Update openai.py

openai.py

@@ -408,21 +408,15 @@ async def chat_completions(request: Request):
                 )
             break
 
-    # Zai ignores the system role — inject into the first user message
-    system_parts: list[str] = []
-    filtered: list[dict] = []
+    # Zai ignores multi-turn context — flatten all messages into a single
+    # user message with <ROLE> tags so the model sees the full conversation.
+    parts: list[str] = []
     for msg in messages:
-        if msg.get("role") == "system":
-            system_parts.append(msg.get("content", ""))
-        else:
-            filtered.append(msg)
-    if system_parts:
-        system_text = "\n".join(system_parts)
-        for i, msg in enumerate(filtered):
-            if msg.get("role") == "user":
-                filtered[i] = {**msg, "content": f"<SYSTEM_PROMPT>{system_text}</SYSTEM_PROMPT>\n{msg.get('content', '')}"}
-                break
-        messages = filtered
+        role = msg.get("role", "user")
+        content = msg.get("content", "") or ""
+        parts.append(f"<{role.upper()}>{content}</{role.upper()}>")
+    flat_content = "\n".join(parts)
+    messages = [{"role": "user", "content": flat_content}]
 
     if not prompt:
         return JSONResponse(