diff --git a/letta/llm_api/openai.py b/letta/llm_api/openai.py
index ee4e7954..ca0c25f2 100644
--- a/letta/llm_api/openai.py
+++ b/letta/llm_api/openai.py
@@ -5,7 +5,7 @@ import requests
 from openai import OpenAI
 
 from letta.llm_api.helpers import add_inner_thoughts_to_functions, convert_to_structured_output, make_post_request
-from letta.local_llm.constants import INNER_THOUGHTS_KWARG, INNER_THOUGHTS_KWARG_DESCRIPTION
+from letta.local_llm.constants import INNER_THOUGHTS_KWARG, INNER_THOUGHTS_KWARG_DESCRIPTION, INNER_THOUGHTS_KWARG_DESCRIPTION_GO_FIRST
 from letta.local_llm.utils import num_tokens_from_functions, num_tokens_from_messages
 from letta.schemas.llm_config import LLMConfig
 from letta.schemas.message import Message as _Message
@@ -96,10 +96,15 @@ def build_openai_chat_completions_request(
     max_tokens: Optional[int],
 ) -> ChatCompletionRequest:
     if functions and llm_config.put_inner_thoughts_in_kwargs:
+        # Special case for LM Studio backend since it needs extra guidance to force out the thoughts first
+        # TODO(fix)
+        inner_thoughts_desc = (
+            INNER_THOUGHTS_KWARG_DESCRIPTION_GO_FIRST if ":1234" in llm_config.model_endpoint else INNER_THOUGHTS_KWARG_DESCRIPTION
+        )
         functions = add_inner_thoughts_to_functions(
             functions=functions,
             inner_thoughts_key=INNER_THOUGHTS_KWARG,
-            inner_thoughts_description=INNER_THOUGHTS_KWARG_DESCRIPTION,
+            inner_thoughts_description=inner_thoughts_desc,
         )
 
     openai_message_list = [
diff --git a/letta/local_llm/constants.py b/letta/local_llm/constants.py
index 03abcc81..f4c66a47 100644
--- a/letta/local_llm/constants.py
+++ b/letta/local_llm/constants.py
@@ -27,6 +27,7 @@ DEFAULT_WRAPPER_NAME = "chatml"
 
 INNER_THOUGHTS_KWARG = "inner_thoughts"
 INNER_THOUGHTS_KWARG_DESCRIPTION = "Deep inner monologue private to you only."
+INNER_THOUGHTS_KWARG_DESCRIPTION_GO_FIRST = f"Deep inner monologue private to you only. Think before you act, so always generate arg '{INNER_THOUGHTS_KWARG}' first before any other arg."
 INNER_THOUGHTS_CLI_SYMBOL = "💭"
 
 ASSISTANT_MESSAGE_CLI_SYMBOL = "🤖"
diff --git a/letta/schemas/message.py b/letta/schemas/message.py
index 4377581a..b865671d 100644
--- a/letta/schemas/message.py
+++ b/letta/schemas/message.py
@@ -1,6 +1,7 @@
 import copy
 import json
 import warnings
+from collections import OrderedDict
 from datetime import datetime, timezone
 from typing import Any, Dict, List, Literal, Optional, Union
 
@@ -33,18 +34,18 @@ def add_inner_thoughts_to_tool_call(
     inner_thoughts_key: str,
 ) -> OpenAIToolCall:
     """Add inner thoughts (arg + value) to a tool call"""
-    # because the kwargs are stored as strings, we need to load then write the JSON dicts
     try:
         # load the args list
         func_args = json.loads(tool_call.function.arguments)
-        # add the inner thoughts to the args list
-        func_args[inner_thoughts_key] = inner_thoughts
+        # create new ordered dict with inner thoughts first
+        ordered_args = OrderedDict({inner_thoughts_key: inner_thoughts})
+        # update with remaining args
+        ordered_args.update(func_args)
         # create the updated tool call (as a string)
         updated_tool_call = copy.deepcopy(tool_call)
-        updated_tool_call.function.arguments = json_dumps(func_args)
+        updated_tool_call.function.arguments = json_dumps(ordered_args)
         return updated_tool_call
     except json.JSONDecodeError as e:
-        # TODO: change to logging
         warnings.warn(f"Failed to put inner thoughts in kwargs: {e}")
         raise e
 
diff --git a/letta/server/server.py b/letta/server/server.py
index 1ef4c407..c9780fdd 100644
--- a/letta/server/server.py
+++ b/letta/server/server.py
@@ -397,11 +397,16 @@ class SyncServer(Server):
             )
         # Attempt to enable LM Studio by default
         if model_settings.lmstudio_base_url:
-            self._enabled_providers.append(
-                LMStudioOpenAIProvider(
-                    base_url=model_settings.lmstudio_base_url,
-                )
+            # Auto-append v1 to the base URL
+            lmstudio_url = (
+                model_settings.lmstudio_base_url
+                if model_settings.lmstudio_base_url.endswith("/v1")
+                else model_settings.lmstudio_base_url + "/v1"
             )
+            # Set the OpenAI API key to something non-empty
+            if model_settings.openai_api_key is None:
+                model_settings.openai_api_key = "DUMMY"
+            self._enabled_providers.append(LMStudioOpenAIProvider(base_url=lmstudio_url))
 
     def load_agent(self, agent_id: str, actor: User, interface: Union[AgentInterface, None] = None) -> Agent:
         """Updated method to load agents from persisted storage"""