diff --git a/src/ai.py b/src/ai.py
index be6356f..949be4e 100644
--- a/src/ai.py
+++ b/src/ai.py
@@ -16,7 +16,8 @@ See Also:
 
 import threading
 import time
-import queue
+import json
+import os
 
 import litellm
 import prof
@@ -29,8 +30,10 @@ TOKEN_KEY = "ai_tokens"
 CHAT_WIN_ID: Optional[int] = None
 # Global message history: {win_id: [{"role": "user"|"assistant", "content": str}, ...]}
 CHAT_HISTORY: Dict[str, List[Dict[str, str]]] = {}
-# Output queue for safe display: (win_id, content)
-OUTPUT_QUEUE: queue.Queue[Tuple[str, str]] = queue.Queue()
+# Output file for safe display
+OUTPUT_FILE = "/tmp/ai_output.txt"
+
+FILE_LOCK = threading.Lock()
 
 # Privacy settings for LiteLLM
 litellm.drop_params = True
@@ -73,36 +76,56 @@ def set_model(model) -> None:
     set_default_model(model)
     prof.cons_show(f"Default model set to: {model}")
 
+def run_completion(window_id: str, model: str, history: list, outfile: str, tokens: dict, file_lock) -> None:
+    """Run AI completion and write response to file."""
+    try:
+        response = litellm.completion(
+            model=model,
+            messages=history,
+            api_key=tokens.get(model.split("/")[0], None),
+        ).choices[0].message.content
+        with file_lock:
+            with open(outfile, "a") as f:
+                json.dump({"win_id": window_id, "content": f"AI: {response}"}, f)
+                f.write("\n")
+    except Exception as e:
+        with file_lock:
+            with open(outfile, "a") as f:
+                json.dump({"win_id": window_id, "content": f"Error: {str(e)}"}, f)
+                f.write("\n")
+
 def handler(win_id: str, message: str) -> None:
-    """Process messages in a chat window using the model from the window title."""
+    """Process messages in a chat window using the model from the window title.
+    
+    Args:
+        win_id: Identifier for the chat window, used to extract the model name.
+        message: The user's message to process.
+    """
     model = win_id.split(" - ", 2)[1]
     prof.win_show(win_id, f"Me: {message}")
-    CHAT_HISTORY.setdefault(win_id, []).append({"role": "user", "content": message})
+    
+    # Create a new history list for this call to avoid modifying shared state
+    current_history = CHAT_HISTORY.get(win_id, []) + [{"role": "user", "content": message}]
 
-    def run_completion():
-        tokens = _get_tokens()
-        try:
-            response = litellm.completion(
-                model=model,
-                messages=CHAT_HISTORY[win_id],
-                api_key=tokens.get(model.split("/")[0], None),
-            ).choices[0].message.content
-            CHAT_HISTORY[win_id].append({"role": "assistant", "content": response})
-            OUTPUT_QUEUE.put_nowait((win_id, f"AI: {response}"))
-        except Exception as e:
-            OUTPUT_QUEUE.put_nowait((win_id, f"Error: {str(e)}"))
-
-    thread = threading.Thread(target=run_completion)
+    thread = threading.Thread(target=run_completion, args=(win_id, model, current_history, OUTPUT_FILE, _get_tokens(), FILE_LOCK))
     thread.start()
 
-
 def process_queued_outputs() -> None:
-    """Process one output from the queue using prof.win_show."""
-    try:
-        win_id, content = OUTPUT_QUEUE.get_nowait()
-        prof.win_show(win_id, content)
-    except queue.Empty:
-        pass
+    """Process outputs from the file using prof.win_show."""
+    if os.path.exists(OUTPUT_FILE):
+        with open(OUTPUT_FILE, "r") as f:
+            lines = f.readlines()
+        outputs = []
+        for line in lines:
+            if line.strip():
+                try:
+                    data = json.loads(line)
+                    outputs.append((data["win_id"], data["content"]))
+                except json.JSONDecodeError:
+                    pass
+        for win_id, content in outputs:
+            prof.win_show(win_id, content)
+        open(OUTPUT_FILE, "w").close()  # truncate the file
 
 
 def create_chat_window(model: str) -> str:
@@ -164,21 +187,12 @@ def correct_message(corrected_text: str) -> None:
         if msg["role"] == "user":
             msg["content"] = corrected_text
             break
-    try:
-        response = litellm.completion(
-            model=model,
-            messages=[{"role": "user", "content": corrected_text}],
-            api_key=_get_tokens().get(model.split("/")[0], None),
-        ).choices[0].message.content
-        CHAT_HISTORY[win_id].append({"role": "assistant", "content": response})
-        prof.win_show(win_id, f"AI: {response}")
-    except Exception as e:
-        prof.cons_show(f"Error: {str(e)}")
+    handler(win_id, corrected_text)
 
 
 def get_default_model() -> str:
-    """Retrieve the default model from settings, defaulting to gpt-3.5-turbo."""
-    return prof.settings_string_get("ai_plugin", DEFAULT_MODEL_KEY, "gpt-3.5-turbo")
+    """Retrieve the default model from settings, defaulting to gpt-5.0."""
+    return prof.settings_string_get("ai_plugin", DEFAULT_MODEL_KEY, "openai/gpt-5.0")
 
 
 def set_default_model(model: str) -> None:
@@ -247,7 +261,7 @@ You can see the list of available models here: https://models.litellm.ai/"""
         "/ai",
         "/ai set token openai sk-xxx",
         "/ai set model gpt-4",
-        "/ai start xai/grok",
+        "/ai start xai/grok-4",
         "/ai clear",
         '/ai correct I has a error',
     ]
@@ -255,8 +269,8 @@ You can see the list of available models here: https://models.litellm.ai/"""
     prof.register_command("/ai", 0, 3, synopsis, description, args, examples, _cmd_ai)
     prof.completer_add("/ai", ["set", "start", "clear", "correct"])
     prof.completer_add("/ai set", ["model", "token"])
-    prof.completer_add("/ai set model", ["openai/gpt-4o-mini", "xai/grok"])
-    prof.completer_add("/ai start", ["xai/grok", "openai/gpt-4o"])
+    prof.completer_add("/ai set model", ["openai/gpt-5", "xai/grok-3-mini"])
+    prof.completer_add("/ai start", ["xai/grok-4", "openai/gpt-4o"])
     prof.completer_add("/ai set token", ["openai", "xai"])
 
-    prof.register_timed(process_queued_outputs, 1)  # 1s interval to process AI message output
+    prof.register_timed(process_queued_outputs, 1)  # 1s interval to process AI message output
\ No newline at end of file