Upload proxy_handler.py

proxy_handler.py

@@ -6,8 +6,9 @@ import json
 import logging
 import re
 import time
-from typing import AsyncGenerator, Dict, Any
+from typing import AsyncGenerator, Dict, Any, Optional
 import httpx
+from fastapi import HTTPException
 from fastapi.responses import StreamingResponse
 
 from config import settings
@@ -15,6 +16,7 @@ from cookie_manager import cookie_manager
 from models import (
     ChatCompletionRequest,
     ChatCompletionResponse,
+    ChatCompletionStreamResponse,
 )
 
 logger = logging.getLogger(__name__)
@@ -22,25 +24,251 @@ logger = logging.getLogger(__name__)
 
 class ProxyHandler:
     def __init__(self):
-        pass
+        # Configure httpx client for streaming support
+        self.client = httpx.AsyncClient(
+            timeout=httpx.Timeout(60.0, read=300.0),  # Longer read timeout for streaming
+            limits=httpx.Limits(max_connections=100, max_keepalive_connections=20),
+            http2=True  # Enable HTTP/2 for better streaming performance
+        )
 
     async def __aenter__(self):
         return self
 
     async def __aexit__(self, exc_type, exc_val, exc_tb):
-        pass
+        await self.client.aclose()
 
-    async def handle_chat_completion(self, request: ChatCompletionRequest):
-        """Handle chat completion request by streaming or aggregating."""
+    def transform_content(self, content: str) -> str:
+        """Transform content by replacing HTML tags and optionally removing think tags"""
+        if not content:
+            return content
+
+        logger.debug(f"SHOW_THINK_TAGS setting: {settings.SHOW_THINK_TAGS}")
+
+        # Optionally remove thinking content based on configuration
+        if not settings.SHOW_THINK_TAGS:
+            logger.debug("Removing thinking content from response")
+            original_length = len(content)
+
+            # Remove <details> blocks (thinking content) - handle both closed and unclosed tags
+            # First try to remove complete <details>...</details> blocks
+            content = re.sub(
+                r"<details[^>]*>.*?</details>", "", content, flags=re.DOTALL
+            )
+
+            # Then remove any remaining <details> opening tags and everything after them until we hit answer content
+            # Look for pattern: <details...><summary>...</summary>...content... and remove the thinking part
+            content = re.sub(
+                r"<details[^>]*>.*?(?=\s*[A-Z]|\s*\d|\s*$)",
+                "",
+                content,
+                flags=re.DOTALL,
+            )
+
+            content = content.strip()
+
+            logger.debug(
+                f"Content length after removing thinking content: {original_length} -> {len(content)}"
+            )
+        else:
+            logger.debug("Keeping thinking content, converting to <think> tags")
+
+            # Replace <details> with <think>
+            content = re.sub(r"<details[^>]*>", "<think>", content)
+            content = content.replace("</details>", "</think>")
+
+            # Remove <summary> tags and their content
+            content = re.sub(r"<summary>.*?</summary>", "", content, flags=re.DOTALL)
+
+            # If there's no closing </think>, add it at the end of thinking content
+            if "<think>" in content and "</think>" not in content:
+                # Find where thinking ends and answer begins
+                think_start = content.find("<think>")
+                if think_start != -1:
+                    # Look for the start of the actual answer (usually starts with a capital letter or number)
+                    answer_match = re.search(r"\n\s*[A-Z0-9]", content[think_start:])
+                    if answer_match:
+                        insert_pos = think_start + answer_match.start()
+                        content = (
+                            content[:insert_pos] + "</think>\n" + content[insert_pos:]
+                        )
+                    else:
+                        content += "</think>"
+
+        return content.strip()
+
+    async def proxy_request(self, request: ChatCompletionRequest) -> Dict[str, Any]:
+        """Proxy request to Z.AI API"""
+        cookie = await cookie_manager.get_next_cookie()
+        if not cookie:
+            raise HTTPException(status_code=503, detail="No available cookies")
+
+        # Transform model name
+        target_model = (
+            settings.UPSTREAM_MODEL
+            if request.model == settings.MODEL_NAME
+            else request.model
+        )
+
+        # Determine if this should be a streaming response
         is_streaming = (
             request.stream if request.stream is not None else settings.DEFAULT_STREAM
         )
 
-        response_generator = self._stream_proxy_logic(request)
+        # Validate parameter compatibility
+        if is_streaming and not settings.SHOW_THINK_TAGS:
+            logger.warning("SHOW_THINK_TAGS=false is ignored for streaming responses")
+
+        # Prepare request data
+        request_data = request.model_dump(exclude_none=True)
+        request_data["model"] = target_model
+
+        # Build request data based on actual Z.AI format from zai-messages.md
+        import uuid
+
+        request_data = {
+            "stream": True,  # Always request streaming from Z.AI for processing
+            "model": target_model,
+            "messages": request_data["messages"],
+            "background_tasks": {"title_generation": True, "tags_generation": True},
+            "chat_id": str(uuid.uuid4()),
+            "features": {
+                "image_generation": False,
+                "code_interpreter": False,
+                "web_search": False,
+                "auto_web_search": False,
+            },
+            "id": str(uuid.uuid4()),
+            "mcp_servers": ["deep-web-search"],
+            "model_item": {"id": target_model, "name": "GLM-4.5", "owned_by": "openai"},
+            "params": {},
+            "tool_servers": [],
+            "variables": {
+                "{{USER_NAME}}": "User",
+                "{{USER_LOCATION}}": "Unknown",
+                "{{CURRENT_DATETIME}}": "2025-08-04 16:46:56",
+            },
+        }
+
+        logger.debug(f"Sending request data: {request_data}")
+
+        headers = {
+            "Content-Type": "application/json",
+            "Authorization": f"Bearer {cookie}",
+            "User-Agent": "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/138.0.0.0 Safari/537.36",
+            "Accept": "application/json, text/event-stream",
+            "Accept-Language": "zh-CN",
+            "sec-ch-ua": '"Not)A;Brand";v="8", "Chromium";v="138", "Google Chrome";v="138"',
+            "sec-ch-ua-mobile": "?0",
+            "sec-ch-ua-platform": '"macOS"',
+            "x-fe-version": "prod-fe-1.0.53",
+            "Origin": "https://chat.z.ai",
+            "Referer": "https://chat.z.ai/c/069723d5-060b-404f-992c-4705f1554c4c",
+        }
+
+        try:
+            # Use client.stream() for TRUE streaming response - this is the key fix!
+            async with self.client.stream(
+                "POST",
+                settings.UPSTREAM_URL,
+                json=request_data,
+                headers=headers,
+                timeout=httpx.Timeout(60.0, read=300.0)
+            ) as response:
+
+                if response.status_code == 401:
+                    await cookie_manager.mark_cookie_failed(cookie)
+                    raise HTTPException(status_code=401, detail="Invalid authentication")
+
+                if response.status_code != 200:
+                    # For streaming, we need to read the error response properly
+                    try:
+                        error_text = await response.aread()
+                        error_detail = error_text.decode('utf-8')
+                    except:
+                        error_detail = f"HTTP {response.status_code}"
+
+                    raise HTTPException(
+                        status_code=response.status_code,
+                        detail=f"Upstream error: {error_detail}",
+                    )
+
+                await cookie_manager.mark_cookie_success(cookie)
+                return {"response": response, "cookie": cookie}
+
+        except httpx.RequestError as e:
+            logger.error(f"Request error: {e}")
+            logger.error(f"Request error type: {type(e).__name__}")
+            logger.error(f"Request URL: {settings.UPSTREAM_URL}")
+            logger.error(f"Request timeout: {self.client.timeout}")
+            await cookie_manager.mark_cookie_failed(cookie)
+            raise HTTPException(
+                status_code=503, detail=f"Upstream service unavailable: {str(e)}"
+            )
+
+    async def process_streaming_response(
+        self, response: httpx.Response
+    ) -> AsyncGenerator[Dict[str, Any], None]:
+        """Process streaming response from Z.AI - TRUE real-time processing"""
+        buffer = ""
+
+        # Use aiter_text with small chunk size for real-time processing
+        async for chunk in response.aiter_text(chunk_size=1024):  # Small chunks for responsiveness
+            if not chunk:
+                continue
+
+            buffer += chunk
+
+            # Process complete lines immediately
+            while "\n" in buffer:
+                line, buffer = buffer.split("\n", 1)
+                line = line.strip()
+
+                if not line.startswith("data: "):
+                    continue
+
+                payload = line[6:].strip()
+                if payload == "[DONE]":
+                    return
+
+                try:
+                    parsed = json.loads(payload)
+
+                    # Check for errors first
+                    if parsed.get("error") or (parsed.get("data", {}).get("error")):
+                        error_detail = (
+                            parsed.get("error", {}).get("detail")
+                            or parsed.get("data", {}).get("error", {}).get("detail")
+                            or "Unknown error from upstream"
+                        )
+                        logger.error(f"Upstream error: {error_detail}")
+                        raise HTTPException(
+                            status_code=400, detail=f"Upstream error: {error_detail}"
+                        )
+
+                    # Clean up response data
+                    if parsed.get("data"):
+                        # Remove unwanted fields for cleaner processing
+                        parsed["data"].pop("edit_index", None)
+                        parsed["data"].pop("edit_content", None)
+
+                    # Yield immediately for real-time streaming
+                    yield parsed
+
+                except json.JSONDecodeError as e:
+                    logger.debug(f"JSON decode error (skipping): {e}")
+                    continue  # Skip non-JSON lines
+
+    async def handle_chat_completion(self, request: ChatCompletionRequest):
+        """Handle chat completion request"""
+        # Determine final streaming mode
+        is_streaming = (
+            request.stream if request.stream is not None else settings.DEFAULT_STREAM
+        )
 
         if is_streaming:
+            # For streaming responses, use direct streaming proxy
             return StreamingResponse(
-                response_generator,
+                self.stream_proxy_response(request),
                 media_type="text/event-stream",
                 headers={
                     "Cache-Control": "no-cache",
@@ -48,159 +276,355 @@ class ProxyHandler:
                 },
             )
         else:
-            full_content = ""
-            final_id = f"chatcmpl-{time.time()}"
-            async for chunk_data in response_generator:
+            # For non-streaming responses, collect all streaming data first
+            chunks = []
+            async for chunk_data in self.stream_proxy_response(request):
                 if chunk_data.startswith("data: ") and not chunk_data.startswith("data: [DONE]"):
                     try:
-                        chunk_json_str = chunk_data[6:].strip()
-                        if not chunk_json_str:
-                            continue
-                        chunk_json = json.loads(chunk_json_str)
-                        final_id = chunk_json.get("id", final_id)
-                        delta = chunk_json.get("choices", [{}])[0].get("delta", {})
-                        if delta and delta.get("content"):
-                            full_content += delta["content"]
-                    except json.JSONDecodeError:
-                        logger.warning(f"Could not decode JSON from stream chunk: {chunk_data}")
+                        chunk_json = json.loads(chunk_data[6:])
+                        if chunk_json.get("choices", [{}])[0].get("delta", {}).get("content"):
+                            chunks.append(chunk_json["choices"][0]["delta"]["content"])
+                    except:
                         continue
-            
+
+            # Combine all content
+            full_content = "".join(chunks)
+
+            # Apply final transformation for non-streaming response
+            transformed_content = self.transform_content(full_content)
+
+            # Return as non-streaming response
+            import time
+            import uuid
             return ChatCompletionResponse(
-                id=final_id,
+                id=f"chatcmpl-{uuid.uuid4().hex[:29]}",
                 created=int(time.time()),
                 model=request.model,
                 choices=[
                     {
                         "index": 0,
-                        "message": {"role": "assistant", "content": full_content.strip()},
+                        "message": {"role": "assistant", "content": transformed_content},
                         "finish_reason": "stop",
                     }
                 ],
             )
 
-    async def _stream_proxy_logic(self, request: ChatCompletionRequest) -> AsyncGenerator[str, None]:
-        """
-        Core logic for streaming from Z.AI and transforming to OpenAI format.
-        """
-        cookie = await cookie_manager.get_next_cookie()
-        if not cookie:
-            error_chunk = {"error": {"message": "No available cookies", "type": "server_error", "code": 503}}
+    async def stream_response(self, response: httpx.Response, model: str) -> AsyncGenerator[str, None]:
+        """Generate TRUE streaming response in OpenAI format - real-time processing"""
+        import uuid
+        import time
+
+        # Generate a unique completion ID
+        completion_id = f"chatcmpl-{uuid.uuid4().hex[:29]}"
+        current_phase = None
+
+        try:
+            # Real-time streaming: process each chunk immediately as it arrives
+            async for parsed in self.process_streaming_response(response):
+                try:
+                    data = parsed.get("data", {})
+                    delta_content = data.get("delta_content", "")
+                    phase = data.get("phase", "")
+
+                    # Track phase changes
+                    if phase != current_phase:
+                        current_phase = phase
+                        logger.debug(f"Phase changed to: {phase}")
+
+                    # Apply filtering based on SHOW_THINK_TAGS and phase
+                    should_send_content = True
+
+                    if not settings.SHOW_THINK_TAGS and phase == "thinking":
+                        # Skip thinking content when SHOW_THINK_TAGS=false
+                        should_send_content = False
+                        logger.debug(f"Skipping thinking content (SHOW_THINK_TAGS=false)")
+
+                    # Process and send content immediately if we should
+                    if delta_content and should_send_content:
+                        # Minimal transformation for real-time streaming
+                        transformed_delta = delta_content
+
+                        if settings.SHOW_THINK_TAGS:
+                            # Simple tag replacement for streaming
+                            transformed_delta = re.sub(r'<details[^>]*>', '<think>', transformed_delta)
+                            transformed_delta = transformed_delta.replace('</details>', '</think>')
+                            # Note: Skip complex regex for streaming performance
+
+                        # Create and send OpenAI-compatible chunk immediately
+                        openai_chunk = {
+                            "id": completion_id,
+                            "object": "chat.completion.chunk",
+                            "created": int(time.time()),
+                            "model": model,
+                            "choices": [{
+                                "index": 0,
+                                "delta": {
+                                    "content": transformed_delta
+                                },
+                                "finish_reason": None
+                            }]
+                        }
+
+                        # Yield immediately for real-time streaming
+                        yield f"data: {json.dumps(openai_chunk)}\n\n"
+
+                except Exception as e:
+                    logger.error(f"Error processing streaming chunk: {e}")
+                    continue
+
+            # Send final completion chunk
+            final_chunk = {
+                "id": completion_id,
+                "object": "chat.completion.chunk",
+                "created": int(time.time()),
+                "model": model,
+                "choices": [{
+                    "index": 0,
+                    "delta": {},
+                    "finish_reason": "stop"
+                }]
+            }
+
+            yield f"data: {json.dumps(final_chunk)}\n\n"
+            yield "data: [DONE]\n\n"
+
+        except Exception as e:
+            logger.error(f"Streaming error: {e}")
+            # Send error in OpenAI format
+            error_chunk = {
+                "error": {
+                    "message": str(e),
+                    "type": "server_error"
+                }
+            }
             yield f"data: {json.dumps(error_chunk)}\n\n"
-            return
-        
-        target_model = settings.UPSTREAM_MODEL
+
+    async def non_stream_response(
+        self, response: httpx.Response, model: str
+    ) -> ChatCompletionResponse:
+        """Generate non-streaming response"""
+        chunks = []
+        async for parsed in self.process_streaming_response(response):
+            chunks.append(parsed)
+            logger.debug(f"Received chunk: {parsed}")  # Debug log
+
+        if not chunks:
+            raise HTTPException(status_code=500, detail="No response from upstream")
+
+        logger.info(f"Total chunks received: {len(chunks)}")
+        logger.debug(f"First chunk structure: {chunks[0] if chunks else 'None'}")
+
+        # Aggregate content based on SHOW_THINK_TAGS setting
+        if settings.SHOW_THINK_TAGS:
+            # Include all content
+            full_content = "".join(
+                chunk.get("data", {}).get("delta_content", "") for chunk in chunks
+            )
+        else:
+            # Only include answer phase content
+            full_content = "".join(
+                chunk.get("data", {}).get("delta_content", "")
+                for chunk in chunks
+                if chunk.get("data", {}).get("phase") == "answer"
+            )
+
+        logger.info(f"Aggregated content length: {len(full_content)}")
+        logger.debug(
+            f"Full aggregated content: {full_content}"
+        )  # Show full content for debugging
+
+        # Apply content transformation (including think tag filtering)
+        transformed_content = self.transform_content(full_content)
+
+        logger.info(f"Transformed content length: {len(transformed_content)}")
+        logger.debug(f"Transformed content: {transformed_content[:200]}...")
+
+        # Create OpenAI-compatible response
+        return ChatCompletionResponse(
+            id=chunks[0].get("data", {}).get("id", "chatcmpl-unknown"),
+            created=int(time.time()),
+            model=model,
+            choices=[
+                {
+                    "index": 0,
+                    "message": {"role": "assistant", "content": transformed_content},
+                    "finish_reason": "stop",
+                }
+            ],
+        )
+
+    async def stream_proxy_response(self, request: ChatCompletionRequest) -> AsyncGenerator[str, None]:
+        """TRUE streaming proxy - direct pass-through with minimal processing"""
         import uuid
-        
-        # KEY FIX: Add "enable_thinking": True to the features object to force the thinking phase.
+        import time
+
+        # Get cookie
+        cookie = await cookie_manager.get_next_cookie()
+        if not cookie:
+            raise HTTPException(status_code=503, detail="No valid authentication available")
+
+        # Prepare request data
+        request_data = request.model_dump(exclude_none=True)
+        target_model = "0727-360B-API"  # Map GLM-4.5 to Z.AI model
+
+        # Build Z.AI request format
         request_data = {
-            "stream": True,
+            "stream": True,  # Always request streaming from Z.AI
             "model": target_model,
-            "messages": request.model_dump(exclude_none=True)["messages"],
+            "messages": request_data["messages"],
+            "background_tasks": {
+                "title_generation": True,
+                "tags_generation": True
+            },
+            "chat_id": str(uuid.uuid4()),
             "features": {
                 "image_generation": False,
                 "code_interpreter": False,
                 "web_search": False,
                 "auto_web_search": False,
-                "enable_thinking": True,  # This is the critical flag.
+                "enable_thinking": True,
             },
             "id": str(uuid.uuid4()),
-            "model_item": {"id": target_model, "name": "GLM-4.5", "owned_by": "openai"},
+            "mcp_servers": ["deep-web-search"],
+            "model_item": {
+                "id": target_model,
+                "name": "GLM-4.5",
+                "owned_by": "openai"
+            },
+            "params": {},
+            "tool_servers": [],
             "variables": {
                 "{{USER_NAME}}": "User",
                 "{{USER_LOCATION}}": "Unknown",
-                "{{CURRENT_DATETIME}}": time.strftime("%Y-%m-%d %H:%M:%S"),
-            },
+                "{{CURRENT_DATETIME}}": "2025-08-04 16:46:56"
+            }
         }
 
         headers = {
-            "Content-Type": "application/json", "Authorization": f"Bearer {cookie}",
+            "Content-Type": "application/json",
+            "Authorization": f"Bearer {cookie}",
             "User-Agent": "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/138.0.0.0 Safari/537.36",
-            "Accept": "application/json, text/event-stream", "Accept-Language": "zh-CN",
-            "sec-ch-ua": '"Not)A;Brand";v="8", "Chromium";v="138", "Google Chrome";v="138"', "sec-ch-ua-mobile": "?0",
-            "sec-ch-ua-platform": '"macOS"', "x-fe-version": "prod-fe-1.0.53", "Origin": "https://chat.z.ai",
-            "Referer": "https://chat.z.ai/",
+            "Accept": "application/json, text/event-stream",
+            "Accept-Language": "zh-CN",
+            "sec-ch-ua": '"Not)A;Brand";v="8", "Chromium";v="138", "Google Chrome";v="138"',
+            "sec-ch-ua-mobile": "?0",
+            "sec-ch-ua-platform": '"macOS"',
+            "x-fe-version": "prod-fe-1.0.53",
+            "Origin": "https://chat.z.ai",
+            "Referer": "https://chat.z.ai/c/069723d5-060b-404f-992c-4705f1554c4c",
         }
 
         completion_id = f"chatcmpl-{uuid.uuid4().hex[:29]}"
-        buffer = ""
+        current_phase = None
+
+        try:
+            # Create a new client for this streaming request to avoid conflicts
+            async with httpx.AsyncClient(
+                timeout=httpx.Timeout(60.0, read=300.0),
+                limits=httpx.Limits(max_connections=100, max_keepalive_connections=20),
+                http2=True
+            ) as stream_client:
+                async with stream_client.stream(
+                    "POST",
+                    settings.UPSTREAM_URL,
+                    json=request_data,
+                    headers=headers
+                ) as response:
+
+                    if response.status_code == 401:
+                        await cookie_manager.mark_cookie_failed(cookie)
+                        raise HTTPException(status_code=401, detail="Invalid authentication")
 
-        async with httpx.AsyncClient(
-            timeout=httpx.Timeout(60.0, read=300.0), limits=httpx.Limits(max_connections=100, max_keepalive_connections=20), http2=True
-        ) as client:
-            try:
-                async with client.stream("POST", settings.UPSTREAM_URL, json=request_data, headers=headers) as response:
                     if response.status_code != 200:
                         await cookie_manager.mark_cookie_failed(cookie)
-                        error_text = await response.aread()
-                        error_detail = error_text.decode('utf-8', errors='ignore')
-                        error_code = response.status_code
-                        error_type = "auth_error" if error_code == 401 else "server_error"
-                        error_chunk = {"error": {"message": f"Upstream error ({error_code}): {error_detail}", "type": error_type, "code": error_code}}
-                        yield f"data: {json.dumps(error_chunk)}\n\n"
-                        return
+                        raise HTTPException(status_code=response.status_code, detail="Upstream error")
 
                     await cookie_manager.mark_cookie_success(cookie)
 
-                    async for chunk in response.aiter_text():
+                    # Process streaming response in real-time
+                    buffer = ""
+                    async for chunk in response.aiter_text(chunk_size=1024):
+                        if not chunk:
+                            continue
+
                         buffer += chunk
+
+                        # Process complete lines immediately
                         while "\n" in buffer:
                             line, buffer = buffer.split("\n", 1)
                             line = line.strip()
 
-                            if not line.startswith("data:"):
-                                continue
-                            
-                            payload_str = line[6:].strip()
-                            if payload_str == "[DONE]":
+                            if not line.startswith("data: "):
                                 continue
 
+                            payload = line[6:].strip()
+                            if payload == "[DONE]":
+                                # Send final chunk and done
+                                final_chunk = {
+                                    "id": completion_id,
+                                    "object": "chat.completion.chunk",
+                                    "created": int(time.time()),
+                                    "model": request.model,
+                                    "choices": [{
+                                        "index": 0,
+                                        "delta": {},
+                                        "finish_reason": "stop"
+                                    }]
+                                }
+                                yield f"data: {json.dumps(final_chunk)}\n\n"
+                                yield "data: [DONE]\n\n"
+                                return
+
                             try:
-                                parsed = json.loads(payload_str)
+                                parsed = json.loads(payload)
                                 data = parsed.get("data", {})
                                 delta_content = data.get("delta_content", "")
-                                phase = data.get("phase", "").strip()
-
-                                if not delta_content:
-                                    continue
-
-                                output_content = ""
-                                if settings.SHOW_THINK_TAGS:
-                                    temp_content = re.sub(r'<details[^>]*>\s*<summary>.*?</summary>', '<think>', delta_content, flags=re.DOTALL)
-                                    output_content = temp_content.replace('</details>', '</think>')
-                                else:
-                                    if phase == "thinking":
-                                        output_content = ""
-                                    elif phase == "answer":
-                                        if '</details>' in delta_content:
-                                            output_content = delta_content.split('</details>', 1)[1]
-                                        else:
-                                            output_content = delta_content
-                                
-                                if output_content:
-                                    yield f"data: {json.dumps(self._create_openai_chunk(completion_id, request.model, output_content))}\n\n"
+                                phase = data.get("phase", "")
+
+                                # Track phase changes
+                                if phase != current_phase:
+                                    current_phase = phase
+                                    logger.debug(f"Phase changed to: {phase}")
+
+                                # Apply filtering based on SHOW_THINK_TAGS and phase
+                                should_send_content = True
+
+                                if not settings.SHOW_THINK_TAGS and phase == "thinking":
+                                    should_send_content = False
+
+                                # Process and send content immediately if we should
+                                if delta_content and should_send_content:
+                                    # Minimal transformation for real-time streaming
+                                    transformed_delta = delta_content
+
+                                    if settings.SHOW_THINK_TAGS:
+                                        # Simple tag replacement for streaming
+                                        transformed_delta = re.sub(r'<details[^>]*>', '<think>', transformed_delta)
+                                        transformed_delta = transformed_delta.replace('</details>', '</think>')
+                                        transformed_delta = re.sub(r"<summary>.*?</summary>", "", transformed_delta, flags=re.DOTALL)
+
+                                    # Create and send OpenAI-compatible chunk immediately
+                                    openai_chunk = {
+                                        "id": completion_id,
+                                        "object": "chat.completion.chunk",
+                                        "created": int(time.time()),
+                                        "model": request.model,
+                                        "choices": [{
+                                            "index": 0,
+                                            "delta": {
+                                                "content": transformed_delta
+                                            },
+                                            "finish_reason": None
+                                        }]
+                                    }
+
+                                    # Yield immediately for real-time streaming
+                                    yield f"data: {json.dumps(openai_chunk)}\n\n"
 
                             except json.JSONDecodeError:
-                                logger.debug(f"Skipping non-JSON data line: {line}")
-                                continue
+                                continue  # Skip non-JSON lines
 
-                final_chunk = self._create_openai_chunk(completion_id, request.model, "", is_final=True)
-                yield f"data: {json.dumps(final_chunk)}\n\n"
-                yield "data: [DONE]\n\n"
-
-            except httpx.RequestError as e:
-                logger.error(f"Streaming request error: {e}")
-                await cookie_manager.mark_cookie_failed(cookie)
-                error_chunk = {"error": {"message": f"Upstream service unavailable: {str(e)}", "type": "server_error"}}
-                yield f"data: {json.dumps(error_chunk)}\n\n"
-
-    def _create_openai_chunk(self, completion_id: str, model: str, content: str, is_final: bool = False) -> Dict[str, Any]:
-        """A helper to create a standard OpenAI stream chunk."""
-        chunk = {
-            "id": completion_id, "object": "chat.completion.chunk", "created": int(time.time()),
-            "model": model, "choices": [{"index": 0, "delta": {}, "finish_reason": None}]
-        }
-        if is_final:
-            chunk["choices"][0]["finish_reason"] = "stop"
-        else:
-            chunk["choices"][0]["delta"] = {"content": content}
-        return chunk
+        except httpx.RequestError as e:
+            logger.error(f"Streaming request error: {e}")
+            await cookie_manager.mark_cookie_failed(cookie)
+            raise HTTPException(status_code=503, detail=f"Upstream service unavailable: {str(e)}")