fix: address Greptile review feedback

- Log provider token counting failures instead of silently swallowing
- Fall back to local tokenizer when provider returns error response
- Map assistant tool_calls to Responses API function_call items
- Concatenate multiple system messages instead of overwriting
- Hide internal error details from proxy API responses
- Narrow exception catch in handler to network/JSON errors only
- Update test to match new fallback behavior

litellm/llms/openai/responses/count_tokens/handler.py

@@ -4,6 +4,7 @@ OpenAI Responses API token counting handler.
 Uses httpx for HTTP requests to OpenAI's /v1/responses/input_tokens endpoint.
 """
 
+import json
 from typing import Any, Dict, List, Optional, Union
 
 import httpx
@@ -96,7 +97,7 @@ class OpenAICountTokensHandler(OpenAICountTokensConfig):
                 status_code=e.response.status_code,
                 message=e.response.text,
             )
-        except Exception as e:
+        except (httpx.RequestError, json.JSONDecodeError) as e:
             verbose_logger.error(f"Error in CountTokens handler: {str(e)}")
             raise OpenAIError(
                 status_code=500,

litellm/llms/openai/responses/count_tokens/transformation.py

@@ -98,7 +98,7 @@ class OpenAICountTokensConfig:
             from system/developer messages.
         """
         input_items: List[Dict[str, Any]] = []
-        instructions: Optional[str] = None
+        instructions_parts: List[str] = []
 
         for msg in messages:
             role = msg.get("role", "")
@@ -107,7 +107,7 @@ class OpenAICountTokensConfig:
             if role in ("system", "developer"):
                 # Extract system/developer messages as instructions
                 if isinstance(content, str):
-                    instructions = content
+                    instructions_parts.append(content)
                 elif isinstance(content, list):
                     # Handle content blocks - extract text
                     text_parts = []
@@ -116,11 +116,23 @@ class OpenAICountTokensConfig:
                             text_parts.append(block.get("text", ""))
                         elif isinstance(block, str):
                             text_parts.append(block)
-                    instructions = "\n".join(text_parts)
+                    instructions_parts.append("\n".join(text_parts))
             elif role == "user":
                 input_items.append({"role": "user", "content": content})
             elif role == "assistant":
-                input_items.append({"role": "assistant", "content": content})
+                # Map tool_calls to Responses API function_call items
+                tool_calls = msg.get("tool_calls")
+                if tool_calls:
+                    for tc in tool_calls:
+                        func = tc.get("function", {})
+                        input_items.append({
+                            "type": "function_call",
+                            "call_id": tc.get("id", ""),
+                            "name": func.get("name", ""),
+                            "arguments": func.get("arguments", ""),
+                        })
+                else:
+                    input_items.append({"role": "assistant", "content": content})
             elif role == "tool":
                 input_items.append({
                     "type": "function_call_output",
@@ -128,4 +140,5 @@ class OpenAICountTokensConfig:
                     "output": content if isinstance(content, str) else str(content),
                 })
 
+        instructions = "\n".join(instructions_parts) if instructions_parts else None
         return input_items, instructions

litellm/main.py

@@ -7609,10 +7609,12 @@ async def acount_tokens(
                     tools=tools,
                     system=system,
                 )
-                if result is not None:
+                if result is not None and not result.error:
                     return result
-    except Exception:
-        pass
+    except Exception as e:
+        verbose_logger.debug(
+            f"Provider token counting failed for model={model}, falling back to local: {e}"
+        )
 
     # Fallback to local tiktoken-based token counting
     local_count = litellm.token_counter(

litellm/proxy/response_api_endpoints/endpoints.py

@@ -517,7 +517,7 @@ async def count_response_input_tokens(
             )
         )
         raise HTTPException(
-            status_code=500, detail={"error": f"Internal server error: {str(e)}"}
+            status_code=500, detail={"error": "Internal server error"}
         )
 
 

tests/test_litellm/test_count_tokens_public_api.py

@@ -133,9 +133,10 @@ def test_acount_tokens_api_error_falls_back():
             )
         )
 
-        # Should return error response, not raise
-        assert result.error is True
-        assert result.status_code == 401
+        # Should fall back to local tokenizer when provider API errors
+        assert result.error is False
+        assert result.tokenizer_type == "local_tokenizer"
+        assert result.total_tokens > 0
 
 
 def test_acount_tokens_no_api_key_falls_back():