chat_handler.py

"""
Chat handling logic for Universal MCP Client - Fixed Version with File Upload Support
"""
import asyncio
import re
import logging
import traceback
from datetime import datetime
from typing import Dict, Any, List, Tuple, Optional
import gradio as gr
from gradio import ChatMessage
from gradio_client import Client
import time
import json
import httpx

from config import AppConfig
from mcp_client import UniversalMCPClient

logger = logging.getLogger(__name__)

class ChatHandler:
    """Handles chat interactions with HF Inference Providers and MCP servers using ChatMessage dataclass"""
    
    def __init__(self, mcp_client: UniversalMCPClient):
        self.mcp_client = mcp_client
        # Initialize the file uploader client for converting local files to public URLs
        try:
            self.uploader_client = Client("abidlabs/file-uploader")
            logger.info("✅ File uploader client initialized")
        except Exception as e:
            logger.error(f"Failed to initialize file uploader: {e}")
            self.uploader_client = None
    
    def _upload_file_to_gradio_server(self, file_path: str) -> str:
        """Upload a file to the Gradio server and get a public URL"""
        if not self.uploader_client:
            logger.error("File uploader client not initialized")
            return file_path
        
        try:
            # Open file in binary mode as your peer discovered
            with open(file_path, "rb") as f_:
                files = [("files", (file_path.split("/")[-1], f_))]
                r = httpx.post(
                    self.uploader_client.upload_url,
                    files=files,
                )
            r.raise_for_status()
            result = r.json()
            uploaded_path = result[0]
            # Construct the full public URL
            public_url = f"{self.uploader_client.src}/gradio_api/file={uploaded_path}"
            logger.info(f"✅ Uploaded {file_path} -> {public_url}")
            return public_url
        except Exception as e:
            logger.error(f"Failed to upload file {file_path}: {e}")
            return file_path  # Return original path as fallback
    
    def process_multimodal_message(self, message: Dict[str, Any], history: List) -> Tuple[List[ChatMessage], Dict[str, Any]]:
        """Enhanced MCP chat function with multimodal input support and ChatMessage formatting"""
        
        if not self.mcp_client.hf_client:
            error_msg = "❌ HuggingFace token not configured. Please set HF_TOKEN environment variable or login."
            history.append(ChatMessage(role="user", content=error_msg))
            history.append(ChatMessage(role="assistant", content=error_msg))
            return history, gr.MultimodalTextbox(value=None, interactive=False)
        
        if not self.mcp_client.current_provider or not self.mcp_client.current_model:
            error_msg = "❌ Please select an inference provider and model first."
            history.append(ChatMessage(role="user", content=error_msg))
            history.append(ChatMessage(role="assistant", content=error_msg))
            return history, gr.MultimodalTextbox(value=None, interactive=False)
        
        # Initialize variables for error handling
        user_text = ""
        user_files = []
        uploaded_file_urls = []  # Store uploaded file URLs
        self.file_url_mapping = {}  # Map local paths to uploaded URLs

        try:
            # Handle multimodal input - message is a dict with 'text' and 'files'
            user_text = message.get("text", "") if message else ""
            user_files = message.get("files", []) if message else []
            
            # Handle case where message might be a string (backward compatibility)
            if isinstance(message, str):
                user_text = message
                user_files = []
            
            logger.info(f"💬 Processing multimodal message:")
            logger.info(f"  📝 Text: {user_text}")
            logger.info(f"  📁 Files: {len(user_files)} files uploaded")
            logger.info(f"  📋 History type: {type(history)}, length: {len(history)}")
            
            # Convert history to ChatMessage objects if needed
            converted_history = []
            for i, msg in enumerate(history):
                try:
                    if isinstance(msg, dict):
                        # Convert dict to ChatMessage for internal processing
                        logger.info(f"  📝 Converting dict message {i}: {msg.get('role', 'unknown')}")
                        converted_history.append(ChatMessage(
                            role=msg.get('role', 'assistant'),
                            content=msg.get('content', ''),
                            metadata=msg.get('metadata', None)
                        ))
                    else:
                        # Already a ChatMessage
                        logger.info(f"  ✅ ChatMessage {i}: {getattr(msg, 'role', 'unknown')}")
                        converted_history.append(msg)
                except Exception as conv_error:
                    logger.error(f"Error converting message {i}: {conv_error}")
                    logger.error(f"Message content: {msg}")
                    # Skip problematic messages
                    continue
            
            history = converted_history
            
            # Upload files and get public URLs
            for file_path in user_files:
                logger.info(f"  📄 Local File: {file_path}")
                try:
                    # Upload file to get public URL
                    uploaded_url = self._upload_file_to_gradio_server(file_path)
                    # Store the mapping
                    self.file_url_mapping[file_path] = uploaded_url
                    uploaded_file_urls.append(uploaded_url)
                    logger.info(f"  ✅ Uploaded File URL: {uploaded_url}")
                    
                    # Add to history with public URL
                    history.append(ChatMessage(role="user", content={"path": uploaded_url}))
                except Exception as upload_error:
                    logger.error(f"Failed to upload file {file_path}: {upload_error}")
                    # Fallback to local path with warning
                    history.append(ChatMessage(role="user", content={"path": file_path}))
                    logger.warning(f"⚠️ Using local path for {file_path} - MCP servers may not be able to access it")
            
            # Add text message if provided
            if user_text and user_text.strip():
                history.append(ChatMessage(role="user", content=user_text))
            
            # If no text and no files, return early
            if not user_text.strip() and not user_files:
                return history, gr.MultimodalTextbox(value=None, interactive=False)
            
            # Create messages for HF Inference API
            messages = self._prepare_hf_messages(history, uploaded_file_urls)
            
            # Process the chat and get structured responses
            response_messages = self._call_hf_api(messages, uploaded_file_urls)
            
            # Add all response messages to history
            history.extend(response_messages)
            
            return history, gr.MultimodalTextbox(value=None, interactive=False)
            
        except Exception as e:
            error_msg = f"❌ Error: {str(e)}"
            logger.error(f"Chat error: {e}")
            logger.error(traceback.format_exc())
            
            # Add user input to history if it exists
            if user_text and user_text.strip():
                history.append(ChatMessage(role="user", content=user_text))
            if user_files:
                for file_path in user_files:
                    history.append(ChatMessage(role="user", content={"path": file_path}))
                    
            history.append(ChatMessage(role="assistant", content=error_msg))
            return history, gr.MultimodalTextbox(value=None, interactive=False)
    
    def _prepare_hf_messages(self, history: List, uploaded_file_urls: List[str] = None) -> List[Dict[str, Any]]:
        """Convert history (ChatMessage or dict) to HF OpenAI-compatible format with multimodal support"""
        messages: List[Dict[str, Any]] = []
        
        # Get optimal context settings for current model/provider
        if self.mcp_client.current_model and self.mcp_client.current_provider:
            context_settings = AppConfig.get_optimal_context_settings(
                self.mcp_client.current_model, 
                self.mcp_client.current_provider,
                len(self.mcp_client.get_enabled_servers())
            )
            max_history = context_settings['recommended_history_limit']
        else:
            max_history = 20  # Fallback
        
        # Convert history to HF API format (text only for context)
        recent_history = history[-max_history:] if len(history) > max_history else history
        
        last_role = None
        is_gpt_oss = AppConfig.is_gpt_oss_model(self.mcp_client.current_model) if self.mcp_client.current_model else False
        for msg in recent_history:
            # Handle both ChatMessage objects and dictionary format for backward compatibility
            if hasattr(msg, 'role'):  # ChatMessage object
                role = msg.role
                content = msg.content
            elif isinstance(msg, dict) and 'role' in msg:  # Dictionary format
                role = msg.get('role')
                content = msg.get('content')
            else:
                continue  # Skip invalid messages
                
            if role == "user":
                if is_gpt_oss:
                    # Text-only content for GPT-OSS (no multimodal parts)
                    if isinstance(content, dict) and "path" in content:
                        file_path = content.get("path", "")
                        # Omit media content; optionally note the upload as text
                        text_piece = ""
                        # Choose to ignore media fully to avoid confusing the model
                    elif isinstance(content, (list, tuple)):
                        text_piece = f"[List: {str(content)[:50]}...]"
                    elif content is None:
                        text_piece = "[Empty]"
                    else:
                        text_piece = str(content)

                    if messages and last_role == "user" and isinstance(messages[-1].get("content"), str):
                        # Concatenate text
                        if text_piece:
                            messages[-1]["content"] = (messages[-1]["content"] + "\n" + text_piece) if messages[-1]["content"] else text_piece
                    else:
                        messages.append({"role": "user", "content": text_piece})
                    last_role = "user"
                else:
                    # Build multimodal user messages with parts (for non-GPT-OSS)
                    part = None
                    if isinstance(content, dict) and "path" in content:
                        file_path = content.get("path", "")
                        if isinstance(file_path, str) and file_path.startswith("http") and AppConfig.is_image_file(file_path):
                            part = {"type": "image_url", "image_url": {"url": file_path}}
                        else:
                            part = {"type": "text", "text": f"[File: {file_path}]"}
                    elif isinstance(content, (list, tuple)):
                        part = {"type": "text", "text": f"[List: {str(content)[:50]}...]"}
                    elif content is None:
                        part = {"type": "text", "text": "[Empty]"}
                    else:
                        part = {"type": "text", "text": str(content)}

                    if messages and last_role == "user" and isinstance(messages[-1].get("content"), list):
                        messages[-1]["content"].append(part)
                    elif messages and last_role == "user" and isinstance(messages[-1].get("content"), str):
                        # Convert existing string content to parts and append
                        existing_text = messages[-1]["content"]
                        messages[-1]["content"] = [{"type": "text", "text": existing_text}, part]
                    else:
                        messages.append({"role": "user", "content": [part]})
                    last_role = "user"

            elif role == "assistant":
                # Assistant content remains text for chat.completions API
                if isinstance(content, dict):
                    text = f"[Object: {str(content)[:50]}...]"
                elif isinstance(content, (list, tuple)):
                    text = f"[List: {str(content)[:50]}...]"
                elif content is None:
                    text = "[Empty]"
                else:
                    text = str(content)
                messages.append({"role": "assistant", "content": text})
                last_role = "assistant"
        
        return messages
    
    def _call_hf_api(self, messages: List[Dict[str, Any]], uploaded_file_urls: List[str] = None) -> List[ChatMessage]:
        """Call HuggingFace Inference API and return structured ChatMessage responses"""
        
        # Check if we have enabled MCP servers to use
        enabled_servers = self.mcp_client.get_enabled_servers()
        if not enabled_servers:
            return self._call_hf_without_mcp(messages)
        else:
            return self._call_hf_with_mcp(messages, uploaded_file_urls)
    
    def _call_hf_without_mcp(self, messages: List[Dict[str, Any]]) -> List[ChatMessage]:
        """Call HF Inference API without MCP servers. Streams tokens for faster feedback."""
        logger.info("💬 No MCP servers available, using streaming HF Inference chat when possible")

        system_prompt = self._get_native_system_prompt()

        # Add system prompt to messages
        if messages and messages[0].get("role") == "system":
            messages[0]["content"] = system_prompt + "\n\n" + messages[0]["content"]
        else:
            messages.insert(0, {"role": "system", "content": system_prompt})

        # Get optimal token settings
        if self.mcp_client.current_model and self.mcp_client.current_provider:
            context_settings = AppConfig.get_optimal_context_settings(
                self.mcp_client.current_model,
                self.mcp_client.current_provider,
                0  # No MCP servers
            )
            max_tokens = context_settings['max_response_tokens']
        else:
            max_tokens = 8192

        # Try streaming first; fall back to non-streaming on error
        try:
            stream = self.mcp_client.generate_chat_completion_stream(messages, **{"max_tokens": max_tokens})
            accumulated = ""
            for chunk in stream:
                try:
                    delta = chunk.choices[0].delta.content or ""
                except Exception:
                    # Some SDK variants stream as message deltas differently
                    delta = getattr(getattr(chunk.choices[0], "delta", None), "content", "") or ""
                if delta:
                    accumulated += delta
            if not accumulated:
                accumulated = "I understand your request and I'm here to help."
            return [ChatMessage(role="assistant", content=accumulated)]
        except Exception as e:
            logger.warning(f"Streaming failed, retrying without stream: {e}")
            try:
                response = self.mcp_client.generate_chat_completion(messages, **{"max_tokens": max_tokens})
                response_text = response.choices[0].message.content
                if not response_text:
                    response_text = "I understand your request and I'm here to help."
                return [ChatMessage(role="assistant", content=response_text)]
            except Exception as e2:
                logger.error(f"HF Inference API call failed: {e2}")
                return [ChatMessage(role="assistant", content=f"❌ API call failed: {str(e2)}")]
    
    def _call_hf_with_mcp(self, messages: List[Dict[str, Any]], uploaded_file_urls: List[str] = None) -> List[ChatMessage]:
        """Call HF Inference API with MCP servers and return structured responses"""
        
        # Enhanced system prompt with multimodal and MCP instructions
        system_prompt = self._get_mcp_system_prompt(uploaded_file_urls)
        
        # Add system prompt to messages
        if messages and messages[0].get("role") == "system":
            messages[0]["content"] = system_prompt + "\n\n" + messages[0]["content"]
        else:
            messages.insert(0, {"role": "system", "content": system_prompt})
        
        # Get optimal token settings
        enabled_servers = self.mcp_client.get_enabled_servers()
        if self.mcp_client.current_model and self.mcp_client.current_provider:
            context_settings = AppConfig.get_optimal_context_settings(
                self.mcp_client.current_model, 
                self.mcp_client.current_provider,
                len(enabled_servers)
            )
            max_tokens = context_settings['max_response_tokens']
        else:
            max_tokens = 8192
        
        # Debug logging
        logger.info(f"📤 Sending {len(messages)} messages to HF Inference API")
        logger.info(f"🔧 Using {len(self.mcp_client.servers)} MCP servers")
        logger.info(f"🤖 Model: {self.mcp_client.current_model} via {self.mcp_client.current_provider}")
        logger.info(f"📏 Max tokens: {max_tokens}")
        
        start_time = time.time()
        
        try:
            # Pass file mapping to MCP client
            if hasattr(self, 'file_url_mapping'):
                self.mcp_client.chat_handler_file_mapping = self.file_url_mapping
                
            # Call HF Inference with MCP tool support - using optimal max_tokens
            response = self.mcp_client.generate_chat_completion_with_mcp_tools(messages, **{"max_tokens": max_tokens})
            
            return self._process_hf_response(response, start_time)
        except Exception as e:
            logger.error(f"HF Inference API call with MCP failed: {e}")
            return [ChatMessage(role="assistant", content=f"❌ API call failed: {str(e)}")]
    
    def _process_hf_response(self, response, start_time: float) -> List[ChatMessage]:
        """Process HF Inference response with simplified media handling and nested errors"""
        chat_messages = []
        
        try:
            response_text = response.choices[0].message.content
            
            if not response_text:
                response_text = "I understand your request and I'm here to help."
            
            # Check if this response includes tool execution info
            if hasattr(response, '_tool_execution'):
                tool_info = response._tool_execution
                logger.info(f"🔧 Processing response with tool execution: {tool_info}")
                
                duration = round(time.time() - start_time, 2)
                tool_id = f"tool_{tool_info['tool']}_{int(time.time())}"
                
                if tool_info['success']:
                    tool_result = str(tool_info['result'])
                    
                    # Extract media URL if present
                    media_url = self._extract_media_url(tool_result, tool_info.get('server', ''))
                    
                    # Create tool usage metadata message
                    chat_messages.append(ChatMessage(
                        role="assistant",
                        content="",
                        metadata={
                            "title": f"🔧 Used {tool_info['tool']}",
                            "status": "done",
                            "duration": duration,
                            "id": tool_id
                        }
                    ))
                    
                    # Add nested success message with the raw result
                    if media_url:
                        result_preview = f"✅ Successfully generated media\nURL: {media_url[:100]}..."
                    else:
                        result_preview = f"✅ Tool executed successfully\nResult: {tool_result[:200]}..."
                    
                    chat_messages.append(ChatMessage(
                        role="assistant",
                        content=result_preview,
                        metadata={
                            "title": "📊 Server Response",
                            "parent_id": tool_id,
                            "status": "done"
                        }
                    ))
                    
                    # Add LLM's descriptive text if present (before media)
                    if response_text and not response_text.startswith('{"use_tool"'):
                        # Clean the response text by removing URLs and tool JSON
                        clean_response = response_text
                        if media_url and media_url in clean_response:
                            clean_response = clean_response.replace(media_url, "").strip()
                        
                        # Remove any remaining JSON tool call patterns
                        clean_response = re.sub(r'\{"use_tool"[^}]+\}', '', clean_response).strip()
                        
                        # Remove all markdown link/image syntax completely
                        clean_response = re.sub(r'!\[([^\]]*)\]\([^)]*\)', '', clean_response)  # Remove image markdown
                        clean_response = re.sub(r'\[([^\]]*)\]\([^)]*\)', '', clean_response)   # Remove link markdown  
                        clean_response = re.sub(r'!\[([^\]]*)\]', '', clean_response)           # Remove broken image refs
                        clean_response = re.sub(r'\[([^\]]*)\]', '', clean_response)            # Remove broken link refs
                        clean_response = re.sub(r'\(\s*\)', '', clean_response)                 # Remove empty parentheses
                        clean_response = clean_response.strip()                                 # Final strip

                        # Only add if there's meaningful text left after cleaning
                        if clean_response and len(clean_response) > 10:
                            chat_messages.append(ChatMessage(
                                role="assistant",
                                content=clean_response
                            ))
                    # Handle media content if present
                    if media_url:
                        # Add media as a separate message - Gradio will auto-detect type
                        chat_messages.append(ChatMessage(
                            role="assistant",
                            content={"path": media_url}
                        ))
                    else:
                        # No media URL found, check if we need to show non-media result
                        if not response_text or response_text.startswith('{"use_tool"'):
                            # Only show result if there wasn't descriptive text from LLM
                            if len(tool_result) > 500:
                                result_preview = f"Operation completed successfully. Result preview: {tool_result[:500]}..."
                            else:
                                result_preview = f"Operation completed successfully. Result: {tool_result}"
                            
                            chat_messages.append(ChatMessage(
                                role="assistant",
                                content=result_preview
                            ))
                            
                else:
                    # Tool execution failed
                    error_details = tool_info['result']
                    
                    # Create main tool message with pending status (error reflected in content)
                    chat_messages.append(ChatMessage(
                        role="assistant",
                        content="",
                        metadata={
                            "title": f"❌ Used {tool_info['tool']}",
                            "status": "pending",
                            "duration": duration,
                            "id": tool_id
                        }
                    ))
                    
                    # Add nested error response from server
                    chat_messages.append(ChatMessage(
                        role="assistant",
                        content=f"❌ Tool execution failed\n```\n{error_details}\n```",
                        metadata={
                            "title": "📊 Server Response",
                            "parent_id": tool_id,
                            "status": "done"
                        }
                    ))
                    
                    # Add suggestions as another nested message
                    chat_messages.append(ChatMessage(
                        role="assistant",
                        content="**Suggestions:**\n• Try modifying your request slightly\n• Wait a moment and try again\n• Use a different MCP server if available",
                        metadata={
                            "title": "💡 Possible Solutions",
                            "parent_id": tool_id,
                            "status": "done"
                        }
                    ))
            else:
                # No tool usage, just return the response
                chat_messages.append(ChatMessage(
                    role="assistant",
                    content=response_text
                ))
            
        except Exception as e:
            logger.error(f"Error processing HF response: {e}")
            logger.error(traceback.format_exc())
            chat_messages.append(ChatMessage(
                role="assistant",
                content="I understand your request and I'm here to help."
            ))
        
        return chat_messages

    def process_multimodal_message_stream(self, message: Dict[str, Any], history: List):
        """Generator that streams assistant output to the UI as it arrives.
        - Streams for plain LLM chats
        - Streams initial planning/tool JSON for MCP flows, executes tool, then streams final answer
        - Attempts to surface reasoning/thinking traces when available
        """
        try:
            # Pre-checks
            if not self.mcp_client.hf_client:
                error_msg = "❌ HuggingFace token not configured. Please set HF_TOKEN environment variable or login."
                history.append(ChatMessage(role="assistant", content=error_msg))
                yield history, gr.MultimodalTextbox(value=None, interactive=False)
                return

            if not self.mcp_client.current_provider or not self.mcp_client.current_model:
                error_msg = "❌ Please select an inference provider and model first."
                history.append(ChatMessage(role="assistant", content=error_msg))
                yield history, gr.MultimodalTextbox(value=None, interactive=False)
                return

            # Parse user input
            user_text = message.get("text", "") if message else ""
            user_files = message.get("files", []) if message else []

            # Upload files and update history similarly to non-stream path
            self.file_url_mapping = {}
            uploaded_file_urls: List[str] = []
            if isinstance(message, str):
                user_text = message
                user_files = []

            if user_files:
                for file_path in user_files:
                    try:
                        uploaded_url = self._upload_file_to_gradio_server(file_path)
                        self.file_url_mapping[file_path] = uploaded_url
                        uploaded_file_urls.append(uploaded_url)
                        history.append(ChatMessage(role="user", content={"path": uploaded_url}))
                    except Exception:
                        history.append(ChatMessage(role="user", content={"path": file_path}))

            if user_text and user_text.strip():
                history.append(ChatMessage(role="user", content=user_text))

            if not user_text.strip() and not user_files:
                yield history, gr.MultimodalTextbox(value=None, interactive=False)
                return

            # Prepare messages for HF
            messages = self._prepare_hf_messages(history, uploaded_file_urls)
            # Choose streaming path based on MCP servers
            if self.mcp_client.get_enabled_servers():
                # Stream with MCP planning/tool execution
                yield from self._stream_with_mcp(messages, uploaded_file_urls, history)
            else:
                # Plain LLM streaming with optional thinking trace
                yield from self._stream_without_mcp(messages, history)
        except Exception as e:
            history.append(ChatMessage(role="assistant", content=f"❌ Error: {str(e)}"))
            yield history, gr.MultimodalTextbox(value=None, interactive=True)

    def _stream_without_mcp(self, messages: List[Dict[str, Any]], history: List):
        """Stream tokens for plain LLM chats; attempts to surface reasoning traces if available."""
        # Add system prompt
        system_prompt = self._get_native_system_prompt()
        if messages and messages[0].get("role") == "system":
            messages[0]["content"] = system_prompt + "\n\n" + messages[0]["content"]
        else:
            messages.insert(0, {"role": "system", "content": system_prompt})

        # Compute max tokens
        if self.mcp_client.current_model and self.mcp_client.current_provider:
            ctx = AppConfig.get_optimal_context_settings(
                self.mcp_client.current_model, self.mcp_client.current_provider, 0
            )
            max_tokens = ctx["max_response_tokens"]
        else:
            max_tokens = 8192

        # Insert placeholders: optional thinking + main assistant
        thinking_index = None
        # Prepare a thinking message only when we actually receive thinking tokens
        history.append(ChatMessage(role="assistant", content=""))
        main_index = len(history) - 1
        yield history, gr.MultimodalTextbox(value=None, interactive=False)

        accumulated = ""
        thinking_accum = ""
        try:
            stream = self.mcp_client.generate_chat_completion_stream(messages, **{"max_tokens": max_tokens})
            for chunk in stream:
                delta = getattr(chunk.choices[0], "delta", None)
                # Reasoning/thinking traces (best-effort extraction)
                reason_delta = None
                if delta is not None:
                    # Some providers expose .reasoning or .thinking
                    reason_delta = (
                        getattr(delta, "reasoning", None)
                        or getattr(delta, "thinking", None)
                    )
                if reason_delta:
                    thinking_accum += str(reason_delta)
                    if thinking_index is None:
                        history.insert(main_index, ChatMessage(
                            role="assistant",
                            content=f"{thinking_accum}",
                            metadata={"title": "🧠 Reasoning", "status": "pending"}
                        ))
                        thinking_index = main_index
                        main_index += 1
                    else:
                        history[thinking_index] = ChatMessage(
                            role="assistant",
                            content=f"{thinking_accum}",
                            metadata={"title": "🧠 Reasoning", "status": "pending"}
                        )

                # Main content
                delta_text = ""
                try:
                    delta_text = delta.content or ""
                except Exception:
                    delta_text = getattr(delta, "content", "") or ""
                if not delta_text:
                    yield history, gr.MultimodalTextbox(value=None, interactive=False)
                    continue
                accumulated += delta_text
                history[main_index] = ChatMessage(role="assistant", content=accumulated)
                yield history, gr.MultimodalTextbox(value=None, interactive=False)
        except Exception as e:
            # Fallback to non-stream
            try:
                resp = self.mcp_client.generate_chat_completion(messages, **{"max_tokens": max_tokens})
                final_text = resp.choices[0].message.content or "I understand your request and I'm here to help."
                history[main_index] = ChatMessage(role="assistant", content=final_text)
                yield history, gr.MultimodalTextbox(value=None, interactive=True)
                return
            except Exception as e2:
                history[main_index] = ChatMessage(role="assistant", content=f"❌ API call failed: {str(e2)}")
                yield history, gr.MultimodalTextbox(value=None, interactive=True)
                return

        # Final yield
        yield history, gr.MultimodalTextbox(value=None, interactive=True)

    def _stream_with_mcp(self, messages: List[Dict[str, Any]], uploaded_file_urls: List[str], history: List):
        """Stream initial planning/tool JSON, execute MCP tool, then stream final response."""
        # Enhanced system prompt with MCP guidance
        system_prompt = self._get_mcp_system_prompt(uploaded_file_urls)
        if messages and messages[0].get("role") == "system":
            messages[0]["content"] = system_prompt + "\n\n" + messages[0]["content"]
        else:
            messages.insert(0, {"role": "system", "content": system_prompt})

        # Compute max tokens taking enabled servers into account
        enabled_servers = self.mcp_client.get_enabled_servers()
        if self.mcp_client.current_model and self.mcp_client.current_provider:
            ctx = AppConfig.get_optimal_context_settings(
                self.mcp_client.current_model, self.mcp_client.current_provider, len(enabled_servers)
            )
            max_tokens = ctx["max_response_tokens"]
        else:
            max_tokens = 8192

        # Placeholders: planning/tool JSON + main assistant
        planning_index = None
        thinking_index = None
        history.append(ChatMessage(role="assistant", content=""))
        main_index = len(history) - 1
        yield history, gr.MultimodalTextbox(value=None, interactive=False)

        text_accum = ""
        tool_json_accum = ""
        in_tool_json = False
        tool_json_detected = False
        try:
            stream = self.mcp_client.generate_chat_completion_stream(messages, **{"max_tokens": max_tokens})
            for chunk in stream:
                delta = getattr(chunk.choices[0], "delta", None)
                # Optional reasoning
                reason_delta = None
                if delta is not None:
                    reason_delta = (
                        getattr(delta, "reasoning", None)
                        or getattr(delta, "thinking", None)
                    )
                if reason_delta:
                    if thinking_index is None:
                        history.insert(main_index, ChatMessage(
                            role="assistant",
                            content=str(reason_delta),
                            metadata={"title": "🧠 Reasoning", "status": "pending"}
                        ))
                        thinking_index = main_index
                        main_index += 1
                    else:
                        history[thinking_index] = ChatMessage(
                            role="assistant",
                            content=(history[thinking_index].content + str(reason_delta)),
                            metadata={"title": "🧠 Reasoning", "status": "pending"}
                        )

                # Main content streaming and tool JSON detection (content-based JSON protocol)
                piece = ""
                try:
                    piece = delta.content or ""
                except Exception:
                    piece = getattr(delta, "content", "") or ""
                if not piece:
                    yield history, gr.MultimodalTextbox(value=None, interactive=False)
                    continue

                # Detect start of tool JSON
                if not tool_json_detected and '{"use_tool":' in piece:
                    in_tool_json = True
                    tool_json_detected = True
                if in_tool_json:
                    tool_json_accum += piece
                    # Initialize planning message
                    if planning_index is None:
                        history.insert(main_index, ChatMessage(
                            role="assistant",
                            content=tool_json_accum,
                            metadata={"title": "🔧 Tool call (planning)", "status": "pending"}
                        ))
                        planning_index = main_index
                        main_index += 1
                    else:
                        history[planning_index] = ChatMessage(
                            role="assistant",
                            content=tool_json_accum,
                            metadata={"title": "🔧 Tool call (planning)", "status": "pending"}
                        )

                    # Try to reconstruct JSON when braces close
                    reconstructed = self.mcp_client._reconstruct_json_from_start(tool_json_accum)
                    if reconstructed:
                        # We have a complete JSON
                        in_tool_json = False
                        # Clean planning content to the reconstructed JSON (for clarity)
                        history[planning_index] = ChatMessage(
                            role="assistant",
                            content=reconstructed,
                            metadata={"title": "🔧 Tool call", "status": "done"}
                        )
                        yield history, gr.MultimodalTextbox(value=None, interactive=False)

                        # Execute tool now
                        import json as _json
                        try:
                            tool_req = _json.loads(reconstructed)
                        except Exception:
                            tool_req = None
                        if tool_req and tool_req.get("use_tool"):
                            server_name = tool_req.get("server")
                            tool_name = tool_req.get("tool")
                            arguments = tool_req.get("arguments", {})

                            # Status message
                            exec_msg = ChatMessage(
                                role="assistant",
                                content=f"Executing {tool_name} on {server_name}…",
                                metadata={"title": "🔧 Tool execution", "status": "pending"}
                            )
                            history.insert(main_index, exec_msg)
                            exec_index = main_index
                            main_index += 1
                            yield history, gr.MultimodalTextbox(value=None, interactive=False)

                            # Replace any local paths with uploaded URLs
                            if hasattr(self, 'file_url_mapping'):
                                for k, v in list(arguments.items()):
                                    if isinstance(v, str) and v.startswith('/tmp/gradio/'):
                                        for lpath, url in self.file_url_mapping.items():
                                            if lpath in v or v in lpath:
                                                arguments[k] = url
                                                break

                            # Run tool (blocking)
                            def _run_tool():
                                loop = asyncio.new_event_loop()
                                asyncio.set_event_loop(loop)
                                try:
                                    return loop.run_until_complete(
                                        self.mcp_client.call_mcp_tool_async(server_name, tool_name, arguments)
                                    )
                                finally:
                                    loop.close()

                            success, result = _run_tool()
                            # Update exec message
                            if success:
                                content = str(result)
                                history[exec_index] = ChatMessage(
                                    role="assistant",
                                    content=content if len(content) < 800 else content[:800] + "…",
                                    metadata={"title": "📊 Server Response", "status": "done"}
                                )
                            else:
                                history[exec_index] = ChatMessage(
                                    role="assistant",
                                    content=f"❌ Tool failed: {result}",
                                    metadata={"title": "📊 Server Response", "status": "done"}
                                )
                            yield history, gr.MultimodalTextbox(value=None, interactive=False)

                            # Start final streamed response using tool result
                            final_messages = messages.copy()
                            # Remove tools instruction portion from system if present
                            if final_messages and final_messages[0].get("role") == "system":
                                sys_text = final_messages[0]["content"]
                                cut = sys_text.split("You have access to the following MCP tools:")[0].strip()
                                final_messages[0]["content"] = cut
                            # Add prior assistant (planning) and user tool result follow-up
                            final_messages.append({"role": "assistant", "content": text_accum})
                            final_messages.append({
                                "role": "user",
                                "content": f"Tool '{tool_name}' from server '{server_name}' completed. Result: {result}. Please provide a helpful response."
                            })

                            # Stream final answer into main message
                            final_accum = ""
                            try:
                                final_stream = self.mcp_client.generate_chat_completion_stream(final_messages, **{"max_tokens": max_tokens})
                                for fchunk in final_stream:
                                    fdelta = getattr(fchunk.choices[0], "delta", None)
                                    ftext = getattr(fdelta, "content", "") if fdelta is not None else ""
                                    if not ftext:
                                        yield history, gr.MultimodalTextbox(value=None, interactive=False)
                                        continue
                                    final_accum += ftext
                                    history[main_index] = ChatMessage(role="assistant", content=(text_accum + final_accum))
                                    yield history, gr.MultimodalTextbox(value=None, interactive=False)
                            except Exception:
                                # Fallback non-stream finalization
                                try:
                                    fresp = self.mcp_client.generate_chat_completion(final_messages, **{"max_tokens": max_tokens})
                                    ftxt = fresp.choices[0].message.content or ""
                                    history[main_index] = ChatMessage(role="assistant", content=(text_accum + ftxt))
                                    yield history, gr.MultimodalTextbox(value=None, interactive=True)
                                    return
                                except Exception as e3:
                                    history[main_index] = ChatMessage(role="assistant", content=(text_accum + f"\n❌ Finalization failed: {e3}"))
                                    yield history, gr.MultimodalTextbox(value=None, interactive=True)
                                    return

                            # Done
                            yield history, gr.MultimodalTextbox(value=None, interactive=True)
                            return
                else:
                    # Normal assistant visible text outside of tool JSON
                    text_accum += piece
                    history[main_index] = ChatMessage(role="assistant", content=text_accum)
                yield history, gr.MultimodalTextbox(value=None, interactive=False)
        except Exception as e:
            # Fallback: Use non-streaming MCP path
            responses = self._call_hf_with_mcp(messages, uploaded_file_urls)
            history.extend(responses)
            yield history, gr.MultimodalTextbox(value=None, interactive=True)
            return

        # If we streamed without any tool usage, finalize
        yield history, gr.MultimodalTextbox(value=None, interactive=True)
    
    def _extract_media_url(self, result_text: str, server_name: str) -> Optional[str]:
        """Extract media URL from MCP response with improved pattern matching"""
        if not isinstance(result_text, str):
            return None
        
        logger.info(f"🔍 Extracting media from result: {result_text[:500]}...")
        
        # Try JSON parsing first
        try:
            if result_text.strip().startswith('[') or result_text.strip().startswith('{'):
                data = json.loads(result_text.strip())
                
                # Handle array format
                if isinstance(data, list) and len(data) > 0:
                    item = data[0]
                    if isinstance(item, dict):
                        # Check for nested media structure
                        for media_type in ['audio', 'video', 'image']:
                            if media_type in item and isinstance(item[media_type], dict):
                                if 'url' in item[media_type]:
                                    url = item[media_type]['url'].strip('\'"')
                                    logger.info(f"🎯 Found {media_type} URL in JSON: {url}")
                                    return url
                        # Check for direct URL
                        if 'url' in item:
                            url = item['url'].strip('\'"')
                            logger.info(f"🎯 Found direct URL in JSON: {url}")
                            return url
                
                # Handle object format
                elif isinstance(data, dict):
                    # Check for nested media structure
                    for media_type in ['audio', 'video', 'image']:
                        if media_type in data and isinstance(data[media_type], dict):
                            if 'url' in data[media_type]:
                                url = data[media_type]['url'].strip('\'"')
                                logger.info(f"🎯 Found {media_type} URL in JSON: {url}")
                                return url
                    # Check for direct URL
                    if 'url' in data:
                        url = data['url'].strip('\'"')
                        logger.info(f"🎯 Found direct URL in JSON: {url}")
                        return url
                        
        except json.JSONDecodeError:
            pass
        
        # Check for Gradio file URLs (common pattern)
        gradio_patterns = [
            r'https://[^/]+\.hf\.space/gradio_api/file=/[^/]+/[^/]+/[^\s"\'<>,]+',
            r'https://[^/]+\.hf\.space/file=[^\s"\'<>,]+',
            r'/gradio_api/file=/[^\s"\'<>,]+'
        ]
        
        for pattern in gradio_patterns:
            match = re.search(pattern, result_text)
            if match:
                url = match.group(0).rstrip('\'",:;')
                logger.info(f"🎯 Found Gradio file URL: {url}")
                return url
        
        # Check for any HTTP URLs with media extensions
        url_pattern = r'https?://[^\s"\'<>]+\.(?:mp3|wav|ogg|m4a|flac|aac|opus|wma|mp4|webm|avi|mov|mkv|m4v|wmv|png|jpg|jpeg|gif|webp|bmp|svg)'
        match = re.search(url_pattern, result_text, re.IGNORECASE)
        if match:
            url = match.group(0)
            logger.info(f"🎯 Found media URL by extension: {url}")
            return url
        
        # Check for data URLs
        if result_text.startswith('data:'):
            logger.info("🎯 Found data URL")
            return result_text
        
        logger.info("❌ No media URL found in result")
        return None
    
    def _get_native_system_prompt(self) -> str:
        """Get system prompt for HF Inference without MCP servers"""
        model_info = AppConfig.AVAILABLE_MODELS.get(self.mcp_client.current_model, {})
        context_length = model_info.get("context_length", 128000)
        
        return f"""You are an AI assistant powered by {self.mcp_client.current_model} via {self.mcp_client.current_provider}. You have native capabilities for:
- **Text Processing**: You can analyze, summarize, translate, and process text directly
- **General Knowledge**: You can answer questions, explain concepts, and have conversations
- **Code Analysis**: You can read, analyze, and explain code
- **Reasoning**: You can perform step-by-step reasoning and problem-solving
- **Context Window**: You have access to {context_length:,} tokens of context
Current time: {datetime.now().strftime('%Y-%m-%d %H:%M:%S')}
Please provide helpful, accurate, and engaging responses to user queries."""
    
    def _get_mcp_system_prompt(self, uploaded_file_urls: List[str] = None) -> str:
        """Get enhanced system prompt for HF Inference with MCP servers"""
        model_info = AppConfig.AVAILABLE_MODELS.get(self.mcp_client.current_model, {})
        context_length = model_info.get("context_length", 128000)
        
        uploaded_files_context = ""
        if uploaded_file_urls:
            uploaded_files_context = f"\n\nFILES UPLOADED BY USER (Public URLs accessible to MCP servers):\n"
            for i, file_url in enumerate(uploaded_file_urls, 1):
                file_name = file_url.split('/')[-1] if '/' in file_url else file_url
                if AppConfig.is_image_file(file_url):
                    file_type = "Image"
                elif AppConfig.is_audio_file(file_url):
                    file_type = "Audio"
                elif AppConfig.is_video_file(file_url):
                    file_type = "Video"
                else:
                    file_type = "File"
                uploaded_files_context += f"{i}. {file_type}: {file_name}\n   URL: {file_url}\n"
        
        # Get available tools with correct names from enabled servers only
        enabled_servers = self.mcp_client.get_enabled_servers()
        tools_info = []
        for server_name, config in enabled_servers.items():
            tools_info.append(f"- **{server_name}**: {config.description}")
        
        return f"""You are an AI assistant powered by {self.mcp_client.current_model} via {self.mcp_client.current_provider}, with access to various MCP tools.
YOUR NATIVE CAPABILITIES:
- **Text Processing**: You can analyze, summarize, translate, and process text directly
- **General Knowledge**: You can answer questions, explain concepts, and have conversations
- **Code Analysis**: You can read, analyze, and explain code
- **Reasoning**: You can perform step-by-step reasoning and problem-solving
- **Context Window**: You have access to {context_length:,} tokens of context
AVAILABLE MCP TOOLS:
You have access to the following MCP servers:
{chr(10).join(tools_info)}
WHEN TO USE MCP TOOLS:
- **Image Generation**: Creating new images from text prompts
- **Image Editing**: Modifying, enhancing, or transforming existing images  
- **Audio Processing**: Transcribing audio, generating speech, audio enhancement
- **Video Processing**: Creating or editing videos
- **Text to Speech**: Converting text to audio
- **Specialized Analysis**: Tasks requiring specific models or APIs
TOOL USAGE FORMAT:
When you need to use an MCP tool, respond with JSON in this exact format:
{{"use_tool": true, "server": "exact_server_name", "tool": "exact_tool_name", "arguments": {{"param": "value"}}}}
IMPORTANT: Always describe what you're going to do BEFORE the JSON tool call. For example:
"I'll generate speech for your text using the TTS tool."
{{"use_tool": true, "server": "text to speech", "tool": "Kokoro_TTS_mcp_test_generate_first", "arguments": {{"text": "hello"}}}}
IMPORTANT TOOL NAME MAPPING:
- For TTS server: use tool name "Kokoro_TTS_mcp_test_generate_first"
- For image generation: use tool name "dalle_3_xl_lora_v2_generate"  
- For video generation: use tool name "ysharma_ltx_video_distilledtext_to_video"
- For letter counting: use tool name "gradio_app_dummy1_letter_counter"
EXACT SERVER NAMES TO USE:
{', '.join([f'"{name}"' for name in enabled_servers.keys()])}
FILE HANDLING FOR MCP TOOLS:
When using MCP tools with uploaded files, always use the public URLs provided above.
These URLs are accessible to remote MCP servers.
{uploaded_files_context}
MEDIA HANDLING:
When tool results contain media URLs (images, audio, videos), the system will automatically embed them as playable media.
IMPORTANT NOTES:
- Always use the EXACT server names and tool names as specified above
- Use proper JSON format for tool calls
- Include all required parameters in arguments
- For file inputs to MCP tools, use the public URLs provided, not local paths
- ALWAYS provide a descriptive message before the JSON tool call
- After tool execution, you can provide additional context or ask if the user needs anything else
Current time: {datetime.now().strftime('%Y-%m-%d %H:%M:%S')}
Current model: {self.mcp_client.current_model} via {self.mcp_client.current_provider}"""