From ff2cfa640ba7d5d2f80111fd62f9b3af5efd7a62 Mon Sep 17 00:00:00 2001
From: PRABHU KIRAN VANDRANKI <72809214+VANDRANKI@users.noreply.github.com>
Date: Tue, 5 May 2026 14:49:26 -0400
Subject: [PATCH] fix(mcp-semantic-filter): use chat format when expanding MCP
 tools for /v1/chat/completions

When mcp_semantic_tool_filter expanded MCP references, it called
_process_mcp_tools_to_openai_format with the default target_format="responses",
which returns flat {"type":"function","name":...} objects (Responses API format).
For /v1/chat/completions backends these must be in the nested
{"type":"function","function":{...}} schema, or strict OpenAI-compatible APIs
(LM Studio, etc.) return 400 BadRequest.

Pass target_format through from _expand_mcp_tools -> _process_mcp_tools_to_openai_format.
In async_pre_call_hook, default to "chat" format and only use "responses" when
call_type=="aresponses". Also fix _get_tool_names_csv to extract tool names from
both the nested (chat) and flat (responses) schemas.
---
 .../proxy/hooks/mcp_semantic_filter/hook.py   | 32 +++++++++++++++----
 1 file changed, 25 insertions(+), 7 deletions(-)

diff --git a/litellm/proxy/hooks/mcp_semantic_filter/hook.py b/litellm/proxy/hooks/mcp_semantic_filter/hook.py
index 4075641d63b..bee3ecccde2 100644
--- a/litellm/proxy/hooks/mcp_semantic_filter/hook.py
+++ b/litellm/proxy/hooks/mcp_semantic_filter/hook.py
@@ -64,12 +64,19 @@ class SemanticToolFilterHook(CustomLogger):
         self,
         tools: List[Any],
         user_api_key_dict: "UserAPIKeyAuth",
+        target_format: str = "chat",
     ) -> List[Dict[str, Any]]:
         """
         Expand MCP references to actual tool definitions.
 
         Reuses LiteLLM_Proxy_MCP_Handler._process_mcp_tools_to_openai_format
         which internally does: parse -> fetch -> filter -> deduplicate -> transform
+
+        Args:
+            tools: List of tool objects which may include MCP references.
+            user_api_key_dict: User auth for access control.
+            target_format: "chat" for ChatCompletions nested ``function`` schema
+                (default), "responses" for Responses API flat schema.
         """
         from litellm.responses.mcp.litellm_proxy_mcp_handler import (
             LiteLLM_Proxy_MCP_Handler,
@@ -87,7 +94,9 @@ class SemanticToolFilterHook(CustomLogger):
             openai_tools,
             _,
         ) = await LiteLLM_Proxy_MCP_Handler._process_mcp_tools_to_openai_format(
-            user_api_key_auth=user_api_key_dict, mcp_tools_with_litellm_proxy=mcp_tools
+            user_api_key_auth=user_api_key_dict,
+            mcp_tools_with_litellm_proxy=mcp_tools,
+            target_format=target_format,  # type: ignore[arg-type]
         )
 
         # Convert Pydantic models to dicts for compatibility
@@ -171,7 +180,12 @@ class SemanticToolFilterHook(CustomLogger):
             )
 
             try:
-                expanded_tools = await self._expand_mcp_tools(tools, user_api_key_dict)
+                # Use "chat" schema (nested function key) for chat completions;
+                # Responses API endpoints expect the flat format.
+                _target_format = "responses" if call_type == "aresponses" else "chat"
+                expanded_tools = await self._expand_mcp_tools(
+                    tools, user_api_key_dict, target_format=_target_format
+                )
 
                 if not expanded_tools:
                     verbose_proxy_logger.warning(
@@ -293,11 +307,15 @@ class SemanticToolFilterHook(CustomLogger):
 
         tool_names = []
         for tool in tools:
-            name = (
-                tool.get("name", "")
-                if isinstance(tool, dict)
-                else getattr(tool, "name", "")
-            )
+            if isinstance(tool, dict):
+                # ChatCompletions format: {"type": "function", "function": {"name": ...}}
+                # Responses API format:   {"type": "function", "name": ...}
+                func = tool.get("function")
+                name = (
+                    func.get("name", "") if isinstance(func, dict) else tool.get("name", "")
+                )
+            else:
+                name = getattr(tool, "name", "")
             if name:
                 tool_names.append(name)