local-openai2anthropic 0.1.0__py3-none-any.whl → 0.3.6__py3-none-any.whl

local_openai2anthropic/router.py

@@ -5,10 +5,8 @@ FastAPI router for Anthropic-compatible Messages API.
 
 import json
 import logging
-import secrets
-import string
 from http import HTTPStatus
-from typing import Any, AsyncGenerator
+from typing import Any, cast
 
 import httpx
 from fastapi import APIRouter, Depends, HTTPException, Request
@@ -26,10 +24,44 @@ from local_openai2anthropic.protocol import (
     MessageCreateParams,
 )
 from local_openai2anthropic.server_tools import ServerToolRegistry
+from local_openai2anthropic.streaming import _convert_result_to_stream, _stream_response
+from local_openai2anthropic.tools import (
+    ServerToolHandler,
+    _add_tool_results_to_messages,
+    _handle_with_server_tools,
+)
+from local_openai2anthropic.utils import (
+    _chunk_text,
+    _count_tokens,
+    _estimate_input_tokens,
+    _generate_server_tool_id,
+    _normalize_usage,
+)
 
 logger = logging.getLogger(__name__)
 router = APIRouter()
 
+# Backward compatibility: re-export functions used by tests
+__all__ = [
+    "router",
+    "get_request_settings",
+    "create_message",
+    "list_models",
+    "count_tokens",
+    "health_check",
+    # Backward compatibility exports
+    "_stream_response",
+    "_convert_result_to_stream",
+    "ServerToolHandler",
+    "_handle_with_server_tools",
+    "_add_tool_results_to_messages",
+    "_generate_server_tool_id",
+    "_normalize_usage",
+    "_count_tokens",
+    "_chunk_text",
+    "_estimate_input_tokens",
+]
+
 
 def get_request_settings(request: Request) -> Settings:
     """Resolve Settings from the running app when available.
@@ -43,504 +75,6 @@ def get_request_settings(request: Request) -> Settings:
     return get_settings()
 
 
-def _generate_server_tool_id() -> str:
-    """Generate Anthropic-style server tool use ID (srvtoolu_...)."""
-    # Generate 24 random alphanumeric characters
-    chars = string.ascii_lowercase + string.digits
-    random_part = ''.join(secrets.choice(chars) for _ in range(24))
-    return f"srvtoolu_{random_part}"
-
-
-async def _stream_response(
-    client: httpx.AsyncClient,
-    url: str,
-    headers: dict,
-    json_data: dict,
-    model: str,
-) -> AsyncGenerator[str, None]:
-    """
-    Stream response from OpenAI and convert to Anthropic format.
-    """
-    try:
-        async with client.stream("POST", url, headers=headers, json=json_data) as response:
-            if response.status_code != 200:
-                error_body = await response.aread()
-                try:
-                    error_json = json.loads(error_body.decode())
-                    error_msg = error_json.get("error", {}).get("message", error_body.decode())
-                except json.JSONDecodeError:
-                    error_msg = error_body.decode()
-
-                error_event = AnthropicErrorResponse(
-                    error=AnthropicError(type="api_error", message=error_msg)
-                )
-                yield f"event: error\ndata: {error_event.model_dump_json()}\n\n"
-                yield "data: [DONE]\n\n"
-                return
-
-            # Process SSE stream
-            first_chunk = True
-            content_block_started = False
-            content_block_index = 0
-            finish_reason = None
-            input_tokens = 0
-            output_tokens = 0
-            message_id = None
-
-            async for line in response.aiter_lines():
-                if not line.startswith("data: "):
-                    continue
-
-                data = line[6:]
-                if data == "[DONE]":
-                    break
-
-                try:
-                    chunk = json.loads(data)
-                except json.JSONDecodeError:
-                    continue
-
-                # First chunk: message_start
-                if first_chunk:
-                    message_id = chunk.get("id", "")
-                    usage = chunk.get("usage", {})
-                    input_tokens = usage.get("prompt_tokens", 0)
-
-                    start_event = {
-                        "type": "message_start",
-                        "message": {
-                            "id": message_id,
-                            "type": "message",
-                            "role": "assistant",
-                            "content": [],
-                            "model": model,
-                            "stop_reason": None,
-                            "stop_sequence": None,
-                            "usage": {
-                                "input_tokens": input_tokens,
-                                "output_tokens": 0,
-                                "cache_creation_input_tokens": None,
-                                "cache_read_input_tokens": None,
-                            },
-                        },
-                    }
-                    yield f"event: message_start\ndata: {json.dumps(start_event)}\n\n"
-                    first_chunk = False
-                    continue
-
-                # Handle usage-only chunks
-                if not chunk.get("choices"):
-                    usage = chunk.get("usage", {})
-                    if usage:
-                        if content_block_started:
-                            yield f"event: content_block_stop\ndata: {json.dumps({'type': 'content_block_stop', 'index': content_block_index})}\n\n"
-                            content_block_started = False
-
-                        stop_reason_map = {"stop": "end_turn", "length": "max_tokens", "tool_calls": "tool_use"}
-                        yield f"event: message_delta\ndata: {json.dumps({'type': 'message_delta', 'delta': {'stop_reason': stop_reason_map.get(finish_reason or 'stop', 'end_turn')}, 'usage': {'input_tokens': usage.get('prompt_tokens', 0), 'output_tokens': usage.get('completion_tokens', 0), 'cache_creation_input_tokens': None, 'cache_read_input_tokens': None}})}\n\n"
-                    continue
-
-                choice = chunk["choices"][0]
-                delta = choice.get("delta", {})
-
-                # Track finish reason
-                if choice.get("finish_reason"):
-                    finish_reason = choice["finish_reason"]
-                    continue
-
-                # Handle content
-                if delta.get("content"):
-                    if not content_block_started:
-                        yield f"event: content_block_start\ndata: {json.dumps({'type': 'content_block_start', 'index': content_block_index, 'content_block': {'type': 'text', 'text': ''}})}\n\n"
-                        content_block_started = True
-
-                    yield f"event: content_block_delta\ndata: {json.dumps({'type': 'content_block_delta', 'index': content_block_index, 'delta': {'type': 'text_delta', 'text': delta['content']}})}\n\n"
-
-                # Handle tool calls
-                if delta.get("tool_calls"):
-                    tool_call = delta["tool_calls"][0]
-
-                    if tool_call.get("id"):
-                        if content_block_started:
-                            yield f"event: content_block_stop\ndata: {json.dumps({'type': 'content_block_stop', 'index': content_block_index})}\n\n"
-                            content_block_started = False
-                            content_block_index += 1
-
-                        yield f"event: content_block_start\ndata: {json.dumps({'type': 'content_block_start', 'index': content_block_index, 'content_block': {'type': 'tool_use', 'id': tool_call['id'], 'name': tool_call.get('function', {}).get('name', ''), 'input': {}}})}\n\n"
-                        content_block_started = True
-
-                    elif tool_call.get("function", {}).get("arguments"):
-                        args = tool_call["function"]["arguments"]
-                        yield f"event: content_block_delta\ndata: {json.dumps({'type': 'content_block_delta', 'index': content_block_index, 'delta': {'type': 'input_json_delta', 'partial_json': args}})}\n\n"
-
-            # Close final content block
-            if content_block_started:
-                yield f"event: content_block_stop\ndata: {json.dumps({'type': 'content_block_stop', 'index': content_block_index})}\n\n"
-
-            # Message stop
-            yield f"event: message_stop\ndata: {json.dumps({'type': 'message_stop'})}\n\n"
-            yield "data: [DONE]\n\n"
-
-    except Exception as e:
-        error_event = AnthropicErrorResponse(
-            error=AnthropicError(type="internal_error", message=str(e))
-        )
-        yield f"event: error\ndata: {error_event.model_dump_json()}\n\n"
-        yield "data: [DONE]\n\n"
-
-
-async def _convert_result_to_stream(
-    result: JSONResponse,
-    model: str,
-) -> AsyncGenerator[str, None]:
-    """Convert a JSONResponse to streaming SSE format."""
-    import time
-    
-    body = json.loads(result.body)
-    message_id = body.get("id", f"msg_{int(time.time() * 1000)}")
-    content = body.get("content", [])
-    usage = body.get("usage", {})
-    stop_reason = body.get("stop_reason", "end_turn")
-    
-    # Map stop_reason
-    stop_reason_map = {"end_turn": "stop", "max_tokens": "length", "tool_use": "tool_calls"}
-    openai_stop_reason = stop_reason_map.get(stop_reason, "stop")
-    
-    # 1. message_start event
-    start_event = {
-        "type": "message_start",
-        "message": {
-            "id": message_id,
-            "type": "message",
-            "role": "assistant",
-            "content": [],
-            "model": model,
-            "stop_reason": None,
-            "stop_sequence": None,
-            "usage": {
-                "input_tokens": usage.get("input_tokens", 0),
-                "output_tokens": 0,
-                "cache_creation_input_tokens": usage.get("cache_creation_input_tokens"),
-                "cache_read_input_tokens": usage.get("cache_read_input_tokens"),
-            },
-        },
-    }
-    yield f"event: message_start\ndata: {json.dumps(start_event)}\n\n"
-    
-    # 2. Process content blocks
-    for i, block in enumerate(content):
-        block_type = block.get("type")
-
-        if block_type == "text":
-            yield f"event: content_block_start\ndata: {json.dumps({'type': 'content_block_start', 'index': i, 'content_block': {'type': 'text', 'text': ''}})}\n\n"
-            text = block.get("text", "")
-            yield f"event: content_block_delta\ndata: {json.dumps({'type': 'content_block_delta', 'index': i, 'delta': {'type': 'text_delta', 'text': text}})}\n\n"
-            yield f"event: content_block_stop\ndata: {json.dumps({'type': 'content_block_stop', 'index': i})}\n\n"
-
-        elif block_type == "tool_use":
-            yield f"event: content_block_start\ndata: {json.dumps({'type': 'content_block_start', 'index': i, 'content_block': {'type': 'tool_use', 'id': block.get('id', ''), 'name': block.get('name', ''), 'input': block.get('input', {})}})}\n\n"
-            yield f"event: content_block_stop\ndata: {json.dumps({'type': 'content_block_stop', 'index': i})}\n\n"
-
-        elif block_type == "server_tool_use":
-            # Preserve official Anthropic block type so clients can count server tool uses.
-            yield f"event: content_block_start\ndata: {json.dumps({'type': 'content_block_start', 'index': i, 'content_block': {'type': 'server_tool_use', 'id': block.get('id', ''), 'name': block.get('name', ''), 'input': block.get('input', {})}})}\n\n"
-            yield f"event: content_block_stop\ndata: {json.dumps({'type': 'content_block_stop', 'index': i})}\n\n"
-
-        elif block_type == "web_search_tool_result":
-            # Stream the tool result as its own content block.
-            # Some clients expect `results`, others expect `content`; include both when possible.
-            tool_result_block = dict(block)
-            if "content" not in tool_result_block and "results" in tool_result_block:
-                tool_result_block["content"] = tool_result_block["results"]
-
-            yield f"event: content_block_start\ndata: {json.dumps({'type': 'content_block_start', 'index': i, 'content_block': tool_result_block})}\n\n"
-            yield f"event: content_block_stop\ndata: {json.dumps({'type': 'content_block_stop', 'index': i})}\n\n"
-    
-    # 3. message_delta with final usage
-    delta_event = {
-        "type": "message_delta",
-        "delta": {"stop_reason": stop_reason},
-        "usage": {
-            "input_tokens": usage.get("input_tokens", 0),
-            "output_tokens": usage.get("output_tokens", 0),
-            "cache_creation_input_tokens": usage.get("cache_creation_input_tokens"),
-            "cache_read_input_tokens": usage.get("cache_read_input_tokens"),
-            "server_tool_use": usage.get("server_tool_use"),
-        },
-    }
-    yield f"event: message_delta\ndata: {json.dumps(delta_event)}\n\n"
-    
-    # 4. message_stop
-    yield f"event: message_stop\ndata: {json.dumps({'type': 'message_stop'})}\n\n"
-    yield "data: [DONE]\n\n"
-
-
-class ServerToolHandler:
-    """Handles server tool execution for non-streaming requests."""
-
-    def __init__(
-        self,
-        server_tools: list[type],
-        configs: dict[str, dict[str, Any]],
-        settings: Settings,
-    ):
-        self.server_tools = {t.tool_name: t for t in server_tools}
-        self.configs = configs
-        self.settings = settings
-        self.usage: dict[str, int] = {}
-
-    def is_server_tool_call(self, tool_call: dict[str, Any]) -> bool:
-        """Check if a tool call is for a server tool."""
-        func_name = tool_call.get("function", {}).get("name")
-        return func_name in self.server_tools
-
-    async def execute_tool(
-        self,
-        tool_call: dict[str, Any],
-    ) -> tuple[list[dict[str, Any]], dict[str, Any]]:
-        """
-        Execute a server tool and return content blocks + tool result message.
-        Returns: (content_blocks, tool_result_message)
-        """
-        func_name = tool_call.get("function", {}).get("name")
-        call_id = tool_call.get("id", "")
-
-        tool_class = self.server_tools[func_name]
-        config = self.configs.get(tool_class.tool_type, {})
-
-        # Extract call arguments
-        args = tool_class.extract_call_args(tool_call)
-        if args is None:
-            args = {}
-
-        # Execute the tool
-        result = await tool_class.execute(call_id, args, config, self.settings)
-
-        # Update usage
-        for key, value in result.usage_increment.items():
-            self.usage[key] = self.usage.get(key, 0) + value
-
-        # Build content blocks
-        content_blocks = tool_class.build_content_blocks(call_id, args, result)
-
-        # Build tool result message for OpenAI
-        tool_result_msg = tool_class.build_tool_result_message(call_id, args, result)
-
-        return content_blocks, tool_result_msg
-
-
-async def _handle_with_server_tools(
-    openai_params: dict[str, Any],
-    url: str,
-    headers: dict[str, str],
-    settings: Settings,
-    server_tools: list[type],
-    model: str,
-) -> JSONResponse:
-    """Handle request with server tool execution loop."""
-    params = dict(openai_params)
-    configs = params.pop("_server_tools_config", {})
-
-    handler = ServerToolHandler(server_tools, configs, settings)
-    accumulated_content: list[dict[str, Any]] = []
-
-    # Get max_uses from configs (default to settings or 5)
-    max_uses = settings.websearch_max_uses
-    for config in configs.values():
-        if config.get("max_uses"):
-            max_uses = config["max_uses"]
-            break
-
-    total_tool_calls = 0
-
-    while True:
-        async with httpx.AsyncClient(timeout=settings.request_timeout) as client:
-            try:
-                # Log full request for debugging
-                logger.info(f"Request body: {json.dumps(params, indent=2, default=str)[:3000]}")
-                
-                response = await client.post(url, headers=headers, json=params)
-
-                if response.status_code != 200:
-                    logger.error(f"OpenAI API error: {response.status_code} - {response.text}")
-                    error_response = AnthropicErrorResponse(
-                        error=AnthropicError(type="api_error", message=response.text)
-                    )
-                    return JSONResponse(
-                        status_code=response.status_code,
-                        content=error_response.model_dump(),
-                    )
-
-                completion_data = response.json()
-                logger.info(f"OpenAI response: {json.dumps(completion_data, indent=2)[:500]}...")
-                from openai.types.chat import ChatCompletion
-                completion = ChatCompletion.model_validate(completion_data)
-
-                # Check for server tool calls
-                server_tool_calls = []
-                other_tool_calls = []
-                
-                tool_calls = completion.choices[0].message.tool_calls
-                logger.info(f"Model returned tool_calls: {len(tool_calls) if tool_calls else 0}")
-
-                if tool_calls:
-                    for tc in tool_calls:
-                        func_name = tc.function.name if tc.function else ""
-                        logger.info(f"  Tool call: {func_name}")
-                        
-                        # Generate Anthropic-style ID for server tools
-                        is_server = handler.is_server_tool_call({
-                            "id": tc.id,
-                            "function": {"name": func_name, "arguments": ""},
-                        })
-                        
-                        # Use Anthropic-style ID for server tools, original ID otherwise
-                        tool_id = _generate_server_tool_id() if is_server else tc.id
-                        
-                        tc_dict = {
-                            "id": tool_id,
-                            "function": {
-                                "name": func_name,
-                                "arguments": tc.function.arguments if tc.function else "{}",
-                            },
-                        }
-                        logger.info(f"    Is server tool: {is_server}, ID: {tool_id}")
-                        if is_server:
-                            server_tool_calls.append(tc_dict)
-                        else:
-                            other_tool_calls.append(tc)
-
-                # No server tool calls - we're done
-                logger.info(f"Server tool calls: {len(server_tool_calls)}, Other: {len(other_tool_calls)}")
-                if not server_tool_calls:
-                    message = convert_openai_to_anthropic(completion, model)
-
-                    if accumulated_content:
-                        message_dict = message.model_dump()
-                        message_dict["content"] = accumulated_content + message_dict.get("content", [])
-                        
-                        if message_dict.get("usage"):
-                            message_dict["usage"]["server_tool_use"] = handler.usage
-                        
-                        # Log full response for debugging
-                        logger.info(f"Response content blocks: {json.dumps(message_dict.get('content', []), ensure_ascii=False)[:1000]}")
-                        logger.info(f"Response usage: {message_dict.get('usage')}")
-                        logger.info(f"Server tool use count: {handler.usage}")
-
-                        return JSONResponse(content=message_dict)
-
-                    return JSONResponse(content=message.model_dump())
-
-                # Check max_uses limit
-                if total_tool_calls >= max_uses:
-                    logger.warning(f"Server tool max_uses ({max_uses}) exceeded")
-                    # Return error for each call
-                    for call in server_tool_calls:
-                        func_name = call.get("function", {}).get("name", "")
-                        tool_class = handler.server_tools.get(func_name)
-                        if tool_class:
-                            from local_openai2anthropic.server_tools import ToolResult
-                            error_result = ToolResult(
-                                success=False,
-                                content=[],
-                                error_code="max_uses_exceeded",
-                            )
-                            error_blocks = tool_class.build_content_blocks(
-                                call["id"],
-                                {},
-                                error_result,
-                            )
-                            accumulated_content.extend(error_blocks)
-
-                    # Continue with modified messages
-                    messages = params.get("messages", [])
-                    messages = _add_tool_results_to_messages(
-                        messages, server_tool_calls, handler, is_error=True
-                    )
-                    params["messages"] = messages
-                    continue
-
-                # Execute server tools
-                messages = params.get("messages", [])
-                assistant_tool_calls = []
-                tool_results = []
-
-                for call in server_tool_calls:
-                    total_tool_calls += 1
-                    content_blocks, tool_result = await handler.execute_tool(call)
-                    accumulated_content.extend(content_blocks)
-
-                    # Track for assistant message
-                    assistant_tool_calls.append({
-                        "id": call["id"],
-                        "type": "function",
-                        "function": {
-                            "name": call["function"]["name"],
-                            "arguments": call["function"]["arguments"],
-                        },
-                    })
-                    tool_results.append(tool_result)
-
-                # Add to messages for next iteration
-                messages = _add_tool_results_to_messages(
-                    messages, assistant_tool_calls, handler, tool_results=tool_results
-                )
-                params["messages"] = messages
-
-            except httpx.TimeoutException:
-                error_response = AnthropicErrorResponse(
-                    error=AnthropicError(type="timeout_error", message="Request timed out")
-                )
-                raise HTTPException(
-                    status_code=HTTPStatus.GATEWAY_TIMEOUT,
-                    detail=error_response.model_dump(),
-                )
-            except httpx.RequestError as e:
-                error_response = AnthropicErrorResponse(
-                    error=AnthropicError(type="connection_error", message=str(e))
-                )
-                raise HTTPException(
-                    status_code=HTTPStatus.BAD_GATEWAY,
-                    detail=error_response.model_dump(),
-                )
-
-
-def _add_tool_results_to_messages(
-    messages: list[dict[str, Any]],
-    tool_calls: list[dict[str, Any]],
-    handler: ServerToolHandler,
-    tool_results: list[dict[str, Any]] | None = None,
-    is_error: bool = False,
-) -> list[dict[str, Any]]:
-    """Add assistant tool call and results to messages."""
-    messages = list(messages)
-
-    # Add assistant message with tool calls
-    # SGLang requires content to be a string, not None
-    assistant_msg: dict[str, Any] = {
-        "role": "assistant",
-        "content": "",  # Empty string instead of None for SGLang compatibility
-        "tool_calls": tool_calls,
-    }
-    messages.append(assistant_msg)
-
-    # Add tool results
-    if is_error:
-        for call in tool_calls:
-            messages.append({
-                "role": "tool",
-                "tool_call_id": call["id"],
-                "content": json.dumps({
-                    "error": "max_uses_exceeded",
-                    "message": "Maximum tool uses exceeded.",
-                }),
-            })
-    elif tool_results:
-        messages.extend(tool_results)
-
-    return messages
-
-
 @router.post(
     "/v1/messages",
     response_model=Message,
@@ -562,13 +96,18 @@ async def create_message(
     try:
         body_bytes = await request.body()
         body_json = json.loads(body_bytes.decode("utf-8"))
+        logger.debug(
+            f"[Anthropic Request] {json.dumps(body_json, ensure_ascii=False, indent=2)}"
+        )
         anthropic_params = body_json
     except json.JSONDecodeError as e:
         logger.error(f"Invalid JSON in request body: {e}")
         error_response = AnthropicErrorResponse(
-            error=AnthropicError(type="invalid_request_error", message=f"Invalid JSON: {e}")
+            error=AnthropicError(
+                type="invalid_request_error", message=f"Invalid JSON: {e}"
+            )
         )
-        return JSONResponse(status_code=422, content=error_response.model_dump())
+        return JSONResponse(status_code=400, content=error_response.model_dump())
     except Exception as e:
         logger.error(f"Failed to parse request body: {e}")
         error_response = AnthropicErrorResponse(
@@ -579,30 +118,40 @@ async def create_message(
     # Validate request shape early (avoid making upstream calls for obviously invalid requests)
     if not isinstance(anthropic_params, dict):
         error_response = AnthropicErrorResponse(
-            error=AnthropicError(type="invalid_request_error", message="Request body must be a JSON object")
+            error=AnthropicError(
+                type="invalid_request_error",
+                message="Request body must be a JSON object",
+            )
         )
-        return JSONResponse(status_code=422, content=error_response.model_dump())
+        return JSONResponse(status_code=400, content=error_response.model_dump())
 
     model_value = anthropic_params.get("model")
     if not isinstance(model_value, str) or not model_value.strip():
         error_response = AnthropicErrorResponse(
-            error=AnthropicError(type="invalid_request_error", message="Model must be a non-empty string")
+            error=AnthropicError(
+                type="invalid_request_error", message="Model must be a non-empty string"
+            )
         )
-        return JSONResponse(status_code=422, content=error_response.model_dump())
+        return JSONResponse(status_code=400, content=error_response.model_dump())
 
     messages_value = anthropic_params.get("messages")
     if not isinstance(messages_value, list) or len(messages_value) == 0:
         error_response = AnthropicErrorResponse(
-            error=AnthropicError(type="invalid_request_error", message="Messages must be a non-empty list")
+            error=AnthropicError(
+                type="invalid_request_error",
+                message="Messages must be a non-empty list",
+            )
         )
-        return JSONResponse(status_code=422, content=error_response.model_dump())
+        return JSONResponse(status_code=400, content=error_response.model_dump())
 
     max_tokens_value = anthropic_params.get("max_tokens")
     if not isinstance(max_tokens_value, int):
         error_response = AnthropicErrorResponse(
-            error=AnthropicError(type="invalid_request_error", message="max_tokens is required")
+            error=AnthropicError(
+                type="invalid_request_error", message="max_tokens is required"
+            )
         )
-        return JSONResponse(status_code=422, content=error_response.model_dump())
+        return JSONResponse(status_code=400, content=error_response.model_dump())
 
     # Check for server tools
     tools = anthropic_params.get("tools", [])
@@ -614,11 +163,17 @@ async def create_message(
 
     # Convert Anthropic params to OpenAI params
     openai_params_obj = convert_anthropic_to_openai(
-        anthropic_params,
+        cast(MessageCreateParams, anthropic_params),
         enabled_server_tools=enabled_server_tools if has_server_tools else None,
     )
     openai_params: dict[str, Any] = dict(openai_params_obj)  # type: ignore
 
+    # Log converted OpenAI request (remove internal fields)
+    log_params = {k: v for k, v in openai_params.items() if not k.startswith("_")}
+    logger.debug(
+        f"[OpenAI Request] {json.dumps(log_params, ensure_ascii=False, indent=2)}"
+    )
+
     stream = openai_params.get("stream", False)
     model = openai_params.get("model", "")
 
@@ -644,7 +199,7 @@ async def create_message(
         result = await _handle_with_server_tools(
             openai_params, url, headers, settings, tool_classes, model
         )
-        
+
         # If original request was streaming, convert result to streaming format
         if stream:
             return StreamingResponse(
@@ -665,8 +220,23 @@ async def create_message(
                 response = await client.post(url, headers=headers, json=openai_params)
 
                 if response.status_code != 200:
+                    raw_text = response.text
+                    try:
+                        if not raw_text:
+                            raw_text = response.content.decode(
+                                "utf-8", errors="replace"
+                            )
+                    except Exception:
+                        raw_text = ""
+                    if not raw_text:
+                        raw_text = response.reason_phrase or ""
+                    error_message = (raw_text or "").strip()
                     error_response = AnthropicErrorResponse(
-                        error=AnthropicError(type="api_error", message=response.text)
+                        error=AnthropicError(
+                            type="api_error",
+                            message=error_message
+                            or f"Upstream API error ({response.status_code})",
+                        )
                     )
                     return JSONResponse(
                         status_code=response.status_code,
@@ -674,27 +244,42 @@ async def create_message(
                     )
 
                 openai_completion = response.json()
+                logger.debug(
+                    f"[OpenAI Response] {json.dumps(openai_completion, ensure_ascii=False, indent=2)}"
+                )
+
                 from openai.types.chat import ChatCompletion
+
                 completion = ChatCompletion.model_validate(openai_completion)
                 anthropic_message = convert_openai_to_anthropic(completion, model)
 
-                return JSONResponse(content=anthropic_message.model_dump())
+                anthropic_response = anthropic_message.model_dump()
+                anthropic_response["usage"] = _normalize_usage(
+                    anthropic_response.get("usage")
+                )
+                logger.debug(
+                    f"[Anthropic Response] {json.dumps(anthropic_response, ensure_ascii=False, indent=2)}"
+                )
+
+                return JSONResponse(content=anthropic_response)
 
             except httpx.TimeoutException:
                 error_response = AnthropicErrorResponse(
-                    error=AnthropicError(type="timeout_error", message="Request timed out")
+                    error=AnthropicError(
+                        type="timeout_error", message="Request timed out"
+                    )
                 )
-                raise HTTPException(
+                return JSONResponse(
                     status_code=HTTPStatus.GATEWAY_TIMEOUT,
-                    detail=error_response.model_dump(),
+                    content=error_response.model_dump(),
                 )
             except httpx.RequestError as e:
                 error_response = AnthropicErrorResponse(
                     error=AnthropicError(type="connection_error", message=str(e))
                 )
-                raise HTTPException(
+                return JSONResponse(
                     status_code=HTTPStatus.BAD_GATEWAY,
-                    detail=error_response.model_dump(),
+                    content=error_response.model_dump(),
                 )
 
 
@@ -727,6 +312,112 @@ async def list_models(
             )
 
 
+@router.post("/v1/messages/count_tokens")
+async def count_tokens(
+    request: Request,
+    settings: Settings = Depends(get_request_settings),
+) -> JSONResponse:
+    """
+    Count tokens in messages without creating a message.
+    Uses tiktoken for local token counting.
+    """
+    try:
+        body_bytes = await request.body()
+        body_json = json.loads(body_bytes.decode("utf-8"))
+        logger.debug(
+            f"[Count Tokens Request] {json.dumps(body_json, ensure_ascii=False, indent=2)}"
+        )
+    except json.JSONDecodeError as e:
+        error_response = AnthropicErrorResponse(
+            error=AnthropicError(
+                type="invalid_request_error", message=f"Invalid JSON: {e}"
+            )
+        )
+        return JSONResponse(status_code=400, content=error_response.model_dump())
+    except Exception as e:
+        error_response = AnthropicErrorResponse(
+            error=AnthropicError(type="invalid_request_error", message=str(e))
+        )
+        return JSONResponse(status_code=400, content=error_response.model_dump())
+
+    # Validate required fields
+    if not isinstance(body_json, dict):
+        error_response = AnthropicErrorResponse(
+            error=AnthropicError(
+                type="invalid_request_error",
+                message="Request body must be a JSON object",
+            )
+        )
+        return JSONResponse(status_code=400, content=error_response.model_dump())
+
+    messages = body_json.get("messages", [])
+    if not isinstance(messages, list):
+        error_response = AnthropicErrorResponse(
+            error=AnthropicError(
+                type="invalid_request_error", message="messages must be a list"
+            )
+        )
+        return JSONResponse(status_code=400, content=error_response.model_dump())
+
+    model = body_json.get("model", "")
+    system = body_json.get("system")
+    tools = body_json.get("tools", [])
+
+    try:
+        # Use tiktoken for token counting
+        import tiktoken  # type: ignore[import-not-found]
+
+        # Map model names to tiktoken encoding
+        # Claude models don't have direct tiktoken encodings, so we use cl100k_base as approximation
+        encoding = tiktoken.get_encoding("cl100k_base")
+
+        total_tokens = 0
+
+        # Count system prompt tokens if present
+        if system:
+            if isinstance(system, str):
+                total_tokens += len(encoding.encode(system))
+            elif isinstance(system, list):
+                for block in system:
+                    if isinstance(block, dict) and block.get("type") == "text":
+                        total_tokens += len(encoding.encode(block.get("text", "")))
+
+        # Count message tokens
+        for msg in messages:
+            content = msg.get("content", "")
+            if isinstance(content, str):
+                total_tokens += len(encoding.encode(content))
+            elif isinstance(content, list):
+                for block in content:
+                    if isinstance(block, dict):
+                        if block.get("type") == "text":
+                            total_tokens += len(encoding.encode(block.get("text", "")))
+                        elif block.get("type") == "image":
+                            # Images are typically counted as a fixed number of tokens
+                            # This is an approximation
+                            total_tokens += 85  # Standard approximation for images
+
+        # Count tool definitions tokens
+        if tools:
+            for tool in tools:
+                tool_def = tool if isinstance(tool, dict) else tool.model_dump()
+                # Rough approximation for tool definitions
+                total_tokens += len(encoding.encode(json.dumps(tool_def)))
+
+        logger.debug(f"[Count Tokens Response] input_tokens: {total_tokens}")
+
+        return JSONResponse(content={"input_tokens": total_tokens})
+
+    except Exception as e:
+        logger.error(f"Token counting error: {e}")
+        error_response = AnthropicErrorResponse(
+            error=AnthropicError(
+                type="internal_error", message=f"Failed to count tokens: {str(e)}"
+            )
+        )
+        return JSONResponse(status_code=500, content=error_response.model_dump())
+
+
 @router.get("/health")
 async def health_check() -> dict[str, str]:
     """Health check endpoint."""