kader 0.1.0__py3-none-any.whl

kader/agent/base.py

@@ -0,0 +1,920 @@
+"""
+Base Agent Implementation.
+
+Defines the BaseAgent class which serves as the foundation for creating specific agents
+with tools, memory, and LLM provider integration.
+"""
+
+from pathlib import Path
+from typing import AsyncIterator, Iterator, Optional, Union
+
+import yaml
+from tenacity import RetryError, stop_after_attempt, wait_exponential
+
+from kader.memory import (
+    ConversationManager,
+    FileSessionManager,
+    SlidingWindowConversationManager,
+)
+from kader.prompts.base import PromptBase
+from kader.providers.base import (
+    BaseLLMProvider,
+    LLMResponse,
+    Message,
+    ModelConfig,
+    StreamChunk,
+)
+from kader.providers.ollama import OllamaProvider
+from kader.tools import BaseTool, ToolRegistry
+
+from .logger import agent_logger
+
+
+class BaseAgent:
+    """
+    Base class for Agents.
+
+    Combines tools, memory, and an LLM provider to perform tasks.
+    Supports synchronous and asynchronous invocation and streaming.
+    Includes built-in retry logic using tenacity.
+    Supports session persistence via FileSessionManager.
+    """
+
+    def __init__(
+        self,
+        name: str,
+        system_prompt: Union[str, PromptBase],
+        tools: Union[list[BaseTool], ToolRegistry, None] = None,
+        provider: Optional[BaseLLMProvider] = None,
+        memory: Optional[ConversationManager] = None,
+        retry_attempts: int = 3,
+        model_name: str = "qwen3-coder:480b-cloud",
+        session_id: Optional[str] = None,
+        use_persistence: bool = False,
+        interrupt_before_tool: bool = True,
+        tool_confirmation_callback: Optional[callable] = None,
+    ) -> None:
+        """
+        Initialize the Base Agent.
+
+        Args:
+            name: Name of the agent.
+            system_prompt: The system prompt definition.
+            tools: List of tools or a ToolRegistry.
+            provider: LLM provider instance. If None, uses OllamaProvider.
+            memory: Conversation/Memory manager. If None, uses SlidingWindowConversationManager.
+            retry_attempts: Number of retry attempts for LLM calls (default: 3).
+            model_name: Default model name if creating a default Ollama provider.
+            session_id: Optional session ID to load/resume.
+            use_persistence: If True, enables session persistence (auto-enabled if session_id provided).
+            interrupt_before_tool: If True, pauses and asks for user confirmation before executing tools.
+            tool_confirmation_callback: Optional callback function for tool confirmation.
+                Signature: (message: str) -> tuple[bool, Optional[str]]
+                Returns (should_execute, user_elaboration_if_declined).
+        """
+        self.name = name
+        self.system_prompt = system_prompt
+        self.retry_attempts = retry_attempts
+        self.interrupt_before_tool = interrupt_before_tool
+        self.tool_confirmation_callback = tool_confirmation_callback
+
+        # Persistence Configuration
+        self.session_id = session_id
+        self.use_persistence = use_persistence or (session_id is not None)
+        self.session_manager = FileSessionManager() if self.use_persistence else None
+
+        # Initialize Logger if agent uses persistence (logs only if there's a session)
+        self.logger_id = None
+        if self.use_persistence:
+            # Only create logger if we have a session_id or if the session manager will create one
+            session_id_for_logger = self.session_id
+            if not session_id_for_logger and self.session_manager:
+                # If no session_id yet but persistence is enabled, we'll get one during _load_session
+                pass  # We'll set up the logger in _load_session if needed
+            if session_id_for_logger:
+                self.logger_id = agent_logger.setup_logger(
+                    self.name, session_id_for_logger
+                )
+
+        # Initialize Provider
+        if provider:
+            self.provider = provider
+        else:
+            self.provider = OllamaProvider(model=model_name)
+
+        # Initialize Memory
+        if memory:
+            self.memory = memory
+        else:
+            self.memory = SlidingWindowConversationManager()
+
+        # Initialize Tools
+        self._tool_registry = ToolRegistry()
+        if tools:
+            if isinstance(tools, ToolRegistry):
+                self._tool_registry = tools
+            elif isinstance(tools, list):
+                for tool in tools:
+                    self._tool_registry.register(tool)
+
+        if self.use_persistence:
+            self._load_session()
+
+        # Propagate session to tools
+        self._propagate_session_to_tools()
+
+        # Update config with tools if provider supports it
+        self._update_provider_tools()
+
+    def _load_session(self) -> None:
+        """Load conversation history from session storage."""
+        if not self.session_manager:
+            return
+
+        if not self.session_id:
+            session = self.session_manager.create_session(self.name)
+            self.session_id = session.session_id
+
+        # Initialize logger if we now have a session_id and logging hasn't been set up yet
+        if self.use_persistence and not self.logger_id and self.session_id:
+            self.logger_id = agent_logger.setup_logger(self.name, self.session_id)
+
+        # Propagate session to tools
+        self._propagate_session_to_tools()
+
+        # Load conversation history
+        try:
+            # We don't check if session exists first because load_conversation
+            # handles missing sessions by returning empty list (usually)
+            # or we catch the error. FileSessionManager.load_conversation returns list[dict].
+            history = self.session_manager.load_conversation(self.session_id)
+            if history:
+                # Add loaded messages to memory
+                # ConversationManager supports adding dicts directly
+                self.memory.add_messages(history)
+        except Exception:
+            # If session doesn't exist or error, we start fresh (or could log warning)
+            # For now, we silently proceed with empty memory
+            pass
+
+    def _propagate_session_to_tools(self) -> None:
+        """Propagate current session ID to all registered tools."""
+        if not self.session_id:
+            return
+
+        for tool in self._tool_registry.tools:
+            if hasattr(tool, "set_session_id"):
+                tool.set_session_id(self.session_id)
+
+    def _save_session(self) -> None:
+        """Save current conversation history to session storage."""
+        if not self.session_manager or not self.session_id:
+            return
+
+        try:
+            # Get all messages from memory
+            # Convert ConversationMessage to dict (using .message property)
+            messages = [msg.message for msg in self.memory.get_messages()]
+            self.session_manager.save_conversation(self.session_id, messages)
+        except Exception:
+            # Log error or handle silently? Best not to crash main flow on save failure
+            pass
+
+    @property
+    def tools_map(self) -> dict[str, BaseTool]:
+        """
+        Get a dictionary mapping tool names to tool instances.
+
+        Returns:
+            Dictionary of {tool_name: tool_instance}
+        """
+        # Access private attribute of ToolRegistry if needed, or iterate
+        # ToolRegistry has .tools property which returns a list
+        return {tool.name: tool for tool in self._tool_registry.tools}
+
+    def _update_provider_tools(self) -> None:
+        """Update the provider's default config with registered tools."""
+        if not self._tool_registry.tools:
+            return
+
+        # We need to update the default config of the provider to include these tools
+        # Since we can't easily modify the internal default_config of the provider cleanly
+        # from here without accessing protected members, strict encapsulation might prevent this.
+        # However, for this implementation, we will pass tools during invoke if they exist.
+        pass
+
+    def _get_run_config(self, config: Optional[ModelConfig] = None) -> ModelConfig:
+        """Prepare execution config with tools."""
+        base_config = config or ModelConfig()
+
+        # If tools are available and not explicitly disabled or overridden
+        if self._tool_registry.tools and not base_config.tools:
+            # Detect provider type to format tools correctly
+            # Defaulting to 'openai' format as it's the de-facto standard
+            provider_type = "openai"
+            if isinstance(self.provider, OllamaProvider):
+                provider_type = "ollama"
+
+            base_config = ModelConfig(
+                temperature=base_config.temperature,
+                max_tokens=base_config.max_tokens,
+                top_p=base_config.top_p,
+                top_k=base_config.top_k,
+                frequency_penalty=base_config.frequency_penalty,
+                stop_sequences=base_config.stop_sequences,
+                stream=base_config.stream,
+                tools=self._tool_registry.to_provider_format(provider_type),
+                tool_choice=base_config.tool_choice,
+                extra=base_config.extra,
+            )
+
+        return base_config
+
+    def _prepare_messages(
+        self, messages: Union[str, list[Message], list[dict]]
+    ) -> list[Message]:
+        """Prepare messages adding system prompt and history."""
+        # Normalize input to list of Message objects
+        input_msgs: list[Message] = []
+        if isinstance(messages, str):
+            input_msgs = [Message.user(messages)]
+        elif isinstance(messages, list):
+            if not messages:
+                pass
+            elif isinstance(messages[0], dict):
+                # Convert dicts to Messages
+                pass  # simplified for now, assuming user passes Message objects or string
+                # But we should handle it better
+                input_msgs = [
+                    Message(**msg) if isinstance(msg, dict) else msg for msg in messages
+                ]
+            else:
+                input_msgs = messages  # Assuming list[Message]
+
+        # Add to memory
+        for msg in input_msgs:
+            self.memory.add_message(msg)
+
+        # Retrieve context (system prompt + history)
+        # 1. Start with System Prompt
+        if isinstance(self.system_prompt, PromptBase):
+            sys_prompt_content = self.system_prompt.resolve_prompt()
+        else:
+            sys_prompt_content = str(self.system_prompt)
+
+        final_messages = [Message.system(sys_prompt_content)]
+
+        # 2. Get history from memory (windowed)
+        # memory.apply_window() returns list[dict], need to convert back to Message
+        history_dicts = self.memory.apply_window()
+
+        # We need to act smart here. invoke/stream usually take the *new* messages
+        # plus history. Memory managers usually store everything.
+        # If we added input_msgs to memory, apply_window should return them too if relevant.
+        # So we just use what Memory gives us.
+
+        for msg_dict in history_dicts:
+            # Basic conversion from dict back to Message
+            # Note: conversation.py Message support might be limited to dicts
+            msg = Message(
+                role=msg_dict.get("role"),
+                content=msg_dict.get("content"),
+                name=msg_dict.get("name"),
+                tool_call_id=msg_dict.get("tool_call_id"),
+                tool_calls=msg_dict.get("tool_calls"),
+            )
+            final_messages.append(msg)
+
+        return final_messages
+
+    def _format_tool_call_for_display(self, tool_call_dict: dict) -> str:
+        """
+        Format a tool call for display to the user.
+
+        Args:
+            tool_call_dict: The tool call dictionary from LLM response.
+
+        Returns:
+            The tool's interruption message.
+        """
+        import json
+
+        fn_info = tool_call_dict.get("function", {})
+        if not fn_info and "name" in tool_call_dict:
+            fn_info = tool_call_dict
+
+        tool_name = fn_info.get("name", "unknown")
+        arguments = fn_info.get("arguments", {})
+
+        # Parse arguments if string
+        if isinstance(arguments, str):
+            try:
+                arguments = json.loads(arguments)
+            except json.JSONDecodeError:
+                pass
+
+        # Get the tool's interruption message if available
+        tool = self._tool_registry.get(tool_name)
+        if tool and isinstance(arguments, dict):
+            return tool.get_interruption_message(**arguments)
+
+        # Fallback for unknown tools
+        return f"execute {tool_name}"
+
+    def _confirm_tool_execution(
+        self, tool_call_dict: dict
+    ) -> tuple[bool, Optional[str]]:
+        """
+        Ask user for confirmation before executing a tool.
+
+        Args:
+            tool_call_dict: The tool call dictionary from LLM response.
+
+        Returns:
+            Tuple of (should_execute: bool, user_input: Optional[str]).
+            If should_execute is False, user_input contains additional context.
+        """
+        display_str = self._format_tool_call_for_display(tool_call_dict)
+
+        # Use callback if provided (e.g., for GUI/TUI)
+        if self.tool_confirmation_callback:
+            return self.tool_confirmation_callback(display_str)
+
+        # Default: use console input
+        print(display_str)
+
+        while True:
+            user_input = input("\nExecute this tool? (yes/no): ").strip().lower()
+
+            if user_input in ("yes", "y"):
+                return True, None
+            elif user_input in ("no", "n"):
+                elaboration = input(
+                    "Please provide more context or instructions: "
+                ).strip()
+                return False, elaboration if elaboration else None
+            else:
+                print("Please enter 'yes' or 'no'.")
+
+    async def _aconfirm_tool_execution(
+        self, tool_call_dict: dict
+    ) -> tuple[bool, Optional[str]]:
+        """
+        Async version - Ask user for confirmation before executing a tool.
+
+        Note: This uses synchronous input() as async stdin is complex.
+        For production use, consider using aioconsole or similar.
+
+        Args:
+            tool_call_dict: The tool call dictionary from LLM response.
+
+        Returns:
+            Tuple of (should_execute: bool, user_input: Optional[str]).
+        """
+        # For simplicity, we use the sync version in async context
+        # In production, use asyncio.to_thread or aioconsole
+        import asyncio
+
+        return await asyncio.to_thread(self._confirm_tool_execution, tool_call_dict)
+
+    def _process_tool_calls(
+        self, response: LLMResponse
+    ) -> Union[list[Message], tuple[bool, str]]:
+        """
+        Execute tool calls from response and return tool messages.
+
+        Args:
+            response: The LLM response containing tool calls.
+
+        Returns:
+            List of Message objects representing tool results, or
+            Tuple of (False, user_input) if user declined tool execution.
+        """
+        tool_messages = []
+        if response.has_tool_calls:
+            for tool_call_dict in response.tool_calls:
+                # Check for interrupt before tool execution
+                if self.interrupt_before_tool:
+                    should_execute, user_input = self._confirm_tool_execution(
+                        tool_call_dict
+                    )
+                    if not should_execute:
+                        # Return the user's elaboration to be processed
+                        return (False, user_input)
+
+                # Need to convert dict to ToolCall object or handle manually
+                # ToolRegistry.run takes ToolCall
+                from kader.tools.base import ToolCall
+
+                # Create ToolCall object
+                # Some providers might differ in specific dict keys, relying on normalization
+                try:
+                    tool_call = ToolCall(
+                        id=tool_call_dict.get("id", ""),
+                        name=tool_call_dict.get("function", {}).get("name", ""),
+                        arguments=tool_call_dict.get("function", {}).get(
+                            "arguments", {}
+                        ),
+                        raw_arguments=str(
+                            tool_call_dict.get("function", {}).get("arguments", {})
+                        ),
+                    )
+                except Exception:
+                    # Fallback or simplified parsing if structure differs
+                    tool_call = ToolCall(
+                        id=tool_call_dict.get("id", ""),
+                        name=tool_call_dict.get("function", {}).get("name", ""),
+                        arguments={},  # Error case
+                    )
+
+                # Execute tool
+                tool_result = self._tool_registry.run(tool_call)
+
+                # add result to memory
+                # But here we just return messages, caller handles memory add
+                tool_msg = Message.tool(
+                    tool_call_id=tool_result.tool_call_id, content=tool_result.content
+                )
+                tool_messages.append(tool_msg)
+
+        return tool_messages
+
+    async def _aprocess_tool_calls(
+        self, response: LLMResponse
+    ) -> Union[list[Message], tuple[bool, str]]:
+        """
+        Async version of _process_tool_calls.
+
+        Returns:
+            List of Message objects representing tool results, or
+            Tuple of (False, user_input) if user declined tool execution.
+        """
+        tool_messages = []
+        if response.has_tool_calls:
+            for tool_call_dict in response.tool_calls:
+                # Check for interrupt before tool execution
+                if self.interrupt_before_tool:
+                    should_execute, user_input = await self._aconfirm_tool_execution(
+                        tool_call_dict
+                    )
+                    if not should_execute:
+                        return (False, user_input)
+
+                from kader.tools.base import ToolCall
+
+                # Check structure - Ollama/OpenAI usually: {'id':..., 'type': 'function', 'function': {'name':.., 'arguments':..}}
+                fn_info = tool_call_dict.get("function", {})
+                if not fn_info and "name" in tool_call_dict:
+                    # Handle flat structure if any
+                    fn_info = tool_call_dict
+
+                tool_call = ToolCall(
+                    id=tool_call_dict.get("id", "call_default"),
+                    name=fn_info.get("name", ""),
+                    arguments=fn_info.get("arguments", {}),
+                )
+
+                # Execute tool async
+                tool_result = await self._tool_registry.arun(tool_call)
+
+                tool_msg = Message.tool(
+                    tool_call_id=tool_result.tool_call_id, content=tool_result.content
+                )
+                tool_messages.append(tool_msg)
+
+        return tool_messages
+
+    # -------------------------------------------------------------------------
+    # Synchronous Methods
+    # -------------------------------------------------------------------------
+
+    def invoke(
+        self, messages: Union[str, list[Message]], config: Optional[ModelConfig] = None
+    ) -> LLMResponse:
+        """
+        Synchronously invoke the agent.
+
+        Handles message preparation, LLM invocation with retries, and tool execution loop.
+        """
+        # Retry decorator wrapper logic
+        # Since tenacity decorators wrap functions, we define an inner function or use the decorator on a method
+        # but we want dynamic retry attempts (from self) which decorators strictly speaking don't support easily without specialized usage.
+        # We will use the functional API of tenacity for dynamic configuration.
+        from tenacity import Retrying
+
+        runner = Retrying(
+            stop=stop_after_attempt(self.retry_attempts),
+            wait=wait_exponential(multiplier=1, min=4, max=10),
+            reraise=True,
+        )
+
+        final_response = None
+
+        # Main Agent Loop (Limit turns to avoid infinite loops)
+        max_turns = 10
+        current_turn = 0
+
+        while current_turn < max_turns:
+            current_turn += 1
+
+            # Prepare full context
+            full_history = self._prepare_messages(messages if current_turn == 1 else [])
+            # Note: _prepare_messages adds input to memory. On subsequent turns (tools),
+            # we don't re-add the user input. self.memory already has it + previous turns.
+
+            # Call LLM with retry
+            try:
+                response = runner(
+                    self.provider.invoke, full_history, self._get_run_config(config)
+                )
+            except RetryError as e:
+                # Should not happen with reraise=True, but just in case
+                raise e
+
+            # Add assistant response to memory
+            self.memory.add_message(response.to_message())
+
+            # Log the interaction if logger is active
+            if self.logger_id:
+                # Extract token usage info if available
+                token_usage = None
+                if hasattr(response, "usage"):
+                    token_usage = {
+                        "prompt_tokens": getattr(response.usage, "prompt_tokens", 0),
+                        "completion_tokens": getattr(
+                            response.usage, "completion_tokens", 0
+                        ),
+                        "total_tokens": getattr(response.usage, "total_tokens", 0),
+                    }
+
+                # Log the LLM response
+                agent_logger.log_llm_response(self.logger_id, str(response.content))
+
+                # Log token usage and calculate cost
+                if token_usage:
+                    agent_logger.log_token_usage(
+                        self.logger_id,
+                        token_usage["prompt_tokens"],
+                        token_usage["completion_tokens"],
+                        token_usage["total_tokens"],
+                    )
+
+                    # Calculate and log cost
+                    agent_logger.calculate_cost(
+                        self.logger_id,
+                        token_usage["prompt_tokens"],
+                        token_usage["completion_tokens"],
+                        getattr(self.provider, "model", ""),
+                    )
+
+            # Save session update
+            if self.use_persistence:
+                self._save_session()
+
+            # Check for tool calls
+            if response.has_tool_calls:
+                tool_result = self._process_tool_calls(response)
+
+                # Check if user declined tool execution
+                if isinstance(tool_result, tuple) and tool_result[0] is False:
+                    # User declined - add their input as a new message and continue
+                    user_elaboration = tool_result[1]
+                    if user_elaboration:
+                        self.memory.add_message(Message.user(user_elaboration))
+                    else:
+                        # User provided no elaboration, return current response
+                        final_response = response
+                        break
+                    continue
+
+                tool_msgs = tool_result
+
+                # Add tool outputs to memory
+                for tm in tool_msgs:
+                    self.memory.add_message(tm)
+
+                    # Log tool usage
+                    if self.logger_id:
+                        # Extract tool name and arguments
+                        tool_name = "unknown"
+                        arguments = {}
+                        if hasattr(tm, "tool_call_id"):
+                            # This is a tool message, need to find the tool name
+                            # We'll check the original response to find the tool
+                            for tool_call in response.tool_calls:
+                                fn_info = tool_call.get("function", {})
+                                if fn_info.get("name"):
+                                    tool_name = fn_info.get("name", "unknown")
+                                    arguments = fn_info.get("arguments", {})
+                                    agent_logger.log_tool_usage(
+                                        self.logger_id, tool_name, arguments
+                                    )
+                                    break
+
+                # Save session update after tool results
+                if self.use_persistence:
+                    self._save_session()
+
+                # Loop continues to feed tool outputs back to LLM
+                continue
+            else:
+                # No tools, final response
+                final_response = response
+                break
+
+        return final_response
+
+    def stream(
+        self, messages: Union[str, list[Message]], config: Optional[ModelConfig] = None
+    ) -> Iterator[StreamChunk]:
+        """
+        Synchronously stream the agent response.
+
+        Note: Tool execution breaks streaming flow typically.
+        If tools are called, we consume the stream to execute tools, then stream the final answer.
+        """
+        # For simplicity in this base implementation, we'll only stream if there are no tool calls initially,
+        # or we buffer if we detect tools. Logic can get complex.
+
+        # Current simplified approach:
+        # 1. Prepare messages
+        full_history = self._prepare_messages(messages)
+
+        # 2. Stream from provider
+        # We need to handle retries for the *start* of the stream
+        from tenacity import Retrying
+
+        runner = Retrying(
+            stop=stop_after_attempt(self.retry_attempts),
+            wait=wait_exponential(multiplier=1, min=4, max=10),
+            reraise=True,
+        )
+
+        # We can't retry the *iteration* easily if it fails mid-stream without complex logic.
+        # We will retry obtaining the iterator.
+        stream_iterator = runner(
+            self.provider.stream, full_history, self._get_run_config(config)
+        )
+
+        yield from stream_iterator
+
+        # Update session at end if needed
+        # Note: Streaming complicates memory/persistence because getting the full message
+        # requires aggregating chunks. The current implementation of base.stream DOES NOT
+        # auto-aggregate into memory (it just yields).
+        # The USER of stream() is responsible for re-assembling the message and adding to memory
+        # if they want history.
+        # BUT, wait. _prepare_messages DOES add input messages to memory.
+        # The RESPONSE is not added here.
+        # TODO: A robust stream implementation should aggregate and save.
+        # For now, we only save the input part since _prepare_messages called it.
+        if self.use_persistence:
+            self._save_session()
+
+    # -------------------------------------------------------------------------
+    # Asynchronous Methods
+    # -------------------------------------------------------------------------
+
+    async def ainvoke(
+        self, messages: Union[str, list[Message]], config: Optional[ModelConfig] = None
+    ) -> LLMResponse:
+        """Asynchronous invocation with retries and tool loop."""
+        from tenacity import AsyncRetrying
+
+        runner = AsyncRetrying(
+            stop=stop_after_attempt(self.retry_attempts),
+            wait=wait_exponential(multiplier=1, min=4, max=10),
+            reraise=True,
+        )
+
+        max_turns = 10
+        current_turn = 0
+        final_response = None
+
+        while current_turn < max_turns:
+            current_turn += 1
+            full_history = self._prepare_messages(messages if current_turn == 1 else [])
+
+            response = await runner(
+                self.provider.ainvoke, full_history, self._get_run_config(config)
+            )
+
+            self.memory.add_message(response.to_message())
+
+            # Log the interaction if logger is active
+            if self.logger_id:
+                # Extract token usage info if available
+                token_usage = None
+                if hasattr(response, "usage"):
+                    token_usage = {
+                        "prompt_tokens": getattr(response.usage, "prompt_tokens", 0),
+                        "completion_tokens": getattr(
+                            response.usage, "completion_tokens", 0
+                        ),
+                        "total_tokens": getattr(response.usage, "total_tokens", 0),
+                    }
+
+                # Log the LLM response
+                agent_logger.log_llm_response(self.logger_id, str(response.content))
+
+                # Log token usage and calculate cost
+                if token_usage:
+                    agent_logger.log_token_usage(
+                        self.logger_id,
+                        token_usage["prompt_tokens"],
+                        token_usage["completion_tokens"],
+                        token_usage["total_tokens"],
+                    )
+
+                    # Calculate and log cost
+                    agent_logger.calculate_cost(
+                        self.logger_id,
+                        token_usage["prompt_tokens"],
+                        token_usage["completion_tokens"],
+                        getattr(self.provider, "model", ""),
+                    )
+
+            # Save session update
+            if self.use_persistence:
+                self._save_session()
+
+            if response.has_tool_calls:
+                tool_result = await self._aprocess_tool_calls(response)
+
+                # Check if user declined tool execution
+                if isinstance(tool_result, tuple) and tool_result[0] is False:
+                    # User declined - add their input as a new message and continue
+                    user_elaboration = tool_result[1]
+                    if user_elaboration:
+                        self.memory.add_message(Message.user(user_elaboration))
+                    else:
+                        final_response = response
+                        break
+                    continue
+
+                tool_msgs = tool_result
+
+                for tm in tool_msgs:
+                    self.memory.add_message(tm)
+
+                    # Log tool usage
+                    if self.logger_id:
+                        # Extract tool name and arguments
+                        tool_name = "unknown"
+                        arguments = {}
+                        if hasattr(tm, "tool_call_id"):
+                            # This is a tool message, need to find the tool name
+                            # We'll check the original response to find the tool
+                            for tool_call in response.tool_calls:
+                                fn_info = tool_call.get("function", {})
+                                if fn_info.get("name"):
+                                    tool_name = fn_info.get("name", "unknown")
+                                    arguments = fn_info.get("arguments", {})
+                                    agent_logger.log_tool_usage(
+                                        self.logger_id, tool_name, arguments
+                                    )
+                                    break
+
+                # Save session update
+                if self.use_persistence:
+                    self._save_session()
+                continue
+            else:
+                final_response = response
+                break
+
+        return final_response
+
+    async def astream(
+        self, messages: Union[str, list[Message]], config: Optional[ModelConfig] = None
+    ) -> AsyncIterator[StreamChunk]:
+        """Asynchronous streaming with memory aggregation."""
+        # Prepare messages
+        full_history = self._prepare_messages(messages)
+
+        # Determine config
+        run_config = self._get_run_config(config)
+
+        # Get stream iterator directly (cannot use tenacity on async generator creation easily)
+        stream_iterator = self.provider.astream(full_history, run_config)
+
+        aggregated_content = ""
+        aggregated_tool_calls = []
+
+        async for chunk in stream_iterator:
+            aggregated_content += chunk.content
+            if chunk.tool_calls:
+                # TODO: robust tool call aggregation if streaming partial JSON
+                # For now, assume provider yields complete tool calls in chunks or we just collect them
+                aggregated_tool_calls.extend(chunk.tool_calls)
+            yield chunk
+
+        # Create Message and add to memory
+        # Note: If no content and no tools, we don't add (or adds empty message)
+
+        # If we have tool calls, we might need to properly format them
+        final_msg = Message(
+            role="assistant",
+            content=aggregated_content,
+            tool_calls=aggregated_tool_calls if aggregated_tool_calls else None,
+        )
+
+        self.memory.add_message(final_msg)
+
+        if self.use_persistence:
+            self._save_session()
+
+    # -------------------------------------------------------------------------
+    # Serialization Methods
+    # -------------------------------------------------------------------------
+
+    def to_yaml(self, path: Union[str, Path]) -> None:
+        """
+        Serialize agent configuration to YAML.
+
+        Args:
+           path: File path to save YAML.
+        """
+        system_prompt_str = (
+            self.system_prompt.resolve_prompt()
+            if isinstance(self.system_prompt, PromptBase)
+            else str(self.system_prompt)
+        )
+        data = {
+            "name": self.name,
+            "system_prompt": system_prompt_str,
+            "retry_attempts": self.retry_attempts,
+            "provider": {
+                "model": self.provider.model,
+                # Add other provider settings if possible
+            },
+            "tools": self._tool_registry.names,
+            # Memory state could be saved here too, but usually configured separately
+        }
+
+        path_obj = Path(path)
+        path_obj.parent.mkdir(parents=True, exist_ok=True)
+
+        with open(path_obj, "w", encoding="utf-8") as f:
+            yaml.dump(data, f, sort_keys=False, default_flow_style=False)
+
+    @classmethod
+    def from_yaml(
+        cls, path: Union[str, Path], tool_registry: Optional[ToolRegistry] = None
+    ) -> "BaseAgent":
+        """
+        Load agent from YAML configuration.
+
+        Args:
+            path: Path to YAML file.
+            tool_registry: Registry containing *available* tools to re-hydrate the agent.
+                           The agent's tools will be selected from this registry based on names in YAML.
+
+        Returns:
+            Instantiated BaseAgent.
+        """
+        path_obj = Path(path)
+        with open(path_obj, "r", encoding="utf-8") as f:
+            data = yaml.safe_load(f)
+
+        name = data.get("name", "unnamed_agent")
+        system_prompt = data.get("system_prompt", "")
+        retry_attempts = data.get("retry_attempts", 3)
+        provider_config = data.get("provider", {})
+        model_name = provider_config.get("model", "qwen3-coder:480b-cloud")
+
+        # Handle persistence parameter
+        # If persistence key exists and is True, use_persistence is True
+        # If persistence key exists and is False, use_persistence is False
+        # If persistence key doesn't exist, use_persistence defaults to True
+        use_persistence = data.get("persistence", True)
+
+        # Reconstruct tools
+        tools = []
+        tool_names = data.get("tools", [])
+
+        # Use provided registry or fallback to default
+        registry = tool_registry
+        if registry is None:
+            # Lazy import to avoid circular dependencies if any
+            try:
+                from kader.tools import get_default_registry
+
+                registry = get_default_registry()
+            except ImportError:
+                pass
+
+        if tool_names and registry:
+            for t_name in tool_names:
+                t = registry.get(t_name)
+                if t:
+                    tools.append(t)
+
+        return cls(
+            name=name,
+            system_prompt=system_prompt,
+            tools=tools,
+            retry_attempts=retry_attempts,
+            model_name=model_name,
+            use_persistence=use_persistence,
+        )