Update proxy_handler.py

proxy_handler.py

@@ -58,26 +58,14 @@ 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:
@@ -85,20 +73,24 @@ class ProxyHandler:
         cleaned_text = re.sub(r'<summary>.*?</summary>|<glm_block.*?</glm_block>|<[^>]*duration="[^"]*"[^>]*>', '', text, flags=re.DOTALL)
         cleaned_text = cleaned_text.replace("</thinking>", "").replace("<Full>", "").replace("</Full>", "")
         cleaned_text = re.sub(r'</?details[^>]*>', '', cleaned_text)
-        # FIX: Corrected typo from re.MULTiline to re.MULTILINE
         cleaned_text = re.sub(r'^\s*>\s*(?!>)', '', cleaned_text, flags=re.MULTILINE)
         cleaned_text = cleaned_text.replace("Thinking…", "")
         return cleaned_text.strip()
 
     def _clean_answer_content(self, text: str) -> str:
         if not text: return ""
-        cleaned_text = re.sub(r'<glm_block.*?</glm_block>|<details[^>]*>.*?</details>|<summary>.*?</summary>', '', text, flags=re.DOTALL)
+        # Enhanced cleaning rules for the final answer
+        # Rule 1: Remove any complete <details> blocks, which contain citations.
+        cleaned_text = re.sub(r'<details[^>]*>.*?</details>', '', text, flags=re.DOTALL)
+        # Rule 2: Remove other known thinking-related tags.
+        cleaned_text = re.sub(r'<glm_block.*?</glm_block>|<summary>.*?</summary>', '', cleaned_text, flags=re.DOTALL)
+        # Rule 3: Remove leftover tag attributes like duration="", which can leak into the answer.
+        cleaned_text = re.sub(r'\s*duration="\d+"[^>]*>', '', cleaned_text)
         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)
@@ -106,39 +98,26 @@ 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,
@@ -146,16 +125,14 @@ 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": chat_id,
-            "id": request_id,
+            "chat_id": str(uuid.uuid4()),
+            "id": str(uuid.uuid4()),
             "features": {
                 "image_generation": False,
                 "web_search": False,
@@ -190,7 +167,9 @@ class ProxyHandler:
             think_open = False
             yielded_think_buffer = "" 
             current_raw_thinking = ""
-            
+            # FIX: Re-introduce the flag to handle the transition chunk correctly.
+            is_first_answer_chunk = True
+
             async def yield_delta(content_type: str, text: str):
                 nonlocal think_open, yielded_think_buffer
                 if content_type == "thinking" and settings.SHOW_THINK_TAGS:
@@ -224,7 +203,6 @@ 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"
@@ -238,34 +216,38 @@ class ProxyHandler:
                         phase = dat.get("phase")
                         content_chunk = dat.get("delta_content") or dat.get("edit_content")
                         if not content_chunk:
-                            # 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
+                            continue
 
                         if phase == "thinking":
                             current_raw_thinking = content_chunk if dat.get("edit_content") is not None else current_raw_thinking + content_chunk
                             async for item in yield_delta("thinking", current_raw_thinking):
                                 yield item
                         elif phase == "answer":
-                            # MODIFICATION: Removed the special handling for the first answer chunk.
-                            # All chunks are now processed uniformly.
-                            # The _clean_answer_content function, called within yield_delta,
-                            # is responsible for cleaning up any remaining tags.
-                            if content_chunk:
-                                async for item in yield_delta("answer", content_chunk):
+                            content_to_process = content_chunk
+                            # FIX: Restore and improve special handling for the first answer chunk.
+                            if is_first_answer_chunk:
+                                # The first answer chunk often contains leftover closing tags from thinking.
+                                # Find the last '</details>' tag and trim everything before and including it.
+                                last_details_pos = content_to_process.rfind('</details>')
+                                if last_details_pos != -1:
+                                    content_to_process = content_to_process[last_details_pos + len('</details>'):]
+                                
+                                # Clean up any other potential leaked raw text from the thinking phase.
+                                # This regex targets the `true" duration...` pattern.
+                                content_to_process = re.sub(r'^[\s\S]*last_tool_call_name="">', '', content_to_process)
+                                
+                                is_first_answer_chunk = False
+                            
+                            if content_to_process:
+                                async for item in yield_delta("answer", content_to_process):
                                     yield item
         except Exception:
             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:
@@ -275,13 +257,9 @@ 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" # Default
+                finish_reason = "stop" 
                 
                 if "choices" in response_data and response_data["choices"]:
                     first_choice = response_data["choices"][0]
@@ -290,6 +268,9 @@ class ProxyHandler:
                     if "finish_reason" in first_choice:
                         finish_reason = first_choice["finish_reason"]
 
+                # Apply 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()),