Update proxy_handler.py

proxy_handler.py

@@ -58,14 +58,26 @@ class ProxyHandler:
         Returns:
             A dictionary with 'signature' and 'timestamp'.
         """
+        # --- Updated logic from signature_generator.py ---
         timestamp_ms = self._get_timestamp_millis()
+        
+        # 1. Base64 encode the message content (t_payload)
         b64_encoded_t = base64.b64encode(t_payload.encode("utf-8")).decode("utf-8")
+
+        # 2. Construct the message string for signing using the b64 encoded payload
         message_string = f"{e_payload}|{b64_encoded_t}|{timestamp_ms}"
+        
+        # 3. Calculate the time bucket 'n' (5-minute intervals)
         n = timestamp_ms // (5 * 60 * 1000) 
+        
+        # 4. Intermediate key derivation
         msg1 = str(n).encode("utf-8")
         intermediate_key = hmac.new(self.primary_secret, msg1, hashlib.sha256).hexdigest()
+
+        # 5. Final signature
         msg2 = message_string.encode("utf-8")
         final_signature = hmac.new(intermediate_key.encode("utf-8"), msg2, hashlib.sha256).hexdigest()
+
         return {"signature": final_signature, "timestamp": timestamp_ms}
 
     def _clean_thinking_content(self, text: str) -> str:
@@ -79,15 +91,13 @@ class ProxyHandler:
 
     def _clean_answer_content(self, text: str) -> str:
         if not text: return ""
-        # This function cleans tags that might appear in any subsequent chunk, not just the first one.
-        cleaned_text = re.sub(r'<details[^>]*>.*?</details>', '', text, flags=re.DOTALL)
-        cleaned_text = re.sub(r'<glm_block.*?</glm_block>|<summary>.*?</summary>', '', cleaned_text, flags=re.DOTALL)
-        cleaned_text = re.sub(r'\s*duration="\d+"[^>]*>', '', cleaned_text)
+        cleaned_text = re.sub(r'<glm_block.*?</glm_block>|<details[^>]*>.*?</details>|<summary>.*?</summary>', '', text, flags=re.DOTALL)
         return cleaned_text
 
     def _serialize_msgs(self, msgs) -> list:
         out = []
         for m in msgs:
+            # Adapting to Pydantic v1/v2 and dicts
             if hasattr(m, "dict"): out.append(m.dict())
             elif hasattr(m, "model_dump"): out.append(m.model_dump())
             elif isinstance(m, dict): out.append(m)
@@ -95,26 +105,39 @@ 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."""
         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
+        chat_id = str(uuid.uuid4())
+        request_id = str(uuid.uuid4())
         
+        # --- NEW Simplified Signature Payload Logic ---
+        user_info = self._parse_jwt_token(ck)
+        user_id = user_info.get("user_id", "")
+        # The reference code uses a separate UUID for user_id in payload, let's follow that.
+        # This seems strange, but let's replicate the reference code exactly.
         payload_user_id = str(uuid.uuid4())
         payload_request_id = str(uuid.uuid4())
         payload_timestamp = str(self._get_timestamp_millis())
         
+        # e: The simplified payload for the signature
         e_payload = f"requestId,{payload_request_id},timestamp,{payload_timestamp},user_id,{payload_user_id}"
+
+        # t: The last message content
         t_payload = ""
         if req.messages:
             last_message = req.messages[-1]
             if isinstance(last_message.content, str):
                 t_payload = last_message.content
 
+        # Generate the signature
         signature_data = self._generate_signature(e_payload, t_payload)
         signature = signature_data["signature"]
         signature_timestamp = signature_data["timestamp"]
         
+        # The reference code sends these as URL parameters, not in the body.
         url_params = {
             "requestId": payload_request_id,
             "timestamp": payload_timestamp,
@@ -122,14 +145,16 @@ class ProxyHandler:
             "signature_timestamp": str(signature_timestamp)
         }
         
+        # Construct URL with query parameters
+        # Note: The reference code has a typo `f"{BASE_URL}/api/chat/completions"`, it should be `z.ai`
         final_url = httpx.URL(settings.UPSTREAM_URL).copy_with(params=url_params)
 
         body = {
             "stream": True,
             "model": model,
             "messages": self._serialize_msgs(req.messages),
-            "chat_id": str(uuid.uuid4()),
-            "id": str(uuid.uuid4()),
+            "chat_id": chat_id,
+            "id": request_id,
             "features": {
                 "image_generation": False,
                 "web_search": False,
@@ -181,7 +206,6 @@ class ProxyHandler:
                     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
-                    # The _clean_answer_content is a general cleaner for subsequent chunks.
                     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"
@@ -200,6 +224,7 @@ class ProxyHandler:
                         line = line.strip()
                         if not line.startswith('data: '): continue
                         payload_str = line[6:]
+                        # The reference code has a special 'done' phase, but the original Z.AI uses [DONE]
                         if payload_str == '[DONE]':
                             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"
@@ -213,7 +238,11 @@ class ProxyHandler:
                         phase = dat.get("phase")
                         content_chunk = dat.get("delta_content") or dat.get("edit_content")
                         if not content_chunk:
-                            continue
+                            # Handle case where chunk is just usage info, etc.
+                            if phase == 'other' and dat.get('usage'):
+                                pass # In streaming, usage might come with the final chunk
+                            else:
+                                continue
 
                         if phase == "thinking":
                             current_raw_thinking = content_chunk if dat.get("edit_content") is not None else current_raw_thinking + content_chunk
@@ -222,15 +251,12 @@ class ProxyHandler:
                         elif phase == "answer":
                             content_to_process = content_chunk
                             if is_first_answer_chunk:
-                                last_bracket_pos = content_to_process.rfind('>')
-                                if last_bracket_pos != -1:
-                                    content_to_process = content_to_process[last_bracket_pos + 1:]
-                                
-                                # FINAL FIX: Remove any leading whitespace after slicing.
-                                content_to_process = content_to_process.lstrip()
-                                
+                                # MODIFICATION START: Fix for lost characters after thinking phase
+                                # Use a robust regex to remove everything up to and including the first '</details>' tag.
+                                # This is more reliable than split() for handling stream boundaries.
+                                content_to_process = re.sub(r'^.*?</details>', '', content_to_process, count=1, flags=re.DOTALL)
+                                # MODIFICATION END
                                 is_first_answer_chunk = False
-                            
                             if content_to_process:
                                 async for item in yield_delta("answer", content_to_process):
                                     yield item
@@ -238,9 +264,12 @@ class ProxyHandler:
             logger.exception("Stream error"); raise
 
     async def non_stream_proxy_response(self, req: ChatCompletionRequest) -> ChatCompletionResponse:
+        # This part of the code can be simplified as well, but let's focus on fixing the streaming first.
+        # The logic will be almost identical to the streaming one.
         ck = None
         try:
             body, headers, ck, url = await self._prep_upstream(req)
+            # For non-stream, set stream to False in the body
             body["stream"] = False
             
             async with self.client.post(url, json=body, headers=headers) as resp:
@@ -250,9 +279,13 @@ class ProxyHandler:
                     raise HTTPException(resp.status_code, f"Upstream error: {error_detail}")
                 
                 await cookie_manager.mark_cookie_success(ck)
+                
+                # Z.AI non-stream response is a single JSON object
                 response_data = resp.json()
+                
+                # We need to adapt Z.AI's response format to OpenAI's format
                 final_content = ""
-                finish_reason = "stop" 
+                finish_reason = "stop" # Default
                 
                 if "choices" in response_data and response_data["choices"]:
                     first_choice = response_data["choices"][0]
@@ -261,9 +294,6 @@ class ProxyHandler:
                     if "finish_reason" in first_choice:
                         finish_reason = first_choice["finish_reason"]
 
-                # Apply general cleaning to non-streamed responses as well.
-                final_content = self._clean_answer_content(final_content)
-
                 return ChatCompletionResponse(
                     id=response_data.get("id", f"chatcmpl-{uuid.uuid4().hex[:29]}"),
                     created=int(time.time()),