jaf-py 2.5.10__py3-none-any.whl → 2.5.11__py3-none-any.whl

jaf/providers/model.py

@@ -16,17 +16,27 @@ from pydantic import BaseModel
 import litellm
 
 from ..core.types import (
-    Agent, ContentRole, Message, ModelProvider, RunConfig, RunState, 
-    CompletionStreamChunk, ToolCallDelta, ToolCallFunctionDelta,
-    MessageContentPart, get_text_content
+    Agent,
+    ContentRole,
+    Message,
+    ModelProvider,
+    RunConfig,
+    RunState,
+    CompletionStreamChunk,
+    ToolCallDelta,
+    ToolCallFunctionDelta,
+    MessageContentPart,
+    get_text_content,
 )
 from ..core.proxy import ProxyConfig
 from ..utils.document_processor import (
-    extract_document_content, is_document_supported, 
-    get_document_description, DocumentProcessingError
+    extract_document_content,
+    is_document_supported,
+    get_document_description,
+    DocumentProcessingError,
 )
 
-Ctx = TypeVar('Ctx')
+Ctx = TypeVar("Ctx")
 
 # Vision model caching
 VISION_MODEL_CACHE_TTL = 5 * 60  # 5 minutes
@@ -34,92 +44,87 @@ VISION_API_TIMEOUT = 3.0  # 3 seconds
 _vision_model_cache: Dict[str, Dict[str, Any]] = {}
 MAX_IMAGE_BYTES = int(os.environ.get("JAF_MAX_IMAGE_BYTES", 8 * 1024 * 1024))
 
+
 async def _is_vision_model(model: str, base_url: str) -> bool:
     """
     Check if a model supports vision capabilities.
-    
+
     Args:
         model: Model name to check
         base_url: Base URL of the LiteLLM server
-        
+
     Returns:
         True if model supports vision, False otherwise
     """
     cache_key = f"{base_url}:{model}"
     cached = _vision_model_cache.get(cache_key)
-    
-    if cached and time.time() - cached['timestamp'] < VISION_MODEL_CACHE_TTL:
-        return cached['supports']
-    
+
+    if cached and time.time() - cached["timestamp"] < VISION_MODEL_CACHE_TTL:
+        return cached["supports"]
+
     try:
         async with httpx.AsyncClient(timeout=VISION_API_TIMEOUT) as client:
             response = await client.get(
-                f"{base_url}/model_group/info",
-                headers={'accept': 'application/json'}
+                f"{base_url}/model_group/info", headers={"accept": "application/json"}
             )
-            
+
             if response.status_code == 200:
                 data = response.json()
                 model_info = None
-                
-                if 'data' in data and isinstance(data['data'], list):
-                    for m in data['data']:
-                        if (m.get('model_group') == model or 
-                            model in str(m.get('model_group', ''))):
+
+                if "data" in data and isinstance(data["data"], list):
+                    for m in data["data"]:
+                        if m.get("model_group") == model or model in str(m.get("model_group", "")):
                             model_info = m
                             break
-                
-                if model_info and 'supports_vision' in model_info:
-                    result = model_info['supports_vision']
-                    _vision_model_cache[cache_key] = {
-                        'supports': result,
-                        'timestamp': time.time()
-                    }
+
+                if model_info and "supports_vision" in model_info:
+                    result = model_info["supports_vision"]
+                    _vision_model_cache[cache_key] = {"supports": result, "timestamp": time.time()}
                     return result
             else:
-                print(f"Warning: Vision API returned status {response.status_code} for model {model}")
-                
+                print(
+                    f"Warning: Vision API returned status {response.status_code} for model {model}"
+                )
+
     except Exception as e:
         print(f"Warning: Vision API error for model {model}: {e}")
-    
+
     # Fallback to known vision models
     known_vision_models = [
-        'gpt-4-vision-preview',
-        'gpt-4o',
-        'gpt-4o-mini',
-        'claude-sonnet-4',
-        'claude-sonnet-4-20250514',
-        'gemini-2.5-flash',
-        'gemini-2.5-pro'
+        "gpt-4-vision-preview",
+        "gpt-4o",
+        "gpt-4o-mini",
+        "claude-sonnet-4",
+        "claude-sonnet-4-20250514",
+        "gemini-2.5-flash",
+        "gemini-2.5-pro",
     ]
-    
+
     is_known_vision_model = any(
-        vision_model.lower() in model.lower()
-        for vision_model in known_vision_models
+        vision_model.lower() in model.lower() for vision_model in known_vision_models
     )
-    
-    _vision_model_cache[cache_key] = {
-        'supports': is_known_vision_model,
-        'timestamp': time.time()
-    }
-    
+
+    _vision_model_cache[cache_key] = {"supports": is_known_vision_model, "timestamp": time.time()}
+
     return is_known_vision_model
 
+
 def make_litellm_provider(
     base_url: str,
     api_key: str = "anything",
     default_timeout: Optional[float] = None,
-    proxy_config: Optional[ProxyConfig] = None
+    proxy_config: Optional[ProxyConfig] = None,
 ) -> ModelProvider[Ctx]:
     """
     Create a LiteLLM-compatible model provider.
-    
+
     Args:
         base_url: Base URL for the LiteLLM server
         api_key: API key (defaults to "anything" for local servers)
         default_timeout: Default timeout for model API calls in seconds
         proxy_config: Optional proxy configuration
-        
+
     Returns:
         ModelProvider instance
     """
@@ -128,48 +133,47 @@ def make_litellm_provider(
         def __init__(self):
             # Default to "anything" if api_key is not provided, for local servers
             effective_api_key = api_key if api_key is not None else "anything"
-            
+
             # Configure HTTP client with proxy support
             client_kwargs = {
                 "base_url": base_url,
                 "api_key": effective_api_key,
             }
-            
+
             if proxy_config:
                 proxies = proxy_config.to_httpx_proxies()
                 if proxies:
                     # Create httpx client with proxy configuration
                     try:
                         # Use the https proxy if available, otherwise http proxy
-                        proxy_url = proxies.get('https://') or proxies.get('http://')
+                        proxy_url = proxies.get("https://") or proxies.get("http://")
                         if proxy_url:
                             http_client = httpx.AsyncClient(proxy=proxy_url)
                             client_kwargs["http_client"] = http_client
                     except Exception as e:
                         print(f"Warning: Could not configure proxy: {e}")
                         # Fall back to environment variables for proxy
-            
+
             self.client = AsyncOpenAI(**client_kwargs)
             self.default_timeout = default_timeout
 
         async def get_completion(
-            self,
-            state: RunState[Ctx],
-            agent: Agent[Ctx, Any],
-            config: RunConfig[Ctx]
+            self, state: RunState[Ctx], agent: Agent[Ctx, Any], config: RunConfig[Ctx]
         ) -> Dict[str, Any]:
             """Get completion from the model."""
 
             # Determine model to use
-            model = (config.model_override or
-                    (agent.model_config.name if agent.model_config else "gpt-4o"))
+            model = config.model_override or (
+                agent.model_config.name if agent.model_config else "gpt-4o"
+            )
 
             # Check if any message contains image content or image attachments
             has_image_content = any(
-                (isinstance(msg.content, list) and
-                 any(part.type == 'image_url' for part in msg.content)) or
-                (msg.attachments and
-                 any(att.kind == 'image' for att in msg.attachments))
+                (
+                    isinstance(msg.content, list)
+                    and any(part.type == "image_url" for part in msg.content)
+                )
+                or (msg.attachments and any(att.kind == "image" for att in msg.attachments))
                 for msg in state.messages
             )
 
@@ -182,51 +186,59 @@ def make_litellm_provider(
                     )
 
             # Create system message
-            system_message = {
-                "role": "system",
-                "content": agent.instructions(state)
-            }
+            system_message = {"role": "system", "content": agent.instructions(state)}
 
             # Convert messages to OpenAI format
             converted_messages = []
             for msg in state.messages:
                 converted_msg = await _convert_message(msg)
                 converted_messages.append(converted_msg)
-            
+
             messages = [system_message] + converted_messages
 
             # Convert tools to OpenAI format
             tools = None
             if agent.tools:
+                # Check if we should inline schema refs
+                inline_refs = (
+                    agent.model_config.inline_tool_schemas if agent.model_config else False
+                )
                 tools = [
                     {
                         "type": "function",
                         "function": {
                             "name": tool.schema.name,
                             "description": tool.schema.description,
-                            "parameters": _pydantic_to_json_schema(tool.schema.parameters),
-                        }
+                            "parameters": _pydantic_to_json_schema(
+                                tool.schema.parameters, inline_refs=inline_refs or False
+                            ),
+                        },
                     }
                     for tool in agent.tools
                 ]
 
             # Determine tool choice behavior
             last_message = state.messages[-1] if state.messages else None
-            is_after_tool_call = last_message and (last_message.role == ContentRole.TOOL or last_message.role == 'tool')
+            is_after_tool_call = last_message and (
+                last_message.role == ContentRole.TOOL or last_message.role == "tool"
+            )
 
             # Prepare request parameters
-            request_params = {
-                "model": model,
-                "messages": messages,
-                "stream": False
-            }
+            request_params = {"model": model, "messages": messages, "stream": False}
 
             # Add optional parameters
             if agent.model_config:
                 if agent.model_config.temperature is not None:
                     request_params["temperature"] = agent.model_config.temperature
-                if agent.model_config.max_tokens is not None:
-                    request_params["max_tokens"] = agent.model_config.max_tokens
+                # Use agent's max_tokens if set, otherwise fall back to config's max_tokens
+                max_tokens = agent.model_config.max_tokens
+                if max_tokens is None:
+                    max_tokens = config.max_tokens
+                if max_tokens is not None:
+                    request_params["max_tokens"] = max_tokens
+            elif config.max_tokens is not None:
+                # No model_config but config has max_tokens
+                request_params["max_tokens"] = config.max_tokens
 
             if tools:
                 request_params["tools"] = tools
@@ -247,12 +259,9 @@ def make_litellm_provider(
             if choice.message.tool_calls:
                 tool_calls = [
                     {
-                        'id': tc.id,
-                        'type': tc.type,
-                        'function': {
-                            'name': tc.function.name,
-                            'arguments': tc.function.arguments
-                        }
+                        "id": tc.id,
+                        "type": tc.type,
+                        "function": {"name": tc.function.name, "arguments": tc.function.arguments},
                     }
                     for tc in choice.message.tool_calls
                 ]
@@ -267,64 +276,64 @@ def make_litellm_provider(
                 }
 
             return {
-                'id': response.id,
-                'created': response.created,
-                'model': response.model,
-                'system_fingerprint': response.system_fingerprint,
-                'message': {
-                    'content': choice.message.content,
-                    'tool_calls': tool_calls
-                },
-                'usage': usage_data,
-                'prompt': messages
+                "id": response.id,
+                "created": response.created,
+                "model": response.model,
+                "system_fingerprint": response.system_fingerprint,
+                "message": {"content": choice.message.content, "tool_calls": tool_calls},
+                "usage": usage_data,
+                "prompt": messages,
             }
 
         async def get_completion_stream(
-            self,
-            state: RunState[Ctx],
-            agent: Agent[Ctx, Any],
-            config: RunConfig[Ctx]
+            self, state: RunState[Ctx], agent: Agent[Ctx, Any], config: RunConfig[Ctx]
         ) -> AsyncIterator[CompletionStreamChunk]:
             """
             Stream completion chunks from the model provider, yielding text deltas and tool-call deltas.
             Uses OpenAI-compatible streaming via LiteLLM endpoint.
             """
             # Determine model to use
-            model = (config.model_override or
-                     (agent.model_config.name if agent.model_config else "gpt-4o"))
+            model = config.model_override or (
+                agent.model_config.name if agent.model_config else "gpt-4o"
+            )
 
             # Create system message
-            system_message = {
-                "role": "system",
-                "content": agent.instructions(state)
-            }
+            system_message = {"role": "system", "content": agent.instructions(state)}
 
-            # Convert messages to OpenAI format  
+            # Convert messages to OpenAI format
             converted_messages = []
             for msg in state.messages:
                 converted_msg = await _convert_message(msg)
                 converted_messages.append(converted_msg)
-            
+
             messages = [system_message] + converted_messages
 
             # Convert tools to OpenAI format
             tools = None
             if agent.tools:
+                # Check if we should inline schema refs
+                inline_refs = (
+                    agent.model_config.inline_tool_schemas if agent.model_config else False
+                )
                 tools = [
                     {
                         "type": "function",
                         "function": {
                             "name": tool.schema.name,
                             "description": tool.schema.description,
-                            "parameters": _pydantic_to_json_schema(tool.schema.parameters),
-                        }
+                            "parameters": _pydantic_to_json_schema(
+                                tool.schema.parameters, inline_refs=inline_refs or False
+                            ),
+                        },
                     }
                     for tool in agent.tools
                 ]
 
             # Determine tool choice behavior
             last_message = state.messages[-1] if state.messages else None
-            is_after_tool_call = last_message and (last_message.role == ContentRole.TOOL or last_message.role == 'tool')
+            is_after_tool_call = last_message and (
+                last_message.role == ContentRole.TOOL or last_message.role == "tool"
+            )
 
             # Prepare request parameters
             request_params: Dict[str, Any] = {
@@ -336,8 +345,15 @@ def make_litellm_provider(
             if agent.model_config:
                 if agent.model_config.temperature is not None:
                     request_params["temperature"] = agent.model_config.temperature
-                if agent.model_config.max_tokens is not None:
-                    request_params["max_tokens"] = agent.model_config.max_tokens
+                # Use agent's max_tokens if set, otherwise fall back to config's max_tokens
+                max_tokens = agent.model_config.max_tokens
+                if max_tokens is None:
+                    max_tokens = config.max_tokens
+                if max_tokens is not None:
+                    request_params["max_tokens"] = max_tokens
+            elif config.max_tokens is not None:
+                # No model_config but config has max_tokens
+                request_params["max_tokens"] = config.max_tokens
 
             if tools:
                 request_params["tools"] = tools
@@ -388,19 +404,20 @@ def make_litellm_provider(
                                     fn = getattr(tc, "function", None)
                                     fn_name = getattr(fn, "name", None) if fn is not None else None
                                     # OpenAI streams "arguments" as incremental deltas
-                                    args_delta = getattr(fn, "arguments", None) if fn is not None else None
+                                    args_delta = (
+                                        getattr(fn, "arguments", None) if fn is not None else None
+                                    )
 
                                     yield CompletionStreamChunk(
                                         tool_call_delta=ToolCallDelta(
                                             index=idx,
                                             id=tc_id,
-                                            type='function',
+                                            type="function",
                                             function=ToolCallFunctionDelta(
-                                                name=fn_name,
-                                                arguments_delta=args_delta
-                                            )
+                                                name=fn_name, arguments_delta=args_delta
+                                            ),
                                         ),
-                                        raw=raw_obj
+                                        raw=raw_obj,
                                     )
                                 except Exception:
                                     # Skip malformed tool-call deltas
@@ -408,26 +425,29 @@ def make_litellm_provider(
 
                     # Completion ended
                     if finish_reason:
-                        yield CompletionStreamChunk(is_done=True, finish_reason=finish_reason, raw=raw_obj)
+                        yield CompletionStreamChunk(
+                            is_done=True, finish_reason=finish_reason, raw=raw_obj
+                        )
                 except Exception:
                     # Skip individual chunk errors, keep streaming
                     continue
 
     return LiteLLMProvider()
 
+
 def make_litellm_sdk_provider(
     api_key: Optional[str] = None,
     model: str = "gpt-3.5-turbo",
     base_url: Optional[str] = None,
     default_timeout: Optional[float] = None,
-    **litellm_kwargs: Any
+    **litellm_kwargs: Any,
 ) -> ModelProvider[Ctx]:
     """
     Create a LiteLLM SDK-based model provider with universal provider support.
-    
+
     LiteLLM automatically detects the provider from the model name and handles
     API key management through environment variables or direct parameters.
-    
+
     Args:
         api_key: API key for the provider (optional, can use env vars)
         model: Model name (e.g., "gpt-4", "claude-3-sonnet", "gemini-pro", "llama2", etc.)
@@ -440,23 +460,23 @@ def make_litellm_sdk_provider(
                          - azure_deployment: "your-deployment" (for Azure OpenAI)
                          - api_base: "https://your-endpoint.com" (custom endpoint)
                          - custom_llm_provider: "custom_provider_name"
-        
+
     Returns:
         ModelProvider instance
-        
+
     Examples:
         # OpenAI
         make_litellm_sdk_provider(api_key="sk-...", model="gpt-4")
-        
+
         # Anthropic Claude
         make_litellm_sdk_provider(api_key="sk-ant-...", model="claude-3-sonnet-20240229")
-        
+
         # Google Gemini
         make_litellm_sdk_provider(model="gemini-pro", vertex_project="my-project")
-        
+
         # Ollama (local)
         make_litellm_sdk_provider(model="ollama/llama2", base_url="http://localhost:11434")
-        
+
         # Azure OpenAI
         make_litellm_sdk_provider(
             model="azure/gpt-4",
@@ -464,13 +484,13 @@ def make_litellm_sdk_provider(
             azure_deployment="gpt-4-deployment",
             api_base="https://your-resource.openai.azure.com"
         )
-        
+
         # Hugging Face
         make_litellm_sdk_provider(
             model="huggingface/microsoft/DialoGPT-medium",
             api_key="hf_..."
         )
-        
+
         # Any custom provider
         make_litellm_sdk_provider(
             model="custom_provider/model-name",
@@ -488,10 +508,7 @@ def make_litellm_sdk_provider(
             self.litellm_kwargs = litellm_kwargs
 
         async def get_completion(
-            self,
-            state: RunState[Ctx],
-            agent: Agent[Ctx, Any],
-            config: RunConfig[Ctx]
+            self, state: RunState[Ctx], agent: Agent[Ctx, Any], config: RunConfig[Ctx]
         ) -> Dict[str, Any]:
             """Get completion from the model using LiteLLM SDK."""
 
@@ -499,10 +516,7 @@ def make_litellm_sdk_provider(
             model_name = config.model_override or self.model
 
             # Create system message
-            system_message = {
-                "role": "system",
-                "content": agent.instructions(state)
-            }
+            system_message = {"role": "system", "content": agent.instructions(state)}
 
             # Convert messages to OpenAI format
             messages = [system_message]
@@ -513,24 +527,26 @@ def make_litellm_sdk_provider(
             # Convert tools to OpenAI format
             tools = None
             if agent.tools:
+                # Check if we should inline schema refs
+                inline_refs = (
+                    agent.model_config.inline_tool_schemas if agent.model_config else False
+                )
                 tools = [
                     {
                         "type": "function",
                         "function": {
                             "name": tool.schema.name,
                             "description": tool.schema.description,
-                            "parameters": _pydantic_to_json_schema(tool.schema.parameters),
-                        }
+                            "parameters": _pydantic_to_json_schema(
+                                tool.schema.parameters, inline_refs=inline_refs or False
+                            ),
+                        },
                     }
                     for tool in agent.tools
                 ]
 
             # Prepare request parameters for LiteLLM
-            request_params = {
-                "model": model_name,
-                "messages": messages,
-                **self.litellm_kwargs
-            }
+            request_params = {"model": model_name, "messages": messages, **self.litellm_kwargs}
 
             # Add API key if provided
             if self.api_key:
@@ -540,8 +556,15 @@ def make_litellm_sdk_provider(
             if agent.model_config:
                 if agent.model_config.temperature is not None:
                     request_params["temperature"] = agent.model_config.temperature
-                if agent.model_config.max_tokens is not None:
-                    request_params["max_tokens"] = agent.model_config.max_tokens
+                # Use agent's max_tokens if set, otherwise fall back to config's max_tokens
+                max_tokens = agent.model_config.max_tokens
+                if max_tokens is None:
+                    max_tokens = config.max_tokens
+                if max_tokens is not None:
+                    request_params["max_tokens"] = max_tokens
+            elif config.max_tokens is not None:
+                # No model_config but config has max_tokens
+                request_params["max_tokens"] = config.max_tokens
 
             if tools:
                 request_params["tools"] = tools
@@ -565,12 +588,9 @@ def make_litellm_sdk_provider(
             if choice.message.tool_calls:
                 tool_calls = [
                     {
-                        'id': tc.id,
-                        'type': tc.type,
-                        'function': {
-                            'name': tc.function.name,
-                            'arguments': tc.function.arguments
-                        }
+                        "id": tc.id,
+                        "type": tc.type,
+                        "function": {"name": tc.function.name, "arguments": tc.function.arguments},
                     }
                     for tc in choice.message.tool_calls
                 ]
@@ -585,23 +605,17 @@ def make_litellm_sdk_provider(
                 }
 
             return {
-                'id': response.id,
-                'created': response.created,
-                'model': response.model,
-                'system_fingerprint': getattr(response, 'system_fingerprint', None),
-                'message': {
-                    'content': choice.message.content,
-                    'tool_calls': tool_calls
-                },
-                'usage': usage_data,
-                'prompt': messages
+                "id": response.id,
+                "created": response.created,
+                "model": response.model,
+                "system_fingerprint": getattr(response, "system_fingerprint", None),
+                "message": {"content": choice.message.content, "tool_calls": tool_calls},
+                "usage": usage_data,
+                "prompt": messages,
             }
 
         async def get_completion_stream(
-            self,
-            state: RunState[Ctx],
-            agent: Agent[Ctx, Any],
-            config: RunConfig[Ctx]
+            self, state: RunState[Ctx], agent: Agent[Ctx, Any], config: RunConfig[Ctx]
         ) -> AsyncIterator[CompletionStreamChunk]:
             """
             Stream completion chunks from the model provider using LiteLLM SDK.
@@ -610,10 +624,7 @@ def make_litellm_sdk_provider(
             model_name = config.model_override or self.model
 
             # Create system message
-            system_message = {
-                "role": "system",
-                "content": agent.instructions(state)
-            }
+            system_message = {"role": "system", "content": agent.instructions(state)}
 
             # Convert messages to OpenAI format
             messages = [system_message]
@@ -624,14 +635,20 @@ def make_litellm_sdk_provider(
             # Convert tools to OpenAI format
             tools = None
             if agent.tools:
+                # Check if we should inline schema refs
+                inline_refs = (
+                    agent.model_config.inline_tool_schemas if agent.model_config else False
+                )
                 tools = [
                     {
                         "type": "function",
                         "function": {
                             "name": tool.schema.name,
                             "description": tool.schema.description,
-                            "parameters": _pydantic_to_json_schema(tool.schema.parameters),
-                        }
+                            "parameters": _pydantic_to_json_schema(
+                                tool.schema.parameters, inline_refs=inline_refs or False
+                            ),
+                        },
                     }
                     for tool in agent.tools
                 ]
@@ -641,7 +658,7 @@ def make_litellm_sdk_provider(
                 "model": model_name,
                 "messages": messages,
                 "stream": True,
-                **self.litellm_kwargs
+                **self.litellm_kwargs,
             }
 
             # Add API key if provided
@@ -652,8 +669,15 @@ def make_litellm_sdk_provider(
             if agent.model_config:
                 if agent.model_config.temperature is not None:
                     request_params["temperature"] = agent.model_config.temperature
-                if agent.model_config.max_tokens is not None:
-                    request_params["max_tokens"] = agent.model_config.max_tokens
+                # Use agent's max_tokens if set, otherwise fall back to config's max_tokens
+                max_tokens = agent.model_config.max_tokens
+                if max_tokens is None:
+                    max_tokens = config.max_tokens
+                if max_tokens is not None:
+                    request_params["max_tokens"] = max_tokens
+            elif config.max_tokens is not None:
+                # No model_config but config has max_tokens
+                request_params["max_tokens"] = config.max_tokens
 
             if tools:
                 request_params["tools"] = tools
@@ -668,12 +692,12 @@ def make_litellm_sdk_provider(
 
             # Stream using litellm
             stream = await litellm.acompletion(**request_params)
-            
+
             async for chunk in stream:
                 try:
                     # Best-effort extraction of raw for debugging
                     try:
-                        raw_obj = chunk.model_dump() if hasattr(chunk, 'model_dump') else None
+                        raw_obj = chunk.model_dump() if hasattr(chunk, "model_dump") else None
                     except Exception:
                         raw_obj = None
 
@@ -702,52 +726,59 @@ def make_litellm_sdk_provider(
                                     tc_id = getattr(tc, "id", None)
                                     fn = getattr(tc, "function", None)
                                     fn_name = getattr(fn, "name", None) if fn is not None else None
-                                    args_delta = getattr(fn, "arguments", None) if fn is not None else None
+                                    args_delta = (
+                                        getattr(fn, "arguments", None) if fn is not None else None
+                                    )
 
                                     yield CompletionStreamChunk(
                                         tool_call_delta=ToolCallDelta(
                                             index=idx,
                                             id=tc_id,
-                                            type='function',
+                                            type="function",
                                             function=ToolCallFunctionDelta(
-                                                name=fn_name,
-                                                arguments_delta=args_delta
-                                            )
+                                                name=fn_name, arguments_delta=args_delta
+                                            ),
                                         ),
-                                        raw=raw_obj
+                                        raw=raw_obj,
                                     )
                                 except Exception:
                                     continue
 
                     # Completion ended
                     if finish_reason:
-                        yield CompletionStreamChunk(is_done=True, finish_reason=finish_reason, raw=raw_obj)
+                        yield CompletionStreamChunk(
+                            is_done=True, finish_reason=finish_reason, raw=raw_obj
+                        )
                 except Exception:
                     continue
 
     return LiteLLMSDKProvider()
 
+
 async def _convert_message(msg: Message) -> Dict[str, Any]:
     """
     Handles all possible role types (string and enum) and content formats.
     """
     # Normalize role to handle both string and enum values
-    role_value = msg.role.value if hasattr(msg.role, 'value') else str(msg.role).lower()
+    role_value = msg.role.value if hasattr(msg.role, "value") else str(msg.role).lower()
 
     # Handle user messages
-    if role_value in ('user', ContentRole.USER.value if hasattr(ContentRole, 'USER') else 'user'):
+    if role_value in ("user", ContentRole.USER.value if hasattr(ContentRole, "USER") else "user"):
         if isinstance(msg.content, list):
             # Multi-part content
             return {
                 "role": "user",
-                "content": [_convert_content_part(part) for part in msg.content]
+                "content": [_convert_content_part(part) for part in msg.content],
             }
         else:
             # Build message with attachments if available
-            return await _build_chat_message_with_attachments('user', msg)
+            return await _build_chat_message_with_attachments("user", msg)
 
     # Handle assistant messages
-    elif role_value in ('assistant', ContentRole.ASSISTANT.value if hasattr(ContentRole, 'ASSISTANT') else 'assistant'):
+    elif role_value in (
+        "assistant",
+        ContentRole.ASSISTANT.value if hasattr(ContentRole, "ASSISTANT") else "assistant",
+    ):
         result = {
             "role": "assistant",
             "content": get_text_content(msg.content) or "",  # Ensure content is never None
@@ -759,10 +790,7 @@ async def _convert_message(msg: Message) -> Dict[str, Any]:
                 {
                     "id": tc.id,
                     "type": tc.type,
-                    "function": {
-                        "name": tc.function.name,
-                        "arguments": tc.function.arguments
-                    }
+                    "function": {"name": tc.function.name, "arguments": tc.function.arguments},
                 }
                 for tc in msg.tool_calls
                 if tc.id and tc.function and tc.function.name  # Validate tool call structure
@@ -771,37 +799,37 @@ async def _convert_message(msg: Message) -> Dict[str, Any]:
         return result
 
     # Handle system messages
-    elif role_value in ('system', ContentRole.SYSTEM.value if hasattr(ContentRole, 'SYSTEM') else 'system'):
-        return {
-            "role": "system",
-            "content": get_text_content(msg.content) or ""
-        }
+    elif role_value in (
+        "system",
+        ContentRole.SYSTEM.value if hasattr(ContentRole, "SYSTEM") else "system",
+    ):
+        return {"role": "system", "content": get_text_content(msg.content) or ""}
 
     # Handle tool messages
-    elif role_value in ('tool', ContentRole.TOOL.value if hasattr(ContentRole, 'TOOL') else 'tool'):
+    elif role_value in ("tool", ContentRole.TOOL.value if hasattr(ContentRole, "TOOL") else "tool"):
         if not msg.tool_call_id:
             raise ValueError(f"Tool message must have tool_call_id. Message: {msg}")
 
         return {
             "role": "tool",
             "content": get_text_content(msg.content) or "",
-            "tool_call_id": msg.tool_call_id
+            "tool_call_id": msg.tool_call_id,
         }
 
     # Handle function messages (legacy support)
-    elif role_value == 'function':
+    elif role_value == "function":
         if not msg.tool_call_id:
             raise ValueError(f"Function message must have tool_call_id. Message: {msg}")
 
         return {
             "role": "function",
             "content": get_text_content(msg.content) or "",
-            "name": getattr(msg, 'name', 'unknown_function')
+            "name": getattr(msg, "name", "unknown_function"),
         }
 
     # Unknown role - provide helpful error message
     else:
-        available_roles = ['user', 'assistant', 'system', 'tool', 'function']
+        available_roles = ["user", "assistant", "system", "tool", "function"]
         raise ValueError(
             f"Unknown message role: {msg.role} (type: {type(msg.role)}). "
             f"Supported roles: {available_roles}. "
@@ -811,46 +839,31 @@ async def _convert_message(msg: Message) -> Dict[str, Any]:
 
 def _convert_content_part(part: MessageContentPart) -> Dict[str, Any]:
     """Convert MessageContentPart to OpenAI format."""
-    if part.type == 'text':
-        return {
-            "type": "text",
-            "text": part.text
-        }
-    elif part.type == 'image_url':
-        return {
-            "type": "image_url",
-            "image_url": part.image_url
-        }
-    elif part.type == 'file':
-        return {
-            "type": "file",
-            "file": part.file
-        }
+    if part.type == "text":
+        return {"type": "text", "text": part.text}
+    elif part.type == "image_url":
+        return {"type": "image_url", "image_url": part.image_url}
+    elif part.type == "file":
+        return {"type": "file", "file": part.file}
     else:
         raise ValueError(f"Unknown content part type: {part.type}")
 
 
-async def _build_chat_message_with_attachments(
-    role: str,
-    msg: Message
-) -> Dict[str, Any]:
+async def _build_chat_message_with_attachments(role: str, msg: Message) -> Dict[str, Any]:
     """
     Build multi-part content for Chat Completions if attachments exist.
     Supports images via image_url and documents via content extraction.
     """
     has_attachments = msg.attachments and len(msg.attachments) > 0
     if not has_attachments:
-        if role == 'assistant':
+        if role == "assistant":
             base_msg = {"role": "assistant", "content": get_text_content(msg.content)}
             if msg.tool_calls:
                 base_msg["tool_calls"] = [
                     {
                         "id": tc.id,
                         "type": tc.type,
-                        "function": {
-                            "name": tc.function.name,
-                            "arguments": tc.function.arguments
-                        }
+                        "function": {"name": tc.function.name, "arguments": tc.function.arguments},
                     }
                     for tc in msg.tool_calls
                 ]
@@ -863,7 +876,7 @@ async def _build_chat_message_with_attachments(
         parts.append({"type": "text", "text": text_content})
 
     for att in msg.attachments:
-        if att.kind == 'image':
+        if att.kind == "image":
             # Prefer explicit URL; otherwise construct a data URL from base64
             url = att.url
             if not url and att.data and att.mime_type:
@@ -871,100 +884,168 @@ async def _build_chat_message_with_attachments(
                 try:
                     # Estimate decoded size (base64 is ~4/3 of decoded size)
                     estimated_size = len(att.data) * 3 // 4
-                    
+
                     if estimated_size > MAX_IMAGE_BYTES:
-                        print(f"Warning: Skipping oversized image ({estimated_size} bytes > {MAX_IMAGE_BYTES}). "
-                              f"Set JAF_MAX_IMAGE_BYTES env var to adjust limit.")
-                        parts.append({
-                            "type": "text",
-                            "text": f"[IMAGE SKIPPED: Size exceeds limit of {MAX_IMAGE_BYTES//1024//1024}MB. "
-                                   f"Image name: {att.name or 'unnamed'}]"
-                        })
+                        print(
+                            f"Warning: Skipping oversized image ({estimated_size} bytes > {MAX_IMAGE_BYTES}). "
+                            f"Set JAF_MAX_IMAGE_BYTES env var to adjust limit."
+                        )
+                        parts.append(
+                            {
+                                "type": "text",
+                                "text": f"[IMAGE SKIPPED: Size exceeds limit of {MAX_IMAGE_BYTES // 1024 // 1024}MB. "
+                                f"Image name: {att.name or 'unnamed'}]",
+                            }
+                        )
                         continue
-                    
+
                     # Create data URL for valid-sized images
                     url = f"data:{att.mime_type};base64,{att.data}"
                 except Exception as e:
                     print(f"Error processing image data: {e}")
-                    parts.append({
-                        "type": "text",
-                        "text": f"[IMAGE ERROR: Failed to process image data. Image name: {att.name or 'unnamed'}]"
-                    })
+                    parts.append(
+                        {
+                            "type": "text",
+                            "text": f"[IMAGE ERROR: Failed to process image data. Image name: {att.name or 'unnamed'}]",
+                        }
+                    )
                     continue
-            
+
             if url:
-                parts.append({
-                    "type": "image_url",
-                    "image_url": {"url": url}
-                })
-        
-        elif att.kind in ['document', 'file']:
+                parts.append({"type": "image_url", "image_url": {"url": url}})
+
+        elif att.kind in ["document", "file"]:
             # Check if attachment has use_litellm_format flag or is a large document
             use_litellm_format = att.use_litellm_format is True
-            
+
             if use_litellm_format and (att.url or att.data):
                 # For now, fall back to content extraction since most providers don't support native file format
                 # TODO: Add provider-specific file format support
-                print(f"Info: LiteLLM format requested for {att.name}, falling back to content extraction")
+                print(
+                    f"Info: LiteLLM format requested for {att.name}, falling back to content extraction"
+                )
                 use_litellm_format = False
-            
+
             if not use_litellm_format:
                 # Extract document content if supported and we have data or URL
                 if is_document_supported(att.mime_type) and (att.data or att.url):
                     try:
                         processed = await extract_document_content(att)
-                        file_name = att.name or 'document'
+                        file_name = att.name or "document"
                         description = get_document_description(att.mime_type)
-                        
-                        parts.append({
-                            "type": "text",
-                            "text": f"DOCUMENT: {file_name} ({description}):\n\n{processed.content}"
-                        })
+
+                        parts.append(
+                            {
+                                "type": "text",
+                                "text": f"DOCUMENT: {file_name} ({description}):\n\n{processed.content}",
+                            }
+                        )
                     except DocumentProcessingError as e:
                         # Fallback to filename if extraction fails
-                        label = att.name or att.format or att.mime_type or 'attachment'
-                        parts.append({
-                            "type": "text",
-                            "text": f"ERROR: Failed to process {att.kind}: {label} ({e})"
-                        })
+                        label = att.name or att.format or att.mime_type or "attachment"
+                        parts.append(
+                            {
+                                "type": "text",
+                                "text": f"ERROR: Failed to process {att.kind}: {label} ({e})",
+                            }
+                        )
                 else:
                     # Unsupported document type - show placeholder
-                    label = att.name or att.format or att.mime_type or 'attachment'
+                    label = att.name or att.format or att.mime_type or "attachment"
                     url_info = f" ({att.url})" if att.url else ""
-                    parts.append({
-                        "type": "text",
-                        "text": f"ATTACHMENT: {att.kind}: {label}{url_info}"
-                    })
+                    parts.append(
+                        {"type": "text", "text": f"ATTACHMENT: {att.kind}: {label}{url_info}"}
+                    )
 
     base_msg = {"role": role, "content": parts}
-    if role == 'assistant' and msg.tool_calls:
+    if role == "assistant" and msg.tool_calls:
         base_msg["tool_calls"] = [
             {
                 "id": tc.id,
                 "type": tc.type,
-                "function": {
-                    "name": tc.function.name,
-                    "arguments": tc.function.arguments
-                }
+                "function": {"name": tc.function.name, "arguments": tc.function.arguments},
             }
             for tc in msg.tool_calls
         ]
-    
+
     return base_msg
 
-def _pydantic_to_json_schema(model_class: type[BaseModel]) -> Dict[str, Any]:
+
+def _resolve_schema_refs(
+    schema: Dict[str, Any], defs: Optional[Dict[str, Any]] = None
+) -> Dict[str, Any]:
+    """
+    Recursively resolve $ref references in a JSON schema by inlining definitions.
+
+    Args:
+        schema: The schema object to process (may contain $ref)
+        defs: The $defs dictionary containing reusable definitions
+
+    Returns:
+        Schema with all references resolved inline
+    """
+    if defs is None:
+        # Extract $defs from root schema if present
+        defs = schema.get("$defs", {})
+
+    # If this is a reference, resolve it
+    if isinstance(schema, dict) and "$ref" in schema:
+        ref_path = schema["$ref"]
+
+        # Handle #/$defs/DefinitionName format
+        if ref_path.startswith("#/$defs/"):
+            def_name = ref_path.split("/")[-1]
+            if def_name in defs:
+                # Recursively resolve the definition (it might have refs too)
+                resolved_def = _resolve_schema_refs(defs[def_name], defs)
+                return resolved_def
+            else:
+                # If definition not found, return the original ref
+                return schema
+        else:
+            # Other ref formats - return as is
+            return schema
+
+    # If this is a dict, recursively process all values
+    if isinstance(schema, dict):
+        result = {}
+        for key, value in schema.items():
+            # Skip $defs as we're inlining them
+            if key == "$defs":
+                continue
+            result[key] = _resolve_schema_refs(value, defs)
+        return result
+
+    # If this is a list, recursively process all items
+    if isinstance(schema, list):
+        return [_resolve_schema_refs(item, defs) for item in schema]
+
+    # For primitive types, return as is
+    return schema
+
+
+def _pydantic_to_json_schema(
+    model_class: type[BaseModel], inline_refs: bool = False
+) -> Dict[str, Any]:
     """
     Convert a Pydantic model to JSON schema for OpenAI tools.
-    
+
     Args:
         model_class: Pydantic model class
-        
+        inline_refs: If True, resolve $refs and inline $defs in the schema
+
     Returns:
         JSON schema dictionary
     """
-    if hasattr(model_class, 'model_json_schema'):
+    if hasattr(model_class, "model_json_schema"):
         # Pydantic v2
-        return model_class.model_json_schema()
+        schema = model_class.model_json_schema()
     else:
         # Pydantic v1 fallback
-        return model_class.schema()
+        schema = model_class.schema()
+
+    # If inline_refs is True, resolve all references
+    if inline_refs:
+        schema = _resolve_schema_refs(schema)
+
+    return schema