,
                'file_type': file_type,
                'file_extension': file_extension,
                'modified_time': stat_info.st_mtime,
                'is_large_file': file_size > 1024 * 1024,  # > 1MB
                'is_very_large_file': file_size > 10 * 1024 * 1024,  # > 10MB
                'first_lines_preview': first_lines,
                'last_lines_preview': last_lines,
                'reading_recommendation': reading_recommendation
            }

            return MCPToolResult(
                success=True,
                data=stats,
                metadata={
                    'analysis_method': 'efficient_sampling' if line_count > 10000 else 'full_analysis'
                }
            )

        except Exception as e:
            logger.error(f"File stats failed: {e}")
            return MCPToolResult(success=False, error=str(e))

    @staticmethod
    def _detect_file_type(file_path: Path, extension: str) -> str:
        """Detect file type based on extension and content"""

        # Extension-based detection
        type_map = {
            '.py': 'python_code',
            '.js': 'javascript_code',
            '.ts': 'typescript_code',
            '.java': 'java_code',
            '.cpp': 'cpp_code',
            '.c': 'c_code',
            '.html': 'html_markup',
            '.css': 'css_stylesheet',
            '.json': 'json_data',
            '.xml': 'xml_data',
            '.yaml': 'yaml_config',
            '.yml': 'yaml_config',
            '.md': 'markdown_document',
            '.txt': 'plain_text',
            '.csv': 'csv_data',
            '.sql': 'sql_code',
            '.sh': 'shell_script',
            '.dockerfile': 'docker_config',
            '.env': 'environment_config'
        }

        if extension in type_map:
            return type_map[extension]

        # Content-based detection for unknown extensions
        try:
            with open(file_path, 'r', encoding='utf-8', errors='ignore') as f:
                first_line = f.readline().strip()

                if first_line.startswith('#!'):
                    return 'executable_script'
                elif first_line.startswith('<?xml'):
                    return 'xml_data'
                elif first_line.startswith('{') or first_line.startswith('['):
                    return 'json_data'
                elif 'DOCTYPE html' in first_line or '<html' in first_line:
                    return 'html_markup'
        except:
            pass

        return 'unknown_text'

    def _format_file_size(self, size_bytes: int) -> str:
        """Format file size in human readable format"""
        if size_bytes < 1024:
            return f"{size_bytes} B"
        elif size_bytes < 1024 * 1024:
            return f"{size_bytes / 1024:.1f} KB"
        elif size_bytes < 1024 * 1024 * 1024:
            return f"{size_bytes / (1024 * 1024):.1f} MB"
        else:
            return f"{size_bytes / (1024 * 1024 * 1024):.1f} GB"

    def _get_reading_recommendation(self, file_size: int, line_count: int,
                                    word_count: int, file_type: str) -> Dict[str, Any]:
        """Provide intelligent recommendations for how to read the file"""

        recommendations = {
            'strategy': 'full_read',
            'reason': 'File is small enough for full reading',
            'alternatives': []
        }

        # Large file strategies
        if file_size > 1024 * 1024:  # > 1MB
            recommendations['strategy'] = 'selective_read'
            recommendations['reason'] = 'File is large, consider targeted approaches'
            recommendations['alternatives'] = [
                'Use file_grep_with_context to search for specific content',
                'Use content_preview to get overview before full read',
                'Use file_read_lines to read specific sections',
                'Content indexing has been disabled'
            ]

        elif line_count > 1000:
            recommendations['strategy'] = 'preview_first'
            recommendations['reason'] = 'Many lines, preview recommended before full read'
            recommendations['alternatives'] = [
                'Use content_preview for quick overview',
                'Use file_grep_with_context for specific searches'
            ]

        # File type specific recommendations
        if file_type in ['json_data', 'xml_data']:
            recommendations['alternatives'].append('Consider parsing structure instead of full text read')
        elif file_type.endswith('_code'):
            recommendations['alternatives'].append('Use grep to find specific functions/classes')
        elif file_type == 'csv_data':
            recommendations['alternatives'].append('Consider reading headers first with file_read_lines')

        return recommendations

    # ================ BASIC FILE TOOLS ================
    def file_write(
            self,
            file_path: str,
            content: str,
            encoding: str = 'utf-8',
            create_dirs: bool = True
    ) -> MCPToolResult:
        """Write content to file"""
        try:
            full_path = self._safe_join(file_path)

            if create_dirs:
                full_path.parent.mkdir(parents=True, exist_ok=True)

            full_path.write_text(content, encoding=encoding)

            return MCPToolResult(
                success=True,
                data=f"Written {len(content)} characters to {file_path}",
                metadata={
                    'file_size': full_path.stat().st_size,
                    'encoding': encoding
                }
            )

        except Exception as e:
            logger.error(f"File write failed: {e}")
            return MCPToolResult(success=False, error=str(e))

    # ================ SEARCH TOOLS ================

    def file_find_by_name(
            self,
            name_pattern: str,
            recursive: bool = True,
            case_sensitive: bool = False,
            max_results: int = 100
    ) -> MCPToolResult:
        """Find files by name pattern"""
        try:
            import fnmatch

            if not case_sensitive:
                name_pattern = name_pattern.lower()

            matches = []
            search_path = self.workspace_path

            def _match_name(file_path: Path) -> bool:
                name = file_path.name
                if not case_sensitive:
                    name = name.lower()

                return fnmatch.fnmatch(name, name_pattern)

            # Search files
            if recursive:
                for file_path in search_path.rglob("*"):
                    if _match_name(file_path):
                        matches.append({
                            'name': file_path.name,
                            'path': str(file_path.relative_to(self.workspace_path)),
                            'type': 'directory' if file_path.is_dir() else 'file',
                            'size': file_path.stat().st_size if file_path.is_file() else None
                        })

                        if len(matches) >= max_results:
                            break
            else:
                for file_path in search_path.iterdir():
                    if _match_name(file_path):
                        matches.append({
                            'name': file_path.name,
                            'path': str(file_path.relative_to(self.workspace_path)),
                            'type': 'directory' if file_path.is_dir() else 'file',
                            'size': file_path.stat().st_size if file_path.is_file() else None
                        })

                        if len(matches) >= max_results:
                            break

            return MCPToolResult(
                success=True,
                data=matches,
                metadata={
                    'pattern': name_pattern,
                    'total_matches': len(matches),
                    'truncated': len(matches) >= max_results
                }
            )

        except Exception as e:
            logger.error(f"File find failed: {e}")
            return MCPToolResult(success=False, error=str(e))

    def file_read_lines(
            self,
            file_path: str,
            start_line: int = 1,
            end_line: int = None,
            max_lines: int = 1000
    ) -> MCPToolResult:
        """
        Read specific line ranges from a file without loading the entire file.
        Perfect for reading specific sections after grep or for large files.
        
        Args:
            file_path: Path to the file
            start_line: Starting line number (1-based)
            end_line: Ending line number (1-based, None for end of file)
            max_lines: Maximum number of lines to read (safety limit)
        """
        try:
            full_path = self._safe_join(file_path)

            if not full_path.exists():
                return MCPToolResult(
                    success=False,
                    error=f"File does not exist: {file_path}"
                )

            if start_line < 1:
                return MCPToolResult(
                    success=False,
                    error="start_line must be >= 1"
                )

            lines_read = []
            current_line = 0

            with open(full_path, 'r', encoding='utf-8', errors='ignore') as f:
                for line in f:
                    current_line += 1

                    # Skip lines before start_line
                    if current_line < start_line:
                        continue

                    # Stop if we've reached end_line
                    if end_line and current_line > end_line:
                        break

                    # Safety check for max_lines
                    if len(lines_read) >= max_lines:
                        break

                    lines_read.append({
                        'line_number': current_line,
                        'content': line.rstrip('\n\r')
                    })

            # Calculate actual end line
            actual_end_line = lines_read[-1]['line_number'] if lines_read else start_line - 1

            return MCPToolResult(
                success=True,
                data={
                    'file_path': file_path,
                    'start_line': start_line,
                    'end_line': actual_end_line,
                    'lines': lines_read,
                    'line_count': len(lines_read)
                },
                metadata={
                    'total_lines_read': len(lines_read),
                    'truncated_due_to_max_lines': len(lines_read) >= max_lines
                }
            )

        except Exception as e:
            logger.error(f"File read lines failed: {e}")
            return MCPToolResult(success=False, error=str(e))

# ================ MCP TOOL SCHEMAS ================

MCP_TOOL_SCHEMAS = {
    "think": {
        "name": "think",
        "description": "Use the tool to think about something. It will not obtain new information or make any changes to the repository, but just log the thought. Use it when complex reasoning or brainstorming is needed.",
        "inputSchema": {
            "type": "object",
            "properties": {
                "thought": {
                    "type": "string",
                    "description": "Your thoughts."
                }
            },
            "required": ["thought"]
        }
    },

    "reflect": {
        "name": "reflect",
        "description": "When multiple attempts yield no progress, use this tool to reflect on previous reasoning and planning, considering possible overlooked clues and exploring more possibilities. It will not obtain new information or make any changes to the repository.",
        "inputSchema": {
            "type": "object",
            "properties": {
                "reflect": {
                    "type": "string",
                    "description": "The specific content of your reflection"
                }
            },
            "required": ["reflect"]
        }
    },

    "batch_web_search": {
        "name": "batch_web_search",
        "description": "Search multiple queries using configurable search API with concurrent processing (no more than 8 search queries)",
        "inputSchema": {
            "type": "object",
            "properties": {
                "queries": {
                    "type": "array",
                    "items": {"type": "string"},
                    "description": "List of search queries"
                },
                "max_results_per_query": {
                    "type": "integer",
                    "default": 4,
                    "description": "Maximum search results per query (limited to 10)"
                },
                "max_workers": {
                    "type": "integer",
                    "default": 5,
                    "description": "Maximum number of concurrent search requests"
                }
            },
            "required": ["queries"]
        }
    },

    "url_crawler": {
        "name": "url_crawler",
        "description": "Extract content from web pages using configurable URL crawler API. Input is a list of documents with metadata including URL and local file path for saving extracted content.",
        "inputSchema": {
            "type": "object",
            "properties": {
                "documents": {
                    "type": "array",
                    "items": {
                        "type": "object",
                        "properties": {
                            "url": {
                                "type": "string",
                                "description": "Web page URL to extract content from"
                            },
                            "file_path": {
                                "type": "string",
                                "description": "Local path to save extracted full text content"
                            },
                            "title": {
                                "type": "string",
                                "description": "Title of the web page"
                            },
                            "time": {
                                "type": "string",
                                "description": "Publication time of the web page"
                            }
                        },
                        "required": ["url", "file_path"]
                    },
                    "description": "List of documents with metadata including URL and save path"
                },
                "max_tokens_per_url": {
                    "type": "integer",
                    "default": 4000,
                    "description": "Maximum tokens per URL result"
                },
                "include_metadata": {
                    "type": "boolean",
                    "default": True,
                    "description": "Whether to include extraction metadata"
                },
                "max_workers": {
                    "type": "integer",
                    "default": 10,
                    "description": "Maximum number of concurrent extraction requests"
                }
            },
            "required": ["documents"]
        }
    },

    "document_qa": {
        "name": "document_qa",
        "description": "Answer questions based on content stored in local files. Each file has a corresponding question. Reads files and uses an AI model to answer each question using the respective file content as context.",
        "inputSchema": {
            "type": "object",
            "properties": {
                "tasks": {
                    "type": "array",
                    "items": {
                        "type": "object",
                        "properties": {
                            "file_path": {
                                "type": "string",
                                "description": "Relative path to the file (relative to workspace root)"
                            },
                            "question": {
                                "type": "string",
                                "description": "Question to ask about this file"
                            }
                        },
                        "required": ["file_path", "question"]
                    },
                    "description": "List of tasks, each containing a file path and a question"
                },
                "max_tokens": {
                    "type": "integer",
                    "default": 2000,
                    "description": "Maximum tokens for the AI response"
                },
                "max_workers": {
                    "type": "integer",
                    "default": 5,
                    "description": "Maximum number of concurrent model API requests"
                }
            },
            "required": ["tasks"]
        }
    },

    "download_files": {
        "name": "download_files",
        "description": "Download files from URLs to the workspace",
        "inputSchema": {
            "type": "object",
            "properties": {
                "urls": {
                    "type": "array",
                    "items": {"type": "string"},
                    "description": "List of URLs to download"
                },
                "target_directory": {
                    "type": "string",
                    "description": "Directory to save files"
                },
                "overwrite": {
                    "type": "boolean",
                    "default": False,
                    "description": "Whether to overwrite existing files"
                },
                "max_file_size_mb": {
                    "type": "integer",
                    "default": 100,
                    "description": "Maximum file size in MB"
                }
            },
            "required": ["urls"]
        }
    },

    "list_workspace": {
        "name": "list_workspace",
        "description": "List files and directories in the workspace",
        "inputSchema": {
            "type": "object",
            "properties": {
                "path": {
                    "type": "string",
                    "description": "Specify the directory path to list, using a relative path"
                },
                "recursive": {
                    "type": "boolean",
                    "default": False,
                    "description": "Whether to list recursively"
                },
                "include_hidden": {
                    "type": "boolean",
                    "default": False,
                    "description": "Whether to include hidden files"
                },
                "max_depth": {
                    "type": "integer",
                    "default": 3,
                    "description": "Maximum recursion depth"
                }
            },
            "required": []
        }
    },

    "str_replace_based_edit_tool": {
        "name": "str_replace_based_edit_tool",
        "description": "Create, view, and edit files with various operations",
        "inputSchema": {
            "type": "object",
            "properties": {
                "action": {
                    "type": "string",
                    "enum": ["create", "view", "str_replace", "insert", "append", "delete"],
                    "description": "Action to perform"
                },
                "file_path": {
                    "type": "string",
                    "description": "Path to the file"
                },
                "content": {
                    "type": "string",
                    "description": "Content for create/insert/append actions"
                },
                "old_str": {
                    "type": "string",
                    "description": "String to replace (for str_replace)"
                },
                "new_str": {
                    "type": "string",
                    "description": "Replacement string (for str_replace)"
                },
                "line_number": {
                    "type": "integer",
                    "description": "Line number for insert action"
                }
            },
            "required": ["action", "file_path"]
        }
    },

    "file_read": {
        "name": "file_read",
        "description": "Read file content",
        "inputSchema": {
            "type": "object",
            "properties": {
                "file_path": {
                    "type": "string",
                    "description": "Relative path to the file (relative to workspace root)"
                },
                "encoding": {
                    "type": "string",
                    "default": "utf-8",
                    "description": "File encoding"
                }
            },
            "required": ["file_path"]
        }
    },

    "load_json": {
        "name": "load_json",
        "description": "Read json format file",
        "inputSchema": {
            "type": "object",
            "properties": {
                "file_path": {
                    "type": "string",
                    "description": "Relative path to the file (relative to workspace root)"
                },
                "encoding": {
                    "type": "string",
                    "default": "utf-8",
                    "description": "File encoding"
                }
            },
            "required": ["file_path"]
        }
    },

    "file_write": {
        "name": "file_write",
        "description": "Write content to file",
        "inputSchema": {
            "type": "object",
            "properties": {
                "file_path": {
                    "type": "string",
                    "description": "Relative path to the file (relative to workspace root)"
                },
                "content": {
                    "type": "string",
                    "description": "Content to write"
                },
                "encoding": {
                    "type": "string",
                    "default": "utf-8",
                    "description": "File encoding"
                },
                "create_dirs": {
                    "type": "boolean",
                    "default": True,
                    "description": "Create parent directories"
                }
            },
            "required": ["file_path", "content"]
        }
    },

    "file_find_by_name": {
        "name": "file_find_by_name",
        "description": "Find files by name pattern",
        "inputSchema": {
            "type": "object",
            "properties": {
                "name_pattern": {
                    "type": "string",
                    "description": "Name pattern to search for"
                },
                "recursive": {
                    "type": "boolean",
                    "default": True,
                    "description": "Search recursively"
                },
                "case_sensitive": {
                    "type": "boolean",
                    "default": False,
                    "description": "Case sensitive search"
                },
                "max_results": {
                    "type": "integer",
                    "default": 100,
                    "description": "Maximum number of results"
                }
            },
            "required": ["name_pattern"]
        }
    },

    "concat_section_files": {
        "name": "concat_section_files",
        "description": "Concatenate the content of the saved section files into a single file",
        "inputSchema": {
            "type": "object",
            "properties": {
                "final_file_path": {
                    "type": "string",
                    "description": "The final file path to save the concatenated content, save the file in the workspace **under the relative path `./report/`**, and specify the final_file_path as `./report/final_report.md`"
                },
                "section_files": {
                    "type": "array",
                    "items": {
                        "type": "object",
                        "properties": {
                            "file_path": {
                                "type": "string",
                                "description": "Relative path to the saved section file"
                            }
                        },
                        "required": ["file_path"]
                    },
                    "description": "List of section files to concatenate"
                }
            },
            "required": ["section_files", "final_file_path"]
        }
    },

    "search_result_classifier": {
        "name": "search_result_classifier",
        "description": "Intelligently classify and organize search result files according to a structured outline for comprehensive long-form content generation. Analyzes files across fouer key dimensions (document time, source authority, core content, and task relevance) and assigns relevant files to appropriate outline sections. Files may be assigned to multiple sections when their content spans different topics.",
        "inputSchema": {
            "type": "object",
            "properties": {
                "outline": {
                    "type": "string",
                    "description": "The outline here must be consistent with the content and structure of the outline generated above"
                },
                "key_files": {
                    "type": "array",
                    "items": {
                        "type": "object",
                        "properties": {
                            "file_path": {
                                "type": "string",
                                "description": "Relative path to the file containing research content"
                            }
                        },
                        "required": ["file_path"]
                    },
                    "description": "List of research files to be classified according to the outline"
                },
                "model": {
                    "type": "string",
                    "default": "pangu_auto",
                    "description": "AI model to use for classification and organization"
                },
                "temperature": {
                    "type": "number",
                    "default": 0.3,
                    "description": "Creativity level for the AI classification (0-1)"
                },
                "max_tokens": {
                    "type": "integer",
                    "default": 2000,
                    "description": "Maximum tokens for the AI response"
                }
            },
            "required": ["key_files", "outline"]
        }
    },

    "document_extract": {
        "name": "document_extract",
        "description": "Multi-dimensional analysis of locally stored files using AI models. Evaluates each file across four key dimensions: web page time extraction, source authority assessment, task relevance evaluation, and core content summarization (~300 words). Provides structured document analysis for research and content evaluation purposes.",
        "inputSchema": {
            "type": "object",
            "properties": {
                "tasks": {
                    "type": "array",
                    "items": {
                        "type": "object",
                        "properties": {
                            "file_path": {
                                "type": "string",
                                "description": "Relative path to the file (relative to workspace root)"
                            },
                            "task": {
                                "type": "string",
                                "description": "The content of the currently executed subtask"
                            }
                        },
                        "required": ["file_path", "task"]
                    },
                    "description": "List of tasks, each containing a file path and the current task"
                },
                "model": {
                    "type": "string",
                    "default": "pangu_auto",
                    "description": "AI model to use for generating answers"
                },
                "temperature": {
                    "type": "number",
                    "default": 0.3,
                    "description": "Creativity level for the AI response (0-1)"
                },
                "max_tokens": {
                    "type": "integer",
                    "default": 2000,
                    "description": "Maximum tokens for the AI response"
                },
                "max_workers": {
                    "type": "integer",
                    "default": 5,
                    "description": "Maximum number of concurrent model API requests"
                }
            },
            "required": ["tasks"]
        }
    },

    "section_writer": {
        "name": "section_writer",
        "description": "Write the current chapter content based on given web information and chapter structure; also consider user questions, completed chapters, and overall outline to ensure content relevance while avoiding duplication or contradictions.",
        "inputSchema": {
            "type": "object",
            "properties": {
                "written_chapters_summary": {
                    "type": "string",
                    "description": "The summary of the written chapters, including the content of the chapters and the reflections on the chapters. Note that this field should be concatenated with the summaries of all previously written chapters with '\\n', and do not modify the original summary. For example, if the current chapter is the third chapter, the value of this field is 'chapter 1 summary \\n chapter 2 summary'. If not, the value is set to 'No previous chapters written yet.'"
                },
                "task_content": {
                    "type": "string",
                    "description": "Detailed description of some requirements for writing the current chapter and avoidance prompts. If there are reflections from the `think` tool on previously written chapters, they can be added to this field."
                },
                "user_query": {
                    "type": "string",
                    "description": "The user query, ensure the drafted content is highly relevant to the user's inquiry."
                },
                "current_chapter_outline": {
                    "type": "string",
                    "description": "This field represents the current chapter structure to be drafted. When composing the chapter content, do not modify content and bold formatting symbols of the existing structure's titles!!!"
                },
                "overall_outline": {
                    "type": "string",
                    "description": "This field represents the overall outline of the article. When drafting the chapter content, you should consider the overall outline to ensure the chapter content is consistent with the overall outline."
                },
                "target_file_path": {
                    "type": "string",
                    "description": "The path to save the chapter content"
                },
                "key_files": {
                    "type": "array",
                    "items": {
                        "type": "object",
                        "properties": {
                            "file_path": {
                                "type": "string",
                                "description": "Relative path to the file containing research content"
                            }
                        },
                        "required": ["file_path"]
                    },
                    "description": "These files are the source materials required for drafting the current chapter."
                },
                "model": {
                    "type": "string",
                    "default": "pangu_auto",
                    "description": "AI model to use for classification and organization"
                },
                "temperature": {
                    "type": "number",
                    "default": 0.3,
                    "description": "Creativity level for the AI classification (0-1)"
                },
                "max_tokens": {
                    "type": "integer",
                    "default": 5000,
                    "description": "Maximum tokens for the AI response"
                },
            },
            "required": ["user_query", "current_chapter_outline", "overall_outline", "target_file_path", "key_files"]
        }
    },

    "file_stats": {
        "name": "file_stats",
        "description": "Get comprehensive file statistics without reading full content - perfect for deciding reading strategy",
        "inputSchema": {
            "type": "object",
            "properties": {
                "file_path": {
                    "type": "string",
                    "description": "Path to the file (relative to workspace)"
                }
            },
            "required": ["file_path"]
        }
    },

    "file_read_lines": {
        "name": "file_read_lines",
        "description": "Read specific line ranges from a file without loading entire file - perfect for large files",
        "inputSchema": {
            "type": "object",
            "properties": {
                "file_path": {
                    "type": "string",
                    "description": "Path to the file"
                },
                "start_line": {
                    "type": "integer",
                    "default": 1,
                    "description": "Starting line number (1-based)"
                },
                "end_line": {
                    "type": "integer",
                    "description": "Ending line number (1-based, None for end of file)"
                },
                "max_lines": {
                    "type": "integer",
                    "default": 1000,
                    "description": "Maximum number of lines to read (safety limit)"
                }
            },
            "required": ["file_path"]
        }
    }

    # NOTE: Task assignment tool schemas removed - these are now built-in methods of PlannerAgent
    # to avoid circular dependency issues with sub-agents trying to create MCP client connections
}


# ================ MAIN INTERFACE ================

def get_tool_schemas() -> Dict[str, Any]:
    """Get all tool schemas for MCP registration"""
    return MCP_TOOL_SCHEMAS