Update proxy_handler.py

proxy_handler.py

@@ -1,10 +1,8 @@
 """
 Proxy handler for Z.AI API requests
 """
-import json, logging, re, time, uuid, base64, hashlib, hmac, urllib.parse
+import json, logging, re, time, uuid, base64, datetime, hashlib, hmac, urllib.parse
 from typing import AsyncGenerator, Dict, Any, Tuple, List
-from datetime import datetime, timezone, timedelta
-
 import httpx
 from fastapi import HTTPException
 from fastapi.responses import StreamingResponse
@@ -23,67 +21,131 @@ class ProxyHandler:
             limits=httpx.Limits(max_connections=100, max_keepalive_connections=20),
             http2=True,
         )
-        # 从JavaScript代码中获取的固定密钥
-        self.primary_secret = "junjie".encode('utf-8')
 
     async def aclose(self):
         if not self.client.is_closed:
             await self.client.aclose()
-            
-    def _get_timestamp_millis(self) -> int:
-        return int(time.time() * 1000)
-
+    
     def _parse_jwt_token(self, token: str) -> Dict[str, str]:
-        """A simple JWT payload decoder to get user ID."""
+        """A simple, dependency-free JWT parser to get the user_id."""
         try:
             parts = token.split('.')
             if len(parts) != 3:
-                return {"userId": ""}
-            # Add padding if necessary
+                return {"user_id": ""}
             payload_b64 = parts[1]
+            # Add padding if necessary for base64 decoding
             payload_b64 += '=' * (-len(payload_b64) % 4)
-            payload_json = base64.urlsafe_b64decode(payload_b64).decode('utf-8')
-            payload = json.loads(payload_json)
-            # The JS code checks multiple keys, 'sub' is the most standard one.
-            return {"userId": payload.get("sub", "")}
-        except Exception as e:
-            logger.warning(f"Failed to parse JWT token to get user ID: {e}")
-            return {"userId": ""}
-
-    def _generate_signature(self, token: str, payload_str: str, mt: str) -> Tuple[str, int]:
-        timestamp_ms = self._get_timestamp_millis()
-        minute_bucket = str(timestamp_ms // 60000)
-
-        level1_data = f"{token}|{minute_bucket}".encode('utf-8')
-        mac1 = hmac.new(self.primary_secret, level1_data, hashlib.sha256)
+            payload_bytes = base64.urlsafe_b64decode(payload_b64)
+            payload = json.loads(payload_bytes)
+            user_id = payload.get("sub", "")
+            return {"user_id": user_id}
+        except Exception:
+            logger.warning("Failed to parse JWT token, continuing without user_id.", exc_info=False)
+            return {"user_id": ""}
+
+    def _construct_payload(self, token: str, user_id: str, chat_id: str, request_id: str) -> Tuple[str, str]:
+        """Constructs the sorted payload string (vl) and URL parameters for signature."""
+        timestamp_ms = str(int(time.time() * 1000))
+        # Hardcoding is fine for these fingerprinting values, mimicking the JS logic
+        data = {
+            'timestamp': timestamp_ms,
+            'requestId': request_id,
+            'user_id': user_id,
+            'token': token,
+            'user_agent': 'Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/141.0.0.0 Safari/537.36',
+            'current_url': f'https://chat.z.ai/c/{chat_id}',
+            'pathname': f'/c/{chat_id}',
+            'timezone': 'Asia/Shanghai',  # Hardcoded for simplicity
+            'timezone_offset': '-480',  # Hardcoded for simplicity (UTC+8)
+            'local_time': datetime.datetime.now().isoformat(),
+            'utc_time': datetime.datetime.utcnow().strftime('%a, %d %b %Y %H:%M:%S GMT'),
+            'version': '0.0.1', 'platform': 'web', 'language': 'zh-CN', 'languages': 'zh-CN,en',
+            'cookie_enabled': 'true', 'screen_width': '2560', 'screen_height': '1440',
+            'screen_resolution': '2560x1440', 'viewport_height': '1328', 'viewport_width': '1342',
+            'viewport_size': '1342x1328', 'color_depth': '24', 'pixel_ratio': '2',
+            'search': '', 'hash': '', 'host': 'chat.z.ai', 'hostname': 'chat.z.ai',
+            'protocol': 'https:', 'referrer': '', 'title': 'Chat with Z.ai - Free AI Chatbot powered by GLM-4.5',
+            'is_mobile': 'false', 'is_touch': 'false', 'max_touch_points': '0',
+            'browser_name': 'Chrome', 'os_name': 'Mac OS'
+        }
+        
+        # Sort keys and create the required string formats
+        sorted_items = sorted(data.items())
+        sorted_payload_str = ','.join([f"{k},{v}" for k, v in sorted_items])
+        url_params_str = urllib.parse.urlencode(dict(sorted_items))
+        
+        return sorted_payload_str, url_params_str
+
+    def _generate_signature(self, vl: str, mt: str, token: str) -> Dict[str, Any]:
+        """Generates the signature based on the provided JS logic."""
+        primary_secret = "junjie"
+        timestamp_ms = int(time.time() * 1000)
+
+        # Use 1-minute buckets as in the JS code (60 seconds * 1000 ms)
+        minute_bucket = timestamp_ms // 60000
+        
+        # Level 1 HMAC to derive key
+        level1_data = f"{token}|{minute_bucket}"
+        mac1 = hmac.new(primary_secret.encode('utf-8'), level1_data.encode('utf-8'), hashlib.sha256)
         derived_key_hex = mac1.hexdigest()
 
-        level2_data = f"{payload_str}|{mt}|{timestamp_ms}".encode('utf-8')
-        mac2 = hmac.new(derived_key_hex.encode('utf-8'), level2_data, hashlib.sha256)
+        # Level 2 HMAC for the final signature
+        level2_data = f"{vl}|{mt}|{timestamp_ms}"
+        mac2 = hmac.new(derived_key_hex.encode('utf-8'), level2_data.encode('utf-8'), hashlib.sha256)
         signature = mac2.hexdigest()
-        
-        return signature, timestamp_ms
 
+        return {"signature": signature, "timestamp": timestamp_ms}
+        
     def _clean_thinking_content(self, text: str) -> str:
-        if not text: return ""
+        """
+        Aggressively cleans raw thinking content strings based on observed patterns
+        from the Z.AI API.
+        """
+        if not text:
+            return ""
+
         cleaned_text = text
+
+        # 1. Remove specific unwanted blocks like tool calls and summaries.
         cleaned_text = re.sub(r'<summary>.*?</summary>', '', cleaned_text, flags=re.DOTALL)
         cleaned_text = re.sub(r'<glm_block.*?</glm_block>', '', cleaned_text, flags=re.DOTALL)
+        
+        # 2. **FIX**: Remove tag-like metadata containing `duration` attribute.
+        # This handles the reported issue: `true" duration="0" ... >`
         cleaned_text = re.sub(r'<[^>]*duration="[^"]*"[^>]*>', '', cleaned_text)
-        cleaned_text = cleaned_text.replace("</thinking>", "").replace("<Full>", "").replace("</Full>", "")
+
+        # 3. Remove specific structural tags, but keep the content between them.
+        cleaned_text = cleaned_text.replace("</thinking>", "")
+        cleaned_text = cleaned_text.replace("<Full>", "")
+        cleaned_text = cleaned_text.replace("</Full>", "")
+        # This regex handles <details>, <details open>, and </details>
         cleaned_text = re.sub(r'</?details[^>]*>', '', cleaned_text)
+
+        # 4. Handle markdown blockquotes, preserving multi-level ones.
         cleaned_text = re.sub(r'^\s*>\s*(?!>)', '', cleaned_text, flags=re.MULTILINE)
+
+        # 5. Remove other known text artifacts.
         cleaned_text = cleaned_text.replace("Thinking…", "")
+        
+        # 6. Final strip to clean up residual whitespace.
         return cleaned_text.strip()
 
     def _clean_answer_content(self, text: str) -> str:
-        if not text: return ""
+        """
+        Cleans unwanted tags from answer content.
+        Does NOT strip whitespace to preserve markdown in streams.
+        """
+        if not text:
+            return ""
+        # Remove tool call blocks
         cleaned_text = re.sub(r'<glm_block.*?</glm_block>', '', text, flags=re.DOTALL)
+        # Remove any residual details/summary blocks that might leak into the answer
         cleaned_text = re.sub(r'<details[^>]*>.*?</details>', '', cleaned_text, flags=re.DOTALL)
         cleaned_text = re.sub(r'<summary>.*?</summary>', '', cleaned_text, flags=re.DOTALL)
         return cleaned_text
 
     def _serialize_msgs(self, msgs) -> list:
+        """Converts message objects to a list of dictionaries."""
         out = []
         for m in msgs:
             if hasattr(m, "dict"): out.append(m.dict())
@@ -93,91 +155,40 @@ class ProxyHandler:
         return out
 
     async def _prep_upstream(self, req: ChatCompletionRequest) -> Tuple[Dict[str, Any], Dict[str, str], str, str]:
-        """Prepares the request body, headers, cookie, and URL for the upstream API."""
+        """Prepares the request body, headers, URL, and cookie for the upstream API."""
         ck = await cookie_manager.get_next_cookie()
         if not ck: raise HTTPException(503, "No available cookies")
 
-        model = settings.UPSTREAM_MODEL if req.model == settings.MODEL_NAME else req.model
+        # 1. Extract necessary info for signature
         chat_id = str(uuid.uuid4())
         request_id = str(uuid.uuid4())
-        user_info = self._parse_jwt_token(ck)
-        user_id = user_info.get("userId", "")
-        
-        body = { "stream": True, "model": model, "messages": self._serialize_msgs(req.messages), "background_tasks": {"title_generation": True, "tags_generation": True}, "chat_id": chat_id, "features": {"image_generation": False, "code_interpreter": False, "web_search": False, "auto_web_search": False, "enable_thinking": True,}, "id": request_id, "mcp_servers": ["deep-web-search"], "model_item": {"id": model, "name": "GLM-4.6", "owned_by": "openai"}, "params": {}, "tool_servers": [], "variables": {"{{USER_NAME}}": "User", "{{USER_LOCATION}}": "Unknown", "{{CURRENT_DATETIME}}": time.strftime("%Y-%m-%d %H:%M:%S"),},}
+        user_id = self._parse_jwt_token(ck).get("user_id", "")
         
-        # --- FIX: Timezone-aware payload generation ---
-        # The JS code uses the local timezone. We'll simulate a common one (Asia/Shanghai, UTC+8).
-        # This is the most likely cause of the signature mismatch.
-        target_tz_offset_hours = 8
-        target_tz = timezone(timedelta(hours=target_tz_offset_hours))
-        now_local = datetime.now(target_tz)
-        now_utc = datetime.now(timezone.utc)
-        
-        # JS getTimezoneOffset returns the difference in minutes from UTC, and the sign is inverted.
-        # e.g., for UTC+8, it returns -480.
-        timezone_offset_minutes = -target_tz_offset_hours * 60
+        last_message = req.messages[-1] if req.messages else None
+        # 'mt' is the content of the last message
+        mt = last_message.content if last_message and isinstance(last_message.content, str) else ""
 
-        payload_data = {
-            'timestamp': str(self._get_timestamp_millis()),
-            'requestId': request_id,
-            'user_id': user_id,
-            'token': ck,
-            'user_agent': "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/141.0.0.0 Safari/537.36",
-            'current_url': f"https://chat.z.ai/c/{chat_id}",
-            'pathname': f"/c/{chat_id}",
-            'timezone': 'Asia/Shanghai',  # Use a specific timezone name
-            'timezone_offset': str(timezone_offset_minutes),
-            'local_time': now_local.isoformat(timespec='milliseconds'),
-            'utc_time': now_utc.strftime('%a, %d %b %Y %H:%M:%S GMT'),
-            'version': '0.0.1',
-            'platform': 'web',
-            'language': 'zh-CN',
-            'languages': 'zh-CN,en',
-            'cookie_enabled': 'true',
-            'screen_width': '2560',
-            'screen_height': '1440',
-            'screen_resolution': '2560x1440',
-            'viewport_height': '1328',
-            'viewport_width': '1342',
-            'viewport_size': '1342x1328',
-            'color_depth': '24',
-            'pixel_ratio': '2',
-            'search': '',
-            'hash': '',
-            'host': 'chat.z.ai',
-            'hostname': 'chat.z.ai',
-            'protocol': 'https:',
-            'referrer': '',
-            'title': 'Chat with Z.ai - Free AI Chatbot powered by GLM-4.5',
-            'is_mobile': 'false',
-            'is_touch': 'false',
-            'max_touch_points': '0',
-            'browser_name': 'Chrome',
-            'os_name': 'Mac OS'
-        }
-        
-        keys = sorted(payload_data.keys())
-        sorted_payload = ",".join([f"{k},{payload_data[k]}" for k in keys])
-        url_params = urllib.parse.urlencode(payload_data)
-
-        last_message_content = ""
-        if req.messages:
-            last_message = req.messages[-1]
-            if isinstance(last_message.content, str):
-                last_message_content = last_message.content
+        # 2. Generate signature components
+        # 'vl' is the sorted payload string
+        vl, url_params = self._construct_payload(ck, user_id, chat_id, request_id)
+        sig_data = self._generate_signature(vl, mt, ck)
+        signature = sig_data["signature"]
+        timestamp = sig_data["timestamp"]
         
-        signature, sig_timestamp = self._generate_signature(ck, sorted_payload, last_message_content)
-        
-        final_url = f"{settings.UPSTREAM_URL}?{url_params}&signature_timestamp={sig_timestamp}"
+        # 3. Construct the final dynamic URL
+        final_url = f"{settings.UPSTREAM_URL}?{url_params}&signature_timestamp={timestamp}"
 
-        headers = { "Content-Type": "application/json", "Authorization": f"Bearer {ck}", "User-Agent": ("Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/141.0.0.0 Safari/537.36"), "Accept": "application/json, text/event-stream", "Accept-Language": "zh-CN", "sec-ch-ua": '"Not)A;Brand";v="8", "Chromium";v="141", "Google Chrome";v="141"', "sec-ch-ua-mobile": "?0", "sec-ch-ua-platform": '"macOS"', "x-fe-version": "prod-fe-1.0.79", "X-Signature": 'ZTpUeXBlRXJyb3I6IENhbm5vdCByZWFkIHByb3BlcnRpZXMgb2YgdW5kZWZpbmVkIChyZWFkaW5nICdjaGlsZE5vZGVzJyk=', "Origin": "https://chat.z.ai", "Referer": "https://chat.z.ai/",}
+        # 4. Prepare body and headers
+        model = settings.UPSTREAM_MODEL if req.model == settings.MODEL_NAME else req.model
+        body = { "stream": True, "model": model, "messages": self._serialize_msgs(req.messages), "background_tasks": {"title_generation": True, "tags_generation": True}, "chat_id": chat_id, "features": {"image_generation": False, "code_interpreter": False, "web_search": False, "auto_web_search": False, "enable_thinking": True,}, "id": request_id, "mcp_servers": ["deep-web-search"], "model_item": {"id": model, "name": "GLM-4.6", "owned_by": "openai"}, "params": {}, "tool_servers": [], "variables": {"{{USER_NAME}}": "User", "{{USER_LOCATION}}": "Unknown", "{{CURRENT_DATETIME}}": time.strftime("%Y-%m-%d %H:%M:%S"),},}
+        headers = { "Content-Type": "application/json", "Authorization": f"Bearer {ck}", "User-Agent": ("Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/141.0.0.0 Safari/537.36"), "Accept": "application/json, text/event-stream", "Accept-Language": "zh-CN", "sec-ch-ua": '"Not)A;Brand";v="8", "Chromium";v="141", "Google Chrome";v="141"', "sec-ch-ua-mobile": "?0", "sec-ch-ua-platform": '"macOS"', "x-fe-version": "prod-fe-1.0.79", "X-Signature": signature, "Origin": "https://chat.z.ai", "Referer": "https://chat.z.ai/",}
         
-        return body, headers, ck, final_url
+        return body, headers, final_url, ck
         
     async def stream_proxy_response(self, req: ChatCompletionRequest) -> AsyncGenerator[str, None]:
         ck = None
         try:
-            body, headers, ck, url = await self._prep_upstream(req)
+            body, headers, url, ck = await self._prep_upstream(req)
             comp_id = f"chatcmpl-{uuid.uuid4().hex[:29]}"
             think_open = False
             yielded_think_buffer = "" 
@@ -190,15 +201,19 @@ class ProxyHandler:
                     if not think_open:
                         yield f"data: {json.dumps({'id': comp_id, 'object': 'chat.completion.chunk', 'created': int(time.time()), 'model': req.model, 'choices': [{'index': 0, 'delta': {'content': '<think>'}, 'finish_reason': None}]})}\n\n"
                         think_open = True
+                    
                     cleaned_full_text = self._clean_thinking_content(text)
-                    delta_to_send = cleaned_full_text[len(yielded_think_buffer):]
+                    delta_to_send = cleaned_full_text[len(yielded_think_buffer):] if cleaned_full_text.startswith(yielded_think_buffer) else cleaned_full_text
+                    
                     if delta_to_send:
                         yield f"data: {json.dumps({'id': comp_id, 'object': 'chat.completion.chunk', 'created': int(time.time()), 'model': req.model, 'choices': [{'index': 0, 'delta': {'content': delta_to_send}, 'finish_reason': None}]})}\n\n"
                     yielded_think_buffer = cleaned_full_text
+
                 elif content_type == "answer":
                     if think_open:
                         yield f"data: {json.dumps({'id': comp_id, 'object': 'chat.completion.chunk', 'created': int(time.time()), 'model': req.model, 'choices': [{'index': 0, 'delta': {'content': '</think>'}, 'finish_reason': None}]})}\n\n"
                         think_open = False
+                    
                     cleaned_text = self._clean_answer_content(text)
                     if cleaned_text:
                         yield f"data: {json.dumps({'id': comp_id, 'object': 'chat.completion.chunk', 'created': int(time.time()), 'model': req.model, 'choices': [{'index': 0, 'delta': {'content': cleaned_text}, 'finish_reason': None}]})}\n\n"
@@ -207,7 +222,6 @@ class ProxyHandler:
                 if resp.status_code != 200:
                     await cookie_manager.mark_cookie_failed(ck); err_body = await resp.aread()
                     err_msg = f"Error: {resp.status_code} - {err_body.decode(errors='ignore')}"
-                    logger.error(f"Upstream error: {err_msg}")
                     err = {"id": comp_id, "object": "chat.completion.chunk", "created": int(time.time()), "model": req.model, "choices": [{"index": 0, "delta": {"content": err_msg}, "finish_reason": "stop"}],}
                     yield f"data: {json.dumps(err)}\n\n"; yield "data: [DONE]\n\n"; return
                 await cookie_manager.mark_cookie_success(ck)
@@ -216,6 +230,7 @@ class ProxyHandler:
                     for line in raw.strip().split('\n'):
                         line = line.strip()
                         if not line.startswith('data: '): continue
+                        
                         payload_str = line[6:]
                         if payload_str == '[DONE]':
                             if think_open: 
@@ -225,16 +240,23 @@ class ProxyHandler:
                             return
                         try:
                             dat = json.loads(payload_str).get("data", {})
-                        except (json.JSONDecodeError, AttributeError): continue
+                        except (json.JSONDecodeError, AttributeError): 
+                            continue
                         
                         phase = dat.get("phase")
                         content_chunk = dat.get("delta_content") or dat.get("edit_content")
-                        if not content_chunk: continue
+
+                        if not content_chunk:
+                            continue
 
                         if phase == "thinking":
-                            current_raw_thinking = content_chunk if dat.get("edit_content") is not None else current_raw_thinking + content_chunk
+                            if dat.get("edit_content") is not None:
+                                current_raw_thinking = content_chunk
+                            else:
+                                current_raw_thinking += content_chunk
                             async for item in yield_delta("thinking", current_raw_thinking):
                                 yield item
+                        
                         elif phase == "answer":
                             content_to_process = content_chunk
                             if is_first_answer_chunk:
@@ -242,6 +264,7 @@ class ProxyHandler:
                                     parts = content_to_process.split('</details>', 1)
                                     content_to_process = parts[1] if len(parts) > 1 else ""
                                 is_first_answer_chunk = False
+                            
                             if content_to_process:
                                 async for item in yield_delta("answer", content_to_process):
                                     yield item
@@ -251,13 +274,12 @@ class ProxyHandler:
     async def non_stream_proxy_response(self, req: ChatCompletionRequest) -> ChatCompletionResponse:
         ck = None
         try:
-            body, headers, ck, url = await self._prep_upstream(req)
+            body, headers, url, ck = await self._prep_upstream(req)
             last_thinking_content = ""
             raw_answer_parts = []
             async with self.client.stream("POST", url, json=body, headers=headers) as resp:
                 if resp.status_code != 200:
                     await cookie_manager.mark_cookie_failed(ck); error_detail = await resp.text()
-                    logger.error(f"Upstream error: {resp.status_code} - {error_detail}")
                     raise HTTPException(resp.status_code, f"Upstream error: {error_detail}")
                 await cookie_manager.mark_cookie_success(ck)
                 
@@ -276,11 +298,17 @@ class ProxyHandler:
                         
                         phase = dat.get("phase")
                         content_chunk = dat.get("delta_content") or dat.get("edit_content")
-                        if not content_chunk: continue
+
+                        if not content_chunk:
+                            continue
 
                         if phase == "thinking":
-                            current_raw_thinking = content_chunk if dat.get("edit_content") is not None else current_raw_thinking + content_chunk
+                            if dat.get("edit_content") is not None:
+                                current_raw_thinking = content_chunk
+                            else:
+                                current_raw_thinking += content_chunk
                             last_thinking_content = current_raw_thinking
+                        
                         elif phase == "answer":
                             content_to_process = content_chunk
                             if is_first_answer_chunk:
@@ -288,6 +316,7 @@ class ProxyHandler:
                                     parts = content_to_process.split('</details>', 1)
                                     content_to_process = parts[1] if len(parts) > 1 else ""
                                 is_first_answer_chunk = False
+                            
                             if content_to_process:
                                 raw_answer_parts.append(content_to_process)
                     else: